Valve's ACO Shader Compiler Under Review For The Mesa Radeon Vulkan Driver

[phoronix]

Phoronix: Valve's ACO Shader Compiler Under Review For The Mesa Radeon Vulkan Driver

The RADV "ACO" shader compiler announced by Valve back in July for the fastest compilation speeds and best possible code generation may soon be hitting mainline Mesa for the open-source AMD Linux graphics stack...

http://www.phoronix.com/scan.php?page=news_item&px=RADV-ACO-Mesa-Merge-Request

[skeevy420]

Very nice. This is one of those milestone days that I've been waiting for. ACO runs the games I've been playing recently very well.

Looks like it's time to update LLVM and Mesa

[ms178]

Very nice. This is one of those milestone days that I've been waiting for. ACO runs the games I've been playing recently very well.

Looks like it's time to update LLVM and Mesa

Soon you don't even need to update LLVM anymore, at least when ACO is feature complete.

[skeevy420]

Soon you don't even need to update LLVM anymore, at least when ACO is feature complete.

If I were just going from mesa master to master a week later I normally don't bother with LLVM, but mesa-master to mesa-aco-master...I suppose I should...

The past week I've been trying 19.3 out with various LLVM versions to see how it was coming along but, overall, it isn't as good as mesa-aco is with Hitman 2 regardless if I'm using LLVM 9 or 10 with 19.3.

[ms178]

If I were just going from mesa master to master a week later I normally don't bother with LLVM, but mesa-master to mesa-aco-master...I suppose I should...

The past week I've been trying 19.3 out with various LLVM versions to see how it was coming along but, overall, it isn't as good as mesa-aco is with Hitman 2 regardless if I'm using LLVM 9 or 10 with 19.3.

My comment was hinting at that ACO is the replacement for LLVM (at least for RADV, possibly radeonSI later). But I guess LLVM will remain a dependency for Mesa. And managing the different versions can be quite a headache. My last journey on my Ryzen/Vega system was plagued with a hard lock as described here, but this was probably due to SDMA issues: https://bugs.freedesktop.org/show_bug.cgi?id=109955

Performance in my games is still better on Windows in most of my games or cannot run on Linux currently (due to Origin). But the experience in Hearts of Iron IV is better on Linux as my 144 Hz monitor flickers in Windows. I still got hope to make the switch on my main system somewhen in the not too distant future...

[brent]

Yeah, the biggest advantage of ACO isn't performance of the compiled shaders or lower compilation time. It's breaking the dependency from LLVM! While Mesa will still use LLVM, it won't be as critical for AMDGPU as it is now. With ACO, we don't need to wait half a year for some compiler backend bugfix anymore.

I'm still convinced that LLVM needs a severly revised release engineering process, though.

[pal666]

Yeah, the biggest advantage of ACO isn't performance of the compiled shaders or lower compilation time. It's breaking the dependency from LLVM!

but that only indirectly affects end users. unlike performance

[brent]

but that only indirectly affects end users. unlike performance

No, of course it does directly affect users. Bugfixes can be shipped much faster to users and with less bureaucracy.

[R41N3R]

Since ACO is out I use it and it works like a charm. I'm really impressed so far and I hope to see it merged to Mesa as soon possible. To depend less from LLVM is greatly appreciated too, it broke Mesa-git several times!

[timofonic]

Please don't merge it until NIR improvements are made and can be used with OpenGL too.

Any news about Intel and others doing an effort equivalent to ACO too? Please make it happen, LLVM sucks because of TOO MANY REASONS.