RadeonSI Completes ACO Compiler Support With Mesa 24.0

[phoronix]

Phoronix: RadeonSI Completes ACO Compiler Support With Mesa 24.0

With the newly-started Mesa 24.0 development cycle a very exciting feature landed today... The ACO compiler integration for the RadeonSI Gallium3D driver has appeared to effectively wrap up for optionally making use of this Valve-developed shader compiler as an alternative to the AMDGPU LLVM shader back-end...

https://www.phoronix.com/news/RadeonSI-ACO-Complete

[geearf]

Could there be a time when the LLVM part is ripped off?
Also would there be a use of ACO for OpenCL?

I'm still surprised that AMD did this work, but that's awesome that we get this kind of collaborative effort!

[ms178]

Great to have one less driver to worry about LLVM and its breaking changes. Some ACO performance tuning would be nice, though. In my testing with Unigine Heaven on a 6950XT, ACO was consistently 2-3% slower. That's not earth shattering, but I hoped for some gains instead.

[davidbepo]

neat, will love to see the benchs

[user1]

Great to have one less driver to worry about LLVM and its breaking changes.

AFAIU the ability to build RadeonSi without LLVM is optional. That means all major distros will probably continue to ship Mesa with LLVM and it also doesn't seem that RadeonSi will switch to ACO by default at least any time soon.

[kiffmet]

ms178 It's worth questioning whether Unigine Heaven does even remotely represent a 2023 workload though. Optimizing drivers specifically for benchmarks usually only benefits their very specific corner cases. Are there nowadays even any OpenGL applications, where a 6950XT doesn't yield enough performance? Funnily enough, the Windows version of Heaven run through Wine is faster with OpenGL (up to 40FPS more in 1080p), than with DXVK for my 6800XT btw.

Also, performance tuning does still happen. Raytracing with said card in the "raytracingreflections" demo from Sascha-Willems made a jump from 860FPS (Mesa 23.2.1) to 1015FPS (Mesa 23.3-rc1) in a maximized window for me (a bit less than 1440p res). AMDVLK-Pro yields 1130FPS in that case, so there's still some catching up to do, but that 18% gain is impressive nonetheless. Quake II RTX, Cyberpunk, Control and other real-world, non-synthetic workloads did benefit too.

[Teggs]

AFAIU the ability to build RadeonSi without LLVM is optional. That means all major distros will probably continue to ship Mesa with LLVM and it also doesn't seem that RadeonSi will switch to ACO by default at least any time soon.

Still it will be interesting to have the option. I have three games still using OpenGL, and one of them is easy to load the GPU at 100%. I intend to test these when 24.0 is available and see what happens in amdgpu_top.

[user1]

Still it will be interesting to have the option.

Yeah, of course. I don't play a lot of games these days and most of my games use OpenGL, so I'll definitely try it when Mesa 24.0 is out.

[FireBurn]

After the Perfeto fix was applied, things ran well on my PRIME system, tested with Heaven, Superposition and BioShock Infinite

[yuq825]

There's some problem when using navi3 GPU with aco, I'm trying to fix them. So you may test on pre navi3 GPUs or wait some time for my fix on navi3 GPUs.