GPU Ray Tracing Performance Comparisons [2021] *spawn*

RTAO is an interesting choice. Is it because you can get away with infrequent BVH updates and mostly static geometry? I wonder why RTAO and SSAO are getting such different results for the underside of the solar panels.

 

So, is that AO based on surfels (GIBS) as being presented by EA at Siggraph some time ago?

 

Because the secondary AO rays don't need to travel entire world, it's very performance friendly.

 

That makes sense. But once you have your Gbuffer and BVH setup it should be relatively straightforward to add other raytraced options. GI requires some voodoo (surfel caches, probe grids etc) but raytraced shadows are basically RTAO but using longer rays.

(And maybe dealing with transparency. And importance sampling lights. Ok maybe it’s not basically RTAO )

 

Isn't GIBS approach totally BVH independent?

If I'm not mistaken, it's using Gbuffer + trapezoidal grids meshes as non-linear acceleration structure

Spoiler

 

That's the acceleration structure for storing and retrieving surfels not scene geometry. When casting rays into the scene to populate those surfels GIBS uses hardware accelerated RT into a BVH.

 

Hmm, I went through the whole GIBS presentation again and I can't find any references to BVH building etc.

Can you please point me to some slide or study to confirm that GIBS is using BVH?

 

I may have missed it, but where did you get the impression that BF2042's AO is based on surfels (GIBS)?

 

It's an implication so it's not explicitly stated in the slides that they use a BVH. The idea behind GIBS is that we have "surfels" which will store our ray traced results in this surface cache. There's two acceleration structures going on. We have a non-uniform trapezoidal grid where we perform lookups on these surfels. Then we have a BVH to trace our rays against to get accurate results which will be stored into the surfels.

GIBS is mostly useful for giving fast and somewhat plausible GI results with dynamic geometry. GIBS operates on the premise that it's faster to build a uniform grid which is why it has a low frametime cost ...

 

So it's a guess, nothing official yet.

 

Yes, that's a guess. That's why you see a question mark at the end.

Anyway, due to fact BF2042 is missing other RT effects and EA/DICE presented surfel approach at Siggraph I assume they implemented (RT)AO/GI via GIBS

 

I don’t know why the lack of other features would imply RTAO was implemented using GIBS. GIBS is a GI solution, not an AO solution.

To your other question on how we know GIBS traverses a BVH…EA said it does. Also the presentation refers to casting rays at geometry in an acceleration structure.

“The solution combines hardware ray tracing with a discretization of scene geometry to cache and amortize lighting calculations across time and space.”

https://www.ea.com/seed/news/siggraph21-global-illumination-surfels

 

Since consoles aren't getting any RT effects in BF2042, it's likely surfel is not used and this approach is based on BVH. The presentation is fairly recent and doubt they would miss an opportunity to use BF2042 as a poster child for the surfel technique.

 

AO doesn't require material shading and the rays can be bound in their traversal length, e.g. 50cm.

So AO is dramatically cheaper than other raytraced options.

 

Apparently, it does work for characters, judging by provided comparison screenshots.
Character model in BVH has to be perfectly aligned with raster pass (so should have the same animation rate and complexity) for precise shading of creases and folds unless they use screen-space AO pass for characters.

Probably shadowing contribution is weighted differently by distance. For RT AO, they would certainly want to capture larger distance AO from large structures, so different weighting factors have to be used.
Or these objects might be missing in BVH or some sort of translucency mask can be applied for them so that they don't receive AO since these surfaces are semi transparent.
Or maybe that's due to view dependency of screen space AO and more correct diffuse appearance of RT AO.

Nah, nobody would bother with RT for 50 cm. rays, that doesn't make much sense since such distance can be captured with screen-space AO. I guess 2 meters is a bare minimum.

 

50cm or 200, the key is that the rays are very short compared to other uses of ray tracing. Which is part of what makes it cheap and should hopefully be a complete replacement for all of the hacks.

 

The rays can be short, for sure, but they are being shot in random directions in the 360° hemisphere centered around normals, so the longer rays can still be quite expensive due to divergent traversal (in case of traversal on SIMD) and incoherent memory accesses, and long rays are pretty much required for the superior IQ in comparison with screen space AO.
Metro Exodus used AO (not GI) for enclosed areas before the EE update and this AO wasn't cheap due to long rays, the same can be said about RE Village's AO, it's also expensive in 1080p where it's done at full res.

 

This page disagrees with you entirely regarding RE:V:

Resident Evil Village Ray Tracing Quality & Performance Impact of Different Settings | Page 4 of 4 | Hardware Times

So which data are you referring to?