Hi!

I am doing my rendering engine.

It contains GBuffer that has no subsamples (only 1 sample per pixel).

I added FXAA post-processing filter and it works good.

I decided to do better and after some research found SMAA filter.

SMAA filter has several modes: 1x, S2x, 4x;

1x works with only 1 sample/pixel, but S2x and 4x are require input texture to be MSAA 2x.

And now I get lost.

Does it mean that all textures inside my GBuffer (Normal texture, Diffuse, Specular, Emissive, HDR light accumulator, and HDR output) should be MSAA 2x also?

I can't understand: after geometry pass, would subsamples in Normal, Diffuse, Specular... textures contain information of different geometries?

In other words: in light pass, would it be possible to run on sub-pixel level of GBuffer or all geometry information will merged to per-pixel data right after geometry pass?

Are there some caveats I should be aware of before moving GBuffer to be MSAA 2x?

Or should I do something different in order to achieve SMAA 4x quality with reasonable performance impact?

Has anyone implemented engine with GBuffer and SMAA?

How does it look like (stages+main textures with or without MSAA)?

It seems I came to a point where tutorials are no longer source for further information

Thanks in advance.

One thing that puzzles me, is where the MSAA subsamples information is stored?

I am curious about this because I can't figure out, how to do straight MSAA between geometry and light passes in deferred renderer.

Common sense says to me that subsample information should be stored somewhere before I do resolve, but I can't find it in a debugger.

All the articles I read say that MSAA increases memory footprint proportionally of samples count.

My engine is capable of forward rendering with 8xMSAA and here the only biggest textures were created:

Name Size Type width height Samples

Forward::DepthStencil D3D11 Texture2D 7916912 R24G8_TYPELESS 1916 1033 8

Forward-BackBuffer DXGI Surface 7916912 R8G8B8A8_UNORM 1916 1033 8

The same texture sizes I can see with sample count = 1 in other, deferred renderer mode:

Name Size Type width height Samples

GBuffer::8_HDR Light Accum D3D11 Texture2D 15833824 R16G16B16A16_FLOAT 1916 1033 1

Deferred::DepthStencil D3D11 Texture2D 7916912 R24G8_TYPELESS 1916 1033 1

Deferred-BackBuffer DXGI Surface 7916912 R8G8B8A8_UNORM 1916 1033 1

GBuffer::0_Normal D3D11 Texture2D 7916912 R10G10B10A2_UNORM 1916 1033 1

I see the same result in VS 2015 and NSight 5.0

In the other words, if I am going to make all textures in GBuffer 2xMSAA, it should increase memory footprint twice.

But right now I can't see size difference between textures created with 8xMSAA and without MSAA.

Any ideas? =)

If it does matter, I have NVidia 660GTX.

I found a way to measure memory footprint for MSAA textures!

Process Explorer can show it on process properties tab:

Deferred renderer has a lot of full-screen textures (about 10), all of them are no-AA,

Forward renderer has only 2 (render target and depth), but it uses twice memory than Deferred.

It seems that VS 2015 and NSight 5.0 debuggers mean something different for size of textures.

So I am ready to convert my GBuffer to 2xMSAA in order to use 4xSMAA!!! :)