fillBlock rendered EVERY pooled voice every sample: 32+32 sequencer +
32+32 bus-audition + 16 live/audition voices, render() called ~144×
per frame × 96 frames per callback — even at idle. The bus-audition
pool (64 voices) is only used by the Toolbox MIDI piano yet ran every
sample during normal playback. This fixed cost, not active-voice count,
dominated the render budget.
Now the pools are scanned once per block into per-bus active flags, and
the per-sample mix loop skips whole idle buses and the entire bus-
audition pool when nothing is being auditioned. The sequencer sets the
flag the instant it triggers a note mid-block, so newly-started notes
still sound sample-accurately within the same block.
On-device profiling (moto g84, 98% battery, real busy track, ~11 sample
voices, 96-frame/2.0ms callback):
before: avg 2.00 ms, xruns +250..+515 per ~500 callbacks (audible drops)
after: avg 1.44 ms, xruns +0 (steady)
Fast-tapping the audition keyboard during playback: avg 1.1-1.4 ms,
xruns +0, voices peaked at 16 then drained cleanly — no stuck notes.
With this, under-load render time equals idle, so voice rendering is no
longer the bottleneck; the remaining ~1.4 ms is fixed JNI + per-channel
FX overhead. A block-based per-voice refactor was evaluated and dropped
as unnecessary (it would not reduce that fixed cost).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
d54fd27c9e