RADV Vulkan vs. RadeonSI OpenGL Performance With Linux 4.13 + Mesa 17.3-dev

[bridgman]

Interesting numbers. I guess we're seeing "lower overhead" vs "less optimization" with the latter gradually going away over time.

[microcode]

so, the only game where radv is better is talos
no, mesa doesn't support windows api(directx)

For what it's worth, there have been state trackers (with varying completeness) for all of the recent D3D versions, and there is a pretty good D3D9 state tracker in Mesa already. It is not inconceivable that Mesa and the matching kernel infrastructure could be ported, and that would provide lots of reasons and samples for developing D3D10, D3D11, and D3D12 impls.

Also, I think on Windows you would be implementing something lower level (DXGI, D3D DDI, etc.) than the D3D API itself, and the D3D library would hook into that to compile shaders and do other such things.

[Termy]

Actually i have a RX480 with a A10-7850k CPU, so that classify as a fast card with a weak cpu ... and at least in mad max, i get better performance in vulkan than opengl

that's what i wanted to know
which version of mesa are you running?

[fagnerln]

OGL requires a lot of single thread performance, if it isn't a problem, it will shine. The only benefit on Vulkan is the multi thread performance (when loses to ogl, it's a driver problem of course), you can't reproduce this on a i7 7700K.

This test would be more interesting in an AMD cpu like A or FX series.

[leipero]

i don't know if i missed it, but i still would prefer to see these kind of benchmarks on a more mundane CPU to see if RADV/vulkan shines when the CPU is the bottleneck

Yeah, look at Dota 2 test, FHD performance on Vulkan are better, 4k different story. And as higuita pointed out, it seems different on other configurations, heck even I with r600g can run Dota 2 with high settings on my prefered resolution (approximately higher than HD 720p) with well over 80FPS smooth, so for those GPU's tested is cake walk to do Dota 2 with highest settings FHD, and they face CPU/game engine limitation and Vulkan helps, but in 4k Vulkan harms performance, so there's room for improvements.

[Apopas]

Vulkan slowly but steadily is catching up! Kudoz!

[artivision]

For what it's worth, there have been state trackers (with varying completeness) for all of the recent D3D versions, and there is a pretty good D3D9 state tracker in Mesa already. It is not inconceivable that Mesa and the matching kernel infrastructure could be ported, and that would provide lots of reasons and samples for developing D3D10, D3D11, and D3D12 impls.

Also, I think on Windows you would be implementing something lower level (DXGI, D3D DDI, etc.) than the D3D API itself, and the D3D library would hook into that to compile shaders and do other such things.

I don't think that is needed anymore. In a 3110m-8670m laptop testing D3D11, i get the same framerate on Linux as on windowz. Tomb Raider 2013 medium: msD3D9=30fps, msD3D11=40fps, nine=30fps, wineD3D11=40fps plus feels more responsive. Rise of the Tomb Raider low: msD3D11=15fps, wineD3D11=15fps plus artifacts - cpu idle and dgpu full.

[andre30correia]

feature ready, only need optimization

[leipero]

I don't think that is needed anymore. In a 3110m-8670m laptop testing D3D11, i get the same framerate on Linux as on windowz. Tomb Raider 2013 medium: msD3D9=30fps, msD3D11=40fps, nine=30fps, wineD3D11=40fps plus feels more responsive. Rise of the Tomb Raider low: msD3D11=15fps, wineD3D11=15fps plus artifacts - cpu idle and dgpu full.

I disagree, that's not proper way to test it (at least for one game), what you need to do is to load game with quite complex geometry, you will see insane performance hit by wine D3D (and I'm speaking about D3D9 here, we can only expect worse with newer versions i guess). I did that type of testing, and as soon as geometry complexity increases, difference that could be as low as few percentage goes over two times lower on wine, and I'm not sure how much high end CPU's could help in those cases.

[pal666]

For what it's worth, there have been state trackers (with varying completeness) for all of the recent D3D versions, and there is a pretty good D3D9 state tracker in Mesa already. It is not inconceivable that Mesa and the matching kernel infrastructure could be ported, and that would provide lots of reasons and samples for developing D3D10, D3D11, and D3D12 impls.

ok, in fantasy world mesa can be used on windows. in real world it can't