PetscSolverFeti.cc 41.3 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);

231 232 233 234 235 236 237 238 239 240 241
    if (subDomainSolver == NULL) {
      if (meshLevel == 0) {
	subDomainSolver = 
	  new SubDomainSolver(meshDistributor, mpiComm, mpiSelfComm);
      } else {
	MeshLevelData& levelData = meshDistributor->getMeshLevelData();

	subDomainSolver = 
	  new SubDomainSolver(meshDistributor, 
			      levelData.getMpiComm(meshLevel - 1),
			      levelData.getMpiComm(meshLevel));
242
	subDomainSolver->setLevel(meshLevel);
243 244
      }
    }
245 246 247

    primalDofMap.init(mpiComm, feSpaces, meshDistributor->getFeSpaces(), 
		      true, true);
248 249
    primalDofMap.setComputeMatIndex(true);
    
250 251
    dualDofMap.init(mpiComm, feSpaces, meshDistributor->getFeSpaces(),
		    false, false);
252 253
    dualDofMap.setComputeMatIndex(true);

254 255
    lagrangeMap.init(mpiComm, feSpaces, meshDistributor->getFeSpaces(),
		     true, true);
256 257
    lagrangeMap.setComputeMatIndex(true);
 
258 259
    localDofMap.init(mpiComm, feSpaces, meshDistributor->getFeSpaces(),
		     false, false);
260 261 262
    localDofMap.setComputeMatIndex(true);

    if (fetiPreconditioner == FETI_DIRICHLET) {
263 264
      interiorDofMap.init(mpiComm, feSpaces,  meshDistributor->getFeSpaces(),
			  false, false);
265 266
      interiorDofMap.setComputeMatIndex(true);
    }
267 268 269 270 271 272 273 274 275 276
  }


  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");
277

278 279 280 281 282 283
    primalDofMap.clear();
    dualDofMap.clear();
    lagrangeMap.clear();
    localDofMap.clear();
    interiorDofMap.clear();

284 285
    for (unsigned int i = 0; i < meshDistributor->getFeSpaces().size(); i++) {
      const FiniteElemSpace *feSpace = meshDistributor->getFeSpace(i);
286

287 288
      createPrimals(feSpace);      
      createDuals(feSpace);
289
      createLagrange(feSpace);      
290 291
      createIndexB(feSpace);
    }
Thomas Witkowski's avatar
Thomas Witkowski committed
292

293 294 295 296 297
    primalDofMap.setDofComm(meshDistributor->getSendDofs(),
			    meshDistributor->getRecvDofs());
    lagrangeMap.setDofComm(meshDistributor->getSendDofs(), 
			   meshDistributor->getRecvDofs());

Thomas Witkowski's avatar
Thomas Witkowski committed
298 299 300 301
    primalDofMap.update();
    dualDofMap.update();
    lagrangeMap.update();
    localDofMap.update();
302
    if (fetiPreconditioner != FETI_NONE)
303
      interiorDofMap.update();
Thomas Witkowski's avatar
Thomas Witkowski committed
304 305 306

    for (unsigned int i = 0; i < meshDistributor->getFeSpaces().size(); i++) {
      const FiniteElemSpace *feSpace = meshDistributor->getFeSpace(i);
307
      MSG("FETI-DP data for %d-ith FE space:\n", i);
Thomas Witkowski's avatar
Thomas Witkowski committed
308

309
      MSG("  nRankPrimals = %d   nOverallPrimals = %d\n", 
310
	  primalDofMap[feSpace].nRankDofs[0], primalDofMap[feSpace].nOverallDofs[0]);
Thomas Witkowski's avatar
Thomas Witkowski committed
311
      
312
      MSG("  nRankDuals = %d  nOverallDuals = %d\n",
313
	  dualDofMap[feSpace].nRankDofs[0], dualDofMap[feSpace].nOverallDofs[0]);
Thomas Witkowski's avatar
Thomas Witkowski committed
314

315
      MSG("  nRankLagrange = %d  nOverallLagrange = %d\n",
316
	  lagrangeMap[feSpace].nRankDofs[0], lagrangeMap[feSpace].nOverallDofs[0]);
317 318 319 320

      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
321
    }
322 323 324


    // If multi level test, inform sub domain solver about coarse space.
325
    subDomainSolver->setDofMapping(&primalDofMap, &localDofMap);
326 327 328
  }


329
  void PetscSolverFeti::createPrimals(const FiniteElemSpace *feSpace)
Thomas Witkowski's avatar
Thomas Witkowski committed
330
  {
331
    FUNCNAME("PetscSolverFeti::createPrimals()");  
332

333 334 335
    // === Define all vertices on the interior boundaries of the macro mesh ===
    // === to be primal variables.                                          ===

336 337 338
    /// Set of DOF indices that are considered to be primal variables.
    
    DofIndexSet primals;
339
    DofContainerSet& vertices = 
340
      meshDistributor->getBoundaryDofInfo(feSpace, meshLevel).geoDofs[VERTEX];
341 342 343
    for (DofContainerSet::iterator it = vertices.begin(); 
	 it != vertices.end(); ++it)
      primals.insert(**it);
344

345 346 347 348

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

349
    for (DofIndexSet::iterator it = primals.begin(); it != primals.end(); ++it)
350
      if (meshDistributor->getDofMap()[feSpace].isRankDof(*it))
351
	primalDofMap[feSpace].insertRankDof(*it);
Thomas Witkowski's avatar
FETI-DP  
Thomas Witkowski committed
352 353
      else
  	primalDofMap[feSpace].insert(*it);
354 355 356
  }


357
  void PetscSolverFeti::createDuals(const FiniteElemSpace *feSpace)
358 359
  {
    FUNCNAME("PetscSolverFeti::createDuals()");
360

361 362 363
    // === Create global index of the dual nodes on each rank. ===

    DofContainer allBoundaryDofs;
364
    meshDistributor->getAllBoundaryDofs(feSpace, meshLevel, allBoundaryDofs);
365 366 367 368 369 370 371 372 373 374 375 376

    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()");

377 378
    // === Create for each dual node that is owned by the rank, the set ===
    // === of ranks that contain this node (denoted by W(x_j)).         ===
379

380
    boundaryDofRanks[feSpace].clear();
381

382
    for (DofComm::Iterator it(meshDistributor->getSendDofs(), meshLevel, feSpace); 
383 384
	 !it.end(); it.nextRank())
      for (; !it.endDofIter(); it.nextDof()) {
385
	if (!isPrimal(feSpace, it.getDofIndex())) {
386 387
	  boundaryDofRanks[feSpace][it.getDofIndex()].insert(mpiRank);
	  boundaryDofRanks[feSpace][it.getDofIndex()].insert(it.getRank());
388 389
	}
      }
390
	
391 392 393 394

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

395
    StdMpi<vector<std::set<int> > > stdMpi(meshDistributor->getMpiComm());
396

397
    for (DofComm::Iterator it(meshDistributor->getSendDofs(), meshLevel, feSpace);
398 399
	 !it.end(); it.nextRank())
      for (; !it.endDofIter(); it.nextDof())
400
	if (!isPrimal(feSpace, it.getDofIndex()))
401
	  stdMpi.getSendData(it.getRank()).push_back(boundaryDofRanks[feSpace][it.getDofIndex()]);
402 403 404

    stdMpi.updateSendDataSize();

405
    for (DofComm::Iterator it(meshDistributor->getRecvDofs(), meshLevel, feSpace); 
406
	 !it.end(); it.nextRank()) {
407
      bool recvFromRank = false;
408
      for (; !it.endDofIter(); it.nextDof()) {
409
	if (!isPrimal(feSpace, it.getDofIndex())) {
410 411 412
	  recvFromRank = true;
	  break;
	}
413
      }
414 415

      if (recvFromRank)
416
	stdMpi.recv(it.getRank());
417
    }
418

419 420
    stdMpi.startCommunication();

421
    for (DofComm::Iterator it(meshDistributor->getRecvDofs(), meshLevel, feSpace); 
422
	 !it.end(); it.nextRank()) {
423
      int i = 0;
424
      for (; !it.endDofIter(); it.nextDof())
425
	if (!isPrimal(feSpace, it.getDofIndex()))
426
	  boundaryDofRanks[feSpace][it.getDofIndex()] = 
427
	    stdMpi.getRecvData(it.getRank())[i++];
428 429
    }

430 431 432
    // === Reserve for each dual node, on the rank that owns this node, the ===
    // === appropriate number of Lagrange constraints.                      ===

433
    int nRankLagrange = 0;
434
    DofMap& dualMap = dualDofMap[feSpace].getMap(0);
435
    for (DofMap::iterator it = dualMap.begin(); it != dualMap.end(); ++it) {
436 437

      if (meshDistributor->getDofMap()[feSpace].isRankDof(it->first)) {
Thomas Witkowski's avatar
Thomas Witkowski committed
438
	lagrangeMap[feSpace].insertRankDof(it->first, nRankLagrange);
439
	int degree = boundaryDofRanks[feSpace][it->first].size();
440
	nRankLagrange += (degree * (degree - 1)) / 2;
Thomas Witkowski's avatar
Thomas Witkowski committed
441 442
      } else {
	lagrangeMap[feSpace].insert(it->first);
443 444
      }
    }
445
    lagrangeMap[feSpace].nRankDofs[0] = nRankLagrange;
446 447 448
  }


449
  void PetscSolverFeti::createIndexB(const FiniteElemSpace *feSpace)
450
  {
451
    FUNCNAME("PetscSolverFeti::createIndexB()");
452

453
    DOFAdmin* admin = feSpace->getAdmin();
454 455 456 457

    // === 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.                                 ===
458

459
    int nLocalInterior = 0;    
460
    for (int i = 0; i < admin->getUsedSize(); i++) {
461
      if (admin->isDofFree(i) == false && 
462
	  isPrimal(feSpace, i) == false &&
463
	  dualDofMap[feSpace].isSet(i) == false) {
464 465
	localDofMap[feSpace].insertRankDof(i, nLocalInterior);

466
	if (fetiPreconditioner != FETI_NONE)
467 468 469
	  interiorDofMap[feSpace].insertRankDof(i, nLocalInterior);

	nLocalInterior++;
470
      }
471
    }
Thomas Witkowski's avatar
FETI-DP  
Thomas Witkowski committed
472
    
473 474
    // === And finally, add the global indicies of all dual nodes. ===

475 476
    for (DofMap::iterator it = dualDofMap[feSpace].getMap(0).begin();
	 it != dualDofMap[feSpace].getMap(0).end(); ++it)
477
      localDofMap[feSpace].insertRankDof(it->first);
478 479 480
  }


481
  void PetscSolverFeti::createMatLagrange(vector<const FiniteElemSpace*> &feSpaces)
482 483 484
  {
    FUNCNAME("PetscSolverFeti::createMatLagrange()");

485 486
    // === Create distributed matrix for Lagrange constraints. ===

487
    MatCreateMPIAIJ(mpiComm,
488 489
		    lagrangeMap.getRankDofs(0), localDofMap.getRankDofs(0),
		    lagrangeMap.getOverallDofs(0), localDofMap.getOverallDofs(0),
490 491 492
		    2, PETSC_NULL, 2, PETSC_NULL,
		    &mat_lagrange);

493 494 495 496 497 498 499
    // === 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
500
    for (unsigned int k = 0; k < feSpaces.size(); k++) {
501
      DofMap &dualMap = dualDofMap[feSpaces[k]].getMap(0);
502

503
      for (DofMap::iterator it = dualMap.begin(); it != dualMap.end(); ++it) {
504
	TEST_EXIT_DBG(boundaryDofRanks[feSpaces[k]].count(it->first))
Thomas Witkowski's avatar
Thomas Witkowski committed
505 506 507 508 509 510
	  ("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.
511 512
	vector<int> W(boundaryDofRanks[feSpaces[k]][it->first].begin(), 
		      boundaryDofRanks[feSpaces[k]][it->first].end());
Thomas Witkowski's avatar
Thomas Witkowski committed
513 514 515 516 517 518
	// 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) {
519
	      int colIndex = localDofMap.getMatIndex(k, it->first);
520
	      double value = (W[i] == mpiRank ? 1.0 : -1.0);
521
	      MatSetValue(mat_lagrange, index, colIndex, value, INSERT_VALUES);
522
	    }
Thomas Witkowski's avatar
Thomas Witkowski committed
523
	    index++;	      
524 525 526 527 528 529 530 531 532 533
	  }
	}
      }
    }

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


534
  void PetscSolverFeti::createSchurPrimalKsp(vector<const FiniteElemSpace*> &feSpaces)
535 536 537
  {
    FUNCNAME("PetscSolverFeti::createSchurPrimal()");

Thomas Witkowski's avatar
Thomas Witkowski committed
538
    if (schurPrimalSolver == 0) {
539 540
      MSG("Create iterative schur primal solver!\n");

541
      schurPrimalData.subSolver = subDomainSolver;
Thomas Witkowski's avatar
Thomas Witkowski committed
542
      
543
      VecCreateMPI(mpiComm, 
544
		   localDofMap.getRankDofs(0), localDofMap.getOverallDofs(0),
Thomas Witkowski's avatar
Thomas Witkowski committed
545
		   &(schurPrimalData.tmp_vec_b));
546
      VecCreateMPI(mpiComm, 
547
		   primalDofMap.getRankDofs(0), primalDofMap.getOverallDofs(0),
Thomas Witkowski's avatar
Thomas Witkowski committed
548 549
		   &(schurPrimalData.tmp_vec_primal));

550
      MatCreateShell(mpiComm,
551 552
		     primalDofMap.getRankDofs(0), primalDofMap.getRankDofs(0), 
		     primalDofMap.getOverallDofs(0), primalDofMap.getOverallDofs(0),
Thomas Witkowski's avatar
Thomas Witkowski committed
553 554 555 556 557
		     &schurPrimalData, 
		     &mat_schur_primal);
      MatShellSetOperation(mat_schur_primal, MATOP_MULT, 
			   (void(*)(void))petscMultMatSchurPrimal);
      
558
      KSPCreate(mpiComm, &ksp_schur_primal);
Thomas Witkowski's avatar
Thomas Witkowski committed
559
      KSPSetOperators(ksp_schur_primal, mat_schur_primal, mat_schur_primal, SAME_NONZERO_PATTERN);
560 561
      KSPSetOptionsPrefix(ksp_schur_primal, "schur_primal_");
      KSPSetType(ksp_schur_primal, KSPGMRES);
Thomas Witkowski's avatar
Thomas Witkowski committed
562 563
      KSPSetFromOptions(ksp_schur_primal);
    } else {
564 565
      MSG("Create direct schur primal solver!\n");

566 567
      double wtime = MPI::Wtime();

568 569 570 571
      int nRowsRankPrimal = primalDofMap.getRankDofs(0);
      int nRowsOverallPrimal = primalDofMap.getOverallDofs(0);
      int nRowsRankB = localDofMap.getRankDofs(0);
      int nRowsOverallB = localDofMap.getOverallDofs(0);
572

Thomas Witkowski's avatar
Thomas Witkowski committed
573
      Mat matBPi;
574
      MatCreateMPIAIJ(mpiComm,
Thomas Witkowski's avatar
Thomas Witkowski committed
575 576 577 578
		      nRowsRankB, nRowsRankPrimal, 
		      nRowsOverallB, nRowsOverallPrimal,
		      30, PETSC_NULL, 30, PETSC_NULL, &matBPi);
      Mat matPrimal;
579

580 581 582
      PetscInt nCols;
      const PetscInt *cols;
      const PetscScalar *values;
Thomas Witkowski's avatar
Thomas Witkowski committed
583

584 585
      map<int, vector<pair<int, double> > > mat_b_primal_cols;

586
      for (int i = 0; i < nRowsRankB; i++) {
587
	PetscInt row = localDofMap.getStartDofs(0) + i;
588
	MatGetRow(subDomainSolver->getMatIntCoarse(), row, &nCols, &cols, &values);
589 590 591 592 593

	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]));

594
	MatRestoreRow(subDomainSolver->getMatIntCoarse(), row, &nCols, &cols, &values);
595 596
      }

597
      TEST_EXIT(static_cast<int>(mat_b_primal_cols.size()) == 
598
		primalDofMap.getLocalDofs(0))
599
	("Should not happen!\n");
600

601 602 603
      for (map<int, vector<pair<int, double> > >::iterator it = mat_b_primal_cols.begin();
	   it != mat_b_primal_cols.end(); ++it) {
	Vec tmpVec;
604
	VecCreateSeq(PETSC_COMM_SELF, nRowsRankB, &tmpVec);
605 606 607 608 609 610 611 612

 	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);

613
	subDomainSolver->solve(tmpVec, tmpVec);
614

Thomas Witkowski's avatar
Thomas Witkowski committed
615
	PetscScalar *tmpValues;
616
	VecGetArray(tmpVec, &tmpValues);
617
	for (int i  = 0; i < nRowsRankB; i++)
Thomas Witkowski's avatar
Thomas Witkowski committed
618
	  MatSetValue(matBPi, 
619
		      localDofMap.getStartDofs(0) + i,
Thomas Witkowski's avatar
Thomas Witkowski committed
620 621 622
		      it->first,
		      tmpValues[i],
		      ADD_VALUES);
623 624 625 626 627
	VecRestoreArray(tmpVec, &tmpValues);

	VecDestroy(&tmpVec);
      }

Thomas Witkowski's avatar
Thomas Witkowski committed
628 629
      MatAssemblyBegin(matBPi, MAT_FINAL_ASSEMBLY);
      MatAssemblyEnd(matBPi, MAT_FINAL_ASSEMBLY);
630 631

      MatDuplicate(subDomainSolver->getMatCoarseCoarse(), MAT_COPY_VALUES, &mat_schur_primal);
632
      MatMatMult(subDomainSolver->getMatCoarseInt(), matBPi, MAT_INITIAL_MATRIX, PETSC_DEFAULT, &matPrimal);
633
      MatAXPY(mat_schur_primal, -1.0, matPrimal, DIFFERENT_NONZERO_PATTERN);
634 635

      MatDestroy(&matPrimal);
Thomas Witkowski's avatar
Thomas Witkowski committed
636
      MatDestroy(&matBPi);
637 638

      MatInfo minfo;
639
      MatGetInfo(mat_schur_primal, MAT_GLOBAL_SUM, &minfo);
640 641
      MSG("Schur primal matrix nnz = %f\n", minfo.nz_used);

642
      KSPCreate(mpiComm, &ksp_schur_primal);
643
      KSPSetOperators(ksp_schur_primal, mat_schur_primal, mat_schur_primal,
644
		      SAME_NONZERO_PATTERN);
645 646 647 648 649 650
      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);
651
      KSPSetFromOptions(ksp_schur_primal);
Thomas Witkowski's avatar
Thomas Witkowski committed
652

653 654
      MSG("Creating Schur primal matrix needed %.5f seconds.\n",
	  MPI::Wtime() - wtime);
Thomas Witkowski's avatar
Thomas Witkowski committed
655
    }
656 657 658 659 660 661 662
  }


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

663
    if (schurPrimalSolver == 0) {
664
      schurPrimalData.subSolver = NULL;
665

666 667 668
      VecDestroy(&schurPrimalData.tmp_vec_b);
      VecDestroy(&schurPrimalData.tmp_vec_primal);
    }
669 670 671
    
    MatDestroy(&mat_schur_primal);
    KSPDestroy(&ksp_schur_primal);
672 673 674
  }


675
  void PetscSolverFeti::createFetiKsp(vector<const FiniteElemSpace*> &feSpaces)
676 677 678
  {
    FUNCNAME("PetscSolverFeti::createFetiKsp()");

679 680
    // === Create FETI-DP solver object. ===

681
    fetiData.mat_lagrange = &mat_lagrange;
682
    fetiData.subSolver = subDomainSolver;
683
    fetiData.ksp_schur_primal = &ksp_schur_primal;
684

685
    VecCreateMPI(mpiComm, 
686
		 localDofMap.getRankDofs(0), localDofMap.getOverallDofs(0),
Thomas Witkowski's avatar
Thomas Witkowski committed
687
		 &(fetiData.tmp_vec_b));
688
    VecCreateMPI(mpiComm,
689
		 lagrangeMap.getRankDofs(0), lagrangeMap.getOverallDofs(0),
Thomas Witkowski's avatar
Thomas Witkowski committed
690
		 &(fetiData.tmp_vec_lagrange));
691
    VecCreateMPI(mpiComm, 
692
		 primalDofMap.getRankDofs(0), primalDofMap.getOverallDofs(0),
693
		 &(fetiData.tmp_vec_primal));
694

695
    MatCreateShell(mpiComm,
696 697
		   lagrangeMap.getRankDofs(0), lagrangeMap.getRankDofs(0),
		   lagrangeMap.getOverallDofs(0), lagrangeMap.getOverallDofs(0),
698
		   &fetiData, &mat_feti);
699 700 701
    MatShellSetOperation(mat_feti, MATOP_MULT, (void(*)(void))petscMultMatFeti);


702
    KSPCreate(mpiComm, &ksp_feti);
703
    KSPSetOperators(ksp_feti, mat_feti, mat_feti, SAME_NONZERO_PATTERN);
704 705 706
    KSPSetOptionsPrefix(ksp_feti, "feti_");
    KSPSetType(ksp_feti, KSPGMRES);
    KSPSetTolerances(ksp_feti, 0, 1e-8, 1e+3, 1000);
707
    KSPSetFromOptions(ksp_feti);
708 709


710
    // === Create FETI-DP preconditioner object. ===
711

712 713 714 715
    if (fetiPreconditioner != FETI_NONE) {
      MatDuplicate(mat_lagrange, MAT_COPY_VALUES, &mat_lagrange_scaled);
      MatScale(mat_lagrange_scaled, 0.5);
    }
716

717 718 719
    switch (fetiPreconditioner) {
    case FETI_DIRICHLET:           
      KSPCreate(PETSC_COMM_SELF, &ksp_interior);
Thomas Witkowski's avatar
Thomas Witkowski committed
720 721
      KSPSetOperators(ksp_interior, mat_interior_interior, mat_interior_interior, 
		      SAME_NONZERO_PATTERN);
722 723 724 725 726 727
      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);
728 729 730 731 732 733 734 735 736
      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;
      
737
      VecCreateMPI(mpiComm, 
738
		   localDofMap.getRankDofs(0),localDofMap.getOverallDofs(0),
739
		   &(fetiDirichletPreconData.tmp_vec_b));      
740 741 742 743 744 745
      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));
746 747

      for (unsigned int i = 0; i < feSpaces.size(); i++) {
748
	DofMap &dualMap = dualDofMap[feSpaces[i]].getMap(0);
749
	for (DofMap::iterator it = dualMap.begin(); it != dualMap.end(); ++it) {
750 751 752 753 754 755 756
	  DegreeOfFreedom d = it->first;
	  int matIndexLocal = localDofMap.getLocalMatIndex(i, d);
	  int matIndexDual = dualDofMap.getLocalMatIndex(i, d);
	  fetiDirichletPreconData.localToDualMap[matIndexLocal] = matIndexDual;
	}
      }

757 758 759 760 761 762 763 764 765 766 767
      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
768
      for (unsigned int i = 0; i < feSpaces.size(); i++) {
769
	DofMap &dualMap = dualDofMap[feSpaces[i]].getMap(0);
770
	for (DofMap::iterator it = dualMap.begin(); it != dualMap.end(); ++it) {
Thomas Witkowski's avatar
Thomas Witkowski committed
771 772 773 774 775 776 777
	  DegreeOfFreedom d = it->first;
	  int matIndexLocal = localDofMap.getLocalMatIndex(i, d);
	  int matIndexDual = dualDofMap.getLocalMatIndex(i, d);
	  fetiLumpedPreconData.localToDualMap[matIndexLocal] = matIndexDual;
	}
      }

778
      VecCreateMPI(mpiComm, 
779 780
		   localDofMap.getRankDofs(0),
		   localDofMap.getOverallDofs(0),
781
		   &(fetiLumpedPreconData.tmp_vec_b));
782 783 784 785
      MatGetVecs(mat_duals_duals, PETSC_NULL, 
		 &(fetiLumpedPreconData.tmp_vec_duals0));
      MatGetVecs(mat_duals_duals, PETSC_NULL, 
		 &(fetiLumpedPreconData.tmp_vec_duals1));
786 787 788 789 790 791

      KSPGetPC(ksp_feti, &precon_feti);
      PCSetType(precon_feti, PCSHELL);
      PCShellSetContext(precon_feti, static_cast<void*>(&fetiLumpedPreconData));
      PCShellSetApply(precon_feti, petscApplyFetiLumpedPrecon);
      
792 793
      break;
    default: