metal

Typed Apple Metal compute bindings, over the Foundation, Metal, and MetalPerformanceShaders frameworks. metal::default_device() returns a Device; chain into new_command_queue(), new_buffer(len), new_library_with_data(bytes), and new_compute_pipeline_state(fn). Every wrapper has a Drop impl that objc_releases the underlying object.

The metal/mps sub-module adds MPS bindings: Apple's pre-tuned matmul, FFT, and softmax. MPSMatrixMultiplication does batched gemm:

import "metal/metal" as metal;
import "metal/mps" as mps;

guard let result::Result::Ok(dev) = metal::default_device() else { return 1; };
let q  = dev.new_command_queue();
let mm = mps::MatrixMultiplication::new(dev, false, false, M, N, K, 1.0f64, 0.0f64);
mm.encode_to(cmd_buf, lhs, rhs, out);

For the CPU fallback path and result-checking reference, see accelerate.