Files
pipelines/app/notes/perf-glass-chrome.md
T
alan b90c7e2d1d Cheaper frosted-glass blurs: sigma 4 and one backdrop read per bar
Blur sigma 16 -> 4 keeps CanvasKit on its direct blur path; under the 84%
fill the look is unchanged (max 13/255 inside the pieces). SegmentedBar's
pieces share a BackdropGroup so Impeller reads the content behind them once.
Adds a scroll frame-time bench, a backdrop census and filler rows to the
full-app probe. CPU CanvasKit, 560x900: run timeline p50 31.8 -> 22.1 ms,
runs list 25.3 -> 19.9 ms.
2026-11-02 00:58:23 +00:00

2.2 KiB

Frosted-glass chrome: cheaper blurs

Every GlassPill (top bar pieces, status strip, page and main tab bars, chat composer, expanded run title) is one BackdropFilter. While content scrolls behind the bars, each re-blurs it every frame.

Change

  • GlassPill.blurSigma 16 -> 4. Under the 84% fill the larger radius adds nothing visible, but above about 4 Skia (CanvasKit) downscales, blurs and upscales each piece every frame.
  • GlassPill uses BackdropFilter.grouped; SegmentedBar puts its row of pieces in one BackdropGroup so Impeller reads the content behind them once instead of once per piece. The expanded title (drawn in the overlay, over the content below the bar) gets its own key. The web renderer ignores backdrop keys.

Measurement

tool/perf/glass_scroll_bench.mjs over the release probe build (?runs=400, ?steps=400 filler rows), 560x900, Material light. Interleaved rounds against a main build because other agents loaded the machine. Medians of 6 rounds of 4 s each, CPU CanvasKit (no WebGL):

scenario blurs on screen frame p50 frame p90 missed vsyncs
Runs list, before 3 25.3 ms 50.0 ms 129
Runs list, after 3 (2 reads) 19.9 ms 38.9 ms 102
Run timeline, before 5 31.8 ms 58.6 ms 159
Run timeline, after 5 (3 reads) 22.1 ms 33.1 ms 91

Without any blur the same frames take about 10-12 ms, so blur was most of the frame. With the M2 Pro GPU (--gpu) frames take about 2-3 ms before and after: the GPU path is not blur-bound on this machine. Android (Impeller) was not measured on a device.

Tried and dropped:

  • Sigma 6 or 8: no measurable gain over 16 on CPU CanvasKit.
  • One backdrop per bar, clipped to the union of its pieces (a custom render object): 5 -> 3 backdrop layers on the run timeline, but no measurable gain over per-piece blurs at sigma 4, so not worth the code.
  • RepaintBoundary around the bars: scrolling repaints only the viewport (a repaint boundary already), so the bars are not repainted.

Pixels (tool/perf/compare_shots.mjs, before vs after at a fixed scroll, both platforms and themes): no pixel differs by more than 13/255, and 0-0.6% differ by more than 2/255. The differences are only inside the pieces.