diff --git a/include/ck/tensor_operation/gpu/device/device_gemm_splitk.hpp b/include/ck/tensor_operation/gpu/device/device_gemm_splitk.hpp index 6407aa7e09..dfc27e726e 100644 --- a/include/ck/tensor_operation/gpu/device/device_gemm_splitk.hpp +++ b/include/ck/tensor_operation/gpu/device/device_gemm_splitk.hpp @@ -20,7 +20,8 @@ template + typename CElementwiseOperation, + typename ComputeType = CDataType> struct DeviceGemmSplitK : public BaseOperator { virtual std::unique_ptr MakeArgumentPointer(const void* p_a, @@ -48,7 +49,8 @@ template + typename CElementwiseOperation, + typename ComputeType = CDataType> using DeviceGemmSplitKPtr = std::unique_ptr>; + CElementwiseOperation, + ComputeType>>; } // namespace device } // namespace tensor_operation diff --git a/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_splitk_c_shuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_splitk_c_shuffle.hpp index e7365daed2..d20b008b88 100644 --- a/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_splitk_c_shuffle.hpp +++ b/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_splitk_c_shuffle.hpp @@ -69,7 +69,8 @@ struct DeviceGemmXdlSplitKCShuffle : public DeviceGemmSplitK + CElementwiseOperation, + ComputeType> { static constexpr auto I0 = Number<0>{}; static constexpr auto I1 = Number<1>{}; diff --git a/library/include/ck/library/tensor_operation_instance/gpu/gemm_splitk.hpp b/library/include/ck/library/tensor_operation_instance/gpu/gemm_splitk.hpp index 4cba0875d9..e14df2d750 100644 --- a/library/include/ck/library/tensor_operation_instance/gpu/gemm_splitk.hpp +++ b/library/include/ck/library/tensor_operation_instance/gpu/gemm_splitk.hpp @@ -98,6 +98,26 @@ void add_device_gemm_xdl_splitk_f16_f8_f16_mk_nk_mn_instances( std::vector>>& instances); + +void add_device_gemm_xdl_splitk_f16_f16_f16_comp_f8_km_kn_mn_instances( + std::vector>>& + instances); + +void add_device_gemm_xdl_splitk_f16_f16_f16_comp_f8_km_nk_mn_instances( + std::vector>>& + instances); + +void add_device_gemm_xdl_splitk_f16_f16_f16_comp_f8_mk_kn_mn_instances( + std::vector>>& + instances); + +void add_device_gemm_xdl_splitk_f16_f16_f16_comp_f8_mk_nk_mn_instances( + std::vector>>& + instances); #endif template + typename CLayout, + typename ComputeType> struct DeviceOperationInstanceFactory< ck::tensor_operation::device::DeviceGemmSplitK> + ck::tensor_operation::element_wise::PassThrough, + ComputeType>> { using DeviceOp = DeviceGemmSplitK; + ck::tensor_operation::element_wise::PassThrough, + ComputeType>; static auto GetInstances() { @@ -158,7 +181,7 @@ struct DeviceOperationInstanceFactory< #endif #ifdef CK_ENABLE_FP16 else if constexpr(is_same_v && is_same_v && - is_same_v) + is_same_v && is_same_v) { if constexpr(is_same_v && is_same_v && is_same_v) @@ -231,6 +254,31 @@ struct DeviceOperationInstanceFactory< add_device_gemm_xdl_splitk_f16_f8_f16_km_nk_mn_instances(op_ptrs); } } + else if constexpr(is_same_v && is_same_v && + is_same_v && is_same_v) + { + if constexpr(is_same_v && is_same_v && + is_same_v) + { + add_device_gemm_xdl_splitk_f16_f16_f16_comp_f8_mk_kn_mn_instances(op_ptrs); + } + else if constexpr(is_same_v && is_same_v && + is_same_v) + { + add_device_gemm_xdl_splitk_f16_f16_f16_comp_f8_mk_nk_mn_instances(op_ptrs); + } + else if constexpr(is_same_v && is_same_v && + is_same_v) + { + add_device_gemm_xdl_splitk_f16_f16_f16_comp_f8_km_kn_mn_instances(op_ptrs); + } + else if constexpr(is_same_v && is_same_v && + is_same_v) + { + add_device_gemm_xdl_splitk_f16_f16_f16_comp_f8_km_nk_mn_instances(op_ptrs); + } + } + #endif return op_ptrs; } diff --git a/library/src/tensor_operation_instance/gpu/gemm_splitk/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/gemm_splitk/CMakeLists.txt index 36c0b50e8e..ec4c27598f 100644 --- a/library/src/tensor_operation_instance/gpu/gemm_splitk/CMakeLists.txt +++ b/library/src/tensor_operation_instance/gpu/gemm_splitk/CMakeLists.txt @@ -15,6 +15,10 @@ list(APPEND GEMM_SPLITK_INSTANCES device_gemm_xdl_splitk_f32_f32_f32_mk_kn_mn_in device_gemm_xdl_splitk_f16_fp8_f16_mk_kn_mn_instance.cpp device_gemm_xdl_splitk_f16_fp8_f16_mk_nk_mn_instance.cpp device_gemm_xdl_splitk_f16_fp8_f16_km_kn_mn_instance.cpp - device_gemm_xdl_splitk_f16_fp8_f16_km_nk_mn_instance.cpp) + device_gemm_xdl_splitk_f16_fp8_f16_km_nk_mn_instance.cpp + device_gemm_xdl_splitk_f16_f16_f16_comp_fp8_mk_kn_mn_instance.cpp + device_gemm_xdl_splitk_f16_f16_f16_comp_fp8_mk_nk_mn_instance.cpp + device_gemm_xdl_splitk_f16_f16_f16_comp_fp8_km_kn_mn_instance.cpp + device_gemm_xdl_splitk_f16_f16_f16_comp_fp8_km_nk_mn_instance.cpp) add_instance_library(device_gemm_splitk_instance ${GEMM_SPLITK_INSTANCES}) diff --git a/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_comp_fp8_km_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_comp_fp8_km_kn_mn_instance.cpp new file mode 100644 index 0000000000..f27e3197bf --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_comp_fp8_km_kn_mn_instance.cpp @@ -0,0 +1,62 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved. + +#include + +#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_xdl_splitk_c_shuffle.hpp" + +#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +using F8 = ck::f8_t; +using F16 = ck::half_t; +using F32 = float; + +using Row = ck::tensor_layout::gemm::RowMajor; +using Col = ck::tensor_layout::gemm::ColumnMajor; + +template +using S = ck::Sequence; + +using PassThrough = ck::tensor_operation::element_wise::PassThrough; + +static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default; + +// Compilation parameters for a[m, k] * b[k, n] = c[m, n] +using device_gemm_xdl_splitk_f16_f16_f16_comp_f8_km_kn_mn_instances = std::tuple< + // clang-format off + //#########################|AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| GEMM| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Compute| + //#########################| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Type| + //#########################| | | | | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| | + //#########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | + DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8, F8>, + DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, true, 1, 1, S<1, 32, 1, 8>, 8, F8>, + DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, true, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, true, 1, 1, S<1, 16, 1, 8>, 8, F8>, + DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8, F8>, + DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, true, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, true, 1, 1, S<1, 32, 1, 4>, 8, F8>, + DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 64, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, true, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, true, 1, 1, S<1, 16, 1, 8>, 8, F8>, + DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 64, 4, 8, 32, 32, 2, 1, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 1, 8, true, 1, 1, S<1, 16, 1, 4>, 8, F8>, + DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 64, 128, 4, 8, 32, 32, 1, 2, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 1, 8, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8, F8> + // clang-format on + >; + +void add_device_gemm_xdl_splitk_f16_f16_f16_comp_f8_km_kn_mn_instances( + std::vector>>& + instances) +{ + add_device_operation_instances(instances, + device_gemm_xdl_splitk_f16_f16_f16_comp_f8_km_kn_mn_instances{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_comp_fp8_km_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_comp_fp8_km_nk_mn_instance.cpp new file mode 100644 index 0000000000..421f90e8c5 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_comp_fp8_km_nk_mn_instance.cpp @@ -0,0 +1,62 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved. + +#include + +#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_xdl_splitk_c_shuffle.hpp" + +#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +using F8 = ck::f8_t; +using F16 = ck::half_t; +using F32 = float; + +using Row = ck::tensor_layout::gemm::RowMajor; +using Col = ck::tensor_layout::gemm::ColumnMajor; + +template +using S = ck::Sequence; + +using PassThrough = ck::tensor_operation::element_wise::PassThrough; + +static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default; + +// Compilation parameters for a[m, k] * b[k, n] = c[m, n] +using device_gemm_xdl_splitk_f16_f16_f16_comp_f8_km_nk_mn_instances = std::tuple< + // clang-format off + //#########################|AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| GEMM| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Compute| + //#########################| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Type| + //#########################| | | | | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| | + //#########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | + DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, true, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, F8>, + DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, true, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, F8>, + DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, true, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8, F8>, + DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, true, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, F8>, + DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, true, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, 1, 1, S<1, 32, 1, 4>, 8, F8>, + DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 64, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, true, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8, F8>, + DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 64, 4, 8, 32, 32, 2, 1, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, true, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, 1, 1, S<1, 16, 1, 4>, 8, F8>, + DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 64, 128, 4, 8, 32, 32, 1, 2, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 1, 8, true, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, F8> + // clang-format on + >; + +void add_device_gemm_xdl_splitk_f16_f16_f16_comp_f8_km_nk_mn_instances( + std::vector>>& + instances) +{ + add_device_operation_instances(instances, + device_gemm_xdl_splitk_f16_f16_f16_comp_f8_km_nk_mn_instances{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_comp_fp8_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_comp_fp8_mk_kn_mn_instance.cpp new file mode 100644 index 0000000000..814a9aabf2 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_comp_fp8_mk_kn_mn_instance.cpp @@ -0,0 +1,83 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved. + +#include + +#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_xdl_splitk_c_shuffle.hpp" + +#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +using F8 = ck::f8_t; +using F16 = ck::half_t; +using F32 = float; + +using Row = ck::tensor_layout::gemm::RowMajor; +using Col = ck::tensor_layout::gemm::ColumnMajor; + +template +using S = ck::Sequence; + +using PassThrough = ck::tensor_operation::element_wise::PassThrough; + +// static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default; +static constexpr auto GemmMNPadding = ck::tensor_operation::device::GemmSpecialization::MNPadding; + +using device_gemm_xdl_splitk_f16_f16_f16_comp_f8_mk_kn_mn_generic_instances = std::tuple< + // clang-format off + //#########################|AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| GEMM| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Compute| + //#########################| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Type| + //#########################| | | | | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| | + //#########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | + DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmMNPadding, 128, 32, 64, 4, 8, 32, 32, 1, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 1, 8, true, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 1, 8, true, 1, 1, S<1, 16, 1, 8>, 2, F8> + // clang-format on + >; + +// Compilation parameters for a[m, k] * b[k, n] = c[m, n] +using device_gemm_xdl_splitk_f16_f16_f16_comp_f8_mk_kn_mn_instances = std::tuple< + // clang-format off + //#########################|AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| GEMM| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Compute| + //#########################| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Type| + //#########################| | | | | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| | + //#########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | + DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmMNPadding, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8, F8>, + DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmMNPadding, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, true, 1, 1, S<1, 32, 1, 8>, 8, F8>, + DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmMNPadding, 128, 128, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, true, 1, 1, S<1, 16, 1, 8>, 8, F8>, + DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmMNPadding, 256, 64, 192, 4, 8, 32, 32, 1, 3, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 48, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8, F8>, + DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmMNPadding, 256, 192, 64, 4, 8, 32, 32, 3, 1, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8, F8>, + DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmMNPadding, 256, 128, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8, F8>, + DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmMNPadding, 128, 128, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, true, 1, 1, S<1, 32, 1, 4>, 8, F8>, + DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmMNPadding, 128, 64, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, true, 1, 1, S<1, 16, 1, 8>, 8, F8>, + DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmMNPadding, 256, 128, 64, 4, 8, 32, 32, 2, 1, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 1, 8, true, 1, 1, S<1, 16, 1, 4>, 8, F8>, + DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmMNPadding, 256, 64, 128, 4, 8, 32, 32, 1, 2, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8, F8>, + DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmMNPadding, 128, 32, 192, 4, 8, 32, 32, 1, 3, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 24, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8, F8>, + DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmMNPadding, 128, 192, 32, 4, 8, 32, 32, 3, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 1, 8, true, 1, 1, S<1, 32, 1, 4>, 8, F8>, + DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmMNPadding, 128, 32, 64, 4, 8, 32, 32, 1, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, true, 1, 1, S<1, 16, 1, 8>, 8, F8>, + DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmMNPadding, 128, 64, 32, 4, 8, 32, 32, 1, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 1, 8, true, 1, 1, S<1, 32, 1, 4>, 8, F8>, + DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmMNPadding, 128, 32, 128, 4, 8, 32, 32, 1, 2, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, true, 1, 1, S<1, 16, 1, 8>, 8, F8>, + DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmMNPadding, 128, 128, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 1, 8, true, 1, 1, S<1, 32, 1, 4>, 8, F8> + // clang-format on + >; + +void add_device_gemm_xdl_splitk_f16_f16_f16_comp_f8_mk_kn_mn_instances( + std::vector>>& + instances) +{ + add_device_operation_instances( + instances, device_gemm_xdl_splitk_f16_f16_f16_comp_f8_mk_kn_mn_generic_instances{}); + add_device_operation_instances(instances, + device_gemm_xdl_splitk_f16_f16_f16_comp_f8_mk_kn_mn_instances{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_comp_fp8_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_comp_fp8_mk_nk_mn_instance.cpp new file mode 100644 index 0000000000..05803cfbec --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_comp_fp8_mk_nk_mn_instance.cpp @@ -0,0 +1,79 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved. + +#include + +#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_xdl_splitk_c_shuffle.hpp" + +#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +using F8 = ck::f8_t; +using F16 = ck::half_t; +using F32 = float; + +using Row = ck::tensor_layout::gemm::RowMajor; +using Col = ck::tensor_layout::gemm::ColumnMajor; + +template +using S = ck::Sequence; + +using PassThrough = ck::tensor_operation::element_wise::PassThrough; + +static constexpr auto MNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding; + +using device_gemm_xdl_splitk_f16_f16_f16_comp_f8_mk_nk_mn_generic_instances = std::tuple< + // clang-format off + //#########################|AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| GEMM| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Compute| + //#########################| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Type| + //#########################| | | | | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| | + //#########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | + DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, MNKPadding, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<1, 4, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 1, 8, true, S<1, 4, 16, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 3, 1, 8, true, 1, 1, S<1, 16, 1, 4>, 2, F8> + // clang-format on + >; + +// Compilation parameters for a[m, k] * b[k, n] = c[m, n] +using device_gemm_xdl_splitk_f16_f16_f16_comp_f8_mk_nk_mn_instances = std::tuple< + // clang-format off + //#########################|AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| GEMM| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Compute| + //#########################| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Type| + //#########################| | | | | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| | + //#########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | + DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, MNKPadding, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 3, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, F8>, + DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, MNKPadding, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 3, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, F8>, + DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, MNKPadding, 128, 128, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 3, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8, F8>, + DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, MNKPadding, 256, 128, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 3, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, F8>, + DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, MNKPadding, 128, 128, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 3, 8, 8, true, 1, 1, S<1, 32, 1, 4>, 8, F8>, + DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, MNKPadding, 128, 64, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 3, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8, F8>, + DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, MNKPadding, 64, 64, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 16, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 3, 8, 8, true, 1, 1, S<1, 16, 1, 4>, 8, F8>, + DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, MNKPadding, 256, 128, 64, 4, 8, 32, 32, 2, 1, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 3, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, F8>, + DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, MNKPadding, 256, 64, 128, 4, 8, 32, 32, 1, 2, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 3, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, F8>, + DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, MNKPadding, 128, 128, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 3, 8, 8, true, 1, 1, S<1, 32, 1, 4>, 8, F8>, + DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, MNKPadding, 128, 32, 128, 4, 8, 32, 32, 1, 2, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 3, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8, F8>, + DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, MNKPadding, 64, 64, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 16, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 3, 8, 8, true, 1, 1, S<1, 16, 1, 4>, 8, F8>, + DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, MNKPadding, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<1, 4, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 16, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 3, 8, 8, true, 1, 1, S<1, 16, 1, 4>, 8, F8> + // clang-format on + >; + +void add_device_gemm_xdl_splitk_f16_f16_f16_comp_f8_mk_nk_mn_instances( + std::vector>>& + instances) +{ + add_device_operation_instances( + instances, device_gemm_xdl_splitk_f16_f16_f16_comp_f8_mk_nk_mn_generic_instances{}); + add_device_operation_instances(instances, + device_gemm_xdl_splitk_f16_f16_f16_comp_f8_mk_nk_mn_instances{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/profiler/include/profiler/profile_gemm_splitk_impl.hpp b/profiler/include/profiler/profile_gemm_splitk_impl.hpp index 495513f665..329a7c07ca 100644 --- a/profiler/include/profiler/profile_gemm_splitk_impl.hpp +++ b/profiler/include/profiler/profile_gemm_splitk_impl.hpp @@ -30,7 +30,8 @@ template + typename CLayout, + typename ComputeType = CDataType> bool profile_gemm_splitk_impl(int do_verification, int init_method, bool do_log, @@ -103,7 +104,8 @@ bool profile_gemm_splitk_impl(int do_verification, CDataType, AElementOp, BElementOp, - CElementOp>; + CElementOp, + ComputeType>; // get device op instances const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory< @@ -120,7 +122,8 @@ bool profile_gemm_splitk_impl(int do_verification, AccDataType, AElementOp, BElementOp, - CElementOp>; + CElementOp, + ComputeType>; auto ref_gemm = ReferenceGemmInstance{}; auto ref_invoker = ref_gemm.MakeInvoker(); diff --git a/profiler/src/profile_gemm_splitk.cpp b/profiler/src/profile_gemm_splitk.cpp index 9c805fc1d1..d8fa3d872a 100644 --- a/profiler/src/profile_gemm_splitk.cpp +++ b/profiler/src/profile_gemm_splitk.cpp @@ -25,6 +25,7 @@ enum struct GemmDataType INT8_INT8_INT8, // 3 F8_F16_F16, // 4 F16_F8_F16, // 5 + F16_F16_F16_F8, // 6 }; #define OP_NAME "gemm_splitk" @@ -35,7 +36,8 @@ int profile_gemm_splitk(int argc, char* argv[]) if(argc != 15) { printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n"); - printf("arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8; 4: f8@f16; 5: f16@f8)\n"); + printf("arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8; 4: f8@f16; 5: f16@f8; 6: f16, " + "comp f8)\n"); printf("arg3: matrix layout (0: A[m, k] * B[k, n] = C[m, n];\n"); printf(" 1: A[m, k] * B[n, k] = C[m, n];\n"); printf(" 2: A[k, m] * B[k, n] = C[m, n];\n"); @@ -80,7 +82,8 @@ int profile_gemm_splitk(int argc, char* argv[]) auto c_type, auto a_layout, auto b_layout, - auto c_layout) { + auto c_layout, + auto compute_type) { using ADataType = decltype(a_type); using BDataType = decltype(b_type); using AccDataType = decltype(acc_type); @@ -90,6 +93,8 @@ int profile_gemm_splitk(int argc, char* argv[]) using BLayout = decltype(b_layout); using CLayout = decltype(c_layout); + using ComputeType = decltype(compute_type); + const int DefaultStrideA = ck::is_same_v ? K : M; const int DefaultStrideB = ck::is_same_v ? N : K; const int DefaultStrideC = ck::is_same_v ? N : M; @@ -100,7 +105,8 @@ int profile_gemm_splitk(int argc, char* argv[]) CDataType, ALayout, BLayout, - CLayout>( + CLayout, + ComputeType>( do_verification, init_method, do_log, @@ -118,68 +124,84 @@ int profile_gemm_splitk(int argc, char* argv[]) if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::MK_KN_MN) { - return profile(F32{}, F32{}, F32{}, F32{}, Row{}, Row{}, Row{}); + return profile(F32{}, F32{}, F32{}, F32{}, Row{}, Row{}, Row{}, F32{}); } else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::MK_NK_MN) { - return profile(F32{}, F32{}, F32{}, F32{}, Row{}, Col{}, Row{}); + return profile(F32{}, F32{}, F32{}, F32{}, Row{}, Col{}, Row{}, F32{}); } else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::KM_KN_MN) { - return profile(F32{}, F32{}, F32{}, F32{}, Col{}, Row{}, Row{}); + return profile(F32{}, F32{}, F32{}, F32{}, Col{}, Row{}, Row{}, F32{}); } else if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::KM_NK_MN) { - return profile(F32{}, F32{}, F32{}, F32{}, Col{}, Col{}, Row{}); + return profile(F32{}, F32{}, F32{}, F32{}, Col{}, Col{}, Row{}, F32{}); } else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_KN_MN) { - return profile(F16{}, F16{}, F32{}, F16{}, Row{}, Row{}, Row{}); + return profile(F16{}, F16{}, F32{}, F16{}, Row{}, Row{}, Row{}, F16{}); } else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_NK_MN) { - return profile(F16{}, F16{}, F32{}, F16{}, Row{}, Col{}, Row{}); + return profile(F16{}, F16{}, F32{}, F16{}, Row{}, Col{}, Row{}, F16{}); } else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::KM_KN_MN) { - return profile(F16{}, F16{}, F32{}, F16{}, Col{}, Row{}, Row{}); + return profile(F16{}, F16{}, F32{}, F16{}, Col{}, Row{}, Row{}, F16{}); } else if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::KM_NK_MN) { - return profile(F16{}, F16{}, F32{}, F16{}, Col{}, Col{}, Row{}); + return profile(F16{}, F16{}, F32{}, F16{}, Col{}, Col{}, Row{}, F16{}); } #if defined CK_ENABLE_FP8 else if(data_type == GemmDataType::F8_F16_F16 && layout == GemmMatrixLayout::MK_KN_MN) { - return profile(F8{}, F16{}, F32{}, F16{}, Row{}, Row{}, Row{}); + return profile(F8{}, F16{}, F32{}, F16{}, Row{}, Row{}, Row{}, F16{}); } else if(data_type == GemmDataType::F8_F16_F16 && layout == GemmMatrixLayout::MK_NK_MN) { - return profile(F8{}, F16{}, F32{}, F16{}, Row{}, Col{}, Row{}); + return profile(F8{}, F16{}, F32{}, F16{}, Row{}, Col{}, Row{}, F16{}); } else if(data_type == GemmDataType::F8_F16_F16 && layout == GemmMatrixLayout::KM_KN_MN) { - return profile(F8{}, F16{}, F32{}, F16{}, Col{}, Row{}, Row{}); + return profile(F8{}, F16{}, F32{}, F16{}, Col{}, Row{}, Row{}, F16{}); } else if(data_type == GemmDataType::F8_F16_F16 && layout == GemmMatrixLayout::KM_NK_MN) { - return profile(F8{}, F16{}, F32{}, F16{}, Col{}, Col{}, Row{}); + return profile(F8{}, F16{}, F32{}, F16{}, Col{}, Col{}, Row{}, F16{}); } else if(data_type == GemmDataType::F16_F8_F16 && layout == GemmMatrixLayout::MK_KN_MN) { - return profile(F16{}, F8{}, F32{}, F16{}, Row{}, Row{}, Row{}); + return profile(F16{}, F8{}, F32{}, F16{}, Row{}, Row{}, Row{}, F16{}); } else if(data_type == GemmDataType::F16_F8_F16 && layout == GemmMatrixLayout::MK_NK_MN) { - return profile(F16{}, F8{}, F32{}, F16{}, Row{}, Col{}, Row{}); + return profile(F16{}, F8{}, F32{}, F16{}, Row{}, Col{}, Row{}, F16{}); } else if(data_type == GemmDataType::F16_F8_F16 && layout == GemmMatrixLayout::KM_KN_MN) { - return profile(F16{}, F8{}, F32{}, F16{}, Col{}, Row{}, Row{}); + return profile(F16{}, F8{}, F32{}, F16{}, Col{}, Row{}, Row{}, F16{}); } else if(data_type == GemmDataType::F16_F8_F16 && layout == GemmMatrixLayout::KM_NK_MN) { - return profile(F16{}, F8{}, F32{}, F16{}, Col{}, Col{}, Row{}); + return profile(F16{}, F8{}, F32{}, F16{}, Col{}, Col{}, Row{}, F16{}); + } + else if(data_type == GemmDataType::F16_F16_F16_F8 && layout == GemmMatrixLayout::MK_KN_MN) + { + return profile(F16{}, F16{}, F32{}, F16{}, Row{}, Row{}, Row{}, F8{}); + } + else if(data_type == GemmDataType::F16_F16_F16_F8 && layout == GemmMatrixLayout::MK_NK_MN) + { + return profile(F16{}, F16{}, F32{}, F16{}, Row{}, Col{}, Row{}, F8{}); + } + else if(data_type == GemmDataType::F16_F16_F16_F8 && layout == GemmMatrixLayout::KM_KN_MN) + { + return profile(F16{}, F16{}, F32{}, F16{}, Col{}, Row{}, Row{}, F8{}); + } + else if(data_type == GemmDataType::F16_F16_F16_F8 && layout == GemmMatrixLayout::KM_NK_MN) + { + return profile(F16{}, F16{}, F32{}, F16{}, Col{}, Col{}, Row{}, F8{}); } #endif else