Problems with PGI and 1.1.1

[David Robertson]

On 5/17/2010 9:54 AM, Gerry Creager wrote:
> Dave
>
> First, PGI 8 should never have seen the light of day. We back-ported to
> 7.2 after a month of sys admin testing. ( was better. 10.3 is good, but
> now 10.5 is out. Keeping up is a pain.

I tried with 10.5 and the tests fail at a different place (ncmpi_inq_var).

I have tried to run nc_test in the totalview debugger. With MPICH and 
PGI 10.5, the segfault occurs at line 638 of src/lib/var.c:

(void) strncpy(name, varp->name->cp, varp->name->nchars);

> I'll have to run the test suite again, but on "real" wrf runs, it
> appears to behave fine.
>
> Are you sure OpenMPI and the mpich flavors were compiled with -fPIC?
> That was a problem I ran into, although it appeared as a compile-time
> error.

All of my MPI implementations are static only so as far as I know -fPIC 
won't do anything.

> gerry
>
>
> David Robertson wrote:
>> Hi,
>>
>> I'm having problems with "make testing" (in nc_test) using PGI. I have
>> tried 3 mpi implementations (MPICH, MPICH2, and Open MPI) and 3
>> versions (8.0, 9.0, and 10.1) of the PGI compiler. All of these
>> compile without errors. However, They all Segfault at different places
>> when I run the testing suite.
>>
>> For example, PGI 10.1 with Open MPI segfaults while testing
>> ncmpi_get_var1_short but 9.0 with Open MPI segfaults while testing
>> ncmpi_get_var_schar. Both the compiler version and MPI implementation
>> seem to affect where nc_test fails.
>>
>> On the same machine I can compile and test with both ifort and
>> gfortran using either MPICH2 or Open MPI. The combination of ifort and
>> MPICH is also successful.
>>
>> Any suggestions would be helpful.
>>
>> Thanks,
>> Dave
>