• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // This file is part of Eigen, a lightweight C++ template library
2 // for linear algebra.
3 //
4 // Copyright (C) 2008 Gael Guennebaud <gael.guennebaud@inria.fr>
5 //
6 // This Source Code Form is subject to the terms of the Mozilla
7 // Public License v. 2.0. If a copy of the MPL was not distributed
8 // with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
9 
10 #ifndef EIGEN_RANDOMSETTER_H
11 #define EIGEN_RANDOMSETTER_H
12 
13 namespace Eigen {
14 
15 /** Represents a std::map
16   *
17   * \see RandomSetter
18   */
19 template<typename Scalar> struct StdMapTraits
20 {
21   typedef int KeyType;
22   typedef std::map<KeyType,Scalar> Type;
23   enum {
24     IsSorted = 1
25   };
26 
setInvalidKeyStdMapTraits27   static void setInvalidKey(Type&, const KeyType&) {}
28 };
29 
30 #ifdef EIGEN_UNORDERED_MAP_SUPPORT
31 /** Represents a std::unordered_map
32   *
33   * To use it you need to both define EIGEN_UNORDERED_MAP_SUPPORT and include the unordered_map header file
34   * yourself making sure that unordered_map is defined in the std namespace.
35   *
36   * For instance, with current version of gcc you can either enable C++0x standard (-std=c++0x) or do:
37   * \code
38   * #include <tr1/unordered_map>
39   * #define EIGEN_UNORDERED_MAP_SUPPORT
40   * namespace std {
41   *   using std::tr1::unordered_map;
42   * }
43   * \endcode
44   *
45   * \see RandomSetter
46   */
47 template<typename Scalar> struct StdUnorderedMapTraits
48 {
49   typedef int KeyType;
50   typedef std::unordered_map<KeyType,Scalar> Type;
51   enum {
52     IsSorted = 0
53   };
54 
setInvalidKeyStdUnorderedMapTraits55   static void setInvalidKey(Type&, const KeyType&) {}
56 };
57 #endif // EIGEN_UNORDERED_MAP_SUPPORT
58 
59 #ifdef _DENSE_HASH_MAP_H_
60 /** Represents a google::dense_hash_map
61   *
62   * \see RandomSetter
63   */
64 template<typename Scalar> struct GoogleDenseHashMapTraits
65 {
66   typedef int KeyType;
67   typedef google::dense_hash_map<KeyType,Scalar> Type;
68   enum {
69     IsSorted = 0
70   };
71 
setInvalidKeyGoogleDenseHashMapTraits72   static void setInvalidKey(Type& map, const KeyType& k)
73   { map.set_empty_key(k); }
74 };
75 #endif
76 
77 #ifdef _SPARSE_HASH_MAP_H_
78 /** Represents a google::sparse_hash_map
79   *
80   * \see RandomSetter
81   */
82 template<typename Scalar> struct GoogleSparseHashMapTraits
83 {
84   typedef int KeyType;
85   typedef google::sparse_hash_map<KeyType,Scalar> Type;
86   enum {
87     IsSorted = 0
88   };
89 
setInvalidKeyGoogleSparseHashMapTraits90   static void setInvalidKey(Type&, const KeyType&) {}
91 };
92 #endif
93 
94 /** \class RandomSetter
95   *
96   * \brief The RandomSetter is a wrapper object allowing to set/update a sparse matrix with random access
97   *
98   * \param SparseMatrixType the type of the sparse matrix we are updating
99   * \param MapTraits a traits class representing the map implementation used for the temporary sparse storage.
100   *                  Its default value depends on the system.
101   * \param OuterPacketBits defines the number of rows (or columns) manage by a single map object
102   *                        as a power of two exponent.
103   *
104   * This class temporarily represents a sparse matrix object using a generic map implementation allowing for
105   * efficient random access. The conversion from the compressed representation to a hash_map object is performed
106   * in the RandomSetter constructor, while the sparse matrix is updated back at destruction time. This strategy
107   * suggest the use of nested blocks as in this example:
108   *
109   * \code
110   * SparseMatrix<double> m(rows,cols);
111   * {
112   *   RandomSetter<SparseMatrix<double> > w(m);
113   *   // don't use m but w instead with read/write random access to the coefficients:
114   *   for(;;)
115   *     w(rand(),rand()) = rand;
116   * }
117   * // when w is deleted, the data are copied back to m
118   * // and m is ready to use.
119   * \endcode
120   *
121   * Since hash_map objects are not fully sorted, representing a full matrix as a single hash_map would
122   * involve a big and costly sort to update the compressed matrix back. To overcome this issue, a RandomSetter
123   * use multiple hash_map, each representing 2^OuterPacketBits columns or rows according to the storage order.
124   * To reach optimal performance, this value should be adjusted according to the average number of nonzeros
125   * per rows/columns.
126   *
127   * The possible values for the template parameter MapTraits are:
128   *  - \b StdMapTraits: corresponds to std::map. (does not perform very well)
129   *  - \b GnuHashMapTraits: corresponds to __gnu_cxx::hash_map (available only with GCC)
130   *  - \b GoogleDenseHashMapTraits: corresponds to google::dense_hash_map (best efficiency, reasonable memory consumption)
131   *  - \b GoogleSparseHashMapTraits: corresponds to google::sparse_hash_map (best memory consumption, relatively good performance)
132   *
133   * The default map implementation depends on the availability, and the preferred order is:
134   * GoogleSparseHashMapTraits, GnuHashMapTraits, and finally StdMapTraits.
135   *
136   * For performance and memory consumption reasons it is highly recommended to use one of
137   * the Google's hash_map implementation. To enable the support for them, you have two options:
138   *  - \#include <google/dense_hash_map> yourself \b before Eigen/Sparse header
139   *  - define EIGEN_GOOGLEHASH_SUPPORT
140   * In the later case the inclusion of <google/dense_hash_map> is made for you.
141   *
142   * \see http://code.google.com/p/google-sparsehash/
143   */
144 template<typename SparseMatrixType,
145          template <typename T> class MapTraits =
146 #if defined _DENSE_HASH_MAP_H_
147           GoogleDenseHashMapTraits
148 #elif defined _HASH_MAP
149           GnuHashMapTraits
150 #else
151           StdMapTraits
152 #endif
153          ,int OuterPacketBits = 6>
154 class RandomSetter
155 {
156     typedef typename SparseMatrixType::Scalar Scalar;
157     typedef typename SparseMatrixType::Index Index;
158 
159     struct ScalarWrapper
160     {
ScalarWrapperScalarWrapper161       ScalarWrapper() : value(0) {}
162       Scalar value;
163     };
164     typedef typename MapTraits<ScalarWrapper>::KeyType KeyType;
165     typedef typename MapTraits<ScalarWrapper>::Type HashMapType;
166     static const int OuterPacketMask = (1 << OuterPacketBits) - 1;
167     enum {
168       SwapStorage = 1 - MapTraits<ScalarWrapper>::IsSorted,
169       TargetRowMajor = (SparseMatrixType::Flags & RowMajorBit) ? 1 : 0,
170       SetterRowMajor = SwapStorage ? 1-TargetRowMajor : TargetRowMajor
171     };
172 
173   public:
174 
175     /** Constructs a random setter object from the sparse matrix \a target
176       *
177       * Note that the initial value of \a target are imported. If you want to re-set
178       * a sparse matrix from scratch, then you must set it to zero first using the
179       * setZero() function.
180       */
RandomSetter(SparseMatrixType & target)181     inline RandomSetter(SparseMatrixType& target)
182       : mp_target(&target)
183     {
184       const Index outerSize = SwapStorage ? target.innerSize() : target.outerSize();
185       const Index innerSize = SwapStorage ? target.outerSize() : target.innerSize();
186       m_outerPackets = outerSize >> OuterPacketBits;
187       if (outerSize&OuterPacketMask)
188         m_outerPackets += 1;
189       m_hashmaps = new HashMapType[m_outerPackets];
190       // compute number of bits needed to store inner indices
191       Index aux = innerSize - 1;
192       m_keyBitsOffset = 0;
193       while (aux)
194       {
195         ++m_keyBitsOffset;
196         aux = aux >> 1;
197       }
198       KeyType ik = (1<<(OuterPacketBits+m_keyBitsOffset));
199       for (Index k=0; k<m_outerPackets; ++k)
200         MapTraits<ScalarWrapper>::setInvalidKey(m_hashmaps[k],ik);
201 
202       // insert current coeffs
203       for (Index j=0; j<mp_target->outerSize(); ++j)
204         for (typename SparseMatrixType::InnerIterator it(*mp_target,j); it; ++it)
205           (*this)(TargetRowMajor?j:it.index(), TargetRowMajor?it.index():j) = it.value();
206     }
207 
208     /** Destructor updating back the sparse matrix target */
~RandomSetter()209     ~RandomSetter()
210     {
211       KeyType keyBitsMask = (1<<m_keyBitsOffset)-1;
212       if (!SwapStorage) // also means the map is sorted
213       {
214         mp_target->setZero();
215         mp_target->makeCompressed();
216         mp_target->reserve(nonZeros());
217         Index prevOuter = -1;
218         for (Index k=0; k<m_outerPackets; ++k)
219         {
220           const Index outerOffset = (1<<OuterPacketBits) * k;
221           typename HashMapType::iterator end = m_hashmaps[k].end();
222           for (typename HashMapType::iterator it = m_hashmaps[k].begin(); it!=end; ++it)
223           {
224             const Index outer = (it->first >> m_keyBitsOffset) + outerOffset;
225             const Index inner = it->first & keyBitsMask;
226             if (prevOuter!=outer)
227             {
228               for (Index j=prevOuter+1;j<=outer;++j)
229                 mp_target->startVec(j);
230               prevOuter = outer;
231             }
232             mp_target->insertBackByOuterInner(outer, inner) = it->second.value;
233           }
234         }
235         mp_target->finalize();
236       }
237       else
238       {
239         VectorXi positions(mp_target->outerSize());
240         positions.setZero();
241         // pass 1
242         for (Index k=0; k<m_outerPackets; ++k)
243         {
244           typename HashMapType::iterator end = m_hashmaps[k].end();
245           for (typename HashMapType::iterator it = m_hashmaps[k].begin(); it!=end; ++it)
246           {
247             const Index outer = it->first & keyBitsMask;
248             ++positions[outer];
249           }
250         }
251         // prefix sum
252         Index count = 0;
253         for (Index j=0; j<mp_target->outerSize(); ++j)
254         {
255           Index tmp = positions[j];
256           mp_target->outerIndexPtr()[j] = count;
257           positions[j] = count;
258           count += tmp;
259         }
260         mp_target->makeCompressed();
261         mp_target->outerIndexPtr()[mp_target->outerSize()] = count;
262         mp_target->resizeNonZeros(count);
263         // pass 2
264         for (Index k=0; k<m_outerPackets; ++k)
265         {
266           const Index outerOffset = (1<<OuterPacketBits) * k;
267           typename HashMapType::iterator end = m_hashmaps[k].end();
268           for (typename HashMapType::iterator it = m_hashmaps[k].begin(); it!=end; ++it)
269           {
270             const Index inner = (it->first >> m_keyBitsOffset) + outerOffset;
271             const Index outer = it->first & keyBitsMask;
272             // sorted insertion
273             // Note that we have to deal with at most 2^OuterPacketBits unsorted coefficients,
274             // moreover those 2^OuterPacketBits coeffs are likely to be sparse, an so only a
275             // small fraction of them have to be sorted, whence the following simple procedure:
276             Index posStart = mp_target->outerIndexPtr()[outer];
277             Index i = (positions[outer]++) - 1;
278             while ( (i >= posStart) && (mp_target->innerIndexPtr()[i] > inner) )
279             {
280               mp_target->valuePtr()[i+1] = mp_target->valuePtr()[i];
281               mp_target->innerIndexPtr()[i+1] = mp_target->innerIndexPtr()[i];
282               --i;
283             }
284             mp_target->innerIndexPtr()[i+1] = inner;
285             mp_target->valuePtr()[i+1] = it->second.value;
286           }
287         }
288       }
289       delete[] m_hashmaps;
290     }
291 
292     /** \returns a reference to the coefficient at given coordinates \a row, \a col */
operator()293     Scalar& operator() (Index row, Index col)
294     {
295       const Index outer = SetterRowMajor ? row : col;
296       const Index inner = SetterRowMajor ? col : row;
297       const Index outerMajor = outer >> OuterPacketBits; // index of the packet/map
298       const Index outerMinor = outer & OuterPacketMask;  // index of the inner vector in the packet
299       const KeyType key = (KeyType(outerMinor)<<m_keyBitsOffset) | inner;
300       return m_hashmaps[outerMajor][key].value;
301     }
302 
303     /** \returns the number of non zero coefficients
304       *
305       * \note According to the underlying map/hash_map implementation,
306       * this function might be quite expensive.
307       */
nonZeros()308     Index nonZeros() const
309     {
310       Index nz = 0;
311       for (Index k=0; k<m_outerPackets; ++k)
312         nz += static_cast<Index>(m_hashmaps[k].size());
313       return nz;
314     }
315 
316 
317   protected:
318 
319     HashMapType* m_hashmaps;
320     SparseMatrixType* mp_target;
321     Index m_outerPackets;
322     unsigned char m_keyBitsOffset;
323 };
324 
325 } // end namespace Eigen
326 
327 #endif // EIGEN_RANDOMSETTER_H
328