For some of the Metal kernels, we use global atomics which causes massive serialization, that affects performance

[himanshunaidu]

Replacing this with atomic per threadgroups and reduction would be a better idea.

Relevant files:

1. WorldPoints.metal
2. MeshPipeline.metal
3. Plane.metal

[himanshunaidu]

We also use global debug counters in some compute kernels that are useful for debugging purposes, but can affect performance in their current state.
Better options would be to do one or both of the following:

1. Add debug compile-time macro
2. Replace per-thread atomics with threadgroup-local counters and then reduction