20000802: netCDF use on SGI Origin 2000 with -i8 -r8

Stephen,

>Date: Tue, 1 Aug 2000 22:43:38 GMT
>From: John Sheldon <address@hidden>
>Organization: UCAR/Unidata
>To: "Stephen Klein" <address@hidden>
>Subject: Re: Steve Emerson's email address
>Keywords: 200008012242.e71MgqT02019

The above message contained the following:

> > I am trying to contact Steve Emerson at Unidata. However, when
> > I send e-mail to address@hidden, the message gets
> > repulsed with the transcript detailed below.  Do I have the
> > correct e-mail address?

You had the correct destination address -- but the wrong sending
address.  I've configured my email system to reject everything from the
hotmail.com domain (except for those on a short list) because 1) the
overwhelming majority of email to me from that domain is spam; and 2)
users should first contact <address@hidden>.

I've added your hotmail address to my list of acceptable senders, so you
can contact me that way now.

> > >My name is Steve Klein and I am from GFDL.  Currently though
> > >I am at NCAR visiting CGD. It was suggested to me by John
> > >Sheldon that you amight have some ideas about the difficulty
> > >I am having invoking netcdf routines on dataproc.
> > >
> > >I am trying to run a set of f90 modules some of which use netcdf
> > >routines installed on the SGI Origin 2000 named dataproc
> > >(which is maintained by SCD).
> > >
> > >Currently when I run the executable the system crashes on a
> > >NF_DEF_VAR command. I have printed out the input to this
> > >command, and my computer support people from GFDL do not
> > >see anything wrong with the input data.  When the
> > >run crashes the executable returns 'Unknown error' as
> > >the netcdf error.  Furthermore trying to use dbx on
> > >the core file crashes dbx.
> > >
> > >I do have a guess that the problem might be related to
> > >the use of 32 vs. 64 bit variables and arithmetic on
> > >dataproc.
> > >
> > >When I compiled the code, I used the -i8 -r8 options
> > >because the f90 code used overloaded interfaces some
> > >which took as arguments floats or integers of double
> > >kind.  Compiling the model without these options fails
> > >as the compiler can't find a specific match for the
> > >overloaded interface routines.  I also added a -default64
> > >option to the f90 so that the model would generate
> > >64 bit objects. The reason that I feel that this may
> > >be an issue is that the netcdf library on dataproc
> > >is stored in a directory /usr/local/lib64/r4i4 which
> > >implies to me that it uses 4byte variables for reals
> > >and integers (hence r4i4).  Furthermore the documentation
> > >pages for the use of netcdf on dataproc indicate some
> > >difficulty for the use of -i8 and -r8 with the
> > >netcdf library.

Sounds like you correctly diagnosed the problem.

The netCDF package supports 64-bit floating-point values but doesn't
support 64-bit integer values *at all*, so there's no possibility of
passing INTEGER*8 arguments through the netCDF API.  It seems to me that
the alternatives are the following:

    1.  Modify the program so that it doesn't pass any 64-bit integers to
        any netCDF routine.

    2.  Enhance the program so that it can be compiled "-i4" (and
        possibly "-r4").

Regards,
Steve Emmerson   <http://www.unidata.ucar.edu>