Optimization B, superseding v0.4.6's per-bus skip. Instead of iterating
each active bus's whole POLYPHONY pool per sample (calling render() on
idle slots, which merely early-out), fillBlock now builds compact lists
of the voices that are actually sounding — once per block — and the
per-sample mix loop touches only those.
Correctness for mid-block starts: the sequencer can trigger a note part
way through a block, so the allocators queue the chosen voice into this
block's list (queueSeq/queueSmp), keeping note timing sample-accurate.
A per-voice generation stamp dedups the append so a voice that finishes
and is re-allocated within one block is never rendered twice (which would
double its amplitude). Live/bus-audition voices are snapshotted at block
start (a manual press mid-block sounds at the next block, ≤2 ms later —
imperceptible, and the pre-existing ~30 ms output buffer dwarfs it).
On-device (moto g84, 98% battery, real busy track, ~13 voices, 2.0 ms
budget), avg render callback:
baseline 2.00 ms (xruns +250..+515) → A 1.44 ms (+0) → B ~1.00 ms (+0)
B removes ~51 inactive-slot render() calls per sample; ART's virtual
dispatch makes each ~90 ns, so this is a real ~30% gain over A (not the
~1% first estimated from the sample track — measurement corrected it).
Fast-tapping audition during playback stays ~1.0-1.3 ms, xruns +0,
voices drain cleanly (no stuck notes, no double-render).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
12c2092843