LLVM 15.0 Planning For Early September Release - Important For RDNA3 Graphics Support

[phoronix]

Phoronix: LLVM 15.0 Planning For Early September Release - Important For RDNA3 Graphics Support

LLVM release manager Tom Stellard of Red Hat has laid out the planned LLVM/Clang 15.0 release schedule for this next major version of this open-source compiler stack...

https://www.phoronix.com/scan.php?page=news_item&px=LLVM-15.0-Release-Plan

[user1]

I've seen rumors in guru3d forums that apart from the improved OpenGL in early pre-release Windows drivers, they also switched to LLVM from the proprietary compiler, because people were reporting slower shader compilation speeds with these drivers. If that's really the case, it doesn't look good because LLVM compilation speed has been abysmal like forever and I can't imagine it will be improved in a short span of time at least to the level of the proprietary compiler.

[JacekJagosz]

If that's really the case, it doesn't look good because LLVM compilation speed has been abysmal like forever and I can't imagine it will be improved in a short span of time at least to the level of the proprietary compiler.

I saw that too and I am actyally excited. It means less fragmentation and now the same compiler is being used on Windows and Linux with AMD's OpenGL, AMDVLK and ROCm.
Now they will have one compiler to focus on, they will put all their efforts into it. Seems exciting for Linux.

[user1]

I saw that too and I am actyally excited. It means less fragmentation and now the same compiler is being used on Windows and Linux with AMD's OpenGL, AMDVLK and ROCm.
Now they will have one compiler to focus on, they will put all their efforts into it. Seems exciting for Linux.

Personally, I'm very glad that AMD finally started caring about OpenGL performance on Windows after all these years. But I don't see how the move to LLVM on their Windows drivers will affect us Linux users. At least if you only care about gaming, out of the 3 compilers, LLVM, pro and ACO (which is not developed by AMD), LLVM is the slowest and ACO (which is already the default on RADV for almost 2 years) is the fastest. RadeonSi still uses LLVM, but eventually it will switch to ACO as well. So unless you also care about compute stuff, I don't see what's so exciting about LLVM.
Btw, ACO is not only the fastest in shader compilation speed, but it also generates better shader code compared to LLVM and pro compiler, which results in slightly higher fps. This was already noticeable when RADV only started using ACO in 2019 (even when it still didn't support all shader stages).

What I'm most excited to see is how AMD's reworked Windows OpenGL will compare to RadeonSi.

[JacekJagosz]

user1 I am wondering if Pro compiler will get replaced by LLVM as well, same as on Windows. They clearly must have had a good reason to switch to LLVM.
And yeah, I want the compute stuff to get better, any help will be useful.

[Alexmitter]

Of all the possible choices, they chose the worst and slowest. Its not like ACO is not right there, fast as hell...

[user1]

Of all the possible choices, they chose the worst and slowest. Its not like ACO is not right there, fast as hell...

Yeah, just imagine how much better other AMD drivers would be with ACO. But from what I understand, there are 2 main reasons why AMD is not interested in ACO:
1. It doesn't support compute (at least for now), while LLVM does.
2. AMD has already invested heavily in LLVM for years, from way before ACO existed.

[user1]

user1 I am wondering if Pro compiler will get replaced by LLVM as well, same as on Windows. They clearly must have had a good reason to switch to LLVM.
And yeah, I want the compute stuff to get better, any help will be useful.

Actually, I remember someone said in an AMD keynote from 2018 (don't remember what it was exactly, I think it was when they announced AMDVLK-open), that they are interested in phasing out their proprietary compiler at some point, so that LLVM will be used everywhere. So basically like you said, this way they will only have to maintain 1 compiler, so less maintenance burden.
At this point however, no one has confirmed that they really switched to LLVM on WIndows. The slower shader compilation may be just a driver bug, since these are early pre-release drivers. One user however, noted that the shader compiler weights more in the new drivers, so that might be a hint of LLVM.

[JacekJagosz]

Actually, I remember someone said in an AMD keynote from 2018 (don't remember what it was exactly, I think it was when they announced AMDVLK-open), that they are interested in phasing out their proprietary compiler at some point, so that LLVM will be used everywhere. So basically like you said, this way they will only have to maintain 1 compiler, so less maintenance burden.
At this point however, no one has confirmed that they really switched to LLVM on WIndows. The slower shader compilation may be just a driver bug, since these are early pre-release drivers. One user however, noted that the shader compiler weights more in the new drivers, so that might be a hint of LLVM.

It that is true, that they fully switched to one compiler, that is awesome. Now even if you use ACO, its developer will be able to see all the tricks AMD introduces. Especially as RADV and ACO are written based on changes to LLVM, before they get hands one new GPUs.
Oh, and not too long ago building ROCm with vanilla LLVM finally became possible, as AMD is upstreaming all the changes from their fork. I am so happy there finally can be less fragmentation, and you can have one compiler for a lot of apps, GPU compute and some GPU graphics.