• 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) 2012 Désiré Nuentsa-Wakam <desire.nuentsa_wakam@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 
11 #include <iostream>
12 #include <fstream>
13 #include "Eigen/SparseCore"
14 #include <bench/BenchTimer.h>
15 #include <cstdlib>
16 #include <string>
17 #include <Eigen/Cholesky>
18 #include <Eigen/Jacobi>
19 #include <Eigen/Householder>
20 #include <Eigen/IterativeLinearSolvers>
21 #include <unsupported/Eigen/IterativeSolvers>
22 #include <Eigen/LU>
23 #include <unsupported/Eigen/SparseExtra>
24 
25 #ifdef EIGEN_CHOLMOD_SUPPORT
26 #include <Eigen/CholmodSupport>
27 #endif
28 
29 #ifdef EIGEN_UMFPACK_SUPPORT
30 #include <Eigen/UmfPackSupport>
31 #endif
32 
33 #ifdef EIGEN_PARDISO_SUPPORT
34 #include <Eigen/PardisoSupport>
35 #endif
36 
37 #ifdef EIGEN_SUPERLU_SUPPORT
38 #include <Eigen/SuperLUSupport>
39 #endif
40 
41 #ifdef EIGEN_PASTIX_SUPPORT
42 #include <Eigen/PaStiXSupport>
43 #endif
44 
45 // CONSTANTS
46 #define EIGEN_UMFPACK  0
47 #define EIGEN_SUPERLU  1
48 #define EIGEN_PASTIX  2
49 #define EIGEN_PARDISO  3
50 #define EIGEN_BICGSTAB  4
51 #define EIGEN_BICGSTAB_ILUT  5
52 #define EIGEN_GMRES 6
53 #define EIGEN_GMRES_ILUT 7
54 #define EIGEN_SIMPLICIAL_LDLT  8
55 #define EIGEN_CHOLMOD_LDLT  9
56 #define EIGEN_PASTIX_LDLT  10
57 #define EIGEN_PARDISO_LDLT  11
58 #define EIGEN_SIMPLICIAL_LLT  12
59 #define EIGEN_CHOLMOD_SUPERNODAL_LLT  13
60 #define EIGEN_CHOLMOD_SIMPLICIAL_LLT  14
61 #define EIGEN_PASTIX_LLT  15
62 #define EIGEN_PARDISO_LLT  16
63 #define EIGEN_CG  17
64 #define EIGEN_CG_PRECOND  18
65 #define EIGEN_ALL_SOLVERS  19
66 
67 using namespace Eigen;
68 using namespace std;
69 
70 struct Stats{
71   ComputationInfo info;
72   double total_time;
73   double compute_time;
74   double solve_time;
75   double rel_error;
76   int memory_used;
77   int iterations;
78   int isavail;
79   int isIterative;
80 };
81 
82 // Global variables for input parameters
83 int MaximumIters; // Maximum number of iterations
84 double RelErr; // Relative error of the computed solution
85 
test_precision()86 template<typename T> inline typename NumTraits<T>::Real test_precision() { return NumTraits<T>::dummy_precision(); }
87 template<> inline float test_precision<float>() { return 1e-3f; }
88 template<> inline double test_precision<double>() { return 1e-6; }
89 template<> inline float test_precision<std::complex<float> >() { return test_precision<float>(); }
90 template<> inline double test_precision<std::complex<double> >() { return test_precision<double>(); }
91 
printStatheader(std::ofstream & out)92 void printStatheader(std::ofstream& out)
93 {
94   int LUcnt = 0;
95   string LUlist =" ", LLTlist = "<TH > LLT", LDLTlist = "<TH > LDLT ";
96 
97 #ifdef EIGEN_UMFPACK_SUPPORT
98   LUlist += "<TH > UMFPACK "; LUcnt++;
99 #endif
100 #ifdef EIGEN_SUPERLU_SUPPORT
101   LUlist += "<TH > SUPERLU "; LUcnt++;
102 #endif
103 #ifdef EIGEN_CHOLMOD_SUPPORT
104   LLTlist += "<TH > CHOLMOD SP LLT<TH > CHOLMOD LLT";
105   LDLTlist += "<TH>CHOLMOD LDLT";
106 #endif
107 #ifdef EIGEN_PARDISO_SUPPORT
108   LUlist += "<TH > PARDISO LU";  LUcnt++;
109   LLTlist += "<TH > PARDISO LLT";
110   LDLTlist += "<TH > PARDISO LDLT";
111 #endif
112 #ifdef EIGEN_PASTIX_SUPPORT
113   LUlist += "<TH > PASTIX LU";  LUcnt++;
114   LLTlist += "<TH > PASTIX LLT";
115   LDLTlist += "<TH > PASTIX LDLT";
116 #endif
117 
118   out << "<TABLE border=\"1\" >\n ";
119   out << "<TR><TH>Matrix <TH> N <TH> NNZ <TH> ";
120   if (LUcnt) out << LUlist;
121   out << " <TH >BiCGSTAB <TH >BiCGSTAB+ILUT"<< "<TH >GMRES+ILUT" <<LDLTlist << LLTlist <<  "<TH> CG "<< std::endl;
122 }
123 
124 
125 template<typename Solver, typename Scalar>
call_solver(Solver & solver,const typename Solver::MatrixType & A,const Matrix<Scalar,Dynamic,1> & b,const Matrix<Scalar,Dynamic,1> & refX)126 Stats call_solver(Solver &solver, const typename Solver::MatrixType& A, const Matrix<Scalar, Dynamic, 1>& b, const Matrix<Scalar, Dynamic, 1>& refX)
127 {
128   Stats stat;
129   Matrix<Scalar, Dynamic, 1> x;
130   BenchTimer timer;
131   timer.reset();
132   timer.start();
133   solver.compute(A);
134   if (solver.info() != Success)
135   {
136     stat.info = NumericalIssue;
137     std::cerr << "Solver failed ... \n";
138     return stat;
139   }
140   timer.stop();
141   stat.compute_time = timer.value();
142 
143   timer.reset();
144   timer.start();
145   x = solver.solve(b);
146   if (solver.info() == NumericalIssue)
147   {
148     stat.info = NumericalIssue;
149     std::cerr << "Solver failed ... \n";
150     return stat;
151   }
152 
153   timer.stop();
154   stat.solve_time = timer.value();
155   stat.total_time = stat.solve_time + stat.compute_time;
156   stat.memory_used = 0;
157   // Verify the relative error
158   if(refX.size() != 0)
159     stat.rel_error = (refX - x).norm()/refX.norm();
160   else
161   {
162     // Compute the relative residual norm
163     Matrix<Scalar, Dynamic, 1> temp;
164     temp = A * x;
165     stat.rel_error = (b-temp).norm()/b.norm();
166   }
167   if ( stat.rel_error > RelErr )
168   {
169     stat.info = NoConvergence;
170     return stat;
171   }
172   else
173   {
174     stat.info = Success;
175     return stat;
176   }
177 }
178 
179 template<typename Solver, typename Scalar>
call_directsolver(Solver & solver,const typename Solver::MatrixType & A,const Matrix<Scalar,Dynamic,1> & b,const Matrix<Scalar,Dynamic,1> & refX)180 Stats call_directsolver(Solver& solver, const typename Solver::MatrixType& A, const Matrix<Scalar, Dynamic, 1>& b, const Matrix<Scalar, Dynamic, 1>& refX)
181 {
182     Stats stat;
183     stat = call_solver(solver, A, b, refX);
184     return stat;
185 }
186 
187 template<typename Solver, typename Scalar>
call_itersolver(Solver & solver,const typename Solver::MatrixType & A,const Matrix<Scalar,Dynamic,1> & b,const Matrix<Scalar,Dynamic,1> & refX)188 Stats call_itersolver(Solver &solver, const typename Solver::MatrixType& A, const Matrix<Scalar, Dynamic, 1>& b, const Matrix<Scalar, Dynamic, 1>& refX)
189 {
190   Stats stat;
191   solver.setTolerance(RelErr);
192   solver.setMaxIterations(MaximumIters);
193 
194   stat = call_solver(solver, A, b, refX);
195   stat.iterations = solver.iterations();
196   return stat;
197 }
198 
printStatItem(Stats * stat,int solver_id,int & best_time_id,double & best_time_val)199 inline void printStatItem(Stats *stat, int solver_id, int& best_time_id, double& best_time_val)
200 {
201   stat[solver_id].isavail = 1;
202 
203   if (stat[solver_id].info == NumericalIssue)
204   {
205     cout << " SOLVER FAILED ... Probably a numerical issue \n";
206     return;
207   }
208   if (stat[solver_id].info == NoConvergence){
209     cout << "REL. ERROR " <<  stat[solver_id].rel_error;
210     if(stat[solver_id].isIterative == 1)
211       cout << " (" << stat[solver_id].iterations << ") \n";
212     return;
213   }
214 
215   // Record the best CPU time
216   if (!best_time_val)
217   {
218     best_time_val = stat[solver_id].total_time;
219     best_time_id = solver_id;
220   }
221   else if (stat[solver_id].total_time < best_time_val)
222   {
223     best_time_val = stat[solver_id].total_time;
224     best_time_id = solver_id;
225   }
226   // Print statistics to standard output
227   if (stat[solver_id].info == Success){
228     cout<< "COMPUTE TIME : " << stat[solver_id].compute_time<< " \n";
229     cout<< "SOLVE TIME : " << stat[solver_id].solve_time<< " \n";
230     cout<< "TOTAL TIME : " << stat[solver_id].total_time<< " \n";
231     cout << "REL. ERROR : " << stat[solver_id].rel_error ;
232     if(stat[solver_id].isIterative == 1) {
233       cout << " (" << stat[solver_id].iterations << ") ";
234     }
235     cout << std::endl;
236   }
237 
238 }
239 
240 
241 /* Print the results from all solvers corresponding to a particular matrix
242  * The best CPU time is printed in bold
243  */
printHtmlStatLine(Stats * stat,int best_time_id,string & statline)244 inline void printHtmlStatLine(Stats *stat, int best_time_id, string& statline)
245 {
246 
247   string markup;
248   ostringstream compute,solve,total,error;
249   for (int i = 0; i < EIGEN_ALL_SOLVERS; i++)
250   {
251     if (stat[i].isavail == 0) continue;
252     if(i == best_time_id)
253       markup = "<TD style=\"background-color:red\">";
254     else
255       markup = "<TD>";
256 
257     if (stat[i].info == Success){
258       compute << markup << stat[i].compute_time;
259       solve << markup << stat[i].solve_time;
260       total << markup << stat[i].total_time;
261       error << " <TD> " << stat[i].rel_error;
262       if(stat[i].isIterative == 1) {
263         error << " (" << stat[i].iterations << ") ";
264       }
265     }
266     else {
267       compute << " <TD> -" ;
268       solve << " <TD> -" ;
269       total << " <TD> -" ;
270       if(stat[i].info == NoConvergence){
271         error << " <TD> "<< stat[i].rel_error ;
272         if(stat[i].isIterative == 1)
273           error << " (" << stat[i].iterations << ") ";
274       }
275       else    error << " <TD> - ";
276     }
277   }
278 
279   statline = "<TH>Compute Time " + compute.str() + "\n"
280                         +  "<TR><TH>Solve Time " + solve.str() + "\n"
281                         +  "<TR><TH>Total Time " + total.str() + "\n"
282                         +"<TR><TH>Error(Iter)" + error.str() + "\n";
283 
284 }
285 
286 template <typename Scalar>
SelectSolvers(const SparseMatrix<Scalar> & A,unsigned int sym,Matrix<Scalar,Dynamic,1> & b,const Matrix<Scalar,Dynamic,1> & refX,Stats * stat)287 int SelectSolvers(const SparseMatrix<Scalar>&A, unsigned int sym, Matrix<Scalar, Dynamic, 1>& b, const Matrix<Scalar, Dynamic, 1>& refX, Stats *stat)
288 {
289   typedef SparseMatrix<Scalar, ColMajor> SpMat;
290   // First, deal with Nonsymmetric and symmetric matrices
291   int best_time_id = 0;
292   double best_time_val = 0.0;
293   //UMFPACK
294   #ifdef EIGEN_UMFPACK_SUPPORT
295   {
296     cout << "Solving with UMFPACK LU ... \n";
297     UmfPackLU<SpMat> solver;
298     stat[EIGEN_UMFPACK] = call_directsolver(solver, A, b, refX);
299     printStatItem(stat, EIGEN_UMFPACK, best_time_id, best_time_val);
300   }
301   #endif
302     //SuperLU
303   #ifdef EIGEN_SUPERLU_SUPPORT
304   {
305     cout << "\nSolving with SUPERLU ... \n";
306     SuperLU<SpMat> solver;
307     stat[EIGEN_SUPERLU] = call_directsolver(solver, A, b, refX);
308     printStatItem(stat, EIGEN_SUPERLU, best_time_id, best_time_val);
309   }
310   #endif
311 
312    // PaStix LU
313   #ifdef EIGEN_PASTIX_SUPPORT
314   {
315     cout << "\nSolving with PASTIX LU ... \n";
316     PastixLU<SpMat> solver;
317     stat[EIGEN_PASTIX] = call_directsolver(solver, A, b, refX) ;
318     printStatItem(stat, EIGEN_PASTIX, best_time_id, best_time_val);
319   }
320   #endif
321 
322    //PARDISO LU
323   #ifdef EIGEN_PARDISO_SUPPORT
324   {
325     cout << "\nSolving with PARDISO LU ... \n";
326     PardisoLU<SpMat>  solver;
327     stat[EIGEN_PARDISO] = call_directsolver(solver, A, b, refX);
328     printStatItem(stat, EIGEN_PARDISO, best_time_id, best_time_val);
329   }
330   #endif
331 
332 
333 
334   //BiCGSTAB
335   {
336     cout << "\nSolving with BiCGSTAB ... \n";
337     BiCGSTAB<SpMat> solver;
338     stat[EIGEN_BICGSTAB] = call_itersolver(solver, A, b, refX);
339     stat[EIGEN_BICGSTAB].isIterative = 1;
340     printStatItem(stat, EIGEN_BICGSTAB, best_time_id, best_time_val);
341   }
342   //BiCGSTAB+ILUT
343   {
344     cout << "\nSolving with BiCGSTAB and ILUT ... \n";
345     BiCGSTAB<SpMat, IncompleteLUT<Scalar> > solver;
346     stat[EIGEN_BICGSTAB_ILUT] = call_itersolver(solver, A, b, refX);
347     stat[EIGEN_BICGSTAB_ILUT].isIterative = 1;
348     printStatItem(stat, EIGEN_BICGSTAB_ILUT, best_time_id, best_time_val);
349   }
350 
351 
352   //GMRES
353 //   {
354 //     cout << "\nSolving with GMRES ... \n";
355 //     GMRES<SpMat> solver;
356 //     stat[EIGEN_GMRES] = call_itersolver(solver, A, b, refX);
357 //     stat[EIGEN_GMRES].isIterative = 1;
358 //     printStatItem(stat, EIGEN_GMRES, best_time_id, best_time_val);
359 //   }
360   //GMRES+ILUT
361   {
362     cout << "\nSolving with GMRES and ILUT ... \n";
363     GMRES<SpMat, IncompleteLUT<Scalar> > solver;
364     stat[EIGEN_GMRES_ILUT] = call_itersolver(solver, A, b, refX);
365     stat[EIGEN_GMRES_ILUT].isIterative = 1;
366     printStatItem(stat, EIGEN_GMRES_ILUT, best_time_id, best_time_val);
367   }
368 
369   // Hermitian and not necessarily positive-definites
370   if (sym != NonSymmetric)
371   {
372     // Internal Cholesky
373     {
374       cout << "\nSolving with Simplicial LDLT ... \n";
375       SimplicialLDLT<SpMat, Lower> solver;
376       stat[EIGEN_SIMPLICIAL_LDLT] = call_directsolver(solver, A, b, refX);
377       printStatItem(stat, EIGEN_SIMPLICIAL_LDLT, best_time_id, best_time_val);
378     }
379 
380     // CHOLMOD
381     #ifdef EIGEN_CHOLMOD_SUPPORT
382     {
383       cout << "\nSolving with CHOLMOD LDLT ... \n";
384       CholmodDecomposition<SpMat, Lower> solver;
385       solver.setMode(CholmodLDLt);
386       stat[EIGEN_CHOLMOD_LDLT] =  call_directsolver(solver, A, b, refX);
387       printStatItem(stat,EIGEN_CHOLMOD_LDLT, best_time_id, best_time_val);
388     }
389     #endif
390 
391     //PASTIX LLT
392     #ifdef EIGEN_PASTIX_SUPPORT
393     {
394       cout << "\nSolving with PASTIX LDLT ... \n";
395       PastixLDLT<SpMat, Lower> solver;
396       stat[EIGEN_PASTIX_LDLT] = call_directsolver(solver, A, b, refX);
397       printStatItem(stat,EIGEN_PASTIX_LDLT, best_time_id, best_time_val);
398     }
399     #endif
400 
401     //PARDISO LLT
402     #ifdef EIGEN_PARDISO_SUPPORT
403     {
404       cout << "\nSolving with PARDISO LDLT ... \n";
405       PardisoLDLT<SpMat, Lower> solver;
406       stat[EIGEN_PARDISO_LDLT] = call_directsolver(solver, A, b, refX);
407       printStatItem(stat,EIGEN_PARDISO_LDLT, best_time_id, best_time_val);
408     }
409     #endif
410   }
411 
412    // Now, symmetric POSITIVE DEFINITE matrices
413   if (sym == SPD)
414   {
415 
416     //Internal Sparse Cholesky
417     {
418       cout << "\nSolving with SIMPLICIAL LLT ... \n";
419       SimplicialLLT<SpMat, Lower> solver;
420       stat[EIGEN_SIMPLICIAL_LLT] = call_directsolver(solver, A, b, refX);
421       printStatItem(stat,EIGEN_SIMPLICIAL_LLT, best_time_id, best_time_val);
422     }
423 
424     // CHOLMOD
425     #ifdef EIGEN_CHOLMOD_SUPPORT
426     {
427       // CholMOD SuperNodal LLT
428       cout << "\nSolving with CHOLMOD LLT (Supernodal)... \n";
429       CholmodDecomposition<SpMat, Lower> solver;
430       solver.setMode(CholmodSupernodalLLt);
431       stat[EIGEN_CHOLMOD_SUPERNODAL_LLT] = call_directsolver(solver, A, b, refX);
432       printStatItem(stat,EIGEN_CHOLMOD_SUPERNODAL_LLT, best_time_id, best_time_val);
433       // CholMod Simplicial LLT
434       cout << "\nSolving with CHOLMOD LLT (Simplicial) ... \n";
435       solver.setMode(CholmodSimplicialLLt);
436       stat[EIGEN_CHOLMOD_SIMPLICIAL_LLT] = call_directsolver(solver, A, b, refX);
437       printStatItem(stat,EIGEN_CHOLMOD_SIMPLICIAL_LLT, best_time_id, best_time_val);
438     }
439     #endif
440 
441     //PASTIX LLT
442     #ifdef EIGEN_PASTIX_SUPPORT
443     {
444       cout << "\nSolving with PASTIX LLT ... \n";
445       PastixLLT<SpMat, Lower> solver;
446       stat[EIGEN_PASTIX_LLT] =  call_directsolver(solver, A, b, refX);
447       printStatItem(stat,EIGEN_PASTIX_LLT, best_time_id, best_time_val);
448     }
449     #endif
450 
451     //PARDISO LLT
452     #ifdef EIGEN_PARDISO_SUPPORT
453     {
454       cout << "\nSolving with PARDISO LLT ... \n";
455       PardisoLLT<SpMat, Lower> solver;
456       stat[EIGEN_PARDISO_LLT] = call_directsolver(solver, A, b, refX);
457       printStatItem(stat,EIGEN_PARDISO_LLT, best_time_id, best_time_val);
458     }
459     #endif
460 
461     // Internal CG
462     {
463       cout << "\nSolving with CG ... \n";
464       ConjugateGradient<SpMat, Lower> solver;
465       stat[EIGEN_CG] = call_itersolver(solver, A, b, refX);
466       stat[EIGEN_CG].isIterative = 1;
467       printStatItem(stat,EIGEN_CG, best_time_id, best_time_val);
468     }
469     //CG+IdentityPreconditioner
470 //     {
471 //       cout << "\nSolving with CG and IdentityPreconditioner ... \n";
472 //       ConjugateGradient<SpMat, Lower, IdentityPreconditioner> solver;
473 //       stat[EIGEN_CG_PRECOND] = call_itersolver(solver, A, b, refX);
474 //       stat[EIGEN_CG_PRECOND].isIterative = 1;
475 //       printStatItem(stat,EIGEN_CG_PRECOND, best_time_id, best_time_val);
476 //     }
477   } // End SPD matrices
478 
479   return best_time_id;
480 }
481 
482 /* Browse all the matrices available in the specified folder
483  * and solve the associated linear system.
484  * The results of each solve are printed in the standard output
485  * and optionally in the provided html file
486  */
487 template <typename Scalar>
Browse_Matrices(const string folder,bool statFileExists,std::string & statFile,int maxiters,double tol)488 void Browse_Matrices(const string folder, bool statFileExists, std::string& statFile, int maxiters, double tol)
489 {
490   MaximumIters = maxiters; // Maximum number of iterations, global variable
491   RelErr = tol;  //Relative residual error  as stopping criterion for iterative solvers
492   MatrixMarketIterator<Scalar> it(folder);
493   Stats stat[EIGEN_ALL_SOLVERS];
494   for ( ; it; ++it)
495   {
496     for (int i = 0; i < EIGEN_ALL_SOLVERS; i++)
497     {
498       stat[i].isavail = 0;
499       stat[i].isIterative = 0;
500     }
501 
502     int best_time_id;
503     cout<< "\n\n===================================================== \n";
504     cout<< " ======  SOLVING WITH MATRIX " << it.matname() << " ====\n";
505     cout<< " =================================================== \n\n";
506     Matrix<Scalar, Dynamic, 1> refX;
507     if(it.hasrefX()) refX = it.refX();
508     best_time_id = SelectSolvers<Scalar>(it.matrix(), it.sym(), it.rhs(), refX, &stat[0]);
509 
510     if(statFileExists)
511     {
512       string statline;
513       printHtmlStatLine(&stat[0], best_time_id, statline);
514       std::ofstream statbuf(statFile.c_str(), std::ios::app);
515       statbuf << "<TR><TH rowspan=\"4\">" << it.matname() << " <TD rowspan=\"4\"> "
516       << it.matrix().rows() << " <TD rowspan=\"4\"> " << it.matrix().nonZeros()<< " "<< statline ;
517       statbuf.close();
518     }
519   }
520 }
521 
522 bool get_options(int argc, char **args, string option, string* value=0)
523 {
524   int idx = 1, found=false;
525   while (idx<argc && !found){
526     if (option.compare(args[idx]) == 0){
527       found = true;
528       if(value) *value = args[idx+1];
529     }
530     idx+=2;
531   }
532   return found;
533 }
534