AMDVLK 2021.Q2.5 Released As A Minor Radeon Vulkan Driver Update

[phoronix]

Phoronix: AMDVLK 2021.Q2.5 Released As A Minor Radeon Vulkan Driver Update

Following the last AMDVLK update at the end of May, AMD has released AMDVLK 2021.Q2.4 as their newest open-source AMD Radeon Vulkan driver release...

https://www.phoronix.com/scan.php?pa...DVLK-2021.Q2.5

[uid313]

Why do AMD even have proprietary Radeon drivers? Why don't they just use open source for everything?
Why don't they just cease development on the proprietary driver and put full focus on the open source driver?

[aufkrawall]

I'd rather see them abandon amdvlk-open and contribute to RADV instead. There seems to be a ton of work to be done for raytracing by Mesa Intel and RADV devs. amdvlk-open development is painfully slow too, it also never lived up to its expectations. Stadia could use -pro driver or also RADV instead. It's so sad that AMD don't see how they hurt not only Mesa and their users, but also themselves as a company...

[mphuZ]

It's so sad that AMD don't see how they hurt not only Mesa and their users, but also themselves as a company...

LOL. How?

[aufkrawall]

LOL. How?

Because amdvlk-open is useless in 99% of the time for users, it regularly misses features/extensions/bugfixes for years and RADV devs are apparently left alone for developing raytracing support?

[user1]

I've noticed that the some of the last few versions of AMDVLK finally added support for some extensions that RADV and other drivers already supoprted for a long time (some over a year ago!). It also seems to me that almost all of those extensions that AMDVLK is late supporting are by Valve.

I think I said it sometime ago, but in 2018 AMD said they have a plan to phase out the proprietary shader compiler, so that LLVM will replace it on all AMD drivers on both Windows and Linux. Maybe that's why they keep developing AMDVLK-open? But then again, the compile time performance of LLVM has barely improved over the last 3 years, so I doubt it will even happen.

[Seebi]

But then again, the compile time performance of LLVM has barely improved over the last 3 years, so I doubt it will even happen.

I heard it has improved quite a bit recently, at least I saw something mentioned in the release notes for 2021.Q2.3.

Improving the compile time of LLVM should also help the OpenGL driver in mesa and OpenCL, so it’s not getting wasted either way.

[user1]

I heard it has improved quite a bit recently, at least I saw something mentioned in the release notes for 2021.Q2.3.

Improving the compile time of LLVM should also help the OpenGL driver in mesa and OpenCL, so it’s not getting wasted either way.

Yeah, I know, The question is how big is the improvement compared to the other 2 compilers. Last time when I tested all the drivers in late 2020, between LLVM and ACO, the compile time performance of the proprietary compiler was much much closer to ACO.

RadeonSI will eventually get ACO as well, so I think in the future when ACO will be mature on RadeonSI, LLVM will become more relevant for compute. Btw, I heard that LLVM is mainly developed by the compute team at AMD.

[clapbr]

RadeonSI will eventually get ACO as well.

Isn't ACO default already in Mesa radeonsi?

[user1]

Isn't ACO default already in Mesa radeonsi?

No, ACO support in RadeonSI is still early wip.
ACO is default in RADV since Mesa 20.2.