Mesa Zink Improvements For OpenGL-On-Vulkan Reportedly Make It Faster Than Radeon OpenGL

[slagiewka]

If you use -vulkan as launch options, it will use dxvk-native

Turns out I just needed persistence. For me adding the flag made the game instantly exit. But... trying it another 2 times made it work. For some reason I also expected to have Vulkan when running `-vulkan` but it's just a DXVK lol. I would never call it native, but Valve do that for ToGL too

Quick test seemed like it stutters more than I'd like to, so that's a pass for me

[cb88]

I meant on Windows. The Intel ARC driver is currently very buggy and slow with D3D9/11 but works fine with Vulkan.

That is because it doesn't even have a native D3D9-11 driver... it only has native DX12 and Vulkan. It uses a fallback layer built into windows to emulate DX9-11 similar to DXVK except built at the OS level with some extra hinting.

[fagnerln]

Quick test seemed like it stutters more than I'd like to, so that's a pass for me

It takes sometime to build the shader cache, once it's built, it runs smooth. The problem is that every time the driver updates it need to create again.

The only issue is that the only available resolution is the native.

[ayumu]

With Zink, on an AMD system you have

Application -> Zink -> RADV -> GPU

or

Application -> RadeonSI -> GPU

Zink being faster than RadeonSI points to some kind of inefficiency in RadeonSI. The issue might be that RSI is very conformant even with odd, old and buggy OpenGL code whereas Zink probably isn't.

It's easy to forget opengl is quite the high level interface, and it is indeed sensible to have vulkan behind it.

While this does point to inefficiencies in radeonsi, there's a more pressing question: Is it worth maintaining that in the first place, when Zink exists?

I don't think so. Vulkan (relatively low level) is the better interface to the hardware, whereas high level languages such as opengl can be built on Vulkan, eliminating pointless duplication of work.

[MadCatX]

It's easy to forget opengl is quite the high level interface, and it is indeed sensible to have vulkan behind it.

Just because it's possible doesn't mean it's sensible.

While this does point to inefficiencies in radeonsi, there's a more pressing question: Is it worth maintaining that in the first place, when Zink exists?

Yes, because radeonsi supports GPUs that don't have Vulkan support.

I don't think so. Vulkan (relatively low level) is the better interface to the hardware, whereas high level languages such as opengl can be built on Vulkan, eliminating pointless duplication of work.

The point of Zink is to not have to develop OpenGL drivers for new GPUs since OpenGL is now considered a legacy API. The biggest challenge might be the implementation of compatibility profiles and all sorts of workarounds to get all of those legacy programs to render correctly. I wonder what the plans are here.

[user1]

Zink has surely a higher overhead than plain OpenGL, but with radeonsi cards has the advantage of leveraging aco compliler in radv, while radeonsi still uses old llvm compiler, and this may be an advantage in some cases.

Yeah, people here don't realize that when we compare RadeonSi to Zink, we also compare LLVM to ACO since it runs on RADV. People forget that ACO may outperform LLVM by 10% or more in some cases.
Personally, I don't share the excitement of Zink replacing RadeonSi because in its current state, it's simply too early, and in the future, it might be faster than native driver in some cases, but in others slower, just like DXVK which is already pretty mature. Zink would've been exciting for me if it had a clear advantage over RadeonSi. Also, RadeonSi is probably the most optimized native GL driver out there.
I'm much more excited in RadeonSi finally adopting ACO, which will give it another performance boost.

[trilean]

With Zink, on an AMD system you have

Application -> Zink -> RADV -> GPU

or

Application -> RadeonSI -> GPU

Zink being faster than RadeonSI points to some kind of inefficiency in RadeonSI. The issue might be that RSI is very conformant even with odd, old and buggy OpenGL code whereas Zink probably isn't.

You're missing the Gallium layer in RadeonSI. Can't find the quote right now, but I read a statement by a Mesa developer (IIRC) somewhere saying that "Vulkan is basically Gallium 2". From that perspective I don't see why Zink should inherently be any slower than RadeonSI.

[MadCatX]

You're missing the Gallium layer in RadeonSI. Can't find the quote right now, but I read a statement by a Mesa developer (IIRC) somewhere saying that "Vulkan is basically Gallium 2". From that perspective I don't see why Zink should inherently be any slower than RadeonSI.

Zink is also a Gallium3D driver. The only difference is that Zink calls Vulkan as the "hardware" backend whereas other drivers call the actual hardware (or rather, its kernel driver).

[yoshi314]

if that is true, it might simply show how much overhead opengl has - which is plausible.

i mean, we've seen how first vulkan games outperformed themselves running on dx/opengl backends. so i would not rule out the possibility of zink being all-around more performant- with some exceptions.