nmatrix-atlas 0.2.0

data/ext/nmatrix/storage/storage.h

@@ -0,0 +1,98 @@
+/////////////////////////////////////////////////////////////////////
+// = NMatrix
+//
+// A linear algebra library for scientific computation in Ruby.
+// NMatrix is part of SciRuby.
+//
+// NMatrix was originally inspired by and derived from NArray, by
+// Masahiro Tanaka: http://narray.rubyforge.org
+//
+// == Copyright Information
+//
+// SciRuby is Copyright (c) 2010 - 2014, Ruby Science Foundation
+// NMatrix is Copyright (c) 2012 - 2014, John Woods and the Ruby Science Foundation
+//
+// Please see LICENSE.txt for additional copyright notices.
+//
+// == Contributing
+//
+// By contributing source code to SciRuby, you agree to be bound by
+// our Contributor Agreement:
+//
+// * https://github.com/SciRuby/sciruby/wiki/Contributor-Agreement
+//
+// == storage.h
+//
+// This file brings together everything in the storage directory.  It should not
+// be included by anything in the storage directory, but should be included by
+// files needing to use the storage code.
+
+#ifndef STORAGE_H
+#define STORAGE_H
+
+/*
+ * Standard Includes
+ */
+
+#include <cstdlib>
+
+/*
+ * Project Includes
+ */
+
+#include "types.h"
+
+#include "data/data.h"
+
+#include "common.h"
+#include "dense/dense.h"
+#include "list/list.h"
+#include "yale/yale.h"
+
+/*
+ * Macros
+ */
+
+#define NMATRIX_DTYPE_IS_COMPLEX(s)		((s->dtype == nm::COMPLEX64) or (s->dtype == nm::COMPLEX128))
+#define NMATRIX_DTYPE_IS_FLOAT(s)			((s->dtype == nm::FLOAT32) or (s->dtype == nm::FLOAT64))
+#define NMATRIX_DTYPE_IS_INTEGER(s)		(s->dtype <= nm::INT64)
+#define NMATRIX_DTYPE_IS_RUBYOBJ(s)		(s->dtype == nm::RUBYOBJ)
+
+
+/*
+ * Types
+ */
+
+
+/*
+ * Data
+ */
+
+namespace nm {
+	const int NUM_STYPES = 3;
+}
+
+extern "C" {
+
+  extern const char* const STYPE_NAMES[nm::NUM_STYPES];
+  extern void (* const STYPE_MARK[nm::NUM_STYPES])(STORAGE*);
+
+  /*
+   * Functions
+   */
+
+  /////////////////////////
+  // Copying and Casting //
+  /////////////////////////
+
+  STORAGE*	  nm_dense_storage_from_list(const STORAGE* right, nm::dtype_t l_dtype, void*);
+  STORAGE*	  nm_dense_storage_from_yale(const STORAGE* right, nm::dtype_t l_dtype, void*);
+  STORAGE*		nm_list_storage_from_dense(const STORAGE* right, nm::dtype_t l_dtype, void*);
+  STORAGE*		nm_list_storage_from_yale(const STORAGE* right,  nm::dtype_t l_dtype, void*);
+  STORAGE*		nm_yale_storage_from_list(const STORAGE* right,  nm::dtype_t l_dtype, void*);
+  STORAGE*		nm_yale_storage_from_dense(const STORAGE* right, nm::dtype_t l_dtype, void*);
+
+} // end of extern "C" block
+
+
+#endif // STORAGE_H

data/ext/nmatrix/storage/yale/class.h

@@ -0,0 +1,1139 @@
+/////////////////////////////////////////////////////////////////////
+// = NMatrix
+//
+// A linear algebra library for scientific computation in Ruby.
+// NMatrix is part of SciRuby.
+//
+// NMatrix was originally inspired by and derived from NArray, by
+// Masahiro Tanaka: http://narray.rubyforge.org
+//
+// == Copyright Information
+//
+// SciRuby is Copyright (c) 2010 - 2014, Ruby Science Foundation
+// NMatrix is Copyright (c) 2012 - 2014, John Woods and the Ruby Science Foundation
+//
+// Please see LICENSE.txt for additional copyright notices.
+//
+// == Contributing
+//
+// By contributing source code to SciRuby, you agree to be bound by
+// our Contributor Agreement:
+//
+// * https://github.com/SciRuby/sciruby/wiki/Contributor-Agreement
+//
+// == class.h
+//
+// Object-oriented interface for Yale.
+//
+
+#ifndef YALE_CLASS_H
+# define YALE_CLASS_H
+
+#include "../dense/dense.h"
+#include "math/transpose.h"
+#include "yale.h"
+
+namespace nm {
+
+
+/*
+ * This class is basically an intermediary for YALE_STORAGE objects which enables us to treat it like a C++ object. It
+ * keeps the src pointer as its s, along with other relevant slice information.
+ *
+ * It's useful for creating iterators and such. It isn't responsible for allocating or freeing its YALE_STORAGE* pointers.
+ */
+
+template <typename D>
+class YaleStorage {
+public:
+  YaleStorage(const YALE_STORAGE* storage)
+   : s(reinterpret_cast<YALE_STORAGE*>(storage->src)),
+     slice(storage != storage->src),
+     slice_shape(storage->shape),
+     slice_offset(storage->offset)
+  {
+    nm_yale_storage_register(storage->src);
+  }
+
+  YaleStorage(const STORAGE* storage)
+   : s(reinterpret_cast<YALE_STORAGE*>(storage->src)),
+     slice(storage != storage->src),
+     slice_shape(storage->shape),
+     slice_offset(storage->offset)
+  {
+    nm_yale_storage_register(reinterpret_cast<STORAGE*>(storage->src));
+  }
+
+  ~YaleStorage() {
+    nm_yale_storage_unregister(s);
+  }
+
+  /* Allows us to do YaleStorage<uint8>::dtype() to get an nm::dtype_t */
+  static nm::dtype_t dtype() {
+    return nm::ctype_to_dtype_enum<D>::value_type;
+  }
+
+
+  bool is_ref() const { return slice; }
+
+  inline D* default_obj_ptr() { return &(a(s->shape[0])); }
+  inline D& default_obj() { return a(s->shape[0]); }
+  inline const D& default_obj() const { return a(s->shape[0]); }
+  inline const D& const_default_obj() const { return a(s->shape[0]); }
+
+
+  /*
+   * Return a Ruby VALUE representation of default_obj()
+   */
+  VALUE const_default_value() const {
+    return nm::yale_storage::nm_rb_dereference(a(s->shape[0]));
+  }
+
+  inline size_t* ija_p()       const       { return reinterpret_cast<size_t*>(s->ija); }
+  inline const size_t& ija(size_t p) const { return ija_p()[p]; }
+  inline size_t& ija(size_t p)             { return ija_p()[p]; }
+  inline D* a_p()         const       { return reinterpret_cast<D*>(s->a); }
+  inline const D& a(size_t p) const   { return a_p()[p]; }
+  inline D& a(size_t p)               { return a_p()[p]; }
+
+  bool real_row_empty(size_t i) const { return ija(i+1) - ija(i) == 0 ? true : false; }
+
+  inline size_t* shape_p()        const { return slice_shape;      }
+  inline size_t  shape(uint8_t d) const { return slice_shape[d];   }
+  inline size_t* real_shape_p() const { return s->shape;           }
+  inline size_t  real_shape(uint8_t d) const { return s->shape[d]; }
+  inline size_t* offset_p()     const { return slice_offset;       }
+  inline size_t  offset(uint8_t d) const { return slice_offset[d]; }
+  inline size_t  capacity() const { return s->capacity;            }
+  inline size_t  size() const { return ija(real_shape(0));         }
+
+
+  /*
+   * Returns true if the value at apos is the default value.
+   * Mainly used for determining if the diagonal contains zeros.
+   */
+  bool is_pos_default_value(size_t apos) const {
+    return (a(apos) == const_default_obj());
+  }
+
+  /*
+   * Given a size-2 array of size_t, representing the shape, determine
+   * the maximum size of YaleStorage arrays.
+   */
+  static size_t max_size(const size_t* shape) {
+    size_t result = shape[0] * shape[1] + 1;
+    if (shape[0] > shape[1])
+      result += shape[0] - shape[1];
+    return result;
+  }
+
+  /*
+   * Minimum size of Yale Storage arrays given some shape.
+   */
+  static size_t min_size(const size_t* shape) {
+    return shape[0]*2 + 1;
+  }
+
+  /*
+   * This is the guaranteed maximum size of the IJA/A arrays of the matrix given its shape.
+   */
+  inline size_t real_max_size() const {
+    return YaleStorage<D>::max_size(real_shape_p());
+  }
+
+  // Binary search between left and right in IJA for column ID real_j. Returns left if not found.
+  size_t real_find_pos(size_t left, size_t right, size_t real_j, bool& found) const {
+    if (left > right) {
+      found = false;
+      return left;
+    }
+
+    size_t mid   = (left + right) / 2;
+    size_t mid_j = ija(mid);
+
+    if (mid_j == real_j) {
+      found = true;
+      return mid;
+    } else if (mid_j > real_j)  return real_find_pos(left, mid - 1, real_j, found);
+    else                        return real_find_pos(mid + 1, right, real_j, found);
+  }
+
+  // Binary search between left and right in IJA for column ID real_j. Essentially finds where the slice should begin,
+  // with no guarantee that there's anything in there.
+  size_t real_find_left_boundary_pos(size_t left, size_t right, size_t real_j) const {
+    if (left > right) return right;
+    if (ija(left) >= real_j) return left;
+
+    size_t mid   = (left + right) / 2;
+    size_t mid_j = ija(mid);
+
+    if (mid_j == real_j)      return mid;
+    else if (mid_j > real_j)  return real_find_left_boundary_pos(left, mid, real_j);
+    else                      return real_find_left_boundary_pos(mid + 1, right, real_j);
+  }
+
+  // Binary search between left and right in IJA for column ID real_j. Essentially finds where the slice should begin,
+  // with no guarantee that there's anything in there.
+  size_t real_find_right_boundary_pos(size_t left, size_t right, size_t real_j) const {
+    if (left > right) return right;
+    if (ija(right) <= real_j) return right;
+
+    size_t mid   = (left + right) / 2;
+    size_t mid_j = ija(mid);
+
+    if (mid_j == real_j)      return mid;
+    else if (mid_j > real_j)  return real_find_right_boundary_pos(left, mid, real_j);
+    else                      return real_find_right_boundary_pos(mid + 1, right, real_j);
+  }
+
+
+  // Binary search for coordinates i,j in the slice. If not found, return -1.
+  std::pair<size_t,bool> find_pos(const std::pair<size_t,size_t>& ij) const {
+    size_t left   = ija(ij.first + offset(0));
+    size_t right  = ija(ij.first + offset(0) + 1) - 1;
+
+    std::pair<size_t, bool> result;
+    result.first = real_find_pos(left, right, ij.second + offset(1), result.second);
+    return result;
+  }
+
+  // Binary search for coordinates i,j in the slice, and return the first position >= j in row i.
+  size_t find_pos_for_insertion(size_t i, size_t j) const {
+    size_t left   = ija(i + offset(0));
+    size_t right  = ija(i + offset(0) + 1) - 1;
+
+    // Check that the right search point is valid. rflbp will check to make sure the left is valid relative to left.
+    if (right > ija(real_shape(0))) {
+      right = ija(real_shape(0))-1;
+    }
+    size_t result = real_find_left_boundary_pos(left, right, j + offset(1));
+    return result;
+  }
+
+  typedef yale_storage::basic_iterator_T<D,D,YaleStorage<D> >              basic_iterator;
+  typedef yale_storage::basic_iterator_T<D,const D,const YaleStorage<D> >  const_basic_iterator;
+
+  typedef yale_storage::stored_diagonal_iterator_T<D,D,YaleStorage<D> >              stored_diagonal_iterator;
+  typedef yale_storage::stored_diagonal_iterator_T<D,const D,const YaleStorage<D> >  const_stored_diagonal_iterator;
+
+  typedef yale_storage::iterator_T<D,D,YaleStorage<D> >                iterator;
+  typedef yale_storage::iterator_T<D,const D,const YaleStorage<D> >    const_iterator;
+
+
+  friend class yale_storage::row_iterator_T<D,D,YaleStorage<D> >;
+  typedef yale_storage::row_iterator_T<D,D,YaleStorage<D> >             row_iterator;
+  typedef yale_storage::row_iterator_T<D,const D,const YaleStorage<D> > const_row_iterator;
+
+  typedef yale_storage::row_stored_iterator_T<D,D,YaleStorage<D>,row_iterator>    row_stored_iterator;
+  typedef yale_storage::row_stored_nd_iterator_T<D,D,YaleStorage<D>,row_iterator> row_stored_nd_iterator;
+  typedef yale_storage::row_stored_iterator_T<D,const D,const YaleStorage<D>,const_row_iterator>       const_row_stored_iterator;
+  typedef yale_storage::row_stored_nd_iterator_T<D,const D,const YaleStorage<D>,const_row_iterator>    const_row_stored_nd_iterator;
+  typedef std::pair<row_iterator,row_stored_nd_iterator>                                               row_nd_iter_pair;
+
+  // Variety of iterator begin and end functions.
+  iterator begin(size_t row = 0)                      {      return iterator(*this, row);                 }
+  iterator row_end(size_t row)                        {      return begin(row+1);                         }
+  iterator end()                                      {      return iterator(*this, shape(0));            }
+  const_iterator cbegin(size_t row = 0) const         {      return const_iterator(*this, row);           }
+  const_iterator crow_end(size_t row) const           {      return cbegin(row+1);                        }
+  const_iterator cend() const                         {      return const_iterator(*this, shape(0));      }
+
+  stored_diagonal_iterator sdbegin(size_t d = 0)      {      return stored_diagonal_iterator(*this, d);   }
+  stored_diagonal_iterator sdend()                    {
+    return stored_diagonal_iterator(*this, std::min( shape(0) + offset(0), shape(1) + offset(1) ) - std::max(offset(0), offset(1)) );
+  }
+  const_stored_diagonal_iterator csdbegin(size_t d = 0) const { return const_stored_diagonal_iterator(*this, d); }
+  const_stored_diagonal_iterator csdend() const        {
+    return const_stored_diagonal_iterator(*this, std::min( shape(0) + offset(0), shape(1) + offset(1) ) - std::max(offset(0), offset(1)) );
+  }
+  row_iterator ribegin(size_t row = 0)                {      return row_iterator(*this, row);             }
+  row_iterator riend()                                {      return row_iterator(*this, shape(0));        }
+  const_row_iterator cribegin(size_t row = 0) const   {      return const_row_iterator(*this, row);       }
+  const_row_iterator criend() const                   {      return const_row_iterator(*this, shape(0));  }
+
+
+  /*
+   * Get a count of the ndnz in the slice as if it were its own matrix.
+   */
+  size_t count_copy_ndnz() const {
+    if (!slice) return s->ndnz; // easy way -- not a slice.
+    size_t count = 0;
+
+    // Visit all stored entries.
+    for (const_row_iterator it = cribegin(); it != criend(); ++it){
+      for (auto jt = it.begin(); jt != it.end(); ++jt) {
+        if (it.i() != jt.j() && *jt != const_default_obj()) ++count;
+      }
+    }
+
+    return count;
+  }
+
+  /*
+   * Returns the iterator for i,j or snd_end() if not found.
+   */
+/*  stored_nondiagonal_iterator find(const std::pair<size_t,size_t>& ij) {
+    std::pair<size_t,bool> find_pos_result = find_pos(ij);
+    if (!find_pos_result.second) return sndend();
+    else return stored_nondiagonal_iterator(*this, ij.first, find_pos_result.first);
+  } */
+
+  /*
+   * Returns a stored_nondiagonal_iterator pointing to the location where some coords i,j should go, or returns their
+   * location if present.
+   */
+  /*std::pair<row_iterator, row_stored_nd_iterator> lower_bound(const std::pair<size_t,size_t>& ij)  {
+    row_iterator it            = ribegin(ij.first);
+    row_stored_nd_iterator jt  = it.lower_bound(ij.second);
+    return std::make_pair(it,jt);
+  } */
+
+  class multi_row_insertion_plan {
+  public:
+    std::vector<size_t>   pos;
+    std::vector<int>      change;
+    int                   total_change; // the net change occurring
+    size_t                num_changes;  // the total number of rows that need to change size
+    multi_row_insertion_plan(size_t rows_in_slice) : pos(rows_in_slice), change(rows_in_slice), total_change(0), num_changes(0) { }
+
+    void add(size_t i, const std::pair<int,size_t>& change_and_pos) {
+      pos[i]        = change_and_pos.second;
+      change[i]     = change_and_pos.first;
+      total_change += change_and_pos.first;
+      if (change_and_pos.first != 0) num_changes++;
+    }
+  };
+
+
+  /*
+   * Find all the information we need in order to modify multiple rows.
+   */
+  multi_row_insertion_plan insertion_plan(row_iterator i, size_t j, size_t* lengths, D* const v, size_t v_size) const {
+    multi_row_insertion_plan p(lengths[0]);
+
+    // v_offset is our offset in the array v. If the user wants to change two elements in each of three rows,
+    // but passes an array of size 3, we need to know that the second insertion plan must start at position
+    // 2 instead of 0; and then the third must start at 1.
+    size_t v_offset = 0;
+    for (size_t m = 0; m < lengths[0]; ++m, ++i) {
+      p.add(m, i.single_row_insertion_plan(j, lengths[1], v, v_size, v_offset));
+    }
+
+    return p;
+  }
+
+
+
+  /*
+   * Insert entries in multiple rows. Slice-setting.
+   */
+  void insert(row_iterator i, size_t j, size_t* lengths, D* const v, size_t v_size) {
+    // Expensive pre-processing step: find all the information we need in order to do insertions.
+    multi_row_insertion_plan p = insertion_plan(i, j, lengths, v, v_size);
+
+    // There are more efficient ways to do this, but this is the low hanging fruit version of the algorithm.
+    // Here's the full problem: http://stackoverflow.com/questions/18753375/algorithm-for-merging-short-lists-into-a-long-vector
+    // --JW
+
+    bool resize = false;
+    size_t sz = size();
+    if (p.num_changes > 1) resize = true; // TODO: There are surely better ways to do this, but I've gone for the low-hanging fruit
+    else if (sz + p.total_change > capacity() || sz + p.total_change <= capacity() / nm::yale_storage::GROWTH_CONSTANT) resize = true;
+
+    if (resize) {
+      update_resize_move_insert(i.i() + offset(0), j + offset(1), lengths, v, v_size, p);
+    } else {
+
+      // Make the necessary modifications, which hopefully can be done in-place.
+      size_t v_offset = 0;
+      //int accum       = 0;
+      for (size_t ii = 0; ii < lengths[0]; ++ii, ++i) {
+        i.insert(row_stored_nd_iterator(i, p.pos[ii]), j, lengths[1], v, v_size, v_offset);
+      }
+    }
+  }
+
+
+  /*
+   * Most Ruby-centric insert function. Accepts coordinate information in slice,
+   * and value information of various types in +right+. This function must evaluate
+   * +right+ and determine what other functions to call in order to properly handle
+   * it.
+   */
+  void insert(SLICE* slice, VALUE right) {
+
+    NM_CONSERVATIVE(nm_register_value(&right));
+
+    std::pair<NMATRIX*,bool> nm_and_free =
+      interpret_arg_as_dense_nmatrix(right, dtype());
+    // Map the data onto D* v
+
+    D*     v;
+    size_t v_size = 1;
+
+    if (nm_and_free.first) {
+      DENSE_STORAGE* s = reinterpret_cast<DENSE_STORAGE*>(nm_and_free.first->storage);
+      v       = reinterpret_cast<D*>(s->elements);
+      v_size  = nm_storage_count_max_elements(s);
+
+    } else if (TYPE(right) == T_ARRAY) {
+      v_size = RARRAY_LEN(right);
+      v      = NM_ALLOC_N(D, v_size);
+      if (dtype() == nm::RUBYOBJ) {
+       nm_register_values(reinterpret_cast<VALUE*>(v), v_size);
+      }
+      for (size_t m = 0; m < v_size; ++m) {
+        rubyval_to_cval(rb_ary_entry(right, m), s->dtype, &(v[m]));
+      }
+      if (dtype() == nm::RUBYOBJ) {
+       nm_unregister_values(reinterpret_cast<VALUE*>(v), v_size);
+      }
+
+    } else {
+      v = reinterpret_cast<D*>(rubyobj_to_cval(right, dtype()));
+    }
+
+    row_iterator i = ribegin(slice->coords[0]);
+
+    if (slice->single || (slice->lengths[0] == 1 && slice->lengths[1] == 1)) { // single entry
+      i.insert(slice->coords[1], *v);
+    } else if (slice->lengths[0] == 1) { // single row, multiple entries
+      i.insert(slice->coords[1], slice->lengths[1], v, v_size);
+    } else { // multiple rows, unknown number of entries
+      insert(i, slice->coords[1], slice->lengths, v, v_size);
+    }
+
+    // Only free v if it was allocated in this function.
+    if (nm_and_free.first) {
+      if (nm_and_free.second) {
+        nm_delete(nm_and_free.first);
+      }
+    } else NM_FREE(v);
+
+    NM_CONSERVATIVE(nm_unregister_value(&right));
+  }
+
+
+  /*
+   * Remove an entry from an already found non-diagonal position.
+   */
+  row_iterator erase(row_iterator it, const row_stored_nd_iterator& position) {
+    it.erase(position);
+    return it;
+  }
+
+
+  /*
+   * Remove an entry from the matrix at the already-located position. If diagonal, just sets to default; otherwise,
+   * actually removes the entry.
+   */
+  row_iterator erase(row_iterator it, const row_stored_iterator& jt) {
+    it.erase((const row_stored_nd_iterator&)jt);
+    return it;
+  }
+
+
+  row_iterator insert(row_iterator it, row_stored_iterator position, size_t j, const D& val) {
+    it.insert(position, j, val);
+    return it;
+  }
+
+
+  /*
+   * Insert an element in column j, using position's p() as the location to insert the new column. i and j will be the
+   * coordinates. This also does a replace if column j is already present.
+   *
+   * Returns true if a new entry was added and false if an entry was replaced.
+   *
+   * Pre-conditions:
+   *   - position.p() must be between ija(real_i) and ija(real_i+1), inclusive, where real_i = i + offset(0)
+   *   - real_i and real_j must not be equal
+   */
+  row_iterator insert(row_iterator it, row_stored_nd_iterator position, size_t j, const D& val) {
+    it.insert(position, j, val);
+    return it;
+  }
+
+
+  /*
+   * Insert n elements v in columns j, using position as a guide. i gives the starting row. If at any time a value in j
+   * decreases,
+   */
+  /*bool insert(stored_iterator position, size_t n, size_t i, size_t* j, DType* v) {
+
+  } */
+
+  /*
+   * A pseudo-insert operation, since the diagonal portion of the A array is constant size.
+   */
+  stored_diagonal_iterator insert(stored_diagonal_iterator position, const D& val) {
+    *position = val;
+    return position;
+  }
+
+
+/*  iterator insert(iterator position, size_t j, const D& val) {
+    if (position.real_i() == position.real_j()) {
+      s->a(position.real_i()) = val;
+      return position;
+    } else {
+      row_iterator it = ribegin(position.i());
+      row_stored_nd_iterator position = it.ndbegin(j);
+      return insert(it, position, j, val);
+    }
+  }*/
+
+
+
+
+  /*
+   * Returns a pointer to the location of some entry in the matrix.
+   *
+   * This is needed for backwards compatibility. We don't really want anyone
+   * to modify the contents of that pointer, because it might be the ZERO location.
+   *
+   * TODO: Change all storage_get functions to return a VALUE once we've put list and
+   * dense in OO mode. ???
+   */
+  inline D* get_single_p(SLICE* slice) {
+    size_t real_i = offset(0) + slice->coords[0],
+           real_j = offset(1) + slice->coords[1];
+
+    if (real_i == real_j)
+      return &(a(real_i));
+
+    if (ija(real_i) == ija(real_i+1))
+      return default_obj_ptr(); // zero pointer
+
+    // binary search for a column's location
+    std::pair<size_t,bool> p = find_pos(std::make_pair(slice->coords[0], slice->coords[1]));
+    if (p.second)
+      return &(a(p.first));
+                       // not found: return default
+    return default_obj_ptr(); // zero pointer
+  }
+
+
+  /*
+   * Allocate a reference pointing to s. Note that even if +this+ is a reference,
+   * we can create a reference within it.
+   *
+   * Note: Make sure you NM_FREE() the result of this call. You can't just cast it
+   * directly into a YaleStorage<D> class.
+   */
+  YALE_STORAGE* alloc_ref(SLICE* slice) {
+    YALE_STORAGE* ns  = NM_ALLOC( YALE_STORAGE );
+
+    ns->dim           = s->dim;
+    ns->offset        = NM_ALLOC_N(size_t, ns->dim);
+    ns->shape         = NM_ALLOC_N(size_t, ns->dim);
+
+    for (size_t d = 0; d < ns->dim; ++d) {
+      ns->offset[d]   = slice->coords[d]  + offset(d);
+      ns->shape[d]    = slice->lengths[d];
+    }
+
+    ns->dtype         = s->dtype;
+    ns->a             = a_p();
+    ns->ija           = ija_p();
+
+    ns->src           = s;
+    s->count++;
+
+    ns->ndnz          = 0;
+    ns->capacity      = 0;
+
+    return ns;
+  }
+
+
+  /*
+   * Allocates and initializes the basic struct (but not IJA or A vectors).
+   */
+  static YALE_STORAGE* alloc(size_t* shape, size_t dim = 2) {
+    YALE_STORAGE* s = NM_ALLOC( YALE_STORAGE );
+
+    s->ndnz         = 0;
+    s->dtype        = dtype();
+    s->shape        = shape;
+    s->offset       = NM_ALLOC_N(size_t, dim);
+    for (size_t d = 0; d < dim; ++d)
+      s->offset[d]  = 0;
+    s->dim          = dim;
+    s->src          = reinterpret_cast<STORAGE*>(s);
+    s->count        = 1;
+
+    return s;
+  }
+
+
+  /*
+   * Create basic storage of same dtype as YaleStorage<D>. Allocates it,
+   * reserves necessary space, but doesn't fill structure at all.
+   */
+  static YALE_STORAGE* create(size_t* shape, size_t reserve) {
+
+    YALE_STORAGE* s = alloc( shape, 2 );
+    size_t max_sz   = YaleStorage<D>::max_size(shape),
+           min_sz   = YaleStorage<D>::min_size(shape);
+
+    if (reserve < min_sz) {
+      s->capacity = min_sz;
+    } else if (reserve > max_sz) {
+      s->capacity = max_sz;
+    } else {
+      s->capacity = reserve;
+    }
+
+    s->ija = NM_ALLOC_N( size_t, s->capacity );
+    s->a   = NM_ALLOC_N( D,      s->capacity );
+
+    return s;
+  }
+
+
+  /*
+   * Clear out the D portion of the A vector (clearing the diagonal and setting
+   * the zero value).
+   */
+  static void clear_diagonal_and_zero(YALE_STORAGE& s, D* init_val = NULL) {
+    D* a  = reinterpret_cast<D*>(s.a);
+
+    // Clear out the diagonal + one extra entry
+    if (init_val) {
+      for (size_t i = 0; i <= s.shape[0]; ++i)
+        a[i] = *init_val;
+    } else {
+      for (size_t i = 0; i <= s.shape[0]; ++i)
+        a[i] = 0;
+    }
+  }
+
+
+  /*
+   * Empty the matrix by initializing the IJA vector and setting the diagonal to 0.
+   *
+   * Called when most YALE_STORAGE objects are created.
+   *
+   * Can't go inside of class YaleStorage because YaleStorage creation requires that
+   * IJA already be initialized.
+   */
+  static void init(YALE_STORAGE& s, D* init_val) {
+    size_t IA_INIT = s.shape[0] + 1;
+    for (size_t m = 0; m < IA_INIT; ++m) {
+      s.ija[m] = IA_INIT;
+    }
+
+    clear_diagonal_and_zero(s, init_val);
+  }
+
+
+  /*
+   * Make a very basic allocation. No structure or copy or anything. It'll be shaped like this
+   * matrix.
+   *
+   * TODO: Combine this with ::create()'s ::alloc(). These are redundant.
+   */
+   template <typename E>
+   YALE_STORAGE* alloc_basic_copy(size_t new_capacity, size_t new_ndnz) const {
+     nm::dtype_t new_dtype = nm::ctype_to_dtype_enum<E>::value_type;
+     YALE_STORAGE* lhs     = NM_ALLOC( YALE_STORAGE );
+     lhs->dim              = s->dim;
+     lhs->shape            = NM_ALLOC_N( size_t, lhs->dim );
+
+     lhs->shape[0]         = shape(0);
+     lhs->shape[1]         = shape(1);
+
+     lhs->offset           = NM_ALLOC_N( size_t, lhs->dim );
+
+     lhs->offset[0]        = 0;
+     lhs->offset[1]        = 0;
+
+     lhs->capacity         = new_capacity;
+     lhs->dtype            = new_dtype;
+     lhs->ndnz             = new_ndnz;
+     lhs->ija              = NM_ALLOC_N( size_t, new_capacity );
+     lhs->a                = NM_ALLOC_N( E,      new_capacity );
+     lhs->src              = lhs;
+     lhs->count            = 1;
+
+     return lhs;
+   }
+
+
+  /*
+   * Make a full matrix structure copy (entries remain uninitialized). Remember to NM_FREE()!
+   */
+  template <typename E>
+  YALE_STORAGE* alloc_struct_copy(size_t new_capacity) const {
+    YALE_STORAGE* lhs     = alloc_basic_copy<E>(new_capacity, count_copy_ndnz());
+    // Now copy the IJA contents
+    if (slice) {
+      rb_raise(rb_eNotImpError, "cannot copy struct due to different offsets");
+    } else {
+      for (size_t m = 0; m < size(); ++m) {
+        lhs->ija[m] = ija(m); // copy indices
+      }
+    }
+    return lhs;
+  }
+
+
+  /*
+   * Copy this slice (or the full matrix if it isn't a slice) into a new matrix which is already allocated, ns.
+   */
+  template <typename E, bool Yield=false>
+  void copy(YALE_STORAGE& ns) const {
+    //nm::dtype_t new_dtype = nm::ctype_to_dtype_enum<E>::value_type;
+    // get the default value for initialization (we'll re-use val for other copies after this)
+    E val = static_cast<E>(const_default_obj());
+
+    // initialize the matrix structure and clear the diagonal so we don't have to
+    // keep track of unwritten entries.
+    YaleStorage<E>::init(ns, &val);
+
+    E* ns_a    = reinterpret_cast<E*>(ns.a);
+    size_t sz  = shape(0) + 1; // current used size of ns
+    nm_yale_storage_register(&ns);
+
+    // FIXME: If diagonals line up, it's probably faster to do this with stored diagonal and stored non-diagonal iterators
+    for (const_row_iterator it = cribegin(); it != criend(); ++it) {
+      for (auto jt = it.begin(); !jt.end(); ++jt) {
+        if (it.i() == jt.j()) {
+          if (Yield)  ns_a[it.i()] = rb_yield(~jt);
+          else        ns_a[it.i()] = static_cast<E>(*jt);
+        } else if (*jt != const_default_obj()) {
+          if (Yield)  ns_a[sz]     = rb_yield(~jt);
+          else        ns_a[sz]     = static_cast<E>(*jt);
+          ns.ija[sz]    = jt.j();
+          ++sz;
+        }
+      }
+      ns.ija[it.i()+1]  = sz;
+    }
+    nm_yale_storage_unregister(&ns);
+
+    //ns.ija[shape(0)] = sz;                // indicate end of last row
+    ns.ndnz          = sz - shape(0) - 1; // update ndnz count
+  }
+
+
+  /*
+   * Allocate a casted copy of this matrix/reference. Remember to NM_FREE() the result!
+   *
+   * If Yield is true, E must be nm::RubyObject, and it will call an rb_yield upon the stored value.
+   */
+  template <typename E, bool Yield = false>
+  YALE_STORAGE* alloc_copy() const {
+    //nm::dtype_t new_dtype = nm::ctype_to_dtype_enum<E>::value_type;
+
+    YALE_STORAGE* lhs;
+    if (slice) {
+      size_t* xshape    = NM_ALLOC_N(size_t, 2);
+      xshape[0]         = shape(0);
+      xshape[1]         = shape(1);
+      size_t ndnz       = count_copy_ndnz();
+      size_t reserve    = shape(0) + ndnz + 1;
+
+//      std::cerr << "reserve = " << reserve << std::endl;
+
+      lhs               = YaleStorage<E>::create(xshape, reserve);
+
+      // FIXME: This should probably be a throw which gets caught outside of the object.
+      if (lhs->capacity < reserve)
+        rb_raise(nm_eStorageTypeError, "conversion failed; capacity of %lu requested, max allowable is %lu", reserve, lhs->capacity);
+
+      // Fill lhs with what's in our current matrix.
+      copy<E, Yield>(*lhs);
+    } else {
+      // Copy the structure and setup the IJA structure.
+      lhs               = alloc_struct_copy<E>(s->capacity);
+
+      E* la = reinterpret_cast<E*>(lhs->a);
+
+      nm_yale_storage_register(lhs);
+      for (size_t m = 0; m < size(); ++m) {
+        if (Yield) {
+	  la[m] = rb_yield(nm::yale_storage::nm_rb_dereference(a(m)));
+	}
+        else       la[m] = static_cast<E>(a(m));
+      }
+      nm_yale_storage_unregister(lhs);
+
+    }
+
+    return lhs;
+  }
+
+  /*
+   * Allocate a transposed copy of the matrix
+   */
+  /*
+   * Allocate a casted copy of this matrix/reference. Remember to NM_FREE() the result!
+   *
+   * If Yield is true, E must be nm::RubyObject, and it will call an rb_yield upon the stored value.
+   */
+  template <typename E, bool Yield = false>
+  YALE_STORAGE* alloc_copy_transposed() const {
+
+    if (slice) {
+      rb_raise(rb_eNotImpError, "please make a copy before transposing");
+    } else {
+      // Copy the structure and setup the IJA structure.
+      size_t* xshape    = NM_ALLOC_N(size_t, 2);
+      xshape[0]         = shape(1);
+      xshape[1]         = shape(0);
+
+      // Take a stab at the number of non-diagonal stored entries we'll have.
+      size_t reserve    = size() - xshape[1] + xshape[0];
+      YALE_STORAGE* lhs = YaleStorage<E>::create(xshape, reserve);
+      E r_init          = static_cast<E>(const_default_obj());
+      YaleStorage<E>::init(*lhs, &r_init);
+
+      nm::yale_storage::transpose_yale<D,E,true,true>(shape(0), shape(1), ija_p(), ija_p(), a_p(), const_default_obj(),
+                                                      lhs->ija, lhs->ija, reinterpret_cast<E*>(lhs->a), r_init);
+      return lhs;
+    }
+
+    return NULL;
+  }
+
+
+  /*
+   * Comparison between two matrices. Does not check size and such -- assumption is that they are the same shape.
+   */
+  template <typename E>
+  bool operator==(const YaleStorage<E>& rhs) const {
+    for (size_t i = 0; i < shape(0); ++i) {
+      typename YaleStorage<D>::const_row_iterator li = cribegin(i);
+      typename YaleStorage<E>::const_row_iterator ri = rhs.cribegin(i);
+
+      size_t j = 0; // keep track of j so we can compare different defaults
+
+      auto lj = li.begin();
+      auto rj = ri.begin();
+      while (!lj.end() || !rj.end()) {
+        if (lj < rj) {
+          if (*lj != rhs.const_default_obj()) return false;
+          ++lj;
+        } else if (rj < lj) {
+          if (const_default_obj() != *rj)     return false;
+          ++rj;
+        } else { // rj == lj
+          if (*lj != *rj) return false;
+          ++lj;
+          ++rj;
+        }
+        ++j;
+      }
+
+      // if we skip an entry (because it's an ndnz in BOTH matrices), we need to compare defaults.
+      // (We know we skipped if lj and rj hit end before j does.)
+      if (j < shape(1) && const_default_obj() != rhs.const_default_obj()) return false;
+
+      ++li;
+      ++ri;
+    }
+
+    return true;
+  }
+
+  /*
+   * Necessary for element-wise operations. The return dtype will be nm::RUBYOBJ.
+   */
+  template <typename E>
+  VALUE map_merged_stored(VALUE klass, nm::YaleStorage<E>& t, VALUE r_init) const {
+    nm_register_value(&r_init);
+    VALUE s_init    = const_default_value(),
+          t_init    = t.const_default_value();
+    nm_register_value(&s_init);
+    nm_register_value(&t_init);
+    
+    // Make a reasonable approximation of the resulting capacity
+    size_t s_ndnz   = count_copy_ndnz(),
+           t_ndnz   = t.count_copy_ndnz();
+    size_t reserve  = shape(0) + std::max(s_ndnz, t_ndnz) + 1;
+
+    size_t* xshape  = NM_ALLOC_N(size_t, 2);
+    xshape[0]       = shape(0);
+    xshape[1]       = shape(1);
+
+    YALE_STORAGE* rs= YaleStorage<nm::RubyObject>::create(xshape, reserve);
+
+    if (r_init == Qnil) {
+      nm_unregister_value(&r_init);
+      r_init       = rb_yield_values(2, s_init, t_init);
+      nm_register_value(&r_init);
+    }
+
+    nm::RubyObject r_init_obj(r_init);
+
+    // Prepare the matrix structure
+    YaleStorage<nm::RubyObject>::init(*rs, &r_init_obj);
+    NMATRIX* m     = nm_create(nm::YALE_STORE, reinterpret_cast<STORAGE*>(rs));
+    nm_register_nmatrix(m);
+    VALUE result   = Data_Wrap_Struct(klass, nm_mark, nm_delete, m);
+    nm_unregister_nmatrix(m);
+    nm_register_value(&result);
+    nm_unregister_value(&r_init);
+
+    RETURN_SIZED_ENUMERATOR_PRE
+    nm_unregister_value(&result);
+    nm_unregister_value(&t_init);
+    nm_unregister_value(&s_init);
+    // No obvious, efficient way to pass a length function as the fourth argument here:
+    RETURN_SIZED_ENUMERATOR(result, 0, 0, 0);
+
+    // Create an object for us to iterate over.
+    YaleStorage<nm::RubyObject> r(rs);
+
+    // Walk down our new matrix, inserting values as we go.
+    for (size_t i = 0; i < xshape[0]; ++i) {
+      YaleStorage<nm::RubyObject>::row_iterator   ri = r.ribegin(i);
+      typename YaleStorage<D>::const_row_iterator si = cribegin(i);
+      typename YaleStorage<E>::const_row_iterator ti = t.cribegin(i);
+
+      auto sj = si.begin();
+      auto tj = ti.begin();
+      auto rj = ri.ndbegin();
+
+      while (sj != si.end() || tj != ti.end()) {
+        VALUE  v;
+        size_t j;
+
+        if (sj < tj) {
+          v = rb_yield_values(2, ~sj, t_init);
+          j = sj.j();
+          ++sj;
+        } else if (tj < sj) {
+          v = rb_yield_values(2, s_init, ~tj);
+          j = tj.j();
+          ++tj;
+        } else {
+          v = rb_yield_values(2, ~sj, ~tj);
+          j = sj.j();
+          ++sj;
+          ++tj;
+        }
+
+        // FIXME: This can be sped up by inserting all at the same time
+        // since it's a new matrix. But that function isn't quite ready
+        // yet.
+        if (j == i) r.a(i) = v;
+        else        rj     = ri.insert(rj, j, v);
+        //RB_P(rb_funcall(result, rb_intern("yale_ija"), 0));
+      }
+    }
+    nm_unregister_value(&result);
+    nm_unregister_value(&t_init);
+    nm_unregister_value(&s_init);
+
+    return result;
+  }
+
+protected:
+  /*
+   * Update row sizes starting with row i
+   */
+  void update_real_row_sizes_from(size_t real_i, int change) {
+    ++real_i;
+    for (; real_i <= real_shape(0); ++real_i) {
+      ija(real_i) += change;
+    }
+  }
+
+
+  /*
+   * Like move_right, but also involving a resize. This updates row sizes as well. This version also takes a plan for
+   * multiple rows, and tries to do them all in one copy. It's used for multi-row slice-setting.
+   *
+   * This also differs from update_resize_move in that it resizes to the exact requested size instead of reserving space.
+   */
+  void update_resize_move_insert(size_t real_i, size_t real_j, size_t* lengths, D* const v, size_t v_size, multi_row_insertion_plan p) {
+    size_t sz      = size(); // current size of the storage vectors
+    size_t new_cap = sz + p.total_change;
+
+    if (new_cap > real_max_size()) {
+      NM_FREE(v);
+      rb_raise(rb_eStandardError, "resize caused by insertion of size %d (on top of current size %lu) would have caused yale matrix size to exceed its maximum (%lu)", p.total_change, sz, real_max_size());
+    }
+
+    if (s->dtype == nm::RUBYOBJ) {
+      nm_register_values(reinterpret_cast<VALUE*>(v), v_size);
+    }
+
+    size_t* new_ija     = NM_ALLOC_N( size_t,new_cap );
+    D* new_a            = NM_ALLOC_N( D,     new_cap );
+
+    // Copy unchanged row pointers first.
+    size_t m = 0;
+    for (; m <= real_i; ++m) {
+      new_ija[m]        = ija(m);
+      new_a[m]          = a(m);
+    }
+
+    // Now copy unchanged locations in IJA and A.
+    size_t q = real_shape(0)+1; // q is the copy-to position.
+    size_t r = real_shape(0)+1; // r is the copy-from position.
+    for (; r < p.pos[0]; ++r, ++q) {
+      new_ija[q]        = ija(r);
+      new_a[q]          = a(r);
+    }
+
+    // For each pos and change in the slice, copy the information prior to the insertion point. Then insert the necessary
+    // information.
+    size_t v_offset = 0;
+    int accum = 0; // keep track of the total change as we go so we can update row information.
+    for (size_t i = 0; i < lengths[0]; ++i, ++m) {
+      for (; r < p.pos[i]; ++r, ++q) {
+        new_ija[q]      = ija(r);
+        new_a[q]        = a(r);
+      }
+
+      // Insert slice data for a single row.
+      for (size_t j = 0; j < lengths[1]; ++j, ++v_offset) {
+        if (v_offset >= v_size) v_offset %= v_size;
+
+        if (j + real_j == i + real_i) { // modify diagonal
+          new_a[real_i + i] = v[v_offset];
+        } else if (v[v_offset] != const_default_obj()) {
+          new_ija[q]        = j + real_j;
+          new_a[q]          = v[v_offset];
+          ++q; // move on to next q location
+        }
+
+        if (r < ija(real_shape(0)) && ija(r) == j + real_j) ++r; // move r forward if the column matches.
+      }
+
+      // Update the row pointer for the current row.
+      accum                += p.change[i];
+      new_ija[m]            = ija(m) + accum;
+      new_a[m]              = a(m); // copy diagonal for this row
+    }
+
+    // Now copy everything subsequent to the last insertion point.
+    for (; r < size(); ++r, ++q) {
+      new_ija[q]            = ija(r);
+      new_a[q]              = a(r);
+    }
+
+    // Update the remaining row pointers and copy remaining diagonals
+    for (; m <= real_shape(0); ++m) {
+      new_ija[m]            = ija(m) + accum;
+      new_a[m]              = a(m);
+    }
+
+    s->capacity = new_cap;
+
+    NM_FREE(s->ija);
+    NM_FREE(s->a);
+
+    if (s->dtype == nm::RUBYOBJ) {
+      nm_unregister_values(reinterpret_cast<VALUE*>(v), v_size);
+    }   
+
+    s->ija      = new_ija;
+    s->a        = reinterpret_cast<void*>(new_a);
+  }
+
+
+
+
+  /*
+   * Like move_right, but also involving a resize. This updates row sizes as well.
+   */
+  void update_resize_move(row_stored_nd_iterator position, size_t real_i, int n) {
+    size_t sz      = size(); // current size of the storage vectors
+    size_t new_cap = n > 0 ? capacity() * nm::yale_storage::GROWTH_CONSTANT
+                           : capacity() / nm::yale_storage::GROWTH_CONSTANT;
+    size_t max_cap = real_max_size();
+
+    if (new_cap > max_cap) {
+      new_cap = max_cap;
+      if (sz + n > max_cap)
+        rb_raise(rb_eStandardError, "resize caused by insertion/deletion of size %d (on top of current size %lu) would have caused yale matrix size to exceed its maximum (%lu)", n, sz, real_max_size());
+    }
+
+    if (new_cap < sz + n) new_cap = sz + n;
+
+    size_t* new_ija     = NM_ALLOC_N( size_t,new_cap );
+    D* new_a            = NM_ALLOC_N( D,     new_cap );
+
+    // Copy unchanged row pointers first.
+    for (size_t m = 0; m <= real_i; ++m) {
+      new_ija[m]        = ija(m);
+      new_a[m]          = a(m);
+    }
+
+    // Now update row pointers following the changed row as we copy the additional values.
+    for (size_t m = real_i + 1; m <= real_shape(0); ++m) {
+      new_ija[m]        = ija(m) + n;
+      new_a[m]          = a(m);
+    }
+
+    // Copy all remaining prior to insertion/removal site
+    for (size_t m = real_shape(0) + 1; m < position.p(); ++m) {
+      new_ija[m]        = ija(m);
+      new_a[m]          = a(m);
+    }
+
+    // Copy all subsequent to insertion/removal site
+    size_t m = position.p();
+    if (n < 0) m -= n;
+
+    for (; m < sz; ++m) {
+      new_ija[m+n]      = ija(m);
+      new_a[m+n]        = a(m);
+    }
+
+    if (s->dtype == nm::RUBYOBJ) {
+      nm_yale_storage_register_a(new_a, new_cap);
+    }
+
+    s->capacity = new_cap;
+
+    NM_FREE(s->ija);
+    NM_FREE(s->a);
+
+    if (s->dtype == nm::RUBYOBJ) {
+      nm_yale_storage_unregister_a(new_a, new_cap);
+    }
+
+    s->ija      = new_ija;
+    s->a        = reinterpret_cast<void*>(new_a);
+  }
+
+
+  /*
+   * Move elements in the IJA and A arrays by n (to the right).
+   * Does not update row sizes.
+   */
+  void move_right(row_stored_nd_iterator position, size_t n) {
+    size_t sz = size();
+    for (size_t m = 0; m < sz - position.p(); ++m) {
+      ija(sz+n-1-m) = ija(sz-1-m);
+      a(sz+n-1-m)   = a(sz-1-m);
+    }
+  }
+
+  /*
+   * Move elements in the IJA and A arrays by n (to the left). Here position gives
+   * the location to move to, and they should come from n to the right.
+   */
+  void move_left(row_stored_nd_iterator position, size_t n) {
+    size_t sz = size();
+    for (size_t m = position.p() + n; m < sz; ++m) {   // work backwards
+      ija(m-n)      = ija(m);
+      a(m-n)        = a(m);
+    }
+  }
+
+  YALE_STORAGE* s;
+  bool          slice;
+  size_t*       slice_shape;
+  size_t*       slice_offset;
+};
+
+} // end of nm namespace
+
+#endif // YALE_CLASS_H