diff --git a/example/65_gemm_multiply_multiply/moe_gemm1.cpp b/example/65_gemm_multiply_multiply/moe_gemm1.cpp index b94dcaf28b..2483e39967 100644 --- a/example/65_gemm_multiply_multiply/moe_gemm1.cpp +++ b/example/65_gemm_multiply_multiply/moe_gemm1.cpp @@ -133,12 +133,12 @@ using BElementOp = PassThrough; static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::Default; static constexpr ck::index_t MPerBlock = 128; +static constexpr ck::index_t MXDLPerWave = 2; +static constexpr ck::index_t NXDLPerWave = 2; static constexpr ck::index_t BLOCKSIZE = 256; static constexpr ck::index_t NPerBlock = 128; static constexpr ck::index_t MNPerXDL = 32; -static constexpr ck::index_t CShuffleMXDLPerWave = MPerBlock / 32; -static constexpr ck::index_t KPerBlock = 256 / sizeof(A0DataType); -static constexpr ck::index_t MXDLPerWave = MPerBlock / 32; //todo fix this constraint +static constexpr ck::index_t KPerBlock = 128 / sizeof(A0DataType); static constexpr ck::index_t AK1 = 16 / sizeof(A0DataType); static constexpr ck::index_t BK1 = 16 / sizeof(B0DataType); static constexpr ck::index_t EVec = 16 / sizeof(EDataType); @@ -164,7 +164,7 @@ using DeviceOpInstance = ck::tensor_operation::device::DeviceMoeGemm // mn_perxdl MNPerXDL, MNPerXDL, // mn_xdlperwave - 2, 2, + MXDLPerWave, NXDLPerWave, // a,b: loadtranfer cluster, cluster order, srcorder,VECDIM, srcpervec, dstpervec, lds_extra // S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, // S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, @@ -173,7 +173,7 @@ using DeviceOpInstance = ck::tensor_operation::device::DeviceMoeGemm // CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| // MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector| // PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl| - 2, 2, S<1, 32, 1, 8>, S, + MXDLPerWave, 1, S<1, 32, 1, 8>, S, ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v1, true, A0DataType>; // kernel 2: 128->32x128x128 // < Row, Col, DsLayout, ELayout, A0DataType, B0DataType, DsDataType, EDataType, AccDataType, CShuffleDataType, AElementOp, BElementOp, CDEElementOp, GemmSpec, 128, 32, 128, 128, 16, 16, 32, 32, 1, 2, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 8>, S<8, 8, 1>, ck::BlockGemmPipelineScheduler::Interwave, ck::BlockGemmPipelineVersion::v1, EDataType>; diff --git a/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle_v3_b_preshuffle.hpp b/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle_v3_b_preshuffle.hpp index ebafbb5d3c..4761ee2026 100644 --- a/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle_v3_b_preshuffle.hpp +++ b/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle_v3_b_preshuffle.hpp @@ -296,7 +296,8 @@ struct DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle } } } - else if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v2) + else if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v2 || + BlkGemmPipelineVer == BlockGemmPipelineVersion::v3) { if(arg.KBatch > 1) { @@ -351,7 +352,7 @@ struct DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle } else { - throw std::runtime_error("todo: only v1 & v2 support now"); + throw std::runtime_error("todo: only v1 v2 and v3 support now"); } } #if 0 @@ -359,6 +360,7 @@ struct DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle { if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v1) { +#if 0 if(arg.KBatch > 1) { if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd) @@ -405,8 +407,29 @@ struct DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle Run(kernel); } } +#endif + if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd) + { + const auto kernel = kernel_gemm_xdl_cshuffle_v3_multi_d_b_preshuffle< + GridwiseGemm, + false, + InMemoryDataOperationEnum::Set, + minimum_occupancy, + TailNumber::Odd>; + Run(kernel); + } + else + { + const auto kernel = kernel_gemm_xdl_cshuffle_v3_multi_d_b_preshuffle< + GridwiseGemm, + false, + InMemoryDataOperationEnum::Set, + minimum_occupancy, + TailNumber::Even>; + Run(kernel); + } } - else if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v2) + else if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v2 || BlkGemmPipelineVer == BlockGemmPipelineVersion::v3) { if(arg.KBatch > 1) { @@ -602,7 +625,9 @@ struct DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle {BlockGemmPipelineScheduler::Interwave, "Interwave"}}; std::map BlkGemmPipelineVersionToString{ - {BlockGemmPipelineVersion::v1, "v1"}, {BlockGemmPipelineVersion::v2, "v2"}}; + {BlockGemmPipelineVersion::v1, "v1"}, + {BlockGemmPipelineVersion::v2, "v2"}, + {BlockGemmPipelineVersion::v3, "v3"}}; // clang-format off str << "DeviceGemmXdlUniversal" diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_multi_d_b_preshuffle.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_multi_d_b_preshuffle.hpp index a594234d7a..238ab14606 100644 --- a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_multi_d_b_preshuffle.hpp +++ b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_multi_d_b_preshuffle.hpp @@ -1521,6 +1521,448 @@ struct GridwiseGemmMultiD_xdl_cshuffle_v3_b_preshuffle } } + template + __device__ static void Run_2Lds(const ADataType* p_a_grid, + const BDataType* p_b_grid, + DsGridPointer& p_ds_grid, + CDataType* p_c_grid, + void* p_shared, + void* p_shared1, + const Problem& problem, + AElementwiseOperation a_element_op, + BElementwiseOperation b_element_op, + CElementwiseOperation c_element_op) + { + const auto block_2_ctile_map = Block2CTileMapDefault{problem.M, problem.N, 4}; + Run_2Lds( + p_a_grid, + p_b_grid, + p_ds_grid, + p_c_grid, + p_shared, + p_shared1, + problem, + a_element_op, + b_element_op, + c_element_op, + block_2_ctile_map); + } + + template + __device__ static void Run_2Lds(const ADataType* p_a_grid, + const BDataType* p_b_grid, + DsGridPointer& p_ds_grid, + CDataType* p_c_grid, + void* p_shared, + void* p_shared1, + const Problem& problem, + AElementwiseOperation a_element_op, + BElementwiseOperation b_element_op, + CElementwiseOperation c_element_op, + const Block2CTileMap& block_2_ctile_map) + { + ignore = b_element_op; + const auto a_grid_desc_ak0_m_ak1 = MakeAGridDescriptor_AK0_M_AK1( + problem.M, problem.MPadded, problem.K, problem.KPadded, problem.StrideA, problem.AK0); + + const auto b_grid_desc_bpreshuffled = + MakeBGridDescriptor_Preshuffled(problem.BN0Shuffled, problem.BK0Shuffled); + const auto c_grid_desc_m_n = MakeCGridDescriptor_M_N( + problem.M, problem.MPadded, problem.N, problem.NPadded, problem.StrideC); + + const auto c_grid_desc_mblock_mperblock_nblock_nperblock = + MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock( + c_grid_desc_m_n, problem.MBlock, problem.NBlock); + + const auto a_grid_buf = make_dynamic_buffer( + p_a_grid, a_grid_desc_ak0_m_ak1.GetElementSpaceSize()); + const auto b_grid_buf = make_dynamic_buffer( + p_b_grid, b_grid_desc_bpreshuffled.GetElementSpaceSize()); + auto c_grid_buf = make_dynamic_buffer( + p_c_grid, c_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize()); + + 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; + } + + const index_t block_m_id = __builtin_amdgcn_readfirstlane(block_work_idx[I0]); + const index_t block_n_id = __builtin_amdgcn_readfirstlane(block_work_idx[I1]); + + // HACK: this force m/n_block_data_idx_on_grid into SGPR + const index_t m_block_data_idx_on_grid = + __builtin_amdgcn_readfirstlane(block_m_id * MPerBlock); + + // N0, K0, Blocksize*KPack + const index_t n_block_data_idx_on_grid = + __builtin_amdgcn_readfirstlane(block_n_id * NXdlPerWave); + + // A matrix in LDS memory, dst of blockwise copy + constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1(); + + // B matrix in LDS memory, dst of blockwise copy + // dummy + constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1(); + + // A matrix blockwise copy + auto a_blockwise_copy = + ThreadGroupTensorSliceTransfer_v4r1, + ABlockTransferThreadClusterLengths_AK0_M_AK1, + ABlockTransferThreadClusterArrangeOrder, + ADataType, + LDSTypeA, + decltype(a_grid_desc_ak0_m_ak1), + decltype(a_block_desc_ak0_m_ak1), + ABlockTransferSrcAccessOrder, + Sequence<0, 1, 2>, + ABlockTransferSrcVectorDim, + 2, + ABlockTransferSrcScalarPerVector, + ABlockTransferDstScalarPerVector_AK1, + 1, + 1, + AThreadTransferSrcResetCoordinateAfterRun, + true, + 2>( + a_grid_desc_ak0_m_ak1, + make_multi_index(0, m_block_data_idx_on_grid, 0), + a_element_op, + a_block_desc_ak0_m_ak1, + make_multi_index(0, 0, 0), + ck::tensor_operation::element_wise::PassThrough{}); + + // Thread-wise copy + // K0 -> N0/NWave -> NWave -> KLane -> NLane -> KPack + auto b_block_buf_ping = make_static_buffer( + b_block_desc_bk0_n_bk1.GetElementSpaceSize()); + auto b_block_buf_pong = make_static_buffer( + b_block_desc_bk0_n_bk1.GetElementSpaceSize()); + auto b_block_bufs = make_tuple(b_block_buf_ping, b_block_buf_pong); + + auto b_blockwise_copy = ThreadwiseTensorSliceTransfer_v2< + BDataType, + BDataType, + decltype(b_grid_desc_bpreshuffled), + decltype(b_block_desc_bk0_n_bk1), + Sequence{}, I1, Number{}, Number{}>, + Sequence<1, 2, 0, 3>, + 3, + BBlockTransferSrcScalarPerVector, + BThreadTransferSrcResetCoordinateAfterRun, + true>(b_grid_desc_bpreshuffled, + make_multi_index(n_block_data_idx_on_grid, + get_warp_local_1d_id() % NWave, + 0, + KPack * (get_thread_local_1d_id() % warpSize))); + + // LDS allocation for A and B: be careful of alignment + // Cast after lds + auto a_block_buf_ping = make_dynamic_buffer( + static_cast(p_shared), a_block_desc_ak0_m_ak1.GetElementSpaceSize()); + auto a_block_buf_pong = make_dynamic_buffer( + static_cast(p_shared1), a_block_desc_ak0_m_ak1.GetElementSpaceSize()); + auto a_block_bufs = make_tuple(a_block_buf_ping, a_block_buf_pong); + + constexpr auto a_block_slice_copy_step = make_multi_index(KPerBlock / AK1Number, 0, 0); + constexpr auto b_block_slice_copy_step = make_multi_index(0, 0, KRepeat, 0); + + // Blockwise GEMM pipeline + static_assert(std::is_default_constructible_v); + auto blockwise_gemm_pipeline = BlockwiseGemmPipe{}; + auto c_thread_buf = blockwise_gemm_pipeline.GetCThreadBuffer(); + + const index_t num_k_block_main_loop = __builtin_amdgcn_readfirstlane( + (a_grid_desc_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2)) / + KPerBlock); + + blockwise_gemm_pipeline.template Run(a_grid_desc_ak0_m_ak1, + a_block_desc_ak0_m_ak1, + a_blockwise_copy, + a_grid_buf, + a_block_bufs, + a_block_slice_copy_step, + b_grid_desc_bpreshuffled, + b_blockwise_copy, + b_grid_buf, + b_block_bufs, + b_block_slice_copy_step, + c_thread_buf, + num_k_block_main_loop); + + // shuffle C and write out + { + static_assert(MXdlPerWave % CShuffleMXdlPerWavePerShuffle == 0 && + NXdlPerWave % CShuffleNXdlPerWavePerShuffle == 0, + "wrong!"); + + constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl); + + // TODO: hacky, fix it! + constexpr auto c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2 = + blockwise_gemm_pipeline.GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(); + + // TODO: hacky, fix it! + // c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp is only used to get lengths + constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp = + blockwise_gemm_pipeline.GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(); + + constexpr auto M0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I0); + constexpr auto N0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I1); + constexpr auto M1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I2); + constexpr auto N1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I3); + constexpr auto M2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I4); + constexpr auto M3 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I5); + constexpr auto M4 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I6); + constexpr auto N2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I7); + + constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock = + GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(); + + auto c_shuffle_block_buf = make_dynamic_buffer( + static_cast(p_shared), + c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize()); + + constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2 = transform_tensor_descriptor( + c_shuffle_block_desc_mblock_mperblock_nblock_nperblock, + make_tuple( + make_freeze_transform(I0), + make_unmerge_transform(make_tuple( + Number{}, // M0 (MXdlPerWave) per shuffle + M1, // M1 = MWave + M2, // M2 * M3 * M4 = MPerXdl + M3, + M4)), + make_freeze_transform(I0), + make_unmerge_transform(make_tuple( + Number{}, // N0 (NXdlPerWave) per shuffle + N1, // N1 = NWave + N2))), // N2 = NPerXdl + make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}), + make_tuple( + Sequence<>{}, Sequence<0, 2, 4, 5, 6>{}, Sequence<>{}, Sequence<1, 3, 7>{})); + + // calculate origin of thread output tensor on global memory + // blockwise GEMM c matrix starting index + const auto c_thread_mtx_on_block = + blockwise_gemm_pipeline.CalculateCThreadOriginDataIndex(I0, I0, I0, I0); + + const index_t m_thread_data_on_block = c_thread_mtx_on_block[I0]; + const index_t n_thread_data_on_block = c_thread_mtx_on_block[I1]; + + const auto m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor = + make_single_stage_tensor_adaptor( + make_tuple(make_merge_transform(make_tuple(M0, M1, M2, M3, M4))), + make_tuple(Sequence<0, 1, 2, 3, 4>{}), + make_tuple(Sequence<0>{})); + + const auto m_thread_data_on_block_idx = + m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor.CalculateBottomIndex( + make_multi_index(m_thread_data_on_block)); + + const auto n_thread_data_on_block_to_n0_n1_n2_adaptor = + make_single_stage_tensor_adaptor( + make_tuple(make_merge_transform(make_tuple(N0, N1, N2))), + make_tuple(Sequence<0, 1, 2>{}), + make_tuple(Sequence<0>{})); + + const auto n_thread_data_on_block_idx = + n_thread_data_on_block_to_n0_n1_n2_adaptor.CalculateBottomIndex( + make_multi_index(n_thread_data_on_block)); + + // shuffle: threadwise copy C from VGPR to LDS + auto c_thread_copy_vgpr_to_lds = + ThreadwiseTensorSliceTransfer_v1r3, + 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{}}; + + using EDataType = CDataType; + + const auto ds_grid_desc_m_n = MakeDsGridDescriptor_M_N( + problem.M, problem.MPadded, problem.N, problem.NPadded, problem.StrideDs); + + const auto ds_grid_desc_mblock_mperblock_nblock_nperblock = + MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock( + ds_grid_desc_m_n, problem.MBlock, problem.NBlock); + + const auto ds_grid_buf = generate_tuple( + [&](auto i) { + return make_dynamic_buffer( + p_ds_grid[i], ds_grid_desc_m_n[i].GetElementSpaceSize()); + }, + Number{}); + + // tuple of reference to C/Ds tensor descriptors + const auto c_ds_desc_refs = concat_tuple_of_reference( + tie(c_shuffle_block_desc_mblock_mperblock_nblock_nperblock), + generate_tie( + [&](auto i) -> const auto& // return type should be reference + { return ds_grid_desc_mblock_mperblock_nblock_nperblock[i]; }, + Number{})); + + // tuple of reference to C/Ds tensor descriptors + const auto c_ds_buf_refs = concat_tuple_of_reference( + tie(c_shuffle_block_buf), + generate_tie( + [&](auto i) -> const auto& // return type should be reference + { return ds_grid_buf[i]; }, + Number{})); + + // tuple of starting index of C/Ds blockwise copy + const auto idx_c_ds_block_begin = container_concat( + make_tuple(make_multi_index(0, 0, 0, 0)), + generate_tuple( + [&](auto) { + return make_multi_index(block_work_idx[I0], 0, block_work_idx[I1], 0); + }, + Number{})); + + const auto e_grid_desc_mblock_mperblock_nblock_nperblock = + c_grid_desc_mblock_mperblock_nblock_nperblock; + + using CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock = + CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock; + const auto EGlobalMemoryDataOperation = CGlobalMemoryDataOperation; + + auto cde_block_copy_lds_and_global = ThreadGroupTensorSliceTransfer_v7r3< + ThisThreadBlock, + decltype(container_concat(make_tuple(CShuffleDataType{}), DsDataType{})), + Tuple, + decltype(c_ds_desc_refs), + decltype(tie(e_grid_desc_mblock_mperblock_nblock_nperblock)), + CElementwiseOperation, + Sequence(EGlobalMemoryDataOperation)>, // FIXME: make Sequence + // support arbitray type + Sequence<1, + CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl, + 1, + CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>, // BlockSliceLengths, + CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, + Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder, + Sequence<0, 1, 2, 3>, // typename SrcDimAccessOrder, + Sequence<0, 1, 2, 3>, // typename DstDimAccessOrder, + 3, // index_t SrcVectorDim, + 3, // index_t DstVectorDim, + CDEShuffleBlockTransferScalarPerVectors, + CShuffleBlockTransferScalarPerVector_NPerBlock, + sequence_merge_t< + Sequence, + uniform_sequence_gen_t>, // ThreadTransferSrcResetCoordinateAfterRunFlags + Sequence> // ThreadTransferDstResetCoordinateAfterRunFlags + {c_ds_desc_refs, + idx_c_ds_block_begin, + tie(e_grid_desc_mblock_mperblock_nblock_nperblock), + make_tuple(make_multi_index(block_m_id, 0, block_n_id, 0)), + c_element_op}; + + // space filling curve for threadwise C in VGPR + constexpr auto sfc_c_vgpr = + SpaceFillingCurve, + Sequence<0, 1, 2, 3, 4, 5, 6, 7>, + Sequence>{}; + + constexpr index_t num_access = sfc_c_vgpr.GetNumOfAccess(); + + // space filling curve for shuffled blockwise C/D/E + constexpr auto sfc_cde_block = + SpaceFillingCurve, + Sequence<0, 2, 1, 3>, + Sequence<1, + CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl, + 1, + CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>>{}; + + static_assert(num_access == sfc_cde_block.GetNumOfAccess(), "wrong!"); + + static_for<0, num_access, 1>{}([&](auto access_id) { + // make sure it's safe to write to LDS + block_sync_lds(); + + // each thread write its data from VGPR to LDS + c_thread_copy_vgpr_to_lds.Run(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2, + sfc_c_vgpr.GetIndexTupleOfNumber(access_id), + c_thread_buf, + c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2, + c_shuffle_block_buf); + + // make sure it's safe to read from LDS + block_sync_lds(); + + // each block copy its data from LDS to global + cde_block_copy_lds_and_global.Run( + c_ds_desc_refs, + c_ds_buf_refs, + tie(e_grid_desc_mblock_mperblock_nblock_nperblock), + tie(c_grid_buf)); + + if constexpr(access_id < num_access - 1) + { + constexpr auto cde_lds_and_global_step = + sfc_cde_block.GetForwardStep(access_id); + + // move on Ds + static_for<0, NumDTensor, 1>{}([&](auto i) { + cde_block_copy_lds_and_global.MoveSrcSliceWindow( + c_ds_desc_refs, i + I1, cde_lds_and_global_step); + }); + + // move on E + cde_block_copy_lds_and_global.MoveDstSliceWindow( + tie(e_grid_desc_mblock_mperblock_nblock_nperblock), + I0, + cde_lds_and_global_step); + } + }); + } + } }; } // namespace ck