[ymote]

The phone shell on the OnePlus 6T: an idle screen stops presenting at display rate, and the transition hitches are split into their causes, with fixes where the cause was found.

Based on perf/mobile-shell (#26). It bumps the makepad pin to OctoSense-org/makepad#6 (fix/android-idle-repaint @ f1db9eb) and adds the shell-side fixes. docs/perf-mobile-shell.md gets a third column.

Idle: the display now gets ~1 frame/s

Root cause (makepad fork).

• GL advances its completed frame serial only when the completion fence is polled, and the GL render path never polled it.
• Released retained-upload allocations therefore stayed in the ledger forever, and render_view set demo_time_repaint on every render (platform/src/os/linux/opengl.rs:409 at ad8f372). Every live pass repainted at display rate.
• Diagnostic build on the phone: submitted=953 completed=0 records=598 terms=allocations.
• The fix polls the fence, and on Android serves the debt on an idle vsync beat without painting. Details in makepad#6.

Shell: the status clock no longer forks date. The sampler thread forked date twice a second, which alone cost 11 % of a core.

idle screen	perf (#26)	this PR
Home	56.2 presents/s, 43 % CPU	1.0 presents/s, 3 % CPU
hosted-AppCard screen	18.6 presents/s (every frame ≥ 33 ms), 34 % CPU	1.0 presents/s, 3 % CPU
Recents	33.7 presents/s	1.2 presents/s
group window	44.1 presents/s	1.0 presents/s

The idle App screen's 66 ms frames were not the AppCard module. They were the forced repaint re-encoding the window gauss pyramid the module's glass had requested.

Frosted backdrop right side up

The fork's gauss helper still asked for a V flip on Android, although GL render targets have been top-left since d1a0eb1cb. It is fixed in the fork, and the shell's partial DrawGaussScene override is removed.

Verified on the phone: the clock behind the shade and Recents reads upright.

Transitions

What the frames are made of.

• makepad#6 adds draw, wait (eglSwapBuffers) and gc channels to Android's frame monitor.
• The burst sampler also reads SurfaceFlinger's acquire-fence time as GPU-ready latency.
• The frames are GPU-bound. On the idle-fix build, open-app spent 25.5 ms/frame in eglSwapBuffers (worst 124 ms) against 9.8 ms of encode, with GPU-ready p50 at 94 ms.

Fixes.

• A hosted module's glass blurs its own capture. WindowFrame gets a CaptureGauss, as the fork's WM has. Before, the AppCard kit's glass requested the whole WM window's pyramid on every frame that recorded the module.
• The opening app zooms its last capture. It is re-recorded once openness settles, instead of a full-screen module draw on every animation frame.
• The wallpaper shader computes one style per pixel.

Results. Burst stats are over intervals under 250 ms; the full tables are in the doc.

scenario	before (#26): fps · p95 ms	this PR: presents/s · burst fps · burst p95 ms · >33
idle-home	56.2 · 33.1	1.0 · — · — · —
open-app(tap AppCard icon)	27.3 · 66.6	6.0 · 31.5 · 49.9 · 5/22
idle-app-screen	18.6 · 66.9	1.0 · — · — · —
swipe-up-home	40.6 · 66.4	7.5 · 40.1 · 33.3 · 1/26
shade-pull+close	41.1 · 50.0	10.2 · 33.7 · 49.9 · 8/46
page-swipe(right then back)	55.3 · 16.8	7.0 · 50.1 · 16.8 · 1/30
recents(swipe-up-hold)	39.9 · 33.6	15.6 · 36.0 · 33.3 · 1/73
recents-idle	33.7 · 49.9	1.2 · 60.1 · 16.6 · 0/1
recents-to-home	48.6 · 33.5	6.2 · 34.1 · 33.3 · 1/21
island-demo(triple-tap clock)	52.9 · 33.3	5.5 · 55.7 · 33.2 · 0/25
group-open(tap Work)	47.3 · 33.5	26.8 · 52.6 · 33.3 · 2/105
group-idle	44.1 · 49.9	1.0 · — · — · —
group-close(tap scrim)	53.7 · 33.3	6.0 · 37.6 · 83.2 · 2/15

What still blocks 55 fps. A temporary A/B build switched parts of the frame off from adb (not committed): the wallpaper shader, the whole backdrop compositor (gauss scene + pyramid), or the home page.

• None of them moves the shade, the Recents hold or Recents → Home beyond noise. Full-screen GPU content is not the bottleneck.
• With the monitor off, the Recents hold is already 52–56 fps. The 36 fps above was measured with PerfGraph on.
• The remaining misses are on the main thread.
  • The shade's finger phase runs at 33/33/50 ms.
  • The swipe home and Recents → Home alternate 16.7/33 ms.
  • A transition's first frame takes 67–150 ms.
  • Monitor per frame: event 6–13 ms (of which the shell's own channels explain ~2), draw 4–6 ms, wait (eglSwapBuffers) 8–15 ms, which together exceed the 16.7 ms budget.
• Next step: a simpleperf CPU profile of the scene recording, which needs a debuggable build.

Tests

• cargo test: 264 passed.
• cargo test --features mobile-only: 261 passed.
• The macOS desktop app draws with the new pin, checked with an in-app frame capture.
• Fork: makepad-platform 252 passed. makepad-widgets has 5 failures that also fail on ad8f372.

Phone

The run set stayon and screen_off_timeout 1800000, and restored both to 30000 and stayon false afterwards. OctoSense is left on Home with this build.

🤖 Generated with Claude Code

[ymote]

Closing as superseded: every commit in this stack (#22–#28) is already on OctoSense-ROM main, where the phone shell continued. Per ADR 0001 the phone layer moves into the shared crates/shell from that copy when OctoSense-Desktop and OctoSense-ROM are merged into one OctoSense repository, so nothing here is lost. #25's octos-app pin is obsolete: AppCard now comes from OctoSense-System-Apps on octos main. The branches are kept.

🤖 Generated with Claude Code