/home/docs/checkouts/readthedocs.org/user_builds/advanced-micro-devices-composable-kernel/checkouts/develop/include/ck/tensor_operation/gpu/grid/gridwise_gemm_wmma.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_wmma.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_wmma.hpp Source File
Go to the documentation of this file.
 // SPDX-License-Identifier: MIT
 // Copyright (c) 2018-2023, 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_wmma.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,
           typename ADataType,
           typename BDataType,
           typename CDataType,
           typename AGridDesc,
           typename BGridDesc,
           typename CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
           typename AElementwiseOperation,
           typename BElementwiseOperation,
           typename CElementwiseOperation,
           typename Block2CTileMap,
           bool HasMainKBlockLoop>
 __global__ void
 #if CK_USE_LAUNCH_BOUNDS
 __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
 #endif
     kernel_gemm_wmma(const ADataType* __restrict__ p_a_grid,
                      const BDataType* __restrict__ p_b_grid,
                      CDataType* __restrict__ p_c_grid,
                      const AGridDesc a_grid_desc,
                      const BGridDesc b_grid_desc,
                      const CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
                          c_grid_desc_mblock_mperblock_nblock_nperblock,
                      const AElementwiseOperation a_element_op,
                      const BElementwiseOperation b_element_op,
                      const CElementwiseOperation c_element_op,
                      const Block2CTileMap block_2_ctile_map)
 {
 #if(defined(__gfx11__) || defined(__gfx12__))
     __shared__ char p_shared[GridwiseGemm::SharedMemTrait::lds_size];
  
     GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid,
                                                   p_b_grid,
                                                   p_c_grid,
                                                   p_shared,
                                                   a_grid_desc,
                                                   b_grid_desc,
                                                   c_grid_desc_mblock_mperblock_nblock_nperblock,
                                                   a_element_op,
                                                   b_element_op,
                                                   c_element_op,
                                                   block_2_ctile_map);
 #else
     ignore = p_a_grid;
     ignore = p_b_grid;
     ignore = p_c_grid;
     ignore = a_grid_desc;
     ignore = b_grid_desc;
     ignore = c_grid_desc_mblock_mperblock_nblock_nperblock;
     ignore = a_element_op;
     ignore = b_element_op;
     ignore = c_element_op;
     ignore = block_2_ctile_map;
 #endif // end of if (defined(__gfx11__))
 }
  
 template <index_t BlockSize,
           typename ADataType,
           typename BDataType,
           typename AccDataType,
           typename CShuffleDataType,
           typename CDataType,
           InMemoryDataOperationEnum CGlobalMemoryDataOperation,
           typename AGridDesc,
           typename BGridDesc,
           typename CGridDesc_M_N,
           typename AElementwiseOperation,
           typename BElementwiseOperation,
           typename CElementwiseOperation,
           index_t MPerBlock,
           index_t NPerBlock,
           index_t KPerBlock,
           index_t MPerWmma,
           index_t NPerWmma,
           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 AEnableLds,
           bool ABlockLdsExtraM,
           typename BBlockTransferThreadClusterLengths_K0_N_K1,
           typename BBlockTransferThreadClusterArrangeOrder,
           typename BBlockTransferSrcAccessOrder,
           index_t BBlockTransferSrcVectorDim,
           index_t BBlockTransferSrcScalarPerVector,
           index_t BBlockTransferDstScalarPerVector_K1,
           bool BThreadTransferSrcResetCoordinateAfterRun,
           bool BEnableLds,
           bool BBlockLdsExtraN,
           index_t CShuffleMRepeatPerShuffle,
           index_t CShuffleNRepeatPerShuffle,
           typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
           index_t CShuffleBlockTransferScalarPerVector_NPerBlock,
           index_t NumGemmKPrefetchStage = 1,
           LoopScheduler LoopSched       = make_default_loop_scheduler(),
           PipelineVersion PipelineVer   = PipelineVersion::v1>
 struct GridwiseGemm_Wmma
 {
     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>{};
  
     // FIX ME: To be deprecated
     static constexpr auto K1 = Number<K1Value>{};
  
     static constexpr auto MWaves = MPerBlock / (MRepeat * MPerWmma);
     static constexpr auto NWaves = NPerBlock / (NRepeat * NPerWmma);
     static constexpr auto WmmaK  = K1 == 16 ? 32 : 16;
  
     using ThisThreadBlock = ThisThreadBlock<BlockSize>;
  
     using GridwiseGemmPipe =
         remove_cvref_t<decltype(GridwiseGemmPipeline_Selector<PipelineVer,
                                                               NumGemmKPrefetchStage,
                                                               LoopSched,
                                                               AEnableLds,
                                                               BEnableLds>())>;
  
     // Describe how data store to (LDS/VGPR) buffer from Global memory
     __host__ __device__ static constexpr auto MakeABlockDescriptor()
     {
         constexpr auto a_block_desc = [&]() {
             if constexpr(AEnableLds)
             {
                 // K0->M->K1 Per Block
                 constexpr auto K0PerBlock    = KPerBlock / K1;
                 constexpr auto max_lds_align = K1;
  
                 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);
                 }
             }
             else
             {
                 constexpr auto A_KRow        = I2;
                 constexpr auto KWmmaPerblock = KPerBlock / WmmaK;
                 constexpr auto K0PerWmma     = WmmaK / A_KRow / K1;
                 // KWmma->MRepeat->MWave->K0PerWmma->KRow->MPerWmma->K1 Per Thread
                 return make_naive_tensor_descriptor(
                     make_tuple(Number<KWmmaPerblock>{},
                                Number<MRepeat>{},
                                I1,
                                Number<K0PerWmma>{},
                                I1,
                                I1,
                                K1),
                     make_tuple(Number<MRepeat>{} * Number<K0PerWmma>{} * K1,
                                Number<K0PerWmma>{} * K1,
                                Number<K0PerWmma>{} * K1,
                                K1,
                                K1,
                                K1,
                                I1));
             }
         }();
  
         return a_block_desc;
     }
  
     __host__ __device__ static constexpr auto MakeBBlockDescriptor()
     {
         constexpr auto b_block_desc = [&]() {
             if constexpr(BEnableLds)
             {
                 // K0->N->K1 Per Block
                 constexpr auto K0PerBlock    = KPerBlock / K1;
                 constexpr auto max_lds_align = K1;
  
                 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);
                 }
             }
             else
             {
  
                 constexpr auto B_KRow        = I2;
                 constexpr auto KWmmaPerblock = KPerBlock / WmmaK;
                 constexpr auto K0PerWmma     = WmmaK / B_KRow / K1;
                 // KWmma->NRepeat->MWave->K0PerWmma->KRow->MPerWmma->K1 Per Thread
                 return make_naive_tensor_descriptor(
                     make_tuple(Number<KWmmaPerblock>{},
                                Number<NRepeat>{},
                                I1,
                                Number<K0PerWmma>{},
                                I1,
                                I1,
                                K1),
                     make_tuple(Number<NRepeat>{} * Number<K0PerWmma>{} * K1,
                                Number<K0PerWmma>{} * K1,
                                Number<K0PerWmma>{} * K1,
                                K1,
                                K1,
                                K1,
                                I1));
             }
         }();
  
         return b_block_desc;
     }
  
     __host__ __device__ static constexpr auto MakeABlockSliceCopyStep()
     {
         constexpr auto a_block_copy_step = [&]() {
             if constexpr(AEnableLds)
             {
                 constexpr auto K0PerBlock = KPerBlock / K1;
  
                 return make_multi_index(K0PerBlock, 0, 0);
             }
             else
             {
                 constexpr auto KWmmaPerBlock = KPerBlock / WmmaK;
  
                 return make_multi_index(KWmmaPerBlock, 0, 0, 0, 0, 0, 0);
             }
         }();
  
         return a_block_copy_step;
     }
  
     __host__ __device__ static constexpr auto MakeBBlockSliceCopyStep()
     {
         constexpr auto b_block_copy_step = [&]() {
             if constexpr(BEnableLds)
             {
                 constexpr auto K0PerBlock = KPerBlock / K1;
  
                 return make_multi_index(K0PerBlock, 0, 0);
             }
             else
             {
                 constexpr auto KWmmaPerBlock = KPerBlock / WmmaK;
  
                 return make_multi_index(KWmmaPerBlock, 0, 0, 0, 0, 0, 0);
             }
         }();
  
         return b_block_copy_step;
     }
  
     // Describe how data read from (LDS/VGPR) buffer
     template <typename ABlockDesc_>
     __host__ __device__ static constexpr auto MakeAWaveDescriptor(const ABlockDesc_&)
     {
  
         constexpr auto a_wave_desc = [&]() {
             if constexpr(AEnableLds)
             {
                 // AK0_M_AK1 -> AK0_MRepeat_Mwaves_AKRow_MPerWmma_AK1
                 constexpr auto A_K0 = ABlockDesc_{}.GetLength(I0);
                 constexpr auto A_K1 = ABlockDesc_{}.GetLength(I2);
 #ifdef __gfx12__
                 constexpr auto A_KRow = I2;
 #else
                 constexpr auto A_KRow = I1;
 #endif
  
                 return transform_tensor_descriptor(
                     ABlockDesc_{},
                     make_tuple(make_unmerge_transform(make_tuple(Number<A_K0 / A_KRow>{}, A_KRow)),
                                make_unmerge_transform(make_tuple(
                                    Number<MRepeat>{}, Number<MWaves>{}, Number<MPerWmma>{})),
                                make_pass_through_transform(Number<A_K1>{})),
                     make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
                     make_tuple(Sequence<0, 3>{}, Sequence<1, 2, 4>{}, Sequence<5>{}));
             }
             else
             {
                 // KWmma_MRepeat_MWave_K0PerWmma_KRow_MPerWmma_K1 -> K0_MRepeat_Mwaves_MPerWmma_K1
                 constexpr auto KWmma     = ABlockDesc_{}.GetLength(I0);
                 constexpr auto K0PerWmma = ABlockDesc_{}.GetLength(I3);
                 constexpr auto A_KRow    = ABlockDesc_{}.GetLength(I4);
                 constexpr auto A_K1      = ABlockDesc_{}.GetLength(I6);
  
                 // Err: merge transform cause non-constexpr issue
  
                 // return transform_tensor_descriptor(
                 //     ABlockDesc_{},
                 //     make_tuple(make_merge_transform(make_tuple(Number<KWmma>{}, I1)),
                 //                make_pass_through_transform(Number<MRepeat>{}),
                 //                make_pass_through_transform(I1),
                 //                make_pass_through_transform(I1),
                 //                make_pass_through_transform(Number<A_K1>{})),
                 //     make_tuple(Sequence<0, 3>{},
                 //                Sequence<1>{},
                 //                Sequence<2>{},
                 //                Sequence<4>{},
                 //                Sequence<5>{}),
                 //     make_tuple(
                 //         Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{},
                 //         Sequence<4>{}));
  
                 // Workaround, Freeze transform
                 return make_naive_tensor_descriptor_packed(make_tuple(Number<KWmma * K0PerWmma>{},
                                                                       Number<MRepeat>{},
                                                                       I1,
                                                                       Number<A_KRow>{},
                                                                       I1,
                                                                       Number<A_K1>{}));
             }
         }();
  
         return a_wave_desc;
     }
  
     template <typename BBlockDesc_>
     __host__ __device__ static constexpr auto MakeBWaveDescriptor(const BBlockDesc_&)
     {
         constexpr auto b_wave_desc = [&]() {
             if constexpr(BEnableLds)
             {
                 // BK0_N_BK1 -> BK0_NRepeat_Nwaves_NPerWmma_BK1
                 constexpr auto B_K0 = BBlockDesc_{}.GetLength(I0);
                 constexpr auto B_K1 = BBlockDesc_{}.GetLength(I2);
 #ifdef __gfx12__
                 constexpr auto B_KRow = I2;
 #else
                 constexpr auto B_KRow = I1;
 #endif
                 return transform_tensor_descriptor(
                     BBlockDesc_{},
                     make_tuple(make_unmerge_transform(make_tuple(Number<B_K0 / B_KRow>{}, B_KRow)),
                                make_unmerge_transform(make_tuple(
                                    Number<NRepeat>{}, Number<NWaves>{}, Number<NPerWmma>{})),
                                make_pass_through_transform(Number<B_K1>{})),
                     make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
                     make_tuple(Sequence<0, 3>{}, Sequence<1, 2, 4>{}, Sequence<5>{}));
             }
             else
             {
                 // KWmma_MRepeat_MWave_K0PerWmma_KRow_MPerWmma_K1 -> K0_MRepeat_Mwaves_MPerWmma_K1
                 constexpr auto KWmma     = BBlockDesc_{}.GetLength(I0);
                 constexpr auto K0PerWmma = BBlockDesc_{}.GetLength(I3);
                 constexpr auto B_KRow    = BBlockDesc_{}.GetLength(I4);
                 constexpr auto B_K1      = BBlockDesc_{}.GetLength(I6);
  
                 // Workaround, Freeze transform
                 return make_naive_tensor_descriptor_packed(make_tuple(Number<KWmma * K0PerWmma>{},
                                                                       Number<NRepeat>{},
                                                                       I1,
                                                                       Number<B_KRow>{},
                                                                       I1,
                                                                       Number<B_K1>{}));
             }
         }();
  
         return b_wave_desc;
     }
  
     __host__ __device__ static constexpr auto
     // *Caution Here repeat is shuffle repeat
     GetCShuffleBlockDescriptor_MShRepeat_MPerShRepeat_NShRepeat_NPerShRepeat()
     {
         constexpr auto c_shuffle_block_desc_mshrepeat_mpershrepeat_nshrepeat_npershrepeat =
             make_naive_tensor_descriptor_packed(
                 make_tuple(I1,
                            Number<CShuffleMRepeatPerShuffle * MWaves * MPerWmma>{},
                            I1,
                            Number<CShuffleNRepeatPerShuffle * NWaves * NPerWmma>{}));
  
         return c_shuffle_block_desc_mshrepeat_mpershrepeat_nshrepeat_npershrepeat;
     }
  
     // block_id to matrix tile idx (m0, n0) mapping are controlled by {M01, N01}
     template <typename Block2CTileMap>
     __host__ __device__ static constexpr bool CheckValidity(const AGridDesc& a_grid_desc,
                                                             const BGridDesc& b_grid_desc,
                                                             const CGridDesc_M_N& c_grid_desc_m_n,
                                                             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 % (MPerWmma * MRepeat) == 0) &&
                           (NPerBlock % (NRepeat * NPerWmma)) == 0,
                       "Invalid tuning param!");
  
         const auto GetAProblemsizeMK = [&]() {
             if constexpr(AEnableLds)
             {
                 return make_tuple(a_grid_desc.GetLength(I1),
                                   a_grid_desc.GetLength(I0) * a_grid_desc.GetLength(I2));
             }
             else
             {
                 return make_tuple(a_grid_desc.GetLength(I1) * a_grid_desc.GetLength(I2) *
                                       a_grid_desc.GetLength(I5),
                                   a_grid_desc.GetLength(I0) * a_grid_desc.GetLength(I3) *
                                       a_grid_desc.GetLength(I4) * a_grid_desc.GetLength(I6));
             }
         };
  
         const auto GetBProblemsizeNK = [&]() {
             if constexpr(BEnableLds)
             {
                 return make_tuple(b_grid_desc.GetLength(I1),
                                   b_grid_desc.GetLength(I0) * b_grid_desc.GetLength(I2));
             }
             else
             {
                 return make_tuple(b_grid_desc.GetLength(I1) * b_grid_desc.GetLength(I2) *
                                       b_grid_desc.GetLength(I5),
                                   b_grid_desc.GetLength(I0) * b_grid_desc.GetLength(I3) *
                                       b_grid_desc.GetLength(I4) * b_grid_desc.GetLength(I6));
             }
         };
  
         const auto M = GetAProblemsizeMK()[I0];
         const auto N = GetBProblemsizeNK()[I0];
         const auto K = GetAProblemsizeMK()[I1];
  
         if(!(M == c_grid_desc_m_n.GetLength(I0) && N == c_grid_desc_m_n.GetLength(I1) &&
              K == GetBProblemsizeNK()[I1]))
         {
             printf("A: MxK = %d x %d, B: NxK = %d x %d, C: MxN = %d x %d\n",
                    GetAProblemsizeMK()[I0],
                    GetAProblemsizeMK()[I1],
                    GetBProblemsizeNK()[I0],
                    GetBProblemsizeNK()[I1],
                    c_grid_desc_m_n.GetLength(I0),
                    c_grid_desc_m_n.GetLength(I1));
             printf("GridwiseOp err: ProblemSize check");
             return false;
         }
  
         if(!(M % MPerBlock == 0 && N % NPerBlock == 0 && K % KPerBlock == 0))
         {
             printf("GridwiseOp err: ProblemSize division");
             return false;
         }
  
         // check gridwise gemm pipeline
         const auto num_k_loop = K / KPerBlock;
  
         if(!GridwiseGemmPipe::IsSupported(num_k_loop))
         {
             printf("GridwiseOp err: Pipeline not support this k_loop");
             return false;
         }
  
         if(!block_2_ctile_map.CheckValidity(c_grid_desc_m_n))
         {
             return false;
         }
  
         // TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
         constexpr long_index_t TwoGB = (long_index_t{1} << 31);
  
         if(!(a_grid_desc.GetElementSpaceSize() * sizeof(ADataType) <= TwoGB &&
              b_grid_desc.GetElementSpaceSize() * sizeof(BDataType) <= TwoGB))
         {
             return false;
         }
         return true;
     }
  
     __host__ __device__ static constexpr bool CalculateHasMainKBlockLoop(index_t K)
     {
         const index_t num_loop = K / KPerBlock;
  
         return GridwiseGemmPipe::CalculateHasMainLoop(num_loop);
     }
  
     __host__ __device__ static constexpr auto
     MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(const CGridDesc_M_N& c_grid_desc_m_n)
     {
         const auto M = c_grid_desc_m_n.GetLength(I0);
         const auto N = c_grid_desc_m_n.GetLength(I1);
  
         const auto MBlock = M / MPerBlock;
         const auto NBlock = N / NPerBlock;
  
         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
     __host__ __device__ static constexpr auto MakeDefaultBlock2CTileMap(
         const CGridDesc_M_N& c_grid_desc_m_n, index_t /* M01 */, index_t /* N01 */)
     {
         return BlockToCTileMap_M00_N0_M01Adapt<MPerBlock, NPerBlock, CGridDesc_M_N>(
             c_grid_desc_m_n);
     }
  
     struct SharedMemTrait
     {
         // LDS allocation for A and B: be careful of alignment
  
         static constexpr auto max_lds_align = K1;
  
         static constexpr auto a_block_space_size_aligned =
             AEnableLds ? math::integer_least_multiple(MakeABlockDescriptor().GetElementSpaceSize(),
                                                       max_lds_align)
                        : 0;
         static constexpr auto b_block_space_size_aligned =
             BEnableLds ? math::integer_least_multiple(MakeBBlockDescriptor().GetElementSpaceSize(),
                                                       max_lds_align)
                        : 0;
  
         static constexpr auto a_block_space_offset = 0;
         static constexpr auto b_block_space_offset = a_block_space_size_aligned;
  
         // LDS allocation for C shuffle in LDS
         static constexpr auto c_shuffle_block_space_size =
             GetCShuffleBlockDescriptor_MShRepeat_MPerShRepeat_NShRepeat_NPerShRepeat()
                 .GetElementSpaceSize();
  
         static constexpr auto c_shuffle_block_space_offset = 0;
  
         static constexpr auto lds_size =
             math::max(c_shuffle_block_space_size * sizeof(CShuffleDataType),
                       a_block_space_size_aligned * sizeof(ADataType) +
                           b_block_space_size_aligned * sizeof(BDataType));
     };
  
     using CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock =
         remove_cvref_t<decltype(MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
             CGridDesc_M_N{}))>;
     using DefaultBlock2CTileMap =
         remove_cvref_t<decltype(MakeDefaultBlock2CTileMap(CGridDesc_M_N{}, 1, 1))>;
  
     template <bool HasMainKBlockLoop, typename Block2CTileMap = DefaultBlock2CTileMap>
     __device__ static void Run(const ADataType* __restrict__ p_a_grid,
                                const BDataType* __restrict__ p_b_grid,
                                CDataType* __restrict__ p_c_grid,
                                void* __restrict__ p_shared,
                                const AGridDesc& a_grid_desc,
                                const BGridDesc& b_grid_desc,
                                const CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock&
                                    c_grid_desc_mblock_mperblock_nblock_nperblock,
                                const AElementwiseOperation& a_element_op,
                                const BElementwiseOperation& b_element_op,
                                const CElementwiseOperation& c_element_op,
                                const Block2CTileMap& block_2_ctile_map)
     {
         // clang-format off
 /*******************************************************************************/
 // Memory buffer zone.
         const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
             p_a_grid, a_grid_desc.GetElementSpaceSize());
         const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
             p_b_grid, b_grid_desc.GetElementSpaceSize());
         auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
             p_c_grid, c_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
  
 /*******************************************************************************/
 // BlockIdx.x -> [BlockId.m, BlockId.n]
         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; }
  
         // Store BlockId 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);
  
 /*******************************************************************************/
 // BlockLevel, A/B Matrix ThreadMapping in WMMA Source buffer, As Destinaion of BlockWise_Copy
         const auto K = [&](){
             if constexpr(AEnableLds){
                 return a_grid_desc.GetLength(I0) * a_grid_desc.GetLength(I2);
             }
             else{
                 return a_grid_desc.GetLength(I0) * a_grid_desc.GetLength(I3) 
                         * a_grid_desc.GetLength(I4) * a_grid_desc.GetLength(I6);
             }
         }();
  
         constexpr auto a_block_desc = MakeABlockDescriptor();
         constexpr auto b_block_desc = MakeBBlockDescriptor();
         
         auto a_block_trait = [&](){
             // A matrix blockwise copy
             if constexpr(AEnableLds)
             {
                 constexpr auto K0PerBlock = KPerBlock/ K1;
                 auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
                     static_cast<ADataType*>(p_shared), 
                     SharedMemTrait::a_block_space_size_aligned);        
  
                 auto a_blockwise_copy =
                     ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
 /* typename SrcElementwiseOperation,              */    AElementwiseOperation,
 /* typename DstElementwiseOperation,              */    ck::tensor_operation::element_wise::PassThrough,
 /* InMemoryDataOperationEnum DstInMemOp,          */    InMemoryDataOperationEnum::Set,
 /* typename BlockSliceLengths,                    */    Sequence<K0PerBlock, MPerBlock, K1>,
 /* typename ThreadClusterLengths,                 */    ABlockTransferThreadClusterLengths_K0_M_K1,
 /* typename ThreadClusterArrangeOrder,            */    ABlockTransferThreadClusterArrangeOrder,
 /* typename SrcData,                              */    ADataType,
 /* typename DstData,                              */    ADataType,
 /* typename SrcDesc,                              */    decltype(a_grid_desc),
 /* typename DstDesc,                              */    decltype(a_block_desc),
 /* typename SrcDimAccessOrder,                    */    ABlockTransferSrcAccessOrder,
 /* typename DstDimAccessOrder,                    */    Sequence<0, 1, 2>,
 /* index_t SrcVectorDim,                          */    ABlockTransferSrcVectorDim,
 /* index_t DstVectorDim,                          */    2,
 /* index_t SrcScalarPerVector,                    */    ABlockTransferSrcScalarPerVector,
 /* index_t DstScalarPerVector,                    */    ABlockTransferDstScalarPerVector_K1,
 /* index_t SrcScalarStrideInVector,               */    1,
 /* index_t DstScalarStrideInVector,               */    1,
 /* bool ThreadTransferSrcResetCoordinateAfterRun, */    AThreadTransferSrcResetCoordinateAfterRun,
 /* bool ThreadTransferDstResetCoordinateAfterRun, */    true,
                                                         NumGemmKPrefetchStage>(
                 a_grid_desc,
                 make_multi_index(0, m_block_data_idx_on_grid, 0),
                 a_element_op,
                 a_block_desc,
                 make_multi_index(0, 0, 0),
                 ck::tensor_operation::element_wise::PassThrough{});
  
                 return make_tuple(a_block_buf, a_blockwise_copy);
             }
             else
             {
                 // Thread-wise copy
                 // KPerBlock/WmmaK -> MRepeat -> MWaves -> K0PerWmma -> KRow -> MPerWmma -> K1
                 constexpr auto KWmmaPerBlock = KPerBlock / WmmaK;
                 constexpr auto K0PerWmma     = WmmaK/2/K1Value;
                 auto a_block_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ADataType>(
                     a_block_desc.GetElementSpaceSize());
                 
                 // Limitation: NumDim of Src and Dst descriptor should be identical
                 auto a_blockwise_copy =
                     ThreadwiseTensorSliceTransfer_v2<ADataType,
                                                      ADataType,
                                                      decltype(a_grid_desc),
                                                      decltype(a_block_desc),
                                                      Sequence<Number<KWmmaPerBlock>{},
                                                               Number<MRepeat>{},
                                                               I1,
                                                               Number<K0PerWmma>{},
                                                               I1,
                                                               I1,
                                                               Number<K1Value>{}>,
                                                      Sequence<0, 1, 2, 3, 4, 5, 6>,
                                                      6,
                                                      ABlockTransferSrcScalarPerVector,
                                                      AThreadTransferSrcResetCoordinateAfterRun,
                                                      true>(
                     a_grid_desc,
                     make_multi_index(0, 
                                      m_block_data_idx_on_grid/(MWaves * MPerWmma), 
                                      get_thread_local_1d_id() / 32,
                                      0,
                                      (get_thread_local_1d_id() % 32 )/ 16, 
                                      get_thread_local_1d_id() % 16,
                                      0));
                                 
                 return make_tuple(a_block_buf, a_blockwise_copy);
             }
         };
  
         auto b_block_trait = [&](){
             if constexpr(BEnableLds)
             {
                 constexpr auto K0PerBlock = KPerBlock/ K1;
                 auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
                     static_cast<BDataType*>(p_shared) + SharedMemTrait::b_block_space_offset, 
                     SharedMemTrait::b_block_space_size_aligned);
  
                 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,
                                                         BDataType,
                                                         BDataType,
                                                         decltype(b_grid_desc),
                                                         decltype(b_block_desc),
                                                         BBlockTransferSrcAccessOrder,
                                                         Sequence<0, 1, 2>,
                                                         BBlockTransferSrcVectorDim,
                                                         2,
                                                         BBlockTransferSrcScalarPerVector,
                                                         BBlockTransferDstScalarPerVector_K1,
                                                         1,
                                                         1,
                                                         BThreadTransferSrcResetCoordinateAfterRun,
                                                         true,
                                                         NumGemmKPrefetchStage>(
                     b_grid_desc,
                     make_multi_index(0, n_block_data_idx_on_grid, 0),
                     b_element_op,
                     b_block_desc,
                     make_multi_index(0, 0, 0),
                     ck::tensor_operation::element_wise::PassThrough{});
                 
                 return make_tuple(b_block_buf, b_blockwise_copy);
             }
             else
             {
                 // Thread-wise copy
                 // KPerBlock/WmmaK -> NRepeat -> NWaves -> WmmaK/K1 -> NPerWmma -> K1
                 constexpr auto KWmmaPerBlock = KPerBlock / WmmaK;
                 constexpr auto K0PerWmma     = WmmaK/2/K1Value;
                 auto b_block_buf = make_static_buffer<AddressSpaceEnum::Vgpr, BDataType>(
                     b_block_desc.GetElementSpaceSize());
                 
                 // Limitation: NumDim of Src and Dst descriptor should be identical
                 auto b_blockwise_copy =
                     ThreadwiseTensorSliceTransfer_v2<BDataType,
                                                      BDataType,
                                                      decltype(b_grid_desc),
                                                      decltype(b_block_desc),
                                                      Sequence<Number<KWmmaPerBlock>{},
                                                               Number<NRepeat>{},
                                                               I1,
                                                               Number<K0PerWmma>{},
                                                               I1,
                                                               I1,
                                                               Number<K1Value>{}>,
                                                      Sequence<0, 1, 2, 3, 4, 5, 6>,
                                                      6,
                                                      BBlockTransferSrcScalarPerVector,
                                                      BThreadTransferSrcResetCoordinateAfterRun,
                                                      true>(
                     b_grid_desc,
                     make_multi_index(0, 
                                      n_block_data_idx_on_grid/(NWaves * NPerWmma), 
                                      get_thread_local_1d_id() / 32,
                                      0,
                                      (get_thread_local_1d_id() % 32 )/ 16, 
                                      get_thread_local_1d_id() % 16,
                                      0));
                                 
                 return make_tuple(b_block_buf, b_blockwise_copy);
             }
         };
  
         auto a_block_buf       = a_block_trait()[I0];
         auto a_blockwise_copy  = a_block_trait()[I1];
  
         auto b_block_buf       = b_block_trait()[I0];
         auto b_blockwise_copy  = b_block_trait()[I1];
 /*******************************************************************************/
         // GEMM
         constexpr auto KPack = math::integer_least_multiple(K1, WmmaK);
  
         auto blockwise_gemm =
             BlockwiseGemmWMMA<BlockSize,
                               ADataType,
                               BDataType,
                               AccDataType,
                               decltype(MakeAWaveDescriptor(a_block_desc)),
                               decltype(MakeBWaveDescriptor(b_block_desc)),
                               MPerBlock,
                               NPerBlock,
                               KPerBlock,
                               MPerWmma,
                               NPerWmma,
                               MRepeat,
                               NRepeat,
                               KPack,
                               AEnableLds,
                               BEnableLds>{};
  
         // Prepare Register for C matrix
         auto c_thread_buf = blockwise_gemm.GetCThreadBuffer();
  
 /*******************************************************************************/        
         // Shift Per SUB_K
         constexpr auto a_block_slice_copy_step = MakeABlockSliceCopyStep();
         constexpr auto b_block_slice_copy_step = MakeBBlockSliceCopyStep();
  
         // gridwise GEMM pipeline
         const index_t KBlockMainLoop = __builtin_amdgcn_readfirstlane(K / KPerBlock);
         GridwiseGemmPipe::template Run<HasMainKBlockLoop>(a_grid_desc,
                                                           a_block_desc,
                                                           a_blockwise_copy,
                                                           a_grid_buf,
                                                           a_block_buf,
                                                           a_block_slice_copy_step,
                                                           b_grid_desc,
                                                           b_block_desc,
                                                           b_blockwise_copy,
                                                           b_grid_buf,
                                                           b_block_buf,
                                                           b_block_slice_copy_step,
                                                           blockwise_gemm,
                                                           c_thread_buf,
                                                           KBlockMainLoop);
 /*******************************************************************************/
         // write out to C, implement shuffle
         {
             // C mapping in single thread.
             constexpr auto c_thread_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs =  
             blockwise_gemm.GetCThreadDescriptor_MRepeat_MWave_MSubGroup_NRepeat_NWave_NThreadPerSubGroup_MAccVgprs();
  
             // C mapping in single block
             constexpr auto c_block_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs_tmp =
                 blockwise_gemm.GetCBlockDescriptor_MRepeat_MWave_MSubGroup_NRepeat_NWave_NThreadPerSubGroup_MAccVgprs();
  
             constexpr auto MWave              = c_block_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs_tmp.GetLength(I1);
             constexpr auto MSubGroup          = c_block_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs_tmp.GetLength(I2);
             constexpr auto NWave              = c_block_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs_tmp.GetLength(I4);
             constexpr auto NThreadPerSubGroup = c_block_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs_tmp.GetLength(I5);
             constexpr auto MAccVgprs          = c_block_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs_tmp.GetLength(I6);
  
             // LDS descriptor, shuffle and write out in MRepeat x NRepeat times
             constexpr auto c_shuffle_block_desc_mshrepeat_mpershrepeat_nshrepeat_npershrepeat =
                 GetCShuffleBlockDescriptor_MShRepeat_MPerShRepeat_NShRepeat_NPerShRepeat();
  
             auto c_shuffle_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
                 static_cast<CShuffleDataType*>(p_shared) + SharedMemTrait::c_shuffle_block_space_offset, 
                 SharedMemTrait::c_shuffle_block_space_size);
  
             constexpr auto c_block_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs = transform_tensor_descriptor(
                 c_shuffle_block_desc_mshrepeat_mpershrepeat_nshrepeat_npershrepeat,
                 make_tuple(
                     make_freeze_transform(I0),
                     make_unmerge_transform(make_tuple(
                         Number<CShuffleMRepeatPerShuffle>{}, // MRepeat per shuffle repeat
                         MWave,                               // MWave
                         MSubGroup,                           // MSubGroup * MAccVgprs = MPerWmma
                         MAccVgprs)),
                     make_freeze_transform(I0),
                     make_unmerge_transform(make_tuple(
                         Number<CShuffleNRepeatPerShuffle>{}, // NRepeat per shuffle repeat
                         NWave,                               // NWave
                         NThreadPerSubGroup))),               // NThreadPerSubGroup = NPerWmma
                 make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
                 make_tuple(Sequence<>{}, Sequence<0, 1, 2, 6>{}, Sequence<>{}, Sequence<3, 4, 5>{}));
  
             // 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);
  
             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_mrepeat_mwave_msubgroup_maccvgprs_adaptor =
             make_single_stage_tensor_adaptor(
                 make_tuple(make_merge_transform(make_tuple(MRepeat, MWave, MSubGroup, MAccVgprs))),
                 make_tuple(Sequence<0, 1, 2, 3>{}),
                 make_tuple(Sequence<0>{}));
  
             const auto n_thread_data_on_block_to_nrepeat_nwave_nthreadpersubgroup_adaptor =
             make_single_stage_tensor_adaptor(
                 make_tuple(make_merge_transform(make_tuple(NRepeat, NWave, NThreadPerSubGroup))),
                 make_tuple(Sequence<0, 1, 2>{}),
                 make_tuple(Sequence<0>{}));
             
             const auto m_thread_data_on_block_idx = m_thread_data_on_block_to_mrepeat_mwave_msubgroup_maccvgprs_adaptor.CalculateBottomIndex(
                 make_multi_index(m_thread_data_on_block));
             
             const auto n_thread_data_on_block_idx = n_thread_data_on_block_to_nrepeat_nwave_nthreadpersubgroup_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<AccDataType,
                                                    CShuffleDataType,
                                                    decltype(c_thread_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs),
                                                    decltype(c_block_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs),
                                                    ck::tensor_operation::element_wise::PassThrough,
                                                    Sequence<CShuffleMRepeatPerShuffle,
                                                             I1,
                                                             I1,
                                                             CShuffleNRepeatPerShuffle,
                                                             I1,
                                                             I1,
                                                             MAccVgprs>,
                                                    Sequence<0, 1, 2, 3, 4, 5, 6>,
                                                    6,
                                                    1, // vector write pixel
                                                    InMemoryDataOperationEnum::Set,
                                                    1,
                                                    true>{
                     c_block_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs,
                     make_multi_index(0,
                                      m_thread_data_on_block_idx[I1],
                                      m_thread_data_on_block_idx[I2],
                                      0,
                                      n_thread_data_on_block_idx[I1],
                                      n_thread_data_on_block_idx[I2],
                                      m_thread_data_on_block_idx[I3]),
                     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,
                          CShuffleMRepeatPerShuffle * MWave * MPerWmma,
                          1,
                          CShuffleNRepeatPerShuffle * NWave * NPerWmma>, // BlockSliceLengths,
                 CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
                 Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
                 CShuffleDataType,        // typename SrcData,
                 CDataType,               // typename DstData,
                 decltype(c_shuffle_block_desc_mshrepeat_mpershrepeat_nshrepeat_npershrepeat),
                 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_mshrepeat_mpershrepeat_nshrepeat_npershrepeat,
                  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 local reg & global memory
             // space filling curve for threadwise C in VGPR
             constexpr auto sfc_c_vgpr =
                 SpaceFillingCurve<Sequence<MRepeat, 1, 1, NRepeat, 1, 1, MAccVgprs>,
                                   Sequence<0, 1, 2, 3, 4, 5, 6>,
                                   Sequence<CShuffleMRepeatPerShuffle,
                                            1,
                                            1,
                                            CShuffleNRepeatPerShuffle,
                                            1,
                                            1,
                                            MAccVgprs>>{};
  
             // 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,
                                            CShuffleMRepeatPerShuffle * MWave * MPerWmma,
                                            1,
                                            CShuffleNRepeatPerShuffle * NWave * NPerWmma>>{};
  
             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_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs,
                                               sfc_c_vgpr.GetIndexTupleOfNumber(access_id),
                                               c_thread_buf,
                                               c_block_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs,
                                               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_mshrepeat_mpershrepeat_nshrepeat_npershrepeat,
                     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);
                 }
             });
         }
         // clang-format on
     }
 };
  
 } // namespace ck