Merge pull request #248 from ROCm/gfx950_mfma

Gfx950 mfma

library/src/tensor_operation_instance/gpu/gemm_universal_streamk/device_gemm_xdl_universal_streamk_f16_f16_f16/device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn.hpp

@@ -34,12 +34,14 @@ static constexpr auto Interwave = BlockGemmPipelineScheduler::Interwave;
 
 template <GemmSpecialization GemmSpec>
 using device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_comp_instances = std::tuple<
-    // clang-format off
+// clang-format off
         //#########################| ALayout| BLayout| CLayout|AData| BData| CData| AccData| Cshuffle|           A|           B|           C|          GEMM| 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|                         Block-wiseGemm|               Block-wiseGemm|
         //#########################|        |        |        | Type|  Type|  Type|    Type|     Type| Elementwise| Elementwise| Elementwise|Specialization|  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_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|                               Pipeline|                     Pipeline|
         //#########################|        |        |        |     |      |      |        |         |   Operation|   Operation|   Operation|              |      |      |      |      |    |    |    |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl|   _NWaveNPerXdl|                              Scheduler|                     Verision|
         //#########################|        |        |        |     |      |      |        |         |            |            |            |              |      |      |      |      |    |    |    |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |                                       |                             |
-        
+#if defined(CK_USE_AMD_MFMA_GFX950)
+        DeviceGemm_Xdl_CShuffle_Streamk_V3<  Row,     Row,     Row,     F16,   F16,  F16,   F32,     F16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   256,   32,   128,  16,   2,  32,   32,    2,    1,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<32, 8, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              2,          0,          1,           1,                   S<1, 32, 1, 8>,               4,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v2>
+#else        
         DeviceGemm_Xdl_CShuffle_Streamk_V3<  Row,     Row,     Row,     F16,   F16,  F16,   F32,     F16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   256,   256,    32,   8,   4,  32,   32,    4,    4,     S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              8,              4,          0,          1,           1,                   S<1, 32, 1, 8>,               8,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>,
         DeviceGemm_Xdl_CShuffle_Streamk_V3<  Row,     Row,     Row,     F16,   F16,  F16,   F32,     F16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   128,   128,    64,   8,   4,  32,   32,    2,    2,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<16, 16, 1>,    S<0, 2, 1>,     S<0, 2, 1>,             1,              8,              4,          0,          1,           1,                   S<1, 32, 1, 8>,               8,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>,
         DeviceGemm_Xdl_CShuffle_Streamk_V3<  Row,     Row,     Row,     F16,   F16,  F16,   F32,     F16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   256,   256,    32,   8,   4,  32,   32,    4,    4,     S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              8,              4,          0,          1,           1,                   S<1, 32, 1, 8>,               8,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v5>,
@@ -50,17 +52,19 @@ using device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_comp_instances = st
         DeviceGemm_Xdl_CShuffle_Streamk_V3<  Row,     Row,     Row,     F16,   F16,  F16,   F32,     F16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   128,   256,    32,   8,   4,  32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              8,              4,          0,          1,           1,                   S<1, 32, 1, 8>,               8,  BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1>,
         DeviceGemm_Xdl_CShuffle_Streamk_V3<  Row,     Row,     Row,     F16,   F16,  F16,   F32,     F16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   256,   128,    32,   8,   4,  32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,          0,          1,           1,                   S<1, 32, 1, 8>,               8,  BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1>,
         DeviceGemm_Xdl_CShuffle_Streamk_V3<  Row,     Row,     Row,     F16,   F16,  F16,   F32,     F16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   128,   128,    64,   8,   4,  32,   32,    2,    2,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<16, 16, 1>,    S<0, 2, 1>,     S<0, 2, 1>,             1,              8,              4,          0,          1,           1,                   S<1, 32, 1, 8>,               8,  BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1>
+#endif // defined(CK_USE_AMD_MFMA_GFX950)
     // clang-format on
     >;
 
 template <BlockGemmPipelineScheduler BlkGemmPipeSched, GemmSpecialization GemmSpec>
 using device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_mem_instances = std::tuple<
-    // clang-format off
+// clang-format off
         //#########################| ALayout| BLayout| CLayout|AData| BData| CData| AccData| Cshuffle|           A|           B|           C|          GEMM| 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|    Block-wiseGemm|               Block-wiseGemm|
         //#########################|        |        |        | Type|  Type|  Type|    Type|     Type| Elementwise| Elementwise| Elementwise|Specialization|  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_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|          Pipeline|                     Pipeline|
         //#########################|        |        |        |     |      |      |        |         |   Operation|   Operation|   Operation|              |      |      |      |      |    |    |    |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl|   _NWaveNPerXdl|         Scheduler|                     Verision|
         //#########################|        |        |        |     |      |      |        |         |            |            |            |              |      |      |      |      |    |    |    |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |                  |                             |
-
+#if defined(CK_USE_AMD_MFMA_GFX950)
+#else        
         // Latency friendly
         DeviceGemm_Xdl_CShuffle_Streamk_V3<  Row,     Row,     Row,     F16,   F16,  F16,   F32,     F16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   128,    32,   16,    64,   8,   4,  16,   16,    1,    1,     S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<16, 8, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              4,          0,          1,           1,                   S<1, 16, 1, 8>,               2,  BlkGemmPipeSched, BlockGemmPipelineVersion::v1>,
         DeviceGemm_Xdl_CShuffle_Streamk_V3<  Row,     Row,     Row,     F16,   F16,  F16,   F32,     F16,      PassThrough, PassThrough, PassThrough,       GemmSpec,    64,    16,   16,   128,   8,   4,  16,   16,    1,    1,     S<16, 4, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<16, 4, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,          0,          1,           1,                   S<1, 16, 1, 4>,               4,  BlkGemmPipeSched, BlockGemmPipelineVersion::v1>,
@@ -83,6 +87,7 @@ using device_gemm_xdl_universal_streamk_f16_f16_f16_mk_kn_mn_mem_instances = std
         DeviceGemm_Xdl_CShuffle_Streamk_V3<  Row,     Row,     Row,     F16,   F16,  F16,   F32,     F16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   128,    32,  128,    64,   8,   4,  32,   32,    1,    2,     S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              8,              4,          0,          1,           1,                   S<1, 16, 1, 8>,               8,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
         DeviceGemm_Xdl_CShuffle_Streamk_V3<  Row,     Row,     Row,     F16,   F16,  F16,   F32,     F16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   256,    16,  256,    64,   8,   4,  16,   16,    1,    4,     S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              8,              4,          0,          1,           1,                   S<1, 16, 1, 16>,              4,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
         DeviceGemm_Xdl_CShuffle_Streamk_V3<  Row,     Row,     Row,     F16,   F16,  F16,   F32,     F16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   256,    32,  256,    64,   8,   4,  32,   32,    1,    2,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              8,              4,          0,          1,           1,                   S<1, 16, 1, 16>,              8,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>
+#endif // defined(CK_USE_AMD_MFMA_GFX950)
     // clang-format on
     >;
 } // namespace instance

library/src/tensor_operation_instance/gpu/gemm_universal_streamk/device_gemm_xdl_universal_streamk_f16_f16_f16/device_gemm_xdl_universal_streamk_f16_f16_f16_mk_nk_mn.hpp

@@ -34,12 +34,13 @@ static constexpr auto Interwave = BlockGemmPipelineScheduler::Interwave;
 
 template <GemmSpecialization GemmSpec>
 using device_gemm_xdl_universal_streamk_f16_f16_f16_mk_nk_mn_comp_instances = std::tuple<
-    // clang-format off
+// clang-format off
         //#########################| ALayout| BLayout| CLayout|AData| BData| CData| AccData| Cshuffle|           A|           B|           C|          GEMM| 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|                         Block-wiseGemm|               Block-wiseGemm|
         //#########################|        |        |        | Type|  Type|  Type|    Type|     Type| Elementwise| Elementwise| Elementwise|Specialization|  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_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|                               Pipeline|                     Pipeline|
         //#########################|        |        |        |     |      |      |        |         |   Operation|   Operation|   Operation|              |      |      |      |      |    |    |    |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl|   _NWaveNPerXdl|                              Scheduler|                     Verision|
         //#########################|        |        |        |     |      |      |        |         |            |            |            |              |      |      |      |      |    |    |    |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |                                       |                             |
-        
+#if defined(CK_USE_AMD_MFMA_GFX950)
+#else        
         // Compute friendly
         DeviceGemm_Xdl_CShuffle_Streamk_V3<  Row,     Col,     Row,     F16,   F16,  F16,   F32,     F16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   256,   256,    32,   8,   8,  32,   32,    4,    4,     S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 32, 1, 8>,               8,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>,
         DeviceGemm_Xdl_CShuffle_Streamk_V3<  Row,     Col,     Row,     F16,   F16,  F16,   F32,     F16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   128,   128,    64,   8,   8,  32,   32,    2,    2,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 32, 1, 8>,               8,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>,
@@ -57,17 +58,19 @@ using device_gemm_xdl_universal_streamk_f16_f16_f16_mk_nk_mn_comp_instances = st
         DeviceGemm_Xdl_CShuffle_Streamk_V3<  Row,     Col,     Row,     F16,   F16,  F16,   F32,     F16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   128,   256,    32,   8,   8,  32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 32, 1, 8>,               8,  BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1>,
         DeviceGemm_Xdl_CShuffle_Streamk_V3<  Row,     Col,     Row,     F16,   F16,  F16,   F32,     F16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   256,   128,    32,   8,   8,  32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 32, 1, 8>,               8,  BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1>,
         DeviceGemm_Xdl_CShuffle_Streamk_V3<  Row,     Col,     Row,     F16,   F16,  F16,   F32,     F16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   128,   128,    64,   8,   8,  32,   32,    2,    2,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 32, 1, 8>,               8,  BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1>
+#endif // defined(CK_USE_AMD_MFMA_GFX950)
     // clang-format on
     >;
 
 template <BlockGemmPipelineScheduler BlkGemmPipeSched, GemmSpecialization GemmSpec>
 using device_gemm_xdl_universal_streamk_f16_f16_f16_mk_nk_mn_mem_instances = std::tuple<
-    // clang-format off
+// clang-format off
         //#########################| ALayout| BLayout| CLayout|AData| BData| CData| AccData| Cshuffle|           A|           B|           C|          GEMM| 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|    Block-wiseGemm|               Block-wiseGemm|
         //#########################|        |        |        | Type|  Type|  Type|    Type|     Type| Elementwise| Elementwise| Elementwise|Specialization|  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_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|          Pipeline|                     Pipeline|
         //#########################|        |        |        |     |      |      |        |         |   Operation|   Operation|   Operation|              |      |      |      |      |    |    |    |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl|   _NWaveNPerXdl|         Scheduler|                     Verision|
         //#########################|        |        |        |     |      |      |        |         |            |            |            |              |      |      |      |      |    |    |    |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |                  |                             |
-
+#if defined(CK_USE_AMD_MFMA_GFX950)
+#else        
         // Latency friendly 
         DeviceGemm_Xdl_CShuffle_Streamk_V3<  Row,     Col,     Row,     F16,   F16,  F16,   F32,     F16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   128,    32,   16,    64,   8,   8,  16,   16,    1,    1,     S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 16, 1, 8>,               2,  BlkGemmPipeSched, BlockGemmPipelineVersion::v1>,
         DeviceGemm_Xdl_CShuffle_Streamk_V3<  Row,     Col,     Row,     F16,   F16,  F16,   F32,     F16,      PassThrough, PassThrough, PassThrough,       GemmSpec,    64,    16,   16,   128,   8,   8,  16,   16,    1,    1,     S<16, 4, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<16, 4, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 16, 1, 4>,               4,  BlkGemmPipeSched, BlockGemmPipelineVersion::v1>,
@@ -90,6 +93,7 @@ using device_gemm_xdl_universal_streamk_f16_f16_f16_mk_nk_mn_mem_instances = std
         DeviceGemm_Xdl_CShuffle_Streamk_V3<  Row,     Col,     Row,     F16,   F16,  F16,   F32,     F16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   128,    32,  128,    64,   8,   8,  32,   32,    1,    2,     S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 16, 1, 8>,               8,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
         DeviceGemm_Xdl_CShuffle_Streamk_V3<  Row,     Col,     Row,     F16,   F16,  F16,   F32,     F16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   256,    16,  256,    64,   8,   8,  16,   16,    1,    4,     S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 16, 1, 16>,              4,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
         DeviceGemm_Xdl_CShuffle_Streamk_V3<  Row,     Col,     Row,     F16,   F16,  F16,   F32,     F16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   256,    32,  256,    64,   8,   8,  32,   32,    1,    2,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 16, 1, 16>,              8,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>
+#endif // defined(CK_USE_AMD_MFMA_GFX950)
     // clang-format on
     >;
 } // namespace instance

library/src/tensor_operation_instance/gpu/grouped_gemm/device_grouped_gemm_xdl_splitk_f16_f8_f16_mk_kn_mn_irregular_instance.cpp

@@ -27,15 +27,18 @@ using S = ck::Sequence<Is...>;
 
 using Empty_Tuple = ck::Tuple<>;
 
-using PassThrough                    = ck::tensor_operation::element_wise::PassThrough;
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+#if !defined(CK_USE_AMD_MFMA_GFX950)
 static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
-
+#endif
 using device_grouped_gemm_xdl_splitk_f16_f8_f16_mk_kn_mn_irregular_tile_instances = std::tuple<
-    // clang-format off
+// 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|
         //################################|       |       |            |       |      |      |        |         |            |      |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+#if defined(CK_USE_AMD_MFMA_GFX950)
+#else
         DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,    F8,     F32,      F16, Empty_Tuple,   F16, 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, PipelineVersion::v1>,
         DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,    F8,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,   128,   256,    32,   8,   8,   32,   32,    2,    4,  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,              4,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8, PipelineVersion::v1>,
         DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,    F8,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,   256,   192,    64,    32,   8,   8,   32,   32,    3,    1,  S<1, 4, 64, 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, 8>,              8, PipelineVersion::v1>,
@@ -98,6 +101,7 @@ using device_grouped_gemm_xdl_splitk_f16_f8_f16_mk_kn_mn_irregular_tile_instance
         DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,    F8,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,    64,    64,    64,    32,   8,   8,   32,   32,    2,    2,  S<1, 4, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 16, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 16, 1, 4>,              8, PipelineVersion::v2>,
         DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,    F8,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,    64,    64,    32,    32,   8,   8,   32,   32,    2,    1,  S<1, 4, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 16, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 16, 1, 4>,              8, PipelineVersion::v2>,
         DeviceGroupedGemmXdlSplitKCShuffle<    Row,    Row, Empty_Tuple,    Row,   F16,    F8,     F32,      F16, Empty_Tuple,   F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding,        1,    64,    32,    64,    32,   8,   8,   32,   32,    1,    2,  S<1, 4, 16, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 16, 1>,  S<0, 1, 3, 2>,  S<0, 1, 3, 2>,             2,              2,              8,         1,           1,           1,               S<1, 16, 1, 4>,              8, PipelineVersion::v2>
+#endif // defined(CK_USE_AMD_MFMA_GFX950)
     // clang-format on
     >;
 

library/src/tensor_operation_instance/gpu/grouped_gemm_tile_loop/device_grouped_gemm_xdl_tile_loop_multiply_bf16_i8_bf16_mk_kn_mn.hpp

@@ -44,11 +44,14 @@ template <typename DsLayout,
           typename CDEElementwiseOp,
           GemmSpecialization GemmSpec = GemmMNKPadding>
 using device_grouped_gemm_xdl_tile_loop_bf16_i8_bf16_mk_kn_mn_comp_instances = std::tuple<
-    // clang-format off
+// 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|
         //###########################################|       |       |            |       |      |      |        |         |            |      |            |            |                 |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |   S<C,D0...,D_N| 
+#if defined(CK_USE_AMD_MFMA_GFX950)
+        DeviceGroupedGemmMultipleDXdlCShuffleTileLoop<   Row,     Row,    DsLayout,    Row,  BF16,    I8,     F32,      F32,  DsDataType,  BF16, PassThrough, PassThrough, CDEElementwiseOp,       GemmSpec,        1,   256,   128,   128,    128,  16,   4,   32,   32,    2,    2,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         0,    S<16, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              8,              4,         0,           1,           1,               S<1, 32, 1, 8>,        S<8,8,8>,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>
+#else
         // DeviceGroupedGemmMultipleDXdlCShuffleTileLoop<   Row,     Row,    DsLayout,    Row,  BF16,    I8,     F32,      F32,  DsDataType,  BF16, PassThrough, PassThrough, CDEElementwiseOp,       GemmSpec,        1,   256,   256,   256,    32,   8,   4,   32,   32,    4,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              8,              4,         0,           1,           1,               S<1, 32, 1, 8>,        S<8,8,1>,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>,
         // DeviceGroupedGemmMultipleDXdlCShuffleTileLoop<   Row,     Row,    DsLayout,    Row,  BF16,    I8,     F32,      F32,  DsDataType,  BF16, PassThrough, PassThrough, CDEElementwiseOp,       GemmSpec,        1,   256,   128,   128,    64,   8,   4,   32,   32,    2,    2,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         0,    S<16, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              8,              4,         0,           1,           1,               S<1, 32, 1, 8>,        S<8,8,1>,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>,
         // DeviceGroupedGemmMultipleDXdlCShuffleTileLoop<   Row,     Row,    DsLayout,    Row,  BF16,    I8,     F32,      F32,  DsDataType,  BF16, PassThrough, PassThrough, CDEElementwiseOp,       GemmSpec,        1,   256,   256,   256,    32,   8,   4,   32,   32,    4,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              8,              4,         0,           1,           1,               S<1, 32, 1, 8>,        S<8,8,1>,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v5>,
@@ -57,6 +60,7 @@ using device_grouped_gemm_xdl_tile_loop_bf16_i8_bf16_mk_kn_mn_comp_instances = s
         // DeviceGroupedGemmMultipleDXdlCShuffleTileLoop<   Row,     Row,    DsLayout,    Row,  BF16,    I8,     F32,      F32,  DsDataType,  BF16, PassThrough, PassThrough, CDEElementwiseOp,       GemmSpec,        1,   256,   128,   128,    64,   8,   4,   32,   32,    2,    2,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         0,    S<16, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              8,              4,         0,           1,           1,               S<1, 32, 1, 8>,        S<8,8,1>,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>,
         // DeviceGroupedGemmMultipleDXdlCShuffleTileLoop<   Row,     Row,    DsLayout,    Row,  BF16,    I8,     F32,      F32,  DsDataType,  BF16, PassThrough, PassThrough, CDEElementwiseOp,       GemmSpec,        1,   256,   128,   256,    32,   8,   4,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              8,              4,         0,           1,           1,               S<1, 32, 1, 8>,        S<8,8,1>,  BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1>,
         DeviceGroupedGemmMultipleDXdlCShuffleTileLoop<   Row,     Row,    DsLayout,    Row,  BF16,    I8,     F32,      F32,  DsDataType,  BF16, PassThrough, PassThrough, CDEElementwiseOp,       GemmSpec,        1,   256,   128,   128,    64,   8,   4,   32,   32,    2,    2,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         0,    S<16, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              8,              4,         0,           1,           1,               S<1, 32, 1, 8>,        S<8,8,8>,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>
+#endif // defined(CK_USE_AMD_MFMA_GFX950)
     // clang-format on
     >;
 
@@ -65,13 +69,14 @@ template <typename DsLayout,
           typename CDEElementwiseOp,
           GemmSpecialization GemmSpec                 = GemmMNKPadding,
           BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave>
-using device_grouped_gemm_xdl_tile_loop_bf16_i8_bf16_mk_kn_mn_mem_instances =
-    std::tuple<
-        // clang-format off
+using device_grouped_gemm_xdl_tile_loop_bf16_i8_bf16_mk_kn_mn_mem_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|
         //###########################################|       |       |            |       |      |      |        |         |            |      |            |            |                 |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |   S<C,D0...,D_N| 
+#if defined(CK_USE_AMD_MFMA_GFX950)
+#else
         // Latency friendly
         // DeviceGroupedGemmMultipleDXdlCShuffleTileLoop<   Row,     Row,    DsLayout,    Row,  BF16,    I8,     F32,      F32,  DsDataType,  BF16, PassThrough, PassThrough, CDEElementwiseOp,       GemmSpec,        1,    64,    16,    16,   256,   8,   4,   16,   16,    1,    1,     S<32, 2, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         0,     S<64, 1, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,             16,              4,          0,          1,           1,               S<1, 16, 1, 4>,        S<4,4,1>,  BlkGemmPipeSched, BlockGemmPipelineVersion::v1>,
         // DeviceGroupedGemmMultipleDXdlCShuffleTileLoop<   Row,     Row,    DsLayout,    Row,  BF16,    I8,     F32,      F32,  DsDataType,  BF16, PassThrough, PassThrough, CDEElementwiseOp,       GemmSpec,        1,   128,    16,    32,   256,   8,   4,   16,   16,    1,    1,     S<32, 4, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         0,     S<64, 2, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,             16,              4,          0,          1,           1,               S<1, 16, 1, 8>,        S<4,4,1>,  BlkGemmPipeSched, BlockGemmPipelineVersion::v1>,
@@ -84,8 +89,9 @@ using device_grouped_gemm_xdl_tile_loop_bf16_i8_bf16_mk_kn_mn_mem_instances =
         // DeviceGroupedGemmMultipleDXdlCShuffleTileLoop<   Row,     Row,    DsLayout,    Row,  BF16,    I8,     F32,      F32,  DsDataType,  BF16, PassThrough, PassThrough, CDEElementwiseOp,       GemmSpec,        1,   128,    32,   128,    64,   8,   4,   32,   32,    1,    2,     S<8, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         0,     S<16, 8, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,             16,              4,          0,          1,           1,               S<1, 16, 1, 8>,        S<8,8,1>,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
         // DeviceGroupedGemmMultipleDXdlCShuffleTileLoop<   Row,     Row,    DsLayout,    Row,  BF16,    I8,     F32,      F32,  DsDataType,  BF16, PassThrough, PassThrough, CDEElementwiseOp,       GemmSpec,        1,   256,    16,   256,    64,   8,   4,   16,   16,    1,    4,     S<8, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         0,    S<16, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,             16,              4,          0,          1,           1,              S<1, 16, 1, 16>,        S<4,4,1>,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
         // DeviceGroupedGemmMultipleDXdlCShuffleTileLoop<   Row,     Row,    DsLayout,    Row,  BF16,    I8,     F32,      F32,  DsDataType,  BF16, PassThrough, PassThrough, CDEElementwiseOp,       GemmSpec,        1,   256,    32,   256,    64,   8,   4,   32,   32,    1,    2,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         0,    S<16, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,             16,              4,          0,          1,           1,              S<1, 16, 1, 16>,        S<8,8,1>,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>
-        // clang-format on
-        >;
+#endif // defined(CK_USE_AMD_MFMA_GFX950)
+       // clang-format on
+    >;
 
 } // namespace instance
 } // namespace device