version 2.1.3
	- Added interp2d and interp3d interpolation functions
	- Added option of nearest-neighbour interpolation

version 2.1.2 (3 Oct 2023)
	- Further bug fixes to reduction of active arrays which did not
	have addequate space allocated by check_space, including "product"
	which requires an additional differential operation per element
	- Fixed out-of-bounds access in test_thread_safe_arrays
	- Slight change to reduce_dimension to avoid incorrect warning
	about ExpressionSize array subscript of -1
	- Fixed broken benchmark/autodiff_benchmark to work with ADOL-C
	- Changed COMPILE_FLAGS argument order in test/Makefile in case
	CPPFLAGS contains Timer.h or other conflicting header file
	- Added benchmark/math_benchmark program

version 2.1.1 (10 April 2022)
	- interp function can perform 1D interpolation of higher
	dimensional Y arrays
	- Bug fix in reduction of an "n" dimensional active array to an
	"n-1" dimensional array: check_space had been forgotten
	- Added Newton-Levenberg[-Marquardt] options to test_minimizer,
	which use the exact Hessian of the Rosenbrock banana function

version 2.1 (5 February 2021)
	- Removed README in favour of README.md

version 2.0.9 (28 January 2021)
	- Fix bug in Array::alignment_offset causing occasional
	crashes reduce and assign operations due to unaligned AVX access,
	now tested in test_packet_operations
	- Added Conjugate-Gradient and L-BFGS minimization methods, both
	bounded and unbounded methods
	- Disabled vectorization on 32-bit ARM NEON targets as there are
	insufficient floating-point intrinsics
	- Fixed interp(x,y,xi) function in case x and y have 0 or 1
	elements

version 2.0.8 (22 August 2020)
	- Added adept_optimize.h header file providing minimization
	capability, initially with the constrained and unconstrained
	Levenberg-Marquardt minimization algorithm
	- Test program test_minimizer tests with the N-dimensional
	Rosenbrock function
	- The Stack member function "jacobian" can now operate on or
	return Adept matrices, rather than solely on raw pointers which
	had to point to data in column-major order
	- Removed "using namespace internal" from several header files so
	that adept namespace is clean
	- Fixed C++98 compatibility

version 2.0.7 (23 June 2020)
	- Added fast, vectorizable exponential function "fastexp", or can
	use as adept::exp if the ADEPT_FAST_EXPONENTIAL preprocessor
	variable is defined
	- Moved all the vector intrinsic stuff to quick_e.h
	- Added ARM-NEON support to quick_e.h
	- Adept is now thread safe on Mac OS versions that support the
	thread_local keyword
	- Fixed bug that caused incorrect differentiation of
	Active<double>/int
	- Preprocessor option ADEPT_INIT_REAL_SNAN and
	ADEPT_INIT_REAL_ZERO initialize real numbers (and complex numbers)
	to signaling NaN or zero, useful for debugging
	- Fixed bug that caused incorrect result of maxval and minval
	applied to active arrays
	- Fixed bug that caused incorrect differentiation of "product"
	function
	- Fixed bug that caused incorrect norm2 for passive vector large
	enough to use vectorization

version 2.0.6 (20 February 2020)
	- Fixed bug in hand-coded adjoint of Toon advection scheme
	(benchmark/advection_schemes_AD.h), as well as other bugs that
	would have prevented the Adjoint and hand-coded adjoints from
	being correct compared to each other
	- Fixed memory leak in Packet.h by ensuring memory is freed in the
	case that neither _POSIX_VERSION nor _MSC_VER are defined
	- Fixed bug in FixedArray.h that prevented active fixed arrays
	from registering themselves with the stack when initialized using
	an initializer list
	- Fixed missing "template" directives in UnaryOperation.h that
	prevented isfinite, isnan and isinf from working correctly on
	arrays
	- Added Array::resize_contigous functions
	- minval and maxval now work correctly with negative and +/-Inf
	arguments; previously minval gave incorrect results even for
	negative arguments
	- Added array_fortran.h to provide the ability to exchange arrays
	between C++/Adept and Fortran, for those Fortran compilers that
	support the 2018 standard
	- Added support for AVX512 vectorization: operations on 16 floats
	and 8 doubles at a time;
	- Added test_packet_operations to check Intel vector intrinsics
	correctly implemented

version 2.0.5 (6 February 2018)
	- Use set_array_print_style(x) to set behaviour of <<Array;
	available are x=PRINT_STYLE_[PLAIN|CSV|CURLY|MATLAB]
	- Fix use of _mm_undefined_ps intrinsic: only use on GCC>=4.9.1
	and Clang if appropriate built-in is present; can't guarantee its
	presence with other compilers
	- Fix writing of active scalar expressions to a stream
	- Added missing fmin/fmax(Expr,Scalar)

version 2.0.4 (8 January 2018)
	- Packet.h copes with undefined _mm_undefined_ps in GCC<4.9.1
	- Fix Packet.h in case SSE2 not enabled
	- ADEPT_FAST preprocessor variable enables
	ADEPT_NO_DIMENSION_CHECKING, ADEPT_NO_ALIAS_CHECKING and
	ADEPT_STACK_THREAD_UNSAFE
	- Divide by scalar now only converts to multiply by (1.0/scalar)
	if scalar is of floating-point type; this fixes indexing with
	"end/2"
	- Fix bug in Packet.h (found by valgrind) to ensure new[] followed
	by delete[] and posix_memalign followed by free
	- Increase initial stack size from 1000 to 1024^2
	- Fixed two bugs in IndexedArray.h that broke indexing a matrix
	with Matrix(int,intVector)
	- Allocated memory in non-OpenMP jacobian_forward is now freed

version 2.0.3 (28 October 2017)
	- Replaced template class "cast" with "expr_cast" to avoid clash
	with Expression's non-template member function; this enables
	compilation with Visual C++.
	- Added adept::have_matrix_multiplication() and
	adept::have_linear_algebra() to test for BLAS and LAPACK
	(respectively) at run-time

version 2.0.2 (21 October 2017)
	- Fixed standards-compliance problem with use of Expression in
	Curiously Recurring Template Pattern, by removing any "static
	const" members that referred to the derived class.  This enabled
	the same code to work with g++, clang++ and the Intel compiler icc.

version 2.0.1 (18 October 2017)
	- Basic passive complex arrays work, tested with
	test/test_complex_arrays
	- Added ADEPT_NO_DIMENSION_CHECKING option
	- Vectorized sqrt, unary-, unary+, max and min
	- Removed the option to vectorize with Packet representing a
	*pair* of SSE2/AVX packed vector; now a Packet can only represent
	a single packed vector. This simplifies maintenance of Packet.h,
	and the pair option offered no performance advantage anyway.
	- Vectorized reduce operations sum, product etc.
	- Many fixes to enable compilation with clang++
	- Fixed FixedArray::operator[] for rank>1

version 2.0 (September 2017)
	- Finalized version for release
	- PDF documentation is no longer installed, so that Git users are
	not obliged to have pdflatex

version 1.9.11 (30 September 2017)
	- Fixed get_gradient member function of Array and FixedArray
	- Added test_array_derivatives test program
	- Fixed indexing of FixedArrays of rank>1
	- Fixed IndexedArray applied to FixedArrays (before had reference
	to temporary dimension object
	- Test and benchmarking programs now work with single precision
	- Stack functions accept Index passed by value rather than
	reference, so that "static const int" passed from FixedArray does
	not need to be explicitly instantiated
	- Active::add_derivative_dependence and
	append_derivative_dependence no longer only accept arguments of
	type "Real"
	- ADEPT_STORAGE_THREAD_SAFE option to protect Storage reference
	counter in multi-threaded environment (C++11 only)
	- Added Array::soft_link() as another means to get thread safety
	- Added test program test_thread_safe_arrays
	- Added adept_reference latex file to doc directory
	- Added "dimensions" function for creating ExpressionSize objects

version 1.9.10 (25 September 2017)
	- Added link syntax A >>= B
	- Added assignment and initialization from initializer_lists for
	Array and FixedArray classes
	- Implemented Fortran-like "count" reduction function
	- Bug fix sending active expression to a stream with "<<"
	- Added "spread<dim>(array,n)" to match Fortran spread(array,dim,n)
	- Added outer_product(x,y)
	- Fixed adept_source.h for non-Unix systems
	- Moved mathematical functions from global to adept namespace
	- Fixed pausable recording and added test_adept_active_pausable
	- Removed unsafe ADEPT_COPY_CONSTRUCTOR_ONLY_ON_RETURN_FROM_FUNCTION
	- C++98 and C++11 correctly take cmath functions from :: and std::
	respectively
	- "make check" now runs test script test/run_tests.sh
	- inv and solve now take general expression arguments
	- Enabled indexed arrays to be assigned to an initializer list
	- BLAS now optional (without it matrix multiplication causes
	run-time exception)
	- Added test_derivatives to test quality of derivatives for all
	mathematical functions
	- Enabled SpecialMatrix and IndexedArray to be assigned to an
	active scalar expression
	- Added fmax and fmin functions (even if C++11 not used)
	- Added atan2 support
	- C++11 on non-Mac platforms uses thread_local keyword instead of
	C++98 compiler extensions
	- Matrix multiplication on active special matrices implemented by
	copying them to a dense Array<2,Real,true>. Very inefficient, but
	it works.
	- Matrix multiplication on inactive triangular and "square"
	matrices now works by converting to them to a dense Array<2,Real,false>.
	- Added alias detection in IndexedArray
	- Alias detection in IndexedArray and SpecialMatrix can be
	deactivated with ADEPT_NO_ALIAS_CHECKING
	- Added "eval" function to evaluate an expression that might be
	subject to aliasing

version 1.9.9 (August 2017)
	- Put on GitHub as rjhogan/Adept-2
	- Added Expression::next_value_contiguous for faster inner loops
	in the case that all expressions have a contiguous and increasing
	inner dimension
	- Preliminary vectorization via Packet class and
	Expression::next_packet
	- Vectorized forward Jacobian calculation using packets
	- Split Expression.h into also UnaryOperation.h and BinaryOperation.h
	- Fixed bug in matmul.h that causes failure if matrix in
	matrix-vector multiplication is strided in both dimensions
	- Added move semantics if C++11 enabled

version 1.9.8 (April 2016):
	- Completed FixedArray.h and tested for active arguments
	- Added array_shortcuts for FixedArrays: (a)VectorX, (a)MatrixXX
	- Added array_shortcuts for Arrays: (a)ArrayXD (for X = 3 to 7)
	- interp permits general Expression arguments

version 1.9.7 (April 2016):
	- Nearly completed FixedArray.h

version 1.9.6 (March 2016):
	- Started FixedArray.h

version 1.9.5 (March 2016):
	- Fixed add_derivative_dependence and append_derivative_dependence
	when applied to elements of arrays
	- Added ADEPT_BOUNDS_CHECKING capability, and fixed IndexedArray
	to work with this
	- Now call BLAS and LAPACK (Fortran) routines, rather than C-BLAS
	and LAPACKE functions
	- Added matrix multiplication benchmark program
	- Added IndexedArray for dimensions up to 7
	- Added Array::data() and Array::const_data() for direct access
	- Added Array::subset(); slightly more concise than using "range"

version 1.9.4 (January 2016):
	- Completed changes to documentation in doc directory
	- Added control/inquiry of settings, e.g. set_max_blas_threads()
	and configuration()

version 1.9.3 (December 2015):
	- Added "max" and "min" as binary operators (note that "maxval"
	and "minval" are reduction operators as in Fortran)

version 1.9.2 (December 2015):
	- Added ActiveConstReference type for active constant references

version 1.9.1 (November 2015):
	- New matmul.h/matmul.cpp - not yet complete

version 1.9.0 (November 2015):
	- SUBSTANTIAL REWRITE TO INCORPORATE ARRAY FUNCTIONALITY

version 1.1 (June 2015):
	- Added ./configure script using autotools
	- Added support for additional mathematical functions: asinh,
	acosh, atanh, expm1, log1p, cbrt, erf, erfc, exp2, log2
	- Changed license from GNU General Public License to Apache
	License, Version 2.0
	- Jacobian calculation uses OpenMP parallelization
	- Removed multiscatter example code
	- New benchmarking program in benchmark/ that compares to other
	automatic differentiation tools if available
	- Fixed bug so that gaps in the gradient list now merge properly
	- Provided capability to compile code without an external library,
	to facilitate porting to Windows
	- Added programs in test/ demonstrating checkpointing,
	thread-safety and compiling without an external library

version 1.0 (September 2013):
	- Very many internal changes and added features
	- Detailed documentation in the doc/ directory
	- Removed the LIFO requirement on the order with which aReal
	objects ought to be created and destroyed
	- For users of version 0.9, the main change to the interface is
	that the Stack::start() member function is no longer supported;
	rather you should call the Stack::new_recording() member function
	*after* the independent variables have been initialized but
	*before* any mathematical operations are performed using them

version 0.9:
	- First public release
