Is it safe to pass a pointer of a non-contiguous array to a Lapack95 routine in Fortran?

Question

I would like to call the syevd Lapack95 subroutine for a non-contigous array as follows:

real :: mat(15000, 15000), vec(15000)
mat=1.d0
associate(eig_vects=>mat(:10000,:10000),eig_vals=>vec(:10000))
   call syevd(eig_vects,eig_evals,'V')
end associate

Is it a safe thing to do this? Earlier I experienced problems with passing non-contigous array slices to a subroutine (the temporary array created by that subroutine could not be allocated). May I expect similar issues when using the eig_vects pointer as an argument?

Use tag fortran for all Fortran questions. Only add version tags to version-specific questions. Do the same for all programming languages. There are very few, who follow the specific version tags. — Vladimir F Героям слава
– Vladimir F Героям слава, Commented Nov 18, 2024 at 6:36
Strange, this is from yesterday? Is it yours? devcodef1.com/news/1427962/… To be fair, I really do not see, what the associate is supposed to achieve, that passing just the subarray doesn't. A temporary array must be created by the compiler in any case, if the dummy argyment is assumed-size. — Vladimir F Героям слава
– Vladimir F Героям слава, Commented Nov 18, 2024 at 6:43
Note also that eig_vects is no more contiguous than mat(:10000,:10000) is. (That is, it's not at all contiguous.) — francescalus
– francescalus, Commented Nov 18, 2024 at 10:59
It is possible that some robot scrapes new questions here and creates some chatbot-based articles about them. — Vladimir F Героям слава
– Vladimir F Героям слава, Commented Nov 18, 2024 at 18:37

PierU · Accepted Answer · 2024-11-18 20:56:47Z

1

Lapack95 is not an independant implementation of Lapack, but rather a wrapper to the the classical Lapack, with a more modern interface. In particular, the dummy array arguments of Lapack95 are assumed-shape (which can be discontiguous), whereas they are assumed-size in classical Lapack (hence contiguous).

Since the end of the call chain (classical Lapack) expects contiguous arrays, a temporary copy has to be made anyway if you want to process discontiguous arrays. The only question is "where":

associate( eig_vects=>mat(:10000,:10000) ) doesn't make any copy, eigs_vects is just a kind of alias here
call syevd(eig_vects,eig_evals,'V') doesn't make any copy either, as the interface of the Lapack95 syevd can handle a discontiguous array.
So, the temporary copy will likely happen on the inner call to the Lapack routine *syevd; but you have no control on how it is performed. For instance, typically the Intel compiler uses the stack for temporary arrays, and it will fails with such a large matrix.

As suggested by @VladimirFГероямслава , you'd better explicitly allocate a contiguous array to work with once for all:

real, allocatable :: eigs_vect(:,:)
allocate( eigs_vect, source=mat(:10000,:10000) )

Note that vec(:10000) is contiguous, so you don't need allocating another array for it.

Note that you could also directly call the classical Lapack routines, as their interfaces can handle this kind of situation without any copy, thanks the decoupling because the size of the problem and of the "leading dimension":

call ssyevd( 'V', 'U', 10000, mat, 15000, vec & 
           , work, lwork, iwork, liwork, stat )

answered Nov 18, 2024 at 20:56

PierU

2,7371 gold badge5 silver badges17 bronze badges

Sign up to request clarification or add additional context in comments.

1 Comment

Vladimir F Героям слава Over a year ago

I think I was wrong and the Lapack95 subroutine will call the ssyevd as you show at the end of your answer, without any temporary copy. Copies may be needed for other calls, but not here.

Collectives™ on Stack Overflow

Is it safe to pass a pointer of a non-contiguous array to a Lapack95 routine in Fortran?

1 Answer 1

1 Comment

Your Answer

Hot Network Questions

Collectives™ on Stack Overflow

1 Answer 1

1 Comment

Your Answer

Sign up or log in

Post as a guest

Related