xref: /aosp_15_r20/external/eigen/Eigen/src/IterativeLinearSolvers/IncompleteLUT.h (revision bf2c37156dfe67e5dfebd6d394bad8b2ab5804d4)
1 // This file is part of Eigen, a lightweight C++ template library
2 // for linear algebra.
3 //
4 // Copyright (C) 2012 Désiré Nuentsa-Wakam <[email protected]>
5 // Copyright (C) 2014 Gael Guennebaud <[email protected]>
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     EIGEN_CONSTEXPR Index rows() const EIGEN_NOEXCEPT { return m_lu.rows(); }
134 
cols()135     EIGEN_CONSTEXPR Index cols() const EIGEN_NOEXCEPT { return m_lu.cols(); }
136 
137     /** \brief Reports whether previous computation was successful.
138       *
139       * \returns \c Success if computation was successful,
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   // To this end, let's symmetrize the pattern and perform AMD on it.
229   SparseMatrix<Scalar,ColMajor, StorageIndex> mat1 = amat;
230   SparseMatrix<Scalar,ColMajor, StorageIndex> mat2 = amat.transpose();
231   // FIXME for a matrix with nearly symmetric pattern, mat2+mat1 is the appropriate choice.
232   //       on the other hand for a really non-symmetric pattern, mat2*mat1 should be preferred...
233   SparseMatrix<Scalar,ColMajor, StorageIndex> AtA = mat2 + mat1;
234   AMDOrdering<StorageIndex> ordering;
235   ordering(AtA,m_P);
236   m_Pinv  = m_P.inverse(); // cache the inverse permutation
237   m_analysisIsOk = true;
238   m_factorizationIsOk = false;
239   m_isInitialized = true;
240 }
241 
242 template <typename Scalar, typename StorageIndex>
243 template<typename _MatrixType>
factorize(const _MatrixType & amat)244 void IncompleteLUT<Scalar,StorageIndex>::factorize(const _MatrixType& amat)
245 {
246   using std::sqrt;
247   using std::swap;
248   using std::abs;
249   using internal::convert_index;
250 
251   eigen_assert((amat.rows() == amat.cols()) && "The factorization should be done on a square matrix");
252   Index n = amat.cols();  // Size of the matrix
253   m_lu.resize(n,n);
254   // Declare Working vectors and variables
255   Vector u(n) ;     // real values of the row -- maximum size is n --
256   VectorI ju(n);   // column position of the values in u -- maximum size  is n
257   VectorI jr(n);   // Indicate the position of the nonzero elements in the vector u -- A zero location is indicated by -1
258 
259   // Apply the fill-reducing permutation
260   eigen_assert(m_analysisIsOk && "You must first call analyzePattern()");
261   SparseMatrix<Scalar,RowMajor, StorageIndex> mat;
262   mat = amat.twistedBy(m_Pinv);
263 
264   // Initialization
265   jr.fill(-1);
266   ju.fill(0);
267   u.fill(0);
268 
269   // number of largest elements to keep in each row:
270   Index fill_in = (amat.nonZeros()*m_fillfactor)/n + 1;
271   if (fill_in > n) fill_in = n;
272 
273   // number of largest nonzero elements to keep in the L and the U part of the current row:
274   Index nnzL = fill_in/2;
275   Index nnzU = nnzL;
276   m_lu.reserve(n * (nnzL + nnzU + 1));
277 
278   // global loop over the rows of the sparse matrix
279   for (Index ii = 0; ii < n; ii++)
280   {
281     // 1 - copy the lower and the upper part of the row i of mat in the working vector u
282 
283     Index sizeu = 1; // number of nonzero elements in the upper part of the current row
284     Index sizel = 0; // number of nonzero elements in the lower part of the current row
285     ju(ii)    = convert_index<StorageIndex>(ii);
286     u(ii)     = 0;
287     jr(ii)    = convert_index<StorageIndex>(ii);
288     RealScalar rownorm = 0;
289 
290     typename FactorType::InnerIterator j_it(mat, ii); // Iterate through the current row ii
291     for (; j_it; ++j_it)
292     {
293       Index k = j_it.index();
294       if (k < ii)
295       {
296         // copy the lower part
297         ju(sizel) = convert_index<StorageIndex>(k);
298         u(sizel) = j_it.value();
299         jr(k) = convert_index<StorageIndex>(sizel);
300         ++sizel;
301       }
302       else if (k == ii)
303       {
304         u(ii) = j_it.value();
305       }
306       else
307       {
308         // copy the upper part
309         Index jpos = ii + sizeu;
310         ju(jpos) = convert_index<StorageIndex>(k);
311         u(jpos) = j_it.value();
312         jr(k) = convert_index<StorageIndex>(jpos);
313         ++sizeu;
314       }
315       rownorm += numext::abs2(j_it.value());
316     }
317 
318     // 2 - detect possible zero row
319     if(rownorm==0)
320     {
321       m_info = NumericalIssue;
322       return;
323     }
324     // Take the 2-norm of the current row as a relative tolerance
325     rownorm = sqrt(rownorm);
326 
327     // 3 - eliminate the previous nonzero rows
328     Index jj = 0;
329     Index len = 0;
330     while (jj < sizel)
331     {
332       // In order to eliminate in the correct order,
333       // we must select first the smallest column index among  ju(jj:sizel)
334       Index k;
335       Index minrow = ju.segment(jj,sizel-jj).minCoeff(&k); // k is relative to the segment
336       k += jj;
337       if (minrow != ju(jj))
338       {
339         // swap the two locations
340         Index j = ju(jj);
341         swap(ju(jj), ju(k));
342         jr(minrow) = convert_index<StorageIndex>(jj);
343         jr(j) = convert_index<StorageIndex>(k);
344         swap(u(jj), u(k));
345       }
346       // Reset this location
347       jr(minrow) = -1;
348 
349       // Start elimination
350       typename FactorType::InnerIterator ki_it(m_lu, minrow);
351       while (ki_it && ki_it.index() < minrow) ++ki_it;
352       eigen_internal_assert(ki_it && ki_it.col()==minrow);
353       Scalar fact = u(jj) / ki_it.value();
354 
355       // drop too small elements
356       if(abs(fact) <= m_droptol)
357       {
358         jj++;
359         continue;
360       }
361 
362       // linear combination of the current row ii and the row minrow
363       ++ki_it;
364       for (; ki_it; ++ki_it)
365       {
366         Scalar prod = fact * ki_it.value();
367         Index j     = ki_it.index();
368         Index jpos  = jr(j);
369         if (jpos == -1) // fill-in element
370         {
371           Index newpos;
372           if (j >= ii) // dealing with the upper part
373           {
374             newpos = ii + sizeu;
375             sizeu++;
376             eigen_internal_assert(sizeu<=n);
377           }
378           else // dealing with the lower part
379           {
380             newpos = sizel;
381             sizel++;
382             eigen_internal_assert(sizel<=ii);
383           }
384           ju(newpos) = convert_index<StorageIndex>(j);
385           u(newpos) = -prod;
386           jr(j) = convert_index<StorageIndex>(newpos);
387         }
388         else
389           u(jpos) -= prod;
390       }
391       // store the pivot element
392       u(len)  = fact;
393       ju(len) = convert_index<StorageIndex>(minrow);
394       ++len;
395 
396       jj++;
397     } // end of the elimination on the row ii
398 
399     // reset the upper part of the pointer jr to zero
400     for(Index k = 0; k <sizeu; k++) jr(ju(ii+k)) = -1;
401 
402     // 4 - partially sort and insert the elements in the m_lu matrix
403 
404     // sort the L-part of the row
405     sizel = len;
406     len = (std::min)(sizel, nnzL);
407     typename Vector::SegmentReturnType ul(u.segment(0, sizel));
408     typename VectorI::SegmentReturnType jul(ju.segment(0, sizel));
409     internal::QuickSplit(ul, jul, len);
410 
411     // store the largest m_fill elements of the L part
412     m_lu.startVec(ii);
413     for(Index k = 0; k < len; k++)
414       m_lu.insertBackByOuterInnerUnordered(ii,ju(k)) = u(k);
415 
416     // store the diagonal element
417     // apply a shifting rule to avoid zero pivots (we are doing an incomplete factorization)
418     if (u(ii) == Scalar(0))
419       u(ii) = sqrt(m_droptol) * rownorm;
420     m_lu.insertBackByOuterInnerUnordered(ii, ii) = u(ii);
421 
422     // sort the U-part of the row
423     // apply the dropping rule first
424     len = 0;
425     for(Index k = 1; k < sizeu; k++)
426     {
427       if(abs(u(ii+k)) > m_droptol * rownorm )
428       {
429         ++len;
430         u(ii + len)  = u(ii + k);
431         ju(ii + len) = ju(ii + k);
432       }
433     }
434     sizeu = len + 1; // +1 to take into account the diagonal element
435     len = (std::min)(sizeu, nnzU);
436     typename Vector::SegmentReturnType uu(u.segment(ii+1, sizeu-1));
437     typename VectorI::SegmentReturnType juu(ju.segment(ii+1, sizeu-1));
438     internal::QuickSplit(uu, juu, len);
439 
440     // store the largest elements of the U part
441     for(Index k = ii + 1; k < ii + len; k++)
442       m_lu.insertBackByOuterInnerUnordered(ii,ju(k)) = u(k);
443   }
444   m_lu.finalize();
445   m_lu.makeCompressed();
446 
447   m_factorizationIsOk = true;
448   m_info = Success;
449 }
450 
451 } // end namespace Eigen
452 
453 #endif // EIGEN_INCOMPLETE_LUT_H
454