ATi is ch**t**g in Filtering

But for the most part, You'd have to expect that D3d Only requires Billinear Samples on most mip map transistions. Point Sampling on Mip Map transistions would Obviously bring back no Certification.

 

Those are the tests DrawPim and you were talking about a page or two ealier? The ones with wildy different mip-levels of vertical and horizontal bars?

 

So rendering with color mipmaps on is now considered "default" while image rendered normally is special case?

 

No, by default the GPU/driver is doing trilinear - as long as no texture is mip-mapped whose mip-levels exceed an unknown contrast value to each other.

Possibly 'real texturing' at all switches off trilinear? Who knows?

 

Specific suggestion for reviewers to use colored mipmaps to verify the behaviour seems kinda slimy.

Ok, can anyone confirm for sure that this behaviour only occurs when application requests autogenerated mipmaps ?

EDIT, i.e. if i do my downfilter myself, then i will get full trilinear on all texture stages where i request it ?

 

Can someone force this "optimized trilinear" on color mip-maps?

Everybody here seems to forget that true trilinear is a trade-off. You get rid of mip map transitions but also the image gets blured in comparison with bilinear by blending with a lower mip map level.

Most FX users said Bri gets a better looking image due to bilinear being sharper.

If ATI can offer a filtering mode that deals with mip map transitions (and not just borders) and gives the added "crispness" of bilinear, I'm all for it.

 

Even bilinear is a tradeoff - information from different texture samples are averaged.

The sharpest possible pictures are renderen with point sampling - albeit with a huge tradeoff regarding overall quality. Not because it is not sharp but because it looks unnatural.[/i]

 

db wrote:

all i can find as a requirement to pass the test concerning textures in the current microsoft version of the whql test specifications is this:

http://www.microsoft.com/whdc/hwtest/pages/specs.mspx

and this:

care to point me to the direction you mentioned?

 

ATi's response

Our algorithm for image analysis-based texture filtering techniques is patent-pending. It works by determining how different one mipmap level is from the next and then applying the appropriate level of filtering. It only applies this optimization to the typical case – specifically, where the mipmaps are generated using box filtering. Atypical situations, where each mipmap could differ significantly from the previous level, receive no optimizations. This includes extreme cases such as colored mipmap levels, which is why tests based on color mipmap levels show different results. Just to be explicit: there is no application detection going on; this just illustrates the sophistication of the algorithm.

http://www.theinquirer.net/?article=15971

Doesn't address why the final output is difference from the Software Reference Rasteriser. What are the implications of this?

 

This has been very interesting. Of course, the opinions here seem to be split as you would expect….with maybe the exception that many of the “fanATIcs’ seem to be more circumspect about this than most “nVidiots†ever were about any of the stuff nVidia has pulled. While there are some things about this “optimization†that are somewhat alike between what ATI & nVidia does, it need to be brought into a bit sharper relief….

nVidias use of brilinier was an attempt to add performance to a sub par architecture that caused major IQ degradation. It was noticeable within seconds, ATI has added a feature to the hardware that improves performance and is not noticeable in use…in fact, has been used for the last year on the RV products with absolutely no complaints… and in fact, the RV series is notable – as are all of ATI’s current products – for having the best IQ in its class. ATI is pushing the envelope of videocard technology here, not trying to cover up a bad product design.

The fact is that what ATI is doing here, technology wise, is a very good thing. If what they say is true, they are giving us as good, if not better quality while increasing the speed. And the fact that the RV’s have been doing this for the last year without any complaints only re-enforces the fact that it works incredibly well. Now, how they have handled it and are handling it is the real crux of the biscuit. Fact is it should be defeatable, period. If for no other reason so that users can get a feel for the technology. And they badly need to explain this new technology as fully as is possible. The enthusiast community needs to have the facts about this ASAP. And this can only help ATI’s case here.

I kind of look at this the exact same way I do with adaptive AF, only it’s better – much better. Where you have a completely obvious change to IQ with adaptive AF over regular AF, with this the differences are, so far, not visible. More speed without any IQ loss…. Kind of a holy grail kind of thing. It’s a case of ATI using intelligence to solve a problem, and I applaud them. And I’m sure that nVidia will follow suit here at some point, just as they have done with adaptive AF. It’s a no brainer…..

 

The main issue is not the actual filtering imo. It's that Ati is telling everyone to use (look at the pictures from the pdf's in this thread) "coloring mip map" tools for checking the quality of their filtering. And then they implement a new mode of filtering that invalidates these tools. Without any explanation and without offering any way of turning these optimizations off.

 

Indeed. ATI know damned well that the filtering used in the coloured mip-map synthetic tests will more than likely be different from the actual filtering applied under normal game conditions.

They may not think it looks different, they may or may not be right, that's not the issue. The synthetic tests they are holding up as demonstrating their IQ is demonstrating the IQ under the best possible conditions, conditions which do not hold for a wide range of real-world situations.

 

That's a pretty valid question actually and we should ask ATI to see what it has to say about that.

 

On the subject of application detection, I think there should be a checkbox option in the driver control panel to allow or disallow the detection, with an explanation beside it that some games require the detection to operate correctly (the explanation could link to the driver help, listing the specific games/apps being detected). This is something that should apply to all companies, i'm not singling anyone out.

The same applies to brilinear filtering and variants. A checkbox enabling or disabling the filtering optimisation should be provided to the user and obeyed by the driver at all times.

Give the user the choice, then respect that choice. If the user then suffers poor IQ or poor performance it is because of his choice and his choice alone and he can remedy it if he so chooses.

It really is that simple.

 

This all optimizing stuff is just silly. Instead of investing R&D time on improving texture filtering quality, IHVs develop more and more techniques to reduce basic trilinear filtering quality even more. This gen hardware offers much worse filter quality than a plain old 2001 Geforce3.

Hell, not even all D3D filtering types in D3D are supported in current hardware. IMO, there would be a lot of things to improve, bicubic filters, gaussian filters, smart filtering and last but not least, single-cycle filters.

 

If we add a checkbox for everything, you are looking at a total usability nightmare though. And big corporations take these things VERY seriously; we are a minority, and the millions of users do not want a control panel filled with lots of different (and most of the times confusing) options.

edit: That does not mean that we shouldn't have the option to do stuff btw. Just looking at it from a different angle.

 

Fact : Reviewers were tricked by ATI about filtering on X800.

Sad.

 

ok, about checking ATI filtering:

1) bench application with textures vs colored mipmaps vs monocolored mipmaps
2) use your own texture, let driver autogenerate mipmaps. Texture has to be prepared in a way that every generated mipmap differs much. Then try to check if transitions are bri/trilinear.

Any method seems valid/possible to you guys?

 

I have to agree with that but, from a marketing POV, if they went straight with this on launch, how many sites would have checked "disable trilinear optimisations" on NV40 drivers and which was fairer?

 

Cripes! For about the first time ever I totally agree with radar!

I wonder if the ATI hardware is capable of doing 'true' trilinear in addition to the optimised filtering?