Can someone sum up the ATI filtering thread please?

Yes . I'm glad a program can tell me the diffrence. Sorry to say I wont be running this program while i'm playing a game right ? Which means its my eyes that must see the diffrence and I can't. With earlier geforce fx drivers i could howerver and still to a small degree can .

Which is why i really dont' understand this witch hunting .

In motion you can't see any diffrence while playing a card .

 

"Should have"? Either it does or it doesn't. It's possible to take samples from higher detail mipmaps and not shimmer you know.

Note that by default the program multiplies the differences by 64! Pretty extreme. Since it appears that the driver is looking for visual differences (not differences that are visible after multiplying by 64), maybe it should default to 1. with a difference scale of 1, I can see a very faint "ghost" image in the difference mode. For all intents and purposes, the two images are identical to the naked eye.

-FUDie

 

My understanding of texture filtering is rudimentary at best, but shouldn't sampling from the more detailed MIPmap lead to less shimmering and blurriness as the screen resolution increases? Is the shimmering (as) noticable at 12x9/12x10 or 16x12?

I'm looking forward to an English translation of that article, Clootie.

 

If you disable mipmapping, i.e. sample from the base texture all the time, then shimmering becomes very apparent. This is due to undersampling. Basically, anytime you undersample, you can be vulnerable to shimmering.

-FUDie

 

Not without taking more samples.

 

It also depends on the data you are sampling. Some data can withstand (some) undersampling, other data can not. Even bilinear filtering can "shimmer" since it's not the perfect reconstruction filter.

-FUDie

 

Sure, but you can't assume that for a game situation. Your filtering technique should be able to handle any data you throw at it.

 

Isn't that why ATI is doing texture analysis? Have we seen evidence that aliasing is now worse than it was? So far, all I have seen are static screenshots from a very difficult texture. Yes, there does appear to be some moire, but there was some to begin with anyway. Look at the shots Dave provided, all three cards had aliasing with the provided texture. It's not obvious to me that the R420 result would be any more apparent in motion given that the differences are so small.

-FUDie

 

Thanks, FUDie, but does a resolution of 16x12 begin to remove undersampling as a problem, or will current "ultra high" textures still shimmer? AFAIK, the point is to be at or below a 1:1 [pixel:texel?] ratio, and I'd have to guess that we're approaching that at 16x12. Or maybe just switching from high- to medium-quality textures will solve this?

Sorry for my total lack of filtering knowledge. I'm just tackling RTR (and Nyquist, box filters, etc.) now. =)

 

I wonder if they'll make a benchmark that multiplies the differences by 64... 1 fps faster equals 64 fps faster, cool.

 

Here’s a summary regarding ATI’s adaptive-Trilinear …

That quote is from the TR interview and this page 6 quote sums things up nicely…

 

Given the same size texture, higher resolution render targets will improve undersampling since the pixel to texel ratio is higher for more pixels. However, things in the distance can still shimmer since you will eventually hit the threshold where the pixel to texel ratio is too low.

-FUDie

 

And what, do you think, ATI is actually doing for texture analysis?

If it was simply high-contrast textures, then the test program that ATI sent to Dave would have "brilinear" disabled for the whole thing. No, it appears that right now, ATI's driver doesn't actually analyze the data that makes up the texture at all. It appears that they just decide whether or not to apply optimizations based upon how the texture is called (i.e. whether or not it's a "managed" texture).

 

And have your Mr. Chalnoth figured this all by yourself Amazing how much one can bullshit with 1/2 knowledge

why don't you write a simple test program and figure this out yourself and then come back to comment?

The excuses of not having an 9600 or X800 or unable to write a d3d program don't cut it....if either of these conditions are true refrain from posting anything authoritatively on this topic..please

Have you even looked properly at the images Dave put up...?

 

Yes. I'm sure that's a patent-worthy algorithm right there.

 

Yeah, right.

Since programs are known to change the texture type when enabling coloured mipmaps...

 

Well, I suppose they may do it by looking at some min/max brightness algorithm, but then that wouldn't always catch colored MIP maps.

Or, it could simply check to see whether or not the higher MIP levels are a simple box filter of the lower ones, but then that would pretty much only cover colored MIP maps.

I somehow doubt they're doing a Fourier transform of each texture to ensure that they only use it on the ones where aliasing can be avoided, as the texture load time would be pretty darned slow. And even then, the algorithm would be better-used on adjusting LOD, not in applying brilinear, as if the fourier transform components are known, then it should be relatively easy to figure out how much one can shift the LOD before aliasing becomes apparent (this would be really nice, though: have an optimal way to adjust LOD on a per-texture basis, and supply developer tools to accurately calculate the appropriate LOD for each texture).

 

I've said "should have" because it's quite subjective, but math is right where, so just check it.

Multiplier set to 64 just for that - so you could spot differences in a second. But it's NOT fixed you can and should modify it. It's feature - not restriction! Problem is when you move slider to 255 you'll see that in some angles more when 80% of image is different in 9800 and 9600/X800.

Ooo, we have a pretty intelligent comment here

Nope, ATI driver tries to analize texture and if mipmaps are different enought - it switches off optimization mode for that texture. It's hard to tell what algo they use. But it's "if not MAX(difference)", probably the same as used in The Compressonator?

 

Well, you couldn't use something like The Compressonator, because MIP maps are of a different size. And a simple difference technique would disable the technique if there's a great difference in colors, and so ATI's program wouldn't have worked.

 

and do you actually know why? i guess you know, too. YEAH BECAUSEOOF CHEATING WITH COLOUR MIPMAPS!!! HARRHARR.

a simple hint, as you possibly still not got a clue: if it is managed, its known to be box-filtered down as it is the drivers job to do it. that means it knows the optimisation can work good there. if its NOT managed, at any time the user (the application) can change any mipmap level, thus, everytime it does that, the driver would have to do image comparisons over several layers (3), and do filtering, and statistics.

and guess what? this can be costy, espencially if its a texture that gets updated often. guess what? thats why they avoid this comparisons.

end result: if the application says "hey, dx/gl, create the mipmaps for me", the driver knows it can optimise as the driver knows how the mipmap looks.
if the app says "hey, dx/gl, i want to handle it myself", the driver doesn't, and the work to check would be huge. so, it doesn't optimize.

its all rather simple to understand, but as i said. you're too dump to do so. you bether feel 1337 at bashing, instead of learning the difference.

it is not an evil plan ati had, to hide their optimisation/cheat. it's NOT. this is not about opinion, this is a fact. you have to learn and accept that.

i'm not against nvidia or ati. i'm just pointing out the huge difference in what the optimisation actually does. nvidias simply lowers quality on the whole image. ati lowers quality where it knows it doesn't really hurt. or tries to.

and besides that, there is the fact still, that trilinear is not the holy grail, and trilinear of managed textures can be optimized lossless. both are independent on how ati implemented it, but both are true as well.

one day you will get mature, chalnoth. but not in this hw generation, i guess.