On Nov 4, 2005, at 11:41 AM, Charles Williams wrote:
I have been using Open MPI in conjunction with PETSc on OSX 10.4,
and had been having trouble with undefined symbols when trying
tests with PETSc:
/usr/bin/ld: Undefined symbols:
_pmpi_wtick__
_pmpi_wtime__
After playing around with things for a while, I realized that these
undefined symbols only appeared in libmpi_f90.a, and were somehow
being pulled into the PETSc libraries when they were built. The
problem occurred when trying to build an executable that included
libraries with these undefined symbols. The simplest solution I
could think of was to edit include/mpif.h.in. In the last two
lines of this file I simply changed:
> double precision MPI_WTIME, MPI_WTICK, PMPI_WTIME, PMPI_WTICK
> external MPI_WTIME, MPI_WTICK, PMPI_WTIME, PMPI_WTICK
to:
< double precision MPI_WTIME, MPI_WTICK
< external MPI_WTIME, MPI_WTICK
This fixed the problem for me, and I am wondering whether this is a
reasonable fix. My understanding is that the functions are not
actually part of open-mpi. Would this approach break anything
else? If not, it would be nice to include it, since the only other
option I can think of for fixing things on OSX 10.4 with the
current build procedures is to strip libraries of undefined symbols
before they are used.
Sorry about that. Some time ago, we decided to make MPI_WTIME and
MPI_WTICK inline functions in C and not provide a profiling interface
for them. Unfortunately, it appears we didn't properly update the
Fortran or C++ interfaces. I committed a change so PMPI_WTICK /
PMPI_WTIME exist and work properly. That should fix your problem.
The change will be in the next 1.0 release candidate, and the nightly
tarballs that will be available tomorrow morning. Release candidates
and betas will be available at the URL below:
http://www.open-mpi.org/software/
Hope this helps,
Brian
--
Brian Barrett
Open MPI developer
http://www.open-mpi.org/
