Alternative AA methods and their comparison with traditional MSAA*

Maybe someone could run Borderlands PC version at 2560x1440, take a screenshot and bicubic resize to 720p show us how it would be with 4xSSAA, or would that be considered 2x supersampling?

 

Could the greater use of this be hindered due to the parallelism of the problem? They say 20ms of SPU time spread over 5 SPUs, Is it highly doubtful that one SPU could do this in 20ms? If so, it could only be used in games that considered it from the get go and left the 4ms of SPU time over 5 at the right time available, so scheduling has to come into it does it not?

 

What about the sub-pixel stuff I would like to see examples of that.

 

When you have 16-33ms to render a frame, you can't spend 20ms for AA, so they have to use multiple SPU's. Considering the SPU time is already reserved for post-processing, reserving them a bit longer for AA shouldn't be that much of a problem.

 

Wow!!! AA is the best I've ever seen! It would look unbelievable in a realistic game such as GT5!

Look at the new blood texture on his chest!!!

 

cant wait to see this in the next uncharted or something, hopefully Infamous 2 also. Anyone know if this works on alpha texture like hairs, grass, fences etc?

 

Hopefully we'll see this in several future PS3 titles.

 

Depends on what buffer you use for edge detection, but the normal way works on everything as it finds and blends the edges on color image.

Problem is when edge is thinner than pixel and there is not enough information to recreate the edge.
In this case it looks pretty close to normal image without AA, including all the flickering.
There is also problems in places where edges cross over each other as the edge information is obscured.

 

Of course you can't do it on one, that's obvious, but I think that it would take one more than 20ms in any case. And yes, but how many games/engines are/going to reserve time on 5 SPUs until the post-processing stage? I'm trying to say that it can't simply be plugged into any other game - GT5 for example. What happens if one of the SPUs is taking longer than it should and you have 4 going at it, job done, and then waiting for the last one to finish the frame because it got stalled by some previous process? I think that the scheduling on this is tricky and not so easily adaptable.

 

Not that i'd know the intricacies of the technology, but logically i would assume that it would be tricky only for games that make good use of the SPUs.
Any game that's released without tapping into all the SPUs as heavily as GOW3 or U2 (and i'm pretty sure there's a lot of them still to this day) would be fine, no?
If one or more SPUs are just sitting there doing nothing, then they should just as easily be running this routine at the right time.

 

I blame the developer for turning me off watching the new game media and remain ignorant - the beheading scene was too much for me last E3 (although I still think there's room for debate on the issue - particularly since some of the cutscenes are rendered in engine but offline according to a recent interview).

GT games have been relatively low on CPU usage in PS2 era. Unless something dramatic changed since GT5P, I'd think SPUs in their case can largely be dedicated to graphic tasks. Whether they already do too many of those to fit AA I wouldn't know though.

Regarding processing load-issues though - postprocessing algorithms are generally easy to tile, so you aren't hard-limited in how you split, if one SPU lags you could offload more tiles to another one that doesn't, and balance out the load (at the very least, minimize the spikes).

 

Wouldn't you have to wait until RSX has finished with the buffer, and not have a lot of frame time left, thus having to launch all the SPUs on the job concurrently or around the same time?

 

With triple buffering, wouldn't you have at least 1 frame period window (16-33ms) that the MLAA task must be completed in?

 

It probably makes more sense in a triple buffer situation, but that would mean that developers would have to change/deal with other things - latency, memory management, etc. All I'm saying is that this type of technique isn't something that you can simply shoehorn into any game. Surely it would require a lot of planning and work?

 

I would think so. The benefits are high quality AA and lesser GPU load. If the developers need/want them, they will figure out a way to "shoehorn" the feature in. The Saboteur team added it and filtered the UI as well. Other teams may find their own partial ways to do it.

 

Well... GT5 uses 4xAA, dropping this could open PD up to add different in-game effects. And I guess this will be quite a lot in terms of time "saved". If it adds one frame of latency at 60Hz, I guess I could live with it.

And, imho, GT is a game that would benefit HUGELY from such an AA method.

 

On the contrary - crips, long view distance games like GT5 will suffer the most from the effects on such screenspace smoothing filters on subpixel detail in the distance.

 

Hmm... could one combine it with a depth buffer and just not filter in the distance?

 

4xMSAA is still the better solution IMHO.

 

Indeed, 4xMSAA and 4xTSAA gives razor sharp edges at native res or below. Only reason I do 8x or even 24x for some games is becouse I can, not becouse it looks cleaner unless one finds some small difference with a magnifying glass. Hower SSAA makes one big difference just becouse there is more detail per pixel.