When is 4X FSAA Not 4X FSAA?

[demalion]

The problem is not "having the option", rather, it's missrepresenting it as 4X FSAA....

It's not the first time it happens.
Remember '4X, performance' FSAA?

Which performance AA option did this before? Besides the Xabre, I was not aware of a similar practice.

 

[Dave Baumann]

R200's 'Performance' SuperSampling appeard to be using upsampled buffers that were not the multiples of the target resolution equal to the FSAA level selected (i.e. if you were at 1024x768 and asked for performance FSAA it would have a super buffer size of 1600x1200, or something, rather than 2045x1536)

 

[Hyp-X]

Which performance AA option did this before? Besides the Xabre, I was not aware of a similar practice.

With the R8500 ATI introduced two kinds of AA: performance and quality.
The quality modes has as many subsamples as the number indicated.
The performance modes had less.

For example the 2x performance mode was 1 x 1.5, 3x performance mode was 1.5 x 1.5.
(as can be seen by the pictures submitted by you in this thread.)

Not as bad as the Xabre though which does 2x AA in both 3x and 4x modes.

 

[demalion]

Which performance AA option did this before? Besides the Xabre, I was not aware of a similar practice.

With the R8500 ATI introduced two kinds of AA: performance and quality.
The quality modes has as many subsamples as the number indicated.
The performance modes had less.

For example the 2x performance mode was 1 x 1.5, 3x performance mode was 1.5 x 1.5.
(as can be seen by the pictures submitted by you in this thread.)

Look at the questions I had, as well.

For context: I consider Quincunx to be a flawed 4x mode, even though the work it is doing is that of 2x sampling. <- correct me if you disagree.

What this seems to be describing is a method of doing "4x" sampling that depends on doing 2x sampling differently on each successive frame. That is not a flawed 4x mode, that is mislabelled 2x mode that depends on counting the output from 2 frames instead of 1.

This also does not appear, to me, to be comparable to Performance mode, nor am I aware of Performance mode having the same negative impact on textures as Quincunx.

Again, note the questions I had in that thread, and if you can answer them as part of clarifying this, I'd appreciate it.

Is it just that you equate all of these equally?

Not as bad as the Xabre though which does 2x AA in both 3x and 4x modes.

I have a problem with equating things with such a broad hand. If I shouldn't, please point out why to me.

 

[Hyp-X]

After all this discussion I realised I failed to write my opinion about this issue.

Actually there was little point using 4x AA on Geforce cards, the IQ improvements over 2x AA was quite small (because it uses an ordered 2x pattern) while the performance hit is quite large.

Now they introduced a new mode in which you have no performance hit in 4x AA over 2x, but the overall quality is worse. (Because it uses an ordered 2x pattern).

So they traded a useless mode - with a useless mode.
Cool.
Way to go.

 

[Hyp-X]

Look at the questions I had, as well.

I did. Maybe I misunderstood them.

For context: I consider Quincunx to be a flawed 4x mode, even though the work it is doing is that of 2x sampling. <- correct me if you disagree.

I never thought you were referring to Quincunx.
Quincunx was never advertised as a 4x mode.
But you consider it a flawed 4x mode.

So it is nVidia's fault of not using 4 samples in Quincunx mode because you conider it a 4x mode even so it was never advertised as such?

<- correct me if you disagree.

Disagree with what? That you consider it a flawed 4x mode?
It depends on the definition of 'flawed'.
But since I don't consider it a 4x mode I doubt I agree with you.

What this seems to be describing is a method of doing "4x" sampling that depends on doing 2x sampling differently on each successive frame. That is not a flawed 4x mode, that is mislabelled 2x mode that depends on counting the output from 2 frames instead of 1.

Correct.
Assuming you are not talking talking about Quincunx but the FX5200 hackery in this paragraph.

This also does not appear, to me, to be comparable to Performance mode, nor am I aware of Performance mode having the same negative impact on textures as Quincunx.

What is <This> in this sentence?

a.) Quincunx
I never said Quincunx is comparable with the Performance mode since I did not talk about Quincunx.

b.) The FX5200 hackery
You say that the FX5200 hackery is not comparable to Performance mode because Quincunx blurs textures?
In this case that sentence doesn't makes sense.

Again, note the questions I had in that thread, and if you can answer them as part of clarifying this, I'd appreciate it.

I'll look into it.

Is it just that you equate all of these equally?

I do not equate them. Where did I said this?
I said it wasn't the first time that an AA mode is misrepresented as '4x'.
Please don't put words in my mouth.

Not as bad as the Xabre though which does 2x AA in both 3x and 4x modes.

I have a problem with equating things with such a broad hand. If I shouldn't, please point out why to me.

Xabre's 3x AA is a 2 sample AA mode.
R200's performance 3x AA is a 1.5 x 1.5 sample AA mode.
The R200 is obviously better as it has 2.25 samples per pixel.
These is outright comparable at the quality level.
(Especially as both are SS.)

The FX5200 4x mode is not comparable at the quality level - so I'm glad I didn't do that...

 

[Hellbinder]

Xabre's 3x AA is a 2 sample AA mode.
R200's performance 3x AA is a 1.5 x 1.5 sample AA mode.
The R200 is obviously better as it has 2.25 samples per pixel.
These is outright comparable at the quality level.

yes.. but in this case the 5200 sample pixels change resulting in serious line crawl. Not the same thing at all to what you are comparing to above.

 

[Hyp-X]

I assume you mean this question:

Performance mode 2x on the 8500 looks like it is has 3 vertical samples...why is that?

Altough I don't have the code of that application, I assume your explanation on how it works in geometry mode fits - it uses subpixel sized triangles (or better quads).
The size of those quads determines how large the dots will be on the output.

At each pixel the quad is placed at a different subpixel location to "probe" that location.

The problem with non-whole subsampling is that the samples will end up at different locations at different pixels.

For example with 1.5x AA there is one pixel with 1 sample and one pixel with 2 samples.
This pattern is repeating every other pixel.

So in theory you should see an 'interlaced' screen where the even lines are indicating the subsample positions of the even line pixels, and the odd lines are indicating the subsample positions of the odd line pixels.

You see this occuring exactly at the two bottom positions.

The top position is different because of how the downsampling works.
Take the 3 samples (A, B, C) in 2 pixels output to the screen. (D, E).
Subsampling does something similar to:
D = (2*A + B)/3
E = (B + 2*C)/3
This means there is one subsample that is contributing to 2 output pixels with half the weight.
This is why at the top position you see a half strength square instead of an 'interlaced' square.

 

[Hyp-X]

]

Xabre's 3x AA is a 2 sample AA mode.
R200's performance 3x AA is a 1.5 x 1.5 sample AA mode.
The R200 is obviously better as it has 2.25 samples per pixel.
These is outright comparable at the quality level.

yes.. but in this case the 5200 sample pixels change resulting in serious line crawl. Not the same thing at all to what you are comparing to above.

I agree.

 

[Hyp-X]

Actually they could have avoided most of the line crawl if they'd use alternating diagonal samples.
I think that the solutions in quality order (best to worst):

1.
X-  -X   
-X  X-

2.
X-  -X
X-  -X

3.
XX  --
--  XX

Now, which one they chose?

PS.: 2. is better than 3. on analog displays only.

 

[demalion]

Look at the questions I had, as well.

I did. Maybe I misunderstood them.

Well, you seem to be equating this "hackery", as you call, it with Performance mode, and to clarify that to me you'd be addressing some of the question I had.

For context: I consider Quincunx to be a flawed 4x mode, even though the work it is doing is that of 2x sampling. <- correct me if you disagree.

I never thought you were referring to Quincunx.

I wasn't, I brought it up for context. <- that because I said "For context:".

Quincunx was never advertised as a 4x mode.
But you consider it a flawed 4x mode.

Yep. I do appear to be mistaken, though, as I thought nVidia had pushed Quincunx as equivalent to 4x AA.

So it is nVidia's fault of not using 4 samples in Quincunx mode because you conider it a 4x mode even so it was never advertised as such?

Nope, I bring it up for something else I don't consider equivalent to this apparent "hackery". This relates to subsequent questions (I'll put (*) for indication).

<- correct me if you disagree.

Disagree with what? That you consider it a flawed 4x mode?
It depends on the definition of 'flawed'.
But since I don't consider it a 4x mode I doubt I agree with you.

Well, it appears that you don't disagree that it is doing the work of 2x sampling, so you've answered the question.
I consider it flawed 4x because it offers improvement at edges that are comparable to 4x AA, and is flawed elsewhere.
...

This also does not appear, to me, to be comparable to Performance mode, nor am I aware of Performance mode having the same negative impact on textures as Quincunx.

What is <This> in this sentence?

a.) Quincunx
I never said Quincunx is comparable with the Performance mode since I did not talk about Quincunx.

(*) Didn't say you were, I brought it up, not you.

b.) The FX5200 hackery
You say that the FX5200 hackery is not comparable to Performance mode because Quincunx blurs textures?

(*) No, I'm saying neither Quincunx nor Performance are comparable to this "hackery". The relation to Quincunx (of the Performance mode, not the hackery) is in the context of my considering Quincunx a flawed 4x mode, and Performance mode as not displaying the flaws to my knowledge. Look at the illustrated sample positions that you referred to in that thread...I don't see that being referred to as Performance 2x as a problem. I do see a problem as referring to this hackery as anything 4x, however.
This is the same argument I maintain about the FX "performance features", in that they are not options, but mislabelled defaults. This hackery takes the problem one step further.
...

Is it just that you equate all of these equally?

I do not equate them. Where did I said this?

To my understanding, you did equate them, as that is how I took your mention in its context. It is no longer how I take it.

I said it wasn't the first time that an AA mode is misrepresented as '4x'.
Please don't put words in my mouth.

Hmm...didn't mean to, and your clarification does seem to indicate that you do not. I think it is a matter of your clarification coming after your initial reply.

Not as bad as the Xabre though which does 2x AA in both 3x and 4x modes.

I have a problem with equating things with such a broad hand. If I shouldn't, please point out why to me.

Xabre's 3x AA is a 2 sample AA mode.
R200's performance 3x AA is a 1.5 x 1.5 sample AA mode.
The R200 is obviously better as it has 2.25 samples per pixel.
These is outright comparable at the quality level.
(Especially as both are SS.)

The FX5200 4x mode is not comparable at the quality level - so I'm glad I didn't do that...

Heh, it is a matter of my misunderstanding your initial reply then?

 

[demalion]

I assume you mean this question:

Performance mode 2x on the 8500 looks like it is has 3 vertical samples...why is that?

Well, more like asking if my explanation of a theoretical (to me) answer was correct. But that's ok, because that's what you answered. EDIT: And how rude of me: Thanks.

My question relates to output, in that the output from this is comparable to 2x. What would be misrepresentative is saying it is 2x jittered (and I do blame ATi for calling their modes jittered when it doesn't do this in, for example, fog). What it is called is 2x Performance.

That's why I consider Quincunx flawed 4x (with a pretty big flaw). For example, I wouldn't have the same problem with Quincunx being called 4x Performance. Well, perhaps a little because of the effect on textures, but I wouldn't have any if it was called Aggressive (in a context where the effect on texture quality is established). That's why I complain about Aggressive when it is compared blindly and without context, but don't blame nVidia for offering it (just for some of the labelling associated with it at various times).

This 2x hackery is just a flat out lie (again, IMO). Sample positions can be switched between frames by other companies.

My problem with bringing up another example besides the Xabre is that the order of the problem is significantly worse for this hackery to my way of thinking.

 

[Hyp-X]

I consider it flawed 4x because it offers improvement at edges that are comparable to 4x AA, and is flawed elsewhere.

So you say it's flawed because it blurs textures.
I can agree with that.
Actually 2x performance SV blurs textures too.
(See my description on how it works.)

On the other hand when you say 'edges that are comparable to 4x AA' you use 4x AA as description of a quality level.
In the right context (which is in this case GeForce cards) it can mean a quality level (the edge quality of the 4x AA of the GeForce cards) in which case I agree that it's comparable. (because the GeForce 4x AA is ordered).

Yet the 4x AA of the GeForce is much worse than the Voodoo5 or the R300 4x AA (even without the gamma correction), so I consider saying that Quincunx is comparable to 4x AA is somewhat misleading.
It's even worse than a sparse sampled 3x AA would be.

...

This is the same argument I maintain about the FX "performance features", in that they are not options, but mislabelled defaults.

That's a good point.
They should make the standard quality mode the default.
(Unfortunately I don't see that happening.)

 

[OpenGL guy]

So what would such a driver do? Blend two frame buffer dumps?

What if you only drew a single frame? Would it do it in 4x or 2x? That would be an interesting experiment.

 

[Colourless]

Heh, funny you should say that. My FSAA Viewer actually only does render a single frame when windowed (unless a WM_PAINT message is received from Windows). Fullscreen though does render multiple frames.

Too see if anything strange is going on (when windowed), you could just keep tapping the space bar. That will invert the colours (and forces the frame to be re-rendered) so you'd be able to see if some frame blending is occuring since eveything should be grey.

Oh yeah, also OpenGL Guy, this isn't exactly related, but I noticed something strange going on with my 9700Pro when I used multisample masking. It seems as if Gamma corrected FSAA is being disabled....

 

[demalion]

I consider it flawed 4x because it offers improvement at edges that are comparable to 4x AA, and is flawed elsewhere.

So you say it's flawed because it blurs textures.
I can agree with that.
Actually 2x performance SV blurs textures too.
(See my description on how it works.)

Compared to 2x SS it blurs textures, not to no AA at all. There is more texture color information being blended for the pixel output. "Super sampled Quincunx" without the trademarked sampling pattern.

...

For your quality discussion, I consider that adequately covered by, say, a reviewer, in recognizing the AA is inferior (in this case, significantly).

And, again, the Quincunx wasn't brought up because I think it would be especially well-labelled as 4x, but because it is markedly better, IMO, than the problem in question.

 

[Hyp-X]

Compared to 2x SS it blurs textures, not to no AA at all. There is more texture color information being blended for the pixel output.

Oops, I forgot about SS.
Yes in most cases it improves textures.

It only has a possible problem with 2D stuff.
(But in most applications that's a problem in any SS modes anyway.)

 

[OpenGL guy]

Oh yeah, also OpenGL Guy, this isn't exactly related, but I noticed something strange going on with my 9700Pro when I used multisample masking. It seems as if Gamma corrected FSAA is being disabled....

Not strange at all... Only way to pass WHQL

 

[Xmas]

Compared to 2x SS it blurs textures, not to no AA at all. There is more texture color information being blended for the pixel output. "Super sampled Quincunx" without the trademarked sampling pattern.

AFAIK Supersampling on R8500 doesn't affect texture LOD, so textures look worse with AA than without.

 

[Xmas]

Oh yeah, also OpenGL Guy, this isn't exactly related, but I noticed something strange going on with my 9700Pro when I used multisample masking. It seems as if Gamma corrected FSAA is being disabled....

Not strange at all... Only way to pass WHQL

Now that is strange
Why would MS require the downsampling to be linear?

btw, I asked this before: Is that gamma-"corrected" downsampling adjustable or does it operate at a fixed gamma value of 2.2 (or similar)?