Files
pipelines/app/notes/perf-round5.md
T
alan 8361322fe2 Keep floating window moves pixel-stable
Round fitted window frames to whole logical pixels and accumulate drag deltas from the gesture start. This prevents floating-point size noise from rebuilding window content while preserving fractional pointer movement and resizing behavior.
2026-11-11 03:28:34 +00:00

215 lines
9.9 KiB
Markdown

# Performance round 5: moving a window does not rebuild its content
Round 5's candidate: every pointer event of a window drag rebuilt every
window's whole content. `WindowScope` carried the body `Rect` and compared the
whole `WindowEntry`, the navigator's `ThemeData` was rebuilt with `copyWith` on
every build, and the root `MaterialPage` was a new object on every build.
## Change
`lib/shell/windows/`:
- `window_scope.dart`: `WindowScope` keeps only `manager` and `entry`, and
notifies only when what content can read changes (id, kind, contentId,
serverKey, `data`). A new `WindowBodyScope` (an `InheritedWidget` with the
body `Rect`) carries the position and size that used to be in `WindowScope`.
- `window_navigator.dart`: `WindowNavigator.theme(context, child:)` became
`WindowNavigator.themeData(theme)`, so the layer can keep the
page-transition `ThemeData`. `WindowPageBox` reads the body from
`WindowBodyScope`, and its page child is a `RepaintBoundary`.
- `window_layer.dart`: `_WindowLayerState` keeps the page-transition theme in a
field and rebuilds it only when `Theme.of(context)` is a different object; it
keeps the root `MaterialPage` in a field (same key) and rebuilds it only when
the kind changes. `WindowFrame` is wrapped in a `RepaintBoundary`.
No behaviour change: frames, snapping, resize limits, popups, pushed pages,
minimise/maximise, focus and the shell's APIs are as before.
`tool/capture/window_probe.dart`: `ProbeContentState.lastSize` records the
`MediaQuery` size the content built at, so tests can see a resize reach it
(no visual change).
## Deterministic build counts
`test/shell/windows/window_rebuild_test.dart`:
- dragging a window's title over 8 pointer moves: the dragged window's content
and the other window's content keep their `builds` count (0 rebuilds);
- resizing a window over 5 pointer moves: the other window keeps its count,
the resized window's content rebuilds and sees the new `MediaQuery` size
(360x380 -> 280x330 body);
- `manager.setContent` retargets a content whose element is keyed stably: the
same state rebuilds with the new contentId.
Against the old implementation (lib changes stashed, test run unchanged): 2
failed — `dragged window rebuilt at move 0` (builds 2 -> 3) and `other window
rebuilt at move 0` (2 -> 3). The new code passes all 3.
## Browser bench
Built twice: the old code (lib/shell/windows changes stashed) into
`build/probe-before`, and this branch into `build/probe-app`:
```
/Users/alan/development/flutter/bin/flutter build web \
--target tool/search_selection_probe_entry.dart --output build/probe-before
/Users/alan/development/flutter/bin/flutter build web \
--target tool/search_selection_probe_entry.dart --output build/probe-app
```
Run, from `app/`, against each build in turn:
```
node tool/perf/interaction_bench.mjs --build build/probe-before --out /tmp/bench \
--gpu --only 'desktop: drag a probe window' --repeat 3
node tool/perf/interaction_bench.mjs --build build/probe-app --out /tmp/bench \
--gpu --only 'desktop: drag a probe window' --repeat 3
```
Three before/after rounds (the machine was shared with other pipeline test
runs; load average 9-14 during these, so frame counts vary. `buildMs/frame =
buildSum / frames`):
| round | code | frames | buildSum | ms/frame | buildMax | longest | over16 | rasterSum |
| --- | --- | --- | --- | --- | --- | --- | --- | --- |
| 1 | before | 50 | 952.2 | 19.0 | 63.5 | 73.9 | 31 | 107.4 |
| 1 | after | 49 | 824.8 | 16.8 | 59.1 | 74.0 | 27 | 107.9 |
| 2 | before | 49 | 615.6 | 12.6 | 25.5 | 28.4 | 15 | 62.1 |
| 2 | after | 51 | 583.0 | 11.4 | 37.4 | 39.9 | 13 | 87.2 |
| 3 | before | 51 | 527.9 | 10.4 | 16.2 | 20.3 | 9 | 68.3 |
| 3 | after | 47 | 606.8 | 12.9 | 24.1 | 26.9 | 18 | 83.5 |
The browser numbers are inconclusive on this loaded machine: only round 1
shows less build work, and the spread between rounds (10.4-19.0 ms/frame on the
same code) is larger than the change. The reliable result is the build count:
the drag went from one content build per pointer event per window (120 builds
over a 40-move drag with three windows) to none. Window chrome and framework
layout still run on every pointer event; that is the remaining build cost in
this scenario, not content.
## Checks
- `tool/memwatch/full.sh 'test/shell/windows test/features/chat/chat_floating_pane_test.dart test/features/terminal'`:
119 passed, 0 failed.
- `tool/analyze.sh`: clean.
## Affected UI states (for ui-review)
`/runs?windows=2` at 1280x800 light: the windows after a title drag
(`drag=250,132>300,150`), after a corner resize
(`drag=483,523>120,80`), and a page pushed inside a window (click `page`) then
dragged: the content must follow the window, stay inside it, and show no stale
or clipped frame.
## Whole-pixel frames
Follow-up: after the change above, a drag still built and laid out every
window's content on every pointer event (14-17 ms of build per frame in the
browser bench). A `Rect` stores left, top, right and bottom, so shifting a
fractional frame changes its width in the last bits (`360.00000000000006`), and
`WindowPageBox` puts that size in a `MediaQuery`: every `MediaQuery` dependent
rebuilt and laid out again on every move. The layer also regenerated its page
theme on every frame, because Material's animated theme hands over a new but
equal `ThemeData` object and `_pageThemeFor` compared with `identical`, so every
theme dependent rebuilt too.
### Change
`lib/shell/windows/`:
- `window_entry.dart`: `fitWindowFrame` rounds left, top, width and height to
whole logical pixels after clamping, keeping the frame inside the bounds and
at the minimum size. A fractional minimum rounds up, so a whole-pixel frame is
never smaller than the minimum the kind asked for (the bounds still win when
they cannot hold it). Whole numbers add and subtract exactly, so a moved
window keeps exactly its size.
- `window_layer.dart`: `_WindowLayerState` remembers the frame a move or resize
started from and the pointer delta summed since, and sets the frame from
`start + total` on every event, so rounding one fractional step after another
cannot lose part of the pointer's movement or drift. The drag state is
cleared in `_moveEnded`, in the corner's `onPanEnd` (`_resizeEnded`, which
still calls `manager.commit`) and on pointer down.
- `window_layer.dart`: `_pageThemeFor` reuses the cached page theme when the app
theme is `==` to the last one, not only identical, so an equal theme from an
animation does not notify every content widget.
Snapping, `manager.commit`, `windowSnapRect` (still fractional where the bounds
are) and a maximised frame are unchanged.
### Tests
`test/shell/windows/window_rebuild_test.dart`:
- dragging a title by eight `Offset(10.3, 4.7)` moves with
`debugOnRebuildDirtyWidget` (reset in a `finally`): no `TextButton`,
`ListView` or `Scrollable` element inside either window's content rebuilds,
the `WindowFrame` size stays exactly `360x420` after every move, and the
window ends within 1 px of `start + 8 x (10.3, 4.7)` (`100,80` -> `182,118`);
- resizing by five `Offset(-10.3, -6.7)` moves: every frame is a whole-pixel
size, the other window does not rebuild, and the resized window's content
rebuilds with the new `MediaQuery` size (within 1 px of
`360x420 - 5 x (10.3, 6.7)`).
`test/shell/windows/window_store_test.dart` also pins the fit: a fractional
minimum (`320.1x240.4`) yields a whole-pixel `321x241` frame, and bounds smaller
than that minimum (`319.6x239.2`) yield `319x239` instead of exceeding the area.
Against the old lib (lib/shell/windows changes stashed, tests unchanged): 2
failed - the drag rebuilt `ListView`, `Scrollable` and `TextButton` at move 1,
and the first resize move gave a fractional `349.70000000000005` width. The new
code passes all 5.
### Browser bench
Built after (`build/probe-app`) and before (`build/probe-before`, lib change
stashed):
```
/Users/alan/development/flutter/bin/flutter build web \
--target tool/search_selection_probe_entry.dart --output build/probe-app
/Users/alan/development/flutter/bin/flutter build web \
--target tool/search_selection_probe_entry.dart --output build/probe-before
```
Then, from `app/`:
```
node tool/perf/interaction_bench.mjs --build build/probe-before --out /tmp/bench \
--gpu --repeat 3 --only 'desktop: drag a probe window'
node tool/perf/interaction_bench.mjs --build build/probe-app --out /tmp/bench \
--gpu --repeat 3 --only 'desktop: drag a probe window'
```
The machine was shared with other pipeline test runs (load average 11.8-75.7
during the rounds; `ms/frame = buildSum / frames`):
| round | load | code | frames | buildSum | ms/frame | buildMax | longest | over16 |
| --- | --- | --- | --- | --- | --- | --- | --- | --- |
| 1 | 22.1 | before | 54 | 995.1 | 18.4 | 77.8 | 80.4 | 29 |
| 1 | 22.1 | after | 42 | 419.2 | 10.0 | 25.8 | 30.3 | 11 |
| 2 | 11.8 | before | 51 | 489.8 | 9.6 | 20.2 | 22.4 | 3 |
| 2 | 11.8 | after | 42 | 549.8 | 13.1 | 44.4 | 59.7 | 14 |
| 3 | 75.7 | before | 48 | 4590.1 | 95.6 | 320.0 | 437.1 | 44 |
| 3 | 75.7 | after | 42 | 1741.3 | 41.5 | 136.3 | 148.9 | 36 |
The browser numbers stay indicative on this loaded machine: round 2 has the
before build ahead, while rounds 1 and 3 (including the heavily loaded one)
show the after build at about half the build time per frame. The deterministic
result is the widget test: a fractional drag no longer rebuilds any content,
and the window size is bit-identical from frame to frame.
### Checks
- `tool/memwatch/full.sh 'test/shell/windows test/features/chat/chat_floating_pane_test.dart test/features/terminal'`:
121 passed, 0 failed.
- `tool/analyze.sh`: clean.
### Affected UI states (for ui-review)
`/runs?windows=2` at 1280x800 light: drag the top window by its title
(`drag=250,132>301,153`), then resize it by its corner: the title bar, body and
edges must stay aligned with no gap or overlap of a pixel, the content must not
reflow during the drag, and the resize corner must stay on the window's lower
right corner.