How to compile OpenCV to use CUDA within a Docker image

In order to compile OpenCV to use CUDA we need to understand first how the C++ build process works:

• Preprocessing: The preprocessor does many things: macro replacements, check for conditional compilation, and insert content whenever a header file is included. Only source files are passed to the preprocessor.
  • input: C++ source code
  • output: file.ii
• Compiler (GCC: GNU COMPILER COLLECTION): Transforms the high-level programming language to a low-level one.
  • input: file.ii (preprocessed file)
  • output (low level language): assembly code/file --> file.s
• Assembler: Machine code representation of your source code.
  • input: assembly code/file --> file.s
  • output: object code --> file.obj / file.o
• Linker: Link the Standard and External libraries and generate the executable file.
  • input: object code --> file.obj / file.o
  • output: file.exe

Now, we can proceed to compile the OpenCV source code in order for it to make some processes directly in a GPU.

1. You need to create a folder to clone OpenCV and OpenCV-contrib repositories there (if their latest versions are too recent, I would suggest trying with the second to last)
2. Check which GPU you have (run nvidia-smi) and its requirements.
   • Here CUDA-GPUS you can find the computing capability of the card.
   • Here CUDA TOOLKIT, CUDA DRIVER AND CUDNN you can find the supported versions regarding the GPU card.
3. Go to Docker hub and choose a Nvidia Docker image matching your GPU requirements.
   • Be aware that there are different types of images, choose one that satisfies your project requirements.
4. Create a Docker container based on the Nvidia Docker image you selected docker run --net=host --runtime=nvidia -it -v <path/to/folder/in/step1>:/main_dir <Nvidia Docker image> /bin/bash (Ex: nvidia/cuda:12.0.0-cudnn8-devel-ubuntu20.04) (Preferably use a Docker image with CUDNN, otherwise you shall install it manually).
5. Once the container is running you have to install:

• Python (Keep in mind to choose the Python version you need):
  • apt-get update && apt-get install -y software-properties-common && rm -rf /var/lib/apt/lists/ && add-apt repository ppa:deadsnakes/ppa
  • apt-get install -y python3.10 python3.10-dev libopencv-dev
  • alias python=python3.10 && alias python3=python3.10
  • update-alternatives --install /usr/bin/python3 python3 /usr/bin/python3.10 1
  • update-alternatives --config python3
• Build OpenCV:
  • Go to the OpenCV folder (cd path/to/opencv/folder)
  • Run mkdir build, then cd build
  • Run cmake
    • Change the GPU architecture CUDA_ARCH_BIN (compute capability from step 2)
    • Change the Python folder path PYTHON_EXECUTABLE
    • Change the CUDA folder path CUDA_TOOLKIT_ROOT_DIR
    • Change the OpenCV contrib modules OPENCV_EXTRA_MODULES_PATH
    • Finally there are some parameters that can be modified to your convenience. Please check the docs to analyze what is best for your project.
      
      Example of a cmake command:
      cmake -D CMAKE_BUILD_TYPE=Release -D BUILD_PNG=OFF -D BUILD_TIFF=OFF -D BUILD_TBB=OFF -D BUILD_JPEG=OFF -D BUILD_JASPER=OFF -D BUILD_ZLIB=OFF -D BUILD_EXAMPLES=OFF -D BUILD_opencv_java=OFF -D BUILD_opencv_python2=OFF -D BUILD_opencv_python3=ON -D ENABLE_NEON=ON -D WITH_OPENCL=OFF -D WITH_OPENMP=OFF -D WITH_FFMPEG=ON -D WITH_GSTREAMER=OFF -D WITH_GSTREAMER_0_10=OFF -D WITH_CUDA=ON -D CUDA_TOOLKIT_ROOT_DIR=/usr/local/cuda -D WITH_GTK=ON -D WITH_VTK=OFF -D WITH_TBB=ON -D WITH_1394=OFF -D WITH_OPENEXR=OFF -D CUDA_ARCH_BIN=7.5 -D CUDA_ARCH_PTX="" -D INSTALL_C_EXAMPLES=OFF -D INSTALL_TESTS=OFF -D WITH_CUDNN=ON -D OPENCV_DNN_CUDA=ON -D ENABLE_FAST_MATH=1 -D CUDA_FAST_MATH=1 -D PYTHON_EXECUTABLE=/usr/bin/python3 -D WITH_CUBLAS=1 -D OPENCV_EXTRA_MODULES_PATH=/main/opencv_contrib/modules ..
  • Run make install

7. If everything went well now you should have a Docker image with CUDA, CUDNN and OpenCV ready to use CUDA! Now, you can exit the container without killing it (Ctrl+p, Ctrl+q) and save a Docker image based on that container docker commit -m "cudaxx-cudnxx-opencvx.x" cudaxx-cudnnxx-opencvx.x:1
   • To check OpenCV build information, print cv2.getBuildInformation() (NVIDIA CUDA AND CUDNN should be marked as YES)
8. Finally you can kill the running container and start using the Docker image for your projects. (Keep in mind that this process can also be summarized in a Dockerfile, but it is always good to understand step by step what is happening during the building process)