This file contains a high-level description of this package's evolution.
Entries are in reverse chronological order (most recent first).

VERSION         COMMENTS
-------         --------
4.2.1		Released ????-??-?? (Note: Jira entries include reference to '[NCF-XX]')
		
		Upgraded doxygen and man-page documentation for ncdump
		and nccopy utilities, including new -w option for
		nccopy with an example.
		[NCF-26]

		Fixed bugs in accessing fixed-length string attributes and variable
		values from non-netCDF-4 HDF5 files.
		[NCF-150]

		Fixed bug in 64-bit offset files with large records
		due to last record variable requiring more than 2**32
		bytes per record.
		[NCF-164]

		Implemented diskless files for both netcdf classic format
                and extended (netcdf-4) format. For netcdf classic files,
                diskless files also support in-memory caching of disk files.
	        For nc_create(), diskless operation performs all operations
		in memory and then optionally persists the results to a file.
	        For nc_open(), diskless operation caches the file in-memory,
		performs all operations on the memory resident version and then
		writes all changes back to the original file.
		[NCF-110][NCF-109][NCF-5]

		Modified ncio mechanism to support multiple ncio packages;
		this is so that it is possible to have e.g. posixio and memio
		operating at the same time.

		Cleanup up a bunch of lint issues (unused variables, etc).

4.2		Released 2012-03-19 (Note: Jira entries include reference to '[NCF-XX]')

		Completely rebuilt the DAP constraint handling.
		This primarily affects users who specify a DAP constraint
		as part of their URL.
		[NCF-120]

		Fixed cause of slow nccopy performance on file systems
		with many records and large disk block size or many
		record variables, by accessing data a record at a time
		instead of a variable at a time.
  		[NCF-142]

		Performance improvement to DAP code to support
                fetching partial variables into the cache; especially
                important when using nc_get_var() API.
                A partial variable is one that has ranges attached
                to the projection variables (e.g. x[1:10][20:21])
		[NCF-157]

		Separate the Fortran and C++ libraries and release
		the C library and ncdump/ncgen/nccopy without Fortran or C++.
		[NCF-24]

		Documentation mostly migrated to Doxygen, from Texinfo.
		[NCF-26]

		Properly convert vara start/count parameters to DAP
		[NCF-105][NCF-106]

		Fixed major wasted space from previous default variable
		chunk sizes algorithm.
                [NCF-81]

		Fixed bug in nccopy, in which compression and chunking
		options were ignored for netCDF-4 input files.
		[NCF-79]

	        Fixed bug in ncgen in which large variables
                (more than 2**18 elements) duplicates the first
                2**18 values into subsequent chunks of data
                [NCF-154].

		Applied Greg Sjaardema's nccopy bug fix, not
		compressing output variables f they were not already
		using compression on the input file when shuffle
		specified. 
		[NCF-162]

		Fixed problem when a URL is provided that contains only a
		host name.
		[NCF-103]

		Fixed behavior of ncgen flags so that -o => -lb
                and, in the absence of any other markers, make
                the default be -k1
                [NCF-158]

		Created a text INSTALL file for netCDF-4.2 release.
		[NCF-161]

		Fixed bug in ncgen for vlen arrays as fields
                of compound types where datalists for those
                types was improperly interpreted
                [NCF-145] (but see NCF-155).

		Improve use of chunk cache in nccopy utility, making it
		practical for rechunking large files.
		[NCF-85]

		Fixed nccopy bug copying a netCDF-4 file with a
		chunksize for an unlimited dimension that is
		larger than the associated dimension size.
		[NCF-139]

		Fixed nccopy bug when rechunking a netCDF-4 file with
		a chunkspec option that doesn't explicitly specify all
		dimensions.
		[NCF-140]

		Fixed bug in netCDF-4 files with non-coordinate
		variable with the same name as a dimension.
		[NCF-141]

		Incorporated Wei Huang's fix for bug where netCDF-4
		sometimes skips over too many values before adding
		fill values to an in-memory buffer.
		[NCF-143]

		Fixed ncgen bug with netCDF-4 variable-length
		constants (H/T to Lynton Appel).
		[NCF-145]

		Incorporated Peter Cao's performance fixes using HDF5
		link iterator for any group with many variables or
		types.
		[NCF-148]

		Incorporated Constantine Khroulev's bug fix for
		invalid usage of MPI_Comm_f2c in nc_create_par.
		[NCF-135]

		Fixed turning off fill values in HDF5 layers when NOFILL
		mode is set in netCDF-4 API (thanks to Karen
		Schuchardt).
		[NCF-151] 		

		Fixed bug with scalar coordinate variables in netCDF-4
		files, causing failure with --enable-extra-tests
		[NCF-149]

		Cleaned up the definition and use of nulldup.
		[NCF-92][NCF-93][NCF-94]

		Fixed various '#include' bugs.
		[NCF-91][NCF-96][NCF-127]

		v2 API functions modified to properly call the external API
		instead of directly calling the netcdf-3 functions.
		[NCF-100]

		Fixed problem with 64-bit offset format where writing more
		than 2**31 records resulted in erroneous NC_EINVALCOORDS error.
		[NCF-101]

		Restored original functionality of ncgen so that a call with
		no flags, only does the syntax check.
		[NCF-104]

		Corrected misc. test bugs
		[NCF-107]

		Modified ncdump to properly output various new types
		(ubyte, ushort, uint, int64, and uint64).
		[NCF-111]

		Fixed incorrect link flag for szip in configure.ac
		[NCF-116]

		ncdump -t now properly parses ISO "T" separator in
		date-time strings.
		[NCF-16]

		ncdump -t "human time" functionality now available for 
		attributes and bounds variables
		[NCF-70]

		In ncdump, add -g option to support selection of groups
		for which data is displayed.
		[NCF-11]

		Now supports bluefire platform
		[NCF-52]

		ncdump now properly displays values of attributes
		of type NC_USHORT as signed shorts
		[NCF-82]

		Rename some code files so that there are no duplicate filenames.
		[NCF-99]

		Demonstration of netCDF-4 Performance Improvement with KNMI Data
		[NCF-113]

		Dimension size in classic model netCDF-4 files now allows
		larger sizes than allowed for 64-bit offset classic files.
		[NCF-117]

		ncdump now reports correct error message when "-x" option
		specifying NcML output is used on netCDF-4 enhanced model input.
		[NCF-129]

		Fixed bug causing infinite loop in ncdump -c of netCDF-4
		file with subgroup with variables using inherited
		dimensions.
		[NCF-136]

4.1.3		2011-06-17

		Replace use of --with-hdf5= and other such configure
		options that violate conventions and causes build
		problems.  Set environment variables CPPFLAGS,
		LDFLAGS, and LD_LIBRARY_PATH instead, before running
		configure script.
		[NCF-20]

		Detect from configure script when szlib is needed
		[NCF-21]

		Fix bug that can silently zero out portions of a file
		when writing data in nofill mode beyond the end of a
		file, crossing disk-block boundaries with region to be
		written while in-memory buffer is in a specific
		state.	This bug was observed disabling fill mode
		using Lustre (or other large blksize file system) and
		writing data slices in reverse order on disk.
		[NCF-22]

		Fix bug that prevents netCDF-4/HDF5 files created with
		netCDF-4.1.2 from being read by earlier versions of
		netCDF or HDF5 versions before 1.8.7.
		[NCF-23]

		Fix bug in configure that did not make the search for
		the xdr library depend on --enable-dap.
		[NCF-41]

		Fix ncgen bug that did not use the value of a _Format
		attribute in the input CDL file to determine the kind
		of output file created, when not specified by the -k
		command-line flag.
		[NCF-42]

		Fix ncgen bug, not properly handling unsigned longlong parsing.
		[NCF-43]

		Fix DAP client code to suppress variables with names
		such as "x.y", which DAP protocol interprets as
		variable "y" inside container "x".  Such variables
		will be invisible when accessed through DAP client.
		[NCF-47]

		Define uint type for unsigned integer, if not
		otherwise available.  Symptom was compile error
		involving uint in putget.c.
		[NCF-49]

		Fix username+password handling in the DAP client code.
		[NCF-50]

		Add test for handling parallel I/O problem from f77
		when user forgets to turn on one of the two MPI flags.
		[NCF-60]

		Resolved "make check" problems when ifort compiler.
		Some "make install" problems remain when using MPI and
		shared libraries.
		[NCF-61]

		Fix problem with f90_def_var not always handle deflate
		setting when compiler was ifort.
		[NCF-67]

		Check that either MPIIO or MPIPOSIX flag is set when
		parallel create or open is called.  Also fix examples
		that didn't set at least one of these flags.
		[NCF-68]

		Improve documentation on handling client-side certificates
		[NCF-48]

		Document that array arguments, except in varm
		functions, must point to contiguous blocks of memory.
		[NCF-69]

		Get netCDF-4 tests working for DLLs generated with mingw.
		[NCF-6]

		Make changes necessary for upgrading to HDF5 1.8.7
		[NCF-66]

4.1.3-rc1	2011-05-06
		Stop looking for xdr if --disable-dap is used.

		Don't try to run (some) fortran configure tests on
		machines with no fortran.

		Allow nccopy to rechunk with chunksizes larger than
		current dimension lengths.

		Initial implementation of CDMREMOTE is complete; needs
		comprehensive testing.

4.1.3-beta1	2011-04-29

		Fixed szlib not linking bug.

		Fixed dreaded "nofill bug", lurking in netCDF classic
		since at least 1999.  Writing more than a disk
		block's worth of data that crossed disk block
		boundaries more than a disk block beyond the end of
		file in nofill mode could zero out recently written
		earlier data that hadn't yet been flushed to disk.

		Changed setting for H5Pset_libver_bounds to ensure
		that all netCDF-4 files can be read by HDF5 1.8.x.

		Merged libncdap3 and libncdap4 into new libdap2 library.
		The suffix dap2 now refers to the dap protocol. This is
		in prep for adding dap4 protocol support.

		Took out --with-hdf5 and related options due to high
		cost of maintaining this non-standard way of finding
		libraries.

4.1.2          2011-03-29

		Changes in build system to support building dlls on
		cygwin/mingw32.

		Changes to fix portability problems and get things
		running on all test platforms.

		Some minor documentation fixes.

		Fixed opendap performance bug for nc_get_vars; required
		adding nc_get_var{s,m} to the dispatch table.

		Now check for libz in configure.ac.

		Fixed some bugs and some performance problems with
		default chunksizes.

4.1.2-beta2    2011-01-11

		Add "-c" option to nccopy to specify chunk sizes used
		in output in terms of list of dimension names.

		Rewrite netCDF-4 attribute put code for a large
		speedup when writing lots of attributes.

		Fix nc-config --libs when static dependent libraries
		are not installed in the same directory as netCDF
		libraries (thanks to Jeff Whitaker).

		Build shared libraries by default, requiring separate
		Fortran library.  Static libraries now built only with
		--disable-shared.

		Refactor of HDF5 file metadata scan for large speedup
		in opening files, especially large files.

		Complete rewrite of the handling of character datalist
		constants. The heuristics are documented in ncgen.1.

		Eliminate use of NC_MAX_DIMS and NC_MAX_VARS in ncdump
		and nccopy, allocating memory as needed and reducing
		their memory footprint.

		Add documentation for new nc_inq_path() function.

		Use hashing to speedup lookups by name for files with
		lots of dimensions and variables (thanks to Greg
		Sjaardema).
		
		Add options to nccopy to support uniform compression
		of variables in output, shuffling, and fixing
		unlimited dimensions.  Documented in nccopy.1 man page
		and User's Guide.

4.1.2-beta1    2010-07-09

		Fix "ncdump -c" bug identifying coordinate variables in groups.

		Fix bug in libsrc/posixio.c when providing sizehint
		larger than default, which then doesn't get used
		(thanks to Harald Anlauf).

		Fix netCDF-4 bug caused when doing enddef/redef and
		then defining coordinate variable out of order.

		Fixed bug in man4 directory automake file which caused
		documentation to be rebuilt after make clean.

		Turned off HDF5 caching when parallel I/O is in use
		because of its memory use.

		Refactoring of netCDF code with dispatch layer to
		decide whether to call netCDF classic, netCDF-4, or
		opendap version of a function.

		Refactoring of netCDF-4 memory internals to reduce
		memory use and end dependence on NC_MAX_DIMS and
		NC_MAX_NAME.

		Modified constraint parser to be more compatible with
		a java version of the parser.

		Modified ncgen to utilize iterators internally; should be
		no user visible effect.

		Fixed two large-file bugs with using classic format or
		64-bit offset format and accessing multidimensional
		variables with more than 2**32 values.

4.1.1          2010-04-01

		Fixed various build issues.

		Fixed various memory bugs.

		Fixed bug for netCDF-4 files with dimensions and coord
		vars written in different orders, with data writes
		interspersed.

		Added test for HDF5-1.8.4 bug.

		Added new C++ API from Lynton Appel.

4.1            2010-01-30

		Much better memory leak checking with valgrind.

		Added per-variable chunk cache control for better
		performance. Use nc_set_var_chunk_cache /
		nf_set_var_chunk_cache / nf90_set_var_chunk_cache to
		set the per-variable cache.

		Automatically set per-variable chunk cache when
		opening a file, or creating a variable, so that the
		cache is big enough for more than one chunk. (Can be
		overridden by user). Settings may be changed with
		configure options --max-default-chunk-size and
		--default-chunks-in-cache.

		Better default chunks size. Now chunks are sized to
		fit inside the DEFAULT_CHUNK_SIZE (settable at
		configure time with --with-default-chunk-size=
		option.)

		Added nccopy utility for converting among netCDF
		format variants or to copy data from DAP servers to
		netCDF files.

		The oc library has been modified to allow the occurrence
		of alias definitions in the DAS, but they will be ignored.

		The old ncgen has been moved to ncgen3 and
		ncgen is now the new ncgen4.

		Modified --enable-remote-tests to be on by default.

		Fixed the nc_get_varm code as applied to DAP data sources.

		Added tests for nc-config.

		Many documentation fixes.

		Added capability to use the parallel-netcdf
		(a.k.a. pnetcdf) library to perform parallel I/O on
		classic and 32-bit offset files. Use the NC_PNETCDF
		mode flag to get parallel I/O for non-netcdf-4 files.

		Added libcf library to netCDF distribution. Turn it on
		with configure option --with-libcf.

		Added capability to read HDF4 files created with the
		SD (Scientific Data) API.

		The DAP support was revised to closely mimic the
		original libnc-dap support.

		Significantly revised the data handling mechanism
		in ncgen4 to more closely mimic the output from the
		original ncgen.

		Added prototype NcML output capability to ncgen4.
		It is specified by the -lcml flag.

		Added capability to read HDF5 files without dimension
		scales. This will allow most existing HDF5 datasets to
		be read by netCDF-4.

		Fixed bug with endianness of default fill values for
		integer types when variables are created with a
		non-native endiannesss and use the default fill value.

		Significant refactoring of HDF5 type handling to
		improve performance and handle complicated nesting of
		types in cross-platform cases.

		Added UDUNITS2 to the distribution. Use --with-udunits
		to build udunits along with netcdf.

		Made changes suggested by HDF5 team to relax
		creation-order requirement (for read-only cases) which
		allows HDF5 1.6.x files to be retrofitted with
		dimension scales, and be readable to netCDF-4.

		Handle duplicate type names within different groups in
		ncdump.  Fix group path handling in absolute and
		relative variable names for "-v" option.

		Added nc-config shell script to help users build
		netCDF programs without having to figure out all the
		compiler options they will need.

		Fixed ncdump -s bug with displaying special attributes
		for classic and 64-bit offset files.

		For writers, nc_sync() now calls fsync() to flush data
		to disk sooner.

		The nc_inq_type() function now works for primitive types.

4.0.1          2009-03-26

		Added optional arguments to F90 API to
		nf90_open/create, nf90_create_var, and
		nf90_inquire_variable so that all netCDF-4 settings
		may be accomplished with optional arguments, instead
		of separate function calls.

		Added control of HDF5 chunk cache to allow for user
		performance tuning.
		
		Added parallel example program in F90.

		Changed default chunking to better handle very large
		variables.

		Made contiguous the default for fixed size data sets
		with no filters.

		Fixed bug in nc_inq_ncid; now it returns NC_ENOGRP if
		the named group is not found.

		Fixed man pages for C and F77 so that netCDF-4 builds
		will result in man pages that document new netCDF-4
		functions.

		Added OPeNDAP support based on a new C-only implementation.
		This is enabled using --enable-dap option and requires
		libcurl. The configure script will attempt to locate libcurl,
		but if it fails, then its location must be specified by the
		--with-curl option.

4.0.1-beta2    2008-12-26

		Changed chunksizes to size_t from int.

		Fixed fill value problem from F77 API.

		Fixed problems in netcdf-4 files with
		multi-dimensional coordinate variables.

		Fixed ncgen to properly handle CDL input that uses
		Windows line endings ("\r\n"), instead of getting a
		syntax error.

		Added "-s" option to ncdump to display performance
		characterisitics of netCDF-4 files as special virtual
		attributes, such as _Chunking, _DeflateLevel, _Format,
		and _Endianness.

		Added "-t" option to ncdump to display times in human
		readable form as strings.  Added code to interpret
		"calendar" attribute according to CF conventions, if
		present, in displaying human-readable times.

		Added experimental version of ncgen4 capable of
		generating netcdf-4 data files and C code for creating
		them. In addition, it supports the special attributes
		_Format, etc.

4.0.1-beta1    2008-10-16

		Fixed Fortran 90 int64 problems.
		
		Rewrote HDF5 read/write code in accordance with
		performance advice from Kent. 
		
		Fixed memory leaks in gets/puts of HDF5 data.
		
		Fixed some broken tests for parallel I/O (i.e. MPI)
		builds.
		
		Fixed some cross-compile problems.

		Rewrote code which placed bogus errors on the HDF5
		error stack, trying to open non-existant attributes
		and variables. Now no HDF5 errors are seen.
		
		Removed man subdirectory. Now man4 subdirectory is
		used for all builds.

		Changed build so that users with access to parallel
		make can use it.

		Added experimental support for accessing data through
		OPeNDAP servers using the DAP protocol (use
		--enable-opendap to build it).

		Fixed ncdump bugs with array field members of compound
		type variables.  Fixed ncdump bug of assuming default
		fill value for data of type unsigned byte.
		
4.0            2008-05-31
		
		Introduced the use of HDF5 as a storage layer, which
		allows use of groups, user-defined types, multiple
		unlimited dimensions, compression, data chunking,
		parallel I/O, and other features. See the netCDF
		Users Guide for more information.

3.6.3          2008-05-31

		In ncdump and ncgen, added CDL support for UTF-8
		encoding of characters in names and for escaped
		special chars in names.  Made sure UTF-8 names are
		normalized using NFC rules before storing or
		comparing.

		Handle IEEE NaNs and infinities in a
		platform-independent way in ncdump output.

		Added support for ARM representation of doubles,
		(thanks to Warren Turkal).

		Fixed bug in C++ API creating 64-bit offset
		files. (See
		http://www.unidata.ucar.edu/software/netcdf/docs/known_problems.html#cxx_64-bit.)

		Fixed bug for variables larger than 4 GB. (See
		http://www.unidata.ucar.edu/software/netcdf/docs/known_problems.html#large_vars_362.)

		Changed the configure.ac to build either 3.6.x or 4.x
		build from the same configure.ac.

		Build now checks gfortran version and handles it
		cleanly, also Portland Group in Intel fortran, with
		various configurations.

		A Fortran netcdf.inc file is now created at build time, based
		on the setting of --disable-v2. 

		Documentation has been fixed in several places.
		
		Upgraded to automake 1.10, autoconf 2.62, and libtool 2.2.2.

		Includes missing Windows Visual Studio build files.

		Fixed missing include of config.h in a C++ test program.

		Fixed maintainer-clean in man directory.

		Fixed --enable-c-only and make check.

		Fixed behavior when opening a zero-length file.

		Many portability enhancements to build cleanly on
		various platforms. 

		Turned on some old test programs which were not being
		used in the build.

3.6.2          2007-03-05

		Released.

3.6.2 beta6    2007-01-20
		
		Fine tuning of build system to properly handle cygwin,
		Mingw, and strange configuration issues.

		Automake 1.10 has a problem with running our tests on
		MinGW, so I'm switching back to automake 1.9.6 for
		this release.

3.6.2 beta5    2006-12-30

		Now netCDF configuration uses autoconf 2.61, and
		automake 1.10. (Thanks to Ralf Wildenhues for the
		patches, and all the autotools help in general!)

		Final major revision of netCDF tutorial before the
		3.6.2 release.

		Now netCDF builds under MinGW, producing a windows DLL
		with the C and F77 APIs. Use the --enable-shared --enable-dll
		--disable-cxx --disable-f90 flags to configure. (C++
		and F90 have never been built as windows DLLs, but
		might be in a future release if there is user
		interest). This has all been documented in the netCDF
		Porting and Installation Guide.

		Now extreme numbers (i.e. those close to the limits of
		their type) can be turned off in nc_test/nf_test, with
		--disable-extreme-numbers. It is turned off
		automatically for Solaris i386 systems. 

		Added --enable-c-only option to configure. This causes
		only the core netCDF-3 C library to be built. It's the
		same as --disable-f77 --disable-cxx --disable-v2
		--disable-utilities.

		Added --disable-utilities to turn off building and
		testing of ncgen/ncdump.

		Fix a long-standing bug in nf90_get_att_text() pointed
		out by Ryo Furue, to make sure resulting string is
		blank-padded on return.  This is fixed in the
		Fortran-90 interface, but is impractical to fix in the
		Fortran-77 interface implemented via cfortran.h.

		Now large file tests are run if --enable-large-file-tests
		is used in the configure.

		For Cray users, the ffio module is used if the
		--enable-ffio option is passed to configure.

		Unrolled loops in byte-swapping code used on
		little-endian platforms to reduce loop overhead. This
		optimization resulted in a 22% speedup for some
		applications accessing floats or ints (e.g. NCO
		utilities ncap and ncbo) and a smaller speedup for
		shorts or doubles.

		Added "-k" option to ncdump and ncgen, for identifying
		and specifying the kind of netCDF file, one of
		"classic", "64-bit-offset", "hdf5", or "hdf5-nc3".
		Removed output of kind of netCDF file in CDL comment
		produced by ncdump.

		Fixed bug of ncdump seg-faulting if invoked
		incorrectly with option like "-c" or "-h" but no file
		name.

3.6.2 beta4    2006-08-15

		Changed F77/F90 man pages from netcdf.3f and
		netcdf.3f90 to netcdf_f77.3 and netcdf_f90.3. Also
		fixed broken install of man pages.

		Changed configure script so that "-g -O2" is no longer
		set as CFLAGS, CXXFLAGS, and FFLAGS by default if a
		GNU compiler is being used. Now nothing is set.
       
		Changed configure script so that fortran flag is set
		in config.h.

		Updated Installation and Porting Guide, C++
		Interface Guide, F77 and F90 Interface Guides.

		Build with static libraries by default.

		Added configure option --enable-separate-fortran,
		which causes the fortran library to be built
		separately. This is turned on automatically for shared
		libraries.

		Improved clarity of error messages.

		Changed configuration to get cygwin DLL and mingw DLL
		builds working, for the C library only (i.e. no
		F77, F90, or C++ APIs).

		Changed type of ncbyte in C++ interface from unsigned
		char to signed char, for consistency with C interface.
		The C++ documentation warned this change would
		eventually occur.
		
		Changed the C++ interface to use only the netCDF-3 C
		interface instead of the older netCDF-2 C interface.
		This has the added benefit that on-the-fly numeric
		conversions are now supported using get methods, for
		example you can get data of any type as double.  When
		using --disable-v2 flag to configure, the C++
		interface can now be built and installed.

3.6.2 beta3    2006-05-24

		Changed to use default prefix of /usr/local instead of
		package-based prefix of previous releases of
		netCDF. Use the --prefix argument to the configure
		script to override the default.

		Made separate fortran library file, instead of
		appending fortran library functions to the C library
		file, if --enable-separate-fortran is used during
		configure (it's turned on automatically if
		--enable-shared is used). If uses, the fortran API
		users must link to *both* the C library and the new
		fortran library, like this: -lnetcdff -lnetcdf

		Added netCDF examples in C, C++, F77, F90, and
		CDL. See the examples subdirectory.

		Added the NetCDF Tutorial.

		Minor fixes to some of the netCDF documentation.

		Made it possible to build without V2 API using
		--disable-v2 from configure.

		Switched to new build system, with automake and
		libtool. Now shared libraries are built (as well as
		static ones) on platforms which support it. For more
		information about shared libraries, see
		http://www.unidata.ucar.edu/software/netcdf/docs/faq.html#shared_intro

		Fixed ncdump crash that happened when no arguments were 
		used.

		Fixed for building with gfortran 4.1.0.

		Important fix for machines whose SIZEOF_SIZE_T !=
		SIZEOF_LONG, such as NEC-SX, thanks to Stephen Leak.

		Fixed C++ on AIX platform.

		Fixed 64-bit builds on AIX platform.

		Removed bad assertion that could be triggered in rare
		cases when reading a small file.

		Added comments in v1hpg.c to clarify purpose of each
		internal function.

		Make sure filesize is determined in nc_close() *after*
		buffers get flushed.

		Fix long-standing problem resulting in files up to 3
		bytes longer than necessary if there is exactly one
		record variable of type byte, char, or short and if
		the number of values per record for that variable is
		not divisible by 4 (or 2 in the case of short).  Now
		the filesize determined from header info by
		NC_calcsize should be correct in all cases.

3.6.1           2006-01-31

		Updated installation manual for 3.6.1.
       
		Changed installation to try to provide correct
		compiler flags for compiling in 64-bit mode on Sun,
		Irix, AIX, and HPUX. (HPUX doesn't work for me,
		however). Now run configure with --enable-64bit to get
		a 64 bit compile.
		
		Fixed long-standing bug that would cause small netCDF
		files to be padded on the end with zero bytes to 4096
		bytes when they were opened and changed.  Now small
		files should stay small after you change a value.

		Fixed bug in assertions in putget.c that would only be
		noticed if you change the manifest constant
		NC_MAX_DIMS in netcdf.h to be different from
		NC_MAX_VAR_DIMS.

		Moved test ftest.F from fortran to nf_test directory,
		and fixed bug in ftest.F which caused it to return 0
		even if tests failed (no tests were failing,
		however). Also renamed some test
		output files to make things a little clearer.

		If open for writing, pad with up to 3 extra zero bytes
		before close to the correct canonical length,
		calculated from the header.  Previously files could be
		short due to not padding when writing in NOFILL mode.

		Doubled arbitrary limits on number of dimensions,
		variables, attributes, and length of names.

		Change name of nc_get_format() to nc_inq_format().
		Add analogous interfaces for nf_inq_format(),
		nf90_inquire(), and NcFile::get_format() to f77, f90,
		and C++ interfaces.  Document new function in texinfo
		files.  Add minimal test to nc_test, nf_test.

3.6.1-beta3     2005-02-17

		Added function nc_get_format(int ncid, int* formatp)
		that returns either NC_FORMAT_CLASSIC or
		NC_FORMAT_64BIT for a CDF1 or CDF2 file, respectively.

		Added test to nc_test that detects whether format
		version was changed after a file is reopened and
		define mode is entered.

		Correctly configure for Intel ifort Fortran compiler on Linux.

3.6.0-p1        2005-02-18

		Fixed bug that changes CDF2 files to CDF1 files if CDF2
		file is reopened for write access and either an
		attribute is changed or define mode is entered.

3.6.1-beta2     2005-1-6
       
		Fixed absoft compile problem. Maybe.

3.6.1-beta1     2005-1-3

		Fixed Cygwin C++ problem.

		Fixed large file problem in MS Visual C++.NET environment.

		More information in installation and porting guide.

3.6.0           2004-12-16

		Added texinfo source for the documentation.

		Added large file tests to Windows directory in distribution.

		Modified win32 visual studio project files so that m4
		is no longer required to build netcdf under visual studio.

		Modified rules.make to use install instead of cp,
		fixing install problem for cygwin users.
		
		Modified configure/install stuff to support HP-UX.

		Modified configure/install stuff to support G95.
       
		In the f90 interface, applied Arnaud Desitter's fixes
		to correct mismatches between scalar and array
		arguments, eliminating (legitimate) complaints by the
		NAGWare f95 compiler.  Also fixed bugs introduced in
		3.6.0-beta5 in the mapped array interfaces.

3.6.0-beta6     2004-10-05

		Fixed AIX 64-bit/largefile install problems.

		Removed FAQ section from netcdf.texi User's Guide, in
		deference to online version that can be kept up to
		date more easily.

3.6.0-beta5     2004-10-04

		Fixed assertion violation on 64-bit platforms when
		size of last fixed size variable exceeds 2^32 - 1.

		Removed another restriction on file size by making
		record size (derived from other sizes, not part of the
		format) an off_t instead of a size_t, when an off_t is
		larger than a size_t.  This permits records to be
		*much* larger in either classic format or
		64-bit-offset format.

		Incorporated patch from Mathis Rosenhauer to improve
		performance of Fortran 90 interface for calls to
		nf90_put_var_TYPE(), nf90_get_var_TYPE(),
		nf90_put_vara_TYPE(), and nf90_get_vara_TYPE()
		functions by not emulating them with the corresponding
		nf90_put_varm_TYPE() and nf90_get_varm_TYPE() calls.

		Added tests for invalid offsets in classic format when
		defining multiple large variables.

		Improved installation ease.  Have configure script use
		Large File Support as a default, if available.

		Add "extra_test" as a target for testing Large File
		Support.

3.6.0-beta3     2004-08-24

		Upgraded to recent autoconf, changed configure to 
		(hopefully) improve installation. Also added macros
		to deal with large file systems.

		Added nf_set_default_format to Fortran interface.
       
		Added testing to the set_default_format functions to
		nc_test and nf_test.

		Added documentation to the man page for
		set_default_format functions.

		Added two new error return codes to C, f77, and f90
		interfaces for invalid dimension size and for bad
		variable size.  Made test for max dimension size
		depend on whether 64-bit offsets used.  Fixed bug with
		dimension sizes between 2^31 and 2^32 (for byte
		variables).

		Fixed ncdump to properly print dimensions larger than
		2^31. 

		Fixed ncgen to properly handle dimensions between 2^31
		and 2^32.

3.6.0-beta2     
		Added -v2 (version 2 format with 64-bit offsets)
		option to ncgen, to specify that generated files or
		generated C/Fortran code should create 64-bit offset
		files.  Also added -x option to ncgen to specify use
		of no-fill mode for fast creation of large files.

		Added function to set default create mode to C
		interface (nc_set_default_create). 

		Added win32 directory, with NET subdirectory to hold
		.NET port of netCDF. To use, open netcdf.sln with
		Visual Studio, and do a clean and then a build of
		either the debug or release builds. Tests will be run
		as part of the build process. VC++ with managed
		extensions is required (i.e. VC++.NET).

		Added windows installer files to build windows binary
		installs. 

3.6.0-beta1     By incorporating Greg Sjaardema's patch, added support
		for 64-bit offset files, which remove many of the
		restrictions relating to very large files (i.e.
		larger than 2 GB.) This introduces a new data format
		for the first time since the original netCDF format
		was introduced.  Files in this new 64-bit offset
		format can't be read by earlier versions of
		netCDF. Users should continue to use the netCDF
		classic format unless they need to create very large
		files.

		The test suite, nc_test, will now be run twice, once for
		netCDF classic format testing, and once for 64-bit offset
		format testing.

		The implementation of the Fortran-77 interface has been
		adapted to version 4.3 of Burkhard Burow's "cfortran.h".

3.6.0-alpha
		Added NEC SX specific optimization for NFILL tunable
		parameter in libsrc/putget.c

		Added support for the ifc Fortran-90 compiler creating 
		files "netcdf.d" and "typesizes.d" (instead of ".mod"
		files).

		Fixed access to iargc and getarg functions from
		Fortran-90 for NAG f90 compiler, contributed by Harald
		Anlauf.

3.5.1   2004-02-03

		Updated INSTALL.html for Mac OS X (Darwin).

		Made the installation of the netCDF Fortran-90 module
		file more robust regarding the name of the file.

		Added support for eight-byte integers in Fortran90
		interface.              

		Increased advisory limits in C netcdf.h and Fortran
		netcdf.inc for maximum number of dimensions,
		variables, and attributes.

		Changed C++ declarations "friend NcFile" to "friend
		class NcFile" in cxx/netcdfcpp.h to conform to
		standard.
		
		Added Dan Schmitt's backward compatible extension to
		the C++ record interface to work with arbitrary
		dimension slices.

		Added C++ documentation note that caller is
		responsible for deleting pointer returned by
		Variable::values() method when no longer needed.

		Made C++ interface more standard; the result may not
		compile on some old pre-standard C++ compilers.

		Fixed bug in ncgen when parsing values of a
		multidimensional char variable that resulted in
		failure to pad a value with nulls on IRIX.

		Fixed ncdump bug adding extra quote to char variable
		data when using -fc or -ff option.              

		Fixed so compiling with -DNO_NETCDF_2 will work for
		building without backward-compatibility netCDF-2
		interfaces.

		Eliminated use of ftruncate(), because it fails on
		FAT32 file systems under Linux.

		Initialized a pointer in putget.m4 (used to generate
		putget.c) that was involved in uninitialized memory
		references when nc_test is run under Purify.  Two
		users had reported seeing crashes resulting from this
		problem in their applications.

		Reverted pointer initializations in putget.m4, after
		testing revealed these caused a performance problem,
		resulting in many extra calls to px_pgin and px_pgout
		when running nc_test.

		Added checking of size of "dimids" vector in function
		nf90_inquire_variable(...) and error-returning if it
		isn't sufficiently capacious.

		Added variable index to ncvarget() and ncattinq() error
		messages and attribute name to ncattinq() error message.

		Tweaked configure script to work with recent C++ compilers.

		Fixed a memory leak in C++ interface, making sure
		NcVar::cur_rec[] gets deleted in NcVar destructor.

		Reimplemented nc_sync() fix of version 3.5.0 to eliminate 
		performance penalty when synchronization is unnecessary.

		Changed order of targets in Makefile to build Fortran
		interface last, as a workaround for problem with make
		on AIX platforms.

3.5.0   2001-03-23

		Added Fortran 90 interface.

		Changed C macro TIMELEN in file cxx/nctst.cpp to
		TIMESTRINGLEN to avoid clash with macro defined on AIX
		systems in /usr/include/time.h.

		Fixed miswriting of netCDF header when exiting define
		mode.  Because the header was always written correctly
		later, this was only a problem if there was another
		reader of the netCDF file.

		Fixed explicit synchronizing between netCDF writer and
		readers via the nc_sync(), nf_sync(), and ncsync()
		functions.

		Fixed a number of bugs related to attempts to support
		shrinking the header in netCDF files when attributes
		are rewritten or deleted.  Also fixed the problem that
		nc__endef() did not work as intended in reserving
		extra space in the file header, since the extra space
		would be compacted again on calling nc_close().

		Fixed the "redef bug" that occurred when nc_enddef()
		or nf_enddef() is called after nc_redef() or
		nf_redef(), the file is growing such that the new
		beginning of a record variable is in the next "chunk",
		and the size of at least one record variable exceeds
		the chunk size (see netcdf.3 man page for a
		description of this tuning parameter and how to set
		it).  This bug resulted in corruption of some values
		in other variables than the one being added.

		The "__" tuning functions for the Fortran interface,
		nf__create, nf__open, and nf__enddef, are now
		documented in the Fortran interface man pages.

		Add an 'uninstall' target to all the Makefiles.
		Dave Glowacki <dglo@SSEC.WISC.EDU> 199810011851.MAA27335

		Added support for multiprocessing on Cray T3E.
		Hooks added by Glenn, but the majority of the work
		was done at NERSC. Also includes changes to ffio
		option specification. Patch rollup provided by
		R. K. Owen <rkowen@Nersc.GOV>. The following functions
		are added to the public interface.
		nc__create_mp()
		nc__open_mp()
		nc_set_base_pe()
		nc_inq_base_pe()

		Fixed makefile URL for Win32 systems in INSTALL file.

		Made test for UNICOS system in the configure script case
		independent.

		Ported to the following systems:
		AIX 4.3 (both /bin/xlc and /usr/vac/bin/xlc compilers)
		IRIX 6.5
		IRIX64 6.5

		Changed the extension of C++ files from ".cc" to ".cpp".
		Renamed the C++ interface header file "netcdfcpp.h"
		instead of "netcdf.hh", changing "netcdf.hh" to
		include "netcdfcpp.h" for backward compatibility.

		Treat "FreeBSD" systems the same as "BSD/OS" system
		w.r.t. Fortran and "whatis" database.

		Corrected manual pages: corrected spelling of "enddef" (was
		"endef") and ensured that the words "index" and "format"
		will be correctly printed.

		Updated support for Fortran-calling-C interface by
		updating "fortran/cfortran.h" from version 3.9 to
		version 4.1.  This new version supports the Portland
		Group Fortran compiler (C macro "pgiFortran")
		and the Absoft Pro Fortran compiler (C macro
		"AbsoftProFortran").

		Corrected use of non-integral-constant-expression
		in specifying size of temporary arrays in file
		"libsrc/ncx_cray.c".

		Added Compaq Alpha Linux workstation example to INSTALL
		file.

		Ported cfortran.h to Cygnus GNU Win32 C compiler (gcc
		for Windows).

		Fixed bug in ncdump using same CDL header name when
		called with multiple files.
		
		Added new NULL data type  NC_NAT (Not A Type) to
		facilitate checking whether a variable object has had
		its type defined yet, for example when working with
		packed values.

		Fixed use of compile-time macro NO_NETCDF_2 so it
		really doesn't include old netCDF-2 interfaces, as
		intended.

		Ported to MacOS X Public Beta (Darwin 1.2/PowerPC).

		Fixed C++ friend declarations to conform to C++ standard.

		Changed INSTALL file to INSTALL.html instead.

3.4     1998-03-09

		Fixed ncx_cray.c to work on all CRAY systems,
		not just CRAY1. Reworked USE_IEG, which was incorrect.
		Reworked short support. Now USE_IEG and otherwise
		both pass t_ncx.

		To better support parallel systems, static and malloc'ed
		scratch areas which were shared in the library
		were eliminated. These were made private and on the
		stack where possible. To support this, the macros
		ALLOC_ONSTACK and  FREE_ONSTACK are defined in onstack.h.

		The buffered i/o system implementation in posixio.c
		was reimplemented to limit the number and size of read()
		or write() system calls and use greater reliance on memory
		to memory copy. This saves a great deal of wall clock time
		on slow (NFS) filesystems, especially during nc_endef().

		Added performance tuning "underbar underbar" interfaces
		nc__open(), nc__create(), and nc__enddef().

		The 'sizehint' contract between the higher
		layers and the ncio layer is consistently enforced. 

		The C++ interface has been updated so that the
		deprecated "nclong" typedef should no longer be
		required, and casts to nclong no longer necessary.  Just
		use int or long as appropriate.  nclong is still
		supported for backwards compatibility.

		The ncdump utility now displays byte values as signed,
		even on platforms where the type corresponding to a C
		char is unsigned (SGI, for example).  Also the ncdump
		and ncgen utilities have been updated to display and
		accept byte attributes as signed numeric values (with a
		"b" suffix) instead of using character constants.

		In libsrc/error.c:nc_strerror(int), explain that
		NC_EBADTYPE applies to "_FillValue type mismatch".

		Some changes to configure scripts (aclocal.m4), 
		macros.make.in and ncgen/Makefile to support
		NEC SUPER-UX 7.2.

		The "usage" messages of ncgen and ncdump include the
		string returned from nc_inq_libvers().

		Corrected some casts in the library so that all phases of
		the arithmetic computing file offsets occurs with "off_t"
		type. This allows certain larger netcdf files to be created
		and read on systems with larger (64bit) off_t.

		In ncgen, multidimensional character variables are now
		padded to the length of last dimension, instead of just
		concatenating them.  This restores an undocumented but
		convenient feature of ncgen under netCDF-2.  Also, a
		syntax error is now reliably reported if the netcdf name
		is omitted in CDL input.

		Fortran and C code generated by ncgen for netCDF
		components whose names contain "-" characters will now
		compile and run correctly instead of causing syntax
		errors.

		The library allows "." characters in names as well
		as "_" and "-" characters. A zero length name "" is
		explicitly not allowed.  The ncgen utility will now
		permit "." characters in CDL names as well.

		Memory leaks in the C++ interface NcVar::as_*() member
		functions and NcFile::add_var() member function are
		fixed.  The documentation was fixed where it indicated
		incorrectly that the library managed value blocks that
		the user is actually responsible for deleting.

		The values of the version 2 Fortran error codes have
		been modified to make the version 2 Fortran interface
		more backward compatible at the source level.

		Added support for systems whose Fortran INTEGER*1 and
		INTEGER*2 types are equivalent to the C "long" type but
		whose C "int" and "long" types differ.  An example of
		such a system is the NEC SX-4 with the "-ew" option to
		the f90 compiler (sheesh, what a system!).

		Fixed Version 2 Fortran compatibility bug: NCVGTG,
		NCVGGC, NCVPTG, and NCVPGC didn't work according to the
		Version 2 documentation if the innermost mapping value
		(i.e. IMAP[1]) was zero (indicating that the netCDF
		structure of the variable should be used).

3.3.1   1997-06-16

		One can now inquire about the number of attributes that a
		variable has using the global variable ID.

		The FORTRAN interface should now work on more systems.
		In particular:

		It should now work with FORTRAN compilers whose
		"integer*1" datatype is either a C "signed char",
		"short", or "int" and whose "integer*2" datatype is
		either a C "short" or "int".

		It should now work with FORTRAN compilers that are
		extremely picky about source code formatting (e.g.
		the NAG f90 compiler).

		The dependency on the non-POSIX utility m4(1) for
		generating the C and FORTRAN manual pages has been
		eliminated.

		EXTERNAL statements have been added to the FORTRAN
		include-file "netcdf.inc" to eliminate excessive
		warnings about "unused" variables (which were actually
		functions) by some compilers (e.g. SunOS 4.1.3's f77(1)
		version 1.x).

		Building the netCDF-3 package no longer requires the
		existence of the Standard C macro RAND_MAX.

		Fixed an ncdump bug resulting in ncdump reporting
		Attempt to convert between text & numbers
		when _FillValue attribute of a character variable set to
		the empty string "".

		Made ncgen tests more stringent and fixed various bugs
		this uncovered.  These included bugs in handling byte
		attributes on platforms on which char is unsigned,
		initializing scalar character variables in generated C
		code under "-c" option, interspersing DATA statements
		with declaration statements in generated Fortran code
		under "-f" option, handling empty string as a value
		correctly in generated C and Fortran, and handling
		escape characters in strings.  The Fortran output under
		the "-f" option was also made less obscure and more
		portable, using automatic conversion with netCDF-3
		interfaces instead of "BYTE", "INTEGER*1", or
		"INTEGER*2" declarations.

		Fixed a C++ interface problem that prevented compiling
		the C++ library with Digital's cxx compiler.

		Made ncgen "make test" report failure and stop if test
		resulted in a failure of generated C or Fortran code.

		The file that you are now reading was created to contain
		a high-level description of the evolution of the
		netCDF-3 package.

3.3     1997-05-15

		The production version of the netCDF-3 package was
		released.

		A comparison of the netCDF-2 and netCDF-3 releases can
		be found in the file COMPATIBILITY.