Tom's GF4 anisotropic scores in his Parhelia review

Wavey, if you mean "8x" as per the game itself then that is correct. But that isn't what it should be for a GF3/GF4. It should be "64" max for the cards. Try this on a Radeon and see what you get (i.e. if you can only get "16" max or higher).

Croteam would never allow "128" as an option (which is an option)... they're not looking that far ahead thinking there will be cards with 128 degree of aniso.

Croteam uses the term "taps" for the 2, 4, 8, 16, 32, 64 & 128 figures. That means number of point samples. Not degree.

[edit]It used to be the case of "degree" for the "Texture Anisotropy" in the original game. The max then was up to "8" for the option for a GF3/GF4. After talking to Croteam for a while over this back in that time, the subsequent patch changed this to 2, 4, 8, so-on-so-forth. In the original game, you could check a bunch of OpenGL driver info by typing /OpenGLInfo() among which the console output will tell you the max aniso degre of the hardware. This variable doesn't seem to available in SS:SE however.
 
Chalnoth said:
Now that I looked at the 2x shots (I didn't before...I saw no reason to...I only compared the 8x shots), I do see very slight differences. What I did see was a few black 'specks' in the 27.51 shot that weren't there in the 29.42 shot. Whether this is more or less correct rendering, I do not know. I couldn't see any difference in texture clarity.
Click to expand...

Yeah, those black spots on the sides of the door bother me, whether it should be there or not is the question.

I dont have this game on my 8500 for a screen shot, so if anyone wants to chime in with it.
 
Croteam uses the term "taps" for the 2, 4, 8, 16, 32, 64 & 128 figures. That means number of point samples. Not degree.
Click to expand...

Well, what they say (or you’ve interpreted them as saying) they are doing is different from what is actually happening as far as I can see, as what appears to be happening is consistent with what I have described above.

If you select a value greater then the degree supported by the card then it default back to the maximum. i.e. if you go into the advanced options and try and select ‘64’ exit and go back into it you’ll see that its gone back to 8 (8x being the maximum GF3/4 can support). Likewise if you do it on a Radeon 8500 the Extreme addon will select ‘16’, if you try to select greater than that then if you go out and back into the advanced options you’ll see it has defaulted back to ‘16’. So it would appear the game is consistent with selecting the degree of filtering rather then the number of ‘taps’ – taps wouldn’t be a consistent numbering system anyway since it also varies dependant on whether you use it in conjunction with Trilinear or Bilinear (this would further be compounded with Radeon only being able to do Aniso in bilinear mode).

[edit] If you look at the 'Tooltip' at the bottom of the page that decribes the option as you hover over it it also states "Sets the Degree of Anisotropy".
 
Ailuros said:
What doesn't make sense here is that looking at your screenshots it's apparent that 29.42 has offset LOD quite a bit further than 28.32. Normally more aggressive (negated) LOD settings hurt performance they don't increase them.
Click to expand...
It depends on the situation. If you can completely avoid minification, then you only need 1 mip level (i.e. the top one). This can be a large savings...

The tradeoff will be increasing texture aliasing, as someone else noted.
 
I'm pretty certain that using MIP mapping actually helps with graphics card performance, as it drastically improves texture cache hits (which would be consistent with the fact that decreasing the LOD Bias, which brings the rendering closer to using only one MIP level, decreases performance).
 
The Radeon and Radeon8500 can go up to 16 bilinear filtered pixels from the source texture for its implementation of anisotropic filtering, ala 16bi X 4samples/Bi = 64 taps. Giving a possible 16:1 anisotropy ratio (if conditions are right).

The GF3/4 can do up to 8 trilinear filtered pixels from the source texture for its implementation, with 8tri X 8samples/tri = 64 taps. Giving a 8:1 anisotropy ratio.

Both can do 64 taps where taps means the number of samples taken from the source texture before rendering a pixel on the screen. In other words the GF3/4 will use 64 texture samples for one frame buffer pixel.

That is how I understand it. Both are capable of 64 taps except the way that it is done is somewhat different. Oh, anisotropy ratio is basically how much you can reliably stretch a square 2d image (texture) onto a none square polygon.
 
But it's obvious that nVidia is compensating for off-angle surfaces, and ATI is not. Additionally, ATI's method has far more texture aliasing. Given the tiny performance hit of ATI's method, and the large amount of texture aliasing in some circumstances, I really believe that it is using fewer taps than nVidia's method.
 
But it's obvious that nVidia is compensating for off-angle surfaces, and ATI is not. Additionally, ATI's method has far more texture aliasing.
Click to expand...

Is that really the case given that nVIDIA now are pushing LOD back as well?
 
Dave, please try another game before drawing that conclusion. Increased LOD was certainl not duplicated in the shots Reverend posted.
 
Chalnoth said:
I'm pretty certain that using MIP mapping actually helps with graphics card performance, as it drastically improves texture cache hits (which would be consistent with the fact that decreasing the LOD Bias, which brings the rendering closer to using only one MIP level, decreases performance).
Click to expand...
This is not always the case, which is why I made my original post.

Let me give you an easy example: trilinear filtering. If you need two levels to do the filter (as nvidia does), then it's quite expensive, but if you shift the LOD so that you minify less frequently, this can help performance as you now only need one texture level in those cases.

Get it now?
 
That depends on two things:

1. Is the graphics hardware capable of doing the same number of trilinear samples as bilinear samples?
2. Exactly how much cache efficiency is lost when removing MIP mapping altogether.

Given the little to no performance hit we've seen from enabling trilinear filtering on modern gtraphics hardware, I really doubt that dropping to one MIP map level would help much, if at all.
 
I really believe that it is using fewer taps than nVidia's method.
Click to expand...

I agree since ATI method also has some AI in not sampling everything at 64 taps - - like walls that are facing you. Still ATI does have a limitation to their implementation when the final rendered polygon angle is approaching 45 degreas rotated clockwise or counterclock wise from view point.
 
Chalnoth said:
Given the little to no performance hit we've seen from enabling trilinear filtering on modern gtraphics hardware, I really doubt that dropping to one MIP map level would help much, if at all.
Click to expand...
I guess I need to spell things out...

My example was for trilinear filtering, but this discussion is about anisotropic filtering. If you need to use multiple texture levels in order to do anisotropic filter and mip filtering (AKA trilinear+anisotropic) then my comments also apply.

I didn't think it was so difficult to figure out.
 
I would think that continually increasing the degree of anisotropy in order to keep the texture from aliasing without MIP mapping would impose much more of a performance hit...

But, it may not be a bad idea if graphics hardware is capable of not enabling trilinear filtering until that second MIP map is used (I'm not sure if current hardware optimizes for this situation...that is, if it always uses up the fillrate required for trilinear, even close to the viewer...although it does make sense that there's less of a memory bandwidth hit when only one MIP map is in use).
 
Chalnoth said:
I would think that continually increasing the degree of anisotropy in order to keep the texture from aliasing without MIP mapping would impose much more of a performance hit...
Click to expand...
I didn't say mipmaps weren't used, just that the LOD can be tweaked so that you use mipmaps less often.

Someone else noted that the aliasing had increased with the new drivers (you can tell the first miplevel is futher away from the colored mipmap shot as well).
But, it may not be a bad idea if graphics hardware is capable of not enabling trilinear filtering until that second MIP map is used (I'm not sure if current hardware optimizes for this situation...that is, if it always uses up the fillrate required for trilinear, even close to the viewer...although it does make sense that there's less of a memory bandwidth hit when only one MIP map is in use).
Click to expand...
nvidia's hardware has traditionally used two levels for trilinear. This method doesn't have to take fillrate, however, it just requires having two levels in the texture cache at once.
 
But when you're dealing with anisotropic filtering, it most likely is using up more fillrate. This would also explain the significantly increased performance hit from enabling trilinear with anisotropic also enabled.
 
OpenGL,

I don't notice any difference in LOD between the driver set 28.32 and 29.42. Is this only a GF4 thing? Plus it would be easy to test to see if lowering LOD will have a positive or negative gain on performance with anisotropic filtering. If I have time I will do some more testing.
 
noko,

When I saw Rev´s screenshots I tested a couple of other games outside of SS:SE and it´s apparent that the NV25 doesn´t use more aggressive LOD in the last drivers either.

Difference being that when you apply the extreme addon in SS:SE it applies 8x level aniso and the texture detail slider gets shifted one notch to sharper, which I don´t believe happened with 28.32. I´ve no idea though if the LOD shift is caused by the game or the drivers, since normally SS games override aniso driver settings.

So far SS:SE is the only game/application I was able to see it happening.

In any case since I don´t use anything outside of 2xRGMS/aniso in SS:SE I prefer to leave LOD at "0". I just can´t stand the added aliasing.
 
Thanks for the update Ailuros. I also use 0 for LOD for the same reason, I don't use RGMS because of performance reasons on my GF3 Ti200.
 
noko said:
I don't notice any difference in LOD between the driver set 28.32 and 29.42. Is this only a GF4 thing? Plus it would be easy to test to see if lowering LOD will have a positive or negative gain on performance with anisotropic filtering. If I have time I will do some more testing.
Click to expand...
I can't say, I was just looking at the screenshots provided.

Let us know what your testing turns up.
 
You must log in or register to reply here.
Back
Top