Arauna GI experiments back
(L) [2007/03/25] [lycium] [Arauna GI experiments] Wayback!you wrote "realtime ray tracing" on an image using 25 samples per pixel and many shadow rays? (btw, 25 samples per pixel ought to look reaaaally smooth for normal ray tracing)
btw, since i guess you haven't looked at irradiance caching, it's definitely not something for realtime use since you dynamically construct an octree as you render the image... no idea how you plan to distribute that across your 8 cores, and even when it's built i'd hardly call the resulting scheme "fast" (let alone realtime) unless you're comparing to brute path tracing.
(L) [2007/03/25] [lycium] [Arauna GI experiments] Wayback!ah yes, that should have been self-evident :|
(L) [2007/03/26] [Phantom] [Arauna GI experiments] Wayback!It's false advertising. I'm doing 25 samples over the hemisphere for each pixel, so it's more like 3 seconds / frame. And that's just for the ambient occlusion; full GI is even slower. It did however produce an interesting image, and I wanted a wallpaper, so I added the ray tracer title screen. [SMILEY Smile]
About this being realtime: I don't see what the problem is with the irradiance cache and the octree? For most pixels I'm just reading the tree, and occasionally I write to it. I think a 10x improvement over naive sampling is possible; for lower resolutions and crude approximations this should give me 10fps on the 8-core.
I did read the papers on irradiance caching (several times in fact); Ward is mentioning temporal coherence, and I was indeed hoping not to construct the whole tree each frame.
_________________
--------------------------------------------------------------
Whatever
(L) [2007/03/26] [bouliiii] [Arauna GI experiments] Wayback!I really think that Interleaved sampling + discontinuity buffering is a good solution (not too bad). You just have to separate self-color of the object (must not be filtered) and irradiance. It is simple and fast.
The problem is for normal mapping of course.
Ben
(L) [2007/03/26] [Phantom] [Arauna GI experiments] Wayback!Ben,
I tried the VPL approach, and indeed it gives good results, even with few VPLs. I did not yet try the discontinuity buffer yet though.
I'm worried about two things, and that's why I am also investigating some other options:
1. I need to use quite some VPLs, because my students are working on a game level. Wald is suggesting ways to ignore lights that will not contribute, but that's a complication.
2. In Wald's thesis, interleaved sampling shows severe artifacts on smaller details (well, not so small actually, the round edge of the table in the office scene).
3. The discontinuity filter must be applied to the final image and will benefit little from multiple cores (I encountered this problem when doing HDR filters in software). On the 8-core, this means that 7 cores will be idling.
So I did some quick tests to compare results of 1-bounce GI and ambient occlusion. In my framework, these are easily exchangable, so I could even support both (depending on the platform). Ambient occlusion is close to 1fps on my system (by limiting the size of the hemisphere and the number of samples), so if Ward is right this could be interactive with the irradiance cache (which maps nicely to all sorts of hemisphere sampling, including ambient occlusion).
Another option that I only just encountered is to gather surfaces by doing a kd-tree proximity query using the hemisphere (just like a photon lookup, only with half the sphere); this quickly yields a set of 'interesting' nearby surfaces that could also be used to approximate ambient occlusion.
I'm in exploration mode as you can see. [SMILEY Smile]
- Jacco.
_________________
--------------------------------------------------------------
Whatever
(L) [2007/03/26] [bouliiii] [Arauna GI experiments] Wayback!I am preparing some screen shots with Yacort to show the results with some scenes (not necessary simple)
Ben
(L) [2007/03/26] [Phantom] [Arauna GI experiments] Wayback!I'm looking forward to those shots!
@Toxie: I never realised that Wald didn't split those. Well in that case it's going to be hard to beat that with an irradiance cache, isn't it?
_________________
--------------------------------------------------------------
Whatever
(L) [2007/03/26] [bouliiii] [Arauna GI experiments] Wayback!BTW, the results are MUCH better if you use textures, but I have no scene with textures ready right now!
Ben
(L) [2007/03/26] [lycium] [Arauna GI experiments] Wayback!you will find irradiance caching can't be threaded (if you somehow do, write a paper for rt07 since you're the keynote speaker!), whereas the awesomely-named discobuffer splits work exactly evenly. if that's not enough motivation on 8 cores, then it will be with more (besides this irradiance caching, especially with gradients to make it not-ugly-as-sin, involves significant mathematics).
btw bouliiii, those are very impressive shots for 1fps! i have to wonder though, if they are perhaps "cherry picked" since 128 VPLs leave a lot of room for big aliasing error. the key statistic is how the framerate varies with virtual light count, which, after some point, i would imagine to be sub-linear; any info in that regard would be really awesome :)
(L) [2007/03/26] [lycium] [Arauna GI experiments] Wayback!with cherry picking i mean something like "showing the best and avoiding the worst"; i haven't thought about how you might choose VPLs by metro-sampling, but certainly that will make the result more "stable".
to give an example of what i mean about the aliasing, i mean the sudden switching on and off of VPL as can be seen here: [LINK http://homepages.paradise.net.nz/nickamy/realtimerad/realtimerad.html]
(L) [2007/03/26] [bouliiii] [Arauna GI experiments] Wayback!Ah ok, I do not specially take some screen shots, but that is sure some are worst than these ones. Otherwise, the sampler I used is really stable and efficient. I have to techreport it ASAP before plubishing the code (I am currently wirting another article and I am beginning to miss time !). I am preparing a screen shot where there are more artifacts.
Ben
(L) [2007/03/26] [lycium] [Arauna GI experiments] Wayback!very much looking forward to it! be sure to make a new thread about it though, since it's a little off-topic here [SMILEY ]
(L) [2007/03/26] [Phantom] [Arauna GI experiments] Wayback!Just implemented the discontinuity buffer. Lycium, I don't see how you could make it multithreaded? I followed the description from 'Interactive Global Illumination', [LINK http://graphics.cs.uni-sb.de/~wald/Publications/2002_IGI/2002_IGI.pdf,] that states (if I interpret correctly) that the discobuffer is simply a NxN box filter where each point is only included if it's normal & distance matches that of the sample point. I can see how a GPU could do this much better than what I have now though. [SMILEY Smile]
Anyway, I tried it on crude ambient occlusion (no time yet to dive in the other approaches in-depth); the effect is that I get a virtually noise-less image now with just 16 rays (per pixel). Still not interactive, but the 8-core at least renders several frames per second.
I hope to have some time to revisit IGI later this week.
_________________
--------------------------------------------------------------
Whatever
(L) [2007/03/26] [lycium] [Arauna GI experiments] Wayback!i imagine your N threads generate the to-be-interleaved samples, then you split the image into segments and process them independently. at least i hope this is correct (and don't see why not), i've never implemented discobuffering (damn i love that term).
with irradiance caching you'd have to lock the octree each time a thread wants to update, and especially near the top of the tree in the beginning, that's probably brutal.
(L) [2007/03/27] [bouliiii] [Arauna GI experiments] Wayback!Hello Jacco,
there is no specific problem for multi-threading, our own implementation is completely multithreaded (damn, I must do this release !!)
The idea is to use deferred shading.
1/ You store normal, pos, dir intersection data (triangle id, instance id and intersection depth) --> easy to multi-thread
SYNCHRO
2/ With the buffers, you compute a discontinuity buffer --> if(pixel == BLACK) no disco otherwise discontinuity
--> this must be made simply and quickly by checking the previous buffers and making the test between the pixel and its three "upper" neighbors (the right one, the top onje and the right top one)
--> easy to multi thread
SYNCHRO
3/ Shading
--> you fill two buffers where you can put the irradiance and the self color. Instead of generating the primary rays, just fetch the data in the previous buffers.
--> which is a bit more technical is to assign for each ray its pool of light source. The best thing to do is to enumerate the light sources during shading and to use a stride equal to the size of the interleaved pattern for each interleaved ray pack.
stg like
for(i = ray->index; i < light_n; i += pattern_size) {
make shading stuffs
}
where 0 =<ray>index <pattern_size (each ray inside a given interleaved pattern has a different index)
SYNCHRO
4/you perform a two pass box filter. The idea: you filter horizontally (resp. vertically) until you find a discontinuity. At the same time, you can blend the filtered irradiance and the self color of the objects.
Et hop!
I hope it may help.
Ben
(the multi-thread can be simply done with tiles for each step. For the shading part, you can make stg different by assigning to each thread a pattern. I did not do that because I think it is either too slow (if each thread fill a entire regular grid (I am refering to the original paper by Keller), the final synchro point may be very expensive since you do not control the granularity of the data)) or too complicated (if you assign for each thread one piece of regular grid then you merge everything ) The good thing is I think, to have once again tiles to give to each thread with a size proportional to the pattern size. Otherwise, it is still easy to mask the inactive rays which are outside the tile (it is what I finally did to let the user choose the tile size AND the pattern)
(L) [2007/03/31] [Phantom] [Arauna GI experiments] Wayback!Ben,
I just realised something: It would be hard (if not impossible) to combine your approach with reflections and refractions, right?
- Jacco.
_________________
--------------------------------------------------------------
Whatever
(L) [2007/03/31] [bouliiii] [Arauna GI experiments] Wayback!If there is a superposition of reflection / diffuse contributions on the same surface, you have to make several passes. I do not see other method. (actually you can also make a diffuse buffer and a reflection buffer or stg like that)
Ben
back
