如何在Jetson上编译PaddlePaddle

本帖子会汇总描述在Jetson上编译PaddlePaddle的方法。

下载代码

1. git clone https://github.com/PaddlePaddle/Paddle

复制代码

Jetpack 4.3 + Paddle v2.0.0-alpha0

1. 
   
2. mkdir build
   
3. cd build
   
4. git checkout v2.0.0-alpha0
   
5. cmake ..  -DWITH_CONTRIB=OFF -DWITH_MKL=OFF -DWITH_MKLDNN=OFF -DWITH_TESTING=OFF -DCMAKE_BUILD_TYPE=Release -DON_INFER=ON -DWITH_PYTHON=OFF -DWITH_XBYAK=OFF -DWITH_NV_JETSON=ON -DWITH_NCCL=OFF
   

复制代码

在编译过程中，你会遇到如最下方日志的错误。
修改/usr/include/aarch64-linux-gnu/NvInferRuntime.h 在IPluginFactory增加

1.     virtual ~IPluginFactory();

复制代码

修改/usr/include/aarch64-linux-gnu/NvInferRuntime.h
把下列代码移到public部分

1. ~IOptimizationProfile() noexcept = default;

复制代码

/home/juns/Paddle/paddle/fluid/inference/tensorrt/plugin/trt_plugin_factory.h:35:7: error: base class ‘class nvinfer1::IPluginFactory’ has accessible non-virtual destructor [-Werror=non-virtual-dtor]
class PluginFactoryTensorRT : public nvinfer1::IPluginFactory,
       ^~~~~~~~~~~~~~~~~~~~~
In file included from /usr/include/c++/7/memory:80:0,
                 from /home/juns/Paddle/paddle/fluid/inference/tensorrt/op_teller.h:16,
                 from /home/juns/Paddle/paddle/fluid/inference/tensorrt/op_teller.cc:15:
/usr/include/c++/7/bits/unique_ptr.h: In instantiation of ‘void std::default_delete<_Tp>:perator()(_Tp*) const [with _Tp = nvinfer1::IOptimizationProfile]’:
/usr/include/c++/7/bits/unique_ptr.h:263:17:   required from ‘std::unique_ptr<_Tp, _Dp>::~unique_ptr() [with _Tp = nvinfer1::IOptimizationProfile; _Dp = std::default_delete<nvinfer1::IOptimizationProfile>]’
/home/juns/Paddle/paddle/fluid/inference/tensorrt/engine.h:138:23:   required from here
/usr/include/c++/7/bits/unique_ptr.h:78:2: error: ‘nvinfer1::IOptimizationProfile::~IOptimizationProfile()’ is protected within this context
  delete __ptr;
  ^~~~~~
In file included from /usr/include/aarch64-linux-gnu/NvInfer.h:53:0,
                 from /home/juns/Paddle/paddle/fluid/inference/tensorrt/engine.h:17,
                 from /home/juns/Paddle/paddle/fluid/inference/tensorrt/op_teller.h:21,
                 from /home/juns/Paddle/paddle/fluid/inference/tensorrt/op_teller.cc:15:
/usr/include/aarch64-linux-gnu/NvInferRuntime.h:1098:5: note: declared protected here
     ~IOptimizationProfile() noexcept = default;

Jetpack 4.4DP + Paddle v2.0.0-alpha0

目前由于Paddle还不能支持CUDNN8，因此目前还不能在Jetpack 4.4DP上使用Paddle最新版本。