lowelllynnduke

Hello Everyone,

This is a fairly generic question, and I promise it may sound like a hardward question at first, but it's not really. Does anyone else notice how slow some of the DX10 samples in the Feb SDK are Specifically those that employ the geometry shader such as the displacement mapped lizard and the example of how to do cube mapping in one pass (when you check "Render Car") They are dramatically slower than other DX10 and DX9 samples. I was shocked when I ran them.

Do you think this is because the DX10 SDK/API is in it's infancy Or the hardware/drivers are still too young I'd ask this in a hardware forum but most of those guys don't know anything specific to the SDK or programming, just the buzz words and marketing numbers that come from Nvidia.

Both of these samples run at about 5-10 FPS on my machine. Not to turn this into a hardware question by any means, but my system is a brand new dual core w/8800 GTX and 100.64 drivers. I expected the GS pipe to be much faster.

The software side of this question is that, if this is the expected performance of the Geometry Shader pipe right now, I need to avoid it in my software projects until it matures. My DX9 code to render a 6 pass cube map blows the pants off of this GS one pass technique and the one pass is touted as being more efficient. :\ I'm throughly confused.

I plan on implementing a few shaders that employ the GS and I'll do my own profiling then, but in the meantime I haven't found a single person/thread/forum/site mention a similar experience.

See, I told you this is really a software question with a few hardware keywords :)

Thanks,




Re: Direct3D 10 Geometry Shader Performance

O.Nyman

From my limited experiments I have determined that the compiler doesn't optimize GS code yet, at least not very well. Simply by unrolling loops one test case went from 60fps to 500fps. Also the accuracy of the maxvertexcount setting affects performance significantly, but I'm not sure if this is something that will change over time.




Re: Direct3D 10 Geometry Shader Performance

Raev

You're not alone ;) - I get similar results on 8800 GTX. The 1 pass render to cube map, displacement mapping, and GS motion blur (with the jumping orc) samples are very slow compared to others. I get the impression some parts of the pipeline are done in software in these demos.

The drivers do seem very young. Whenever I do a logic mistake in my code (like forgetting to set proper input layout before a draw call), the drivers hang up and OS restarts them as they "stop responding" with a screen going blank for a second ;).

Well, at least you get a feeling of being DX10 pioneer in exchange ;]





Re: Direct3D 10 Geometry Shader Performance

lowelllynnduke

Thanks for the responses. That's good to know. :) Yes the GS motion blur is slow as well. Plus, I get the exact same thing - one wrong code move and my screen goes blank and a Vista message says that the Nvidia driver has been restarted. LOL. Well, good then. Thanks guys. I'll probably shy away from the GS for the time being, I've got a bunch of DX9 stuff still to port so I'll revisit it later.

Happy pioneering!

-L






Re: Direct3D 10 Geometry Shader Performance

WolfgangEngel

there is also a chance that first-gen DX10 hardware does not offer enough performance for the geometry shader :-) ... so we have to wait until the second generation is out.
This will work in the same way as with DX 9. Everyone has a graphics card ID database and we will switch on and off DX10 features depending on what the graphics card supports. A cap bits system would help a bit here, because than we could track new cards easier. With the ID system we have to constantly patch as long as a new graphics card comes to market and at some point the game won't run anymore with the latest and greatest graphics cards because we gave up patching.





Re: Direct3D 10 Geometry Shader Performance

lunatic

did anyone of you tryied the shader workshop
in my case the geometry shader volume shadows just do not work
the adjacent vertices are just wrong




Re: Direct3D 10 Geometry Shader Performance

TiPiou

same here lunatic.
I don't know if their adjacency generation algorithm is in cause... since the cast shadows example with the dwarf in armor works fine... but this tower in the workshop doesn't have correct adjacency information.




Re: Direct3D 10 Geometry Shader Performance

lunatic

thanks
nice to read, i'm not the only one
i allready thought i could be the card