PVR voxel PS_3 demo on ati and nvidia hardware, useful for benchmarking?

[PP]

I was testing this demo on a 7800 GT and X1800 XL, and found that ATI hardware is much faster on almost all render methods, particullary lit1 & 3 (like 1:3 proportion). Do you find it useful like synthetic bench, or could be misleading?

 

[cloudscapes]

I can't get it to run at all. Tells me I either need a 3.0 PS card or to enable rasterizing (which also doesn't work).

X850 XT with latest Cats.

EDIT: just read you had a X1800 and not an X800 like I previously thought I read.

 

[PP]

Yes, it's for PS3 boards only.

I'm using FW 78.01 for GT and Cat. 8.17 for XL. I'd like to use a newer ATI driver, but included one was that. I saw version 8.18 on station-drivers, but it's modded with a custom .inf and it doesn't work.

And well, does someone knows why this demo is so hard for the 7800 GT?

 

[Tridam]

If I remember well enough the demo sources (I looked at it right after the 6800 launch, but not tried with X1k), it uses complex shaders with loops, calls, breaks and branches and multiple levels of these. Stuff like

loop
   if
      if
         call
      endif
      if
         break
      endif
   endif
endloop

GeForce 6 and 7 should start crying like babies with such shaders

 

[PP]

Oh, that explains a lot I'll use it in my review, don't know if it'll be representative of future SM3 games, but seems fair if performance loss is caused by that. I wondered if it was some kind of driver bug, even thinking nvidia had 1,5 years to tweak them.

 

[trinibwoy]

Hehe - maybe there's something to ATi's "SM3.0 done right!" slogan :smile:

 

[Ailuros]

It's a shame R520 can't run the Cloth demo calling for vertex texturing; it also has pretty complex shaders and dynamic branching.

 

[Jawed]

I guess the cloth demo is an ideal opportunity for Humus to write a R2VB version - should it be possible

And while we're discussing the cloth demo, why is the cloth simulation being run in pixel shaders rather than vertex shaders? Is it because there are more pixel pipelines (i.e. it'll run faster)?

Jawed

 

[Ailuros]

And while we're discussing the cloth demo, why is the cloth simulation being run in pixel shaders rather than vertex shaders?

Where does that one come from?

 

[Jawed]

Where does that one come from?

From reading the shader code.

The actual "physics" simulation of the cloth is in pixel shaders, not vertex shaders.

e.g. PSClothPhase-24OPT.psh

Jawed

 

[Ailuros]

Don't you think the PS would be pretty "helpless" if it wouldn't setup first the texture coordinates in VS3.0 during the simulation phase?

 

[akira888]

And while we're discussing the cloth demo, why is the cloth simulation being run in pixel shaders rather than vertex shaders? Is it because there are more pixel pipelines (i.e. it'll run faster)?

Because IIRC it encodes the cloth's current physical state into two textures that store the positions and normals for a given number of points on a grid, and the cloth simulation fragment shader samples the values of the proximate neighbors of the current point being recalculated, determines the new position and normal of that point, and then outputs this into two render targets that represent the updated state of the cloth.

The fact that the bulk of the calculation is done at the fragment level seems, to me, to be determined more by the inherent structure of the algorithm rather than any speed considerations. I really can't see a sensible way to rewrite the demo to run predominately vertex-side, but that may have more do with my own imbecility than anything else.

 

[Humus]

And while we're discussing the cloth demo, why is the cloth simulation being run in pixel shaders rather than vertex shaders? Is it because there are more pixel pipelines (i.e. it'll run faster)?

Mainly because vertex shaders has no means for storing its results. It has to pass through the pixel shader anyway. So it makes more sense to do the job there directly.

 

[Matasar]

Do you have some numbers how much faster the XL are ?