Isn't this the Lahey/Fujitsu compiler already supported in fcompiler/lahey.py?

@rgommers
Thanks for your reply.

Isn't this the Lahey/Fujitsu compiler already supported in fcompiler/lahey.py?

No, it's different from the lahey/fujitsu compiler.
The new compiler registered is developed by Fujitsu and is used in A64FX on Fugaku. There is currently no HP that introduces this compiler.

@rgommers

Could you add this info to the docstring of the file?

OK, I will add docstring to 'fujitsu.py'.
I'm planning to publish a following article introducing a64fx:
https://www.fujitsu.com/global/products/computing/servers/supercomputer/a64fx/

One question out of interest - if frt going to stay? Based on this article I'd expect gfortran support too, which may then become the default?

As you say, we can use the gcc compiler on Fugaku, but we recommend using the Fujitsu compiler to maximize the performance of fugaku.

I just agree to support fugaku's gfortran in numpy.
I can't determine if a special implementation about gfortran is needed, so I'll ask the person in charge to consider gfortran support.

As you say, we can use the gcc compiler on Fugaku, but we recommend using the Fujitsu compiler to maximize the performance of fugaku.

That makes sense.

I just agree to support fugaku's gfortran in numpy.

Sounds good, thank you.

Thanks for merge!

@t-karatsu this PR added compiler support but performance is still poor because the fujitsu linear algebra library is not detected.

Of course, one can build e.g. openblas with the fujitsu compiler and it get's better, but it would be better to use the fujitsu linear algebra library, no? As I understand it, it's always (?) bundled with the compiler?

The patch in https://gist.github.com/migueldiascosta/3bf319b9eecd888957db50933d2d701f worked for me at Fugaku, and performance is indeed better - is this something you've worked on and/or are interested in testing/reviewing?