Unexpected behavior for OpenMP collapse clause

[chrisdaley]

Hi all,

Could someone please confirm whether I have found a compiler bug for
the OpenMP collapse clause?

I have attached two test programs that use the OpenMP collapse clause
to divide iterations from two loops. I find that the OpenMP collapse
clause works as I expect when the loop limits are hard-coded values
(see program value_limits.F90), but not when I use loop limits that
are determined from an array (see program array_limits.F90) -- the
array containing the loop limits is shared, but is not modified within
the nested loops. When I run the affected program array_limits.F90
the number of executed loop iterations is undefined -- sometimes it is
1 iteration and sometimes it seems to tend to infinity. A subset of
the code from the affected program array_limits.F90 is shown below:

limits(1) = 1
limits(2) = 3
!$omp parallel default(none) private(i,j) shared(limits)
!$omp do collapse(2) schedule(static)
do j = limits(1),limits(2)
do i = limits(1),limits(2)
print *, 'Thread', omp_get_thread_num(), " of", &
omp_get_num_threads(), " has i", i, " and j", j
end do
end do
!$omp end do
!$omp end parallel


I have been using an x86_64 version of Intel compiler version 12.0.2
with Linux Kernel 2.6.18-194.el5.

The attachments are: value_limits.F90, array_limits.F90, Makefile,
README.

Thanks for any feedback,
Chris

[TimP]

My own experience agrees with yours, that the collapse is reliable only for the case where the outer loop count is known at compile time. Thanks for submitting a test case; maybe this can get some attention.

[rreis]

It works fine with gfortran... I mean, it worked.