A scene with 200 spawned gold pickups, each with a Pickup behavior of
97 nodes, dropped the game to about 14 fps. The graphs themselves ran
only a dozen nodes per pickup and frame. The cost was the lookup: for
every data input the interpreter searched all edges and all nodes of
the graph linearly and computed the port shape of the node again.
- New module core/graph/graphIndex.ts builds tables once per graph:
nodes by id, edges by node and port, node types. Port shapes are
cached as well.
- BehaviorSystem shares one cache between all its instances and
clears it in syncFromModel. A graph changed during a run applies
from the next frame on, as before.
- An interpreter without a cache, as in most tests, searches as
before.
Measured on the scene Performance test (headless per frame, browser
fps on a throwaway copy of the project):
- Headless: about 21 ms down to about 3.6 ms per frame.
- Browser, about 505 entities: 13.6 fps up to 23 fps.
- Browser, about 320 entities: 32.3 fps up to 48 fps.
- After the fix behaviors take about 15 % of the browser profile.
The rest is the editor mirroring the game into the model and
rebuilding its panels, which is a separate change.
Tests:
- ./check.sh: 192 files, 2786 tests, CHECK EXIT 0.
- Full e2e run (local): ALL PASSED, 1551 checks.
- tests/core/graph-index.test.ts: 2 tests. The tables return the
same as the search, in the same edge order. A graph changed during
a run applies from the next frame.
- Counter-check: without clearing the cache in syncFromModel the
second test fails with ['before'].
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
