Merge remote-tracking branch 'origin' into aosewski/ggemm_multi_d2

ae20247a · Adam Osewski · d1f7a3cf · a776978c · ae20247a · ae20247a
Commit ae20247a authored Feb 29, 2024 by Adam Osewski
20 changed files
--- a/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
@@ -5,6 +5,8 @@ list(APPEND GROUPED_GEMM_FIXED_NK_INSTANCES device_grouped_gemm_xdl_fixed_nk_f16
                                            device_grouped_gemm_xdl_fixed_nk_f16_fp8_f16_mk_kn_mn_instance.cpp
                                            device_grouped_gemm_xdl_fixed_nk_f16_fp8_f16_mk_nk_mn_instance.cpp
                                            device_grouped_gemm_xdl_fixed_nk_f16_i8_f16_mk_kn_mn_instance.cpp
-                                            device_grouped_gemm_xdl_fixed_nk_f16_i8_f16_mk_nk_mn_instance.cpp)
+                                            device_grouped_gemm_xdl_fixed_nk_f16_i8_f16_mk_nk_mn_instance.cpp
+                                            device_grouped_gemm_xdl_fixed_nk_bf16_i8_bf16_mk_kn_mn_instance.cpp
+                                            device_grouped_gemm_xdl_fixed_nk_bf16_i8_bf16_mk_nk_mn_instance.cpp)

 add_instance_library(device_grouped_gemm_fixed_nk_instance ${GROUPED_GEMM_FIXED_NK_INSTANCES})
--- a/library/src/tensor_operation_instance/gpu/grouped_gemm_fixed_nk/device_grouped_gemm_xdl_fixed_nk_bf16_i8_bf16_mk_kn_mn_instance.cpp
+++ b/library/src/tensor_operation_instance/gpu/grouped_gemm_fixed_nk/device_grouped_gemm_xdl_fixed_nk_bf16_i8_bf16_mk_kn_mn_instance.cpp
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl_fixed_nk.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using I8   = int8_t;
+using BF16 = ck::bhalf_t;
+using F32  = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using DsDataType                     = ck::Tuple<>;
+using DsLayout                       = ck::Tuple<>;
+using PassThrough                    = ck::tensor_operation::element_wise::PassThrough;
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+using device_grouped_gemm_xdl_fixed_nk_bf16_i8_bf16_mk_kn_mn_irregular_tile_instances = std::tuple<
+    // clang-format off
+        //############################|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|      DsData| EData|           A|           B|           C|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //############################| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|        Type|  Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //############################|       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //############################|       |       |            |       |      |      |        |         |            |      |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedGemm_Xdl_Fixed_NK<    Row,    Row,    DsLayout,    Row,   BF16,    I8,     F32,      F32,  DsDataType,   BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2, S< 1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl_Fixed_NK<    Row,    Row,    DsLayout,    Row,   BF16,    I8,     F32,      F32,  DsDataType,   BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,   128,   128,    32,   8,   8,   32,   32,    2,    2, S< 1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl_Fixed_NK<    Row,    Row,    DsLayout,    Row,   BF16,    I8,     F32,      F32,  DsDataType,   BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   2,   32,   32,    2,    1, S< 1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 16,16, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl_Fixed_NK<    Row,    Row,    DsLayout,    Row,   BF16,    I8,     F32,      F32,  DsDataType,   BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1, S< 1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              1,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl_Fixed_NK<    Row,    Row,    DsLayout,    Row,   BF16,    I8,     F32,      F32,  DsDataType,   BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   2,   32,   32,    1,    2, S< 1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              2,         0,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl_Fixed_NK<    Row,    Row,    DsLayout,    Row,   BF16,    I8,     F32,      F32,  DsDataType,   BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2, S< 1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl_Fixed_NK<    Row,    Row,    DsLayout,    Row,   BF16,    I8,     F32,      F32,  DsDataType,   BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   2,   32,   32,    2,    2, S< 1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 16, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              2,         0,           1,           1,               S<1, 32, 1, 4>,              8>,
+        DeviceGroupedGemm_Xdl_Fixed_NK<    Row,    Row,    DsLayout,    Row,   BF16,    I8,     F32,      F32,  DsDataType,   BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,   128,    64,    32,   8,   8,   32,   32,    2,    2, S< 1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8>,
+        DeviceGroupedGemm_Xdl_Fixed_NK<    Row,    Row,    DsLayout,    Row,   BF16,    I8,     F32,      F32,  DsDataType,   BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   2,   32,   32,    2,    2, S< 1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              2,         0,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl_Fixed_NK<    Row,    Row,    DsLayout,    Row,   BF16,    I8,     F32,      F32,  DsDataType,   BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,    64,   128,    32,   8,   8,   32,   32,    2,    2, S< 1, 4, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 32, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              4,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>
+    // clang-format on
+    >;
+
+void add_device_grouped_gemm_xdl_fixed_nk_bf16_i8_bf16_mk_kn_mn_instances(
+    std::vector<std::unique_ptr<DeviceGroupedGemmFixedNK<Row,
+                                                         Row,
+                                                         DsLayout,
+                                                         Row,
+                                                         BF16,
+                                                         I8,
+                                                         DsDataType,
+                                                         BF16,
+                                                         PassThrough,
+                                                         PassThrough,
+                                                         PassThrough>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_grouped_gemm_xdl_fixed_nk_bf16_i8_bf16_mk_kn_mn_irregular_tile_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
--- a/library/src/tensor_operation_instance/gpu/grouped_gemm_fixed_nk/device_grouped_gemm_xdl_fixed_nk_bf16_i8_bf16_mk_nk_mn_instance.cpp
+++ b/library/src/tensor_operation_instance/gpu/grouped_gemm_fixed_nk/device_grouped_gemm_xdl_fixed_nk_bf16_i8_bf16_mk_nk_mn_instance.cpp
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl_fixed_nk.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using I8   = int8_t;
+using BF16 = ck::bhalf_t;
+using F32  = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using DsDataType                     = ck::Tuple<>;
+using DsLayout                       = ck::Tuple<>;
+using PassThrough                    = ck::tensor_operation::element_wise::PassThrough;
+static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+using device_grouped_gemm_xdl_fixed_nk_bf16_i8_bf16_mk_nk_mn_irregular_tile_instances = std::tuple<
+    // clang-format off
+        //############################|      A|      B|          Ds|      E| AData| BData| AccData| CShuffle|      DsData| EData|           A|           B|         CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+        //############################| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|        Type|  Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+        //############################|       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+        //############################|       |       |            |       |      |      |        |         |            |      |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        DeviceGroupedGemm_Xdl_Fixed_NK<    Row,    Col,    DsLayout,    Row,  BF16,    I8,     F32,      F32,  DsDataType,  BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,   128,   256,    64,   8,   8,   32,   32,    2,    4,  S<1, 8, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl_Fixed_NK<    Row,    Col,    DsLayout,    Row,  BF16,    I8,     F32,      F32,  DsDataType,  BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,   128,   128,    64,   8,   8,   32,   32,    2,    2,  S<1, 8, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl_Fixed_NK<    Row,    Col,    DsLayout,    Row,  BF16,    I8,     F32,      F32,  DsDataType,  BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,   128,    64,    64,   8,   8,   32,   32,    2,    1,  S<1, 8, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl_Fixed_NK<    Row,    Col,    DsLayout,    Row,  BF16,    I8,     F32,      F32,  DsDataType,  BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,    64,   128,    64,   8,   8,   32,   32,    1,    2,  S<1, 8, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 32, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl_Fixed_NK<    Row,    Col,    DsLayout,    Row,  BF16,    I8,     F32,      F32,  DsDataType,  BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,   128,   128,    64,   8,   8,   32,   32,    4,    2,  S<1, 8, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl_Fixed_NK<    Row,    Col,    DsLayout,    Row,  BF16,    I8,     F32,      F32,  DsDataType,  BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,   128,    64,    64,   8,   8,   32,   32,    2,    2,  S<1, 8, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8>,
+        DeviceGroupedGemm_Xdl_Fixed_NK<    Row,    Col,    DsLayout,    Row,  BF16,    I8,     F32,      F32,  DsDataType,  BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,    64,   128,    64,   8,   8,   32,   32,    2,    2,  S<1, 8, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl_Fixed_NK<    Row,    Col,    DsLayout,    Row,  BF16,    I8,     F32,      F32,  DsDataType,  BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,   128,    32,    64,   8,   8,   32,   32,    2,    1,  S<1, 8, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 4>,              8>,
+        DeviceGroupedGemm_Xdl_Fixed_NK<    Row,    Col,    DsLayout,    Row,  BF16,    I8,     F32,      F32,  DsDataType,  BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,    32,   128,    64,   8,   8,   32,   32,    1,    2,  S<1, 8, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl_Fixed_NK<    Row,    Col,    DsLayout,    Row,  BF16,    I8,     F32,      F32,  DsDataType,  BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   128,    32,   256,    64,   8,   8,   32,   32,    1,    4,  S<1, 8, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>,
+        DeviceGroupedGemm_Xdl_Fixed_NK<    Row,    Col,    DsLayout,    Row,  BF16,    I8,     F32,      F32,  DsDataType,  BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,    64,    64,    64,    64,   8,   8,   32,   32,    2,    2,  S<1, 8,  8, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8,  8, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,              8>,
+        DeviceGroupedGemm_Xdl_Fixed_NK<    Row,    Col,    DsLayout,    Row,  BF16,    I8,     F32,      F32,  DsDataType,  BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,    64,    64,    32,    64,   8,   8,   32,   32,    2,    1,  S<1, 8,  8, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8,  8, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,              8>,
+        DeviceGroupedGemm_Xdl_Fixed_NK<    Row,    Col,    DsLayout,    Row,  BF16,    I8,     F32,      F32,  DsDataType,  BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,    64,    32,    64,    64,   8,   8,   32,   32,    1,    2,  S<1, 8,  8, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 8,  8, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 16, 1, 4>,              8>
+    // clang-format on
+    >;
+
+void add_device_grouped_gemm_xdl_fixed_nk_bf16_i8_bf16_mk_nk_mn_instances(
+    std::vector<std::unique_ptr<DeviceGroupedGemmFixedNK<Row,
+                                                         Col,
+                                                         DsLayout,
+                                                         Row,
+                                                         BF16,
+                                                         I8,
+                                                         DsDataType,
+                                                         BF16,
+                                                         PassThrough,
+                                                         PassThrough,
+                                                         PassThrough>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_grouped_gemm_xdl_fixed_nk_bf16_i8_bf16_mk_nk_mn_irregular_tile_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
--- a/library/src/tensor_operation_instance/gpu/permute_scale/CMakeLists.txt
+++ b/library/src/tensor_operation_instance/gpu/permute_scale/CMakeLists.txt
 add_instance_library(device_permute_scale_instance 
-	device_permute_scale_instances.cpp)
+	device_permute_scale_1d_instances.cpp
+	device_permute_scale_2d_instances.cpp
+	device_permute_scale_3d_instances.cpp
+	device_permute_scale_4d_instances.cpp
+	device_permute_scale_5d_instances.cpp
+	device_permute_scale_6d_instances.cpp)
--- a/library/src/tensor_operation_instance/gpu/permute_scale/device_permute_scale_1d_instances.cpp
+++ b/library/src/tensor_operation_instance/gpu/permute_scale/device_permute_scale_1d_instances.cpp
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/gpu/permute_scale/device_permute_scale_instances.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_permute_scale_1d_f16_instances(
+    std::vector<std::unique_ptr<
+        DeviceElementwise<ck::Tuple<F16>, ck::Tuple<F16>, Pass, UnaryOp, Scale, 1>>>& instances)
+{
+    add_device_operation_instances(instances, device_permute_scale_f16_instances<1>{});
+}
+
+void add_device_permute_scale_1d_f32_instances(
+    std::vector<std::unique_ptr<
+        DeviceElementwise<ck::Tuple<F32>, ck::Tuple<F32>, Pass, UnaryOp, Scale, 1>>>& instances)
+{
+    add_device_operation_instances(instances, device_permute_scale_f32_instances<1>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
--- a/library/src/tensor_operation_instance/gpu/permute_scale/device_permute_scale_2d_instances.cpp
+++ b/library/src/tensor_operation_instance/gpu/permute_scale/device_permute_scale_2d_instances.cpp
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/gpu/permute_scale/device_permute_scale_instances.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_permute_scale_2d_f16_instances(
+    std::vector<std::unique_ptr<
+        DeviceElementwise<ck::Tuple<F16>, ck::Tuple<F16>, Pass, UnaryOp, Scale, 2>>>& instances)
+{
+    add_device_operation_instances(instances, device_permute_scale_f16_instances<2>{});
+}
+
+void add_device_permute_scale_2d_f32_instances(
+    std::vector<std::unique_ptr<
+        DeviceElementwise<ck::Tuple<F32>, ck::Tuple<F32>, Pass, UnaryOp, Scale, 2>>>& instances)
+{
+    add_device_operation_instances(instances, device_permute_scale_f32_instances<2>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
--- a/library/src/tensor_operation_instance/gpu/permute_scale/device_permute_scale_3d_instances.cpp
+++ b/library/src/tensor_operation_instance/gpu/permute_scale/device_permute_scale_3d_instances.cpp
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/gpu/permute_scale/device_permute_scale_instances.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_permute_scale_3d_f16_instances(
+    std::vector<std::unique_ptr<
+        DeviceElementwise<ck::Tuple<F16>, ck::Tuple<F16>, Pass, UnaryOp, Scale, 3>>>& instances)
+{
+    add_device_operation_instances(instances, device_permute_scale_f16_instances<3>{});
+}
+
+void add_device_permute_scale_3d_f32_instances(
+    std::vector<std::unique_ptr<
+        DeviceElementwise<ck::Tuple<F32>, ck::Tuple<F32>, Pass, UnaryOp, Scale, 3>>>& instances)
+{
+    add_device_operation_instances(instances, device_permute_scale_f32_instances<3>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
--- a/library/src/tensor_operation_instance/gpu/permute_scale/device_permute_scale_4d_instances.cpp
+++ b/library/src/tensor_operation_instance/gpu/permute_scale/device_permute_scale_4d_instances.cpp
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/gpu/permute_scale/device_permute_scale_instances.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_permute_scale_4d_f16_instances(
+    std::vector<std::unique_ptr<
+        DeviceElementwise<ck::Tuple<F16>, ck::Tuple<F16>, Pass, UnaryOp, Scale, 4>>>& instances)
+{
+    add_device_operation_instances(instances, device_permute_scale_f16_instances<4>{});
+}
+
+void add_device_permute_scale_4d_f32_instances(
+    std::vector<std::unique_ptr<
+        DeviceElementwise<ck::Tuple<F32>, ck::Tuple<F32>, Pass, UnaryOp, Scale, 4>>>& instances)
+{
+    add_device_operation_instances(instances, device_permute_scale_f32_instances<4>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
--- a/library/src/tensor_operation_instance/gpu/permute_scale/device_permute_scale_5d_instances.cpp
+++ b/library/src/tensor_operation_instance/gpu/permute_scale/device_permute_scale_5d_instances.cpp
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/gpu/permute_scale/device_permute_scale_instances.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_permute_scale_5d_f16_instances(
+    std::vector<std::unique_ptr<
+        DeviceElementwise<ck::Tuple<F16>, ck::Tuple<F16>, Pass, UnaryOp, Scale, 5>>>& instances)
+{
+    add_device_operation_instances(instances, device_permute_scale_f16_instances<5>{});
+}
+
+void add_device_permute_scale_5d_f32_instances(
+    std::vector<std::unique_ptr<
+        DeviceElementwise<ck::Tuple<F32>, ck::Tuple<F32>, Pass, UnaryOp, Scale, 5>>>& instances)
+{
+    add_device_operation_instances(instances, device_permute_scale_f32_instances<5>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
--- a/library/src/tensor_operation_instance/gpu/permute_scale/device_permute_scale_6d_instances.cpp
+++ b/library/src/tensor_operation_instance/gpu/permute_scale/device_permute_scale_6d_instances.cpp
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/gpu/permute_scale/device_permute_scale_instances.hpp"
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_permute_scale_6d_f16_instances(
+    std::vector<std::unique_ptr<
+        DeviceElementwise<ck::Tuple<F16>, ck::Tuple<F16>, Pass, UnaryOp, Scale, 6>>>& instances)
+{
+    add_device_operation_instances(instances, device_permute_scale_f16_instances<6>{});
+}
+
+void add_device_permute_scale_6d_f32_instances(
+    std::vector<std::unique_ptr<
+        DeviceElementwise<ck::Tuple<F32>, ck::Tuple<F32>, Pass, UnaryOp, Scale, 6>>>& instances)
+{
+    add_device_operation_instances(instances, device_permute_scale_f32_instances<6>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
--- a/profiler/README.md
+++ b/profiler/README.md
@@ -37,9 +37,9 @@ Best Perf: 1.1933 ms, 107.977 TFlops, 79.0848 GB/s
 ################          op datatype  in_layout   wei_layout  out_layout  verify  init  log  repeat  N__ K___ C___ Y X Hi__ Wi__ Strides Dilations LeftPads RightPads
 ./bin/ckProfiler  conv2d_fwd        1          1            1           1       1     1    0       5  128  256  192 3 3   71   71     2 2       1 1      1 1       1 1
 ```
-
 Result (MI100 @ 1087Mhz, 133.5TFlops peak FP16)
-```
+
+```bash
 in_n_c_hi_wi: dim 4, lengths {128, 192, 71, 71}, strides {967872, 1, 13632, 192}
 wei_k_c_y_x: dim 4, lengths {256, 192, 3, 3}, strides {1728, 1, 576, 192}
 out_n_k_ho_wo: dim 4, lengths {128, 256, 36, 36}, strides {331776, 1, 9216, 256}
@@ -104,6 +104,7 @@ arg.b_grid_desc_k0_n0_n1_k1_{2048, 4096, 2}
 arg.e_grid_desc_m_n_{ 4096, 4096}
 ....
 Best Perf: 58.0306 ms, 37.8942 TFlops, 27.7545 GB/s
+```
 ## Profile grouped convolution backward data kernels
 ```bash
 # arg1: tensor operation (grouped_conv_bwd_data: Grouped Convolution Backward Data)
@@ -129,10 +130,11 @@ Best Perf: 58.0306 ms, 37.8942 TFlops, 27.7545 GB/s
 ################                   op   datatype  layout  verify  init  log  time  Ndims  G  N   K   C  Y  X  Hi  Wi  Sy  Sx  Dy  Dx  LeftPy  LeftPx  RightPy  RightPx
 ./bin/ckProfiler grouped_conv_bwd_data          1       0       1     1    0     1      2 32  4 192 192  3  3  28  28   1   1   1   1       1       1        1        1

- ```
+```

 Result (MI100, FP16, GNHWC_GKYXC_GNHWK)
-```
+
+```bash
 out: dim 5, lengths {32, 4, 192, 28, 28}, strides {602112, 150528, 1, 5376, 192}
 wei: dim 5, lengths {32, 192, 192, 3, 3}, strides {331776, 1728, 1, 576, 192}
 in: dim 5, lengths {32, 4, 192, 28, 28}, strides {602112, 150528, 1, 5376, 192}
@@ -173,10 +175,11 @@ GB/s: 127.947
 ################                   op   datatype  layout  verify  init  log  time  Ndims  G   N   K   C  Y  X  Hi  Wi  Sy  Sx  Dy  Dx  LeftPy  LeftPx  RightPy  RightPx  SplitK
 ./bin/ckProfiler grouped_conv_bwd_weight         1       1      0     1    0     1      2 32 256 256 512  3  3  28  28   1   1   1   1       1       0        0        0       1

- ```
+```

 Result (MI100, FP16, GNHWC_GKYXC_GNHWK)
-```
+
+```bash
 input: dim 5, lengths {32, 512, 1024, 28, 28}, strides {411041792, 802816, 1, 28672, 1024}
 weight: dim 5, lengths {32, 512, 1024, 3, 3}, strides {4718592, 9216, 1, 3072, 1024}
 output: dim 5, lengths {32, 512, 512, 26, 26}, strides {177209344, 346112, 1, 13312, 512}
@@ -190,8 +193,9 @@ GB/s: 69.2301
 Note: This kernel use atomic add, this will cause output buffer to be accumulated multiple times, causing verification failure. To work around it, do not use CK's own timer and do verification at the same time.

 ## Profile image to column/column to image kernels
+
 ```bash
-# arg1: tensor operation (" OP_NAME ": " OP_DESC ")
+# arg1: tensor operation ( conv_tensor_rearrange : Conv Tensor Rearrange )
 # arg2: data type (0: Input fp32, Weight fp32, Output fp32
 #                  1: Input fp16, Weight fp16, Output fp16
 #                  2: Input bf16, Weight bf16, Output bf16
@@ -216,10 +220,11 @@ Note: This kernel use atomic add, this will cause output buffer to be accumulate
 ################                   op   datatype  layout  verify  init  log  time opType Ndims  G   N   K   C  Y  X  Hi  Wi  Sy  Sx  Dy  Dx  LeftPy  LeftPx  RightPy  RightPx
 ./bin/ckProfiler conv_tensor_rearrange          0       0       0     1    0     1      0     2  1 256   1 512  3  3   28  28   1   1   1   1        0       0       0        0

- ```
+```

 Result (MI210, FP32, NHWC)
-```
+
+```bash
 input: dim 5, lengths {1, 256, 512, 28, 28}, strides {102760448, 401408, 1, 14336, 512}
 output: dim 2, lengths {173056, 4608}, strides {4608, 1}
 ....
@@ -229,3 +234,30 @@ avg_time: 3.12326
 GB/s: 2042.59
 ```
 Note: Column to image kernel adds to the output memory, this will cause output buffer to be accumulated multiple times, causing verification failure. To work around it, do not use CK's own timer and do verification at the same time.
+
+## Profile Permute scale kernels
+
+```bash
+# arg1: tensor operation ( permute_scale : Permute Scale )
+# arg2: data type (0: Input fp32, Output fp32
+#                  1: Input fp16, Output fp16
+# arg4: verification (0: no, 1: yes)
+# arg5: initialization (0: no init, 1: integer value, 2: decimal value)
+# arg6: print tensor value (0: no; 1: yes)
+# arg7: time kernel (0: no, 1: yes)
+# from arg8: tensor lengths
+#            input strides
+#            output strides
+
+################            op datatype  verify  init  log  time  dim0 dim1 dim2 in_stride0 in_stride1 in_stride2 out_stride0 out_stride1 out_stride2
+./bin/ckProfiler permute_scale        0       1     1    0     1    64   64   64       4096         64          1           1          64        4096
+```
+
+Result (MI100, FP32)
+
+```bash
+A: dim 3, lengths {64, 64, 64}, strides {4096, 64, 1}
+B: dim 3, lengths {64, 64, 64}, strides {1, 64, 4096}
+....
+Best perf = 0.0146878 ms, 142.782 GB/s, DeviceElementwiseNormalizationImpl<3, 2>
+```
--- a/profiler/include/profiler/profile_elementwise_layernorm_impl.hpp
+++ b/profiler/include/profiler/profile_elementwise_layernorm_impl.hpp
@@ -233,7 +233,7 @@ bool profile_elementwise_layernorm_impl(int do_verification,
            y_dev.FromDevice(y.mData.data());

            bool pass =
-                ck::utils::check_err(y.mData, host_y.mData, "Error: Incorrect results", 1e-3, 1e-3);
+                ck::utils::check_err(y.mData, host_y.mData, "Error: Incorrect results", 5e-3, 5e-3);

            if(do_log)
            {

--- a/profiler/include/profiler/profile_gemm_add_impl.hpp
+++ b/profiler/include/profiler/profile_gemm_add_impl.hpp
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iomanip>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/gemm_add.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+
+namespace ck {
+namespace profiler {
+
+template <typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename D0DataType,
+          typename EDataType,
+          typename ALayout,
+          typename BLayout,
+          typename D0Layout,
+          typename ELayout>
+bool profile_gemm_add_impl(int do_verification,
+                           int init_method,
+                           bool /*do_log*/,
+                           bool time_kernel,
+                           int M,
+                           int N,
+                           int K,
+                           int StrideA,
+                           int StrideB,
+                           int StrideD0,
+                           int StrideE)
+{
+    auto f_host_tensor_descriptor =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            using namespace ck::literals;
+
+            if(is_same<decltype(layout), tensor_layout::gemm::RowMajor>::value)
+            {
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
+            }
+            else
+            {
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
+            }
+        };
+
+    Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
+    Tensor<D0DataType> d0_m_n(f_host_tensor_descriptor(M, N, StrideD0, D0Layout{}));
+    Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+    Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+
+    std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
+    std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
+    std::cout << "d0_m_n: " << d0_m_n.mDesc << std::endl;
+    std::cout << "e_m_n: " << e_m_n_device_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        d0_m_n.GenerateTensorValue(GeneratorTensor_2<D0DataType>{-5, 5});
+        break;
+    default:
+        a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        d0_m_n.GenerateTensorValue(GeneratorTensor_3<D0DataType>{0.0, 1.0});
+    }
+
+    using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+    using Add         = ck::tensor_operation::element_wise::Add;
+
+    using AElementOp   = PassThrough;
+    using BElementOp   = PassThrough;
+    using CDEElementOp = Add;
+
+    const auto a_element_op   = AElementOp{};
+    const auto b_element_op   = BElementOp{};
+    const auto cde_element_op = CDEElementOp{};
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGemmMultipleD<
+        ALayout,
+        BLayout,
+        ck::Tuple<D0Layout>,
+        ELayout,
+        ADataType,
+        BDataType,
+        ck::Tuple<D0DataType>,
+        EDataType,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::Add>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    // run reference
+    if(do_verification)
+    {
+        Tensor<AccDataType> c_m_n({M, N});
+
+        using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                                BDataType,
+                                                                                AccDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                BElementOp,
+                                                                                PassThrough>;
+
+        auto ref_gemm    = ReferenceGemmInstance{};
+        auto ref_invoker = ref_gemm.MakeInvoker();
+
+        auto ref_argument =
+            ref_gemm.MakeArgument(a_m_k, b_k_n, c_m_n, a_element_op, b_element_op, PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        for(int m = 0; m < M; ++m)
+        {
+            for(int n = 0; n < N; ++n)
+            {
+                cde_element_op(e_m_n_host_result(m, n), c_m_n(m, n), d0_m_n(m, n));
+            }
+        }
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem d0_m_n_device_buf(sizeof(D0DataType) * d0_m_n.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a_m_k.mData.data());
+    b_device_buf.ToDevice(b_k_n.mData.data());
+    d0_m_n_device_buf.ToDevice(d0_m_n.mData.data());
+
+    std::string best_op_name;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    bool pass = true;
+
+    // profile device operation instances
+    for(auto& op_ptr : op_ptrs)
+    {
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            a_device_buf.GetDeviceBuffer(),
+            b_device_buf.GetDeviceBuffer(),
+            std::array<const void*, 1>{d0_m_n_device_buf.GetDeviceBuffer()},
+            e_device_buf.GetDeviceBuffer(),
+            M,
+            N,
+            K,
+            StrideA,
+            StrideB,
+            std::array<ck::index_t, 1>{StrideD0},
+            StrideE,
+            a_element_op,
+            b_element_op,
+            cde_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            // re-init E to zero before profiling a kernel
+            e_device_buf.SetZero();
+
+            float ave_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+            std::size_t flop = std::size_t(2) * M * N * K;
+
+            std::size_t num_btype =
+                sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(EDataType) * M * N;
+
+            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+
+            if(do_verification)
+            {
+                e_device_buf.FromDevice(e_m_n_device_result.mData.data());
+
+                pass = pass && ck::utils::check_err(e_m_n_device_result, e_m_n_host_result);
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
+              << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    return pass;
+}
+
+} // namespace profiler
+} // namespace ck
--- a/profiler/include/profiler/profile_gemm_add_relu_impl.hpp
+++ b/profiler/include/profiler/profile_gemm_add_relu_impl.hpp
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iomanip>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/gemm_add_relu.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+
+namespace ck {
+namespace profiler {
+
+template <typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename D0DataType,
+          typename EDataType,
+          typename ALayout,
+          typename BLayout,
+          typename D0Layout,
+          typename ELayout>
+bool profile_gemm_add_relu_impl(int do_verification,
+                                int init_method,
+                                bool /*do_log*/,
+                                bool time_kernel,
+                                int M,
+                                int N,
+                                int K,
+                                int StrideA,
+                                int StrideB,
+                                int StrideD0,
+                                int StrideE)
+{
+    auto f_host_tensor_descriptor =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            using namespace ck::literals;
+
+            if(is_same<decltype(layout), tensor_layout::gemm::RowMajor>::value)
+            {
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
+            }
+            else
+            {
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
+            }
+        };
+
+    Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
+    Tensor<D0DataType> d0_m_n(f_host_tensor_descriptor(M, N, StrideD0, D0Layout{}));
+    Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+    Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+
+    std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
+    std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
+    std::cout << "d0_m_n: " << d0_m_n.mDesc << std::endl;
+    std::cout << "e_m_n: " << e_m_n_device_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        d0_m_n.GenerateTensorValue(GeneratorTensor_2<D0DataType>{-5, 5});
+        break;
+    default:
+        a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        d0_m_n.GenerateTensorValue(GeneratorTensor_3<D0DataType>{0.0, 1.0});
+    }
+
+    using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+    using AddRelu     = ck::tensor_operation::element_wise::AddRelu;
+
+    using AElementOp   = PassThrough;
+    using BElementOp   = PassThrough;
+    using CDEElementOp = AddRelu;
+
+    const auto a_element_op   = AElementOp{};
+    const auto b_element_op   = BElementOp{};
+    const auto cde_element_op = CDEElementOp{};
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGemmMultipleD<
+        ALayout,
+        BLayout,
+        ck::Tuple<D0Layout>,
+        ELayout,
+        ADataType,
+        BDataType,
+        ck::Tuple<D0DataType>,
+        EDataType,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::AddRelu>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    // run reference
+    if(do_verification)
+    {
+        Tensor<AccDataType> c_m_n({M, N});
+
+        using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                                BDataType,
+                                                                                AccDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                BElementOp,
+                                                                                PassThrough>;
+
+        auto ref_gemm    = ReferenceGemmInstance{};
+        auto ref_invoker = ref_gemm.MakeInvoker();
+
+        auto ref_argument =
+            ref_gemm.MakeArgument(a_m_k, b_k_n, c_m_n, a_element_op, b_element_op, PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        for(int m = 0; m < M; ++m)
+        {
+            for(int n = 0; n < N; ++n)
+            {
+                cde_element_op(e_m_n_host_result(m, n), c_m_n(m, n), d0_m_n(m, n));
+            }
+        }
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem d0_m_n_device_buf(sizeof(D0DataType) * d0_m_n.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a_m_k.mData.data());
+    b_device_buf.ToDevice(b_k_n.mData.data());
+    d0_m_n_device_buf.ToDevice(d0_m_n.mData.data());
+
+    std::string best_op_name;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    bool pass = true;
+
+    // profile device operation instances
+    for(auto& op_ptr : op_ptrs)
+    {
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            a_device_buf.GetDeviceBuffer(),
+            b_device_buf.GetDeviceBuffer(),
+            std::array<const void*, 1>{d0_m_n_device_buf.GetDeviceBuffer()},
+            e_device_buf.GetDeviceBuffer(),
+            M,
+            N,
+            K,
+            StrideA,
+            StrideB,
+            std::array<ck::index_t, 1>{StrideD0},
+            StrideE,
+            a_element_op,
+            b_element_op,
+            cde_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            // re-init E to zero before profiling a kernel
+            e_device_buf.SetZero();
+
+            float ave_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+            std::size_t flop = std::size_t(2) * M * N * K;
+
+            std::size_t num_btype =
+                sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(EDataType) * M * N;
+
+            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+
+            if(do_verification)
+            {
+                e_device_buf.FromDevice(e_m_n_device_result.mData.data());
+
+                pass = pass && ck::utils::check_err(e_m_n_device_result, e_m_n_host_result);
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
+              << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    return pass;
+}
+
+} // namespace profiler
+} // namespace ck
--- a/profiler/include/profiler/profile_gemm_add_silu_impl.hpp
+++ b/profiler/include/profiler/profile_gemm_add_silu_impl.hpp
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iomanip>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/gemm_add_silu.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+
+namespace ck {
+namespace profiler {
+
+template <typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename D0DataType,
+          typename EDataType,
+          typename ALayout,
+          typename BLayout,
+          typename D0Layout,
+          typename ELayout>
+bool profile_gemm_add_silu_impl(int do_verification,
+                                int init_method,
+                                bool /*do_log*/,
+                                bool time_kernel,
+                                int M,
+                                int N,
+                                int K,
+                                int StrideA,
+                                int StrideB,
+                                int StrideD0,
+                                int StrideE)
+{
+    auto f_host_tensor_descriptor =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            using namespace ck::literals;
+
+            if(is_same<decltype(layout), tensor_layout::gemm::RowMajor>::value)
+            {
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
+            }
+            else
+            {
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
+            }
+        };
+
+    Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
+    Tensor<D0DataType> d0_m_n(f_host_tensor_descriptor(M, N, StrideD0, D0Layout{}));
+    Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+    Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+
+    std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
+    std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
+    std::cout << "d0_m_n: " << d0_m_n.mDesc << std::endl;
+    std::cout << "e_m_n: " << e_m_n_device_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        d0_m_n.GenerateTensorValue(GeneratorTensor_2<D0DataType>{-5, 5});
+        break;
+    default:
+        a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        d0_m_n.GenerateTensorValue(GeneratorTensor_3<D0DataType>{0.0, 1.0});
+    }
+
+    using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+    using AddRelu     = ck::tensor_operation::element_wise::AddSilu;
+
+    using AElementOp   = PassThrough;
+    using BElementOp   = PassThrough;
+    using CDEElementOp = AddRelu;
+
+    const auto a_element_op   = AElementOp{};
+    const auto b_element_op   = BElementOp{};
+    const auto cde_element_op = CDEElementOp{};
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGemmMultipleD<
+        ALayout,
+        BLayout,
+        ck::Tuple<D0Layout>,
+        ELayout,
+        ADataType,
+        BDataType,
+        ck::Tuple<D0DataType>,
+        EDataType,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::AddSilu>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    // run reference
+    if(do_verification)
+    {
+        Tensor<AccDataType> c_m_n({M, N});
+
+        using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                                BDataType,
+                                                                                AccDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                BElementOp,
+                                                                                PassThrough>;
+
+        auto ref_gemm    = ReferenceGemmInstance{};
+        auto ref_invoker = ref_gemm.MakeInvoker();
+
+        auto ref_argument =
+            ref_gemm.MakeArgument(a_m_k, b_k_n, c_m_n, a_element_op, b_element_op, PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        for(int m = 0; m < M; ++m)
+        {
+            for(int n = 0; n < N; ++n)
+            {
+                cde_element_op(e_m_n_host_result(m, n), c_m_n(m, n), d0_m_n(m, n));
+            }
+        }
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem d0_m_n_device_buf(sizeof(D0DataType) * d0_m_n.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a_m_k.mData.data());
+    b_device_buf.ToDevice(b_k_n.mData.data());
+    d0_m_n_device_buf.ToDevice(d0_m_n.mData.data());
+
+    std::string best_op_name;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    bool pass = true;
+
+    // profile device operation instances
+    for(auto& op_ptr : op_ptrs)
+    {
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            a_device_buf.GetDeviceBuffer(),
+            b_device_buf.GetDeviceBuffer(),
+            std::array<const void*, 1>{d0_m_n_device_buf.GetDeviceBuffer()},
+            e_device_buf.GetDeviceBuffer(),
+            M,
+            N,
+            K,
+            StrideA,
+            StrideB,
+            std::array<ck::index_t, 1>{StrideD0},
+            StrideE,
+            a_element_op,
+            b_element_op,
+            cde_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            // re-init E to zero before profiling a kernel
+            e_device_buf.SetZero();
+
+            float ave_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+            std::size_t flop = std::size_t(2) * M * N * K;
+
+            std::size_t num_btype =
+                sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(EDataType) * M * N;
+
+            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+
+            if(do_verification)
+            {
+                e_device_buf.FromDevice(e_m_n_device_result.mData.data());
+
+                pass = pass && ck::utils::check_err(e_m_n_device_result, e_m_n_host_result);
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
+              << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    return pass;
+}
+
+} // namespace profiler
+} // namespace ck
--- a/profiler/include/profiler/profile_grouped_gemm_fixed_nk_impl.hpp
+++ b/profiler/include/profiler/profile_grouped_gemm_fixed_nk_impl.hpp
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iomanip>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_gemm_fixed_nk.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/grouped_gemm_fixed_nk.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/convolution_parameter.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/utility/fill.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+
+namespace ck {
+namespace profiler {
+
+template <typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename AccDataType,
+          typename ALayout,
+          typename BLayout,
+          typename CLayout>
+bool profile_grouped_gemm_fixed_nk_impl(int do_verification,
+                                        int init_method,
+                                        bool do_log,
+                                        bool time_kernel,
+                                        const std::vector<int>& Ms,
+                                        const std::vector<int>& Ns,
+                                        const std::vector<int>& Ks,
+                                        const std::vector<int>& StrideAs,
+                                        const std::vector<int>& StrideBs,
+                                        const std::vector<int>& StrideCs,
+                                        int kbatch   = 1,
+                                        int n_warmup = 1,
+                                        int n_iter   = 10)
+{
+    bool pass = true;
+
+    auto f_host_tensor_descriptor =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            using namespace ck::literals;
+
+            if(is_same<decltype(layout), tensor_layout::gemm::RowMajor>::value)
+            {
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
+            }
+            else
+            {
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
+            }
+        };
+
+    std::size_t group_count = Ms.size();
+
+    if(!(group_count == Ns.size() && group_count == Ks.size() && group_count == StrideAs.size() &&
+         group_count == StrideBs.size() && group_count == StrideCs.size()))
+    {
+        throw std::runtime_error("wrong! inconsistent M/N/Ks, StrideA/B/Cs size\n");
+    }
+
+    std::vector<Tensor<ADataType>> a_m_k;
+    std::vector<Tensor<BDataType>> b_k_n;
+    std::vector<Tensor<CDataType>> c_m_n_host_results;
+    std::vector<Tensor<CDataType>> c_m_n_device_results;
+
+    for(std::size_t i = 0; i < group_count; i++)
+    {
+        a_m_k.push_back(
+            Tensor<ADataType>(f_host_tensor_descriptor(Ms[i], Ks[i], StrideAs[i], ALayout{})));
+        b_k_n.push_back(
+            Tensor<BDataType>(f_host_tensor_descriptor(Ks[i], Ns[i], StrideBs[i], BLayout{})));
+
+        c_m_n_device_results.push_back(
+            Tensor<CDataType>(f_host_tensor_descriptor(Ms[i], Ns[i], StrideCs[i], CLayout{})));
+
+        c_m_n_host_results.push_back(
+            Tensor<CDataType>(f_host_tensor_descriptor(Ms[i], Ns[i], StrideCs[i], CLayout{})));
+#if DEBUG_LOG
+        std::cout << "group: " << i << " a_m_k[" << i << "]:" << a_m_k[i].mDesc << ", b_k_n[" << i
+                  << "]:" << b_k_n[i].mDesc << ", c_m_n_device_results[" << i
+                  << "]:" << c_m_n_device_results[i].mDesc << std::endl;
+#endif // DEBUG_LOG
+        std::size_t num_thread = 1;
+        switch(init_method)
+        {
+        case 0: break;
+        case 1:
+            a_m_k[i].GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5}, num_thread);
+            b_k_n[i].GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5}, num_thread);
+            break;
+        default:
+            a_m_k[i].GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0}, num_thread);
+            b_k_n[i].GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5}, num_thread);
+        }
+    }
+
+    using AElementOp = ck::tensor_operation::element_wise::PassThrough;
+    using BElementOp = ck::tensor_operation::element_wise::PassThrough;
+    using CElementOp = ck::tensor_operation::element_wise::PassThrough;
+
+    const auto a_element_op = AElementOp{};
+    const auto b_element_op = BElementOp{};
+    const auto c_element_op = CElementOp{};
+
+    using DeviceMemPtr = std::unique_ptr<DeviceMem>;
+    std::vector<DeviceMemPtr> a_device_buf, b_device_buf, c_device_buf;
+
+    a_device_buf.reserve(group_count);
+    b_device_buf.reserve(group_count);
+    c_device_buf.reserve(group_count);
+
+    std::vector<const void*> p_a, p_b;
+    std::vector<void*> p_c;
+
+    p_a.reserve(group_count);
+    p_b.reserve(group_count);
+    p_c.reserve(group_count);
+
+    std::vector<ck::tensor_operation::device::GemmDesc> gemm_descs;
+    gemm_descs.reserve(group_count);
+
+    std::vector<ck::tensor_operation::device::GroupedGemmKernelArgument<1>>
+        grouped_gemm_kernel_args_;
+    grouped_gemm_kernel_args_.reserve(group_count);
+
+    for(std::size_t i = 0; i < group_count; i++)
+    {
+        a_device_buf.emplace_back(
+            std::make_unique<DeviceMem>(sizeof(ADataType) * a_m_k[i].mDesc.GetElementSpaceSize()));
+        b_device_buf.emplace_back(
+            std::make_unique<DeviceMem>(sizeof(BDataType) * b_k_n[i].mDesc.GetElementSpaceSize()));
+        c_device_buf.emplace_back(std::make_unique<DeviceMem>(
+            sizeof(CDataType) * c_m_n_device_results[i].mDesc.GetElementSpaceSize()));
+
+        a_device_buf[i]->ToDevice(a_m_k[i].mData.data());
+        b_device_buf[i]->ToDevice(b_k_n[i].mData.data());
+
+        gemm_descs.push_back({Ms[i], Ns[i], Ks[i], StrideAs[i], StrideBs[i], StrideCs[i], {}});
+
+        p_a.push_back(a_device_buf[i]->GetDeviceBuffer());
+        p_b.push_back(b_device_buf[i]->GetDeviceBuffer());
+        p_c.push_back(c_device_buf[i]->GetDeviceBuffer());
+
+        grouped_gemm_kernel_args_.push_back({a_device_buf[i]->GetDeviceBuffer(),
+                                             b_device_buf[i]->GetDeviceBuffer(),
+                                             {},
+                                             c_device_buf[i]->GetDeviceBuffer(),
+                                             Ms[i],
+                                             Ns[i],
+                                             Ks[i],
+                                             StrideAs[i],
+                                             StrideBs[i],
+                                             {},
+                                             StrideCs[i]});
+    }
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGroupedGemmFixedNK<ALayout,
+                                                                            BLayout,
+                                                                            ck::Tuple<>,
+                                                                            CLayout,
+                                                                            ADataType,
+                                                                            BDataType,
+                                                                            ck::Tuple<>,
+                                                                            CDataType,
+                                                                            AElementOp,
+                                                                            BElementOp,
+                                                                            CElementOp>;
+
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    if(op_ptrs.size() <= 0)
+    {
+        throw std::runtime_error("wrong! no device GEMM instance found");
+    }
+
+    std::string best_gemm_name;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+    float best_kbatch     = 0;
+
+    auto p_ds = std::vector<std::array<const void*, 0>>{};
+
+    if(do_verification)
+    {
+        for(std::size_t i = 0; i < gemm_descs.size(); i++)
+        {
+            using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                                    BDataType,
+                                                                                    CDataType,
+                                                                                    AccDataType,
+                                                                                    AElementOp,
+                                                                                    BElementOp,
+                                                                                    CElementOp>;
+
+            auto ref_gemm    = ReferenceGemmInstance{};
+            auto ref_invoker = ref_gemm.MakeInvoker();
+
+            auto ref_argument = ref_gemm.MakeArgument(a_m_k[i],
+                                                      b_k_n[i],
+                                                      c_m_n_host_results[i],
+                                                      a_element_op,
+                                                      b_element_op,
+                                                      c_element_op);
+
+            ref_invoker.Run(ref_argument);
+        }
+    }
+
+    // profile device GEMM instances
+    for(auto& gemm_ptr : op_ptrs)
+    {
+        auto argument_ptr =
+            gemm_ptr->MakeArgumentPointer(p_a,
+                                          p_b,
+                                          p_ds,
+                                          p_c,
+                                          gemm_descs,
+                                          ck::tensor_operation::element_wise::PassThrough{},
+                                          ck::tensor_operation::element_wise::PassThrough{},
+                                          ck::tensor_operation::element_wise::PassThrough{});
+
+        auto invoker_ptr = gemm_ptr->MakeInvokerPointer();
+
+        DeviceMem gemm_desc_workspace(gemm_ptr->GetWorkSpaceSize(argument_ptr.get()));
+
+        DeviceMem grouped_gemm_kernel_args_dev(
+            gemm_ptr->GetDeviceKernelArgSize(argument_ptr.get()));
+
+        hipGetErrorString(hipMemcpy(grouped_gemm_kernel_args_dev.GetDeviceBuffer(),
+                                    grouped_gemm_kernel_args_.data(),
+                                    gemm_ptr->GetDeviceKernelArgSize(argument_ptr.get()),
+                                    hipMemcpyHostToDevice));
+
+        gemm_ptr->SetWorkSpacePointer(argument_ptr.get(), gemm_desc_workspace.GetDeviceBuffer());
+
+        gemm_ptr->SetDeviceKernelArgs(argument_ptr.get(),
+                                      grouped_gemm_kernel_args_dev.GetDeviceBuffer());
+
+        std::string gemm_name = gemm_ptr->GetTypeString();
+
+        std::vector<int> kbatch_list = {1, 2, 4, 8, 12, 16, 20, 24, 32, 48, 64};
+
+        if(kbatch > 0)
+        {
+            kbatch_list = {kbatch};
+        }
+
+        for(std::size_t j = 0; j < kbatch_list.size(); j++)
+        {
+
+            auto kbatch_curr = kbatch_list[j];
+
+            gemm_ptr->SetKBatch(argument_ptr.get(), kbatch_curr);
+
+            if(gemm_ptr->IsSupportedArgument(argument_ptr.get()))
+            {
+                for(std::size_t i = 0; i < gemm_descs.size(); i++)
+                    c_device_buf[i]->SetZero();
+
+                invoker_ptr->Run(argument_ptr.get(),
+                                 StreamConfig{nullptr, false, 0, n_warmup, n_iter});
+
+                if(do_verification)
+                {
+                    bool instance_pass = true;
+                    for(std::size_t i = 0; i < gemm_descs.size(); i++)
+                    {
+
+                        c_device_buf[i]->FromDevice(c_m_n_device_results[i].mData.data());
+
+                        if(std::is_same_v<CDataType, ck::half_t> && kbatch_curr > 1)
+                        {
+                            instance_pass =
+                                instance_pass && ck::utils::check_err(c_m_n_device_results[i],
+                                                                      c_m_n_host_results[i],
+                                                                      "Error: Incorrect results!",
+                                                                      0.06);
+                        }
+                        else
+                        {
+                            instance_pass =
+                                instance_pass && ck::utils::check_err(c_m_n_device_results[i],
+                                                                      c_m_n_host_results[i]);
+                        }
+
+                        if(do_log)
+                        {
+                            LogRangeAsType<float>(std::cout << "a : ", a_m_k[i].mData, ",")
+                                << std::endl;
+                            LogRangeAsType<float>(std::cout << "b: ", b_k_n[i].mData, ",")
+                                << std::endl;
+                            LogRangeAsType<float>(
+                                std::cout << "c_device: ", c_m_n_device_results[i].mData, ",")
+                                << std::endl;
+                            LogRangeAsType<float>(
+                                std::cout << "c_host  : ", c_m_n_host_results[i].mData, ",")
+                                << std::endl;
+                        }
+                    }
+
+                    std::cout << "Instance: " << gemm_name << " verification "
+                              << (instance_pass ? "SUCCEED" : "FAILED") << std::endl;
+
+                    pass = pass && instance_pass;
+                }
+
+                float ave_time = invoker_ptr->Run(
+                    argument_ptr.get(), StreamConfig{nullptr, time_kernel, 0, n_warmup, n_iter});
+
+                if(time_kernel)
+                {
+                    std::size_t flop = 0, num_btype = 0;
+                    for(std::size_t i = 0; i < gemm_descs.size(); i++)
+                    {
+                        flop += std::size_t(2) * Ms[i] * Ns[i] * Ks[i];
+
+                        num_btype += sizeof(ADataType) * Ms[i] * Ks[i] +
+                                     sizeof(BDataType) * Ks[i] * Ns[i] +
+                                     sizeof(CDataType) * Ms[i] * Ns[i];
+                    }
+
+                    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+                    float gb_per_sec = num_btype / 1.E6 / ave_time;
+                    std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops
+                              << " TFlops, " << gb_per_sec << " GB/s, " << gemm_name << ", KBatch "
+                              << kbatch_curr << std::endl;
+
+                    if(tflops > best_tflops)
+                    {
+                        best_gemm_name  = gemm_name;
+                        best_tflops     = tflops;
+                        best_ave_time   = ave_time;
+                        best_gb_per_sec = gb_per_sec;
+                        best_kbatch     = kbatch_curr;
+                    }
+                }
+            }
+            else
+            {
+                std::cout << "Instance: " << gemm_name << ", does not support this GEMM problem"
+                          << std::endl;
+            }
+        }
+    }
+
+    if(time_kernel)
+    {
+        std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
+                  << best_gb_per_sec << " GB/s, " << best_gemm_name << ", KBatch = " << best_kbatch
+                  << std::endl;
+    }
+    return pass;
+}
+
+} // namespace profiler
+} // namespace ck
--- a/test/permute_scale/test_permute_scale_impl.hpp
+++ b/test/permute_scale/test_permute_scale_impl.hpp
-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
-
-#pragma once
-
-#include <iomanip>
-#include <random>
-
-#include "ck/ck.hpp"
-#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
-#include "ck/tensor_operation/gpu/device/device_elementwise_scale.hpp"
-#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
-#include "ck/tensor_operation/gpu/device/impl/device_elementwise_scale_impl.hpp"
-
-#include "ck/library/tensor_operation_instance/gpu/permute_scale.hpp"
-
-#include "ck/library/utility/check_err.hpp"
-#include "ck/library/utility/device_memory.hpp"
-#include "ck/library/utility/host_tensor.hpp"
-#include "ck/library/utility/host_tensor_generator.hpp"
-#include "ck/library/utility/literals.hpp"
-
-namespace ck {
-template <typename HostTensorA, typename HostTensorB, typename FunctorA, typename FunctorB>
-void host_elementwise4D(HostTensorB& B_nhwc,
-                        const HostTensorA& A_nchw,
-                        FunctorA functor_a,
-                        FunctorB functor_b,
-                        float scale)
-{
-    std::size_t N = A_nchw.mDesc.GetLengths()[0];
-    std::size_t C = A_nchw.mDesc.GetLengths()[1];
-    std::size_t H = A_nchw.mDesc.GetLengths()[2];
-    std::size_t W = A_nchw.mDesc.GetLengths()[3];
-    for(std::size_t w = 0; w < W; ++w)
-        for(std::size_t h = 0; h < H; ++h)
-            for(std::size_t c = 0; c < C; ++c)
-                for(std::size_t n = 0; n < N; ++n)
-                {
-                    using tmp_type   = ck::remove_reference_t<decltype(B_nhwc(0, 0))>;
-                    tmp_type tmp_val = 0;
-                    auto a_val       = A_nchw.mData[(n) + (c * N) + (h * C * N) + (w * H * C * N)];
-                    functor_b(tmp_val, a_val);
-                    functor_a(B_nhwc.mData[(n) + (c * W * H * N) + (h * N) + (w * H * N)],
-                              scale * tmp_val);
-                }
-}
-
-template <typename ADataType, typename BDataType, index_t NumDim>
-bool test_permute_scale_impl(int do_verification,
-                             int init_method,
-                             bool do_log,
-                             bool time_kernel,
-                             std::vector<index_t> lengths)
-{
-    bool pass = true;
-
-    using ElementOp = ck::tensor_operation::element_wise::PassThrough;
-    using UnaryOp   = ck::tensor_operation::element_wise::UnarySquare;
-    using Scale     = ck::tensor_operation::element_wise::Scale;
-    float scale     = 2.f;
-
-    index_t N = lengths[0];
-    index_t C = lengths[1];
-    index_t H = lengths[2];
-    index_t W = lengths[3];
-
-    std::vector<ck::index_t> nchw = {N, C, H, W};
-    std::vector<ck::index_t> nhwc = {N, H, W, C};
-    Tensor<ADataType> a(nchw);
-    Tensor<BDataType> b(nhwc);
-    Tensor<BDataType> host_b(nhwc);
-
-    std::array<ck::index_t, 4> ab_lengths;
-
-    std::array<ck::index_t, 4> a_strides = {1,
-                                            static_cast<int>(nchw[0]),
-                                            static_cast<int>(nchw[0] * nchw[1]),
-                                            static_cast<int>(nchw[0] * nchw[1] * nchw[2])};
-
-    std::array<ck::index_t, 4> b_strides = {1,
-                                            static_cast<int>(nhwc[0] * nhwc[1] * nhwc[2]),
-                                            static_cast<int>(nhwc[0]),
-                                            static_cast<int>(nhwc[0] * nhwc[1])};
-    ck::ranges::copy(nchw, ab_lengths.begin());
-
-    std::cout << "A: " << a.mDesc << std::endl;
-    std::cout << "B: " << b.mDesc << std::endl;
-
-    switch(init_method)
-    {
-    case 0: break;
-    case 1: a.GenerateTensorValue(GeneratorTensor_2<ADataType>{-1, 2}); break;
-    default: // a.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0}
-        std::mt19937 gen(11939);
-        std::uniform_int_distribution<int> dis(0, 1);
-        auto i = 0;
-        for(std::size_t w = 0; w < a.mDesc.GetLengths()[3]; ++w)
-            for(std::size_t h = 0; h < a.mDesc.GetLengths()[2]; ++h)
-                for(std::size_t c = 0; c < a.mDesc.GetLengths()[1]; ++c)
-                    for(std::size_t n = 0; n < a.mDesc.GetLengths()[0]; ++n)
-                    {
-                        a.mData[(n * nchw[1] * nchw[2] * nchw[3]) + (c * nchw[2] * nchw[3]) +
-                                (h * nchw[3]) + w] = i;
-                        i                          = dis(gen);
-                    }
-    }
-
-    DeviceMem a_device_buf(sizeof(ADataType) * a.mDesc.GetElementSpaceSize());
-    DeviceMem b_device_buf(sizeof(BDataType) * b.mDesc.GetElementSpaceSize());
-
-    a_device_buf.ToDevice(a.mData.data());
-
-    std::array<const void*, 1> input = {a_device_buf.GetDeviceBuffer()};
-    std::array<void*, 1> output      = {b_device_buf.GetDeviceBuffer()};
-    using DeviceOp = ck::tensor_operation::device::DeviceElementwise<ck::Tuple<ADataType>,
-                                                                     ck::Tuple<BDataType>,
-                                                                     ElementOp,
-                                                                     UnaryOp,
-                                                                     Scale,
-                                                                     NumDim>;
-
-    // get device op instances
-    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
-        DeviceOp>::GetInstances();
-
-    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
-
-    std::string best_instance_name;
-    float best_ave_time   = std::numeric_limits<float>::max();
-    float best_gb_per_sec = 0;
-    float best_tflops     = 0;
-
-    if(do_verification)
-    {
-        host_elementwise4D(host_b, a, ElementOp{}, UnaryOp{}, scale);
-    }
-
-    for(auto& op_ptr : op_ptrs)
-    {
-        auto argument_ptr = op_ptr->MakeArgumentPointer(ab_lengths,
-                                                        {a_strides},
-                                                        {b_strides},
-                                                        input,
-                                                        output,
-                                                        ElementOp{},
-                                                        UnaryOp{},
-                                                        Scale{scale});
-
-        auto invoker_ptr = op_ptr->MakeInvokerPointer();
-
-        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
-        {
-            b_device_buf.SetZero();
-
-            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
-
-            if(do_verification)
-            {
-                b_device_buf.FromDevice(b.mData.data());
-
-                pass &= ck::utils::check_err(
-                    b.mData, host_b.mData, "Error: Incorrect results b", 1e-3, 1e-3);
-
-                if(do_log)
-                {
-                    LogRangeAsType<float>(std::cout << "a : ", a.mData, ",") << std::endl;
-                    LogRangeAsType<float>(std::cout << "b: ", b.mData, ",") << std::endl;
-                }
-            }
-
-            std::string op_name = op_ptr->GetTypeString();
-
-            float ave_time =
-                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
-
-            std::size_t flop = std::size_t(2) * nchw[0] * nchw[1] * nchw[2] * nchw[3];
-
-            std::size_t num_btype = sizeof(ADataType) * (nchw[0] * nchw[1] * nchw[2] * nchw[3]) +
-                                    sizeof(BDataType) * (nchw[0] * nchw[1] * nchw[2] * nchw[3]);
-
-            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
-
-            float gb_per_sec = num_btype / 1.E6 / ave_time;
-
-            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
-                      << gb_per_sec << " GB/s, " << op_name << std::endl;
-
-            if(tflops > best_tflops)
-            {
-                best_instance_name = op_name;
-                best_tflops        = tflops;
-                best_ave_time      = ave_time;
-                best_gb_per_sec    = gb_per_sec;
-            }
-        }
-        else
-        {
-            std::cout << op_ptr->GetTypeString() << " does not support this problem" << std::endl;
-        }
-    }
-    if(time_kernel)
-    {
-        LogRange(std::cout << "length = ", lengths, ",") << ", ";
-        std::cout << "best perf = " << best_ave_time << " ms, " << best_gb_per_sec << " GB/s, "
-                  << best_instance_name << std::endl;
-    }
-
-    return true;
-}
-
-} // namespace ck
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iomanip>
+#include <random>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_elementwise_scale.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_elementwise_scale_impl.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/permute_scale.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+
+namespace ck {
+template <typename HostTensorA,
+          typename HostTensorB,
+          typename AElementOp,
+          typename BElementOp,
+          typename ScaleElementOp>
+void reference_permute_scale(HostTensorB& b_tensor,
+                             const HostTensorA& a_tensor,
+                             AElementOp a_tensor_op,
+                             BElementOp b_tensor_op,
+                             ScaleElementOp scale_op)
+{
+    b_tensor.ForEach([&](auto& self, auto idx) {
+        auto tmp_val = a_tensor(idx);
+        b_tensor_op(tmp_val, tmp_val);
+        scale_op(tmp_val, tmp_val);
+        a_tensor_op(self(idx), tmp_val);
+    });
+}
+
+namespace profiler {
+
+template <typename ADataType, typename BDataType, index_t NumDim>
+bool profile_permute_scale_impl(int do_verification,
+                                int init_method,
+                                bool do_log,
+                                bool time_kernel,
+                                std::vector<index_t> lengths_vector,
+                                std::vector<index_t> input_strides_vector,
+                                std::vector<index_t> output_strides_vector)
+{
+    bool pass           = true;
+    bool instance_found = false;
+
+    using ElementOp = ck::tensor_operation::element_wise::PassThrough;
+    using UnaryOp   = ck::tensor_operation::element_wise::UnarySquare;
+    using Scale     = ck::tensor_operation::element_wise::Scale;
+    float scale     = 2.f;
+
+    Tensor<ADataType> a(lengths_vector, input_strides_vector);
+    Tensor<BDataType> b(lengths_vector, output_strides_vector);
+    Tensor<BDataType> host_b(lengths_vector, output_strides_vector);
+
+    std::cout << "A: " << a.mDesc << std::endl;
+    std::cout << "B: " << b.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1: a.GenerateTensorValue(GeneratorTensor_2<ADataType>{-1, 2}); break;
+    default: a.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0}); break;
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a.mData.data());
+
+    std::array<const void*, 1> input = {a_device_buf.GetDeviceBuffer()};
+    std::array<void*, 1> output      = {b_device_buf.GetDeviceBuffer()};
+    using DeviceOp = ck::tensor_operation::device::DeviceElementwise<ck::Tuple<ADataType>,
+                                                                     ck::Tuple<BDataType>,
+                                                                     ElementOp,
+                                                                     UnaryOp,
+                                                                     Scale,
+                                                                     NumDim>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    std::string best_instance_name;
+    float best_ave_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+    float best_tflops     = 0;
+
+    if(do_verification)
+    {
+        reference_permute_scale(host_b, a, ElementOp{}, UnaryOp{}, Scale{scale});
+    }
+
+    auto copy = [](const auto& x, auto& y) { std::copy(x.begin(), x.end(), y.begin()); };
+    std::array<ck::index_t, NumDim> lengths{};
+    std::array<ck::index_t, NumDim> input_strides{};
+    std::array<ck::index_t, NumDim> output_strides{};
+    copy(lengths_vector, lengths);
+    copy(input_strides_vector, input_strides);
+    copy(output_strides_vector, output_strides);
+
+    for(auto& op_ptr : op_ptrs)
+    {
+        auto argument_ptr = op_ptr->MakeArgumentPointer(lengths,
+                                                        {input_strides},
+                                                        {output_strides},
+                                                        input,
+                                                        output,
+                                                        ElementOp{},
+                                                        UnaryOp{},
+                                                        Scale{scale});
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            instance_found = true;
+
+            b_device_buf.SetZero();
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+
+            if(do_verification)
+            {
+                b_device_buf.FromDevice(b.mData.data());
+
+                pass &= ck::utils::check_err(
+                    b.mData, host_b.mData, "Error: Incorrect results b", 1e-3, 1e-3);
+
+                if(do_log)
+                {
+                    LogRangeAsType<float>(std::cout << "a : ", a.mData, ",") << std::endl;
+                    LogRangeAsType<float>(std::cout << "b: ", b.mData, ",") << std::endl;
+                }
+            }
+
+            std::string op_name = op_ptr->GetTypeString();
+
+            float ave_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+            std::size_t flop = std::size_t(2) * a.mDesc.GetElementSpaceSize() / sizeof(ADataType);
+
+            std::size_t num_btype = sizeof(ADataType) * a.mDesc.GetElementSpaceSize() +
+                                    sizeof(BDataType) * b.mDesc.GetElementSpaceSize();
+
+            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_instance_name = op_name;
+                best_tflops        = tflops;
+                best_ave_time      = ave_time;
+                best_gb_per_sec    = gb_per_sec;
+            }
+        }
+        else
+        {
+            std::cout << op_ptr->GetTypeString() << " does not support this problem" << std::endl;
+        }
+    }
+    if(time_kernel)
+    {
+        std::cout << "Best perf = " << best_ave_time << " ms, " << best_gb_per_sec << " GB/s, "
+                  << best_instance_name << std::endl;
+    }
+
+    return pass && instance_found;
+}
+
+} // namespace profiler
+} // namespace ck
--- a/profiler/src/CMakeLists.txt
+++ b/profiler/src/CMakeLists.txt
@@ -32,6 +32,7 @@ set(PROFILER_SOURCES
    profile_grouped_conv_bwd_data.cpp
    profile_conv_tensor_rearrange.cpp
    profile_transpose.cpp
+    profile_permute_scale.cpp
 )

 if(DL_KERNELS)
@@ -43,11 +44,15 @@ if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
  list(APPEND PROFILER_SOURCES profile_gemm_fastgelu.cpp)
  list(APPEND PROFILER_SOURCES profile_gemm_streamk.cpp)
  list(APPEND PROFILER_SOURCES profile_gemm_bilinear.cpp)
+  list(APPEND PROFILER_SOURCES profile_gemm_add.cpp) 
  list(APPEND PROFILER_SOURCES profile_gemm_add_fastgelu.cpp)
+  list(APPEND PROFILER_SOURCES profile_gemm_add_relu.cpp)
+  list(APPEND PROFILER_SOURCES profile_gemm_add_silu.cpp)
  list(APPEND PROFILER_SOURCES profile_gemm_add_add_fastgelu.cpp)
  list(APPEND PROFILER_SOURCES profile_gemm_add_relu_add_layernorm.cpp)
  list(APPEND PROFILER_SOURCES profile_batched_gemm_add_relu_gemm_add.cpp)
  list(APPEND PROFILER_SOURCES profile_grouped_gemm.cpp)
+  list(APPEND PROFILER_SOURCES profile_grouped_gemm_fixed_nk.cpp)
  list(APPEND PROFILER_SOURCES profile_grouped_gemm_fastgelu.cpp)
  list(APPEND PROFILER_SOURCES profile_grouped_gemm_multiple_d_splitk.cpp)
 endif()
@@ -97,6 +102,7 @@ target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_bwd_d
 target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_image_to_column_instance)
 target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_column_to_image_instance)
 target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_transpose_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_permute_scale_instance)

 if(DTYPES MATCHES "fp32" OR DTYPES MATCHES "fp64" OR NOT DEFINED DTYPES)
  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_contraction_bilinear_instance)
@@ -110,7 +116,10 @@ if(DL_KERNELS)
 endif()

 if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_instance)
  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_fastgelu_instance)
+  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_relu_instance)
+  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_silu_instance)
  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_relu_add_layernorm_instance)
  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_bilinear_instance)
  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_add_fastgelu_instance)
@@ -119,6 +128,7 @@ if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_gemm_instance)
  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_add_relu_gemm_add_instance)
  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_instance)
+  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_fixed_nk_instance)
  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_fastgelu_instance)
  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_gemm_multiple_d_instance)
 endif()

--- a/profiler/src/profile_gemm_add.cpp
+++ b/profiler/src/profile_gemm_add.cpp
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "profiler/profile_gemm_add_impl.hpp"
+#include "profiler_operation_registry.hpp"
+
+#define OP_NAME "gemm_add"
+#define OP_DESC "GEMM+Add"
+
+using INT8 = int8_t;
+using BF16 = ck::bhalf_t;
+
+int profile_gemm_add(int argc, char* argv[])
+{
+    enum struct MatrixLayout
+    {
+        MK_KN_MN_MN, // 0
+        MK_NK_MN_MN, // 1
+        KM_KN_MN_MN, // 2
+        KM_NK_MN_MN, // 3
+    };
+
+    enum struct MatrixDataType
+    {
+        F16_INT8_F16_F16,    // 0
+        BF16_INT8_BF16_BF16, // 1
+    };
+
+    if(argc != 15)
+    {
+        // clang-format off
+        printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
+        printf("arg2: data type (0: f16&i8 1: bf16&i8)\n");
+        printf("arg3: matrix layout (0: E[m, n] = ReLU(A[m, k] * B[k, n] + D0[m, n]);\n");
+        printf("                     1: E[m, n] = ReLU(A[m, k] * B[n, k] + D0[m, n]);\n");
+        printf("                     2: E[m, n] = ReLU(A[k, m] * B[k, n] + D0[m, n]);\n");
+        printf("                     3: E[m, n] = ReLU(A[k, m] * B[n, k] + D0[m, n]))\n");
+        printf("arg4: verification (0: no; 1: yes)\n");
+        printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n");
+        printf("arg6: print tensor value (0: no; 1: yes)\n");
+        printf("arg7: time kernel (0=no, 1=yes)\n");
+        printf("arg8 to 14: M, N, K, StrideA, StrideB, StrideD0, StrideE\n");
+        // clang-format on
+        exit(1);
+    }
+
+    const auto data_type       = static_cast<MatrixDataType>(std::stoi(argv[2]));
+    const auto layout          = static_cast<MatrixLayout>(std::stoi(argv[3]));
+    const bool do_verification = std::stoi(argv[4]);
+    const int init_method      = std::stoi(argv[5]);
+    const bool do_log          = std::stoi(argv[6]);
+    const bool time_kernel     = std::stoi(argv[7]);
+
+    const int M = std::stoi(argv[8]);
+    const int N = std::stoi(argv[9]);
+    const int K = std::stoi(argv[10]);
+
+    const int StrideA  = std::stoi(argv[11]);
+    const int StrideB  = std::stoi(argv[12]);
+    const int StrideD0 = std::stoi(argv[13]);
+    const int StrideE  = std::stoi(argv[14]);
+
+    using F16 = ck::half_t;
+    using F32 = float;
+
+    using Row = ck::tensor_layout::gemm::RowMajor;
+    // using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+    auto profile = [&](auto a_type,
+                       auto b_type,
+                       auto acc_type,
+                       auto d0_type,
+                       auto e_type,
+                       auto a_layout,
+                       auto b_layout,
+                       auto d0_layout,
+                       auto e_layout) {
+        using ADataType   = decltype(a_type);
+        using BDataType   = decltype(b_type);
+        using AccDataType = decltype(acc_type);
+        using D0DataType  = decltype(d0_type);
+        using EDataType   = decltype(e_type);
+
+        using ALayout  = decltype(a_layout);
+        using BLayout  = decltype(b_layout);
+        using D0Layout = decltype(d0_layout);
+        using ELayout  = decltype(e_layout);
+
+        const int DefaultStrideA  = ck::is_same_v<ALayout, Row> ? K : M;
+        const int DefaultStrideB  = ck::is_same_v<BLayout, Row> ? N : K;
+        const int DefaultStrideD0 = ck::is_same_v<D0Layout, Row> ? N : M;
+        const int DefaultStrideE  = ck::is_same_v<ELayout, Row> ? N : M;
+
+        bool pass = ck::profiler::profile_gemm_add_impl<ADataType,
+                                                        BDataType,
+                                                        AccDataType,
+                                                        D0DataType,
+                                                        EDataType,
+                                                        ALayout,
+                                                        BLayout,
+                                                        D0Layout,
+                                                        ELayout>(
+            do_verification,
+            init_method,
+            do_log,
+            time_kernel,
+            M,
+            N,
+            K,
+            (StrideA < 0) ? DefaultStrideA : StrideA,
+            (StrideB < 0) ? DefaultStrideB : StrideB,
+            (StrideD0 < 0) ? DefaultStrideD0 : StrideD0,
+            (StrideE < 0) ? DefaultStrideE : StrideE);
+
+        return pass ? 0 : 1;
+    };
+
+    if(data_type == MatrixDataType::F16_INT8_F16_F16 && layout == MatrixLayout::MK_KN_MN_MN)
+    {
+        return profile(F16{}, INT8{}, F32{}, F16{}, F16{}, Row{}, Row{}, Row{}, Row{});
+    }
+    else if(data_type == MatrixDataType::BF16_INT8_BF16_BF16 && layout == MatrixLayout::MK_KN_MN_MN)
+    {
+        return profile(BF16{}, INT8{}, F32{}, BF16{}, BF16{}, Row{}, Row{}, Row{}, Row{});
+    }
+    else
+    {
+        std::cout << "this data_type & layout is not implemented" << std::endl;
+
+        return 1;
+    }
+}
+
+REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_gemm_add);
--- a/profiler/src/profile_gemm_add_fastgelu.cpp
+++ b/profiler/src/profile_gemm_add_fastgelu.cpp
 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.

 #include <iostream>
 #include <numeric>
@@ -12,6 +12,9 @@
 #define OP_NAME "gemm_add_fastgelu"
 #define OP_DESC "GEMM+Add+FastGeLU"

+using INT8 = int8_t;
+using BF16 = ck::bhalf_t;
+
 int profile_gemm_add_fastgelu(int argc, char* argv[])
 {
    enum struct MatrixLayout
@@ -28,13 +31,15 @@ int profile_gemm_add_fastgelu(int argc, char* argv[])
        F16_F16_F16_F16,     // 1
        BF16_BF16_BF16_BF16, // 2
        INT8_INT8_INT8_INT8, // 3
+        F16_INT8_F16_F16,    // 4
+        BF16_INT8_BF16_BF16, // 5
    };

    if(argc != 15)
    {
        // clang-format off
        printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
-        printf("arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8)\n");
+        printf("arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8; 4: f16&i8 5: bf16&i8)\n");
        printf("arg3: matrix layout (0: E[m, n] = FastGeLU(A[m, k] * B[k, n] + D0[m, n]);\n");
        printf("                     1: E[m, n] = FastGeLU(A[m, k] * B[n, k] + D0[m, n]);\n");
        printf("                     2: E[m, n] = FastGeLU(A[k, m] * B[k, n] + D0[m, n]);\n");
@@ -135,6 +140,14 @@ int profile_gemm_add_fastgelu(int argc, char* argv[])
    {
        return profile(F16{}, F16{}, F32{}, F16{}, F16{}, Col{}, Col{}, Row{}, Row{});
    }
+    else if(data_type == MatrixDataType::F16_INT8_F16_F16 && layout == MatrixLayout::MK_KN_MN_MN)
+    {
+        return profile(F16{}, INT8{}, F32{}, F16{}, F16{}, Row{}, Row{}, Row{}, Row{});
+    }
+    else if(data_type == MatrixDataType::BF16_INT8_BF16_BF16 && layout == MatrixLayout::MK_KN_MN_MN)
+    {
+        return profile(BF16{}, INT8{}, F32{}, BF16{}, BF16{}, Row{}, Row{}, Row{}, Row{});
+    }
    else
    {
        std::cout << "this data_type & layout is not implemented" << std::endl;