setuptools-cpp-cuda 0.1.6

The setuptools-cpp-cuda is a module that extends setuptools functionality for building hybrid C++ and CUDA extensions

Setuptools C++ CUDA

The setuptools-cpp-cuda is a module that extends setuptools functionality for building hybrid C++ and CUDA extensions for Python wrapper modules.

---

Table of Contents

• Summary
• Features
• Installation
• Usage
• Issues
• License
• Acknowledgements

Summary

This project meant to be a soft solution to include mixed c++/CUDA extensions in your projects, no matter if you are using old python version (3.6+) or old GPU drivers (sm/compute arch 3.0+).

Features

• Python version >= 3.6 .
• SM(StreamingMultiprocessor)/Compute architecture >= 3.0 .
• Cython compatible but not mandatory.
• Any CUDA version (since you can configure nvcc flags).
• Preloaded flags for cpp and CUDA compilers.
• Mixed compilations (.cpp and .cu files can be included in a single extension).
• Advanced find_cuda features (automatically try to find the CUDAHOME directory).
• Include NVIDIA Management Library (NVML) capabilities info.

Installation

pip install setuptools-cpp-cuda

Usage

Add the library to your project configuration files ("pyproject.toml" and/or "setup.py/.cfg").

1. Example for "legacy build" (old python versions with setuptools < 61.0.0):

setup.py

from pathlib import Path
from setuptools import setup
from setuptools_cpp_cuda import CUDAExtension, BuildExtension, fix_dll

cuda_ext_path = Path('src/my_cuda_package/cuda_ext')
cuda_ext = CUDAExtension(
    name='my_cuda_package.cuda_ext',
    include_dirs=[cuda_ext_path / 'include'],
    sources=[
        cuda_ext_path / 'cuda_ext.cu',
        cuda_ext_path / 'cuda_ext_wrapper.cpp',
    ],
    libraries=fix_dll(['cudart']),  # Use fix_dll() only for Windows compatibility (check documentation for more info).
    extra_compile_args={
        'cxx': ['-g'],  # cpp compiler flags
        'nvcc': ['-O2'],  # nvcc flags
    },
)

setup(
    name='my-cuda-package',
    version='0.0.1',
    install_requires=['numpy', ],
    extras_require={'cython': ['cython'], },
    ext_modules=[cuda_ext],
    cmdclass={'build_ext': BuildExtension},
)

You can also use pyproject.toml with Flit making a custom build-backend.

2. Example for "pyproject.toml build" (with setuptools >= 61.0.0):

pyproject.toml

[build-system]
requires = ["setuptools-cpp-cuda", "flit_core >=3.2,<4", "wheel", "cython"]
build-backend = "flit_core.buildapi"

[project]
name = "my-cuda-package"
dependencies = ["numpy"]
dynamic = ["version", "description"]
# ...

And configure the setup.py for the different extensions you want to use:

setup.py

from pathlib import Path
from setuptools import setup
from setuptools_cpp_cuda import CUDAExtension, BuildExtension, fix_dll

cuda_ext_path = Path('src/my_cuda_package/cuda_ext')
cuda_ext = CUDAExtension(
    name='my_cuda_package.cuda_ext',
    include_dirs=[cuda_ext_path / 'include'],
    sources=[
        cuda_ext_path / 'cuda_ext.cu',
        cuda_ext_path / 'cuda_ext_wrapper.cpp',
    ],
    libraries=fix_dll(['cudart']),  # Use fix_dll() only for Windows compatibility (check documentation for more info).
    extra_compile_args={
        'cxx': ['-g'],  # cpp compiler flags
        'nvcc': ['-O2'],  # nvcc flags
    },
)

setup(
    ext_modules=[cuda_ext],
    cmdclass={'build_ext': BuildExtension},
)

Issues

If you receive a EnvironmentError exception you should set CUDAHOME environment variable pointing to the CUDA installation path. This would happen if the find_cuda() method is not capable of locate it. As reference the directory should contain:

CUDAHOME
├── bin
│   └── nvcc
├── include
│   └── cudart.h
├── lib
└── nvml

License

setuptools-cpp-cuda is distributed under the terms of the MIT license.

Acknowledgements

The package is based on cpp_extension, but it also includes:

• Support for deprecated older architectures (from sm / compute 3.0).
• Improved find_cuda system.
• Pathlib library and Windows missing dll support.