Re: Problem with OpenMP Reduction Clause

[Benjamin Réveillé <benjamin dot reveille at gmail dot com>]

Le Mardi 2 Mai 2006 18:43, Steve Kargl a écrit :
> On Tue, May 02, 2006 at 05:34:46PM +0200, Benjamin R?veill? wrote:
> > On 5/2/06, Diego Novillo <dnovillo@redhat.com> wrote:
> >>Benjamin R?veill? wrote:
> >>>   integer, dimension(n), intent(out) :: sumarray
> >>>   !
> >>>   integer :: i,j
> >>>   !
> >>>   sumarray(:)=0
> >>
> >>2.8.3.  Restrictions to the reduction clause:
> >>
> >>? Fortran pointers, Cray pointers, assumed-size arrays and allocatable
> >>arrays may not appear in a reduction clause.
> >
> > Yep that is exactly what I also read...
> >
> >>I'm not too familiar with Fortran, but sumarray seems to be an
> >>assumed-size array here?  If so, then we should probably emit a
> >>diagnostic.  In either case, we shouldn't ICE, could you file a PR?
> >
> > I'm not too familiar either... But I tried looking it up...
> > And I thought
> > integer, dimension(*), intent(out) :: sumarray  --> would make
> > sumarray assumed-size
> >
> >
> > c.f. http://www.nsc.liu.se/~boein/f77to90/a9.html
> > ############################################################
> > assumed-size array:
> > The array has a variable dimension, like in Fortran 77, through that
> > both the array name and the parameters of the dimension are dummy
> > arguments, except the last dimension, which is given by a *
> >
> >                SUBROUTINE SUB (A, NA1, NA2)
> >                REAL, DIMENSION (NA1, NA2,*) :: A
> >
> > In Fortran 66 the concept "assumed-size array" was not defined, but it
> > was simulated by placing the digit "1" where the "*" is in Fortran 77.
> > This custom violates the index check and is of course forbidden by
> > modern compilers, for example the NAG Fortran 90 compiler. Many old
> > programs still use this way of simulating dynamic memory allocation.
> > ############################################################
> >
> > What about this other variant:
> > array_reduction2.f90
> > program array_reduction
> >   implicit none
> >   integer, parameter :: n=10,m=1000
> >   integer :: i
> >   !
> >   call foo(n,m)
> >
> > end program array_reduction
> >
> > subroutine foo(n,m)
> >   use omp_lib, only : omp_get_thread_num
> >   implicit none
> >   integer, intent(in) :: n,m
> >   !
> >   integer :: i,j
> >   integer, dimension(n) :: sumarray
> >   !
> >   sumarray(:)=0
> > !$OMP PARALLEL DEFAULT(shared)
> >   print*,'hello from thread ',omp_get_thread_num()
> > !$OMP DO PRIVATE(j,i), REDUCTION(+:sumarray)
> >   do j=1,m
> >      do i=1,n
> >         sumarray(i)=sumarray(i)+i
> >      end do
> >   end do
> > !$OMP END DO
> > !$OMP END PARALLEL
> >   do i=1,n
> >      print*,sumarray(i)
> >   end do
> > end subroutine foo
> >
> > Which passes on intel, xlf, pgi but segfaults on gfortran.
> >
> > Do I file a seperate PR Or put both test cases in the PR ???
>
> File 1 bug report with both test programs.  Note that
> one program should issue an error and the other
> should compile.   You'll probably want to add the
> URL to the head of this thread, and you might want
> to ping Jakub, who knows gfoftran's openmp implementation
> better than anyone else, to ensure that he see the problem.

Hello Steve

Thanks for the answer.
Given Richard's answer (i.e. sumarray is not an assumed-size array) shouldn't 
both compile (as they do with the other compilers) ? If not which one should 
compile and which one should give an error ?

Ben