PetscSolverFeti.cc 43.3 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13
//
// Software License for AMDiS
//
// Copyright (c) 2010 Dresden University of Technology 
// All rights reserved.
// Authors: Simon Vey, Thomas Witkowski et al.
//
// This file is part of AMDiS
//
// See also license.opensource.txt in the distribution.


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

namespace AMDiS {

  using namespace std;

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

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

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

    return 0;
  }


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

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

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

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

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

    return 0;
  }


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

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


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

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

87 88 89 90 91 92
    PetscScalar *local_b, *local_duals;
    VecGetArray(data->tmp_vec_b, &local_b);
    VecGetArray(data->tmp_vec_duals0, &local_duals);

    for (int i = nLocalB - nLocalDuals, j = 0; i < nLocalB; i++, j++)
      local_duals[j] = local_b[i];
93 94

    VecRestoreArray(data->tmp_vec_b, &local_b);
95
    VecRestoreArray(data->tmp_vec_duals0, &local_duals);
96 97


98
    // === K_DD - K_DI inv(K_II) K_ID ===
99

100
    MatMult(*(data->mat_duals_duals), data->tmp_vec_duals0, data->tmp_vec_duals1);
101

102
    MatMult(*(data->mat_interior_duals), data->tmp_vec_duals0, data->tmp_vec_interior);
103
    KSPSolve(*(data->ksp_interior), data->tmp_vec_interior, data->tmp_vec_interior);
104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137
    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);

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

    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);
138 139


140
    // === Restriction of the B nodes to the boundary nodes. ===
141

142 143 144 145
    int nLocalB;
    int nLocalDuals;
    VecGetLocalSize(data->tmp_vec_b, &nLocalB);
    VecGetLocalSize(data->tmp_vec_duals0, &nLocalDuals);
146

147
    PetscScalar *local_b, *local_duals;
148
    VecGetArray(data->tmp_vec_b, &local_b);
149
    VecGetArray(data->tmp_vec_duals0, &local_duals);
150

151 152
    for (int i = nLocalB - nLocalDuals, j = 0; i < nLocalB; i++, j++)
      local_duals[j] = local_b[i];
153 154

    VecRestoreArray(data->tmp_vec_b, &local_b);
155 156 157 158 159 160 161
    VecRestoreArray(data->tmp_vec_duals0, &local_duals);


    // === K_DD ===

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

162

163
    // === Prolongation from local dual nodes to B nodes.
164

165 166 167 168 169
    VecGetArray(data->tmp_vec_b, &local_b);
    VecGetArray(data->tmp_vec_duals1, &local_duals);

    for (int i = nLocalB - nLocalDuals, j = 0; i < nLocalB; i++, j++)
      local_b[i] = local_duals[j];
170

171 172
    VecRestoreArray(data->tmp_vec_b, &local_b);
    VecRestoreArray(data->tmp_vec_duals0, &local_duals);
173

174 175

    // Multiply with scaled Lagrange constraint matrix.
176 177 178 179 180 181
    MatMult(*(data->mat_lagrange_scaled), data->tmp_vec_b, y);

    return 0;
  }


182 183
  PetscSolverFeti::PetscSolverFeti()
    : PetscSolver(),
Thomas Witkowski's avatar
Thomas Witkowski committed
184
      schurPrimalSolver(0)
185 186 187 188 189 190
  {
    FUNCNAME("PetscSolverFeti::PetscSolverFeti()");

    string preconditionerName = "";
    Parameters::get("parallel->solver->precon", preconditionerName);
    if (preconditionerName == "" || preconditionerName == "none") {
191
      MSG("Create FETI-DP solver with no preconditioner!\n");
192 193
      fetiPreconditioner = FETI_NONE;
    } else if (preconditionerName == "dirichlet") {
194
      MSG("Create FETI-DP solver with Dirichlet preconditioner!\n");
195 196
      fetiPreconditioner = FETI_DIRICHLET;
    } else if (preconditionerName == "lumped") {
197
      MSG("Create FETI-DP solver with lumped preconditioner!\n");
198 199
      fetiPreconditioner = FETI_LUMPED;
    } else {
Thomas Witkowski's avatar
Thomas Witkowski committed
200 201
      ERROR_EXIT("Preconditioner \"%s\" not available!\n", 
		 preconditionerName.c_str());
202
    }
Thomas Witkowski's avatar
Thomas Witkowski committed
203 204 205 206 207

    Parameters::get("parallel->feti->schur primal solver", schurPrimalSolver);
    TEST_EXIT(schurPrimalSolver == 0 || schurPrimalSolver == 1)
      ("Wrong solver \"%d\"for the Schur primal complement!\n", 
       schurPrimalSolver);
208 209 210
  }


211
  void PetscSolverFeti::updateDofData()
212 213
  {
    FUNCNAME("PetscSolverFeti::updateDofData()");
214 215 216

    TEST_EXIT(meshDistributor->getFeSpace()->getBasisFcts()->getDegree() == 1)
      ("Works for linear basis functions only!\n");
217
   
218 219 220 221
    for (unsigned int i = 0; i < meshDistributor->getFeSpaces().size(); i++) {
      const FiniteElemSpace *feSpace = meshDistributor->getFeSpace(i);
      createPrimals(feSpace);      
      createDuals(feSpace);
222
      createLagrange(feSpace);      
223 224
      createIndexB(feSpace);
    }
Thomas Witkowski's avatar
Thomas Witkowski committed
225

226 227 228 229 230
    primalDofMap.setDofComm(meshDistributor->getSendDofs(),
			    meshDistributor->getRecvDofs());
    lagrangeMap.setDofComm(meshDistributor->getSendDofs(), 
			   meshDistributor->getRecvDofs());

Thomas Witkowski's avatar
Thomas Witkowski committed
231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248
    primalDofMap.update();
    dualDofMap.update();
    lagrangeMap.update();
    localDofMap.update();

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

      MSG("nRankPrimals = %d   nOverallPrimals = %d\n", 
	  primalDofMap[feSpace].nRankDofs, primalDofMap[feSpace].nOverallDofs);
      
      MSG("nRankDuals = %d   nOverallDuals = %d\n",
	  dualDofMap[feSpace].nRankDofs, 
	  dualDofMap[feSpace].nOverallDofs);

      MSG("nRankLagrange = %d  nOverallLagrange = %d\n",
	  lagrangeMap[feSpace].nRankDofs, 
	  lagrangeMap[feSpace].nOverallDofs);
249 250 251 252

      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
253
    }
254 255 256
  }


257
  void PetscSolverFeti::createPrimals(const FiniteElemSpace *feSpace)
Thomas Witkowski's avatar
Thomas Witkowski committed
258
  {
259
    FUNCNAME("PetscSolverFeti::createPrimals()");  
260

261 262 263
    // === Define all vertices on the interior boundaries of the macro mesh ===
    // === to be primal variables.                                          ===

264 265 266
    /// Set of DOF indices that are considered to be primal variables.
    
    DofIndexSet primals;
267
    DofContainerSet& vertices = 
268
      meshDistributor->getBoundaryDofInfo(feSpace).geoDofs[VERTEX];
269 270 271 272
    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);
273

274 275 276 277

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

278
    for (DofIndexSet::iterator it = primals.begin(); it != primals.end(); ++it)
279 280
      if (meshDistributor->getIsRankDof(feSpace, *it))
	primalDofMap[feSpace].insertRankDof(*it);
Thomas Witkowski's avatar
FETI-DP  
Thomas Witkowski committed
281 282
      else
  	primalDofMap[feSpace].insert(*it);
283 284 285
  }


286
  void PetscSolverFeti::createDuals(const FiniteElemSpace *feSpace)
287 288
  {
    FUNCNAME("PetscSolverFeti::createDuals()");
289

290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305
    // === 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()");

306 307
    // === Create for each dual node that is owned by the rank, the set ===
    // === of ranks that contain this node (denoted by W(x_j)).         ===
308 309 310

    boundaryDofRanks.clear();

311 312 313
    for (DofComm::Iterator it(meshDistributor->getSendDofs(), feSpace); 
	 !it.end(); it.nextRank())
      for (; !it.endDofIter(); it.nextDof()) {
314
	if (!isPrimal(feSpace, it.getDofIndex())) {
315 316
	  boundaryDofRanks[it.getDofIndex()].insert(mpiRank);
	  boundaryDofRanks[it.getDofIndex()].insert(it.getRank());
317 318
	}
      }
319
	
320 321 322 323

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

324
    StdMpi<vector<std::set<int> > > stdMpi(meshDistributor->getMpiComm());
325 326 327 328

    for (DofComm::Iterator it(meshDistributor->getSendDofs(), feSpace);
	 !it.end(); it.nextRank())
      for (; !it.endDofIter(); it.nextDof())
329
	if (!isPrimal(feSpace, it.getDofIndex()))
330
	  stdMpi.getSendData(it.getRank()).push_back(boundaryDofRanks[it.getDofIndex()]);
331 332 333

    stdMpi.updateSendDataSize();

334 335
    for (DofComm::Iterator it(meshDistributor->getRecvDofs(), feSpace); 
	 !it.end(); it.nextRank()) {
336
      bool recvFromRank = false;
337
      for (; !it.endDofIter(); it.nextDof()) {
338
	if (!isPrimal(feSpace, it.getDofIndex())) {
339 340 341
	  recvFromRank = true;
	  break;
	}
342
      }
343 344

      if (recvFromRank)
345
	stdMpi.recv(it.getRank());
346
    }
347

348 349
    stdMpi.startCommunication();

350 351
    for (DofComm::Iterator it(meshDistributor->getRecvDofs(), feSpace); 
	 !it.end(); it.nextRank()) {
352
      int i = 0;
353
      for (; !it.endDofIter(); it.nextDof())
354
	if (!isPrimal(feSpace, it.getDofIndex()))
355 356
	  boundaryDofRanks[it.getDofIndex()] = 
	    stdMpi.getRecvData(it.getRank())[i++];
357 358 359
    }


360 361 362
    // === Reserve for each dual node, on the rank that owns this node, the ===
    // === appropriate number of Lagrange constraints.                      ===

363
    int nRankLagrange = 0;
Thomas Witkowski's avatar
Thomas Witkowski committed
364 365
    map<DegreeOfFreedom, MultiIndex>& dualMap = dualDofMap[feSpace].getMap();
    for (map<DegreeOfFreedom, MultiIndex>::iterator it = dualMap.begin(); it != dualMap.end(); ++it) {
366
      if (meshDistributor->getIsRankDof(feSpace, it->first)) {
Thomas Witkowski's avatar
Thomas Witkowski committed
367
	lagrangeMap[feSpace].insertRankDof(it->first, nRankLagrange);
368
	int degree = boundaryDofRanks[it->first].size();
369
	nRankLagrange += (degree * (degree - 1)) / 2;
Thomas Witkowski's avatar
Thomas Witkowski committed
370 371
      } else {
	lagrangeMap[feSpace].insert(it->first);
372 373
      }
    }
374
    lagrangeMap[feSpace].nRankDofs = nRankLagrange;
375 376 377
  }


378
  void PetscSolverFeti::createIndexB(const FiniteElemSpace *feSpace)
379
  {
380
    FUNCNAME("PetscSolverFeti::createIndexB()");
381

382
    DOFAdmin* admin = feSpace->getAdmin();
383 384 385 386

    // === 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.                                 ===
387

388
    nLocalInterior[feSpace] = 0;    
389
    for (int i = 0; i < admin->getUsedSize(); i++) {
390
      if (admin->isDofFree(i) == false && 
391
	  isPrimal(feSpace, i) == false &&
392
	  dualDofMap[feSpace].isSet(i) == false) {
393 394
	localDofMap[feSpace].insertRankDof(i, nLocalInterior[feSpace]);
	nLocalInterior[feSpace]++;
395
      }
396
    }
Thomas Witkowski's avatar
FETI-DP  
Thomas Witkowski committed
397
    
398 399
    // === And finally, add the global indicies of all dual nodes. ===

Thomas Witkowski's avatar
Thomas Witkowski committed
400
    for (map<DegreeOfFreedom, MultiIndex>::iterator it = dualDofMap[feSpace].getMap().begin();
401
	 it != dualDofMap[feSpace].getMap().end(); ++it)
402
      localDofMap[feSpace].insertRankDof(it->first);
403 404 405
  }


406
  void PetscSolverFeti::createMatLagrange(vector<const FiniteElemSpace*> &feSpaces)
407 408 409
  {
    FUNCNAME("PetscSolverFeti::createMatLagrange()");

410 411
    // === Create distributed matrix for Lagrange constraints. ===

412
    MatCreateMPIAIJ(PETSC_COMM_WORLD,
413 414
		    lagrangeMap.getRankDofs(), localDofMap.getRankDofs(),
		    lagrangeMap.getOverallDofs(), localDofMap.getOverallDofs(),	
415 416 417
		    2, PETSC_NULL, 2, PETSC_NULL,
		    &mat_lagrange);

418 419 420 421 422 423 424
    // === 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
425 426 427
    for (unsigned int k = 0; k < feSpaces.size(); k++) {
      map<DegreeOfFreedom, MultiIndex> &dualMap = 
	dualDofMap[feSpaces[k]].getMap();
428

Thomas Witkowski's avatar
Thomas Witkowski committed
429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446
      for (map<DegreeOfFreedom, MultiIndex>::iterator it = dualMap.begin(); 
	   it != dualMap.end(); ++it) {
	TEST_EXIT_DBG(boundaryDofRanks.count(it->first))
	  ("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.
	vector<int> W(boundaryDofRanks[it->first].begin(), 
		      boundaryDofRanks[it->first].end());
	// 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) {
	      int colIndex = localDofMap.mapGlobal(it->first, k);
447
	      double value = (W[i] == mpiRank ? 1.0 : -1.0);
Thomas Witkowski's avatar
Thomas Witkowski committed
448
	      MatSetValue(mat_lagrange, index, colIndex, value, 
449
			  INSERT_VALUES);
450
	    }
Thomas Witkowski's avatar
Thomas Witkowski committed
451
	    index++;	      
452 453 454 455 456 457 458 459 460 461
	  }
	}
      }
    }

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


462
  void PetscSolverFeti::createSchurPrimalKsp(vector<const FiniteElemSpace*> &feSpaces)
463 464 465
  {
    FUNCNAME("PetscSolverFeti::createSchurPrimal()");

Thomas Witkowski's avatar
Thomas Witkowski committed
466
    if (schurPrimalSolver == 0) {
467 468
      MSG("Create iterative schur primal solver!\n");

Thomas Witkowski's avatar
Thomas Witkowski committed
469 470 471 472 473 474
      schurPrimalData.mat_primal_primal = &mat_primal_primal;
      schurPrimalData.mat_primal_b = &mat_primal_b;
      schurPrimalData.mat_b_primal = &mat_b_primal;
      schurPrimalData.fetiSolver = this;
      
      VecCreateMPI(PETSC_COMM_WORLD, 
475
		   localDofMap.getRankDofs(), localDofMap.getOverallDofs(),
Thomas Witkowski's avatar
Thomas Witkowski committed
476 477
		   &(schurPrimalData.tmp_vec_b));
      VecCreateMPI(PETSC_COMM_WORLD, 
478
		   primalDofMap.getRankDofs(), primalDofMap.getOverallDofs(),
Thomas Witkowski's avatar
Thomas Witkowski committed
479 480 481
		   &(schurPrimalData.tmp_vec_primal));

      MatCreateShell(PETSC_COMM_WORLD,
482 483
		     primalDofMap.getRankDofs(), primalDofMap.getRankDofs(), 
		     primalDofMap.getOverallDofs(), primalDofMap.getOverallDofs(),
Thomas Witkowski's avatar
Thomas Witkowski committed
484 485 486 487 488 489 490 491 492 493
		     &schurPrimalData, 
		     &mat_schur_primal);
      MatShellSetOperation(mat_schur_primal, MATOP_MULT, 
			   (void(*)(void))petscMultMatSchurPrimal);
      
      KSPCreate(PETSC_COMM_WORLD, &ksp_schur_primal);
      KSPSetOperators(ksp_schur_primal, mat_schur_primal, mat_schur_primal, SAME_NONZERO_PATTERN);
      KSPSetOptionsPrefix(ksp_schur_primal, "solver_sp_");
      KSPSetFromOptions(ksp_schur_primal);
    } else {
494 495
      MSG("Create direct schur primal solver!\n");

496 497 498 499
      TEST_EXIT(ksp_b)("No solver object for local problem!\n");

      double wtime = MPI::Wtime();

500 501
      int nRowsRankPrimal = primalDofMap.getRankDofs();
      int nRowsOverallPrimal = primalDofMap.getOverallDofs();
502 503
      int nRowsRankB = localDofMap.getRankDofs();
      int nRowsOverallB = localDofMap.getOverallDofs();
504

Thomas Witkowski's avatar
Thomas Witkowski committed
505 506 507 508 509 510
      Mat matBPi;
      MatCreateMPIAIJ(PETSC_COMM_WORLD,
		      nRowsRankB, nRowsRankPrimal, 
		      nRowsOverallB, nRowsOverallPrimal,
		      30, PETSC_NULL, 30, PETSC_NULL, &matBPi);
      Mat matPrimal;
511

512 513 514
      PetscInt nCols;
      const PetscInt *cols;
      const PetscScalar *values;
Thomas Witkowski's avatar
Thomas Witkowski committed
515

516 517
      map<int, vector<pair<int, double> > > mat_b_primal_cols;

518 519
      for (int i = 0; i < nRowsRankB; i++) {
	PetscInt row = localDofMap.getStartDofs() + i;
520 521 522 523 524 525
	MatGetRow(mat_b_primal, row, &nCols, &cols, &values);

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

Thomas Witkowski's avatar
Thomas Witkowski committed
526
	MatRestoreRow(mat_b_primal, row, &nCols, &cols, &values);
527 528 529 530 531
      }

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

532 533
      TEST_EXIT(static_cast<int>(mat_b_primal_cols.size()) == 
		primalDofMap.getLocalDofs())
534
	("Should not happen!\n");
535

536 537 538
      for (map<int, vector<pair<int, double> > >::iterator it = mat_b_primal_cols.begin();
	   it != mat_b_primal_cols.end(); ++it) {
	Vec tmpVec;
539
	VecCreateSeq(PETSC_COMM_SELF, nRowsRankB, &tmpVec);
540 541 542 543 544 545 546 547 548 549

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

	VecAssemblyBegin(tmpVec);
	VecAssemblyEnd(tmpVec);

       	KSPSolve(ksp_b, tmpVec, tmpVec);

Thomas Witkowski's avatar
Thomas Witkowski committed
550
	PetscScalar *tmpValues;
551
	VecGetArray(tmpVec, &tmpValues);
552
	for (int i  = 0; i < nRowsRankB; i++)
Thomas Witkowski's avatar
Thomas Witkowski committed
553
	  MatSetValue(matBPi, 
554
		      localDofMap.getStartDofs() + i,
Thomas Witkowski's avatar
Thomas Witkowski committed
555 556 557
		      it->first,
		      tmpValues[i],
		      ADD_VALUES);
558 559 560 561 562
	VecRestoreArray(tmpVec, &tmpValues);

	VecDestroy(&tmpVec);
      }

Thomas Witkowski's avatar
Thomas Witkowski committed
563 564 565
      MatAssemblyBegin(matBPi, MAT_FINAL_ASSEMBLY);
      MatAssemblyEnd(matBPi, MAT_FINAL_ASSEMBLY);
      MatMatMult(mat_primal_b, matBPi, MAT_INITIAL_MATRIX, PETSC_DEFAULT, &matPrimal);
566 567 568
      MatAXPY(mat_primal_primal, -1.0, matPrimal, DIFFERENT_NONZERO_PATTERN);

      MatDestroy(&matPrimal);
Thomas Witkowski's avatar
Thomas Witkowski committed
569
      MatDestroy(&matBPi);
570 571 572 573 574 575 576 577 578 579

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

      KSPCreate(PETSC_COMM_WORLD, &ksp_schur_primal);
      KSPSetOperators(ksp_schur_primal, mat_primal_primal, 
		      mat_primal_primal, SAME_NONZERO_PATTERN);
      KSPSetOptionsPrefix(ksp_schur_primal, "solver_sp_");
      KSPSetFromOptions(ksp_schur_primal);
Thomas Witkowski's avatar
Thomas Witkowski committed
580

581 582
      MSG("Creating Schur primal matrix needed %.5f seconds.\n",
	  MPI::Wtime() - wtime);
Thomas Witkowski's avatar
Thomas Witkowski committed
583
    }
584 585 586 587 588 589 590
  }


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

591 592 593 594 595
    if (schurPrimalSolver == 0) {
      schurPrimalData.mat_primal_primal = PETSC_NULL;
      schurPrimalData.mat_primal_b = PETSC_NULL;
      schurPrimalData.mat_b_primal = PETSC_NULL;
      schurPrimalData.fetiSolver = NULL;
596

597 598
      VecDestroy(&schurPrimalData.tmp_vec_b);
      VecDestroy(&schurPrimalData.tmp_vec_primal);
599

600 601 602 603 604
      MatDestroy(&mat_schur_primal);
      KSPDestroy(&ksp_schur_primal);
    } else {
      KSPDestroy(&ksp_schur_primal);
    }
605 606 607
  }


608
  void PetscSolverFeti::createFetiKsp(vector<const FiniteElemSpace*> &feSpaces)
609 610 611
  {
    FUNCNAME("PetscSolverFeti::createFetiKsp()");

612 613
    // === Create FETI-DP solver object. ===

614 615 616
    fetiData.mat_primal_b = &mat_primal_b;
    fetiData.mat_b_primal = &mat_b_primal;
    fetiData.mat_lagrange = &mat_lagrange;
Thomas Witkowski's avatar
Thomas Witkowski committed
617
    fetiData.fetiSolver = this;
618
    fetiData.ksp_schur_primal = &ksp_schur_primal;
619

620
    VecCreateMPI(PETSC_COMM_WORLD, 
621
		 localDofMap.getRankDofs(), localDofMap.getOverallDofs(),
Thomas Witkowski's avatar
Thomas Witkowski committed
622 623
		 &(fetiData.tmp_vec_b));
    VecCreateMPI(PETSC_COMM_WORLD,
624
		 lagrangeMap.getRankDofs(), lagrangeMap.getOverallDofs(),
Thomas Witkowski's avatar
Thomas Witkowski committed
625
		 &(fetiData.tmp_vec_lagrange));
626
    VecCreateMPI(PETSC_COMM_WORLD, 
627
		 primalDofMap.getRankDofs(), primalDofMap.getOverallDofs(),
628
		 &(fetiData.tmp_vec_primal));
629 630

    MatCreateShell(PETSC_COMM_WORLD,
631 632
		   lagrangeMap.getRankDofs(), lagrangeMap.getRankDofs(),
		   lagrangeMap.getOverallDofs(), lagrangeMap.getOverallDofs(),
633
		   &fetiData, &mat_feti);
634 635 636 637 638 639 640
    MatShellSetOperation(mat_feti, MATOP_MULT, (void(*)(void))petscMultMatFeti);


    KSPCreate(PETSC_COMM_WORLD, &ksp_feti);
    KSPSetOperators(ksp_feti, mat_feti, mat_feti, SAME_NONZERO_PATTERN);
    KSPSetOptionsPrefix(ksp_feti, "solver_feti_");
    KSPSetFromOptions(ksp_feti);
641 642


643
    // === Create FETI-DP preconditioner object. ===
644

645 646 647 648
    if (fetiPreconditioner != FETI_NONE) {
      MatDuplicate(mat_lagrange, MAT_COPY_VALUES, &mat_lagrange_scaled);
      MatScale(mat_lagrange_scaled, 0.5);
    }
649

650 651 652
    switch (fetiPreconditioner) {
    case FETI_DIRICHLET:           
      KSPCreate(PETSC_COMM_SELF, &ksp_interior);
Thomas Witkowski's avatar
Thomas Witkowski committed
653 654
      KSPSetOperators(ksp_interior, mat_interior_interior, mat_interior_interior, 
		      SAME_NONZERO_PATTERN);
655 656 657 658 659 660 661 662 663 664
      KSPSetOptionsPrefix(ksp_interior, "solver_interior_");
      KSPSetFromOptions(ksp_interior);
            
      fetiDirichletPreconData.mat_lagrange_scaled = &mat_lagrange_scaled;
      fetiDirichletPreconData.mat_interior_interior = &mat_interior_interior;
      fetiDirichletPreconData.mat_duals_duals = &mat_duals_duals;
      fetiDirichletPreconData.mat_interior_duals = &mat_interior_duals;
      fetiDirichletPreconData.mat_duals_interior = &mat_duals_interior;
      fetiDirichletPreconData.ksp_interior = &ksp_interior;
      
665
      VecCreateMPI(PETSC_COMM_WORLD, 
666
		   localDofMap.getRankDofs(),localDofMap.getOverallDofs(),
667
		   &(fetiDirichletPreconData.tmp_vec_b));      
668 669 670 671 672 673
      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));
674 675 676 677 678 679 680 681 682 683 684 685
      
      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;

686
      VecCreateMPI(PETSC_COMM_WORLD, 
687 688
		   localDofMap.getRankDofs(),
		   localDofMap.getOverallDofs(),
689
		   &(fetiLumpedPreconData.tmp_vec_b));
690 691 692 693
      MatGetVecs(mat_duals_duals, PETSC_NULL, 
		 &(fetiLumpedPreconData.tmp_vec_duals0));
      MatGetVecs(mat_duals_duals, PETSC_NULL, 
		 &(fetiLumpedPreconData.tmp_vec_duals1));
694 695 696 697 698 699

      KSPGetPC(ksp_feti, &precon_feti);
      PCSetType(precon_feti, PCSHELL);
      PCShellSetContext(precon_feti, static_cast<void*>(&fetiLumpedPreconData));
      PCShellSetApply(precon_feti, petscApplyFetiLumpedPrecon);
      
700 701
      break;
    default:
702 703
      break;
    }
704 705 706 707 708 709 710
  }
  

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

711 712
    // === Destroy FETI-DP solver object. ===

713 714 715
    fetiData.mat_primal_b = PETSC_NULL;
    fetiData.mat_b_primal = PETSC_NULL;
    fetiData.mat_lagrange = PETSC_NULL;
Thomas Witkowski's avatar
Thomas Witkowski committed
716
    fetiData.fetiSolver = NULL;
717
    fetiData.ksp_schur_primal = PETSC_NULL;
718

Thomas Witkowski's avatar
Thomas Witkowski committed
719 720
    VecDestroy(&fetiData.tmp_vec_b);
    VecDestroy(&fetiData.tmp_vec_lagrange);
721
    VecDestroy(&fetiData.tmp_vec_primal);
722 723
    MatDestroy(&mat_feti);
    KSPDestroy(&ksp_feti);
724 725


726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753
    // === 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;
754 755
    default:
      break;
756
    }
757 758 759 760 761 762 763 764 765
  }


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

766
    // === Get local part of the solution for B variables. ===
767 768 769 770 771

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


772 773
    // === Create scatter to get solutions of all primal nodes that are ===
    // === contained in rank's domain.                                  ===
774
    
Thomas Witkowski's avatar
Thomas Witkowski committed
775 776
    vector<const FiniteElemSpace*> feSpaces = getFeSpaces(&vec);

777
    vector<PetscInt> globalIsIndex, localIsIndex;
Thomas Witkowski's avatar
Thomas Witkowski committed
778 779
    globalIsIndex.reserve(primalDofMap.getLocalDofs());
    localIsIndex.reserve(primalDofMap.getLocalDofs());
780 781

    {
Thomas Witkowski's avatar
Thomas Witkowski committed
782 783 784 785 786 787 788 789
      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++);
790 791
	}
      }
Thomas Witkowski's avatar
Thomas Witkowski committed
792 793

      TEST_EXIT_DBG(cnt == primalDofMap.getLocalDofs())("Should not happen!\n");
794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818
    }
    
    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);

819 820 821
    ISDestroy(&globalIs);
    ISDestroy(&localIs);    
    VecScatterDestroy(&primalScatter);    
822 823 824 825

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

826
    // === And copy from PETSc local vectors to the DOF vectors. ===
827

Thomas Witkowski's avatar
Thomas Witkowski committed
828 829
    int cnt = 0;
    for (int i = 0; i < vec.getSize(); i++) {
830 831
      DOFVector<double>& dofVec = *(vec.getDOFVector(i));

Thomas Witkowski's avatar
Thomas Witkowski committed
832 833
      for (map<DegreeOfFreedom, MultiIndex>::iterator it = localDofMap[feSpaces[i]].getMap().begin();
	   it != localDofMap[feSpaces[i]].getMap().end(); ++it) {
834
	int petscIndex = localDofMap.mapLocal(it->first, i);
835 836 837
	dofVec[it->first] = localSolB[petscIndex];
      }

Thomas Witkowski's avatar
Thomas Witkowski committed
838 839 840
      for (map<DegreeOfFreedom, MultiIndex>::iterator it = primalDofMap[feSpaces[i]].getMap().begin();
	   it != primalDofMap[feSpaces[i]].getMap().end(); ++it)
	dofVec[it->first] = localSolPrimal[cnt++];
841 842 843 844
    }

    VecRestoreArray(vec_sol_b, &localSolB);
    VecRestoreArray(local_sol_primal, &localSolPrimal);
845
    VecDestroy(&local_sol_primal);
846 847 848
  }


849
  void PetscSolverFeti::fillPetscMatrix(Matrix<DOFMatrix*> *mat)
850 851
  {
    FUNCNAME("PetscSolverFeti::fillPetscMatrix()");   
852

853 854
    // === Create all sets and indices. ===

Thomas Witkowski's avatar
Thomas Witkowski committed
855 856
    vector<const FiniteElemSpace*> feSpaces = getFeSpaces(mat);

857 858 859 860
    primalDofMap.init(mpiComm, feSpaces, true, true);
    dualDofMap.init(mpiComm, feSpaces, false, false);
    lagrangeMap.init(mpiComm, feSpaces, true, true);
    localDofMap.init(mpiComm, feSpaces, false, false);
861

862 863
    updateDofData();

864 865
    // === Create matrices for the FETI-DP method. ===

866 867
    int nRowsRankB = localDofMap.getRankDofs();
    int nRowsOverallB = localDofMap.getOverallDofs();
868 869 870
    int nRowsRankPrimal = primalDofMap.getRankDofs();
    int nRowsOverallPrimal = primalDofMap.getOverallDofs();
    int nRowsDual = dualDofMap.getRankDofs();
871 872
    int nRowsInterior = nRowsRankB - nRowsDual;

Thomas Witkowski's avatar
Thomas Witkowski committed
873 874
    MatCreateSeqAIJ(PETSC_COMM_SELF, nRowsRankB, nRowsRankB, 30, PETSC_NULL,
		    &mat_b_b);
875 876

    MatCreateMPIAIJ(PETSC_COMM_WORLD,
877 878
		    nRowsRankPrimal, nRowsRankPrimal, 
		    nRowsOverallPrimal, nRowsOverallPrimal,
879
		    30, PETSC_NULL, 30, PETSC_NULL, &mat_primal_primal);
880 881

    MatCreateMPIAIJ(PETSC_COMM_WORLD,
882 883
		    nRowsRankB, nRowsRankPrimal, 
		    nRowsOverallB, nRowsOverallPrimal,
884
		    30, PETSC_NULL, 30, PETSC_NULL, &mat_b_primal);
885 886

    MatCreateMPIAIJ(PETSC_COMM_WORLD,
887 888
		    nRowsRankPrimal, nRowsRankB,
		    nRowsOverallPrimal, nRowsOverallB,
Thomas Witkowski's avatar
Thomas Witkowski committed
889
		    15, PETSC_NULL, 15, PETSC_NULL, &mat_primal_b);
890

891

892 893 894 895
    // === Create matrices for FETI-DP preconditioner. ===

    if (fetiPreconditioner != FETI_NONE)
      MatCreateSeqAIJ(PETSC_COMM_SELF,
896
		      nRowsDual, nRowsDual, 30, PETSC_NULL,
897 898 899 900
		      &mat_duals_duals);

    if (fetiPreconditioner == FETI_DIRICHLET) {
      MatCreateSeqAIJ(PETSC_COMM_SELF,
901
		      nRowsInterior, nRowsInterior, 30, PETSC_NULL,
902 903 904
		      &mat_interior_interior);
      
      MatCreateSeqAIJ(PETSC_COMM_SELF,
905
		      nRowsInterior, nRowsDual, 30, PETSC_NULL,
906 907 908
		      &mat_interior_duals);
      
      MatCreateSeqAIJ(PETSC_COMM_SELF,
909
		      nRowsDual, nRowsInterior, 30, PETSC_NULL,
910 911
		      &mat_duals_interior);
    }
912

913 914
    
    // === Prepare traverse of sequentially created matrices. ===
915 916 917 918 919 920 921 922 923 924 925 926 927 928 929

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

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

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

930 931 932 933 934 935 936
    vector<int> colsLocal, colsLocalOther;
    vector<double> valuesLocal, valuesLocalOther;
    colsLocal.reserve(300);
    colsLocalOther.reserve(300);
    valuesLocal.reserve(300);
    valuesLocalOther.reserve(300);

937 938 939 940

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

Thomas Witkowski's avatar
Thomas Witkowski committed
941
    int nComponents = mat->getSize();
942 943 944 945 946 947 948 949 950 951 952
    for (int i = 0; i < nComponents; i++) {
      for (int j = 0; j < nComponents; j++) {
	if (!(*mat)[i][j])
	  continue;

	traits::col<Matrix>::type col((*mat)[i][j]->getBaseMatrix());
	traits::const_value<Matrix>::type value((*mat)[i][j]->getBaseMatrix());
	
	// Traverse all rows.
	for (cursor_type cursor = begin<row>((*mat)[i][j]->getBaseMatrix()), 
	       cend = end<row>((*mat)[i][j]->getBaseMatrix()); cursor != cend; ++cursor) {
953

Thomas Witkowski's avatar
Thomas Witkowski committed
954
	  bool rowPrimal = isPrimal(feSpaces[i], *cursor);
955
  
956 957
	  cols.clear();
	  colsOther.clear();
958
	  values.clear();