PetscSolverFeti.cc 44.1 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
      subDomainSolver(NULL),
193
194
195
      meshLevel(0),
      rStartInterior(0),
      nGlobalOverallInterior(0)
196
197
198
199
200
201
  {
    FUNCNAME("PetscSolverFeti::PetscSolverFeti()");

    string preconditionerName = "";
    Parameters::get("parallel->solver->precon", preconditionerName);
    if (preconditionerName == "" || preconditionerName == "none") {
202
      MSG("Create FETI-DP solver with no preconditioner!\n");
203
204
      fetiPreconditioner = FETI_NONE;
    } else if (preconditionerName == "dirichlet") {
205
      MSG("Create FETI-DP solver with Dirichlet preconditioner!\n");
206
207
      fetiPreconditioner = FETI_DIRICHLET;
    } else if (preconditionerName == "lumped") {
208
      MSG("Create FETI-DP solver with lumped preconditioner!\n");
209
210
      fetiPreconditioner = FETI_LUMPED;
    } else {
Thomas Witkowski's avatar
Thomas Witkowski committed
211
212
      ERROR_EXIT("Preconditioner \"%s\" not available!\n", 
		 preconditionerName.c_str());
213
    }
Thomas Witkowski's avatar
Thomas Witkowski committed
214
215
216
217
218

    Parameters::get("parallel->feti->schur primal solver", schurPrimalSolver);
    TEST_EXIT(schurPrimalSolver == 0 || schurPrimalSolver == 1)
      ("Wrong solver \"%d\"for the Schur primal complement!\n", 
       schurPrimalSolver);
219
220

    Parameters::get("parallel->multi level test", multiLevelTest);
221
222
    if (multiLevelTest)
      meshLevel = 1;
223
224
225
  }


226
  void PetscSolverFeti::initialize(vector<const FiniteElemSpace*> feSpaces)
227
  {
228
229
    FUNCNAME("PetscSolverFeti::initialize()");

230
231
232
    TEST_EXIT_DBG(meshLevel + 1 == meshDistributor->getMeshLevelData().getLevelNumber())
      ("Mesh hierarchy does not contain %d levels!\n", meshLevel + 1);

233
234
    MeshLevelData& levelData = meshDistributor->getMeshLevelData();

235
236
237
238
239
240
241
242
243
    if (subDomainSolver == NULL) {
      if (meshLevel == 0) {
	subDomainSolver = 
	  new SubDomainSolver(meshDistributor, mpiComm, mpiSelfComm);
      } else {
	subDomainSolver = 
	  new SubDomainSolver(meshDistributor, 
			      levelData.getMpiComm(meshLevel - 1),
			      levelData.getMpiComm(meshLevel));
244
	subDomainSolver->setLevel(meshLevel);
245
246
      }
    }
247

248
249
    primalDofMap.init(levelData,
		      feSpaces, meshDistributor->getFeSpaces(), 
250
		      true, true);
251
252
    primalDofMap.setComputeMatIndex(true);
    
253
254
    dualDofMap.init(levelData, 
		    feSpaces, meshDistributor->getFeSpaces(),
255
		    false, false);
256
257
    dualDofMap.setComputeMatIndex(true);

258
259
    lagrangeMap.init(levelData, 
		     feSpaces, meshDistributor->getFeSpaces(),
260
		     true, true);
261
    lagrangeMap.setComputeMatIndex(true);
262
263
264
265
266
267
268
269
270
271
272
273

    if (meshLevel == 0) {
      localDofMap.init(levelData, 
		       feSpaces, meshDistributor->getFeSpaces(),
		       false, false);
      localDofMap.setComputeMatIndex(true);
    } else {
      localDofMap.init(levelData, 
		       feSpaces, meshDistributor->getFeSpaces(),
		       true, true);
      localDofMap.setComputeMatIndex(true);
    }
274
275

    if (fetiPreconditioner == FETI_DIRICHLET) {
276
277
278
      TEST_EXIT(meshLevel == 0)
	("Dirichlet preconditioner not yet implemented for multilevel FETI-DP\n");

279
280
      interiorDofMap.init(levelData, 
			  feSpaces,  meshDistributor->getFeSpaces(),
281
			  false, false);
282
283
      interiorDofMap.setComputeMatIndex(true);
    }
284
285
286
287
288
289
290
291
292
293
  }


  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");
294

295
296
    MeshLevelData& levelData = meshDistributor->getMeshLevelData();

297
298
299
300
    primalDofMap.clear();
    dualDofMap.clear();
    lagrangeMap.clear();
    localDofMap.clear();
301
302
303
    if (fetiPreconditioner == FETI_DIRICHLET)
      interiorDofMap.clear();

304
305
    primalDofMap.setDofComm(meshDistributor->getDofComm());
    lagrangeMap.setDofComm(meshDistributor->getDofComm());
306

307
308
309
310
    primalDofMap.setMpiComm(levelData.getMpiComm(0), 0);
    dualDofMap.setMpiComm(levelData.getMpiComm(0), 0);
    lagrangeMap.setMpiComm(levelData.getMpiComm(0), 0);    
    localDofMap.setMpiComm(levelData.getMpiComm(meshLevel), meshLevel);
311
312
    if (fetiPreconditioner != FETI_NONE)
      interiorDofMap.setMpiComm(levelData.getMpiComm(meshLevel), meshLevel);
313
314
315
316

    if (meshLevel > 0)
      localDofMap.setDofComm(meshDistributor->getDofCommSd());

317
318
    for (unsigned int i = 0; i < meshDistributor->getFeSpaces().size(); i++) {
      const FiniteElemSpace *feSpace = meshDistributor->getFeSpace(i);
319

320
321
      createPrimals(feSpace);  

322
      createDuals(feSpace);      
323

324
325
      createIndexB(feSpace);
    }
Thomas Witkowski's avatar
Thomas Witkowski committed
326
327
328
329

    primalDofMap.update();
    dualDofMap.update();
    localDofMap.update();
330
    if (fetiPreconditioner != FETI_NONE)
331
      interiorDofMap.update();
Thomas Witkowski's avatar
Thomas Witkowski committed
332

333
334
335
336
337
338
339
    for (unsigned int i = 0; i < meshDistributor->getFeSpaces().size(); i++) {
      const FiniteElemSpace *feSpace = meshDistributor->getFeSpace(i);
      createLagrange(feSpace);
    }
    lagrangeMap.update();


340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372

    // === ===

    if (meshLevel == 0) {
      rStartInterior = 0;
      nGlobalOverallInterior = localDofMap.getOverallDofs();
    } else {
      MeshLevelData& levelData = meshDistributor->getMeshLevelData();

      MSG("RANK %d FROM %d\n",
	  levelData.getMpiComm(1).Get_rank(),
	  levelData.getMpiComm(1).Get_size());

      int groupRowsInterior = 0;
      if (levelData.getMpiComm(1).Get_rank() == 0)
	groupRowsInterior = localDofMap.getOverallDofs();

      mpi::getDofNumbering(mpiComm, groupRowsInterior,
			   rStartInterior, nGlobalOverallInterior);

      int tmp = 0;
      if (levelData.getMpiComm(1).Get_rank() == 0)
	tmp = rStartInterior;

      levelData.getMpiComm(1).Allreduce(&tmp, &rStartInterior, 1, MPI_INT, MPI_SUM);

      MSG("COMM TEST FETI-DP: %d %d %d %d %d\n",
	  levelData.getMpiComm(1).Get_size(),
	  localDofMap.getRankDofs(),
	  localDofMap.getOverallDofs(),
	  nGlobalOverallInterior, rStartInterior);
    }

373
    MSG("FETI-DP data created on mesh level %d\n", meshLevel);
Thomas Witkowski's avatar
Thomas Witkowski committed
374
375
    for (unsigned int i = 0; i < meshDistributor->getFeSpaces().size(); i++) {
      const FiniteElemSpace *feSpace = meshDistributor->getFeSpace(i);
376
      MSG("FETI-DP data for %d-ith FE space:\n", i);
Thomas Witkowski's avatar
Thomas Witkowski committed
377

378
      MSG("  nRankPrimals = %d   nOverallPrimals = %d\n", 
379
380
	  primalDofMap[feSpace].nRankDofs, 
	  primalDofMap[feSpace].nOverallDofs);
Thomas Witkowski's avatar
Thomas Witkowski committed
381
      
382
      MSG("  nRankDuals = %d  nOverallDuals = %d\n",
383
384
	  dualDofMap[feSpace].nRankDofs, 
	  dualDofMap[feSpace].nOverallDofs);
Thomas Witkowski's avatar
Thomas Witkowski committed
385

386
      MSG("  nRankLagrange = %d  nOverallLagrange = %d\n",
387
388
	  lagrangeMap[feSpace].nRankDofs, 
	  lagrangeMap[feSpace].nOverallDofs);
389

390
      TEST_EXIT_DBG(localDofMap[feSpace].size() + primalDofMap[feSpace].size() == 
391
392
		    static_cast<unsigned int>(feSpace->getAdmin()->getUsedDofs()))
	("Should not happen!\n");
Thomas Witkowski's avatar
Thomas Witkowski committed
393
    }
394
395
396


    // If multi level test, inform sub domain solver about coarse space.
397
    subDomainSolver->setDofMapping(&primalDofMap, &localDofMap);
398
399
400
  }


401
  void PetscSolverFeti::createPrimals(const FiniteElemSpace *feSpace)
Thomas Witkowski's avatar
Thomas Witkowski committed
402
  {
403
    FUNCNAME("PetscSolverFeti::createPrimals()");  
404

405
406
407
    // === Define all vertices on the interior boundaries of the macro mesh ===
    // === to be primal variables.                                          ===

408
409
    /// Set of DOF indices that are considered to be primal variables.
    
410
    DofContainerSet& vertices = 
411
      meshDistributor->getBoundaryDofInfo(feSpace, meshLevel).geoDofs[VERTEX];
412
413

    DofIndexSet primals;
414
415
416
    for (DofContainerSet::iterator it = vertices.begin(); 
	 it != vertices.end(); ++it)
      primals.insert(**it);
417

418
419
420
421

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

422
    for (DofIndexSet::iterator it = primals.begin(); it != primals.end(); ++it)
423
424
      if (meshDistributor->getDofMap()[feSpace].isRankDof(*it))
	primalDofMap[feSpace].insertRankDof(*it);
Thomas Witkowski's avatar
FETI-DP    
Thomas Witkowski committed
425
      else
426
  	primalDofMap[feSpace].insertNonRankDof(*it);
427
428
429
  }


430
  void PetscSolverFeti::createDuals(const FiniteElemSpace *feSpace)
431
432
  {
    FUNCNAME("PetscSolverFeti::createDuals()");
433

434
435
436
    // === Create global index of the dual nodes on each rank. ===

    DofContainer allBoundaryDofs;
437
    meshDistributor->getAllBoundaryDofs(feSpace, meshLevel, allBoundaryDofs);
438
439
440
441

    for (DofContainer::iterator it = allBoundaryDofs.begin();
	 it != allBoundaryDofs.end(); ++it)
      if (!isPrimal(feSpace, **it))
442
	dualDofMap[feSpace].insertRankDof(**it);
443
444
445
446
447
448
449
  }

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

450
451
    boundaryDofRanks[feSpace].clear();

452
    if (dualDofMap[feSpace].nLocalDofs == 0)
453
454
      return;

455
456
    // === Create for each dual node that is owned by the rank, the set ===
    // === of ranks that contain this node (denoted by W(x_j)).         ===
457

458
    for (DofComm::Iterator it(meshDistributor->getDofComm().getSendDofs(), meshLevel, feSpace); 
459
460
	 !it.end(); it.nextRank())
      for (; !it.endDofIter(); it.nextDof()) {
461
	if (!isPrimal(feSpace, it.getDofIndex())) {
462
463
	  boundaryDofRanks[feSpace][it.getDofIndex()].insert(mpiRank);
	  boundaryDofRanks[feSpace][it.getDofIndex()].insert(it.getRank());
464
465
	}
      }
466
	
467
468
469
470

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

471
    StdMpi<vector<std::set<int> > > stdMpi(meshDistributor->getMpiComm());
472

473
    for (DofComm::Iterator it(meshDistributor->getDofComm().getSendDofs(), meshLevel, feSpace);
474
475
	 !it.end(); it.nextRank())
      for (; !it.endDofIter(); it.nextDof())
476
	if (!isPrimal(feSpace, it.getDofIndex()))
477
	  stdMpi.getSendData(it.getRank()).push_back(boundaryDofRanks[feSpace][it.getDofIndex()]);
478
479
480

    stdMpi.updateSendDataSize();

481
    for (DofComm::Iterator it(meshDistributor->getDofComm().getRecvDofs(), meshLevel, feSpace); 
482
	 !it.end(); it.nextRank()) {
483
      bool recvFromRank = false;
484
      for (; !it.endDofIter(); it.nextDof()) {
485
	if (!isPrimal(feSpace, it.getDofIndex())) {
486
487
488
	  recvFromRank = true;
	  break;
	}
489
      }
490
491

      if (recvFromRank)
492
	stdMpi.recv(it.getRank());
493
    }
494

495
496
    stdMpi.startCommunication();

497
    for (DofComm::Iterator it(meshDistributor->getDofComm().getRecvDofs(), meshLevel, feSpace); 
498
	 !it.end(); it.nextRank()) {
499
      int i = 0;
500
      for (; !it.endDofIter(); it.nextDof())
501
	if (!isPrimal(feSpace, it.getDofIndex()))
502
	  boundaryDofRanks[feSpace][it.getDofIndex()] = 
503
	    stdMpi.getRecvData(it.getRank())[i++];
504
505
    }

506
507
508
    // === Reserve for each dual node, on the rank that owns this node, the ===
    // === appropriate number of Lagrange constraints.                      ===

509
    int nRankLagrange = 0;
510
    DofMap& dualMap = dualDofMap[feSpace].getMap();
511
    for (DofMap::iterator it = dualMap.begin(); it != dualMap.end(); ++it) {
512

513
514
      if (meshDistributor->getDofMap()[feSpace].isRankDof(it->first)) {
	lagrangeMap[feSpace].insertRankDof(it->first, nRankLagrange);
515
	int degree = boundaryDofRanks[feSpace][it->first].size();
516
	nRankLagrange += (degree * (degree - 1)) / 2;
Thomas Witkowski's avatar
Thomas Witkowski committed
517
      } else {
518
	lagrangeMap[feSpace].insertNonRankDof(it->first);
519
520
      }
    }
521
    lagrangeMap[feSpace].nRankDofs = nRankLagrange;
522
523
524
  }


525
  void PetscSolverFeti::createIndexB(const FiniteElemSpace *feSpace)
526
  {
527
    FUNCNAME("PetscSolverFeti::createIndexB()");
528

529
    DOFAdmin* admin = feSpace->getAdmin();
530
531
532
533

    // === 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.                                 ===
534

535
    int nLocalInterior = 0;    
536
    for (int i = 0; i < admin->getUsedSize(); i++) {
537
      if (admin->isDofFree(i) == false && 
538
	  isPrimal(feSpace, i) == false &&
539
	  isDual(feSpace, i) == false) {
540
541
	if (meshLevel == 0) {
	  localDofMap[feSpace].insertRankDof(i, nLocalInterior);
542

543
544
545
546
547
548
549
550
551
	  if (fetiPreconditioner != FETI_NONE)
	    interiorDofMap[feSpace].insertRankDof(i, nLocalInterior);
	  
	  nLocalInterior++;	  
	} else {
	  if (meshDistributor->getDofMapSd()[feSpace].isRankDof(i))
	    localDofMap[feSpace].insertRankDof(i);
	  else
	    localDofMap[feSpace].insertNonRankDof(i);
552
553
554

	  TEST_EXIT_DBG(fetiPreconditioner == FETI_NONE)
	    ("Not yet implemnted!\n");
555
	}
556
      }
557
    }
Thomas Witkowski's avatar
FETI-DP    
Thomas Witkowski committed
558
    
559
560
    // === And finally, add the global indicies of all dual nodes. ===

561
562
563
564
565
566
567
568
569
570
    for (DofMap::iterator it = dualDofMap[feSpace].getMap().begin();
	 it != dualDofMap[feSpace].getMap().end(); ++it)
      if (meshLevel == 0)
	localDofMap[feSpace].insertRankDof(it->first);
      else {
	if (meshDistributor->getDofMapSd()[feSpace].isRankDof(it->first))
	  localDofMap[feSpace].insertRankDof(it->first);
	else 
	  localDofMap[feSpace].insertNonRankDof(it->first);
      }
571
572
573
  }


574
  void PetscSolverFeti::createMatLagrange(vector<const FiniteElemSpace*> &feSpaces)
575
576
577
  {
    FUNCNAME("PetscSolverFeti::createMatLagrange()");

578
579
    // === Create distributed matrix for Lagrange constraints. ===

580
    MatCreateMPIAIJ(mpiComm,
581
582
583
		    lagrangeMap.getRankDofs(), 
		    localDofMap.getRankDofs(),
		    lagrangeMap.getOverallDofs(), 
584
		    nGlobalOverallInterior,
585
586
587
		    2, PETSC_NULL, 2, PETSC_NULL,
		    &mat_lagrange);

588
589
590
591
592
593
594
    // === 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
595
    for (unsigned int k = 0; k < feSpaces.size(); k++) {
596
      DofMap &dualMap = dualDofMap[feSpaces[k]].getMap();
597

598
      for (DofMap::iterator it = dualMap.begin(); it != dualMap.end(); ++it) {
599
	TEST_EXIT_DBG(boundaryDofRanks[feSpaces[k]].count(it->first))
Thomas Witkowski's avatar
Thomas Witkowski committed
600
601
602
	  ("Should not happen!\n");
	
	// Global index of the first Lagrange constriant for this node.
603
	int index = lagrangeMap.getMatIndex(k, it->first);
Thomas Witkowski's avatar
Thomas Witkowski committed
604
605
	
	// Copy set of all ranks that contain this dual node.
606
607
	vector<int> W(boundaryDofRanks[feSpaces[k]][it->first].begin(), 
		      boundaryDofRanks[feSpaces[k]][it->first].end());
Thomas Witkowski's avatar
Thomas Witkowski committed
608
609
610
611
612
613
	// 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) {
614
	      int colIndex = localDofMap.getMatIndex(k, it->first);
615
616
617
618

	      if (meshLevel > 0)
		colIndex += rStartInterior;

619
	      double value = (W[i] == mpiRank ? 1.0 : -1.0);
620
	      MatSetValue(mat_lagrange, index, colIndex, value, INSERT_VALUES);
621
	    }
Thomas Witkowski's avatar
Thomas Witkowski committed
622
	    index++;	      
623
624
625
626
627
628
629
630
631
632
	  }
	}
      }
    }

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


633
  void PetscSolverFeti::createSchurPrimalKsp(vector<const FiniteElemSpace*> &feSpaces)
634
  {
635
    FUNCNAME("PetscSolverFeti::createSchurPrimalKsp()");
636

Thomas Witkowski's avatar
Thomas Witkowski committed
637
    if (schurPrimalSolver == 0) {
638
639
      MSG("Create iterative schur primal solver!\n");

640
      schurPrimalData.subSolver = subDomainSolver;
Thomas Witkowski's avatar
Thomas Witkowski committed
641
      
642
      VecCreateMPI(mpiComm, 
643
		   localDofMap.getRankDofs(), 
644
		   nGlobalOverallInterior,
Thomas Witkowski's avatar
Thomas Witkowski committed
645
		   &(schurPrimalData.tmp_vec_b));
646
      VecCreateMPI(mpiComm, 
647
648
		   primalDofMap.getRankDofs(), 
		   primalDofMap.getOverallDofs(),
Thomas Witkowski's avatar
Thomas Witkowski committed
649
650
		   &(schurPrimalData.tmp_vec_primal));

651
      MatCreateShell(mpiComm,
652
653
654
655
		     primalDofMap.getRankDofs(), 
		     primalDofMap.getRankDofs(), 
		     primalDofMap.getOverallDofs(), 
		     primalDofMap.getOverallDofs(),
Thomas Witkowski's avatar
Thomas Witkowski committed
656
657
658
659
660
		     &schurPrimalData, 
		     &mat_schur_primal);
      MatShellSetOperation(mat_schur_primal, MATOP_MULT, 
			   (void(*)(void))petscMultMatSchurPrimal);
      
661
      KSPCreate(mpiComm, &ksp_schur_primal);
Thomas Witkowski's avatar
Thomas Witkowski committed
662
      KSPSetOperators(ksp_schur_primal, mat_schur_primal, mat_schur_primal, SAME_NONZERO_PATTERN);
663
664
      KSPSetOptionsPrefix(ksp_schur_primal, "schur_primal_");
      KSPSetType(ksp_schur_primal, KSPGMRES);
Thomas Witkowski's avatar
Thomas Witkowski committed
665
666
      KSPSetFromOptions(ksp_schur_primal);
    } else {
667
668
      MSG("Create direct schur primal solver!\n");

669
670
      double wtime = MPI::Wtime();

671
672
673
      TEST_EXIT_DBG(meshLevel == 0)
	("Does not support for multilevel, check usage of localDofMap.\n");

674
675
676
      int nRowsRankPrimal = primalDofMap.getRankDofs();
      int nRowsOverallPrimal = primalDofMap.getOverallDofs();
      int nRowsRankB = localDofMap.getRankDofs();
677

Thomas Witkowski's avatar
Thomas Witkowski committed
678
      Mat matBPi;
679
      MatCreateMPIAIJ(mpiComm,
Thomas Witkowski's avatar
Thomas Witkowski committed
680
		      nRowsRankB, nRowsRankPrimal, 
681
		      nGlobalOverallInterior, nRowsOverallPrimal,
Thomas Witkowski's avatar
Thomas Witkowski committed
682
683
		      30, PETSC_NULL, 30, PETSC_NULL, &matBPi);
      Mat matPrimal;
684

685
686
687
      PetscInt nCols;
      const PetscInt *cols;
      const PetscScalar *values;
Thomas Witkowski's avatar
Thomas Witkowski committed
688

689
690
      map<int, vector<pair<int, double> > > mat_b_primal_cols;

691
      for (int i = 0; i < nRowsRankB; i++) {
692
	PetscInt row = localDofMap.getStartDofs() + i;
693
	MatGetRow(subDomainSolver->getMatIntCoarse(), row, &nCols, &cols, &values);
694
695
696
697
698

	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]));

699
	MatRestoreRow(subDomainSolver->getMatIntCoarse(), row, &nCols, &cols, &values);
700
701
      }

702
      TEST_EXIT(static_cast<int>(mat_b_primal_cols.size()) == 
703
		primalDofMap.getLocalDofs())
704
	("Should not happen!\n");
705

706
707
708
      for (map<int, vector<pair<int, double> > >::iterator it = mat_b_primal_cols.begin();
	   it != mat_b_primal_cols.end(); ++it) {
	Vec tmpVec;
709
	VecCreateSeq(PETSC_COMM_SELF, nRowsRankB, &tmpVec);
710
711
712
713
714
715
716
717

 	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);

718
	subDomainSolver->solve(tmpVec, tmpVec);
719

Thomas Witkowski's avatar
Thomas Witkowski committed
720
	PetscScalar *tmpValues;
721
	VecGetArray(tmpVec, &tmpValues);
722
	for (int i  = 0; i < nRowsRankB; i++)
Thomas Witkowski's avatar
Thomas Witkowski committed
723
	  MatSetValue(matBPi, 
724
		      localDofMap.getStartDofs() + i,
Thomas Witkowski's avatar
Thomas Witkowski committed
725
726
727
		      it->first,
		      tmpValues[i],
		      ADD_VALUES);
728
729
730
731
732
	VecRestoreArray(tmpVec, &tmpValues);

	VecDestroy(&tmpVec);
      }

Thomas Witkowski's avatar
Thomas Witkowski committed
733
734
      MatAssemblyBegin(matBPi, MAT_FINAL_ASSEMBLY);
      MatAssemblyEnd(matBPi, MAT_FINAL_ASSEMBLY);
735
736

      MatDuplicate(subDomainSolver->getMatCoarseCoarse(), MAT_COPY_VALUES, &mat_schur_primal);
737
      MatMatMult(subDomainSolver->getMatCoarseInt(), matBPi, MAT_INITIAL_MATRIX, PETSC_DEFAULT, &matPrimal);
738
      MatAXPY(mat_schur_primal, -1.0, matPrimal, DIFFERENT_NONZERO_PATTERN);
739
740

      MatDestroy(&matPrimal);
Thomas Witkowski's avatar
Thomas Witkowski committed
741
      MatDestroy(&matBPi);
742
743

      MatInfo minfo;
744
      MatGetInfo(mat_schur_primal, MAT_GLOBAL_SUM, &minfo);
745
746
      MSG("Schur primal matrix nnz = %f\n", minfo.nz_used);

747
      KSPCreate(mpiComm, &ksp_schur_primal);
748
      KSPSetOperators(ksp_schur_primal, mat_schur_primal, mat_schur_primal,
749
		      SAME_NONZERO_PATTERN);
750
751
752
753
754
755
      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);
756
      KSPSetFromOptions(ksp_schur_primal);
Thomas Witkowski's avatar
Thomas Witkowski committed
757

758
759
      MSG("Creating Schur primal matrix needed %.5f seconds.\n",
	  MPI::Wtime() - wtime);
Thomas Witkowski's avatar
Thomas Witkowski committed
760
    }
761
762
763
764
765
766
767
  }


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

768
    if (schurPrimalSolver == 0) {
769
      schurPrimalData.subSolver = NULL;
770

771
772
773
      VecDestroy(&schurPrimalData.tmp_vec_b);
      VecDestroy(&schurPrimalData.tmp_vec_primal);
    }
774
775
776
    
    MatDestroy(&mat_schur_primal);
    KSPDestroy(&ksp_schur_primal);
777
778
779
  }


780
  void PetscSolverFeti::createFetiKsp(vector<const FiniteElemSpace*> &feSpaces)
781
782
783
  {
    FUNCNAME("PetscSolverFeti::createFetiKsp()");

784
785
    // === Create FETI-DP solver object. ===

786
    fetiData.mat_lagrange = &mat_lagrange;
787
    fetiData.subSolver = subDomainSolver;
788
    fetiData.ksp_schur_primal = &ksp_schur_primal;
789

790
    VecCreateMPI(mpiComm, 
791
		 localDofMap.getRankDofs(), 
792
		 nGlobalOverallInterior,
Thomas Witkowski's avatar
Thomas Witkowski committed
793
		 &(fetiData.tmp_vec_b));
794
    VecCreateMPI(mpiComm,
795
796
		 lagrangeMap.getRankDofs(), 
		 lagrangeMap.getOverallDofs(),
Thomas Witkowski's avatar
Thomas Witkowski committed
797
		 &(fetiData.tmp_vec_lagrange));
798
    VecCreateMPI(mpiComm, 
799
800
		 primalDofMap.getRankDofs(), 
		 primalDofMap.getOverallDofs(),
801
		 &(fetiData.tmp_vec_primal));
802

803
    MatCreateShell(mpiComm,
804
805
806
807
		   lagrangeMap.getRankDofs(), 
		   lagrangeMap.getRankDofs(),
		   lagrangeMap.getOverallDofs(), 
		   lagrangeMap.getOverallDofs(),
808
		   &fetiData, &mat_feti);
809
810
811
    MatShellSetOperation(mat_feti, MATOP_MULT, (void(*)(void))petscMultMatFeti);


812
    KSPCreate(mpiComm, &ksp_feti);
813
    KSPSetOperators(ksp_feti, mat_feti, mat_feti, SAME_NONZERO_PATTERN);
814
815
816
    KSPSetOptionsPrefix(ksp_feti, "feti_");
    KSPSetType(ksp_feti, KSPGMRES);
    KSPSetTolerances(ksp_feti, 0, 1e-8, 1e+3, 1000);
817
    KSPSetFromOptions(ksp_feti);
818
819


820
    // === Create FETI-DP preconditioner object. ===
821

822
823
824
825
    if (fetiPreconditioner != FETI_NONE) {
      MatDuplicate(mat_lagrange, MAT_COPY_VALUES, &mat_lagrange_scaled);
      MatScale(mat_lagrange_scaled, 0.5);
    }
826

827
    switch (fetiPreconditioner) {
828
829
830
831
    case FETI_DIRICHLET:      
      TEST_EXIT_DBG(meshLevel == 0)
	("Check for localDofMap.getLocalMatIndex, which should not work for multilevel FETI-DP!\n");

832
      KSPCreate(PETSC_COMM_SELF, &ksp_interior);
Thomas Witkowski's avatar
Thomas Witkowski committed
833
834
      KSPSetOperators(ksp_interior, mat_interior_interior, mat_interior_interior, 
		      SAME_NONZERO_PATTERN);
835
836
837
838
839
840
      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);
841
842
843
844
845
846
847
848
849
      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;
      
850
      VecCreateMPI(mpiComm, 
851
		   localDofMap.getRankDofs(),
852
		   nGlobalOverallInterior,
853
		   &(fetiDirichletPreconData.tmp_vec_b));      
854
855
856
857
858
859
      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));
860

861
862
863
      TEST_EXIT_DBG(meshLevel == 0)
	("Should not happen, check usage of localDofMap!\n");

864
      for (unsigned int i = 0; i < feSpaces.size(); i++) {
865
	DofMap &dualMap = dualDofMap[feSpaces[i]].getMap();
866
	for (DofMap::iterator it = dualMap.begin(); it != dualMap.end(); ++it) {
867
	  DegreeOfFreedom d = it->first;	  
868
869
	  int matIndexLocal = localDofMap.getLocalMatIndex(i, d);
	  int matIndexDual = dualDofMap.getLocalMatIndex(i, d);
870
871
872
873
	  fetiDirichletPreconData.localToDualMap[matIndexLocal] = matIndexDual;
	}
      }

874
875
876
877
878
879
880
881
882
883
884
      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
885
      for (unsigned int i = 0; i < feSpaces.size(); i++) {
886
	DofMap &dualMap = dualDofMap[feSpaces[i]].getMap();
887
	for (DofMap::iterator it = dualMap.begin(); it != dualMap.end(); ++it) {
Thomas Witkowski's avatar
Thomas Witkowski committed
888
	  DegreeOfFreedom d = it->first;
889
890
	  int matIndexLocal = localDofMap.getLocalMatIndex(i, d);
	  int matIndexDual = dualDofMap.getLocalMatIndex(i, d);
Thomas Witkowski's avatar
Thomas Witkowski committed
891
892
893
894
	  fetiLumpedPreconData.localToDualMap[matIndexLocal] = matIndexDual;
	}
      }

895
      VecCreateMPI(mpiComm, 
896
897
		   localDofMap.getRankDofs(),
		   localDofMap.getOverallDofs(),
898
		   &(fetiLumpedPreconData.tmp_vec_b));
899
900
901
902
      MatGetVecs(mat_duals_duals, PETSC_NULL, 
		 &(fetiLumpedPreconData.tmp_vec_duals0));
      MatGetVecs(mat_duals_duals, PETSC_NULL, 
		 &(fetiLumpedPreconData.tmp_vec_duals1));
903
904
905
906
907
908

      KSPGetPC(ksp_feti, &precon_feti);
      PCSetType(precon_feti, PCSHELL);
      PCShellSetContext(precon_feti, static_cast<void*>(&fetiLumpedPreconData));
      PCShellSetApply(precon_feti, petscApplyFetiLumpedPrecon);
      
909
910
      break;
    default:
911
912
      break;
    }
913
914
915
916
917
918
919
  }
  

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

920
921
    // === Destroy FETI-DP solver object. ===

922
    fetiData.mat_lagrange = PETSC_NULL;
923
    fetiData.subSolver = NULL;
924
    fetiData.ksp_schur_primal = PETSC_NULL;
925

Thomas Witkowski's avatar
Thomas Witkowski committed
926
927
    VecDestroy(&fetiData.tmp_vec_b);
    VecDestroy(&fetiData.tmp_vec_lagrange);
928
    VecDestroy(&fetiData.tmp_vec_primal);