Something to try: Geometry Based D3D FSAA Viewer

[Bambers]

Gf3Ti200, Det. 40.72, 6xS FSAA, 32x32, 1/8

The pattern looks identical to pocketmoon's shot but mine has different grey tones, atleast on my monitor. (Could it be jpg?)

Might be that the weighting on your shot is different. When screenshots were taken of 6xS in the first drivers it was in it was mentioned that the right hand 2 samples had double the weight of the others. pocketmoons shot is on newer drivers so maybe they've changed the weighting.

 

[Basic]

To be a bit more complete. Jittered sampling isn't without merit for textures, it's just not so obvious. It can actuallly (with a modified mip level algorithm) be used to get some "free" AF. Maybe not as good as "real" AF, and it will be angle dependent, but better than the AF effect you get from ordered grid texture sampling.

 

[Althornin]

To be a bit more complete. Jittered sampling isn't without merit for textures, it's just not so obvious. It can actuallly (with a modified mip level algorithm) be used to get some "free" AF. Maybe not as good as "real" AF, and it will be angle dependent, but better than the AF effect you get from ordered grid texture sampling.

What i dont understand is why jittered AA would be more "expensive" to use than ordered grid. You are still taking the same number of extra samples, right?
Or are you suggesting that this MAY just be an ATI specific implemintation perf hit?
Or maybe this is not the reason at all?

 

[3dcgi]

To be a bit more complete. Jittered sampling isn't without merit for textures, it's just not so obvious. It can actuallly (with a modified mip level algorithm) be used to get some "free" AF. Maybe not as good as "real" AF, and it will be angle dependent, but better than the AF effect you get from ordered grid texture sampling.

What i dont understand is why jittered AA would be more "expensive" to use than ordered grid. You are still taking the same number of extra samples, right?
Or are you suggesting that this MAY just be an ATI specific implemintation perf hit?
Or maybe this is not the reason at all?

It's more expensive because everything is on a grid, even jittered. It's just that jittered uses a much finer grid which requires more precision. More precision means more bits and more hardware.

 

[Basic]

I don't claim to know how it's calculated in ATI's cards, or any elses either. But there is some optimizations that could be done.

You'll have to take the same number of samples for both methods, but you could save on the texcoord calculation part. Instead of doing the calculation {s, t} = {a00*x+a01*y+a02, a10*x+a11*y+a12} / (a20*x+a21*y+a22) for every (sub)pixel, they might have made some "hyperbolic interpolator" that spits out the texcoords for four pixels in parallel. (The main benefit is to get rid of reciprocals.) You could probably do local linearisations of the equation to reduce calcs, and still get good accuracy. And you would probably find the simplest solution with a fixed, preferably orded, grid.

 

[Nagorak]

In theory, it sure would be better for diagonals than that R9700P pattern Wavey posted. Too bad the eye barely cares about diagonals

Uttar

Isn't that the whole point of doing a rotated grid to begin with? And isn't it more due to the fact that stuff that's almost diagonal isn't that bad to begin with (due to the stairs being closer), so it's better optimizing for near horizontal/vertical and letting diagonals be less optimal?

 

[Althornin]

It's more expensive because everything is on a grid, even jittered. It's just that jittered uses a much finer grid which requires more precision. More precision means more bits and more hardware.

No, i understand that.
The real question is: What makes it "mopre expensive" on ATI's hardware?
If the hardware is already there to handle it, what possible performance gain do you get from not using it?