From 8fe0377e86c83111f8be570b14252dfdab954f14 Mon Sep 17 00:00:00 2001 From: rocking5566 Date: Fri, 20 May 2022 10:56:56 +0800 Subject: [PATCH] Gemm reduce max (#209) * [What] Rename the example [Why] Prepare to add unary reduction * Add global oparation to the parameter * Add atomicmax * Fix compile error * Support atomicMax (hip library) * Rename the reduction example * Fix target name * use p_d1_grid as the indicator directly * Prevent performance issue. Let passthrough handle it. * Implement the function template the specialize the float2 * No need to separate into two lines * Remove empty line * add comment * Fix compile error due to merge from develop * make the implementation of atomic_max / atomic_add explicit for each datatype * Refine typo * For future CI test * Fix compiler error in ckProfiler * Merge commit 'de2769e3a6695b38a20529261273ddc5cdaab2fe' * simply use remove_pointer * Rename type and var * Refine example * Modify reducemax example * Fix bug in reduction * Change initialize range * Implement F64 version of atomicMax * Move reduction code together * Add buffer atomic_max * Fix coding style by clang-format * Integrate new api of DeviceGemmReduce_Xdl_CShuffle * Integrate Batch gemm reduction * Fix example * fix example * clean up * Fix batch gemm tensor operation * Fix coding style * Fix template augument * Fix clang format * Keep flexible of different stride for each D tensor * Fix compile error for ckProfiler * Fix typo * [What] Fix naming [Why] Prepare to add out elementop * Add DoutElementOp Co-authored-by: Chao Liu Co-authored-by: rocking [ROCm/composable_kernel commit: 0ffe956ab1c1a8e128c2d6e419de68fcc1a8b5ff] --- example/16_gemm_reduce/CMakeLists.txt | 3 +- .../gemm_reduce_xdl_max_fp16.cpp | 249 +++++++++++++++++ ...=> gemm_reduce_xdl_sum_squaresum_fp16.cpp} | 88 +++--- .../batched_gemm_reduce_xdl_fp16.cpp | 89 ++++--- example/CMakeLists.txt | 2 +- include/ck/config.hpp | 23 +- ...evice_batched_gemm_reduce_xdl_cshuffle.hpp | 144 +++++----- .../gpu/device/device_gemm_reduce.hpp | 51 ++-- .../device_gemm_reduce_xdl_cshuffle.hpp | 89 ++++--- .../gridwise_gemm_reduce_xdl_cshuffle_v1.hpp | 252 +++++++++--------- include/ck/utility/amd_buffer_addressing.hpp | 108 ++++++++ include/ck/utility/common_header.hpp | 2 +- include/ck/utility/dynamic_buffer.hpp | 42 ++- .../utility/generic_memory_space_atomic.hpp | 97 +++++++ .../generic_memory_space_atomic_add.hpp | 44 --- include/ck/utility/type.hpp | 3 + .../include/ck/library/host_tensor/device.hpp | 30 ++- .../gpu/batched_gemm_reduce/CMakeLists.txt | 2 +- ...6_f16_f16_f32_f32_gkm_gkn_gmn_instance.cpp | 64 +++-- ...6_f16_f16_f32_f32_gkm_gnk_gmn_instance.cpp | 64 +++-- ...6_f16_f16_f32_f32_gmk_gkn_gmn_instance.cpp | 64 +++-- ...6_f16_f16_f32_f32_gmk_gnk_gmn_instance.cpp | 58 ++-- ..._f16_f16_f16_f32_f32_km_kn_mn_instance.cpp | 64 +++-- ..._f16_f16_f16_f32_f32_km_nk_mn_instance.cpp | 64 +++-- ..._f16_f16_f16_f32_f32_mk_kn_mn_instance.cpp | 64 +++-- ..._f16_f16_f16_f32_f32_mk_nk_mn_instance.cpp | 58 ++-- .../profile_batched_gemm_reduce_impl.hpp | 53 ++-- profiler/include/profile_gemm_reduce_impl.hpp | 53 ++-- 28 files changed, 1298 insertions(+), 626 deletions(-) create mode 100644 example/16_gemm_reduce/gemm_reduce_xdl_max_fp16.cpp rename example/16_gemm_reduce/{gemm_reduce_xdl_fp16.cpp => gemm_reduce_xdl_sum_squaresum_fp16.cpp} (61%) create mode 100644 include/ck/utility/generic_memory_space_atomic.hpp delete mode 100644 include/ck/utility/generic_memory_space_atomic_add.hpp diff --git a/example/16_gemm_reduce/CMakeLists.txt b/example/16_gemm_reduce/CMakeLists.txt index 08d37b34a6..5441247a56 100644 --- a/example/16_gemm_reduce/CMakeLists.txt +++ b/example/16_gemm_reduce/CMakeLists.txt @@ -1 +1,2 @@ -add_example_executable(example_gemm_reduce_xdl_fp16 gemm_reduce_xdl_fp16.cpp) +add_example_executable(example_gemm_reduce_xdl_max_fp16 gemm_reduce_xdl_max_fp16.cpp) +add_example_executable(example_gemm_reduce_xdl_sum_squaresum_fp16 gemm_reduce_xdl_sum_squaresum_fp16.cpp) diff --git a/example/16_gemm_reduce/gemm_reduce_xdl_max_fp16.cpp b/example/16_gemm_reduce/gemm_reduce_xdl_max_fp16.cpp new file mode 100644 index 0000000000..4d837c4675 --- /dev/null +++ b/example/16_gemm_reduce/gemm_reduce_xdl_max_fp16.cpp @@ -0,0 +1,249 @@ +#include +#include +#include +#include +#include + +#include "check_err.hpp" +#include "config.hpp" +#include "device.hpp" +#include "host_tensor.hpp" +#include "host_tensor_generator.hpp" +#include "device_tensor.hpp" +#include "device_gemm_reduce_xdl_cshuffle.hpp" +#include "element_wise_operation.hpp" +#include "reference_gemm.hpp" +#include "gemm_specialization.hpp" +#include "element_wise_reduce_operation.hpp" + +template +using S = ck::Sequence; + +using F16 = ck::half_t; +using F32 = float; +using F64 = double; + +using Row = ck::tensor_layout::gemm::RowMajor; +using Col = ck::tensor_layout::gemm::ColumnMajor; + +using ADataType = F16; +using BDataType = F16; +using CDataType = F16; +using ReduceAccDataType = F32; +using DDataType = F64; +using DPtrsGlobal = ck::Tuple; + +using ALayout = ck::tensor_layout::gemm::RowMajor; +using BLayout = ck::tensor_layout::gemm::ColumnMajor; +using CLayout = ck::tensor_layout::gemm::RowMajor; + +using AElementOp = ck::tensor_operation::element_wise::PassThrough; +using BElementOp = ck::tensor_operation::element_wise::PassThrough; +using CElementOp = ck::tensor_operation::element_wise::PassThrough; +using DsReduceOp = ck::Tuple>; +using DsElementOp = ck::Tuple< + ck::tensor_operation::element_wise::UnaryIdentic>; +using DGlobalMemOp = + ck::InMemoryDataOperationEnumSequence; + +static constexpr auto GemmSpecialization = + ck::tensor_operation::device::GemmSpecialization::Default; + +// clang-format off +using DeviceGemmReduceInstance = ck::tensor_operation::device::DeviceGemmReduce_Xdl_CShuffle +//######| ALayout| BLayout| CLayout|AData| BData| CData| GemmAcc| CShuffle| ReduceAcc| DData| A| B| C| Dxs| DxsInEleOp| DxsOutEleOp| D| GEMM| NumGemmK| 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| CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy| +//######| | | | Type| Type| Type| DataType| DataType| DataType| Type Tuple| Elementwise| Elementwise| Elementwise| Reduce| | | MemoryData| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths| SrcDstScalarPerVector| SrcDstScalarPerVector| +//######| | | | | | | | | | | Operation| Operation| Operation| Operation| | | Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NPerBlock| _NPerBlock| _MPerBlock_NPerBlock| _NPerBlock| _MPerBlock| +//######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | + < Row, Col, Row, F16, F16, F16, F32, F32, ReduceAccDataType, DPtrsGlobal, AElementOp, BElementOp, CElementOp, DsReduceOp, DsElementOp, DsElementOp, DGlobalMemOp, GemmSpecialization, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>; +// clang-format on + +using ReferenceGemmInstance = ck::tensor_operation::host:: + ReferenceGemm; + +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 = 4096; + ck::index_t StrideB = 4096; + ck::index_t StrideC = 4096; + + if(argc == 1) + { + // do nothing + } + 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 == 10) + { + 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]); + StrideC = std::stoi(argv[9]); + } + else + { + printf("arg1: verification (0=no, 1=yes)\n"); + printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"); + printf("arg3: run kernel # of times (>1)\n"); + printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC\n"); + exit(0); + } + + auto f_host_tensor_descriptor = + [](std::size_t row, std::size_t col, std::size_t stride, auto layout) { + if(std::is_same::value) + { + return HostTensorDescriptor(std::vector({row, col}), + std::vector({stride, 1})); + } + else + { + return HostTensorDescriptor(std::vector({row, col}), + std::vector({1, stride})); + } + }; + + Tensor a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{})); + Tensor b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{})); + + Tensor c_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{})); + Tensor d_m_host_result( + HostTensorDescriptor(std::vector({static_cast(M)}))); + + Tensor c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{})); + Tensor d_m_device_result( + HostTensorDescriptor(std::vector({static_cast(M)}))); + + std::cout << "a_m_k: " << a_m_k.mDesc << std::endl; + std::cout << "b_k_n: " << b_k_n.mDesc << std::endl; + std::cout << "c_m_n: " << c_m_n_host_result.mDesc << std::endl; + std::cout << "d_m: " << d_m_host_result.mDesc << std::endl; + + switch(init_method) + { + case 0: break; + case 1: + a_m_k.GenerateTensorValue(GeneratorTensor_2{-5, 5}); + b_k_n.GenerateTensorValue(GeneratorTensor_2{-5, 5}); + break; + default: + a_m_k.GenerateTensorValue(GeneratorTensor_3{0.0, 1.0}); + b_k_n.GenerateTensorValue(GeneratorTensor_3{-0.5, 0.5}); + break; + } + + DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpace()); + DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpace()); + DeviceMem c_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpace()); + DeviceMem d_device_buf(sizeof(DDataType) * d_m_device_result.mDesc.GetElementSpace()); + + a_device_buf.ToDevice(a_m_k.mData.data()); + b_device_buf.ToDevice(b_k_n.mData.data()); + + auto a_element_op = AElementOp{}; + auto b_element_op = BElementOp{}; + auto c_element_op = CElementOp{}; + auto ds_element_op = DsElementOp{}; + auto p_ds_global = ck::make_tuple(static_cast(d_device_buf.GetDeviceBuffer())); + + // do GEMM + auto gemm = DeviceGemmReduceInstance{}; + auto invoker = gemm.MakeInvoker(); + auto argument = gemm.MakeArgument(static_cast(a_device_buf.GetDeviceBuffer()), + static_cast(b_device_buf.GetDeviceBuffer()), + static_cast(c_device_buf.GetDeviceBuffer()), + p_ds_global, + M, + N, + K, + StrideA, + StrideB, + StrideC, + a_element_op, + b_element_op, + c_element_op, + ds_element_op, + ds_element_op); + + if(!gemm.IsSupportedArgument(argument)) + { + throw std::runtime_error( + "wrong! device_gemm with the specified compilation parameters does " + "not support this GEMM problem"); + } + + // init D + d_device_buf.SetValue(ck::NumericLimits::Lowest()); + + float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel}); + + std::size_t flop = std::size_t(2) * M * N * K; + std::size_t num_btype = + sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(CDataType) * 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, " + << gemm.GetTypeString() << std::endl; + + bool pass = true; + + if(do_verification) + { + c_device_buf.FromDevice(c_m_n_device_result.mData.data()); + d_device_buf.FromDevice(d_m_device_result.mData.data()); + + auto ref_gemm = ReferenceGemmInstance{}; + auto ref_invoker = ref_gemm.MakeInvoker(); + + auto ref_argument = ref_gemm.MakeArgument( + a_m_k, b_k_n, c_m_n_host_result, a_element_op, b_element_op, c_element_op); + + ref_invoker.Run(ref_argument); + + auto d_reduce_op = DsReduceOp{}[ck::Number<0>{}]; + + for(int m = 0; m < M; ++m) + { + ReduceAccDataType d_acc = d_reduce_op.GetReductionZeroVal(); + + for(int n = 0; n < N; ++n) + d_reduce_op(d_acc, c_m_n_host_result(m, n)); + + d_m_host_result(m) = d_acc; + } + + pass = ck::utils::check_err(c_m_n_device_result.mData, + c_m_n_host_result.mData, + "Error: Incorrect results c") && + ck::utils::check_err(d_m_device_result.mData, + d_m_host_result.mData, + "Error: Incorrect results d", + 1e-3, + 1e-3); + } + + return pass ? 0 : 1; +} diff --git a/example/16_gemm_reduce/gemm_reduce_xdl_fp16.cpp b/example/16_gemm_reduce/gemm_reduce_xdl_sum_squaresum_fp16.cpp similarity index 61% rename from example/16_gemm_reduce/gemm_reduce_xdl_fp16.cpp rename to example/16_gemm_reduce/gemm_reduce_xdl_sum_squaresum_fp16.cpp index 860d9eea2a..dff9c02f44 100644 --- a/example/16_gemm_reduce/gemm_reduce_xdl_fp16.cpp +++ b/example/16_gemm_reduce/gemm_reduce_xdl_sum_squaresum_fp16.cpp @@ -3,7 +3,7 @@ #include #include #include -#include + #include "check_err.hpp" #include "config.hpp" #include "device.hpp" @@ -26,10 +26,12 @@ using F32 = float; using Row = ck::tensor_layout::gemm::RowMajor; using Col = ck::tensor_layout::gemm::ColumnMajor; -using ADataType = F16; -using BDataType = F16; -using CDataType = F16; -using DDataType = F32; +using ADataType = F16; +using BDataType = F16; +using CDataType = F16; +using ReduceAccDataType = F32; +using DDataType = F32; +using DPtrsGlobal = ck::Tuple; using ALayout = ck::tensor_layout::gemm::RowMajor; using BLayout = ck::tensor_layout::gemm::ColumnMajor; @@ -38,20 +40,31 @@ using CLayout = ck::tensor_layout::gemm::RowMajor; using AElementOp = ck::tensor_operation::element_wise::PassThrough; using BElementOp = ck::tensor_operation::element_wise::PassThrough; using CElementOp = ck::tensor_operation::element_wise::PassThrough; -using D0ReduceOp = ck::reduce::Add; -using D1ReduceOp = ck::reduce::Add; -using D1ElementOp = ck::tensor_operation::element_wise::UnarySquare; +using D0ReduceOp = ck::reduce::Add; +using D1ReduceOp = ck::reduce::Add; +using DxsReduceOp = ck::Tuple; + +using UnaryIdenticElementOp = + ck::tensor_operation::element_wise::UnaryIdentic; +using UnarySquareElementOp = + ck::tensor_operation::element_wise::UnarySquare; +using DxsInElementOp = ck::Tuple; +using DxsOutElementOp = ck::Tuple; + +using DGlobalMemOp = + ck::InMemoryDataOperationEnumSequence; static constexpr auto GemmSpecialization = ck::tensor_operation::device::GemmSpecialization::Default; // clang-format off using DeviceGemmReduceInstance = ck::tensor_operation::device::DeviceGemmReduce_Xdl_CShuffle -//######| ALayout| BLayout| CLayout|AData| BData| CData| GemmAcc| CShuffle| ReduceAcc| DData| A| B| C| D0| D1| D1EleOp| GEMM| NumGemmK| 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| CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy| -//######| | | | Type| Type| Type| DataType| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Reduce| Reduce| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths| SrcDstScalarPerVector| SrcDstScalarPerVector| -//######| | | | | | | | | | | Operation| Operation| Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NPerBlock| _NPerBlock| _MPerBlock_NPerBlock| _NPerBlock| _MPerBlock| -//######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | - < Row, Col, Row, F16, F16, F16, F32, F32, F32, F32, AElementOp, BElementOp, CElementOp, D0ReduceOp, D1ReduceOp, D1ElementOp, GemmSpecialization, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>; +//######| ALayout| BLayout| CLayout|AData| BData| CData| GemmAcc| CShuffle| ReduceAcc| DData| A| B| C| Dxs| DxsInEleOp| DxsOutEleOp| D| GEMM| NumGemmK| 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| CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy| +//######| | | | Type| Type| Type| DataType| DataType| DataType| Type Tuple| Elementwise| Elementwise| Elementwise| Reduce| | | MemoryData| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths| SrcDstScalarPerVector| SrcDstScalarPerVector| +//######| | | | | | | | | | | Operation| Operation| Operation| Operation| | | Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NPerBlock| _NPerBlock| _MPerBlock_NPerBlock| _NPerBlock| _MPerBlock| +//######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | + < Row, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, AElementOp, BElementOp, CElementOp, DxsReduceOp, DxsInElementOp, DxsOutElementOp, DGlobalMemOp, GemmSpecialization, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>; // clang-format on using ReferenceGemmInstance = ck::tensor_operation::host:: @@ -162,10 +175,11 @@ int main(int argc, char* argv[]) a_device_buf.ToDevice(a_m_k.mData.data()); b_device_buf.ToDevice(b_k_n.mData.data()); - auto a_element_op = AElementOp{}; - auto b_element_op = BElementOp{}; - auto c_element_op = CElementOp{}; - auto d1_element_op = D1ElementOp{}; + auto a_element_op = AElementOp{}; + auto b_element_op = BElementOp{}; + auto c_element_op = CElementOp{}; + auto dxs_global = ck::make_tuple(static_cast(d0_device_buf.GetDeviceBuffer()), + static_cast(d1_device_buf.GetDeviceBuffer())); // do GEMM auto gemm = DeviceGemmReduceInstance{}; @@ -173,8 +187,7 @@ int main(int argc, char* argv[]) auto argument = gemm.MakeArgument(static_cast(a_device_buf.GetDeviceBuffer()), static_cast(b_device_buf.GetDeviceBuffer()), static_cast(c_device_buf.GetDeviceBuffer()), - static_cast(d0_device_buf.GetDeviceBuffer()), - static_cast(d1_device_buf.GetDeviceBuffer()), + dxs_global, M, N, K, @@ -184,7 +197,8 @@ int main(int argc, char* argv[]) a_element_op, b_element_op, c_element_op, - d1_element_op); + DxsInElementOp{}, + DxsOutElementOp{}); if(!gemm.IsSupportedArgument(argument)) { @@ -213,6 +227,7 @@ int main(int argc, char* argv[]) << gemm.GetTypeString() << std::endl; bool pass = true; + if(do_verification) { c_device_buf.FromDevice(c_m_n_device_result.mData.data()); @@ -237,10 +252,12 @@ int main(int argc, char* argv[]) for(int n = 0; n < N; ++n) { - float d0_val = ck::type_convert(c_m_n_host_result(m, n)); - float d1_val; + float c_val = ck::type_convert(c_m_n_host_result(m, n)); + float d0_val = 0; + float d1_val = 0; - d1_element_op(d1_val, d0_val); + UnaryIdenticElementOp{}(d0_val, c_val); + UnarySquareElementOp{}(d1_val, c_val); d0_reduce_op(d0_acc, d0_val); d1_reduce_op(d1_acc, d1_val); } @@ -249,18 +266,19 @@ int main(int argc, char* argv[]) d1_m_host_result(m) = ck::type_convert(d1_acc); } - pass &= ck::utils::check_err( - c_m_n_device_result.mData, c_m_n_host_result.mData, "Error: Incorrect results c"); - pass &= ck::utils::check_err(d0_m_device_result.mData, - d0_m_host_result.mData, - "Error: Incorrect results d0", - 1e-3, - 1e-3); - pass &= ck::utils::check_err(d1_m_device_result.mData, - d1_m_host_result.mData, - "Error: Incorrect results d1", - 1e-3, - 1e-3); + pass = ck::utils::check_err(c_m_n_device_result.mData, + c_m_n_host_result.mData, + "Error: Incorrect results c") && + ck::utils::check_err(d0_m_device_result.mData, + d0_m_host_result.mData, + "Error: Incorrect results d0", + 1e-4, + 1e-5) && + ck::utils::check_err(d1_m_device_result.mData, + d1_m_host_result.mData, + "Error: Incorrect results d1", + 1e-3, + 1e-5); } return pass ? 0 : 1; diff --git a/example/18_batched_gemm_reduce/batched_gemm_reduce_xdl_fp16.cpp b/example/18_batched_gemm_reduce/batched_gemm_reduce_xdl_fp16.cpp index d993c8e8d1..df63053c80 100644 --- a/example/18_batched_gemm_reduce/batched_gemm_reduce_xdl_fp16.cpp +++ b/example/18_batched_gemm_reduce/batched_gemm_reduce_xdl_fp16.cpp @@ -25,10 +25,12 @@ using F32 = float; using Row = ck::tensor_layout::gemm::RowMajor; using Col = ck::tensor_layout::gemm::ColumnMajor; -using ADataType = F16; -using BDataType = F16; -using CDataType = F16; -using DDataType = F32; +using ADataType = F16; +using BDataType = F16; +using CDataType = F16; +using ReduceAccDataType = F32; +using DDataType = F32; +using DPtrsGlobal = ck::Tuple; using ALayout = ck::tensor_layout::gemm::RowMajor; using BLayout = ck::tensor_layout::gemm::ColumnMajor; @@ -37,20 +39,31 @@ using CLayout = ck::tensor_layout::gemm::RowMajor; using AElementOp = ck::tensor_operation::element_wise::PassThrough; using BElementOp = ck::tensor_operation::element_wise::PassThrough; using CElementOp = ck::tensor_operation::element_wise::PassThrough; -using D0ReduceOp = ck::reduce::Add; -using D1ReduceOp = ck::reduce::Add; -using D1ElementOp = ck::tensor_operation::element_wise::UnarySquare; +using D0ReduceOp = ck::reduce::Add; +using D1ReduceOp = ck::reduce::Add; +using DxsReduceOp = ck::Tuple; + +using UnaryIdenticElementOp = + ck::tensor_operation::element_wise::UnaryIdentic; +using UnarySquareElementOp = + ck::tensor_operation::element_wise::UnarySquare; +using DxsInElementOp = ck::Tuple; +using DxsOutElementOp = ck::Tuple; + +using DGlobalMemOp = + ck::InMemoryDataOperationEnumSequence; static constexpr auto GemmSpecialization = ck::tensor_operation::device::GemmSpecialization::Default; // clang-format off using DeviceBatchedGemmReduceInstance = ck::tensor_operation::device::DeviceBatchedGemmReduce_Xdl_CShuffle -//######| ALayout| BLayout| CLayout|AData| BData| CData| GemmAcc| CShuffle| ReduceAcc| DData| A| B| C| D0| D1| D1EleOp| GEMM| NumGemmK| 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| CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy| -//######| | | | Type| Type| Type| DataType| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Reduce| Reduce| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths| SrcDstScalarPerVector| SrcDstScalarPerVector| -//######| | | | | | | | | | | Operation| Operation| Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NPerBlock| _NPerBlock| _MPerBlock_NPerBlock| _NPerBlock| _MPerBlock| -//######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | - < Row, Col, Row, F16, F16, F16, F32, F32, F32, F32, AElementOp, BElementOp, CElementOp, D0ReduceOp, D1ReduceOp, D1ElementOp, GemmSpecialization, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>; +//######| ALayout| BLayout| CLayout|AData| BData| CData| GemmAcc| CShuffle| ReduceAcc| DData| A| B| C| Dxs| DxsInEleOp| DxsOutEleOp| D| GEMM| NumGemmK| 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| CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy| +//######| | | | Type| Type| Type| DataType| DataType| DataType| Type Tuple| Elementwise| Elementwise| Elementwise| Reduce| | | MemoryData| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths| SrcDstScalarPerVector| SrcDstScalarPerVector| +//######| | | | | | | | | | | Operation| Operation| Operation| Operation| | | Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NPerBlock| _NPerBlock| _MPerBlock_NPerBlock| _NPerBlock| _MPerBlock| +//######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | + < Row, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, AElementOp, BElementOp, CElementOp, DxsReduceOp, DxsInElementOp, DxsOutElementOp, DGlobalMemOp, GemmSpecialization, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>; // clang-format on using ReferenceBatchedGemmInstance = ck::tensor_operation::host:: @@ -170,12 +183,11 @@ int main(int argc, char* argv[]) a_device_buf.ToDevice(a_g_m_k.mData.data()); b_device_buf.ToDevice(b_g_k_n.mData.data()); - auto a_element_op = AElementOp{}; - auto b_element_op = BElementOp{}; - auto c_element_op = CElementOp{}; - auto d0_reduce_op = D0ReduceOp{}; - auto d1_reduce_op = D1ReduceOp{}; - auto d1_element_op = D1ElementOp{}; + auto a_element_op = AElementOp{}; + auto b_element_op = BElementOp{}; + auto c_element_op = CElementOp{}; + auto dxs_global = ck::make_tuple(static_cast(d0_device_buf.GetDeviceBuffer()), + static_cast(d1_device_buf.GetDeviceBuffer())); // do GEMM auto batched_gemm = DeviceBatchedGemmReduceInstance{}; @@ -184,8 +196,7 @@ int main(int argc, char* argv[]) batched_gemm.MakeArgument(static_cast(a_device_buf.GetDeviceBuffer()), static_cast(b_device_buf.GetDeviceBuffer()), static_cast(c_device_buf.GetDeviceBuffer()), - static_cast(d0_device_buf.GetDeviceBuffer()), - static_cast(d1_device_buf.GetDeviceBuffer()), + dxs_global, M, N, K, @@ -195,7 +206,8 @@ int main(int argc, char* argv[]) a_element_op, b_element_op, c_element_op, - d1_element_op, + DxsInElementOp{}, + DxsOutElementOp{}, BatchCount); if(!batched_gemm.IsSupportedArgument(argument)) @@ -240,6 +252,9 @@ int main(int argc, char* argv[]) ref_invoker.Run(ref_argument); + auto d0_reduce_op = D0ReduceOp{}; + auto d1_reduce_op = D1ReduceOp{}; + for(int batch = 0; batch < BatchCount; ++batch) { for(int m = 0; m < M; ++m) @@ -249,10 +264,12 @@ int main(int argc, char* argv[]) for(int n = 0; n < N; ++n) { - float d0_val = ck::type_convert(c_g_m_n_host_result(batch, m, n)); - float d1_val; + float c_val = ck::type_convert(c_g_m_n_host_result(batch, m, n)); + float d0_val = 0; + float d1_val = 0; - d1_element_op(d1_val, d0_val); + UnaryIdenticElementOp{}(d0_val, c_val); + UnarySquareElementOp{}(d1_val, c_val); d0_reduce_op(d0_acc, d0_val); d1_reduce_op(d1_acc, d1_val); } @@ -262,17 +279,19 @@ int main(int argc, char* argv[]) } } - pass &= ck::utils::check_err(c_g_m_n_host_result.mData, c_g_m_n_device_result.mData); - pass &= ck::utils::check_err(d0_g_m_device_result.mData, - d0_g_m_host_result.mData, - "Error: Incorrect results! D0", - 1e-3, - 1e-3); - pass &= ck::utils::check_err(d1_g_m_device_result.mData, - d1_g_m_host_result.mData, - "Error: Incorrect results! D1", - 1e-3, - 1e-3); + pass = ck::utils::check_err(c_g_m_n_host_result.mData, + c_g_m_n_device_result.mData, + "Error: Incorrect results c") && + ck::utils::check_err(d0_g_m_device_result.mData, + d0_g_m_host_result.mData, + "Error: Incorrect results! D0", + 1e-4, + 1e-5) && + ck::utils::check_err(d1_g_m_device_result.mData, + d1_g_m_host_result.mData, + "Error: Incorrect results! D1", + 1e-3, + 1e-5); } return pass ? 0 : 1; diff --git a/example/CMakeLists.txt b/example/CMakeLists.txt index 1f4ed01de1..8661591b3f 100644 --- a/example/CMakeLists.txt +++ b/example/CMakeLists.txt @@ -33,7 +33,7 @@ function(add_example_executable_no_testing EXAMPLE_NAME FILE_NAME) add_executable(${EXAMPLE_NAME} ${FILE_NAME}) target_link_libraries(${EXAMPLE_NAME} PRIVATE host_tensor) add_dependencies(examples ${EXAMPLE_NAME}) -endfunction(add_example_executable EXAMPLE_NAME) +endfunction(add_example_executable_no_testing EXAMPLE_NAME) add_subdirectory(01_gemm) add_subdirectory(02_gemm_alpha_beta) diff --git a/include/ck/config.hpp b/include/ck/config.hpp index 710cd552d7..6699640424 100644 --- a/include/ck/config.hpp +++ b/include/ck/config.hpp @@ -76,6 +76,12 @@ #define CK_USE_AMD_BUFFER_ATOMIC_ADD_FLOAT 0 #endif +#if defined(__gfx90a__) // for GPU code +#define CK_USE_AMD_BUFFER_ATOMIC_MAX_FLOAT64 1 +#else +#define CK_USE_AMD_BUFFER_ATOMIC_MAX_FLOAT64 0 +#endif + // inline asm #define CK_USE_AMD_INLINE_ASM 1 @@ -91,10 +97,11 @@ // experimental feature: static tensor descriptor #define CK_EXPERIMENTAL_STATIC_TENSOR_DESCRIPTOR 0 -// experimental feature: buffer load/store/atomic-add OOB trick +// experimental feature: buffer load/store/atomic-add/ OOB trick #define CK_EXPERIMENTAL_USE_BUFFER_LOAD_OOB_CHECK_OFFSET_TRICK 0 #define CK_EXPERIMENTAL_USE_BUFFER_STORE_OOB_CHECK_OFFSET_TRICK 1 #define CK_EXPERIMENTAL_USE_BUFFER_ATOMIC_ADD_OOB_CHECK_OFFSET_TRICK 1 +#define CK_EXPERIMENTAL_USE_BUFFER_ATOMIC_MAX_OOB_CHECK_OFFSET_TRICK 1 // experimental feature: in-regsiter sub-dword transpose #define CK_EXPERIMENTAL_USE_IN_REGISTER_SUB_DWORD_TRANSPOSE 1 @@ -142,9 +149,23 @@ enum struct InMemoryDataOperationEnum { Set, AtomicAdd, + AtomicMax, Add }; +template +struct InMemoryDataOperationEnumSequence +{ + static constexpr int mSize = sizeof...(Is); + + __host__ __device__ static constexpr InMemoryDataOperationEnum At(int I) + { + // the last dummy element is to prevent compiler complain about empty array, when mSize = 0 + const InMemoryDataOperationEnum mData[mSize + 1] = {Is..., InMemoryDataOperationEnum::Set}; + return mData[I]; + } +}; + // TODO: no longer needed, remove this enum struct ActivTypeEnum { diff --git a/include/ck/tensor_operation/gpu/device/device_batched_gemm_reduce_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/device_batched_gemm_reduce_xdl_cshuffle.hpp index a6408007ed..273225c20a 100644 --- a/include/ck/tensor_operation/gpu/device/device_batched_gemm_reduce_xdl_cshuffle.hpp +++ b/include/ck/tensor_operation/gpu/device/device_batched_gemm_reduce_xdl_cshuffle.hpp @@ -17,11 +17,12 @@ namespace device { template (compute_base_ptr_of_batch_.GetCBasePtr(g_idx))); - const long_index_t d0_batch_offset = __builtin_amdgcn_readfirstlane( - static_cast(compute_base_ptr_of_batch_.GetD0BasePtr(g_idx))); - const long_index_t d1_batch_offset = __builtin_amdgcn_readfirstlane( - static_cast(compute_base_ptr_of_batch_.GetD1BasePtr(g_idx))); + static_for<0, p_ds_grid.Size(), 1>{}([&](auto In) { + const long_index_t d_batch_offset = __builtin_amdgcn_readfirstlane( + static_cast(compute_base_ptr_of_batch_.GetDBasePtr(g_idx, In))); + p_ds_grid(In) = p_ds_grid(In) + d_batch_offset; + }); __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()]; GridwiseGemm::template Run(p_a_grid + a_batch_offset, p_b_grid + b_batch_offset, p_c_grid + c_batch_offset, - p_d0_grid + d0_batch_offset, - p_d1_grid + d1_batch_offset, + p_ds_grid, p_shared, a_element_op, b_element_op, c_element_op, - d1_element_op, + dxs_in_element_op, + dxs_out_element_op, a_grid_desc_ak0_m_ak1, b_grid_desc_bk0_n_bk1, c_grid_desc_mblock_mperblock_nblock_nperblock, @@ -90,13 +92,13 @@ __global__ void ignore = p_a_grid; ignore = p_b_grid; ignore = p_c_grid; - ignore = p_d0_grid; - ignore = p_d1_grid; + ignore = p_ds_grid; ignore = batch_count; ignore = a_element_op; ignore = b_element_op; ignore = c_element_op; - ignore = d1_element_op; + ignore = dxs_in_element_op; + ignore = dxs_out_element_op; ignore = a_grid_desc_ak0_m_ak1; ignore = b_grid_desc_bk0_n_bk1; ignore = c_grid_desc_mblock_mperblock_nblock_nperblock; @@ -118,13 +120,14 @@ template -struct DeviceBatchedGemmReduce_Xdl_CShuffle : public DeviceGemmReduce + DxsInElementwiseOperation, + DxsOutElementwiseOperation> { using DeviceOp = DeviceBatchedGemmReduce_Xdl_CShuffle; @@ -508,13 +513,11 @@ struct DeviceBatchedGemmReduce_Xdl_CShuffle : public DeviceGemmReduce(BatchStrideC_); } - __host__ __device__ constexpr long_index_t GetD0BasePtr(index_t g_idx) const + template + __host__ __device__ constexpr long_index_t GetDBasePtr(index_t g_idx, + Number reduction_idx) const { - return g_idx * static_cast(BatchStrideD0_); - } - - __host__ __device__ constexpr long_index_t GetD1BasePtr(index_t g_idx) const - { - return g_idx * static_cast(BatchStrideD1_); + // TODO - Support sequence of StrideD in MakeArgument() + (void)reduction_idx; + return g_idx * static_cast(BatchStrideD_); } private: index_t BatchStrideA_; index_t BatchStrideB_; index_t BatchStrideC_; - index_t BatchStrideD0_; - index_t BatchStrideD1_; + index_t BatchStrideD_; }; // GridwiseGemm @@ -558,15 +559,15 @@ struct DeviceBatchedGemmReduce_Xdl_CShuffle : public DeviceGemmReduce(a_grid_desc_ak0_m_ak1_.GetElementSpaceSize()), type_convert(b_grid_desc_bk0_n_bk1_.GetElementSpaceSize()), type_convert(c_grid_desc_m_n_.GetElementSpaceSize()), - type_convert(d_grid_desc_m_.GetElementSpaceSize()), type_convert(d_grid_desc_m_.GetElementSpaceSize())}, block_2_ctile_map_{}, a_element_op_{a_element_op}, b_element_op_{b_element_op}, c_element_op_{c_element_op}, - d1_element_op_{d1_element_op} + dxs_in_element_op_{dxs_in_element_op}, + dxs_out_element_op_{dxs_out_element_op} { if(GridwiseGemm::CheckValidity( a_grid_desc_ak0_m_ak1_, b_grid_desc_bk0_n_bk1_, c_grid_desc_m_n_)) @@ -670,8 +670,7 @@ struct DeviceBatchedGemmReduce_Xdl_CShuffle : public DeviceGemmReduce MakeArgumentPointer(const void* p_a, const void* p_b, void* p_c, - void* p_d0, - void* p_d1, + DPtrsGlobal p_dxs, index_t MRaw, index_t NRaw, index_t KRaw, @@ -904,14 +905,14 @@ struct DeviceBatchedGemmReduce_Xdl_CShuffle : public DeviceGemmReduce(static_cast(p_a), static_cast(p_b), static_cast(p_c), - static_cast(p_d0), - static_cast(p_d1), + p_dxs, MRaw, NRaw, KRaw, @@ -921,7 +922,8 @@ struct DeviceBatchedGemmReduce_Xdl_CShuffle : public DeviceGemmReduce + typename DxsInElementwiseOperation, + typename DxsOutElementwiseOperation> struct DeviceGemmReduce : public BaseOperator { - virtual std::unique_ptr MakeArgumentPointer(const void* p_a, - const void* p_b, - void* p_c, - void* p_d0, - void* p_d1, - ck::index_t M, - ck::index_t N, - ck::index_t K, - ck::index_t StrideA, - ck::index_t StrideB, - ck::index_t StrideC, - AElementwiseOperation a_element_op, - BElementwiseOperation b_element_op, - CElementwiseOperation c_element_op, - D1ElementwiseOperation d1_element_op, - ck::index_t BatchCount = 1) = 0; + virtual std::unique_ptr + MakeArgumentPointer(const void* p_a, + const void* p_b, + void* p_c, + DPtrsGlobal p_dxs, + ck::index_t M, + ck::index_t N, + ck::index_t K, + ck::index_t StrideA, + ck::index_t StrideB, + ck::index_t StrideC, + AElementwiseOperation a_element_op, + BElementwiseOperation b_element_op, + CElementwiseOperation c_element_op, + DxsInElementwiseOperation dxs_in_element_op, + DxsOutElementwiseOperation dxs_out_element_op, + ck::index_t BatchCount = 1) = 0; virtual std::unique_ptr MakeInvokerPointer() = 0; }; -template -using DeviceGemmReducePtr = std::unique_ptr +using DeviceGemmReducePtr = std::unique_ptr>; + DxsInElementwiseOperation, + DxsOutElementwiseOperation>>; } // namespace device } // namespace tensor_operation diff --git a/include/ck/tensor_operation/gpu/device/device_gemm_reduce_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/device/device_gemm_reduce_xdl_cshuffle.hpp index 69c29b72d3..e8f48f9ba3 100644 --- a/include/ck/tensor_operation/gpu/device/device_gemm_reduce_xdl_cshuffle.hpp +++ b/include/ck/tensor_operation/gpu/device/device_gemm_reduce_xdl_cshuffle.hpp @@ -26,13 +26,14 @@ template -struct DeviceGemmReduce_Xdl_CShuffle : public DeviceGemmReduce + DxsInElementwiseOperation, + DxsOutElementwiseOperation> { using DeviceOp = DeviceGemmReduce_Xdl_CShuffle; @@ -380,15 +383,15 @@ struct DeviceGemmReduce_Xdl_CShuffle : public DeviceGemmReduce MakeArgumentPointer(const void* p_a, const void* p_b, void* p_c, - void* p_d0, - void* p_d1, + DPtrsGlobal p_dxs, index_t MRaw, index_t NRaw, index_t KRaw, @@ -695,14 +699,14 @@ struct DeviceGemmReduce_Xdl_CShuffle : public DeviceGemmReduce(static_cast(p_a), static_cast(p_b), static_cast(p_c), - static_cast(p_d0), - static_cast(p_d1), + p_dxs, MRaw, NRaw, KRaw, @@ -712,7 +716,8 @@ struct DeviceGemmReduce_Xdl_CShuffle : public DeviceGemmReduce(p_a_grid, p_b_grid, p_c_grid, - p_d0_grid, - p_d1_grid, + p_ds_grid, p_shared, a_element_op, b_element_op, c_element_op, - d1_element_op, + dxs_in_element_op, + dxs_out_element_op, a_grid_desc_ak0_m_ak1, b_grid_desc_bk0_n_bk1, c_grid_desc_mblock_mperblock_nblock_nperblock, @@ -69,12 +70,12 @@ __global__ void ignore = p_a_grid; ignore = p_b_grid; ignore = p_c_grid; - ignore = p_d0_grid; - ignore = p_d1_grid; + ignore = p_ds_grid; ignore = a_element_op; ignore = b_element_op; ignore = c_element_op; - ignore = d1_element_op; + ignore = dxs_in_element_op; + ignore = dxs_out_element_op; ignore = a_grid_desc_ak0_m_ak1; ignore = b_grid_desc_bk0_n_bk1; ignore = c_grid_desc_mblock_mperblock_nblock_nperblock; @@ -88,15 +89,15 @@ template ( p_c_grid, c_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize()); - auto d0_grid_buf = make_dynamic_buffer( - p_d0_grid, d_grid_desc_mblock_mperblock.GetElementSpaceSize()); - auto d1_grid_buf = make_dynamic_buffer( - p_d1_grid, d_grid_desc_mblock_mperblock.GetElementSpaceSize()); // divide block work by [M, N] const auto block_work_idx = @@ -527,7 +524,7 @@ struct GridwiseGemmReduce_k0mk1_k0nk1_mn_xdl_cshuffle_v1 c_thread_buf, num_k_block_main_loop); - // shuffle C and write out + // shuffle C + reduction + write out { static_assert(MXdlPerWave % CShuffleMXdlPerWavePerShuffle == 0 && NXdlPerWave % CShuffleNXdlPerWavePerShuffle == 0, @@ -666,6 +663,29 @@ struct GridwiseGemmReduce_k0mk1_k0nk1_mn_xdl_cshuffle_v1 make_multi_index(block_work_idx[I0], 0, block_work_idx[I1], 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>{}; + + // space filling curve for shuffled blockwise C in global mem + constexpr auto sfc_c_global = + SpaceFillingCurve, + Sequence<0, 2, 1, 3>, + Sequence<1, + CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl, + 1, + CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>>{}; + + // TODO: this should be implemented as a blockwise reduction // LDS c_reduce_block_desc_mperblock_nperblock constexpr auto c_reduce_block_desc_mperblock_nperblock = transform_tensor_descriptor( c_shuffle_block_desc_mblock_mperblock_nblock_nperblock, @@ -716,16 +736,9 @@ struct GridwiseGemmReduce_k0mk1_k0nk1_mn_xdl_cshuffle_v1 constexpr auto d_reduce_thread_desc_mblock_mperblock = make_naive_tensor_descriptor_packed(make_tuple(I1, Number{})); - // TODO: this should be implemented as a blockwise reduction auto c_reduce_thread_buf = make_static_buffer( c_reduce_thread_desc_mperblock_nperblock.GetElementSpaceSize()); - auto d0_thread_buf = make_static_buffer( - d_reduce_thread_desc_mperblock.GetElementSpaceSize()); - - auto d1_thread_buf = make_static_buffer( - d_reduce_thread_desc_mperblock.GetElementSpaceSize()); - // reduce: threadwise copy from LDS to VGPR constexpr auto c_reduce_thread_cluster_desc = make_cluster_descriptor( CReduceThreadClusterLengths_MPerBlock_NPerBlock{}, Sequence<1, 0>{}); @@ -749,47 +762,29 @@ struct GridwiseGemmReduce_k0mk1_k0nk1_mn_xdl_cshuffle_v1 1, true>{c_reduce_block_desc_mperblock_nperblock, c_reduce_thread_data_idx_begin}; - // reduce: copy from VGPR to global - auto d0_reduce_thread_copy_vgpr_to_global = ThreadwiseTensorSliceTransfer_v1r3< - FloatReduceAcc, - FloatD, - decltype(d_reduce_thread_desc_mblock_mperblock), - decltype(d_grid_desc_mblock_mperblock), - ck::tensor_operation::element_wise::PassThrough, - Sequence<1, mreduce_per_thread>, - Sequence<0, 1>, - 1, - CReduceThreadVgpr2GlobalCopySrcDstScalarPerVector_MPerBlock, - DGlobalMemoryDataOperation, - 1, - false>{d_grid_desc_mblock_mperblock, - make_multi_index(block_work_idx[I0], // mblock - c_reduce_thread_data_idx_begin[I0]), // mperblock - ck::tensor_operation::element_wise::PassThrough{}}; + auto dxs_reduce_thread_copy_vgpr_to_global = generate_tuple( + [&](auto I) { + auto p_d_grid = p_ds_grid[I]; + auto d_out_element_op = dxs_out_element_op[I]; - auto d1_reduce_thread_copy_vgpr_to_global = d0_reduce_thread_copy_vgpr_to_global; - - // space filling curve for threadwise C in VGPR - constexpr auto sfc_c_vgpr = - SpaceFillingCurve, - Sequence<0, 1, 2, 3, 4, 5, 6, 7>, - Sequence>{}; - - // space filling curve for shuffled blockwise C in global mem - constexpr auto sfc_c_global = - SpaceFillingCurve, - Sequence<0, 2, 1, 3>, - Sequence<1, - CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl, - 1, - CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>>{}; + return ThreadwiseTensorSliceTransfer_v1r3< + FloatReduceAcc, + remove_pointer_t, + decltype(d_reduce_thread_desc_mblock_mperblock), + decltype(d_grid_desc_mblock_mperblock), + decltype(d_out_element_op), + Sequence<1, mreduce_per_thread>, + Sequence<0, 1>, + 1, + CReduceThreadVgpr2GlobalCopySrcDstScalarPerVector_MPerBlock, + DGlobalMemoryDataOperation::At(I), + 1, + false>{d_grid_desc_mblock_mperblock, + make_multi_index(block_work_idx[I0], // mblock + c_reduce_thread_data_idx_begin[I0]), // mperblock + d_out_element_op}; + }, + Number{}); constexpr index_t num_access = sfc_c_vgpr.GetNumOfAccess(); @@ -816,64 +811,73 @@ struct GridwiseGemmReduce_k0mk1_k0nk1_mn_xdl_cshuffle_v1 c_grid_desc_mblock_mperblock_nblock_nperblock, c_grid_buf); - using ThreadwiseReduce_D0 = - ThreadwiseReduction; - - using ThreadwiseReduce_D1 = - ThreadwiseReduction; - - const auto d0_zeroVal = D0ReduceOperation::GetReductionZeroVal(); - const auto d1_zeroVal = D0ReduceOperation::GetReductionZeroVal(); - - static_for<0, mreduce_per_thread, 1>{}( - [&](auto I) { d0_thread_buf(I) = d0_zeroVal; }); - static_for<0, mreduce_per_thread, 1>{}( - [&](auto I) { d1_thread_buf(I) = d1_zeroVal; }); - - // reduce + // TODO - extract following into reduction_blockwise { - // copy from LDS to VGPR c_reduce_thread_copy_lds_to_vgpr.Run(c_reduce_block_desc_mperblock_nperblock, c_shuffle_block_buf, c_reduce_thread_desc_mperblock_nperblock, make_tuple(I0, I0), c_reduce_thread_buf); - // reduce in VGPR - ThreadwiseReduce_D0::Reduce(c_reduce_thread_buf, d0_thread_buf); + static_for<0, p_ds_grid.Size(), 1>{}([&](auto In) { + auto& p_d_grid = p_ds_grid[In]; - static_for<0, mreduce_per_thread, 1>{}([&](auto im) { - static_for<0, nreduce_per_thread, 1>{}([&](auto in) { - constexpr auto offset = - Number{}; + auto d_grid_buf = make_dynamic_buffer( + p_d_grid, d_grid_desc_mblock_mperblock.GetElementSpaceSize()); - d1_element_op(c_reduce_thread_buf(offset), c_reduce_thread_buf(offset)); + auto d_thread_buf = + make_static_buffer( + d_reduce_thread_desc_mperblock.GetElementSpaceSize()); + + auto& d_in_element_op = dxs_in_element_op[In]; + + auto& d_reduce_thread_copy_vgpr_to_global = + dxs_reduce_thread_copy_vgpr_to_global(In); + + using DReduceOperation = remove_cvref_t; + using ThreadwiseReduce = + ThreadwiseReduction; + + // Global write Gemm shuffle + reduction + const auto d_zeroVal = DReduceOperation::GetReductionZeroVal(); + + static_for<0, mreduce_per_thread, 1>{}( + [&](auto I) { d_thread_buf(I) = d_zeroVal; }); + + // reduce in VGPR + static_for<0, mreduce_per_thread, 1>{}([&](auto im) { + static_for<0, nreduce_per_thread, 1>{}([&](auto in) { + constexpr auto offset = + Number{}; + + d_in_element_op(c_reduce_thread_buf(offset), + c_reduce_thread_buf(offset)); + }); }); + + ThreadwiseReduce::Reduce(c_reduce_thread_buf, d_thread_buf); + + // copy from VGPR to Global + d_reduce_thread_copy_vgpr_to_global.Run( + d_reduce_thread_desc_mblock_mperblock, + make_tuple(I0, I0), + d_thread_buf, + d_grid_desc_mblock_mperblock, + d_grid_buf); + + if constexpr(access_id < num_access - 1) + { + constexpr auto c_global_step = sfc_c_global.GetForwardStep(access_id); + d_reduce_thread_copy_vgpr_to_global.MoveDstSliceWindow( + d_grid_desc_mblock_mperblock, + make_tuple(c_global_step[I0], c_global_step[I1])); + } }); - - ThreadwiseReduce_D1::Reduce(c_reduce_thread_buf, d1_thread_buf); - - // copy from VGPR to Global - d0_reduce_thread_copy_vgpr_to_global.Run(d_reduce_thread_desc_mblock_mperblock, - make_tuple(I0, I0), - d0_thread_buf, - d_grid_desc_mblock_mperblock, - d0_grid_buf); - - d1_reduce_thread_copy_vgpr_to_global.Run(d_reduce_thread_desc_mblock_mperblock, - make_tuple(I0, I0), - d1_thread_buf, - d_grid_desc_mblock_mperblock, - d1_grid_buf); } if constexpr(access_id < num_access - 1) @@ -883,18 +887,10 @@ struct GridwiseGemmReduce_k0mk1_k0nk1_mn_xdl_cshuffle_v1 // move on C c_shuffle_block_copy_lds_to_global.MoveDstSliceWindow( c_grid_desc_mblock_mperblock_nblock_nperblock, c_global_step); - - // move on D0 - d0_reduce_thread_copy_vgpr_to_global.MoveDstSliceWindow( - d_grid_desc_mblock_mperblock, - make_tuple(c_global_step[I0], c_global_step[I1])); - - // move on D1 - d1_reduce_thread_copy_vgpr_to_global.MoveDstSliceWindow( - d_grid_desc_mblock_mperblock, - make_tuple(c_global_step[I0], c_global_step[I1])); } }); + + // Reduction } } }; diff --git a/include/ck/utility/amd_buffer_addressing.hpp b/include/ck/utility/amd_buffer_addressing.hpp index 53c24b9a98..6831658fc9 100644 --- a/include/ck/utility/amd_buffer_addressing.hpp +++ b/include/ck/utility/amd_buffer_addressing.hpp @@ -258,6 +258,14 @@ __device__ float llvm_amdgcn_raw_buffer_atomic_add_fp32( index_t soffset, index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.atomic.fadd.f32"); +// buffer atomic-add fp32 +__device__ double +llvm_amdgcn_raw_buffer_atomic_max_fp64(double vdata, + int32x4_t rsrc, // dst_wave_buffer_resource + int voffset, // dst_thread_addr_offset + int soffset, // dst_wave_addr_offset + int glc_slc) __asm("llvm.amdgcn.raw.buffer.atomic.fmax.f64"); + template __device__ typename vector_type::type amd_buffer_load_impl(int32x4_t src_wave_buffer_resource, index_t src_thread_addr_offset, @@ -915,6 +923,71 @@ __device__ void amd_buffer_atomic_add_impl(const typename vector_type::typ } } +template +__device__ void amd_buffer_atomic_max_impl(const typename vector_type::type src_thread_data, + int32x4_t dst_wave_buffer_resource, + index_t dst_thread_addr_offset, + index_t dst_wave_addr_offset) +{ + static_assert((is_same::value && (N == 1 || N == 2 || N == 4)), + "wrong! not implemented"); + if constexpr(is_same::value) + { + if constexpr(N == 1) + { + llvm_amdgcn_raw_buffer_atomic_max_fp64(src_thread_data, + dst_wave_buffer_resource, + dst_thread_addr_offset, + dst_wave_addr_offset, + 0); + } + else if constexpr(N == 2) + { + vector_type tmp{src_thread_data}; + + llvm_amdgcn_raw_buffer_atomic_max_fp64(tmp.AsType()[Number<0>{}], + dst_wave_buffer_resource, + dst_thread_addr_offset, + dst_wave_addr_offset, + 0); + + llvm_amdgcn_raw_buffer_atomic_max_fp64(tmp.AsType()[Number<1>{}], + dst_wave_buffer_resource, + dst_thread_addr_offset, + dst_wave_addr_offset + sizeof(double), + 0); + } + else if constexpr(N == 4) + { + vector_type tmp{src_thread_data}; + + llvm_amdgcn_raw_buffer_atomic_max_fp64(tmp.AsType()[Number<0>{}], + dst_wave_buffer_resource, + dst_thread_addr_offset, + dst_wave_addr_offset, + 0); + + llvm_amdgcn_raw_buffer_atomic_max_fp64(tmp.AsType()[Number<1>{}], + dst_wave_buffer_resource, + dst_thread_addr_offset, + dst_wave_addr_offset + sizeof(double), + 0); + + llvm_amdgcn_raw_buffer_atomic_max_fp64(tmp.AsType()[Number<2>{}], + dst_wave_buffer_resource, + dst_thread_addr_offset, + dst_wave_addr_offset + 2 * sizeof(double), + 0); + + llvm_amdgcn_raw_buffer_atomic_max_fp64(tmp.AsType()[Number<3>{}], + dst_wave_buffer_resource, + dst_thread_addr_offset, + dst_wave_addr_offset + 3 * sizeof(double), + 0); + } + } +} + // buffer_load requires: // 1) p_src_wave must point to global memory space // 2) p_src_wave must be a wavewise pointer. @@ -1046,4 +1119,39 @@ amd_buffer_atomic_add(const typename vector_type_maker::type::type src_thr #endif } +// buffer_atomic_max requires: +// 1) p_dst_wave must point to global memory +// 2) p_dst_wave must be a wavewise pointer. +// It is user's responsibility to make sure that is true. +template +__device__ void +amd_buffer_atomic_max(const typename vector_type_maker::type::type src_thread_data, + T* p_dst_wave, + const index_t dst_thread_element_offset, + const bool dst_thread_element_valid, + const index_t dst_element_space_size) +{ + const int32x4_t dst_wave_buffer_resource = + make_wave_buffer_resource(p_dst_wave, dst_element_space_size); + + index_t dst_thread_addr_offset = dst_thread_element_offset * sizeof(T); + + using vector_t = typename vector_type_maker::type::type; + using scalar_t = typename scalar_type::type; + constexpr index_t vector_size = scalar_type::vector_size; + +#if CK_EXPERIMENTAL_USE_BUFFER_ATOMIC_MAX_OOB_CHECK_OFFSET_TRICK + uint32_t dst_addr_shift = dst_thread_element_valid ? 0 : 0x7fffffff; + + amd_buffer_atomic_max_impl( + src_thread_data, dst_wave_buffer_resource, dst_addr_shift + dst_thread_addr_offset, 0); +#else + if(dst_thread_element_valid) + { + amd_buffer_atomic_max_impl( + src_thread_data, dst_wave_buffer_resource, dst_thread_addr_offset, 0); + } +#endif +} + } // namespace ck diff --git a/include/ck/utility/common_header.hpp b/include/ck/utility/common_header.hpp index 539263703b..34c0a7821b 100644 --- a/include/ck/utility/common_header.hpp +++ b/include/ck/utility/common_header.hpp @@ -32,7 +32,7 @@ #include "debug.hpp" #include "amd_buffer_addressing.hpp" -#include "generic_memory_space_atomic_add.hpp" +#include "generic_memory_space_atomic.hpp" #include "get_id.hpp" #include "synchronization.hpp" #include "amd_address_space.hpp" diff --git a/include/ck/utility/dynamic_buffer.hpp b/include/ck/utility/dynamic_buffer.hpp index c00982dfff..0ad78423fe 100644 --- a/include/ck/utility/dynamic_buffer.hpp +++ b/include/ck/utility/dynamic_buffer.hpp @@ -3,7 +3,7 @@ #include "enable_if.hpp" #include "c_style_pointer_cast.hpp" #include "amd_buffer_addressing.hpp" -#include "generic_memory_space_atomic_add.hpp" +#include "generic_memory_space_atomic.hpp" namespace ck { @@ -125,6 +125,10 @@ struct DynamicBuffer { this->template AtomicAdd(i, is_valid_element, x); } + else if constexpr(Op == InMemoryDataOperationEnum::AtomicMax) + { + this->template AtomicMax(i, is_valid_element, x); + } else if constexpr(Op == InMemoryDataOperationEnum::Add) { auto tmp = this->template Get(i, is_valid_element); @@ -326,6 +330,42 @@ struct DynamicBuffer } } + template >::type, + typename scalar_type>::type>::value, + bool>::type = false> + __host__ __device__ void AtomicMax(index_t i, bool is_valid_element, const X& x) + { + // X contains multiple T + constexpr index_t scalar_per_t_vector = scalar_type>::vector_size; + + constexpr index_t scalar_per_x_vector = scalar_type>::vector_size; + + static_assert(scalar_per_x_vector % scalar_per_t_vector == 0, + "wrong! X should contain multiple T"); + + static_assert(GetAddressSpace() == AddressSpaceEnum::Global, "only support global mem"); + +#if CK_USE_AMD_BUFFER_ATOMIC_MAX_FLOAT64 + using scalar_t = typename scalar_type>::type; + bool constexpr use_amd_buffer_addressing = is_same_v, double>; +#else + bool constexpr use_amd_buffer_addressing = false; +#endif + + if constexpr(use_amd_buffer_addressing) + { + constexpr index_t t_per_x = scalar_per_x_vector / scalar_per_t_vector; + + amd_buffer_atomic_max, t_per_x>( + x, p_data_, i, is_valid_element, element_space_size_); + } + else if(is_valid_element) + { + atomic_max(c_style_pointer_cast(&p_data_[i]), x); + } + } + __host__ __device__ static constexpr bool IsStaticBuffer() { return false; } __host__ __device__ static constexpr bool IsDynamicBuffer() { return true; } diff --git a/include/ck/utility/generic_memory_space_atomic.hpp b/include/ck/utility/generic_memory_space_atomic.hpp new file mode 100644 index 0000000000..712d815f52 --- /dev/null +++ b/include/ck/utility/generic_memory_space_atomic.hpp @@ -0,0 +1,97 @@ +#pragma once +#include "data_type.hpp" + +namespace ck { + +// Caution: DO NOT REMOVE +// intentionally have only declaration but no definition to cause compilation failure when trying to +// instantiate this template. The purpose is to make the implementation of atomic_add explicit for +// each datatype. +template +__device__ X atomic_add(X* p_dst, const X& x); + +template <> +__device__ int32_t atomic_add(int32_t* p_dst, const int32_t& x) +{ + return atomicAdd(p_dst, x); +} + +template <> +__device__ uint32_t atomic_add(uint32_t* p_dst, const uint32_t& x) +{ + return atomicAdd(p_dst, x); +} + +template <> +__device__ float atomic_add(float* p_dst, const float& x) +{ + return atomicAdd(p_dst, x); +} + +template <> +__device__ float2_t atomic_add(float2_t* p_dst, const float2_t& x) +{ + constexpr auto I0 = Number<0>{}; + constexpr auto I1 = Number<1>{}; + + const vector_type vx{x}; + vector_type vy{0}; + + vy.template AsType()(I0) = + atomicAdd(c_style_pointer_cast(p_dst), vx.template AsType()[I0]); + vy.template AsType()(I1) = + atomicAdd(c_style_pointer_cast(p_dst) + 1, vx.template AsType()[I1]); + + return vy.template AsType()[I0]; +} + +// Caution: DO NOT REMOVE +// intentionally have only declaration but no definition to cause compilation failure when trying to +// instantiate this template. The purpose is to make the implementation of atomic_max explicit for +// each datatype. + +template +__device__ X atomic_max(X* p_dst, const X& x); + +template <> +__device__ int32_t atomic_max(int32_t* p_dst, const int32_t& x) +{ + return atomicMax(p_dst, x); +} + +template <> +__device__ uint32_t atomic_max(uint32_t* p_dst, const uint32_t& x) +{ + return atomicMax(p_dst, x); +} + +template <> +__device__ float atomic_max(float* p_dst, const float& x) +{ + return atomicMax(p_dst, x); +} + +template <> +__device__ double atomic_max(double* p_dst, const double& x) +{ + return atomicMax(p_dst, x); +} + +template <> +__device__ float2_t atomic_max(float2_t* p_dst, const float2_t& x) +{ + constexpr auto I0 = Number<0>{}; + constexpr auto I1 = Number<1>{}; + + const vector_type vx{x}; + vector_type vy{0}; + + vy.template AsType()(I0) = + atomicMax(c_style_pointer_cast(p_dst), vx.template AsType()[I0]); + vy.template AsType()(I1) = + atomicMax(c_style_pointer_cast(p_dst) + 1, vx.template AsType()[I1]); + + return vy.template AsType()[I0]; +} + +} // namespace ck diff --git a/include/ck/utility/generic_memory_space_atomic_add.hpp b/include/ck/utility/generic_memory_space_atomic_add.hpp deleted file mode 100644 index 8ee2081776..0000000000 --- a/include/ck/utility/generic_memory_space_atomic_add.hpp +++ /dev/null @@ -1,44 +0,0 @@ -#pragma once -#include "data_type.hpp" - -namespace ck { - -template -__device__ X atomic_add(X* p_dst, const X& x); - -template <> -__device__ int32_t atomic_add(int32_t* p_dst, const int32_t& x) -{ - return atomicAdd(p_dst, x); -} - -template <> -__device__ uint32_t atomic_add(uint32_t* p_dst, const uint32_t& x) -{ - return atomicAdd(p_dst, x); -} - -template <> -__device__ float atomic_add(float* p_dst, const float& x) -{ - return atomicAdd(p_dst, x); -} - -template <> -__device__ float2_t atomic_add(float2_t* p_dst, const float2_t& x) -{ - constexpr auto I0 = Number<0>{}; - constexpr auto I1 = Number<1>{}; - - const vector_type vx{x}; - vector_type vy{0}; - - vy.template AsType()(I0) = - atomicAdd(c_style_pointer_cast(p_dst), vx.template AsType()[I0]); - vy.template AsType()(I1) = - atomicAdd(c_style_pointer_cast(p_dst) + 1, vx.template AsType()[I1]); - - return vy.template AsType()[I0]; -} - -} // namespace ck diff --git a/include/ck/utility/type.hpp b/include/ck/utility/type.hpp index e212c82232..ee3189ebe5 100644 --- a/include/ck/utility/type.hpp +++ b/include/ck/utility/type.hpp @@ -29,6 +29,9 @@ using remove_cv_t = typename std::remove_cv::type; template using remove_cvref_t = remove_cv_t>; +template +using remove_pointer_t = typename std::remove_pointer::type; + template inline constexpr bool is_pointer_v = std::is_pointer::value; diff --git a/library/include/ck/library/host_tensor/device.hpp b/library/include/ck/library/host_tensor/device.hpp index d549b14c8c..990d2f98b3 100644 --- a/library/include/ck/library/host_tensor/device.hpp +++ b/library/include/ck/library/host_tensor/device.hpp @@ -10,6 +10,15 @@ #include "stream_config.hpp" #include "ck/options.hpp" +template +__global__ void set_buffer_value(T* p, T x, uint64_t buffer_element_size) +{ + for(uint64_t i = threadIdx.x; i < buffer_element_size; i += blockDim.x) + { + p[i] = x; + } +} + inline void hip_check_error(hipError_t x) { if(x != hipSuccess) @@ -30,6 +39,16 @@ struct DeviceMem void ToDevice(const void* p); void FromDevice(void* p); void SetZero(); + template + void SetValue(T x) + { + if(mMemSize % sizeof(T) != 0) + { + throw std::runtime_error("wrong! not entire DeviceMem will be set"); + } + + set_buffer_value<<<1, 1024>>>(static_cast(mpDeviceBuf), x, mMemSize / sizeof(T)); + } ~DeviceMem(); void* mpDeviceBuf; @@ -74,8 +93,7 @@ float launch_and_time_kernel(const StreamConfig& stream_config, printf("Warm up 1 time\n"); // warm up - hipLaunchKernelGGL( - kernel, grid_dim, block_dim, lds_byte, stream_config.stream_id_, args...); + kernel<<>>(args...); printf("Start running %d times...\n", nrepeat); @@ -84,8 +102,7 @@ float launch_and_time_kernel(const StreamConfig& stream_config, for(int i = 0; i < nrepeat; ++i) { - hipLaunchKernelGGL( - kernel, grid_dim, block_dim, lds_byte, stream_config.stream_id_, args...); + kernel<<>>(args...); } timer.End(); @@ -94,13 +111,12 @@ float launch_and_time_kernel(const StreamConfig& stream_config, } else { - hipLaunchKernelGGL( - kernel, grid_dim, block_dim, lds_byte, stream_config.stream_id_, args...); + kernel<<>>(args...); return 0; } #else - hipLaunchKernelGGL(kernel, grid_dim, block_dim, lds_byte, stream_config.stream_id_, args...); + kernel<<>>(args...); return 0; #endif diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_reduce/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/batched_gemm_reduce/CMakeLists.txt index 67a3c15d00..0606df01f1 100644 --- a/library/src/tensor_operation_instance/gpu/batched_gemm_reduce/CMakeLists.txt +++ b/library/src/tensor_operation_instance/gpu/batched_gemm_reduce/CMakeLists.txt @@ -1,4 +1,4 @@ -set(DEVICE_BATCHED_GEMM_REDUCE_INSTANCE_SOURCE +set(DEVICE_BATCHED_GEMM_REDUCE_INSTANCE_SOURCE device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gkn_gmn_instance.cpp device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gnk_gmn_instance.cpp device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gkn_gmn_instance.cpp diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gkn_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gkn_gmn_instance.cpp index 3653169921..322b0ddaf5 100644 --- a/library/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gkn_gmn_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gkn_gmn_instance.cpp @@ -10,8 +10,9 @@ namespace tensor_operation { namespace device { namespace device_gemm_instance { -using F16 = ck::half_t; -using F32 = float; +using F16 = ck::half_t; +using F32 = float; +using DPtrsGlobal = ck::Tuple; using Row = ck::tensor_layout::gemm::RowMajor; using Col = ck::tensor_layout::gemm::ColumnMajor; @@ -21,41 +22,52 @@ using S = ck::Sequence; using PassThrough = ck::tensor_operation::element_wise::PassThrough; using ReduceSum = ck::reduce::Add; -using Square = ck::tensor_operation::element_wise::UnarySquare; +using ReduceOps = ck::Tuple; + +using Identity = ck::tensor_operation::element_wise::UnaryIdentic; +using Square = ck::tensor_operation::element_wise::UnarySquare; +using DInElementOps = ck::Tuple; +using DOutElementOps = ck::Tuple; + +using ReduceMemOp = ck::InMemoryDataOperationEnumSequence; static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default; // c[g, m, n] = a[g, m, k] * b[g, n, k] -// d0[g, m] = reduce0(c[g, m, n]) -// d1[g, m] = reduce1(c[g, m, n]) using device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gkn_gmn_instances = std::tuple< // clang-format off - //##################################| ALayout| BLayout| CLayout|AData| BData| CData| GemmAcc| CShuffle| ReduceAcc| DData| A| B| C| D0| D1| D1EleOp| GEMM| NumGemmK| 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| CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy| - //##################################| | | | Type| Type| Type| DataType| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Reduce| Reduce| Spacialization| Prefetch| 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_MPerBlock| ScalarPerVector| ThreadClusterLengths| SrcDstScalarPerVector| SrcDstScalarPerVector| - //##################################| | | | | | | | | | | Operation| Operation| Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NPerBlock| _NPerBlock| _MPerBlock_NPerBlock| _NPerBlock| _MPerBlock| - //##################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | - DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 256, 128, 32, 2, 2, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, - DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, - DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 256, 32, 4, 4, 32, 32, 2, 4, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, - DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, - DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 128, 128, 32, 2, 2, 32, 32, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>, - DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>, - DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 128, 32, 2, 2, 32, 32, 2, 2, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, - DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, - DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 128, 64, 32, 2, 2, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>, - DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>, - DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 64, 128, 32, 2, 2, 32, 32, 2, 2, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>, - DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>, - DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 64, 32, 2, 2, 32, 32, 2, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 16, 1, 4>, 8, S<64, 4>, 4, 1>, - DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, 1, 1, S<1, 16, 1, 4>, 8, S<64, 4>, 4, 1>, - DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 64, 128, 32, 2, 2, 32, 32, 1, 2, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, - DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1> + //##################################| ALayout| BLayout| CLayout|AData| BData| CData| GemmAcc| CShuffle| ReduceAcc| DData| A| B| C| Dxs| DxsInEleOp| DxsOutEleOp| D| GEMM| NumGemmK| 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| CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy| + //##################################| | | | Type| Type| Type| DataType| DataType| DataType| Type Tuple| Elementwise| Elementwise| Elementwise| Reduce| | | MemoryData|Spacialization| Prefetch| 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_MPerBlock| ScalarPerVector| ThreadClusterLengths| SrcDstScalarPerVector| SrcDstScalarPerVector| + //##################################| | | | | | | | | | | Operation| Operation| Operation| Operation| | | Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NPerBlock| _NPerBlock| _MPerBlock_NPerBlock| _NPerBlock| _MPerBlock| + //##################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | + DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 256, 128, 32, 2, 2, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, + DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, + DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 256, 32, 4, 4, 32, 32, 2, 4, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, + DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, + DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 128, 128, 32, 2, 2, 32, 32, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>, + DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>, + DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 128, 32, 2, 2, 32, 32, 2, 2, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, + DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, + DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 128, 64, 32, 2, 2, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>, + DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>, + DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 64, 128, 32, 2, 2, 32, 32, 2, 2, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>, + DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>, + DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 64, 32, 2, 2, 32, 32, 2, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 16, 1, 4>, 8, S<64, 4>, 4, 1>, + DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, 1, 1, S<1, 16, 1, 4>, 8, S<64, 4>, 4, 1>, + DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 64, 128, 32, 2, 2, 32, 32, 1, 2, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, + DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1> // clang-format on >; void add_device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gkn_gmn_instances( - std::vector>& instances) + std::vector>& instances) { add_device_operation_instances( instances, diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gnk_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gnk_gmn_instance.cpp index 070056980d..bdc5aebe1a 100644 --- a/library/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gnk_gmn_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gnk_gmn_instance.cpp @@ -10,8 +10,9 @@ namespace tensor_operation { namespace device { namespace device_gemm_instance { -using F16 = ck::half_t; -using F32 = float; +using F16 = ck::half_t; +using F32 = float; +using DPtrsGlobal = ck::Tuple; using Row = ck::tensor_layout::gemm::RowMajor; using Col = ck::tensor_layout::gemm::ColumnMajor; @@ -21,41 +22,52 @@ using S = ck::Sequence; using PassThrough = ck::tensor_operation::element_wise::PassThrough; using ReduceSum = ck::reduce::Add; -using Square = ck::tensor_operation::element_wise::UnarySquare; +using ReduceOps = ck::Tuple; + +using Identity = ck::tensor_operation::element_wise::UnaryIdentic; +using Square = ck::tensor_operation::element_wise::UnarySquare; +using DInElementOps = ck::Tuple; +using DOutElementOps = ck::Tuple; + +using ReduceMemOp = ck::InMemoryDataOperationEnumSequence; static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default; // c[g, m, n] = a[g, m, k] * b[g, n, k] -// d0[g, m] = reduce0(c[g, m, n]) -// d1[g, m] = reduce1(c[g, m, n]) using device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gnk_gmn_instances = std::tuple< // clang-format off - //##################################| ALayout| BLayout| CLayout|AData| BData| CData| GemmAcc| CShuffle| ReduceAcc| DData| A| B| C| D0| D1| D1EleOp| GEMM| NumGemmK| 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| CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy| - //##################################| | | | Type| Type| Type| DataType| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Reduce| Reduce| Spacialization| Prefetch| 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_MPerBlock| ScalarPerVector| ThreadClusterLengths| SrcDstScalarPerVector| SrcDstScalarPerVector| - //##################################| | | | | | | | | | | Operation| Operation| Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NPerBlock| _NPerBlock| _MPerBlock_NPerBlock| _NPerBlock| _MPerBlock| - //##################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | - DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 256, 128, 32, 2, 8, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, - DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, - DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 256, 32, 2, 8, 32, 32, 2, 4, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, - DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, - DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 128, 128, 32, 2, 8, 32, 32, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>, - DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>, - DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 128, 32, 2, 8, 32, 32, 2, 2, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, - DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, - DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 128, 64, 32, 2, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>, - DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>, - DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 64, 128, 32, 2, 8, 32, 32, 2, 2, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>, - DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>, - DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 64, 32, 2, 8, 32, 32, 2, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 4>, 8, S<64, 4>, 4, 1>, - DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 4>, 8, S<64, 4>, 4, 1>, - DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 64, 128, 32, 2, 8, 32, 32, 1, 2, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, - DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1> + //##################################| ALayout| BLayout| CLayout|AData| BData| CData| GemmAcc| CShuffle| ReduceAcc| DData| A| B| C| Dxs| DxsInEleOp| DxsOutEleOp| D| GEMM| NumGemmK| 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| CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy| + //##################################| | | | Type| Type| Type| DataType| DataType| DataType| Type Tuple| Elementwise| Elementwise| Elementwise| Reduce| | | MemoryData|Spacialization| Prefetch| 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_MPerBlock| ScalarPerVector| ThreadClusterLengths| SrcDstScalarPerVector| SrcDstScalarPerVector| + //##################################| | | | | | | | | | | Operation| Operation| Operation| Operation| | | Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NPerBlock| _NPerBlock| _MPerBlock_NPerBlock| _NPerBlock| _MPerBlock| + //##################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | + DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 256, 128, 32, 2, 8, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, + DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, + DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 256, 32, 2, 8, 32, 32, 2, 4, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, + DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, + DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 128, 128, 32, 2, 8, 32, 32, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>, + DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>, + DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 128, 32, 2, 8, 32, 32, 2, 2, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, + DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, + DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 128, 64, 32, 2, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>, + DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>, + DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 64, 128, 32, 2, 8, 32, 32, 2, 2, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>, + DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>, + DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 64, 32, 2, 8, 32, 32, 2, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 4>, 8, S<64, 4>, 4, 1>, + DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 4>, 8, S<64, 4>, 4, 1>, + DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 64, 128, 32, 2, 8, 32, 32, 1, 2, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, + DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1> // clang-format on >; void add_device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gnk_gmn_instances( - std::vector>& instances) + std::vector>& instances) { add_device_operation_instances( instances, diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gkn_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gkn_gmn_instance.cpp index f242b3c12e..df51cb617b 100644 --- a/library/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gkn_gmn_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gkn_gmn_instance.cpp @@ -10,8 +10,9 @@ namespace tensor_operation { namespace device { namespace device_gemm_instance { -using F16 = ck::half_t; -using F32 = float; +using F16 = ck::half_t; +using F32 = float; +using DPtrsGlobal = ck::Tuple; using Row = ck::tensor_layout::gemm::RowMajor; using Col = ck::tensor_layout::gemm::ColumnMajor; @@ -21,41 +22,52 @@ using S = ck::Sequence; using PassThrough = ck::tensor_operation::element_wise::PassThrough; using ReduceSum = ck::reduce::Add; -using Square = ck::tensor_operation::element_wise::UnarySquare; +using ReduceOps = ck::Tuple; + +using Identity = ck::tensor_operation::element_wise::UnaryIdentic; +using Square = ck::tensor_operation::element_wise::UnarySquare; +using DInElementOps = ck::Tuple; +using DOutElementOps = ck::Tuple; + +using ReduceMemOp = ck::InMemoryDataOperationEnumSequence; static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default; // c[g, m, n] = a[g, m, k] * b[g, n, k] -// d0[g, m] = reduce0(c[g, m, n]) -// d1[g, m] = reduce1(c[g, m, n]) using device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gkn_gmn_instances = std::tuple< // clang-format off - //##################################| ALayout| BLayout| CLayout| AData| BData| CData| GemmAcc| CShuffle| ReduceAcc| DData| A| B| C| D0| D1| D1EleOp| GEMM| NumGemmK| 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| CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy| - //##################################| | | | Type| Type| Type| DataType| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Reduce| Reduce| Spacialization| Prefetch| 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_MPerBlock| ScalarPerVector| ThreadClusterLengths| SrcDstScalarPerVector| SrcDstScalarPerVector| - //##################################| | | | | | | | | | | Operation| Operation| Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NPerBlock| _NPerBlock| _MPerBlock_NPerBlock| _NPerBlock| _MPerBlock| - //##################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | - DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 256, 128, 32, 8, 2, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, - DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, - DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 256, 32, 8, 2, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, - DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, - DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 128, 128, 32, 8, 2, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>, - DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>, - DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 128, 32, 8, 2, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, - DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, - DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 128, 64, 32, 8, 2, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>, - DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>, - DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 64, 128, 32, 8, 2, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>, - DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>, - DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 64, 32, 8, 2, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 16, 1, 4>, 8, S<64, 4>, 4, 1>, - DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, 1, 1, S<1, 16, 1, 4>, 8, S<64, 4>, 4, 1>, - DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 64, 128, 32, 8, 2, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, - DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1> + //##################################| ALayout| BLayout| CLayout| AData| BData| CData| GemmAcc| CShuffle| ReduceAcc| DData| A| B| C| Dxs| DxsInEleOp| DxsOutEleOp| D| GEMM| NumGemmK| 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| CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy| + //##################################| | | | Type| Type| Type| DataType| DataType| DataType| Type Tuple| Elementwise| Elementwise| Elementwise| Reduce| | | MemoryData|Spacialization| Prefetch| 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_MPerBlock| ScalarPerVector| ThreadClusterLengths| SrcDstScalarPerVector| SrcDstScalarPerVector| + //##################################| | | | | | | | | | | Operation| Operation| Operation| Operation| | | Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NPerBlock| _NPerBlock| _MPerBlock_NPerBlock| _NPerBlock| _MPerBlock| + //##################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | + DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 256, 128, 32, 8, 2, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, + DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, + DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 256, 32, 8, 2, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, + DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, + DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 128, 128, 32, 8, 2, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>, + DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>, + DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 128, 32, 8, 2, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, + DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, + DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 128, 64, 32, 8, 2, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>, + DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>, + DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 64, 128, 32, 8, 2, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>, + DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>, + DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 64, 32, 8, 2, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 16, 1, 4>, 8, S<64, 4>, 4, 1>, + DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, 1, 1, S<1, 16, 1, 4>, 8, S<64, 4>, 4, 1>, + DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 64, 128, 32, 8, 2, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, + DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1> // clang-format on >; void add_device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gkn_gmn_instances( - std::vector>& instances) + std::vector>& instances) { add_device_operation_instances( instances, diff --git a/library/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gnk_gmn_instance.cpp b/library/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gnk_gmn_instance.cpp index cbf3c16171..10afddb5c6 100644 --- a/library/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gnk_gmn_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gnk_gmn_instance.cpp @@ -10,8 +10,9 @@ namespace tensor_operation { namespace device { namespace device_gemm_instance { -using F16 = ck::half_t; -using F32 = float; +using F16 = ck::half_t; +using F32 = float; +using DPtrsGlobal = ck::Tuple; using Row = ck::tensor_layout::gemm::RowMajor; using Col = ck::tensor_layout::gemm::ColumnMajor; @@ -21,38 +22,49 @@ using S = ck::Sequence; using PassThrough = ck::tensor_operation::element_wise::PassThrough; using ReduceSum = ck::reduce::Add; -using Square = ck::tensor_operation::element_wise::UnarySquare; +using ReduceOps = ck::Tuple; + +using Identity = ck::tensor_operation::element_wise::UnaryIdentic; +using Square = ck::tensor_operation::element_wise::UnarySquare; +using DInElementOps = ck::Tuple; +using DOutElementOps = ck::Tuple; + +using ReduceMemOp = ck::InMemoryDataOperationEnumSequence; static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default; // c[g, m, n] = a[g, m, k] * b[g, n, k] -// d0[g, m] = reduce0(c[g, m, n]) -// d1[g, m] = reduce1(c[g, m, n]) using device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gnk_gmn_instances = std::tuple< // clang-format off - //##################################| ALayout| BLayout| CLayout|AData| BData| CData| GemmAcc| CShuffle| ReduceAcc| DData| A| B| C| D0| D1| D1EleOp| GEMM| NumGemmK| 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| CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy| - //##################################| | | | Type| Type| Type| DataType| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Reduce| Reduce| Spacialization| Prefetch| 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_MPerBlock| ScalarPerVector| ThreadClusterLengths| SrcDstScalarPerVector| SrcDstScalarPerVector| - //##################################| | | | | | | | | | | Operation| Operation| Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NPerBlock| _NPerBlock| _MPerBlock_NPerBlock| _NPerBlock| _MPerBlock| - //##################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | - DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, - DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, - DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>, - DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, - DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>, - DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>, - DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 4>, 8, S<32, 2>, 4, 1>, - DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, - DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, - DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 128, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>, - DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 32, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>, - DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 4>, 8, S<32, 2>, 4, 1>, - DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 64, 32, 64, 32, 8, 8, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 4>, 8, S<32, 2>, 4, 1> + //##################################| ALayout| BLayout| CLayout|AData| BData| CData| GemmAcc| CShuffle| ReduceAcc| DData| A| B| C| Dxs| DxsInEleOp| DxsOutEleOp| D| GEMM| NumGemmK| 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| CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy| + //##################################| | | | Type| Type| Type| DataType| DataType| DataType| Type Tuple| Elementwise| Elementwise| Elementwise| Reduce| | | MemoryData|Spacialization| Prefetch| 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_MPerBlock| ScalarPerVector| ThreadClusterLengths| SrcDstScalarPerVector| SrcDstScalarPerVector| + //##################################| | | | | | | | | | | Operation| Operation| Operation| Operation| | | Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NPerBlock| _NPerBlock| _MPerBlock_NPerBlock| _NPerBlock| _MPerBlock| + //##################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | + DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, + DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, + DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>, + DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, + DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>, + DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>, + DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 4>, 8, S<32, 2>, 4, 1>, + DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, + DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, + DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 128, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>, + DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 32, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>, + DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 4>, 8, S<32, 2>, 4, 1>, + DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 64, 32, 64, 32, 8, 8, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 4>, 8, S<32, 2>, 4, 1> // clang-format on >; void add_device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gnk_gmn_instances( - std::vector>& instances) + std::vector>& instances) { add_device_operation_instances( instances, diff --git a/library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_kn_mn_instance.cpp index 2f1509b6c8..33660c0481 100644 --- a/library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_kn_mn_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_kn_mn_instance.cpp @@ -10,8 +10,9 @@ namespace tensor_operation { namespace device { namespace device_gemm_instance { -using F16 = ck::half_t; -using F32 = float; +using F16 = ck::half_t; +using F32 = float; +using DPtrsGlobal = ck::Tuple; using Row = ck::tensor_layout::gemm::RowMajor; using Col = ck::tensor_layout::gemm::ColumnMajor; @@ -21,40 +22,51 @@ using S = ck::Sequence; using PassThrough = ck::tensor_operation::element_wise::PassThrough; using ReduceSum = ck::reduce::Add; -using Square = ck::tensor_operation::element_wise::UnarySquare; +using ReduceOps = ck::Tuple; + +using Identity = ck::tensor_operation::element_wise::UnaryIdentic; +using Square = ck::tensor_operation::element_wise::UnarySquare; +using DInElementOps = ck::Tuple; +using DOutElementOps = ck::Tuple; + +using ReduceMemOp = ck::InMemoryDataOperationEnumSequence; static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default; // c[m, n] = a[k, m] * b[k, n] -// d0[m] = reduce0(c[m, n]) -// d1[m] = reduce1(c[m, n]) using device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_kn_mn_instances = std::tuple< // clang-format off - //###########################| ALayout| BLayout| CLayout|AData| BData| CData| GemmAcc| CShuffle| ReduceAcc| DData| A| B| C| D0| D1| D1EleOp| GEMM| NumGemmK| 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| CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy| - //###########################| | | | Type| Type| Type| DataType| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Reduce| Reduce| Spacialization| Prefetch| 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_MPerBlock| ScalarPerVector| ThreadClusterLengths| SrcDstScalarPerVector| SrcDstScalarPerVector| - //###########################| | | | | | | | | | | Operation| Operation| Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NPerBlock| _NPerBlock| _MPerBlock_NPerBlock| _NPerBlock| _MPerBlock| - //###########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | - DeviceGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 256, 128, 32, 2, 2, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, - DeviceGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, - DeviceGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 256, 32, 2, 2, 32, 32, 2, 4, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, - DeviceGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, - DeviceGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 128, 128, 32, 2, 2, 32, 32, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>, - DeviceGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>, - DeviceGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 128, 32, 2, 2, 32, 32, 2, 2, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, - DeviceGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, - DeviceGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 128, 64, 32, 2, 2, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>, - DeviceGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>, - DeviceGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 64, 128, 32, 2, 2, 32, 32, 2, 2, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>, - DeviceGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>, - DeviceGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 64, 32, 2, 2, 32, 32, 2, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 16, 1, 4>, 8, S<64, 4>, 4, 1>, - DeviceGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, 1, 1, S<1, 16, 1, 4>, 8, S<64, 4>, 4, 1>, - DeviceGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 64, 128, 32, 2, 2, 32, 32, 1, 2, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, - DeviceGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1> + //###########################| ALayout| BLayout| CLayout|AData| BData| CData| GemmAcc| CShuffle| ReduceAcc| DData| A| B| C| Dxs| DxsInEleOp| DxsOutEleOp| D| GEMM| NumGemmK| 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| CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy| + //###########################| | | | Type| Type| Type| DataType| DataType| DataType| Type Tuple| Elementwise| Elementwise| Elementwise| Reduce| | | MemoryData|Spacialization| Prefetch| 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_MPerBlock| ScalarPerVector| ThreadClusterLengths| SrcDstScalarPerVector| SrcDstScalarPerVector| + //###########################| | | | | | | | | | | Operation| Operation| Operation| Operation| | | Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NPerBlock| _NPerBlock| _MPerBlock_NPerBlock| _NPerBlock| _MPerBlock| + //###########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | + DeviceGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 256, 128, 32, 2, 2, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, + DeviceGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, + DeviceGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 256, 32, 2, 2, 32, 32, 2, 4, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, + DeviceGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, + DeviceGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 128, 128, 32, 2, 2, 32, 32, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>, + DeviceGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>, + DeviceGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 128, 32, 2, 2, 32, 32, 2, 2, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, + DeviceGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, + DeviceGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 128, 64, 32, 2, 2, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>, + DeviceGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>, + DeviceGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 64, 128, 32, 2, 2, 32, 32, 2, 2, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>, + DeviceGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>, + DeviceGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 64, 32, 2, 2, 32, 32, 2, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 16, 1, 4>, 8, S<64, 4>, 4, 1>, + DeviceGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, 1, 1, S<1, 16, 1, 4>, 8, S<64, 4>, 4, 1>, + DeviceGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 64, 128, 32, 2, 2, 32, 32, 1, 2, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, + DeviceGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1> // clang-format on >; void add_device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_kn_mn_instances( - std::vector>& instances) + std::vector>& instances) { add_device_operation_instances( instances, device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_kn_mn_instances{}); diff --git a/library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_nk_mn_instance.cpp index c3e04287e4..bd8766a617 100644 --- a/library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_nk_mn_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_nk_mn_instance.cpp @@ -10,8 +10,9 @@ namespace tensor_operation { namespace device { namespace device_gemm_instance { -using F16 = ck::half_t; -using F32 = float; +using F16 = ck::half_t; +using F32 = float; +using DPtrsGlobal = ck::Tuple; using Row = ck::tensor_layout::gemm::RowMajor; using Col = ck::tensor_layout::gemm::ColumnMajor; @@ -21,40 +22,51 @@ using S = ck::Sequence; using PassThrough = ck::tensor_operation::element_wise::PassThrough; using ReduceSum = ck::reduce::Add; -using Square = ck::tensor_operation::element_wise::UnarySquare; +using ReduceOps = ck::Tuple; + +using Identity = ck::tensor_operation::element_wise::UnaryIdentic; +using Square = ck::tensor_operation::element_wise::UnarySquare; +using DInElementOps = ck::Tuple; +using DOutElementOps = ck::Tuple; + +using ReduceMemOp = ck::InMemoryDataOperationEnumSequence; static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default; // c[m, n] = a[k, m] * b[n, k] -// d0[m] = reduce0(c[m, n]) -// d1[m] = reduce1(c[m, n]) using device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_nk_mn_instances = std::tuple< // clang-format off - //###########################| ALayout| BLayout| CLayout|AData| BData| CData| GemmAcc| CShuffle| ReduceAcc| DData| A| B| C| D0| D1| D1EleOp| GEMM| NumGemmK| 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| CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy| - //###########################| | | | Type| Type| Type| DataType| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Reduce| Reduce| Spacialization| Prefetch| 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_MPerBlock| ScalarPerVector| ThreadClusterLengths| SrcDstScalarPerVector| SrcDstScalarPerVector| - //###########################| | | | | | | | | | | Operation| Operation| Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NPerBlock| _NPerBlock| _MPerBlock_NPerBlock| _NPerBlock| _MPerBlock| - //###########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | - DeviceGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 256, 128, 32, 2, 8, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, - DeviceGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, - DeviceGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 256, 32, 2, 8, 32, 32, 2, 4, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, - DeviceGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, - DeviceGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 128, 128, 32, 2, 8, 32, 32, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>, - DeviceGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>, - DeviceGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 128, 32, 2, 8, 32, 32, 2, 2, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, - DeviceGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, - DeviceGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 128, 64, 32, 2, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>, - DeviceGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>, - DeviceGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 64, 128, 32, 2, 8, 32, 32, 2, 2, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>, - DeviceGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>, - DeviceGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 64, 32, 2, 8, 32, 32, 2, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 4>, 8, S<64, 4>, 4, 1>, - DeviceGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 4>, 8, S<64, 4>, 4, 1>, - DeviceGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 64, 128, 32, 2, 8, 32, 32, 1, 2, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, - DeviceGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1> + //###########################| ALayout| BLayout| CLayout|AData| BData| CData| GemmAcc| CShuffle| ReduceAcc| DData| A| B| C| Dxs| DxsInEleOp| DxsOutEleOp| D| GEMM| NumGemmK| 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| CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy| + //###########################| | | | Type| Type| Type| DataType| DataType| DataType| Type Tuple| Elementwise| Elementwise| Elementwise| Reduce| | | MemoryData|Spacialization| Prefetch| 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_MPerBlock| ScalarPerVector| ThreadClusterLengths| SrcDstScalarPerVector| SrcDstScalarPerVector| + //###########################| | | | | | | | | | | Operation| Operation| Operation| Operation| | | Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NPerBlock| _NPerBlock| _MPerBlock_NPerBlock| _NPerBlock| _MPerBlock| + //###########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | + DeviceGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 256, 128, 32, 2, 8, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, + DeviceGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, + DeviceGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 256, 32, 2, 8, 32, 32, 2, 4, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, + DeviceGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, + DeviceGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 128, 128, 32, 2, 8, 32, 32, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>, + DeviceGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>, + DeviceGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 128, 32, 2, 8, 32, 32, 2, 2, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, + DeviceGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, + DeviceGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 128, 64, 32, 2, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>, + DeviceGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>, + DeviceGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 64, 128, 32, 2, 8, 32, 32, 2, 2, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>, + DeviceGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>, + DeviceGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 64, 32, 2, 8, 32, 32, 2, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 4>, 8, S<64, 4>, 4, 1>, + DeviceGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 4>, 8, S<64, 4>, 4, 1>, + DeviceGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 64, 128, 32, 2, 8, 32, 32, 1, 2, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, + DeviceGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1> // clang-format on >; void add_device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_nk_mn_instances( - std::vector>& instances) + std::vector>& instances) { add_device_operation_instances( instances, device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_nk_mn_instances{}); diff --git a/library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_kn_mn_instance.cpp index e845c3bf82..c04431c1e0 100644 --- a/library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_kn_mn_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_kn_mn_instance.cpp @@ -10,8 +10,9 @@ namespace tensor_operation { namespace device { namespace device_gemm_instance { -using F16 = ck::half_t; -using F32 = float; +using F16 = ck::half_t; +using F32 = float; +using DPtrsGlobal = ck::Tuple; using Row = ck::tensor_layout::gemm::RowMajor; using Col = ck::tensor_layout::gemm::ColumnMajor; @@ -21,40 +22,51 @@ using S = ck::Sequence; using PassThrough = ck::tensor_operation::element_wise::PassThrough; using ReduceSum = ck::reduce::Add; -using Square = ck::tensor_operation::element_wise::UnarySquare; +using ReduceOps = ck::Tuple; + +using Identity = ck::tensor_operation::element_wise::UnaryIdentic; +using Square = ck::tensor_operation::element_wise::UnarySquare; +using DInElementOps = ck::Tuple; +using DOutElementOps = ck::Tuple; + +using ReduceMemOp = ck::InMemoryDataOperationEnumSequence; static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default; // c[m, n] = a[m, k] * b[n, k] -// d0[m] = reduce0(c[m, n]) -// d1[m] = reduce1(c[m, n]) using device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_kn_mn_instances = std::tuple< // clang-format off - //###########################| ALayout| BLayout| CLayout| AData| BData| CData| GemmAcc| CShuffle| ReduceAcc| DData| A| B| C| D0| D1| D1EleOp| GEMM| NumGemmK| 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| CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy| - //###########################| | | | Type| Type| Type| DataType| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Reduce| Reduce| Spacialization| Prefetch| 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_MPerBlock| ScalarPerVector| ThreadClusterLengths| SrcDstScalarPerVector| SrcDstScalarPerVector| - //###########################| | | | | | | | | | | Operation| Operation| Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NPerBlock| _NPerBlock| _MPerBlock_NPerBlock| _NPerBlock| _MPerBlock| - //###########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | - DeviceGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 256, 128, 32, 8, 2, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, - DeviceGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, - DeviceGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 256, 32, 8, 2, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, - DeviceGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, - DeviceGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 128, 128, 32, 8, 2, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>, - DeviceGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>, - DeviceGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 128, 32, 8, 2, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, - DeviceGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, - DeviceGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 128, 64, 32, 8, 2, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>, - DeviceGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>, - DeviceGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 64, 128, 32, 8, 2, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>, - DeviceGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>, - DeviceGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 64, 32, 8, 2, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 16, 1, 4>, 8, S<64, 4>, 4, 1>, - DeviceGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, 1, 1, S<1, 16, 1, 4>, 8, S<64, 4>, 4, 1>, - DeviceGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 64, 128, 32, 8, 2, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, - DeviceGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1> + //###########################| ALayout| BLayout| CLayout| AData| BData| CData| GemmAcc| CShuffle| ReduceAcc| DData| A| B| C| Dxs| DxsInEleOp| DxsOutEleOp| D| GEMM| NumGemmK| 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| CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy| + //###########################| | | | Type| Type| Type| DataType| DataType| DataType| Type Tuple| Elementwise| Elementwise| Elementwise| Reduce| | | MemoryData|Spacialization| Prefetch| 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_MPerBlock| ScalarPerVector| ThreadClusterLengths| SrcDstScalarPerVector| SrcDstScalarPerVector| + //###########################| | | | | | | | | | | Operation| Operation| Operation| Operation| | | Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NPerBlock| _NPerBlock| _MPerBlock_NPerBlock| _NPerBlock| _MPerBlock| + //###########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | + DeviceGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 256, 128, 32, 8, 2, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, + DeviceGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, + DeviceGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 256, 32, 8, 2, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, + DeviceGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, + DeviceGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 128, 128, 32, 8, 2, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>, + DeviceGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>, + DeviceGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 128, 32, 8, 2, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, + DeviceGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, + DeviceGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 128, 64, 32, 8, 2, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>, + DeviceGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>, + DeviceGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 64, 128, 32, 8, 2, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>, + DeviceGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>, + DeviceGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 64, 32, 8, 2, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 16, 1, 4>, 8, S<64, 4>, 4, 1>, + DeviceGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, 1, 1, S<1, 16, 1, 4>, 8, S<64, 4>, 4, 1>, + DeviceGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 64, 128, 32, 8, 2, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, + DeviceGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1> // clang-format on >; void add_device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_kn_mn_instances( - std::vector>& instances) + std::vector>& instances) { add_device_operation_instances( instances, device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_kn_mn_instances{}); diff --git a/library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_nk_mn_instance.cpp index a356170789..ebd89e5975 100644 --- a/library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_nk_mn_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_nk_mn_instance.cpp @@ -10,8 +10,9 @@ namespace tensor_operation { namespace device { namespace device_gemm_instance { -using F16 = ck::half_t; -using F32 = float; +using F16 = ck::half_t; +using F32 = float; +using DPtrsGlobal = ck::Tuple; using Row = ck::tensor_layout::gemm::RowMajor; using Col = ck::tensor_layout::gemm::ColumnMajor; @@ -21,37 +22,48 @@ using S = ck::Sequence; using PassThrough = ck::tensor_operation::element_wise::PassThrough; using ReduceSum = ck::reduce::Add; -using Square = ck::tensor_operation::element_wise::UnarySquare; +using ReduceOps = ck::Tuple; + +using Identity = ck::tensor_operation::element_wise::UnaryIdentic; +using Square = ck::tensor_operation::element_wise::UnarySquare; +using DInElementOps = ck::Tuple; +using DOutElementOps = ck::Tuple; + +using ReduceMemOp = ck::InMemoryDataOperationEnumSequence; static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default; // c[m, n] = a[m, k] * b[n, k] -// d0[m] = reduce0(c[m, n]) -// d1[m] = reduce1(c[m, n]) using device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_nk_mn_instances = std::tuple< // clang-format off - //###########################| ALayout| BLayout| CLayout|AData| BData| CData| GemmAcc| CShuffle| ReduceAcc| DData| A| B| C| D0| D1| D1EleOp| GEMM| NumGemmK| 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| CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy| - //###########################| | | | Type| Type| Type| DataType| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Reduce| Reduce| Spacialization| Prefetch| 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_MPerBlock| ScalarPerVector| ThreadClusterLengths| SrcDstScalarPerVector| SrcDstScalarPerVector| - //###########################| | | | | | | | | | | Operation| Operation| Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NPerBlock| _NPerBlock| _MPerBlock_NPerBlock| _NPerBlock| _MPerBlock| - //###########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | - DeviceGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, - DeviceGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, - DeviceGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>, - DeviceGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, - DeviceGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>, - DeviceGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>, - DeviceGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 4>, 8, S<32, 2>, 4, 1>, - DeviceGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, - DeviceGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, - DeviceGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 128, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>, - DeviceGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 32, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>, - DeviceGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 4>, 8, S<32, 2>, 4, 1>, - DeviceGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 64, 32, 64, 32, 8, 8, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 4>, 8, S<32, 2>, 4, 1> + //###########################| ALayout| BLayout| CLayout|AData| BData| CData| GemmAcc| CShuffle| ReduceAcc| DData| A| B| C| Dxs| DxsInEleOp| DxsOutEleOp| D| GEMM| NumGemmK| 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| CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy| + //###########################| | | | Type| Type| Type| DataType| DataType| DataType| Type Tuple| Elementwise| Elementwise| Elementwise| Reduce| | | MemoryData| Spacialization| Prefetch| 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_MPerBlock| ScalarPerVector| ThreadClusterLengths| SrcDstScalarPerVector| SrcDstScalarPerVector| + //###########################| | | | | | | | | | | Operation| Operation| Operation| Operation| | | Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NPerBlock| _NPerBlock| _MPerBlock_NPerBlock| _NPerBlock| _MPerBlock| + //###########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | + DeviceGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, + DeviceGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, + DeviceGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>, + DeviceGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, + DeviceGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>, + DeviceGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>, + DeviceGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 4>, 8, S<32, 2>, 4, 1>, + DeviceGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, + DeviceGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>, + DeviceGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 128, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>, + DeviceGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 32, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>, + DeviceGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 4>, 8, S<32, 2>, 4, 1>, + DeviceGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 64, 32, 64, 32, 8, 8, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 4>, 8, S<32, 2>, 4, 1> // clang-format on >; void add_device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_nk_mn_instances( - std::vector>& instances) + std::vector>& instances) { add_device_operation_instances( instances, device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_nk_mn_instances{}); diff --git a/profiler/include/profile_batched_gemm_reduce_impl.hpp b/profiler/include/profile_batched_gemm_reduce_impl.hpp index bd74dbf459..56ca2cbebe 100644 --- a/profiler/include/profile_batched_gemm_reduce_impl.hpp +++ b/profiler/include/profile_batched_gemm_reduce_impl.hpp @@ -17,11 +17,21 @@ namespace tensor_operation { namespace device { namespace device_gemm_instance { +using F32 = float; +using F16 = ck::half_t; +using DPtrsGlobal = ck::Tuple; +using Identity = ck::tensor_operation::element_wise::UnaryIdentic; +using Square = ck::tensor_operation::element_wise::UnarySquare; +using DInElementOps = ck::Tuple; +using DOutElementOps = ck::Tuple; + using DeviceGemmReduceNoOpPtr = ck::tensor_operation::device::DeviceGemmReducePtr< + DPtrsGlobal, ck::tensor_operation::element_wise::PassThrough, ck::tensor_operation::element_wise::PassThrough, ck::tensor_operation::element_wise::PassThrough, - ck::tensor_operation::element_wise::UnarySquare>; + DInElementOps, + DOutElementOps>; void add_device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gkn_gmn_instances( std::vector&); @@ -119,19 +129,25 @@ bool profile_batched_gemm_reduce_impl(int do_verification, b_g_k_n.GenerateTensorValue(GeneratorTensor_3{-0.5, 0.5}, num_thread); } - using AElementOp = ck::tensor_operation::element_wise::PassThrough; - using BElementOp = ck::tensor_operation::element_wise::PassThrough; - using CElementOp = ck::tensor_operation::element_wise::PassThrough; - using D0ReduceOp = ck::reduce::Add; - using D1ReduceOp = ck::reduce::Add; - using D1ElementOp = ck::tensor_operation::element_wise::UnarySquare; + using AElementOp = ck::tensor_operation::element_wise::PassThrough; + using BElementOp = ck::tensor_operation::element_wise::PassThrough; + using CElementOp = ck::tensor_operation::element_wise::PassThrough; + using D0ReduceOp = ck::reduce::Add; + using D1ReduceOp = ck::reduce::Add; + using UnaryIdenticElementOp = + ck::tensor_operation::element_wise::UnaryIdentic; + using UnarySquareElementOp = + ck::tensor_operation::element_wise::UnarySquare; + using DxsInElementOps = ck::Tuple; + using DxsOutElementOps = ck::Tuple; - const auto a_element_op = AElementOp{}; - const auto b_element_op = BElementOp{}; - const auto c_element_op = CElementOp{}; - const auto d0_reduce_op = D0ReduceOp{}; - const auto d1_reduce_op = D1ReduceOp{}; - const auto d1_element_op = D1ElementOp{}; + const auto a_element_op = AElementOp{}; + const auto b_element_op = BElementOp{}; + const auto c_element_op = CElementOp{}; + const auto dxs_in_element_op = DxsInElementOps{}; + const auto dxs_out_element_op = DxsOutElementOps{}; + const auto d0_reduce_op = D0ReduceOp{}; + const auto d1_reduce_op = D1ReduceOp{}; if(do_verification) { @@ -163,7 +179,7 @@ bool profile_batched_gemm_reduce_impl(int do_verification, float d0_val = ck::type_convert(c_g_m_n_host_result(batch, m, n)); float d1_val; - d1_element_op(d1_val, d0_val); + UnarySquareElementOp{}(d1_val, d0_val); d0_reduce_op(d0_acc, d0_val); d1_reduce_op(d1_acc, d1_val); } @@ -180,6 +196,9 @@ bool profile_batched_gemm_reduce_impl(int do_verification, DeviceMem d0_device_buf(sizeof(DDataType) * d0_g_m_device_result.mDesc.GetElementSpace()); DeviceMem d1_device_buf(sizeof(DDataType) * d1_g_m_device_result.mDesc.GetElementSpace()); + auto dxs_global = ck::make_tuple(static_cast(d0_device_buf.GetDeviceBuffer()), + static_cast(d1_device_buf.GetDeviceBuffer())); + a_device_buf.ToDevice(a_g_m_k.mData.data()); b_device_buf.ToDevice(b_g_k_n.mData.data()); @@ -241,8 +260,7 @@ bool profile_batched_gemm_reduce_impl(int do_verification, gemm_ptr->MakeArgumentPointer(static_cast(a_device_buf.GetDeviceBuffer()), static_cast(b_device_buf.GetDeviceBuffer()), static_cast(c_device_buf.GetDeviceBuffer()), - static_cast(d0_device_buf.GetDeviceBuffer()), - static_cast(d1_device_buf.GetDeviceBuffer()), + dxs_global, M, N, K, @@ -252,7 +270,8 @@ bool profile_batched_gemm_reduce_impl(int do_verification, a_element_op, b_element_op, c_element_op, - d1_element_op, + dxs_in_element_op, + dxs_out_element_op, BatchCount); auto invoker_ptr = gemm_ptr->MakeInvokerPointer(); diff --git a/profiler/include/profile_gemm_reduce_impl.hpp b/profiler/include/profile_gemm_reduce_impl.hpp index d034c9f750..97d0f2523b 100644 --- a/profiler/include/profile_gemm_reduce_impl.hpp +++ b/profiler/include/profile_gemm_reduce_impl.hpp @@ -16,11 +16,21 @@ namespace tensor_operation { namespace device { namespace device_gemm_instance { +using F32 = float; +using F16 = ck::half_t; +using DPtrsGlobal = ck::Tuple; +using Identity = ck::tensor_operation::element_wise::UnaryIdentic; +using Square = ck::tensor_operation::element_wise::UnarySquare; +using DInElementOps = ck::Tuple; +using DOutElementOps = ck::Tuple; + using DeviceGemmReduceNoOpPtr = ck::tensor_operation::device::DeviceGemmReducePtr< + DPtrsGlobal, ck::tensor_operation::element_wise::PassThrough, ck::tensor_operation::element_wise::PassThrough, ck::tensor_operation::element_wise::PassThrough, - ck::tensor_operation::element_wise::UnarySquare>; + DInElementOps, + DOutElementOps>; void add_device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_kn_mn_instances( std::vector&); @@ -112,19 +122,25 @@ bool profile_gemm_reduce_impl(int do_verification, b_k_n.GenerateTensorValue(GeneratorTensor_3{-0.5, 0.5}, num_thread); } - using AElementOp = ck::tensor_operation::element_wise::PassThrough; - using BElementOp = ck::tensor_operation::element_wise::PassThrough; - using CElementOp = ck::tensor_operation::element_wise::PassThrough; - using D0ReduceOp = ck::reduce::Add; - using D1ReduceOp = ck::reduce::Add; - using D1ElementOp = ck::tensor_operation::element_wise::UnarySquare; + using AElementOp = ck::tensor_operation::element_wise::PassThrough; + using BElementOp = ck::tensor_operation::element_wise::PassThrough; + using CElementOp = ck::tensor_operation::element_wise::PassThrough; + using D0ReduceOp = ck::reduce::Add; + using D1ReduceOp = ck::reduce::Add; + using UnaryIdenticElementOp = + ck::tensor_operation::element_wise::UnaryIdentic; + using UnarySquareElementOp = + ck::tensor_operation::element_wise::UnarySquare; + using DxsInElementOps = ck::Tuple; + using DxsOutElementOps = ck::Tuple; - const auto a_element_op = AElementOp{}; - const auto b_element_op = BElementOp{}; - const auto c_element_op = CElementOp{}; - const auto d0_reduce_op = D0ReduceOp{}; - const auto d1_reduce_op = D1ReduceOp{}; - const auto d1_element_op = D1ElementOp{}; + const auto a_element_op = AElementOp{}; + const auto b_element_op = BElementOp{}; + const auto c_element_op = CElementOp{}; + const auto dxs_in_element_op = DxsInElementOps{}; + const auto dxs_out_element_op = DxsOutElementOps{}; + const auto d0_reduce_op = D0ReduceOp{}; + const auto d1_reduce_op = D1ReduceOp{}; if(do_verification) { @@ -149,7 +165,7 @@ bool profile_gemm_reduce_impl(int do_verification, float d0_val = ck::type_convert(c_m_n_host_result(m, n)); float d1_val; - d1_element_op(d1_val, d0_val); + UnarySquareElementOp{}(d1_val, d0_val); d0_reduce_op(d0_acc, d0_val); d1_reduce_op(d1_acc, d1_val); } @@ -165,6 +181,9 @@ bool profile_gemm_reduce_impl(int do_verification, DeviceMem d0_device_buf(sizeof(DDataType) * d0_m_device_result.mDesc.GetElementSpace()); DeviceMem d1_device_buf(sizeof(DDataType) * d1_m_device_result.mDesc.GetElementSpace()); + auto dxs_global = ck::make_tuple(static_cast(d0_device_buf.GetDeviceBuffer()), + static_cast(d1_device_buf.GetDeviceBuffer())); + a_device_buf.ToDevice(a_m_k.mData.data()); b_device_buf.ToDevice(b_k_n.mData.data()); @@ -226,8 +245,7 @@ bool profile_gemm_reduce_impl(int do_verification, gemm_ptr->MakeArgumentPointer(static_cast(a_device_buf.GetDeviceBuffer()), static_cast(b_device_buf.GetDeviceBuffer()), static_cast(c_device_buf.GetDeviceBuffer()), - static_cast(d0_device_buf.GetDeviceBuffer()), - static_cast(d1_device_buf.GetDeviceBuffer()), + dxs_global, M, N, K, @@ -237,7 +255,8 @@ bool profile_gemm_reduce_impl(int do_verification, a_element_op, b_element_op, c_element_op, - d1_element_op); + dxs_in_element_op, + dxs_out_element_op); auto invoker_ptr = gemm_ptr->MakeInvokerPointer();