Multiple Importance Sampling >> The one-sample model from Veech

[S L]

Hi, 

I am trying to improve the MIS performance on my GPU render. Looking for help if somebody did similar things. While I don't have any RTX hardware, reducing hit-test for MIS is crucial for me.

The typical MIS requires at least 2 samples per bounce, yet 2 hit-test with the BVH tree, 1 for Light sampling, 1 for BSDF sampling. However, that's way too costing for my old GPU.

The paper from Veach has shown some hope. He introduced the one-sample model for MIS.

While, I don't understand mathematically how could it produce the same result as typical MIS.

After some research, it seem nobody is using or discussing this method. That's weird.

Does it mathematically correct to randomly select Light or BSDF sampling with some kind of weight function? What might be the best approach?

--

Best Regards

[Matt Pharr]

The one sample method does work (for example, the 4th edition of pbrt uses it in the volume scattering code, where only one wavelength is sampled but MIS is performed over all wavelength’s probabilities.) If you work through the math it ends up being one of those things where you do some computation 1/n of the time, but when you do it, you weight it by a factor of n, so it’s correct in expectation.

For a GPU ray tracer, the problem is that you still have to do the expensive thing (tracing a closest-hit ray) some fraction of the time. Only doing it sometimes may not save much if other threads in the warp are idle…

Matt

<image.png>

--

Best Regards

--
You received this message because you are subscribed to the Google Groups "pbrt" group.
To unsubscribe from this group and stop receiving emails from it, send an email to pbrt+uns...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/pbrt/CAHES6s7UazMwqTXqV9A8dZpGbgx%3D9j28fnQ6UekL9ORtzNVStA%40mail.gmail.com.