CUDA distutils

Introduction

After reading McGibbon's npcuda-example, I set out on an attempt at creating extension modules by creating a separate nvcc cuda compiler to the distutils process.

Obviously, most of the idea is inspired from npcuda-example. But instead of doing runtime hot-patches to the compiler object, this goes via the OO way, in the hope of keeping things cleaner.

Example Usage

Taking the setup.py of npcuda-example/cython/setup.py as an example, here's what it would take to add support for building extension modules having CUDA kernels.

import os
from setuptools import setup
from distutils.extension import Extension
import cudistutils as cud
import numpy as np
env = cud.CudaEnv()
ext = Extension("gpuadder",
                sources=["src/manager.cu", "wrapper.pyx"],
                library_dirs=[env.lib64],
                libraries=env.base_cuda_libs,
                language="c++",
                runtime_library_dirs=[env.lib64],
                extra_compile_args={"gcc"  : [],
                                    "nvcc" : env.default_nvcc_opts()},
                include_dirs = [np.get_include(), env.include, "src"])
setup(name="gpuadder",
      author="Robert McGibbon",
      version="0.1",
      ext_modules = [ext],
      cmdclass={"build_ext": cud.cuda_build_ext},
      zip_safe=False)

As can be seen, most of your setup.py remains similar to writing C/C++ extension modules. Only change comes with passing 'extra_compile_args', which is needed since the underlying compiler class switches between C/C++ compiler and cuda compiler based on the input source file extension.

The project is hosted here. Hopefully will spend some time to get it accessible via 'pip install'.