FarCry Performance Revisited: ATI Strikes Back with Sh 2b

[Luminescent]

Interesting post, whql, although I'm not sure if the game Deano was referring to is pixel shader limited. Deano followed up with his initial post with the following claim:

Its probably more our fault than nVidia's to be honest, we built it pretty much on 9700 and 9800, so its not surprising its runs better on an X800. The NV40 path is nowhere as optimized at the moment. We seem to encounter a fair few bugs on the NV40 that cause us to run slower code paths.

I was referring to pixel shader performance, specifically, which is generally determined by specialized benchmarks. I do know that anisotropic filtering changes things a little, being that R420 has independent texture address and sampling processors.

 

[wzh100]

i don't think the shader model 3.0 is too much faster then shader model 2.0b.

 

[Luminescent]

The difference between SM 3.0 and 2.0b, assuming both paths are used for performance optimization, is minimal, if Far Cry 1.2 benchmarks are anything to go by.

 

[Mintmaster]

When shadow mapping with multiple jittered samples using bilinear filtered PCF versus point sampled PCF makes surprisingly little difference in appearance. Here are some images from some tests I did.

<snip>

Before taking a look at the differences I was inclined to think that using bilinear filtered PCF on the jittered samples would look significantly better and that it was a mistake on ATI's part to not include hardware support for it. Now, I am not so sure that bilinear filtered PCF is very useful because it does not make any visible difference when using jittered sampling and because shadows maps sampled with just bilinear PCF look ugly in comparison to ones sampled with jittered PCF.

Thanks for the pics! Looks pretty good.

That's what I figured it would look like. My guess is that Carmack wants to make the jittering at least a few texels wide because an ugly artifact is the stepping/crawling of shadow edges for slowly moving objects when the shadow map resolution isn't too high. This way the stepping will only affect one sample.

It might make more of a difference when there's only 4 samples, though. Do you have any results for that?