DX11 MSAA

[NaturalViolence]

Gabbyjay Wrote:AFAIK AMD drivers support SGSSAA too.

4xSGSSAA, for example, essentially is RGSSAA, as you end up with a rotated grid for optimal EER. It's only scaled a bit, so you wont get vast subpixel-free areas between to neighbouring pixels. A rotated grid is a form of sparse grid in this case.

You can also see this on nVidia cards, when you have a look at the sample mask patterns for MSAA/SGSSAA:
http://pics.computerbase.de/1/5/8/8/6/41.png

That image is referring to a 7900 GTX. Different GPUs use different sample patterns at the hardware level for the same AA method at the driver level. As far as I know SG wasn't implemented yet on the geforce 7 series.

If AMD does support SGSSAA they still have a completely different implementation of it that's not going to look the same.

[Gabbyjay]

The 4x subpixel-mask looks the same for AMD and nVidia, it's only mirrored. You can check it out yourself with FSAA viewer.
(It is still the same sample pattern on today's cards the image was only an example!)

Compare:
http://pics.computerbase.de/2/6/9/3/9/11.png (Radeon 5800)

and
http://pics.computerbase.de/2/8/9/4/1/49-s.png (GTX 480)




Here's the 8x sparse grid on an ATI 5800:
http://pics.computerbase.de/2/6/9/3/9/12.png

Its clearly a sparse grid!


(This one looks different on nVidia-cards, of course, since there is no "optimal" pattern in this case. Sometimes the nVidia-Version has the advantage, sometimes the ATI one, depending on the angle of the edge. All in all they perform about the same, quality-wise.)


Edit:
Also notice in the image how both the samples for geometry and textures are exactly the same.
This only makes sense, cause they use a sparse grid in MSAA for years now.
Why should they use a different pattern for SSAA? That dont make sense.

[NaturalViolence]

Gabbyjay Wrote:All in all they perform about the same, quality-wise.

Having tried both my eyes beg to differ. Not only does AMDs SSAA produce noticeably poorer IQ in some games it even causes strange artifacts such as missing pieces of transparent textures sometimes. I don't know how either one is implemented at the driver/hardware level but clearly there are some key differences in the implementation. It seems as if they both use RG for 2x and 4x and SG for 8x with current cards.

Gabbyjay Wrote:Why should they use a different pattern for SSAA? That dont make sense.

They don't. All I'm saying is that Nvidia and AMDs SSAA implementations are different. We don't even know if either one does the resolve in hardware or software for example.

[Gabbyjay]

Quote:Having tried both my eyes beg to differ. Not only does AMDs SSAA produce noticeably poorer IQ in some games it even causes strange artifacts such as missing pieces of transparent textures sometimes.

You are comparing the end result now (many factors can come into this one) - my comparison was on the 4x sample patterns, as i wrote.
Both nVidia's and AMD's 4x sample pattern provide the same results, because they ARE the same, as can be easily prooved by using FSAA-viewer. I already provided screenshots for this.

It is the same pattern for both companies, only it is mirrored, which does not make a difference here.

Quote:It seems as if they both use RG for 2x and 4x and SG for 8x with current cards.

I already tried to explained that.

The rotated Grid in the 2x and 4x modes are a form of sparse grid.
To put it in other words:
The optimal sample pattern for 2x or 4x SGSSAA looks like a rotated grid (but is scaled compared to a rotated standard OGSSAA pattern), as this is already the optimal sparse sampling mask for that sample count for best EER.
Got it?
If not, try it yourself using a sheet of paper and a pen, its not that hard to figure out the pattern.

Quote: All I'm saying is that Nvidia and AMDs SSAA implementations are different. We don't even know if either one does the resolve in hardware or software for example.

No, you said AMD dont use SGSSAA and use only rotated grid.

That is not true.

The 4x mask already slightly differs from a rotated standard OG (with equal distance between each subpixel), if you look closely.

The 8x mask is not rotated at all, as i showed with my picture in the last post.

You can clearly see the (seemlingly) irregular nature of the subpixels.


To sum it up:
Both their sampling masks perform the same (8x differs but its just another variation for the solution), and both companies use SGSSAA.


And they use a sparse grid for years, even before they included SGSSAA, as the multisampling used and uses the same patterns.




If you consider the AA process at whole and not just the sampling masks, there can be further differences, of course. No argument here.

[NaturalViolence]

Gabbyjay Wrote:If you consider the AA process at whole and not just the sampling masks, there can be further differences, of course. No argument here.

As long as you agree to that I'll concede that you were right about the sample masks.

[Gabbyjay]

Sure.
I never wanted to say the end result would be the same on both cards. I was only referring to the sample pattern as such. Simple misunderstanding.


Another thing:
Is there any information about the OGL-Plugin's CSAA regarding amount of color- and coverage-samples?

It's a bit confusing as these options use the same names as the ones in the nVidia drivers (or nVidia inspector), yet they seem to have a different meaning.
8xQ is 8x MSAA without any CS in inspector, in Dolphin there is 8xQ CSAA. There's also 8x CSAA in Dolphin, and 8x.
nVidia uses 8x for 4 color and 4 coverage samples, where 8x in Dolphin seems to be plain MSAA.
This is confusing, is there an overview clearing this up?