arm平台上的MNN编译与运行

0.成果物

直接获取成果物见：https://download.csdn.net/download/u012824853/87867665

以下为编译、运行过程

1.编译准备

在GitHub - alibaba/MNN: MNN is a blazing fast, lightweight deep learning framework, battle-tested by business-critical use cases in Alibaba下载code->downloadZIP

unzip MNN-master.zipcd MNN-mastervim CMakeLists.txt

将其中的MNN_BUILD_DEMO的option改为ON

2.生成Makefile

查看arm的架构，如armv7、v8等，赋值给CMAKE_SYSTEM_PROCESSOR

确定编译结果保存文件的路径赋值给CMAKE_INSTALL_PREFIX

确定编译工具链的位置，赋值给CMAKE_C_COMPILER、CMAKE_CXX_COMPILER

whereis arm-linux-gnueabihf-gcc
arm-linux-gnueabihf-gcc: /root/arm-linux-compiler/gcc-linaro-12.2.1-2023.01-x86_64_arm-linux-gnueabihf/bin/arm-linux-gnueabihf-gcc

whereis arm-linux-gnueabihf-g++
arm-linux-gnueabihf-g++: /root/arm-linux-compiler/gcc-linaro-12.2.1-2023.01-x86_64_arm-linux-gnueabihf/bin/arm-linux-gnueabihf-g++

mkdir install && mkdir build && cd build
cmake .. \
-DCMAKE_BUILD_TYPE=Release \
-DMNN_BUILD_DEMO=ON \
-DMNN_USE_INT8_FAST=true \
-DMNN_BUILD_TEST=true\
-DMNN_OPENCL=ON \
-DCMAKE_SYSTEM_NAME=Linux \
-DCMAKE_SYSTEM_VERSION=1 \
-DCMAKE_SYSTEM_PROCESSOR=armv7 \
-DCMAKE_INSTALL_PREFIX=/root/mnn/MNN-master/install \
-DCMAKE_C_COMPILER=/root/arm-linux-compiler/gcc-linaro-12.2.1-2023.01-x86_64_arm-linux-gnueabihf/bin/arm-linux-gnueabihf-gcc \
-DCMAKE_CXX_COMPILER=/root/arm-linux-compiler/gcc-linaro-12.2.1-2023.01-x86_64_arm-linux-gnueabihf/bin/arm-linux-gnueabihf-g++

3.编译

执行make

如果遇到报错：

/root/mnn/MNN-master/source/math/Vec.hpp:342:71: error: cannot convert ‘int16x8_t’ to ‘int8x16_t’342 |         auto m0m1 = vtrnq_s8(reinterpret_cast<int8x16_t>(vec0.value), reinterpret_cast<int16x8_t>(vec1.value));|                                                                       ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~|                                                                       ||                                                                       int16x8_t
In file included from /root/mnn/MNN-master/source/math/Vec.hpp:15:
/root/arm-linux-compiler/gcc-linaro-12.2.1-2023.01-x86_64_arm-linux-gnueabihf/lib/gcc/arm-linux-gnueabihf/12.2.1/include/arm_neon.h:9532:36: note:   initializing argument 2 of ‘int8x16x2_t vtrnq_s8(int8x16_t, int8x16_t)’9532 | vtrnq_s8 (int8x16_t __a, int8x16_t __b)|                          ~~~~~~~~~~^~~
/root/mnn/MNN-master/source/math/Vec.hpp:343:71: error: cannot convert ‘int16x8_t’ to ‘int8x16_t’343 |         auto m2m3 = vtrnq_s8(reinterpret_cast<int8x16_t>(vec2.value), reinterpret_cast<int16x8_t>(vec3.value));|                                                                       ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~|                                                                       ||                                                                       int16x8_t

则将/root/mnn/MNN-master/source/math/Vec.hpp的342、343行

从

auto m0m1 = vtrnq_s8(reinterpret_cast<int8x16_t>(vec0.value), reinterpret_cast<int16x8_t>(vec1.value));
auto m2m3 = vtrnq_s8(reinterpret_cast<int8x16_t>(vec2.value), reinterpret_cast<int16x8_t>(vec3.value));

改为

auto m0m1 = vtrnq_s8(reinterpret_cast<int8x16_t>(vec0.value), reinterpret_cast<int8x16_t>(vec1.value));
auto m2m3 = vtrnq_s8(reinterpret_cast<int8x16_t>(vec2.value), reinterpret_cast<int8x16_t>(vec3.value));

重新make -j4

如果发生make进度卡住的情况，则不要加-j4

成果物是libMNN.so

4.模型准备

在Model Zoo · 语雀中找合适的模型，比如：GitHub - Linzaer/Ultra-Light-Fast-Generic-Face-Detector-1MB: 💎1MB lightweight face detection model (1MB轻量级人脸检测模型)

code->downloadZIP，解压

将刚才编译的build的目录/MNN-master/build下的libMNN.so拷贝到Ultra-Light-Fast-Generic-Face-Detector-1MB-master/MNN/mnn/lib下

在Ultra-Light-Fast-Generic-Face-Detector-1MB-master/MNN/CMakeLists.txt中添加：

set(OpenCV_DIR /opencv-3.4.1/build)

opencv3.4.1在ARM平台上的编译方法见：

opencv arm交叉编译与仿真验证详细流程_opencv交叉编译arm_I_am_Damon的博客-CSDN博客

将Ultra-Light-Fast-Generic-Face-Detector-1MB-master/MNN/src/main.cpp中的cv:imshow和cv::waitkey注释掉，因为嵌入式端一般不能显示图像

在Ultra-Light-Fast-Generic-Face-Detector-1MB-master/MNN/下执行：

mkdir build;
cd build;
cmake .. -DCMAKE_SYSTEM_PROCESSOR=armv7 \
-DCMAKE_C_COMPILER=/root/arm-linux-compiler/gcc-linaro-12.2.1-2023.01-x86_64_arm-linux-gnueabihf/bin/arm-linux-gnueabihf-gcc \
-DCMAKE_CXX_COMPILER=/root/arm-linux-compiler/gcc-linaro-12.2.1-2023.01-x86_64_arm-linux-gnueabihf/bin/arm-linux-gnueabihf-g++
make -j4;

可生成可执行程序：Ultra-face-mnn

5.板端运行

将opencv生成的所有动态库拷贝到执行设备（arm一般是嵌入式板子）上的/lib/目录下

将libMNN.so也拷贝到执行设备上的/lib/目录下

然后按照Ultra-Light-Fast-Generic-Face-Detector-1MB-master/MNN/README.md执行测试用例：

./Ultra-face-mnn ./model/version-RFB/RFB-320.mnn ./imgs/1.jpg

即可生成代码检测框的图片: