ATi is ch**t**g in Filtering

Does brilinear cause that 'crawling' you get on horizontal textures(Most noticable) in a game when you are running? It's hard to accurately describe. It's as if some are focused and some are not, or some suddenly come into focus.

 

Well, I´ve received some additional information from ID (JC) about the Q3 engine and colored mipmaps:

This is indeed a "cheat" that both major vendors now do. Instead of always
sampling the two adjacent mip map levels and doing a full blend between
them, they have plateaus where only a single mip level is sampled, reducing
the average samples from 8 to about 6.

It is actually a pretty sensible performance enhancement, with minimal
visual issues. However, having drivers analyze mip map uploads to hide the
cheat is an unfortunate consequence.

The colored mip map option in Q3 should have absolutely zero performance
impact in the absense of performance options like this.

I´ve also run some benches on NV40 in Call Of Duty to see what´s happening there (with the P4 system again... ). 8xAF by driver.

NV40 Norm/Colored (with TrilOpt On)
16x12: 108,9 / 108,7

NV40 Norm/Colored (with TrilOpt Off)
16x12: 82,6 / 82,4

X800 XT Norm/Colored
16x12: 96,3 / 84,8

The Computerbase story started with the COD benches. Now we got the note from ID that there should be no difference in Q3 (so I assume this is also the same for COD). But regarding the additional comments from JC there seems to be much more going on...

Lars - THG

 

That's the third time you've said that in this thread alone.You're starting to sound like a broken record. I got it the first time thanks.

 

As I mentioned - the information within the texture can affect how it can be sampled, colouring each mipmap level changes the sampled information and can potentially change the behaviour.

 

Well I spent 10 minutes flipping the two images back and forth focusing on the plates you outlined, I can say they are different, but I don't know that I can say one is better than the other.

It's certainly nowhere near as noticeable as the brilinear used by the FX cards in UT2k3, at least in a screenshot.

 

I would like to see a shot from an FX board in the same position.

 

But this would mean that ATI is doing Brilinear!? In full tilinear there should´ne be different behaviours or what do you mean?

Demirug at 3DC also did some comparisons with and without coloredmip screenshots in UT. There are enough areas where the textures appear normal and there is a difference (the bril bands) between the X800 shots. This behaviour does not appear at NV40 shots.

Lars

 

Feel free to serve yourself.
edit: clickable smiley
edit2: 'Twas not my idea, to call this demo 'cheat'. Leonidas over at 3DCenter did it, when he documented brilinear on the FX ;D

But, as someone over at R3D said to Hellbender: This is not a cross-vendor issue.
edit: I won't say, your like HB, though!

 

ati and nvidia both cheat, its pretty simple. i just hope ati will stop here, and not go any further

 

That's exactly what i meant, when i said i wouldn't judge in anyones stead.

 

You're assuming it is a cheat.

I have yet to see any screen shots from games proving there is a cheat.

Why is that?

 

The problem is that you can´t show lowered filtering quality by using colored mipmaps - because the guess is that this mode is beeing detected and served with higher quality.

Of course we can discuss if it´s OK when it´s not vissible. But it´s worth finding out what the driver does. That´s what is going on at the moment.

Lars - THG

 

If anything is being "detected" then I believe its the information that requires sampling - I don't believe the driver is inherantly changing behaviour when mip map colouring is on, but the coloured mips have different sampling requirements.

FYI - WHQL DCT does have tests for trilinear filtering.

 

3dFX mip-map dithering?

 

Hm.. this would make sense. But in COD there´s no performance loss with colored mipmaps with NV40, the X800 XT looses pretty much 11 FPS. So I´m curious on the official explanation. Could it be something that hits ATI only in general, like more alpha test passes in UT?

Lars

 

Would this test be detectable? I have been wondering if this is true than how come it has eluded M$'s WHQL

 

I keep seeing the 'C' word thrown about, and as far as I am concerned it's much too early to say that for sure.

As far as I am concerned, one of the major differences between a "cheat" and an "optimization" is openness. For example, "brilinear" is a cheat when you call it 'trilinear' and it's an optimization when you call it "brilinear." It may or may not be a valid optimization, but until we have evidence of deception it's not a cheat.

I have yet to see anything from ATI regarding this anomaly. Until I do I can't call it a cheat or an optimization or even a bug. Something is definitely happening, but right now we don't know exactly what or why, and until we do we should avoid loaded terms.

 

DrawPim already mentioned that one, but came to a different conclusion as to whether the content to be filtered is influencing the behaviour of texture filtering applied.

I guess i need a good night's sleep before i might think of all the consequences this post could imply in a malicious mind.

 

What if your driver / graphics chip analyses the contents of the sampled information and makes decisions based on that - some of the times it will decide that, given the information in the sampled data, it can alter its behaviour dependant on that information and offer something that may not be "full" trilinear, but something else like "fast trilinear" (box filtering), for example; however mipmap colouring, due to the nature of the constantly varying samples, may end up a being a pathological case, constantly forcing full trilinear - hence there is a performance difference between coloured and non coloured mip maps.

Perhaps it hasn't - perhaps it within the bounds of those tests.

 

Sumarizing:

1.- The output of colored mipmaps shows that both cards are doing trilineal in the Q3 engine color test.

2.- NV40's performance does not drop with colored mipmaps, but R420's does.

This implies that, assuming JC's statement, there is not an " absense of performance options like this" in the way R420 is processing the data.

I cannot say the same for NV40 since there is no performance drop.

Just logic.

Am I right?