Re: netCDF

> Date: Mon, 18 Nov 1996 16:43:22 -0800
> From: address@hidden (P.J.Flatau)
> To: address@hidden
> Subject: netCDF

Hi Piotr,

> I just wanted to use NetCDF to write  complex array. I have noticed
> your recent reply to similar question. 
> 
> Clearly, there are applications where complex numbers are important
> (take for example spectral GCM models which uses complex FFT).
> 
> Anyway, what is your prescription for writting complex numbers.
> 
> I guess the question about writting complex number reduces to
> how do I write 2 real arrays, i.e. how do I skip every second 
> element. If there was  istart, icount, iskip --- it would be easy.

Well, one way to do it is to just define an n by 2 netCDF variable of
type real:

    netcdf oct {
    dimensions:
            n = 8 ;
            ri = 2 ;          // real (1) or imaginary (2) part
    variables:
            float c(n, ri) ;  // eighth roots of 1 in complex plane

    data:

     c =
      1, 0,
      0.7071068, 0.7071068,
      0, 1,
      -0.7071068, 0.7071068,
      -1, 0,
      -0.7071068, -0.7071068,
      0, -1,
      -0.7071068, 0.7071068 ;
    }

From FORTRAN, you should just be able to access complex arrays directly
this way, without any kind of skipping, since the storage order for a
complex array is with the real and imaginary parts of each value in
successive locations.  Thus from FORTRAN with the above structure, you
should be able to read all eight complex values into a single complex
array with one call to NCVGT:

      COMPLEX C(8)
      INTEGER NCID, VARID, RCODE
      ...
      CALL NCVGT(NCID, VARID, 1, 8, C, RCODE)

and similarly for higher-dimensional arrays.

If you needed to deal with many complex arrays this way, it would
probably be a good idea to adopt a conventional variable attribute to
identify the variables as complex, so that programmers in C or other
languages with no complex type could understand your intent, e.g.

        float:type = "complex";

If you want to skip every second element and just read the real parts,
for example, use NCVGTG and specify the stride as 2.

> More general comment: why not provide some simple
> FORTRAN routines which would elimiminate need to use
> "id" variables  in the same way as it is done in IDL
> (i.e. using character name). It is difficult to use NetCDF 
> as it is.
> 
> E.g.
> 
> ncdf_varget,idnc,'wavea',wavea

Good question.  We had thought about doing it this way, and the original
HDF SDS interface used variable names instead of variable IDs, for
example.

The reason we don't do this in the FORTRAN and C interfaces is
efficiency.  If you use long variable names and have lots of variables,
each I/O access would require looking up an offset by variable name,
which even in the best case (hashing on the variable string and using
the hash index for direct lookup) would require extra time proportional
to the length of the variable name string on every data access.  With a
less sophisticated lookup, it would require additional time proportional
to the number of variables, or the log of the number of variables if
they are kept sorted.  This extra time could be significant lots of
small accesses.

In the C++ interface and the upcoming Java interface, no variable ID is
required, since each variable is just an object and the variable
referring to the object can be used directly.

--Russ

_____________________________________________________________________

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