使用Agner Fog向量类在MSVC编译下性能骤降的问题求助

我在测试原生SSE/AVX函数与Agner Fog's vector class（我的CPU支持SSE和AVX指令）的性能差异时，GNU编译器（版本12.2）下结果完全符合预期，但切换到MSVC（VS2019/Clion环境）编译后，所有向量类实现的性能比原生SSE/AVX函数慢10倍甚至更多。

我的CMake编译配置如下：

cmake_minimum_required(VERSION 3.21)
project(testSIMD)

set(CMAKE_MODULE_PATH ${CMAKE_MODULE_PATH} "${CMAKE_SOURCE_DIR}/external_modules/cmakeFind/")

set(CMAKE_CXX_STANDARD 17)

set(USE_VECTORIZE ON)  

if (NOT VECTOR_TYPE)        
    set(VECTOR_TYPE full_vectorize)
endif (NOT VECTOR_TYPE)
if (VECTOR_TYPE STREQUAL "none")
    set(VECTOR_TYPE none)
endif (VECTOR_TYPE STREQUAL "none")
if (VECTOR_TYPE STREQUAL "default")
    set(VECTOR_TYPE default)
endif (VECTOR_TYPE STREQUAL "default")


# Set vectorization flags for a few compilers
set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS}")
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS}")


find_package(VectorFOG REQUIRED)

#Pushing vector flags onto compiler flags for try compiles
set(CMAKE_C_FLAGS_SAVE ${CMAKE_C_FLAGS})
set(CMAKE_CXX_FLAGS_SAVE ${CMAKE_CXX_FLAGS})
set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} ${VECTOR_C_FLAGS}")
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${VECTOR_CXX_FLAGS}")

try_compile(HAVE_FOG_VECTOR_CLASS "${CMAKE_BINARY_DIR}" "${CMAKE_SOURCE_DIR}/fogvectorclasstest.cpp")

if(HAVE_FOG_VECTOR_CLASS)
    message(STATUS "Trying Fog Vector Class -- works")
    add_definitions(-DHAVE_FOG_VECTOR_CLASS)
    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -DMAX_VECTOR_SIZE=512")
    set (FILE_LIST ${FILE_LIST}
            kahan_fog_vector.cpp
            kahan_fog_vector8.cpp)
else()
    message(STATUS "Trying Fog Vector Class -- fails")
endif()


# Set vectorization flags for GNU compiler
if (CMAKE_CXX_COMPILER_ID STREQUAL "GNU") # using GCC
    if (USE_VECTORIZE)
        set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -g -O3 -fstrict-aliasing -fopenmp-simd -march=native -mtune=native -ffast-math -ftree-vectorize -fopt-info-vec-optimized")
    else ()
        set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -g -O3 -fstrict-aliasing -march=native -mtune=native -ffast-math -fno-tree-vectorize -fopt-info-vec-optimized")
    endif ()
    if ("${CMAKE_CXX_COMPILER_VERSION}" VERSION_GREATER "7.4.0")    
        set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -mprefer-vector-width=512")
    endif ("${CMAKE_CXX_COMPILER_VERSION}" VERSION_GREATER "7.4.0")
elseif ("${CMAKE_CXX_COMPILER_ID}" STREQUAL "Clang") # using Clang
    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -g -O3 -fstrict-aliasing -fvectorize -march=native -mtune=native -ffast-math -Rpass-analysis=loop-vectorize")
elseif ("${CMAKE_CXX_COMPILER_ID}" STREQUAL "MSVC")
    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} /O3")
    add_compile_options(/arch:AVX2)    
    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} /Qvec-report:2")
endif ()

可能的原因及修复方案 #

• 缺少快速数学优化选项：
  MSVC默认的浮点精度策略会限制向量优化，需添加/fp:fast选项（对应GCC的-ffast-math），这是Agner Fog向量类发挥性能的关键。修改MSVC分支的配置：

  elseif ("${CMAKE_CXX_COMPILER_ID}" STREQUAL "MSVC")
      set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} /O3 /fp:fast")
      add_compile_options(/arch:AVX2)    
      set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} /Qvec-report:2")
  endif ()
  

• MAX_VECTOR_SIZE定义格式错误：
  MSVC使用/D而非GCC的-D来定义宏，需要针对编译器区分宏定义方式：

  if(HAVE_FOG_VECTOR_CLASS)
      message(STATUS "Trying Fog Vector Class -- works")
      add_definitions(-DHAVE_FOG_VECTOR_CLASS)
      if (CMAKE_CXX_COMPILER_ID STREQUAL "MSVC")
          set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} /DMAX_VECTOR_SIZE=512")
      else()
          set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -DMAX_VECTOR_SIZE=512")
      endif()
      set (FILE_LIST ${FILE_LIST}
              kahan_fog_vector.cpp
              kahan_fog_vector8.cpp)
  else()
      message(STATUS "Trying Fog Vector Class -- fails")
  endif()
  

• VectorFOG的编译标志未保留：
  try_compile后恢复了原编译标志，导致VECTOR_CXX_FLAGS没有应用到最终编译中，需在MSVC分支中添加：

  set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${VECTOR_CXX_FLAGS}")
  

• 检查向量化报告：
  查看/Qvec-report:2的输出，确认向量类相关代码是否被成功向量化。如果未被向量化，可在循环前添加#pragma loop(enable_vectorization)强制MSVC进行向量化。

内容的提问来源于stack exchange，提问作者hamster_watcher