Array.h

Go to the documentation of this file.

//# Array.h: A templated N-D Array class with zero origin
//# Copyright (C) 1993,1994,1995,1996,1997,1998,1999,2000,2001,2002,2003,2015
//# Associated Universities, Inc. Washington DC, USA,
//# National Astronomical Observatory of Japan
//# 2-21-1, Osawa, Mitaka, Tokyo, 181-8588, Japan.
//#
//# This library is free software; you can redistribute it and/or modify it
//# under the terms of the GNU Library General Public License as published by
//# the Free Software Foundation; either version 2 of the License, or (at your
//# option) any later version.
//#
//# This library is distributed in the hope that it will be useful, but WITHOUT
//# ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
//# FITNESS FOR A PARTICULAR PURPOSE.  See the GNU Library General Public
//# License for more details.
//#
//# You should have received a copy of the GNU Library General Public License
//# along with this library; if not, write to the Free Software Foundation,
//# Inc., 675 Massachusetts Ave, Cambridge, MA 02139, USA.
//#
//# Correspondence concerning AIPS++ should be addressed as follows:
//#        Internet email: casa-feedback@nrao.edu.
//#        Postal address: AIPS++ Project Office
//#                        National Radio Astronomy Observatory
//#                        520 Edgemont Road
//#                        Charlottesville, VA 22903-2475 USA
 
#ifndef CASA_ARRAY_2_H
#define CASA_ARRAY_2_H
 
//# Includes
#include "ArrayBase.h"
#include "ArrayError.h"
#include "IPosition.h"
#include "MaskLogiArrFwd.h"
#include "Storage.h"
 
#include <complex>
#include <iterator>
#include <initializer_list>
#include <type_traits>
 
namespace casacore { //#Begin casa namespace
 
// <summary> A templated N-D %Array class with zero origin. </summary>
 
// Array<T> is a templated, N-dimensional, %Array class. The origin is zero,
// but by default indices are zero-based. This Array class is the
// base class for the Vector, Matrix, and Cube subclasses.
//
// Indexing into the array, and positions in general, are given with IPosition
// (essentially a vector of integers) objects. That is, an N-dimensional 
// array requires a length-N IPosition to define a position within the array.
// Unlike C, indexing is done with (), not []. Also, the storage order
// is the same as in FORTRAN, i.e. memory varies most rapidly with the first
// index.
// <srcblock>
//                                     // axisLengths = [1,2,3,4,5]
// IPosition axisLengths(5, 1, 2, 3, 4, 5); 
// Array<int> ai(axisLengths);         // ai is a 5 dimensional array of
//                                     // integers; indices are 0-based
//                                     // => ai.nelements() == 120
// Array<int> ai2(axisLengths);        // The first element is at index 0
// IPosition zero(5); zero = 0;        // [0,0,0,0,0]
// //...
// </srcblock>
// Indexing into an N-dimensional array is relatively expensive. Normally
// you will index into a Vector, Matrix, or Cube. These may be obtained from
// an N-dimensional array by creating a reference, or by using an 
// ArrayIterator. The "shape" of the array is an IPosition which gives the
// length of each axis.
//
// An Array may be standalone, or it may refer to another array, or to
// part of another array (by refer we mean that if you change a pixel in 
// the current array, a pixel in the referred to array also changes, i.e.
// they share underlying storage).
// <note role=warning>
//        One way one array can reference another is through the copy 
//        constructor. While this might be what you want, you should
//        probably use the reference() member function to make it explicit.
//        The copy constructor is used when arguments are passed by value;
//        normally functions should not pass Arrays by value, rather they
//        should pass a reference or a const reference. On the positive
//        side, returning an array from a function is efficient since no
//        copying need be done.
// </note>
//
// Aside from the explicit reference() member function, a user will
// most commonly encounter an array which references another array
// when he takes an array slice (or section). A slice is a sub-region of
// an array (which might also have a stride: every nth row, every mth column,
// ...).
// <srcblock>
// IPosition lengths(3,10,20,30);
// Array<int> ai(lengths);         // A 10x20x30 cube
// Cube<int> ci;
// //...
// ci.reference(ai1);              // ci and ai now reference the same
//                                 // storage
// ci(0,0,0) = 123;                // Can use Cube indexing
// ci.xyPlane(2) = 0;              //     and other member functions
// IPosition zero(3,0,0,0);
// assert(ai(zero) == 123);        // true because ai, ci are references
// //...
// Array<int> subArray;
// IPosition blc(3,0,0,0), trc(3,5,5,5);
// subArray.reference(ai(blc, trc));
// subArray = 10;                  // All of subArray, which is the
//                                 // subcube from 0,0,0 to 5,5,5 in
//                                 // ai, has the value 10.
// </srcblock>
// While the last example has an array slice referenced explicitly by another
// array variable, normally the user will often only use the slice as
// a temporary in an expresion, for example:
// <srcblock>
//   Array<Complex> array;
//   IPosition blc, trc, offset;
//   //...
//   // Copy from one region of the array into another
//   array(blc, trc) = array(blc+offset, trc+offset);
// </srcblock>
//
// The Array classes are intended to operate on relatively large
// amounts of data. While they haven't been extensively tuned yet,
// they are relatively efficient in terms of speed. Presently they
// are not space efficient -- the overhead is about 15 words. While
// this will be improved (probably to about 1/2 that), these array
// classes are not appropriate for very large numbers of very small
// arrays. The Block<T> class may be what you want in this circumstance.
//
// Element by element mathematical and logical operations are available
// for arrays (defined in aips/ArrayMath.h and aips/ArrayLogical.h).
// Because arithmetic and logical functions are split out, it is possible
// to create an Array<T> (and hence Vector<T> etc) for any type T that has
// a default constructor, assignment operator, and copy constructor. In
// particular, Array<String> works.
//
// If compiled with the preprocessor symbol AIPS_DEBUG symbol, array
// consistency ("invariants") will be checked in most member
// functions, and indexing will be range-checked. This should not be
// defined for production runs.
//
// <note role=tip>
// Most of the data members and functions which are "protected" should
// likely become "private".
// </note>
//
// <todo asof="1999/12/30">
//   <li> Integrate into the Lattice hierarchy
//   <li> Factor out the common functions (shape etc) into a type-independent
//        base class.
// </todo>
 
template<typename T> class Array : public ArrayBase
{
public:
    // Result has dimensionality of zero, and  nelements is zero.
    Array();
 
    // Create an array of the given shape, i.e. after construction
    // array.ndim() == shape.nelements() and array.shape() == shape.
    // The origin of the Array is zero.
    // Storage is allocated by <src>DefaultAllocator<T></src>.
    // Without initPolicy parameter, the initialization of elements depends on type <src>T</src>.
    // When <src>T</src> is a fundamental type like <src>int</src>, elements are NOT initialized.
    // When <src>T</src> is a class type like <src>casacore::Complex</src> or <src>std::string</src>, elements are initialized.
    // This inconsistent behavior confuses programmers and make it hard to write efficient and generic code using template.
    // Especially when <src>T</src> is of type <src>Complex</src> or <src>DComplex</src> and it is unnecessary to initialize,
    // provide initPolicy with value <src>NO_INIT</src> to skip the initialization.
    // Therefore, it is strongly recommended to explicitly provide initPolicy parameter,
    explicit Array(const IPosition &shape);
 
    // Create an array of the given shape and initialize it with the
    // initial value.
    // Storage is allocated by <src>DefaultAllocator<T></src>.
    Array(const IPosition &shape, const T &initialValue);
    
    // This is a tag for the constructor that may be used to construct an uninitialized Array.
    static struct uninitializedType{} uninitialized;
    
    // Constructor to create an uninitialized array. This constructor can for example
    // be called with:
    // <srcblock>
    //   Array<int> a(shape, Array<int>::uninitialized);
    // </srcblock>
    Array(const IPosition& shape, uninitializedType);
    
    // Construct a one-dimensional array from an initializer list.
    // Example:
    // <srcblock>
    //   Array<int> a({5, 6, 7, 8});
    // </srcblock>
    Array(std::initializer_list<T> list);
 
    // After construction, this and other reference the same storage.
    Array(const Array<T> &other);
 
    // Source will be empty after this call.
    Array(Array<T>&& source) noexcept;
 
    // Create an Array of a given shape from a pointer.
    // If <src>policy</src> is <src>COPY</src>, storage of a new copy is allocated by <src>DefaultAllocator<T></src>.
    // If <src>policy</src> is <src>TAKE_OVER</src>, <src>storage</src> will be destructed and released by the specified allocator.
    // <srcblock>
    //   FILE *fp = ...;
    //   typedef DefaultAllocator<int> Alloc;
    //   Alloc::type alloc;
    //   IPosition shape(1, 10);
    //   int *ptr = alloc.allocate(shape.product());
    //   size_t nread = fread(ptr, sizeof(int), shape.product(), fp);
    //   Array<int> ai(shape, ptr, TAKE_OVER);
    // </srcblock>
    Array(const IPosition &shape, T *storage, StorageInitPolicy policy = COPY);
 
    // Create an Array of a given shape from a pointer. Because the pointer
    // is const, a copy is always made.
    // The copy is allocated by <src>DefaultAllocator<T></src>.
    Array(const IPosition &shape, const T *storage);
 
    // Construct an array from an iterator and a shape.
    template<typename InputIterator>
    Array(const IPosition &shape, InputIterator startIter);
    
    // Frees up storage only if this array was the last reference to it.
    virtual ~Array() noexcept;
 
    // Make an empty array of the same template type.
    virtual std::unique_ptr<ArrayBase> makeArray() const override;
 
    // Assign the other array to this array.
    // If the shapes mismatch, this array is resized.
    // <group>
    void assign (const Array<T>& other);
    
    void assignBase (const ArrayBase& other, bool checkType=true) override;
    // </group>
 
    // Set every element of the array to "value." Also could use the
    // assignment operator which assigns an array from a scalar.
    void set(const T &value);
 
    // Apply the function to every element of the array. This modifies
    // the array in place.
    // (TODO this version made the other versions of apply() redundant)
    template<typename Callable>
    void apply(Callable function);
 
    // After invocation, this array and other reference the same storage. That
    // is, modifying an element through one will show up in the other. The
    // arrays appear to be identical; they have the same shape.
    // <br>Please note that this function makes it possible to reference a
    // const Array, thus effectively it makes a const Array non-const.
    // Although this may seem undesirable at first sight, it is necessary to
    // be able to make references to temporary Array objects, in particular to 
    // Array slices. Otherwise one first needs to use the copy constructor.
    //# The const has been introduced on 2005-Mar-31 because of the hassle
    //# involved in calling the copy ctor before reference.
    virtual void reference(const Array<T> &other);
 
    // Copy the values in other to this. If the array on the left hand
    // side has no elements, then it is resized to be the same size as
    // as the array on the right hand side. Otherwise, the arrays must
    // conform (same shapes).
    // <srcblock>
    // IPosition shape(2,10,10);     // some shape
    // Array<double> ad(shape);
    // //...
    // Array<double> ad2;            // N.B. ad2.nelements() == 0
    // ad2 = ad;                     // ad2 resizes, then elements
    //                               //     are copied.
    // shape = 20;
    // Array<double> ad3(shape);
    // ad3 = ad;                     // Error: arrays do not conform
    // </srcblock>
    // Note that the assign function can be used to assign a
    // non-conforming array.
    Array<T>& assign_conforming(const Array<T>& other)
    {
      return assign_conforming_implementation(other, std::is_copy_assignable<T>());
    }
    
    // Copy to this those values in marray whose corresponding elements
    // in marray's mask are true.
    // <thrown>
    //    <li> ArrayConformanceError
    // </thrown>
    //
    Array<T>& assign_conforming (const MaskedArray<T>& marray);
 
    // TODO we should change the semantics
    Array<T>& operator=(const Array<T>& other)
    { return assign_conforming(other); }
 
    // Calls assign_conforming().
    Array<T>& operator=(const MaskedArray<T>& marray)
    { return assign_conforming(marray); }
    
    // The move operator takes the storage from the given array. After moving an
    // Array, the source Array will be left empty.
    Array<T>& operator=(Array<T>&& other);
 
    // Set every element of this array to "value". In other words, a scalar
    // behaves as if it were a constant conformant array.
    Array<T>& operator=(const T& value);
 
    // This makes a copy of the array and returns it. This can be
    // useful for, e.g. making working copies of function arguments
    // that you can write into.
    // <srcblock>
    // void someFunction(const Array<int> &arg)
    // {
    //     Array<int> tmp(arg.copy());
    //     // ...
    // }
    // </srcblock>
    // Note that since the copy constructor makes a reference, if we just
    // created used to copy constructor, modifying "tmp" would also
    // modify "arg". Clearly another alternative would simply be:
    // <srcblock>
    // void someFunction(const Array<int> &arg)
    // {
    //     Array<int> tmp;
    //     tmp = arg;
    //     // ...
    // }
    // </srcblock>
    // which likely would be simpler to understand. (Should copy() 
    // be deprecated and removed?)
    //
    // TODO deprecate
    Array<T> copy() const;
    
    // This function copies the matching part of from array to this array.
    // The matching part is the part with the minimum size for each axis.
    // E.g. if this array has shape [4,5,6] and from array has shape [7,3],
    // the matching part has shape [4,3].
    // <br>Note it is used by the resize function if
    // <src>copyValues==true</src>.
    void copyMatchingPart (const Array<T>& from);
 
    // This ensures that this array does not reference any other storage.
    // <note role=tip>
    //        When a section is taken of an array with non-unity strides,
    //        storage can be wasted if the array, which originally contained
    //        all the data, goes away. unique() also reclaims storage. This
    //        is an optimization users don't normally need to understand.
    //
    //        <srcblock>
    //        IPosition shape(...), blc(...), trc(...), inc(...);
    //        Array<float> af(shape);
    //        inc = 2; // or anything > 1
    //        Array<float> aSection.reference(af(blc, trc, inc));
    //        af.reference(anotherArray);
    //        // aSection now references storage that has a stride
    //        // in it, but nothing else is. Storage is wasted.
    //        aSection.unique();
    //        </srcblock>
    // </note>
    void unique();
 
    // Create an STL vector from an Array. The created vector is a linear
    // representation of the Array memory. See
    // <linkto class=Vector>Vector</linkto>  for
    // details of the operation and its reverse (i.e. creating a 
    // <src>Vector</src> from a <src>vector</src>), and for details of
    // definition and instantiation.
    // <group>
    template <class U>
    void tovector(std::vector<T, U>& out) const;
    std::vector<T> tovector() const;
    // </group>
 
    // It is occasionally useful to have an array which access the same
    // storage appear to have a different shape. For example,
    // turning an N-dimensional array into a Vector.
    // <br>When the array data are contiguous, the array can be reshaped
    // to any form as long as the number of elements stays the same.
    // When not contiguous, it is only possible to remove or add axes
    // with length 1.
    // <srcblock>
    // IPosition squareShape(2,5,5);
    // Array<float> square(squareShape);
    // IPosition lineShape(1,25);
    // Vector<float> line(square.reform(lineShape));
    // // "square"'s storage may now be accessed through Vector "line"
    // </srcblock>
    Array<T> reform(const IPosition& shape) const;
 
    // Having an array that can be reused without requiring reallocation can
    // be useful for large arrays.  The method reformOrResize permits this
    // usage.
    //
    // The reformOrResize method first attempts to reform the matrix so that
    // it reuses the existing storage for an array with a new shape.  If the
    // existing storage will not hold the new shape, then the method will
    // resize the array when resizeIfNeeded is true; if a resize is needed and
    // resizeIfNeeded is false, then an ArrayConformanceError is thrown.  The
    // copyDataIfNeeded parameter is passed to resize if resizing is performed.
    // resizePercentage is the percent of additional storage to be addeed when
    // a resize is performed; this allows the allocations to be amortized when
    // the caller expects to be calling this method again in the future.  The
    // parameter is used to define an allocation shape which is larger than
    // the newShape by increasing the last dimension by resizePercentage percent
    // (i.e., lastDim = (lastDim * (100 + resizePercentage)) / 100).  If
    // resizePercentage <= 0 then resizing uses newShape as-is.  Returns true
    // if resizing (allocation) was performed.
    //
    // To truncate the array so that it no longer holds additional storage,
    // use the resize method.
    //
    // Array may not be shared with another Array object during this call.
    // Exception thrown if it is shared.
 
    bool reformOrResize (const IPosition & newShape,
                         size_t resizePercentage = 0,
                         bool resizeIfNeeded = true);
 
    // Use this method to extend or reduce the last dimension of an array.  If
    // sufficient excess capacity exists then the bookkeeping is adjusted to 
    // support the new shape.  If insufficient storage exists then a new array
    // is allocated (unless resizeIfNeeded is false; then an exception is thrown).
    // If resizing is not required then the data remains untouched; if resizing
    // is required then the data is copied into the new storage.  The resizePercentage
    // works the same as for reformOrResize (see above).  This method never releases
    // extra storage; use "resize" to do this.  Array may not be sharing storage
    // with another array at call time; an exception will be thrown if the array is shared.
    // Returns true if the array was extension required a Array<T>::resize operation.
 
    bool adjustLastAxis (const IPosition & newShape,
                 size_t resizePercentage = 0, 
                         bool resizeIfNeeded = true);
 
    // Returns the number of elements allocated.  This value is >= to the value returned
    // by size().
 
    size_t capacity () const; 
 
    // These member functions remove degenerate (ie. length==1) axes from
    // Arrays.  Only axes greater than startingAxis are considered (normally
    // one wants to remove trailing axes). The first two of these functions
    // return an Array reference with axes removed. The latter two functions
    // let this Array object reference the 'other' array with degenerated axes
    // removed.
    // <br>
    // Unless throwIfError is false, an exception will be thrown if
    // startingAxis exceeds the array's dimensionality.
    // <br>
    // The functions with argument <src>ignoreAxes</src> do
    // not consider the axes given in that argument. In this way it can be
    // achieved that degenerate axes are kept.
    // <note role=caution> When the two functions returning <src>void</src>
    // are invoked on a derived object (e.g. Matrix), an exception is
    // thrown if removing the degenerate axes from other does not result
    // in a correct number of axes.
    // </note>
    // <group>
    Array<T> nonDegenerate(size_t startingAxis=0, bool throwIfError=true) const;
    Array<T> nonDegenerate(const IPosition& ignoreAxes) const;
    void nonDegenerate(const Array<T> &other, size_t startingAxis=0,
               bool throwIfError=true);
    void nonDegenerate(const Array<T> &other, const IPosition &ignoreAxes)
        { doNonDegenerate (other, ignoreAxes); }
    // </group> 
 
    // Remove degenerate axes from this Array object.
    // Note it does not make sense to use these functions on a derived object
    // like Matrix, because it is not possible to remove axes from them.
    // <group>
    void removeDegenerate(size_t startingAxis=0,
                          bool throwIfError=true);
    void removeDegenerate(const IPosition &ignoreAxes);
    // </group>
 
    // This member function returns an Array reference with the specified
    // number of extra axes, all of length one, appended to the end of the
    // Array. Note that the <src>reform</src> function can also be
    // used to add extra axes.
    // <group>
    const Array<T> addDegenerate(size_t numAxes) const;
    Array<T> addDegenerate(size_t numAxes);
    // </group>
 
    // Make this array a different shape. If <src>copyValues==true</src>
    // the old values are copied over to the new array.
    // Copying is done on a per axis basis, thus a subsection with the
    // minimum of the old and new shape is copied.
    // <br>Resize without argument is equal to resize(IPosition()).
    // <br>It is important to note that if multiple Array objects
    // reference the same data storage, this Array object still references
    // the same data storage as the other Array objects if the shape does
    // not change. Otherwise this Array object references newly allocated
    // storage, while the other Array objects still reference the existing
    // data storage.
    // <br>If you want to be sure that the data storage of this Array object
    // is not referenced by other Array objects, the function unique should
    // be called first.
    // <group>
    void resize();
    
    void resize(const IPosition &newShape, bool copyValues=false) override;
    // </group>
 
    // Access a single element of the array. This is relatively
    // expensive. Extensive indexing should be done through one
    // of the Array specializations (Vector, Matrix, Cube).
    // <group>
    T &operator()(const IPosition &);
    const T &operator()(const IPosition &) const;
    // </group>
 
    // Get a reference to an array section extending
    // from start to end (inclusive).
    // <group>
    Array<T> operator()(const IPosition &start,
                        const IPosition &end);
    const Array<T> operator()(const IPosition &start,
                              const IPosition &end) const;
    // Along the ith axis, every inc[i]'th element is chosen.
    Array<T> operator()(const IPosition &start,
                        const IPosition &end,
            const IPosition &inc);
    const Array<T> operator()(const IPosition &start,
                              const IPosition &end,
                              const IPosition &inc) const;
    // </group>
 
    // Get a reference to an array section using a Slicer.
    // <group>
    Array<T> operator()(const Slicer&);
    const Array<T> operator()(const Slicer&) const;
    // </group>
 
    // Get a reference to a section of an array.
    // This is the same as operator(), but can be used in a type-agnostic way.
    std::unique_ptr<ArrayBase> getSection (const Slicer&) const override;
 
    // Get the subset given by the i-th value of the last axis. So for a cube
    // it returns the i-th xy plane. For a Matrix it returns the i-th row.
    // The returned array references the original array data; its dimensionality
    // is one less. For a 1-dim array it still returns a 1-dim array.
    // <note>This function should not be used in tight loops as it is (much)
    // slower than iterating using begin() and end(), ArrayIter, or
    // ArrayAccessor.</note>
    Array<T> operator[] (size_t i) const;
 
    // Get the diagonal of each matrix part in the full array.
    // The matrices are taken using axes firstAxes and firstAxis+1.
    // diag==0 is main diagonal; diag>0 above the main diagonal; diag<0 below.
    Array<T> diagonals (size_t firstAxis=0, long long diag=0) const;
 
    // The array is masked by the input LogicalArray.
    // This mask must conform to the array.
    // <group>
    const MaskedArray<T> operator() (const LogicalArray &mask) const;
    MaskedArray<T> operator() (const LogicalArray &mask);
    // </group>
 
    // The array is masked by the input MaskedLogicalArray.
    // The mask is effectively the AND of the internal LogicalArray
    // and the internal mask of the MaskedLogicalArray.
    // The MaskedLogicalArray must conform to the array.
    // <group>
    const MaskedArray<T> operator() (const MaskedLogicalArray &mask) const;
    MaskedArray<T> operator() (const MaskedLogicalArray &mask);
    // </group>
 
    // The number of references the underlying storage has assigned to it.
    // It is 1 unless there are outstanding references to the storage (e.g.,
    // through a slice). Normally you have no need to do this since the
    // arrays handle all of the references for you.
    // NB: Even when nrefs()==1, the array might be shared, because the
    // the storage itself might be shared. Therefore, this function should
    // not be used outside debugging.
    // TODO make protected.
    size_t nrefs() const;
 
    // Check to see if the Array is consistent. This is about the same thing
    // as checking for invariants. If AIPS_DEBUG is defined, this is invoked
    // after construction and on entry to most member functions.
    virtual bool ok() const override;
 
    // Are the shapes identical?
    // <group>
    bool conform (const Array<T> &other) const
      { return conform2(other); }
    bool conform (const MaskedArray<T> &other) const;
    // </group>
 
    // Get a pointer to the beginning of the array.
    // Note that the array may not be contiguous.
    // <group>
    T* data()
      { return begin_p; }
    const T* data() const
      { return begin_p; }
    // </group>
 
    // Generally use of this should be shunned, except to use a FORTRAN routine
    // or something similar. Because you can't know the state of the underlying
    // data layout (in particular, if there are increments) sometimes the
    // pointer returned will be to a copy, but often this won't be necessary.
    // A boolean is returned which tells you if this is a copy (and hence the
    // storage must be deleted). Note that if you don't do anything unusual,
    // getStorage followed by freeStorage or putStorage will do the deletion
    // for you (if required). e.g.:
    // <srcblock>
    // Array<int> a(shape); ...
    // bool deleteIt; int *storage = a.getStorage(deleteIt);
    // foo(storage, a.nelements()); a.puStorage(storage, deleteIt);
    // // or a.freeStorage(storage, deleteIt) if a is const.
    // </srcblock>
    // NB: However, if you only use getStorage, you will have to delete the
    // pointer yourself using freeStorage().
    //
    // It would probably be useful to have corresponding "copyin" "copyout"
    // functions that used a user supplied buffer.
    // Note that deleteIt is set in this function.
    // <group>
    T *getStorage(bool& deleteIt);
    const T *getStorage(bool& deleteIt) const
    {
      // The cast is OK because the return pointer will be cast to const
      return const_cast<Array<T>*>(this)->getStorage(deleteIt);
    }
    void *getVStorage(bool &deleteIt) override;
    const void *getVStorage(bool &deleteIt) const override;
    // </group>
 
    // putStorage() is normally called after a call to getStorage() (cf).
    // The "storage" pointer is set to zero.
    void putStorage(T *&storage, bool deleteAndCopy);
    void putVStorage(void *&storage, bool deleteAndCopy) override;
 
    // If deleteIt is set, delete "storage". Normally freeStorage calls
    // will follow calls to getStorage. The reason the pointer is "const"
    // is because only const pointers are released from const arrays.
    // The "storage" pointer is set to zero.
    // TODO this function can not be const for stateful allocators
    void freeStorage(const T *&storage, bool deleteIt) const;
    void freeVStorage(const void *&storage, bool deleteIt) const override;
 
    // Replace the data values with those in the pointer <src>storage</src>.
    // The results are undefined if storage does not point at nelements() or
    // more data elements. After takeStorage() is called, <src>nrefs()</src>
    // is 1.
    // <group>
    // If <src>policy</src> is <src>COPY</src>, storage of a new copy is allocated by <src>allocator</src>.
    // If <src>policy</src> is <src>TAKE_OVER</src>, <src>storage</src> will be destructed and released by <src>allocator</src>.
    virtual void takeStorage(const IPosition &shape, T *storage,
        StorageInitPolicy policy = COPY);
 
    // Since the pointer is const, a copy is always taken.
    // Storage of a new copy is allocated by the specified allocator.
    virtual void takeStorage(const IPosition &shape, const T *storage);
    // </group>
 
 
    // Used to iterate through Arrays. Derived classes VectorIterator and
    // MatrixIterator are probably more useful.
    friend class ArrayIterator<T>;
 
    // Create an ArrayIterator object of the correct type.
    std::unique_ptr<ArrayPositionIterator> makeIterator(size_t byDim) const override;
 
    // <group name=STL-iterator>
    // See the function begin() and end() for a detailed description
    // of the STL iterator capability.
    class BaseIteratorSTL
    {
    public:
      // Create the begin const_iterator object for an Array.
      explicit BaseIteratorSTL (const Array<T>&);
      // Create the end const_iterator object for an Array.
      // It also acts as the default constructor.
      explicit BaseIteratorSTL (const T* end = 0)
    : itsPos(end), itsLineEnd(0), itsLineIncr(0), itsLineAxis(0),
      itsArray(0), itsContig(false) {}
 
      void nextElem()
      {
    itsPos++;
    if (!itsContig) {
      itsPos += itsLineIncr;
      if (itsPos > itsLineEnd) increment();
    }
      }
      void nextLine()
      {
    itsPos = itsLineEnd;
    increment();
      }
 
      bool operator== (const BaseIteratorSTL& other) const
        { return itsPos == other.itsPos; }
 
      bool operator!= (const BaseIteratorSTL& other) const
        { return itsPos != other.itsPos; }
 
      T* getPos()
        { return const_cast<T*>(itsPos); }
 
      friend std::ostream& operator<< (std::ostream& os, const BaseIteratorSTL& iter)
        { os << iter.itsPos; return os; }
 
    protected:
      // Increment iterator for a non-contiguous array.
      void increment();
 
      const T*  itsPos;
      const T*  itsLineEnd;
      size_t    itsLineIncr;
      size_t      itsLineAxis;
      IPosition itsCurPos;
      IPosition itsLastPos;
      const Array<T>* itsArray;
      bool      itsContig;
    };
 
    class IteratorSTL: public BaseIteratorSTL
    {
    public:
      // <group name=STL-iterator-typedefs>
      typedef T                 value_type;
      typedef value_type*       pointer;
      typedef value_type&       reference;
      typedef std::size_t       size_type;
      typedef ptrdiff_t         difference_type;
      typedef std::forward_iterator_tag iterator_category;
      // </group>
 
      // Create the begin iterator object for an Array.
      explicit IteratorSTL (Array<T>& arr)
    : BaseIteratorSTL (arr) {}
      // Create the end iterator object for an Array.
      // It also acts as the default constructor.
      explicit IteratorSTL (const T* end = 0)
    : BaseIteratorSTL (end) {}
 
      const IteratorSTL& operator++()
      {
    this->nextElem();
    return *this;
      }
      IteratorSTL operator++(int)
      {
    IteratorSTL old(*this);
    this->nextElem();
    return old;
      }
 
      T& operator*()
        { return *this->getPos(); }
      T* operator->()
        { return this->getPos(); }
    };
 
  class ConstIteratorSTL: public BaseIteratorSTL
    {
    public:
      // <group name=STL-const-iterator-typedefs>
      typedef T                 value_type;
      typedef const value_type* pointer;
      typedef const value_type& reference;
      typedef std::size_t       size_type;
      typedef ptrdiff_t         difference_type;
      typedef std::forward_iterator_tag iterator_category;
      // </group>
 
      // Create the begin const_iterator object for an Array.
      explicit ConstIteratorSTL (const Array<T>& arr)
    : BaseIteratorSTL (arr) {}
      // Create the end const_iterator object for an Array.
      // It also acts as the default constructor.
      explicit ConstIteratorSTL (const T* end = 0)
    : BaseIteratorSTL (end) {}
      // Create from a non-const iterator.
      ConstIteratorSTL (const IteratorSTL& iter)
    : BaseIteratorSTL (iter) {}
 
      const ConstIteratorSTL& operator++()
      {
    this->nextElem();
    return *this;
      }
      ConstIteratorSTL operator++(int)
      {
    ConstIteratorSTL old(*this);
    this->nextElem();
    return old;
      }
 
      const T& operator*() const
        { return *this->itsPos; }
      const T* operator->()
        { return this->itsPos; }
 
      const T* pos() const
        { return this->itsPos; }
    };
    // </group>
 
    // Define the STL-style iterator functions (only forward iterator).
    // It makes it possible to iterate through all data elements of an array
    // and to use it common STL functions.
    // The end() function is relatively expensive, so it should not be
    // used inside a for statement. It is much better to call it beforehand
    // as shown in the example below. Furthermore it is very important to
    // use <src>++iter</src>, because <src>iter++</src> is 4 times slower.
    // <srcblock>
    //  Array<int> arr(shape);
    //  Array<int>::iterator iterend(arr.end());
    //  for (Array<int>::iterator iter=arr.begin(); iter!=iterend; ++iter) {
    //    *iter += 1;
    //  }
    // </srcblock>
    // The Array class supports random access, so in principle a random
    // iterator could be implemented, but its performance would not be great,
    // especially for non-contiguous arrays.
    // <br>Some other STL like functions exist for performance reasons.
    // If the array is contiguous, it is possible to use the
    // <src>cbegin</src> and <src>cend</src> functions which are
    // about 10% faster.
    // <group name=iterator-typedefs>
    // STL-style typedefs.
    // <group>
    
    // Element type
    typedef T value_type;
    // TODO This is how std containers define a reference type, but
    // the name 'reference' is already taken by a method.
    // typedef T&       reference;
    typedef const T& const_reference;
    // Pointer to an element type
    typedef T* pointer;
    // Constant pointer to the element type
    typedef const T* const_pointer;
    typedef IteratorSTL      iterator;
    typedef ConstIteratorSTL const_iterator;
    typedef T*               contiter;
    typedef const T*         const_contiter;
    // </group>
    // Get the begin iterator object for any array.
    // <group>
    iterator begin()
        { return iterator (*this); }
    const_iterator begin() const
        { return const_iterator (*this); }
    iterator end()
        { return iterator(end_p); }
    const_iterator end() const
        { return const_iterator(end_p); }
    // </group>
 
    // Get the begin iterator object for a contiguous array.
    // <group>
    contiter cbegin()
        { return begin_p; }
    const_contiter cbegin() const
        { return begin_p; }
    contiter cend()
        { return end_p; }
    const_contiter cend() const
        { return end_p; }
    // </group>
 
    // </group>
 
 
private:
  // Implementation of constructor taking a Shape, a Templated parameter and an allocator.
  // This method implements it for when T is integral, in which case the templated parameter
  // is the initial value.
  template<typename Integral>
  Array(const IPosition &shape, Integral startIter, std::true_type /*is_integral*/ );
  
  // Implementation of constructor taking a Shape, a Templated parameter and an allocator.
  // This method implements it for when T is NOT integral, in which case the templated parameter
  // is an iterator.
  template<typename InputIterator>
  Array(const IPosition &shape, InputIterator startIter, std::false_type /*is_integral*/ );
  
  // Implementation for assign for copyable types
  Array<T>& assign_conforming_implementation (const Array<T>& other, std::true_type);
  // Implementation for assign for non-copyable types: can not be assigned
  Array<T>& assign_conforming_implementation (const Array<T>&, std::false_type)
  {
    throw ArrayError("Can not assign from non-copyable object");
  }
  static void copyToContiguousStorage(T *dst, Array<T> const& src, std::true_type);
  static void copyToContiguousStorage(T*, Array<T> const&, std::false_type)
  {
    throw ArrayError("Can not coy from non-copyable object");      
  }
  
  // An Array is unique when the container is shared and when nrefs==1.
  bool isUnique() const
  {
    return !data_p->is_shared() && nrefs()==1;
  }
    
protected:
     // Source will be empty with given shape after this call.
    Array(Array<T>&& source, const IPosition& shapeForSource) noexcept;
    
    template<typename ST>
    friend void swap(Array<ST>& left, Array<ST>& right);
    
    // Swap this array with another array.
    // Normally, casacore::swap() should be used instead.
    void swap(Array<T>& other);
 
    // pre/post processing hook of takeStorage() for subclasses.
    virtual void preTakeStorage(const IPosition &) {}
    
    virtual void postTakeStorage() {}
    
    // This function is called when this array is about to be resized, before
    // any work is done. Subclasses can throw an exception if the size doesn't
    // match, e.g. if a Matrix is resized to have 3 dimensions.
    // Before this function existed, assign-like functions had to be virtual.
    // However, for non-copyable types, assign can't exist. This is fine for
    // non-virtual methods (they just can't be called), but virtual methods
    // cause the who class to fail to be instantiatable.
    void checkBeforeResize(const IPosition &newShape)
    {
      if(fixedDimensionality() !=0 && newShape.size() != fixedDimensionality())
        throw(ArrayNDimError(fixedDimensionality(), newShape.size(),
          std::string("Invalid size given to ") + typeid(*this).name() +
          ": should have dimensionality of " + std::to_string(fixedDimensionality())));
    }
 
    // Subclasses can return their dimensionality. The Array class will make sure
    // that whenever it is resized, the dimensionality is checked.
    // Array's constructors do not check the dimensionality, because the subclass
    // hasn't been created yet at that point. Subclasses should therefore make
    // sure to call the constructors appropriately.
    // For classes that return 0, any resize will be allowed.
    virtual size_t fixedDimensionality() const { return 0; }
    
    virtual void checkAssignableType(ArrayBase& arrayBase) const
    {
      const Array<T>* pa = dynamic_cast<const Array<T>*>(&arrayBase);
      if (pa == nullptr) {
        throw ArrayError("ArrayBase& has incorrect template type");
      }
    }
    
    static void copyToContiguousStorage(T *dst, Array<T> const& src)
    {
      copyToContiguousStorage(dst, src, std::is_copy_assignable<T>());
    }
 
    // Remove the degenerate axes from the Array object.
    // This is the implementation of the nonDegenerate functions.
    // It has a different name to be able to make it virtual without having
    // the "hide virtual function" message when compiling derived classes.
    virtual void doNonDegenerate(const Array<T> &other,
                                 const IPosition &ignoreAxes);
 
 
    // Shared pointer to a Storage that contains the data.
    std::shared_ptr<arrays_internal::Storage<T>> data_p;
 
    // This pointer is adjusted to point to the first element of the array.
    // It is not necessarily the same thing as data->storage() since
    // this array might be a section, e.g. have a blc which shifts us forward
    // into the block.
    T *begin_p;
 
    // The end for an STL-style iteration.
    T* end_p;
 
 
    // Fill the steps and the end for a derived class.
    void makeSteps()
      { baseMakeSteps(); this->setEndIter(); }
 
    // Set the end iterator.
    void setEndIter()
      { end_p = (nels_p==0 ? 0 : (contiguous_p  ?  begin_p + nels_p :
                   begin_p + size_t(length_p(ndim()-1)) * steps_p(ndim()-1))); }
};
 
// Swap the first array with the second.
// This is more efficient than std::swap()
template<typename T>
void swap(Array<T>& first, Array<T>& second)
{
  first.swap(second);
}
 
//# Declare extern templates for often used types.
extern template class Array<bool>;
extern template class Array<char>;
extern template class Array<unsigned char>;
extern template class Array<short>;
extern template class Array<unsigned short>;
extern template class Array<int>;
extern template class Array<long long>;
extern template class Array<float>;
extern template class Array<double>;
 
}//#End casa namespace
 
#include "ArrayStr.h"
#include "Array.tcc"
 
#endif