ATi is ch**t**g in Filtering

[davepermen]

as far as i can see this algo does indeed replace the ordinary way to do full trilinear filtering in some cases. the only reason why the bits are different is simple: round-off-errors. this always happens if you shuffle / replace EQUAL operations in a filter. this is something we have to face all the time.

because of that, i don't see all the complains. ati found a an optimisation, where certain ops are useless for the trilinear, so it doesn't always need to do the full job (but result in the same output, except for roundoffs). this is all fine.

this is like saying a mult instruction is not allowed to simply return 0 if one of the ops is 0, because it would not calculate it then. there is no difference (in this case, not even a roundoff), so whats the mather? and if one of the ops is one, you're allowed to directly return the other op, instead of calculate *1.

this is about what this algo does (except, more complex, much more clever, and with a lot of work behind the scenes, to get it done..).

 

[bloodbob]

as far as i can see this algo does indeed replace the ordinary way to do full trilinear filtering in some cases. the only reason why the bits are different is simple: round-off-errors. this always happens if you shuffle / replace EQUAL operations in a filter. this is something we have to face all the time.

There is no comparision operations in trilinear filtering so how come we are talking about equal operations??

If I am just going insane again. Maybe you can post the pseudo code for trilinear filter as per any paper you wish to reference.

 

[christoph]

@davepermen
why do you quote me/respond totally out of context?

 

[davepermen]

first: ati states it only uses it there, where it results in (more or less) equal output. (and thats the big difference to brilinear, wich spreads everywhere..).

second: the fact that its in since over a year and nobody noted hints that this is really working (while, understandable, an unfair fact, for the ones that now feel cheated... ).

prove i'm wrong. proof there IS visibile difference between trilinear and atis "if its not visiible we can use a faster mode" filtering, a.k.a. trylinear. the only thing that got proofen is, there are measurable difference between different renderers. but as i stated, this was the case since years. so what? no renderer works the same way, none gets the same output. but its not really visible.

 

[davepermen]

@davepermen
why do you quote me/respond totally out of context?

i replied to your post right above my post.. may have missed something.. i'll check


nope, haven't seen some problem. whats wrong?

 

[bloodbob]

And why do you make one post after another instead of doing it all in one post?

 

[davepermen]

And why do you make one post after another instead of doing it all in one post?

because thats how the cookie crumbles, or what ever. want to directly personally attack now? cool.. flamewar level..

reason why i post like that? because its easier. press reply, have the quote in, reply. read next, press reply, have the quote in, reply. works quite fast.

 

[CarstenS]

first: ati states it only uses it there, where it results in (more or less) equal output. (and thats the big difference to brilinear, wich spreads everywhere..).

What if that weren't the case?

edit:

second: the fact that its in since over a year and nobody noted hints that this is really working (while, understandable, an unfair fact, for the ones that now feel cheated... ).

Well, maybe two facts could play a role here:

first: It's been field-testet in a mid-Range to low-cost Card. Traditionally, reviewers do not look as closely there as they do with the 500-Bucks-Flagships.

second: Everyone was busy looking for Cheats in nVidias line-up. ATi's had the stigma of being kind of a reference. If both shots did differ in any significant way, odds were, nVidia was blamed for rendering it wrong or "cutting corners".

 

[Hanners]

first: ati states it only uses it there, where it results in (more or less) equal output. (and thats the big difference to brilinear, wich spreads everywhere..).

What if that weren't the case?

Then hopefully people will be kind enough to provide evidence.

 

[Dave Baumann]

first: ati states it only uses it there, where it results in (more or less) equal output. (and thats the big difference to brilinear, wich spreads everywhere..).

What if that weren't the case?

As I've said a couple of time, NVIDIA's mode has improved significantly since its introduction (although, curiously, you might like to notice the effects with AF - you get more "Bilinear" areas with higher AF in the FX boards), however according to what I've heard there is no adaptivity (although there certainly may be per app tuning).

 

[davepermen]

first: ati states it only uses it there, where it results in (more or less) equal output. (and thats the big difference to brilinear, wich spreads everywhere..).

What if that weren't the case?

there are two parts of a company: pr, and the rest. its not the pr departement, that states this. the pr states its all full trilinear (wich, sort, is true, but, in pr language only with SMARTFILTER to enhance trilinear..).

the non-pr part of the company states this:

we have a faster way than trilinear, wich looks identical (escept roundoff) in SOME cases. we detect those cases, and use it instead of trilinear. if there is ANY evidence the faster way does NOT behave as expected, we move back to full ordinary trilinear, the way it was done before.

one such case, where it falls back, is coloured mipmaps. another one would be mipmaps of normalmaps (wich have other rules on how to filter them down, so apps do it on their own, if at all).

one case where they can guarantee the filterer works, is, if the app requests automatic generation of mipmaps (GL_SGIS_generate_mipmaps, i think, in opengl). there, the driver has to handle the job, and thus he knows the faster filtering can be directly applied.



i do believe in official non-pr-statements of both nvidia and ati. till now, they never proved to be wrong. (and if they would, they could even expect lawsuits.. so they don't lie normally)

 

[davepermen]

first: ati states it only uses it there, where it results in (more or less) equal output. (and thats the big difference to brilinear, wich spreads everywhere..).

What if that weren't the case?

As I've said a couple of time, NVIDIA's mode has improved significantly since its introduction (although, curiously, you might like to notice the effects with AF - you get more "Bilinear" areas with higher AF in the FX boards), however according to what I've heard there is no adaptivity (although there certainly may be per app tuning).

i never said brilinear is ugly. it's just a mather of fact that it happens everywhere (and thus, happens to get ugly in coloured mipmaps..). atis way is more smart at detecting where we there would be loss in quality if we take the faster way.

i for myself think brilinear is quite a fine thing. happens due the fact that i don't believe much in trilinear anyways..

 

[CarstenS]

first: ati states it only uses it there, where it results in (more or less) equal output. (and thats the big difference to brilinear, wich spreads everywhere..).

What if that weren't the case?

As I've said a couple of time, NVIDIA's mode has improved significantly since its introduction (although, curiously, you might like to notice the effects with AF - you get more "Bilinear" areas with higher AF in the FX boards), however according to what I've heard there is no adaptivity (although there certainly may be per app tuning).

I do not like nVidias way of not-giving-me-trilinear-when-i-asked-for-it any better nor any worse.
If i push the "add" Button on my calculator, i expect it to "add", not to guess a (granted, pretty close) approximation of the result.

 

[davepermen]

then you should never calculate on a pc (or read up on floatingpoint).

most people don't want trilinear. they just think they know trilinear is sort of a holy grail for good graphics (e-penis-enlargement as well), and thus, want it, because they want good graphics. most don't have a clue what its really about. all they want is good graphics.

but of course, the result should be, for cases where trilinear is needed, trilinear.

problem is, do you care HOW your calculator adds? do you care about the implementation? because, in atis case, THATS what we discuss. we ALWAYS have roundoff errors. this is ALWAYS the case. NOTHING renders the way the refrast renders. every aa works different, every anyso works different. why not texsamplers?

 

[croc_mak]

If i push the "add" Button on my calculator, i expect it to "add", not to guess a (granted, pretty close) approximation of the result.

This is actually a good eg:- for why some of the shader optimizations that were done last year by Nvidia are bad..

Not to insult anyone..but wrt to mipmapping and trilinear expecting anything else mathematically other than a big guess is plenty stupid.

Infact, the whole discussion is about as stupid as the Nvidia cheating on lod on Nv40 discussion a while ago...Why hasn't it occured to the "computerbase" and others that there are highly VISIBLE differences between NV3X and NV40 on mipmap boundaries and also make CLAIMs that it is cheating or some other optimization that Nvidia misled by not revealing? I'm making a bad attempt to be sarcastic here..incase you haven't noticed


Why don't everyone expect the same with FSAA?( I expect an "add" example given above)...AFAIK there is no exact equation that needs to be executed to any bit of accuracy with lod calculation either - in DirectX or Opengl

Honestly - the more I read this the more I'm convinced that this discussion appears to be based on some emotional issues with ATI than any rational issues....which is sad since it's bringing beyond3d to the levels of the rest of the forums and will result in driving away the knowledgeable and useful folks

 

[KimB]

AFAIK there is no exact equation that needs to be executed to any bit of accuracy with lod calculation either - in DirectX or Opengl

There is. It's in the spec. There are acceptable LOD calculations, and there are unacceptable ones. It's all very well-defined, though I'm a bit tired to dig up the spec myself. If you're curious, you can download the OpenGL spec from http://www.opengl.org

 

[Suspicious]

Few questions:

I am new here and before someone says I am a <f a n b o y> I currently have/use Sapphire Radeon 9600 Pro card.

Wit that said and out of the way I have few questions:

1. If "trylinear" filtering existed in ATI drivers from Catalyst 3.4 why do we use *ANY* ATI card/driver output as reference in comparing? Why not compare images with software reference rasterizer?

2. Why should I accept this "optimization" just because I have Radeon 9600 Pro? I shelled 180 EUR for this card when I bought it and I also expect it to do full trilinear when I ask it to do it.

3. How can anyone believe ATI that there is no exact definition of trilinear filtering when there is?

4. Even if it does not degrade IQ (which IMHO has yet to be proven) how can anyone think it is OK when it gave unfair advantage to ATI because they asked reviewers to turn off the similar if not the same kind of optimization in competitors product thus biasing all benchmarks in their favor?

And my 2 cents on WHQL signing and ATIs poor excuse:

1. WHQL signature means just that the driver has passed stress test and it does not have anything to do with IQ. There is no specific WHQL for display drivers, WHQL testing is the same for all drivers.

2. If it is true that Microsoft worked close with ATI when it accepted their FP24 as a DirectX standard instead of FP32 how can we rely on them being neutral when doing IQ testing if they do it at all?

 

[reever]

I am new here and before someone says I am a

Uh yeah. finish that thought....

4. Even if it does not degrade IQ (which IMHO has yet to be proven)

It hasn't been proven that it *does* degrade IQ....

 

[Suspicious]

Uh yeah. finish that thought....

Didn't know certain words are filtered out.

It hasn't been proven that it *does* degrade IQ....

What is your problem? Of course it hasn't been proven when you got nothing valid to compare it with!

 

[Bry]

Re: Few questions:

4. Even if it does not degrade IQ (which IMHO has yet to be proven) how can anyone think it is OK when it gave unfair advantage to ATI because they asked reviewers to turn off the similar if not the same kind of optimization in competitors product thus biasing all benchmarks in their favor?

ATI's competitor did not have the same kind of optimization, nor was it similar. They never used TRI at all..it was all Bri. Where as ATI is claiming Tri when IQ difference would be noted. And Bri (or whatever they use) when no IQ difference would be noted. If this is the case, And this is what has not yet been proven, then I see nothing wrong with this. And would think it would be a great thing because it can give us some performance boosts without degrading IQ. Which is what we all want.

2. If it is true that Microsoft worked close with ATI when it accepted their FP24 as a DirectX standard instead of FP32 how can we rely on them being neutral when doing IQ testing if they do it at all?

I think you got this backwards. NV is the one that pushed FP-32. FP-24 has been the DX standard since the beginning. The odd thing is after NV pushed to get this standard approved. They could not even run it properly. IE it was just a checkbox PR stint to try to sell more of their cards. Yet people now screem because ATI does not support it..

The only issue I have with this so far, and it may change after dave lets us know his findings, is why ATI did not publically acknowledge this. And use the public to help them tweek it to make sure everything works properly. I.E. We could let them know if in a certain game it is reducing to Bri when it should stay Tri. Obviously with this being a general optimization and not a game specific. It could run into problems with a few games. If this was the case and we found out about it. There is a possibility that they could make changes to make it work right in that game. Even if it had to be game specific. But in doing it the way I described. I am sure all but pure NVidiots would not complain as it still gave us the best IQ and best speed possible.
Also the fact that ATI's PR claimed full tri at all times when it is obvious that that was not true. [/quote]