Almost all of the grunt work was done by Codex. I implemented a few simple WebGPU kernels as an example, and also wrote a few example shims for supporting fixed-point. Codex did a huge amount of heavy lifting implementing various gnarly operations and primitives in WGSL.
The performance win is pretty huge. Even for really simple shapes, just being able to run dual contouring in a massively parallel environment gives like a ~3x speed improvement.