When enough is enough (AF quality on g70)

Why do software developers request LOD bias anyway?

 

I'd like to know this too.

Ailurous, very nice writeup - why'd you go with PDF?

 

I enjoyed the writeup as well. Thanks!

Nite_Hawk

 

Excellent work Smithers.

 

Check out this article: http://www.bit-tech.net/news/2005/09/08/cat58_shimmering/

Did ATI decide that the over optimizations are necessary as well?

 

I am not aware of any differences in image quality between Catalyst 5.7 and 5.8 in any of the CatalystAI modes. As far as I am aware our internal IQ testing did not show any changes, and I am not aware of any changes in the driver that would have caused any difference.

There are no filtering optimisations applied in ATI drivers when CatalystAI is disabled. The filtering optimisations applied in other Catalyst AI modes did not change at all between 5.7 and 5.8 as far as I am aware, although I will double check.

 

So I'll see the same 'swimming' textures in Catalyst 5.7 too? Even though I've never noticed them before. I sit down and play games rather than watch timedemos go past on the screen. There was definitely something different going on as I stated in my article.

I think the reason why I couldn't see the shimmering in the videos that I'd created was because I'd taken them while I was riding on the jetski/hovercraft/boat/whatever you want to call it at full speed.

 

Because we still need some time until the new site is up and running. I could wait until then, but by then the write-up would had been old and redundant.

I figure it was faster on your HDD than a 10 page html or not?

 

Is that a trick question of some sort? The layman here is confused

***edit: oh and thanks folks for all the kind words

 

Have you tried taking an image screenshot comparison between each set of drivers? If there is "more shimmer" between one and the other then the images should be doing things differently.

 

It used to be usefull as a hack before aniso was widely available. These days aniso actually solves the problem better and is about the same cost and changing the MIP bias.

I have no Idea why devs change it now.

 

Hyp-X: thanks

http://www.3dconcept.ch/chips/chiptable2nde.htm
http://www.hwupgrade.com/skvideo/video_cards_comparison/index2.html
http://www.extra.hu/pcpage/videokartyak/savage/menu.htm.frame

so why is it mentioned anywhere?

 

Because people didn't analyse things to the same degree back then? Fast Trilinear only really became apparent after Kyro was released - at this point people were looking a little deeper into IQ and other tools, but not necessarily getting all the facts (in the case of Kyro's fast Trilinear Q3 mip-maps would appear to be only Bilinear when coloured because Fast Trilinear only uses one mipmap laevel with multiple samples - Kyro only had this enabled when compression was used though, which is why it showed up in Q3); people later hooked on to S3 having "Fast Trilinear" / "Box Filtering" after Kyro was released. Ask anyone back then what Anisotropic filtering was, as well, and they would probably have blinked twice and stared blankly at you.

 

Like Riva TNT had single cycle anisotropic filtering, according to the 3DConcept table

 

Precisely.. hence my stipulation on the last page (and lots of interesting replies, but fully supporting my assertion since then).

How many GF3 reviews/website articles fired up RegEdit and filled out the DefaultLogAniso value and listed "4-tap, 8-tap, 16-tap" numbers and screenshots? A whole lot of them!

We've seen a pretty drastic change in ethic- from the desire to do whatever was necessary to create an IQ-crown, even if it meant hunting and poking around in the registry to get it... but suddenly now, the registry is taboo.. and even specific "Quality" driver options are to ONLY be used/explored while being totally blind to what's on the screen.

It seems a little odd how things have shifted to perversely the opposite way. Which way do you think offerred the most value to the "average" gamer? How about the "average" Beyond3D nerd?

 

Just another small historical flip-back: back then it made more sense to enable 2x SSAA (even with just bilinear) over non-adaptive 2xAF available on some accelerators back then. The overall gain was bigger since you'd not only get an effect quite similar to 2xAF on textures (well at least on one axis with 2xSSAA) but you'd also get some polygon edge AA at the same time. I don't recall what the performance drop for 2xAF for the GF2-line was, but fill-rate back then was definitely better invested in 2x Supersampling compared to 2xAF. At least IMO.

***edit: that said, if the future (when hardware gets more powerful) holds equivalent texture quality for us as 4*4 OGSS on GF's, without of course counting seconds per frame as with high sample densities of strict Supersampling, then I'll be more than just happy.

 

I disagree. The performance drop with 2xAF on GF2 was quite low, because it was bandwidth-starved anyway and its texture filtering rate was way higher than its pixel fillrate.

I know of no chip with non-adaptive 2xAF, btw. Which ones do you mean?

 

Damn my weak memory; I just checked B3D's GF2 review and the drop for 2xAF was at least twice as low as with 2xSSAA.

How about if I say, that AF was less adaptive back then? Does that sound more accurate?

 

Necessary? Hell no. It's a bug.

 

This can't be any "speed optimization", because performance (test1, test2) is slightly worse, too...