Geforce FX Bilinear Anisotropic Filtering Question ??

[Xmas]

? Since when does "Application" mean "High Quality"? I'd assume it means not to force any settings and let the application set everything, like it does for my drivers. Does it intuitively mean something else to you?

Indeed, that 'Application' setting is misnamed. But I can tell you for sure that it means "High Quality", i.e. identical to GF3/4 AF. 'Balanced' is worse, and 'Aggressive' even more so. You could call it 'almost bilinear'.

edit: I'm not so sure this slider affects AA at all.

 

[Brent]

i'll run some tests to see if its bi or tri and whats going on

man, so many more things to do with the card, i'll be busy for weeks

Tx ...shouldn't take long (can't remember console command for Serious Sam though) or Quake 3 for that matter.

/colormiplevels 1 = Quake3

Serious Sam 2 is somethin like hightlight miplevels i'll have to look it up

 

[demalion]

Hmm...you mean that there might be a lower performing mode than Anand's "balanced" benchmark results?

On the plus side, GF FX balanced mode looks good in screen shots.

I still don't know what exact mode Hard OCP used for the GF FX aniso (did I miss it, or is my hinting to Brent about this again justified?... ).

 

[Joe DeFuria]

Yes...now I am utterly confused.

I never heard anything about "3" different Aniso quality levels. Where did that come from?

 

[Doomtrooper]

More is better..right

 

[Brent]

We used Performance-Balanced in all the tests

according to the NVIDIA documentation Performance-Balanced is the equivelent of ATI's Quality Aniso mode

 

[Ailuros]

I was referring to the D3D "optimizations" where you'd limit certain mipmap levels (or texturing stages if that's teh correct term here).

The idea of it is that you initially pick Level 8x aniso for instance and after that you can optimize per texturing stage (one texturing stage at least have to be left unchecked).

It can look somewhat like that:

texturing stage 0 (unchecked; thus it'll get 8x aniso)
texturing stage 1,2,3 (checked; where you can pick to either use no aniso at all or any other level between 2x and 4x)

That way you'll get anisotropic filtering only on base textures. Additionally there's another box where you can limit the optimized stages (in the example above 1,2,3) to just bilinear.

That application/balanced/performance mode seems to be a Rivatuner carry over from Unwinder's openGL patches, because he had already a simple performance/quality mode in his application, whereby if the drivers where patched with the Anisobooster OGL patch, applications where forced to use whatever setting the user might have set in RT to (performance/quality).

The SS engine w/o the patch would default automatically to quality settings, whereby with the patch it gets forced to use whatever setting you prefer.

DaveH,

I'd still think that it's rotationally invariant. I'm in the mid of a system upgrade so I can't really fool around with that stuff right now. I'd suggest anyone attempting to get around it, can use Xmas' little anisotesting utility and/or fool around with some LOD adjustments.

A simple LOD bias offset in UT2003 costs in performance on GF's, wether aniso is on or off is irrelevant.

 

[OpenGL guy]

We used Performance-Balanced in all the tests

according to the NVIDIA documentation Performance-Balanced is the equivelent of ATI's Quality Aniso mode

That still doesn't answer the question of what "Performance-Balanced" actually does. If it's not doing mip-linear (i.e. trilinear) filtering, then it's not equivalent.

 

[Ailuros]

Xmas,

I really haven't looked much into it, but is there actually a performance difference between application and balanced? If yes which applications/games?

 

[Crusher]

So balanced is adaptive+trilinear, and aggressive is adaptive+bilinear perhaps? And application is non-adaptive trilinear, unless otherwise specified by the application?

 

[demalion]

We used Performance-Balanced in all the tests

according to the NVIDIA documentation Performance-Balanced is the equivelent of ATI's Quality Aniso mode

Was that info in the review anywhere (EDIT: that Performance-Balanced was used in all the tests)? I was looking at Test Setup and didn't see the info I would have expected there clearly stating what setting would be used, and I didn't notice them being specified around the graphs.

 

[Xmas]

Xmas,

I really haven't looked much into it, but is there actually a performance difference between application and balanced? If yes which applications/games?

I don't know, I haven't got a card to test

I can only tell that there is a visual difference between the three settings, at least when the application requests AF.

 

[Dave H]

aggressive is adaptive+bilinear perhaps?

Going by Anand's screenshot (and the very small performance hit), there's a lot more to aggressive than just angle-dependent adaptive + bilinear. Or maybe I should say a lot less. From the screenshot and performance, aggressive 8x might be nothing more than...2x?

Or maybe there's some complicated algorithm going on here, or severe positive LOD bias, or something...but the end result is just ugly (albeit better than no AF). Absolutely no comparison with ATI's performance (adaptive + bilinear) AF.

 

[Ailuros]

I don't know, I haven't got a card to test

I can only tell that there is a visual difference between the three settings, at least when the application requests AF.

Has the algorithm really changed between NV25 and NV30?

I can tell as much: Rivatuner flips from performance to quality only when I push the slider to "application". Balanced and aggressive apply to performance mode in RT (that of course with the driver unpatched).

Take them with a grain of salt since those are from a NV25. I used trilinear and texture LOD in the game at "normal" (drivers at "0"), unless noted otherwise:

Aggressive mode:

http://users.otenet.gr/~ailuros/AGR.jpg

Balanced mode:

http://users.otenet.gr/~ailuros/BAL.jpg

Application:

http://users.otenet.gr/~ailuros/APP.jpg

Aggressive mode + (-0.5LOD)

http://users.otenet.gr/~ailuros/AGRLOD.jpg

Don't know if that helps any.

 

[KimB]

The slider at the bottom is to determine the number of samples used in the filtering. AFAIK the balanced/performance is supposed to switch between the "filter everything the same" full-on mode that the GeForce 4 used, and an angle-dependent adaptive algorithm like ATI uses, for increased speed. Whether they force bilinear along with that or not, I don't know, but I haven't heard any mention of that happening.

Well, for one thing, it's not going to switch to an angle-dependent algorithm, as that's based upon the anisotropic degree selection algorithm, which is hard-wired into the core. I don't think that's going to be changed.

I think that the more likely thing is that the maximum degree of anisotropy is being limited, in some situations. Right now it doesn't appear that the "aggressive" mode is very useful. An examination with more mature drivers later would be nice. I would guess that the goal of this mode is to disable or reduce anisotropic for textures that shouldn't need it (such as lightmaps). Right now it doesn't look any good.

 

[KimB]

Has the algorithm really changed between NV25 and NV30?

Well, the Performance/Aggressive switch doesn't appear to make any difference on my GeForce4, in benchmarks or visual quality (I used Quake3 for the tests....pardon me for not posting shots, but they are rather uninteresting).

 

[Ailuros]

Has the algorithm really changed between NV25 and NV30?

Well, the Performance/Aggressive switch doesn't appear to make any difference on my GeForce4, in benchmarks or visual quality (I used Quake3 for the tests....pardon me for not posting shots, but they are rather uninteresting).

Exept lightmaps; I suggest you try some benchmarks between application and either balanced or aggressive.

Application= RT quality mode
Balanced/Aggressive= RT performance mode

 

[Althornin]

Well, for one thing, it's not going to switch to an angle-dependent algorithm, as that's based upon the anisotropic degree selection algorithm, which is hard-wired into the core. I don't think that's going to be changed.

How do you know this?
why do you think this?
evidence?

 

[THe_KELRaTH]

I think Aggressive is point sampling, balanced is Bilinear and the adaptive texture alogorithm is that, as well as using ATI's - use where necessary, it also drops down to point sampling to keep the fps up. Some may remember they showed this off in 40.41 drivers. I wouldn't be suprised if that was a test to see users reaction.
With PS setting on it also increased the 3Dmark score.

The problem here is that if you try to compare ATI's AF and NV FX using an ingame screenshot you are not going to see the real difference. ATI's adaptive alogorithm is based on using a predefined quality and then using it where necessary so when you take a screenshot it's a still of what you are seeing in motion.
NV's FX version is rather different as although it's similar to ATI's in respect of only using AF where needed it also changes in quality too, therefore in motion it could easily drop to point sampling but when taking a screenshot what you get is max quality AF (based on your driver setting).

I also think this is why they didn't use standard naming conventions as reviewers would quickly dismiss aggressive mode if it were called point sampling and balanced "aka Bilinear" would only be compared to ATI's bilinear mode and not ATI's trilinear mode.

Looking at the examples on Anandtech, specifically the x8AF NV Balanced v ATI x8 Performance I would also wonder if ATI and NV have a different definition on "Using it where Necessary" as it can be seen clearly that ATI's version is having more effect.

 

[ram]

Below are some Geforce FX comparison screenshots made by PC Games Hardware prior benchmarking with a cool tool developed by Xmas. The tool is available for download here and is able to show the degree of anisotropy in relation of the polygon angle. The colours show the different mip maps. As the lod bias and therefore the change of the mip map versions usaually is tied to the degree of anisotropy, "closer" mipmaps mean a lower degree of anisotropy.

As you can see, the Geforce FX "balanced" and "aggressive" modes seem to work angle dependent, but only on some mip maps. The R300 in contrast works angle dependent on all mip maps. The "old" application mode is the same as on the NV2x series, which is slow but offers the highest quality.

The "aggressive" mode, while not switching back to bilinear, is doing something close to this. The interpolation area between the mipmaps is quite small. Nvidia advices to use the aggressive mode for comparison benchmarks to the R300 performance mode.