Radiosity on ATI hardware?

[davepermen]

better talk about global illumination, and the different components of it. mainly in this case, indirect lighting (or diffuse indirect lighting, or what ever). radiosity is too well known and interpreted as the algorithm. and this algorithm simply isn't useful for a realtime situation, it scales too bad.

but different parts of the gi equation can be handled with different algorithms, and on openrt.de they show how to handle about all major parts of it. there are quite some algorithms for parts of the equation. from shadows, to different ways to handle reflections (more or less acurate), refractions, even caustics by today. but the diffuse indirect illumination still hasn't found a good (dynamic) way to do realtime on rastericers. only some partial precomputed ways are available till now.

and i have a different opinion than chrisATI about "hacks". i don't consider a solution to be a hack, if it is scalable to get as close to the correct result as you want, without showing up any bias. but yes, there isn't any realtime engine yet showing such a rendering solution that can scale up to a "100% perfect" image if you set quality to maximum. in raytracing solutions, with photonmapping or other approaches, it is (nearly by default) doable. thus i consider such solutions not as a hack.

anything that needs precomputations is, in my eyes, a hack. espencially with the future being dynamic, physically correct scenes, such hacks get very visible.

can't wait for the GPGPU support on the new ati cards to speed up my raytracing stuff

 

[Ho_Ho]

can't wait for the GPGPU support on the new ati cards to speed up my raytracing stuff

I wait for someone to come up with a good algorithm for building KD-trees with GPU's
So far most solutions are being run on single CPU and therefore dynamic scenes are not exactly easy to do with ray tracing. Some newer studies have shown that BVH's are better than KD-trees for doing GPU based ray tracing but with R520 excellent branching abilities KD-trees should be quite efficient.

 

[Andrew Lauritzen]

I wait for someone to come up with a good algorithm for building KD-trees with GPU's

It has been done... in fact to my knowledge the current fastest raytracer runs on the GPU using Sh (http://libsh.org).

See http://portal.acm.org/citation.cfm?id=1071866.1071869.

 

[nelg]

ShIT. Someone has to say it. Andy are you a Sh IT specialist?

 

[squarewithin]

It has been done... in fact to my knowledge the current fastest raytracer runs on the GPU using Sh (http://libsh.org).

See http://portal.acm.org/citation.cfm?id=1071866.1071869.

Ummm... that paper says that they trace over an order of magnitude less rays per second than an optimized CPU raytracer. How is the GPU speeding anything up in this case? That also calls your claim of the fastest raytracer being on a GPU into question.

 

[Ho_Ho]

It has been done... in fact to my knowledge the current fastest raytracer runs on the GPU using Sh (http://libsh.org).

See http://portal.acm.org/citation.cfm?id=1071866.1071869.

I haven't read the paper yet but in their abstract they say they use kd-trees but I see no reference of building them on GPU. Using is relatively simple and easily parallelizable. Building them is not a trivial task and desn't scale well.

Currently I know only one way to parallelize building KDtrees and that is to build each tree separately but that won't help much when a complicated object needs its tree to rebuilt. One possibility would be to use key framed animation but in a world of rag dolls that wouldn't look so good

 

[davepermen]

yeah, using kd-trees in a pixelshader shouldn't be a big issue. and with the fast branching hardware of ati, it should be very fast.
generating them is quite another topic... animation is still not solved that nice. thoughts about updating the tree, having some less-optimal but animatable tree direcly, or what ever, come up, but actual implementation tests are still missing..

well well..

but i'm definitely interested in a performance comparison between a highly optimised stanford bunny raytraced in software, and on the new ati cards. (oh, and the rasterised version as well ). i bet the ati card can beat a single cpu by quite some margin, if done right.

 

[Ho_Ho]

but i'm definitely interested in a performance comparison between a highly optimised stanford bunny raytraced in software, and on the new ati cards. (oh, and the rasterised version as well ). i bet the ati card can beat a single cpu by quite some margin, if done right.

Why test with a relatively low poly bunny (~70k)? Most newer games have about as much polygons visible per frame as the entire bunny. Why not throw a powerplant at it with all of its 12.5M polygons to see what ray tracers are really good for: high polycount scene rendering.

According to my calculations using the same data structure as Wald described in his thesis storing only the tirangle intersection data (no KD-tree or actual triangle coordinates!) it takes roughly 600MiB for the powerplant. Huge memory useage is one thing that doesn't make GPU ray tracing very attractive in my oppinion. Perhaps using AGP aperture or unlimited texture memory that NV's Quadros* have help a little but I'm not sure how effective it is compared to storing everything in vram. I know their prototype hardware could live fine with some kiB's of cache and 4MiB's of on board memory and streaming all required data over the PCI (not PCIe) bus from PC's memory.

*)perhaps ATI professional cards have it too, I haven't checked


a bit OT, why can't I edit my messages? It sais in the FAQ that there is an edit button but I haven't seen it on any of my messages. I have seen other changing their messages so it should be possible. Or is there some rule that applies for newbies that I somehow missed?

 

[davepermen]

well, the bunny would've just been a start. i thought of it because i know we can run that one about realtime on a todays cpu, completely in software. would be nice to see how hw compares to it, thats all. of course, powerplant and similar things would be lovely, too ur unreal engine 3 models, or what ever (not the ingame, but the real models )

no clue why you can't edit. i'll test if i can edit here.

[edit]yes, i can edit..[/edit]

 

[bloodbob]

but i'm definitely interested in a performance comparison between a highly optimised stanford bunny raytraced in software, and on the new ati cards. (oh, and the rasterised version as well ). i bet the ati card can beat a single cpu by quite some margin, if done right.

Yeah and then try some hardware accelerated ray tracing.

 

[davepermen]

Yeah and then try some hardware accelerated ray tracing.

i'm only talking about raytracing here anyways.

 

[AlphaWolf]

a bit OT, why can't I edit my messages? It sais in the FAQ that there is an edit button but I haven't seen it on any of my messages. I have seen other changing their messages so it should be possible. Or is there some rule that applies for newbies that I somehow missed?

Yes its a new member thing, you need 5days registered, 10 reputation points (the default) and 10 posts(not counting general and feedback posts) to make junior member. I think junior can edit but not sure.

Info on the member policy is here.

 

[Ho_Ho]

i'm only talking about raytracing here anyways.

Well, there are at least two HW accelerators for ray tracing:
First SaarCOR Prototype
Second Prototype: RPU

They are extreemly efficient compared to CPUs or GPUs. I think for the same transistor count and MHz as R520 they would simply annihilate it in ray tracing.


Thanks for the usergoups link. I had no idea there is such thing.

 

[davepermen]

hehe, i know those..

 

[Matasar]

Im lost, did anyone answer the question in first post ?
Some say NV has that as an extension and it would speed up "radiosity" big time and other cards except GF7 just dont have the power to do it.

 

[davepermen]

Im lost, did anyone answer the question in first post ?
Some say NV has that as an extension and it would speed up "radiosity" big time and other cards except GF7 just dont have the power to do it.

it got answered if you read the thread: there is no way radiosity (the algorithm) can be done in a useful-for-games manner on any todays (and, the way the algorithm works, future) hardware.

the fact that nvidia has it on their page is just some marketing fuzz. they possibly put that in because of some PRECOMPUTED radiance transfer demos, that run in realtime (they, in realtime, show precomputed data.. wow ).

and no, there is no reason the ati hardware can not handle any solution or fake solution nvidia has to show anything. quite the other way round..

the ati card is definitely capable of hardware accelerate radiosity calculators, thus could speed up the lightmap generator in a q3 level editor (or similar stuff). it's gpgpu capabilities are still waiting to really show up.

simple said: the dude who told the OP, that nvidia can handle radiosity while ati can't, simply said something wrong.

 

[Matasar]

Thanks for your answer!

 

[Subtlesnake]

http://www.gamedev.net/reference/programming/features/rtradiosity2/default.asp

Would this work?

 

[davepermen]

yeah, but afai remember, this is simply some approach similar to cubemapping. but i'm unsure, just looked at the link, haven't read (and it's long time ago). nothing to do with real radiosity (but has some of the effects of it). and definitely not an "automatic all purpose" way. means a solution. only a hack.

 

[Ho_Ho]

http://www.gamedev.net/reference/programming/features/rtradiosity2/default.asp

Would this work?

Probably not very well.
Those two articles were pure theory with hints that he has implemented it but for some reason he had didn't show any images, at least I couldn't find any. I wonder why.