Hi3516编译 OpenCV

落日映苍穹つ 2023-07-01 11:59 6阅读 0赞

为Hi35xx处理器编译 OpenCV 时会遇到 NEON 无法启用的问题，报错信息如下：

CMake Error at cmake/OpenCVCompilerOptimizations.cmake:535 (message):
      Required baseline optimization is not supported: NEON
      (CPU_BASELINE_REQUIRE=;VFPV3;NEON)
    Call Stack (most recent call first):
      cmake/OpenCVCompilerOptions.cmake:292 (include)
      CMakeLists.txt:594 (include)

cpu 支持的特性仅检测到了 VFV3而没有 NEON。

--   CPU/HW features:
    --     Baseline:                    VFPV3
    --       requested:                 DETECT
    --       required:                  VFPV3 NEON

报错出现在 [OpenCVCompilerOptimizations.cmake][] 中， [OpenCVUtils.cmake][] 会对 cpu 特性进行检查。即使在跳过检查强制启用 NEON 也不行。

几经测试发现，编译时添加`-mfloat-abi=softfp`选项可解决该问题。

VFP 是 ARM 的“矢量浮点”单元（Floating Point Unit），执行硬件加速的浮点计算。由于历史原因并且为了与 ARM RVCT 套件匹配，GCC 的 FPU 和 ABI 选项并非完全正交。`-mfloat-abi =`选项控制 ABI 以及是否可以使用浮点指令。可用的选项有：

*  `soft`：全软浮点，编译器不会生成真实的FPU指令，并且忽略`-mfpu =`。FPU 操作由编译器模拟生成函数调用以在整数寄存器中传递 FP 参数（float, double）（一个用于浮点，一对用于双精度寄存器）。
 *  `softfp`：使用软浮点 ABI 的硬件浮点。编译器可以根据选择的FPU类型（`-mfpu =`）确定何时以及是否生成模拟的或实际的 FPU 指令。
 *  `hard`：完整的硬件浮点，FP 参数直接在 FPU 寄存器中传递。如果没有在`-mfpu=`中定义的FPU（或该 FPU的超集），则无法运行。

## build.sh ##

编译脚本如下：

#!/bin/bash
    set -e # Exit immediately if a command exits with a non-zero status.
    BuildDir=arm_build
    ToolChain=/path/to/arm-himix200-linux
    if [ ! -d "$BuildDir" ]; then
      # Take action if $BuildDir doesn‘t exists. 
      echo "create ${BuildDir}..."
      mkdir -p ${BuildDir}
    fi
    cd ${BuildDir}
    
    echo "building OpenCV3"
    
    cmake -D CMAKE_BUILD_TYPE=RELEASE \
      -D BUILD_SHARED_LIBS=ON \
      -D CMAKE_FIND_ROOT_PATH=${ToolChain}/target\
      -D CMAKE_TOOLCHAIN_FILE=../platforms/linux/arm-gnueabi.toolchain.cmake\
      -D CMAKE_C_COMPILER=${ToolChain}/bin/arm-himix200-linux-gcc-6.3.0\
      -D CMAKE_CXX_COMPILER=${ToolChain}/bin/arm-himix200-linux-g++\
      -D CMAKE_CXX_FLAGS="-mfloat-abi=softfp"\
      -D CMAKE_C_FLAGS="-mfloat-abi=softfp"\
      -D CMAKE_INSTALL_PREFIX=${BuildDir} \
      -D WITH_TBB=ON \
      -D WITH_EIGEN=ON \
      -D BUILD_ZLIB=ON \
      -D BUILD_TIFF=ON \
      -D BUILD_JASPER=ON \
      -D BUILD_JPEG=ON \
      -D BUILD_PNG=ON \
      -D ENABLE_NEON=ON \
      -D ENABLE_VFPV3=ON \
      -D WITH_LIBV4L=ON \
      -D BUILD_opencv_python=OFF\
      -D ENABLE_PRECOMPILED_HEADERS=OFF \
      -D BUILD_EXAMPLES=OFF \
      -D BUILD_TESTS=OFF \
      -D BUILD_PERF_TESTS=OFF \
      -D BUILD_WITH_DEBUG_INFO=OFF \
      -D BUILD_DOCS=OFF \
      -D WITH_OPENCL=OFF \
      ../
    make -j$(nproc)

## 参考资料： ##

*  [Ccache (简体中文)][Ccache]
 *  [通过 ccache 改善协同构建时间][ccache]
 *  [cmake cross-compile for andrid meeting “Compiler doesn’t support baseline optimization flags”][cmake cross-compile for andrid meeting _Compiler doesn_t support baseline optimization flags]
 *  [ARM Cortex-A 处理器和 GCC 命令行][ARM Cortex-A _ GCC]
 *  [在编译器中使能了neon（-mfpu=neon）之后，还能同时使用fpu计算64bit浮点吗？][neon_-mfpu_neon_fpu_64bit]
 *  [HI3516的编译参数-mcpu=cortex-a7、-mfloat-abi=softfp和-mfpu=neon-vfpv4][HI3516_-mcpu_cortex-a7_-mfloat-abi_softfp_-mfpu_neon-vfpv4]
 *  [‘std::exception’ error when cross-compile OpenCV 4][std_exception_ error when cross-compile OpenCV 4]
 *  [Upgrading to G++ 4.8 - exception\_ptr.h Does not support exception propagation][Upgrading to G_ 4.8 - exception_ptr.h Does not support exception propagation]
 *  [G++ Link-time optimisation on Mac - compiler/linker bug?][G_ Link-time optimisation on Mac - compiler_linker bug]
 *  [cpu: cortex-m: compiling vectors.c with LTO triggers compiler bug \#5775][cpu_ cortex-m_ compiling vectors.c with LTO triggers compiler bug _5775]
 *  [移植opencv3.1.0到hi3516a][opencv3.1.0_hi3516a]
 *  [海思hi3516移植OpenCV2.4.9][hi3516_OpenCV2.4.9]
 *  [cmake使用教程（十）-关于file][cmake_-_file]
 *  [NEON™ Support in Compilation Tools][NEON_ Support in Compilation Tools]
 *  [\[dpdk-dev,v3,2/4\] arm: detect NEON cpu feature by checking \_\_ARM\_NEON][dpdk-dev_v3_2_4_ arm_ detect NEON cpu feature by checking _ARM_NEON]
 *  [ARM crypto instructions and \_\_ARM\_FEATURE\_CRYPTO macro][ARM crypto instructions and _ARM_FEATURE_CRYPTO macro]
 *  [Optimizing opencv 2.4 and 3.1 in Raspberry Pi 3][]
 *  [What is carotene\_o4t and why does it cause issues with compiling a program using opencv with Qt on Android?][What is carotene_o4t and why does it cause issues with compiling a program using opencv with Qt on Android]
 *  [Simple C++ Symbol Visibility Demo][Simple C_ Symbol Visibility Demo]
 *  [gcc的fvisibility选项 \#2][gcc_fvisibility_ _2]
 *  [控制符号的可见性][Link 1]
 *  [Why is the new C++ visibility support so useful?][Why is the new C_ visibility support so useful]
 *  [C++ -fvisibility=hidden -fvisibility-inlines-hidden][C_ -fvisibility_hidden -fvisibility-inlines-hidden]
 *  [target\_compile\_definitions和target\_compile\_options中第二个参数的含义][target_compile_definitions_target_compile_options]
 *  [target\_compile\_definitions for multiple CMake targets?][target_compile_definitions for multiple CMake targets]
 *  [GCC编译之如何控制共享文件导出符号][GCC]
 *  [Demystifying ARM Floating Point Compiler Options][]
 *  [liqi-c/tengine-nnie][liqi-c_tengine-nnie]
 *  [How To Cross-Compile Clang/LLVM using Clang/LLVM][How To Cross-Compile Clang_LLVM using Clang_LLVM]
 *  [\[ARM\]ARM平台处理器简介-ARMv7][ARM_ARM_-ARMv7]
 *  [Building OpenCV 3.4.6 for Xavier - is VFPV3 supported?][Building OpenCV 3.4.6 for Xavier - is VFPV3 supported]
 *  [ArmHardFloatPortVfpComparison][]
 *  [himix200.toolchain.cmake][]
 *  [SylixOS ARMv7A 处理器性能测试与改进][SylixOS ARMv7A]
 *  [移植OpenCV4.1.0到Hi3519AV100上][OpenCV4.1.0_Hi3519AV100]
 *  [OpenCV arm용으로 포팅하기][OpenCV arm_]
 *  [Building OpenCV for ARM Cortex-A8][]
 *  [Cross compilation for ARM based Linux systems][]
 *  [Cross compiling OpenCV 4 for Raspberry Pi and BeagleBone Black][]
 *  [armv7 build with neon on error, and resize will be too slow \#13886][armv7 build with neon on error_ and resize will be too slow _13886]
 *  [Optimizing opencv 3.4 in Raspberry Pi 3 using GStreamer][]
 *  [Enabling optimization options for Raspberry Pi 3 with OpenCV 4.0.0][]
 *  [HI3516D之Caffe移植][HI3516D_Caffe]
 *  [Failed to cross-compile with some"undefined reference" errors \#11878][Failed to cross-compile with some_undefined reference_ errors _11878]

[OpenCVCompilerOptimizations.cmake]: https://github.com/opencv/opencv/blob/master/cmake/OpenCVCompilerOptions.cmake
[OpenCVUtils.cmake]: https://github.com/opencv/opencv/blob/master/cmake/OpenCVUtils.cmake
[Ccache]: https://wiki.archlinux.org/index.php/Ccache_%28%E7%AE%80%E4%BD%93%E4%B8%AD%E6%96%87%29
[ccache]: https://www.ibm.com/developerworks/cn/linux/l-ccache.html
[cmake cross-compile for andrid meeting _Compiler doesn_t support baseline optimization flags]: https://answers.opencv.org/question/214636/cmake-cross-compile-for-andrid-meeting-compiler-doesnt-support-baseline-optimization-flags/
[ARM Cortex-A _ GCC]: https://community.arm.com/cn/b/blog/posts/arm-cortex-a-gcc
[neon_-mfpu_neon_fpu_64bit]: https://community.arm.com/cn/f/discussions/9924/neon--mfpu-neon-fpu-64bit
[HI3516_-mcpu_cortex-a7_-mfloat-abi_softfp_-mfpu_neon-vfpv4]: https://blog.csdn.net/u014470361/article/details/87931856
[std_exception_ error when cross-compile OpenCV 4]: https://stackoverflow.com/questions/55684583/stdexception-error-when-cross-compile-opencv-4
[Upgrading to G_ 4.8 - exception_ptr.h Does not support exception propagation]: https://stackoverflow.com/questions/45400971/upgrading-to-g-4-8-exception-ptr-h-does-not-support-exception-propagation/45479950#45479950
[G_ Link-time optimisation on Mac - compiler_linker bug]: https://stackoverflow.com/questions/32037069/g-link-time-optimisation-on-mac-compiler-linker-bug
[cpu_ cortex-m_ compiling vectors.c with LTO triggers compiler bug _5775]: https://github.com/RIOT-OS/RIOT/issues/5775
[opencv3.1.0_hi3516a]: https://blog.csdn.net/brightming/article/details/50563565
[hi3516_OpenCV2.4.9]: https://blog.csdn.net/ggqhit/article/details/82556190
[cmake_-_file]: https://juejin.im/post/5b3ecfef6fb9a04f8c5ebab5
[NEON_ Support in Compilation Tools]: http://infocenter.arm.com/help/topic/com.arm.doc.dht0004a/DHT0004A_neon_support_in_compilation_tools.pdf
[dpdk-dev_v3_2_4_ arm_ detect NEON cpu feature by checking _ARM_NEON]: http://patchwork.dpdk.org/patch/11607/
[ARM crypto instructions and _ARM_FEATURE_CRYPTO macro]: https://stackoverflow.com/questions/37236645/arm-crypto-instructions-and-arm-feature-crypto-macro?rq=1
[Optimizing opencv 2.4 and 3.1 in Raspberry Pi 3]: https://ebenezertechs.com/optimizing-opencv-2-4-and-3-1-in-raspberry-pi-3/
[What is carotene_o4t and why does it cause issues with compiling a program using opencv with Qt on Android]: https://stackoverflow.com/questions/41665655/what-is-carotene-o4t-and-why-does-it-cause-issues-with-compiling-a-program-using
[Simple C_ Symbol Visibility Demo]: https://labjack.com/news/simple-cpp-symbol-visibility-demo
[gcc_fvisibility_ _2]: https://github.com/chenpengcong/blog/issues/2
[Link 1]: http://my.huhoo.net/archives/2010/03/post_52.html
[Why is the new C_ visibility support so useful]: https://gcc.gnu.org/wiki/Visibility
[C_ -fvisibility_hidden -fvisibility-inlines-hidden]: https://stackoverflow.com/questions/3570355/c-fvisibility-hidden-fvisibility-inlines-hidden
[target_compile_definitions_target_compile_options]: https://blog.csdn.net/qq_34369618/article/details/96358204
[target_compile_definitions for multiple CMake targets]: https://stackoverflow.com/questions/47611319/target-compile-definitions-for-multiple-cmake-targets
[GCC]: https://www.cnblogs.com/persistentsnail/archive/2012/05/19/3294855.html
[Demystifying ARM Floating Point Compiler Options]: https://embeddedartistry.com/blog/2017/10/11/demystifying-arm-floating-point-compiler-options/
[liqi-c_tengine-nnie]: https://github.com/liqi-c/tengine-nnie
[How To Cross-Compile Clang_LLVM using Clang_LLVM]: https://llvm.org/docs/HowToCrossCompileLLVM.html
[ARM_ARM_-ARMv7]: http://zhaohongjian000.is-programmer.com/posts/34688.html
[Building OpenCV 3.4.6 for Xavier - is VFPV3 supported]: https://forums.developer.nvidia.com/t/building-opencv-3-4-6-for-xavier-is-vfpv3-supported/111542
[ArmHardFloatPortVfpComparison]: https://wiki.debian.org/ArmHardFloatPort/VfpComparison#A.22softfp.22
[himix200.toolchain.cmake]: https://github.com/Tencent/ncnn/blob/master/toolchains/himix200.toolchain.cmake
[SylixOS ARMv7A]: https://jiaojinxing.github.io/2015/07/17/SylixOS-ARMv7A-%E5%A4%84%E7%90%86%E5%99%A8%E6%80%A7%E8%83%BD%E6%B5%8B%E8%AF%95%E4%B8%8E%E6%94%B9%E8%BF%9B/
[OpenCV4.1.0_Hi3519AV100]: https://www.weibo.com/ttarticle/p/show?id=2313501000014390724396253284
[OpenCV arm_]: https://gist.github.com/RaySuhyunLee/394a116e006e7296c03a0f97c7ad37cd
[Building OpenCV for ARM Cortex-A8]: https://processors.wiki.ti.com/index.php/Building_OpenCV_for_ARM_Cortex-A8
[Cross compilation for ARM based Linux systems]: https://docs.opencv.org/master/d0/d76/tutorial_arm_crosscompile_with_cmake.html
[Cross compiling OpenCV 4 for Raspberry Pi and BeagleBone Black]: https://solarianprogrammer.com/2018/12/18/cross-compile-opencv-raspberry-pi-raspbian/
[armv7 build with neon on error_ and resize will be too slow _13886]: https://github.com/opencv/opencv/issues/13886
[Optimizing opencv 3.4 in Raspberry Pi 3 using GStreamer]: https://ebenezertechs.com/optimizing-opencv-3-4-in-raspberry-pi-3-using-gstreamer/
[Enabling optimization options for Raspberry Pi 3 with OpenCV 4.0.0]: https://answers.opencv.org/question/203740/enabling-optimization-options-for-raspberry-pi-3-with-opencv-400/
[HI3516D_Caffe]: https://blog.csdn.net/sanallen/article/details/78862387
[Failed to cross-compile with some_undefined reference_ errors _11878]: https://github.com/opencv/opencv/issues/11878