AMD's Open-Source Radeon Driver After Four Years

[marek]

I certainly DON'T think that everybody should move over to Gallium3D drivers on older hardware as for many people, that'd be a loss of features (no MSAA, no Hyper-Z, less 3D performance, etc.)... User's shouldn't have to sacrifice features...

I can tell you for sure that even though MSAA is missing and most of HyperZ is disabled by default (you can enable it though) and the performance isn't at 100% of the level of Catalyst, the open r300g driver has much more features in total. There are over 50 more OpenGL extensions in r300g than in Catalyst 9.3 on the same hardware. That's a ton of features. See for yourself.

[liam]

"cutting edge feature"

LOL ??? not really its a poor man feature real man get this: Rotated Sample-Super-Sample-AA RS-SSAAx8

but this killer cutting edge feature kill all hardware thats why they do MLAA the poor mans AA..

Cutting edge in that it is new, looks pretty much as good as MSAA and is apparently faster.
http://www.youtube.com/watch?v=d31oi1OOKbM

[smitty3268]

"cutting edge feature"

LOL ??? not really its a poor man feature real man get this: Rotated Sample-Super-Sample-AA RS-SSAAx8

but this killer cutting edge feature kill all hardware thats why they do MLAA the poor mans AA..

This is a strange definition of cutting edge. By that logic, the new smartphones and tablets this year can't be cutting edge because i can already buy a 3Ghz PC.

[BlackStar]

I'm very (pleasantly) surprised this is coming to Linux so quickly.I also like the idea of MLAA. A nice, simple post-processing idea that can be applied to any scene (though, unlike BlackStar I don't think it is variable only across resolution since it seems very much edge/color contrast dependent).

As far as I know, the filter has two passes: edge detection and blurring. The kernel for each pass gets executed exactly once for each pixel on the framebuffer, regardless of the amount of edges in the scene (unlike MSAA which gets executed per edge). Now, performance might be somewhat correlation with the amount of edges if the shader uses branches (which I doubt) but even if this is the case... 12 times faster than MSAA 8x? This wouldn't happen if the amount of edges affected MLAA significantly.

[smitty3268]

As far as I know, the filter has two passes: edge detection and blurring. The kernel for each pass gets executed exactly once for each pixel on the framebuffer, regardless of the amount of edges in the scene (unlike MSAA which gets executed per edge). Now, performance might be somewhat correlation with the amount of edges if the shader uses branches (which I doubt) but even if this is the case... 12 times faster than MSAA 8x? This wouldn't happen if the amount of edges affected MLAA significantly.

There appear to be 3 passes, actually: Edge detection, blend weights, and smooth edges.

Also, i think the 2nd and 3rd passes are based on edges, or at least are optimized to a degree for non-edges. But I'm not very familiar with graphics programming.

The implementation for Mesa is in TGSI here: http://lists.freedesktop.org/archive...st/010629.html

which is based on the Jimenez MLAA code, which you can view in a more readable DX10 version here: https://github.com/iryoku/jimenez-ml...haders/MLAA.fx

Maybe someone smarter than me can make more sense out of it.

[curaga]

As far as I know, the filter has two passes: edge detection and blurring. The kernel for each pass gets executed exactly once for each pixel on the framebuffer, regardless of the amount of edges in the scene (unlike MSAA which gets executed per edge). Now, performance might be somewhat correlation with the amount of edges if the shader uses branches (which I doubt) but even if this is the case... 12 times faster than MSAA 8x? This wouldn't happen if the amount of edges affected MLAA significantly.

Three passes, and only the first (edge detection) is executed for all pixels. The second and third pass are only executed for the edge-marked pixels (via the stencil), which brings a big speedup compared to running them on all pixels.

The second pass is what makes it so much better in quality than many of the other implementations; instead of a constant blur, it depends on the type of aliasing. I recommend the fxaa vs mlaa article @ digitalfoundry, you can see how much more fxaa blurs.

[Sidicas]

Cutting edge in that it is new, looks pretty much as good as MSAA and is apparently faster.
http://www.youtube.com/watch?v=d31oi1OOKbM

As a gamer on nvidia hardware (under Windows) I've found that MSAA and CSAA fall short while FXAA and MLAA really nail it..
The major problem with MSAA & CSAA is that it doesn't anti-alias shaders very well (if at all)... So you have these very bright pixels on the edges of objects that are caused by shiny objects but the bright pixels aren't anti-aliased properly because the shaders are at a different level than MSAA/CSAA.. So at decent resolutions (1080p) and very high texture & lightning details you can see some course grainy pixels on the texture caused by the bright reflective lighting on the edges of shiny objects (like wet steps).. MLAA and FXAA fixes that while you can crank MSAA and CSAA to the max all day long and it will do nothing to fix those annoying coarse white pixels along the edges of very shiny objects (which can really detract from the realism, IMO).

[Qaridarium]

This is a strange definition of cutting edge. By that logic, the new smartphones and tablets this year can't be cutting edge because i can already buy a 3Ghz PC.

my point isn't in that way. my point is RS-SSAAx8 beats MLAA in "Quality"

RS-SSAAx8 is the best of the best of the best!

[BlackStar]

Three passes, and only the first (edge detection) is executed for all pixels. The second and third pass are only executed for the edge-marked pixels (via the stencil), which brings a big speedup compared to running them on all pixels.

I am missing some secret sauce here. How does the stencil get written? Via GL_ARB_stencil_shader_export (is this even supported on Mesa? It requires GLSL 1.40!) And if yes, why doesn't the loss of early z-tests destroy performance?

The second pass is what makes it so much better in quality than many of the other implementations; instead of a constant blur, it depends on the type of aliasing. I recommend the fxaa vs mlaa article @ digitalfoundry, you can see how much more fxaa blurs.

Thanks, will do.

Edit: I missed the part where this is written in TGSI rather than GLSL. But still, the hardware limitations should be identical. Need to think about this some more.

[curaga]

I am missing some secret sauce here. How does the stencil get written? Via GL_ARB_stencil_shader_export (is this even supported on Mesa? It requires GLSL 1.40!) And if yes, why doesn't the loss of early z-tests destroy performance?

Via

glClear(GL_STENCIL_BUFFER_BIT);
glStencilFunc(GL_ALWAYS, 1, ~0);
glStencilOp(GL_KEEP, GL_KEEP, GL_REPLACE);
// pass 1 here

Since pass 1 calls discard for non-edge pixels, the stencil is not marked for those. Early-Z isn't lost this way. This is quite a smart optimization, my jaw dropped too when I first saw it.