The noblest of all dogs is the hot dog. It feeds the hand that bites it - Hot Dog Restaurant
|
|
|
Intel pushes Raytracing again
YOU MAY RECOGNISE the name Daniel Pohl from his raytraced Quake games and various talks surrounding those efforts. At GDC, the talk was familiar once again, and more progress is being made toward that glorious raycast future.
For those not in the know, there are two methods of drawing a picture, the common one is called rasterisation, and it is what you see in about 99.9% of games out there. The less common one is called raytracing, and it is much more accurate, but also vastly more computationally intensive.
Pohl wrote a version of Quake 3, later 4 as well, that used a raytraced engine instead of a rasteriser. Both methods have advantages and disadvantages, but in a situation bound by computational resources, rasterisation wins at the cost of pixel perfect quality.
The nice thing about Raytracing is that is is pixel perfect, what you see is what you see, or something to that effect. The problem is that for every pixel, you must trace a ray from your eye to the first polygon on the screen. If that point can draw a line to the nearest light, it is lit, if not, it is shadowed.
In the case of reflective surfaces, it then casts another ray at the right angle and colours the pixel it came from the correct fashion. Multiple bounces can require multiple pixels, and if you think about it, a 1024 * 1024 screen takes over a million initial rays. This is not a trivial number, but you can do it, and you can do the multiple reflections in a sane way. This is not always the case with rasterisation.
There are also some interesting effects that are hard to do if you are not raytracing, shadows tend to be the big problem here. Most game engines cry if you try to make them cast shadows on a stairway, but with raytracing, it is no problem. Same with refraction in semi-transparent surfaces. With raytracing, you just cast another ray at the angle specified by the substance's index of refraction.
Other than that, you can do a lot of other fun tweaks. Collision detection is another one, since you are casting rays that intersect with surfaces, you can do collision detection for essentially free with the code you already have.
Camera portals are another similar hack, hard to do but easy with rays. With raytracing, when a ray hits a portal, you just add an offset to it. As Pohl puts it, three lines of code. If you need to do pixel perfect recursion, portals in portals in portals in .... where was I? Oh yeah, in portals and more portals, it is the "same three lines of code". They were demonstrating 140 deep portal recursion with perfect accuracy, try that in the game Portal and watch your CPU melt.
In the end, we are not there yet. With a 16 core system, you can match what a mid-range GPU does now in terms of pixel pushing, but beat it on quality and effects. This won't cut for games yet, but there is much work to be done in optimisation, and Larrabee will change the landscape. µ
Tags: Intel
Share this:
Share
Wilson hedjek, could you give me more detail about who this raytracing Bay area group is? As you can see from my following comment I have particular interest in it.


I try to check on what these guys are doing sometimes. Trying to get to talk to them is fairly difficult. I thought they already were trying to speed it up through GPU programming.

Notice Intel is promising 3D graphics again, and I imagine this could have something to do with it.

About ten years ago I came up with a near photo realistic ray tracing technique that should be able to theoretically render Hollywood like animation on an single CPU at 1080p for my OS graphics API with some minor hardware. Even without hardware, the GPU might still be rendering DVD like frame live. Even full featured hardware might be much less complex than any current GPU graphics card, and deliver many times the performance for the amount of hardware. The other thing that Raytracing suffers from, when used with high quality modeling techniques, like 3D volumetric pixels, is the shear size of the models compared to using polygons of lesser quality. With my 3D pixel technique I am reasonably confident I can get model sizes to within striking distance of existing polygons techniques with as much flexibility. I still think up new refinements for the technologies every now and then.

I have had to put it on hold because of sickness and other reasons, but imagine what it could do on Nvidia/ATI GPU's.

I have tried to contact the RTRT/Polh people a few times about working with them, but from memory I don't know if I even got much of a reply. Probably pretty confronting when somebody emails you in saying that they have an technique that performs as well in software to what you do in hardware, and in hardware as good as any graphic card for less hardware complexity. Maybe I should contact Intel instead.

What bay area firm is that:

I would be interested in hearing which Bay area firm is doing that 3D hardware? Sometime ago there was an bunch of New Zealanders showing an true volumetric free floating in the air technique (no mirrors, no fog/vapor screen, no enclosure etc) over here looking for funding. A good friend of mine, who is now passed away, met them, and they had publicly used the technology (from what I gather at an festival, this is really the backwoods of the technological world, people do think that it is a good idea to be so open). They eventually went to the US and last my friend heard from them, disappeared in the Bay Area, near where an large naval research centre was. Within 6 months. or an year. after that the joint armed, whatever, services announced that they would be using holograms on the battle field.

Another interesting thing that happened a little after that, was that a acquaintance who was into some particular computing technology, around that location, had been to an firm developing optical computing technology to an advanced stage, so advanced it was supposed to be on the market within an short time many years ago. He was investigating an partnership with them, and knew about what he was being told and shown. I had speculated previously to my dead friend, about the 3D technology, that the only way they could do what they were doing, because of processing requirements with conventional 3D pixel techniques, was if they were using optical computing, which nobody was supposed to have developed to that stage, I got the acknowledgement look and slight nod to that. I speculate that this group was related to the 3D group, and I would like to find out about this other 3D raytrace Bay area group.
I doubt ray tracing will be the end of GPU's. GPU's could be reconfigured to handle ray tracing very nicely, and I'm guessing that with all the shader language, some wonderful new hybrid could run without too much chip modification. The two technologies are slowly merging. They have been for many years, and even studios like Pixar are using ray tracing as one of many rendering techniques. There are many things (like motion blur) that ray tracing stinks at, and many techniques that are handled well by both ray tracing and raster graphics. For example, both now have techniques for global illumination, which drastically improves the lighting over any simple ray tracing that is/was being done. If you want examples of that, check out radiosity.
Raytracing is not this magical holy grail of graphics. It does not automatically give you perfect photorealistic images. Take a look at a 20 year old raytrace image that everyone went ooh and ahh over back in the day.

With fully programmable shader pipelines, raster systems are Turing-complete. This means anything that can be done in a raytracer can be done in a rasterizer. It's just a question of efficiency. Likewise radiosity has yet another set of good and bad cases.

Yes volume shadows are easier raytraced. Ok now try soft shadows. Oh you want anti-aliasing? Better get a lot more cores ready. 

The hardware accelerator market could have really gone either way. There were demos of raytrace accelerators about the same time, but it didn't pan out.

If it had, we'd probably have had Quake 2 with Doom 3's shadowing system, but we'd just now be gasping in amazement at the first hardware fast enough to run a game at a resolution above 1024x768.
I recently heard a rumor that a small graphics started up in the Bay Area has figured out how to do Real-Time Ray Tracing at 1080P at over 100 fps. Unlike the pseudo RTRT Quake 4 / Daniel Pohl / Intel demo we've all seen, I heard they can really do it. Can anyone confirm this? Has anyone seen this? Who are these guys? What will this mean to the industry, gaming, design, fx, et al? Could this be possible?
Why would this have to be done on the CPU?
Instead of GPU's being used for rasterisation type graphics why can't their huge parallel processing power be used to do ray tracing on a GPGPU type of basis? God how much power do you need with quad crossfire you could have 1280 stream processors and 2GB of GDDR4 memory to assist.
Does ray tracing spell the end of the graphics card industry? If quality of graphics is superior using ray tracing, won't the massively multi core chips coming from Intel and AMD soon make graphics cards redundant? Surely specialized chips like Cell can already do this. If so, why isn't a cell graphics card being released?
...but how far do you follow it?

For example: recursion depth and light-source area & quality settings?

I'm not trying to nay-say this development but I think it's probably got a lot further to go than you'd think before it really makes any practical difference, from listening to the hype.

Pure ray-tracing _is_ capable of photographic quality but it usually takes > an hour per (dvd qual) frame (depending upon your farm) so it's a bit difficult to treat this as the breakthrough it claims to be.

R-T is one app where mpp _really_ makes sense.