Radeon RADV "ACO" Performance On Mesa 19.3 Looking Good

[ntropy]

Aco 19.3 on 5.4 kernel is an explosion of frames on the Vega: https://youtu.be/0-kEjQTRkUM are amd LRU bulk moves enabled yet?

[shmerl]

AMD should accept this compiler in their AMDVLK and radeonsi.
Otherwise ACO will be useless.

radv / aco focus on gaming use case most. amdvlk might have more compute requirements not covered by aco. Just a guess.

[nuetzel]

And mesa master branch is not even containing all the available improvements. As it was mentioned in the merge forum thread:


Anyway, great to see that even that, it managed to obtain better performance.
Thank you for the testing, Michael!
And thank you Valve & Co. for your efforts, much appreciated!

Thank you Daniel (Bas and Rhys)!!!

[agd5f]

AMD should accept this compiler in their AMDVLK and radeonsi.

I wonder how much improved LLVM would be if the same effort that was expended to write a new compiler from scratch had been put into LLVM instead.

[tildearrow]

From https://github.com/RadeonOpenCompute/ROCm/wiki it seems not.

Therefore it is a typo.

[tuxayo]

Thanks for these benchmarks. If I understand correctly, the NIR patches aren't used here right?

Should we also expect more performance from them?

[castlefox]

I probably missed it but why is there a lack of Navi support with this??

[LeJimster]

I wonder how much improved LLVM would be if the same effort that was expended to write a new compiler from scratch had been put into LLVM instead.

I'm certainly no expert, but wasn't the reasoning that LLVM is such a large project, that to optimize it for one thing without breaking it for other use cases would have required much more time and resources. In the end they decided (rightly or wrongly) to create something new with one focus.

I sometimes wonder what could have been if AMDVLK had been open sourced a year or so earlier. Seems we have a lot of projects doing the same thing.

[bridgman]

From https://github.com/RadeonOpenCompute/ROCm/wiki it seems not.

GFX8 also includes Tonga, Carrizo, Stony Ridge and Iceland/Topaz (small laptop only part)

[smitty3268]

I wonder how much improved LLVM would be if the same effort that was expended to write a new compiler from scratch had been put into LLVM instead.

Well, I guess we'll never know for sure, but the ACO devs seem to think not much. They're arguing that doing a whole new compiler was easier than getting changes into LLVM.