rk3399+ubuntu：基于深度学习框架ncnn对比clang和gcc性能

安装clang

apt-get install clang

编译方式

修改toolchains/aarch64-linux-gnu.toolchain.cmake，内容如下：

# set cross-compiled system type, it's better not use the type which cmake cannot recognized.
SET ( CMAKE_SYSTEM_NAME Linux )
SET ( CMAKE_SYSTEM_PROCESSOR aarch64 )

# 注意这里的编译器改为clang
SET ( CMAKE_C_COMPILER "clang" )
SET ( CMAKE_CXX_COMPILER "clang++" )

# set searching rules for cross-compiler
SET ( CMAKE_FIND_ROOT_PATH_MODE_PROGRAM NEVER )
SET ( CMAKE_FIND_ROOT_PATH_MODE_LIBRARY ONLY )
SET ( CMAKE_FIND_ROOT_PATH_MODE_INCLUDE ONLY )

# set ${CMAKE_C_FLAGS} and ${CMAKE_CXX_FLAGS}flag for cross-compiled process
SET ( CMAKE_CXX_FLAGS "-std=c++11 -march=armv8-a -fopenmp ${CMAKE_CXX_FLAGS}" )

# other settings
add_definitions(-D__ARM_NEON)
add_definitions(-D__ANDROID__)
SET ( ANDROID true)

然后：

cd $(ncnn root)
mkdir build
cd build
cmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/aarch64-linux-gnu-clang.toolchain.cmake ../
make -j6

编译完毕，把$(ncnnroot)/build/benchmark/benchmark复制到$(ncnn root)/benchmark/目录下

简单性能测试

clang编译的ncnn在小核和大核上都有肉眼可见的优势。特别是小核上，能优化的和大核差不多。大核上优化空间反而比较小。这点很神奇。

个别模型clang上会差一点。

综合来说，用clang不亏。

树莓派的性能在这里基本不能看了。