ATi is ch**t**g in Filtering

[Exxtreme]

The 9800XT does not display this behaviour, as the computerbase benchmarks clearly show.

The R300, R350 and R360 are not capable to do brilinear filtering. This *cough* feature *cough* needs special hardware support.

 

[Bjorn]

I'm just glad nvidia seems to have learned from there past mistakes.
The NV40 looks to give us what we ask for.
Top notch performance and top IQ to boot.
Nvidia has the right idea now to just give us the power to lower
IQ for better performance.
I'm thinking I'm going with the 6800GT.
NV nor ATI is getting me off 5 bills this go around.
I'm going for bang for the buck.

I can't fully agree with that. If Nvidia would have had an option for disabling the angle dependant AF then yes. But as it is, they have lower AF quality on the NV4X then on the GF4. A 3+ old card.

But ok, at least they took a step in the right direction with the NV4X. Seems like Ati did the other way around.

 

[Dave Baumann]

Isn't that a non objective? If you don't have full tri in front of you, you can't compare. If you don't know, how can you complain?

People have noticed the difference between FX's filtering and the filtering on their previous boards - you always have you last card to compare against.

Now, do you, DB, see the difference? If yes, that's relevant.

Can't say that I have yet. As I said, I'll have to look on some of the UT2003 maps that I've used previously to have a look again.

Anyway, just had a call from ATI Europe asking about this, they are taking it seriously. You probably won't hear anything into Canada/Santa Clara look into it though.

 

[Borsti]

Well, there´s something else going on here. I was running a comparison of normal/colored with the X800 XT in UT2004. 8xAF was forced by control panel. As you know, in this mode, X800 is filtering pure bilinear here!

Here´re the results:

Normal/ColoredMip (8xAF by Driver - filters Colored mipmaps bilinear)

10x7: 132,68 / 132,90
12x10: 131,06 / 125,46
16x12: 118,42 / 99,56

Normal/ColoredMip (8xAF by App - filters colored mipmaps trilinear)

10x7: 132,83 / 131,69
12x10: 128,98 / 112,63
16x12: 107,37 / 84,48

(Don´t wonder about the higher numbers than before, but I changed the testbed system from P4 to FX-51.)

You can see that the results are lower even when pure bilinear is used (by CP)!! So the difference in performance can´t be a result from going from Tril to Bril. I don´t say Computerbase did something wrong but there seems to be more here. As the guy from Epic said, there should´nt be a performance difference at all when using colored mipmaps and now we see it even with bilinear!?!? This can´t be a result of a change in the filtering.

PS: I did not do any changes to the texture stages in these tests.

Lars - THG

 

[bloodbob]

Well, there´s something else going on here. I was running a comparison of normal/colored with the X800 XT in UT2004. 8xAF was forced by control panel. As you know, in this mode, X800 is filtering pure bilinear here!

So you selected Preformance AF on the control panel?? ( Quality AF does trilinear on teh first stage still ).

 

[Exxtreme]

Borsti,

the reason for the differences between AF via CP and AF via application is that the AF via CP doesn't activate trilinear AF for all texture stages. Only textures on stage 0 are filtered with trilinear AF. The other textures will be filtered with bilinear AF.

 

[Veridian3]

Its also worth chainging the device ID to see if it behaves differently...

 

[Evildeus]

People have noticed the difference between FX's filtering and the filtering on their previous boards - you always have you last card to compare against.

Well, did they use AF before, or even now? 9600 are mid-range owners, i don't think they would go from high to middle, so were they? Is this optimisation still used in 9600s btw?

Can't say that I have yet. As I said, I'll have to look on some of the UT2003 maps that I've used previously to have a look again.

If i look at your SS, there's not much differencies between the X800/59**, how come people clearly see the differencies with GFFX and not X800? Did you, DB, see the differencies using GFFX?

Anyway, just had a call from ATI Europe asking about this, they are taking it seriously. You probably won't hear anything into Canada/Santa Clara look into it though.

Of course, they are using tricks, oups sorry "bugs" to cover on what the hardware does indeed, why wouldn't they take that seriously? I'm sure we won't see some official before some time.

But as the X800 has been during a long time labelled a 4*9600, don't you think it's a correct behaviour from the X800? I mean it doesn't really matter if it's full tri or bri if it's labelled as so. If the X800 is a big derivative of the 9600s, i would expect the bri to be the official filtering from the x800s.

 

[Borsti]

Well, there´s something else going on here. I was running a comparison of normal/colored with the X800 XT in UT2004. 8xAF was forced by control panel. As you know, in this mode, X800 is filtering pure bilinear here!

So you selected Preformance AF on the control panel?? ( Quality AF does trilinear on teh first stage still ).

No, I selected Quality. But the ATI driver filters the colored mipmaps bilinear anyway in UT03/04 when AF is forced only by the CP. But I can do another run to see if there´s also a difference when using performance.

PS: I did not wonder about the difference between AF by driver and app but on that there´s even a difference when using AF by driver (because the colored-mips are already bilinear in this mode). So where should the perf-difference between colored/non-colored mipmaps come from?

Lars - THG

 

[bloodbob]

Well, there´s something else going on here. I was running a comparison of normal/colored with the X800 XT in UT2004. 8xAF was forced by control panel. As you know, in this mode, X800 is filtering pure bilinear here!

So you selected Preformance AF on the control panel?? ( Quality AF does trilinear on teh first stage still ).

No, I selected Quality. But the ATI driver filters bilinear anyway in UT03/04 when AF is forced only by the CP.

Lars - THG

Lars your wrong It does bilinear filter on tex stage 1-7 but does trilinear on stage 0.

PS. I believe you can specify which layer you whish to do the colour mip-maping on in UT2K4. Anyway download either the D3D or OGL AF tester FORCE AF to 16X or whatever with quality and check out the difference between texture stages.

Read here if you don't believe me http://www.beyond3d.com/forum/viewtopic.php?p=274841#274841

 

[tEd]

Well, there´s something else going on here. I was running a comparison of normal/colored with the X800 XT in UT2004. 8xAF was forced by control panel. As you know, in this mode, X800 is filtering pure bilinear here!

Here´re the results:

Normal/ColoredMip (8xAF by Driver - filters Colored mipmaps bilinear)

10x7: 132,68 / 132,90
12x10: 131,06 / 125,46
16x12: 118,42 / 99,56

Normal/ColoredMip (8xAF by App - filters colored mipmaps trilinear)

10x7: 132,83 / 131,69
12x10: 128,98 / 112,63
16x12: 107,37 / 84,48

(Don´t wonder about the higher numbers than before, but I changed the testbed system from P4 to FX-51.)

You can see that the results are lower even when pure bilinear is used (by CP)!! So the difference in performance can´t be a result from going from Tril to Bril. I don´t say Computerbase did something wrong but there seems to be more here. As the guy from Epic said, there should´nt be a performance difference at all when using colored mipmaps and now we see it even with bilinear!?!? This can´t be a result of a change in the filtering.

PS: I did not do any changes to the texture stages in these tests.

Lars - THG

with CP you mean perf. or quality?

the results doesn't surprise me much

Using CP 8x AF quality mode:

normal rendering
1stage brilinear all other stages bilinear

colored mipmap rendering
1stage trilinear all other stages bilinear

Using App. 8x AF:

normal rendering
All stages briliniear

colored mipmap rendering
all stages trilinear

 

[Hanners]

But as the X800 has been during a long time labelled a 4*9600, don't you think it's a correct behaviour from the X800? I mean it doesn't really matter if it's full tri or bri if it's labelled as so. If the X800 is a big derivative of the 9600s, i would expect the bri to be the official filtering from the x800s.

'Brilinear' wasn't the default filtering mode for the 9600 if the texture slider was set to its highest quality, you had to dial it down a notch to use it if I remember correctly. I would have expected the X800 to behave the same.

 

[Dave Baumann]

Well, did they use AF before, or even now? 9600 are mid-range owners, i don't think they would go from high to middle, so were they? Is this optimisation still used in 9600s btw?

Gah! This is not an AF issue, this is a study on Trilinear filering.

If i look at your SS, there's not much differencies between the X800/59**, how come people clearly see the differencies with GFFX and not X800? Did you, DB, see the differencies using GFFX?

The SS:SE maps are easy to use for bit comparison images because they are saved and its easy to alter the setting in SS:SE, but they are not the best for detecting this type of thing by eyeballing it.

Of course, they are using tricks, oups sorry "bugs" to cover on what the hardware does indeed, why wouldn't they take that seriously? I'm sure we won't see some official before some time.

Well, at the moment, they (PR) don't know what it is - I was explaining to them the apparent issue.

But as the X800 has been during a long time labelled a 4*9600, don't you think it's a correct behaviour from the X800? I mean it doesn't really matter if it's full tri or bri if it's labelled as so. If the X800 is a big derivative of the 9600s, i would expect the bri to be the official filtering from the x800s.

R420 isn't "4 9600's" since there are large chunks of the pipeline that are in R420 but not is RV3x0. The texture kernals appear to be the same but then there are some fixes in RV3x0 that weren't in R3x0 (such as the AF LOD bug that manifested itself in UT2003), as well as the "Brilinear" capabilities.

 

[Borsti]

Ok, it seems I need some more coffee this morning

Here are new numbers.

AF forced by driver. 8x AF, Performance (this should result in pure bilinear on all stages)

10x7: 132,85 / 132,84
12x10: 132,12 / 131,08
16x12: 124,68 / 113,85

Now same setting with changed Mipmap Detail Level from H-Quality to High Performance)

10x7: 132,54 / 132,71
12x10: 131,78 / 131,15
16x12: 129,9 / 123,70

The last on with changing Textures from H-Quality to High Performance (Mipmal details back to default)

10x7: 131,22 / 131,92
12x10: 131,18 / 130,69
16x12: 127,84 / 117,41

Well the whole discussion started because the guys at Computerbase found out that the X800 is loosing performance when colored mipmaps are used. The suspicion is that ATI is using a lower filtering by default and switches to full tril only when colored mipmaps are used.

This can´t be the reason since there´s even a performace difference when you´re already using pure bilinear by default!?!? And we have the quote from Epic that there should´nt be a difference at all. Maybe it´s because the lack of coffee this morning but I´m getting a bit confused

Lars - THG

 

[dan2097]

Borsti: Looks like the mystery deepens 8)

 

[Evildeus]

'Brilinear' wasn't the default filtering mode for the 9600 if the texture slider was set to its highest quality, you had to dial it down a notch to use it if I remember correctly. I would have expected the X800 to behave the same.

Ok thanks Hanners.

 

[Evildeus]

Gah! This is not an AF issue, this is a study on Trilinear filering.

Yeah sorry, confusion
But Hanners post just confirms that there's no reason that people should complain, as they need to change the default filtering mode.

The SS:SE maps are easy to use for bit comparison images because they are saved and its easy to alter the setting in SS:SE, but they are not the best for detecting this type of thing by eyeballing it.

Ok, thanks for the explanation.

Well, at the moment, they (PR) don't know what it is - I was explaining to them the apparent issue.

Well, PR people don't know this kind of issue, of course.

R420 isn't "4 9600's" since there are large chunks of the pipeline that are in R420 but not is RV3x0. The texture kernals appear to be the same but then there are some fixes in RV3x0 that weren't in R3x0 (such as the AF LOD bug that manifested itself in UT2003), as well as the "Brilinear" capabilities.

I didn't say it is, i said that it seems to derivate from.

 

[Marc]

Well the whole discussion started because the guys at Computerbase found out that the X800 is loosing performance when colored mipmaps are used.

And the discussion still going on because there is a real difference in the trilinear :/

Cand you do the same test on R3xx ?

And take a look at the differences beetween R420 8x AF Performance shoot and R3xx same shot via compressonator ?

 

[mikechai]

Could it be the driver detects when colored mipmap is used and force trilinear no matter what the CP setup or application request?
________
Fake weed

 

[Hanners]

Yeah sorry, confusion
But Hanners post just confirms that there's no reason that people should complain, as they need to change the default filtering mode.

On the 9600 yes, but I'm assuming from the shots posted here that it seems to be the default mode on X800 boards, hence the complaints. I should have made that bit clearer in my last post.