GCC 12 Compiler Optimization Tuning With The AMD Ryzen Threadripper 3990X

[phoronix]

Phoronix: GCC 12 Compiler Optimization Tuning With The AMD Ryzen Threadripper 3990X

Following the recent discussions about -O3'ing the Linux kernel and other compiler optimizations, a request came in to see some fresh GCC compiler optimization benchmarks with the recently released GCC 12. So here is a fresh look at various GCC optimization levels up through -Ofast as well as with link-time optimizations (LTO) and "-march=native" tuning on the new GCC 12 with the mature AMD Ryzen Threadripper 3990X platform.

https://www.phoronix.com/scan.php?page=article&item=gcc12-optimize-threadripper&num=1

[birdie]

Michael

Would be nice to see the Geometric Mean Of All Test Results not as numbers but as percentages, e.g. 100% for -O2 and then the others as improvements or regressions over -O2 (100%). Makes understanding the results a lot easier than performing mental gymnastics. Thank you.

[skeevy420]

Damn. So -Ofast really do go brrrrrttttttt

In other news it's nice to know that I've probably been doing it wrong being conservative with -O2 -march=znver2 with my Zen 2

[V1tol]

Damn. So -Ofast really do go brrrrrttttttt

-Ofast is not really safe. I tried it with something safe and compiled dxvk with it - immediately received glitched graphics and GPU hang . So -O3 -march=znver3 -flto=auto is my option here

[Markopolo]

as others have noted, Ofast is not safe. GCC specifically states it enables optimizations that may not be safe. for example it enables fast math and allow-store-data-races which are the type of optimizations that may break things but only be visible under extreme load / rare edge cases.

I might have missed it but you should probably caution users about that before showing plots with it at the top…

that O3 consistently beats O2 with native was a welcome surprise to me. Thanks for all the benchmarks!

[skeevy420]

-Ofast is not really safe. I tried it with something safe and compiled dxvk with it - immediately received glitched graphics and GPU hang . So -O3 -march=znver3 -flto=auto is my option here

Not a surprise whatsoever. I'm probably gonna bump it back to -O3 by default after seeing these results.

Michael That post reminded me that there was an AOCL release a week and a half ago that I don't think Phoronix covered.

[skeevy420]

that O3 consistently beats O2 with native was a welcome surprise to me. Thanks for all the benchmarks!

Ditto.

[SigHunter]

but what about the bigger binaries and the buffer bloat ppl where whining about?

[GonVas]

Great article, is it possible to test with PGO? Really interested how much faster it is.

[CochainComplex]

Not a surprise whatsoever. I'm probably gonna bump it back to -O3 by default after seeing these results.

Michael That post reminded me that there was an AOCL release a week and a half ago that I don't think Phoronix covered.

I hope aocc 3.3 will follow soon?