AMDVLK 2020.Q3.2 Radeon Vulkan Driver Christened Early Due To Bugs

[phoronix]

Phoronix: AMDVLK 2020.Q3.2 Radeon Vulkan Driver Christened Early Due To Bugs

While it was just two days ago that AMDVLK 2020.Q3.1 debuted and normally there is a two to three week release cadence for these open-source AMD Radeon Vulkan driver code drops, this morning was already met by the debut of AMDVLK 2020.Q3.2...

http://www.phoronix.com/scan.php?page=news_item&px=AMDVLK-2020.Q3.2

[user1]

I think there was a talk here recently about AMDVLK eventually replacing AMDVLK-Pro (LLVM replacing the proprietary compiler on both Windows and Linux). Just want to ask out of curiosity when will that happen eventually? I have tested both drivers with DXVK games and it seems to me that at least in compile time performance the proprietary compiler is still way ahead of LLVM.

[smitty3268]

I think there was a talk here recently about AMDVLK eventually replacing AMDVLK-Pro (LLVM replacing the proprietary compiler on both Windows and Linux). Just want to ask out of curiosity when will that happen eventually? I have tested both drivers with DXVK games and it seems to me that at least in compile time performance the proprietary compiler is still way ahead of LLVM.

I don't think it's happening anytime soon. They've had it as a long term goal for years now. As you saw, the compile performance of LLVM is really bad. That's half of what led to ACO.

Also, you can expect there will always be a pro driver, so it's not getting entirely replaced, just overhauled with some of the code currently restricted to AMDVLK. They use the pro drivers to create validated releases for their professional users. The AMDVLK drivers are just kind of snapshots thrown over the wall every couple weeks.

[user1]

I don't think it's happening anytime soon. They've had it as a long term goal for years now. As you saw, the compile performance of LLVM is really bad. That's half of what led to ACO.

Also, you can expect there will always be a pro driver, so it's not getting entirely replaced, just overhauled with some of the code currently restricted to AMDVLK. They use the pro drivers to create validated releases for their professional users. The AMDVLK drivers are just kind of snapshots thrown over the wall every couple weeks.

Yeah, the first time I heard about this plan was back in 2018. I mean during these 2 years the compile time performance of LLVM definitely improved a little bit, but the progress is very slow because the proprietary compiler is still far ahead of LLVM in compile time performance and I still experience huge amount of stutter in DXVK games with LLVM which makes some games almost unplayable. ACO is even faster. I've seen some AMD developers are not very fond about ACO, they said that instead of reinventing the wheel, those resources could be spent on improving LLVM. Well, I remember reading a Valve post and they said that they had experience with LLVM before ACO, and they came to the conclusion that LLVM has some serious issues that are really hard to fix in order to improve performance and in this case it's just easier developing your own compiler. But I do understand AMD's position in this regard because by the time ACO appeared, AMD has already put a huge amount of effort in LLVM.