Archives Discussions

[hax666]

DirectX 9, we are resolving MSAA z-buffer to non-MSAA zbuffer, and then using it for rendering with RenderTarget texture. This helps us to lower fillrate during postprocessing (some of our postprocessing effects require z-compare). All surfaces have same width/height, zbuffer is not discardable, StretchRect() is called between EndScene()/BeginScene().
FPS is 3-4. Works fine on nVidia cards.
Is this driver issue, or hardware just can't handle this efficiently?

[avk]

A RV530/535 chips has very limited texture pipes - 4, which is significally lower than on many nVidia chips of the same price range. BTW, what nVidia cards you use to compare?

[hax666]

I don't see how copying between video surfaces are connected to texture pipes.

Works fine on gf6600-8800.

[avk]

I think that it will be useful to test another Radeon chips to find out the source of the problem. Who knows how exactly performs operation you mentioned...

[hax666]

Radeon 1600: FPS is 3-4, dest Zbuffer contains garbage.
Radeon 2600: FPS is 30, dest Zbuffer contains garbage.
I suppose this functionality was never tested.

[avk]

Different chips, different characteristics (pipes, cache ...). BTW, which resultions and drivers were used during testing?

[hax666]

Radeon 1600 - Catalyst 7.1 and Catalyst 8.3. Radeon 2600 - Catalyst 8.3. In any case, dest zbuffer contains garbage. How exactly resolving MSAA zbuffer should happen is not well defined in DX docs, but it is legal. Just first sample or samples average is ok for our case.

[avk]

... And what about screen resolutions?

[hax666]

1024x768x32 fullscreen, 800x600x32 in window. I think resolution does not affect this.

[avk]

Who knows... Radeon 1600 is a low-level card, so it can have a very little on-die cache, which can lead to low performance on high resolutions, especially with MSAA enabled.