• 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 Guillaume Saupin <guillaume.saupin@cea.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_SKYLINEINPLACELU_H
11 #define EIGEN_SKYLINEINPLACELU_H
12 
13 namespace Eigen {
14 
15 /** \ingroup Skyline_Module
16  *
17  * \class SkylineInplaceLU
18  *
19  * \brief Inplace LU decomposition of a skyline matrix and associated features
20  *
21  * \param MatrixType the type of the matrix of which we are computing the LU factorization
22  *
23  */
24 template<typename MatrixType>
25 class SkylineInplaceLU {
26 protected:
27     typedef typename MatrixType::Scalar Scalar;
28     typedef typename MatrixType::Index Index;
29 
30     typedef typename NumTraits<typename MatrixType::Scalar>::Real RealScalar;
31 
32 public:
33 
34     /** Creates a LU object and compute the respective factorization of \a matrix using
35      * flags \a flags. */
36     SkylineInplaceLU(MatrixType& matrix, int flags = 0)
m_flags(flags)37     : /*m_matrix(matrix.rows(), matrix.cols()),*/ m_flags(flags), m_status(0), m_lu(matrix) {
38         m_precision = RealScalar(0.1) * Eigen::dummy_precision<RealScalar > ();
39         m_lu.IsRowMajor ? computeRowMajor() : compute();
40     }
41 
42     /** Sets the relative threshold value used to prune zero coefficients during the decomposition.
43      *
44      * Setting a value greater than zero speeds up computation, and yields to an imcomplete
45      * factorization with fewer non zero coefficients. Such approximate factors are especially
46      * useful to initialize an iterative solver.
47      *
48      * Note that the exact meaning of this parameter might depends on the actual
49      * backend. Moreover, not all backends support this feature.
50      *
51      * \sa precision() */
setPrecision(RealScalar v)52     void setPrecision(RealScalar v) {
53         m_precision = v;
54     }
55 
56     /** \returns the current precision.
57      *
58      * \sa setPrecision() */
precision()59     RealScalar precision() const {
60         return m_precision;
61     }
62 
63     /** Sets the flags. Possible values are:
64      *  - CompleteFactorization
65      *  - IncompleteFactorization
66      *  - MemoryEfficient
67      *  - one of the ordering methods
68      *  - etc...
69      *
70      * \sa flags() */
setFlags(int f)71     void setFlags(int f) {
72         m_flags = f;
73     }
74 
75     /** \returns the current flags */
flags()76     int flags() const {
77         return m_flags;
78     }
79 
setOrderingMethod(int m)80     void setOrderingMethod(int m) {
81         m_flags = m;
82     }
83 
orderingMethod()84     int orderingMethod() const {
85         return m_flags;
86     }
87 
88     /** Computes/re-computes the LU factorization */
89     void compute();
90     void computeRowMajor();
91 
92     /** \returns the lower triangular matrix L */
93     //inline const MatrixType& matrixL() const { return m_matrixL; }
94 
95     /** \returns the upper triangular matrix U */
96     //inline const MatrixType& matrixU() const { return m_matrixU; }
97 
98     template<typename BDerived, typename XDerived>
99     bool solve(const MatrixBase<BDerived> &b, MatrixBase<XDerived>* x,
100             const int transposed = 0) const;
101 
102     /** \returns true if the factorization succeeded */
succeeded(void)103     inline bool succeeded(void) const {
104         return m_succeeded;
105     }
106 
107 protected:
108     RealScalar m_precision;
109     int m_flags;
110     mutable int m_status;
111     bool m_succeeded;
112     MatrixType& m_lu;
113 };
114 
115 /** Computes / recomputes the in place LU decomposition of the SkylineInplaceLU.
116  * using the default algorithm.
117  */
118 template<typename MatrixType>
119 //template<typename _Scalar>
compute()120 void SkylineInplaceLU<MatrixType>::compute() {
121     const size_t rows = m_lu.rows();
122     const size_t cols = m_lu.cols();
123 
124     eigen_assert(rows == cols && "We do not (yet) support rectangular LU.");
125     eigen_assert(!m_lu.IsRowMajor && "LU decomposition does not work with rowMajor Storage");
126 
127     for (Index row = 0; row < rows; row++) {
128         const double pivot = m_lu.coeffDiag(row);
129 
130         //Lower matrix Columns update
131         const Index& col = row;
132         for (typename MatrixType::InnerLowerIterator lIt(m_lu, col); lIt; ++lIt) {
133             lIt.valueRef() /= pivot;
134         }
135 
136         //Upper matrix update -> contiguous memory access
137         typename MatrixType::InnerLowerIterator lIt(m_lu, col);
138         for (Index rrow = row + 1; rrow < m_lu.rows(); rrow++) {
139             typename MatrixType::InnerUpperIterator uItPivot(m_lu, row);
140             typename MatrixType::InnerUpperIterator uIt(m_lu, rrow);
141             const double coef = lIt.value();
142 
143             uItPivot += (rrow - row - 1);
144 
145             //update upper part  -> contiguous memory access
146             for (++uItPivot; uIt && uItPivot;) {
147                 uIt.valueRef() -= uItPivot.value() * coef;
148 
149                 ++uIt;
150                 ++uItPivot;
151             }
152             ++lIt;
153         }
154 
155         //Upper matrix update -> non contiguous memory access
156         typename MatrixType::InnerLowerIterator lIt3(m_lu, col);
157         for (Index rrow = row + 1; rrow < m_lu.rows(); rrow++) {
158             typename MatrixType::InnerUpperIterator uItPivot(m_lu, row);
159             const double coef = lIt3.value();
160 
161             //update lower part ->  non contiguous memory access
162             for (Index i = 0; i < rrow - row - 1; i++) {
163                 m_lu.coeffRefLower(rrow, row + i + 1) -= uItPivot.value() * coef;
164                 ++uItPivot;
165             }
166             ++lIt3;
167         }
168         //update diag -> contiguous
169         typename MatrixType::InnerLowerIterator lIt2(m_lu, col);
170         for (Index rrow = row + 1; rrow < m_lu.rows(); rrow++) {
171 
172             typename MatrixType::InnerUpperIterator uItPivot(m_lu, row);
173             typename MatrixType::InnerUpperIterator uIt(m_lu, rrow);
174             const double coef = lIt2.value();
175 
176             uItPivot += (rrow - row - 1);
177             m_lu.coeffRefDiag(rrow) -= uItPivot.value() * coef;
178             ++lIt2;
179         }
180     }
181 }
182 
183 template<typename MatrixType>
computeRowMajor()184 void SkylineInplaceLU<MatrixType>::computeRowMajor() {
185     const size_t rows = m_lu.rows();
186     const size_t cols = m_lu.cols();
187 
188     eigen_assert(rows == cols && "We do not (yet) support rectangular LU.");
189     eigen_assert(m_lu.IsRowMajor && "You're trying to apply rowMajor decomposition on a ColMajor matrix !");
190 
191     for (Index row = 0; row < rows; row++) {
192         typename MatrixType::InnerLowerIterator llIt(m_lu, row);
193 
194 
195         for (Index col = llIt.col(); col < row; col++) {
196             if (m_lu.coeffExistLower(row, col)) {
197                 const double diag = m_lu.coeffDiag(col);
198 
199                 typename MatrixType::InnerLowerIterator lIt(m_lu, row);
200                 typename MatrixType::InnerUpperIterator uIt(m_lu, col);
201 
202 
203                 const Index offset = lIt.col() - uIt.row();
204 
205 
206                 Index stop = offset > 0 ? col - lIt.col() : col - uIt.row();
207 
208                 //#define VECTORIZE
209 #ifdef VECTORIZE
210                 Map<VectorXd > rowVal(lIt.valuePtr() + (offset > 0 ? 0 : -offset), stop);
211                 Map<VectorXd > colVal(uIt.valuePtr() + (offset > 0 ? offset : 0), stop);
212 
213 
214                 Scalar newCoeff = m_lu.coeffLower(row, col) - rowVal.dot(colVal);
215 #else
216                 if (offset > 0) //Skip zero value of lIt
217                     uIt += offset;
218                 else //Skip zero values of uIt
219                     lIt += -offset;
220                 Scalar newCoeff = m_lu.coeffLower(row, col);
221 
222                 for (Index k = 0; k < stop; ++k) {
223                     const Scalar tmp = newCoeff;
224                     newCoeff = tmp - lIt.value() * uIt.value();
225                     ++lIt;
226                     ++uIt;
227                 }
228 #endif
229 
230                 m_lu.coeffRefLower(row, col) = newCoeff / diag;
231             }
232         }
233 
234         //Upper matrix update
235         const Index col = row;
236         typename MatrixType::InnerUpperIterator uuIt(m_lu, col);
237         for (Index rrow = uuIt.row(); rrow < col; rrow++) {
238 
239             typename MatrixType::InnerLowerIterator lIt(m_lu, rrow);
240             typename MatrixType::InnerUpperIterator uIt(m_lu, col);
241             const Index offset = lIt.col() - uIt.row();
242 
243             Index stop = offset > 0 ? rrow - lIt.col() : rrow - uIt.row();
244 
245 #ifdef VECTORIZE
246             Map<VectorXd > rowVal(lIt.valuePtr() + (offset > 0 ? 0 : -offset), stop);
247             Map<VectorXd > colVal(uIt.valuePtr() + (offset > 0 ? offset : 0), stop);
248 
249             Scalar newCoeff = m_lu.coeffUpper(rrow, col) - rowVal.dot(colVal);
250 #else
251             if (offset > 0) //Skip zero value of lIt
252                 uIt += offset;
253             else //Skip zero values of uIt
254                 lIt += -offset;
255             Scalar newCoeff = m_lu.coeffUpper(rrow, col);
256             for (Index k = 0; k < stop; ++k) {
257                 const Scalar tmp = newCoeff;
258                 newCoeff = tmp - lIt.value() * uIt.value();
259 
260                 ++lIt;
261                 ++uIt;
262             }
263 #endif
264             m_lu.coeffRefUpper(rrow, col) = newCoeff;
265         }
266 
267 
268         //Diag matrix update
269         typename MatrixType::InnerLowerIterator lIt(m_lu, row);
270         typename MatrixType::InnerUpperIterator uIt(m_lu, row);
271 
272         const Index offset = lIt.col() - uIt.row();
273 
274 
275         Index stop = offset > 0 ? lIt.size() : uIt.size();
276 #ifdef VECTORIZE
277         Map<VectorXd > rowVal(lIt.valuePtr() + (offset > 0 ? 0 : -offset), stop);
278         Map<VectorXd > colVal(uIt.valuePtr() + (offset > 0 ? offset : 0), stop);
279         Scalar newCoeff = m_lu.coeffDiag(row) - rowVal.dot(colVal);
280 #else
281         if (offset > 0) //Skip zero value of lIt
282             uIt += offset;
283         else //Skip zero values of uIt
284             lIt += -offset;
285         Scalar newCoeff = m_lu.coeffDiag(row);
286         for (Index k = 0; k < stop; ++k) {
287             const Scalar tmp = newCoeff;
288             newCoeff = tmp - lIt.value() * uIt.value();
289             ++lIt;
290             ++uIt;
291         }
292 #endif
293         m_lu.coeffRefDiag(row) = newCoeff;
294     }
295 }
296 
297 /** Computes *x = U^-1 L^-1 b
298  *
299  * If \a transpose is set to SvTranspose or SvAdjoint, the solution
300  * of the transposed/adjoint system is computed instead.
301  *
302  * Not all backends implement the solution of the transposed or
303  * adjoint system.
304  */
305 template<typename MatrixType>
306 template<typename BDerived, typename XDerived>
solve(const MatrixBase<BDerived> & b,MatrixBase<XDerived> * x,const int transposed)307 bool SkylineInplaceLU<MatrixType>::solve(const MatrixBase<BDerived> &b, MatrixBase<XDerived>* x, const int transposed) const {
308     const size_t rows = m_lu.rows();
309     const size_t cols = m_lu.cols();
310 
311 
312     for (Index row = 0; row < rows; row++) {
313         x->coeffRef(row) = b.coeff(row);
314         Scalar newVal = x->coeff(row);
315         typename MatrixType::InnerLowerIterator lIt(m_lu, row);
316 
317         Index col = lIt.col();
318         while (lIt.col() < row) {
319 
320             newVal -= x->coeff(col++) * lIt.value();
321             ++lIt;
322         }
323 
324         x->coeffRef(row) = newVal;
325     }
326 
327 
328     for (Index col = rows - 1; col > 0; col--) {
329         x->coeffRef(col) = x->coeff(col) / m_lu.coeffDiag(col);
330 
331         const Scalar x_col = x->coeff(col);
332 
333         typename MatrixType::InnerUpperIterator uIt(m_lu, col);
334         uIt += uIt.size()-1;
335 
336 
337         while (uIt) {
338             x->coeffRef(uIt.row()) -= x_col * uIt.value();
339             //TODO : introduce --operator
340             uIt += -1;
341         }
342 
343 
344     }
345     x->coeffRef(0) = x->coeff(0) / m_lu.coeffDiag(0);
346 
347     return true;
348 }
349 
350 } // end namespace Eigen
351 
352 #endif // EIGEN_SKYLINELU_H
353