PetscSolverFeti.cc 41.4 KB
Newer Older
1
2
3
4
5
6
7
8
9
10
11
12
//
// 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.


13
#include "AMDiS.h"
14
#include "parallel/PetscSolverFeti.h"
15
#include "parallel/PetscSolverFetiStructs.h"
16
17
#include "parallel/StdMpi.h"
#include "parallel/MpiHelper.h"
18
#include "parallel/SubDomainSolver.h"
19
#include "io/VtkWriter.h"
20
21
22
23
24

namespace AMDiS {

  using namespace std;

25
26
27
28
29
30
31
  // 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);
32
    SchurPrimalData* data = static_cast<SchurPrimalData*>(ctx);
33

34
35
36
37
    MatMult(data->subSolver->getMatIntCoarse(), x, data->tmp_vec_b);
    data->subSolver->solveGlobal(data->tmp_vec_b, data->tmp_vec_b);
    MatMult(data->subSolver->getMatCoarseInt(), data->tmp_vec_b, data->tmp_vec_primal);
    MatMult(data->subSolver->getMatCoarseCoarse(), x, y);
38
39
40
41
42
43
44
45
46
    VecAXPBY(y, -1.0, 1.0, data->tmp_vec_primal);

    return 0;
  }


  // y = mat * x
  int petscMultMatFeti(Mat mat, Vec x, Vec y)
  {
47
48
    //    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)
49
50
51

    void *ctx;
    MatShellGetContext(mat, &ctx);
52
    FetiData* data = static_cast<FetiData*>(ctx);
53

Thomas Witkowski's avatar
Thomas Witkowski committed
54
    MatMultTranspose(*(data->mat_lagrange), x, data->tmp_vec_b);
55
    data->subSolver->solveGlobal(data->tmp_vec_b, data->tmp_vec_b);
Thomas Witkowski's avatar
Thomas Witkowski committed
56
    MatMult(*(data->mat_lagrange), data->tmp_vec_b, data->tmp_vec_lagrange);
Thomas Witkowski's avatar
Thomas Witkowski committed
57

58
    MatMult(data->subSolver->getMatCoarseInt(), data->tmp_vec_b, data->tmp_vec_primal);
59
    KSPSolve(*(data->ksp_schur_primal), data->tmp_vec_primal, data->tmp_vec_primal);
60
   MatMult(data->subSolver->getMatIntCoarse(), data->tmp_vec_primal, data->tmp_vec_b);
61
    data->subSolver->solveGlobal(data->tmp_vec_b, data->tmp_vec_b);
Thomas Witkowski's avatar
Thomas Witkowski committed
62
    MatMult(*(data->mat_lagrange), data->tmp_vec_b, y);
63

Thomas Witkowski's avatar
Thomas Witkowski committed
64
    VecAXPBY(y, 1.0, 1.0, data->tmp_vec_lagrange);
65
66
67
68
69

    return 0;
  }


70
  // y = PC * x
71
  PetscErrorCode petscApplyFetiDirichletPrecon(PC pc, Vec x, Vec y)
72
  {
73
    // Get data for the preconditioner
74
75
    void *ctx;
    PCShellGetContext(pc, &ctx);
76
    FetiDirichletPreconData* data = static_cast<FetiDirichletPreconData*>(ctx);
77

78
    // Multiply with scaled Lagrange constraint matrix.
79
80
81
    MatMultTranspose(*(data->mat_lagrange_scaled), x, data->tmp_vec_b);


82
    // === Restriction of the B nodes to the boundary nodes. ===
83

84
85
86
87
    int nLocalB;
    int nLocalDuals;
    VecGetLocalSize(data->tmp_vec_b, &nLocalB);
    VecGetLocalSize(data->tmp_vec_duals0, &nLocalDuals);
88

89
90
91
92
    PetscScalar *local_b, *local_duals;
    VecGetArray(data->tmp_vec_b, &local_b);
    VecGetArray(data->tmp_vec_duals0, &local_duals);

93
94
95
    for (map<int, int>::iterator it = data->localToDualMap.begin();
	 it != data->localToDualMap.end(); ++it)
      local_duals[it->second] = local_b[it->first];
96
97

    VecRestoreArray(data->tmp_vec_b, &local_b);
98
    VecRestoreArray(data->tmp_vec_duals0, &local_duals);
99
100


101
    // === K_DD - K_DI inv(K_II) K_ID ===
102

103
    MatMult(*(data->mat_duals_duals), data->tmp_vec_duals0, data->tmp_vec_duals1);
104

105
    MatMult(*(data->mat_interior_duals), data->tmp_vec_duals0, data->tmp_vec_interior);
106
    KSPSolve(*(data->ksp_interior), data->tmp_vec_interior, data->tmp_vec_interior);
107
108
109
110
111
112
113
114
115
116
    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);

117
118
119
    for (map<int, int>::iterator it = data->localToDualMap.begin();
	 it != data->localToDualMap.end(); ++it)
      local_b[it->first] = local_duals[it->second];
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141

    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);
142
143


144
    // === Restriction of the B nodes to the boundary nodes. ===
145

146
147
148
149
    int nLocalB;
    int nLocalDuals;
    VecGetLocalSize(data->tmp_vec_b, &nLocalB);
    VecGetLocalSize(data->tmp_vec_duals0, &nLocalDuals);
150

151
    PetscScalar *local_b, *local_duals;
152
    VecGetArray(data->tmp_vec_b, &local_b);
153
    VecGetArray(data->tmp_vec_duals0, &local_duals);
154

Thomas Witkowski's avatar
Thomas Witkowski committed
155
156
157
    for (map<int, int>::iterator it = data->localToDualMap.begin();
	 it != data->localToDualMap.end(); ++it)
      local_duals[it->second] = local_b[it->first];
158
159

    VecRestoreArray(data->tmp_vec_b, &local_b);
160
161
162
163
164
165
166
    VecRestoreArray(data->tmp_vec_duals0, &local_duals);


    // === K_DD ===

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

167

168
    // === Prolongation from local dual nodes to B nodes.
169

170
171
172
    VecGetArray(data->tmp_vec_b, &local_b);
    VecGetArray(data->tmp_vec_duals1, &local_duals);

Thomas Witkowski's avatar
Thomas Witkowski committed
173
174
175
    for (map<int, int>::iterator it = data->localToDualMap.begin();
	 it != data->localToDualMap.end(); ++it)
      local_b[it->first] = local_duals[it->second];
176

177
178
    VecRestoreArray(data->tmp_vec_b, &local_b);
    VecRestoreArray(data->tmp_vec_duals0, &local_duals);
179

180
181

    // Multiply with scaled Lagrange constraint matrix.
182
183
184
185
186
187
    MatMult(*(data->mat_lagrange_scaled), data->tmp_vec_b, y);

    return 0;
  }


188
189
  PetscSolverFeti::PetscSolverFeti()
    : PetscSolver(),
190
      schurPrimalSolver(0),
191
      multiLevelTest(false),
192
193
      subDomainSolver(NULL),
      meshLevel(0)
194
195
196
197
198
199
  {
    FUNCNAME("PetscSolverFeti::PetscSolverFeti()");

    string preconditionerName = "";
    Parameters::get("parallel->solver->precon", preconditionerName);
    if (preconditionerName == "" || preconditionerName == "none") {
200
      MSG("Create FETI-DP solver with no preconditioner!\n");
201
202
      fetiPreconditioner = FETI_NONE;
    } else if (preconditionerName == "dirichlet") {
203
      MSG("Create FETI-DP solver with Dirichlet preconditioner!\n");
204
205
      fetiPreconditioner = FETI_DIRICHLET;
    } else if (preconditionerName == "lumped") {
206
      MSG("Create FETI-DP solver with lumped preconditioner!\n");
207
208
      fetiPreconditioner = FETI_LUMPED;
    } else {
Thomas Witkowski's avatar
Thomas Witkowski committed
209
210
      ERROR_EXIT("Preconditioner \"%s\" not available!\n", 
		 preconditionerName.c_str());
211
    }
Thomas Witkowski's avatar
Thomas Witkowski committed
212
213
214
215
216

    Parameters::get("parallel->feti->schur primal solver", schurPrimalSolver);
    TEST_EXIT(schurPrimalSolver == 0 || schurPrimalSolver == 1)
      ("Wrong solver \"%d\"for the Schur primal complement!\n", 
       schurPrimalSolver);
217
218

    Parameters::get("parallel->multi level test", multiLevelTest);
219
220
    if (multiLevelTest)
      meshLevel = 1;
221
222
223
  }


224
  void PetscSolverFeti::initialize(vector<const FiniteElemSpace*> feSpaces)
225
  {
226
227
    FUNCNAME("PetscSolverFeti::initialize()");

228
229
230
    TEST_EXIT_DBG(meshLevel + 1 == meshDistributor->getMeshLevelData().getLevelNumber())
      ("Mesh hierarchy does not contain %d levels!\n", meshLevel + 1);

231
232
    MeshLevelData& levelData = meshDistributor->getMeshLevelData();

233
234
235
236
237
238
239
240
241
    if (subDomainSolver == NULL) {
      if (meshLevel == 0) {
	subDomainSolver = 
	  new SubDomainSolver(meshDistributor, mpiComm, mpiSelfComm);
      } else {
	subDomainSolver = 
	  new SubDomainSolver(meshDistributor, 
			      levelData.getMpiComm(meshLevel - 1),
			      levelData.getMpiComm(meshLevel));
242
	subDomainSolver->setLevel(meshLevel);
243
244
      }
    }
245

246
247
    primalDofMap.init(levelData,
		      feSpaces, meshDistributor->getFeSpaces(), 
248
		      true, true);
249
250
    primalDofMap.setComputeMatIndex(true);
    
251
252
    dualDofMap.init(levelData, 
		    feSpaces, meshDistributor->getFeSpaces(),
253
		    false, false);
254
255
    dualDofMap.setComputeMatIndex(true);

256
257
    lagrangeMap.init(levelData, 
		     feSpaces, meshDistributor->getFeSpaces(),
258
		     true, true);
259
260
    lagrangeMap.setComputeMatIndex(true);
 
261
262
    localDofMap.init(levelData, 
		     feSpaces, meshDistributor->getFeSpaces(),
263
		     false, false);
264
265
266
    localDofMap.setComputeMatIndex(true);

    if (fetiPreconditioner == FETI_DIRICHLET) {
267
268
      interiorDofMap.init(levelData, 
			  feSpaces,  meshDistributor->getFeSpaces(),
269
			  false, false);
270
271
      interiorDofMap.setComputeMatIndex(true);
    }
272
273
274
275
276
277
278
279
280
281
  }


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

    TEST_EXIT(meshDistributor)("No mesh distributor object defined!\n");
    TEST_EXIT(meshDistributor->getFeSpaces().size() > 0)
      ("No FE space defined in mesh distributor!\n");
282

283
284
285
286
287
288
    primalDofMap.clear();
    dualDofMap.clear();
    lagrangeMap.clear();
    localDofMap.clear();
    interiorDofMap.clear();

289
290
    for (unsigned int i = 0; i < meshDistributor->getFeSpaces().size(); i++) {
      const FiniteElemSpace *feSpace = meshDistributor->getFeSpace(i);
291

292
293
      createPrimals(feSpace);      
      createDuals(feSpace);
294
      createLagrange(feSpace);      
295
296
      createIndexB(feSpace);
    }
Thomas Witkowski's avatar
Thomas Witkowski committed
297

298
299
300
301
302
    primalDofMap.setDofComm(meshDistributor->getSendDofs(),
			    meshDistributor->getRecvDofs());
    lagrangeMap.setDofComm(meshDistributor->getSendDofs(), 
			   meshDistributor->getRecvDofs());

Thomas Witkowski's avatar
Thomas Witkowski committed
303
304
305
306
    primalDofMap.update();
    dualDofMap.update();
    lagrangeMap.update();
    localDofMap.update();
307
    if (fetiPreconditioner != FETI_NONE)
308
      interiorDofMap.update();
Thomas Witkowski's avatar
Thomas Witkowski committed
309
310
311

    for (unsigned int i = 0; i < meshDistributor->getFeSpaces().size(); i++) {
      const FiniteElemSpace *feSpace = meshDistributor->getFeSpace(i);
312
      MSG("FETI-DP data for %d-ith FE space:\n", i);
Thomas Witkowski's avatar
Thomas Witkowski committed
313

314
      MSG("  nRankPrimals = %d   nOverallPrimals = %d\n", 
315
	  primalDofMap[feSpace].nRankDofs[0], primalDofMap[feSpace].nOverallDofs[0]);
Thomas Witkowski's avatar
Thomas Witkowski committed
316
      
317
      MSG("  nRankDuals = %d  nOverallDuals = %d\n",
318
	  dualDofMap[feSpace].nRankDofs[0], dualDofMap[feSpace].nOverallDofs[0]);
Thomas Witkowski's avatar
Thomas Witkowski committed
319

320
      MSG("  nRankLagrange = %d  nOverallLagrange = %d\n",
321
	  lagrangeMap[feSpace].nRankDofs[0], lagrangeMap[feSpace].nOverallDofs[0]);
322
323
324
325

      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
326
    }
327
328
329


    // If multi level test, inform sub domain solver about coarse space.
330
    subDomainSolver->setDofMapping(&primalDofMap, &localDofMap);
331
332
333
  }


334
  void PetscSolverFeti::createPrimals(const FiniteElemSpace *feSpace)
Thomas Witkowski's avatar
Thomas Witkowski committed
335
  {
336
    FUNCNAME("PetscSolverFeti::createPrimals()");  
337

338
339
340
    // === Define all vertices on the interior boundaries of the macro mesh ===
    // === to be primal variables.                                          ===

341
342
343
    /// Set of DOF indices that are considered to be primal variables.
    
    DofIndexSet primals;
344
    DofContainerSet& vertices = 
345
      meshDistributor->getBoundaryDofInfo(feSpace, meshLevel).geoDofs[VERTEX];
346
347
348
    for (DofContainerSet::iterator it = vertices.begin(); 
	 it != vertices.end(); ++it)
      primals.insert(**it);
349

350
351
352
353

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

354
    for (DofIndexSet::iterator it = primals.begin(); it != primals.end(); ++it)
355
      if (meshDistributor->getDofMap()[feSpace].isRankDof(*it))
356
	primalDofMap[feSpace].insertRankDof(0, *it);
Thomas Witkowski's avatar
FETI-DP  
Thomas Witkowski committed
357
      else
358
  	primalDofMap[feSpace].insertNonRankDof(0, *it);
359
360
361
  }


362
  void PetscSolverFeti::createDuals(const FiniteElemSpace *feSpace)
363
364
  {
    FUNCNAME("PetscSolverFeti::createDuals()");
365

366
367
368
    // === Create global index of the dual nodes on each rank. ===

    DofContainer allBoundaryDofs;
369
    meshDistributor->getAllBoundaryDofs(feSpace, meshLevel, allBoundaryDofs);
370
371
372
373

    for (DofContainer::iterator it = allBoundaryDofs.begin();
	 it != allBoundaryDofs.end(); ++it)
      if (!isPrimal(feSpace, **it))
374
	dualDofMap[feSpace].insertRankDof(0, **it);
375
376
377
378
379
380
381
  }

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

382
383
    // === Create for each dual node that is owned by the rank, the set ===
    // === of ranks that contain this node (denoted by W(x_j)).         ===
384

385
    boundaryDofRanks[feSpace].clear();
386

387
    for (DofComm::Iterator it(meshDistributor->getSendDofs(), meshLevel, feSpace); 
388
389
	 !it.end(); it.nextRank())
      for (; !it.endDofIter(); it.nextDof()) {
390
	if (!isPrimal(feSpace, it.getDofIndex())) {
391
392
	  boundaryDofRanks[feSpace][it.getDofIndex()].insert(mpiRank);
	  boundaryDofRanks[feSpace][it.getDofIndex()].insert(it.getRank());
393
394
	}
      }
395
	
396
397
398
399

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

400
    StdMpi<vector<std::set<int> > > stdMpi(meshDistributor->getMpiComm());
401

402
    for (DofComm::Iterator it(meshDistributor->getSendDofs(), meshLevel, feSpace);
403
404
	 !it.end(); it.nextRank())
      for (; !it.endDofIter(); it.nextDof())
405
	if (!isPrimal(feSpace, it.getDofIndex()))
406
	  stdMpi.getSendData(it.getRank()).push_back(boundaryDofRanks[feSpace][it.getDofIndex()]);
407
408
409

    stdMpi.updateSendDataSize();

410
    for (DofComm::Iterator it(meshDistributor->getRecvDofs(), meshLevel, feSpace); 
411
	 !it.end(); it.nextRank()) {
412
      bool recvFromRank = false;
413
      for (; !it.endDofIter(); it.nextDof()) {
414
	if (!isPrimal(feSpace, it.getDofIndex())) {
415
416
417
	  recvFromRank = true;
	  break;
	}
418
      }
419
420

      if (recvFromRank)
421
	stdMpi.recv(it.getRank());
422
    }
423

424
425
    stdMpi.startCommunication();

426
    for (DofComm::Iterator it(meshDistributor->getRecvDofs(), meshLevel, feSpace); 
427
	 !it.end(); it.nextRank()) {
428
      int i = 0;
429
      for (; !it.endDofIter(); it.nextDof())
430
	if (!isPrimal(feSpace, it.getDofIndex()))
431
	  boundaryDofRanks[feSpace][it.getDofIndex()] = 
432
	    stdMpi.getRecvData(it.getRank())[i++];
433
434
    }

435
436
437
    // === Reserve for each dual node, on the rank that owns this node, the ===
    // === appropriate number of Lagrange constraints.                      ===

438
    int nRankLagrange = 0;
439
    DofMap& dualMap = dualDofMap[feSpace].getMap(0);
440
    for (DofMap::iterator it = dualMap.begin(); it != dualMap.end(); ++it) {
441
442

      if (meshDistributor->getDofMap()[feSpace].isRankDof(it->first)) {
443
	lagrangeMap[feSpace].insertRankDof(0, it->first, nRankLagrange);
444
	int degree = boundaryDofRanks[feSpace][it->first].size();
445
	nRankLagrange += (degree * (degree - 1)) / 2;
Thomas Witkowski's avatar
Thomas Witkowski committed
446
      } else {
447
	lagrangeMap[feSpace].insertNonRankDof(0, it->first);
448
449
      }
    }
450
    lagrangeMap[feSpace].nRankDofs[0] = nRankLagrange;
451
452
453
  }


454
  void PetscSolverFeti::createIndexB(const FiniteElemSpace *feSpace)
455
  {
456
    FUNCNAME("PetscSolverFeti::createIndexB()");
457

458
    DOFAdmin* admin = feSpace->getAdmin();
459
460
461
462

    // === 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.                                 ===
463

464
    int nLocalInterior = 0;    
465
    for (int i = 0; i < admin->getUsedSize(); i++) {
466
      if (admin->isDofFree(i) == false && 
467
	  isPrimal(feSpace, i) == false &&
468
	  dualDofMap[feSpace].isSet(i) == false) {
469
	localDofMap[feSpace].insertRankDof(0, i, nLocalInterior);
470

471
	if (fetiPreconditioner != FETI_NONE)
472
	  interiorDofMap[feSpace].insertRankDof(0, i, nLocalInterior);
473
474

	nLocalInterior++;
475
      }
476
    }
Thomas Witkowski's avatar
FETI-DP  
Thomas Witkowski committed
477
    
478
479
    // === And finally, add the global indicies of all dual nodes. ===

480
481
    for (DofMap::iterator it = dualDofMap[feSpace].getMap(0).begin();
	 it != dualDofMap[feSpace].getMap(0).end(); ++it)
482
      localDofMap[feSpace].insertRankDof(0, it->first);
483
484
485
  }


486
  void PetscSolverFeti::createMatLagrange(vector<const FiniteElemSpace*> &feSpaces)
487
488
489
  {
    FUNCNAME("PetscSolverFeti::createMatLagrange()");

490
491
    // === Create distributed matrix for Lagrange constraints. ===

492
    MatCreateMPIAIJ(mpiComm,
493
494
		    lagrangeMap.getRankDofs(0), localDofMap.getRankDofs(0),
		    lagrangeMap.getOverallDofs(0), localDofMap.getOverallDofs(0),
495
496
497
		    2, PETSC_NULL, 2, PETSC_NULL,
		    &mat_lagrange);

498
499
500
501
502
503
504
    // === 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
505
    for (unsigned int k = 0; k < feSpaces.size(); k++) {
506
      DofMap &dualMap = dualDofMap[feSpaces[k]].getMap(0);
507

508
      for (DofMap::iterator it = dualMap.begin(); it != dualMap.end(); ++it) {
509
	TEST_EXIT_DBG(boundaryDofRanks[feSpaces[k]].count(it->first))
Thomas Witkowski's avatar
Thomas Witkowski committed
510
511
512
513
514
515
	  ("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.
516
517
	vector<int> W(boundaryDofRanks[feSpaces[k]][it->first].begin(), 
		      boundaryDofRanks[feSpaces[k]][it->first].end());
Thomas Witkowski's avatar
Thomas Witkowski committed
518
519
520
521
522
523
	// 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) {
524
	      int colIndex = localDofMap.getMatIndex(k, it->first);
525
	      double value = (W[i] == mpiRank ? 1.0 : -1.0);
526
	      MatSetValue(mat_lagrange, index, colIndex, value, INSERT_VALUES);
527
	    }
Thomas Witkowski's avatar
Thomas Witkowski committed
528
	    index++;	      
529
530
531
532
533
534
535
536
537
538
	  }
	}
      }
    }

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


539
  void PetscSolverFeti::createSchurPrimalKsp(vector<const FiniteElemSpace*> &feSpaces)
540
541
542
  {
    FUNCNAME("PetscSolverFeti::createSchurPrimal()");

Thomas Witkowski's avatar
Thomas Witkowski committed
543
    if (schurPrimalSolver == 0) {
544
545
      MSG("Create iterative schur primal solver!\n");

546
      schurPrimalData.subSolver = subDomainSolver;
Thomas Witkowski's avatar
Thomas Witkowski committed
547
      
548
      VecCreateMPI(mpiComm, 
549
		   localDofMap.getRankDofs(0), localDofMap.getOverallDofs(0),
Thomas Witkowski's avatar
Thomas Witkowski committed
550
		   &(schurPrimalData.tmp_vec_b));
551
      VecCreateMPI(mpiComm, 
552
		   primalDofMap.getRankDofs(0), primalDofMap.getOverallDofs(0),
Thomas Witkowski's avatar
Thomas Witkowski committed
553
554
		   &(schurPrimalData.tmp_vec_primal));

555
      MatCreateShell(mpiComm,
556
557
		     primalDofMap.getRankDofs(0), primalDofMap.getRankDofs(0), 
		     primalDofMap.getOverallDofs(0), primalDofMap.getOverallDofs(0),
Thomas Witkowski's avatar
Thomas Witkowski committed
558
559
560
561
562
		     &schurPrimalData, 
		     &mat_schur_primal);
      MatShellSetOperation(mat_schur_primal, MATOP_MULT, 
			   (void(*)(void))petscMultMatSchurPrimal);
      
563
      KSPCreate(mpiComm, &ksp_schur_primal);
Thomas Witkowski's avatar
Thomas Witkowski committed
564
      KSPSetOperators(ksp_schur_primal, mat_schur_primal, mat_schur_primal, SAME_NONZERO_PATTERN);
565
566
      KSPSetOptionsPrefix(ksp_schur_primal, "schur_primal_");
      KSPSetType(ksp_schur_primal, KSPGMRES);
Thomas Witkowski's avatar
Thomas Witkowski committed
567
568
      KSPSetFromOptions(ksp_schur_primal);
    } else {
569
570
      MSG("Create direct schur primal solver!\n");

571
572
      double wtime = MPI::Wtime();

573
574
575
576
      int nRowsRankPrimal = primalDofMap.getRankDofs(0);
      int nRowsOverallPrimal = primalDofMap.getOverallDofs(0);
      int nRowsRankB = localDofMap.getRankDofs(0);
      int nRowsOverallB = localDofMap.getOverallDofs(0);
577

Thomas Witkowski's avatar
Thomas Witkowski committed
578
      Mat matBPi;
579
      MatCreateMPIAIJ(mpiComm,
Thomas Witkowski's avatar
Thomas Witkowski committed
580
581
582
583
		      nRowsRankB, nRowsRankPrimal, 
		      nRowsOverallB, nRowsOverallPrimal,
		      30, PETSC_NULL, 30, PETSC_NULL, &matBPi);
      Mat matPrimal;
584

585
586
587
      PetscInt nCols;
      const PetscInt *cols;
      const PetscScalar *values;
Thomas Witkowski's avatar
Thomas Witkowski committed
588

589
590
      map<int, vector<pair<int, double> > > mat_b_primal_cols;

591
      for (int i = 0; i < nRowsRankB; i++) {
592
	PetscInt row = localDofMap.getStartDofs(0) + i;
593
	MatGetRow(subDomainSolver->getMatIntCoarse(), row, &nCols, &cols, &values);
594
595
596
597
598

	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]));

599
	MatRestoreRow(subDomainSolver->getMatIntCoarse(), row, &nCols, &cols, &values);
600
601
      }

602
      TEST_EXIT(static_cast<int>(mat_b_primal_cols.size()) == 
603
		primalDofMap.getLocalDofs(0))
604
	("Should not happen!\n");
605

606
607
608
      for (map<int, vector<pair<int, double> > >::iterator it = mat_b_primal_cols.begin();
	   it != mat_b_primal_cols.end(); ++it) {
	Vec tmpVec;
609
	VecCreateSeq(PETSC_COMM_SELF, nRowsRankB, &tmpVec);
610
611
612
613
614
615
616
617

 	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);

618
	subDomainSolver->solve(tmpVec, tmpVec);
619

Thomas Witkowski's avatar
Thomas Witkowski committed
620
	PetscScalar *tmpValues;
621
	VecGetArray(tmpVec, &tmpValues);
622
	for (int i  = 0; i < nRowsRankB; i++)
Thomas Witkowski's avatar
Thomas Witkowski committed
623
	  MatSetValue(matBPi, 
624
		      localDofMap.getStartDofs(0) + i,
Thomas Witkowski's avatar
Thomas Witkowski committed
625
626
627
		      it->first,
		      tmpValues[i],
		      ADD_VALUES);
628
629
630
631
632
	VecRestoreArray(tmpVec, &tmpValues);

	VecDestroy(&tmpVec);
      }

Thomas Witkowski's avatar
Thomas Witkowski committed
633
634
      MatAssemblyBegin(matBPi, MAT_FINAL_ASSEMBLY);
      MatAssemblyEnd(matBPi, MAT_FINAL_ASSEMBLY);
635
636

      MatDuplicate(subDomainSolver->getMatCoarseCoarse(), MAT_COPY_VALUES, &mat_schur_primal);
637
      MatMatMult(subDomainSolver->getMatCoarseInt(), matBPi, MAT_INITIAL_MATRIX, PETSC_DEFAULT, &matPrimal);
638
      MatAXPY(mat_schur_primal, -1.0, matPrimal, DIFFERENT_NONZERO_PATTERN);
639
640

      MatDestroy(&matPrimal);
Thomas Witkowski's avatar
Thomas Witkowski committed
641
      MatDestroy(&matBPi);
642
643

      MatInfo minfo;
644
      MatGetInfo(mat_schur_primal, MAT_GLOBAL_SUM, &minfo);
645
646
      MSG("Schur primal matrix nnz = %f\n", minfo.nz_used);

647
      KSPCreate(mpiComm, &ksp_schur_primal);
648
      KSPSetOperators(ksp_schur_primal, mat_schur_primal, mat_schur_primal,
649
		      SAME_NONZERO_PATTERN);
650
651
652
653
654
655
      KSPSetOptionsPrefix(ksp_schur_primal, "schur_primal_");
      KSPSetType(ksp_schur_primal, KSPPREONLY);
      PC pc_schur_primal;      
      KSPGetPC(ksp_schur_primal, &pc_schur_primal);
      PCSetType(pc_schur_primal, PCLU);
      PCFactorSetMatSolverPackage(pc_schur_primal, MATSOLVERMUMPS);
656
      KSPSetFromOptions(ksp_schur_primal);
Thomas Witkowski's avatar
Thomas Witkowski committed
657

658
659
      MSG("Creating Schur primal matrix needed %.5f seconds.\n",
	  MPI::Wtime() - wtime);
Thomas Witkowski's avatar
Thomas Witkowski committed
660
    }
661
662
663
664
665
666
667
  }


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

668
    if (schurPrimalSolver == 0) {
669
      schurPrimalData.subSolver = NULL;
670

671
672
673
      VecDestroy(&schurPrimalData.tmp_vec_b);
      VecDestroy(&schurPrimalData.tmp_vec_primal);
    }
674
675
676
    
    MatDestroy(&mat_schur_primal);
    KSPDestroy(&ksp_schur_primal);
677
678
679
  }


680
  void PetscSolverFeti::createFetiKsp(vector<const FiniteElemSpace*> &feSpaces)
681
682
683
  {
    FUNCNAME("PetscSolverFeti::createFetiKsp()");

684
685
    // === Create FETI-DP solver object. ===

686
    fetiData.mat_lagrange = &mat_lagrange;
687
    fetiData.subSolver = subDomainSolver;
688
    fetiData.ksp_schur_primal = &ksp_schur_primal;
689

690
    VecCreateMPI(mpiComm, 
691
		 localDofMap.getRankDofs(0), localDofMap.getOverallDofs(0),
Thomas Witkowski's avatar
Thomas Witkowski committed
692
		 &(fetiData.tmp_vec_b));
693
    VecCreateMPI(mpiComm,
694
		 lagrangeMap.getRankDofs(0), lagrangeMap.getOverallDofs(0),
Thomas Witkowski's avatar
Thomas Witkowski committed
695
		 &(fetiData.tmp_vec_lagrange));
696
    VecCreateMPI(mpiComm, 
697
		 primalDofMap.getRankDofs(0), primalDofMap.getOverallDofs(0),
698
		 &(fetiData.tmp_vec_primal));
699

700
    MatCreateShell(mpiComm,
701
702
		   lagrangeMap.getRankDofs(0), lagrangeMap.getRankDofs(0),
		   lagrangeMap.getOverallDofs(0), lagrangeMap.getOverallDofs(0),
703
		   &fetiData, &mat_feti);
704
705
706
    MatShellSetOperation(mat_feti, MATOP_MULT, (void(*)(void))petscMultMatFeti);


707
    KSPCreate(mpiComm, &ksp_feti);
708
    KSPSetOperators(ksp_feti, mat_feti, mat_feti, SAME_NONZERO_PATTERN);
709
710
711
    KSPSetOptionsPrefix(ksp_feti, "feti_");
    KSPSetType(ksp_feti, KSPGMRES);
    KSPSetTolerances(ksp_feti, 0, 1e-8, 1e+3, 1000);
712
    KSPSetFromOptions(ksp_feti);
713
714


715
    // === Create FETI-DP preconditioner object. ===
716

717
718
719
720
    if (fetiPreconditioner != FETI_NONE) {
      MatDuplicate(mat_lagrange, MAT_COPY_VALUES, &mat_lagrange_scaled);
      MatScale(mat_lagrange_scaled, 0.5);
    }
721

722
723
724
    switch (fetiPreconditioner) {
    case FETI_DIRICHLET:           
      KSPCreate(PETSC_COMM_SELF, &ksp_interior);
Thomas Witkowski's avatar
Thomas Witkowski committed
725
726
      KSPSetOperators(ksp_interior, mat_interior_interior, mat_interior_interior, 
		      SAME_NONZERO_PATTERN);
727
728
729
730
731
732
      KSPSetOptionsPrefix(ksp_interior, "precon_interior_");
      KSPSetType(ksp_interior, KSPPREONLY);
      PC pc_interior;
      KSPGetPC(ksp_interior, &pc_interior);
      PCSetType(pc_interior, PCLU);
      PCFactorSetMatSolverPackage(pc_interior, MATSOLVERUMFPACK);
733
734
735
736
737
738
739
740
741
      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;
      
742
      VecCreateMPI(mpiComm, 
743
		   localDofMap.getRankDofs(0),localDofMap.getOverallDofs(0),
744
		   &(fetiDirichletPreconData.tmp_vec_b));      
745
746
747
748
749
750
      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));
751
752

      for (unsigned int i = 0; i < feSpaces.size(); i++) {
753
	DofMap &dualMap = dualDofMap[feSpaces[i]].getMap(0);
754
	for (DofMap::iterator it = dualMap.begin(); it != dualMap.end(); ++it) {
755
756
757
758
759
760
761
	  DegreeOfFreedom d = it->first;
	  int matIndexLocal = localDofMap.getLocalMatIndex(i, d);
	  int matIndexDual = dualDofMap.getLocalMatIndex(i, d);
	  fetiDirichletPreconData.localToDualMap[matIndexLocal] = matIndexDual;
	}
      }

762
763
764
765
766
767
768
769
770
771
772
      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;

Thomas Witkowski's avatar
Thomas Witkowski committed
773
      for (unsigned int i = 0; i < feSpaces.size(); i++) {
774
	DofMap &dualMap = dualDofMap[feSpaces[i]].getMap(0);
775
	for (DofMap::iterator it = dualMap.begin(); it != dualMap.end(); ++it) {
Thomas Witkowski's avatar
Thomas Witkowski committed
776
777
778
779
780
781
782
	  DegreeOfFreedom d = it->first;
	  int matIndexLocal = localDofMap.getLocalMatIndex(i, d);
	  int matIndexDual = dualDofMap.getLocalMatIndex(i, d);
	  fetiLumpedPreconData.localToDualMap[matIndexLocal] = matIndexDual;
	}
      }

783
      VecCreateMPI(mpiComm, 
784
785
		   localDofMap.getRankDofs(0),
		   localDofMap.getOverallDofs(0),
786
		   &(fetiLumpedPreconData.tmp_vec_b));
787
788
789
790
      MatGetVecs(mat_duals_duals, PETSC_NULL, 
		 &(fetiLumpedPreconData.tmp_vec_duals0));
      MatGetVecs(mat_duals_duals, PETSC_NULL, 
		 &(fetiLumpedPreconData.tmp_vec_duals1));
791
792
793
794
795
796

      KSPGetPC(ksp_feti, &precon_feti);
      PCSetType(precon_feti, PCSHELL);
      PCShellSetContext(precon_feti, static_cast<void*>(&fetiLumpedPreconData));
      PCShellSetApply(precon_feti, petscApplyFetiLumpedPrecon);
      
797
798
      break;
    default:
799
800
      break;
    }
801
802
803
804
805
806
807
  }
  

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

808
809
    // === Destroy FETI-DP solver object. ===

810
    fetiData.mat_lagrange = PETSC_NULL;
811
    fetiData.subSolver = NULL;
812
    fetiData.ksp_schur_primal = PETSC_NULL;
813

Thomas Witkowski's avatar
Thomas Witkowski committed
814
815
    VecDestroy(&fetiData.tmp_vec_b);
    VecDestroy(&fetiData.tmp_vec_lagrange);
816
    VecDestroy(&fetiData.tmp_vec_primal);
817
818
    MatDestroy(&mat_feti);
    KSPDestroy(&ksp_feti);
819
820


821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
    // === 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;