Quincunx & FSAA: how is it changed between Geforces

[Above]

Hello to the intellectuals at beyond3d.

I am very impressed with the Geforce4 based on beyond3d and nvnews reviews with regard to antialiasing capabilities. Both in cost and quality they surpass preceeding generations.

Quincunx on the Geforce4 seems to have less performance penalty than it did on the Geforce3. I have seen users say that they are unable to take screenshots of it in the usual manner, i.e. the image does not therefore lie in the framebuffer but is hardware implemented at the output stage or "postfiltered". Is Quincunx changed by software drivers, or is the function embedded in hardware in the Geforce 4? 4xs may be hacked for on the Geforce3 by driver settings.ad that it is sharper. I have seen this in screenshots in an nvnews review (how do they capture a screen that does not display in the framebuffer?)

I want to know if Quincunx on the GF4 is good in motion. How does it handle texture shimmers. Is it a good alternative to the 2x, 4x and 4xs modes, or will its sampling algorithm always fall short because it uses no enhanced resolution. Quincunx may not stand up to 4xs, but on the other hand it comes very cheap.

 

[Dave Baumann]

FSAA on GeForce 4 took a leaf out of the 3dfx technology books.

Normally with FSAA you have your FSAA buffer or buffers which contains information for all the pixel samples; so, for instance, if you are running 4XFSAA there will be 4 times the information in the frame buffer than will actually be displayed. When the frame is fully rendered its normal for each of the subsample to be read into the chip, averaged (to produce the final pixel colour from all the samples data), and then pushed out to the frame buffer again before being sent to the RAMDAC for display – this means that the FSAA sample buffer(s) is read into the chip then passed back out to another buffer that the size of the display resolution which is then sent back to the chip for the RAMDAC to display it.

What Voodoo 5 did, and now GeForce 4 is miss out the last step – it actually combines (averages) the FSAA sub sampled in the RAMDAC when the frame is rendered which means it misses on passing a buffer into and out of frame buffer RAM meaning less time is lost, but more importantly (in most cases) less bandwidth. It follows that Quincunx is also able to do the colour combination (averaging) in the RAMDAC as well which is why it produces results faster then the previous generation of GeForce.

 

[Mintmaster]

Hello to the intellectuals at beyond3d.

I am very impressed with the Geforce4 based on beyond3d and nvnews reviews with regard to antialiasing capabilities. Both in cost and quality they surpass preceeding generations.

Quincunx on the Geforce4 seems to have less performance penalty than it did on the Geforce3. I have seen users say that they are unable to take screenshots of it in the usual manner, i.e. the image does not therefore lie in the framebuffer but is hardware implemented at the output stage or "postfiltered". Is Quincunx changed by software drivers, or is the function embedded in hardware in the Geforce 4? 4xs may be hacked for on the Geforce3 by driver settings.ad that it is sharper. I have seen this in screenshots in an nvnews review (how do they capture a screen that does not display in the framebuffer?)

I want to know if Quincunx on the GF4 is good in motion. How does it handle texture shimmers. Is it a good alternative to the 2x, 4x and 4xs modes, or will its sampling algorithm always fall short because it uses no enhanced resolution. Quincunx may not stand up to 4xs, but on the other hand it comes very cheap.

The Geforce3/4's FSAA is multisampling, and so it only samples each texture once per screen pixel. The result is that multisampling does one thing and one thing only - smooth jaggies, and even that it doesn't get right 100% of the time. Some jaggies are from alpha compare functions, as is used in MANY games for leaves on a tree, but multisampling won't do jack for these. It smooths the jaggies on the edges and intersection of polygons only.

Texture aliasing, therefore, is not addressed by multisampling. The idea is that with proper filtering there should be no texture aliasing, at least for traditional textures without bump-mapping or any dependant texture reads. While this may be true, you will still get a worse image than with proper super-sampling, as that reduces texture shimmer and sharpens textures at the same time due to a higher LOD.

As for Quincunx, this is a blur filter applied to pixels after multisampling. This is a real cheap-ass way of removing jaggies, as you could essentially do the exact same thing with a bad, blurry monitor. However, because the blur spreads half of the samples to 4 adjacent pixels, texture aliasing will be reduced a bit.

Personally, I think multisampling is only a temporary solution to removing jaggies. As more complex pixel shaders start popping up, supersampling will show why its greater performance loss is not for nothing.

 

[Ailuros]

Mintmaster,

What happens when you force the applications to skip alpha tests and modify anisotropic algorithms to not pose any unwanted side-effects with pixel shaders?

Of course is Supersampling superior to Multisampling but everything comes at a price. Besides MSAA combined with advanced texture filtering is no way near as bad as many try to picture it.

Ideally I´d like T-buffer quality in today´s hardware too, but with the performance penalty Multisampling has.

 

[Above]

Interesting information about the Geforce4, thank you people. I am still curious about the newer Quincunx sampling and why it looks sharper. Is this change implemented at hardware or software level? It is irritating when "new" features are ones merely obstructed on older hardware through drivers. The idea is to run the older cheaper Geforce3 as much in the same way as the Geforce4 as possible, merely accepting lower framerates.

 

[Randell]

I always understood it to be a hardware change to the pixels sampled.

 

[Sharkfood]

One of the biggest changes I've seen in the Quincunx method is one that appears to be related to using a post-filter.

What this does is makes it impossible to capture Quincunx in screenshots. What you get is not what the final result looks like. What this has done is lead to a ration of websites publishing Quincunx screenshots with "See? It's not blurry!" when the screenshots only have the multisampling effect applied and not the post-filter blur.

You can proof/test this by using a game such as Tribes or similar that has a small font/transparent overlayed text window in-game. Set Quincunx and snapshot with and without Quincunx. The text/chat window is a perfect example as the font is heavily blurred in-game, but using HypersnapDX, CTRL-PRNTSCN, FRAPS screenshot utility and even the games own internal screenshot function- the text in the Quincunx screenshots is razor sharp with none of the blurring that occurs in-game. In fact, it's nearly identical to 2xRGMS screenshots.

So what it comes down to is this- either lots of websites are publishing screenshots of 2xRGMS labeled as Quincunx as the blend is post-filter, or they have discovered some new/undocumented way to get the Quincunx output from the framebuffer. I'd testify it is the former as many screenshots I've seen as Quincunx simply do not represent what my Visionek GF4 Ti4600 displays on-screen/in-game as being represented at many websites.

So it looks like you may need to simply try one of these cards to formulate your own opinion on it. There is no denying that it is improved from the GF3's version of Quincunx, but the improvement is no way proportional to the improvement shown in screenshots.

Cheers,
-Shark

 

[MikeC]

Hello Above,

My observations with Quincunx antialiasing on the GeForce4 are that it has improved over the GeForce3 in both performance and image quality. Note that when I mention image quality in this message, I'm only referring to the "blur" effect.

When I initially previewed the GeForce4 Ti 4600 back in February, Quincunx performance was much faster on the GeForce4, but the image quality was similar to that of the GeForce3. I even commented in the preview by stating that "the new sampling pattern doesn't have much of an affect on Quincunx antialiasing in this scene as the quality is comparable on both cards."

I created the following java applet to use for the comparison in that preview:

http://www.nvnews.net/previews/geforce4/images/page_4/scene1_1024x768_comparison.shtml

Upon reviewing the GeForce4 Ti 4200 some months later using an updated Detonator driver, I performed a similar comparison and found that image quality improved under OpenGL:

http://www.nvnews.net/previews/geforce4_ti_4200/images/q3_aa_comparison.shtml

but not under Direct3D:

http://www.nvnews.net/previews/geforce4_ti_4200/images/il2_aa_comparison.shtml

Texture shimmering continues to occur when using Quincunx, but 4XS antialiasing, which is only available under Direct3D, is able to partially eliminate it. While Quincunx is not the perfect antialiasing solution, it does allow one to play many games with it enabled at resolutions as high as 1280x1024 using the GeForce4 Ti. There are a number of discussions on game play with Quincunx in the nV News graphics card forum and many GeForce4 Ti owners agree that Quincunx is "playable" at high resolutions.

As graphics cards continue to become more powerful and offer new features, image quality has become an important topic. Along with a few other reviewers, I devoted more time to image quality comparisons when the GeForce3 made its debut. I'm continuing the trend in order to give our visitors pertinent information that can be useful in purchasing decisions. Which leads me to my next point.

By now you may have noticed that I intentionally left out discussing the technology behind Quincunx. While I strive to enhance my knowledge of 3D graphics, I'm still not qualified to discuss much of the underlying technology. My policy has always been to refrain from writing about a subject unless it's accurate (to the best of my knowledge). While a forum can be a valuable communications tool, it can also be harmful since there's a great deal of unsubstantiated material posted (not necessarily in this thread).

 

[Dave Baumann]

While a forum can be a valuable communications tool, it can also be harmful since there's a great deal of unsubstantiated material being posted.

FYI - what I've posted above is just a description of what has been confirmed by nVIDIA in a recent interview / quote.

 

[darkblu]

While a forum can be a valuable communications tool, it can also be harmful since there's a great deal of unsubstantiated material being posted.

FYI - what I've posted above is just a description of what has been confirmed by nVIDIA in a recent interview / quote.

interesting info, Dave. btw, that brings forth the following question - how is rendering to texture handled in this case - can the geforce perform any form of ms-antialiased render-to-tex? which, btw, reminds me of the solution 3dfx had for their t-buffer aa: despite the fact that vsa100 didn't have a 'render-to-tex' capability (IIRC), it still handled reading-back from a render-target surfaces correctly, the gpu blending the pixels being read on-board (i.e. not by the DAC) the same way as they were blended by the DAC, and that was transparent for the application (except for the fact that reading back from render-targets w/ aa enabled was really slow)

 

[Sharkfood]

MikeC-

I created the following java applet to use for the comparison in that preview:

Did you spend some time ensuring what your comparison is based upon is not missing the apparently post-filtered blend of Quincunx on the GF4? What capture/screenshot process did you use?

I only say this from my knowledge of my previous post. As I continue to pour over screenshots, they have little to no resemblence to what is shown on-screen. My speculation is, again, a post-filter process for the neighbor blend portion, but the end result is screenshots from all known methods I've tried result in overly sharp, unrealistic views of the Quincunx output of the GF4.

I used Tribes as an example, as the text window is testament for the issue here. With Quincunx enabled, the text is adversely effected with a small amount of blur that naturally occurs in the process... yet taking screenshots, I have been unable to get the *real* image quality output.

Here's the text window in question with NO AA:

And here is a screenshot with Quincunx enabled:

Just firing up Tribes with Quincunx enabled shows the text to not be nearly as clear as the screenshot illustrates, hence my query. I've been trying to figure out what kind of screengrab to use, from Hypersnap, to FRAPS, to CTRL-PRNTSCN, to games internal snapshot- all get a nice, sharp, non-neighbor blend image.

Any insights or reproduction information would be appreciated. For now, I can't honestly say I've seen a realistic approximation of Quincunx to accurately depict it's image quality at this stage. If it is indeed using a post-filter process (like the old Voodoo cards did for 16-bit blend), then we may need a new screenshot tool with the algorithm included for post-processing screenshots like Hypersnap did for years for 3dfx.

Cheers,
-Shark

 

[John Reynolds]

The eVGA GF4 Ti4400 I just bought a few weeks ago exhibits noticeable blurring with Quincunx enabled. So much so that it's unusable for me.

 

[Sharkfood]

I'm of the same opinion John.. but screenshots don't seem to capture it!

I see all these nice, sharp screenshots- then fire up the actual games and go "hmmm. this isn't anything like how the screenshots depict.."

I would have sensed some 'foul play' if I didnt take a bunch myself and discover my own shots are nice and razor sharp as well!

I'm still speculating it's using post-filter for the blend. This would be a logical assumption since it's not in the framebuffer, or at least in no way that I've figured out how to capture. My text shots above illustrate this quite well.

 

[Randell]

The eVGA GF4 Ti4400 I just bought a few weeks ago exhibits noticeable blurring with Quincunx enabled. So much so that it's unusable for me.

but John you are a dyed in the wall nVidiot now, how can you say such things

How is 4xS?

How is Morrowind? Can you use 2xRGMS at all on your new rig with it?

 

[Ailuros]

Shark,

Interesting. Do you have post-processing enabled in Hypersnap? I haven't tried all the options myself, but I'll try to see if using 1x2 or 4x4 options make any difference.

 

[John Reynolds]

The eVGA GF4 Ti4400 I just bought a few weeks ago exhibits noticeable blurring with Quincunx enabled. So much so that it's unusable for me.

but John you are a dyed in the wall nVidiot now, how can you say such things

Heh, so I've heard.

How is 4xS?

How is Morrowind? Can you use 2xRGMS at all on your new rig with it?

4xS is nice, probably the first edge AA that actually comes close to the V5's 4xAA. Unfortunately, it's not usuable with any degree of aniso over 2x in most recent shooters (even NOLF, for example), though those setttings are fine for non-shooters (Freedom Force, Sacrifice, etc.) My Morrowind settings are 12x9x32 with 2xAA and 4x aniso, all in-game settings maxed. I suspect, however, that those settings are largely playable thanks to the P4 2.4B and PC1066 RDRAM I'm running.

 

[Sharkfood]

but I'll try to see if using 1x2 or 4x4 options make any difference.

Well, those will make a difference if they actually apply. But I dont think Quincunx's algorithm is as simple as 3dfx's 1x2 or 4x1 or whatever neighbor blends.. so ya'd be trading one inaccurate capture for another.

 

[jb]

I suspect, however, that those settings are largely playable thanks to the P4 2.4B and PC1066 RDRAM I'm running.

Y'a think?

 

[Randell]

John, I wont quote as i know Dave doenst like nested quotes

I guessed your P4 rig would help give you the edge when it comes to stuff like that.

Still only RPG I'm doing right now is DAoC and I just cant get time to do another.

 

[Ailuros]

but I'll try to see if using 1x2 or 4x4 options make any difference.

Well, those will make a difference if they actually apply. But I dont think Quincunx's algorithm is as simple as 3dfx's 1x2 or 4x1 or whatever neighbor blends.. so ya'd be trading one inaccurate capture for another.

Eck I have to correct my own brainfart, Hypersnap (at least the version I have) has 1x4 and 2x2 post filtering options for voodoos.

I tried all of them and I'm not sure if it's my impression yet 2x2 comes out a bit brighter (since it uses a 1.3 gamma value vs. 1.0) so it doesn't leave the algorithm completely indifferent I guess.

I tried to capture screenshots from the SS:SE menu (openGL), and the resulting images no matter what screencapturing method used has little in common what clarity concerns to the real output. Haven't tried D3D in the said game yet NFS5 doesn't show the same behaviour. (nothing else than bilinear enabled in all occassions).

edit: typos