Page 1 of 1

lapack f90 routines or not?

PostPosted: Tue Jul 04, 2006 6:04 am
by alexatlapack
Hi all,
we actually revise our dft simulation package code and try to get rid of some warnings and pitfalls.
our code is written in f90.
for some compilers it has been discussed, that they eventually make a copy of some matrices when calling lapack.
my question now is: does there exist some f90 interface which prevents from that? does a stable f90 interface exist at all? (this could be interesting in order to get a consistent code....)

thanks for any hint
alex

PostPosted: Tue Jul 04, 2006 11:49 am
by Piotr Luszczek
You can have a look at LAPACK95 and LAPACK3E. But from your
perspective they will be just thin layers on top of old LAPACK.

The problem with any kind of solution is the lack of BLAS interface
that would support Fortran 90. Imagine that the whole LAPACK was
rewritten in Fortran 90. As unlikely as it may be, it still doesn't
solve your problem. You potentially avoid a copy when calling
LAPACK but LAPACK will be making copies when calling BLAS (old
BLAS).

Until there is BLAS in Fortran 90 there is no point in having fancy
LAPACK interface. It's just better to make one copy when you call
LAPACK then to make multiple copies when you call BLAS inside LAPACK.

And considering the fact that it took some 30 years to make Fotran 77
BLAS a widely accepted standard, I would not bet much money on
having Fortran 90 BLAS any time soon.

PostPosted: Tue Jul 04, 2006 11:53 am
by alexatlapack
... that sounds reasonable...