Conversation

To continue sharing rusty things from last year, I did some experimenting with Vulkan and put the code here: github.com/sjb3d/caldera/. Has a very basic render graph and some probably terrible use of rust async/.await.

github.com

GitHub - sjb3d/caldera: Vulkan and rust experiments, including a spectral path tracer using Vulkan...

Vulkan and rust experiments, including a spectral path tracer using Vulkan ray tracing extensions - GitHub - sjb3d/caldera: Vulkan and rust experiments, including a spectral path tracer using Vulka...

Simon Brown

@sjb3d

Jan 30

The goal was to try out Vulkan ray tracing/mesh shaders/bindless/etc, but I also implemented "Coherent Parallel Hashing" (ggg.udg.edu/publicacions/U) since it is a useful technique for GPU hash tables. Code for that is here: github.com/sjb3d/caldera/

Simon Brown

@sjb3d

Jan 30

I also got a bit carried away and implemented a spectral path tracer using Vulkan ray tracing here: github.com/sjb3d/caldera/. It implements "Practical Hash-based Owen Scrambling" for sampling, and a few different BRDFs. Some renders below:

199

Replying to

Looks great! What spectral representation did you use?

Replying to

Thanks, for reflectance data it is piecewise linear data baked to textures. Where I only have RGB I convert using the method by Brian Smits. Integration importance samples 3 wavelengths and traces them together (no dispersion yet).

Replying to

Are you using hero wavelength sampling for integration? Sampling 4 wavelengths rather than 3 significantly improves colour noise at no extra cost (assuming simd)

Replying to

I am importance sampling 3 wavelengths independently (the code does compare with hero wavelength sampling though). Is running on GPU so 3 wavelengths is a perf win vs 4. :)

Replying to

Ah interesting. Is the perf win during integration or in conversion to/from RGB?

Simon Brown

@sjb3d

Replying to

@anderslanglands

I just mean that there is no unused SIMD lane, going from 3 to 4 will be more instructions and register pressure, so there will be perf cost. I tried to compare with hero wavelength sampling (for F10 illuminant as a test) here: github.com/sjb3d/caldera/

7:32 PM · Jan 30, 2022Twitter Web App

Replying to

Ah so you’re doing CMIS, cool. Would be interesting to see if doing 4 wavelengths would still end up more efficient than 3 even though each path is higher cost - did you try that already?

Replying to

I did trace 4 for a while, but I think at the time colour noise was not my primary source of noise, so I switched to 3 to make some packing of data between shaders smaller. The switch was a while ago, but I could attempt to make the shaders support either again...