ringkernel-metal

Apple Metal backend for RingKernel.

Status

This crate is scaffolded - the API is defined but implementation is pending.

Overview

This crate will provide GPU compute support for RingKernel on Apple platforms using the Metal framework. It targets macOS, iOS, and Apple Silicon devices.

Requirements (Planned)

macOS 10.15+ or iOS 13+
Apple Silicon (M1/M2/M3) or compatible AMD GPU
Metal feature enabled

Features (Planned)

Event-driven kernel execution via Metal compute shaders
MSL (Metal Shading Language) shader support
Apple Silicon unified memory architecture optimization
iOS deployment support

Limitations

Like WebGPU, Metal does not support CUDA-style cooperative groups, so persistent kernels will use event-driven execution with host-side dispatch loops.

Usage (API Preview)

use ringkernel_metal::MetalRuntime;

#[tokio::main]
async fn main() -> Result<(), Box<dyn std::error::Error>> {
    if !ringkernel_metal::is_metal_available() {
        eprintln!("Metal not available");
        return Ok(());
    }

    let runtime = MetalRuntime::new().await?;
    let kernel = runtime.launch("compute", Default::default()).await?;

    // Process messages...

    kernel.terminate().await?;
    runtime.shutdown().await?;
    Ok(())
}

Current State

The crate currently provides:

Stub MetalRuntime that returns BackendUnavailable errors
MSL kernel template (RING_KERNEL_MSL_TEMPLATE)
is_metal_available() runtime check

Contributing

Contributions to implement the Metal backend are welcome. Key work items:

Device enumeration via metal::Device
Compute pipeline creation
Buffer management with unified memory
Kernel dispatch and synchronization

License

Apache-2.0

ringkernel-metal 0.1.3