From e5863fd619825cd3f850602e816608cff311afc1 Mon Sep 17 00:00:00 2001
From: Jing Zhang <jizha@amd.com>
Date: Sun, 5 Nov 2023 02:49:02 +0000
Subject: [PATCH] fixed layout

---
 composable_kernel/CHANGELOG.md => CHANGELOG.md      |   0
 composable_kernel/CITATION.cff => CITATION.cff      |   0
 composable_kernel/CMakeLists.txt => CMakeLists.txt  |   0
 .../CONTRIBUTORS.md => CONTRIBUTORS.md              |   0
 .../Config.cmake.in => Config.cmake.in              |   0
 composable_kernel/Dockerfile => Dockerfile          |   0
 composable_kernel/Jenkinsfile => Jenkinsfile        |   0
 composable_kernel/LICENSE => LICENSE                |   0
 composable_kernel/README.md => README.md            |   0
 .../01_gemm/CMakeLists.txt                          |   0
 .../01_gemm/gemm.cpp                                |   0
 .../02_gemm_add_add_fastgelu/CMakeLists.txt         |   0
 .../gemm_add_add_fastgelu.cpp                       |   0
 .../gemm_add_add_fastgelu_generic.cpp               |   0
 .../02_gemm_add_add_fastgelu/gemm_add_fastgelu.cpp  |   0
 .../gemm_add_fastgelu_generic.cpp                   |   0
 .../02_gemm_add_add_fastgelu/gemm_fastgelu.cpp      |   0
 .../gemm_fastgelu_generic.cpp                       |   0
 .../03_gemm_layernorm/CMakeLists.txt                |   0
 .../gemm_add_add_layernorm_naive.cpp                |   0
 .../gemm_add_relu_add_layernorm_welford.cpp         |   0
 .../04_contraction/CMakeLists.txt                   |   0
 .../04_contraction/contraction_bilinear_fp32.cpp    |   0
 .../04_contraction/contraction_bilinear_fp64.cpp    |   0
 .../contraction_g1m2n3k1_add_xdl_fp16.cpp           |   0
 .../04_contraction/contraction_scale_fp32.cpp       |   0
 .../04_contraction/contraction_scale_fp64.cpp       |   0
 .../05_layernorm/CMakeLists.txt                     |   0
 .../05_layernorm/layernorm2d.cpp                    |   0
 .../06_softmax/CMakeLists.txt                       |   0
 .../06_softmax/softmax4d.cpp                        |   0
 .../07_grouped_convnd_fwd/CMakeLists.txt            |   0
 .../07_grouped_convnd_fwd/grouped_conv1d_fwd.cpp    |   0
 .../07_grouped_convnd_fwd/grouped_conv2d_fwd.cpp    |   0
 .../08_fused_attention/CMakeLists.txt               |   0
 .../08_fused_attention/fused_attention.cpp          |   0
 .../08_fused_attention/fused_attention_bias.cpp     |   0
 .../09_quantization/CMakeLists.txt                  |   0
 ...conv2d_fwd_bias_relu_perchannel_quantization.cpp |   0
 .../conv2d_fwd_bias_relu_perlayer_quantization.cpp  |   0
 ...conv2d_fwd_bias_tanh_perchannel_quantization.cpp |   0
 .../conv2d_fwd_bias_tanh_perlayer_quantization.cpp  |   0
 .../conv2d_fwd_perchannel_quantization.cpp          |   0
 .../conv2d_fwd_perlayer_quantization.cpp            |   0
 .../09_quantization/gemm_quantization.cpp           |   0
 .../10_grouped_convnd_bwd_data/CMakeLists.txt       |   0
 .../grouped_conv2d_bwd_data.cpp                     |   0
 .../grouped_conv3d_bwd_data.cpp                     |   0
 ...rouped_conv3d_bwd_data_input_fp16_comp_bf8f8.cpp |   0
 .../11_grouped_conv_bwd_weight/CMakeLists.txt       |   0
 .../11_grouped_conv_bwd_weight/common.hpp           |   0
 .../grouped_conv1d_bwd_weight_fp16.cpp              |   0
 .../grouped_conv2d_bwd_weight_fp16.cpp              |   0
 .../grouped_conv3d_bwd_weight_fp16.cpp              |   0
 .../grouped_conv3d_bwd_weight_fp32.cpp              |   0
 .../12_elementwise_normalization/CMakeLists.txt     |   0
 .../elementwise_layernorm2d.cpp                     |   0
 .../13_batchnorm/CMakeLists.txt                     |   0
 .../13_batchnorm/batchnorm_bwd_nhwc.cpp             |   0
 .../13_batchnorm/batchnorm_fwd_nhwc.cpp             |   0
 .../13_batchnorm/batchnorm_infer_nhwc.cpp           |   0
 .../14_instance_id/CMakeLists.txt                   |   0
 .../14_instance_id/batchnorm_fwd_instance_id.cpp    |   0
 .../15_convnd_bwd_data/CMakeLists.txt               |   0
 .../15_convnd_bwd_data/common.hpp                   |   0
 .../15_convnd_bwd_data/conv3d_bwd_data_fp16.cpp     |   0
 .../15_convnd_bwd_data/conv3d_bwd_data_fp32.cpp     |   0
 .../15_gemm_add_multiply/CMakeLists.txt             |   0
 .../15_gemm_add_multiply/gemm_add_multiply.cpp      |   0
 .../15_reduce/CMakeLists.txt                        |   0
 .../15_reduce/reduce_nhwc_c.cpp                     |   0
 .../16_convnd_fwd/CMakeLists.txt                    |   0
 .../16_convnd_fwd/common.hpp                        |   0
 .../16_convnd_fwd/conv3d_fwd_fp16.cpp               |   0
 .../16_convnd_fwd/conv3d_fwd_fp16_comp_fp8.cpp      |   0
 .../16_convnd_fwd/conv3d_fwd_fp32.cpp               |   0
 .../17_grouped_gemm_fastgelu/CMakeLists.txt         |   0
 .../grouped_gemm_fastgelu.cpp                       |   0
 .../18_groupnorm/CMakeLists.txt                     |   0
 .../18_groupnorm/groupnorm_swish.cpp                |   0
 .../19_pool/CMakeLists.txt                          |   0
 .../19_pool/avg_pool3d_bwd.cpp                      |   0
 .../19_pool/avg_pool3d_fwd.cpp                      |   0
 .../19_pool/max_pool2d_bwd.cpp                      |   0
 .../19_pool/max_pool2d_fwd.cpp                      |   0
 .../20_splitk_gemm/CMakeLists.txt                   |   0
 .../20_splitk_gemm/splitK_gemm_fp16_f8.cpp          |   0
 .../21_grouped_gemm_bias/CMakeLists.txt             |   0
 .../grouped_gemm_fixed_nk_bias_fp16.cpp             |   0
 .../22_grouped_gemm/CMakeLists.txt                  |   0
 .../22_grouped_gemm/grouped_gemm_fixed_nk_fp16.cpp  |   0
 .../22_grouped_gemm/grouped_gemm_fixed_nk_fp8.cpp   |   0
 .../22_grouped_gemm/grouped_gemm_fixed_nk_i8.cpp    |   0
 .../22_im2col_col2im/CMakeLists.txt                 |   0
 .../22_im2col_col2im/column_to_image.cpp            |   0
 .../22_im2col_col2im/image_to_column.cpp            |   0
 .../CMakeLists.txt                                  |   0
 .../client_example => client_example}/README.md     |   0
 {composable_kernel/cmake => cmake}/Analyzers.cmake  |   0
 {composable_kernel/cmake => cmake}/ClangTidy.cmake  |   0
 {composable_kernel/cmake => cmake}/CppCheck.cmake   |   0
 {composable_kernel/cmake => cmake}/DoxygenDoc.cmake |   0
 .../cmake => cmake}/EnableCompilerWarnings.cmake    |   0
 .../cmake => cmake}/TargetFlags.cmake               |   0
 {composable_kernel/cmake => cmake}/googletest.cmake |   0
 .../dev-requirements.txt => dev-requirements.txt    |   0
 .../docs => docs}/API_Reference_Guide.rst           |   0
 .../docs => docs}/Contributors_Guide.rst            |   0
 .../docs => docs}/Supported_Primitives_Guide.rst    |   0
 {composable_kernel/docs => docs}/conf.py            |   0
 .../docs => docs}/data/ck_component.png             | Bin
 {composable_kernel/docs => docs}/data/ck_layer.png  | Bin
 {composable_kernel/docs => docs}/dockerhub.rst      |   0
 {composable_kernel/docs => docs}/doxygen/Doxyfile   |   0
 {composable_kernel/docs => docs}/index.rst          |   0
 {composable_kernel/docs => docs}/license.rst        |   0
 {composable_kernel/docs => docs}/refs.bib           |   0
 {composable_kernel/docs => docs}/sphinx/_toc.yml.in |   0
 .../docs => docs}/sphinx/requirements.in            |   0
 .../docs => docs}/sphinx/requirements.txt           |   0
 .../docs => docs}/tutorial_hello_world.rst          |   0
 .../example => example}/01_gemm/CMakeLists.txt      |   0
 .../example => example}/01_gemm/README.md           |   0
 .../example => example}/01_gemm/common.hpp          |   0
 .../example => example}/01_gemm/gemm_dl_fp16.cpp    |   0
 .../example => example}/01_gemm/gemm_dl_fp32.cpp    |   0
 .../example => example}/01_gemm/gemm_dl_int4.cpp    |   0
 .../example => example}/01_gemm/gemm_dl_int8.cpp    |   0
 .../example => example}/01_gemm/gemm_dpp_fp16.cpp   |   0
 .../example => example}/01_gemm/gemm_wmma_fp16.cpp  |   0
 .../example => example}/01_gemm/gemm_xdl_bf16.cpp   |   0
 .../01_gemm/gemm_xdl_bf16_rtn.cpp                   |   0
 .../example => example}/01_gemm/gemm_xdl_fp16.cpp   |   0
 .../01_gemm/gemm_xdl_fp16_fp8.cpp                   |   0
 .../example => example}/01_gemm/gemm_xdl_fp64.cpp   |   0
 .../example => example}/01_gemm/gemm_xdl_fp8.cpp    |   0
 .../01_gemm/gemm_xdl_fp8_bf8.cpp                    |   0
 .../example => example}/01_gemm/gemm_xdl_int4.cpp   |   0
 .../example => example}/01_gemm/gemm_xdl_int8.cpp   |   0
 .../01_gemm/gemm_xdl_skip_b_lds_fp16.cpp            |   0
 .../01_gemm/gemm_xdl_streamk.cpp                    |   0
 .../01_gemm/gemm_xdl_wavelet_fp16.cpp               |   0
 .../01_gemm/run_gemm_example.inc                    |   0
 .../02_gemm_bilinear/CMakeLists.txt                 |   0
 .../example => example}/02_gemm_bilinear/README.md  |   0
 .../02_gemm_bilinear/gemm_bilinear_wmma_fp16.cpp    |   0
 .../02_gemm_bilinear/gemm_bilinear_wmma_int8.cpp    |   0
 .../02_gemm_bilinear/gemm_bilinear_xdl_fp16.cpp     |   0
 .../03_gemm_bias_relu/CMakeLists.txt                |   0
 .../example => example}/03_gemm_bias_relu/README.md |   0
 .../03_gemm_bias_relu/gemm_bias_relu_xdl_fp16.cpp   |   0
 .../04_gemm_add_add_fastgelu/CMakeLists.txt         |   0
 .../04_gemm_add_add_fastgelu/README.md              |   0
 .../04_gemm_add_add_fastgelu/common.hpp             |   0
 .../gemm_add_add_fastgelu_xdl_bf16.cpp              |   0
 .../gemm_add_add_fastgelu_xdl_fp16.cpp              |   0
 .../gemm_add_add_fastgelu_xdl_fp32.cpp              |   0
 .../gemm_add_add_fastgelu_xdl_int4.cpp              |   0
 .../gemm_add_add_fastgelu_xdl_int8.cpp              |   0
 .../run_gemm_add_add_fastgelu_example.inc           |   0
 .../09_convnd_fwd/CMakeLists.txt                    |   0
 .../example => example}/09_convnd_fwd/README.md     |   0
 .../09_convnd_fwd/convnd_fwd_common.hpp             |   0
 .../09_convnd_fwd/convnd_fwd_dl_common.hpp          |   0
 .../09_convnd_fwd/convnd_fwd_dl_fp16.cpp            |   0
 .../09_convnd_fwd/convnd_fwd_dl_fp32.cpp            |   0
 .../09_convnd_fwd/convnd_fwd_dl_int8.cpp            |   0
 .../09_convnd_fwd/convnd_fwd_xdl_bf16.cpp           |   0
 .../09_convnd_fwd/convnd_fwd_xdl_fp16.cpp           |   0
 .../09_convnd_fwd/convnd_fwd_xdl_fp32.cpp           |   0
 .../09_convnd_fwd/convnd_fwd_xdl_fp64.cpp           |   0
 .../09_convnd_fwd/convnd_fwd_xdl_int8.cpp           |   0
 .../09_convnd_fwd/run_convnd_fwd_dl_example.inc     |   0
 .../09_convnd_fwd/run_convnd_fwd_example.inc        |   0
 .../CMakeLists.txt                                  |   0
 .../common.hpp                                      |   0
 .../convnd_fwd_max_xdl_bf16.cpp                     |   0
 .../convnd_fwd_max_xdl_fp16.cpp                     |   0
 .../convnd_fwd_max_xdl_fp32.cpp                     |   0
 .../convnd_fwd_max_xdl_int4.cpp                     |   0
 .../convnd_fwd_max_xdl_int8.cpp                     |   0
 .../run_convnd_fwd_max_example.inc                  |   0
 .../example => example}/12_reduce/CMakeLists.txt    |   0
 .../example => example}/12_reduce/README.md         |   0
 .../12_reduce/reduce_blockwise.cpp                  |   0
 .../12_reduce/reduce_blockwise_impl.hpp             |   0
 .../12_reduce/reduce_blockwise_two_call.cpp         |   0
 .../12_reduce/reduce_example_common.hpp             |   0
 .../12_reduce/reduce_multiblock_atomic_add.cpp      |   0
 .../12_reduce/reduce_multiblock_atomic_add_impl.hpp |   0
 .../13_pool2d_fwd/CMakeLists.txt                    |   0
 .../example => example}/13_pool2d_fwd/README.md     |   0
 .../13_pool2d_fwd/pool2d_fwd_common.hpp             |   0
 .../13_pool2d_fwd/pool2d_fwd_fp16.cpp               |   0
 .../13_pool2d_fwd/pool2d_fwd_fp32.cpp               |   0
 .../14_gemm_quantization/CMakeLists.txt             |   0
 .../gemm_dl_quantization_int8.cpp                   |   0
 .../gemm_xdl_bias_relu_quantization_int8.cpp        |   0
 .../gemm_xdl_quantization_int8.cpp                  |   0
 .../15_grouped_gemm/CMakeLists.txt                  |   0
 .../example => example}/15_grouped_gemm/README.md   |   0
 .../grouped_gemm_multiple_d_dl_fp16.cpp             |   0
 .../15_grouped_gemm/grouped_gemm_xdl_bf16.cpp       |   0
 .../grouped_gemm_xdl_fixed_nk_bias_fp16.cpp         |   0
 .../grouped_gemm_xdl_fixed_nk_fp16.cpp              |   0
 .../grouped_gemm_xdl_fixed_nk_fp8.cpp               |   0
 .../15_grouped_gemm/grouped_gemm_xdl_fp16.cpp       |   0
 .../15_grouped_gemm/grouped_gemm_xdl_fp32.cpp       |   0
 .../15_grouped_gemm/grouped_gemm_xdl_int4.cpp       |   0
 .../15_grouped_gemm/grouped_gemm_xdl_int8.cpp       |   0
 .../grouped_gemm_xdl_splitk_fp16.cpp                |   0
 .../15_grouped_gemm/run_grouped_gemm_example.inc    |   0
 .../16_gemm_multi_d_multi_reduces/CMakeLists.txt    |   0
 .../gemm_add_add_mean_meansquare_xdl_fp16.cpp       |   0
 .../gemm_add_addsquare_xdl_int8.cpp                 |   0
 .../gemm_max_xdl_bf16.cpp                           |   0
 .../gemm_max_xdl_fp16.cpp                           |   0
 .../gemm_max_xdl_fp32.cpp                           |   0
 .../gemm_max_xdl_int4.cpp                           |   0
 .../gemm_max_xdl_int8.cpp                           |   0
 .../gemm_mean_meansquare_xdl_bf16.cpp               |   0
 .../gemm_mean_meansquare_xdl_fp16.cpp               |   0
 .../gemm_mean_meansquare_xdl_fp32.cpp               |   0
 .../gemm_reduce_xdl_common.hpp                      |   0
 .../17_convnd_bwd_data/CMakeLists.txt               |   0
 .../17_convnd_bwd_data/README.md                    |   0
 .../17_convnd_bwd_data/convnd_bwd_data_common.hpp   |   0
 .../17_convnd_bwd_data/convnd_bwd_data_dl_fp16.cpp  |   0
 .../17_convnd_bwd_data/convnd_bwd_data_xdl_fp16.cpp |   0
 .../18_batched_gemm_reduce/CMakeLists.txt           |   0
 .../batched_gemm_reduce_xdl_fp16.cpp                |   0
 .../19_binary_elementwise/CMakeLists.txt            |   0
 .../broadcast_add_2d_amn_bn.cpp                     |   0
 .../broadcast_add_3d_am_bmnk.cpp                    |   0
 .../19_binary_elementwise/elementwise_add_1d.cpp    |   0
 .../19_binary_elementwise/elementwise_add_4d.cpp    |   0
 .../20_grouped_conv_bwd_weight/CMakeLists.txt       |   0
 .../20_grouped_conv_bwd_weight/common.hpp           |   0
 .../grouped_conv_bwd_weight_dl_fp16.cpp             |   0
 .../grouped_conv_bwd_weight_wmma_fp16.cpp           |   0
 .../grouped_conv_bwd_weight_xdl_bf16.cpp            |   0
 .../grouped_conv_bwd_weight_xdl_fp16.cpp            |   0
 ...rouped_conv_bwd_weight_xdl_fp16_comp_bf8_fp8.cpp |   0
 .../run_grouped_conv_bwd_weight_example.inc         |   0
 .../21_gemm_layernorm/CMakeLists.txt                |   0
 .../gemm_bias_relu_add_layernorm_xdl_naive_fp16.cpp |   0
 ...emm_bias_relu_add_layernorm_xdl_welford_fp16.cpp |   0
 .../gemm_layernorm_xdl_naive_fp16.cpp               |   0
 .../gemm_xdl_layernorm_naive_single_kernel_fp16.cpp |   0
 .../example => example}/22_cgemm/CMakeLists.txt     |   0
 .../example => example}/22_cgemm/cgemm_xdl_bf16.cpp |   0
 .../22_cgemm/cgemm_xdl_common.hpp                   |   0
 .../example => example}/22_cgemm/cgemm_xdl_fp16.cpp |   0
 .../example => example}/22_cgemm/cgemm_xdl_fp32.cpp |   0
 .../example => example}/22_cgemm/cgemm_xdl_int4.cpp |   0
 .../example => example}/22_cgemm/cgemm_xdl_int8.cpp |   0
 .../example => example}/23_softmax/CMakeLists.txt   |   0
 .../example => example}/23_softmax/README.md        |   0
 .../23_softmax/softmax_blockwise.cpp                |   0
 .../24_batched_gemm/CMakeLists.txt                  |   0
 .../24_batched_gemm/batched_gemm_xdl_bf16.cpp       |   0
 .../24_batched_gemm/batched_gemm_xdl_fp16.cpp       |   0
 .../24_batched_gemm/batched_gemm_xdl_fp32.cpp       |   0
 .../24_batched_gemm/batched_gemm_xdl_int4.cpp       |   0
 .../24_batched_gemm/batched_gemm_xdl_int8.cpp       |   0
 .../24_batched_gemm/run_batched_gemm_example.inc    |   0
 .../25_gemm_bias_e_permute/CMakeLists.txt           |   0
 .../gemm_bias_e_permute_g1m2n3k1_xdl_fp16.cpp       |   0
 .../gemm_bias_e_permute_g1m3n2k1_xdl_fp16.cpp       |   0
 .../26_contraction/CMakeLists.txt                   |   0
 .../example => example}/26_contraction/README.md    |   0
 .../contraction_bilinear_xdl_fp32.cpp               |   0
 .../contraction_bilinear_xdl_fp64.cpp               |   0
 .../26_contraction/contraction_scale_xdl_fp32.cpp   |   0
 .../26_contraction/contraction_scale_xdl_fp64.cpp   |   0
 .../example => example}/27_layernorm/CMakeLists.txt |   0
 .../example => example}/27_layernorm/common.hpp     |   0
 .../27_layernorm/layernorm_fp16.cpp                 |   0
 .../27_layernorm/layernorm_splitk_fp16.cpp          |   0
 .../27_layernorm/run_layernorm_example.inc          |   0
 .../28_grouped_gemm_bias_e_permute/CMakeLists.txt   |   0
 .../grouped_gemm_bias_e_permute_xdl_fp16.cpp        |   0
 .../29_batched_gemm_bias_e_permute/CMakeLists.txt   |   0
 .../batched_gemm_bias_e_permute_wmma_fp16.cpp       |   0
 .../batched_gemm_bias_e_permute_xdl_fp16.cpp        |   0
 .../30_grouped_conv_fwd_multiple_d/CMakeLists.txt   |   0
 .../30_grouped_conv_fwd_multiple_d/README.md        |   0
 .../30_grouped_conv_fwd_multiple_d/common.hpp       |   0
 .../30_grouped_conv_fwd_multiple_d/common_wmma.hpp  |   0
 .../grouped_conv_fwd_bias_relu_add_wmma_fp16.cpp    |   0
 .../grouped_conv_fwd_bias_relu_add_wmma_int8.cpp    |   0
 .../grouped_conv_fwd_bias_relu_add_xdl_bf16.cpp     |   0
 .../grouped_conv_fwd_bias_relu_add_xdl_fp16.cpp     |   0
 .../grouped_conv_fwd_bias_relu_add_xdl_fp32.cpp     |   0
 .../grouped_conv_fwd_bias_relu_add_xdl_int4.cpp     |   0
 .../grouped_conv_fwd_bias_relu_add_xdl_int8.cpp     |   0
 .../grouped_conv_fwd_xdl_fp16.cpp                   |   0
 .../run_grouped_conv_fwd_bias_relu_add_example.inc  |   0
 ..._grouped_conv_fwd_bias_relu_add_wmma_example.inc |   0
 .../run_grouped_conv_fwd_example.inc                |   0
 .../31_batched_gemm_gemm/CMakeLists.txt             |   0
 .../batched_gemm_gemm_xdl_bf16.cpp                  |   0
 .../batched_gemm_gemm_xdl_fp16.cpp                  |   0
 .../batched_gemm_gemm_xdl_fp32.cpp                  |   0
 .../batched_gemm_gemm_xdl_int4.cpp                  |   0
 .../batched_gemm_gemm_xdl_int8.cpp                  |   0
 .../run_batched_gemm_gemm_example.inc               |   0
 .../CMakeLists.txt                                  |   0
 ...triangle_scale_softmax_gemm_permute_xdl_fp16.cpp |   0
 ...hed_gemm_scale_softmax_gemm_permute_xdl_bf16.cpp |   0
 ...hed_gemm_scale_softmax_gemm_permute_xdl_fp16.cpp |   0
 .../batched_gemm_scale_softmax_gemm_xdl_bf16.cpp    |   0
 .../batched_gemm_scale_softmax_gemm_xdl_fp16.cpp    |   0
 ...triangle_scale_softmax_gemm_permute_xdl_fp16.cpp |   0
 ...ped_gemm_scale_softmax_gemm_permute_xdl_fp16.cpp |   0
 .../run_batched_gemm_scale_softmax_gemm.inc         |   0
 .../run_batched_gemm_scale_softmax_gemm_permute.inc |   0
 .../run_grouped_gemm_scale_softmax_gemm_permute.inc |   0
 .../33_multiple_reduce/CMakeLists.txt               |   0
 .../33_multiple_reduce/README.md                    |   0
 .../33_multiple_reduce/dual_reduce_common.hpp       |   0
 .../33_multiple_reduce/dual_reduce_multiblock.cpp   |   0
 .../33_multiple_reduce/dual_reduce_threadwise.cpp   |   0
 .../example => example}/34_batchnorm/CMakeLists.txt |   0
 .../example => example}/34_batchnorm/README.md      |   0
 .../34_batchnorm/batchnorm_backward_nhwc.cpp        |   0
 .../34_batchnorm/batchnorm_common.hpp               |   0
 .../batchnorm_forward_inferring_nhwc.cpp            |   0
 .../batchnorm_forward_training_nhwc.cpp             |   0
 .../batchnorm_forward_training_nhwc_obsolete.cpp    |   0
 .../34_batchnorm/batchnorm_infer_impl.hpp           |   0
 .../35_splitK_gemm/CMakeLists.txt                   |   0
 .../35_splitK_gemm/run_splitK_gemm_example.inc      |   0
 .../35_splitK_gemm/splitK_gemm_xdl_bf16.cpp         |   0
 .../35_splitK_gemm/splitK_gemm_xdl_fp16.cpp         |   0
 .../35_splitK_gemm/splitK_gemm_xdl_fp32.cpp         |   0
 .../35_splitK_gemm/splitK_gemm_xdl_int4.cpp         |   0
 .../35_splitK_gemm/splitK_gemm_xdl_int8.cpp         |   0
 .../36_sparse_embedding/CMakeLists.txt              |   0
 .../sparse_embedding3_forward_layernorm.cpp         |   0
 .../CMakeLists.txt                                  |   0
 .../batched_gemm_add_add_relu_gemm_add_xdl_fp16.cpp |   0
 .../CMakeLists.txt                                  |   0
 .../38_grouped_conv_bwd_data_multiple_d/common.hpp  |   0
 .../grouped_conv_bwd_data_bias_relu_xdl_fp16.cpp    |   0
 .../grouped_conv_bwd_data_wmma_fp16.cpp             |   0
 .../grouped_conv_bwd_data_xdl_fp16.cpp              |   0
 .../run_grouped_conv_bwd_data_bias_relu_example.inc |   0
 .../run_grouped_conv_bwd_data_example.inc           |   0
 .../example => example}/39_permute/CMakeLists.txt   |   0
 .../example => example}/39_permute/common.hpp       |   0
 .../39_permute/permute_1xHxW_fp16.cpp               |   0
 .../39_permute/permute_HxWx4_fp16.cpp               |   0
 .../39_permute/permute_NxHxW_fp16.cpp               |   0
 .../39_permute/run_permute_bundle_example.inc       |   0
 .../39_permute/run_permute_element_example.inc      |   0
 .../40_conv2d_fwd_quantization/CMakeLists.txt       |   0
 .../40_conv2d_fwd_quantization/common.hpp           |   0
 ...wd_dl_bias_relu_perchannel_quantization_int8.cpp |   0
 ..._fwd_dl_bias_relu_perlayer_quantization_int8.cpp |   0
 ...wd_dl_bias_tanh_perchannel_quantization_int8.cpp |   0
 ..._fwd_dl_bias_tanh_perlayer_quantization_int8.cpp |   0
 .../conv2d_fwd_dl_perchannel_quantization_int8.cpp  |   0
 .../conv2d_fwd_dl_perlayer_quantization_int8.cpp    |   0
 ...d_xdl_bias_relu_perchannel_quantization_int8.cpp |   0
 ...fwd_xdl_bias_relu_perlayer_quantization_int8.cpp |   0
 .../conv2d_fwd_xdl_perchannel_quantization_int8.cpp |   0
 .../conv2d_fwd_xdl_perlayer_quantization_int8.cpp   |   0
 ...v2d_fwd_bias_perchannel_quantization_example.inc |   0
 ...onv2d_fwd_bias_perlayer_quantization_example.inc |   0
 ...n_conv2d_fwd_perchannel_quantization_example.inc |   0
 ...run_conv2d_fwd_perlayer_quantization_example.inc |   0
 .../41_grouped_conv_conv_fwd/CMakeLists.txt         |   0
 .../grouped_conv_conv_fwd_xdl_bf16.cpp              |   0
 .../grouped_conv_conv_fwd_xdl_fp16.cpp              |   0
 .../grouped_conv_conv_fwd_xdl_fp32.cpp              |   0
 .../grouped_conv_conv_fwd_xdl_int4.cpp              |   0
 .../grouped_conv_conv_fwd_xdl_int8.cpp              |   0
 .../run_grouped_conv_conv_fwd_example.inc           |   0
 .../example => example}/42_groupnorm/CMakeLists.txt |   0
 .../example => example}/42_groupnorm/common.hpp     |   0
 .../42_groupnorm/groupnorm_sigmoid_mul_fp16.cpp     |   0
 .../42_groupnorm/groupnorm_splitk_fp16.cpp          |   0
 .../42_groupnorm/groupnorm_swish_fp16.cpp           |   0
 .../42_groupnorm/run_groupnorm_example.inc          |   0
 .../43_splitk_gemm_bias_e_permute/CMakeLists.txt    |   0
 .../splitk_gemm_bias_e_permute_xdl_fp16.cpp         |   0
 .../splitk_gemm_bias_e_permute_xdl_fp32.cpp         |   0
 .../44_elementwise_permute/CMakeLists.txt           |   0
 .../elementwise_permute_4D_fp16.cpp                 |   0
 .../elementwise_permute_4D_fp16_2d.cpp              |   0
 .../45_elementwise_normalization/CMakeLists.txt     |   0
 .../elementwise_layernorm_blockwise.cpp             |   0
 .../46_gemm_add_multiply/CMakeLists.txt             |   0
 .../46_gemm_add_multiply/README.md                  |   0
 .../46_gemm_add_multiply/common.hpp                 |   0
 .../gemm_add_multiply_dl_fp16.cpp                   |   0
 .../gemm_add_multiply_xdl_fp16.cpp                  |   0
 .../run_gemm_add_multiply_example.inc               |   0
 .../CMakeLists.txt                                  |   0
 .../gemm_bias_softmax_gemm_permute.cpp              |   0
 .../48_pool3d_fwd/CMakeLists.txt                    |   0
 .../48_pool3d_fwd/pool3d_fwd_common.hpp             |   0
 .../48_pool3d_fwd/pool3d_fwd_fp16.cpp               |   0
 .../49_maxpool2d_bwd/CMakeLists.txt                 |   0
 .../49_maxpool2d_bwd/maxpool2d_bwd_bf16.cpp         |   0
 .../49_maxpool2d_bwd/maxpool2d_bwd_common.hpp       |   0
 .../49_maxpool2d_bwd/maxpool2d_bwd_fp16.cpp         |   0
 .../49_maxpool2d_bwd/maxpool2d_bwd_fp32.cpp         |   0
 .../50_put_element/CMakeLists.txt                   |   0
 .../50_put_element/put_element_fp16.cpp             |   0
 .../51_avgpool3d_bwd/CMakeLists.txt                 |   0
 .../51_avgpool3d_bwd/avgpool3d_bwd_bf16.cpp         |   0
 .../51_avgpool3d_bwd/avgpool3d_bwd_common.hpp       |   0
 .../51_avgpool3d_bwd/avgpool3d_bwd_fp16.cpp         |   0
 .../51_avgpool3d_bwd/avgpool3d_bwd_fp32.cpp         |   0
 .../52_im2col_col2im/CMakeLists.txt                 |   0
 .../52_im2col_col2im/column_to_image_f32.cpp        |   0
 .../example => example}/52_im2col_col2im/common.hpp |   0
 .../52_im2col_col2im/image_to_column_f32.cpp        |   0
 .../53_gemv_splitk/CMakeLists.txt                   |   0
 .../example => example}/53_gemv_splitk/README.md    |   0
 .../example => example}/53_gemv_splitk/common.hpp   |   0
 .../53_gemv_splitk/gemv_splitk_fp16.cpp             |   0
 .../53_gemv_splitk/run_gemv_splitk_example.inc      |   0
 .../54_tall_and_skinny_gemm_splitk/CMakeLists.txt   |   0
 .../54_tall_and_skinny_gemm_splitk/README.md        |   0
 .../54_tall_and_skinny_gemm_splitk/common.hpp       |   0
 .../run_tall_and_skinny_gemm_splitk_example.inc     |   0
 .../tall_and_skinny_gemm_splitk_fp16.cpp            |   0
 .../60_gemm_multi_ABD/CMakeLists.txt                |   0
 .../60_gemm_multi_ABD/gemm_multi_ABD_xdl_fp16.cpp   |   0
 .../61_contraction_multi_ABD/CMakeLists.txt         |   0
 .../contraction_multi_ABD_xdl_fp16.cpp              |   0
 .../62_conv_fwd_activ/CMakeLists.txt                |   0
 .../62_conv_fwd_activ/convnd_fwd_activ_common.hpp   |   0
 .../62_conv_fwd_activ/convnd_fwd_xdl_abs_fp16.cpp   |   0
 .../convnd_fwd_xdl_clippedrelu_fp16.cpp             |   0
 .../62_conv_fwd_activ/convnd_fwd_xdl_elu_fp16.cpp   |   0
 .../convnd_fwd_xdl_leakyrelu_fp16.cpp               |   0
 .../62_conv_fwd_activ/convnd_fwd_xdl_pow_fp16.cpp   |   0
 .../62_conv_fwd_activ/convnd_fwd_xdl_relu_fp16.cpp  |   0
 .../convnd_fwd_xdl_sigmoid_fp16.cpp                 |   0
 .../convnd_fwd_xdl_softrelu_fp16.cpp                |   0
 .../62_conv_fwd_activ/convnd_fwd_xdl_tanh_fp16.cpp  |   0
 .../run_convnd_fwd_activ_example.inc                |   0
 .../example => example}/CMakeLists.txt              |   0
 {composable_kernel/include => include}/ck/ck.hpp    |   0
 .../include => include}/ck/config.h.in              |   0
 .../ck/host_utility/device_prop.hpp                 |   0
 .../ck/host_utility/hip_check_error.hpp             |   0
 .../include => include}/ck/host_utility/io.hpp      |   0
 .../ck/host_utility/kernel_launch.hpp               |   0
 .../ck/host_utility/stream_utility.hpp              |   0
 ...olution3d_into_gemm_v4r4r4_ndhwc_kzyxc_ndhwk.hpp |   0
 .../include => include}/ck/stream_config.hpp        |   0
 .../include => include}/ck/tensor/static_tensor.hpp |   0
 .../ck/tensor_description/cluster_descriptor.hpp    |   0
 .../ck/tensor_description/multi_index_transform.hpp |   0
 .../multi_index_transform_helper.hpp                |   0
 .../ck/tensor_description/tensor_adaptor.hpp        |   0
 .../ck/tensor_description/tensor_descriptor.hpp     |   0
 .../tensor_description/tensor_descriptor_helper.hpp |   0
 .../tensor_space_filling_curve.hpp                  |   0
 .../gpu/block/blockwise_gemm_dl_v2r3.hpp            |   0
 .../gpu/block/blockwise_gemm_dlops_v2r2.hpp         |   0
 .../gpu/block/blockwise_gemm_dlops_v3.hpp           |   0
 .../gpu/block/blockwise_gemm_dpp.hpp                |   0
 .../gpu/block/blockwise_gemm_wmma.hpp               |   0
 .../gpu/block/blockwise_gemm_xdlops.hpp             |   0
 .../gpu/block/blockwise_gemm_xdlops_skip_b_lds.hpp  |   0
 .../gpu/block/blockwise_softmax.hpp                 |   0
 .../block/blockwise_tensor_slice_transfer_v5r1.hpp  |   0
 .../gpu/block/blockwise_welford.hpp                 |   0
 .../gpu/block/reduction_functions_blockwise.hpp     |   0
 .../thread_group_tensor_slice_transfer_v4r1.hpp     |   0
 .../thread_group_tensor_slice_transfer_v6r1.hpp     |   0
 .../thread_group_tensor_slice_transfer_v6r1r2.hpp   |   0
 .../thread_group_tensor_slice_transfer_v6r2.hpp     |   0
 .../thread_group_tensor_slice_transfer_v6r3.hpp     |   0
 .../block/thread_group_tensor_slice_transfer_v7.hpp |   0
 .../thread_group_tensor_slice_transfer_v7r2.hpp     |   0
 .../gpu/device/conv_tensor_rearrange_op.hpp         |   0
 .../convolution_backward_data_specialization.hpp    |   0
 .../convolution_backward_weight_specialization.hpp  |   0
 .../device/convolution_forward_specialization.hpp   |   0
 .../gpu/device/device_avgpool_bwd.hpp               |   0
 .../ck/tensor_operation/gpu/device/device_base.hpp  |   0
 .../device_batched_contraction_multiple_d.hpp       |   0
 .../gpu/device/device_batched_gemm.hpp              |   0
 .../gpu/device/device_batched_gemm_e_permute.hpp    |   0
 .../gpu/device/device_batched_gemm_gemm.hpp         |   0
 .../gpu/device/device_batched_gemm_multi_d.hpp      |   0
 ...vice_batched_gemm_multiple_d_gemm_multiple_d.hpp |   0
 .../gpu/device/device_batched_gemm_softmax_gemm.hpp |   0
 .../device_batched_gemm_softmax_gemm_permute.hpp    |   0
 .../gpu/device/device_batchnorm_backward.hpp        |   0
 .../gpu/device/device_batchnorm_forward.hpp         |   0
 .../gpu/device/device_batchnorm_infer.hpp           |   0
 .../ck/tensor_operation/gpu/device/device_cgemm.hpp |   0
 .../gpu/device/device_contraction_multiple_abd.hpp  |   0
 .../gpu/device/device_contraction_multiple_d.hpp    |   0
 .../gpu/device/device_conv_bwd_data.hpp             |   0
 .../tensor_operation/gpu/device/device_conv_fwd.hpp |   0
 .../gpu/device/device_conv_fwd_bias_activation.hpp  |   0
 .../device/device_conv_fwd_bias_activation_add.hpp  |   0
 .../gpu/device/device_conv_tensor_rearrange.hpp     |   0
 .../gpu/device/device_elementwise.hpp               |   0
 .../gpu/device/device_elementwise_normalization.hpp |   0
 .../ck/tensor_operation/gpu/device/device_gemm.hpp  |   0
 .../gpu/device/device_gemm_bias_e_permute.hpp       |   0
 .../gpu/device/device_gemm_multiple_abd.hpp         |   0
 .../gpu/device/device_gemm_multiple_d.hpp           |   0
 .../gpu/device/device_gemm_multiple_d_layernorm.hpp |   0
 .../device/device_gemm_multiple_d_multiple_r.hpp    |   0
 .../gpu/device/device_gemm_reduce.hpp               |   0
 .../gpu/device/device_gemm_splitk.hpp               |   0
 .../gpu/device/device_gemm_streamk.hpp              |   0
 .../device_grouped_contraction_multiple_d.hpp       |   0
 .../device_grouped_conv_bwd_data_multiple_d.hpp     |   0
 .../gpu/device/device_grouped_conv_bwd_weight.hpp   |   0
 .../gpu/device/device_grouped_conv_fwd.hpp          |   0
 .../device/device_grouped_conv_fwd_multiple_d.hpp   |   0
 .../gpu/device/device_grouped_gemm.hpp              |   0
 .../gpu/device/device_grouped_gemm_fixed_nk.hpp     |   0
 .../device_grouped_gemm_softmax_gemm_permute.hpp    |   0
 .../gpu/device/device_grouped_gemm_splitk.hpp       |   0
 .../gpu/device/device_max_pool_bwd.hpp              |   0
 .../gpu/device/device_multiple_reduce.hpp           |   0
 .../gpu/device/device_normalization.hpp             |   0
 .../tensor_operation/gpu/device/device_permute.hpp  |   0
 .../tensor_operation/gpu/device/device_pool_fwd.hpp |   0
 .../gpu/device/device_put_element.hpp               |   0
 .../tensor_operation/gpu/device/device_reduce.hpp   |   0
 .../tensor_operation/gpu/device/device_softmax.hpp  |   0
 .../device/device_splitk_contraction_multiple_d.hpp |   0
 .../gpu/device/device_tall_and_skinny_gemm.hpp      |   0
 .../gpu/device/gemm_specialization.hpp              |   0
 .../impl/device_avgpool3d_bwd_ndhwc_ndhwc.hpp       |   0
 ...batched_contraction_multiple_d_wmma_cshuffle.hpp |   0
 ..._batched_contraction_multiple_d_xdl_cshuffle.hpp |   0
 .../impl/device_batched_gemm_e_permute_xdl.hpp      |   0
 .../impl/device_batched_gemm_gemm_xdl_cshuffle.hpp  |   0
 .../device/impl/device_batched_gemm_multi_d_xdl.hpp |   0
 .../impl/device_batched_gemm_multiple_d_dl.hpp      |   0
 ...gemm_multiple_d_gemm_multiple_d_xdl_cshuffle.hpp |   0
 .../device_batched_gemm_reduce_xdl_cshuffle.hpp     |   0
 ...tched_gemm_softmax_gemm_permute_xdl_cshuffle.hpp |   0
 ...evice_batched_gemm_softmax_gemm_xdl_cshuffle.hpp |   0
 .../gpu/device/impl/device_batched_gemm_xdl.hpp     |   0
 .../device/impl/device_batchnorm_backward_impl.hpp  |   0
 .../device/impl/device_batchnorm_forward_impl.hpp   |   0
 .../impl/device_batchnorm_forward_impl_obsolete.hpp |   0
 .../device/impl/device_cgemm_4gemm_xdl_cshuffle.hpp |   0
 .../gpu/device/impl/device_column_to_image_impl.hpp |   0
 ...device_contraction_multiple_abd_xdl_cshuffle.hpp |   0
 .../device_contraction_multiple_d_xdl_cshuffle.hpp  |   0
 ...backward_weight_xdl_c_shuffle_nhwc_kyxc_nhwk.hpp |   0
 .../device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk.hpp   |   0
 ...c_shuffle_bias_activation_add_nhwc_kyxc_nhwk.hpp |   0
 ...xdl_c_shuffle_bias_activation_nhwc_kyxc_nhwk.hpp |   0
 ...vice_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk.hpp |   0
 .../impl/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk.hpp   |   0
 .../device_conv3d_fwd_naive_ndhwc_kzyxc_ndhwk.hpp   |   0
 .../device_conv3d_fwd_xdl_ndhwc_kzyxc_ndhwk.hpp     |   0
 .../impl/device_convnd_bwd_data_nwc_kxc_nwk_dl.hpp  |   0
 .../impl/device_convnd_bwd_data_nwc_kxc_nwk_xdl.hpp |   0
 .../gpu/device/impl/device_elementwise_2d_impl.hpp  |   0
 .../gpu/device/impl/device_elementwise_impl.hpp     |   0
 .../impl/device_elementwise_normalization_impl.hpp  |   0
 .../device_gemm_bias_add_reduce_xdl_cshuffle.hpp    |   0
 .../gpu/device/impl/device_gemm_dl.hpp              |   0
 .../gpu/device/impl/device_gemm_dpp.hpp             |   0
 .../impl/device_gemm_multiple_abd_xdl_cshuffle.hpp  |   0
 .../gpu/device/impl/device_gemm_multiple_d_dl.hpp   |   0
 ...evice_gemm_multiple_d_layernorm_xdl_cshuffle.hpp |   0
 ...vice_gemm_multiple_d_multiple_r_xdl_cshuffle.hpp |   0
 .../impl/device_gemm_multiple_d_wmma_cshuffle.hpp   |   0
 .../impl/device_gemm_multiple_d_xdl_cshuffle.hpp    |   0
 .../device/impl/device_gemm_reduce_xdl_cshuffle.hpp |   0
 .../gpu/device/impl/device_gemm_wmma.hpp            |   0
 .../gpu/device/impl/device_gemm_xdl.hpp             |   0
 .../gpu/device/impl/device_gemm_xdl_cshuffle.hpp    |   0
 .../impl/device_gemm_xdl_layernorm_cshuffle.hpp     |   0
 .../gpu/device/impl/device_gemm_xdl_skip_b_lds.hpp  |   0
 .../impl/device_gemm_xdl_splitk_c_shuffle.hpp       |   0
 .../gpu/device/impl/device_gemm_xdl_streamk.hpp     |   0
 .../impl/device_gemm_xdl_waveletmodel_cshuffle.hpp  |   0
 ..._grouped_contraction_multiple_d_xdl_cshuffle.hpp |   0
 ...ouped_conv_bwd_data_multiple_d_wmma_cshuffle.hpp |   0
 ...ped_conv_bwd_data_multiple_d_xdl_cshuffle_v1.hpp |   0
 .../impl/device_grouped_conv_bwd_weight_dl.hpp      |   0
 ...device_grouped_conv_bwd_weight_wmma_cshuffle.hpp |   0
 .../device_grouped_conv_bwd_weight_xdl_cshuffle.hpp |   0
 ...rouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp |   0
 .../device_grouped_conv_fwd_dl_nhwc_kyxc_nhwk.hpp   |   0
 ...evice_grouped_conv_fwd_multiple_d_multiple_r.hpp |   0
 ..._conv_fwd_multiple_d_multiple_r_xdl_cshuffle.hpp |   0
 ...ce_grouped_conv_fwd_multiple_d_wmma_cshuffle.hpp |   0
 ...ice_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp |   0
 .../gpu/device/impl/device_grouped_conv_utils.hpp   |   0
 .../impl/device_grouped_gemm_multiple_d_dl.hpp      |   0
 ...ouped_gemm_softmax_gemm_permute_xdl_cshuffle.hpp |   0
 .../gpu/device/impl/device_grouped_gemm_xdl.hpp     |   0
 .../impl/device_grouped_gemm_xdl_fixed_nk.hpp       |   0
 .../device_grouped_gemm_xdl_splitk_cshuffle.hpp     |   0
 .../gpu/device/impl/device_image_to_column_impl.hpp |   0
 .../gpu/device/impl/device_max_pool_bwd_impl.hpp    |   0
 .../impl/device_multiple_reduce_multiblock.hpp      |   0
 .../impl/device_multiple_reduce_threadwise.hpp      |   0
 .../gpu/device/impl/device_normalization_impl.hpp   |   0
 .../impl/device_normalization_splitk_impl.hpp       |   0
 .../gpu/device/impl/device_permute_impl.hpp         |   0
 .../gpu/device/impl/device_pool2d_fwd_nhwc_nhwc.hpp |   0
 .../device/impl/device_pool3d_fwd_ndhwc_ndhwc.hpp   |   0
 .../gpu/device/impl/device_put_element_impl.hpp     |   0
 .../gpu/device/impl/device_reduce_common.hpp        |   0
 .../gpu/device/impl/device_reduce_multiblock.hpp    |   0
 .../gpu/device/impl/device_reduce_threadwise.hpp    |   0
 .../gpu/device/impl/device_softmax_impl.hpp         |   0
 .../device_sparse_embeddings_forward_layernorm.hpp  |   0
 ...e_splitk_contraction_multiple_d_xdl_cshuffle.hpp |   0
 .../impl/device_tall_and_skinny_gemm_splitk.hpp     |   0
 .../gpu/device/masking_specialization.hpp           |   0
 .../tensor_operation/gpu/device/matrix_padder.hpp   |   0
 .../gpu/device/reduction_operator_mapping.hpp       |   0
 .../tensor_operation/gpu/device/tensor_layout.hpp   |   0
 .../gpu/device/tensor_specialization.hpp            |   0
 .../tensor_operation/gpu/device/welford_helper.hpp  |   0
 .../gpu/element/binary_element_wise_operation.hpp   |   0
 .../gpu/element/element_wise_operation.hpp          |   0
 .../gpu/element/quantization_operation.hpp          |   0
 .../gpu/element/unary_element_wise_operation.hpp    |   0
 .../gridwise_multiblock_batchnorm_forward.hpp       |   0
 ..._reduce_second_half_batchnorm_backward_final.hpp |   0
 .../gridwise_multiblock_welford_first_half.hpp      |   0
 ...second_half_batchnorm_forward_final_obsolete.hpp |   0
 ...ord_second_half_multiblock_reduce_first_half.hpp |   0
 .../gpu/grid/block_to_ctile_map.hpp                 |   0
 ...m_multiple_d_welford_first_half_xdl_cshuffle.hpp |   0
 .../gridwise_welford_second_half_layernorm2d.hpp    |   0
 .../gridwise_2d_multiple_reduction_multiblock.hpp   |   0
 .../gridwise_2d_multiple_reduction_threadwise.hpp   |   0
 .../gpu/grid/gridwise_2d_reduction_multiblock.hpp   |   0
 .../gpu/grid/gridwise_2d_reduction_threadwise.hpp   |   0
 .../gridwise_batched_gemm_gemm_xdl_cshuffle_v1.hpp  |   0
 ...m_multiple_d_gemm_multiple_d_xdl_cshuffle_v1.hpp |   0
 ...gemm_multiple_d_softmax_gemm_xdl_cshuffle_v1.hpp |   0
 ...se_batched_gemm_softmax_gemm_xdl_cshuffle_v1.hpp |   0
 ...ridwise_batchnorm_backward_blockwise_welford.hpp |   0
 ...gridwise_batchnorm_forward_blockwise_welford.hpp |   0
 .../gpu/grid/gridwise_elementwise_1d.hpp            |   0
 .../gpu/grid/gridwise_elementwise_2d.hpp            |   0
 ...dwise_elementwise_layernorm_welford_variance.hpp |   0
 ...ridwise_gemm_bias_add_reduce_xdl_cshuffle_v1.hpp |   0
 .../gpu/grid/gridwise_gemm_dl_multiple_d.hpp        |   0
 .../gpu/grid/gridwise_gemm_dl_v1r3.hpp              |   0
 .../tensor_operation/gpu/grid/gridwise_gemm_dpp.hpp |   0
 .../gridwise_gemm_multiple_abd_xdl_cshuffle.hpp     |   0
 ...wise_gemm_multiple_d_multiple_r_xdl_cshuffle.hpp |   0
 .../grid/gridwise_gemm_multiple_d_wmma_cshuffle.hpp |   0
 .../grid/gridwise_gemm_multiple_d_xdl_cshuffle.hpp  |   0
 ...gridwise_gemm_multiple_d_xdl_splitk_cshuffle.hpp |   0
 .../gpu/grid/gridwise_gemm_pipeline_selector.hpp    |   0
 .../gpu/grid/gridwise_gemm_pipeline_v1.hpp          |   0
 .../gpu/grid/gridwise_gemm_pipeline_v2.hpp          |   0
 .../gpu/grid/gridwise_gemm_pipeline_v3.hpp          |   0
 .../grid/gridwise_gemm_reduce_xdl_cshuffle_v1.hpp   |   0
 ...ridwise_gemm_split_k_multiple_d_xdl_cshuffle.hpp |   0
 ...wise_gemm_split_k_multiple_d_xdl_cshuffle_v2.hpp |   0
 .../gpu/grid/gridwise_gemm_waveletmodel.hpp         |   0
 .../gpu/grid/gridwise_gemm_wmma.hpp                 |   0
 .../gpu/grid/gridwise_gemm_xdl_cshuffle_v1.hpp      |   0
 .../gridwise_gemm_xdl_layernorm_cshuffle_v1.hpp     |   0
 .../gridwise_gemm_xdl_waveletmodel_cshuffle.hpp     |   0
 .../gpu/grid/gridwise_gemm_xdlops_bwd_weight.hpp    |   0
 .../gpu/grid/gridwise_gemm_xdlops_skip_b_lds_v1.hpp |   0
 .../gpu/grid/gridwise_gemm_xdlops_streamk.hpp       |   0
 .../gpu/grid/gridwise_gemm_xdlops_v2r3.hpp          |   0
 .../gpu/grid/gridwise_gemm_xdlops_v2r4.hpp          |   0
 .../gpu/grid/gridwise_gemm_xdlops_v2r4r2.hpp        |   0
 .../gpu/grid/gridwise_gemm_xdlops_v3r1.hpp          |   0
 .../gpu/grid/gridwise_gemm_xdlops_v3r2.hpp          |   0
 .../gpu/grid/gridwise_gemm_xdlops_v3r3.hpp          |   0
 .../gpu/grid/gridwise_gemv_splitk.hpp               |   0
 .../tensor_operation/gpu/grid/gridwise_permute.hpp  |   0
 .../gpu/grid/gridwise_put_element_1d.hpp            |   0
 .../gpu/grid/gridwise_set_buffer_value.hpp          |   0
 .../gpu/grid/gridwise_set_multiple_buffer_value.hpp |   0
 .../tensor_operation/gpu/grid/gridwise_softmax.hpp  |   0
 ...gridwise_sparse_embeddings_forward_layernorm.hpp |   0
 .../grid/gridwise_tall_and_skinny_gemm_splitk.hpp   |   0
 .../gpu/grid/gridwise_tensor_rearrange.hpp          |   0
 .../gridwise_normalization_naive_variance.hpp       |   0
 .../gridwise_normalization_selector.hpp             |   0
 .../gridwise_normalization_splitk_1st.hpp           |   0
 .../gridwise_normalization_splitk_2nd.hpp           |   0
 .../gridwise_normalization_welford_variance.hpp     |   0
 .../gpu/thread/reduction_functions_threadwise.hpp   |   0
 .../gpu/thread/threadwise_contraction_dl.hpp        |   0
 .../gpu/thread/threadwise_gemm_dlops_v3.hpp         |   0
 .../gpu/thread/threadwise_tensor_slice_set.hpp      |   0
 .../gpu/thread/threadwise_tensor_slice_transfer.hpp |   0
 .../threadwise_tensor_slice_transfer_v3r1.hpp       |   0
 .../threadwise_tensor_slice_transfer_v4r1.hpp       |   0
 .../threadwise_tensor_slice_transfer_v5r1.hpp       |   0
 .../threadwise_tensor_slice_transfer_v6r1.hpp       |   0
 .../threadwise_tensor_slice_transfer_v6r1r2.hpp     |   0
 .../threadwise_tensor_slice_transfer_v6r2.hpp       |   0
 .../threadwise_tensor_slice_transfer_v6r3.hpp       |   0
 .../thread/threadwise_tensor_slice_transfer_v7.hpp  |   0
 .../threadwise_tensor_slice_transfer_v7r2.hpp       |   0
 .../gpu/thread/threadwise_welford.hpp               |   0
 .../ck/tensor_operation/gpu/warp/dpp_gemm.hpp       |   0
 .../ck/tensor_operation/gpu/warp/wmma_gemm.hpp      |   0
 .../ck/tensor_operation/gpu/warp/xdlops_gemm.hpp    |   0
 .../transform_contraction_to_gemm.hpp               |   0
 .../transform_conv_bwd_data_to_gemm_v1.hpp          |   0
 .../transform_conv_fwd_to_gemm.hpp                  |   0
 .../ck/utility/amd_address_space.hpp                |   0
 .../ck/utility/amd_buffer_addressing.hpp            |   0
 .../include => include}/ck/utility/amd_gemm_dpp.hpp |   0
 .../ck/utility/amd_inline_asm.hpp                   |   0
 .../ck/utility/amd_wave_read_first_lane.hpp         |   0
 .../include => include}/ck/utility/amd_wmma.hpp     |   0
 .../include => include}/ck/utility/amd_xdlops.hpp   |   0
 .../include => include}/ck/utility/array.hpp        |   0
 .../ck/utility/array_multi_index.hpp                |   0
 .../ck/utility/c_style_pointer_cast.hpp             |   0
 .../ck/utility/common_header.hpp                    |   0
 .../ck/utility/container_element_picker.hpp         |   0
 .../ck/utility/container_helper.hpp                 |   0
 .../include => include}/ck/utility/data_type.hpp    |   0
 .../include => include}/ck/utility/debug.hpp        |   0
 .../ck/utility/dynamic_buffer.hpp                   |   0
 .../include => include}/ck/utility/enable_if.hpp    |   0
 .../include => include}/ck/utility/f8_utils.hpp     |   0
 .../include => include}/ck/utility/functional.hpp   |   0
 .../include => include}/ck/utility/functional2.hpp  |   0
 .../include => include}/ck/utility/functional3.hpp  |   0
 .../include => include}/ck/utility/functional4.hpp  |   0
 .../ck/utility/generic_memory_space_atomic.hpp      |   0
 .../include => include}/ck/utility/get_id.hpp       |   0
 .../include => include}/ck/utility/get_shift.hpp    |   0
 .../include => include}/ck/utility/ignore.hpp       |   0
 .../ck/utility/inner_product.hpp                    |   0
 .../ck/utility/inner_product_dpp8.hpp               |   0
 .../ck/utility/integral_constant.hpp                |   0
 .../include => include}/ck/utility/is_detected.hpp  |   0
 .../ck/utility/is_known_at_compile_time.hpp         |   0
 .../ck/utility/loop_scheduler.hpp                   |   0
 .../ck/utility/magic_division.hpp                   |   0
 .../include => include}/ck/utility/math.hpp         |   0
 .../include => include}/ck/utility/math_v2.hpp      |   0
 .../include => include}/ck/utility/multi_index.hpp  |   0
 .../include => include}/ck/utility/number.hpp       |   0
 .../include => include}/ck/utility/random_gen.hpp   |   0
 .../ck/utility/reduction_common.hpp                 |   0
 .../ck/utility/reduction_enums.hpp                  |   0
 .../ck/utility/reduction_functions_accumulate.hpp   |   0
 .../ck/utility/reduction_operator.hpp               |   0
 .../include => include}/ck/utility/sequence.hpp     |   0
 .../ck/utility/sequence_helper.hpp                  |   0
 .../include => include}/ck/utility/span.hpp         |   0
 .../ck/utility/static_buffer.hpp                    |   0
 .../ck/utility/statically_indexed_array.hpp         |   0
 .../statically_indexed_array_multi_index.hpp        |   0
 .../ck/utility/synchronization.hpp                  |   0
 .../include => include}/ck/utility/thread_group.hpp |   0
 .../ck/utility/transpose_vectors.hpp                |   0
 .../include => include}/ck/utility/tuple.hpp        |   0
 .../include => include}/ck/utility/tuple_helper.hpp |   0
 .../include => include}/ck/utility/type.hpp         |   0
 .../include => include}/ck/utility/type_convert.hpp |   0
 .../ck/utility/workgroup_barrier.hpp                |   0
 .../ck/utility/workgroup_synchronization.hpp        |   0
 .../include => include}/ck/version.h.in             |   0
 .../library => library}/CMakeLists.txt              |   0
 .../cpu/reference_avgpool_bwd.hpp                   |   0
 .../cpu/reference_batched_gemm.hpp                  |   0
 .../cpu/reference_batchnorm_backward.hpp            |   0
 .../cpu/reference_batchnorm_forward.hpp             |   0
 .../cpu/reference_batchnorm_infer.hpp               |   0
 .../cpu/reference_cgemm.hpp                         |   0
 .../cpu/reference_column_to_image.hpp               |   0
 .../cpu/reference_contraction.hpp                   |   0
 .../cpu/reference_conv_bwd_data.hpp                 |   0
 .../cpu/reference_conv_bwd_weight.hpp               |   0
 .../cpu/reference_conv_fwd.hpp                      |   0
 .../cpu/reference_conv_fwd_bias_activation.hpp      |   0
 .../cpu/reference_conv_fwd_bias_activation_add.hpp  |   0
 .../cpu/reference_gemm.hpp                          |   0
 .../cpu/reference_gemm_layernorm.hpp                |   0
 .../cpu/reference_groupnorm.hpp                     |   0
 .../cpu/reference_image_to_column.hpp               |   0
 .../cpu/reference_layernorm.hpp                     |   0
 .../cpu/reference_maxpool_bwd.hpp                   |   0
 .../cpu/reference_pool_fwd.hpp                      |   0
 .../cpu/reference_reduce.hpp                        |   0
 .../cpu/reference_softmax.hpp                       |   0
 ...eference_sparse_embedding3_forward_layernorm.hpp |   0
 .../gpu/naive_conv_fwd.hpp                          |   0
 .../add_device_operation_instance.hpp               |   0
 .../device_operation_instance_factory.hpp           |   0
 .../gpu/avg_pool3d_bwd.hpp                          |   0
 .../tensor_operation_instance/gpu/batched_gemm.hpp  |   0
 .../gpu/batched_gemm_add_relu_gemm_add.hpp          |   0
 .../gpu/batched_gemm_bias_permute.hpp               |   0
 .../gpu/batched_gemm_bias_softmax_gemm_permute.hpp  |   0
 .../gpu/batched_gemm_gemm.hpp                       |   0
 .../gpu/batched_gemm_multi_d.hpp                    |   0
 .../gpu/batched_gemm_softmax_gemm.hpp               |   0
 .../gpu/batched_gemm_softmax_gemm_permute.hpp       |   0
 .../gpu/batchnorm_backward.hpp                      |   0
 .../gpu/batchnorm_forward.hpp                       |   0
 .../gpu/batchnorm_infer.hpp                         |   0
 .../gpu/contraction_bilinear.hpp                    |   0
 .../gpu/contraction_scale.hpp                       |   0
 .../gpu/conv_tensor_rearrange.hpp                   |   0
 .../device_column_to_image_instance.hpp             |   0
 .../device_image_to_column_instance.hpp             |   0
 .../gpu/convolution_backward_data.hpp               |   0
 .../gpu/convolution_forward.hpp                     |   0
 .../gpu/device_elementwise_instance.hpp             |   0
 .../gpu/device_gemm_mean_squaremean_instance.hpp    |   0
 .../gpu/elementwise_normalization.hpp               |   0
 .../library/tensor_operation_instance/gpu/gemm.hpp  |   0
 .../gpu/gemm_add_add_fastgelu.hpp                   |   0
 .../gpu/gemm_add_fastgelu.hpp                       |   0
 .../gpu/gemm_add_multiply.hpp                       |   0
 .../gpu/gemm_add_relu_add_layernorm.hpp             |   0
 .../tensor_operation_instance/gpu/gemm_bilinear.hpp |   0
 .../tensor_operation_instance/gpu/gemm_fastgelu.hpp |   0
 .../gpu/gemm_multiply_add.hpp                       |   0
 .../tensor_operation_instance/gpu/gemm_splitk.hpp   |   0
 .../tensor_operation_instance/gpu/gemm_streamk.hpp  |   0
 .../tensor_operation_instance/gpu/gemv_splitk.hpp   |   0
 .../device_grouped_conv_bwd_data_wmma_instance.hpp  |   0
 .../device_grouped_conv_bwd_data_xdl_instance.hpp   |   0
 .../device_grouped_conv_bwd_weight_dl_instance.hpp  |   0
 ...device_grouped_conv_bwd_weight_wmma_instance.hpp |   0
 .../device_grouped_conv_bwd_weight_xdl_instance.hpp |   0
 .../device_grouped_conv2d_fwd_wmma_instance.hpp     |   0
 .../device_grouped_conv_fwd_dl_instance.hpp         |   0
 .../device_grouped_conv_fwd_wmma_instance.hpp       |   0
 .../device_grouped_conv_fwd_xdl_instance.hpp        |   0
 .../gpu/grouped_convolution_backward_data.hpp       |   0
 .../gpu/grouped_convolution_backward_weight.hpp     |   0
 .../gpu/grouped_convolution_forward.hpp             |   0
 .../tensor_operation_instance/gpu/grouped_gemm.hpp  |   0
 .../gpu/grouped_gemm_bias.hpp                       |   0
 .../gpu/grouped_gemm_fastgelu.hpp                   |   0
 .../gpu/grouped_gemm_fixed_nk.hpp                   |   0
 .../gpu/image_to_column.hpp                         |   0
 .../tensor_operation_instance/gpu/max_pool_bwd.hpp  |   0
 .../tensor_operation_instance/gpu/normalization.hpp |   0
 .../gpu/normalization_swish.hpp                     |   0
 .../tensor_operation_instance/gpu/pool3d_fwd.hpp    |   0
 .../gpu/quantization/gemm_quantization.hpp          |   0
 ...olution_bias_forward_perchannel_quantization.hpp |   0
 ...nvolution_bias_forward_perlayer_quantization.hpp |   0
 ..._convolution_forward_perchannel_quantization.hpp |   0
 ...ed_convolution_forward_perlayer_quantization.hpp |   0
 .../gpu/reduce/device_reduce_instance.hpp           |   0
 .../gpu/reduce/device_reduce_instance_blockwise.hpp |   0
 ...ce_reduce_instance_blockwise_b16_f32_b16_add.hpp |   0
 ...e_reduce_instance_blockwise_b16_f32_b16_amax.hpp |   0
 ...ce_reduce_instance_blockwise_b16_f32_b16_avg.hpp |   0
 ...ce_reduce_instance_blockwise_b16_f32_b16_max.hpp |   0
 ...ce_reduce_instance_blockwise_b16_f32_b16_min.hpp |   0
 ..._reduce_instance_blockwise_b16_f32_b16_norm2.hpp |   0
 ...e_reduce_instance_blockwise_f16_f16_f16_amax.hpp |   0
 ...ce_reduce_instance_blockwise_f16_f16_f16_max.hpp |   0
 ...ce_reduce_instance_blockwise_f16_f16_f16_min.hpp |   0
 ...ce_reduce_instance_blockwise_f16_f32_f16_add.hpp |   0
 ...ce_reduce_instance_blockwise_f16_f32_f16_avg.hpp |   0
 ..._reduce_instance_blockwise_f16_f32_f16_norm2.hpp |   0
 ...ce_reduce_instance_blockwise_f32_f32_f32_add.hpp |   0
 ...e_reduce_instance_blockwise_f32_f32_f32_amax.hpp |   0
 ...ce_reduce_instance_blockwise_f32_f32_f32_avg.hpp |   0
 ...ce_reduce_instance_blockwise_f32_f32_f32_max.hpp |   0
 ...ce_reduce_instance_blockwise_f32_f32_f32_min.hpp |   0
 ..._reduce_instance_blockwise_f32_f32_f32_norm2.hpp |   0
 ...ce_reduce_instance_blockwise_f32_f64_f32_add.hpp |   0
 ...ce_reduce_instance_blockwise_f32_f64_f32_avg.hpp |   0
 ..._reduce_instance_blockwise_f32_f64_f32_norm2.hpp |   0
 ...ce_reduce_instance_blockwise_f64_f64_f64_add.hpp |   0
 ...e_reduce_instance_blockwise_f64_f64_f64_amax.hpp |   0
 ...ce_reduce_instance_blockwise_f64_f64_f64_avg.hpp |   0
 ...ce_reduce_instance_blockwise_f64_f64_f64_max.hpp |   0
 ...ce_reduce_instance_blockwise_f64_f64_f64_min.hpp |   0
 ..._reduce_instance_blockwise_f64_f64_f64_norm2.hpp |   0
 ...vice_reduce_instance_blockwise_i8_i32_i8_add.hpp |   0
 ...vice_reduce_instance_blockwise_i8_i32_i8_avg.hpp |   0
 ...vice_reduce_instance_blockwise_i8_i8_i8_amax.hpp |   0
 ...evice_reduce_instance_blockwise_i8_i8_i8_max.hpp |   0
 ...evice_reduce_instance_blockwise_i8_i8_i8_min.hpp |   0
 .../reduce/device_reduce_instance_impl_common.hpp   |   0
 ...device_reduce_instance_multiblock_atomic_add.hpp |   0
 ...stance_multiblock_atomic_add_b16_f32_f32_add.hpp |   0
 ...stance_multiblock_atomic_add_b16_f32_f32_avg.hpp |   0
 ...stance_multiblock_atomic_add_f16_f32_f32_add.hpp |   0
 ...stance_multiblock_atomic_add_f16_f32_f32_avg.hpp |   0
 ...stance_multiblock_atomic_add_f32_f32_f32_add.hpp |   0
 ...stance_multiblock_atomic_add_f32_f32_f32_avg.hpp |   0
 ...stance_multiblock_atomic_add_f32_f64_f32_add.hpp |   0
 ...stance_multiblock_atomic_add_f32_f64_f32_avg.hpp |   0
 ...stance_multiblock_atomic_add_f64_f64_f64_add.hpp |   0
 ...stance_multiblock_atomic_add_f64_f64_f64_avg.hpp |   0
 .../reduce/device_reduce_instance_threadwise.hpp    |   0
 ...e_reduce_instance_threadwise_b16_f32_b16_add.hpp |   0
 ..._reduce_instance_threadwise_b16_f32_b16_amax.hpp |   0
 ...e_reduce_instance_threadwise_b16_f32_b16_avg.hpp |   0
 ...e_reduce_instance_threadwise_b16_f32_b16_max.hpp |   0
 ...e_reduce_instance_threadwise_b16_f32_b16_min.hpp |   0
 ...reduce_instance_threadwise_b16_f32_b16_norm2.hpp |   0
 ..._reduce_instance_threadwise_f16_f16_f16_amax.hpp |   0
 ...e_reduce_instance_threadwise_f16_f16_f16_max.hpp |   0
 ...e_reduce_instance_threadwise_f16_f16_f16_min.hpp |   0
 ...e_reduce_instance_threadwise_f16_f32_f16_add.hpp |   0
 ...e_reduce_instance_threadwise_f16_f32_f16_avg.hpp |   0
 ...reduce_instance_threadwise_f16_f32_f16_norm2.hpp |   0
 ...e_reduce_instance_threadwise_f32_f32_f32_add.hpp |   0
 ..._reduce_instance_threadwise_f32_f32_f32_amax.hpp |   0
 ...e_reduce_instance_threadwise_f32_f32_f32_avg.hpp |   0
 ...e_reduce_instance_threadwise_f32_f32_f32_max.hpp |   0
 ...e_reduce_instance_threadwise_f32_f32_f32_min.hpp |   0
 ...reduce_instance_threadwise_f32_f32_f32_norm2.hpp |   0
 ...e_reduce_instance_threadwise_f32_f64_f32_add.hpp |   0
 ...e_reduce_instance_threadwise_f32_f64_f32_avg.hpp |   0
 ...reduce_instance_threadwise_f32_f64_f32_norm2.hpp |   0
 ...e_reduce_instance_threadwise_f64_f64_f64_add.hpp |   0
 ..._reduce_instance_threadwise_f64_f64_f64_amax.hpp |   0
 ...e_reduce_instance_threadwise_f64_f64_f64_avg.hpp |   0
 ...e_reduce_instance_threadwise_f64_f64_f64_max.hpp |   0
 ...e_reduce_instance_threadwise_f64_f64_f64_min.hpp |   0
 ...reduce_instance_threadwise_f64_f64_f64_norm2.hpp |   0
 ...ice_reduce_instance_threadwise_i8_i32_i8_add.hpp |   0
 ...ice_reduce_instance_threadwise_i8_i32_i8_avg.hpp |   0
 ...ice_reduce_instance_threadwise_i8_i8_i8_amax.hpp |   0
 ...vice_reduce_instance_threadwise_i8_i8_i8_max.hpp |   0
 ...vice_reduce_instance_threadwise_i8_i8_i8_min.hpp |   0
 .../tensor_operation_instance/gpu/reduce/reduce.hpp |   0
 .../tensor_operation_instance/gpu/softmax.hpp       |   0
 ...evice_softmax_f16_f16_instance_rank3_reduce1.hpp |   0
 ...evice_softmax_f16_f16_instance_rank3_reduce2.hpp |   0
 ...evice_softmax_f16_f16_instance_rank3_reduce3.hpp |   0
 ...evice_softmax_f16_f16_instance_rank4_reduce1.hpp |   0
 ...evice_softmax_f16_f16_instance_rank4_reduce2.hpp |   0
 ...evice_softmax_f16_f16_instance_rank4_reduce3.hpp |   0
 ...evice_softmax_f16_f16_instance_rank4_reduce4.hpp |   0
 .../device_softmax_f16_f16_instance_type.hpp        |   0
 ...evice_softmax_f32_f32_instance_rank3_reduce1.hpp |   0
 ...evice_softmax_f32_f32_instance_rank3_reduce2.hpp |   0
 ...evice_softmax_f32_f32_instance_rank3_reduce3.hpp |   0
 ...evice_softmax_f32_f32_instance_rank4_reduce1.hpp |   0
 ...evice_softmax_f32_f32_instance_rank4_reduce2.hpp |   0
 ...evice_softmax_f32_f32_instance_rank4_reduce3.hpp |   0
 ...evice_softmax_f32_f32_instance_rank4_reduce4.hpp |   0
 .../device_softmax_f32_f32_instance_type.hpp        |   0
 .../gpu/softmax/device_softmax_instance.hpp         |   0
 .../gpu/tall_and_skinny_gemm_splitk.hpp             |   0
 .../include/ck/library/utility/algorithm.hpp        |   0
 .../include/ck/library/utility/check_err.hpp        |   0
 .../include/ck/library/utility/conv_common.hpp      |   0
 .../convolution_host_tensor_descriptor_helper.hpp   |   0
 .../ck/library/utility/convolution_parameter.hpp    |   0
 .../include/ck/library/utility/device_memory.hpp    |   0
 .../include/ck/library/utility/fill.hpp             |   0
 .../include/ck/library/utility/host_common_util.hpp |   0
 .../include/ck/library/utility/host_gemm.hpp        |   0
 .../include/ck/library/utility/host_tensor.hpp      |   0
 .../ck/library/utility/host_tensor_generator.hpp    |   0
 .../include/ck/library/utility/iterator.hpp         |   0
 .../include/ck/library/utility/literals.hpp         |   0
 .../include/ck/library/utility/numeric.hpp          |   0
 .../include/ck/library/utility/ranges.hpp           |   0
 .../tensor_operation_instance/gpu/CMakeLists.txt    |   0
 .../gpu/avg_pool3d_bwd/CMakeLists.txt               |   0
 .../avg_pool3d_bwd_ndhwc_instance_common.hpp        |   0
 .../device_avg_pool3d_bwd_ndhwc_bf16_instance.cpp   |   0
 .../device_avg_pool3d_bwd_ndhwc_f16_instance.cpp    |   0
 .../device_avg_pool3d_bwd_ndhwc_f32_instance.cpp    |   0
 .../gpu/batched_gemm/CMakeLists.txt                 |   0
 ...gemm_xdl_bf16_bf16_bf16_gkm_gkn_gmn_instance.cpp |   0
 ...gemm_xdl_bf16_bf16_bf16_gkm_gnk_gmn_instance.cpp |   0
 ...gemm_xdl_bf16_bf16_bf16_gmk_gkn_gmn_instance.cpp |   0
 ...gemm_xdl_bf16_bf16_bf16_gmk_gnk_gmn_instance.cpp |   0
 ...ed_gemm_xdl_f16_f16_f16_gkm_gkn_gmn_instance.cpp |   0
 ...ed_gemm_xdl_f16_f16_f16_gkm_gnk_gmn_instance.cpp |   0
 ...ed_gemm_xdl_f16_f16_f16_gmk_gkn_gmn_instance.cpp |   0
 ...ed_gemm_xdl_f16_f16_f16_gmk_gnk_gmn_instance.cpp |   0
 ...ed_gemm_xdl_f32_f32_f32_gkm_gkn_gmn_instance.cpp |   0
 ...ed_gemm_xdl_f32_f32_f32_gkm_gnk_gmn_instance.cpp |   0
 ...ed_gemm_xdl_f32_f32_f32_gmk_gkn_gmn_instance.cpp |   0
 ...ed_gemm_xdl_f32_f32_f32_gmk_gnk_gmn_instance.cpp |   0
 ...gemm_xdl_int8_int8_int8_gkm_gkn_gmn_instance.cpp |   0
 ...gemm_xdl_int8_int8_int8_gkm_gnk_gmn_instance.cpp |   0
 ...gemm_xdl_int8_int8_int8_gmk_gkn_gmn_instance.cpp |   0
 ...gemm_xdl_int8_int8_int8_gmk_gnk_gmn_instance.cpp |   0
 .../batched_gemm_add_relu_gemm_add/CMakeLists.txt   |   0
 ...fle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp |   0
 ...fle_f16_f16_f16_f16_gmk_gnk_gon_gmo_instance.cpp |   0
 .../gpu/batched_gemm_bias_permute/CMakeLists.txt    |   0
 ...n3_k1_xdl_c_shuffle_f16_f16_f16_f16_instance.cpp |   0
 .../gpu/batched_gemm_gemm/CMakeLists.txt            |   0
 ...fle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp |   0
 ...fle_f16_f16_f16_f16_gmk_gnk_gon_gmo_instance.cpp |   0
 .../gpu/batched_gemm_multi_d/CMakeLists.txt         |   0
 ..._multi_d_dl_f16_f16_f16_gkm_gkn_gmn_instance.cpp |   0
 ...l_f16_f16_f16_gkm_gkn_gmn_irregular_instance.cpp |   0
 ..._multi_d_dl_f16_f16_f16_gkm_gnk_gmn_instance.cpp |   0
 ...l_f16_f16_f16_gkm_gnk_gmn_irregular_instance.cpp |   0
 ..._multi_d_dl_f16_f16_f16_gmk_gkn_gmn_instance.cpp |   0
 ...l_f16_f16_f16_gmk_gkn_gmn_irregular_instance.cpp |   0
 ..._multi_d_dl_f16_f16_f16_gmk_gnk_gmn_instance.cpp |   0
 ...l_f16_f16_f16_gmk_gnk_gmn_irregular_instance.cpp |   0
 ...emm_multi_d_dl_i8_i8_i8_gkm_gkn_gmn_instance.cpp |   0
 ...d_dl_i8_i8_i8_gkm_gkn_gmn_irregular_instance.cpp |   0
 ...emm_multi_d_dl_i8_i8_i8_gkm_gnk_gmn_instance.cpp |   0
 ...d_dl_i8_i8_i8_gkm_gnk_gmn_irregular_instance.cpp |   0
 ...emm_multi_d_dl_i8_i8_i8_gmk_gkn_gmn_instance.cpp |   0
 ...d_dl_i8_i8_i8_gmk_gkn_gmn_irregular_instance.cpp |   0
 ...emm_multi_d_dl_i8_i8_i8_gmk_gnk_gmn_instance.cpp |   0
 ...d_dl_i8_i8_i8_gmk_gnk_gmn_irregular_instance.cpp |   0
 .../gpu/batched_gemm_reduce/CMakeLists.txt          |   0
 ...fle_f16_f16_f16_f32_f32_gkm_gkn_gmn_instance.cpp |   0
 ...fle_f16_f16_f16_f32_f32_gkm_gnk_gmn_instance.cpp |   0
 ...fle_f16_f16_f16_f32_f32_gmk_gkn_gmn_instance.cpp |   0
 ...fle_f16_f16_f16_f32_f32_gmk_gnk_gmn_instance.cpp |   0
 .../gpu/batched_gemm_softmax_gemm/CMakeLists.txt    |   0
 ...fle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp |   0
 .../CMakeLists.txt                                  |   0
 ...bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instance.cpp |   0
 ...fle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp |   0
 ...bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instance.cpp |   0
 ...fle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp |   0
 .../gpu/batchnorm/CMakeLists.txt                    |   0
 .../device_batchnorm_backward_bf16_instance.cpp     |   0
 .../device_batchnorm_backward_f16_instance.cpp      |   0
 .../device_batchnorm_backward_f32_instance.cpp      |   0
 .../device_batchnorm_backward_f64_instance.cpp      |   0
 .../device_batchnorm_forward_bf16_instance.cpp      |   0
 .../device_batchnorm_forward_f16_instance.cpp       |   0
 .../device_batchnorm_forward_f32_instance.cpp       |   0
 .../device_batchnorm_forward_f64_instance.cpp       |   0
 .../device_batchnorm_infer_bf16_instance.cpp        |   0
 .../device_batchnorm_infer_f16_instance.cpp         |   0
 .../device_batchnorm_infer_f32_instance.cpp         |   0
 .../device_batchnorm_infer_f64_instance.cpp         |   0
 .../gpu/column_to_image/CMakeLists.txt              |   0
 .../device_column_to_image_nhwc_1d_instance.cpp     |   0
 .../device_column_to_image_nhwc_2d_instance.cpp     |   0
 .../device_column_to_image_nhwc_3d_instance.cpp     |   0
 .../gpu/contraction_bilinear/CMakeLists.txt         |   0
 ..._xdl_c_shuffle_f32_f32_f32_f32_kknn_instance.cpp |   0
 ..._xdl_c_shuffle_f32_f32_f32_f32_knnn_instance.cpp |   0
 ..._xdl_c_shuffle_f32_f32_f32_f32_mknn_instance.cpp |   0
 ..._xdl_c_shuffle_f32_f32_f32_f32_mnnn_instance.cpp |   0
 ..._xdl_c_shuffle_f64_f64_f64_f64_kknn_instance.cpp |   0
 ..._xdl_c_shuffle_f64_f64_f64_f64_knnn_instance.cpp |   0
 ..._xdl_c_shuffle_f64_f64_f64_f64_mknn_instance.cpp |   0
 ..._xdl_c_shuffle_f64_f64_f64_f64_mnnn_instance.cpp |   0
 .../gpu/contraction_scale/CMakeLists.txt            |   0
 ...n2_k2_xdl_c_shuffle_f32_f32_f32_kkn_instance.cpp |   0
 ...n2_k2_xdl_c_shuffle_f32_f32_f32_knn_instance.cpp |   0
 ...n2_k2_xdl_c_shuffle_f32_f32_f32_mkn_instance.cpp |   0
 ...n2_k2_xdl_c_shuffle_f32_f32_f32_mnn_instance.cpp |   0
 ...n2_k2_xdl_c_shuffle_f64_f64_f64_kkn_instance.cpp |   0
 ...n2_k2_xdl_c_shuffle_f64_f64_f64_knn_instance.cpp |   0
 ...n2_k2_xdl_c_shuffle_f64_f64_f64_mkn_instance.cpp |   0
 ...n2_k2_xdl_c_shuffle_f64_f64_f64_mnn_instance.cpp |   0
 .../gpu/conv1d_bwd_data/CMakeLists.txt              |   0
 ...onv1d_bwd_data_xdl_nwc_kxc_nwk_bf16_instance.cpp |   0
 ...conv1d_bwd_data_xdl_nwc_kxc_nwk_f16_instance.cpp |   0
 ...conv1d_bwd_data_xdl_nwc_kxc_nwk_f32_instance.cpp |   0
 ...onv1d_bwd_data_xdl_nwc_kxc_nwk_int8_instance.cpp |   0
 .../gpu/conv2d_bwd_data/CMakeLists.txt              |   0
 ...nv2d_bwd_data_dl_nhwc_kyxc_nhwk_f16_instance.cpp |   0
 ...nv2d_bwd_data_dl_nhwc_kyxc_nhwk_f32_instance.cpp |   0
 ...v2d_bwd_data_dl_nhwc_kyxc_nhwk_int8_instance.cpp |   0
 ...2d_bwd_data_xdl_nhwc_kyxc_nhwk_bf16_instance.cpp |   0
 ...v2d_bwd_data_xdl_nhwc_kyxc_nhwk_f16_instance.cpp |   0
 ...v2d_bwd_data_xdl_nhwc_kyxc_nhwk_f32_instance.cpp |   0
 ...2d_bwd_data_xdl_nhwc_kyxc_nhwk_int8_instance.cpp |   0
 .../gpu/conv2d_fwd/CMakeLists.txt                   |   0
 ...wd_xdl_c_shuffle_nhwc_kyxc_nhwk_f16_instance.cpp |   0
 ..._conv2d_fwd_xdl_nhwc_kyxc_nhwk_bf16_instance.cpp |   0
 ...e_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f16_instance.cpp |   0
 ...e_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f32_instance.cpp |   0
 ..._conv2d_fwd_xdl_nhwc_kyxc_nhwk_int8_instance.cpp |   0
 .../gpu/conv2d_fwd_bias_relu/CMakeLists.txt         |   0
 ...huffle_bias_relu_nhwc_kyxc_nhwk_f16_instance.cpp |   0
 .../gpu/conv2d_fwd_bias_relu_add/CMakeLists.txt     |   0
 ...le_bias_relu_add_nhwc_kyxc_nhwk_f16_instance.cpp |   0
 .../gpu/conv3d_bwd_data/CMakeLists.txt              |   0
 ...bwd_data_xdl_ndhwc_kzyxc_ndhwk_bf16_instance.cpp |   0
 ..._bwd_data_xdl_ndhwc_kzyxc_ndhwk_f16_instance.cpp |   0
 ..._bwd_data_xdl_ndhwc_kzyxc_ndhwk_f32_instance.cpp |   0
 ...bwd_data_xdl_ndhwc_kzyxc_ndhwk_int8_instance.cpp |   0
 .../gpu/elementwise/CMakeLists.txt                  |   0
 .../gpu/elementwise/device_normalize_instance.cpp   |   0
 .../gpu/elementwise_normalization/CMakeLists.txt    |   0
 ...evice_elementwise_normalization_f16_instance.cpp |   0
 .../gpu/gemm/CMakeLists.txt                         |   0
 ...device_gemm_dl_f16_f16_f16_km_kn_mn_instance.cpp |   0
 ...m_dl_f16_f16_f16_km_kn_mn_irregular_instance.cpp |   0
 ...device_gemm_dl_f16_f16_f16_km_nk_mn_instance.cpp |   0
 ...m_dl_f16_f16_f16_km_nk_mn_irregular_instance.cpp |   0
 ...device_gemm_dl_f16_f16_f16_mk_kn_mn_instance.cpp |   0
 ...m_dl_f16_f16_f16_mk_kn_mn_irregular_instance.cpp |   0
 ...device_gemm_dl_f16_f16_f16_mk_nk_mn_instance.cpp |   0
 ...m_dl_f16_f16_f16_mk_nk_mn_irregular_instance.cpp |   0
 ...device_gemm_dl_f32_f32_f32_km_kn_mn_instance.cpp |   0
 ...device_gemm_dl_f32_f32_f32_km_nk_mn_instance.cpp |   0
 ...device_gemm_dl_f32_f32_f32_mk_kn_mn_instance.cpp |   0
 ...device_gemm_dl_f32_f32_f32_mk_nk_mn_instance.cpp |   0
 .../device_gemm_dl_i8_i8_i8_km_kn_mn_instance.cpp   |   0
 ...gemm_dl_i8_i8_i8_km_kn_mn_irregular_instance.cpp |   0
 .../device_gemm_dl_i8_i8_i8_km_nk_mn_instance.cpp   |   0
 ...gemm_dl_i8_i8_i8_km_nk_mn_irregular_instance.cpp |   0
 .../device_gemm_dl_i8_i8_i8_mk_kn_mn_instance.cpp   |   0
 ...gemm_dl_i8_i8_i8_mk_kn_mn_irregular_instance.cpp |   0
 .../device_gemm_dl_i8_i8_i8_mk_nk_mn_instance.cpp   |   0
 ...gemm_dl_i8_i8_i8_mk_nk_mn_irregular_instance.cpp |   0
 ...evice_gemm_dpp_f16_f16_f16_km_kn_mn_instance.cpp |   0
 ..._dpp_f16_f16_f16_km_kn_mn_irregular_instance.cpp |   0
 ...evice_gemm_dpp_f16_f16_f16_km_nk_mn_instance.cpp |   0
 ..._dpp_f16_f16_f16_km_nk_mn_irregular_instance.cpp |   0
 ...evice_gemm_dpp_f16_f16_f16_mk_kn_mn_instance.cpp |   0
 ..._dpp_f16_f16_f16_mk_kn_mn_irregular_instance.cpp |   0
 ...evice_gemm_dpp_f16_f16_f16_mk_nk_mn_instance.cpp |   0
 ..._dpp_f16_f16_f16_mk_nk_mn_irregular_instance.cpp |   0
 ...huffle_2_stage_f16_f16_f16_mk_nk_mn_instance.cpp |   0
 ...l_c_shuffle_bf16_bf16_bf16_km_kn_mn_instance.cpp |   0
 ...l_c_shuffle_bf16_bf16_bf16_km_nk_mn_instance.cpp |   0
 ...l_c_shuffle_bf16_bf16_bf16_mk_kn_mn_instance.cpp |   0
 ...l_c_shuffle_bf16_bf16_bf16_mk_nk_mn_instance.cpp |   0
 ..._xdl_c_shuffle_f16_f16_f16_km_kn_mn_instance.cpp |   0
 ..._xdl_c_shuffle_f16_f16_f16_km_nk_mn_instance.cpp |   0
 ..._xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instance.cpp |   0
 ..._xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instance.cpp |   0
 ..._xdl_c_shuffle_f32_f32_f32_km_kn_mn_instance.cpp |   0
 ..._xdl_c_shuffle_f32_f32_f32_km_nk_mn_instance.cpp |   0
 ..._xdl_c_shuffle_f32_f32_f32_mk_kn_mn_instance.cpp |   0
 ..._xdl_c_shuffle_f32_f32_f32_mk_nk_mn_instance.cpp |   0
 ..._xdl_c_shuffle_fp8_fp8_fp8_km_kn_mn_instance.cpp |   0
 ..._xdl_c_shuffle_fp8_fp8_fp8_km_nk_mn_instance.cpp |   0
 ..._xdl_c_shuffle_fp8_fp8_fp8_mk_kn_mn_instance.cpp |   0
 ..._xdl_c_shuffle_fp8_fp8_fp8_mk_nk_mn_instance.cpp |   0
 ...emm_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instance.cpp |   0
 ...emm_xdl_c_shuffle_i8_i8_i8_km_nk_mn_instance.cpp |   0
 ...emm_xdl_c_shuffle_i8_i8_i8_mk_kn_mn_instance.cpp |   0
 ...emm_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instance.cpp |   0
 .../gpu/gemm/device_gemm_xdl_f16_f16_f16/common.hpp |   0
 .../km_kn_mn_add_instance.cpp                       |   0
 .../km_kn_mn_default_pipeline_v1_instance.cpp       |   0
 .../km_kn_mn_default_pipeline_v2_instance.cpp       |   0
 .../km_kn_mn_default_pipeline_v2_opt_instance.cpp   |   0
 .../km_kn_mn_interwave_pipeline_v1_instance.cpp     |   0
 ...kn_mn_irregular_default_pipeline_v1_instance.cpp |   0
 ...kn_mn_irregular_default_pipeline_v2_instance.cpp |   0
 ..._mn_irregular_interwave_pipeline_v1_instance.cpp |   0
 .../km_nk_mn_add_instance.cpp                       |   0
 .../km_nk_mn_default_pipeline_v1_instance.cpp       |   0
 .../km_nk_mn_default_pipeline_v2_instance.cpp       |   0
 .../km_nk_mn_default_pipeline_v2_opt_instance.cpp   |   0
 .../km_nk_mn_interwave_pipeline_v1_instance.cpp     |   0
 ...nk_mn_irregular_default_pipeline_v1_instance.cpp |   0
 ...nk_mn_irregular_default_pipeline_v2_instance.cpp |   0
 ..._mn_irregular_interwave_pipeline_v1_instance.cpp |   0
 .../mk_kn_mn_add_instance.cpp                       |   0
 .../mk_kn_mn_default_pipeline_v1_instance.cpp       |   0
 .../mk_kn_mn_default_pipeline_v2_instance.cpp       |   0
 .../mk_kn_mn_default_pipeline_v2_opt_instance.cpp   |   0
 .../mk_kn_mn_interwave_pipeline_v1_instance.cpp     |   0
 ...kn_mn_irregular_default_pipeline_v1_instance.cpp |   0
 ...kn_mn_irregular_default_pipeline_v2_instance.cpp |   0
 ..._mn_irregular_interwave_pipeline_v1_instance.cpp |   0
 .../mk_nk_mn_add_instance.cpp                       |   0
 .../mk_nk_mn_default_pipeline_v1_instance.cpp       |   0
 .../mk_nk_mn_default_pipeline_v2_instance.cpp       |   0
 .../mk_nk_mn_default_pipeline_v2_opt_instance.cpp   |   0
 .../mk_nk_mn_interwave_pipeline_v1_instance.cpp     |   0
 ...nk_mn_irregular_default_pipeline_v1_instance.cpp |   0
 ...nk_mn_irregular_default_pipeline_v2_instance.cpp |   0
 ..._mn_irregular_interwave_pipeline_v1_instance.cpp |   0
 ...evice_gemm_xdl_f32_f32_f32_km_kn_mn_instance.cpp |   0
 ...evice_gemm_xdl_f32_f32_f32_km_nk_mn_instance.cpp |   0
 ...evice_gemm_xdl_f32_f32_f32_mk_kn_mn_instance.cpp |   0
 ...evice_gemm_xdl_f32_f32_f32_mk_nk_mn_instance.cpp |   0
 ...evice_gemm_xdl_f64_f64_f64_km_kn_mn_instance.cpp |   0
 ...evice_gemm_xdl_f64_f64_f64_km_nk_mn_instance.cpp |   0
 ...evice_gemm_xdl_f64_f64_f64_mk_kn_mn_instance.cpp |   0
 ...evice_gemm_xdl_f64_f64_f64_mk_nk_mn_instance.cpp |   0
 .../gpu/gemm_add_add_fastgelu/CMakeLists.txt        |   0
 ..._f16_f16_f16_f16_f16_km_kn_mn_mn_mn_instance.cpp |   0
 ..._f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instance.cpp |   0
 ..._f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instance.cpp |   0
 ..._f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instance.cpp |   0
 .../gpu/gemm_add_fastgelu/CMakeLists.txt            |   0
 ...shuffle_f16_f16_f16_f16_km_kn_mn_mn_instance.cpp |   0
 ...shuffle_f16_f16_f16_f16_km_nk_mn_mn_instance.cpp |   0
 ...shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instance.cpp |   0
 ...shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instance.cpp |   0
 .../gpu/gemm_add_multiply/CMakeLists.txt            |   0
 ..._f16_f16_f16_f16_f16_km_kn_mn_mn_mn_instance.cpp |   0
 ..._f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instance.cpp |   0
 ..._f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instance.cpp |   0
 ..._f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instance.cpp |   0
 .../gpu/gemm_add_relu_add_layernorm/CMakeLists.txt  |   0
 ...huffle_layernorm_f16_km_kn_mn_mn_mn_instance.cpp |   0
 ...huffle_layernorm_f16_km_nk_mn_mn_mn_instance.cpp |   0
 ...huffle_layernorm_f16_mk_kn_mn_mn_mn_instance.cpp |   0
 ...huffle_layernorm_f16_mk_nk_mn_mn_mn_instance.cpp |   0
 .../gpu/gemm_bias_add_reduce/CMakeLists.txt         |   0
 ...huffle_f16_f16_f16_f32_f32_km_kn_mn_instance.cpp |   0
 ...huffle_f16_f16_f16_f32_f32_km_nk_mn_instance.cpp |   0
 ...huffle_f16_f16_f16_f32_f32_mk_kn_mn_instance.cpp |   0
 ...huffle_f16_f16_f16_f32_f32_mk_nk_mn_instance.cpp |   0
 .../gpu/gemm_bilinear/CMakeLists.txt                |   0
 ...a_c_shuffle_i8_i8_i8_i8_km_kn_mn_mn_instance.cpp |   0
 ...a_c_shuffle_i8_i8_i8_i8_km_nk_mn_mn_instance.cpp |   0
 ...a_c_shuffle_i8_i8_i8_i8_mk_kn_mn_mn_instance.cpp |   0
 ...a_c_shuffle_i8_i8_i8_i8_mk_nk_mn_mn_instance.cpp |   0
 ...shuffle_f16_f16_f16_f16_km_kn_mn_mn_instance.cpp |   0
 ...shuffle_f16_f16_f16_f16_km_nk_mn_mn_instance.cpp |   0
 ...shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instance.cpp |   0
 ...shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instance.cpp |   0
 .../gpu/gemm_fastgelu/CMakeLists.txt                |   0
 ..._xdl_c_shuffle_f16_f16_f16_km_kn_mn_instance.cpp |   0
 ..._xdl_c_shuffle_f16_f16_f16_km_nk_mn_instance.cpp |   0
 ..._xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instance.cpp |   0
 ..._xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instance.cpp |   0
 .../gpu/gemm_multiply_add/CMakeLists.txt            |   0
 ..._f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instance.cpp |   0
 ..._f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instance.cpp |   0
 ...e_f16_f8_f32_f32_f16_mk_kn_mn_mn_mn_instance.cpp |   0
 ...e_f16_f8_f32_f32_f16_mk_nk_mn_mn_mn_instance.cpp |   0
 .../gpu/gemm_reduce/CMakeLists.txt                  |   0
 ...huffle_f16_f16_f16_f32_f32_km_kn_mn_instance.cpp |   0
 ...huffle_f16_f16_f16_f32_f32_km_nk_mn_instance.cpp |   0
 ...huffle_f16_f16_f16_f32_f32_mk_kn_mn_instance.cpp |   0
 ...huffle_f16_f16_f16_f32_f32_mk_nk_mn_instance.cpp |   0
 .../gpu/gemm_splitk/CMakeLists.txt                  |   0
 ...plitk_f16_f16_f16_comp_fp8_km_kn_mn_instance.cpp |   0
 ...plitk_f16_f16_f16_comp_fp8_km_nk_mn_instance.cpp |   0
 ...plitk_f16_f16_f16_comp_fp8_mk_kn_mn_instance.cpp |   0
 ...plitk_f16_f16_f16_comp_fp8_mk_nk_mn_instance.cpp |   0
 ...emm_xdl_splitk_f16_f16_f16_km_kn_mn_instance.cpp |   0
 ...emm_xdl_splitk_f16_f16_f16_km_nk_mn_instance.cpp |   0
 ...emm_xdl_splitk_f16_f16_f16_mk_kn_mn_instance.cpp |   0
 ...emm_xdl_splitk_f16_f16_f16_mk_nk_mn_instance.cpp |   0
 ...emm_xdl_splitk_f16_fp8_f16_km_kn_mn_instance.cpp |   0
 ...emm_xdl_splitk_f16_fp8_f16_km_nk_mn_instance.cpp |   0
 ...emm_xdl_splitk_f16_fp8_f16_mk_kn_mn_instance.cpp |   0
 ...emm_xdl_splitk_f16_fp8_f16_mk_nk_mn_instance.cpp |   0
 ...emm_xdl_splitk_f32_f32_f32_km_kn_mn_instance.cpp |   0
 ...emm_xdl_splitk_f32_f32_f32_km_nk_mn_instance.cpp |   0
 ...emm_xdl_splitk_f32_f32_f32_mk_kn_mn_instance.cpp |   0
 ...emm_xdl_splitk_f32_f32_f32_mk_nk_mn_instance.cpp |   0
 ...emm_xdl_splitk_fp8_f16_f16_km_kn_mn_instance.cpp |   0
 ...emm_xdl_splitk_fp8_f16_f16_km_nk_mn_instance.cpp |   0
 ...emm_xdl_splitk_fp8_f16_f16_mk_kn_mn_instance.cpp |   0
 ...emm_xdl_splitk_fp8_f16_f16_mk_nk_mn_instance.cpp |   0
 .../gpu/gemm_streamk/CMakeLists.txt                 |   0
 ...mm_xdl_streamk_f16_f16_f16_mk_kn_mn_instance.cpp |   0
 .../gpu/gemv_splitk/CMakeLists.txt                  |   0
 ...ce_gemv_splitk_f16_f16_f16_mk_kn_mn_instance.cpp |   0
 ...ce_gemv_splitk_f16_f16_f16_mk_nk_mn_instance.cpp |   0
 .../gpu/grouped_conv1d_bwd_weight/CMakeLists.txt    |   0
 ...d_bwd_weight_dl_gnwc_gkxc_gnwk_bf16_instance.cpp |   0
 ...1d_bwd_weight_dl_gnwc_gkxc_gnwk_f16_instance.cpp |   0
 ...1d_bwd_weight_dl_gnwc_gkxc_gnwk_f32_instance.cpp |   0
 ...d_bwd_weight_dl_nwgc_gkxc_nwgk_bf16_instance.cpp |   0
 ...1d_bwd_weight_dl_nwgc_gkxc_nwgk_f16_instance.cpp |   0
 ...1d_bwd_weight_dl_nwgc_gkxc_nwgk_f32_instance.cpp |   0
 ..._bwd_weight_xdl_gnwc_gkxc_gnwk_bf16_instance.cpp |   0
 ...d_bwd_weight_xdl_gnwc_gkxc_gnwk_f16_instance.cpp |   0
 ...d_bwd_weight_xdl_gnwc_gkxc_gnwk_f32_instance.cpp |   0
 .../gpu/grouped_conv1d_fwd/CMakeLists.txt           |   0
 ..._conv1d_fwd_xdl_gnwc_gkxc_gnwk_bf16_instance.cpp |   0
 ...d_conv1d_fwd_xdl_gnwc_gkxc_gnwk_f16_instance.cpp |   0
 ...d_conv1d_fwd_xdl_gnwc_gkxc_gnwk_f32_instance.cpp |   0
 ..._conv1d_fwd_xdl_gnwc_gkxc_gnwk_int8_instance.cpp |   0
 .../gpu/grouped_conv2d_bwd_data/CMakeLists.txt      |   0
 ..._wmma_gnhwc_gkyxc_gnhwk_f16_1x1s1p0_instance.cpp |   0
 ...bwd_data_wmma_gnhwc_gkyxc_gnhwk_f16_instance.cpp |   0
 ...a_wmma_gnhwc_gkyxc_gnhwk_i8_1x1s1p0_instance.cpp |   0
 ..._bwd_data_wmma_gnhwc_gkyxc_gnhwk_i8_instance.cpp |   0
 ..._wmma_nhwgc_gkyxc_nhwgk_f16_1x1s1p0_instance.cpp |   0
 ...bwd_data_wmma_nhwgc_gkyxc_nhwgk_f16_instance.cpp |   0
 ...a_wmma_nhwgc_gkyxc_nhwgk_i8_1x1s1p0_instance.cpp |   0
 ..._bwd_data_wmma_nhwgc_gkyxc_nhwgk_i8_instance.cpp |   0
 ...bwd_data_xdl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp |   0
 ..._bwd_data_xdl_gnhwc_gkyxc_gnhwk_f16_instance.cpp |   0
 ..._bwd_data_xdl_gnhwc_gkyxc_gnhwk_f32_instance.cpp |   0
 ...bwd_data_xdl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp |   0
 ..._bwd_data_xdl_nhwgc_gkyxc_nhwgk_f16_instance.cpp |   0
 ..._bwd_data_xdl_nhwgc_gkyxc_nhwgk_f32_instance.cpp |   0
 .../gpu/grouped_conv2d_bwd_weight/CMakeLists.txt    |   0
 ...wd_weight_dl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp |   0
 ...bwd_weight_dl_gnhwc_gkyxc_gnhwk_f16_instance.cpp |   0
 ...bwd_weight_dl_gnhwc_gkyxc_gnhwk_f32_instance.cpp |   0
 ...wd_weight_dl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp |   0
 ...bwd_weight_dl_nhwgc_gkyxc_nhwgk_f16_instance.cpp |   0
 ...bwd_weight_dl_nhwgc_gkyxc_nhwgk_f32_instance.cpp |   0
 ...d_weight_xdl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp |   0
 ...wd_weight_xdl_gnhwc_gkyxc_gnhwk_f16_instance.cpp |   0
 ...wd_weight_xdl_gnhwc_gkyxc_gnhwk_f32_instance.cpp |   0
 ...d_weight_xdl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp |   0
 ...wd_weight_xdl_nhwgc_gkyxc_nhwgk_f16_instance.cpp |   0
 ...wd_weight_xdl_nhwgc_gkyxc_nhwgk_f32_instance.cpp |   0
 .../gpu/grouped_conv2d_fwd/CMakeLists.txt           |   0
 ...nv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_f16_instance.cpp |   0
 ...onv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_i8_instance.cpp |   0
 ...conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f16_instance.cpp |   0
 ...conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f32_instance.cpp |   0
 ...conv2d_fwd_dl_nhwgc_gkyxc_nhwgk_f16_instance.cpp |   0
 ...conv2d_fwd_dl_nhwgc_gkyxc_nhwgk_f32_instance.cpp |   0
 ...wd_wmma_gnhwc_gkyxc_gnhwk_f16_1x1p0_instance.cpp |   0
 ..._wmma_gnhwc_gkyxc_gnhwk_f16_1x1s1p0_instance.cpp |   0
 ...nv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_f16_instance.cpp |   0
 ...fwd_wmma_gnhwc_gkyxc_gnhwk_f16_oddc_instance.cpp |   0
 ...fwd_wmma_gnhwc_gkyxc_gnhwk_i8_1x1p0_instance.cpp |   0
 ...d_wmma_gnhwc_gkyxc_gnhwk_i8_1x1s1p0_instance.cpp |   0
 ...onv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_i8_instance.cpp |   0
 ..._fwd_wmma_gnhwc_gkyxc_gnhwk_i8_oddc_instance.cpp |   0
 ...wd_wmma_nhwgc_gkyxc_nhwgk_f16_1x1p0_instance.cpp |   0
 ..._wmma_nhwgc_gkyxc_nhwgk_f16_1x1s1p0_instance.cpp |   0
 ...nv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_f16_instance.cpp |   0
 ...fwd_wmma_nhwgc_gkyxc_nhwgk_f16_oddc_instance.cpp |   0
 ...fwd_wmma_nhwgc_gkyxc_nhwgk_i8_1x1p0_instance.cpp |   0
 ...d_wmma_nhwgc_gkyxc_nhwgk_i8_1x1s1p0_instance.cpp |   0
 ...onv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_i8_instance.cpp |   0
 ..._fwd_wmma_nhwgc_gkyxc_nhwgk_i8_oddc_instance.cpp |   0
 ...nv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp |   0
 ...onv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f16_instance.cpp |   0
 ...onv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f32_instance.cpp |   0
 ...nv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp |   0
 ...onv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_instance.cpp |   0
 ...onv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_instance.cpp |   0
 .../gpu/grouped_conv3d_bwd_data/CMakeLists.txt      |   0
 ...ma_gndhwc_gkzyxc_gndhwk_f16_1x1s1p0_instance.cpp |   0
 ..._data_wmma_gndhwc_gkzyxc_gndhwk_f16_instance.cpp |   0
 ...mma_gndhwc_gkzyxc_gndhwk_i8_1x1s1p0_instance.cpp |   0
 ...d_data_wmma_gndhwc_gkzyxc_gndhwk_i8_instance.cpp |   0
 ...ma_ndhwgc_gkzyxc_ndhwgk_f16_1x1s1p0_instance.cpp |   0
 ..._data_wmma_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp |   0
 ...mma_ndhwgc_gkzyxc_ndhwgk_i8_1x1s1p0_instance.cpp |   0
 ...d_data_wmma_ndhwgc_gkzyxc_ndhwgk_i8_instance.cpp |   0
 ..._data_xdl_gndhwc_gkzyxc_gndhwk_bf16_instance.cpp |   0
 ...d_data_xdl_gndhwc_gkzyxc_gndhwk_f16_instance.cpp |   0
 ...d_data_xdl_gndhwc_gkzyxc_gndhwk_f32_instance.cpp |   0
 ..._data_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp |   0
 ...d_data_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp |   0
 ...d_data_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp |   0
 ...gkzyxc_ndhwgk_input_f16_comp_bf8_f8_instance.cpp |   0
 .../gpu/grouped_conv3d_bwd_weight/CMakeLists.txt    |   0
 ...weight_dl_gndhwc_gkzyxc_gndhwk_bf16_instance.cpp |   0
 ..._weight_dl_gndhwc_gkzyxc_gndhwk_f16_instance.cpp |   0
 ..._weight_dl_gndhwc_gkzyxc_gndhwk_f32_instance.cpp |   0
 ...weight_dl_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp |   0
 ..._weight_dl_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp |   0
 ..._weight_dl_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp |   0
 ...ma_gndhwc_gkzyxc_gndhwk_f16_1x1s1p0_instance.cpp |   0
 ...eight_wmma_gndhwc_gkzyxc_gndhwk_f16_instance.cpp |   0
 ...mma_gndhwc_gkzyxc_gndhwk_i8_1x1s1p0_instance.cpp |   0
 ...weight_wmma_gndhwc_gkzyxc_gndhwk_i8_instance.cpp |   0
 ...ma_ndhwgc_gkzyxc_ndhwgk_f16_1x1s1p0_instance.cpp |   0
 ...eight_wmma_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp |   0
 ...mma_ndhwgc_gkzyxc_ndhwgk_i8_1x1s1p0_instance.cpp |   0
 ...weight_wmma_ndhwgc_gkzyxc_ndhwgk_i8_instance.cpp |   0
 ...eight_xdl_gndhwc_gkzyxc_gndhwk_bf16_instance.cpp |   0
 ...weight_xdl_gndhwc_gkzyxc_gndhwk_f16_instance.cpp |   0
 ...weight_xdl_gndhwc_gkzyxc_gndhwk_f32_instance.cpp |   0
 ...eight_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp |   0
 ...hwgc_gkzyxc_ndhwgk_f16_comp_bf8_fp8_instance.cpp |   0
 ...weight_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp |   0
 ...weight_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp |   0
 .../gpu/grouped_conv3d_fwd/CMakeLists.txt           |   0
 ...wmma_gndhwc_gkzyxc_gndhwk_f16_1x1p0_instance.cpp |   0
 ...ma_gndhwc_gkzyxc_gndhwk_f16_1x1s1p0_instance.cpp |   0
 ...d_fwd_wmma_gndhwc_gkzyxc_gndhwk_f16_instance.cpp |   0
 ..._wmma_gndhwc_gkzyxc_gndhwk_f16_oddc_instance.cpp |   0
 ..._wmma_gndhwc_gkzyxc_gndhwk_i8_1x1p0_instance.cpp |   0
 ...mma_gndhwc_gkzyxc_gndhwk_i8_1x1s1p0_instance.cpp |   0
 ...3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_i8_instance.cpp |   0
 ...d_wmma_gndhwc_gkzyxc_gndhwk_i8_oddc_instance.cpp |   0
 ...wmma_ndhwgc_gkzyxc_ndhwgk_f16_1x1p0_instance.cpp |   0
 ...ma_ndhwgc_gkzyxc_ndhwgk_f16_1x1s1p0_instance.cpp |   0
 ...d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp |   0
 ..._wmma_ndhwgc_gkzyxc_ndhwgk_f16_oddc_instance.cpp |   0
 ..._wmma_ndhwgc_gkzyxc_ndhwgk_i8_1x1p0_instance.cpp |   0
 ...mma_ndhwgc_gkzyxc_ndhwgk_i8_1x1s1p0_instance.cpp |   0
 ...3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_i8_instance.cpp |   0
 ...d_wmma_ndhwgc_gkzyxc_ndhwgk_i8_oddc_instance.cpp |   0
 ...d_fwd_xdl_gndhwc_gkzyxc_gndhwk_bf16_instance.cpp |   0
 ...3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_f16_instance.cpp |   0
 ...3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_f32_instance.cpp |   0
 ...d_fwd_xdl_gndhwc_gkzyxc_gndhwk_int8_instance.cpp |   0
 ...d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp |   0
 ...l_ndhwgc_gkzyxc_ndhwgk_f16_comp_fp8_instance.cpp |   0
 ...3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp |   0
 ...3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp |   0
 ...d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_int8_instance.cpp |   0
 .../gpu/grouped_gemm/CMakeLists.txt                 |   0
 ...ouped_gemm_xdl_f16_f16_f16_km_kn_mn_instance.cpp |   0
 ...ouped_gemm_xdl_f16_f16_f16_km_nk_mn_instance.cpp |   0
 ...ouped_gemm_xdl_f16_f16_f16_mk_kn_mn_instance.cpp |   0
 ...ouped_gemm_xdl_f16_f16_f16_mk_nk_mn_instance.cpp |   0
 ...emm_xdl_splitk_f16_f16_f16_mk_kn_mn_instance.cpp |   0
 ...litk_f16_f16_f16_mk_kn_mn_irregular_instance.cpp |   0
 ...emm_xdl_splitk_f16_f16_f16_mk_nk_mn_instance.cpp |   0
 ...litk_f16_f16_f16_mk_nk_mn_irregular_instance.cpp |   0
 .../gpu/grouped_gemm_bias/CMakeLists.txt            |   0
 ..._fixed_nk_bias_f16_f16_f16_mk_kn_mn_instance.cpp |   0
 ..._fixed_nk_bias_f16_f16_f16_mk_nk_mn_instance.cpp |   0
 ..._fixed_nk_bias_f16_f16_f32_mk_kn_mn_instance.cpp |   0
 ..._fixed_nk_bias_f16_f16_f32_mk_nk_mn_instance.cpp |   0
 .../gpu/grouped_gemm_fastgelu/CMakeLists.txt        |   0
 ...m_fastgelu_xdl_f16_f16_f16_km_kn_mn_instance.cpp |   0
 ...m_fastgelu_xdl_f16_f16_f16_km_nk_mn_instance.cpp |   0
 ...m_fastgelu_xdl_f16_f16_f16_mk_kn_mn_instance.cpp |   0
 ...m_fastgelu_xdl_f16_f16_f16_mk_nk_mn_instance.cpp |   0
 .../gpu/grouped_gemm_fixed_nk/CMakeLists.txt        |   0
 ...m_xdl_fixed_nk_f16_f16_f16_mk_kn_mn_instance.cpp |   0
 ...m_xdl_fixed_nk_f16_f16_f16_mk_nk_mn_instance.cpp |   0
 ...m_xdl_fixed_nk_f16_fp8_f16_mk_kn_mn_instance.cpp |   0
 ...m_xdl_fixed_nk_f16_fp8_f16_mk_nk_mn_instance.cpp |   0
 ...mm_xdl_fixed_nk_f16_i8_f16_mk_kn_mn_instance.cpp |   0
 ...mm_xdl_fixed_nk_f16_i8_f16_mk_nk_mn_instance.cpp |   0
 .../gpu/image_to_column/CMakeLists.txt              |   0
 .../device_image_to_column_nhwc_1d_instance.cpp     |   0
 .../device_image_to_column_nhwc_2d_instance.cpp     |   0
 .../device_image_to_column_nhwc_3d_instance.cpp     |   0
 .../gpu/max_pool_bwd/CMakeLists.txt                 |   0
 .../device_max_pool_bwd_bf16_instance.cpp           |   0
 .../device_max_pool_bwd_f16_instance.cpp            |   0
 .../device_max_pool_bwd_f32_instance.cpp            |   0
 .../max_pool_bwd/max_pool_bwd_instance_common.hpp   |   0
 .../gpu/normalization/CMakeLists.txt                |   0
 .../normalization/device_groupnorm_f16_instance.cpp |   0
 .../normalization/device_groupnorm_f32_instance.cpp |   0
 ...ice_groupnorm_swish_f16_f32_f32_f16_instance.cpp |   0
 .../device_groupnorm_swish_f16_instance.cpp         |   0
 .../device_groupnorm_swish_f32_instance.cpp         |   0
 .../device_layernorm2d_f16_instance.cpp             |   0
 .../device_layernorm2d_f32_instance.cpp             |   0
 .../device_layernorm4d_f16_instance.cpp             |   0
 .../device_layernorm4d_f32_instance.cpp             |   0
 .../normalization/normalization_instance_common.hpp |   0
 .../gpu/pool3d_fwd/CMakeLists.txt                   |   0
 .../device_avg_pool3d_fwd_ndhwc_bf16_instance.cpp   |   0
 .../device_avg_pool3d_fwd_ndhwc_f16_instance.cpp    |   0
 .../device_avg_pool3d_fwd_ndhwc_f32_instance.cpp    |   0
 .../device_max_pool3d_fwd_ndhwc_bf16_instance.cpp   |   0
 .../device_max_pool3d_fwd_ndhwc_f16_instance.cpp    |   0
 .../device_max_pool3d_fwd_ndhwc_f32_instance.cpp    |   0
 .../gpu/pool3d_fwd/pool_fwd_instance_common.hpp     |   0
 .../gpu/quantization/CMakeLists.txt                 |   0
 .../conv2d_fwd/conv2d_quantization_common.hpp       |   0
 ...l_bias_perchannel_quantization_int8_instance.cpp |   0
 ..._dl_bias_perlayer_quantization_int8_instance.cpp |   0
 .../conv2d_fwd/device_conv2d_dl_int8_instance.hpp   |   0
 ...v2d_dl_perchannel_quantization_int8_instance.cpp |   0
 ...onv2d_dl_perlayer_quantization_int8_instance.cpp |   0
 ...l_bias_perchannel_quantization_int8_instance.cpp |   0
 ...xdl_bias_perlayer_quantization_int8_instance.cpp |   0
 .../conv2d_fwd/device_conv2d_xdl_int8_instance.hpp  |   0
 ...2d_xdl_perchannel_quantization_int8_instance.cpp |   0
 ...nv2d_xdl_perlayer_quantization_int8_instance.cpp |   0
 ..._quantization_dl_c_shuffle_i8_i8_i8_instance.hpp |   0
 ...tion_dl_c_shuffle_i8_i8_i8_km_kn_mn_instance.cpp |   0
 ...tion_dl_c_shuffle_i8_i8_i8_km_nk_mn_instance.cpp |   0
 ...tion_dl_c_shuffle_i8_i8_i8_mk_kn_mn_instance.cpp |   0
 ...tion_dl_c_shuffle_i8_i8_i8_mk_nk_mn_instance.cpp |   0
 ...quantization_xdl_c_shuffle_i8_i8_i8_instance.hpp |   0
 ...ion_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instance.cpp |   0
 ...ion_xdl_c_shuffle_i8_i8_i8_km_nk_mn_instance.cpp |   0
 ...ion_xdl_c_shuffle_i8_i8_i8_mk_kn_mn_instance.cpp |   0
 ...ion_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instance.cpp |   0
 .../quantization/gemm/gemm_quantization_common.hpp  |   0
 .../gpu/reduce/CMakeLists.txt                       |   0
 ...ce_reduce_instance_blockwise_b16_f32_b16_add.cpp |   0
 ...e_reduce_instance_blockwise_b16_f32_b16_amax.cpp |   0
 ...ce_reduce_instance_blockwise_b16_f32_b16_avg.cpp |   0
 ...ce_reduce_instance_blockwise_b16_f32_b16_max.cpp |   0
 ...ce_reduce_instance_blockwise_b16_f32_b16_min.cpp |   0
 ..._reduce_instance_blockwise_b16_f32_b16_norm2.cpp |   0
 ...e_reduce_instance_blockwise_f16_f16_f16_amax.cpp |   0
 ...ce_reduce_instance_blockwise_f16_f16_f16_max.cpp |   0
 ...ce_reduce_instance_blockwise_f16_f16_f16_min.cpp |   0
 ...ce_reduce_instance_blockwise_f16_f32_f16_add.cpp |   0
 ...ce_reduce_instance_blockwise_f16_f32_f16_avg.cpp |   0
 ..._reduce_instance_blockwise_f16_f32_f16_norm2.cpp |   0
 ...ce_reduce_instance_blockwise_f32_f32_f32_add.cpp |   0
 ...e_reduce_instance_blockwise_f32_f32_f32_amax.cpp |   0
 ...ce_reduce_instance_blockwise_f32_f32_f32_avg.cpp |   0
 ...ce_reduce_instance_blockwise_f32_f32_f32_max.cpp |   0
 ...ce_reduce_instance_blockwise_f32_f32_f32_min.cpp |   0
 ..._reduce_instance_blockwise_f32_f32_f32_norm2.cpp |   0
 ...ce_reduce_instance_blockwise_f32_f64_f32_add.cpp |   0
 ...ce_reduce_instance_blockwise_f32_f64_f32_avg.cpp |   0
 ..._reduce_instance_blockwise_f32_f64_f32_norm2.cpp |   0
 ...ce_reduce_instance_blockwise_f64_f64_f64_add.cpp |   0
 ...e_reduce_instance_blockwise_f64_f64_f64_amax.cpp |   0
 ...ce_reduce_instance_blockwise_f64_f64_f64_avg.cpp |   0
 ...ce_reduce_instance_blockwise_f64_f64_f64_max.cpp |   0
 ...ce_reduce_instance_blockwise_f64_f64_f64_min.cpp |   0
 ..._reduce_instance_blockwise_f64_f64_f64_norm2.cpp |   0
 ...vice_reduce_instance_blockwise_i8_i32_i8_add.cpp |   0
 ...vice_reduce_instance_blockwise_i8_i32_i8_avg.cpp |   0
 ...vice_reduce_instance_blockwise_i8_i8_i8_amax.cpp |   0
 ...evice_reduce_instance_blockwise_i8_i8_i8_max.cpp |   0
 ...evice_reduce_instance_blockwise_i8_i8_i8_min.cpp |   0
 ...stance_multiblock_atomic_add_b16_f32_f32_add.cpp |   0
 ...stance_multiblock_atomic_add_b16_f32_f32_avg.cpp |   0
 ...stance_multiblock_atomic_add_f16_f32_f32_add.cpp |   0
 ...stance_multiblock_atomic_add_f16_f32_f32_avg.cpp |   0
 ...stance_multiblock_atomic_add_f32_f32_f32_add.cpp |   0
 ...stance_multiblock_atomic_add_f32_f32_f32_avg.cpp |   0
 ...stance_multiblock_atomic_add_f32_f64_f32_add.cpp |   0
 ...stance_multiblock_atomic_add_f32_f64_f32_avg.cpp |   0
 ...stance_multiblock_atomic_add_f64_f64_f64_add.cpp |   0
 ...stance_multiblock_atomic_add_f64_f64_f64_avg.cpp |   0
 ...e_reduce_instance_threadwise_b16_f32_b16_add.cpp |   0
 ..._reduce_instance_threadwise_b16_f32_b16_amax.cpp |   0
 ...e_reduce_instance_threadwise_b16_f32_b16_avg.cpp |   0
 ...e_reduce_instance_threadwise_b16_f32_b16_max.cpp |   0
 ...e_reduce_instance_threadwise_b16_f32_b16_min.cpp |   0
 ...reduce_instance_threadwise_b16_f32_b16_norm2.cpp |   0
 ..._reduce_instance_threadwise_f16_f16_f16_amax.cpp |   0
 ...e_reduce_instance_threadwise_f16_f16_f16_max.cpp |   0
 ...e_reduce_instance_threadwise_f16_f16_f16_min.cpp |   0
 ...e_reduce_instance_threadwise_f16_f32_f16_add.cpp |   0
 ...e_reduce_instance_threadwise_f16_f32_f16_avg.cpp |   0
 ...reduce_instance_threadwise_f16_f32_f16_norm2.cpp |   0
 ...e_reduce_instance_threadwise_f32_f32_f32_add.cpp |   0
 ..._reduce_instance_threadwise_f32_f32_f32_amax.cpp |   0
 ...e_reduce_instance_threadwise_f32_f32_f32_avg.cpp |   0
 ...e_reduce_instance_threadwise_f32_f32_f32_max.cpp |   0
 ...e_reduce_instance_threadwise_f32_f32_f32_min.cpp |   0
 ...reduce_instance_threadwise_f32_f32_f32_norm2.cpp |   0
 ...e_reduce_instance_threadwise_f32_f64_f32_add.cpp |   0
 ...e_reduce_instance_threadwise_f32_f64_f32_avg.cpp |   0
 ...reduce_instance_threadwise_f32_f64_f32_norm2.cpp |   0
 ...e_reduce_instance_threadwise_f64_f64_f64_add.cpp |   0
 ..._reduce_instance_threadwise_f64_f64_f64_amax.cpp |   0
 ...e_reduce_instance_threadwise_f64_f64_f64_avg.cpp |   0
 ...e_reduce_instance_threadwise_f64_f64_f64_max.cpp |   0
 ...e_reduce_instance_threadwise_f64_f64_f64_min.cpp |   0
 ...reduce_instance_threadwise_f64_f64_f64_norm2.cpp |   0
 ...ice_reduce_instance_threadwise_i8_i32_i8_add.cpp |   0
 ...ice_reduce_instance_threadwise_i8_i32_i8_avg.cpp |   0
 ...ice_reduce_instance_threadwise_i8_i8_i8_amax.cpp |   0
 ...vice_reduce_instance_threadwise_i8_i8_i8_max.cpp |   0
 ...vice_reduce_instance_threadwise_i8_i8_i8_min.cpp |   0
 .../gpu/softmax/CMakeLists.txt                      |   0
 ...evice_softmax_f16_f16_instance_rank3_reduce1.cpp |   0
 ...evice_softmax_f16_f16_instance_rank3_reduce2.cpp |   0
 ...evice_softmax_f16_f16_instance_rank3_reduce3.cpp |   0
 ...evice_softmax_f16_f16_instance_rank4_reduce1.cpp |   0
 ...evice_softmax_f16_f16_instance_rank4_reduce2.cpp |   0
 ...evice_softmax_f16_f16_instance_rank4_reduce3.cpp |   0
 ...evice_softmax_f16_f16_instance_rank4_reduce4.cpp |   0
 ...evice_softmax_f32_f32_instance_rank3_reduce1.cpp |   0
 ...evice_softmax_f32_f32_instance_rank3_reduce2.cpp |   0
 ...evice_softmax_f32_f32_instance_rank3_reduce3.cpp |   0
 ...evice_softmax_f32_f32_instance_rank4_reduce1.cpp |   0
 ...evice_softmax_f32_f32_instance_rank4_reduce2.cpp |   0
 ...evice_softmax_f32_f32_instance_rank4_reduce3.cpp |   0
 ...evice_softmax_f32_f32_instance_rank4_reduce4.cpp |   0
 .../gpu/tall_and_skinny_gemm_splitk/CMakeLists.txt  |   0
 ...ny_gemm_splitk_f16_f16_f16_mk_kn_mn_instance.cpp |   0
 ...ny_gemm_splitk_f16_f16_f16_mk_nk_mn_instance.cpp |   0
 .../library => library}/src/utility/CMakeLists.txt  |   0
 .../src/utility/convolution_parameter.cpp           |   0
 .../src/utility/device_memory.cpp                   |   0
 .../library => library}/src/utility/host_tensor.cpp |   0
 .../profiler => profiler}/CMakeLists.txt            |   0
 {composable_kernel/profiler => profiler}/README.md  |   0
 .../include/profiler/data_type_enum.hpp             |   0
 .../profiler/profile_avg_pool3d_bwd_impl.hpp        |   0
 .../profile_batched_gemm_add_relu_gemm_add_impl.hpp |   0
 ..._batched_gemm_bias_softmax_gemm_permute_impl.hpp |   0
 .../profiler/profile_batched_gemm_gemm_impl.hpp     |   0
 .../include/profiler/profile_batched_gemm_impl.hpp  |   0
 .../profiler/profile_batched_gemm_reduce_impl.hpp   |   0
 .../profile_batched_gemm_softmax_gemm_impl.hpp      |   0
 ...ofile_batched_gemm_softmax_gemm_permute_impl.hpp |   0
 .../profiler/profile_batchnorm_backward_impl.hpp    |   0
 .../profiler/profile_batchnorm_forward_impl.hpp     |   0
 .../profiler/profile_batchnorm_infer_impl.hpp       |   0
 .../include/profiler/profile_contraction_impl.hpp   |   0
 .../include/profiler/profile_contraction_utils.hpp  |   0
 .../include/profiler/profile_conv_bwd_data_impl.hpp |   0
 .../profile_conv_fwd_bias_relu_add_impl.hpp         |   0
 .../profiler/profile_conv_fwd_bias_relu_impl.hpp    |   0
 .../include/profiler/profile_conv_fwd_impl.hpp      |   0
 .../profiler/profile_conv_tensor_rearrange_impl.hpp |   0
 .../profiler/profile_elementwise_layernorm_impl.hpp |   0
 .../profiler/profile_gemm_add_add_fastgelu_impl.hpp |   0
 .../profiler/profile_gemm_add_fastgelu_impl.hpp     |   0
 .../profiler/profile_gemm_add_multiply_impl.hpp     |   0
 .../profile_gemm_add_relu_add_layernorm_impl.hpp    |   0
 .../profiler/profile_gemm_bias_add_reduce_impl.hpp  |   0
 .../include/profiler/profile_gemm_bilinear_impl.hpp |   0
 .../include/profiler/profile_gemm_fastgelu_impl.hpp |   0
 .../include/profiler/profile_gemm_impl.hpp          |   0
 .../profiler/profile_gemm_multiply_add_impl.hpp     |   0
 .../include/profiler/profile_gemm_reduce_impl.hpp   |   0
 .../include/profiler/profile_gemm_splitk_impl.hpp   |   0
 .../include/profiler/profile_gemm_streamk_impl.hpp  |   0
 .../include/profiler/profile_gemv_splitk_impl.hpp   |   0
 .../profiler/profile_grouped_conv_bwd_data_impl.hpp |   0
 .../profile_grouped_conv_bwd_weight_impl.hpp        |   0
 .../profiler/profile_grouped_conv_fwd_impl.hpp      |   0
 .../profiler/profile_grouped_gemm_fastgelu_impl.hpp |   0
 .../include/profiler/profile_grouped_gemm_impl.hpp  |   0
 .../include/profiler/profile_groupnorm_impl.hpp     |   0
 .../include/profiler/profile_layernorm_impl.hpp     |   0
 .../profiler/profile_max_pool3d_bwd_impl.hpp        |   0
 .../include/profiler/profile_pool3d_fwd_impl.hpp    |   0
 .../include/profiler/profile_reduce_impl.hpp        |   0
 .../include/profiler/profile_softmax_impl.hpp       |   0
 .../profile_tall_and_skinny_gemm_splitk_impl.hpp    |   0
 .../profiler => profiler}/src/CMakeLists.txt        |   0
 .../src/profile_avg_pool3d_bwd.cpp                  |   0
 .../src/profile_batched_gemm.cpp                    |   0
 .../src/profile_batched_gemm_add_relu_gemm_add.cpp  |   0
 .../src/profile_batched_gemm_gemm.cpp               |   0
 .../src/profile_batched_gemm_multi_d.cpp            |   0
 .../src/profile_batched_gemm_reduce.cpp             |   0
 .../src/profile_batchnorm_bwd.cpp                   |   0
 .../src/profile_batchnorm_fwd.cpp                   |   0
 .../src/profile_batchnorm_infer.cpp                 |   0
 .../src/profile_contraction_bilinear.cpp            |   0
 .../src/profile_contraction_scale.cpp               |   0
 .../src/profile_conv_bwd_data.cpp                   |   0
 .../profiler => profiler}/src/profile_conv_fwd.cpp  |   0
 .../src/profile_conv_fwd_bias_relu.cpp              |   0
 .../src/profile_conv_fwd_bias_relu_add.cpp          |   0
 .../src/profile_conv_tensor_rearrange.cpp           |   0
 .../profiler => profiler}/src/profile_gemm.cpp      |   0
 .../src/profile_gemm_add_add_fastgelu.cpp           |   0
 .../src/profile_gemm_add_fastgelu.cpp               |   0
 .../src/profile_gemm_add_multiply.cpp               |   0
 .../src/profile_gemm_add_relu_add_layernorm.cpp     |   0
 .../src/profile_gemm_bias_add_reduce.cpp            |   0
 .../src/profile_gemm_bilinear.cpp                   |   0
 .../src/profile_gemm_fastgelu.cpp                   |   0
 .../src/profile_gemm_multiply_add.cpp               |   0
 .../src/profile_gemm_reduce.cpp                     |   0
 .../src/profile_gemm_splitk.cpp                     |   0
 .../src/profile_gemm_streamk.cpp                    |   0
 .../src/profile_gemv_splitk.cpp                     |   0
 .../src/profile_grouped_conv_bwd_data.cpp           |   0
 .../src/profile_grouped_conv_bwd_weight.cpp         |   0
 .../src/profile_grouped_conv_fwd.cpp                |   0
 .../src/profile_grouped_gemm.cpp                    |   0
 .../src/profile_grouped_gemm_fastgelu.cpp           |   0
 .../profiler => profiler}/src/profile_groupnorm.cpp |   0
 .../profiler => profiler}/src/profile_layernorm.cpp |   0
 .../src/profile_max_pool3d_bwd.cpp                  |   0
 .../src/profile_max_pool3d_fwd.cpp                  |   0
 .../profiler => profiler}/src/profile_reduce.cpp    |   0
 .../profiler => profiler}/src/profile_softmax.cpp   |   0
 .../src/profile_tall_and_skinny_gemm_splitk.cpp     |   0
 .../profiler => profiler}/src/profiler.cpp          |   0
 .../src/profiler_operation_registry.hpp             |   0
 composable_kernel/rbuild.ini => rbuild.ini          |   0
 .../requirements.txt => requirements.txt            |   0
 .../script => script}/check_copyright_year.sh       |   0
 .../script => script}/clang-format-overwrite.sh     |   0
 .../script => script}/cmake-ck-dev.sh               |   0
 .../script => script}/cmake-ck-release.sh           |   0
 {composable_kernel/script => script}/count_vgpr.sh  |   0
 .../script => script}/hipclang_opt.sh               |   0
 .../script => script}/install_precommit.sh          |   0
 .../script => script}/parse_perf_data.py            |   0
 .../script => script}/process_perf_data.py          |   0
 .../script => script}/process_perf_data.sh          |   0
 .../script => script}/process_qa_data.sh            |   0
 .../script => script}/profile_batched_gemm.sh       |   0
 .../script => script}/profile_conv_bwd_data.sh      |   0
 .../script => script}/profile_conv_fwd.sh           |   0
 .../script => script}/profile_gemm.sh               |   0
 .../script => script}/profile_gemm_bilinear.sh      |   0
 .../script => script}/profile_grouped_gemm.sh       |   0
 .../script => script}/profile_onnx_gemm.sh          |   0
 .../script => script}/profile_reduce_no_index.sh    |   0
 .../script => script}/profile_reduce_with_index.sh  |   0
 .../script => script}/profile_resnet50.sh           |   0
 .../script => script}/profile_splitK_gemm.sh        |   0
 .../script => script}/run_full_performance_tests.sh |   0
 .../script => script}/run_performance_tests.sh      |   0
 .../script => script}/test_convnd_fwd.sh            |   0
 .../script => script}/test_reduce_no_index.sh       |   0
 .../script => script}/test_reduce_with_index.sh     |   0
 .../script => script}/uninstall_precommit.sh        |   0
 {composable_kernel/test => test}/CMakeLists.txt     |   0
 .../test => test}/batched_gemm/CMakeLists.txt       |   0
 .../batched_gemm/batched_gemm_bf16.cpp              |   0
 .../batched_gemm/batched_gemm_fp16.cpp              |   0
 .../batched_gemm/batched_gemm_fp32.cpp              |   0
 .../batched_gemm/batched_gemm_int8.cpp              |   0
 .../batched_gemm/test_batched_gemm.cpp              |   0
 .../test => test}/batched_gemm_gemm/CMakeLists.txt  |   0
 .../test_batched_gemm_gemm_fp16.cpp                 |   0
 .../test_batched_gemm_gemm_util.hpp                 |   0
 .../batched_gemm_multi_d/CMakeLists.txt             |   0
 .../test_batched_gemm_multi_d_dl.cpp                |   0
 .../batched_gemm_reduce/CMakeLists.txt              |   0
 .../batched_gemm_reduce_fp16.cpp                    |   0
 .../batched_gemm_softmax_gemm/CMakeLists.txt        |   0
 .../test_batched_gemm_softmax_gemm_fp16.cpp         |   0
 .../test_batched_gemm_softmax_gemm_util.hpp         |   0
 .../CMakeLists.txt                                  |   0
 ..._batched_gemm_bias_softmax_gemm_permute_bf16.cpp |   0
 ..._batched_gemm_bias_softmax_gemm_permute_fp16.cpp |   0
 ..._batched_gemm_bias_softmax_gemm_permute_util.hpp |   0
 .../test_batched_gemm_softmax_gemm_permute_bf16.cpp |   0
 .../test_batched_gemm_softmax_gemm_permute_fp16.cpp |   0
 .../test_batched_gemm_softmax_gemm_permute_util.hpp |   0
 .../test => test}/batchnorm/CMakeLists.txt          |   0
 .../batchnorm/batchnorm_bwd_rank_4.cpp              |   0
 .../batchnorm/batchnorm_fwd_rank_4.cpp              |   0
 .../batchnorm/batchnorm_infer_rank_4.cpp            |   0
 .../block_swizzle_test/block_swizzle_test.cpp       |   0
 .../test => test}/block_swizzle_test/rebuild.sh     |   0
 .../test => test}/block_swizzle_test/simple_args.h  |   0
 .../test => test}/block_to_ctile_map/CMakeLists.txt |   0
 .../block_to_ctile_map/test_block_to_ctile_map.cpp  |   0
 .../test => test}/contraction/CMakeLists.txt        |   0
 .../test => test}/contraction/test_contraction.cpp  |   0
 .../contraction/test_contraction_interface.cpp      |   0
 .../conv_tensor_rearrange/CMakeLists.txt            |   0
 .../test_conv_tensor_rearrange.cpp                  |   0
 .../test_conv_tensor_rearrange_interface.cpp        |   0
 .../test => test}/conv_util/CMakeLists.txt          |   0
 .../test => test}/conv_util/conv_util.cpp           |   0
 .../test => test}/convnd_bwd_data/CMakeLists.txt    |   0
 .../convnd_bwd_data/convnd_bwd_data.cpp             |   0
 .../test => test}/convnd_fwd/CMakeLists.txt         |   0
 .../test => test}/convnd_fwd/convnd_fwd.cpp         |   0
 .../test => test}/data_type/CMakeLists.txt          |   0
 .../test => test}/data_type/test_bf8.cpp            |   0
 .../test => test}/data_type/test_fp8.cpp            |   0
 .../test => test}/data_type/test_int4.cpp           |   0
 .../test => test}/data_type/type_convert_const.cpp  |   0
 .../elementwise_normalization/CMakeLists.txt        |   0
 .../test_elementwise_layernorm_fp16.cpp             |   0
 .../test => test}/gemm/CMakeLists.txt               |   0
 {composable_kernel/test => test}/gemm/gemm_bf16.cpp |   0
 {composable_kernel/test => test}/gemm/gemm_fp16.cpp |   0
 {composable_kernel/test => test}/gemm/gemm_fp32.cpp |   0
 {composable_kernel/test => test}/gemm/gemm_fp64.cpp |   0
 {composable_kernel/test => test}/gemm/gemm_int8.cpp |   0
 .../test => test}/gemm/gemm_standalone_xdl_fp16.cpp |   0
 {composable_kernel/test => test}/gemm/gemm_util.hpp |   0
 .../gemm/instance/gemm_f16_nn_instance.cpp          |   0
 .../gemm/instance/gemm_f16_nn_instance.hpp          |   0
 .../gemm/instance/gemm_f16_nt_instance.cpp          |   0
 .../gemm/instance/gemm_f16_nt_instance.hpp          |   0
 .../gemm/instance/gemm_f16_tn_instance.cpp          |   0
 .../gemm/instance/gemm_f16_tn_instance.hpp          |   0
 .../gemm/instance/gemm_f16_tt_instance.cpp          |   0
 .../gemm/instance/gemm_f16_tt_instance.hpp          |   0
 .../gemm/instance/gemm_wavelet_f16_tn_instance.cpp  |   0
 .../gemm/instance/gemm_wavelet_f16_tn_instance.hpp  |   0
 .../test => test}/gemm/run_gemm_test.inc            |   0
 .../test => test}/gemm_layernorm/CMakeLists.txt     |   0
 .../test_gemm_add_relu_add_layernorm_fp16.cpp       |   0
 .../test => test}/gemm_reduce/CMakeLists.txt        |   0
 .../test => test}/gemm_reduce/gemm_reduce_fp16.cpp  |   0
 .../test => test}/gemm_split_k/CMakeLists.txt       |   0
 .../test => test}/gemm_split_k/test_gemm_splitk.cpp |   0
 .../gemm_split_k/test_gemm_splitk_ut_cases.inc      |   0
 .../gemm_split_k/test_gemm_splitk_util.hpp          |   0
 .../grouped_convnd_bwd_data/CMakeLists.txt          |   0
 .../test_grouped_convnd_bwd_data.cpp                |   0
 .../test_grouped_convnd_bwd_data_interface_wmma.cpp |   0
 .../test_grouped_convnd_bwd_data_interface_xdl.cpp  |   0
 .../grouped_convnd_bwd_weight/CMakeLists.txt        |   0
 .../test_grouped_convnd_bwd_weight.cpp              |   0
 ...est_grouped_convnd_bwd_weight_interface_wmma.cpp |   0
 ...test_grouped_convnd_bwd_weight_interface_xdl.cpp |   0
 .../test => test}/grouped_convnd_fwd/CMakeLists.txt |   0
 .../grouped_convnd_fwd/grouped_convnd_fwd.cpp       |   0
 .../grouped_convnd_fwd/test_grouped_convnd_fwd.cpp  |   0
 .../test => test}/grouped_gemm/CMakeLists.txt       |   0
 .../grouped_gemm/test_grouped_gemm_interface.cpp    |   0
 .../grouped_gemm/test_grouped_gemm_splitk.cpp       |   0
 .../grouped_gemm/test_grouped_gemm_ut_cases.inc     |   0
 .../grouped_gemm/test_grouped_gemm_util.hpp         |   0
 .../test => test}/image_to_column/CMakeLists.txt    |   0
 .../image_to_column/test_image_to_column.cpp        |   0
 .../test_image_to_column_interface.cpp              |   0
 .../magic_number_division/CMakeLists.txt            |   0
 .../magic_number_division/magic_number_division.cpp |   0
 .../test => test}/normalization/CMakeLists.txt      |   0
 .../normalization/test_groupnorm_fp16.cpp           |   0
 .../normalization/test_groupnorm_fp32.cpp           |   0
 .../normalization/test_layernorm2d_fp16.cpp         |   0
 .../normalization/test_layernorm2d_fp32.cpp         |   0
 .../test => test}/pool/CMakeLists.txt               |   0
 .../test => test}/pool/test_avg_pool3d_bwd.cpp      |   0
 .../test => test}/pool/test_avg_pool3d_fwd.cpp      |   0
 .../test => test}/pool/test_max_pool3d_bwd.cpp      |   0
 .../test => test}/pool/test_max_pool3d_fwd.cpp      |   0
 .../test => test}/pool/test_pool_fwd_common.hpp     |   0
 .../test => test}/reduce/CMakeLists.txt             |   0
 .../test => test}/reduce/reduce_no_index.cpp        |   0
 .../test => test}/reduce/reduce_with_index.cpp      |   0
 .../test => test}/reference_conv_fwd/CMakeLists.txt |   0
 .../reference_conv_fwd/reference_conv_fwd.cpp       |   0
 .../test => test}/softmax/CMakeLists.txt            |   0
 .../softmax/test_softmax_interface.cpp              |   0
 .../test => test}/softmax/test_softmax_rank3.cpp    |   0
 .../test => test}/softmax/test_softmax_rank4.cpp    |   0
 .../test => test}/softmax/test_softmax_ut_cases.inc |   0
 .../test => test}/softmax/test_softmax_util.hpp     |   0
 .../space_filling_curve/CMakeLists.txt              |   0
 .../space_filling_curve/space_filling_curve.cpp     |   0
 .../test => test}/wmma_op/CMakeLists.txt            |   0
 .../test => test}/wmma_op/wmma_op.cpp               |   0
 .../test => test}/wmma_op/wmma_op_util.hpp          |   0
 1828 files changed, 0 insertions(+), 0 deletions(-)
 rename composable_kernel/CHANGELOG.md => CHANGELOG.md (100%)
 rename composable_kernel/CITATION.cff => CITATION.cff (100%)
 rename composable_kernel/CMakeLists.txt => CMakeLists.txt (100%)
 rename composable_kernel/CONTRIBUTORS.md => CONTRIBUTORS.md (100%)
 rename composable_kernel/Config.cmake.in => Config.cmake.in (100%)
 rename composable_kernel/Dockerfile => Dockerfile (100%)
 rename composable_kernel/Jenkinsfile => Jenkinsfile (100%)
 rename composable_kernel/LICENSE => LICENSE (100%)
 rename composable_kernel/README.md => README.md (100%)
 rename {composable_kernel/client_example => client_example}/01_gemm/CMakeLists.txt (100%)
 rename {composable_kernel/client_example => client_example}/01_gemm/gemm.cpp (100%)
 rename {composable_kernel/client_example => client_example}/02_gemm_add_add_fastgelu/CMakeLists.txt (100%)
 rename {composable_kernel/client_example => client_example}/02_gemm_add_add_fastgelu/gemm_add_add_fastgelu.cpp (100%)
 rename {composable_kernel/client_example => client_example}/02_gemm_add_add_fastgelu/gemm_add_add_fastgelu_generic.cpp (100%)
 rename {composable_kernel/client_example => client_example}/02_gemm_add_add_fastgelu/gemm_add_fastgelu.cpp (100%)
 rename {composable_kernel/client_example => client_example}/02_gemm_add_add_fastgelu/gemm_add_fastgelu_generic.cpp (100%)
 rename {composable_kernel/client_example => client_example}/02_gemm_add_add_fastgelu/gemm_fastgelu.cpp (100%)
 rename {composable_kernel/client_example => client_example}/02_gemm_add_add_fastgelu/gemm_fastgelu_generic.cpp (100%)
 rename {composable_kernel/client_example => client_example}/03_gemm_layernorm/CMakeLists.txt (100%)
 rename {composable_kernel/client_example => client_example}/03_gemm_layernorm/gemm_add_add_layernorm_naive.cpp (100%)
 rename {composable_kernel/client_example => client_example}/03_gemm_layernorm/gemm_add_relu_add_layernorm_welford.cpp (100%)
 rename {composable_kernel/client_example => client_example}/04_contraction/CMakeLists.txt (100%)
 rename {composable_kernel/client_example => client_example}/04_contraction/contraction_bilinear_fp32.cpp (100%)
 rename {composable_kernel/client_example => client_example}/04_contraction/contraction_bilinear_fp64.cpp (100%)
 rename {composable_kernel/client_example => client_example}/04_contraction/contraction_g1m2n3k1_add_xdl_fp16.cpp (100%)
 rename {composable_kernel/client_example => client_example}/04_contraction/contraction_scale_fp32.cpp (100%)
 rename {composable_kernel/client_example => client_example}/04_contraction/contraction_scale_fp64.cpp (100%)
 rename {composable_kernel/client_example => client_example}/05_layernorm/CMakeLists.txt (100%)
 rename {composable_kernel/client_example => client_example}/05_layernorm/layernorm2d.cpp (100%)
 rename {composable_kernel/client_example => client_example}/06_softmax/CMakeLists.txt (100%)
 rename {composable_kernel/client_example => client_example}/06_softmax/softmax4d.cpp (100%)
 rename {composable_kernel/client_example => client_example}/07_grouped_convnd_fwd/CMakeLists.txt (100%)
 rename {composable_kernel/client_example => client_example}/07_grouped_convnd_fwd/grouped_conv1d_fwd.cpp (100%)
 rename {composable_kernel/client_example => client_example}/07_grouped_convnd_fwd/grouped_conv2d_fwd.cpp (100%)
 rename {composable_kernel/client_example => client_example}/08_fused_attention/CMakeLists.txt (100%)
 rename {composable_kernel/client_example => client_example}/08_fused_attention/fused_attention.cpp (100%)
 rename {composable_kernel/client_example => client_example}/08_fused_attention/fused_attention_bias.cpp (100%)
 rename {composable_kernel/client_example => client_example}/09_quantization/CMakeLists.txt (100%)
 rename {composable_kernel/client_example => client_example}/09_quantization/conv2d_fwd_bias_relu_perchannel_quantization.cpp (100%)
 rename {composable_kernel/client_example => client_example}/09_quantization/conv2d_fwd_bias_relu_perlayer_quantization.cpp (100%)
 rename {composable_kernel/client_example => client_example}/09_quantization/conv2d_fwd_bias_tanh_perchannel_quantization.cpp (100%)
 rename {composable_kernel/client_example => client_example}/09_quantization/conv2d_fwd_bias_tanh_perlayer_quantization.cpp (100%)
 rename {composable_kernel/client_example => client_example}/09_quantization/conv2d_fwd_perchannel_quantization.cpp (100%)
 rename {composable_kernel/client_example => client_example}/09_quantization/conv2d_fwd_perlayer_quantization.cpp (100%)
 rename {composable_kernel/client_example => client_example}/09_quantization/gemm_quantization.cpp (100%)
 rename {composable_kernel/client_example => client_example}/10_grouped_convnd_bwd_data/CMakeLists.txt (100%)
 rename {composable_kernel/client_example => client_example}/10_grouped_convnd_bwd_data/grouped_conv2d_bwd_data.cpp (100%)
 rename {composable_kernel/client_example => client_example}/10_grouped_convnd_bwd_data/grouped_conv3d_bwd_data.cpp (100%)
 rename {composable_kernel/client_example => client_example}/10_grouped_convnd_bwd_data/grouped_conv3d_bwd_data_input_fp16_comp_bf8f8.cpp (100%)
 rename {composable_kernel/client_example => client_example}/11_grouped_conv_bwd_weight/CMakeLists.txt (100%)
 rename {composable_kernel/client_example => client_example}/11_grouped_conv_bwd_weight/common.hpp (100%)
 rename {composable_kernel/client_example => client_example}/11_grouped_conv_bwd_weight/grouped_conv1d_bwd_weight_fp16.cpp (100%)
 rename {composable_kernel/client_example => client_example}/11_grouped_conv_bwd_weight/grouped_conv2d_bwd_weight_fp16.cpp (100%)
 rename {composable_kernel/client_example => client_example}/11_grouped_conv_bwd_weight/grouped_conv3d_bwd_weight_fp16.cpp (100%)
 rename {composable_kernel/client_example => client_example}/11_grouped_conv_bwd_weight/grouped_conv3d_bwd_weight_fp32.cpp (100%)
 rename {composable_kernel/client_example => client_example}/12_elementwise_normalization/CMakeLists.txt (100%)
 rename {composable_kernel/client_example => client_example}/12_elementwise_normalization/elementwise_layernorm2d.cpp (100%)
 rename {composable_kernel/client_example => client_example}/13_batchnorm/CMakeLists.txt (100%)
 rename {composable_kernel/client_example => client_example}/13_batchnorm/batchnorm_bwd_nhwc.cpp (100%)
 rename {composable_kernel/client_example => client_example}/13_batchnorm/batchnorm_fwd_nhwc.cpp (100%)
 rename {composable_kernel/client_example => client_example}/13_batchnorm/batchnorm_infer_nhwc.cpp (100%)
 rename {composable_kernel/client_example => client_example}/14_instance_id/CMakeLists.txt (100%)
 rename {composable_kernel/client_example => client_example}/14_instance_id/batchnorm_fwd_instance_id.cpp (100%)
 rename {composable_kernel/client_example => client_example}/15_convnd_bwd_data/CMakeLists.txt (100%)
 rename {composable_kernel/client_example => client_example}/15_convnd_bwd_data/common.hpp (100%)
 rename {composable_kernel/client_example => client_example}/15_convnd_bwd_data/conv3d_bwd_data_fp16.cpp (100%)
 rename {composable_kernel/client_example => client_example}/15_convnd_bwd_data/conv3d_bwd_data_fp32.cpp (100%)
 rename {composable_kernel/client_example => client_example}/15_gemm_add_multiply/CMakeLists.txt (100%)
 rename {composable_kernel/client_example => client_example}/15_gemm_add_multiply/gemm_add_multiply.cpp (100%)
 rename {composable_kernel/client_example => client_example}/15_reduce/CMakeLists.txt (100%)
 rename {composable_kernel/client_example => client_example}/15_reduce/reduce_nhwc_c.cpp (100%)
 rename {composable_kernel/client_example => client_example}/16_convnd_fwd/CMakeLists.txt (100%)
 rename {composable_kernel/client_example => client_example}/16_convnd_fwd/common.hpp (100%)
 rename {composable_kernel/client_example => client_example}/16_convnd_fwd/conv3d_fwd_fp16.cpp (100%)
 rename {composable_kernel/client_example => client_example}/16_convnd_fwd/conv3d_fwd_fp16_comp_fp8.cpp (100%)
 rename {composable_kernel/client_example => client_example}/16_convnd_fwd/conv3d_fwd_fp32.cpp (100%)
 rename {composable_kernel/client_example => client_example}/17_grouped_gemm_fastgelu/CMakeLists.txt (100%)
 rename {composable_kernel/client_example => client_example}/17_grouped_gemm_fastgelu/grouped_gemm_fastgelu.cpp (100%)
 rename {composable_kernel/client_example => client_example}/18_groupnorm/CMakeLists.txt (100%)
 rename {composable_kernel/client_example => client_example}/18_groupnorm/groupnorm_swish.cpp (100%)
 rename {composable_kernel/client_example => client_example}/19_pool/CMakeLists.txt (100%)
 rename {composable_kernel/client_example => client_example}/19_pool/avg_pool3d_bwd.cpp (100%)
 rename {composable_kernel/client_example => client_example}/19_pool/avg_pool3d_fwd.cpp (100%)
 rename {composable_kernel/client_example => client_example}/19_pool/max_pool2d_bwd.cpp (100%)
 rename {composable_kernel/client_example => client_example}/19_pool/max_pool2d_fwd.cpp (100%)
 rename {composable_kernel/client_example => client_example}/20_splitk_gemm/CMakeLists.txt (100%)
 rename {composable_kernel/client_example => client_example}/20_splitk_gemm/splitK_gemm_fp16_f8.cpp (100%)
 rename {composable_kernel/client_example => client_example}/21_grouped_gemm_bias/CMakeLists.txt (100%)
 rename {composable_kernel/client_example => client_example}/21_grouped_gemm_bias/grouped_gemm_fixed_nk_bias_fp16.cpp (100%)
 rename {composable_kernel/client_example => client_example}/22_grouped_gemm/CMakeLists.txt (100%)
 rename {composable_kernel/client_example => client_example}/22_grouped_gemm/grouped_gemm_fixed_nk_fp16.cpp (100%)
 rename {composable_kernel/client_example => client_example}/22_grouped_gemm/grouped_gemm_fixed_nk_fp8.cpp (100%)
 rename {composable_kernel/client_example => client_example}/22_grouped_gemm/grouped_gemm_fixed_nk_i8.cpp (100%)
 rename {composable_kernel/client_example => client_example}/22_im2col_col2im/CMakeLists.txt (100%)
 rename {composable_kernel/client_example => client_example}/22_im2col_col2im/column_to_image.cpp (100%)
 rename {composable_kernel/client_example => client_example}/22_im2col_col2im/image_to_column.cpp (100%)
 rename {composable_kernel/client_example => client_example}/CMakeLists.txt (100%)
 rename {composable_kernel/client_example => client_example}/README.md (100%)
 rename {composable_kernel/cmake => cmake}/Analyzers.cmake (100%)
 rename {composable_kernel/cmake => cmake}/ClangTidy.cmake (100%)
 rename {composable_kernel/cmake => cmake}/CppCheck.cmake (100%)
 rename {composable_kernel/cmake => cmake}/DoxygenDoc.cmake (100%)
 rename {composable_kernel/cmake => cmake}/EnableCompilerWarnings.cmake (100%)
 rename {composable_kernel/cmake => cmake}/TargetFlags.cmake (100%)
 rename {composable_kernel/cmake => cmake}/googletest.cmake (100%)
 rename composable_kernel/dev-requirements.txt => dev-requirements.txt (100%)
 rename {composable_kernel/docs => docs}/API_Reference_Guide.rst (100%)
 rename {composable_kernel/docs => docs}/Contributors_Guide.rst (100%)
 rename {composable_kernel/docs => docs}/Supported_Primitives_Guide.rst (100%)
 rename {composable_kernel/docs => docs}/conf.py (100%)
 rename {composable_kernel/docs => docs}/data/ck_component.png (100%)
 rename {composable_kernel/docs => docs}/data/ck_layer.png (100%)
 rename {composable_kernel/docs => docs}/dockerhub.rst (100%)
 rename {composable_kernel/docs => docs}/doxygen/Doxyfile (100%)
 rename {composable_kernel/docs => docs}/index.rst (100%)
 rename {composable_kernel/docs => docs}/license.rst (100%)
 rename {composable_kernel/docs => docs}/refs.bib (100%)
 rename {composable_kernel/docs => docs}/sphinx/_toc.yml.in (100%)
 rename {composable_kernel/docs => docs}/sphinx/requirements.in (100%)
 rename {composable_kernel/docs => docs}/sphinx/requirements.txt (100%)
 rename {composable_kernel/docs => docs}/tutorial_hello_world.rst (100%)
 rename {composable_kernel/example => example}/01_gemm/CMakeLists.txt (100%)
 rename {composable_kernel/example => example}/01_gemm/README.md (100%)
 rename {composable_kernel/example => example}/01_gemm/common.hpp (100%)
 rename {composable_kernel/example => example}/01_gemm/gemm_dl_fp16.cpp (100%)
 rename {composable_kernel/example => example}/01_gemm/gemm_dl_fp32.cpp (100%)
 rename {composable_kernel/example => example}/01_gemm/gemm_dl_int4.cpp (100%)
 rename {composable_kernel/example => example}/01_gemm/gemm_dl_int8.cpp (100%)
 rename {composable_kernel/example => example}/01_gemm/gemm_dpp_fp16.cpp (100%)
 rename {composable_kernel/example => example}/01_gemm/gemm_wmma_fp16.cpp (100%)
 rename {composable_kernel/example => example}/01_gemm/gemm_xdl_bf16.cpp (100%)
 rename {composable_kernel/example => example}/01_gemm/gemm_xdl_bf16_rtn.cpp (100%)
 rename {composable_kernel/example => example}/01_gemm/gemm_xdl_fp16.cpp (100%)
 rename {composable_kernel/example => example}/01_gemm/gemm_xdl_fp16_fp8.cpp (100%)
 rename {composable_kernel/example => example}/01_gemm/gemm_xdl_fp64.cpp (100%)
 rename {composable_kernel/example => example}/01_gemm/gemm_xdl_fp8.cpp (100%)
 rename {composable_kernel/example => example}/01_gemm/gemm_xdl_fp8_bf8.cpp (100%)
 rename {composable_kernel/example => example}/01_gemm/gemm_xdl_int4.cpp (100%)
 rename {composable_kernel/example => example}/01_gemm/gemm_xdl_int8.cpp (100%)
 rename {composable_kernel/example => example}/01_gemm/gemm_xdl_skip_b_lds_fp16.cpp (100%)
 rename {composable_kernel/example => example}/01_gemm/gemm_xdl_streamk.cpp (100%)
 rename {composable_kernel/example => example}/01_gemm/gemm_xdl_wavelet_fp16.cpp (100%)
 rename {composable_kernel/example => example}/01_gemm/run_gemm_example.inc (100%)
 rename {composable_kernel/example => example}/02_gemm_bilinear/CMakeLists.txt (100%)
 rename {composable_kernel/example => example}/02_gemm_bilinear/README.md (100%)
 rename {composable_kernel/example => example}/02_gemm_bilinear/gemm_bilinear_wmma_fp16.cpp (100%)
 rename {composable_kernel/example => example}/02_gemm_bilinear/gemm_bilinear_wmma_int8.cpp (100%)
 rename {composable_kernel/example => example}/02_gemm_bilinear/gemm_bilinear_xdl_fp16.cpp (100%)
 rename {composable_kernel/example => example}/03_gemm_bias_relu/CMakeLists.txt (100%)
 rename {composable_kernel/example => example}/03_gemm_bias_relu/README.md (100%)
 rename {composable_kernel/example => example}/03_gemm_bias_relu/gemm_bias_relu_xdl_fp16.cpp (100%)
 rename {composable_kernel/example => example}/04_gemm_add_add_fastgelu/CMakeLists.txt (100%)
 rename {composable_kernel/example => example}/04_gemm_add_add_fastgelu/README.md (100%)
 rename {composable_kernel/example => example}/04_gemm_add_add_fastgelu/common.hpp (100%)
 rename {composable_kernel/example => example}/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_bf16.cpp (100%)
 rename {composable_kernel/example => example}/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_fp16.cpp (100%)
 rename {composable_kernel/example => example}/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_fp32.cpp (100%)
 rename {composable_kernel/example => example}/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_int4.cpp (100%)
 rename {composable_kernel/example => example}/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_int8.cpp (100%)
 rename {composable_kernel/example => example}/04_gemm_add_add_fastgelu/run_gemm_add_add_fastgelu_example.inc (100%)
 rename {composable_kernel/example => example}/09_convnd_fwd/CMakeLists.txt (100%)
 rename {composable_kernel/example => example}/09_convnd_fwd/README.md (100%)
 rename {composable_kernel/example => example}/09_convnd_fwd/convnd_fwd_common.hpp (100%)
 rename {composable_kernel/example => example}/09_convnd_fwd/convnd_fwd_dl_common.hpp (100%)
 rename {composable_kernel/example => example}/09_convnd_fwd/convnd_fwd_dl_fp16.cpp (100%)
 rename {composable_kernel/example => example}/09_convnd_fwd/convnd_fwd_dl_fp32.cpp (100%)
 rename {composable_kernel/example => example}/09_convnd_fwd/convnd_fwd_dl_int8.cpp (100%)
 rename {composable_kernel/example => example}/09_convnd_fwd/convnd_fwd_xdl_bf16.cpp (100%)
 rename {composable_kernel/example => example}/09_convnd_fwd/convnd_fwd_xdl_fp16.cpp (100%)
 rename {composable_kernel/example => example}/09_convnd_fwd/convnd_fwd_xdl_fp32.cpp (100%)
 rename {composable_kernel/example => example}/09_convnd_fwd/convnd_fwd_xdl_fp64.cpp (100%)
 rename {composable_kernel/example => example}/09_convnd_fwd/convnd_fwd_xdl_int8.cpp (100%)
 rename {composable_kernel/example => example}/09_convnd_fwd/run_convnd_fwd_dl_example.inc (100%)
 rename {composable_kernel/example => example}/09_convnd_fwd/run_convnd_fwd_example.inc (100%)
 rename {composable_kernel/example => example}/10_convnd_fwd_multiple_d_multiple_reduce/CMakeLists.txt (100%)
 rename {composable_kernel/example => example}/10_convnd_fwd_multiple_d_multiple_reduce/common.hpp (100%)
 rename {composable_kernel/example => example}/10_convnd_fwd_multiple_d_multiple_reduce/convnd_fwd_max_xdl_bf16.cpp (100%)
 rename {composable_kernel/example => example}/10_convnd_fwd_multiple_d_multiple_reduce/convnd_fwd_max_xdl_fp16.cpp (100%)
 rename {composable_kernel/example => example}/10_convnd_fwd_multiple_d_multiple_reduce/convnd_fwd_max_xdl_fp32.cpp (100%)
 rename {composable_kernel/example => example}/10_convnd_fwd_multiple_d_multiple_reduce/convnd_fwd_max_xdl_int4.cpp (100%)
 rename {composable_kernel/example => example}/10_convnd_fwd_multiple_d_multiple_reduce/convnd_fwd_max_xdl_int8.cpp (100%)
 rename {composable_kernel/example => example}/10_convnd_fwd_multiple_d_multiple_reduce/run_convnd_fwd_max_example.inc (100%)
 rename {composable_kernel/example => example}/12_reduce/CMakeLists.txt (100%)
 rename {composable_kernel/example => example}/12_reduce/README.md (100%)
 rename {composable_kernel/example => example}/12_reduce/reduce_blockwise.cpp (100%)
 rename {composable_kernel/example => example}/12_reduce/reduce_blockwise_impl.hpp (100%)
 rename {composable_kernel/example => example}/12_reduce/reduce_blockwise_two_call.cpp (100%)
 rename {composable_kernel/example => example}/12_reduce/reduce_example_common.hpp (100%)
 rename {composable_kernel/example => example}/12_reduce/reduce_multiblock_atomic_add.cpp (100%)
 rename {composable_kernel/example => example}/12_reduce/reduce_multiblock_atomic_add_impl.hpp (100%)
 rename {composable_kernel/example => example}/13_pool2d_fwd/CMakeLists.txt (100%)
 rename {composable_kernel/example => example}/13_pool2d_fwd/README.md (100%)
 rename {composable_kernel/example => example}/13_pool2d_fwd/pool2d_fwd_common.hpp (100%)
 rename {composable_kernel/example => example}/13_pool2d_fwd/pool2d_fwd_fp16.cpp (100%)
 rename {composable_kernel/example => example}/13_pool2d_fwd/pool2d_fwd_fp32.cpp (100%)
 rename {composable_kernel/example => example}/14_gemm_quantization/CMakeLists.txt (100%)
 rename {composable_kernel/example => example}/14_gemm_quantization/gemm_dl_quantization_int8.cpp (100%)
 rename {composable_kernel/example => example}/14_gemm_quantization/gemm_xdl_bias_relu_quantization_int8.cpp (100%)
 rename {composable_kernel/example => example}/14_gemm_quantization/gemm_xdl_quantization_int8.cpp (100%)
 rename {composable_kernel/example => example}/15_grouped_gemm/CMakeLists.txt (100%)
 rename {composable_kernel/example => example}/15_grouped_gemm/README.md (100%)
 rename {composable_kernel/example => example}/15_grouped_gemm/grouped_gemm_multiple_d_dl_fp16.cpp (100%)
 rename {composable_kernel/example => example}/15_grouped_gemm/grouped_gemm_xdl_bf16.cpp (100%)
 rename {composable_kernel/example => example}/15_grouped_gemm/grouped_gemm_xdl_fixed_nk_bias_fp16.cpp (100%)
 rename {composable_kernel/example => example}/15_grouped_gemm/grouped_gemm_xdl_fixed_nk_fp16.cpp (100%)
 rename {composable_kernel/example => example}/15_grouped_gemm/grouped_gemm_xdl_fixed_nk_fp8.cpp (100%)
 rename {composable_kernel/example => example}/15_grouped_gemm/grouped_gemm_xdl_fp16.cpp (100%)
 rename {composable_kernel/example => example}/15_grouped_gemm/grouped_gemm_xdl_fp32.cpp (100%)
 rename {composable_kernel/example => example}/15_grouped_gemm/grouped_gemm_xdl_int4.cpp (100%)
 rename {composable_kernel/example => example}/15_grouped_gemm/grouped_gemm_xdl_int8.cpp (100%)
 rename {composable_kernel/example => example}/15_grouped_gemm/grouped_gemm_xdl_splitk_fp16.cpp (100%)
 rename {composable_kernel/example => example}/15_grouped_gemm/run_grouped_gemm_example.inc (100%)
 rename {composable_kernel/example => example}/16_gemm_multi_d_multi_reduces/CMakeLists.txt (100%)
 rename {composable_kernel/example => example}/16_gemm_multi_d_multi_reduces/gemm_add_add_mean_meansquare_xdl_fp16.cpp (100%)
 rename {composable_kernel/example => example}/16_gemm_multi_d_multi_reduces/gemm_add_addsquare_xdl_int8.cpp (100%)
 rename {composable_kernel/example => example}/16_gemm_multi_d_multi_reduces/gemm_max_xdl_bf16.cpp (100%)
 rename {composable_kernel/example => example}/16_gemm_multi_d_multi_reduces/gemm_max_xdl_fp16.cpp (100%)
 rename {composable_kernel/example => example}/16_gemm_multi_d_multi_reduces/gemm_max_xdl_fp32.cpp (100%)
 rename {composable_kernel/example => example}/16_gemm_multi_d_multi_reduces/gemm_max_xdl_int4.cpp (100%)
 rename {composable_kernel/example => example}/16_gemm_multi_d_multi_reduces/gemm_max_xdl_int8.cpp (100%)
 rename {composable_kernel/example => example}/16_gemm_multi_d_multi_reduces/gemm_mean_meansquare_xdl_bf16.cpp (100%)
 rename {composable_kernel/example => example}/16_gemm_multi_d_multi_reduces/gemm_mean_meansquare_xdl_fp16.cpp (100%)
 rename {composable_kernel/example => example}/16_gemm_multi_d_multi_reduces/gemm_mean_meansquare_xdl_fp32.cpp (100%)
 rename {composable_kernel/example => example}/16_gemm_multi_d_multi_reduces/gemm_reduce_xdl_common.hpp (100%)
 rename {composable_kernel/example => example}/17_convnd_bwd_data/CMakeLists.txt (100%)
 rename {composable_kernel/example => example}/17_convnd_bwd_data/README.md (100%)
 rename {composable_kernel/example => example}/17_convnd_bwd_data/convnd_bwd_data_common.hpp (100%)
 rename {composable_kernel/example => example}/17_convnd_bwd_data/convnd_bwd_data_dl_fp16.cpp (100%)
 rename {composable_kernel/example => example}/17_convnd_bwd_data/convnd_bwd_data_xdl_fp16.cpp (100%)
 rename {composable_kernel/example => example}/18_batched_gemm_reduce/CMakeLists.txt (100%)
 rename {composable_kernel/example => example}/18_batched_gemm_reduce/batched_gemm_reduce_xdl_fp16.cpp (100%)
 rename {composable_kernel/example => example}/19_binary_elementwise/CMakeLists.txt (100%)
 rename {composable_kernel/example => example}/19_binary_elementwise/broadcast_add_2d_amn_bn.cpp (100%)
 rename {composable_kernel/example => example}/19_binary_elementwise/broadcast_add_3d_am_bmnk.cpp (100%)
 rename {composable_kernel/example => example}/19_binary_elementwise/elementwise_add_1d.cpp (100%)
 rename {composable_kernel/example => example}/19_binary_elementwise/elementwise_add_4d.cpp (100%)
 rename {composable_kernel/example => example}/20_grouped_conv_bwd_weight/CMakeLists.txt (100%)
 rename {composable_kernel/example => example}/20_grouped_conv_bwd_weight/common.hpp (100%)
 rename {composable_kernel/example => example}/20_grouped_conv_bwd_weight/grouped_conv_bwd_weight_dl_fp16.cpp (100%)
 rename {composable_kernel/example => example}/20_grouped_conv_bwd_weight/grouped_conv_bwd_weight_wmma_fp16.cpp (100%)
 rename {composable_kernel/example => example}/20_grouped_conv_bwd_weight/grouped_conv_bwd_weight_xdl_bf16.cpp (100%)
 rename {composable_kernel/example => example}/20_grouped_conv_bwd_weight/grouped_conv_bwd_weight_xdl_fp16.cpp (100%)
 rename {composable_kernel/example => example}/20_grouped_conv_bwd_weight/grouped_conv_bwd_weight_xdl_fp16_comp_bf8_fp8.cpp (100%)
 rename {composable_kernel/example => example}/20_grouped_conv_bwd_weight/run_grouped_conv_bwd_weight_example.inc (100%)
 rename {composable_kernel/example => example}/21_gemm_layernorm/CMakeLists.txt (100%)
 rename {composable_kernel/example => example}/21_gemm_layernorm/gemm_bias_relu_add_layernorm_xdl_naive_fp16.cpp (100%)
 rename {composable_kernel/example => example}/21_gemm_layernorm/gemm_bias_relu_add_layernorm_xdl_welford_fp16.cpp (100%)
 rename {composable_kernel/example => example}/21_gemm_layernorm/gemm_layernorm_xdl_naive_fp16.cpp (100%)
 rename {composable_kernel/example => example}/21_gemm_layernorm/gemm_xdl_layernorm_naive_single_kernel_fp16.cpp (100%)
 rename {composable_kernel/example => example}/22_cgemm/CMakeLists.txt (100%)
 rename {composable_kernel/example => example}/22_cgemm/cgemm_xdl_bf16.cpp (100%)
 rename {composable_kernel/example => example}/22_cgemm/cgemm_xdl_common.hpp (100%)
 rename {composable_kernel/example => example}/22_cgemm/cgemm_xdl_fp16.cpp (100%)
 rename {composable_kernel/example => example}/22_cgemm/cgemm_xdl_fp32.cpp (100%)
 rename {composable_kernel/example => example}/22_cgemm/cgemm_xdl_int4.cpp (100%)
 rename {composable_kernel/example => example}/22_cgemm/cgemm_xdl_int8.cpp (100%)
 rename {composable_kernel/example => example}/23_softmax/CMakeLists.txt (100%)
 rename {composable_kernel/example => example}/23_softmax/README.md (100%)
 rename {composable_kernel/example => example}/23_softmax/softmax_blockwise.cpp (100%)
 rename {composable_kernel/example => example}/24_batched_gemm/CMakeLists.txt (100%)
 rename {composable_kernel/example => example}/24_batched_gemm/batched_gemm_xdl_bf16.cpp (100%)
 rename {composable_kernel/example => example}/24_batched_gemm/batched_gemm_xdl_fp16.cpp (100%)
 rename {composable_kernel/example => example}/24_batched_gemm/batched_gemm_xdl_fp32.cpp (100%)
 rename {composable_kernel/example => example}/24_batched_gemm/batched_gemm_xdl_int4.cpp (100%)
 rename {composable_kernel/example => example}/24_batched_gemm/batched_gemm_xdl_int8.cpp (100%)
 rename {composable_kernel/example => example}/24_batched_gemm/run_batched_gemm_example.inc (100%)
 rename {composable_kernel/example => example}/25_gemm_bias_e_permute/CMakeLists.txt (100%)
 rename {composable_kernel/example => example}/25_gemm_bias_e_permute/gemm_bias_e_permute_g1m2n3k1_xdl_fp16.cpp (100%)
 rename {composable_kernel/example => example}/25_gemm_bias_e_permute/gemm_bias_e_permute_g1m3n2k1_xdl_fp16.cpp (100%)
 rename {composable_kernel/example => example}/26_contraction/CMakeLists.txt (100%)
 rename {composable_kernel/example => example}/26_contraction/README.md (100%)
 rename {composable_kernel/example => example}/26_contraction/contraction_bilinear_xdl_fp32.cpp (100%)
 rename {composable_kernel/example => example}/26_contraction/contraction_bilinear_xdl_fp64.cpp (100%)
 rename {composable_kernel/example => example}/26_contraction/contraction_scale_xdl_fp32.cpp (100%)
 rename {composable_kernel/example => example}/26_contraction/contraction_scale_xdl_fp64.cpp (100%)
 rename {composable_kernel/example => example}/27_layernorm/CMakeLists.txt (100%)
 rename {composable_kernel/example => example}/27_layernorm/common.hpp (100%)
 rename {composable_kernel/example => example}/27_layernorm/layernorm_fp16.cpp (100%)
 rename {composable_kernel/example => example}/27_layernorm/layernorm_splitk_fp16.cpp (100%)
 rename {composable_kernel/example => example}/27_layernorm/run_layernorm_example.inc (100%)
 rename {composable_kernel/example => example}/28_grouped_gemm_bias_e_permute/CMakeLists.txt (100%)
 rename {composable_kernel/example => example}/28_grouped_gemm_bias_e_permute/grouped_gemm_bias_e_permute_xdl_fp16.cpp (100%)
 rename {composable_kernel/example => example}/29_batched_gemm_bias_e_permute/CMakeLists.txt (100%)
 rename {composable_kernel/example => example}/29_batched_gemm_bias_e_permute/batched_gemm_bias_e_permute_wmma_fp16.cpp (100%)
 rename {composable_kernel/example => example}/29_batched_gemm_bias_e_permute/batched_gemm_bias_e_permute_xdl_fp16.cpp (100%)
 rename {composable_kernel/example => example}/30_grouped_conv_fwd_multiple_d/CMakeLists.txt (100%)
 rename {composable_kernel/example => example}/30_grouped_conv_fwd_multiple_d/README.md (100%)
 rename {composable_kernel/example => example}/30_grouped_conv_fwd_multiple_d/common.hpp (100%)
 rename {composable_kernel/example => example}/30_grouped_conv_fwd_multiple_d/common_wmma.hpp (100%)
 rename {composable_kernel/example => example}/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_wmma_fp16.cpp (100%)
 rename {composable_kernel/example => example}/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_wmma_int8.cpp (100%)
 rename {composable_kernel/example => example}/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_xdl_bf16.cpp (100%)
 rename {composable_kernel/example => example}/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_xdl_fp16.cpp (100%)
 rename {composable_kernel/example => example}/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_xdl_fp32.cpp (100%)
 rename {composable_kernel/example => example}/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_xdl_int4.cpp (100%)
 rename {composable_kernel/example => example}/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_xdl_int8.cpp (100%)
 rename {composable_kernel/example => example}/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_xdl_fp16.cpp (100%)
 rename {composable_kernel/example => example}/30_grouped_conv_fwd_multiple_d/run_grouped_conv_fwd_bias_relu_add_example.inc (100%)
 rename {composable_kernel/example => example}/30_grouped_conv_fwd_multiple_d/run_grouped_conv_fwd_bias_relu_add_wmma_example.inc (100%)
 rename {composable_kernel/example => example}/30_grouped_conv_fwd_multiple_d/run_grouped_conv_fwd_example.inc (100%)
 rename {composable_kernel/example => example}/31_batched_gemm_gemm/CMakeLists.txt (100%)
 rename {composable_kernel/example => example}/31_batched_gemm_gemm/batched_gemm_gemm_xdl_bf16.cpp (100%)
 rename {composable_kernel/example => example}/31_batched_gemm_gemm/batched_gemm_gemm_xdl_fp16.cpp (100%)
 rename {composable_kernel/example => example}/31_batched_gemm_gemm/batched_gemm_gemm_xdl_fp32.cpp (100%)
 rename {composable_kernel/example => example}/31_batched_gemm_gemm/batched_gemm_gemm_xdl_int4.cpp (100%)
 rename {composable_kernel/example => example}/31_batched_gemm_gemm/batched_gemm_gemm_xdl_int8.cpp (100%)
 rename {composable_kernel/example => example}/31_batched_gemm_gemm/run_batched_gemm_gemm_example.inc (100%)
 rename {composable_kernel/example => example}/32_batched_gemm_scale_softmax_gemm/CMakeLists.txt (100%)
 rename {composable_kernel/example => example}/32_batched_gemm_scale_softmax_gemm/batched_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16.cpp (100%)
 rename {composable_kernel/example => example}/32_batched_gemm_scale_softmax_gemm/batched_gemm_scale_softmax_gemm_permute_xdl_bf16.cpp (100%)
 rename {composable_kernel/example => example}/32_batched_gemm_scale_softmax_gemm/batched_gemm_scale_softmax_gemm_permute_xdl_fp16.cpp (100%)
 rename {composable_kernel/example => example}/32_batched_gemm_scale_softmax_gemm/batched_gemm_scale_softmax_gemm_xdl_bf16.cpp (100%)
 rename {composable_kernel/example => example}/32_batched_gemm_scale_softmax_gemm/batched_gemm_scale_softmax_gemm_xdl_fp16.cpp (100%)
 rename {composable_kernel/example => example}/32_batched_gemm_scale_softmax_gemm/grouped_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16.cpp (100%)
 rename {composable_kernel/example => example}/32_batched_gemm_scale_softmax_gemm/grouped_gemm_scale_softmax_gemm_permute_xdl_fp16.cpp (100%)
 rename {composable_kernel/example => example}/32_batched_gemm_scale_softmax_gemm/run_batched_gemm_scale_softmax_gemm.inc (100%)
 rename {composable_kernel/example => example}/32_batched_gemm_scale_softmax_gemm/run_batched_gemm_scale_softmax_gemm_permute.inc (100%)
 rename {composable_kernel/example => example}/32_batched_gemm_scale_softmax_gemm/run_grouped_gemm_scale_softmax_gemm_permute.inc (100%)
 rename {composable_kernel/example => example}/33_multiple_reduce/CMakeLists.txt (100%)
 rename {composable_kernel/example => example}/33_multiple_reduce/README.md (100%)
 rename {composable_kernel/example => example}/33_multiple_reduce/dual_reduce_common.hpp (100%)
 rename {composable_kernel/example => example}/33_multiple_reduce/dual_reduce_multiblock.cpp (100%)
 rename {composable_kernel/example => example}/33_multiple_reduce/dual_reduce_threadwise.cpp (100%)
 rename {composable_kernel/example => example}/34_batchnorm/CMakeLists.txt (100%)
 rename {composable_kernel/example => example}/34_batchnorm/README.md (100%)
 rename {composable_kernel/example => example}/34_batchnorm/batchnorm_backward_nhwc.cpp (100%)
 rename {composable_kernel/example => example}/34_batchnorm/batchnorm_common.hpp (100%)
 rename {composable_kernel/example => example}/34_batchnorm/batchnorm_forward_inferring_nhwc.cpp (100%)
 rename {composable_kernel/example => example}/34_batchnorm/batchnorm_forward_training_nhwc.cpp (100%)
 rename {composable_kernel/example => example}/34_batchnorm/batchnorm_forward_training_nhwc_obsolete.cpp (100%)
 rename {composable_kernel/example => example}/34_batchnorm/batchnorm_infer_impl.hpp (100%)
 rename {composable_kernel/example => example}/35_splitK_gemm/CMakeLists.txt (100%)
 rename {composable_kernel/example => example}/35_splitK_gemm/run_splitK_gemm_example.inc (100%)
 rename {composable_kernel/example => example}/35_splitK_gemm/splitK_gemm_xdl_bf16.cpp (100%)
 rename {composable_kernel/example => example}/35_splitK_gemm/splitK_gemm_xdl_fp16.cpp (100%)
 rename {composable_kernel/example => example}/35_splitK_gemm/splitK_gemm_xdl_fp32.cpp (100%)
 rename {composable_kernel/example => example}/35_splitK_gemm/splitK_gemm_xdl_int4.cpp (100%)
 rename {composable_kernel/example => example}/35_splitK_gemm/splitK_gemm_xdl_int8.cpp (100%)
 rename {composable_kernel/example => example}/36_sparse_embedding/CMakeLists.txt (100%)
 rename {composable_kernel/example => example}/36_sparse_embedding/sparse_embedding3_forward_layernorm.cpp (100%)
 rename {composable_kernel/example => example}/37_batched_gemm_add_add_relu_gemm_add/CMakeLists.txt (100%)
 rename {composable_kernel/example => example}/37_batched_gemm_add_add_relu_gemm_add/batched_gemm_add_add_relu_gemm_add_xdl_fp16.cpp (100%)
 rename {composable_kernel/example => example}/38_grouped_conv_bwd_data_multiple_d/CMakeLists.txt (100%)
 rename {composable_kernel/example => example}/38_grouped_conv_bwd_data_multiple_d/common.hpp (100%)
 rename {composable_kernel/example => example}/38_grouped_conv_bwd_data_multiple_d/grouped_conv_bwd_data_bias_relu_xdl_fp16.cpp (100%)
 rename {composable_kernel/example => example}/38_grouped_conv_bwd_data_multiple_d/grouped_conv_bwd_data_wmma_fp16.cpp (100%)
 rename {composable_kernel/example => example}/38_grouped_conv_bwd_data_multiple_d/grouped_conv_bwd_data_xdl_fp16.cpp (100%)
 rename {composable_kernel/example => example}/38_grouped_conv_bwd_data_multiple_d/run_grouped_conv_bwd_data_bias_relu_example.inc (100%)
 rename {composable_kernel/example => example}/38_grouped_conv_bwd_data_multiple_d/run_grouped_conv_bwd_data_example.inc (100%)
 rename {composable_kernel/example => example}/39_permute/CMakeLists.txt (100%)
 rename {composable_kernel/example => example}/39_permute/common.hpp (100%)
 rename {composable_kernel/example => example}/39_permute/permute_1xHxW_fp16.cpp (100%)
 rename {composable_kernel/example => example}/39_permute/permute_HxWx4_fp16.cpp (100%)
 rename {composable_kernel/example => example}/39_permute/permute_NxHxW_fp16.cpp (100%)
 rename {composable_kernel/example => example}/39_permute/run_permute_bundle_example.inc (100%)
 rename {composable_kernel/example => example}/39_permute/run_permute_element_example.inc (100%)
 rename {composable_kernel/example => example}/40_conv2d_fwd_quantization/CMakeLists.txt (100%)
 rename {composable_kernel/example => example}/40_conv2d_fwd_quantization/common.hpp (100%)
 rename {composable_kernel/example => example}/40_conv2d_fwd_quantization/conv2d_fwd_dl_bias_relu_perchannel_quantization_int8.cpp (100%)
 rename {composable_kernel/example => example}/40_conv2d_fwd_quantization/conv2d_fwd_dl_bias_relu_perlayer_quantization_int8.cpp (100%)
 rename {composable_kernel/example => example}/40_conv2d_fwd_quantization/conv2d_fwd_dl_bias_tanh_perchannel_quantization_int8.cpp (100%)
 rename {composable_kernel/example => example}/40_conv2d_fwd_quantization/conv2d_fwd_dl_bias_tanh_perlayer_quantization_int8.cpp (100%)
 rename {composable_kernel/example => example}/40_conv2d_fwd_quantization/conv2d_fwd_dl_perchannel_quantization_int8.cpp (100%)
 rename {composable_kernel/example => example}/40_conv2d_fwd_quantization/conv2d_fwd_dl_perlayer_quantization_int8.cpp (100%)
 rename {composable_kernel/example => example}/40_conv2d_fwd_quantization/conv2d_fwd_xdl_bias_relu_perchannel_quantization_int8.cpp (100%)
 rename {composable_kernel/example => example}/40_conv2d_fwd_quantization/conv2d_fwd_xdl_bias_relu_perlayer_quantization_int8.cpp (100%)
 rename {composable_kernel/example => example}/40_conv2d_fwd_quantization/conv2d_fwd_xdl_perchannel_quantization_int8.cpp (100%)
 rename {composable_kernel/example => example}/40_conv2d_fwd_quantization/conv2d_fwd_xdl_perlayer_quantization_int8.cpp (100%)
 rename {composable_kernel/example => example}/40_conv2d_fwd_quantization/run_conv2d_fwd_bias_perchannel_quantization_example.inc (100%)
 rename {composable_kernel/example => example}/40_conv2d_fwd_quantization/run_conv2d_fwd_bias_perlayer_quantization_example.inc (100%)
 rename {composable_kernel/example => example}/40_conv2d_fwd_quantization/run_conv2d_fwd_perchannel_quantization_example.inc (100%)
 rename {composable_kernel/example => example}/40_conv2d_fwd_quantization/run_conv2d_fwd_perlayer_quantization_example.inc (100%)
 rename {composable_kernel/example => example}/41_grouped_conv_conv_fwd/CMakeLists.txt (100%)
 rename {composable_kernel/example => example}/41_grouped_conv_conv_fwd/grouped_conv_conv_fwd_xdl_bf16.cpp (100%)
 rename {composable_kernel/example => example}/41_grouped_conv_conv_fwd/grouped_conv_conv_fwd_xdl_fp16.cpp (100%)
 rename {composable_kernel/example => example}/41_grouped_conv_conv_fwd/grouped_conv_conv_fwd_xdl_fp32.cpp (100%)
 rename {composable_kernel/example => example}/41_grouped_conv_conv_fwd/grouped_conv_conv_fwd_xdl_int4.cpp (100%)
 rename {composable_kernel/example => example}/41_grouped_conv_conv_fwd/grouped_conv_conv_fwd_xdl_int8.cpp (100%)
 rename {composable_kernel/example => example}/41_grouped_conv_conv_fwd/run_grouped_conv_conv_fwd_example.inc (100%)
 rename {composable_kernel/example => example}/42_groupnorm/CMakeLists.txt (100%)
 rename {composable_kernel/example => example}/42_groupnorm/common.hpp (100%)
 rename {composable_kernel/example => example}/42_groupnorm/groupnorm_sigmoid_mul_fp16.cpp (100%)
 rename {composable_kernel/example => example}/42_groupnorm/groupnorm_splitk_fp16.cpp (100%)
 rename {composable_kernel/example => example}/42_groupnorm/groupnorm_swish_fp16.cpp (100%)
 rename {composable_kernel/example => example}/42_groupnorm/run_groupnorm_example.inc (100%)
 rename {composable_kernel/example => example}/43_splitk_gemm_bias_e_permute/CMakeLists.txt (100%)
 rename {composable_kernel/example => example}/43_splitk_gemm_bias_e_permute/splitk_gemm_bias_e_permute_xdl_fp16.cpp (100%)
 rename {composable_kernel/example => example}/43_splitk_gemm_bias_e_permute/splitk_gemm_bias_e_permute_xdl_fp32.cpp (100%)
 rename {composable_kernel/example => example}/44_elementwise_permute/CMakeLists.txt (100%)
 rename {composable_kernel/example => example}/44_elementwise_permute/elementwise_permute_4D_fp16.cpp (100%)
 rename {composable_kernel/example => example}/44_elementwise_permute/elementwise_permute_4D_fp16_2d.cpp (100%)
 rename {composable_kernel/example => example}/45_elementwise_normalization/CMakeLists.txt (100%)
 rename {composable_kernel/example => example}/45_elementwise_normalization/elementwise_layernorm_blockwise.cpp (100%)
 rename {composable_kernel/example => example}/46_gemm_add_multiply/CMakeLists.txt (100%)
 rename {composable_kernel/example => example}/46_gemm_add_multiply/README.md (100%)
 rename {composable_kernel/example => example}/46_gemm_add_multiply/common.hpp (100%)
 rename {composable_kernel/example => example}/46_gemm_add_multiply/gemm_add_multiply_dl_fp16.cpp (100%)
 rename {composable_kernel/example => example}/46_gemm_add_multiply/gemm_add_multiply_xdl_fp16.cpp (100%)
 rename {composable_kernel/example => example}/46_gemm_add_multiply/run_gemm_add_multiply_example.inc (100%)
 rename {composable_kernel/example => example}/47_gemm_bias_softmax_gemm_permute/CMakeLists.txt (100%)
 rename {composable_kernel/example => example}/47_gemm_bias_softmax_gemm_permute/gemm_bias_softmax_gemm_permute.cpp (100%)
 rename {composable_kernel/example => example}/48_pool3d_fwd/CMakeLists.txt (100%)
 rename {composable_kernel/example => example}/48_pool3d_fwd/pool3d_fwd_common.hpp (100%)
 rename {composable_kernel/example => example}/48_pool3d_fwd/pool3d_fwd_fp16.cpp (100%)
 rename {composable_kernel/example => example}/49_maxpool2d_bwd/CMakeLists.txt (100%)
 rename {composable_kernel/example => example}/49_maxpool2d_bwd/maxpool2d_bwd_bf16.cpp (100%)
 rename {composable_kernel/example => example}/49_maxpool2d_bwd/maxpool2d_bwd_common.hpp (100%)
 rename {composable_kernel/example => example}/49_maxpool2d_bwd/maxpool2d_bwd_fp16.cpp (100%)
 rename {composable_kernel/example => example}/49_maxpool2d_bwd/maxpool2d_bwd_fp32.cpp (100%)
 rename {composable_kernel/example => example}/50_put_element/CMakeLists.txt (100%)
 rename {composable_kernel/example => example}/50_put_element/put_element_fp16.cpp (100%)
 rename {composable_kernel/example => example}/51_avgpool3d_bwd/CMakeLists.txt (100%)
 rename {composable_kernel/example => example}/51_avgpool3d_bwd/avgpool3d_bwd_bf16.cpp (100%)
 rename {composable_kernel/example => example}/51_avgpool3d_bwd/avgpool3d_bwd_common.hpp (100%)
 rename {composable_kernel/example => example}/51_avgpool3d_bwd/avgpool3d_bwd_fp16.cpp (100%)
 rename {composable_kernel/example => example}/51_avgpool3d_bwd/avgpool3d_bwd_fp32.cpp (100%)
 rename {composable_kernel/example => example}/52_im2col_col2im/CMakeLists.txt (100%)
 rename {composable_kernel/example => example}/52_im2col_col2im/column_to_image_f32.cpp (100%)
 rename {composable_kernel/example => example}/52_im2col_col2im/common.hpp (100%)
 rename {composable_kernel/example => example}/52_im2col_col2im/image_to_column_f32.cpp (100%)
 rename {composable_kernel/example => example}/53_gemv_splitk/CMakeLists.txt (100%)
 rename {composable_kernel/example => example}/53_gemv_splitk/README.md (100%)
 rename {composable_kernel/example => example}/53_gemv_splitk/common.hpp (100%)
 rename {composable_kernel/example => example}/53_gemv_splitk/gemv_splitk_fp16.cpp (100%)
 rename {composable_kernel/example => example}/53_gemv_splitk/run_gemv_splitk_example.inc (100%)
 rename {composable_kernel/example => example}/54_tall_and_skinny_gemm_splitk/CMakeLists.txt (100%)
 rename {composable_kernel/example => example}/54_tall_and_skinny_gemm_splitk/README.md (100%)
 rename {composable_kernel/example => example}/54_tall_and_skinny_gemm_splitk/common.hpp (100%)
 rename {composable_kernel/example => example}/54_tall_and_skinny_gemm_splitk/run_tall_and_skinny_gemm_splitk_example.inc (100%)
 rename {composable_kernel/example => example}/54_tall_and_skinny_gemm_splitk/tall_and_skinny_gemm_splitk_fp16.cpp (100%)
 rename {composable_kernel/example => example}/60_gemm_multi_ABD/CMakeLists.txt (100%)
 rename {composable_kernel/example => example}/60_gemm_multi_ABD/gemm_multi_ABD_xdl_fp16.cpp (100%)
 rename {composable_kernel/example => example}/61_contraction_multi_ABD/CMakeLists.txt (100%)
 rename {composable_kernel/example => example}/61_contraction_multi_ABD/contraction_multi_ABD_xdl_fp16.cpp (100%)
 rename {composable_kernel/example => example}/62_conv_fwd_activ/CMakeLists.txt (100%)
 rename {composable_kernel/example => example}/62_conv_fwd_activ/convnd_fwd_activ_common.hpp (100%)
 rename {composable_kernel/example => example}/62_conv_fwd_activ/convnd_fwd_xdl_abs_fp16.cpp (100%)
 rename {composable_kernel/example => example}/62_conv_fwd_activ/convnd_fwd_xdl_clippedrelu_fp16.cpp (100%)
 rename {composable_kernel/example => example}/62_conv_fwd_activ/convnd_fwd_xdl_elu_fp16.cpp (100%)
 rename {composable_kernel/example => example}/62_conv_fwd_activ/convnd_fwd_xdl_leakyrelu_fp16.cpp (100%)
 rename {composable_kernel/example => example}/62_conv_fwd_activ/convnd_fwd_xdl_pow_fp16.cpp (100%)
 rename {composable_kernel/example => example}/62_conv_fwd_activ/convnd_fwd_xdl_relu_fp16.cpp (100%)
 rename {composable_kernel/example => example}/62_conv_fwd_activ/convnd_fwd_xdl_sigmoid_fp16.cpp (100%)
 rename {composable_kernel/example => example}/62_conv_fwd_activ/convnd_fwd_xdl_softrelu_fp16.cpp (100%)
 rename {composable_kernel/example => example}/62_conv_fwd_activ/convnd_fwd_xdl_tanh_fp16.cpp (100%)
 rename {composable_kernel/example => example}/62_conv_fwd_activ/run_convnd_fwd_activ_example.inc (100%)
 rename {composable_kernel/example => example}/CMakeLists.txt (100%)
 rename {composable_kernel/include => include}/ck/ck.hpp (100%)
 rename {composable_kernel/include => include}/ck/config.h.in (100%)
 rename {composable_kernel/include => include}/ck/host_utility/device_prop.hpp (100%)
 rename {composable_kernel/include => include}/ck/host_utility/hip_check_error.hpp (100%)
 rename {composable_kernel/include => include}/ck/host_utility/io.hpp (100%)
 rename {composable_kernel/include => include}/ck/host_utility/kernel_launch.hpp (100%)
 rename {composable_kernel/include => include}/ck/host_utility/stream_utility.hpp (100%)
 rename {composable_kernel/include => include}/ck/problem_transform/transform_forward_convolution3d_into_gemm_v4r4r4_ndhwc_kzyxc_ndhwk.hpp (100%)
 rename {composable_kernel/include => include}/ck/stream_config.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor/static_tensor.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_description/cluster_descriptor.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_description/multi_index_transform.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_description/multi_index_transform_helper.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_description/tensor_adaptor.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_description/tensor_descriptor.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_description/tensor_descriptor_helper.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_description/tensor_space_filling_curve.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/block/blockwise_gemm_dl_v2r3.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/block/blockwise_gemm_dlops_v2r2.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/block/blockwise_gemm_dlops_v3.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/block/blockwise_gemm_dpp.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/block/blockwise_gemm_wmma.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/block/blockwise_gemm_xdlops_skip_b_lds.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/block/blockwise_softmax.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/block/blockwise_tensor_slice_transfer_v5r1.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/block/blockwise_welford.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/block/reduction_functions_blockwise.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r1.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r1r2.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r2.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r3.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v7.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v7r2.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/conv_tensor_rearrange_op.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/convolution_backward_data_specialization.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/convolution_backward_weight_specialization.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/device_avgpool_bwd.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/device_base.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/device_batched_contraction_multiple_d.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/device_batched_gemm.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/device_batched_gemm_e_permute.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/device_batched_gemm_gemm.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/device_batched_gemm_multi_d.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/device_batched_gemm_multiple_d_gemm_multiple_d.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/device_batched_gemm_softmax_gemm.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/device_batched_gemm_softmax_gemm_permute.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/device_batchnorm_backward.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/device_batchnorm_forward.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/device_batchnorm_infer.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/device_cgemm.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/device_contraction_multiple_abd.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/device_contraction_multiple_d.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/device_conv_bwd_data.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/device_conv_fwd.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/device_conv_fwd_bias_activation.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/device_conv_fwd_bias_activation_add.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/device_conv_tensor_rearrange.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/device_elementwise.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/device_elementwise_normalization.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/device_gemm.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/device_gemm_bias_e_permute.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/device_gemm_multiple_abd.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/device_gemm_multiple_d_layernorm.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/device_gemm_multiple_d_multiple_r.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/device_gemm_reduce.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/device_gemm_splitk.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/device_gemm_streamk.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/device_grouped_contraction_multiple_d.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/device_grouped_conv_bwd_data_multiple_d.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/device_grouped_conv_bwd_weight.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/device_grouped_conv_fwd.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_d.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/device_grouped_gemm.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/device_grouped_gemm_fixed_nk.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/device_grouped_gemm_softmax_gemm_permute.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/device_grouped_gemm_splitk.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/device_max_pool_bwd.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/device_multiple_reduce.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/device_normalization.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/device_permute.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/device_pool_fwd.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/device_put_element.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/device_reduce.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/device_softmax.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/device_splitk_contraction_multiple_d.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/device_tall_and_skinny_gemm.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/gemm_specialization.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/impl/device_avgpool3d_bwd_ndhwc_ndhwc.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/impl/device_batched_contraction_multiple_d_wmma_cshuffle.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/impl/device_batched_contraction_multiple_d_xdl_cshuffle.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/impl/device_batched_gemm_e_permute_xdl.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/impl/device_batched_gemm_gemm_xdl_cshuffle.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/impl/device_batched_gemm_multi_d_xdl.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_dl.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_gemm_multiple_d_xdl_cshuffle.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/impl/device_batched_gemm_reduce_xdl_cshuffle.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_xdl_cshuffle.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/impl/device_batched_gemm_xdl.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/impl/device_batchnorm_backward_impl.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/impl/device_batchnorm_forward_impl.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/impl/device_batchnorm_forward_impl_obsolete.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/impl/device_cgemm_4gemm_xdl_cshuffle.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/impl/device_column_to_image_impl.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/impl/device_contraction_multiple_abd_xdl_cshuffle.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/impl/device_contraction_multiple_d_xdl_cshuffle.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/impl/device_conv2d_backward_weight_xdl_c_shuffle_nhwc_kyxc_nhwk.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/impl/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/impl/device_conv2d_fwd_xdl_c_shuffle_bias_activation_add_nhwc_kyxc_nhwk.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/impl/device_conv2d_fwd_xdl_c_shuffle_bias_activation_nhwc_kyxc_nhwk.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/impl/device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/impl/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/impl/device_conv3d_fwd_naive_ndhwc_kzyxc_ndhwk.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/impl/device_conv3d_fwd_xdl_ndhwc_kzyxc_ndhwk.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/impl/device_convnd_bwd_data_nwc_kxc_nwk_dl.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/impl/device_convnd_bwd_data_nwc_kxc_nwk_xdl.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/impl/device_elementwise_2d_impl.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/impl/device_elementwise_impl.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/impl/device_elementwise_normalization_impl.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/impl/device_gemm_bias_add_reduce_xdl_cshuffle.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/impl/device_gemm_dl.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/impl/device_gemm_dpp.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_abd_xdl_cshuffle.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_dl.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_layernorm_xdl_cshuffle.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_multiple_r_xdl_cshuffle.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_wmma_cshuffle.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/impl/device_gemm_reduce_xdl_cshuffle.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/impl/device_gemm_wmma.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/impl/device_gemm_xdl.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_layernorm_cshuffle.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_skip_b_lds.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_splitk_c_shuffle.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_streamk.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_waveletmodel_cshuffle.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/impl/device_grouped_contraction_multiple_d_xdl_cshuffle.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_data_multiple_d_wmma_cshuffle.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_data_multiple_d_xdl_cshuffle_v1.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_dl.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_wmma_cshuffle.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_xdl_cshuffle.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_dl_nhwc_kyxc_nhwk.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_multiple_r.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_multiple_r_xdl_cshuffle.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_wmma_cshuffle.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/impl/device_grouped_conv_utils.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_multiple_d_dl.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_softmax_gemm_permute_xdl_cshuffle.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl_fixed_nk.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl_splitk_cshuffle.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/impl/device_image_to_column_impl.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/impl/device_max_pool_bwd_impl.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/impl/device_multiple_reduce_multiblock.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/impl/device_multiple_reduce_threadwise.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/impl/device_normalization_impl.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/impl/device_normalization_splitk_impl.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/impl/device_permute_impl.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/impl/device_pool2d_fwd_nhwc_nhwc.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/impl/device_pool3d_fwd_ndhwc_ndhwc.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/impl/device_put_element_impl.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/impl/device_reduce_common.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/impl/device_reduce_multiblock.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/impl/device_reduce_threadwise.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/impl/device_softmax_impl.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/impl/device_sparse_embeddings_forward_layernorm.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/impl/device_splitk_contraction_multiple_d_xdl_cshuffle.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/impl/device_tall_and_skinny_gemm_splitk.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/masking_specialization.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/matrix_padder.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/tensor_layout.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/tensor_specialization.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/device/welford_helper.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/element/element_wise_operation.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/element/quantization_operation.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_batchnorm_forward.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_reduce_second_half_batchnorm_backward_final.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_welford_first_half.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_welford_second_half_batchnorm_forward_final_obsolete.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_welford_second_half_multiblock_reduce_first_half.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/grid/gemm_layernorm/gridwise_gemm_multiple_d_welford_first_half_xdl_cshuffle.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/grid/gemm_layernorm/gridwise_welford_second_half_layernorm2d.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/grid/gridwise_2d_multiple_reduction_multiblock.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/grid/gridwise_2d_multiple_reduction_threadwise.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/grid/gridwise_2d_reduction_multiblock.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/grid/gridwise_2d_reduction_threadwise.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_gemm_xdl_cshuffle_v1.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_multiple_d_gemm_multiple_d_xdl_cshuffle_v1.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_multiple_d_softmax_gemm_xdl_cshuffle_v1.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_softmax_gemm_xdl_cshuffle_v1.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/grid/gridwise_batchnorm_backward_blockwise_welford.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/grid/gridwise_batchnorm_forward_blockwise_welford.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/grid/gridwise_elementwise_1d.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/grid/gridwise_elementwise_2d.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/grid/gridwise_elementwise_layernorm_welford_variance.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/grid/gridwise_gemm_bias_add_reduce_xdl_cshuffle_v1.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/grid/gridwise_gemm_dl_multiple_d.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/grid/gridwise_gemm_dl_v1r3.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/grid/gridwise_gemm_dpp.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_abd_xdl_cshuffle.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_multiple_r_xdl_cshuffle.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_wmma_cshuffle.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_xdl_cshuffle.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_xdl_splitk_cshuffle.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_selector.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v1.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v2.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v3.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/grid/gridwise_gemm_reduce_xdl_cshuffle_v1.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/grid/gridwise_gemm_split_k_multiple_d_xdl_cshuffle.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/grid/gridwise_gemm_split_k_multiple_d_xdl_cshuffle_v2.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/grid/gridwise_gemm_waveletmodel.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/grid/gridwise_gemm_wmma.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v1.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_layernorm_cshuffle_v1.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_waveletmodel_cshuffle.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_bwd_weight.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_skip_b_lds_v1.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_streamk.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r3.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r4.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r4r2.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r1.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r2.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r3.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/grid/gridwise_gemv_splitk.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/grid/gridwise_permute.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/grid/gridwise_put_element_1d.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/grid/gridwise_set_buffer_value.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/grid/gridwise_set_multiple_buffer_value.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/grid/gridwise_softmax.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/grid/gridwise_sparse_embeddings_forward_layernorm.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/grid/gridwise_tall_and_skinny_gemm_splitk.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/grid/gridwise_tensor_rearrange.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/grid/normalization/gridwise_normalization_naive_variance.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/grid/normalization/gridwise_normalization_selector.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/grid/normalization/gridwise_normalization_splitk_1st.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/grid/normalization/gridwise_normalization_splitk_2nd.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/grid/normalization/gridwise_normalization_welford_variance.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/thread/reduction_functions_threadwise.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/thread/threadwise_contraction_dl.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/thread/threadwise_gemm_dlops_v3.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_set.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v3r1.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v4r1.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v5r1.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v6r1.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v6r1r2.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v6r2.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v6r3.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v7.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v7r2.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/thread/threadwise_welford.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/warp/dpp_gemm.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/warp/wmma_gemm.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/gpu/warp/xdlops_gemm.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/operator_transform/transform_contraction_to_gemm.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/operator_transform/transform_conv_bwd_data_to_gemm_v1.hpp (100%)
 rename {composable_kernel/include => include}/ck/tensor_operation/operator_transform/transform_conv_fwd_to_gemm.hpp (100%)
 rename {composable_kernel/include => include}/ck/utility/amd_address_space.hpp (100%)
 rename {composable_kernel/include => include}/ck/utility/amd_buffer_addressing.hpp (100%)
 rename {composable_kernel/include => include}/ck/utility/amd_gemm_dpp.hpp (100%)
 rename {composable_kernel/include => include}/ck/utility/amd_inline_asm.hpp (100%)
 rename {composable_kernel/include => include}/ck/utility/amd_wave_read_first_lane.hpp (100%)
 rename {composable_kernel/include => include}/ck/utility/amd_wmma.hpp (100%)
 rename {composable_kernel/include => include}/ck/utility/amd_xdlops.hpp (100%)
 rename {composable_kernel/include => include}/ck/utility/array.hpp (100%)
 rename {composable_kernel/include => include}/ck/utility/array_multi_index.hpp (100%)
 rename {composable_kernel/include => include}/ck/utility/c_style_pointer_cast.hpp (100%)
 rename {composable_kernel/include => include}/ck/utility/common_header.hpp (100%)
 rename {composable_kernel/include => include}/ck/utility/container_element_picker.hpp (100%)
 rename {composable_kernel/include => include}/ck/utility/container_helper.hpp (100%)
 rename {composable_kernel/include => include}/ck/utility/data_type.hpp (100%)
 rename {composable_kernel/include => include}/ck/utility/debug.hpp (100%)
 rename {composable_kernel/include => include}/ck/utility/dynamic_buffer.hpp (100%)
 rename {composable_kernel/include => include}/ck/utility/enable_if.hpp (100%)
 rename {composable_kernel/include => include}/ck/utility/f8_utils.hpp (100%)
 rename {composable_kernel/include => include}/ck/utility/functional.hpp (100%)
 rename {composable_kernel/include => include}/ck/utility/functional2.hpp (100%)
 rename {composable_kernel/include => include}/ck/utility/functional3.hpp (100%)
 rename {composable_kernel/include => include}/ck/utility/functional4.hpp (100%)
 rename {composable_kernel/include => include}/ck/utility/generic_memory_space_atomic.hpp (100%)
 rename {composable_kernel/include => include}/ck/utility/get_id.hpp (100%)
 rename {composable_kernel/include => include}/ck/utility/get_shift.hpp (100%)
 rename {composable_kernel/include => include}/ck/utility/ignore.hpp (100%)
 rename {composable_kernel/include => include}/ck/utility/inner_product.hpp (100%)
 rename {composable_kernel/include => include}/ck/utility/inner_product_dpp8.hpp (100%)
 rename {composable_kernel/include => include}/ck/utility/integral_constant.hpp (100%)
 rename {composable_kernel/include => include}/ck/utility/is_detected.hpp (100%)
 rename {composable_kernel/include => include}/ck/utility/is_known_at_compile_time.hpp (100%)
 rename {composable_kernel/include => include}/ck/utility/loop_scheduler.hpp (100%)
 rename {composable_kernel/include => include}/ck/utility/magic_division.hpp (100%)
 rename {composable_kernel/include => include}/ck/utility/math.hpp (100%)
 rename {composable_kernel/include => include}/ck/utility/math_v2.hpp (100%)
 rename {composable_kernel/include => include}/ck/utility/multi_index.hpp (100%)
 rename {composable_kernel/include => include}/ck/utility/number.hpp (100%)
 rename {composable_kernel/include => include}/ck/utility/random_gen.hpp (100%)
 rename {composable_kernel/include => include}/ck/utility/reduction_common.hpp (100%)
 rename {composable_kernel/include => include}/ck/utility/reduction_enums.hpp (100%)
 rename {composable_kernel/include => include}/ck/utility/reduction_functions_accumulate.hpp (100%)
 rename {composable_kernel/include => include}/ck/utility/reduction_operator.hpp (100%)
 rename {composable_kernel/include => include}/ck/utility/sequence.hpp (100%)
 rename {composable_kernel/include => include}/ck/utility/sequence_helper.hpp (100%)
 rename {composable_kernel/include => include}/ck/utility/span.hpp (100%)
 rename {composable_kernel/include => include}/ck/utility/static_buffer.hpp (100%)
 rename {composable_kernel/include => include}/ck/utility/statically_indexed_array.hpp (100%)
 rename {composable_kernel/include => include}/ck/utility/statically_indexed_array_multi_index.hpp (100%)
 rename {composable_kernel/include => include}/ck/utility/synchronization.hpp (100%)
 rename {composable_kernel/include => include}/ck/utility/thread_group.hpp (100%)
 rename {composable_kernel/include => include}/ck/utility/transpose_vectors.hpp (100%)
 rename {composable_kernel/include => include}/ck/utility/tuple.hpp (100%)
 rename {composable_kernel/include => include}/ck/utility/tuple_helper.hpp (100%)
 rename {composable_kernel/include => include}/ck/utility/type.hpp (100%)
 rename {composable_kernel/include => include}/ck/utility/type_convert.hpp (100%)
 rename {composable_kernel/include => include}/ck/utility/workgroup_barrier.hpp (100%)
 rename {composable_kernel/include => include}/ck/utility/workgroup_synchronization.hpp (100%)
 rename {composable_kernel/include => include}/ck/version.h.in (100%)
 rename {composable_kernel/library => library}/CMakeLists.txt (100%)
 rename {composable_kernel/library => library}/include/ck/library/reference_tensor_operation/cpu/reference_avgpool_bwd.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/reference_tensor_operation/cpu/reference_batchnorm_backward.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/reference_tensor_operation/cpu/reference_batchnorm_forward.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/reference_tensor_operation/cpu/reference_batchnorm_infer.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/reference_tensor_operation/cpu/reference_cgemm.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/reference_tensor_operation/cpu/reference_column_to_image.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/reference_tensor_operation/cpu/reference_contraction.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/reference_tensor_operation/cpu/reference_conv_bwd_data.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/reference_tensor_operation/cpu/reference_conv_bwd_weight.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/reference_tensor_operation/cpu/reference_conv_fwd.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/reference_tensor_operation/cpu/reference_conv_fwd_bias_activation.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/reference_tensor_operation/cpu/reference_conv_fwd_bias_activation_add.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/reference_tensor_operation/cpu/reference_gemm.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/reference_tensor_operation/cpu/reference_gemm_layernorm.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/reference_tensor_operation/cpu/reference_groupnorm.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/reference_tensor_operation/cpu/reference_image_to_column.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/reference_tensor_operation/cpu/reference_layernorm.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/reference_tensor_operation/cpu/reference_maxpool_bwd.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/reference_tensor_operation/cpu/reference_pool_fwd.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/reference_tensor_operation/cpu/reference_reduce.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/reference_tensor_operation/cpu/reference_softmax.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/reference_tensor_operation/cpu/reference_sparse_embedding3_forward_layernorm.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/reference_tensor_operation/gpu/naive_conv_fwd.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/add_device_operation_instance.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/device_operation_instance_factory.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/avg_pool3d_bwd.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/batched_gemm.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/batched_gemm_add_relu_gemm_add.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/batched_gemm_bias_permute.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/batched_gemm_bias_softmax_gemm_permute.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/batched_gemm_gemm.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/batched_gemm_multi_d.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/batched_gemm_softmax_gemm.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/batchnorm_backward.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/batchnorm_forward.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/batchnorm_infer.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/contraction_bilinear.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/contraction_scale.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/conv_tensor_rearrange.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/conv_tensor_rearrange/device_column_to_image_instance.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/conv_tensor_rearrange/device_image_to_column_instance.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/convolution_backward_data.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/convolution_forward.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/device_elementwise_instance.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/device_gemm_mean_squaremean_instance.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/elementwise_normalization.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/gemm.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/gemm_add_add_fastgelu.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/gemm_add_fastgelu.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/gemm_add_multiply.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/gemm_add_relu_add_layernorm.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/gemm_bilinear.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/gemm_fastgelu.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/gemm_multiply_add.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/gemm_splitk.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/gemm_streamk.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/gemv_splitk.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_data/device_grouped_conv_bwd_data_wmma_instance.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_data/device_grouped_conv_bwd_data_xdl_instance.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_dl_instance.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_wmma_instance.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_xdl_instance.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv2d_fwd_wmma_instance.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_dl_instance.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_wmma_instance.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_instance.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_data.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_weight.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/grouped_gemm.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/grouped_gemm_bias.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/grouped_gemm_fastgelu.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/grouped_gemm_fixed_nk.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/image_to_column.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/max_pool_bwd.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/normalization.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/normalization_swish.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/pool3d_fwd.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/quantization/gemm_quantization.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_bias_forward_perchannel_quantization.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_bias_forward_perlayer_quantization.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_forward_perchannel_quantization.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_forward_perlayer_quantization.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_add.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_amax.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_avg.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_max.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_min.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_norm2.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_amax.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_max.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_min.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_add.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_avg.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_norm2.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_add.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_amax.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_avg.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_max.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_min.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_norm2.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_add.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_avg.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_norm2.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_add.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_amax.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_avg.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_max.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_min.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_norm2.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i32_i8_add.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i32_i8_avg.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_amax.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_max.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_min.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_impl_common.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_b16_f32_f32_add.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_b16_f32_f32_avg.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f16_f32_f32_add.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f16_f32_f32_avg.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f32_f32_add.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f32_f32_avg.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f64_f32_add.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f64_f32_avg.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f64_f64_f64_add.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f64_f64_f64_avg.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_add.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_amax.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_avg.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_max.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_min.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_norm2.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_amax.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_max.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_min.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_add.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_avg.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_norm2.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_add.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_amax.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_avg.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_max.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_min.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_norm2.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_add.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_avg.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_norm2.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_add.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_amax.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_avg.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_max.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_min.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_norm2.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i32_i8_add.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i32_i8_avg.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_amax.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_max.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_min.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/reduce/reduce.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/softmax.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce1.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce2.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce3.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce1.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce2.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce3.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce4.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_type.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce1.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce2.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce3.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce1.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce2.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce3.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce4.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_type.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_instance.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/tensor_operation_instance/gpu/tall_and_skinny_gemm_splitk.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/utility/algorithm.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/utility/check_err.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/utility/conv_common.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/utility/convolution_host_tensor_descriptor_helper.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/utility/convolution_parameter.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/utility/device_memory.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/utility/fill.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/utility/host_common_util.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/utility/host_gemm.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/utility/host_tensor.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/utility/host_tensor_generator.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/utility/iterator.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/utility/literals.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/utility/numeric.hpp (100%)
 rename {composable_kernel/library => library}/include/ck/library/utility/ranges.hpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/CMakeLists.txt (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/avg_pool3d_bwd/CMakeLists.txt (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/avg_pool3d_bwd/avg_pool3d_bwd_ndhwc_instance_common.hpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/avg_pool3d_bwd/device_avg_pool3d_bwd_ndhwc_bf16_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/avg_pool3d_bwd/device_avg_pool3d_bwd_ndhwc_f16_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/avg_pool3d_bwd/device_avg_pool3d_bwd_ndhwc_f32_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/batched_gemm/CMakeLists.txt (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_bf16_bf16_bf16_gkm_gkn_gmn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_bf16_bf16_bf16_gkm_gnk_gmn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_bf16_bf16_bf16_gmk_gkn_gmn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_bf16_bf16_bf16_gmk_gnk_gmn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f16_f16_f16_gkm_gkn_gmn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f16_f16_f16_gkm_gnk_gmn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f16_f16_f16_gmk_gkn_gmn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f16_f16_f16_gmk_gnk_gmn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f32_f32_f32_gkm_gkn_gmn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f32_f32_f32_gkm_gnk_gmn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f32_f32_f32_gmk_gkn_gmn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f32_f32_f32_gmk_gnk_gmn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_int8_int8_int8_gkm_gkn_gmn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_int8_int8_int8_gkm_gnk_gmn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_int8_int8_int8_gmk_gkn_gmn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_int8_int8_int8_gmk_gnk_gmn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/batched_gemm_add_relu_gemm_add/CMakeLists.txt (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/batched_gemm_add_relu_gemm_add/device_batched_gemm_add_relu_gemm_add_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/batched_gemm_add_relu_gemm_add/device_batched_gemm_add_relu_gemm_add_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gon_gmo_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/batched_gemm_bias_permute/CMakeLists.txt (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/batched_gemm_bias_permute/device_batched_gemm_bias_permute_m2_n3_k1_xdl_c_shuffle_f16_f16_f16_f16_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/batched_gemm_gemm/CMakeLists.txt (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/batched_gemm_gemm/device_batched_gemm_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/batched_gemm_gemm/device_batched_gemm_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gon_gmo_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/batched_gemm_multi_d/CMakeLists.txt (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gkn_gmn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gkn_gmn_irregular_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gnk_gmn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gnk_gmn_irregular_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gkn_gmn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gkn_gmn_irregular_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gnk_gmn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gnk_gmn_irregular_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gkn_gmn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gkn_gmn_irregular_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gnk_gmn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gnk_gmn_irregular_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gkn_gmn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gkn_gmn_irregular_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gnk_gmn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gnk_gmn_irregular_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/batched_gemm_reduce/CMakeLists.txt (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gkn_gmn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gnk_gmn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gkn_gmn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gnk_gmn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm/CMakeLists.txt (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm/device_batched_gemm_softmax_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/CMakeLists.txt (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/batchnorm/CMakeLists.txt (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_backward_bf16_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_backward_f16_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_backward_f32_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_backward_f64_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_forward_bf16_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_forward_f16_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_forward_f32_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_forward_f64_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_infer_bf16_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_infer_f16_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_infer_f32_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_infer_f64_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/column_to_image/CMakeLists.txt (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/column_to_image/device_column_to_image_nhwc_1d_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/column_to_image/device_column_to_image_nhwc_2d_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/column_to_image/device_column_to_image_nhwc_3d_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/contraction_bilinear/CMakeLists.txt (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_kknn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_knnn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mknn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mnnn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_kknn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_knnn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_mknn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_mnnn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/contraction_scale/CMakeLists.txt (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_kkn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_knn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mkn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mnn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_kkn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_knn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_mkn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_mnn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/conv1d_bwd_data/CMakeLists.txt (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/conv1d_bwd_data/device_conv1d_bwd_data_xdl_nwc_kxc_nwk_bf16_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/conv1d_bwd_data/device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f16_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/conv1d_bwd_data/device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f32_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/conv1d_bwd_data/device_conv1d_bwd_data_xdl_nwc_kxc_nwk_int8_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/conv2d_bwd_data/CMakeLists.txt (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_f16_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_f32_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_int8_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_bf16_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f16_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f32_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_int8_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/conv2d_fwd/CMakeLists.txt (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/conv2d_fwd/device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_f16_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/conv2d_fwd/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_bf16_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/conv2d_fwd/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f16_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/conv2d_fwd/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f32_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/conv2d_fwd/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_int8_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/conv2d_fwd_bias_relu/CMakeLists.txt (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/conv2d_fwd_bias_relu/device_conv2d_fwd_xdl_c_shuffle_bias_relu_nhwc_kyxc_nhwk_f16_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/conv2d_fwd_bias_relu_add/CMakeLists.txt (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/conv2d_fwd_bias_relu_add/device_conv2d_fwd_xdl_c_shuffle_bias_relu_add_nhwc_kyxc_nhwk_f16_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/conv3d_bwd_data/CMakeLists.txt (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/conv3d_bwd_data/device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_bf16_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/conv3d_bwd_data/device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f16_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/conv3d_bwd_data/device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f32_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/conv3d_bwd_data/device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_int8_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/elementwise/CMakeLists.txt (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/elementwise/device_normalize_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/elementwise_normalization/CMakeLists.txt (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/elementwise_normalization/device_elementwise_normalization_f16_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/CMakeLists.txt (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_km_kn_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_km_kn_mn_irregular_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_km_nk_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_km_nk_mn_irregular_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_mk_kn_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_mk_kn_mn_irregular_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_mk_nk_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_mk_nk_mn_irregular_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f32_f32_f32_km_kn_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f32_f32_f32_km_nk_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f32_f32_f32_mk_kn_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f32_f32_f32_mk_nk_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_km_kn_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_km_kn_mn_irregular_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_km_nk_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_km_nk_mn_irregular_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_mk_kn_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_mk_kn_mn_irregular_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_mk_nk_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_mk_nk_mn_irregular_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/device_gemm_dpp_f16_f16_f16_km_kn_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/device_gemm_dpp_f16_f16_f16_km_kn_mn_irregular_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/device_gemm_dpp_f16_f16_f16_km_nk_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/device_gemm_dpp_f16_f16_f16_km_nk_mn_irregular_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/device_gemm_dpp_f16_f16_f16_mk_kn_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/device_gemm_dpp_f16_f16_f16_mk_kn_mn_irregular_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/device_gemm_dpp_f16_f16_f16_mk_nk_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/device_gemm_dpp_f16_f16_f16_mk_nk_mn_irregular_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_2_stage_f16_f16_f16_mk_nk_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_kn_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_nk_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_kn_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_nk_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f32_f32_f32_km_kn_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f32_f32_f32_km_nk_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f32_f32_f32_mk_kn_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f32_f32_f32_mk_nk_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_km_kn_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_km_nk_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_mk_kn_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_mk_nk_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_i8_i8_i8_km_nk_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_i8_i8_i8_mk_kn_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/common.hpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_add_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_default_pipeline_v1_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_default_pipeline_v2_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_default_pipeline_v2_opt_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_interwave_pipeline_v1_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_irregular_default_pipeline_v1_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_irregular_default_pipeline_v2_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_irregular_interwave_pipeline_v1_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_add_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_default_pipeline_v1_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_default_pipeline_v2_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_default_pipeline_v2_opt_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_interwave_pipeline_v1_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_irregular_default_pipeline_v1_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_irregular_default_pipeline_v2_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_irregular_interwave_pipeline_v1_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_add_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_default_pipeline_v1_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_default_pipeline_v2_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_default_pipeline_v2_opt_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_interwave_pipeline_v1_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_irregular_default_pipeline_v1_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_irregular_default_pipeline_v2_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_irregular_interwave_pipeline_v1_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_add_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_default_pipeline_v1_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_default_pipeline_v2_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_default_pipeline_v2_opt_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_interwave_pipeline_v1_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_irregular_default_pipeline_v1_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_irregular_default_pipeline_v2_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_irregular_interwave_pipeline_v1_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f32_f32_f32_km_kn_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f32_f32_f32_km_nk_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f32_f32_f32_mk_kn_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f32_f32_f32_mk_nk_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f64_f64_f64_km_kn_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f64_f64_f64_km_nk_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f64_f64_f64_mk_kn_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f64_f64_f64_mk_nk_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/CMakeLists.txt (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm_add_fastgelu/CMakeLists.txt (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm_add_multiply/CMakeLists.txt (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm_add_multiply/device_gemm_add_multiply_xdl_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm_add_multiply/device_gemm_add_multiply_xdl_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm_add_multiply/device_gemm_add_multiply_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm_add_multiply/device_gemm_add_multiply_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm_add_relu_add_layernorm/CMakeLists.txt (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm_add_relu_add_layernorm/device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_kn_mn_mn_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm_add_relu_add_layernorm/device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_nk_mn_mn_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm_add_relu_add_layernorm/device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_kn_mn_mn_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm_add_relu_add_layernorm/device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_nk_mn_mn_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm_bias_add_reduce/CMakeLists.txt (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm_bias_add_reduce/device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f32_f32_km_kn_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm_bias_add_reduce/device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f32_f32_km_nk_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm_bias_add_reduce/device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f32_f32_mk_kn_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm_bias_add_reduce/device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f32_f32_mk_nk_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm_bilinear/CMakeLists.txt (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_wmma_c_shuffle_i8_i8_i8_i8_km_kn_mn_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_wmma_c_shuffle_i8_i8_i8_i8_km_nk_mn_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_wmma_c_shuffle_i8_i8_i8_i8_mk_kn_mn_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_wmma_c_shuffle_i8_i8_i8_i8_mk_nk_mn_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm_fastgelu/CMakeLists.txt (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm_multiply_add/CMakeLists.txt (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm_multiply_add/device_gemm_multiply_add_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm_multiply_add/device_gemm_multiply_add_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm_multiply_add/device_gemm_multiply_add_xdl_c_shuffle_f16_f8_f32_f32_f16_mk_kn_mn_mn_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm_multiply_add/device_gemm_multiply_add_xdl_c_shuffle_f16_f8_f32_f32_f16_mk_nk_mn_mn_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm_reduce/CMakeLists.txt (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_kn_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_nk_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_kn_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_nk_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm_splitk/CMakeLists.txt (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_comp_fp8_km_kn_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_comp_fp8_km_nk_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_comp_fp8_mk_kn_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_comp_fp8_mk_nk_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_km_kn_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_km_nk_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_fp8_f16_km_kn_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_fp8_f16_km_nk_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_fp8_f16_mk_kn_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_fp8_f16_mk_nk_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f32_f32_f32_km_kn_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f32_f32_f32_km_nk_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f32_f32_f32_mk_kn_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f32_f32_f32_mk_nk_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_fp8_f16_f16_km_kn_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_fp8_f16_f16_km_nk_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_fp8_f16_f16_mk_kn_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_fp8_f16_f16_mk_nk_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm_streamk/CMakeLists.txt (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemm_streamk/device_gemm_xdl_streamk_f16_f16_f16_mk_kn_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemv_splitk/CMakeLists.txt (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemv_splitk/device_gemv_splitk_f16_f16_f16_mk_kn_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/gemv_splitk/device_gemv_splitk_f16_f16_f16_mk_nk_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/CMakeLists.txt (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/dl/device_grouped_conv1d_bwd_weight_dl_gnwc_gkxc_gnwk_bf16_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/dl/device_grouped_conv1d_bwd_weight_dl_gnwc_gkxc_gnwk_f16_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/dl/device_grouped_conv1d_bwd_weight_dl_gnwc_gkxc_gnwk_f32_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/dl/device_grouped_conv1d_bwd_weight_dl_nwgc_gkxc_nwgk_bf16_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/dl/device_grouped_conv1d_bwd_weight_dl_nwgc_gkxc_nwgk_f16_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/dl/device_grouped_conv1d_bwd_weight_dl_nwgc_gkxc_nwgk_f32_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/xdl/device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_bf16_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/xdl/device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_f16_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/xdl/device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_f32_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv1d_fwd/CMakeLists.txt (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv1d_fwd/xdl/device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_bf16_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv1d_fwd/xdl/device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_f16_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv1d_fwd/xdl/device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_f32_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv1d_fwd/xdl/device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_int8_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/CMakeLists.txt (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/wmma/device_grouped_conv2d_bwd_data_wmma_gnhwc_gkyxc_gnhwk_f16_1x1s1p0_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/wmma/device_grouped_conv2d_bwd_data_wmma_gnhwc_gkyxc_gnhwk_f16_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/wmma/device_grouped_conv2d_bwd_data_wmma_gnhwc_gkyxc_gnhwk_i8_1x1s1p0_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/wmma/device_grouped_conv2d_bwd_data_wmma_gnhwc_gkyxc_gnhwk_i8_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/wmma/device_grouped_conv2d_bwd_data_wmma_nhwgc_gkyxc_nhwgk_f16_1x1s1p0_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/wmma/device_grouped_conv2d_bwd_data_wmma_nhwgc_gkyxc_nhwgk_f16_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/wmma/device_grouped_conv2d_bwd_data_wmma_nhwgc_gkyxc_nhwgk_i8_1x1s1p0_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/wmma/device_grouped_conv2d_bwd_data_wmma_nhwgc_gkyxc_nhwgk_i8_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/xdl/device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/xdl/device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_f16_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/xdl/device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_f32_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/xdl/device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/xdl/device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_f16_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/xdl/device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_f32_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/CMakeLists.txt (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/dl/device_grouped_conv2d_bwd_weight_dl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/dl/device_grouped_conv2d_bwd_weight_dl_gnhwc_gkyxc_gnhwk_f16_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/dl/device_grouped_conv2d_bwd_weight_dl_gnhwc_gkyxc_gnhwk_f32_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/dl/device_grouped_conv2d_bwd_weight_dl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/dl/device_grouped_conv2d_bwd_weight_dl_nhwgc_gkyxc_nhwgk_f16_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/dl/device_grouped_conv2d_bwd_weight_dl_nhwgc_gkyxc_nhwgk_f32_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/xdl/device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/xdl/device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f16_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/xdl/device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f32_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/xdl/device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/xdl/device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f16_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/xdl/device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f32_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/CMakeLists.txt (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_f16_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_i8_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/dl/device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f16_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/dl/device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f32_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/dl/device_grouped_conv2d_fwd_dl_nhwgc_gkyxc_nhwgk_f16_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/dl/device_grouped_conv2d_fwd_dl_nhwgc_gkyxc_nhwgk_f32_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_f16_1x1p0_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_f16_1x1s1p0_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_f16_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_f16_oddc_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_i8_1x1p0_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_i8_1x1s1p0_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_i8_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_i8_oddc_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_f16_1x1p0_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_f16_1x1s1p0_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_f16_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_f16_oddc_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_i8_1x1p0_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_i8_1x1s1p0_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_i8_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_i8_oddc_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f16_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f32_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/CMakeLists.txt (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/wmma/device_grouped_conv3d_bwd_data_wmma_gndhwc_gkzyxc_gndhwk_f16_1x1s1p0_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/wmma/device_grouped_conv3d_bwd_data_wmma_gndhwc_gkzyxc_gndhwk_f16_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/wmma/device_grouped_conv3d_bwd_data_wmma_gndhwc_gkzyxc_gndhwk_i8_1x1s1p0_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/wmma/device_grouped_conv3d_bwd_data_wmma_gndhwc_gkzyxc_gndhwk_i8_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/wmma/device_grouped_conv3d_bwd_data_wmma_ndhwgc_gkzyxc_ndhwgk_f16_1x1s1p0_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/wmma/device_grouped_conv3d_bwd_data_wmma_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/wmma/device_grouped_conv3d_bwd_data_wmma_ndhwgc_gkzyxc_ndhwgk_i8_1x1s1p0_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/wmma/device_grouped_conv3d_bwd_data_wmma_ndhwgc_gkzyxc_ndhwgk_i8_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/xdl/device_grouped_conv3d_bwd_data_xdl_gndhwc_gkzyxc_gndhwk_bf16_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/xdl/device_grouped_conv3d_bwd_data_xdl_gndhwc_gkzyxc_gndhwk_f16_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/xdl/device_grouped_conv3d_bwd_data_xdl_gndhwc_gkzyxc_gndhwk_f32_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/xdl/device_grouped_conv3d_bwd_data_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/xdl/device_grouped_conv3d_bwd_data_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/xdl/device_grouped_conv3d_bwd_data_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/xdl/device_grouped_conv3d_bwd_data_xdl_ndhwgc_gkzyxc_ndhwgk_input_f16_comp_bf8_f8_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/CMakeLists.txt (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/dl/device_grouped_conv3d_bwd_weight_dl_gndhwc_gkzyxc_gndhwk_bf16_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/dl/device_grouped_conv3d_bwd_weight_dl_gndhwc_gkzyxc_gndhwk_f16_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/dl/device_grouped_conv3d_bwd_weight_dl_gndhwc_gkzyxc_gndhwk_f32_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/dl/device_grouped_conv3d_bwd_weight_dl_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/dl/device_grouped_conv3d_bwd_weight_dl_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/dl/device_grouped_conv3d_bwd_weight_dl_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/device_grouped_conv3d_bwd_weight_wmma_gndhwc_gkzyxc_gndhwk_f16_1x1s1p0_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/device_grouped_conv3d_bwd_weight_wmma_gndhwc_gkzyxc_gndhwk_f16_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/device_grouped_conv3d_bwd_weight_wmma_gndhwc_gkzyxc_gndhwk_i8_1x1s1p0_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/device_grouped_conv3d_bwd_weight_wmma_gndhwc_gkzyxc_gndhwk_i8_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_f16_1x1s1p0_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_i8_1x1s1p0_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_i8_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/xdl/device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_bf16_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/xdl/device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_f16_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/xdl/device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_f32_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/xdl/device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/xdl/device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f16_comp_bf8_fp8_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/xdl/device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/xdl/device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/CMakeLists.txt (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_f16_1x1p0_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_f16_1x1s1p0_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_f16_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_f16_oddc_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_i8_1x1p0_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_i8_1x1s1p0_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_i8_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_i8_oddc_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_f16_1x1p0_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_f16_1x1s1p0_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_f16_oddc_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_i8_1x1p0_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_i8_1x1s1p0_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_i8_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_i8_oddc_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/xdl/device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_bf16_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/xdl/device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_f16_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/xdl/device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_f32_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/xdl/device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_int8_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/xdl/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/xdl/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f16_comp_fp8_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/xdl/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/xdl/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/xdl/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_int8_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_gemm/CMakeLists.txt (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_f16_f16_f16_km_kn_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_f16_f16_f16_km_nk_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_f16_f16_f16_mk_kn_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_f16_f16_f16_mk_nk_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_irregular_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_irregular_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_gemm_bias/CMakeLists.txt (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_gemm_bias/device_grouped_gemm_xdl_fixed_nk_bias_f16_f16_f16_mk_kn_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_gemm_bias/device_grouped_gemm_xdl_fixed_nk_bias_f16_f16_f16_mk_nk_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_gemm_bias/device_grouped_gemm_xdl_fixed_nk_bias_f16_f16_f32_mk_kn_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_gemm_bias/device_grouped_gemm_xdl_fixed_nk_bias_f16_f16_f32_mk_nk_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_gemm_fastgelu/CMakeLists.txt (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_gemm_fastgelu/device_grouped_gemm_fastgelu_xdl_f16_f16_f16_km_kn_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_gemm_fastgelu/device_grouped_gemm_fastgelu_xdl_f16_f16_f16_km_nk_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_gemm_fastgelu/device_grouped_gemm_fastgelu_xdl_f16_f16_f16_mk_kn_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_gemm_fastgelu/device_grouped_gemm_fastgelu_xdl_f16_f16_f16_mk_nk_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_gemm_fixed_nk/CMakeLists.txt (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_gemm_fixed_nk/device_grouped_gemm_xdl_fixed_nk_f16_f16_f16_mk_kn_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_gemm_fixed_nk/device_grouped_gemm_xdl_fixed_nk_f16_f16_f16_mk_nk_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_gemm_fixed_nk/device_grouped_gemm_xdl_fixed_nk_f16_fp8_f16_mk_kn_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_gemm_fixed_nk/device_grouped_gemm_xdl_fixed_nk_f16_fp8_f16_mk_nk_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_gemm_fixed_nk/device_grouped_gemm_xdl_fixed_nk_f16_i8_f16_mk_kn_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/grouped_gemm_fixed_nk/device_grouped_gemm_xdl_fixed_nk_f16_i8_f16_mk_nk_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/image_to_column/CMakeLists.txt (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/image_to_column/device_image_to_column_nhwc_1d_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/image_to_column/device_image_to_column_nhwc_2d_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/image_to_column/device_image_to_column_nhwc_3d_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/max_pool_bwd/CMakeLists.txt (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/max_pool_bwd/device_max_pool_bwd_bf16_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/max_pool_bwd/device_max_pool_bwd_f16_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/max_pool_bwd/device_max_pool_bwd_f32_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/max_pool_bwd/max_pool_bwd_instance_common.hpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/normalization/CMakeLists.txt (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/normalization/device_groupnorm_f16_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/normalization/device_groupnorm_f32_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/normalization/device_groupnorm_swish_f16_f32_f32_f16_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/normalization/device_groupnorm_swish_f16_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/normalization/device_groupnorm_swish_f32_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/normalization/device_layernorm2d_f16_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/normalization/device_layernorm2d_f32_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/normalization/device_layernorm4d_f16_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/normalization/device_layernorm4d_f32_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/normalization/normalization_instance_common.hpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/pool3d_fwd/CMakeLists.txt (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/pool3d_fwd/device_avg_pool3d_fwd_ndhwc_bf16_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/pool3d_fwd/device_avg_pool3d_fwd_ndhwc_f16_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/pool3d_fwd/device_avg_pool3d_fwd_ndhwc_f32_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/pool3d_fwd/device_max_pool3d_fwd_ndhwc_bf16_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/pool3d_fwd/device_max_pool3d_fwd_ndhwc_f16_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/pool3d_fwd/device_max_pool3d_fwd_ndhwc_f32_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/pool3d_fwd/pool_fwd_instance_common.hpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/quantization/CMakeLists.txt (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/conv2d_quantization_common.hpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_dl_bias_perchannel_quantization_int8_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_dl_bias_perlayer_quantization_int8_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_dl_int8_instance.hpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_dl_perchannel_quantization_int8_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_dl_perlayer_quantization_int8_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_xdl_bias_perchannel_quantization_int8_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_xdl_bias_perlayer_quantization_int8_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_xdl_int8_instance.hpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_xdl_perchannel_quantization_int8_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_xdl_perlayer_quantization_int8_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_dl_c_shuffle_i8_i8_i8_instance.hpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_dl_c_shuffle_i8_i8_i8_km_kn_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_dl_c_shuffle_i8_i8_i8_km_nk_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_dl_c_shuffle_i8_i8_i8_mk_kn_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_dl_c_shuffle_i8_i8_i8_mk_nk_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_instance.hpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_km_nk_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_mk_kn_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/quantization/gemm/gemm_quantization_common.hpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/reduce/CMakeLists.txt (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_add.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_amax.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_avg.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_max.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_min.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_norm2.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_amax.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_max.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_min.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_add.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_avg.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_norm2.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_add.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_amax.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_avg.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_max.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_min.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_norm2.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_add.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_avg.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_norm2.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_add.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_amax.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_avg.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_max.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_min.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_norm2.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i32_i8_add.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i32_i8_avg.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_amax.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_max.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_min.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_b16_f32_f32_add.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_b16_f32_f32_avg.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f16_f32_f32_add.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f16_f32_f32_avg.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f32_f32_add.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f32_f32_avg.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f64_f32_add.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f64_f32_avg.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f64_f64_f64_add.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f64_f64_f64_avg.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_add.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_amax.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_avg.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_max.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_min.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_norm2.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_amax.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_max.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_min.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_add.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_avg.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_norm2.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_add.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_amax.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_avg.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_max.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_min.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_norm2.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_add.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_avg.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_norm2.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_add.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_amax.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_avg.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_max.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_min.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_norm2.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i32_i8_add.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i32_i8_avg.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_amax.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_max.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_min.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/softmax/CMakeLists.txt (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce1.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce2.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce3.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce1.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce2.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce3.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce4.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce1.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce2.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce3.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce1.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce2.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce3.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce4.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/tall_and_skinny_gemm_splitk/CMakeLists.txt (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/tall_and_skinny_gemm_splitk/device_tall_and_skinny_gemm_splitk_f16_f16_f16_mk_kn_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/tensor_operation_instance/gpu/tall_and_skinny_gemm_splitk/device_tall_and_skinny_gemm_splitk_f16_f16_f16_mk_nk_mn_instance.cpp (100%)
 rename {composable_kernel/library => library}/src/utility/CMakeLists.txt (100%)
 rename {composable_kernel/library => library}/src/utility/convolution_parameter.cpp (100%)
 rename {composable_kernel/library => library}/src/utility/device_memory.cpp (100%)
 rename {composable_kernel/library => library}/src/utility/host_tensor.cpp (100%)
 rename {composable_kernel/profiler => profiler}/CMakeLists.txt (100%)
 rename {composable_kernel/profiler => profiler}/README.md (100%)
 rename {composable_kernel/profiler => profiler}/include/profiler/data_type_enum.hpp (100%)
 rename {composable_kernel/profiler => profiler}/include/profiler/profile_avg_pool3d_bwd_impl.hpp (100%)
 rename {composable_kernel/profiler => profiler}/include/profiler/profile_batched_gemm_add_relu_gemm_add_impl.hpp (100%)
 rename {composable_kernel/profiler => profiler}/include/profiler/profile_batched_gemm_bias_softmax_gemm_permute_impl.hpp (100%)
 rename {composable_kernel/profiler => profiler}/include/profiler/profile_batched_gemm_gemm_impl.hpp (100%)
 rename {composable_kernel/profiler => profiler}/include/profiler/profile_batched_gemm_impl.hpp (100%)
 rename {composable_kernel/profiler => profiler}/include/profiler/profile_batched_gemm_reduce_impl.hpp (100%)
 rename {composable_kernel/profiler => profiler}/include/profiler/profile_batched_gemm_softmax_gemm_impl.hpp (100%)
 rename {composable_kernel/profiler => profiler}/include/profiler/profile_batched_gemm_softmax_gemm_permute_impl.hpp (100%)
 rename {composable_kernel/profiler => profiler}/include/profiler/profile_batchnorm_backward_impl.hpp (100%)
 rename {composable_kernel/profiler => profiler}/include/profiler/profile_batchnorm_forward_impl.hpp (100%)
 rename {composable_kernel/profiler => profiler}/include/profiler/profile_batchnorm_infer_impl.hpp (100%)
 rename {composable_kernel/profiler => profiler}/include/profiler/profile_contraction_impl.hpp (100%)
 rename {composable_kernel/profiler => profiler}/include/profiler/profile_contraction_utils.hpp (100%)
 rename {composable_kernel/profiler => profiler}/include/profiler/profile_conv_bwd_data_impl.hpp (100%)
 rename {composable_kernel/profiler => profiler}/include/profiler/profile_conv_fwd_bias_relu_add_impl.hpp (100%)
 rename {composable_kernel/profiler => profiler}/include/profiler/profile_conv_fwd_bias_relu_impl.hpp (100%)
 rename {composable_kernel/profiler => profiler}/include/profiler/profile_conv_fwd_impl.hpp (100%)
 rename {composable_kernel/profiler => profiler}/include/profiler/profile_conv_tensor_rearrange_impl.hpp (100%)
 rename {composable_kernel/profiler => profiler}/include/profiler/profile_elementwise_layernorm_impl.hpp (100%)
 rename {composable_kernel/profiler => profiler}/include/profiler/profile_gemm_add_add_fastgelu_impl.hpp (100%)
 rename {composable_kernel/profiler => profiler}/include/profiler/profile_gemm_add_fastgelu_impl.hpp (100%)
 rename {composable_kernel/profiler => profiler}/include/profiler/profile_gemm_add_multiply_impl.hpp (100%)
 rename {composable_kernel/profiler => profiler}/include/profiler/profile_gemm_add_relu_add_layernorm_impl.hpp (100%)
 rename {composable_kernel/profiler => profiler}/include/profiler/profile_gemm_bias_add_reduce_impl.hpp (100%)
 rename {composable_kernel/profiler => profiler}/include/profiler/profile_gemm_bilinear_impl.hpp (100%)
 rename {composable_kernel/profiler => profiler}/include/profiler/profile_gemm_fastgelu_impl.hpp (100%)
 rename {composable_kernel/profiler => profiler}/include/profiler/profile_gemm_impl.hpp (100%)
 rename {composable_kernel/profiler => profiler}/include/profiler/profile_gemm_multiply_add_impl.hpp (100%)
 rename {composable_kernel/profiler => profiler}/include/profiler/profile_gemm_reduce_impl.hpp (100%)
 rename {composable_kernel/profiler => profiler}/include/profiler/profile_gemm_splitk_impl.hpp (100%)
 rename {composable_kernel/profiler => profiler}/include/profiler/profile_gemm_streamk_impl.hpp (100%)
 rename {composable_kernel/profiler => profiler}/include/profiler/profile_gemv_splitk_impl.hpp (100%)
 rename {composable_kernel/profiler => profiler}/include/profiler/profile_grouped_conv_bwd_data_impl.hpp (100%)
 rename {composable_kernel/profiler => profiler}/include/profiler/profile_grouped_conv_bwd_weight_impl.hpp (100%)
 rename {composable_kernel/profiler => profiler}/include/profiler/profile_grouped_conv_fwd_impl.hpp (100%)
 rename {composable_kernel/profiler => profiler}/include/profiler/profile_grouped_gemm_fastgelu_impl.hpp (100%)
 rename {composable_kernel/profiler => profiler}/include/profiler/profile_grouped_gemm_impl.hpp (100%)
 rename {composable_kernel/profiler => profiler}/include/profiler/profile_groupnorm_impl.hpp (100%)
 rename {composable_kernel/profiler => profiler}/include/profiler/profile_layernorm_impl.hpp (100%)
 rename {composable_kernel/profiler => profiler}/include/profiler/profile_max_pool3d_bwd_impl.hpp (100%)
 rename {composable_kernel/profiler => profiler}/include/profiler/profile_pool3d_fwd_impl.hpp (100%)
 rename {composable_kernel/profiler => profiler}/include/profiler/profile_reduce_impl.hpp (100%)
 rename {composable_kernel/profiler => profiler}/include/profiler/profile_softmax_impl.hpp (100%)
 rename {composable_kernel/profiler => profiler}/include/profiler/profile_tall_and_skinny_gemm_splitk_impl.hpp (100%)
 rename {composable_kernel/profiler => profiler}/src/CMakeLists.txt (100%)
 rename {composable_kernel/profiler => profiler}/src/profile_avg_pool3d_bwd.cpp (100%)
 rename {composable_kernel/profiler => profiler}/src/profile_batched_gemm.cpp (100%)
 rename {composable_kernel/profiler => profiler}/src/profile_batched_gemm_add_relu_gemm_add.cpp (100%)
 rename {composable_kernel/profiler => profiler}/src/profile_batched_gemm_gemm.cpp (100%)
 rename {composable_kernel/profiler => profiler}/src/profile_batched_gemm_multi_d.cpp (100%)
 rename {composable_kernel/profiler => profiler}/src/profile_batched_gemm_reduce.cpp (100%)
 rename {composable_kernel/profiler => profiler}/src/profile_batchnorm_bwd.cpp (100%)
 rename {composable_kernel/profiler => profiler}/src/profile_batchnorm_fwd.cpp (100%)
 rename {composable_kernel/profiler => profiler}/src/profile_batchnorm_infer.cpp (100%)
 rename {composable_kernel/profiler => profiler}/src/profile_contraction_bilinear.cpp (100%)
 rename {composable_kernel/profiler => profiler}/src/profile_contraction_scale.cpp (100%)
 rename {composable_kernel/profiler => profiler}/src/profile_conv_bwd_data.cpp (100%)
 rename {composable_kernel/profiler => profiler}/src/profile_conv_fwd.cpp (100%)
 rename {composable_kernel/profiler => profiler}/src/profile_conv_fwd_bias_relu.cpp (100%)
 rename {composable_kernel/profiler => profiler}/src/profile_conv_fwd_bias_relu_add.cpp (100%)
 rename {composable_kernel/profiler => profiler}/src/profile_conv_tensor_rearrange.cpp (100%)
 rename {composable_kernel/profiler => profiler}/src/profile_gemm.cpp (100%)
 rename {composable_kernel/profiler => profiler}/src/profile_gemm_add_add_fastgelu.cpp (100%)
 rename {composable_kernel/profiler => profiler}/src/profile_gemm_add_fastgelu.cpp (100%)
 rename {composable_kernel/profiler => profiler}/src/profile_gemm_add_multiply.cpp (100%)
 rename {composable_kernel/profiler => profiler}/src/profile_gemm_add_relu_add_layernorm.cpp (100%)
 rename {composable_kernel/profiler => profiler}/src/profile_gemm_bias_add_reduce.cpp (100%)
 rename {composable_kernel/profiler => profiler}/src/profile_gemm_bilinear.cpp (100%)
 rename {composable_kernel/profiler => profiler}/src/profile_gemm_fastgelu.cpp (100%)
 rename {composable_kernel/profiler => profiler}/src/profile_gemm_multiply_add.cpp (100%)
 rename {composable_kernel/profiler => profiler}/src/profile_gemm_reduce.cpp (100%)
 rename {composable_kernel/profiler => profiler}/src/profile_gemm_splitk.cpp (100%)
 rename {composable_kernel/profiler => profiler}/src/profile_gemm_streamk.cpp (100%)
 rename {composable_kernel/profiler => profiler}/src/profile_gemv_splitk.cpp (100%)
 rename {composable_kernel/profiler => profiler}/src/profile_grouped_conv_bwd_data.cpp (100%)
 rename {composable_kernel/profiler => profiler}/src/profile_grouped_conv_bwd_weight.cpp (100%)
 rename {composable_kernel/profiler => profiler}/src/profile_grouped_conv_fwd.cpp (100%)
 rename {composable_kernel/profiler => profiler}/src/profile_grouped_gemm.cpp (100%)
 rename {composable_kernel/profiler => profiler}/src/profile_grouped_gemm_fastgelu.cpp (100%)
 rename {composable_kernel/profiler => profiler}/src/profile_groupnorm.cpp (100%)
 rename {composable_kernel/profiler => profiler}/src/profile_layernorm.cpp (100%)
 rename {composable_kernel/profiler => profiler}/src/profile_max_pool3d_bwd.cpp (100%)
 rename {composable_kernel/profiler => profiler}/src/profile_max_pool3d_fwd.cpp (100%)
 rename {composable_kernel/profiler => profiler}/src/profile_reduce.cpp (100%)
 rename {composable_kernel/profiler => profiler}/src/profile_softmax.cpp (100%)
 rename {composable_kernel/profiler => profiler}/src/profile_tall_and_skinny_gemm_splitk.cpp (100%)
 rename {composable_kernel/profiler => profiler}/src/profiler.cpp (100%)
 rename {composable_kernel/profiler => profiler}/src/profiler_operation_registry.hpp (100%)
 rename composable_kernel/rbuild.ini => rbuild.ini (100%)
 rename composable_kernel/requirements.txt => requirements.txt (100%)
 rename {composable_kernel/script => script}/check_copyright_year.sh (100%)
 rename {composable_kernel/script => script}/clang-format-overwrite.sh (100%)
 rename {composable_kernel/script => script}/cmake-ck-dev.sh (100%)
 rename {composable_kernel/script => script}/cmake-ck-release.sh (100%)
 rename {composable_kernel/script => script}/count_vgpr.sh (100%)
 rename {composable_kernel/script => script}/hipclang_opt.sh (100%)
 rename {composable_kernel/script => script}/install_precommit.sh (100%)
 rename {composable_kernel/script => script}/parse_perf_data.py (100%)
 rename {composable_kernel/script => script}/process_perf_data.py (100%)
 rename {composable_kernel/script => script}/process_perf_data.sh (100%)
 rename {composable_kernel/script => script}/process_qa_data.sh (100%)
 rename {composable_kernel/script => script}/profile_batched_gemm.sh (100%)
 rename {composable_kernel/script => script}/profile_conv_bwd_data.sh (100%)
 rename {composable_kernel/script => script}/profile_conv_fwd.sh (100%)
 rename {composable_kernel/script => script}/profile_gemm.sh (100%)
 rename {composable_kernel/script => script}/profile_gemm_bilinear.sh (100%)
 rename {composable_kernel/script => script}/profile_grouped_gemm.sh (100%)
 rename {composable_kernel/script => script}/profile_onnx_gemm.sh (100%)
 rename {composable_kernel/script => script}/profile_reduce_no_index.sh (100%)
 rename {composable_kernel/script => script}/profile_reduce_with_index.sh (100%)
 rename {composable_kernel/script => script}/profile_resnet50.sh (100%)
 rename {composable_kernel/script => script}/profile_splitK_gemm.sh (100%)
 rename {composable_kernel/script => script}/run_full_performance_tests.sh (100%)
 rename {composable_kernel/script => script}/run_performance_tests.sh (100%)
 rename {composable_kernel/script => script}/test_convnd_fwd.sh (100%)
 rename {composable_kernel/script => script}/test_reduce_no_index.sh (100%)
 rename {composable_kernel/script => script}/test_reduce_with_index.sh (100%)
 rename {composable_kernel/script => script}/uninstall_precommit.sh (100%)
 rename {composable_kernel/test => test}/CMakeLists.txt (100%)
 rename {composable_kernel/test => test}/batched_gemm/CMakeLists.txt (100%)
 rename {composable_kernel/test => test}/batched_gemm/batched_gemm_bf16.cpp (100%)
 rename {composable_kernel/test => test}/batched_gemm/batched_gemm_fp16.cpp (100%)
 rename {composable_kernel/test => test}/batched_gemm/batched_gemm_fp32.cpp (100%)
 rename {composable_kernel/test => test}/batched_gemm/batched_gemm_int8.cpp (100%)
 rename {composable_kernel/test => test}/batched_gemm/test_batched_gemm.cpp (100%)
 rename {composable_kernel/test => test}/batched_gemm_gemm/CMakeLists.txt (100%)
 rename {composable_kernel/test => test}/batched_gemm_gemm/test_batched_gemm_gemm_fp16.cpp (100%)
 rename {composable_kernel/test => test}/batched_gemm_gemm/test_batched_gemm_gemm_util.hpp (100%)
 rename {composable_kernel/test => test}/batched_gemm_multi_d/CMakeLists.txt (100%)
 rename {composable_kernel/test => test}/batched_gemm_multi_d/test_batched_gemm_multi_d_dl.cpp (100%)
 rename {composable_kernel/test => test}/batched_gemm_reduce/CMakeLists.txt (100%)
 rename {composable_kernel/test => test}/batched_gemm_reduce/batched_gemm_reduce_fp16.cpp (100%)
 rename {composable_kernel/test => test}/batched_gemm_softmax_gemm/CMakeLists.txt (100%)
 rename {composable_kernel/test => test}/batched_gemm_softmax_gemm/test_batched_gemm_softmax_gemm_fp16.cpp (100%)
 rename {composable_kernel/test => test}/batched_gemm_softmax_gemm/test_batched_gemm_softmax_gemm_util.hpp (100%)
 rename {composable_kernel/test => test}/batched_gemm_softmax_gemm_permute/CMakeLists.txt (100%)
 rename {composable_kernel/test => test}/batched_gemm_softmax_gemm_permute/test_batched_gemm_bias_softmax_gemm_permute_bf16.cpp (100%)
 rename {composable_kernel/test => test}/batched_gemm_softmax_gemm_permute/test_batched_gemm_bias_softmax_gemm_permute_fp16.cpp (100%)
 rename {composable_kernel/test => test}/batched_gemm_softmax_gemm_permute/test_batched_gemm_bias_softmax_gemm_permute_util.hpp (100%)
 rename {composable_kernel/test => test}/batched_gemm_softmax_gemm_permute/test_batched_gemm_softmax_gemm_permute_bf16.cpp (100%)
 rename {composable_kernel/test => test}/batched_gemm_softmax_gemm_permute/test_batched_gemm_softmax_gemm_permute_fp16.cpp (100%)
 rename {composable_kernel/test => test}/batched_gemm_softmax_gemm_permute/test_batched_gemm_softmax_gemm_permute_util.hpp (100%)
 rename {composable_kernel/test => test}/batchnorm/CMakeLists.txt (100%)
 rename {composable_kernel/test => test}/batchnorm/batchnorm_bwd_rank_4.cpp (100%)
 rename {composable_kernel/test => test}/batchnorm/batchnorm_fwd_rank_4.cpp (100%)
 rename {composable_kernel/test => test}/batchnorm/batchnorm_infer_rank_4.cpp (100%)
 rename {composable_kernel/test => test}/block_swizzle_test/block_swizzle_test.cpp (100%)
 rename {composable_kernel/test => test}/block_swizzle_test/rebuild.sh (100%)
 rename {composable_kernel/test => test}/block_swizzle_test/simple_args.h (100%)
 rename {composable_kernel/test => test}/block_to_ctile_map/CMakeLists.txt (100%)
 rename {composable_kernel/test => test}/block_to_ctile_map/test_block_to_ctile_map.cpp (100%)
 rename {composable_kernel/test => test}/contraction/CMakeLists.txt (100%)
 rename {composable_kernel/test => test}/contraction/test_contraction.cpp (100%)
 rename {composable_kernel/test => test}/contraction/test_contraction_interface.cpp (100%)
 rename {composable_kernel/test => test}/conv_tensor_rearrange/CMakeLists.txt (100%)
 rename {composable_kernel/test => test}/conv_tensor_rearrange/test_conv_tensor_rearrange.cpp (100%)
 rename {composable_kernel/test => test}/conv_tensor_rearrange/test_conv_tensor_rearrange_interface.cpp (100%)
 rename {composable_kernel/test => test}/conv_util/CMakeLists.txt (100%)
 rename {composable_kernel/test => test}/conv_util/conv_util.cpp (100%)
 rename {composable_kernel/test => test}/convnd_bwd_data/CMakeLists.txt (100%)
 rename {composable_kernel/test => test}/convnd_bwd_data/convnd_bwd_data.cpp (100%)
 rename {composable_kernel/test => test}/convnd_fwd/CMakeLists.txt (100%)
 rename {composable_kernel/test => test}/convnd_fwd/convnd_fwd.cpp (100%)
 rename {composable_kernel/test => test}/data_type/CMakeLists.txt (100%)
 rename {composable_kernel/test => test}/data_type/test_bf8.cpp (100%)
 rename {composable_kernel/test => test}/data_type/test_fp8.cpp (100%)
 rename {composable_kernel/test => test}/data_type/test_int4.cpp (100%)
 rename {composable_kernel/test => test}/data_type/type_convert_const.cpp (100%)
 rename {composable_kernel/test => test}/elementwise_normalization/CMakeLists.txt (100%)
 rename {composable_kernel/test => test}/elementwise_normalization/test_elementwise_layernorm_fp16.cpp (100%)
 rename {composable_kernel/test => test}/gemm/CMakeLists.txt (100%)
 rename {composable_kernel/test => test}/gemm/gemm_bf16.cpp (100%)
 rename {composable_kernel/test => test}/gemm/gemm_fp16.cpp (100%)
 rename {composable_kernel/test => test}/gemm/gemm_fp32.cpp (100%)
 rename {composable_kernel/test => test}/gemm/gemm_fp64.cpp (100%)
 rename {composable_kernel/test => test}/gemm/gemm_int8.cpp (100%)
 rename {composable_kernel/test => test}/gemm/gemm_standalone_xdl_fp16.cpp (100%)
 rename {composable_kernel/test => test}/gemm/gemm_util.hpp (100%)
 rename {composable_kernel/test => test}/gemm/instance/gemm_f16_nn_instance.cpp (100%)
 rename {composable_kernel/test => test}/gemm/instance/gemm_f16_nn_instance.hpp (100%)
 rename {composable_kernel/test => test}/gemm/instance/gemm_f16_nt_instance.cpp (100%)
 rename {composable_kernel/test => test}/gemm/instance/gemm_f16_nt_instance.hpp (100%)
 rename {composable_kernel/test => test}/gemm/instance/gemm_f16_tn_instance.cpp (100%)
 rename {composable_kernel/test => test}/gemm/instance/gemm_f16_tn_instance.hpp (100%)
 rename {composable_kernel/test => test}/gemm/instance/gemm_f16_tt_instance.cpp (100%)
 rename {composable_kernel/test => test}/gemm/instance/gemm_f16_tt_instance.hpp (100%)
 rename {composable_kernel/test => test}/gemm/instance/gemm_wavelet_f16_tn_instance.cpp (100%)
 rename {composable_kernel/test => test}/gemm/instance/gemm_wavelet_f16_tn_instance.hpp (100%)
 rename {composable_kernel/test => test}/gemm/run_gemm_test.inc (100%)
 rename {composable_kernel/test => test}/gemm_layernorm/CMakeLists.txt (100%)
 rename {composable_kernel/test => test}/gemm_layernorm/test_gemm_add_relu_add_layernorm_fp16.cpp (100%)
 rename {composable_kernel/test => test}/gemm_reduce/CMakeLists.txt (100%)
 rename {composable_kernel/test => test}/gemm_reduce/gemm_reduce_fp16.cpp (100%)
 rename {composable_kernel/test => test}/gemm_split_k/CMakeLists.txt (100%)
 rename {composable_kernel/test => test}/gemm_split_k/test_gemm_splitk.cpp (100%)
 rename {composable_kernel/test => test}/gemm_split_k/test_gemm_splitk_ut_cases.inc (100%)
 rename {composable_kernel/test => test}/gemm_split_k/test_gemm_splitk_util.hpp (100%)
 rename {composable_kernel/test => test}/grouped_convnd_bwd_data/CMakeLists.txt (100%)
 rename {composable_kernel/test => test}/grouped_convnd_bwd_data/test_grouped_convnd_bwd_data.cpp (100%)
 rename {composable_kernel/test => test}/grouped_convnd_bwd_data/test_grouped_convnd_bwd_data_interface_wmma.cpp (100%)
 rename {composable_kernel/test => test}/grouped_convnd_bwd_data/test_grouped_convnd_bwd_data_interface_xdl.cpp (100%)
 rename {composable_kernel/test => test}/grouped_convnd_bwd_weight/CMakeLists.txt (100%)
 rename {composable_kernel/test => test}/grouped_convnd_bwd_weight/test_grouped_convnd_bwd_weight.cpp (100%)
 rename {composable_kernel/test => test}/grouped_convnd_bwd_weight/test_grouped_convnd_bwd_weight_interface_wmma.cpp (100%)
 rename {composable_kernel/test => test}/grouped_convnd_bwd_weight/test_grouped_convnd_bwd_weight_interface_xdl.cpp (100%)
 rename {composable_kernel/test => test}/grouped_convnd_fwd/CMakeLists.txt (100%)
 rename {composable_kernel/test => test}/grouped_convnd_fwd/grouped_convnd_fwd.cpp (100%)
 rename {composable_kernel/test => test}/grouped_convnd_fwd/test_grouped_convnd_fwd.cpp (100%)
 rename {composable_kernel/test => test}/grouped_gemm/CMakeLists.txt (100%)
 rename {composable_kernel/test => test}/grouped_gemm/test_grouped_gemm_interface.cpp (100%)
 rename {composable_kernel/test => test}/grouped_gemm/test_grouped_gemm_splitk.cpp (100%)
 rename {composable_kernel/test => test}/grouped_gemm/test_grouped_gemm_ut_cases.inc (100%)
 rename {composable_kernel/test => test}/grouped_gemm/test_grouped_gemm_util.hpp (100%)
 rename {composable_kernel/test => test}/image_to_column/CMakeLists.txt (100%)
 rename {composable_kernel/test => test}/image_to_column/test_image_to_column.cpp (100%)
 rename {composable_kernel/test => test}/image_to_column/test_image_to_column_interface.cpp (100%)
 rename {composable_kernel/test => test}/magic_number_division/CMakeLists.txt (100%)
 rename {composable_kernel/test => test}/magic_number_division/magic_number_division.cpp (100%)
 rename {composable_kernel/test => test}/normalization/CMakeLists.txt (100%)
 rename {composable_kernel/test => test}/normalization/test_groupnorm_fp16.cpp (100%)
 rename {composable_kernel/test => test}/normalization/test_groupnorm_fp32.cpp (100%)
 rename {composable_kernel/test => test}/normalization/test_layernorm2d_fp16.cpp (100%)
 rename {composable_kernel/test => test}/normalization/test_layernorm2d_fp32.cpp (100%)
 rename {composable_kernel/test => test}/pool/CMakeLists.txt (100%)
 rename {composable_kernel/test => test}/pool/test_avg_pool3d_bwd.cpp (100%)
 rename {composable_kernel/test => test}/pool/test_avg_pool3d_fwd.cpp (100%)
 rename {composable_kernel/test => test}/pool/test_max_pool3d_bwd.cpp (100%)
 rename {composable_kernel/test => test}/pool/test_max_pool3d_fwd.cpp (100%)
 rename {composable_kernel/test => test}/pool/test_pool_fwd_common.hpp (100%)
 rename {composable_kernel/test => test}/reduce/CMakeLists.txt (100%)
 rename {composable_kernel/test => test}/reduce/reduce_no_index.cpp (100%)
 rename {composable_kernel/test => test}/reduce/reduce_with_index.cpp (100%)
 rename {composable_kernel/test => test}/reference_conv_fwd/CMakeLists.txt (100%)
 rename {composable_kernel/test => test}/reference_conv_fwd/reference_conv_fwd.cpp (100%)
 rename {composable_kernel/test => test}/softmax/CMakeLists.txt (100%)
 rename {composable_kernel/test => test}/softmax/test_softmax_interface.cpp (100%)
 rename {composable_kernel/test => test}/softmax/test_softmax_rank3.cpp (100%)
 rename {composable_kernel/test => test}/softmax/test_softmax_rank4.cpp (100%)
 rename {composable_kernel/test => test}/softmax/test_softmax_ut_cases.inc (100%)
 rename {composable_kernel/test => test}/softmax/test_softmax_util.hpp (100%)
 rename {composable_kernel/test => test}/space_filling_curve/CMakeLists.txt (100%)
 rename {composable_kernel/test => test}/space_filling_curve/space_filling_curve.cpp (100%)
 rename {composable_kernel/test => test}/wmma_op/CMakeLists.txt (100%)
 rename {composable_kernel/test => test}/wmma_op/wmma_op.cpp (100%)
 rename {composable_kernel/test => test}/wmma_op/wmma_op_util.hpp (100%)

diff --git a/composable_kernel/CHANGELOG.md b/CHANGELOG.md
similarity index 100%
rename from composable_kernel/CHANGELOG.md
rename to CHANGELOG.md
diff --git a/composable_kernel/CITATION.cff b/CITATION.cff
similarity index 100%
rename from composable_kernel/CITATION.cff
rename to CITATION.cff
diff --git a/composable_kernel/CMakeLists.txt b/CMakeLists.txt
similarity index 100%
rename from composable_kernel/CMakeLists.txt
rename to CMakeLists.txt
diff --git a/composable_kernel/CONTRIBUTORS.md b/CONTRIBUTORS.md
similarity index 100%
rename from composable_kernel/CONTRIBUTORS.md
rename to CONTRIBUTORS.md
diff --git a/composable_kernel/Config.cmake.in b/Config.cmake.in
similarity index 100%
rename from composable_kernel/Config.cmake.in
rename to Config.cmake.in
diff --git a/composable_kernel/Dockerfile b/Dockerfile
similarity index 100%
rename from composable_kernel/Dockerfile
rename to Dockerfile
diff --git a/composable_kernel/Jenkinsfile b/Jenkinsfile
similarity index 100%
rename from composable_kernel/Jenkinsfile
rename to Jenkinsfile
diff --git a/composable_kernel/LICENSE b/LICENSE
similarity index 100%
rename from composable_kernel/LICENSE
rename to LICENSE
diff --git a/composable_kernel/README.md b/README.md
similarity index 100%
rename from composable_kernel/README.md
rename to README.md
diff --git a/composable_kernel/client_example/01_gemm/CMakeLists.txt b/client_example/01_gemm/CMakeLists.txt
similarity index 100%
rename from composable_kernel/client_example/01_gemm/CMakeLists.txt
rename to client_example/01_gemm/CMakeLists.txt
diff --git a/composable_kernel/client_example/01_gemm/gemm.cpp b/client_example/01_gemm/gemm.cpp
similarity index 100%
rename from composable_kernel/client_example/01_gemm/gemm.cpp
rename to client_example/01_gemm/gemm.cpp
diff --git a/composable_kernel/client_example/02_gemm_add_add_fastgelu/CMakeLists.txt b/client_example/02_gemm_add_add_fastgelu/CMakeLists.txt
similarity index 100%
rename from composable_kernel/client_example/02_gemm_add_add_fastgelu/CMakeLists.txt
rename to client_example/02_gemm_add_add_fastgelu/CMakeLists.txt
diff --git a/composable_kernel/client_example/02_gemm_add_add_fastgelu/gemm_add_add_fastgelu.cpp b/client_example/02_gemm_add_add_fastgelu/gemm_add_add_fastgelu.cpp
similarity index 100%
rename from composable_kernel/client_example/02_gemm_add_add_fastgelu/gemm_add_add_fastgelu.cpp
rename to client_example/02_gemm_add_add_fastgelu/gemm_add_add_fastgelu.cpp
diff --git a/composable_kernel/client_example/02_gemm_add_add_fastgelu/gemm_add_add_fastgelu_generic.cpp b/client_example/02_gemm_add_add_fastgelu/gemm_add_add_fastgelu_generic.cpp
similarity index 100%
rename from composable_kernel/client_example/02_gemm_add_add_fastgelu/gemm_add_add_fastgelu_generic.cpp
rename to client_example/02_gemm_add_add_fastgelu/gemm_add_add_fastgelu_generic.cpp
diff --git a/composable_kernel/client_example/02_gemm_add_add_fastgelu/gemm_add_fastgelu.cpp b/client_example/02_gemm_add_add_fastgelu/gemm_add_fastgelu.cpp
similarity index 100%
rename from composable_kernel/client_example/02_gemm_add_add_fastgelu/gemm_add_fastgelu.cpp
rename to client_example/02_gemm_add_add_fastgelu/gemm_add_fastgelu.cpp
diff --git a/composable_kernel/client_example/02_gemm_add_add_fastgelu/gemm_add_fastgelu_generic.cpp b/client_example/02_gemm_add_add_fastgelu/gemm_add_fastgelu_generic.cpp
similarity index 100%
rename from composable_kernel/client_example/02_gemm_add_add_fastgelu/gemm_add_fastgelu_generic.cpp
rename to client_example/02_gemm_add_add_fastgelu/gemm_add_fastgelu_generic.cpp
diff --git a/composable_kernel/client_example/02_gemm_add_add_fastgelu/gemm_fastgelu.cpp b/client_example/02_gemm_add_add_fastgelu/gemm_fastgelu.cpp
similarity index 100%
rename from composable_kernel/client_example/02_gemm_add_add_fastgelu/gemm_fastgelu.cpp
rename to client_example/02_gemm_add_add_fastgelu/gemm_fastgelu.cpp
diff --git a/composable_kernel/client_example/02_gemm_add_add_fastgelu/gemm_fastgelu_generic.cpp b/client_example/02_gemm_add_add_fastgelu/gemm_fastgelu_generic.cpp
similarity index 100%
rename from composable_kernel/client_example/02_gemm_add_add_fastgelu/gemm_fastgelu_generic.cpp
rename to client_example/02_gemm_add_add_fastgelu/gemm_fastgelu_generic.cpp
diff --git a/composable_kernel/client_example/03_gemm_layernorm/CMakeLists.txt b/client_example/03_gemm_layernorm/CMakeLists.txt
similarity index 100%
rename from composable_kernel/client_example/03_gemm_layernorm/CMakeLists.txt
rename to client_example/03_gemm_layernorm/CMakeLists.txt
diff --git a/composable_kernel/client_example/03_gemm_layernorm/gemm_add_add_layernorm_naive.cpp b/client_example/03_gemm_layernorm/gemm_add_add_layernorm_naive.cpp
similarity index 100%
rename from composable_kernel/client_example/03_gemm_layernorm/gemm_add_add_layernorm_naive.cpp
rename to client_example/03_gemm_layernorm/gemm_add_add_layernorm_naive.cpp
diff --git a/composable_kernel/client_example/03_gemm_layernorm/gemm_add_relu_add_layernorm_welford.cpp b/client_example/03_gemm_layernorm/gemm_add_relu_add_layernorm_welford.cpp
similarity index 100%
rename from composable_kernel/client_example/03_gemm_layernorm/gemm_add_relu_add_layernorm_welford.cpp
rename to client_example/03_gemm_layernorm/gemm_add_relu_add_layernorm_welford.cpp
diff --git a/composable_kernel/client_example/04_contraction/CMakeLists.txt b/client_example/04_contraction/CMakeLists.txt
similarity index 100%
rename from composable_kernel/client_example/04_contraction/CMakeLists.txt
rename to client_example/04_contraction/CMakeLists.txt
diff --git a/composable_kernel/client_example/04_contraction/contraction_bilinear_fp32.cpp b/client_example/04_contraction/contraction_bilinear_fp32.cpp
similarity index 100%
rename from composable_kernel/client_example/04_contraction/contraction_bilinear_fp32.cpp
rename to client_example/04_contraction/contraction_bilinear_fp32.cpp
diff --git a/composable_kernel/client_example/04_contraction/contraction_bilinear_fp64.cpp b/client_example/04_contraction/contraction_bilinear_fp64.cpp
similarity index 100%
rename from composable_kernel/client_example/04_contraction/contraction_bilinear_fp64.cpp
rename to client_example/04_contraction/contraction_bilinear_fp64.cpp
diff --git a/composable_kernel/client_example/04_contraction/contraction_g1m2n3k1_add_xdl_fp16.cpp b/client_example/04_contraction/contraction_g1m2n3k1_add_xdl_fp16.cpp
similarity index 100%
rename from composable_kernel/client_example/04_contraction/contraction_g1m2n3k1_add_xdl_fp16.cpp
rename to client_example/04_contraction/contraction_g1m2n3k1_add_xdl_fp16.cpp
diff --git a/composable_kernel/client_example/04_contraction/contraction_scale_fp32.cpp b/client_example/04_contraction/contraction_scale_fp32.cpp
similarity index 100%
rename from composable_kernel/client_example/04_contraction/contraction_scale_fp32.cpp
rename to client_example/04_contraction/contraction_scale_fp32.cpp
diff --git a/composable_kernel/client_example/04_contraction/contraction_scale_fp64.cpp b/client_example/04_contraction/contraction_scale_fp64.cpp
similarity index 100%
rename from composable_kernel/client_example/04_contraction/contraction_scale_fp64.cpp
rename to client_example/04_contraction/contraction_scale_fp64.cpp
diff --git a/composable_kernel/client_example/05_layernorm/CMakeLists.txt b/client_example/05_layernorm/CMakeLists.txt
similarity index 100%
rename from composable_kernel/client_example/05_layernorm/CMakeLists.txt
rename to client_example/05_layernorm/CMakeLists.txt
diff --git a/composable_kernel/client_example/05_layernorm/layernorm2d.cpp b/client_example/05_layernorm/layernorm2d.cpp
similarity index 100%
rename from composable_kernel/client_example/05_layernorm/layernorm2d.cpp
rename to client_example/05_layernorm/layernorm2d.cpp
diff --git a/composable_kernel/client_example/06_softmax/CMakeLists.txt b/client_example/06_softmax/CMakeLists.txt
similarity index 100%
rename from composable_kernel/client_example/06_softmax/CMakeLists.txt
rename to client_example/06_softmax/CMakeLists.txt
diff --git a/composable_kernel/client_example/06_softmax/softmax4d.cpp b/client_example/06_softmax/softmax4d.cpp
similarity index 100%
rename from composable_kernel/client_example/06_softmax/softmax4d.cpp
rename to client_example/06_softmax/softmax4d.cpp
diff --git a/composable_kernel/client_example/07_grouped_convnd_fwd/CMakeLists.txt b/client_example/07_grouped_convnd_fwd/CMakeLists.txt
similarity index 100%
rename from composable_kernel/client_example/07_grouped_convnd_fwd/CMakeLists.txt
rename to client_example/07_grouped_convnd_fwd/CMakeLists.txt
diff --git a/composable_kernel/client_example/07_grouped_convnd_fwd/grouped_conv1d_fwd.cpp b/client_example/07_grouped_convnd_fwd/grouped_conv1d_fwd.cpp
similarity index 100%
rename from composable_kernel/client_example/07_grouped_convnd_fwd/grouped_conv1d_fwd.cpp
rename to client_example/07_grouped_convnd_fwd/grouped_conv1d_fwd.cpp
diff --git a/composable_kernel/client_example/07_grouped_convnd_fwd/grouped_conv2d_fwd.cpp b/client_example/07_grouped_convnd_fwd/grouped_conv2d_fwd.cpp
similarity index 100%
rename from composable_kernel/client_example/07_grouped_convnd_fwd/grouped_conv2d_fwd.cpp
rename to client_example/07_grouped_convnd_fwd/grouped_conv2d_fwd.cpp
diff --git a/composable_kernel/client_example/08_fused_attention/CMakeLists.txt b/client_example/08_fused_attention/CMakeLists.txt
similarity index 100%
rename from composable_kernel/client_example/08_fused_attention/CMakeLists.txt
rename to client_example/08_fused_attention/CMakeLists.txt
diff --git a/composable_kernel/client_example/08_fused_attention/fused_attention.cpp b/client_example/08_fused_attention/fused_attention.cpp
similarity index 100%
rename from composable_kernel/client_example/08_fused_attention/fused_attention.cpp
rename to client_example/08_fused_attention/fused_attention.cpp
diff --git a/composable_kernel/client_example/08_fused_attention/fused_attention_bias.cpp b/client_example/08_fused_attention/fused_attention_bias.cpp
similarity index 100%
rename from composable_kernel/client_example/08_fused_attention/fused_attention_bias.cpp
rename to client_example/08_fused_attention/fused_attention_bias.cpp
diff --git a/composable_kernel/client_example/09_quantization/CMakeLists.txt b/client_example/09_quantization/CMakeLists.txt
similarity index 100%
rename from composable_kernel/client_example/09_quantization/CMakeLists.txt
rename to client_example/09_quantization/CMakeLists.txt
diff --git a/composable_kernel/client_example/09_quantization/conv2d_fwd_bias_relu_perchannel_quantization.cpp b/client_example/09_quantization/conv2d_fwd_bias_relu_perchannel_quantization.cpp
similarity index 100%
rename from composable_kernel/client_example/09_quantization/conv2d_fwd_bias_relu_perchannel_quantization.cpp
rename to client_example/09_quantization/conv2d_fwd_bias_relu_perchannel_quantization.cpp
diff --git a/composable_kernel/client_example/09_quantization/conv2d_fwd_bias_relu_perlayer_quantization.cpp b/client_example/09_quantization/conv2d_fwd_bias_relu_perlayer_quantization.cpp
similarity index 100%
rename from composable_kernel/client_example/09_quantization/conv2d_fwd_bias_relu_perlayer_quantization.cpp
rename to client_example/09_quantization/conv2d_fwd_bias_relu_perlayer_quantization.cpp
diff --git a/composable_kernel/client_example/09_quantization/conv2d_fwd_bias_tanh_perchannel_quantization.cpp b/client_example/09_quantization/conv2d_fwd_bias_tanh_perchannel_quantization.cpp
similarity index 100%
rename from composable_kernel/client_example/09_quantization/conv2d_fwd_bias_tanh_perchannel_quantization.cpp
rename to client_example/09_quantization/conv2d_fwd_bias_tanh_perchannel_quantization.cpp
diff --git a/composable_kernel/client_example/09_quantization/conv2d_fwd_bias_tanh_perlayer_quantization.cpp b/client_example/09_quantization/conv2d_fwd_bias_tanh_perlayer_quantization.cpp
similarity index 100%
rename from composable_kernel/client_example/09_quantization/conv2d_fwd_bias_tanh_perlayer_quantization.cpp
rename to client_example/09_quantization/conv2d_fwd_bias_tanh_perlayer_quantization.cpp
diff --git a/composable_kernel/client_example/09_quantization/conv2d_fwd_perchannel_quantization.cpp b/client_example/09_quantization/conv2d_fwd_perchannel_quantization.cpp
similarity index 100%
rename from composable_kernel/client_example/09_quantization/conv2d_fwd_perchannel_quantization.cpp
rename to client_example/09_quantization/conv2d_fwd_perchannel_quantization.cpp
diff --git a/composable_kernel/client_example/09_quantization/conv2d_fwd_perlayer_quantization.cpp b/client_example/09_quantization/conv2d_fwd_perlayer_quantization.cpp
similarity index 100%
rename from composable_kernel/client_example/09_quantization/conv2d_fwd_perlayer_quantization.cpp
rename to client_example/09_quantization/conv2d_fwd_perlayer_quantization.cpp
diff --git a/composable_kernel/client_example/09_quantization/gemm_quantization.cpp b/client_example/09_quantization/gemm_quantization.cpp
similarity index 100%
rename from composable_kernel/client_example/09_quantization/gemm_quantization.cpp
rename to client_example/09_quantization/gemm_quantization.cpp
diff --git a/composable_kernel/client_example/10_grouped_convnd_bwd_data/CMakeLists.txt b/client_example/10_grouped_convnd_bwd_data/CMakeLists.txt
similarity index 100%
rename from composable_kernel/client_example/10_grouped_convnd_bwd_data/CMakeLists.txt
rename to client_example/10_grouped_convnd_bwd_data/CMakeLists.txt
diff --git a/composable_kernel/client_example/10_grouped_convnd_bwd_data/grouped_conv2d_bwd_data.cpp b/client_example/10_grouped_convnd_bwd_data/grouped_conv2d_bwd_data.cpp
similarity index 100%
rename from composable_kernel/client_example/10_grouped_convnd_bwd_data/grouped_conv2d_bwd_data.cpp
rename to client_example/10_grouped_convnd_bwd_data/grouped_conv2d_bwd_data.cpp
diff --git a/composable_kernel/client_example/10_grouped_convnd_bwd_data/grouped_conv3d_bwd_data.cpp b/client_example/10_grouped_convnd_bwd_data/grouped_conv3d_bwd_data.cpp
similarity index 100%
rename from composable_kernel/client_example/10_grouped_convnd_bwd_data/grouped_conv3d_bwd_data.cpp
rename to client_example/10_grouped_convnd_bwd_data/grouped_conv3d_bwd_data.cpp
diff --git a/composable_kernel/client_example/10_grouped_convnd_bwd_data/grouped_conv3d_bwd_data_input_fp16_comp_bf8f8.cpp b/client_example/10_grouped_convnd_bwd_data/grouped_conv3d_bwd_data_input_fp16_comp_bf8f8.cpp
similarity index 100%
rename from composable_kernel/client_example/10_grouped_convnd_bwd_data/grouped_conv3d_bwd_data_input_fp16_comp_bf8f8.cpp
rename to client_example/10_grouped_convnd_bwd_data/grouped_conv3d_bwd_data_input_fp16_comp_bf8f8.cpp
diff --git a/composable_kernel/client_example/11_grouped_conv_bwd_weight/CMakeLists.txt b/client_example/11_grouped_conv_bwd_weight/CMakeLists.txt
similarity index 100%
rename from composable_kernel/client_example/11_grouped_conv_bwd_weight/CMakeLists.txt
rename to client_example/11_grouped_conv_bwd_weight/CMakeLists.txt
diff --git a/composable_kernel/client_example/11_grouped_conv_bwd_weight/common.hpp b/client_example/11_grouped_conv_bwd_weight/common.hpp
similarity index 100%
rename from composable_kernel/client_example/11_grouped_conv_bwd_weight/common.hpp
rename to client_example/11_grouped_conv_bwd_weight/common.hpp
diff --git a/composable_kernel/client_example/11_grouped_conv_bwd_weight/grouped_conv1d_bwd_weight_fp16.cpp b/client_example/11_grouped_conv_bwd_weight/grouped_conv1d_bwd_weight_fp16.cpp
similarity index 100%
rename from composable_kernel/client_example/11_grouped_conv_bwd_weight/grouped_conv1d_bwd_weight_fp16.cpp
rename to client_example/11_grouped_conv_bwd_weight/grouped_conv1d_bwd_weight_fp16.cpp
diff --git a/composable_kernel/client_example/11_grouped_conv_bwd_weight/grouped_conv2d_bwd_weight_fp16.cpp b/client_example/11_grouped_conv_bwd_weight/grouped_conv2d_bwd_weight_fp16.cpp
similarity index 100%
rename from composable_kernel/client_example/11_grouped_conv_bwd_weight/grouped_conv2d_bwd_weight_fp16.cpp
rename to client_example/11_grouped_conv_bwd_weight/grouped_conv2d_bwd_weight_fp16.cpp
diff --git a/composable_kernel/client_example/11_grouped_conv_bwd_weight/grouped_conv3d_bwd_weight_fp16.cpp b/client_example/11_grouped_conv_bwd_weight/grouped_conv3d_bwd_weight_fp16.cpp
similarity index 100%
rename from composable_kernel/client_example/11_grouped_conv_bwd_weight/grouped_conv3d_bwd_weight_fp16.cpp
rename to client_example/11_grouped_conv_bwd_weight/grouped_conv3d_bwd_weight_fp16.cpp
diff --git a/composable_kernel/client_example/11_grouped_conv_bwd_weight/grouped_conv3d_bwd_weight_fp32.cpp b/client_example/11_grouped_conv_bwd_weight/grouped_conv3d_bwd_weight_fp32.cpp
similarity index 100%
rename from composable_kernel/client_example/11_grouped_conv_bwd_weight/grouped_conv3d_bwd_weight_fp32.cpp
rename to client_example/11_grouped_conv_bwd_weight/grouped_conv3d_bwd_weight_fp32.cpp
diff --git a/composable_kernel/client_example/12_elementwise_normalization/CMakeLists.txt b/client_example/12_elementwise_normalization/CMakeLists.txt
similarity index 100%
rename from composable_kernel/client_example/12_elementwise_normalization/CMakeLists.txt
rename to client_example/12_elementwise_normalization/CMakeLists.txt
diff --git a/composable_kernel/client_example/12_elementwise_normalization/elementwise_layernorm2d.cpp b/client_example/12_elementwise_normalization/elementwise_layernorm2d.cpp
similarity index 100%
rename from composable_kernel/client_example/12_elementwise_normalization/elementwise_layernorm2d.cpp
rename to client_example/12_elementwise_normalization/elementwise_layernorm2d.cpp
diff --git a/composable_kernel/client_example/13_batchnorm/CMakeLists.txt b/client_example/13_batchnorm/CMakeLists.txt
similarity index 100%
rename from composable_kernel/client_example/13_batchnorm/CMakeLists.txt
rename to client_example/13_batchnorm/CMakeLists.txt
diff --git a/composable_kernel/client_example/13_batchnorm/batchnorm_bwd_nhwc.cpp b/client_example/13_batchnorm/batchnorm_bwd_nhwc.cpp
similarity index 100%
rename from composable_kernel/client_example/13_batchnorm/batchnorm_bwd_nhwc.cpp
rename to client_example/13_batchnorm/batchnorm_bwd_nhwc.cpp
diff --git a/composable_kernel/client_example/13_batchnorm/batchnorm_fwd_nhwc.cpp b/client_example/13_batchnorm/batchnorm_fwd_nhwc.cpp
similarity index 100%
rename from composable_kernel/client_example/13_batchnorm/batchnorm_fwd_nhwc.cpp
rename to client_example/13_batchnorm/batchnorm_fwd_nhwc.cpp
diff --git a/composable_kernel/client_example/13_batchnorm/batchnorm_infer_nhwc.cpp b/client_example/13_batchnorm/batchnorm_infer_nhwc.cpp
similarity index 100%
rename from composable_kernel/client_example/13_batchnorm/batchnorm_infer_nhwc.cpp
rename to client_example/13_batchnorm/batchnorm_infer_nhwc.cpp
diff --git a/composable_kernel/client_example/14_instance_id/CMakeLists.txt b/client_example/14_instance_id/CMakeLists.txt
similarity index 100%
rename from composable_kernel/client_example/14_instance_id/CMakeLists.txt
rename to client_example/14_instance_id/CMakeLists.txt
diff --git a/composable_kernel/client_example/14_instance_id/batchnorm_fwd_instance_id.cpp b/client_example/14_instance_id/batchnorm_fwd_instance_id.cpp
similarity index 100%
rename from composable_kernel/client_example/14_instance_id/batchnorm_fwd_instance_id.cpp
rename to client_example/14_instance_id/batchnorm_fwd_instance_id.cpp
diff --git a/composable_kernel/client_example/15_convnd_bwd_data/CMakeLists.txt b/client_example/15_convnd_bwd_data/CMakeLists.txt
similarity index 100%
rename from composable_kernel/client_example/15_convnd_bwd_data/CMakeLists.txt
rename to client_example/15_convnd_bwd_data/CMakeLists.txt
diff --git a/composable_kernel/client_example/15_convnd_bwd_data/common.hpp b/client_example/15_convnd_bwd_data/common.hpp
similarity index 100%
rename from composable_kernel/client_example/15_convnd_bwd_data/common.hpp
rename to client_example/15_convnd_bwd_data/common.hpp
diff --git a/composable_kernel/client_example/15_convnd_bwd_data/conv3d_bwd_data_fp16.cpp b/client_example/15_convnd_bwd_data/conv3d_bwd_data_fp16.cpp
similarity index 100%
rename from composable_kernel/client_example/15_convnd_bwd_data/conv3d_bwd_data_fp16.cpp
rename to client_example/15_convnd_bwd_data/conv3d_bwd_data_fp16.cpp
diff --git a/composable_kernel/client_example/15_convnd_bwd_data/conv3d_bwd_data_fp32.cpp b/client_example/15_convnd_bwd_data/conv3d_bwd_data_fp32.cpp
similarity index 100%
rename from composable_kernel/client_example/15_convnd_bwd_data/conv3d_bwd_data_fp32.cpp
rename to client_example/15_convnd_bwd_data/conv3d_bwd_data_fp32.cpp
diff --git a/composable_kernel/client_example/15_gemm_add_multiply/CMakeLists.txt b/client_example/15_gemm_add_multiply/CMakeLists.txt
similarity index 100%
rename from composable_kernel/client_example/15_gemm_add_multiply/CMakeLists.txt
rename to client_example/15_gemm_add_multiply/CMakeLists.txt
diff --git a/composable_kernel/client_example/15_gemm_add_multiply/gemm_add_multiply.cpp b/client_example/15_gemm_add_multiply/gemm_add_multiply.cpp
similarity index 100%
rename from composable_kernel/client_example/15_gemm_add_multiply/gemm_add_multiply.cpp
rename to client_example/15_gemm_add_multiply/gemm_add_multiply.cpp
diff --git a/composable_kernel/client_example/15_reduce/CMakeLists.txt b/client_example/15_reduce/CMakeLists.txt
similarity index 100%
rename from composable_kernel/client_example/15_reduce/CMakeLists.txt
rename to client_example/15_reduce/CMakeLists.txt
diff --git a/composable_kernel/client_example/15_reduce/reduce_nhwc_c.cpp b/client_example/15_reduce/reduce_nhwc_c.cpp
similarity index 100%
rename from composable_kernel/client_example/15_reduce/reduce_nhwc_c.cpp
rename to client_example/15_reduce/reduce_nhwc_c.cpp
diff --git a/composable_kernel/client_example/16_convnd_fwd/CMakeLists.txt b/client_example/16_convnd_fwd/CMakeLists.txt
similarity index 100%
rename from composable_kernel/client_example/16_convnd_fwd/CMakeLists.txt
rename to client_example/16_convnd_fwd/CMakeLists.txt
diff --git a/composable_kernel/client_example/16_convnd_fwd/common.hpp b/client_example/16_convnd_fwd/common.hpp
similarity index 100%
rename from composable_kernel/client_example/16_convnd_fwd/common.hpp
rename to client_example/16_convnd_fwd/common.hpp
diff --git a/composable_kernel/client_example/16_convnd_fwd/conv3d_fwd_fp16.cpp b/client_example/16_convnd_fwd/conv3d_fwd_fp16.cpp
similarity index 100%
rename from composable_kernel/client_example/16_convnd_fwd/conv3d_fwd_fp16.cpp
rename to client_example/16_convnd_fwd/conv3d_fwd_fp16.cpp
diff --git a/composable_kernel/client_example/16_convnd_fwd/conv3d_fwd_fp16_comp_fp8.cpp b/client_example/16_convnd_fwd/conv3d_fwd_fp16_comp_fp8.cpp
similarity index 100%
rename from composable_kernel/client_example/16_convnd_fwd/conv3d_fwd_fp16_comp_fp8.cpp
rename to client_example/16_convnd_fwd/conv3d_fwd_fp16_comp_fp8.cpp
diff --git a/composable_kernel/client_example/16_convnd_fwd/conv3d_fwd_fp32.cpp b/client_example/16_convnd_fwd/conv3d_fwd_fp32.cpp
similarity index 100%
rename from composable_kernel/client_example/16_convnd_fwd/conv3d_fwd_fp32.cpp
rename to client_example/16_convnd_fwd/conv3d_fwd_fp32.cpp
diff --git a/composable_kernel/client_example/17_grouped_gemm_fastgelu/CMakeLists.txt b/client_example/17_grouped_gemm_fastgelu/CMakeLists.txt
similarity index 100%
rename from composable_kernel/client_example/17_grouped_gemm_fastgelu/CMakeLists.txt
rename to client_example/17_grouped_gemm_fastgelu/CMakeLists.txt
diff --git a/composable_kernel/client_example/17_grouped_gemm_fastgelu/grouped_gemm_fastgelu.cpp b/client_example/17_grouped_gemm_fastgelu/grouped_gemm_fastgelu.cpp
similarity index 100%
rename from composable_kernel/client_example/17_grouped_gemm_fastgelu/grouped_gemm_fastgelu.cpp
rename to client_example/17_grouped_gemm_fastgelu/grouped_gemm_fastgelu.cpp
diff --git a/composable_kernel/client_example/18_groupnorm/CMakeLists.txt b/client_example/18_groupnorm/CMakeLists.txt
similarity index 100%
rename from composable_kernel/client_example/18_groupnorm/CMakeLists.txt
rename to client_example/18_groupnorm/CMakeLists.txt
diff --git a/composable_kernel/client_example/18_groupnorm/groupnorm_swish.cpp b/client_example/18_groupnorm/groupnorm_swish.cpp
similarity index 100%
rename from composable_kernel/client_example/18_groupnorm/groupnorm_swish.cpp
rename to client_example/18_groupnorm/groupnorm_swish.cpp
diff --git a/composable_kernel/client_example/19_pool/CMakeLists.txt b/client_example/19_pool/CMakeLists.txt
similarity index 100%
rename from composable_kernel/client_example/19_pool/CMakeLists.txt
rename to client_example/19_pool/CMakeLists.txt
diff --git a/composable_kernel/client_example/19_pool/avg_pool3d_bwd.cpp b/client_example/19_pool/avg_pool3d_bwd.cpp
similarity index 100%
rename from composable_kernel/client_example/19_pool/avg_pool3d_bwd.cpp
rename to client_example/19_pool/avg_pool3d_bwd.cpp
diff --git a/composable_kernel/client_example/19_pool/avg_pool3d_fwd.cpp b/client_example/19_pool/avg_pool3d_fwd.cpp
similarity index 100%
rename from composable_kernel/client_example/19_pool/avg_pool3d_fwd.cpp
rename to client_example/19_pool/avg_pool3d_fwd.cpp
diff --git a/composable_kernel/client_example/19_pool/max_pool2d_bwd.cpp b/client_example/19_pool/max_pool2d_bwd.cpp
similarity index 100%
rename from composable_kernel/client_example/19_pool/max_pool2d_bwd.cpp
rename to client_example/19_pool/max_pool2d_bwd.cpp
diff --git a/composable_kernel/client_example/19_pool/max_pool2d_fwd.cpp b/client_example/19_pool/max_pool2d_fwd.cpp
similarity index 100%
rename from composable_kernel/client_example/19_pool/max_pool2d_fwd.cpp
rename to client_example/19_pool/max_pool2d_fwd.cpp
diff --git a/composable_kernel/client_example/20_splitk_gemm/CMakeLists.txt b/client_example/20_splitk_gemm/CMakeLists.txt
similarity index 100%
rename from composable_kernel/client_example/20_splitk_gemm/CMakeLists.txt
rename to client_example/20_splitk_gemm/CMakeLists.txt
diff --git a/composable_kernel/client_example/20_splitk_gemm/splitK_gemm_fp16_f8.cpp b/client_example/20_splitk_gemm/splitK_gemm_fp16_f8.cpp
similarity index 100%
rename from composable_kernel/client_example/20_splitk_gemm/splitK_gemm_fp16_f8.cpp
rename to client_example/20_splitk_gemm/splitK_gemm_fp16_f8.cpp
diff --git a/composable_kernel/client_example/21_grouped_gemm_bias/CMakeLists.txt b/client_example/21_grouped_gemm_bias/CMakeLists.txt
similarity index 100%
rename from composable_kernel/client_example/21_grouped_gemm_bias/CMakeLists.txt
rename to client_example/21_grouped_gemm_bias/CMakeLists.txt
diff --git a/composable_kernel/client_example/21_grouped_gemm_bias/grouped_gemm_fixed_nk_bias_fp16.cpp b/client_example/21_grouped_gemm_bias/grouped_gemm_fixed_nk_bias_fp16.cpp
similarity index 100%
rename from composable_kernel/client_example/21_grouped_gemm_bias/grouped_gemm_fixed_nk_bias_fp16.cpp
rename to client_example/21_grouped_gemm_bias/grouped_gemm_fixed_nk_bias_fp16.cpp
diff --git a/composable_kernel/client_example/22_grouped_gemm/CMakeLists.txt b/client_example/22_grouped_gemm/CMakeLists.txt
similarity index 100%
rename from composable_kernel/client_example/22_grouped_gemm/CMakeLists.txt
rename to client_example/22_grouped_gemm/CMakeLists.txt
diff --git a/composable_kernel/client_example/22_grouped_gemm/grouped_gemm_fixed_nk_fp16.cpp b/client_example/22_grouped_gemm/grouped_gemm_fixed_nk_fp16.cpp
similarity index 100%
rename from composable_kernel/client_example/22_grouped_gemm/grouped_gemm_fixed_nk_fp16.cpp
rename to client_example/22_grouped_gemm/grouped_gemm_fixed_nk_fp16.cpp
diff --git a/composable_kernel/client_example/22_grouped_gemm/grouped_gemm_fixed_nk_fp8.cpp b/client_example/22_grouped_gemm/grouped_gemm_fixed_nk_fp8.cpp
similarity index 100%
rename from composable_kernel/client_example/22_grouped_gemm/grouped_gemm_fixed_nk_fp8.cpp
rename to client_example/22_grouped_gemm/grouped_gemm_fixed_nk_fp8.cpp
diff --git a/composable_kernel/client_example/22_grouped_gemm/grouped_gemm_fixed_nk_i8.cpp b/client_example/22_grouped_gemm/grouped_gemm_fixed_nk_i8.cpp
similarity index 100%
rename from composable_kernel/client_example/22_grouped_gemm/grouped_gemm_fixed_nk_i8.cpp
rename to client_example/22_grouped_gemm/grouped_gemm_fixed_nk_i8.cpp
diff --git a/composable_kernel/client_example/22_im2col_col2im/CMakeLists.txt b/client_example/22_im2col_col2im/CMakeLists.txt
similarity index 100%
rename from composable_kernel/client_example/22_im2col_col2im/CMakeLists.txt
rename to client_example/22_im2col_col2im/CMakeLists.txt
diff --git a/composable_kernel/client_example/22_im2col_col2im/column_to_image.cpp b/client_example/22_im2col_col2im/column_to_image.cpp
similarity index 100%
rename from composable_kernel/client_example/22_im2col_col2im/column_to_image.cpp
rename to client_example/22_im2col_col2im/column_to_image.cpp
diff --git a/composable_kernel/client_example/22_im2col_col2im/image_to_column.cpp b/client_example/22_im2col_col2im/image_to_column.cpp
similarity index 100%
rename from composable_kernel/client_example/22_im2col_col2im/image_to_column.cpp
rename to client_example/22_im2col_col2im/image_to_column.cpp
diff --git a/composable_kernel/client_example/CMakeLists.txt b/client_example/CMakeLists.txt
similarity index 100%
rename from composable_kernel/client_example/CMakeLists.txt
rename to client_example/CMakeLists.txt
diff --git a/composable_kernel/client_example/README.md b/client_example/README.md
similarity index 100%
rename from composable_kernel/client_example/README.md
rename to client_example/README.md
diff --git a/composable_kernel/cmake/Analyzers.cmake b/cmake/Analyzers.cmake
similarity index 100%
rename from composable_kernel/cmake/Analyzers.cmake
rename to cmake/Analyzers.cmake
diff --git a/composable_kernel/cmake/ClangTidy.cmake b/cmake/ClangTidy.cmake
similarity index 100%
rename from composable_kernel/cmake/ClangTidy.cmake
rename to cmake/ClangTidy.cmake
diff --git a/composable_kernel/cmake/CppCheck.cmake b/cmake/CppCheck.cmake
similarity index 100%
rename from composable_kernel/cmake/CppCheck.cmake
rename to cmake/CppCheck.cmake
diff --git a/composable_kernel/cmake/DoxygenDoc.cmake b/cmake/DoxygenDoc.cmake
similarity index 100%
rename from composable_kernel/cmake/DoxygenDoc.cmake
rename to cmake/DoxygenDoc.cmake
diff --git a/composable_kernel/cmake/EnableCompilerWarnings.cmake b/cmake/EnableCompilerWarnings.cmake
similarity index 100%
rename from composable_kernel/cmake/EnableCompilerWarnings.cmake
rename to cmake/EnableCompilerWarnings.cmake
diff --git a/composable_kernel/cmake/TargetFlags.cmake b/cmake/TargetFlags.cmake
similarity index 100%
rename from composable_kernel/cmake/TargetFlags.cmake
rename to cmake/TargetFlags.cmake
diff --git a/composable_kernel/cmake/googletest.cmake b/cmake/googletest.cmake
similarity index 100%
rename from composable_kernel/cmake/googletest.cmake
rename to cmake/googletest.cmake
diff --git a/composable_kernel/dev-requirements.txt b/dev-requirements.txt
similarity index 100%
rename from composable_kernel/dev-requirements.txt
rename to dev-requirements.txt
diff --git a/composable_kernel/docs/API_Reference_Guide.rst b/docs/API_Reference_Guide.rst
similarity index 100%
rename from composable_kernel/docs/API_Reference_Guide.rst
rename to docs/API_Reference_Guide.rst
diff --git a/composable_kernel/docs/Contributors_Guide.rst b/docs/Contributors_Guide.rst
similarity index 100%
rename from composable_kernel/docs/Contributors_Guide.rst
rename to docs/Contributors_Guide.rst
diff --git a/composable_kernel/docs/Supported_Primitives_Guide.rst b/docs/Supported_Primitives_Guide.rst
similarity index 100%
rename from composable_kernel/docs/Supported_Primitives_Guide.rst
rename to docs/Supported_Primitives_Guide.rst
diff --git a/composable_kernel/docs/conf.py b/docs/conf.py
similarity index 100%
rename from composable_kernel/docs/conf.py
rename to docs/conf.py
diff --git a/composable_kernel/docs/data/ck_component.png b/docs/data/ck_component.png
similarity index 100%
rename from composable_kernel/docs/data/ck_component.png
rename to docs/data/ck_component.png
diff --git a/composable_kernel/docs/data/ck_layer.png b/docs/data/ck_layer.png
similarity index 100%
rename from composable_kernel/docs/data/ck_layer.png
rename to docs/data/ck_layer.png
diff --git a/composable_kernel/docs/dockerhub.rst b/docs/dockerhub.rst
similarity index 100%
rename from composable_kernel/docs/dockerhub.rst
rename to docs/dockerhub.rst
diff --git a/composable_kernel/docs/doxygen/Doxyfile b/docs/doxygen/Doxyfile
similarity index 100%
rename from composable_kernel/docs/doxygen/Doxyfile
rename to docs/doxygen/Doxyfile
diff --git a/composable_kernel/docs/index.rst b/docs/index.rst
similarity index 100%
rename from composable_kernel/docs/index.rst
rename to docs/index.rst
diff --git a/composable_kernel/docs/license.rst b/docs/license.rst
similarity index 100%
rename from composable_kernel/docs/license.rst
rename to docs/license.rst
diff --git a/composable_kernel/docs/refs.bib b/docs/refs.bib
similarity index 100%
rename from composable_kernel/docs/refs.bib
rename to docs/refs.bib
diff --git a/composable_kernel/docs/sphinx/_toc.yml.in b/docs/sphinx/_toc.yml.in
similarity index 100%
rename from composable_kernel/docs/sphinx/_toc.yml.in
rename to docs/sphinx/_toc.yml.in
diff --git a/composable_kernel/docs/sphinx/requirements.in b/docs/sphinx/requirements.in
similarity index 100%
rename from composable_kernel/docs/sphinx/requirements.in
rename to docs/sphinx/requirements.in
diff --git a/composable_kernel/docs/sphinx/requirements.txt b/docs/sphinx/requirements.txt
similarity index 100%
rename from composable_kernel/docs/sphinx/requirements.txt
rename to docs/sphinx/requirements.txt
diff --git a/composable_kernel/docs/tutorial_hello_world.rst b/docs/tutorial_hello_world.rst
similarity index 100%
rename from composable_kernel/docs/tutorial_hello_world.rst
rename to docs/tutorial_hello_world.rst
diff --git a/composable_kernel/example/01_gemm/CMakeLists.txt b/example/01_gemm/CMakeLists.txt
similarity index 100%
rename from composable_kernel/example/01_gemm/CMakeLists.txt
rename to example/01_gemm/CMakeLists.txt
diff --git a/composable_kernel/example/01_gemm/README.md b/example/01_gemm/README.md
similarity index 100%
rename from composable_kernel/example/01_gemm/README.md
rename to example/01_gemm/README.md
diff --git a/composable_kernel/example/01_gemm/common.hpp b/example/01_gemm/common.hpp
similarity index 100%
rename from composable_kernel/example/01_gemm/common.hpp
rename to example/01_gemm/common.hpp
diff --git a/composable_kernel/example/01_gemm/gemm_dl_fp16.cpp b/example/01_gemm/gemm_dl_fp16.cpp
similarity index 100%
rename from composable_kernel/example/01_gemm/gemm_dl_fp16.cpp
rename to example/01_gemm/gemm_dl_fp16.cpp
diff --git a/composable_kernel/example/01_gemm/gemm_dl_fp32.cpp b/example/01_gemm/gemm_dl_fp32.cpp
similarity index 100%
rename from composable_kernel/example/01_gemm/gemm_dl_fp32.cpp
rename to example/01_gemm/gemm_dl_fp32.cpp
diff --git a/composable_kernel/example/01_gemm/gemm_dl_int4.cpp b/example/01_gemm/gemm_dl_int4.cpp
similarity index 100%
rename from composable_kernel/example/01_gemm/gemm_dl_int4.cpp
rename to example/01_gemm/gemm_dl_int4.cpp
diff --git a/composable_kernel/example/01_gemm/gemm_dl_int8.cpp b/example/01_gemm/gemm_dl_int8.cpp
similarity index 100%
rename from composable_kernel/example/01_gemm/gemm_dl_int8.cpp
rename to example/01_gemm/gemm_dl_int8.cpp
diff --git a/composable_kernel/example/01_gemm/gemm_dpp_fp16.cpp b/example/01_gemm/gemm_dpp_fp16.cpp
similarity index 100%
rename from composable_kernel/example/01_gemm/gemm_dpp_fp16.cpp
rename to example/01_gemm/gemm_dpp_fp16.cpp
diff --git a/composable_kernel/example/01_gemm/gemm_wmma_fp16.cpp b/example/01_gemm/gemm_wmma_fp16.cpp
similarity index 100%
rename from composable_kernel/example/01_gemm/gemm_wmma_fp16.cpp
rename to example/01_gemm/gemm_wmma_fp16.cpp
diff --git a/composable_kernel/example/01_gemm/gemm_xdl_bf16.cpp b/example/01_gemm/gemm_xdl_bf16.cpp
similarity index 100%
rename from composable_kernel/example/01_gemm/gemm_xdl_bf16.cpp
rename to example/01_gemm/gemm_xdl_bf16.cpp
diff --git a/composable_kernel/example/01_gemm/gemm_xdl_bf16_rtn.cpp b/example/01_gemm/gemm_xdl_bf16_rtn.cpp
similarity index 100%
rename from composable_kernel/example/01_gemm/gemm_xdl_bf16_rtn.cpp
rename to example/01_gemm/gemm_xdl_bf16_rtn.cpp
diff --git a/composable_kernel/example/01_gemm/gemm_xdl_fp16.cpp b/example/01_gemm/gemm_xdl_fp16.cpp
similarity index 100%
rename from composable_kernel/example/01_gemm/gemm_xdl_fp16.cpp
rename to example/01_gemm/gemm_xdl_fp16.cpp
diff --git a/composable_kernel/example/01_gemm/gemm_xdl_fp16_fp8.cpp b/example/01_gemm/gemm_xdl_fp16_fp8.cpp
similarity index 100%
rename from composable_kernel/example/01_gemm/gemm_xdl_fp16_fp8.cpp
rename to example/01_gemm/gemm_xdl_fp16_fp8.cpp
diff --git a/composable_kernel/example/01_gemm/gemm_xdl_fp64.cpp b/example/01_gemm/gemm_xdl_fp64.cpp
similarity index 100%
rename from composable_kernel/example/01_gemm/gemm_xdl_fp64.cpp
rename to example/01_gemm/gemm_xdl_fp64.cpp
diff --git a/composable_kernel/example/01_gemm/gemm_xdl_fp8.cpp b/example/01_gemm/gemm_xdl_fp8.cpp
similarity index 100%
rename from composable_kernel/example/01_gemm/gemm_xdl_fp8.cpp
rename to example/01_gemm/gemm_xdl_fp8.cpp
diff --git a/composable_kernel/example/01_gemm/gemm_xdl_fp8_bf8.cpp b/example/01_gemm/gemm_xdl_fp8_bf8.cpp
similarity index 100%
rename from composable_kernel/example/01_gemm/gemm_xdl_fp8_bf8.cpp
rename to example/01_gemm/gemm_xdl_fp8_bf8.cpp
diff --git a/composable_kernel/example/01_gemm/gemm_xdl_int4.cpp b/example/01_gemm/gemm_xdl_int4.cpp
similarity index 100%
rename from composable_kernel/example/01_gemm/gemm_xdl_int4.cpp
rename to example/01_gemm/gemm_xdl_int4.cpp
diff --git a/composable_kernel/example/01_gemm/gemm_xdl_int8.cpp b/example/01_gemm/gemm_xdl_int8.cpp
similarity index 100%
rename from composable_kernel/example/01_gemm/gemm_xdl_int8.cpp
rename to example/01_gemm/gemm_xdl_int8.cpp
diff --git a/composable_kernel/example/01_gemm/gemm_xdl_skip_b_lds_fp16.cpp b/example/01_gemm/gemm_xdl_skip_b_lds_fp16.cpp
similarity index 100%
rename from composable_kernel/example/01_gemm/gemm_xdl_skip_b_lds_fp16.cpp
rename to example/01_gemm/gemm_xdl_skip_b_lds_fp16.cpp
diff --git a/composable_kernel/example/01_gemm/gemm_xdl_streamk.cpp b/example/01_gemm/gemm_xdl_streamk.cpp
similarity index 100%
rename from composable_kernel/example/01_gemm/gemm_xdl_streamk.cpp
rename to example/01_gemm/gemm_xdl_streamk.cpp
diff --git a/composable_kernel/example/01_gemm/gemm_xdl_wavelet_fp16.cpp b/example/01_gemm/gemm_xdl_wavelet_fp16.cpp
similarity index 100%
rename from composable_kernel/example/01_gemm/gemm_xdl_wavelet_fp16.cpp
rename to example/01_gemm/gemm_xdl_wavelet_fp16.cpp
diff --git a/composable_kernel/example/01_gemm/run_gemm_example.inc b/example/01_gemm/run_gemm_example.inc
similarity index 100%
rename from composable_kernel/example/01_gemm/run_gemm_example.inc
rename to example/01_gemm/run_gemm_example.inc
diff --git a/composable_kernel/example/02_gemm_bilinear/CMakeLists.txt b/example/02_gemm_bilinear/CMakeLists.txt
similarity index 100%
rename from composable_kernel/example/02_gemm_bilinear/CMakeLists.txt
rename to example/02_gemm_bilinear/CMakeLists.txt
diff --git a/composable_kernel/example/02_gemm_bilinear/README.md b/example/02_gemm_bilinear/README.md
similarity index 100%
rename from composable_kernel/example/02_gemm_bilinear/README.md
rename to example/02_gemm_bilinear/README.md
diff --git a/composable_kernel/example/02_gemm_bilinear/gemm_bilinear_wmma_fp16.cpp b/example/02_gemm_bilinear/gemm_bilinear_wmma_fp16.cpp
similarity index 100%
rename from composable_kernel/example/02_gemm_bilinear/gemm_bilinear_wmma_fp16.cpp
rename to example/02_gemm_bilinear/gemm_bilinear_wmma_fp16.cpp
diff --git a/composable_kernel/example/02_gemm_bilinear/gemm_bilinear_wmma_int8.cpp b/example/02_gemm_bilinear/gemm_bilinear_wmma_int8.cpp
similarity index 100%
rename from composable_kernel/example/02_gemm_bilinear/gemm_bilinear_wmma_int8.cpp
rename to example/02_gemm_bilinear/gemm_bilinear_wmma_int8.cpp
diff --git a/composable_kernel/example/02_gemm_bilinear/gemm_bilinear_xdl_fp16.cpp b/example/02_gemm_bilinear/gemm_bilinear_xdl_fp16.cpp
similarity index 100%
rename from composable_kernel/example/02_gemm_bilinear/gemm_bilinear_xdl_fp16.cpp
rename to example/02_gemm_bilinear/gemm_bilinear_xdl_fp16.cpp
diff --git a/composable_kernel/example/03_gemm_bias_relu/CMakeLists.txt b/example/03_gemm_bias_relu/CMakeLists.txt
similarity index 100%
rename from composable_kernel/example/03_gemm_bias_relu/CMakeLists.txt
rename to example/03_gemm_bias_relu/CMakeLists.txt
diff --git a/composable_kernel/example/03_gemm_bias_relu/README.md b/example/03_gemm_bias_relu/README.md
similarity index 100%
rename from composable_kernel/example/03_gemm_bias_relu/README.md
rename to example/03_gemm_bias_relu/README.md
diff --git a/composable_kernel/example/03_gemm_bias_relu/gemm_bias_relu_xdl_fp16.cpp b/example/03_gemm_bias_relu/gemm_bias_relu_xdl_fp16.cpp
similarity index 100%
rename from composable_kernel/example/03_gemm_bias_relu/gemm_bias_relu_xdl_fp16.cpp
rename to example/03_gemm_bias_relu/gemm_bias_relu_xdl_fp16.cpp
diff --git a/composable_kernel/example/04_gemm_add_add_fastgelu/CMakeLists.txt b/example/04_gemm_add_add_fastgelu/CMakeLists.txt
similarity index 100%
rename from composable_kernel/example/04_gemm_add_add_fastgelu/CMakeLists.txt
rename to example/04_gemm_add_add_fastgelu/CMakeLists.txt
diff --git a/composable_kernel/example/04_gemm_add_add_fastgelu/README.md b/example/04_gemm_add_add_fastgelu/README.md
similarity index 100%
rename from composable_kernel/example/04_gemm_add_add_fastgelu/README.md
rename to example/04_gemm_add_add_fastgelu/README.md
diff --git a/composable_kernel/example/04_gemm_add_add_fastgelu/common.hpp b/example/04_gemm_add_add_fastgelu/common.hpp
similarity index 100%
rename from composable_kernel/example/04_gemm_add_add_fastgelu/common.hpp
rename to example/04_gemm_add_add_fastgelu/common.hpp
diff --git a/composable_kernel/example/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_bf16.cpp b/example/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_bf16.cpp
similarity index 100%
rename from composable_kernel/example/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_bf16.cpp
rename to example/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_bf16.cpp
diff --git a/composable_kernel/example/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_fp16.cpp b/example/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_fp16.cpp
similarity index 100%
rename from composable_kernel/example/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_fp16.cpp
rename to example/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_fp16.cpp
diff --git a/composable_kernel/example/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_fp32.cpp b/example/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_fp32.cpp
similarity index 100%
rename from composable_kernel/example/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_fp32.cpp
rename to example/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_fp32.cpp
diff --git a/composable_kernel/example/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_int4.cpp b/example/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_int4.cpp
similarity index 100%
rename from composable_kernel/example/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_int4.cpp
rename to example/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_int4.cpp
diff --git a/composable_kernel/example/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_int8.cpp b/example/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_int8.cpp
similarity index 100%
rename from composable_kernel/example/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_int8.cpp
rename to example/04_gemm_add_add_fastgelu/gemm_add_add_fastgelu_xdl_int8.cpp
diff --git a/composable_kernel/example/04_gemm_add_add_fastgelu/run_gemm_add_add_fastgelu_example.inc b/example/04_gemm_add_add_fastgelu/run_gemm_add_add_fastgelu_example.inc
similarity index 100%
rename from composable_kernel/example/04_gemm_add_add_fastgelu/run_gemm_add_add_fastgelu_example.inc
rename to example/04_gemm_add_add_fastgelu/run_gemm_add_add_fastgelu_example.inc
diff --git a/composable_kernel/example/09_convnd_fwd/CMakeLists.txt b/example/09_convnd_fwd/CMakeLists.txt
similarity index 100%
rename from composable_kernel/example/09_convnd_fwd/CMakeLists.txt
rename to example/09_convnd_fwd/CMakeLists.txt
diff --git a/composable_kernel/example/09_convnd_fwd/README.md b/example/09_convnd_fwd/README.md
similarity index 100%
rename from composable_kernel/example/09_convnd_fwd/README.md
rename to example/09_convnd_fwd/README.md
diff --git a/composable_kernel/example/09_convnd_fwd/convnd_fwd_common.hpp b/example/09_convnd_fwd/convnd_fwd_common.hpp
similarity index 100%
rename from composable_kernel/example/09_convnd_fwd/convnd_fwd_common.hpp
rename to example/09_convnd_fwd/convnd_fwd_common.hpp
diff --git a/composable_kernel/example/09_convnd_fwd/convnd_fwd_dl_common.hpp b/example/09_convnd_fwd/convnd_fwd_dl_common.hpp
similarity index 100%
rename from composable_kernel/example/09_convnd_fwd/convnd_fwd_dl_common.hpp
rename to example/09_convnd_fwd/convnd_fwd_dl_common.hpp
diff --git a/composable_kernel/example/09_convnd_fwd/convnd_fwd_dl_fp16.cpp b/example/09_convnd_fwd/convnd_fwd_dl_fp16.cpp
similarity index 100%
rename from composable_kernel/example/09_convnd_fwd/convnd_fwd_dl_fp16.cpp
rename to example/09_convnd_fwd/convnd_fwd_dl_fp16.cpp
diff --git a/composable_kernel/example/09_convnd_fwd/convnd_fwd_dl_fp32.cpp b/example/09_convnd_fwd/convnd_fwd_dl_fp32.cpp
similarity index 100%
rename from composable_kernel/example/09_convnd_fwd/convnd_fwd_dl_fp32.cpp
rename to example/09_convnd_fwd/convnd_fwd_dl_fp32.cpp
diff --git a/composable_kernel/example/09_convnd_fwd/convnd_fwd_dl_int8.cpp b/example/09_convnd_fwd/convnd_fwd_dl_int8.cpp
similarity index 100%
rename from composable_kernel/example/09_convnd_fwd/convnd_fwd_dl_int8.cpp
rename to example/09_convnd_fwd/convnd_fwd_dl_int8.cpp
diff --git a/composable_kernel/example/09_convnd_fwd/convnd_fwd_xdl_bf16.cpp b/example/09_convnd_fwd/convnd_fwd_xdl_bf16.cpp
similarity index 100%
rename from composable_kernel/example/09_convnd_fwd/convnd_fwd_xdl_bf16.cpp
rename to example/09_convnd_fwd/convnd_fwd_xdl_bf16.cpp
diff --git a/composable_kernel/example/09_convnd_fwd/convnd_fwd_xdl_fp16.cpp b/example/09_convnd_fwd/convnd_fwd_xdl_fp16.cpp
similarity index 100%
rename from composable_kernel/example/09_convnd_fwd/convnd_fwd_xdl_fp16.cpp
rename to example/09_convnd_fwd/convnd_fwd_xdl_fp16.cpp
diff --git a/composable_kernel/example/09_convnd_fwd/convnd_fwd_xdl_fp32.cpp b/example/09_convnd_fwd/convnd_fwd_xdl_fp32.cpp
similarity index 100%
rename from composable_kernel/example/09_convnd_fwd/convnd_fwd_xdl_fp32.cpp
rename to example/09_convnd_fwd/convnd_fwd_xdl_fp32.cpp
diff --git a/composable_kernel/example/09_convnd_fwd/convnd_fwd_xdl_fp64.cpp b/example/09_convnd_fwd/convnd_fwd_xdl_fp64.cpp
similarity index 100%
rename from composable_kernel/example/09_convnd_fwd/convnd_fwd_xdl_fp64.cpp
rename to example/09_convnd_fwd/convnd_fwd_xdl_fp64.cpp
diff --git a/composable_kernel/example/09_convnd_fwd/convnd_fwd_xdl_int8.cpp b/example/09_convnd_fwd/convnd_fwd_xdl_int8.cpp
similarity index 100%
rename from composable_kernel/example/09_convnd_fwd/convnd_fwd_xdl_int8.cpp
rename to example/09_convnd_fwd/convnd_fwd_xdl_int8.cpp
diff --git a/composable_kernel/example/09_convnd_fwd/run_convnd_fwd_dl_example.inc b/example/09_convnd_fwd/run_convnd_fwd_dl_example.inc
similarity index 100%
rename from composable_kernel/example/09_convnd_fwd/run_convnd_fwd_dl_example.inc
rename to example/09_convnd_fwd/run_convnd_fwd_dl_example.inc
diff --git a/composable_kernel/example/09_convnd_fwd/run_convnd_fwd_example.inc b/example/09_convnd_fwd/run_convnd_fwd_example.inc
similarity index 100%
rename from composable_kernel/example/09_convnd_fwd/run_convnd_fwd_example.inc
rename to example/09_convnd_fwd/run_convnd_fwd_example.inc
diff --git a/composable_kernel/example/10_convnd_fwd_multiple_d_multiple_reduce/CMakeLists.txt b/example/10_convnd_fwd_multiple_d_multiple_reduce/CMakeLists.txt
similarity index 100%
rename from composable_kernel/example/10_convnd_fwd_multiple_d_multiple_reduce/CMakeLists.txt
rename to example/10_convnd_fwd_multiple_d_multiple_reduce/CMakeLists.txt
diff --git a/composable_kernel/example/10_convnd_fwd_multiple_d_multiple_reduce/common.hpp b/example/10_convnd_fwd_multiple_d_multiple_reduce/common.hpp
similarity index 100%
rename from composable_kernel/example/10_convnd_fwd_multiple_d_multiple_reduce/common.hpp
rename to example/10_convnd_fwd_multiple_d_multiple_reduce/common.hpp
diff --git a/composable_kernel/example/10_convnd_fwd_multiple_d_multiple_reduce/convnd_fwd_max_xdl_bf16.cpp b/example/10_convnd_fwd_multiple_d_multiple_reduce/convnd_fwd_max_xdl_bf16.cpp
similarity index 100%
rename from composable_kernel/example/10_convnd_fwd_multiple_d_multiple_reduce/convnd_fwd_max_xdl_bf16.cpp
rename to example/10_convnd_fwd_multiple_d_multiple_reduce/convnd_fwd_max_xdl_bf16.cpp
diff --git a/composable_kernel/example/10_convnd_fwd_multiple_d_multiple_reduce/convnd_fwd_max_xdl_fp16.cpp b/example/10_convnd_fwd_multiple_d_multiple_reduce/convnd_fwd_max_xdl_fp16.cpp
similarity index 100%
rename from composable_kernel/example/10_convnd_fwd_multiple_d_multiple_reduce/convnd_fwd_max_xdl_fp16.cpp
rename to example/10_convnd_fwd_multiple_d_multiple_reduce/convnd_fwd_max_xdl_fp16.cpp
diff --git a/composable_kernel/example/10_convnd_fwd_multiple_d_multiple_reduce/convnd_fwd_max_xdl_fp32.cpp b/example/10_convnd_fwd_multiple_d_multiple_reduce/convnd_fwd_max_xdl_fp32.cpp
similarity index 100%
rename from composable_kernel/example/10_convnd_fwd_multiple_d_multiple_reduce/convnd_fwd_max_xdl_fp32.cpp
rename to example/10_convnd_fwd_multiple_d_multiple_reduce/convnd_fwd_max_xdl_fp32.cpp
diff --git a/composable_kernel/example/10_convnd_fwd_multiple_d_multiple_reduce/convnd_fwd_max_xdl_int4.cpp b/example/10_convnd_fwd_multiple_d_multiple_reduce/convnd_fwd_max_xdl_int4.cpp
similarity index 100%
rename from composable_kernel/example/10_convnd_fwd_multiple_d_multiple_reduce/convnd_fwd_max_xdl_int4.cpp
rename to example/10_convnd_fwd_multiple_d_multiple_reduce/convnd_fwd_max_xdl_int4.cpp
diff --git a/composable_kernel/example/10_convnd_fwd_multiple_d_multiple_reduce/convnd_fwd_max_xdl_int8.cpp b/example/10_convnd_fwd_multiple_d_multiple_reduce/convnd_fwd_max_xdl_int8.cpp
similarity index 100%
rename from composable_kernel/example/10_convnd_fwd_multiple_d_multiple_reduce/convnd_fwd_max_xdl_int8.cpp
rename to example/10_convnd_fwd_multiple_d_multiple_reduce/convnd_fwd_max_xdl_int8.cpp
diff --git a/composable_kernel/example/10_convnd_fwd_multiple_d_multiple_reduce/run_convnd_fwd_max_example.inc b/example/10_convnd_fwd_multiple_d_multiple_reduce/run_convnd_fwd_max_example.inc
similarity index 100%
rename from composable_kernel/example/10_convnd_fwd_multiple_d_multiple_reduce/run_convnd_fwd_max_example.inc
rename to example/10_convnd_fwd_multiple_d_multiple_reduce/run_convnd_fwd_max_example.inc
diff --git a/composable_kernel/example/12_reduce/CMakeLists.txt b/example/12_reduce/CMakeLists.txt
similarity index 100%
rename from composable_kernel/example/12_reduce/CMakeLists.txt
rename to example/12_reduce/CMakeLists.txt
diff --git a/composable_kernel/example/12_reduce/README.md b/example/12_reduce/README.md
similarity index 100%
rename from composable_kernel/example/12_reduce/README.md
rename to example/12_reduce/README.md
diff --git a/composable_kernel/example/12_reduce/reduce_blockwise.cpp b/example/12_reduce/reduce_blockwise.cpp
similarity index 100%
rename from composable_kernel/example/12_reduce/reduce_blockwise.cpp
rename to example/12_reduce/reduce_blockwise.cpp
diff --git a/composable_kernel/example/12_reduce/reduce_blockwise_impl.hpp b/example/12_reduce/reduce_blockwise_impl.hpp
similarity index 100%
rename from composable_kernel/example/12_reduce/reduce_blockwise_impl.hpp
rename to example/12_reduce/reduce_blockwise_impl.hpp
diff --git a/composable_kernel/example/12_reduce/reduce_blockwise_two_call.cpp b/example/12_reduce/reduce_blockwise_two_call.cpp
similarity index 100%
rename from composable_kernel/example/12_reduce/reduce_blockwise_two_call.cpp
rename to example/12_reduce/reduce_blockwise_two_call.cpp
diff --git a/composable_kernel/example/12_reduce/reduce_example_common.hpp b/example/12_reduce/reduce_example_common.hpp
similarity index 100%
rename from composable_kernel/example/12_reduce/reduce_example_common.hpp
rename to example/12_reduce/reduce_example_common.hpp
diff --git a/composable_kernel/example/12_reduce/reduce_multiblock_atomic_add.cpp b/example/12_reduce/reduce_multiblock_atomic_add.cpp
similarity index 100%
rename from composable_kernel/example/12_reduce/reduce_multiblock_atomic_add.cpp
rename to example/12_reduce/reduce_multiblock_atomic_add.cpp
diff --git a/composable_kernel/example/12_reduce/reduce_multiblock_atomic_add_impl.hpp b/example/12_reduce/reduce_multiblock_atomic_add_impl.hpp
similarity index 100%
rename from composable_kernel/example/12_reduce/reduce_multiblock_atomic_add_impl.hpp
rename to example/12_reduce/reduce_multiblock_atomic_add_impl.hpp
diff --git a/composable_kernel/example/13_pool2d_fwd/CMakeLists.txt b/example/13_pool2d_fwd/CMakeLists.txt
similarity index 100%
rename from composable_kernel/example/13_pool2d_fwd/CMakeLists.txt
rename to example/13_pool2d_fwd/CMakeLists.txt
diff --git a/composable_kernel/example/13_pool2d_fwd/README.md b/example/13_pool2d_fwd/README.md
similarity index 100%
rename from composable_kernel/example/13_pool2d_fwd/README.md
rename to example/13_pool2d_fwd/README.md
diff --git a/composable_kernel/example/13_pool2d_fwd/pool2d_fwd_common.hpp b/example/13_pool2d_fwd/pool2d_fwd_common.hpp
similarity index 100%
rename from composable_kernel/example/13_pool2d_fwd/pool2d_fwd_common.hpp
rename to example/13_pool2d_fwd/pool2d_fwd_common.hpp
diff --git a/composable_kernel/example/13_pool2d_fwd/pool2d_fwd_fp16.cpp b/example/13_pool2d_fwd/pool2d_fwd_fp16.cpp
similarity index 100%
rename from composable_kernel/example/13_pool2d_fwd/pool2d_fwd_fp16.cpp
rename to example/13_pool2d_fwd/pool2d_fwd_fp16.cpp
diff --git a/composable_kernel/example/13_pool2d_fwd/pool2d_fwd_fp32.cpp b/example/13_pool2d_fwd/pool2d_fwd_fp32.cpp
similarity index 100%
rename from composable_kernel/example/13_pool2d_fwd/pool2d_fwd_fp32.cpp
rename to example/13_pool2d_fwd/pool2d_fwd_fp32.cpp
diff --git a/composable_kernel/example/14_gemm_quantization/CMakeLists.txt b/example/14_gemm_quantization/CMakeLists.txt
similarity index 100%
rename from composable_kernel/example/14_gemm_quantization/CMakeLists.txt
rename to example/14_gemm_quantization/CMakeLists.txt
diff --git a/composable_kernel/example/14_gemm_quantization/gemm_dl_quantization_int8.cpp b/example/14_gemm_quantization/gemm_dl_quantization_int8.cpp
similarity index 100%
rename from composable_kernel/example/14_gemm_quantization/gemm_dl_quantization_int8.cpp
rename to example/14_gemm_quantization/gemm_dl_quantization_int8.cpp
diff --git a/composable_kernel/example/14_gemm_quantization/gemm_xdl_bias_relu_quantization_int8.cpp b/example/14_gemm_quantization/gemm_xdl_bias_relu_quantization_int8.cpp
similarity index 100%
rename from composable_kernel/example/14_gemm_quantization/gemm_xdl_bias_relu_quantization_int8.cpp
rename to example/14_gemm_quantization/gemm_xdl_bias_relu_quantization_int8.cpp
diff --git a/composable_kernel/example/14_gemm_quantization/gemm_xdl_quantization_int8.cpp b/example/14_gemm_quantization/gemm_xdl_quantization_int8.cpp
similarity index 100%
rename from composable_kernel/example/14_gemm_quantization/gemm_xdl_quantization_int8.cpp
rename to example/14_gemm_quantization/gemm_xdl_quantization_int8.cpp
diff --git a/composable_kernel/example/15_grouped_gemm/CMakeLists.txt b/example/15_grouped_gemm/CMakeLists.txt
similarity index 100%
rename from composable_kernel/example/15_grouped_gemm/CMakeLists.txt
rename to example/15_grouped_gemm/CMakeLists.txt
diff --git a/composable_kernel/example/15_grouped_gemm/README.md b/example/15_grouped_gemm/README.md
similarity index 100%
rename from composable_kernel/example/15_grouped_gemm/README.md
rename to example/15_grouped_gemm/README.md
diff --git a/composable_kernel/example/15_grouped_gemm/grouped_gemm_multiple_d_dl_fp16.cpp b/example/15_grouped_gemm/grouped_gemm_multiple_d_dl_fp16.cpp
similarity index 100%
rename from composable_kernel/example/15_grouped_gemm/grouped_gemm_multiple_d_dl_fp16.cpp
rename to example/15_grouped_gemm/grouped_gemm_multiple_d_dl_fp16.cpp
diff --git a/composable_kernel/example/15_grouped_gemm/grouped_gemm_xdl_bf16.cpp b/example/15_grouped_gemm/grouped_gemm_xdl_bf16.cpp
similarity index 100%
rename from composable_kernel/example/15_grouped_gemm/grouped_gemm_xdl_bf16.cpp
rename to example/15_grouped_gemm/grouped_gemm_xdl_bf16.cpp
diff --git a/composable_kernel/example/15_grouped_gemm/grouped_gemm_xdl_fixed_nk_bias_fp16.cpp b/example/15_grouped_gemm/grouped_gemm_xdl_fixed_nk_bias_fp16.cpp
similarity index 100%
rename from composable_kernel/example/15_grouped_gemm/grouped_gemm_xdl_fixed_nk_bias_fp16.cpp
rename to example/15_grouped_gemm/grouped_gemm_xdl_fixed_nk_bias_fp16.cpp
diff --git a/composable_kernel/example/15_grouped_gemm/grouped_gemm_xdl_fixed_nk_fp16.cpp b/example/15_grouped_gemm/grouped_gemm_xdl_fixed_nk_fp16.cpp
similarity index 100%
rename from composable_kernel/example/15_grouped_gemm/grouped_gemm_xdl_fixed_nk_fp16.cpp
rename to example/15_grouped_gemm/grouped_gemm_xdl_fixed_nk_fp16.cpp
diff --git a/composable_kernel/example/15_grouped_gemm/grouped_gemm_xdl_fixed_nk_fp8.cpp b/example/15_grouped_gemm/grouped_gemm_xdl_fixed_nk_fp8.cpp
similarity index 100%
rename from composable_kernel/example/15_grouped_gemm/grouped_gemm_xdl_fixed_nk_fp8.cpp
rename to example/15_grouped_gemm/grouped_gemm_xdl_fixed_nk_fp8.cpp
diff --git a/composable_kernel/example/15_grouped_gemm/grouped_gemm_xdl_fp16.cpp b/example/15_grouped_gemm/grouped_gemm_xdl_fp16.cpp
similarity index 100%
rename from composable_kernel/example/15_grouped_gemm/grouped_gemm_xdl_fp16.cpp
rename to example/15_grouped_gemm/grouped_gemm_xdl_fp16.cpp
diff --git a/composable_kernel/example/15_grouped_gemm/grouped_gemm_xdl_fp32.cpp b/example/15_grouped_gemm/grouped_gemm_xdl_fp32.cpp
similarity index 100%
rename from composable_kernel/example/15_grouped_gemm/grouped_gemm_xdl_fp32.cpp
rename to example/15_grouped_gemm/grouped_gemm_xdl_fp32.cpp
diff --git a/composable_kernel/example/15_grouped_gemm/grouped_gemm_xdl_int4.cpp b/example/15_grouped_gemm/grouped_gemm_xdl_int4.cpp
similarity index 100%
rename from composable_kernel/example/15_grouped_gemm/grouped_gemm_xdl_int4.cpp
rename to example/15_grouped_gemm/grouped_gemm_xdl_int4.cpp
diff --git a/composable_kernel/example/15_grouped_gemm/grouped_gemm_xdl_int8.cpp b/example/15_grouped_gemm/grouped_gemm_xdl_int8.cpp
similarity index 100%
rename from composable_kernel/example/15_grouped_gemm/grouped_gemm_xdl_int8.cpp
rename to example/15_grouped_gemm/grouped_gemm_xdl_int8.cpp
diff --git a/composable_kernel/example/15_grouped_gemm/grouped_gemm_xdl_splitk_fp16.cpp b/example/15_grouped_gemm/grouped_gemm_xdl_splitk_fp16.cpp
similarity index 100%
rename from composable_kernel/example/15_grouped_gemm/grouped_gemm_xdl_splitk_fp16.cpp
rename to example/15_grouped_gemm/grouped_gemm_xdl_splitk_fp16.cpp
diff --git a/composable_kernel/example/15_grouped_gemm/run_grouped_gemm_example.inc b/example/15_grouped_gemm/run_grouped_gemm_example.inc
similarity index 100%
rename from composable_kernel/example/15_grouped_gemm/run_grouped_gemm_example.inc
rename to example/15_grouped_gemm/run_grouped_gemm_example.inc
diff --git a/composable_kernel/example/16_gemm_multi_d_multi_reduces/CMakeLists.txt b/example/16_gemm_multi_d_multi_reduces/CMakeLists.txt
similarity index 100%
rename from composable_kernel/example/16_gemm_multi_d_multi_reduces/CMakeLists.txt
rename to example/16_gemm_multi_d_multi_reduces/CMakeLists.txt
diff --git a/composable_kernel/example/16_gemm_multi_d_multi_reduces/gemm_add_add_mean_meansquare_xdl_fp16.cpp b/example/16_gemm_multi_d_multi_reduces/gemm_add_add_mean_meansquare_xdl_fp16.cpp
similarity index 100%
rename from composable_kernel/example/16_gemm_multi_d_multi_reduces/gemm_add_add_mean_meansquare_xdl_fp16.cpp
rename to example/16_gemm_multi_d_multi_reduces/gemm_add_add_mean_meansquare_xdl_fp16.cpp
diff --git a/composable_kernel/example/16_gemm_multi_d_multi_reduces/gemm_add_addsquare_xdl_int8.cpp b/example/16_gemm_multi_d_multi_reduces/gemm_add_addsquare_xdl_int8.cpp
similarity index 100%
rename from composable_kernel/example/16_gemm_multi_d_multi_reduces/gemm_add_addsquare_xdl_int8.cpp
rename to example/16_gemm_multi_d_multi_reduces/gemm_add_addsquare_xdl_int8.cpp
diff --git a/composable_kernel/example/16_gemm_multi_d_multi_reduces/gemm_max_xdl_bf16.cpp b/example/16_gemm_multi_d_multi_reduces/gemm_max_xdl_bf16.cpp
similarity index 100%
rename from composable_kernel/example/16_gemm_multi_d_multi_reduces/gemm_max_xdl_bf16.cpp
rename to example/16_gemm_multi_d_multi_reduces/gemm_max_xdl_bf16.cpp
diff --git a/composable_kernel/example/16_gemm_multi_d_multi_reduces/gemm_max_xdl_fp16.cpp b/example/16_gemm_multi_d_multi_reduces/gemm_max_xdl_fp16.cpp
similarity index 100%
rename from composable_kernel/example/16_gemm_multi_d_multi_reduces/gemm_max_xdl_fp16.cpp
rename to example/16_gemm_multi_d_multi_reduces/gemm_max_xdl_fp16.cpp
diff --git a/composable_kernel/example/16_gemm_multi_d_multi_reduces/gemm_max_xdl_fp32.cpp b/example/16_gemm_multi_d_multi_reduces/gemm_max_xdl_fp32.cpp
similarity index 100%
rename from composable_kernel/example/16_gemm_multi_d_multi_reduces/gemm_max_xdl_fp32.cpp
rename to example/16_gemm_multi_d_multi_reduces/gemm_max_xdl_fp32.cpp
diff --git a/composable_kernel/example/16_gemm_multi_d_multi_reduces/gemm_max_xdl_int4.cpp b/example/16_gemm_multi_d_multi_reduces/gemm_max_xdl_int4.cpp
similarity index 100%
rename from composable_kernel/example/16_gemm_multi_d_multi_reduces/gemm_max_xdl_int4.cpp
rename to example/16_gemm_multi_d_multi_reduces/gemm_max_xdl_int4.cpp
diff --git a/composable_kernel/example/16_gemm_multi_d_multi_reduces/gemm_max_xdl_int8.cpp b/example/16_gemm_multi_d_multi_reduces/gemm_max_xdl_int8.cpp
similarity index 100%
rename from composable_kernel/example/16_gemm_multi_d_multi_reduces/gemm_max_xdl_int8.cpp
rename to example/16_gemm_multi_d_multi_reduces/gemm_max_xdl_int8.cpp
diff --git a/composable_kernel/example/16_gemm_multi_d_multi_reduces/gemm_mean_meansquare_xdl_bf16.cpp b/example/16_gemm_multi_d_multi_reduces/gemm_mean_meansquare_xdl_bf16.cpp
similarity index 100%
rename from composable_kernel/example/16_gemm_multi_d_multi_reduces/gemm_mean_meansquare_xdl_bf16.cpp
rename to example/16_gemm_multi_d_multi_reduces/gemm_mean_meansquare_xdl_bf16.cpp
diff --git a/composable_kernel/example/16_gemm_multi_d_multi_reduces/gemm_mean_meansquare_xdl_fp16.cpp b/example/16_gemm_multi_d_multi_reduces/gemm_mean_meansquare_xdl_fp16.cpp
similarity index 100%
rename from composable_kernel/example/16_gemm_multi_d_multi_reduces/gemm_mean_meansquare_xdl_fp16.cpp
rename to example/16_gemm_multi_d_multi_reduces/gemm_mean_meansquare_xdl_fp16.cpp
diff --git a/composable_kernel/example/16_gemm_multi_d_multi_reduces/gemm_mean_meansquare_xdl_fp32.cpp b/example/16_gemm_multi_d_multi_reduces/gemm_mean_meansquare_xdl_fp32.cpp
similarity index 100%
rename from composable_kernel/example/16_gemm_multi_d_multi_reduces/gemm_mean_meansquare_xdl_fp32.cpp
rename to example/16_gemm_multi_d_multi_reduces/gemm_mean_meansquare_xdl_fp32.cpp
diff --git a/composable_kernel/example/16_gemm_multi_d_multi_reduces/gemm_reduce_xdl_common.hpp b/example/16_gemm_multi_d_multi_reduces/gemm_reduce_xdl_common.hpp
similarity index 100%
rename from composable_kernel/example/16_gemm_multi_d_multi_reduces/gemm_reduce_xdl_common.hpp
rename to example/16_gemm_multi_d_multi_reduces/gemm_reduce_xdl_common.hpp
diff --git a/composable_kernel/example/17_convnd_bwd_data/CMakeLists.txt b/example/17_convnd_bwd_data/CMakeLists.txt
similarity index 100%
rename from composable_kernel/example/17_convnd_bwd_data/CMakeLists.txt
rename to example/17_convnd_bwd_data/CMakeLists.txt
diff --git a/composable_kernel/example/17_convnd_bwd_data/README.md b/example/17_convnd_bwd_data/README.md
similarity index 100%
rename from composable_kernel/example/17_convnd_bwd_data/README.md
rename to example/17_convnd_bwd_data/README.md
diff --git a/composable_kernel/example/17_convnd_bwd_data/convnd_bwd_data_common.hpp b/example/17_convnd_bwd_data/convnd_bwd_data_common.hpp
similarity index 100%
rename from composable_kernel/example/17_convnd_bwd_data/convnd_bwd_data_common.hpp
rename to example/17_convnd_bwd_data/convnd_bwd_data_common.hpp
diff --git a/composable_kernel/example/17_convnd_bwd_data/convnd_bwd_data_dl_fp16.cpp b/example/17_convnd_bwd_data/convnd_bwd_data_dl_fp16.cpp
similarity index 100%
rename from composable_kernel/example/17_convnd_bwd_data/convnd_bwd_data_dl_fp16.cpp
rename to example/17_convnd_bwd_data/convnd_bwd_data_dl_fp16.cpp
diff --git a/composable_kernel/example/17_convnd_bwd_data/convnd_bwd_data_xdl_fp16.cpp b/example/17_convnd_bwd_data/convnd_bwd_data_xdl_fp16.cpp
similarity index 100%
rename from composable_kernel/example/17_convnd_bwd_data/convnd_bwd_data_xdl_fp16.cpp
rename to example/17_convnd_bwd_data/convnd_bwd_data_xdl_fp16.cpp
diff --git a/composable_kernel/example/18_batched_gemm_reduce/CMakeLists.txt b/example/18_batched_gemm_reduce/CMakeLists.txt
similarity index 100%
rename from composable_kernel/example/18_batched_gemm_reduce/CMakeLists.txt
rename to example/18_batched_gemm_reduce/CMakeLists.txt
diff --git a/composable_kernel/example/18_batched_gemm_reduce/batched_gemm_reduce_xdl_fp16.cpp b/example/18_batched_gemm_reduce/batched_gemm_reduce_xdl_fp16.cpp
similarity index 100%
rename from composable_kernel/example/18_batched_gemm_reduce/batched_gemm_reduce_xdl_fp16.cpp
rename to example/18_batched_gemm_reduce/batched_gemm_reduce_xdl_fp16.cpp
diff --git a/composable_kernel/example/19_binary_elementwise/CMakeLists.txt b/example/19_binary_elementwise/CMakeLists.txt
similarity index 100%
rename from composable_kernel/example/19_binary_elementwise/CMakeLists.txt
rename to example/19_binary_elementwise/CMakeLists.txt
diff --git a/composable_kernel/example/19_binary_elementwise/broadcast_add_2d_amn_bn.cpp b/example/19_binary_elementwise/broadcast_add_2d_amn_bn.cpp
similarity index 100%
rename from composable_kernel/example/19_binary_elementwise/broadcast_add_2d_amn_bn.cpp
rename to example/19_binary_elementwise/broadcast_add_2d_amn_bn.cpp
diff --git a/composable_kernel/example/19_binary_elementwise/broadcast_add_3d_am_bmnk.cpp b/example/19_binary_elementwise/broadcast_add_3d_am_bmnk.cpp
similarity index 100%
rename from composable_kernel/example/19_binary_elementwise/broadcast_add_3d_am_bmnk.cpp
rename to example/19_binary_elementwise/broadcast_add_3d_am_bmnk.cpp
diff --git a/composable_kernel/example/19_binary_elementwise/elementwise_add_1d.cpp b/example/19_binary_elementwise/elementwise_add_1d.cpp
similarity index 100%
rename from composable_kernel/example/19_binary_elementwise/elementwise_add_1d.cpp
rename to example/19_binary_elementwise/elementwise_add_1d.cpp
diff --git a/composable_kernel/example/19_binary_elementwise/elementwise_add_4d.cpp b/example/19_binary_elementwise/elementwise_add_4d.cpp
similarity index 100%
rename from composable_kernel/example/19_binary_elementwise/elementwise_add_4d.cpp
rename to example/19_binary_elementwise/elementwise_add_4d.cpp
diff --git a/composable_kernel/example/20_grouped_conv_bwd_weight/CMakeLists.txt b/example/20_grouped_conv_bwd_weight/CMakeLists.txt
similarity index 100%
rename from composable_kernel/example/20_grouped_conv_bwd_weight/CMakeLists.txt
rename to example/20_grouped_conv_bwd_weight/CMakeLists.txt
diff --git a/composable_kernel/example/20_grouped_conv_bwd_weight/common.hpp b/example/20_grouped_conv_bwd_weight/common.hpp
similarity index 100%
rename from composable_kernel/example/20_grouped_conv_bwd_weight/common.hpp
rename to example/20_grouped_conv_bwd_weight/common.hpp
diff --git a/composable_kernel/example/20_grouped_conv_bwd_weight/grouped_conv_bwd_weight_dl_fp16.cpp b/example/20_grouped_conv_bwd_weight/grouped_conv_bwd_weight_dl_fp16.cpp
similarity index 100%
rename from composable_kernel/example/20_grouped_conv_bwd_weight/grouped_conv_bwd_weight_dl_fp16.cpp
rename to example/20_grouped_conv_bwd_weight/grouped_conv_bwd_weight_dl_fp16.cpp
diff --git a/composable_kernel/example/20_grouped_conv_bwd_weight/grouped_conv_bwd_weight_wmma_fp16.cpp b/example/20_grouped_conv_bwd_weight/grouped_conv_bwd_weight_wmma_fp16.cpp
similarity index 100%
rename from composable_kernel/example/20_grouped_conv_bwd_weight/grouped_conv_bwd_weight_wmma_fp16.cpp
rename to example/20_grouped_conv_bwd_weight/grouped_conv_bwd_weight_wmma_fp16.cpp
diff --git a/composable_kernel/example/20_grouped_conv_bwd_weight/grouped_conv_bwd_weight_xdl_bf16.cpp b/example/20_grouped_conv_bwd_weight/grouped_conv_bwd_weight_xdl_bf16.cpp
similarity index 100%
rename from composable_kernel/example/20_grouped_conv_bwd_weight/grouped_conv_bwd_weight_xdl_bf16.cpp
rename to example/20_grouped_conv_bwd_weight/grouped_conv_bwd_weight_xdl_bf16.cpp
diff --git a/composable_kernel/example/20_grouped_conv_bwd_weight/grouped_conv_bwd_weight_xdl_fp16.cpp b/example/20_grouped_conv_bwd_weight/grouped_conv_bwd_weight_xdl_fp16.cpp
similarity index 100%
rename from composable_kernel/example/20_grouped_conv_bwd_weight/grouped_conv_bwd_weight_xdl_fp16.cpp
rename to example/20_grouped_conv_bwd_weight/grouped_conv_bwd_weight_xdl_fp16.cpp
diff --git a/composable_kernel/example/20_grouped_conv_bwd_weight/grouped_conv_bwd_weight_xdl_fp16_comp_bf8_fp8.cpp b/example/20_grouped_conv_bwd_weight/grouped_conv_bwd_weight_xdl_fp16_comp_bf8_fp8.cpp
similarity index 100%
rename from composable_kernel/example/20_grouped_conv_bwd_weight/grouped_conv_bwd_weight_xdl_fp16_comp_bf8_fp8.cpp
rename to example/20_grouped_conv_bwd_weight/grouped_conv_bwd_weight_xdl_fp16_comp_bf8_fp8.cpp
diff --git a/composable_kernel/example/20_grouped_conv_bwd_weight/run_grouped_conv_bwd_weight_example.inc b/example/20_grouped_conv_bwd_weight/run_grouped_conv_bwd_weight_example.inc
similarity index 100%
rename from composable_kernel/example/20_grouped_conv_bwd_weight/run_grouped_conv_bwd_weight_example.inc
rename to example/20_grouped_conv_bwd_weight/run_grouped_conv_bwd_weight_example.inc
diff --git a/composable_kernel/example/21_gemm_layernorm/CMakeLists.txt b/example/21_gemm_layernorm/CMakeLists.txt
similarity index 100%
rename from composable_kernel/example/21_gemm_layernorm/CMakeLists.txt
rename to example/21_gemm_layernorm/CMakeLists.txt
diff --git a/composable_kernel/example/21_gemm_layernorm/gemm_bias_relu_add_layernorm_xdl_naive_fp16.cpp b/example/21_gemm_layernorm/gemm_bias_relu_add_layernorm_xdl_naive_fp16.cpp
similarity index 100%
rename from composable_kernel/example/21_gemm_layernorm/gemm_bias_relu_add_layernorm_xdl_naive_fp16.cpp
rename to example/21_gemm_layernorm/gemm_bias_relu_add_layernorm_xdl_naive_fp16.cpp
diff --git a/composable_kernel/example/21_gemm_layernorm/gemm_bias_relu_add_layernorm_xdl_welford_fp16.cpp b/example/21_gemm_layernorm/gemm_bias_relu_add_layernorm_xdl_welford_fp16.cpp
similarity index 100%
rename from composable_kernel/example/21_gemm_layernorm/gemm_bias_relu_add_layernorm_xdl_welford_fp16.cpp
rename to example/21_gemm_layernorm/gemm_bias_relu_add_layernorm_xdl_welford_fp16.cpp
diff --git a/composable_kernel/example/21_gemm_layernorm/gemm_layernorm_xdl_naive_fp16.cpp b/example/21_gemm_layernorm/gemm_layernorm_xdl_naive_fp16.cpp
similarity index 100%
rename from composable_kernel/example/21_gemm_layernorm/gemm_layernorm_xdl_naive_fp16.cpp
rename to example/21_gemm_layernorm/gemm_layernorm_xdl_naive_fp16.cpp
diff --git a/composable_kernel/example/21_gemm_layernorm/gemm_xdl_layernorm_naive_single_kernel_fp16.cpp b/example/21_gemm_layernorm/gemm_xdl_layernorm_naive_single_kernel_fp16.cpp
similarity index 100%
rename from composable_kernel/example/21_gemm_layernorm/gemm_xdl_layernorm_naive_single_kernel_fp16.cpp
rename to example/21_gemm_layernorm/gemm_xdl_layernorm_naive_single_kernel_fp16.cpp
diff --git a/composable_kernel/example/22_cgemm/CMakeLists.txt b/example/22_cgemm/CMakeLists.txt
similarity index 100%
rename from composable_kernel/example/22_cgemm/CMakeLists.txt
rename to example/22_cgemm/CMakeLists.txt
diff --git a/composable_kernel/example/22_cgemm/cgemm_xdl_bf16.cpp b/example/22_cgemm/cgemm_xdl_bf16.cpp
similarity index 100%
rename from composable_kernel/example/22_cgemm/cgemm_xdl_bf16.cpp
rename to example/22_cgemm/cgemm_xdl_bf16.cpp
diff --git a/composable_kernel/example/22_cgemm/cgemm_xdl_common.hpp b/example/22_cgemm/cgemm_xdl_common.hpp
similarity index 100%
rename from composable_kernel/example/22_cgemm/cgemm_xdl_common.hpp
rename to example/22_cgemm/cgemm_xdl_common.hpp
diff --git a/composable_kernel/example/22_cgemm/cgemm_xdl_fp16.cpp b/example/22_cgemm/cgemm_xdl_fp16.cpp
similarity index 100%
rename from composable_kernel/example/22_cgemm/cgemm_xdl_fp16.cpp
rename to example/22_cgemm/cgemm_xdl_fp16.cpp
diff --git a/composable_kernel/example/22_cgemm/cgemm_xdl_fp32.cpp b/example/22_cgemm/cgemm_xdl_fp32.cpp
similarity index 100%
rename from composable_kernel/example/22_cgemm/cgemm_xdl_fp32.cpp
rename to example/22_cgemm/cgemm_xdl_fp32.cpp
diff --git a/composable_kernel/example/22_cgemm/cgemm_xdl_int4.cpp b/example/22_cgemm/cgemm_xdl_int4.cpp
similarity index 100%
rename from composable_kernel/example/22_cgemm/cgemm_xdl_int4.cpp
rename to example/22_cgemm/cgemm_xdl_int4.cpp
diff --git a/composable_kernel/example/22_cgemm/cgemm_xdl_int8.cpp b/example/22_cgemm/cgemm_xdl_int8.cpp
similarity index 100%
rename from composable_kernel/example/22_cgemm/cgemm_xdl_int8.cpp
rename to example/22_cgemm/cgemm_xdl_int8.cpp
diff --git a/composable_kernel/example/23_softmax/CMakeLists.txt b/example/23_softmax/CMakeLists.txt
similarity index 100%
rename from composable_kernel/example/23_softmax/CMakeLists.txt
rename to example/23_softmax/CMakeLists.txt
diff --git a/composable_kernel/example/23_softmax/README.md b/example/23_softmax/README.md
similarity index 100%
rename from composable_kernel/example/23_softmax/README.md
rename to example/23_softmax/README.md
diff --git a/composable_kernel/example/23_softmax/softmax_blockwise.cpp b/example/23_softmax/softmax_blockwise.cpp
similarity index 100%
rename from composable_kernel/example/23_softmax/softmax_blockwise.cpp
rename to example/23_softmax/softmax_blockwise.cpp
diff --git a/composable_kernel/example/24_batched_gemm/CMakeLists.txt b/example/24_batched_gemm/CMakeLists.txt
similarity index 100%
rename from composable_kernel/example/24_batched_gemm/CMakeLists.txt
rename to example/24_batched_gemm/CMakeLists.txt
diff --git a/composable_kernel/example/24_batched_gemm/batched_gemm_xdl_bf16.cpp b/example/24_batched_gemm/batched_gemm_xdl_bf16.cpp
similarity index 100%
rename from composable_kernel/example/24_batched_gemm/batched_gemm_xdl_bf16.cpp
rename to example/24_batched_gemm/batched_gemm_xdl_bf16.cpp
diff --git a/composable_kernel/example/24_batched_gemm/batched_gemm_xdl_fp16.cpp b/example/24_batched_gemm/batched_gemm_xdl_fp16.cpp
similarity index 100%
rename from composable_kernel/example/24_batched_gemm/batched_gemm_xdl_fp16.cpp
rename to example/24_batched_gemm/batched_gemm_xdl_fp16.cpp
diff --git a/composable_kernel/example/24_batched_gemm/batched_gemm_xdl_fp32.cpp b/example/24_batched_gemm/batched_gemm_xdl_fp32.cpp
similarity index 100%
rename from composable_kernel/example/24_batched_gemm/batched_gemm_xdl_fp32.cpp
rename to example/24_batched_gemm/batched_gemm_xdl_fp32.cpp
diff --git a/composable_kernel/example/24_batched_gemm/batched_gemm_xdl_int4.cpp b/example/24_batched_gemm/batched_gemm_xdl_int4.cpp
similarity index 100%
rename from composable_kernel/example/24_batched_gemm/batched_gemm_xdl_int4.cpp
rename to example/24_batched_gemm/batched_gemm_xdl_int4.cpp
diff --git a/composable_kernel/example/24_batched_gemm/batched_gemm_xdl_int8.cpp b/example/24_batched_gemm/batched_gemm_xdl_int8.cpp
similarity index 100%
rename from composable_kernel/example/24_batched_gemm/batched_gemm_xdl_int8.cpp
rename to example/24_batched_gemm/batched_gemm_xdl_int8.cpp
diff --git a/composable_kernel/example/24_batched_gemm/run_batched_gemm_example.inc b/example/24_batched_gemm/run_batched_gemm_example.inc
similarity index 100%
rename from composable_kernel/example/24_batched_gemm/run_batched_gemm_example.inc
rename to example/24_batched_gemm/run_batched_gemm_example.inc
diff --git a/composable_kernel/example/25_gemm_bias_e_permute/CMakeLists.txt b/example/25_gemm_bias_e_permute/CMakeLists.txt
similarity index 100%
rename from composable_kernel/example/25_gemm_bias_e_permute/CMakeLists.txt
rename to example/25_gemm_bias_e_permute/CMakeLists.txt
diff --git a/composable_kernel/example/25_gemm_bias_e_permute/gemm_bias_e_permute_g1m2n3k1_xdl_fp16.cpp b/example/25_gemm_bias_e_permute/gemm_bias_e_permute_g1m2n3k1_xdl_fp16.cpp
similarity index 100%
rename from composable_kernel/example/25_gemm_bias_e_permute/gemm_bias_e_permute_g1m2n3k1_xdl_fp16.cpp
rename to example/25_gemm_bias_e_permute/gemm_bias_e_permute_g1m2n3k1_xdl_fp16.cpp
diff --git a/composable_kernel/example/25_gemm_bias_e_permute/gemm_bias_e_permute_g1m3n2k1_xdl_fp16.cpp b/example/25_gemm_bias_e_permute/gemm_bias_e_permute_g1m3n2k1_xdl_fp16.cpp
similarity index 100%
rename from composable_kernel/example/25_gemm_bias_e_permute/gemm_bias_e_permute_g1m3n2k1_xdl_fp16.cpp
rename to example/25_gemm_bias_e_permute/gemm_bias_e_permute_g1m3n2k1_xdl_fp16.cpp
diff --git a/composable_kernel/example/26_contraction/CMakeLists.txt b/example/26_contraction/CMakeLists.txt
similarity index 100%
rename from composable_kernel/example/26_contraction/CMakeLists.txt
rename to example/26_contraction/CMakeLists.txt
diff --git a/composable_kernel/example/26_contraction/README.md b/example/26_contraction/README.md
similarity index 100%
rename from composable_kernel/example/26_contraction/README.md
rename to example/26_contraction/README.md
diff --git a/composable_kernel/example/26_contraction/contraction_bilinear_xdl_fp32.cpp b/example/26_contraction/contraction_bilinear_xdl_fp32.cpp
similarity index 100%
rename from composable_kernel/example/26_contraction/contraction_bilinear_xdl_fp32.cpp
rename to example/26_contraction/contraction_bilinear_xdl_fp32.cpp
diff --git a/composable_kernel/example/26_contraction/contraction_bilinear_xdl_fp64.cpp b/example/26_contraction/contraction_bilinear_xdl_fp64.cpp
similarity index 100%
rename from composable_kernel/example/26_contraction/contraction_bilinear_xdl_fp64.cpp
rename to example/26_contraction/contraction_bilinear_xdl_fp64.cpp
diff --git a/composable_kernel/example/26_contraction/contraction_scale_xdl_fp32.cpp b/example/26_contraction/contraction_scale_xdl_fp32.cpp
similarity index 100%
rename from composable_kernel/example/26_contraction/contraction_scale_xdl_fp32.cpp
rename to example/26_contraction/contraction_scale_xdl_fp32.cpp
diff --git a/composable_kernel/example/26_contraction/contraction_scale_xdl_fp64.cpp b/example/26_contraction/contraction_scale_xdl_fp64.cpp
similarity index 100%
rename from composable_kernel/example/26_contraction/contraction_scale_xdl_fp64.cpp
rename to example/26_contraction/contraction_scale_xdl_fp64.cpp
diff --git a/composable_kernel/example/27_layernorm/CMakeLists.txt b/example/27_layernorm/CMakeLists.txt
similarity index 100%
rename from composable_kernel/example/27_layernorm/CMakeLists.txt
rename to example/27_layernorm/CMakeLists.txt
diff --git a/composable_kernel/example/27_layernorm/common.hpp b/example/27_layernorm/common.hpp
similarity index 100%
rename from composable_kernel/example/27_layernorm/common.hpp
rename to example/27_layernorm/common.hpp
diff --git a/composable_kernel/example/27_layernorm/layernorm_fp16.cpp b/example/27_layernorm/layernorm_fp16.cpp
similarity index 100%
rename from composable_kernel/example/27_layernorm/layernorm_fp16.cpp
rename to example/27_layernorm/layernorm_fp16.cpp
diff --git a/composable_kernel/example/27_layernorm/layernorm_splitk_fp16.cpp b/example/27_layernorm/layernorm_splitk_fp16.cpp
similarity index 100%
rename from composable_kernel/example/27_layernorm/layernorm_splitk_fp16.cpp
rename to example/27_layernorm/layernorm_splitk_fp16.cpp
diff --git a/composable_kernel/example/27_layernorm/run_layernorm_example.inc b/example/27_layernorm/run_layernorm_example.inc
similarity index 100%
rename from composable_kernel/example/27_layernorm/run_layernorm_example.inc
rename to example/27_layernorm/run_layernorm_example.inc
diff --git a/composable_kernel/example/28_grouped_gemm_bias_e_permute/CMakeLists.txt b/example/28_grouped_gemm_bias_e_permute/CMakeLists.txt
similarity index 100%
rename from composable_kernel/example/28_grouped_gemm_bias_e_permute/CMakeLists.txt
rename to example/28_grouped_gemm_bias_e_permute/CMakeLists.txt
diff --git a/composable_kernel/example/28_grouped_gemm_bias_e_permute/grouped_gemm_bias_e_permute_xdl_fp16.cpp b/example/28_grouped_gemm_bias_e_permute/grouped_gemm_bias_e_permute_xdl_fp16.cpp
similarity index 100%
rename from composable_kernel/example/28_grouped_gemm_bias_e_permute/grouped_gemm_bias_e_permute_xdl_fp16.cpp
rename to example/28_grouped_gemm_bias_e_permute/grouped_gemm_bias_e_permute_xdl_fp16.cpp
diff --git a/composable_kernel/example/29_batched_gemm_bias_e_permute/CMakeLists.txt b/example/29_batched_gemm_bias_e_permute/CMakeLists.txt
similarity index 100%
rename from composable_kernel/example/29_batched_gemm_bias_e_permute/CMakeLists.txt
rename to example/29_batched_gemm_bias_e_permute/CMakeLists.txt
diff --git a/composable_kernel/example/29_batched_gemm_bias_e_permute/batched_gemm_bias_e_permute_wmma_fp16.cpp b/example/29_batched_gemm_bias_e_permute/batched_gemm_bias_e_permute_wmma_fp16.cpp
similarity index 100%
rename from composable_kernel/example/29_batched_gemm_bias_e_permute/batched_gemm_bias_e_permute_wmma_fp16.cpp
rename to example/29_batched_gemm_bias_e_permute/batched_gemm_bias_e_permute_wmma_fp16.cpp
diff --git a/composable_kernel/example/29_batched_gemm_bias_e_permute/batched_gemm_bias_e_permute_xdl_fp16.cpp b/example/29_batched_gemm_bias_e_permute/batched_gemm_bias_e_permute_xdl_fp16.cpp
similarity index 100%
rename from composable_kernel/example/29_batched_gemm_bias_e_permute/batched_gemm_bias_e_permute_xdl_fp16.cpp
rename to example/29_batched_gemm_bias_e_permute/batched_gemm_bias_e_permute_xdl_fp16.cpp
diff --git a/composable_kernel/example/30_grouped_conv_fwd_multiple_d/CMakeLists.txt b/example/30_grouped_conv_fwd_multiple_d/CMakeLists.txt
similarity index 100%
rename from composable_kernel/example/30_grouped_conv_fwd_multiple_d/CMakeLists.txt
rename to example/30_grouped_conv_fwd_multiple_d/CMakeLists.txt
diff --git a/composable_kernel/example/30_grouped_conv_fwd_multiple_d/README.md b/example/30_grouped_conv_fwd_multiple_d/README.md
similarity index 100%
rename from composable_kernel/example/30_grouped_conv_fwd_multiple_d/README.md
rename to example/30_grouped_conv_fwd_multiple_d/README.md
diff --git a/composable_kernel/example/30_grouped_conv_fwd_multiple_d/common.hpp b/example/30_grouped_conv_fwd_multiple_d/common.hpp
similarity index 100%
rename from composable_kernel/example/30_grouped_conv_fwd_multiple_d/common.hpp
rename to example/30_grouped_conv_fwd_multiple_d/common.hpp
diff --git a/composable_kernel/example/30_grouped_conv_fwd_multiple_d/common_wmma.hpp b/example/30_grouped_conv_fwd_multiple_d/common_wmma.hpp
similarity index 100%
rename from composable_kernel/example/30_grouped_conv_fwd_multiple_d/common_wmma.hpp
rename to example/30_grouped_conv_fwd_multiple_d/common_wmma.hpp
diff --git a/composable_kernel/example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_wmma_fp16.cpp b/example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_wmma_fp16.cpp
similarity index 100%
rename from composable_kernel/example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_wmma_fp16.cpp
rename to example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_wmma_fp16.cpp
diff --git a/composable_kernel/example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_wmma_int8.cpp b/example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_wmma_int8.cpp
similarity index 100%
rename from composable_kernel/example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_wmma_int8.cpp
rename to example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_wmma_int8.cpp
diff --git a/composable_kernel/example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_xdl_bf16.cpp b/example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_xdl_bf16.cpp
similarity index 100%
rename from composable_kernel/example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_xdl_bf16.cpp
rename to example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_xdl_bf16.cpp
diff --git a/composable_kernel/example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_xdl_fp16.cpp b/example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_xdl_fp16.cpp
similarity index 100%
rename from composable_kernel/example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_xdl_fp16.cpp
rename to example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_xdl_fp16.cpp
diff --git a/composable_kernel/example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_xdl_fp32.cpp b/example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_xdl_fp32.cpp
similarity index 100%
rename from composable_kernel/example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_xdl_fp32.cpp
rename to example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_xdl_fp32.cpp
diff --git a/composable_kernel/example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_xdl_int4.cpp b/example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_xdl_int4.cpp
similarity index 100%
rename from composable_kernel/example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_xdl_int4.cpp
rename to example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_xdl_int4.cpp
diff --git a/composable_kernel/example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_xdl_int8.cpp b/example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_xdl_int8.cpp
similarity index 100%
rename from composable_kernel/example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_xdl_int8.cpp
rename to example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_xdl_int8.cpp
diff --git a/composable_kernel/example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_xdl_fp16.cpp b/example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_xdl_fp16.cpp
similarity index 100%
rename from composable_kernel/example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_xdl_fp16.cpp
rename to example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_xdl_fp16.cpp
diff --git a/composable_kernel/example/30_grouped_conv_fwd_multiple_d/run_grouped_conv_fwd_bias_relu_add_example.inc b/example/30_grouped_conv_fwd_multiple_d/run_grouped_conv_fwd_bias_relu_add_example.inc
similarity index 100%
rename from composable_kernel/example/30_grouped_conv_fwd_multiple_d/run_grouped_conv_fwd_bias_relu_add_example.inc
rename to example/30_grouped_conv_fwd_multiple_d/run_grouped_conv_fwd_bias_relu_add_example.inc
diff --git a/composable_kernel/example/30_grouped_conv_fwd_multiple_d/run_grouped_conv_fwd_bias_relu_add_wmma_example.inc b/example/30_grouped_conv_fwd_multiple_d/run_grouped_conv_fwd_bias_relu_add_wmma_example.inc
similarity index 100%
rename from composable_kernel/example/30_grouped_conv_fwd_multiple_d/run_grouped_conv_fwd_bias_relu_add_wmma_example.inc
rename to example/30_grouped_conv_fwd_multiple_d/run_grouped_conv_fwd_bias_relu_add_wmma_example.inc
diff --git a/composable_kernel/example/30_grouped_conv_fwd_multiple_d/run_grouped_conv_fwd_example.inc b/example/30_grouped_conv_fwd_multiple_d/run_grouped_conv_fwd_example.inc
similarity index 100%
rename from composable_kernel/example/30_grouped_conv_fwd_multiple_d/run_grouped_conv_fwd_example.inc
rename to example/30_grouped_conv_fwd_multiple_d/run_grouped_conv_fwd_example.inc
diff --git a/composable_kernel/example/31_batched_gemm_gemm/CMakeLists.txt b/example/31_batched_gemm_gemm/CMakeLists.txt
similarity index 100%
rename from composable_kernel/example/31_batched_gemm_gemm/CMakeLists.txt
rename to example/31_batched_gemm_gemm/CMakeLists.txt
diff --git a/composable_kernel/example/31_batched_gemm_gemm/batched_gemm_gemm_xdl_bf16.cpp b/example/31_batched_gemm_gemm/batched_gemm_gemm_xdl_bf16.cpp
similarity index 100%
rename from composable_kernel/example/31_batched_gemm_gemm/batched_gemm_gemm_xdl_bf16.cpp
rename to example/31_batched_gemm_gemm/batched_gemm_gemm_xdl_bf16.cpp
diff --git a/composable_kernel/example/31_batched_gemm_gemm/batched_gemm_gemm_xdl_fp16.cpp b/example/31_batched_gemm_gemm/batched_gemm_gemm_xdl_fp16.cpp
similarity index 100%
rename from composable_kernel/example/31_batched_gemm_gemm/batched_gemm_gemm_xdl_fp16.cpp
rename to example/31_batched_gemm_gemm/batched_gemm_gemm_xdl_fp16.cpp
diff --git a/composable_kernel/example/31_batched_gemm_gemm/batched_gemm_gemm_xdl_fp32.cpp b/example/31_batched_gemm_gemm/batched_gemm_gemm_xdl_fp32.cpp
similarity index 100%
rename from composable_kernel/example/31_batched_gemm_gemm/batched_gemm_gemm_xdl_fp32.cpp
rename to example/31_batched_gemm_gemm/batched_gemm_gemm_xdl_fp32.cpp
diff --git a/composable_kernel/example/31_batched_gemm_gemm/batched_gemm_gemm_xdl_int4.cpp b/example/31_batched_gemm_gemm/batched_gemm_gemm_xdl_int4.cpp
similarity index 100%
rename from composable_kernel/example/31_batched_gemm_gemm/batched_gemm_gemm_xdl_int4.cpp
rename to example/31_batched_gemm_gemm/batched_gemm_gemm_xdl_int4.cpp
diff --git a/composable_kernel/example/31_batched_gemm_gemm/batched_gemm_gemm_xdl_int8.cpp b/example/31_batched_gemm_gemm/batched_gemm_gemm_xdl_int8.cpp
similarity index 100%
rename from composable_kernel/example/31_batched_gemm_gemm/batched_gemm_gemm_xdl_int8.cpp
rename to example/31_batched_gemm_gemm/batched_gemm_gemm_xdl_int8.cpp
diff --git a/composable_kernel/example/31_batched_gemm_gemm/run_batched_gemm_gemm_example.inc b/example/31_batched_gemm_gemm/run_batched_gemm_gemm_example.inc
similarity index 100%
rename from composable_kernel/example/31_batched_gemm_gemm/run_batched_gemm_gemm_example.inc
rename to example/31_batched_gemm_gemm/run_batched_gemm_gemm_example.inc
diff --git a/composable_kernel/example/32_batched_gemm_scale_softmax_gemm/CMakeLists.txt b/example/32_batched_gemm_scale_softmax_gemm/CMakeLists.txt
similarity index 100%
rename from composable_kernel/example/32_batched_gemm_scale_softmax_gemm/CMakeLists.txt
rename to example/32_batched_gemm_scale_softmax_gemm/CMakeLists.txt
diff --git a/composable_kernel/example/32_batched_gemm_scale_softmax_gemm/batched_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16.cpp b/example/32_batched_gemm_scale_softmax_gemm/batched_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16.cpp
similarity index 100%
rename from composable_kernel/example/32_batched_gemm_scale_softmax_gemm/batched_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16.cpp
rename to example/32_batched_gemm_scale_softmax_gemm/batched_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16.cpp
diff --git a/composable_kernel/example/32_batched_gemm_scale_softmax_gemm/batched_gemm_scale_softmax_gemm_permute_xdl_bf16.cpp b/example/32_batched_gemm_scale_softmax_gemm/batched_gemm_scale_softmax_gemm_permute_xdl_bf16.cpp
similarity index 100%
rename from composable_kernel/example/32_batched_gemm_scale_softmax_gemm/batched_gemm_scale_softmax_gemm_permute_xdl_bf16.cpp
rename to example/32_batched_gemm_scale_softmax_gemm/batched_gemm_scale_softmax_gemm_permute_xdl_bf16.cpp
diff --git a/composable_kernel/example/32_batched_gemm_scale_softmax_gemm/batched_gemm_scale_softmax_gemm_permute_xdl_fp16.cpp b/example/32_batched_gemm_scale_softmax_gemm/batched_gemm_scale_softmax_gemm_permute_xdl_fp16.cpp
similarity index 100%
rename from composable_kernel/example/32_batched_gemm_scale_softmax_gemm/batched_gemm_scale_softmax_gemm_permute_xdl_fp16.cpp
rename to example/32_batched_gemm_scale_softmax_gemm/batched_gemm_scale_softmax_gemm_permute_xdl_fp16.cpp
diff --git a/composable_kernel/example/32_batched_gemm_scale_softmax_gemm/batched_gemm_scale_softmax_gemm_xdl_bf16.cpp b/example/32_batched_gemm_scale_softmax_gemm/batched_gemm_scale_softmax_gemm_xdl_bf16.cpp
similarity index 100%
rename from composable_kernel/example/32_batched_gemm_scale_softmax_gemm/batched_gemm_scale_softmax_gemm_xdl_bf16.cpp
rename to example/32_batched_gemm_scale_softmax_gemm/batched_gemm_scale_softmax_gemm_xdl_bf16.cpp
diff --git a/composable_kernel/example/32_batched_gemm_scale_softmax_gemm/batched_gemm_scale_softmax_gemm_xdl_fp16.cpp b/example/32_batched_gemm_scale_softmax_gemm/batched_gemm_scale_softmax_gemm_xdl_fp16.cpp
similarity index 100%
rename from composable_kernel/example/32_batched_gemm_scale_softmax_gemm/batched_gemm_scale_softmax_gemm_xdl_fp16.cpp
rename to example/32_batched_gemm_scale_softmax_gemm/batched_gemm_scale_softmax_gemm_xdl_fp16.cpp
diff --git a/composable_kernel/example/32_batched_gemm_scale_softmax_gemm/grouped_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16.cpp b/example/32_batched_gemm_scale_softmax_gemm/grouped_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16.cpp
similarity index 100%
rename from composable_kernel/example/32_batched_gemm_scale_softmax_gemm/grouped_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16.cpp
rename to example/32_batched_gemm_scale_softmax_gemm/grouped_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16.cpp
diff --git a/composable_kernel/example/32_batched_gemm_scale_softmax_gemm/grouped_gemm_scale_softmax_gemm_permute_xdl_fp16.cpp b/example/32_batched_gemm_scale_softmax_gemm/grouped_gemm_scale_softmax_gemm_permute_xdl_fp16.cpp
similarity index 100%
rename from composable_kernel/example/32_batched_gemm_scale_softmax_gemm/grouped_gemm_scale_softmax_gemm_permute_xdl_fp16.cpp
rename to example/32_batched_gemm_scale_softmax_gemm/grouped_gemm_scale_softmax_gemm_permute_xdl_fp16.cpp
diff --git a/composable_kernel/example/32_batched_gemm_scale_softmax_gemm/run_batched_gemm_scale_softmax_gemm.inc b/example/32_batched_gemm_scale_softmax_gemm/run_batched_gemm_scale_softmax_gemm.inc
similarity index 100%
rename from composable_kernel/example/32_batched_gemm_scale_softmax_gemm/run_batched_gemm_scale_softmax_gemm.inc
rename to example/32_batched_gemm_scale_softmax_gemm/run_batched_gemm_scale_softmax_gemm.inc
diff --git a/composable_kernel/example/32_batched_gemm_scale_softmax_gemm/run_batched_gemm_scale_softmax_gemm_permute.inc b/example/32_batched_gemm_scale_softmax_gemm/run_batched_gemm_scale_softmax_gemm_permute.inc
similarity index 100%
rename from composable_kernel/example/32_batched_gemm_scale_softmax_gemm/run_batched_gemm_scale_softmax_gemm_permute.inc
rename to example/32_batched_gemm_scale_softmax_gemm/run_batched_gemm_scale_softmax_gemm_permute.inc
diff --git a/composable_kernel/example/32_batched_gemm_scale_softmax_gemm/run_grouped_gemm_scale_softmax_gemm_permute.inc b/example/32_batched_gemm_scale_softmax_gemm/run_grouped_gemm_scale_softmax_gemm_permute.inc
similarity index 100%
rename from composable_kernel/example/32_batched_gemm_scale_softmax_gemm/run_grouped_gemm_scale_softmax_gemm_permute.inc
rename to example/32_batched_gemm_scale_softmax_gemm/run_grouped_gemm_scale_softmax_gemm_permute.inc
diff --git a/composable_kernel/example/33_multiple_reduce/CMakeLists.txt b/example/33_multiple_reduce/CMakeLists.txt
similarity index 100%
rename from composable_kernel/example/33_multiple_reduce/CMakeLists.txt
rename to example/33_multiple_reduce/CMakeLists.txt
diff --git a/composable_kernel/example/33_multiple_reduce/README.md b/example/33_multiple_reduce/README.md
similarity index 100%
rename from composable_kernel/example/33_multiple_reduce/README.md
rename to example/33_multiple_reduce/README.md
diff --git a/composable_kernel/example/33_multiple_reduce/dual_reduce_common.hpp b/example/33_multiple_reduce/dual_reduce_common.hpp
similarity index 100%
rename from composable_kernel/example/33_multiple_reduce/dual_reduce_common.hpp
rename to example/33_multiple_reduce/dual_reduce_common.hpp
diff --git a/composable_kernel/example/33_multiple_reduce/dual_reduce_multiblock.cpp b/example/33_multiple_reduce/dual_reduce_multiblock.cpp
similarity index 100%
rename from composable_kernel/example/33_multiple_reduce/dual_reduce_multiblock.cpp
rename to example/33_multiple_reduce/dual_reduce_multiblock.cpp
diff --git a/composable_kernel/example/33_multiple_reduce/dual_reduce_threadwise.cpp b/example/33_multiple_reduce/dual_reduce_threadwise.cpp
similarity index 100%
rename from composable_kernel/example/33_multiple_reduce/dual_reduce_threadwise.cpp
rename to example/33_multiple_reduce/dual_reduce_threadwise.cpp
diff --git a/composable_kernel/example/34_batchnorm/CMakeLists.txt b/example/34_batchnorm/CMakeLists.txt
similarity index 100%
rename from composable_kernel/example/34_batchnorm/CMakeLists.txt
rename to example/34_batchnorm/CMakeLists.txt
diff --git a/composable_kernel/example/34_batchnorm/README.md b/example/34_batchnorm/README.md
similarity index 100%
rename from composable_kernel/example/34_batchnorm/README.md
rename to example/34_batchnorm/README.md
diff --git a/composable_kernel/example/34_batchnorm/batchnorm_backward_nhwc.cpp b/example/34_batchnorm/batchnorm_backward_nhwc.cpp
similarity index 100%
rename from composable_kernel/example/34_batchnorm/batchnorm_backward_nhwc.cpp
rename to example/34_batchnorm/batchnorm_backward_nhwc.cpp
diff --git a/composable_kernel/example/34_batchnorm/batchnorm_common.hpp b/example/34_batchnorm/batchnorm_common.hpp
similarity index 100%
rename from composable_kernel/example/34_batchnorm/batchnorm_common.hpp
rename to example/34_batchnorm/batchnorm_common.hpp
diff --git a/composable_kernel/example/34_batchnorm/batchnorm_forward_inferring_nhwc.cpp b/example/34_batchnorm/batchnorm_forward_inferring_nhwc.cpp
similarity index 100%
rename from composable_kernel/example/34_batchnorm/batchnorm_forward_inferring_nhwc.cpp
rename to example/34_batchnorm/batchnorm_forward_inferring_nhwc.cpp
diff --git a/composable_kernel/example/34_batchnorm/batchnorm_forward_training_nhwc.cpp b/example/34_batchnorm/batchnorm_forward_training_nhwc.cpp
similarity index 100%
rename from composable_kernel/example/34_batchnorm/batchnorm_forward_training_nhwc.cpp
rename to example/34_batchnorm/batchnorm_forward_training_nhwc.cpp
diff --git a/composable_kernel/example/34_batchnorm/batchnorm_forward_training_nhwc_obsolete.cpp b/example/34_batchnorm/batchnorm_forward_training_nhwc_obsolete.cpp
similarity index 100%
rename from composable_kernel/example/34_batchnorm/batchnorm_forward_training_nhwc_obsolete.cpp
rename to example/34_batchnorm/batchnorm_forward_training_nhwc_obsolete.cpp
diff --git a/composable_kernel/example/34_batchnorm/batchnorm_infer_impl.hpp b/example/34_batchnorm/batchnorm_infer_impl.hpp
similarity index 100%
rename from composable_kernel/example/34_batchnorm/batchnorm_infer_impl.hpp
rename to example/34_batchnorm/batchnorm_infer_impl.hpp
diff --git a/composable_kernel/example/35_splitK_gemm/CMakeLists.txt b/example/35_splitK_gemm/CMakeLists.txt
similarity index 100%
rename from composable_kernel/example/35_splitK_gemm/CMakeLists.txt
rename to example/35_splitK_gemm/CMakeLists.txt
diff --git a/composable_kernel/example/35_splitK_gemm/run_splitK_gemm_example.inc b/example/35_splitK_gemm/run_splitK_gemm_example.inc
similarity index 100%
rename from composable_kernel/example/35_splitK_gemm/run_splitK_gemm_example.inc
rename to example/35_splitK_gemm/run_splitK_gemm_example.inc
diff --git a/composable_kernel/example/35_splitK_gemm/splitK_gemm_xdl_bf16.cpp b/example/35_splitK_gemm/splitK_gemm_xdl_bf16.cpp
similarity index 100%
rename from composable_kernel/example/35_splitK_gemm/splitK_gemm_xdl_bf16.cpp
rename to example/35_splitK_gemm/splitK_gemm_xdl_bf16.cpp
diff --git a/composable_kernel/example/35_splitK_gemm/splitK_gemm_xdl_fp16.cpp b/example/35_splitK_gemm/splitK_gemm_xdl_fp16.cpp
similarity index 100%
rename from composable_kernel/example/35_splitK_gemm/splitK_gemm_xdl_fp16.cpp
rename to example/35_splitK_gemm/splitK_gemm_xdl_fp16.cpp
diff --git a/composable_kernel/example/35_splitK_gemm/splitK_gemm_xdl_fp32.cpp b/example/35_splitK_gemm/splitK_gemm_xdl_fp32.cpp
similarity index 100%
rename from composable_kernel/example/35_splitK_gemm/splitK_gemm_xdl_fp32.cpp
rename to example/35_splitK_gemm/splitK_gemm_xdl_fp32.cpp
diff --git a/composable_kernel/example/35_splitK_gemm/splitK_gemm_xdl_int4.cpp b/example/35_splitK_gemm/splitK_gemm_xdl_int4.cpp
similarity index 100%
rename from composable_kernel/example/35_splitK_gemm/splitK_gemm_xdl_int4.cpp
rename to example/35_splitK_gemm/splitK_gemm_xdl_int4.cpp
diff --git a/composable_kernel/example/35_splitK_gemm/splitK_gemm_xdl_int8.cpp b/example/35_splitK_gemm/splitK_gemm_xdl_int8.cpp
similarity index 100%
rename from composable_kernel/example/35_splitK_gemm/splitK_gemm_xdl_int8.cpp
rename to example/35_splitK_gemm/splitK_gemm_xdl_int8.cpp
diff --git a/composable_kernel/example/36_sparse_embedding/CMakeLists.txt b/example/36_sparse_embedding/CMakeLists.txt
similarity index 100%
rename from composable_kernel/example/36_sparse_embedding/CMakeLists.txt
rename to example/36_sparse_embedding/CMakeLists.txt
diff --git a/composable_kernel/example/36_sparse_embedding/sparse_embedding3_forward_layernorm.cpp b/example/36_sparse_embedding/sparse_embedding3_forward_layernorm.cpp
similarity index 100%
rename from composable_kernel/example/36_sparse_embedding/sparse_embedding3_forward_layernorm.cpp
rename to example/36_sparse_embedding/sparse_embedding3_forward_layernorm.cpp
diff --git a/composable_kernel/example/37_batched_gemm_add_add_relu_gemm_add/CMakeLists.txt b/example/37_batched_gemm_add_add_relu_gemm_add/CMakeLists.txt
similarity index 100%
rename from composable_kernel/example/37_batched_gemm_add_add_relu_gemm_add/CMakeLists.txt
rename to example/37_batched_gemm_add_add_relu_gemm_add/CMakeLists.txt
diff --git a/composable_kernel/example/37_batched_gemm_add_add_relu_gemm_add/batched_gemm_add_add_relu_gemm_add_xdl_fp16.cpp b/example/37_batched_gemm_add_add_relu_gemm_add/batched_gemm_add_add_relu_gemm_add_xdl_fp16.cpp
similarity index 100%
rename from composable_kernel/example/37_batched_gemm_add_add_relu_gemm_add/batched_gemm_add_add_relu_gemm_add_xdl_fp16.cpp
rename to example/37_batched_gemm_add_add_relu_gemm_add/batched_gemm_add_add_relu_gemm_add_xdl_fp16.cpp
diff --git a/composable_kernel/example/38_grouped_conv_bwd_data_multiple_d/CMakeLists.txt b/example/38_grouped_conv_bwd_data_multiple_d/CMakeLists.txt
similarity index 100%
rename from composable_kernel/example/38_grouped_conv_bwd_data_multiple_d/CMakeLists.txt
rename to example/38_grouped_conv_bwd_data_multiple_d/CMakeLists.txt
diff --git a/composable_kernel/example/38_grouped_conv_bwd_data_multiple_d/common.hpp b/example/38_grouped_conv_bwd_data_multiple_d/common.hpp
similarity index 100%
rename from composable_kernel/example/38_grouped_conv_bwd_data_multiple_d/common.hpp
rename to example/38_grouped_conv_bwd_data_multiple_d/common.hpp
diff --git a/composable_kernel/example/38_grouped_conv_bwd_data_multiple_d/grouped_conv_bwd_data_bias_relu_xdl_fp16.cpp b/example/38_grouped_conv_bwd_data_multiple_d/grouped_conv_bwd_data_bias_relu_xdl_fp16.cpp
similarity index 100%
rename from composable_kernel/example/38_grouped_conv_bwd_data_multiple_d/grouped_conv_bwd_data_bias_relu_xdl_fp16.cpp
rename to example/38_grouped_conv_bwd_data_multiple_d/grouped_conv_bwd_data_bias_relu_xdl_fp16.cpp
diff --git a/composable_kernel/example/38_grouped_conv_bwd_data_multiple_d/grouped_conv_bwd_data_wmma_fp16.cpp b/example/38_grouped_conv_bwd_data_multiple_d/grouped_conv_bwd_data_wmma_fp16.cpp
similarity index 100%
rename from composable_kernel/example/38_grouped_conv_bwd_data_multiple_d/grouped_conv_bwd_data_wmma_fp16.cpp
rename to example/38_grouped_conv_bwd_data_multiple_d/grouped_conv_bwd_data_wmma_fp16.cpp
diff --git a/composable_kernel/example/38_grouped_conv_bwd_data_multiple_d/grouped_conv_bwd_data_xdl_fp16.cpp b/example/38_grouped_conv_bwd_data_multiple_d/grouped_conv_bwd_data_xdl_fp16.cpp
similarity index 100%
rename from composable_kernel/example/38_grouped_conv_bwd_data_multiple_d/grouped_conv_bwd_data_xdl_fp16.cpp
rename to example/38_grouped_conv_bwd_data_multiple_d/grouped_conv_bwd_data_xdl_fp16.cpp
diff --git a/composable_kernel/example/38_grouped_conv_bwd_data_multiple_d/run_grouped_conv_bwd_data_bias_relu_example.inc b/example/38_grouped_conv_bwd_data_multiple_d/run_grouped_conv_bwd_data_bias_relu_example.inc
similarity index 100%
rename from composable_kernel/example/38_grouped_conv_bwd_data_multiple_d/run_grouped_conv_bwd_data_bias_relu_example.inc
rename to example/38_grouped_conv_bwd_data_multiple_d/run_grouped_conv_bwd_data_bias_relu_example.inc
diff --git a/composable_kernel/example/38_grouped_conv_bwd_data_multiple_d/run_grouped_conv_bwd_data_example.inc b/example/38_grouped_conv_bwd_data_multiple_d/run_grouped_conv_bwd_data_example.inc
similarity index 100%
rename from composable_kernel/example/38_grouped_conv_bwd_data_multiple_d/run_grouped_conv_bwd_data_example.inc
rename to example/38_grouped_conv_bwd_data_multiple_d/run_grouped_conv_bwd_data_example.inc
diff --git a/composable_kernel/example/39_permute/CMakeLists.txt b/example/39_permute/CMakeLists.txt
similarity index 100%
rename from composable_kernel/example/39_permute/CMakeLists.txt
rename to example/39_permute/CMakeLists.txt
diff --git a/composable_kernel/example/39_permute/common.hpp b/example/39_permute/common.hpp
similarity index 100%
rename from composable_kernel/example/39_permute/common.hpp
rename to example/39_permute/common.hpp
diff --git a/composable_kernel/example/39_permute/permute_1xHxW_fp16.cpp b/example/39_permute/permute_1xHxW_fp16.cpp
similarity index 100%
rename from composable_kernel/example/39_permute/permute_1xHxW_fp16.cpp
rename to example/39_permute/permute_1xHxW_fp16.cpp
diff --git a/composable_kernel/example/39_permute/permute_HxWx4_fp16.cpp b/example/39_permute/permute_HxWx4_fp16.cpp
similarity index 100%
rename from composable_kernel/example/39_permute/permute_HxWx4_fp16.cpp
rename to example/39_permute/permute_HxWx4_fp16.cpp
diff --git a/composable_kernel/example/39_permute/permute_NxHxW_fp16.cpp b/example/39_permute/permute_NxHxW_fp16.cpp
similarity index 100%
rename from composable_kernel/example/39_permute/permute_NxHxW_fp16.cpp
rename to example/39_permute/permute_NxHxW_fp16.cpp
diff --git a/composable_kernel/example/39_permute/run_permute_bundle_example.inc b/example/39_permute/run_permute_bundle_example.inc
similarity index 100%
rename from composable_kernel/example/39_permute/run_permute_bundle_example.inc
rename to example/39_permute/run_permute_bundle_example.inc
diff --git a/composable_kernel/example/39_permute/run_permute_element_example.inc b/example/39_permute/run_permute_element_example.inc
similarity index 100%
rename from composable_kernel/example/39_permute/run_permute_element_example.inc
rename to example/39_permute/run_permute_element_example.inc
diff --git a/composable_kernel/example/40_conv2d_fwd_quantization/CMakeLists.txt b/example/40_conv2d_fwd_quantization/CMakeLists.txt
similarity index 100%
rename from composable_kernel/example/40_conv2d_fwd_quantization/CMakeLists.txt
rename to example/40_conv2d_fwd_quantization/CMakeLists.txt
diff --git a/composable_kernel/example/40_conv2d_fwd_quantization/common.hpp b/example/40_conv2d_fwd_quantization/common.hpp
similarity index 100%
rename from composable_kernel/example/40_conv2d_fwd_quantization/common.hpp
rename to example/40_conv2d_fwd_quantization/common.hpp
diff --git a/composable_kernel/example/40_conv2d_fwd_quantization/conv2d_fwd_dl_bias_relu_perchannel_quantization_int8.cpp b/example/40_conv2d_fwd_quantization/conv2d_fwd_dl_bias_relu_perchannel_quantization_int8.cpp
similarity index 100%
rename from composable_kernel/example/40_conv2d_fwd_quantization/conv2d_fwd_dl_bias_relu_perchannel_quantization_int8.cpp
rename to example/40_conv2d_fwd_quantization/conv2d_fwd_dl_bias_relu_perchannel_quantization_int8.cpp
diff --git a/composable_kernel/example/40_conv2d_fwd_quantization/conv2d_fwd_dl_bias_relu_perlayer_quantization_int8.cpp b/example/40_conv2d_fwd_quantization/conv2d_fwd_dl_bias_relu_perlayer_quantization_int8.cpp
similarity index 100%
rename from composable_kernel/example/40_conv2d_fwd_quantization/conv2d_fwd_dl_bias_relu_perlayer_quantization_int8.cpp
rename to example/40_conv2d_fwd_quantization/conv2d_fwd_dl_bias_relu_perlayer_quantization_int8.cpp
diff --git a/composable_kernel/example/40_conv2d_fwd_quantization/conv2d_fwd_dl_bias_tanh_perchannel_quantization_int8.cpp b/example/40_conv2d_fwd_quantization/conv2d_fwd_dl_bias_tanh_perchannel_quantization_int8.cpp
similarity index 100%
rename from composable_kernel/example/40_conv2d_fwd_quantization/conv2d_fwd_dl_bias_tanh_perchannel_quantization_int8.cpp
rename to example/40_conv2d_fwd_quantization/conv2d_fwd_dl_bias_tanh_perchannel_quantization_int8.cpp
diff --git a/composable_kernel/example/40_conv2d_fwd_quantization/conv2d_fwd_dl_bias_tanh_perlayer_quantization_int8.cpp b/example/40_conv2d_fwd_quantization/conv2d_fwd_dl_bias_tanh_perlayer_quantization_int8.cpp
similarity index 100%
rename from composable_kernel/example/40_conv2d_fwd_quantization/conv2d_fwd_dl_bias_tanh_perlayer_quantization_int8.cpp
rename to example/40_conv2d_fwd_quantization/conv2d_fwd_dl_bias_tanh_perlayer_quantization_int8.cpp
diff --git a/composable_kernel/example/40_conv2d_fwd_quantization/conv2d_fwd_dl_perchannel_quantization_int8.cpp b/example/40_conv2d_fwd_quantization/conv2d_fwd_dl_perchannel_quantization_int8.cpp
similarity index 100%
rename from composable_kernel/example/40_conv2d_fwd_quantization/conv2d_fwd_dl_perchannel_quantization_int8.cpp
rename to example/40_conv2d_fwd_quantization/conv2d_fwd_dl_perchannel_quantization_int8.cpp
diff --git a/composable_kernel/example/40_conv2d_fwd_quantization/conv2d_fwd_dl_perlayer_quantization_int8.cpp b/example/40_conv2d_fwd_quantization/conv2d_fwd_dl_perlayer_quantization_int8.cpp
similarity index 100%
rename from composable_kernel/example/40_conv2d_fwd_quantization/conv2d_fwd_dl_perlayer_quantization_int8.cpp
rename to example/40_conv2d_fwd_quantization/conv2d_fwd_dl_perlayer_quantization_int8.cpp
diff --git a/composable_kernel/example/40_conv2d_fwd_quantization/conv2d_fwd_xdl_bias_relu_perchannel_quantization_int8.cpp b/example/40_conv2d_fwd_quantization/conv2d_fwd_xdl_bias_relu_perchannel_quantization_int8.cpp
similarity index 100%
rename from composable_kernel/example/40_conv2d_fwd_quantization/conv2d_fwd_xdl_bias_relu_perchannel_quantization_int8.cpp
rename to example/40_conv2d_fwd_quantization/conv2d_fwd_xdl_bias_relu_perchannel_quantization_int8.cpp
diff --git a/composable_kernel/example/40_conv2d_fwd_quantization/conv2d_fwd_xdl_bias_relu_perlayer_quantization_int8.cpp b/example/40_conv2d_fwd_quantization/conv2d_fwd_xdl_bias_relu_perlayer_quantization_int8.cpp
similarity index 100%
rename from composable_kernel/example/40_conv2d_fwd_quantization/conv2d_fwd_xdl_bias_relu_perlayer_quantization_int8.cpp
rename to example/40_conv2d_fwd_quantization/conv2d_fwd_xdl_bias_relu_perlayer_quantization_int8.cpp
diff --git a/composable_kernel/example/40_conv2d_fwd_quantization/conv2d_fwd_xdl_perchannel_quantization_int8.cpp b/example/40_conv2d_fwd_quantization/conv2d_fwd_xdl_perchannel_quantization_int8.cpp
similarity index 100%
rename from composable_kernel/example/40_conv2d_fwd_quantization/conv2d_fwd_xdl_perchannel_quantization_int8.cpp
rename to example/40_conv2d_fwd_quantization/conv2d_fwd_xdl_perchannel_quantization_int8.cpp
diff --git a/composable_kernel/example/40_conv2d_fwd_quantization/conv2d_fwd_xdl_perlayer_quantization_int8.cpp b/example/40_conv2d_fwd_quantization/conv2d_fwd_xdl_perlayer_quantization_int8.cpp
similarity index 100%
rename from composable_kernel/example/40_conv2d_fwd_quantization/conv2d_fwd_xdl_perlayer_quantization_int8.cpp
rename to example/40_conv2d_fwd_quantization/conv2d_fwd_xdl_perlayer_quantization_int8.cpp
diff --git a/composable_kernel/example/40_conv2d_fwd_quantization/run_conv2d_fwd_bias_perchannel_quantization_example.inc b/example/40_conv2d_fwd_quantization/run_conv2d_fwd_bias_perchannel_quantization_example.inc
similarity index 100%
rename from composable_kernel/example/40_conv2d_fwd_quantization/run_conv2d_fwd_bias_perchannel_quantization_example.inc
rename to example/40_conv2d_fwd_quantization/run_conv2d_fwd_bias_perchannel_quantization_example.inc
diff --git a/composable_kernel/example/40_conv2d_fwd_quantization/run_conv2d_fwd_bias_perlayer_quantization_example.inc b/example/40_conv2d_fwd_quantization/run_conv2d_fwd_bias_perlayer_quantization_example.inc
similarity index 100%
rename from composable_kernel/example/40_conv2d_fwd_quantization/run_conv2d_fwd_bias_perlayer_quantization_example.inc
rename to example/40_conv2d_fwd_quantization/run_conv2d_fwd_bias_perlayer_quantization_example.inc
diff --git a/composable_kernel/example/40_conv2d_fwd_quantization/run_conv2d_fwd_perchannel_quantization_example.inc b/example/40_conv2d_fwd_quantization/run_conv2d_fwd_perchannel_quantization_example.inc
similarity index 100%
rename from composable_kernel/example/40_conv2d_fwd_quantization/run_conv2d_fwd_perchannel_quantization_example.inc
rename to example/40_conv2d_fwd_quantization/run_conv2d_fwd_perchannel_quantization_example.inc
diff --git a/composable_kernel/example/40_conv2d_fwd_quantization/run_conv2d_fwd_perlayer_quantization_example.inc b/example/40_conv2d_fwd_quantization/run_conv2d_fwd_perlayer_quantization_example.inc
similarity index 100%
rename from composable_kernel/example/40_conv2d_fwd_quantization/run_conv2d_fwd_perlayer_quantization_example.inc
rename to example/40_conv2d_fwd_quantization/run_conv2d_fwd_perlayer_quantization_example.inc
diff --git a/composable_kernel/example/41_grouped_conv_conv_fwd/CMakeLists.txt b/example/41_grouped_conv_conv_fwd/CMakeLists.txt
similarity index 100%
rename from composable_kernel/example/41_grouped_conv_conv_fwd/CMakeLists.txt
rename to example/41_grouped_conv_conv_fwd/CMakeLists.txt
diff --git a/composable_kernel/example/41_grouped_conv_conv_fwd/grouped_conv_conv_fwd_xdl_bf16.cpp b/example/41_grouped_conv_conv_fwd/grouped_conv_conv_fwd_xdl_bf16.cpp
similarity index 100%
rename from composable_kernel/example/41_grouped_conv_conv_fwd/grouped_conv_conv_fwd_xdl_bf16.cpp
rename to example/41_grouped_conv_conv_fwd/grouped_conv_conv_fwd_xdl_bf16.cpp
diff --git a/composable_kernel/example/41_grouped_conv_conv_fwd/grouped_conv_conv_fwd_xdl_fp16.cpp b/example/41_grouped_conv_conv_fwd/grouped_conv_conv_fwd_xdl_fp16.cpp
similarity index 100%
rename from composable_kernel/example/41_grouped_conv_conv_fwd/grouped_conv_conv_fwd_xdl_fp16.cpp
rename to example/41_grouped_conv_conv_fwd/grouped_conv_conv_fwd_xdl_fp16.cpp
diff --git a/composable_kernel/example/41_grouped_conv_conv_fwd/grouped_conv_conv_fwd_xdl_fp32.cpp b/example/41_grouped_conv_conv_fwd/grouped_conv_conv_fwd_xdl_fp32.cpp
similarity index 100%
rename from composable_kernel/example/41_grouped_conv_conv_fwd/grouped_conv_conv_fwd_xdl_fp32.cpp
rename to example/41_grouped_conv_conv_fwd/grouped_conv_conv_fwd_xdl_fp32.cpp
diff --git a/composable_kernel/example/41_grouped_conv_conv_fwd/grouped_conv_conv_fwd_xdl_int4.cpp b/example/41_grouped_conv_conv_fwd/grouped_conv_conv_fwd_xdl_int4.cpp
similarity index 100%
rename from composable_kernel/example/41_grouped_conv_conv_fwd/grouped_conv_conv_fwd_xdl_int4.cpp
rename to example/41_grouped_conv_conv_fwd/grouped_conv_conv_fwd_xdl_int4.cpp
diff --git a/composable_kernel/example/41_grouped_conv_conv_fwd/grouped_conv_conv_fwd_xdl_int8.cpp b/example/41_grouped_conv_conv_fwd/grouped_conv_conv_fwd_xdl_int8.cpp
similarity index 100%
rename from composable_kernel/example/41_grouped_conv_conv_fwd/grouped_conv_conv_fwd_xdl_int8.cpp
rename to example/41_grouped_conv_conv_fwd/grouped_conv_conv_fwd_xdl_int8.cpp
diff --git a/composable_kernel/example/41_grouped_conv_conv_fwd/run_grouped_conv_conv_fwd_example.inc b/example/41_grouped_conv_conv_fwd/run_grouped_conv_conv_fwd_example.inc
similarity index 100%
rename from composable_kernel/example/41_grouped_conv_conv_fwd/run_grouped_conv_conv_fwd_example.inc
rename to example/41_grouped_conv_conv_fwd/run_grouped_conv_conv_fwd_example.inc
diff --git a/composable_kernel/example/42_groupnorm/CMakeLists.txt b/example/42_groupnorm/CMakeLists.txt
similarity index 100%
rename from composable_kernel/example/42_groupnorm/CMakeLists.txt
rename to example/42_groupnorm/CMakeLists.txt
diff --git a/composable_kernel/example/42_groupnorm/common.hpp b/example/42_groupnorm/common.hpp
similarity index 100%
rename from composable_kernel/example/42_groupnorm/common.hpp
rename to example/42_groupnorm/common.hpp
diff --git a/composable_kernel/example/42_groupnorm/groupnorm_sigmoid_mul_fp16.cpp b/example/42_groupnorm/groupnorm_sigmoid_mul_fp16.cpp
similarity index 100%
rename from composable_kernel/example/42_groupnorm/groupnorm_sigmoid_mul_fp16.cpp
rename to example/42_groupnorm/groupnorm_sigmoid_mul_fp16.cpp
diff --git a/composable_kernel/example/42_groupnorm/groupnorm_splitk_fp16.cpp b/example/42_groupnorm/groupnorm_splitk_fp16.cpp
similarity index 100%
rename from composable_kernel/example/42_groupnorm/groupnorm_splitk_fp16.cpp
rename to example/42_groupnorm/groupnorm_splitk_fp16.cpp
diff --git a/composable_kernel/example/42_groupnorm/groupnorm_swish_fp16.cpp b/example/42_groupnorm/groupnorm_swish_fp16.cpp
similarity index 100%
rename from composable_kernel/example/42_groupnorm/groupnorm_swish_fp16.cpp
rename to example/42_groupnorm/groupnorm_swish_fp16.cpp
diff --git a/composable_kernel/example/42_groupnorm/run_groupnorm_example.inc b/example/42_groupnorm/run_groupnorm_example.inc
similarity index 100%
rename from composable_kernel/example/42_groupnorm/run_groupnorm_example.inc
rename to example/42_groupnorm/run_groupnorm_example.inc
diff --git a/composable_kernel/example/43_splitk_gemm_bias_e_permute/CMakeLists.txt b/example/43_splitk_gemm_bias_e_permute/CMakeLists.txt
similarity index 100%
rename from composable_kernel/example/43_splitk_gemm_bias_e_permute/CMakeLists.txt
rename to example/43_splitk_gemm_bias_e_permute/CMakeLists.txt
diff --git a/composable_kernel/example/43_splitk_gemm_bias_e_permute/splitk_gemm_bias_e_permute_xdl_fp16.cpp b/example/43_splitk_gemm_bias_e_permute/splitk_gemm_bias_e_permute_xdl_fp16.cpp
similarity index 100%
rename from composable_kernel/example/43_splitk_gemm_bias_e_permute/splitk_gemm_bias_e_permute_xdl_fp16.cpp
rename to example/43_splitk_gemm_bias_e_permute/splitk_gemm_bias_e_permute_xdl_fp16.cpp
diff --git a/composable_kernel/example/43_splitk_gemm_bias_e_permute/splitk_gemm_bias_e_permute_xdl_fp32.cpp b/example/43_splitk_gemm_bias_e_permute/splitk_gemm_bias_e_permute_xdl_fp32.cpp
similarity index 100%
rename from composable_kernel/example/43_splitk_gemm_bias_e_permute/splitk_gemm_bias_e_permute_xdl_fp32.cpp
rename to example/43_splitk_gemm_bias_e_permute/splitk_gemm_bias_e_permute_xdl_fp32.cpp
diff --git a/composable_kernel/example/44_elementwise_permute/CMakeLists.txt b/example/44_elementwise_permute/CMakeLists.txt
similarity index 100%
rename from composable_kernel/example/44_elementwise_permute/CMakeLists.txt
rename to example/44_elementwise_permute/CMakeLists.txt
diff --git a/composable_kernel/example/44_elementwise_permute/elementwise_permute_4D_fp16.cpp b/example/44_elementwise_permute/elementwise_permute_4D_fp16.cpp
similarity index 100%
rename from composable_kernel/example/44_elementwise_permute/elementwise_permute_4D_fp16.cpp
rename to example/44_elementwise_permute/elementwise_permute_4D_fp16.cpp
diff --git a/composable_kernel/example/44_elementwise_permute/elementwise_permute_4D_fp16_2d.cpp b/example/44_elementwise_permute/elementwise_permute_4D_fp16_2d.cpp
similarity index 100%
rename from composable_kernel/example/44_elementwise_permute/elementwise_permute_4D_fp16_2d.cpp
rename to example/44_elementwise_permute/elementwise_permute_4D_fp16_2d.cpp
diff --git a/composable_kernel/example/45_elementwise_normalization/CMakeLists.txt b/example/45_elementwise_normalization/CMakeLists.txt
similarity index 100%
rename from composable_kernel/example/45_elementwise_normalization/CMakeLists.txt
rename to example/45_elementwise_normalization/CMakeLists.txt
diff --git a/composable_kernel/example/45_elementwise_normalization/elementwise_layernorm_blockwise.cpp b/example/45_elementwise_normalization/elementwise_layernorm_blockwise.cpp
similarity index 100%
rename from composable_kernel/example/45_elementwise_normalization/elementwise_layernorm_blockwise.cpp
rename to example/45_elementwise_normalization/elementwise_layernorm_blockwise.cpp
diff --git a/composable_kernel/example/46_gemm_add_multiply/CMakeLists.txt b/example/46_gemm_add_multiply/CMakeLists.txt
similarity index 100%
rename from composable_kernel/example/46_gemm_add_multiply/CMakeLists.txt
rename to example/46_gemm_add_multiply/CMakeLists.txt
diff --git a/composable_kernel/example/46_gemm_add_multiply/README.md b/example/46_gemm_add_multiply/README.md
similarity index 100%
rename from composable_kernel/example/46_gemm_add_multiply/README.md
rename to example/46_gemm_add_multiply/README.md
diff --git a/composable_kernel/example/46_gemm_add_multiply/common.hpp b/example/46_gemm_add_multiply/common.hpp
similarity index 100%
rename from composable_kernel/example/46_gemm_add_multiply/common.hpp
rename to example/46_gemm_add_multiply/common.hpp
diff --git a/composable_kernel/example/46_gemm_add_multiply/gemm_add_multiply_dl_fp16.cpp b/example/46_gemm_add_multiply/gemm_add_multiply_dl_fp16.cpp
similarity index 100%
rename from composable_kernel/example/46_gemm_add_multiply/gemm_add_multiply_dl_fp16.cpp
rename to example/46_gemm_add_multiply/gemm_add_multiply_dl_fp16.cpp
diff --git a/composable_kernel/example/46_gemm_add_multiply/gemm_add_multiply_xdl_fp16.cpp b/example/46_gemm_add_multiply/gemm_add_multiply_xdl_fp16.cpp
similarity index 100%
rename from composable_kernel/example/46_gemm_add_multiply/gemm_add_multiply_xdl_fp16.cpp
rename to example/46_gemm_add_multiply/gemm_add_multiply_xdl_fp16.cpp
diff --git a/composable_kernel/example/46_gemm_add_multiply/run_gemm_add_multiply_example.inc b/example/46_gemm_add_multiply/run_gemm_add_multiply_example.inc
similarity index 100%
rename from composable_kernel/example/46_gemm_add_multiply/run_gemm_add_multiply_example.inc
rename to example/46_gemm_add_multiply/run_gemm_add_multiply_example.inc
diff --git a/composable_kernel/example/47_gemm_bias_softmax_gemm_permute/CMakeLists.txt b/example/47_gemm_bias_softmax_gemm_permute/CMakeLists.txt
similarity index 100%
rename from composable_kernel/example/47_gemm_bias_softmax_gemm_permute/CMakeLists.txt
rename to example/47_gemm_bias_softmax_gemm_permute/CMakeLists.txt
diff --git a/composable_kernel/example/47_gemm_bias_softmax_gemm_permute/gemm_bias_softmax_gemm_permute.cpp b/example/47_gemm_bias_softmax_gemm_permute/gemm_bias_softmax_gemm_permute.cpp
similarity index 100%
rename from composable_kernel/example/47_gemm_bias_softmax_gemm_permute/gemm_bias_softmax_gemm_permute.cpp
rename to example/47_gemm_bias_softmax_gemm_permute/gemm_bias_softmax_gemm_permute.cpp
diff --git a/composable_kernel/example/48_pool3d_fwd/CMakeLists.txt b/example/48_pool3d_fwd/CMakeLists.txt
similarity index 100%
rename from composable_kernel/example/48_pool3d_fwd/CMakeLists.txt
rename to example/48_pool3d_fwd/CMakeLists.txt
diff --git a/composable_kernel/example/48_pool3d_fwd/pool3d_fwd_common.hpp b/example/48_pool3d_fwd/pool3d_fwd_common.hpp
similarity index 100%
rename from composable_kernel/example/48_pool3d_fwd/pool3d_fwd_common.hpp
rename to example/48_pool3d_fwd/pool3d_fwd_common.hpp
diff --git a/composable_kernel/example/48_pool3d_fwd/pool3d_fwd_fp16.cpp b/example/48_pool3d_fwd/pool3d_fwd_fp16.cpp
similarity index 100%
rename from composable_kernel/example/48_pool3d_fwd/pool3d_fwd_fp16.cpp
rename to example/48_pool3d_fwd/pool3d_fwd_fp16.cpp
diff --git a/composable_kernel/example/49_maxpool2d_bwd/CMakeLists.txt b/example/49_maxpool2d_bwd/CMakeLists.txt
similarity index 100%
rename from composable_kernel/example/49_maxpool2d_bwd/CMakeLists.txt
rename to example/49_maxpool2d_bwd/CMakeLists.txt
diff --git a/composable_kernel/example/49_maxpool2d_bwd/maxpool2d_bwd_bf16.cpp b/example/49_maxpool2d_bwd/maxpool2d_bwd_bf16.cpp
similarity index 100%
rename from composable_kernel/example/49_maxpool2d_bwd/maxpool2d_bwd_bf16.cpp
rename to example/49_maxpool2d_bwd/maxpool2d_bwd_bf16.cpp
diff --git a/composable_kernel/example/49_maxpool2d_bwd/maxpool2d_bwd_common.hpp b/example/49_maxpool2d_bwd/maxpool2d_bwd_common.hpp
similarity index 100%
rename from composable_kernel/example/49_maxpool2d_bwd/maxpool2d_bwd_common.hpp
rename to example/49_maxpool2d_bwd/maxpool2d_bwd_common.hpp
diff --git a/composable_kernel/example/49_maxpool2d_bwd/maxpool2d_bwd_fp16.cpp b/example/49_maxpool2d_bwd/maxpool2d_bwd_fp16.cpp
similarity index 100%
rename from composable_kernel/example/49_maxpool2d_bwd/maxpool2d_bwd_fp16.cpp
rename to example/49_maxpool2d_bwd/maxpool2d_bwd_fp16.cpp
diff --git a/composable_kernel/example/49_maxpool2d_bwd/maxpool2d_bwd_fp32.cpp b/example/49_maxpool2d_bwd/maxpool2d_bwd_fp32.cpp
similarity index 100%
rename from composable_kernel/example/49_maxpool2d_bwd/maxpool2d_bwd_fp32.cpp
rename to example/49_maxpool2d_bwd/maxpool2d_bwd_fp32.cpp
diff --git a/composable_kernel/example/50_put_element/CMakeLists.txt b/example/50_put_element/CMakeLists.txt
similarity index 100%
rename from composable_kernel/example/50_put_element/CMakeLists.txt
rename to example/50_put_element/CMakeLists.txt
diff --git a/composable_kernel/example/50_put_element/put_element_fp16.cpp b/example/50_put_element/put_element_fp16.cpp
similarity index 100%
rename from composable_kernel/example/50_put_element/put_element_fp16.cpp
rename to example/50_put_element/put_element_fp16.cpp
diff --git a/composable_kernel/example/51_avgpool3d_bwd/CMakeLists.txt b/example/51_avgpool3d_bwd/CMakeLists.txt
similarity index 100%
rename from composable_kernel/example/51_avgpool3d_bwd/CMakeLists.txt
rename to example/51_avgpool3d_bwd/CMakeLists.txt
diff --git a/composable_kernel/example/51_avgpool3d_bwd/avgpool3d_bwd_bf16.cpp b/example/51_avgpool3d_bwd/avgpool3d_bwd_bf16.cpp
similarity index 100%
rename from composable_kernel/example/51_avgpool3d_bwd/avgpool3d_bwd_bf16.cpp
rename to example/51_avgpool3d_bwd/avgpool3d_bwd_bf16.cpp
diff --git a/composable_kernel/example/51_avgpool3d_bwd/avgpool3d_bwd_common.hpp b/example/51_avgpool3d_bwd/avgpool3d_bwd_common.hpp
similarity index 100%
rename from composable_kernel/example/51_avgpool3d_bwd/avgpool3d_bwd_common.hpp
rename to example/51_avgpool3d_bwd/avgpool3d_bwd_common.hpp
diff --git a/composable_kernel/example/51_avgpool3d_bwd/avgpool3d_bwd_fp16.cpp b/example/51_avgpool3d_bwd/avgpool3d_bwd_fp16.cpp
similarity index 100%
rename from composable_kernel/example/51_avgpool3d_bwd/avgpool3d_bwd_fp16.cpp
rename to example/51_avgpool3d_bwd/avgpool3d_bwd_fp16.cpp
diff --git a/composable_kernel/example/51_avgpool3d_bwd/avgpool3d_bwd_fp32.cpp b/example/51_avgpool3d_bwd/avgpool3d_bwd_fp32.cpp
similarity index 100%
rename from composable_kernel/example/51_avgpool3d_bwd/avgpool3d_bwd_fp32.cpp
rename to example/51_avgpool3d_bwd/avgpool3d_bwd_fp32.cpp
diff --git a/composable_kernel/example/52_im2col_col2im/CMakeLists.txt b/example/52_im2col_col2im/CMakeLists.txt
similarity index 100%
rename from composable_kernel/example/52_im2col_col2im/CMakeLists.txt
rename to example/52_im2col_col2im/CMakeLists.txt
diff --git a/composable_kernel/example/52_im2col_col2im/column_to_image_f32.cpp b/example/52_im2col_col2im/column_to_image_f32.cpp
similarity index 100%
rename from composable_kernel/example/52_im2col_col2im/column_to_image_f32.cpp
rename to example/52_im2col_col2im/column_to_image_f32.cpp
diff --git a/composable_kernel/example/52_im2col_col2im/common.hpp b/example/52_im2col_col2im/common.hpp
similarity index 100%
rename from composable_kernel/example/52_im2col_col2im/common.hpp
rename to example/52_im2col_col2im/common.hpp
diff --git a/composable_kernel/example/52_im2col_col2im/image_to_column_f32.cpp b/example/52_im2col_col2im/image_to_column_f32.cpp
similarity index 100%
rename from composable_kernel/example/52_im2col_col2im/image_to_column_f32.cpp
rename to example/52_im2col_col2im/image_to_column_f32.cpp
diff --git a/composable_kernel/example/53_gemv_splitk/CMakeLists.txt b/example/53_gemv_splitk/CMakeLists.txt
similarity index 100%
rename from composable_kernel/example/53_gemv_splitk/CMakeLists.txt
rename to example/53_gemv_splitk/CMakeLists.txt
diff --git a/composable_kernel/example/53_gemv_splitk/README.md b/example/53_gemv_splitk/README.md
similarity index 100%
rename from composable_kernel/example/53_gemv_splitk/README.md
rename to example/53_gemv_splitk/README.md
diff --git a/composable_kernel/example/53_gemv_splitk/common.hpp b/example/53_gemv_splitk/common.hpp
similarity index 100%
rename from composable_kernel/example/53_gemv_splitk/common.hpp
rename to example/53_gemv_splitk/common.hpp
diff --git a/composable_kernel/example/53_gemv_splitk/gemv_splitk_fp16.cpp b/example/53_gemv_splitk/gemv_splitk_fp16.cpp
similarity index 100%
rename from composable_kernel/example/53_gemv_splitk/gemv_splitk_fp16.cpp
rename to example/53_gemv_splitk/gemv_splitk_fp16.cpp
diff --git a/composable_kernel/example/53_gemv_splitk/run_gemv_splitk_example.inc b/example/53_gemv_splitk/run_gemv_splitk_example.inc
similarity index 100%
rename from composable_kernel/example/53_gemv_splitk/run_gemv_splitk_example.inc
rename to example/53_gemv_splitk/run_gemv_splitk_example.inc
diff --git a/composable_kernel/example/54_tall_and_skinny_gemm_splitk/CMakeLists.txt b/example/54_tall_and_skinny_gemm_splitk/CMakeLists.txt
similarity index 100%
rename from composable_kernel/example/54_tall_and_skinny_gemm_splitk/CMakeLists.txt
rename to example/54_tall_and_skinny_gemm_splitk/CMakeLists.txt
diff --git a/composable_kernel/example/54_tall_and_skinny_gemm_splitk/README.md b/example/54_tall_and_skinny_gemm_splitk/README.md
similarity index 100%
rename from composable_kernel/example/54_tall_and_skinny_gemm_splitk/README.md
rename to example/54_tall_and_skinny_gemm_splitk/README.md
diff --git a/composable_kernel/example/54_tall_and_skinny_gemm_splitk/common.hpp b/example/54_tall_and_skinny_gemm_splitk/common.hpp
similarity index 100%
rename from composable_kernel/example/54_tall_and_skinny_gemm_splitk/common.hpp
rename to example/54_tall_and_skinny_gemm_splitk/common.hpp
diff --git a/composable_kernel/example/54_tall_and_skinny_gemm_splitk/run_tall_and_skinny_gemm_splitk_example.inc b/example/54_tall_and_skinny_gemm_splitk/run_tall_and_skinny_gemm_splitk_example.inc
similarity index 100%
rename from composable_kernel/example/54_tall_and_skinny_gemm_splitk/run_tall_and_skinny_gemm_splitk_example.inc
rename to example/54_tall_and_skinny_gemm_splitk/run_tall_and_skinny_gemm_splitk_example.inc
diff --git a/composable_kernel/example/54_tall_and_skinny_gemm_splitk/tall_and_skinny_gemm_splitk_fp16.cpp b/example/54_tall_and_skinny_gemm_splitk/tall_and_skinny_gemm_splitk_fp16.cpp
similarity index 100%
rename from composable_kernel/example/54_tall_and_skinny_gemm_splitk/tall_and_skinny_gemm_splitk_fp16.cpp
rename to example/54_tall_and_skinny_gemm_splitk/tall_and_skinny_gemm_splitk_fp16.cpp
diff --git a/composable_kernel/example/60_gemm_multi_ABD/CMakeLists.txt b/example/60_gemm_multi_ABD/CMakeLists.txt
similarity index 100%
rename from composable_kernel/example/60_gemm_multi_ABD/CMakeLists.txt
rename to example/60_gemm_multi_ABD/CMakeLists.txt
diff --git a/composable_kernel/example/60_gemm_multi_ABD/gemm_multi_ABD_xdl_fp16.cpp b/example/60_gemm_multi_ABD/gemm_multi_ABD_xdl_fp16.cpp
similarity index 100%
rename from composable_kernel/example/60_gemm_multi_ABD/gemm_multi_ABD_xdl_fp16.cpp
rename to example/60_gemm_multi_ABD/gemm_multi_ABD_xdl_fp16.cpp
diff --git a/composable_kernel/example/61_contraction_multi_ABD/CMakeLists.txt b/example/61_contraction_multi_ABD/CMakeLists.txt
similarity index 100%
rename from composable_kernel/example/61_contraction_multi_ABD/CMakeLists.txt
rename to example/61_contraction_multi_ABD/CMakeLists.txt
diff --git a/composable_kernel/example/61_contraction_multi_ABD/contraction_multi_ABD_xdl_fp16.cpp b/example/61_contraction_multi_ABD/contraction_multi_ABD_xdl_fp16.cpp
similarity index 100%
rename from composable_kernel/example/61_contraction_multi_ABD/contraction_multi_ABD_xdl_fp16.cpp
rename to example/61_contraction_multi_ABD/contraction_multi_ABD_xdl_fp16.cpp
diff --git a/composable_kernel/example/62_conv_fwd_activ/CMakeLists.txt b/example/62_conv_fwd_activ/CMakeLists.txt
similarity index 100%
rename from composable_kernel/example/62_conv_fwd_activ/CMakeLists.txt
rename to example/62_conv_fwd_activ/CMakeLists.txt
diff --git a/composable_kernel/example/62_conv_fwd_activ/convnd_fwd_activ_common.hpp b/example/62_conv_fwd_activ/convnd_fwd_activ_common.hpp
similarity index 100%
rename from composable_kernel/example/62_conv_fwd_activ/convnd_fwd_activ_common.hpp
rename to example/62_conv_fwd_activ/convnd_fwd_activ_common.hpp
diff --git a/composable_kernel/example/62_conv_fwd_activ/convnd_fwd_xdl_abs_fp16.cpp b/example/62_conv_fwd_activ/convnd_fwd_xdl_abs_fp16.cpp
similarity index 100%
rename from composable_kernel/example/62_conv_fwd_activ/convnd_fwd_xdl_abs_fp16.cpp
rename to example/62_conv_fwd_activ/convnd_fwd_xdl_abs_fp16.cpp
diff --git a/composable_kernel/example/62_conv_fwd_activ/convnd_fwd_xdl_clippedrelu_fp16.cpp b/example/62_conv_fwd_activ/convnd_fwd_xdl_clippedrelu_fp16.cpp
similarity index 100%
rename from composable_kernel/example/62_conv_fwd_activ/convnd_fwd_xdl_clippedrelu_fp16.cpp
rename to example/62_conv_fwd_activ/convnd_fwd_xdl_clippedrelu_fp16.cpp
diff --git a/composable_kernel/example/62_conv_fwd_activ/convnd_fwd_xdl_elu_fp16.cpp b/example/62_conv_fwd_activ/convnd_fwd_xdl_elu_fp16.cpp
similarity index 100%
rename from composable_kernel/example/62_conv_fwd_activ/convnd_fwd_xdl_elu_fp16.cpp
rename to example/62_conv_fwd_activ/convnd_fwd_xdl_elu_fp16.cpp
diff --git a/composable_kernel/example/62_conv_fwd_activ/convnd_fwd_xdl_leakyrelu_fp16.cpp b/example/62_conv_fwd_activ/convnd_fwd_xdl_leakyrelu_fp16.cpp
similarity index 100%
rename from composable_kernel/example/62_conv_fwd_activ/convnd_fwd_xdl_leakyrelu_fp16.cpp
rename to example/62_conv_fwd_activ/convnd_fwd_xdl_leakyrelu_fp16.cpp
diff --git a/composable_kernel/example/62_conv_fwd_activ/convnd_fwd_xdl_pow_fp16.cpp b/example/62_conv_fwd_activ/convnd_fwd_xdl_pow_fp16.cpp
similarity index 100%
rename from composable_kernel/example/62_conv_fwd_activ/convnd_fwd_xdl_pow_fp16.cpp
rename to example/62_conv_fwd_activ/convnd_fwd_xdl_pow_fp16.cpp
diff --git a/composable_kernel/example/62_conv_fwd_activ/convnd_fwd_xdl_relu_fp16.cpp b/example/62_conv_fwd_activ/convnd_fwd_xdl_relu_fp16.cpp
similarity index 100%
rename from composable_kernel/example/62_conv_fwd_activ/convnd_fwd_xdl_relu_fp16.cpp
rename to example/62_conv_fwd_activ/convnd_fwd_xdl_relu_fp16.cpp
diff --git a/composable_kernel/example/62_conv_fwd_activ/convnd_fwd_xdl_sigmoid_fp16.cpp b/example/62_conv_fwd_activ/convnd_fwd_xdl_sigmoid_fp16.cpp
similarity index 100%
rename from composable_kernel/example/62_conv_fwd_activ/convnd_fwd_xdl_sigmoid_fp16.cpp
rename to example/62_conv_fwd_activ/convnd_fwd_xdl_sigmoid_fp16.cpp
diff --git a/composable_kernel/example/62_conv_fwd_activ/convnd_fwd_xdl_softrelu_fp16.cpp b/example/62_conv_fwd_activ/convnd_fwd_xdl_softrelu_fp16.cpp
similarity index 100%
rename from composable_kernel/example/62_conv_fwd_activ/convnd_fwd_xdl_softrelu_fp16.cpp
rename to example/62_conv_fwd_activ/convnd_fwd_xdl_softrelu_fp16.cpp
diff --git a/composable_kernel/example/62_conv_fwd_activ/convnd_fwd_xdl_tanh_fp16.cpp b/example/62_conv_fwd_activ/convnd_fwd_xdl_tanh_fp16.cpp
similarity index 100%
rename from composable_kernel/example/62_conv_fwd_activ/convnd_fwd_xdl_tanh_fp16.cpp
rename to example/62_conv_fwd_activ/convnd_fwd_xdl_tanh_fp16.cpp
diff --git a/composable_kernel/example/62_conv_fwd_activ/run_convnd_fwd_activ_example.inc b/example/62_conv_fwd_activ/run_convnd_fwd_activ_example.inc
similarity index 100%
rename from composable_kernel/example/62_conv_fwd_activ/run_convnd_fwd_activ_example.inc
rename to example/62_conv_fwd_activ/run_convnd_fwd_activ_example.inc
diff --git a/composable_kernel/example/CMakeLists.txt b/example/CMakeLists.txt
similarity index 100%
rename from composable_kernel/example/CMakeLists.txt
rename to example/CMakeLists.txt
diff --git a/composable_kernel/include/ck/ck.hpp b/include/ck/ck.hpp
similarity index 100%
rename from composable_kernel/include/ck/ck.hpp
rename to include/ck/ck.hpp
diff --git a/composable_kernel/include/ck/config.h.in b/include/ck/config.h.in
similarity index 100%
rename from composable_kernel/include/ck/config.h.in
rename to include/ck/config.h.in
diff --git a/composable_kernel/include/ck/host_utility/device_prop.hpp b/include/ck/host_utility/device_prop.hpp
similarity index 100%
rename from composable_kernel/include/ck/host_utility/device_prop.hpp
rename to include/ck/host_utility/device_prop.hpp
diff --git a/composable_kernel/include/ck/host_utility/hip_check_error.hpp b/include/ck/host_utility/hip_check_error.hpp
similarity index 100%
rename from composable_kernel/include/ck/host_utility/hip_check_error.hpp
rename to include/ck/host_utility/hip_check_error.hpp
diff --git a/composable_kernel/include/ck/host_utility/io.hpp b/include/ck/host_utility/io.hpp
similarity index 100%
rename from composable_kernel/include/ck/host_utility/io.hpp
rename to include/ck/host_utility/io.hpp
diff --git a/composable_kernel/include/ck/host_utility/kernel_launch.hpp b/include/ck/host_utility/kernel_launch.hpp
similarity index 100%
rename from composable_kernel/include/ck/host_utility/kernel_launch.hpp
rename to include/ck/host_utility/kernel_launch.hpp
diff --git a/composable_kernel/include/ck/host_utility/stream_utility.hpp b/include/ck/host_utility/stream_utility.hpp
similarity index 100%
rename from composable_kernel/include/ck/host_utility/stream_utility.hpp
rename to include/ck/host_utility/stream_utility.hpp
diff --git a/composable_kernel/include/ck/problem_transform/transform_forward_convolution3d_into_gemm_v4r4r4_ndhwc_kzyxc_ndhwk.hpp b/include/ck/problem_transform/transform_forward_convolution3d_into_gemm_v4r4r4_ndhwc_kzyxc_ndhwk.hpp
similarity index 100%
rename from composable_kernel/include/ck/problem_transform/transform_forward_convolution3d_into_gemm_v4r4r4_ndhwc_kzyxc_ndhwk.hpp
rename to include/ck/problem_transform/transform_forward_convolution3d_into_gemm_v4r4r4_ndhwc_kzyxc_ndhwk.hpp
diff --git a/composable_kernel/include/ck/stream_config.hpp b/include/ck/stream_config.hpp
similarity index 100%
rename from composable_kernel/include/ck/stream_config.hpp
rename to include/ck/stream_config.hpp
diff --git a/composable_kernel/include/ck/tensor/static_tensor.hpp b/include/ck/tensor/static_tensor.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor/static_tensor.hpp
rename to include/ck/tensor/static_tensor.hpp
diff --git a/composable_kernel/include/ck/tensor_description/cluster_descriptor.hpp b/include/ck/tensor_description/cluster_descriptor.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_description/cluster_descriptor.hpp
rename to include/ck/tensor_description/cluster_descriptor.hpp
diff --git a/composable_kernel/include/ck/tensor_description/multi_index_transform.hpp b/include/ck/tensor_description/multi_index_transform.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_description/multi_index_transform.hpp
rename to include/ck/tensor_description/multi_index_transform.hpp
diff --git a/composable_kernel/include/ck/tensor_description/multi_index_transform_helper.hpp b/include/ck/tensor_description/multi_index_transform_helper.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_description/multi_index_transform_helper.hpp
rename to include/ck/tensor_description/multi_index_transform_helper.hpp
diff --git a/composable_kernel/include/ck/tensor_description/tensor_adaptor.hpp b/include/ck/tensor_description/tensor_adaptor.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_description/tensor_adaptor.hpp
rename to include/ck/tensor_description/tensor_adaptor.hpp
diff --git a/composable_kernel/include/ck/tensor_description/tensor_descriptor.hpp b/include/ck/tensor_description/tensor_descriptor.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_description/tensor_descriptor.hpp
rename to include/ck/tensor_description/tensor_descriptor.hpp
diff --git a/composable_kernel/include/ck/tensor_description/tensor_descriptor_helper.hpp b/include/ck/tensor_description/tensor_descriptor_helper.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_description/tensor_descriptor_helper.hpp
rename to include/ck/tensor_description/tensor_descriptor_helper.hpp
diff --git a/composable_kernel/include/ck/tensor_description/tensor_space_filling_curve.hpp b/include/ck/tensor_description/tensor_space_filling_curve.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_description/tensor_space_filling_curve.hpp
rename to include/ck/tensor_description/tensor_space_filling_curve.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/block/blockwise_gemm_dl_v2r3.hpp b/include/ck/tensor_operation/gpu/block/blockwise_gemm_dl_v2r3.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/block/blockwise_gemm_dl_v2r3.hpp
rename to include/ck/tensor_operation/gpu/block/blockwise_gemm_dl_v2r3.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/block/blockwise_gemm_dlops_v2r2.hpp b/include/ck/tensor_operation/gpu/block/blockwise_gemm_dlops_v2r2.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/block/blockwise_gemm_dlops_v2r2.hpp
rename to include/ck/tensor_operation/gpu/block/blockwise_gemm_dlops_v2r2.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/block/blockwise_gemm_dlops_v3.hpp b/include/ck/tensor_operation/gpu/block/blockwise_gemm_dlops_v3.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/block/blockwise_gemm_dlops_v3.hpp
rename to include/ck/tensor_operation/gpu/block/blockwise_gemm_dlops_v3.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/block/blockwise_gemm_dpp.hpp b/include/ck/tensor_operation/gpu/block/blockwise_gemm_dpp.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/block/blockwise_gemm_dpp.hpp
rename to include/ck/tensor_operation/gpu/block/blockwise_gemm_dpp.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/block/blockwise_gemm_wmma.hpp b/include/ck/tensor_operation/gpu/block/blockwise_gemm_wmma.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/block/blockwise_gemm_wmma.hpp
rename to include/ck/tensor_operation/gpu/block/blockwise_gemm_wmma.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp b/include/ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp
rename to include/ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/block/blockwise_gemm_xdlops_skip_b_lds.hpp b/include/ck/tensor_operation/gpu/block/blockwise_gemm_xdlops_skip_b_lds.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/block/blockwise_gemm_xdlops_skip_b_lds.hpp
rename to include/ck/tensor_operation/gpu/block/blockwise_gemm_xdlops_skip_b_lds.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/block/blockwise_softmax.hpp b/include/ck/tensor_operation/gpu/block/blockwise_softmax.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/block/blockwise_softmax.hpp
rename to include/ck/tensor_operation/gpu/block/blockwise_softmax.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/block/blockwise_tensor_slice_transfer_v5r1.hpp b/include/ck/tensor_operation/gpu/block/blockwise_tensor_slice_transfer_v5r1.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/block/blockwise_tensor_slice_transfer_v5r1.hpp
rename to include/ck/tensor_operation/gpu/block/blockwise_tensor_slice_transfer_v5r1.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/block/blockwise_welford.hpp b/include/ck/tensor_operation/gpu/block/blockwise_welford.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/block/blockwise_welford.hpp
rename to include/ck/tensor_operation/gpu/block/blockwise_welford.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/block/reduction_functions_blockwise.hpp b/include/ck/tensor_operation/gpu/block/reduction_functions_blockwise.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/block/reduction_functions_blockwise.hpp
rename to include/ck/tensor_operation/gpu/block/reduction_functions_blockwise.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp b/include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp
rename to include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r1.hpp b/include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r1.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r1.hpp
rename to include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r1.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r1r2.hpp b/include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r1r2.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r1r2.hpp
rename to include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r1r2.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r2.hpp b/include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r2.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r2.hpp
rename to include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r2.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r3.hpp b/include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r3.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r3.hpp
rename to include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r3.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v7.hpp b/include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v7.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v7.hpp
rename to include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v7.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v7r2.hpp b/include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v7r2.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v7r2.hpp
rename to include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v7r2.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/conv_tensor_rearrange_op.hpp b/include/ck/tensor_operation/gpu/device/conv_tensor_rearrange_op.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/conv_tensor_rearrange_op.hpp
rename to include/ck/tensor_operation/gpu/device/conv_tensor_rearrange_op.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/convolution_backward_data_specialization.hpp b/include/ck/tensor_operation/gpu/device/convolution_backward_data_specialization.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/convolution_backward_data_specialization.hpp
rename to include/ck/tensor_operation/gpu/device/convolution_backward_data_specialization.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/convolution_backward_weight_specialization.hpp b/include/ck/tensor_operation/gpu/device/convolution_backward_weight_specialization.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/convolution_backward_weight_specialization.hpp
rename to include/ck/tensor_operation/gpu/device/convolution_backward_weight_specialization.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp b/include/ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp
rename to include/ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/device_avgpool_bwd.hpp b/include/ck/tensor_operation/gpu/device/device_avgpool_bwd.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/device_avgpool_bwd.hpp
rename to include/ck/tensor_operation/gpu/device/device_avgpool_bwd.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/device_base.hpp b/include/ck/tensor_operation/gpu/device/device_base.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/device_base.hpp
rename to include/ck/tensor_operation/gpu/device/device_base.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/device_batched_contraction_multiple_d.hpp b/include/ck/tensor_operation/gpu/device/device_batched_contraction_multiple_d.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/device_batched_contraction_multiple_d.hpp
rename to include/ck/tensor_operation/gpu/device/device_batched_contraction_multiple_d.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/device_batched_gemm.hpp b/include/ck/tensor_operation/gpu/device/device_batched_gemm.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/device_batched_gemm.hpp
rename to include/ck/tensor_operation/gpu/device/device_batched_gemm.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/device_batched_gemm_e_permute.hpp b/include/ck/tensor_operation/gpu/device/device_batched_gemm_e_permute.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/device_batched_gemm_e_permute.hpp
rename to include/ck/tensor_operation/gpu/device/device_batched_gemm_e_permute.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/device_batched_gemm_gemm.hpp b/include/ck/tensor_operation/gpu/device/device_batched_gemm_gemm.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/device_batched_gemm_gemm.hpp
rename to include/ck/tensor_operation/gpu/device/device_batched_gemm_gemm.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/device_batched_gemm_multi_d.hpp b/include/ck/tensor_operation/gpu/device/device_batched_gemm_multi_d.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/device_batched_gemm_multi_d.hpp
rename to include/ck/tensor_operation/gpu/device/device_batched_gemm_multi_d.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/device_batched_gemm_multiple_d_gemm_multiple_d.hpp b/include/ck/tensor_operation/gpu/device/device_batched_gemm_multiple_d_gemm_multiple_d.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/device_batched_gemm_multiple_d_gemm_multiple_d.hpp
rename to include/ck/tensor_operation/gpu/device/device_batched_gemm_multiple_d_gemm_multiple_d.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/device_batched_gemm_softmax_gemm.hpp b/include/ck/tensor_operation/gpu/device/device_batched_gemm_softmax_gemm.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/device_batched_gemm_softmax_gemm.hpp
rename to include/ck/tensor_operation/gpu/device/device_batched_gemm_softmax_gemm.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/device_batched_gemm_softmax_gemm_permute.hpp b/include/ck/tensor_operation/gpu/device/device_batched_gemm_softmax_gemm_permute.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/device_batched_gemm_softmax_gemm_permute.hpp
rename to include/ck/tensor_operation/gpu/device/device_batched_gemm_softmax_gemm_permute.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/device_batchnorm_backward.hpp b/include/ck/tensor_operation/gpu/device/device_batchnorm_backward.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/device_batchnorm_backward.hpp
rename to include/ck/tensor_operation/gpu/device/device_batchnorm_backward.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/device_batchnorm_forward.hpp b/include/ck/tensor_operation/gpu/device/device_batchnorm_forward.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/device_batchnorm_forward.hpp
rename to include/ck/tensor_operation/gpu/device/device_batchnorm_forward.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/device_batchnorm_infer.hpp b/include/ck/tensor_operation/gpu/device/device_batchnorm_infer.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/device_batchnorm_infer.hpp
rename to include/ck/tensor_operation/gpu/device/device_batchnorm_infer.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/device_cgemm.hpp b/include/ck/tensor_operation/gpu/device/device_cgemm.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/device_cgemm.hpp
rename to include/ck/tensor_operation/gpu/device/device_cgemm.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/device_contraction_multiple_abd.hpp b/include/ck/tensor_operation/gpu/device/device_contraction_multiple_abd.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/device_contraction_multiple_abd.hpp
rename to include/ck/tensor_operation/gpu/device/device_contraction_multiple_abd.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/device_contraction_multiple_d.hpp b/include/ck/tensor_operation/gpu/device/device_contraction_multiple_d.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/device_contraction_multiple_d.hpp
rename to include/ck/tensor_operation/gpu/device/device_contraction_multiple_d.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/device_conv_bwd_data.hpp b/include/ck/tensor_operation/gpu/device/device_conv_bwd_data.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/device_conv_bwd_data.hpp
rename to include/ck/tensor_operation/gpu/device/device_conv_bwd_data.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/device_conv_fwd.hpp b/include/ck/tensor_operation/gpu/device/device_conv_fwd.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/device_conv_fwd.hpp
rename to include/ck/tensor_operation/gpu/device/device_conv_fwd.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/device_conv_fwd_bias_activation.hpp b/include/ck/tensor_operation/gpu/device/device_conv_fwd_bias_activation.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/device_conv_fwd_bias_activation.hpp
rename to include/ck/tensor_operation/gpu/device/device_conv_fwd_bias_activation.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/device_conv_fwd_bias_activation_add.hpp b/include/ck/tensor_operation/gpu/device/device_conv_fwd_bias_activation_add.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/device_conv_fwd_bias_activation_add.hpp
rename to include/ck/tensor_operation/gpu/device/device_conv_fwd_bias_activation_add.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/device_conv_tensor_rearrange.hpp b/include/ck/tensor_operation/gpu/device/device_conv_tensor_rearrange.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/device_conv_tensor_rearrange.hpp
rename to include/ck/tensor_operation/gpu/device/device_conv_tensor_rearrange.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/device_elementwise.hpp b/include/ck/tensor_operation/gpu/device/device_elementwise.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/device_elementwise.hpp
rename to include/ck/tensor_operation/gpu/device/device_elementwise.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/device_elementwise_normalization.hpp b/include/ck/tensor_operation/gpu/device/device_elementwise_normalization.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/device_elementwise_normalization.hpp
rename to include/ck/tensor_operation/gpu/device/device_elementwise_normalization.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/device_gemm.hpp b/include/ck/tensor_operation/gpu/device/device_gemm.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/device_gemm.hpp
rename to include/ck/tensor_operation/gpu/device/device_gemm.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/device_gemm_bias_e_permute.hpp b/include/ck/tensor_operation/gpu/device/device_gemm_bias_e_permute.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/device_gemm_bias_e_permute.hpp
rename to include/ck/tensor_operation/gpu/device/device_gemm_bias_e_permute.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/device_gemm_multiple_abd.hpp b/include/ck/tensor_operation/gpu/device/device_gemm_multiple_abd.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/device_gemm_multiple_abd.hpp
rename to include/ck/tensor_operation/gpu/device/device_gemm_multiple_abd.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp b/include/ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp
rename to include/ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/device_gemm_multiple_d_layernorm.hpp b/include/ck/tensor_operation/gpu/device/device_gemm_multiple_d_layernorm.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/device_gemm_multiple_d_layernorm.hpp
rename to include/ck/tensor_operation/gpu/device/device_gemm_multiple_d_layernorm.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/device_gemm_multiple_d_multiple_r.hpp b/include/ck/tensor_operation/gpu/device/device_gemm_multiple_d_multiple_r.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/device_gemm_multiple_d_multiple_r.hpp
rename to include/ck/tensor_operation/gpu/device/device_gemm_multiple_d_multiple_r.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/device_gemm_reduce.hpp b/include/ck/tensor_operation/gpu/device/device_gemm_reduce.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/device_gemm_reduce.hpp
rename to include/ck/tensor_operation/gpu/device/device_gemm_reduce.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/device_gemm_splitk.hpp b/include/ck/tensor_operation/gpu/device/device_gemm_splitk.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/device_gemm_splitk.hpp
rename to include/ck/tensor_operation/gpu/device/device_gemm_splitk.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/device_gemm_streamk.hpp b/include/ck/tensor_operation/gpu/device/device_gemm_streamk.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/device_gemm_streamk.hpp
rename to include/ck/tensor_operation/gpu/device/device_gemm_streamk.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/device_grouped_contraction_multiple_d.hpp b/include/ck/tensor_operation/gpu/device/device_grouped_contraction_multiple_d.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/device_grouped_contraction_multiple_d.hpp
rename to include/ck/tensor_operation/gpu/device/device_grouped_contraction_multiple_d.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/device_grouped_conv_bwd_data_multiple_d.hpp b/include/ck/tensor_operation/gpu/device/device_grouped_conv_bwd_data_multiple_d.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/device_grouped_conv_bwd_data_multiple_d.hpp
rename to include/ck/tensor_operation/gpu/device/device_grouped_conv_bwd_data_multiple_d.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/device_grouped_conv_bwd_weight.hpp b/include/ck/tensor_operation/gpu/device/device_grouped_conv_bwd_weight.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/device_grouped_conv_bwd_weight.hpp
rename to include/ck/tensor_operation/gpu/device/device_grouped_conv_bwd_weight.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/device_grouped_conv_fwd.hpp b/include/ck/tensor_operation/gpu/device/device_grouped_conv_fwd.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/device_grouped_conv_fwd.hpp
rename to include/ck/tensor_operation/gpu/device/device_grouped_conv_fwd.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_d.hpp b/include/ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_d.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_d.hpp
rename to include/ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_d.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/device_grouped_gemm.hpp b/include/ck/tensor_operation/gpu/device/device_grouped_gemm.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/device_grouped_gemm.hpp
rename to include/ck/tensor_operation/gpu/device/device_grouped_gemm.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/device_grouped_gemm_fixed_nk.hpp b/include/ck/tensor_operation/gpu/device/device_grouped_gemm_fixed_nk.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/device_grouped_gemm_fixed_nk.hpp
rename to include/ck/tensor_operation/gpu/device/device_grouped_gemm_fixed_nk.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/device_grouped_gemm_softmax_gemm_permute.hpp b/include/ck/tensor_operation/gpu/device/device_grouped_gemm_softmax_gemm_permute.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/device_grouped_gemm_softmax_gemm_permute.hpp
rename to include/ck/tensor_operation/gpu/device/device_grouped_gemm_softmax_gemm_permute.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/device_grouped_gemm_splitk.hpp b/include/ck/tensor_operation/gpu/device/device_grouped_gemm_splitk.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/device_grouped_gemm_splitk.hpp
rename to include/ck/tensor_operation/gpu/device/device_grouped_gemm_splitk.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/device_max_pool_bwd.hpp b/include/ck/tensor_operation/gpu/device/device_max_pool_bwd.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/device_max_pool_bwd.hpp
rename to include/ck/tensor_operation/gpu/device/device_max_pool_bwd.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/device_multiple_reduce.hpp b/include/ck/tensor_operation/gpu/device/device_multiple_reduce.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/device_multiple_reduce.hpp
rename to include/ck/tensor_operation/gpu/device/device_multiple_reduce.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/device_normalization.hpp b/include/ck/tensor_operation/gpu/device/device_normalization.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/device_normalization.hpp
rename to include/ck/tensor_operation/gpu/device/device_normalization.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/device_permute.hpp b/include/ck/tensor_operation/gpu/device/device_permute.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/device_permute.hpp
rename to include/ck/tensor_operation/gpu/device/device_permute.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/device_pool_fwd.hpp b/include/ck/tensor_operation/gpu/device/device_pool_fwd.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/device_pool_fwd.hpp
rename to include/ck/tensor_operation/gpu/device/device_pool_fwd.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/device_put_element.hpp b/include/ck/tensor_operation/gpu/device/device_put_element.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/device_put_element.hpp
rename to include/ck/tensor_operation/gpu/device/device_put_element.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/device_reduce.hpp b/include/ck/tensor_operation/gpu/device/device_reduce.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/device_reduce.hpp
rename to include/ck/tensor_operation/gpu/device/device_reduce.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/device_softmax.hpp b/include/ck/tensor_operation/gpu/device/device_softmax.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/device_softmax.hpp
rename to include/ck/tensor_operation/gpu/device/device_softmax.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/device_splitk_contraction_multiple_d.hpp b/include/ck/tensor_operation/gpu/device/device_splitk_contraction_multiple_d.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/device_splitk_contraction_multiple_d.hpp
rename to include/ck/tensor_operation/gpu/device/device_splitk_contraction_multiple_d.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/device_tall_and_skinny_gemm.hpp b/include/ck/tensor_operation/gpu/device/device_tall_and_skinny_gemm.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/device_tall_and_skinny_gemm.hpp
rename to include/ck/tensor_operation/gpu/device/device_tall_and_skinny_gemm.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/gemm_specialization.hpp b/include/ck/tensor_operation/gpu/device/gemm_specialization.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/gemm_specialization.hpp
rename to include/ck/tensor_operation/gpu/device/gemm_specialization.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_avgpool3d_bwd_ndhwc_ndhwc.hpp b/include/ck/tensor_operation/gpu/device/impl/device_avgpool3d_bwd_ndhwc_ndhwc.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_avgpool3d_bwd_ndhwc_ndhwc.hpp
rename to include/ck/tensor_operation/gpu/device/impl/device_avgpool3d_bwd_ndhwc_ndhwc.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_batched_contraction_multiple_d_wmma_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_batched_contraction_multiple_d_wmma_cshuffle.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_batched_contraction_multiple_d_wmma_cshuffle.hpp
rename to include/ck/tensor_operation/gpu/device/impl/device_batched_contraction_multiple_d_wmma_cshuffle.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_batched_contraction_multiple_d_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_batched_contraction_multiple_d_xdl_cshuffle.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_batched_contraction_multiple_d_xdl_cshuffle.hpp
rename to include/ck/tensor_operation/gpu/device/impl/device_batched_contraction_multiple_d_xdl_cshuffle.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_e_permute_xdl.hpp b/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_e_permute_xdl.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_e_permute_xdl.hpp
rename to include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_e_permute_xdl.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_gemm_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_gemm_xdl_cshuffle.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_gemm_xdl_cshuffle.hpp
rename to include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_gemm_xdl_cshuffle.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_multi_d_xdl.hpp b/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_multi_d_xdl.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_multi_d_xdl.hpp
rename to include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_multi_d_xdl.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_dl.hpp b/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_dl.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_dl.hpp
rename to include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_dl.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_gemm_multiple_d_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_gemm_multiple_d_xdl_cshuffle.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_gemm_multiple_d_xdl_cshuffle.hpp
rename to include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_gemm_multiple_d_xdl_cshuffle.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_reduce_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_reduce_xdl_cshuffle.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_reduce_xdl_cshuffle.hpp
rename to include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_reduce_xdl_cshuffle.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle.hpp
rename to include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_xdl_cshuffle.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_xdl_cshuffle.hpp
rename to include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_xdl_cshuffle.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_xdl.hpp b/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_xdl.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_xdl.hpp
rename to include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_xdl.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_batchnorm_backward_impl.hpp b/include/ck/tensor_operation/gpu/device/impl/device_batchnorm_backward_impl.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_batchnorm_backward_impl.hpp
rename to include/ck/tensor_operation/gpu/device/impl/device_batchnorm_backward_impl.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_batchnorm_forward_impl.hpp b/include/ck/tensor_operation/gpu/device/impl/device_batchnorm_forward_impl.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_batchnorm_forward_impl.hpp
rename to include/ck/tensor_operation/gpu/device/impl/device_batchnorm_forward_impl.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_batchnorm_forward_impl_obsolete.hpp b/include/ck/tensor_operation/gpu/device/impl/device_batchnorm_forward_impl_obsolete.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_batchnorm_forward_impl_obsolete.hpp
rename to include/ck/tensor_operation/gpu/device/impl/device_batchnorm_forward_impl_obsolete.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_cgemm_4gemm_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_cgemm_4gemm_xdl_cshuffle.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_cgemm_4gemm_xdl_cshuffle.hpp
rename to include/ck/tensor_operation/gpu/device/impl/device_cgemm_4gemm_xdl_cshuffle.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_column_to_image_impl.hpp b/include/ck/tensor_operation/gpu/device/impl/device_column_to_image_impl.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_column_to_image_impl.hpp
rename to include/ck/tensor_operation/gpu/device/impl/device_column_to_image_impl.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_contraction_multiple_abd_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_contraction_multiple_abd_xdl_cshuffle.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_contraction_multiple_abd_xdl_cshuffle.hpp
rename to include/ck/tensor_operation/gpu/device/impl/device_contraction_multiple_abd_xdl_cshuffle.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_contraction_multiple_d_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_contraction_multiple_d_xdl_cshuffle.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_contraction_multiple_d_xdl_cshuffle.hpp
rename to include/ck/tensor_operation/gpu/device/impl/device_contraction_multiple_d_xdl_cshuffle.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_conv2d_backward_weight_xdl_c_shuffle_nhwc_kyxc_nhwk.hpp b/include/ck/tensor_operation/gpu/device/impl/device_conv2d_backward_weight_xdl_c_shuffle_nhwc_kyxc_nhwk.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_conv2d_backward_weight_xdl_c_shuffle_nhwc_kyxc_nhwk.hpp
rename to include/ck/tensor_operation/gpu/device/impl/device_conv2d_backward_weight_xdl_c_shuffle_nhwc_kyxc_nhwk.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk.hpp b/include/ck/tensor_operation/gpu/device/impl/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk.hpp
rename to include/ck/tensor_operation/gpu/device/impl/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_conv2d_fwd_xdl_c_shuffle_bias_activation_add_nhwc_kyxc_nhwk.hpp b/include/ck/tensor_operation/gpu/device/impl/device_conv2d_fwd_xdl_c_shuffle_bias_activation_add_nhwc_kyxc_nhwk.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_conv2d_fwd_xdl_c_shuffle_bias_activation_add_nhwc_kyxc_nhwk.hpp
rename to include/ck/tensor_operation/gpu/device/impl/device_conv2d_fwd_xdl_c_shuffle_bias_activation_add_nhwc_kyxc_nhwk.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_conv2d_fwd_xdl_c_shuffle_bias_activation_nhwc_kyxc_nhwk.hpp b/include/ck/tensor_operation/gpu/device/impl/device_conv2d_fwd_xdl_c_shuffle_bias_activation_nhwc_kyxc_nhwk.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_conv2d_fwd_xdl_c_shuffle_bias_activation_nhwc_kyxc_nhwk.hpp
rename to include/ck/tensor_operation/gpu/device/impl/device_conv2d_fwd_xdl_c_shuffle_bias_activation_nhwc_kyxc_nhwk.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk.hpp b/include/ck/tensor_operation/gpu/device/impl/device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk.hpp
rename to include/ck/tensor_operation/gpu/device/impl/device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk.hpp b/include/ck/tensor_operation/gpu/device/impl/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk.hpp
rename to include/ck/tensor_operation/gpu/device/impl/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_conv3d_fwd_naive_ndhwc_kzyxc_ndhwk.hpp b/include/ck/tensor_operation/gpu/device/impl/device_conv3d_fwd_naive_ndhwc_kzyxc_ndhwk.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_conv3d_fwd_naive_ndhwc_kzyxc_ndhwk.hpp
rename to include/ck/tensor_operation/gpu/device/impl/device_conv3d_fwd_naive_ndhwc_kzyxc_ndhwk.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_conv3d_fwd_xdl_ndhwc_kzyxc_ndhwk.hpp b/include/ck/tensor_operation/gpu/device/impl/device_conv3d_fwd_xdl_ndhwc_kzyxc_ndhwk.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_conv3d_fwd_xdl_ndhwc_kzyxc_ndhwk.hpp
rename to include/ck/tensor_operation/gpu/device/impl/device_conv3d_fwd_xdl_ndhwc_kzyxc_ndhwk.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_convnd_bwd_data_nwc_kxc_nwk_dl.hpp b/include/ck/tensor_operation/gpu/device/impl/device_convnd_bwd_data_nwc_kxc_nwk_dl.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_convnd_bwd_data_nwc_kxc_nwk_dl.hpp
rename to include/ck/tensor_operation/gpu/device/impl/device_convnd_bwd_data_nwc_kxc_nwk_dl.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_convnd_bwd_data_nwc_kxc_nwk_xdl.hpp b/include/ck/tensor_operation/gpu/device/impl/device_convnd_bwd_data_nwc_kxc_nwk_xdl.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_convnd_bwd_data_nwc_kxc_nwk_xdl.hpp
rename to include/ck/tensor_operation/gpu/device/impl/device_convnd_bwd_data_nwc_kxc_nwk_xdl.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_elementwise_2d_impl.hpp b/include/ck/tensor_operation/gpu/device/impl/device_elementwise_2d_impl.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_elementwise_2d_impl.hpp
rename to include/ck/tensor_operation/gpu/device/impl/device_elementwise_2d_impl.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_elementwise_impl.hpp b/include/ck/tensor_operation/gpu/device/impl/device_elementwise_impl.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_elementwise_impl.hpp
rename to include/ck/tensor_operation/gpu/device/impl/device_elementwise_impl.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_elementwise_normalization_impl.hpp b/include/ck/tensor_operation/gpu/device/impl/device_elementwise_normalization_impl.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_elementwise_normalization_impl.hpp
rename to include/ck/tensor_operation/gpu/device/impl/device_elementwise_normalization_impl.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_gemm_bias_add_reduce_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_gemm_bias_add_reduce_xdl_cshuffle.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_gemm_bias_add_reduce_xdl_cshuffle.hpp
rename to include/ck/tensor_operation/gpu/device/impl/device_gemm_bias_add_reduce_xdl_cshuffle.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_gemm_dl.hpp b/include/ck/tensor_operation/gpu/device/impl/device_gemm_dl.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_gemm_dl.hpp
rename to include/ck/tensor_operation/gpu/device/impl/device_gemm_dl.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_gemm_dpp.hpp b/include/ck/tensor_operation/gpu/device/impl/device_gemm_dpp.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_gemm_dpp.hpp
rename to include/ck/tensor_operation/gpu/device/impl/device_gemm_dpp.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_abd_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_abd_xdl_cshuffle.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_abd_xdl_cshuffle.hpp
rename to include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_abd_xdl_cshuffle.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_dl.hpp b/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_dl.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_dl.hpp
rename to include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_dl.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_layernorm_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_layernorm_xdl_cshuffle.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_layernorm_xdl_cshuffle.hpp
rename to include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_layernorm_xdl_cshuffle.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_multiple_r_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_multiple_r_xdl_cshuffle.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_multiple_r_xdl_cshuffle.hpp
rename to include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_multiple_r_xdl_cshuffle.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_wmma_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_wmma_cshuffle.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_wmma_cshuffle.hpp
rename to include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_wmma_cshuffle.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp
rename to include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_gemm_reduce_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_gemm_reduce_xdl_cshuffle.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_gemm_reduce_xdl_cshuffle.hpp
rename to include/ck/tensor_operation/gpu/device/impl/device_gemm_reduce_xdl_cshuffle.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_gemm_wmma.hpp b/include/ck/tensor_operation/gpu/device/impl/device_gemm_wmma.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_gemm_wmma.hpp
rename to include/ck/tensor_operation/gpu/device/impl/device_gemm_wmma.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl.hpp b/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl.hpp
rename to include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp
rename to include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_layernorm_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_layernorm_cshuffle.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_layernorm_cshuffle.hpp
rename to include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_layernorm_cshuffle.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_skip_b_lds.hpp b/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_skip_b_lds.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_skip_b_lds.hpp
rename to include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_skip_b_lds.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_splitk_c_shuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_splitk_c_shuffle.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_splitk_c_shuffle.hpp
rename to include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_splitk_c_shuffle.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_streamk.hpp b/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_streamk.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_streamk.hpp
rename to include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_streamk.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_waveletmodel_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_waveletmodel_cshuffle.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_waveletmodel_cshuffle.hpp
rename to include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_waveletmodel_cshuffle.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_grouped_contraction_multiple_d_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_grouped_contraction_multiple_d_xdl_cshuffle.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_grouped_contraction_multiple_d_xdl_cshuffle.hpp
rename to include/ck/tensor_operation/gpu/device/impl/device_grouped_contraction_multiple_d_xdl_cshuffle.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_data_multiple_d_wmma_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_data_multiple_d_wmma_cshuffle.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_data_multiple_d_wmma_cshuffle.hpp
rename to include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_data_multiple_d_wmma_cshuffle.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_data_multiple_d_xdl_cshuffle_v1.hpp b/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_data_multiple_d_xdl_cshuffle_v1.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_data_multiple_d_xdl_cshuffle_v1.hpp
rename to include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_data_multiple_d_xdl_cshuffle_v1.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_dl.hpp b/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_dl.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_dl.hpp
rename to include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_dl.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_wmma_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_wmma_cshuffle.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_wmma_cshuffle.hpp
rename to include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_wmma_cshuffle.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_xdl_cshuffle.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_xdl_cshuffle.hpp
rename to include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_xdl_cshuffle.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp b/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp
rename to include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_dl_nhwc_kyxc_nhwk.hpp b/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_dl_nhwc_kyxc_nhwk.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_dl_nhwc_kyxc_nhwk.hpp
rename to include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_dl_nhwc_kyxc_nhwk.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_multiple_r.hpp b/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_multiple_r.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_multiple_r.hpp
rename to include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_multiple_r.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_multiple_r_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_multiple_r_xdl_cshuffle.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_multiple_r_xdl_cshuffle.hpp
rename to include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_multiple_r_xdl_cshuffle.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_wmma_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_wmma_cshuffle.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_wmma_cshuffle.hpp
rename to include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_wmma_cshuffle.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp
rename to include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_utils.hpp b/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_utils.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_utils.hpp
rename to include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_utils.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_multiple_d_dl.hpp b/include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_multiple_d_dl.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_multiple_d_dl.hpp
rename to include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_multiple_d_dl.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_softmax_gemm_permute_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_softmax_gemm_permute_xdl_cshuffle.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_softmax_gemm_permute_xdl_cshuffle.hpp
rename to include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_softmax_gemm_permute_xdl_cshuffle.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl.hpp b/include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl.hpp
rename to include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl_fixed_nk.hpp b/include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl_fixed_nk.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl_fixed_nk.hpp
rename to include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl_fixed_nk.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl_splitk_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl_splitk_cshuffle.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl_splitk_cshuffle.hpp
rename to include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl_splitk_cshuffle.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_image_to_column_impl.hpp b/include/ck/tensor_operation/gpu/device/impl/device_image_to_column_impl.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_image_to_column_impl.hpp
rename to include/ck/tensor_operation/gpu/device/impl/device_image_to_column_impl.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_max_pool_bwd_impl.hpp b/include/ck/tensor_operation/gpu/device/impl/device_max_pool_bwd_impl.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_max_pool_bwd_impl.hpp
rename to include/ck/tensor_operation/gpu/device/impl/device_max_pool_bwd_impl.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_multiple_reduce_multiblock.hpp b/include/ck/tensor_operation/gpu/device/impl/device_multiple_reduce_multiblock.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_multiple_reduce_multiblock.hpp
rename to include/ck/tensor_operation/gpu/device/impl/device_multiple_reduce_multiblock.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_multiple_reduce_threadwise.hpp b/include/ck/tensor_operation/gpu/device/impl/device_multiple_reduce_threadwise.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_multiple_reduce_threadwise.hpp
rename to include/ck/tensor_operation/gpu/device/impl/device_multiple_reduce_threadwise.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_normalization_impl.hpp b/include/ck/tensor_operation/gpu/device/impl/device_normalization_impl.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_normalization_impl.hpp
rename to include/ck/tensor_operation/gpu/device/impl/device_normalization_impl.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_normalization_splitk_impl.hpp b/include/ck/tensor_operation/gpu/device/impl/device_normalization_splitk_impl.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_normalization_splitk_impl.hpp
rename to include/ck/tensor_operation/gpu/device/impl/device_normalization_splitk_impl.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_permute_impl.hpp b/include/ck/tensor_operation/gpu/device/impl/device_permute_impl.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_permute_impl.hpp
rename to include/ck/tensor_operation/gpu/device/impl/device_permute_impl.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_pool2d_fwd_nhwc_nhwc.hpp b/include/ck/tensor_operation/gpu/device/impl/device_pool2d_fwd_nhwc_nhwc.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_pool2d_fwd_nhwc_nhwc.hpp
rename to include/ck/tensor_operation/gpu/device/impl/device_pool2d_fwd_nhwc_nhwc.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_pool3d_fwd_ndhwc_ndhwc.hpp b/include/ck/tensor_operation/gpu/device/impl/device_pool3d_fwd_ndhwc_ndhwc.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_pool3d_fwd_ndhwc_ndhwc.hpp
rename to include/ck/tensor_operation/gpu/device/impl/device_pool3d_fwd_ndhwc_ndhwc.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_put_element_impl.hpp b/include/ck/tensor_operation/gpu/device/impl/device_put_element_impl.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_put_element_impl.hpp
rename to include/ck/tensor_operation/gpu/device/impl/device_put_element_impl.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_reduce_common.hpp b/include/ck/tensor_operation/gpu/device/impl/device_reduce_common.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_reduce_common.hpp
rename to include/ck/tensor_operation/gpu/device/impl/device_reduce_common.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_reduce_multiblock.hpp b/include/ck/tensor_operation/gpu/device/impl/device_reduce_multiblock.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_reduce_multiblock.hpp
rename to include/ck/tensor_operation/gpu/device/impl/device_reduce_multiblock.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_reduce_threadwise.hpp b/include/ck/tensor_operation/gpu/device/impl/device_reduce_threadwise.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_reduce_threadwise.hpp
rename to include/ck/tensor_operation/gpu/device/impl/device_reduce_threadwise.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_softmax_impl.hpp b/include/ck/tensor_operation/gpu/device/impl/device_softmax_impl.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_softmax_impl.hpp
rename to include/ck/tensor_operation/gpu/device/impl/device_softmax_impl.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_sparse_embeddings_forward_layernorm.hpp b/include/ck/tensor_operation/gpu/device/impl/device_sparse_embeddings_forward_layernorm.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_sparse_embeddings_forward_layernorm.hpp
rename to include/ck/tensor_operation/gpu/device/impl/device_sparse_embeddings_forward_layernorm.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_splitk_contraction_multiple_d_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_splitk_contraction_multiple_d_xdl_cshuffle.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_splitk_contraction_multiple_d_xdl_cshuffle.hpp
rename to include/ck/tensor_operation/gpu/device/impl/device_splitk_contraction_multiple_d_xdl_cshuffle.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_tall_and_skinny_gemm_splitk.hpp b/include/ck/tensor_operation/gpu/device/impl/device_tall_and_skinny_gemm_splitk.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/impl/device_tall_and_skinny_gemm_splitk.hpp
rename to include/ck/tensor_operation/gpu/device/impl/device_tall_and_skinny_gemm_splitk.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/masking_specialization.hpp b/include/ck/tensor_operation/gpu/device/masking_specialization.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/masking_specialization.hpp
rename to include/ck/tensor_operation/gpu/device/masking_specialization.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/matrix_padder.hpp b/include/ck/tensor_operation/gpu/device/matrix_padder.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/matrix_padder.hpp
rename to include/ck/tensor_operation/gpu/device/matrix_padder.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp b/include/ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp
rename to include/ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/tensor_layout.hpp b/include/ck/tensor_operation/gpu/device/tensor_layout.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/tensor_layout.hpp
rename to include/ck/tensor_operation/gpu/device/tensor_layout.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/tensor_specialization.hpp b/include/ck/tensor_operation/gpu/device/tensor_specialization.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/tensor_specialization.hpp
rename to include/ck/tensor_operation/gpu/device/tensor_specialization.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/device/welford_helper.hpp b/include/ck/tensor_operation/gpu/device/welford_helper.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/device/welford_helper.hpp
rename to include/ck/tensor_operation/gpu/device/welford_helper.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp b/include/ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp
rename to include/ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/element/element_wise_operation.hpp b/include/ck/tensor_operation/gpu/element/element_wise_operation.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/element/element_wise_operation.hpp
rename to include/ck/tensor_operation/gpu/element/element_wise_operation.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/element/quantization_operation.hpp b/include/ck/tensor_operation/gpu/element/quantization_operation.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/element/quantization_operation.hpp
rename to include/ck/tensor_operation/gpu/element/quantization_operation.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp b/include/ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp
rename to include/ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_batchnorm_forward.hpp b/include/ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_batchnorm_forward.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_batchnorm_forward.hpp
rename to include/ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_batchnorm_forward.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_reduce_second_half_batchnorm_backward_final.hpp b/include/ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_reduce_second_half_batchnorm_backward_final.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_reduce_second_half_batchnorm_backward_final.hpp
rename to include/ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_reduce_second_half_batchnorm_backward_final.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_welford_first_half.hpp b/include/ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_welford_first_half.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_welford_first_half.hpp
rename to include/ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_welford_first_half.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_welford_second_half_batchnorm_forward_final_obsolete.hpp b/include/ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_welford_second_half_batchnorm_forward_final_obsolete.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_welford_second_half_batchnorm_forward_final_obsolete.hpp
rename to include/ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_welford_second_half_batchnorm_forward_final_obsolete.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_welford_second_half_multiblock_reduce_first_half.hpp b/include/ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_welford_second_half_multiblock_reduce_first_half.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_welford_second_half_multiblock_reduce_first_half.hpp
rename to include/ck/tensor_operation/gpu/grid/batchnorm_multiblock/gridwise_multiblock_welford_second_half_multiblock_reduce_first_half.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp b/include/ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp
rename to include/ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/grid/gemm_layernorm/gridwise_gemm_multiple_d_welford_first_half_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/grid/gemm_layernorm/gridwise_gemm_multiple_d_welford_first_half_xdl_cshuffle.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/grid/gemm_layernorm/gridwise_gemm_multiple_d_welford_first_half_xdl_cshuffle.hpp
rename to include/ck/tensor_operation/gpu/grid/gemm_layernorm/gridwise_gemm_multiple_d_welford_first_half_xdl_cshuffle.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/grid/gemm_layernorm/gridwise_welford_second_half_layernorm2d.hpp b/include/ck/tensor_operation/gpu/grid/gemm_layernorm/gridwise_welford_second_half_layernorm2d.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/grid/gemm_layernorm/gridwise_welford_second_half_layernorm2d.hpp
rename to include/ck/tensor_operation/gpu/grid/gemm_layernorm/gridwise_welford_second_half_layernorm2d.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_2d_multiple_reduction_multiblock.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_2d_multiple_reduction_multiblock.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_2d_multiple_reduction_multiblock.hpp
rename to include/ck/tensor_operation/gpu/grid/gridwise_2d_multiple_reduction_multiblock.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_2d_multiple_reduction_threadwise.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_2d_multiple_reduction_threadwise.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_2d_multiple_reduction_threadwise.hpp
rename to include/ck/tensor_operation/gpu/grid/gridwise_2d_multiple_reduction_threadwise.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_2d_reduction_multiblock.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_2d_reduction_multiblock.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_2d_reduction_multiblock.hpp
rename to include/ck/tensor_operation/gpu/grid/gridwise_2d_reduction_multiblock.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_2d_reduction_threadwise.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_2d_reduction_threadwise.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_2d_reduction_threadwise.hpp
rename to include/ck/tensor_operation/gpu/grid/gridwise_2d_reduction_threadwise.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_gemm_xdl_cshuffle_v1.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_gemm_xdl_cshuffle_v1.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_gemm_xdl_cshuffle_v1.hpp
rename to include/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_gemm_xdl_cshuffle_v1.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_multiple_d_gemm_multiple_d_xdl_cshuffle_v1.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_multiple_d_gemm_multiple_d_xdl_cshuffle_v1.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_multiple_d_gemm_multiple_d_xdl_cshuffle_v1.hpp
rename to include/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_multiple_d_gemm_multiple_d_xdl_cshuffle_v1.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_multiple_d_softmax_gemm_xdl_cshuffle_v1.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_multiple_d_softmax_gemm_xdl_cshuffle_v1.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_multiple_d_softmax_gemm_xdl_cshuffle_v1.hpp
rename to include/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_multiple_d_softmax_gemm_xdl_cshuffle_v1.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_softmax_gemm_xdl_cshuffle_v1.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_softmax_gemm_xdl_cshuffle_v1.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_softmax_gemm_xdl_cshuffle_v1.hpp
rename to include/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_softmax_gemm_xdl_cshuffle_v1.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_batchnorm_backward_blockwise_welford.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_batchnorm_backward_blockwise_welford.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_batchnorm_backward_blockwise_welford.hpp
rename to include/ck/tensor_operation/gpu/grid/gridwise_batchnorm_backward_blockwise_welford.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_batchnorm_forward_blockwise_welford.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_batchnorm_forward_blockwise_welford.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_batchnorm_forward_blockwise_welford.hpp
rename to include/ck/tensor_operation/gpu/grid/gridwise_batchnorm_forward_blockwise_welford.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_elementwise_1d.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_elementwise_1d.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_elementwise_1d.hpp
rename to include/ck/tensor_operation/gpu/grid/gridwise_elementwise_1d.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_elementwise_2d.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_elementwise_2d.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_elementwise_2d.hpp
rename to include/ck/tensor_operation/gpu/grid/gridwise_elementwise_2d.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_elementwise_layernorm_welford_variance.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_elementwise_layernorm_welford_variance.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_elementwise_layernorm_welford_variance.hpp
rename to include/ck/tensor_operation/gpu/grid/gridwise_elementwise_layernorm_welford_variance.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_bias_add_reduce_xdl_cshuffle_v1.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_bias_add_reduce_xdl_cshuffle_v1.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_bias_add_reduce_xdl_cshuffle_v1.hpp
rename to include/ck/tensor_operation/gpu/grid/gridwise_gemm_bias_add_reduce_xdl_cshuffle_v1.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_dl_multiple_d.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_dl_multiple_d.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_dl_multiple_d.hpp
rename to include/ck/tensor_operation/gpu/grid/gridwise_gemm_dl_multiple_d.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_dl_v1r3.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_dl_v1r3.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_dl_v1r3.hpp
rename to include/ck/tensor_operation/gpu/grid/gridwise_gemm_dl_v1r3.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_dpp.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_dpp.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_dpp.hpp
rename to include/ck/tensor_operation/gpu/grid/gridwise_gemm_dpp.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_abd_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_abd_xdl_cshuffle.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_abd_xdl_cshuffle.hpp
rename to include/ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_abd_xdl_cshuffle.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_multiple_r_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_multiple_r_xdl_cshuffle.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_multiple_r_xdl_cshuffle.hpp
rename to include/ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_multiple_r_xdl_cshuffle.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_wmma_cshuffle.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_wmma_cshuffle.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_wmma_cshuffle.hpp
rename to include/ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_wmma_cshuffle.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_xdl_cshuffle.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_xdl_cshuffle.hpp
rename to include/ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_xdl_cshuffle.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_xdl_splitk_cshuffle.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_xdl_splitk_cshuffle.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_xdl_splitk_cshuffle.hpp
rename to include/ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_xdl_splitk_cshuffle.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_selector.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_selector.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_selector.hpp
rename to include/ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_selector.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v1.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v1.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v1.hpp
rename to include/ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v1.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v2.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v2.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v2.hpp
rename to include/ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v2.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v3.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v3.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v3.hpp
rename to include/ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v3.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_reduce_xdl_cshuffle_v1.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_reduce_xdl_cshuffle_v1.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_reduce_xdl_cshuffle_v1.hpp
rename to include/ck/tensor_operation/gpu/grid/gridwise_gemm_reduce_xdl_cshuffle_v1.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_split_k_multiple_d_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_split_k_multiple_d_xdl_cshuffle.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_split_k_multiple_d_xdl_cshuffle.hpp
rename to include/ck/tensor_operation/gpu/grid/gridwise_gemm_split_k_multiple_d_xdl_cshuffle.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_split_k_multiple_d_xdl_cshuffle_v2.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_split_k_multiple_d_xdl_cshuffle_v2.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_split_k_multiple_d_xdl_cshuffle_v2.hpp
rename to include/ck/tensor_operation/gpu/grid/gridwise_gemm_split_k_multiple_d_xdl_cshuffle_v2.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_waveletmodel.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_waveletmodel.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_waveletmodel.hpp
rename to include/ck/tensor_operation/gpu/grid/gridwise_gemm_waveletmodel.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_wmma.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_wmma.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_wmma.hpp
rename to include/ck/tensor_operation/gpu/grid/gridwise_gemm_wmma.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v1.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v1.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v1.hpp
rename to include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v1.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_layernorm_cshuffle_v1.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_layernorm_cshuffle_v1.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_layernorm_cshuffle_v1.hpp
rename to include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_layernorm_cshuffle_v1.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_waveletmodel_cshuffle.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_waveletmodel_cshuffle.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_waveletmodel_cshuffle.hpp
rename to include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_waveletmodel_cshuffle.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_bwd_weight.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_bwd_weight.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_bwd_weight.hpp
rename to include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_bwd_weight.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_skip_b_lds_v1.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_skip_b_lds_v1.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_skip_b_lds_v1.hpp
rename to include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_skip_b_lds_v1.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_streamk.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_streamk.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_streamk.hpp
rename to include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_streamk.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r3.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r3.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r3.hpp
rename to include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r3.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r4.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r4.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r4.hpp
rename to include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r4.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r4r2.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r4r2.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r4r2.hpp
rename to include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r4r2.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r1.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r1.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r1.hpp
rename to include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r1.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r2.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r2.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r2.hpp
rename to include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r2.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r3.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r3.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r3.hpp
rename to include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r3.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemv_splitk.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemv_splitk.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_gemv_splitk.hpp
rename to include/ck/tensor_operation/gpu/grid/gridwise_gemv_splitk.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_permute.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_permute.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_permute.hpp
rename to include/ck/tensor_operation/gpu/grid/gridwise_permute.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_put_element_1d.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_put_element_1d.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_put_element_1d.hpp
rename to include/ck/tensor_operation/gpu/grid/gridwise_put_element_1d.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_set_buffer_value.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_set_buffer_value.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_set_buffer_value.hpp
rename to include/ck/tensor_operation/gpu/grid/gridwise_set_buffer_value.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_set_multiple_buffer_value.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_set_multiple_buffer_value.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_set_multiple_buffer_value.hpp
rename to include/ck/tensor_operation/gpu/grid/gridwise_set_multiple_buffer_value.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_softmax.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_softmax.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_softmax.hpp
rename to include/ck/tensor_operation/gpu/grid/gridwise_softmax.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_sparse_embeddings_forward_layernorm.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_sparse_embeddings_forward_layernorm.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_sparse_embeddings_forward_layernorm.hpp
rename to include/ck/tensor_operation/gpu/grid/gridwise_sparse_embeddings_forward_layernorm.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_tall_and_skinny_gemm_splitk.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_tall_and_skinny_gemm_splitk.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_tall_and_skinny_gemm_splitk.hpp
rename to include/ck/tensor_operation/gpu/grid/gridwise_tall_and_skinny_gemm_splitk.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_tensor_rearrange.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_tensor_rearrange.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/grid/gridwise_tensor_rearrange.hpp
rename to include/ck/tensor_operation/gpu/grid/gridwise_tensor_rearrange.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/grid/normalization/gridwise_normalization_naive_variance.hpp b/include/ck/tensor_operation/gpu/grid/normalization/gridwise_normalization_naive_variance.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/grid/normalization/gridwise_normalization_naive_variance.hpp
rename to include/ck/tensor_operation/gpu/grid/normalization/gridwise_normalization_naive_variance.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/grid/normalization/gridwise_normalization_selector.hpp b/include/ck/tensor_operation/gpu/grid/normalization/gridwise_normalization_selector.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/grid/normalization/gridwise_normalization_selector.hpp
rename to include/ck/tensor_operation/gpu/grid/normalization/gridwise_normalization_selector.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/grid/normalization/gridwise_normalization_splitk_1st.hpp b/include/ck/tensor_operation/gpu/grid/normalization/gridwise_normalization_splitk_1st.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/grid/normalization/gridwise_normalization_splitk_1st.hpp
rename to include/ck/tensor_operation/gpu/grid/normalization/gridwise_normalization_splitk_1st.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/grid/normalization/gridwise_normalization_splitk_2nd.hpp b/include/ck/tensor_operation/gpu/grid/normalization/gridwise_normalization_splitk_2nd.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/grid/normalization/gridwise_normalization_splitk_2nd.hpp
rename to include/ck/tensor_operation/gpu/grid/normalization/gridwise_normalization_splitk_2nd.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/grid/normalization/gridwise_normalization_welford_variance.hpp b/include/ck/tensor_operation/gpu/grid/normalization/gridwise_normalization_welford_variance.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/grid/normalization/gridwise_normalization_welford_variance.hpp
rename to include/ck/tensor_operation/gpu/grid/normalization/gridwise_normalization_welford_variance.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/thread/reduction_functions_threadwise.hpp b/include/ck/tensor_operation/gpu/thread/reduction_functions_threadwise.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/thread/reduction_functions_threadwise.hpp
rename to include/ck/tensor_operation/gpu/thread/reduction_functions_threadwise.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/thread/threadwise_contraction_dl.hpp b/include/ck/tensor_operation/gpu/thread/threadwise_contraction_dl.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/thread/threadwise_contraction_dl.hpp
rename to include/ck/tensor_operation/gpu/thread/threadwise_contraction_dl.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/thread/threadwise_gemm_dlops_v3.hpp b/include/ck/tensor_operation/gpu/thread/threadwise_gemm_dlops_v3.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/thread/threadwise_gemm_dlops_v3.hpp
rename to include/ck/tensor_operation/gpu/thread/threadwise_gemm_dlops_v3.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_set.hpp b/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_set.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_set.hpp
rename to include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_set.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp b/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp
rename to include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v3r1.hpp b/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v3r1.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v3r1.hpp
rename to include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v3r1.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v4r1.hpp b/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v4r1.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v4r1.hpp
rename to include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v4r1.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v5r1.hpp b/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v5r1.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v5r1.hpp
rename to include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v5r1.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v6r1.hpp b/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v6r1.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v6r1.hpp
rename to include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v6r1.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v6r1r2.hpp b/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v6r1r2.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v6r1r2.hpp
rename to include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v6r1r2.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v6r2.hpp b/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v6r2.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v6r2.hpp
rename to include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v6r2.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v6r3.hpp b/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v6r3.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v6r3.hpp
rename to include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v6r3.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v7.hpp b/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v7.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v7.hpp
rename to include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v7.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v7r2.hpp b/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v7r2.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v7r2.hpp
rename to include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v7r2.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/thread/threadwise_welford.hpp b/include/ck/tensor_operation/gpu/thread/threadwise_welford.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/thread/threadwise_welford.hpp
rename to include/ck/tensor_operation/gpu/thread/threadwise_welford.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/warp/dpp_gemm.hpp b/include/ck/tensor_operation/gpu/warp/dpp_gemm.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/warp/dpp_gemm.hpp
rename to include/ck/tensor_operation/gpu/warp/dpp_gemm.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/warp/wmma_gemm.hpp b/include/ck/tensor_operation/gpu/warp/wmma_gemm.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/warp/wmma_gemm.hpp
rename to include/ck/tensor_operation/gpu/warp/wmma_gemm.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/gpu/warp/xdlops_gemm.hpp b/include/ck/tensor_operation/gpu/warp/xdlops_gemm.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/gpu/warp/xdlops_gemm.hpp
rename to include/ck/tensor_operation/gpu/warp/xdlops_gemm.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/operator_transform/transform_contraction_to_gemm.hpp b/include/ck/tensor_operation/operator_transform/transform_contraction_to_gemm.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/operator_transform/transform_contraction_to_gemm.hpp
rename to include/ck/tensor_operation/operator_transform/transform_contraction_to_gemm.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/operator_transform/transform_conv_bwd_data_to_gemm_v1.hpp b/include/ck/tensor_operation/operator_transform/transform_conv_bwd_data_to_gemm_v1.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/operator_transform/transform_conv_bwd_data_to_gemm_v1.hpp
rename to include/ck/tensor_operation/operator_transform/transform_conv_bwd_data_to_gemm_v1.hpp
diff --git a/composable_kernel/include/ck/tensor_operation/operator_transform/transform_conv_fwd_to_gemm.hpp b/include/ck/tensor_operation/operator_transform/transform_conv_fwd_to_gemm.hpp
similarity index 100%
rename from composable_kernel/include/ck/tensor_operation/operator_transform/transform_conv_fwd_to_gemm.hpp
rename to include/ck/tensor_operation/operator_transform/transform_conv_fwd_to_gemm.hpp
diff --git a/composable_kernel/include/ck/utility/amd_address_space.hpp b/include/ck/utility/amd_address_space.hpp
similarity index 100%
rename from composable_kernel/include/ck/utility/amd_address_space.hpp
rename to include/ck/utility/amd_address_space.hpp
diff --git a/composable_kernel/include/ck/utility/amd_buffer_addressing.hpp b/include/ck/utility/amd_buffer_addressing.hpp
similarity index 100%
rename from composable_kernel/include/ck/utility/amd_buffer_addressing.hpp
rename to include/ck/utility/amd_buffer_addressing.hpp
diff --git a/composable_kernel/include/ck/utility/amd_gemm_dpp.hpp b/include/ck/utility/amd_gemm_dpp.hpp
similarity index 100%
rename from composable_kernel/include/ck/utility/amd_gemm_dpp.hpp
rename to include/ck/utility/amd_gemm_dpp.hpp
diff --git a/composable_kernel/include/ck/utility/amd_inline_asm.hpp b/include/ck/utility/amd_inline_asm.hpp
similarity index 100%
rename from composable_kernel/include/ck/utility/amd_inline_asm.hpp
rename to include/ck/utility/amd_inline_asm.hpp
diff --git a/composable_kernel/include/ck/utility/amd_wave_read_first_lane.hpp b/include/ck/utility/amd_wave_read_first_lane.hpp
similarity index 100%
rename from composable_kernel/include/ck/utility/amd_wave_read_first_lane.hpp
rename to include/ck/utility/amd_wave_read_first_lane.hpp
diff --git a/composable_kernel/include/ck/utility/amd_wmma.hpp b/include/ck/utility/amd_wmma.hpp
similarity index 100%
rename from composable_kernel/include/ck/utility/amd_wmma.hpp
rename to include/ck/utility/amd_wmma.hpp
diff --git a/composable_kernel/include/ck/utility/amd_xdlops.hpp b/include/ck/utility/amd_xdlops.hpp
similarity index 100%
rename from composable_kernel/include/ck/utility/amd_xdlops.hpp
rename to include/ck/utility/amd_xdlops.hpp
diff --git a/composable_kernel/include/ck/utility/array.hpp b/include/ck/utility/array.hpp
similarity index 100%
rename from composable_kernel/include/ck/utility/array.hpp
rename to include/ck/utility/array.hpp
diff --git a/composable_kernel/include/ck/utility/array_multi_index.hpp b/include/ck/utility/array_multi_index.hpp
similarity index 100%
rename from composable_kernel/include/ck/utility/array_multi_index.hpp
rename to include/ck/utility/array_multi_index.hpp
diff --git a/composable_kernel/include/ck/utility/c_style_pointer_cast.hpp b/include/ck/utility/c_style_pointer_cast.hpp
similarity index 100%
rename from composable_kernel/include/ck/utility/c_style_pointer_cast.hpp
rename to include/ck/utility/c_style_pointer_cast.hpp
diff --git a/composable_kernel/include/ck/utility/common_header.hpp b/include/ck/utility/common_header.hpp
similarity index 100%
rename from composable_kernel/include/ck/utility/common_header.hpp
rename to include/ck/utility/common_header.hpp
diff --git a/composable_kernel/include/ck/utility/container_element_picker.hpp b/include/ck/utility/container_element_picker.hpp
similarity index 100%
rename from composable_kernel/include/ck/utility/container_element_picker.hpp
rename to include/ck/utility/container_element_picker.hpp
diff --git a/composable_kernel/include/ck/utility/container_helper.hpp b/include/ck/utility/container_helper.hpp
similarity index 100%
rename from composable_kernel/include/ck/utility/container_helper.hpp
rename to include/ck/utility/container_helper.hpp
diff --git a/composable_kernel/include/ck/utility/data_type.hpp b/include/ck/utility/data_type.hpp
similarity index 100%
rename from composable_kernel/include/ck/utility/data_type.hpp
rename to include/ck/utility/data_type.hpp
diff --git a/composable_kernel/include/ck/utility/debug.hpp b/include/ck/utility/debug.hpp
similarity index 100%
rename from composable_kernel/include/ck/utility/debug.hpp
rename to include/ck/utility/debug.hpp
diff --git a/composable_kernel/include/ck/utility/dynamic_buffer.hpp b/include/ck/utility/dynamic_buffer.hpp
similarity index 100%
rename from composable_kernel/include/ck/utility/dynamic_buffer.hpp
rename to include/ck/utility/dynamic_buffer.hpp
diff --git a/composable_kernel/include/ck/utility/enable_if.hpp b/include/ck/utility/enable_if.hpp
similarity index 100%
rename from composable_kernel/include/ck/utility/enable_if.hpp
rename to include/ck/utility/enable_if.hpp
diff --git a/composable_kernel/include/ck/utility/f8_utils.hpp b/include/ck/utility/f8_utils.hpp
similarity index 100%
rename from composable_kernel/include/ck/utility/f8_utils.hpp
rename to include/ck/utility/f8_utils.hpp
diff --git a/composable_kernel/include/ck/utility/functional.hpp b/include/ck/utility/functional.hpp
similarity index 100%
rename from composable_kernel/include/ck/utility/functional.hpp
rename to include/ck/utility/functional.hpp
diff --git a/composable_kernel/include/ck/utility/functional2.hpp b/include/ck/utility/functional2.hpp
similarity index 100%
rename from composable_kernel/include/ck/utility/functional2.hpp
rename to include/ck/utility/functional2.hpp
diff --git a/composable_kernel/include/ck/utility/functional3.hpp b/include/ck/utility/functional3.hpp
similarity index 100%
rename from composable_kernel/include/ck/utility/functional3.hpp
rename to include/ck/utility/functional3.hpp
diff --git a/composable_kernel/include/ck/utility/functional4.hpp b/include/ck/utility/functional4.hpp
similarity index 100%
rename from composable_kernel/include/ck/utility/functional4.hpp
rename to include/ck/utility/functional4.hpp
diff --git a/composable_kernel/include/ck/utility/generic_memory_space_atomic.hpp b/include/ck/utility/generic_memory_space_atomic.hpp
similarity index 100%
rename from composable_kernel/include/ck/utility/generic_memory_space_atomic.hpp
rename to include/ck/utility/generic_memory_space_atomic.hpp
diff --git a/composable_kernel/include/ck/utility/get_id.hpp b/include/ck/utility/get_id.hpp
similarity index 100%
rename from composable_kernel/include/ck/utility/get_id.hpp
rename to include/ck/utility/get_id.hpp
diff --git a/composable_kernel/include/ck/utility/get_shift.hpp b/include/ck/utility/get_shift.hpp
similarity index 100%
rename from composable_kernel/include/ck/utility/get_shift.hpp
rename to include/ck/utility/get_shift.hpp
diff --git a/composable_kernel/include/ck/utility/ignore.hpp b/include/ck/utility/ignore.hpp
similarity index 100%
rename from composable_kernel/include/ck/utility/ignore.hpp
rename to include/ck/utility/ignore.hpp
diff --git a/composable_kernel/include/ck/utility/inner_product.hpp b/include/ck/utility/inner_product.hpp
similarity index 100%
rename from composable_kernel/include/ck/utility/inner_product.hpp
rename to include/ck/utility/inner_product.hpp
diff --git a/composable_kernel/include/ck/utility/inner_product_dpp8.hpp b/include/ck/utility/inner_product_dpp8.hpp
similarity index 100%
rename from composable_kernel/include/ck/utility/inner_product_dpp8.hpp
rename to include/ck/utility/inner_product_dpp8.hpp
diff --git a/composable_kernel/include/ck/utility/integral_constant.hpp b/include/ck/utility/integral_constant.hpp
similarity index 100%
rename from composable_kernel/include/ck/utility/integral_constant.hpp
rename to include/ck/utility/integral_constant.hpp
diff --git a/composable_kernel/include/ck/utility/is_detected.hpp b/include/ck/utility/is_detected.hpp
similarity index 100%
rename from composable_kernel/include/ck/utility/is_detected.hpp
rename to include/ck/utility/is_detected.hpp
diff --git a/composable_kernel/include/ck/utility/is_known_at_compile_time.hpp b/include/ck/utility/is_known_at_compile_time.hpp
similarity index 100%
rename from composable_kernel/include/ck/utility/is_known_at_compile_time.hpp
rename to include/ck/utility/is_known_at_compile_time.hpp
diff --git a/composable_kernel/include/ck/utility/loop_scheduler.hpp b/include/ck/utility/loop_scheduler.hpp
similarity index 100%
rename from composable_kernel/include/ck/utility/loop_scheduler.hpp
rename to include/ck/utility/loop_scheduler.hpp
diff --git a/composable_kernel/include/ck/utility/magic_division.hpp b/include/ck/utility/magic_division.hpp
similarity index 100%
rename from composable_kernel/include/ck/utility/magic_division.hpp
rename to include/ck/utility/magic_division.hpp
diff --git a/composable_kernel/include/ck/utility/math.hpp b/include/ck/utility/math.hpp
similarity index 100%
rename from composable_kernel/include/ck/utility/math.hpp
rename to include/ck/utility/math.hpp
diff --git a/composable_kernel/include/ck/utility/math_v2.hpp b/include/ck/utility/math_v2.hpp
similarity index 100%
rename from composable_kernel/include/ck/utility/math_v2.hpp
rename to include/ck/utility/math_v2.hpp
diff --git a/composable_kernel/include/ck/utility/multi_index.hpp b/include/ck/utility/multi_index.hpp
similarity index 100%
rename from composable_kernel/include/ck/utility/multi_index.hpp
rename to include/ck/utility/multi_index.hpp
diff --git a/composable_kernel/include/ck/utility/number.hpp b/include/ck/utility/number.hpp
similarity index 100%
rename from composable_kernel/include/ck/utility/number.hpp
rename to include/ck/utility/number.hpp
diff --git a/composable_kernel/include/ck/utility/random_gen.hpp b/include/ck/utility/random_gen.hpp
similarity index 100%
rename from composable_kernel/include/ck/utility/random_gen.hpp
rename to include/ck/utility/random_gen.hpp
diff --git a/composable_kernel/include/ck/utility/reduction_common.hpp b/include/ck/utility/reduction_common.hpp
similarity index 100%
rename from composable_kernel/include/ck/utility/reduction_common.hpp
rename to include/ck/utility/reduction_common.hpp
diff --git a/composable_kernel/include/ck/utility/reduction_enums.hpp b/include/ck/utility/reduction_enums.hpp
similarity index 100%
rename from composable_kernel/include/ck/utility/reduction_enums.hpp
rename to include/ck/utility/reduction_enums.hpp
diff --git a/composable_kernel/include/ck/utility/reduction_functions_accumulate.hpp b/include/ck/utility/reduction_functions_accumulate.hpp
similarity index 100%
rename from composable_kernel/include/ck/utility/reduction_functions_accumulate.hpp
rename to include/ck/utility/reduction_functions_accumulate.hpp
diff --git a/composable_kernel/include/ck/utility/reduction_operator.hpp b/include/ck/utility/reduction_operator.hpp
similarity index 100%
rename from composable_kernel/include/ck/utility/reduction_operator.hpp
rename to include/ck/utility/reduction_operator.hpp
diff --git a/composable_kernel/include/ck/utility/sequence.hpp b/include/ck/utility/sequence.hpp
similarity index 100%
rename from composable_kernel/include/ck/utility/sequence.hpp
rename to include/ck/utility/sequence.hpp
diff --git a/composable_kernel/include/ck/utility/sequence_helper.hpp b/include/ck/utility/sequence_helper.hpp
similarity index 100%
rename from composable_kernel/include/ck/utility/sequence_helper.hpp
rename to include/ck/utility/sequence_helper.hpp
diff --git a/composable_kernel/include/ck/utility/span.hpp b/include/ck/utility/span.hpp
similarity index 100%
rename from composable_kernel/include/ck/utility/span.hpp
rename to include/ck/utility/span.hpp
diff --git a/composable_kernel/include/ck/utility/static_buffer.hpp b/include/ck/utility/static_buffer.hpp
similarity index 100%
rename from composable_kernel/include/ck/utility/static_buffer.hpp
rename to include/ck/utility/static_buffer.hpp
diff --git a/composable_kernel/include/ck/utility/statically_indexed_array.hpp b/include/ck/utility/statically_indexed_array.hpp
similarity index 100%
rename from composable_kernel/include/ck/utility/statically_indexed_array.hpp
rename to include/ck/utility/statically_indexed_array.hpp
diff --git a/composable_kernel/include/ck/utility/statically_indexed_array_multi_index.hpp b/include/ck/utility/statically_indexed_array_multi_index.hpp
similarity index 100%
rename from composable_kernel/include/ck/utility/statically_indexed_array_multi_index.hpp
rename to include/ck/utility/statically_indexed_array_multi_index.hpp
diff --git a/composable_kernel/include/ck/utility/synchronization.hpp b/include/ck/utility/synchronization.hpp
similarity index 100%
rename from composable_kernel/include/ck/utility/synchronization.hpp
rename to include/ck/utility/synchronization.hpp
diff --git a/composable_kernel/include/ck/utility/thread_group.hpp b/include/ck/utility/thread_group.hpp
similarity index 100%
rename from composable_kernel/include/ck/utility/thread_group.hpp
rename to include/ck/utility/thread_group.hpp
diff --git a/composable_kernel/include/ck/utility/transpose_vectors.hpp b/include/ck/utility/transpose_vectors.hpp
similarity index 100%
rename from composable_kernel/include/ck/utility/transpose_vectors.hpp
rename to include/ck/utility/transpose_vectors.hpp
diff --git a/composable_kernel/include/ck/utility/tuple.hpp b/include/ck/utility/tuple.hpp
similarity index 100%
rename from composable_kernel/include/ck/utility/tuple.hpp
rename to include/ck/utility/tuple.hpp
diff --git a/composable_kernel/include/ck/utility/tuple_helper.hpp b/include/ck/utility/tuple_helper.hpp
similarity index 100%
rename from composable_kernel/include/ck/utility/tuple_helper.hpp
rename to include/ck/utility/tuple_helper.hpp
diff --git a/composable_kernel/include/ck/utility/type.hpp b/include/ck/utility/type.hpp
similarity index 100%
rename from composable_kernel/include/ck/utility/type.hpp
rename to include/ck/utility/type.hpp
diff --git a/composable_kernel/include/ck/utility/type_convert.hpp b/include/ck/utility/type_convert.hpp
similarity index 100%
rename from composable_kernel/include/ck/utility/type_convert.hpp
rename to include/ck/utility/type_convert.hpp
diff --git a/composable_kernel/include/ck/utility/workgroup_barrier.hpp b/include/ck/utility/workgroup_barrier.hpp
similarity index 100%
rename from composable_kernel/include/ck/utility/workgroup_barrier.hpp
rename to include/ck/utility/workgroup_barrier.hpp
diff --git a/composable_kernel/include/ck/utility/workgroup_synchronization.hpp b/include/ck/utility/workgroup_synchronization.hpp
similarity index 100%
rename from composable_kernel/include/ck/utility/workgroup_synchronization.hpp
rename to include/ck/utility/workgroup_synchronization.hpp
diff --git a/composable_kernel/include/ck/version.h.in b/include/ck/version.h.in
similarity index 100%
rename from composable_kernel/include/ck/version.h.in
rename to include/ck/version.h.in
diff --git a/composable_kernel/library/CMakeLists.txt b/library/CMakeLists.txt
similarity index 100%
rename from composable_kernel/library/CMakeLists.txt
rename to library/CMakeLists.txt
diff --git a/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_avgpool_bwd.hpp b/library/include/ck/library/reference_tensor_operation/cpu/reference_avgpool_bwd.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_avgpool_bwd.hpp
rename to library/include/ck/library/reference_tensor_operation/cpu/reference_avgpool_bwd.hpp
diff --git a/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp b/library/include/ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp
rename to library/include/ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp
diff --git a/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_batchnorm_backward.hpp b/library/include/ck/library/reference_tensor_operation/cpu/reference_batchnorm_backward.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_batchnorm_backward.hpp
rename to library/include/ck/library/reference_tensor_operation/cpu/reference_batchnorm_backward.hpp
diff --git a/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_batchnorm_forward.hpp b/library/include/ck/library/reference_tensor_operation/cpu/reference_batchnorm_forward.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_batchnorm_forward.hpp
rename to library/include/ck/library/reference_tensor_operation/cpu/reference_batchnorm_forward.hpp
diff --git a/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_batchnorm_infer.hpp b/library/include/ck/library/reference_tensor_operation/cpu/reference_batchnorm_infer.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_batchnorm_infer.hpp
rename to library/include/ck/library/reference_tensor_operation/cpu/reference_batchnorm_infer.hpp
diff --git a/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_cgemm.hpp b/library/include/ck/library/reference_tensor_operation/cpu/reference_cgemm.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_cgemm.hpp
rename to library/include/ck/library/reference_tensor_operation/cpu/reference_cgemm.hpp
diff --git a/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_column_to_image.hpp b/library/include/ck/library/reference_tensor_operation/cpu/reference_column_to_image.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_column_to_image.hpp
rename to library/include/ck/library/reference_tensor_operation/cpu/reference_column_to_image.hpp
diff --git a/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_contraction.hpp b/library/include/ck/library/reference_tensor_operation/cpu/reference_contraction.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_contraction.hpp
rename to library/include/ck/library/reference_tensor_operation/cpu/reference_contraction.hpp
diff --git a/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_conv_bwd_data.hpp b/library/include/ck/library/reference_tensor_operation/cpu/reference_conv_bwd_data.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_conv_bwd_data.hpp
rename to library/include/ck/library/reference_tensor_operation/cpu/reference_conv_bwd_data.hpp
diff --git a/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_conv_bwd_weight.hpp b/library/include/ck/library/reference_tensor_operation/cpu/reference_conv_bwd_weight.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_conv_bwd_weight.hpp
rename to library/include/ck/library/reference_tensor_operation/cpu/reference_conv_bwd_weight.hpp
diff --git a/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_conv_fwd.hpp b/library/include/ck/library/reference_tensor_operation/cpu/reference_conv_fwd.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_conv_fwd.hpp
rename to library/include/ck/library/reference_tensor_operation/cpu/reference_conv_fwd.hpp
diff --git a/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_conv_fwd_bias_activation.hpp b/library/include/ck/library/reference_tensor_operation/cpu/reference_conv_fwd_bias_activation.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_conv_fwd_bias_activation.hpp
rename to library/include/ck/library/reference_tensor_operation/cpu/reference_conv_fwd_bias_activation.hpp
diff --git a/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_conv_fwd_bias_activation_add.hpp b/library/include/ck/library/reference_tensor_operation/cpu/reference_conv_fwd_bias_activation_add.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_conv_fwd_bias_activation_add.hpp
rename to library/include/ck/library/reference_tensor_operation/cpu/reference_conv_fwd_bias_activation_add.hpp
diff --git a/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_gemm.hpp b/library/include/ck/library/reference_tensor_operation/cpu/reference_gemm.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_gemm.hpp
rename to library/include/ck/library/reference_tensor_operation/cpu/reference_gemm.hpp
diff --git a/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_gemm_layernorm.hpp b/library/include/ck/library/reference_tensor_operation/cpu/reference_gemm_layernorm.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_gemm_layernorm.hpp
rename to library/include/ck/library/reference_tensor_operation/cpu/reference_gemm_layernorm.hpp
diff --git a/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_groupnorm.hpp b/library/include/ck/library/reference_tensor_operation/cpu/reference_groupnorm.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_groupnorm.hpp
rename to library/include/ck/library/reference_tensor_operation/cpu/reference_groupnorm.hpp
diff --git a/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_image_to_column.hpp b/library/include/ck/library/reference_tensor_operation/cpu/reference_image_to_column.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_image_to_column.hpp
rename to library/include/ck/library/reference_tensor_operation/cpu/reference_image_to_column.hpp
diff --git a/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_layernorm.hpp b/library/include/ck/library/reference_tensor_operation/cpu/reference_layernorm.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_layernorm.hpp
rename to library/include/ck/library/reference_tensor_operation/cpu/reference_layernorm.hpp
diff --git a/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_maxpool_bwd.hpp b/library/include/ck/library/reference_tensor_operation/cpu/reference_maxpool_bwd.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_maxpool_bwd.hpp
rename to library/include/ck/library/reference_tensor_operation/cpu/reference_maxpool_bwd.hpp
diff --git a/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_pool_fwd.hpp b/library/include/ck/library/reference_tensor_operation/cpu/reference_pool_fwd.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_pool_fwd.hpp
rename to library/include/ck/library/reference_tensor_operation/cpu/reference_pool_fwd.hpp
diff --git a/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_reduce.hpp b/library/include/ck/library/reference_tensor_operation/cpu/reference_reduce.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_reduce.hpp
rename to library/include/ck/library/reference_tensor_operation/cpu/reference_reduce.hpp
diff --git a/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_softmax.hpp b/library/include/ck/library/reference_tensor_operation/cpu/reference_softmax.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_softmax.hpp
rename to library/include/ck/library/reference_tensor_operation/cpu/reference_softmax.hpp
diff --git a/composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_sparse_embedding3_forward_layernorm.hpp b/library/include/ck/library/reference_tensor_operation/cpu/reference_sparse_embedding3_forward_layernorm.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/reference_tensor_operation/cpu/reference_sparse_embedding3_forward_layernorm.hpp
rename to library/include/ck/library/reference_tensor_operation/cpu/reference_sparse_embedding3_forward_layernorm.hpp
diff --git a/composable_kernel/library/include/ck/library/reference_tensor_operation/gpu/naive_conv_fwd.hpp b/library/include/ck/library/reference_tensor_operation/gpu/naive_conv_fwd.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/reference_tensor_operation/gpu/naive_conv_fwd.hpp
rename to library/include/ck/library/reference_tensor_operation/gpu/naive_conv_fwd.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/add_device_operation_instance.hpp b/library/include/ck/library/tensor_operation_instance/add_device_operation_instance.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/add_device_operation_instance.hpp
rename to library/include/ck/library/tensor_operation_instance/add_device_operation_instance.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/device_operation_instance_factory.hpp b/library/include/ck/library/tensor_operation_instance/device_operation_instance_factory.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/device_operation_instance_factory.hpp
rename to library/include/ck/library/tensor_operation_instance/device_operation_instance_factory.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/avg_pool3d_bwd.hpp b/library/include/ck/library/tensor_operation_instance/gpu/avg_pool3d_bwd.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/avg_pool3d_bwd.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/avg_pool3d_bwd.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/batched_gemm.hpp b/library/include/ck/library/tensor_operation_instance/gpu/batched_gemm.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/batched_gemm.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/batched_gemm.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_add_relu_gemm_add.hpp b/library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_add_relu_gemm_add.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_add_relu_gemm_add.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_add_relu_gemm_add.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_bias_permute.hpp b/library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_bias_permute.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_bias_permute.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_bias_permute.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_bias_softmax_gemm_permute.hpp b/library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_bias_softmax_gemm_permute.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_bias_softmax_gemm_permute.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_bias_softmax_gemm_permute.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_gemm.hpp b/library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_gemm.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_gemm.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_gemm.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_multi_d.hpp b/library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_multi_d.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_multi_d.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_multi_d.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_softmax_gemm.hpp b/library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_softmax_gemm.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_softmax_gemm.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_softmax_gemm.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute.hpp b/library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/batchnorm_backward.hpp b/library/include/ck/library/tensor_operation_instance/gpu/batchnorm_backward.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/batchnorm_backward.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/batchnorm_backward.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/batchnorm_forward.hpp b/library/include/ck/library/tensor_operation_instance/gpu/batchnorm_forward.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/batchnorm_forward.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/batchnorm_forward.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/batchnorm_infer.hpp b/library/include/ck/library/tensor_operation_instance/gpu/batchnorm_infer.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/batchnorm_infer.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/batchnorm_infer.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/contraction_bilinear.hpp b/library/include/ck/library/tensor_operation_instance/gpu/contraction_bilinear.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/contraction_bilinear.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/contraction_bilinear.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/contraction_scale.hpp b/library/include/ck/library/tensor_operation_instance/gpu/contraction_scale.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/contraction_scale.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/contraction_scale.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/conv_tensor_rearrange.hpp b/library/include/ck/library/tensor_operation_instance/gpu/conv_tensor_rearrange.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/conv_tensor_rearrange.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/conv_tensor_rearrange.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/conv_tensor_rearrange/device_column_to_image_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/conv_tensor_rearrange/device_column_to_image_instance.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/conv_tensor_rearrange/device_column_to_image_instance.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/conv_tensor_rearrange/device_column_to_image_instance.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/conv_tensor_rearrange/device_image_to_column_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/conv_tensor_rearrange/device_image_to_column_instance.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/conv_tensor_rearrange/device_image_to_column_instance.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/conv_tensor_rearrange/device_image_to_column_instance.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/convolution_backward_data.hpp b/library/include/ck/library/tensor_operation_instance/gpu/convolution_backward_data.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/convolution_backward_data.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/convolution_backward_data.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/convolution_forward.hpp b/library/include/ck/library/tensor_operation_instance/gpu/convolution_forward.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/convolution_forward.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/convolution_forward.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/device_elementwise_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/device_elementwise_instance.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/device_elementwise_instance.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/device_elementwise_instance.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/device_gemm_mean_squaremean_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/device_gemm_mean_squaremean_instance.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/device_gemm_mean_squaremean_instance.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/device_gemm_mean_squaremean_instance.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/elementwise_normalization.hpp b/library/include/ck/library/tensor_operation_instance/gpu/elementwise_normalization.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/elementwise_normalization.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/elementwise_normalization.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/gemm.hpp b/library/include/ck/library/tensor_operation_instance/gpu/gemm.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/gemm.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/gemm.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/gemm_add_add_fastgelu.hpp b/library/include/ck/library/tensor_operation_instance/gpu/gemm_add_add_fastgelu.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/gemm_add_add_fastgelu.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/gemm_add_add_fastgelu.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/gemm_add_fastgelu.hpp b/library/include/ck/library/tensor_operation_instance/gpu/gemm_add_fastgelu.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/gemm_add_fastgelu.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/gemm_add_fastgelu.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/gemm_add_multiply.hpp b/library/include/ck/library/tensor_operation_instance/gpu/gemm_add_multiply.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/gemm_add_multiply.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/gemm_add_multiply.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/gemm_add_relu_add_layernorm.hpp b/library/include/ck/library/tensor_operation_instance/gpu/gemm_add_relu_add_layernorm.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/gemm_add_relu_add_layernorm.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/gemm_add_relu_add_layernorm.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/gemm_bilinear.hpp b/library/include/ck/library/tensor_operation_instance/gpu/gemm_bilinear.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/gemm_bilinear.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/gemm_bilinear.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/gemm_fastgelu.hpp b/library/include/ck/library/tensor_operation_instance/gpu/gemm_fastgelu.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/gemm_fastgelu.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/gemm_fastgelu.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/gemm_multiply_add.hpp b/library/include/ck/library/tensor_operation_instance/gpu/gemm_multiply_add.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/gemm_multiply_add.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/gemm_multiply_add.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/gemm_splitk.hpp b/library/include/ck/library/tensor_operation_instance/gpu/gemm_splitk.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/gemm_splitk.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/gemm_splitk.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/gemm_streamk.hpp b/library/include/ck/library/tensor_operation_instance/gpu/gemm_streamk.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/gemm_streamk.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/gemm_streamk.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/gemv_splitk.hpp b/library/include/ck/library/tensor_operation_instance/gpu/gemv_splitk.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/gemv_splitk.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/gemv_splitk.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_data/device_grouped_conv_bwd_data_wmma_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_data/device_grouped_conv_bwd_data_wmma_instance.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_data/device_grouped_conv_bwd_data_wmma_instance.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_data/device_grouped_conv_bwd_data_wmma_instance.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_data/device_grouped_conv_bwd_data_xdl_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_data/device_grouped_conv_bwd_data_xdl_instance.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_data/device_grouped_conv_bwd_data_xdl_instance.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_data/device_grouped_conv_bwd_data_xdl_instance.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_dl_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_dl_instance.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_dl_instance.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_dl_instance.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_wmma_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_wmma_instance.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_wmma_instance.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_wmma_instance.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_xdl_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_xdl_instance.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_xdl_instance.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_xdl_instance.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv2d_fwd_wmma_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv2d_fwd_wmma_instance.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv2d_fwd_wmma_instance.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv2d_fwd_wmma_instance.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_dl_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_dl_instance.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_dl_instance.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_dl_instance.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_wmma_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_wmma_instance.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_wmma_instance.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_wmma_instance.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_instance.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_instance.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_instance.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_data.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_data.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_data.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_data.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_weight.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_weight.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_weight.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_weight.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/grouped_gemm.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_gemm.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/grouped_gemm.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/grouped_gemm.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/grouped_gemm_bias.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_gemm_bias.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/grouped_gemm_bias.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/grouped_gemm_bias.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/grouped_gemm_fastgelu.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_gemm_fastgelu.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/grouped_gemm_fastgelu.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/grouped_gemm_fastgelu.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/grouped_gemm_fixed_nk.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_gemm_fixed_nk.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/grouped_gemm_fixed_nk.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/grouped_gemm_fixed_nk.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/image_to_column.hpp b/library/include/ck/library/tensor_operation_instance/gpu/image_to_column.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/image_to_column.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/image_to_column.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/max_pool_bwd.hpp b/library/include/ck/library/tensor_operation_instance/gpu/max_pool_bwd.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/max_pool_bwd.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/max_pool_bwd.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/normalization.hpp b/library/include/ck/library/tensor_operation_instance/gpu/normalization.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/normalization.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/normalization.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/normalization_swish.hpp b/library/include/ck/library/tensor_operation_instance/gpu/normalization_swish.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/normalization_swish.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/normalization_swish.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/pool3d_fwd.hpp b/library/include/ck/library/tensor_operation_instance/gpu/pool3d_fwd.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/pool3d_fwd.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/pool3d_fwd.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/quantization/gemm_quantization.hpp b/library/include/ck/library/tensor_operation_instance/gpu/quantization/gemm_quantization.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/quantization/gemm_quantization.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/quantization/gemm_quantization.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_bias_forward_perchannel_quantization.hpp b/library/include/ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_bias_forward_perchannel_quantization.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_bias_forward_perchannel_quantization.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_bias_forward_perchannel_quantization.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_bias_forward_perlayer_quantization.hpp b/library/include/ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_bias_forward_perlayer_quantization.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_bias_forward_perlayer_quantization.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_bias_forward_perlayer_quantization.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_forward_perchannel_quantization.hpp b/library/include/ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_forward_perchannel_quantization.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_forward_perchannel_quantization.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_forward_perchannel_quantization.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_forward_perlayer_quantization.hpp b/library/include/ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_forward_perlayer_quantization.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_forward_perlayer_quantization.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_forward_perlayer_quantization.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_add.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_add.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_add.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_add.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_amax.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_amax.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_amax.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_amax.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_avg.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_avg.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_avg.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_avg.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_max.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_max.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_max.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_max.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_min.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_min.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_min.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_min.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_norm2.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_norm2.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_norm2.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_norm2.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_amax.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_amax.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_amax.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_amax.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_max.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_max.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_max.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_max.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_min.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_min.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_min.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_min.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_add.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_add.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_add.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_add.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_avg.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_avg.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_avg.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_avg.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_norm2.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_norm2.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_norm2.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_norm2.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_add.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_add.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_add.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_add.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_amax.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_amax.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_amax.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_amax.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_avg.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_avg.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_avg.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_avg.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_max.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_max.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_max.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_max.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_min.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_min.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_min.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_min.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_norm2.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_norm2.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_norm2.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_norm2.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_add.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_add.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_add.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_add.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_avg.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_avg.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_avg.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_avg.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_norm2.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_norm2.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_norm2.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_norm2.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_add.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_add.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_add.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_add.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_amax.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_amax.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_amax.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_amax.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_avg.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_avg.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_avg.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_avg.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_max.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_max.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_max.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_max.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_min.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_min.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_min.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_min.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_norm2.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_norm2.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_norm2.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_norm2.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i32_i8_add.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i32_i8_add.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i32_i8_add.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i32_i8_add.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i32_i8_avg.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i32_i8_avg.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i32_i8_avg.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i32_i8_avg.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_amax.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_amax.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_amax.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_amax.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_max.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_max.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_max.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_max.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_min.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_min.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_min.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_min.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_impl_common.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_impl_common.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_impl_common.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_impl_common.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_b16_f32_f32_add.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_b16_f32_f32_add.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_b16_f32_f32_add.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_b16_f32_f32_add.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_b16_f32_f32_avg.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_b16_f32_f32_avg.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_b16_f32_f32_avg.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_b16_f32_f32_avg.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f16_f32_f32_add.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f16_f32_f32_add.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f16_f32_f32_add.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f16_f32_f32_add.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f16_f32_f32_avg.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f16_f32_f32_avg.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f16_f32_f32_avg.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f16_f32_f32_avg.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f32_f32_add.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f32_f32_add.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f32_f32_add.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f32_f32_add.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f32_f32_avg.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f32_f32_avg.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f32_f32_avg.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f32_f32_avg.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f64_f32_add.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f64_f32_add.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f64_f32_add.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f64_f32_add.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f64_f32_avg.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f64_f32_avg.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f64_f32_avg.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f64_f32_avg.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f64_f64_f64_add.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f64_f64_f64_add.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f64_f64_f64_add.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f64_f64_f64_add.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f64_f64_f64_avg.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f64_f64_f64_avg.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f64_f64_f64_avg.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f64_f64_f64_avg.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_add.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_add.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_add.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_add.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_amax.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_amax.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_amax.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_amax.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_avg.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_avg.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_avg.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_avg.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_max.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_max.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_max.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_max.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_min.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_min.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_min.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_min.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_norm2.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_norm2.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_norm2.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_norm2.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_amax.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_amax.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_amax.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_amax.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_max.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_max.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_max.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_max.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_min.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_min.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_min.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_min.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_add.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_add.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_add.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_add.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_avg.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_avg.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_avg.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_avg.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_norm2.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_norm2.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_norm2.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_norm2.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_add.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_add.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_add.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_add.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_amax.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_amax.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_amax.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_amax.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_avg.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_avg.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_avg.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_avg.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_max.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_max.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_max.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_max.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_min.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_min.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_min.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_min.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_norm2.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_norm2.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_norm2.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_norm2.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_add.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_add.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_add.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_add.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_avg.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_avg.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_avg.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_avg.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_norm2.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_norm2.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_norm2.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_norm2.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_add.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_add.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_add.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_add.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_amax.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_amax.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_amax.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_amax.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_avg.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_avg.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_avg.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_avg.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_max.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_max.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_max.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_max.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_min.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_min.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_min.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_min.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_norm2.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_norm2.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_norm2.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_norm2.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i32_i8_add.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i32_i8_add.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i32_i8_add.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i32_i8_add.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i32_i8_avg.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i32_i8_avg.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i32_i8_avg.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i32_i8_avg.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_amax.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_amax.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_amax.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_amax.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_max.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_max.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_max.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_max.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_min.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_min.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_min.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_min.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/reduce.hpp b/library/include/ck/library/tensor_operation_instance/gpu/reduce/reduce.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/reduce/reduce.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/reduce/reduce.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax.hpp b/library/include/ck/library/tensor_operation_instance/gpu/softmax.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/softmax.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce1.hpp b/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce1.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce1.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce1.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce2.hpp b/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce2.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce2.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce2.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce3.hpp b/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce3.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce3.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce3.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce1.hpp b/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce1.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce1.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce1.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce2.hpp b/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce2.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce2.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce2.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce3.hpp b/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce3.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce3.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce3.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce4.hpp b/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce4.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce4.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce4.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_type.hpp b/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_type.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_type.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_type.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce1.hpp b/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce1.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce1.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce1.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce2.hpp b/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce2.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce2.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce2.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce3.hpp b/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce3.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce3.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce3.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce1.hpp b/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce1.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce1.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce1.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce2.hpp b/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce2.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce2.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce2.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce3.hpp b/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce3.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce3.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce3.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce4.hpp b/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce4.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce4.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce4.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_type.hpp b/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_type.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_type.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_type.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_instance.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_instance.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/softmax/device_softmax_instance.hpp
diff --git a/composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/tall_and_skinny_gemm_splitk.hpp b/library/include/ck/library/tensor_operation_instance/gpu/tall_and_skinny_gemm_splitk.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/tensor_operation_instance/gpu/tall_and_skinny_gemm_splitk.hpp
rename to library/include/ck/library/tensor_operation_instance/gpu/tall_and_skinny_gemm_splitk.hpp
diff --git a/composable_kernel/library/include/ck/library/utility/algorithm.hpp b/library/include/ck/library/utility/algorithm.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/utility/algorithm.hpp
rename to library/include/ck/library/utility/algorithm.hpp
diff --git a/composable_kernel/library/include/ck/library/utility/check_err.hpp b/library/include/ck/library/utility/check_err.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/utility/check_err.hpp
rename to library/include/ck/library/utility/check_err.hpp
diff --git a/composable_kernel/library/include/ck/library/utility/conv_common.hpp b/library/include/ck/library/utility/conv_common.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/utility/conv_common.hpp
rename to library/include/ck/library/utility/conv_common.hpp
diff --git a/composable_kernel/library/include/ck/library/utility/convolution_host_tensor_descriptor_helper.hpp b/library/include/ck/library/utility/convolution_host_tensor_descriptor_helper.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/utility/convolution_host_tensor_descriptor_helper.hpp
rename to library/include/ck/library/utility/convolution_host_tensor_descriptor_helper.hpp
diff --git a/composable_kernel/library/include/ck/library/utility/convolution_parameter.hpp b/library/include/ck/library/utility/convolution_parameter.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/utility/convolution_parameter.hpp
rename to library/include/ck/library/utility/convolution_parameter.hpp
diff --git a/composable_kernel/library/include/ck/library/utility/device_memory.hpp b/library/include/ck/library/utility/device_memory.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/utility/device_memory.hpp
rename to library/include/ck/library/utility/device_memory.hpp
diff --git a/composable_kernel/library/include/ck/library/utility/fill.hpp b/library/include/ck/library/utility/fill.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/utility/fill.hpp
rename to library/include/ck/library/utility/fill.hpp
diff --git a/composable_kernel/library/include/ck/library/utility/host_common_util.hpp b/library/include/ck/library/utility/host_common_util.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/utility/host_common_util.hpp
rename to library/include/ck/library/utility/host_common_util.hpp
diff --git a/composable_kernel/library/include/ck/library/utility/host_gemm.hpp b/library/include/ck/library/utility/host_gemm.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/utility/host_gemm.hpp
rename to library/include/ck/library/utility/host_gemm.hpp
diff --git a/composable_kernel/library/include/ck/library/utility/host_tensor.hpp b/library/include/ck/library/utility/host_tensor.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/utility/host_tensor.hpp
rename to library/include/ck/library/utility/host_tensor.hpp
diff --git a/composable_kernel/library/include/ck/library/utility/host_tensor_generator.hpp b/library/include/ck/library/utility/host_tensor_generator.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/utility/host_tensor_generator.hpp
rename to library/include/ck/library/utility/host_tensor_generator.hpp
diff --git a/composable_kernel/library/include/ck/library/utility/iterator.hpp b/library/include/ck/library/utility/iterator.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/utility/iterator.hpp
rename to library/include/ck/library/utility/iterator.hpp
diff --git a/composable_kernel/library/include/ck/library/utility/literals.hpp b/library/include/ck/library/utility/literals.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/utility/literals.hpp
rename to library/include/ck/library/utility/literals.hpp
diff --git a/composable_kernel/library/include/ck/library/utility/numeric.hpp b/library/include/ck/library/utility/numeric.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/utility/numeric.hpp
rename to library/include/ck/library/utility/numeric.hpp
diff --git a/composable_kernel/library/include/ck/library/utility/ranges.hpp b/library/include/ck/library/utility/ranges.hpp
similarity index 100%
rename from composable_kernel/library/include/ck/library/utility/ranges.hpp
rename to library/include/ck/library/utility/ranges.hpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/CMakeLists.txt
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/CMakeLists.txt
rename to library/src/tensor_operation_instance/gpu/CMakeLists.txt
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/avg_pool3d_bwd/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/avg_pool3d_bwd/CMakeLists.txt
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/avg_pool3d_bwd/CMakeLists.txt
rename to library/src/tensor_operation_instance/gpu/avg_pool3d_bwd/CMakeLists.txt
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/avg_pool3d_bwd/avg_pool3d_bwd_ndhwc_instance_common.hpp b/library/src/tensor_operation_instance/gpu/avg_pool3d_bwd/avg_pool3d_bwd_ndhwc_instance_common.hpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/avg_pool3d_bwd/avg_pool3d_bwd_ndhwc_instance_common.hpp
rename to library/src/tensor_operation_instance/gpu/avg_pool3d_bwd/avg_pool3d_bwd_ndhwc_instance_common.hpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/avg_pool3d_bwd/device_avg_pool3d_bwd_ndhwc_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/avg_pool3d_bwd/device_avg_pool3d_bwd_ndhwc_bf16_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/avg_pool3d_bwd/device_avg_pool3d_bwd_ndhwc_bf16_instance.cpp
rename to library/src/tensor_operation_instance/gpu/avg_pool3d_bwd/device_avg_pool3d_bwd_ndhwc_bf16_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/avg_pool3d_bwd/device_avg_pool3d_bwd_ndhwc_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/avg_pool3d_bwd/device_avg_pool3d_bwd_ndhwc_f16_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/avg_pool3d_bwd/device_avg_pool3d_bwd_ndhwc_f16_instance.cpp
rename to library/src/tensor_operation_instance/gpu/avg_pool3d_bwd/device_avg_pool3d_bwd_ndhwc_f16_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/avg_pool3d_bwd/device_avg_pool3d_bwd_ndhwc_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/avg_pool3d_bwd/device_avg_pool3d_bwd_ndhwc_f32_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/avg_pool3d_bwd/device_avg_pool3d_bwd_ndhwc_f32_instance.cpp
rename to library/src/tensor_operation_instance/gpu/avg_pool3d_bwd/device_avg_pool3d_bwd_ndhwc_f32_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/batched_gemm/CMakeLists.txt
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/CMakeLists.txt
rename to library/src/tensor_operation_instance/gpu/batched_gemm/CMakeLists.txt
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_bf16_bf16_bf16_gkm_gkn_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_bf16_bf16_bf16_gkm_gkn_gmn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_bf16_bf16_bf16_gkm_gkn_gmn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_bf16_bf16_bf16_gkm_gkn_gmn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_bf16_bf16_bf16_gkm_gnk_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_bf16_bf16_bf16_gkm_gnk_gmn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_bf16_bf16_bf16_gkm_gnk_gmn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_bf16_bf16_bf16_gkm_gnk_gmn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_bf16_bf16_bf16_gmk_gkn_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_bf16_bf16_bf16_gmk_gkn_gmn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_bf16_bf16_bf16_gmk_gkn_gmn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_bf16_bf16_bf16_gmk_gkn_gmn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_bf16_bf16_bf16_gmk_gnk_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_bf16_bf16_bf16_gmk_gnk_gmn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_bf16_bf16_bf16_gmk_gnk_gmn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_bf16_bf16_bf16_gmk_gnk_gmn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f16_f16_f16_gkm_gkn_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f16_f16_f16_gkm_gkn_gmn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f16_f16_f16_gkm_gkn_gmn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f16_f16_f16_gkm_gkn_gmn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f16_f16_f16_gkm_gnk_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f16_f16_f16_gkm_gnk_gmn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f16_f16_f16_gkm_gnk_gmn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f16_f16_f16_gkm_gnk_gmn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f16_f16_f16_gmk_gkn_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f16_f16_f16_gmk_gkn_gmn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f16_f16_f16_gmk_gkn_gmn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f16_f16_f16_gmk_gkn_gmn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f16_f16_f16_gmk_gnk_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f16_f16_f16_gmk_gnk_gmn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f16_f16_f16_gmk_gnk_gmn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f16_f16_f16_gmk_gnk_gmn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f32_f32_f32_gkm_gkn_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f32_f32_f32_gkm_gkn_gmn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f32_f32_f32_gkm_gkn_gmn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f32_f32_f32_gkm_gkn_gmn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f32_f32_f32_gkm_gnk_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f32_f32_f32_gkm_gnk_gmn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f32_f32_f32_gkm_gnk_gmn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f32_f32_f32_gkm_gnk_gmn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f32_f32_f32_gmk_gkn_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f32_f32_f32_gmk_gkn_gmn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f32_f32_f32_gmk_gkn_gmn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f32_f32_f32_gmk_gkn_gmn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f32_f32_f32_gmk_gnk_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f32_f32_f32_gmk_gnk_gmn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f32_f32_f32_gmk_gnk_gmn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f32_f32_f32_gmk_gnk_gmn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_int8_int8_int8_gkm_gkn_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_int8_int8_int8_gkm_gkn_gmn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_int8_int8_int8_gkm_gkn_gmn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_int8_int8_int8_gkm_gkn_gmn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_int8_int8_int8_gkm_gnk_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_int8_int8_int8_gkm_gnk_gmn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_int8_int8_int8_gkm_gnk_gmn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_int8_int8_int8_gkm_gnk_gmn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_int8_int8_int8_gmk_gkn_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_int8_int8_int8_gmk_gkn_gmn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_int8_int8_int8_gmk_gkn_gmn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_int8_int8_int8_gmk_gkn_gmn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_int8_int8_int8_gmk_gnk_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_int8_int8_int8_gmk_gnk_gmn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_int8_int8_int8_gmk_gnk_gmn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_int8_int8_int8_gmk_gnk_gmn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_add_relu_gemm_add/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/batched_gemm_add_relu_gemm_add/CMakeLists.txt
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_add_relu_gemm_add/CMakeLists.txt
rename to library/src/tensor_operation_instance/gpu/batched_gemm_add_relu_gemm_add/CMakeLists.txt
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_add_relu_gemm_add/device_batched_gemm_add_relu_gemm_add_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_add_relu_gemm_add/device_batched_gemm_add_relu_gemm_add_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_add_relu_gemm_add/device_batched_gemm_add_relu_gemm_add_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp
rename to library/src/tensor_operation_instance/gpu/batched_gemm_add_relu_gemm_add/device_batched_gemm_add_relu_gemm_add_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_add_relu_gemm_add/device_batched_gemm_add_relu_gemm_add_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gon_gmo_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_add_relu_gemm_add/device_batched_gemm_add_relu_gemm_add_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gon_gmo_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_add_relu_gemm_add/device_batched_gemm_add_relu_gemm_add_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gon_gmo_instance.cpp
rename to library/src/tensor_operation_instance/gpu/batched_gemm_add_relu_gemm_add/device_batched_gemm_add_relu_gemm_add_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gon_gmo_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_bias_permute/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/batched_gemm_bias_permute/CMakeLists.txt
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_bias_permute/CMakeLists.txt
rename to library/src/tensor_operation_instance/gpu/batched_gemm_bias_permute/CMakeLists.txt
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_bias_permute/device_batched_gemm_bias_permute_m2_n3_k1_xdl_c_shuffle_f16_f16_f16_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_bias_permute/device_batched_gemm_bias_permute_m2_n3_k1_xdl_c_shuffle_f16_f16_f16_f16_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_bias_permute/device_batched_gemm_bias_permute_m2_n3_k1_xdl_c_shuffle_f16_f16_f16_f16_instance.cpp
rename to library/src/tensor_operation_instance/gpu/batched_gemm_bias_permute/device_batched_gemm_bias_permute_m2_n3_k1_xdl_c_shuffle_f16_f16_f16_f16_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_gemm/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/batched_gemm_gemm/CMakeLists.txt
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_gemm/CMakeLists.txt
rename to library/src/tensor_operation_instance/gpu/batched_gemm_gemm/CMakeLists.txt
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_gemm/device_batched_gemm_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_gemm/device_batched_gemm_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_gemm/device_batched_gemm_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp
rename to library/src/tensor_operation_instance/gpu/batched_gemm_gemm/device_batched_gemm_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_gemm/device_batched_gemm_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gon_gmo_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_gemm/device_batched_gemm_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gon_gmo_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_gemm/device_batched_gemm_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gon_gmo_instance.cpp
rename to library/src/tensor_operation_instance/gpu/batched_gemm_gemm/device_batched_gemm_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gon_gmo_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/CMakeLists.txt
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/CMakeLists.txt
rename to library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/CMakeLists.txt
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gkn_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gkn_gmn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gkn_gmn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gkn_gmn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gkn_gmn_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gkn_gmn_irregular_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gkn_gmn_irregular_instance.cpp
rename to library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gkn_gmn_irregular_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gnk_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gnk_gmn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gnk_gmn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gnk_gmn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gnk_gmn_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gnk_gmn_irregular_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gnk_gmn_irregular_instance.cpp
rename to library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gnk_gmn_irregular_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gkn_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gkn_gmn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gkn_gmn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gkn_gmn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gkn_gmn_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gkn_gmn_irregular_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gkn_gmn_irregular_instance.cpp
rename to library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gkn_gmn_irregular_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gnk_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gnk_gmn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gnk_gmn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gnk_gmn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gnk_gmn_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gnk_gmn_irregular_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gnk_gmn_irregular_instance.cpp
rename to library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gnk_gmn_irregular_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gkn_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gkn_gmn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gkn_gmn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gkn_gmn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gkn_gmn_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gkn_gmn_irregular_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gkn_gmn_irregular_instance.cpp
rename to library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gkn_gmn_irregular_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gnk_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gnk_gmn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gnk_gmn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gnk_gmn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gnk_gmn_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gnk_gmn_irregular_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gnk_gmn_irregular_instance.cpp
rename to library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gnk_gmn_irregular_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gkn_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gkn_gmn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gkn_gmn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gkn_gmn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gkn_gmn_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gkn_gmn_irregular_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gkn_gmn_irregular_instance.cpp
rename to library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gkn_gmn_irregular_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gnk_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gnk_gmn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gnk_gmn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gnk_gmn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gnk_gmn_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gnk_gmn_irregular_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gnk_gmn_irregular_instance.cpp
rename to library/src/tensor_operation_instance/gpu/batched_gemm_multi_d/device_batched_gemm_multi_d_dl_i8_i8_i8_gmk_gnk_gmn_irregular_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_reduce/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/batched_gemm_reduce/CMakeLists.txt
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_reduce/CMakeLists.txt
rename to library/src/tensor_operation_instance/gpu/batched_gemm_reduce/CMakeLists.txt
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gkn_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gkn_gmn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gkn_gmn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gkn_gmn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gnk_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gnk_gmn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gnk_gmn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gnk_gmn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gkn_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gkn_gmn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gkn_gmn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gkn_gmn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gnk_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gnk_gmn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gnk_gmn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gnk_gmn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm/CMakeLists.txt
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm/CMakeLists.txt
rename to library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm/CMakeLists.txt
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm/device_batched_gemm_softmax_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm/device_batched_gemm_softmax_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm/device_batched_gemm_softmax_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp
rename to library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm/device_batched_gemm_softmax_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/CMakeLists.txt
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/CMakeLists.txt
rename to library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/CMakeLists.txt
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instance.cpp
rename to library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp
rename to library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instance.cpp
rename to library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp
rename to library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/batchnorm/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/batchnorm/CMakeLists.txt
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/batchnorm/CMakeLists.txt
rename to library/src/tensor_operation_instance/gpu/batchnorm/CMakeLists.txt
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_backward_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_backward_bf16_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_backward_bf16_instance.cpp
rename to library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_backward_bf16_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_backward_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_backward_f16_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_backward_f16_instance.cpp
rename to library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_backward_f16_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_backward_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_backward_f32_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_backward_f32_instance.cpp
rename to library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_backward_f32_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_backward_f64_instance.cpp b/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_backward_f64_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_backward_f64_instance.cpp
rename to library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_backward_f64_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_forward_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_forward_bf16_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_forward_bf16_instance.cpp
rename to library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_forward_bf16_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_forward_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_forward_f16_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_forward_f16_instance.cpp
rename to library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_forward_f16_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_forward_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_forward_f32_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_forward_f32_instance.cpp
rename to library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_forward_f32_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_forward_f64_instance.cpp b/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_forward_f64_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_forward_f64_instance.cpp
rename to library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_forward_f64_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_infer_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_infer_bf16_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_infer_bf16_instance.cpp
rename to library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_infer_bf16_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_infer_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_infer_f16_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_infer_f16_instance.cpp
rename to library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_infer_f16_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_infer_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_infer_f32_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_infer_f32_instance.cpp
rename to library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_infer_f32_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_infer_f64_instance.cpp b/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_infer_f64_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_infer_f64_instance.cpp
rename to library/src/tensor_operation_instance/gpu/batchnorm/device_batchnorm_infer_f64_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/column_to_image/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/column_to_image/CMakeLists.txt
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/column_to_image/CMakeLists.txt
rename to library/src/tensor_operation_instance/gpu/column_to_image/CMakeLists.txt
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/column_to_image/device_column_to_image_nhwc_1d_instance.cpp b/library/src/tensor_operation_instance/gpu/column_to_image/device_column_to_image_nhwc_1d_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/column_to_image/device_column_to_image_nhwc_1d_instance.cpp
rename to library/src/tensor_operation_instance/gpu/column_to_image/device_column_to_image_nhwc_1d_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/column_to_image/device_column_to_image_nhwc_2d_instance.cpp b/library/src/tensor_operation_instance/gpu/column_to_image/device_column_to_image_nhwc_2d_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/column_to_image/device_column_to_image_nhwc_2d_instance.cpp
rename to library/src/tensor_operation_instance/gpu/column_to_image/device_column_to_image_nhwc_2d_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/column_to_image/device_column_to_image_nhwc_3d_instance.cpp b/library/src/tensor_operation_instance/gpu/column_to_image/device_column_to_image_nhwc_3d_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/column_to_image/device_column_to_image_nhwc_3d_instance.cpp
rename to library/src/tensor_operation_instance/gpu/column_to_image/device_column_to_image_nhwc_3d_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/contraction_bilinear/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/contraction_bilinear/CMakeLists.txt
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/contraction_bilinear/CMakeLists.txt
rename to library/src/tensor_operation_instance/gpu/contraction_bilinear/CMakeLists.txt
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_kknn_instance.cpp b/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_kknn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_kknn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_kknn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_knnn_instance.cpp b/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_knnn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_knnn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_knnn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mknn_instance.cpp b/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mknn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mknn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mknn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mnnn_instance.cpp b/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mnnn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mnnn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_f32_mnnn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_kknn_instance.cpp b/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_kknn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_kknn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_kknn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_knnn_instance.cpp b/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_knnn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_knnn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_knnn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_mknn_instance.cpp b/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_mknn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_mknn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_mknn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_mnnn_instance.cpp b/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_mnnn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_mnnn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/contraction_bilinear/device_contraction_bilinear_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_f64_mnnn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/contraction_scale/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/contraction_scale/CMakeLists.txt
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/contraction_scale/CMakeLists.txt
rename to library/src/tensor_operation_instance/gpu/contraction_scale/CMakeLists.txt
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_kkn_instance.cpp b/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_kkn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_kkn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_kkn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_knn_instance.cpp b/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_knn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_knn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_knn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mkn_instance.cpp b/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mkn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mkn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mkn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mnn_instance.cpp b/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mnn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mnn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f32_f32_f32_mnn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_kkn_instance.cpp b/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_kkn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_kkn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_kkn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_knn_instance.cpp b/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_knn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_knn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_knn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_mkn_instance.cpp b/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_mkn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_mkn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_mkn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_mnn_instance.cpp b/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_mnn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_mnn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/contraction_scale/device_contraction_scale_m2_n2_k2_xdl_c_shuffle_f64_f64_f64_mnn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/conv1d_bwd_data/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/conv1d_bwd_data/CMakeLists.txt
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/conv1d_bwd_data/CMakeLists.txt
rename to library/src/tensor_operation_instance/gpu/conv1d_bwd_data/CMakeLists.txt
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/conv1d_bwd_data/device_conv1d_bwd_data_xdl_nwc_kxc_nwk_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/conv1d_bwd_data/device_conv1d_bwd_data_xdl_nwc_kxc_nwk_bf16_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/conv1d_bwd_data/device_conv1d_bwd_data_xdl_nwc_kxc_nwk_bf16_instance.cpp
rename to library/src/tensor_operation_instance/gpu/conv1d_bwd_data/device_conv1d_bwd_data_xdl_nwc_kxc_nwk_bf16_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/conv1d_bwd_data/device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/conv1d_bwd_data/device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f16_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/conv1d_bwd_data/device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f16_instance.cpp
rename to library/src/tensor_operation_instance/gpu/conv1d_bwd_data/device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f16_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/conv1d_bwd_data/device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/conv1d_bwd_data/device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f32_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/conv1d_bwd_data/device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f32_instance.cpp
rename to library/src/tensor_operation_instance/gpu/conv1d_bwd_data/device_conv1d_bwd_data_xdl_nwc_kxc_nwk_f32_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/conv1d_bwd_data/device_conv1d_bwd_data_xdl_nwc_kxc_nwk_int8_instance.cpp b/library/src/tensor_operation_instance/gpu/conv1d_bwd_data/device_conv1d_bwd_data_xdl_nwc_kxc_nwk_int8_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/conv1d_bwd_data/device_conv1d_bwd_data_xdl_nwc_kxc_nwk_int8_instance.cpp
rename to library/src/tensor_operation_instance/gpu/conv1d_bwd_data/device_conv1d_bwd_data_xdl_nwc_kxc_nwk_int8_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/CMakeLists.txt
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/CMakeLists.txt
rename to library/src/tensor_operation_instance/gpu/conv2d_bwd_data/CMakeLists.txt
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_f16_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_f16_instance.cpp
rename to library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_f16_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_f32_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_f32_instance.cpp
rename to library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_f32_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_int8_instance.cpp b/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_int8_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_int8_instance.cpp
rename to library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_dl_nhwc_kyxc_nhwk_int8_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_bf16_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_bf16_instance.cpp
rename to library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_bf16_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f16_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f16_instance.cpp
rename to library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f16_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f32_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f32_instance.cpp
rename to library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f32_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_int8_instance.cpp b/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_int8_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_int8_instance.cpp
rename to library/src/tensor_operation_instance/gpu/conv2d_bwd_data/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_int8_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_fwd/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/conv2d_fwd/CMakeLists.txt
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_fwd/CMakeLists.txt
rename to library/src/tensor_operation_instance/gpu/conv2d_fwd/CMakeLists.txt
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_fwd/device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/conv2d_fwd/device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_f16_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_fwd/device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_f16_instance.cpp
rename to library/src/tensor_operation_instance/gpu/conv2d_fwd/device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_f16_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_fwd/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/conv2d_fwd/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_bf16_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_fwd/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_bf16_instance.cpp
rename to library/src/tensor_operation_instance/gpu/conv2d_fwd/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_bf16_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_fwd/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/conv2d_fwd/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f16_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_fwd/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f16_instance.cpp
rename to library/src/tensor_operation_instance/gpu/conv2d_fwd/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f16_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_fwd/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/conv2d_fwd/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f32_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_fwd/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f32_instance.cpp
rename to library/src/tensor_operation_instance/gpu/conv2d_fwd/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_f32_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_fwd/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_int8_instance.cpp b/library/src/tensor_operation_instance/gpu/conv2d_fwd/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_int8_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_fwd/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_int8_instance.cpp
rename to library/src/tensor_operation_instance/gpu/conv2d_fwd/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_int8_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_fwd_bias_relu/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/conv2d_fwd_bias_relu/CMakeLists.txt
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_fwd_bias_relu/CMakeLists.txt
rename to library/src/tensor_operation_instance/gpu/conv2d_fwd_bias_relu/CMakeLists.txt
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_fwd_bias_relu/device_conv2d_fwd_xdl_c_shuffle_bias_relu_nhwc_kyxc_nhwk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/conv2d_fwd_bias_relu/device_conv2d_fwd_xdl_c_shuffle_bias_relu_nhwc_kyxc_nhwk_f16_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_fwd_bias_relu/device_conv2d_fwd_xdl_c_shuffle_bias_relu_nhwc_kyxc_nhwk_f16_instance.cpp
rename to library/src/tensor_operation_instance/gpu/conv2d_fwd_bias_relu/device_conv2d_fwd_xdl_c_shuffle_bias_relu_nhwc_kyxc_nhwk_f16_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_fwd_bias_relu_add/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/conv2d_fwd_bias_relu_add/CMakeLists.txt
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_fwd_bias_relu_add/CMakeLists.txt
rename to library/src/tensor_operation_instance/gpu/conv2d_fwd_bias_relu_add/CMakeLists.txt
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_fwd_bias_relu_add/device_conv2d_fwd_xdl_c_shuffle_bias_relu_add_nhwc_kyxc_nhwk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/conv2d_fwd_bias_relu_add/device_conv2d_fwd_xdl_c_shuffle_bias_relu_add_nhwc_kyxc_nhwk_f16_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/conv2d_fwd_bias_relu_add/device_conv2d_fwd_xdl_c_shuffle_bias_relu_add_nhwc_kyxc_nhwk_f16_instance.cpp
rename to library/src/tensor_operation_instance/gpu/conv2d_fwd_bias_relu_add/device_conv2d_fwd_xdl_c_shuffle_bias_relu_add_nhwc_kyxc_nhwk_f16_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/conv3d_bwd_data/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/conv3d_bwd_data/CMakeLists.txt
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/conv3d_bwd_data/CMakeLists.txt
rename to library/src/tensor_operation_instance/gpu/conv3d_bwd_data/CMakeLists.txt
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/conv3d_bwd_data/device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/conv3d_bwd_data/device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_bf16_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/conv3d_bwd_data/device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_bf16_instance.cpp
rename to library/src/tensor_operation_instance/gpu/conv3d_bwd_data/device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_bf16_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/conv3d_bwd_data/device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/conv3d_bwd_data/device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f16_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/conv3d_bwd_data/device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f16_instance.cpp
rename to library/src/tensor_operation_instance/gpu/conv3d_bwd_data/device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f16_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/conv3d_bwd_data/device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/conv3d_bwd_data/device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f32_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/conv3d_bwd_data/device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f32_instance.cpp
rename to library/src/tensor_operation_instance/gpu/conv3d_bwd_data/device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_f32_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/conv3d_bwd_data/device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_int8_instance.cpp b/library/src/tensor_operation_instance/gpu/conv3d_bwd_data/device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_int8_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/conv3d_bwd_data/device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_int8_instance.cpp
rename to library/src/tensor_operation_instance/gpu/conv3d_bwd_data/device_conv3d_bwd_data_xdl_ndhwc_kzyxc_ndhwk_int8_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/elementwise/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/elementwise/CMakeLists.txt
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/elementwise/CMakeLists.txt
rename to library/src/tensor_operation_instance/gpu/elementwise/CMakeLists.txt
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/elementwise/device_normalize_instance.cpp b/library/src/tensor_operation_instance/gpu/elementwise/device_normalize_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/elementwise/device_normalize_instance.cpp
rename to library/src/tensor_operation_instance/gpu/elementwise/device_normalize_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/elementwise_normalization/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/elementwise_normalization/CMakeLists.txt
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/elementwise_normalization/CMakeLists.txt
rename to library/src/tensor_operation_instance/gpu/elementwise_normalization/CMakeLists.txt
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/elementwise_normalization/device_elementwise_normalization_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/elementwise_normalization/device_elementwise_normalization_f16_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/elementwise_normalization/device_elementwise_normalization_f16_instance.cpp
rename to library/src/tensor_operation_instance/gpu/elementwise_normalization/device_elementwise_normalization_f16_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/gemm/CMakeLists.txt
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/CMakeLists.txt
rename to library/src/tensor_operation_instance/gpu/gemm/CMakeLists.txt
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_km_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_km_kn_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_km_kn_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_km_kn_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_km_kn_mn_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_km_kn_mn_irregular_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_km_kn_mn_irregular_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_km_kn_mn_irregular_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_km_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_km_nk_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_km_nk_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_km_nk_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_km_nk_mn_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_km_nk_mn_irregular_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_km_nk_mn_irregular_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_km_nk_mn_irregular_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_mk_kn_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_mk_kn_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_mk_kn_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_mk_kn_mn_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_mk_kn_mn_irregular_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_mk_kn_mn_irregular_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_mk_kn_mn_irregular_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_mk_nk_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_mk_nk_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_mk_nk_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_mk_nk_mn_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_mk_nk_mn_irregular_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_mk_nk_mn_irregular_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_mk_nk_mn_irregular_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f32_f32_f32_km_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f32_f32_f32_km_kn_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f32_f32_f32_km_kn_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f32_f32_f32_km_kn_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f32_f32_f32_km_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f32_f32_f32_km_nk_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f32_f32_f32_km_nk_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f32_f32_f32_km_nk_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f32_f32_f32_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f32_f32_f32_mk_kn_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f32_f32_f32_mk_kn_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f32_f32_f32_mk_kn_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f32_f32_f32_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f32_f32_f32_mk_nk_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f32_f32_f32_mk_nk_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f32_f32_f32_mk_nk_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_km_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_km_kn_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_km_kn_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_km_kn_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_km_kn_mn_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_km_kn_mn_irregular_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_km_kn_mn_irregular_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_km_kn_mn_irregular_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_km_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_km_nk_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_km_nk_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_km_nk_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_km_nk_mn_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_km_nk_mn_irregular_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_km_nk_mn_irregular_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_km_nk_mn_irregular_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_mk_kn_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_mk_kn_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_mk_kn_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_mk_kn_mn_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_mk_kn_mn_irregular_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_mk_kn_mn_irregular_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_mk_kn_mn_irregular_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_mk_nk_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_mk_nk_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_mk_nk_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_mk_nk_mn_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_mk_nk_mn_irregular_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_mk_nk_mn_irregular_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_i8_i8_i8_mk_nk_mn_irregular_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dpp_f16_f16_f16_km_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dpp_f16_f16_f16_km_kn_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dpp_f16_f16_f16_km_kn_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm/device_gemm_dpp_f16_f16_f16_km_kn_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dpp_f16_f16_f16_km_kn_mn_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dpp_f16_f16_f16_km_kn_mn_irregular_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dpp_f16_f16_f16_km_kn_mn_irregular_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm/device_gemm_dpp_f16_f16_f16_km_kn_mn_irregular_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dpp_f16_f16_f16_km_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dpp_f16_f16_f16_km_nk_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dpp_f16_f16_f16_km_nk_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm/device_gemm_dpp_f16_f16_f16_km_nk_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dpp_f16_f16_f16_km_nk_mn_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dpp_f16_f16_f16_km_nk_mn_irregular_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dpp_f16_f16_f16_km_nk_mn_irregular_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm/device_gemm_dpp_f16_f16_f16_km_nk_mn_irregular_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dpp_f16_f16_f16_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dpp_f16_f16_f16_mk_kn_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dpp_f16_f16_f16_mk_kn_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm/device_gemm_dpp_f16_f16_f16_mk_kn_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dpp_f16_f16_f16_mk_kn_mn_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dpp_f16_f16_f16_mk_kn_mn_irregular_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dpp_f16_f16_f16_mk_kn_mn_irregular_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm/device_gemm_dpp_f16_f16_f16_mk_kn_mn_irregular_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dpp_f16_f16_f16_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dpp_f16_f16_f16_mk_nk_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dpp_f16_f16_f16_mk_nk_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm/device_gemm_dpp_f16_f16_f16_mk_nk_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dpp_f16_f16_f16_mk_nk_mn_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dpp_f16_f16_f16_mk_nk_mn_irregular_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dpp_f16_f16_f16_mk_nk_mn_irregular_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm/device_gemm_dpp_f16_f16_f16_mk_nk_mn_irregular_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_2_stage_f16_f16_f16_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_2_stage_f16_f16_f16_mk_nk_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_2_stage_f16_f16_f16_mk_nk_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_2_stage_f16_f16_f16_mk_nk_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_kn_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_kn_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_kn_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_nk_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_nk_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_nk_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_kn_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_kn_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_kn_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_nk_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_nk_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_nk_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f32_f32_f32_km_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f32_f32_f32_km_kn_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f32_f32_f32_km_kn_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f32_f32_f32_km_kn_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f32_f32_f32_km_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f32_f32_f32_km_nk_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f32_f32_f32_km_nk_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f32_f32_f32_km_nk_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f32_f32_f32_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f32_f32_f32_mk_kn_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f32_f32_f32_mk_kn_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f32_f32_f32_mk_kn_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f32_f32_f32_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f32_f32_f32_mk_nk_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f32_f32_f32_mk_nk_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_f32_f32_f32_mk_nk_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_km_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_km_kn_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_km_kn_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_km_kn_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_km_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_km_nk_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_km_nk_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_km_nk_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_mk_kn_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_mk_kn_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_mk_kn_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_mk_nk_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_mk_nk_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_fp8_fp8_fp8_mk_nk_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_i8_i8_i8_km_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_i8_i8_i8_km_nk_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_i8_i8_i8_km_nk_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_i8_i8_i8_km_nk_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_i8_i8_i8_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_i8_i8_i8_mk_kn_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_i8_i8_i8_mk_kn_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_i8_i8_i8_mk_kn_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/common.hpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/common.hpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/common.hpp
rename to library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/common.hpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_add_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_add_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_add_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_add_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_default_pipeline_v1_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_default_pipeline_v1_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_default_pipeline_v1_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_default_pipeline_v1_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_default_pipeline_v2_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_default_pipeline_v2_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_default_pipeline_v2_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_default_pipeline_v2_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_default_pipeline_v2_opt_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_default_pipeline_v2_opt_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_default_pipeline_v2_opt_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_default_pipeline_v2_opt_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_interwave_pipeline_v1_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_interwave_pipeline_v1_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_interwave_pipeline_v1_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_interwave_pipeline_v1_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_irregular_default_pipeline_v1_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_irregular_default_pipeline_v1_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_irregular_default_pipeline_v1_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_irregular_default_pipeline_v1_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_irregular_default_pipeline_v2_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_irregular_default_pipeline_v2_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_irregular_default_pipeline_v2_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_irregular_default_pipeline_v2_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_irregular_interwave_pipeline_v1_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_irregular_interwave_pipeline_v1_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_irregular_interwave_pipeline_v1_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_kn_mn_irregular_interwave_pipeline_v1_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_add_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_add_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_add_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_add_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_default_pipeline_v1_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_default_pipeline_v1_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_default_pipeline_v1_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_default_pipeline_v1_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_default_pipeline_v2_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_default_pipeline_v2_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_default_pipeline_v2_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_default_pipeline_v2_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_default_pipeline_v2_opt_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_default_pipeline_v2_opt_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_default_pipeline_v2_opt_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_default_pipeline_v2_opt_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_interwave_pipeline_v1_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_interwave_pipeline_v1_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_interwave_pipeline_v1_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_interwave_pipeline_v1_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_irregular_default_pipeline_v1_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_irregular_default_pipeline_v1_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_irregular_default_pipeline_v1_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_irregular_default_pipeline_v1_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_irregular_default_pipeline_v2_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_irregular_default_pipeline_v2_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_irregular_default_pipeline_v2_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_irregular_default_pipeline_v2_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_irregular_interwave_pipeline_v1_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_irregular_interwave_pipeline_v1_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_irregular_interwave_pipeline_v1_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/km_nk_mn_irregular_interwave_pipeline_v1_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_add_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_add_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_add_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_add_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_default_pipeline_v1_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_default_pipeline_v1_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_default_pipeline_v1_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_default_pipeline_v1_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_default_pipeline_v2_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_default_pipeline_v2_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_default_pipeline_v2_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_default_pipeline_v2_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_default_pipeline_v2_opt_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_default_pipeline_v2_opt_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_default_pipeline_v2_opt_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_default_pipeline_v2_opt_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_interwave_pipeline_v1_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_interwave_pipeline_v1_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_interwave_pipeline_v1_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_interwave_pipeline_v1_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_irregular_default_pipeline_v1_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_irregular_default_pipeline_v1_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_irregular_default_pipeline_v1_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_irregular_default_pipeline_v1_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_irregular_default_pipeline_v2_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_irregular_default_pipeline_v2_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_irregular_default_pipeline_v2_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_irregular_default_pipeline_v2_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_irregular_interwave_pipeline_v1_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_irregular_interwave_pipeline_v1_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_irregular_interwave_pipeline_v1_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_kn_mn_irregular_interwave_pipeline_v1_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_add_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_add_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_add_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_add_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_default_pipeline_v1_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_default_pipeline_v1_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_default_pipeline_v1_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_default_pipeline_v1_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_default_pipeline_v2_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_default_pipeline_v2_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_default_pipeline_v2_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_default_pipeline_v2_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_default_pipeline_v2_opt_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_default_pipeline_v2_opt_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_default_pipeline_v2_opt_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_default_pipeline_v2_opt_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_interwave_pipeline_v1_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_interwave_pipeline_v1_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_interwave_pipeline_v1_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_interwave_pipeline_v1_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_irregular_default_pipeline_v1_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_irregular_default_pipeline_v1_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_irregular_default_pipeline_v1_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_irregular_default_pipeline_v1_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_irregular_default_pipeline_v2_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_irregular_default_pipeline_v2_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_irregular_default_pipeline_v2_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_irregular_default_pipeline_v2_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_irregular_interwave_pipeline_v1_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_irregular_interwave_pipeline_v1_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_irregular_interwave_pipeline_v1_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f16_f16_f16/mk_nk_mn_irregular_interwave_pipeline_v1_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f32_f32_f32_km_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f32_f32_f32_km_kn_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f32_f32_f32_km_kn_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f32_f32_f32_km_kn_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f32_f32_f32_km_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f32_f32_f32_km_nk_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f32_f32_f32_km_nk_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f32_f32_f32_km_nk_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f32_f32_f32_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f32_f32_f32_mk_kn_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f32_f32_f32_mk_kn_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f32_f32_f32_mk_kn_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f32_f32_f32_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f32_f32_f32_mk_nk_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f32_f32_f32_mk_nk_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f32_f32_f32_mk_nk_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f64_f64_f64_km_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f64_f64_f64_km_kn_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f64_f64_f64_km_kn_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f64_f64_f64_km_kn_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f64_f64_f64_km_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f64_f64_f64_km_nk_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f64_f64_f64_km_nk_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f64_f64_f64_km_nk_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f64_f64_f64_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f64_f64_f64_mk_kn_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f64_f64_f64_mk_kn_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f64_f64_f64_mk_kn_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f64_f64_f64_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f64_f64_f64_mk_nk_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f64_f64_f64_mk_nk_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f64_f64_f64_mk_nk_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/CMakeLists.txt
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/CMakeLists.txt
rename to library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/CMakeLists.txt
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/CMakeLists.txt
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/CMakeLists.txt
rename to library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/CMakeLists.txt
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_add_multiply/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/gemm_add_multiply/CMakeLists.txt
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm_add_multiply/CMakeLists.txt
rename to library/src/tensor_operation_instance/gpu/gemm_add_multiply/CMakeLists.txt
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_add_multiply/device_gemm_add_multiply_xdl_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_add_multiply/device_gemm_add_multiply_xdl_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm_add_multiply/device_gemm_add_multiply_xdl_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm_add_multiply/device_gemm_add_multiply_xdl_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_add_multiply/device_gemm_add_multiply_xdl_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_add_multiply/device_gemm_add_multiply_xdl_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm_add_multiply/device_gemm_add_multiply_xdl_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm_add_multiply/device_gemm_add_multiply_xdl_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_add_multiply/device_gemm_add_multiply_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_add_multiply/device_gemm_add_multiply_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm_add_multiply/device_gemm_add_multiply_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm_add_multiply/device_gemm_add_multiply_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_add_multiply/device_gemm_add_multiply_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_add_multiply/device_gemm_add_multiply_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm_add_multiply/device_gemm_add_multiply_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm_add_multiply/device_gemm_add_multiply_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_add_relu_add_layernorm/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/gemm_add_relu_add_layernorm/CMakeLists.txt
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm_add_relu_add_layernorm/CMakeLists.txt
rename to library/src/tensor_operation_instance/gpu/gemm_add_relu_add_layernorm/CMakeLists.txt
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_add_relu_add_layernorm/device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_kn_mn_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_add_relu_add_layernorm/device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_kn_mn_mn_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm_add_relu_add_layernorm/device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_kn_mn_mn_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm_add_relu_add_layernorm/device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_kn_mn_mn_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_add_relu_add_layernorm/device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_nk_mn_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_add_relu_add_layernorm/device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_nk_mn_mn_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm_add_relu_add_layernorm/device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_nk_mn_mn_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm_add_relu_add_layernorm/device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_km_nk_mn_mn_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_add_relu_add_layernorm/device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_kn_mn_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_add_relu_add_layernorm/device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_kn_mn_mn_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm_add_relu_add_layernorm/device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_kn_mn_mn_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm_add_relu_add_layernorm/device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_kn_mn_mn_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_add_relu_add_layernorm/device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_nk_mn_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_add_relu_add_layernorm/device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_nk_mn_mn_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm_add_relu_add_layernorm/device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_nk_mn_mn_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm_add_relu_add_layernorm/device_gemm_add_relu_add_xdl_c_shuffle_layernorm_f16_mk_nk_mn_mn_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_bias_add_reduce/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/gemm_bias_add_reduce/CMakeLists.txt
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm_bias_add_reduce/CMakeLists.txt
rename to library/src/tensor_operation_instance/gpu/gemm_bias_add_reduce/CMakeLists.txt
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_bias_add_reduce/device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f32_f32_km_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_bias_add_reduce/device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f32_f32_km_kn_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm_bias_add_reduce/device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f32_f32_km_kn_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm_bias_add_reduce/device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f32_f32_km_kn_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_bias_add_reduce/device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f32_f32_km_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_bias_add_reduce/device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f32_f32_km_nk_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm_bias_add_reduce/device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f32_f32_km_nk_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm_bias_add_reduce/device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f32_f32_km_nk_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_bias_add_reduce/device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f32_f32_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_bias_add_reduce/device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f32_f32_mk_kn_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm_bias_add_reduce/device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f32_f32_mk_kn_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm_bias_add_reduce/device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f32_f32_mk_kn_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_bias_add_reduce/device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f32_f32_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_bias_add_reduce/device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f32_f32_mk_nk_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm_bias_add_reduce/device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f32_f32_mk_nk_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm_bias_add_reduce/device_gemm_bias_add_mean_squaremean_xdl_cshuffle_f16_f16_f16_f32_f32_mk_nk_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_bilinear/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/gemm_bilinear/CMakeLists.txt
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm_bilinear/CMakeLists.txt
rename to library/src/tensor_operation_instance/gpu/gemm_bilinear/CMakeLists.txt
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_wmma_c_shuffle_i8_i8_i8_i8_km_kn_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_wmma_c_shuffle_i8_i8_i8_i8_km_kn_mn_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_wmma_c_shuffle_i8_i8_i8_i8_km_kn_mn_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_wmma_c_shuffle_i8_i8_i8_i8_km_kn_mn_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_wmma_c_shuffle_i8_i8_i8_i8_km_nk_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_wmma_c_shuffle_i8_i8_i8_i8_km_nk_mn_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_wmma_c_shuffle_i8_i8_i8_i8_km_nk_mn_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_wmma_c_shuffle_i8_i8_i8_i8_km_nk_mn_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_wmma_c_shuffle_i8_i8_i8_i8_mk_kn_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_wmma_c_shuffle_i8_i8_i8_i8_mk_kn_mn_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_wmma_c_shuffle_i8_i8_i8_i8_mk_kn_mn_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_wmma_c_shuffle_i8_i8_i8_i8_mk_kn_mn_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_wmma_c_shuffle_i8_i8_i8_i8_mk_nk_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_wmma_c_shuffle_i8_i8_i8_i8_mk_nk_mn_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_wmma_c_shuffle_i8_i8_i8_i8_mk_nk_mn_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_wmma_c_shuffle_i8_i8_i8_i8_mk_nk_mn_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm_bilinear/device_gemm_bilinear_xdl_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_fastgelu/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/gemm_fastgelu/CMakeLists.txt
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm_fastgelu/CMakeLists.txt
rename to library/src/tensor_operation_instance/gpu/gemm_fastgelu/CMakeLists.txt
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_multiply_add/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/gemm_multiply_add/CMakeLists.txt
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm_multiply_add/CMakeLists.txt
rename to library/src/tensor_operation_instance/gpu/gemm_multiply_add/CMakeLists.txt
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_multiply_add/device_gemm_multiply_add_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_multiply_add/device_gemm_multiply_add_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm_multiply_add/device_gemm_multiply_add_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm_multiply_add/device_gemm_multiply_add_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_multiply_add/device_gemm_multiply_add_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_multiply_add/device_gemm_multiply_add_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm_multiply_add/device_gemm_multiply_add_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm_multiply_add/device_gemm_multiply_add_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_multiply_add/device_gemm_multiply_add_xdl_c_shuffle_f16_f8_f32_f32_f16_mk_kn_mn_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_multiply_add/device_gemm_multiply_add_xdl_c_shuffle_f16_f8_f32_f32_f16_mk_kn_mn_mn_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm_multiply_add/device_gemm_multiply_add_xdl_c_shuffle_f16_f8_f32_f32_f16_mk_kn_mn_mn_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm_multiply_add/device_gemm_multiply_add_xdl_c_shuffle_f16_f8_f32_f32_f16_mk_kn_mn_mn_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_multiply_add/device_gemm_multiply_add_xdl_c_shuffle_f16_f8_f32_f32_f16_mk_nk_mn_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_multiply_add/device_gemm_multiply_add_xdl_c_shuffle_f16_f8_f32_f32_f16_mk_nk_mn_mn_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm_multiply_add/device_gemm_multiply_add_xdl_c_shuffle_f16_f8_f32_f32_f16_mk_nk_mn_mn_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm_multiply_add/device_gemm_multiply_add_xdl_c_shuffle_f16_f8_f32_f32_f16_mk_nk_mn_mn_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_reduce/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/gemm_reduce/CMakeLists.txt
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm_reduce/CMakeLists.txt
rename to library/src/tensor_operation_instance/gpu/gemm_reduce/CMakeLists.txt
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_kn_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_kn_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_kn_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_nk_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_nk_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_nk_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_kn_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_kn_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_kn_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_nk_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_nk_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_nk_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_splitk/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/gemm_splitk/CMakeLists.txt
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm_splitk/CMakeLists.txt
rename to library/src/tensor_operation_instance/gpu/gemm_splitk/CMakeLists.txt
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_comp_fp8_km_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_comp_fp8_km_kn_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_comp_fp8_km_kn_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_comp_fp8_km_kn_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_comp_fp8_km_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_comp_fp8_km_nk_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_comp_fp8_km_nk_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_comp_fp8_km_nk_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_comp_fp8_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_comp_fp8_mk_kn_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_comp_fp8_mk_kn_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_comp_fp8_mk_kn_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_comp_fp8_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_comp_fp8_mk_nk_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_comp_fp8_mk_nk_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_comp_fp8_mk_nk_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_km_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_km_kn_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_km_kn_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_km_kn_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_km_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_km_nk_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_km_nk_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_km_nk_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_fp8_f16_km_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_fp8_f16_km_kn_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_fp8_f16_km_kn_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_fp8_f16_km_kn_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_fp8_f16_km_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_fp8_f16_km_nk_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_fp8_f16_km_nk_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_fp8_f16_km_nk_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_fp8_f16_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_fp8_f16_mk_kn_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_fp8_f16_mk_kn_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_fp8_f16_mk_kn_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_fp8_f16_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_fp8_f16_mk_nk_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_fp8_f16_mk_nk_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_fp8_f16_mk_nk_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f32_f32_f32_km_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f32_f32_f32_km_kn_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f32_f32_f32_km_kn_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f32_f32_f32_km_kn_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f32_f32_f32_km_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f32_f32_f32_km_nk_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f32_f32_f32_km_nk_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f32_f32_f32_km_nk_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f32_f32_f32_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f32_f32_f32_mk_kn_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f32_f32_f32_mk_kn_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f32_f32_f32_mk_kn_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f32_f32_f32_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f32_f32_f32_mk_nk_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f32_f32_f32_mk_nk_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f32_f32_f32_mk_nk_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_fp8_f16_f16_km_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_fp8_f16_f16_km_kn_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_fp8_f16_f16_km_kn_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_fp8_f16_f16_km_kn_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_fp8_f16_f16_km_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_fp8_f16_f16_km_nk_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_fp8_f16_f16_km_nk_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_fp8_f16_f16_km_nk_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_fp8_f16_f16_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_fp8_f16_f16_mk_kn_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_fp8_f16_f16_mk_kn_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_fp8_f16_f16_mk_kn_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_fp8_f16_f16_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_fp8_f16_f16_mk_nk_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_fp8_f16_f16_mk_nk_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_fp8_f16_f16_mk_nk_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_streamk/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/gemm_streamk/CMakeLists.txt
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm_streamk/CMakeLists.txt
rename to library/src/tensor_operation_instance/gpu/gemm_streamk/CMakeLists.txt
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemm_streamk/device_gemm_xdl_streamk_f16_f16_f16_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_streamk/device_gemm_xdl_streamk_f16_f16_f16_mk_kn_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemm_streamk/device_gemm_xdl_streamk_f16_f16_f16_mk_kn_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemm_streamk/device_gemm_xdl_streamk_f16_f16_f16_mk_kn_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemv_splitk/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/gemv_splitk/CMakeLists.txt
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemv_splitk/CMakeLists.txt
rename to library/src/tensor_operation_instance/gpu/gemv_splitk/CMakeLists.txt
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemv_splitk/device_gemv_splitk_f16_f16_f16_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemv_splitk/device_gemv_splitk_f16_f16_f16_mk_kn_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemv_splitk/device_gemv_splitk_f16_f16_f16_mk_kn_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemv_splitk/device_gemv_splitk_f16_f16_f16_mk_kn_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/gemv_splitk/device_gemv_splitk_f16_f16_f16_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemv_splitk/device_gemv_splitk_f16_f16_f16_mk_nk_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/gemv_splitk/device_gemv_splitk_f16_f16_f16_mk_nk_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/gemv_splitk/device_gemv_splitk_f16_f16_f16_mk_nk_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/CMakeLists.txt
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/CMakeLists.txt
rename to library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/CMakeLists.txt
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/dl/device_grouped_conv1d_bwd_weight_dl_gnwc_gkxc_gnwk_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/dl/device_grouped_conv1d_bwd_weight_dl_gnwc_gkxc_gnwk_bf16_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/dl/device_grouped_conv1d_bwd_weight_dl_gnwc_gkxc_gnwk_bf16_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/dl/device_grouped_conv1d_bwd_weight_dl_gnwc_gkxc_gnwk_bf16_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/dl/device_grouped_conv1d_bwd_weight_dl_gnwc_gkxc_gnwk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/dl/device_grouped_conv1d_bwd_weight_dl_gnwc_gkxc_gnwk_f16_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/dl/device_grouped_conv1d_bwd_weight_dl_gnwc_gkxc_gnwk_f16_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/dl/device_grouped_conv1d_bwd_weight_dl_gnwc_gkxc_gnwk_f16_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/dl/device_grouped_conv1d_bwd_weight_dl_gnwc_gkxc_gnwk_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/dl/device_grouped_conv1d_bwd_weight_dl_gnwc_gkxc_gnwk_f32_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/dl/device_grouped_conv1d_bwd_weight_dl_gnwc_gkxc_gnwk_f32_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/dl/device_grouped_conv1d_bwd_weight_dl_gnwc_gkxc_gnwk_f32_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/dl/device_grouped_conv1d_bwd_weight_dl_nwgc_gkxc_nwgk_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/dl/device_grouped_conv1d_bwd_weight_dl_nwgc_gkxc_nwgk_bf16_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/dl/device_grouped_conv1d_bwd_weight_dl_nwgc_gkxc_nwgk_bf16_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/dl/device_grouped_conv1d_bwd_weight_dl_nwgc_gkxc_nwgk_bf16_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/dl/device_grouped_conv1d_bwd_weight_dl_nwgc_gkxc_nwgk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/dl/device_grouped_conv1d_bwd_weight_dl_nwgc_gkxc_nwgk_f16_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/dl/device_grouped_conv1d_bwd_weight_dl_nwgc_gkxc_nwgk_f16_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/dl/device_grouped_conv1d_bwd_weight_dl_nwgc_gkxc_nwgk_f16_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/dl/device_grouped_conv1d_bwd_weight_dl_nwgc_gkxc_nwgk_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/dl/device_grouped_conv1d_bwd_weight_dl_nwgc_gkxc_nwgk_f32_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/dl/device_grouped_conv1d_bwd_weight_dl_nwgc_gkxc_nwgk_f32_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/dl/device_grouped_conv1d_bwd_weight_dl_nwgc_gkxc_nwgk_f32_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/xdl/device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/xdl/device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_bf16_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/xdl/device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_bf16_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/xdl/device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_bf16_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/xdl/device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/xdl/device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_f16_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/xdl/device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_f16_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/xdl/device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_f16_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/xdl/device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/xdl/device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_f32_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/xdl/device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_f32_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/xdl/device_grouped_conv1d_bwd_weight_xdl_gnwc_gkxc_gnwk_f32_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv1d_fwd/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/grouped_conv1d_fwd/CMakeLists.txt
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv1d_fwd/CMakeLists.txt
rename to library/src/tensor_operation_instance/gpu/grouped_conv1d_fwd/CMakeLists.txt
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv1d_fwd/xdl/device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv1d_fwd/xdl/device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_bf16_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv1d_fwd/xdl/device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_bf16_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv1d_fwd/xdl/device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_bf16_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv1d_fwd/xdl/device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv1d_fwd/xdl/device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_f16_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv1d_fwd/xdl/device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_f16_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv1d_fwd/xdl/device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_f16_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv1d_fwd/xdl/device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv1d_fwd/xdl/device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_f32_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv1d_fwd/xdl/device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_f32_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv1d_fwd/xdl/device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_f32_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv1d_fwd/xdl/device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_int8_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv1d_fwd/xdl/device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_int8_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv1d_fwd/xdl/device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_int8_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv1d_fwd/xdl/device_grouped_conv1d_fwd_xdl_gnwc_gkxc_gnwk_int8_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/CMakeLists.txt
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/CMakeLists.txt
rename to library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/CMakeLists.txt
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/wmma/device_grouped_conv2d_bwd_data_wmma_gnhwc_gkyxc_gnhwk_f16_1x1s1p0_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/wmma/device_grouped_conv2d_bwd_data_wmma_gnhwc_gkyxc_gnhwk_f16_1x1s1p0_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/wmma/device_grouped_conv2d_bwd_data_wmma_gnhwc_gkyxc_gnhwk_f16_1x1s1p0_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/wmma/device_grouped_conv2d_bwd_data_wmma_gnhwc_gkyxc_gnhwk_f16_1x1s1p0_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/wmma/device_grouped_conv2d_bwd_data_wmma_gnhwc_gkyxc_gnhwk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/wmma/device_grouped_conv2d_bwd_data_wmma_gnhwc_gkyxc_gnhwk_f16_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/wmma/device_grouped_conv2d_bwd_data_wmma_gnhwc_gkyxc_gnhwk_f16_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/wmma/device_grouped_conv2d_bwd_data_wmma_gnhwc_gkyxc_gnhwk_f16_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/wmma/device_grouped_conv2d_bwd_data_wmma_gnhwc_gkyxc_gnhwk_i8_1x1s1p0_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/wmma/device_grouped_conv2d_bwd_data_wmma_gnhwc_gkyxc_gnhwk_i8_1x1s1p0_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/wmma/device_grouped_conv2d_bwd_data_wmma_gnhwc_gkyxc_gnhwk_i8_1x1s1p0_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/wmma/device_grouped_conv2d_bwd_data_wmma_gnhwc_gkyxc_gnhwk_i8_1x1s1p0_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/wmma/device_grouped_conv2d_bwd_data_wmma_gnhwc_gkyxc_gnhwk_i8_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/wmma/device_grouped_conv2d_bwd_data_wmma_gnhwc_gkyxc_gnhwk_i8_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/wmma/device_grouped_conv2d_bwd_data_wmma_gnhwc_gkyxc_gnhwk_i8_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/wmma/device_grouped_conv2d_bwd_data_wmma_gnhwc_gkyxc_gnhwk_i8_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/wmma/device_grouped_conv2d_bwd_data_wmma_nhwgc_gkyxc_nhwgk_f16_1x1s1p0_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/wmma/device_grouped_conv2d_bwd_data_wmma_nhwgc_gkyxc_nhwgk_f16_1x1s1p0_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/wmma/device_grouped_conv2d_bwd_data_wmma_nhwgc_gkyxc_nhwgk_f16_1x1s1p0_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/wmma/device_grouped_conv2d_bwd_data_wmma_nhwgc_gkyxc_nhwgk_f16_1x1s1p0_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/wmma/device_grouped_conv2d_bwd_data_wmma_nhwgc_gkyxc_nhwgk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/wmma/device_grouped_conv2d_bwd_data_wmma_nhwgc_gkyxc_nhwgk_f16_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/wmma/device_grouped_conv2d_bwd_data_wmma_nhwgc_gkyxc_nhwgk_f16_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/wmma/device_grouped_conv2d_bwd_data_wmma_nhwgc_gkyxc_nhwgk_f16_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/wmma/device_grouped_conv2d_bwd_data_wmma_nhwgc_gkyxc_nhwgk_i8_1x1s1p0_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/wmma/device_grouped_conv2d_bwd_data_wmma_nhwgc_gkyxc_nhwgk_i8_1x1s1p0_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/wmma/device_grouped_conv2d_bwd_data_wmma_nhwgc_gkyxc_nhwgk_i8_1x1s1p0_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/wmma/device_grouped_conv2d_bwd_data_wmma_nhwgc_gkyxc_nhwgk_i8_1x1s1p0_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/wmma/device_grouped_conv2d_bwd_data_wmma_nhwgc_gkyxc_nhwgk_i8_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/wmma/device_grouped_conv2d_bwd_data_wmma_nhwgc_gkyxc_nhwgk_i8_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/wmma/device_grouped_conv2d_bwd_data_wmma_nhwgc_gkyxc_nhwgk_i8_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/wmma/device_grouped_conv2d_bwd_data_wmma_nhwgc_gkyxc_nhwgk_i8_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/xdl/device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/xdl/device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/xdl/device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/xdl/device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/xdl/device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/xdl/device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_f16_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/xdl/device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_f16_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/xdl/device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_f16_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/xdl/device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/xdl/device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_f32_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/xdl/device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_f32_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/xdl/device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_f32_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/xdl/device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/xdl/device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/xdl/device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/xdl/device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/xdl/device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/xdl/device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_f16_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/xdl/device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_f16_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/xdl/device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_f16_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/xdl/device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/xdl/device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_f32_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/xdl/device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_f32_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/xdl/device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_f32_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/CMakeLists.txt
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/CMakeLists.txt
rename to library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/CMakeLists.txt
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/dl/device_grouped_conv2d_bwd_weight_dl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/dl/device_grouped_conv2d_bwd_weight_dl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/dl/device_grouped_conv2d_bwd_weight_dl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/dl/device_grouped_conv2d_bwd_weight_dl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/dl/device_grouped_conv2d_bwd_weight_dl_gnhwc_gkyxc_gnhwk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/dl/device_grouped_conv2d_bwd_weight_dl_gnhwc_gkyxc_gnhwk_f16_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/dl/device_grouped_conv2d_bwd_weight_dl_gnhwc_gkyxc_gnhwk_f16_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/dl/device_grouped_conv2d_bwd_weight_dl_gnhwc_gkyxc_gnhwk_f16_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/dl/device_grouped_conv2d_bwd_weight_dl_gnhwc_gkyxc_gnhwk_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/dl/device_grouped_conv2d_bwd_weight_dl_gnhwc_gkyxc_gnhwk_f32_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/dl/device_grouped_conv2d_bwd_weight_dl_gnhwc_gkyxc_gnhwk_f32_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/dl/device_grouped_conv2d_bwd_weight_dl_gnhwc_gkyxc_gnhwk_f32_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/dl/device_grouped_conv2d_bwd_weight_dl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/dl/device_grouped_conv2d_bwd_weight_dl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/dl/device_grouped_conv2d_bwd_weight_dl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/dl/device_grouped_conv2d_bwd_weight_dl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/dl/device_grouped_conv2d_bwd_weight_dl_nhwgc_gkyxc_nhwgk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/dl/device_grouped_conv2d_bwd_weight_dl_nhwgc_gkyxc_nhwgk_f16_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/dl/device_grouped_conv2d_bwd_weight_dl_nhwgc_gkyxc_nhwgk_f16_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/dl/device_grouped_conv2d_bwd_weight_dl_nhwgc_gkyxc_nhwgk_f16_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/dl/device_grouped_conv2d_bwd_weight_dl_nhwgc_gkyxc_nhwgk_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/dl/device_grouped_conv2d_bwd_weight_dl_nhwgc_gkyxc_nhwgk_f32_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/dl/device_grouped_conv2d_bwd_weight_dl_nhwgc_gkyxc_nhwgk_f32_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/dl/device_grouped_conv2d_bwd_weight_dl_nhwgc_gkyxc_nhwgk_f32_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/xdl/device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/xdl/device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/xdl/device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/xdl/device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/xdl/device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/xdl/device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f16_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/xdl/device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f16_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/xdl/device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f16_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/xdl/device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/xdl/device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f32_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/xdl/device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f32_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/xdl/device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f32_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/xdl/device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/xdl/device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/xdl/device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/xdl/device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/xdl/device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/xdl/device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f16_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/xdl/device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f16_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/xdl/device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f16_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/xdl/device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/xdl/device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f32_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/xdl/device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f32_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/xdl/device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f32_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/CMakeLists.txt
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/CMakeLists.txt
rename to library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/CMakeLists.txt
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_f16_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_f16_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_f16_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_i8_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_i8_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_i8_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_i8_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/dl/device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/dl/device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f16_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/dl/device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f16_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/dl/device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f16_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/dl/device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/dl/device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f32_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/dl/device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f32_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/dl/device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f32_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/dl/device_grouped_conv2d_fwd_dl_nhwgc_gkyxc_nhwgk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/dl/device_grouped_conv2d_fwd_dl_nhwgc_gkyxc_nhwgk_f16_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/dl/device_grouped_conv2d_fwd_dl_nhwgc_gkyxc_nhwgk_f16_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/dl/device_grouped_conv2d_fwd_dl_nhwgc_gkyxc_nhwgk_f16_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/dl/device_grouped_conv2d_fwd_dl_nhwgc_gkyxc_nhwgk_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/dl/device_grouped_conv2d_fwd_dl_nhwgc_gkyxc_nhwgk_f32_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/dl/device_grouped_conv2d_fwd_dl_nhwgc_gkyxc_nhwgk_f32_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/dl/device_grouped_conv2d_fwd_dl_nhwgc_gkyxc_nhwgk_f32_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_f16_1x1p0_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_f16_1x1p0_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_f16_1x1p0_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_f16_1x1p0_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_f16_1x1s1p0_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_f16_1x1s1p0_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_f16_1x1s1p0_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_f16_1x1s1p0_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_f16_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_f16_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_f16_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_f16_oddc_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_f16_oddc_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_f16_oddc_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_f16_oddc_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_i8_1x1p0_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_i8_1x1p0_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_i8_1x1p0_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_i8_1x1p0_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_i8_1x1s1p0_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_i8_1x1s1p0_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_i8_1x1s1p0_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_i8_1x1s1p0_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_i8_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_i8_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_i8_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_i8_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_i8_oddc_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_i8_oddc_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_i8_oddc_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_gnhwc_gkyxc_gnhwk_i8_oddc_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_f16_1x1p0_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_f16_1x1p0_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_f16_1x1p0_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_f16_1x1p0_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_f16_1x1s1p0_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_f16_1x1s1p0_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_f16_1x1s1p0_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_f16_1x1s1p0_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_f16_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_f16_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_f16_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_f16_oddc_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_f16_oddc_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_f16_oddc_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_f16_oddc_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_i8_1x1p0_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_i8_1x1p0_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_i8_1x1p0_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_i8_1x1p0_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_i8_1x1s1p0_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_i8_1x1s1p0_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_i8_1x1s1p0_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_i8_1x1s1p0_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_i8_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_i8_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_i8_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_i8_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_i8_oddc_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_i8_oddc_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_i8_oddc_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/wmma/device_grouped_conv2d_fwd_wmma_nhwgc_gkyxc_nhwgk_i8_oddc_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f16_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f16_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f16_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f32_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f32_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f32_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/CMakeLists.txt
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/CMakeLists.txt
rename to library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/CMakeLists.txt
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/wmma/device_grouped_conv3d_bwd_data_wmma_gndhwc_gkzyxc_gndhwk_f16_1x1s1p0_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/wmma/device_grouped_conv3d_bwd_data_wmma_gndhwc_gkzyxc_gndhwk_f16_1x1s1p0_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/wmma/device_grouped_conv3d_bwd_data_wmma_gndhwc_gkzyxc_gndhwk_f16_1x1s1p0_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/wmma/device_grouped_conv3d_bwd_data_wmma_gndhwc_gkzyxc_gndhwk_f16_1x1s1p0_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/wmma/device_grouped_conv3d_bwd_data_wmma_gndhwc_gkzyxc_gndhwk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/wmma/device_grouped_conv3d_bwd_data_wmma_gndhwc_gkzyxc_gndhwk_f16_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/wmma/device_grouped_conv3d_bwd_data_wmma_gndhwc_gkzyxc_gndhwk_f16_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/wmma/device_grouped_conv3d_bwd_data_wmma_gndhwc_gkzyxc_gndhwk_f16_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/wmma/device_grouped_conv3d_bwd_data_wmma_gndhwc_gkzyxc_gndhwk_i8_1x1s1p0_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/wmma/device_grouped_conv3d_bwd_data_wmma_gndhwc_gkzyxc_gndhwk_i8_1x1s1p0_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/wmma/device_grouped_conv3d_bwd_data_wmma_gndhwc_gkzyxc_gndhwk_i8_1x1s1p0_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/wmma/device_grouped_conv3d_bwd_data_wmma_gndhwc_gkzyxc_gndhwk_i8_1x1s1p0_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/wmma/device_grouped_conv3d_bwd_data_wmma_gndhwc_gkzyxc_gndhwk_i8_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/wmma/device_grouped_conv3d_bwd_data_wmma_gndhwc_gkzyxc_gndhwk_i8_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/wmma/device_grouped_conv3d_bwd_data_wmma_gndhwc_gkzyxc_gndhwk_i8_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/wmma/device_grouped_conv3d_bwd_data_wmma_gndhwc_gkzyxc_gndhwk_i8_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/wmma/device_grouped_conv3d_bwd_data_wmma_ndhwgc_gkzyxc_ndhwgk_f16_1x1s1p0_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/wmma/device_grouped_conv3d_bwd_data_wmma_ndhwgc_gkzyxc_ndhwgk_f16_1x1s1p0_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/wmma/device_grouped_conv3d_bwd_data_wmma_ndhwgc_gkzyxc_ndhwgk_f16_1x1s1p0_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/wmma/device_grouped_conv3d_bwd_data_wmma_ndhwgc_gkzyxc_ndhwgk_f16_1x1s1p0_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/wmma/device_grouped_conv3d_bwd_data_wmma_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/wmma/device_grouped_conv3d_bwd_data_wmma_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/wmma/device_grouped_conv3d_bwd_data_wmma_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/wmma/device_grouped_conv3d_bwd_data_wmma_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/wmma/device_grouped_conv3d_bwd_data_wmma_ndhwgc_gkzyxc_ndhwgk_i8_1x1s1p0_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/wmma/device_grouped_conv3d_bwd_data_wmma_ndhwgc_gkzyxc_ndhwgk_i8_1x1s1p0_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/wmma/device_grouped_conv3d_bwd_data_wmma_ndhwgc_gkzyxc_ndhwgk_i8_1x1s1p0_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/wmma/device_grouped_conv3d_bwd_data_wmma_ndhwgc_gkzyxc_ndhwgk_i8_1x1s1p0_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/wmma/device_grouped_conv3d_bwd_data_wmma_ndhwgc_gkzyxc_ndhwgk_i8_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/wmma/device_grouped_conv3d_bwd_data_wmma_ndhwgc_gkzyxc_ndhwgk_i8_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/wmma/device_grouped_conv3d_bwd_data_wmma_ndhwgc_gkzyxc_ndhwgk_i8_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/wmma/device_grouped_conv3d_bwd_data_wmma_ndhwgc_gkzyxc_ndhwgk_i8_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/xdl/device_grouped_conv3d_bwd_data_xdl_gndhwc_gkzyxc_gndhwk_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/xdl/device_grouped_conv3d_bwd_data_xdl_gndhwc_gkzyxc_gndhwk_bf16_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/xdl/device_grouped_conv3d_bwd_data_xdl_gndhwc_gkzyxc_gndhwk_bf16_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/xdl/device_grouped_conv3d_bwd_data_xdl_gndhwc_gkzyxc_gndhwk_bf16_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/xdl/device_grouped_conv3d_bwd_data_xdl_gndhwc_gkzyxc_gndhwk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/xdl/device_grouped_conv3d_bwd_data_xdl_gndhwc_gkzyxc_gndhwk_f16_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/xdl/device_grouped_conv3d_bwd_data_xdl_gndhwc_gkzyxc_gndhwk_f16_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/xdl/device_grouped_conv3d_bwd_data_xdl_gndhwc_gkzyxc_gndhwk_f16_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/xdl/device_grouped_conv3d_bwd_data_xdl_gndhwc_gkzyxc_gndhwk_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/xdl/device_grouped_conv3d_bwd_data_xdl_gndhwc_gkzyxc_gndhwk_f32_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/xdl/device_grouped_conv3d_bwd_data_xdl_gndhwc_gkzyxc_gndhwk_f32_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/xdl/device_grouped_conv3d_bwd_data_xdl_gndhwc_gkzyxc_gndhwk_f32_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/xdl/device_grouped_conv3d_bwd_data_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/xdl/device_grouped_conv3d_bwd_data_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/xdl/device_grouped_conv3d_bwd_data_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/xdl/device_grouped_conv3d_bwd_data_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/xdl/device_grouped_conv3d_bwd_data_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/xdl/device_grouped_conv3d_bwd_data_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/xdl/device_grouped_conv3d_bwd_data_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/xdl/device_grouped_conv3d_bwd_data_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/xdl/device_grouped_conv3d_bwd_data_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/xdl/device_grouped_conv3d_bwd_data_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/xdl/device_grouped_conv3d_bwd_data_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/xdl/device_grouped_conv3d_bwd_data_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/xdl/device_grouped_conv3d_bwd_data_xdl_ndhwgc_gkzyxc_ndhwgk_input_f16_comp_bf8_f8_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/xdl/device_grouped_conv3d_bwd_data_xdl_ndhwgc_gkzyxc_ndhwgk_input_f16_comp_bf8_f8_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/xdl/device_grouped_conv3d_bwd_data_xdl_ndhwgc_gkzyxc_ndhwgk_input_f16_comp_bf8_f8_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/xdl/device_grouped_conv3d_bwd_data_xdl_ndhwgc_gkzyxc_ndhwgk_input_f16_comp_bf8_f8_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/CMakeLists.txt
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/CMakeLists.txt
rename to library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/CMakeLists.txt
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/dl/device_grouped_conv3d_bwd_weight_dl_gndhwc_gkzyxc_gndhwk_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/dl/device_grouped_conv3d_bwd_weight_dl_gndhwc_gkzyxc_gndhwk_bf16_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/dl/device_grouped_conv3d_bwd_weight_dl_gndhwc_gkzyxc_gndhwk_bf16_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/dl/device_grouped_conv3d_bwd_weight_dl_gndhwc_gkzyxc_gndhwk_bf16_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/dl/device_grouped_conv3d_bwd_weight_dl_gndhwc_gkzyxc_gndhwk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/dl/device_grouped_conv3d_bwd_weight_dl_gndhwc_gkzyxc_gndhwk_f16_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/dl/device_grouped_conv3d_bwd_weight_dl_gndhwc_gkzyxc_gndhwk_f16_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/dl/device_grouped_conv3d_bwd_weight_dl_gndhwc_gkzyxc_gndhwk_f16_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/dl/device_grouped_conv3d_bwd_weight_dl_gndhwc_gkzyxc_gndhwk_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/dl/device_grouped_conv3d_bwd_weight_dl_gndhwc_gkzyxc_gndhwk_f32_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/dl/device_grouped_conv3d_bwd_weight_dl_gndhwc_gkzyxc_gndhwk_f32_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/dl/device_grouped_conv3d_bwd_weight_dl_gndhwc_gkzyxc_gndhwk_f32_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/dl/device_grouped_conv3d_bwd_weight_dl_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/dl/device_grouped_conv3d_bwd_weight_dl_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/dl/device_grouped_conv3d_bwd_weight_dl_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/dl/device_grouped_conv3d_bwd_weight_dl_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/dl/device_grouped_conv3d_bwd_weight_dl_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/dl/device_grouped_conv3d_bwd_weight_dl_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/dl/device_grouped_conv3d_bwd_weight_dl_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/dl/device_grouped_conv3d_bwd_weight_dl_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/dl/device_grouped_conv3d_bwd_weight_dl_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/dl/device_grouped_conv3d_bwd_weight_dl_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/dl/device_grouped_conv3d_bwd_weight_dl_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/dl/device_grouped_conv3d_bwd_weight_dl_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/device_grouped_conv3d_bwd_weight_wmma_gndhwc_gkzyxc_gndhwk_f16_1x1s1p0_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/device_grouped_conv3d_bwd_weight_wmma_gndhwc_gkzyxc_gndhwk_f16_1x1s1p0_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/device_grouped_conv3d_bwd_weight_wmma_gndhwc_gkzyxc_gndhwk_f16_1x1s1p0_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/device_grouped_conv3d_bwd_weight_wmma_gndhwc_gkzyxc_gndhwk_f16_1x1s1p0_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/device_grouped_conv3d_bwd_weight_wmma_gndhwc_gkzyxc_gndhwk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/device_grouped_conv3d_bwd_weight_wmma_gndhwc_gkzyxc_gndhwk_f16_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/device_grouped_conv3d_bwd_weight_wmma_gndhwc_gkzyxc_gndhwk_f16_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/device_grouped_conv3d_bwd_weight_wmma_gndhwc_gkzyxc_gndhwk_f16_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/device_grouped_conv3d_bwd_weight_wmma_gndhwc_gkzyxc_gndhwk_i8_1x1s1p0_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/device_grouped_conv3d_bwd_weight_wmma_gndhwc_gkzyxc_gndhwk_i8_1x1s1p0_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/device_grouped_conv3d_bwd_weight_wmma_gndhwc_gkzyxc_gndhwk_i8_1x1s1p0_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/device_grouped_conv3d_bwd_weight_wmma_gndhwc_gkzyxc_gndhwk_i8_1x1s1p0_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/device_grouped_conv3d_bwd_weight_wmma_gndhwc_gkzyxc_gndhwk_i8_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/device_grouped_conv3d_bwd_weight_wmma_gndhwc_gkzyxc_gndhwk_i8_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/device_grouped_conv3d_bwd_weight_wmma_gndhwc_gkzyxc_gndhwk_i8_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/device_grouped_conv3d_bwd_weight_wmma_gndhwc_gkzyxc_gndhwk_i8_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_f16_1x1s1p0_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_f16_1x1s1p0_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_f16_1x1s1p0_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_f16_1x1s1p0_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_i8_1x1s1p0_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_i8_1x1s1p0_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_i8_1x1s1p0_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_i8_1x1s1p0_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_i8_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_i8_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_i8_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_i8_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/xdl/device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/xdl/device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_bf16_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/xdl/device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_bf16_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/xdl/device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_bf16_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/xdl/device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/xdl/device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_f16_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/xdl/device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_f16_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/xdl/device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_f16_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/xdl/device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/xdl/device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_f32_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/xdl/device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_f32_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/xdl/device_grouped_conv3d_bwd_weight_xdl_gndhwc_gkzyxc_gndhwk_f32_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/xdl/device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/xdl/device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/xdl/device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/xdl/device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/xdl/device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f16_comp_bf8_fp8_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/xdl/device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f16_comp_bf8_fp8_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/xdl/device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f16_comp_bf8_fp8_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/xdl/device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f16_comp_bf8_fp8_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/xdl/device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/xdl/device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/xdl/device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/xdl/device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/xdl/device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/xdl/device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/xdl/device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/xdl/device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/CMakeLists.txt
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/CMakeLists.txt
rename to library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/CMakeLists.txt
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_f16_1x1p0_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_f16_1x1p0_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_f16_1x1p0_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_f16_1x1p0_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_f16_1x1s1p0_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_f16_1x1s1p0_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_f16_1x1s1p0_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_f16_1x1s1p0_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_f16_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_f16_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_f16_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_f16_oddc_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_f16_oddc_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_f16_oddc_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_f16_oddc_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_i8_1x1p0_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_i8_1x1p0_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_i8_1x1p0_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_i8_1x1p0_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_i8_1x1s1p0_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_i8_1x1s1p0_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_i8_1x1s1p0_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_i8_1x1s1p0_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_i8_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_i8_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_i8_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_i8_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_i8_oddc_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_i8_oddc_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_i8_oddc_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_gndhwc_gkzyxc_gndhwk_i8_oddc_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_f16_1x1p0_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_f16_1x1p0_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_f16_1x1p0_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_f16_1x1p0_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_f16_1x1s1p0_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_f16_1x1s1p0_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_f16_1x1s1p0_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_f16_1x1s1p0_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_f16_oddc_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_f16_oddc_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_f16_oddc_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_f16_oddc_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_i8_1x1p0_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_i8_1x1p0_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_i8_1x1p0_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_i8_1x1p0_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_i8_1x1s1p0_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_i8_1x1s1p0_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_i8_1x1s1p0_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_i8_1x1s1p0_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_i8_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_i8_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_i8_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_i8_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_i8_oddc_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_i8_oddc_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_i8_oddc_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/wmma/device_grouped_conv3d_fwd_wmma_ndhwgc_gkzyxc_ndhwgk_i8_oddc_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/xdl/device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/xdl/device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_bf16_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/xdl/device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_bf16_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/xdl/device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_bf16_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/xdl/device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/xdl/device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_f16_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/xdl/device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_f16_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/xdl/device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_f16_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/xdl/device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/xdl/device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_f32_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/xdl/device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_f32_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/xdl/device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_f32_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/xdl/device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_int8_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/xdl/device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_int8_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/xdl/device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_int8_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/xdl/device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_int8_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/xdl/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/xdl/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/xdl/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/xdl/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/xdl/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f16_comp_fp8_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/xdl/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f16_comp_fp8_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/xdl/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f16_comp_fp8_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/xdl/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f16_comp_fp8_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/xdl/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/xdl/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/xdl/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/xdl/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/xdl/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/xdl/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/xdl/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/xdl/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/xdl/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_int8_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/xdl/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_int8_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/xdl/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_int8_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/xdl/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_int8_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_gemm/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/grouped_gemm/CMakeLists.txt
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_gemm/CMakeLists.txt
rename to library/src/tensor_operation_instance/gpu/grouped_gemm/CMakeLists.txt
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_f16_f16_f16_km_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_f16_f16_f16_km_kn_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_f16_f16_f16_km_kn_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_f16_f16_f16_km_kn_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_f16_f16_f16_km_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_f16_f16_f16_km_nk_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_f16_f16_f16_km_nk_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_f16_f16_f16_km_nk_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_f16_f16_f16_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_f16_f16_f16_mk_kn_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_f16_f16_f16_mk_kn_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_f16_f16_f16_mk_kn_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_f16_f16_f16_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_f16_f16_f16_mk_nk_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_f16_f16_f16_mk_nk_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_f16_f16_f16_mk_nk_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_irregular_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_irregular_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_irregular_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_irregular_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_irregular_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_irregular_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_gemm_bias/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/grouped_gemm_bias/CMakeLists.txt
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_gemm_bias/CMakeLists.txt
rename to library/src/tensor_operation_instance/gpu/grouped_gemm_bias/CMakeLists.txt
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_gemm_bias/device_grouped_gemm_xdl_fixed_nk_bias_f16_f16_f16_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_gemm_bias/device_grouped_gemm_xdl_fixed_nk_bias_f16_f16_f16_mk_kn_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_gemm_bias/device_grouped_gemm_xdl_fixed_nk_bias_f16_f16_f16_mk_kn_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_gemm_bias/device_grouped_gemm_xdl_fixed_nk_bias_f16_f16_f16_mk_kn_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_gemm_bias/device_grouped_gemm_xdl_fixed_nk_bias_f16_f16_f16_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_gemm_bias/device_grouped_gemm_xdl_fixed_nk_bias_f16_f16_f16_mk_nk_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_gemm_bias/device_grouped_gemm_xdl_fixed_nk_bias_f16_f16_f16_mk_nk_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_gemm_bias/device_grouped_gemm_xdl_fixed_nk_bias_f16_f16_f16_mk_nk_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_gemm_bias/device_grouped_gemm_xdl_fixed_nk_bias_f16_f16_f32_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_gemm_bias/device_grouped_gemm_xdl_fixed_nk_bias_f16_f16_f32_mk_kn_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_gemm_bias/device_grouped_gemm_xdl_fixed_nk_bias_f16_f16_f32_mk_kn_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_gemm_bias/device_grouped_gemm_xdl_fixed_nk_bias_f16_f16_f32_mk_kn_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_gemm_bias/device_grouped_gemm_xdl_fixed_nk_bias_f16_f16_f32_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_gemm_bias/device_grouped_gemm_xdl_fixed_nk_bias_f16_f16_f32_mk_nk_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_gemm_bias/device_grouped_gemm_xdl_fixed_nk_bias_f16_f16_f32_mk_nk_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_gemm_bias/device_grouped_gemm_xdl_fixed_nk_bias_f16_f16_f32_mk_nk_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_gemm_fastgelu/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/grouped_gemm_fastgelu/CMakeLists.txt
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_gemm_fastgelu/CMakeLists.txt
rename to library/src/tensor_operation_instance/gpu/grouped_gemm_fastgelu/CMakeLists.txt
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_gemm_fastgelu/device_grouped_gemm_fastgelu_xdl_f16_f16_f16_km_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_gemm_fastgelu/device_grouped_gemm_fastgelu_xdl_f16_f16_f16_km_kn_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_gemm_fastgelu/device_grouped_gemm_fastgelu_xdl_f16_f16_f16_km_kn_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_gemm_fastgelu/device_grouped_gemm_fastgelu_xdl_f16_f16_f16_km_kn_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_gemm_fastgelu/device_grouped_gemm_fastgelu_xdl_f16_f16_f16_km_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_gemm_fastgelu/device_grouped_gemm_fastgelu_xdl_f16_f16_f16_km_nk_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_gemm_fastgelu/device_grouped_gemm_fastgelu_xdl_f16_f16_f16_km_nk_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_gemm_fastgelu/device_grouped_gemm_fastgelu_xdl_f16_f16_f16_km_nk_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_gemm_fastgelu/device_grouped_gemm_fastgelu_xdl_f16_f16_f16_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_gemm_fastgelu/device_grouped_gemm_fastgelu_xdl_f16_f16_f16_mk_kn_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_gemm_fastgelu/device_grouped_gemm_fastgelu_xdl_f16_f16_f16_mk_kn_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_gemm_fastgelu/device_grouped_gemm_fastgelu_xdl_f16_f16_f16_mk_kn_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_gemm_fastgelu/device_grouped_gemm_fastgelu_xdl_f16_f16_f16_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_gemm_fastgelu/device_grouped_gemm_fastgelu_xdl_f16_f16_f16_mk_nk_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_gemm_fastgelu/device_grouped_gemm_fastgelu_xdl_f16_f16_f16_mk_nk_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_gemm_fastgelu/device_grouped_gemm_fastgelu_xdl_f16_f16_f16_mk_nk_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_gemm_fixed_nk/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/grouped_gemm_fixed_nk/CMakeLists.txt
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_gemm_fixed_nk/CMakeLists.txt
rename to library/src/tensor_operation_instance/gpu/grouped_gemm_fixed_nk/CMakeLists.txt
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_gemm_fixed_nk/device_grouped_gemm_xdl_fixed_nk_f16_f16_f16_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_gemm_fixed_nk/device_grouped_gemm_xdl_fixed_nk_f16_f16_f16_mk_kn_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_gemm_fixed_nk/device_grouped_gemm_xdl_fixed_nk_f16_f16_f16_mk_kn_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_gemm_fixed_nk/device_grouped_gemm_xdl_fixed_nk_f16_f16_f16_mk_kn_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_gemm_fixed_nk/device_grouped_gemm_xdl_fixed_nk_f16_f16_f16_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_gemm_fixed_nk/device_grouped_gemm_xdl_fixed_nk_f16_f16_f16_mk_nk_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_gemm_fixed_nk/device_grouped_gemm_xdl_fixed_nk_f16_f16_f16_mk_nk_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_gemm_fixed_nk/device_grouped_gemm_xdl_fixed_nk_f16_f16_f16_mk_nk_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_gemm_fixed_nk/device_grouped_gemm_xdl_fixed_nk_f16_fp8_f16_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_gemm_fixed_nk/device_grouped_gemm_xdl_fixed_nk_f16_fp8_f16_mk_kn_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_gemm_fixed_nk/device_grouped_gemm_xdl_fixed_nk_f16_fp8_f16_mk_kn_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_gemm_fixed_nk/device_grouped_gemm_xdl_fixed_nk_f16_fp8_f16_mk_kn_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_gemm_fixed_nk/device_grouped_gemm_xdl_fixed_nk_f16_fp8_f16_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_gemm_fixed_nk/device_grouped_gemm_xdl_fixed_nk_f16_fp8_f16_mk_nk_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_gemm_fixed_nk/device_grouped_gemm_xdl_fixed_nk_f16_fp8_f16_mk_nk_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_gemm_fixed_nk/device_grouped_gemm_xdl_fixed_nk_f16_fp8_f16_mk_nk_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_gemm_fixed_nk/device_grouped_gemm_xdl_fixed_nk_f16_i8_f16_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_gemm_fixed_nk/device_grouped_gemm_xdl_fixed_nk_f16_i8_f16_mk_kn_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_gemm_fixed_nk/device_grouped_gemm_xdl_fixed_nk_f16_i8_f16_mk_kn_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_gemm_fixed_nk/device_grouped_gemm_xdl_fixed_nk_f16_i8_f16_mk_kn_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/grouped_gemm_fixed_nk/device_grouped_gemm_xdl_fixed_nk_f16_i8_f16_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_gemm_fixed_nk/device_grouped_gemm_xdl_fixed_nk_f16_i8_f16_mk_nk_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/grouped_gemm_fixed_nk/device_grouped_gemm_xdl_fixed_nk_f16_i8_f16_mk_nk_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/grouped_gemm_fixed_nk/device_grouped_gemm_xdl_fixed_nk_f16_i8_f16_mk_nk_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/image_to_column/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/image_to_column/CMakeLists.txt
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/image_to_column/CMakeLists.txt
rename to library/src/tensor_operation_instance/gpu/image_to_column/CMakeLists.txt
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/image_to_column/device_image_to_column_nhwc_1d_instance.cpp b/library/src/tensor_operation_instance/gpu/image_to_column/device_image_to_column_nhwc_1d_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/image_to_column/device_image_to_column_nhwc_1d_instance.cpp
rename to library/src/tensor_operation_instance/gpu/image_to_column/device_image_to_column_nhwc_1d_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/image_to_column/device_image_to_column_nhwc_2d_instance.cpp b/library/src/tensor_operation_instance/gpu/image_to_column/device_image_to_column_nhwc_2d_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/image_to_column/device_image_to_column_nhwc_2d_instance.cpp
rename to library/src/tensor_operation_instance/gpu/image_to_column/device_image_to_column_nhwc_2d_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/image_to_column/device_image_to_column_nhwc_3d_instance.cpp b/library/src/tensor_operation_instance/gpu/image_to_column/device_image_to_column_nhwc_3d_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/image_to_column/device_image_to_column_nhwc_3d_instance.cpp
rename to library/src/tensor_operation_instance/gpu/image_to_column/device_image_to_column_nhwc_3d_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/max_pool_bwd/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/max_pool_bwd/CMakeLists.txt
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/max_pool_bwd/CMakeLists.txt
rename to library/src/tensor_operation_instance/gpu/max_pool_bwd/CMakeLists.txt
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/max_pool_bwd/device_max_pool_bwd_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/max_pool_bwd/device_max_pool_bwd_bf16_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/max_pool_bwd/device_max_pool_bwd_bf16_instance.cpp
rename to library/src/tensor_operation_instance/gpu/max_pool_bwd/device_max_pool_bwd_bf16_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/max_pool_bwd/device_max_pool_bwd_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/max_pool_bwd/device_max_pool_bwd_f16_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/max_pool_bwd/device_max_pool_bwd_f16_instance.cpp
rename to library/src/tensor_operation_instance/gpu/max_pool_bwd/device_max_pool_bwd_f16_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/max_pool_bwd/device_max_pool_bwd_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/max_pool_bwd/device_max_pool_bwd_f32_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/max_pool_bwd/device_max_pool_bwd_f32_instance.cpp
rename to library/src/tensor_operation_instance/gpu/max_pool_bwd/device_max_pool_bwd_f32_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/max_pool_bwd/max_pool_bwd_instance_common.hpp b/library/src/tensor_operation_instance/gpu/max_pool_bwd/max_pool_bwd_instance_common.hpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/max_pool_bwd/max_pool_bwd_instance_common.hpp
rename to library/src/tensor_operation_instance/gpu/max_pool_bwd/max_pool_bwd_instance_common.hpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/normalization/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/normalization/CMakeLists.txt
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/normalization/CMakeLists.txt
rename to library/src/tensor_operation_instance/gpu/normalization/CMakeLists.txt
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/normalization/device_groupnorm_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/normalization/device_groupnorm_f16_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/normalization/device_groupnorm_f16_instance.cpp
rename to library/src/tensor_operation_instance/gpu/normalization/device_groupnorm_f16_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/normalization/device_groupnorm_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/normalization/device_groupnorm_f32_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/normalization/device_groupnorm_f32_instance.cpp
rename to library/src/tensor_operation_instance/gpu/normalization/device_groupnorm_f32_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/normalization/device_groupnorm_swish_f16_f32_f32_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/normalization/device_groupnorm_swish_f16_f32_f32_f16_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/normalization/device_groupnorm_swish_f16_f32_f32_f16_instance.cpp
rename to library/src/tensor_operation_instance/gpu/normalization/device_groupnorm_swish_f16_f32_f32_f16_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/normalization/device_groupnorm_swish_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/normalization/device_groupnorm_swish_f16_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/normalization/device_groupnorm_swish_f16_instance.cpp
rename to library/src/tensor_operation_instance/gpu/normalization/device_groupnorm_swish_f16_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/normalization/device_groupnorm_swish_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/normalization/device_groupnorm_swish_f32_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/normalization/device_groupnorm_swish_f32_instance.cpp
rename to library/src/tensor_operation_instance/gpu/normalization/device_groupnorm_swish_f32_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/normalization/device_layernorm2d_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/normalization/device_layernorm2d_f16_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/normalization/device_layernorm2d_f16_instance.cpp
rename to library/src/tensor_operation_instance/gpu/normalization/device_layernorm2d_f16_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/normalization/device_layernorm2d_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/normalization/device_layernorm2d_f32_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/normalization/device_layernorm2d_f32_instance.cpp
rename to library/src/tensor_operation_instance/gpu/normalization/device_layernorm2d_f32_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/normalization/device_layernorm4d_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/normalization/device_layernorm4d_f16_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/normalization/device_layernorm4d_f16_instance.cpp
rename to library/src/tensor_operation_instance/gpu/normalization/device_layernorm4d_f16_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/normalization/device_layernorm4d_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/normalization/device_layernorm4d_f32_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/normalization/device_layernorm4d_f32_instance.cpp
rename to library/src/tensor_operation_instance/gpu/normalization/device_layernorm4d_f32_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/normalization/normalization_instance_common.hpp b/library/src/tensor_operation_instance/gpu/normalization/normalization_instance_common.hpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/normalization/normalization_instance_common.hpp
rename to library/src/tensor_operation_instance/gpu/normalization/normalization_instance_common.hpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/pool3d_fwd/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/pool3d_fwd/CMakeLists.txt
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/pool3d_fwd/CMakeLists.txt
rename to library/src/tensor_operation_instance/gpu/pool3d_fwd/CMakeLists.txt
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/pool3d_fwd/device_avg_pool3d_fwd_ndhwc_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/pool3d_fwd/device_avg_pool3d_fwd_ndhwc_bf16_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/pool3d_fwd/device_avg_pool3d_fwd_ndhwc_bf16_instance.cpp
rename to library/src/tensor_operation_instance/gpu/pool3d_fwd/device_avg_pool3d_fwd_ndhwc_bf16_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/pool3d_fwd/device_avg_pool3d_fwd_ndhwc_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/pool3d_fwd/device_avg_pool3d_fwd_ndhwc_f16_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/pool3d_fwd/device_avg_pool3d_fwd_ndhwc_f16_instance.cpp
rename to library/src/tensor_operation_instance/gpu/pool3d_fwd/device_avg_pool3d_fwd_ndhwc_f16_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/pool3d_fwd/device_avg_pool3d_fwd_ndhwc_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/pool3d_fwd/device_avg_pool3d_fwd_ndhwc_f32_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/pool3d_fwd/device_avg_pool3d_fwd_ndhwc_f32_instance.cpp
rename to library/src/tensor_operation_instance/gpu/pool3d_fwd/device_avg_pool3d_fwd_ndhwc_f32_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/pool3d_fwd/device_max_pool3d_fwd_ndhwc_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/pool3d_fwd/device_max_pool3d_fwd_ndhwc_bf16_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/pool3d_fwd/device_max_pool3d_fwd_ndhwc_bf16_instance.cpp
rename to library/src/tensor_operation_instance/gpu/pool3d_fwd/device_max_pool3d_fwd_ndhwc_bf16_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/pool3d_fwd/device_max_pool3d_fwd_ndhwc_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/pool3d_fwd/device_max_pool3d_fwd_ndhwc_f16_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/pool3d_fwd/device_max_pool3d_fwd_ndhwc_f16_instance.cpp
rename to library/src/tensor_operation_instance/gpu/pool3d_fwd/device_max_pool3d_fwd_ndhwc_f16_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/pool3d_fwd/device_max_pool3d_fwd_ndhwc_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/pool3d_fwd/device_max_pool3d_fwd_ndhwc_f32_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/pool3d_fwd/device_max_pool3d_fwd_ndhwc_f32_instance.cpp
rename to library/src/tensor_operation_instance/gpu/pool3d_fwd/device_max_pool3d_fwd_ndhwc_f32_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/pool3d_fwd/pool_fwd_instance_common.hpp b/library/src/tensor_operation_instance/gpu/pool3d_fwd/pool_fwd_instance_common.hpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/pool3d_fwd/pool_fwd_instance_common.hpp
rename to library/src/tensor_operation_instance/gpu/pool3d_fwd/pool_fwd_instance_common.hpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/quantization/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/quantization/CMakeLists.txt
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/quantization/CMakeLists.txt
rename to library/src/tensor_operation_instance/gpu/quantization/CMakeLists.txt
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/conv2d_quantization_common.hpp b/library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/conv2d_quantization_common.hpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/conv2d_quantization_common.hpp
rename to library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/conv2d_quantization_common.hpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_dl_bias_perchannel_quantization_int8_instance.cpp b/library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_dl_bias_perchannel_quantization_int8_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_dl_bias_perchannel_quantization_int8_instance.cpp
rename to library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_dl_bias_perchannel_quantization_int8_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_dl_bias_perlayer_quantization_int8_instance.cpp b/library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_dl_bias_perlayer_quantization_int8_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_dl_bias_perlayer_quantization_int8_instance.cpp
rename to library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_dl_bias_perlayer_quantization_int8_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_dl_int8_instance.hpp b/library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_dl_int8_instance.hpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_dl_int8_instance.hpp
rename to library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_dl_int8_instance.hpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_dl_perchannel_quantization_int8_instance.cpp b/library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_dl_perchannel_quantization_int8_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_dl_perchannel_quantization_int8_instance.cpp
rename to library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_dl_perchannel_quantization_int8_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_dl_perlayer_quantization_int8_instance.cpp b/library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_dl_perlayer_quantization_int8_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_dl_perlayer_quantization_int8_instance.cpp
rename to library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_dl_perlayer_quantization_int8_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_xdl_bias_perchannel_quantization_int8_instance.cpp b/library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_xdl_bias_perchannel_quantization_int8_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_xdl_bias_perchannel_quantization_int8_instance.cpp
rename to library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_xdl_bias_perchannel_quantization_int8_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_xdl_bias_perlayer_quantization_int8_instance.cpp b/library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_xdl_bias_perlayer_quantization_int8_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_xdl_bias_perlayer_quantization_int8_instance.cpp
rename to library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_xdl_bias_perlayer_quantization_int8_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_xdl_int8_instance.hpp b/library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_xdl_int8_instance.hpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_xdl_int8_instance.hpp
rename to library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_xdl_int8_instance.hpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_xdl_perchannel_quantization_int8_instance.cpp b/library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_xdl_perchannel_quantization_int8_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_xdl_perchannel_quantization_int8_instance.cpp
rename to library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_xdl_perchannel_quantization_int8_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_xdl_perlayer_quantization_int8_instance.cpp b/library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_xdl_perlayer_quantization_int8_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_xdl_perlayer_quantization_int8_instance.cpp
rename to library/src/tensor_operation_instance/gpu/quantization/conv2d_fwd/device_conv2d_xdl_perlayer_quantization_int8_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_dl_c_shuffle_i8_i8_i8_instance.hpp b/library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_dl_c_shuffle_i8_i8_i8_instance.hpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_dl_c_shuffle_i8_i8_i8_instance.hpp
rename to library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_dl_c_shuffle_i8_i8_i8_instance.hpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_dl_c_shuffle_i8_i8_i8_km_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_dl_c_shuffle_i8_i8_i8_km_kn_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_dl_c_shuffle_i8_i8_i8_km_kn_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_dl_c_shuffle_i8_i8_i8_km_kn_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_dl_c_shuffle_i8_i8_i8_km_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_dl_c_shuffle_i8_i8_i8_km_nk_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_dl_c_shuffle_i8_i8_i8_km_nk_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_dl_c_shuffle_i8_i8_i8_km_nk_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_dl_c_shuffle_i8_i8_i8_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_dl_c_shuffle_i8_i8_i8_mk_kn_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_dl_c_shuffle_i8_i8_i8_mk_kn_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_dl_c_shuffle_i8_i8_i8_mk_kn_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_dl_c_shuffle_i8_i8_i8_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_dl_c_shuffle_i8_i8_i8_mk_nk_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_dl_c_shuffle_i8_i8_i8_mk_nk_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_dl_c_shuffle_i8_i8_i8_mk_nk_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_instance.hpp b/library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_instance.hpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_instance.hpp
rename to library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_instance.hpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_km_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_km_nk_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_km_nk_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_km_nk_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_mk_kn_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_mk_kn_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_mk_kn_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/quantization/gemm/device_gemm_quantization_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/quantization/gemm/gemm_quantization_common.hpp b/library/src/tensor_operation_instance/gpu/quantization/gemm/gemm_quantization_common.hpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/quantization/gemm/gemm_quantization_common.hpp
rename to library/src/tensor_operation_instance/gpu/quantization/gemm/gemm_quantization_common.hpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/reduce/CMakeLists.txt
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/reduce/CMakeLists.txt
rename to library/src/tensor_operation_instance/gpu/reduce/CMakeLists.txt
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_add.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_add.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_add.cpp
rename to library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_add.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_amax.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_amax.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_amax.cpp
rename to library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_amax.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_avg.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_avg.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_avg.cpp
rename to library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_avg.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_max.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_max.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_max.cpp
rename to library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_max.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_min.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_min.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_min.cpp
rename to library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_min.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_norm2.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_norm2.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_norm2.cpp
rename to library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_norm2.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_amax.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_amax.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_amax.cpp
rename to library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_amax.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_max.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_max.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_max.cpp
rename to library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_max.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_min.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_min.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_min.cpp
rename to library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_min.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_add.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_add.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_add.cpp
rename to library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_add.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_avg.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_avg.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_avg.cpp
rename to library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_avg.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_norm2.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_norm2.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_norm2.cpp
rename to library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_norm2.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_add.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_add.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_add.cpp
rename to library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_add.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_amax.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_amax.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_amax.cpp
rename to library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_amax.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_avg.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_avg.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_avg.cpp
rename to library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_avg.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_max.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_max.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_max.cpp
rename to library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_max.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_min.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_min.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_min.cpp
rename to library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_min.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_norm2.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_norm2.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_norm2.cpp
rename to library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_norm2.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_add.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_add.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_add.cpp
rename to library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_add.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_avg.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_avg.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_avg.cpp
rename to library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_avg.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_norm2.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_norm2.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_norm2.cpp
rename to library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_norm2.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_add.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_add.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_add.cpp
rename to library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_add.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_amax.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_amax.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_amax.cpp
rename to library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_amax.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_avg.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_avg.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_avg.cpp
rename to library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_avg.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_max.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_max.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_max.cpp
rename to library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_max.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_min.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_min.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_min.cpp
rename to library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_min.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_norm2.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_norm2.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_norm2.cpp
rename to library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_norm2.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i32_i8_add.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i32_i8_add.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i32_i8_add.cpp
rename to library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i32_i8_add.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i32_i8_avg.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i32_i8_avg.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i32_i8_avg.cpp
rename to library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i32_i8_avg.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_amax.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_amax.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_amax.cpp
rename to library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_amax.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_max.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_max.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_max.cpp
rename to library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_max.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_min.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_min.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_min.cpp
rename to library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_min.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_b16_f32_f32_add.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_b16_f32_f32_add.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_b16_f32_f32_add.cpp
rename to library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_b16_f32_f32_add.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_b16_f32_f32_avg.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_b16_f32_f32_avg.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_b16_f32_f32_avg.cpp
rename to library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_b16_f32_f32_avg.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f16_f32_f32_add.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f16_f32_f32_add.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f16_f32_f32_add.cpp
rename to library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f16_f32_f32_add.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f16_f32_f32_avg.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f16_f32_f32_avg.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f16_f32_f32_avg.cpp
rename to library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f16_f32_f32_avg.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f32_f32_add.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f32_f32_add.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f32_f32_add.cpp
rename to library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f32_f32_add.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f32_f32_avg.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f32_f32_avg.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f32_f32_avg.cpp
rename to library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f32_f32_avg.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f64_f32_add.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f64_f32_add.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f64_f32_add.cpp
rename to library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f64_f32_add.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f64_f32_avg.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f64_f32_avg.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f64_f32_avg.cpp
rename to library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f64_f32_avg.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f64_f64_f64_add.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f64_f64_f64_add.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f64_f64_f64_add.cpp
rename to library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f64_f64_f64_add.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f64_f64_f64_avg.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f64_f64_f64_avg.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f64_f64_f64_avg.cpp
rename to library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f64_f64_f64_avg.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_add.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_add.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_add.cpp
rename to library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_add.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_amax.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_amax.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_amax.cpp
rename to library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_amax.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_avg.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_avg.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_avg.cpp
rename to library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_avg.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_max.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_max.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_max.cpp
rename to library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_max.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_min.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_min.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_min.cpp
rename to library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_min.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_norm2.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_norm2.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_norm2.cpp
rename to library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_norm2.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_amax.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_amax.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_amax.cpp
rename to library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_amax.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_max.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_max.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_max.cpp
rename to library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_max.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_min.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_min.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_min.cpp
rename to library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_min.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_add.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_add.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_add.cpp
rename to library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_add.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_avg.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_avg.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_avg.cpp
rename to library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_avg.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_norm2.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_norm2.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_norm2.cpp
rename to library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_norm2.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_add.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_add.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_add.cpp
rename to library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_add.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_amax.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_amax.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_amax.cpp
rename to library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_amax.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_avg.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_avg.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_avg.cpp
rename to library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_avg.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_max.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_max.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_max.cpp
rename to library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_max.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_min.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_min.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_min.cpp
rename to library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_min.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_norm2.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_norm2.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_norm2.cpp
rename to library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_norm2.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_add.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_add.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_add.cpp
rename to library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_add.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_avg.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_avg.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_avg.cpp
rename to library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_avg.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_norm2.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_norm2.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_norm2.cpp
rename to library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_norm2.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_add.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_add.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_add.cpp
rename to library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_add.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_amax.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_amax.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_amax.cpp
rename to library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_amax.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_avg.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_avg.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_avg.cpp
rename to library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_avg.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_max.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_max.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_max.cpp
rename to library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_max.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_min.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_min.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_min.cpp
rename to library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_min.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_norm2.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_norm2.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_norm2.cpp
rename to library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_norm2.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i32_i8_add.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i32_i8_add.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i32_i8_add.cpp
rename to library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i32_i8_add.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i32_i8_avg.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i32_i8_avg.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i32_i8_avg.cpp
rename to library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i32_i8_avg.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_amax.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_amax.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_amax.cpp
rename to library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_amax.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_max.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_max.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_max.cpp
rename to library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_max.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_min.cpp b/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_min.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_min.cpp
rename to library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_min.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/softmax/CMakeLists.txt
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/softmax/CMakeLists.txt
rename to library/src/tensor_operation_instance/gpu/softmax/CMakeLists.txt
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce1.cpp b/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce1.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce1.cpp
rename to library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce1.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce2.cpp b/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce2.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce2.cpp
rename to library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce2.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce3.cpp b/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce3.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce3.cpp
rename to library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank3_reduce3.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce1.cpp b/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce1.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce1.cpp
rename to library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce1.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce2.cpp b/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce2.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce2.cpp
rename to library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce2.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce3.cpp b/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce3.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce3.cpp
rename to library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce3.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce4.cpp b/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce4.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce4.cpp
rename to library/src/tensor_operation_instance/gpu/softmax/device_softmax_f16_f16_instance_rank4_reduce4.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce1.cpp b/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce1.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce1.cpp
rename to library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce1.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce2.cpp b/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce2.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce2.cpp
rename to library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce2.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce3.cpp b/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce3.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce3.cpp
rename to library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank3_reduce3.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce1.cpp b/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce1.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce1.cpp
rename to library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce1.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce2.cpp b/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce2.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce2.cpp
rename to library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce2.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce3.cpp b/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce3.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce3.cpp
rename to library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce3.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce4.cpp b/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce4.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce4.cpp
rename to library/src/tensor_operation_instance/gpu/softmax/device_softmax_f32_f32_instance_rank4_reduce4.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/tall_and_skinny_gemm_splitk/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/tall_and_skinny_gemm_splitk/CMakeLists.txt
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/tall_and_skinny_gemm_splitk/CMakeLists.txt
rename to library/src/tensor_operation_instance/gpu/tall_and_skinny_gemm_splitk/CMakeLists.txt
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/tall_and_skinny_gemm_splitk/device_tall_and_skinny_gemm_splitk_f16_f16_f16_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/tall_and_skinny_gemm_splitk/device_tall_and_skinny_gemm_splitk_f16_f16_f16_mk_kn_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/tall_and_skinny_gemm_splitk/device_tall_and_skinny_gemm_splitk_f16_f16_f16_mk_kn_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/tall_and_skinny_gemm_splitk/device_tall_and_skinny_gemm_splitk_f16_f16_f16_mk_kn_mn_instance.cpp
diff --git a/composable_kernel/library/src/tensor_operation_instance/gpu/tall_and_skinny_gemm_splitk/device_tall_and_skinny_gemm_splitk_f16_f16_f16_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/tall_and_skinny_gemm_splitk/device_tall_and_skinny_gemm_splitk_f16_f16_f16_mk_nk_mn_instance.cpp
similarity index 100%
rename from composable_kernel/library/src/tensor_operation_instance/gpu/tall_and_skinny_gemm_splitk/device_tall_and_skinny_gemm_splitk_f16_f16_f16_mk_nk_mn_instance.cpp
rename to library/src/tensor_operation_instance/gpu/tall_and_skinny_gemm_splitk/device_tall_and_skinny_gemm_splitk_f16_f16_f16_mk_nk_mn_instance.cpp
diff --git a/composable_kernel/library/src/utility/CMakeLists.txt b/library/src/utility/CMakeLists.txt
similarity index 100%
rename from composable_kernel/library/src/utility/CMakeLists.txt
rename to library/src/utility/CMakeLists.txt
diff --git a/composable_kernel/library/src/utility/convolution_parameter.cpp b/library/src/utility/convolution_parameter.cpp
similarity index 100%
rename from composable_kernel/library/src/utility/convolution_parameter.cpp
rename to library/src/utility/convolution_parameter.cpp
diff --git a/composable_kernel/library/src/utility/device_memory.cpp b/library/src/utility/device_memory.cpp
similarity index 100%
rename from composable_kernel/library/src/utility/device_memory.cpp
rename to library/src/utility/device_memory.cpp
diff --git a/composable_kernel/library/src/utility/host_tensor.cpp b/library/src/utility/host_tensor.cpp
similarity index 100%
rename from composable_kernel/library/src/utility/host_tensor.cpp
rename to library/src/utility/host_tensor.cpp
diff --git a/composable_kernel/profiler/CMakeLists.txt b/profiler/CMakeLists.txt
similarity index 100%
rename from composable_kernel/profiler/CMakeLists.txt
rename to profiler/CMakeLists.txt
diff --git a/composable_kernel/profiler/README.md b/profiler/README.md
similarity index 100%
rename from composable_kernel/profiler/README.md
rename to profiler/README.md
diff --git a/composable_kernel/profiler/include/profiler/data_type_enum.hpp b/profiler/include/profiler/data_type_enum.hpp
similarity index 100%
rename from composable_kernel/profiler/include/profiler/data_type_enum.hpp
rename to profiler/include/profiler/data_type_enum.hpp
diff --git a/composable_kernel/profiler/include/profiler/profile_avg_pool3d_bwd_impl.hpp b/profiler/include/profiler/profile_avg_pool3d_bwd_impl.hpp
similarity index 100%
rename from composable_kernel/profiler/include/profiler/profile_avg_pool3d_bwd_impl.hpp
rename to profiler/include/profiler/profile_avg_pool3d_bwd_impl.hpp
diff --git a/composable_kernel/profiler/include/profiler/profile_batched_gemm_add_relu_gemm_add_impl.hpp b/profiler/include/profiler/profile_batched_gemm_add_relu_gemm_add_impl.hpp
similarity index 100%
rename from composable_kernel/profiler/include/profiler/profile_batched_gemm_add_relu_gemm_add_impl.hpp
rename to profiler/include/profiler/profile_batched_gemm_add_relu_gemm_add_impl.hpp
diff --git a/composable_kernel/profiler/include/profiler/profile_batched_gemm_bias_softmax_gemm_permute_impl.hpp b/profiler/include/profiler/profile_batched_gemm_bias_softmax_gemm_permute_impl.hpp
similarity index 100%
rename from composable_kernel/profiler/include/profiler/profile_batched_gemm_bias_softmax_gemm_permute_impl.hpp
rename to profiler/include/profiler/profile_batched_gemm_bias_softmax_gemm_permute_impl.hpp
diff --git a/composable_kernel/profiler/include/profiler/profile_batched_gemm_gemm_impl.hpp b/profiler/include/profiler/profile_batched_gemm_gemm_impl.hpp
similarity index 100%
rename from composable_kernel/profiler/include/profiler/profile_batched_gemm_gemm_impl.hpp
rename to profiler/include/profiler/profile_batched_gemm_gemm_impl.hpp
diff --git a/composable_kernel/profiler/include/profiler/profile_batched_gemm_impl.hpp b/profiler/include/profiler/profile_batched_gemm_impl.hpp
similarity index 100%
rename from composable_kernel/profiler/include/profiler/profile_batched_gemm_impl.hpp
rename to profiler/include/profiler/profile_batched_gemm_impl.hpp
diff --git a/composable_kernel/profiler/include/profiler/profile_batched_gemm_reduce_impl.hpp b/profiler/include/profiler/profile_batched_gemm_reduce_impl.hpp
similarity index 100%
rename from composable_kernel/profiler/include/profiler/profile_batched_gemm_reduce_impl.hpp
rename to profiler/include/profiler/profile_batched_gemm_reduce_impl.hpp
diff --git a/composable_kernel/profiler/include/profiler/profile_batched_gemm_softmax_gemm_impl.hpp b/profiler/include/profiler/profile_batched_gemm_softmax_gemm_impl.hpp
similarity index 100%
rename from composable_kernel/profiler/include/profiler/profile_batched_gemm_softmax_gemm_impl.hpp
rename to profiler/include/profiler/profile_batched_gemm_softmax_gemm_impl.hpp
diff --git a/composable_kernel/profiler/include/profiler/profile_batched_gemm_softmax_gemm_permute_impl.hpp b/profiler/include/profiler/profile_batched_gemm_softmax_gemm_permute_impl.hpp
similarity index 100%
rename from composable_kernel/profiler/include/profiler/profile_batched_gemm_softmax_gemm_permute_impl.hpp
rename to profiler/include/profiler/profile_batched_gemm_softmax_gemm_permute_impl.hpp
diff --git a/composable_kernel/profiler/include/profiler/profile_batchnorm_backward_impl.hpp b/profiler/include/profiler/profile_batchnorm_backward_impl.hpp
similarity index 100%
rename from composable_kernel/profiler/include/profiler/profile_batchnorm_backward_impl.hpp
rename to profiler/include/profiler/profile_batchnorm_backward_impl.hpp
diff --git a/composable_kernel/profiler/include/profiler/profile_batchnorm_forward_impl.hpp b/profiler/include/profiler/profile_batchnorm_forward_impl.hpp
similarity index 100%
rename from composable_kernel/profiler/include/profiler/profile_batchnorm_forward_impl.hpp
rename to profiler/include/profiler/profile_batchnorm_forward_impl.hpp
diff --git a/composable_kernel/profiler/include/profiler/profile_batchnorm_infer_impl.hpp b/profiler/include/profiler/profile_batchnorm_infer_impl.hpp
similarity index 100%
rename from composable_kernel/profiler/include/profiler/profile_batchnorm_infer_impl.hpp
rename to profiler/include/profiler/profile_batchnorm_infer_impl.hpp
diff --git a/composable_kernel/profiler/include/profiler/profile_contraction_impl.hpp b/profiler/include/profiler/profile_contraction_impl.hpp
similarity index 100%
rename from composable_kernel/profiler/include/profiler/profile_contraction_impl.hpp
rename to profiler/include/profiler/profile_contraction_impl.hpp
diff --git a/composable_kernel/profiler/include/profiler/profile_contraction_utils.hpp b/profiler/include/profiler/profile_contraction_utils.hpp
similarity index 100%
rename from composable_kernel/profiler/include/profiler/profile_contraction_utils.hpp
rename to profiler/include/profiler/profile_contraction_utils.hpp
diff --git a/composable_kernel/profiler/include/profiler/profile_conv_bwd_data_impl.hpp b/profiler/include/profiler/profile_conv_bwd_data_impl.hpp
similarity index 100%
rename from composable_kernel/profiler/include/profiler/profile_conv_bwd_data_impl.hpp
rename to profiler/include/profiler/profile_conv_bwd_data_impl.hpp
diff --git a/composable_kernel/profiler/include/profiler/profile_conv_fwd_bias_relu_add_impl.hpp b/profiler/include/profiler/profile_conv_fwd_bias_relu_add_impl.hpp
similarity index 100%
rename from composable_kernel/profiler/include/profiler/profile_conv_fwd_bias_relu_add_impl.hpp
rename to profiler/include/profiler/profile_conv_fwd_bias_relu_add_impl.hpp
diff --git a/composable_kernel/profiler/include/profiler/profile_conv_fwd_bias_relu_impl.hpp b/profiler/include/profiler/profile_conv_fwd_bias_relu_impl.hpp
similarity index 100%
rename from composable_kernel/profiler/include/profiler/profile_conv_fwd_bias_relu_impl.hpp
rename to profiler/include/profiler/profile_conv_fwd_bias_relu_impl.hpp
diff --git a/composable_kernel/profiler/include/profiler/profile_conv_fwd_impl.hpp b/profiler/include/profiler/profile_conv_fwd_impl.hpp
similarity index 100%
rename from composable_kernel/profiler/include/profiler/profile_conv_fwd_impl.hpp
rename to profiler/include/profiler/profile_conv_fwd_impl.hpp
diff --git a/composable_kernel/profiler/include/profiler/profile_conv_tensor_rearrange_impl.hpp b/profiler/include/profiler/profile_conv_tensor_rearrange_impl.hpp
similarity index 100%
rename from composable_kernel/profiler/include/profiler/profile_conv_tensor_rearrange_impl.hpp
rename to profiler/include/profiler/profile_conv_tensor_rearrange_impl.hpp
diff --git a/composable_kernel/profiler/include/profiler/profile_elementwise_layernorm_impl.hpp b/profiler/include/profiler/profile_elementwise_layernorm_impl.hpp
similarity index 100%
rename from composable_kernel/profiler/include/profiler/profile_elementwise_layernorm_impl.hpp
rename to profiler/include/profiler/profile_elementwise_layernorm_impl.hpp
diff --git a/composable_kernel/profiler/include/profiler/profile_gemm_add_add_fastgelu_impl.hpp b/profiler/include/profiler/profile_gemm_add_add_fastgelu_impl.hpp
similarity index 100%
rename from composable_kernel/profiler/include/profiler/profile_gemm_add_add_fastgelu_impl.hpp
rename to profiler/include/profiler/profile_gemm_add_add_fastgelu_impl.hpp
diff --git a/composable_kernel/profiler/include/profiler/profile_gemm_add_fastgelu_impl.hpp b/profiler/include/profiler/profile_gemm_add_fastgelu_impl.hpp
similarity index 100%
rename from composable_kernel/profiler/include/profiler/profile_gemm_add_fastgelu_impl.hpp
rename to profiler/include/profiler/profile_gemm_add_fastgelu_impl.hpp
diff --git a/composable_kernel/profiler/include/profiler/profile_gemm_add_multiply_impl.hpp b/profiler/include/profiler/profile_gemm_add_multiply_impl.hpp
similarity index 100%
rename from composable_kernel/profiler/include/profiler/profile_gemm_add_multiply_impl.hpp
rename to profiler/include/profiler/profile_gemm_add_multiply_impl.hpp
diff --git a/composable_kernel/profiler/include/profiler/profile_gemm_add_relu_add_layernorm_impl.hpp b/profiler/include/profiler/profile_gemm_add_relu_add_layernorm_impl.hpp
similarity index 100%
rename from composable_kernel/profiler/include/profiler/profile_gemm_add_relu_add_layernorm_impl.hpp
rename to profiler/include/profiler/profile_gemm_add_relu_add_layernorm_impl.hpp
diff --git a/composable_kernel/profiler/include/profiler/profile_gemm_bias_add_reduce_impl.hpp b/profiler/include/profiler/profile_gemm_bias_add_reduce_impl.hpp
similarity index 100%
rename from composable_kernel/profiler/include/profiler/profile_gemm_bias_add_reduce_impl.hpp
rename to profiler/include/profiler/profile_gemm_bias_add_reduce_impl.hpp
diff --git a/composable_kernel/profiler/include/profiler/profile_gemm_bilinear_impl.hpp b/profiler/include/profiler/profile_gemm_bilinear_impl.hpp
similarity index 100%
rename from composable_kernel/profiler/include/profiler/profile_gemm_bilinear_impl.hpp
rename to profiler/include/profiler/profile_gemm_bilinear_impl.hpp
diff --git a/composable_kernel/profiler/include/profiler/profile_gemm_fastgelu_impl.hpp b/profiler/include/profiler/profile_gemm_fastgelu_impl.hpp
similarity index 100%
rename from composable_kernel/profiler/include/profiler/profile_gemm_fastgelu_impl.hpp
rename to profiler/include/profiler/profile_gemm_fastgelu_impl.hpp
diff --git a/composable_kernel/profiler/include/profiler/profile_gemm_impl.hpp b/profiler/include/profiler/profile_gemm_impl.hpp
similarity index 100%
rename from composable_kernel/profiler/include/profiler/profile_gemm_impl.hpp
rename to profiler/include/profiler/profile_gemm_impl.hpp
diff --git a/composable_kernel/profiler/include/profiler/profile_gemm_multiply_add_impl.hpp b/profiler/include/profiler/profile_gemm_multiply_add_impl.hpp
similarity index 100%
rename from composable_kernel/profiler/include/profiler/profile_gemm_multiply_add_impl.hpp
rename to profiler/include/profiler/profile_gemm_multiply_add_impl.hpp
diff --git a/composable_kernel/profiler/include/profiler/profile_gemm_reduce_impl.hpp b/profiler/include/profiler/profile_gemm_reduce_impl.hpp
similarity index 100%
rename from composable_kernel/profiler/include/profiler/profile_gemm_reduce_impl.hpp
rename to profiler/include/profiler/profile_gemm_reduce_impl.hpp
diff --git a/composable_kernel/profiler/include/profiler/profile_gemm_splitk_impl.hpp b/profiler/include/profiler/profile_gemm_splitk_impl.hpp
similarity index 100%
rename from composable_kernel/profiler/include/profiler/profile_gemm_splitk_impl.hpp
rename to profiler/include/profiler/profile_gemm_splitk_impl.hpp
diff --git a/composable_kernel/profiler/include/profiler/profile_gemm_streamk_impl.hpp b/profiler/include/profiler/profile_gemm_streamk_impl.hpp
similarity index 100%
rename from composable_kernel/profiler/include/profiler/profile_gemm_streamk_impl.hpp
rename to profiler/include/profiler/profile_gemm_streamk_impl.hpp
diff --git a/composable_kernel/profiler/include/profiler/profile_gemv_splitk_impl.hpp b/profiler/include/profiler/profile_gemv_splitk_impl.hpp
similarity index 100%
rename from composable_kernel/profiler/include/profiler/profile_gemv_splitk_impl.hpp
rename to profiler/include/profiler/profile_gemv_splitk_impl.hpp
diff --git a/composable_kernel/profiler/include/profiler/profile_grouped_conv_bwd_data_impl.hpp b/profiler/include/profiler/profile_grouped_conv_bwd_data_impl.hpp
similarity index 100%
rename from composable_kernel/profiler/include/profiler/profile_grouped_conv_bwd_data_impl.hpp
rename to profiler/include/profiler/profile_grouped_conv_bwd_data_impl.hpp
diff --git a/composable_kernel/profiler/include/profiler/profile_grouped_conv_bwd_weight_impl.hpp b/profiler/include/profiler/profile_grouped_conv_bwd_weight_impl.hpp
similarity index 100%
rename from composable_kernel/profiler/include/profiler/profile_grouped_conv_bwd_weight_impl.hpp
rename to profiler/include/profiler/profile_grouped_conv_bwd_weight_impl.hpp
diff --git a/composable_kernel/profiler/include/profiler/profile_grouped_conv_fwd_impl.hpp b/profiler/include/profiler/profile_grouped_conv_fwd_impl.hpp
similarity index 100%
rename from composable_kernel/profiler/include/profiler/profile_grouped_conv_fwd_impl.hpp
rename to profiler/include/profiler/profile_grouped_conv_fwd_impl.hpp
diff --git a/composable_kernel/profiler/include/profiler/profile_grouped_gemm_fastgelu_impl.hpp b/profiler/include/profiler/profile_grouped_gemm_fastgelu_impl.hpp
similarity index 100%
rename from composable_kernel/profiler/include/profiler/profile_grouped_gemm_fastgelu_impl.hpp
rename to profiler/include/profiler/profile_grouped_gemm_fastgelu_impl.hpp
diff --git a/composable_kernel/profiler/include/profiler/profile_grouped_gemm_impl.hpp b/profiler/include/profiler/profile_grouped_gemm_impl.hpp
similarity index 100%
rename from composable_kernel/profiler/include/profiler/profile_grouped_gemm_impl.hpp
rename to profiler/include/profiler/profile_grouped_gemm_impl.hpp
diff --git a/composable_kernel/profiler/include/profiler/profile_groupnorm_impl.hpp b/profiler/include/profiler/profile_groupnorm_impl.hpp
similarity index 100%
rename from composable_kernel/profiler/include/profiler/profile_groupnorm_impl.hpp
rename to profiler/include/profiler/profile_groupnorm_impl.hpp
diff --git a/composable_kernel/profiler/include/profiler/profile_layernorm_impl.hpp b/profiler/include/profiler/profile_layernorm_impl.hpp
similarity index 100%
rename from composable_kernel/profiler/include/profiler/profile_layernorm_impl.hpp
rename to profiler/include/profiler/profile_layernorm_impl.hpp
diff --git a/composable_kernel/profiler/include/profiler/profile_max_pool3d_bwd_impl.hpp b/profiler/include/profiler/profile_max_pool3d_bwd_impl.hpp
similarity index 100%
rename from composable_kernel/profiler/include/profiler/profile_max_pool3d_bwd_impl.hpp
rename to profiler/include/profiler/profile_max_pool3d_bwd_impl.hpp
diff --git a/composable_kernel/profiler/include/profiler/profile_pool3d_fwd_impl.hpp b/profiler/include/profiler/profile_pool3d_fwd_impl.hpp
similarity index 100%
rename from composable_kernel/profiler/include/profiler/profile_pool3d_fwd_impl.hpp
rename to profiler/include/profiler/profile_pool3d_fwd_impl.hpp
diff --git a/composable_kernel/profiler/include/profiler/profile_reduce_impl.hpp b/profiler/include/profiler/profile_reduce_impl.hpp
similarity index 100%
rename from composable_kernel/profiler/include/profiler/profile_reduce_impl.hpp
rename to profiler/include/profiler/profile_reduce_impl.hpp
diff --git a/composable_kernel/profiler/include/profiler/profile_softmax_impl.hpp b/profiler/include/profiler/profile_softmax_impl.hpp
similarity index 100%
rename from composable_kernel/profiler/include/profiler/profile_softmax_impl.hpp
rename to profiler/include/profiler/profile_softmax_impl.hpp
diff --git a/composable_kernel/profiler/include/profiler/profile_tall_and_skinny_gemm_splitk_impl.hpp b/profiler/include/profiler/profile_tall_and_skinny_gemm_splitk_impl.hpp
similarity index 100%
rename from composable_kernel/profiler/include/profiler/profile_tall_and_skinny_gemm_splitk_impl.hpp
rename to profiler/include/profiler/profile_tall_and_skinny_gemm_splitk_impl.hpp
diff --git a/composable_kernel/profiler/src/CMakeLists.txt b/profiler/src/CMakeLists.txt
similarity index 100%
rename from composable_kernel/profiler/src/CMakeLists.txt
rename to profiler/src/CMakeLists.txt
diff --git a/composable_kernel/profiler/src/profile_avg_pool3d_bwd.cpp b/profiler/src/profile_avg_pool3d_bwd.cpp
similarity index 100%
rename from composable_kernel/profiler/src/profile_avg_pool3d_bwd.cpp
rename to profiler/src/profile_avg_pool3d_bwd.cpp
diff --git a/composable_kernel/profiler/src/profile_batched_gemm.cpp b/profiler/src/profile_batched_gemm.cpp
similarity index 100%
rename from composable_kernel/profiler/src/profile_batched_gemm.cpp
rename to profiler/src/profile_batched_gemm.cpp
diff --git a/composable_kernel/profiler/src/profile_batched_gemm_add_relu_gemm_add.cpp b/profiler/src/profile_batched_gemm_add_relu_gemm_add.cpp
similarity index 100%
rename from composable_kernel/profiler/src/profile_batched_gemm_add_relu_gemm_add.cpp
rename to profiler/src/profile_batched_gemm_add_relu_gemm_add.cpp
diff --git a/composable_kernel/profiler/src/profile_batched_gemm_gemm.cpp b/profiler/src/profile_batched_gemm_gemm.cpp
similarity index 100%
rename from composable_kernel/profiler/src/profile_batched_gemm_gemm.cpp
rename to profiler/src/profile_batched_gemm_gemm.cpp
diff --git a/composable_kernel/profiler/src/profile_batched_gemm_multi_d.cpp b/profiler/src/profile_batched_gemm_multi_d.cpp
similarity index 100%
rename from composable_kernel/profiler/src/profile_batched_gemm_multi_d.cpp
rename to profiler/src/profile_batched_gemm_multi_d.cpp
diff --git a/composable_kernel/profiler/src/profile_batched_gemm_reduce.cpp b/profiler/src/profile_batched_gemm_reduce.cpp
similarity index 100%
rename from composable_kernel/profiler/src/profile_batched_gemm_reduce.cpp
rename to profiler/src/profile_batched_gemm_reduce.cpp
diff --git a/composable_kernel/profiler/src/profile_batchnorm_bwd.cpp b/profiler/src/profile_batchnorm_bwd.cpp
similarity index 100%
rename from composable_kernel/profiler/src/profile_batchnorm_bwd.cpp
rename to profiler/src/profile_batchnorm_bwd.cpp
diff --git a/composable_kernel/profiler/src/profile_batchnorm_fwd.cpp b/profiler/src/profile_batchnorm_fwd.cpp
similarity index 100%
rename from composable_kernel/profiler/src/profile_batchnorm_fwd.cpp
rename to profiler/src/profile_batchnorm_fwd.cpp
diff --git a/composable_kernel/profiler/src/profile_batchnorm_infer.cpp b/profiler/src/profile_batchnorm_infer.cpp
similarity index 100%
rename from composable_kernel/profiler/src/profile_batchnorm_infer.cpp
rename to profiler/src/profile_batchnorm_infer.cpp
diff --git a/composable_kernel/profiler/src/profile_contraction_bilinear.cpp b/profiler/src/profile_contraction_bilinear.cpp
similarity index 100%
rename from composable_kernel/profiler/src/profile_contraction_bilinear.cpp
rename to profiler/src/profile_contraction_bilinear.cpp
diff --git a/composable_kernel/profiler/src/profile_contraction_scale.cpp b/profiler/src/profile_contraction_scale.cpp
similarity index 100%
rename from composable_kernel/profiler/src/profile_contraction_scale.cpp
rename to profiler/src/profile_contraction_scale.cpp
diff --git a/composable_kernel/profiler/src/profile_conv_bwd_data.cpp b/profiler/src/profile_conv_bwd_data.cpp
similarity index 100%
rename from composable_kernel/profiler/src/profile_conv_bwd_data.cpp
rename to profiler/src/profile_conv_bwd_data.cpp
diff --git a/composable_kernel/profiler/src/profile_conv_fwd.cpp b/profiler/src/profile_conv_fwd.cpp
similarity index 100%
rename from composable_kernel/profiler/src/profile_conv_fwd.cpp
rename to profiler/src/profile_conv_fwd.cpp
diff --git a/composable_kernel/profiler/src/profile_conv_fwd_bias_relu.cpp b/profiler/src/profile_conv_fwd_bias_relu.cpp
similarity index 100%
rename from composable_kernel/profiler/src/profile_conv_fwd_bias_relu.cpp
rename to profiler/src/profile_conv_fwd_bias_relu.cpp
diff --git a/composable_kernel/profiler/src/profile_conv_fwd_bias_relu_add.cpp b/profiler/src/profile_conv_fwd_bias_relu_add.cpp
similarity index 100%
rename from composable_kernel/profiler/src/profile_conv_fwd_bias_relu_add.cpp
rename to profiler/src/profile_conv_fwd_bias_relu_add.cpp
diff --git a/composable_kernel/profiler/src/profile_conv_tensor_rearrange.cpp b/profiler/src/profile_conv_tensor_rearrange.cpp
similarity index 100%
rename from composable_kernel/profiler/src/profile_conv_tensor_rearrange.cpp
rename to profiler/src/profile_conv_tensor_rearrange.cpp
diff --git a/composable_kernel/profiler/src/profile_gemm.cpp b/profiler/src/profile_gemm.cpp
similarity index 100%
rename from composable_kernel/profiler/src/profile_gemm.cpp
rename to profiler/src/profile_gemm.cpp
diff --git a/composable_kernel/profiler/src/profile_gemm_add_add_fastgelu.cpp b/profiler/src/profile_gemm_add_add_fastgelu.cpp
similarity index 100%
rename from composable_kernel/profiler/src/profile_gemm_add_add_fastgelu.cpp
rename to profiler/src/profile_gemm_add_add_fastgelu.cpp
diff --git a/composable_kernel/profiler/src/profile_gemm_add_fastgelu.cpp b/profiler/src/profile_gemm_add_fastgelu.cpp
similarity index 100%
rename from composable_kernel/profiler/src/profile_gemm_add_fastgelu.cpp
rename to profiler/src/profile_gemm_add_fastgelu.cpp
diff --git a/composable_kernel/profiler/src/profile_gemm_add_multiply.cpp b/profiler/src/profile_gemm_add_multiply.cpp
similarity index 100%
rename from composable_kernel/profiler/src/profile_gemm_add_multiply.cpp
rename to profiler/src/profile_gemm_add_multiply.cpp
diff --git a/composable_kernel/profiler/src/profile_gemm_add_relu_add_layernorm.cpp b/profiler/src/profile_gemm_add_relu_add_layernorm.cpp
similarity index 100%
rename from composable_kernel/profiler/src/profile_gemm_add_relu_add_layernorm.cpp
rename to profiler/src/profile_gemm_add_relu_add_layernorm.cpp
diff --git a/composable_kernel/profiler/src/profile_gemm_bias_add_reduce.cpp b/profiler/src/profile_gemm_bias_add_reduce.cpp
similarity index 100%
rename from composable_kernel/profiler/src/profile_gemm_bias_add_reduce.cpp
rename to profiler/src/profile_gemm_bias_add_reduce.cpp
diff --git a/composable_kernel/profiler/src/profile_gemm_bilinear.cpp b/profiler/src/profile_gemm_bilinear.cpp
similarity index 100%
rename from composable_kernel/profiler/src/profile_gemm_bilinear.cpp
rename to profiler/src/profile_gemm_bilinear.cpp
diff --git a/composable_kernel/profiler/src/profile_gemm_fastgelu.cpp b/profiler/src/profile_gemm_fastgelu.cpp
similarity index 100%
rename from composable_kernel/profiler/src/profile_gemm_fastgelu.cpp
rename to profiler/src/profile_gemm_fastgelu.cpp
diff --git a/composable_kernel/profiler/src/profile_gemm_multiply_add.cpp b/profiler/src/profile_gemm_multiply_add.cpp
similarity index 100%
rename from composable_kernel/profiler/src/profile_gemm_multiply_add.cpp
rename to profiler/src/profile_gemm_multiply_add.cpp
diff --git a/composable_kernel/profiler/src/profile_gemm_reduce.cpp b/profiler/src/profile_gemm_reduce.cpp
similarity index 100%
rename from composable_kernel/profiler/src/profile_gemm_reduce.cpp
rename to profiler/src/profile_gemm_reduce.cpp
diff --git a/composable_kernel/profiler/src/profile_gemm_splitk.cpp b/profiler/src/profile_gemm_splitk.cpp
similarity index 100%
rename from composable_kernel/profiler/src/profile_gemm_splitk.cpp
rename to profiler/src/profile_gemm_splitk.cpp
diff --git a/composable_kernel/profiler/src/profile_gemm_streamk.cpp b/profiler/src/profile_gemm_streamk.cpp
similarity index 100%
rename from composable_kernel/profiler/src/profile_gemm_streamk.cpp
rename to profiler/src/profile_gemm_streamk.cpp
diff --git a/composable_kernel/profiler/src/profile_gemv_splitk.cpp b/profiler/src/profile_gemv_splitk.cpp
similarity index 100%
rename from composable_kernel/profiler/src/profile_gemv_splitk.cpp
rename to profiler/src/profile_gemv_splitk.cpp
diff --git a/composable_kernel/profiler/src/profile_grouped_conv_bwd_data.cpp b/profiler/src/profile_grouped_conv_bwd_data.cpp
similarity index 100%
rename from composable_kernel/profiler/src/profile_grouped_conv_bwd_data.cpp
rename to profiler/src/profile_grouped_conv_bwd_data.cpp
diff --git a/composable_kernel/profiler/src/profile_grouped_conv_bwd_weight.cpp b/profiler/src/profile_grouped_conv_bwd_weight.cpp
similarity index 100%
rename from composable_kernel/profiler/src/profile_grouped_conv_bwd_weight.cpp
rename to profiler/src/profile_grouped_conv_bwd_weight.cpp
diff --git a/composable_kernel/profiler/src/profile_grouped_conv_fwd.cpp b/profiler/src/profile_grouped_conv_fwd.cpp
similarity index 100%
rename from composable_kernel/profiler/src/profile_grouped_conv_fwd.cpp
rename to profiler/src/profile_grouped_conv_fwd.cpp
diff --git a/composable_kernel/profiler/src/profile_grouped_gemm.cpp b/profiler/src/profile_grouped_gemm.cpp
similarity index 100%
rename from composable_kernel/profiler/src/profile_grouped_gemm.cpp
rename to profiler/src/profile_grouped_gemm.cpp
diff --git a/composable_kernel/profiler/src/profile_grouped_gemm_fastgelu.cpp b/profiler/src/profile_grouped_gemm_fastgelu.cpp
similarity index 100%
rename from composable_kernel/profiler/src/profile_grouped_gemm_fastgelu.cpp
rename to profiler/src/profile_grouped_gemm_fastgelu.cpp
diff --git a/composable_kernel/profiler/src/profile_groupnorm.cpp b/profiler/src/profile_groupnorm.cpp
similarity index 100%
rename from composable_kernel/profiler/src/profile_groupnorm.cpp
rename to profiler/src/profile_groupnorm.cpp
diff --git a/composable_kernel/profiler/src/profile_layernorm.cpp b/profiler/src/profile_layernorm.cpp
similarity index 100%
rename from composable_kernel/profiler/src/profile_layernorm.cpp
rename to profiler/src/profile_layernorm.cpp
diff --git a/composable_kernel/profiler/src/profile_max_pool3d_bwd.cpp b/profiler/src/profile_max_pool3d_bwd.cpp
similarity index 100%
rename from composable_kernel/profiler/src/profile_max_pool3d_bwd.cpp
rename to profiler/src/profile_max_pool3d_bwd.cpp
diff --git a/composable_kernel/profiler/src/profile_max_pool3d_fwd.cpp b/profiler/src/profile_max_pool3d_fwd.cpp
similarity index 100%
rename from composable_kernel/profiler/src/profile_max_pool3d_fwd.cpp
rename to profiler/src/profile_max_pool3d_fwd.cpp
diff --git a/composable_kernel/profiler/src/profile_reduce.cpp b/profiler/src/profile_reduce.cpp
similarity index 100%
rename from composable_kernel/profiler/src/profile_reduce.cpp
rename to profiler/src/profile_reduce.cpp
diff --git a/composable_kernel/profiler/src/profile_softmax.cpp b/profiler/src/profile_softmax.cpp
similarity index 100%
rename from composable_kernel/profiler/src/profile_softmax.cpp
rename to profiler/src/profile_softmax.cpp
diff --git a/composable_kernel/profiler/src/profile_tall_and_skinny_gemm_splitk.cpp b/profiler/src/profile_tall_and_skinny_gemm_splitk.cpp
similarity index 100%
rename from composable_kernel/profiler/src/profile_tall_and_skinny_gemm_splitk.cpp
rename to profiler/src/profile_tall_and_skinny_gemm_splitk.cpp
diff --git a/composable_kernel/profiler/src/profiler.cpp b/profiler/src/profiler.cpp
similarity index 100%
rename from composable_kernel/profiler/src/profiler.cpp
rename to profiler/src/profiler.cpp
diff --git a/composable_kernel/profiler/src/profiler_operation_registry.hpp b/profiler/src/profiler_operation_registry.hpp
similarity index 100%
rename from composable_kernel/profiler/src/profiler_operation_registry.hpp
rename to profiler/src/profiler_operation_registry.hpp
diff --git a/composable_kernel/rbuild.ini b/rbuild.ini
similarity index 100%
rename from composable_kernel/rbuild.ini
rename to rbuild.ini
diff --git a/composable_kernel/requirements.txt b/requirements.txt
similarity index 100%
rename from composable_kernel/requirements.txt
rename to requirements.txt
diff --git a/composable_kernel/script/check_copyright_year.sh b/script/check_copyright_year.sh
similarity index 100%
rename from composable_kernel/script/check_copyright_year.sh
rename to script/check_copyright_year.sh
diff --git a/composable_kernel/script/clang-format-overwrite.sh b/script/clang-format-overwrite.sh
similarity index 100%
rename from composable_kernel/script/clang-format-overwrite.sh
rename to script/clang-format-overwrite.sh
diff --git a/composable_kernel/script/cmake-ck-dev.sh b/script/cmake-ck-dev.sh
similarity index 100%
rename from composable_kernel/script/cmake-ck-dev.sh
rename to script/cmake-ck-dev.sh
diff --git a/composable_kernel/script/cmake-ck-release.sh b/script/cmake-ck-release.sh
similarity index 100%
rename from composable_kernel/script/cmake-ck-release.sh
rename to script/cmake-ck-release.sh
diff --git a/composable_kernel/script/count_vgpr.sh b/script/count_vgpr.sh
similarity index 100%
rename from composable_kernel/script/count_vgpr.sh
rename to script/count_vgpr.sh
diff --git a/composable_kernel/script/hipclang_opt.sh b/script/hipclang_opt.sh
similarity index 100%
rename from composable_kernel/script/hipclang_opt.sh
rename to script/hipclang_opt.sh
diff --git a/composable_kernel/script/install_precommit.sh b/script/install_precommit.sh
similarity index 100%
rename from composable_kernel/script/install_precommit.sh
rename to script/install_precommit.sh
diff --git a/composable_kernel/script/parse_perf_data.py b/script/parse_perf_data.py
similarity index 100%
rename from composable_kernel/script/parse_perf_data.py
rename to script/parse_perf_data.py
diff --git a/composable_kernel/script/process_perf_data.py b/script/process_perf_data.py
similarity index 100%
rename from composable_kernel/script/process_perf_data.py
rename to script/process_perf_data.py
diff --git a/composable_kernel/script/process_perf_data.sh b/script/process_perf_data.sh
similarity index 100%
rename from composable_kernel/script/process_perf_data.sh
rename to script/process_perf_data.sh
diff --git a/composable_kernel/script/process_qa_data.sh b/script/process_qa_data.sh
similarity index 100%
rename from composable_kernel/script/process_qa_data.sh
rename to script/process_qa_data.sh
diff --git a/composable_kernel/script/profile_batched_gemm.sh b/script/profile_batched_gemm.sh
similarity index 100%
rename from composable_kernel/script/profile_batched_gemm.sh
rename to script/profile_batched_gemm.sh
diff --git a/composable_kernel/script/profile_conv_bwd_data.sh b/script/profile_conv_bwd_data.sh
similarity index 100%
rename from composable_kernel/script/profile_conv_bwd_data.sh
rename to script/profile_conv_bwd_data.sh
diff --git a/composable_kernel/script/profile_conv_fwd.sh b/script/profile_conv_fwd.sh
similarity index 100%
rename from composable_kernel/script/profile_conv_fwd.sh
rename to script/profile_conv_fwd.sh
diff --git a/composable_kernel/script/profile_gemm.sh b/script/profile_gemm.sh
similarity index 100%
rename from composable_kernel/script/profile_gemm.sh
rename to script/profile_gemm.sh
diff --git a/composable_kernel/script/profile_gemm_bilinear.sh b/script/profile_gemm_bilinear.sh
similarity index 100%
rename from composable_kernel/script/profile_gemm_bilinear.sh
rename to script/profile_gemm_bilinear.sh
diff --git a/composable_kernel/script/profile_grouped_gemm.sh b/script/profile_grouped_gemm.sh
similarity index 100%
rename from composable_kernel/script/profile_grouped_gemm.sh
rename to script/profile_grouped_gemm.sh
diff --git a/composable_kernel/script/profile_onnx_gemm.sh b/script/profile_onnx_gemm.sh
similarity index 100%
rename from composable_kernel/script/profile_onnx_gemm.sh
rename to script/profile_onnx_gemm.sh
diff --git a/composable_kernel/script/profile_reduce_no_index.sh b/script/profile_reduce_no_index.sh
similarity index 100%
rename from composable_kernel/script/profile_reduce_no_index.sh
rename to script/profile_reduce_no_index.sh
diff --git a/composable_kernel/script/profile_reduce_with_index.sh b/script/profile_reduce_with_index.sh
similarity index 100%
rename from composable_kernel/script/profile_reduce_with_index.sh
rename to script/profile_reduce_with_index.sh
diff --git a/composable_kernel/script/profile_resnet50.sh b/script/profile_resnet50.sh
similarity index 100%
rename from composable_kernel/script/profile_resnet50.sh
rename to script/profile_resnet50.sh
diff --git a/composable_kernel/script/profile_splitK_gemm.sh b/script/profile_splitK_gemm.sh
similarity index 100%
rename from composable_kernel/script/profile_splitK_gemm.sh
rename to script/profile_splitK_gemm.sh
diff --git a/composable_kernel/script/run_full_performance_tests.sh b/script/run_full_performance_tests.sh
similarity index 100%
rename from composable_kernel/script/run_full_performance_tests.sh
rename to script/run_full_performance_tests.sh
diff --git a/composable_kernel/script/run_performance_tests.sh b/script/run_performance_tests.sh
similarity index 100%
rename from composable_kernel/script/run_performance_tests.sh
rename to script/run_performance_tests.sh
diff --git a/composable_kernel/script/test_convnd_fwd.sh b/script/test_convnd_fwd.sh
similarity index 100%
rename from composable_kernel/script/test_convnd_fwd.sh
rename to script/test_convnd_fwd.sh
diff --git a/composable_kernel/script/test_reduce_no_index.sh b/script/test_reduce_no_index.sh
similarity index 100%
rename from composable_kernel/script/test_reduce_no_index.sh
rename to script/test_reduce_no_index.sh
diff --git a/composable_kernel/script/test_reduce_with_index.sh b/script/test_reduce_with_index.sh
similarity index 100%
rename from composable_kernel/script/test_reduce_with_index.sh
rename to script/test_reduce_with_index.sh
diff --git a/composable_kernel/script/uninstall_precommit.sh b/script/uninstall_precommit.sh
similarity index 100%
rename from composable_kernel/script/uninstall_precommit.sh
rename to script/uninstall_precommit.sh
diff --git a/composable_kernel/test/CMakeLists.txt b/test/CMakeLists.txt
similarity index 100%
rename from composable_kernel/test/CMakeLists.txt
rename to test/CMakeLists.txt
diff --git a/composable_kernel/test/batched_gemm/CMakeLists.txt b/test/batched_gemm/CMakeLists.txt
similarity index 100%
rename from composable_kernel/test/batched_gemm/CMakeLists.txt
rename to test/batched_gemm/CMakeLists.txt
diff --git a/composable_kernel/test/batched_gemm/batched_gemm_bf16.cpp b/test/batched_gemm/batched_gemm_bf16.cpp
similarity index 100%
rename from composable_kernel/test/batched_gemm/batched_gemm_bf16.cpp
rename to test/batched_gemm/batched_gemm_bf16.cpp
diff --git a/composable_kernel/test/batched_gemm/batched_gemm_fp16.cpp b/test/batched_gemm/batched_gemm_fp16.cpp
similarity index 100%
rename from composable_kernel/test/batched_gemm/batched_gemm_fp16.cpp
rename to test/batched_gemm/batched_gemm_fp16.cpp
diff --git a/composable_kernel/test/batched_gemm/batched_gemm_fp32.cpp b/test/batched_gemm/batched_gemm_fp32.cpp
similarity index 100%
rename from composable_kernel/test/batched_gemm/batched_gemm_fp32.cpp
rename to test/batched_gemm/batched_gemm_fp32.cpp
diff --git a/composable_kernel/test/batched_gemm/batched_gemm_int8.cpp b/test/batched_gemm/batched_gemm_int8.cpp
similarity index 100%
rename from composable_kernel/test/batched_gemm/batched_gemm_int8.cpp
rename to test/batched_gemm/batched_gemm_int8.cpp
diff --git a/composable_kernel/test/batched_gemm/test_batched_gemm.cpp b/test/batched_gemm/test_batched_gemm.cpp
similarity index 100%
rename from composable_kernel/test/batched_gemm/test_batched_gemm.cpp
rename to test/batched_gemm/test_batched_gemm.cpp
diff --git a/composable_kernel/test/batched_gemm_gemm/CMakeLists.txt b/test/batched_gemm_gemm/CMakeLists.txt
similarity index 100%
rename from composable_kernel/test/batched_gemm_gemm/CMakeLists.txt
rename to test/batched_gemm_gemm/CMakeLists.txt
diff --git a/composable_kernel/test/batched_gemm_gemm/test_batched_gemm_gemm_fp16.cpp b/test/batched_gemm_gemm/test_batched_gemm_gemm_fp16.cpp
similarity index 100%
rename from composable_kernel/test/batched_gemm_gemm/test_batched_gemm_gemm_fp16.cpp
rename to test/batched_gemm_gemm/test_batched_gemm_gemm_fp16.cpp
diff --git a/composable_kernel/test/batched_gemm_gemm/test_batched_gemm_gemm_util.hpp b/test/batched_gemm_gemm/test_batched_gemm_gemm_util.hpp
similarity index 100%
rename from composable_kernel/test/batched_gemm_gemm/test_batched_gemm_gemm_util.hpp
rename to test/batched_gemm_gemm/test_batched_gemm_gemm_util.hpp
diff --git a/composable_kernel/test/batched_gemm_multi_d/CMakeLists.txt b/test/batched_gemm_multi_d/CMakeLists.txt
similarity index 100%
rename from composable_kernel/test/batched_gemm_multi_d/CMakeLists.txt
rename to test/batched_gemm_multi_d/CMakeLists.txt
diff --git a/composable_kernel/test/batched_gemm_multi_d/test_batched_gemm_multi_d_dl.cpp b/test/batched_gemm_multi_d/test_batched_gemm_multi_d_dl.cpp
similarity index 100%
rename from composable_kernel/test/batched_gemm_multi_d/test_batched_gemm_multi_d_dl.cpp
rename to test/batched_gemm_multi_d/test_batched_gemm_multi_d_dl.cpp
diff --git a/composable_kernel/test/batched_gemm_reduce/CMakeLists.txt b/test/batched_gemm_reduce/CMakeLists.txt
similarity index 100%
rename from composable_kernel/test/batched_gemm_reduce/CMakeLists.txt
rename to test/batched_gemm_reduce/CMakeLists.txt
diff --git a/composable_kernel/test/batched_gemm_reduce/batched_gemm_reduce_fp16.cpp b/test/batched_gemm_reduce/batched_gemm_reduce_fp16.cpp
similarity index 100%
rename from composable_kernel/test/batched_gemm_reduce/batched_gemm_reduce_fp16.cpp
rename to test/batched_gemm_reduce/batched_gemm_reduce_fp16.cpp
diff --git a/composable_kernel/test/batched_gemm_softmax_gemm/CMakeLists.txt b/test/batched_gemm_softmax_gemm/CMakeLists.txt
similarity index 100%
rename from composable_kernel/test/batched_gemm_softmax_gemm/CMakeLists.txt
rename to test/batched_gemm_softmax_gemm/CMakeLists.txt
diff --git a/composable_kernel/test/batched_gemm_softmax_gemm/test_batched_gemm_softmax_gemm_fp16.cpp b/test/batched_gemm_softmax_gemm/test_batched_gemm_softmax_gemm_fp16.cpp
similarity index 100%
rename from composable_kernel/test/batched_gemm_softmax_gemm/test_batched_gemm_softmax_gemm_fp16.cpp
rename to test/batched_gemm_softmax_gemm/test_batched_gemm_softmax_gemm_fp16.cpp
diff --git a/composable_kernel/test/batched_gemm_softmax_gemm/test_batched_gemm_softmax_gemm_util.hpp b/test/batched_gemm_softmax_gemm/test_batched_gemm_softmax_gemm_util.hpp
similarity index 100%
rename from composable_kernel/test/batched_gemm_softmax_gemm/test_batched_gemm_softmax_gemm_util.hpp
rename to test/batched_gemm_softmax_gemm/test_batched_gemm_softmax_gemm_util.hpp
diff --git a/composable_kernel/test/batched_gemm_softmax_gemm_permute/CMakeLists.txt b/test/batched_gemm_softmax_gemm_permute/CMakeLists.txt
similarity index 100%
rename from composable_kernel/test/batched_gemm_softmax_gemm_permute/CMakeLists.txt
rename to test/batched_gemm_softmax_gemm_permute/CMakeLists.txt
diff --git a/composable_kernel/test/batched_gemm_softmax_gemm_permute/test_batched_gemm_bias_softmax_gemm_permute_bf16.cpp b/test/batched_gemm_softmax_gemm_permute/test_batched_gemm_bias_softmax_gemm_permute_bf16.cpp
similarity index 100%
rename from composable_kernel/test/batched_gemm_softmax_gemm_permute/test_batched_gemm_bias_softmax_gemm_permute_bf16.cpp
rename to test/batched_gemm_softmax_gemm_permute/test_batched_gemm_bias_softmax_gemm_permute_bf16.cpp
diff --git a/composable_kernel/test/batched_gemm_softmax_gemm_permute/test_batched_gemm_bias_softmax_gemm_permute_fp16.cpp b/test/batched_gemm_softmax_gemm_permute/test_batched_gemm_bias_softmax_gemm_permute_fp16.cpp
similarity index 100%
rename from composable_kernel/test/batched_gemm_softmax_gemm_permute/test_batched_gemm_bias_softmax_gemm_permute_fp16.cpp
rename to test/batched_gemm_softmax_gemm_permute/test_batched_gemm_bias_softmax_gemm_permute_fp16.cpp
diff --git a/composable_kernel/test/batched_gemm_softmax_gemm_permute/test_batched_gemm_bias_softmax_gemm_permute_util.hpp b/test/batched_gemm_softmax_gemm_permute/test_batched_gemm_bias_softmax_gemm_permute_util.hpp
similarity index 100%
rename from composable_kernel/test/batched_gemm_softmax_gemm_permute/test_batched_gemm_bias_softmax_gemm_permute_util.hpp
rename to test/batched_gemm_softmax_gemm_permute/test_batched_gemm_bias_softmax_gemm_permute_util.hpp
diff --git a/composable_kernel/test/batched_gemm_softmax_gemm_permute/test_batched_gemm_softmax_gemm_permute_bf16.cpp b/test/batched_gemm_softmax_gemm_permute/test_batched_gemm_softmax_gemm_permute_bf16.cpp
similarity index 100%
rename from composable_kernel/test/batched_gemm_softmax_gemm_permute/test_batched_gemm_softmax_gemm_permute_bf16.cpp
rename to test/batched_gemm_softmax_gemm_permute/test_batched_gemm_softmax_gemm_permute_bf16.cpp
diff --git a/composable_kernel/test/batched_gemm_softmax_gemm_permute/test_batched_gemm_softmax_gemm_permute_fp16.cpp b/test/batched_gemm_softmax_gemm_permute/test_batched_gemm_softmax_gemm_permute_fp16.cpp
similarity index 100%
rename from composable_kernel/test/batched_gemm_softmax_gemm_permute/test_batched_gemm_softmax_gemm_permute_fp16.cpp
rename to test/batched_gemm_softmax_gemm_permute/test_batched_gemm_softmax_gemm_permute_fp16.cpp
diff --git a/composable_kernel/test/batched_gemm_softmax_gemm_permute/test_batched_gemm_softmax_gemm_permute_util.hpp b/test/batched_gemm_softmax_gemm_permute/test_batched_gemm_softmax_gemm_permute_util.hpp
similarity index 100%
rename from composable_kernel/test/batched_gemm_softmax_gemm_permute/test_batched_gemm_softmax_gemm_permute_util.hpp
rename to test/batched_gemm_softmax_gemm_permute/test_batched_gemm_softmax_gemm_permute_util.hpp
diff --git a/composable_kernel/test/batchnorm/CMakeLists.txt b/test/batchnorm/CMakeLists.txt
similarity index 100%
rename from composable_kernel/test/batchnorm/CMakeLists.txt
rename to test/batchnorm/CMakeLists.txt
diff --git a/composable_kernel/test/batchnorm/batchnorm_bwd_rank_4.cpp b/test/batchnorm/batchnorm_bwd_rank_4.cpp
similarity index 100%
rename from composable_kernel/test/batchnorm/batchnorm_bwd_rank_4.cpp
rename to test/batchnorm/batchnorm_bwd_rank_4.cpp
diff --git a/composable_kernel/test/batchnorm/batchnorm_fwd_rank_4.cpp b/test/batchnorm/batchnorm_fwd_rank_4.cpp
similarity index 100%
rename from composable_kernel/test/batchnorm/batchnorm_fwd_rank_4.cpp
rename to test/batchnorm/batchnorm_fwd_rank_4.cpp
diff --git a/composable_kernel/test/batchnorm/batchnorm_infer_rank_4.cpp b/test/batchnorm/batchnorm_infer_rank_4.cpp
similarity index 100%
rename from composable_kernel/test/batchnorm/batchnorm_infer_rank_4.cpp
rename to test/batchnorm/batchnorm_infer_rank_4.cpp
diff --git a/composable_kernel/test/block_swizzle_test/block_swizzle_test.cpp b/test/block_swizzle_test/block_swizzle_test.cpp
similarity index 100%
rename from composable_kernel/test/block_swizzle_test/block_swizzle_test.cpp
rename to test/block_swizzle_test/block_swizzle_test.cpp
diff --git a/composable_kernel/test/block_swizzle_test/rebuild.sh b/test/block_swizzle_test/rebuild.sh
similarity index 100%
rename from composable_kernel/test/block_swizzle_test/rebuild.sh
rename to test/block_swizzle_test/rebuild.sh
diff --git a/composable_kernel/test/block_swizzle_test/simple_args.h b/test/block_swizzle_test/simple_args.h
similarity index 100%
rename from composable_kernel/test/block_swizzle_test/simple_args.h
rename to test/block_swizzle_test/simple_args.h
diff --git a/composable_kernel/test/block_to_ctile_map/CMakeLists.txt b/test/block_to_ctile_map/CMakeLists.txt
similarity index 100%
rename from composable_kernel/test/block_to_ctile_map/CMakeLists.txt
rename to test/block_to_ctile_map/CMakeLists.txt
diff --git a/composable_kernel/test/block_to_ctile_map/test_block_to_ctile_map.cpp b/test/block_to_ctile_map/test_block_to_ctile_map.cpp
similarity index 100%
rename from composable_kernel/test/block_to_ctile_map/test_block_to_ctile_map.cpp
rename to test/block_to_ctile_map/test_block_to_ctile_map.cpp
diff --git a/composable_kernel/test/contraction/CMakeLists.txt b/test/contraction/CMakeLists.txt
similarity index 100%
rename from composable_kernel/test/contraction/CMakeLists.txt
rename to test/contraction/CMakeLists.txt
diff --git a/composable_kernel/test/contraction/test_contraction.cpp b/test/contraction/test_contraction.cpp
similarity index 100%
rename from composable_kernel/test/contraction/test_contraction.cpp
rename to test/contraction/test_contraction.cpp
diff --git a/composable_kernel/test/contraction/test_contraction_interface.cpp b/test/contraction/test_contraction_interface.cpp
similarity index 100%
rename from composable_kernel/test/contraction/test_contraction_interface.cpp
rename to test/contraction/test_contraction_interface.cpp
diff --git a/composable_kernel/test/conv_tensor_rearrange/CMakeLists.txt b/test/conv_tensor_rearrange/CMakeLists.txt
similarity index 100%
rename from composable_kernel/test/conv_tensor_rearrange/CMakeLists.txt
rename to test/conv_tensor_rearrange/CMakeLists.txt
diff --git a/composable_kernel/test/conv_tensor_rearrange/test_conv_tensor_rearrange.cpp b/test/conv_tensor_rearrange/test_conv_tensor_rearrange.cpp
similarity index 100%
rename from composable_kernel/test/conv_tensor_rearrange/test_conv_tensor_rearrange.cpp
rename to test/conv_tensor_rearrange/test_conv_tensor_rearrange.cpp
diff --git a/composable_kernel/test/conv_tensor_rearrange/test_conv_tensor_rearrange_interface.cpp b/test/conv_tensor_rearrange/test_conv_tensor_rearrange_interface.cpp
similarity index 100%
rename from composable_kernel/test/conv_tensor_rearrange/test_conv_tensor_rearrange_interface.cpp
rename to test/conv_tensor_rearrange/test_conv_tensor_rearrange_interface.cpp
diff --git a/composable_kernel/test/conv_util/CMakeLists.txt b/test/conv_util/CMakeLists.txt
similarity index 100%
rename from composable_kernel/test/conv_util/CMakeLists.txt
rename to test/conv_util/CMakeLists.txt
diff --git a/composable_kernel/test/conv_util/conv_util.cpp b/test/conv_util/conv_util.cpp
similarity index 100%
rename from composable_kernel/test/conv_util/conv_util.cpp
rename to test/conv_util/conv_util.cpp
diff --git a/composable_kernel/test/convnd_bwd_data/CMakeLists.txt b/test/convnd_bwd_data/CMakeLists.txt
similarity index 100%
rename from composable_kernel/test/convnd_bwd_data/CMakeLists.txt
rename to test/convnd_bwd_data/CMakeLists.txt
diff --git a/composable_kernel/test/convnd_bwd_data/convnd_bwd_data.cpp b/test/convnd_bwd_data/convnd_bwd_data.cpp
similarity index 100%
rename from composable_kernel/test/convnd_bwd_data/convnd_bwd_data.cpp
rename to test/convnd_bwd_data/convnd_bwd_data.cpp
diff --git a/composable_kernel/test/convnd_fwd/CMakeLists.txt b/test/convnd_fwd/CMakeLists.txt
similarity index 100%
rename from composable_kernel/test/convnd_fwd/CMakeLists.txt
rename to test/convnd_fwd/CMakeLists.txt
diff --git a/composable_kernel/test/convnd_fwd/convnd_fwd.cpp b/test/convnd_fwd/convnd_fwd.cpp
similarity index 100%
rename from composable_kernel/test/convnd_fwd/convnd_fwd.cpp
rename to test/convnd_fwd/convnd_fwd.cpp
diff --git a/composable_kernel/test/data_type/CMakeLists.txt b/test/data_type/CMakeLists.txt
similarity index 100%
rename from composable_kernel/test/data_type/CMakeLists.txt
rename to test/data_type/CMakeLists.txt
diff --git a/composable_kernel/test/data_type/test_bf8.cpp b/test/data_type/test_bf8.cpp
similarity index 100%
rename from composable_kernel/test/data_type/test_bf8.cpp
rename to test/data_type/test_bf8.cpp
diff --git a/composable_kernel/test/data_type/test_fp8.cpp b/test/data_type/test_fp8.cpp
similarity index 100%
rename from composable_kernel/test/data_type/test_fp8.cpp
rename to test/data_type/test_fp8.cpp
diff --git a/composable_kernel/test/data_type/test_int4.cpp b/test/data_type/test_int4.cpp
similarity index 100%
rename from composable_kernel/test/data_type/test_int4.cpp
rename to test/data_type/test_int4.cpp
diff --git a/composable_kernel/test/data_type/type_convert_const.cpp b/test/data_type/type_convert_const.cpp
similarity index 100%
rename from composable_kernel/test/data_type/type_convert_const.cpp
rename to test/data_type/type_convert_const.cpp
diff --git a/composable_kernel/test/elementwise_normalization/CMakeLists.txt b/test/elementwise_normalization/CMakeLists.txt
similarity index 100%
rename from composable_kernel/test/elementwise_normalization/CMakeLists.txt
rename to test/elementwise_normalization/CMakeLists.txt
diff --git a/composable_kernel/test/elementwise_normalization/test_elementwise_layernorm_fp16.cpp b/test/elementwise_normalization/test_elementwise_layernorm_fp16.cpp
similarity index 100%
rename from composable_kernel/test/elementwise_normalization/test_elementwise_layernorm_fp16.cpp
rename to test/elementwise_normalization/test_elementwise_layernorm_fp16.cpp
diff --git a/composable_kernel/test/gemm/CMakeLists.txt b/test/gemm/CMakeLists.txt
similarity index 100%
rename from composable_kernel/test/gemm/CMakeLists.txt
rename to test/gemm/CMakeLists.txt
diff --git a/composable_kernel/test/gemm/gemm_bf16.cpp b/test/gemm/gemm_bf16.cpp
similarity index 100%
rename from composable_kernel/test/gemm/gemm_bf16.cpp
rename to test/gemm/gemm_bf16.cpp
diff --git a/composable_kernel/test/gemm/gemm_fp16.cpp b/test/gemm/gemm_fp16.cpp
similarity index 100%
rename from composable_kernel/test/gemm/gemm_fp16.cpp
rename to test/gemm/gemm_fp16.cpp
diff --git a/composable_kernel/test/gemm/gemm_fp32.cpp b/test/gemm/gemm_fp32.cpp
similarity index 100%
rename from composable_kernel/test/gemm/gemm_fp32.cpp
rename to test/gemm/gemm_fp32.cpp
diff --git a/composable_kernel/test/gemm/gemm_fp64.cpp b/test/gemm/gemm_fp64.cpp
similarity index 100%
rename from composable_kernel/test/gemm/gemm_fp64.cpp
rename to test/gemm/gemm_fp64.cpp
diff --git a/composable_kernel/test/gemm/gemm_int8.cpp b/test/gemm/gemm_int8.cpp
similarity index 100%
rename from composable_kernel/test/gemm/gemm_int8.cpp
rename to test/gemm/gemm_int8.cpp
diff --git a/composable_kernel/test/gemm/gemm_standalone_xdl_fp16.cpp b/test/gemm/gemm_standalone_xdl_fp16.cpp
similarity index 100%
rename from composable_kernel/test/gemm/gemm_standalone_xdl_fp16.cpp
rename to test/gemm/gemm_standalone_xdl_fp16.cpp
diff --git a/composable_kernel/test/gemm/gemm_util.hpp b/test/gemm/gemm_util.hpp
similarity index 100%
rename from composable_kernel/test/gemm/gemm_util.hpp
rename to test/gemm/gemm_util.hpp
diff --git a/composable_kernel/test/gemm/instance/gemm_f16_nn_instance.cpp b/test/gemm/instance/gemm_f16_nn_instance.cpp
similarity index 100%
rename from composable_kernel/test/gemm/instance/gemm_f16_nn_instance.cpp
rename to test/gemm/instance/gemm_f16_nn_instance.cpp
diff --git a/composable_kernel/test/gemm/instance/gemm_f16_nn_instance.hpp b/test/gemm/instance/gemm_f16_nn_instance.hpp
similarity index 100%
rename from composable_kernel/test/gemm/instance/gemm_f16_nn_instance.hpp
rename to test/gemm/instance/gemm_f16_nn_instance.hpp
diff --git a/composable_kernel/test/gemm/instance/gemm_f16_nt_instance.cpp b/test/gemm/instance/gemm_f16_nt_instance.cpp
similarity index 100%
rename from composable_kernel/test/gemm/instance/gemm_f16_nt_instance.cpp
rename to test/gemm/instance/gemm_f16_nt_instance.cpp
diff --git a/composable_kernel/test/gemm/instance/gemm_f16_nt_instance.hpp b/test/gemm/instance/gemm_f16_nt_instance.hpp
similarity index 100%
rename from composable_kernel/test/gemm/instance/gemm_f16_nt_instance.hpp
rename to test/gemm/instance/gemm_f16_nt_instance.hpp
diff --git a/composable_kernel/test/gemm/instance/gemm_f16_tn_instance.cpp b/test/gemm/instance/gemm_f16_tn_instance.cpp
similarity index 100%
rename from composable_kernel/test/gemm/instance/gemm_f16_tn_instance.cpp
rename to test/gemm/instance/gemm_f16_tn_instance.cpp
diff --git a/composable_kernel/test/gemm/instance/gemm_f16_tn_instance.hpp b/test/gemm/instance/gemm_f16_tn_instance.hpp
similarity index 100%
rename from composable_kernel/test/gemm/instance/gemm_f16_tn_instance.hpp
rename to test/gemm/instance/gemm_f16_tn_instance.hpp
diff --git a/composable_kernel/test/gemm/instance/gemm_f16_tt_instance.cpp b/test/gemm/instance/gemm_f16_tt_instance.cpp
similarity index 100%
rename from composable_kernel/test/gemm/instance/gemm_f16_tt_instance.cpp
rename to test/gemm/instance/gemm_f16_tt_instance.cpp
diff --git a/composable_kernel/test/gemm/instance/gemm_f16_tt_instance.hpp b/test/gemm/instance/gemm_f16_tt_instance.hpp
similarity index 100%
rename from composable_kernel/test/gemm/instance/gemm_f16_tt_instance.hpp
rename to test/gemm/instance/gemm_f16_tt_instance.hpp
diff --git a/composable_kernel/test/gemm/instance/gemm_wavelet_f16_tn_instance.cpp b/test/gemm/instance/gemm_wavelet_f16_tn_instance.cpp
similarity index 100%
rename from composable_kernel/test/gemm/instance/gemm_wavelet_f16_tn_instance.cpp
rename to test/gemm/instance/gemm_wavelet_f16_tn_instance.cpp
diff --git a/composable_kernel/test/gemm/instance/gemm_wavelet_f16_tn_instance.hpp b/test/gemm/instance/gemm_wavelet_f16_tn_instance.hpp
similarity index 100%
rename from composable_kernel/test/gemm/instance/gemm_wavelet_f16_tn_instance.hpp
rename to test/gemm/instance/gemm_wavelet_f16_tn_instance.hpp
diff --git a/composable_kernel/test/gemm/run_gemm_test.inc b/test/gemm/run_gemm_test.inc
similarity index 100%
rename from composable_kernel/test/gemm/run_gemm_test.inc
rename to test/gemm/run_gemm_test.inc
diff --git a/composable_kernel/test/gemm_layernorm/CMakeLists.txt b/test/gemm_layernorm/CMakeLists.txt
similarity index 100%
rename from composable_kernel/test/gemm_layernorm/CMakeLists.txt
rename to test/gemm_layernorm/CMakeLists.txt
diff --git a/composable_kernel/test/gemm_layernorm/test_gemm_add_relu_add_layernorm_fp16.cpp b/test/gemm_layernorm/test_gemm_add_relu_add_layernorm_fp16.cpp
similarity index 100%
rename from composable_kernel/test/gemm_layernorm/test_gemm_add_relu_add_layernorm_fp16.cpp
rename to test/gemm_layernorm/test_gemm_add_relu_add_layernorm_fp16.cpp
diff --git a/composable_kernel/test/gemm_reduce/CMakeLists.txt b/test/gemm_reduce/CMakeLists.txt
similarity index 100%
rename from composable_kernel/test/gemm_reduce/CMakeLists.txt
rename to test/gemm_reduce/CMakeLists.txt
diff --git a/composable_kernel/test/gemm_reduce/gemm_reduce_fp16.cpp b/test/gemm_reduce/gemm_reduce_fp16.cpp
similarity index 100%
rename from composable_kernel/test/gemm_reduce/gemm_reduce_fp16.cpp
rename to test/gemm_reduce/gemm_reduce_fp16.cpp
diff --git a/composable_kernel/test/gemm_split_k/CMakeLists.txt b/test/gemm_split_k/CMakeLists.txt
similarity index 100%
rename from composable_kernel/test/gemm_split_k/CMakeLists.txt
rename to test/gemm_split_k/CMakeLists.txt
diff --git a/composable_kernel/test/gemm_split_k/test_gemm_splitk.cpp b/test/gemm_split_k/test_gemm_splitk.cpp
similarity index 100%
rename from composable_kernel/test/gemm_split_k/test_gemm_splitk.cpp
rename to test/gemm_split_k/test_gemm_splitk.cpp
diff --git a/composable_kernel/test/gemm_split_k/test_gemm_splitk_ut_cases.inc b/test/gemm_split_k/test_gemm_splitk_ut_cases.inc
similarity index 100%
rename from composable_kernel/test/gemm_split_k/test_gemm_splitk_ut_cases.inc
rename to test/gemm_split_k/test_gemm_splitk_ut_cases.inc
diff --git a/composable_kernel/test/gemm_split_k/test_gemm_splitk_util.hpp b/test/gemm_split_k/test_gemm_splitk_util.hpp
similarity index 100%
rename from composable_kernel/test/gemm_split_k/test_gemm_splitk_util.hpp
rename to test/gemm_split_k/test_gemm_splitk_util.hpp
diff --git a/composable_kernel/test/grouped_convnd_bwd_data/CMakeLists.txt b/test/grouped_convnd_bwd_data/CMakeLists.txt
similarity index 100%
rename from composable_kernel/test/grouped_convnd_bwd_data/CMakeLists.txt
rename to test/grouped_convnd_bwd_data/CMakeLists.txt
diff --git a/composable_kernel/test/grouped_convnd_bwd_data/test_grouped_convnd_bwd_data.cpp b/test/grouped_convnd_bwd_data/test_grouped_convnd_bwd_data.cpp
similarity index 100%
rename from composable_kernel/test/grouped_convnd_bwd_data/test_grouped_convnd_bwd_data.cpp
rename to test/grouped_convnd_bwd_data/test_grouped_convnd_bwd_data.cpp
diff --git a/composable_kernel/test/grouped_convnd_bwd_data/test_grouped_convnd_bwd_data_interface_wmma.cpp b/test/grouped_convnd_bwd_data/test_grouped_convnd_bwd_data_interface_wmma.cpp
similarity index 100%
rename from composable_kernel/test/grouped_convnd_bwd_data/test_grouped_convnd_bwd_data_interface_wmma.cpp
rename to test/grouped_convnd_bwd_data/test_grouped_convnd_bwd_data_interface_wmma.cpp
diff --git a/composable_kernel/test/grouped_convnd_bwd_data/test_grouped_convnd_bwd_data_interface_xdl.cpp b/test/grouped_convnd_bwd_data/test_grouped_convnd_bwd_data_interface_xdl.cpp
similarity index 100%
rename from composable_kernel/test/grouped_convnd_bwd_data/test_grouped_convnd_bwd_data_interface_xdl.cpp
rename to test/grouped_convnd_bwd_data/test_grouped_convnd_bwd_data_interface_xdl.cpp
diff --git a/composable_kernel/test/grouped_convnd_bwd_weight/CMakeLists.txt b/test/grouped_convnd_bwd_weight/CMakeLists.txt
similarity index 100%
rename from composable_kernel/test/grouped_convnd_bwd_weight/CMakeLists.txt
rename to test/grouped_convnd_bwd_weight/CMakeLists.txt
diff --git a/composable_kernel/test/grouped_convnd_bwd_weight/test_grouped_convnd_bwd_weight.cpp b/test/grouped_convnd_bwd_weight/test_grouped_convnd_bwd_weight.cpp
similarity index 100%
rename from composable_kernel/test/grouped_convnd_bwd_weight/test_grouped_convnd_bwd_weight.cpp
rename to test/grouped_convnd_bwd_weight/test_grouped_convnd_bwd_weight.cpp
diff --git a/composable_kernel/test/grouped_convnd_bwd_weight/test_grouped_convnd_bwd_weight_interface_wmma.cpp b/test/grouped_convnd_bwd_weight/test_grouped_convnd_bwd_weight_interface_wmma.cpp
similarity index 100%
rename from composable_kernel/test/grouped_convnd_bwd_weight/test_grouped_convnd_bwd_weight_interface_wmma.cpp
rename to test/grouped_convnd_bwd_weight/test_grouped_convnd_bwd_weight_interface_wmma.cpp
diff --git a/composable_kernel/test/grouped_convnd_bwd_weight/test_grouped_convnd_bwd_weight_interface_xdl.cpp b/test/grouped_convnd_bwd_weight/test_grouped_convnd_bwd_weight_interface_xdl.cpp
similarity index 100%
rename from composable_kernel/test/grouped_convnd_bwd_weight/test_grouped_convnd_bwd_weight_interface_xdl.cpp
rename to test/grouped_convnd_bwd_weight/test_grouped_convnd_bwd_weight_interface_xdl.cpp
diff --git a/composable_kernel/test/grouped_convnd_fwd/CMakeLists.txt b/test/grouped_convnd_fwd/CMakeLists.txt
similarity index 100%
rename from composable_kernel/test/grouped_convnd_fwd/CMakeLists.txt
rename to test/grouped_convnd_fwd/CMakeLists.txt
diff --git a/composable_kernel/test/grouped_convnd_fwd/grouped_convnd_fwd.cpp b/test/grouped_convnd_fwd/grouped_convnd_fwd.cpp
similarity index 100%
rename from composable_kernel/test/grouped_convnd_fwd/grouped_convnd_fwd.cpp
rename to test/grouped_convnd_fwd/grouped_convnd_fwd.cpp
diff --git a/composable_kernel/test/grouped_convnd_fwd/test_grouped_convnd_fwd.cpp b/test/grouped_convnd_fwd/test_grouped_convnd_fwd.cpp
similarity index 100%
rename from composable_kernel/test/grouped_convnd_fwd/test_grouped_convnd_fwd.cpp
rename to test/grouped_convnd_fwd/test_grouped_convnd_fwd.cpp
diff --git a/composable_kernel/test/grouped_gemm/CMakeLists.txt b/test/grouped_gemm/CMakeLists.txt
similarity index 100%
rename from composable_kernel/test/grouped_gemm/CMakeLists.txt
rename to test/grouped_gemm/CMakeLists.txt
diff --git a/composable_kernel/test/grouped_gemm/test_grouped_gemm_interface.cpp b/test/grouped_gemm/test_grouped_gemm_interface.cpp
similarity index 100%
rename from composable_kernel/test/grouped_gemm/test_grouped_gemm_interface.cpp
rename to test/grouped_gemm/test_grouped_gemm_interface.cpp
diff --git a/composable_kernel/test/grouped_gemm/test_grouped_gemm_splitk.cpp b/test/grouped_gemm/test_grouped_gemm_splitk.cpp
similarity index 100%
rename from composable_kernel/test/grouped_gemm/test_grouped_gemm_splitk.cpp
rename to test/grouped_gemm/test_grouped_gemm_splitk.cpp
diff --git a/composable_kernel/test/grouped_gemm/test_grouped_gemm_ut_cases.inc b/test/grouped_gemm/test_grouped_gemm_ut_cases.inc
similarity index 100%
rename from composable_kernel/test/grouped_gemm/test_grouped_gemm_ut_cases.inc
rename to test/grouped_gemm/test_grouped_gemm_ut_cases.inc
diff --git a/composable_kernel/test/grouped_gemm/test_grouped_gemm_util.hpp b/test/grouped_gemm/test_grouped_gemm_util.hpp
similarity index 100%
rename from composable_kernel/test/grouped_gemm/test_grouped_gemm_util.hpp
rename to test/grouped_gemm/test_grouped_gemm_util.hpp
diff --git a/composable_kernel/test/image_to_column/CMakeLists.txt b/test/image_to_column/CMakeLists.txt
similarity index 100%
rename from composable_kernel/test/image_to_column/CMakeLists.txt
rename to test/image_to_column/CMakeLists.txt
diff --git a/composable_kernel/test/image_to_column/test_image_to_column.cpp b/test/image_to_column/test_image_to_column.cpp
similarity index 100%
rename from composable_kernel/test/image_to_column/test_image_to_column.cpp
rename to test/image_to_column/test_image_to_column.cpp
diff --git a/composable_kernel/test/image_to_column/test_image_to_column_interface.cpp b/test/image_to_column/test_image_to_column_interface.cpp
similarity index 100%
rename from composable_kernel/test/image_to_column/test_image_to_column_interface.cpp
rename to test/image_to_column/test_image_to_column_interface.cpp
diff --git a/composable_kernel/test/magic_number_division/CMakeLists.txt b/test/magic_number_division/CMakeLists.txt
similarity index 100%
rename from composable_kernel/test/magic_number_division/CMakeLists.txt
rename to test/magic_number_division/CMakeLists.txt
diff --git a/composable_kernel/test/magic_number_division/magic_number_division.cpp b/test/magic_number_division/magic_number_division.cpp
similarity index 100%
rename from composable_kernel/test/magic_number_division/magic_number_division.cpp
rename to test/magic_number_division/magic_number_division.cpp
diff --git a/composable_kernel/test/normalization/CMakeLists.txt b/test/normalization/CMakeLists.txt
similarity index 100%
rename from composable_kernel/test/normalization/CMakeLists.txt
rename to test/normalization/CMakeLists.txt
diff --git a/composable_kernel/test/normalization/test_groupnorm_fp16.cpp b/test/normalization/test_groupnorm_fp16.cpp
similarity index 100%
rename from composable_kernel/test/normalization/test_groupnorm_fp16.cpp
rename to test/normalization/test_groupnorm_fp16.cpp
diff --git a/composable_kernel/test/normalization/test_groupnorm_fp32.cpp b/test/normalization/test_groupnorm_fp32.cpp
similarity index 100%
rename from composable_kernel/test/normalization/test_groupnorm_fp32.cpp
rename to test/normalization/test_groupnorm_fp32.cpp
diff --git a/composable_kernel/test/normalization/test_layernorm2d_fp16.cpp b/test/normalization/test_layernorm2d_fp16.cpp
similarity index 100%
rename from composable_kernel/test/normalization/test_layernorm2d_fp16.cpp
rename to test/normalization/test_layernorm2d_fp16.cpp
diff --git a/composable_kernel/test/normalization/test_layernorm2d_fp32.cpp b/test/normalization/test_layernorm2d_fp32.cpp
similarity index 100%
rename from composable_kernel/test/normalization/test_layernorm2d_fp32.cpp
rename to test/normalization/test_layernorm2d_fp32.cpp
diff --git a/composable_kernel/test/pool/CMakeLists.txt b/test/pool/CMakeLists.txt
similarity index 100%
rename from composable_kernel/test/pool/CMakeLists.txt
rename to test/pool/CMakeLists.txt
diff --git a/composable_kernel/test/pool/test_avg_pool3d_bwd.cpp b/test/pool/test_avg_pool3d_bwd.cpp
similarity index 100%
rename from composable_kernel/test/pool/test_avg_pool3d_bwd.cpp
rename to test/pool/test_avg_pool3d_bwd.cpp
diff --git a/composable_kernel/test/pool/test_avg_pool3d_fwd.cpp b/test/pool/test_avg_pool3d_fwd.cpp
similarity index 100%
rename from composable_kernel/test/pool/test_avg_pool3d_fwd.cpp
rename to test/pool/test_avg_pool3d_fwd.cpp
diff --git a/composable_kernel/test/pool/test_max_pool3d_bwd.cpp b/test/pool/test_max_pool3d_bwd.cpp
similarity index 100%
rename from composable_kernel/test/pool/test_max_pool3d_bwd.cpp
rename to test/pool/test_max_pool3d_bwd.cpp
diff --git a/composable_kernel/test/pool/test_max_pool3d_fwd.cpp b/test/pool/test_max_pool3d_fwd.cpp
similarity index 100%
rename from composable_kernel/test/pool/test_max_pool3d_fwd.cpp
rename to test/pool/test_max_pool3d_fwd.cpp
diff --git a/composable_kernel/test/pool/test_pool_fwd_common.hpp b/test/pool/test_pool_fwd_common.hpp
similarity index 100%
rename from composable_kernel/test/pool/test_pool_fwd_common.hpp
rename to test/pool/test_pool_fwd_common.hpp
diff --git a/composable_kernel/test/reduce/CMakeLists.txt b/test/reduce/CMakeLists.txt
similarity index 100%
rename from composable_kernel/test/reduce/CMakeLists.txt
rename to test/reduce/CMakeLists.txt
diff --git a/composable_kernel/test/reduce/reduce_no_index.cpp b/test/reduce/reduce_no_index.cpp
similarity index 100%
rename from composable_kernel/test/reduce/reduce_no_index.cpp
rename to test/reduce/reduce_no_index.cpp
diff --git a/composable_kernel/test/reduce/reduce_with_index.cpp b/test/reduce/reduce_with_index.cpp
similarity index 100%
rename from composable_kernel/test/reduce/reduce_with_index.cpp
rename to test/reduce/reduce_with_index.cpp
diff --git a/composable_kernel/test/reference_conv_fwd/CMakeLists.txt b/test/reference_conv_fwd/CMakeLists.txt
similarity index 100%
rename from composable_kernel/test/reference_conv_fwd/CMakeLists.txt
rename to test/reference_conv_fwd/CMakeLists.txt
diff --git a/composable_kernel/test/reference_conv_fwd/reference_conv_fwd.cpp b/test/reference_conv_fwd/reference_conv_fwd.cpp
similarity index 100%
rename from composable_kernel/test/reference_conv_fwd/reference_conv_fwd.cpp
rename to test/reference_conv_fwd/reference_conv_fwd.cpp
diff --git a/composable_kernel/test/softmax/CMakeLists.txt b/test/softmax/CMakeLists.txt
similarity index 100%
rename from composable_kernel/test/softmax/CMakeLists.txt
rename to test/softmax/CMakeLists.txt
diff --git a/composable_kernel/test/softmax/test_softmax_interface.cpp b/test/softmax/test_softmax_interface.cpp
similarity index 100%
rename from composable_kernel/test/softmax/test_softmax_interface.cpp
rename to test/softmax/test_softmax_interface.cpp
diff --git a/composable_kernel/test/softmax/test_softmax_rank3.cpp b/test/softmax/test_softmax_rank3.cpp
similarity index 100%
rename from composable_kernel/test/softmax/test_softmax_rank3.cpp
rename to test/softmax/test_softmax_rank3.cpp
diff --git a/composable_kernel/test/softmax/test_softmax_rank4.cpp b/test/softmax/test_softmax_rank4.cpp
similarity index 100%
rename from composable_kernel/test/softmax/test_softmax_rank4.cpp
rename to test/softmax/test_softmax_rank4.cpp
diff --git a/composable_kernel/test/softmax/test_softmax_ut_cases.inc b/test/softmax/test_softmax_ut_cases.inc
similarity index 100%
rename from composable_kernel/test/softmax/test_softmax_ut_cases.inc
rename to test/softmax/test_softmax_ut_cases.inc
diff --git a/composable_kernel/test/softmax/test_softmax_util.hpp b/test/softmax/test_softmax_util.hpp
similarity index 100%
rename from composable_kernel/test/softmax/test_softmax_util.hpp
rename to test/softmax/test_softmax_util.hpp
diff --git a/composable_kernel/test/space_filling_curve/CMakeLists.txt b/test/space_filling_curve/CMakeLists.txt
similarity index 100%
rename from composable_kernel/test/space_filling_curve/CMakeLists.txt
rename to test/space_filling_curve/CMakeLists.txt
diff --git a/composable_kernel/test/space_filling_curve/space_filling_curve.cpp b/test/space_filling_curve/space_filling_curve.cpp
similarity index 100%
rename from composable_kernel/test/space_filling_curve/space_filling_curve.cpp
rename to test/space_filling_curve/space_filling_curve.cpp
diff --git a/composable_kernel/test/wmma_op/CMakeLists.txt b/test/wmma_op/CMakeLists.txt
similarity index 100%
rename from composable_kernel/test/wmma_op/CMakeLists.txt
rename to test/wmma_op/CMakeLists.txt
diff --git a/composable_kernel/test/wmma_op/wmma_op.cpp b/test/wmma_op/wmma_op.cpp
similarity index 100%
rename from composable_kernel/test/wmma_op/wmma_op.cpp
rename to test/wmma_op/wmma_op.cpp
diff --git a/composable_kernel/test/wmma_op/wmma_op_util.hpp b/test/wmma_op/wmma_op_util.hpp
similarity index 100%
rename from composable_kernel/test/wmma_op/wmma_op_util.hpp
rename to test/wmma_op/wmma_op_util.hpp
-- 
GitLab