Future of MSAA?

[Simon F]

But how do they get blended? Is there some form of accumulation buffer?

Retina-buffer(tm)?

Exactly. I think I'd much prefer to see some form of digital LP filter applied first.

 

[KimB]

And you should understand that if there's motion, they can't be blended out.

That's the point.
Temporal AA may work only if there's no movement in a scene and your display device (CRT or TFT) refreshes very slow.

I think that's an inaccurate statement. First of all, for a CRT, you want a high refresh for ATI's temporal dithering to work best. For LCD's a long response time really shouldn't make much difference, except for very high contrast edges.

Now, the description of tempral dithered samples being averaged in the eye is only accurate for a completely static scene. When it's in motion, however, it's a different story. This doesn't mean that temporal dithering is useless. What it does mean is that, for example, 2x with 2x temporal dithering is not equivalent to 4x AA, even at high framerates.

 

[Cryect]

And you should understand that if there's motion, they can't be blended out.

That's the point.
Temporal AA may work only if there's no movement in a scene and your display device (CRT or TFT) refreshes very slow.

I think you mean refreshes very fast but has a slow response time to refreshes (slow phosphors in the case of CRT's).

If you have a slow refresh rate Temporal AA flickers worse The slow response time allows for the accumulation to occur.

 

[KimB]

Heh. Ideally, a CRT will go what is effectively completely black between refreshes of the screen. There's no reason to desire motion trails on any display device.

 

[Cryect]

Heh. Ideally, a CRT will go what is effectively completely black between refreshes of the screen. There's no reason to desire motion trails on any display device.

Thats true but unfortunately as they have been putting better phosphors into monitors so you can have refresh rates unfortunately this has a detrimental effect on lower refresh rates. I've noticed older monitors I have no issues with 60Hz refresh rate on them but newer monitors <80Hz on a lot still has flickering issues.

BTW you can qualitatively with a couple monitors next to each other kinda tell how long they stay black by watching them while brushing your teeth with an electric toothbrush, its kinda funky cause your brain stabilizes the monitor casing but it looks like the image is jumping all over the place since its only emitting light part of the time.

 

[KimB]

Thats true but unfortunately as they have been putting better phosphors into monitors so you can have refresh rates unfortunately this has a detrimental effect on lower refresh rates. I've noticed older monitors I have no issues with 60Hz refresh rate on them but newer monitors <80Hz on a lot still has flickering issues.

Sure, but if they are made to run at the higher refresh rates this isn't an issue, is it?

 

[Cryect]

Nah its not really just was commenting on it since its a minor issue if you are using older stuff that doesn't run higher than 60Hz (ex VGA)

 

[KimB]

Nah its not really just was commenting on it since its a minor issue if you are using older stuff that doesn't run higher than 60Hz (ex VGA)

Well, you can always force higher refreshes with various sorts of software...

 

[flick556]

I always liked HAA(half antialiasing) I have seen it in one or two software renders, you simply render at twice the required resolution and then downsize to the desired resolution. The results are pretty good. I think this should already be possable in consoles given thier low res and it should soon be possable to render double 1024 res for pc. If this ever worked at acceptable speeds it could replace current systems, though I'm not sure if current AA is based on this idea or not.

 

[KimB]

I always liked HAA(half antialiasing) I have seen it in one or two software renders, you simply render at twice the required resolution and then downsize to the desired resolution. The results are pretty good. I think this should already be possable in consoles given thier low res and it should soon be possable to render double 1024 res for pc. If this ever worked at acceptable speeds it could replace current systems, though I'm not sure if current AA is based on this idea or not.

No. This would be the exact same thing as 4x OGSS. We've been there, done that. It's typically not worth the performance hit, particularly not compared to modern MSAA implementations.

 

[flick556]

I always liked HAA(half antialiasing) I have seen it in one or two software renders, you simply render at twice the required resolution and then downsize to the desired resolution. The results are pretty good. I think this should already be possable in consoles given thier low res and it should soon be possable to render double 1024 res for pc. If this ever worked at acceptable speeds it could replace current systems, though I'm not sure if current AA is based on this idea or not.

No. This would be the exact same thing as 4x OGSS. We've been there, done that. It's typically not worth the performance hit, particularly not compared to modern MSAA implementations.

Good to know; how was the quality, it's hard for me to compare given the differences between hardware and software renders and I have never seen 4x OGSS before. It seems like such a simple solution that could produce very good results. Though the performance hit would be high.

 

[Cryect]

Quality isn't that great in OGSS (ordered grid super sampling). The issue is its easy for edges to not really get as much antialiasing as possible in OGSS. If you have something thats perfectly horizontal or vertical in 4x OGSS it will only ever get the equilvalent of 2x AA since it always hits two of the samples at the same time.

RGSS (Rotated Grid Super Sampling) is a little better in that its picks an angle less likely for edges to align up with. Then move each of the points slightly so its not possible if an edge is moving slowly and continous to not go over multiple sample points at the same time.

Best overall with super sampling is every pixel has their own sampling pattern or if thats not possible have say 16 or 64 different sampling patterns applied to every 4x4 pixel block or 8x8 block.

MSAA (Multisampling AA) is preferred in general because it only breaks pixels into 4 pixels if an edge goes through that pixel and then once it breaks into 4 pixels it of course takes 4 times the processing there. Since the screen isn't mostly edges this can easily perform much better with a lot less bandwidthm and pixel processing hit. Of course as pixels approach the size of pixels we shouldn't use MSAA since there is a slight overhead hit for handling if a pixel is being super sampled or not.

 

[KimB]

Good to know; how was the quality, it's hard for me to compare given the differences between hardware and software renders and I have never seen 4x OGSS before. It seems like such a simple solution that could produce very good results. Though the performance hit would be high.

This was the FSAA used with the GeForce 256/GeForce2 series of processors.

 

[KimB]

MSAA (Multisampling AA) is preferred in general because it only breaks pixels into 4 pixels if an edge goes through that pixel and then once it breaks into 4 pixels it of course takes 4 times the processing there.

This isn't true. MSAA never computes more than one color value per triangle per pixel. The reason why you get some extra cost at triangle edges is that when more than one triangle intersects the pixel, more than one color value ends up being calculated. So it's not an automatic 4x performance hit for that pixel.

 

[Cryect]

MSAA (Multisampling AA) is preferred in general because it only breaks pixels into 4 pixels if an edge goes through that pixel and then once it breaks into 4 pixels it of course takes 4 times the processing there.

This isn't true. MSAA never computes more than one color value per triangle per pixel. The reason why you get some extra cost at triangle edges is that when more than one triangle intersects the pixel, more than one color value ends up being calculated. So it's not an automatic 4x performance hit for that pixel.

Okay, thanks makes sense. So just bandwidth hit then if a pixel gets hit by an edge. And really its not then that more than one color value is being calculated since its possible that would have occured assuming its overdraw.

 

[KimB]

No, that's not it, either. For each triangle that shares the same pixel, one color value must be calculated.

 

[Fodder]

Very interesting thread guys. Anyone have the spare time to explain to me (in not too technical terms if possible :?) why it is MSAA can't AA alpha textures?

 

[KimB]

MSAA only determines coverage based upon depth tests. Alpha testing is based upon per-pixel, texture-dependent information, and thus would require that the entire pixel be calculated multiple times in order to do any anti-aliasing (independent of the software design: it is possible to anti-alias alpha tested-surfaces in a variety of ways in software).

 

[Ailuros]

I was actually hoping to read a couple of good ideas for possible future antialiasing algorithms...

Anyway I personally never liked or like Quincunx or 4x 9-tap and call it a matter of taste if you prefer. I just can't stand the blur the additional filters add in both cases to the scenery, even worse in the latter case.

Quick experiment in SS:SE (Direct3D) and in an obnoxiously low resolution on purpose and with some text added to the mix:

2xRGMS

Quincunx

2xOGSS

(all pulled through print screen)

Another thing is that increasing the resolution will not eliminate aliasing; in order to get rid of aliasing you have to antialias.

 

[KimB]

Personally, it's the blurring of the text with Quincunx that really gets me. I don't think I'd really notice the slight blurring of the scenery.