Bug#961976: mumps: provide 64-bit build

[Drew Parsons]

Source: mumps
Version: 5.7.3-3
Followup-For: Bug #961976
Control: severity -1 wishlist
Control: block -1 by 961108

I tried building full 64-bit mumps packages (mumps-64i, following
labelling used for scotch).

The libraries build without problem with -DINTSIZE64 -DPORD_INTSIZE64
(and -fdefault-integer-8).

But tests fail:

examples$ LD_LIBRARY_PATH=../libmumps64i mpirun -n 2 ./dsimpletest
[sandy][[29665,1],0][../../../../../../opal/mca/btl/tcp/btl_tcp_proc.c:266:mca_btl_tcp_proc_create_interface_graph] Unable to find reachable pairing between local and remote interfaces
[sandy][[29665,1],1][../../../../../../opal/mca/btl/tcp/btl_tcp_proc.c:266:mca_btl_tcp_proc_create_interface_graph] Unable to find reachable pairing between local and remote interfaces
 ** ERROR RETURN ** FROM DMUMPS INFO(1)=   -3
 ** INFO(2)=               0
 ** ERROR RETURN ** FROM DMUMPS INFO(1)=   -3
 ** INFO(2)=               0
At line 29 of file dsimpletest.F (unit = 5, file = 'stdin')
Fortran runtime error: End of file

Error termination. Backtrace:
#0  0x7f56904277d5 in ???
#1  0x7f5690428399 in ???
#2  0x7f5690428f0f in ???
#3  0x7f56906e3400 in ???
#4  0x7f56906dbe3e in ???
#5  0x7f56906dcbba in ???
#6  0x560bf74b8376 in mumps_test
	at ./dsimpletest.F:29
#7  0x560bf74b81a4 in main
	at ./dsimpletest.F:73
--------------------------------------------------------------------------
prterun detected that one or more processes exited with non-zero status,
thus causing the job to be terminated. The first process to do so was:

   Process name: [prterun-sandy-94219 at 1,1]
   Exit code:    2
--------------------------------------------------------------------------


Error return value -3 from DUMPS dsimpletest.F l.29
comes from dmumps_driver.F l.532
checking if JOBMIN and JOBMAX are the same in consecutive runs of
MPI_ALLREDUCE.

C     -------------------------
C     Check if value of JOB is
C     the same on all processes
C     -------------------------
      CALL MPI_ALLREDUCE(JOB,JOBMIN,1,MPI_INTEGER,MPI_MAX,
     &                   id%COMM,IERR)
      CALL MPI_ALLREDUCE(JOB,JOBMAX,1,MPI_INTEGER,MPI_MIN,
     &                   id%COMM,IERR)
      IF ( JOBMIN .NE. JOBMAX ) THEN
        id%INFO(1) = -3.
        id%INFO(2) = JOB
        GOTO 499
      END IF


The check fails indicating JOBMIN and JOBMAX are not the same, when
they should be.

I take this to mean MPI_ALLREDUCE is failing to process 64-bit integers,
that is our MPI implementation (openmpi 5.0.8) cannot handle 64-bit
integers directly (has not been configured for 64-bit integers).

If that's the case then we won't be able to proceed implementation
full 64-bit MUMPS (-DINTSIZE64) until we have a 64-bit MPI library.
The request has been raised in Bug#961108, but doesn't look like it
will be so easy to fulfil.

Hence I'm shifting the bug severity for this INTSIZE64 MUMPS to wishlist.