Metal Matrices

Just a fun project I've been working on. I'm using the metal-rs crate to interface with Metal/MSL. I'm thinking of expanding this into something more than matrix operations.

Tooling

• Enter the dev environment with direnv allow (uses .envrc + flake.nix).
• The flake provides a stable Rust toolchain (rustc, cargo, clippy, rustfmt, rust-src) and just.
• Build shaders with just (default recipe) or just all.
• Clean generated shader artifacts with just clean.
• Metal shader compilation needs full Xcode selected via xcode-select (metal/metallib are not in CLT-only installs).

Compute Backend API

• Matrix<T, M, N>::multiply_with(&other, ComputeOptions) supports CPU, GPU, or Auto.
• Matrix<T, M, N>::multiply(&other) defaults to CPU for all numeric types.
• ComputeOptions::gpu().with_fallback(true) gives explicit GPU-first behavior with CPU fallback.
• Integer scalar types (i8, u8, i16, u16, i32, u32, i64, u64) always work on CPU.
• Current GPU path is implemented for f32 only; requesting GPU for unsupported scalar types returns a clear error unless fallback is enabled.

Benchmark Report

• Run just bench-report to execute Criterion benchmarks and generate a condensed statistical Markdown table.
• Output file: docs/benchmark-comparison.md.
• The report includes mean latency with 95% CI, relative speedup vs baseline, permutation-test p-values, and Cliff's delta effect size.
• Run just bench-scaling-plots to benchmark multiple matrix sizes and generate plots + Markdown summary.
• Outputs: docs/benchmark-scaling.md and docs/plots/matrix-scaling-*.png.