Anti-Aliasing types for Next Gen Consoles

[Silent_Buddha]

God, I hope the next gen consoles will be able to do better than MLAA/SMAA/TXAA/etc. They are good for the performance penalty, but ultimately do very little, IMO. Then again I guess I'm just used to far better edge AA combined with sharp textures than any post process AA can possibly hope to accomplish.

It's one of those things that always makes console gaming difficult to justify for myself.

I find it notable that even the creator of FXAA isn't planning on using FXAA in his console engine. He's planning on using 2x MSAA (much better).

Regards,
SB

 

[Inuhanyou]

That's the point, they are good for the performance penalty and you need to optimize for performance when working with consoles. There are other more efficient anti aliasing techniques besides MSAA which is pretty outdated at this point and in my opinion, not worth the performance hit.

For Agni, its as simple as toning down that excessive 8XMSAA technique down to maybe 2XMSAA with a dynamic FXAA injector. Or combining MLAA/SMAA/TXAA as needed.

When console gaming, your hooked up to the TV and sitting a reasonable distance away so as to not see the super minute details of the image. Its different from sitting up close to a monitor.

 

[jlippo]

I find it notable that even the creator of FXAA isn't planning on using FXAA in his console engine. He's planning on using 2x MSAA (much better).

He is going for a rotated grid rendering to ~2x sized buffer and doing custom resolve.
This has nothing to do with MSAA.

 

[Billy Idol]

God, I hope the next gen consoles will be able to do better than MLAA/SMAA/TXAA/etc. They are good for the performance penalty, but ultimately do very little, IMO. Then again I guess I'm just used to far better edge AA combined with sharp textures than any post process AA can possibly hope to accomplish.

It's one of those things that always makes console gaming difficult to justify for myself.

I find it notable that even the creator of FXAA isn't planning on using FXAA in his console engine. He's planning on using 2x MSAA (much better).

Regards,
SB

Er...TXAA is not compareable to MLAA and SMAA. It is only available for 600 NVidia GPUs is based on MSAA and work very well. If next gen features TXAA quality of jaggie reduction...I would be happy.

 

[Silent_Buddha]

He is going for a rotated grid rendering to ~2x sized buffer and doing custom resolve.
This has nothing to do with MSAA.

Ah, that's right, I'd forgotten that he was going for 2X RGSSAA, even better. That does seem to lend some credence to the consoles potentially being CPU limited. If that is the case, there will be GPU performance to spare for SSAA.

In which case, hell, I wouldn't mind if all Orbis and Durango next gen games came out at 540p with RGSSAA. Screw 720p and 1080p in that case.

Regards,
SB

 

[Shifty Geezer]

Ah, that's right, I'd forgotten that he was going for 2X RGSSAA, even better.

2x SSAA will give far worse jaggie reduction than higher levels of MSAA or the post FX AAs.

 

[Arwin]

There were definitely some scenes in Uncharted where post-AA didn't work well, such as the forest in France. That was horrible, and I get the feeling that would have looked better with no AA at all. In many other areas though it looked fantastic. It's still early days, but having native 1080p decreases the burden on AA solutions significantly.

 

[Shifty Geezer]

Post only is no good because it can't resolve sub-pixel geometry. SSAA is just a waste of effort though. Combine post AA with subsample resolution in a clever AA technique, also factoring in temporal variations, and very high quality should be obtainable with relatively little effort.

That's where a truly programmable architecture would be beneficial. Things like stochastic sampling of geometry edges at high framerates could all but eliminate perceptible jaggies at little effort, but the nature of graphics rendering doesn't really support that yet AFAIK. We're mostly limited to what the hardware allows, save post techniques, so it'll be MSAA+whatever algos to smooth things out. When post-AA works, it is extremely effective. Just need to solve the lack of subpixel resolution issues.

 

[Silent_Buddha]

2x SSAA will give far worse jaggie reduction than higher levels of MSAA or the post FX AAs.

Yes, for any given level of RGSSAA (OGSSAA is much worse) or MSAA, the quality of the edges should be similar. RGSSAA has multiple benefits however when it comes to cleaning up other rendering artifacts (specular, transparent texture edges, etc. for example) that MSAA doesn't touch.

I have yet to see an implementation of post FX AA (shader based AA) give a reasonably good image. And that includes what I've seen on PS3. I would still take 2x MSAA w/transparency AA over any of the post AA processes.

Where post AA processes come in somewhat handy sometimes is when combined with MSAA. When used in that way the quality can come sorta/kinda close to RGSSAA, albeit with a much "softer" image.

Really, from everything I've seen of software AA solutions, I'd be just as happy without them. That doesn't mean I don't think there's potential there. Just that I have yet to see an implementation that is actually good.

Regards,
SB

 

[Jedi2016]

All these acronyms are starting to get confusing. SSSGGAA is so much better than RGSSGSSAA and so forth. There's probably one guy that not only knows what they all stand for, but exactly how each one of them works.

Honestly, the only two that I know completely how they work is SSAA and MSAA. I have a rough idea of how FXAA and SMAA work, but I don't know what the acronyms stand for. And that's about it.

 

[Grall]

SSSGGAA is so much better than RGSSGSSAA and so forth.

You nub, everybody knows TSSXGSSFAAAA is much better thgan SSSGGAA!

I have a rough idea of how FXAA and SMAA work, but I don't know what the acronyms stand for. And that's about it.

Same thing here. It seems the difference in how these post-process AA techniques work is fairly minute, so why even bother with the separate labels? It feels like it's just free advertising for nvidia to keep repeating their proprietary four-letter acronym.

 

[Boxercide]

All these acronyms are starting to get confusing. SSSGGAA is so much better than RGSSGSSAA and so forth. There's probably one guy that not only knows what they all stand for, but exactly how each one of them works.

Honestly, the only two that I know completely how they work is SSAA and MSAA. I have a rough idea of how FXAA and SMAA work, but I don't know what the acronyms stand for. And that's about it.

SSAA-Super Sampling. Rendering in a much higher resolution and then downscaling to the one you're using.

RGAA-a type of SSAA where you rotate your samples so as not to line up perfectly on the axes for better coverage.

FXAA-post process intelligent blur; means fast approximate AA

SMAA-another post process, aims to prevent pixel crawl by taking a look at sub pixel artifacts; acronym means sub-pixel morphological AA

MSAA-bog standard multi-sampling AA

TXAA-honestly I'm not really up to par on exactly what this is. I thought is combines temporal AA with MSAA and maybe FXAA but I'm pretty sure that's me talking out of my bum

temporal AA using a form a last frame blending and the eye's natural reaction to motion to try and smooth jaggies, often with a half pixel offset

QAA-is 2xMSAA but also some neighboring color samples. Gives the impression of 4xMSAA but blurs a LOT of fine detail

 

[Gitaroo]

QAA should be gone next gen, no more nVidia GPU. I think MSAA + SMAA combo might be best to stay? 2XMSAA + SMAA should be very good combo no? Assuming all games aims at higher than 1080p. And eventually down to just SMAA later in the console life cycle due to performance.

 

[Shifty Geezer]

I hope AA goes smarter than even that. Independent AA line drawing (used in at least one racer this gen), proper filtering in the shader code, and subsample accurate image reconstruction on geometry.

 

[Arwin]

I hope AA goes smarter than even that. Independent AA line drawing (used in at least one racer this gen), proper filtering in the shader code, and subsample accurate image reconstruction on geometry.

Or perhaps we can get to a point where every visible pixel will have all its position, shading and so on information available at any time, including perhaps a certain margin of culled pixels?

 

[Shifty Geezer]

Or perhaps we can get to a point where every visible pixel will have all its position, shading and so on information available at any time...

That's effectively a massive database of values to be processed and composited at will; an evolution of the concept of the deferred renderer.

 

[Silent_Buddha]

I hope AA goes smarter than even that. Independent AA line drawing (used in at least one racer this gen), proper filtering in the shader code, and subsample accurate image reconstruction on geometry.

Yes, an intelligent combination of MSAA + intelligent compute based AA might offer a decent tradeoff in performance and quality versus brute force RGSSAA. This is assuming that developers will properly implement MSAA with deferred renderers (which their GPUs fully support).

The main problem that I have with any current compute based AA approach is that it tends to soften the image overly much without adequately and consistently addressing polygon edges.

All these acronyms are starting to get confusing. SSSGGAA is so much better than RGSSGSSAA and so forth. There's probably one guy that not only knows what they all stand for, but exactly how each one of them works.

Honestly, the only two that I know completely how they work is SSAA and MSAA. I have a rough idea of how FXAA and SMAA work, but I don't know what the acronyms stand for. And that's about it.

damienw addressed most of that above. But just wanted to clarify two of the ones I mentioned.

SSAA as mentioned is just short for Super Sampled Antialiasing. It's easy to think of as rendering at a higher resolution then downsampling to a lower resolution although it doesn't have to be done like that. So 2x means render at twice the resolution and then reduce.

OGSSAA is Ordered Grid SSAA. Or dumb SSAA. It's decent but retains some of the rendering artifacts that you'd have with just rendering at native resolution. This is basically what people think of when you just say SSAA.

RGSSAA (RGAA that damienw used) is Rotated Grid SSAA. This is where you take X number of subpixel around the target pixel, rotate them then resolve. This produces a much nicer image and resolves many rendering artifacts that remain when using regular OGSSAA. Before Anistropic Filtering was good and performant, RGSSAA did a lot to clean up texture shimmering and other texture related artifacts as well. In many ways a lot of people consider RGSSAA as the holy grail of antialiasing. Getting that sort of quality without the associated performance hit would be fantastic.

Regards,
SB

 

[Shifty Geezer]

Stochastic sampling would be even better, but good luck implementing that.

 

[Silent_Buddha]

Stochastic sampling would be even better, but good luck implementing that.

Well yes, methods for random or pseudo random sample patterns would be fantastic. But require frame rates higher than 60 FPS which isn't possible due to the majority of displays configured to deal with, at most, 60 FPS video streams.

AMD had experimented with something along those lines briefly but it didn't catch on due to the high FPS required to have a good experience. And even then it was just alternating two different sample patterns.

It was also limited to polygon edges as it was just an MSAA variant. But the quality was superb for very low performance cost (roughly doubling visual MSAA at half the cost) as long as you could maintain the required FPS. Considering that would mean having a minimum FPS of 60 FPS, it was always pretty hit or miss.

Regards,
SB

 

[XpiderMX]

From vgleaks:

Antialiasing

The Durango GPU supports 2x, 4x, and 8x MSAA levels. It also implements a modified type of MSAA known as compressed AA. Compressed AA decouples two notions of sample:

Coverage sample–One of several screenspace positions generated by rasterization of one pixel
Surface sample– One of several entries representing a single pixel in a color or depth/stencil surface



Traditionally, coverage samples and surface samples match up one to one. In standard 4xMSAA, for example, a triangle may cover from zero to four samples of any given pixel, and a depth and a color are recorded for each covered sample.

Under compressed AA, there can be more coverage samples than surface samples. In other words, a triangle may still cover several screenspace locations per pixel, but the GPU does not allocate enough render target space to store a unique depth and color for each location. Hardware logic determines how to combine data from multiple coverage samples. In areas of the screen with extensive subpixel detail, this data reduction process is lossy, but the errors are generally unobjectionable. Compressed AA combines most of the quality benefits of high MSAA levels with the relaxed space requirements of lower MSAA levels.