device_batched_gemm_multi_d_xdl.hpp Source File

device_batched_gemm_multi_d_xdl.hpp Source File#

Composable Kernel: device_batched_gemm_multi_d_xdl.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_batched_gemm_multi_d.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
#include "ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_xdl_cshuffle.hpp"
#include "ck/host_utility/device_prop.hpp"
#include "ck/host_utility/kernel_launch.hpp"
#include "ck/host_utility/io.hpp"
 
namespace ck {
namespace tensor_operation {
namespace device {
 
/*
 * \brief Wrapper function of GridwiseGemm::Run to realize BatchedGEMM.
 *
 * \tparam ComputePtrOffsetOfBatch Class that computes the base pointer offsets of A, B, C matrix
 * given the batch. For example, ComputePtrOffsetOfStridedBatch() computes the offsets of evenly
 * strided batched, but we can easily extend to other layouts. The returned offset can be either \p
 * index_t or \p long_index_t. If it returns \p long_index_t, we are not subject to the 2GB
 * limitations.
 *
 * \tparam Block2ETileMap Block2ETileMap::CalculateBottomIndex() takes in id of a workgroup and
 * returns the 2D index of the tile that it computes. \see
 * GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3::Run().
 *
 * \note Using \p ComputePtrOffsetOfBatch gives us the flexibility that 2 workgroups can compute 2
 * tiles from different matrices. Keep in mind that these 2 matrices can share the same grid
 * descriptor (like in BatchedGEMM), or use their own grid descriptors (in GroupedGemm). \link
 * impl/device_conv3d_fwd_xdl_ndhwc_kzyxc_ndhwk.hpp kernel_gemm_xdlops_v2r3_for_conv3d \endlink for
 * \link DeviceConv3d \endlink uses the same concept, but currently does NOT encapsulate the
 * computing of pointer offset into \p ComputePtrOffsetOfStridedBatch.
 *
 * \note \p Block2ETileMap allows customized mapping between a workgroup and the C-tile it computes.
 * Together with \p ComputePtrOffsetOfBatch, we can reuse GridwiseGemm (and GridwiseGemm fusion ) to
 * realize BatchedGemm and GroupedGemm (and the corresponding GEMM fusion).
 *
 */
template <typename GridwiseGemm,
          typename ABDataType,
          typename DsPointer,
          typename EDataType,
          typename AElementwiseOperation,
          typename BElementwiseOperation,
          typename CDEElementwiseOperation,
          typename AGridDesc_AK0_M_AK1,
          typename BGridDesc_BK0_N_BK1,
          typename DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
          typename EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
          typename ComputePtrOffsetOfBatch,
          typename Block2ETileMap,
          bool HasMainKBlockLoop>
__global__ void
#if CK_USE_LAUNCH_BOUNDS
__launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
#endif
    kernel_batched_gemm_xdl(const ABDataType* __restrict__ p_a_grid,
                            const ABDataType* __restrict__ p_b_grid,
                            DsPointer p_ds_grid,
                            EDataType* __restrict__ p_e_grid,
                            const index_t batch_count,
                            const AElementwiseOperation a_element_op,
                            const BElementwiseOperation b_element_op,
                            const CDEElementwiseOperation cde_element_op,
                            const AGridDesc_AK0_M_AK1 a_grid_desc_k0_m_k1,
                            const BGridDesc_BK0_N_BK1 b_grid_desc_k0_n_k1,
                            const DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
                                ds_grid_desc_mblock_mperblock_nblock_nperblock,
                            const EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock
                                e_grid_desc_mblock_mperblock_nblock_nperblock_,
                            const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch,
                            const Block2ETileMap block_2_etile_map)
{
#if defined(__gfx9__) || defined(__gfx11__) || defined(__gfx12__)
    if constexpr(GridwiseGemm::template IsValidCompilationParameter<>())
    {
        const index_t num_blocks_per_batch =
            __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
        const index_t g_idx =
            __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);
 
        const long_index_t a_batch_offset = __builtin_amdgcn_readfirstlane(
            static_cast<long_index_t>(compute_ptr_offset_of_batch.GetAPtrOffset(g_idx)));
        const long_index_t b_batch_offset = __builtin_amdgcn_readfirstlane(
            static_cast<long_index_t>(compute_ptr_offset_of_batch.GetBPtrOffset(g_idx)));
        const long_index_t e_batch_offset = __builtin_amdgcn_readfirstlane(
            static_cast<long_index_t>(compute_ptr_offset_of_batch.GetEPtrOffset(g_idx)));
 
        const auto ds_batch_offset = compute_ptr_offset_of_batch.GetDsPtrOffset(g_idx);
 
        __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
 
        DsPointer p_ds_grid_grp;
 
        static constexpr index_t NumDTensor =
            DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock::Size();
 
        static_for<0, NumDTensor, 1>{}(
            [&](auto i) { p_ds_grid_grp(i) = p_ds_grid[i] + ds_batch_offset[i]; });
 
        GridwiseGemm::template Run<HasMainKBlockLoop, InMemoryDataOperationEnum::Set>(
            p_a_grid + a_batch_offset,
            p_b_grid + b_batch_offset,
            p_ds_grid_grp,
            p_e_grid + e_batch_offset,
            p_shared,
            a_element_op,
            b_element_op,
            cde_element_op,
            a_grid_desc_k0_m_k1,
            b_grid_desc_k0_n_k1,
            ds_grid_desc_mblock_mperblock_nblock_nperblock,
            e_grid_desc_mblock_mperblock_nblock_nperblock_,
            block_2_etile_map);
    }
#else
    ignore = p_a_grid;
    ignore = p_b_grid;
    ignore = p_ds_grid;
    ignore = p_e_grid;
    ignore = batch_count;
    ignore = a_grid_desc_k0_m_k1;
    ignore = b_grid_desc_k0_n_k1;
    ignore = ds_grid_desc_mblock_mperblock_nblock_nperblock;
    ignore = e_grid_desc_mblock_mperblock_nblock_nperblock_;
    ignore = a_element_op;
    ignore = b_element_op;
    ignore = cde_element_op;
    ignore = compute_ptr_offset_of_batch;
    ignore = block_2_etile_map;
#endif
}
 
template <typename ALayout,
          typename BLayout,
          typename DsLayout,
          typename ELayout,
          typename ADataType,
          typename BDataType,
          typename AccDataType,
          typename CShuffleDataType,
          typename DsDataType,
          typename EDataType,
          typename AElementwiseOperation,
          typename BElementwiseOperation,
          typename CDEElementwiseOperation,
          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 CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
          index_t CDEBlockTransferScalarPerVector_NPerBlock,
          LoopScheduler LoopSched = make_default_loop_scheduler()>
struct DeviceBatchedGemmMultiD_Xdl : public DeviceBatchedGemmMultiD<ALayout,
                                                                    BLayout,
                                                                    DsLayout,
                                                                    ELayout,
                                                                    ADataType,
                                                                    BDataType,
                                                                    DsDataType,
                                                                    EDataType,
                                                                    AElementwiseOperation,
                                                                    BElementwiseOperation,
                                                                    CDEElementwiseOperation>
{
    using DeviceOp = DeviceBatchedGemmMultiD_Xdl;
 
    GET_NXDL_PER_WAVE_IMPL
    static constexpr auto NXdlPerWave64 = GetNXdlPerWave<true>();
    static constexpr auto NXdlPerWave32 = GetNXdlPerWave<false>();
 
    static constexpr index_t NumDTensor = DsDataType::Size();
 
    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 matrix_padder =
        MatrixPadder<GemmSpec, index_t, index_t, index_t>{MPerBlock, NPerBlock, KPerBlock};
 
    static auto MakeAGridDescriptor_M_K(index_t MRaw, index_t KRaw, index_t StrideA)
    {
        const auto a_grid_desc_mraw_kraw = [&]() {
            if constexpr(is_same_v<tensor_layout::gemm::RowMajor, ALayout>)
            {
                return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
                                                    make_tuple(StrideA, I1));
            }
            else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, ALayout>)
            {
                return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
                                                    make_tuple(I1, StrideA));
            }
        }();
 
        return matrix_padder.PadADescriptor_M_K(a_grid_desc_mraw_kraw);
    }
 
    static auto MakeBGridDescriptor_N_K(index_t KRaw, index_t NRaw, index_t StrideB)
    {
        const auto b_grid_desc_nraw_kraw = [&]() {
            if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
            {
                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
                                                    make_tuple(I1, StrideB));
            }
            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
            {
                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
                                                    make_tuple(StrideB, I1));
            }
        }();
 
        return matrix_padder.PadBDescriptor_N_K(b_grid_desc_nraw_kraw);
    }
 
    template <typename ELay>
    static auto MakeEGridDescriptor_M_N(index_t MRaw, index_t NRaw, index_t StrideE)
    {
        const auto e_grid_desc_mraw_nraw = [&]() {
            if constexpr(is_same<tensor_layout::gemm::RowMajor, ELay>::value)
            {
                return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
                                                    make_tuple(StrideE, I1));
            }
            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, ELay>::value)
            {
                return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
                                                    make_tuple(I1, StrideE));
            }
        }();
 
        return matrix_padder.PadCDescriptor_M_N(e_grid_desc_mraw_nraw);
    }
 
    static auto MakeDsGridDescriptor_M_N(const std::array<index_t, NumDTensor>& MRaws,
                                         const std::array<index_t, NumDTensor>& NRaws,
                                         const std::array<index_t, NumDTensor>& DsStride)
    {
        return generate_tuple(
            [&](auto i) {
                using DLayout = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
 
                return DeviceOp::MakeEGridDescriptor_M_N<DLayout>(MRaws[i], NRaws[i], DsStride[i]);
            },
            Number<NumDTensor>{});
    }
 
    using AGridDesc_M_K  = decltype(MakeAGridDescriptor_M_K(1, 1, 1));
    using BGridDesc_N_K  = decltype(MakeBGridDescriptor_N_K(1, 1, 1));
    using DsGridDesc_M_N = remove_cvref_t<decltype(MakeDsGridDescriptor_M_N({}, {}, {}))>;
    using EGridDesc_M_N  = decltype(MakeEGridDescriptor_M_N<ELayout>(1, 1, 1));
 
    struct ComputePtrOffsetOfStridedBatch
    {
        ComputePtrOffsetOfStridedBatch(index_t BatchStrideA,
                                       index_t BatchStrideB,
                                       std::array<ck::index_t, NumDTensor> BatchStrideDs,
                                       index_t BatchStrideE)
            : BatchStrideA_(BatchStrideA),
              BatchStrideB_(BatchStrideB),
              BatchStrideDs_(BatchStrideDs),
              BatchStrideE_(BatchStrideE)
        {
        }
 
        __host__ __device__ constexpr long_index_t GetAPtrOffset(index_t g_idx) const
        {
            return g_idx * static_cast<long_index_t>(BatchStrideA_);
        }
 
        __host__ __device__ constexpr long_index_t GetBPtrOffset(index_t g_idx) const
        {
            return g_idx * static_cast<long_index_t>(BatchStrideB_);
        }
 
        __host__ __device__ constexpr auto GetDsPtrOffset(index_t g_idx) const
        {
            std::array<long_index_t, NumDTensor> ds_offset;
            static_for<0, NumDTensor, 1>{}([&](auto i) {
                ds_offset[i] = g_idx * static_cast<long_index_t>(BatchStrideDs_[i]);
            });
            return ds_offset;
        }
 
        __host__ __device__ constexpr long_index_t GetEPtrOffset(index_t g_idx) const
        {
            return g_idx * static_cast<long_index_t>(BatchStrideE_);
        }
 
        private:
        index_t BatchStrideA_;
        index_t BatchStrideB_;
        std::array<ck::index_t, NumDTensor> BatchStrideDs_;
        index_t BatchStrideE_;
    };
 
    using ComputeDataType = ADataType;
 
    template <index_t NXdlPerWave_>
    using GridwiseGemmBase = GridwiseGemmMultipleD_xdl_cshuffle<
        ADataType, // TODO: distinguish A/B datatype
        BDataType,
        ComputeDataType,
        AccDataType,
        CShuffleDataType,
        DsDataType,
        EDataType,
        AElementwiseOperation,
        BElementwiseOperation,
        CDEElementwiseOperation,
        NumGemmKPrefetchStage,
        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,
        CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
        CDEBlockTransferScalarPerVector_NPerBlock,
        LoopSched>;
    using GridwiseGemm64 = GridwiseGemmBase<math::max(NXdlPerWave64, 1)>;
    using GridwiseGemm32 = GridwiseGemmBase<NXdlPerWave32>;
 
    // desc for blockwise copy
    using AGridDesc_AK0_M_AK1 =
        remove_cvref_t<decltype(GridwiseGemm64::MakeDefaultAGridDescriptor_AK0_M_AK1(
            AGridDesc_M_K{}))>;
    using BGridDesc_BK0_N_BK1 =
        remove_cvref_t<decltype(GridwiseGemm64::MakeDefaultBGridDescriptor_BK0_N_BK1(
            BGridDesc_N_K{}))>;
    using DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock = remove_cvref_t<
        decltype(GridwiseGemm64::MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
            DsGridDesc_M_N{}))>;
    using EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock = remove_cvref_t<
        decltype(GridwiseGemm64::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
            EGridDesc_M_N{}))>;
 
    // block-to-e-tile map
    using Block2ETileMap =
        remove_cvref_t<decltype(GridwiseGemm64::MakeDefaultBlock2ETileMap(EGridDesc_M_N{}))>;
 
    // Argument
    struct Argument : public BaseArgument
    {
        template <typename GridwiseGemm>
        void init_ds_e_grid_desc()
        {
            if(GridwiseGemm::CheckValidity(a_grid_desc_m_k_,
                                           b_grid_desc_n_k_,
                                           ds_grid_desc_m_n_,
                                           e_grid_desc_m_n_,
                                           block_2_etile_map_))
            {
                ds_grid_desc_mblock_mperblock_nblock_nperblock_ =
                    GridwiseGemm::MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
                        ds_grid_desc_m_n_);
 
                e_grid_desc_mblock_mperblock_nblock_nperblock_ =
                    GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
                        e_grid_desc_m_n_);
            }
        }
        Argument(const void* p_a_grid,
                 const void* p_b_grid,
                 std::array<const void*, NumDTensor> p_ds_grid,
                 void* p_e_grid,
                 index_t MRaw,
                 index_t NRaw,
                 index_t KRaw,
                 index_t Batch,
                 index_t StrideA,
                 index_t StrideB,
                 const std::array<ck::index_t, NumDTensor>& StrideDs,
                 index_t StrideE,
                 index_t BatchStrideA,
                 index_t BatchStrideB,
                 const std::array<ck::index_t, NumDTensor>& BatchStrideDs,
                 index_t BatchStrideE,
                 AElementwiseOperation a_element_op,
                 BElementwiseOperation b_element_op,
                 CDEElementwiseOperation cde_element_op)
            : p_a_grid_{static_cast<const ADataType*>(p_a_grid)},
              p_b_grid_{static_cast<const BDataType*>(p_b_grid)},
              p_ds_grid_{},
              p_e_grid_{static_cast<EDataType*>(p_e_grid)},
              Batch_(Batch),
              a_grid_desc_m_k_{DeviceOp::MakeAGridDescriptor_M_K(MRaw, KRaw, StrideA)},
              b_grid_desc_n_k_{DeviceOp::MakeBGridDescriptor_N_K(KRaw, NRaw, StrideB)},
              ds_grid_desc_m_n_{},
              e_grid_desc_m_n_{DeviceOp::MakeEGridDescriptor_M_N<ELayout>(MRaw, NRaw, StrideE)},
              a_grid_desc_ak0_m_ak1_{
                  GridwiseGemm64::MakeDefaultAGridDescriptor_AK0_M_AK1(a_grid_desc_m_k_)},
              b_grid_desc_bk0_n_bk1_{
                  GridwiseGemm64::MakeDefaultBGridDescriptor_BK0_N_BK1(b_grid_desc_n_k_)},
              ds_grid_desc_mblock_mperblock_nblock_nperblock_{},
              e_grid_desc_mblock_mperblock_nblock_nperblock_{},
              compute_ptr_offset_of_batch_{BatchStrideA, BatchStrideB, BatchStrideDs, BatchStrideE},
              block_2_etile_map_{GridwiseGemm64::MakeDefaultBlock2ETileMap(e_grid_desc_m_n_)},
              a_element_op_{a_element_op},
              b_element_op_{b_element_op},
              cde_element_op_{cde_element_op}
        {
            // populate pointer, desc for Ds
            static_for<0, NumDTensor, 1>{}([&](auto i) {
                using DLayout   = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
                using DDataType = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
 
                // D pointer
                p_ds_grid_(i) = static_cast<const DDataType*>(p_ds_grid[i]);
 
                // D desc
                ds_grid_desc_m_n_(i) =
                    DeviceOp::MakeEGridDescriptor_M_N<DLayout>(MRaw, NRaw, StrideDs[i]);
            });
 
            // populate desc for Ds/E
            if(get_warp_size() == 64)
            {
                if constexpr(NXdlPerWave64 > 0)
                {
                    init_ds_e_grid_desc<GridwiseGemm64>();
                }
            }
            else
            {
                if constexpr(NXdlPerWave32 > 0)
                {
                    init_ds_e_grid_desc<GridwiseGemm32>();
                }
            }
        }
 
        void Print() const
        {
            std::cout << "A[M, K]: " << a_grid_desc_m_k_ << std::endl;
            std::cout << "B[N, K]: " << b_grid_desc_n_k_ << std::endl;
            static_for<0, NumDTensor, 1>{}(
                [&](auto i) { std::cout << "Ds[M, N]: " << ds_grid_desc_m_n_[i] << std::endl; });
            std::cout << "E[M, N]: " << e_grid_desc_m_n_ << std::endl;
        }
 
        //  private:
        // pointers
        const ADataType* p_a_grid_;
        const BDataType* p_b_grid_;
        typename GridwiseGemm64::DsGridPointer p_ds_grid_;
        EDataType* p_e_grid_;
 
        // Batch
        index_t Batch_;
 
        // tensor descriptors for problem definiton
        AGridDesc_M_K a_grid_desc_m_k_;
        BGridDesc_N_K b_grid_desc_n_k_;
        DsGridDesc_M_N ds_grid_desc_m_n_;
        EGridDesc_M_N e_grid_desc_m_n_;
 
        // tensor descriptors for block/thread-wise copy
        AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1_;
        BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1_;
        DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock
            ds_grid_desc_mblock_mperblock_nblock_nperblock_;
        EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock e_grid_desc_mblock_mperblock_nblock_nperblock_;
 
        // for calculating batch offset
        ComputePtrOffsetOfStridedBatch compute_ptr_offset_of_batch_;
 
        // block-to-e-tile map
        Block2ETileMap block_2_etile_map_;
 
        // element-wise op
        AElementwiseOperation a_element_op_;
        BElementwiseOperation b_element_op_;
        CDEElementwiseOperation cde_element_op_;
    };
 
    // Invoker
    struct Invoker : public BaseInvoker
    {
        using Argument = DeviceBatchedGemmMultiD_Xdl::Argument;
 
        template <typename GridwiseGemm>
        float RunImp(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
        {
            if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_m_k_,
                                            arg.b_grid_desc_n_k_,
                                            arg.ds_grid_desc_m_n_,
                                            arg.e_grid_desc_m_n_,
                                            arg.block_2_etile_map_))
            {
                throw std::runtime_error("wrong! GridwiseGemm has invalid setting");
            }
 
            const index_t grid_size =
                arg.block_2_etile_map_.CalculateGridSize(arg.e_grid_desc_m_n_) * arg.Batch_;
 
            const auto K =
                arg.a_grid_desc_ak0_m_ak1_.GetLength(I0) * arg.a_grid_desc_ak0_m_ak1_.GetLength(I2);
 
            auto launch_kernel = [&](auto has_main_k_block_loop) {
                constexpr bool has_main_loop = has_main_k_block_loop.value;
 
                const auto kernel =
                    kernel_batched_gemm_xdl<GridwiseGemm,
                                            ADataType, // TODO: distiguish A/B datatype
                                            typename GridwiseGemm::DsGridPointer,
                                            EDataType,
                                            AElementwiseOperation,
                                            BElementwiseOperation,
                                            CDEElementwiseOperation,
                                            DeviceOp::AGridDesc_AK0_M_AK1,
                                            DeviceOp::BGridDesc_BK0_N_BK1,
                                            DeviceOp::DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
                                            DeviceOp::EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
                                            ComputePtrOffsetOfStridedBatch,
                                            Block2ETileMap,
                                            has_main_loop>;
 
                return launch_and_time_kernel(stream_config,
                                              kernel,
                                              dim3(grid_size),
                                              dim3(BlockSize),
                                              0,
                                              arg.p_a_grid_,
                                              arg.p_b_grid_,
                                              arg.p_ds_grid_,
                                              arg.p_e_grid_,
                                              arg.Batch_,
                                              arg.a_element_op_,
                                              arg.b_element_op_,
                                              arg.cde_element_op_,
                                              arg.a_grid_desc_ak0_m_ak1_,
                                              arg.b_grid_desc_bk0_n_bk1_,
                                              arg.ds_grid_desc_mblock_mperblock_nblock_nperblock_,
                                              arg.e_grid_desc_mblock_mperblock_nblock_nperblock_,
                                              arg.compute_ptr_offset_of_batch_,
                                              arg.block_2_etile_map_);
            };
 
            if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
            {
                return launch_kernel(integral_constant<bool, true>{});
            }
            else
            {
                return launch_kernel(integral_constant<bool, false>{});
            }
        }
 
        INVOKER_RUN_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 bool IsSupportedArgument(const Argument& arg)
    {
        if(!ck::is_xdl_wmma_supported<ADataType, BDataType, MPerXDL, NPerXDL>())
        {
            return false;
        }
        if(get_warp_size() == 64)
        {
            if constexpr(NXdlPerWave64 > 0)
            {
                return GridwiseGemm64::CheckValidity(arg.a_grid_desc_m_k_,
                                                     arg.b_grid_desc_n_k_,
                                                     arg.ds_grid_desc_m_n_,
                                                     arg.e_grid_desc_m_n_,
                                                     arg.block_2_etile_map_);
            }
        }
        else
        {
            if constexpr(NXdlPerWave32 > 0)
            {
                return GridwiseGemm32::CheckValidity(arg.a_grid_desc_m_k_,
                                                     arg.b_grid_desc_n_k_,
                                                     arg.ds_grid_desc_m_n_,
                                                     arg.e_grid_desc_m_n_,
                                                     arg.block_2_etile_map_);
            }
        }
        return false;
    }
 
    // polymorphic
    bool IsSupportedArgument(const BaseArgument* p_arg) override
    {
        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
    }
 
    static auto MakeArgument(const void* p_a,
                             const void* p_b,
                             const std::array<const void*, NumDTensor>& p_ds,
                             void* p_e,
                             index_t M,
                             index_t N,
                             index_t K,
                             index_t Batch,
                             index_t StrideA,
                             index_t StrideB,
                             const std::array<index_t, NumDTensor>& StrideDs,
                             index_t StrideE,
                             index_t BatchStrideA,
                             index_t BatchStrideB,
                             const std::array<ck::index_t, NumDTensor>& BatchStrideDs,
                             index_t BatchStrideE,
                             AElementwiseOperation a_element_op,
                             BElementwiseOperation b_element_op,
                             CDEElementwiseOperation cde_element_op)
    {
        return Argument{p_a,
                        p_b,
                        p_ds,
                        p_e,
                        M,
                        N,
                        K,
                        Batch,
                        StrideA,
                        StrideB,
                        StrideDs,
                        StrideE,
                        BatchStrideA,
                        BatchStrideB,
                        BatchStrideDs,
                        BatchStrideE,
                        a_element_op,
                        b_element_op,
                        cde_element_op};
    }
 
    static auto MakeInvoker() { return Invoker{}; }
 
    // polymorphic
    std::unique_ptr<BaseArgument>
    MakeArgumentPointer(const void* p_a,
                        const void* p_b,
                        const std::array<const void*, NumDTensor>& p_ds,
                        void* p_e,
                        index_t M,
                        index_t N,
                        index_t K,
                        index_t Batch,
                        index_t StrideA,
                        index_t StrideB,
                        const std::array<ck::index_t, NumDTensor>& StrideDs,
                        index_t StrideE,
                        index_t BatchStrideA,
                        index_t BatchStrideB,
                        const std::array<ck::index_t, NumDTensor>& BatchStrideDs,
                        index_t BatchStrideE,
                        AElementwiseOperation a_element_op,
                        BElementwiseOperation b_element_op,
                        CDEElementwiseOperation cde_element_op) override
    {
        return std::make_unique<Argument>(p_a,
                                          p_b,
                                          p_ds,
                                          p_e,
                                          M,
                                          N,
                                          K,
                                          Batch,
                                          StrideA,
                                          StrideB,
                                          StrideDs,
                                          StrideE,
                                          BatchStrideA,
                                          BatchStrideB,
                                          BatchStrideDs,
                                          BatchStrideE,
                                          a_element_op,
                                          b_element_op,
                                          cde_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();
 
        // clang-format off
        str << "DeviceBatchedGemmMultiD_Xdl"
            << "<"
            << BlockSize << ", "
            << MPerBlock << ", "
            << NPerBlock << ", "
            << KPerBlock << ", "
            << AK1 << ", "
            << BK1 << ", "
            << getGemmSpecializationString(GemmSpec)
            << ">";
        // clang-format on
 
        return str.str();
    }
};
 
} // namespace device
} // namespace tensor_operation
} // namespace ck