"just spoke to an ATi guy and an nVidia guy"

[kyleb]

Not identical, but equivalent.

i don't see how you can argue equivalence. even if the sample patterns were just mirror images of each other, certain angles would favor one patten more so than the other. regardless, the patterns are not simply horizontally flipped, but they are slightly skewed from each other as well.


*edited for euan, who was right on both counts.

 

[euan]

Not identical, but equivalent.

i don't see how you can argue eqivalentce. even if the sample patrens were just mirror images of eachother, certain angles would favor one patren moreso than the other. regardless, the patrens are not simply horizonaly fliped, but they are slighty skewed from eachother as well.

Eitehr ist time fur bed, or youve ben on teh drinc.

 

[Quitch]

Nah, he's obviously just started using an egonomic keyboard.

 

[Xmas]

Not identical, but equivalent.

i don't see how you can argue eqivalentce. even if the sample patrens were just mirror images of eachother, certain angles would favor one patren moreso than the other.

Huh? Sure, if you mirror the sample pattern, you mirror the angles where the pattern does a good or bad job respecively, as well. But neither one is better or worse than the other overall. Hence, equivalence.

regardless, the patrens are not simply horizonaly fliped, but they are slighty skewed from eachother as well.

They aren't skewed. The position of the texture sampling point (or the center of the pixel) is different, but other than that they are mirror images. I checked it by mirroring and layering the patterns over each other with PSP.
http://www.beyond3d.com/reviews/ati/r420_x800/index.php?p=13#aa
http://www.beyond3d.com/previews/nvidia/nv40/index.php?p=17

 

[kyleb]

so, the texture sampling point is the center of the pixel for both ati's and nvidia's multisampling, but they are different? i don't follow you here. it seems to me that for them to be different, at least one would have to be taking the texture sample from somewhere other than the center of the pixel. since they are both in the clearly center, it seems to me that the geometry sampling points are where the difference is. am i missing something here?


as for your comment that mirroring the sample patten results in a situation where "either one is better or worse than the other overall"; this assumes that the mirrored angles, where each pattern does a good or bad job respectively, are equally common in practice. while i am no expert on the science of antialising by any means, it seems reasonable to me to assume that some angles are more common in modern gaming than others; hence, one sample pattren would be more effective over all when compared to its mirrored version.

 

[Rys]

I'm guessing Xmas's thinking is somewhat similar to a "sum of all histories" approach, where over time, both approaches will produce an equal quality of image. One gives better quality with some geometry, one gives better quality with other geometry, but assuming (and it seems reasonable to do so) that geometry that favours either sample grid is no more likely in a scene than other geometry, neither one will turn up noticeably better image quality than the other, given their inherent similarities (the equivalence).

That's how I see it.

Rys

 

[Xmas]

so, the texture sampling point is the center of the pixel for both ati's and nvidia's multisampling, but they are different? i don't follow you here. it seems to me that for them to be different, at least one would have to be taking the texture sample from somewhere other than the center of the pixel. since they are both in the clearly center, it seems to me that the geometry sampling points are where the difference is. am i missing something here?

It doesn't really matter what you take as a reference point, the pixel center, the texture sampling point, or the geometry sampling points. Fact is, if you mirror one of the patterns, you can see that the geometry sampling points are at the same positions relative to each other. The pixel center and the texture sampling point are not. But that doesn't affect the quality of smooth edges, it only slightly affects the texture sampling and "shifts" the triangles a bit.

as for your comment that mirroring the sample patten results in a situation where "either one is better or worse than the other overall"; this assumes that the mirrored angles, where each pattern does a good or bad job respectively, are equally common in practice. while i am no expert on the science of antialising by any means, it seems reasonable to me to assume that some angles are more common in modern gaming than others; hence, one sample pattren would be more effective over all when compared to its mirrored version.

"Some angles are more common" - this may be true if you mean that e.g. horizontal or vertical edges are more common than 45° ones. But a 30° angle "to the left" is usually equally common as a 30° angle "to the right" (30° to 150°). There's no reason why one would occur more often than the other.

Rys's explanation is very good.

 

[kyleb]

yet, for instance, in first person shooters we commonly have our gun protruding from the right side of the screen at about a 30° angle; but what do we have to make the opposing angle equally common?


as for the sample points; lets assume a situation were we have geometry sampling points which are at the same positions relative to each other as in ati's x2 but shifted to where one geometry sampling point is moved to being barely diagonally left and below the texture sample point, leaving the other geometry sampling point all the way in the upper right corner. would not such set of sampling locations land on different geometry when compared to the other? would this not be the same case in the difference between ati's x2 and a horizontally mirrored version of nvidia's x2, but with the difference in ati's and nvidia's simply being less extreme?

oh and i also wanted to point out that while you are correct in saying that the "geometry sampling points are at the same positions relative to each other" for x2; the x4 patterns are offset slightly more.

 

[Xmas]

yet, for instance, in first person shooters we commonly have our gun protruding from the right side of the screen at about a 30° angle; but what do we have to make the opposing angle equally common?

You have a point there. But many FPS games let you choose the weapon hand. And it depends a lot on the weapons design which pattern looks better.

as for the sample points; lets assume a situation were we have geometry sampling points which are at the same positions relative to each other as in ati's x2 but shifted to where one geometry sampling point is moved to being barely diagonally left and below the texture sample point, leaving the other geometry sampling point all the way in the upper right corner. would not such set of sampling locations land on different geometry when compared to the other? would this not be the same case in the difference between ati's x2 and a horizontally mirrored version of nvidia's x2, but with the difference in ati's and nvidia's simply being less extreme?

Yes, there is this difference, and that's what I meant by saying it shifts triangles a bit.

NVidia changed the position of the texture sampling point from NV20 to NV25. In NV20, it is the same as the top left geometry sampling point. In NV25 it is in the center of the geometry sampling points. This didn't change edge quality overall.

oh and i also wanted to point out that while you are correct in saying that the "geometry sampling points are at the same positions relative to each other" for x2; the x4 patterns are offset slightly more.

1 pixel in a 100 pixel high image - ok, but very slightly