Why do we still have fixed function AA and filtering?

[OpenGL guy]

I would bet on it.
Static screenshots like the pinwheel type that we have seen in several reviews show that ATI's 6x has better edge quality than nVidias 8x

Sorry, for some reason, I was thinking of texture aliasing.

Anyway, 8xS vs. ATI's 6x should be relatively similar in edge AA quality, as far as the sample patterns are concerned (each one should win at different angles). One crucial difference, however, is that ATI's uses a gamma-correction algorithm, which will skew it slightly in ATI's favor.

Except that the sample positions used in 8xS are so poor that it barely looks better than ATI's 4x AA (if you ignore gamma correction). If you factor in gamma correction, ATI's 4x AA looks better.

 

[KimB]

Except that the sample positions used in 8xS are so poor that it barely looks better than ATI's 4x AA (if you ignore gamma correction). If you factor in gamma correction, ATI's 4x AA looks better.

I'm not sure that you can say that, at least, not by comparing the above shots. Because ATI's 4x AA has the gamma correction factored in.

If you're going to ignore the gamma correction, you'd have to develop a software test to show the edge effects of a custom AA sampling pattern.

Side note: I still think ATI should offer a way to customize the gamma correction of their FSAA. I don't know if it's possible in hardware right now, but it sure would be a nice addition. Not every monitor has the same voltage/brightness characteristics.

 

[Dave Baumann]

Side note: I still think ATI should offer a way to customize the gamma correction of their FSAA. I don't know if it's possible in hardware right now, but it sure would be a nice addition. Not every monitor has the same voltage/brightness characteristics.

AFAIK, you can - any game that has gamma controls will alter the gamma for AA as well.

 

[Ailuros]

Anyway, 8xS vs. ATI's 6x should be relatively similar in edge AA quality, as far as the sample patterns are concerned (each one should win at different angles). One crucial difference, however, is that ATI's uses a gamma-correction algorithm, which will skew it slightly in ATI's favor.

Not from what I can tell. I'd personally place 8xS damn close to 4x sparse, in some departments slightly better. 6x sparse is in a league of it's own.

Here's the 6x sparse shot, same spot:

Zoom in, in all three of them, the differences are there.

 

[Ailuros]

I'm not sure that you can say that, at least, not by comparing the above shots. Because ATI's 4x AA has the gamma correction factored in.

Even w/o gamma correction it's still a 4*4 vs a 4*4 grid. 6*6 is one noticable step above and not just in the former shots, but in real time too.

Let's blame it all on the gamma correction.

 

[KimB]

Even w/o gamma correction it's still a 4*4 vs a 4*4 grid. 6*6 is one noticable step above and not just in the former shots, but in real time too.

Let's blame it all on the gamma correction.

I really don't understand how you can say that the size of the grid is what's important here.

 

[Ailuros]

I really don't understand how you can say that the size of the grid is what's important here.

I've seen both in realtime if that helps at the other and not just in one game. If you'd like to believe that 8xS is a rough equivalent to 6x sparse than it's your problem.

If the size of the grid would be as irrelevant, than there wouldn't be much sense to compare sparse or rotated with ordered grid patterns either. Same amount of samples, same output.

8xS is still just a combination of 2xRGMS with 4xOGSS and that's effectively 2*2 + 2*2 = 4*4.

 

[KimB]

If the size of the grid would be as irrelevant, than there wouldn't be much sense to compare sparse or rotated with ordered grid patterns either. Same amount of samples, same output.

Because the grid is nothing more than a visual construct of what's happening. There's no real reason that the actual algorithm ATI uses must have anything to do with any sort of grid.

The important things to deal with when looking at AA are the sampling method (SSAA, MSAA), whether or not it's full screen (FAA vs. anything else), the number of samples, and the sample pattern.

The fact that the 8xS mode is effectively on a 4x4 grid is nothing more than a mathematical construct. It has no bearing in either in the performance or the visual quality of the technique.

 

[OpenGL guy]

8xS is still just a combination of 2xRGMS with 4xOGSS and that's effectively 2*2 + 2*2 = 4*4.

Maybe I'm not understanding your notation, but 2*2 + 2*2 = 4*2. 4*4 would be 16 samples

Also, above where you talk about 6x AA and use an equation like (1024*6)*(768*6), this doesn't make sense to me as you don't increase X and Y each by a factor of 6 to get 6x AA. 6x AA is 6 samples per pixel so it should read (1024*768)*6.

If I've misinterpreted something, please let me know.

 

[KimB]

8xS is still just a combination of 2xRGMS with 4xOGSS and that's effectively 2*2 + 2*2 = 4*4.

Maybe I'm not understanding your notation, but 2*2 + 2*2 = 4*2. 4*4 would be 16 samples

I think he's trying to say it's 8 samples laid out in a 4x4 grid (half of the grid positions not sampled)

 

[OpenGL guy]

8xS is still just a combination of 2xRGMS with 4xOGSS and that's effectively 2*2 + 2*2 = 4*4.

Maybe I'm not understanding your notation, but 2*2 + 2*2 = 4*2. 4*4 would be 16 samples

I think he's trying to say it's 8 samples laid out in a 4x4 grid (half of the grid positions not sampled)

But that's not how the positions are arranged in 8xS. In 8xS they are like:

x x x x

x x x x

which is why the edge quality isn't so hot.

 

[KimB]

But that's not how the positions are arranged in 8xS. In 8xS they are like:

x x x x

x x x x

which is why the edge quality isn't so hot.

I'm pretty sure that's 8x, not 8xS. I thought 8xS was:

.x.x
x.x.
.x.x
x.x.

 

[OpenGL guy]

But that's not how the positions are arranged in 8xS. In 8xS they are like:

x x x x

x x x x

which is why the edge quality isn't so hot.

I'm pretty sure that's 8x, not 8xS. I thought 8xS was:

.x.x
x.x.
.x.x
x.x.

I don't think so. That would require 2x MSAA w/ 4x SSAA. 8xS is 4x MSAA w/ 2x SSAA. At least that's what was shown with the AA tester a while back.

 

[Dave Baumann]

5800 8xS:

http://www.beyond3d.com/previews/nvidia/gffxu/index.php?p=17

5900 8X

http://www.beyond3d.com//previews/nvidia/nv35/index.php?p=6

(Do I need really need to keep reminding people these are all in our articles???!)

 

[Ailuros]

We do read your articles Wavey.

Also, above where you talk about 6x AA and use an equation like (1024*6)*(768*6), this doesn't make sense to me as you don't increase X and Y each by a factor of 6 to get 6x AA. 6x AA is 6 samples per pixel so it should read (1024*768)*6.

That´s not what I meant, but rather the edge equivalent resolution. If I´m still wrong feel free to correct me.

Maybe I'm not understanding your notation, but 2*2 + 2*2 = 4*2. 4*4 would be 16 samples.

There has been quite some confusion between the official 8x and unofficial 8xS mode.

8x = 4xOGMS + 2xOGSS (4*2)

8xS = 2xRGMS + 4xOGSS (4*4)

(images borrowed from 3DCenter.de and this article
http://www.3dcenter.de/artikel/geforcefx_aa_modi/index_english.php

The fact that the 8xS mode is effectively on a 4x4 grid is nothing more than a mathematical construct. It has no bearing in either in the performance or the visual quality of the technique.

It´s still in between 4x and 6x sparse grid, yet not equivalent to 6xAA. Based on the grid pattern alone 2xRGMS and 4xOGMS might have identical grid values, yet the latter is slightly better than the first. Still it´s nowhere near 4xRGMS. Take a CPU limited flight sim, use an outside camera and start rotating the camera around it and/or bank the plane in both directions and it´s pretty easy to see which is in fact better or worse.

It does have enough bearing to understand where a method is lacklustering. 4xS is 2*4; big news if anyone says that it has a weakness on the x axis.

If NVIDIA has any plans to stick to those hybrid methods in the next generation and not introduce anything higher than 2xRGMS, they deserve to be shot.

 

[Arun]

If NVIDIA has any plans to stick to those hybrid methods in the next generation and not introduce anything higher than 2xRGMS, they deserve to be shot.

Don't worry, even NVIDIA, as strange as it seems based on history, doesn't like to be shot
Speculation: The "conservative" goal would be 16x OGMS, and base your 8x/4x/2x patterns on that grid ( like they do for 4x / 2x today ), with GC.

But maybe, with all the bad press they got, they're focusing on beating ATI in this category, not just being equal...


Uttar

 

[Ailuros]

How many Z/stencil units in order to get 16x sample MSAA?

That´s the first question that pops into my mind, other than that I still don´t personally like the idea of ordered grid patterns, whatever the IHV. Since I´d speculate that the transistor increase between 16xOGMS and 8x sparse should be pretty close, I´d rather opt for the latter.

16xOGMS is still 4*4.

 

[3dcgi]

How many Z/stencil units in order to get 16x sample MSAA?

That´s the first question that pops into my mind, other than that I still don´t personally like the idea of ordered grid patterns, whatever the IHV. Since I´d speculate that the transistor increase between 16xOGMS and 8x sparse should be pretty close, I´d rather opt for the latter.

16xOGMS is still 4*4.

You could make a design that does 16x MSAA with only one Z/stencil unit or you could have up to 16. More realistically you would probably have 4 or 8 and loop to support 16x.

Regarding the transistor increase, 8x sparse sampling likely requires more transistors than 16x ordered sampling because it uses an 8x8 grid. However the real bottleneck is probably memory bandwidth, in which case 8x spare sampling is better if you assume equal z compression between methods.

 

[Ailuros]

More realistically you would probably have 4 or 8 and loop to support 16x.

Interesting. Does the loop theoretically affect performance?

 

[cho]

hmm, the NVIDIA 16XAA = [4X(4X OGMS)] AA