The M2 perf work (non-blocking sync, exact-GPU-completion watermark recycle, clockwise PLS measured throughput-positive at N≈32-128) has no automated guard. A regression would only surface by hand.
Goal: a perf workflow on the self-hosted GPU runner that:
- Builds
--release (release rive libs; timings are only meaningful optimized).
- Runs a fixed scene/N sweep and records frame-time / throughput.
- Compares against a committed baseline with a tolerance band; fails on regression beyond threshold.
Depends on: the self-hosted GPU runner issue. Needs a stable baseline file + a noise-tolerant comparison to avoid flaky failures. Manual-dispatch first, then consider gating release PRs.
The M2 perf work (non-blocking sync, exact-GPU-completion watermark recycle, clockwise PLS measured throughput-positive at N≈32-128) has no automated guard. A regression would only surface by hand.
Goal: a
perfworkflow on the self-hosted GPU runner that:--release(release rive libs; timings are only meaningful optimized).Depends on: the self-hosted GPU runner issue. Needs a stable baseline file + a noise-tolerant comparison to avoid flaky failures. Manual-dispatch first, then consider gating release PRs.