Performance: the reference budgets (rc.1)
What "within the documented reference budgets" means for tono, and how the numbers are produced:
- Benches live in
crates/tono-core/benches/render.rs(criterion). - The CI workflow can run them by explicit manual dispatch and publishes the results — no per-commit gate, by design: shared CI runners are too noisy for hard thresholds.
- Read reports against the budgets below: a sustained 2× regression on any of them is a release blocker; a few percent is runner weather.
- Reference numbers measured on Apple Silicon (macOS-aarch64, release); CI numbers differ in absolute terms and track in relative ones.
| Bench | What it measures | rc.1 reference |
|---|---|---|
render/blip_osc_env | sine + ADSR micro-render (0.3 s) | ~0.31 ms |
render/tracks_mix_automation_master_reverb | full mixer pass incl. lanes + master chain | ~3.5 ms |
render/seq_piano | the additive piano voice (1 s) | ~8.8 ms |
render/fx_chain_reverb_delay_compress | a wet FX chain | ~1.9 ms |
streaming/streamgraph_fill_512 | one 512-frame block through the streaming renderer | ~0.89 ms |
compile/song_to_program | Song::compile of a representative multi-track song | ~0.21 ms |
scheduling/performance_fill_512 | one 512-frame block through Performance::fill | ~19 µs |
mixing/tracks_stems_8 | render_stems on an 8-track mix with a bus | ~72 ms |
Check the real-time headroom
- A 512-frame block at 48 kHz lasts 10.7 ms of wall time.
Performance::fillrenders it in ~19 µs (~0.2% of block time); the full streaming render of the same block is ~0.89 ms (~8%).- The command-queue and voice budgets are static and documented (4096 scheduled commands, estimates-driven preallocation), so the engine never draws on more than its headroom under legal load.
- The soak (
tests/soak.rs, on-demand) proves it end to end with zero engine-originated underruns. - Offline render cost is dominated by content, not the framework: a piano second is ~8.8 ms (~113× real time), a wet FX chain ~1.9 ms, an 8-track stem decomposition ~72 ms — all comfortably inside authoring-loop latency.
The engine-5 premium, documented and accepted
- The deterministic kernels (
det) are pure f64 polynomial evaluation, so transcendental-heavy micro-renders pay a measured premium over platform libm: the blip bench sits ~2× its pre-engine-5 baseline, the wet FX chain ~1.4×. - This is the price of cross-platform byte-identity, accepted deliberately: it applies only to documents stamped
engine: 5— older revisions keep their libm paths and their old numbers. - The premium is bounded by the table above; no bench crosses its budget because of it. Engine ≤ 4 paths are byte- and speed-identical to before.
Bound memory at compile time
Compile-time estimates bound the runtime:
estimates.frames— the exact render length,round(duration × sample_rate).estimates.memory_bytes—frames × 8(the stereo f32 output).peak_voices— per-track max overlap summed: a voice-pool size that never steals.
The estimates test (tests/estimates.rs) proves the bounds against real renders.