R520 will support 8x AA?

[Ailuros]

No, it won't. Yes, you're doubling the sample density, but that doesn't mean you're doubling the benefit. The benefit of AA is the removal of situations where aliasing is noticeable. Many more situations are removed in the lower range of AA than in the higher range (e.g. 16x sparse would be virtually indistinguishable from 32x sparse).

16x and 32x sample densities play in an entirely different league and you know it very well. I'm not so sure more than 16x sparsed samples make actually more sense than to move straight to a semi-stochastic algorithm instead.

Even 10.000 samples would be hard to disginuish compared to 16x in some cases, like that one:

http://web.onetel.net.uk/~simonnihal/assorted3d/samppat.html

If the sample positions will be optimal with 8x sparsed there very well be quite a few cases where it will make a disguinshable difference compared to 4x sparsed.

Are you sure that judging the EER is going to be a practical means of judging quality?

Since the subject is obviously about Multisampling I don't see why not; under the presupposition that the sample placement is optimal or close to optimal.

If you get patterns like those:

http://www.3dcenter.de/artikel/anti-aliasing-masken/index6.php

....it's natural that it won't have the results it should have.

Here's an example of a good 16x sparsed sampled grid:

http://www.beyond3d.com/forum/showthread.php?p=46158#post46158

and here 8x,6x and 4x sparsed sampled grids:

http://www.beyond3d.com/forum/showthread.php?p=44389#post44389

 

[no-X]

A 12x12 grid is already supporting 12x AA firaly optimally.

Yes, but if the grid size isn't an exact multiple* of number of samples used per pixel, the result can't be good. Just try to draw 12*12 grid and fit 8 samples into it. It's something like to use 3 samples with 4x4 grid or 6 samples with 8x8 grid etc.

(* or is better to say integer multiple?)

 

[Geo]

I think they might, but possibly more as a by-blow to optimizing 4x as single loop more than anything else.

 

[KimB]

16x and 32x sample densities play in an entirely different league and you know it very well. I'm not so sure more than 16x sparsed samples make actually more sense than to move straight to a semi-stochastic algorithm instead.

If you mean that 32x will look much better than 16x when used with stochastic sampling, then yes, they are in a different league. But that's not the point I was trying to make.

You'd be very, very hard-pressed to find a difference between 16x sparse and 32x sparse in any current game. As you increase the number of samples, you decrease the number of situations where aliasing is still apparent. So it only stands to reason that doubling the number of samples won't give the same effect at every stepping.

So the move from 2x AA to 4x AA is going to be much more impressive than a move from 4x to 8x (assuming nothing but sparse sample patterns used).

 

[Dave Baumann]

Yes, but if the grid size isn't an exact multiple* of number of samples used per pixel, the result can't be good.

Sorry, do you have some actual basis for this? The most important things are to have a reasonably wide dispersal, and not to have samples crossing each other on the horizontal or vertical axis, as you then start reaching the point o diminishing returns.

 

[hoom]

http://web.onetel.net.uk/~simonnihal/assorted3d/samppat.html

Yes.master.I.obey.

 

[no-X]

Sorry, do you have some actual basis for this? The most important things are to have a reasonably wide dispersal, and not to have samples crossing each other on the horizontal or vertical axis, as you then start reaching the point o diminishing returns.

I will not argue, but the result of 8xAA/12x12grid can't be better than 8xAA/8x8 (or 16x16/24x24 etc.) grid. Especially on near vertical/horizontal lines. I like continuous color gradient more than irregular jumps.

detail:



You can try any other sample pattern, but it will never be better than / or competitive with 8x/8x8-16x16-24x24.

edit: arrrse: we talk about multisampling - your post si about moire (=textures)

 

[Dave Baumann]

You've only picked one particular gradient there. What sample pattern is being used for 8x?

 

[Ailuros]

If you mean that 32x will look much better than 16x when used with stochastic sampling, then yes, they are in a different league. But that's not the point I was trying to make.

You'd be very, very hard-pressed to find a difference between 16x sparse and 32x sparse in any current game. As you increase the number of samples, you decrease the number of situations where aliasing is still apparent. So it only stands to reason that doubling the number of samples won't give the same effect at every stepping.

So the move from 2x AA to 4x AA is going to be much more impressive than a move from 4x to 8x (assuming nothing but sparse sample patterns used).

More or less the same with anisotropic filtering samples, with the only other difference that there algorithms are adaptive (whereby "adaptive" has become a too wide term for my taste lately).

Just because I can come up with a ton of situations where it's hard to see a difference between 4x and 8x AF, it doesn't mean that the difference isn't there or that there isn't a need for either 16x or even higher sample densities in the future.

I admit that this gets somewhat silly though with the tons of alpha tests and abysmally bad texturing I can still see in recent games. Implementations like Transparency-AA are obviously a step in the right direction, but I haven't yet understood why IHVs have to cover for "cheap" game content in the end.

If there wouldn't be natural performance restrictions I'd personally have the result on textures 4*4 SSAA delivers, even if the EER stays at 4*4. That doesn't change the fact though that the difference between 4xS and 8xS isn't just subtle as another example and that on both axes and purely of course in terms of polygon edges.

 

[KimB]

Just because I can come up with a ton of situations where it's hard to see a difference between 4x and 8x AF, it doesn't mean that the difference isn't there or that there isn't a need for either 16x or even higher sample densities in the future.

Right, but 2x the samples isn't 2x the visual quality. Fortunately, neither is 2x the samples half the performance. Personally, I don't expect a whole lot more to happen with respect to AA for quite some time.

That is, I expect that within two-three years, ATI will implement something similar to nVidia's transparency AA (selective supersampling...alpha mask could come in driver much sooner). I'm expecting nVidia's next architecture (that should be out at the end of next year) to implement sparse sampling at greater than 4 samples per pixel. But I don't expect much more than that.

 

[Ailuros]

Nothing against that, but multi-board configs already allow quite high sample densities with quite high edge equivalent resolutions. 16x SLi-AA for instance delivers already an 8*16 EER, albeit the sample distribution might need a few improvements over time.

I'm expecting nVidia's next architecture (that should be out at the end of next year) to implement sparse sampling at greater than 4 samples per pixel. But I don't expect much more than that.

And in would mean what exactly for ultra high end SLi systems and especially in terms of EER considering the former paragraph?

What if also Transparency AA sample densities adapt to the increased amount of MSAA samples on single board sollutions? Something like 8xMSAA + 8xTSAA?

***edit: if Vista comes out at the end of next year than yes; if not they wouldn't want to miss that train....

 

[no-X]

You've only picked one particular gradient there.?

Yes, but isn't this example of angle, which needs AA more than any other?

What sample pattern is being used for 8x?

It is not very important which one is chosen, I tried more of them and the result is very similar in this case:



You are true, that 8x/12x12 is possible, but 8x8 etc. solution is better on my opinion and gives much better result for low angles, which are mostly rated. (RG is better than OG only at low angles and despite it is RG rated as superior to OG - our situation is quite similar)

I only hope you aren't pointing, that R520 will use 8x/12x12 solution

 

[KimB]

And in would mean what exactly for ultra high end SLi systems and especially in terms of EER considering the former paragraph?

Ah, you don't need a higher-resolution grid for SLI. Consider: let's imagine that the pixel center for the second card is offset a half pixel sample away from the first card's pixel center. This would allow all of the second card's samples to lie off of the first card's gridlines, such that a, for example, 8x8 grid would turn into an effective 16x16 grid.

What if also Transparency AA sample densities adapt to the increased amount of MSAA samples on single board sollutions? Something like 8xMSAA + 8xTSAA?

That would be an interesting change, but we'll see. Such a thing would make the implementation more complex, and I doubt it'd be that great for the end user.

 

[Dave Baumann]

Yes, but isn't this example of angle, which needs AA more than any other?

Its one, change the angle slightly (which could be another just as important) and the gradients will change in both cases. Have you tried playing around with other sample positions for 8x on a 12x12 grid, ones that are slighty more ordered (well, less random) than the sample pattern used before?

You are true, that 8x/12x12 is possible, but 8x8 etc. solution is better on my opinion and gives much better result for low angles

Any implementation would have to cope with multiple AA types.

I only hope you aren't pointing, that R520 will use 8x/12x12 solution

Not a clue what it supports.

 

[Ailuros]

Ah, you don't need a higher-resolution grid for SLI. Consider: let's imagine that the pixel center for the second card is offset a half pixel sample away from the first card's pixel center. This would allow all of the second card's samples to lie off of the first card's gridlines, such that a, for example, 8x8 grid would turn into an effective 16x16grid.

With more than 4x MSAA samples per clock, a much higher EER could be achieved. Exactly the reason why 8xS is so much better than 4xS.

That would be an interesting change, but we'll see. Such a thing would make the implementation more complex, and I doubt it'd be that great for the end user.

If they'd reach 8xMSAA/clk I don't see why not.