Is S3's AF dodgy?

[vnet]

What do you think of S3's AF?
I saw at Techbits some screenshot comparison of S3's AF compared to the GeForce 4 Ti4200 and the Radeon 9600 and the S3's AF looks very weird indeed, to say the least.
Below are some screenshots

Left: GeForce 4 Ti4200 | Center: S3 Omnichrome | Right: Radeon 9600

1x AF

2x Anisotropy

4x Anisotropy

8x Anisotropy

16x Anisotropy (no GF4)

Is it just me or is there something weird going on? Their AF looks ugly. The color doesn't look nice and it seems to be just affecting the LOD only, the GF4 or 9600 AF look much better.

 

[Ostsol]

Actually, it looks pretty good. . . except for the fact that it's bilinear filtered. . .

 

[vnet]

I thought we were passed the bilinear-AF from the 8500 era....

 

[Scali]

IIRC, S3 uses a different filtering method for trilinear, similar to the one used in the Kyro at the time.
This may also have its effect here, since it takes more samples from the higher LOD texture, meaning you get more white and less other colours.
These screenshots are utterly useless for judging the actual quality delivered by the filter, which is far more interesting than looking at which LOD it takes its samples from. Especially since the colours indicate that at higher AF levels, more samples are taken from the highest level than with other cards, it could well be that the S3 is giving higher quality than the others, rather than lower quality.
You'd need a tool such as the 'tunnel viewer' in 3DMark to see the actual quality of the filter.
See this article: http://www.xbitlabs.com/articles/video/display/s3-deltachrome_13.html

 

[Dio]

S3 had, in the Savage3D/Savage4 days, a trilinear filtering algorithm that is incompatible with colour mip levels (which is not to say it is necessarily of lower quality; it is of equivalent quality to some typical trilinear filtered cases).

I do not know if they are still using it now, but it might explain some of the differences in those screenshots.

 

[Dave Baumann]

Fast Trilinear - they patented it, I can't see them not using it!

 

[vnet]

Left: GeForce 4 Ti4200 | Center: S3 Omnichrome | Right: Radeon 9600

2x Anisotropy

 

[OpenGL guy]

Left: GeForce 4 Ti4200 | Center: S3 Omnichrome | Right: Radeon 9600

2x Anisotropy

http://www.techbits.ca/images/reviews/omni/omni_2x_s.jpg
http://www.techbits.ca/images/reviews/omni/atiaf2x_s.jpg
http://www.techbits.ca/images/reviews/omni/gf_2x_s.jpg

You have the order mixed up. According to the names of the images, it's S3/ATI/NVIDIA. Also, why does the S3 shot look like trilinear and the others like bilinear?

 

[Scali]

Again, I ask what the purpose is of these screenshots.
They do not show the quality of the filtering, they merely show where the texture units take their samples from.
I get the feeling that this thread exists mainly to discredit S3, based on false information. For example:

Is it just me or is there something weird going on? Their AF looks ugly. The color doesn't look nice and it seems to be just affecting the LOD only, the GF4 or 9600 AF look much better.

This is complete nonsense, since you cannot see anything about the quality in the presented screenshots, as pointed out before.
This sort of thing annoys me.

If you take a look at the link to the xbitlabs article I've posted earlier, you will see that while the mipmap-access pattern looks as if there is no trilinear filtering going on, there are no mip-bands visible in the actual texturing on screen. So there is indeed trilinear filtering applied.
The same goes for AF, the only difference between S3 and ATi/NV seems to be a slightly higher LOD bias on S3's side. I suppose it's rather subjective whether you think the higher or the lower bias is better quality, and perhaps you can adjust that in the drivers too.
Unlike ATi (and NV4x?), S3 also doesn't have the 'inconvenient angle' problem, and unlike NV, it doesn't use brilinear filtering, so that's plus in terms of filtering quality.

In short, I suggest you learn how to use the tools before you try to pass judgement.

 

[Rolf N]

The screenshots are dodgy
I have seen mipmap transitions in both of these AF testers (OpenGL and Direct3D versions) on a DeltaChrome myself. The young padawans at Techbits should have bothered with the settings for a few seconds longer, and maybe then they would have seen them, too.

That "fast trilinear" thing is indeed implemented in the DeltaChrome, but it's not "always on". The chip is capable of "classic" trilinear as well (two cycles), and the driver appears to do some decisions based on texture content.

The AF isn't perfect IMO. Tthe GeforceFX line (with disabled "optimizations") does it slightly better.
But you won't be able to see the reason for this opinion of mine in these shots. For that you'd need to look at it in motion.