The performance panel showed Physics and Editor as the large parts of
a frame in a scene of 222 entities. Nearly all of it came from a
tilemap that only stands there.
- Whoever writes cells in place now says so with cellsChanged. Only
then does cellsVersion compare the content. The comparison ran about
60 times per frame and was the largest single item at 14 percent.
- The four writers call it: the painter during and at the end of a
stroke, PaintTilesCommand on execute and on undo. A new writer that
forgets it stays invisible.
- The game view is drawn once per frame in play. With an animated tile
in view it was drawn twice, once by the run and once by animate().
animate() now gets the redraw count from the start of the tick.
- TilemapField builds the geometry of a mesh only when its source
changes: the cells, the tile size, the origin or the image. For that
groupByTile and layerRows return the same objects for the same list.
- PhysicsSystem skips a tilemap whose source and switch are unchanged.
The switch is kept because a map that starts switched off would
otherwise have dead colliders once it is switched on.
- World.syncToModel looks entities and assets up in a map per pass
instead of a linear search per object.
Measured on stein2 with a copy of the reference project, ms per frame:
Physics 2.64 to 0.97, Drawing 6.89 to 2.56, Editor 1.95 to 1.31.
Left open: syncToModel still mirrors every entity every frame. That
is its own issue.
Tests:
- Gate: 198 files, 2887 tests (2883 + 4), exit 0
- e2e 1-17,28,33,39-43,59: 810 checks passed, 40 has +3
- Smokes: 24 / 19 / 21, exit 0
- Three unit tests wrote cells directly and now call cellsChanged
- Eleven counter-probes fail. Two did not at first: the stroke needed
a check of its own, the physics test has to start with the map
switched off. That last one was repeated on stein2 only.
- e2e total 1613 is computed, not measured in a full run
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A project with three tilemap layers, two of them with collision on
water and hill tiles, ran at 33 to 52 fps. A CPU profile showed that
almost all of the time went into the collision geometry of the map.
It was rebuilt from scratch every frame although nothing had changed.
The animated water makes the views redraw every frame, so the cost
was never skipped.
- New module tilemapMemo keeps what is derived from the cells: the
tiles of a layer, the blocking cells, the rectangles and polygons.
It is reused as long as the content stays the same.
- A brush stroke paints into the existing cells object. The memo
therefore compares the content cell by cell, not the identity.
The comparison walks two arrays in insertion order.
- Clipped outlines are kept per outline list. The model always
replaces that list instead of changing it.
- PhysicsSystem no longer rebuilds the shapes of a tilemap each tick
while its cells, placement and settings are unchanged. Before, it
also built a JSON key from all world shapes every tick.
- The overlaid layers after Set tile in a run stay stable. Otherwise
the memo would miss on every call after the first edit.
The #184 work only helped Move and raycasts, not this path.
Measured on stein2 (i5-7500T, HD 630), script time per frame:
- Scene, nothing selected: 17.7 ms, 52 fps -> 5.8 ms, 60 fps
- Tilemap selected: 26.2 ms, 38 fps -> 8.6 ms, 60 fps
- Play: 29.9 ms, 33 fps -> 8.7 ms, 60 fps
About 3 ms per frame remain. The content comparison costs roughly
60 us per 700-cell layer and runs about 40 times a frame. The mesh
geometries in TilemapField are still rebuilt every frame.
Tests:
- ./check.sh: 162 files, 2377 tests, CHECK EXIT 0 (before 2370)
- e2e 1-5 and 39-43: ALL PASSED, 289 checks
- 7 new unit tests in tests/core/tilemap-memo.test.ts
- 5 counter checks (content ignored, no memo, outline by kind only,
physics always rebuilds, layer settings ignored) each fail their
test. All originals are back, checked with cmp.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
adminandClaude Opus 5.5