/home/docs/checkouts/readthedocs.org/user_builds/advanced-micro-devices-composable-kernel/checkouts/docs-7.0.1/include/ck/tensor_operation/gpu/grid/gridwise_gemm_dl_multiple_d.hpp Source File

/home/docs/checkouts/readthedocs.org/user_builds/advanced-micro-devices-composable-kernel/checkouts/docs-7.0.1/include/ck/tensor_operation/gpu/grid/gridwise_gemm_dl_multiple_d.hpp Source File#

Composable Kernel: /home/docs/checkouts/readthedocs.org/user_builds/advanced-micro-devices-composable-kernel/checkouts/docs-7.0.1/include/ck/tensor_operation/gpu/grid/gridwise_gemm_dl_multiple_d.hpp Source File
Go to the documentation of this file.
 // SPDX-License-Identifier: MIT
 // Copyright (c) 2018-2024, 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_dl_v2r3.hpp"
 #include "ck/tensor_operation/gpu/block/blockwise_tensor_slice_transfer_v5r1.hpp"
 #include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v7.hpp"
 #include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
 #include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_set.hpp"
 #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
  
 namespace ck {
  
 template <index_t BlockSize,
           typename FloatAB,
           typename FloatAcc,
           typename DsDataType,
           typename FloatC,
           typename AElementwiseOperation,
           typename BElementwiseOperation,
           typename CDEElementwiseOperation,
           InMemoryDataOperationEnum CGlobalMemoryDataOperation,
           typename AGridDesc_K0_M_K1,
           typename BGridDesc_K0_N_K1,
           typename CGridDesc_M_N,
           index_t MPerBlock,
           index_t NPerBlock,
           index_t K0PerBlock,
           index_t K1Value,
           index_t M1PerThreadM111,
           index_t N1PerThreadN111,
           index_t KPerThread,
           typename M11N11ThreadClusterM110Xs,
           typename M11N11ThreadClusterN110Xs,
           typename ABlockTransferThreadSliceLengths_K0_M0_M1_K1,
           typename ABlockTransferThreadClusterLengths_K0_M0_M1_K1,
           typename ABlockTransferThreadClusterArrangeOrder,
           typename ABlockTransferSrcAccessOrder,
           typename ABlockTransferSrcVectorTensorLengths_K0_M0_M1_K1,
           typename ABlockTransferSrcVectorTensorContiguousDimOrder,
           typename ABlockTransferDstVectorTensorLengths_K0_M0_M1_K1,
           typename BBlockTransferThreadSliceLengths_K0_N0_N1_K1,
           typename BBlockTransferThreadClusterLengths_K0_N0_N1_K1,
           typename BBlockTransferThreadClusterArrangeOrder,
           typename BBlockTransferSrcAccessOrder,
           typename BBlockTransferSrcVectorTensorLengths_K0_N0_N1_K1,
           typename BBlockTransferSrcVectorTensorContiguousDimOrder,
           typename BBlockTransferDstVectorTensorLengths_K0_N0_N1_K1,
           typename CThreadTransferSrcDstAccessOrder,
           index_t CThreadTransferSrcDstVectorDim,
           index_t CThreadTransferDstScalarPerVector>
 struct GridwiseGemmDlMultipleD_km_kn_mn
 {
     static constexpr index_t NumDTensor = DsDataType::Size();
  
     static constexpr auto I0 = Number<0>{};
     static constexpr auto I1 = Number<1>{};
     static constexpr auto I2 = Number<2>{};
     static constexpr auto I3 = Number<3>{};
  
     // K1 should be Number<...>
     static constexpr auto K1 = Number<K1Value>{};
  
     // ck::Tuple<const D0DataType*, const D1DataType*, ...>
     static constexpr auto MakeDsGridPointer()
     {
         return generate_tuple(
             [&](auto i) {
                 using DDataType = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
  
                 return static_cast<const DDataType*>(nullptr);
             },
             Number<NumDTensor>{});
     }
  
     __host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
     {
         // TODO: change this. I think it needs multi-dimensional alignment
         constexpr auto max_lds_align = K1;
  
         // TODO: check alignment
         // A matrix in LDS memory, dst of blockwise copy
         constexpr auto a_block_desc_k_m = make_naive_tensor_descriptor_aligned(
             make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1), max_lds_align);
  
         // TODO: check alignment
         // B matrix in LDS memory, dst of blockwise copy
         constexpr auto b_block_desc_k_n = make_naive_tensor_descriptor_aligned(
             make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1), max_lds_align);
  
         // TODO: check alignment
         // LDS allocation for A and B: be careful of alignment
         constexpr auto a_block_aligned_space_size =
             math::integer_least_multiple(a_block_desc_k_m.GetElementSpaceSize(), max_lds_align);
  
         constexpr auto b_block_aligned_space_size =
             math::integer_least_multiple(b_block_desc_k_n.GetElementSpaceSize(), max_lds_align);
  
         return 2 * (a_block_aligned_space_size + b_block_aligned_space_size) * sizeof(FloatAB);
     }
  
     __host__ __device__ static constexpr bool
     CheckValidity(const AGridDesc_K0_M_K1& a_grid_desc_k0_m_k1,
                   const BGridDesc_K0_N_K1& b_grid_desc_k0_n_k1,
                   const CGridDesc_M_N& c_grid_desc_m_n)
     {
         constexpr long_index_t TwoGB = (long_index_t{1} << 31);
  
         if(!(a_grid_desc_k0_m_k1.GetElementSpaceSize() * sizeof(FloatAB) <= TwoGB &&
              b_grid_desc_k0_n_k1.GetElementSpaceSize() * sizeof(FloatAB) <= TwoGB &&
              c_grid_desc_m_n.GetElementSpaceSize() * sizeof(FloatC) <= TwoGB))
         {
             return false;
         }
  
         const auto M  = a_grid_desc_k0_m_k1.GetLength(I1);
         const auto N  = b_grid_desc_k0_n_k1.GetLength(I1);
         const auto K0 = a_grid_desc_k0_m_k1.GetLength(I0);
  
         // TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
  
         return (M == c_grid_desc_m_n.GetLength(I0) && N == c_grid_desc_m_n.GetLength(I1) &&
                 K0 == b_grid_desc_k0_n_k1.GetLength(I0) &&
                 K1 == a_grid_desc_k0_m_k1.GetLength(I2) &&
                 K1 == b_grid_desc_k0_n_k1.GetLength(I2)) &&
                (M % MPerBlock == 0 && N % NPerBlock == 0 && K0 % K0PerBlock == 0);
     }
  
     __host__ __device__ static constexpr index_t CalculateGridSize(index_t M, index_t N)
     {
         const index_t grid_size = (M / MPerBlock) * (N / NPerBlock);
  
         return grid_size;
     }
  
     __host__ __device__ static constexpr bool CalculateHasMainKBlockLoop(index_t K0)
     {
         const bool has_main_k_block_loop = (K0 + K0PerBlock) / (2 * K0PerBlock) > 1;
  
         return has_main_k_block_loop;
     }
  
     __host__ __device__ static constexpr bool CalculateHasDoubleTailKBlockLoop(index_t K0)
     {
         const bool has_double_tail_k_block_loop = (K0 / K0PerBlock) % 2 == 0;
  
         return has_double_tail_k_block_loop;
     }
  
     __host__ __device__ static constexpr auto
     MakeAGridDescriptor_K0_M0_M1_K1(const AGridDesc_K0_M_K1& a_grid_desc_k0_m_k1)
     {
         const auto K0 = a_grid_desc_k0_m_k1.GetLength(I0);
         const auto M  = a_grid_desc_k0_m_k1.GetLength(I1);
  
         const auto M1 = Number<MPerBlock>{};
         const auto M0 = M / M1;
  
         const auto a_grid_desc_k0_m0_m1_k1 =
             transform_tensor_descriptor(a_grid_desc_k0_m_k1,
                                         make_tuple(make_pass_through_transform(K0),
                                                    make_unmerge_transform(make_tuple(M0, M1)),
                                                    make_pass_through_transform(K1)),
                                         make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
                                         make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{}));
  
         return a_grid_desc_k0_m0_m1_k1;
     }
  
     __host__ __device__ static constexpr auto
     MakeBGridDescriptor_K0_N0_N1_K1(const BGridDesc_K0_N_K1& b_grid_desc_k0_n_k1)
     {
         const auto K0 = b_grid_desc_k0_n_k1.GetLength(I0);
         const auto N  = b_grid_desc_k0_n_k1.GetLength(I1);
  
         const auto N1 = Number<NPerBlock>{};
         const auto N0 = N / N1;
  
         const auto b_grid_desc_k0_n0_n1_k1 =
             transform_tensor_descriptor(b_grid_desc_k0_n_k1,
                                         make_tuple(make_pass_through_transform(K0),
                                                    make_unmerge_transform(make_tuple(N0, N1)),
                                                    make_pass_through_transform(K1)),
                                         make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
                                         make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{}));
  
         return b_grid_desc_k0_n0_n1_k1;
     }
  
     // E desc for destination in blockwise copy
     template <typename CGridDesc_M_N_>
     __host__ __device__ static constexpr auto
     MakeCGridDescriptor_M0_M10_M11_N0_N10_N11(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);
  
         constexpr auto M1 = Number<MPerBlock>{};
         constexpr auto N1 = Number<NPerBlock>{};
  
         const auto M0 = M / M1;
         const auto N0 = N / N1;
  
         constexpr auto M11 =
             Number<container_reduce(M11N11ThreadClusterM110Xs{}, math::multiplies{}, I1) *
                    M1PerThreadM111>{};
         constexpr auto N11 =
             Number<container_reduce(M11N11ThreadClusterN110Xs{}, math::multiplies{}, I1) *
                    N1PerThreadN111>{};
  
         constexpr auto M10 = M1 / M11;
         constexpr auto N10 = N1 / N11;
  
         const auto c_grid_desc_m0_m10_m11_n0_n10_n11 = transform_tensor_descriptor(
             c_grid_desc_m_n,
             make_tuple(make_unmerge_transform(make_tuple(M0, M10, M11)),
                        make_unmerge_transform(make_tuple(N0, N10, N11))),
             make_tuple(Sequence<0>{}, Sequence<1>{}),
             make_tuple(Sequence<0, 1, 2>{}, Sequence<3, 4, 5>{}));
  
         return c_grid_desc_m0_m10_m11_n0_n10_n11;
     }
  
     // Ds desc for source in blockwise copy
     template <typename DsGridDesc_M_N>
     __host__ __device__ static constexpr auto
     MakeDsGridDescriptor_M0_M10_M11_N0_N10_N11(const DsGridDesc_M_N& ds_grid_desc_m_n)
     {
         return generate_tuple(
             [&](auto i) { return MakeCGridDescriptor_M0_M10_M11_N0_N10_N11(ds_grid_desc_m_n[i]); },
             Number<NumDTensor>{});
     }
     // 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)
     {
         return BlockToCTileMap_M00_N00_M01_N01<MPerBlock, NPerBlock, CGridDesc_M_N>(
             c_grid_desc_m_n);
     }
  
     using AGridDesc_K0_M0_M1_K1 = decltype(MakeAGridDescriptor_K0_M0_M1_K1(AGridDesc_K0_M_K1{}));
     using BGridDesc_K0_N0_N1_K1 = decltype(MakeBGridDescriptor_K0_N0_N1_K1(BGridDesc_K0_N_K1{}));
     using CGridDesc_M0_M10_M11_N0_N10_N11 =
         decltype(MakeCGridDescriptor_M0_M10_M11_N0_N10_N11(CGridDesc_M_N{}));
  
     using DsGridPointer = decltype(MakeDsGridPointer());
  
     template <typename DsGridDesc_M0_M10_M11_N0_N10_N11,
               bool HasMainKBlockLoop,
               bool HasDoubleTailKBlockLoop,
               typename Block2CTileMap>
     __device__ static void
     Run(const FloatAB* __restrict__ p_a_grid,
         const FloatAB* __restrict__ p_b_grid,
         DsGridPointer p_ds_grid,
         FloatC* __restrict__ p_c_grid,
         void* __restrict__ p_shared_block,
         const AElementwiseOperation&,
         const BElementwiseOperation&,
         const CDEElementwiseOperation& cde_element_op,
         const AGridDesc_K0_M0_M1_K1& a_grid_desc_k0_m0_m1_k1,
         const BGridDesc_K0_N0_N1_K1& b_grid_desc_k0_n0_n1_k1,
         const DsGridDesc_M0_M10_M11_N0_N10_N11& ds_grid_desc_m0_m10_m11_n0_n10_n11,
         const CGridDesc_M0_M10_M11_N0_N10_N11& c_grid_desc_m0_m10_m11_n0_n10_n11,
         const Block2CTileMap& block_2_ctile_map,
         integral_constant<bool, HasMainKBlockLoop>,
         integral_constant<bool, HasDoubleTailKBlockLoop>)
     {
         const auto a_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
             p_a_grid, a_grid_desc_k0_m0_m1_k1.GetElementSpaceSize());
         const auto b_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
             p_b_grid, b_grid_desc_k0_n0_n1_k1.GetElementSpaceSize());
         auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
             p_c_grid, c_grid_desc_m0_m10_m11_n0_n10_n11.GetElementSpaceSize());
  
         // divide block work by [M, N]
         const auto c_m0_n0_block_cluster_idx =
             block_2_ctile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
  
         // HACK: this force index data into SGPR
         const index_t im0 = __builtin_amdgcn_readfirstlane(c_m0_n0_block_cluster_idx[I0]);
         const index_t in0 = __builtin_amdgcn_readfirstlane(c_m0_n0_block_cluster_idx[I1]);
  
         if(!block_2_ctile_map.ValidCTileIndex(
                make_tuple(im0, in0),
                make_tuple(c_grid_desc_m0_m10_m11_n0_n10_n11.GetLength(I0),
                           c_grid_desc_m0_m10_m11_n0_n10_n11.GetLength(I3))))
         {
             return;
         }
  
         // TODO: change this. I think it needs multi-dimensional alignment
         constexpr auto max_lds_align = K1;
  
         // TODO: check alignment
         // A matrix in LDS memory, dst of blockwise copy
         //   be careful of LDS alignment
         constexpr auto a_block_desc_k0_m0_m1_k1 = make_naive_tensor_descriptor_aligned(
             make_tuple(Number<K0PerBlock>{}, I1, Number<MPerBlock>{}, K1), max_lds_align);
  
         // TODO: check alignment
         // B matrix in LDS memory, dst of blockwise copy
         //   be careful of LDS alignment
         constexpr auto b_block_desc_k0_n0_n1_k1 = make_naive_tensor_descriptor_aligned(
             make_tuple(Number<K0PerBlock>{}, I1, Number<NPerBlock>{}, K1), max_lds_align);
  
         // TODO: check alignment
         // A matrix in LDS memory, for blockwise GEMM
         constexpr auto a_k0_m_k1_block_desc = make_naive_tensor_descriptor_aligned(
             make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1), max_lds_align);
  
         // TODO: check alignment
         // B matrix in LDS memory, for blockwise GEMM
         constexpr auto b_k0_n_k1_block_desc = make_naive_tensor_descriptor_aligned(
             make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1), max_lds_align);
  
         static_assert(a_block_desc_k0_m0_m1_k1.GetElementSpaceSize() ==
                           a_k0_m_k1_block_desc.GetElementSpaceSize() &&
                       b_block_desc_k0_n0_n1_k1.GetElementSpaceSize() ==
                           b_k0_n_k1_block_desc.GetElementSpaceSize() &&
                       "wrong!");
  
         // A matrix blockwise copy
         auto a_blockwise_copy = BlockwiseTensorSliceTransfer_v5r1<
             BlockSize,
             InMemoryDataOperationEnum::Set,
             Sequence<K0PerBlock, 1, MPerBlock, K1.value>,
             ABlockTransferThreadSliceLengths_K0_M0_M1_K1,
             ABlockTransferThreadClusterLengths_K0_M0_M1_K1,
             ABlockTransferThreadClusterArrangeOrder,
             FloatAB,
             FloatAB,
             remove_reference_t<decltype(a_grid_desc_k0_m0_m1_k1)>,
             decltype(a_block_desc_k0_m0_m1_k1),
             ABlockTransferSrcAccessOrder,
             Sequence<0, 1, 2, 3>,
             ABlockTransferSrcVectorTensorLengths_K0_M0_M1_K1, // SrcVectorTensorLengths
             ABlockTransferDstVectorTensorLengths_K0_M0_M1_K1, // DstVectorTensorLengths
             ABlockTransferSrcVectorTensorContiguousDimOrder,  // SrcVectorTensorContiguousDimOrder
             Sequence<0, 1, 2, 3>,                             // DstVectorTensorContiguousDimOrder
             false,
             true>(a_grid_desc_k0_m0_m1_k1,
                   make_multi_index(0, im0, 0, 0),
                   a_block_desc_k0_m0_m1_k1,
                   make_multi_index(0, 0, 0, 0));
  
         // B matrix blockwise copy
         auto b_blockwise_copy = BlockwiseTensorSliceTransfer_v5r1<
             BlockSize,
             InMemoryDataOperationEnum::Set,
             Sequence<K0PerBlock, 1, NPerBlock, K1.value>,
             BBlockTransferThreadSliceLengths_K0_N0_N1_K1,
             BBlockTransferThreadClusterLengths_K0_N0_N1_K1,
             BBlockTransferThreadClusterArrangeOrder,
             FloatAB,
             FloatAB,
             remove_reference_t<decltype(b_grid_desc_k0_n0_n1_k1)>,
             decltype(b_block_desc_k0_n0_n1_k1),
             BBlockTransferSrcAccessOrder,
             Sequence<0, 1, 2, 3>,
             BBlockTransferSrcVectorTensorLengths_K0_N0_N1_K1, // SrcVectorTensorLengths
             BBlockTransferDstVectorTensorLengths_K0_N0_N1_K1, // DstVectorTensorLengths
             BBlockTransferSrcVectorTensorContiguousDimOrder,  // SrcVectorTensorContiguousDimOrder
             Sequence<0, 1, 2, 3>,                             // DstVectorTensorContiguousDimOrder
             false,
             true>(b_grid_desc_k0_n0_n1_k1,
                   make_multi_index(0, in0, 0, 0),
                   b_block_desc_k0_n0_n1_k1,
                   make_multi_index(0, 0, 0, 0));
  
         // GEMM definition
         //   c_mtx += transpose(a_mtx) * b_mtx
         //     a_mtx[K0PerBlock, MPerBlock] is in LDS
         //     b_mtx[KPerBlocl, NPerBlock] is in LDS
         //     c_mtx[MPerBlock, NPerBlock] is distributed among threads, and saved in
         //       register
         const auto blockwise_gemm =
             BlockwiseGemmDl_A_BK0_BM_BK1_B_BK0_BN_BK1_C_BM0_BM1_BN0_BN1_pipeline_BM0_2_BN0_2<
                 BlockSize,
                 FloatAB,
                 FloatAB,
                 FloatAcc,
                 decltype(a_k0_m_k1_block_desc),
                 decltype(b_k0_n_k1_block_desc),
                 M1PerThreadM111,
                 N1PerThreadN111,
                 KPerThread,
                 M11N11ThreadClusterM110Xs,
                 M11N11ThreadClusterN110Xs,
                 M1PerThreadM111,
                 N1PerThreadN111>{};
  
         constexpr auto c_m10_m11_n10_n11_thread_tensor_lengths =
             decltype(blockwise_gemm)::GetCThreadTensorLengths_BM0_BM1_BN0_BN1();
  
         constexpr auto c_thread_desc_m10_m11_n10_n11 = make_naive_tensor_descriptor_packed(
             sequence_to_tuple_of_number(c_m10_m11_n10_n11_thread_tensor_lengths));
  
         // LDS allocation for A and B: be careful of alignment
         constexpr auto a_block_aligned_space_size = math::integer_least_multiple(
             a_block_desc_k0_m0_m1_k1.GetElementSpaceSize(), max_lds_align);
  
         constexpr auto b_block_aligned_space_size = math::integer_least_multiple(
             b_block_desc_k0_n0_n1_k1.GetElementSpaceSize(), max_lds_align);
  
         FloatAB* p_a_block_double = static_cast<FloatAB*>(p_shared_block);
         FloatAB* p_b_block_double =
             static_cast<FloatAB*>(p_shared_block) + 2 * a_block_aligned_space_size;
  
         // register allocation for output
         auto c_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatAcc>(
             c_thread_desc_m10_m11_n10_n11.GetElementSpaceSize());
  
         // Initialize C
         c_thread_buf.Clear();
  
         constexpr auto a_block_slice_copy_step = make_multi_index(K0PerBlock, 0, 0, 0);
         constexpr auto b_block_slice_copy_step = make_multi_index(K0PerBlock, 0, 0, 0);
  
         auto a_block_even_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
             p_a_block_double, a_block_desc_k0_m0_m1_k1.GetElementSpaceSize());
         auto b_block_even_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
             p_b_block_double, b_block_desc_k0_n0_n1_k1.GetElementSpaceSize());
  
         auto a_block_odd_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
             p_a_block_double + a_block_aligned_space_size,
             a_block_desc_k0_m0_m1_k1.GetElementSpaceSize());
         auto b_block_odd_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
             p_b_block_double + b_block_aligned_space_size,
             b_block_desc_k0_n0_n1_k1.GetElementSpaceSize());
  
         // LDS double buffer: preload data into LDS
         {
             a_blockwise_copy.RunRead(a_grid_desc_k0_m0_m1_k1, a_global_buf);
             b_blockwise_copy.RunRead(b_grid_desc_k0_n0_n1_k1, b_global_buf);
  
             a_blockwise_copy.RunWrite(a_block_desc_k0_m0_m1_k1, a_block_even_buf);
             b_blockwise_copy.RunWrite(b_block_desc_k0_n0_n1_k1, b_block_even_buf);
         }
  
         if constexpr(HasMainKBlockLoop)
         {
             const auto K0 = a_grid_desc_k0_m0_m1_k1.GetLength(I0);
  
             index_t k_block_data_begin = 0;
  
             // LDS double buffer: main body
             // use Do-While loop instead of For loop to simplify control flow
             do
             {
                 // even iteration
                 a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc_k0_m0_m1_k1,
                                                     a_block_slice_copy_step);
                 b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc_k0_n0_n1_k1,
                                                     b_block_slice_copy_step);
  
                 // LDS doubel buffer: load next data from device mem
                 a_blockwise_copy.RunRead(a_grid_desc_k0_m0_m1_k1, a_global_buf);
                 b_blockwise_copy.RunRead(b_grid_desc_k0_n0_n1_k1, b_global_buf);
  
                 block_sync_lds();
  
                 // LDS double buffer: GEMM on current data
                 blockwise_gemm.Run(c_thread_desc_m10_m11_n10_n11,
                                    a_block_even_buf,
                                    b_block_even_buf,
                                    c_thread_buf);
  
                 // LDS double buffer: store next data to LDS
                 a_blockwise_copy.RunWrite(a_block_desc_k0_m0_m1_k1, a_block_odd_buf);
                 b_blockwise_copy.RunWrite(b_block_desc_k0_n0_n1_k1, b_block_odd_buf);
  
                 // odd iteration
                 a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc_k0_m0_m1_k1,
                                                     a_block_slice_copy_step);
                 b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc_k0_n0_n1_k1,
                                                     b_block_slice_copy_step);
  
                 // LDS doubel buffer: load next data from device mem
                 a_blockwise_copy.RunRead(a_grid_desc_k0_m0_m1_k1, a_global_buf);
                 b_blockwise_copy.RunRead(b_grid_desc_k0_n0_n1_k1, b_global_buf);
  
                 block_sync_lds();
  
                 // LDS double buffer: GEMM on current data
                 blockwise_gemm.Run(
                     c_thread_desc_m10_m11_n10_n11, a_block_odd_buf, b_block_odd_buf, c_thread_buf);
  
                 // LDS double buffer: store next data to LDS
                 a_blockwise_copy.RunWrite(a_block_desc_k0_m0_m1_k1, a_block_even_buf);
                 b_blockwise_copy.RunWrite(b_block_desc_k0_n0_n1_k1, b_block_even_buf);
  
                 k_block_data_begin += 2 * K0PerBlock;
             } while(k_block_data_begin < K0 - 2 * K0PerBlock);
         }
  
         // LDS double buffer: tail
         if constexpr(HasDoubleTailKBlockLoop) // if has 2 iteration left
         {
             a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc_k0_m0_m1_k1, a_block_slice_copy_step);
             b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc_k0_n0_n1_k1, b_block_slice_copy_step);
  
             block_sync_lds();
  
             // LDS double buffer: load last data from device mem
             a_blockwise_copy.RunRead(a_grid_desc_k0_m0_m1_k1, a_global_buf);
             b_blockwise_copy.RunRead(b_grid_desc_k0_n0_n1_k1, b_global_buf);
  
             // LDS double buffer: GEMM on 2nd-last data
             blockwise_gemm.Run(
                 c_thread_desc_m10_m11_n10_n11, a_block_even_buf, b_block_even_buf, c_thread_buf);
  
             // LDS double buffer: store last data to LDS
             a_blockwise_copy.RunWrite(a_block_desc_k0_m0_m1_k1, a_block_odd_buf);
             b_blockwise_copy.RunWrite(b_block_desc_k0_n0_n1_k1, b_block_odd_buf);
  
             block_sync_lds();
  
             // LDS double buffer: GEMM on last data
             blockwise_gemm.Run(
                 c_thread_desc_m10_m11_n10_n11, a_block_odd_buf, b_block_odd_buf, c_thread_buf);
         }
         else // if has 1 iteration left
         {
             __syncthreads();
  
             // LDS double buffer: GEMM on last data
             blockwise_gemm.Run(
                 c_thread_desc_m10_m11_n10_n11, a_block_even_buf, b_block_even_buf, c_thread_buf);
         }
  
         // output: register to global memory
         {
             constexpr auto c_thread_desc_m0_m10_m11_n0_n10_n11 =
                 make_naive_tensor_descriptor_packed(
                     make_tuple(I1,
                                Number<c_m10_m11_n10_n11_thread_tensor_lengths[I0]>{},
                                Number<c_m10_m11_n10_n11_thread_tensor_lengths[I1]>{},
                                I1,
                                Number<c_m10_m11_n10_n11_thread_tensor_lengths[I2]>{},
                                Number<c_m10_m11_n10_n11_thread_tensor_lengths[I3]>{}));
  
             const auto c_m10_m11_n10_n11_thread_origin_idx_on_block =
                 blockwise_gemm.CalculateCThreadOriginOnBlock_BM0_BM1_BN0_BN1(
                     get_thread_local_1d_id());
  
             const auto ds_grid_buf = generate_tuple(
                 [&](auto i) {
                     return make_dynamic_buffer<AddressSpaceEnum::Global>(
                         p_ds_grid[i], ds_grid_desc_m0_m10_m11_n0_n10_n11[i].GetElementSpaceSize());
                 },
                 Number<NumDTensor>{});
  
             auto ds_thread_buf = generate_tuple(
                 [&](auto i) {
                     using DDataType = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
  
                     return StaticBuffer<AddressSpaceEnum::Vgpr,
                                         DDataType,
                                         c_m10_m11_n10_n11_thread_tensor_lengths[I3],
                                         true>{};
                 },
                 Number<NumDTensor>{});
  
             auto ds_threadwise_copy = generate_tuple(
                 [&](auto i) {
                     using DDataType = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
  
                     return ThreadwiseTensorSliceTransfer_v2<
                         DDataType,
                         DDataType,
                         decltype(ds_grid_desc_m0_m10_m11_n0_n10_n11[i]),
                         decltype(c_thread_desc_m0_m10_m11_n0_n10_n11),
                         Sequence<I1,
                                  I1,
                                  I1,
                                  I1,
                                  I1,
                                  Number<c_m10_m11_n10_n11_thread_tensor_lengths[I3]>{}>,
                         CThreadTransferSrcDstAccessOrder,
                         CThreadTransferSrcDstVectorDim,
                         CThreadTransferDstScalarPerVector,
                         1,
                         false>(ds_grid_desc_m0_m10_m11_n0_n10_n11[i],
                                make_multi_index(im0,
                                                 c_m10_m11_n10_n11_thread_origin_idx_on_block[I0],
                                                 c_m10_m11_n10_n11_thread_origin_idx_on_block[I1],
                                                 in0,
                                                 c_m10_m11_n10_n11_thread_origin_idx_on_block[I2],
                                                 c_m10_m11_n10_n11_thread_origin_idx_on_block[I3]));
                 },
                 Number<NumDTensor>{});
  
             static_for<0, c_m10_m11_n10_n11_thread_tensor_lengths[I0], 1>{}([&](auto m10) {
                 static_for<0, c_m10_m11_n10_n11_thread_tensor_lengths[I1], 1>{}([&](auto m11) {
                     static_for<0, c_m10_m11_n10_n11_thread_tensor_lengths[I2], 1>{}([&](auto n10) {
                         // load d matrix data
                         static_for<0, NumDTensor, 1>{}([&](auto i) {
                             ds_threadwise_copy(i).Run(ds_grid_desc_m0_m10_m11_n0_n10_n11[i],
                                                       ds_grid_buf[i],
                                                       c_thread_desc_m0_m10_m11_n0_n10_n11,
                                                       make_tuple(I0, I0, I0, I0, I0, I0),
                                                       ds_thread_buf(i));
                         });
                         // cal element op
                         static_for<0, c_m10_m11_n10_n11_thread_tensor_lengths[I3], 1>{}(
                             [&](auto i) {
                                 // get reference to src data
                                 const auto src_data_refs = generate_tie(
                                     // return type should be lvalue
                                     [&](auto iSrc) -> const auto& {
                                         return ds_thread_buf[iSrc][i];
                                     },
                                     Number<NumDTensor>{});
  
                                 // get reference to dst data
                                 constexpr index_t c_offset =
                                     c_thread_desc_m0_m10_m11_n0_n10_n11.CalculateOffset(
                                         make_tuple(0, m10, m11, 0, n10, i));
                                 auto dst_data_refs = generate_tie(
                                     // return type should be lvalue
                                     [&](auto) -> auto& { return c_thread_buf(Number<c_offset>{}); },
                                     Number<2>{});
  
                                 unpack2(cde_element_op, dst_data_refs, src_data_refs);
                             });
  
                         static_for<0, NumDTensor, 1>{}([&](auto i) {
                             ds_threadwise_copy(i).MoveSrcSliceWindow(
                                 ds_grid_desc_m0_m10_m11_n0_n10_n11[i],
                                 make_multi_index(0, 0, 0, 0, 1, 0));
                         });
                     });
                     static_for<0, NumDTensor, 1>{}([&](auto i) {
                         ds_threadwise_copy(i).MoveSrcSliceWindow(
                             ds_grid_desc_m0_m10_m11_n0_n10_n11[i],
                             make_multi_index(
                                 0, 0, 1, 0, -c_m10_m11_n10_n11_thread_tensor_lengths[I2], 0));
                     });
                 });
                 static_for<0, NumDTensor, 1>{}([&](auto i) {
                     ds_threadwise_copy(i).MoveSrcSliceWindow(
                         ds_grid_desc_m0_m10_m11_n0_n10_n11[i],
                         make_multi_index(
                             0, 1, -c_m10_m11_n10_n11_thread_tensor_lengths[I1], 0, 0, 0));
                 });
             });
  
             ThreadwiseTensorSliceTransfer_v1r3<
                 FloatAcc,
                 FloatC,
                 decltype(c_thread_desc_m0_m10_m11_n0_n10_n11),
                 decltype(c_grid_desc_m0_m10_m11_n0_n10_n11),
                 ck::tensor_operation::element_wise::PassThrough,
                 Sequence<1,
                          c_m10_m11_n10_n11_thread_tensor_lengths[I0],
                          c_m10_m11_n10_n11_thread_tensor_lengths[I1],
                          1,
                          c_m10_m11_n10_n11_thread_tensor_lengths[I2],
                          c_m10_m11_n10_n11_thread_tensor_lengths[I3]>,
                 CThreadTransferSrcDstAccessOrder,
                 CThreadTransferSrcDstVectorDim,
                 CThreadTransferDstScalarPerVector,
                 CGlobalMemoryDataOperation,
                 1,
                 true>{c_grid_desc_m0_m10_m11_n0_n10_n11,
                       make_multi_index(im0,
                                        c_m10_m11_n10_n11_thread_origin_idx_on_block[I0],
                                        c_m10_m11_n10_n11_thread_origin_idx_on_block[I1],
                                        in0,
                                        c_m10_m11_n10_n11_thread_origin_idx_on_block[I2],
                                        c_m10_m11_n10_n11_thread_origin_idx_on_block[I3]),
                       ck::tensor_operation::element_wise::PassThrough{}}
                 .Run(c_thread_desc_m0_m10_m11_n0_n10_n11,
                      make_tuple(I0, I0, I0, I0, I0, I0),
                      c_thread_buf,
                      c_grid_desc_m0_m10_m11_n0_n10_n11,
                      c_grid_buf);
         }
     }
 };
  
 } // namespace ck