[eeZah7Ux]

When implementing similar algorithms on a load-balancing reverse proxy cache-friendliness and lockless reads and updates in the "load scores" can be way more important than optimizing for CPU cycles.

Another option would be to map coords into very small 2D tiles and than look up "tile -> shard number" with a simple, large, array. A variable number of tiles would land into the same shard. Fast and lockless reads and writes. I wonder if it could be faster.