mirror of
https://github.com/Unidata/netcdf-c.git
synced 2024-11-27 07:30:33 +08:00
84 lines
2.3 KiB
Plaintext
84 lines
2.3 KiB
Plaintext
#if 0
|
|
$Id: nfconfig.in,v 1.1.1.1 2009/07/06 15:06:29 ed Exp $
|
|
#endif
|
|
|
|
|
|
#ifndef UD_NETCDF_CPP_INC
|
|
#define UD_NETCDF_CPP_INC
|
|
|
|
|
|
#if 0
|
|
Do not have C-style comments in here because this file is processed
|
|
by both the FORTRAN compiler (for the nf_test/ stuff) and the C
|
|
compiler (for the FORTRAN-callable interface routines) and some
|
|
FORTRAN preprocessors do not understand the /*...*/ syntax.
|
|
#endif
|
|
|
|
|
|
#if 0
|
|
The following macros define the supplementary FORTRAN arithmetic
|
|
datatypes beyond the standard INTEGER, REAL, and DOUBLEPRECISION --
|
|
ostensibly corresponding to 8-bit and 16-bit integers, respectively.
|
|
For example:
|
|
|
|
#define NF_INT1_T byte
|
|
#define NF_INT2_T integer*2
|
|
|
|
These are the types of the relevant arguments in the NF_*_INT1() and
|
|
NF_*_INT2() netCDF FORTRAN function calls. The word "ostensibly"
|
|
is used advisedly: on some systems an "integer*2" datatype,
|
|
nevertheless, occupies 64 bits (we are not making this up).
|
|
|
|
If your FORTRAN system does not have the respective supplementary
|
|
datatype, then do not define the corresponding macro.
|
|
#endif
|
|
#undef NF_INT1_T
|
|
#undef NF_INT2_T
|
|
|
|
|
|
#if 0
|
|
Define the following NF_*_IS_C_* macros appropriatly for your system.
|
|
The "INT1", "INT2" and "INT" after the "NF_" refer to the NF_INT1_T
|
|
FORTRAN datatype, the NF_INT2_T FORTRAN datatype, and the INTEGER
|
|
FORTRAN datatype, respectively. If the respective FORTRAN datatype
|
|
does not exist, then do not define the corresponding macro.
|
|
#endif
|
|
#undef NF_INT1_IS_C_SIGNED_CHAR
|
|
#undef NF_INT1_IS_C_SHORT
|
|
#undef NF_INT1_IS_C_INT
|
|
#undef NF_INT1_IS_C_LONG
|
|
#undef NF_INT2_IS_C_SHORT
|
|
#undef NF_INT2_IS_C_INT
|
|
#undef NF_INT2_IS_C_LONG
|
|
#undef NF_INT_IS_C_INT
|
|
#undef NF_INT_IS_C_LONG
|
|
#undef NF_REAL_IS_C_FLOAT
|
|
#undef NF_REAL_IS_C_DOUBLE
|
|
#undef NF_DOUBLEPRECISION_IS_C_DOUBLE
|
|
#undef NF_DOUBLEPRECISION_IS_C_FLOAT
|
|
|
|
|
|
#if 0
|
|
Whether the system uses something besides the IEEE floating-point
|
|
format to represent floating-point values.
|
|
#endif
|
|
#undef NO_IEEE_FLOAT
|
|
|
|
|
|
#if 0
|
|
END OF CUSTOMIZATION
|
|
#endif
|
|
|
|
|
|
#if 0
|
|
FORTRAN data types corresponding to netCDF version 2 "byte" and "short"
|
|
data types (e.g. INTEGER*1, INTEGER*2). See file "ftest.F" for usage.
|
|
#endif
|
|
#if !defined(NO_NETCDF_2)
|
|
# undef NCBYTE_T
|
|
# undef NCSHORT_T
|
|
#endif
|
|
|
|
|
|
#endif
|