Support NeoGAF

[mrklaw]

Not sure how important the z-only pass is, but that's one of the few times you'll see all 48 ALUs running as VS. .

Its potentially a huge advantage for Xenos, if RSX can't do something similar.

It effectively lets you calculate occluded polygons, so you only have to pass visible pixels to the shaders.

So if you had a car driving in front of a wall, it doesn't have to render the wall pixels.

Depending how good normal culling is, that could save enough pixels to effectively be more powerful than RSX in pixel shading - eg 30% slower, but draws 40% less pixels.

 

[kaching]

Do they really want games early in development to be seen or would a smarter strategy be to leave those thoughts of Killzone and Motorstorm in peoples heads with the promise that they will deliver to keep them from getting a 360. Now if they can get a level running in realtime of Killzone that looks just like the e3 presentation then this is all a moot point.

It's not as if KZ and Motorstorm were shown by themselves, that that's all we've seen of the PS3 projects so far. At the very same press conference they already DID show less impressive looking content (in terms of scope and overall pushing the visual envelope), such as Eyedentity, I8, the GT next footage, etc.

 

[TristenCanSing]

Hey, you're right!

Haha, that's neat! Did you make that yourself?

 

[Pimpwerx]

Its potentially a huge advantage for Xenos, if RSX can't do something similar.

It effectively lets you calculate occluded polygons, so you only have to pass visible pixels to the shaders.

So if you had a car driving in front of a wall, it doesn't have to render the wall pixels.

Depending how good normal culling is, that could save enough pixels to effectively be more powerful than RSX in pixel shading - eg 30% slower, but draws 40% less pixels.

All next-gen systems should be able to do this. It's a function of VS power. Xenos has an advantage in that its VS throughput when running all 48 ALUs in VS mode should be higher than the 8 VS in RSX. This is ignoring Cell, just comparing the GPUs. But there's no reason you couldn't do this on Cell, which is much better equipped for this than either RSX or Xenos.

Xenos will close the gap through efficiency, not b/c it can do anything drastically different than RSX. One thing it might have over RSX is a tesselation unit. Didn't read anything about that in G70, so RSX might be lacking it too. The chips will have more in common than not. Like RSX may or may not be able to support fp16/fp32 efficiently, but Xenos can still do fp10, and maybe fp16 with some compromises (not sure how it effects throughput/tiles).

I wish the companies would go ahead and reveal all the info up-front, b/c there are some things like SSS and "radiosity" on G70 that weren't explained in enough detail, and the PPP and HDR on Xenos weren't explained well either. PEACE.

 

[Lord Error]

Its potentially a huge advantage for Xenos, if RSX can't do something similar.

Z-only pass, according to people at B3D is trivial, and not vertex shader heavy operation at all (which surprised me as well).

 

[mrklaw]

All next-gen systems should be able to do this. It's a function of VS power. Xenos has an advantage in that its VS throughput when running all 48 ALUs in VS mode should be higher than the 8 VS in RSX. .

exactly. Xenos is shat on in almost all other stats. But in the one example where you would use all VS setup, it is significantly better than RSX. I guess you're right though that in this case CELL could perform the same task.

 

[dorio]

Z-only pass, according to people at B3D is trivial, and not vertex shader heavy operation at all (which surprised me as well).

If its so trivial I don't see why everyone isn't doing it.

 

[Nostromo]

It's trivial! you don't need a NG GPU to do it, you can do it even on 3DFX voodoo.
First pass you draw everything (but non opaque stuff) only on the zbuffer, second pass you re-submit all the geometry setting ztest as EQUAL e you render everything.
This way one can completely avoid opaque ovedraw cause only incoming pixels who have a z coordinate equal to their counterpart on the z-buffer will pass (and will be shaded)
Rendering everything in 2 passes has a cost and it makes sense only if pixel shading cost is very high and you don't want to shade a lot of unseen (due to occlusion) pixels, that's why most games don't use it
Doom3 adopts the same rendering scheme, but there are other games that use the same technique (anyway it's a well known rendering scheme in the field, Carmack has not invented it).
NG games would probably use the same technique since we'd expect a very heavy use of shaders.
Xenos has special support for it and during the first pass it also tags rendering commands in the display list with tiles tags, so during the second 'real' rendering pass only triangles batches that lie in the current tile are resubmitted thus savind bandwith and ALUs cycles

 

[OmniGamer]

:lol

Finally, it got a laugh out of someone!

 

[dorio]

Rendering everything in 2 passes has a cost and it makes sense only if pixel shading cost is very high and you don't want to shade a lot of unseen (due to occlusion) pixels, that's why most games don't use it

So it's not trivial since it requires 2 passes.

 

[Nostromo]

So it's not trivial since it requires 2 passes.

From a games programmer point of view is trivial, you have just to draw everything two times
If you can do it one time..you can do it two times too, it's very simple, any 3d engine coder out there can confirm this

 

[dorio]

From a games programmer point of view is trivial, you have just to draw everything two times
If you can do it one time..you can do it two times too, it's very simple, any 3d engine coder out there can confirm this

Having to render every vertex twice definitely isn't trivial for the hardware.

 

[Nostromo]

That's why I've said it has a cost
Since the cost to vertex shade a scene 2 times should be (if we talk about NG games) vastly inferior to the cost to shade the same pixels MANY times if we're using long pixel shaders, to obtain better performance games developers HAVE to make this choice.
In fact one the first games that makes a very heavy use of shaders to use this technique is Doom3, this is not by any meaning a coincidence

 

[Lord Error]

Actually in this very thread (re: z-pass):

99% Trivial transforms, generally completely vertex setup limited. Xenos or not, this is just an example of a situation where most shader ALUs will be sitting idle.

 

[dorio]

Actually in this very thread (re: z-pass):

99% Trivial transforms, generally completely vertex setup limited. Xenos or not, this is just an example of a situation where most shader ALUs will be sitting idle.

Xenos doesn't have pixel shaders, they're a chameleon. They can be anything you want them to be. If there are free shaders left over during the z-pass then they will be used to render the pixels from the upcoming frame.

 

[gofreak]

Sort of related: some B3D guys snooped some work Sony is doing on improving Cg for PS3.

http://alumnus.caltech.edu/~heirich/activities.htm#Computer Graphics and Gaming
For those who like white papers: http://alumnus.caltech.edu/~heirich/eghw05_1003.pdf

It's pretty technical, I won't get a chance to read till tomorrow, but may be of interest to some here.

 

[Nostromo]

some B3D guys snooped some work Sony

LOL..that's me! If ony GAF admins would let me use my old nickname..

About Xenos ALUs..they could stall too in the first pass if the rasterizer can't keep up with the tremendous number of transformed vertices (thus primitives) ALUs can shade.
Xenos architecture removes some bottleneck, not every bottleneck you can found in a GPU