Method for prerendering splats to improve performance about unitygaussiansplatting HOT 2 CLOSED

That’s absolutely not the bottleneck. Calculating “the gaussian” part is just one dot product and one exponent, which is way cheaper than sampling the texture.

The slow parts of the rendering right now are: 1) sorting could be made faster, 2) raw amount of alpha blending when tons of splats overlap, and tiny gaussians being suboptimal for the GPU — this can be addressed by writing a compute shader tiled raaterizer, and 3) there’s no culling of gaussians at all, something that could be implemented.

All in all, there are quite obvious ways how to speed this up, “just” they need to get done. But replacing gaussian falloff with a texture lookup is not one of those ways :)

from unitygaussiansplatting.

I appreciate you for taking the time to comment and enlighten. It's a good reminder to myself that armchair optimization is rarely worth it's weight in words!

from unitygaussiansplatting.