Alternative AA methods and their comparison with traditional MSAA*

From the same blog:

I am curious to know in which case the traditional MSAA resolve does not generate smooth gradients with HDR images because as long as one performs tone mapping before the resolve I am not aware of any problem with it. Unfortunately the images that have released so far make me think that the "standard MSAA" was not properly tested.

Ignoring the temporal component of the algorithm from its description it sounds like a post-tone mapping MSAA resolve done with a wider resolve filter (Mitchell, Lanczos, etc..) that can give better edge smoothing and less aliasing due to the reconstruction filter at the price of a slightly blurrier image.

 

Why not just implement custom resove? Works on any DX10-capable hardware.

 

You can use (fe.) the 2x MSAA resolve vs. the normal resolve to predict a higher 4x MSAA resolve. It's a form of linear prediction, similar to predictive compressors. Naively this would need two depth-buffers, but you can get statistically valid information just from the lower resolve.

As he refutes post & edge, it's probably not that, but's another option that could be tried, PEAA (Predictive Edge Anti Aliasing). I'm surprised the relation of fake-AA to image compression and reconstruction hasn't been explored yet.

 

That's what I was referring to (a custom resolve where something else than a 1-pixel wide box filter is used..) as long as it is applied after tone mapping.

 

Yeah, he is probably comparing TXAA to broken MSAA implementations like what DICE uses in BF3 and BFBC2.
There are certainly examples out there in real games where MSAA + HDR is handled properly. Two Worlds II comes to mind.

I am still interested in how this will compare to regular MSAA, especially in the cases of transparency and shader aliasing. I can live with a slightly blurred image if it means no jaggies or high frequency noise like we see in the jungles of Crysis.

 

Is performance that bad that so many developers do not apply MSAA properly? Or is there a technical hurdle/tradeoff?

 

I really wish they'd just publish a technical or academic paper on what they're doing (maybe this is forth-coming), particularly if there is actually any hardware changes involved. From the description in the TechReport article though it sounds like the idea may not be entirely baked at this point, so maybe marketing just got their hands on it a bit too early. In any case the comments in the article seemed consistent with the assumption in this thread that it's some form of MSAA + custom resolve, which is good, but hopefully there's some new insight as well.

 

There's definitely going to be some tradeoffs between memory, performance, and quality (just like anything else in life). Typically you want tone mapping to be the last step in your post processing pipeline, so that you can do all of your other post processing operations in linear HDR space. There's also typically a few other steps after rendering or lighting opaques that can be cheaper if you resolve beforehand, such as transparents, particles, upscaling low-res particles, deferred fog, etc. All of those things become more expensive (and possibly consume more memory) if you need to keep your individual subsamples around so that you can postpone your resolve until the tone mapping step.

 

This is a good question. How much more of a performance hit would MSAA take in BF3 if they went ahead and did it correctly? Because honestly the way it is now I hardly see the point; it still incurs a huge performance hit and the quality is not so great.

 

SMAA 1x and S2x added into the newest Project CARS Build.