Supersampling versus downsampling

[Rennn]

Supersampling (SGSSAA or OGSSAA) and downsampling both have similar results, but in case anyone is interested, I'm going to run through what I've noticed so far. I could be wrong, as this is just from my personal benchmarks and limited testing...

 

OGSSAA

-Very effective at eliminating aliasing

-Can improve scene and texture clarity, especially when used together with a low LoD Bias

-Relatively low performance loss

-OGSSAA works in Mass Effect, Dishonored, and other UE3 games with deferred lighting that are normally incompatible with sampled antialiasing

-Does not alter the size of HUD elements

 

SGSSAA

-Very effective at eliminating aliasing, but not quite as effective as OGSSAA

-Can improve scene and texture clarity, especially when used together with a low LoD Bias

-High performance loss

-SGSSAA also works in games that are normally incompatible with sampled antialiasing

-Does not alter the size of HUD elements

-Seems to retain a slightly sharper image than OGSSAA, in return for the higher performance cost

 

Downsampling

-A downsample of 1440p (2X compared to 1080p) eliminates most jaggies, and is similar to 2X supersampling

-A downsample of 2160p (4X compared to 1080p) eliminates all jaggies, and is similar to 4X supersampling (albeit with a higher FPS cost)

-The most effective option for increasing scene clarity, especially in DX9 games

-Very high performance loss

-Improves AO quality

-Downsampling also works in games that are normally incompatible with sampled antialiasing, but is even compatible with games that do not support any hardware AA

-Alters the size of some HUD elements in some games

[FMod]

Downsampling

-The most effective option for increasing scene clarity, especially in DX9 games

Depends a lot.

 

There's no nice way of putting it: without the crutch of another AA mode, namely either MS or shader AA, a downsampled image looks just bad.

In 2x downsample (1440p), it's almost as jaggy as the original and edges look inferior to simple 4x MSAA, despite similar performance impact to 4x SSAA.

 

Downsampling 4x does look fine in all regards, but its edges still only ~=8xCSAA, at the price of significant compatibility and performance issues (can be painful to even enable).

 

Are you using correct compatibility bits for SSAA modes, BTW? Could be that you only got transparency SSAA active.

 

It's rather DX10 and DX11 games that don't support custom AA modes where you want to use downsampling. Downsampling is reliable at reducing the image quality impact of crutches such as shader-based AA - FXAA, TXAA, etc - which most such games use, and it works there.

 

Sparse vs Ordered grid AA depends entirely on the specific game, both performance wise and quality wise - SGSSAA works well in games that support MSAA and work well with it. Usually there seems to be a positive correlation between performance and quality.

 

With very few exceptions, all DX9 and below games support MSAA forced through drivers. These exceptions are essentially DX11 era engines with a dumbed-down DX9 mode if for some reason native DX11 doesn't work or was cut for consoles.

 

Games that support MSAA, also support either hybrid AA such as 16xS or SGSSAA, and will work correctly with one or both of them. While downsampling combined with proper AA methods will yield pretty good results there, hybrid AA just works better.

 

Although downsampling does have some utility if you want to use SLI AA. But ultimately the difference between AA on the edges comes down to just how well they can be rendered, and there's a limit on that.

Edited September 6, 2013 by FMod

[Rennn]

 

With very few exceptions, all DX9 and below games support MSAA forced through drivers. These exceptions are essentially DX11 era engines with a dumbed-down DX9 mode if for some reason native DX11 doesn't work or was cut for consoles.

 

 

Virtually all DX9 games that use the Unreal 3 Engine (and there are a lot of them; Dishonored, Mass Effect 1, Borderlands, Borderlands 2, Zeno Clash 2, etc.) are incompatible with MSAA, but work properly with both SGSSAA and OGSSAA. I was able to get Dishonored working with a specific compatibility flag for MSAA, but that was still an even greater performance drop than 4X OGSSAA, with obviously lower quality.

 

I was lead to believe that SGSSAA is intended to have a lower performance impact than OGSSAA. However, in all my tests so far, OGSSAA has been far more efficient than SGSSAA. Perhaps that's because I've been testing in primarily DX9 and DX10 Unreal 3 Engine games, with the exception of DAO.

 

Downsampling is only useful in limited circumstances. For example, rendering at 2560x1440, combined with FXAA, is effective at banishing aliasing with minimal amounts of blur and reducing ambient occlusion flicker in Zeno Clash 2.

Edited September 6, 2013 by Rennn

[FMod]

Virtually all DX9 games that use the Unreal 3 Engine...

Mostly that one.

 

Unreal Engine 3 is still very modern, despite its apparent age. It's based on essentially the same technology as newer all-DX11 engines, even if games don't use it to full potential.

 

 

I was lead to believe that SGSSAA is intended to have a lower performance impact than OGSSAA. However, in all my tests so far, OGSSAA has been far more efficient than SGSSAA.

SGSSAA is an additional tier on top of MSAA. Essentially it uses MSAA to get its reference points and then supersamples using them.

 

In games that don't properly work with MSAA, you're going to get either quality loss or severe performance impact from MSAA being forced on a deferred shading based engine.

 

Engines that don't support MSAA should generally use conventional (OG) SSAA. Forcing "blind" MSAA processing doesn't contribute much to sampling accuracy (it does change the grid pattern, however) and the following SSAA technique is the same.

 

Performance impact should be ~same otherwise.

 

 

Downsampling is only useful in limited circumstances. For example, rendering at 2560x1440, combined with FXAA...

Yes.

Shader AA + downsampling works pretty well, because downsampling itself is a poor AA technique, while shader AA is nothing other than a more intelligent form of blur.

 

So blurring the scene with FXAA (or another shader) and then downsampling in turn reduces that blur.

[Rennn]

 

Engines that don't support MSAA should generally use conventional (OG) SSAA. Forcing "blind" MSAA processing doesn't contribute much to sampling accuracy (it does change the grid pattern, however) and the following SSAA technique is the same.

 

Interesting. Should I stop using MSAA along with OGSSAA, then? I was using 4x MSAA to gain reference points for 4x OGSSAA in Mass Effect 1, but based on what you said, is that only useful with SGSSAA?

Now I need to go test this...

Edited September 7, 2013 by Rennn

[FMod]

Interesting. Should I stop using MSAA along with OGSSAA, then?

If you're using the proper hybrid mode? Not necessarily. MSAA+SSAA (each 4x) will result in essentially 16x MSAA on the edges with 4x SSAA.

 

They will operate independently. There's no need to synchronize them, e.g. 4x or 8x MSAA with 2x SSAA is fine.

However MSAA has to be supported by the engine to be useful.

 

 

I was using 4x MSAA to gain reference points for 4x OGSSAA in Mass Effect 1, but based on what you said, is that only useful with SGSSAA?

It only works that way with SGSSAA.

 

In games that support MSAA independently, though, MSAA will also apply its effects on edges.

 

In games that don't, MSAA will glitch and either not work at all (with OGSSAA) or work full-scene (with SGSSAA and correct "hacked" compatibility bits) to be later upgraded by SSAA.

 

The main benefit of SGSSAA is that it works pretty well at 2x, while OG doesn't at all (needs 1x2 or 2x1) and downsampling can create local aliasing.

Edited September 7, 2013 by FMod