Zink OpenGL-On-Vulkan vs. RadeonSI OpenGL Performance As Of January 2021

[peppercats]

I expected it to be slower than native, but not that much. Seems something is off.

It's unfinished. They're still working on feature compliance, performance comes after.

[R41N3R]

Thanks for benchmarking Zink. Quite interesting to see it now and to follow its improvements. At this stage the main focus is to get to OpenGL 4.6. Afterwards I hope the performance improvements will be tracked more closely.

[zboszor]

Now with those 300-500 fps values in some benchmarks, it would really be interesting to measure Mesa 21.1-git against NVIDIA again.

[trilean]

I'd expect the native-vs-Zink performance to eventually be in the same ballpark as DXVK, i.e. 80+%. And at the rate Zink is improving it probably won't be too long before we see those numbers. Keep up the good work!

[pal666]

lol @ people who dreamed about replacing radeonsi with zink

[pal666]

Very interesting! I suppose the reason for the performance difference is that zink is more tuned for the intel vulkan driver?

no the reason is zink has inherent unsolvable overhead

[ssokolow]

no the reason is zink has inherent unsolvable overhead

That's like saying that using OS APIs has inherent unsolvable overhead over twiddling hardware registers directly. OpenGL is a significantly higher-level API than Vulkan, so it's inherently going to have some overhead, but not necessarily as much as going from DirectX 11 or below to OpenGL, which are equally high level.

Vulkan is designed to more closely represent how modern GPUs work so, for all we know, there might be a way to optimize Zink so that it performs on par. It all depends on how much effort is wasted "moving off to the side and then back again" from going OpenGL→Vulkan→Hardware as opposed to OpenGL→Hardware. (Do Mesa Vulkan drivers still translate to NIR before going to hardware or does Vulkan replace NIR in the pipeline?)

[andre30correia]

as expect need a lot of work in performance but right the important thing is putting this working in all apps and games

[geearf]

That's like saying that using OS APIs has inherent unsolvable overhead over twiddling hardware registers directly. OpenGL is a significantly higher-level API than Vulkan, so it's inherently going to have some overhead, but not necessarily as much as going from DirectX 11 or below to OpenGL, which are equally high level.

Vulkan is designed to more closely represent how modern GPUs work so, for all we know, there might be a way to optimize Zink so that it performs on par. It all depends on how much effort is wasted "moving off to the side and then back again" from going OpenGL→Vulkan→Hardware as opposed to OpenGL→Hardware. (Do Mesa Vulkan drivers still translate to NIR before going to hardware or does Vulkan replace NIR in the pipeline?)

NIR can't be replaced by Vulkan you're thinking of SPIR-V, unless you want Vulkan to replace Gallium3D?

[pal666]

That's like saying that using OS APIs has inherent unsolvable overhead over twiddling hardware registers directly.

os brings you resource arbitration, zink brings you only slowdown

OpenGL is a significantly higher-level API than Vulkan, so it's inherently going to have some overhead

this is completely irrelevant. both zink and radeonsi start with opengl. overhead is zink converting gallium to vulkan which then converts it to hardware, radeonsi converts gallium to hardware without extra vulkan step

Vulkan is designed to more closely represent how modern GPUs work so, for all we know, there might be a way to optimize Zink so that it performs on par.

you can't design extra step to work on par with nothing