libDF: enable wasm_simd on rustfft and realfft

[czoli1976]

Summary

Both rustfft and realfft ship comprehensive WASM SIMD paths gated behind a non-default wasm_simd feature flag. Without it, FFTs fall back to scalar on wasm32 even when the build has +simd128 enabled. For libDF's STFT path (which goes through realfft, with rustfft as a transitive dep with default-features = false), both crates' flags must be enabled for the SIMD code to actually flow through to runtime.

Changes

libDF/Cargo.toml:

• rustfft: keep ^6.1.0 constraint (the lockfile-resolved 6.2.0 already has the wasm_simd feature), add features = ["wasm_simd"].
• realfft: bump constraint from ^3.1.0 to ^3.5 (the wasm_simd feature was first added in 3.5.0), add features = ["wasm_simd"].

Cargo.lock: realfft 3.3.0 → 3.5.0, with workspace-wide cascade. The lockfile diff is large (~80 crate version changes) because libDF's resolution hasn't been refreshed in a while; touching realfft triggers re-resolution of dependents. If the noise is unwanted, reverting just the lockfile and leaving downstream cargo update to handle it is safe — the libDF/Cargo.toml change is what matters for behavior.

Impact

On wasm32, the WASM SIMD path produces ~2.4× faster FFTs at DFN3-relevant sizes. Isolated bench (wasmtime / Cranelift, M-class Mac) on FFT round-trips:

FFT size	scalar	with wasm_simd	Δ
n=480 (DFN3 STFT)	3042 ns	1280 ns	-58%
n=512 (pow2)	2601 ns	1146 ns	-56%
n=1024 (pow2)	5906 ns	2482 ns	-58%
n=2048	12051 ns	5324 ns	-56%
n=4096	26149 ns	11270 ns	-57%

End-to-end DFN3 RTF impact is small (within noise on wasmtime, ~1.9% on Node 22 / V8) because FFT is roughly 1% of DFN3's per-hop budget at 48 kHz with hop=480. The gain is much larger for FFT-bound workloads (Whisper, vocoders, music processing, pure DSP code).

The flag is a no-op on non-wasm targets — native ARM (NEON) and x86 (AVX/SSE) paths are unchanged.

Testing

RUSTFLAGS='-C target-feature=+simd128' cargo build --release --target=wasm32-wasip1

succeeds with the new constraint. Standalone DFN3 RTF bench against vendored libDF (30s synthetic test signal, 3 timed passes per build, min-of-3, hop_size=480, sr=48000):

Runtime	Baseline RTF	With wasm_simd	Δ
wasmtime / Cranelift	0.0533	0.0529	-0.7% (within noise)
Node 22 / V8	0.0520	0.0510	-1.9%

Companion PR

Tract has the same gap at the workspace level: sonos/tract#2205. The two PRs are independent — both should land for the full SIMD path to flow on wasm32 builds that use libDF.

🤖 Generated with Claude Code