Nvidia GF4 Aniso and FSAA

[KimB]

Try again, please...I think my firewall was messing things up (It tends not to give permission to programs started before the firewall...only programs started after the firewall...).

 

[SvP]

OK I spent about a minute switching between the images, but didn't see the difference.

I mostly looked at the end of the floor where's a wall. The transition form floor to wall is equally blurry in both shots. It definitely should't be.

There are couple of short-sized lines near the above-mentioned wall; the length of "lines" between floor-plates do not change. (There seems to be one where the road turns right. The length of it is about third of the plate in both shots).

Could you make another images where there's a red circle around the area where the difference is?

 

[KimB]

Look at the walls on the sides of the floor.

 

[MikeC]

Some clarifications:
Quincunx does not appear to blur textures in the least on my GeForce4

When I compared antialiasing on the GeForce4 Ti 4200, I found that Quincunx offered better image quality under OpenGL versus Direct 3D. Compare the IL-2 Sturmovik PNG screenshots in the preview which were from the 28.32 Detonator drivers.

http://www.nvnews.net/previews/geforce4_ti_4200/

When I previewed the GeForce4 Ti 4600 back in February using the 27.30 Detonator drivers, Quincunx looked like it did on the GeForce3.

 

[SvP]

OK, I tried to make these images equal-sized.

http://svp_1.tripod.com/4x-chal.jpg
http://svp_1.tripod.com/8x-chal.jpg

From these images it seems that there is something different between them. It could be due to better filtering, but it also could be that the "noise" is simply in different places Definitely not something I'd call working, functional, good enough and "oh-so-much-better-than-anything-else" filtering

 

[KimB]

Well, I made a custom level that would show a more obvious difference, here it is:

<a href=http://169.237.254.233/atest0x.jpg>No aniso</a>
<a href=http://169.237.254.233/atest2x.jpg>2-degree aniso</a>
<a href=http://169.237.254.233/atest4x.jpg>4-degree aniso</a>
<a href=http://169.237.254.233/atest8x.jpg>8-degree aniso</a>

If you'll note, I did my best to pick a worst-case texture for texture filtering, and there are noticeable moire effects. I would be very interested to see how a Radeon 8500 does in this situation. Just open the level, and take the shot (press F9) without moving the mouse. I messed up in the start position, so you can't even walk around...(You'll have to turn "Tournament Mode" on) but it works well enough for visual comparisons. Here's the level:

http://169.237.254.233/AnisoTest.zip

 

[SvP]

Now I see it

Edit: if only the impact on image quality in real games would be in line with performance impact...

 

[Sharkfood]

Um, no, this is the way anisotropic works. Here's a link to how it's implemented in OpenGL:

Uhm, yes, it is how anisotropic filtering works. Here's a better link to the definition:
http://www.sgi.com/software/opengl/advanced98/notes/node37.html

This has nothing to do with subjective viewing of the images. Some images simply do not need higher degrees of anisotropic filtering. ... (rest of long winded excuse making omitted)

You are absolutely right- if a texture consists of nothing but a solid color, then anisotropic filtering will serve no purpose. But I don't know of anyone that would waste texture mem when shading would be simple for the case you have described.

The bottom line is- anisotropic filtering *is* broken on the Gf4 in Direct3D. There is no amount of flag waving, superscaled, 800x magnified pixel-movements from a different observation viewing angle or whathave you involved. Simply revert a GF3 back to the 12.90 drivers and this much becomes painfully obvious... It was functioning properly at that stage (again, in Direct3D).

Your UT level shows exactly what the problem is- it's either a bug or a cheap method to tune performance... if I take your word for it and assume you took those shots in Direct3D... seems a safe bet since 8x is at least working to some degree in OGL and would likely be a tad more pronounced. I'm too lazy to deal with the *forever* install of UT on the GF4 box, although I do have it on the rig with the 8500:
http://shark_food.tripod.com/modrugs/modrugs.html

Quite the posterchild case for moire-patterned texture, but the 8500 has a clear effect moving the isolinear point of detail with each new gradient level of anisotropic filtering applied. The GF3 used to do fairly similar with older drivers as well.

Cheers,
-Shark

 

[LLB]

Comparing the 8500 16x shot with the GF 8x on Chalnoth's UT level it's apparent that the GF4 is doing a much better job regarding the amount of moire that is visible.
Frankly the 8500 looks pretty appalling in comparison.

Why is there so much moire with the 8500?

LLB

 

[Althornin]

Because its only bilinearly sampled aniso, not trilinearly sampled.
l

 

[Bambers]

Atis agreesive LOD bias adds to it as well.

 

[KimB]

You are absolutely right- if a texture consists of nothing but a solid color, then anisotropic filtering will serve no purpose.

It doesn't even take that. What it does take is an image that doesn't have steeply-angled textures. In Morrowind, for example, you'll probably find it rather hard to discern the higher levels of anisotropy when outdoors due to the limited draw distance.

The bottom line is- anisotropic filtering *is* broken on the Gf4 in Direct3D. There is no amount of flag waving, superscaled, 800x magnified pixel-movements from a different observation viewing angle or whathave you involved.

And yet the shots I posted show the contrary...do you have a GeForce3 that you'd like to compare to prove your point? Btw, here's some OpenGL shots:

<a href=http://169.237.254.233/anisogl4x.jpg>4x Aniso</a>
<a href=http://169.237.254.233/anisogl8x.jpg>8x Aniso</a>

Yes, there is slightly more difference in the OpenGL shots, though I think that's just because of slightly different LOD settings. I'll look at it a bit more closely soon, but it currently looks like the Direct3D shots show slightly higher-detail MIP maps.

Update: Yes, it most definitely looks like this is the case.

Quite the posterchild case for moire-patterned texture, but the 8500 has a clear effect moving the isolinear point of detail with each new gradient level of anisotropic filtering applied. The GF3 used to do fairly similar with older drivers as well.

One thing that's interesting to me is that most of the texture aliasing occurs straight ahead in this shot, not to the far left or right. I imagine you'd see the exact opposite scenario if I angled the shot along either the texture U or V line, as we've seen before on Serious Sam screens.

 

[Hellbinder]

I really was not trying to insite yet another riot between Nvidia and ATi....

What I really wanted to know about Aniso, is what if anything can Nvidia do to speed up their implimentation. Or if they could via drivers *hack* a faster method.

As it stands now though, looking at 4Xs, who needs Aniso? I actually found that I like Types screens with straight 4xs, more than 4xs-Aniso. Perhaps in motion it makes a bigger difference.

I have officially ordered a MSI GF4 Ti 4200. It will be here on friday, it is being pared with my new P4 2.4ghz (533) system arriving on thursday. I am giving my Radeon 8500 to my sister for now. Mostly becuase I need Video In to use with my new JVC 920u Digital camcorder.

To tell you the truth, I am really excited about seeing the card in action once i overclock it to about 295/295. Just in time for the UT2003 demo comming in a few more days. :

 

[KimB]

4xS isn't as effective on textures as anisotropic. I'm reasonably certain it's kind of like 2x MSAA + 2x SSAA. So, it should be roughly equivalent to 2-degree anisotropic.

Additionally, the speed hit for 4xS + aniso is pretty bad compared to the other FSAA forms + aniso.

 

[John Reynolds]

I just got my new rig up 'n running last week (a P4 2.4B, 512MB of PC1066 RDRAM, and a GF4 Ti4400) and I'm finding that I really like 4xS and 2x aniso (LOD at -05 to -1, depending on the game) in Morrowind, Sacrifice, Freedom Force, Independence War 2, and Aliens vs. Predator 2. And I often use those two settings in conjunction with high resolutions, such as either 12x9 or 12x10, while still getting very playable frame rates.

 

[SirPauly]

That certainly is a tough texture but adding ATI's AA helps clean up texturing artefacts and when coupled together does a solid job on this worse-case example while moving.

Here is a shot with x4 AA and with x2 AA with anisotrophy.

x2 AA:

http://sirpauly.homestead.com/files/snap020.jpg

X4 AA:

http://sirpauly.homestead.com/files/Anisotrophy.jpg

 

[ben6]

Actually I got a interesting description of 4XS. Remember last year at E3, where Nvidia said they might go 4X RGMS , well this is basically it. The S stands for Shifted . They shifted the position of the subpixels so it's almost a rotated grid and added another sample. That's Tony Tamasi's description of 4XS.

 

[Dave Baumann]

I think we've been over the sampling pattern of 4xS many times now - its 2X vertical superampling each with 45 degree 2X multisampling. When the samples are combined it gives a slightly shifted sampling pattern edge sampling pattern and the 2X vertical supersampling accounts for the extra texture detail.

Ram has a digram that illustrates this and we've also shown it at 3DGPU with Basic's FSAA tester. I would post the images up again but its v late and I'm off to bed!

 

[KimB]

That certainly is a tough texture but adding ATI's AA helps clean up texturing artefacts and when coupled together does a solid job on this worse-case example while moving.

Too bad it still doesn't completely clear 'em up, and ATI's FSAA isn't very useful because of its performance.

ATI *really* should have done a much better job on the MIP LOD selection. Of course, nVidia's isn't perfect, either, as you can see from my screenshots.

 

[Sharkfood]

LLB-

Why is there so much moire with the 8500?

Because the 8500 doesnt make performance compromises with LOD Bias like the NVIDIA cards do. We already had this discussion earlier. I took shots at default LOD Bias for both cards, but this can, of course, be reduced down.

Chalnoth-

In Morrowind, for example, you'll probably find it rather hard to discern the higher levels of anisotropy when outdoors due to the limited draw distance.

Reduced Z-distance is grandstanding, old chap. It shouldnt take a mipmap a million miles away at 200x magnification to see the difference between 4x and 8x anisotropic filtering. It surely doesnt in OpenGL regardless of the textures (barring full-solid color textures, of course).

And yet the shots I posted show the contrary...

The shots your provided show the EXACT same thing- little to no change by doubling the anisotropic filtering samples. If it takes a mipmap so totally out of clipping range in a game without clipping range to show the effectivity of doubling texel samples, this proves the point quite well actually.

Especially given how the GF3 in early drivers, the 8500 and SGI/USparc graphic workstations do not exhibit this "GF4-only" behavior. Just the GF4.

Too bad it still doesn't completely clear 'em up, and ATI's FSAA isn't very useful because of its performance.

You really need to try an 8500. From your commentary it's plainly obvious you have *never* used, seen or operated an 8500. This much is clear. And from that basis, I discount your opinions as such from someone that has zero experience, and this rings in more and more true with every post you place.

John-

and I'm finding that I really like 4xS and 2x aniso (LOD at -05 to -1, depending on the game) in Morrowind, Sacrifice, Freedom Force, Independence War 2, and Aliens vs. Predator 2. And I often use those two settings in conjunction with high resolutions, such as either 12x9 or 12x10, while still getting very playable frame rates.

I love the quality and use the same settings, but just hate the performance. I'm really waiting for the R300/NV30 in hopes to see about 15-20% improvement with similar visuals. That would most definately be a "sweet spot" if we could have this level of visuals and 40+ fps to boot.