Android on xgDebug的博客

使用ncnn布署pytorch模型到Android手机

Sat, 28 Sep 2024 05:08:49 +0000

使用 ncnn 布署 pytorch 模型到 Android 手机

编译 NCNN 时要打开显卡支持 vulkan 是针对 gpu 的 -DNCNN_VULKAN=ON
MobileNetV3

編譯成 MT 時要打開 CMAKE 0091 特性

cmake_minimum_required(VERSION 3.20.0)
cmake_policy(SET CMP0091 NEW)
set(CMAKE_MSVC_RUNTIME_LIBRARY "MultiThreaded$<$:Debug>")
project("client-project")

训练 YOLO

\Envs\torch\Scripts\activate.ps1
python train.py --batch 6 --workers 2 --imgsz 960 --epochs 300 --data "\Core\yaml\data.yaml" --cfg "\Core\yaml\cfg.yaml" --weights \Core\weights\best.pt --device 0

转换模型

from torch import nn
import torch.utils.model_zoo as model_zoo
import torch.onnx
from libs import define
from libs.net import Net
from libs.dataset import ImageDataset
import os

test_data = ImageDataset(define.testPath,False)
test_loader = torch.utils.data.DataLoader( test_data, batch_size=1, shuffle=True)

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = Net(out_dim=19).to(device)
model.load_state_dict(torch.load( "./widget/last.pt" ))
model.eval()

def saveOnnx():
    for data, target in test_loader:
        data, target = data.to(device), target.to(device)
        label = target.long()
        y = model(data)
        # Export the model
        torch.onnx.export(model,                   # model being run
                        data,                      # model input (or a tuple for multiple inputs)
                        "./widget/best.onnx",            # where to save the model (can be a file or file-like object)
                        export_params=True,        # store the trained parameter weights inside the model file
                        opset_version=10,          # the ONNX version to export the model to
                        do_constant_folding=True,  # whether to execute constant folding for optimization
                        input_names = ['input'],   # the model's input names
                        output_names = ['output'],  # the model's output names
                        dynamic_axes={'input' : {0 : 'batch_size'},    # variable lenght axes
                                        'output' : {0 : 'batch_size'}})

        traced_script_module = torch.jit.trace(model, data)
        return

saveOnnx()
# 转换
os.system("python -m onnxsim ./widget/best.onnx ./widgetbest-sim.onnx")
os.system("./bin/onnx2ncnn.exe ./widget/best-sim.onnx ./widget/best.param ./widget/best.bin")
os.system("./bin/ncnnoptimize.exe ./widget/best.param ./widget/best.bin ./widget/best-opt.param ./widget/best-opt.bin 65536")

python .\export.py --weights weights/best.pt --img 960 --batch 1 --train
python -m onnxsim best.onnx best-sim.onnx
.\onnx2ncnn.exe best-sim.onnx best.param best.bin
ncnnoptimize best.param best.bin best-opt.param best-opt.bin 65536

Git clone ncnn repo with submodule

$ git clone https://github.com/Tencent/ncnn.git
$ cd ncnn
$ git submodule update --init

Build for Linux

Install required build dependencies:

git
g++
cmake
protocol buffer (protobuf) headers files and protobuf compiler
vulkan header files and loader library
glslang
(optional) opencv # For building examples

Generally if you have Intel, AMD or Nvidia GPU from last 10 years, Vulkan can be easily used.

On some systems there are no Vulkan drivers easily available at the moment (October 2020), so you might need to disable use of Vulkan on them. This applies to Raspberry Pi 3 (but there is experimental open source Vulkan driver in the works, which is not ready yet). Nvidia Tegra series devices (like Nvidia Jetson) should support Vulkan. Ensure you have most recent software installed for best expirience.

On Debian, Ubuntu or Raspberry Pi OS, you can install all required dependencies using:

sudo apt install build-essential git cmake libprotobuf-dev protobuf-compiler libvulkan-dev vulkan-utils libopencv-dev

To use Vulkan backend install Vulkan header files, a vulkan driver loader, GLSL to SPIR-V compiler and vulkaninfo tool. Preferably from your distribution repositories. Alternatively download and install full Vulkan SDK (about 200MB in size; it contains all header files, documentation and prebuilt loader, as well some extra tools and source code of everything) from https://vulkan.lunarg.com/sdk/home

wget https://sdk.lunarg.com/sdk/download/1.2.189.0/linux/vulkansdk-linux-x86_64-1.2.189.0.tar.gz?Human=true -O vulkansdk-linux-x86_64-1.2.189.0.tar.gz
tar -xf vulkansdk-linux-x86_64-1.2.189.0.tar.gz
export VULKAN_SDK=$(pwd)/1.2.189.0/x86_64

To use Vulkan after building ncnn later, you will also need to have Vulkan driver for your GPU. For AMD and Intel GPUs these can be found in Mesa graphics driver, which usually is installed by default on all distros (i.e. sudo apt install mesa-vulkan-drivers on Debian/Ubuntu). For Nvidia GPUs the proprietary Nvidia driver must be downloaded and installed (some distros will allow easier installation in some way). After installing Vulkan driver, confirm Vulkan libraries and driver are working, by using vulkaninfo or vulkaninfo | grep deviceType, it should list GPU device type. If there are more than one GPU installed (including the case of integrated GPU and discrete GPU, commonly found in laptops), you might need to note the order of devices to use later on.

Nvidia Jetson devices the Vulkan support should be present in Nvidia provided SDK (Jetpack) or prebuild OS images.

Raspberry Pi Vulkan drivers do exists, but are not mature. You are free to experiment at your own discretion, and report results and performance.

cd ncnn
mkdir -p build
cd build
cmake -DCMAKE_BUILD_TYPE=Release -DNCNN_VULKAN=ON -DNCNN_SYSTEM_GLSLANG=ON -DNCNN_BUILD_EXAMPLES=ON ..
make -j$(nproc)

You can add -GNinja to cmake above to use Ninja build system (invoke build using ninja or cmake --build .).

For Nvidia Jetson devices, add -DCMAKE_TOOLCHAIN_FILE=../toolchains/jetson.toolchain.cmake to cmake.

For Rasberry Pi 3, add -DCMAKE_TOOLCHAIN_FILE=../toolchains/pi3.toolchain.cmake -DPI3=ON to cmake. You can also consider disabling Vulkan support as the Vulkan drivers for Rasberry Pi are still not mature, but it doesn’t hurt to build the support in, but not use it.

Verify build by running some examples:

cd ../examples
../build/examples/squeezenet ../images/256-ncnn.png
[0 AMD RADV FIJI (LLVM 10.0.1)]  queueC=1[4]  queueG=0[1]  queueT=0[1]
[0 AMD RADV FIJI (LLVM 10.0.1)]  bugsbn1=0  buglbia=0  bugcopc=0  bugihfa=0
[0 AMD RADV FIJI (LLVM 10.0.1)]  fp16p=1  fp16s=1  fp16a=0  int8s=1  int8a=1
532 = 0.163452
920 = 0.093140
716 = 0.061584

You can also run benchmarks (the 4th argument is a GPU device index to use, refer to vulkaninfo, if you have more than one GPU):

cd ../benchmark
../build/benchmark/benchncnn 10 $(nproc) 0 0
[0 AMD RADV FIJI (LLVM 10.0.1)]  queueC=1[4]  queueG=0[1]  queueT=0[1]
[0 AMD RADV FIJI (LLVM 10.0.1)]  bugsbn1=0  buglbia=0  bugcopc=0  bugihfa=0
[0 AMD RADV FIJI (LLVM 10.0.1)]  fp16p=1  fp16s=1  fp16a=0  int8s=1  int8a=1
num_threads = 4
powersave = 0
gpu_device = 0
cooling_down = 1
          squeezenet  min =    4.68  max =    4.99  avg =    4.85
     squeezenet_int8  min =   38.52  max =   66.90  avg =   48.52
...

To run benchmarks on a CPU, set the 5th argument to -1.

Build for Windows x64 using Visual Studio Community 2017

Download and Install Visual Studio Community 2017 from https://visualstudio.microsoft.com/vs/community/

Start the command prompt: Start → Programs → Visual Studio 2017 → Visual Studio Tools → x64 Native Tools Command Prompt for VS 2017

Download protobuf-3.4.0 from https://github.com/google/protobuf/archive/v3.4.0.zip

Build protobuf library:

cd 
mkdir build
cd build
cmake -G"NMake Makefiles" -DCMAKE_BUILD_TYPE=Release -DCMAKE_INSTALL_PREFIX=%cd%/install -Dprotobuf_BUILD_TESTS=OFF -Dprotobuf_MSVC_STATIC_RUNTIME=OFF ../cmake
nmake
nmake install

(optional) Download and install Vulkan SDK from https://vulkan.lunarg.com/sdk/home

Build ncnn library (replace protobuf-root-dir with a proper path):

cd 
mkdir -p build
cd build
cmake -G"NMake Makefiles" -DCMAKE_BUILD_TYPE=Release -DCMAKE_INSTALL_PREFIX=%cd%/install -DProtobuf_INCLUDE_DIR=/build/install/include -DProtobuf_LIBRARIES=/build/install/lib/libprotobuf.lib -DProtobuf_PROTOC_EXECUTABLE=/build/install/bin/protoc.exe -DNCNN_VULKAN=ON ..
nmake
nmake install

Note: To speed up compilation process on multi core machines, configuring cmake to use jom or ninja using -G flag is recommended.

Build for macOS

First install Xcode or Xcode Command Line Tools according to your needs.

Then install protobuf and libomp via homebrew

brew install protobuf libomp

Download and install Vulkan SDK from https://vulkan.lunarg.com/sdk/home

wget https://sdk.lunarg.com/sdk/download/1.2.189.0/mac/vulkansdk-macos-1.2.189.0.dmg?Human=true -O vulkansdk-macos-1.2.189.0.dmg
hdiutil attach vulkansdk-macos-1.2.189.0.dmg
sudo /Volumes/vulkansdk-macos-1.2.189.0/InstallVulkan.app/Contents/MacOS/InstallVulkan --root `pwd`/vulkansdk-macos-1.2.189.0 --accept-licenses --default-answer --confirm-command install
hdiutil detach /Volumes/vulkansdk-macos-1.2.189.0

# setup env
export VULKAN_SDK=`pwd`/vulkansdk-macos-1.2.189.0/macOS

cd 
mkdir -p build
cd build

cmake -DCMAKE_OSX_ARCHITECTURES="x86_64;arm64" \
    -DVulkan_INCLUDE_DIR=`pwd`/../vulkansdk-macos-1.2.189.0/MoltenVK/include \
    -DVulkan_LIBRARY=`pwd`/../vulkansdk-macos-1.2.189.0/MoltenVK/dylib/macOS/libMoltenVK.dylib \
    -DNCNN_VULKAN=ON -DNCNN_BUILD_EXAMPLES=ON ..

cmake --build . -j 4
cmake --build . --target install

Note: If you encounter libomp related errors during installation, you can also check our GitHub Actions at here to install and use openmp.

Build for ARM Cortex-A family with cross-compiling

Download ARM toolchain from https://developer.arm.com/open-source/gnu-toolchain/gnu-a/downloads

export PATH=":${PATH}"

Alternatively install a cross-compiler provided by the distribution (i.e. on Debian / Ubuntu, you can do sudo apt install g++-arm-linux-gnueabi g++-arm-linux-gnueabihf g++-aarch64-linux-gnu).

Depending on your needs build one or more of the below targets.

AArch32 target with soft float (arm-linux-gnueabi)

cd 
mkdir -p build-arm-linux-gnueabi
cd build-arm-linux-gnueabi
cmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/arm-linux-gnueabi.toolchain.cmake ..
make -j$(nproc)

AArch32 target with hard float (arm-linux-gnueabihf)

cd 
mkdir -p build-arm-linux-gnueabihf
cd build-arm-linux-gnueabihf
cmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/arm-linux-gnueabihf.toolchain.cmake ..
make -j$(nproc)

AArch64 GNU/Linux target (aarch64-linux-gnu)

cd 
mkdir -p build-aarch64-linux-gnu
cd build-aarch64-linux-gnu
cmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/aarch64-linux-gnu.toolchain.cmake ..
make -j$(nproc)

Build for Hisilicon platform with cross-compiling

Download and install Hisilicon SDK. The toolchain should be in /opt/hisi-linux/x86-arm

cd 
mkdir -p build
cd build

# Choose one cmake toolchain file depends on your target platform
cmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/hisiv300.toolchain.cmake ..
cmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/hisiv500.toolchain.cmake ..
cmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/himix100.toolchain.cmake ..
cmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/himix200.toolchain.cmake ..

make -j$(nproc)
make install

Build for Android

You can use the pre-build ncnn-android-lib.zip from https://github.com/Tencent/ncnn/releases

Download Android NDK from http://developer.android.com/ndk/downloads/index.html and install it, for example:

unzip android-ndk-r21d-linux-x86_64.zip
export ANDROID_NDK=

(optional) remove the hardcoded debug flag in Android NDK android-ndk issue

# open $ANDROID_NDK/build/cmake/android.toolchain.cmake
# delete "-g" line
list(APPEND ANDROID_COMPILER_FLAGS
  -g
  -DANDROID

Build armv7 library

cd 
mkdir -p build-android-armv7
cd build-android-armv7

cmake -DCMAKE_TOOLCHAIN_FILE="$ANDROID_NDK/build/cmake/android.toolchain.cmake" \
    -DANDROID_ABI="armeabi-v7a" -DANDROID_ARM_NEON=ON \
    -DANDROID_PLATFORM=android-14 ..

# If you want to enable Vulkan, platform api version >= android-24 is needed
cmake -DCMAKE_TOOLCHAIN_FILE="$ANDROID_NDK/build/cmake/android.toolchain.cmake" \
  -DANDROID_ABI="armeabi-v7a" -DANDROID_ARM_NEON=ON \
  -DANDROID_PLATFORM=android-24 -DNCNN_VULKAN=ON ..

make -j$(nproc)
make install

Pick build-android-armv7/install folder for further JNI usage.

Build aarch64 library:

cd 
mkdir -p build-android-aarch64
cd build-android-aarch64

cmake -DCMAKE_TOOLCHAIN_FILE="$ANDROID_NDK/build/cmake/android.toolchain.cmake"\
  -DANDROID_ABI="arm64-v8a" \
  -DANDROID_PLATFORM=android-21 ..

# If you want to enable Vulkan, platform api version >= android-24 is needed
cmake -DCMAKE_TOOLCHAIN_FILE="$ANDROID_NDK/build/cmake/android.toolchain.cmake" \
  -DANDROID_ABI="arm64-v8a" \
  -DANDROID_PLATFORM=android-24 -DNCNN_VULKAN=ON ..

make -j$(nproc)
make install

Pick build-android-aarch64/install folder for further JNI usage.

Build for iOS on macOS with xcode

You can use the pre-build ncnn.framework glslang.framework and openmp.framework from https://github.com/Tencent/ncnn/releases

Install xcode

You can replace -DENABLE_BITCODE=0 to -DENABLE_BITCODE=1 in the following cmake arguments if you want to build bitcode enabled libraries.

Download and install openmp for multithreading inference feature on iPhoneOS

wget https://github.com/llvm/llvm-project/releases/download/llvmorg-11.0.0/openmp-11.0.0.src.tar.xz
tar -xf openmp-11.0.0.src.tar.xz
cd openmp-11.0.0.src

# apply some compilation fix
sed -i'' -e '/.size __kmp_unnamed_critical_addr/d' runtime/src/z_Linux_asm.S
sed -i'' -e 's/__kmp_unnamed_critical_addr/___kmp_unnamed_critical_addr/g' runtime/src/z_Linux_asm.S

mkdir -p build-ios
cd build-ios

cmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/ios.toolchain.cmake -DCMAKE_BUILD_TYPE=Release -DCMAKE_INSTALL_PREFIX=install \
    -DIOS_PLATFORM=OS -DENABLE_BITCODE=0 -DENABLE_ARC=0 -DENABLE_VISIBILITY=0 -DIOS_ARCH="armv7;arm64;arm64e" \
    -DPERL_EXECUTABLE=/usr/local/bin/perl \
    -DLIBOMP_ENABLE_SHARED=OFF -DLIBOMP_OMPT_SUPPORT=OFF -DLIBOMP_USE_HWLOC=OFF ..

cmake --build . -j 4
cmake --build . --target install

# copy openmp library and header files to xcode toolchain sysroot
sudo cp install/include/* /Applications/Xcode.app/Contents/Developer/Platforms/iPhoneOS.platform/Developer/SDKs/iPhoneOS.sdk/usr/include
sudo cp install/lib/libomp.a /Applications/Xcode.app/Contents/Developer/Platforms/iPhoneOS.platform/Developer/SDKs/iPhoneOS.sdk/usr/lib

Download and install openmp for multithreading inference feature on iPhoneSimulator

wget https://github.com/llvm/llvm-project/releases/download/llvmorg-11.0.0/openmp-11.0.0.src.tar.xz
tar -xf openmp-11.0.0.src.tar.xz
cd openmp-11.0.0.src

# apply some compilation fix
sed -i'' -e '/.size __kmp_unnamed_critical_addr/d' runtime/src/z_Linux_asm.S
sed -i'' -e 's/__kmp_unnamed_critical_addr/___kmp_unnamed_critical_addr/g' runtime/src/z_Linux_asm.S

mkdir -p build-ios-sim
cd build-ios-sim

cmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/ios.toolchain.cmake -DCMAKE_BUILD_TYPE=Release -DCMAKE_INSTALL_PREFIX=install \
    -DIOS_PLATFORM=SIMULATOR -DENABLE_BITCODE=0 -DENABLE_ARC=0 -DENABLE_VISIBILITY=0 -DIOS_ARCH="i386;x86_64" \
    -DPERL_EXECUTABLE=/usr/local/bin/perl \
    -DLIBOMP_ENABLE_SHARED=OFF -DLIBOMP_OMPT_SUPPORT=OFF -DLIBOMP_USE_HWLOC=OFF ..

cmake --build . -j 4
cmake --build . --target install

# copy openmp library and header files to xcode toolchain sysroot
sudo cp install/include/* /Applications/Xcode.app/Contents/Developer/Platforms/iPhoneSimulator.platform/Developer/SDKs/iPhoneSimulator.sdk/usr/include
sudo cp install/lib/libomp.a /Applications/Xcode.app/Contents/Developer/Platforms/iPhoneSimulator.platform/Developer/SDKs/iPhoneSimulator.sdk/usr/lib

Package openmp framework:

cd 

mkdir -p openmp.framework/Versions/A/Headers
mkdir -p openmp.framework/Versions/A/Resources
ln -s A openmp.framework/Versions/Current
ln -s Versions/Current/Headers openmp.framework/Headers
ln -s Versions/Current/Resources openmp.framework/Resources
ln -s Versions/Current/openmp openmp.framework/openmp
lipo -create build-ios/install/lib/libomp.a build-ios-sim/install/lib/libomp.a -o openmp.framework/Versions/A/openmp
cp -r build-ios/install/include/* openmp.framework/Versions/A/Headers/
sed -e 's/__NAME__/openmp/g' -e 's/__IDENTIFIER__/org.llvm.openmp/g' -e 's/__VERSION__/11.0/g' Info.plist > openmp.framework/Versions/A/Resources/Info.plist

Download and install Vulkan SDK from https://vulkan.lunarg.com/sdk/home

wget https://sdk.lunarg.com/sdk/download/1.2.189.0/mac/vulkansdk-macos-1.2.189.0.dmg?Human=true -O vulkansdk-macos-1.2.189.0.dmg
hdiutil attach vulkansdk-macos-1.2.189.0.dmg
sudo /Volumes/vulkansdk-macos-1.2.189.0/InstallVulkan.app/Contents/MacOS/InstallVulkan --root `pwd`/vulkansdk-macos-1.2.189.0 --accept-licenses --default-answer --confirm-command install
hdiutil detach /Volumes/vulkansdk-macos-1.2.189.0

# setup env
export VULKAN_SDK=`pwd`/vulkansdk-macos-1.2.189.0/macOS

Build library for iPhoneOS:

cd 
mkdir -p build-ios
cd build-ios

cmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/ios.toolchain.cmake -DIOS_PLATFORM=OS -DIOS_ARCH="armv7;arm64;arm64e" \
    -DENABLE_BITCODE=0 -DENABLE_ARC=0 -DENABLE_VISIBILITY=0 \
    -DOpenMP_C_FLAGS="-Xclang -fopenmp" -DOpenMP_CXX_FLAGS="-Xclang -fopenmp" \
    -DOpenMP_C_LIB_NAMES="libomp" -DOpenMP_CXX_LIB_NAMES="libomp" \
    -DOpenMP_libomp_LIBRARY="/Applications/Xcode.app/Contents/Developer/Platforms/iPhoneOS.platform/Developer/SDKs/iPhoneOS.sdk/usr/lib/libomp.a" \
    -DNCNN_BUILD_BENCHMARK=OFF ..

# vulkan is only available on arm64 devices
cmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/ios.toolchain.cmake -DIOS_PLATFORM=OS64 -DIOS_ARCH="arm64;arm64e" \
    -DENABLE_BITCODE=0 -DENABLE_ARC=0 -DENABLE_VISIBILITY=0 \
    -DOpenMP_C_FLAGS="-Xclang -fopenmp" -DOpenMP_CXX_FLAGS="-Xclang -fopenmp" \
    -DOpenMP_C_LIB_NAMES="libomp" -DOpenMP_CXX_LIB_NAMES="libomp" \
    -DOpenMP_libomp_LIBRARY="/Applications/Xcode.app/Contents/Developer/Platforms/iPhoneOS.platform/Developer/SDKs/iPhoneOS.sdk/usr/lib/libomp.a" \
    -DVulkan_INCLUDE_DIR=`pwd`/../vulkansdk-macos-1.2.189.0/MoltenVK/include \
    -DVulkan_LIBRARY=`pwd`/../vulkansdk-macos-1.2.189.0/MoltenVK/dylib/iOS/libMoltenVK.dylib \
    -DNCNN_VULKAN=ON -DNCNN_BUILD_BENCHMARK=OFF ..

cmake --build . -j 4
cmake --build . --target install

Build library for iPhoneSimulator:

cd 
mkdir -p build-ios-sim
cd build-ios-sim

cmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/ios.toolchain.cmake -DIOS_PLATFORM=SIMULATOR -DIOS_ARCH="i386;x86_64" \
    -DENABLE_BITCODE=0 -DENABLE_ARC=0 -DENABLE_VISIBILITY=0 \
    -DOpenMP_C_FLAGS="-Xclang -fopenmp" -DOpenMP_CXX_FLAGS="-Xclang -fopenmp" \
    -DOpenMP_C_LIB_NAMES="libomp" -DOpenMP_CXX_LIB_NAMES="libomp" \
    -DOpenMP_libomp_LIBRARY="/Applications/Xcode.app/Contents/Developer/Platforms/iPhoneSimulator.platform/Developer/SDKs/iPhoneSimulator.sdk/usr/lib/libomp.a" \
    -DNCNN_BUILD_BENCHMARK=OFF ..

cmake --build . -j 4
cmake --build . --target install

Package glslang framework:

cd 

mkdir -p glslang.framework/Versions/A/Headers
mkdir -p glslang.framework/Versions/A/Resources
ln -s A glslang.framework/Versions/Current
ln -s Versions/Current/Headers glslang.framework/Headers
ln -s Versions/Current/Resources glslang.framework/Resources
ln -s Versions/Current/glslang glslang.framework/glslang
libtool -static build-ios/install/lib/libglslang.a build-ios/install/lib/libSPIRV.a build-ios/install/lib/libOGLCompiler.a build-ios/install/lib/libOSDependent.a -o build-ios/install/lib/libglslang_combined.a
libtool -static build-ios-sim/install/lib/libglslang.a build-ios-sim/install/lib/libSPIRV.a build-ios-sim/install/lib/libOGLCompiler.a build-ios-sim/install/lib/libOSDependent.a -o build-ios-sim/install/lib/libglslang_combined.a
lipo -create build-ios/install/lib/libglslang_combined.a build-ios-sim/install/lib/libglslang_combined.a -o glslang.framework/Versions/A/glslang
cp -r build/install/include/glslang glslang.framework/Versions/A/Headers/
sed -e 's/__NAME__/glslang/g' -e 's/__IDENTIFIER__/org.khronos.glslang/g' -e 's/__VERSION__/1.0/g' Info.plist > glslang.framework/Versions/A/Resources/Info.plist

Package ncnn framework:

cd 

mkdir -p ncnn.framework/Versions/A/Headers
mkdir -p ncnn.framework/Versions/A/Resources
ln -s A ncnn.framework/Versions/Current
ln -s Versions/Current/Headers ncnn.framework/Headers
ln -s Versions/Current/Resources ncnn.framework/Resources
ln -s Versions/Current/ncnn ncnn.framework/ncnn
lipo -create build-ios/install/lib/libncnn.a build-ios-sim/install/lib/libncnn.a -o ncnn.framework/Versions/A/ncnn
cp -r build-ios/install/include/* ncnn.framework/Versions/A/Headers/
sed -e 's/__NAME__/ncnn/g' -e 's/__IDENTIFIER__/com.tencent.ncnn/g' -e 's/__VERSION__/1.0/g' Info.plist > ncnn.framework/Versions/A/Resources/Info.plist

Pick ncnn.framework glslang.framework and openmp.framework folder for app development.

Build for WebAssembly

Install Emscripten

git clone https://github.com/emscripten-core/emsdk.git
cd emsdk
./emsdk install 2.0.8
./emsdk activate 2.0.8

source emsdk/emsdk_env.sh

Build without any extension for general compatibility:

mkdir -p build
cd build
cmake -DCMAKE_TOOLCHAIN_FILE=../emsdk/upstream/emscripten/cmake/Modules/Platform/Emscripten.cmake \
    -DNCNN_THREADS=OFF -DNCNN_OPENMP=OFF -DNCNN_SIMPLEOMP=OFF -DNCNN_RUNTIME_CPU=OFF -DNCNN_SSE2=OFF -DNCNN_AVX2=OFF -DNCNN_AVX=OFF \
    -DNCNN_BUILD_TOOLS=OFF -DNCNN_BUILD_EXAMPLES=OFF -DNCNN_BUILD_BENCHMARK=OFF ..
cmake --build . -j 4
cmake --build . --target install

Build with WASM SIMD extension:

mkdir -p build-simd
cd build-simd
cmake -DCMAKE_TOOLCHAIN_FILE=../emsdk/upstream/emscripten/cmake/Modules/Platform/Emscripten.cmake \
    -DNCNN_THREADS=OFF -DNCNN_OPENMP=OFF -DNCNN_SIMPLEOMP=OFF -DNCNN_RUNTIME_CPU=OFF -DNCNN_SSE2=ON -DNCNN_AVX2=OFF -DNCNN_AVX=OFF \
    -DNCNN_BUILD_TOOLS=OFF -DNCNN_BUILD_EXAMPLES=OFF -DNCNN_BUILD_BENCHMARK=OFF ..
cmake --build . -j 4
cmake --build . --target install

Build with WASM Thread extension:

mkdir -p build-threads
cd build-threads
cmake -DCMAKE_TOOLCHAIN_FILE=../emsdk/upstream/emscripten/cmake/Modules/Platform/Emscripten.cmake \
    -DNCNN_THREADS=ON -DNCNN_OPENMP=ON -DNCNN_SIMPLEOMP=ON -DNCNN_RUNTIME_CPU=OFF -DNCNN_SSE2=OFF -DNCNN_AVX2=OFF -DNCNN_AVX=OFF \
    -DNCNN_BUILD_TOOLS=OFF -DNCNN_BUILD_EXAMPLES=OFF -DNCNN_BUILD_BENCHMARK=OFF ..
cmake --build . -j 4
cmake --build . --target install

Build with WASM SIMD and Thread extension:

mkdir -p build-simd-threads
cd build-simd-threads
cmake -DCMAKE_TOOLCHAIN_FILE=../emsdk/upstream/emscripten/cmake/Modules/Platform/Emscripten.cmake \
    -DNCNN_THREADS=ON -DNCNN_OPENMP=ON -DNCNN_SIMPLEOMP=ON -DNCNN_RUNTIME_CPU=OFF -DNCNN_SSE2=ON -DNCNN_AVX2=OFF -DNCNN_AVX=OFF \
    -DNCNN_BUILD_TOOLS=OFF -DNCNN_BUILD_EXAMPLES=OFF -DNCNN_BUILD_BENCHMARK=OFF ..
cmake --build . -j 4
cmake --build . --target install

Pick build-XYZ/install folder for further usage.

Build for AllWinner D1

Download c906 toolchain package from https://occ.t-head.cn/community/download?id=3913221581316624384

tar -xf riscv64-linux-x86_64-20210512.tar.gz
export RISCV_ROOT_PATH=/home/nihui/osd/riscv64-linux-x86_64-20210512

Build ncnn with riscv-v vector and simpleocv enabled:

mkdir -p build-c906
cd build-c906
cmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/c906.toolchain.cmake \
    -DCMAKE_BUILD_TYPE=relwithdebinfo -DNCNN_OPENMP=OFF -DNCNN_THREADS=OFF -DNCNN_RUNTIME_CPU=OFF -DNCNN_RVV=ON \
    -DNCNN_SIMPLEOCV=ON -DNCNN_BUILD_EXAMPLES=ON ..
cmake --build . -j 4
cmake --build . --target install

Pick build-c906/install folder for further usage.

You can upload binary inside build-c906/examples folder and run on D1 board for testing.

Build for Loongson 2K1000

For gcc version < 8.5, you need to fix msa.h header for workaround msa fmadd bug.

Open /usr/lib/gcc/mips64el-linux-gnuabi64/8/include/msa.h, find __msa_fmadd_w and apply changes as the following

// #define __msa_fmadd_w __builtin_msa_fmadd_w
#define __msa_fmadd_w(a, b, c) __builtin_msa_fmadd_w(c, b, a)

Build ncnn with mips msa and simpleocv enabled:

mkdir -p build
cd build
cmake -DNCNN_DISABLE_RTTI=ON -DNCNN_DISABLE_EXCEPTION=ON -DNCNN_RUNTIME_CPU=OFF -DNCNN_MSA=ON -DNCNN_MMI=ON -DNCNN_SIMPLEOCV=ON ..
cmake --build . -j 2
cmake --build . --target install

Pick build/install folder for further usage.

You can run binary inside build/examples folder for testing.

Build for Termux on Android

Install app Termux on your phone,and install Ubuntu in Termux.

If you want use ssh, just install openssh in Termux

pkg install proot-distro
proot-distro install ubuntu

or you can see what system can be installed using proot-distro list

while you install ubuntu successfully, using proot-distro login ubuntu to login Ubuntu.

Then make ncnn,no need to install any other dependencies.

git clone https://github.com/Tencent/ncnn.git
cd ncnn
git submodule update --init
mkdir -p build
cd build
cmake -DCMAKE_BUILD_TYPE=Release -DNCNN_BUILD_EXAMPLES=ON -DNCNN_PLATFORM_API=OFF -DNCNN_SIMPLEOCV=ON ..
make -j$(nproc)

Then you can run a test

on my Pixel 3 XL using Qualcomm 845,cant load 256-ncnn.png

cd ../examples
../build/examples/squeezenet ../images/128-ncnn.png

在Android 11 上使用 LLDV 调试原生程序

Sat, 28 Sep 2024 04:32:12 +0000

手机端

PUSH 调试服务器到手机

adb push lldb-server /data/local/tmp
chmod 755 /data/local/tmp/lldb-server

启动调试器服务

/data/local/tmp/lldb-server platform --listen "*:8888" --server

电脑端

端口转发

adb forward tcp:8888 tcp:8888

启动 LLDB

.\lldb.exe

查看支持平台

platform list

选 ANDROID 平台

platform select remote-android

连接到手机手机序列号: 9643e0ec0604 要换成当前调试的手机,使用 adb devices 查看序列号

platform connect connect://9643e0ec0604:8888

查看当前正在运行的进程

platform process list

附加上去

attach 9053

下断点

b send

跑起来

查看线程列表

thread list

查看调用栈

bt

用 ncnn 部署 PyTorch 模型到 Android 手机

Fri, 11 Aug 2023 01:18:11 +0000

使用 ncnn 将 pytorch 模型部署到安卓手机

在编译 NCNN 时开启图形卡支持。Vulkan 用于 GPU，设置 -DNCNN_VULKAN=ON。
MobileNetV3

在编译到 MT 时开启 CMAKE 0091 特性

cmake_minimum_required(VERSION 3.20.0)
cmake_policy(SET CMP0091 NEW)
set(CMAKE_MSVC_RUNTIME_LIBRARY "MultiThreaded$<$:Debug>")
project("client-project")

训练 YOLO

\Envs\torch\Scripts\activate.ps1
python train.py --batch 6 --workers 2 --imgsz 960 --epochs 300 --data "\Core\yaml\data.yaml" --cfg "\Core\yaml\cfg.yaml" --weights \ Core\weights\best.pt --device 0

模型转换

from torch import nn
import torch.utils.model_zoo as model_zoo
import torch.onnx
from libs import define
from libs.net import Net
from libs.dataset import ImageDataset
import os

test_data = ImageDataset(define.testPath,False)
test_loader = torch.utils.data.DataLoader( test_data, batch_size=1, shuffle=True)

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = Net(out_dim=19).to(device)
model.load_state_dict(torch.load( "./widget/last.pt" ))
model.eval()

def saveOnnx():
    for data, target in test_loader:
        data, target = data.to(device), target.to(device)
        label = target.long()
        y = model(data)
        # 导出模型
        torch.onnx.export(model, # 正在运行的模型
                        data, # 模型输入（或多个输入的元组）
                        "./widget/best.onnx", # 保存模型的位置（可以是文件或文件对象）
                        export_params=True, # 将训练好的参数权重存储在模型文件中
                        opset_version=10, # 导出模型的 ONNX 版本
                        do_constant_folding=True, # 是否执行常量折叠以进行优化
                        input_names = ['input'], # 模型的输入名称
                        output_names = ['output'], # 模型的输出名称
                        dynamic_axes={'input': {0:'batch_size'}, # 可变长度轴
                                        'output': {0:'batch_size'}})

        traced_script_module = torch.jit.trace(model, data)
        return

saveOnnx()
# 转换
os.system("python -m onnxsim ./widget/best.onnx ./widgetbest-sim.onnx")
os.system("./bin/onnx2ncnn.exe ./widget/best-sim.onnx ./widget/best.param ./widget/best.bin")
os.system("./bin/ncnnoptimize.exe ./widget/best.param ./widget/best.bin ./widget/best-opt.param ./widget/best-opt.bin 65536")

python .\export.py --weights weights/best.pt --img 960 --batch 1 --train
python -m onnxsim best.onnx best-sim.onnx
.\onnx2ncnn.exe best-sim.onnx best.param best.bin
ncnnoptimize best.param best.bin best-opt.param best-opt.bin 65536

Git 克隆 ncnn 仓库及子模块

$ git clone https://github.com/Tencent/ncnn.git
$ cd ncnn
$ git submodule update --init

为 Linux 构建

安装所需的构建依赖项：

git
g++
cmake
protocol buffer (protobuf) 头文件和 protobuf 编译器
vulkan 头文件和加载器库
glslang
(可选) opencv # 用于构建示例

一般来说，如果你拥有过去 10 年的 Intel、AMD 或 Nvidia GPU，Vulkan 就可以轻松使用。

在某些系统上，目前（2020 年 10 月）可能没有易于获得的 Vulkan 驱动，因此你可能需要禁用 Vulkan 的使用。这适用于 Raspberry Pi 3（但正在开发中有一个实验性的开源 Vulkan 驱动，尚未准备好）。Nvidia Tegra 系列设备（如 Nvidia Jetson）应该支持 Vulkan。确保安装了最新的软件以获得最佳体验。

在 Debian、Ubuntu 或 Raspberry Pi OS 上，你可以使用以下命令安装所有必需的依赖项：

sudo apt install build-essential git cmake libprotobuf-dev protobuf-compiler libvulkan-dev vulkan-utils libopencv-dev

要使用 Vulkan 后端，你需要安装 Vulkan 头文件、一个 Vulkan 驱动加载器、GLSL 到 SPIR-V 编译器和 vulkaninfo 工具。最好从你的发行版仓库中获取。或者，从 https://vulkan.lunarg.com/sdk/home 下载并安装完整的 Vulkan SDK（大小约为 200MB；它包含所有头文件、文档和预构建加载器，以及一些额外的工具和所有内容的源代码）。

wget https://sdk.lunarg.com/sdk/download/1.2.189.0/linux/vulkansdk-linux-x86_64-1.2.189.0.tar.gz?Human=true -O vulkansdk-linux-x86_64-1.2.189.0.tar.gz
tar -xf vulkansdk-linux-x86_64-1.2.189.0.tar.gz
export VULKAN_SDK=$(pwd)/1.2.189.0/x86_64

在稍后构建 ncnn 时使用 Vulkan，你还需要为你的 GPU 拥有 Vulkan 驱动。对于 AMD 和 Intel GPU，这些可以在 Mesa 图形驱动中找到，通常在所有发行版中默认安装（例如，在 Debian/Ubuntu 上使用 sudo apt install mesa-vulkan-drivers）。对于 Nvidia GPU，必须下载并安装专有 Nvidia 驱动（某些发行版会以更简单的方式允许安装）。安装 Vulkan 驱动后，使用 vulkaninfo 或 vulkaninfo | grep deviceType 确认 Vulkan 库和驱动是否正常工作，它应该列出 GPU 设备类型。如果安装了多个 GPU（包括集成 GPU 和独立 GPU 的情况，常见于笔记本电脑），你可能需要记下设备的顺序以便稍后使用。

Nvidia Jetson 设备中的 Vulkan 支持应包含在 Nvidia 提供的 SDK (Jetpack) 或预构建的操作系统镜像中。

Raspberry Pi 的 Vulkan 驱动是存在的，但尚未成熟。你可以根据自己的判断进行实验，并报告结果和性能。

cd ncnn
mkdir -p build
cd build
cmake -DCMAKE_BUILD_TYPE=Release -DNCNN_VULKAN=ON -DNCNN_SYSTEM_GLSLANG=ON -DNCNN_BUILD_EXAMPLES=ON ..
make -j$(nproc)

你可以向上面的 cmake 添加 -GNinja 来使用 Ninja 构建系统（使用 ninja 或 cmake --build . 调用构建）。

对于 Nvidia Jetson 设备，向 cmake 添加 -DCMAKE_TOOLCHAIN_FILE=../toolchains/jetson.toolchain.cmake。

对于 Rasberry Pi 3，向 cmake 添加 -DCMAKE_TOOLCHAIN_FILE=../toolchains/pi3.toolchain.cmake -DPI3=ON。你也可以考虑禁用 Vulkan 支持，因为 Raspberry Pi 的 Vulkan 驱动仍不成熟，但启用支持并未使用它也不会有坏处。

通过运行一些示例来验证构建：

cd ../examples
../build/examples/squeezenet ../images/256-ncnn.png
[0 AMD RADV FIJI (LLVM 10.0.1)] queueC=1[4] queueG=0[1] queueT=0[1]
[0 AMD RADV FIJI (LLVM 10.0.1)] bugsbn1=0 buglbia=0 bugcopc=0 bugihfa=0
[0 AMD RADV FIJI (LLVM 10.0.1)] fp16p=1 fp16s=1 fp16a=0 int8s=1 int8a=1
532 = 0.163452
920 = 0.093140
716 = 0.061584

你也可以运行基准测试（第 4 个参数是使用的 GPU 设备索引，参考 vulkaninfo，如果你有多个 GPU）：

cd ../benchmark
../build/benchmark/benchncnn 10 $(nproc) 0 0
[0 AMD RADV FIJI (LLVM 10.0.1)] queueC=1[4] queueG=0[1] queueT=0[1]
[0 AMD RADV FIJI (LLVM 10.0.1)] bugsbn1=0 buglbia=0 bugcopc=0 bugihfa=0
[0 AMD RADV FIJI (LLVM 10.0.1)] fp16p=1 fp16s=1 fp16a=0 int8s=1 int8a=1
num_threads = 4
powersave = 0
gpu_device = 0
cooling_down = 1
squeezenet min = 4.68 max = 4.99 avg = 4.85
squeezenet_int8 min = 38.52 max = 66.90 avg = 48.52

要在 CPU 上运行基准测试，将第 5 个参数设置为 -1。

使用 Visual Studio Community 2017 为 Windows x64 构建

从 https://visualstudio.microsoft.com/vs/community/ 下载并安装 Visual Studio Community 2017。

启动命令提示符：Start → Programs → Visual Studio 2017 → Visual Studio Tools → x64 Native Tools Command Prompt for VS 2017

从 https://github.com/google/protobuf/archive/v3.4.0.zip 下载 protobuf-3.4.0。

构建 protobuf 库：

cd 
mkdir build
cd build
cmake -G"NMake Makefiles" -DCMAKE_BUILD_TYPE=Release -DCMAKE_INSTALL_PREFIX=%cd%/install -Dprotobuf_BUILD_TESTS=OFF -Dprotobuf_MSVC_STATIC_RUNTIME=OFF ../cmake
nmake
nmake install

(可选) 从 https://vulkan.lunarg.com/sdk/home 下载并安装 Vulkan SDK。

构建 ncnn 库（将 protobuf-root-dir 替换为正确的路径）：

cd 
mkdir -p build
cd build
cmake -G"NMake Makefiles" -DCMAKE_BUILD_TYPE=Release -DCMAKE_INSTALL_PREFIX=%cd%/install -DProtobuf_INCLUDE_DIR=/build/install/include -DProtobuf_LIBRARIES=/build/install/lib/libprotobuf.lib -DProtobuf_PROTOC_EXECUTABLE=/build/install/bin/protoc.exe -DNCNN_VULKAN=ON ..
nmake
nmake install

注意：为了加快多核机器上的编译过程，推荐使用 -G 标志配置 cmake 来使用 jom 或 ninja。

为 macOS 构建

首先根据你的需求安装 Xcode 或 Xcode 命令行工具。

然后通过 homebrew 安装 protobuf 和 libomp。

brew install protobuf libomp

从 https://vulkan.lunarg.com/sdk/home 下载并安装 Vulkan SDK。

wget https://sdk.lunarg.com/sdk/download/1.2.189.0/mac/vulkansdk-macos-1.2.189.0.dmg?Human=true -O vulkansdk-macos-1.2.189.0.dmg
hdiutil attach vulkansdk-macos-1.2.189.0.dmg
sudo /Volumes/vulkansdk-macos-1.2.189.0/InstallVulkan.app/Contents/MacOS/InstallVulkan --root `pwd`/vulkansdk-macos-1.2.189.0 --accept-licenses --default-answer --confirm-command install
hdiutil detach /Volumes/vulkansdk-macos-1.2.189.0

# 设置环境变量

export VULKAN_SDK=`pwd`/vulkansdk-macos-1.2.189.0/macOS

cd 
mkdir -p build
cd build

cmake -DCMAKE_OSX_ARCHITECTURES="x86_64;arm64" \
 -DVulkan_INCLUDE_DIR=`pwd`/../vulkansdk-macos-1.2.189.0/MoltenVK/include \
 -DVulkan_LIBRARY=`pwd`/../vulkansdk-macos-1.2.189.0/MoltenVK/dylib/macOS/libMoltenVK.dylib \
 -DNCNN_VULKAN=ON -DNCNN_BUILD_EXAMPLES=ON ..

cmake --build . -j 4
cmake --build . --target install

注意：如果在安装过程中遇到与 libomp 相关的错误，你也可以查看我们的 GitHub Actions 此处来安装和使用 openmp。

使用交叉编译为 ARM Cortex-A 系列构建

从 https://developer.arm.com/open-source/gnu-toolchain/gnu-a/downloads 下载 ARM 工具链。

export PATH=":${PATH}"

或者安装发行版提供的交叉编译器（例如，在 Debian / Ubuntu 上，你可以执行 sudo apt install g++-arm-linux-gnueabi g++-arm-linux-gnueabihf g++-aarch64-linux-gnu）。

根据你的需求，构建以下一个或多个目标。

AArch32 软浮点目标 (arm-linux-gnueabi)

cd 
mkdir -p build-arm-linux-gnueabi
cd build-arm-linux-gnueabi
cmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/arm-linux-gnueabi.toolchain.cmake ..
make -j$(nproc)

AArch32 硬浮点目标 (arm-linux-gnueabihf)

cd 
mkdir -p build-arm-linux-gnueabihf
cd build-arm-linux-gnueabihf
cmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/arm-linux-gnueabihf.toolchain.cmake ..
make -j$(nproc)

AArch64 GNU/Linux 目标 (aarch64-linux-gnu)

cd 
mkdir -p build-aarch64-linux-gnu
cd build-aarch64-linux-gnu
cmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/aarch64-linux-gnu.toolchain.cmake ..
make -j$(nproc)

使用交叉编译为 Hisilicon 平台构建

下载并安装 Hisilicon SDK。工具链应位于 /opt/hisi-linux/x86-arm。

cd 
mkdir -p build
cd build

# 选择一个 cmake 工具链文件，取决于你的目标平台

cmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/hisiv300.toolchain.cmake ..
cmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/hisiv500.toolchain.cmake ..
cmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/himix100.toolchain.cmake ..
cmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/himix200.toolchain.cmake ..

make -j$(nproc)
make install

为 Android 构建

你可以使用来自 https://github.com/Tencent/ncnn/releases 的预构建 ncnn-android-lib.zip。

从 http://developer.android.com/ndk/downloads/index.html 下载 Android NDK 并安装它，例如：

unzip android-ndk-r21d-linux-x86_64.zip
export ANDROID_NDK=

(可选) 移除 Android NDK 中的硬编码调试标志 android-ndk issue。

# 打开 $ANDROID_NDK/build/cmake/android.toolchain.cmake

# 删除 "-g" 行

list(APPEND ANDROID_COMPILER_FLAGS
-g
-DANDROID

构建 armv7 库

cd 
mkdir -p build-android-armv7
cd build-android-armv7

cmake -DCMAKE_TOOLCHAIN_FILE="$ANDROID_NDK/build/cmake/android.toolchain.cmake" \
 -DANDROID_ABI="armeabi-v7a" -DANDROID_ARM_NEON=ON \
 -DANDROID_PLATFORM=android-14 ..

# 如果你想启用 Vulkan，则需要平台 API 版本 >= android-24

cmake -DCMAKE_TOOLCHAIN_FILE="$ANDROID_NDK/build/cmake/android.toolchain.cmake" \
 -DANDROID_ABI="armeabi-v7a" -DANDROID_ARM_NEON=ON \
 -DANDROID_PLATFORM=android-24 -DNCNN_VULKAN=ON ..

make -j$(nproc)
make install

选择 build-android-armv7/install 文件夹用于进一步的 JNI 使用。

构建 aarch64 库：

cd 
mkdir -p build-android-aarch64
cd build-android-aarch64

cmake -DCMAKE_TOOLCHAIN_FILE="$ANDROID_NDK/build/cmake/android.toolchain.cmake"\
 -DANDROID_ABI="arm64-v8a" \
 -DANDROID_PLATFORM=android-21 ..

# 如果你想启用 Vulkan，则需要平台 API 版本 >= android-24

cmake -DCMAKE_TOOLCHAIN_FILE="$ANDROID_NDK/build/cmake/android.toolchain.cmake" \
 -DANDROID_ABI="arm64-v8a" \
 -DANDROID_PLATFORM=android-24 -DNCNN_VULKAN=ON ..

make -j$(nproc)
make install

选择 build-android-aarch64/install 文件夹用于进一步的 JNI 使用。

在 macOS 上使用 xcode 为 iOS 构建

你可以使用来自 https://github.com/Tencent/ncnn/releases 的预构建 ncnn.framework、glslang.framework 和 openmp.framework。

安装 xcode

如果你想构建启用 bitcode 的库，可以将以下 cmake 参数中的 -DENABLE_BITCODE=0 替换为 -DENABLE_BITCODE=1。

下载并安装 openmp 以启用 iPhoneOS 上的多线程推理功能

wget https://github.com/llvm/llvm-project/releases/download/llvmorg-11.0.0/openmp-11.0.0.src.tar.xz
tar -xf openmp-11.0.0.src.tar.xz
cd openmp-11.0.0.src

# 应用一些编译修复

sed -i'' -e '/.size **kmp_unnamed_critical_addr/d' runtime/src/z_Linux_asm.S
sed -i'' -e 's/**kmp_unnamed_critical_addr/\_\_\_kmp_unnamed_critical_addr/g' runtime/src/z_Linux_asm.S

mkdir -p build-ios
cd build-ios

cmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/ios.toolchain.cmake -DCMAKE_BUILD_TYPE=Release -DCMAKE_INSTALL_PREFIX=install \
 -DIOS_PLATFORM=OS -DENABLE_BITCODE=0 -DENABLE_ARC=0 -DENABLE_VISIBILITY=0 -DIOS_ARCH="armv7;arm64;arm64e" \
 -DPERL_EXECUTABLE=/usr/local/bin/perl \
 -DLIBOMP_ENABLE_SHARED=OFF -DLIBOMP_OMPT_SUPPORT=OFF -DLIBOMP_USE_HWLOC=OFF ..

cmake --build . -j 4
cmake --build . --target install

# 将 openmp 库和头文件复制到 xcode 工具链 sysroot

sudo cp install/include/\* /Applications/Xcode.app/Contents/Developer/Platforms/iPhoneOS.platform/Developer/SDKs/iPhoneOS.sdk/usr/include
sudo cp install/lib/libomp.a /Applications/Xcode.app/Contents/Developer/Platforms/iPhoneOS.platform/Developer/SDKs/iPhoneOS.sdk/usr/lib

下载并安装 openmp 以启用 iPhoneSimulator 上的多线程推理功能

wget https://github.com/llvm/llvm-project/releases/download/llvmorg-11.0.0/openmp-11.0.0.src.tar.xz
tar -xf openmp-11.0.0.src.tar.xz
cd openmp-11.0.0.src

# 应用一些编译修复

sed -i'' -e '/.size **kmp_unnamed_critical_addr/d' runtime/src/z_Linux_asm.S
sed -i'' -e 's/**kmp_unnamed_critical_addr/\_\_\_kmp_unnamed_critical_addr/g' runtime/src/z_Linux_asm.S

mkdir -p build-ios-sim
cd build-ios-sim

cmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/ios.toolchain.cmake -DCMAKE_BUILD_TYPE=Release -DCMAKE_INSTALL_PREFIX=install \
 -DIOS_PLATFORM=SIMULATOR -DENABLE_BITCODE=0 -DENABLE_ARC=0 -DENABLE_VISIBILITY=0 -DIOS_ARCH="i386;x86_64" \
 -DPERL_EXECUTABLE=/usr/local/bin/perl \
 -DLIBOMP_ENABLE_SHARED=OFF -DLIBOMP_OMPT_SUPPORT=OFF -DLIBOMP_USE_HWLOC=OFF ..

cmake --build . -j 4
cmake --build . --target install

# 将 openmp 库和头文件复制到 xcode 工具链 sysroot

sudo cp install/include/\* /Applications/Xcode.app/Contents/Developer/Platforms/iPhoneSimulator.platform/Developer/SDKs/iPhoneSimulator.sdk/usr/include
sudo cp install/lib/libomp.a /Applications/Xcode.app/Contents/Developer/Platforms/iPhoneSimulator.platform/Developer/SDKs/iPhoneSimulator.sdk/usr/lib

打包 openmp 框架：

cd 

mkdir -p openmp.framework/Versions/A/Headers
mkdir -p openmp.framework/Versions/A/Resources
ln -s A openmp.framework/Versions/Current
ln -s Versions/Current/Headers openmp.framework/Headers
ln -s Versions/Current/Resources openmp.framework/Resources
ln -s Versions/Current/openmp openmp.framework/openmp
lipo -create build-ios/install/lib/libomp.a build-ios-sim/install/lib/libomp.a -o openmp.framework/Versions/A/openmp
cp -r build-ios/install/include/\* openmp.framework/Versions/A/Headers/
sed -e 's/**NAME**/openmp/g' -e 's/**IDENTIFIER**/org.llvm.openmp/g' -e 's/**VERSION**/11.0/g' Info.plist > openmp.framework/Versions/A/Resources/Info.plist

下载并安装 Vulkan SDK。

wget https://sdk.lunarg.com/sdk/download/1.2.189.0/mac/vulkansdk-macos-1.2.189.0.dmg?Human=true -O vulkansdk-macos-1.2.189.0.dmg
hdiutil attach vulkansdk-macos-1.2.189.0.dmg
sudo /Volumes/vulkansdk-macos-1.2.189.0/InstallVulkan.app/Contents/MacOS/InstallVulkan --root `pwd`/vulkansdk-macos-1.2.189.0 --accept-licenses --default-answer --confirm-command install
hdiutil detach /Volumes/vulkansdk-macos-1.2.189.0

# 设置环境变量

export VULKAN_SDK=`pwd`/vulkansdk-macos-1.2.189.0/macOS

为 iPhoneOS 构建库：

cd 
mkdir -p build-ios
cd build-ios

cmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/ios.toolchain.cmake -DIOS_PLATFORM=OS -DIOS_ARCH="armv7;arm64;arm64e" \
 -DENABLE_BITCODE=0 -DENABLE_ARC=0 -DENABLE_VISIBILITY=0 \
 -DOpenMP_C_FLAGS="-Xclang -fopenmp" -DOpenMP_CXX_FLAGS="-Xclang -fopenmp" \
 -DOpenMP_C_LIB_NAMES="libomp" -DOpenMP_CXX_LIB_NAMES="libomp" \
 -DOpenMP_libomp_LIBRARY="/Applications/Xcode.app/Contents/Developer/Platforms/iPhoneOS.platform/Developer/SDKs/iPhoneOS.sdk/usr/lib/libomp.a" \
 -DNCNN_BUILD_BENCHMARK=OFF ..

# Vulkan 仅在 arm64 设备上可用

cmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/ios.toolchain.cmake -DIOS_PLATFORM=OS64 -DIOS_ARCH="arm64;arm64e" \
 -DENABLE_BITCODE=0 -DENABLE_ARC=0 -DENABLE_VISIBILITY=0 \
 -DOpenMP_C_FLAGS="-Xclang -fopenmp" -DOpenMP_CXX_FLAGS="-Xclang -fopenmp" \
 -DOpenMP_C_LIB_NAMES="libomp" -DOpenMP_CXX_LIB_NAMES="libomp" \
 -DOpenMP_libomp_LIBRARY="/Applications/Xcode.app/Contents/Developer/Platforms/iPhoneOS.platform/Developer/SDKs/iPhoneOS.sdk/usr/lib/libomp.a" \
 -DVulkan_INCLUDE_DIR=`pwd`/../vulkansdk-macos-1.2.189.0/MoltenVK/include \
 -DVulkan_LIBRARY=`pwd`/../vulkansdk-macos-1.2.189.0/MoltenVK/dylib/iOS/libMoltenVK.dylib \
 -DNCNN_VULKAN=ON -DNCNN_BUILD_BENCHMARK=OFF ..

cmake --build . -j 4
cmake --build . --target install

为 iPhoneSimulator 构建库：

cd 
mkdir -p build-ios-sim
cd build-ios-sim

cmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/ios.toolchain.cmake -DIOS_PLATFORM=SIMULATOR -DIOS_ARCH="i386;x86_64" \
 -DENABLE_BITCODE=0 -DENABLE_ARC=0 -DENABLE_VISIBILITY=0 \
 -DOpenMP_C_FLAGS="-Xclang -fopenmp" -DOpenMP_CXX_FLAGS="-Xclang -fopenmp" \
 -DOpenMP_C_LIB_NAMES="libomp" -DOpenMP_CXX_LIB_NAMES="libomp" \
 -DOpenMP_libomp_LIBRARY="/Applications/Xcode.app/Contents/Developer/Platforms/iPhoneSimulator.platform/Developer/SDKs/iPhoneSimulator.sdk/usr/lib/libomp.a" \
 -DNCNN_BUILD_BENCHMARK=OFF ..

cmake --build . -j 4
cmake --build . --target install

打包 glslang 框架：

cd 

mkdir -p glslang.framework/Versions/A/Headers
mkdir -p glslang.framework/Versions/A/Resources
ln -s A glslang.framework/Versions/Current
ln -s Versions/Current/Headers glslang.framework/Headers
ln -s Versions/Current/Resources glslang.framework/Resources
ln -s Versions/Current/glslang glslang.framework/glslang
libtool -static build-ios/install/lib/libglslang.a build-ios/install/lib/libSPIRV.a build-ios/install/lib/libOGLCompiler.a build-ios/install/lib/libOSDependent.a -o build-ios/install/lib/libglslang_combined.a
libtool -static build-ios-sim/install/lib/libglslang.a build-ios-sim/install/lib/libSPIRV.a build-ios-sim/install/lib/libOGLCompiler.a build-ios-sim/install/lib/libOSDependent.a -o build-ios-sim/install/lib/libglslang_combined.a
lipo -create build-ios/install/lib/libglslang_combined.a build-ios-sim/install/lib/libglslang_combined.a -o glslang.framework/Versions/A/glslang
cp -r build/install/include/glslang glslang.framework/Versions/A/Headers/
sed -e 's/**NAME**/glslang/g' -e 's/**IDENTIFIER**/org.khronos.glslang/g' -e 's/**VERSION**/1.0/g' Info.plist > glslang.framework/Versions/A/Resources/Info.plist

打包 ncnn 框架：

cd 

mkdir -p ncnn.framework/Versions/A/Headers
mkdir -p ncnn.framework/Versions/A/Resources
ln -s A ncnn.framework/Versions/Current
ln -s Versions/Current/Headers ncnn.framework/Headers
ln -s Versions/Current/Resources ncnn.framework/Resources
ln -s Versions/Current/ncnn ncnn.framework/ncnn
lipo -create build-ios/install/lib/libncnn.a build-ios-sim/install/lib/libncnn.a -o ncnn.framework/Versions/A/ncnn
cp -r build-ios/install/include/\* ncnn.framework/Versions/A/Headers/
sed -e 's/**NAME**/ncnn/g' -e 's/**IDENTIFIER**/com.tencent.ncnn/g' -e 's/**VERSION**/1.0/g' Info.plist > ncnn.framework/Versions/A/Resources/Info.plist

选择 ncnn.framework、glslang.framework 和 openmp.framework 文件夹进行应用开发。

为 WebAssembly 构建

安装 Emscripten

git clone https://github.com/emscripten-core/emsdk.git
cd emsdk
./emsdk install 2.0.8
./emsdk activate 2.0.8

source emsdk/emsdk_env.sh

不带任何扩展名构建以实现通用兼容性：

mkdir -p build
cd build
cmake -DCMAKE_TOOLCHAIN_FILE=../emsdk/upstream/emscripten/cmake/Modules/Platform/Emscripten.cmake \
 -DNCNN_THREADS=OFF -DNCNN_OPENMP=OFF -DNCNN_SIMPLEOMP=OFF -DNCNN_RUNTIME_CPU=OFF -DNCNN_SSE2=OFF -DNCNN_AVX2=OFF -DNCNN_AVX=OFF \
 -DNCNN_BUILD_TOOLS=OFF -DNCNN_BUILD_EXAMPLES=OFF -DNCNN_BUILD_BENCHMARK=OFF ..
cmake --build . -j 4
cmake --build . --target install

带 WASM SIMD 扩展构建：

mkdir -p build-simd
cd build-simd
cmake -DCMAKE_TOOLCHAIN_FILE=../emsdk/upstream/emscripten/cmake/Modules/Platform/Emscripten.cmake \
 -DNCNN_THREADS=OFF -DNCNN_OPENMP=OFF -DNCNN_SIMPLEOMP=OFF -DNCNN_RUNTIME_CPU=OFF -DNCNN_SSE2=ON -DNCNN_AVX2=OFF -DNCNN_AVX=OFF \
 -DNCNN_BUILD_TOOLS=OFF -DNCNN_BUILD_EXAMPLES=OFF -DNCNN_BUILD_BENCHMARK=OFF ..
cmake --build . -j 4
cmake --build . --target install

带 WASM Thread 扩展构建：

mkdir -p build-threads
cd build-threads
cmake -DCMAKE_TOOLCHAIN_FILE=../emsdk/upstream/emscripten/cmake/Modules/Platform/Emscripten.cmake \
 -DNCNN_THREADS=ON -DNCNN_OPENMP=ON -DNCNN_SIMPLEOMP=ON -DNCNN_RUNTIME_CPU=OFF -DNCNN_SSE2=OFF -DNCNN_AVX2=OFF -DNCNN_AVX=OFF \
 -DNCNN_BUILD_TOOLS=OFF -DNCNN_BUILD_EXAMPLES=OFF -DNCNN_BUILD_BENCHMARK=OFF ..
cmake --build . -j 4
cmake --build . --target install

带 WASM SIMD 和 Thread 扩展构建：

mkdir -p build-simd-threads
cd build-simd-threads
cmake -DCMAKE_TOOLCHAIN_FILE=../emsdk/upstream/emscripten/cmake/Modules/Platform/Emscripten.cmake \
 -DNCNN_THREADS=ON -DNCNN_OPENMP=ON -DNCNN_SIMPLEOMP=ON -DNCNN_RUNTIME_CPU=OFF -DNCNN_SSE2=ON -DNCNN_AVX2=OFF -DNCNN_AVX=OFF \
 -DNCNN_BUILD_TOOLS=OFF -DNCNN_BUILD_EXAMPLES=OFF -DNCNN_BUILD_BENCHMARK=OFF ..
cmake --build . -j 4
cmake --build . --target install

选择 build-XYZ/install 文件夹用于进一步使用。

为 AllWinner D1 构建

从 https://occ.t-head.cn/community/download?id=3913221581316624384 下载 c906 工具链包。

tar -xf riscv64-linux-x86_64-20210512.tar.gz
export RISCV_ROOT_PATH=/home/nihui/osd/riscv64-linux-x86_64-20210512

启用 riscv-v 向量和 simpleocv 的 ncnn 构建：

mkdir -p build-c906
cd build-c906
cmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/c906.toolchain.cmake \
 -DCMAKE_BUILD_TYPE=relwithdebinfo -DNCNN_OPENMP=OFF -DNCNN_THREADS=OFF -DNCNN_RUNTIME_CPU=OFF -DNCNN_RVV=ON \
 -DNCNN_SIMPLEOCV=ON -DNCNN_BUILD_EXAMPLES=ON ..
cmake --build . -j 4
cmake --build . --target install

选择 build-c906/install 文件夹用于进一步使用。

你可以将二进制文件上传到 build-c906/examples 文件夹并在 D1 板上运行进行测试。

为 Loongson 2K1000 构建

对于 gcc 版本 <= 8.5，你需要修复 msa.h 头文件以解决 msa fmadd 错误。

打开 /usr/lib/gcc/mips64el-linux-gnuabi64/8/include/msa.h，找到 __msa_fmadd_w 并应用以下更改：

// #define **msa_fmadd_w **builtin_msa_fmadd_w
#define **msa_fmadd_w(a, b, c) **builtin_msa_fmadd_w(c, b, a)

启用 mips msa 和 simpleocv 的 ncnn 构建：

mkdir -p build
cd build
cmake -DNCNN_DISABLE_RTTI=ON -DNCNN_DISABLE_EXCEPTION=ON -DNCNN_RUNTIME_CPU=OFF -DNCNN_MSA=ON -DNCNN_MMI=ON -DNCNN_SIMPLEOCV=ON ..
cmake --build . -j 2
cmake --build . --target install

选择 build/install 文件夹用于进一步使用。

你可以运行 build/examples 文件夹中的二进制文件进行测试。

为 Android 上的 Termux 构建

在手机上安装 Termux 应用，并在 Termux 中安装 Ubuntu。

如果你想使用 ssh，只需在 Termux 中安装 openssh。

pkg install proot-distro
proot-distro install ubuntu

或者你可以使用 proot-distro list 查看可以安装的系统。

在成功安装 Ubuntu 后，使用 proot-distro login ubuntu 登录 Ubuntu。

然后构建 ncnn，无需安装任何其他依赖项。

git clone https://github.com/Tencent/ncnn.git
cd ncnn
git submodule update --init
mkdir -p build
cd build
cmake -DCMAKE_BUILD_TYPE=Release -DNCNN_BUILD_EXAMPLES=ON -DNCNN_PLATFORM_API=OFF -DNCNN_SIMPLEOCV=ON ..
make -j$(nproc)

然后你可以运行测试：

在我的 Pixel 3 XL 上使用 Qualcomm 845，无法加载 256-ncnn.png

cd ../examples
../build/examples/squeezenet ../images/128-ncnn.png

在 Android 11 上用 LLDV 调试原生应用

Fri, 11 Aug 2023 01:14:06 +0000

Mobile

将 PUSH 调试服务器推送到手机

adb push lldb-server /data/local/tmp
chmod 755 /data/local/tmp/lldb-server

启动调试器服务

/data/local/tmp/lldb-server platform --listen "*:8888" --server

+++

电脑端

端口转发

adb forward tcp:8888 tcp:8888

启动 LLDB

. \lldb.exe

查看支持的平台

platform list

选择 ANDROID 平台

platform select remote-android

连接到手机

手机序列号: 9643e0ec0604 要更改为当前调试的手机，请使用 adb devices 检查序列号

platform connect connect://9643e0ec0604:8888

查看当前运行的进程

platform process list

附加到

attach 9053

断点

b send

运行 (继续执行)

查看线程列表

thread list

查看调用堆栈

bt

Android on xgDebug的博客

使用ncnn布署pytorch模型到Android手机

使用 ncnn 布署 pytorch 模型到 Android 手机

編譯成 MT 時要打開 CMAKE 0091 特性

训练 YOLO

转换模型

Git clone ncnn repo with submodule

Build for Linux

Build for Windows x64 using Visual Studio Community 2017

Build for macOS

Build for ARM Cortex-A family with cross-compiling

Build for Hisilicon platform with cross-compiling

Build for Android

Build for iOS on macOS with xcode

Build for WebAssembly

Build for AllWinner D1

Build for Loongson 2K1000

Build for Termux on Android

在Android 11 上使用 LLDV 调试原生程序

手机端

PUSH 调试服务器到手机

启动调试器服务

电脑端

端口转发

启动 LLDB

查看支持平台

选 ANDROID 平台

连接到手机 手机序列号: 9643e0ec0604 要换成当前调试的手机,使用 adb devices 查看序列号

查看当前正在运行的进程

附加上去

下断点

跑起来

查看线程列表

查看调用栈

用 ncnn 部署 PyTorch 模型到 Android 手机

使用 ncnn 将 pytorch 模型部署到安卓手机

在编译到 MT 时开启 CMAKE 0091 特性

训练 YOLO

模型转换

Git 克隆 ncnn 仓库及子模块

为 Linux 构建

使用 Visual Studio Community 2017 为 Windows x64 构建

为 macOS 构建

使用交叉编译为 ARM Cortex-A 系列构建

使用交叉编译为 Hisilicon 平台构建

为 Android 构建

在 macOS 上使用 xcode 为 iOS 构建

为 WebAssembly 构建

为 AllWinner D1 构建

为 Loongson 2K1000 构建

为 Android 上的 Termux 构建

在 Android 11 上用 LLDV 调试原生应用

Mobile

将 PUSH 调试服务器推送到手机

启动调试器服务

电脑端

端口转发

启动 LLDB

查看支持的平台

选择 ANDROID 平台

连接到手机

查看当前运行的进程

附加到

断点

运行 (继续执行)

查看线程列表

查看调用堆栈

连接到手机手机序列号: 9643e0ec0604 要换成当前调试的手机,使用 adb devices 查看序列号