diff --git a/example/65_gemm_multiply_multiply/CMakeLists.txt b/example/65_gemm_multiply_multiply/CMakeLists.txt new file mode 100644 index 0000000000..f3594d1535 --- /dev/null +++ b/example/65_gemm_multiply_multiply/CMakeLists.txt @@ -0,0 +1 @@ +add_example_executable(example_gemm_multiply_multiply_xdl_fp16 gemm_multiply_multiply_xdl_fp16.cpp) diff --git a/example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp16.cpp b/example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp16.cpp new file mode 100644 index 0000000000..b0e75a5594 --- /dev/null +++ b/example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp16.cpp @@ -0,0 +1,274 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved. + +#include +#include +#include +#include + +#include "ck/ck.hpp" +#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp" +#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle_v3.hpp" +#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp" +#include "ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp" + +#include "ck/library/utility/device_memory.hpp" +#include "ck/library/utility/host_tensor.hpp" +#include "ck/library/utility/host_tensor_generator.hpp" +#include "ck/library/utility/literals.hpp" +#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp" +#include "ck/library/utility/check_err.hpp" + +#include "ck/utility/blkgemmpipe_scheduler.hpp" + +template +using S = ck::Sequence; + +using F16 = ck::half_t; +using FP8 = ck::f8_t; +using F32 = float; + +using Row = ck::tensor_layout::gemm::RowMajor; +using Col = ck::tensor_layout::gemm::ColumnMajor; + +using A0DataType = FP8; +using B0DataType = FP8; +using AccDataType = F32; +using CShuffleDataType = F32; +using D0DataType = F32; +using D1DataType = F32; +using DsDataType = ck::Tuple; +using EDataType = F16; + +using A0Layout = Row; +using B0Layout = Col; +using D0Layout = Row; +using D1Layout = Col; +using DsLayout = ck::Tuple; +using ELayout = Row; + +struct MultiplyMultiply +{ + template + __host__ __device__ constexpr void + operator()(E& e, const C& c, const D0& d0, const D1& d1) const; + + template <> + __host__ __device__ constexpr void operator()( + ck::half_t& e, const float& c, const float& d0, const float& d1) const + { + const float x0_f = c * d0 * d1; + + e = ck::type_convert(x0_f); + } +}; + +using PassThrough = ck::tensor_operation::element_wise::PassThrough; + +using AElementOp = PassThrough; +using BElementOp = PassThrough; +using CDEElementOp = MultiplyMultiply; + +static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::Default; + +using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultiD_Xdl_CShuffle_V3 + // clang-format off +///######| ALayout| BLayout| DsLayout| ELayout| AData| BData| DsData| EData| AccData| CShuffle| A| B| CDE| GEMM| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| +///######| | | | | Type| Type| Type| Type| Type| DataType| Elementwise| Elementwise| Elementwise| Spacialization| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector| +///######| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl| +///######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | S| +///###### RRR + ///< Row, Row, DsLayout, ELayout, A0DataType, B0DataType, DsDataType, EDataType, AccDataType, CShuffleDataType, AElementOp, BElementOp, CDEElementOp, GemmSpec, 256, 256, 128, 64, 16, 4, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, ck::BlockGemmPipelineScheduler::Interwave, ck::BlockGemmPipelineVersion::v1, FP8>; +///###### RCR + < Row, Col, DsLayout, ELayout, A0DataType, B0DataType, DsDataType, EDataType, AccDataType, CShuffleDataType, AElementOp, BElementOp, CDEElementOp, GemmSpec, 256, 256, 128, 64, 16, 16, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, ck::BlockGemmPipelineScheduler::Interwave, ck::BlockGemmPipelineVersion::v1, FP8>; +// clang-format on + +int main(int argc, char* argv[]) +{ + bool do_verification = true; + int init_method = 1; + bool time_kernel = false; + + // GEMM shape + ck::index_t M = 3840; + ck::index_t N = 4096; + ck::index_t K = 4096; + + ck::index_t StrideA = K; + ck::index_t StrideB = N; + ck::index_t StrideD = 0; + ck::index_t StrideE = N; + + if(argc == 1) + { + // use default case + } + else if(argc == 4) + { + do_verification = std::stoi(argv[1]); + init_method = std::stoi(argv[2]); + time_kernel = std::stoi(argv[3]); + } + else if(argc == 11) + { + do_verification = std::stoi(argv[1]); + init_method = std::stoi(argv[2]); + time_kernel = std::stoi(argv[3]); + + M = std::stoi(argv[4]); + N = std::stoi(argv[5]); + K = std::stoi(argv[6]); + + StrideA = std::stoi(argv[7]); + StrideB = std::stoi(argv[8]); + StrideD = std::stoi(argv[9]); + StrideE = std::stoi(argv[10]); + } + else + { + printf("arg1: verification (0=no, 1=yes)\n"); + printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"); + printf("arg3: time kernel (0=no, 1=yes)\n"); + printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideD, StrideE\n"); + exit(0); + } + + auto f_host_tensor_descriptor = + [](std::size_t row, std::size_t col, std::size_t stride, auto layout) { + using namespace ck::literals; + + if(std::is_same::value) + { + return HostTensorDescriptor({row, col}, {stride, 1_uz}); + } + else + { + return HostTensorDescriptor({row, col}, {1_uz, stride}); + } + }; + + Tensor a0_m_k(f_host_tensor_descriptor(M, K, StrideA, A0Layout{})); + Tensor b0_k_n(f_host_tensor_descriptor(K, N, StrideB, B0Layout{})); + Tensor d0_m_n(f_host_tensor_descriptor(M, N, StrideD, D0Layout{})); + Tensor d1_m_n(f_host_tensor_descriptor(M, N, StrideD, D1Layout{})); + Tensor e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{})); + Tensor e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{})); + + std::cout << "a0_m_k: " << a0_m_k.mDesc << std::endl; + std::cout << "b0_k_n: " << b0_k_n.mDesc << std::endl; + std::cout << "d1_m_n: " << d1_m_n.mDesc << std::endl; + std::cout << "d0_m_n: " << d0_m_n.mDesc << std::endl; + std::cout << "e_m_n: " << e_m_n_host_result.mDesc << std::endl; + + switch(init_method) + { + case 0: break; + case 1: + a0_m_k.GenerateTensorValue(GeneratorTensor_2{-5, 5}); + b0_k_n.GenerateTensorValue(GeneratorTensor_2{-5, 5}); + d0_m_n.GenerateTensorValue(GeneratorTensor_2{-5, 5}); + d1_m_n.GenerateTensorValue(GeneratorTensor_2{-5, 5}); + break; + default: + a0_m_k.GenerateTensorValue(GeneratorTensor_3{0.0, 1.0}); + b0_k_n.GenerateTensorValue(GeneratorTensor_3{-0.5, 0.5}); + d0_m_n.GenerateTensorValue(GeneratorTensor_3{-0.5, 0.5}); + d1_m_n.GenerateTensorValue(GeneratorTensor_3{-0.5, 0.5}); + } + + DeviceMem a0_device_buf(sizeof(A0DataType) * a0_m_k.mDesc.GetElementSpaceSize()); + DeviceMem b0_device_buf(sizeof(B0DataType) * b0_k_n.mDesc.GetElementSpaceSize()); + DeviceMem d0_device_buf(sizeof(D0DataType) * d0_m_n.mDesc.GetElementSpaceSize()); + DeviceMem d1_device_buf(sizeof(D1DataType) * d1_m_n.mDesc.GetElementSpaceSize()); + DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize()); + + a0_device_buf.ToDevice(a0_m_k.mData.data()); + b0_device_buf.ToDevice(b0_k_n.mData.data()); + d0_device_buf.ToDevice(d0_m_n.mData.data()); + d1_device_buf.ToDevice(d1_m_n.mData.data()); + e_device_buf.ToDevice(e_m_n_device_result.mData.data()); + + auto a_element_op = AElementOp{}; + auto b_element_op = BElementOp{}; + auto cde_element_op = CDEElementOp{}; + + constexpr ck::index_t NumDTensor = DsDataType::Size(); + + constexpr auto I0 = ck::Number<0>{}; + + // do GEMM + auto device_op = DeviceOpInstance{}; + auto invoker = device_op.MakeInvoker(); + auto argument = + device_op.MakeArgument(a0_device_buf.GetDeviceBuffer(), + b0_device_buf.GetDeviceBuffer(), + std::array{d0_device_buf.GetDeviceBuffer(), + d1_device_buf.GetDeviceBuffer()}, + e_device_buf.GetDeviceBuffer(), + M, + N, + K, + StrideA, + StrideB, + std::array{I0, I0}, + StrideE, + a_element_op, + b_element_op, + cde_element_op); + + if(!device_op.IsSupportedArgument(argument)) + { + throw std::runtime_error( + "wrong! device_gemm with the specified compilation parameters does " + "not support this GEMM problem"); + } + + float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel, 20, 50}); + + std::size_t flop = std::size_t(2) * M * N * K; + std::size_t num_btype = + sizeof(A0DataType) * M * K + sizeof(B0DataType) * K * N + sizeof(EDataType) * M * N; + + float tflops = static_cast(flop) / 1.E9 / ave_time; + + float gb_per_sec = num_btype / 1.E6 / ave_time; + + std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s" + << std::endl; + + e_device_buf.FromDevice(e_m_n_device_result.mData.data()); + + if(do_verification) + { + Tensor c_m_n({M, N}); + + using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm; + auto ref_gemm = ReferenceGemmInstance{}; + auto ref_invoker = ref_gemm.MakeInvoker(); + + auto ref_argument = ref_gemm.MakeArgument( + a0_m_k, b0_k_n, c_m_n, PassThrough{}, PassThrough{}, PassThrough{}); + + ref_invoker.Run(ref_argument); + + for(int m = 0; m < M; ++m) + { + for(int n = 0; n < N; ++n) + { + cde_element_op(e_m_n_host_result(m, n), c_m_n(m, n), d0_m_n(m, n), d1_m_n(m, n)); + } + } + + e_device_buf.FromDevice(e_m_n_device_result.mData.data()); + + return ck::utils::check_err(e_m_n_device_result, e_m_n_host_result) ? 0 : 1; + } + + return 0; +} diff --git a/include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v7r3.hpp b/include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v7r3.hpp new file mode 100644 index 0000000000..46d0c6ac2e --- /dev/null +++ b/include/ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v7r3.hpp @@ -0,0 +1,220 @@ +// 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/tensor_descriptor.hpp" +#include "ck/tensor_description/tensor_descriptor_helper.hpp" +#include "ck/tensor_description/cluster_descriptor.hpp" +#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v7r3.hpp" +#include "ck/utility/is_detected.hpp" + +namespace ck { + +// Thread-group level multi-source, multi-destination tensor slice data movement +// Assume: +// 1. All sources and destinations are DynamicBuffer +// 2. Same VectorDim and ScalerPerVector for all sources and destinations +// 3. DstInMemOps are per destination tensor +// 4. ThreadTransferSrcResetCoordinateAfterRunFlags are per source tensor +// 5. ThreadTransferDstResetCoordinateAfterRunFlags are per destination tensor +// +// Does following things to avoid scratch memory issue +// 1. Pass tensor descritpors by reference (or tuple of references) +// 2. Does not keep reference to tensor descriptor +// 3. Does not construct new tensor coordinate when call Run() +template + typename SliceLengths, + typename ThreadClusterLengths, + typename ThreadClusterArrangeOrder, + typename SrcDimAccessOrder, + typename DstDimAccessOrder, + index_t SrcVectorDim, + index_t DstVectorDim, + typename SrcScalarPerVectors, + index_t DstScalarPerVector, + typename ThreadTransferSrcResetCoordinateAfterRunFlags, + typename ThreadTransferDstResetCoordinateAfterRunFlags, + index_t NumThreadScratch = 1> +struct ThreadGroupTensorSliceTransfer_v7r3 +{ + static constexpr index_t nDim = + remove_cvref_t>::GetNumOfDimension(); + + static constexpr index_t nSrc = remove_cvref_t::Size(); + static constexpr index_t nDst = remove_cvref_t::Size(); + + using Index = MultiIndex; + + static constexpr auto thread_slice_lengths = SliceLengths{} / ThreadClusterLengths{}; + + __device__ constexpr ThreadGroupTensorSliceTransfer_v7r3( + const SrcDescs& src_descs, + const StaticallyIndexedArray& src_block_slice_origins, + const DstDescs& dst_descs, + const StaticallyIndexedArray& dst_block_slice_origins, + const ElementwiseOperation& element_op) + : threadwise_transfer_(src_descs, + StaticallyIndexedArray{}, + dst_descs, + StaticallyIndexedArray{}, + element_op) + { + static_assert(nSrc == SrcDatas::Size() && nSrc == SrcDescs::Size() && + nSrc == ThreadTransferSrcResetCoordinateAfterRunFlags::Size() && + nDst == DstDatas::Size() && nDst == DstDescs::Size() && + nDst == ThreadTransferDstResetCoordinateAfterRunFlags::Size(), + "wrong!"); + + static_for<0, nSrc, 1>{}([&](auto i) { + static_assert( + nDim == remove_cvref_t>::GetNumOfDimension(), + "wrong!"); + }); + + static_for<0, nDst, 1>{}([&](auto i) { + static_assert( + nDim == remove_cvref_t>::GetNumOfDimension(), + "wrong!"); + }); + + static_assert(nDim == ThreadClusterLengths::Size() && + nDim == ThreadClusterArrangeOrder::Size() && + nDim == SrcDimAccessOrder::Size() && nDim == DstDimAccessOrder::Size(), + "wrong! nDim not consistent"); + + static_assert( + is_same{}, + "wrong! threads should be mapped to cover entire slicing window"); + + static_assert(ThreadGroup::GetNumOfThread() >= thread_cluster_desc_.GetElementSize(), + "wrong! ThreadGroup::GetNumOfThread() too small"); + + if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or + ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize()) + { + const auto thread_cluster_idx = thread_cluster_desc_.CalculateBottomIndex( + make_multi_index(ThreadGroup::GetThreadId())); + + const auto thread_data_idx_begin = thread_cluster_idx * thread_slice_lengths; + + const auto src_thread_slice_origins = generate_tuple( + [&](auto i) { return src_block_slice_origins[i] + thread_data_idx_begin; }, + Number{}); + + const auto dst_thread_slice_origins = generate_tuple( + [&](auto i) { return dst_block_slice_origins[i] + thread_data_idx_begin; }, + Number{}); + + threadwise_transfer_.SetSrcSliceOrigins(src_descs, src_thread_slice_origins); + threadwise_transfer_.SetDstSliceOrigins(dst_descs, dst_thread_slice_origins); + } + } + + template + __device__ void RunRead(const SrcDescs& src_descs, + const SrcBuffers& src_bufs, + Number thread_scratch_id = Number{}) + { + if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or + ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize()) + { + threadwise_transfer_.RunRead(src_descs, src_bufs, thread_scratch_id); + } + } + + template + using is_tuple = decltype(std::declval().IsTuple()); + + template + __device__ void RunWrite(const DstDescs& dst_descs, + DstBuffers dst_bufs, + Number thread_scratch_id = Number{}) + { + if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or + ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize()) + { + if constexpr(is_detected::value) + threadwise_transfer_.RunWrite(dst_descs, dst_bufs, thread_scratch_id); + else + threadwise_transfer_.RunWrite(dst_descs, tie(dst_bufs), thread_scratch_id); + } + } + + template + __device__ void Run(const SrcDescs& src_descs, + const SrcBuffers& src_bufs, + const DstDescs& dst_descs, + DstBuffers dst_bufs) + { + RunRead(src_descs, src_bufs); + RunWrite(dst_descs, dst_bufs); + } + + template + __device__ void + MoveSrcSliceWindow(const SrcDescs& src_descs, Number iSrc, const Index& step) + { + if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or + ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize()) + { + threadwise_transfer_.MoveSrcSliceWindow(src_descs, iSrc, step); + } + } + + __device__ void MoveSrcSliceWindow(const SrcDescs& src_descs, const Index& step) + { + static_for<0, SrcDescs::Size(), 1>{}( + [&](auto i) { MoveSrcSliceWindow(src_descs, i, step); }); + } + + template + __device__ void + MoveDstSliceWindow(const DstDescs& dst_descs, Number iDst, const Index& step) + { + if(ThreadGroup::GetNumOfThread() == thread_cluster_desc_.GetElementSize() or + ThreadGroup::GetThreadId() < thread_cluster_desc_.GetElementSize()) + { + threadwise_transfer_.MoveDstSliceWindow(dst_descs, iDst, step); + } + } + + __device__ void MoveDstSliceWindow(const DstDescs& dst_descs, const Index& step) + { + static_for<0, DstDescs::Size(), 1>{}( + [&](auto i) { MoveDstSliceWindow(dst_descs, i, step); }); + } + + private: + static constexpr auto thread_cluster_desc_ = + make_cluster_descriptor(ThreadClusterLengths{}, ThreadClusterArrangeOrder{}); + + using ThreadwiseTransfer = + ThreadwiseTensorSliceTransfer_v7r3; + + ThreadwiseTransfer threadwise_transfer_; +}; + +} // namespace ck diff --git a/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle_v3.hpp b/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle_v3.hpp new file mode 100644 index 0000000000..2275d83641 --- /dev/null +++ b/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle_v3.hpp @@ -0,0 +1,730 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved. + +#pragma once + +#include +#include + +#include "ck/utility/common_header.hpp" +#include "ck/tensor_description/tensor_descriptor.hpp" +#include "ck/tensor_description/tensor_descriptor_helper.hpp" +#include "ck/tensor_operation/gpu/device/tensor_layout.hpp" +#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp" +#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp" +#include "ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_multi_d.hpp" +#include "ck/host_utility/device_prop.hpp" +#include "ck/host_utility/kernel_launch.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { + +template +struct DeviceGemmMultiD_Xdl_CShuffle_V3 : public DeviceGemmMultipleD +{ + static constexpr index_t NumDTensor = DsDataType::Size(); + + // GridwiseGemm + using GridwiseGemm = GridwiseGemm_xdl_cshuffle_v3< + ALayout, + BLayout, + DsLayout, + CLayout, + ADataType, + BDataType, + GemmAccDataType, + CShuffleDataType, + DsDataType, + CDataType, + AElementwiseOperation, + BElementwiseOperation, + CElementwiseOperation, + GemmSpec, + BlockSize, + MPerBlock, + NPerBlock, + KPerBlock, + AK1, + BK1, + MPerXDL, + NPerXDL, + MXdlPerWave, + NXdlPerWave, + ABlockTransferThreadClusterLengths_AK0_M_AK1, + ABlockTransferThreadClusterArrangeOrder, + ABlockTransferSrcAccessOrder, + ABlockTransferSrcVectorDim, + ABlockTransferSrcScalarPerVector, + ABlockTransferDstScalarPerVector_AK1, + false, + ABlockLdsExtraM, + BBlockTransferThreadClusterLengths_BK0_N_BK1, + BBlockTransferThreadClusterArrangeOrder, + BBlockTransferSrcAccessOrder, + BBlockTransferSrcVectorDim, + BBlockTransferSrcScalarPerVector, + BBlockTransferDstScalarPerVector_BK1, + false, + BBlockLdsExtraN, + CShuffleMXdlPerWavePerShuffle, + CShuffleNXdlPerWavePerShuffle, + CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, + CDEShuffleBlockTransferScalarPerVectors, + BlkGemmPipeSched, + BlkGemmPipelineVer, + ComputeTypeA, + ComputeTypeB, + LDSTypeA, + LDSTypeB>; + + using Argument = typename GridwiseGemm::Argument; + + // Invoker + struct Invoker : public BaseInvoker + { + float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{}) + { + if(stream_config.log_level_ > 0) + { + arg.Print(); + } + + if(!GridwiseGemm::CheckValidity(arg)) + { + throw std::runtime_error("wrong! GridwiseGemm has invalid setting"); + } + + index_t gdx, gdy, gdz; + std::tie(gdx, gdy, gdz) = GridwiseGemm::CalculateGridSize(arg.M, arg.N, arg.KBatch); + + float ave_time = 0; + + index_t k_grain = arg.KBatch * KPerBlock; + index_t K_split = (arg.K + k_grain - 1) / k_grain * KPerBlock; + + const bool has_main_k_block_loop = GridwiseGemm::CalculateHasMainKBlockLoop(K_split); + + const auto Run = [&](const auto& kernel) { + if(arg.KBatch > 1) + hipGetErrorString(hipMemsetAsync(arg.p_c_grid, + 0, + arg.M * arg.N * sizeof(CDataType), + stream_config.stream_id_)); + + ave_time = launch_and_time_kernel( + stream_config, kernel, dim3(gdx, gdy, gdz), dim3(BlockSize), 0, arg); + }; + + constexpr index_t minimum_occupancy = + BlkGemmPipeSched == BlockGemmPipelineScheduler::Intrawave ? 1 : 2; + + if(has_main_k_block_loop) + { + // Tail number always full + if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v1 || + BlkGemmPipelineVer == BlockGemmPipelineVersion::v3) + { +#if 0 + if(arg.KBatch > 1) + { + const auto kernel = + kernel_gemm_xdl_cshuffle_v3; + Run(kernel); + } + else +#endif + { + const auto kernel = + kernel_gemm_xdl_cshuffle_v3; + Run(kernel); + } + } + // Tail number could be One to Seven + else if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v2) + { +#if 0 + if(arg.KBatch > 1) + { + if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::One) + { + const auto kernel = + kernel_gemm_xdl_cshuffle_v3; + Run(kernel); + } + else if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == + TailNumber::Full) + { + const auto kernel = + kernel_gemm_xdl_cshuffle_v3; + Run(kernel); + } + + if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 2) + { + if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Two) + { + const auto kernel = kernel_gemm_xdl_cshuffle_v3< + GridwiseGemm, + true, + InMemoryDataOperationEnum::AtomicAdd, + minimum_occupancy, + TailNumber::Two>; + Run(kernel); + } + } + + if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 3) + { + if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == + TailNumber::Three) + { + const auto kernel = kernel_gemm_xdl_cshuffle_v3< + GridwiseGemm, + true, + InMemoryDataOperationEnum::AtomicAdd, + minimum_occupancy, + TailNumber::Three>; + Run(kernel); + } + } + + if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 4) + { + if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == + TailNumber::Four) + { + const auto kernel = kernel_gemm_xdl_cshuffle_v3< + GridwiseGemm, + true, + InMemoryDataOperationEnum::AtomicAdd, + minimum_occupancy, + TailNumber::Four>; + Run(kernel); + } + } + + if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 5) + { + if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == + TailNumber::Five) + { + const auto kernel = kernel_gemm_xdl_cshuffle_v3< + GridwiseGemm, + true, + InMemoryDataOperationEnum::AtomicAdd, + minimum_occupancy, + TailNumber::Five>; + Run(kernel); + } + } + + if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 6) + { + if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Six) + { + const auto kernel = kernel_gemm_xdl_cshuffle_v3< + GridwiseGemm, + true, + InMemoryDataOperationEnum::AtomicAdd, + minimum_occupancy, + TailNumber::Six>; + Run(kernel); + } + } + + if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 7) + { + if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == + TailNumber::Seven) + { + const auto kernel = kernel_gemm_xdl_cshuffle_v3< + GridwiseGemm, + true, + InMemoryDataOperationEnum::AtomicAdd, + minimum_occupancy, + TailNumber::Seven>; + Run(kernel); + } + } + } + else +#endif + { + if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::One) + { + const auto kernel = + kernel_gemm_xdl_cshuffle_v3; + Run(kernel); + } + else if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == + TailNumber::Full) + { + const auto kernel = + kernel_gemm_xdl_cshuffle_v3; + Run(kernel); + } + + if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 2) + { + if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Two) + { + const auto kernel = + kernel_gemm_xdl_cshuffle_v3; + Run(kernel); + } + } + + if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 3) + { + if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == + TailNumber::Three) + { + const auto kernel = + kernel_gemm_xdl_cshuffle_v3; + Run(kernel); + } + } + + if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 4) + { + if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == + TailNumber::Four) + { + const auto kernel = + kernel_gemm_xdl_cshuffle_v3; + Run(kernel); + } + } + + if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 5) + { + if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == + TailNumber::Five) + { + const auto kernel = + kernel_gemm_xdl_cshuffle_v3; + Run(kernel); + } + } + + if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 6) + { + if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Six) + { + const auto kernel = + kernel_gemm_xdl_cshuffle_v3; + Run(kernel); + } + } + + if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 7) + { + if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == + TailNumber::Seven) + { + const auto kernel = + kernel_gemm_xdl_cshuffle_v3; + Run(kernel); + } + } + } + } + // Tail number could be Odd or Even + else if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v4) + { +#if 0 + if(arg.KBatch > 1) + { + if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd) + { + const auto kernel = kernel_gemm_xdl_cshuffle_v3_2lds< + GridwiseGemm, + true, + InMemoryDataOperationEnum::AtomicAdd, + minimum_occupancy, + TailNumber::Odd>; + Run(kernel); + } + else + { + const auto kernel = kernel_gemm_xdl_cshuffle_v3_2lds< + GridwiseGemm, + true, + InMemoryDataOperationEnum::AtomicAdd, + minimum_occupancy, + TailNumber::Even>; + Run(kernel); + } + } + else +#endif + { + if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd) + { + const auto kernel = + kernel_gemm_xdl_cshuffle_v3_2lds; + Run(kernel); + } + else + { + const auto kernel = + kernel_gemm_xdl_cshuffle_v3_2lds; + Run(kernel); + } + } + } + else + { +#if 0 + if(arg.KBatch > 1) + { + if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd) + { + const auto kernel = + kernel_gemm_xdl_cshuffle_v3; + Run(kernel); + } + else + { + const auto kernel = + kernel_gemm_xdl_cshuffle_v3; + Run(kernel); + } + } + else +#endif + { + if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd) + { + const auto kernel = + kernel_gemm_xdl_cshuffle_v3; + Run(kernel); + } + else + { + const auto kernel = + kernel_gemm_xdl_cshuffle_v3; + Run(kernel); + } + } + } + } + else + { + // Tail number always 1 + if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v1) + { +#if 0 + if(arg.KBatch > 1) + { + const auto kernel = + kernel_gemm_xdl_cshuffle_v3; + Run(kernel); + } + else +#endif + { + const auto kernel = + kernel_gemm_xdl_cshuffle_v3; + Run(kernel); + } + } + } + + return ave_time; + } + + // polymorphic + float Run(const BaseArgument* p_arg, + const StreamConfig& stream_config = StreamConfig{}) override + { + return Run(*dynamic_cast(p_arg), stream_config); + } + }; + + static constexpr bool IsValidCompilationParameter() + { + // TODO: properly implement this check + return true; + } + + static bool IsSupportedArgument(const Argument& arg) + { + if(!ck::is_xdl_supported()) + { + return false; + } + + if((arg.K % AK1 != 0 || arg.K % BK1 != 0) && !(GemmSpec == GemmSpecialization::MKPadding || + GemmSpec == GemmSpecialization::NKPadding || + GemmSpec == GemmSpecialization::MNKPadding || + GemmSpec == GemmSpecialization::KPadding)) + { + return false; + } + + return GridwiseGemm::CheckValidity(arg); + } + + // polymorphic + bool IsSupportedArgument(const BaseArgument* p_arg) override + { + return IsSupportedArgument(*dynamic_cast(p_arg)); + } + + static auto MakeArgument(const void* p_a, + const void* p_b, + std::array p_ds, + void* p_c, + index_t M, + index_t N, + index_t K, + index_t StrideA, + index_t StrideB, + std::array StrideDs, + index_t StrideC, + AElementwiseOperation a_element_op, + BElementwiseOperation b_element_op, + CElementwiseOperation c_element_op) + { + return Argument{static_cast(p_a), + static_cast(p_b), + p_ds, + static_cast(p_c), + M, + N, + K, + StrideA, + StrideB, + StrideDs, + StrideC, + 1, + a_element_op, + b_element_op, + c_element_op}; + } + + static auto MakeInvoker() { return Invoker{}; } + + // polymorphic + std::unique_ptr MakeArgumentPointer(const void* p_a, + const void* p_b, + std::array p_ds, + void* p_c, + index_t M, + index_t N, + index_t K, + index_t StrideA, + index_t StrideB, + std::array StrideDs, + index_t StrideC, + AElementwiseOperation a_element_op, + BElementwiseOperation b_element_op, + CElementwiseOperation c_element_op) override + { + return std::make_unique(static_cast(p_a), + static_cast(p_b), + p_ds, + static_cast(p_c), + M, + N, + K, + StrideA, + StrideB, + StrideDs, + StrideC, + 1, + a_element_op, + b_element_op, + c_element_op); + } + + // polymorphic + std::unique_ptr MakeInvokerPointer() override + { + return std::make_unique(Invoker{}); + } + + // polymorphic + std::string GetTypeString() const override + { + auto str = std::stringstream(); + + std::map BlkGemmPipelineSchedulerToString{ + {BlockGemmPipelineScheduler::Intrawave, "Intrawave"}, + {BlockGemmPipelineScheduler::Interwave, "Interwave"}}; + + std::map BlkGemmPipelineVersionToString{ + {BlockGemmPipelineVersion::v1, "v1"}, + {BlockGemmPipelineVersion::v2, "v2"}, + {BlockGemmPipelineVersion::v3, "v3"}, + {BlockGemmPipelineVersion::v4, "v4"}, + {BlockGemmPipelineVersion::v5, "v5"}}; + + // clang-format off + str << "DeviceGemmXdlUniversal" + << "<" + << getGemmSpecializationString(GemmSpec) << ", " + << std::string(ALayout::name)[0] + << std::string(BLayout::name)[0] + << std::string(CLayout::name)[0] + << ">" + << " BlkSize: " + << BlockSize << ", " + << "BlkTile: " + << MPerBlock<<"x"< +__global__ void +#if CK_USE_LAUNCH_BOUNDS + __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, MinimumOccupancy) +#endif + // __attribute__((amdgpu_waves_per_eu(1, 1))) + kernel_gemm_xdl_cshuffle_v3(typename GridwiseGemm::Argument karg) +{ +#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \ + defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)) + __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()]; + + auto splitk_batch_offset = typename GridwiseGemm::SplitKBatchOffset(karg); + + GridwiseGemm::template Run( + karg.p_a_grid + splitk_batch_offset.a_k_split_offset, + karg.p_b_grid + splitk_batch_offset.b_k_split_offset, + karg.p_ds_grid, + karg.p_c_grid, + p_shared, + karg, + karg.a_element_op, + karg.b_element_op, + karg.c_element_op); +#else + ignore = karg; +#endif // end of if (defined(__gfx908__) || defined(__gfx90a__)) +} + +template +__global__ void +#if CK_USE_LAUNCH_BOUNDS + __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, MinimumOccupancy) +#endif + // __attribute__((amdgpu_waves_per_eu(1, 1))) + kernel_gemm_xdl_cshuffle_v3_2lds(typename GridwiseGemm::Argument karg) +{ +#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \ + defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)) + // Pass two lds pointer is the key to tell compiler that ds_read/write + // operate on different lds chunk at same time without order dependecy + __shared__ char p_shared_0[GridwiseGemm::GetSharedMemoryNumberOfByte()]; + __shared__ char p_shared_1[GridwiseGemm::GetSharedMemoryNumberOfByte()]; + + auto splitk_batch_offset = typename GridwiseGemm::SplitKBatchOffset(karg); + + GridwiseGemm::template Run_2Lds( + karg.p_a_grid + splitk_batch_offset.a_k_split_offset, + karg.p_b_grid + splitk_batch_offset.b_k_split_offset, + karg.p_ds_grid, + karg.p_c_grid, + p_shared_0, + p_shared_1, + karg, + karg.a_element_op, + karg.b_element_op, + karg.c_element_op); +#else + ignore = karg; +#endif // end of if (defined(__gfx908__) || defined(__gfx90a__)) +} + +template +struct GridwiseGemm_xdl_cshuffle_v3 +{ + 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>{}; + + static constexpr auto CShuffleBlockTransferScalarPerVector_NPerBlock = + CDEShuffleBlockTransferScalarPerVectors{}[I0]; + + // K1 should be Number<...> + static constexpr auto AK0Number = Number{}; + static constexpr auto BK0Number = Number{}; + static constexpr auto AK1Number = Number{}; + static constexpr auto BK1Number = Number{}; + + static constexpr index_t NumDTensor = DsDataType::Size(); + + static constexpr auto MakeDsGridPointer() + { + return generate_tuple( + [&](auto i) { + using DDataType = remove_cvref_t>; + + return static_cast(nullptr); + }, + Number{}); + } + + using DsGridPointer = decltype(MakeDsGridPointer()); + + static constexpr index_t KPack = math::max( + math::lcm(AK1Number, BK1Number), + MfmaSelector::selected_mfma.k_per_blk); + + using ThisThreadBlock = ThisThreadBlock; + + __host__ static auto CalculateGridSize(index_t M, index_t N, index_t KBatch) + { + return std::make_tuple(Block2CTileMap::CalculateGridSize(M, N), 1, KBatch); + } + + __host__ static auto CalculateMPadded(index_t M) + { + return math::integer_least_multiple(M, MPerBlock); + } + + __host__ static auto CalculateNPadded(index_t N) + { + return math::integer_least_multiple(N, NPerBlock); + } + + __host__ static auto CalculateKPadded(index_t K) + { + return math::integer_divide_ceil(K, KPerBlock) * KPerBlock; + } + + __host__ static auto CalculateAK0Padded(index_t K, index_t K_Batch = 1) + { + auto K_t = K_Batch * KPerBlock; + return (K + K_t - 1) / K_t * (KPerBlock / AK1Value); + } + + __host__ static auto CalculateBK0Padded(index_t K, index_t K_Batch = 1) + { + auto K_t = K_Batch * KPerBlock; + return (K + K_t - 1) / K_t * (KPerBlock / BK1Value); + } + + __host__ static auto CalculateKPadded(index_t K, index_t K_Batch = 1) + { + auto K_t = K_Batch * KPerBlock; + return (K + K_t - 1) / K_t * KPerBlock; + } + + __host__ static auto CalculateKRead(index_t K, index_t K_Batch = 1) + { + constexpr auto KReadVec = math::lcm(AK1Number, BK1Number); + auto K_t = K_Batch * KReadVec; + return (K + K_t - 1) / K_t * KReadVec; + } + + __host__ static auto CalculateMBlock(index_t M) + { + return math::integer_divide_ceil(M, MPerBlock); + } + + __host__ static auto CalculateNBlock(index_t N) + { + return math::integer_divide_ceil(N, NPerBlock); + } + + template + __host__ __device__ static constexpr auto MakeGemmMmaTileDescriptor(const TileDesc_K0_MN_K1&) + { + constexpr index_t K0 = TileDesc_K0_MN_K1{}.GetLength(Number<0>{}); + constexpr index_t K1 = TileDesc_K0_MN_K1{}.GetLength(Number<2>{}); + + return transform_tensor_descriptor( + TileDesc_K0_MN_K1{}, + make_tuple(make_merge_transform_v3_division_mod(make_tuple(Number{}, Number{})), + make_unmerge_transform(make_tuple( + Number{}, Number{}, Number{}))), + make_tuple(Sequence<0, 2>{}, Sequence<1>{}), + make_tuple(Sequence<3>{}, Sequence<0, 1, 2>{})); + } + + __device__ static auto MakeAGridDescriptor_AK0_M_AK1( + index_t M, index_t MPad, index_t K, index_t KPad, index_t StrideA, index_t AK0) + { + const auto a_grid_desc_mraw_kraw = [&]() { + if constexpr(is_same_v) + { + return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(StrideA, I1)); + } + else if constexpr(is_same_v) + { + return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(I1, StrideA)); + } + }(); + + using GemmSpecialization = tensor_operation::device::GemmSpecialization; + + if constexpr(GemmSpec == GemmSpecialization::MKPadding || + GemmSpec == GemmSpecialization::MNKPadding) + { + // pad both M and K + const auto a_grid_desc_m_k = + transform_tensor_descriptor(a_grid_desc_mraw_kraw, + make_tuple(make_right_pad_transform(M, MPad - M), + make_right_pad_transform(K, KPad - K)), + make_tuple(Sequence<0>{}, Sequence<1>{}), + make_tuple(Sequence<0>{}, Sequence<1>{})); + + const auto a_grid_desc_ak0_m_ak1 = transform_tensor_descriptor( + a_grid_desc_m_k, + make_tuple(make_unmerge_transform(make_tuple(AK0, AK1Value)), + make_pass_through_transform(MPad)), + make_tuple(Sequence<1>{}, Sequence<0>{}), + make_tuple(Sequence<0, 2>{}, Sequence<1>{})); + + return a_grid_desc_ak0_m_ak1; + } + else if constexpr(GemmSpec == GemmSpecialization::MPadding || + GemmSpec == GemmSpecialization::MNPadding) + { + // pad M, but not K + const auto a_grid_desc_ak0_m_ak1 = transform_tensor_descriptor( + a_grid_desc_mraw_kraw, + make_tuple(make_unmerge_transform(make_tuple(AK0, AK1Value)), + make_right_pad_transform(M, MPad - M)), + make_tuple(Sequence<1>{}, Sequence<0>{}), + make_tuple(Sequence<0, 2>{}, Sequence<1>{})); + + return a_grid_desc_ak0_m_ak1; + } + else if constexpr(GemmSpec == GemmSpecialization::KPadding || + GemmSpec == GemmSpecialization::NKPadding) + { + // pad K, but not M + const auto a_grid_desc_m_k = transform_tensor_descriptor( + a_grid_desc_mraw_kraw, + make_tuple(make_pass_through_transform(M), make_right_pad_transform(K, KPad - K)), + make_tuple(Sequence<0>{}, Sequence<1>{}), + make_tuple(Sequence<0>{}, Sequence<1>{})); + + const auto a_grid_desc_ak0_m_ak1 = transform_tensor_descriptor( + a_grid_desc_m_k, + make_tuple(make_unmerge_transform(make_tuple(AK0, AK1Value)), + make_pass_through_transform(M)), + make_tuple(Sequence<1>{}, Sequence<0>{}), + make_tuple(Sequence<0, 2>{}, Sequence<1>{})); + + return a_grid_desc_ak0_m_ak1; + } + else + { + // not pad M or K + const auto a_grid_desc_ak0_m_ak1 = transform_tensor_descriptor( + a_grid_desc_mraw_kraw, + make_tuple(make_unmerge_transform(make_tuple(AK0, AK1Value)), + make_pass_through_transform(M)), + make_tuple(Sequence<1>{}, Sequence<0>{}), + make_tuple(Sequence<0, 2>{}, Sequence<1>{})); + + return a_grid_desc_ak0_m_ak1; + } + } + + __device__ static auto MakeBGridDescriptor_BK0_N_BK1( + index_t K, index_t KPad, index_t N, index_t NPad, index_t StrideB, index_t BK0) + { + const auto b_grid_desc_nraw_kraw = [&]() { + if constexpr(is_same::value) + { + return make_naive_tensor_descriptor(make_tuple(N, K), make_tuple(I1, StrideB)); + } + else if constexpr(is_same::value) + { + return make_naive_tensor_descriptor(make_tuple(N, K), make_tuple(StrideB, I1)); + } + }(); + + using GemmSpecialization = tensor_operation::device::GemmSpecialization; + + if constexpr(GemmSpec == GemmSpecialization::NKPadding || + GemmSpec == GemmSpecialization::MNKPadding) + { + // pad both N and K + const auto b_grid_desc_n_k = + transform_tensor_descriptor(b_grid_desc_nraw_kraw, + make_tuple(make_right_pad_transform(N, NPad - N), + make_right_pad_transform(K, KPad - K)), + make_tuple(Sequence<0>{}, Sequence<1>{}), + make_tuple(Sequence<0>{}, Sequence<1>{})); + + const auto b_grid_desc_bk0_n_bk1 = transform_tensor_descriptor( + b_grid_desc_n_k, + make_tuple(make_unmerge_transform(make_tuple(BK0, BK1Value)), + make_pass_through_transform(NPad)), + make_tuple(Sequence<1>{}, Sequence<0>{}), + make_tuple(Sequence<0, 2>{}, Sequence<1>{})); + + return b_grid_desc_bk0_n_bk1; + } + else if constexpr(GemmSpec == GemmSpecialization::NPadding || + GemmSpec == GemmSpecialization::MNPadding) + { + // pad N, but not K + const auto b_grid_desc_bk0_n_bk1 = transform_tensor_descriptor( + b_grid_desc_nraw_kraw, + make_tuple(make_unmerge_transform(make_tuple(BK0, BK1Value)), + make_right_pad_transform(N, NPad - N)), + make_tuple(Sequence<1>{}, Sequence<0>{}), + make_tuple(Sequence<0, 2>{}, Sequence<1>{})); + + return b_grid_desc_bk0_n_bk1; + } + else if constexpr(GemmSpec == GemmSpecialization::KPadding || + GemmSpec == GemmSpecialization::MKPadding) + { + // pad K, but not N + const auto b_grid_desc_n_k = transform_tensor_descriptor( + b_grid_desc_nraw_kraw, + make_tuple(make_pass_through_transform(N), make_right_pad_transform(K, KPad - K)), + make_tuple(Sequence<0>{}, Sequence<1>{}), + make_tuple(Sequence<0>{}, Sequence<1>{})); + + const auto b_grid_desc_bk0_n_bk1 = transform_tensor_descriptor( + b_grid_desc_n_k, + make_tuple(make_unmerge_transform(make_tuple(BK0, BK1Value)), + make_pass_through_transform(N)), + make_tuple(Sequence<1>{}, Sequence<0>{}), + make_tuple(Sequence<0, 2>{}, Sequence<1>{})); + + return b_grid_desc_bk0_n_bk1; + } + else + { + // not pad N or K + const auto b_grid_desc_bk0_n_bk1 = transform_tensor_descriptor( + b_grid_desc_nraw_kraw, + make_tuple(make_unmerge_transform(make_tuple(BK0, BK1Value)), + make_pass_through_transform(N)), + make_tuple(Sequence<1>{}, Sequence<0>{}), + make_tuple(Sequence<0, 2>{}, Sequence<1>{})); + + return b_grid_desc_bk0_n_bk1; + } + } + + template + __host__ __device__ static constexpr auto + MakeAMmaTileDescriptor_M0_M1_M2_K(const ABlockDesc_AK0_M_AK1&) + { + constexpr index_t MWaves = MPerBlock / (MXdlPerWave * MPerXdl); + + return MakeGemmMmaTileDescriptor(ABlockDesc_AK0_M_AK1{}); + } + + template + __host__ __device__ static constexpr auto + MakeBMmaTileDescriptor_N0_N1_N2_K(const BBlockDesc_BK0_N_BK1&) + { + constexpr index_t NWaves = NPerBlock / (NXdlPerWave * NPerXdl); + + return MakeGemmMmaTileDescriptor(BBlockDesc_BK0_N_BK1{}); + } + + template + __host__ __device__ static auto + MakeCGridDescriptor_M_N(index_t M, index_t MPad, index_t N, index_t NPad, index_t StrideC) + { + const auto c_grid_desc_mraw_nraw = [&]() { + if constexpr(is_same::value) + { + return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(StrideC, I1)); + } + else if constexpr(is_same::value) + { + return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(I1, StrideC)); + } + }(); + + using GemmSpecialization = tensor_operation::device::GemmSpecialization; + + if constexpr(GemmSpec == GemmSpecialization::MNPadding || + GemmSpec == GemmSpecialization::MNKPadding) + { + // pad M and N + return transform_tensor_descriptor(c_grid_desc_mraw_nraw, + make_tuple(make_right_pad_transform(M, MPad - M), + make_right_pad_transform(N, NPad - N)), + make_tuple(Sequence<0>{}, Sequence<1>{}), + make_tuple(Sequence<0>{}, Sequence<1>{})); + } + else if constexpr(GemmSpec == GemmSpecialization::MPadding || + GemmSpec == GemmSpecialization::MKPadding) + { + // pad M, but not N + return transform_tensor_descriptor( + c_grid_desc_mraw_nraw, + make_tuple(make_right_pad_transform(M, MPad - M), make_pass_through_transform(N)), + make_tuple(Sequence<0>{}, Sequence<1>{}), + make_tuple(Sequence<0>{}, Sequence<1>{})); + } + else if constexpr(GemmSpec == GemmSpecialization::NPadding || + GemmSpec == GemmSpecialization::NKPadding) + { + // pad N, but not M + return transform_tensor_descriptor( + c_grid_desc_mraw_nraw, + make_tuple(make_pass_through_transform(M), make_right_pad_transform(N, NPad - N)), + make_tuple(Sequence<0>{}, Sequence<1>{}), + make_tuple(Sequence<0>{}, Sequence<1>{})); + } + else + { + // not pad M or N + return c_grid_desc_mraw_nraw; + } + } + + __host__ __device__ static auto MakeDsGridDescriptor_M_N( + index_t M, index_t MPad, index_t N, index_t NPad, std::array StrideDs) + { + return generate_tuple( + [&](auto i) { + using DLayout = remove_cvref_t>; + return MakeCGridDescriptor_M_N(M, MPad, N, NPad, StrideDs[i]); + }, + Number{}); + } + + template + __device__ static constexpr auto MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock( + const DsGridDesc& ds_grid_desc_m_n, index_t MBlock, index_t NBlock) + { + return generate_tuple( + [&](auto i) { + return MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock( + ds_grid_desc_m_n[i], MBlock, NBlock); + }, + Number{}); + } + + using DsGridDesc_M_N = remove_cvref_t; + + struct Problem + { + __host__ Problem(index_t M_, + index_t N_, + index_t K_, + index_t StrideA_, + index_t StrideB_, + std::array StrideDs_, + index_t StrideC_, + index_t KBatch_) + : M{M_}, + N{N_}, + K{K_}, + StrideA{StrideA_}, + StrideB{StrideB_}, + StrideDs{StrideDs_}, + StrideC{StrideC_}, + KBatch{KBatch_}, + MPadded{CalculateMPadded(M_)}, + NPadded{CalculateNPadded(N_)}, + KRead{CalculateKRead(K_, KBatch_)}, + KPadded{CalculateKPadded(K_, KBatch_)}, + AK0{CalculateAK0Padded(K_, KBatch_)}, + BK0{CalculateBK0Padded(K_, KBatch_)}, + MBlock{CalculateMBlock(M_)}, + NBlock{CalculateNBlock(N_)} + { + } + + __host__ void Print() const + { + std::cout << "problem {" + << "M:" << M << ", " + << "N:" << N << ", " + << "K:" << K << ", " + << "SA:" << StrideA << ", " + << "SB:" << StrideB << ", " + << "SC:" << StrideC << ", " + << "MP:" << MPadded << ", " + << "NP:" << NPadded << ", " + << "KRead:" << KRead << ", " + << "KP:" << KPadded << ", " + << "AK0:" << AK0 << ", " + << "BK0:" << BK0 << ", " + << "MBlock: " << MBlock << ", " + << "NBlock: " << NBlock << "}" << std::endl; + } + + index_t M; + index_t N; + index_t K; + index_t StrideA; + index_t StrideB; + std::array StrideDs; + index_t StrideC; + index_t KBatch; + index_t MPadded; + index_t NPadded; + index_t KRead; + index_t KPadded; + index_t AK0; + index_t BK0; + index_t MBlock; + index_t NBlock; + }; + + // Argument + struct Argument : public tensor_operation::device::BaseArgument, public Problem + { + __host__ Argument(const ADataType* p_a_grid_, + const BDataType* p_b_grid_, + std::array p_ds_grid_, + CDataType* p_c_grid_, + index_t M_, + index_t N_, + index_t K_, + index_t StrideA_, + index_t StrideB_, + std::array StrideDs_, + index_t StrideC_, + index_t k_batch_, + AElementwiseOperation a_element_op_, + BElementwiseOperation b_element_op_, + CElementwiseOperation c_element_op_) + : Problem{M_, N_, K_, StrideA_, StrideB_, StrideDs_, StrideC_, k_batch_}, + p_a_grid{p_a_grid_}, + p_b_grid{p_b_grid_}, + p_ds_grid{}, + p_c_grid{p_c_grid_}, + a_element_op{a_element_op_}, + b_element_op{b_element_op_}, + c_element_op{c_element_op_} + { + + // populate pointer, desc for Ds + static_for<0, NumDTensor, 1>{}([&](auto i) { + using DDataType_ = remove_cvref_t>; + + // D pointer + p_ds_grid(i) = static_cast(p_ds_grid_[i]); + }); + } + + const ADataType* p_a_grid; + const BDataType* p_b_grid; + DsGridPointer p_ds_grid; + CDataType* p_c_grid; + + const AElementwiseOperation a_element_op; + const BElementwiseOperation b_element_op; + const CElementwiseOperation c_element_op; + }; + + struct SplitKBatchOffset + { + __device__ SplitKBatchOffset(Argument& karg) + { + if constexpr(is_same_v) + { + a_k_split_offset = blockIdx.z * karg.KRead; + } + else if constexpr(is_same_v) + { + a_k_split_offset = blockIdx.z * karg.KRead * karg.M; + } + + if constexpr(is_same_v) + { + b_k_split_offset = blockIdx.z * karg.KRead * karg.N; + } + else if constexpr(is_same_v) + { + b_k_split_offset = blockIdx.z * karg.KRead; + } + + if(blockIdx.z < static_cast(karg.KBatch - 1)) + { + karg.K = karg.KRead; + } + else + { + karg.K = karg.K - karg.KRead * (karg.KBatch - 1); + } + } + + index_t a_k_split_offset; + index_t b_k_split_offset; + }; + + __device__ static constexpr auto GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1() + { + // A matrix in LDS memory, dst of blockwise copy + if constexpr(ABlockLdsExtraM) + { + return make_naive_tensor_descriptor( + make_tuple(AK0Number, Number{}, AK1Number), + make_tuple(AK1Number, Number{}, I1)); + } + // xor tensor transformation request more unnecessary vgpr usage, would cause register spill + // in some cases. + else if constexpr(is_same::value) + { + constexpr auto MLdsLayer = 32 * 4 / KPerBlock / sizeof(LDSTypeA) < 1 + ? 1 + : 32 * 4 / KPerBlock / sizeof(LDSTypeA); + constexpr auto a_lds_block_desc = make_naive_tensor_descriptor( + make_tuple( + AK0Number * Number{}, Number{}, AK1Number), + make_tuple(AK1Number, Number{}, I1)); + + constexpr auto a_lds_block_desc_permuted = transform_tensor_descriptor( + a_lds_block_desc, + make_tuple(make_xor_transform(make_tuple(Number{}, + Number{})), + make_pass_through_transform(AK1Number)), + make_tuple(Sequence<1, 0>{}, Sequence<2>{}), + make_tuple(Sequence<1, 0>{}, Sequence<2>{})); + + constexpr auto a_lds_block_desc_ak0_mldslayer_m_ak1 = transform_tensor_descriptor( + a_lds_block_desc_permuted, + make_tuple(make_unmerge_transform(make_tuple(AK0Number, Number{})), + make_pass_through_transform(Number{}), + make_pass_through_transform(AK1Number)), + make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}), + make_tuple(Sequence<0, 2>{}, Sequence<1>{}, Sequence<3>{})); + + constexpr auto a_lds_block_desc_ak0_m_ak1 = transform_tensor_descriptor( + a_lds_block_desc_ak0_mldslayer_m_ak1, + make_tuple(make_pass_through_transform(AK0Number), + make_merge_transform_v3_division_mod( + make_tuple(Number{}, Number{})), + make_pass_through_transform(AK1Number)), + make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{}), + make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{})); + + return a_lds_block_desc_ak0_m_ak1; + } + else // ColumnMajor A + { + // kfold and mpair dimension is not always required. + // more dimension in merge_transform increase the difficulty of generating immarg offset + // for compiler. + constexpr auto M0 = ABlockTransferThreadClusterLengths_AK0_M_AK1{}.At(I1); + constexpr auto M1 = MPerBlock / M0; + + constexpr auto KThreadWrite = ABlockTransferThreadClusterLengths_AK0_M_AK1{}.At(I0); + constexpr auto K0PerThreadWrite = AK0Number / KThreadWrite; + constexpr auto KThreadRead = 64 / MPerXdl; + constexpr auto K0PerThreadRead = AK0Number / KThreadRead; + + constexpr auto kfold = (AK1Number * M0 * sizeof(LDSTypeA) > 128) + ? 1 + : 128 / (AK1Number * M0 * sizeof(LDSTypeA)); + constexpr auto KThreadReadPerm = + (kfold * K0PerThreadWrite / K0PerThreadRead) > 1 + ? KThreadRead / (kfold * K0PerThreadWrite / K0PerThreadRead) + : KThreadRead; + + // 1<=mpair<=n0 + constexpr auto mpair = (AK1Number * MPerXdl * sizeof(LDSTypeA) > 128) + ? 1 + : ((128 / (AK1Number * MPerXdl * sizeof(LDSTypeA))) > M0 + ? M0 + : 128 / (AK1Number * MPerXdl * sizeof(LDSTypeA))); + + constexpr auto a_lds_block_desc = make_naive_tensor_descriptor_packed( + make_tuple(Number{}, + Number{}, + Number{}, + Number{}, + Number{}, + AK1Number)); + + constexpr auto a_lds_block_desc_permuted = transform_tensor_descriptor( + a_lds_block_desc, + make_tuple( + make_pass_through_transform(Number{}), + make_pass_through_transform(Number{}), + make_xor_transform( + make_tuple(Number{}, Number{})), + make_pass_through_transform(Number{}), + make_pass_through_transform(AK1Number)), + make_tuple( + Sequence<0>{}, Sequence<1>{}, Sequence<2, 3>{}, Sequence<4>{}, Sequence<5>{}), + make_tuple( + Sequence<0>{}, Sequence<1>{}, Sequence<2, 3>{}, Sequence<4>{}, Sequence<5>{})); + + constexpr auto a_lds_block_desc_unmerged = transform_tensor_descriptor( + a_lds_block_desc_permuted, + make_tuple( + make_pass_through_transform(Number{}), + make_pass_through_transform(Number{}), + make_unmerge_transform(make_tuple(Number{}, Number{})), + make_unmerge_transform(make_tuple(Number{}, Number{})), + make_pass_through_transform(Number{}), + make_pass_through_transform(AK1Number)), + make_tuple(Sequence<0>{}, + Sequence<1>{}, + Sequence<2>{}, + Sequence<3>{}, + Sequence<4>{}, + Sequence<5>{}), + make_tuple(Sequence<1>{}, + Sequence<2>{}, + Sequence<0, 3>{}, + Sequence<4, 5>{}, + Sequence<6>{}, + Sequence<7>{})); + + constexpr auto a_lds_block_desc_ak0_m_ak1 = transform_tensor_descriptor( + a_lds_block_desc_unmerged, + make_tuple(make_merge_transform_v3_division_mod( + make_tuple(Number{}, + Number{}, + Number{}, + Number{})), + make_merge_transform_v3_division_mod( + make_tuple(Number{}, Number{}, Number{})), + make_pass_through_transform(AK1Number)), + make_tuple(Sequence<0, 1, 4, 2>{}, Sequence<5, 6, 3>{}, Sequence<7>{}), + make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{})); + + return a_lds_block_desc_ak0_m_ak1; + } + } + + __device__ static constexpr auto GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1() + { + // B matrix in LDS memory, dst of blockwise copy + if constexpr(BBlockLdsExtraN) + { + return make_naive_tensor_descriptor( + make_tuple(BK0Number, Number{}, BK1Number), + make_tuple(BK1Number, Number{}, I1)); + } + else if constexpr(is_same::value) + { + // NLdsLayer * K0 as logical Bank + constexpr auto NLdsLayer = 32 * 4 / KPerBlock / sizeof(LDSTypeB) < 1 + ? 1 + : 32 * 4 / KPerBlock / sizeof(LDSTypeB); + ; + constexpr auto b_lds_block_desc = make_naive_tensor_descriptor( + make_tuple( + BK0Number * Number{}, Number{}, BK1Number), + make_tuple(BK1Number, Number{}, I1)); + + constexpr auto b_lds_block_desc_permuted = transform_tensor_descriptor( + b_lds_block_desc, + make_tuple(make_xor_transform(make_tuple(Number{}, + Number{})), + make_pass_through_transform(BK1Number)), + make_tuple(Sequence<1, 0>{}, Sequence<2>{}), + make_tuple(Sequence<1, 0>{}, Sequence<2>{})); + + constexpr auto b_lds_block_desc_bk0_nldslayer_n_bk1 = transform_tensor_descriptor( + b_lds_block_desc_permuted, + make_tuple(make_unmerge_transform(make_tuple(BK0Number, Number{})), + make_pass_through_transform(Number{}), + make_pass_through_transform(BK1Number)), + make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}), + make_tuple(Sequence<0, 2>{}, Sequence<1>{}, Sequence<3>{})); + + constexpr auto b_lds_block_desc_bk0_n_bk1 = transform_tensor_descriptor( + b_lds_block_desc_bk0_nldslayer_n_bk1, + make_tuple(make_pass_through_transform(BK0Number), + make_merge_transform_v3_division_mod( + make_tuple(Number{}, Number{})), + make_pass_through_transform(BK1Number)), + make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{}), + make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{})); + + return b_lds_block_desc_bk0_n_bk1; + } + else // RowMajor B + { + constexpr auto N0 = BBlockTransferThreadClusterLengths_BK0_N_BK1{}.At(I1); + constexpr auto N1 = NPerBlock / N0; + + constexpr auto KThreadWrite = BBlockTransferThreadClusterLengths_BK0_N_BK1{}.At(I0); + constexpr auto K0PerThreadWrite = BK0Number / KThreadWrite; + constexpr auto KThreadRead = 64 / NPerXdl; + constexpr auto K0PerThreadRead = BK0Number / KThreadRead; + + constexpr auto kfold = (BK1Number * N0 * sizeof(LDSTypeB) > 128) + ? 1 + : 128 / (BK1Number * N0 * sizeof(LDSTypeB)); + constexpr auto KThreadReadPerm = + (kfold * K0PerThreadWrite / K0PerThreadRead) > 1 + ? KThreadRead / (kfold * K0PerThreadWrite / K0PerThreadRead) + : KThreadRead; + + // 1<=npair<=n0 + constexpr auto npair = (BK1Number * NPerXdl * sizeof(LDSTypeB) > 128) + ? 1 + : ((128 / (BK1Number * NPerXdl * sizeof(LDSTypeB))) > N0 + ? N0 + : 128 / (BK1Number * NPerXdl * sizeof(LDSTypeB))); + + constexpr auto b_lds_block_desc = make_naive_tensor_descriptor_packed( + make_tuple(Number{}, + Number{}, + Number{}, + Number{}, + Number{}, + BK1Number)); + + constexpr auto b_lds_block_desc_permuted = transform_tensor_descriptor( + b_lds_block_desc, + make_tuple( + make_pass_through_transform(Number{}), + make_pass_through_transform(Number{}), + make_xor_transform( + make_tuple(Number{}, Number{})), + make_pass_through_transform(Number{}), + make_pass_through_transform(BK1Number)), + make_tuple( + Sequence<0>{}, Sequence<1>{}, Sequence<2, 3>{}, Sequence<4>{}, Sequence<5>{}), + make_tuple( + Sequence<0>{}, Sequence<1>{}, Sequence<2, 3>{}, Sequence<4>{}, Sequence<5>{})); + + constexpr auto b_lds_block_desc_unmerged = transform_tensor_descriptor( + b_lds_block_desc_permuted, + make_tuple( + make_pass_through_transform(Number{}), + make_pass_through_transform(Number{}), + make_unmerge_transform(make_tuple(Number{}, Number{})), + make_unmerge_transform(make_tuple(Number{}, Number{})), + make_pass_through_transform(Number{}), + make_pass_through_transform(BK1Number)), + make_tuple(Sequence<0>{}, + Sequence<1>{}, + Sequence<2>{}, + Sequence<3>{}, + Sequence<4>{}, + Sequence<5>{}), + make_tuple(Sequence<1>{}, + Sequence<2>{}, + Sequence<0, 3>{}, + Sequence<4, 5>{}, + Sequence<6>{}, + Sequence<7>{})); + + constexpr auto b_lds_block_desc_bk0_n_bk1 = transform_tensor_descriptor( + b_lds_block_desc_unmerged, + make_tuple(make_merge_transform_v3_division_mod( + make_tuple(Number{}, + Number{}, + Number{}, + Number{})), + make_merge_transform_v3_division_mod( + make_tuple(Number{}, Number{}, Number{})), + make_pass_through_transform(BK1Number)), + make_tuple(Sequence<0, 1, 4, 2>{}, Sequence<5, 6, 3>{}, Sequence<7>{}), + make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{})); + + return b_lds_block_desc_bk0_n_bk1; + } + } + + __device__ static constexpr auto GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock() + { + constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl); + constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl); + + constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock = + make_naive_tensor_descriptor_packed( + make_tuple(I1, + Number{}, + I1, + Number{})); + + return c_shuffle_block_desc_mblock_mperblock_nblock_nperblock; + } + + using BlockwiseGemmPipe = + remove_cvref_t())>; + + __device__ static constexpr index_t GetSharedMemoryNumberOfByte() + { + // LDS allocation for A and B: be careful of alignment + constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1(); + constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1(); + + // lds max alignment + constexpr auto max_lds_align = math::lcm(AK1Number, BK1Number); + + constexpr auto a_block_space_size_aligned = math::integer_least_multiple( + a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align); + + constexpr auto b_block_space_size_aligned = math::integer_least_multiple( + b_block_desc_bk0_n_bk1.GetElementSpaceSize(), max_lds_align); + + // LDS allocation for C shuffle in LDS + constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock = + GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(); + + constexpr auto c_block_size = + c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize(); + + return math::max((a_block_space_size_aligned * sizeof(LDSTypeA) + + b_block_space_size_aligned * sizeof(LDSTypeB)), + c_block_size * sizeof(CShuffleDataType)); + } + + // block_id to matrix tile idx (m0, n0) mapping are controlled by {M01, N01} + __host__ static constexpr bool CheckValidity(const Argument& karg) + { + static_assert((MPerBlock % (MPerXdl * MXdlPerWave) == 0) && + (NPerBlock % (NXdlPerWave * NPerXdl)) == 0, + "Invalid tuning param!"); + + if constexpr(!(GemmSpec == tensor_operation::device::GemmSpecialization::MPadding || + GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding || + GemmSpec == tensor_operation::device::GemmSpecialization::MKPadding || + GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding)) + { + if(!(karg.M % MPerBlock == 0)) + { +#if DEBUG_LOG + std::cout << "Arg M value is not a multiple of MPerBlock! M: " << karg.M << " " + << __FILE__ << ":" << __LINE__ << ", in function: " << __func__ + << std::endl; + +#endif // DEBUG_LOG + return false; + } + } + + if constexpr(!(GemmSpec == tensor_operation::device::GemmSpecialization::NPadding || + GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding || + GemmSpec == tensor_operation::device::GemmSpecialization::NKPadding || + GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding)) + { + if(!(karg.N % NPerBlock == 0)) + { +#if DEBUG_LOG + std::cout << "Arg N value is not a multiple of NPerBlock! N: " << karg.N << " " + << __FILE__ << ":" << __LINE__ << ", in function: " << __func__ + << std::endl; + +#endif // DEBUG_LOG + return false; + } + } + + if constexpr(!(GemmSpec == tensor_operation::device::GemmSpecialization::KPadding || + GemmSpec == tensor_operation::device::GemmSpecialization::MKPadding || + GemmSpec == tensor_operation::device::GemmSpecialization::NKPadding || + GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding)) + { + + auto K_t = karg.KBatch * KPerBlock; + if(!(karg.K % K_t == 0)) + { +#if DEBUG_LOG + std::cout << "Arg K value is not a multiple of K_Batch * K0PerBlock * K1! K: " + << karg.K << " " << __FILE__ << ":" << __LINE__ + << ", in function: " << __func__ << std::endl; + +#endif // DEBUG_LOG + return false; + } + } + else + { + constexpr auto KReadVec = math::lcm(AK1Number, BK1Number); + auto K_t = karg.KBatch * KReadVec; + auto KReadPadSplited = math::integer_divide_ceil(karg.K, K_t) * KReadVec; + if((KReadPadSplited * (karg.KBatch - 1)) >= karg.K) + { + return false; + } + } + + if constexpr(is_same::value) + { + if(karg.K % ABlockTransferSrcScalarPerVector != 0) + { +#if DEBUG_LOG + std::cout << "Arg K (" << karg.K + << ") value is not a multiple of ABlockTransferSrcScalarPerVector (" + << ABlockTransferSrcScalarPerVector << " )! " << __FILE__ << ":" + << __LINE__ << ", in function: " << __func__ << std::endl; + +#endif // DEBUG_LOG + return false; + } + } + else + { + if(karg.M % ABlockTransferSrcScalarPerVector != 0) + { +#if DEBUG_LOG + std::cout << "Arg M (" << karg.M + << ") value is not a multiple of ABlockTransferSrcScalarPerVector (" + << ABlockTransferSrcScalarPerVector << " )! " << __FILE__ << ":" + << __LINE__ << ", in function: " << __func__ << std::endl; + +#endif // DEBUG_LOG + return false; + } + } + + if constexpr(is_same::value) + { + if(karg.N % BBlockTransferSrcScalarPerVector != 0) + { +#if DEBUG_LOG + std::cout << "Arg N (" << karg.N + << ") value is not a multiple of BBlockTransferSrcScalarPerVector (" + << BBlockTransferSrcScalarPerVector << " )! " << __FILE__ << ":" + << __LINE__ << ", in function: " << __func__ << std::endl; + +#endif // DEBUG_LOG + return false; + } + } + else + { + if(karg.K % BBlockTransferSrcScalarPerVector != 0) + { +#if DEBUG_LOG + std::cout << "Arg K (" << karg.K + << ") value is not a multiple of BBlockTransferSrcScalarPerVector (" + << BBlockTransferSrcScalarPerVector << " )! " << __FILE__ << ":" + << __LINE__ << ", in function: " << __func__ << std::endl; + +#endif // DEBUG_LOG + return false; + } + } + + if constexpr(is_same::value) + { + if(karg.N % CShuffleBlockTransferScalarPerVector_NPerBlock != 0) + { +#if DEBUG_LOG + std::cout << "Arg N (" << karg.N + << ") value is not a multiple of " + "CShuffleBlockTransferScalarPerVector_NPerBlock (" + << CShuffleBlockTransferScalarPerVector_NPerBlock << " )! " << __FILE__ + << ":" << __LINE__ << ", in function: " << __func__ << std::endl; + +#endif // DEBUG_LOG + return false; + } + } + else + { + if(karg.M % CShuffleBlockTransferScalarPerVector_NPerBlock != 0) + { +#if DEBUG_LOG + std::cout << "Arg M (" << karg.M + << ") value is not a multiple of " + "CShuffleBlockTransferScalarPerVector_NPerBlock (" + << CShuffleBlockTransferScalarPerVector_NPerBlock << " )! " << __FILE__ + << ":" << __LINE__ << ", in function: " << __func__ << std::endl; + +#endif // DEBUG_LOG + return false; + } + } + + // check gridwise gemm pipeline + const auto num_k_loop = karg.AK0 / (KPerBlock / AK1Value); + + if constexpr(BlkGemmPipelineVer != BlockGemmPipelineVersion::v1) + { + if(num_k_loop <= BlockwiseGemmPipe::PrefetchStages) + { + return false; + } + } + + // TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc) + return true; + } + + __host__ static constexpr bool CalculateHasMainKBlockLoop(index_t K) + { + const index_t num_loop = K / KPerBlock; + + return BlockwiseGemmPipe::BlockHasHotloop(num_loop); + } + + __host__ static constexpr TailNumber CalculateKBlockLoopTailNum(index_t K) + { + const index_t num_loop = K / KPerBlock; + + return BlockwiseGemmPipe::BlockLoopTailNum(num_loop); + } + + template + __device__ static constexpr auto MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock( + const CGridDesc& c_grid_desc_m_n, index_t MBlock, index_t NBlock) + { + 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{})), + make_unmerge_transform(make_tuple(NBlock, Number{}))), + 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 + // if arch = gfx942 + using Block2CTileMap = BlockToCTileMap_Grouped_M00_N0_M01Adapt<8, MPerBlock, NPerBlock>; + // using Block2CTileMap = BlockToCTileMap_3DGrid_KSplit; + + template + __device__ static void Run(const ADataType* p_a_grid, + const BDataType* p_b_grid, + DsGridPointer& p_ds_grid, + CDataType* p_c_grid, + void* p_shared, + const Problem& problem, + AElementwiseOperation a_element_op, + BElementwiseOperation b_element_op, + CElementwiseOperation c_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_bk0_n_bk1 = MakeBGridDescriptor_BK0_N_BK1( + problem.K, problem.KPadded, problem.N, problem.NPadded, problem.StrideB, problem.BK0); + 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_bk0_n_bk1.GetElementSpaceSize()); + auto c_grid_buf = make_dynamic_buffer( + p_c_grid, c_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize()); + + // divide block work by [M, N] + const auto block_2_ctile_map = Block2CTileMap{problem.M, problem.N, 4}; + + 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); + + const index_t n_block_data_idx_on_grid = + __builtin_amdgcn_readfirstlane(block_n_id * NPerBlock); + + // lds max alignment + constexpr auto max_lds_align = math::lcm(AK1Number, BK1Number); + + // 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 + 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, + BlockwiseGemmPipe::GlobalBufferNum>( + 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{}); + + // B matrix blockwise copy + auto b_blockwise_copy = + ThreadGroupTensorSliceTransfer_v4r1, + BBlockTransferThreadClusterLengths_BK0_N_BK1, + BBlockTransferThreadClusterArrangeOrder, + BDataType, + LDSTypeB, + decltype(b_grid_desc_bk0_n_bk1), + decltype(b_block_desc_bk0_n_bk1), + BBlockTransferSrcAccessOrder, + Sequence<0, 1, 2>, + BBlockTransferSrcVectorDim, + 2, + BBlockTransferSrcScalarPerVector, + BBlockTransferDstScalarPerVector_BK1, + 1, + 1, + BThreadTransferSrcResetCoordinateAfterRun, + true, + BlockwiseGemmPipe::GlobalBufferNum>( + b_grid_desc_bk0_n_bk1, + make_multi_index(0, n_block_data_idx_on_grid, 0), + b_element_op, + b_block_desc_bk0_n_bk1, + make_multi_index(0, 0, 0), + ck::tensor_operation::element_wise::PassThrough{}); + + // LDS allocation for A and B: be careful of alignment + constexpr auto a_block_space_size_aligned = math::integer_least_multiple( + a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align); + + // Cast after lds + auto a_block_buf = make_dynamic_buffer( + static_cast(p_shared), a_block_desc_ak0_m_ak1.GetElementSpaceSize()); + + auto b_block_buf = make_dynamic_buffer( + static_cast(p_shared) + + a_block_space_size_aligned * sizeof(LDSTypeA) / sizeof(LDSTypeB), + b_block_desc_bk0_n_bk1.GetElementSpaceSize()); + + constexpr auto a_block_slice_copy_step = make_multi_index(KPerBlock / AK1Number, 0, 0); + constexpr auto b_block_slice_copy_step = make_multi_index(KPerBlock / BK1Number, 0, 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_buf, + a_block_slice_copy_step, + b_grid_desc_bk0_n_bk1, + b_block_desc_bk0_n_bk1, + b_blockwise_copy, + b_grid_buf, + b_block_buf, + 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); + constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl); + + // 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); + } + }); + } + } + + 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_0, + void* p_shared_1, + const Problem& problem, + AElementwiseOperation a_element_op, + BElementwiseOperation b_element_op, + CElementwiseOperation c_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_bk0_n_bk1 = MakeBGridDescriptor_BK0_N_BK1( + problem.K, problem.KPadded, problem.N, problem.NPadded, problem.StrideB, problem.BK0); + 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_bk0_n_bk1.GetElementSpaceSize()); + auto c_grid_buf = make_dynamic_buffer( + p_c_grid, c_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize()); + + // divide block work by [M, N] + const auto block_2_ctile_map = Block2CTileMap{problem.M, problem.N, 4}; + + 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); + + const index_t n_block_data_idx_on_grid = + __builtin_amdgcn_readfirstlane(block_n_id * NPerBlock); + + // lds max alignment + constexpr auto max_lds_align = math::lcm(AK1Number, BK1Number); + + // 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 + 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, + BlockwiseGemmPipe::GlobalBufferNum>( + 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{}); + + // B matrix blockwise copy + auto b_blockwise_copy = + ThreadGroupTensorSliceTransfer_v4r1, + BBlockTransferThreadClusterLengths_BK0_N_BK1, + BBlockTransferThreadClusterArrangeOrder, + BDataType, + LDSTypeB, + decltype(b_grid_desc_bk0_n_bk1), + decltype(b_block_desc_bk0_n_bk1), + BBlockTransferSrcAccessOrder, + Sequence<0, 1, 2>, + BBlockTransferSrcVectorDim, + 2, + BBlockTransferSrcScalarPerVector, + BBlockTransferDstScalarPerVector_BK1, + 1, + 1, + BThreadTransferSrcResetCoordinateAfterRun, + true, + BlockwiseGemmPipe::GlobalBufferNum>( + b_grid_desc_bk0_n_bk1, + make_multi_index(0, n_block_data_idx_on_grid, 0), + b_element_op, + b_block_desc_bk0_n_bk1, + make_multi_index(0, 0, 0), + ck::tensor_operation::element_wise::PassThrough{}); + + // LDS allocation for A and B: be careful of alignment + constexpr auto a_block_space_size_aligned = math::integer_least_multiple( + a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align); + + auto a_block_buf_ping = make_dynamic_buffer( + static_cast(p_shared_0), a_block_desc_ak0_m_ak1.GetElementSpaceSize()); + + auto b_block_buf_ping = make_dynamic_buffer( + static_cast(p_shared_0) + + a_block_space_size_aligned * sizeof(LDSTypeA) / sizeof(LDSTypeB), + b_block_desc_bk0_n_bk1.GetElementSpaceSize()); + + auto a_block_buf_pong = make_dynamic_buffer( + static_cast(p_shared_1), a_block_desc_ak0_m_ak1.GetElementSpaceSize()); + + auto b_block_buf_pong = make_dynamic_buffer( + static_cast(p_shared_1) + + a_block_space_size_aligned * sizeof(LDSTypeA) / sizeof(LDSTypeB), + b_block_desc_bk0_n_bk1.GetElementSpaceSize()); + + auto a_block_bufs = make_tuple(a_block_buf_ping, a_block_buf_pong); + auto b_block_bufs = make_tuple(b_block_buf_ping, b_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(KPerBlock / BK1Number, 0, 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_bk0_n_bk1, + b_block_desc_bk0_n_bk1, + 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); + constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl); + + // 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_0), + 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 diff --git a/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v7r3.hpp b/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v7r3.hpp new file mode 100644 index 0000000000..ea074144b6 --- /dev/null +++ b/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v7r3.hpp @@ -0,0 +1,648 @@ +// 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/tensor_descriptor.hpp" +#include "ck/tensor_description/tensor_descriptor_helper.hpp" +#include "ck/tensor_description/tensor_space_filling_curve.hpp" +#include "ck/utility/is_detected.hpp" +#include "ck/tensor/static_tensor.hpp" + +#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_util.hpp" + +namespace ck { +// Thread-level multi-source, multi-destination tensor slice data movement +// Assume: +// 1. All sources and destinations are DynamicBuffer +// 2. Same VectorDim and ScalerPerVector for all sources and destinations +// 3. DstInMemOps are per destination tensor +// 4. ThreadTransferSrcResetCoordinateAfterRunFlags are per source tensor +// 5. ThreadTransferDstResetCoordinateAfterRunFlags are per destination tensor +// 6. Does not need to know src_descs and dst_descs at compile-time +// 7. Does not need to know src_slice_origins and dst_slice_origins at compile-time, +// +// Does following things to avoid scratch memory issue +// 1. Use StaticallyIndexedArray or vector_type instead of C array for thread buffer +// 2. Pass tensor descritpors by reference (or tuple of references) +// 3. Does not keep reference to tensor descriptor +// 4. Does not construct new tensor coordinate when call Run() +template + typename SliceLengths, + typename SrcDimAccessOrder, + typename DstDimAccessOrder, + index_t SrcVectorDim, + index_t DstVectorDim, + typename SrcScalarPerVectors, + index_t DstScalarPerVector, + typename SrcResetCoordinateAfterRunFlags, // Sequence + typename DstResetCoordinateAfterRunFlags, // Sequence + index_t NumThreadScratch = 1> +struct ThreadwiseTensorSliceTransfer_v7r3 +{ + static constexpr auto I0 = Number<0>{}; + + static constexpr auto SrcScalarPerVector = SrcScalarPerVectors{}[I0]; + + static constexpr index_t nDim = SliceLengths::Size(); + + static constexpr index_t nSrc = SrcDescs::Size(); + static constexpr index_t nDst = DstDescs::Size(); + + using Index = MultiIndex; + + // return a tuple of coordiantes for a tuple of tensor + template = false> + static constexpr auto MakeCoordinates(const Descs& descs, const Indices& indices) + { + return generate_tuple([&](auto i) { return make_tensor_coordinate(descs[i], indices[i]); }, + Number{}); + } + + using SrcCoords = decltype(MakeCoordinates(SrcDescs{}, StaticallyIndexedArray{})); + using DstCoords = decltype(MakeCoordinates(DstDescs{}, StaticallyIndexedArray{})); + + // scalar per access on each dim + // FIXME: don't use lambda_scalar_per_access + static constexpr auto src_scalar_per_access = generate_sequence( + detail::lambda_scalar_per_access{}, Number{}); + + static constexpr auto dst_scalar_per_access = generate_sequence( + detail::lambda_scalar_per_access{}, Number{}); + + using SrcSpaceFillingCurve = SpaceFillingCurve, + false>; + + using DstSpaceFillingCurve = SpaceFillingCurve, + false>; + + __device__ constexpr ThreadwiseTensorSliceTransfer_v7r3( + const SrcDescs& src_descs, + const StaticallyIndexedArray& src_slice_origins, + const DstDescs& dst_descs, + const StaticallyIndexedArray& dst_slice_origins, + const ElementwiseOperation& element_op) + : src_coords_(MakeCoordinates(src_descs, src_slice_origins)), + dst_coords_(MakeCoordinates(dst_descs, dst_slice_origins)), + element_op_(element_op) + { + static_assert(SliceLengths::At(Number{}) % SrcScalarPerVector == 0, + "wrong! cannot evenly divide"); + + static_assert(SliceLengths::At(Number{}) % DstScalarPerVector == 0, + "wrong! cannot evenly divide"); + } + + template = false> + __device__ void SetSrcSliceOrigins(const SrcDescs& src_descs, + const Indices& src_slice_origin_idxs) + { + static_for<0, nSrc, 1>{}([&](auto i) { + src_coords_(i) = make_tensor_coordinate(src_descs[i], src_slice_origin_idxs[i]); + }); + } + + template = false> + __device__ void SetDstSliceOrigins(const DstDescs& dst_descs, + const Indices& dst_slice_origin_idxs) + { + static_for<0, nDst, 1>{}([&](auto i) { + dst_coords_(i) = make_tensor_coordinate(dst_descs[i], dst_slice_origin_idxs[i]); + }); + } + + template + __device__ static auto generate_vectors() + { + auto data_types = DataTypes{}; + + constexpr index_t num = data_types.Size(); + + return generate_tuple( + [&](auto i) { + using DataType = remove_cvref_t; + + return vector_type_maker_t{}; + }, + Number{}); + } + + // SrcDescs: Tuple + // SrcBuffers: Tuple + template = false> + __device__ void RunRead(const SrcDescs& src_descs, + const SrcBuffers& src_bufs, + Number thread_scratch_id = Number{}) + { + // loop over space-filling curve + static_for<0, src_num_access, 1>{}([&](auto iAccess) { + auto src_vectors = generate_vectors(); + auto elm_vectors = generate_vectors(); + + bool oob_val = true; + + // copy data from src_bufs into src_vectors + static_for<0, nSrc, 1>{}([&](auto i) { + using src_vector_t = typename remove_cvref_t::type; + + const bool is_src_valid = + coordinate_has_valid_offset_assuming_visible_index_is_valid(src_descs[i], + src_coords_[i]); + + oob_val = oob_val & is_src_valid; + + if constexpr(SrcScalarPerVectors{}[i] == 1) + { + auto data_types = SrcDatas{}; + using DataType = remove_cvref_t; + const auto tmp = + src_bufs[i].template Get(src_coords_[i].GetOffset(), true); + + static_for<0, SrcScalarPerVector, 1>{}( + [&](auto j) { src_vectors(i).template AsType()(j) = tmp; }); + } + else + { + src_vectors(i).template AsType()(I0) = + src_bufs[i].template Get(src_coords_[i].GetOffset(), true); + } + }); + + constexpr auto get_elem_op_vec_len = []() { + if constexpr(is_detected::value) + { + if constexpr(decltype(element_op_)::is_pack8_invocable) + return math::min(8, SrcScalarPerVector); + } + if constexpr(is_detected::value) + { + if constexpr(decltype(element_op_)::is_pack4_invocable) + return math::min(4, SrcScalarPerVector); + } + if constexpr(is_detected::value) + { + if constexpr(decltype(element_op_)::is_pack2_invocable) + return math::min(2, SrcScalarPerVector); + } + return 1; + }; + + constexpr index_t elem_op_vec_len = get_elem_op_vec_len(); + + // apply pointwise function + static_for<0, SrcScalarPerVector / elem_op_vec_len, 1>{}([&](auto i) { + // get reference to src data + const auto src_data_refs = generate_tie( + // return type should be lvalue + [&](auto iSrc) -> const auto& { + using SrcData = remove_cvref_t>; + + using elem_op_vec_t = typename vector_type::type; + + return src_vectors[iSrc].template AsType()[i]; + }, + Number{}); + + // get reference to dst data + auto dst_data_refs = generate_tie( + // return type should be lvalue + [&](auto iDst) -> auto& { + using DstData = remove_cvref_t>; + + using elem_op_vec_t = typename vector_type::type; + + return elm_vectors(iDst).template AsType()(i); + }, + Number{}); + + // apply pointwise function + // pointwise function signature: + // element_op_(dst_data_refs[I0], + // dst_data_refs[I1], + // ..., + // src_data_refs[I0], + // src_data_refs[I1], + // ...) + unpack2(element_op_, dst_data_refs, src_data_refs); + }); + + elm_vectors_tuple_(thread_scratch_id)(iAccess) = elm_vectors; + oob_vectors_tuple_(thread_scratch_id)(iAccess) = oob_val; + + // move coordinate + if constexpr(iAccess.value != src_num_access - 1) + { + constexpr auto forward_step = SrcSpaceFillingCurve::GetForwardStep(iAccess); + + static_for<0, nSrc, 1>{}([&](auto i) { + move_tensor_coordinate(src_descs[i], + src_coords_(i), + make_tensor_coordinate_step(src_descs[i], forward_step)); + }); + } + }); + + // move coordinate back to slice origin (or not) + static_for<0, nSrc, 1>{}([&](auto i) { + if constexpr(SrcResetCoordinateAfterRunFlags::At(i)) + { + const auto src_reset_step = + make_tensor_coordinate_step(src_descs[i], GetSrcCoordinateResetStep()); + + move_tensor_coordinate(src_descs[i], src_coords_(i), src_reset_step); + } + }); + } + +#if 1 + template + __device__ void OOBCheck(Number thread_scratch_id = Number{}) + { + // loop over space-filling curve + static_for<0, src_num_access, 1>{}([&](auto iAccess) { + auto elm_vectors = elm_vectors_tuple_[thread_scratch_id][iAccess]; + auto oob_val = oob_vectors_tuple_[thread_scratch_id][iAccess]; + + static_for<0, nDst, 1>{}([&](auto i) { + using elm_vector_t = typename remove_cvref_t::type; + elm_vectors(i).template AsType()(I0) = + oob_val ? elm_vectors(i).template AsType()[I0] : elm_vector_t{0}; + }); + + elm_vectors_tuple_(thread_scratch_id)(iAccess) = elm_vectors; + }); + } +#endif + + template + __device__ void + TransposeFromElmToDst(Number thread_scratch_id = Number{}) + { + using DstData = remove_cvref_t; + + using ElmThreadScratch = + StaticTensorTupleOfVectorBuffer; + using DstThreadScratch = + StaticTensorTupleOfVectorBuffer; + + ElmThreadScratch elm_thread_scratch_; + DstThreadScratch dst_thread_scratch_; + + elm_thread_scratch_.data_ = + bit_cast(elm_vectors_tuple_[thread_scratch_id]); + + if constexpr(SrcVectorDim != DstVectorDim && + ((is_same>::value && + SrcScalarPerVector % 2 == 0 && DstScalarPerVector % 2 == 0) || + (is_same>::value && + SrcScalarPerVector % 4 == 0 && DstScalarPerVector % 4 == 0) || + (is_same>::value && + SrcScalarPerVector % 4 == 0 && DstScalarPerVector % 4 == 0))) + { + // each transpose does + // DstScalarPerVector # of src vectors in src_thread_scratch_ + // SrcScalarPerVector # of dst vectors in dst_thread_scratch_ + constexpr index_t num_src_vector = Number{}; + constexpr index_t num_dst_vector = Number{}; + + // Assume SrcVectorDim is not the same as DstVectorDim, so we do transpose + // TODO: make this logic generic for all scenario + + constexpr auto src_scalar_step_in_vector = generate_sequence( + detail::lambda_scalar_step_in_vector{}, Number{}); + + constexpr auto dst_scalar_step_in_vector = generate_sequence( + detail::lambda_scalar_step_in_vector{}, Number{}); + + constexpr auto scalar_per_access = generate_sequence( + detail::lambda_scalar_per_access_for_src_and_dst{}, + Number{}); + + constexpr auto access_lengths = SliceLengths{} / scalar_per_access; + + static_ford{}([&](auto access_idx) { + constexpr auto data_idx = access_idx * scalar_per_access; + + constexpr auto data_idx_seq = generate_sequence_v2( + [&](auto i) { return Number{}; }, Number{}); + + using src_vector_t = vector_type_maker_t; + using dst_vector_t = vector_type_maker_t; + + // get DstScalarPerVector # of read-only references to src vectors from + // src_thread_scratch_ + const auto src_vector_refs = generate_tie( + [&](auto i) -> const src_vector_t& { + // i increment corresponds to movement in DstVectorDim + return elm_thread_scratch_.GetVectorTypeReference( + data_idx_seq + i * dst_scalar_step_in_vector); + }, + Number{}); + + // get SrcScalarPerVector # of references to dst vectors from + // dst_thread_scratch_ + auto dst_vector_refs = generate_tie( + [&](auto i) -> dst_vector_t& { + // i increment corresponds to movement in SrcVectorDim + return dst_thread_scratch_.GetVectorTypeReference( + data_idx_seq + i * src_scalar_step_in_vector); + }, + Number{}); + + // do data transpose + transpose_vectors{}( + src_vector_refs, dst_vector_refs); + }); + } + else + { + static_ford{}( + [&](auto idx) { dst_thread_scratch_(idx) = elm_thread_scratch_[idx]; }); + } + + dst_vectors_tuple_(thread_scratch_id) = bit_cast(dst_thread_scratch_.data_); + } + + // DstDescs: Tuple + // DstBuffers: Tuple + template = false> + __device__ void RunWrite(const DstDescs& dst_descs, + DstBuffers dst_bufs, + Number thread_scratch_id = Number{}) + { + OOBCheck(thread_scratch_id); + TransposeFromElmToDst(thread_scratch_id); + + // loop over space-filling curve + static_for<0, dst_num_access, 1>{}([&](auto iAccess) { + auto dst_vectors = dst_vectors_tuple_[thread_scratch_id][iAccess]; + + // copy data from buf_vectors into dst_bufs + static_for<0, nDst, 1>{}([&](auto i) { + using dst_vector_t = typename remove_cvref_t::type; + + const bool is_dst_valid = + coordinate_has_valid_offset_assuming_visible_index_is_valid(dst_descs[i], + dst_coords_[i]); + + constexpr InMemoryDataOperationEnum DstInMemOp = + static_cast(DstInMemOps::At(i.value)); + + dst_bufs(i).template Update( + dst_coords_[i].GetOffset(), + is_dst_valid, + dst_vectors[i].template AsType()[I0]); + }); + + // move coordinate + if constexpr(iAccess.value != dst_num_access - 1) + { + constexpr auto forward_step = DstSpaceFillingCurve::GetForwardStep(iAccess); + + static_for<0, nDst, 1>{}([&](auto i) { + move_tensor_coordinate(dst_descs[i], + dst_coords_(i), + make_tensor_coordinate_step(dst_descs[i], forward_step)); + }); + } + }); + + static_for<0, nDst, 1>{}([&](auto i) { + if constexpr(DstResetCoordinateAfterRunFlags::At(i)) + { + const auto dst_reset_step = + make_tensor_coordinate_step(dst_descs[i], GetDstCoordinateResetStep()); + + move_tensor_coordinate(dst_descs[i], dst_coords_(i), dst_reset_step); + } + }); + } + + // SrcDescs: Tuple + // SrcBuffers: Tuple + // DstDescs: Tuple + // DstBuffers: Tuple + template = false> + __device__ void Run(const SrcDescs& src_descs, + const SrcBuffers& src_bufs, + const DstDescs& dst_descs, + DstBuffers dst_bufs) + { + RunRead(src_descs, src_bufs); + RunWrite(dst_descs, dst_bufs); + } + + __device__ static constexpr auto GetSrcCoordinateResetStep() + { + if constexpr(src_num_access == 0) + { + return typename SrcSpaceFillingCurve::Index{}; + } + else + { + return SrcSpaceFillingCurve::GetStepBetween(Number{}, Number<0>{}); + } + } + + __device__ static constexpr auto GetDstCoordinateResetStep() + { + if constexpr(dst_num_access == 0) + { + return typename DstSpaceFillingCurve::Index{}; + } + else + { + return DstSpaceFillingCurve::GetStepBetween(Number{}, Number<0>{}); + } + } + + __device__ static constexpr auto GetSrcThreadScratchDescriptor() + { + // constexpr auto src_scalar_per_access = generate_sequence( + // detail::lambda_scalar_per_access{}, + // Number{}); + + constexpr auto src_access_lengths = SliceLengths{} / src_scalar_per_access; + + constexpr auto src_access_lengths_and_vector_length = container_push_back( + sequence_to_tuple_of_number(src_access_lengths), Number{}); + + // 1st stage of transforms + constexpr auto desc0 = + make_naive_tensor_descriptor_packed(src_access_lengths_and_vector_length); + + // 2nd stage of transforms + constexpr auto transforms = generate_tuple( + [&](auto i) { + if constexpr(i == SrcVectorDim) + { + return make_merge_transform_v3_division_mod( + make_tuple(src_access_lengths_and_vector_length[i], + src_access_lengths_and_vector_length[Number{}])); + } + else + { + return make_pass_through_transform(src_access_lengths_and_vector_length[i]); + } + }, + Number{}); + + constexpr auto low_dim_idss = generate_tuple( + [&](auto i) { + if constexpr(i == SrcVectorDim) + { + return Sequence{}; + } + else + { + return Sequence{}; + } + }, + Number{}); + + constexpr auto up_dim_idss = + generate_tuple([&](auto i) { return Sequence{}; }, Number{}); + + return transform_tensor_descriptor(desc0, transforms, low_dim_idss, up_dim_idss); + } + + __device__ static constexpr auto GetDstThreadScratchDescriptor() + { + // 1st stage of transforms + // constexpr auto dst_scalar_per_access = generate_sequence( + // detail::lambda_scalar_per_access{}, + // Number{}); + + constexpr auto dst_access_lengths = SliceLengths{} / dst_scalar_per_access; + + constexpr auto dst_access_lengths_and_vector_length = container_push_back( + sequence_to_tuple_of_number(dst_access_lengths), Number{}); + + constexpr auto desc0 = + make_naive_tensor_descriptor_packed(dst_access_lengths_and_vector_length); + + // 2nd stage of transforms + constexpr auto transforms = generate_tuple( + [&](auto i) { + if constexpr(i == DstVectorDim) + { + return make_merge_transform_v3_division_mod( + make_tuple(dst_access_lengths_and_vector_length[i], + dst_access_lengths_and_vector_length[Number{}])); + } + else + { + return make_pass_through_transform(dst_access_lengths_and_vector_length[i]); + } + }, + Number{}); + + constexpr auto low_dim_idss = generate_tuple( + [&](auto i) { + if constexpr(i == DstVectorDim) + { + return Sequence{}; + } + else + { + return Sequence{}; + } + }, + Number{}); + + constexpr auto up_dim_idss = + generate_tuple([&](auto i) { return Sequence{}; }, Number{}); + + return transform_tensor_descriptor(desc0, transforms, low_dim_idss, up_dim_idss); + } + + // src_slice_origin_step_idx need to be known at compile-time, for performance reason + template + __device__ void MoveSrcSliceWindow(const SrcDescs& src_descs, + Number iSrc, + const Index& src_slice_origin_step_idx) + { + // if src coord was not reset by RunRead(), then need to adjust the step here + const auto adjusted_step_idx = + SrcResetCoordinateAfterRunFlags::At(iSrc) + ? src_slice_origin_step_idx + : src_slice_origin_step_idx + GetSrcCoordinateResetStep(); + + // is it OK to construct a new step every time? + const auto adjusted_step = make_tensor_coordinate_step(src_descs[iSrc], adjusted_step_idx); + + move_tensor_coordinate(src_descs[iSrc], src_coords_(iSrc), adjusted_step); + } + + // dst_slice_origin_step_idx need to be known at compile-time, for performance reason + template + __device__ void MoveDstSliceWindow(const DstDescs& dst_descs, + Number iDst, + const Index& dst_slice_origin_step_idx) + { + // if dst coord was not reset by Run(), then need to adjust the step here + const auto adjusted_step_idx = + DstResetCoordinateAfterRunFlags::At(iDst) + ? dst_slice_origin_step_idx + : dst_slice_origin_step_idx + GetDstCoordinateResetStep(); + + // is it OK to construct a new step every time? + const auto adjusted_step = make_tensor_coordinate_step(dst_descs[iDst], adjusted_step_idx); + + move_tensor_coordinate(dst_descs[iDst], dst_coords_(iDst), adjusted_step); + } + + private: + using SrcVectorsType = decltype(generate_vectors()); + using ElmVectorsType = decltype(generate_vectors()); + using DstVectorsType = decltype(generate_vectors()); + + static constexpr auto src_num_access = SrcSpaceFillingCurve::GetNumOfAccess(); + static constexpr auto dst_num_access = DstSpaceFillingCurve::GetNumOfAccess(); + + using ElmVectorTuple = StaticallyIndexedArray; + using DstVectorTuple = StaticallyIndexedArray; + + StaticallyIndexedArray elm_vectors_tuple_; + StaticallyIndexedArray dst_vectors_tuple_; + + using OOBVectorTuple = StaticallyIndexedArray; + StaticallyIndexedArray oob_vectors_tuple_; + + SrcCoords src_coords_; + DstCoords dst_coords_; + const ElementwiseOperation element_op_; +}; + +} // namespace ck