From 28889caea5d1cb466227bff32680f13125193e09 Mon Sep 17 00:00:00 2001 From: Tao Luo Date: Mon, 8 Oct 2018 15:20:44 +0800 Subject: [PATCH] disable EIGEN_FAST_MATH and use_fast_math test=develop --- CMakeLists.txt | 1 + cmake/cuda.cmake | 5 ++++- cmake/external/eigen.cmake | 8 ++++++++ .../fluid/inference/tests/api/analyzer_resnet50_tester.cc | 3 +++ 4 files changed, 16 insertions(+), 1 deletion(-) diff --git a/CMakeLists.txt b/CMakeLists.txt index d43df124bde..24262c1821d 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -72,6 +72,7 @@ option(WITH_INFERENCE "Compile fluid inference library" ON) option(WITH_INFERENCE_API_TEST "Test fluid inference high-level api interface" OFF) option(WITH_SYSTEM_BLAS "Use system blas library" OFF) option(PY_VERSION "Compile PaddlePaddle with python3 support" ${PY_VERSION}) +option(WITH_FAST_MATH "Make use of fast math library" OFF) # PY_VERSION if(NOT PY_VERSION) diff --git a/cmake/cuda.cmake b/cmake/cuda.cmake index 03c73786a6c..f507bb41a11 100644 --- a/cmake/cuda.cmake +++ b/cmake/cuda.cmake @@ -175,7 +175,10 @@ list(APPEND CUDA_NVCC_FLAGS "-std=c++11") list(APPEND CUDA_NVCC_FLAGS "-Xcompiler -fPIC") endif(NOT WIN32) -list(APPEND CUDA_NVCC_FLAGS "--use_fast_math") +if(WITH_FAST_MATH) + # Make use of fast math library. https://docs.nvidia.com/cuda/cuda-compiler-driver-nvcc/index.html + list(APPEND CUDA_NVCC_FLAGS "--use_fast_math") +endif() # in cuda9, suppress cuda warning on eigen list(APPEND CUDA_NVCC_FLAGS "-w") # Set :expt-relaxed-constexpr to suppress Eigen warnings diff --git a/cmake/external/eigen.cmake b/cmake/external/eigen.cmake index e029300eee9..573ad5e5f06 100644 --- a/cmake/external/eigen.cmake +++ b/cmake/external/eigen.cmake @@ -3,6 +3,14 @@ INCLUDE(ExternalProject) SET(EIGEN_SOURCE_DIR ${THIRD_PARTY_PATH}/eigen3) SET(EIGEN_INCLUDE_DIR ${EIGEN_SOURCE_DIR}/src/extern_eigen3) INCLUDE_DIRECTORIES(${EIGEN_INCLUDE_DIR}) +if(NOT WITH_FAST_MATH) + # EIGEN_FAST_MATH: https://eigen.tuxfamily.org/dox/TopicPreprocessorDirectives.html + # enables some optimizations which might affect the accuracy of the result. + # This currently enables the SSE vectorization of sin() and cos(), + # and speedups sqrt() for single precision. + # Defined to 1 by default. Define it to 0 to disable. + add_definitions(-DEIGEN_FAST_MATH=0) +endif() if(WITH_AMD_GPU) ExternalProject_Add( diff --git a/paddle/fluid/inference/tests/api/analyzer_resnet50_tester.cc b/paddle/fluid/inference/tests/api/analyzer_resnet50_tester.cc index 290fb007d8b..8add7a59da6 100644 --- a/paddle/fluid/inference/tests/api/analyzer_resnet50_tester.cc +++ b/paddle/fluid/inference/tests/api/analyzer_resnet50_tester.cc @@ -27,6 +27,9 @@ void SetConfig(AnalysisConfig *cfg) { cfg->device = 0; cfg->enable_ir_optim = true; cfg->specify_input_name = true; +#ifdef PADDLE_WITH_MKLDNN + cfg->_use_mkldnn = true; +#endif } void SetInput(std::vector> *inputs) { -- GitLab