gridwise_gemm_xdlops_v2r4.hpp Source File

gridwise_gemm_xdlops_v2r4.hpp Source File#

Composable Kernel: gridwise_gemm_xdlops_v2r4.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_v1.hpp"
#include "ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp"
#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.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,
          typename FloatAB,
          typename FloatC,
          typename ABK0MK1GridDesc,
          typename BBK0NK1GridDesc,
          typename CM0N0M1N1M2M3M4N2GridDesc,
          typename AElementwiseOperation,
          typename BElementwiseOperation,
          typename CElementwiseOperation,
          typename CBlockClusterAdaptor,
          bool HasMainKBlockLoop>
__global__ void
#if CK_USE_LAUNCH_BOUNDS
__launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
#endif
    kernel_gemm_xdlops_v2r4(const FloatAB* __restrict__ p_a_grid,
                            const FloatAB* __restrict__ p_b_grid,
                            FloatC* __restrict__ p_c_grid,
                            const ABK0MK1GridDesc a_b_k0_m_k1_grid_desc,
                            const BBK0NK1GridDesc b_b_k0_n_k1_grid_desc,
                            const CM0N0M1N1M2M3M4N2GridDesc c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc,
                            const AElementwiseOperation a_element_op,
                            const BElementwiseOperation b_element_op,
                            const CElementwiseOperation c_element_op,
                            const CBlockClusterAdaptor c_block_cluster_adaptor)
{
#ifdefined(__gfx908__) || defined(__gfx90a__) || defined(__gfx94__) || defined(__gfx11__) || \
    defined(__gfx12__)
    if constexpr(GridwiseGemm::template IsValidCompilationParameter<>())
    {
        constexpr index_t shared_block_size =
            GridwiseGemm::GetSharedMemoryNumberOfByte() / sizeof(FloatAB);
 
        __shared__ FloatAB p_shared_block[shared_block_size];
 
        GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid,
                                                      p_b_grid,
                                                      p_c_grid,
                                                      p_shared_block,
                                                      a_b_k0_m_k1_grid_desc,
                                                      b_b_k0_n_k1_grid_desc,
                                                      c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc,
                                                      a_element_op,
                                                      b_element_op,
                                                      c_element_op,
                                                      c_block_cluster_adaptor);
    }
#else
    ignore = p_a_grid;
    ignore = p_b_grid;
    ignore = p_c_grid;
    ignore = a_b_k0_m_k1_grid_desc;
    ignore = b_b_k0_n_k1_grid_desc;
    ignore = c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc;
    ignore = a_element_op;
    ignore = b_element_op;
    ignore = c_element_op;
    ignore = c_block_cluster_adaptor;
#endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
}
 
template <index_t BlockSize,
          typename FloatAB,
          typename FloatAcc,
          typename FloatC,
          InMemoryDataOperationEnum CGlobalMemoryDataOperation,
          typename ABK0MK1GridDesc,
          typename BBK0NK1GridDesc,
          typename CMNGridDesc,
          typename AElementwiseOperation,
          typename BElementwiseOperation,
          typename CElementwiseOperation,
          index_t MPerBlock,
          index_t NPerBlock,
          index_t K0PerBlock,
          index_t MPerXDL,
          index_t NPerXDL,
          index_t K1Value,
          index_t MRepeat,
          index_t NRepeat,
          typename ABlockTransferThreadClusterLengths_K0_M_K1,
          typename ABlockTransferThreadClusterArrangeOrder,
          typename ABlockTransferSrcAccessOrder,
          index_t ABlockTransferSrcVectorDim,
          index_t ABlockTransferSrcScalarPerVector,
          index_t ABlockTransferDstScalarPerVector_K1,
          bool AThreadTransferSrcResetCoordinateAfterRun,
          bool ABlockLdsExtraM,
          typename BBlockTransferThreadClusterLengths_K0_N_K1,
          typename BBlockTransferThreadClusterArrangeOrder,
          typename BBlockTransferSrcAccessOrder,
          index_t BBlockTransferSrcVectorDim,
          index_t BBlockTransferSrcScalarPerVector,
          index_t BBlockTransferDstScalarPerVector_K1,
          bool BThreadTransferSrcResetCoordinateAfterRun,
          bool BBlockLdsExtraN,
          typename CThreadTransferSrcDstAccessOrder,
          index_t CThreadTransferSrcDstVectorDim,
          index_t CThreadTransferDstScalarPerVector>
struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4
{
    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 K1 = Number<K1Value>{};
 
    using ThisThreadBlock = ThisThreadBlock<BlockSize>;
 
    __host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
    {
        constexpr auto max_lds_align = K1;
 
        // A matrix in LDS memory, dst of blockwise copy
        constexpr auto a_k0_m_k1_block_desc = [&]() {
            if constexpr(ABlockLdsExtraM)
            {
                return make_naive_tensor_descriptor(
                    make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1),
                    make_tuple(Number<MPerBlock + 1>{} * K1, K1, I1));
            }
            else
            {
                return make_naive_tensor_descriptor_aligned(
                    make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1), max_lds_align);
            }
        }();
 
        // B matrix in LDS memory, dst of blockwise copy
        constexpr auto b_k0_n_k1_block_desc = [&]() {
            if constexpr(BBlockLdsExtraN)
            {
                return make_naive_tensor_descriptor(
                    make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1),
                    make_tuple(Number<NPerBlock + 1>{} * K1, K1, I1));
            }
            else
            {
                return make_naive_tensor_descriptor_aligned(
                    make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1), max_lds_align);
            }
        }();
 
        // LDS allocation for A and B: be careful of alignment
        constexpr auto a_block_space_size =
            math::integer_least_multiple(a_k0_m_k1_block_desc.GetElementSpaceSize(), max_lds_align);
 
        constexpr auto b_block_space_size =
            math::integer_least_multiple(b_k0_n_k1_block_desc.GetElementSpaceSize(), max_lds_align);
 
        return (a_block_space_size + b_block_space_size) * sizeof(FloatAB);
    }
 
    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}
    template <typename Block2CTileMap>
    __host__ __device__ static constexpr bool
    CheckValidity(const ABK0MK1GridDesc& a_b_k0_m_k1_grid_desc,
                  const BBK0NK1GridDesc& b_b_k0_n_k1_grid_desc,
                  const CMNGridDesc& c_m_n_grid_desc,
                  const Block2CTileMap& block_2_ctile_map)
    {
        static_assert(is_known_at_compile_time<remove_cv_t<decltype(K1)>>::value,
                      "wrong! K1 need to be known at compile-time");
 
        static_assert((MPerBlock % (MPerXDL * MRepeat) == 0) &&
                          (NPerBlock % (NRepeat * NPerXDL)) == 0,
                      "Invalid tuning param!");
 
        const auto M      = a_b_k0_m_k1_grid_desc.GetLength(I2);
        const auto N      = b_b_k0_n_k1_grid_desc.GetLength(I2);
        const auto K0     = a_b_k0_m_k1_grid_desc.GetLength(I1);
        const auto KBatch = a_b_k0_m_k1_grid_desc.GetLength(I0);
 
        if(!(M == c_m_n_grid_desc.GetLength(I0) && N == c_m_n_grid_desc.GetLength(I1) &&
             K0 == b_b_k0_n_k1_grid_desc.GetLength(I1) &&
             K1 == a_b_k0_m_k1_grid_desc.GetLength(I3) &&
             K1 == b_b_k0_n_k1_grid_desc.GetLength(I3) &&
             KBatch == b_b_k0_n_k1_grid_desc.GetLength(I0)))
            return false;
 
        if(!(M % MPerBlock == 0 && N % NPerBlock == 0 && K0 % K0PerBlock == 0))
            return false;
 
        if(!block_2_ctile_map.CheckValidity(c_m_n_grid_desc))
        {
            return false;
        }
 
        // TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
        return true;
    }
 
    __host__ __device__ static constexpr bool CalculateHasMainK0BlockLoop(index_t K0)
    {
        const bool has_main_k0_block_loop = K0 > K0PerBlock;
 
        return has_main_k0_block_loop;
    }
 
    __host__ __device__ static constexpr auto
    MakeCM0N0M1N1M2M3M4N2GridDescriptor(const CMNGridDesc& c_m_n_grid_desc)
    {
        constexpr auto max_lds_align = K1;
 
        // A matrix in LDS memory, dst of blockwise copy
        constexpr auto a_k0_m_k1_block_desc = [&]() {
            if constexpr(ABlockLdsExtraM)
            {
                return make_naive_tensor_descriptor(
                    make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1),
                    make_tuple(Number<MPerBlock + 1>{} * K1, K1, I1));
            }
            else
            {
                return make_naive_tensor_descriptor_aligned(
                    make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1), max_lds_align);
            }
        }();
 
        // B matrix in LDS memory, dst of blockwise copy
        constexpr auto b_k0_n_k1_block_desc = [&]() {
            if constexpr(BBlockLdsExtraN)
            {
                return make_naive_tensor_descriptor(
                    make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1),
                    make_tuple(Number<NPerBlock + 1>{} * K1, K1, I1));
            }
            else
            {
                return make_naive_tensor_descriptor_aligned(
                    make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1), max_lds_align);
            }
        }();
 
        using BlockwiseGemm =
            BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<BlockSize,
                                                                FloatAB,
                                                                FloatAcc,
                                                                decltype(a_k0_m_k1_block_desc),
                                                                decltype(b_k0_n_k1_block_desc),
                                                                MPerXDL,
                                                                NPerXDL,
                                                                MRepeat,
                                                                NRepeat,
                                                                K1>;
 
        return BlockwiseGemm::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(c_m_n_grid_desc);
    }
 
    // return block_id to C matrix tile idx (m0, n0) mapping
    __host__ __device__ static constexpr auto MakeCBlockClusterAdaptor(
        const CMNGridDesc& c_m_n_grid_desc, index_t /* M01 */, index_t /* N01 */, index_t KBatch)
    {
        return BlockToCTileMap_KSplit_M00_N0_M01Adapt<MPerBlock, NPerBlock, CMNGridDesc>(
            c_m_n_grid_desc, 8, KBatch);
    }
 
    using CM0N0M1N1M2M3M4N2GridDesc = decltype(MakeCM0N0M1N1M2M3M4N2GridDescriptor(CMNGridDesc{}));
    using CBlockClusterAdaptor      = decltype(MakeCBlockClusterAdaptor(CMNGridDesc{}, 1, 1, 1));
 
    template <bool HasMainKBlockLoop>
    __device__ static void Run(const FloatAB* __restrict__ p_a_grid,
                               const FloatAB* __restrict__ p_b_grid,
                               FloatC* __restrict__ p_c_grid,
                               FloatAB* __restrict__ p_shared_block,
                               const ABK0MK1GridDesc& a_b_k0_m_k1_grid_desc,
                               const BBK0NK1GridDesc& b_b_k0_n_k1_grid_desc,
                               const CM0N0M1N1M2M3M4N2GridDesc& c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc,
                               const AElementwiseOperation& a_element_op,
                               const BElementwiseOperation& b_element_op,
                               const CElementwiseOperation& c_element_op,
                               const CBlockClusterAdaptor& c_block_cluster_adaptor)
    {
        const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
            p_a_grid, a_b_k0_m_k1_grid_desc.GetElementSpaceSize());
        const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
            p_b_grid, b_b_k0_n_k1_grid_desc.GetElementSpaceSize());
        auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
            p_c_grid, c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc.GetElementSpaceSize());
 
        const auto K0 = a_b_k0_m_k1_grid_desc.GetLength(I1);
 
        // divide block work by [M, N]
        const auto block_work_idx =
            c_block_cluster_adaptor.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
 
        if(!c_block_cluster_adaptor.ValidCTileIndex(
               make_tuple(block_work_idx[I1], block_work_idx[I2]),
               make_tuple(c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc.GetLength(I0),
                          c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc.GetLength(I1))))
        {
            return;
        }
 
        const index_t k_batch_id = block_work_idx[I0];
 
        // 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[I1] * MPerBlock);
 
        const index_t n_block_data_idx_on_grid =
            __builtin_amdgcn_readfirstlane(block_work_idx[I2] * NPerBlock);
 
        // lds max alignment
        constexpr auto max_lds_align = K1;
 
        // A matrix in LDS memory, dst of blockwise copy
        constexpr auto a_k0_m_k1_block_desc = [&]() {
            if constexpr(ABlockLdsExtraM)
            {
                return make_naive_tensor_descriptor(
                    make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1),
                    make_tuple(Number<MPerBlock + 1>{} * K1, K1, I1));
            }
            else
            {
                return make_naive_tensor_descriptor_aligned(
                    make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1), max_lds_align);
            }
        }();
 
        constexpr auto a_b_k0_m_k1_block_desc = [&]() {
            if constexpr(ABlockLdsExtraM)
            {
                return make_naive_tensor_descriptor(
                    make_tuple(Number<1>{}, Number<K0PerBlock>{}, Number<MPerBlock>{}, K1),
                    make_tuple(Number<K0PerBlock>{} * Number<MPerBlock + 1>{} * K1,
                               Number<MPerBlock + 1>{} * K1,
                               K1,
                               I1));
            }
            else
            {
                return make_naive_tensor_descriptor_aligned(
                    make_tuple(Number<1>{}, Number<K0PerBlock>{}, Number<MPerBlock>{}, K1),
                    max_lds_align);
            }
        }();
        // B matrix in LDS memory, dst of blockwise copy
        constexpr auto b_k0_n_k1_block_desc = [&]() {
            if constexpr(BBlockLdsExtraN)
            {
                return make_naive_tensor_descriptor(
                    make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1),
                    make_tuple(Number<NPerBlock + 1>{} * K1, K1, I1));
            }
            else
            {
                return make_naive_tensor_descriptor_aligned(
                    make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1), max_lds_align);
            }
        }();
 
        constexpr auto b_b_k0_n_k1_block_desc = [&]() {
            if constexpr(BBlockLdsExtraN)
            {
                return make_naive_tensor_descriptor(
                    make_tuple(Number<1>{}, Number<K0PerBlock>{}, Number<NPerBlock>{}, K1),
                    make_tuple(Number<K0PerBlock>{} * Number<NPerBlock + 1>{} * K1,
                               Number<NPerBlock + 1>{} * K1,
                               K1,
                               I1));
            }
            else
            {
                return make_naive_tensor_descriptor_aligned(
                    make_tuple(Number<1>{}, Number<K0PerBlock>{}, Number<NPerBlock>{}, K1),
                    max_lds_align);
            }
        }();
        // A matrix blockwise copy
        auto a_blockwise_copy =
            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
                                                AElementwiseOperation,
                                                ck::tensor_operation::element_wise::PassThrough,
                                                InMemoryDataOperationEnum::Set,
                                                Sequence<1, K0PerBlock, MPerBlock, K1>,
                                                ABlockTransferThreadClusterLengths_K0_M_K1,
                                                ABlockTransferThreadClusterArrangeOrder,
                                                FloatAB,
                                                FloatAB,
                                                decltype(a_b_k0_m_k1_grid_desc),
                                                decltype(a_b_k0_m_k1_block_desc),
                                                ABlockTransferSrcAccessOrder,
                                                Sequence<0, 2, 1, 3>,
                                                ABlockTransferSrcVectorDim,
                                                3,
                                                ABlockTransferSrcScalarPerVector,
                                                ABlockTransferDstScalarPerVector_K1,
                                                1,
                                                1,
                                                AThreadTransferSrcResetCoordinateAfterRun,
                                                true>(
                a_b_k0_m_k1_grid_desc,
                make_multi_index(k_batch_id, 0, m_block_data_idx_on_grid, 0),
                a_element_op,
                a_b_k0_m_k1_block_desc,
                make_multi_index(0, 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<1, K0PerBlock, NPerBlock, K1>,
                                                BBlockTransferThreadClusterLengths_K0_N_K1,
                                                BBlockTransferThreadClusterArrangeOrder,
                                                FloatAB,
                                                FloatAB,
                                                decltype(b_b_k0_n_k1_grid_desc),
                                                decltype(b_b_k0_n_k1_block_desc),
                                                BBlockTransferSrcAccessOrder,
                                                Sequence<0, 2, 1, 3>,
                                                BBlockTransferSrcVectorDim,
                                                3,
                                                BBlockTransferSrcScalarPerVector,
                                                BBlockTransferDstScalarPerVector_K1,
                                                1,
                                                1,
                                                BThreadTransferSrcResetCoordinateAfterRun,
                                                true>(
                b_b_k0_n_k1_grid_desc,
                make_multi_index(k_batch_id, 0, n_block_data_idx_on_grid, 0),
                b_element_op,
                b_b_k0_n_k1_block_desc,
                make_multi_index(0, 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
 
        auto blockwise_gemm =
            BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<BlockSize,
                                                                FloatAB,
                                                                FloatAcc,
                                                                decltype(a_k0_m_k1_block_desc),
                                                                decltype(b_k0_n_k1_block_desc),
                                                                MPerXDL,
                                                                NPerXDL,
                                                                MRepeat,
                                                                NRepeat,
                                                                K1>{};
 
        auto c_thread_buf = blockwise_gemm.GetCThreadBuffer();
 
        // LDS allocation for A and B: be careful of alignment
        constexpr auto a_block_space_size =
            math::integer_least_multiple(a_k0_m_k1_block_desc.GetElementSpaceSize(), max_lds_align);
 
        FloatAB* p_a_block = p_shared_block;
        FloatAB* p_b_block = p_shared_block + a_block_space_size;
 
        constexpr auto a_block_slice_copy_step = make_multi_index(0, K0PerBlock, 0, 0);
        constexpr auto b_block_slice_copy_step = make_multi_index(0, K0PerBlock, 0, 0);
 
        auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
            p_a_block, a_k0_m_k1_block_desc.GetElementSpaceSize());
        auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
            p_b_block, b_k0_n_k1_block_desc.GetElementSpaceSize());
 
        // preload data into LDS
        {
            a_blockwise_copy.RunRead(a_b_k0_m_k1_grid_desc, a_grid_buf);
            b_blockwise_copy.RunRead(b_b_k0_n_k1_grid_desc, b_grid_buf);
 
            a_blockwise_copy.RunWrite(a_b_k0_m_k1_block_desc, a_block_buf);
            b_blockwise_copy.RunWrite(b_b_k0_n_k1_block_desc, b_block_buf);
        }
 
        // Initialize C
        c_thread_buf.Clear();
 
        // main body
        if constexpr(HasMainKBlockLoop)
        {
            index_t k0_block_data_begin = 0;
 
            do
            {
                a_blockwise_copy.MoveSrcSliceWindow(a_b_k0_m_k1_grid_desc, a_block_slice_copy_step);
                b_blockwise_copy.MoveSrcSliceWindow(b_b_k0_n_k1_grid_desc, b_block_slice_copy_step);
 
                a_blockwise_copy.RunRead(a_b_k0_m_k1_grid_desc, a_grid_buf);
 
                block_sync_lds();
 
                b_blockwise_copy.RunRead(b_b_k0_n_k1_grid_desc, b_grid_buf);
 
                blockwise_gemm.Run(a_block_buf, b_block_buf, c_thread_buf);
 
                block_sync_lds();
 
                a_blockwise_copy.RunWrite(a_b_k0_m_k1_block_desc, a_block_buf);
                b_blockwise_copy.RunWrite(b_b_k0_n_k1_block_desc, b_block_buf);
 
                k0_block_data_begin += K0PerBlock;
            } while(k0_block_data_begin < (K0 - K0PerBlock));
        }
 
        // tail
        {
            block_sync_lds();
 
            blockwise_gemm.Run(a_block_buf, b_block_buf, c_thread_buf);
        }
 
        // output: register to global memory
        {
            constexpr auto c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc =
                blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
 
            constexpr auto M0 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I0);
            constexpr auto N0 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I1);
            constexpr auto M1 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I2);
            constexpr auto N1 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I3);
            constexpr auto M2 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I4);
            constexpr auto M3 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I5);
            constexpr auto M4 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I6);
            constexpr auto N2 = c_m0_n0_m1_n1_m2_m3_m4_n2_block_desc.GetLength(I7);
 
            constexpr auto c_m0_n0_m1_n1_m2_m3_m4_n2_thread_desc =
                make_naive_tensor_descriptor_packed(make_tuple(
                    Number<M0>{}, Number<N0>{}, I1, I1, Number<M2>{}, I1, Number<M4>{}, I1));
 
            // calculate origin of thread output tensor on global memory
            //     blockwise GEMM c matrix starting index
            const auto c_thread_mtx_on_block =
                blockwise_gemm.CalculateCThreadOriginDataIndex(I0, I0, I0, I0);
 
            const index_t m_thread_data_on_grid =
                m_block_data_idx_on_grid + c_thread_mtx_on_block[I0];
 
            const index_t n_thread_data_on_grid =
                n_block_data_idx_on_grid + c_thread_mtx_on_block[I1];
 
            const auto m_thread_data_on_grid_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_grid_idx =
                m_thread_data_on_grid_to_m0_m1_m2_m3_m4_adaptor.CalculateBottomIndex(
                    make_multi_index(m_thread_data_on_grid));
 
            const auto n_thread_data_on_grid_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_grid_idx =
                n_thread_data_on_grid_to_n0_n1_n2_adaptor.CalculateBottomIndex(
                    make_multi_index(n_thread_data_on_grid));
 
            auto c_thread_copy =
                ThreadwiseTensorSliceTransfer_v1r3<FloatAcc,
                                                   FloatC,
                                                   decltype(c_m0_n0_m1_n1_m2_m3_m4_n2_thread_desc),
                                                   decltype(c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc),
                                                   CElementwiseOperation,
                                                   Sequence<M0, N0, I1, I1, M2, I1, M4, I1>,
                                                   CThreadTransferSrcDstAccessOrder,
                                                   CThreadTransferSrcDstVectorDim,
                                                   CThreadTransferDstScalarPerVector,
                                                   CGlobalMemoryDataOperation,
                                                   1,
                                                   true>{
 
                    c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc,
                    make_multi_index(m_thread_data_on_grid_idx[I0],
                                     n_thread_data_on_grid_idx[I0],
                                     m_thread_data_on_grid_idx[I1],
                                     n_thread_data_on_grid_idx[I1],
                                     m_thread_data_on_grid_idx[I2],
                                     m_thread_data_on_grid_idx[I3],
                                     m_thread_data_on_grid_idx[I4],
                                     n_thread_data_on_grid_idx[I2]),
                    c_element_op};
 
            c_thread_copy.Run(c_m0_n0_m1_n1_m2_m3_m4_n2_thread_desc,
                              make_tuple(I0, I0, I0, I0, I0, I0, I0, I0),
                              c_thread_buf,
                              c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc,
                              c_grid_buf);
        }
    }
};
 
} // namespace ck