mel: precompute FFT twiddle factors and reuse scratch - #65
haydonryan wants to merge 2 commits into
Conversation
Replace the per-frame iterative radix-2 FFT's per-butterfly twiddle recomputation with a cached FftPlan (bit-reversal + per-stage twiddle table built once per FFT length, thread_local scratch, no per-frame allocation). The transform stays in double and the table is generated bit-identically to the old iterative complex-multiply, so the output is bit-identical. Measured: ~8.5% faster mel frontend on the 110m (bit-identical). Assisted-by: AI:DeepSeek-V4-Flash
localai-org-maint-bot
left a comment
localai-org-maint-bot
left a comment
There was a problem hiding this comment.
rfft can be called with different FFT lengths on the same thread, but static thread_local FftPlan plan(n) permanently caches only the first length. The comment says “one plan per distinct n”, while the implementation has one plan total. A later call with another power-of-two length passes differently sized input to apply; it reads n_ values and silently produces the wrong transform (or reads past the shorter input). Please key the cache by n or rebuild when plan.n() != n, and add a regression test that calls rfft with two lengths in both orders on the same thread.
The optimization itself looks useful; this lifetime/cache mismatch is the blocker.
50cae1a to
5366723
Compare
The FFT plan cache was a single thread_local FftPlan built for the FIRST length seen, so a later rfft() call with a different power-of-two length reused the wrong plan and transformed the wrong number of samples. Rebuild the cached plan whenever n changes (zero steady-state cost: one int compare). Add tests/test_fft.cpp which calls rfft with several lengths in both orders on one thread to guard this. Assisted-by: AI:DeepSeek-V4-Flash
5366723 to
a3d1e4c
Compare
localai-org-maint-bot
left a comment
localai-org-maint-bot
left a comment
There was a problem hiding this comment.
The cache now rebuilds when the FFT length changes, and the same-thread regression covers multiple lengths in both directions. The focused and full project gates are green. This resolves my blocker. @mudler, this is good to merge.
2 participants
Replace the per-frame iterative radix-2 FFT's per-butterfly twiddle recomputation with a cached FftPlan (bit-reversal + per-stage twiddle table built once per FFT length, thread_local scratch, no per-frame allocation). The transform stays in double and the table is generated bit-identically to the old iterative complex-multiply, so the output is bit-identical.
Measured: ~8.5% faster mel frontend on the 110m (bit-identical).
Assisted-by: AI:DeepSeek-V4-Flash