MUDA - CUDA on Metal

MUDA provides a CUDA-like programming interface that runs on macOS by leveraging Apple's Metal framework. This project enables CUDA applications to be ported to macOS with minimal code changes, bridging the gap between NVIDIA CUDA and Apple's Metal compute capabilities.

• ✅ Provide CUDA runtime API compatibility layer
• ✅ Enable CUDA code migration to macOS/Metal
• ✅ Support FFT operations via Apple's vDSP
• ✅ Maintain familiar CUDA programming patterns

block-beta
    columns 1
    block
        A["Your Application<br/>CUDA-compatible code"]
        style A fill:#f9f,stroke:#333,stroke-width:2px
    end
    block
        B["CUDA Runtime API<br/>cudart - CUDA Runtime Library<br/>cudaMalloc, cudaMemcpy, etc."]
        C["CUDA Driver API<br/>cuda - Driver Interface<br/>cuModuleLoad, cuLaunchKernel, etc."]
        style B fill:#bbf,stroke:#333,stroke-width:2px
        style C fill:#bbf,stroke:#333,stroke-width:2px
    end
    block
        D["metal-cpp<br/>Apple's Metal C++ Wrapper"]
        E["Metal<br/>Apple's GPU Computing Framework"]
        style D fill:#bfb,stroke:#333,stroke-width:2px
        style E fill:#bfb,stroke:#333,stroke-width:2px
    end
    block
        F["GPU Hardware<br/>Apple Silicon / Intel Mac"]
        style F fill:#86f,stroke:#333,stroke-width:2px
    end

• macOS 12.0+ (Monterey or later)
• Xcode 14.0+ with command-line tools
• Apple Silicon (M1/M2/M3) or Intel Mac with Metal

# Clone the repository
git clone https://github.com/yourusername/muda.git
cd muda

# configure
cmake -B build

# build
cmake --build build

#include <cuda.h>
#include <cuda_runtime_api.h>

int main() {
    // Initialize CUDA
    int deviceCount;
    cudaGetDeviceCount(&deviceCount);

    if (deviceCount == 0) {
        return CUDA_ERROR_NO_DEVICE;
    }

    // Set device
    cudaSetDevice(0);

    // Allocate memory
    float* d_data;
    cudaMalloc(&d_data, sizeof(float) * 1024);

    // Load and launch kernel
    CUmodule module;
    cuModuleLoad(&module, "mykernel.metallib");

    CUfunction kernel;
    cuModuleGetFunction(&kernel, module, "myKernel");

    cuLaunchKernel(kernel, 1, 1, 1, 256, 1, 1, 0, nullptr, nullptr, nullptr);

    // Cleanup
    cudaFree(d_data);
    cuModuleUnload(module);

    return cudaSuccess;
}

MUDA requires Metal shader libraries (.metallib files). Compile your shaders using metal:

# Compile .metal to .metallib
xcrun metal myshader.metal -o myshader.air

#include <metal_stdlib>
using namespace metal;

[[kernel]] void myKernel(device float* data [[buffer(0)]],
                     uint id [[thread_position_in_grid]]) {
    data[id] = data[id] * 2.0f;
}

• ⚠️ NVIDIA-specific features (Tensor Cores, cuBLAS, cuDNN) are not available
• ⚠️ PTX assembly is not supported; use Metal Shading Language
• ⚠️ Some advanced CUDA features may have different semantics
• ⚠️ Performance characteristics differ from NVIDIA GPUs

Contributions are welcome! Please read our contributing guidelines before submitting PRs.

1. Do NOT modify metal-cpp/ - This is Apple's official wrapper
2. Follow the existing code style
3. Add tests for new functionality
4. Update documentation accordingly

This project is licensed under the MIT License - see the LICENSE file for details.

• Apple for Metal and metal-cpp
• NVIDIA for the CUDA API specification
• Apple's Accelerate framework for vDSP FFT implementation