Restore wasm SIMD autovectorization in ResonatorBank hot loop

[jhartquist]

Summary

Fixes a wasm perf issue diagnosed in #1: target-feature=+simd128 was enabled, but the hot EWMA loop in ResonatorBank::process_sample was only half-vectorizing — LLVM emitted v128 loads/stores but scalarized every f32::mul_add into per-lane fmaf calls (wasm has no vector FMA). The effect was a ~10× slowdown vs what the code looked like it was doing.

Credit to @pengowray for the diagnosis and reproducer.

The fix (three changes in bank.rs)

1. mul_add(a, b, c) helper, cfg-gated to use a*b + c on wasm32+simd128 and f32::mul_add everywhere else. On wasm this gives up fused rounding to keep the vector loop; on native (x86 FMA, aarch64 NEON) the fused instruction is preserved.
2. Batched process_samples that chunks up to the next stabilization boundary, so the modulo check no longer fires inside the per-sample loop.
3. Slice-hoisted process_sample_inner binds ten &mut [f32] locals of known length n once per call. This lets LLVM drop bounds checks, hoist the length-min across all backing Vecs, and trust disjointedness inside the loop.

Perf

Browser bench (examples/web-bench):

Before:

# bins	ns / sample	μs / quantum	% budget
88	730	93	3.50%
264	2097	268	10.07%
440	3449	442	16.56%
880	6888	882	33.06%

After:

# bins	ns / sample	μs / quantum	% budget
88	54	7	0.26%
264	143	18	0.68%
440	238	30	1.14%
880	470	60	2.26%

Speedup: ~13.5× at 88 bins, ~14.7× at 264/880, ~14.5× at 440.

Native bench (cargo bench --bench bank, aarch64): within noise at every bin count (tiny −2% at 88 bins, ±0% elsewhere). Native path intentionally unchanged.

Release

Bumps workspace version 0.1.0 → 0.1.1. Re-unifies resonators-py with the workspace version now that both are at 0.1.1; pyproject.toml uses dynamic = ["version"] so maturin pulls from Cargo.toml.

[Copilot]

Pull request overview

Restores WASM SIMD autovectorization in the ResonatorBank per-sample hot loop by avoiding f32::mul_add on wasm32+simd128, and tightens the sample-processing path to keep stabilization checks out of the inner loop while hoisting slices to help LLVM remove bounds checks.

Changes:

• Add a mul_add(a,b,c) helper that uses a*b + c on wasm32+simd128 (preserving f32::mul_add elsewhere) to avoid wasm SIMD scalarization.
• Rework process_samples to batch up to the next stabilization boundary and factor the hot loop into process_sample_inner with slice-hoisted locals.
• Bump workspace version to 0.1.1 and align resonators-py versioning via dynamic = ["version"] and version.workspace = true.

Reviewed changes

Copilot reviewed 5 out of 6 changed files in this pull request and generated no comments.

Show a summary per file

File	Description
crates/resonators/src/bank.rs	Refactors the hot loop to restore wasm SIMD autovectorization and reduces per-sample overhead by batching stabilization checks.
crates/resonators-py/pyproject.toml	Switches Python package version to dynamic so maturin reads it from Cargo metadata.
crates/resonators-py/Cargo.toml	Unifies Python binding crate version with the workspace version.
Cargo.toml	Bumps workspace package version to 0.1.1.
Cargo.lock	Updates locked package versions to 0.1.1.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[pengowray]

Ah, that's a more elegant solution, getting autovectorization to kick in again, and much better performance too