From 5bdc993dbf90de10b737d71c91b58d33545550be Mon Sep 17 00:00:00 2001 From: Anca Hamuraru Date: Wed, 30 Apr 2025 13:02:57 +0000 Subject: [PATCH 01/11] Fixed cmake errors related to gemm_bilinear. Previously, if the above flags are set, cmake build fails: GPU_TARGETS="gfx1100;gfx1201" -D DTYPES="fp16;bf16;fp8" --- .../gpu/CMakeLists.txt | 9 ++++++++ profiler/src/CMakeLists.txt | 22 +++++++++++++++++-- 2 files changed, 29 insertions(+), 2 deletions(-) diff --git a/library/src/tensor_operation_instance/gpu/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/CMakeLists.txt index fe35d9ca76..a6cdd21bd6 100755 --- a/library/src/tensor_operation_instance/gpu/CMakeLists.txt +++ b/library/src/tensor_operation_instance/gpu/CMakeLists.txt @@ -279,6 +279,15 @@ FOREACH(subdir_path ${dir_list}) message("Found gemm_multiply_multiply_f8 instances, but gfx94/gfx95 not on the target list. Skipping.") set(add_inst 0) endif() + if ("${cmake_instance}" MATCHES "gemm_bilinear") + set(add_inst 0) + if((SUPPORTED_GPU_TARGETS MATCHES "gfx9") AND (DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)) + set(add_inst 1) + endif() + if((SUPPORTED_GPU_TARGETS MATCHES "gfx11" OR SUPPORTED_GPU_TARGETS MATCHES "gfx12") AND (DTYPES MATCHES "i8" OR NOT DEFINED DTYPES)) + set(add_inst 1) + endif() + endif() if((add_inst EQUAL 1)) get_filename_component(target_dir ${subdir_path} NAME) add_subdirectory(${target_dir}) diff --git a/profiler/src/CMakeLists.txt b/profiler/src/CMakeLists.txt index 17c8c277eb..b8f1077dff 100644 --- a/profiler/src/CMakeLists.txt +++ b/profiler/src/CMakeLists.txt @@ -71,10 +71,19 @@ if(SUPPORTED_GPU_TARGETS MATCHES "gfx9") endif() -if(SUPPORTED_GPU_TARGETS MATCHES "gfx11" OR SUPPORTED_GPU_TARGETS MATCHES "gfx12" OR SUPPORTED_GPU_TARGETS MATCHES "gfx9") +if(SUPPORTED_GPU_TARGETS MATCHES "gfx9") if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES) list(APPEND PROFILER_SOURCES profile_gemm_bilinear.cpp) endif() +endif() + +if(SUPPORTED_GPU_TARGETS MATCHES "gfx11" OR SUPPORTED_GPU_TARGETS MATCHES "gfx12") + if(DTYPES MATCHES "i8" OR NOT DEFINED DTYPES) + list(APPEND PROFILER_SOURCES profile_gemm_bilinear.cpp) + endif() +endif() + +if(SUPPORTED_GPU_TARGETS MATCHES "gfx11" OR SUPPORTED_GPU_TARGETS MATCHES "gfx12" OR SUPPORTED_GPU_TARGETS MATCHES "gfx9") list(APPEND PROFILER_SOURCES profile_gemm_universal.cpp) list(APPEND PROFILER_SOURCES profile_grouped_conv_fwd.cpp) list(APPEND PROFILER_SOURCES profile_grouped_conv_bwd_data.cpp) @@ -165,10 +174,19 @@ if(SUPPORTED_GPU_TARGETS MATCHES "gfx9") target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_fwd_convinvscale_instance) endif() -if(SUPPORTED_GPU_TARGETS MATCHES "gfx9" OR SUPPORTED_GPU_TARGETS MATCHES "gfx11" OR SUPPORTED_GPU_TARGETS MATCHES "gfx12") +if(SUPPORTED_GPU_TARGETS MATCHES "gfx9") if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES) target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_bilinear_instance) endif() +endif() + +if(SUPPORTED_GPU_TARGETS MATCHES "gfx11" OR SUPPORTED_GPU_TARGETS MATCHES "gfx12") + if(DTYPES MATCHES "i8" OR NOT DEFINED DTYPES) + target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_bilinear_instance) + endif() +endif() + +if(SUPPORTED_GPU_TARGETS MATCHES "gfx9" OR SUPPORTED_GPU_TARGETS MATCHES "gfx11" OR SUPPORTED_GPU_TARGETS MATCHES "gfx12") target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_instance) target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_fwd_instance) target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_bwd_data_instance) From 12176616b6b0c24d96c2bec0262e44bf7c2b9a79 Mon Sep 17 00:00:00 2001 From: Anca Hamuraru Date: Wed, 30 Apr 2025 13:03:42 +0000 Subject: [PATCH 02/11] Fixed cmake build errors related to test_fp8 --- test/data_type/CMakeLists.txt | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/test/data_type/CMakeLists.txt b/test/data_type/CMakeLists.txt index 8a0f631b39..8f6e9a0d15 100644 --- a/test/data_type/CMakeLists.txt +++ b/test/data_type/CMakeLists.txt @@ -16,15 +16,15 @@ if (CK_USE_OCP_FP8) add_gtest_executable(test_fp8_ocp test_fp8_ocp.cpp) if(result EQUAL 0) target_link_libraries(test_fp8_ocp PRIVATE utility) + add_dependencies(test_fp8 test_fp8_ocp) endif() add_gtest_executable(test_bf8_ocp test_bf8_ocp.cpp) if(result EQUAL 0) target_link_libraries(test_bf8_ocp PRIVATE utility) + add_dependencies(test_fp8 test_bf8_ocp) endif() - add_dependencies(test_fp8 test_fp8_ocp) - add_dependencies(test_fp8 test_bf8_ocp) endif() if (CK_USE_FNUZ_FP8) From df929f094ab0ec7551321e8f5b3d6ced7b4ad361 Mon Sep 17 00:00:00 2001 From: Anca Hamuraru Date: Thu, 1 May 2025 12:51:21 +0000 Subject: [PATCH 03/11] Updates to support mixed precision (cherry picked from commit e65d71180393e7b66169c56565a6bac740427de6) Co-authored-by: Anca Hamuraru --- .../gpu/block/blockwise_gemm_pipeline_wmmaops_base.hpp | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_wmmaops_base.hpp b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_wmmaops_base.hpp index a63d32802e..0adf9e7241 100644 --- a/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_wmmaops_base.hpp +++ b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_wmmaops_base.hpp @@ -282,7 +282,7 @@ struct BlockwiseGemmWmmaops_pipeline_base using AThreadCopy = ThreadwiseTensorSliceTransfer_v4, @@ -293,7 +293,7 @@ struct BlockwiseGemmWmmaops_pipeline_base using BThreadCopy = ThreadwiseTensorSliceTransfer_v4, From 03c0446ed237bc2639d2399a419eb2fe32aefb71 Mon Sep 17 00:00:00 2001 From: Anca Hamuraru Date: Thu, 1 May 2025 12:52:18 +0000 Subject: [PATCH 04/11] Adding support for RRR, F8xF16xF16 gemm_universal_wmma - wip (cherry picked from commit f8c06322df0abcbd5945a56cdf5bffe56480f9f0) Co-authored-by: Anca Hamuraru --- .../blockwise_gemm_pipeline_wmmaops_base.hpp | 2 +- .../gpu/gemm_universal.hpp | 11 ++++ .../gpu/gemm_universal_wmma.inc | 7 +- .../gpu/gemm_universal/CMakeLists.txt | 4 ++ ...emm_wmma_universal_f8_f16_f16_km_kn_mn.hpp | 65 +++++++++++++++++++ ...f16_f16_km_kn_mn_comp_default_instance.cpp | 24 +++++++ 6 files changed, 111 insertions(+), 2 deletions(-) create mode 100644 library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f8_f16_f16/device_gemm_wmma_universal_f8_f16_f16_km_kn_mn.hpp create mode 100644 library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f8_f16_f16/device_gemm_wmma_universal_f8_f16_f16_km_kn_mn_comp_default_instance.cpp diff --git a/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_wmmaops_base.hpp b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_wmmaops_base.hpp index 0adf9e7241..0147a890dd 100644 --- a/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_wmmaops_base.hpp +++ b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_wmmaops_base.hpp @@ -61,7 +61,7 @@ struct BlockwiseGemmWmmaops_pipeline_base static_assert(KPack % (B_K1 * B_KRow) == 0, "wrong!"); static constexpr auto wmma_gemm = - WmmaGemm{}; + WmmaGemm{}; static constexpr index_t KRepeat = KPerBlock / KPack; diff --git a/library/include/ck/library/tensor_operation_instance/gpu/gemm_universal.hpp b/library/include/ck/library/tensor_operation_instance/gpu/gemm_universal.hpp index 79212e16dd..dd1be28d97 100644 --- a/library/include/ck/library/tensor_operation_instance/gpu/gemm_universal.hpp +++ b/library/include/ck/library/tensor_operation_instance/gpu/gemm_universal.hpp @@ -128,6 +128,17 @@ struct DeviceOperationInstanceFactory< } } #endif +#if(defined(CK_ENABLE_FP16) && defined(CK_ENABLE_FP8)) + if constexpr(is_same_v && is_same_v && + is_same_v) + { + if constexpr(is_same_v && is_same_v && + is_same_v) + { + add_device_gemm_wmma_universal_f8_f16_f16_mk_kn_mn_comp_default_instances(op_ptrs); + } + } +#endif #endif // CK_USE_WMMA #ifdef CK_USE_XDL diff --git a/library/include/ck/library/tensor_operation_instance/gpu/gemm_universal_wmma.inc b/library/include/ck/library/tensor_operation_instance/gpu/gemm_universal_wmma.inc index 1396437326..599898caa5 100644 --- a/library/include/ck/library/tensor_operation_instance/gpu/gemm_universal_wmma.inc +++ b/library/include/ck/library/tensor_operation_instance/gpu/gemm_universal_wmma.inc @@ -61,7 +61,12 @@ void add_device_gemm_wmma_universal_f8_f8_bf16_mk_nk_mn_comp_default_instances( DeviceGemmV2>>& instances); #endif - +#if(defined(CK_ENABLE_FP16) && defined(CK_ENABLE_FP8)) +void add_device_gemm_wmma_universal_f8_f16_f16_mk_kn_mn_comp_default_instances( + std::vector>>& + instances); +#endif } // namespace instance } // namespace device } // namespace tensor_operation diff --git a/library/src/tensor_operation_instance/gpu/gemm_universal/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/gemm_universal/CMakeLists.txt index 18eeefa522..e81bd03660 100644 --- a/library/src/tensor_operation_instance/gpu/gemm_universal/CMakeLists.txt +++ b/library/src/tensor_operation_instance/gpu/gemm_universal/CMakeLists.txt @@ -12,6 +12,8 @@ list(APPEND GEMM_UNIVERSAL_INSTANCES device_gemm_wmma_universal_bf16_bf16_bf16/device_gemm_wmma_universal_bf16_bf16_bf16_km_kn_mn_comp_default_instance.cpp device_gemm_wmma_universal_bf16_bf16_bf16/device_gemm_wmma_universal_bf16_bf16_bf16_km_nk_mn_comp_default_instance.cpp + device_gemm_wmma_universal_f8_f16_f16/device_gemm_wmma_universal_f8_f16_f16_km_kn_mn_comp_default_instance.cpp + device_gemm_xdl_universal_f16_f16_f16/device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn_comp_default_instance.cpp device_gemm_xdl_universal_f16_f16_f16/device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn_comp_kpadding_instance.cpp device_gemm_xdl_universal_f16_f16_f16/device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn_comp_mnpadding_instance.cpp @@ -77,6 +79,8 @@ set_source_files_properties(device_gemm_wmma_universal_bf16_bf16_bf16/device_gem set_source_files_properties(device_gemm_wmma_universal_bf16_bf16_bf16/device_gemm_wmma_universal_bf16_bf16_bf16_km_kn_mn_comp_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1") set_source_files_properties(device_gemm_wmma_universal_bf16_bf16_bf16/device_gemm_wmma_universal_bf16_bf16_bf16_km_nk_mn_comp_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1") +set_source_files_properties(device_gemm_wmma_universal_f8_f16_f16/device_gemm_wmma_universal_f8_f16_f16_km_kn_mn_comp_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1") + set_source_files_properties(device_gemm_xdl_universal_f16_f16_f16/device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn_comp_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1") set_source_files_properties(device_gemm_xdl_universal_f16_f16_f16/device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn_comp_kpadding_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1") set_source_files_properties(device_gemm_xdl_universal_f16_f16_f16/device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn_comp_mnpadding_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1") diff --git a/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f8_f16_f16/device_gemm_wmma_universal_f8_f16_f16_km_kn_mn.hpp b/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f8_f16_f16/device_gemm_wmma_universal_f8_f16_f16_km_kn_mn.hpp new file mode 100644 index 0000000000..0040ecda52 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f8_f16_f16/device_gemm_wmma_universal_f8_f16_f16_km_kn_mn.hpp @@ -0,0 +1,65 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved. + +#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_gemm_wmma_cshuffle_v3.hpp" + +#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +using F8 = f8_t; +using F16 = half_t; +using F32 = float; + +using Row = tensor_layout::gemm::RowMajor; +using Col = tensor_layout::gemm::ColumnMajor; + +template +using S = Sequence; + +using PassThrough = element_wise::PassThrough; + +static constexpr auto GemmDefault = GemmSpecialization::Default; +static constexpr auto GemmKPadding = GemmSpecialization::KPadding; +static constexpr auto GemmMNPadding = GemmSpecialization::MNPadding; +static constexpr auto GemmMNKPadding = GemmSpecialization::MNKPadding; + +static constexpr auto Intrawave = BlockGemmPipelineScheduler::Intrawave; +static constexpr auto Interwave = BlockGemmPipelineScheduler::Interwave; + +template +using device_gemm_wmma_universal_f8_f16_f16_km_kn_mn_comp_instances = + std::tuple< + // clang-format off + //#########################| ALayout| BLayout| CLayout|AData| BData| CData| AccData| CShuffle| A| B| C| GemmSpec| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MRepeat| NRepeat| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CShuffleBlockTransfer| CShuffleBlockTransfer| BlkGemm| BlkGemm| + //#########################| | | | Type| Type| Type| Type| DataType| Elementwise| Elementwise| Elementwise| | Size| Block| Block| Block| | | Wmma| Wmma| | | ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraN| MRepeat| NRepeat| ClusterLengths| ScalarPerVector| PipeSched| PipelineVer| + //#########################| | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | Lengths_AK0_M_AK1| ArrangeOrder| | | PerVector| PerVector_AK1| | Lengths_BK0_N_BK1| ArrangeOrder| | | PerVector| PerVector_BK1| | PerShuffle| PerShuffle| _MBlock_MPerBlock| _NPerBlock| | | + //#########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | _NBlock_NPerBlock| | | | + DeviceGemm_Wmma_CShuffleV3< Col, Row, Row, F8, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 32, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Col, Row, Row, F8, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 64, 64, 8, 8, 16, 16, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Col, Row, Row, F8, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 64, 32, 32, 8, 8, 16, 16, 4, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Col, Row, Row, F8, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 32, 16, 16, 32, 8, 8, 16, 16, 1, 1, S<2, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<2, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 16, 1, 2>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Col, Row, Row, F8, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 256, 64, 8, 8, 16, 16, 4, 4, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Col, Row, Row, F8, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 64, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Col, Row, Row, F8, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 160, 64, 8, 8, 16, 16, 2, 5, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 64, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Col, Row, Row, F8, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 128, 32, 8, 8, 16, 16, 4, 4, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Col, Row, Row, F8, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 256, 64, 64, 8, 8, 16, 16, 8, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Col, Row, Row, F8, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 64, 256, 64, 8, 8, 16, 16, 2, 8, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Col, Row, Row, F8, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 64, 80, 64, 8, 8, 16, 16, 1, 5, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 64, 1, 2>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Col, Row, Row, F8, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 64, 64, 8, 8, 16, 16, 1, 2, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Col, Row, Row, F8, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 64, 32, 64, 8, 8, 16, 16, 4, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Col, Row, Row, F8, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 32, 64, 64, 8, 8, 16, 16, 2, 2, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Col, Row, Row, F8, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 32, 16, 32, 64, 8, 8, 16, 16, 1, 2, S<2, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<2, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 16, 1, 2>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Col, Row, Row, F8, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 32, 16, 16, 64, 8, 8, 16, 16, 1, 1, S<2, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<2, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 16, 1, 2>, 8, Intrawave, BlockGemmPipelineVersion::v3> + // clang-format on + >; +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f8_f16_f16/device_gemm_wmma_universal_f8_f16_f16_km_kn_mn_comp_default_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f8_f16_f16/device_gemm_wmma_universal_f8_f16_f16_km_kn_mn_comp_default_instance.cpp new file mode 100644 index 0000000000..90b9ad8e64 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f8_f16_f16/device_gemm_wmma_universal_f8_f16_f16_km_kn_mn_comp_default_instance.cpp @@ -0,0 +1,24 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved. + +#include "device_gemm_wmma_universal_f8_f16_f16_km_kn_mn.hpp" +#include "ck/host_utility/device_prop.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +void add_device_gemm_wmma_universal_f8_f16_f16_km_kn_mn_comp_default_instances( + std::vector>>& + instances) +{ + add_device_operation_instances( + instances, device_gemm_wmma_universal_f8_f16_f16_km_kn_mn_comp_instances{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck From 977d8a66494e63c37356214e48e4e386781d6faf Mon Sep 17 00:00:00 2001 From: Anca Hamuraru Date: Thu, 1 May 2025 12:53:15 +0000 Subject: [PATCH 05/11] Added support for F8xF16xF16 to gemm_wmma_universal (cherry picked from commit 15c851de6daa513a12c2e3af299bab0176175fb5) Co-authored-by: Anca Hamuraru --- .../gpu/gemm_universal.hpp | 15 ++++ .../gpu/gemm_universal_wmma.inc | 15 ++++ .../gpu/gemm_universal/CMakeLists.txt | 6 ++ ...emm_wmma_universal_f8_f16_f16_km_nk_mn.hpp | 65 ++++++++++++++++++ ...f16_f16_km_nk_mn_comp_default_instance.cpp | 24 +++++++ ...emm_wmma_universal_f8_f16_f16_mk_kn_mn.hpp | 68 +++++++++++++++++++ ...f16_f16_mk_kn_mn_comp_default_instance.cpp | 24 +++++++ ...emm_wmma_universal_f8_f16_f16_mk_nk_mn.hpp | 65 ++++++++++++++++++ ...f16_f16_mk_nk_mn_comp_default_instance.cpp | 24 +++++++ .../test_gemm_universal_wmma_fp16.cpp | 31 +++++++++ 10 files changed, 337 insertions(+) create mode 100644 library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f8_f16_f16/device_gemm_wmma_universal_f8_f16_f16_km_nk_mn.hpp create mode 100644 library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f8_f16_f16/device_gemm_wmma_universal_f8_f16_f16_km_nk_mn_comp_default_instance.cpp create mode 100644 library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f8_f16_f16/device_gemm_wmma_universal_f8_f16_f16_mk_kn_mn.hpp create mode 100644 library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f8_f16_f16/device_gemm_wmma_universal_f8_f16_f16_mk_kn_mn_comp_default_instance.cpp create mode 100644 library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f8_f16_f16/device_gemm_wmma_universal_f8_f16_f16_mk_nk_mn.hpp create mode 100644 library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f8_f16_f16/device_gemm_wmma_universal_f8_f16_f16_mk_nk_mn_comp_default_instance.cpp diff --git a/library/include/ck/library/tensor_operation_instance/gpu/gemm_universal.hpp b/library/include/ck/library/tensor_operation_instance/gpu/gemm_universal.hpp index dd1be28d97..68a37fb420 100644 --- a/library/include/ck/library/tensor_operation_instance/gpu/gemm_universal.hpp +++ b/library/include/ck/library/tensor_operation_instance/gpu/gemm_universal.hpp @@ -137,6 +137,21 @@ struct DeviceOperationInstanceFactory< { add_device_gemm_wmma_universal_f8_f16_f16_mk_kn_mn_comp_default_instances(op_ptrs); } + else if constexpr(is_same_v && is_same_v && + is_same_v) + { + add_device_gemm_wmma_universal_f8_f16_f16_mk_nk_mn_comp_default_instances(op_ptrs); + } + else if constexpr(is_same_v && is_same_v && + is_same_v) + { + add_device_gemm_wmma_universal_f8_f16_f16_km_kn_mn_comp_default_instances(op_ptrs); + } + else if constexpr(is_same_v && is_same_v && + is_same_v) + { + add_device_gemm_wmma_universal_f8_f16_f16_km_nk_mn_comp_default_instances(op_ptrs); + } } #endif #endif // CK_USE_WMMA diff --git a/library/include/ck/library/tensor_operation_instance/gpu/gemm_universal_wmma.inc b/library/include/ck/library/tensor_operation_instance/gpu/gemm_universal_wmma.inc index 599898caa5..7ae315eb17 100644 --- a/library/include/ck/library/tensor_operation_instance/gpu/gemm_universal_wmma.inc +++ b/library/include/ck/library/tensor_operation_instance/gpu/gemm_universal_wmma.inc @@ -66,6 +66,21 @@ void add_device_gemm_wmma_universal_f8_f16_f16_mk_kn_mn_comp_default_instances( std::vector>>& instances); + +void add_device_gemm_wmma_universal_f8_f16_f16_mk_nk_mn_comp_default_instances( + std::vector>>& + instances); + +void add_device_gemm_wmma_universal_f8_f16_f16_km_kn_mn_comp_default_instances( + std::vector>>& + instances); + +void add_device_gemm_wmma_universal_f8_f16_f16_km_nk_mn_comp_default_instances( + std::vector>>& + instances); #endif } // namespace instance } // namespace device diff --git a/library/src/tensor_operation_instance/gpu/gemm_universal/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/gemm_universal/CMakeLists.txt index e81bd03660..e33562ad59 100644 --- a/library/src/tensor_operation_instance/gpu/gemm_universal/CMakeLists.txt +++ b/library/src/tensor_operation_instance/gpu/gemm_universal/CMakeLists.txt @@ -12,7 +12,10 @@ list(APPEND GEMM_UNIVERSAL_INSTANCES device_gemm_wmma_universal_bf16_bf16_bf16/device_gemm_wmma_universal_bf16_bf16_bf16_km_kn_mn_comp_default_instance.cpp device_gemm_wmma_universal_bf16_bf16_bf16/device_gemm_wmma_universal_bf16_bf16_bf16_km_nk_mn_comp_default_instance.cpp + device_gemm_wmma_universal_f8_f16_f16/device_gemm_wmma_universal_f8_f16_f16_mk_kn_mn_comp_default_instance.cpp + device_gemm_wmma_universal_f8_f16_f16/device_gemm_wmma_universal_f8_f16_f16_mk_nk_mn_comp_default_instance.cpp device_gemm_wmma_universal_f8_f16_f16/device_gemm_wmma_universal_f8_f16_f16_km_kn_mn_comp_default_instance.cpp + device_gemm_wmma_universal_f8_f16_f16/device_gemm_wmma_universal_f8_f16_f16_km_nk_mn_comp_default_instance.cpp device_gemm_xdl_universal_f16_f16_f16/device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn_comp_default_instance.cpp device_gemm_xdl_universal_f16_f16_f16/device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn_comp_kpadding_instance.cpp @@ -79,7 +82,10 @@ set_source_files_properties(device_gemm_wmma_universal_bf16_bf16_bf16/device_gem set_source_files_properties(device_gemm_wmma_universal_bf16_bf16_bf16/device_gemm_wmma_universal_bf16_bf16_bf16_km_kn_mn_comp_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1") set_source_files_properties(device_gemm_wmma_universal_bf16_bf16_bf16/device_gemm_wmma_universal_bf16_bf16_bf16_km_nk_mn_comp_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1") +set_source_files_properties(device_gemm_wmma_universal_f8_f16_f16/device_gemm_wmma_universal_f8_f16_f16_mk_kn_mn_comp_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1") +set_source_files_properties(device_gemm_wmma_universal_f8_f16_f16/device_gemm_wmma_universal_f8_f16_f16_mk_nk_mn_comp_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1") set_source_files_properties(device_gemm_wmma_universal_f8_f16_f16/device_gemm_wmma_universal_f8_f16_f16_km_kn_mn_comp_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1") +set_source_files_properties(device_gemm_wmma_universal_f8_f16_f16/device_gemm_wmma_universal_f8_f16_f16_km_nk_mn_comp_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1") set_source_files_properties(device_gemm_xdl_universal_f16_f16_f16/device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn_comp_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1") set_source_files_properties(device_gemm_xdl_universal_f16_f16_f16/device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn_comp_kpadding_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1") diff --git a/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f8_f16_f16/device_gemm_wmma_universal_f8_f16_f16_km_nk_mn.hpp b/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f8_f16_f16/device_gemm_wmma_universal_f8_f16_f16_km_nk_mn.hpp new file mode 100644 index 0000000000..06f05c721b --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f8_f16_f16/device_gemm_wmma_universal_f8_f16_f16_km_nk_mn.hpp @@ -0,0 +1,65 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved. + +#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_gemm_wmma_cshuffle_v3.hpp" + +#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +using F8 = f8_t; +using F16 = half_t; +using F32 = float; + +using Row = tensor_layout::gemm::RowMajor; +using Col = tensor_layout::gemm::ColumnMajor; + +template +using S = Sequence; + +using PassThrough = element_wise::PassThrough; + +static constexpr auto GemmDefault = GemmSpecialization::Default; +static constexpr auto GemmKPadding = GemmSpecialization::KPadding; +static constexpr auto GemmMNPadding = GemmSpecialization::MNPadding; +static constexpr auto GemmMNKPadding = GemmSpecialization::MNKPadding; + +static constexpr auto Intrawave = BlockGemmPipelineScheduler::Intrawave; +static constexpr auto Interwave = BlockGemmPipelineScheduler::Interwave; + +template +using device_gemm_wmma_universal_f8_f16_f16_km_nk_mn_comp_instances = + std::tuple< + // clang-format off + //#########################| ALayout| BLayout| CLayout|AData| BData| CData| AccData| CShuffle| A| B| C| GemmSpec| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MRepeat| NRepeat| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CShuffleBlockTransfer| CShuffleBlockTransfer| BlkGemm| BlkGemm| + //#########################| | | | Type| Type| Type| Type| DataType| Elementwise| Elementwise| Elementwise| | Size| Block| Block| Block| | | Wmma| Wmma| | | ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraN| MRepeat| NRepeat| ClusterLengths| ScalarPerVector| PipeSched| PipelineVer| + //#########################| | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | Lengths_AK0_M_AK1| ArrangeOrder| | | PerVector| PerVector_AK1| | Lengths_BK0_N_BK1| ArrangeOrder| | | PerVector| PerVector_BK1| | PerShuffle| PerShuffle| _MBlock_MPerBlock| _NPerBlock| | | + //#########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | _NBlock_NPerBlock| | | | + DeviceGemm_Wmma_CShuffleV3< Col, Col, Row, F8, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 32, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Col, Col, Row, F8, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 64, 64, 8, 8, 16, 16, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Col, Col, Row, F8, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 64, 32, 32, 8, 8, 16, 16, 4, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Col, Col, Row, F8, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 32, 16, 16, 32, 8, 8, 16, 16, 1, 1, S<2, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 2>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Col, Col, Row, F8, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 256, 64, 8, 8, 16, 16, 4, 4, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Col, Col, Row, F8, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 64, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Col, Col, Row, F8, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 160, 64, 8, 8, 16, 16, 2, 5, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 64, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Col, Col, Row, F8, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 128, 32, 8, 8, 16, 16, 4, 4, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Col, Col, Row, F8, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 256, 64, 64, 8, 8, 16, 16, 8, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Col, Col, Row, F8, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 64, 256, 64, 8, 8, 16, 16, 2, 8, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Col, Col, Row, F8, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 64, 80, 64, 8, 8, 16, 16, 1, 5, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 64, 1, 2>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Col, Col, Row, F8, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 64, 64, 8, 8, 16, 16, 1, 2, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Col, Col, Row, F8, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 64, 32, 64, 8, 8, 16, 16, 4, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Col, Col, Row, F8, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 32, 64, 64, 8, 8, 16, 16, 2, 2, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Col, Col, Row, F8, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 32, 16, 32, 64, 8, 8, 16, 16, 1, 2, S<2, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 2>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Col, Col, Row, F8, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 32, 16, 16, 64, 8, 8, 16, 16, 1, 1, S<2, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 2>, 8, Intrawave, BlockGemmPipelineVersion::v3> + // clang-format on + >; +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f8_f16_f16/device_gemm_wmma_universal_f8_f16_f16_km_nk_mn_comp_default_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f8_f16_f16/device_gemm_wmma_universal_f8_f16_f16_km_nk_mn_comp_default_instance.cpp new file mode 100644 index 0000000000..5fa17f6f45 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f8_f16_f16/device_gemm_wmma_universal_f8_f16_f16_km_nk_mn_comp_default_instance.cpp @@ -0,0 +1,24 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved. + +#include "device_gemm_wmma_universal_f8_f16_f16_km_nk_mn.hpp" +#include "ck/host_utility/device_prop.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +void add_device_gemm_wmma_universal_f8_f16_f16_km_nk_mn_comp_default_instances( + std::vector>>& + instances) +{ + add_device_operation_instances( + instances, device_gemm_wmma_universal_f8_f16_f16_km_nk_mn_comp_instances{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f8_f16_f16/device_gemm_wmma_universal_f8_f16_f16_mk_kn_mn.hpp b/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f8_f16_f16/device_gemm_wmma_universal_f8_f16_f16_mk_kn_mn.hpp new file mode 100644 index 0000000000..20b196834a --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f8_f16_f16/device_gemm_wmma_universal_f8_f16_f16_mk_kn_mn.hpp @@ -0,0 +1,68 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved. + +#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_gemm_wmma_cshuffle_v3.hpp" + +#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +using F8 = f8_t; +using F16 = half_t; +using F32 = float; + +using Row = tensor_layout::gemm::RowMajor; +using Col = tensor_layout::gemm::ColumnMajor; + +template +using S = Sequence; + +using PassThrough = element_wise::PassThrough; + +static constexpr auto GemmDefault = GemmSpecialization::Default; +static constexpr auto GemmKPadding = GemmSpecialization::KPadding; +static constexpr auto GemmMNPadding = GemmSpecialization::MNPadding; +static constexpr auto GemmMNKPadding = GemmSpecialization::MNKPadding; + +static constexpr auto Intrawave = BlockGemmPipelineScheduler::Intrawave; +static constexpr auto Interwave = BlockGemmPipelineScheduler::Interwave; + +template +using device_gemm_wmma_universal_f8_f16_f16_mk_kn_mn_comp_instances = + std::tuple< + // clang-format off + //#########################| ALayout| BLayout| CLayout|AData| BData| CData| AccData| CShuffle| A| B| C| GemmSpec| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MRepeat| NRepeat| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CShuffleBlockTransfer| CShuffleBlockTransfer| BlkGemm| BlkGemm| + //#########################| | | | Type| Type| Type| Type| DataType| Elementwise| Elementwise| Elementwise| | Size| Block| Block| Block| | | Wmma| Wmma| | | ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraN| MRepeat| NRepeat| ClusterLengths| ScalarPerVector| PipeSched| PipelineVer| + //#########################| | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | Lengths_AK0_M_AK1| ArrangeOrder| | | PerVector| PerVector_AK1| | Lengths_BK0_N_BK1| ArrangeOrder| | | PerVector| PerVector_BK1| | PerShuffle| PerShuffle| _MBlock_MPerBlock| _NPerBlock| | | + //#########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | _NBlock_NPerBlock| | | | + DeviceGemm_Wmma_CShuffleV3< Row, Row, Row, F8, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 32, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Row, Row, F8, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 64, 64, 8, 8, 16, 16, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Row, Row, F8, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 64, 32, 32, 8, 8, 16, 16, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Row, Row, F8, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 32, 16, 16, 32, 8, 8, 16, 16, 1, 1, S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<2, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 16, 1, 2>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Row, Row, F8, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 256, 64, 8, 8, 16, 16, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Row, Row, F8, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 64, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Row, Row, F8, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 160, 64, 8, 8, 16, 16, 2, 5, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 64, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Row, Row, F8, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 128, 32, 8, 8, 16, 16, 4, 4, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Row, Row, F8, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 256, 64, 64, 8, 8, 16, 16, 8, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Row, Row, F8, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 64, 256, 64, 8, 8, 16, 16, 2, 8, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Row, Row, F8, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 64, 80, 64, 8, 8, 16, 16, 1, 5, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 64, 1, 2>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Row, Row, F8, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 64, 64, 8, 8, 16, 16, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Row, Row, F8, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 64, 32, 64, 8, 8, 16, 16, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Row, Row, F8, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 32, 64, 64, 8, 8, 16, 16, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Row, Row, F8, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 32, 16, 32, 64, 8, 8, 16, 16, 1, 2, S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<2, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 16, 1, 2>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Row, Row, F8, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 32, 16, 16, 64, 8, 8, 16, 16, 1, 1, S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<2, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 16, 1, 2>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + // Configurations used during development, mainly for testing + DeviceGemm_Wmma_CShuffleV3< Row, Row, Row, F8, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 32, 16, 16, 64, 8, 8, 16, 16, 1, 1, S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<2, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 16, 1, 2>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Row, Row, F8, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 64, 64, 32, 8, 8, 16, 16, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 1, 1, 1, S<1, 32, 1, 2>, 8, Intrawave, BlockGemmPipelineVersion::v3> + // clang-format on + >; +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f8_f16_f16/device_gemm_wmma_universal_f8_f16_f16_mk_kn_mn_comp_default_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f8_f16_f16/device_gemm_wmma_universal_f8_f16_f16_mk_kn_mn_comp_default_instance.cpp new file mode 100644 index 0000000000..3af30df47a --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f8_f16_f16/device_gemm_wmma_universal_f8_f16_f16_mk_kn_mn_comp_default_instance.cpp @@ -0,0 +1,24 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved. + +#include "device_gemm_wmma_universal_f8_f16_f16_mk_kn_mn.hpp" +#include "ck/host_utility/device_prop.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +void add_device_gemm_wmma_universal_f8_f16_f16_mk_kn_mn_comp_default_instances( + std::vector>>& + instances) +{ + add_device_operation_instances( + instances, device_gemm_wmma_universal_f8_f16_f16_mk_kn_mn_comp_instances{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f8_f16_f16/device_gemm_wmma_universal_f8_f16_f16_mk_nk_mn.hpp b/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f8_f16_f16/device_gemm_wmma_universal_f8_f16_f16_mk_nk_mn.hpp new file mode 100644 index 0000000000..da91beb161 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f8_f16_f16/device_gemm_wmma_universal_f8_f16_f16_mk_nk_mn.hpp @@ -0,0 +1,65 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved. + +#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_gemm_wmma_cshuffle_v3.hpp" + +#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +using F8 = f8_t; +using F16 = half_t; +using F32 = float; + +using Row = tensor_layout::gemm::RowMajor; +using Col = tensor_layout::gemm::ColumnMajor; + +template +using S = Sequence; + +using PassThrough = element_wise::PassThrough; + +static constexpr auto GemmDefault = GemmSpecialization::Default; +static constexpr auto GemmKPadding = GemmSpecialization::KPadding; +static constexpr auto GemmMNPadding = GemmSpecialization::MNPadding; +static constexpr auto GemmMNKPadding = GemmSpecialization::MNKPadding; + +static constexpr auto Intrawave = BlockGemmPipelineScheduler::Intrawave; +static constexpr auto Interwave = BlockGemmPipelineScheduler::Interwave; + +template +using device_gemm_wmma_universal_f8_f16_f16_mk_nk_mn_comp_instances = + std::tuple< + // clang-format off + //#########################| ALayout| BLayout| CLayout|AData| BData| CData| AccData| CShuffle| A| B| C| GemmSpec| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MRepeat| NRepeat| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CShuffleBlockTransfer| CShuffleBlockTransfer| BlkGemm| BlkGemm| + //#########################| | | | Type| Type| Type| Type| DataType| Elementwise| Elementwise| Elementwise| | Size| Block| Block| Block| | | Wmma| Wmma| | | ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraN| MRepeat| NRepeat| ClusterLengths| ScalarPerVector| PipeSched| PipelineVer| + //#########################| | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | Lengths_AK0_M_AK1| ArrangeOrder| | | PerVector| PerVector_AK1| | Lengths_BK0_N_BK1| ArrangeOrder| | | PerVector| PerVector_BK1| | PerShuffle| PerShuffle| _MBlock_MPerBlock| _NPerBlock| | | + //#########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | _NBlock_NPerBlock| | | | + DeviceGemm_Wmma_CShuffleV3< Row, Col, Row, F8, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 32, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Col, Row, F8, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 64, 64, 8, 8, 16, 16, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Col, Row, F8, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 64, 32, 32, 8, 8, 16, 16, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Col, Row, F8, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 32, 16, 16, 32, 8, 8, 16, 16, 1, 1, S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 2>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Col, Row, F8, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 256, 64, 8, 8, 16, 16, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Col, Row, F8, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 64, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Col, Row, F8, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 160, 64, 8, 8, 16, 16, 2, 5, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 64, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Col, Row, F8, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 128, 32, 8, 8, 16, 16, 4, 4, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Col, Row, F8, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 256, 64, 64, 8, 8, 16, 16, 8, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Col, Row, F8, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 64, 256, 64, 8, 8, 16, 16, 2, 8, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Col, Row, F8, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 64, 80, 64, 8, 8, 16, 16, 1, 5, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 64, 1, 2>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Col, Row, F8, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 64, 64, 8, 8, 16, 16, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Col, Row, F8, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 64, 32, 64, 8, 8, 16, 16, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Col, Row, F8, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 32, 64, 64, 8, 8, 16, 16, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Col, Row, F8, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 32, 16, 32, 64, 8, 8, 16, 16, 1, 2, S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 2>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Col, Row, F8, F16, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 32, 16, 16, 64, 8, 8, 16, 16, 1, 1, S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 2>, 8, Intrawave, BlockGemmPipelineVersion::v3> + // clang-format on + >; +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f8_f16_f16/device_gemm_wmma_universal_f8_f16_f16_mk_nk_mn_comp_default_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f8_f16_f16/device_gemm_wmma_universal_f8_f16_f16_mk_nk_mn_comp_default_instance.cpp new file mode 100644 index 0000000000..57a4bbd3c7 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f8_f16_f16/device_gemm_wmma_universal_f8_f16_f16_mk_nk_mn_comp_default_instance.cpp @@ -0,0 +1,24 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved. + +#include "device_gemm_wmma_universal_f8_f16_f16_mk_nk_mn.hpp" +#include "ck/host_utility/device_prop.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +void add_device_gemm_wmma_universal_f8_f16_f16_mk_nk_mn_comp_default_instances( + std::vector>>& + instances) +{ + add_device_operation_instances( + instances, device_gemm_wmma_universal_f8_f16_f16_mk_nk_mn_comp_instances{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/test/gemm_universal/test_gemm_universal_wmma_fp16.cpp b/test/gemm_universal/test_gemm_universal_wmma_fp16.cpp index 1adee41ed2..c7b2f78cd0 100644 --- a/test/gemm_universal/test_gemm_universal_wmma_fp16.cpp +++ b/test/gemm_universal/test_gemm_universal_wmma_fp16.cpp @@ -7,6 +7,7 @@ #include "ck/tensor_operation/gpu/device/tensor_layout.hpp" #include "test_gemm_universal_util.hpp" +using F8 = ck::f8_t; using F16 = ck::half_t; using F32 = float; @@ -39,19 +40,49 @@ class TestGemmUniversal_FP16_MK_NK { }; +template +class TestGemmUniversal_FP16_KM_KN + : public ck::test::TestGemmUniversal, Tuple>::type> +{ +}; + // clang-format off using KernelTypes_MK_KN = ::testing::Types< // ADataType, BDataType, ComputeDataType, CDataType +#if defined(CK_ENABLE_FP8) + std::tuple< F8, F16, F16, F16>, +#endif std::tuple< F16, F16, F16, F16> >; using KernelTypes_MK_NK = ::testing::Types< // ADataType, BDataType, ComputeDataType, CDataType +#if defined(CK_ENABLE_FP8) + std::tuple< F8, F16, F16, F16>, +#endif std::tuple< F16, F16, F16, F16> >; + +using KernelTypes_KM_NK = ::testing::Types< + // ADataType, BDataType, ComputeDataType, CDataType +#if defined(CK_ENABLE_FP8) + std::tuple< F8, F16, F16, F16>, +#endif + std::tuple< F16, F16, F16, F16> + >; + +using KernelTypes_KM_KN = ::testing::Types< + // ADataType, BDataType, ComputeDataType, CDataType +#if defined(CK_ENABLE_FP8) + std::tuple< F8, F16, F16, F16>, +#endif + std::tuple< F16, F16, F16, F16> + >; // clang-format on TYPED_TEST_SUITE(TestGemmUniversal_FP16_MK_KN, KernelTypes_MK_KN); TYPED_TEST_SUITE(TestGemmUniversal_FP16_MK_NK, KernelTypes_MK_NK); +TYPED_TEST_SUITE(TestGemmUniversal_FP16_KM_NK, KernelTypes_KM_NK); +TYPED_TEST_SUITE(TestGemmUniversal_FP16_KM_KN, KernelTypes_KM_KN); #include "test_gemm_universal_ut_cases_fp16.inc" From 636fbd5517faacbb1ba56f0072e6ce1a78e8261a Mon Sep 17 00:00:00 2001 From: Anca Hamuraru Date: Mon, 5 May 2025 08:47:30 +0000 Subject: [PATCH 06/11] Added support for F16xF8xF16 to gemm_wmma_universal --- .../gpu/gemm_universal.hpp | 25 +++++++ .../gpu/gemm_universal_wmma.inc | 20 ++++++ .../gpu/gemm_universal/CMakeLists.txt | 10 +++ ...emm_wmma_universal_f16_f8_f16_km_kn_mn.hpp | 65 ++++++++++++++++++ ..._f8_f16_km_kn_mn_comp_default_instance.cpp | 24 +++++++ ...emm_wmma_universal_f16_f8_f16_km_nk_mn.hpp | 65 ++++++++++++++++++ ..._f8_f16_km_nk_mn_comp_default_instance.cpp | 24 +++++++ ...emm_wmma_universal_f16_f8_f16_mk_kn_mn.hpp | 68 +++++++++++++++++++ ..._f8_f16_mk_kn_mn_comp_default_instance.cpp | 24 +++++++ ...emm_wmma_universal_f16_f8_f16_mk_nk_mn.hpp | 65 ++++++++++++++++++ ..._f8_f16_mk_nk_mn_comp_default_instance.cpp | 24 +++++++ .../test_gemm_universal_wmma_fp16.cpp | 2 + 12 files changed, 416 insertions(+) create mode 100644 library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f16_f8_f16/device_gemm_wmma_universal_f16_f8_f16_km_kn_mn.hpp create mode 100644 library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f16_f8_f16/device_gemm_wmma_universal_f16_f8_f16_km_kn_mn_comp_default_instance.cpp create mode 100644 library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f16_f8_f16/device_gemm_wmma_universal_f16_f8_f16_km_nk_mn.hpp create mode 100644 library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f16_f8_f16/device_gemm_wmma_universal_f16_f8_f16_km_nk_mn_comp_default_instance.cpp create mode 100644 library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f16_f8_f16/device_gemm_wmma_universal_f16_f8_f16_mk_kn_mn.hpp create mode 100644 library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f16_f8_f16/device_gemm_wmma_universal_f16_f8_f16_mk_kn_mn_comp_default_instance.cpp create mode 100644 library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f16_f8_f16/device_gemm_wmma_universal_f16_f8_f16_mk_nk_mn.hpp create mode 100644 library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f16_f8_f16/device_gemm_wmma_universal_f16_f8_f16_mk_nk_mn_comp_default_instance.cpp diff --git a/library/include/ck/library/tensor_operation_instance/gpu/gemm_universal.hpp b/library/include/ck/library/tensor_operation_instance/gpu/gemm_universal.hpp index 68a37fb420..0e94be9f39 100644 --- a/library/include/ck/library/tensor_operation_instance/gpu/gemm_universal.hpp +++ b/library/include/ck/library/tensor_operation_instance/gpu/gemm_universal.hpp @@ -153,6 +153,31 @@ struct DeviceOperationInstanceFactory< add_device_gemm_wmma_universal_f8_f16_f16_km_nk_mn_comp_default_instances(op_ptrs); } } + + if constexpr(is_same_v && is_same_v && + is_same_v) + { + if constexpr(is_same_v && is_same_v && + is_same_v) + { + add_device_gemm_wmma_universal_f16_f8_f16_mk_kn_mn_comp_default_instances(op_ptrs); + } + else if constexpr(is_same_v && is_same_v && + is_same_v) + { + add_device_gemm_wmma_universal_f16_f8_f16_mk_nk_mn_comp_default_instances(op_ptrs); + } + else if constexpr(is_same_v && is_same_v && + is_same_v) + { + add_device_gemm_wmma_universal_f16_f8_f16_km_kn_mn_comp_default_instances(op_ptrs); + } + else if constexpr(is_same_v && is_same_v && + is_same_v) + { + add_device_gemm_wmma_universal_f16_f8_f16_km_nk_mn_comp_default_instances(op_ptrs); + } + } #endif #endif // CK_USE_WMMA diff --git a/library/include/ck/library/tensor_operation_instance/gpu/gemm_universal_wmma.inc b/library/include/ck/library/tensor_operation_instance/gpu/gemm_universal_wmma.inc index 7ae315eb17..78006c992c 100644 --- a/library/include/ck/library/tensor_operation_instance/gpu/gemm_universal_wmma.inc +++ b/library/include/ck/library/tensor_operation_instance/gpu/gemm_universal_wmma.inc @@ -81,6 +81,26 @@ void add_device_gemm_wmma_universal_f8_f16_f16_km_nk_mn_comp_default_instances( std::vector>>& instances); + +void add_device_gemm_wmma_universal_f16_f8_f16_mk_kn_mn_comp_default_instances( + std::vector>>& + instances); + +void add_device_gemm_wmma_universal_f16_f8_f16_mk_nk_mn_comp_default_instances( + std::vector>>& + instances); + +void add_device_gemm_wmma_universal_f16_f8_f16_km_kn_mn_comp_default_instances( + std::vector>>& + instances); + +void add_device_gemm_wmma_universal_f16_f8_f16_km_nk_mn_comp_default_instances( + std::vector>>& + instances); #endif } // namespace instance } // namespace device diff --git a/library/src/tensor_operation_instance/gpu/gemm_universal/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/gemm_universal/CMakeLists.txt index e33562ad59..d4ace298ef 100644 --- a/library/src/tensor_operation_instance/gpu/gemm_universal/CMakeLists.txt +++ b/library/src/tensor_operation_instance/gpu/gemm_universal/CMakeLists.txt @@ -16,6 +16,11 @@ list(APPEND GEMM_UNIVERSAL_INSTANCES device_gemm_wmma_universal_f8_f16_f16/device_gemm_wmma_universal_f8_f16_f16_mk_nk_mn_comp_default_instance.cpp device_gemm_wmma_universal_f8_f16_f16/device_gemm_wmma_universal_f8_f16_f16_km_kn_mn_comp_default_instance.cpp device_gemm_wmma_universal_f8_f16_f16/device_gemm_wmma_universal_f8_f16_f16_km_nk_mn_comp_default_instance.cpp + + device_gemm_wmma_universal_f16_f8_f16/device_gemm_wmma_universal_f16_f8_f16_mk_kn_mn_comp_default_instance.cpp + device_gemm_wmma_universal_f16_f8_f16/device_gemm_wmma_universal_f16_f8_f16_mk_nk_mn_comp_default_instance.cpp + device_gemm_wmma_universal_f16_f8_f16/device_gemm_wmma_universal_f16_f8_f16_km_kn_mn_comp_default_instance.cpp + device_gemm_wmma_universal_f16_f8_f16/device_gemm_wmma_universal_f16_f8_f16_km_nk_mn_comp_default_instance.cpp device_gemm_xdl_universal_f16_f16_f16/device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn_comp_default_instance.cpp device_gemm_xdl_universal_f16_f16_f16/device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn_comp_kpadding_instance.cpp @@ -87,6 +92,11 @@ set_source_files_properties(device_gemm_wmma_universal_f8_f16_f16/device_gemm_wm set_source_files_properties(device_gemm_wmma_universal_f8_f16_f16/device_gemm_wmma_universal_f8_f16_f16_km_kn_mn_comp_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1") set_source_files_properties(device_gemm_wmma_universal_f8_f16_f16/device_gemm_wmma_universal_f8_f16_f16_km_nk_mn_comp_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1") +set_source_files_properties(device_gemm_wmma_universal_f16_f8_f16/device_gemm_wmma_universal_f16_f8_f16_mk_kn_mn_comp_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1") +set_source_files_properties(device_gemm_wmma_universal_f16_f8_f16/device_gemm_wmma_universal_f16_f8_f16_mk_nk_mn_comp_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1") +set_source_files_properties(device_gemm_wmma_universal_f16_f8_f16/device_gemm_wmma_universal_f16_f8_f16_km_kn_mn_comp_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1") +set_source_files_properties(device_gemm_wmma_universal_f16_f8_f16/device_gemm_wmma_universal_f16_f8_f16_km_nk_mn_comp_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1") + set_source_files_properties(device_gemm_xdl_universal_f16_f16_f16/device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn_comp_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1") set_source_files_properties(device_gemm_xdl_universal_f16_f16_f16/device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn_comp_kpadding_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1") set_source_files_properties(device_gemm_xdl_universal_f16_f16_f16/device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn_comp_mnpadding_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1") diff --git a/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f16_f8_f16/device_gemm_wmma_universal_f16_f8_f16_km_kn_mn.hpp b/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f16_f8_f16/device_gemm_wmma_universal_f16_f8_f16_km_kn_mn.hpp new file mode 100644 index 0000000000..c98cd429dd --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f16_f8_f16/device_gemm_wmma_universal_f16_f8_f16_km_kn_mn.hpp @@ -0,0 +1,65 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved. + +#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_gemm_wmma_cshuffle_v3.hpp" + +#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +using F8 = f8_t; +using F16 = half_t; +using F32 = float; + +using Row = tensor_layout::gemm::RowMajor; +using Col = tensor_layout::gemm::ColumnMajor; + +template +using S = Sequence; + +using PassThrough = element_wise::PassThrough; + +static constexpr auto GemmDefault = GemmSpecialization::Default; +static constexpr auto GemmKPadding = GemmSpecialization::KPadding; +static constexpr auto GemmMNPadding = GemmSpecialization::MNPadding; +static constexpr auto GemmMNKPadding = GemmSpecialization::MNKPadding; + +static constexpr auto Intrawave = BlockGemmPipelineScheduler::Intrawave; +static constexpr auto Interwave = BlockGemmPipelineScheduler::Interwave; + +template +using device_gemm_wmma_universal_f16_f8_f16_km_kn_mn_comp_instances = + std::tuple< + // clang-format off + //#########################| ALayout| BLayout| CLayout|AData| BData| CData| AccData| CShuffle| A| B| C| GemmSpec| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MRepeat| NRepeat| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CShuffleBlockTransfer| CShuffleBlockTransfer| BlkGemm| BlkGemm| + //#########################| | | | Type| Type| Type| Type| DataType| Elementwise| Elementwise| Elementwise| | Size| Block| Block| Block| | | Wmma| Wmma| | | ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraN| MRepeat| NRepeat| ClusterLengths| ScalarPerVector| PipeSched| PipelineVer| + //#########################| | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | Lengths_AK0_M_AK1| ArrangeOrder| | | PerVector| PerVector_AK1| | Lengths_BK0_N_BK1| ArrangeOrder| | | PerVector| PerVector_BK1| | PerShuffle| PerShuffle| _MBlock_MPerBlock| _NPerBlock| | | + //#########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | _NBlock_NPerBlock| | | | + DeviceGemm_Wmma_CShuffleV3< Col, Row, Row, F16, F8, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 32, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Col, Row, Row, F16, F8, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 64, 64, 8, 8, 16, 16, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Col, Row, Row, F16, F8, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 64, 32, 32, 8, 8, 16, 16, 4, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Col, Row, Row, F16, F8, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 32, 16, 16, 32, 8, 8, 16, 16, 1, 1, S<2, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<2, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 16, 1, 2>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Col, Row, Row, F16, F8, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 256, 64, 8, 8, 16, 16, 4, 4, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Col, Row, Row, F16, F8, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 64, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Col, Row, Row, F16, F8, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 160, 64, 8, 8, 16, 16, 2, 5, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 64, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Col, Row, Row, F16, F8, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 128, 32, 8, 8, 16, 16, 4, 4, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Col, Row, Row, F16, F8, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 256, 64, 64, 8, 8, 16, 16, 8, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Col, Row, Row, F16, F8, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 64, 256, 64, 8, 8, 16, 16, 2, 8, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Col, Row, Row, F16, F8, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 64, 80, 64, 8, 8, 16, 16, 1, 5, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 64, 1, 2>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Col, Row, Row, F16, F8, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 64, 64, 8, 8, 16, 16, 1, 2, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Col, Row, Row, F16, F8, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 64, 32, 64, 8, 8, 16, 16, 4, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Col, Row, Row, F16, F8, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 32, 64, 64, 8, 8, 16, 16, 2, 2, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Col, Row, Row, F16, F8, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 32, 16, 32, 64, 8, 8, 16, 16, 1, 2, S<2, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<2, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 16, 1, 2>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Col, Row, Row, F16, F8, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 32, 16, 16, 64, 8, 8, 16, 16, 1, 1, S<2, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<2, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 16, 1, 2>, 8, Intrawave, BlockGemmPipelineVersion::v3> + // clang-format on + >; +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f16_f8_f16/device_gemm_wmma_universal_f16_f8_f16_km_kn_mn_comp_default_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f16_f8_f16/device_gemm_wmma_universal_f16_f8_f16_km_kn_mn_comp_default_instance.cpp new file mode 100644 index 0000000000..cfd0a7aa8b --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f16_f8_f16/device_gemm_wmma_universal_f16_f8_f16_km_kn_mn_comp_default_instance.cpp @@ -0,0 +1,24 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved. + +#include "device_gemm_wmma_universal_f16_f8_f16_km_kn_mn.hpp" +#include "ck/host_utility/device_prop.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +void add_device_gemm_wmma_universal_f16_f8_f16_km_kn_mn_comp_default_instances( + std::vector>>& + instances) +{ + add_device_operation_instances( + instances, device_gemm_wmma_universal_f16_f8_f16_km_kn_mn_comp_instances{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f16_f8_f16/device_gemm_wmma_universal_f16_f8_f16_km_nk_mn.hpp b/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f16_f8_f16/device_gemm_wmma_universal_f16_f8_f16_km_nk_mn.hpp new file mode 100644 index 0000000000..dbc0434d9a --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f16_f8_f16/device_gemm_wmma_universal_f16_f8_f16_km_nk_mn.hpp @@ -0,0 +1,65 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved. + +#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_gemm_wmma_cshuffle_v3.hpp" + +#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +using F8 = f8_t; +using F16 = half_t; +using F32 = float; + +using Row = tensor_layout::gemm::RowMajor; +using Col = tensor_layout::gemm::ColumnMajor; + +template +using S = Sequence; + +using PassThrough = element_wise::PassThrough; + +static constexpr auto GemmDefault = GemmSpecialization::Default; +static constexpr auto GemmKPadding = GemmSpecialization::KPadding; +static constexpr auto GemmMNPadding = GemmSpecialization::MNPadding; +static constexpr auto GemmMNKPadding = GemmSpecialization::MNKPadding; + +static constexpr auto Intrawave = BlockGemmPipelineScheduler::Intrawave; +static constexpr auto Interwave = BlockGemmPipelineScheduler::Interwave; + +template +using device_gemm_wmma_universal_f16_f8_f16_km_nk_mn_comp_instances = + std::tuple< + // clang-format off + //#########################| ALayout| BLayout| CLayout|AData| BData| CData| AccData| CShuffle| A| B| C| GemmSpec| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MRepeat| NRepeat| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CShuffleBlockTransfer| CShuffleBlockTransfer| BlkGemm| BlkGemm| + //#########################| | | | Type| Type| Type| Type| DataType| Elementwise| Elementwise| Elementwise| | Size| Block| Block| Block| | | Wmma| Wmma| | | ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraN| MRepeat| NRepeat| ClusterLengths| ScalarPerVector| PipeSched| PipelineVer| + //#########################| | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | Lengths_AK0_M_AK1| ArrangeOrder| | | PerVector| PerVector_AK1| | Lengths_BK0_N_BK1| ArrangeOrder| | | PerVector| PerVector_BK1| | PerShuffle| PerShuffle| _MBlock_MPerBlock| _NPerBlock| | | + //#########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | _NBlock_NPerBlock| | | | + DeviceGemm_Wmma_CShuffleV3< Col, Col, Row, F16, F8, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 32, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Col, Col, Row, F16, F8, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 64, 64, 8, 8, 16, 16, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Col, Col, Row, F16, F8, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 64, 32, 32, 8, 8, 16, 16, 4, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Col, Col, Row, F16, F8, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 32, 16, 16, 32, 8, 8, 16, 16, 1, 1, S<2, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 2>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Col, Col, Row, F16, F8, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 256, 64, 8, 8, 16, 16, 4, 4, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Col, Col, Row, F16, F8, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 64, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Col, Col, Row, F16, F8, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 160, 64, 8, 8, 16, 16, 2, 5, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 64, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Col, Col, Row, F16, F8, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 128, 32, 8, 8, 16, 16, 4, 4, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Col, Col, Row, F16, F8, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 256, 64, 64, 8, 8, 16, 16, 8, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Col, Col, Row, F16, F8, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 64, 256, 64, 8, 8, 16, 16, 2, 8, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Col, Col, Row, F16, F8, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 64, 80, 64, 8, 8, 16, 16, 1, 5, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 64, 1, 2>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Col, Col, Row, F16, F8, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 64, 64, 8, 8, 16, 16, 1, 2, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Col, Col, Row, F16, F8, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 64, 32, 64, 8, 8, 16, 16, 4, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Col, Col, Row, F16, F8, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 32, 64, 64, 8, 8, 16, 16, 2, 2, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Col, Col, Row, F16, F8, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 32, 16, 32, 64, 8, 8, 16, 16, 1, 2, S<2, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 2>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Col, Col, Row, F16, F8, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 32, 16, 16, 64, 8, 8, 16, 16, 1, 1, S<2, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 2>, 8, Intrawave, BlockGemmPipelineVersion::v3> + // clang-format on + >; +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f16_f8_f16/device_gemm_wmma_universal_f16_f8_f16_km_nk_mn_comp_default_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f16_f8_f16/device_gemm_wmma_universal_f16_f8_f16_km_nk_mn_comp_default_instance.cpp new file mode 100644 index 0000000000..1f736e775b --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f16_f8_f16/device_gemm_wmma_universal_f16_f8_f16_km_nk_mn_comp_default_instance.cpp @@ -0,0 +1,24 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved. + +#include "device_gemm_wmma_universal_f16_f8_f16_km_nk_mn.hpp" +#include "ck/host_utility/device_prop.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +void add_device_gemm_wmma_universal_f16_f8_f16_km_nk_mn_comp_default_instances( + std::vector>>& + instances) +{ + add_device_operation_instances( + instances, device_gemm_wmma_universal_f16_f8_f16_km_nk_mn_comp_instances{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f16_f8_f16/device_gemm_wmma_universal_f16_f8_f16_mk_kn_mn.hpp b/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f16_f8_f16/device_gemm_wmma_universal_f16_f8_f16_mk_kn_mn.hpp new file mode 100644 index 0000000000..b976d5dc0d --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f16_f8_f16/device_gemm_wmma_universal_f16_f8_f16_mk_kn_mn.hpp @@ -0,0 +1,68 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved. + +#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_gemm_wmma_cshuffle_v3.hpp" + +#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +using F8 = f8_t; +using F16 = half_t; +using F32 = float; + +using Row = tensor_layout::gemm::RowMajor; +using Col = tensor_layout::gemm::ColumnMajor; + +template +using S = Sequence; + +using PassThrough = element_wise::PassThrough; + +static constexpr auto GemmDefault = GemmSpecialization::Default; +static constexpr auto GemmKPadding = GemmSpecialization::KPadding; +static constexpr auto GemmMNPadding = GemmSpecialization::MNPadding; +static constexpr auto GemmMNKPadding = GemmSpecialization::MNKPadding; + +static constexpr auto Intrawave = BlockGemmPipelineScheduler::Intrawave; +static constexpr auto Interwave = BlockGemmPipelineScheduler::Interwave; + +template +using device_gemm_wmma_universal_f16_f8_f16_mk_kn_mn_comp_instances = + std::tuple< + // clang-format off + //#########################| ALayout| BLayout| CLayout|AData| BData| CData| AccData| CShuffle| A| B| C| GemmSpec| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MRepeat| NRepeat| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CShuffleBlockTransfer| CShuffleBlockTransfer| BlkGemm| BlkGemm| + //#########################| | | | Type| Type| Type| Type| DataType| Elementwise| Elementwise| Elementwise| | Size| Block| Block| Block| | | Wmma| Wmma| | | ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraN| MRepeat| NRepeat| ClusterLengths| ScalarPerVector| PipeSched| PipelineVer| + //#########################| | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | Lengths_AK0_M_AK1| ArrangeOrder| | | PerVector| PerVector_AK1| | Lengths_BK0_N_BK1| ArrangeOrder| | | PerVector| PerVector_BK1| | PerShuffle| PerShuffle| _MBlock_MPerBlock| _NPerBlock| | | + //#########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | _NBlock_NPerBlock| | | | + DeviceGemm_Wmma_CShuffleV3< Row, Row, Row, F16, F8, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 32, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Row, Row, F16, F8, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 64, 64, 8, 8, 16, 16, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Row, Row, F16, F8, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 64, 32, 32, 8, 8, 16, 16, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Row, Row, F16, F8, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 32, 16, 16, 32, 8, 8, 16, 16, 1, 1, S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<2, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 16, 1, 2>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Row, Row, F16, F8, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 256, 64, 8, 8, 16, 16, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Row, Row, F16, F8, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 64, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Row, Row, F16, F8, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 160, 64, 8, 8, 16, 16, 2, 5, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 64, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Row, Row, F16, F8, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 128, 32, 8, 8, 16, 16, 4, 4, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Row, Row, F16, F8, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 256, 64, 64, 8, 8, 16, 16, 8, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Row, Row, F16, F8, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 64, 256, 64, 8, 8, 16, 16, 2, 8, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Row, Row, F16, F8, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 64, 80, 64, 8, 8, 16, 16, 1, 5, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 64, 1, 2>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Row, Row, F16, F8, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 64, 64, 8, 8, 16, 16, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Row, Row, F16, F8, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 64, 32, 64, 8, 8, 16, 16, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Row, Row, F16, F8, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 32, 64, 64, 8, 8, 16, 16, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Row, Row, F16, F8, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 32, 16, 32, 64, 8, 8, 16, 16, 1, 2, S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<2, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 16, 1, 2>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Row, Row, F16, F8, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 32, 16, 16, 64, 8, 8, 16, 16, 1, 1, S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<2, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 16, 1, 2>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + // Configurations used during development, mainly for testing + DeviceGemm_Wmma_CShuffleV3< Row, Row, Row, F16, F8, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 32, 16, 16, 64, 8, 8, 16, 16, 1, 1, S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<2, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 16, 1, 2>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Row, Row, F16, F8, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 64, 64, 32, 8, 8, 16, 16, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 1, 1, 1, S<1, 32, 1, 2>, 8, Intrawave, BlockGemmPipelineVersion::v3> + // clang-format on + >; +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f16_f8_f16/device_gemm_wmma_universal_f16_f8_f16_mk_kn_mn_comp_default_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f16_f8_f16/device_gemm_wmma_universal_f16_f8_f16_mk_kn_mn_comp_default_instance.cpp new file mode 100644 index 0000000000..08f9cb533b --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f16_f8_f16/device_gemm_wmma_universal_f16_f8_f16_mk_kn_mn_comp_default_instance.cpp @@ -0,0 +1,24 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved. + +#include "device_gemm_wmma_universal_f16_f8_f16_mk_kn_mn.hpp" +#include "ck/host_utility/device_prop.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +void add_device_gemm_wmma_universal_f16_f8_f16_mk_kn_mn_comp_default_instances( + std::vector>>& + instances) +{ + add_device_operation_instances( + instances, device_gemm_wmma_universal_f16_f8_f16_mk_kn_mn_comp_instances{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f16_f8_f16/device_gemm_wmma_universal_f16_f8_f16_mk_nk_mn.hpp b/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f16_f8_f16/device_gemm_wmma_universal_f16_f8_f16_mk_nk_mn.hpp new file mode 100644 index 0000000000..c93204cec9 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f16_f8_f16/device_gemm_wmma_universal_f16_f8_f16_mk_nk_mn.hpp @@ -0,0 +1,65 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved. + +#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_gemm_wmma_cshuffle_v3.hpp" + +#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +using F8 = f8_t; +using F16 = half_t; +using F32 = float; + +using Row = tensor_layout::gemm::RowMajor; +using Col = tensor_layout::gemm::ColumnMajor; + +template +using S = Sequence; + +using PassThrough = element_wise::PassThrough; + +static constexpr auto GemmDefault = GemmSpecialization::Default; +static constexpr auto GemmKPadding = GemmSpecialization::KPadding; +static constexpr auto GemmMNPadding = GemmSpecialization::MNPadding; +static constexpr auto GemmMNKPadding = GemmSpecialization::MNKPadding; + +static constexpr auto Intrawave = BlockGemmPipelineScheduler::Intrawave; +static constexpr auto Interwave = BlockGemmPipelineScheduler::Interwave; + +template +using device_gemm_wmma_universal_f16_f8_f16_mk_nk_mn_comp_instances = + std::tuple< + // clang-format off + //#########################| ALayout| BLayout| CLayout|AData| BData| CData| AccData| CShuffle| A| B| C| GemmSpec| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MRepeat| NRepeat| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CShuffleBlockTransfer| CShuffleBlockTransfer| BlkGemm| BlkGemm| + //#########################| | | | Type| Type| Type| Type| DataType| Elementwise| Elementwise| Elementwise| | Size| Block| Block| Block| | | Wmma| Wmma| | | ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraN| MRepeat| NRepeat| ClusterLengths| ScalarPerVector| PipeSched| PipelineVer| + //#########################| | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | Lengths_AK0_M_AK1| ArrangeOrder| | | PerVector| PerVector_AK1| | Lengths_BK0_N_BK1| ArrangeOrder| | | PerVector| PerVector_BK1| | PerShuffle| PerShuffle| _MBlock_MPerBlock| _NPerBlock| | | + //#########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | _NBlock_NPerBlock| | | | + DeviceGemm_Wmma_CShuffleV3< Row, Col, Row, F16, F8, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 32, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Col, Row, F16, F8, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 64, 64, 8, 8, 16, 16, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Col, Row, F16, F8, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 64, 32, 32, 8, 8, 16, 16, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Col, Row, F16, F8, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 32, 16, 16, 32, 8, 8, 16, 16, 1, 1, S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 2>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Col, Row, F16, F8, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 256, 64, 8, 8, 16, 16, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Col, Row, F16, F8, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 64, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Col, Row, F16, F8, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 160, 64, 8, 8, 16, 16, 2, 5, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 64, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Col, Row, F16, F8, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 128, 32, 8, 8, 16, 16, 4, 4, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Col, Row, F16, F8, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 256, 64, 64, 8, 8, 16, 16, 8, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Col, Row, F16, F8, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 64, 256, 64, 8, 8, 16, 16, 2, 8, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Col, Row, F16, F8, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 64, 80, 64, 8, 8, 16, 16, 1, 5, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 64, 1, 2>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Col, Row, F16, F8, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 64, 64, 8, 8, 16, 16, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Col, Row, F16, F8, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 64, 32, 64, 8, 8, 16, 16, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Col, Row, F16, F8, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 32, 64, 64, 8, 8, 16, 16, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Col, Row, F16, F8, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 32, 16, 32, 64, 8, 8, 16, 16, 1, 2, S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 2>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Col, Row, F16, F8, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 32, 16, 16, 64, 8, 8, 16, 16, 1, 1, S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 2>, 8, Intrawave, BlockGemmPipelineVersion::v3> + // clang-format on + >; +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f16_f8_f16/device_gemm_wmma_universal_f16_f8_f16_mk_nk_mn_comp_default_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f16_f8_f16/device_gemm_wmma_universal_f16_f8_f16_mk_nk_mn_comp_default_instance.cpp new file mode 100644 index 0000000000..5a3fd38c2f --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f16_f8_f16/device_gemm_wmma_universal_f16_f8_f16_mk_nk_mn_comp_default_instance.cpp @@ -0,0 +1,24 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved. + +#include "device_gemm_wmma_universal_f16_f8_f16_mk_nk_mn.hpp" +#include "ck/host_utility/device_prop.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +void add_device_gemm_wmma_universal_f16_f8_f16_mk_nk_mn_comp_default_instances( + std::vector>>& + instances) +{ + add_device_operation_instances( + instances, device_gemm_wmma_universal_f16_f8_f16_mk_nk_mn_comp_instances{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/test/gemm_universal/test_gemm_universal_wmma_fp16.cpp b/test/gemm_universal/test_gemm_universal_wmma_fp16.cpp index c7b2f78cd0..ae9be8566c 100644 --- a/test/gemm_universal/test_gemm_universal_wmma_fp16.cpp +++ b/test/gemm_universal/test_gemm_universal_wmma_fp16.cpp @@ -67,6 +67,7 @@ using KernelTypes_KM_NK = ::testing::Types< // ADataType, BDataType, ComputeDataType, CDataType #if defined(CK_ENABLE_FP8) std::tuple< F8, F16, F16, F16>, + std::tuple< F16, F8, F16, F16>, #endif std::tuple< F16, F16, F16, F16> >; @@ -75,6 +76,7 @@ using KernelTypes_KM_KN = ::testing::Types< // ADataType, BDataType, ComputeDataType, CDataType #if defined(CK_ENABLE_FP8) std::tuple< F8, F16, F16, F16>, + std::tuple< F16, F8, F16, F16>, #endif std::tuple< F16, F16, F16, F16> >; From c5d99d22959bc21901d93205d37110449b869caf Mon Sep 17 00:00:00 2001 From: Anca Hamuraru Date: Mon, 5 May 2025 09:55:29 +0000 Subject: [PATCH 07/11] Added support for BF16xI4xBF16 to gemm_wmma_universal (cherry picked from commit c6a4a69d2d43d59bae8bdabfae80d648646f217e) Co-authored-by: Anca Hamuraru --- .../gpu/gemm_universal.hpp | 18 +++++ .../gpu/gemm_universal_wmma.inc | 10 +++ .../gpu/gemm_universal/CMakeLists.txt | 6 ++ ...m_wmma_universal_bf16_i4_bf16_km_nk_mn.hpp | 66 +++++++++++++++++++ ...i4_bf16_km_nk_mn_comp_default_instance.cpp | 24 +++++++ ...m_wmma_universal_bf16_i4_bf16_mk_nk_mn.hpp | 66 +++++++++++++++++++ ...i4_bf16_mk_nk_mn_comp_default_instance.cpp | 24 +++++++ .../test_gemm_universal_wmma_bf16.cpp | 4 ++ 8 files changed, 218 insertions(+) create mode 100644 library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_bf16_i4_bf16/device_gemm_wmma_universal_bf16_i4_bf16_km_nk_mn.hpp create mode 100644 library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_bf16_i4_bf16/device_gemm_wmma_universal_bf16_i4_bf16_km_nk_mn_comp_default_instance.cpp create mode 100644 library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_bf16_i4_bf16/device_gemm_wmma_universal_bf16_i4_bf16_mk_nk_mn.hpp create mode 100644 library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_bf16_i4_bf16/device_gemm_wmma_universal_bf16_i4_bf16_mk_nk_mn_comp_default_instance.cpp diff --git a/library/include/ck/library/tensor_operation_instance/gpu/gemm_universal.hpp b/library/include/ck/library/tensor_operation_instance/gpu/gemm_universal.hpp index 0e94be9f39..6aa6088c17 100644 --- a/library/include/ck/library/tensor_operation_instance/gpu/gemm_universal.hpp +++ b/library/include/ck/library/tensor_operation_instance/gpu/gemm_universal.hpp @@ -128,6 +128,24 @@ struct DeviceOperationInstanceFactory< } } #endif +#if(defined(CK_ENABLE_BF16) && defined(CK_ENABLE_FP8)) + if constexpr(is_same_v && is_same_v && + is_same_v) + { + if constexpr(is_same_v && is_same_v && + is_same_v) + { + add_device_gemm_wmma_universal_bf16_i4_bf16_mk_nk_mn_comp_default_instances( + op_ptrs); + } + else if constexpr(is_same_v && is_same_v && + is_same_v) + { + add_device_gemm_wmma_universal_bf16_i4_bf16_km_nk_mn_comp_default_instances( + op_ptrs); + } + } +#endif #if(defined(CK_ENABLE_FP16) && defined(CK_ENABLE_FP8)) if constexpr(is_same_v && is_same_v && is_same_v) diff --git a/library/include/ck/library/tensor_operation_instance/gpu/gemm_universal_wmma.inc b/library/include/ck/library/tensor_operation_instance/gpu/gemm_universal_wmma.inc index 78006c992c..72f56f5dc0 100644 --- a/library/include/ck/library/tensor_operation_instance/gpu/gemm_universal_wmma.inc +++ b/library/include/ck/library/tensor_operation_instance/gpu/gemm_universal_wmma.inc @@ -56,6 +56,16 @@ void add_device_gemm_wmma_universal_f8_f8_bf16_mk_kn_mn_comp_default_instances( DeviceGemmV2>>& instances); +void add_device_gemm_wmma_universal_bf16_i4_bf16_mk_nk_mn_comp_default_instances( + std::vector>>& + instances); + +void add_device_gemm_wmma_universal_bf16_i4_bf16_km_nk_mn_comp_default_instances( + std::vector>>& + instances); + void add_device_gemm_wmma_universal_f8_f8_bf16_mk_nk_mn_comp_default_instances( std::vector>>& diff --git a/library/src/tensor_operation_instance/gpu/gemm_universal/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/gemm_universal/CMakeLists.txt index d4ace298ef..c857949424 100644 --- a/library/src/tensor_operation_instance/gpu/gemm_universal/CMakeLists.txt +++ b/library/src/tensor_operation_instance/gpu/gemm_universal/CMakeLists.txt @@ -12,6 +12,9 @@ list(APPEND GEMM_UNIVERSAL_INSTANCES device_gemm_wmma_universal_bf16_bf16_bf16/device_gemm_wmma_universal_bf16_bf16_bf16_km_kn_mn_comp_default_instance.cpp device_gemm_wmma_universal_bf16_bf16_bf16/device_gemm_wmma_universal_bf16_bf16_bf16_km_nk_mn_comp_default_instance.cpp + device_gemm_wmma_universal_bf16_i4_bf16/device_gemm_wmma_universal_bf16_i4_bf16_mk_nk_mn_comp_default_instance.cpp + device_gemm_wmma_universal_bf16_i4_bf16/device_gemm_wmma_universal_bf16_i4_bf16_km_nk_mn_comp_default_instance.cpp + device_gemm_wmma_universal_f8_f16_f16/device_gemm_wmma_universal_f8_f16_f16_mk_kn_mn_comp_default_instance.cpp device_gemm_wmma_universal_f8_f16_f16/device_gemm_wmma_universal_f8_f16_f16_mk_nk_mn_comp_default_instance.cpp device_gemm_wmma_universal_f8_f16_f16/device_gemm_wmma_universal_f8_f16_f16_km_kn_mn_comp_default_instance.cpp @@ -87,6 +90,9 @@ set_source_files_properties(device_gemm_wmma_universal_bf16_bf16_bf16/device_gem set_source_files_properties(device_gemm_wmma_universal_bf16_bf16_bf16/device_gemm_wmma_universal_bf16_bf16_bf16_km_kn_mn_comp_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1") set_source_files_properties(device_gemm_wmma_universal_bf16_bf16_bf16/device_gemm_wmma_universal_bf16_bf16_bf16_km_nk_mn_comp_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1") +set_source_files_properties(device_gemm_wmma_universal_bf16_i4_bf16/device_gemm_wmma_universal_bf16_i4_bf16_mk_nk_mn_comp_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1") +set_source_files_properties(device_gemm_wmma_universal_bf16_i4_bf16/device_gemm_wmma_universal_bf16_i4_bf16_km_nk_mn_comp_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1") + set_source_files_properties(device_gemm_wmma_universal_f8_f16_f16/device_gemm_wmma_universal_f8_f16_f16_mk_kn_mn_comp_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1") set_source_files_properties(device_gemm_wmma_universal_f8_f16_f16/device_gemm_wmma_universal_f8_f16_f16_mk_nk_mn_comp_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1") set_source_files_properties(device_gemm_wmma_universal_f8_f16_f16/device_gemm_wmma_universal_f8_f16_f16_km_kn_mn_comp_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1") diff --git a/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_bf16_i4_bf16/device_gemm_wmma_universal_bf16_i4_bf16_km_nk_mn.hpp b/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_bf16_i4_bf16/device_gemm_wmma_universal_bf16_i4_bf16_km_nk_mn.hpp new file mode 100644 index 0000000000..e1b85c554d --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_bf16_i4_bf16/device_gemm_wmma_universal_bf16_i4_bf16_km_nk_mn.hpp @@ -0,0 +1,66 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved. + +#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_gemm_wmma_cshuffle_v3.hpp" + +#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +using I4 = pk_i4_t; +using F16 = half_t; +using BF16 = bhalf_t; +using F32 = float; + +using Row = tensor_layout::gemm::RowMajor; +using Col = tensor_layout::gemm::ColumnMajor; + +template +using S = Sequence; + +using PassThrough = element_wise::PassThrough; + +static constexpr auto GemmDefault = GemmSpecialization::Default; +static constexpr auto GemmKPadding = GemmSpecialization::KPadding; +static constexpr auto GemmMNPadding = GemmSpecialization::MNPadding; +static constexpr auto GemmMNKPadding = GemmSpecialization::MNKPadding; + +static constexpr auto Intrawave = BlockGemmPipelineScheduler::Intrawave; +static constexpr auto Interwave = BlockGemmPipelineScheduler::Interwave; + +template +using device_gemm_wmma_universal_bf16_i4_bf16_km_nk_mn_comp_instances = + std::tuple< + // clang-format off + //#########################| ALayout| BLayout| CLayout|AData| BData| CData| AccData| CShuffle| A| B| C| GemmSpec| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MRepeat| NRepeat| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CShuffleBlockTransfer| CShuffleBlockTransfer| BlkGemm| BlkGemm| + //#########################| | | | Type| Type| Type| Type| DataType| Elementwise| Elementwise| Elementwise| | Size| Block| Block| Block| | | Wmma| Wmma| | | ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraN| MRepeat| NRepeat| ClusterLengths| ScalarPerVector| PipeSched| PipelineVer| + //#########################| | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | Lengths_AK0_M_AK1| ArrangeOrder| | | PerVector| PerVector_AK1| | Lengths_BK0_N_BK1| ArrangeOrder| | | PerVector| PerVector_BK1| | PerShuffle| PerShuffle| _MBlock_MPerBlock| _NPerBlock| | | + //#########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | _NBlock_NPerBlock| | | | + DeviceGemm_Wmma_CShuffleV3< Col, Col, Row, BF16, I4, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 32, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Col, Col, Row, BF16, I4, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 64, 64, 8, 8, 16, 16, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Col, Col, Row, BF16, I4, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 64, 32, 32, 8, 8, 16, 16, 4, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Col, Col, Row, BF16, I4, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 32, 16, 16, 32, 8, 8, 16, 16, 1, 1, S<2, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 2>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Col, Col, Row, BF16, I4, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 256, 64, 8, 8, 16, 16, 4, 4, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Col, Col, Row, BF16, I4, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 64, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Col, Col, Row, BF16, I4, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 160, 64, 8, 8, 16, 16, 2, 5, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 64, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Col, Col, Row, BF16, I4, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 128, 32, 8, 8, 16, 16, 4, 4, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Col, Col, Row, BF16, I4, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 256, 64, 64, 8, 8, 16, 16, 8, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Col, Col, Row, BF16, I4, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 64, 256, 64, 8, 8, 16, 16, 2, 8, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Col, Col, Row, BF16, I4, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 64, 80, 64, 8, 8, 16, 16, 1, 5, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 64, 1, 2>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Col, Col, Row, BF16, I4, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 64, 64, 8, 8, 16, 16, 1, 2, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Col, Col, Row, BF16, I4, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 64, 32, 64, 8, 8, 16, 16, 4, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Col, Col, Row, BF16, I4, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 32, 64, 64, 8, 8, 16, 16, 2, 2, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Col, Col, Row, BF16, I4, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 32, 16, 32, 64, 8, 8, 16, 16, 1, 2, S<2, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 2>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Col, Col, Row, BF16, I4, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 32, 16, 16, 64, 8, 8, 16, 16, 1, 1, S<2, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 2>, 8, Intrawave, BlockGemmPipelineVersion::v3> + // clang-format on + >; +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_bf16_i4_bf16/device_gemm_wmma_universal_bf16_i4_bf16_km_nk_mn_comp_default_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_bf16_i4_bf16/device_gemm_wmma_universal_bf16_i4_bf16_km_nk_mn_comp_default_instance.cpp new file mode 100644 index 0000000000..0ab06a49e4 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_bf16_i4_bf16/device_gemm_wmma_universal_bf16_i4_bf16_km_nk_mn_comp_default_instance.cpp @@ -0,0 +1,24 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved. + +#include "device_gemm_wmma_universal_bf16_i4_bf16_km_nk_mn.hpp" +#include "ck/host_utility/device_prop.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +void add_device_gemm_wmma_universal_bf16_i4_bf16_km_nk_mn_comp_default_instances( + std::vector>>& + instances) +{ + add_device_operation_instances( + instances, device_gemm_wmma_universal_bf16_i4_bf16_km_nk_mn_comp_instances{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_bf16_i4_bf16/device_gemm_wmma_universal_bf16_i4_bf16_mk_nk_mn.hpp b/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_bf16_i4_bf16/device_gemm_wmma_universal_bf16_i4_bf16_mk_nk_mn.hpp new file mode 100644 index 0000000000..c13cf6e9bf --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_bf16_i4_bf16/device_gemm_wmma_universal_bf16_i4_bf16_mk_nk_mn.hpp @@ -0,0 +1,66 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved. + +#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_gemm_wmma_cshuffle_v3.hpp" + +#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +using I4 = pk_i4_t; +using F16 = half_t; +using BF16 = bhalf_t; +using F32 = float; + +using Row = tensor_layout::gemm::RowMajor; +using Col = tensor_layout::gemm::ColumnMajor; + +template +using S = Sequence; + +using PassThrough = element_wise::PassThrough; + +static constexpr auto GemmDefault = GemmSpecialization::Default; +static constexpr auto GemmKPadding = GemmSpecialization::KPadding; +static constexpr auto GemmMNPadding = GemmSpecialization::MNPadding; +static constexpr auto GemmMNKPadding = GemmSpecialization::MNKPadding; + +static constexpr auto Intrawave = BlockGemmPipelineScheduler::Intrawave; +static constexpr auto Interwave = BlockGemmPipelineScheduler::Interwave; + +template +using device_gemm_wmma_universal_bf16_i4_bf16_mk_nk_mn_comp_instances = + std::tuple< + // clang-format off + //#########################| ALayout| BLayout| CLayout|AData| BData| CData| AccData| CShuffle| A| B| C| GemmSpec| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MRepeat| NRepeat| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CShuffleBlockTransfer| CShuffleBlockTransfer| BlkGemm| BlkGemm| + //#########################| | | | Type| Type| Type| Type| DataType| Elementwise| Elementwise| Elementwise| | Size| Block| Block| Block| | | Wmma| Wmma| | | ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraN| MRepeat| NRepeat| ClusterLengths| ScalarPerVector| PipeSched| PipelineVer| + //#########################| | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | Lengths_AK0_M_AK1| ArrangeOrder| | | PerVector| PerVector_AK1| | Lengths_BK0_N_BK1| ArrangeOrder| | | PerVector| PerVector_BK1| | PerShuffle| PerShuffle| _MBlock_MPerBlock| _NPerBlock| | | + //#########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | _NBlock_NPerBlock| | | | + DeviceGemm_Wmma_CShuffleV3< Row, Col, Row, BF16, I4, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 32, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Col, Row, BF16, I4, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 64, 64, 8, 8, 16, 16, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Col, Row, BF16, I4, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 64, 32, 32, 8, 8, 16, 16, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Col, Row, BF16, I4, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 32, 16, 16, 32, 8, 8, 16, 16, 1, 1, S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 2>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Col, Row, BF16, I4, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 256, 64, 8, 8, 16, 16, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Col, Row, BF16, I4, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 64, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Col, Row, BF16, I4, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 160, 64, 8, 8, 16, 16, 2, 5, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 64, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Col, Row, BF16, I4, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 128, 32, 8, 8, 16, 16, 4, 4, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Col, Row, BF16, I4, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 256, 64, 64, 8, 8, 16, 16, 8, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Col, Row, BF16, I4, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 64, 256, 64, 8, 8, 16, 16, 2, 8, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Col, Row, BF16, I4, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 64, 80, 64, 8, 8, 16, 16, 1, 5, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 64, 1, 2>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Col, Row, BF16, I4, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 64, 64, 8, 8, 16, 16, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Col, Row, BF16, I4, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 64, 32, 64, 8, 8, 16, 16, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Col, Row, BF16, I4, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 32, 64, 64, 8, 8, 16, 16, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Col, Row, BF16, I4, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 32, 16, 32, 64, 8, 8, 16, 16, 1, 2, S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 2>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Col, Row, BF16, I4, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmSpec, 32, 16, 16, 64, 8, 8, 16, 16, 1, 1, S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 2>, 8, Intrawave, BlockGemmPipelineVersion::v3> + // clang-format on + >; +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_bf16_i4_bf16/device_gemm_wmma_universal_bf16_i4_bf16_mk_nk_mn_comp_default_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_bf16_i4_bf16/device_gemm_wmma_universal_bf16_i4_bf16_mk_nk_mn_comp_default_instance.cpp new file mode 100644 index 0000000000..6d550374f7 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_bf16_i4_bf16/device_gemm_wmma_universal_bf16_i4_bf16_mk_nk_mn_comp_default_instance.cpp @@ -0,0 +1,24 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved. + +#include "device_gemm_wmma_universal_bf16_i4_bf16_mk_nk_mn.hpp" +#include "ck/host_utility/device_prop.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +void add_device_gemm_wmma_universal_bf16_i4_bf16_mk_nk_mn_comp_default_instances( + std::vector>>& + instances) +{ + add_device_operation_instances( + instances, device_gemm_wmma_universal_bf16_i4_bf16_mk_nk_mn_comp_instances{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/test/gemm_universal/test_gemm_universal_wmma_bf16.cpp b/test/gemm_universal/test_gemm_universal_wmma_bf16.cpp index 22376a8599..63eede6e40 100644 --- a/test/gemm_universal/test_gemm_universal_wmma_bf16.cpp +++ b/test/gemm_universal/test_gemm_universal_wmma_bf16.cpp @@ -7,6 +7,7 @@ #include "ck/tensor_operation/gpu/device/tensor_layout.hpp" #include "test_gemm_universal_util.hpp" +using I4 = pk_i4_t; using BF16 = ck::bhalf_t; using F32 = float; @@ -58,6 +59,9 @@ using KernelTypes_MK_KN = ::testing::Types< using KernelTypes_MK_NK = ::testing::Types< // ADataType, BDataType, ComputeDataType, CDataType +#if defined(CK_ENABLE_FP8) + std::tuple< BF16, I4, BF16, F16>, +#endif std::tuple< BF16, BF16, BF16, BF16> >; From 55f160278babab041cb296facc1554a53b94084e Mon Sep 17 00:00:00 2001 From: Anca Hamuraru Date: Mon, 5 May 2025 10:36:59 +0000 Subject: [PATCH 08/11] Added support for F16xI4xF16 to gemm_wmma_universal --- .../gpu/gemm_universal.hpp | 17 +++++ .../gpu/gemm_universal_wmma.inc | 10 +++ .../gpu/gemm_universal/CMakeLists.txt | 6 ++ ...emm_wmma_universal_f16_i4_f16_km_nk_mn.hpp | 65 +++++++++++++++++++ ..._i4_f16_km_nk_mn_comp_default_instance.cpp | 24 +++++++ ...emm_wmma_universal_f16_i4_f16_mk_nk_mn.hpp | 65 +++++++++++++++++++ ..._i4_f16_mk_nk_mn_comp_default_instance.cpp | 24 +++++++ .../test_gemm_universal_wmma_fp16.cpp | 2 + 8 files changed, 213 insertions(+) create mode 100644 library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f16_i4_f16/device_gemm_wmma_universal_f16_i4_f16_km_nk_mn.hpp create mode 100644 library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f16_i4_f16/device_gemm_wmma_universal_f16_i4_f16_km_nk_mn_comp_default_instance.cpp create mode 100644 library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f16_i4_f16/device_gemm_wmma_universal_f16_i4_f16_mk_nk_mn.hpp create mode 100644 library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f16_i4_f16/device_gemm_wmma_universal_f16_i4_f16_mk_nk_mn_comp_default_instance.cpp diff --git a/library/include/ck/library/tensor_operation_instance/gpu/gemm_universal.hpp b/library/include/ck/library/tensor_operation_instance/gpu/gemm_universal.hpp index 6aa6088c17..91b9056249 100644 --- a/library/include/ck/library/tensor_operation_instance/gpu/gemm_universal.hpp +++ b/library/include/ck/library/tensor_operation_instance/gpu/gemm_universal.hpp @@ -196,6 +196,23 @@ struct DeviceOperationInstanceFactory< add_device_gemm_wmma_universal_f16_f8_f16_km_nk_mn_comp_default_instances(op_ptrs); } } + + if constexpr(is_same_v && is_same_v && + is_same_v) + { + if constexpr(is_same_v && is_same_v && + is_same_v) + { + add_device_gemm_wmma_universal_f16_i4_f16_mk_nk_mn_comp_default_instances( + op_ptrs); + } + else if constexpr(is_same_v && is_same_v && + is_same_v) + { + add_device_gemm_wmma_universal_f16_i4_f16_km_nk_mn_comp_default_instances( + op_ptrs); + } + } #endif #endif // CK_USE_WMMA diff --git a/library/include/ck/library/tensor_operation_instance/gpu/gemm_universal_wmma.inc b/library/include/ck/library/tensor_operation_instance/gpu/gemm_universal_wmma.inc index 72f56f5dc0..faf42cec03 100644 --- a/library/include/ck/library/tensor_operation_instance/gpu/gemm_universal_wmma.inc +++ b/library/include/ck/library/tensor_operation_instance/gpu/gemm_universal_wmma.inc @@ -111,6 +111,16 @@ void add_device_gemm_wmma_universal_f16_f8_f16_km_nk_mn_comp_default_instances( std::vector>>& instances); + +void add_device_gemm_wmma_universal_f16_i4_f16_mk_nk_mn_comp_default_instances( + std::vector>>& + instances); + +void add_device_gemm_wmma_universal_f16_i4_f16_km_nk_mn_comp_default_instances( + std::vector>>& + instances); #endif } // namespace instance } // namespace device diff --git a/library/src/tensor_operation_instance/gpu/gemm_universal/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/gemm_universal/CMakeLists.txt index c857949424..5849c6652b 100644 --- a/library/src/tensor_operation_instance/gpu/gemm_universal/CMakeLists.txt +++ b/library/src/tensor_operation_instance/gpu/gemm_universal/CMakeLists.txt @@ -15,6 +15,9 @@ list(APPEND GEMM_UNIVERSAL_INSTANCES device_gemm_wmma_universal_bf16_i4_bf16/device_gemm_wmma_universal_bf16_i4_bf16_mk_nk_mn_comp_default_instance.cpp device_gemm_wmma_universal_bf16_i4_bf16/device_gemm_wmma_universal_bf16_i4_bf16_km_nk_mn_comp_default_instance.cpp + device_gemm_wmma_universal_f16_i4_f16/device_gemm_wmma_universal_f16_i4_f16_mk_nk_mn_comp_default_instance.cpp + device_gemm_wmma_universal_f16_i4_f16/device_gemm_wmma_universal_f16_i4_f16_km_nk_mn_comp_default_instance.cpp + device_gemm_wmma_universal_f8_f16_f16/device_gemm_wmma_universal_f8_f16_f16_mk_kn_mn_comp_default_instance.cpp device_gemm_wmma_universal_f8_f16_f16/device_gemm_wmma_universal_f8_f16_f16_mk_nk_mn_comp_default_instance.cpp device_gemm_wmma_universal_f8_f16_f16/device_gemm_wmma_universal_f8_f16_f16_km_kn_mn_comp_default_instance.cpp @@ -93,6 +96,9 @@ set_source_files_properties(device_gemm_wmma_universal_bf16_bf16_bf16/device_gem set_source_files_properties(device_gemm_wmma_universal_bf16_i4_bf16/device_gemm_wmma_universal_bf16_i4_bf16_mk_nk_mn_comp_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1") set_source_files_properties(device_gemm_wmma_universal_bf16_i4_bf16/device_gemm_wmma_universal_bf16_i4_bf16_km_nk_mn_comp_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1") +set_source_files_properties(device_gemm_wmma_universal_f16_i4_bf16/device_gemm_wmma_universal_f16_i4_bf16_mk_nk_mn_comp_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1") +set_source_files_properties(device_gemm_wmma_universal_f16_i4_bf16/device_gemm_wmma_universal_f16_i4_bf16_km_nk_mn_comp_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1") + set_source_files_properties(device_gemm_wmma_universal_f8_f16_f16/device_gemm_wmma_universal_f8_f16_f16_mk_kn_mn_comp_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1") set_source_files_properties(device_gemm_wmma_universal_f8_f16_f16/device_gemm_wmma_universal_f8_f16_f16_mk_nk_mn_comp_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1") set_source_files_properties(device_gemm_wmma_universal_f8_f16_f16/device_gemm_wmma_universal_f8_f16_f16_km_kn_mn_comp_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1") diff --git a/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f16_i4_f16/device_gemm_wmma_universal_f16_i4_f16_km_nk_mn.hpp b/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f16_i4_f16/device_gemm_wmma_universal_f16_i4_f16_km_nk_mn.hpp new file mode 100644 index 0000000000..f7efa23f31 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f16_i4_f16/device_gemm_wmma_universal_f16_i4_f16_km_nk_mn.hpp @@ -0,0 +1,65 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved. + +#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_gemm_wmma_cshuffle_v3.hpp" + +#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +using I4 = pk_i4_t; +using F16 = half_t; +using F32 = float; + +using Row = tensor_layout::gemm::RowMajor; +using Col = tensor_layout::gemm::ColumnMajor; + +template +using S = Sequence; + +using PassThrough = element_wise::PassThrough; + +static constexpr auto GemmDefault = GemmSpecialization::Default; +static constexpr auto GemmKPadding = GemmSpecialization::KPadding; +static constexpr auto GemmMNPadding = GemmSpecialization::MNPadding; +static constexpr auto GemmMNKPadding = GemmSpecialization::MNKPadding; + +static constexpr auto Intrawave = BlockGemmPipelineScheduler::Intrawave; +static constexpr auto Interwave = BlockGemmPipelineScheduler::Interwave; + +template +using device_gemm_wmma_universal_f16_i4_f16_km_nk_mn_comp_instances = + std::tuple< + // clang-format off + //#########################| ALayout| BLayout| CLayout|AData| BData| CData| AccData| CShuffle| A| B| C| GemmSpec| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MRepeat| NRepeat| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CShuffleBlockTransfer| CShuffleBlockTransfer| BlkGemm| BlkGemm| + //#########################| | | | Type| Type| Type| Type| DataType| Elementwise| Elementwise| Elementwise| | Size| Block| Block| Block| | | Wmma| Wmma| | | ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraN| MRepeat| NRepeat| ClusterLengths| ScalarPerVector| PipeSched| PipelineVer| + //#########################| | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | Lengths_AK0_M_AK1| ArrangeOrder| | | PerVector| PerVector_AK1| | Lengths_BK0_N_BK1| ArrangeOrder| | | PerVector| PerVector_BK1| | PerShuffle| PerShuffle| _MBlock_MPerBlock| _NPerBlock| | | + //#########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | _NBlock_NPerBlock| | | | + DeviceGemm_Wmma_CShuffleV3< Col, Col, Row, F16, I4, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 32, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Col, Col, Row, F16, I4, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 64, 64, 8, 8, 16, 16, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Col, Col, Row, F16, I4, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 64, 32, 32, 8, 8, 16, 16, 4, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Col, Col, Row, F16, I4, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 32, 16, 16, 32, 8, 8, 16, 16, 1, 1, S<2, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 2>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Col, Col, Row, F16, I4, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 256, 64, 8, 8, 16, 16, 4, 4, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Col, Col, Row, F16, I4, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 64, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Col, Col, Row, F16, I4, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 160, 64, 8, 8, 16, 16, 2, 5, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 64, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Col, Col, Row, F16, I4, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 128, 32, 8, 8, 16, 16, 4, 4, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Col, Col, Row, F16, I4, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 256, 64, 64, 8, 8, 16, 16, 8, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Col, Col, Row, F16, I4, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 64, 256, 64, 8, 8, 16, 16, 2, 8, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Col, Col, Row, F16, I4, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 64, 80, 64, 8, 8, 16, 16, 1, 5, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 64, 1, 2>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Col, Col, Row, F16, I4, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 64, 64, 8, 8, 16, 16, 1, 2, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Col, Col, Row, F16, I4, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 64, 32, 64, 8, 8, 16, 16, 4, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Col, Col, Row, F16, I4, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 32, 64, 64, 8, 8, 16, 16, 2, 2, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Col, Col, Row, F16, I4, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 32, 16, 32, 64, 8, 8, 16, 16, 1, 2, S<2, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 2>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Col, Col, Row, F16, I4, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 32, 16, 16, 64, 8, 8, 16, 16, 1, 1, S<2, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 2>, 8, Intrawave, BlockGemmPipelineVersion::v3> + // clang-format on + >; +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f16_i4_f16/device_gemm_wmma_universal_f16_i4_f16_km_nk_mn_comp_default_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f16_i4_f16/device_gemm_wmma_universal_f16_i4_f16_km_nk_mn_comp_default_instance.cpp new file mode 100644 index 0000000000..df6719d605 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f16_i4_f16/device_gemm_wmma_universal_f16_i4_f16_km_nk_mn_comp_default_instance.cpp @@ -0,0 +1,24 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved. + +#include "device_gemm_wmma_universal_f16_i4_f16_km_nk_mn.hpp" +#include "ck/host_utility/device_prop.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +void add_device_gemm_wmma_universal_f16_i4_f16_km_nk_mn_comp_default_instances( + std::vector>>& + instances) +{ + add_device_operation_instances( + instances, device_gemm_wmma_universal_f16_i4_f16_km_nk_mn_comp_instances{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f16_i4_f16/device_gemm_wmma_universal_f16_i4_f16_mk_nk_mn.hpp b/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f16_i4_f16/device_gemm_wmma_universal_f16_i4_f16_mk_nk_mn.hpp new file mode 100644 index 0000000000..8ae2bc510c --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f16_i4_f16/device_gemm_wmma_universal_f16_i4_f16_mk_nk_mn.hpp @@ -0,0 +1,65 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved. + +#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_gemm_wmma_cshuffle_v3.hpp" + +#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +using I4 = pk_i4_t; +using F16 = half_t; +using F32 = float; + +using Row = tensor_layout::gemm::RowMajor; +using Col = tensor_layout::gemm::ColumnMajor; + +template +using S = Sequence; + +using PassThrough = element_wise::PassThrough; + +static constexpr auto GemmDefault = GemmSpecialization::Default; +static constexpr auto GemmKPadding = GemmSpecialization::KPadding; +static constexpr auto GemmMNPadding = GemmSpecialization::MNPadding; +static constexpr auto GemmMNKPadding = GemmSpecialization::MNKPadding; + +static constexpr auto Intrawave = BlockGemmPipelineScheduler::Intrawave; +static constexpr auto Interwave = BlockGemmPipelineScheduler::Interwave; + +template +using device_gemm_wmma_universal_f16_i4_f16_mk_nk_mn_comp_instances = + std::tuple< + // clang-format off + //#########################| ALayout| BLayout| CLayout|AData| BData| CData| AccData| CShuffle| A| B| C| GemmSpec| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MRepeat| NRepeat| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CShuffleBlockTransfer| CShuffleBlockTransfer| BlkGemm| BlkGemm| + //#########################| | | | Type| Type| Type| Type| DataType| Elementwise| Elementwise| Elementwise| | Size| Block| Block| Block| | | Wmma| Wmma| | | ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraN| MRepeat| NRepeat| ClusterLengths| ScalarPerVector| PipeSched| PipelineVer| + //#########################| | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | Lengths_AK0_M_AK1| ArrangeOrder| | | PerVector| PerVector_AK1| | Lengths_BK0_N_BK1| ArrangeOrder| | | PerVector| PerVector_BK1| | PerShuffle| PerShuffle| _MBlock_MPerBlock| _NPerBlock| | | + //#########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | _NBlock_NPerBlock| | | | + DeviceGemm_Wmma_CShuffleV3< Row, Col, Row, F16, I4, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 32, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Col, Row, F16, I4, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 64, 64, 8, 8, 16, 16, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Col, Row, F16, I4, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 64, 32, 32, 8, 8, 16, 16, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Col, Row, F16, I4, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 32, 16, 16, 32, 8, 8, 16, 16, 1, 1, S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 2>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Col, Row, F16, I4, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 256, 64, 8, 8, 16, 16, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Col, Row, F16, I4, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 64, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Col, Row, F16, I4, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 160, 64, 8, 8, 16, 16, 2, 5, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 64, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Col, Row, F16, I4, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 128, 32, 8, 8, 16, 16, 4, 4, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Col, Row, F16, I4, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 256, 64, 64, 8, 8, 16, 16, 8, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Col, Row, F16, I4, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 64, 256, 64, 8, 8, 16, 16, 2, 8, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Col, Row, F16, I4, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 64, 80, 64, 8, 8, 16, 16, 1, 5, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 64, 1, 2>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Col, Row, F16, I4, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 64, 64, 8, 8, 16, 16, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Col, Row, F16, I4, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 64, 32, 64, 8, 8, 16, 16, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Col, Row, F16, I4, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 32, 64, 64, 8, 8, 16, 16, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Col, Row, F16, I4, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 32, 16, 32, 64, 8, 8, 16, 16, 1, 2, S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 2>, 8, Intrawave, BlockGemmPipelineVersion::v3>, + DeviceGemm_Wmma_CShuffleV3< Row, Col, Row, F16, I4, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmSpec, 32, 16, 16, 64, 8, 8, 16, 16, 1, 1, S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<2, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 2>, 8, Intrawave, BlockGemmPipelineVersion::v3> + // clang-format on + >; +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f16_i4_f16/device_gemm_wmma_universal_f16_i4_f16_mk_nk_mn_comp_default_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f16_i4_f16/device_gemm_wmma_universal_f16_i4_f16_mk_nk_mn_comp_default_instance.cpp new file mode 100644 index 0000000000..42c00b4e86 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f16_i4_f16/device_gemm_wmma_universal_f16_i4_f16_mk_nk_mn_comp_default_instance.cpp @@ -0,0 +1,24 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved. + +#include "device_gemm_wmma_universal_f16_i4_f16_mk_nk_mn.hpp" +#include "ck/host_utility/device_prop.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +void add_device_gemm_wmma_universal_f16_i4_f16_mk_nk_mn_comp_default_instances( + std::vector>>& + instances) +{ + add_device_operation_instances( + instances, device_gemm_wmma_universal_f16_i4_f16_mk_nk_mn_comp_instances{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/test/gemm_universal/test_gemm_universal_wmma_fp16.cpp b/test/gemm_universal/test_gemm_universal_wmma_fp16.cpp index ae9be8566c..4e17f7168d 100644 --- a/test/gemm_universal/test_gemm_universal_wmma_fp16.cpp +++ b/test/gemm_universal/test_gemm_universal_wmma_fp16.cpp @@ -7,6 +7,7 @@ #include "ck/tensor_operation/gpu/device/tensor_layout.hpp" #include "test_gemm_universal_util.hpp" +using I4 = ck::pk_i4_t; using F8 = ck::f8_t; using F16 = ck::half_t; @@ -68,6 +69,7 @@ using KernelTypes_KM_NK = ::testing::Types< #if defined(CK_ENABLE_FP8) std::tuple< F8, F16, F16, F16>, std::tuple< F16, F8, F16, F16>, + std::tuple< F16, I4, F16, F16>, #endif std::tuple< F16, F16, F16, F16> >; From d892b5a0e93923d9594cdfe652ab3d4122dd9eb1 Mon Sep 17 00:00:00 2001 From: Anca Hamuraru Date: Thu, 8 May 2025 10:49:47 +0000 Subject: [PATCH 09/11] Fixed IsSupportedArgument to check ComputeTypeA, ComputeTypeB instead of ADataType, BDataType --- .../gpu/device/impl/device_gemm_wmma_cshuffle_v3.hpp | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/include/ck/tensor_operation/gpu/device/impl/device_gemm_wmma_cshuffle_v3.hpp b/include/ck/tensor_operation/gpu/device/impl/device_gemm_wmma_cshuffle_v3.hpp index 1ef8a9b8ad..bbef34316c 100644 --- a/include/ck/tensor_operation/gpu/device/impl/device_gemm_wmma_cshuffle_v3.hpp +++ b/include/ck/tensor_operation/gpu/device/impl/device_gemm_wmma_cshuffle_v3.hpp @@ -405,8 +405,8 @@ struct DeviceGemm_Wmma_CShuffleV3 : public DeviceGemmV2 || std::is_same_v || - std::is_same_v || std::is_same_v) + if constexpr(std::is_same_v || std::is_same_v || + std::is_same_v || std::is_same_v) { if(ck::is_gfx11_supported()) { From 97e024918ee0abfd3f3233590c28ba7e6471643a Mon Sep 17 00:00:00 2001 From: Anca Hamuraru Date: Thu, 8 May 2025 10:50:49 +0000 Subject: [PATCH 10/11] Added missing test class for FP16_KM_NK --- test/gemm_universal/test_gemm_universal_wmma_fp16.cpp | 6 ++++++ 1 file changed, 6 insertions(+) diff --git a/test/gemm_universal/test_gemm_universal_wmma_fp16.cpp b/test/gemm_universal/test_gemm_universal_wmma_fp16.cpp index 4e17f7168d..734dfe72d7 100644 --- a/test/gemm_universal/test_gemm_universal_wmma_fp16.cpp +++ b/test/gemm_universal/test_gemm_universal_wmma_fp16.cpp @@ -47,6 +47,12 @@ class TestGemmUniversal_FP16_KM_KN { }; +template +class TestGemmUniversal_FP16_KM_NK + : public ck::test::TestGemmUniversal, Tuple>::type> +{ +}; + // clang-format off using KernelTypes_MK_KN = ::testing::Types< // ADataType, BDataType, ComputeDataType, CDataType From 744262dd888a4d795425a0f9bc6baa4e7edfdca1 Mon Sep 17 00:00:00 2001 From: Anca Hamuraru Date: Thu, 8 May 2025 13:48:18 +0000 Subject: [PATCH 11/11] Pre-commit hooks fixes --- .../tensor_operation_instance/gpu/gemm_universal.hpp | 10 ++++------ .../device_gemm_wmma_universal_f16_f8_f16_km_kn_mn.hpp | 2 +- .../device_gemm_wmma_universal_f16_f8_f16_km_nk_mn.hpp | 2 +- .../device_gemm_wmma_universal_f16_f8_f16_mk_kn_mn.hpp | 2 +- .../device_gemm_wmma_universal_f16_f8_f16_mk_nk_mn.hpp | 2 +- .../device_gemm_wmma_universal_f8_f16_f16_km_kn_mn.hpp | 2 +- .../device_gemm_wmma_universal_f8_f16_f16_km_nk_mn.hpp | 2 +- .../device_gemm_wmma_universal_f8_f16_f16_mk_kn_mn.hpp | 2 +- .../device_gemm_wmma_universal_f8_f16_f16_mk_nk_mn.hpp | 2 +- test/gemm_universal/test_gemm_universal_wmma_bf16.cpp | 2 +- 10 files changed, 13 insertions(+), 15 deletions(-) diff --git a/library/include/ck/library/tensor_operation_instance/gpu/gemm_universal.hpp b/library/include/ck/library/tensor_operation_instance/gpu/gemm_universal.hpp index 91b9056249..75ab945f99 100644 --- a/library/include/ck/library/tensor_operation_instance/gpu/gemm_universal.hpp +++ b/library/include/ck/library/tensor_operation_instance/gpu/gemm_universal.hpp @@ -133,7 +133,7 @@ struct DeviceOperationInstanceFactory< is_same_v) { if constexpr(is_same_v && is_same_v && - is_same_v) + is_same_v) { add_device_gemm_wmma_universal_bf16_i4_bf16_mk_nk_mn_comp_default_instances( op_ptrs); @@ -201,16 +201,14 @@ struct DeviceOperationInstanceFactory< is_same_v) { if constexpr(is_same_v && is_same_v && - is_same_v) + is_same_v) { - add_device_gemm_wmma_universal_f16_i4_f16_mk_nk_mn_comp_default_instances( - op_ptrs); + add_device_gemm_wmma_universal_f16_i4_f16_mk_nk_mn_comp_default_instances(op_ptrs); } else if constexpr(is_same_v && is_same_v && is_same_v) { - add_device_gemm_wmma_universal_f16_i4_f16_km_nk_mn_comp_default_instances( - op_ptrs); + add_device_gemm_wmma_universal_f16_i4_f16_km_nk_mn_comp_default_instances(op_ptrs); } } #endif diff --git a/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f16_f8_f16/device_gemm_wmma_universal_f16_f8_f16_km_kn_mn.hpp b/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f16_f8_f16/device_gemm_wmma_universal_f16_f8_f16_km_kn_mn.hpp index c98cd429dd..d25f00db12 100644 --- a/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f16_f8_f16/device_gemm_wmma_universal_f16_f8_f16_km_kn_mn.hpp +++ b/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f16_f8_f16/device_gemm_wmma_universal_f16_f8_f16_km_kn_mn.hpp @@ -13,7 +13,7 @@ namespace tensor_operation { namespace device { namespace instance { -using F8 = f8_t; +using F8 = f8_t; using F16 = half_t; using F32 = float; diff --git a/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f16_f8_f16/device_gemm_wmma_universal_f16_f8_f16_km_nk_mn.hpp b/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f16_f8_f16/device_gemm_wmma_universal_f16_f8_f16_km_nk_mn.hpp index dbc0434d9a..ad56d50e43 100644 --- a/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f16_f8_f16/device_gemm_wmma_universal_f16_f8_f16_km_nk_mn.hpp +++ b/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f16_f8_f16/device_gemm_wmma_universal_f16_f8_f16_km_nk_mn.hpp @@ -13,7 +13,7 @@ namespace tensor_operation { namespace device { namespace instance { -using F8 = f8_t; +using F8 = f8_t; using F16 = half_t; using F32 = float; diff --git a/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f16_f8_f16/device_gemm_wmma_universal_f16_f8_f16_mk_kn_mn.hpp b/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f16_f8_f16/device_gemm_wmma_universal_f16_f8_f16_mk_kn_mn.hpp index b976d5dc0d..2feadad794 100644 --- a/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f16_f8_f16/device_gemm_wmma_universal_f16_f8_f16_mk_kn_mn.hpp +++ b/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f16_f8_f16/device_gemm_wmma_universal_f16_f8_f16_mk_kn_mn.hpp @@ -13,7 +13,7 @@ namespace tensor_operation { namespace device { namespace instance { -using F8 = f8_t; +using F8 = f8_t; using F16 = half_t; using F32 = float; diff --git a/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f16_f8_f16/device_gemm_wmma_universal_f16_f8_f16_mk_nk_mn.hpp b/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f16_f8_f16/device_gemm_wmma_universal_f16_f8_f16_mk_nk_mn.hpp index c93204cec9..4fa4560741 100644 --- a/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f16_f8_f16/device_gemm_wmma_universal_f16_f8_f16_mk_nk_mn.hpp +++ b/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f16_f8_f16/device_gemm_wmma_universal_f16_f8_f16_mk_nk_mn.hpp @@ -13,7 +13,7 @@ namespace tensor_operation { namespace device { namespace instance { -using F8 = f8_t; +using F8 = f8_t; using F16 = half_t; using F32 = float; diff --git a/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f8_f16_f16/device_gemm_wmma_universal_f8_f16_f16_km_kn_mn.hpp b/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f8_f16_f16/device_gemm_wmma_universal_f8_f16_f16_km_kn_mn.hpp index 0040ecda52..ea8db0c0d5 100644 --- a/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f8_f16_f16/device_gemm_wmma_universal_f8_f16_f16_km_kn_mn.hpp +++ b/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f8_f16_f16/device_gemm_wmma_universal_f8_f16_f16_km_kn_mn.hpp @@ -13,7 +13,7 @@ namespace tensor_operation { namespace device { namespace instance { -using F8 = f8_t; +using F8 = f8_t; using F16 = half_t; using F32 = float; diff --git a/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f8_f16_f16/device_gemm_wmma_universal_f8_f16_f16_km_nk_mn.hpp b/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f8_f16_f16/device_gemm_wmma_universal_f8_f16_f16_km_nk_mn.hpp index 06f05c721b..1d5972dabf 100644 --- a/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f8_f16_f16/device_gemm_wmma_universal_f8_f16_f16_km_nk_mn.hpp +++ b/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f8_f16_f16/device_gemm_wmma_universal_f8_f16_f16_km_nk_mn.hpp @@ -13,7 +13,7 @@ namespace tensor_operation { namespace device { namespace instance { -using F8 = f8_t; +using F8 = f8_t; using F16 = half_t; using F32 = float; diff --git a/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f8_f16_f16/device_gemm_wmma_universal_f8_f16_f16_mk_kn_mn.hpp b/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f8_f16_f16/device_gemm_wmma_universal_f8_f16_f16_mk_kn_mn.hpp index 20b196834a..c58ee3d4d4 100644 --- a/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f8_f16_f16/device_gemm_wmma_universal_f8_f16_f16_mk_kn_mn.hpp +++ b/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f8_f16_f16/device_gemm_wmma_universal_f8_f16_f16_mk_kn_mn.hpp @@ -13,7 +13,7 @@ namespace tensor_operation { namespace device { namespace instance { -using F8 = f8_t; +using F8 = f8_t; using F16 = half_t; using F32 = float; diff --git a/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f8_f16_f16/device_gemm_wmma_universal_f8_f16_f16_mk_nk_mn.hpp b/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f8_f16_f16/device_gemm_wmma_universal_f8_f16_f16_mk_nk_mn.hpp index da91beb161..a0a16931a9 100644 --- a/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f8_f16_f16/device_gemm_wmma_universal_f8_f16_f16_mk_nk_mn.hpp +++ b/library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_wmma_universal_f8_f16_f16/device_gemm_wmma_universal_f8_f16_f16_mk_nk_mn.hpp @@ -13,7 +13,7 @@ namespace tensor_operation { namespace device { namespace instance { -using F8 = f8_t; +using F8 = f8_t; using F16 = half_t; using F32 = float; diff --git a/test/gemm_universal/test_gemm_universal_wmma_bf16.cpp b/test/gemm_universal/test_gemm_universal_wmma_bf16.cpp index 63eede6e40..83d7eab1ad 100644 --- a/test/gemm_universal/test_gemm_universal_wmma_bf16.cpp +++ b/test/gemm_universal/test_gemm_universal_wmma_bf16.cpp @@ -7,7 +7,7 @@ #include "ck/tensor_operation/gpu/device/tensor_layout.hpp" #include "test_gemm_universal_util.hpp" -using I4 = pk_i4_t; +using I4 = pk_i4_t; using BF16 = ck::bhalf_t; using F32 = float;