Overview of Library Changes Between netCDF Version 2 and 3:

NetCDF version 3 includes a complete rewrite of the netCDF library. It
is about twice as fast as the previous version. The netCDF file format
is unchanged, so files written with version 3 can be read with version 2
code and vice versa.

The library is now written in Standard C. For example, prototypes are
used throughout as well as "const" qualifiers where appropriate. You
*must* have a Standard C compiler to compile this version.

Rewriting the library to use Standard C offered an opportunity to
implement an improved C interface that provides some significant
benefits:

   * Type safety, by eliminating the use of generic void* pointers;

   * Automatic type conversions, by eliminating the undesirable coupling
     between the language-independent external netCDF types (NC_BYTE, ...,
     NC_DOUBLE) and language-dependent internal data types (char, ...,
     double);

   * Support for future enhancements, by eliminating obstacles to the clean
     addition of support for packed data and multithreading;

   * More standard error behavior, by uniformly communicating an error
     status back to the calling program in the return value of each
     function.

It is not necessary to rewrite programs that use the version
2 interface, because the netCDF-3 library includes a backward
compatibility interface that supports all the old functions, globals,
and behavior. We are hoping that the benefits of the new interface will
be an incentive to use it in new netCDF applications. It is possible to
convert old applications to the new interface incrementally, replacing
netCDF-2 calls with the corresponding netCDF-3 calls one at a time. If
you want to check that only netCDF-3 calls are used in an application, a
preprocessor macro (NO_NETCDF_2) is available for that purpose.

Other changes in the implementation of netCDF result in improved
portability, maintainability, and performance on most platforms. A clean
separation between I/O and type layers facilitates platform-specific
optimizations. The new library no longer uses a vendor-provided XDR
library, which simplifies linking programs that use netCDF and speeds up
data access significantly in most cases. A rewrite of the install script
provides a simpler installation procedure that no longer generates new
Makefiles for each platform.

----------------------------------------------------------------------------
The New C Interface:

First, here's an example of C code that uses the netCDF-2 interface:

    void *bufferp;
    nc_type xtype;
    ncvarinq(ncid, varid, ..., &xtype, ...
    ...
    /* allocate bufferp based on dimensions and type */
    ...
    if (ncvarget(ncid, varid, start, count, bufferp) == -1) {
        (void) fprintf(stderr, "Some error message, error code is %d\n"
                       ncerr);
        /* deal with it */
        ...
    }
    switch(xtype) {
        /* deal with the data, according to type */
    ....
    case  NC_FLOAT:
        fanalyze((float *)bufferp);
        break;
    case NC_DOUBLE:
        danalyze((double *)bufferp);
        break;
    }

Here's how you might handle this with the new netCDF-3 interface:

    /*
     * I want to use doubles for my analysis.
     */
    double dbuf[NDOUBLES];
    int status;

    /* So, I use the function that gets the data as doubles. */
    status = nc_get_vara_double(ncid, varid, start, count, dbuf)
    if (status != NC_NOERR) {
       (void) fprintf(stderr, "Some error message, %s\n",
                      nc_strerror(status));
        /* deal with it */
        ...
    }
    danalyze(dbuf);

The example above illustrates changes in function names, data type
conversion, and error handling, discussed in detail in the sections below.

----------------------------------------------------------------------------
The New FORTRAN interface:

The new FORTRAN interface is almost one-for-one identical with the new C
interface.  For example, the following symbols are defined by their
respective interfaces:

    C Interface                         FORTRAN Interface
    -----------                         -----------------

    nc_get_vara_double                  nf_get_vara_double
    nc_get_vara_float                   nf_get_vara_real
    nc_put_var_int                      nf_put_var_int
    nc_inq_libvers                      nf_inq_libvers
    nc_strerror                         nf_strerror
    nc_open                             nf_open
    nc_create                           nf_create
    nc_sync                             nf_sync

----------------------------------------------------------------------------
Naming Conventions:
 
The netCDF-3 library employs a new naming convention, intended to make
netCDF programs more readable.  For example, the name of the function
to rename a variable is now "nc_rename_var" (FORTRAN: "nf_rename_var")
instead of the previous "ncvarrename" (FORTRAN: "ncvren").

All function names begin with "nc_" (FORTRAN: "nf_").  The second
part of the name is a verb, like "get", "put", "inq" (for inquire),
or "open".  The third part of the name is typically the object of the
verb: for example "dim", "var", or "att" for functions dealing with
dimensions, variables, or attributes.  To distinguish the various I/O
operations for variables, "var" gets a single character modifier:

  "var"  access to the entire array

  "var1" access to a single value

  "vara" access to an array or array section

  "vars" strided access to a subsample of values

  "varm" mapped access to values not contiguous in memory

At the end of the name for variable and attribute functions, there is
component indicating the type of the final argument: "text", "uchar",
"schar", "short", "int", "long", "float", or "double" (FORTRAN: "text",
"int1", "int2", "int", "real", or "double").  This part of the function
name indicates the type of the data container you are using in your
program: character string, "unsigned char", "signed char", and so on.

----------------------------------------------------------------------------
Type Conversion:

With the new interface, users need not be aware of the external type of
numeric variables, since automatic conversion to or from any desired
numeric type is now available. You can use this feature to simplify
code, by making it independent of external types. The elimination of
void* pointers provides detection of type errors at compile time that
could not be detected with the previous interface. Programs may be made
more robust with the new interface, because they need not be changed to
accommodate a change to the external type of a variable.

If conversion to or from an external numeric type is necessary, it
is handled by the library. This automatic conversion and separation
of external data representation from internal data types will become
even more important in netCDF version 4, when new external types will
be added for packed data for which there is no natural corresponding
internal type (for example, arrays of 11-bit values).

Converting from one numeric type to another may result in an error if
the target type is not capable of representing the converted value. (In
netCDF-2, such overflows can only happen in the XDR layer.) For example,
a short may not be able to hold data stored externally as an NC_FLOAT
(an IEEE floating-point number). When accessing an array of values, an
ERANGE error is returned if one or more values are out of the range of
representable values, but other values are converted properly.

Note that mere loss of precision in type conversion does not return
an error. Thus, if you read double precision values into a "long",
for example, no error results unless the magnitude of the double
precision value exceeds the representable range of "long"s on your
platform. Similarly, if you read a large integer into a "float"
incapable of representing all the bits of the integer in its mantissa,
this loss of precision will not result in an error.  If you want to
avoid such precision loss, check the external types of the variables
you access to make sure you use an internal type that has a compatible
precision.

The new interface distinguishes arrays of characters intended to
represent text strings from arrays of 8-bit bytes intended to represent
small integers. The interface supports the internal types "text",
"uchar", and "schar", intended for text strings, unsigned byte values,
and signed byte values.

The "_uchar" and "_schar" functions were introduced in netCDF-3 to
eliminate an ambiguity, and support both signed and unsigned byte
data. In netCDF-2, whether the external NC_BYTE type represented signed
or unsigned values was left up to the user. In netcdf-3, we treat
NC_BYTE as signed for the purposes of conversion to "short", "int",
"long", "float", or "double". (Of course, no conversion takes place
when the internal type is "signed char".) In the _uchar functions, we
treat NC_BYTE as if it were unsigned. Thus, no ERANGE error can occur
converting between NC_BYTE and "unsigned char".

----------------------------------------------------------------------------
Error handling:

The new interface handles errors differently than netCDF-2. In the old
interface, the default behavior when an error was detected was to print
an error message and exit. To get control of error handling, you had
to set flag bits in a global variable, "ncopts", and to determine the
cause of an error, you had to test the value of another global variable
"ncerr".

In the new interface, functions return an integer status that indicates
not only success or failure, but also the cause of the error. The global
variables "ncerr" and "ncopt" have been eliminated. The library will
never try to print anything, nor will it call exit() (unless you are
using the netCDF version 2 compatibility functions). You will have to
check the function return status and do this yourself. We eliminated
these globals in the interest of supporting parallel (multiprocessor)
execution cleanly, as well as reducing the number of assumptions
about the environment where netCDF is being used. The new behavior
should provide better support for using netCDF as a hidden layer in
applications that have their own GUI interface.

----------------------------------------------------------------------------
NC_LONG and NC_INT:

Where the netCDF-2 interface used NC_LONG to identify an external data
type corresponding to 32-bit integers, the new interface uses NC_INT
instead.  NC_LONG is defined to have the same value as NC_INT for
backward compatibility, but it should not be used in new code. With new
64-bit platforms using long for 64-bit integers, we would like to reduce
the confusion caused by this name clash. Note that there is still no
netCDF external data type corresponding to 64-bit integers.

----------------------------------------------------------------------------
What's Missing?

The new C interface omits three "record I/O" functions, "ncrecput",
"ncrecget", and "ncrecinq", from the netCDF-2 interface, although
these functions are still supported via the netCDF-2 compatibility
interface. This means you may have to replace one record-oriented call
with multiple type-specific calls, one for each record variable. For
example, a single call to "ncrecput" can always be replaced by multiple
calls to the appropriate "nc_put_var" functions, one call for each
variable accessed. The record-oriented functions were omitted, because
there is no simple way to provide type-safety and automatic type
conversion for such an interface.

There also is no function corresponding to the "nctypelen" function
from the version 2 interface. The separation of internal and external
types and the new type-conversion interfaces make "nctypelen"
unnecessary. Since users read into and write out of native types, the
"sizeof" operator is perfectly adequate to determine how much space to
allocate for a value.

In the previous library, there was no checking that the characters used
in the name of a netCDF object were compatible with the CDL restrictions
that only alphanumeric characters, "_" and "-" are permitted in
names. Now this restriction is enforced by the library on creation of
new dimensions, variables, and attributes, but previously existing
components with names like "@@# ^&* !" will still work OK.

----------------------------------------------------------------------------
Other Changes:

There are two new functions in netCDF-3 that don't correspond to
any netCDF-2 functions: "nc_inq_libvers" and "nc_strerror". The
version of the netCDF library in use is returned as a string by
"nc_inq_libvers". An error message corresponding to the status returned
by a netCDF function call is returned as a string by the "nc_strerror"
function.

A new NC_SHARE flag is available for use in an "nc_open" or "nc_create"
call, to suppress the default buffering of accesses. The use of
NC_SHARE for concurrent access to a netCDF file means you don't have to
call "nc_sync" after every access to make sure that disk updates are
synchronous.

The version 2 interface had a single inquiry function, "ncvarinq",
for getting the name, type, and shape of a variable. Similarly, only
a single inquiry function was available for getting information about
a dimension, an attribute, or a netCDF dataset. When you only wanted
a subset of this information, you had to provide NULL arguments as
placeholders for the unneeded information. The new interface includes
additional inquire functions that return each item separately, so errors
are less likely from miscounting arguments.

The new test program, "nc_test", is run on installation and tests
all the functions in the new interface. A capability has been added
to specify an argument to "nc_test" that will test a read-only
implementations of the netCDF interface.

The previous implementation returned an error when 0-valued count
components were specified in "ncvarput" and "ncvarget" calls. This
restriction has been removed, so that now functions in the "nc_put_var"
and "nc_get_var" families may be called with 0-valued count components,
resulting in no data being accessed. Although this may seem useless, it
simplifies some programs to not treat 0-valued counts as a special case.

The previous implementation returned an error when the same dimension
was used more than once in specifying the shape of a variable in
"ncvardef". This restriction is relaxed in the netCDF-3 implementation,
because an autocorrelation matrix is a good example where using the same
dimension twice makes sense.

In the new interface, units for the imap argument to the mapped access
functions "nc_put_varm_*" and "nc_put_varm_*" are now in terms of the
number of data elements of the desired internal type, not in terms of
bytes as in the netCDF version-2 mapped access interfaces.
