device_gemm_multiple_abd_xdl_cshuffle.hpp Source File

device_gemm_multiple_abd_xdl_cshuffle.hpp Source File#

Composable Kernel: device_gemm_multiple_abd_xdl_cshuffle.hpp Source File
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
 
#pragma once
 
#include <iostream>
#include <sstream>
 
#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_v2.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_multi_abd.hpp"
#include "ck/host_utility/device_prop.hpp"
#include "ck/host_utility/kernel_launch.hpp"
#include "ck/tensor_operation/gpu/device/device_gemm_multiple_abd.hpp"
 
namespace ck {
namespace tensor_operation {
namespace device {
 
template <typename AsLayout,
          typename BsLayout,
          typename DsLayout,
          typename CLayout,
          typename AsDataType,
          typename BsDataType,
          typename GemmAccDataType,
          typename CShuffleDataType,
          typename DsDataType,
          typename CDataType,
          typename AElementwiseOperation,
          typename BElementwiseOperation,
          typename CElementwiseOperation,
          GemmSpecialization GemmSpec,
          index_t NumGemmKPrefetchStage,
          index_t BlockSize,
          index_t MPerBlock,
          index_t NPerBlock,
          index_t KPerBlock,
          index_t AK1,
          index_t BK1,
          index_t MPerXDL,
          index_t NPerXDL,
          index_t MXdlPerWave,
          index_t NXdlPerWave,
          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
          typename ABlockTransferThreadClusterArrangeOrder,
          typename ABlockTransferSrcAccessOrder,
          index_t ABlockTransferSrcVectorDim,
          index_t ABlockTransferSrcScalarPerVector,
          index_t ABlockTransferDstScalarPerVector_AK1,
          bool ABlockLdsExtraM,
          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
          typename BBlockTransferThreadClusterArrangeOrder,
          typename BBlockTransferSrcAccessOrder,
          index_t BBlockTransferSrcVectorDim,
          index_t BBlockTransferSrcScalarPerVector,
          index_t BBlockTransferDstScalarPerVector_BK1,
          bool BBlockLdsExtraN,
          index_t CShuffleMXdlPerWavePerShuffle,
          index_t CShuffleNXdlPerWavePerShuffle,
          typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
          index_t CShuffleBlockTransferScalarPerVector_NPerBlock,
          BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave,
          BlockGemmPipelineVersion BlkGemmPipelineVer = BlockGemmPipelineVersion::v1,
          typename ComputeTypeA                       = CDataType,
          typename ComputeTypeB                       = ComputeTypeA>
struct DeviceGemmMultipleABD_Xdl_CShuffle : public DeviceGemmMultipleABD<AsLayout,
                                                                         BsLayout,
                                                                         DsLayout,
                                                                         CLayout,
                                                                         AsDataType,
                                                                         BsDataType,
                                                                         DsDataType,
                                                                         CDataType,
                                                                         AElementwiseOperation,
                                                                         BElementwiseOperation,
                                                                         CElementwiseOperation>
{
    GET_NXDL_PER_WAVE_IMPL
    static constexpr auto NXdlPerWave64 = GetNXdlPerWave<true>();
    static constexpr auto NXdlPerWave32 = GetNXdlPerWave<false>();
 
    static constexpr index_t NumATensor = AsDataType::Size();
    static constexpr index_t NumBTensor = BsDataType::Size();
    static constexpr index_t NumDTensor = DsDataType::Size();
 
    using ALayout = remove_cvref_t<tuple_element_t<0, AsLayout>>;
    using BLayout = remove_cvref_t<tuple_element_t<0, BsLayout>>;
 
    // GridwiseGemm
    template <index_t NXdlPerWave_>
    using GridwiseGemmBase = GridwiseGemm_xdl_cshuffle_v3<
        ALayout,
        BLayout,
        CLayout,
        AsDataType,
        BsDataType,
        GemmAccDataType,
        CShuffleDataType,
        DsDataType,
        CDataType,
        AElementwiseOperation,
        BElementwiseOperation,
        CElementwiseOperation,
        GemmSpec,
        BlockSize,
        MPerBlock,
        NPerBlock,
        KPerBlock,
        AK1,
        BK1,
        MPerXDL,
        NPerXDL,
        MXdlPerWave,
        NXdlPerWave_,
        ABlockTransferThreadClusterLengths_AK0_M_AK1,
        ABlockTransferThreadClusterArrangeOrder,
        ABlockTransferSrcAccessOrder,
        ABlockTransferSrcVectorDim,
        ABlockTransferSrcScalarPerVector,
        ABlockTransferDstScalarPerVector_AK1,
        false,
        ABlockLdsExtraM,
        BBlockTransferThreadClusterLengths_BK0_N_BK1,
        BBlockTransferThreadClusterArrangeOrder,
        BBlockTransferSrcAccessOrder,
        BBlockTransferSrcVectorDim,
        BBlockTransferSrcScalarPerVector,
        BBlockTransferDstScalarPerVector_BK1,
        false,
        BBlockLdsExtraN,
        CShuffleMXdlPerWavePerShuffle,
        CShuffleNXdlPerWavePerShuffle,
        CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
        CShuffleBlockTransferScalarPerVector_NPerBlock,
        BlkGemmPipeSched,
        BlkGemmPipelineVer,
        ComputeTypeA,
        ComputeTypeB>;
    using GridwiseGemm64 = GridwiseGemmBase<math::max(NXdlPerWave64, 1)>;
    using GridwiseGemm32 = GridwiseGemmBase<NXdlPerWave32>;
 
    using Argument = typename GridwiseGemm64::Argument;
 
    // Invoker
    struct Invoker : public BaseInvoker
    {
        template <typename GridwiseGemm>
        float RunImp(const typename GridwiseGemm::Argument& arg,
                     const StreamConfig& stream_config = StreamConfig{})
        {
            if(stream_config.log_level_ > 0)
            {
                arg.Print();
            }
 
            if(!GridwiseGemm::CheckValidity(arg))
            {
                throw std::runtime_error("wrong! GridwiseGemm has invalid setting");
            }
 
            index_t gdx, gdy, gdz;
            std::tie(gdx, gdy, gdz) = GridwiseGemm::CalculateGridSize(arg.M, arg.N, arg.KBatch);
 
            float ave_time = 0;
 
            index_t k_grain = arg.KBatch * KPerBlock;
            index_t K_split = (arg.K + k_grain - 1) / k_grain * KPerBlock;
 
            const bool has_main_k_block_loop = GridwiseGemm::CalculateHasMainKBlockLoop(K_split);
 
            const auto Run = [&](const auto& kernel) {
                if(arg.KBatch > 1)
                    hipGetErrorString(hipMemsetAsync(arg.p_c_grid,
                                                     0,
                                                     arg.M * arg.N * sizeof(CDataType),
                                                     stream_config.stream_id_));
 
                ave_time = launch_and_time_kernel(
                    stream_config, kernel, dim3(gdx, gdy, gdz), dim3(BlockSize), 0, arg);
            };
 
            constexpr index_t minimum_occupancy =
                BlkGemmPipeSched == BlockGemmPipelineScheduler::Intrawave ? 1 : 2;
 
            if(has_main_k_block_loop)
            {
                // Tail number always full
                if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v1 ||
                             BlkGemmPipelineVer == BlockGemmPipelineVersion::v3)
                {
#if 0
                    if(arg.KBatch > 1)
                    {
                        const auto kernel =
                            kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
                                                        true,
                                                        InMemoryDataOperationEnum::AtomicAdd,
                                                        minimum_occupancy>;
                        Run(kernel);
                    }
                    else
#endif
                    {
                        const auto kernel =
                            kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
                                                        true,
                                                        InMemoryDataOperationEnum::Set,
                                                        minimum_occupancy>;
                        Run(kernel);
                    }
                }
                // Tail number could be One to Seven
                else if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v2)
                {
#if 0
                    if(arg.KBatch > 1)
                    {
                        if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::One)
                        {
                            const auto kernel =
                                kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
                                                            true,
                                                            InMemoryDataOperationEnum::AtomicAdd,
                                                            minimum_occupancy,
                                                            TailNumber::One>;
                            Run(kernel);
                        }
                        else if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) ==
                                TailNumber::Full)
                        {
                            const auto kernel =
                                kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
                                                            true,
                                                            InMemoryDataOperationEnum::AtomicAdd,
                                                            minimum_occupancy,
                                                            TailNumber::Full>;
                            Run(kernel);
                        }
 
                        if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 2)
                        {
                            if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Two)
                            {
                                const auto kernel = kernel_gemm_xdl_cshuffle_v3<
                                    GridwiseGemm,
                                    true,
                                    InMemoryDataOperationEnum::AtomicAdd,
                                    minimum_occupancy,
                                    TailNumber::Two>;
                                Run(kernel);
                            }
                        }
 
                        if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 3)
                        {
                            if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) ==
                               TailNumber::Three)
                            {
                                const auto kernel = kernel_gemm_xdl_cshuffle_v3<
                                    GridwiseGemm,
                                    true,
                                    InMemoryDataOperationEnum::AtomicAdd,
                                    minimum_occupancy,
                                    TailNumber::Three>;
                                Run(kernel);
                            }
                        }
 
                        if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 4)
                        {
                            if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) ==
                               TailNumber::Four)
                            {
                                const auto kernel = kernel_gemm_xdl_cshuffle_v3<
                                    GridwiseGemm,
                                    true,
                                    InMemoryDataOperationEnum::AtomicAdd,
                                    minimum_occupancy,
                                    TailNumber::Four>;
                                Run(kernel);
                            }
                        }
 
                        if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 5)
                        {
                            if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) ==
                               TailNumber::Five)
                            {
                                const auto kernel = kernel_gemm_xdl_cshuffle_v3<
                                    GridwiseGemm,
                                    true,
                                    InMemoryDataOperationEnum::AtomicAdd,
                                    minimum_occupancy,
                                    TailNumber::Five>;
                                Run(kernel);
                            }
                        }
 
                        if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 6)
                        {
                            if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Six)
                            {
                                const auto kernel = kernel_gemm_xdl_cshuffle_v3<
                                    GridwiseGemm,
                                    true,
                                    InMemoryDataOperationEnum::AtomicAdd,
                                    minimum_occupancy,
                                    TailNumber::Six>;
                                Run(kernel);
                            }
                        }
 
                        if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 7)
                        {
                            if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) ==
                               TailNumber::Seven)
                            {
                                const auto kernel = kernel_gemm_xdl_cshuffle_v3<
                                    GridwiseGemm,
                                    true,
                                    InMemoryDataOperationEnum::AtomicAdd,
                                    minimum_occupancy,
                                    TailNumber::Seven>;
                                Run(kernel);
                            }
                        }
                    }
                    else
#endif
                    {
                        if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::One)
                        {
                            const auto kernel =
                                kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
                                                            true,
                                                            InMemoryDataOperationEnum::Set,
                                                            minimum_occupancy,
                                                            TailNumber::One>;
                            Run(kernel);
                        }
                        else if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) ==
                                TailNumber::Full)
                        {
                            const auto kernel =
                                kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
                                                            true,
                                                            InMemoryDataOperationEnum::Set,
                                                            minimum_occupancy,
                                                            TailNumber::Full>;
                            Run(kernel);
                        }
 
                        if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 2)
                        {
                            if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Two)
                            {
                                const auto kernel =
                                    kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
                                                                true,
                                                                InMemoryDataOperationEnum::Set,
                                                                minimum_occupancy,
                                                                TailNumber::Two>;
                                Run(kernel);
                            }
                        }
 
                        if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 3)
                        {
                            if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) ==
                               TailNumber::Three)
                            {
                                const auto kernel =
                                    kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
                                                                true,
                                                                InMemoryDataOperationEnum::Set,
                                                                minimum_occupancy,
                                                                TailNumber::Three>;
                                Run(kernel);
                            }
                        }
 
                        if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 4)
                        {
                            if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) ==
                               TailNumber::Four)
                            {
                                const auto kernel =
                                    kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
                                                                true,
                                                                InMemoryDataOperationEnum::Set,
                                                                minimum_occupancy,
                                                                TailNumber::Four>;
                                Run(kernel);
                            }
                        }
 
                        if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 5)
                        {
                            if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) ==
                               TailNumber::Five)
                            {
                                const auto kernel =
                                    kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
                                                                true,
                                                                InMemoryDataOperationEnum::Set,
                                                                minimum_occupancy,
                                                                TailNumber::Five>;
                                Run(kernel);
                            }
                        }
 
                        if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 6)
                        {
                            if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Six)
                            {
                                const auto kernel =
                                    kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
                                                                true,
                                                                InMemoryDataOperationEnum::Set,
                                                                minimum_occupancy,
                                                                TailNumber::Six>;
                                Run(kernel);
                            }
                        }
 
                        if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 7)
                        {
                            if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) ==
                               TailNumber::Seven)
                            {
                                const auto kernel =
                                    kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
                                                                true,
                                                                InMemoryDataOperationEnum::Set,
                                                                minimum_occupancy,
                                                                TailNumber::Seven>;
                                Run(kernel);
                            }
                        }
                    }
                }
                // Tail number could be Odd or Even
                else if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v4)
                {
#if 0
                    if(arg.KBatch > 1)
                    {
                        if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd)
                        {
                            const auto kernel = kernel_gemm_xdl_cshuffle_v3_2lds<
                                GridwiseGemm,
                                true,
                                InMemoryDataOperationEnum::AtomicAdd,
                                minimum_occupancy,
                                TailNumber::Odd>;
                            Run(kernel);
                        }
                        else
                        {
                            const auto kernel = kernel_gemm_xdl_cshuffle_v3_2lds<
                                GridwiseGemm,
                                true,
                                InMemoryDataOperationEnum::AtomicAdd,
                                minimum_occupancy,
                                TailNumber::Even>;
                            Run(kernel);
                        }
                    }
                    else
#endif
                    {
                        if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd)
                        {
                            const auto kernel =
                                kernel_gemm_xdl_cshuffle_v3_2lds<GridwiseGemm,
                                                                 true,
                                                                 InMemoryDataOperationEnum::Set,
                                                                 minimum_occupancy,
                                                                 TailNumber::Odd>;
                            Run(kernel);
                        }
                        else
                        {
                            const auto kernel =
                                kernel_gemm_xdl_cshuffle_v3_2lds<GridwiseGemm,
                                                                 true,
                                                                 InMemoryDataOperationEnum::Set,
                                                                 minimum_occupancy,
                                                                 TailNumber::Even>;
                            Run(kernel);
                        }
                    }
                }
                else
                {
#if 0
                    if(arg.KBatch > 1)
                    {
                        if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd)
                        {
                            const auto kernel =
                                kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
                                                            true,
                                                            InMemoryDataOperationEnum::AtomicAdd,
                                                            minimum_occupancy,
                                                            TailNumber::Odd>;
                            Run(kernel);
                        }
                        else
                        {
                            const auto kernel =
                                kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
                                                            true,
                                                            InMemoryDataOperationEnum::AtomicAdd,
                                                            minimum_occupancy,
                                                            TailNumber::Even>;
                            Run(kernel);
                        }
                    }
                    else
#endif
                    {
                        if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd)
                        {
                            const auto kernel =
                                kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
                                                            true,
                                                            InMemoryDataOperationEnum::Set,
                                                            minimum_occupancy,
                                                            TailNumber::Odd>;
                            Run(kernel);
                        }
                        else
                        {
                            const auto kernel =
                                kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
                                                            true,
                                                            InMemoryDataOperationEnum::Set,
                                                            minimum_occupancy,
                                                            TailNumber::Even>;
                            Run(kernel);
                        }
                    }
                }
            }
            else
            {
                // Tail number always 1
                if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v1)
                {
#if 0
                    if(arg.KBatch > 1)
                    {
                        const auto kernel =
                            kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
                                                        false,
                                                        InMemoryDataOperationEnum::AtomicAdd,
                                                        minimum_occupancy>;
                        Run(kernel);
                    }
                    else
#endif
                    {
                        const auto kernel =
                            kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
                                                        false,
                                                        InMemoryDataOperationEnum::Set,
                                                        minimum_occupancy>;
                        Run(kernel);
                    }
                }
            }
 
            return ave_time;
        }
 
        INVOKER_RUN3_IMPL
 
        // polymorphic
        float Run(const BaseArgument* p_arg,
                  const StreamConfig& stream_config = StreamConfig{}) override
        {
            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
        }
    };
 
    static constexpr bool IsValidCompilationParameter()
    {
        // TODO: properly implement this check
        return true;
    }
 
    static bool IsSupportedArgument(const Argument& arg)
    {
        if(!ck::is_xdl_wmma_supported<ComputeTypeA, ComputeTypeB, MPerXDL, NPerXDL>())
        {
            return false;
        }
        if((arg.K % AK1 != 0 || arg.K % BK1 != 0) && !(GemmSpec == GemmSpecialization::MKPadding ||
                                                       GemmSpec == GemmSpecialization::NKPadding ||
                                                       GemmSpec == GemmSpecialization::MNKPadding ||
                                                       GemmSpec == GemmSpecialization::KPadding))
        {
            return false;
        }
 
        if(get_warp_size() == 64)
        {
            if constexpr(NXdlPerWave64 > 0)
            {
                return GridwiseGemm64::CheckValidity(arg);
            }
        }
        else
        {
            if constexpr(NXdlPerWave32 > 0)
            {
                return GridwiseGemm32::CheckValidity(
                    reinterpret_cast<const typename GridwiseGemm32::Argument&>(arg));
            }
        }
        return false;
    }
 
    // polymorphic
    bool IsSupportedArgument(const BaseArgument* p_arg) override
    {
        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
    }
 
    static auto MakeArgument(std::array<const void*, NumATensor> p_as,
                             std::array<const void*, NumBTensor> p_bs,
                             std::array<const void*, NumDTensor> p_ds,
                             void* p_e,
                             index_t MRaw,
                             index_t NRaw,
                             index_t KRaw,
                             std::array<index_t, NumATensor> StrideAs,
                             std::array<index_t, NumBTensor> StrideBs,
                             std::array<index_t, NumDTensor> StrideDs,
                             index_t StrideE,
                             AElementwiseOperation a_element_op,
                             BElementwiseOperation b_element_op,
                             CElementwiseOperation c_element_op)
    {
 
        static_for<0, NumATensor, 1>{}([&](auto i) {
            using ALayout_ = remove_cvref_t<tuple_element_t<i.value, AsLayout>>;
 
            static_assert(is_same<ALayout_, ALayout>::value, "");
        });
 
        static_for<0, NumBTensor, 1>{}([&](auto i) {
            using BLayout_ = remove_cvref_t<tuple_element_t<i.value, BsLayout>>;
 
            static_assert(is_same<BLayout_, BLayout>::value, "");
        });
 
        static_for<0, NumDTensor, 1>{}([&](auto i) {
            using DLayout_ = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
 
            static_assert(is_same<DLayout_, CLayout>::value, "");
        });
 
        return Argument{p_as,
                        p_bs,
                        p_ds,
                        p_e,
                        MRaw,
                        NRaw,
                        KRaw,
                        StrideAs,
                        StrideBs,
                        StrideDs,
                        StrideE,
                        1,
                        a_element_op,
                        b_element_op,
                        c_element_op};
    }
 
    static auto MakeInvoker() { return Invoker{}; }
 
    // polymorphic
    std::unique_ptr<BaseArgument> MakeArgumentPointer(std::array<const void*, NumATensor> p_as,
                                                      std::array<const void*, NumBTensor> p_bs,
                                                      std::array<const void*, NumDTensor> p_ds,
                                                      void* p_e,
                                                      index_t MRaw,
                                                      index_t NRaw,
                                                      index_t KRaw,
                                                      std::array<ck::index_t, NumATensor> StrideAs,
                                                      std::array<ck::index_t, NumBTensor> StrideBs,
                                                      std::array<ck::index_t, NumDTensor> StrideDs,
                                                      index_t StrideE,
                                                      AElementwiseOperation a_element_op,
                                                      BElementwiseOperation b_element_op,
                                                      CElementwiseOperation c_element_op) override
    {
        return std::make_unique<Argument>(p_as,
                                          p_bs,
                                          p_ds,
                                          p_e,
                                          MRaw,
                                          NRaw,
                                          KRaw,
                                          StrideAs,
                                          StrideBs,
                                          StrideDs,
                                          StrideE,
                                          1,
                                          a_element_op,
                                          b_element_op,
                                          c_element_op);
    }
 
    // polymorphic
    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
    {
        return std::make_unique<Invoker>(Invoker{});
    }
 
    // polymorphic
    std::string GetTypeString() const override
    {
        auto str = std::stringstream();
 
        std::map<BlockGemmPipelineScheduler, std::string> BlkGemmPipelineSchedulerToString{
            {BlockGemmPipelineScheduler::Intrawave, "Intrawave"},
            {BlockGemmPipelineScheduler::Interwave, "Interwave"}};
 
        std::map<BlockGemmPipelineVersion, std::string> BlkGemmPipelineVersionToString{
            {BlockGemmPipelineVersion::v1, "v1"},
            {BlockGemmPipelineVersion::v2, "v2"},
            {BlockGemmPipelineVersion::v3, "v3"},
            {BlockGemmPipelineVersion::v4, "v4"},
            {BlockGemmPipelineVersion::v5, "v5"}};
 
        // clang-format off
        str << "DeviceGemmXdlUniversal"
            << "<"
            << getGemmSpecializationString(GemmSpec) << ", "
            << std::string(ALayout::name)[0]
            << std::string(BLayout::name)[0]
            << std::string(CLayout::name)[0]
            << ">"
            << " BlkSize: "
            << BlockSize << ", "
            << "BlkTile: "
            << MPerBlock<<"x"<<NPerBlock<<"x"<<KPerBlock << ", "
            << "WaveTile: "
            << MPerXDL<<"x"<<NPerXDL << ", "
            << "WaveMap: "
            << MXdlPerWave<<"x" << NXdlPerWave<<", "
            << "VmemReadVec: "
            << ABlockTransferSrcScalarPerVector<<"x"<<BBlockTransferSrcScalarPerVector<<", "
            << "BlkGemmPipelineScheduler: "
            << BlkGemmPipelineSchedulerToString[BlkGemmPipeSched] << ", "
            << "BlkGemmPipelineVersion: "
            << BlkGemmPipelineVersionToString[BlkGemmPipelineVer] << ", "
            << "BlkGemmPipelinePrefetchStages: "
            << GridwiseGemm64::BlockwiseGemmPipe::PrefetchStages;
        // clang-format on
 
        return str.str();
    }
};
 
} // namespace device
} // namespace tensor_operation
} // namespace ck