Zink OpenGL-On-Vulkan Sees More Performance Tuning, Regression Fixing

[phoronix]

Phoronix: Zink OpenGL-On-Vulkan Sees More Performance Tuning, Regression Fixing

Following my recent RADV+Zink vs. RadeonSI OpenGL benchmarking for various games and workloads, Valve's Zink lead developer Mike Blumenkrantz was hopping on some of the benchmarks where this generic OpenGL-on-Vulkan implementation lagged behind the RadeonSI Gallium3D driver...

https://www.phoronix.com/news/Zink-Fixing-ET-Bioshock

[tildearrow]

Typo: twekas

[Quackdoc]

gotta go fast indeed lmao

[R41N3R]

Nice improvements! But it seems like the BioShock benchmark of Phoronix wasn't using radeonsi but dxvk... so the benchmark run of Phoronix was wrong in this case?

Quote from Mike's blog: "On RadeonSI I was unable to repro anything remotely approaching the figures above—at most, I was able to get about half the claimed FPS. When switching to the Proton+DXVK mode, however, that’s around the performance I see."

[rmfx]

Is vk extension shader_object being used in the benchmarks ?

If not because radv gets it later in mesa 23.2, I will be curious to see the difference it provides for Zink.

[pinguinpc]

Is vk extension shader_object being used in the benchmarks ?

If not because radv gets it later in mesa 23.2, I will be curious to see the difference it provides for Zink.

however ext_shader_object extension is only avalaible on lvp aka lavapipe until dont stay on anv or radv, maybe nvidia propietary blob for disgrace must be have support now

[emblemparade]

I have decided that every time I see a post about Mike Blumenkrantz I'm going to respond with an adoring fan comment. Here goes this one: Mike for President. Of ... graphic APIs.

[Amaranth]

however ext_shader_object extension is only avalaible on lvp aka lavapipe until dont stay on and or radv, maybe nvidia propietary blob for disgrace must be have support now

You can use this today on Intel and AMD drivers if you use this extension layer https://github.com/KhronosGroup/Vulk...bject_layer.md in your app. It won't be as efficient as having it implemented natively but I suspect the mesa implementation will end up being similar to this, they like to build things in terms of other features so they can be shared between all drivers.