AA/AF enhancements

[sonix666]

SGI has always been saying in their documentation that random sampling AA should at least use 16 samples to look good. Doing random sampling at 4 or 6 samples would look ugly in my opinion.

 

[Ailuros]

Please guys, "lose" not "loose". The first is "to misplace something", the second is "not tight". I'm seeing this mistake so often now even I'm beginning to lose my mind!!!

Guilty as charged and not trying to hide either *sticks hand in the air*.

Too much of a wild goose chase I figure, hence the addiction to the double O

Considering I'm a greek, minor quirks like that should be forgiven *hangs head*

 

[Arun]

A question about FAA:

After relooking at Matrox's FAA paper, they say they use 16x SSAA when the hardware determines you need AA.
They also state that in a typical scene, you've often got 5% fragments.

So, that means you've got: 100 + 80% = 180% normal fillrate needs
That means you've got 55% effective fillrate! Nearly as if you were using 2x SSAA! Looking at performance numbers, it does indeed look similar from what you'd expect out of a 2x SSAA performance hit...

Now, my question is simple: would be possible to use MSAA instead of SSAA, and possibly get much less of a performance hit, in the order of 2x MSAA instead, or even less?

Also, I'd guess implementing GC on FAA is possible, isn't it? Would be nice to see that...


Uttar

 

[KimB]

The problem with FAA is that the edge detection on the Perhelia is not perfect. I don't believe it ever can be. This makes the AA look excellent most of the time, but there is a total lack of AA every once in a while.

Personally, I'd much rather have a form of FSAA that gets all edges at a lower quality than one that misses an edge here and there.

 

[KimB]

SGI has always been saying in their documentation that random sampling AA should at least use 16 samples to look good. Doing random sampling at 4 or 6 samples would look ugly in my opinion.

Well, for totally random sampling, I don't doubt it.

But for specifically-chosen and varying sample patterns, I think around 4 samples per pixel could be enough.

 

[Xmas]

Uttar, I'm pretty sure they are not using SSAA, even if they call it SSAA in the paper. Supersampling for polygon edges only doesn't make sense at all.

 

[Arun]

Uttar, I'm pretty sure they are not using SSAA, even if they call it SSAA in the paper. Supersampling for polygon edges only doesn't make sense at all.

Hmm, yes, I think you're right there.
While they DO use the term "supersample" , they also say it "does not impact fillrate".
Bah, bad wording I guess. Sorry for the stupid question then...

I must admit I never really understood how FAA worked. So, from my understanding, you got a fragment buffer in *memory*, not cache, which stores the subpixel coverage map of every fragment? How exactly would it be organized, for example? AFAIK, they don't use a fixed amount of memory, so how does it work?


Uttar

 

[Randell]

Considering I'm a greek, minor quirks like that should be forgiven *hangs head*

which is why I let it go Ail. You know how I love to pick up on your fowl spelling

 

[Bambers]

This reminds me of this paper :

That reminds me a bit of the whitepaper on how the AA on the 8500 works (well should have worked anyway, it never turned out like that)

The pdf is still somewhere on atis site I think.

 

[KimB]

Hmm, yes, I think you're right there.
While they DO use the term "supersample" , they also say it "does not impact fillrate".
Bah, bad wording I guess. Sorry for the stupid question then...

The fragment detection doesn't impact fillrate.

Fragment anti-aliasing does.

While the whitepaper does not state whether or not fragment AA uses supersampling, it seems to make sense that it would have to. Otherwise Matrox would be able to use a more advanced form of FSAA than simple 4x supersampling when "normal" FSAA is selected (Btw, I am aware that this argument isn't perfect, but I think that this, combined with the performance hit of FAA, makes it seem highly likely that supersampling is used on fragments).

Edit:
Whitepaper can be found here.

 

[Xmas]

While the whitepaper does not state whether or not fragment AA uses supersampling, it seems to make sense that it would have to. Otherwise Matrox would be able to use a more advanced form of FSAA than simple 4x supersampling when "normal" FSAA is selected.

I don't buy that argument. 4x4 multisampling is even simpler than 4x4 supersampling in this case.

 

[KimB]

I don't buy that argument. 4x4 multisampling is even simpler than 4x4 supersampling in this case.

Except then it should easily be able to do 4x multisampling when using "normal" FSAA.

 

[Xmas]

Except then it should easily be able to do 4x multisampling when using "normal" FSAA.

Not necessarily. But most probably it's not even "full multisampling", just storing two colors and a coverage mask or something similar is more likely.

 

[Dave Baumann]

Multi-sampling is effectively supersampling at polygon edges anyway.

 

[Xmas]

Multi-sampling is effectively supersampling at polygon edges anyway.

With 2x AA, or with very small polygons, yes. But the more samples you use, the less it resembles supersampling.

 

[OpenGL guy]

Multi-sampling is effectively supersampling at polygon edges anyway.

Actually, no. Imagine drawing a quad as two triangles with a checkboard texture applied. If you actually did supersampling on the edges, then you would see a seam where the two triangle meet. This doesn't happen with multisampling because the same texture sample is used regardless of which triangle the edge pixel falls in (sample point is always in the center).

Supersampling applied to the same quad would cause the edges of the checkers to be filtered, this also doesn't happen with multisampling.

 

[Testiculus Giganticus]

As I said, Matrox only opened the way, that`s all.It is only a first try-and it is far from perfect.Don`t think that FAA will go on exactly like Matrox did it, it`s only going to evolve.I think that we can get edge detection algorithms to a level that ensures that the number of "misses" in a scene is under 5%- I think that it`s more than OK.And yes, the drop can be reduced by using a different way of getting the AA samples

 

[Ailuros]

Considering I'm a greek, minor quirks like that should be forgiven *hangs head*

which is why I let it go Ail. You know how I love to pick up on your fowl spelling

**the crowd may excuse another OT**

Oh I also type relatively fast for an average user and no I don´t look at the keyboard either (even dumber excuses). Feel free to not let it go from now on buddy and all you´ve got to do is edit your text as often as possible

j/k of course

 

[3dcgi]

While the whitepaper does not state whether or not fragment AA uses supersampling, it seems to make sense that it would have to. Otherwise Matrox would be able to use a more advanced form of FSAA than simple 4x supersampling when "normal" FSAA is selected (Btw, I am aware that this argument isn't perfect, but I think that this, combined with the performance hit of FAA, makes it seem highly likely that supersampling is used on fragments).

That argument doesn't make any sense to me. The observations make the argument for the actual AA technique to be more like multisampling as Xmas said. If the hardware was able to 16x supersample pixels the fall back supersampling mode would have used a spare sample pattern.

Also, the only way for an edge detection algorithm to catch all edges including intersection edges is to do full scene AA or to transform all the edges before starting rasterization. At least those are the only ways I know of.

 

[OpenGL guy]

Also, the only way for an edge detection algorithm to catch all edges including intersection edges is to do full scene AA or to transform all the edges before starting rasterization. At least those are the only ways I know of.

Or just do MSAA MSAA will catch intersections. The reason is that each subsample has it's own Z value. Apparently, Matrox's FAA uses the same Z value for all subsamples so it will break down on intersections.