gridwise_gemm_xdl_cshuffle_v2.hpp Source File

gridwise_gemm_xdl_cshuffle_v2.hpp Source File#

Composable Kernel: gridwise_gemm_xdl_cshuffle_v2.hpp Source File
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
 
#pragma once
 
#include "ck/utility/common_header.hpp"
#include "ck/tensor_description/multi_index_transform_helper.hpp"
#include "ck/tensor_description/tensor_descriptor.hpp"
#include "ck/tensor_description/tensor_descriptor_helper.hpp"
#include "ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp"
#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_selector.hpp"
#include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops.hpp"
#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp"
#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r1.hpp"
#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
 
namespace ck {
 
template <typename GridwiseGemm, bool HasMainKBlockLoop, index_t TailNum = 3>
__global__ void
#if CK_USE_LAUNCH_BOUNDS
__launch_bounds__(CK_MAX_THREAD_PER_BLOCK, 1)
#endif
    // __attribute__((amdgpu_waves_per_eu(1, 1)))
    kernel_gemm_xdl_cshuffle_v2(typename GridwiseGemm::Argument karg)
{
#if defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx94__) || defined(__gfx11__) || \
    defined(__gfx12__)
    if constexpr(GridwiseGemm::template IsValidCompilationParameter<>())
    {
        // Pass two lds pointer is the key to tell compiler that ds_read/write
        // operate on different lds chunk at same time without order dependecy
        __shared__ char p_shared_0[GridwiseGemm::GetSharedMemoryNumberOfByte()];
        __shared__ char p_shared_1[GridwiseGemm::GetSharedMemoryNumberOfByte()];
 
        GridwiseGemm::template Run<HasMainKBlockLoop, TailNum>(
            karg.p_a_grid, karg.p_b_grid, karg.p_c_grid, p_shared_0, p_shared_1, karg);
    }
#else
    ignore = karg;
#endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
}
 
template <typename GridwiseGemm,
          typename FloatA,
          typename FloatB,
          typename FloatC,
          bool HasMainKBlockLoop>
__global__ void
#if CK_USE_LAUNCH_BOUNDS
__launch_bounds__(CK_MAX_THREAD_PER_BLOCK, 1)
#endif
    kernel_gemm_xdl_cshuffle_v2(const FloatA* p_a_grid,
                                const FloatB* p_b_grid,
                                FloatC* p_c_grid,
                                typename GridwiseGemm::Problem problem)
{
#if defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx94__) || defined(__gfx11__) || \
    defined(__gfx12__)
    if constexpr(GridwiseGemm::template IsValidCompilationParameter<>())
    {
        __shared__ char p_shared_0[GridwiseGemm::GetSharedMemoryNumberOfByte()];
        __shared__ char p_shared_1[GridwiseGemm::GetSharedMemoryNumberOfByte()];
 
        GridwiseGemm::template Run<HasMainKBlockLoop>(
            p_a_grid, p_b_grid, p_c_grid, p_shared_0, p_shared_1, problem);
    }
#else
    ignore = p_a_grid;
    ignore = p_b_grid;
    ignore = p_c_grid;
    ignore = problem;
#endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
}
 
template <typename ALayout,
          typename BLayout,
          typename CLayout,
          typename FloatA,
          typename FloatB,
          typename FloatGemmAcc,
          typename FloatCShuffle,
          typename FloatC,
          typename AElementwiseOperation,
          typename BElementwiseOperation,
          typename CElementwiseOperation,
          tensor_operation::device::GemmSpecialization GemmSpec,
          InMemoryDataOperationEnum CGlobalMemoryDataOperation,
          index_t NumGemmKPrefetchStage,
          index_t BlockSize,
          index_t MPerBlock,
          index_t NPerBlock,
          index_t KPerBlock,
          index_t AK1Value,
          index_t BK1Value,
          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 AThreadTransferSrcResetCoordinateAfterRun,
          index_t ABlockLdsExtraM,
          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
          typename BBlockTransferThreadClusterArrangeOrder,
          typename BBlockTransferSrcAccessOrder,
          index_t BBlockTransferSrcVectorDim,
          index_t BBlockTransferSrcScalarPerVector,
          index_t BBlockTransferDstScalarPerVector_BK1,
          bool BThreadTransferSrcResetCoordinateAfterRun,
          index_t BBlockLdsExtraN,
          index_t CShuffleMXdlPerWavePerShuffle,
          index_t CShuffleNXdlPerWavePerShuffle,
          typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
          index_t CShuffleBlockTransferScalarPerVector_NPerBlock,
          LoopScheduler LoopSched,
          PipelineVersion PipelineVer = PipelineVersion::v1,
          typename ComputeTypeA       = FloatC,
          typename ComputeTypeB       = ComputeTypeA>
struct GridwiseGemm_xdl_cshuffle_v2
{
    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 I4 = Number<4>{};
    static constexpr auto I5 = Number<5>{};
    static constexpr auto I6 = Number<6>{};
    static constexpr auto I7 = Number<7>{};
 
    // K1 should be Number<...>
    static constexpr auto AK0Number = Number<KPerBlock / AK1Value>{};
    static constexpr auto BK0Number = Number<KPerBlock / BK1Value>{};
    static constexpr auto AK1Number = Number<AK1Value>{};
    static constexpr auto BK1Number = Number<BK1Value>{};
 
    using ThisThreadBlock = ThisThreadBlock<BlockSize>;
 
    __host__ static auto CalculateGridSize(index_t M, index_t N)
    {
        return std::make_tuple(Block2CTileMap::CalculateGridSize(M, N), 1, 1);
    }
 
    __host__ static auto CalculateMPadded(index_t M)
    {
        return math::integer_divide_ceil(M, MPerBlock) * MPerBlock;
    }
 
    __host__ static auto CalculateNPadded(index_t N)
    {
        return math::integer_divide_ceil(N, NPerBlock) * NPerBlock;
    }
 
    __host__ static auto CalculateKPadded(index_t K)
    {
        return math::integer_divide_ceil(K, KPerBlock) * KPerBlock;
    }
 
    __host__ static auto CalculateAK0(index_t K)
    {
        using GemmSpecialization = tensor_operation::device::GemmSpecialization;
 
        if constexpr(GemmSpec == GemmSpecialization::MKPadding ||
                     GemmSpec == GemmSpecialization::MNKPadding ||
                     GemmSpec == GemmSpecialization::KPadding ||
                     GemmSpec == GemmSpecialization::NKPadding)
        {
            return CalculateKPadded(K) / AK1Value;
        }
        else
        {
            return K / AK1Value;
        }
    }
 
    __host__ static auto CalculateBK0(index_t K)
    {
        using GemmSpecialization = tensor_operation::device::GemmSpecialization;
 
        if constexpr(GemmSpec == GemmSpecialization::NKPadding ||
                     GemmSpec == GemmSpecialization::MNKPadding ||
                     GemmSpec == GemmSpecialization::KPadding ||
                     GemmSpec == GemmSpecialization::MKPadding)
        {
            return CalculateKPadded(K) / BK1Value;
        }
        else
        {
            return K / BK1Value;
        }
    }
 
    __host__ static auto CalculateMBlock(index_t M)
    {
        return math::integer_divide_floor(M, MPerBlock);
    }
 
    __host__ static auto CalculateNBlock(index_t N)
    {
        return math::integer_divide_floor(N, NPerBlock);
    }
 
    template <index_t MNXdlPerWave, index_t MNWaves, index_t MNPerXdl, typename TileDesc_K0_MN_K1>
    __host__ __device__ static constexpr auto MakeGemmMmaTileDescriptor(const TileDesc_K0_MN_K1&)
    {
        constexpr index_t K0 = TileDesc_K0_MN_K1{}.GetLength(Number<0>{});
        constexpr index_t K1 = TileDesc_K0_MN_K1{}.GetLength(Number<2>{});
 
        return transform_tensor_descriptor(
            TileDesc_K0_MN_K1{},
            make_tuple(make_merge_transform_v3_division_mod(make_tuple(Number<K0>{}, Number<K1>{})),
                       make_unmerge_transform(make_tuple(
                           Number<MNXdlPerWave>{}, Number<MNWaves>{}, Number<MNPerXdl>{}))),
            make_tuple(Sequence<0, 2>{}, Sequence<1>{}),
            make_tuple(Sequence<3>{}, Sequence<0, 1, 2>{}));
    }
 
    __device__ static auto MakeAGridDescriptor_AK0_M_AK1(
        index_t M, index_t MPad, index_t K, index_t KPad, index_t StrideA, index_t AK0)
    {
        const auto a_grid_desc_mraw_kraw = [&]() {
            if constexpr(is_same_v<tensor_layout::gemm::RowMajor, ALayout>)
            {
                return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(StrideA, I1));
            }
            else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, ALayout>)
            {
                return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(I1, StrideA));
            }
        }();
 
        using GemmSpecialization = tensor_operation::device::GemmSpecialization;
 
        if constexpr(GemmSpec == GemmSpecialization::MKPadding ||
                     GemmSpec == GemmSpecialization::MNKPadding)
        {
            // pad both M and K
            const auto a_grid_desc_m_k =
                transform_tensor_descriptor(a_grid_desc_mraw_kraw,
                                            make_tuple(make_right_pad_transform(M, MPad - M),
                                                       make_right_pad_transform(K, KPad - K)),
                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
 
            const auto a_grid_desc_ak0_m_ak1 = transform_tensor_descriptor(
                a_grid_desc_m_k,
                make_tuple(make_unmerge_transform(make_tuple(AK0, AK1Value)),
                           make_pass_through_transform(MPad)),
                make_tuple(Sequence<1>{}, Sequence<0>{}),
                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
 
            return a_grid_desc_ak0_m_ak1;
        }
        else if constexpr(GemmSpec == GemmSpecialization::MPadding ||
                          GemmSpec == GemmSpecialization::MNPadding)
        {
            // pad M, but not K
            const auto a_grid_desc_ak0_m_ak1 = transform_tensor_descriptor(
                a_grid_desc_mraw_kraw,
                make_tuple(make_unmerge_transform(make_tuple(AK0, AK1Value)),
                           make_right_pad_transform(M, MPad - M)),
                make_tuple(Sequence<1>{}, Sequence<0>{}),
                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
 
            return a_grid_desc_ak0_m_ak1;
        }
        else if constexpr(GemmSpec == GemmSpecialization::KPadding ||
                          GemmSpec == GemmSpecialization::NKPadding)
        {
            // pad K, but not M
            const auto a_grid_desc_m_k = transform_tensor_descriptor(
                a_grid_desc_mraw_kraw,
                make_tuple(make_pass_through_transform(M), make_right_pad_transform(K, KPad - K)),
                make_tuple(Sequence<0>{}, Sequence<1>{}),
                make_tuple(Sequence<0>{}, Sequence<1>{}));
 
            const auto a_grid_desc_ak0_m_ak1 = transform_tensor_descriptor(
                a_grid_desc_m_k,
                make_tuple(make_unmerge_transform(make_tuple(AK0, AK1Value)),
                           make_pass_through_transform(M)),
                make_tuple(Sequence<1>{}, Sequence<0>{}),
                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
 
            return a_grid_desc_ak0_m_ak1;
        }
        else
        {
            // not pad M or K
            const auto a_grid_desc_ak0_m_ak1 = transform_tensor_descriptor(
                a_grid_desc_mraw_kraw,
                make_tuple(make_unmerge_transform(make_tuple(AK0, AK1Value)),
                           make_pass_through_transform(M)),
                make_tuple(Sequence<1>{}, Sequence<0>{}),
                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
 
            return a_grid_desc_ak0_m_ak1;
        }
    }
 
    __device__ static auto MakeBGridDescriptor_BK0_N_BK1(
        index_t K, index_t KPad, index_t N, index_t NPad, index_t StrideB, index_t BK0)
    {
        const auto b_grid_desc_nraw_kraw = [&]() {
            if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
            {
                return make_naive_tensor_descriptor(make_tuple(N, K), make_tuple(I1, StrideB));
            }
            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
            {
                return make_naive_tensor_descriptor(make_tuple(N, K), make_tuple(StrideB, I1));
            }
        }();
 
        using GemmSpecialization = tensor_operation::device::GemmSpecialization;
 
        if constexpr(GemmSpec == GemmSpecialization::NKPadding ||
                     GemmSpec == GemmSpecialization::MNKPadding)
        {
            // pad both N and K
            const auto b_grid_desc_n_k =
                transform_tensor_descriptor(b_grid_desc_nraw_kraw,
                                            make_tuple(make_right_pad_transform(N, NPad - N),
                                                       make_right_pad_transform(K, KPad - K)),
                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
 
            const auto b_grid_desc_bk0_n_bk1 = transform_tensor_descriptor(
                b_grid_desc_n_k,
                make_tuple(make_unmerge_transform(make_tuple(BK0, BK1Value)),
                           make_pass_through_transform(NPad)),
                make_tuple(Sequence<1>{}, Sequence<0>{}),
                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
 
            return b_grid_desc_bk0_n_bk1;
        }
        else if constexpr(GemmSpec == GemmSpecialization::NPadding ||
                          GemmSpec == GemmSpecialization::MNPadding)
        {
            // pad N, but not K
            const auto b_grid_desc_bk0_n_bk1 = transform_tensor_descriptor(
                b_grid_desc_nraw_kraw,
                make_tuple(make_unmerge_transform(make_tuple(BK0, BK1Value)),
                           make_right_pad_transform(N, NPad - N)),
                make_tuple(Sequence<1>{}, Sequence<0>{}),
                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
 
            return b_grid_desc_bk0_n_bk1;
        }
        else if constexpr(GemmSpec == GemmSpecialization::KPadding ||
                          GemmSpec == GemmSpecialization::MKPadding)
        {
            // pad K, but not N
            const auto b_grid_desc_n_k = transform_tensor_descriptor(
                b_grid_desc_nraw_kraw,
                make_tuple(make_pass_through_transform(N), make_right_pad_transform(K, KPad - K)),
                make_tuple(Sequence<0>{}, Sequence<1>{}),
                make_tuple(Sequence<0>{}, Sequence<1>{}));
 
            const auto b_grid_desc_bk0_n_bk1 = transform_tensor_descriptor(
                b_grid_desc_n_k,
                make_tuple(make_unmerge_transform(make_tuple(BK0, BK1Value)),
                           make_pass_through_transform(N)),
                make_tuple(Sequence<1>{}, Sequence<0>{}),
                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
 
            return b_grid_desc_bk0_n_bk1;
        }
        else
        {
            // not pad N or K
            const auto b_grid_desc_bk0_n_bk1 = transform_tensor_descriptor(
                b_grid_desc_nraw_kraw,
                make_tuple(make_unmerge_transform(make_tuple(BK0, BK1Value)),
                           make_pass_through_transform(N)),
                make_tuple(Sequence<1>{}, Sequence<0>{}),
                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
 
            return b_grid_desc_bk0_n_bk1;
        }
    }
 
    template <typename ABlockDesc_AK0_M_AK1>
    __host__ __device__ static constexpr auto
    MakeAMmaTileDescriptor_M0_M1_M2_K(const ABlockDesc_AK0_M_AK1&)
    {
        constexpr index_t MWaves = MPerBlock / (MXdlPerWave * MPerXdl);
 
        return MakeGemmMmaTileDescriptor<MXdlPerWave, MWaves, MPerXdl>(ABlockDesc_AK0_M_AK1{});
    }
 
    template <typename BBlockDesc_BK0_N_BK1>
    __host__ __device__ static constexpr auto
    MakeBMmaTileDescriptor_N0_N1_N2_K(const BBlockDesc_BK0_N_BK1&)
    {
        constexpr index_t NWaves = NPerBlock / (NXdlPerWave * NPerXdl);
 
        return MakeGemmMmaTileDescriptor<NXdlPerWave, NWaves, NPerXdl>(BBlockDesc_BK0_N_BK1{});
    }
 
    __host__ __device__ static auto
    MakeCGridDescriptor_M_N(index_t M, index_t MPad, index_t N, index_t NPad, index_t StrideC)
    {
        const auto c_grid_desc_mraw_nraw = [&]() {
            if constexpr(is_same<tensor_layout::gemm::RowMajor, CLayout>::value)
            {
                return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(StrideC, I1));
            }
            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, CLayout>::value)
            {
                return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(I1, StrideC));
            }
        }();
 
        using GemmSpecialization = tensor_operation::device::GemmSpecialization;
 
        if constexpr(GemmSpec == GemmSpecialization::MNPadding ||
                     GemmSpec == GemmSpecialization::MNKPadding)
        {
            // pad M and N
            return transform_tensor_descriptor(c_grid_desc_mraw_nraw,
                                               make_tuple(make_right_pad_transform(M, MPad - M),
                                                          make_right_pad_transform(N, NPad - N)),
                                               make_tuple(Sequence<0>{}, Sequence<1>{}),
                                               make_tuple(Sequence<0>{}, Sequence<1>{}));
        }
        else if constexpr(GemmSpec == GemmSpecialization::MPadding ||
                          GemmSpec == GemmSpecialization::MKPadding)
        {
            // pad M, but not N
            return transform_tensor_descriptor(
                c_grid_desc_mraw_nraw,
                make_tuple(make_right_pad_transform(M, MPad - M), make_pass_through_transform(N)),
                make_tuple(Sequence<0>{}, Sequence<1>{}),
                make_tuple(Sequence<0>{}, Sequence<1>{}));
        }
        else if constexpr(GemmSpec == GemmSpecialization::NPadding ||
                          GemmSpec == GemmSpecialization::NKPadding)
        {
            // pad N, but not M
            return transform_tensor_descriptor(
                c_grid_desc_mraw_nraw,
                make_tuple(make_pass_through_transform(M), make_right_pad_transform(N, NPad - N)),
                make_tuple(Sequence<0>{}, Sequence<1>{}),
                make_tuple(Sequence<0>{}, Sequence<1>{}));
        }
        else
        {
            // not pad M or N
            return c_grid_desc_mraw_nraw;
        }
    }
 
    struct Problem
    {
        __host__ Problem(index_t M_,
                         index_t N_,
                         index_t K_,
                         index_t StrideA_,
                         index_t StrideB_,
                         index_t StrideC_)
            : M{M_},
              N{N_},
              K{K_},
              StrideA{StrideA_},
              StrideB{StrideB_},
              StrideC{StrideC_},
              MPadded{CalculateMPadded(M_)},
              NPadded{CalculateNPadded(N_)},
              KPadded{CalculateKPadded(K_)},
              AK0{CalculateAK0(K_)},
              BK0{CalculateBK0(K_)},
              MBlock{CalculateMBlock(M_)},
              NBlock{CalculateNBlock(N_)}
        {
        }
 
        __host__ void Print() const
        {
            std::cout << "problem {" << "M:" << M << ", " << "N:" << N << ", " << "K:" << K << ", "
                      << "SA:" << StrideA << ", " << "SB:" << StrideB << ", " << "SC:" << StrideC
                      << ", " << "MP:" << MPadded << ", " << "NP:" << NPadded << ", "
                      << "KP:" << KPadded << ", " << "AK0:" << AK0 << ", " << "BK0:" << BK0 << ", "
                      << "MBlock: " << MBlock << ", " << "NBlock: " << NBlock << "}" << std::endl;
        }
 
        index_t M;
        index_t N;
        index_t K;
        index_t StrideA;
        index_t StrideB;
        index_t StrideC;
        index_t MPadded;
        index_t NPadded;
        index_t KPadded;
        index_t AK0;
        index_t BK0;
        index_t MBlock;
        index_t NBlock;
    };
 
    // Argument
    struct Argument : public tensor_operation::device::BaseArgument, public Problem
    {
        __host__ Argument(const FloatA* p_a_grid_,
                          const FloatB* p_b_grid_,
                          FloatC* p_c_grid_,
                          index_t M_,
                          index_t N_,
                          index_t K_,
                          index_t StrideA_,
                          index_t StrideB_,
                          index_t StrideC_)
            : Problem{M_, N_, K_, StrideA_, StrideB_, StrideC_},
              p_a_grid{p_a_grid_},
              p_b_grid{p_b_grid_},
              p_c_grid{p_c_grid_}
        {
        }
 
        const FloatA* p_a_grid;
        const FloatB* p_b_grid;
        FloatC* p_c_grid;
    };
 
    // FIXME: pass GridwiseGemmPipe as a template arguement into GridwiseGemm
    using GridwiseGemmPipe = remove_cvref_t<
        decltype(GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage, LoopSched>())>;
 
    __device__ static constexpr auto GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1()
    {
        // A matrix in LDS memory, dst of blockwise copy
        return make_naive_tensor_descriptor(
            make_tuple(AK0Number, Number<MPerBlock>{}, AK1Number),
            make_tuple(Number<MPerBlock + ABlockLdsExtraM>{} * AK1Number, AK1Number, I1));
    }
 
    __device__ static constexpr auto GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1()
    {
        // B matrix in LDS memory, dst of blockwise copy
        return make_naive_tensor_descriptor(
            make_tuple(BK0Number, Number<NPerBlock>{}, BK1Number),
            make_tuple(Number<NPerBlock + BBlockLdsExtraN>{} * BK1Number, BK1Number, I1));
    }
 
    __device__ static constexpr auto GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock()
    {
        constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
        constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
 
        constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
            make_naive_tensor_descriptor_packed(
                make_tuple(I1,
                           Number<CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl>{},
                           I1,
                           Number<CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>{}));
 
        return c_shuffle_block_desc_mblock_mperblock_nblock_nperblock;
    }
 
    __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
    {
        // LDS allocation for A and B: be careful of alignment
        constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
        constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
 
        // lds max alignment
        constexpr auto max_lds_align = math::lcm(AK1Number, BK1Number);
 
        constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
            a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
 
        constexpr auto b_block_space_size_aligned = math::integer_least_multiple(
            b_block_desc_bk0_n_bk1.GetElementSpaceSize(), max_lds_align);
 
        // LDS allocation for C shuffle in LDS
        constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
            GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
 
        constexpr auto c_block_size =
            c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize();
 
        return math::max((a_block_space_size_aligned * sizeof(ComputeTypeA) +
                          b_block_space_size_aligned * sizeof(ComputeTypeB)),
                         c_block_size * sizeof(FloatCShuffle));
    }
 
    template <
        InMemoryDataOperationEnum CGlobalMemoryDataOperation_ = InMemoryDataOperationEnum::Set>
    __device__ static bool constexpr IsValidCompilationParameter()
    {
        return ck::tensor_operation::device::IsValidGemmCompilationParameter<
            BlockSize,
            MPerBlock,
            NPerBlock,
            MPerXdl,
            NPerXdl,
            MXdlPerWave,
            NXdlPerWave,
            FloatC,
            CGlobalMemoryDataOperation>();
    }
 
    // block_id to matrix tile idx (m0, n0) mapping are controlled by {M01, N01}
    __host__ static constexpr bool CheckValidity(const Problem& problem)
    {
        static_assert((MPerBlock % (MPerXdl * MXdlPerWave) == 0) &&
                          (NPerBlock % (NXdlPerWave * NPerXdl)) == 0,
                      "Invalid tuning param!");
 
        if constexpr(!(GemmSpec == tensor_operation::device::GemmSpecialization::MPadding ||
                       GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding ||
                       GemmSpec == tensor_operation::device::GemmSpecialization::MKPadding ||
                       GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding))
        {
            if(!(problem.M % MPerBlock == 0))
            {
                return false;
            }
        }
 
        if constexpr(!(GemmSpec == tensor_operation::device::GemmSpecialization::NPadding ||
                       GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding ||
                       GemmSpec == tensor_operation::device::GemmSpecialization::NKPadding ||
                       GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding))
        {
            if(!(problem.N % NPerBlock == 0))
            {
                return false;
            }
        }
 
        if constexpr(GemmSpec == tensor_operation::device::GemmSpecialization::MKPadding ||
                     GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding ||
                     GemmSpec == tensor_operation::device::GemmSpecialization::KPadding ||
                     GemmSpec == tensor_operation::device::GemmSpecialization::NKPadding)
        {
            if(!(CalculateKPadded(problem.K) % AK1Value == 0) ||
               !(CalculateKPadded(problem.K) % BK1Value == 0))
            {
                return false;
            }
        }
        else
        {
            if(!(problem.K % AK1Value == 0) || !(problem.K % BK1Value == 0))
            {
                return false;
            }
        }
 
        if constexpr(is_same<tensor_layout::gemm::RowMajor, ALayout>::value)
        {
            if(problem.K % ABlockTransferSrcScalarPerVector != 0)
            {
                return false;
            }
        }
        else
        {
            if(problem.M % ABlockTransferSrcScalarPerVector != 0)
            {
                return false;
            }
        }
 
        if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
        {
            if(problem.N % BBlockTransferSrcScalarPerVector != 0)
            {
                return false;
            }
        }
        else
        {
            if(problem.K % BBlockTransferSrcScalarPerVector != 0)
            {
                return false;
            }
        }
 
        if constexpr(is_same<tensor_layout::gemm::RowMajor, CLayout>::value)
        {
            if(problem.N % CShuffleBlockTransferScalarPerVector_NPerBlock != 0)
            {
                return false;
            }
        }
        else
        {
            if(problem.M % CShuffleBlockTransferScalarPerVector_NPerBlock != 0)
            {
                return false;
            }
        }
 
        // check gridwise gemm pipeline
        const auto num_k_loop = (CalculateAK0(problem.K) * AK1Value) / KPerBlock;
 
        if(num_k_loop < 4)
        {
            return false;
        }
 
        // TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
        return true;
    }
 
    __host__ static constexpr bool CalculateHasMainKBlockLoop(index_t K)
    {
        const index_t num_loop = K / KPerBlock;
 
        return num_loop > 3;
    }
 
    __host__ static constexpr index_t CalculateKBlockLoopTailNum(index_t K)
    {
        const index_t num_loop = K / KPerBlock;
 
        if(num_loop % 2 == 1)
            return 3;
        else
            return 2;
    }
 
    template <typename CGridDesc>
    __device__ static constexpr auto MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
        const CGridDesc& c_grid_desc_m_n, index_t MBlock, index_t NBlock)
    {
        const auto c_grid_desc_mblock_mperblock_nblock_nperblock = transform_tensor_descriptor(
            c_grid_desc_m_n,
            make_tuple(make_unmerge_transform(make_tuple(MBlock, Number<MPerBlock>{})),
                       make_unmerge_transform(make_tuple(NBlock, Number<NPerBlock>{}))),
            make_tuple(Sequence<0>{}, Sequence<1>{}),
            make_tuple(Sequence<0, 1>{}, Sequence<2, 3>{}));
 
        return c_grid_desc_mblock_mperblock_nblock_nperblock;
    }
 
    // return block_id to C matrix tile idx (m0, n0) mapping
    // if arch = gfx942
    using Block2CTileMap = BlockToCTileMap_Grouped_M00_N0_M01Adapt<8, MPerBlock, NPerBlock>;
 
    template <bool HasMainKBlockLoop, index_t TailNum = 3>
    __device__ static void Run(const FloatA* p_a_grid,
                               const FloatB* p_b_grid,
                               FloatC* p_c_grid,
                               void* p_shared_0,
                               void* p_shared_1,
                               const Problem& problem)
    {
        const auto a_grid_desc_ak0_m_ak1 = MakeAGridDescriptor_AK0_M_AK1(
            problem.M, problem.MPadded, problem.K, problem.KPadded, problem.StrideA, problem.AK0);
        const auto b_grid_desc_bk0_n_bk1 = MakeBGridDescriptor_BK0_N_BK1(
            problem.K, problem.KPadded, problem.N, problem.NPadded, problem.StrideB, problem.BK0);
        const auto c_grid_desc_m_n = MakeCGridDescriptor_M_N(
            problem.M, problem.MPadded, problem.N, problem.NPadded, problem.StrideC);
 
        const auto c_grid_desc_mblock_mperblock_nblock_nperblock =
            MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
                c_grid_desc_m_n, problem.MBlock, problem.NBlock);
 
        const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
            p_a_grid, a_grid_desc_ak0_m_ak1.GetElementSpaceSize());
        const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
            p_b_grid, b_grid_desc_bk0_n_bk1.GetElementSpaceSize());
        auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
            p_c_grid, c_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
 
        const AElementwiseOperation a_element_op{};
        const BElementwiseOperation b_element_op{};
        const CElementwiseOperation c_element_op{};
 
        // divide block work by [M, N]
        const auto block_2_ctile_map = Block2CTileMap{problem.M, problem.N, 4};
 
        const auto block_work_idx =
            block_2_ctile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
 
        if(!block_2_ctile_map.ValidCTileIndex(
               block_work_idx,
               make_tuple(c_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I0),
                          c_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I2))))
        {
            return;
        }
#if 0
        if(threadIdx.x == 0){
            printf("Hardware assigned No. %03d workgroup of logical C tile (%02d, %02d) on %d th XCC Die, %d th SE, %d th CU\n",
               get_block_1d_id(),
               block_work_idx[I0],
               block_work_idx[I1],
               __smid()>>6 & 0xf,
               __smid()>>4 & 0x3,
               __smid() & 0xf);
        }
#endif
        // HACK: this force m/n_block_data_idx_on_grid into SGPR
        const index_t m_block_data_idx_on_grid =
            __builtin_amdgcn_readfirstlane(block_work_idx[I0] * MPerBlock);
 
        const index_t n_block_data_idx_on_grid =
            __builtin_amdgcn_readfirstlane(block_work_idx[I1] * NPerBlock);
 
        // lds max alignment
        constexpr auto max_lds_align = math::lcm(AK1Number, BK1Number);
 
        // A matrix in LDS memory, dst of blockwise copy
        constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
 
        // B matrix in LDS memory, dst of blockwise copy
        constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
 
        // A matrix blockwise copy
        auto a_blockwise_copy =
            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
                                                AElementwiseOperation,
                                                ck::tensor_operation::element_wise::PassThrough,
                                                InMemoryDataOperationEnum::Set,
                                                Sequence<AK0Number, MPerBlock, AK1Number>,
                                                ABlockTransferThreadClusterLengths_AK0_M_AK1,
                                                ABlockTransferThreadClusterArrangeOrder,
                                                FloatA,
                                                ComputeTypeA,
                                                decltype(a_grid_desc_ak0_m_ak1),
                                                decltype(a_block_desc_ak0_m_ak1),
                                                ABlockTransferSrcAccessOrder,
                                                Sequence<1, 0, 2>,
                                                ABlockTransferSrcVectorDim,
                                                2,
                                                ABlockTransferSrcScalarPerVector,
                                                ABlockTransferDstScalarPerVector_AK1,
                                                1,
                                                1,
                                                AThreadTransferSrcResetCoordinateAfterRun,
                                                true>(
                a_grid_desc_ak0_m_ak1,
                make_multi_index(0, m_block_data_idx_on_grid, 0),
                a_element_op,
                a_block_desc_ak0_m_ak1,
                make_multi_index(0, 0, 0),
                ck::tensor_operation::element_wise::PassThrough{});
 
        // B matrix blockwise copy
        auto b_blockwise_copy =
            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
                                                BElementwiseOperation,
                                                ck::tensor_operation::element_wise::PassThrough,
                                                InMemoryDataOperationEnum::Set,
                                                Sequence<BK0Number, NPerBlock, BK1Number>,
                                                BBlockTransferThreadClusterLengths_BK0_N_BK1,
                                                BBlockTransferThreadClusterArrangeOrder,
                                                FloatB,
                                                ComputeTypeB,
                                                decltype(b_grid_desc_bk0_n_bk1),
                                                decltype(b_block_desc_bk0_n_bk1),
                                                BBlockTransferSrcAccessOrder,
                                                Sequence<1, 0, 2>,
                                                BBlockTransferSrcVectorDim,
                                                2,
                                                BBlockTransferSrcScalarPerVector,
                                                BBlockTransferDstScalarPerVector_BK1,
                                                1,
                                                1,
                                                BThreadTransferSrcResetCoordinateAfterRun,
                                                true>(
                b_grid_desc_bk0_n_bk1,
                make_multi_index(0, n_block_data_idx_on_grid, 0),
                b_element_op,
                b_block_desc_bk0_n_bk1,
                make_multi_index(0, 0, 0),
                ck::tensor_operation::element_wise::PassThrough{});
 
        // GEMM definition
        //   c_mtx += transpose(a_mtx) * b_mtx
        //     a_mtx[K0PerBlock, MPerBlock] is in LDS
        //     b_mtx[K0PerBlock, NPerBlock] is in LDS
        //     c_mtx[MPerBlock, NPerBlock] is distributed among threads, and saved in
        //       register
        // sanity check
        constexpr auto lcm_AK1_BK1 = math::lcm(AK1Number, BK1Number);
        constexpr bool is_single_rate_mfma =
            (((is_same<ComputeTypeA, half_t>::value || is_same<ComputeTypeA, bhalf_t>::value) &&
              lcm_AK1_BK1 <= 4) ||
             (is_same<ComputeTypeA, int8_t>::value && lcm_AK1_BK1 <= 8) ||
             ((is_same<ComputeTypeA, f8_t>::value || is_same<ComputeTypeA, bf8_t>::value) &&
              lcm_AK1_BK1 < 32))
                ? true
                : false;
        constexpr auto is_scale_mfma = false;
        constexpr index_t KPack      = math::max(lcm_AK1_BK1,
                                            MfmaSelector<ComputeTypeA,
                                                              MPerXdl,
                                                              NPerXdl,
                                                              ComputeTypeA,
                                                              is_single_rate_mfma,
                                                              is_scale_mfma>::selected_mfma.k_per_blk);
 
        // auto blockwise_gemm = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector<
        //     BlockSize,
        //     ComputeType,
        //     FloatGemmAcc,
        //     decltype(a_block_desc_ak0_m_ak1),
        //     decltype(b_block_desc_bk0_n_bk1),
        //     MPerXdl,
        //     NPerXdl,
        //     MXdlPerWave,
        //     NXdlPerWave,
        //     KPack,
        //     LoopSched>();
        auto blockwise_gemm_pipeline = BlockwiseGemmXdlops_pipeline_v4<
            BlockSize,
            ComputeTypeA,
            FloatGemmAcc,
            decltype(a_block_desc_ak0_m_ak1),
            decltype(b_block_desc_bk0_n_bk1),
            decltype(MakeAMmaTileDescriptor_M0_M1_M2_K(a_block_desc_ak0_m_ak1)),
            decltype(MakeBMmaTileDescriptor_N0_N1_N2_K(b_block_desc_bk0_n_bk1)),
            MPerBlock,
            NPerBlock,
            KPerBlock,
            MPerXdl,
            NPerXdl,
            MXdlPerWave,
            NXdlPerWave,
            KPack>{}; // TransposeC
 
        auto c_thread_buf = blockwise_gemm_pipeline.GetCThreadBuffer();
 
        // LDS allocation for A and B: be careful of alignment
        constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
            a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
 
        auto a_block_buf_ping = make_dynamic_buffer<AddressSpaceEnum::Lds>(
            static_cast<ComputeTypeA*>(p_shared_0), a_block_desc_ak0_m_ak1.GetElementSpaceSize());
 
        auto b_block_buf_ping = make_dynamic_buffer<AddressSpaceEnum::Lds>(
            static_cast<ComputeTypeB*>(p_shared_0) + a_block_space_size_aligned,
            b_block_desc_bk0_n_bk1.GetElementSpaceSize());
 
        auto a_block_buf_pong = make_dynamic_buffer<AddressSpaceEnum::Lds>(
            static_cast<ComputeTypeA*>(p_shared_1), a_block_desc_ak0_m_ak1.GetElementSpaceSize());
 
        auto b_block_buf_pong = make_dynamic_buffer<AddressSpaceEnum::Lds>(
            static_cast<ComputeTypeB*>(p_shared_1) + a_block_space_size_aligned,
            b_block_desc_bk0_n_bk1.GetElementSpaceSize());
 
        auto a_block_bufs = make_tuple(a_block_buf_ping, a_block_buf_pong);
        auto b_block_bufs = make_tuple(b_block_buf_ping, b_block_buf_pong);
 
        constexpr auto a_block_slice_copy_step = make_multi_index(KPerBlock / AK1Number, 0, 0);
        constexpr auto b_block_slice_copy_step = make_multi_index(KPerBlock / BK1Number, 0, 0);
 
        // gridwise GEMM pipeline
        static_assert(std::is_default_constructible_v<GridwiseGemmPipe>);
        // const auto gridwise_gemm_pipeline = GridwiseGemmPipe{};
 
        const index_t num_k_block_main_loop = __builtin_amdgcn_readfirstlane(
            (a_grid_desc_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2)) /
            KPerBlock);
 
        blockwise_gemm_pipeline.template Run<HasMainKBlockLoop, TailNum>(a_grid_desc_ak0_m_ak1,
                                                                         a_block_desc_ak0_m_ak1,
                                                                         a_blockwise_copy,
                                                                         a_grid_buf,
                                                                         a_block_bufs,
                                                                         a_block_slice_copy_step,
                                                                         b_grid_desc_bk0_n_bk1,
                                                                         b_block_desc_bk0_n_bk1,
                                                                         b_blockwise_copy,
                                                                         b_grid_buf,
                                                                         b_block_bufs,
                                                                         b_block_slice_copy_step,
                                                                         c_thread_buf,
                                                                         num_k_block_main_loop);
 
        // shuffle C and write out
        {
            static_assert(MXdlPerWave % CShuffleMXdlPerWavePerShuffle == 0 &&
                              NXdlPerWave % CShuffleNXdlPerWavePerShuffle == 0,
                          "wrong!");
 
            constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
            constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
 
            // TODO: hacky, fix it!
            constexpr auto c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2 =
                blockwise_gemm_pipeline.GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
 
            // TODO: hacky, fix it!
            // c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp is only used to get lengths
            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp =
                blockwise_gemm_pipeline.GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
 
            constexpr auto M0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I0);
            constexpr auto N0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I1);
            constexpr auto M1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I2);
            constexpr auto N1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I3);
            constexpr auto M2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I4);
            constexpr auto M3 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I5);
            constexpr auto M4 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I6);
            constexpr auto N2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I7);
 
            constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
                GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
 
            auto c_shuffle_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
                static_cast<FloatCShuffle*>(p_shared_0),
                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
 
            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2 = transform_tensor_descriptor(
                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
                make_tuple(
                    make_freeze_transform(I0),
                    make_unmerge_transform(make_tuple(
                        Number<CShuffleMXdlPerWavePerShuffle>{}, // M0 (MXdlPerWave) per shuffle
                        M1,                                      // M1 = MWave
                        M2,                                      // M2 * M3 * M4 = MPerXdl
                        M3,
                        M4)),
                    make_freeze_transform(I0),
                    make_unmerge_transform(make_tuple(
                        Number<CShuffleNXdlPerWavePerShuffle>{}, // N0 (NXdlPerWave) per shuffle
                        N1,                                      // N1 = NWave
                        N2))),                                   // N2 = NPerXdl
                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
                make_tuple(
                    Sequence<>{}, Sequence<0, 2, 4, 5, 6>{}, Sequence<>{}, Sequence<1, 3, 7>{}));
 
            // calculate origin of thread output tensor on global memory
            //     blockwise GEMM c matrix starting index
            const auto c_thread_mtx_on_block =
                blockwise_gemm_pipeline.CalculateCThreadOriginDataIndex(I0, I0, I0, I0);
 
            const index_t m_thread_data_on_block = c_thread_mtx_on_block[I0];
            const index_t n_thread_data_on_block = c_thread_mtx_on_block[I1];
 
            const auto m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor =
                make_single_stage_tensor_adaptor(
                    make_tuple(make_merge_transform(make_tuple(M0, M1, M2, M3, M4))),
                    make_tuple(Sequence<0, 1, 2, 3, 4>{}),
                    make_tuple(Sequence<0>{}));
 
            const auto m_thread_data_on_block_idx =
                m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor.CalculateBottomIndex(
                    make_multi_index(m_thread_data_on_block));
 
            const auto n_thread_data_on_block_to_n0_n1_n2_adaptor =
                make_single_stage_tensor_adaptor(
                    make_tuple(make_merge_transform(make_tuple(N0, N1, N2))),
                    make_tuple(Sequence<0, 1, 2>{}),
                    make_tuple(Sequence<0>{}));
 
            const auto n_thread_data_on_block_idx =
                n_thread_data_on_block_to_n0_n1_n2_adaptor.CalculateBottomIndex(
                    make_multi_index(n_thread_data_on_block));
 
            // shuffle: threadwise copy C from VGPR to LDS
            auto c_thread_copy_vgpr_to_lds =
                ThreadwiseTensorSliceTransfer_v1r3<FloatGemmAcc,
                                                   FloatCShuffle,
                                                   decltype(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2),
                                                   decltype(c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2),
                                                   ck::tensor_operation::element_wise::PassThrough,
                                                   Sequence<CShuffleMXdlPerWavePerShuffle,
                                                            CShuffleNXdlPerWavePerShuffle,
                                                            I1,
                                                            I1,
                                                            M2,
                                                            I1,
                                                            M4,
                                                            I1>,
                                                   Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
                                                   7,
                                                   1,
                                                   InMemoryDataOperationEnum::Set,
                                                   1,
                                                   true>{
                    c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
                    make_multi_index(0,
                                     0,
                                     m_thread_data_on_block_idx[I1],
                                     n_thread_data_on_block_idx[I1],
                                     m_thread_data_on_block_idx[I2],
                                     m_thread_data_on_block_idx[I3],
                                     m_thread_data_on_block_idx[I4],
                                     n_thread_data_on_block_idx[I2]),
                    ck::tensor_operation::element_wise::PassThrough{}};
 
            // shuffle: blockwise copy C from LDS to global
            auto c_shuffle_block_copy_lds_to_global = ThreadGroupTensorSliceTransfer_v6r1<
                ThisThreadBlock,            // ThreadGroup
                CElementwiseOperation,      // ElementwiseOperation,
                CGlobalMemoryDataOperation, // DstInMemOp,
                Sequence<1,
                         CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
                         1,
                         CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>, // BlockSliceLengths,
                CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
                Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
                FloatCShuffle,        // typename SrcData,
                FloatC,               // typename DstData,
                decltype(c_shuffle_block_desc_mblock_mperblock_nblock_nperblock),
                decltype(c_grid_desc_mblock_mperblock_nblock_nperblock),
                Sequence<0, 1, 2, 3>,                           // typename DimAccessOrder,
                3,                                              // index_t VectorDim,
                CShuffleBlockTransferScalarPerVector_NPerBlock, // index_t ScalarPerVector,
                true,  // bool ThreadTransferSrcResetCoordinateAfterRun,
                false> // bool ThreadTransferDstResetCoordinateAfterRun>
                {c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
                 make_multi_index(0, 0, 0, 0),
                 c_grid_desc_mblock_mperblock_nblock_nperblock,
                 make_multi_index(block_work_idx[I0], 0, block_work_idx[I1], 0),
                 c_element_op};
 
            // space filling curve for threadwise C in VGPR
            constexpr auto sfc_c_vgpr =
                SpaceFillingCurve<Sequence<MXdlPerWave, NXdlPerWave, 1, 1, M2, 1, M4, 1>,
                                  Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
                                  Sequence<CShuffleMXdlPerWavePerShuffle,
                                           CShuffleNXdlPerWavePerShuffle,
                                           1,
                                           1,
                                           M2,
                                           1,
                                           M4,
                                           1>>{};
 
            // space filling curve for shuffled blockwise C in global mem
            constexpr auto sfc_c_global =
                SpaceFillingCurve<Sequence<1, MPerBlock, 1, NPerBlock>,
                                  Sequence<0, 2, 1, 3>,
                                  Sequence<1,
                                           CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
                                           1,
                                           CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>>{};
 
            constexpr index_t num_access = sfc_c_vgpr.GetNumOfAccess();
 
            static_assert(num_access == sfc_c_global.GetNumOfAccess(), "wrong!");
 
            static_for<0, num_access, 1>{}([&](auto access_id) {
                // make sure it's safe to write to LDS
                block_sync_lds();
 
                // each thread write its data from VGPR to LDS
                c_thread_copy_vgpr_to_lds.Run(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2,
                                              sfc_c_vgpr.GetIndexTupleOfNumber(access_id),
                                              c_thread_buf,
                                              c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
                                              c_shuffle_block_buf);
 
                // make sure it's safe to read from LDS
                block_sync_lds();
 
                // each block copy its data from LDS to global
                c_shuffle_block_copy_lds_to_global.Run(
                    c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
                    c_shuffle_block_buf,
                    c_grid_desc_mblock_mperblock_nblock_nperblock,
                    c_grid_buf);
 
                if constexpr(access_id < num_access - 1)
                {
                    constexpr auto c_global_step = sfc_c_global.GetForwardStep(access_id);
 
                    // move on C
                    c_shuffle_block_copy_lds_to_global.MoveDstSliceWindow(
                        c_grid_desc_mblock_mperblock_nblock_nperblock, c_global_step);
                }
            });
        }
    }
};
 
} // namespace ck