Hanners at EB: "Temporal anti-aliasing on R3x0 investig

[Killer-Kris]

I just had an interesting thought, since these new AA modes have different sample patterns than the purely spatial modes, is it not to likely that ATI has tools to compare the two modes and that is how they settled on the new sample patterns? Then in addition to that, is ATI going to provide reviewers with these tools for the release of the R420 so that they can show the improved image quality?

 

[Basic]

I'd say the opposite.
The way the new patterns look indicate that they haven't examined them practically with any good tools.

The first rule for those patterns should be that each single frame should be rendered with an optimal pattern. Then switch between different optimal patterns from frame to frame, with a slight shift to distribute all samples evenly in x and y direction (or at least as evenly as possible).

Ie With 6x AA there are 36 optimal patterns. So in 6x2T you can select two, and shift one of them 1/12 pixel in x and y direction. You can select two that give a rather good 12xAA pattern together, but it's more important that the two 6x patterns are optimal by themself that that the 12x pattern is optimal.

A lot of the flickering people see would be removed that way.

 

[Ylandro]

The first rule for those patterns should be that each single frame should be rendered with an optimal pattern.

The single frame patterns are not optimal for AA....
But who says these patterns aren't optimal for reducing flickering

Do you have any way to compare?

 

[vember]

as I see it, and as ylandrom pointed out (i think), is that the problem that stops temporal AA from being the best thing ever is that the entire backbuffer has to share the same AA-pattern.. If the GPU was able to choose a sample pattern randomly for each pixel, all the flicker would be gone and the edges wuld just be a little noisy instead. Much easier on the eyes. It would look like film!

Any thoughts about this? I reckon changing pattern on a per-pixel basis isn't that easily implemented in hardware and that's why we don't see it happening..

 

[Basic]

The single frame patterns are not optimal for AA....
But who says these patterns aren't optimal for reducing flickering

Do you have any way to compare?

If you want the pattern to be optimal to reduce flickering, then the patterns should be designed so that the intensity change as little as possible when switching. And that's best done by doing what I described.

What's done currenty doesn't cut it. Some of the patterns used are really horrible, like 4x patterns that are worse than a usual 2x pattern.

I don't have any other tools than FSAAviewer. But I'd say that it's farly obvious that an AAed edge can be a lot better than switching between two staircases that look much worse than the non-temporal edge AA.

 

[Dio]

Being able to change the pattern on a per-pixel basis doesn't help - it would mean straight edges might not look straight any more.

 

[Basic]

Agree, but it might not be so bad together with tempAA. The "lumps" would hop back and forth between the images, which can make the pixel flickering on two adjacent pixels 180º out of phase. That should reduce the perceived flicker.

So think of a checker board with pixel sized squares. Put one of the optimal patterns (as I described above) on white squares, and the other on black. Switch white and black squares to the next frame.

I don't dare to say anything definite without seeing it live though.

But I guess you can't comment on that (yet).

 

[vember]

Being able to change the pattern on a per-pixel basis doesn't help - it would mean straight edges might not look straight any more.

well, that's not really a problem now, is it?.. it has worked in offline CG.. Noise around the lines is present in photos as well and I don't see anyone complaining..

 

[akira888]

vember, in theatrical film-quality offline rendering typically 32+ samples per pixel are taken - that's quite different than the realtime world of 2-8 samples per pixel.

 

[Killer-Kris]

So does anyone know if this is going to become an official feature for the R3x0 series, or is it only going to be enabled officially for R420? I sure hope they enable it for the older chips since I'm more likely to continue to buy from a company that gives me a great deal of value even long after it's been sold.

Well either way, things sure look like they'll get interesting in the next couple of days.

 

[vember]

vember, in theatrical film-quality offline rendering typically 32+ samples per pixel are taken - that's quite different than the realtime world of 2-8 samples per pixel.

I'm not saying that it will be perfect, but I suspect it will be way better than the flicker we suffer of with the temporal AA used now. The GPU would still want to use predefined "optimal" patterns to choose from, the only difference would be that the pattern is selected on a per-pixel/quad basis. The magnitude of the temporal intensity variation would be the same as in the current temporal AA, but since the AA-pattern of each pixel wouldn't have to be directly correlated with the rest of the screen it wouldn't be apparent as flickery. And unless I'm mistaken it would work notibly better when the player is moving as well. However, it might be tricky to select good patterns on a per-pixel basis if you only want two patterns to toggle between in order to avoid the effect from being too noticable. But still, I'd take noise over flicker anyday.

 

[banksie]

So does anyone know if this is going to become an official feature for the R3x0 series, or is it only going to be enabled officially for R420? I sure hope they enable it for the older chips since I'm more likely to continue to buy from a company that gives me a great deal of value even long after it's been sold.

Given that this feature has been present and potentially enabled for the last two driver sets. Also given that ATi still seem to be persuing performance tweaks for their 8500 cards in the recent-ish driver builds then I would say the chances are very good.