PetscSolverFeti.cc 43.6 KB
Newer Older
1
2
3
4
5
6
7
8
9
10
11
12
13
//
// Software License for AMDiS
//
// Copyright (c) 2010 Dresden University of Technology 
// All rights reserved.
// Authors: Simon Vey, Thomas Witkowski et al.
//
// This file is part of AMDiS
//
// See also license.opensource.txt in the distribution.


#include "parallel/PetscSolverFeti.h"
14
#include "parallel/PetscSolverFetiStructs.h"
15
16
#include "parallel/StdMpi.h"
#include "parallel/MpiHelper.h"
17
#include "io/VtkWriter.h"
18
19
20
21
22

namespace AMDiS {

  using namespace std;

23
24
25
26
27
28
29
  // y = mat * x
  int petscMultMatSchurPrimal(Mat mat, Vec x, Vec y)
  {
    // S_PiPi = K_PiPi - K_PiB inv(K_BB) K_BPi

    void *ctx;
    MatShellGetContext(mat, &ctx);
30
    SchurPrimalData* data = static_cast<SchurPrimalData*>(ctx);
31
32

    MatMult(*(data->mat_b_primal), x, data->tmp_vec_b);
Thomas Witkowski's avatar
Thomas Witkowski committed
33
    data->fetiSolver->solveLocalProblem(data->tmp_vec_b, data->tmp_vec_b);
34
35
36
37
38
39
40
41
42
43
44
    MatMult(*(data->mat_primal_b), data->tmp_vec_b, data->tmp_vec_primal);
    MatMult(*(data->mat_primal_primal), x, y);
    VecAXPBY(y, -1.0, 1.0, data->tmp_vec_primal);

    return 0;
  }


  // y = mat * x
  int petscMultMatFeti(Mat mat, Vec x, Vec y)
  {
45
46
    //    F = L inv(K_BB) trans(L) + L inv(K_BB) K_BPi inv(S_PiPi) K_PiB inv(K_BB) trans(L)
    // => F = L [I + inv(K_BB) K_BPi inv(S_PiPi) K_PiB] inv(K_BB) trans(L)
47
48
49

    void *ctx;
    MatShellGetContext(mat, &ctx);
50
    FetiData* data = static_cast<FetiData*>(ctx);
51

Thomas Witkowski's avatar
Thomas Witkowski committed
52
53
54
    MatMultTranspose(*(data->mat_lagrange), x, data->tmp_vec_b);
    data->fetiSolver->solveLocalProblem(data->tmp_vec_b, data->tmp_vec_b);
    MatMult(*(data->mat_lagrange), data->tmp_vec_b, data->tmp_vec_lagrange);
Thomas Witkowski's avatar
Thomas Witkowski committed
55

Thomas Witkowski's avatar
Thomas Witkowski committed
56
    MatMult(*(data->mat_primal_b), data->tmp_vec_b, data->tmp_vec_primal);
57
    KSPSolve(*(data->ksp_schur_primal), data->tmp_vec_primal, data->tmp_vec_primal);
Thomas Witkowski's avatar
Thomas Witkowski committed
58
59
60
    MatMult(*(data->mat_b_primal), data->tmp_vec_primal, data->tmp_vec_b);
    data->fetiSolver->solveLocalProblem(data->tmp_vec_b, data->tmp_vec_b);
    MatMult(*(data->mat_lagrange), data->tmp_vec_b, y);
61

Thomas Witkowski's avatar
Thomas Witkowski committed
62
    VecAXPBY(y, 1.0, 1.0, data->tmp_vec_lagrange);
63
64
65
66
67

    return 0;
  }


68
  // y = PC * x
69
  PetscErrorCode petscApplyFetiDirichletPrecon(PC pc, Vec x, Vec y)
70
  {
71
    // Get data for the preconditioner
72
73
    void *ctx;
    PCShellGetContext(pc, &ctx);
74
    FetiDirichletPreconData* data = static_cast<FetiDirichletPreconData*>(ctx);
75

76
    // Multiply with scaled Lagrange constraint matrix.
77
78
79
    MatMultTranspose(*(data->mat_lagrange_scaled), x, data->tmp_vec_b);


80
    // === Restriction of the B nodes to the boundary nodes. ===
81

82
83
84
85
    int nLocalB;
    int nLocalDuals;
    VecGetLocalSize(data->tmp_vec_b, &nLocalB);
    VecGetLocalSize(data->tmp_vec_duals0, &nLocalDuals);
86

87
88
89
90
    PetscScalar *local_b, *local_duals;
    VecGetArray(data->tmp_vec_b, &local_b);
    VecGetArray(data->tmp_vec_duals0, &local_duals);

91
92
93
    for (map<int, int>::iterator it = data->localToDualMap.begin();
	 it != data->localToDualMap.end(); ++it)
      local_duals[it->second] = local_b[it->first];
94
95

    VecRestoreArray(data->tmp_vec_b, &local_b);
96
    VecRestoreArray(data->tmp_vec_duals0, &local_duals);
97
98


99
    // === K_DD - K_DI inv(K_II) K_ID ===
100

101
    MatMult(*(data->mat_duals_duals), data->tmp_vec_duals0, data->tmp_vec_duals1);
102

103
    MatMult(*(data->mat_interior_duals), data->tmp_vec_duals0, data->tmp_vec_interior);
104
    KSPSolve(*(data->ksp_interior), data->tmp_vec_interior, data->tmp_vec_interior);
105
106
107
108
109
110
111
112
113
114
    MatMult(*(data->mat_duals_interior), data->tmp_vec_interior, data->tmp_vec_duals0);

    VecAXPBY(data->tmp_vec_duals0, 1.0, -1.0, data->tmp_vec_duals1);


    // === Prolongation from local dual nodes to B nodes.

    VecGetArray(data->tmp_vec_b, &local_b);
    VecGetArray(data->tmp_vec_duals0, &local_duals);

115
116
117
    for (map<int, int>::iterator it = data->localToDualMap.begin();
	 it != data->localToDualMap.end(); ++it)
      local_b[it->first] = local_duals[it->second];
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139

    VecRestoreArray(data->tmp_vec_b, &local_b);
    VecRestoreArray(data->tmp_vec_duals0, &local_duals);


    // Multiply with scaled Lagrange constraint matrix.
    MatMult(*(data->mat_lagrange_scaled), data->tmp_vec_b, y);

    return 0;
  }


  // y = PC * x
  PetscErrorCode petscApplyFetiLumpedPrecon(PC pc, Vec x, Vec y)
  {
    // Get data for the preconditioner
    void *ctx;
    PCShellGetContext(pc, &ctx);
    FetiLumpedPreconData* data = static_cast<FetiLumpedPreconData*>(ctx);

    // Multiply with scaled Lagrange constraint matrix.
    MatMultTranspose(*(data->mat_lagrange_scaled), x, data->tmp_vec_b);
140
141


142
    // === Restriction of the B nodes to the boundary nodes. ===
143

144
145
146
147
    int nLocalB;
    int nLocalDuals;
    VecGetLocalSize(data->tmp_vec_b, &nLocalB);
    VecGetLocalSize(data->tmp_vec_duals0, &nLocalDuals);
148

149
    PetscScalar *local_b, *local_duals;
150
    VecGetArray(data->tmp_vec_b, &local_b);
151
    VecGetArray(data->tmp_vec_duals0, &local_duals);
152

153
154
    for (int i = nLocalB - nLocalDuals, j = 0; i < nLocalB; i++, j++)
      local_duals[j] = local_b[i];
155
156

    VecRestoreArray(data->tmp_vec_b, &local_b);
157
158
159
160
161
162
163
    VecRestoreArray(data->tmp_vec_duals0, &local_duals);


    // === K_DD ===

    MatMult(*(data->mat_duals_duals), data->tmp_vec_duals0, data->tmp_vec_duals1);

164

165
    // === Prolongation from local dual nodes to B nodes.
166

167
168
169
170
171
    VecGetArray(data->tmp_vec_b, &local_b);
    VecGetArray(data->tmp_vec_duals1, &local_duals);

    for (int i = nLocalB - nLocalDuals, j = 0; i < nLocalB; i++, j++)
      local_b[i] = local_duals[j];
172

173
174
    VecRestoreArray(data->tmp_vec_b, &local_b);
    VecRestoreArray(data->tmp_vec_duals0, &local_duals);
175

176
177

    // Multiply with scaled Lagrange constraint matrix.
178
179
180
181
182
183
    MatMult(*(data->mat_lagrange_scaled), data->tmp_vec_b, y);

    return 0;
  }


184
185
  PetscSolverFeti::PetscSolverFeti()
    : PetscSolver(),
Thomas Witkowski's avatar
Thomas Witkowski committed
186
      schurPrimalSolver(0)
187
188
189
190
191
192
  {
    FUNCNAME("PetscSolverFeti::PetscSolverFeti()");

    string preconditionerName = "";
    Parameters::get("parallel->solver->precon", preconditionerName);
    if (preconditionerName == "" || preconditionerName == "none") {
193
      MSG("Create FETI-DP solver with no preconditioner!\n");
194
195
      fetiPreconditioner = FETI_NONE;
    } else if (preconditionerName == "dirichlet") {
196
      MSG("Create FETI-DP solver with Dirichlet preconditioner!\n");
197
198
      fetiPreconditioner = FETI_DIRICHLET;
    } else if (preconditionerName == "lumped") {
199
      MSG("Create FETI-DP solver with lumped preconditioner!\n");
200
201
      fetiPreconditioner = FETI_LUMPED;
    } else {
Thomas Witkowski's avatar
Thomas Witkowski committed
202
203
      ERROR_EXIT("Preconditioner \"%s\" not available!\n", 
		 preconditionerName.c_str());
204
    }
Thomas Witkowski's avatar
Thomas Witkowski committed
205
206
207
208
209

    Parameters::get("parallel->feti->schur primal solver", schurPrimalSolver);
    TEST_EXIT(schurPrimalSolver == 0 || schurPrimalSolver == 1)
      ("Wrong solver \"%d\"for the Schur primal complement!\n", 
       schurPrimalSolver);
210
211
212
  }


213
  void PetscSolverFeti::updateDofData()
214
215
  {
    FUNCNAME("PetscSolverFeti::updateDofData()");
216
217
218

    TEST_EXIT(meshDistributor->getFeSpace()->getBasisFcts()->getDegree() == 1)
      ("Works for linear basis functions only!\n");
219
   
220
221
222
223
    for (unsigned int i = 0; i < meshDistributor->getFeSpaces().size(); i++) {
      const FiniteElemSpace *feSpace = meshDistributor->getFeSpace(i);
      createPrimals(feSpace);      
      createDuals(feSpace);
224
      createLagrange(feSpace);      
225
226
      createIndexB(feSpace);
    }
Thomas Witkowski's avatar
Thomas Witkowski committed
227

228
229
230
231
232
    primalDofMap.setDofComm(meshDistributor->getSendDofs(),
			    meshDistributor->getRecvDofs());
    lagrangeMap.setDofComm(meshDistributor->getSendDofs(), 
			   meshDistributor->getRecvDofs());

Thomas Witkowski's avatar
Thomas Witkowski committed
233
234
235
236
    primalDofMap.update();
    dualDofMap.update();
    lagrangeMap.update();
    localDofMap.update();
237
238
    if (fetiPreconditioner == FETI_DIRICHLET)
      interiorDofMap.update();
Thomas Witkowski's avatar
Thomas Witkowski committed
239
240
241
242
243
244
245
246

    for (unsigned int i = 0; i < meshDistributor->getFeSpaces().size(); i++) {
      const FiniteElemSpace *feSpace = meshDistributor->getFeSpace(i);

      MSG("nRankPrimals = %d   nOverallPrimals = %d\n", 
	  primalDofMap[feSpace].nRankDofs, primalDofMap[feSpace].nOverallDofs);
      
      MSG("nRankDuals = %d   nOverallDuals = %d\n",
247
	  dualDofMap[feSpace].nRankDofs, dualDofMap[feSpace].nOverallDofs);
Thomas Witkowski's avatar
Thomas Witkowski committed
248
249

      MSG("nRankLagrange = %d  nOverallLagrange = %d\n",
250
	  lagrangeMap[feSpace].nRankDofs, lagrangeMap[feSpace].nOverallDofs);
251
252
253
254

      TEST_EXIT_DBG(localDofMap[feSpace].size() + primalDofMap[feSpace].size() == 
		    static_cast<unsigned int>(feSpace->getAdmin()->getUsedDofs()))
	("Should not happen!\n");
Thomas Witkowski's avatar
Thomas Witkowski committed
255
    }
256
257
258
  }


259
  void PetscSolverFeti::createPrimals(const FiniteElemSpace *feSpace)
Thomas Witkowski's avatar
Thomas Witkowski committed
260
  {
261
    FUNCNAME("PetscSolverFeti::createPrimals()");  
262

263
264
265
    // === Define all vertices on the interior boundaries of the macro mesh ===
    // === to be primal variables.                                          ===

266
267
268
    /// Set of DOF indices that are considered to be primal variables.
    
    DofIndexSet primals;
269
    DofContainerSet& vertices = 
270
      meshDistributor->getBoundaryDofInfo(feSpace).geoDofs[VERTEX];
271
272
273
274
    TEST_EXIT_DBG(vertices.size())("No primal vertices on this rank!\n");
    for (DofContainerSet::iterator it = vertices.begin(); 
	 it != vertices.end(); ++it)
      primals.insert(**it);
275

276
277
278
279

    // === Calculate the number of primals that are owned by the rank and ===
    // === create local indices of the primals starting at zero.          ===

280
    for (DofIndexSet::iterator it = primals.begin(); it != primals.end(); ++it)
281
282
      if (meshDistributor->getIsRankDof(feSpace, *it))
	primalDofMap[feSpace].insertRankDof(*it);
Thomas Witkowski's avatar
FETI-DP  
Thomas Witkowski committed
283
284
      else
  	primalDofMap[feSpace].insert(*it);
285
286
287
  }


288
  void PetscSolverFeti::createDuals(const FiniteElemSpace *feSpace)
289
290
  {
    FUNCNAME("PetscSolverFeti::createDuals()");
291

292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
    // === Create global index of the dual nodes on each rank. ===

    DofContainer allBoundaryDofs;
    meshDistributor->getAllBoundaryDofs(feSpace, allBoundaryDofs);

    for (DofContainer::iterator it = allBoundaryDofs.begin();
	 it != allBoundaryDofs.end(); ++it)
      if (!isPrimal(feSpace, **it))
	dualDofMap[feSpace].insertRankDof(**it);
  }

  
  void PetscSolverFeti::createLagrange(const FiniteElemSpace *feSpace)
  {
    FUNCNAME("PetscSolverFeti::createLagrange()");

308
309
    // === Create for each dual node that is owned by the rank, the set ===
    // === of ranks that contain this node (denoted by W(x_j)).         ===
310

311
    boundaryDofRanks[feSpace].clear();
312

313
314
315
    for (DofComm::Iterator it(meshDistributor->getSendDofs(), feSpace); 
	 !it.end(); it.nextRank())
      for (; !it.endDofIter(); it.nextDof()) {
316
	if (!isPrimal(feSpace, it.getDofIndex())) {
317
318
	  boundaryDofRanks[feSpace][it.getDofIndex()].insert(mpiRank);
	  boundaryDofRanks[feSpace][it.getDofIndex()].insert(it.getRank());
319
320
	}
      }
321
	
322
323
324
325

    // === Communicate these sets for all rank owned dual nodes to other ===
    // === ranks that also have this node.                               ===

326
    StdMpi<vector<std::set<int> > > stdMpi(meshDistributor->getMpiComm());
327
328
329
330

    for (DofComm::Iterator it(meshDistributor->getSendDofs(), feSpace);
	 !it.end(); it.nextRank())
      for (; !it.endDofIter(); it.nextDof())
331
	if (!isPrimal(feSpace, it.getDofIndex()))
332
	  stdMpi.getSendData(it.getRank()).push_back(boundaryDofRanks[feSpace][it.getDofIndex()]);
333
334
335

    stdMpi.updateSendDataSize();

336
337
    for (DofComm::Iterator it(meshDistributor->getRecvDofs(), feSpace); 
	 !it.end(); it.nextRank()) {
338
      bool recvFromRank = false;
339
      for (; !it.endDofIter(); it.nextDof()) {
340
	if (!isPrimal(feSpace, it.getDofIndex())) {
341
342
343
	  recvFromRank = true;
	  break;
	}
344
      }
345
346

      if (recvFromRank)
347
	stdMpi.recv(it.getRank());
348
    }
349

350
351
    stdMpi.startCommunication();

352
353
    for (DofComm::Iterator it(meshDistributor->getRecvDofs(), feSpace); 
	 !it.end(); it.nextRank()) {
354
      int i = 0;
355
      for (; !it.endDofIter(); it.nextDof())
356
	if (!isPrimal(feSpace, it.getDofIndex()))
357
	  boundaryDofRanks[feSpace][it.getDofIndex()] = 
358
	    stdMpi.getRecvData(it.getRank())[i++];
359
360
361
    }


362
363
364
    // === Reserve for each dual node, on the rank that owns this node, the ===
    // === appropriate number of Lagrange constraints.                      ===

365
    int nRankLagrange = 0;
Thomas Witkowski's avatar
Thomas Witkowski committed
366
367
    map<DegreeOfFreedom, MultiIndex>& dualMap = dualDofMap[feSpace].getMap();
    for (map<DegreeOfFreedom, MultiIndex>::iterator it = dualMap.begin(); it != dualMap.end(); ++it) {
368
      if (meshDistributor->getIsRankDof(feSpace, it->first)) {
Thomas Witkowski's avatar
Thomas Witkowski committed
369
	lagrangeMap[feSpace].insertRankDof(it->first, nRankLagrange);
370
	int degree = boundaryDofRanks[feSpace][it->first].size();
371
	nRankLagrange += (degree * (degree - 1)) / 2;
Thomas Witkowski's avatar
Thomas Witkowski committed
372
373
      } else {
	lagrangeMap[feSpace].insert(it->first);
374
375
      }
    }
376
    lagrangeMap[feSpace].nRankDofs = nRankLagrange;
377
378
379
  }


380
  void PetscSolverFeti::createIndexB(const FiniteElemSpace *feSpace)
381
  {
382
    FUNCNAME("PetscSolverFeti::createIndexB()");
383

384
    DOFAdmin* admin = feSpace->getAdmin();
385
386
387
388

    // === To ensure that all interior node on each rank are listen first in ===
    // === the global index of all B nodes, insert all interior nodes first, ===
    // === without defining a correct index.                                 ===
389

390
    int nLocalInterior = 0;    
391
    for (int i = 0; i < admin->getUsedSize(); i++) {
392
      if (admin->isDofFree(i) == false && 
393
	  isPrimal(feSpace, i) == false &&
394
	  dualDofMap[feSpace].isSet(i) == false) {
395
396
397
398
399
400
	localDofMap[feSpace].insertRankDof(i, nLocalInterior);

	if (fetiPreconditioner == FETI_DIRICHLET)
	  interiorDofMap[feSpace].insertRankDof(i, nLocalInterior);

	nLocalInterior++;
401
      }
402
    }
Thomas Witkowski's avatar
FETI-DP  
Thomas Witkowski committed
403
    
404
405
    // === And finally, add the global indicies of all dual nodes. ===

Thomas Witkowski's avatar
Thomas Witkowski committed
406
    for (map<DegreeOfFreedom, MultiIndex>::iterator it = dualDofMap[feSpace].getMap().begin();
407
	 it != dualDofMap[feSpace].getMap().end(); ++it)
408
      localDofMap[feSpace].insertRankDof(it->first);
409
410
411
  }


412
  void PetscSolverFeti::createMatLagrange(vector<const FiniteElemSpace*> &feSpaces)
413
414
415
  {
    FUNCNAME("PetscSolverFeti::createMatLagrange()");

416
417
    // === Create distributed matrix for Lagrange constraints. ===

418
    MatCreateMPIAIJ(PETSC_COMM_WORLD,
419
420
		    lagrangeMap.getRankDofs(), localDofMap.getRankDofs(),
		    lagrangeMap.getOverallDofs(), localDofMap.getOverallDofs(),	
421
422
423
		    2, PETSC_NULL, 2, PETSC_NULL,
		    &mat_lagrange);

424
425
426
427
428
429
430
    // === Create for all duals the corresponding Lagrange constraints. On ===
    // === each rank we traverse all pairs (n, m) of ranks, with n < m,    ===
    // === that contain this node. If the current rank number is r, and    ===
    // === n == r, the rank sets 1.0 for the corresponding constraint, if  ===
    // === m == r, than the rank sets -1.0 for the corresponding           ===
    // === constraint.                                                     ===

Thomas Witkowski's avatar
Thomas Witkowski committed
431
432
433
    for (unsigned int k = 0; k < feSpaces.size(); k++) {
      map<DegreeOfFreedom, MultiIndex> &dualMap = 
	dualDofMap[feSpaces[k]].getMap();
434

Thomas Witkowski's avatar
Thomas Witkowski committed
435
436
      for (map<DegreeOfFreedom, MultiIndex>::iterator it = dualMap.begin(); 
	   it != dualMap.end(); ++it) {
437
	TEST_EXIT_DBG(boundaryDofRanks[feSpaces[k]].count(it->first))
Thomas Witkowski's avatar
Thomas Witkowski committed
438
439
440
441
442
443
	  ("Should not happen!\n");
	
	// Global index of the first Lagrange constriant for this node.
	int index = lagrangeMap.getMatIndex(k, it->first);
	
	// Copy set of all ranks that contain this dual node.
444
445
	vector<int> W(boundaryDofRanks[feSpaces[k]][it->first].begin(), 
		      boundaryDofRanks[feSpaces[k]][it->first].end());
Thomas Witkowski's avatar
Thomas Witkowski committed
446
447
448
449
450
451
	// Number of ranks that contain this dual node.
	int degree = W.size();
	
	for (int i = 0; i < degree; i++) {
	  for (int j = i + 1; j < degree; j++) {
	    if (W[i] == mpiRank || W[j] == mpiRank) {
452
	      int colIndex = localDofMap.getMatIndex(k, it->first);
453
	      double value = (W[i] == mpiRank ? 1.0 : -1.0);
454
	      MatSetValue(mat_lagrange, index, colIndex, value, INSERT_VALUES);
455
	    }
Thomas Witkowski's avatar
Thomas Witkowski committed
456
	    index++;	      
457
458
459
460
461
462
463
464
465
466
	  }
	}
      }
    }

    MatAssemblyBegin(mat_lagrange, MAT_FINAL_ASSEMBLY);
    MatAssemblyEnd(mat_lagrange, MAT_FINAL_ASSEMBLY);
  }


467
  void PetscSolverFeti::createSchurPrimalKsp(vector<const FiniteElemSpace*> &feSpaces)
468
469
470
  {
    FUNCNAME("PetscSolverFeti::createSchurPrimal()");

Thomas Witkowski's avatar
Thomas Witkowski committed
471
    if (schurPrimalSolver == 0) {
472
473
      MSG("Create iterative schur primal solver!\n");

Thomas Witkowski's avatar
Thomas Witkowski committed
474
475
476
477
478
479
      schurPrimalData.mat_primal_primal = &mat_primal_primal;
      schurPrimalData.mat_primal_b = &mat_primal_b;
      schurPrimalData.mat_b_primal = &mat_b_primal;
      schurPrimalData.fetiSolver = this;
      
      VecCreateMPI(PETSC_COMM_WORLD, 
480
		   localDofMap.getRankDofs(), localDofMap.getOverallDofs(),
Thomas Witkowski's avatar
Thomas Witkowski committed
481
482
		   &(schurPrimalData.tmp_vec_b));
      VecCreateMPI(PETSC_COMM_WORLD, 
483
		   primalDofMap.getRankDofs(), primalDofMap.getOverallDofs(),
Thomas Witkowski's avatar
Thomas Witkowski committed
484
485
486
		   &(schurPrimalData.tmp_vec_primal));

      MatCreateShell(PETSC_COMM_WORLD,
487
488
		     primalDofMap.getRankDofs(), primalDofMap.getRankDofs(), 
		     primalDofMap.getOverallDofs(), primalDofMap.getOverallDofs(),
Thomas Witkowski's avatar
Thomas Witkowski committed
489
490
491
492
493
494
495
496
497
498
		     &schurPrimalData, 
		     &mat_schur_primal);
      MatShellSetOperation(mat_schur_primal, MATOP_MULT, 
			   (void(*)(void))petscMultMatSchurPrimal);
      
      KSPCreate(PETSC_COMM_WORLD, &ksp_schur_primal);
      KSPSetOperators(ksp_schur_primal, mat_schur_primal, mat_schur_primal, SAME_NONZERO_PATTERN);
      KSPSetOptionsPrefix(ksp_schur_primal, "solver_sp_");
      KSPSetFromOptions(ksp_schur_primal);
    } else {
499
500
      MSG("Create direct schur primal solver!\n");

501
502
503
504
      TEST_EXIT(ksp_b)("No solver object for local problem!\n");

      double wtime = MPI::Wtime();

505
506
      int nRowsRankPrimal = primalDofMap.getRankDofs();
      int nRowsOverallPrimal = primalDofMap.getOverallDofs();
507
508
      int nRowsRankB = localDofMap.getRankDofs();
      int nRowsOverallB = localDofMap.getOverallDofs();
509

Thomas Witkowski's avatar
Thomas Witkowski committed
510
511
512
513
514
515
      Mat matBPi;
      MatCreateMPIAIJ(PETSC_COMM_WORLD,
		      nRowsRankB, nRowsRankPrimal, 
		      nRowsOverallB, nRowsOverallPrimal,
		      30, PETSC_NULL, 30, PETSC_NULL, &matBPi);
      Mat matPrimal;
516

517
518
519
      PetscInt nCols;
      const PetscInt *cols;
      const PetscScalar *values;
Thomas Witkowski's avatar
Thomas Witkowski committed
520

521
522
      map<int, vector<pair<int, double> > > mat_b_primal_cols;

523
524
      for (int i = 0; i < nRowsRankB; i++) {
	PetscInt row = localDofMap.getStartDofs() + i;
525
526
527
528
529
530
	MatGetRow(mat_b_primal, row, &nCols, &cols, &values);

	for (int j = 0; j < nCols; j++)
	  if (values[j] != 0.0)
	    mat_b_primal_cols[cols[j]].push_back(make_pair(i, values[j]));

Thomas Witkowski's avatar
Thomas Witkowski committed
531
	MatRestoreRow(mat_b_primal, row, &nCols, &cols, &values);
532
533
534
535
536
      }

      int maxLocalPrimal = mat_b_primal_cols.size();
      mpi::globalMax(maxLocalPrimal);

537
538
      TEST_EXIT(static_cast<int>(mat_b_primal_cols.size()) == 
		primalDofMap.getLocalDofs())
539
	("Should not happen!\n");
540

541
542
543
      for (map<int, vector<pair<int, double> > >::iterator it = mat_b_primal_cols.begin();
	   it != mat_b_primal_cols.end(); ++it) {
	Vec tmpVec;
544
	VecCreateSeq(PETSC_COMM_SELF, nRowsRankB, &tmpVec);
545
546
547
548
549
550
551
552
553
554

 	for (unsigned int i = 0; i < it->second.size(); i++) 
 	  VecSetValue(tmpVec, 
 		      it->second[i].first, it->second[i].second, INSERT_VALUES);

	VecAssemblyBegin(tmpVec);
	VecAssemblyEnd(tmpVec);

       	KSPSolve(ksp_b, tmpVec, tmpVec);

Thomas Witkowski's avatar
Thomas Witkowski committed
555
	PetscScalar *tmpValues;
556
	VecGetArray(tmpVec, &tmpValues);
557
	for (int i  = 0; i < nRowsRankB; i++)
Thomas Witkowski's avatar
Thomas Witkowski committed
558
	  MatSetValue(matBPi, 
559
		      localDofMap.getStartDofs() + i,
Thomas Witkowski's avatar
Thomas Witkowski committed
560
561
562
		      it->first,
		      tmpValues[i],
		      ADD_VALUES);
563
564
565
566
567
	VecRestoreArray(tmpVec, &tmpValues);

	VecDestroy(&tmpVec);
      }

Thomas Witkowski's avatar
Thomas Witkowski committed
568
569
570
      MatAssemblyBegin(matBPi, MAT_FINAL_ASSEMBLY);
      MatAssemblyEnd(matBPi, MAT_FINAL_ASSEMBLY);
      MatMatMult(mat_primal_b, matBPi, MAT_INITIAL_MATRIX, PETSC_DEFAULT, &matPrimal);
571
572
573
      MatAXPY(mat_primal_primal, -1.0, matPrimal, DIFFERENT_NONZERO_PATTERN);

      MatDestroy(&matPrimal);
Thomas Witkowski's avatar
Thomas Witkowski committed
574
      MatDestroy(&matBPi);
575
576
577
578
579
580
581
582
583
584

      MatInfo minfo;
      MatGetInfo(mat_primal_primal, MAT_GLOBAL_SUM, &minfo);
      MSG("Schur primal matrix nnz = %f\n", minfo.nz_used);

      KSPCreate(PETSC_COMM_WORLD, &ksp_schur_primal);
      KSPSetOperators(ksp_schur_primal, mat_primal_primal, 
		      mat_primal_primal, SAME_NONZERO_PATTERN);
      KSPSetOptionsPrefix(ksp_schur_primal, "solver_sp_");
      KSPSetFromOptions(ksp_schur_primal);
Thomas Witkowski's avatar
Thomas Witkowski committed
585

586
587
      MSG("Creating Schur primal matrix needed %.5f seconds.\n",
	  MPI::Wtime() - wtime);
Thomas Witkowski's avatar
Thomas Witkowski committed
588
    }
589
590
591
592
593
594
595
  }


  void PetscSolverFeti::destroySchurPrimalKsp()
  {
    FUNCNAME("PetscSolverFeti::destroySchurPrimal()");

596
597
598
599
600
    if (schurPrimalSolver == 0) {
      schurPrimalData.mat_primal_primal = PETSC_NULL;
      schurPrimalData.mat_primal_b = PETSC_NULL;
      schurPrimalData.mat_b_primal = PETSC_NULL;
      schurPrimalData.fetiSolver = NULL;
601

602
603
      VecDestroy(&schurPrimalData.tmp_vec_b);
      VecDestroy(&schurPrimalData.tmp_vec_primal);
604

605
606
607
608
609
      MatDestroy(&mat_schur_primal);
      KSPDestroy(&ksp_schur_primal);
    } else {
      KSPDestroy(&ksp_schur_primal);
    }
610
611
612
  }


613
  void PetscSolverFeti::createFetiKsp(vector<const FiniteElemSpace*> &feSpaces)
614
615
616
  {
    FUNCNAME("PetscSolverFeti::createFetiKsp()");

617
618
    // === Create FETI-DP solver object. ===

619
620
621
    fetiData.mat_primal_b = &mat_primal_b;
    fetiData.mat_b_primal = &mat_b_primal;
    fetiData.mat_lagrange = &mat_lagrange;
Thomas Witkowski's avatar
Thomas Witkowski committed
622
    fetiData.fetiSolver = this;
623
    fetiData.ksp_schur_primal = &ksp_schur_primal;
624

625
    VecCreateMPI(PETSC_COMM_WORLD, 
626
		 localDofMap.getRankDofs(), localDofMap.getOverallDofs(),
Thomas Witkowski's avatar
Thomas Witkowski committed
627
628
		 &(fetiData.tmp_vec_b));
    VecCreateMPI(PETSC_COMM_WORLD,
629
		 lagrangeMap.getRankDofs(), lagrangeMap.getOverallDofs(),
Thomas Witkowski's avatar
Thomas Witkowski committed
630
		 &(fetiData.tmp_vec_lagrange));
631
    VecCreateMPI(PETSC_COMM_WORLD, 
632
		 primalDofMap.getRankDofs(), primalDofMap.getOverallDofs(),
633
		 &(fetiData.tmp_vec_primal));
634
635

    MatCreateShell(PETSC_COMM_WORLD,
636
637
		   lagrangeMap.getRankDofs(), lagrangeMap.getRankDofs(),
		   lagrangeMap.getOverallDofs(), lagrangeMap.getOverallDofs(),
638
		   &fetiData, &mat_feti);
639
640
641
642
643
644
645
    MatShellSetOperation(mat_feti, MATOP_MULT, (void(*)(void))petscMultMatFeti);


    KSPCreate(PETSC_COMM_WORLD, &ksp_feti);
    KSPSetOperators(ksp_feti, mat_feti, mat_feti, SAME_NONZERO_PATTERN);
    KSPSetOptionsPrefix(ksp_feti, "solver_feti_");
    KSPSetFromOptions(ksp_feti);
646
647


648
    // === Create FETI-DP preconditioner object. ===
649

650
651
652
653
    if (fetiPreconditioner != FETI_NONE) {
      MatDuplicate(mat_lagrange, MAT_COPY_VALUES, &mat_lagrange_scaled);
      MatScale(mat_lagrange_scaled, 0.5);
    }
654

655
656
657
    switch (fetiPreconditioner) {
    case FETI_DIRICHLET:           
      KSPCreate(PETSC_COMM_SELF, &ksp_interior);
Thomas Witkowski's avatar
Thomas Witkowski committed
658
659
      KSPSetOperators(ksp_interior, mat_interior_interior, mat_interior_interior, 
		      SAME_NONZERO_PATTERN);
660
661
662
663
664
665
666
667
668
669
      KSPSetOptionsPrefix(ksp_interior, "solver_interior_");
      KSPSetFromOptions(ksp_interior);
            
      fetiDirichletPreconData.mat_lagrange_scaled = &mat_lagrange_scaled;
      fetiDirichletPreconData.mat_interior_interior = &mat_interior_interior;
      fetiDirichletPreconData.mat_duals_duals = &mat_duals_duals;
      fetiDirichletPreconData.mat_interior_duals = &mat_interior_duals;
      fetiDirichletPreconData.mat_duals_interior = &mat_duals_interior;
      fetiDirichletPreconData.ksp_interior = &ksp_interior;
      
670
      VecCreateMPI(PETSC_COMM_WORLD, 
671
		   localDofMap.getRankDofs(),localDofMap.getOverallDofs(),
672
		   &(fetiDirichletPreconData.tmp_vec_b));      
673
674
675
676
677
678
      MatGetVecs(mat_duals_duals, PETSC_NULL, 
		 &(fetiDirichletPreconData.tmp_vec_duals0));
      MatGetVecs(mat_duals_duals, PETSC_NULL, 
		 &(fetiDirichletPreconData.tmp_vec_duals1));
      MatGetVecs(mat_interior_interior, PETSC_NULL, 
		 &(fetiDirichletPreconData.tmp_vec_interior));
679
680
681
682
683
684
685
686
687
688
689
690
691

      for (unsigned int i = 0; i < feSpaces.size(); i++) {
	map<DegreeOfFreedom, MultiIndex> &dualMap = 
	  dualDofMap[feSpaces[i]].getMap();
	for (map<DegreeOfFreedom, MultiIndex>::iterator it = dualMap.begin(); 
	     it != dualMap.end(); ++it) {
	  DegreeOfFreedom d = it->first;
	  int matIndexLocal = localDofMap.getLocalMatIndex(i, d);
	  int matIndexDual = dualDofMap.getLocalMatIndex(i, d);
	  fetiDirichletPreconData.localToDualMap[matIndexLocal] = matIndexDual;
	}
      }

692
693
694
695
696
697
698
699
700
701
702
      KSPGetPC(ksp_feti, &precon_feti);
      PCSetType(precon_feti, PCSHELL);
      PCShellSetContext(precon_feti, static_cast<void*>(&fetiDirichletPreconData));
      PCShellSetApply(precon_feti, petscApplyFetiDirichletPrecon);
      
      break;

    case FETI_LUMPED:
      fetiLumpedPreconData.mat_lagrange_scaled = &mat_lagrange_scaled;
      fetiLumpedPreconData.mat_duals_duals = &mat_duals_duals;

703
      VecCreateMPI(PETSC_COMM_WORLD, 
704
705
		   localDofMap.getRankDofs(),
		   localDofMap.getOverallDofs(),
706
		   &(fetiLumpedPreconData.tmp_vec_b));
707
708
709
710
      MatGetVecs(mat_duals_duals, PETSC_NULL, 
		 &(fetiLumpedPreconData.tmp_vec_duals0));
      MatGetVecs(mat_duals_duals, PETSC_NULL, 
		 &(fetiLumpedPreconData.tmp_vec_duals1));
711
712
713
714
715
716

      KSPGetPC(ksp_feti, &precon_feti);
      PCSetType(precon_feti, PCSHELL);
      PCShellSetContext(precon_feti, static_cast<void*>(&fetiLumpedPreconData));
      PCShellSetApply(precon_feti, petscApplyFetiLumpedPrecon);
      
717
718
      break;
    default:
719
720
      break;
    }
721
722
723
724
725
726
727
  }
  

  void PetscSolverFeti::destroyFetiKsp()
  {
    FUNCNAME("PetscSolverFeti::destroyFetiKsp()");

728
729
    // === Destroy FETI-DP solver object. ===

730
731
732
    fetiData.mat_primal_b = PETSC_NULL;
    fetiData.mat_b_primal = PETSC_NULL;
    fetiData.mat_lagrange = PETSC_NULL;
Thomas Witkowski's avatar
Thomas Witkowski committed
733
    fetiData.fetiSolver = NULL;
734
    fetiData.ksp_schur_primal = PETSC_NULL;
735

Thomas Witkowski's avatar
Thomas Witkowski committed
736
737
    VecDestroy(&fetiData.tmp_vec_b);
    VecDestroy(&fetiData.tmp_vec_lagrange);
738
    VecDestroy(&fetiData.tmp_vec_primal);
739
740
    MatDestroy(&mat_feti);
    KSPDestroy(&ksp_feti);
741
742


743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
    // === Destroy FETI-DP preconditioner object. ===

    switch (fetiPreconditioner) {
    case FETI_DIRICHLET:           
      KSPDestroy(&ksp_interior);

      fetiDirichletPreconData.mat_lagrange_scaled = NULL;
      fetiDirichletPreconData.mat_interior_interior = NULL;
      fetiDirichletPreconData.mat_duals_duals = NULL;
      fetiDirichletPreconData.mat_interior_duals = NULL;
      fetiDirichletPreconData.mat_duals_interior = NULL;
      fetiDirichletPreconData.ksp_interior = NULL;
      
      VecDestroy(&fetiDirichletPreconData.tmp_vec_b);
      VecDestroy(&fetiDirichletPreconData.tmp_vec_duals0);
      VecDestroy(&fetiDirichletPreconData.tmp_vec_duals1);
      VecDestroy(&fetiDirichletPreconData.tmp_vec_interior);
      MatDestroy(&mat_lagrange_scaled);
      break;

    case FETI_LUMPED:
      fetiLumpedPreconData.mat_lagrange_scaled = NULL;
      fetiLumpedPreconData.mat_duals_duals = NULL;

      VecDestroy(&fetiLumpedPreconData.tmp_vec_b);
      VecDestroy(&fetiLumpedPreconData.tmp_vec_duals0);
      VecDestroy(&fetiLumpedPreconData.tmp_vec_duals1);
      break;
771
772
    default:
      break;
773
    }
774
775
776
777
778
779
780
781
782
  }


  void PetscSolverFeti::recoverSolution(Vec &vec_sol_b,
					Vec &vec_sol_primal,
					SystemVector &vec)
  {
    FUNCNAME("PetscSolverFeti::recoverSolution()");

783
    // === Get local part of the solution for B variables. ===
784
785
786
787
788

    PetscScalar *localSolB;
    VecGetArray(vec_sol_b, &localSolB);


789
790
    // === Create scatter to get solutions of all primal nodes that are ===
    // === contained in rank's domain.                                  ===
791
    
Thomas Witkowski's avatar
Thomas Witkowski committed
792
793
    vector<const FiniteElemSpace*> feSpaces = getFeSpaces(&vec);

794
    vector<PetscInt> globalIsIndex, localIsIndex;
Thomas Witkowski's avatar
Thomas Witkowski committed
795
796
    globalIsIndex.reserve(primalDofMap.getLocalDofs());
    localIsIndex.reserve(primalDofMap.getLocalDofs());
797
798

    {
Thomas Witkowski's avatar
Thomas Witkowski committed
799
800
801
802
803
804
805
806
      int cnt = 0;
      for (unsigned int i = 0; i < feSpaces.size(); i++) {
	map<DegreeOfFreedom, MultiIndex>& dofMap = 
	  primalDofMap[feSpaces[i]].getMap();
	for (map<DegreeOfFreedom, MultiIndex>::iterator it = dofMap.begin();
	     it != dofMap.end(); ++it) {
	  globalIsIndex.push_back(primalDofMap.getMatIndex(i, it->first));
	  localIsIndex.push_back(cnt++);
807
808
	}
      }
Thomas Witkowski's avatar
Thomas Witkowski committed
809
810

      TEST_EXIT_DBG(cnt == primalDofMap.getLocalDofs())("Should not happen!\n");
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
    }
    
    IS globalIs, localIs;
    ISCreateGeneral(PETSC_COMM_SELF, 
		    globalIsIndex.size(), 
		    &(globalIsIndex[0]),
		    PETSC_USE_POINTER,
		    &globalIs);

    ISCreateGeneral(PETSC_COMM_SELF, 
		    localIsIndex.size(), 
		    &(localIsIndex[0]),
		    PETSC_USE_POINTER,
		    &localIs);

    Vec local_sol_primal;
    VecCreateSeq(PETSC_COMM_SELF, localIsIndex.size(), &local_sol_primal);

    VecScatter primalScatter;
    VecScatterCreate(vec_sol_primal, globalIs, local_sol_primal, localIs, &primalScatter);
    VecScatterBegin(primalScatter, vec_sol_primal, local_sol_primal, 
		    INSERT_VALUES, SCATTER_FORWARD);
    VecScatterEnd(primalScatter, vec_sol_primal, local_sol_primal, 
		  INSERT_VALUES, SCATTER_FORWARD);

836
837
838
    ISDestroy(&globalIs);
    ISDestroy(&localIs);    
    VecScatterDestroy(&primalScatter);    
839
840
841
842

    PetscScalar *localSolPrimal;
    VecGetArray(local_sol_primal, &localSolPrimal);

843
    // === And copy from PETSc local vectors to the DOF vectors. ===
844

Thomas Witkowski's avatar
Thomas Witkowski committed
845
846
    int cnt = 0;
    for (int i = 0; i < vec.getSize(); i++) {
847
848
      DOFVector<double>& dofVec = *(vec.getDOFVector(i));

Thomas Witkowski's avatar
Thomas Witkowski committed
849
850
      for (map<DegreeOfFreedom, MultiIndex>::iterator it = localDofMap[feSpaces[i]].getMap().begin();
	   it != localDofMap[feSpaces[i]].getMap().end(); ++it) {
851
	int petscIndex = localDofMap.getLocalMatIndex(i, it->first);
852
853
854
	dofVec[it->first] = localSolB[petscIndex];
      }

Thomas Witkowski's avatar
Thomas Witkowski committed
855
856
857
      for (map<DegreeOfFreedom, MultiIndex>::iterator it = primalDofMap[feSpaces[i]].getMap().begin();
	   it != primalDofMap[feSpaces[i]].getMap().end(); ++it)
	dofVec[it->first] = localSolPrimal[cnt++];
858
859
860
861
    }

    VecRestoreArray(vec_sol_b, &localSolB);
    VecRestoreArray(local_sol_primal, &localSolPrimal);
862
    VecDestroy(&local_sol_primal);
863
864
865
  }


866
  void PetscSolverFeti::fillPetscMatrix(Matrix<DOFMatrix*> *mat)
867
868
  {
    FUNCNAME("PetscSolverFeti::fillPetscMatrix()");   
869
    
870
    // === Create all sets and indices. ===
871
    
Thomas Witkowski's avatar
Thomas Witkowski committed
872
    vector<const FiniteElemSpace*> feSpaces = getFeSpaces(mat);
873
    
874
875
876
877
    primalDofMap.init(mpiComm, feSpaces, true, true);
    dualDofMap.init(mpiComm, feSpaces, false, false);
    lagrangeMap.init(mpiComm, feSpaces, true, true);
    localDofMap.init(mpiComm, feSpaces, false, false);
878
879
    if (fetiPreconditioner == FETI_DIRICHLET)
      interiorDofMap.init(mpiComm, feSpaces, false, false);
880

881
    updateDofData();
882
    
883
    // === Create matrices for the FETI-DP method. ===
884
    
885
886
    int nRowsRankB = localDofMap.getRankDofs();
    int nRowsOverallB = localDofMap.getOverallDofs();
887
888
    int nRowsRankPrimal = primalDofMap.getRankDofs();
    int nRowsOverallPrimal = primalDofMap.getOverallDofs();
889
    
Thomas Witkowski's avatar
Thomas Witkowski committed
890
891
    MatCreateSeqAIJ(PETSC_COMM_SELF, nRowsRankB, nRowsRankB, 30, PETSC_NULL,
		    &mat_b_b);
892
    
893
    MatCreateMPIAIJ(PETSC_COMM_WORLD,
894
895
		    nRowsRankPrimal, nRowsRankPrimal, 
		    nRowsOverallPrimal, nRowsOverallPrimal,
896
		    30, PETSC_NULL, 30, PETSC_NULL, &mat_primal_primal);
897
    
898
    MatCreateMPIAIJ(PETSC_COMM_WORLD,
899
900
		    nRowsRankB, nRowsRankPrimal, 
		    nRowsOverallB, nRowsOverallPrimal,
901
		    30, PETSC_NULL, 30, PETSC_NULL, &mat_b_primal);
902
    
903
    MatCreateMPIAIJ(PETSC_COMM_WORLD,
904
905
		    nRowsRankPrimal, nRowsRankB,
		    nRowsOverallPrimal, nRowsOverallB,
Thomas Witkowski's avatar
Thomas Witkowski committed
906
		    15, PETSC_NULL, 15, PETSC_NULL, &mat_primal_b);
907
908
    
    
909
    // === Create matrices for FETI-DP preconditioner. ===
910
911
912
    
    if (fetiPreconditioner != FETI_NONE) {
      int nRowsDual = dualDofMap.getRankDofs();
913
914

      MatCreateSeqAIJ(PETSC_COMM_SELF,
915
		      nRowsDual, nRowsDual, 30, PETSC_NULL,
916
917
		      &mat_duals_duals);

918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
      if (fetiPreconditioner == FETI_DIRICHLET) {
	int nRowsInterior = interiorDofMap.getRankDofs();

	MatCreateSeqAIJ(PETSC_COMM_SELF,
			nRowsInterior, nRowsInterior, 30, PETSC_NULL,
			&mat_interior_interior);
	
	MatCreateSeqAIJ(PETSC_COMM_SELF,
			nRowsInterior, nRowsDual, 30, PETSC_NULL,
			&mat_interior_duals);
	
	MatCreateSeqAIJ(PETSC_COMM_SELF,
			nRowsDual, nRowsInterior, 30, PETSC_NULL,
			&mat_duals_interior);
      }
933
    }
934

935
936
    
    // === Prepare traverse of sequentially created matrices. ===
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951

    using mtl::tag::row; using mtl::tag::nz; using mtl::begin; using mtl::end;
    namespace traits = mtl::traits;
    typedef DOFMatrix::base_matrix_type Matrix;

    typedef traits::range_generator<row, Matrix>::type cursor_type;
    typedef traits::range_generator<nz, cursor_type>::type icursor_type;

    vector<int> cols, colsOther;
    vector<double> values, valuesOther;
    cols.reserve(300);
    colsOther.reserve(300);
    values.reserve(300);
    valuesOther.reserve(300);

952
953
954
955
956
957
958
    vector<int> colsLocal, colsLocalOther;
    vector<double> valuesLocal, valuesLocalOther;
    colsLocal.reserve(300);
    colsLocalOther.reserve(300);
    valuesLocal.reserve(300);
    valuesLocalOther.reserve(300);

959
960
961
962

    // === Traverse all sequentially created matrices and add the values to ===
    // === the global PETSc matrices.                                       ===

Thomas Witkowski's avatar
Thomas Witkowski committed
963
    int nComponents = mat->getSize();
964
965
966
967
968
969
970
971
972
973
974
    for (int i = 0; i < nComponents; i++) {
      for (int j = 0; j < nComponents; j++) {
	if (!(*mat)[i][j])
	  continue;

	traits::col<Matrix>::type col((*mat)[i][j]->getBaseMatrix());
	traits::const_value<Matrix>::type value((*mat)[i][j]->getBaseMatrix());
	
	// Traverse all rows.
	for (cursor_type cursor = begin<row>((*mat)[i][j]->getBaseMatrix()), 
	       cend = end<row>((*mat)[i][j]->getBaseMatrix()); cursor != cend; ++cursor) {
975

Thomas Witkowski's avatar
Thomas Witkowski committed
976
	  bool rowPrimal = isPrimal(feSpaces[i], *cursor);
977