[MJP]

That's pretty neat that they trace voxels. It's probably also useful for other things to have a voxelized representation of the scene. I wonder how much time they spend per-frame voxelizing things.

[Solarus]

Would their GI implementation be closer to what Ubisoft is doing with Far Cry 3 PC or is this steps above that as well?

[CNCAddict]

Apparently this new GI technique also handles AO natively and is shockingly close to path tracing..only instead of tracing rays, it's tracing cones of various sizes depending on the needs and accuracy one desires. For specular the cones need to be narrow, for diffuse very wide.

It looks like Epic is actually ahead of Crytek now in their lighting method with their relatively accurate reflections, GI, and AO all in one algorithm!

http://blog.icare3d.org/2012/05/gtc-...ed-sparse.html

[fellix]

Yea, the AO looks much more precise than all the screen-space methods out there. Finally a good looking real-time implementation.

[doob]

Watching the presentation about the octree-based sparse voxelization for real time global illumination, when he reaches around 15:40 there's an odd "nvidia only" written with regards to the atomicadds and im no programmer, but i was under the impression that AMD/ATI is capable of such as well (isn't it a requisite for directx11 compliance?). And what do they mean by native?

Is there anything regarding AMD's 7xxx architecture that would pose a problem to this technique?

Edit: Apologies for my naivety regarding the questions/curiosity in the matter, but wanted to add that from a common consumer point of view, reading stuff like that casts doubt on short-med term regarding the features of current competitor products, even more so considering how popular unreal engine was, is and will be. So any light on the matter from knowledgeable people is very appreciated.

[AlexV]

Originally Posted by doob

Watching the presentation about the octree-based sparse voxelization for real time global illumination, when he reaches around 15:40 there's an odd "nvidia only" written with regards to the atomicadds and im no programmer, but i was under the impression that AMD/ATI is capable of such as well (isn't it a requisite for directx11 compliance?). And what do they mean by native?

They're probably thinking about float atomics, which only NV does currently. Interestingly, that's also outside of DX spec IIRC (not sure about 11.1), and exposed in GL/CL via extensions, so reliance on it is curious. Of course just doing atomic exchanges is doable, since you'd just cast the result to float even if you're doing it through int atomics, but the rest of the arithmetic is more hairy. You can work around lack of float atomics it by manual scatter gather and sync at kernel boundary (not the best solution, mind you), or try to remember how float math works and rely on bitwise ops to emulate it.

[doob]

Would those workarounds lead to meaningfull performance impact to the point of unplayability? Can you or someone else try to guess-estimate(guessing appart from epic itself, pretty hard to do, anyway just in case it is, roughly) the performance hit with best case scenario/workaround if say NVidia was rendering a scene at 30fps relative to AMD?

[3dcgi]

Epic probably implemented the algorithm differently. Especially since that presentation was very recent. It's probably pointless to guess and extrapolate that to UE4 performance.

[ninelven]

Even with atomics, the quoted performance was 16 FPS @ 1920x1080 with a GTX 680. I'm guessing that is without any actual game code running as well. That isn't really viable for actual games today or even the near future. Maybe for the Maxwell generation though...

[MJP]

*deleted*

[jlippo]

Originally Posted by MJP

That's pretty neat that they trace voxels. It's probably also useful for other things to have a voxelized representation of the scene. I wonder how much time they spend per-frame voxelizing things.

I wouldn't be surprised if they use the information for particle collisions and such.
Shadow rendering might be nice as well, either direct tracing or rendering lot's of shadow maps in one pass.

[Rodéric]

Latest Voxel stuff I saw when I was looking up voxels and SVO earlier this year was :
http://maverick.inria.fr/Membres/Cyril.Crassin/

Been thinking about using SVO in a game, but can't really figure out whether it would be a win (memory, perf or quality wise) with a rasterizer.
(Also assumed it was linked to Carmack's idea of "MegaGeometry"[probably means virtual geometry like virtual textures], but I found other ways to do the same with vertices.)

[Bumpyride]

I was kind of expecting more discussion here. I mean, the voxel cone tracing algorithm is awesome. It's clever and efficient and now it's implemented in what's probably the most licensed game engine. Seems like a big deal to me.

Anyway, there are a few things here that haven't come up yet that I'm curious about people's impressions of:

1) Tessellation. Obviously the mountains scream it and the main character's spikiness looks obvious, but are these displacement maps on conventional meshes? Are they using any higher order surfaces in general? It doesn't look like it to me.

2) Textures. There isn't a mind-blowing amount of texture detail here so its hard to say if there's any virtualized texture tech being used - at least nothing beyond ue3.

3) AA. No jaggies visible to me in either video. Most likely MSAA but the shader aliasing looks solid as well. It's hard to say from a youtube video.

4) Light emitting particles. I've looked at this a few times and I can't really figure out if the particles are emitters in the GI solution. The room with the 'ice' particles does turn blue but it's hard to say if that's light from the sphere that's generating them or the particles themselves. The mass of them in the 'fire' element doesn't seem to be reflected in the lighting around the room to me.

5) How is that lava flow being handled? Looks like a pretty solid (not perfect but solid) fluid sim to me. Pretty cool if it's generalized and they can just drop a fluid in there like that.

That's about all I got.

[doob]

Originally Posted by ninelven

Even with atomics, the quoted performance was 16 FPS @ 1920x1080 with a GTX 680. I'm guessing that is without any actual game code running as well. That isn't really viable for actual games today or even the near future. Maybe for the Maxwell generation though...

They seem to be already using it in UE4 and you can hear the speaker of the development walkthrough video refering to their voxelised global ilumination features. And the technique doesn't seem to rely on the cpu to any significant degree. All the heavy work is done on the gpu, the octree data is all kept in the gpu('s) memory (edit: actually he says global video memory, so i guess it can spill into system memory? Still don't see how the cpu/game code would be that relevant also keep in mind they'r running it on the editor with eventually allot of game code running&editing).

Now how did they achieve it at acceptable performance (guessing around 30fps)? Maybe by trading off precision/quality for performance. The other presentation on NVidia does mention when talking about the glossy reflections that lowering the resolution/precision/ammount of specular reflection cones gives performance increase. Perhaps that was one way they achieved it for UE4 elemental demo.

[ninelven]

Now how did they achieve it at acceptable performance (guessing around 30fps)? Maybe by trading off precision/quality for performance

Doesn't really matter, even 30 fps on a GTX680 isn't really usable (b/c most people don't have a 680+). I expect it will be some time before that level of performance is considered average, at least another node shrink.

And yes, game code will still slow it down.

It's nice; I just don't expect to see it really used in titles for a few years.

[Davros]

Didnt want to start a new thread for this :

This video is from a Square Enix Real Time tech demo to demonstrate the capabilities of their new, "Luminous," engine.

[Ike Turner]

Originally Posted by Bumpyride

I was kind of expecting more discussion here. I mean, the voxel cone tracing algorithm is awesome. It's clever and efficient and now it's implemented in what's probably the most licensed game engine. Seems like a big deal to me.

Anyway, there are a few things here that haven't come up yet that I'm curious about people's impressions of:

1) Tessellation. Obviously the mountains scream it and the main character's spikiness looks obvious, but are these displacement maps on conventional meshes? Are they using any higher order surfaces in general? It doesn't look like it to me.

2) Textures. There isn't a mind-blowing amount of texture detail here so its hard to say if there's any virtualized texture tech being used - at least nothing beyond ue3.

3) AA. No jaggies visible to me in either video. Most likely MSAA but the shader aliasing looks solid as well. It's hard to say from a youtube video.

4) Light emitting particles. I've looked at this a few times and I can't really figure out if the particles are emitters in the GI solution. The room with the 'ice' particles does turn blue but it's hard to say if that's light from the sphere that's generating them or the particles themselves. The mass of them in the 'fire' element doesn't seem to be reflected in the lighting around the room to me.

5) How is that lava flow being handled? Looks like a pretty solid (not perfect but solid) fluid sim to me. Pretty cool if it's generalized and they can just drop a fluid in there like that.

That's about all I got.

1) The polycount is shockingly low in this particular demo. Just look closely at the frame buffer grabs provided by NVIDIA or even the first wireframe shots posted in the Wired unveiling article a few weeksago.

2) Indeed, texture quality is also surprisingly low.

3) Jaggies are everywhere is you watch the firect feed video provided by NVIDIA (.MOV file on Geforce.com) and the original non shopped screenshots.

4) The particle system is all over the place IMO. First it's fairly apparent that what was shown was nothing more than NVIDIA's APEX Turbelence (http://developer.nvidia.com/apex-turbulence) which was alreacy showcased at GDC 2012 in UE3. The other physics based effects where also done via PhysX according to Tim Swenney. So no really Epic tech here it seems unfortunitelly. You will also notice that ther's no AO (Ambiant Occlusion) at all in the demo. Seems like lights emmited by the particles have no effect on the GI solution right now.

5) Once again: most probably NVIDIA PhysX.

[Angelcurio]

Originally Posted by Davros

Didnt want to start a new thread for this :

This video is from a Square Enix Real Time tech demo to demonstrate the capabilities of their new, "Luminous," engine.

The red particles look somewhat akward on that video, specifically in the 0:44 -0:47 mark. Aside from that, i think it looks really good.

[milk]

Originally Posted by Ike Turner

1) The polycount is shockingly low in this particular demo. Just look closely at the frame buffer grabs provided by NVIDIA or even the first wireframe shots posted in the Wired unveiling article a few weeksago.

As I said, that was a disappointment to me, but on another watching today, I think there might be some displacement mapping going on in the lava which seems to have true geometric features, so maybe there it is. Again epic is using tessellation on things where artifacts end up more unnoticed.

Originally Posted by Ike Turner

You will also notice that ther's no AO (Ambiant Occlusion) at all in the demo.

Well, there is no actual "Ambient Occlusion" to speak of, but their voxel GI is naturally occludes their dynamic ambient light bounces, so you could say they do have some low frequency world space directional ambient occlusion. But I do believe adding a good HD SSDO solution there to complement their real time radiosity would have increased the lighting quality in that demo a good notch. The same way crytek's SSAO turns their even lower frequency light propagation volumes (three cascades of a 32x32 grid around the camera) into something that truly fits in well with the scene.
If well done it could make it hard to tell when the high detail screen space trick ends and the softer world space correct occlusion begins.

[jlippo]

Originally Posted by doob

Now how did they achieve it at acceptable performance (guessing around 30fps)? Maybe by trading off precision/quality for performance. The other presentation on NVidia does mention when talking about the glossy reflections that lowering the resolution/precision/ammount of specular reflection cones gives performance increase. Perhaps that was one way they achieved it for UE4 elemental demo.

Remember that the voxel structure is pre-filtered.
If the cone is wide they do not have to trace so deep in the voxel tree and sample as much, so it ends up faster than tracing a tight cone.

[Bumpyride]

Originally Posted by milk

As I said, that was a disappointment to me, but on another watching today, I think there might be some displacement mapping going on in the lava which seems to have true geometric features, so maybe there it is. Again epic is using tessellation on things where artifacts end up more unnoticed.



Well, there is no actual "Ambient Occlusion" to speak of, but their voxel GI is naturally occludes their dynamic ambient light bounces, so you could say they do have some low frequency world space directional ambient occlusion. But I do believe adding a good HD SSDO solution there to complement their real time radiosity would have increased the lighting quality in that demo a good notch. The same way crytek's SSAO turns their even lower frequency light propagation volumes (three cascades of a 32x32 grid around the camera) into something that truly fits in well with the scene.
If well done it could make it hard to tell when the high detail screen space trick ends and the softer world space correct occlusion begins.

I agree. The apparent lack of AO in tight spaces (like the base of the structure in the 'planets' room) suggests that they may be using a really coarse voxelized representation for the GI. If that's the case, it would benefit from some SSAO of some kind or another to fill in the high frequency details. The base of that structure in particular looks jarring to me.

The geometry detail looks alright to me. It's obviously selective displacement mapping though. Such as the main character verses the bricks on the stairs.

[Bumpyride]

Originally Posted by Ike Turner

1) The polycount is shockingly low in this particular demo. Just look closely at the frame buffer grabs provided by NVIDIA or even the first wireframe shots posted in the Wired unveiling article a few weeksago.

2) Indeed, texture quality is also surprisingly low.

3) Jaggies are everywhere is you watch the firect feed video provided by NVIDIA (.MOV file on Geforce.com) and the original non shopped screenshots.

4) The particle system is all over the place IMO. First it's fairly apparent that what was shown was nothing more than NVIDIA's APEX Turbelence (http://developer.nvidia.com/apex-turbulence) which was alreacy showcased at GDC 2012 in UE3. The other physics based effects where also done via PhysX according to Tim Swenney. So no really Epic tech here it seems unfortunitelly. You will also notice that ther's no AO (Ambiant Occlusion) at all in the demo. Seems like lights emmited by the particles have no effect on the GI solution right now.

5) Once again: most probably NVIDIA PhysX.

Thanks for the turbulence links. I agree with you, that is what it looks like. I wouldn't expect them to use each particle as an emitter but putting a colored line source light along the particles' vector field attractors (what they look like to me, not sure what they're really called) would be an easy solution that would probably fit with the voxel lighting system.

I don't think it's a major knock against them that these are NVIDIA technologies. They're pretty cool and it's nice to see it integrated into unreal engine.

I'm disappointed to hear about the jaggies, though. I was hoping that would be the first place we would start with graphics this generation.

[Arnage]

This might explain the current lack of AA:

Originally Posted by Timothy Lottes on his blog

FYI TXAA is not in any of the screens or videos for UE3 or UE4. But you will be seeing TXAA soon...

Source

[Bryant]

What bothers me more than the low texture resolution is the fact that it's 2012 and we're still using linear interpolation for texture filtering.

[Andrew Lauritzen]

Originally Posted by Bryant

What bothers me more than the low texture resolution is the fact that it's 2012 and we're still using linear interpolation for texture filtering.

What do you want to be using? Bicubic? That's obviously only going to make much difference for mag filtering, which is only relevant when you *are* using low texture resolution.