编译 NCNN
1. 编译 CMake
NCNN 编译要求 CMake 版本大于 3.13（当前 2023-02-02），而 Ubuntu 18 源很快停止维护，将维持 3.10 版本，所以我们需要从源码安装。
卸载 CMake：
bash
apt-get remove cmake安装 OpenSSL 库：
bash
sudo apt-get install openssl libssl-dev在 CMake: download 页面找到最新版本，下载源代码：
bash
wget https://github.com/Kitware/CMake/releases/download/v3.25.2/cmake-3.25.2.tar.gz
tar xzvf cmake-3.25.2.tar.gz
cd cmake-3.25.2编译安装：
bash
./bootstrap
make -j $(nproc)
sudo make install下面测试 CMake 版本：
bash
cmake --version2. 编译 NCNN
2.1 正常编译
克隆仓库，如果克隆失败建议使用代理：
bash
git clone https://github.com/Tencent/ncnn.git
cd ncnn
git submodule update --init现在准备编译：
bash
mkdir -p build
cd build
cmake \
-DCMAKE_BUILD_TYPE=Release \
-DCMAKE_TOOLCHAIN_FILE=../toolchains/jetson.toolchain.cmake \
-DNCNN_VULKAN=ON \
-DNCNN_SIMPLEOCV=ON \
-DNCNN_BUILD_EXAMPLES=ON \
..
make -j $(nproc)CUDA 错误
TODO 在未指定 -DNCNN_SIMPLEOCV=ON 时将使用系统的 OpenCV，如果系统的 OpenCV 包含 CUDA 支持，那么此时需要指定 CUDA 的位置信息。但是通过各种方法仍然不能编译通过，出现找不到 CUDA 库信息的报告。当前使用 CUDA 10.2。
2.2 使用 Clang 套件
需要编辑 jetson.toolchain.cmake 来禁用默认的 GCC 套件：
cmake
# set(CMAKE_C_COMPILER "aarch64-linux-gnu-gcc")
# set(CMAKE_CXX_COMPILER "aarch64-linux-gnu-g++")
# set ${CMAKE_C_FLAGS} and ${CMAKE_CXX_FLAGS}flag for cross-compiled process
# -march=armv8-a could work on Jetson, but will compile without some extra cpu features
# set(CMAKE_C_FLAGS "-march=native ${CMAKE_C_FLAGS}")
# set(CMAKE_CXX_FLAGS "-march=native ${CMAKE_CXX_FLAGS}")下面使用 Clang 套件编译：
bash
cmake \
-DCMAKE_CXX_COMPILER=$(which clang++) \
-DCMAKE_C_COMPILER=$(which clang) \
-DCMAKE_LINKER=$(which lld) \
-DCMAKE_BUILD_TYPE=Release \
-DCMAKE_TOOLCHAIN_FILE=../toolchains/jetson.toolchain.cmake \
-DNCNN_VULKAN=ON \
-DNCNN_SIMPLEOCV=ON \
-DNCNN_BUILD_EXAMPLES=ON \
..
make -j $(nproc)段错误
TODO 该测试使用 Clang 套件，并包含 Vulkan 支持，执行任何测试程序输出结果后不能正确停止，抛出段错误：
bash
cd ../examples
../build/examples/squeezenet ../images/128-ncnn.png输出：
log
[0 NVIDIA Tegra X1 (nvgpu)] queueC=0[16] queueG=0[16] queueT=0[16]
[0 NVIDIA Tegra X1 (nvgpu)] bugsbn1=0 bugbilz=0 bugcopc=0 bugihfa=0
[0 NVIDIA Tegra X1 (nvgpu)] fp16-p/s/a=1/1/1 int8-p/s/a=1/1/1
[0 NVIDIA Tegra X1 (nvgpu)] subgroup=32 basic=1 vote=1 ballot=1 shuffle=1
532 = 0.138184
920 = 0.061310
598 = 0.049622
Segmentation fault (core dumped)暂时未解决。猜测和 Vulkan 有关，未使用 Vulkan 的 CPU 版本没有此错误。
3. 基准测试
bash
cd ../benchmark
../build/benchmark/benchncnn 10 $(nproc) 0 0含 Vulkan 的版本在 Jetson Nano（B01）上的测试结果如下（2023-02-02）：
log
[0 NVIDIA Tegra X1 (nvgpu)] queueC=0[16] queueG=0[16] queueT=0[16]
[0 NVIDIA Tegra X1 (nvgpu)] bugsbn1=0 bugbilz=0 bugcopc=0 bugihfa=0
[0 NVIDIA Tegra X1 (nvgpu)] fp16-p/s/a=1/1/1 int8-p/s/a=1/1/1
[0 NVIDIA Tegra X1 (nvgpu)] subgroup=32 basic=1 vote=1 ballot=1 shuffle=1
loop_count = 10
num_threads = 4
powersave = 0
gpu_device = 0
cooling_down = 1
squeezenet min = 9.32 max = 9.46 avg = 9.37
squeezenet_int8 min = 25.15 max = 25.30 avg = 25.23
mobilenet min = 15.72 max = 15.86 avg = 15.79
mobilenet_int8 min = 27.42 max = 27.72 avg = 27.57
mobilenet_v2 min = 12.43 max = 12.53 avg = 12.48
mobilenet_v3 min = 13.13 max = 13.26 avg = 13.19
shufflenet min = 10.93 max = 11.01 avg = 10.96
shufflenet_v2 min = 9.82 max = 10.01 avg = 9.86
mnasnet min = 13.01 max = 13.25 avg = 13.07
proxylessnasnet min = 14.27 max = 14.41 avg = 14.32
efficientnet_b0 min = 20.98 max = 21.22 avg = 21.09
efficientnetv2_b0 min = 38.90 max = 39.22 avg = 39.02
regnety_400m min = 17.12 max = 17.30 avg = 17.19
blazeface min = 3.24 max = 3.29 avg = 3.26
googlenet min = 30.27 max = 30.49 avg = 30.34
googlenet_int8 min = 79.61 max = 80.18 avg = 79.85
resnet18 min = 27.47 max = 27.57 avg = 27.52
resnet18_int8 min = 59.07 max = 59.74 avg = 59.40
alexnet min = 48.73 max = 49.26 avg = 48.96
vgg16 min = 142.72 max = 144.31 avg = 143.05
vgg16_int8 min = 286.85 max = 288.51 avg = 287.74
resnet50 min = 71.55 max = 71.96 avg = 71.71
resnet50_int8 min = 148.30 max = 149.12 avg = 148.60
squeezenet_ssd min = 32.92 max = 33.23 avg = 33.05
squeezenet_ssd_int8 min = 65.86 max = 66.57 avg = 66.14
mobilenet_ssd min = 34.70 max = 35.01 avg = 34.86
mobilenet_ssd_int8 min = 56.45 max = 56.75 avg = 56.63
mobilenet_yolo min = 70.04 max = 70.67 avg = 70.39
mobilenetv2_yolov3 min = 45.47 max = 45.76 avg = 45.61
yolov4-tiny min = 66.89 max = 67.40 avg = 67.12
nanodet_m min = 18.73 max = 18.88 avg = 18.81
yolo-fastest-1.1 min = 11.22 max = 11.39 avg = 11.30
yolo-fastestv2 min = 9.76 max = 9.89 avg = 9.81
vision_transformer min = 2682.28 max = 2697.72 avg = 2689.92
FastestDet min = 9.94 max = 10.10 avg = 10.03不含 Vulkan 的 CPU 版本在 Jetson Nano 上的测试结果如下（2023-02-02）：
log
loop_count = 10
num_threads = 4
powersave = 0
gpu_device = 0
cooling_down = 1
squeezenet min = 28.98 max = 29.28 avg = 29.11
squeezenet_int8 min = 25.10 max = 25.53 avg = 25.23
mobilenet min = 39.80 max = 40.09 avg = 39.96
mobilenet_int8 min = 27.29 max = 27.63 avg = 27.49
mobilenet_v2 min = 31.51 max = 31.98 avg = 31.65
mobilenet_v3 min = 26.49 max = 26.86 avg = 26.69
shufflenet min = 20.04 max = 20.58 avg = 20.26
shufflenet_v2 min = 17.02 max = 17.60 avg = 17.19
mnasnet min = 29.49 max = 30.02 avg = 29.62
proxylessnasnet min = 34.17 max = 34.86 avg = 34.42
efficientnet_b0 min = 59.06 max = 59.87 avg = 59.28
efficientnetv2_b0 min = 65.72 max = 91.38 avg = 68.73
regnety_400m min = 56.01 max = 56.48 avg = 56.21
blazeface min = 5.67 max = 5.84 avg = 5.74
googlenet min = 90.85 max = 91.61 avg = 91.17
googlenet_int8 min = 78.91 max = 79.58 avg = 79.19
resnet18 min = 80.18 max = 82.15 avg = 80.64
resnet18_int8 min = 58.98 max = 77.46 avg = 61.45
alexnet min = 60.00 max = 61.35 avg = 60.46
vgg16 min = 377.00 max = 382.35 avg = 379.10
vgg16_int8 min = 285.44 max = 288.74 avg = 286.98
resnet50 min = 196.79 max = 197.87 avg = 197.23
resnet50_int8 min = 148.09 max = 149.84 avg = 148.88
squeezenet_ssd min = 82.21 max = 82.94 avg = 82.54
squeezenet_ssd_int8 min = 65.03 max = 65.51 avg = 65.25
mobilenet_ssd min = 83.00 max = 83.69 avg = 83.28
mobilenet_ssd_int8 min = 56.43 max = 56.95 avg = 56.65
mobilenet_yolo min = 185.10 max = 186.02 avg = 185.55
mobilenetv2_yolov3 min = 111.98 max = 112.51 avg = 112.23
yolov4-tiny min = 156.89 max = 158.15 avg = 157.47
nanodet_m min = 40.70 max = 41.68 avg = 41.11
yolo-fastest-1.1 min = 22.65 max = 22.91 avg = 22.79
yolo-fastestv2 min = 18.55 max = 18.95 avg = 18.65
vision_transformer min = 3163.91 max = 3259.01 avg = 3206.24
FastestDet min = 19.80 max = 20.10 avg = 19.90下面是详细比较（平均值）：
| 网络名称 | GPU(Vulkan) 平均 | CPU 平均 |
|---|---|---|
| squeezenet | 9.37 | 29.11 |
| squeezenet_int8 | 25.23 | 25.23 |
| mobilenet | 15.79 | 39.96 |
| mobilenet_int8 | 27.57 | 27.49 |
| mobilenet_v2 | 12.48 | 31.65 |
| mobilenet_v3 | 13.19 | 26.69 |
| shufflenet | 10.96 | 20.26 |
| shufflenet_v2 | 9.86 | 17.19 |
| mnasnet | 13.07 | 29.62 |
| proxylessnasnet | 14.32 | 34.42 |
| efficientnet_b0 | 21.09 | 59.28 |
| efficientnetv2_b0 | 39.02 | 68.73 |
| regnety_400m | 17.19 | 56.21 |
| blazeface | 3.26 | 5.74 |
| googlenet | 30.34 | 91.17 |
| googlenet_int8 | 79.85 | 79.19 |
| resnet18 | 27.52 | 80.64 |
| resnet18_int8 | 59.40 | 61.45 |
| alexnet | 48.96 | 60.46 |
| vgg16 | 143.05 | 379.10 |
| vgg16_int8 | 287.74 | 286.98 |
| resnet50 | 71.71 | 197.23 |
| resnet50_int8 | 148.60 | 148.88 |
| squeezenet_ssd | 33.05 | 82.54 |
| squeezenet_ssd_int8 | 66.14 | 65.25 |
| mobilenet_ssd | 34.86 | 83.28 |
| mobilenet_ssd_int8 | 56.63 | 56.65 |
| mobilenet_yolo | 70.39 | 185.55 |
| mobilenetv2_yolov3 | 45.61 | 112.23 |
| yolov4-tiny | 67.12 | 157.47 |
| nanodet_m | 18.81 | 41.11 |
| yolo-fastest-1.1 | 11.30 | 22.79 |
| yolo-fastestv2 | 9.81 | 18.65 |
| vision_transformer | 2689.92 | 3206.24 |
| FastestDet | 10.03 | 19.90 |