Image Quality and Framebuffer Analysis for Available/release build Games *Read the first post*

[Rolf N]

It sure does reduces the jagged edges rather effectively compared to 2*MSAA, the blur is another story.

For some games its pleasing to look at & the blur is minor enough to negate (like TRU, AC2 PS3), for some games it suits the environment rather nicely (like KZ2 & cel shaded games) ...but for some games its a destructive process as far as sharpness goes. This is why I believe if done right, QAA can be a rather good solution.

Also aliasing & shimmering hurt the IQ just as much as a blur does...& this is where it ultimately comes down to what you'd prefer, or which case is more sensitive to your eyes:

1) A more aliased, shimmering look.
2) A less aliased & sort of blurry look.

2xMSAA is option 3, has the exact same performance and memory requirements as Quincunx AA, the same degree of anti-aliasing, but no blur. You'd love it!

QAA is not just 2*MSAA+blur btw, the edge smoothing quality is identical to 4*MSAA but one of the samples here is a blurred sample. This is the cause for the blur in IQ.

It is actually 2xMSAA+blur. It pulls in AA samples from three neighbouring pixels, in addition to the 2 AA samples that actually belong there because they lie inside the pixel boundaries. It assigns the highest weight (1/2) to one of the two inside samples, and splits up the rest of the weight evenly between the secondary inside sample and the three outside samples (1/8 each).

http://developer.nvidia.com/object/geforce3_faq.html
(this is the old Geforce 3 grid; Geforce 4 and newer have it offset by a quarter pixel)

The Quincunx filter spans over multiple output domains, all weights are positive and their sum is 1. This happens to be the definition of a blur filter (or "low pass" if you're more of an audio person).

A GPU with Quincunx resolve going on cannot display a sharp edge (as seen in the box image), no matter what, not even if a sharp edge is the most precise rendition of the image details.

The base 2xMSAA is the only thing here that actually anti-aliases the image. Smoothing is not anti-aliasing, you see.

 

[nightshade]

2xMSAA is option 3, has the exact same performance and memory requirements as Quincunx AA, the same degree of anti-aliasing, but no blur. You'd love it!

Actually...option 1 is 2*MSAA. And I don't like it in many cases cause its not enough for the most part. (especially for high contrast games)

I thought it was a well known & agreed fact, since even digital foundry compares QAA to 4*MSAA when it comes to the ability to reduce jagged edges. And from what I recall QAA is also "slighty" more costly than 2*MSAA, if I read it right here few months ago.

 

[assurdum]

Quincunx AA on PS3 - two geometry sample points are used just like 2xMSAA (so the same storage cost), but it also uses 3 samples belonging to neighbouring pixels (regardless of a polygon edge) to the right and below of the original texture sampled point (see sample pattern image for clarity). The result is a blurring of the entire image, but higher perceived polygon AA. Consider a texture of higher frequency components, lots of different colour patterns. The current pixel's two geometry sample points indicate the pixel is entirely within one polygon. However, the three neighbouring sub-samples are still accounted for in the final pixel, hence the overall image blur. It's explained in the topic of rendering native resolution of game.
From what I have seen until now blur not reach any AA edge...a part costum edge detected blur of The saboteur...

 

[assurdum]

2xMSAA is option 3, has the exact same performance and memory requirements as Quincunx AA, the same degree of anti-aliasing, but no blur. You'd love it!
It is actually 2xMSAA+blur. It pulls in AA samples from three neighbouring pixels, in addition to the 2 AA samples that actually belong there because they lie inside the pixel boundaries. It assigns the highest weight (1/2) to one of the two inside samples, and splits up the rest of the weight evenly between the secondary inside sample and the three outside samples (1/8 each).

http://developer.nvidia.com/object/geforce3_faq.html
(this is the old Geforce 3 grid; Geforce 4 and newer have it offset by a quarter pixel)

The Quincunx filter spans over multiple output domains, all weights are positive and their sum is 1. This happens to be the definition of a blur filter (or "low pass" if you're more of an audio person).

A GPU with Quincunx resolve going on cannot display a sharp edge (as seen in the box image), no matter what, not even if a sharp edge is the most precise rendition of the image details.

The base 2xMSAA is the only thing here that actually anti-aliases the image. Smoothing is not anti-aliasing, you see .

No? Since...? Ah you refer to the technical means...to be honest I don't see the difference through a 'normal' AA which adding colour pixels and the 'horrible' blur edge detected when the resulted reach it's the same. The difference it's just that one is more expensive & the other less. But it seem more expensive is ever associated to better quality. However the discussion here stay again to personal taste than to technical matter.

 

[Rolf N]

No? Since...? Ah you refer to the technical means...to be honest I don't see the difference through a 'normal' AA which adding colour pixels and the 'horrible' blur edge detected when the resulted reach it's the same. The difference it's just that one is more expensive & the other less. But it seem more expensive is ever associated to better quality. However the discussion here stay again to personal taste than to technical matter.

I don't think anti-aliasing basics are a taste issue at all. Anti-aliasing in practice involves blending together multiple samples, yes. Some of the pixels in the end result appear averaged, smoothed, yes. But that does not mean the opposite is true, that everything that appears smoothed must be anti-aliased.

Aliasing occurs when fine details slip through a coarse net and are not represented in a result. What you want to do here is include more information about those fine details in the image, not dampen the information you already have. Full-screen smoothing is a way to make aliasing harder to detect, true ... by making everything harder to detect.

If a transitionary pixel at a polygon edge is (roughly) halfway between two colors, that should be because the edge lies (roughly) halfway between the pixel boundaries. The blending is proportional to the distance of the edge from the pixel grid. OTOH if an edge coincides perfectly with the pixel grid, there's no need to average the colors.

But in Quincunx two neighbouring pixels can never have more than 75% contrast between them. All pixels are averages of wider surroundings. The "lesser" sample in the Quincunx pattern contributes its color equally to four pixels, even though it is measured in a place that lies entirely outside of the boundaries of three of these four pixels. This does nothing to increase the precision of edge positions.

Equating uniform blur with anti-aliasing is equivalent to saying "sharp detail cannot exist".

And again, if uniform blur made everything better, we'd use it everywhere. But we don't. You'll notice the B3D forum reply box is delineated with a sharp dark grey line. That's not wrong. That's not aliasing. It's just sharp and precise.

 

[Laa-Yosh]

Rolf N, your explanations are highly appreciated.

 

[Neb]

How would Quincunx algorithm compare to say ATIs wide tent filter?

 

[SG79]

2xMSAA is option 3, has the exact same performance and memory requirements as Quincunx AA, the same degree of anti-aliasing, but no blur. You'd love it!

I was called crazy for saying that I'd rather have no AA at all if QAA is the only option It left such a bad taste in my mouth when I saw the two console versions of RE5.

Seriously, is there a good technical reason why QAA is still used? Killzone 2 excepted, no first party PS3 game uses it. Is it easier to implement?

 

[zed]

Seriously, is there a good technical reason why QAA is still used? Killzone 2 excepted, no first party PS3 game uses it. Is it easier to implement?

QAA is actually slower than 2xAA, thus the developers believe it looks better thus they choose it.
Personally I dont know why they dont give everyone option
0xAA,2xAA,QAA,4xAA and let the user do the AA vs framerate tradeoff (true alpha to coverage effects etc may be affected but hey more power to the user)

 

[SG79]

QAA is actually slower than 2xAA, thus the developers believe it looks better thus they choose it.

Slower

 

[nightshade]

Seriously, is there a good technical reason why QAA is still used? Killzone 2 excepted, no first party PS3 game uses it. Is it easier to implement?

Resistance 1 & Resistance 2 use QAA....while not really a first party tittle, but still really close since Insomniac games do tech share with other first party studios.

 

[N_B]

I was called crazy for saying that I'd rather have no AA at all if QAA is the only option It left such a bad taste in my mouth when I saw the two console versions of RE5.

Seriously, is there a good technical reason why QAA is still used? Killzone 2 excepted, no first party PS3 game uses it. Is it easier to implement?

A non technical reason would be just as good as a technical reason here. All you need to do is look at the Quincunx comparison here to see why:

http://upsilandre.free.fr/images/Quincunx.jpg

The perceived additional edge smoothing of quincunx is important, and it's far more likely that someone will notice extra aliasing than an extra bit of blur. I've heard more complaints about sub HD games being jaggy than blurry for example, and I support that sentiment.

 

[TheWretched]

Keep in mind, that Joe Sixpack doesn't see that kind of blur... especially, if he can't compare both versions simultaneously.

And since the blur is only slightly, and the benefit of less jaggies "huge", I guess it is a viable compromise. Seeing jaggy hell like Fallout 3 again on PS3 makes me feel uneasy, since I played most games on PC recently. And there, I can usually switch to 4x or even more AA, which makes all the difference. The blur, albeit not nice, is making the game appear better than it actually is.

 

[Neb]

Depends on game how it will affect. Some have major IQ penalties and others less and might even look better becouse it looks softer.

Tested with Wide Tent 2xAA ("6xAA").

2xAA/2xAA "6xAA" wide tent/4xAA/8xAA (Half-Life 2 CM)
http://img263.imageshack.us/img263/383/hl2hra2xaa.jpg
http://img683.imageshack.us/img683/4058/hl2hrb2xaawt.jpg
http://img706.imageshack.us/img706/4849/hl2hrb4xaa.jpg
http://img37.imageshack.us/img37/8198/hl2hrd8xaa.jpg

2xAA/2xAA "6xAA" wide tent/4xAA (FC2)
http://img97.imageshack.us/img97/8841/farcry2b2xaa.jpg
http://img706.imageshack.us/img706/707/farcry2a2xaawt.jpg
http://img405.imageshack.us/img405/7905/farcry2c4xaa.jpg

 

[Neb]

Slower

Barely any slower.

 

[nightshade]

Depends on game how it will affect. Some have major IQ penalties and others less and might even look better becouse it looks softer.

Tested with Wide Tent 2xAA ("6xAA").

2xAA/2xAA "6xAA" wide tent/4xAA/8xAA (Half-Life 2 CM)
http://img263.imageshack.us/img263/383/hl2hra2xaa.jpg
http://img683.imageshack.us/img683/4058/hl2hrb2xaawt.jpg
http://img706.imageshack.us/img706/4849/hl2hrb4xaa.jpg
http://img37.imageshack.us/img37/8198/hl2hrd8xaa.jpg

2xAA/2xAA "6xAA" wide tent/4xAA (FC2)
http://img97.imageshack.us/img97/8841/farcry2b2xaa.jpg
http://img706.imageshack.us/img706/707/farcry2a2xaawt.jpg
http://img405.imageshack.us/img405/7905/farcry2c4xaa.jpg

I for one would definitely choose the wide tent versions over 2*AA in both case. (Infact I may even choose it over 4*AA in case of FC2, cause it'll help with the vegetation).

I'll say one thing though, in screenshots it'll look like a whole world of difference, since we tend to look at all places at once, while playing the game itself its much less noticeable because you spend much less time looking away from the center of the screen where your reticule is & ofcourse you are looking at moving images while playing.

P.S.(My left eye is kinda weak btw & I still don't wear glasses, so lots of things look blurry to me in real life too...sort of like my eyes already having a built in QAA filter running /gross exaggeration)

 

[MJP]

QAA is actually slower than 2xAA, thus the developers believe it looks better thus they choose it.
Personally I dont know why they dont give everyone option
0xAA,2xAA,QAA,4xAA and let the user do the AA vs framerate tradeoff (true alpha to coverage effects etc may be affected but hey more power to the user)

You're not serious about either of those comments...are you?

 

[Arwin]

Zed, quincunx on the PS3 is pretty much free!

 

[_phil_]

Zed, quincunx on the PS3 is pretty much free!

Uh ? 3-4 ms is not exactly free...

 

[betan]

All it leads to is detail loss of every pixel. It does not and cannot blur aliased edges any differently than it blurs non-aliased edges and texture detail. It's a purely destructive process.

If Quincunx enhances quality, we should all get our eyes lasered to include a bit of myopia. But it does not, and we should not.

edit:
To demonstrate, how many aliased edges do you see in this image?

2xMSAA is option 3, has the exact same performance and memory requirements as Quincunx AA, the same degree of anti-aliasing, but no blur. You'd love it!
It is actually 2xMSAA+blur. It pulls in AA samples from three neighbouring pixels, in addition to the 2 AA samples that actually belong there because they lie inside the pixel boundaries. It assigns the highest weight (1/2) to one of the two inside samples, and splits up the rest of the weight evenly between the secondary inside sample and the three outside samples (1/8 each).

http://developer.nvidia.com/object/geforce3_faq.html
(this is the old Geforce 3 grid; Geforce 4 and newer have it offset by a quarter pixel)

The Quincunx filter spans over multiple output domains, all weights are positive and their sum is 1. This happens to be the definition of a blur filter (or "low pass" if you're more of an audio person).

A GPU with Quincunx resolve going on cannot display a sharp edge (as seen in the box image), no matter what, not even if a sharp edge is the most precise rendition of the image details.

The base 2xMSAA is the only thing here that actually anti-aliases the image. Smoothing is not anti-aliasing, you see.

Since you basically keep saying the same thing (that it's only a destructive process, correct me if I'm wrong) I will simply address all here if you don't mind.

First let's recall AA we are talking about here is loosely the act of doing something to get rid of aliasing. MSAA for example is trying to address the issue for edge aliasing since mathematical edges practically introduce infinite frequency that cannot be displayed on any finite pixel, thus introduce aliasing.

The real AA would be getting rid of those extra frequencies on analog/continues space before sampling at the resolution of the display. Getting rid of extra frequencies generally means applying a low-pass filter. Concisely AA is applying low-pass filter before sampling (or something equivalent).

The reason blind low-pass after sampling doesn't work is basically the low-pass band your final signal can carry is already "corrupted" by aliasing.

Because how rasterisation works real AA is not practical. The second best thing is emulating the continues space with a higher resolution discrete space (aliasing of which is less then aliasing of lower resolution sampling), and do the downsampling from there. This is called SSAA also how bullshots are created. Downscaling is applying low-pass filtering and undersampling in this order (which is also what proper interpolation is equivalent to).

Since this is also expensive, MSAA tries to emulate SSAA for edges only. So What MSAA does when resolving pixels is low-pass filter (average 2 or more samples) and then only pick one (subsample).
Of course that low pass filter is a poor one, and the aliasing for higher resolution samples is generally significant but not a bad effort overall.

The moral of the story is MSAA already "blurs" before subsampling. QAA just uses a different low-pass filter, something arguably stronger (but not necessarily better). And it's not equivalent to MSAA + blur for edge pixels (since those extra samples are potentially from different surfaces).
Hence, you cannot come up with a blur filter to apply after 2xMSAA that gives the same edge quality as QAA.