Unresolved references to mpi_wtime_ building mpich 1.2.5 with lf95

[wootletootle]

A similar question (brought up by R.T) received
the following reply from Lahey Support:

>These unresolved symbols occur when the >mpich install is running test programs to see if >everything works. The problem seems to occur >because the configure program is unable to >figure out that these functions are not present.
>To resolve the problem, edit the Fortran include >files mpif.h and mpef.h in the ~/include directory >and comment out or remove references to these >functions.

However it doesn't seem that simple to me.
There are references to mpi_wtime and the
other symbols right throughout the code. It is certainly not sufficient to comment out a few lines in the include files.

Could Lahey Support comment further on this
topic? Is there an alternative resolution to the problem?

It isn't just the mpich test programs
which complain, the symbols are also undefined
in the mpich libraries and I am unable to
compile my application program (GeoFEM).

Thanks for your help,
Roger Young
r.young@irl.cri.nz

[tzeis]

Hmmm. I am not sure what is going on with your installation. Here are steps I took, which seemed to be successful.
1) Ran the configure script for mpich-1.2.5 using the command:
./configure -f90=lf95 -fc=lf95 -f90linker=lf95
2) Ran make for mpich. During the course of the make, saw unresolved references for mpe_wtime, but ignored them.
3)In the geofem-v6.0.1 directory copied the config.h file for linux-ix86 from the config directory.
4) In numerous geofem makefiles, changed the value for MPIDIR and INSTDIR to point to the installation directories. The value for F90 should be mpif90.
The program appeared to make correctly. I was able to get the test programs to the point where they would compile and run, which indicates that the link problem was not occurring.

[tleger]

I have a similar problem as wootletootle, except with two other (research) codes that use mpi_wtime to time portions of their execution. When I built mpich-1.2.5 I used the following configuration line:

./configure -c++=gcc -cc=gcc -fc=lf95 -f90=lf95 --enable-traceback --prefix=/research/mpich-1.2.5-lf95

This is for a beowulf cluster running Mandrake 8.2 linux and we just switched to using Lahey from Portland Group. While Lahey is much better in key areas, I can say that the Portland Group compilers didn't suffer from this problem with mpi_wtime in mpich-1.2.5.

Regards, Tim Leger

[wootletootle]

Thanks for your replies to this thread.
However I have to say that I have not
yet succeeded in compiling my application.
There are 3 elements: lf95, mpich, and the application (geofem).

It may be that there are two different
sources of error, mpe_wtime and
mpi_wtime. Regarding the latter I have
received the following information from
Daniel Katz (Daniel.S.Katz@jpl.nasa.gov):

>in src/fortran/src/wtimef.c, change the one >instance of mpi_time to
>mpi_wtime. this is a bug in mpich-1.2.5.2

After making this change the mpi_wtime
error was eliminated (the mpe_wtime
error remained, however I ignored it as suggested by <tzeis>).

The mpich installation then completed
apart from some of the example files
(pi3f90.f90 on line 48 referring to MPI_BCAST).

However the compilation of the application (geofem) did not succeed, lGeoUtil was
not compiled. Tracing back this leads to
the sequence

lGeoUtil --> geofem_arg.f90 -->
geofem_iargc --> farg.f [in mpich]
--> IARGC()

I read somewhere else (Tim Prince) about an IARGC() problem with reference to the Intel compiler. Is there some connection?

What is not clear to me is where the problem lies, is it in the compiler
(lf95), in mpich, or in geofem??

Roger Young
r.young@irl.cri.nz