diff --git a/include/ck/tensor_operation/gpu/device/device_gemm_xdl_splitk.hpp b/include/ck/tensor_operation/gpu/device/device_gemm_xdl_splitk.hpp deleted file mode 100644 index 62832c3a71..0000000000 --- a/include/ck/tensor_operation/gpu/device/device_gemm_xdl_splitk.hpp +++ /dev/null @@ -1,646 +0,0 @@ -// SPDX-License-Identifier: MIT -// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved. - -#pragma once - -#include -#include - -#include "ck/utility/common_header.hpp" -#include "ck/tensor_description/tensor_descriptor.hpp" -#include "ck/tensor_description/tensor_descriptor_helper.hpp" -#include "ck/tensor_operation/gpu/device/tensor_layout.hpp" -#include "ck/tensor_operation/gpu/device/device_gemm_splitk.hpp" -#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp" -#include "ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r4.hpp" -#include "ck/host_utility/device_prop.hpp" -#include "ck/host_utility/kernel_launch.hpp" - -namespace ck { -namespace tensor_operation { -namespace device { - -template -struct DeviceGemmXdlSplitK : public DeviceGemmSplitK -{ - static constexpr auto I0 = Number<0>{}; - static constexpr auto I1 = Number<1>{}; - static constexpr auto I2 = Number<2>{}; - static constexpr auto I3 = Number<3>{}; - - static constexpr auto K1Number = Number{}; - - static auto - MakeAGridDescriptor_KBatch_K0_M_K1(index_t M, index_t K, index_t StrideA, int KBatch, int KPad) - { - assert(KPad % (K1 * KBatch) == 0); - - const index_t K0 = KPad / (K1 * KBatch); - - const auto a_grid_desc_m_k = [&]() { - if constexpr(is_same::value) - { - return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(StrideA, I1)); - } - else if constexpr(is_same::value) - { - return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(I1, StrideA)); - } - }(); - - const auto a_grid_desc_m_kpad = transform_tensor_descriptor( - a_grid_desc_m_k, - make_tuple(make_pass_through_transform(M), make_right_pad_transform(K, KPad - K)), - make_tuple(Sequence<0>{}, Sequence<1>{}), - make_tuple(Sequence<1>{}, Sequence<0>{})); - - if constexpr(GemmSpec == GemmSpecialization::MNPadding) - { - const auto PadM = (MPerBlock - M % MPerBlock) % MPerBlock; - return transform_tensor_descriptor( - a_grid_desc_m_kpad, - make_tuple(make_unmerge_transform(make_tuple(KBatch, K0, K1Number)), - make_right_pad_transform(M, PadM)), - make_tuple(Sequence<1>{}, Sequence<0>{}), - make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{})); - } - else - { - return transform_tensor_descriptor( - a_grid_desc_m_kpad, - make_tuple(make_unmerge_transform(make_tuple(KBatch, K0, K1Number)), - make_pass_through_transform(M)), - make_tuple(Sequence<1>{}, Sequence<0>{}), - make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{})); - } - } - - static auto - MakeBGridDescriptor_KBatch_K0_N_K1(index_t K, index_t N, index_t StrideB, int KBatch, int KPad) - { - assert(KPad % (K1 * KBatch) == 0); - - const index_t K0 = KPad / (K1 * KBatch); - - const auto b_grid_desc_k_n = [&]() { - if constexpr(is_same::value) - { - return make_naive_tensor_descriptor(make_tuple(K, N), make_tuple(StrideB, I1)); - } - else if constexpr(is_same::value) - { - return make_naive_tensor_descriptor(make_tuple(K, N), make_tuple(I1, StrideB)); - } - }(); - - const auto b_grid_desc_kpad_n = transform_tensor_descriptor( - b_grid_desc_k_n, - make_tuple(make_right_pad_transform(K, KPad - K), make_pass_through_transform(N)), - make_tuple(Sequence<0>{}, Sequence<1>{}), - make_tuple(Sequence<0>{}, Sequence<1>{})); - - if constexpr(GemmSpec == GemmSpecialization::MNPadding) - { - const auto PadN = (NPerBlock - N % NPerBlock) % NPerBlock; - return transform_tensor_descriptor( - b_grid_desc_kpad_n, - make_tuple(make_unmerge_transform(make_tuple(KBatch, K0, K1Number)), - make_right_pad_transform(N, PadN)), - make_tuple(Sequence<0>{}, Sequence<1>{}), - make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{})); - } - else - { - return transform_tensor_descriptor( - b_grid_desc_kpad_n, - make_tuple(make_unmerge_transform(make_tuple(KBatch, K0, K1Number)), - make_pass_through_transform(N)), - make_tuple(Sequence<0>{}, Sequence<1>{}), - make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{})); - } - } - - static auto MakeCGridDescriptor_M_N(index_t M, index_t N, index_t StrideC) - { - const auto c_grid_desc_m_n = [&]() { - if constexpr(is_same::value) - { - return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(StrideC, I1)); - } - else if constexpr(is_same::value) - { - return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(I1, StrideC)); - } - }(); - - if constexpr(GemmSpec == GemmSpecialization::MNPadding) - { - const auto PadM = (MPerBlock - M % MPerBlock) % MPerBlock; - const auto PadN = (NPerBlock - N % NPerBlock) % NPerBlock; - - return transform_tensor_descriptor( - c_grid_desc_m_n, - make_tuple(make_right_pad_transform(M, PadM), make_right_pad_transform(N, PadN)), - make_tuple(Sequence<0>{}, Sequence<1>{}), - make_tuple(Sequence<0>{}, Sequence<1>{})); - } - else - { - - return transform_tensor_descriptor( - c_grid_desc_m_n, - make_tuple(make_pass_through_transform(M), make_pass_through_transform(N)), - make_tuple(Sequence<0>{}, Sequence<1>{}), - make_tuple(Sequence<0>{}, Sequence<1>{})); - } - } - - static auto GetKPad(index_t K, index_t KBatch) - { - const index_t K0 = math::integer_divide_ceil(K, K1 * K0PerBlock * KBatch) * K0PerBlock; - const index_t KPad = KBatch * K0 * K1; - return KPad; - } - - using AGridDesc_K0_M_K1 = decltype(MakeAGridDescriptor_KBatch_K0_M_K1(1, 1, 1, 1, 1)); - using BGridDesc_K0_N_K1 = decltype(MakeBGridDescriptor_KBatch_K0_N_K1(1, 1, 1, 1, 1)); - using CGridDesc_M_N = decltype(MakeCGridDescriptor_M_N(1, 1, 1)); - - // GridwiseGemm - using GridwiseGemm = GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4< - BlockSize, - ADataType, // TODO: distinguish A/B datatype - AccDataType, - CDataType, - InMemoryDataOperationEnum::Set, - AGridDesc_K0_M_K1, - BGridDesc_K0_N_K1, - CGridDesc_M_N, - AElementwiseOperation, - BElementwiseOperation, - CElementwiseOperation, - MPerBlock, - NPerBlock, - K0PerBlock, - MPerXDL, - NPerXDL, - K1, - MXdlPerWave, - NXdlPerWave, - ABlockTransferThreadClusterLengths_K0_M_K1, - ABlockTransferThreadClusterArrangeOrder, - ABlockTransferSrcAccessOrder, - ABlockTransferSrcVectorDim, - ABlockTransferSrcScalarPerVector, - ABlockTransferDstScalarPerVector_K1, - false, // AThreadTransferSrcResetCoordinateAfterRun, - ABlockLdsAddExtraM, - BBlockTransferThreadClusterLengths_K0_N_K1, - BBlockTransferThreadClusterArrangeOrder, - BBlockTransferSrcAccessOrder, - BBlockTransferSrcVectorDim, - BBlockTransferSrcScalarPerVector, - BBlockTransferDstScalarPerVector_K1, - false, // BThreadTransferSrcResetCoordinateAfterRun, - BBlockLdsAddExtraN, - Sequence<0, 2, 4, 5, 6, 1, 3, 7>, // CThreadTransferSrcDstAccessOrder, - CThreadTransferSrcDstVectorDim, - CThreadTransferDstScalarPerVector>; - - // GridwiseGemm - using GridwiseGemmAtomicAdd = GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4< - BlockSize, - ADataType, // TODO: distinguish A/B datatype - AccDataType, - CDataType, - InMemoryDataOperationEnum::AtomicAdd, - AGridDesc_K0_M_K1, - BGridDesc_K0_N_K1, - CGridDesc_M_N, - AElementwiseOperation, - BElementwiseOperation, - CElementwiseOperation, - MPerBlock, - NPerBlock, - K0PerBlock, - MPerXDL, - NPerXDL, - K1, - MXdlPerWave, - NXdlPerWave, - ABlockTransferThreadClusterLengths_K0_M_K1, - ABlockTransferThreadClusterArrangeOrder, - ABlockTransferSrcAccessOrder, - ABlockTransferSrcVectorDim, - ABlockTransferSrcScalarPerVector, - ABlockTransferDstScalarPerVector_K1, - false, // AThreadTransferSrcResetCoordinateAfterRun, - ABlockLdsAddExtraM, - BBlockTransferThreadClusterLengths_K0_N_K1, - BBlockTransferThreadClusterArrangeOrder, - BBlockTransferSrcAccessOrder, - BBlockTransferSrcVectorDim, - BBlockTransferSrcScalarPerVector, - BBlockTransferDstScalarPerVector_K1, - false, // BThreadTransferSrcResetCoordinateAfterRun, - BBlockLdsAddExtraN, - Sequence<0, 2, 4, 5, 6, 1, 3, 7>, // CThreadTransferSrcDstAccessOrder, - CThreadTransferSrcDstVectorDim, - CThreadTransferDstScalarPerVector>; - - using CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2 = - decltype(GridwiseGemm::MakeCM0N0M1N1M2M3M4N2GridDescriptor(CGridDesc_M_N{})); - - using Block2CTileMap = - decltype(GridwiseGemm::MakeCBlockClusterAdaptor(CGridDesc_M_N{}, 1, 1, 1)); - - // Argument - struct Argument : public BaseArgument - { - Argument(const ADataType* p_a_grid, - const BDataType* p_b_grid, - CDataType* p_c_grid, - index_t M, - index_t N, - index_t K, - index_t StrideA, - index_t StrideB, - index_t StrideC, - index_t M01, - index_t N01, - AElementwiseOperation a_element_op, - BElementwiseOperation b_element_op, - CElementwiseOperation c_element_op, - index_t k_batch) - : p_a_grid_{p_a_grid}, - p_b_grid_{p_b_grid}, - p_c_grid_{p_c_grid}, - a_grid_desc_kbatch_k0_m_k1_{}, - b_grid_desc_kbatch_k0_n_k1_{}, - c_grid_desc_m_n_{}, - c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_{}, - block_2_ctile_map_{}, - M01_{M01}, - N01_{N01}, - a_element_op_{a_element_op}, - b_element_op_{b_element_op}, - c_element_op_{c_element_op}, - k_batch_{k_batch} - { - int KPad = DeviceGemmXdlSplitK::GetKPad(K, k_batch_); - - a_grid_desc_kbatch_k0_m_k1_ = DeviceGemmXdlSplitK::MakeAGridDescriptor_KBatch_K0_M_K1( - M, K, StrideA, k_batch_, KPad); - b_grid_desc_kbatch_k0_n_k1_ = DeviceGemmXdlSplitK::MakeBGridDescriptor_KBatch_K0_N_K1( - K, N, StrideB, k_batch_, KPad); - c_grid_desc_m_n_ = DeviceGemmXdlSplitK::MakeCGridDescriptor_M_N(M, N, StrideC); - - block_2_ctile_map_ = - GridwiseGemm::MakeCBlockClusterAdaptor(c_grid_desc_m_n_, M01, N01, k_batch_); - - if(GridwiseGemm::CheckValidity(a_grid_desc_kbatch_k0_m_k1_, - b_grid_desc_kbatch_k0_n_k1_, - c_grid_desc_m_n_, - block_2_ctile_map_)) - { - c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_ = - GridwiseGemm::MakeCM0N0M1N1M2M3M4N2GridDescriptor(c_grid_desc_m_n_); - } - } - - // private: - const ADataType* p_a_grid_; - const BDataType* p_b_grid_; - CDataType* p_c_grid_; - AGridDesc_K0_M_K1 a_grid_desc_kbatch_k0_m_k1_; - BGridDesc_K0_N_K1 b_grid_desc_kbatch_k0_n_k1_; - CGridDesc_M_N c_grid_desc_m_n_; - CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2 c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_; - Block2CTileMap block_2_ctile_map_; - index_t M01_; - index_t N01_; - AElementwiseOperation a_element_op_; - BElementwiseOperation b_element_op_; - CElementwiseOperation c_element_op_; - index_t k_batch_; - }; - - // Invoker - struct Invoker : public BaseInvoker - { - using Argument = DeviceGemmXdlSplitK::Argument; - - void ShowInfo(const Argument& arg) - { - std::cout << "arg.a_grid_desc_kbatch_k0_m_k1_{" - << arg.a_grid_desc_kbatch_k0_m_k1_.GetLength(I0) << ", " - << arg.a_grid_desc_kbatch_k0_m_k1_.GetLength(I1) << ", " - << arg.a_grid_desc_kbatch_k0_m_k1_.GetLength(I2) << ", " - << arg.a_grid_desc_kbatch_k0_m_k1_.GetLength(I3) << "}" << std::endl; - - std::cout << "arg.b_grid_desc_kbatch_k0_n_k1_{" - << arg.b_grid_desc_kbatch_k0_n_k1_.GetLength(I0) << ", " - << arg.b_grid_desc_kbatch_k0_n_k1_.GetLength(I1) << ", " - << arg.b_grid_desc_kbatch_k0_n_k1_.GetLength(I2) << ", " - << arg.b_grid_desc_kbatch_k0_n_k1_.GetLength(I3) << "}" << std::endl; - - std::cout << "arg.c_grid_desc_m_n_{ " << arg.c_grid_desc_m_n_.GetLength(I0) << ", " - << arg.c_grid_desc_m_n_.GetLength(I1) << "}" << std::endl; - } - - float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{}) - { - ShowInfo(arg); - - const auto kbatch = arg.a_grid_desc_kbatch_k0_m_k1_.GetLength(I0); - - if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_kbatch_k0_m_k1_, - arg.b_grid_desc_kbatch_k0_n_k1_, - arg.c_grid_desc_m_n_, - arg.block_2_ctile_map_)) - { - throw std::runtime_error( - "wrong! GridwiseGemm_km_kn_m0m1n0n1_xdlops_v2r3 has invalid setting"); - } - - const index_t grid_size = - arg.block_2_ctile_map_.CalculateGridSize(arg.c_grid_desc_m_n_); - - const auto K0 = arg.a_grid_desc_kbatch_k0_m_k1_.GetLength(I1); - - const bool has_main_k0_block_loop = GridwiseGemm::CalculateHasMainK0BlockLoop(K0); - - float ave_time = 0; - - const auto Run = [&](const auto& kernel) { - // FIXME: this should be moved outside of DeviceOp - hipGetErrorString( - hipMemset(arg.p_c_grid_, - 0, - arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_.GetElementSpaceSize() * - sizeof(CDataType))); - - ave_time = launch_and_time_kernel(stream_config, - kernel, - dim3(grid_size), - dim3(BlockSize), - 0, - arg.p_a_grid_, - arg.p_b_grid_, - arg.p_c_grid_, - arg.a_grid_desc_kbatch_k0_m_k1_, - arg.b_grid_desc_kbatch_k0_n_k1_, - arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_, - arg.a_element_op_, - arg.b_element_op_, - arg.c_element_op_, - arg.block_2_ctile_map_); - }; - - if(has_main_k0_block_loop) - { - if(kbatch == 1) - { - const auto kernel = kernel_gemm_xdlops_v2r4< - GridwiseGemm, - ADataType, // TODO: distiguish A/B datatype - CDataType, - remove_reference_t, - remove_reference_t, - remove_reference_t, - AElementwiseOperation, - BElementwiseOperation, - CElementwiseOperation, - remove_reference_t, - true>; - - Run(kernel); - } - else - { - const auto kernel = kernel_gemm_xdlops_v2r4< - GridwiseGemmAtomicAdd, - ADataType, // TODO: distiguish A/B datatype - CDataType, - remove_reference_t, - remove_reference_t, - remove_reference_t, - AElementwiseOperation, - BElementwiseOperation, - CElementwiseOperation, - remove_reference_t, - true>; - - Run(kernel); - } - } - else - { - if(kbatch == 1) - { - const auto kernel = kernel_gemm_xdlops_v2r4< - GridwiseGemm, - ADataType, // TODO: distiguish A/B datatype - CDataType, - remove_reference_t, - remove_reference_t, - remove_reference_t, - AElementwiseOperation, - BElementwiseOperation, - CElementwiseOperation, - remove_reference_t, - false>; - - Run(kernel); - } - else - { - const auto kernel = kernel_gemm_xdlops_v2r4< - GridwiseGemmAtomicAdd, - ADataType, // TODO: distiguish A/B datatype - CDataType, - remove_reference_t, - remove_reference_t, - remove_reference_t, - AElementwiseOperation, - BElementwiseOperation, - CElementwiseOperation, - remove_reference_t, - false>; - - Run(kernel); - } - } - - return ave_time; - } - - // polymorphic - float Run(const BaseArgument* p_arg, - const StreamConfig& stream_config = StreamConfig{}) override - { - return Run(*dynamic_cast(p_arg), stream_config); - } - }; - - static constexpr bool IsValidCompilationParameter() - { - // TODO: properly implement this check - return true; - } - - static bool IsSupportedArgument(const Argument& arg) - { - if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a")) - { - return false; - } - - return GridwiseGemm::CheckValidity(arg.a_grid_desc_kbatch_k0_m_k1_, - arg.b_grid_desc_kbatch_k0_n_k1_, - arg.c_grid_desc_m_n_, - arg.block_2_ctile_map_); - } - - // polymorphic - bool IsSupportedArgument(const BaseArgument* p_arg) override - { - return IsSupportedArgument(*dynamic_cast(p_arg)); - } - - static auto MakeArgument(const ADataType* p_a, - const BDataType* p_b, - CDataType* p_c, - index_t M, - index_t N, - index_t K, - index_t StrideA, - index_t StrideB, - index_t StrideC, - AElementwiseOperation a_element_op, - BElementwiseOperation b_element_op, - CElementwiseOperation c_element_op, - index_t KBatch) - { - return Argument{p_a, - p_b, - p_c, - M, - N, - K, - StrideA, - StrideB, - StrideC, - 1, - 1, - a_element_op, - b_element_op, - c_element_op, - KBatch}; - } - - static auto MakeInvoker() { return Invoker{}; } - - // polymorphic - std::unique_ptr MakeArgumentPointer(const void* p_a, - const void* p_b, - void* p_c, - index_t M, - index_t N, - index_t K, - index_t StrideA, - index_t StrideB, - index_t StrideC, - AElementwiseOperation a_element_op, - BElementwiseOperation b_element_op, - CElementwiseOperation c_element_op, - ck::index_t KBatch = 1) override - { - return std::make_unique(static_cast(p_a), - static_cast(p_b), - static_cast(p_c), - M, - N, - K, - StrideA, - StrideB, - StrideC, - 1, - 1, - a_element_op, - b_element_op, - c_element_op, - KBatch); - } - - // polymorphic - std::unique_ptr MakeInvokerPointer() override - { - return std::make_unique(Invoker{}); - } - - // polymorphic - std::string GetTypeString() const override - { - auto str = std::stringstream(); - - // clang-format off - str << "DeviceGemmXdlSplitK" - << "<" - << BlockSize << ", " - << MPerBlock << ", " - << NPerBlock << ", " - << K0PerBlock - << ">"; - // clang-format on - - return str.str(); - } -}; - -} // namespace device -} // namespace tensor_operation -} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f32_f32_f32_km_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f32_f32_f32_km_kn_mn_instance.cpp index 051ff652b9..f9b05aba0e 100644 --- a/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f32_f32_f32_km_kn_mn_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f32_f32_f32_km_kn_mn_instance.cpp @@ -6,7 +6,7 @@ #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/device_gemm_xdl_splitk.hpp" +#include "ck/tensor_operation/gpu/device/device_gemm_xdl_splitk_c_shuffle.hpp" #include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" @@ -31,18 +31,18 @@ static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecializa // Compilation parameters for a[k, m] * b[k, n] = c[m, n] using device_gemm_xdl_splitk_f32_f32_f32_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| CThreadTransfer| CThreadTransfer| - //#################| 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| SrcDstVectorDim| DstScalar| - //#################| | | | | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | | PerVector| - //#################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | - DeviceGemmXdlSplitK< F32, F32, F32, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 256, 128, 4, 4, 32, 32, 4, 2, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 4, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 4, true, 7, 1>, - DeviceGemmXdlSplitK< F32, F32, F32, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 256, 4, 4, 32, 32, 2, 4, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 4, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 4, true, 7, 1>, - DeviceGemmXdlSplitK< F32, F32, F32, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 128, 4, 4, 32, 32, 4, 2, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 4, true, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 4, true, 7, 1>, - DeviceGemmXdlSplitK< F32, F32, F32, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 4, 4, 32, 32, 2, 2, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 4, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 4, true, 7, 1>, - DeviceGemmXdlSplitK< F32, F32, F32, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 4, true, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 4, true, 7, 1>, - DeviceGemmXdlSplitK< F32, F32, F32, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 64, 128, 4, 4, 32, 32, 2, 2, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 4, true, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 4, true, 7, 1>, - DeviceGemmXdlSplitK< F32, F32, F32, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 64, 4, 4, 32, 32, 2, 1, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 4, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 1, 4, true, 7, 1>, - DeviceGemmXdlSplitK< F32, F32, F32, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 64, 128, 4, 4, 32, 32, 1, 2, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 1, 4, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 4, true, 7, 1> + //#########################|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| + //#########################| 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| + //#########################| | | | | | | | 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< F32, F32, F32, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 256, 128, 4, 4, 32, 32, 4, 2, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 4, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 4, true, 1, 1, S<1, 32, 1, 8>, 4>, + DeviceGemmXdlSplitKCShuffle< F32, F32, F32, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 256, 4, 4, 32, 32, 2, 4, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 4, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 8>, 4>, + DeviceGemmXdlSplitKCShuffle< F32, F32, F32, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 128, 4, 4, 32, 32, 4, 2, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 4, true, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 8>, 4>, + DeviceGemmXdlSplitKCShuffle< F32, F32, F32, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 4, 4, 32, 32, 2, 2, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 4, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 4, true, 1, 1, S<1, 32, 1, 8>, 4>, + DeviceGemmXdlSplitKCShuffle< F32, F32, F32, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 4, true, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 4, true, 1, 1, S<1, 32, 1, 4>, 4>, + DeviceGemmXdlSplitKCShuffle< F32, F32, F32, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 64, 128, 4, 4, 32, 32, 2, 2, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 4, true, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 8>, 4>, + DeviceGemmXdlSplitKCShuffle< F32, F32, F32, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 64, 4, 4, 32, 32, 2, 1, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 4, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 1, 4, true, 1, 1, S<1, 16, 1, 4>, 4>, + DeviceGemmXdlSplitKCShuffle< F32, F32, F32, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 64, 128, 4, 4, 32, 32, 1, 2, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 1, 4, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 4, true, 1, 1, S<1, 32, 1, 8>, 4> // clang-format on >; diff --git a/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f32_f32_f32_km_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f32_f32_f32_km_nk_mn_instance.cpp index 5d3cbf896b..c375befdd1 100644 --- a/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f32_f32_f32_km_nk_mn_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f32_f32_f32_km_nk_mn_instance.cpp @@ -6,7 +6,7 @@ #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/device_gemm_xdl_splitk.hpp" +#include "ck/tensor_operation/gpu/device/device_gemm_xdl_splitk_c_shuffle.hpp" #include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" @@ -31,18 +31,18 @@ static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecializa // Compilation parameters for a[k, m] * b[n, k] = c[m, n] using device_gemm_xdl_splitk_f32_f32_f32_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| CThreadTransfer| CThreadTransfer| - //#################| 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| SrcDstVectorDim| DstScalar| - //#################| | | | | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | | PerVector| - //#################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | - DeviceGemmXdlSplitK< F32, F32, F32, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 256, 128, 4, 4, 32, 32, 4, 2, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 4, true, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 4, 4, true, 7, 1>, - DeviceGemmXdlSplitK< F32, F32, F32, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 256, 4, 4, 32, 32, 2, 4, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 4, true, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 4, 4, true, 7, 1>, - DeviceGemmXdlSplitK< F32, F32, F32, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 128, 4, 4, 32, 32, 4, 2, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 4, true, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 4, 4, true, 7, 1>, - DeviceGemmXdlSplitK< F32, F32, F32, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 4, 4, 32, 32, 2, 2, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 4, true, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 4, 4, true, 7, 1>, - DeviceGemmXdlSplitK< F32, F32, F32, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 4, true, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 4, 4, true, 7, 1>, - DeviceGemmXdlSplitK< F32, F32, F32, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 64, 128, 4, 4, 32, 32, 2, 2, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 4, true, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 4, 4, true, 7, 1>, - DeviceGemmXdlSplitK< F32, F32, F32, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 64, 4, 4, 32, 32, 2, 1, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 4, true, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 4, 4, true, 7, 1>, - DeviceGemmXdlSplitK< F32, F32, F32, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 64, 128, 4, 4, 32, 32, 1, 2, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 1, 4, true, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 4, 4, true, 7, 1> + //#########################|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| + //#########################| 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| + //#########################| | | | | | | | 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< F32, F32, F32, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 256, 128, 4, 4, 32, 32, 4, 2, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 4, true, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 4, 4, true, 1, 1, S<1, 32, 1, 8>, 4>, + DeviceGemmXdlSplitKCShuffle< F32, F32, F32, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 256, 4, 4, 32, 32, 2, 4, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 4, true, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 4, 4, true, 1, 1, S<1, 32, 1, 8>, 4>, + DeviceGemmXdlSplitKCShuffle< F32, F32, F32, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 128, 4, 4, 32, 32, 4, 2, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 4, true, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 4, 4, true, 1, 1, S<1, 16, 1, 8>, 4>, + DeviceGemmXdlSplitKCShuffle< F32, F32, F32, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 4, 4, 32, 32, 2, 2, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 4, true, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 4, 4, true, 1, 1, S<1, 32, 1, 8>, 4>, + DeviceGemmXdlSplitKCShuffle< F32, F32, F32, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 4, true, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 4, 4, true, 1, 1, S<1, 32, 1, 4>, 4>, + DeviceGemmXdlSplitKCShuffle< F32, F32, F32, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 64, 128, 4, 4, 32, 32, 2, 2, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 4, true, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 4, 4, true, 1, 1, S<1, 16, 1, 8>, 4>, + DeviceGemmXdlSplitKCShuffle< F32, F32, F32, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 64, 4, 4, 32, 32, 2, 1, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 4, true, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 4, 4, true, 1, 1, S<1, 16, 1, 4>, 4>, + DeviceGemmXdlSplitKCShuffle< F32, F32, F32, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 64, 128, 4, 4, 32, 32, 1, 2, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 1, 4, true, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 4, 4, true, 1, 1, S<1, 32, 1, 8>, 4> // clang-format on >; diff --git a/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f32_f32_f32_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f32_f32_f32_mk_kn_mn_instance.cpp index 9a9b05a326..299686521a 100644 --- a/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f32_f32_f32_mk_kn_mn_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f32_f32_f32_mk_kn_mn_instance.cpp @@ -6,7 +6,7 @@ #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/device_gemm_xdl_splitk.hpp" +#include "ck/tensor_operation/gpu/device/device_gemm_xdl_splitk_c_shuffle.hpp" #include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" @@ -26,28 +26,23 @@ using S = ck::Sequence; using PassThrough = ck::tensor_operation::element_wise::PassThrough; -static constexpr auto GemmMNPadding = ck::tensor_operation::device::GemmSpecialization::MNPadding; +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_f32_f32_f32_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| CThreadTransfer| CThreadTransfer| - //###################| 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| SrcDstVectorDim| DstScalar| - //###################| | | | | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | | PerVector| - //###################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | - DeviceGemmXdlSplitK< F32, F32, F32, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmMNPadding, 256, 96, 128, 4, 8, 16, 16, 3, 4, S<1, 4, 32, 2>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 4, 4, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 4, true, 7, 1>, - DeviceGemmXdlSplitK< F32, F32, F32, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmMNPadding, 256, 256, 128, 4, 4, 32, 32, 4, 2, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 4, 4, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 4, true, 7, 1>, - DeviceGemmXdlSplitK< F32, F32, F32, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmMNPadding, 256, 128, 256, 4, 4, 32, 32, 2, 4, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 4, 4, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 4, true, 7, 1>, - DeviceGemmXdlSplitK< F32, F32, F32, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmMNPadding, 128, 128, 128, 4, 4, 32, 32, 4, 2, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 4, 4, true, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 4, true, 7, 1>, - DeviceGemmXdlSplitK< F32, F32, F32, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmMNPadding, 256, 128, 128, 4, 4, 32, 32, 2, 2, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 4, 4, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 4, true, 7, 1>, - DeviceGemmXdlSplitK< F32, F32, F32, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmMNPadding, 128, 128, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 4, 4, true, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 4, true, 7, 1>, - DeviceGemmXdlSplitK< F32, F32, F32, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmMNPadding, 128, 64, 128, 4, 4, 32, 32, 2, 2, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 4, 4, true, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 4, true, 7, 1>, - DeviceGemmXdlSplitK< F32, F32, F32, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmMNPadding, 256, 128, 64, 4, 4, 32, 32, 2, 1, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 4, 4, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 1, 4, true, 7, 1>, - DeviceGemmXdlSplitK< F32, F32, F32, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmMNPadding, 256, 64, 128, 4, 4, 32, 32, 1, 2, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 4, 4, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 4, true, 7, 1>, - DeviceGemmXdlSplitK< F32, F32, F32, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmMNPadding, 256, 32, 256, 4, 4, 32, 32, 1, 2, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 4, 4, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 4, true, 7, 1>, - DeviceGemmXdlSplitK< F32, F32, F32, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmMNPadding, 128, 32, 128, 4, 4, 32, 32, 1, 2, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 4, 4, true, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 4, true, 7, 1>, - DeviceGemmXdlSplitK< F32, F32, F32, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmMNPadding, 256, 16, 256, 4, 4, 16, 16, 1, 4, S<1, 4, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 4, 4, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 4, true, 7, 1>, - DeviceGemmXdlSplitK< F32, F32, F32, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmMNPadding, 128, 16, 128, 4, 4, 16, 16, 1, 4, S<1, 4, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 4, 4, true, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 4, true, 7, 1> + //#########################|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| + //#########################| 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| + //#########################| | | | | | | | 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< F32, F32, F32, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 256, 128, 4, 4, 32, 32, 4, 2, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 4, 4, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 4, true, 1, 1, S<1, 32, 1, 8>, 4>, + DeviceGemmXdlSplitKCShuffle< F32, F32, F32, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 256, 4, 4, 32, 32, 2, 4, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 4, 4, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 8>, 4>, + DeviceGemmXdlSplitKCShuffle< F32, F32, F32, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 128, 4, 4, 32, 32, 4, 2, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 4, 4, true, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 8>, 4>, + DeviceGemmXdlSplitKCShuffle< F32, F32, F32, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 4, 4, 32, 32, 2, 2, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 4, 4, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 4, true, 1, 1, S<1, 32, 1, 8>, 4>, + DeviceGemmXdlSplitKCShuffle< F32, F32, F32, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 4, 4, true, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 4, true, 1, 1, S<1, 32, 1, 4>, 4>, + DeviceGemmXdlSplitKCShuffle< F32, F32, F32, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 64, 128, 4, 4, 32, 32, 2, 2, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 4, 4, true, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 8>, 4>, + DeviceGemmXdlSplitKCShuffle< F32, F32, F32, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 64, 4, 4, 32, 32, 2, 1, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 4, 4, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 1, 4, true, 1, 1, S<1, 16, 1, 4>, 4>, + DeviceGemmXdlSplitKCShuffle< F32, F32, F32, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 64, 128, 4, 4, 32, 32, 1, 2, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 4, 4, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 4, true, 1, 1, S<1, 32, 1, 8>, 4> // clang-format on >; diff --git a/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f32_f32_f32_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f32_f32_f32_mk_nk_mn_instance.cpp index 50dc93051d..3786743e72 100644 --- a/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f32_f32_f32_mk_nk_mn_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f32_f32_f32_mk_nk_mn_instance.cpp @@ -6,7 +6,7 @@ #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/device_gemm_xdl_splitk.hpp" +#include "ck/tensor_operation/gpu/device/device_gemm_xdl_splitk_c_shuffle.hpp" #include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" @@ -31,23 +31,23 @@ static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecializa // Compilation parameters for a[m, k] * b[n, k] = c[m, n] using device_gemm_xdl_splitk_f32_f32_f32_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| CThreadTransfer| CThreadTransfer| - //#################| 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| SrcDstVectorDim| DstScalar| - //#################| | | | | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | | PerVector| - //#################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | - DeviceGemmXdlSplitK< F32, F32, F32, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 256, 128, 4, 4, 32, 32, 4, 2, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 4, 4, true, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 4, 4, true, 7, 1>, - DeviceGemmXdlSplitK< F32, F32, F32, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 256, 4, 4, 32, 32, 2, 4, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 4, 4, true, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 4, 4, true, 7, 1>, - DeviceGemmXdlSplitK< F32, F32, F32, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 128, 4, 4, 32, 32, 4, 2, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 4, 4, true, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 4, 4, true, 7, 1>, - DeviceGemmXdlSplitK< F32, F32, F32, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 4, 4, 32, 32, 2, 2, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 4, 4, true, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 4, 4, true, 7, 1>, - DeviceGemmXdlSplitK< F32, F32, F32, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 4, 4, true, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 4, 4, true, 7, 1>, - DeviceGemmXdlSplitK< F32, F32, F32, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 64, 128, 4, 4, 32, 32, 2, 2, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 4, 4, true, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 4, 4, true, 7, 1>, - DeviceGemmXdlSplitK< F32, F32, F32, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 4, 4, true, S<1, 4, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 4, 4, true, 7, 1>, - DeviceGemmXdlSplitK< F32, F32, F32, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 64, 4, 4, 32, 32, 2, 1, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 4, 4, true, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 4, 4, true, 7, 1>, - DeviceGemmXdlSplitK< F32, F32, F32, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 64, 128, 4, 4, 32, 32, 1, 2, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 4, 4, true, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 4, 4, true, 7, 1>, - DeviceGemmXdlSplitK< F32, F32, F32, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 32, 4, 4, 32, 32, 2, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 4, 4, true, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 4, 4, true, 7, 1>, - DeviceGemmXdlSplitK< F32, F32, F32, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 32, 128, 4, 4, 32, 32, 1, 2, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 4, 4, true, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 4, 4, true, 7, 1>, - DeviceGemmXdlSplitK< F32, F32, F32, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 32, 4, 4, 32, 32, 2, 1, S<1, 4, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 4, 4, true, S<1, 4, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 4, 4, true, 7, 1>, - DeviceGemmXdlSplitK< F32, F32, F32, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 32, 64, 4, 4, 32, 32, 1, 2, S<1, 4, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 4, 4, true, S<1, 4, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 4, 4, true, 7, 1> + //#########################|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| + //#########################| 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| + //#########################| | | | | | | | 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< F32, F32, F32, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 256, 128, 4, 4, 32, 32, 4, 2, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 4, 4, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 3, 4, 4, true, 1, 1, S<1, 32, 1, 8>, 4>, + DeviceGemmXdlSplitKCShuffle< F32, F32, F32, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 256, 4, 4, 32, 32, 2, 4, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 4, 4, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 3, 4, 4, true, 1, 1, S<1, 32, 1, 8>, 4>, + DeviceGemmXdlSplitKCShuffle< F32, F32, F32, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 128, 4, 4, 32, 32, 4, 2, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 4, 4, true, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 3, 4, 4, true, 1, 1, S<1, 16, 1, 8>, 4>, + DeviceGemmXdlSplitKCShuffle< F32, F32, F32, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 4, 4, 32, 32, 2, 2, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 4, 4, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 3, 4, 4, true, 1, 1, S<1, 32, 1, 8>, 4>, + DeviceGemmXdlSplitKCShuffle< F32, F32, F32, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 4, 4, true, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 3, 4, 4, true, 1, 1, S<1, 32, 1, 4>, 4>, + DeviceGemmXdlSplitKCShuffle< F32, F32, F32, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 64, 128, 4, 4, 32, 32, 2, 2, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 4, 4, true, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 3, 4, 4, true, 1, 1, S<1, 16, 1, 8>, 4>, + DeviceGemmXdlSplitKCShuffle< F32, F32, F32, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 4, 4, true, S<1, 4, 16, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 3, 4, 4, true, 1, 1, S<1, 16, 1, 4>, 4>, + DeviceGemmXdlSplitKCShuffle< F32, F32, F32, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 64, 4, 4, 32, 32, 2, 1, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 4, 4, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 3, 4, 4, true, 1, 1, S<1, 32, 1, 8>, 4>, + DeviceGemmXdlSplitKCShuffle< F32, F32, F32, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 64, 128, 4, 4, 32, 32, 1, 2, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 4, 4, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 3, 4, 4, true, 1, 1, S<1, 32, 1, 8>, 4>, + DeviceGemmXdlSplitKCShuffle< F32, F32, F32, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 32, 4, 4, 32, 32, 2, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 4, 4, true, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 3, 4, 4, true, 1, 1, S<1, 32, 1, 4>, 4>, + DeviceGemmXdlSplitKCShuffle< F32, F32, F32, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 32, 128, 4, 4, 32, 32, 1, 2, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 4, 4, true, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 3, 4, 4, true, 1, 1, S<1, 16, 1, 8>, 4>, + DeviceGemmXdlSplitKCShuffle< F32, F32, F32, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 32, 4, 4, 32, 32, 2, 1, S<1, 4, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 4, 4, true, S<1, 4, 16, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 3, 4, 4, true, 1, 1, S<1, 16, 1, 4>, 4>, + DeviceGemmXdlSplitKCShuffle< F32, F32, F32, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 32, 64, 4, 4, 32, 32, 1, 2, S<1, 4, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 4, 4, true, S<1, 4, 16, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 3, 4, 4, true, 1, 1, S<1, 16, 1, 4>, 4> // clang-format on >; diff --git a/test/gemm_split_k/gemm_split_k.cpp b/test/gemm_split_k/gemm_split_k.cpp index e03cd4fa19..0a4cc2311f 100644 --- a/test/gemm_split_k/gemm_split_k.cpp +++ b/test/gemm_split_k/gemm_split_k.cpp @@ -8,7 +8,6 @@ #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/device_gemm_xdl_splitk.hpp" #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp" #include "ck/library/tensor_operation_instance/gpu/gemm_splitk.hpp"