Trilinear Filtering Comparison(R420 vs R360 vs NV40 vs NV38)
- Thread starter mikechai
- Start date
Mintmaster
Veteran
Stryyder said:
Auto: What they're doing now, using brilinear when they see fit. Probably the best option for the average consumer.
Always on: Will be similar to NVidia's optimizations. Note that ATI stated they're not doing this all the time, only when they're analysis deems it fit. One of Lars' benchmarks shows that NV40 gains more from brilinear than ATI does. This option will let them use the optimization for all textures, and will help them in benchmarks.
Always off: Traditional trilinear. Some people notice the mip-map transitions of brilinear, so this option lets them keep the full trilinear on if they want.
Mintmaster
Veteran
Stryyder, may I ask what you're sighing at? I think I made perfect sense.
If ATI is getting so much flak right now, why not have the capability to enable the optimisation on all textures? They should get a notable performance boost.
In some games you may notice the difference, and many people are complaining about these high end cards not doing real trilinear, so why not have to ability to turn it off?
If ATI is getting so much flak right now, why not have the capability to enable the optimisation on all textures? They should get a notable performance boost.
In some games you may notice the difference, and many people are complaining about these high end cards not doing real trilinear, so why not have to ability to turn it off?
MrGaribaldi
Regular
Re: I'm confused
While I share you concern wrt "apple-to-apples" comparisons, I'm not sure I agree with your conclusion, as this will only gain the IHV's and not the customers.
I believe that the customers should have the most options to choose from, and thus having only one slider to control everything would severly limit those choices. What if someone has no problem with bilinear filtering, but can't stand jaggies? Or what if the person don't want to use FSAA at all? Or want to use TruForm and turn down everything else so the hit won't be too high? How will that person be able to change the settings to cater to his/her need?
Just because we're seeing more and more reviewers who are holding shoot-outs, and believe that as long as 4X AA is used on both cards the result will be identical, doesn't mean we should take away options from the customers.
<rant>
I think it'd be better to get the reviewers to do either what B3D does (no shoot-out), something along the lines of what Brent does (set a limit for the FPS and increase IQ until you hit that limit), or go back to what some sites used to do and ignore what the setting is called and check which settings are the closest.
With the exception of the first option, these options require more work from the person doing the testing, but if the tester does it properly should give his/her readers a much better review than what is delivered today.
</rant>
No argument here
Seiko said:
While I share you concern wrt "apple-to-apples" comparisons, I'm not sure I agree with your conclusion, as this will only gain the IHV's and not the customers.
I believe that the customers should have the most options to choose from, and thus having only one slider to control everything would severly limit those choices. What if someone has no problem with bilinear filtering, but can't stand jaggies? Or what if the person don't want to use FSAA at all? Or want to use TruForm and turn down everything else so the hit won't be too high? How will that person be able to change the settings to cater to his/her need?
Just because we're seeing more and more reviewers who are holding shoot-outs, and believe that as long as 4X AA is used on both cards the result will be identical, doesn't mean we should take away options from the customers.
<rant>
I think it'd be better to get the reviewers to do either what B3D does (no shoot-out), something along the lines of what Brent does (set a limit for the FPS and increase IQ until you hit that limit), or go back to what some sites used to do and ignore what the setting is called and check which settings are the closest.
With the exception of the first option, these options require more work from the person doing the testing, but if the tester does it properly should give his/her readers a much better review than what is delivered today.
</rant>
Seiko said:
No argument here
Quitch said:
Mipmaps are not a fudge, not if you've built your mimaps with correctly designed filters so that each mipmap contains information from different spatial frequency bands in what's known as a tight frame (no information loss). Trilinear follows the right philosophy but I now understand that the current "double bilinear" in two successive mipmaps is not good. Texture crawl comes from too high spatial frequency information given the resolution of the pixels we can draw on the screen. According to scale space theory, you can interpolate two texels, one from any two successive scales (levels in a mipmap) to actually give you the ideal spatial frequency for displaying on the screen: not too high so you don't get texture aliasing and not too low so that you lose info and the drawn pixels become blurred. The texel in the higher level mipmap has to be obtained by bilinear interpolation in some circumstances (for texel correspondence between mipmap levels). We actually need less bilinear in the lower (finer) mipmap and more bilinear (to obtain texel correspondence) in the higher (coarser) mipmap. Does that make sense?
Chalnoth said:
Ooops, sorry for using that terminology. The way I understand ATI's new trylinear is that they do bilinear on a mipmap and take less samples from the higher mipmap to save on texture lookups when they do not need to. One way to actually do adaptive scale filtering is not to do any bilinear on the current mipmap (since that info is already available at the higher mipmap) and instead use a bilinear interp to obtain the corresponding texel at the higher mipmap (3/4 of texels are missing if the higher mipmap has been subsampled by a factor of two in u,v). I'm suggesting only one tex lookup for current mipmap, more tex lookup for the interp in the higher mipmap. Then interpolate between the two texel values at the different mipmap levels depending on which intermediate scale (continuous mip levels instead of discrete ones) you want.
I'm still not exactly sure what you're saying. But I don't think that it would be benficial to drop to anything below basic bilinear filtering (4 texels linearly-interpolated).
Are you saying that with anisotropic, it would be beneficial to take fewer bilinear samples from the lower-detail MIP map, and more samples from the higher-detail MIP map?
Are you saying that with anisotropic, it would be beneficial to take fewer bilinear samples from the lower-detail MIP map, and more samples from the higher-detail MIP map?
MrGaribaldi
Regular
Re: I'm confused
Which is why I outlined three different approaches, which should most needs.
1. No shoot-out, just review the card
2. A Brent-esque approach, set fps limit, change settings to achieve
3. "Old-school" (at least to me), ignore setting name, compare IQ to find best settings. (Give examples and bench all/most settings)
I feel that if hardware sites used one of these methods, instead of the "regular" method (2x = 2x no matter what), we'd see an increase in review quality.
Note: These are not meant to be a be-all-end-all list of "acceptable" review methods, as I'm sure there are other ways which are just as good/better, but rather as a pointer in which way I'd prefer reviews to evolve....
Bjorn said:
Which is why I outlined three different approaches, which should most needs.
1. No shoot-out, just review the card
2. A Brent-esque approach, set fps limit, change settings to achieve
3. "Old-school" (at least to me), ignore setting name, compare IQ to find best settings. (Give examples and bench all/most settings)
I feel that if hardware sites used one of these methods, instead of the "regular" method (2x = 2x no matter what), we'd see an increase in review quality.
Note: These are not meant to be a be-all-end-all list of "acceptable" review methods, as I'm sure there are other ways which are just as good/better, but rather as a pointer in which way I'd prefer reviews to evolve....
Chalnoth said:
Aliasing occurs when you subsample an image which has got too high spatial frequencies. A bilinear filter can be used as a cheap means of removing these high frequencies (but it's not that great at it).
Say you have got a high detail texture that can only be sampled at 1:1 and we call it level 0. If you subsample, aliasing is going to occur. You can use a properly designed low-pass filter, or a gaussian, to remove enough high frequency so that the texture can be safely subsampled at 1:4 without any aliasing. We do not perform any subsampling yet and call this version level 1.
What if you would like to subsample the texture at 1:2 (for filling a polygon)? You can interpolate between a texel at level 0 and a texel at 1 and if you get the interpolation weights right, you can get the same result as a low-pass filtering for safely subsampling at 1:2 (without having to do the low-pass filtering).
Usually in mipmap levels, the higher (my convention is coarser but substitute as necessary) level would have been subsampled. So if the sample that you require at level 1 is not available, you can use bilinear interpolation to obtain its value from the neighbouring samples that are available.
In the end, you only need two samples, one from the current mipmap level and one from the level higher up. If that last texel sample is not available, you use bilinear to obtain it. The trick is the interpolation weights between the two levels.
Does that make better sense?
Mintmaster said:
Why create what I am sure ATI believes to be a superior filtering method that actually improves performance just to turn it off? Whats the business reason? ATI is telling us this is there new method, we think its better if anyone shows us proof to the contrary we will fix the issue.
You people need to stop hoping for an option to shut it off, the most ATI will do is to give reviewers the tool that DB has for IQ comparisons. Unless someone can prove in multiple cases that the algorithm produces worse results you are not going to see an option to shut it off period.
