Page 1 of 2

complex rendering

Posted: Fri Feb 03, 2012 11:16 am
by mpeterson
enclosed a first testshot of our new rt-engine. its a head lamp with around 2.5mio ploygons.
here we used a combination of our multi-frustum traversal + ray sorting/streaming. on a dual
sandy-bridge system 6fps are possible at 1920x1080. there is no progessive rendering like
elsewhere. the final quality is available at 6fps.


4-8 spp (adaptive filter importance sampling)
15 bounces
mc hdr-sampling (hdr-size: 28000x14000)
12 different shader (datasets comes with 62 materials)
full hd
2.5 mio polygons
final quality (see image) in around 150ms


Re: complex rendering

Posted: Fri Feb 03, 2012 11:27 am
by jbikker
Wow, amazing quality at that speed! Is this Whitted-style ray tracing?

Re: complex rendering

Posted: Fri Feb 03, 2012 12:30 pm
by mpeterson
jbikker wrote:Wow, amazing quality at that speed! Is this Whitted-style ray tracing?
more or less yes. we trace refl/refr at the same time (no mc here). so its exploding after a time (recursion depth 15).
we block/discard rays when their energy drops below a threshold. hdr is pre-processed and mc sampled with 16 rays per hit
on diffuse surfaces.


Re: complex rendering

Posted: Fri Feb 03, 2012 2:22 pm
by ingenious
Ah, BMW X6... :)

Very good image indeed! And yeah, you have to trace both reflection and refraction if you don't want excessive noise. I'd really love to see a video how this quality is achieved in real time.

Re: complex rendering

Posted: Tue Feb 28, 2012 3:45 pm
by spectral

It is hard to beleive :-P , so you got the CPU renderer speed world record :-P


Re: complex rendering

Posted: Sat Mar 03, 2012 3:35 pm
by SonKim

How is that possible, 150 ms? You must be from the future...a time traveler with advance rendering knowledge! Do you plan on commercializing your rt-engine? I would love to see some more shaders :D

Re: complex rendering

Posted: Sat Mar 03, 2012 5:33 pm
by graphicsMan
Okay. This is beyond amazing. Just eye rays: 28000*14000*4/0.15 = 10 billion rays per second. If you estimate that with a ray depth of 15 and with exponential increase due to splitting at glass interfaces, we'll go easy and say that the average path length is 10 (probably still on the low side, even accounting for direct background hits). That means you're tracing 100 billion rays per second.

The best numbers I've seen elsewhere are about a billion rays per second in the completely coherent case. Two orders of magnitude difference. Two. That is a hell of a difference. This is on 16 cores? Over 6 billion rays per second per core?

How well does your algorithm work for very incoherent rays? Scratch that. I just saw your other thread claiming > 500 Mrays/s for coherent rays and > 100 Mrays/s for incoherent rays. These seem like reasonable numbers. Is there a typo in your 28000x14000? Is it supposed to be 2800x1400? Still, these are good numbers.

Re: complex rendering

Posted: Sat Mar 03, 2012 9:20 pm
by franz
According to his description, 28000x14000 is the resolution of his environment map; rendering speed should be pretty much independent of that parameter. Still 6 fps at 1920x1080 with 15 bounces is kind of cool.


Re: complex rendering

Posted: Sat Mar 03, 2012 9:54 pm
by Serendipity
Hold your horses: He said that the HDR they are using for lighting is 28000x14000 pixel - which is large but also pretty much irrelevant since also smaller resolutions don´t fit into any cache.
The render resolution was FullHD, so it´s 1920 x 1080 x 4spp x 6fps x 15 depth ( not considering the 2 rays per glass hit because that would be insane) = 746.496.000 Rays/s - which is still impressive for a dual sandy bridge (i am assuming it is a 2 x 6 Core system, which would mean 24 virtual cores). I didn´t count the 16 mc sampled env "things" because everything that is diffuse in that scene is essentially black so probably isn´t sampled at all and the glossy parts are not sampled either (at least it looks that way since the results are wrong).

But the performance is really impressive.

Re: complex rendering

Posted: Sun Mar 04, 2012 4:54 pm
by graphicsMan
Wow, somehow I really botched my reading of this description. Doh! I thought it said the rendered image was 28kx14k. At 1080p, that's far more reasonable, but still impressive.

Is the sorting/streaming using something similar to Garanzha et al?