This update optimizes an existing feature (Tilemap Collision Indexing).
This optimization will apply to [nearly] everyone using procedural generation tilemaps, infinite tilemaps, multiplayer tilemaps, particularly large tilemaps (especially those dyanmic in nature) or who otherwise intend to index collisions before the tiles are loaded.
Benchmarking:
Using the standard approach, indexing collisions for a 500x450 tileset took 2530ms. By pre-building the collision index array in its entirety, I was able to reduce that to 259ms. However, this implementation reduces it to 8 ms. Larger tilemaps would be exponentially affected.
There are some considerations to make here since there are better implementations, but it would require deprecating some existing code. The advantage would be that the standard approach itself would be optimized, (thus avoiding the sort of 2500+ms costs here), rather than fairly obscured as it is with this implementation.
Refine the loop to only scan through the drawn boundaries that are being tested against (except when skipping cull). This eliminates a huge bottleneck / lookup time from the original loop each frame.
The update should affect tilemaps exponentially based on their size. It reduced my culling time to 1/70th for 1000x1000 tilemaps, and by about 1/20th on 250x250 tilemaps. It also reduced my requestAnimationFrame time from over 16ms per frame to ~1ms for the larger map.
The fix from https://github.com/photonstorm/phaser/pull/3354/ leads to some tiles getting culled when camera is zoomed as the camera.scrollX/camera.scrollY-Properties are needed in unscaled Coordinates (World-Coordinates?) in CullTile.js
Error gets bigger with distance from (0|0)