Why is anisotropic filtering so slow on the GF4?

[BD2003]

2x Ani filtering knocks the framerate down more than quincunx AA. It was much less of a difference on the GF3. Wtf?

And when will nvidia get their act together and release a driver that does Anisotropic as fast as ATI? Sure Ati doesnt have "true" anisotropic, but can you actually tell the difference? The lowest level of aniso on my 7500 blows away everything but the highest on the geforces...

 

[butch84]

can anybody explain to me what exactly anisotropic filtering does? I understand the concept of antialiasing, but not anisotropic filtering . . . . damn its hard to spell!!

 

[richleader]

Simply put, it makes the transition from detailed textures to lower quality mip-maps smoother. Rather than in bilinear or trilinear where you can see abrupt lines where these transitions occur, anisotropic filtering makes a smooth sphere of texture gradation. Higher levels of filtering expand that sphere so textures have more detail further away. It can also reduce texture shimmering in far away objects--I think Reverend at 3dpulpit has a good video of the effect in the game Sacrifice--some people have called it "FSAA for textures" not accurate, perhaps, but it's close to the effect.

I can't speak for the geforce 4 (I have a Geforce 3 classic, use 4x ani, happily, with little hit), but I don't remember the anisotropic benches in any of the reviews, here or otherwise, reflecting the kind of experience that you're having.

 

[BFG10K]

And when will nvidia get their act together and release a driver that does Anisotropic as fast as ATI?

If by "getting its act together" you mean not having true anisotropic and not being able to utilise trilinear at the same time then probably never. Still, it'd be nice to have an "approximated ansotropic" option to make it like ATi's (although not with bilinear filtering though).

but can you actually tell the difference?

Yes, especially with that ugly bilinear filtering that you're forced to use in conjunction with it.

can anybody explain to me what exactly anisotropic filtering does?

Anisotropic filtering keeps textures sharp for longer distances, eliminates distortions caused by viewing textures at high angles and reduces texture shimmering, especially at long range. Generally makes everything appear sharper and clearer.

FSAA for textures" not accurate, perhaps, but it's close to the effect.

It's not even close at all and in fact it's the exact opposite. FSAA blurs things while anisotropic sharpens them; also anisotropic does not remove jagged edges.

 

[JellyBaby]

does Anisotropic as fast as ATI

BS. ATI's implementation, like BFG10K said, is a joke. It's not applied in many circumstances where Nvidia's *is*. ATI cheats, as usual, to produce those lovely numbers.

 

[oldfart]

GF4 anio performance discussion @ nV News

 

[BD2003]

I dont care if its not "true" anisotropic. All I know is that for the performance hit, its amazing. I'm not just reading numbers, I've seen it myself. You dont need aniso on every surface. Having it on most surfaces is a waste.

Sure it looks better, but if its 5 times slower, whats the point?

And I have no idea what youre talking about with the trilinear. It wasnt very hard to do both at the same time on my 7500.

 

[43st]

ATI cheats

But, as many know, it does look better. So if they cheat at making better image quality then that's fine by me.

 

[JellyBaby]

With ATI you're not getting the truth. Do they provide an option for "full aniso" vs. "half-baked aniso"? For me personally I want a video card that fully implements its features, doesn't change it's implementation based on the current game running, and doesn't fudge benchmarks.

 

[MattB]

ATI's implementation doesn't really cheat, it just turns off if the angle of textures is anything but 90 degrees. Whereas NVIDIA's implementation works all the time, and using trilinear. If you really can't tell the difference between bilinear ripmapping aniso and trilinear true aniso, then you may want to get an appointment for an eye exam. Both ATI's and NVIDIA's implementation are good though, it's better than no anisotropic filtering. I use both cards (R8500 128MB which I just posted a review of on my site today, and a GF4 Ti4600), and I like NVIDIA's implementation because it looks better all around and doesn't display pixel poppings like ATI's implementation, but ATI's implementation I like too because even at 128-tap it looks pretty good (over no aniso anyways) and the performance hit is negligable.

 

[BFG10K]

I dont care if its not "true" anisotropic. All I know is that for the performance hit, its amazing.

Well to be fair I agree with you to a certain extent. As I said before I would quite like nVidia to have a "fast anisotropic" driver option that does what ATi's solution does. The most common places that people look for the difference is on the floor straight in front of them and on walls which are straight and at 90 degree angles to the floor and ATi's method is still in action there.

However the biggest problem with the Radeon is its lack of trilinear filtering when anisotropic filtering is enabled. That is something that is unnacceptable to me.

But, as many know, it [ATI's method] does look better

Uh, no it doesn't.

 

[richleader]

t's not even close at all and in fact it's the exact opposite. FSAA blurs things while anisotropic sharpens them; also anisotropic does not remove jagged edges.

This is so very typical of you BFG, I'm surprised that you even bothered to read my post if you went on to answer the same questions with less detail. At any rate, my use of the quote was describing the use of anisotropic filtering to reduce shimmer and popping, something that FSAA does as well. If you want to get technical, I'd say that your assertion that FSAA "blurs" things is doubly idiotic (outside of Quincunx, at any rate) because an anti-aliased image contains more information than a non anti-aliased one--hence the need for greater sampling. I'm pretty sure most people could tell the difference between 4x anti-aliasing and a gaussian blur with a 2 pixel radius...

 

[BFG10K]

I'd say that your assertion that FSAA "blurs" things is doubly idiotic (outside of Quincunx, at any rate)

And I'd say your completely wrong. All current consumer brands of FSAA blur the image because when you break it down all you're really doing is blending/sampling multiple pixels/colours into one. This blending eliminates aliasing (and a few other things) at the cost of blurring the entire image.

because an anti-aliased image contains more information than a non anti-aliased one

Of course it does which is exactly why the blurring's there. When you've got multiple colours going into one colour, the original colour is diluted into a mixture containing the combination of the sampled ones.

 

[richleader]

Well, I'm still going to assert that "blurring" is more innapropriate than my use of a quote used by more than a few in the industry which carries the sense of anisotropic filterings results to many people's eyes--though I'll spare everyone by not breaking out webster's and the equations behind doing real blur filters and only say that a non-fsaa image isn't correct, it's only an aproximation of the mathematical truth (as is a FSAA image) so by definition, it cannot be blurred as there is no "original" to base that claim from or compare to.

 

[richleader]

And yes, to head off the inevitable retort, I am aware that background pixels can be more defined and "sharp" as well in a non-FSAA image, but that is a question of pixel priority, not "correctness." After all, cards don't ship with a LOD bias of -50...

Engaging both FSAA and anisotrophic filtering do allow one to push back the LOD to some degree without many of the detrimental visual effects that defined and "sharp" (BLURRY BAD!!!) pixels can actually add to the image.

 

[BFG10K]

and only say that a non-fsaa image isn't correct, it's only an aproximation of the mathematical truth (as is a FSAA image) so by definition,

Absolutely correct but the higher the resolution, the more correct and accurate the image is. If you had an infinite resolution the final rendering would be 100% correct. OTOH with FSAA it'll never be correct no matter how many extra samples you take or how much extra blending you do because at the end of the day you have to scale the image back down to fit in a limited resolution.

it cannot be blurred as there is no "original" to base that claim from or compare to.

The blur can be compared to two things:

(1) The image without FSAA to the image with FSAA at the same resolution (eg 1024 x 768 x 0 FSAA to 1024 x 768 x 4 FSAA).
(2) The image without FSAA at a resolution multiplied by the FSAA factor to the image with FSAA (eg 640 x 480 x 2 FSAA to 1280 x 960 x 0 FSAA).

In both cases it's plainly obvious that the non-FSAA images are sharper than the FSAA images (although they're not necessarily higher image quality).

but that is a question of pixel priority, not "correctness."

That's right - a correctness level that is increased as the resolution increases without the negative blurring effect of FSAA.

Engaging both FSAA and anisotrophic filtering do allow one to push back the LOD to some degree without many of the detrimental visual effects that defined and "sharp" (BLURRY BAD!!!) pixels can actually add to the image.

I'm well aware that anisotropic filtering counters the blurring of FSAA and in fact you have just admitted that the two are completely opposite in that respect because you've used this very example.