[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: 971229: question about compiling netCDF on Crays

> >From: address@hidden (Mary Haley)
> >Subject: question about compiling netCDF on Crays
> >Organization: NCAR/SCD
> >Keywords: 199712291733.KAA18810 netCDF 3.3.1 Cray

Hi Mary,

> In your INSTALL notes for building netcdf 3.3.1 on ouray, you mention
> the following:
> 
>     In the following, the "-i64" and "-dp" options are necessary in
>     order to make FORTRAN INTEGER and DOUBLE PRECISION datatypes
>     equivalent to C "int" and "double", respectively.
> 
> Is this required or just recommended? The reason I ask is because a
> user on ouray was trying to write some Fortran double precision
> variables to a netCDF file, but it's not getting the correct results
> presumably because the options above cause double precision to behave
> like C doubles, which would make them 8 bytes and not 16 bytes. Is
> this correct?

NetCDF only supports external representations for IEEE single-precision
(4 bytes) and IEEE double precision (8 bytes) floating-point numbers.
There is no straightforward way to store a 16-byte floating-point number
in a netCDF file, since there is no external data type corresponding to
a floating-point type with that much precision.  It would be possible to
store such a number using type punning as an array of 4 4-byte integers,
for example, but this would require writing a layer on top of netCDF to
make it practical, and no other applications for dealing with netCDF data
would understand this convention.

We used the "-dp" option on SCD's recommendation, since someone up there
told us that's the way the majority of Fortran programs were compiled,
and that no one used the 128-bit double precision types because their
arithmetic was very slow, implemented in software rather than hardware.

> Should I tell the user he can't write double precision variables from
> Fortran, or could I actually recompile the netCDF library on ouray
> without these options and give him the new library for his own use?

You can write double precision variables from Fortran using the netCDF-3
interfaces, but their values will only be stored as IEEE 64-bit floats,
so precision will be lost from Cray's 128-bit floats.

--Russ

_____________________________________________________________________

Russ Rew                                         UCAR Unidata Program
address@hidden                     http://www.unidata.ucar.edu