AMD AOCC 1.2.1 Compiler Flings Flang Fixes

[phoronix]

Phoronix: AMD AOCC 1.2.1 Compiler Flings Flang Fixes

AMD released a minor update to their AMD Optimizing C/C++ Compiler...

http://www.phoronix.com/scan.php?page=news_item&px=AMD-AOCC-1.2.1-Released

[tildearrow]

Typo:

Fortran compiler thath as been in

[celrod]

If you benchmark Flang vs gfortran on Ryzen or Skylake-X, I want to point out that gfortran-8 (and the rest of gcc) prefers 256 and 128 bit vectors respectively, instead of 512 and 256 like LLVM-66 compilers. You can conntrol this with -mprefer-vector-width=
Maybe gcc's preference is faster for general purpose computing, but when it comes to number crunching. Number crunching loves vectorization, and I've found the wider vectors much faster more often than not -- even if avx-512 downclocks, or Ryzen only has 2 128-bit fma units.

[slacka]

I'm positive AMD's AOCC has less than 1% of the market share, while in my sphere, clang is quite popular. So if they want to see Zen performing better, these improvement need to make it into clang.

Is AOCC developed by AMD's clang team? Have any of their past improvement's been upstreamed?

[Royi]

AMD really needs to support OpenBLAS to create a real alternative to Intel MKL.
They also need to create SVML and Intel IPP equivalent which are optimized to Zen.

Those are responsible to significant part of Intel advantage.