# [Neural Network CUDA](https://github.com/godweiyang/NN-CUDA-Example)

We provide several ways to compile the CUDA kernels and their cpp wrappers, including setuptools and cmake.

## Code Structure

```
├── csrc
│   |── cpu
|   |     ├── nms_3d_cpu.cpp
|   |     └── vision.h
│   |── cuda
|   |     ├── deformable_conv_3d_cuda.cu
|   |     ├── modulated_deformable_conv_3d_cuda.cu
|   |     ├── nms_3d.cu
|   |     ├── overlap_3d.cu
|   |     ├── ROIAlign_3d_cuda.cu
|   |     └── vision.h
│   |── custom_ops
|   |     └── custom_ops.cpp
│   |── deformable_conv_3d.h
│   |── modulated_deformable_conv_3d.h
│   |── nms_3d.h
│   |── overlap_3d.h
│   └── ROIAlign_3d.h
├── custom_op
│   └── setup.py
│── deploy
│   └── custom_op_cmake
│         |── CMakeLists.txt
│         |── custom_ops.cpp
│         |── nms_3d.cu
│         |── nms_3d.h
│         |── ROIAlign_3d.h
│         |── ROIAlign_3d_cuda.cu
│         └── vision.h
└── layers
    |── lib
    |     └── custom_ops.cpp
    |── __init__.py
    |── deformable_conv_3d.py
    |── modulated_deformable_conv_3d.py
    |── nms_3d.py
    |── overlap_3d.py
    └── roi_align_3d.py
```



## PyTorch

### Compile cpp and cuda

- Setuptools

    ```shell
    # 用于python端训练
    cd custom_op
    python setup.py install
    ```

- CMake

    ```shell
    # 用于c++端部署，亦可用于python端训练
    # 在nvcr.io/nvidia/pytorch镜像内编译，版本与部署一致，例如：20.08
    cd deploy/custom_op_cmake
    mkdir build
    cd build
    cmake ../ -DCMAKE_PREFIX_PATH=/opt/conda/lib/python3.6/site-packages/torch
    make -j 8
    ```