Gaussian Splats on the GPU

A Metal rasterizer for Gaussian splats in Houdini Solaris, 7 to 12 times faster than the multithreaded CPU renderer.

Techniques

A Metal compute tile-fill kernel with unified-memory buffers on Apple Silicon
Opt-in: it appears as a second viewport renderer, Particle Field (Gaussian Splat) (GPU), next to the CPU one
Bit-exact against the CPU reference. It writes colour only for now, and sorting, projection and binning stay on the CPU
Three Hydra delegate bugs fixed along the way: the schema never loaded, the viewport flashed on every camera move, and parallel projection

Splats	CPU fps	GPU fps	Speedup
10k	27.2	193.2	7.1x
100k	2.5	30.5	12.1x
500k	0.5	5.7	12.2x
1M	0.2	2.4	10.9x

Per-pixel fill, GPU against the multithreaded CPU, at 1280 by 720 on Apple Silicon.

Built on houdini-gsplat by Plattipus (MIT). Splat capture: DayLily by Natural Ai, CC BY 4.0.