Interesting new AA modes for GF3s and GF4s

[EvilEngine]

Using the D3D tubes AA demo, I tested 4XS AA to see if there was an improvement in current 40.72 WHQL drivers.
No change really at all in that mode, but being the daring guy I am, I used Rivatuner to set the AA mode to 7, (5 = 4xs) and look what happened:
http://evilengine.apokalyptik.com/special/screenshots/AAmode.jpg

4XS:
http://evilengine.apokalyptik.com/special/screenshots/4xs.jpg

Level 7:
http://evilengine.apokalyptik.com/special/screenshots/7x.jpg

*edited for Ichneumon*

[siconik]

Perhaps it the 6SX that we have been hearing about?

[siconik]

Wasn't there a program that showed exactly what samples are taken? I can't seem to find it.

[Althornin]

or what is the 4x+9tap gaussian blur?

[EvilEngine]

Heres a 4x9tap shot, noway compares to the mysterious level 7:
http://evilengine.apokalyptik.com/special/screenshots/9tap.jpg

[Joe DeFuria]

That level 7 looks really nice...

Can you force that in all DX apps? Can you run some benchmarks against the other modes?

[Nappe1]

btw, anyone still has Classic Radeon??

I find out some time ago that Radeon (R100) core supports Smoothvision 5x and 6x modes. Or at least using registry AAmode values from Radeon 8500 drivers, I got those modes working on my AIW.

I never found out time to test some more and my 32 Megger AIW was limited on 5x mode to 640x480 and in 6x mode to whopping 512x384, so it was rather not so good idea. :) on older games with 64MB vivo / Radeon 7500, those modes could have been handy...

umm... has anyone yet checked, if this same works between Radeon 8500 and Radeon 9500 cards too?

[Basic]

http://hem.passagen.se/basic3/fora/beyond3d/fsaaview.htm

Just Do It. :D

[siconik]

Yeah, that's the one I've been looking for. Can anyone try it with this new FSAA mode and tell us what you see?

[EvilEngine]

Level 7
http://nemesis.segfault.ws:81/~evilengine/Screenshots/FSAAVIEW.png

*edited for Ichneumon*

[EvilEngine]

Heres Level 6 if anyone's interested. I don't know if this is good or bad.
http://nemesis.segfault.ws:81/~evilengine/Screenshots/level6.png

*edited for Ichneumon*

[Basic]

About level 7:

Hmm, I've seen those blurry dots before.
I explicitly set the filtering to point sampling, but that's the look you get with bilinear. Have anybody heard of any forced filtering?

Either way.
Textures are sampled in a 2x2 grid.
Both edges has 0, 25%, 50%, 75%, 100% sequences, with occational single pixels at 12.5%, 37.5%, 67.5% , and 87.5%.
This implys two pixels in each of four rows, and two pixels in each of four columns.

This could be explained with a 2xRGMS that then is downsampled by 2x2.

[edit]
Btw, good move to use png.

[Basic]

About level 6:

It could be 2xRGMS * 1.5x2xOGSS

Or in other words this pattern distributed over two pixels horizontaly:

A--B--C--
--a--b--c
D--E--F--
--d--e--f

'A' and 'a' come from the same multisample, 'B' and 'b' from another, and so on...

AaBDdE are mixed to the left pixel (double weight to 'B' and 'E' for some reason).
bCceFf are mixed to the right pixel (double weight on 'c' and 'f').

But there were room for misinterpretations, so I could be wrong.

[Cless]

About level 7:

Hmm, I've seen those blurry dots before.
I explicitly set the filtering to point sampling, but that's the look you get with bilinear. Have anybody heard of any forced filtering?


Could the Forced Anisotropic filtering specified in RivaTuner shot be the problem?

-Colourless (with an inactive account that he can't reactivate after he attempted to change his email address)

[EvilEngine]

Oh, I didnt know you were suppossed to turn off aniso for this test, sorry.
Anyways here is level 7 again with aniso set to "application controlled".
http://evilengine.apokalyptik.com/special/screenshots/level72.png
LOL oops, *edit*, posted the image with no AA.

*edited for Ichneumon*

[Bambers]

Can you could get this in opengl as the 0.3 ogl only version has a useful geometry mode.

Both of basics suggestions for level 6 and level 7 have been suggestions in the other threads for the nv30s 6xS and 8x.

8x

_S_____S____
____________
____S_____S_
____________
_S_____S____
____________
____S_____S_

Which could be quite expensive considering it has 4x supersampling. Maybe nvidia feels the nv30 has the fillrate to handle this.

So 6xS would be

___S___
_S___S_
___S___
_S___S_

and

_S___S_
___S___
_S___S_
___S___

for alternating pixels which does a very nice job of explaining the 'dithering' of the AA on the bottom edge. The colour graduations on the edges are a bit odd though. The vetical edge has shades at 0%, 25%, 50% and 100% with single pixels at 12.5%, 37.5%, and 75%. Shouldn't these be even and at pecentages like 16.6 and 33 etc?

For the horizontal edge there is alternating between 12.5% and 37.1% then a solid gradient at 50% followed by alternating between 62.1% and 87.1%. Not the kind of values of sixths that the pattern would imply. :-?

EDIT: Didn't notice Basic said the right 2 samples have double weight. Possibly a driver bug?

Also why would 8x be able to be used in opengl too despite being a combination of supersampling and multisampling?

EDIT: fixed 8x pattern

[aths]

The new Modes are 6xS and 8xS, well desribed in our 3DCenter-Article. (http://www.3dcenter.org/artikel/2002/11-20_b.php)


There is even already a new aTuner-Version (http://www.3dcenter.org/atuner/) with Support of 6x and 8x.

[Dave Baumann]

Also why would 8x be able to be used in opengl too despite being a combination of supersampling and multisampling?

I can't quite tell from the outputs from Basics program, but could it be because the texture sample from the two pipelines producing each of the 4 samples is sampled from the same position?

However, to me its a little disappointing that the AA sampling pipeline for NV30 bascially hasn't changed since NV20.

[Xmas]

for alternating pixels which does a very nice job of explaining the 'dithering' of the AA on the bottom edge. The colour graduations on the edges are a bit odd though. The vetical edge has shades at 0%, 25%, 50% and 100% with single pixels at 12.5%, 37.5%, and 75%. Shouldn't these be even and at pecentages like 16.6 and 33 etc?

For the horizontal edge there is alternating between 12.5% and 37.1% then a solid gradient at 50% followed by alternating between 62.1% and 87.1%. Not the kind of values of sixths that the pattern would imply. :-?

You're not considering taking a weighted average of the samples. After some thinking I come to the following conclusion:

___1___
_1___2_
___1___
_1___2_

and

_1___2_
___1___
_1___2_
___1___

I don't know exactly what's the sense behind this (besides the obvious divide by 8 simplicity), but it fits the result.

Maybe Basic could write a "mirrored" version of his program, this would show if my assumptions are correct.

[Ichneumon]

Off topic:

Would people posting large images on the forum be kind enough to simply link to them? The phpBB forum is really annoying that it doesn't resize the non-image posts in the thread to be normal sized, and it makes reading threads really hard to scroll left and right all the time to read each message.

[Bambers]

lol xmas posted while I was editing my post :-?

I was thinking that the odd weighting could just be down to the fact that this part of the drivers isn't properly finished yet.

[Xmas]

I don't think it's a bug. It's much easier to divide by eight than by six. And Quincunx already divides by 8.

[Crusher]

Then wouldn't this make more sense?

___1___
_2___1_
___1___
_1___2_

[TempestX]

The latest version of aTuner has the modes already built in for DirectX for the secret modes. Madden 2003 on my machine at 1024x768 32 bit runs pretty smooth on my Ti 4400 in 6x mode. 8x mode runs but I drop down to 800x600 32 bit.

My CPU is Athlon XP 1900+ at 1680 140 MHz bus.

[Edited] Screenshots of aTuner with the additional FSAA modes in this MadOnion link:

http://discuss.madonion.com/forum/showflat.pl?Cat=&Board=techdisplayadapters&Number=1475768&page=0&view=collapsed&sb=5&o=0&fpart=1

[Snyder]

Screenshots in this MadOnion link:

http://discuss.madonion.com/forum/showflat.pl?Cat=&Board=techdisplayadapters&Number= 1475768&page=0&view=collapsed&sb=5&o=0&fpart=1

Are you sure you didn't mistake them guys signatures for screenshots? :wink:

Seriously, I didn't find them...

[TempestX]

Snyder...Sorry about that....Screenshots of aTuner with the added FSAA modes....

[AdamK47]

OpenGL anyone?

[zymbalist]

Will this tweaking enable Ti 4400 or 4600 to catch up the lost ground with Radeon 9500 Pro or 9700 in FSAA and Anisotropic Filtering??

[BRiT]

zym, no, those cards will always be slower with lower quality AA (ie: non-gamma corrected). These new modes are even more taxing (higher performance hit) than the officially supported modes.

[Geeforcer]

Will this tweaking enable Ti 4400 or 4600 to catch up the lost ground with Radeon 9500 Pro or 9700 in FSAA and Anisotropic Filtering??

No. If you use one of the new AA modes and max aniso, you will get excellent image quality - but it will be an order of magnitude slower then Radeon 9700. You will get very good performance if you use low aniso and AA setting, but then it will look much worse then Radeon 9700. Considering that you get either great IQ or great performance but not both, I would hardly call that catching up.