python_cuda

POC for using CUDA from Python via pybind11

Source files

• *.cu: They contain the GPU kernel codes and the C++ interface definitions.
• vector.hpp: A header for the Vector class wrapping the numpy array and the CUDA pointer.
• vector.cpp: The implementation of the Vector class.
• vector_functions.cuh: A header file for the C++ interfaces of the CUDA functions defined in the .cu files.
• exceptions.hpp: A header file for the C++ exceptions.
• python_cuda.cpp: It contains the Python interfaces.

Build process

I am sure there are nicer ways do to compile and link the code, but here is the one used by the build.sh:

• It compiles the .cu files to .o files with nvcc
• It compiles the .cpp files to .o files with nvcc
• It links the .o files to a python_cuda.so file that can be used from Python with nvcc

Docker

I use Docker for several reasons, one of them being easier dependency management (compared to doing it directly in Windows).

To create the docker image, you can use the following command:

docker build -t python_cuda_dev .

And to run the container with the current directory mounted to /workspace, you can use the following command:

docker run --rm -ti --gpus all -v .:/workspace python_cuda_dev

An appropriate NVIDIA GPU is required to use the project, and you need to assign it to the docker container. This is done by the "--gpus all" flag, which makes all of the gpus available for the container. A possible indicator of no available gpu is the following runtime error while trying to use compiled cuda code:

RuntimeError: CUDA driver version is insufficient for CUDA runtime version