CPU sort that enables WebGPU export

[WindyYam]

Not a real issue, but rather some info on work.
I was originally try to proceed on the stochastic branch's work to investigate on sort-free implementation that would potentially work on WebGPU.
Well, at the end there's no a very good solution to the noise. Simply do subpixel sample would greatly affect performance.

We all know GPU radix sort is not supported yet on WebGPU, but right now sort is the only quality solution. So at the end I went back to alpha blend and CPU sort, at the end I came up with a solution:

1. Use octree nodes to manage the scene splats
2. Do culling through octree to filter out visible nodes
3. Coarse sort the visible nodes
4. In each node, sort the splats inside using CPU threads in parellel(since splats between nodes are isolated)
5. Preserve the sort result in each node, so we only update the node sort result when camera position changes a bit threshold

I put it in my fork, hope it can help someone https://github.com/WindyYam/UnityGaussianSplatting
For anyone who want to try: https://windyyam.github.io/unity_splat_test_deploy/

[aras-p]

@WindyYam wow, this looks very interesting! I was hoping that someday GPU sorting would "just work" on WebGPU deployment platform, but before that happens I guess this fallback would be interesting to have. Would you want to package that up as a proper pull request?

[WindyYam]

@WindyYam wow, this looks very interesting! I was hoping that someday GPU sorting would "just work" on WebGPU deployment platform, but before that happens I guess this fallback would be interesting to have. Would you want to package that up as a proper pull request?

Would be a bit tricky, as there seem to have some issue with the branch out in compute shader build on the WebGPU, even if I put the macro switch "if SHADER_API_DX12" or "if SHADER_API_WEBGPU", so I have to remove the compute shader part to make it build and run. I'll try if there are other ways when I got time