/home/docs/checkouts/readthedocs.org/user_builds/advanced-micro-devices-composable-kernel/checkouts/develop/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_streamk.hpp Source File

/home/docs/checkouts/readthedocs.org/user_builds/advanced-micro-devices-composable-kernel/checkouts/develop/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_streamk.hpp Source File#

Composable Kernel: /home/docs/checkouts/readthedocs.org/user_builds/advanced-micro-devices-composable-kernel/checkouts/develop/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_streamk.hpp Source File
Go to the documentation of this file.
 // SPDX-License-Identifier: MIT
 // Copyright (c) 2018-2022, 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/block/thread_group_tensor_slice_transfer_v6r1r2.hpp"
 #include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
 #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
 #include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v3.hpp"
 #include "ck/utility/workgroup_barrier.hpp"
 #include "ck/utility/reduction_functions_accumulate.hpp"
  
 namespace ck {
  
 template <typename GridwiseGemm>
 __global__ void
 #if CK_USE_LAUNCH_BOUNDS
 __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
 #endif
     kernel_gemm_xdlops_streamk(const typename GridwiseGemm::FloatAB* p_a_grid,
                                const typename GridwiseGemm::FloatAB* p_b_grid,
                                typename GridwiseGemm::FloatC* p_c_grid,
                                void* p_workspace,
                                index_t M,
                                index_t N,
                                index_t K,
                                index_t StrideA,
                                index_t StrideB,
                                index_t StrideC,
                                typename GridwiseGemm::Block2CTileMap block_mapping)
 {
 #if(defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx94__))
     constexpr index_t shared_size = GridwiseGemm::GetSharedMemoryNumberOfByte();
  
     __shared__ uint8_t p_shared[shared_size];
  
     GridwiseGemm::Run(p_a_grid,
                       p_b_grid,
                       p_c_grid,
                       p_workspace,
                       M,
                       N,
                       K,
                       StrideA,
                       StrideB,
                       StrideC,
                       block_mapping,
                       static_cast<void*>(p_shared));
 #else
     ignore = p_a_grid;
     ignore = p_b_grid;
     ignore = p_c_grid;
     ignore = p_workspace;
     ignore = M;
     ignore = N;
     ignore = K;
     ignore = StrideA;
     ignore = StrideB;
     ignore = StrideC;
     ignore = block_mapping;
 #endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
 }
  
 template <index_t BlockSize,
           typename Block2CTileMap_,
           typename FloatAB_,
           typename FloatAcc_,
           typename FloatC_,
           typename ALayout,
           typename BLayout,
           typename CLayout,
           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,
           index_t ABlockLdsExtraM,
           typename BBlockTransferThreadClusterLengths_K0_N_K1,
           typename BBlockTransferThreadClusterArrangeOrder,
           typename BBlockTransferSrcAccessOrder,
           index_t BBlockTransferSrcVectorDim,
           index_t BBlockTransferSrcScalarPerVector,
           index_t BBlockTransferDstScalarPerVector_K1,
           bool BThreadTransferSrcResetCoordinateAfterRun,
           index_t BBlockLdsExtraN,
           index_t CShuffleMRepeatPerShuffle,
           index_t CShuffleNRepeatPerShuffle,
           index_t CBlockTransferScalarPerVector_NWaveNPerXDL,
           typename CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock>
 struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_streamk
 {
     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>{};
     static constexpr auto M01       = 1;
     static constexpr auto N01       = 1;
     static constexpr auto KPerBlock = K0PerBlock * K1;
  
     using ThisThreadBlock = ThisThreadBlock<BlockSize>;
     using FloatAcc        = FloatAcc_;
     using FloatCShuffle   = FloatAcc;
  
     using Block2CTileMap = Block2CTileMap_;
     using FloatAB        = FloatAB_;
     using FloatC         = FloatC_;
  
     struct Argument : public ck::tensor_operation::device::BaseArgument
     {
         const FloatAB* p_a_grid;
         const FloatAB* 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;
         Block2CTileMap block_mapping;
  
         Argument(const FloatAB* p_a_grid_,
                  const FloatAB* 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_,
                  uint32_t num_cu,
                  uint32_t occupancy,
                  uint32_t num_sk_blocks_)
             : p_a_grid(p_a_grid_),
               p_b_grid(p_b_grid_),
               p_c_grid(p_c_grid_),
               M(M_),
               N(N_),
               K(K_),
               StrideA(StrideA_),
               StrideB(StrideB_),
               StrideC(StrideC_),
               block_mapping(M, N, K, num_cu, occupancy, num_sk_blocks_)
         {
         }
  
         void Print() const
         {
             std::cout << "arg {" << "M:" << M << ", " << "N:" << N << ", " << "K:" << K << ", "
                       << "SA:" << StrideA << ", " << "SB:" << StrideB << ", " << "SC:" << StrideC
                       << std::endl;
         }
     };
  
     __host__ __device__ static auto CalculateGridSize(const Argument& karg)
     {
         return std::make_tuple(math::integer_divide_ceil(karg.N, NPerBlock),
                                math::integer_divide_ceil(karg.M, MPerBlock),
                                karg.k_batch);
     }
  
     __host__ __device__ static auto CalculateK0(index_t KPad) { return KPad / K1; }
  
     __host__ __device__ static auto
     MakeAGridDescriptor_K0_M_K1(index_t M, index_t MPad, index_t K, index_t KPad, index_t StrideA)
     {
         const index_t K0 = CalculateK0(KPad);
  
         const auto a_grid_desc_m_k = [&]() {
             if constexpr(is_same<tensor_layout::gemm::RowMajor, ALayout>::value)
             {
                 return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(StrideA, I1));
             }
             else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, ALayout>::value)
             {
                 return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(I1, StrideA));
             }
         }();
  
         const auto a_grid_desc_m_kpad = transform_tensor_descriptor(
             a_grid_desc_m_k,
             make_tuple(make_pass_through_transform(M), make_right_pad_transform(K, KPad - K)),
             make_tuple(Sequence<0>{}, Sequence<1>{}),
             make_tuple(Sequence<0>{}, Sequence<1>{}));
  
         return transform_tensor_descriptor(a_grid_desc_m_kpad,
                                            make_tuple(make_unmerge_transform(make_tuple(K0, K1)),
                                                       make_right_pad_transform(M, MPad - M)),
                                            make_tuple(Sequence<1>{}, Sequence<0>{}),
                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
     }
  
     __host__ __device__ static auto
     MakeBGridDescriptor_K0_N_K1(index_t K, index_t KPad, index_t N, index_t NPad, index_t StrideB)
     {
         const index_t K0 = CalculateK0(KPad);
  
         const auto b_grid_desc_k_n = [&]() {
             if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
             {
                 return make_naive_tensor_descriptor(make_tuple(K, N), make_tuple(StrideB, I1));
             }
             else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
             {
                 return make_naive_tensor_descriptor(make_tuple(K, N), make_tuple(I1, StrideB));
             }
         }();
  
         const auto b_grid_desc_kpad_n = transform_tensor_descriptor(
             b_grid_desc_k_n,
             make_tuple(make_right_pad_transform(K, KPad - K), make_pass_through_transform(N)),
             make_tuple(Sequence<0>{}, Sequence<1>{}),
             make_tuple(Sequence<0>{}, Sequence<1>{}));
  
         return transform_tensor_descriptor(b_grid_desc_kpad_n,
                                            make_tuple(make_unmerge_transform(make_tuple(K0, K1)),
                                                       make_right_pad_transform(N, NPad - N)),
                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
     }
  
     __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_m_n = [&]() {
             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));
             }
         }();
  
         return transform_tensor_descriptor(c_grid_desc_m_n,
                                            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>{}));
     }
  
     __host__ __device__ static constexpr auto GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1()
     {
         // A matrix in LDS memory, dst of blockwise copy
         return make_naive_tensor_descriptor(
             make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1),
             make_tuple(Number<MPerBlock + ABlockLdsExtraM>{} * K1, K1, I1));
     }
  
     __host__ __device__ static constexpr auto GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1()
     {
         // B matrix in LDS memory, dst of blockwise copy
         return make_naive_tensor_descriptor(
             make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1),
             make_tuple(Number<NPerBlock + BBlockLdsExtraN>{} * K1, K1, I1));
     }
  
     __host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
     {
         constexpr auto max_lds_align = K1;
  
         // LDS allocation for A and B: be careful of alignment
         constexpr auto a_block_desc_k0_m_k1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
         constexpr auto b_block_desc_k0_n_k1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
  
         constexpr auto a_block_space_size_aligned =
             math::integer_least_multiple(a_block_desc_k0_m_k1.GetElementSpaceSize(), max_lds_align);
  
         constexpr auto b_block_space_size_aligned =
             math::integer_least_multiple(b_block_desc_k0_n_k1.GetElementSpaceSize(), max_lds_align);
  
         constexpr auto c_block_size =
             GetCBlockDescriptor_MBlock_MPerShuffle_NBlock_NPerShuffle().GetElementSpaceSize();
  
         return math::max((a_block_space_size_aligned + b_block_space_size_aligned) *
                              sizeof(FloatAB),
                          c_block_size * sizeof(FloatCShuffle));
     }
  
     __host__ __device__ static constexpr bool CheckValidity(const Argument& karg)
     {
         if constexpr(is_same<tensor_layout::gemm::RowMajor, ALayout>::value)
         {
             if(karg.K % ABlockTransferSrcScalarPerVector != 0)
                 return false;
         }
         else
         {
             if(karg.M % ABlockTransferSrcScalarPerVector != 0)
                 return false;
         }
  
         if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
         {
             if(karg.N % BBlockTransferSrcScalarPerVector != 0)
                 return false;
         }
         else
         {
             if(karg.K % BBlockTransferSrcScalarPerVector != 0)
                 return false;
         }
  
         if constexpr(is_same<tensor_layout::gemm::RowMajor, CLayout>::value)
         {
             if(karg.N % CBlockTransferScalarPerVector_NWaveNPerXDL != 0)
                 return false;
         }
         else
         {
             if(karg.M % CBlockTransferScalarPerVector_NWaveNPerXDL != 0)
                 return false;
         }
  
         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;
     }
  
     template <typename CGridDesc>
     __host__ __device__ static constexpr auto
     MakeCGridDesc_MBlock_MPerBlock_NBlock_NPerBlock(const CGridDesc& c_m_n_grid_desc)
     {
         const auto M = c_m_n_grid_desc.GetLength(I0);
         const auto N = c_m_n_grid_desc.GetLength(I1);
  
         const auto MBlock = M / MPerBlock;
         const auto NBlock = N / NPerBlock;
  
         return transform_tensor_descriptor(
             c_m_n_grid_desc,
             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 block_id to C matrix tile idx (m0, n0) mapping
     template <typename CGridDesc>
     __host__ __device__ static constexpr auto MakeCBlockClusterAdaptor(
         const CGridDesc& c_m_n_grid_desc, index_t /* M01 */, index_t /* N01 */, index_t KBatch)
     {
         return BlockToCTileMap_KSplit_M00_N0_M01Adapt<MPerBlock, NPerBlock, CGridDesc>(
             c_m_n_grid_desc, 8, KBatch);
     }
  
     __host__ __device__ static constexpr auto
     GetCBlockDescriptor_MBlock_MPerShuffle_NBlock_NPerShuffle()
     {
         constexpr index_t MWave = MPerBlock / (MRepeat * MPerXDL);
         constexpr index_t NWave = NPerBlock / (NRepeat * NPerXDL);
  
         return make_naive_tensor_descriptor_packed(
             make_tuple(I1,
                        Number<CShuffleMRepeatPerShuffle * MWave * MPerXDL>{},
                        I1,
                        Number<CShuffleNRepeatPerShuffle * NWave * NPerXDL>{}));
     }
  
     __host__ __device__ static constexpr auto
     GetCBlockDescriptor_MShuffleRepeat_MPerShuffle_NShuffleRepeat_NPerShuffle()
     {
         constexpr index_t MWave = MPerBlock / (MRepeat * MPerXDL);
         constexpr index_t NWave = NPerBlock / (NRepeat * NPerXDL);
  
         return make_naive_tensor_descriptor_packed(
             make_tuple(Number<MRepeat / CShuffleMRepeatPerShuffle>{},
                        Number<CShuffleMRepeatPerShuffle * MWave * MPerXDL>{},
                        Number<NRepeat / CShuffleNRepeatPerShuffle>{},
                        Number<CShuffleNRepeatPerShuffle * NWave * NPerXDL>{}));
     }
  
     __host__ __device__ static constexpr auto GetClusterLengthReduction()
     {
         // TODO: assume C is row major
         // TODO: we always first loop over N, then M
         constexpr auto NPerBlockPow2 = math::next_power_of_two<NPerBlock>();
         constexpr auto NPerBlockReduction =
             NPerBlockPow2 / CBlockTransferScalarPerVector_NWaveNPerXDL;
         constexpr auto MPerBlockReduction =
             (BlockSize + NPerBlockReduction - 1) / NPerBlockReduction;
         return Sequence<MPerBlockReduction, NPerBlockReduction>{};
     }
  
     __host__ __device__ static constexpr auto GetPartialAccBlockDescriptor()
     {
         const auto c_partial_acc_block_m_n = [&]() {
             if constexpr(is_same<tensor_layout::gemm::RowMajor, CLayout>::value)
             {
                 return make_naive_tensor_descriptor(make_tuple(MPerBlock, NPerBlock),
                                                     make_tuple(NPerBlock, I1));
             }
             else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, CLayout>::value)
             {
                 return make_naive_tensor_descriptor(make_tuple(MPerBlock, NPerBlock),
                                                     make_tuple(I1, MPerBlock));
             }
         }();
         return c_partial_acc_block_m_n;
     }
  
     using CGridDesc_M_N = remove_cvref_t<decltype(MakeCGridDescriptor_M_N(1, 1, 1, 1, 1))>;
  
     __device__ static void Run(const FloatAB* p_a_grid,
                                const FloatAB* p_b_grid,
                                FloatC* p_c_grid,
                                void* p_workspace,
                                index_t M,
                                index_t N,
                                index_t K,
                                index_t StrideA,
                                index_t StrideB,
                                index_t StrideC,
                                Block2CTileMap block_mapping,
                                void* __restrict__ p_shared_block)
     {
         uint32_t m        = M;
         uint32_t n        = N;
         uint32_t k        = K;
         uint32_t pad_m    = (m + MPerBlock - 1) / MPerBlock * MPerBlock;
         uint32_t pad_n    = (n + NPerBlock - 1) / NPerBlock * NPerBlock;
         uint32_t pad_k    = (k + KPerBlock - 1) / KPerBlock * KPerBlock;
         uint32_t stride_a = StrideA;
         uint32_t stride_b = StrideB;
         uint32_t stride_c = StrideC;
  
         const auto a_k0_m_k1_grid_desc = MakeAGridDescriptor_K0_M_K1(m, pad_m, k, pad_k, stride_a);
         const auto b_k0_n_k1_grid_desc = MakeBGridDescriptor_K0_N_K1(k, pad_k, n, pad_n, stride_b);
         const auto c_grid_desc_m_n     = MakeCGridDescriptor_M_N(m, pad_m, n, pad_n, stride_c);
  
         const auto c_grid_desc_mblock_mperblock_nblock_nperblock =
             MakeCGridDesc_MBlock_MPerBlock_NBlock_NPerBlock(c_grid_desc_m_n);
         const AElementwiseOperation a_element_op = AElementwiseOperation{};
         const BElementwiseOperation b_element_op = BElementwiseOperation{};
         const CElementwiseOperation c_element_op = CElementwiseOperation{};
  
         const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
             p_a_grid, a_k0_m_k1_grid_desc.GetElementSpaceSize());
         const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
             p_b_grid, b_k0_n_k1_grid_desc.GetElementSpaceSize());
         auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
             p_c_grid, c_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
  
         // lds max alignment
         constexpr auto max_lds_align = K1;
  
         // A matrix in LDS memory, dst of blockwise copy
         constexpr auto a_block_desc_k0_m_k1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
  
         // B matrix in LDS memory, dst of blockwise copy
         constexpr auto b_block_desc_k0_n_k1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
  
         auto blockwise_gemm =
             BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<BlockSize,
                                                                 FloatAB,
                                                                 FloatAB,
                                                                 FloatAcc,
                                                                 decltype(a_block_desc_k0_m_k1),
                                                                 decltype(b_block_desc_k0_n_k1),
                                                                 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_block_desc_k0_m_k1.GetElementSpaceSize(), max_lds_align);
  
         FloatAB* p_a_block = static_cast<FloatAB*>(p_shared_block);
         FloatAB* p_b_block = static_cast<FloatAB*>(p_shared_block) + a_block_space_size;
  
         constexpr auto a_block_slice_copy_step = make_multi_index(K0PerBlock, 0, 0);
         constexpr auto b_block_slice_copy_step = make_multi_index(K0PerBlock, 0, 0);
  
         auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
             p_a_block, a_block_desc_k0_m_k1.GetElementSpaceSize());
         auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
             p_b_block, b_block_desc_k0_n_k1.GetElementSpaceSize());
  
         // gridwise GEMM pipeline
         const auto gridwise_gemm_pipeline = GridwiseGemmPipeline_v3();
  
         uint32_t block_idx = block_mapping.get_block_idx();
         bool is_sk_block   = block_idx < block_mapping.sk_num_blocks;
         bool is_dp_block   = block_idx >= block_mapping.dp_start_block_idx &&
                            block_idx < block_mapping.reduction_start_block_idx;
         bool is_reduction_block = block_idx >= block_mapping.reduction_start_block_idx;
         bool is_padding_block   = block_idx >= block_mapping.sk_num_blocks &&
                                 block_idx < block_mapping.dp_start_block_idx;
         uint32_t iter_start, iter_end;
         block_mapping.get_block_itr(block_idx, iter_start, iter_end);
         uint32_t total_iter_length = iter_end - iter_start;
  
         if(is_padding_block)
             return;
  
         uint32_t* p_semaphore =
             reinterpret_cast<uint32_t*>(reinterpret_cast<char*>(p_workspace) +
                                         block_mapping.get_workspace_size_for_acc(sizeof(FloatAcc)));
  
         if constexpr(Block2CTileMap::ReductionStrategy == StreamKReductionStrategy::Reduction)
         {
             if(is_reduction_block)
             {
                 // descriptors
                 constexpr auto cluster_length_reduce = GetClusterLengthReduction();
                 constexpr auto reduce_desc = make_cluster_descriptor(cluster_length_reduce);
                 const auto reduce_thread_cluster_idx =
                     reduce_desc.CalculateBottomIndex(make_multi_index(get_thread_local_1d_id()));
                 const auto thread_m_cluster_id = reduce_thread_cluster_idx[I0];
                 const auto thread_n_cluster_id = reduce_thread_cluster_idx[I1];
  
                 constexpr auto MReduceIters =
                     math::integer_divide_ceil(Number<MPerBlock>{}, cluster_length_reduce.At(I0));
                 constexpr auto NReduceIters = math::integer_divide_ceil(
                     Number<NPerBlock>{},
                     cluster_length_reduce.At(I1) *
                         Number<CBlockTransferScalarPerVector_NWaveNPerXDL>{});
  
                 constexpr auto acc_thread_buf_load_desc = make_naive_tensor_descriptor_packed(
                     make_tuple(I1, Number<CBlockTransferScalarPerVector_NWaveNPerXDL>{}));
                 constexpr auto acc_thread_buf_store_desc = make_naive_tensor_descriptor_packed(
                     make_tuple(I1, I1, I1, Number<CBlockTransferScalarPerVector_NWaveNPerXDL>{}));
  
                 constexpr auto c_partial_acc_block_m_n = GetPartialAccBlockDescriptor();
  
                 constexpr auto partial_acc_load_step_n = make_multi_index(
                     0, cluster_length_reduce.At(I1) * CBlockTransferScalarPerVector_NWaveNPerXDL);
                 constexpr auto partial_acc_load_step_n_reverse =
                     make_multi_index(0,
                                      -1 * cluster_length_reduce.At(I1).value * (NReduceIters - 1) *
                                          CBlockTransferScalarPerVector_NWaveNPerXDL);
                 constexpr auto partial_acc_load_step_m =
                     make_multi_index(cluster_length_reduce.At(I0), 0);
  
                 constexpr auto partial_acc_store_step_n = make_multi_index(
                     0,
                     0,
                     0,
                     cluster_length_reduce.At(I1) * CBlockTransferScalarPerVector_NWaveNPerXDL);
                 constexpr auto partial_acc_store_step_n_reverse =
                     make_multi_index(0,
                                      0,
                                      0,
                                      -1 * cluster_length_reduce.At(I1).value * (NReduceIters - 1) *
                                          CBlockTransferScalarPerVector_NWaveNPerXDL);
                 constexpr auto partial_acc_store_step_m =
                     make_multi_index(0, cluster_length_reduce.At(I0), 0, 0);
  
                 StaticBuffer<AddressSpaceEnum::Vgpr,
                              FloatAcc,
                              CBlockTransferScalarPerVector_NWaveNPerXDL,
                              true>
                     parcial_acc_buf;
                 StaticBuffer<AddressSpaceEnum::Vgpr,
                              FloatAcc,
                              CBlockTransferScalarPerVector_NWaveNPerXDL,
                              true>
                     acc_buf;
  
                 // start to compute
                 auto reduction_idx = blockIdx.x - block_mapping.reduction_start_block_idx;
                 auto spatial_idx   = block_mapping.tile_to_spatial(reduction_idx, m, n);
  
                 workgroup_barrier wg_barrier(p_semaphore);
  
                 uint32_t tile_acc_offset_start =
                     block_mapping.get_acc_buffer_offset_from_tile(reduction_idx);
                 uint32_t tile_acc_offset_end =
                     block_mapping.get_acc_buffer_offset_from_tile(reduction_idx + 1);
  
                 auto acc_load = ThreadwiseTensorSliceTransfer_v2<
                     FloatAcc,                                                // SrcData,
                     FloatAcc,                                                // DstData,
                     decltype(c_partial_acc_block_m_n),                       // SrcDesc,
                     decltype(acc_thread_buf_load_desc),                      // DstDesc,
                     Sequence<1, CBlockTransferScalarPerVector_NWaveNPerXDL>, // SliceLengths,
                     Sequence<0, 1>,                                          // DimAccessOrder,
                     1,                                                       // SrcVectorDim,
                     CBlockTransferScalarPerVector_NWaveNPerXDL,              // SrcScalarPerVector,
                     1,    // SrcScalarStrideInVector,
                     false // SrcResetCoordinateAfterRun,
                     >{c_partial_acc_block_m_n,
                       make_multi_index(thread_m_cluster_id,
                                        thread_n_cluster_id *
                                            CBlockTransferScalarPerVector_NWaveNPerXDL)};
  
                 auto acc_store = ThreadwiseTensorSliceTransfer_v1r3<
                     FloatAcc,                                                // SrcData,
                     FloatC,                                                  // DstData,
                     decltype(acc_thread_buf_store_desc),                     // SrcDesc,
                     decltype(c_grid_desc_mblock_mperblock_nblock_nperblock), // DstDesc,
                     CElementwiseOperation, // ElementwiseOperation,
                     Sequence<1, 1, 1, CBlockTransferScalarPerVector_NWaveNPerXDL>, // SliceLengths,
                     Sequence<0, 1, 2, 3>,                       // DimAccessOrder,
                     3,                                          // DstVectorDim,
                     CBlockTransferScalarPerVector_NWaveNPerXDL, // DstScalarPerVector,
                     InMemoryDataOperationEnum::Set, // InMemoryDataOperationEnum DstInMemOp,
                     1,                              // DstScalarStrideInVector,
                     false                           // DstResetCoordinateAfterRun,
                     >{c_grid_desc_mblock_mperblock_nblock_nperblock,
                       make_multi_index(__builtin_amdgcn_readfirstlane(spatial_idx[I0]),
                                        thread_m_cluster_id,
                                        __builtin_amdgcn_readfirstlane(spatial_idx[I1]),
                                        thread_n_cluster_id *
                                            CBlockTransferScalarPerVector_NWaveNPerXDL),
                       CElementwiseOperation{}};
  
                 // block synchronization
                 wg_barrier.wait_eq(reduction_idx, tile_acc_offset_end - tile_acc_offset_start);
  
 #if 0
                 if(threadIdx.x == 0) {
                     printf("bid:%d, rid:%d, os:%d,%d, spatial:%d,%d\n", static_cast<int>(blockIdx.x),
                         reduction_idx, __builtin_amdgcn_readfirstlane(tile_acc_offset_start), __builtin_amdgcn_readfirstlane(tile_acc_offset_end),
                         __builtin_amdgcn_readfirstlane(spatial_idx[I0]),
                         __builtin_amdgcn_readfirstlane(spatial_idx[I1]));
                 }
 #endif
  
                 using Accumulation = ck::detail::
                     AccumulateWithNanCheck<false /*PropagateNan*/, reduce::Add, FloatAcc>;
  
                 for(int i_m = 0; i_m < MReduceIters; i_m++)
                 {
                     static_for<0, NReduceIters, 1>{}([&](auto i_n_reduce) {
                         acc_buf.Clear();
                         for(auto i = tile_acc_offset_start; i < tile_acc_offset_end; i++)
                         {
                             auto c_partial_acc_buf =
                                 make_dynamic_buffer<AddressSpaceEnum::Global,
                                                     AmdBufferCoherenceEnum::GLC>(
                                     reinterpret_cast<FloatAcc*>(p_workspace) +
                                         i * c_partial_acc_block_m_n.GetElementSpaceSize(),
                                     c_partial_acc_block_m_n.GetElementSpaceSize());
  
                             acc_load.Run(c_partial_acc_block_m_n,
                                          c_partial_acc_buf,
                                          acc_thread_buf_load_desc,
                                          make_tuple(I0, I0),
                                          parcial_acc_buf);
  
                             static_for<0, CBlockTransferScalarPerVector_NWaveNPerXDL, 1>{}(
                                 [&](auto i_vec) {
                                     constexpr auto offset =
                                         acc_thread_buf_load_desc.CalculateOffset(
                                             make_tuple(0, i_vec));
                                     Accumulation::Calculate(acc_buf(Number<offset>{}),
                                                             parcial_acc_buf[Number<offset>{}]);
                                 });
                         }
  
                         if(thread_n_cluster_id * CBlockTransferScalarPerVector_NWaveNPerXDL <
                            NPerBlock)
                         {
                             acc_store.Run(acc_thread_buf_store_desc,
                                           make_tuple(I0, I0, I0, I0),
                                           acc_buf,
                                           c_grid_desc_mblock_mperblock_nblock_nperblock,
                                           c_grid_buf);
                         }
                         if constexpr(NReduceIters != 1)
                         {
                             if constexpr(i_n_reduce != (NReduceIters - 1))
                             {
                                 acc_load.MoveSrcSliceWindow(c_partial_acc_block_m_n,
                                                             partial_acc_load_step_n);
                                 acc_store.MoveDstSliceWindow(
                                     c_grid_desc_mblock_mperblock_nblock_nperblock,
                                     partial_acc_store_step_n);
                             }
                             else
                             {
                                 acc_load.MoveSrcSliceWindow(c_partial_acc_block_m_n,
                                                             partial_acc_load_step_n_reverse);
                                 acc_store.MoveDstSliceWindow(
                                     c_grid_desc_mblock_mperblock_nblock_nperblock,
                                     partial_acc_store_step_n_reverse);
                             }
                         }
                     });
                     {
                         acc_load.MoveSrcSliceWindow(c_partial_acc_block_m_n,
                                                     partial_acc_load_step_m);
                         acc_store.MoveDstSliceWindow(c_grid_desc_mblock_mperblock_nblock_nperblock,
                                                      partial_acc_store_step_m);
                     }
                 }
                 return;
             }
         }
  
         // offset for last acc buffer of this block
         uint32_t block_acc_offset =
             (block_mapping.get_acc_buffer_offset_from_block(block_idx + 1) - 1) * MPerBlock *
             NPerBlock;
  
         while(true)
         {
             uint32_t current_iter_length = __builtin_amdgcn_readfirstlane(
                 block_mapping.get_current_iter_length(iter_start, iter_end, total_iter_length));
             uint32_t tile_idx, iter_offset;
             block_mapping.get_tile_idx_with_offset(iter_end - 1, tile_idx, iter_offset);
             iter_offset = __builtin_amdgcn_readfirstlane(iter_offset - current_iter_length + 1);
             auto spatial_idx = block_mapping.tile_to_spatial(tile_idx, m, n);
  
             const index_t m_block_data_idx_on_grid =
                 __builtin_amdgcn_readfirstlane(spatial_idx[I0] * MPerBlock);
  
             const index_t n_block_data_idx_on_grid =
                 __builtin_amdgcn_readfirstlane(spatial_idx[I1] * NPerBlock);
  
             const index_t k0_block_data_idx_on_grid =
                 __builtin_amdgcn_readfirstlane(iter_offset * K0PerBlock);
  
             // A matrix blockwise copy
             auto a_blockwise_copy =
                 ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
                                                     AElementwiseOperation,
                                                     ck::tensor_operation::element_wise::PassThrough,
                                                     InMemoryDataOperationEnum::Set,
                                                     Sequence<K0PerBlock, MPerBlock, K1>,
                                                     ABlockTransferThreadClusterLengths_K0_M_K1,
                                                     ABlockTransferThreadClusterArrangeOrder,
                                                     FloatAB,
                                                     FloatAB,
                                                     decltype(a_k0_m_k1_grid_desc),
                                                     decltype(a_block_desc_k0_m_k1),
                                                     ABlockTransferSrcAccessOrder,
                                                     Sequence<1, 0, 2>,
                                                     ABlockTransferSrcVectorDim,
                                                     2,
                                                     ABlockTransferSrcScalarPerVector,
                                                     ABlockTransferDstScalarPerVector_K1,
                                                     1,
                                                     1,
                                                     AThreadTransferSrcResetCoordinateAfterRun,
                                                     true>(
                     a_k0_m_k1_grid_desc,
                     make_multi_index(k0_block_data_idx_on_grid, m_block_data_idx_on_grid, 0),
                     a_element_op,
                     a_block_desc_k0_m_k1,
                     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<K0PerBlock, NPerBlock, K1>,
                                                     BBlockTransferThreadClusterLengths_K0_N_K1,
                                                     BBlockTransferThreadClusterArrangeOrder,
                                                     FloatAB,
                                                     FloatAB,
                                                     decltype(b_k0_n_k1_grid_desc),
                                                     decltype(b_block_desc_k0_n_k1),
                                                     BBlockTransferSrcAccessOrder,
                                                     Sequence<1, 0, 2>,
                                                     BBlockTransferSrcVectorDim,
                                                     2,
                                                     BBlockTransferSrcScalarPerVector,
                                                     BBlockTransferDstScalarPerVector_K1,
                                                     1,
                                                     1,
                                                     BThreadTransferSrcResetCoordinateAfterRun,
                                                     true>(
                     b_k0_n_k1_grid_desc,
                     make_multi_index(k0_block_data_idx_on_grid, n_block_data_idx_on_grid, 0),
                     b_element_op,
                     b_block_desc_k0_n_k1,
                     make_multi_index(0, 0, 0),
                     ck::tensor_operation::element_wise::PassThrough{});
  
             const index_t num_k_block_main_loop = current_iter_length;
  
             gridwise_gemm_pipeline.Run(a_k0_m_k1_grid_desc,
                                        a_block_desc_k0_m_k1,
                                        a_blockwise_copy,
                                        a_grid_buf,
                                        a_block_buf,
                                        a_block_slice_copy_step,
                                        b_k0_n_k1_grid_desc,
                                        b_block_desc_k0_n_k1,
                                        b_blockwise_copy,
                                        b_grid_buf,
                                        b_block_buf,
                                        b_block_slice_copy_step,
                                        blockwise_gemm,
                                        c_thread_buf,
                                        num_k_block_main_loop);
  
             // output: register to global memory
             {
                 constexpr index_t MWave = MPerBlock / (MRepeat * MPerXDL);
                 constexpr index_t NWave = NPerBlock / (NRepeat * NPerXDL);
  
                 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 c_m0_n0_m1_n1_m2_m3_m4_n2_thread_desc =
                     blockwise_gemm.GetCThreadDescriptor_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_block_desc_mblock_mpershuffle_nblock_npershuffle =
                     GetCBlockDescriptor_MBlock_MPerShuffle_NBlock_NPerShuffle();
  
                 constexpr auto c_block_desc_mshuffle_mpershuffle_nshuffle_npershuffle =
                     GetCBlockDescriptor_MShuffleRepeat_MPerShuffle_NShuffleRepeat_NPerShuffle();
  
                 auto c_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
                     reinterpret_cast<FloatCShuffle*>(p_shared_block),
                     c_block_desc_mblock_mpershuffle_nblock_npershuffle.GetElementSpaceSize());
  
                 auto c_partial_acc_buf =
                     make_dynamic_buffer<AddressSpaceEnum::Global, AmdBufferCoherenceEnum::GLC>(
                         reinterpret_cast<FloatAcc*>(p_workspace) + block_acc_offset,
                         c_block_desc_mshuffle_mpershuffle_nshuffle_npershuffle
                             .GetElementSpaceSize());
  
                 constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2 = transform_tensor_descriptor(
                     c_block_desc_mblock_mpershuffle_nblock_npershuffle,
                     make_tuple(make_freeze_transform(I0), // freeze mblock
                                make_unmerge_transform(
                                    make_tuple(CShuffleMRepeatPerShuffle,
                                               M1,
                                               M2,
                                               M3,
                                               M4)),       // M1 = MWave, M2 * M3 * M4 = MPerXDL
                                make_freeze_transform(I0), // freeze nblock
                                make_unmerge_transform(
                                    make_tuple(CShuffleNRepeatPerShuffle,
                                               N1,
                                               N2))), // M1 = MWave, M2 * M3 * M4 = MPerXDL
                     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.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));
  
                 // VGPR to LDS
                 auto c_thread_copy_vgpr_to_lds = ThreadwiseTensorSliceTransfer_v1r3<
                     FloatAcc,
                     FloatCShuffle,
                     decltype(c_m0_n0_m1_n1_m2_m3_m4_n2_thread_desc),
                     decltype(c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2),
                     ck::tensor_operation::element_wise::PassThrough,
                     Sequence<CShuffleMRepeatPerShuffle,
                              CShuffleNRepeatPerShuffle,
                              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{}};
  
                 // LDS to global
                 auto c_block_copy_lds_to_global = ThreadGroupTensorSliceTransfer_v6r1r2<
                     ThisThreadBlock,       // index_t BlockSize,
                     CElementwiseOperation, // ElementwiseOperation,
                                            // InMemoryDataOperationEnum::Set, // DstInMemOp,
                     Sequence<1,
                              CShuffleMRepeatPerShuffle * MWave * MPerXDL,
                              1,
                              CShuffleNRepeatPerShuffle * NWave * NPerXDL>, // BlockSliceLengths,
                     CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
                     Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
                     FloatCShuffle,        // typename SrcData,
                     FloatC,               // typename DstData,
                     decltype(c_block_desc_mblock_mpershuffle_nblock_npershuffle),
                     decltype(c_grid_desc_mblock_mperblock_nblock_nperblock),
                     Sequence<0, 1, 2, 3>,                       // typename DimAccessOrder,
                     3,                                          // index_t VectorDim,
                     CBlockTransferScalarPerVector_NWaveNPerXDL, // index_t ScalarPerVector,
                     false, // bool ThreadTransferSrcResetCoordinateAfterRun,
                     false> // bool ThreadTransferDstResetCoordinateAfterRun
                     {c_block_desc_mblock_mpershuffle_nblock_npershuffle,
                      make_multi_index(0, 0, 0, 0),
                      c_grid_desc_mblock_mperblock_nblock_nperblock,
                      make_multi_index(__builtin_amdgcn_readfirstlane(spatial_idx[I0]),
                                       0,
                                       __builtin_amdgcn_readfirstlane(spatial_idx[I1]),
                                       0),
                      c_element_op};
  
                 // LDS to global partial acc
                 auto c_block_copy_lds_to_partial_acc = ThreadGroupTensorSliceTransfer_v6r1r2<
                     ThisThreadBlock,       // index_t BlockSize,
                     CElementwiseOperation, // ElementwiseOperation,
                                            // InMemoryDataOperationEnum::Set, // DstInMemOp,
                     Sequence<1,
                              CShuffleMRepeatPerShuffle * MWave * MPerXDL,
                              1,
                              CShuffleNRepeatPerShuffle * NWave * NPerXDL>, // BlockSliceLengths,
                     CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
                     Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
                     FloatCShuffle,        // typename SrcData,
                     FloatCShuffle,        // typename DstData,
                     decltype(c_block_desc_mblock_mpershuffle_nblock_npershuffle),
                     decltype(c_block_desc_mshuffle_mpershuffle_nshuffle_npershuffle),
                     Sequence<0, 1, 2, 3>,                       // typename DimAccessOrder,
                     3,                                          // index_t VectorDim,
                     CBlockTransferScalarPerVector_NWaveNPerXDL, // index_t ScalarPerVector,
                     false, // bool ThreadTransferSrcResetCoordinateAfterRun, => need to be false,
                            // othre wise has scratch
                     false> // bool ThreadTransferDstResetCoordinateAfterRun, => need to be false,
                            // othre wise has scratch
                     {c_block_desc_mblock_mpershuffle_nblock_npershuffle,
                      make_multi_index(0, 0, 0, 0),
                      c_block_desc_mshuffle_mpershuffle_nshuffle_npershuffle,
                      make_multi_index(0, 0, 0, 0),
                      c_element_op};
  
                 constexpr auto mxdlperwave_forward_step =
                     make_multi_index(0, CShuffleMRepeatPerShuffle * MWave * MPerXDL, 0, 0);
                 constexpr auto nxdlperwave_forward_step =
                     make_multi_index(0, 0, 0, CShuffleNRepeatPerShuffle * NWave * NPerXDL);
                 constexpr auto nxdlperwave_backward_step =
                     make_multi_index(0, 0, 0, -CShuffleNRepeatPerShuffle * NWave * NPerXDL);
  
                 static_for<0, MRepeat, CShuffleMRepeatPerShuffle>{}([&](auto mxdlperwave_iter) {
                     constexpr auto mxdlperwave = mxdlperwave_iter;
  
                     static_for<0, NRepeat, CShuffleNRepeatPerShuffle>{}([&](auto nxdlperwave_iter) {
                         constexpr bool nxdlperwave_forward_sweep =
                             (mxdlperwave % (2 * CShuffleMRepeatPerShuffle) == 0);
  
                         constexpr index_t nxdlperwave_value =
                             nxdlperwave_forward_sweep
                                 ? nxdlperwave_iter
                                 : (NRepeat - nxdlperwave_iter - CShuffleNRepeatPerShuffle);
  
                         constexpr auto nxdlperwave = Number<nxdlperwave_value>{};
  
                         // make sure it's safe to do ds_write
                         block_sync_lds();
  
                         // VGPR to LDS
                         c_thread_copy_vgpr_to_lds.Run(
                             c_m0_n0_m1_n1_m2_m3_m4_n2_thread_desc,
                             make_tuple(mxdlperwave, nxdlperwave, I0, I0, I0, I0, I0, I0),
                             c_thread_buf,
                             c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
                             c_block_buf);
  
                         // make sure it's safe to do ds_read
                         block_sync_lds();
  
                         c_block_copy_lds_to_global.SetSrcSliceOrigin(
                             c_block_desc_mblock_mpershuffle_nblock_npershuffle,
                             make_tuple(0, 0, 0, 0));
  
                         // LDS to global
                         if(is_dp_block)
                             c_block_copy_lds_to_global.template Run<decltype(c_block_buf),
                                                                     decltype(c_grid_buf),
                                                                     InMemoryDataOperationEnum::Set>(
                                 c_block_desc_mblock_mpershuffle_nblock_npershuffle,
                                 c_block_buf,
                                 c_grid_desc_mblock_mperblock_nblock_nperblock,
                                 c_grid_buf);
                         else if(is_sk_block)
                         {
                             if constexpr(Block2CTileMap::ReductionStrategy ==
                                          StreamKReductionStrategy::Reduction)
                             {
                                 // constexpr offset
                                 c_block_copy_lds_to_partial_acc.SetSrcSliceOrigin(
                                     c_block_desc_mblock_mpershuffle_nblock_npershuffle,
                                     make_tuple(0, 0, 0, 0));
  
                                 c_block_copy_lds_to_partial_acc.SetDstSliceOrigin(
                                     c_block_desc_mshuffle_mpershuffle_nshuffle_npershuffle,
                                     make_tuple(mxdlperwave.value, 0, nxdlperwave.value, 0));
  
                                 c_block_copy_lds_to_partial_acc
                                     .template Run<decltype(c_block_buf),
                                                   decltype(c_partial_acc_buf),
                                                   InMemoryDataOperationEnum::Set>(
                                         c_block_desc_mblock_mpershuffle_nblock_npershuffle,
                                         c_block_buf,
                                         c_block_desc_mshuffle_mpershuffle_nshuffle_npershuffle,
                                         c_partial_acc_buf);
                             }
                             else if constexpr(Block2CTileMap::ReductionStrategy ==
                                               StreamKReductionStrategy::Atomic)
                             {
                                 c_block_copy_lds_to_global
                                     .template Run<decltype(c_block_buf),
                                                   decltype(c_grid_buf),
                                                   InMemoryDataOperationEnum::AtomicAdd>(
                                         c_block_desc_mblock_mpershuffle_nblock_npershuffle,
                                         c_block_buf,
                                         c_grid_desc_mblock_mperblock_nblock_nperblock,
                                         c_grid_buf);
                             }
                         }
  
                         // move on nxdlperwave dimension
                         if constexpr(nxdlperwave_forward_sweep &&
                                      (nxdlperwave < NRepeat - CShuffleNRepeatPerShuffle))
                         {
                             c_block_copy_lds_to_global.MoveDstSliceWindow(
                                 c_grid_desc_mblock_mperblock_nblock_nperblock,
                                 nxdlperwave_forward_step);
                         }
                         else if constexpr((!nxdlperwave_forward_sweep) && (nxdlperwave > 0))
                         {
                             c_block_copy_lds_to_global.MoveDstSliceWindow(
                                 c_grid_desc_mblock_mperblock_nblock_nperblock,
                                 nxdlperwave_backward_step);
                         }
                     });
  
                     // move on mxdlperwave dimension
                     if constexpr(mxdlperwave < MRepeat - CShuffleMRepeatPerShuffle)
                     {
                         c_block_copy_lds_to_global.MoveDstSliceWindow(
                             c_grid_desc_mblock_mperblock_nblock_nperblock,
                             mxdlperwave_forward_step);
                     }
                 });
  
                 if constexpr(Block2CTileMap::ReductionStrategy ==
                              StreamKReductionStrategy::Reduction)
                 {
                     if(is_sk_block)
                     {
                         // increase the counter for this tile
                         workgroup_barrier wg_barrier(p_semaphore);
                         wg_barrier.inc(tile_idx);
                     }
                 }
             }
  
             // exit condition
             iter_end -= current_iter_length;
             if(iter_end <= iter_start)
                 break;
  
             if constexpr(Block2CTileMap::ReductionStrategy == StreamKReductionStrategy::Reduction)
             {
                 block_acc_offset -= MPerBlock * NPerBlock;
             }
             // make sure next loop LDS is ready for use
             block_sync_lds();
         }
     }
  
     template <typename Layout>
     struct LStr
     {
         static std::string Get() { return ""; }
     };
  
     template <>
     struct LStr<ck::tensor_layout::gemm::RowMajor>
     {
         static std::string Get() { return "R"; }
     };
  
     template <>
     struct LStr<ck::tensor_layout::gemm::ColumnMajor>
     {
         static std::string Get() { return "C"; }
     };
  
     static std::string GetTypeString()
     {
         auto str = std::stringstream();
  
         // clang-format off
         str << "GemmXdlStreamK_"
             << std::string(ALayout::name)[0]
             << std::string(BLayout::name)[0]
             << std::string(CLayout::name)[0]
             << "_"
             << "B" << BlockSize << "_"
             << "Vec" << ABlockTransferSrcScalarPerVector << "x"
             << BBlockTransferSrcScalarPerVector << "x"
             << CBlockTransferScalarPerVector_NWaveNPerXDL << "_"
             << MPerBlock << "x"
             << NPerBlock << "x"
             << K0PerBlock << "x"
             << K1 ;
         // clang-format on
  
         return str.str();
     }
 };
  
 } // namespace ck