ATi is ch**t**g in Filtering

Because it would be way too expensive for the driver to check dynamic textures the optimizations are probably not applied there...

Sounds like a plan...

 

BTW, if IHVs want to bring us all those new cool wellperforming filtering modes, why not do it the usual way with introducing an API extension or cap, and leave the legacy modes alone.

 

Well, if they would, there'd be nothin to win, since it's impossible to write engines using imaginary caps.

 

borst 1 more question. how come you only benched farcry at 1026 resolution??? lookign foward to ur reply

 

Why don't you leave your questions about thg's benchmarking practices out of this thread, it's really getting annoying.

 

You dont consider this an interesting development? ATI possibly not doing what they advertise?

Sure, you mean you fogot the post you responded to that had a summary on pasge 17?

Here is what was said if you forgot.

Nice try and dont forget your response on page 26 which is just 3 pages before you asking me in a sarcastic way to summarize what has happened.

I think it is quite apparent you have read the entire thread and thus your request was bunk and instead written to invoke a response.

 

than dont read them. its not addressed to you and its a perfectly valid question.

 

My "thanks for the recap" was 9 pages back from your "interesting development" post, I thought something had came up in those 9 pages that I had missed.

 

Ok, so it sounds like my theory wouldn't apply... Now I'm probably out of my league here, but despite the fact that the texture is compressed couldn't the hardware check if a box filter was used? After all, at the time of performing a trilinear computation couldn't it
do a comparison between the original texture and a computed box filtered texture? If they're equivalent, then mark it with a flag indicating that next time it can use an optimization so it only needs to be done once per mip level pair.

Or would this be too complicated to implement -- perhaps the autogenerated case which no_way suggested is correct.

(edited for clarity)

 

So let me get this straight - ATI is doing something to optimize the performance of their trilinear filtering. The algorithm is adaptive so it applies only as much trilinear filtering as is necessary to preserve the image quality. No one has been able to show it has any visible effect on image quality unless you take a difference image and maximize the contrast, thus creating something which barely resembles the original image. And they've even offered to address any such situation if one can be found - this would imply pretty strongly that they've already looked for such situations themselves.

So people's response is to call them cheaters (because it works on colored mip levels rather than breaking down) and demand they add yet another option in their control panel that will have no effect on image quality but will lower performance. I just don't see how this makes any sense.

 

One solution I can think of is to have a texture where one half is a normal texture and the mip-maps on that half are also normal, then have the other half coloured. Take another texture that uses the same pattern as the normal half on other texture and then do a comparison. If such detection is occuring, the half-and-half texture should receive full trilinear while the other one shouldn't, right? This would also work towards finding out how noticable the difference really is.

EDIT: This would also allow us to see if what ATI is doing is an adaptive algorithm whereby the type of trilinear is determined is on a per-sample basis or on a per-texture basis. The former would definitely be interesting. 8)

 

Why don't you respect this thread instead and PM Lars with Qs that may be "valid" but are otherwise totally unrelated to the topic at hand? As I previously requested, please don't make it harder than it already is to get through a thread that's already at 32 pages (a great portion of which are not on-topic). TIA.

Not quite. We haven't yet determined that the (presumably software) algorithm applies "only as much tri as nec'y to preserve the IQ."

Actually, I believe Quasar has said more than once that he can notice the difference in motion, though I'm not sure he has yet labelled the resultant IQ as inferior (though the algorithm does appear to produce inferior filtering than "true" trilinear in that the MIP-map transition seems to be smaller). I do agree with Dave that the fact that the 9600P has had similar optimisations for quite a while (as he indicated back when he uncovered nV's UT2K4-specific brilinear) and no one's really complained is a positive sign (though I don't know if the 9600 series default to HQ [tri] or just Q [bri] Texture Quality).

Again, I think it may be a bit hasty to take ATi's word for it that they're satisfied with the results, or even that they'd change things if enough people demanded it (surely there've been more than a few requests for SSAA since R300 came out?).

Actually, people just want control over the IQ, which is the reason why nV included a "disable Tri optimisations" checkbox in their drivers. And it is a bit deceptive to parade colormips as indicative of their tri quality when it seems the MIP-map transitions are altered in more "normal" usage (e.g., a floor or wall of repeating textures).

 

Thanks for the clarification, I don't know how I missed your explanation earlier. So this seems to give some credence to this method by ATI I think. The majority of the complaint seems to be coming only from high contrast images which doesn't represent anykind of real gaming.

Now rigging textures that don't represent anykind of real game application to exposed this new type of optimized filtering seems flawed from the onset. Real game textures being looked at seems to be where the attention should be at. Does it have obvious mipmap lines, does aliasing increase around mipmap transition points during motion I think should be the focus.

 

Offhand, I think that matter has been in debate for ages. The IQ tends to get split up into many different sectors (not just AA and AF, but different methods within them as well) and the "Pure IQ" label stuck to one company or another at each spot.

...and then, of course, it all gets played out over the performance levels.

But yes, I agree that this technique should have been more fully documented and explained SOMEWHERE, especially before a new generation of high-end cards came out to be scrupulously analyzed. IQ has gotten much more analysis of late, and the reviewer and enthusiast reaction has been much more knee-jerk. Starting off knowing the most about a situation is much better than having to peel the onion apart layer by layer through a company's PR department.

 

So if ATI is using an algorithm that checks LOD difference between mip map boundaries and varies the filtering done, this means games that use more mip map levels will in effect perform better on the X800 series, since there are more mipmap levels the differences between each mip level isn't that great so full trilinear can be dodged. But a game using very few mip levels will mean the difference in mip levels are greater and so trilinear will be used in order to keep quality at the cost of performance.

So i would assume older games that use less mip levels take a greater hit in comparions to newer titles which i've noticed use more mip levels.

Looking at some benchies with older games and AF on the x800 it seems to be the case. I guess this is a test someone can perform, ie change the number of mip levels and bench the fps difference. If it is adaptive it should perform like i expect.

BTW I'm a firm believer that if one selects high quality they should get high quality, not some optimisation/approximation of high quality. there isn't anyway you can deny that ATI were advertising their technique as full high quality trilinear, when in reality it isn't. Because of the lying and decieving i catergorise it as a cheat. If ATI on the otherhand offered this mode as driver feature and advertised the benefits of the technique first hand, they could of used it as a marketing feature. Instead they have proven they are no better than nvidia when it comes to integrity and honesty.

 

None of this makes sense at all. Textures with mipmaps normally have the full range; i.e. a 2^n x 2^m texture will have max {n,m} levels. You can't "skip" a mipmap in the middle of the chain.

Again, this doesn't make sense.

-FUDie

 

thinking about it yeah it probably doesn't make sense, but does anyone else think the depth placement of mip mapping bounderies and the number of em would impact this filtering technique? Then again i guess since ATI hasn't release bugger all info on what exactly this method does it's all just speculation.

 

what makes you beleive that ATI isn't offering high quality when you select high quality? Do you have any proof? All I saw in the whole thread was that X800 image looks different than 9800 - when zoomed up and bit-wise diffed....And many posters claimed that they like X800 quality better than 9800

 

Hey so am I, but I tend to worry more about what is actually seen on screen rather than the exact technical despcription of each effect being used to display that image. To me if an image looks high quality (no jaggies, no banding, no mip map transitions, etc etc) then IT IS high quality. So far I've not seen one shred of evidence to show that IQ has been reduced and from my perspective having to resort to image bit wise image comaprison in image editior apps is a bit desperate.

As for the marketing docs, well they ARE PR material what do you expect?