Trilinear vs Trylinear vs Brilinear analysis

It's a driver level feature, so it could be implemented, barring intellectual property issues 3-4 years down the road if ATI's patent gets granted and they choose to go after nV. I imagine nV has it's plate full with the compiler at this point tho.
 
demalion said:
They use this determination to propose what seems to be a skewed final evaluation of aliasing and detail (...) that is flawed not in what it is based on (the LOD examination is excellent, AFAICS), but what it ignores, i.e. that the mip map sampled from is not the only factor determining the resulting detail and aliasing.
And other are?
demalion said:
The methodology behind the samples taken and the interaction with the comprehensive LOD determination investigation seems completely unexamined, and even the "guess" conclusions offered seem to direct away from recognizing that at all by evaluating in their absence.
Also, of course, if the sampling methodologies involved do not counteract the LOD observations, ignoring them might not be a significant issue. I think IHVs should be providing explanations to address this if their methodologies take this into account.
Seems I'm loosing you here...
demalion said:
I was disappointed and puzzled, and remain so, that they went through the trouble of offering a program tool in conjunction with the article towards the goal of evaluating the filtering issues, and completely bypassed, AFAICS, any opportunity for something like "moving" across the texture so the user could actually get an in motion evaluation.
Pressing middle and right mouse buttons will allows you to move image plane. 8) You either missed it or explain your wishes, please...
 
Clootie said:
...And other are?
The next part continues what you quoted, and is where I actually tried to provide an answer...
demalion said:
The methodology behind the samples taken and the interaction with the comprehensive LOD determination investigation seems completely unexamined, and even the "guess" conclusions offered seem to direct away from recognizing that at all by evaluating in their absence.
Also, of course, if the sampling methodologies involved do not counteract the LOD observations, ignoring them might not be a significant issue. I think IHVs should be providing explanations to address this if their methodologies take this into account.
Seems I'm loosing you here...
Well, I'm not sure how I'm losing you...
You don't think the way you sample from the selected mip map, and the method used to calculate the output, are important factors in determining the output aliasing and detail level?
To me, this is akin to saying 4 sample AA = 4 sample AA, because of the sample count, and ignoring anything beyond that. The hardware implementation for AF seems to have even more opportunities for deviation in characteristics.
demalion said:
I was disappointed and puzzled, and remain so, that they went through the trouble of offering a program tool in conjunction with the article towards the goal of evaluating the filtering issues, and completely bypassed, AFAICS, any opportunity for something like "moving" across the texture so the user could actually get an in motion evaluation.
Pressing middle and right mouse buttons will allows you to move image plane. 8) You either missed it or explain your wishes, please...
:!: Both missed it and was looking for something else. :LOL:
I was looking for something that moved you along the texture at a fixed speed, like some sort of button or menu option to do the movement, so you could watch the texture for aliasing...I'd found the left mouse button option intuitively, but the only other mouse twisting option I'd considered for what I was looking for was the mouse wheel.
I even looked for a readme file, but I unfortunately didn't look in the "About" menu option, because I didn't expect instructions there (if I had, I'd have repeated my request for something like 3dmark 03's filtering quality test to be implemented).
Take a look at the 3dmark 03 filtering quality test, where you can hold down a button or a key and it will move the plane in a common case usage, and make it a significant part of your evaluation of aliasing.
Also, textures that wrap around would be helpful...there are many wallpapers that would serve better than the included textures.
 
demalion said:
You don't think the way you sample from the selected mip map, and the method used to calculate the output, are important factors in determining the output aliasing and detail level?
To me, this is akin to saying 4 sample AA = 4 sample AA, because of the sample count, and ignoring anything beyond that. The hardware implementation for AF seems to have even more opportunities for deviation in characteristics.
And if you look at the article again it examines only plain trilinear filtering (i.e. bilinear+Mip-linear, not aniso) - it's actually stated in article couple of times. And bilinear filtering in this context is of equal quality for both chips. As for difference in infuence of optimized trilinear on different chips in case of anisotriopic filtering - IMO it will be even less visible, as mip-map boundaries pushed back and taking more texels in filter kernel will lead to less differences between trilinearly interpolated samples.

demalion said:
:!: Both missed it and was looking for something else. :LOL:
I was looking for something that moved you along the texture at a fixed speed, like some sort of button or menu option to do the movement, so you could watch the texture for aliasing. ... (if I had, I'd have repeated my request for something like 3dmark 03's filtering quality test to be implemented).
:LOL: I was afraid that you'll wish something like this. No promizes throw. :)
 
http://www.ixbt.com/video2/nv40-rx800-5-p1.shtml
:)
Качество фильтраций: FarCry
Качество фильтраций: Need For Speed: Underground
Качество фильтраций: Пираты Карибского моря
Качество фильтраций: Unreal II
Качество фильтраций: Unreal Tournament 2004
Качество фильтраций: Serious Sam: The Second Encounter
Качество фильтраций: Painkiller
 
chavvdarrr said:
JFYI you may want to take a look at NEXT article http://www.ixbt.com/video2/nv40-rx800-4.shtml :)

Thank you. Finally someone posts a link with some shots relevant to the matter a hand. What we need to see are black/white mipmaps (where ATI adaptive Tri takes place) with the X800 set to application-preference-filtering (ie. Full Trilinear on all texture stages).

The shots near the bottom of the page that meet this criteria show the X800 is doing an excellent job on the mipmap transitions. (the 3 sets of shots with the X800 set to app - 4-8-16xAF.)


I actually think the mipmap transitions on the X800 are smoother than the 9800. It’s easy to see that NV’s “Brilinear” is doing a worse job of filtering here.

Conclusion : NV’s Tri-optimizations should be turned off when benching against the X800. Either that or we need to lower the settings on the X800 -- towards performance.

It would be interesting to see these black/white mipmap shots with the X800 slider on the performance settings to get an idea where we would need to set things against NV’s Brilinear.
 
Blastman said:
I actually think the mipmap transitions on the X800 are smoother than the 9800. It’s easy to see that NV’s “Brilinearâ€￾ is doing a worse job of filtering here.

I'm of the opinion that the 9800 and 6800 without optimizations are pretty similar and the same with X800 vs 6800 brilinear. (no AF screenshots).

AF4 (app): 9800, 6800 FT best, then X800 and last 6800 with opt.
AF8 (app): see above

Edit:

AF16 (app): 9800, 6800 FT best, X800 not as good but a lot less visible transitions then 6800 opt (brilinear).

Nvidias brilinear seems to be a lot better then Ati's trylinear at P and Q.
 
demalion said:
Take a look at the 3dmark 03 filtering quality test, where you can hold down a button or a key and it will move the plane in a common case usage, and make it a significant part of your evaluation of aliasing.
Ok, you can download updated version. Now, in fullscreen mode, keyboard cursor keys allows to move image plane at constant speed. But rotation analizis should be taken with the great care, as filtering applied to left/right parts of plane will be different in most cases.
 
Bjorn said:
Blastman said:
I actually think the mipmap transitions on the X800 are smoother than the 9800. It’s easy to see that NV’s “Brilinear” is doing a worse job of filtering here.

I'm of the opinion that the 9800 and 6800 without optimizations are pretty similar and the same with X800 vs 6800 brilinear. (no AF screenshots).

AF4 (app): 9800, 6800 FT best, then X800 and last 6800 with opt.
AF8 (app): see above

Edit:

AF16 (app): 9800, 6800 FT best, X800 not as good but a lot less visible transitions then 6800 opt (brilinear).

Nvidias brilinear seems to be a lot better then Ati's trylinear at P and Q.


I don't hink that's the case at all. Are you looking at the right things?

Look at the …


X800 ANIS 8 APP, 45d.
6800 ANIS 8 APP, 45d.

…shots.

It looks like the mipmaps are pushed slightly farther back on the 6800 shot but the transition is clearly smoother on the X800. In fact, I can see what looks like … “slight” …mipmap transitions on the 6800’s Full-Tri in that 8xAF shot.

Edit: The 6800 has shorter and harsher transitions between the mipmaps.
 
Blastman said:
That’s clearly not the case. Are you looking at the right things?

Look at the …


X800 ANIS 8 APP, 45d.
6800 ANIS 8 APP, 45d.

…shots.

It looks like the mipmaps are pushed slightly farther back on the 6800 shot but the transition is clearly smoother on the X800. In fact, I can see what looks like … “slightâ€￾ …mipmap transitions on the 6800’s Full-Tri in that 8xAF shot.

I agree that it seems that the mipmaps are pushed further back on the 6800 full tri at 8X AF. Something that could explain the "slight" mipmap transition. Though i still prefer if over the X800 which has more visible transitions closer to the "viewer" imo, 9800 XT has the least amount of visible transition And Trylinear looks much better then brilinear with those amounts of AF, at least in these screenshots. But i think it's very obvious that trylinear != trilinear when you look at 4X AF.
 
Clootie, I think you should try reading point 2 as a whole, and then read the summary again. I''m aware that the only thing it looked at was isotropic filtering and that it offered conclusions based on it, as I said...I think you're reading as if my conclusion was that the article was bad and dismissed that completely, instead of saying why I found fault with that.

BTW, a subtle point about your "gpgpu" mention (perhaps some subtletly was lost in translation?): shouldn't ATI's behavior of using trilinear when provided mip maps for other than "image quality", and nVidia's user controlled trilinear optimization switch, fulfill your concern? Your discussion seems to fail to put the gpgpu concern for doing trilinear when "ordered" into the context of what you are examining. It looks odd, because both mentioned cards seem to take steps to avoid the concern you seem to raise being a factor, yet your mention seems to introduce the concern without recognizing that.

...

And thanks for looking at the functionality I'm suggesting. :)

EDIT: Thanks for IMPLEMENTING the functionality I suggested, downloading now. The key implementation makes, with some set up, exactly the examination I wanted possible. The only thing lacking :)P) is an option to set a (or maybe several) fixed orientation of the texture, and orientation and altitude of the cameara, to facilitate consistent reproduction between users (and maybe controllable speed, though the selected speed looks well suited to exposing aliasing, AFAICS), and an article and a new set of textures examining it.
This would, IMO, be a big deal, as I think it would be exactly the tool needed to center a comprehensive article around, and solve the "how to show the user the in motion aliasing or lack of it" problem.
 
demalion, seems you would liked "global" conclusion, but I wanted to stay as close as possible to trilinear/brilinear issue. If you want to know mine personal opinion - NV standart trilinear is the best current implementation. And if you think about gpgpu usage of it - it's most usefull as LOD value is linearly interpolated (this is not true in ATI/refrast case).
It's not possible to force ATI to implement linear LOD calculation as their current hardware just can't physically do it. And it's actually saves them some transistors. I can't really blame them as DX RefRast does the same thing. But it's possible to force them to include standart trilinear checkbox in ControlPanel.
What I've tried to show in this article is what kind of quality/speed trade off you can expect both in NV and ATI cases. By the way showing that it's incorrect to compare even standart trilinear on both of the chips.
...
As for utility - it's already possible to store/restore current state (it's written to INI file). But futher analizis is going to be more subjective and less objective as different people have different opinions; different gamma settings on their monitors; blurry CRT's and sharp LCD screens. And many more factors that's not possible to predict :oops:
 
i doubt you can force ATI to put trilinear in the control panel if you remember the reviews with 61.11 drivers. A lot of reviewers were unable to see that the remove tri optimizations was broken and benched as if it was working. Why do you expect a company to start a competition with a penality ?
 
So you understand why Nv remove the full tri from 61.11? :oops:

I want both to deliver full tri, or both will go for brilinear...
 
Bjorn said:
Blastman said:
I actually think the mipmap transitions on the X800 are smoother than the 9800. It’s easy to see that NV’s “Brilinearâ€￾ is doing a worse job of filtering here.

I'm of the opinion that the 9800 and 6800 without optimizations are pretty similar and the same with X800 vs 6800 brilinear. (no AF screenshots).

AF4 (app): 9800, 6800 FT best, then X800 and last 6800 with opt.
AF8 (app): see above

Edit:

AF16 (app): 9800, 6800 FT best, X800 not as good but a lot less visible transitions then 6800 opt (brilinear).

Nvidias brilinear seems to be a lot better then Ati's trylinear at P and Q.
It seems obvious that the P and Q settings were the same. We know that ATI does trilinear on stage 0 only in Q mode, so either the tester did something wrong, or they used a stage other than 0 for their analysis.

-FUDie
 
PatrickL said:
Stop trying to twist my words. Thanks :rolleyes:
Well, that's what you said. "Why do you expect a company to start a competition with a penality ?"
So either it's true for Nv and Ati or false for both on tri/brilinear. So either both do full either both do bri ;)
 
Ati should let us turn it off . But put out buggy drivers that wont let us really tun it off and there ya go :p
 
Back
Top