2022年7月版——在“哪吒”上跑AI 全志D1 ncnn框架移植笔记

giao LV 5

ncnn框架是当前边缘计算的主流框架，关于全志D1运行ncnn框架的笔记，nihui大佬曾经于2021年发表了相关的帖子。但是目前工具链和ncnn版本有所变化，因此在大佬的基础上将，完整过程记录如下。

相关链接：
1.https://bbs.aw-ol.com/topic/705/
2.https://zhuanlan.zhihu.com/p/386312071?utm_source=qq&utm_medium=social&utm_oi=872955404320141312

1 准备交叉编译工具链

目前使用的交叉工具链是v2.2.6版本，具体下载链接见：https://occ.t-head.cn/community/download?id=4046947553902661632

下载这个文件：Xuantie-900-gcc-linux-5.10.4-glibc-x86_64-V2.2.6-20220516.tar.gz

tar -xf Xuantie-900-gcc-linux-5.10.4-glibc-x86_64-V2.2.6-20220516.tar.gz
export RISCV_ROOT_PATH=/home/{你的用户名}/Xuantie-900-gcc-linux-5.10.4-glibc-x86_64-V2.2.6

2 下载和编译ncnn

这边的toolchain file选用c906-v226.toolchain.cmake，其余的不用修改，在toolchain file里面都已经改好了。其中可能有路径问题，根据ncnn文件夹，实际所在的路径修改即可。

git clone https://github.com/Tencent/ncnn.git
cd ncnn
mkdir build-c906
cd build-c906
cmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/c906-v226.toolchain.cmake -DCMAKE_BUILD_TYPE=release -DNCNN_OPENMP=OFF -DNCNN_THREADS=OFF -DNCNN_RUNTIME_CPU=OFF -DNCNN_RVV=ON -DNCNN_SIMPLEOCV=ON -DNCNN_BUILD_EXAMPLES=ON ..
make -j4

3 测试benchncnn

目前2.0版本的TinaLinux 执行 ncnn 程序已经不会有问题

将 ncnn/build-c906/benchmark/benchncnn 和 ncnn/benchmark/*.param 拷贝到 d1开发板上，路径结构如下：

chmod 0777 benchncnn 
./benchncnn 4 1 0 -1 0

oot@TinaLinux:~/ncnn# ./benchncnn 4 1 0 -1 0
syscall error -1
loop_count = 4
num_threads = 1
powersave = 0
gpu_device = -1
cooling_down = 0
fopen squeezenet.param failed
network graph not ready
Segmentation fault
root@TinaLinux:~/ncnn# ./benchncnn 4 1 0 -1 0
syscall error -1
loop_count = 4
num_threads = 1
powersave = 0
gpu_device = -1
cooling_down = 0
          squeezenet  min =  382.77  max =  385.92  avg =  384.66
     squeezenet_int8  min = 8975.18  max = 8979.84  avg = 8977.72
           mobilenet  min =  702.76  max =  704.76  avg =  703.74
      mobilenet_int8  min = 29501.19  max = 29546.07  avg = 29514.59
        mobilenet_v2  min =  451.34  max =  452.27  avg =  451.78
        mobilenet_v3  min =  358.12  max =  359.24  avg =  358.42
          shufflenet  min =  527.04  max =  528.48  avg =  527.69
       shufflenet_v2  min =  310.96  max =  312.30  avg =  311.59
             mnasnet  min =  435.13  max =  437.02  avg =  435.94
     proxylessnasnet  min =  473.05  max =  476.37  avg =  475.19
     efficientnet_b0  min =  601.75  max =  610.41  avg =  606.18
   efficientnetv2_b0  min =  859.37  max =  861.75  avg =  860.67
        regnety_400m  min =  653.71  max =  654.24  avg =  653.88
           blazeface  min =  154.08  max =  154.67  avg =  154.38
           googlenet  min = 1404.80  max = 1409.50  avg = 1407.33
      googlenet_int8  min = 42586.56  max = 42608.04  avg = 42597.36
            resnet18  min = 1047.81  max = 1050.59  avg = 1049.40
       resnet18_int8  min = 45011.41  max = 45185.10  avg = 45098.12
             alexnet  min =  955.34  max =  956.98  avg =  956.06
               vgg16  min = 5355.40  max = 5356.05  avg = 5355.75

如果一切运行正常，则会出现上述信息。