From c5d99d22959bc21901d93205d37110449b869caf Mon Sep 17 00:00:00 2001 From: Anca Hamuraru Date: Mon, 5 May 2025 09:55:29 +0000 Subject: [PATCH] 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> >;