AMD AOCC 3.2 Helps Squeeze A Bit More Out Of Zen 3

[phoronix]

Phoronix: AMD AOCC 3.2 Helps Squeeze A Bit More Out Of Zen 3

Released earlier this month was AMD's AOCC 3.2 compiler based on LLVM/Clang/Flang that provides optimized support for AMD Zen processors. I've been running some benchmarks of AOCC 3.2 compared to prior AMD Optimizing C/C++ Compiler releases and this newest release has been helpful in squeezing a bit more out of Zen 3.

https://www.phoronix.com/vr.php?view=30792

[Setif]

It would be nice if you make some benchmarks of AOCC/flang vs GCC/gfortran!

[pkese]

This is mostly useless without comparison to baseline gcc and llvm compilers.

[skeevy420]

Wanna know how I know y'all didn't read the article?

On the holiday benchmarking agenda is also a fresh comparison looking at AOCC 3.2 performance against upstream LLVM/Clang and GCC along with the possible Ryzen comparison.

[FireBurn]

I wonder if there's any improvements with big projects like Chromium when built with this

[perpetually high]

I was curious and compiled just a hello world C program and was surprised that the AOCC-compiled binary was a third of the GCC-compiled binary.


-rwxrwxr-x 1 15952 Dec 22 08:50 hello
-rwxrwxr-x 1 5768 Dec 22 08:49 hello-aocc

So mostly all performance improvements with no regressions from Michael's tests and the binaries are smaller. Sounds like a huge win/win.

[jjmcwill2003]

I was curious and compiled just a hello world C program and was surprised that the AOCC-compiled binary was a third of the GCC-compiled binary.


-rwxrwxr-x 1 15952 Dec 22 08:50 hello
-rwxrwxr-x 1 5768 Dec 22 08:49 hello-aocc

So mostly all performance improvements with no regressions from Michael's tests and the binaries are smaller. Sounds like a huge win/win.

If AOCC is based on clang, then you should also try compiling your hello world program with vanilla clang and compare file sizes. The file size differences between AOCC and gcc may come from architectural difference in the gcc runtime vs clang's.

[brucethemoose]

Wanna know how I know y'all didn't read the article?

Indeed, but the point does stand. This is basically a regression test benchmark, where I imagine most Phoronix readers are more interested in the difference between AOCC 3.2 and mainstream llvm 14.