More Texture Filtering

[Dave Baumann]

Its "Guess what generated these" time...

 

[Rasmoo]

Still stuck with an R8500 & a Ti4200, so these are wild guesses of course

As the pictures appear:
R3xx (good and fuzzy) - GF FX (quick progression, short transitions)
GF FX
R3xx (longer transitions)

 

[Dave H]

My guess...

...GF4Ti with 44.03 drivers.

First pic demonstrates that it's not doing trilinear filtering of detail textures in UT2003 Quality mode, which is the interesting piece of news Dave wants us to note. Second and third I'm not positive about, but I'm guessing Performance and Quality (with no AF), demonstrating that Performance is almost-bilinear and Quality is trilinear...with the exception of UT2003 detail textures.

Can we turn this one into a contest with 9800 Pros as the prizes too?

 

[vb]

Why aren't they getting the same mip map level?

is it a LOD Bias problem?

 

[Dave Baumann]

The LOD differences are because UT2003 has different LOD biases for the default and detail textures, and the drivers are interpretting these correctly (which is a clue).

 

[K.I.L.E.R]

Dave impies that the stuffed up pics are nVIDIA's and the good ones are Ati's or vice versa. Thanks for the clue Dave.

 

[Dave Baumann]

All images are from the same board.

 

[Dave Baumann]

For reference, here' one from the "Balanced" (performance) mode taken from our NV30 preview:

 

[K.I.L.E.R]

Let's see:

filt2.jpg: bilinear
filt3.jpg: trilinear


Whoops - sorry KILER, pressed the wrong button. Didn't mean to kill your message!! :? DB.

 

[Rasmoo]

How many guesses do we get then?

NV30 [Added] / NV31 / NV35
UT2003 done with 43.45 drivers (not the "evil" 44.03 drivers), varying texture level (dunno what exactly to expect, so I'll just copy Dave B here)
AF shot 1: Aggressive mode
AF shot 2: Balanced mode

[Added] Would there be any way of distinguishing the different NV3x's?

 

[vb]

nv35

same driver

trilinear on texture 0,
"improved bilinear" on detail texture
(or the other way around)

 

[jb]

Gee Dave, I did not realize how much of a difference one texture can make ie in that UT test level I thought you migh see some difference...but thats a lot more than "some"

 

[Dave Baumann]

Both are Trilinear, this is what Bilinear would look like.

 

[Dave H]

I've got it! All are NV3x. The first is UT2003, Quality, no AF, demonstrating the near-bilinear on the detail textures.

The third is Quality, no AF, with Xmas' filtering test tunnel.

The second is Quality, no AF, with Xmas' filtering test tunnel.....






......hacked to trip the driver's application detection of UT2003!!!!!



(Man, if I got this right I'm not settling for anything less than the Shuttle PC plus the 9800.)

 

[Dave Baumann]

No. These are from a 9600.

The point I'm demonstrating is that ATI do have a similar feature to NVIDIA's Bi/Tilinear mix functionality. I can't find it enabled under R300/R350 but RV350 does have this under DirectX (of course, there was also a less advanced form in R200). Anyone with a 9600 can play around with the texture preference slider and it will do virtually exactly the same thing as NVIDIA's does, except if you have it at high quality it will alway do full Trilinear (at least, AFAIK - I've not seen an instance where it doesn't as yet).

So when [H] concludes "NVIDIA has seemingly found a way to do less work doing Trilinear Filtering than ATI while producing an IQ that easily comparable with ATI's.", with the 9600 this isn't the case, since it probable that a similar IQ level can be found.

Brent - in your upcoming 9600 review will you be investigating the texture quality slider under D3D to see if you can get a similar UT2003 quality to NVIDIA and then benchmarking them that way?

 

[Rasmoo]

No. These are from a 9600.

Golly. And those AF shots look so much more like the NV30 shots than the R350 shots. All done by a texture preference slider?

I guess it makes my condition of "assuming the R3xx hardware is capable of mimicking NV3x behaviour" (different thread) always true.

 

[rubank]

I fiddled around with these settings in D3D AF tester a week ago, and I saw the same result. 9600pro and Cat 3.4.

I must say, that the NV30 AF test Dave put up for reference looks much worse than the comparable 9600 one, at a closer look.
I don´t know if this has anything to do with the one taken from D3D and the other from OGL though.

 

[WaltC]

No. These are from a 9600.

The point I'm demonstrating is that ATI do have a similar feature to NVIDIA's Bi/Tilinear mix functionality. I can't find it enabled under R300/R350 but RV350 does have this under DirectX (of course, there was also a less advanced form in R200). Anyone with a 9600 can play around with the texture preference slider and it will do virtually exactly the same thing as NVIDIA's does, except if you have it at high quality it will alway do full Trilinear (at least, AFAIK - I've not seen an instance where it doesn't as yet).

I think the bi-tri mix is perfectly OK for any IHV when it's restricted to a "performance" setting in which the end user is aware he's running in a performance IQ mode. The thing about Det behavior in UT2K3 that I dislike is the fact that even when the application calls for full trilinear the Dets don't give it--but provide a "performance blend" instead.

So when [H] concludes "NVIDIA has seemingly found a way to do less work doing Trilinear Filtering than ATI while producing an IQ that easily comparable with ATI's.", with the 9600 this isn't the case, since it probable that a similar IQ level can be found.

What I object to about the [H] statement above is that it characterizes the behavior of the Dets running UT2K3 as though it was universally applicable to all 3D games. In a game other than UT2K3, a game in which the Dets actually provide full trilinear support, are the Dets "doing less work" than the Cats? I can't see how that would be the case, since in that event the Dets would be doing full trilinear just like the Cats. Sure, in UT2K3 the Dets are doing less work because the drivers are doing less trilinear filtering--and so performance is better. What is remarkable about that? Brent...?

 

[Myrmecophagavir]

I think the bi-tri mix is perfectly OK for any IHV when it's restricted to a "performance" setting in which the end user is aware he's running in a performance IQ mode.

Does that include ATI's "Quality" AF control panel setting, which only does trilinear filtering on the first texture stage?

The thing about Det behavior in UT2K3 that I dislike is the fact that even when the application calls for full trilinear the Dets don't give it--but provide a "performance blend" instead.

So the same could be said for the Cats...

 

[Ostsol]

I think any such forms of optmization are okay as long as the drivers provide the user with the option of turning them off. The reported lack of an "application preference" option in NVidia drivers is unacceptable.