Skip to content

Latest commit

 

History

History
129 lines (97 loc) · 4.28 KB

README_cn.md

File metadata and controls

129 lines (97 loc) · 4.28 KB
Raw

飞桨自定义接入硬件后端(Custom CPU)

简体中文 | English | 日本語

请参考以下步骤进行硬件后端(Custom CPU)的编译安装与验证

一、环境准备与源码同步

# 1) 拉取镜像，注意此镜像仅为开发环境，镜像中不包含预编译的飞桨安装包
#    此镜像的构建脚本与 dockerfile 位于 tools/dockerfile 目录下
docker pull registry.baidubce.com/device/paddle-cpu:ubuntu20-x86_64-gcc84-py310
docker pull registry.baidubce.com/device/paddle-cpu:ubuntu20-aarch64-gcc84-py310

# 2) 参考如下命令启动容器
docker run -it --name paddle-dev-cpu -v `pwd`:/workspace \
  --network=host --shm-size=128G --workdir=/workspace \
  --cap-add=SYS_PTRACE --security-opt seccomp=unconfined \
  registry.baidubce.com/device/paddle-cpu:ubuntu20-$(uname -m)-gcc84-py310 /bin/bash

# 3) 克隆源码，注意 PaddleCustomDevice 依赖 PaddlePaddle 主框架源码
git clone --recursive https://github.com/PaddlePaddle/PaddleCustomDevice
cd PaddleCustomDevice

# 4) 请执行以下命令，以保证 checkout 最新的 PaddlePaddle 主框架源码
git submodule sync
git submodule update --remote --init --recursive

二、编译安装

# 进入硬件后端(Custom CPU)目录
cd backends/custom_cpu

# 编译之前需要先保证环境下装有Paddle WHL包，可以直接安装CPU版本
pip install paddlepaddle==0.0.0 -f https://www.paddlepaddle.org.cn/whl/linux/cpu-mkl/develop.html

# 创建编译目录并编译
mkdir build && cd build

cmake ..
make -j8

# 编译产出在dist路径下，使用pip安装
pip install dist/paddle_custom_cpu*.whl

三、功能验证

# 列出可用硬件后端
python -c "import paddle; print(paddle.device.get_all_custom_device_type())"
# 期待输出以下结果
['custom_cpu']

# 运行简单模型
python ../tests/test_MNIST_model.py
# 期待输出以下类似结果
... ...
Epoch 0 step 0, Loss = [2.2956038], Accuracy = 0.15625
Epoch 0 step 100, Loss = [2.1552896], Accuracy = 0.3125
Epoch 0 step 200, Loss = [2.1177733], Accuracy = 0.4375
Epoch 0 step 300, Loss = [2.0089214], Accuracy = 0.53125
Epoch 0 step 400, Loss = [2.0845466], Accuracy = 0.421875
Epoch 0 step 500, Loss = [2.0473], Accuracy = 0.453125
Epoch 0 step 600, Loss = [1.8561764], Accuracy = 0.71875
Epoch 0 step 700, Loss = [1.9915285], Accuracy = 0.53125
Epoch 0 step 800, Loss = [1.8925955], Accuracy = 0.640625
Epoch 0 step 900, Loss = [1.8199624], Accuracy = 0.734375

四、使用PaddleInference

重新编译插件

# 编译PaddleInference
git clone https://github.com/PaddlePaddle/Paddle.git
git clone https://github.com/ronny1996/Paddle-Inference-Demo.git

mkdir -p Paddle/build
pushd Paddle/build

cmake .. -DPY_VERSION=3.7 -DWITH_GPU=OFF -DWITH_TESTING=ON -DCMAKE_BUILD_TYPE=Release -DON_INFER=ON -DWITH_MKL=ON -DWITH_CUSTOM_DEVICE=ON

make -j8

popd
cp -R Paddle/build/paddle_inference_install_dir Paddle-Inference-Demo/c++/lib/paddle_inference
export PADDLE_INFERENCE_LIB_DIR=$(realpath Paddle-Inference-Demo/c++/lib/paddle_inference/paddle/lib)

# 编译插件
mkdir -p PaddleCustomDevice/backends/custom_cpu/build
pushd PaddleCustomDevice/backends/custom_cpu/build

cmake .. -DON_INFER=ON -DPADDLE_INFERENCE_LIB_DIR=${PADDLE_INFERENCE_LIB_DIR}
make -j8

# 指定插件路径
export CUSTOM_DEVICE_ROOT=$PWD
popd

使用 PaddleInference

pushd Paddle-Inference-Demo/c++/resnet50

# 修改 resnet50_test.cc，使用 config.EnableCustomDevice("custom_cpu", 0) 接口替换 config.EnableUseGpu(100, 0)

bash run.sh

期待输出以下类似结果

I0713 09:02:38.808723 24792 resnet50_test.cc:74] run avg time is 297.75 ms
I0713 09:02:38.808859 24792 resnet50_test.cc:89] 0 : 8.76192e-29
I0713 09:02:38.808894 24792 resnet50_test.cc:89] 100 : 8.76192e-29
I0713 09:02:38.808904 24792 resnet50_test.cc:89] 200 : 8.76192e-29
I0713 09:02:38.808912 24792 resnet50_test.cc:89] 300 : 8.76192e-29
I0713 09:02:38.808920 24792 resnet50_test.cc:89] 400 : 8.76192e-29
I0713 09:02:38.808928 24792 resnet50_test.cc:89] 500 : 8.76192e-29
I0713 09:02:38.808936 24792 resnet50_test.cc:89] 600 : 1.05766e-19
I0713 09:02:38.808945 24792 resnet50_test.cc:89] 700 : 2.04093e-23
I0713 09:02:38.808954 24792 resnet50_test.cc:89] 800 : 3.85255e-25
I0713 09:02:38.808961 24792 resnet50_test.cc:89] 900 : 8.76192e-29