Regarding comparing GLSL Pathtracer with a CUDA Pathtracer

[mobeen]

Hi all,
Just populating the forum with my first post here. This is a current WIP that I work in my spare times. This is an extension to the wonderful work of Evan Wallace (http://madebyevan.com/webgl-path-tracing/).

First a basic snapshot.


Now the videos.
http://www.youtube.com/embed/PMj4kEaZhRY
http://www.youtube.com/embed/1FuZhTeI-9o
http://www.youtube.com/embed/bdteeIknjPk
http://www.youtube.com/embed/aPDXRrBRwRc

I want to know the fastest CUDA pathtracer to compare against. Can anyone suggest some CUDA pathtracer for comparison?

[sirpalee]

I don't think that any comparison is worth it, with these very simple cases. On the other hand, with complex scenes CUDA should be faster since you have more control over the kernels, register count and everything else

[mobeen]

Hi sirpalee,
You are right this is just basic stuff here what good is a path tracer without hi-res meshes. The only reason I ask this question was to know what the current fastest CUDA pathtracer is and what is the expected output from it. I have seen some amazing work being done by Jacco Biker in Brigade 2. It seems to me that it is the fastest available CUDA pathtracer at the moment. Are there any more options?

[jbikker]

I did something similar for the Kajiya demo, in CUDA. It's not using an acceleration structure (apart from a few bounding boxes, but since they are nested, that doesn't qualify as a bounding volume hierarchy ). Video is here:

Actually that's still the old scene, without elipsoids. O well.

EDIT: Youtube tags now work.

[mobeen]

WOW JBikker, This is amazing stuff. Could u tell me a bit more about the Kajiya demo setup? Is this demo available for download I want to give it a try. Also would it be possible for u to share the scene setups like the camera settings, sphere radii,box dimensions their positions, eta etc?

[jbikker]

I am not sure if I posted the sources before, but you can get them here:

http://igad.nhtv.nl/~bikker/files/kajiya.zip

All coordinates are in there. That version also does the elipsoids properly, and is very similar to Kajiya's original scene.
You may also want to look around for a demo that I did for a demo compo at our university. It's called 'Simplex Paternitas' and uses the same tracer code. It is a bit more flexible on the C++ end though. Links:

http://igad.nhtv.nl/~bikker/files/Simpl ... s%20v3.rar
http://igad.nhtv.nl/~bikker/files/Simpl ... %20src.rar

Youtube:

[mobeen]

Hi jbikker,
I ran your Kajiya demo and it is very nicely done. Great work.

Thanks,
Mobeen

[beason]

You reminded me... I found a subtle bug in the webgl path tracer: it doesn't scale reflected rays' colors by n-dot-l.
I just emailed him about it, but in the mean time you can see the fix in action here (using cosine weighted distribution which lets you drop n-dot-l factor):

http://www.rhythm.com/~beason/webgl/Web ... acing.html

[mobeen]

Hi Beason,
Thanks for the reply. Could u tell me what changes you did and where? In the original code, he simply changes the origin and direction of the ray, the surfaceColor is never altered.

[beason]

mobeen,

Compare webgl-path-tracing.js with the original file, webgl-path-tracing.js.orig, here:

http://www.rhythm.com/~beason/webgl/Web ... ing_files/