Optix 6.0 Path Tracing demo

[jbikker]

Hi all,

Since the release of Optix 6.0 two days ago I have been playing a bit with the new RTX support. Here's a demo that should run on Maxwell, Pascal and Turing, which allows a direct comparison of the ray tracing performance of these platforms. Spoiler: Turing fares well.

http://www.cs.uu.nl/docs/vakken/magr/20 ... _v0.1a.zip
http://www.cs.uu.nl/docs/vakken/magr/20 ... _v0.2a.zip
http://www.cs.uu.nl/docs/vakken/magr/20 ... _v0.3a.zip

New in version 0.3a:

- pre-RTX hardware will use a faster Optix Prime path (+~25%).
- TAA + sharpening for more temporal stability.
- A-Trous wavelet filtering for better performance (~25fps on 1060).
- Autonomous demo with spline path.

Instructions: toggle filter with F1, TAA with F2.

This is an 'under construction' demo which may or may not crash on your machine and so on. More updates will follow.

NOTE: this demo requires *very* recent drivers (like, about than a week old).

- Jacco.

https://youtu.be/BDvkoSDTUxc

sponza2a.jpg (126.35 KiB) Viewed 405 times

[Tristan]

Thanks for sharing.

[jbikker]

Updated first post with new demo.

[olliej]

Would you mind if I used your version of the sponza model to test my own renderer?

The various open licensed versions I've found lack the detailing and have some questionable texture coords.

Also it's nice to be able to be able to compare exact scene rendering to an actual good renderer

[jbikker]

Please go ahead. But I think this is straight from McGuire's page?

[olliej]

It is indeed -- I wonder if I dismissed it due to my RT not being able to handle it earlier.

Now to work out why my photon mapping is weird and blown out.

[olliej]

@jbikker: ok, possibly dumb question - how do you handle convert the emission coefficients from the wavefront into optix? I'm doing the dumbest thing possible with and doing ke * light colour ?

[jbikker]

Well I think this deserves its own topic. I don't handle this at all; the demo uses unidirectional path tracing only, no photon mapping.
But what I got for photon mapping is: lightcolor * dot( D, N ) / lightPdf, where D is a cosine-weighted random direction with pdf = (dot( D, N ) / PI) / lightArea and N is the normal at the light source. Note that the dot(D,N) cancels out in this case, but I prefer to leave it in to make the application of the pdf clear.

[olliej]

Okie dokie

I'm still trying to work out what it should be doing (this is my first real attempt to implement PM). I'm getting a huge blow out on the direct photons, and a rapid drop in the propagation following the first bounce. (I'd search the board but photon mapping is apparently too common to use as a search term )