PetscSolverFeti.cc 41.1 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13
//
// Software License for AMDiS
//
// Copyright (c) 2010 Dresden University of Technology 
// All rights reserved.
// Authors: Simon Vey, Thomas Witkowski et al.
//
// This file is part of AMDiS
//
// See also license.opensource.txt in the distribution.


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

namespace AMDiS {

  using namespace std;

24 25 26 27 28 29 30
  // 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);
31
    SchurPrimalData* data = static_cast<SchurPrimalData*>(ctx);
32

33 34 35 36
    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);
37 38 39 40 41 42 43 44 45
    VecAXPBY(y, -1.0, 1.0, data->tmp_vec_primal);

    return 0;
  }


  // y = mat * x
  int petscMultMatFeti(Mat mat, Vec x, Vec y)
  {
46 47
    //    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)
48 49 50

    void *ctx;
    MatShellGetContext(mat, &ctx);
51
    FetiData* data = static_cast<FetiData*>(ctx);
52

Thomas Witkowski's avatar
Thomas Witkowski committed
53
    MatMultTranspose(*(data->mat_lagrange), x, data->tmp_vec_b);
54
    data->subSolver->solveGlobal(data->tmp_vec_b, data->tmp_vec_b);
Thomas Witkowski's avatar
Thomas Witkowski committed
55
    MatMult(*(data->mat_lagrange), data->tmp_vec_b, data->tmp_vec_lagrange);
Thomas Witkowski's avatar
Thomas Witkowski committed
56

57
    MatMult(data->subSolver->getMatCoarseInt(), data->tmp_vec_b, data->tmp_vec_primal);
58
    KSPSolve(*(data->ksp_schur_primal), data->tmp_vec_primal, data->tmp_vec_primal);
59
   MatMult(data->subSolver->getMatIntCoarse(), data->tmp_vec_primal, data->tmp_vec_b);
60
    data->subSolver->solveGlobal(data->tmp_vec_b, data->tmp_vec_b);
Thomas Witkowski's avatar
Thomas Witkowski committed
61
    MatMult(*(data->mat_lagrange), data->tmp_vec_b, y);
62

Thomas Witkowski's avatar
Thomas Witkowski committed
63
    VecAXPBY(y, 1.0, 1.0, data->tmp_vec_lagrange);
64 65 66 67 68

    return 0;
  }


69
  // y = PC * x
70
  PetscErrorCode petscApplyFetiDirichletPrecon(PC pc, Vec x, Vec y)
71
  {
72
    // Get data for the preconditioner
73 74
    void *ctx;
    PCShellGetContext(pc, &ctx);
75
    FetiDirichletPreconData* data = static_cast<FetiDirichletPreconData*>(ctx);
76

77
    // Multiply with scaled Lagrange constraint matrix.
78 79 80
    MatMultTranspose(*(data->mat_lagrange_scaled), x, data->tmp_vec_b);


81
    // === Restriction of the B nodes to the boundary nodes. ===
82

83 84 85 86
    int nLocalB;
    int nLocalDuals;
    VecGetLocalSize(data->tmp_vec_b, &nLocalB);
    VecGetLocalSize(data->tmp_vec_duals0, &nLocalDuals);
87

88 89 90 91
    PetscScalar *local_b, *local_duals;
    VecGetArray(data->tmp_vec_b, &local_b);
    VecGetArray(data->tmp_vec_duals0, &local_duals);

92 93 94
    for (map<int, int>::iterator it = data->localToDualMap.begin();
	 it != data->localToDualMap.end(); ++it)
      local_duals[it->second] = local_b[it->first];
95 96

    VecRestoreArray(data->tmp_vec_b, &local_b);
97
    VecRestoreArray(data->tmp_vec_duals0, &local_duals);
98 99


100
    // === K_DD - K_DI inv(K_II) K_ID ===
101

102
    MatMult(*(data->mat_duals_duals), data->tmp_vec_duals0, data->tmp_vec_duals1);
103

104
    MatMult(*(data->mat_interior_duals), data->tmp_vec_duals0, data->tmp_vec_interior);
105
    KSPSolve(*(data->ksp_interior), data->tmp_vec_interior, data->tmp_vec_interior);
106 107 108 109 110 111 112 113 114 115
    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);

116 117 118
    for (map<int, int>::iterator it = data->localToDualMap.begin();
	 it != data->localToDualMap.end(); ++it)
      local_b[it->first] = local_duals[it->second];
119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140

    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);
141 142


143
    // === Restriction of the B nodes to the boundary nodes. ===
144

145 146 147 148
    int nLocalB;
    int nLocalDuals;
    VecGetLocalSize(data->tmp_vec_b, &nLocalB);
    VecGetLocalSize(data->tmp_vec_duals0, &nLocalDuals);
149

150
    PetscScalar *local_b, *local_duals;
151
    VecGetArray(data->tmp_vec_b, &local_b);
152
    VecGetArray(data->tmp_vec_duals0, &local_duals);
153

Thomas Witkowski's avatar
Thomas Witkowski committed
154 155 156
    for (map<int, int>::iterator it = data->localToDualMap.begin();
	 it != data->localToDualMap.end(); ++it)
      local_duals[it->second] = local_b[it->first];
157 158

    VecRestoreArray(data->tmp_vec_b, &local_b);
159 160 161 162 163 164 165
    VecRestoreArray(data->tmp_vec_duals0, &local_duals);


    // === K_DD ===

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

166

167
    // === Prolongation from local dual nodes to B nodes.
168

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

Thomas Witkowski's avatar
Thomas Witkowski committed
172 173 174
    for (map<int, int>::iterator it = data->localToDualMap.begin();
	 it != data->localToDualMap.end(); ++it)
      local_b[it->first] = local_duals[it->second];
175

176 177
    VecRestoreArray(data->tmp_vec_b, &local_b);
    VecRestoreArray(data->tmp_vec_duals0, &local_duals);
178

179 180

    // Multiply with scaled Lagrange constraint matrix.
181 182 183 184 185 186
    MatMult(*(data->mat_lagrange_scaled), data->tmp_vec_b, y);

    return 0;
  }


187 188
  PetscSolverFeti::PetscSolverFeti()
    : PetscSolver(),
189
      schurPrimalSolver(0),
190
      multiLevelTest(false),
191 192
      subDomainSolver(NULL),
      meshLevel(0)
193 194 195 196 197 198
  {
    FUNCNAME("PetscSolverFeti::PetscSolverFeti()");

    string preconditionerName = "";
    Parameters::get("parallel->solver->precon", preconditionerName);
    if (preconditionerName == "" || preconditionerName == "none") {
199
      MSG("Create FETI-DP solver with no preconditioner!\n");
200 201
      fetiPreconditioner = FETI_NONE;
    } else if (preconditionerName == "dirichlet") {
202
      MSG("Create FETI-DP solver with Dirichlet preconditioner!\n");
203 204
      fetiPreconditioner = FETI_DIRICHLET;
    } else if (preconditionerName == "lumped") {
205
      MSG("Create FETI-DP solver with lumped preconditioner!\n");
206 207
      fetiPreconditioner = FETI_LUMPED;
    } else {
Thomas Witkowski's avatar
Thomas Witkowski committed
208 209
      ERROR_EXIT("Preconditioner \"%s\" not available!\n", 
		 preconditionerName.c_str());
210
    }
Thomas Witkowski's avatar
Thomas Witkowski committed
211 212 213 214 215

    Parameters::get("parallel->feti->schur primal solver", schurPrimalSolver);
    TEST_EXIT(schurPrimalSolver == 0 || schurPrimalSolver == 1)
      ("Wrong solver \"%d\"for the Schur primal complement!\n", 
       schurPrimalSolver);
216 217

    Parameters::get("parallel->multi level test", multiLevelTest);
218 219
    if (multiLevelTest)
      meshLevel = 1;
220 221 222
  }


223
  void PetscSolverFeti::initialize(vector<const FiniteElemSpace*> feSpaces)
224
  {
225 226
    FUNCNAME("PetscSolverFeti::initialize()");

227 228 229
    TEST_EXIT_DBG(meshLevel + 1 == meshDistributor->getMeshLevelData().getLevelNumber())
      ("Mesh hierarchy does not contain %d levels!\n", meshLevel + 1);

230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253
    if (subDomainSolver == NULL)
      subDomainSolver = new SubDomainSolver(meshDistributor, mpiComm, mpiSelfComm);

    primalDofMap.init(mpiComm, feSpaces, meshDistributor->getFeSpaces(), 
		      true, true);
    dualDofMap.init(mpiComm, feSpaces, meshDistributor->getFeSpaces(),
		    false, false);
    lagrangeMap.init(mpiComm, feSpaces, meshDistributor->getFeSpaces(),
		     true, true);
    localDofMap.init(mpiComm, feSpaces, meshDistributor->getFeSpaces(),
		     false, false);
    if (fetiPreconditioner == FETI_DIRICHLET)
      interiorDofMap.init(mpiComm, feSpaces,  meshDistributor->getFeSpaces(),
			  false, false);
  }


  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");
254

255 256 257 258 259 260
    primalDofMap.clear();
    dualDofMap.clear();
    lagrangeMap.clear();
    localDofMap.clear();
    interiorDofMap.clear();

261 262
    for (unsigned int i = 0; i < meshDistributor->getFeSpaces().size(); i++) {
      const FiniteElemSpace *feSpace = meshDistributor->getFeSpace(i);
263

264 265
      createPrimals(feSpace);      
      createDuals(feSpace);
266
      createLagrange(feSpace);      
267 268
      createIndexB(feSpace);
    }
Thomas Witkowski's avatar
Thomas Witkowski committed
269

270 271 272 273 274
    primalDofMap.setDofComm(meshDistributor->getSendDofs(),
			    meshDistributor->getRecvDofs());
    lagrangeMap.setDofComm(meshDistributor->getSendDofs(), 
			   meshDistributor->getRecvDofs());

Thomas Witkowski's avatar
Thomas Witkowski committed
275 276 277 278
    primalDofMap.update();
    dualDofMap.update();
    lagrangeMap.update();
    localDofMap.update();
279
    if (fetiPreconditioner != FETI_NONE)
280
      interiorDofMap.update();
Thomas Witkowski's avatar
Thomas Witkowski committed
281 282 283

    for (unsigned int i = 0; i < meshDistributor->getFeSpaces().size(); i++) {
      const FiniteElemSpace *feSpace = meshDistributor->getFeSpace(i);
284
      MSG("FETI-DP data for %d-ith FE space:\n", i);
Thomas Witkowski's avatar
Thomas Witkowski committed
285

286
      MSG("  nRankPrimals = %d   nOverallPrimals = %d\n", 
Thomas Witkowski's avatar
Thomas Witkowski committed
287 288
	  primalDofMap[feSpace].nRankDofs, primalDofMap[feSpace].nOverallDofs);
      
289
      MSG("  nRankDuals = %d  nOverallDuals = %d\n",
290
	  dualDofMap[feSpace].nRankDofs, dualDofMap[feSpace].nOverallDofs);
Thomas Witkowski's avatar
Thomas Witkowski committed
291

292
      MSG("  nRankLagrange = %d  nOverallLagrange = %d\n",
293
	  lagrangeMap[feSpace].nRankDofs, lagrangeMap[feSpace].nOverallDofs);
294 295 296 297

      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
298
    }
299 300 301


    // If multi level test, inform sub domain solver about coarse space.
302
    subDomainSolver->setDofMapping(&primalDofMap, &localDofMap);
303 304 305
  }


306
  void PetscSolverFeti::createPrimals(const FiniteElemSpace *feSpace)
Thomas Witkowski's avatar
Thomas Witkowski committed
307
  {
308
    FUNCNAME("PetscSolverFeti::createPrimals()");  
309

310 311 312
    // === Define all vertices on the interior boundaries of the macro mesh ===
    // === to be primal variables.                                          ===

313 314 315
    /// Set of DOF indices that are considered to be primal variables.
    
    DofIndexSet primals;
316
    DofContainerSet& vertices = 
317
      meshDistributor->getBoundaryDofInfo(feSpace).geoDofs[VERTEX];
318 319 320 321
    TEST_EXIT_DBG(vertices.size())("No primal vertices on this rank!\n");
    for (DofContainerSet::iterator it = vertices.begin(); 
	 it != vertices.end(); ++it)
      primals.insert(**it);
322

323 324 325 326

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

327
    for (DofIndexSet::iterator it = primals.begin(); it != primals.end(); ++it)
328 329
      if (meshDistributor->getIsRankDof(feSpace, *it))
	primalDofMap[feSpace].insertRankDof(*it);
Thomas Witkowski's avatar
FETI-DP  
Thomas Witkowski committed
330 331
      else
  	primalDofMap[feSpace].insert(*it);
332 333 334
  }


335
  void PetscSolverFeti::createDuals(const FiniteElemSpace *feSpace)
336 337
  {
    FUNCNAME("PetscSolverFeti::createDuals()");
338

339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354
    // === Create global index of the dual nodes on each rank. ===

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

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

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

355 356
    // === Create for each dual node that is owned by the rank, the set ===
    // === of ranks that contain this node (denoted by W(x_j)).         ===
357

358
    boundaryDofRanks[feSpace].clear();
359

360 361 362
    for (DofComm::Iterator it(meshDistributor->getSendDofs(), feSpace); 
	 !it.end(); it.nextRank())
      for (; !it.endDofIter(); it.nextDof()) {
363
	if (!isPrimal(feSpace, it.getDofIndex())) {
364 365
	  boundaryDofRanks[feSpace][it.getDofIndex()].insert(mpiRank);
	  boundaryDofRanks[feSpace][it.getDofIndex()].insert(it.getRank());
366 367
	}
      }
368
	
369 370 371 372

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

373
    StdMpi<vector<std::set<int> > > stdMpi(meshDistributor->getMpiComm());
374 375 376 377

    for (DofComm::Iterator it(meshDistributor->getSendDofs(), feSpace);
	 !it.end(); it.nextRank())
      for (; !it.endDofIter(); it.nextDof())
378
	if (!isPrimal(feSpace, it.getDofIndex()))
379
	  stdMpi.getSendData(it.getRank()).push_back(boundaryDofRanks[feSpace][it.getDofIndex()]);
380 381 382

    stdMpi.updateSendDataSize();

383 384
    for (DofComm::Iterator it(meshDistributor->getRecvDofs(), feSpace); 
	 !it.end(); it.nextRank()) {
385
      bool recvFromRank = false;
386
      for (; !it.endDofIter(); it.nextDof()) {
387
	if (!isPrimal(feSpace, it.getDofIndex())) {
388 389 390
	  recvFromRank = true;
	  break;
	}
391
      }
392 393

      if (recvFromRank)
394
	stdMpi.recv(it.getRank());
395
    }
396

397 398
    stdMpi.startCommunication();

399 400
    for (DofComm::Iterator it(meshDistributor->getRecvDofs(), feSpace); 
	 !it.end(); it.nextRank()) {
401
      int i = 0;
402
      for (; !it.endDofIter(); it.nextDof())
403
	if (!isPrimal(feSpace, it.getDofIndex()))
404
	  boundaryDofRanks[feSpace][it.getDofIndex()] = 
405
	    stdMpi.getRecvData(it.getRank())[i++];
406 407 408
    }


409 410 411
    // === Reserve for each dual node, on the rank that owns this node, the ===
    // === appropriate number of Lagrange constraints.                      ===

412
    int nRankLagrange = 0;
Thomas Witkowski's avatar
Thomas Witkowski committed
413 414
    map<DegreeOfFreedom, MultiIndex>& dualMap = dualDofMap[feSpace].getMap();
    for (map<DegreeOfFreedom, MultiIndex>::iterator it = dualMap.begin(); it != dualMap.end(); ++it) {
415
      if (meshDistributor->getIsRankDof(feSpace, it->first)) {
Thomas Witkowski's avatar
Thomas Witkowski committed
416
	lagrangeMap[feSpace].insertRankDof(it->first, nRankLagrange);
417
	int degree = boundaryDofRanks[feSpace][it->first].size();
418
	nRankLagrange += (degree * (degree - 1)) / 2;
Thomas Witkowski's avatar
Thomas Witkowski committed
419 420
      } else {
	lagrangeMap[feSpace].insert(it->first);
421 422
      }
    }
423
    lagrangeMap[feSpace].nRankDofs = nRankLagrange;
424 425 426
  }


427
  void PetscSolverFeti::createIndexB(const FiniteElemSpace *feSpace)
428
  {
429
    FUNCNAME("PetscSolverFeti::createIndexB()");
430

431
    DOFAdmin* admin = feSpace->getAdmin();
432 433 434 435

    // === 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.                                 ===
436

437
    int nLocalInterior = 0;    
438
    for (int i = 0; i < admin->getUsedSize(); i++) {
439
      if (admin->isDofFree(i) == false && 
440
	  isPrimal(feSpace, i) == false &&
441
	  dualDofMap[feSpace].isSet(i) == false) {
442 443
	localDofMap[feSpace].insertRankDof(i, nLocalInterior);

444
	if (fetiPreconditioner != FETI_NONE)
445 446 447
	  interiorDofMap[feSpace].insertRankDof(i, nLocalInterior);

	nLocalInterior++;
448
      }
449
    }
Thomas Witkowski's avatar
FETI-DP  
Thomas Witkowski committed
450
    
451 452
    // === And finally, add the global indicies of all dual nodes. ===

Thomas Witkowski's avatar
Thomas Witkowski committed
453
    for (map<DegreeOfFreedom, MultiIndex>::iterator it = dualDofMap[feSpace].getMap().begin();
454
	 it != dualDofMap[feSpace].getMap().end(); ++it)
455
      localDofMap[feSpace].insertRankDof(it->first);
456 457 458
  }


459
  void PetscSolverFeti::createMatLagrange(vector<const FiniteElemSpace*> &feSpaces)
460 461 462
  {
    FUNCNAME("PetscSolverFeti::createMatLagrange()");

463 464
    // === Create distributed matrix for Lagrange constraints. ===

465
    MatCreateMPIAIJ(mpiComm,
466 467
		    lagrangeMap.getRankDofs(), localDofMap.getRankDofs(),
		    lagrangeMap.getOverallDofs(), localDofMap.getOverallDofs(),	
468 469 470
		    2, PETSC_NULL, 2, PETSC_NULL,
		    &mat_lagrange);

471 472 473 474 475 476 477
    // === 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
478 479 480
    for (unsigned int k = 0; k < feSpaces.size(); k++) {
      map<DegreeOfFreedom, MultiIndex> &dualMap = 
	dualDofMap[feSpaces[k]].getMap();
481

Thomas Witkowski's avatar
Thomas Witkowski committed
482 483
      for (map<DegreeOfFreedom, MultiIndex>::iterator it = dualMap.begin(); 
	   it != dualMap.end(); ++it) {
484
	TEST_EXIT_DBG(boundaryDofRanks[feSpaces[k]].count(it->first))
Thomas Witkowski's avatar
Thomas Witkowski committed
485 486 487 488 489 490
	  ("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.
491 492
	vector<int> W(boundaryDofRanks[feSpaces[k]][it->first].begin(), 
		      boundaryDofRanks[feSpaces[k]][it->first].end());
Thomas Witkowski's avatar
Thomas Witkowski committed
493 494 495 496 497 498
	// 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) {
499
	      int colIndex = localDofMap.getMatIndex(k, it->first);
500
	      double value = (W[i] == mpiRank ? 1.0 : -1.0);
501
	      MatSetValue(mat_lagrange, index, colIndex, value, INSERT_VALUES);
502
	    }
Thomas Witkowski's avatar
Thomas Witkowski committed
503
	    index++;	      
504 505 506 507 508 509 510 511 512 513
	  }
	}
      }
    }

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


514
  void PetscSolverFeti::createSchurPrimalKsp(vector<const FiniteElemSpace*> &feSpaces)
515 516 517
  {
    FUNCNAME("PetscSolverFeti::createSchurPrimal()");

Thomas Witkowski's avatar
Thomas Witkowski committed
518
    if (schurPrimalSolver == 0) {
519 520
      MSG("Create iterative schur primal solver!\n");

521
      schurPrimalData.subSolver = subDomainSolver;
Thomas Witkowski's avatar
Thomas Witkowski committed
522
      
523
      VecCreateMPI(mpiComm, 
524
		   localDofMap.getRankDofs(), localDofMap.getOverallDofs(),
Thomas Witkowski's avatar
Thomas Witkowski committed
525
		   &(schurPrimalData.tmp_vec_b));
526
      VecCreateMPI(mpiComm, 
527
		   primalDofMap.getRankDofs(), primalDofMap.getOverallDofs(),
Thomas Witkowski's avatar
Thomas Witkowski committed
528 529
		   &(schurPrimalData.tmp_vec_primal));

530
      MatCreateShell(mpiComm,
531 532
		     primalDofMap.getRankDofs(), primalDofMap.getRankDofs(), 
		     primalDofMap.getOverallDofs(), primalDofMap.getOverallDofs(),
Thomas Witkowski's avatar
Thomas Witkowski committed
533 534 535 536 537
		     &schurPrimalData, 
		     &mat_schur_primal);
      MatShellSetOperation(mat_schur_primal, MATOP_MULT, 
			   (void(*)(void))petscMultMatSchurPrimal);
      
538
      KSPCreate(mpiComm, &ksp_schur_primal);
Thomas Witkowski's avatar
Thomas Witkowski committed
539
      KSPSetOperators(ksp_schur_primal, mat_schur_primal, mat_schur_primal, SAME_NONZERO_PATTERN);
540 541
      KSPSetOptionsPrefix(ksp_schur_primal, "schur_primal_");
      KSPSetType(ksp_schur_primal, KSPGMRES);
Thomas Witkowski's avatar
Thomas Witkowski committed
542 543
      KSPSetFromOptions(ksp_schur_primal);
    } else {
544 545
      MSG("Create direct schur primal solver!\n");

546 547
      double wtime = MPI::Wtime();

548 549
      int nRowsRankPrimal = primalDofMap.getRankDofs();
      int nRowsOverallPrimal = primalDofMap.getOverallDofs();
550 551
      int nRowsRankB = localDofMap.getRankDofs();
      int nRowsOverallB = localDofMap.getOverallDofs();
552

Thomas Witkowski's avatar
Thomas Witkowski committed
553
      Mat matBPi;
554
      MatCreateMPIAIJ(mpiComm,
Thomas Witkowski's avatar
Thomas Witkowski committed
555 556 557 558
		      nRowsRankB, nRowsRankPrimal, 
		      nRowsOverallB, nRowsOverallPrimal,
		      30, PETSC_NULL, 30, PETSC_NULL, &matBPi);
      Mat matPrimal;
559

560 561 562
      PetscInt nCols;
      const PetscInt *cols;
      const PetscScalar *values;
Thomas Witkowski's avatar
Thomas Witkowski committed
563

564 565
      map<int, vector<pair<int, double> > > mat_b_primal_cols;

566 567
      for (int i = 0; i < nRowsRankB; i++) {
	PetscInt row = localDofMap.getStartDofs() + i;
568
	MatGetRow(subDomainSolver->getMatIntCoarse(), row, &nCols, &cols, &values);
569 570 571 572 573

	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]));

574
	MatRestoreRow(subDomainSolver->getMatIntCoarse(), row, &nCols, &cols, &values);
575 576
      }

577 578
      TEST_EXIT(static_cast<int>(mat_b_primal_cols.size()) == 
		primalDofMap.getLocalDofs())
579
	("Should not happen!\n");
580

581 582 583
      for (map<int, vector<pair<int, double> > >::iterator it = mat_b_primal_cols.begin();
	   it != mat_b_primal_cols.end(); ++it) {
	Vec tmpVec;
584
	VecCreateSeq(PETSC_COMM_SELF, nRowsRankB, &tmpVec);
585 586 587 588 589 590 591 592

 	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);

593
	subDomainSolver->solve(tmpVec, tmpVec);
594

Thomas Witkowski's avatar
Thomas Witkowski committed
595
	PetscScalar *tmpValues;
596
	VecGetArray(tmpVec, &tmpValues);
597
	for (int i  = 0; i < nRowsRankB; i++)
Thomas Witkowski's avatar
Thomas Witkowski committed
598
	  MatSetValue(matBPi, 
599
		      localDofMap.getStartDofs() + i,
Thomas Witkowski's avatar
Thomas Witkowski committed
600 601 602
		      it->first,
		      tmpValues[i],
		      ADD_VALUES);
603 604 605 606 607
	VecRestoreArray(tmpVec, &tmpValues);

	VecDestroy(&tmpVec);
      }

Thomas Witkowski's avatar
Thomas Witkowski committed
608 609
      MatAssemblyBegin(matBPi, MAT_FINAL_ASSEMBLY);
      MatAssemblyEnd(matBPi, MAT_FINAL_ASSEMBLY);
610 611

      MatDuplicate(subDomainSolver->getMatCoarseCoarse(), MAT_COPY_VALUES, &mat_schur_primal);
612
      MatMatMult(subDomainSolver->getMatCoarseInt(), matBPi, MAT_INITIAL_MATRIX, PETSC_DEFAULT, &matPrimal);
613
      MatAXPY(mat_schur_primal, -1.0, matPrimal, DIFFERENT_NONZERO_PATTERN);
614 615

      MatDestroy(&matPrimal);
Thomas Witkowski's avatar
Thomas Witkowski committed
616
      MatDestroy(&matBPi);
617 618

      MatInfo minfo;
619
      MatGetInfo(mat_schur_primal, MAT_GLOBAL_SUM, &minfo);
620 621
      MSG("Schur primal matrix nnz = %f\n", minfo.nz_used);

622
      KSPCreate(mpiComm, &ksp_schur_primal);
623
      KSPSetOperators(ksp_schur_primal, mat_schur_primal, mat_schur_primal,
624
		      SAME_NONZERO_PATTERN);
625 626 627 628 629 630
      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);
631
      KSPSetFromOptions(ksp_schur_primal);
Thomas Witkowski's avatar
Thomas Witkowski committed
632

633 634
      MSG("Creating Schur primal matrix needed %.5f seconds.\n",
	  MPI::Wtime() - wtime);
Thomas Witkowski's avatar
Thomas Witkowski committed
635
    }
636 637 638 639 640 641 642
  }


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

643
    if (schurPrimalSolver == 0) {
644
      schurPrimalData.subSolver = NULL;
645

646 647 648
      VecDestroy(&schurPrimalData.tmp_vec_b);
      VecDestroy(&schurPrimalData.tmp_vec_primal);
    }
649 650 651
    
    MatDestroy(&mat_schur_primal);
    KSPDestroy(&ksp_schur_primal);
652 653 654
  }


655
  void PetscSolverFeti::createFetiKsp(vector<const FiniteElemSpace*> &feSpaces)
656 657 658
  {
    FUNCNAME("PetscSolverFeti::createFetiKsp()");

659 660
    // === Create FETI-DP solver object. ===

661
    fetiData.mat_lagrange = &mat_lagrange;
662
    fetiData.subSolver = subDomainSolver;
663
    fetiData.ksp_schur_primal = &ksp_schur_primal;
664

665
    VecCreateMPI(mpiComm, 
666
		 localDofMap.getRankDofs(), localDofMap.getOverallDofs(),
Thomas Witkowski's avatar
Thomas Witkowski committed
667
		 &(fetiData.tmp_vec_b));
668
    VecCreateMPI(mpiComm,
669
		 lagrangeMap.getRankDofs(), lagrangeMap.getOverallDofs(),
Thomas Witkowski's avatar
Thomas Witkowski committed
670
		 &(fetiData.tmp_vec_lagrange));
671
    VecCreateMPI(mpiComm, 
672
		 primalDofMap.getRankDofs(), primalDofMap.getOverallDofs(),
673
		 &(fetiData.tmp_vec_primal));
674

675
    MatCreateShell(mpiComm,
676 677
		   lagrangeMap.getRankDofs(), lagrangeMap.getRankDofs(),
		   lagrangeMap.getOverallDofs(), lagrangeMap.getOverallDofs(),
678
		   &fetiData, &mat_feti);
679 680 681
    MatShellSetOperation(mat_feti, MATOP_MULT, (void(*)(void))petscMultMatFeti);


682
    KSPCreate(mpiComm, &ksp_feti);
683
    KSPSetOperators(ksp_feti, mat_feti, mat_feti, SAME_NONZERO_PATTERN);
684 685 686
    KSPSetOptionsPrefix(ksp_feti, "feti_");
    KSPSetType(ksp_feti, KSPGMRES);
    KSPSetTolerances(ksp_feti, 0, 1e-8, 1e+3, 1000);
687
    KSPSetFromOptions(ksp_feti);
688 689


690
    // === Create FETI-DP preconditioner object. ===
691

692 693 694 695
    if (fetiPreconditioner != FETI_NONE) {
      MatDuplicate(mat_lagrange, MAT_COPY_VALUES, &mat_lagrange_scaled);
      MatScale(mat_lagrange_scaled, 0.5);
    }
696

697 698 699
    switch (fetiPreconditioner) {
    case FETI_DIRICHLET:           
      KSPCreate(PETSC_COMM_SELF, &ksp_interior);
Thomas Witkowski's avatar
Thomas Witkowski committed
700 701
      KSPSetOperators(ksp_interior, mat_interior_interior, mat_interior_interior, 
		      SAME_NONZERO_PATTERN);
702 703 704 705 706 707
      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);
708 709 710 711 712 713 714 715 716
      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;
      
717
      VecCreateMPI(mpiComm, 
718
		   localDofMap.getRankDofs(),localDofMap.getOverallDofs(),
719
		   &(fetiDirichletPreconData.tmp_vec_b));      
720 721 722 723 724 725
      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));
726 727 728 729 730 731 732 733 734 735 736 737 738

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

739 740 741 742 743 744 745 746 747 748 749
      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
750 751 752 753 754 755 756 757 758 759 760 761
      for (unsigned int i = 0; i < feSpaces.size(); i++) {
	map<DegreeOfFreedom, MultiIndex> &dualMap = 
	  dualDofMap[feSpaces[i]].getMap();
	for (map<DegreeOfFreedom, MultiIndex>::iterator it = dualMap.begin(); 
	     it != dualMap.end(); ++it) {
	  DegreeOfFreedom d = it->first;
	  int matIndexLocal = localDofMap.getLocalMatIndex(i, d);
	  int matIndexDual = dualDofMap.getLocalMatIndex(i, d);
	  fetiLumpedPreconData.localToDualMap[matIndexLocal] = matIndexDual;
	}
      }

762
      VecCreateMPI(mpiComm, 
763 764
		   localDofMap.getRankDofs(),
		   localDofMap.getOverallDofs(),
765
		   &(fetiLumpedPreconData.tmp_vec_b));
766 767 768 769
      MatGetVecs(mat_duals_duals, PETSC_NULL, 
		 &(fetiLumpedPreconData.tmp_vec_duals0));
      MatGetVecs(mat_duals_duals, PETSC_NULL, 
		 &(fetiLumpedPreconData.tmp_vec_duals1));
770 771 772 773 774 775

      KSPGetPC(ksp_feti, &precon_feti);
      PCSetType(precon_feti, PCSHELL);
      PCShellSetContext(precon_feti, static_cast<void*>(&fetiLumpedPreconData));
      PCShellSetApply(precon_feti, petscApplyFetiLumpedPrecon);
      
776 777
      break;
    default:
778 779
      break;
    }
780 781 782 783 784 785 786
  }
  

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

787 788
    // === Destroy FETI-DP solver object. ===

789
    fetiData.mat_lagrange = PETSC_NULL;
790
    fetiData.subSolver = NULL;
791
    fetiData.ksp_schur_primal = PETSC_NULL;
792

Thomas Witkowski's avatar
Thomas Witkowski committed
793 794
    VecDestroy(&fetiData.tmp_vec_b);
    VecDestroy(&fetiData.tmp_vec_lagrange);
795
    VecDestroy(&fetiData.tmp_vec_primal);
796 797
    MatDestroy(&mat_feti);
    KSPDestroy(&ksp_feti);
798 799


800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827
    // === 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;
828 829
    default:
      break;
830
    }
831 832 833 834 835 836 837 838 839
  }


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

840
    // === Get local part of the solution for B variables. ===
841 842 843 844 845

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


846 847
    // === Create scatter to get solutions of all primal nodes that are ===
    // === contained in rank's domain.                                  ===
848
    
Thomas Witkowski's avatar
Thomas Witkowski committed
849 850
    vector<const FiniteElemSpace*> feSpaces = getFeSpaces(&vec);

851
    vector<PetscInt> globalIsIndex, localIsIndex;
Thomas Witkowski's avatar
Thomas Witkowski committed
852 853
    globalIsIndex.reserve(primalDofMap.getLocalDofs());
    localIsIndex.reserve(primalDofMap.getLocalDofs());
854 855

    {
Thomas Witkowski's avatar
Thomas Witkowski committed
856 857 858 859 860 861 862 863
      int cnt = 0;
      for (unsigned int i = 0; i < feSpaces.size(); i++) {
	map<DegreeOfFreedom, MultiIndex>& dofMap = 
	  primalDofMap[feSpaces[i]].getMap();
	for (map<DegreeOfFreedom, MultiIndex>::iterator it = dofMap.begin();
	     it != dofMap.end(); ++it) {
	  globalIsIndex.push_back(primalDofMap.getMatIndex(i, it->first));
	  localIsIndex.push_back(cnt++);
864 865
	}
      }
866
      
Thomas Witkowski's avatar
Thomas Witkowski committed
867
      TEST_EXIT_DBG(cnt == primalDofMap.getLocalDofs())("Should not happen!\n");
868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892
    }
    
    IS globalIs, localIs;
    ISCreateGeneral(PETSC_COMM_SELF, 
		    globalIsIndex.size(), 
		    &(globalIsIndex[0]),
		    PETSC_USE_POINTER,
		    &globalIs);

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

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

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