[petsc-dev] BuildSystem performance

[Jed Brown]

Jed Brown <jedbrown at mcs.anl.gov> writes:

> Surely we can find it within our heart to confirm that there are no
> missing functions in BlasLapack in less than 12 seconds.  We are testing
> the same few functions many times:
>
>      12 char ddot();
>      12 char dgeev();
>      12 char dgetrs();
>      13 char dgeev_();
>      14 char df2cblaslapack311_id();
>      14 char dgetrs_();
>      25 char ddot_();
>
> This is a system with normal underscore name mangling and -llapack
> -lblas, with no missing functions.  At worst, I would expect a few
> failures with one function, then only one test to confirm that the same
> mangling works for all the others.

Why do we test these first?  Testing these accounts for half of the
BlasLapack run-time, or 6% of the overall run-time.  Can we at least
move the top three to the end because they are so unlikely to appear?

    yield ('IRIX Mathematics library', None, 'libcomplib.sgimath.a', 1)
    yield ('Another IRIX Mathematics library', None, 'libscs.a', 1)
    yield ('Compaq/Alpha Mathematics library', None, 'libcxml.a', 1)
    # IBM ESSL locations
    yield ('IBM ESSL Mathematics library', None, 'libessl.a', 1)
    yield ('IBM ESSL Mathematics library for Blue Gene', None, 'libesslbg.a', 2)


Does it really make sense to look for ESSL this way?  Every machine puts
it in a different place and it often needs other stuff like -lmass.