Painting on the 60x60 map of a project with fifty objects dropped the
editor to 20 to 39 fps. A profile of a brush stroke showed four causes,
none of them the brush itself.
- Each frame of a stroke previews the model. The asset and hierarchy
panels then built every row anew, so every label became a new text
and a new upload to the GPU. One run spent 46 % of the main thread
in those uploads. ListView.setItems now keeps a row whose item draws
the same (sameListItem) and builds only the rows that changed.
- The brush drew the whole scene on every pointer event, on top of the
preview. Now the preview draws what the stroke changed, once a frame.
The rectangle tool asks for one drawing a frame for its box.
- An animated tile made the view draw in the tick and the preview drew
it again in the same frame. animate now asks just before the frame
is drawn and stays out when the view was drawn since the tick began.
- layoutSymbols ran on every drawing and costs up to the cube of the
symbols. SymbolLayoutMemo keeps the spots while no anchor moves.
- A painted cell made the map read every "x,y" key again, and once
more for each of the eight neighbours of every cell. layerCells reads
a layer once per content into an index by place. layerTiles and the
blocking cells take their neighbours from there (neighboursIn).
parseCellKey reads a key without split, with the same results.
The test hook sceneRedraws counts how often the scene view was drawn.
Not done: a painted cell still works out the tiles and blocking cells
of the whole map again, so that part grows with the map.
Tests:
- Brush stroke on a copy of the project (stein2, Chrome): 20 to 39 fps
before, 47.5 after, 55 to 57 without painting. Labels drawn again
during the stroke 2012 before, none after. Scene drawings per frame
1.32 before, 1 after.
- One painted cell, headless: 28 to 16 ms on 60x60, 109 to 49 ms on
the same map doubled to 120x120.
- frame-preview: 2 new, red with the old pointer redraw.
- tilemap: 3 new (key reading, index, refresh), list-ping 1 new,
symbol-layout 1 new. 5 single reversals turned each red.
- e2e 39,40 (melektaus-atx): ALL PASSED, 107 checks (1 new, a fast
stroke: 6 drawings in 6 frames, no label drawn again). With the row
reuse and the pointer redraw reverted it fails: 24 drawings in
3 frames, and the panels draw their labels again.
- Gate (melektaus-atx): CHECK EXIT 0, 260 files, 4038 tests.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Deleting a tile left the cells painted with it in the tilemaps. They
drew nothing, blocked nothing and did not show in the delete dialog,
but still read as a tile for "Tile at" and stopped a fill.
- DeleteAssetCommand empties the cells of a deleted tile on every
layer of every tilemap in the same step. Undo puts the tile and its
cells back together.
- The delete dialog counts them, for example "3 cell(s) on 1
tilemap(s) are painted with it and become empty." (cellsPainted).
- presentTile treats a cell naming a missing tile as empty. "Tile at"
(World.tileAt) and floodFill use it, so cells already in existing
files read and fill as empty. There is no migration.
- The brush, the rectangle and the fill check the tile in hand. A
tile deleted after it was picked no longer paints new invisible
cells, the editor says "Choose a tile to paint with." instead.
- Unchanged on purpose: the picker still reports such a cell as
naming a tile that no longer exists, and erasing removes it.
Tests:
- ./check.sh: 195 files, 2823 tests, CHECK EXIT 0 (+6).
- GM_STEPS=39,40: ALL PASSED, 100 checks (4 new in 40t).
- Full e2e sum is now 1590 (1586 confirmed by CI on round 8, plus 4).
- Counter-checks, all fail as expected: 5 in unit tests (delete
without emptying, undo without cells, fill and "Tile at" without
presentTile, wrong map count) and 3 in the browser (dialog without
count, delete without emptying, brush without the check).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Recurring cleanup round over the code since round 6 (056a622), 105
commits. Behaviour is unchanged. Duplicated code in src went from 29
clones with 229 lines (0.32 %) to 14 clones with 109 lines (0.15 %),
measured with jscpd and the same settings as round 6.
- BehaviorSystem.ts shrinks from 941 to 589 lines. Collecting the
pulses (keys, timers, collisions, animations) moved to
behaviorPulses.ts. The NodeContext the nodes receive moved to
behaviorContext.ts. Imports from outside stay the same, TimeScaled
is still exported from BehaviorSystem.ts.
- PointKindButton and drawPointGrip replace the "Point:" button with
its menu and the point and handle drawing that the curve and the
vector art tab each had on their own.
- lerpVec, cubicAt and splitCubic in src/core/util/vec.ts replace four
copies of the cubic curve and two de Casteljau splits.
- mapNested in Properties.ts is the one walk over lists and structs
that liveObjects and objectRefs did separately. withFieldRenamed
and withoutField replace three copies of renaming a struct field
and two of removing one.
- jsonBody and staleRevision build the JSON body and turn a 409 into
StaleRevisionError for all stores. The auth store turns a 401 into
its error in one place.
- Smaller helpers: structAssetFrom for the two "Make struct type"
commands, shownObject in the inspector, selectedItem for the size
gizmo and the polygon corners, solidTiles in the tilemap collision,
requireVariablesAsset and requireBehaviorVariable.
- Left as is on purpose: the plain forwarding from the collision and
physics systems to Solids, the pointer wiring of the canvases and
the identical last lines of the two make commands.
- Removed unused code: tilemapTileIds and tilemapsUsing were never
called since #128 and #131. EditorState.runtimeProject became unused
with GameRun in #266. The assetId getter of the make commands was
read only by a test.
- 60 values used only in their own file lost their export. The shader
and particle node catalogues stay exported as in round 6.
- Five comments that told history now describe the current state. The
doc comment of BehaviorSystem sits on the class again.
- README lists the struct, clip area and curve assets and mentions the
game link from #266.
Findings:
- Deleting a tile that is painted on a tilemap leaves its cells in
place, and they draw nothing. The delete dialog does not mention the
tilemaps. That was probably the job of the unused tilemapsUsing.
Left as is, because fixing it changes behaviour.
- The solid filter for tile surfaces was untested. A bent filter
stayed green. A new test in tests/core/surface.test.ts now fails
against it.
Tests:
- ./check.sh: 195 files, 2815 tests, exit 0 (one test removed with
tilemapsUsing, one added for tile surfaces).
- Full e2e run: ALL PASSED, 1580 checks.
- Smokes jumprun, tetris, actor: 24, 19 and 21 checks, all passed.
The example project files were restored byte for byte afterwards.
- Counter-checks, 12, all fail as expected: 10 in unit tests (among
them mapNested, splitCubic, cubicAt, staleRevision, the stay pulse,
the late timer and the time scale), 2 in the browser (Point button
always enabled fails 47, selectedItem without a hit fails 29).
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>
A behavior could move sprites around but had no way to look at or
change the tilemap it was standing on. Doors, breakable blocks and
"is there ground ahead?" checks all need that.
- New nodes Tile at, Set tile, World to cell and Cell to world. The
conversion honours position, rotation, scale and mirroring of the
placed map. A small Tile literal node lets a graph compare what it
read against a fixed tile.
- The layer is picked by name from all tilemap layers in the project
through a new field kind. "First layer" takes the map's first one.
Because layer is an input, a text can also be wired in.
- Each placement keeps only its edited cells in liveTilemap.ts,
layered over the asset's cells. A second placement of the same map
stays untouched, and the asset itself is never written.
- A read sees a Set tile from earlier in the same flow. Drawing picks
up the edit at once and neighbouring autotile edges adapt.
- Solids see the edits straight away, so a Move in the same frame
already hits the new cell. Rapier rebuilds its shapes on the next
step.
- After Stop the map is back to what was painted. The editor run shows
the edit in the scene view while it lasts and does not touch the
project.
- An unwired tilemap input means no map, like Get position. Cells
outside the grid are refused and done reports false.
Tests:
- ./check.sh: 155 files, 2271 tests, CHECK EXIT 0 (before 2245).
- tests/core/tile-access.test.ts (14), one physics test for a floor
set at runtime, catalog entries for the five nodes.
- e2e 39,40: all passed, 86 checks (before 81), five new in 40s.
- Counter-checks: collision ignoring edits fails 3 unit tests, the
solids cache missing same-tick edits fails only the "at once" test,
drawing ignoring edits fails the 2 drawing checks, the editor run
not reporting a new tile fails only its check.
- 40s uses a still tile. The animated fixture tile redraws every frame
and hid the last counter-check at first.
- Check total 1351 + 5 = 1356, computed. CI confirmed 1351 at cb32b4b.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
adminandClaude Opus 5.5