Re: 64 bit IEEE Netcdf

> It is a while since we have been in contact - I have not been doing much
> netCDF work.  In fact the current state of my fan code is embarassing.
>
> Per Nyberg (NEC) has installed netCDF on our new NEC SX-4/16 with much
> appreciated help of Glenn Davis.  I have agreed to assist Per with netCDF.
>
> The main remaining task is providing for f77 IEEE-64-bit floating point mode.
> The f77 compiler provides 3 options for REAL: IEEE-32-bit, IEEE-64-bit &
> cray-64-bit.  Per has implemented libraries for IEEE-32-bit & cray-64-bit,
but
> not IEEE-64-bit. I have taken on this task.
> What we want is as follows:
> C int 32 bits
> C long 64 bits
> C float 32 bits
> C double 64 bits
> f77 INTEGER 64 bits
> f77 REAL 64 bits
> f77 DOUBLE 64 bits

I'm not clear on what the real goal is here.

I understand that he wants the calculations done in IEEE 64 bit registers.

- Does he want to simply wish to avoid converting  declarations
 to REAL*8 or DOUBLE PRECISION?

- Does he want to store the results on disk as NC_FLOAT or as NC_DOUBLE?

- Does he require that the netcdf-2 fortran interface for NC_FLOAT take
a REAL argument, even when REAL is 64 bits?


>
> I did a 'configure' as follows (under ksh) which seemed to work OK:
>
> CC='cc -Xa' \
> CFLAGS=-h2 \
> CXX='' \
> FC=f77 \
> FFLAGS='-float0 -ew' \
> ./configure
>
> Then I did a 'make' which produced the following message:
> "ncx.c", line 1119: warning: #error detected: "ix_float implementation"
> This comes from following section of code:
>
> #elif _SX && _FLOAT2
> ...
> #else
> #error "ix_float implementation"
> #endif
>
> But it is not clear to me what preceding #if commences this cpp conditional.
> In fact I suspect it may be missing!

ncx.c is built from ncx.m4.
The sequence of cases in ncx.m4 is as follows...

#if X_SIZEOF_FLOAT == SIZEOF_FLOAT && !defined(NO_IEEE_FLOAT)
 ... the usual case
#elif vax
 ... vax float
#elif defined(_CRAY)
 ... cray architectures
 nested
   #if defined(NO_IEEE_FLOAT)
    ... Cray floating point
   #else
   ... IEEE Cray with only doubles
   #endif
#elif _SX && _FLOAT2
 SX computer pretending to be a cray
#else
#error "ix_float implementation"
#endif

I'm sorry this is so convuluted, it is a messy world.
It could be better organized, especially since there is a
cray specific file ncx_cray which works better for non IEEE
crays.

The reason you are getting stumped is that you are building in
a mode where you are
 IEEE float (NO_IEEE_FLOAT is undefined)
 X_SIZEOF_FLOAT != SIZEOF_FLOAT (you only have 8 byte IEEE numbers
and you are not a CRAY.
Given the CFLAGS and FFLAGS you present above, I think there must be
some FLMOD enviroment variable in force that is hosing you up.
The default C build on that machine (float0) should give you what you
want in the C lib without error. I think you can force the issue
by explictly saying float0 to the C compiler.

I think that configuration for the FORTRAN interface will handle the case where
a FORTRAN REAL is 8 bytes and the C 'float' remains 4 bytes.
You should be able to get what you want without changing any code.
Just be sure that float0 is the mode in force when you configure and
compile the C code.
Adding the -ew as you have done to the FFLAGS and the FORTRAN configuration
should figure out to use the C double to hold FORTRAN REAL.

Optimization (vectorization) is more interesting.
If conversion between (64 bit) REAL and NC_FLOAT is desired,
ncx_putn_float_double() and ncx_getn_float_double() will be the
critical sections, and you will probably want to do a vector implementation
with the appropriate pragma or whatever.
If converting between REAL and NC_DOUBLE is desired.
ncx_putn_double_double() and ncx_getn_double_double() will be critical.


> I do not know how the netCDF version 3 Fortran interface works.  What library
> code is involved?

Like the new nc_ C library routines, there are NF_ FORTRAN library routines
which are type specific. These are defined in src/fortran/*.c in terms of the
C routines using an excessive amount of C preprocessor magic.
The old netcdf-2 fortran routines are also defined in
src/fortran/fort-v2compat.c.

Hope this helps.

If you are getting paid excessively, I want a cut. :-).
(I have suffered greatly making netcdf work on an SX4 in
Houston for some NCAR climate guys.)

BTW, what did you think about the proposed performance tuning
interfaces? I sent some email about it in October and didn't hear
back from you.

-glenn