[AlStrong]

Originally Posted by ToTTenTranz

But wouldn't the higher-resolution rendering + downscaling provide the best possible visual outcome?

Of course.

Is the performance difference so big that it doesn't justify pursuing this method over others that bring some disadvantages to the scene?

Well, you're increasing the texture sampling and shading requirements linearly.

[Deadly Towers]

Originally Posted by ToTTenTranz

Silly question:
As higher DPI screens become standardized, I think antialiasing will eventually lose its importance at anything above 2xMSAA + FXAA.
.

I'm not sure because higher dpi means that you could be shading more detail than the eye can see and still have noticeable aliasing (see John Hable's post on hyperacuity: http://filmicgames.com/archives/698)

[swaaye]

I sometimes downscale 2720x1536 on my 1360x768 TV and it's about as good as 4x ordered grid SSAA. It's not even close to as good as AMD's RGSSAA modes. I definitely prefer it over FXAA and the like though.

I can see the post process AA options being the preferred route because it's cheap and easy. This combined with 1080p being a lot nicer than 640-720p will be a noticeable improvement for the console folk.

[Shifty Geezer]

4x SSAA is only going ot provide the same edge antialiasing as 4xMSAA. For really smooth edges, algorithmic processes are more effective. The trick is filtering the edges and not the surfaces/textures.

[ToTTenTranz]

Originally Posted by swaaye

I sometimes downscale 2720x1536 on my 1360x768 TV and it's about as good as 4x ordered grid SSAA.

From a PC? How do you do that?

[swaaye]

Originally Posted by ToTTenTranz

From a PC? How do you do that?

The NVIDIA custom resolution trick. You define a huge rez and then let the GPU downscale it to your native rez. It works very reliably but while it helps it's only about as good as OGSSAA. But considering other AA rarely works these days, or you can't use transparency AA, and shader effect aliasing is everywhere, and the current FXAA/MLAA stuff sucks, the downscale trick is useful.

And while MSAA is better for poly edges, the downscale trick arguably improves the consistency of the whole picture. It is supersampling afterall. That counts for something even if it's not RGSSAA. Sometimes I add on 2x MSAA for a little extra aliasing killing but it's hard to notice.

[PixResearch]

Last time I dug around in Catalyst Control Centre there was an option to enable supersample AA. It does a great job on edges but I believe it blurs the textures by softening all detail in the frame:

http://www.hardocp.com/article/2009/..._card_review/5

"An old feature being brought back to life on the Radeon HD 5800 series is Supersample AA! Yes, full Supersample AA is now selectable in Catalyst Control Center. You will be able to select whether Multisample AA or Supersample AA is used in all your games. This of course demands a LOT from the hardware, but it has the performance to at least allow some level of it in most games, at lower resolutions. We will show you how this looks in games on the next page. It is great to have this option now present."


"Supersampling AA definitely works in reducing aliasing, on all textures, polygons and even eliminates shader aliasing. However, as you can see above in all three screenshot comparisons, enabling Supersampling AA also causes a reduction in texture quality. By enabling even the lowest 2X Supersampling AA we found that textures started to look blurry, losing their detail and crispness. This, unfortunately, is the nature of Supersampling AA. It literally is a full screen AA method that eliminates jaggies on EVERYTHING, but at the same time degrades texture quality."

[RedVi]

Originally Posted by ToTTenTranz

Ok sorry.

But wouldn't the higher-resolution rendering + downscaling provide the best possible visual outcome?
Is the performance difference so big that it doesn't justify pursuing this method over others that bring some disadvantages to the scene?



BTW, why isn't this a default feature for graphics drivers in PC games?
How hard would it be to "trick" the games into accepting huge resolutions and then the driver would handle the downscaling to the monitor's highest possible native resolution (as we see in videos)?

I know I would definitely enjoy that in less demanding games like console ports.

Adaptive SSAA is already a better solution than this IMO. Less of an overhead with the same advantage of being able to sample from a higher resolution image.

[Simon F]

Originally Posted by ToTTenTranz

Silly question:
Why not just render at a very high resolution (say, QFHD 2160p or higher?) and then use a high-quality hardware downscaler to 1080p?

No it wouldn't. To see this you need to take the fourier transform of a rectangular grid of points (representing your "QFHD 2160p or higher" source image) and comparing the frequency space result with the transform of a "less regular" placement of sample locations. The latter should shift the unrepresentably high frequencies (i.e those that alias) further towards the blue end of the spectrum which make them easier to filter out.

[swaaye]

Originally Posted by RedVi

Adaptive SSAA is already a better solution than this IMO. Less of an overhead with the same advantage of being able to sample from a higher resolution image.

Are you referring to some sort of special SSAA or to AMD's Adaptive AA (alpha texture anti-aliasing)? Combining MSAA and Adaptive/Transparency AA still misses a lot of aliasing of various effects in a current-day game. One has to go far back in time to find a game that is thoroughly anti-aliased by MSAA+AAA/TAA.

[Shifty Geezer]

I don't know if adaptive SSAA is implemented in any GPU solution, but it's a technique that effectively measures what you're rendering and adds samples as needed. It was used in the Cell terrain demo many years ago. You don't waste cycles rendering 16 samples of blue sky or similar grass, but you'd render 16 samples per pixel along edges or noisy surfaces.

[jlippo]

Originally Posted by Shifty Geezer

I don't know if adaptive SSAA is implemented in any GPU solution, but it's a technique that effectively measures what you're rendering and adds samples as needed. It was used in the Cell terrain demo many years ago. You don't waste cycles rendering 16 samples of blue sky or similar grass, but you'd render 16 samples per pixel along edges or noisy surfaces.

This should be quite easily possible with any deferred renderer using MSAA.
You already have to find edges for secondary pass to perform super sampling.
In this phase you could mark certain materials or search areas which have strong normal variation for extra sampling.

IMHO.
Proper way is to reduce the shading aliasing is in the shader itself, this way you do not have sampling limits, you can do it analytically or prefilter data.. (Clean/Lean mapping..etc.)

[Shifty Geezer]

Fractal talk moved here.

[almighty]

They'll be using the newer versions of FXAA/MLAA but MSAA is still the best option there is..

Why I love being a PC gamer, MSAA+TrSAA is the king of IQ!

And it's even better now TrSAA is fully supported in DX9/10/11

[MJP]

Originally Posted by jlippo

This should be quite easily possible with any deferred renderer using MSAA.
You already have to find edges for secondary pass to perform super sampling.
In this phase you could mark certain materials or search areas which have strong normal variation for extra sampling.

IMHO.
Proper way is to reduce the shading aliasing is in the shader itself, this way you do not have sampling limits, you can do it analytically or prefilter data.. (Clean/Lean mapping..etc.)

Well the problem is that you don't want to just supersample the lighting pass, you also want to supersample the shader that outputs the G-Buffer properties (and also the alpha test, if you're using that). This is trickier, unless you decide per shader/material what's going to be run at per-sample frequency. It can also be really expensive.

[user542745831]

Originally Posted by ToTTenTranz

From a PC? How do you do that?

http://translate.google.com/translat...hl=en&ie=UTF-8

http://forums.guru3d.com/showthread.php?t=346325

http://www.pcgameshardware.com/aid,6...sampling/News/
http://www.pcgameshardware.com/aid,7...sampling/News/

[swaaye]

Originally Posted by jlippo

IMHO.
Proper way is to reduce the shading aliasing is in the shader itself, this way you do not have sampling limits, you can do it analytically or prefilter data.. (Clean/Lean mapping..etc.)

I've been reading this on B3D for years now but why aren't game developers doing this now? I suppose it's just another result of the old consoles holding everything back?

Originally Posted by almighty

They'll be using the newer versions of FXAA/MLAA but MSAA is still the best option there is.. Why I love being a PC gamer, MSAA+TrSAA is the king of IQ!

Quite a few games still won't work with forced MSAA/SSAA. AFAIK, NVIDIA and AMD have to profile games and set the driver up to have forced AA work in most cases these days. Sometimes they don't bother. I find it completely ridiculous that we have not gotten past this crap yet. FXAA and MLAA have probably made things even worse because now developers can just easily drop in a fairly low overhead full screen effect that looks like nice AA (in screenshots anyway).

[Shifty Geezer]

Originally Posted by swaaye

I've been reading this on B3D for years now but why aren't game developers doing this now? I suppose it's just another result of the old consoles holding everything back?

Lack of shading power for the consoles, and if you're usign the same shaders on PC they'll also be unfiltered.

Quite a few games still won't work with forced MSAA/SSAA. AFAIK, NVIDIA and AMD have to profile games and set the driver up to have forced AA work in most cases these days.

I find it odd the games all run on DirectX and yet can't have a simple flag to toggle exposed to the users.

[ERP]

Comes back to an issue I have with forcing anything in the driver.

Yes you get features like AA in older games, but what you really do as the driver writer is assume responsibility for the feature.
.
Later as rendering techniques evolve and as a result the driver hacks no longer work devs still feel no real need to implement the feature. Why should they, if the game is popular enough the driver writers will have to hack the functionality in for them.

Back when every rendering pipeline was based on OGL 1.X driver overrides made a lot of sense, they were just overrides for global state values. Now when there are so many parts of the pipeline programmable, you're pretty much guaranteed any generic approach will fail on any renderer that pushes the boundaries.

I have the same issue with the automatic stereoscopic 3D feature in the NVidia drivers, you need it short term, but it's a hack and not a long term solution and in the end all it will really do is delay adoption by devs.

[Fafalada]

Originally Posted by ERP

but it's a hack and not a long term solution and in the end all it will really do is delay adoption by devs.

I have a lot of dislike for driver overrides for the simple reason that they make what's already not much of a platform(to put it mildly) into something even less predictable, as you can't even trust the things application sets itself to stick anymore.
And responsibility always (at least partially)gets pushed on the devs, whenever a game is "incompatible" with driver overrides too.

[swaaye]

It sounds like a nightmare. I don't see how it would ever go away though because end users expect to be able to force options and the IHVs use it as selling points.

The forced driver features have been around forever. I remember Riva 128 being troublesome because of its auto mipmap generation that messed up some games. I think NV backpedaled on that feature in later drivers.

[ebola]

can multiple-render-targets implement the old 3dfx t-buffer idea efficiently (i.e. multiple instances of the same triangle rendered simultaneously fetching the same textures, fetching the geometry once) .. if so do any demos exist

[lightarrow]

Why cant they have a separate feature for racing games that give just the whole car say 16xaa while the rest of the environment has only 4xaa? Or same with weapons & hud.

[jlippo]

Originally Posted by lightarrow

Why cant they have a separate feature for racing games that give just the whole car say 16xaa while the rest of the environment has only 4xaa? Or same with weapons & hud.

Of course you can do it.
Problem is the memory consumption etc. for the 16x pass and the actual combination pass.