在 xr806 上用 ncnn 跑神经网络 mnist
0x0 介绍 xr806 和 ncnn
XR806是全志科技旗下子公司广州芯之联研发设计的一款支持WiFi和BLE的高集成度无线MCU芯片，支持鸿蒙L0系统
https://github.com/Tencent/ncnn
ncnn是腾讯开源的高性能神经网络推理框架，无第三方依赖，跨平台，具备非常好的可移植性，允许零拷贝加载模型节省内存
https://aijishu.com/e/1120000000256667
又能从极术社区白嫖开发板玩啦！
这次要把 ncnn 移植到这款 MCU 芯片
给白嫖就给移植，就像乐鑫科技送的 esp32c3 ^^:D
0x1 配环境，编译，烧录
配代理
shell
export http_proxy="http://127.0.0.1:4567"
export https_proxy="http://127.0.0.1:4567"下载 repo 工具，配 PATH
shell
mkdir -p bin
curl https://storage.googleapis.com/git-repo-downloads/repo > bin/repo
chmod a+rx bin/repo
PATH="`pwd`/bin:$PATH"下载 hb 工具，配 PATH
shell
pip install --user ohos-build
PATH="$HOME/.local/bin:$PATH"下载 openharmony 源码
有时候会遇到错误 error.GitError: manifest rev-list ('^e263a2d3e4557381c32fa31ce24384048d07d682', 'HEAD', '--'): fatal: bad object HEAD，多次重试即可
shell
repo init -u https://gitee.com/openharmony-sig/manifest.git -b OpenHarmony_1.0.1_release --no-repo-verify -m devboard_xr806.xml
repo sync -c
repo forall -c 'git lfs pull'下载 arm toolchain
shell
wget -c https://developer.arm.com/-/media/Files/downloads/gnu-rm/10-2020q4/gcc-arm-none-eabi-10-2020-q4-major-x86_64-linux.tar.bz2
tar -xf gcc-arm-none-eabi-10-2020-q4-major-x86_64-linux.tar.bz2更新 toolchain 信息
不要用最新版本的工具链，会编译失败
shell
sed -i "s@~/tools/gcc-arm-none-eabi-10-2020-q4-major/bin/arm-none-eabi-@`pwd`/gcc-arm-none-eabi-10-2020-q4-major/bin/arm-none-eabi-@g" device/xradio/xr806/liteos_m/config.gni
sed -i "s@~/tools/gcc-arm-none-eabi-10-2020-q4-major/bin@`pwd`/gcc-arm-none-eabi-10-2020-q4-major/bin@g" device/xradio/xr806/xr_skylark/gcc.mk修正 SDKconfig.gni 生成，修正 copy 脚本
shell
sed -i "s@open('\.{0}/@open('{0}/@g" device/xradio/xr806/xr_skylark/config.py
sed -i "s@open('\.{0}/@open('{0}/@g" device/xradio/xr806/libcopy.py编译默认的固件
menuconfig 直接 EXIT，后面再选择 wifi_skylark
shell
cd device/xradio/xr806/xr_skylark
cp project/demo/audio_demo/gcc/deconfig .config
make build_clean
make menuconfig
make lib -j
cd -
hb set
hb build -f烧录固件
usb 连接 xr806 开发板，此时会出现 /dev/ttyUSB0 设备（多个设备也可能是别的编号）
打开 device/xradio/xr806/xr_skylark/tools/settings.ini，修改固件路径，默认是 strImagePath = ../out/xr_system.img，iBaud 保持默认 iBaud = 921600
root 权限执行 phoenixMC 工具，开始烧录
shell
cd device/xradio/xr806/xr_skylark/tools
su
./phoenixMC串口查看输出
root 权限执行 tio，添加 -m INLCRNL参数将 \n自动调整为 \r\n防止输出乱掉
shell
su
tio -m INLCRNL /dev/ttyUSB0按下 xr806 上的 reset 按钮（usb接口右边的按钮），tio 里会输出信息
0x2 编译出 hello world
打开 device/xradio/xr806/BUILD.gn，启用 deps += "ohosdemo:ohosdemo"
打开 device/xradio/xr806/ohosdemo/BUILD.gn，启用 deps = "hello_demo:app_hello"
重新上一节编译默认的固件 操作，串口会每秒打印 hello world
0x3 精简和编译 ncnn
配置 cmake toolchain
下载最新 ncnn 源码
shell
git clone https://github.com/Tencent/ncnn.git在 toolchains 目录新建 xr806.toolchain.cmake，设置 toolchain 目录和编译器路径
为避免引入架构相关的代码，处理器设为 xtensa
从前面生成的 device/xradio/xr806/liteos_m/SDKconfig.gni中摘抄有用的目标编译参数，设为默认编译参数
cmake
set(CMAKE_SYSTEM_NAME Generic)
set(CMAKE_SYSTEM_PROCESSOR xtensa)
set(CMAKE_C_COMPILER "/home/nihui/osd/xr806/gcc-arm-none-eabi-10-2020-q4-major/bin/arm-none-eabi-gcc")
set(CMAKE_CXX_COMPILER "/home/nihui/osd/xr806/gcc-arm-none-eabi-10-2020-q4-major/bin/arm-none-eabi-g++")
set(CMAKE_FIND_ROOT_PATH "/home/nihui/osd/xr806/gcc-arm-none-eabi-10-2020-q4-major/arm-none-eabi")
set(CMAKE_TRY_COMPILE_TARGET_TYPE STATIC_LIBRARY)
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_FIND_ROOT_PATH_MODE_PACKAGE ONLY)
set(CMAKE_C_FLAGS "-mcpu=cortex-m33 -mtune=cortex-m33 -march=armv8-m.main+dsp -mfpu=fpv5-sp-d16 -mfloat-abi=softfp -mcmse -mthumb -fno-common")
set(CMAKE_CXX_FLAGS "-mcpu=cortex-m33 -mtune=cortex-m33 -march=armv8-m.main+dsp -mfpu=fpv5-sp-d16 -mfloat-abi=softfp -mcmse -mthumb -fno-common")
# cache flags
set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS}" CACHE STRING "c flags")
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS}" CACHE STRING "c++ flags")精简 ncnn 功能，减小二进制体积
参考文档 https://github.com/Tencent/ncnn/wiki/build-minimal-library 在 xr806.toolchain.cmake追加 ncnn cmake 编译选项
去除 bf16，int8，图像处理，文件加载模型，多线程，平台相关功能，C接口，c++ rtti exception
去除 mnist 模型中用不到的算子
开启 NCNN_SIMPLESTL 以便在没有 c++ stl 的 xr806 上使用 c++ stl 特性
cmake
option(NCNN_INSTALL_SDK "" ON)
option(NCNN_PIXEL_ROTATE "" OFF)
option(NCNN_PIXEL_AFFINE "" OFF)
option(NCNN_PIXEL_DRAWING "" OFF)
option(NCNN_BUILD_BENCHMARK "" OFF)
option(NCNN_BUILD_TESTS "" OFF)
option(NCNN_BUILD_TOOLS "" OFF)
option(NCNN_BUILD_EXAMPLES "" OFF)
option(NCNN_DISABLE_RTTI "" ON)
option(NCNN_DISABLE_EXCEPTION "" ON)
option(NCNN_BF16 "" OFF)
option(NCNN_INT8 "" OFF)
option(NCNN_THREADS "" OFF)
option(NCNN_OPENMP "" OFF)
option(NCNN_STDIO "" OFF)
option(NCNN_STRING "" OFF)
option(NCNN_C_API "" OFF)
option(NCNN_PLATFORM_API "" OFF)
option(NCNN_RUNTIME_CPU "" OFF)
option(NCNN_SIMPLESTL "" ON)
option(WITH_LAYER_absval "" OFF)
option(WITH_LAYER_argmax "" OFF)
option(WITH_LAYER_batchnorm "" OFF)
option(WITH_LAYER_bias "" OFF)
option(WITH_LAYER_bnll "" OFF)
option(WITH_LAYER_concat "" OFF)
option(WITH_LAYER_convolution "" ON)
option(WITH_LAYER_crop "" OFF)
option(WITH_LAYER_deconvolution "" OFF)
option(WITH_LAYER_dropout "" OFF)
option(WITH_LAYER_eltwise "" OFF)
option(WITH_LAYER_elu "" OFF)
option(WITH_LAYER_embed "" OFF)
option(WITH_LAYER_exp "" OFF)
option(WITH_LAYER_flatten "" ON)
option(WITH_LAYER_innerproduct "" ON)
option(WITH_LAYER_input "" ON)
option(WITH_LAYER_log "" OFF)
option(WITH_LAYER_lrn "" OFF)
option(WITH_LAYER_memorydata "" ON)
option(WITH_LAYER_mvn "" OFF)
option(WITH_LAYER_pooling "" ON)
option(WITH_LAYER_power "" OFF)
option(WITH_LAYER_prelu "" OFF)
option(WITH_LAYER_proposal "" OFF)
option(WITH_LAYER_reduction "" OFF)
option(WITH_LAYER_relu "" ON)
option(WITH_LAYER_reshape "" ON)
option(WITH_LAYER_roipooling "" OFF)
option(WITH_LAYER_scale "" OFF)
option(WITH_LAYER_sigmoid "" OFF)
option(WITH_LAYER_slice "" OFF)
option(WITH_LAYER_softmax "" OFF)
option(WITH_LAYER_split "" ON)
option(WITH_LAYER_spp "" OFF)
option(WITH_LAYER_tanh "" OFF)
option(WITH_LAYER_threshold "" OFF)
option(WITH_LAYER_tile "" OFF)
option(WITH_LAYER_rnn "" OFF)
option(WITH_LAYER_lstm "" OFF)
option(WITH_LAYER_binaryop "" ON)
option(WITH_LAYER_unaryop "" OFF)
option(WITH_LAYER_convolutiondepthwise "" OFF)
option(WITH_LAYER_padding "" ON)
option(WITH_LAYER_squeeze "" OFF)
option(WITH_LAYER_expanddims "" OFF)
option(WITH_LAYER_normalize "" OFF)
option(WITH_LAYER_permute "" OFF)
option(WITH_LAYER_priorbox "" OFF)
option(WITH_LAYER_detectionoutput "" OFF)
option(WITH_LAYER_interp "" OFF)
option(WITH_LAYER_deconvolutiondepthwise "" OFF)
option(WITH_LAYER_shufflechannel "" OFF)
option(WITH_LAYER_instancenorm "" OFF)
option(WITH_LAYER_clip "" OFF)
option(WITH_LAYER_reorg "" OFF)
option(WITH_LAYER_yolodetectionoutput "" OFF)
option(WITH_LAYER_quantize "" OFF)
option(WITH_LAYER_dequantize "" OFF)
option(WITH_LAYER_yolov3detectionoutput "" OFF)
option(WITH_LAYER_psroipooling "" OFF)
option(WITH_LAYER_roialign "" OFF)
option(WITH_LAYER_packing "" ON)
option(WITH_LAYER_requantize "" OFF)
option(WITH_LAYER_cast "" ON)
option(WITH_LAYER_hardsigmoid "" OFF)
option(WITH_LAYER_selu "" OFF)
option(WITH_LAYER_hardswish "" OFF)
option(WITH_LAYER_noop "" OFF)
option(WITH_LAYER_pixelshuffle "" OFF)
option(WITH_LAYER_deepcopy "" OFF)
option(WITH_LAYER_mish "" OFF)
option(WITH_LAYER_statisticspooling "" OFF)
option(WITH_LAYER_swish "" OFF)
option(WITH_LAYER_gemm "" ON)
option(WITH_LAYER_groupnorm "" OFF)
option(WITH_LAYER_layernorm "" OFF)
option(WITH_LAYER_softplus "" OFF)
option(WITH_LAYER_gru "" OFF)
option(WITH_LAYER_multiheadattention "" OFF)
option(WITH_LAYER_gelu "" OFF)
option(WITH_LAYER_convolution1d "" OFF)
option(WITH_LAYER_pooling1d "" OFF)
option(WITH_LAYER_convolutiondepthwise1d "" OFF)
option(WITH_LAYER_convolution3d "" OFF)
option(WITH_LAYER_convolutiondepthwise3d "" OFF)
option(WITH_LAYER_pooling3d "" OFF)进一步精简 ncnn 代码，减小二进制体积
打开 src/layer/fused_activation.h，删除除 relu 激活外的其他激活函数实现，避免调用 exp log 等 xr806 上没有的数学函数
打开 src/layer/binaryop.cpp，删除 forward 中除 ADD 外的其他二元操作，减小 binaryop 算子二进制体积
编译 xr806 平台的 ncnn
shell
mkdir build-xr806
cd build-xr806
cmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/xr806.toolchain.cmake -DCMAKE_BUILD_TYPE=Release ..
make -j4
make install编译出的 ncnn 库和头文件在 build-xr806/install 中，libncnn.a 大小约为 234KB
0x4 实现 mnist 手写数字识别
配置 c++ 工程
仿照前面 hello world demo，新建 ncnn demo 工程，main.c 重命名为 main.cpp
将前面编译出的 build-xr806/install 目录拷贝到 ncnn_demo 中
BUILD.gn 中新加 cflags_cc 是 C++ 编译参数，include_dirs 增加 ncnn 头文件目录
python
import("//device/xradio/xr806/liteos_m/config.gni")
static_library("app_ncnn") {
configs = []
sources = [
"src/main.cpp",
]
cflags = board_cflags
cflags_cc = board_cflags
include_dirs = board_include_dirs
include_dirs += [
"//kernel/liteos_m/kernel/arch/include",
"src/ncnn/include",
]
}实现 ncnn mnist
参考 https://github.com/Tencent/ncnn 的相关教程，从训练的模型转换出 ncnn 模型
使用 ncnnoptimize 优化模型架构，使用 ncnn2mem 将模型转换为可嵌入程序的静态数组
为了后续能直接引用加载模型，使用 fp32 类型
shell
ncnnoptimize mnist_1.param mnist_1.bin mnist_1_opt.param mnist_1_opt.bin 0
ncnn2mem mnist_1_opt.param mnist_1_opt.bin mnist_1.id.h mnist_1_bin.h引用加载模型，直接使用程序 rodata 数据，加载不会再申请堆内存，适合 xr806 这样的运行内存较小的设备
cpp
#include "ncnn/benchmark.h"
#include "ncnn/cpu.h"
#include "ncnn/net.h"
#include "ncnn/mat.h"
#include "mnist_1_bin.h"
ncnn::Net net;
net.load_param(mnist_1_opt_param_bin);
net.load_model(mnist_1_opt_bin);预先读取 28x28 灰度图的像素数据，存放在 const unsigned char IN_IMG[]数组中
在推理前后使用 ncnn::get_current_timeapi 获得当前 ms 时间戳，算出推理耗时
cpp
ncnn::Mat in = ncnn::Mat::from_pixels(IN_IMG, ncnn::Mat::PIXEL_GRAY, 28, 28);
ncnn::Mat out;
double start = ncnn::get_current_time();
ncnn::Extractor ex = net.create_extractor();
ex.input(0, in);
ex.extract(17, out);
double end = ncnn::get_current_time();
float time_ms = end - start;输出 10 分类结果，找到概率最高的识别结果
cpp
int gussed = -1;
float guss_exp = -10000000;
for (int i = 0; i < 10; i++)
{
printf("%d: %.2f\n", i, out[i]);
if (out[i] > guss_exp)
{
gussed = i;
guss_exp = out[i];
}
}
printf("I think it is number %d!\n", gussed);以上代码封装好后，在 main.cpp MainThread 调用
0x5 编译含有 ncnn mnist 的固件，新的问题和应对
链接失败，找不到 operator new 等 C++ 底层实现
menuconfig 中的 C++ support其实并没什么用，只要编译 ncnn 时开启 NCNN_SIMPLESTL 选项即可
链接失败，找不到 exp pow log 等数学函数实现
似乎是没有 libm，就直接把 ncnn 中使用的代码删除吧，主要就是 src/fused_activation.h的 sigmoid 这些，模型没用到就不需要
链接失败，找不到 ncnn 的实现
没有链接 libncnn.a，应该是 BUILD.gn 没写好，但我也不知道怎么写，于是直接修改 device/xradio/xr806/libcopy.py
在 APPlib = filenames_filter(FnList,'libapp_','.a')后面添加
python
os.system("cp /home/nihui/osd/xr806/device/xradio/xr806/ohosdemo/ncnn_demo/src/ncnn/lib/libncnn.a /home/nihui/osd/xr806/device/xradio/xr806/xr_skylark/lib/ohos/")
APPlib.append("libncnn.a")修复 image map
在重复前面编译过程时，发生 bin0 bin1 重叠问题，打开 device/xradio/xr806/xr_skylark/project/demo/wlan_ble_demo/image/xr806/image_wlan_ble.cfg，将自动生成的 device/xradio/xr806/xr_skylark/project/demo/wlan_ble_demo/image/xr806/image_auto_cal.cfg文件内容覆盖前者，再次编译解决
无法烧录，救活 xr806
在烧录含有 ncnn mnist 固件后，加载 mnist 模型发生 malloc 失败，此后出现 cpu exception 6 并显示一大堆崩坏的指针，死机，也无法烧录新固件
参考 https://xr806.docs.aw-ol.com/study/soft_flash/ 最后的方法，用镊子接触 GNB 和 PN2，并按下 reset 按钮，再烧录即可成功
去除 wifi 蓝牙模块
因为资源实在紧张，要去除 wifi 蓝牙
又因为对 openharmony 不熟悉，于是暴力删除 device/xradio/xr806/adapter/hals/communication文件夹
此时编译出错，找到 build/lite/config/component/lite_component.gni
修改，不让添加 communication/wifi_lite
python
# add component features
foreach(feature_label, features) {
print(feature_label)
if (feature_label != "//device/xradio/xr806/adapter/hals/communication/wifi_lite/wifiservice:wifiservice") {
deps += [ feature_label ]
}
}menuconfig 去掉 wifi 蓝牙的相关功能
再次编译解决，烧录固件后，不再 malloc 失败，可以正常推理 ncnn mnist
0x6 xr806 ncnn mnist 推理实测
整理的项目
https://github.com/nihui/ncnn_on_xr806
https://github.com/nihui/ncnn_on_esp32
耗时非常稳定，测得算力几乎持平 esp32c3
