[MITgcm-support] MNC oddities on Cray T3E

[Mark Hadfield]

Thanks to Ed Hill and Constantinos Evangelinos for their responses 
(though I am a little hurt by Ed's dissing of our T3E).

I take Constantinos' point about the desirability of doing type 
conversions explicitly. Also, I have realised since posting the original 
message that MNC_CW_RL_W_OFFSET includes some code to handle missing 
values and that presumably should be type sensitive.

However on further reflection I think that there is one (and only one) 
correct & robust way to declare the variables to be passed to the 
NF_PUT_* functions and that is to match the declarations of those 
functions in the netCDF library, eg:

      DOUBLEPRECISION  resh_d( MNC_MAX_BUFF )
      REAL  resh_r( MNC_MAX_BUFF )
      INTEGER resh_i( MNC_MAX_BUFF )

This avoids the need to worry about whether REAL is REAL*4 or REAL*8, 
etc and avoids the need for #ifdefs. It still means that we might be 
passing 8-byte reals into a netCDF function that will write them to a 
netCDF FLOAT variable, but that is unavoidable.

Constantinos suggested I could use the preprocessor to #define REAL*4 to 
REAL*8. I think this would break the ability to read 32-bit data via 
MDSIO. There's also the mismatch between Fortran case-insensitivity and 
cpp case-sensitivity.

RE the possibility of rebuilding the netCDF library on the Cray T3E to 
make 32-bit the default  precision, I did manage to do this, but it 
involves changing 8 or so files and there are still some rough edges. 
(The problem is not with the real numbers but with the integers, as a 
Fortran INTEGER now matches a C short, a possibility that the netCDF 
build system has not considered.) So I think this is a second-best solution.

-- 
Mark Hadfield          "Ka puwaha te tai nei, Hoea tahi tatou"
m.hadfield at niwa.co.nz
National Institute for Water and Atmospheric Research (NIWA)