Stochastic alpha blend (and discussion on CPU sort that enables WebGPU support)

[WindyYam]

I see your implementation basically reject the most compute shader, move them to per-vertex. I guess it is possible to complete reject compute shader by doing stochastic alpha blending, then every thing running as pure opaque dots, thus Z-test would natually work, no sort needed. it might be possible to even make it work on WebGL

[wuyize25]

Thanks for the suggestion! ​My main concern with stochastic alpha blending is the visual artifacting (noise/shimmering) it can produce, which lowers the rendering quality and requires a filter like TAA. Due to these, it's not on my immediate roadmap. Additionally, if you really want to make it work under WebGL, you could also consider performing a radix sort on the CPU to replace the existing GPU sorting pass. It should be fast enough.

[WindyYam]

Thanks for the suggestion! ​My main concern with stochastic alpha blending is the visual artifacting (noise/shimmering) it can produce, which lowers the rendering quality and requires a filter like TAA. Due to these, it's not on my immediate roadmap. Additionally, if you really want to make it work under WebGL, you could also consider performing a radix sort on the CPU to replace the existing GPU sorting pass. It should be fast enough.

Yes, I tried, I can't resolve the noise quality issue on stochastic alpha

FYI I've done a version that do sort in CPU, which allows export in WebGPU (not WebGL unfortunately as we still need some structurebuffer support) https://github.com/WindyYam/UnityGaussianSplatting

Key insight is using octree to manage scene splats since they are all static. Then we can do camera culling to filter out visible octree nodes, do coarse sort on visible octree nodes, and on each node do parellel sorting, also add some trick to preserve sort result when camera movement to that node is not significant.