ashvardanianashvardanian
31

PyBindToGPUs

screenshot of PyBindToGPUs

Parallel Computing starter project to build GPU & CPU kernels in CUDA & C++ and call them from Python without a single line of CMake using PyBind11

GithubDemo

Overview:

The C++ & CUDA Starter Kit for Python Developers is an essential tool for anyone involved in high-performance computing. With a streamlined workflow for prototyping algorithms in Python before porting them to C++ and CUDA, it alleviates the common headaches associated with configuring build tools for heterogeneous code and hardware. This kit simplifies the development process significantly, enabling developers to focus on optimizing their algorithms rather than getting bogged down by configuration complexities.

This project provides a pre-configured environment that fosters swift development and testing. By supporting both CPU and GPU implementations seamlessly, it promises to enhance productivity while ensuring that developers can leverage the full potential of CUDA and OpenMP for their computational needs.

Features:

  • Pre-Configured Environment: Set up effortlessly with only setup.py and requirements-{cpu,gpu}.txt, eliminating the cumbersome configuration typically associated with such projects.
  • Support for Parallelism: Utilizes OpenMP for CPU parallelism and CUDA for GPU acceleration, enabling efficient performance across different hardware configurations.
  • Utilization of CCCL Libraries: Integrates libraries like Thrust and CUB to simplify code development, allowing for cleaner implementations of complex algorithms.
  • Baseline Implementation Provided: Includes a basic Python implementation using Numba for array accumulation and matrix multiplication, providing a clear starting point for optimization.
  • Customized Debugging Setup: For VSCode users, the tasks.json is pre-configured to support debugging for both CPU and GPU code, enhancing the development experience.
  • Flexible Installation Options: Users can easily fork or clone the repository, making it convenient to adapt the starter kit for personal projects.
  • Readily Adaptable Workflow: Designed for simplicity, guiding developers through implementing baseline and optimized algorithms efficiently.
  • Extensive Learning Materials: Comes with a compilation of resources for beginners and advanced users alike, helping them deepen their understanding of CUDA and GPGPU concepts.