Skip to content

performance.md

Latest commit

 

History

History
39 lines (30 loc) · 1.57 KB

performance.md

File metadata and controls

39 lines (30 loc) · 1.57 KB
Raw

Meld performance

Performance is part of the API contract. Planning is separated from painting so expensive work is cacheable and measurable.

Measurement

melos bootstrap
melos run benchmark

The benchmark executes repeated plans for a representative line-icon pair and reports timing and cache hit/miss counts. Keep the machine, Dart SDK, and build mode fixed when comparing results.

Runtime practices

  • Share one MeldEngine for repeated icon pairs.
  • Call prewarm before the first interactive frame when pairs are known.
  • Use adaptive sampling for general UI and raise maxPointCount only for detailed artwork.
  • Keep MeldDiagnosticsOverlay out of release layouts.
  • Use MeldPaintStyle.original to retain a filled glyph's source intent instead of inflating it into stroke paths.
  • MeldIcon controllers share one frame ticker per Flutter isolate and the painter reuses its Path and Paint instances. Avoid rebuilding large ancestor subtrees from a per-frame listener when targeting 120 Hz displays.

Plan and sample caches are separate bounded LRU caches with both entry and approximate byte limits. MeldCacheStats.bytes exposes the retained budget so production diagnostics can detect pressure before it becomes a GC problem.

Regression policy

The default target is a desktop plan P95 below 2 ms for typical icon pairs. If the benchmark exceeds the target, inspect sample count, residual, and subpath count in diagnostics before changing quality settings. Record visual or performance changes in GOALS.md with a fixture and acceptance criterion.