Nvidia's MFAA demo (BF4)

[f1sherman]

Looking same or better than 4xMSAA. Performance hit of 2xMSAA:

https://www.youtube.com/watch?v=eg5iKiQcRMc

I expect there will be cases where 4xMSAA will look better, but overall MFAA should be the more desirable solution.

 

[Silverforce11]

MFAA = 4x MSAA when still, 2x MSAA when in movement.

It's not magic that the -1n frame as a reference would be off in fast movement thus that cache will be flushed and not used, else you get massive motion blur.

They are betting that when you move fast, you won't notice the difference between 2 vs 4x MSAA, but when you stand still or for screenshots, you will certainly notice it.

Which is true.

They should have called it Dynamic AA or Adaptive AA: AAA! Triple A is widely recognized as a good thing. This is a case of marketing failure since they did not capitalize on the naming better.

 

[Phynaz]

Why do the colors change when they are switching between aa types? Thinking mxaa is just another post process blur filter.

 

[f1sherman]

MFAA = 4x MSAA when still, 2x MSAA when in movement.

It's not magic that the -1n frame as a reference would be off in fast movement thus that cache will be flushed and not used, else you get massive motion blur.

They are betting that when you move fast, you won't notice the difference between 2 vs 4x MSAA, but when you stand still or for screenshots, you will certainly notice it.

Which is true.

They should have called it Dynamic AA or Adaptive AA: AAA! Triple A is widely recognized as a good thing. This is a case of marketing failure since they did not capitalize on the naming better.

You are betting that -1 frame has no information that can be used to build the better animation centered around 0 (zero) frame.
This is totaly counter-intuitive to me

SMAA T2x and TXAA are looking BETTER in motion and they are also collecting samples over time/ from previous frames?
It's only natural that taking samples over time should average your animation better than time fixed collection.

 

[f1sherman]

And yes, that sampling process sounds like it can result in blurring even easier than tradition AA methods.
Yet I see no apparent blurring in BF4 demo, and here MFAA looks like a clear winner to me.

Then again nothing is free so it could be that some of what you're saying is true.

They should have called it Dynamic AA or Adaptive AA: AAA! Triple A is widely recognized as a good thing. This is a case of marketing failure since they did not capitalize on the naming better.

LOL that aa name is already taken. Metro...

 

[Silverforce11]

You are betting that -1 frame has no information that can be used to build the better animation centered around 0 (zero) frame.
This is totaly counter-intuitive to me

MSAA breaks the scene into grids where it then samples points.

In a fast moving scene, the -1n frame may not fall into the same grid as the n frame.

This is why I think NV's drivers ignore the -1n frame when it does not align, defaulting to 2x MSAA in motion.

Some will call it cheating, but its just clever optimizations. Why? Because I generally don't notice the difference between 2 vs 4x MSAA when I move around in FPS. So while my GPU is grunting away doing 4x MSAA, the final IQ to me is not obvious, ie. it's wasted power.

It should definitely be Adaptive AA because thats exactly what it is, switching between 2 and 4x MSAA depending on the scene.

Edit: The BF4 demo should work exactly like 4x MSAA since its a stationary scene in terms of polygon data, the dynamic part is the rain & water which is probably shaders.

 

[SirPauly]

Thanks for that!


nVidia offered,"MFAA is applied everywhere," and pointed out transparency advantages!

 

[f1sherman]

@Silverforce11

I get what you are saying, and this could very well be happening due to no blurring.
But it's still a bold claim.

Your proposition sounds like a clever but less exciting optimization than I would hope for.

@SirPauly

idd, full screen MFAA should be a clear winner with all those games that miss alpha textures.

 

[Subyman]

Silverforce, what you are saying is that MFAA is just MSAA, but applied differently depending on movement? That seems to go against what the presenter was saying since he said MFAA offered advantages to MSAA in alpha transparency in that it sampled not just the edges. It would seem more is going on that just swapping MSAA 2x and 4x.

 

[ams23]

MFAA requires second gen Maxwell hardware to leverage the much more flexible and much more programmable pixel and sub-pixel sampling patterns that can be used on these GPU's. In theory, MFAA can provide better image quality than MSAA (and vice versa), but it is still too early to say anything definitively at this point.

The best description I have seen so far is from overclockersclub review:

"Multi Frame Sampled AA or MFAA: "Game developers and GPU vendors are increasingly implementing more advanced forms of anti-aliasing (AA) to enhance image quality. GM2xx GPUs have a number of new features for much more flexible sampling, enabling further advancements in AA quality and efficiency. Today’s GPUs ship with fixed sample patterns for AA that are stored in ROM. When gamers select 2x or 4xMSAA for example, the pre-stored sample patterns are used. While many current games implement deferred, post-processed AA techniques such as FXAA or SMAA, there are still many others that continue to use traditional hardware-based multi-sample AA (MSAA). GM2xx GPUs support multi-pixel programmable sampling for rasterization, providing opportunities for more flexible and novel AA techniques to be implemented in the context of both deferred and conventional forward rendering."

"With programmable sample positions, the ROMs that were used to store the standard sample positions are replaced with RAMs. The RAMs may be programmed with the standard patterns, but the driver or application may also load the RAMs with custom positions which are free to vary from frame to frame or even within a frame. In a 16x16 grid per pixel, we have 256 different locations to choose from for each sample. We’ve also extended this programmable sample location support to span multiple pixels, so for example in 4x MSAA rendering, all 16 samples within a 2x2 pixel footprint can be located arbitrarily. This sample randomization can greatly reduce the quantization artifacts (like stair-stepping) that occur with more traditional forms of AA. These freely specified sampling positions may be used in the development of effective new algorithms."

"NVIDIA engineers have recently developed new AA algorithms that vary, in interleaved fashion, the sample patterns used per pixel either spatially in a single frame (where, for example, each successive pixel uses one of four different 4xAA sample patterns) or interleaved across multiple frames in time. Multi-Frame Sampled AA (MFAA) is a new AA technique that alternates AA sample patterns both temporally and spatially to produce the best image quality while still offering a performance advantage compared to traditional MSAA. The final result can deliver image quality approaching that of 8xAA at roughly the cost of 4xAA, or 4xAA quality at roughly the cost of 2xAA. Below we have a few images that show 4XMSAA implemented in BF4 and on the right the same scene with MFAA enabled."

All that said, there is no doubt that perf. and perf. per watt can improve significantly with MFAA vs. MSAA:

 

[Silverforce11]

Silverforce, what you are saying is that MFAA is just MSAA, but applied differently depending on movement? That seems to go against what the presenter was saying since he said MFAA offered advantages to MSAA in alpha transparency in that it sampled not just the edges. It would seem more is going on that just swapping MSAA 2x and 4x.

I only take it at what JH was presenting during the Game 24 speech. He was explaining the theory behind it.

But sure, they can add a post AA filter to it as well that handles transparency. Have to see it in detailed and in motion to tell whether there is a trade-off for IQ.

If it does not blur and produces 4x MSAA equivalent results, then its a great AA mode for sure! I am anti-post AA because nearly every example in games were rubbish, completely ruining the IQ with the blurfest to the point ultra quality textures look like low/medium.

 

[Anarchist420]

i don't see anything wrong with the properly rotated grid super sampling Nixxes/Crystal D used for Tomb Raider or anything wrong with driver-forced SGSSAA.

sounds to me like MFAA is some more me-too stuff or something that wasn't as important as something else that could be invented or something that could be fixed.

otoh, if MFAA is more programmability, then it is a step in the right direction to get devs used to programming everything. but single precision won't be enough. no fixed functions (except good ones like lag-free 1:1 pm in the on-die display logic) along w/ non-crippled 64 bit float precision with pro-speed pro-precision instructions in the compute shader is a good idea, especially for the future. fixed function and crippling double precision performance doesnt allow for all that could be.

 

[BFG10K]

The temporal part of this is absolutely useless, just like TXAA and TAA. I don't understand why nVidia keeps pushing out this kind of garbage given the artifacts and/or blur associated with this kind of blending.

OTOH the spatial part is a very good solution, similar to what the Radeon 8500 was doing back in the day. Spatial 2xMFAA (for example) should look better than 2xMSAA (probably 2.5x or 3x MSAA IQ-wise) while running at the same speed.

Compatibility-wise it should be no better than regular MSAA except 2xMFAA (temporal), which I'd expect to work in any game.

 

[BrightCandle]

The whole point of temporal AA is to remove the shimmering that comes with MSAA. You can remove the jaggies but introduce another artefact while moving. TXAA removes that shimmer but at the cost of blur and quite a lot of performance. If MFAA can remove the shimmer and not cost performance (but still blur) then its a trade off that might be worth making.

 

[BFG10K]

The whole point of temporal AA is to remove the shimmering that comes with MSAA. You can remove the jaggies but introduce another artefact while moving.

MSAA doesn't cause shimmering. The shimmering is what's left where MSAA has no effect; it stands out more because of the reduction to polygon aliasing.

 

[SirPauly]

The whole point of temporal AA is to remove the shimmering that comes with MSAA. You can remove the jaggies but introduce another artefact while moving. TXAA removes that shimmer but at the cost of blur and quite a lot of performance. If MFAA can remove the shimmer and not cost performance (but still blur) then its a trade off that might be worth making.

One may use a third party tool to add sharpness; based on smoothness and softness is subjective with tolerant levels differing.

TXAA was welcome based on it tries to tackle temporal aliasing, while being efficient when compared to traditional or conventional SSAA methods.

 

[sontin]

The temporal part of this is absolutely useless, just like TXAA and TAA. I don't understand why nVidia keeps pushing out this kind of garbage given the artifacts and/or blur associated with this kind of blending.

The only temporal part is the calculation of the sample points from the previous frame. There is no introduction of blur.

 

[SirPauly]

There was no blur with ATI's temporal AA but if one set the frame-rate threshold manually too low, one may see flickering where the aa positioning was taking place.

 

[BFG10K]

The only temporal part is the calculation of the sample points from the previous frame. There is no introduction of blur.

Not true - they run a filter to combine n -1 with n, just like TXAA.

 

[sontin]

Not true - they run a filter to combine n -1 with n, just like TXAA.

They dont combine Frame-1 with Frame n. They use the result of Frame-1 to calculate the geometry color value of the pixel in Frame n.

 

[Erenhardt]

They dont combine Frame-1 with Frame n. They use the result of Frame-1 to calculate the geometry color value of the pixel in Frame n.

Which is new motion blur technique!

 

[BFG10K]

They dont combine Frame-1 with Frame n. They use the result of Frame-1 to calculate the geometry color value of the pixel in Frame n.

Err, if they're combining data from the previous frame with the current frame in whatever capacity, that's the definition of a blend.

Also their video states a filter runs to combine the frames, and repeatedly references pixels in a finished state. Finished frames have no sub-pixel information suitable for geometry calculations.

 

[Keysplayr]

Regardless of how it is accomplished, is the finished result equivalent to a level of MFAA visual quality at less performance hit or is it not?

 

[sontin]

Err, if they're combining data from the previous frame with the current frame in whatever capacity, that's the definition of a blend.

Also their video states a filter runs to combine the frames, and repeatedly references pixels in a finished state. Finished frames have no sub-pixel information suitable for geometry calculations.

They use only those pixels which are covered by the sample points and blend them offer to Frame n. Blur is the same as it would be with MSAA.

 

[BFG10K]

They use only those pixels which are covered by the sample points and blend them offer to Frame n.

This statement makes no sense whatsoever. Every pixel has at least one sample point, even without AA. Without a sample you couldn't make the pixel in the first place.

Furthermore, a pixel covers a sample point, not the other way around.

Can you please provide a link to any document where you're getting this information from? Thanks.