Improving Data Locality for Efficient In-Core Path Tracing

[jbikker]

The paper "Improving Data Locality for Efficient In-Core Path Tracing" has been accepted for CG Forum and is available for early online view as of today. Apparently I am allowed to distribute digital copies, upon request. So no link here, but feel free to PM.

Abstract:

In this paper, we investigate the efficiency of ray queries on the CPU in the context of path tracing, where ray distributions are mostly random. We show that existing schemes that exploit data locality to improve ray tracing efficiency fail to do so beyond the first diffuse bounce, and analyze the cause for this. We then present an alternative scheme inspired by the work of Pharr et al. in which we improve data locality by using a data-centric breadth-first approach. We show that our scheme improves on state-of-the-art performance for ray distributions in a path tracer.

[Merax]

I don't have enough posts to send a PM, but I'd be interested in reading the paper if you can PM me

[fpsunflower]

Me too!

[Merax]

Thanks for the link, the paper was interesting and I like your idea for grouping the rays.

I'm torn between being impressed that the speed up is as good as you'd get from a new generation of CPUs and disappointed that it's not more considering how big the speed difference is between cache and main memory.

[fpsunflower]

Thanks!

Do you have a follow up planned that goes into the details of the GPU implementation? I would be curious to see a comparison to the Garanzha and Loop scheme.

[franz]

Hey Jacco,

I'd be interested to get a copy of that paper as well, if you don't mind. Thanks!

Cheers,
Franz

[HungryBear]

Any chances to get a link /copy too? Thanks in advance

[Brigadir]

Please, can you send/PM me this paper?

[jbikker]

Do you have a follow up planned that goes into the details of the GPU implementation? I would be curious to see a comparison to the Garanzha and Loop scheme.

I didn't attempt a GPU implementation yet. Perhaps somewhere in the future.

[TheSFReader]

The paper "Improving Data Locality for Efficient In-Core Path Tracing" has been accepted for CG Forum and is available for early online view as of today. Apparently I am allowed to distribute digital copies, upon request. So no link here, but feel free to PM.

Abstract:

In this paper, we investigate the efficiency of ray queries on the CPU in the context of path tracing, where ray distributions are mostly random. We show that existing schemes that exploit data locality to improve ray tracing efficiency fail to do so beyond the first diffuse bounce, and analyze the cause for this. We then present an alternative scheme inspired by the work of Pharr et al. in which we improve data locality by using a data-centric breadth-first approach. We show that our scheme improves on state-of-the-art performance for ray distributions in a path tracer.

Received the link. Thank you. Can we discuss it publicly, or shoud discussions remain "hidden" ?