1 // This file is part of Eigen, a lightweight C++ template library
2 // for linear algebra.
3 //
4 // Copyright (C) 2012 Désiré Nuentsa-Wakam <desire.nuentsa_wakam@inria.fr>
5 // Copyright (C) 2014 Gael Guennebaud <gael.guennebaud@inria.fr>
6 //
7 // This Source Code Form is subject to the terms of the Mozilla
8 // Public License v. 2.0. If a copy of the MPL was not distributed
9 // with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
10
11 #ifndef EIGEN_INCOMPLETE_LUT_H
12 #define EIGEN_INCOMPLETE_LUT_H
13
14
15 namespace Eigen {
16
17 namespace internal {
18
19 /** \internal
20 * Compute a quick-sort split of a vector
21 * On output, the vector row is permuted such that its elements satisfy
22 * abs(row(i)) >= abs(row(ncut)) if i<ncut
23 * abs(row(i)) <= abs(row(ncut)) if i>ncut
24 * \param row The vector of values
25 * \param ind The array of index for the elements in @p row
26 * \param ncut The number of largest elements to keep
27 **/
28 template <typename VectorV, typename VectorI>
QuickSplit(VectorV & row,VectorI & ind,Index ncut)29 Index QuickSplit(VectorV &row, VectorI &ind, Index ncut)
30 {
31 typedef typename VectorV::RealScalar RealScalar;
32 using std::swap;
33 using std::abs;
34 Index mid;
35 Index n = row.size(); /* length of the vector */
36 Index first, last ;
37
38 ncut--; /* to fit the zero-based indices */
39 first = 0;
40 last = n-1;
41 if (ncut < first || ncut > last ) return 0;
42
43 do {
44 mid = first;
45 RealScalar abskey = abs(row(mid));
46 for (Index j = first + 1; j <= last; j++) {
47 if ( abs(row(j)) > abskey) {
48 ++mid;
49 swap(row(mid), row(j));
50 swap(ind(mid), ind(j));
51 }
52 }
53 /* Interchange for the pivot element */
54 swap(row(mid), row(first));
55 swap(ind(mid), ind(first));
56
57 if (mid > ncut) last = mid - 1;
58 else if (mid < ncut ) first = mid + 1;
59 } while (mid != ncut );
60
61 return 0; /* mid is equal to ncut */
62 }
63
64 }// end namespace internal
65
66 /** \ingroup IterativeLinearSolvers_Module
67 * \class IncompleteLUT
68 * \brief Incomplete LU factorization with dual-threshold strategy
69 *
70 * \implsparsesolverconcept
71 *
72 * During the numerical factorization, two dropping rules are used :
73 * 1) any element whose magnitude is less than some tolerance is dropped.
74 * This tolerance is obtained by multiplying the input tolerance @p droptol
75 * by the average magnitude of all the original elements in the current row.
76 * 2) After the elimination of the row, only the @p fill largest elements in
77 * the L part and the @p fill largest elements in the U part are kept
78 * (in addition to the diagonal element ). Note that @p fill is computed from
79 * the input parameter @p fillfactor which is used the ratio to control the fill_in
80 * relatively to the initial number of nonzero elements.
81 *
82 * The two extreme cases are when @p droptol=0 (to keep all the @p fill*2 largest elements)
83 * and when @p fill=n/2 with @p droptol being different to zero.
84 *
85 * References : Yousef Saad, ILUT: A dual threshold incomplete LU factorization,
86 * Numerical Linear Algebra with Applications, 1(4), pp 387-402, 1994.
87 *
88 * NOTE : The following implementation is derived from the ILUT implementation
89 * in the SPARSKIT package, Copyright (C) 2005, the Regents of the University of Minnesota
90 * released under the terms of the GNU LGPL:
91 * http://www-users.cs.umn.edu/~saad/software/SPARSKIT/README
92 * However, Yousef Saad gave us permission to relicense his ILUT code to MPL2.
93 * See the Eigen mailing list archive, thread: ILUT, date: July 8, 2012:
94 * http://listengine.tuxfamily.org/lists.tuxfamily.org/eigen/2012/07/msg00064.html
95 * alternatively, on GMANE:
96 * http://comments.gmane.org/gmane.comp.lib.eigen/3302
97 */
98 template <typename _Scalar, typename _StorageIndex = int>
99 class IncompleteLUT : public SparseSolverBase<IncompleteLUT<_Scalar, _StorageIndex> >
100 {
101 protected:
102 typedef SparseSolverBase<IncompleteLUT> Base;
103 using Base::m_isInitialized;
104 public:
105 typedef _Scalar Scalar;
106 typedef _StorageIndex StorageIndex;
107 typedef typename NumTraits<Scalar>::Real RealScalar;
108 typedef Matrix<Scalar,Dynamic,1> Vector;
109 typedef Matrix<StorageIndex,Dynamic,1> VectorI;
110 typedef SparseMatrix<Scalar,RowMajor,StorageIndex> FactorType;
111
112 enum {
113 ColsAtCompileTime = Dynamic,
114 MaxColsAtCompileTime = Dynamic
115 };
116
117 public:
118
IncompleteLUT()119 IncompleteLUT()
120 : m_droptol(NumTraits<Scalar>::dummy_precision()), m_fillfactor(10),
121 m_analysisIsOk(false), m_factorizationIsOk(false)
122 {}
123
124 template<typename MatrixType>
125 explicit IncompleteLUT(const MatrixType& mat, const RealScalar& droptol=NumTraits<Scalar>::dummy_precision(), int fillfactor = 10)
m_droptol(droptol)126 : m_droptol(droptol),m_fillfactor(fillfactor),
127 m_analysisIsOk(false),m_factorizationIsOk(false)
128 {
129 eigen_assert(fillfactor != 0);
130 compute(mat);
131 }
132
rows()133 Index rows() const { return m_lu.rows(); }
134
cols()135 Index cols() const { return m_lu.cols(); }
136
137 /** \brief Reports whether previous computation was successful.
138 *
139 * \returns \c Success if computation was succesful,
140 * \c NumericalIssue if the matrix.appears to be negative.
141 */
info()142 ComputationInfo info() const
143 {
144 eigen_assert(m_isInitialized && "IncompleteLUT is not initialized.");
145 return m_info;
146 }
147
148 template<typename MatrixType>
149 void analyzePattern(const MatrixType& amat);
150
151 template<typename MatrixType>
152 void factorize(const MatrixType& amat);
153
154 /**
155 * Compute an incomplete LU factorization with dual threshold on the matrix mat
156 * No pivoting is done in this version
157 *
158 **/
159 template<typename MatrixType>
compute(const MatrixType & amat)160 IncompleteLUT& compute(const MatrixType& amat)
161 {
162 analyzePattern(amat);
163 factorize(amat);
164 return *this;
165 }
166
167 void setDroptol(const RealScalar& droptol);
168 void setFillfactor(int fillfactor);
169
170 template<typename Rhs, typename Dest>
_solve_impl(const Rhs & b,Dest & x)171 void _solve_impl(const Rhs& b, Dest& x) const
172 {
173 x = m_Pinv * b;
174 x = m_lu.template triangularView<UnitLower>().solve(x);
175 x = m_lu.template triangularView<Upper>().solve(x);
176 x = m_P * x;
177 }
178
179 protected:
180
181 /** keeps off-diagonal entries; drops diagonal entries */
182 struct keep_diag {
operatorkeep_diag183 inline bool operator() (const Index& row, const Index& col, const Scalar&) const
184 {
185 return row!=col;
186 }
187 };
188
189 protected:
190
191 FactorType m_lu;
192 RealScalar m_droptol;
193 int m_fillfactor;
194 bool m_analysisIsOk;
195 bool m_factorizationIsOk;
196 ComputationInfo m_info;
197 PermutationMatrix<Dynamic,Dynamic,StorageIndex> m_P; // Fill-reducing permutation
198 PermutationMatrix<Dynamic,Dynamic,StorageIndex> m_Pinv; // Inverse permutation
199 };
200
201 /**
202 * Set control parameter droptol
203 * \param droptol Drop any element whose magnitude is less than this tolerance
204 **/
205 template<typename Scalar, typename StorageIndex>
setDroptol(const RealScalar & droptol)206 void IncompleteLUT<Scalar,StorageIndex>::setDroptol(const RealScalar& droptol)
207 {
208 this->m_droptol = droptol;
209 }
210
211 /**
212 * Set control parameter fillfactor
213 * \param fillfactor This is used to compute the number @p fill_in of largest elements to keep on each row.
214 **/
215 template<typename Scalar, typename StorageIndex>
setFillfactor(int fillfactor)216 void IncompleteLUT<Scalar,StorageIndex>::setFillfactor(int fillfactor)
217 {
218 this->m_fillfactor = fillfactor;
219 }
220
221 template <typename Scalar, typename StorageIndex>
222 template<typename _MatrixType>
analyzePattern(const _MatrixType & amat)223 void IncompleteLUT<Scalar,StorageIndex>::analyzePattern(const _MatrixType& amat)
224 {
225 // Compute the Fill-reducing permutation
226 // Since ILUT does not perform any numerical pivoting,
227 // it is highly preferable to keep the diagonal through symmetric permutations.
228 #ifndef EIGEN_MPL2_ONLY
229 // To this end, let's symmetrize the pattern and perform AMD on it.
230 SparseMatrix<Scalar,ColMajor, StorageIndex> mat1 = amat;
231 SparseMatrix<Scalar,ColMajor, StorageIndex> mat2 = amat.transpose();
232 // FIXME for a matrix with nearly symmetric pattern, mat2+mat1 is the appropriate choice.
233 // on the other hand for a really non-symmetric pattern, mat2*mat1 should be prefered...
234 SparseMatrix<Scalar,ColMajor, StorageIndex> AtA = mat2 + mat1;
235 AMDOrdering<StorageIndex> ordering;
236 ordering(AtA,m_P);
237 m_Pinv = m_P.inverse(); // cache the inverse permutation
238 #else
239 // If AMD is not available, (MPL2-only), then let's use the slower COLAMD routine.
240 SparseMatrix<Scalar,ColMajor, StorageIndex> mat1 = amat;
241 COLAMDOrdering<StorageIndex> ordering;
242 ordering(mat1,m_Pinv);
243 m_P = m_Pinv.inverse();
244 #endif
245
246 m_analysisIsOk = true;
247 m_factorizationIsOk = false;
248 m_isInitialized = true;
249 }
250
251 template <typename Scalar, typename StorageIndex>
252 template<typename _MatrixType>
factorize(const _MatrixType & amat)253 void IncompleteLUT<Scalar,StorageIndex>::factorize(const _MatrixType& amat)
254 {
255 using std::sqrt;
256 using std::swap;
257 using std::abs;
258 using internal::convert_index;
259
260 eigen_assert((amat.rows() == amat.cols()) && "The factorization should be done on a square matrix");
261 Index n = amat.cols(); // Size of the matrix
262 m_lu.resize(n,n);
263 // Declare Working vectors and variables
264 Vector u(n) ; // real values of the row -- maximum size is n --
265 VectorI ju(n); // column position of the values in u -- maximum size is n
266 VectorI jr(n); // Indicate the position of the nonzero elements in the vector u -- A zero location is indicated by -1
267
268 // Apply the fill-reducing permutation
269 eigen_assert(m_analysisIsOk && "You must first call analyzePattern()");
270 SparseMatrix<Scalar,RowMajor, StorageIndex> mat;
271 mat = amat.twistedBy(m_Pinv);
272
273 // Initialization
274 jr.fill(-1);
275 ju.fill(0);
276 u.fill(0);
277
278 // number of largest elements to keep in each row:
279 Index fill_in = (amat.nonZeros()*m_fillfactor)/n + 1;
280 if (fill_in > n) fill_in = n;
281
282 // number of largest nonzero elements to keep in the L and the U part of the current row:
283 Index nnzL = fill_in/2;
284 Index nnzU = nnzL;
285 m_lu.reserve(n * (nnzL + nnzU + 1));
286
287 // global loop over the rows of the sparse matrix
288 for (Index ii = 0; ii < n; ii++)
289 {
290 // 1 - copy the lower and the upper part of the row i of mat in the working vector u
291
292 Index sizeu = 1; // number of nonzero elements in the upper part of the current row
293 Index sizel = 0; // number of nonzero elements in the lower part of the current row
294 ju(ii) = convert_index<StorageIndex>(ii);
295 u(ii) = 0;
296 jr(ii) = convert_index<StorageIndex>(ii);
297 RealScalar rownorm = 0;
298
299 typename FactorType::InnerIterator j_it(mat, ii); // Iterate through the current row ii
300 for (; j_it; ++j_it)
301 {
302 Index k = j_it.index();
303 if (k < ii)
304 {
305 // copy the lower part
306 ju(sizel) = convert_index<StorageIndex>(k);
307 u(sizel) = j_it.value();
308 jr(k) = convert_index<StorageIndex>(sizel);
309 ++sizel;
310 }
311 else if (k == ii)
312 {
313 u(ii) = j_it.value();
314 }
315 else
316 {
317 // copy the upper part
318 Index jpos = ii + sizeu;
319 ju(jpos) = convert_index<StorageIndex>(k);
320 u(jpos) = j_it.value();
321 jr(k) = convert_index<StorageIndex>(jpos);
322 ++sizeu;
323 }
324 rownorm += numext::abs2(j_it.value());
325 }
326
327 // 2 - detect possible zero row
328 if(rownorm==0)
329 {
330 m_info = NumericalIssue;
331 return;
332 }
333 // Take the 2-norm of the current row as a relative tolerance
334 rownorm = sqrt(rownorm);
335
336 // 3 - eliminate the previous nonzero rows
337 Index jj = 0;
338 Index len = 0;
339 while (jj < sizel)
340 {
341 // In order to eliminate in the correct order,
342 // we must select first the smallest column index among ju(jj:sizel)
343 Index k;
344 Index minrow = ju.segment(jj,sizel-jj).minCoeff(&k); // k is relative to the segment
345 k += jj;
346 if (minrow != ju(jj))
347 {
348 // swap the two locations
349 Index j = ju(jj);
350 swap(ju(jj), ju(k));
351 jr(minrow) = convert_index<StorageIndex>(jj);
352 jr(j) = convert_index<StorageIndex>(k);
353 swap(u(jj), u(k));
354 }
355 // Reset this location
356 jr(minrow) = -1;
357
358 // Start elimination
359 typename FactorType::InnerIterator ki_it(m_lu, minrow);
360 while (ki_it && ki_it.index() < minrow) ++ki_it;
361 eigen_internal_assert(ki_it && ki_it.col()==minrow);
362 Scalar fact = u(jj) / ki_it.value();
363
364 // drop too small elements
365 if(abs(fact) <= m_droptol)
366 {
367 jj++;
368 continue;
369 }
370
371 // linear combination of the current row ii and the row minrow
372 ++ki_it;
373 for (; ki_it; ++ki_it)
374 {
375 Scalar prod = fact * ki_it.value();
376 Index j = ki_it.index();
377 Index jpos = jr(j);
378 if (jpos == -1) // fill-in element
379 {
380 Index newpos;
381 if (j >= ii) // dealing with the upper part
382 {
383 newpos = ii + sizeu;
384 sizeu++;
385 eigen_internal_assert(sizeu<=n);
386 }
387 else // dealing with the lower part
388 {
389 newpos = sizel;
390 sizel++;
391 eigen_internal_assert(sizel<=ii);
392 }
393 ju(newpos) = convert_index<StorageIndex>(j);
394 u(newpos) = -prod;
395 jr(j) = convert_index<StorageIndex>(newpos);
396 }
397 else
398 u(jpos) -= prod;
399 }
400 // store the pivot element
401 u(len) = fact;
402 ju(len) = convert_index<StorageIndex>(minrow);
403 ++len;
404
405 jj++;
406 } // end of the elimination on the row ii
407
408 // reset the upper part of the pointer jr to zero
409 for(Index k = 0; k <sizeu; k++) jr(ju(ii+k)) = -1;
410
411 // 4 - partially sort and insert the elements in the m_lu matrix
412
413 // sort the L-part of the row
414 sizel = len;
415 len = (std::min)(sizel, nnzL);
416 typename Vector::SegmentReturnType ul(u.segment(0, sizel));
417 typename VectorI::SegmentReturnType jul(ju.segment(0, sizel));
418 internal::QuickSplit(ul, jul, len);
419
420 // store the largest m_fill elements of the L part
421 m_lu.startVec(ii);
422 for(Index k = 0; k < len; k++)
423 m_lu.insertBackByOuterInnerUnordered(ii,ju(k)) = u(k);
424
425 // store the diagonal element
426 // apply a shifting rule to avoid zero pivots (we are doing an incomplete factorization)
427 if (u(ii) == Scalar(0))
428 u(ii) = sqrt(m_droptol) * rownorm;
429 m_lu.insertBackByOuterInnerUnordered(ii, ii) = u(ii);
430
431 // sort the U-part of the row
432 // apply the dropping rule first
433 len = 0;
434 for(Index k = 1; k < sizeu; k++)
435 {
436 if(abs(u(ii+k)) > m_droptol * rownorm )
437 {
438 ++len;
439 u(ii + len) = u(ii + k);
440 ju(ii + len) = ju(ii + k);
441 }
442 }
443 sizeu = len + 1; // +1 to take into account the diagonal element
444 len = (std::min)(sizeu, nnzU);
445 typename Vector::SegmentReturnType uu(u.segment(ii+1, sizeu-1));
446 typename VectorI::SegmentReturnType juu(ju.segment(ii+1, sizeu-1));
447 internal::QuickSplit(uu, juu, len);
448
449 // store the largest elements of the U part
450 for(Index k = ii + 1; k < ii + len; k++)
451 m_lu.insertBackByOuterInnerUnordered(ii,ju(k)) = u(k);
452 }
453 m_lu.finalize();
454 m_lu.makeCompressed();
455
456 m_factorizationIsOk = true;
457 m_info = Success;
458 }
459
460 } // end namespace Eigen
461
462 #endif // EIGEN_INCOMPLETE_LUT_H
463