sillyCUDA

A compatibility runtime to execute CUDA code on Apple Silicon chips by translating calls to Metal.

📖 About

sillyCUDA is an open-source research project designed to bridge the gap between NVIDIA's proprietary CUDA ecosystem and Apple's Metal API. The goal is to build a translation layer that intercepts standard CUDA calls and executes them on Apple Silicon (M1/M2/M3/M4) GPUs using Metal Compute.

This project is primarily an educational journey into GPU architecture, driver development, and cross-platform compatibility. It is not intended to replace NVIDIA hardware for high-performance enterprise workloads, but rather to allow students, researchers, and "Mac enjoyers" to run and study CUDA kernels without needing a dedicated green team GPU.

Read the full introduction story on Medium: > Introducing sillyCUDA: My Quest to Learn CUDA as a Mac Enjoyer

🏗 Architecture

sillyCUDA operates by mimicking the CUDA C API and translating logic in real-time. The architecture consists of three main components:

1. The Runtime (cuda_runtime_api):
   • Intercepts standard calls like cudaMalloc and cudaLaunchKernel.
   • Manages memory allocation via MTLBuffer instead of NVIDIA drivers.
2. The Translator:
   • Parses CUDA kernels (C++ dialect).
   • Generates an Abstract Syntax Tree (AST).
   • Transpiles logic into valid Metal Shading Language (MSL).
3. The Metal Backend:
   • Handles the execution pipeline on Apple Silicon.
   • Manages command queues and compute states.

🚀 Status & Roadmap

Currently, the project is in the initial development phase.

• Phase 1: Mastering GPU Architecture & Basics (Vector Add, Matrix Mult).
• Phase 2: Native Metal reimplementation of CUDA concepts.
• Phase 3 (Current Goal): The "Vertical Slice" — getting a standard Vector Addition CUDA program to compile and run on macOS without modifying source code.

Scope & Limitations (v1.0 Goals)

Feature	Status
Basic memory management (malloc, memcpy)	✅ Planned
Kernel launches	✅ Planned
Simple math operations	✅ Planned
Texture memory	❌ Out of scope
Advanced shared memory	❌ Out of scope
Highly optimized cuBLAS libraries	❌ Out of scope

Target Performance: ~50% of native Metal execution speed.

🛠 Building & Installation

Requirements: macOS 13+, Xcode Command Line Tools, CMake.

# Clone the repository
git clone [https://github.com/lwi00/sillyCUDA.git](https://github.com/lwi00/sillyCUDA.git)
cd sillyCUDA

# Create build directory
mkdir build && cd build

# Configure and build
cmake ..
make

🤝 Contributing

This is an open journey! Whether you are a student learning compilers, a graphics engineer, contributions are welcome.

• Report Bugs: Open an issue if you find something broken (there will be many).

• Discuss: Share ideas on how to improve the translation layer.

📄 License

This project is licensed under the Apache 2.0 License. See the LICENSE file for details.

Made with ❤️ and ☕️ by lwi00.