History.txt

=== 0.0.1 / 2012-04-10

* 1 major enhancement

  * Initial alpha release

=== 0.0.2 / 2012-09-21

* 15 major enhancements

  * Second alpha release

  * Rewrote NMatrix in C++0x and C++11 using templates, namespaces;
    removed Ruby generators and CAST parser

  * Added preliminary C API

  * Slicing and referencing support for dense and list matrices (by
    @flipback)

  * BLAS level-3 xTRSM algorithm added for rationals and BLAS types

  * LAPACK support added, including partially working xGETRF
    subroutine

  * Element-wise comparisons now return byte-matrices

  * Element-wise operations on list matrices may alter the default
    value of the return matrix

  * Element-wise division behaves like Ruby division

  * Improved MATLAB .MAT v5 file reading

  * clang support

  * `==` operator now used for matrix equality, `=~` and `!~` for
    element-wise comparisons

  * Dense `each` returns an Enumerator when called without a block

  * Sped up list storage item deletion, fixed bugs

  * List matrix-to-hash conversion with `to_h`

* Note: Element-wise list operations current disabled

=== 0.0.3 / 2013-01-18

* 8 major enhancements

  * Matrix-scalar operations (dense, list)

  * Shortcuts for matrix creation (by @agarie)

  * Access to most ATLAS-implemented LAPACK functions for those
    with ATLAS' CLAPACK interface: xGETRF, xGETRI, xGETRS, xGESV,
    xPOTRF, xPOTRI, xPOTRS, xPOSV, xLASWP, xSCAL, xLAUUM

  * Access to additional ATLAS-implemented BLAS functions: xTRMM,
    xSYRK, xHERK, xROT, xROTG

  * Non-ATLAS versions of CLAPACK functions: xLASWP, xSCAL, xLAUUM,
    xROT

  * Matrix inversion (LU and Cholesky; requires CLAPACK)

  * LU factoring with and without CLAPACK

  * Native matrix I/O for dense (supporting upper, lower, hermitian,
    skew, symmetric, and general) and yale (general only); excludes
    Ruby objects currently

* 2 bug fixes:

  * Yale-to-list casting

  * Now requires packable-1.3.5 or higher, fixing a problem with
    MATLAB .mat v5 file I/O (specific to doubles)

=== 0.0.4 / 2013-05-17

* 3 major enhancements

  * Added a more user-friendly interface for cblas_rot in the form of
    NMatrix::BLAS::rot

  * Added to_hash for Yale matrices

  * Improved source code documentation (by @agarie)

* 4 minor enhancements

  * Spec clean-up (by @masaomi)

  * Made it possible to request a different itype internally for Yale
    matrices

  * Improved space usage of Yale slice-by-copying, which was
    requesting more space than needed

  * Improved compile-time Mac OS X and Ubuntu library searching

* 8 bug fixes:

  * NMatrix::BLAS::gemv segfaults

  * Fixed Yale matrix slice-by-copy write error where default itypes
    (which are based on shape) differ, and a separate problem where
    incorrect IJA and A entries were written.

  * NVector-scalar operations and NVector-NVector element-wise
    options now return an NVector instead of an NMatrix

  * Addressed problems with segmentation faults during iteration (by
    @cjfuller)

  * Addressed Ubuntu/Debian installation problems (incompatibility
    with apt-supplied atlas)

  * Fixed transpose behavior following slice-by-reference (by
    @cjfuller)

  * Fixed gem install command in Rakefile (by @jpmckinney)

  * Fixed Spanish language compile issue (by @imcsk8 and @agarie)

=== 0.0.5 / 2013-07-09

* 4 major enhancements

  * NVector orientation is now controlled by its shape, not by the
    @orientation property

  * NVector default orientation is now a row vector rather than a
    column, as this is more efficient for Yale storage

  * NVector objects may now be created with dtypes other than dense

  * Exposure of additional ATLAS-implemented BLAS functions,
    including native rational and Ruby object support, for xANUM (sum
    of the absolute values of a vector) and xNRM2 (2-norm of a
    vector); and Ruby helper functions BLAS::anum and BLAS::nrm2
    which should do more careful argument sanity checking

* 9 minor enhancements

  * Added #yale_vector_insert to NMatrix::YaleFunctions, to speed up
    insertion of multiple items into a Yale matrix

  * Added #yale_nd_row, #yale_nd_row_as_hash, #yale_nd_row_as_array,
    #yale_nd_row_as_set, #yale_nd_row_as_sorted_set, #yale_row,
    #yale_row_as_hash, #yale_row_as_array, #yale_row_as_set,
    #yale_row_as_sorted_set, #yale_nd_row_size to
    NMatrix::YaleFunctions in order to speed up getting multiple
    items from some row of a Yale matrix

  * Improved #yale_ija, #yale_a, #yale_d by allowing an optional
    index argument, which returns a single array value instead of
    copying and returning the entire array

  * Improved sorting algorithm for Yale matrix multiplication;
    instead of selection sort, now uses quicksort; and subs in
    insertion sort for small partitions

  * Slicing a single row or column now returns an NVector instead
    of an NMatrix (does not yet work for n-dimensional matrices)

  * Improved function documentation for NVector and NMatrix

  * Added #min, #max, #to_a, #shuffle, #shuffle!, #absolute_sum,
    #norm2 functions to NVector

  * Aliased missing dimension of NVector#each_stored_with_indices to
    #each_stored_with_index, which only yields a value and i or j
    (not both i and j) depending on the vector orientation

  * Added #each_row, #each_column to NMatrix

* 5 bug fixes:

  * Dense iterators now return self (an NMatrix) in order to be
    consistent with Ruby Array behavior (by @cjfuller)

  * Fixed Yale resize problem (by @v0dro)

  * Fixed Yale nx1 times 1xn multiplication problem

  * Fixed Yale sorting-following-multiplication problem

  * NMatrix::read() now raises an exception when asked to read a file 
    that does not exist

=== 0.0.6 / 2013-08-09

* 8 major enhancements:

  * Refactored iteration, so that each storage type now has each of:
    #each, #each_with_indices, #each_stored_with_indices

  * Added element-wise power function (**) for dense matrices (by
    @agarie)

  * Dramatically improved matrix element-wise and scalar functions so
    C++ templates are no longer necessary; element-wise operations
    may now be written in protected Ruby methods that look like
    NMatrix#__list_elementwise_op__ and NMatrix#__list_scalar_op__

  * Element-wise and scalar operations that might return a true or
    false now return Ruby matrices

  * Yale element-wise and scalar operations have been added

  * Yale is now allowed to have a non-zero default (specifically to
    permit true-false matrices and nil sparse bases)

  * Dramatically improved NMatrix#cast to allow for hashed options
    including a :default for list and yale casts

  * Dramatically improved speed of compilation

* 14 minor enhancements:

  * Improved documentation for exposed BLAS and LAPACK functions

  * Allowed for use of BLAS::rot without cloning x and y (in-place
    plane rotation); removed unnecessary test of unfriendly version

  * Added more user-friendly cblas_xrotg implementation: BLAS::rotg

  * Moved NMatrix::YaleFunctions::yale_vector_insert to
    NMatrix#__yale_vector_set__, which is more consistent with
    behavior

  * Changed notations for protected stype-specific functions, which
    now look like __stype_method_name__, e.g., __yale_vector_set__

  * Added NMatrix#__list_default_value__ protected function to get
    the initial (sparse) value for list matrices

  * Changed behavior and names of NMatrix::YaleFunctions methods
    which get column indices and cell contents by row, so that they
    now expect the :keys option (like Hash#keys) instead of :array,
    which doesn't make sense; name changes are as follows:
      yale_row_as_sorted_set -> yale_ja_d_keys_sorted_set_at
      yale_row_as_set -> yale_ja_d_keys_set_at
      yale_row_as_array -> yale_ja_d_keys_at
      yale_nd_row_as_sorted_set -> yale_ja_sorted_set_at
      yale_nd_row_as_set -> yale_ja_set_at
      yale_nd_row_as_array -> yale_ja_at
    Aliases are included but will be removed without notice.

  * Added NVector#sorted_indices and #binned_sorted_indices for use
    when running k-nearest neighbor searches on a distance matrix

  * Added NVector::logspace shortcut function (analogous to
    NVector::linspace)

  * Cleaned up code by removing monkey patches that we stopped using
    circa v0.0.2 (Array#min, Array#max, String#constantize,
    String#camelize, String#underscore)

  * Re-enabled element-wise mod (%) method

  * Added NMatrix::guess_dtype class method, which allows you to
    figure out what dtype (roughly) should be used for any given
    Ruby value (e.g., 3)

  * String and nil objects in NMatrix cells are no longer outlawed
    (but are not supported), provided they are of the :object dtype

  * NMatrix#diag shortcut for specifying sparse matrix with a user-
    specified diagonal array (by @ryanmt)

* 3 bug fixes:

  * Corrected BLAS::rot bounds checking on optional n argument

  * Removed BLAS::rotg and BLAS::nrm2 code for rational numbers, as
    both involve a square root

  * Repaired list matrix element-wise functions

=== 0.0.7 / 2013-08-22

* 6 major enhancements:

  * Eliminated NVector in favor of NMatrix objects with
    #effective_dim smaller than #dim; added NVector-like
    functionality to NMatrix, sometimes with #respond_to? dependent
    upon matrix dimensions; allowed for NVector.new to continue
    to function as before, but now returns an NMatrix instead

  * Began major re-factoring of headers for math.cpp

  * Added two singular value decomposition functions for dense
    NMatrix objects, #gesvd and #gesdd, for floating point and
    complex dtypes

  * Added additional slicing shorthand, which uses hashes (e.g.,
    n[0=>3,2] for n[0..3,2]), which may eventually allow users to use
    n[0:3,2] notation instead (needs Ruby core patch)

  * #[] and #[]= calls no longer need index on those matrix shape
    elements which are equal to 1 (e.g., vectors)

  * Yale slicing-by-reference has been added

* 18 minor enhancements:

  * Added user-friendly NMatrix::LAPACK::laswp method

  * Added NMatrix#permute_columns! and NMatrix#permute_columns

  * Added NMatrix#abs to compute element-wise absolute values, and
    #abs_dtype to determine the dtype returned by a call to #abs on a
    given matrix (needed for RSpec)

  * Added NMatrix#size to compute the total number of cells in an
    NMatrix object (solely accounting for the shape, not sparsity)

  * Added RSpec monkey patches for #be_within to work with NMatrix
    objects; usable by requiring "nmatrix/rspec"

  * Added experimental NMatrix::LAPACK::lapack_gesvd function (which
    does NOT depend upon CLAPACK) (by @ryanmt and @mohawkjohn)

  * Added experimental non-LAPACK-dependent function
    NMatrix::LAPACK::lapack_gesdd

  * Added NMatrix#supershape method for getting the shape of a
    slice's parent or other ancestor, which may be useful for calling
    ATLAS and LAPACK functions on slices

  * Aliased NMatrix[] to function like N[] shortcut for matrix
    creation (by @agarie)

  * Added #layer for matrices with dimension greater than two
    (corresponds to #row and #column)

  * Added #rank and #each_rank generics for #row/#column/#layer and
    #each_row/#each_column/#each_layer respectively (#each_rank
    replaces #each_along_dim)

  * Replaced #reduce_along_dim with #inject_rank

  * NMatrix#to_a now works for up to two dimensional matrices; and
    returns a flattened array for single-row matrices

  * NMatrix#to_flat_a now returns a flattened array of all entries

  * Re-organized NMatrix Ruby sources into multiple files: math.rb
    for instance methods which are mathematical in nature or are
    essentially ATLAS/LAPACK/BLAS calls, enumerate.rb for methods
    involving enumeration; and existing shortcuts.rb for convenience
    functions and shortcut constructors, and nmatrix.rb for core
    functionality (#inspect, #to_h, #to_a, #to_s, #pretty_print,
    and so on)

  * Improved #pretty_print, which now also prints layers (the third
    dimension in a 3D matrix)

  * Re-factored some of dense slicing to remove some redundant code

  * Added shortcut functions #list?, #dense?, #yale? for quickly
    testing matrix storage type

* 5 bug fixes:

  * Fixed compilation problem involving <typeinfo> and <vector> STL
    headers

  * Fixed NMatrix#inverse problem with non-square matrices

  * Fixed invalid read problem detected by Valgrind for
    Yale element-wise subtraction in spec

  * Fixed conversion from Ruby object to Complex and Rational

  * Fixed memory leak in slicing

=== 0.0.8 / 2013-08-23

* 2 bug fixes:

  * Fixed Ubuntu compilation bug caused by math header file
    refactoring

  * Fixed pry version error which manifests on some systems but not
    others

=== 0.0.9 / 2013-09-18

* 5 major enhancements:

  * Re-factored NMatrix constructor

  * Improved usability of NMatrix shortcut constructor options
    (e.g., #zeros, #ones, #random, etc.) using new NMatrix
    constructor

  * Left-assignment of slices for all matrix storage types (uses a
    dense cast, or accepts an array or single value)

  * Re-factored Yale into a more object-oriented and less confusing
    set of header files

  * Enabled Travis CI (by @cjfuller)

* 4 minor enhancements:

  * Renamed some specs in order to change their test order, so that
    critical tests fail first (particularly in the case of
    segmentation faults)

  * Default dtype is now :object when no initial values are
    specified

  * Deprecated NVector#initialize and a number of other unnecessary
    NVector functionality

  * Made Ubuntu compilation significantly easier (by @cjfuller)

* 2 bug fixes:

  * nil values in matrices are now pretty printed as "nil"

  * Casting from dense to Yale now properly accepts the default
    value option

=== 0.1.0.rc1 / 2013-12-28

* 4 major enhancements:

  * Improved garbage collection strategy for partial object creation
    (i.e., when VALUEs are allocated but not registered right away),
    which in addition to fixing numerous bugs should prevent some new
    bugs from arising in the future (by @cjfuller)

  * Implemented list storage transpose

  * Implemented generic n-dimensional transpose

  * Implemented == comparison between differing matrix stypes

* 9 minor enhancements:

  * User-friendly #gesvd and #gesdd updates (by @ryanmt)

  * Added experimental #yale_row_key_intersection function for expert
    recommendation problems

  * Added additional *indgen shortcuts and changed behavior for some;
    now, #cindgen for :complex64, #zindgen for :complex128, #findgen
    for :float32, #dindgen for :float64, #rindgen for :rational128,
    and #rbindgen for Ruby objects (which contain integers); also,
    removed code repetition

  * Changed #stddev to use elementwise #sqrt instead of a manual map
    block (by @cjfuller)

  * Added alias from MATLAB `load_mat` method to `load` for
    consistency with the MatrixMarket loader

  * Improved organization by moving list and yale code into storage/
    subdirectories

  * Added NMatrix#potrf! and NMatrix#getrf, which are instance methods
    for calling CLAPACK functions (NMatrix#getrf! already existed)

  * Added GCC installation instructions for Mac OS X Mavericks, and
    updated the old installation instructions for Mac OS X (both
    found in scripts/)

  * Switched NMatrix::VERSION to work more like Rails::VERSION, with
    support for MAJOR, MINOR, TINY, and PRE

  * Added #concat, #hconcat, #vconcat, and #dconcat for joining
    matrices together

* 16 bug fixes:

  * Spec revisions for lapack_gesdd and lapack_gesvd (by @ryanmt)

  * Fixed two double-free problems (by @cjfuller and @mohawkjohn)

  * Fixed contiguous array marking fencepost error

  * Fixed C/C++ API compatibility problem preventing rb/gsl linking

  * Resolved a number of compiler warnings, including one return-type
    problem that would likely have become a garbage collection error
    (if it wasn't already)

  * Fixed -O3 optimization problems

  * Restored NMatrix#asum, #nrm2, #binned_sorted_indices,
    #sorted_indices which were inadvertantly removed by NVector
    deprecation; have not tested

  * Experimental #yale_nd_row and functions which call it now checks
    range of argument to prevent segfault

  * Fixed :* shortcut for a full list dimension (by @cjfuller)

  * Fixed list construction problem which occurred when an initial
    value array was provided (by @cjfuller)

  * Fixed #inject issue with list and yale matrices of two dimensions
    (by @cjfuller)

  * Fixed several garbage collection problems (also listed under
    enhancements) (by @cjfuller)

  * Updated object cleaning target in extconf.rb

  * Fixed possible compilation problem on Mavericks with Xcode 5.02

  * Fixed errors involving undefined symbols, unresolved symbols, and
    lazy symbol binding

  * Improved LAPACK and BLAS header selection for Ubuntu/Debian
    systems with ATLAS (by @mvz)

=== 0.1.0.rc2 / 2014-03-12

* No major enhancements.

* 14 minor enhancements:

  * Implemented negative-index slicing (by @rajatkapoor)

  * Added reader for Point Cloud Library's PCD format

  * Added Ruby 2.1 support (including Travis CI testing)

  * Implemented LAPACK-independent exact inverse calculation for
    dense matrices of size 2x2 and 3x3, as well as

  * Added NMatrix::has_clapack? method to determine whether CLAPACK
    support has been compiled in

  * Improved conformance of specs to RSpec best practices (by 
    @duggiefresh)

  * Travis CI now updates the IRC channel when a check passes (by
    @agarie)

  * Added NMatrix#data_pointer, which returns the memory address of
    the stored elements in a matrix (generally for use with FFI and
    other libraries that need pointers)

  * Made NMatrix#clone_structure a public method (was protected)

  * Added :scale option for NMatrix::random to handle non-floating
    point forms

  * Added complex support to NMatrix::random

  * Aliased NMatrix::random to NMatrix::rand

  * Added NMatrix#reshape! for in-place reshape of dense matrices (by
    @rajatkapoor)

  * Implemented unary negation of matrices

* 6 bug fixes:

  * Fixed dot product operation on 1-dimensional matrices (by @rve
    and @cjfuller)

  * Fixed segfault on 1-dimensional matrix transpose (by @cjfuller)

  * Fixed compile error with Ruby 2.1 (by @diminish7)

  * Fixed regression which wasn't causing any problems but was
    counter to design: stride was declared prior to data storage for
    dense matrix storage

  * Fixed Rakefile problem which was causing specs to run twice in a
    row with each call to rake spec

  * NMatrix::random now raises an exception when rational matrices
    are requested

=== 0.1.0.rc3 / 2014-03-27

* No major enhancements.

* 2 minor enhancements:

  * Exposed NMatrix::LAPACK::geev for LAPACK's xGEEV

  * Added out-of-place complex conjugate for dense and yale storage 
    (by @rve)

* 1 bug fixes:

  * Fixed critical bug with transposing a matrix reference slice (by
    @rajatkapoor)

=== 0.1.0.rc4 / 2014-07-24

* No major enhancements.

* 1 minor enhancement:

  * NMatrix#floor and #ceil implemented (by @v0dro)

* 2 bug fixes:

  * Disallowed out-of-bounds rank calls (by @andrewcsmith)

  * Fixed rspec 3.0 conflict with rspec-longrun 1.0.1

=== 0.1.0.rc5 / 2014-08-01

* No major enhancements.

* 1 minor enhancements:

  * Added optional extension for generating homogeneous
    transformation matrices for rotations and translations in three
    dimensions

* 3 bug fixes:

  * Fixed rake install (by @duggiefresh)

  * Fixed API problems which prevented NMatrix from working with
    the SciRuby rb-gsl fork

  * Fixed Yale #inject behavior (by @yoongkang)

=== 0.1.0 / 2014-12-11

* 3 major enhancements:

  * Updated to BSD 3-clause license

  * Removed Ruby 1.9.2 support; now requires Ruby 1.9.3 or 
    higher (by @v0dro)

  * Added Gauss-Jordan elimination for calculation of
    matrix inverses (by @v0dro)

* 6 minor enhancements:

  * Added trace method for square matrices

  * Improved Array#to_nm monkey patch so matrices and
    arrays can be interconverted easily, without need for a
    shape argument (by @andrewcsmith)

  * Added Harwell-Boeing and Fortran matrix format parsers
    (by @v0dro)

  * Removed soon-to-be-deprecated autoloads and replaced
    with a more robust method (by @duggiefresh)

  * Updated rake install task to use Bundler's GemHelper
    install task (by @duggiefresh)

  * Moved packable requirement from Gemfile to gemspec
    (by @andrewcsmith)

* 3 bug fixes:

  * Corrected Ubuntu clapack functionality checking, which
    should fix most functions which rely on the ATLAS
    version of clapack

  * Corrected NMatrix::gesdd workspace request size (by
    @yoongkang)

  * Fixed definition of NMatrix#asum for one-by-one
    matrices (by @andrewcsmith)

=== 0.2.0 / 2015-08-24

* 2 major enhancements:

  * External libraries are now linked via optional plugins,
    removing ATLAS dependencies (by @wlevine)

  * Made it possible to use NMatrix together with NArray (by
    @blackwinter)

* 9 minor enhancements:

  * Removed rational types (by @wlevine)

  * Added block-diagonal matrix generation method,
    NMatrix.block_diagonal (by @agisga)

  * Added Kronecker product method #kron_prod (by @agisga)

  * Made #permute_columns usage more intuitive (@agisga)

  * Added #pow method to raise a matrix to an integer power
    (by @v0dro)

  * Added #meshgrid method (by @dilcom)

  * Added #hessenberg method, for reducing matrices to upper
    Hessenberg form (by @v0dro)

  * Added calculation of correlation matrix with #corr and
    covariance using #cov (by @v0dro)

  * Added method for returning matrix diagonal, #diagonal
    (by @v0dro)

* 11 bug fixes:

  * Fixed #== operator problems (by @wlevine)

  * Fixed BLAS.gemv (by @wlevine)

  * Fixed #asum for single element complex matrices (by
    @wlevine)

  * Fixed determinant calculation (by @wlevine)

  * Fixed division by zero (by @spacewander)

  * Fixed NMatrix#respond_to? so it would accept two
    arguments properly (by @ktns)

  * Fixed NMatrix#hermitian? (by @agisga)

  * Fixed #gesdd, #gesvd (by @wlevine)

  * Fixed #symmetric? (by @ktns)

  * Made rdoc a development dependency to resolve dependency
    conflicts (by @matt-liu)

  * Fixed bug where Array#to_nm would alter the array (by
    @andrewcsmith)

=== 0.2.1 / 2016-01-18

* 3 major enhancements:
  
  * New plugin nmatrix-fftw for wrapping over FFTW (by
     @v0dro)
  
  * Change Ruby Array C pointer references to be compatible
    with Ruby 2.3.0 (by @mrkn)

  * NMatrix can now be compiled with clang (by @mrkn)

* 4 minor enhancements: 

  * Improved Travis configs to test on Linux and OSX with
    and without plugins (by @mrkn)

  * Added non-abbreviated versions to the options of
    NMatrix#solve; added more docs (by @agisga)

  * Added several specialized algorithms to NMatrix#solve
    for more efficient solving of linear systems with upper
    or lower triangular matrices (by @agisga)

  * Remove redundant C implementation of
    NMatrix#complex_conjugate (by @yoongkang)

* 4 bug fixes:

  * Fixed memory leak in math.cpp (inverse()) (by @lokeshh)

  * Check if optional permute parameter in NMatrix#transpose
    is an Array to prevent unexpected disappearing-parameter
    behavior (by @firemind)
    
  * Moved rubyobj_from_cval function into `nm` namespace
    from C-linkage to fix a C compile time error (by @mrkn)

  * Fixed undefined variable 'idefaults' in lapacke extconf
    (by @agisga)

=== 0.2.2 / 2016-07-22

* No major enhancements.

* 15 minor enhancements:

  * Added Hilbert and inverse Hilbert matrix functions
    #hilbert and #inv_hilbert (by @arafatk)

  * Added NMatrix.linspace constructor for generating a 
    vector with linearly spaced elements (by @gau27)

  * Added NMatrix.logspace constructor for generating a
    vector with logarithmically spaced elements (by @gau27)

  * Improved Travis configs (by @v0dro)

  * Added C API documentation and included ruby_constants.h
    in C API include files (by @v0dro)

  * Added #magic function to create magic square matrices
    (by @shahsaurabh0605)

  * Added NMatrix#last (by @gau27)

  * Added QR factorization by exposing LAPACK functions
    GEQRF, ORMQR, UNMQR (by @gau27)

  * Made templates a little smarter for those functions
    which require a separate return dtype by adding the
    MagnitudeDType template typename; and added a magnitude
    function to replace std::abs and abs to make complex
    and real templates more generic (by @mohawkjohn)

  * Added #adjugate and #adjugate! functions (by @sujithvm)

  * Added #scale and #scale! methods by exposing BLAS SCAL
    (by @lds56)

  * Re-factored type comparisons to use RB_TYPE_P instead of
    TYPE(obj) (by @mrkn)

  * Updated license to BSD 3-clause (by @gau27)

  * Cleaned up gem installation settings and dependencies
    (by @mrkn)

  * DRYed up extconf script (by @mrkn)

* 15 bug fixes:

  * Fixed offsets and changed limits in TRSM to follow the
    Fortran implementation (by @lokeshh), and adjusted
    triangular #solve accordingly (by @agisga)

  * Fixed NRM2 (CBLAS 2-norm) for complex types (by
    @lokeshh)

  * Fixed NRM2 divide-by-zero bug (by @arafatk)

  * Fixed #reshape! to work when changing dimensionality (by
    @wlevine)

  * Fixed ambiguous references by making proper use of
    namespace qualifiers for dtypes, allowing compilation in
    Windows using the mingw64 toolchain available through
    msys2 (by @preetpalS)

  * Replaced all uses of u_int8_t with uint8_t and added static
    assertions to prevent ambiguous use of u_intX_t types
    (by @preetpalS)

  * Added workaround in extconf script for Windows use of a
    different name for the null device (by @preetpalS)

  * Updated deprecated RSpec code and other miscellaneous
    cleanups (by @yui-knk)

  * Removed incomplete support of Hash as an argument for
    NMatrix#[] (by @yui-knk)

  * Fixed typo in slicing exception error message (by
    @mohawkjohn)

  * Fixed #concat implementation for case of differing
    sizes along concatenation dimension (by @alno)

  * Ensured dtype is preserved by #repeat (by @alno)

  * Fixed #det_exact for :object dtype (by @isuruf)

  * Stopped using deprecated register storage class
    specifier (by @mrkn)

  * Fixed clang/clang++ compiler selection by forcing use of
    clang++ when clang is used (by @mrkn)

=== 0.2.3 / 2016-07-25

* No major enhancements.

* No minor enhancements.

* 1 bug fix:

  * Fixed gem installation problem caused by mkmf
    abstraction (by @mrkn)

=== 0.2.4 / 2017-12-14

* No major enhancements.

* 2 minor enhancements:

  * Eliminated code reuse in math.rb between JRuby and MRI
    versions of library (by @prasunanand)

  * Slightly simplified #positive_definite? (by
    @prasunanand)

* 2 bug fixes:

  * Fixed compilation problem on Mac OS X High Sierra (by
    @mohawkjohn)

  * Fixed failing #block_diagonal spec (due to missing
    Array#sum) (by @mohawkjohn)