Too much AA?

It would be one hell of a case where there would be any significant performance cost on a iPad4 or even iPad3 for 4x Multisampling even in 960*640.

The iPad4 loses in GL2.5 (which is quite a demanding case compared to today's mobile games) around 12% of performance in 2048*1536: http://www.glbenchmark.com/phonedetails.jsp?benchmark=glpro25&D=Apple+iPad+4. while the iPad3 barely 7% (probably because there's a bigger bottleneck elsewhere): http://www.glbenchmark.com/phonedetails.jsp?benchmark=glpro25&D=Apple+iPad+3

While a large crop of games on those two aren't actually rendered in their native resolution, are there any games that scale as low as 960*640?

What you can do however on an alternative SFF mobile device like the Galaxy S3 if you have fillrate to spare to enable 16xAA, which is a combination of 4xMSAA + 4xSSAA.

 

Samsung SyncMaster 1100DF. Thank God my desk is equally huge. If you don't have enough space on the desk it's a pretty bad idea; when I bought it I had to unbox it because it didn't fit even with backseats dropped into my car

 

He he, i had a 22" Mitsubishi CRT ( Professional series, i had take it from my work ), with the 5 BNC cables ( colors, sync etc ), still regret he have die one day .. Huge is an euphemism.

 

I have a huge beast too, running in 1280x960. My OS prevents me from using higher than 85Hz else I think I'd do 1400x1050 100Hz. The way it's set, it's good enough for my eyes.
iiyama 22", a vision master pro 510.

 

Blur is a problem with your reconstruction filter, not the number of samples contributing to the reconstruction. I've never heard someone dog 192kHz audio for being 'too many samples', even though you can get a better 44/48kHz output after processing; I've never heard anyone dog floating point math for being 'too accurate' for lighting calculations, either.

 

I haven't seen anyone over use the SMAA in SweetFX - mostly because it's very hard to over use SMAA.
I have seen plenty of my users who abuse the other effects though. Some people have the mistaken notion that more = better and crank up sharpening and vibrance to ridiculously high values.

Most other types of AA are also either very hard or impossible to abuse.
At least when it comes to image quality - it's easy to kill the performance by applying lots of AA.

The exceptions are probably AMD MLAA, FXAA and TXAA which can blur the image.
Some careful tweaking of their settings and a little bit of post-sharpening can however reduce that to the point where it's very acceptable and no longer annoying.

Among the screen-space AA solutions I still prefer SMAA because it doesn't have this problem at all and it's quality / performance cost ratio is awesome - especially with SMAA T2x

 

Anyone played Inside by Playdead?

Best game AA implementation I've ever seen.
Zero pixel crawling. Superbly stable image. Slightly softened, but with the right measure.

http://www.gdcvault.com/play/1022970/Temporal-Reprojection-Anti-Aliasing-in

 

Inside's temporal AA implementation is pretty basic. Not as fancy as the newest AAA implementations, but it works well with their content. They also use TAA very well with stochastic techniques (shadows, reflections, volumetrics). End result is fantastic.

 

While I agree that it really looks very neat, I also agree that content is their key to success: They have an environment, where they are able to very carefully manage all that critical corners (and do so very successfully): No high-frequency content and a rather shallow depth of field focussing on the protagonist. Great job nevertheless.

 

INSIDE is pure perfection in terms of visuals... it's frightening how meticulous the work was...but they had the luxury (and funds) to take 6 years to craft this masterpiece.

 

Didn't mean to imply otherwise.

 

Has anyone claimed what pixel density is needed to remove aliasing ?

 

It depends entirely on the content. You can perfectly render a square with no AA at all, and a chain link fence at a distance might require dozens of samples to render correctly (to the human eye, technically it will still be imperfect). To attain a truly aliasing free image on all possible content, the number of samples you need approaches infinity.

It reminds me of "what framerate is good enough", though I think that question is easier to answer with some simple tests.

 

Plus it depends on eyesight...

My prime example always has been the sub-pixel-level plane in the far far distance in a flight sim. At some (varying) point it will pop in and out of the rendered image. The display's resolution only varies the distance to the viewport at which this occurs. And while people argue that the eyesight has finite resolution and a static chain-link fence may look 'perfect', movement and popping in/out always triggers our evolutional reflexes.

 

Yes, motion is key. A modest amount of AA can mostly eliminate any visible aliasing in a still image but crawling jaggies and flickering fences stick out like a sore boner.

 

They also released source code.
https://github.com/playdeadgames/temporal

Not that I know.

You would need display which can properly show an anisotropic surface.
CD without AA is a nightmare.

 

From wiki
"In the field of computer graphics, an anisotropic surface changes in appearance as it rotates about its geometric normal, as is the case with velvet."
a any tv/monitor can display velvet

 

Agreed that if flat surface has proper anisotropic shader which handles AA, it is easily displayable.
Just modelling surface and sending it to screen, not so much.

 

INSIDE looks very good. But this is a tech forum, so I had to point out the technical details.

Their TAA implementation has the following minor shortcomings:
- Bilinear fetch instead of bicubic from the history buffer. Catmull-Rom bicubic is used by best implementations. Bilinear causes slightly more blurred result.
- RGB color space neighborhood clamp instead of YCrCb or YCoCg. RGB clipping causes slightly more color ghosting in some failure cases.

Both of these quality improvements cost extra cycles, but cost can be mitigated by processing TAA in tiles by compute shader (load to LDS -> transform color space once per pixel = 9x savings to transform ALU). Similarly LDS neighborhood (and depth) loads result in 9x reduction in those memory fetches -> bicubic fetches are much less of a bottleneck.

This is the most advanced TAA implementation currently available:
http://advances.realtimerendering.com/s2016/Filmic SMAA v7.pptx

I recommend watching the video from page 93. It clearly shows how much better bicubic is versus bilinear (in history sampling). Jorge's version also has lots of improvements to the sampling and resolve (anti-ghosting & anti-flickering) steps. The simpler algorithm used in INSIDE causes errors in moving object edges. Playdead had to blur these errors with extra motion blur. It works well in their case, but if you have content with high contrast textures, all that extra blur starts causing various problems.