gpu.cpp provides a C++ library for portable, low-level GPU computation using the WebGPU specification. It targets individual developers and researchers seeking to integrate GPU acceleration into native applications with minimal boilerplate and dependencies, enabling fast iteration cycles across diverse hardware.

How It Works

The library leverages WebGPU as a cross-platform, low-level GPU API, allowing C++ code to run on GPUs from various vendors (Nvidia, Intel, AMD) via Vulkan, Metal, or DirectX backends. It emphasizes a small API surface area and minimal dependencies, offering a header-only C++ interface. For convenience, it bundles a pre-built binary of Google's Dawn WebGPU implementation, avoiding lengthy build times for the WebGPU backend during development.

Quick Start & Requirements

• Install/Run: make from the top-level directory.
• Prerequisites: clang++ (C++17), python3, make. Linux requires Vulkan drivers.
• Setup: The first run automatically downloads the Dawn shared library. Build and execution take a few seconds.
• Details: examples/hello_world/run.cpp

Highlighted Details

• Lightweight, header-only C++ core implementation (~1000 lines).
• Pre-built Dawn WebGPU implementation for rapid setup.
• Supports MacOS, Linux, and Windows (WSL).
• Example matrix multiplication achieves ~3.5+ TFLOPs on a Macbook Pro M1 Max.

Maintenance & Community

• Developed by Answer.AI.
• Community channel: #gpu-cpp on AnswerDotAI Discord.
• Contact: @austinvhuang on X.

Licensing & Compatibility

• License: MIT (for gpu.cpp itself). Dawn's license is not explicitly stated in the README but is typically BSD-style.
• Compatibility: Suitable for commercial use and linking with closed-source applications.

Limitations & Caveats

Browser targets are a short-term priority and not yet tested. The library currently lacks a reusable kernel library, with WGSL implementations residing within examples.