ATi is ch**t**g in Filtering

Well ATI’s own documentation claims high quality is using full trilinear, something they criticised nvidia for not doing, in the end we find out that what they were claiming as high quality (ie Full trilinear in AF) isn’t what the card is doing. Quality is subjective, everyone has different quality monitors which can highlight and hide rendering errors or problems, and bilinear artefacts I find are noticeable quite a bit more in motion, in a screenshot it may look ok but in motion I’m sure you’ll notice some differences. But I’m certain that most people consider high quality AF uses full trilinear all the time, anything doing different in my mind is a hybrid trade off of quality for performance. I’m sure that the cards can do full trilinear all the time, which says to me one when puts the slider in high quality, they are not getting the absolute best quality the card is able to produce, then there’s the whole deception and lying thing that makes is even worse.

 

what documents ?

All i have read from them is that these optimizations don't affect the image quality and if there is a case where it does we should report it so it could be fixed


and if you want to see what full trilinar looks like we should use color mip maps .

I don't see where they say it doesn't provide the same image quality you'd get from trillinar .


For that matter I remember seeing refrence rastner (Sp?) images compared to nvidia and ati cards of the time (9500+ and fx series) showing all 3 images differering .



So how do we go about this ?

I know the image quality with nvidia solution on the fx (can't speak of the 6800 ultra since i don't own it) is visable to my naked eye . Yet ati's is not .

IN fact even just blowing up the pictures and comparing them in 99% of cases shows a diffrence.

YOu then have to subtract the picture from the 9800xt picture.

What results would we get by doing that with the 6800 cards ?

Would it not also be wrong compared to the 9800xt pic ?

 

While I'm all for a low hit alternative to just Bilinear I do think that both Nvidia and ATI should also include manual options for the user to select the highest quality irrespective of performance hit.

It does seem that rather than ATI v Nvidia creating "Competition" to benefit the end user this new generation is all about keeping the prices high and lowering IQ so that users see higher fps numbers. Both manufacturers have dropped IQ for performance - but with the argument that it's either no lower or not as low as the competitors.

 

and ? I read no where that it reduces anything to bilinear. Just that its something in between.

That is the way its been explained in all of this .

Also this is providing the same image quality as trilinear according to ati and all the benchmarking sites that have noted its image quality

 

If it's producing the same image quality, then the diffs that are being done with the 9800xt that uses full trilinear should be same, but they are not so image quality isn't the same and like it or not ATI is using "TRICKS" to get similar quality, which according to them in their own documentation is unacceptable. Basically it's all pot kettle black.

 

so does nvidia do "full trilinear"?

Sorry but again, does nv40 and nv35 do "full trilinear filtering" when "TriOpt" or something like that (I don't use nvidia hardware) is turned off?
Please forgive me if this question is so naive or old subject but I really don't know whether nobody knows or nobody want to answer.

 

Is the image quality of the geforce 6800 ultra the same as the 9800xt ?

How about the 9800pro ? Is that the same ? The 9700pro ?

For that matter when did the 9800xt become The refernce image ?


I find this all interesting as everything ati said was bad they didn't do

I have already said running a whole bunch of cards 9600xts , 9700pros , x800pros , geforce fx 5800 the fx 5800 looked worse to my sister adn no one else and all the radeons looked exactly the same .

I will admit that I'm not very good with quality as i have not trained my self to see all the diffrences but my sister has working as a 3d graphic designer (well getting her degree in it) and obviously many other people who test this cards and compare iq have not been able to see it even with images blown up 200% .

THey have to take the images and subtract them from each other and then blow it up again to see the diffrence.

Perhaps i'm reading something wrong. Perhaps your just trying to make a bigger deal out of this than it is .

Wonder which it is

 

Re: so does nvidia do "full trilinear"?

NV35 does not do trilinear nor ever will (not unless someone finds a registry entry that enables it on the new drivers). NV40 does not do trilinear when the optimization is turned on (and it is turned on by default)

And IIRC, reviewers were asked to benchmark with default settings ie. with optimization turned on.

 

What makes you think 9800xt is the golden standard. Everything I saw seemed to indicate X800 had better filtering. If anything it appears as though ATI is using "tricks" to get better performance and better quality...

Based on what I could gather from the thread and the shitty translation from the german website

- colored mips have lower performance than regular mips..My take..So, what?

- 9800 xt is slightly different than X800. X800 is more closer to 9600...Ok..again so what?

and the conclusion...ATI is cheating..ATI is cheating...actually makes me wonder why this thread ran for so long...and don't even get me started on Borsti's tests (X+X)/2 = X doesn't matter what hardware you run it on..unless it's some very ancient hardware that really sucks

 

You forget the fact that Ati tells reviewers to use "color mipmapping tools" to determine quality. Which this method invalidates. And they're also talking about not using "filtering tricks" that reduce the workload when they're in fact doing the same thing themselves. It remains to be seen if this is reducing IQ visibly though.

 

Maybe ATI should market it as "Trilinear-Plus"

 

No different hardware implementation so there are differences of course.

With AF they are actually the same, nothing changed in AF between the 9700pro and 9800pro.

It’s reference imo because it’s uses very much the same angle dependant AF technique and we know that it does full trilinear with no tricks, fyi with colour maps the 9800xt and X800 produce the same results.

What planet are you on? They claimed they were using full trilinear and got busted by using a brilinear technique. End of story. You do know the meaning of “pot kettle black�

So everyone here is meant to take your sisters word as gospel? lol. Forget reference rasterisers, they should just stick your sister in a room and run tests against her.

Well that’s what people had to do with nvidia’s shader hacks and nobody gave em slack for doing it? Why should ATI be treated any different when they have shown the same sort of misconduct?

Well if it wasn’t such a big deal then why didn’t ATI come out in the first place and be honest with their new technique? Obviously they thought it was a big deal to hide it and give the impression that the X800 series was super fast with so called “Trilinear AFâ€. There lies the significance, the constant marketing deception that these graphics companies are doing, which people like you think is ok?

 

Ah well actually its all in the wording. You see, they term some tricks as "unacceptable" and those are most definately bad. But these tricks that they are using fall into the "acceptable" category so thats OK

 

I have to agree with jvd.

Just doing bitwise comparisons is not a valid method to determine high quality. (just to determine differences, but that is not the same as quality)
Nobody will claim that the 9800 and 6800 are cheating because they don't produce an image that is bit-for-bit the same as the reference render, or the same as each others images.

If we accept that, and claim that all three images are high quality, then we should also accept that there can be images from the X800 and 9800 which are not bit-for-bit the same as each other or the reference render, and that that does not mean that both images are not high quality.

The real question then becomes: How do we determine if an image is high quality.

Unfortunately most people in this thread don't seem able to consider that question at the moment and just whine that ATI misled them about the techniques they used to attain high quality.


IMO the most important part of this discussion is if the IQ is still high quality.
Now I don't care a bit which technique is used. High IQ is not something defined by the usage of algoritme A or B. High IQ is something that is defined by the end result. That is the only thing that really counts.

Somewhere in this thread someone has mentione that full trilinear is not always the best solution for high quality. But because it is the best solution in MOST cases, it is used for everything.
Now suppose you have a drives that doesn't use full trilinear in those rare situations were it is not the best solution. Does that mean that IQ has fallen beause full trilinear was not used? Of course not!

What if you have a situation were full trilinear doesn't add anything at all compared to a simpler/faster solution. And you ONLY use that solution in THOSE situations. Does that mean that IQ has fallen because full trilinear was not used. Again it has not.

Conclusion: it is the end result that counts and not the method used.

It does pose a problem for reviewers because it makes it harder for them to judge IQ. But is THAT a reason to call ATI cheaters and to let them remove a algoritme that improves performance without sacrificing IQ?
I should hope not!

So let's focus the rest of this discussion on the real issue here: Do we get lower IQ?

So far I see people unable to tell a difference in quality.
I see a few people saying that they think IQ is better.
I see nobody saying that they think IQ is lower.

(Quasar doesn't want to say what he thinks. His only answer is that he won't answer because if he says he thinks it is lower, people won't believe him. But that doesn't mean that he actually thinks it is lower.
To me it sounds like the typical answer of someone who doesn't see any lower IQ but does not want to admit that)

We'll have to find way's to determine the worst case scenario for this technique and see if we see differences.
And we can forget all those images with coloured mipmaps, because they won't tell us anything anymore. We will have to look at the actually end-result from now on.

 

So in conclusion:
- Ati does an adaptative filtering
- This filtering invalidates color mipmaps tests
- This filtering has not shown any IQ defficiencies atm
- This filtering offer better FPS
- We still don't have any good methodology to show if there's any differencies in IQ.

I'm i right?

 

so why compare the x800 to the 9800xt ?

It is a diffrent hardware implementation. THe closest one we have to the x800 is the 9600xt . Perhaps you should compare there .

Are you sure ? Have you taken pictures and subtracted them from each other and then blew up the picture to view ?

Yet we know the x800 is closest to the 9600 series in these things and infact by ati's own words they pushed back the angle dependency in the new x800s . Which invalidate this doesn't it ?

well instead of little attacks at me and poking fun you can now explain away all the reviews who said the image quality looked the same and was very good ?

Oh wait you can't

Please point out these examples

as in all cases you can clearly see the diffrence esp in farcry .

From others posts you can see ati admits it . Look at the slide under What can be done .


ALso note that ati says to see triliner to use color mipmaps. What they have lied about i don't know .

THey have obviously been using this for a very long time in the 9600 series and no one has complained


This isn't just blowing up the image and seeing a diffrence. This is having to caculate the diffrence.

The one pic which should show this looking worse actually shows a sharper texture and the image quality looks better !!

 

Why would this method invalidate color mipmapping tools? Color mipmapping tools tell just one side of the story..Just like they always did..Whether ATI has this optimization or not - it's very simplistic to assume an implementations quality just using the colored mip tests that people have accustomed to using - as I noted before - there are some well known techniques to take advantage of signals with certain properties

a small hint for example...if say you already have a color value from a miplevel at value "10" for eg:- and you somehow already know that what you are going to fetch next is a value of "5" - why wouldn't you just shift 10 right by a bit and generate 5 and save expensive fetches?

what if say similar set of mathematical relationships allow you to implement higher quality filters at lower-cost - since you can do trick math without full math units...

Anyone with signal processing background and basic understanding of computer graphics could be having some serious fun at all the so called IQ analysis being done here

The basic assumption that's wrong is that filtering quality is independent of the test signal - how good certain filter over the other is dependent on the input signal.

I will not be surprised if the ATI docs you are referring to are written by marketing departments who have not much deeper understanding than most review sites that use these tools...sorry..I wouldn't necessarily attribute for ATI to be that smart enough to be deliberately mis-leading reviewers..I could be wrong though

 

(Edit: Oh yeah, I agree )

I wouldnt say it invalidates it as such. But...

The algorithm is adaptive so will give full trilinear when mipmaps are of a certain difference or more. So coloured mipmaps would give an indication of what happens for real mipmaps that cross this threshold.

However where is this threshold? It could be set to such an aggressive level that the only textures that get full trilinear are those that are as different as the coloured mipmaps, and how often would that appear in a real scenario?

Also is it a threshold or does it alter the regions depth for trilinear filtering as a function of mipmap difference?

I think there is much fun to be had for some people here trying to uncover whats going on.

I would like to see the one sugestion tried (by Ostel I think) and try using coloured mipmaps where only half of it is coloured.

 

You agree or disagree?

Note that it's the best situation, all the others are mixed tri/bilinear. If we could speak of lie, it would be a lie of omission.