use -march=native rather than -xHost for Intel oneAPI compilers >= 2025.0

[boegel]

fixes #4744

[Micket]

I have been using march=native for intel-2024a, but mostly tested on zen4, where things seem fine. I guess we should build some random crap on icelake nodes and compare the two flags.

I'll ask my colleageus to help try it out.

[Micket]

I tried some miscellenous code from debian language shootout game.

fannkuchredux fortran - no difference
spectralnorm c - no difference
pidigits fortan - no difference
nbody c++ - no difference
spectralnorm fortran - 3x difference. I don't know what's happening with that one case. https://benchmarksgame-team.pages.debian.net/benchmarksgame/program/spectralnorm-ifx-3.html

That test actually becomes better when I don't even specify -march at all (but -xHost is better)

[ohmanm@alvis2 shootout]$ ifx -O3 -ipo -qopenmp spectralnorm.ifx-3.f90  -o spectralnorm.ifx-3.ifx_run
[ohmanm@alvis2 shootout]$ time ./spectralnorm.ifx-3.ifx_run2 10000
1.274224153

real    0m0.701s
user    0m6.277s
sys     0m0.205s
[ohmanm@alvis2 shootout]$ ifx -O3 -march=native -ipo -qopenmp spectralnorm.ifx-3.f90  -o spectralnorm.ifx-3.ifx_run
[ohmanm@alvis2 shootout]$ time ./spectralnorm.ifx-3.ifx_run 10000
1.274224153

real    0m1.793s
user    0m17.510s
sys     0m0.220s

Ugh, why couldn't this just be simple.

edit: This was with intel/2024a on Intel Ice Lake + Skylake. On AMD Zen4, binaries built with -xHOST didn't work at all (it does run when built with -march=native)

[boegel]

This result is reason enough for me to put this on hold, and keep using -xHOST for now, until we learn more...

[Micket]

Just some testing some more, I think the test examples i ran from shootout aren't that great for testing this. Many of them aren't really affected by advanced instructions at all, and when they are, it seems to me that xHost wins.

Maybe the -march=... flags are just using llvms optimization passes, while xhost does the "intel specific magic"? They are definitely not doing the same code gen.
But the counter example where it fails hard is basically just some trivial dot-product loops. And -march=native there seems to be such a crappy job that it's worse than -march=generic, so something is going horrible wrong with spectralnorm example when using Fortran.

But to be clear xHost is not viable option on non-intel CPUs (the binaries wont run).
Those with mixed systems should probably consider -march=... -ax=.... as that would make a fat binary that dispatches on intel hardware.

I wonder if we could just set this conditionally? I looked for a nice clean way to detect if it's a amd system, but I found nothing useful.

The base branch was changed.