Skip to content

GitLab

Commit d0b9a467 authored by Chao Liu's avatar Chao Liu
Browse files

Merge remote-tracking branch 'origin/develop' into improve_pipeline

parents c8f6d5d1 8eca05a6
Hide whitespace changes
Inline Side-by-side
Showing with 801 additions and 872 deletions
library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_kn_mn_instance.cpp
View file @ d0b9a467
......@@ -2,7 +2,7 @@
#include "config.hpp"
#include "device_gemm_reduce_xdl_cshuffle.hpp"
#include "element_wise_operation.hpp"
#include "element_wise_reduce_operation.hpp"
#include "reduction_operator.hpp"
#include "device_operation_instance.hpp"
namespace ck {
......@@ -19,9 +19,9 @@ using Col = ck::tensor_layout::gemm::ColumnMajor;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ReduceSum = ck::tensor_operation::element_wise::ReduceSum;
using ReduceSquareSum = ck::tensor_operation::element_wise::ReduceSquareSum;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ReduceSum = ck::reduce::Add<F32>;
using Square = ck::tensor_operation::element_wise::UnarySquare<F32, F32, false>;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
......@@ -30,33 +30,31 @@ static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecializa
// 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| 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|
//###########################| 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, ReduceSquareSum, 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, ReduceSquareSum, 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, ReduceSquareSum, 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, ReduceSquareSum, 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, ReduceSquareSum, 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, ReduceSquareSum, 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, ReduceSquareSum, 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, ReduceSquareSum, 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, ReduceSquareSum, 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, ReduceSquareSum, 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, ReduceSquareSum, 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, ReduceSquareSum, 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, ReduceSquareSum, 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, ReduceSquareSum, 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, ReduceSquareSum, 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, ReduceSquareSum, 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>
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>
// clang-format on
>;
void add_device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_kn_mn_instances(
std::vector<
DeviceGemmReducePtr<PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSquareSum>>&
instances)
std::vector<DeviceGemmReducePtr<PassThrough, PassThrough, PassThrough, Square>>& instances)
{
add_device_operation_instances(
instances, device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_kn_mn_instances{});
......
library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_nk_mn_instance.cpp
View file @ d0b9a467
......@@ -2,7 +2,7 @@
#include "config.hpp"
#include "device_gemm_reduce_xdl_cshuffle.hpp"
#include "element_wise_operation.hpp"
#include "element_wise_reduce_operation.hpp"
#include "reduction_operator.hpp"
#include "device_operation_instance.hpp"
namespace ck {
......@@ -19,9 +19,9 @@ using Col = ck::tensor_layout::gemm::ColumnMajor;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ReduceSum = ck::tensor_operation::element_wise::ReduceSum;
using ReduceSquareSum = ck::tensor_operation::element_wise::ReduceSquareSum;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ReduceSum = ck::reduce::Add<F32>;
using Square = ck::tensor_operation::element_wise::UnarySquare<F32, F32, false>;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
......@@ -30,33 +30,31 @@ static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecializa
// 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| 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|
//###########################| 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, ReduceSquareSum, 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, ReduceSquareSum, 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, ReduceSquareSum, 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, ReduceSquareSum, 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, ReduceSquareSum, 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, ReduceSquareSum, 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, ReduceSquareSum, 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, ReduceSquareSum, 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, ReduceSquareSum, 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, ReduceSquareSum, 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, ReduceSquareSum, 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, ReduceSquareSum, 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, ReduceSquareSum, 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, ReduceSquareSum, 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, ReduceSquareSum, 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, ReduceSquareSum, 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>
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>
// clang-format on
>;
void add_device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_nk_mn_instances(
std::vector<
DeviceGemmReducePtr<PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSquareSum>>&
instances)
std::vector<DeviceGemmReducePtr<PassThrough, PassThrough, PassThrough, Square>>& instances)
{
add_device_operation_instances(
instances, device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_nk_mn_instances{});
......
library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_kn_mn_instance.cpp
View file @ d0b9a467
......@@ -2,7 +2,7 @@
#include "config.hpp"
#include "device_gemm_reduce_xdl_cshuffle.hpp"
#include "element_wise_operation.hpp"
#include "element_wise_reduce_operation.hpp"
#include "reduction_operator.hpp"
#include "device_operation_instance.hpp"
namespace ck {
......@@ -19,9 +19,9 @@ using Col = ck::tensor_layout::gemm::ColumnMajor;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ReduceSum = ck::tensor_operation::element_wise::ReduceSum;
using ReduceSquareSum = ck::tensor_operation::element_wise::ReduceSquareSum;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ReduceSum = ck::reduce::Add<F32>;
using Square = ck::tensor_operation::element_wise::UnarySquare<F32, F32, false>;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
......@@ -30,33 +30,31 @@ static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecializa
// 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| 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|
//###########################| 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, ReduceSquareSum, 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, ReduceSquareSum, 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, ReduceSquareSum, 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, ReduceSquareSum, 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, ReduceSquareSum, 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, ReduceSquareSum, 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, ReduceSquareSum, 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, ReduceSquareSum, 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, ReduceSquareSum, 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, ReduceSquareSum, 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, ReduceSquareSum, 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, ReduceSquareSum, 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, ReduceSquareSum, 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, ReduceSquareSum, 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, ReduceSquareSum, 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, ReduceSquareSum, 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>
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>
// clang-format on
>;
void add_device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_kn_mn_instances(
std::vector<
DeviceGemmReducePtr<PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSquareSum>>&
instances)
std::vector<DeviceGemmReducePtr<PassThrough, PassThrough, PassThrough, Square>>& instances)
{
add_device_operation_instances(
instances, device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_kn_mn_instances{});
......
library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_nk_mn_instance.cpp
View file @ d0b9a467
......@@ -2,7 +2,7 @@
#include "config.hpp"
#include "device_gemm_reduce_xdl_cshuffle.hpp"
#include "element_wise_operation.hpp"
#include "element_wise_reduce_operation.hpp"
#include "reduction_operator.hpp"
#include "device_operation_instance.hpp"
namespace ck {
......@@ -19,9 +19,9 @@ using Col = ck::tensor_layout::gemm::ColumnMajor;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ReduceSum = ck::tensor_operation::element_wise::ReduceSum;
using ReduceSquareSum = ck::tensor_operation::element_wise::ReduceSquareSum;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ReduceSum = ck::reduce::Add<F32>;
using Square = ck::tensor_operation::element_wise::UnarySquare<F32, F32, false>;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
......@@ -30,30 +30,28 @@ static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecializa
// 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| 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|
//###########################| 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, ReduceSquareSum, 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, ReduceSquareSum, 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, ReduceSquareSum, 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, ReduceSquareSum, 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, ReduceSquareSum, 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, ReduceSquareSum, 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, ReduceSquareSum, 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, ReduceSquareSum, 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, ReduceSquareSum, 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, ReduceSquareSum, 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, ReduceSquareSum, 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, ReduceSquareSum, 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, ReduceSquareSum, 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>
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>
// clang-format on
>;
void add_device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_nk_mn_instances(
std::vector<
DeviceGemmReducePtr<PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSquareSum>>&
instances)
std::vector<DeviceGemmReducePtr<PassThrough, PassThrough, PassThrough, Square>>& instances)
{
add_device_operation_instances(
instances, device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_nk_mn_instances{});
......
profiler/include/profile_batched_gemm_reduce_impl.hpp
View file @ d0b9a467
......@@ -8,7 +8,7 @@
#include "tensor_layout.hpp"
#include "device_tensor.hpp"
#include "element_wise_operation.hpp"
#include "element_wise_reduce_operation.hpp"
#include "reduction_operator.hpp"
#include "device_gemm_reduce.hpp"
#include "reference_batched_gemm.hpp"
......@@ -21,8 +21,7 @@ using DeviceGemmReduceNoOpPtr = ck::tensor_operation::device::DeviceGemmReducePt
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::ReduceSum,
ck::tensor_operation::element_wise::ReduceSquareSum>;
ck::tensor_operation::element_wise::UnarySquare<float, float, false>>;
void add_device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gkn_gmn_instances(
std::vector<DeviceGemmReduceNoOpPtr>&);
......@@ -120,17 +119,19 @@ bool profile_batched_gemm_reduce_impl(int do_verification,
b_g_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-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::tensor_operation::element_wise::ReduceSum;
using D1ReduceOp = ck::tensor_operation::element_wise::ReduceSquareSum;
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<float>;
using D1ReduceOp = ck::reduce::Add<float>;
using D1ElementOp = ck::tensor_operation::element_wise::UnarySquare<float, float, false>;
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 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{};
if(do_verification)
{
......@@ -154,17 +155,21 @@ bool profile_batched_gemm_reduce_impl(int do_verification,
{
for(int m = 0; m < M; ++m)
{
float d0_acc = d0_reduce_op.GetReduceZeroValue();
float d1_acc = d1_reduce_op.GetReduceZeroValue();
float d0_acc = d0_reduce_op.GetReductionZeroVal();
float d1_acc = d1_reduce_op.GetReductionZeroVal();
for(int n = 0; n < N; ++n)
{
d0_reduce_op.Reduce(d0_acc, c_g_m_n_host_result(batch, m, n));
d1_reduce_op.Reduce(d1_acc, c_g_m_n_host_result(batch, m, n));
float d0_val = ck::type_convert<float>(c_g_m_n_host_result(batch, m, n));
float d1_val;
d1_element_op(d1_val, d0_val);
d0_reduce_op(d0_acc, d0_val);
d1_reduce_op(d1_acc, d1_val);
}
d0_g_m_host_result(batch, m) = d0_acc;
d1_g_m_host_result(batch, m) = d1_acc;
d0_g_m_host_result(batch, m) = ck::type_convert<DDataType>(d0_acc);
d1_g_m_host_result(batch, m) = ck::type_convert<DDataType>(d1_acc);
}
}
}
......@@ -247,8 +252,7 @@ bool profile_batched_gemm_reduce_impl(int do_verification,
a_element_op,
b_element_op,
c_element_op,
d0_reduce_op,
d1_reduce_op,
d1_element_op,
BatchCount);
auto invoker_ptr = gemm_ptr->MakeInvokerPointer();
......
profiler/include/profile_gemm_reduce_impl.hpp
View file @ d0b9a467
......@@ -7,7 +7,7 @@
#include "tensor_layout.hpp"
#include "device_tensor.hpp"
#include "element_wise_operation.hpp"
#include "element_wise_reduce_operation.hpp"
#include "reduction_operator.hpp"
#include "device_gemm_reduce.hpp"
#include "reference_gemm.hpp"
......@@ -20,8 +20,7 @@ using DeviceGemmReduceNoOpPtr = ck::tensor_operation::device::DeviceGemmReducePt
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::ReduceSum,
ck::tensor_operation::element_wise::ReduceSquareSum>;
ck::tensor_operation::element_wise::UnarySquare<float, float, false>>;
void add_device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_kn_mn_instances(
std::vector<DeviceGemmReduceNoOpPtr>&);
......@@ -113,17 +112,19 @@ bool profile_gemm_reduce_impl(int do_verification,
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-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::tensor_operation::element_wise::ReduceSum;
using D1ReduceOp = ck::tensor_operation::element_wise::ReduceSquareSum;
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<float>;
using D1ReduceOp = ck::reduce::Add<float>;
using D1ElementOp = ck::tensor_operation::element_wise::UnarySquare<float, float, false>;
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 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{};
if(do_verification)
{
......@@ -140,17 +141,21 @@ bool profile_gemm_reduce_impl(int do_verification,
for(int m = 0; m < M; ++m)
{
float d0_acc = d0_reduce_op.GetReduceZeroValue();
float d1_acc = d1_reduce_op.GetReduceZeroValue();
float d0_acc = d0_reduce_op.GetReductionZeroVal();
float d1_acc = d1_reduce_op.GetReductionZeroVal();
for(int n = 0; n < N; ++n)
{
d0_reduce_op.Reduce(d0_acc, c_m_n_host_result(m, n));
d1_reduce_op.Reduce(d1_acc, c_m_n_host_result(m, n));
float d0_val = ck::type_convert<float>(c_m_n_host_result(m, n));
float d1_val;
d1_element_op(d1_val, d0_val);
d0_reduce_op(d0_acc, d0_val);
d1_reduce_op(d1_acc, d1_val);
}
d0_m_host_result(m) = d0_acc;
d1_m_host_result(m) = d1_acc;
d0_m_host_result(m) = ck::type_convert<DDataType>(d0_acc);
d1_m_host_result(m) = ck::type_convert<DDataType>(d1_acc);
}
}
......@@ -232,8 +237,7 @@ bool profile_gemm_reduce_impl(int do_verification,
a_element_op,
b_element_op,
c_element_op,
d0_reduce_op,
d1_reduce_op);
d1_element_op);
auto invoker_ptr = gemm_ptr->MakeInvokerPointer();
......
script/clang-format-overwrite.sh
View file @ d0b9a467
find . -name deps -prune -o -name build -prune -o -iname '*.h' -o -iname '*.hpp' -o -iname '*.cpp' -o -iname '*.h.in' -o -iname '*.hpp.in' -o -iname '*.cpp.in' -o -iname '*.cl' -o -iname '*.cuh' -o -iname '*.cu' | xargs -n 1 -P 16 -I{} -t sh -c 'clang-format-10 -i -style=file {}'
#find . -name deps -prune -o -name build -prune -o -iname '*.h' -o -iname '*.hpp' -o -iname '*.cpp' -o -iname '*.h.in' -o -iname '*.hpp.in' -o -iname '*.cpp.in' -o -iname '*.cl' -o -iname '*.cuh' -o -iname '*.cu' | xargs -n 1 -P 16 -I{} -t sh -c 'clang-format-10 -i -style=file {}'
git status --porcelain | awk '$1 != "D" && (match($2, "\\.cpp|hpp")) {print $2}' | xargs -n 1 -P 16 -I{} -t sh -c 'clang-format-10 -i -style=file {}'
test/CMakeLists.txt
View file @ d0b9a467
......@@ -24,6 +24,7 @@ include_directories(BEFORE
add_custom_target(check COMMAND ${CMAKE_CTEST_COMMAND} --output-on-failure -C ${CMAKE_CFG_INTDIR})
add_custom_target(tests)
function(add_test_executable TEST_NAME)
message("adding test ${TEST_NAME}")
add_executable(${TEST_NAME} ${ARGN})
......@@ -32,6 +33,20 @@ function(add_test_executable TEST_NAME)
add_dependencies(check ${TEST_NAME})
endfunction(add_test_executable TEST_NAME)
include(GoogleTest)
function(add_gtest_executable TEST_NAME)
message("adding gtest ${TEST_NAME}")
add_executable(${TEST_NAME} ${ARGN})
add_dependencies(tests ${TEST_NAME})
add_dependencies(check ${TEST_NAME})
# suppress gtest warnings
target_compile_options(${TEST_NAME} PRIVATE -Wno-global-constructors)
target_link_libraries(${TEST_NAME} PRIVATE gtest_main)
gtest_discover_tests(${TEST_NAME})
endfunction(add_gtest_executable TEST_NAME)
add_subdirectory(magic_number_division)
add_subdirectory(space_filling_curve)
add_subdirectory(conv_util)
......
test/conv_util/CMakeLists.txt
View file @ d0b9a467
add_test_executable(test_conv_util conv_util.cpp)
add_gtest_executable(test_conv_util conv_util.cpp)
target_link_libraries(test_conv_util PRIVATE host_tensor conv_fwd_util)
test/conv_util/conv_util.cpp
View file @ d0b9a467
#include <iostream>
#include <string>
#include <vector>
#include "gtest/gtest.h"
#include "config.hpp"
#include "conv_fwd_util.hpp"
......@@ -9,196 +10,194 @@
namespace {
bool test_conv_params_get_output_spatial_lengths()
class TestConvUtil : public ::testing::Test
{
bool res{true};
// -------------------------- default 2D ------------------------------------
public:
void SetNDParams(std::size_t ndims)
{
conv_params.num_dim_spatial = ndims;
conv_params.filter_spatial_lengths = std::vector<ck::index_t>(ndims, 3);
conv_params.input_spatial_lengths = std::vector<ck::index_t>(ndims, 71);
conv_params.conv_filter_strides = std::vector<ck::index_t>(ndims, 2);
conv_params.conv_filter_dilations = std::vector<ck::index_t>(ndims, 1);
conv_params.input_left_pads = std::vector<ck::index_t>(ndims, 1);
conv_params.input_right_pads = std::vector<ck::index_t>(ndims, 1);
}
protected:
// ------- default 2D -------
// input NCHW {128,192,71,71},
// weights KCYX {256,192,3,3},
// stride {2,2},
// dilations {1,1},
// padding {{1,1}, {1,1}}
ck::utils::conv::ConvParams conv_params;
};
} // namespace
TEST_F(TestConvUtil, ConvParamsGetOutputSpatialLengths2D)
{
ck::utils::conv::ConvParams conv_params;
std::vector<ck::index_t> out_spatial_len = conv_params.GetOutputSpatialLengths();
res = ck::utils::check_err(out_spatial_len,
std::vector<ck::index_t>{36, 36},
"Error: ConvParams 2D default constructor.");
EXPECT_TRUE(ck::utils::check_err(out_spatial_len,
std::vector<ck::index_t>{36, 36},
"Error: ConvParams 2D default constructor."));
conv_params.conv_filter_strides = std::vector<ck::index_t>{1, 1};
out_spatial_len = conv_params.GetOutputSpatialLengths();
res = ck::utils::check_err(
out_spatial_len, std::vector<ck::index_t>{71, 71}, "Error: ConvParams 2D stride {1,1}.");
EXPECT_TRUE(ck::utils::check_err(
out_spatial_len, std::vector<ck::index_t>{71, 71}, "Error: ConvParams 2D stride {1,1}."));
conv_params.conv_filter_strides = std::vector<ck::index_t>{2, 2};
conv_params.input_left_pads = std::vector<ck::index_t>{2, 2};
conv_params.input_right_pads = std::vector<ck::index_t>{2, 2};
out_spatial_len = conv_params.GetOutputSpatialLengths();
res = ck::utils::check_err(out_spatial_len,
std::vector<ck::index_t>{37, 37},
"Error: ConvParams 2D padding left/right {2,2}.");
EXPECT_TRUE(ck::utils::check_err(out_spatial_len,
std::vector<ck::index_t>{37, 37},
"Error: ConvParams 2D padding left/right {2,2}."));
conv_params.conv_filter_dilations = std::vector<ck::index_t>{2, 2};
out_spatial_len = conv_params.GetOutputSpatialLengths();
res = ck::utils::check_err(
out_spatial_len, std::vector<ck::index_t>{36, 36}, "Error: ConvParams 2D dilation {2,2}.");
EXPECT_TRUE(ck::utils::check_err(
out_spatial_len, std::vector<ck::index_t>{36, 36}, "Error: ConvParams 2D dilation {2,2}."));
conv_params.conv_filter_strides = std::vector<ck::index_t>{3, 3};
conv_params.input_left_pads = std::vector<ck::index_t>{1, 1};
conv_params.input_right_pads = std::vector<ck::index_t>{1, 1};
conv_params.conv_filter_dilations = std::vector<ck::index_t>{2, 2};
out_spatial_len = conv_params.GetOutputSpatialLengths();
res =
EXPECT_TRUE(
ck::utils::check_err(out_spatial_len,
std::vector<ck::index_t>{23, 23},
"Error: ConvParams 2D strides{3,3}, padding {1,1}, dilations {2,2}.");
"Error: ConvParams 2D strides{3,3}, padding {1,1}, dilations {2,2}."));
}
// -------------------------- 1D ------------------------------------
conv_params.num_dim_spatial = 1;
conv_params.filter_spatial_lengths = std::vector<ck::index_t>{3};
conv_params.input_spatial_lengths = std::vector<ck::index_t>{71};
conv_params.conv_filter_strides = std::vector<ck::index_t>{2};
conv_params.conv_filter_dilations = std::vector<ck::index_t>{1};
conv_params.input_left_pads = std::vector<ck::index_t>{1};
conv_params.input_right_pads = std::vector<ck::index_t>{1};
TEST_F(TestConvUtil, ConvParamsGetOutputSpatialLengths1D)
{
SetNDParams(1);
out_spatial_len = conv_params.GetOutputSpatialLengths();
res = ck::utils::check_err(
out_spatial_len, std::vector<ck::index_t>{36}, "Error: ConvParams 1D.");
std::vector<ck::index_t> out_spatial_len = conv_params.GetOutputSpatialLengths();
EXPECT_TRUE(ck::utils::check_err(
out_spatial_len, std::vector<ck::index_t>{36}, "Error: ConvParams 1D."));
conv_params.conv_filter_strides = std::vector<ck::index_t>{1};
out_spatial_len = conv_params.GetOutputSpatialLengths();
res = ck::utils::check_err(
out_spatial_len, std::vector<ck::index_t>{71}, "Error: ConvParams 1D stride {1}.");
EXPECT_TRUE(ck::utils::check_err(
out_spatial_len, std::vector<ck::index_t>{71}, "Error: ConvParams 1D stride {1}."));
conv_params.conv_filter_strides = std::vector<ck::index_t>{2};
conv_params.input_left_pads = std::vector<ck::index_t>{2};
conv_params.input_right_pads = std::vector<ck::index_t>{2};
out_spatial_len = conv_params.GetOutputSpatialLengths();
res = ck::utils::check_err(out_spatial_len,
std::vector<ck::index_t>{37},
"Error: ConvParams 1D padding left/right {2}.");
EXPECT_TRUE(ck::utils::check_err(out_spatial_len,
std::vector<ck::index_t>{37},
"Error: ConvParams 1D padding left/right {2}."));
conv_params.conv_filter_dilations = std::vector<ck::index_t>{2};
out_spatial_len = conv_params.GetOutputSpatialLengths();
res = ck::utils::check_err(
out_spatial_len, std::vector<ck::index_t>{36}, "Error: ConvParams 1D dilation {2}.");
EXPECT_TRUE(ck::utils::check_err(
out_spatial_len, std::vector<ck::index_t>{36}, "Error: ConvParams 1D dilation {2}."));
conv_params.conv_filter_strides = std::vector<ck::index_t>{3};
conv_params.input_left_pads = std::vector<ck::index_t>{1};
conv_params.input_right_pads = std::vector<ck::index_t>{1};
conv_params.conv_filter_dilations = std::vector<ck::index_t>{2};
out_spatial_len = conv_params.GetOutputSpatialLengths();
res = ck::utils::check_err(out_spatial_len,
std::vector<ck::index_t>{23},
"Error: ConvParams 1D strides{3}, padding {1}, dilations {2}.");
// -------------------------- 3D ------------------------------------
conv_params.num_dim_spatial = 3;
conv_params.filter_spatial_lengths = std::vector<ck::index_t>{3, 3, 3};
conv_params.input_spatial_lengths = std::vector<ck::index_t>{71, 71, 71};
conv_params.conv_filter_strides = std::vector<ck::index_t>{2, 2, 2};
conv_params.conv_filter_dilations = std::vector<ck::index_t>{1, 1, 1};
conv_params.input_left_pads = std::vector<ck::index_t>{1, 1, 1};
conv_params.input_right_pads = std::vector<ck::index_t>{1, 1, 1};
out_spatial_len = conv_params.GetOutputSpatialLengths();
res = ck::utils::check_err(
out_spatial_len, std::vector<ck::index_t>{36, 36, 36}, "Error: ConvParams 3D.");
EXPECT_TRUE(
ck::utils::check_err(out_spatial_len,
std::vector<ck::index_t>{23},
"Error: ConvParams 1D strides{3}, padding {1}, dilations {2}."));
}
TEST_F(TestConvUtil, ConvParamsGetOutputSpatialLengths3D)
{
SetNDParams(3);
std::vector<ck::index_t> out_spatial_len = conv_params.GetOutputSpatialLengths();
EXPECT_TRUE(ck::utils::check_err(
out_spatial_len, std::vector<ck::index_t>{36, 36, 36}, "Error: ConvParams 3D."));
conv_params.conv_filter_strides = std::vector<ck::index_t>{1, 1, 1};
out_spatial_len = conv_params.GetOutputSpatialLengths();
res = ck::utils::check_err(out_spatial_len,
std::vector<ck::index_t>{71, 71, 71},
"Error: ConvParams 3D stride {1, 1, 1}.");
EXPECT_TRUE(ck::utils::check_err(out_spatial_len,
std::vector<ck::index_t>{71, 71, 71},
"Error: ConvParams 3D stride {1, 1, 1}."));
conv_params.conv_filter_strides = std::vector<ck::index_t>{2, 2, 2};
conv_params.input_left_pads = std::vector<ck::index_t>{2, 2, 2};
conv_params.input_right_pads = std::vector<ck::index_t>{2, 2, 2};
out_spatial_len = conv_params.GetOutputSpatialLengths();
res = ck::utils::check_err(out_spatial_len,
std::vector<ck::index_t>{37, 37, 37},
"Error: ConvParams 3D padding left/right {2, 2, 2}.");
EXPECT_TRUE(ck::utils::check_err(out_spatial_len,
std::vector<ck::index_t>{37, 37, 37},
"Error: ConvParams 3D padding left/right {2, 2, 2}."));
conv_params.conv_filter_dilations = std::vector<ck::index_t>{2, 2, 2};
out_spatial_len = conv_params.GetOutputSpatialLengths();
res = ck::utils::check_err(out_spatial_len,
std::vector<ck::index_t>{36, 36, 36},
"Error: ConvParams 3D dilation {2, 2, 2}.");
EXPECT_TRUE(ck::utils::check_err(out_spatial_len,
std::vector<ck::index_t>{36, 36, 36},
"Error: ConvParams 3D dilation {2, 2, 2}."));
conv_params.conv_filter_strides = std::vector<ck::index_t>{3, 3, 3};
conv_params.input_left_pads = std::vector<ck::index_t>{1, 1, 1};
conv_params.input_right_pads = std::vector<ck::index_t>{1, 1, 1};
conv_params.conv_filter_dilations = std::vector<ck::index_t>{2, 2, 2};
out_spatial_len = conv_params.GetOutputSpatialLengths();
res = ck::utils::check_err(
EXPECT_TRUE(ck::utils::check_err(
out_spatial_len,
std::vector<ck::index_t>{23, 23, 23},
"Error: ConvParams 3D strides{3, 3, 3}, padding {1, 1, 1}, dilations {2, 2, 2}.");
return res;
"Error: ConvParams 3D strides{3, 3, 3}, padding {1, 1, 1}, dilations {2, 2, 2}."));
}
bool test_get_host_tensor_descriptor()
TEST(ConvUtil, GetHostTensorDescriptor)
{
bool res{true};
namespace tl = ck::tensor_layout::convolution;
std::vector<std::size_t> dims{2, 3, 4, 5};
HostTensorDescriptor h = ck::utils::conv::get_host_tensor_descriptor(dims, tl::NHWC{});
res =
ck::utils::check_err(h.GetLengths(), {2, 3, 4, 5}, "Error: wrong NHWC dimensions lengths!");
res = ck::utils::check_err(
h.GetStrides(), {3 * 4 * 5, 1, 3 * 5, 3}, "Error: wrong NHWC dimensions strides!");
EXPECT_TRUE(ck::utils::check_err(
h.GetLengths(), {2, 3, 4, 5}, "Error: wrong NHWC dimensions lengths!"));
EXPECT_TRUE(ck::utils::check_err(
h.GetStrides(), {3 * 4 * 5, 1, 3 * 5, 3}, "Error: wrong NHWC dimensions strides!"));
h = ck::utils::conv::get_host_tensor_descriptor(dims, tl::NCHW{});
res =
ck::utils::check_err(h.GetLengths(), {2, 3, 4, 5}, "Error: wrong NCHW dimensions lengths!");
res = ck::utils::check_err(
h.GetStrides(), {3 * 4 * 5, 4 * 5, 5, 1}, "Error: wrong NCHW dimensions strides!");
EXPECT_TRUE(ck::utils::check_err(
h.GetLengths(), {2, 3, 4, 5}, "Error: wrong NCHW dimensions lengths!"));
EXPECT_TRUE(ck::utils::check_err(
h.GetStrides(), {3 * 4 * 5, 4 * 5, 5, 1}, "Error: wrong NCHW dimensions strides!"));
dims = std::vector<std::size_t>{2, 3, 4};
h = ck::utils::conv::get_host_tensor_descriptor(dims, tl::NWC{});
res = ck::utils::check_err(h.GetLengths(), {2, 3, 4}, "Error: wrong NWC dimensions lengths!");
res =
ck::utils::check_err(h.GetStrides(), {3 * 4, 1, 3}, "Error: wrong NWC dimensions strides!");
EXPECT_TRUE(
ck::utils::check_err(h.GetLengths(), {2, 3, 4}, "Error: wrong NWC dimensions lengths!"));
EXPECT_TRUE(ck::utils::check_err(
h.GetStrides(), {3 * 4, 1, 3}, "Error: wrong NWC dimensions strides!"));
h = ck::utils::conv::get_host_tensor_descriptor(dims, tl::NCW{});
res = ck::utils::check_err(h.GetLengths(), {2, 3, 4}, "Error: wrong NCW dimensions lengths!");
res =
ck::utils::check_err(h.GetStrides(), {3 * 4, 4, 1}, "Error: wrong NCW dimensions strides!");
h = ck::utils::conv::get_host_tensor_descriptor(dims, tl::NCW{});
EXPECT_TRUE(
ck::utils::check_err(h.GetLengths(), {2, 3, 4}, "Error: wrong NCW dimensions lengths!"));
EXPECT_TRUE(ck::utils::check_err(
h.GetStrides(), {3 * 4, 4, 1}, "Error: wrong NCW dimensions strides!"));
dims = std::vector<std::size_t>{2, 3, 4, 5, 6};
h = ck::utils::conv::get_host_tensor_descriptor(dims, tl::NDHWC{});
res = ck::utils::check_err(h.GetLengths(), dims, "Error: wrong NDHWC dimensions lengths!");
res = ck::utils::check_err(h.GetStrides(),
{3 * 4 * 5 * 6, // N
1, // C
3 * 5 * 6, // D
3 * 6, // H
3}, // W
"Error: wrong NDHWC dimensions strides!");
h = ck::utils::conv::get_host_tensor_descriptor(dims, tl::NCDHW{});
res = ck::utils::check_err(h.GetLengths(), dims, "Error: wrong NCDHW dimensions lengths!");
res = ck::utils::check_err(h.GetStrides(),
{3 * 4 * 5 * 6, // N
4 * 5 * 6, // C
5 * 6, // D
6, // H
1}, // W
"Error: wrong NCDHW dimensions strides!");
return res;
}
} // namespace
int main(void)
{
bool res = test_conv_params_get_output_spatial_lengths();
std::cout << "test_conv_params_get_output_spatial_lengths ..... "
<< (res ? "SUCCESS" : "FAILURE") << std::endl;
res = test_get_host_tensor_descriptor();
std::cout << "test_get_host_tensor_descriptor ..... " << (res ? "SUCCESS" : "FAILURE")
<< std::endl;
return res ? 0 : 1;
EXPECT_TRUE(
ck::utils::check_err(h.GetLengths(), dims, "Error: wrong NDHWC dimensions lengths!"));
EXPECT_TRUE(ck::utils::check_err(h.GetStrides(),
{3 * 4 * 5 * 6, // N
1, // C
3 * 5 * 6, // D
3 * 6, // H
3}, // W
"Error: wrong NDHWC dimensions strides!"));
h = ck::utils::conv::get_host_tensor_descriptor(dims, tl::NCDHW{});
EXPECT_TRUE(
ck::utils::check_err(h.GetLengths(), dims, "Error: wrong NCDHW dimensions lengths!"));
EXPECT_TRUE(ck::utils::check_err(h.GetStrides(),
{3 * 4 * 5 * 6, // N
4 * 5 * 6, // C
5 * 6, // D
6, // H
1}, // W
"Error: wrong NCDHW dimensions strides!"));
}
test/convnd_fwd/CMakeLists.txt
View file @ d0b9a467
add_custom_target(test_convnd_fwd)
add_test_executable(test_conv1d_fwd conv1d_fwd.cpp)
add_gtest_executable(test_conv1d_fwd conv1d_fwd.cpp)
target_link_libraries(test_conv1d_fwd PRIVATE host_tensor device_conv1d_fwd_instance conv_fwd_util)
target_link_libraries(test_conv1d_fwd PRIVATE )
add_dependencies(test_convnd_fwd test_conv1d_fwd)
add_test_executable(test_conv2d_fwd conv2d_fwd.cpp)
add_gtest_executable(test_conv2d_fwd conv2d_fwd.cpp)
target_link_libraries(test_conv2d_fwd PRIVATE host_tensor device_conv2d_fwd_instance conv_fwd_util)
add_dependencies(test_convnd_fwd test_conv2d_fwd)
add_test_executable(test_conv3d_fwd conv3d_fwd.cpp)
add_gtest_executable(test_conv3d_fwd conv3d_fwd.cpp)
target_link_libraries(test_conv3d_fwd PRIVATE host_tensor device_conv3d_fwd_instance conv_fwd_util)
add_dependencies(test_convnd_fwd test_conv3d_fwd)
test/convnd_fwd/conv1d_fwd.cpp
View file @ d0b9a467
......@@ -2,6 +2,7 @@
#include <stdexcept>
#include <tuple>
#include <vector>
#include "gtest/gtest.h"
#include "data_type.hpp"
#include "element_wise_operation.hpp"
......@@ -10,7 +11,8 @@
namespace {
bool test_conv1D_nwc()
template <typename T>
bool test_conv1d_nwc_instances(const std::vector<test::conv::DeviceConvFwdNoOpPtr>& conv_ptrs)
{
using namespace std::placeholders;
using namespace ck::utils;
......@@ -18,31 +20,24 @@ bool test_conv1D_nwc()
ck::utils::conv::ConvParams params;
params.num_dim_spatial = 1;
params.N = 2;
params.K = 16;
params.C = 4;
params.filter_spatial_lengths = std::vector<ck::index_t>{3};
params.input_spatial_lengths = std::vector<ck::index_t>{16};
params.conv_filter_strides = std::vector<ck::index_t>{1};
params.input_spatial_lengths = std::vector<ck::index_t>{71};
params.conv_filter_strides = std::vector<ck::index_t>{2};
params.conv_filter_dilations = std::vector<ck::index_t>{1};
params.input_left_pads = std::vector<ck::index_t>{1};
params.input_right_pads = std::vector<ck::index_t>{1};
std::vector<test::conv::DeviceConvFwdNoOpPtr> conv_ptrs;
test::conv::get_test_convolution_fwd_instance<1>(conv_ptrs);
conv::ConvFwdOpInstance<float, float, float, ctl::NWC, ctl::KCX, ctl::NWK> conv_instance(
params);
conv::ConvFwdOpInstance<T, T, T, ctl::NWC, ctl::KCX, ctl::NWK> conv_instance(params);
auto reference_conv_fwd_fun = std::bind(
conv::run_reference_convolution_forward<1, float, float, float>, params, _1, _2, _3);
OpInstanceRunEngine<float, float, float> run_engine(conv_instance, reference_conv_fwd_fun);
run_engine.SetAtol(1e-5);
run_engine.SetRtol(1e-4);
auto reference_conv_fwd_fun =
std::bind(conv::run_reference_convolution_forward<1, T, T, T>, params, _1, _2, _3);
OpInstanceRunEngine<T, T, T> run_engine(conv_instance, reference_conv_fwd_fun);
return run_engine.Test(conv_ptrs);
}
template <typename T>
bool test_conv1d_nwc_instances(const std::vector<test::conv::DeviceConvFwdNoOpPtr>& conv_ptrs)
} // anonymous namespace
TEST(Conv1DFwdNWC, TestConv1D)
{
using namespace std::placeholders;
using namespace ck::utils;
......@@ -50,65 +45,49 @@ bool test_conv1d_nwc_instances(const std::vector<test::conv::DeviceConvFwdNoOpPt
ck::utils::conv::ConvParams params;
params.num_dim_spatial = 1;
params.N = 2;
params.K = 16;
params.C = 4;
params.filter_spatial_lengths = std::vector<ck::index_t>{3};
params.input_spatial_lengths = std::vector<ck::index_t>{71};
params.conv_filter_strides = std::vector<ck::index_t>{2};
params.input_spatial_lengths = std::vector<ck::index_t>{16};
params.conv_filter_strides = std::vector<ck::index_t>{1};
params.conv_filter_dilations = std::vector<ck::index_t>{1};
params.input_left_pads = std::vector<ck::index_t>{1};
params.input_right_pads = std::vector<ck::index_t>{1};
conv::ConvFwdOpInstance<T, T, T, ctl::NWC, ctl::KCX, ctl::NWK> conv_instance(params);
auto reference_conv_fwd_fun =
std::bind(conv::run_reference_convolution_forward<1, T, T, T>, params, _1, _2, _3);
OpInstanceRunEngine<T, T, T> run_engine(conv_instance, reference_conv_fwd_fun);
return run_engine.Test(conv_ptrs);
}
std::vector<test::conv::DeviceConvFwdNoOpPtr> conv_ptrs;
test::conv::get_test_convolution_fwd_instance<1>(conv_ptrs);
conv::ConvFwdOpInstance<float, float, float, ctl::NWC, ctl::KCX, ctl::NWK> conv_instance(
params);
bool test_conv1d_nwc_bf16_instances()
{
return test_conv1d_nwc_instances<ck::bhalf_t>(
ck::utils::conv::ConvolutionFwdInstances<ck::bhalf_t, ck::bhalf_t, ck::bhalf_t>::Get<1>());
auto reference_conv_fwd_fun = std::bind(
conv::run_reference_convolution_forward<1, float, float, float>, params, _1, _2, _3);
OpInstanceRunEngine<float, float, float> run_engine(conv_instance, reference_conv_fwd_fun);
run_engine.SetAtol(1e-5);
run_engine.SetRtol(1e-4);
EXPECT_TRUE(run_engine.Test(conv_ptrs));
}
bool test_conv1d_nwc_f16_instances()
TEST(Conv1DFwdNWC, Bf16Iinstances)
{
return test_conv1d_nwc_instances<ck::half_t>(
ck::utils::conv::ConvolutionFwdInstances<ck::half_t, ck::half_t, ck::half_t>::Get<1>());
EXPECT_TRUE(test_conv1d_nwc_instances<ck::bhalf_t>(
ck::utils::conv::ConvolutionFwdInstances<ck::bhalf_t, ck::bhalf_t, ck::bhalf_t>::Get<1>()));
}
bool test_conv1d_nwc_f32_instances()
TEST(Conv1DFwdNWC, F16Instances)
{
return test_conv1d_nwc_instances<float>(
ck::utils::conv::ConvolutionFwdInstances<float, float, float>::Get<1>());
EXPECT_TRUE(test_conv1d_nwc_instances<ck::half_t>(
ck::utils::conv::ConvolutionFwdInstances<ck::half_t, ck::half_t, ck::half_t>::Get<1>()));
}
bool test_conv1d_nwc_int8_instances()
TEST(Conv1DFwdNWC, F32Instances)
{
return test_conv1d_nwc_instances<int8_t>(
ck::utils::conv::ConvolutionFwdInstances<int8_t, int8_t, int8_t>::Get<1>());
EXPECT_TRUE(test_conv1d_nwc_instances<float>(
ck::utils::conv::ConvolutionFwdInstances<float, float, float>::Get<1>()));
}
} // anonymous namespace
int main()
TEST(Conv1DFwdNWC, Int8Instances)
{
bool res{true};
res = test_conv1D_nwc();
std::cout << "test_conv1D_nwc ..... " << (res ? "SUCCESS" : "FAILURE") << std::endl;
res = test_conv1d_nwc_bf16_instances();
std::cout << "\nTestConv1DNWCBF16Instances ..... " << (res ? "SUCCESS" : "FAILURE")
<< std::endl;
res = test_conv1d_nwc_f16_instances();
std::cout << "\ntest_conv1d_nwc_f16_instances ..... " << (res ? "SUCCESS" : "FAILURE")
<< std::endl;
res = test_conv1d_nwc_f32_instances();
std::cout << "\ntest_conv1d_nwc_f32_instances ..... " << (res ? "SUCCESS" : "FAILURE")
<< std::endl;
res = test_conv1d_nwc_int8_instances();
std::cout << "\ntest_conv1d_nwc_int8_instances ..... " << (res ? "SUCCESS" : "FAILURE")
<< std::endl;
return res ? 0 : 1;
EXPECT_TRUE(test_conv1d_nwc_instances<int8_t>(
ck::utils::conv::ConvolutionFwdInstances<int8_t, int8_t, int8_t>::Get<1>()));
}
test/convnd_fwd/conv2d_fwd.cpp
View file @ d0b9a467
......@@ -2,6 +2,7 @@
#include <iostream>
#include <tuple>
#include <vector>
#include "gtest/gtest.h"
#include "data_type.hpp"
#include "element_wise_operation.hpp"
......@@ -10,30 +11,6 @@
namespace {
bool test_conv2d_nhwc()
{
using namespace std::placeholders;
using namespace ck::utils;
ck::utils::conv::ConvParams params;
params.N = 2;
params.K = 16;
params.C = 4;
params.input_spatial_lengths = std::vector<ck::index_t>{16, 16};
params.conv_filter_strides = std::vector<ck::index_t>{1, 1};
std::vector<test::conv::DeviceConvFwdNoOpPtr> conv_ptrs;
test::conv::get_test_convolution_fwd_instance<2>(conv_ptrs);
conv::ConvFwdOpInstance<float, float, float> conv_instance(params);
auto reference_conv_fwd_fun = std::bind(
conv::run_reference_convolution_forward<2, float, float, float>, params, _1, _2, _3);
OpInstanceRunEngine<float, float, float> run_engine(conv_instance, reference_conv_fwd_fun);
run_engine.SetAtol(1e-5);
run_engine.SetRtol(1e-4);
return run_engine.Test(conv_ptrs);
}
template <typename T>
bool test_conv2d_nhwc_instances(const std::vector<test::conv::DeviceConvFwdNoOpPtr>& conv_ptrs)
{
......@@ -57,50 +34,58 @@ bool test_conv2d_nhwc_instances(const std::vector<test::conv::DeviceConvFwdNoOpP
return run_engine.Test(conv_ptrs);
}
bool test_conv2d_nhwc_bf16_instances()
} // anonymous namespace
TEST(Conv2DFwdNHWC, TestConv2D)
{
return test_conv2d_nhwc_instances<ck::bhalf_t>(
ck::utils::conv::ConvolutionFwdInstances<ck::bhalf_t, ck::bhalf_t, ck::bhalf_t>::Get<2>());
using namespace std::placeholders;
using namespace ck::utils;
ck::utils::conv::ConvParams params;
params.N = 2;
params.K = 16;
params.C = 4;
params.input_spatial_lengths = std::vector<ck::index_t>{16, 16};
params.conv_filter_strides = std::vector<ck::index_t>{1, 1};
std::vector<test::conv::DeviceConvFwdNoOpPtr> conv_ptrs;
test::conv::get_test_convolution_fwd_instance<2>(conv_ptrs);
conv::ConvFwdOpInstance<float, float, float> conv_instance(params);
auto reference_conv_fwd_fun = std::bind(
conv::run_reference_convolution_forward<2, float, float, float>, params, _1, _2, _3);
OpInstanceRunEngine<float, float, float> run_engine(conv_instance, reference_conv_fwd_fun);
run_engine.SetAtol(1e-5);
run_engine.SetRtol(1e-4);
EXPECT_TRUE(run_engine.Test(conv_ptrs));
}
bool test_conv2d_nhwc_f16_instances()
TEST(Conv2DFwdNHWC, Bf16Instances)
{
return test_conv2d_nhwc_instances<ck::half_t>(
ck::utils::conv::ConvolutionFwdInstances<ck::half_t, ck::half_t, ck::half_t>::Get<2>());
EXPECT_TRUE(test_conv2d_nhwc_instances<ck::bhalf_t>(
ck::utils::conv::ConvolutionFwdInstances<ck::bhalf_t, ck::bhalf_t, ck::bhalf_t>::Get<2>()));
}
bool test_conv2d_nhwc_f32_instances()
TEST(Conv2DFwdNHWC, F16Instances)
{
return test_conv2d_nhwc_instances<float>(
ck::utils::conv::ConvolutionFwdInstances<float, float, float>::Get<2>());
EXPECT_TRUE(test_conv2d_nhwc_instances<ck::half_t>(
ck::utils::conv::ConvolutionFwdInstances<ck::half_t, ck::half_t, ck::half_t>::Get<2>()));
}
bool test_conv2d_nhwc_int8_instances()
TEST(Conv2DFwdNHWC, BF32Instances)
{
return test_conv2d_nhwc_instances<int8_t>(
ck::utils::conv::ConvolutionFwdInstances<int8_t, int8_t, int8_t>::Get<2>());
EXPECT_TRUE(test_conv2d_nhwc_instances<float>(
ck::utils::conv::ConvolutionFwdInstances<float, float, float>::Get<2>()));
}
} // anonymous namespace
TEST(Conv2DFwdNHWC, F32Instances)
{
EXPECT_TRUE(test_conv2d_nhwc_instances<float>(
ck::utils::conv::ConvolutionFwdInstances<float, float, float>::Get<2>()));
}
int main()
TEST(Conv2DFwdNHWC, Int8Instances)
{
bool res{true};
res = test_conv2d_nhwc();
std::cout << "test_conv2d_nhwc ..... " << (res ? "SUCCESS" : "FAILURE") << std::endl;
res = test_conv2d_nhwc_bf16_instances();
std::cout << "\ntest_conv2d_nhwc_bf16_instances ..... " << (res ? "SUCCESS" : "FAILURE")
<< std::endl;
res = test_conv2d_nhwc_f16_instances();
std::cout << "\ntest_conv2d_nhwc_f16_instances ....." << (res ? "SUCCESS" : "FAILURE")
<< std::endl;
res = test_conv2d_nhwc_f32_instances();
std::cout << "\ntest_conv2d_nhwc_f32_instances ..... " << (res ? "SUCCESS" : "FAILURE")
<< std::endl;
res = test_conv2d_nhwc_int8_instances();
std::cout << "\ntest_conv2d_nhwc_int8_instances ..... " << (res ? "SUCCESS" : "FAILURE")
<< std::endl;
return res ? 0 : 1;
EXPECT_TRUE(test_conv2d_nhwc_instances<int8_t>(
ck::utils::conv::ConvolutionFwdInstances<int8_t, int8_t, int8_t>::Get<2>()));
}
test/convnd_fwd/conv3d_fwd.cpp
View file @ d0b9a467
......@@ -3,6 +3,7 @@
#include <stdexcept>
#include <tuple>
#include <vector>
#include "gtest/gtest.h"
#include "data_type.hpp"
#include "element_wise_operation.hpp"
......@@ -11,7 +12,34 @@
namespace {
bool test_conv3d_ndhwc()
template <typename T>
bool test_conv3d_ndhwc_instances(const std::vector<test::conv::DeviceConvFwdNoOpPtr>& conv_ptrs)
{
using namespace std::placeholders;
using namespace ck::utils;
namespace ctl = ck::tensor_layout::convolution;
conv::ConvParams params;
params.N = 64;
params.num_dim_spatial = 3;
params.filter_spatial_lengths = std::vector<ck::index_t>{3, 3, 2};
params.input_spatial_lengths = std::vector<ck::index_t>{32, 32, 2};
params.conv_filter_strides = std::vector<ck::index_t>{2, 2, 2};
params.conv_filter_dilations = std::vector<ck::index_t>{1, 1, 1};
params.input_left_pads = std::vector<ck::index_t>{1, 1, 1};
params.input_right_pads = std::vector<ck::index_t>{1, 1, 1};
conv::ConvFwdOpInstance<T, T, T, ctl::NDHWC, ctl::KZYXC, ctl::NDHWK> conv_instance(params);
auto reference_conv_fwd_fun =
std::bind(conv::run_reference_convolution_forward<3, T, T, T>, params, _1, _2, _3);
OpInstanceRunEngine<T, T, T> run_engine(conv_instance, reference_conv_fwd_fun);
return run_engine.Test(conv_ptrs);
}
} // anonymous namespace
TEST(Conv3DFwdNDHWC, TestConv3D)
{
using namespace std::placeholders;
using namespace ck::utils;
......@@ -39,10 +67,10 @@ bool test_conv3d_ndhwc()
OpInstanceRunEngine<float, float, float> run_engine(conv_instance, reference_conv_fwd_fun);
run_engine.SetAtol(1e-5);
run_engine.SetRtol(1e-4);
return run_engine.Test(conv_ptrs);
EXPECT_TRUE(run_engine.Test(conv_ptrs));
}
bool test_conv3d_ndhwc_2gb_input()
TEST(Conv3DFwdNDHWC, InputOver2GB)
{
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using namespace ck::utils;
......@@ -79,10 +107,10 @@ bool test_conv3d_ndhwc_2gb_input()
PassThrough{},
PassThrough{},
PassThrough{});
return !(conv_ptrs.back()->IsSupportedArgument(arg.get()));
EXPECT_FALSE(conv_ptrs.back()->IsSupportedArgument(arg.get()));
}
bool test_conv3d_ndhwc_2gb_filters()
TEST(Conv3DFwdNDHWC, FiltersOver2GB)
{
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using namespace ck::utils;
......@@ -119,10 +147,10 @@ bool test_conv3d_ndhwc_2gb_filters()
PassThrough{},
PassThrough{},
PassThrough{});
return !(conv_ptrs.back()->IsSupportedArgument(arg.get()));
EXPECT_FALSE(conv_ptrs.back()->IsSupportedArgument(arg.get()));
}
bool test_conv3d_ndhwc_2gb_output()
TEST(Conv3DFwdNDHWC, OutputOver2GB)
{
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using namespace ck::utils;
......@@ -158,88 +186,29 @@ bool test_conv3d_ndhwc_2gb_output()
PassThrough{},
PassThrough{},
PassThrough{});
return !(conv_ptrs.back()->IsSupportedArgument(arg.get()));
}
template <typename T>
bool test_conv3d_ndhwc_instances(const std::vector<test::conv::DeviceConvFwdNoOpPtr>& conv_ptrs)
{
using namespace std::placeholders;
using namespace ck::utils;
namespace ctl = ck::tensor_layout::convolution;
conv::ConvParams params;
params.N = 64;
params.num_dim_spatial = 3;
params.filter_spatial_lengths = std::vector<ck::index_t>{3, 3, 2};
params.input_spatial_lengths = std::vector<ck::index_t>{32, 32, 2};
params.conv_filter_strides = std::vector<ck::index_t>{2, 2, 2};
params.conv_filter_dilations = std::vector<ck::index_t>{1, 1, 1};
params.input_left_pads = std::vector<ck::index_t>{1, 1, 1};
params.input_right_pads = std::vector<ck::index_t>{1, 1, 1};
conv::ConvFwdOpInstance<T, T, T, ctl::NDHWC, ctl::KZYXC, ctl::NDHWK> conv_instance(params);
auto reference_conv_fwd_fun =
std::bind(conv::run_reference_convolution_forward<3, T, T, T>, params, _1, _2, _3);
OpInstanceRunEngine<T, T, T> run_engine(conv_instance, reference_conv_fwd_fun);
return run_engine.Test(conv_ptrs);
EXPECT_FALSE(conv_ptrs.back()->IsSupportedArgument(arg.get()));
}
bool test_conv3d_ndhwc_bf16_instances()
TEST(Conv3DFwdNDHWC, Bf16Instances)
{
return test_conv3d_ndhwc_instances<ck::bhalf_t>(
ck::utils::conv::ConvolutionFwdInstances<ck::bhalf_t, ck::bhalf_t, ck::bhalf_t>::Get<3>());
EXPECT_TRUE(test_conv3d_ndhwc_instances<ck::bhalf_t>(
ck::utils::conv::ConvolutionFwdInstances<ck::bhalf_t, ck::bhalf_t, ck::bhalf_t>::Get<3>()));
}
bool test_conv3d_ndhwc_f16_instances()
TEST(Conv3DFwdNDHWC, F16Instances)
{
return test_conv3d_ndhwc_instances<ck::half_t>(
ck::utils::conv::ConvolutionFwdInstances<ck::half_t, ck::half_t, ck::half_t>::Get<3>());
EXPECT_TRUE(test_conv3d_ndhwc_instances<ck::half_t>(
ck::utils::conv::ConvolutionFwdInstances<ck::half_t, ck::half_t, ck::half_t>::Get<3>()));
}
bool test_conv3d_ndhwc_f32_instances()
TEST(Conv3DFwdNDHWC, F32Instances)
{
return test_conv3d_ndhwc_instances<float>(
ck::utils::conv::ConvolutionFwdInstances<float, float, float>::Get<3>());
EXPECT_TRUE(test_conv3d_ndhwc_instances<float>(
ck::utils::conv::ConvolutionFwdInstances<float, float, float>::Get<3>()));
}
bool test_conv3d_ndhwc_int8_instances()
{
return test_conv3d_ndhwc_instances<int8_t>(
ck::utils::conv::ConvolutionFwdInstances<int8_t, int8_t, int8_t>::Get<3>());
}
} // anonymous namespace
int main()
TEST(Conv3DFwdNDHWC, Int8Instances)
{
bool res{true};
res = test_conv3d_ndhwc();
std::cout << "test_conv3d_ndhwc ..... " << (res ? "SUCCESS" : "FAILURE") << std::endl;
res = test_conv3d_ndhwc_2gb_input();
std::cout << "\ntest_conv3d_ndhwc_2gb_input ..... " << (res ? "SUCCESS" : "FAILURE")
<< std::endl;
res = test_conv3d_ndhwc_2gb_filters();
std::cout << "\ntest_conv3d_ndhwc_2gb_filters ..... " << (res ? "SUCCESS" : "FAILURE")
<< std::endl;
res = test_conv3d_ndhwc_2gb_output();
std::cout << "\ntest_conv3d_ndhwc_2gb_output ..... " << (res ? "SUCCESS" : "FAILURE")
<< std::endl;
res = test_conv3d_ndhwc_bf16_instances();
std::cout << "\ntest_conv3d_ndhwc_bf16_instances ..... " << (res ? "SUCCESS" : "FAILURE")
<< std::endl;
res = test_conv3d_ndhwc_f16_instances();
std::cout << "\ntest_conv3d_ndhwc_f16_instances ..... " << (res ? "SUCCESS" : "FAILURE")
<< std::endl;
res = test_conv3d_ndhwc_f32_instances();
std::cout << "\ntest_conv3d_ndhwc_f32_instances ..... " << (res ? "SUCCESS" : "FAILURE")
<< std::endl;
res = test_conv3d_ndhwc_int8_instances();
std::cout << "\ntest_conv3d_ndhwc_int8_instances ..... " << (res ? "SUCCESS" : "FAILURE")
<< std::endl;
return res ? 0 : 1;
EXPECT_TRUE(test_conv3d_ndhwc_instances<int8_t>(
ck::utils::conv::ConvolutionFwdInstances<int8_t, int8_t, int8_t>::Get<3>()));
}
test/gemm/gemm_util.hpp
View file @ d0b9a467
#ifndef GEMM_UTILS_HPP
#define GEMM_UTILS_HPP
#include "check_err.hpp"
#include "config.hpp"
#include "device.hpp"
#include "host_tensor.hpp"
#include "host_tensor_generator.hpp"
#include "reference_gemm.hpp"
#include "tensor_layout.hpp"
namespace ck {
namespace gemm_util {
struct GemmParams
{
GemmParams()
: M(1024), N(1024), K(1024), StrideA(1024), StrideB(1024), StrideC(1024), alpha(1), beta(0)
{
}
ck::index_t M;
ck::index_t N;
ck::index_t K;
ck::index_t StrideA;
ck::index_t StrideB;
ck::index_t StrideC;
float alpha;
float beta;
};
template <typename GemmInstance,
typename ADataType,
typename BDataType,
typename CDataType,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CElementwiseOperation>
void RunHostGEMM(const Tensor<ADataType>& A,
const Tensor<BDataType>& B,
Tensor<CDataType>& C,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CElementwiseOperation c_element_op)
{
auto ref_gemm = GemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument = ref_gemm.MakeArgument(A, B, C, a_element_op, b_element_op, c_element_op);
ref_invoker.Run(ref_argument);
}
template <typename DeviceGemmPtr_,
typename ADataType,
typename BDataType,
typename CDataType,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CElementwiseOperation>
void RunDeviceGEMM(DeviceGemmPtr_& gemmPtr,
const ck::gemm_util::GemmParams& params,
const Tensor<ADataType>& A,
const Tensor<BDataType>& B,
Tensor<CDataType>& C,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CElementwiseOperation c_element_op)
{
DeviceMem a_m_k_device_buf(sizeof(ADataType) * A.mDesc.GetElementSpace());
DeviceMem b_k_n_device_buf(sizeof(BDataType) * B.mDesc.GetElementSpace());
DeviceMem c_m_n_device_buf(sizeof(CDataType) * C.mDesc.GetElementSpace());
a_m_k_device_buf.ToDevice(A.mData.data());
b_k_n_device_buf.ToDevice(B.mData.data());
auto invoker_ptr = gemmPtr->MakeInvokerPointer();
auto argument_ptr =
gemmPtr->MakeArgumentPointer(static_cast<ADataType*>(a_m_k_device_buf.GetDeviceBuffer()),
static_cast<BDataType*>(b_k_n_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c_m_n_device_buf.GetDeviceBuffer()),
params.M,
params.N,
params.K,
params.StrideA,
params.StrideB,
params.StrideC,
a_element_op,
b_element_op,
c_element_op);
if(!gemmPtr->IsSupportedArgument(argument_ptr.get()))
{
throw std::runtime_error(
"wrong! device_gemm with the specified compilation parameters does "
"not support this GEMM problem");
}
invoker_ptr->Run(argument_ptr.get());
c_m_n_device_buf.FromDevice(C.mData.data());
}
template <typename DeviceGemmPtr_,
typename ADataType,
typename BDataType,
typename CDataType,
typename ALayout,
typename BLayout,
typename CLayout,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CElementwiseOperation>
struct TestGemm
{
auto PrepareGemmTensor(const ck::gemm_util::GemmParams& params)
{
auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
{
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({stride, 1}));
}
else
{
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({1, stride}));
}
};
Tensor<ADataType> a_m_k(
f_host_tensor_descriptor(params.M, params.K, params.StrideA, ALayout{}));
Tensor<BDataType> b_k_n(
f_host_tensor_descriptor(params.K, params.N, params.StrideB, BLayout{}));
Tensor<CDataType> c_m_n_host_result(
f_host_tensor_descriptor(params.M, params.N, params.StrideC, CLayout{}));
Tensor<CDataType> c_m_n_device_result(
f_host_tensor_descriptor(params.M, params.N, params.StrideC, CLayout{}));
auto f_generate_tensor_value = [](auto desc, auto type) {
using dataType = decltype(type);
if(std::is_same<dataType, int8_t>::value)
{
desc.GenerateTensorValue(GeneratorTensor_2<int8_t>{-5, 5});
}
else
{
desc.GenerateTensorValue(GeneratorTensor_3<dataType>{-0.5, 0.5});
}
};
f_generate_tensor_value(a_m_k, ADataType{});
f_generate_tensor_value(b_k_n, BDataType{});
return std::make_tuple(a_m_k, b_k_n, c_m_n_host_result, c_m_n_device_result);
}
auto operator()(DeviceGemmPtr_& gemmPtr)
{
std::cout << "ALayout = " << ALayout{}.name << ", BLayout = " << BLayout{}.name
<< ", CLayout = " << CLayout{}.name << std::endl;
std::cout << gemmPtr->GetTypeString() << std::endl;
// Arrange
ck::gemm_util::GemmParams params;
params.M = 1024;
params.N = 1024;
params.K = 1024;
params.StrideA = 1024;
params.StrideB = 1024;
params.StrideC = 1024;
auto host_tensors = PrepareGemmTensor(params);
const Tensor<ADataType>& a = std::get<0>(host_tensors);
const Tensor<BDataType>& b = std::get<1>(host_tensors);
Tensor<CDataType>& c_host = std::get<2>(host_tensors);
Tensor<CDataType>& c_device = std::get<3>(host_tensors);
auto a_element_op = AElementwiseOperation{};
auto b_element_op = BElementwiseOperation{};
auto c_element_op = CElementwiseOperation{};
using ReferenceGemmInstance =
ck::tensor_operation::host::ReferenceGemm<ADataType,
BDataType,
CDataType,
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation>;
ck::gemm_util::RunHostGEMM<ReferenceGemmInstance>(
a, b, c_host, a_element_op, b_element_op, c_element_op);
// Act
ck::gemm_util::RunDeviceGEMM(
gemmPtr, params, a, b, c_device, a_element_op, b_element_op, c_element_op);
// Assert
bool res = false;
if(std::is_same<CDataType, float>::value)
{
res = ck::utils::check_err(c_device.mData, c_host.mData);
std::cout << (res ? "SUCCESS" : "FAILURE") << std::endl;
}
else if(std::is_same<CDataType, ck::half_t>::value)
{
res = ck::utils::check_err(c_device.mData, c_host.mData);
std::cout << (res ? "SUCCESS" : "FAILURE") << std::endl;
}
else if(std::is_same<CDataType, int8_t>::value)
{
res = ck::utils::check_err(c_device.mData, c_host.mData);
std::cout << (res ? "SUCCESS" : "FAILURE") << std::endl;
}
return res;
}
};
template <typename DeviceGemmPtr_,
typename ALayout,
typename BLayout,
typename CLayout,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CElementwiseOperation>
struct TestGemmBF16
{
using BF16 = ck::bhalf_t;
auto PrepareGemmTensorBF16(const ck::gemm_util::GemmParams& params)
{
auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
{
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({stride, 1}));
}
else
{
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({1, stride}));
}
};
// use fp32 host kernel to verify bf16 device kernel
Tensor<BF16> a_m_k_bf16(
f_host_tensor_descriptor(params.M, params.K, params.StrideA, ALayout{}));
Tensor<BF16> b_k_n_bf16(
f_host_tensor_descriptor(params.K, params.N, params.StrideB, BLayout{}));
Tensor<BF16> c_m_n_device_bf16(
f_host_tensor_descriptor(params.M, params.N, params.StrideC, CLayout{}));
Tensor<float> a_m_k_fp32(
f_host_tensor_descriptor(params.M, params.K, params.StrideA, ALayout{}));
Tensor<float> b_k_n_fp32(
f_host_tensor_descriptor(params.K, params.N, params.StrideB, BLayout{}));
Tensor<float> c_m_n_host_fp32(
f_host_tensor_descriptor(params.M, params.N, params.StrideC, CLayout{}));
Tensor<float> c_m_n_device_fp32(
f_host_tensor_descriptor(params.M, params.N, params.StrideC, CLayout{}));
a_m_k_bf16.GenerateTensorValue(GeneratorTensor_3<BF16>{-0.5, 0.5});
b_k_n_bf16.GenerateTensorValue(GeneratorTensor_3<BF16>{-0.5, 0.5});
bf16_to_f32_(a_m_k_bf16, a_m_k_fp32);
bf16_to_f32_(b_k_n_bf16, b_k_n_fp32);
return std::make_tuple(a_m_k_bf16,
b_k_n_bf16,
c_m_n_device_bf16,
a_m_k_fp32,
b_k_n_fp32,
c_m_n_host_fp32,
c_m_n_device_fp32);
}
auto operator()(DeviceGemmPtr_& gemmPtr)
{
// Arrange
ck::gemm_util::GemmParams params;
params.M = 1024;
params.N = 1024;
params.K = 1024;
params.StrideA = 1024;
params.StrideB = 1024;
params.StrideC = 1024;
auto host_tensors = PrepareGemmTensorBF16(params);
const Tensor<BF16>& a_bf16 = std::get<0>(host_tensors);
const Tensor<BF16>& b_bf16 = std::get<1>(host_tensors);
Tensor<BF16>& c_device_bf16 = std::get<2>(host_tensors);
Tensor<float>& a_fp32 = std::get<3>(host_tensors);
Tensor<float>& b_fp32 = std::get<4>(host_tensors);
Tensor<float>& c_host_fp32 = std::get<5>(host_tensors);
Tensor<float>& c_device_fp32 = std::get<6>(host_tensors);
auto a_element_op = AElementwiseOperation{};
auto b_element_op = BElementwiseOperation{};
auto c_element_op = CElementwiseOperation{};
// use fp32 host kernel to verify bf16 device kernel
using ReferenceGemmInstance =
ck::tensor_operation::host::ReferenceGemm<float,
float,
float,
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation>;
ck::gemm_util::RunHostGEMM<ReferenceGemmInstance>(
a_fp32, b_fp32, c_host_fp32, a_element_op, b_element_op, c_element_op);
// Act
ck::gemm_util::RunDeviceGEMM(gemmPtr,
params,
a_bf16,
b_bf16,
c_device_bf16,
a_element_op,
b_element_op,
c_element_op);
bf16_to_f32_(c_device_bf16, c_device_fp32);
// Assert
bool res = ck::utils::check_err(
c_device_fp32.mData, c_host_fp32.mData, "Error: incorrect results!", 1e-2f, 1e-3f);
std::cout << (res ? "SUCCESS" : "FAILURE") << std::endl;
return res;
};
};
} // namespace gemm_util
} // namespace ck
#endif
#ifndef GEMM_UTILS_HPP
#define GEMM_UTILS_HPP
#include "check_err.hpp"
#include "config.hpp"
#include "device.hpp"
#include "host_tensor.hpp"
#include "host_tensor_generator.hpp"
#include "reference_gemm.hpp"
#include "tensor_layout.hpp"
namespace ck {
namespace gemm_util {
struct GemmParams
{
GemmParams()
: M(1024), N(1024), K(1024), StrideA(1024), StrideB(1024), StrideC(1024), alpha(1), beta(0)
{
}
ck::index_t M;
ck::index_t N;
ck::index_t K;
ck::index_t StrideA;
ck::index_t StrideB;
ck::index_t StrideC;
float alpha;
float beta;
};
template <typename GemmInstance,
typename ADataType,
typename BDataType,
typename CDataType,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CElementwiseOperation>
void RunHostGEMM(const Tensor<ADataType>& A,
const Tensor<BDataType>& B,
Tensor<CDataType>& C,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CElementwiseOperation c_element_op)
{
auto ref_gemm = GemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument = ref_gemm.MakeArgument(A, B, C, a_element_op, b_element_op, c_element_op);
ref_invoker.Run(ref_argument);
}
template <typename DeviceGemmPtr_,
typename ADataType,
typename BDataType,
typename CDataType,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CElementwiseOperation>
void RunDeviceGEMM(DeviceGemmPtr_& gemmPtr,
const ck::gemm_util::GemmParams& params,
const Tensor<ADataType>& A,
const Tensor<BDataType>& B,
Tensor<CDataType>& C,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CElementwiseOperation c_element_op)
{
DeviceMem a_m_k_device_buf(sizeof(ADataType) * A.mDesc.GetElementSpace());
DeviceMem b_k_n_device_buf(sizeof(BDataType) * B.mDesc.GetElementSpace());
DeviceMem c_m_n_device_buf(sizeof(CDataType) * C.mDesc.GetElementSpace());
a_m_k_device_buf.ToDevice(A.mData.data());
b_k_n_device_buf.ToDevice(B.mData.data());
auto invoker_ptr = gemmPtr->MakeInvokerPointer();
auto argument_ptr =
gemmPtr->MakeArgumentPointer(static_cast<ADataType*>(a_m_k_device_buf.GetDeviceBuffer()),
static_cast<BDataType*>(b_k_n_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c_m_n_device_buf.GetDeviceBuffer()),
params.M,
params.N,
params.K,
params.StrideA,
params.StrideB,
params.StrideC,
a_element_op,
b_element_op,
c_element_op);
if(!gemmPtr->IsSupportedArgument(argument_ptr.get()))
{
throw std::runtime_error(
"wrong! device_gemm with the specified compilation parameters does "
"not support this GEMM problem");
}
invoker_ptr->Run(argument_ptr.get());
c_m_n_device_buf.FromDevice(C.mData.data());
}
template <typename DeviceGemmPtr_,
typename ADataType,
typename BDataType,
typename CDataType,
typename ALayout,
typename BLayout,
typename CLayout,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CElementwiseOperation>
struct TestGemm
{
auto PrepareGemmTensor(const ck::gemm_util::GemmParams& params)
{
auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
{
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({stride, 1}));
}
else
{
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({1, stride}));
}
};
Tensor<ADataType> a_m_k(
f_host_tensor_descriptor(params.M, params.K, params.StrideA, ALayout{}));
Tensor<BDataType> b_k_n(
f_host_tensor_descriptor(params.K, params.N, params.StrideB, BLayout{}));
Tensor<CDataType> c_m_n_host_result(
f_host_tensor_descriptor(params.M, params.N, params.StrideC, CLayout{}));
Tensor<CDataType> c_m_n_device_result(
f_host_tensor_descriptor(params.M, params.N, params.StrideC, CLayout{}));
auto f_generate_tensor_value = [](auto& tensor, auto type) {
using dataType = decltype(type);
tensor.GenerateTensorValue(GeneratorTensor_2<dataType>{-5, 5});
};
f_generate_tensor_value(a_m_k, ADataType{});
f_generate_tensor_value(b_k_n, BDataType{});
return std::make_tuple(a_m_k, b_k_n, c_m_n_host_result, c_m_n_device_result);
}
auto operator()(DeviceGemmPtr_& gemmPtr)
{
std::cout << "ALayout = " << ALayout{}.name << ", BLayout = " << BLayout{}.name
<< ", CLayout = " << CLayout{}.name << std::endl;
std::cout << gemmPtr->GetTypeString() << std::endl;
// Arrange
ck::gemm_util::GemmParams params;
params.M = 1024;
params.N = 1024;
params.K = 1024;
params.StrideA = 1024;
params.StrideB = 1024;
params.StrideC = 1024;
auto host_tensors = PrepareGemmTensor(params);
const Tensor<ADataType>& a = std::get<0>(host_tensors);
const Tensor<BDataType>& b = std::get<1>(host_tensors);
Tensor<CDataType>& c_host = std::get<2>(host_tensors);
Tensor<CDataType>& c_device = std::get<3>(host_tensors);
auto a_element_op = AElementwiseOperation{};
auto b_element_op = BElementwiseOperation{};
auto c_element_op = CElementwiseOperation{};
using ReferenceGemmInstance =
ck::tensor_operation::host::ReferenceGemm<ADataType,
BDataType,
CDataType,
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation>;
ck::gemm_util::RunHostGEMM<ReferenceGemmInstance>(
a, b, c_host, a_element_op, b_element_op, c_element_op);
// Act
ck::gemm_util::RunDeviceGEMM(
gemmPtr, params, a, b, c_device, a_element_op, b_element_op, c_element_op);
// Assert
bool res = false;
if(std::is_same<CDataType, float>::value)
{
res = ck::utils::check_err(c_device.mData, c_host.mData);
std::cout << (res ? "SUCCESS" : "FAILURE") << std::endl;
}
else if(std::is_same<CDataType, ck::half_t>::value)
{
res = ck::utils::check_err(c_device.mData, c_host.mData);
std::cout << (res ? "SUCCESS" : "FAILURE") << std::endl;
}
else if(std::is_same<CDataType, int8_t>::value)
{
res = ck::utils::check_err(c_device.mData, c_host.mData);
std::cout << (res ? "SUCCESS" : "FAILURE") << std::endl;
}
return res;
}
};
template <typename DeviceGemmPtr_,
typename ALayout,
typename BLayout,
typename CLayout,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CElementwiseOperation>
struct TestGemmBF16
{
using BF16 = ck::bhalf_t;
auto PrepareGemmTensorBF16(const ck::gemm_util::GemmParams& params)
{
auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
{
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({stride, 1}));
}
else
{
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({1, stride}));
}
};
// use fp32 host kernel to verify bf16 device kernel
Tensor<BF16> a_m_k_bf16(
f_host_tensor_descriptor(params.M, params.K, params.StrideA, ALayout{}));
Tensor<BF16> b_k_n_bf16(
f_host_tensor_descriptor(params.K, params.N, params.StrideB, BLayout{}));
Tensor<BF16> c_m_n_device_bf16(
f_host_tensor_descriptor(params.M, params.N, params.StrideC, CLayout{}));
Tensor<float> a_m_k_fp32(
f_host_tensor_descriptor(params.M, params.K, params.StrideA, ALayout{}));
Tensor<float> b_k_n_fp32(
f_host_tensor_descriptor(params.K, params.N, params.StrideB, BLayout{}));
Tensor<float> c_m_n_host_fp32(
f_host_tensor_descriptor(params.M, params.N, params.StrideC, CLayout{}));
Tensor<float> c_m_n_device_fp32(
f_host_tensor_descriptor(params.M, params.N, params.StrideC, CLayout{}));
a_m_k_bf16.GenerateTensorValue(GeneratorTensor_3<BF16>{-0.5, 0.5});
b_k_n_bf16.GenerateTensorValue(GeneratorTensor_3<BF16>{-0.5, 0.5});
bf16_to_f32_(a_m_k_bf16, a_m_k_fp32);
bf16_to_f32_(b_k_n_bf16, b_k_n_fp32);
return std::make_tuple(a_m_k_bf16,
b_k_n_bf16,
c_m_n_device_bf16,
a_m_k_fp32,
b_k_n_fp32,
c_m_n_host_fp32,
c_m_n_device_fp32);
}
auto operator()(DeviceGemmPtr_& gemmPtr)
{
// Arrange
ck::gemm_util::GemmParams params;
params.M = 1024;
params.N = 1024;
params.K = 1024;
params.StrideA = 1024;
params.StrideB = 1024;
params.StrideC = 1024;
auto host_tensors = PrepareGemmTensorBF16(params);
const Tensor<BF16>& a_bf16 = std::get<0>(host_tensors);
const Tensor<BF16>& b_bf16 = std::get<1>(host_tensors);
Tensor<BF16>& c_device_bf16 = std::get<2>(host_tensors);
Tensor<float>& a_fp32 = std::get<3>(host_tensors);
Tensor<float>& b_fp32 = std::get<4>(host_tensors);
Tensor<float>& c_host_fp32 = std::get<5>(host_tensors);
Tensor<float>& c_device_fp32 = std::get<6>(host_tensors);
auto a_element_op = AElementwiseOperation{};
auto b_element_op = BElementwiseOperation{};
auto c_element_op = CElementwiseOperation{};
// use fp32 host kernel to verify bf16 device kernel
using ReferenceGemmInstance =
ck::tensor_operation::host::ReferenceGemm<float,
float,
float,
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation>;
ck::gemm_util::RunHostGEMM<ReferenceGemmInstance>(
a_fp32, b_fp32, c_host_fp32, a_element_op, b_element_op, c_element_op);
// Act
ck::gemm_util::RunDeviceGEMM(gemmPtr,
params,
a_bf16,
b_bf16,
c_device_bf16,
a_element_op,
b_element_op,
c_element_op);
bf16_to_f32_(c_device_bf16, c_device_fp32);
// Assert
bool res = ck::utils::check_err(
c_device_fp32.mData, c_host_fp32.mData, "Error: incorrect results!", 1e-2f, 1e-3f);
std::cout << (res ? "SUCCESS" : "FAILURE") << std::endl;
return res;
};
};
} // namespace gemm_util
} // namespace ck
#endif
test/reference_conv_fwd/CMakeLists.txt
View file @ d0b9a467
add_test_executable(test_reference_conv_fwd reference_conv_fwd.cpp)
add_gtest_executable(test_reference_conv_fwd reference_conv_fwd.cpp)
target_link_libraries(test_reference_conv_fwd PRIVATE host_tensor conv_fwd_util)
test/reference_conv_fwd/reference_conv_fwd.cpp
View file @ d0b9a467
......@@ -4,6 +4,7 @@
#include <numeric>
#include <type_traits>
#include <vector>
#include "gtest/gtest.h"
#include "check_err.hpp"
#include "config.hpp"
......@@ -82,13 +83,13 @@ run_reference_convolution_forward(const ck::utils::conv::ConvParams& params,
OutElementOp{});
ref_invoker.Run(ref_argument);
// std::cout <<"output: " << host_output.mDesc << std::endl << host_output.mData << std::endl;
return host_output;
}
bool test_conv2d_nhwc()
} // anonymous namespace
TEST(ReferenceConvolutionFWD, Conv2DNHWC)
{
bool res{true};
ck::utils::conv::ConvParams params;
params.N = 1;
params.K = 1;
......@@ -118,11 +119,14 @@ bool test_conv2d_nhwc()
472.5,
490.5,
508.5};
res = res && ck::utils::check_err(out_tensor.mDesc.GetLengths(),
ref_dims,
"Error: wrong output tensor dimensions!");
res = res && ck::utils::check_err(out_tensor.mData, ref_data, "Error: incorrect results!");
EXPECT_TRUE(ck::utils::check_err(
out_tensor.mDesc.GetLengths(), ref_dims, "Error: wrong output tensor dimensions!"));
EXPECT_TRUE(ck::utils::check_err(out_tensor.mData, ref_data, "Error: incorrect results!"));
}
TEST(ReferenceConvolutionFWD, Conv2DNHWCStridesDilationsPadding)
{
ck::utils::conv::ConvParams params;
params.N = 1;
params.K = 2;
params.C = 2;
......@@ -133,25 +137,21 @@ bool test_conv2d_nhwc()
params.input_left_pads = std::vector<ck::index_t>{1, 1};
params.input_right_pads = std::vector<ck::index_t>{1, 1};
out_tensor = run_reference_convolution_forward<2>(params);
ref_dims = std::vector<std::size_t>{1, 2, 5, 5};
ref_data = std::vector<float>{
auto out_tensor = run_reference_convolution_forward<2>(params);
std::vector<std::size_t> ref_dims = std::vector<std::size_t>{1, 2, 5, 5};
std::vector<float> ref_data{
210., 210., 327., 327., 351., 351., 375., 375., 399., 399.,
459., 459., 706.5, 706.5, 742.5, 742.5, 778.5, 778.5, 814.5, 814.5,
747., 747., 1138.5, 1138.5, 1174.5, 1174.5, 1210.5, 1210.5, 1246.5, 1246.5,
1035., 1035., 1570.5, 1570.5, 1606.5, 1606.5, 1642.5, 1642.5, 1678.5, 1678.5,
1323., 1323., 2002.5, 2002.5, 2038.5, 2038.5, 2074.5, 2074.5, 2110.5, 2110.5};
res = res && ck::utils::check_err(out_tensor.mDesc.GetLengths(),
ref_dims,
"Error: wrong output tensor dimensions!");
res = res && ck::utils::check_err(out_tensor.mData, ref_data, "Error: incorrect results!");
return res;
EXPECT_TRUE(ck::utils::check_err(
out_tensor.mDesc.GetLengths(), ref_dims, "Error: wrong output tensor dimensions!"));
EXPECT_TRUE(ck::utils::check_err(out_tensor.mData, ref_data, "Error: incorrect results!"));
}
bool test_conv1d_nwc()
TEST(ReferenceConvolutionFWD, Conv1DNWC)
{
bool res{true};
ck::utils::conv::ConvParams params;
params.num_dim_spatial = 1;
params.N = 1;
......@@ -174,11 +174,14 @@ bool test_conv1d_nwc()
ck::tensor_layout::convolution::NWK>(params);
std::vector<std::size_t> ref_dims{1, 1, 4};
std::vector<float> ref_data{7.5, 13.5, 19.5, 25.5};
res = res && ck::utils::check_err(out_tensor.mDesc.GetLengths(),
ref_dims,
"Error: wrong output tensor dimensions!");
res = res && ck::utils::check_err(out_tensor.mData, ref_data, "Error: incorrect results!");
EXPECT_TRUE(ck::utils::check_err(
out_tensor.mDesc.GetLengths(), ref_dims, "Error: wrong output tensor dimensions!"));
EXPECT_TRUE(ck::utils::check_err(out_tensor.mData, ref_data, "Error: incorrect results!"));
}
TEST(ReferenceConvolutionFWD, Conv1DNWCStridesDilationsPadding)
{
ck::utils::conv::ConvParams params;
params.num_dim_spatial = 1;
params.N = 1;
params.K = 2;
......@@ -190,20 +193,24 @@ bool test_conv1d_nwc()
params.input_left_pads = std::vector<ck::index_t>{1};
params.input_right_pads = std::vector<ck::index_t>{1};
out_tensor = run_reference_convolution_forward<1,
float,
float,
float,
ck::tensor_layout::convolution::NWC,
ck::tensor_layout::convolution::KXC,
ck::tensor_layout::convolution::NWK>(params);
ref_dims = std::vector<std::size_t>{1, 2, 5};
ref_data = std::vector<float>{9., 9., 19.5, 19.5, 31.5, 31.5, 43.5, 43.5, 55.5, 55.5};
res = res && ck::utils::check_err(out_tensor.mDesc.GetLengths(),
ref_dims,
"Error: wrong output tensor dimensions!");
res = res && ck::utils::check_err(out_tensor.mData, ref_data, "Error: incorrect results!");
auto out_tensor =
run_reference_convolution_forward<1,
float,
float,
float,
ck::tensor_layout::convolution::NWC,
ck::tensor_layout::convolution::KXC,
ck::tensor_layout::convolution::NWK>(params);
std::vector<std::size_t> ref_dims{1, 2, 5};
std::vector<float> ref_data{9., 9., 19.5, 19.5, 31.5, 31.5, 43.5, 43.5, 55.5, 55.5};
EXPECT_TRUE(ck::utils::check_err(
out_tensor.mDesc.GetLengths(), ref_dims, "Error: wrong output tensor dimensions!"));
EXPECT_TRUE(ck::utils::check_err(out_tensor.mData, ref_data, "Error: incorrect results!"));
}
TEST(ReferenceConvolutionFWD, Conv1DNWCSameOutputSize)
{
ck::utils::conv::ConvParams params;
params.num_dim_spatial = 1;
params.N = 2;
params.K = 16;
......@@ -224,8 +231,8 @@ bool test_conv1d_nwc()
ck::tensor_layout::convolution::NWK>(
params, ck::utils::FillMonotonicSeq<float>{0.f, 0.1f});
ref_dims = std::vector<std::size_t>{2, 16, 16};
ref_data = std::vector<float>{
std::vector<std::size_t> ref_dims{2, 16, 16};
std::vector<float> ref_data{
1.4, 1.4, 1.4, 1.4, 1.4, 1.4, 1.4, 1.4,
1.4, 1.4, 1.4, 1.4, 1.4, 1.4, 1.4, 1.4,
3.3, 3.3, 3.3, 3.3, 3.3, 3.3, 3.3, 3.3,
......@@ -290,17 +297,13 @@ bool test_conv1d_nwc()
72.9, 72.9, 72.9, 72.9, 72.9, 72.9, 72.9, 72.9,
49.4, 49.4, 49.4, 49.4, 49.4, 49.4, 49.4, 49.4,
49.4, 49.4, 49.4, 49.4, 49.4, 49.4, 49.4, 49.4};
res = res && ck::utils::check_err(out_tensor2.mDesc.GetLengths(),
ref_dims,
"Error: wrong output tensor dimensions!");
res = res && ck::utils::check_err(out_tensor2.mData, ref_data, "Error: incorrect results!");
return res;
EXPECT_TRUE(ck::utils::check_err(
out_tensor2.mDesc.GetLengths(), ref_dims, "Error: wrong output tensor dimensions!"));
EXPECT_TRUE(ck::utils::check_err(out_tensor2.mData, ref_data, "Error: incorrect results!"));
}
bool test_conv3d_ncdhw()
TEST(ReferenceConvolutionFWD, Conv3DNCDHW)
{
bool res{true};
ck::utils::conv::ConvParams params;
params.num_dim_spatial = 3;
params.N = 1;
......@@ -331,12 +334,17 @@ bool test_conv3d_ncdhw()
634.5, 637.2, 639.9, 642.60004, 650.7, 653.4, 656.10004, 658.8,
699.3, 702., 704.7, 707.4, 715.5, 718.2, 720.9, 723.60004,
731.7, 734.4001, 737.10004, 739.8, 747.9001, 750.60004, 753.3, 756.};
res = res && ck::utils::check_err(out_tensor.mDesc.GetLengths(),
ref_dims,
"Error [case 1]: wrong output tensor dimensions!");
res = res &&
ck::utils::check_err(out_tensor.mData, ref_data, "Error [case 1]: incorrect results!");
EXPECT_TRUE(ck::utils::check_err(out_tensor.mDesc.GetLengths(),
ref_dims,
"Error [case 1]: wrong output tensor dimensions!"));
EXPECT_TRUE(
ck::utils::check_err(out_tensor.mData, ref_data, "Error [case 1]: incorrect results!"));
}
TEST(ReferenceConvolutionFWD, Conv3DNCDHWStridesDilations)
{
ck::utils::conv::ConvParams params;
params.num_dim_spatial = 3;
params.N = 1;
params.K = 2;
params.C = 2;
......@@ -347,16 +355,16 @@ bool test_conv3d_ncdhw()
params.input_left_pads = std::vector<ck::index_t>{0, 0, 0};
params.input_right_pads = std::vector<ck::index_t>{0, 0, 0};
out_tensor = run_reference_convolution_forward<3,
float,
float,
float,
ck::tensor_layout::convolution::NCDHW,
ck::tensor_layout::convolution::KCZYX,
ck::tensor_layout::convolution::NKDHW>(
auto out_tensor = run_reference_convolution_forward<3,
float,
float,
float,
ck::tensor_layout::convolution::NCDHW,
ck::tensor_layout::convolution::KCZYX,
ck::tensor_layout::convolution::NKDHW>(
params, ck::utils::FillMonotonicSeq<float>{0.f, 0.1f});
ref_dims = std::vector<std::size_t>{1, 2, 4, 4, 4};
ref_data = std::vector<float>{
std::vector<std::size_t> ref_dims{1, 2, 4, 4, 4};
std::vector<float> ref_data{
2756.7002, 2764.7998, 2772.9001, 2781., 2853.9001, 2862., 2870.1, 2878.2002,
2951.1, 2959.2002, 2967.2998, 2975.4001, 3048.2998, 3056.4001, 3064.5, 3072.6,
3923.1, 3931.2, 3939.2998, 3947.4, 4020.2998, 4028.4001, 4036.5002, 4044.5999,
......@@ -373,26 +381,9 @@ bool test_conv3d_ncdhw()
5283.9004, 5292., 5300.0996, 5308.2, 5381.0996, 5389.2, 5397.3, 5405.4004,
6255.9004, 6264.0005, 6272.1, 6280.2, 6353.1, 6361.2, 6369.301, 6377.4,
6450.301, 6458.4, 6466.5, 6474.6, 6547.5, 6555.6, 6563.699, 6571.801};
res = res && ck::utils::check_err(out_tensor.mDesc.GetLengths(),
ref_dims,
"Error [case 2]: wrong output tensor dimensions!");
res =
res && ck::utils::check_err(
out_tensor.mData, ref_data, "Error [case 2]: incorrect results!", 1e-4f, 1e-6f);
return res;
}
} // anonymous namespace
int main(void)
{
bool res{true};
res = test_conv2d_nhwc();
std::cout << "test_conv2d_nhwc ..... " << (res ? "SUCCESS" : "FAILURE") << std::endl;
res = test_conv1d_nwc();
std::cout << "TestConv1DNHWC ..... " << (res ? "SUCCESS" : "FAILURE") << std::endl;
res = test_conv3d_ncdhw();
std::cout << "test_conv3d_ncdhw ..... " << (res ? "SUCCESS" : "FAILURE") << std::endl;
return res ? 0 : 1;
EXPECT_TRUE(ck::utils::check_err(out_tensor.mDesc.GetLengths(),
ref_dims,
"Error [case 2]: wrong output tensor dimensions!"));
EXPECT_TRUE(ck::utils::check_err(
out_tensor.mData, ref_data, "Error [case 2]: incorrect results!", 1e-4f, 1e-6f));
}
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment