Open Problems discussion

(L) [2007/09/18] [Zakalwe] [Open Problems discussion] Wayback!

Due to prior commitments I couldn't make RT07. One of the parts I was really looking forward to was the Open Problems section. I thought therefore we could have our own discussion here.

Here's the lists from RT06 and 07 amalgamated into one. I've made some start on discussion.

Soft Shadows

Traditionally these require a lot of extra rays being shot. These days, having realtime soft shadows is a big selling point in many games (see FEAR for example). We can of course fake it by little more than blurring a properly cast shadow, but one of the beauties of ray tracing is getting away from these types of hacks so prevalent in rasterisation.

Level of Detail.

Another "hack", but in this case I believe a more elegant one. With careful use resulting renders would be nearly, or perhaps completely indistinguishable from non-LOD renders, if the proxies for the full model are used in a proper manner. After all, why waste time intersecting a blob of 10 nearly identical triangles across the stadium for one pixel when a proxy of 1 will give the same result?

Animation (Dynamic models/API/Scene graph)

Robustness and Smoothness.

Efficient AntiAliasing.

I'm interested in what the problems are seen to be regarding adaptive sub-sampling and why this was mentioned.

SIMD >4.

Why is this listed as a problem? Surely this is more of a solution waiting for a problem and that problem is packet tracing. If our >4 SIMD registers can be broken down into hi/lo constituent parts which SIMD operators can still be applied to still then even better. I'm sure many of us would adapt our code to a 512bit wide register in minutes and be glad of it.

Particle/Volume effects.

Displacement mapped surfaces (including hair, fur)..

Hierarchal Geometry..

Faster RT on GPUs (special architectures).

Memory Bandwidth.

A real problem which will require some serious work and the real limiting factor in ray tracing. It's compounded by the introduction of more cores, dynamic scenes and massive models. Ideally we should be scaling nearly linearly with the addition of more processing units. Current results are not showing that.

Real-world cameras.

RT distribution?

Educate Game developers.

Temporal Aliasing.

Advanced lighting and materials..

Incoherent Rays

Still a problem when we want to make our sharers more complex. There's a reason most RT engines concentrate on primary and shadow rays and that's because, as we all know, it's easy to stick them in packets.

(L) [2007/09/18] [Phantom] [Open Problems discussion] Wayback!

Allow me to comment, since I DID make it to RT07. [SMILEY Smile]

Adaptive sub-sampling: In my keynote, I mentioned that adaptive supersampling is very much related (imo) to the adaptive subsampling that we need to get things like GI / soft shadows / AO fast. Basically: How do we efficiently bring back the number of rays for certain expensive, but generally low-frequent 'effects'? I was mentioning this because we are approaching the point where RTRT is 'fast enough' as a basic tool, so we need to start looking at what we want to use this tool for. Later someone mentioned (during a literal walk in the park) that this 'tool' may consist of a hammer and a single nail, and now we expect to build a house. [SMILEY Smile]

SIMD >4: Apparently, some people expect some hardware with larger than 128bit SIMD registers. I guess they are looking for ways to exploit that efficiently, although I think that should be straightforward.

Bandwidth: There was another GPU ray tracer at RT07 (http://www.mpi-inf.mpg.de/~guenther/BVHonGPU/index.html), and it's author claimed that it was not memory bound. Basically, we concluded that it's thus using too much computation, since it doesn't run any faster than Horn's tracer, which WAS memory bound (big time). So basically, the bandwidth problem is hitting the GPU hard.

Educate game developers: If you look around on gamedev.net and other fora for (wanabe) gamedevelopers, it appears that most developers are not aware (and perhaps not interested anymore) in the state of the art of RTRT. They once worshipped the RealStorm guys (see their guestbook), but now, the GPU reigns. If RTRT is ever going to take over the world, we need to tap into the huge potential of the game development community and gamedev hobbyists. These guys have worked wonders before (again, realstorm) and they can do it again, if they just would believe. This would also pave the way for industry adaptation: RTRT may not beat the GPU in terms of raw image quality within the next three years, but it sure is a lot more fun to program for. But no-one seems to know that.

And finally, one that was on the beamer and not on the chalkboard: "Don't forget the killer RTRT demo". NVidia is reaching millions with it's awesome demos. RTRT can reach millions too if only we had a kick-ass demo that hides RTRT problems and emphasizes it's advantages. Success of such a demo is just as much a matter of getting good designers on the job, as good developers. This appears to be rather though.
_________________
--------------------------------------------------------------

Whatever

(L) [2007/09/19] [Zakalwe] [Open Problems discussion] Wayback!

Killer Demos(TM) will require eye candy and lots of it. Currently in RTRT there's nothing that can't be done (or faked reasonably well) via rasterisation on a GPU. We just haven't got the required horsepower/methods yet that will let us unleash the really cool things in raytracing to any great level (reflection, refraction, atmospheric effects, complex camera models etc).

Soft Shadows

I've had a quick first read through the beam-tracing paper in the other thread and despite my early misgivings over the method. I'm pretty impressed at first glance.

Open Problems discussion back