trilinear, anisotropic filtering?

[CJP]

Are trilinear and anisotropic filtering related? How should you configure them to work right on a fast system?

 

[gururu]

they are different modes for AF.
trilinear is higher quality AF. the other option is bilinear, which is of lesser quality.
performance settings will force bilinear, whereas quality settings will enable bilinear and trilinear
depending on what the application chooses to use.

 

[Metalloid]

I think that bilinear, trilinear, and AF are all different. Someone please correct me if I am wrong.

Bilinear being worst quality and AF being best.

 

[BFG10K]

Bilinear, trilinear and anisotropic are all filtering modes.

• Bilinear works by sampling 2 x 2 = 4 texels on the x/y axis. It serves to smooth out the texels on flat surfaces
• Trilinear works by sampling 2 x 2 x 2 = 8 texels across the x/y/z axis. It's like bilinear filtering except it also works with depth (z) to create smoother blending between the different texture sizes (mip-maps).
• Anisotropic filtering samples 16/32/64/128 texels depending on the setting and eliminates blurriness and distortions on high angled polygons because it's a non-linear form of sampling.

Bilinear and trilinear are linked and you can only have one or the other at once as strictly speaking bilinear is a subset of trilinear. Anisotropic filtering OTOH is independent of the two and can operate at any setting while either of the others operate as well.

In terms of speed and image quality, from fastest to slowest, and from worst quality to best:
(1) Bilinear.
(2) Trilinear.
(3) Bilinear and anisotropic.
(4) Trilinear and anisotropic.

What settings you use depends entirely on the game and on your system.

 

[CJP]

Thanks BFG, Gururu, and Metalloid.

 

[BenSkywalker]

Anisotropic filtering samples 16/32/64/128 texels depending on the setting and eliminates blurriness and distortions on high angled polygons because it's a non-linear form of sampling.

Actually, 8 sample anisotropic is also available on all modern hardware AFAIK(certainly is on the R300 boards- 2x Performance is 8 sample). 6, 4 and 2 sample anisotropic are also possible utilizing point filtering(not that anyone would want to, but it is possible).

Anisotropic simply means that you are sampling in a non isotropic(non sqaure) sampling pattern. Anisotropic should reduce texture aliasing('shimmering') significantly on any really good implementation. The Radeon9500/9700/9800 and GFFX line of boards do not have very good implementations, the NV2X boards do, although they are significantly slower and don't offer as high of settings that the R300 core boards do, the R200 core boards have horrendously poor anisotropic(NV1X boards are pretty much the same implementation as the NV2X boards but slower still and with even lower settings).

 

[BFG10K]

Actually, 8 sample anisotropic is also available on all modern hardware AFAIK

Well it probably is but it's largely a Red Herring, just like point filtering is. Even if you could get it to work it would look practically the same as no anisotropic at all.

2x Performance is 8 sample).

Only at certain Z-angles; at 90 degrees and 45 degrees it'll be at full strength (16 tap). Non-R300 boards will only be at full strength at 90 degrees.

This is the nature of ATi's adaptive anisotropic and the performance/quality doesn't factor into the equation at all. That setting affects bilinear/trilinear sampling for the anisotropic filtering, not the sampling strength across the different angles.

6, 4 and 2 sample anisotropic are also possible utilizing point filtering(not that anyone would want to, but it is possible).

Using point filtering on a 3D accelerator is an utter travesty.

The Radeon9500/9700/9800 and GFFX line of boards do not have very good implementations,

I think the R300's implementation is by far the best in terms of performance and image quality ratios. All GF cards are either butt ugly or far too slow. In my experience the R300's 16x performance mode looks far better than the Ti4600's 8x full mode and believe me, I've run a lot of tests in a wide variety of games. The R300's performance mode is almost perfect and it's extremely rare for it to exhibit any problems at all and when it does they're extremely minor. Plus it's amazingly fast.

the R200 core boards have horrendously poor anisotropic(

Well yes and no. I agree that on some occasions certain surfaces look crap but at the same time the algorithm makes it lightning fast on boards that don't have the same level of power as R300 chipsets, meaning all Radeon users can enjoy full anisotropic filtering for almost free. And think how advanced it was when the original Radeon could use 16x anisotropic filtering for practically zero performance hit. This was something that most consumers had never really seen or had access to before.

 

[BenSkywalker]

Only at certain Z-angles; at 90 degrees and 45 degrees it'll be at full strength (16 tap).

Performance 2x is 8 sample at full strength. Performance mode is bilinear.

I my experience the R300's 16x performance mode looks far better than the Ti4600's 8x full mode and believe me, I've run a lot of tests in a wide variety of games.

The R300 has very noticeable texture aliasing running the higher levels of anisotropic filtering. Reduction of texture aliasing is supposed to be the primary concern of anisotropic, increasing the LOD bias is secondary. If you adjust the LOD bias to the point where noticeable aliasing is gone, it offers less crisp textures then the NV2X boards. Performance mode is horrible IMO. It makes me feel like I'm playing on a Voodoo3, bilinear filtering isn't something I find acceptable in 03.

 

[Genx87]

NV30s Application mode is the best bar none. Their Balanced and aggressive modes have issues.

R300s 16x AF basically gives you 16x @ 90, 8x @ 22, and 2x for all other angles. Eventually when you start to see more curved surfaces this will probably catch up to the R300.

 

[gururu]

ok, i may stand corrected, but someone please verify if this is true.
I based my response on the Rage3d tweak which states word for word under Direct3D for
anisotropic mode:
"Performance mode forces bilinear sampling for AF. Quality mode supports Bilinear or Trilinear
sampling based on the settings in the application."

also, it'd help if any clarifications be clearer than those attempted in the thread thus far, because I
know not much about this subject.

 

[BFG10K]

Performance 2x is 8 sample at full strength.

I see what you mean now. Sorry, I was thinking about something else. However the R300's method must be doing something extra because the mip-map lines are near non-existant compared to previous Radeon boards (a lower LOD perhaps).

The R300 has very noticeable texture aliasing running the higher levels of anisotropic filtering.

Which is usually cured by high resolutions and/or reducing the LOD slider. I use the former.

Performance mode is horrible IMO. It makes me feel like I'm playing on a Voodoo3, bilinear filtering isn't something I find acceptable in 03.

99% of the time I can't see the mip-map lines at all and if I do see them, they're very slight. On pre-R300 boards this was an issue sometimes but R300 boards rarely have this issue at all.

The 9700 Pro running under 16x peformance always looks far better than a Ti4600 running at 8x in my experience.

NV30s Application mode is the best bar none.

Nope, because it's far too slow.

R300s 16x AF basically gives you 16x @ 90, 8x @ 22, and 2x for all other angles.

That's not true at all. I've tested full 360 degree rotations on my card and I can tell you that it's not 2x at other angles.

It's 16x at 90 degrees and 4x at 22.5, and any angle in between it's between those two values. I've also tested a wide range of games with heavily uneven floors and I never saw anything resembling 2x when using 16x.

"Performance mode forces bilinear sampling for AF. Quality mode supports Bilinear or Trilinear
sampling based on the settings in the application."

That's correct. If you leave the setting on quality you can control bilinear or trilinear through the application. If you use performance the card always drops to bilinear.

 

[Genx87]

"Nope, because it's far too slow."

I was talking image quality.

16x AF on R300

FX 8x AF

Notice the jaggies on the R350 while the FX is a cylinder under application mode.

"As we progress through the levels of anisotropic filtering we can see that only certain angles get the full levels of filtering on the R350, as it was with R300. We can see that the 90 degree angle portions have the full level of filtering applied, while the 22 degree portions are left at about 8x; the rest of the filtering appears to be at about 2x."

R350 review

 

[BFG10K]

The Beyond3D tests only rotated around the Z-axis and only rendered a tunnel which appeared to have no depth at all. This is a very one-dimensional test and never happens in a proper 3D game, unless of course you're staring right in front of a flat wall.

A far better test would have been to use the tiled plane, increased the distance to maximum, and then rotate it 80 degrees around the Y-axis so that it was giving the impression of real depth.