PetscSolverGlobalMatrix.cc 30.5 KB
Newer Older
Thomas Witkowski's avatar
Thomas Witkowski committed
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"
Thomas Witkowski's avatar
Thomas Witkowski committed
14 15 16 17 18 19
#include "parallel/PetscSolverGlobalMatrix.h"
#include "parallel/StdMpi.h"
#include "parallel/MpiHelper.h"

namespace AMDiS {

20
  void PetscSolverGlobalMatrix::fillPetscMatrix(Matrix<DOFMatrix*> *mat)
Thomas Witkowski's avatar
Thomas Witkowski committed
21 22 23
  {
    FUNCNAME("PetscSolverGlobalMatrix::fillPetscMatrix()");

24
    if (coarseSpaceMap != NULL) {
Thomas Witkowski's avatar
Thomas Witkowski committed
25
      updateSubdomainData();
26 27 28 29
      fillPetscMatrixWithCoarseSpace(mat);
      return;
    }

Thomas Witkowski's avatar
Thomas Witkowski committed
30
    TEST_EXIT_DBG(meshDistributor)("No mesh distributor object defined!\n");
31
    TEST_EXIT_DBG(interiorMap)("No parallel mapping object defined!\n");
Thomas Witkowski's avatar
Thomas Witkowski committed
32
    TEST_EXIT_DBG(mat)("No DOF matrix defined!\n");
33
    
Thomas Witkowski's avatar
Thomas Witkowski committed
34
    double wtime = MPI::Wtime();
35

36 37 38
    
    // === Check if mesh was changed and, in this case, recompute matrix ===
    // === nnz structure and matrix indices.                             ===
39

Thomas Witkowski's avatar
Thomas Witkowski committed
40
    int recvAllValues = 0;
41 42
    int sendValue = 
      static_cast<int>(meshDistributor->getLastMeshChangeIndex() != lastMeshNnz);
43
    mpiCommGlobal.Allreduce(&sendValue, &recvAllValues, 1, MPI_INT, MPI_SUM);
Thomas Witkowski's avatar
Thomas Witkowski committed
44

45
    if (!d_nnz || recvAllValues != 0 || alwaysCreateNnzStructure) {
46
      vector<const FiniteElemSpace*> feSpaces = getFeSpaces(mat);
47 48
      interiorMap->setComputeMatIndex(true, true);
      interiorMap->update(feSpaces);
49

Thomas Witkowski's avatar
Thomas Witkowski committed
50 51 52 53 54 55 56
      if (d_nnz) {
	delete [] d_nnz;
	d_nnz = NULL;
	delete [] o_nnz;
	o_nnz = NULL;
      }

Thomas Witkowski's avatar
Thomas Witkowski committed
57
      updateSubdomainData();
Thomas Witkowski's avatar
Thomas Witkowski committed
58 59 60 61 62
      createPetscNnzStructure(mat);
      lastMeshNnz = meshDistributor->getLastMeshChangeIndex();
    }


63 64
    // === Create PETSc vector (solution and a temporary vector). ===

65 66
    int nRankRows = interiorMap->getRankDofs();
    int nOverallRows = interiorMap->getOverallDofs();
67

68
    VecCreateMPI(mpiCommGlobal, nRankRows, nOverallRows, &petscSolVec);
69 70


Thomas Witkowski's avatar
Thomas Witkowski committed
71 72
    // === Create PETSc matrix with the computed nnz data structure. ===

73
    MatCreateMPIAIJ(mpiCommGlobal, nRankRows, nRankRows, 
74
 		    nOverallRows, nOverallRows,
75
		    0, d_nnz, 0, o_nnz, &matIntInt);
76

Thomas Witkowski's avatar
Thomas Witkowski committed
77 78 79 80 81 82
#if (DEBUG != 0)
    MSG("Fill petsc matrix 1 needed %.5f seconds\n", MPI::Wtime() - wtime);
#endif

#if (DEBUG != 0)
    int a, b;
83 84 85
    MatGetOwnershipRange(matIntInt, &a, &b);
    TEST_EXIT(a == interiorMap->getStartDofs())("Wrong matrix ownership range!\n");
    TEST_EXIT(b == interiorMap->getStartDofs() + nRankRows)
Thomas Witkowski's avatar
Thomas Witkowski committed
86 87 88 89 90 91
      ("Wrong matrix ownership range!\n");
#endif


    // === Transfer values from DOF matrices to the PETSc matrix. === 

92
    int nComponents = mat->getNumRows();
Thomas Witkowski's avatar
Thomas Witkowski committed
93 94 95
    for (int i = 0; i < nComponents; i++)
      for (int j = 0; j < nComponents; j++)
	if ((*mat)[i][j])
96
	  setDofMatrix((*mat)[i][j], i, j);
Thomas Witkowski's avatar
Thomas Witkowski committed
97 98 99 100 101

#if (DEBUG != 0)
    MSG("Fill petsc matrix 2 needed %.5f seconds\n", MPI::Wtime() - wtime);
#endif

102 103
    MatAssemblyBegin(matIntInt, MAT_FINAL_ASSEMBLY);
    MatAssemblyEnd(matIntInt, MAT_FINAL_ASSEMBLY);
Thomas Witkowski's avatar
Thomas Witkowski committed
104

105
    // === Init PETSc solver. ===
106 107 108 109 110 111 112 113
    KSPCreate(mpiCommGlobal, &kspInterior);
    KSPGetPC(kspInterior, &pcInterior);
    KSPSetOperators(kspInterior, matIntInt, matIntInt, SAME_NONZERO_PATTERN); 
    KSPSetTolerances(kspInterior, 0.0, 1e-8, PETSC_DEFAULT, PETSC_DEFAULT);
    KSPSetType(kspInterior, KSPBCGS);
    KSPSetOptionsPrefix(kspInterior, kspPrefix.c_str());
    KSPSetFromOptions(kspInterior);
    PCSetFromOptions(pcInterior);
114 115 116

    // Do not delete the solution vector, use it for the initial guess.
    if (!zeroStartVector)
117
      KSPSetInitialGuessNonzero(kspInterior, PETSC_TRUE);
118

119 120 121 122
    MSG("Fill petsc matrix needed %.5f seconds\n", MPI::Wtime() - wtime);
  }


123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315
  void PetscSolverGlobalMatrix::fillPetscMatrixWithCoarseSpace(Matrix<DOFMatrix*> *mat)
  {
    FUNCNAME("PetscSolverGlobalMatrix::fillPetscMatrixWithCoarseSpace()");
    
    vector<const FiniteElemSpace*> feSpaces = getFeSpaces(mat);

    int nRowsRankInterior = interiorMap->getRankDofs();
    int nRowsOverallInterior = interiorMap->getOverallDofs();

    if (subdomainLevel == 0) {
      MatCreateSeqAIJ(mpiCommLocal, nRowsRankInterior, nRowsRankInterior,
		      60, PETSC_NULL, &matIntInt);
    } else {
      MatCreateMPIAIJ(mpiCommLocal, 
		      nRowsRankInterior, nRowsRankInterior,
		      nRowsOverallInterior, nRowsOverallInterior,
		      60, PETSC_NULL, 60, PETSC_NULL, &matIntInt);
    }

    if (coarseSpaceMap) {
      int nRowsRankCoarse = coarseSpaceMap->getRankDofs();
      int nRowsOverallCoarse = coarseSpaceMap->getOverallDofs();

      MatCreateMPIAIJ(mpiCommGlobal,
		      nRowsRankCoarse, nRowsRankCoarse,
		      nRowsOverallCoarse, nRowsOverallCoarse,
		      60, PETSC_NULL, 60, PETSC_NULL, &matCoarseCoarse);
      
      MatCreateMPIAIJ(mpiCommGlobal,
		      nRowsRankCoarse, nRowsRankInterior,
		      nRowsOverallCoarse, nGlobalOverallInterior,
		      60, PETSC_NULL, 60, PETSC_NULL, &matCoarseInt);
      
      MatCreateMPIAIJ(mpiCommGlobal,
		      nRowsRankInterior, nRowsRankCoarse,
		      nGlobalOverallInterior, nRowsOverallCoarse,
		      60, PETSC_NULL, 60, PETSC_NULL, &matIntCoarse);
    }

    // === Prepare traverse of sequentially created matrices. ===

    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);

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

    int nComponents = mat->getSize();
    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) {

	  bool rowPrimal = isCoarseSpace(feSpaces[i], *cursor);
  
	  cols.clear();
	  colsOther.clear();
	  values.clear();	  
	  valuesOther.clear();

	  // Traverse all columns.
	  for (icursor_type icursor = begin<nz>(cursor), icend = end<nz>(cursor); 
	       icursor != icend; ++icursor) {

	    bool colPrimal = isCoarseSpace(feSpaces[j], col(*icursor));

	    if (colPrimal) {
	      if (rowPrimal) {
		cols.push_back(col(*icursor));
		values.push_back(value(*icursor));
	      } else {
		colsOther.push_back(col(*icursor));
		valuesOther.push_back(value(*icursor));
	      }
	    } else {
	      if (rowPrimal) {
		colsOther.push_back(col(*icursor));
		valuesOther.push_back(value(*icursor));
	      } else {
		cols.push_back(col(*icursor));
		values.push_back(value(*icursor));
	      }
	    }
	  }  // for each nnz in row


	  // === Set matrix values. ===

	  if (rowPrimal) {
	    int rowIndex = coarseSpaceMap->getMatIndex(i, *cursor);
	    for (unsigned int k = 0; k < cols.size(); k++)
	      cols[k] = coarseSpaceMap->getMatIndex(j, cols[k]);

	    MatSetValues(matCoarseCoarse, 1, &rowIndex, cols.size(),
			 &(cols[0]), &(values[0]), ADD_VALUES);

	    if (colsOther.size()) {
	      if (subdomainLevel == 0) {
		for (unsigned int k = 0; k < colsOther.size(); k++)
		  colsOther[k] = interiorMap->getMatIndex(j, colsOther[k]);
	      } else {
		for (unsigned int k = 0; k < colsOther.size(); k++)
		  colsOther[k] = 
		    interiorMap->getMatIndex(j, colsOther[k]) + rStartInterior;
	      }
 	      
	      MatSetValues(matCoarseInt, 1, &rowIndex, colsOther.size(),
 			   &(colsOther[0]), &(valuesOther[0]), ADD_VALUES);
	    }
	  } else {
	    int localRowIndex = 
	      (subdomainLevel == 0 ? interiorMap->getLocalMatIndex(i, *cursor) :
	       interiorMap->getMatIndex(i, *cursor));

	    for (unsigned int k = 0; k < cols.size(); k++) {
	      if (subdomainLevel == 0)
		cols[k] = interiorMap->getLocalMatIndex(j, cols[k]);
	      else
		cols[k] = interiorMap->getMatIndex(j, cols[k]);
	    }
	    
  	    MatSetValues(matIntInt, 1, &localRowIndex, cols.size(),
  			 &(cols[0]), &(values[0]), ADD_VALUES);

	    if (colsOther.size()) {
	      int globalRowIndex = interiorMap->getMatIndex(i, *cursor);

	      if (subdomainLevel != 0)
		globalRowIndex += rStartInterior;

	      for (unsigned int k = 0; k < colsOther.size(); k++)
		colsOther[k] = coarseSpaceMap->getMatIndex(j, colsOther[k]);

  	      MatSetValues(matIntCoarse, 1, &globalRowIndex, colsOther.size(),
  			   &(colsOther[0]), &(valuesOther[0]), ADD_VALUES);
	    }
	  }
	} 
      }
    }

    // === Start global assembly procedure. ===

    MatAssemblyBegin(matIntInt, MAT_FINAL_ASSEMBLY);
    MatAssemblyEnd(matIntInt, MAT_FINAL_ASSEMBLY);

    if (coarseSpaceMap) {
      MatAssemblyBegin(matCoarseCoarse, MAT_FINAL_ASSEMBLY);
      MatAssemblyEnd(matCoarseCoarse, MAT_FINAL_ASSEMBLY);
      
      MatAssemblyBegin(matIntCoarse, MAT_FINAL_ASSEMBLY);
      MatAssemblyEnd(matIntCoarse, MAT_FINAL_ASSEMBLY);
      
      MatAssemblyBegin(matCoarseInt, MAT_FINAL_ASSEMBLY);
      MatAssemblyEnd(matCoarseInt, MAT_FINAL_ASSEMBLY);
    }


    // === Create solver for the non primal (thus local) variables. ===

    KSPCreate(mpiCommLocal, &kspInterior);
    KSPSetOperators(kspInterior, matIntInt, matIntInt, SAME_NONZERO_PATTERN);
    KSPSetOptionsPrefix(kspInterior, "interior_");
    KSPSetType(kspInterior, KSPPREONLY);
    PC pcInterior;
    KSPGetPC(kspInterior, &pcInterior);
    PCSetType(pcInterior, PCLU);
    if (subdomainLevel == 0)
      PCFactorSetMatSolverPackage(pcInterior, MATSOLVERUMFPACK);
    else
      PCFactorSetMatSolverPackage(pcInterior, MATSOLVERMUMPS);
    KSPSetFromOptions(kspInterior);  
  }


316 317 318
  void PetscSolverGlobalMatrix::fillPetscRhs(SystemVector *vec)
  {
    FUNCNAME("PetscSolverGlobalMatrix::fillPetscRhs()");
Thomas Witkowski's avatar
Thomas Witkowski committed
319

Thomas Witkowski's avatar
Thomas Witkowski committed
320 321 322 323
    VecCreateMPI(mpiCommGlobal, 
		 interiorMap->getRankDofs(), 
		 nGlobalOverallInterior,
		 &rhsInterior);
324

Thomas Witkowski's avatar
Thomas Witkowski committed
325 326 327 328 329 330
    if (coarseSpaceMap) 
      VecCreateMPI(mpiCommGlobal, 
		   coarseSpaceMap->getRankDofs(), 
		   coarseSpaceMap->getOverallDofs(),
		   &rhsCoarseSpace);
    
331
    TEST_EXIT_DBG(vec)("No DOF vector defined!\n");
332
    TEST_EXIT_DBG(interiorMap)("No parallel DOF map defined!\n");
Thomas Witkowski's avatar
Thomas Witkowski committed
333
    
334

Thomas Witkowski's avatar
Thomas Witkowski committed
335 336 337 338 339 340 341
    if (coarseSpaceMap) {
      fillPetscRhsWithCoarseSpace(vec);
    } else {
      // === Transfer values from DOF vector to the PETSc vector. === 
      for (int i = 0; i < vec->getSize(); i++)
	setDofVector(rhsInterior, vec->getDOFVector(i), i);
    }
Thomas Witkowski's avatar
Thomas Witkowski committed
342

343 344
    VecAssemblyBegin(rhsInterior);
    VecAssemblyEnd(rhsInterior);
345

Thomas Witkowski's avatar
Thomas Witkowski committed
346 347 348 349 350
    if (coarseSpaceMap) {
      VecAssemblyBegin(rhsCoarseSpace);
      VecAssemblyEnd(rhsCoarseSpace);
    }

351
    if (removeRhsNullSpace) {
Thomas Witkowski's avatar
Thomas Witkowski committed
352 353
      TEST_EXIT_DBG(coarseSpaceMap == NULL)("Not supported!\n");

354
      MSG("Remove constant null space from the RHS!\n");
355
      MatNullSpace sp;
356 357
      MatNullSpaceCreate(mpiCommGlobal, PETSC_TRUE, 0, PETSC_NULL, &sp);
      MatNullSpaceRemove(sp, rhsInterior, PETSC_NULL);
358 359
      MatNullSpaceDestroy(&sp);
    }
Thomas Witkowski's avatar
Thomas Witkowski committed
360 361 362
  }


363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383
  void PetscSolverGlobalMatrix::fillPetscRhsWithCoarseSpace(SystemVector *vec)
  {
    FUNCNAME("SubDomainSolver::fillPetscRhs()");

    for (int i = 0; i < vec->getSize(); i++) {
      const FiniteElemSpace *feSpace = vec->getDOFVector(i)->getFeSpace();
      DOFVector<double>::Iterator dofIt(vec->getDOFVector(i), USED_DOFS);
      for (dofIt.reset(); !dofIt.end(); ++dofIt) {
	int index = dofIt.getDOFIndex();
	if (isCoarseSpace(feSpace, index)) {	  
	  index = coarseSpaceMap->getMatIndex(i, index);
	  VecSetValue(rhsCoarseSpace, index, *dofIt, ADD_VALUES);
	} else {
	  index = interiorMap->getMatIndex(i, index) + rStartInterior;
	  VecSetValue(rhsInterior, index, *dofIt, ADD_VALUES);
	}      
      }
    }
  }


384 385
  void PetscSolverGlobalMatrix::solvePetscMatrix(SystemVector &vec, 
						 AdaptInfo *adaptInfo)
Thomas Witkowski's avatar
Thomas Witkowski committed
386 387 388 389 390 391 392 393 394 395
  {
    FUNCNAME("PetscSolverGlobalMatrix::solvePetscMatrix()");

    int nComponents = vec.getSize();

    // === Set old solution to be initiual guess for PETSc solver. ===
    if (!zeroStartVector) {
      VecSet(petscSolVec, 0.0);
      
      for (int i = 0; i < nComponents; i++)
396
	setDofVector(petscSolVec, vec.getDOFVector(i), i, true);
Thomas Witkowski's avatar
Thomas Witkowski committed
397 398 399 400 401 402
      
      VecAssemblyBegin(petscSolVec);
      VecAssemblyEnd(petscSolVec);
    }

    // PETSc.
403
    solve(rhsInterior, petscSolVec);
Thomas Witkowski's avatar
Thomas Witkowski committed
404 405 406 407 408

    // === Transfere values from PETSc's solution vectors to the DOF vectors. ===
    PetscScalar *vecPointer;
    VecGetArray(petscSolVec, &vecPointer);

409
    int c = 0;
Thomas Witkowski's avatar
Thomas Witkowski committed
410
    for (int i = 0; i < nComponents; i++) {
411
      DOFVector<double> &dv = *(vec.getDOFVector(i));
412
      DofMap& d = (*interiorMap)[dv.getFeSpace()].getMap();
413 414 415
      for (DofMap::iterator it = d.begin(); it != d.end(); ++it)
	if (it->second.local != -1)
	  dv[it->first] = vecPointer[c++];
Thomas Witkowski's avatar
Thomas Witkowski committed
416 417 418 419 420 421 422 423
    }

    VecRestoreArray(petscSolVec, &vecPointer);


    // === Synchronize DOFs at common DOFs, i.e., DOFs that correspond to ===
    // === more than one partition.                                       ===
    meshDistributor->synchVector(vec);
424 425 426
  }


427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464
  void PetscSolverGlobalMatrix::solveGlobal(Vec &rhs, Vec &sol)
  {
    FUNCNAME("PetscSolverGlobalMatrix::solveGlobal()");

    Vec tmp;
    if (mpiCommLocal.Get_size() == 1)
      VecCreateSeq(mpiCommLocal, interiorMap->getRankDofs(), &tmp);
    else
      VecCreateMPI(mpiCommLocal,
		   interiorMap->getRankDofs(),
		   interiorMap->getOverallDofs(),
		   &tmp);

    PetscScalar *tmpValues, *rhsValues;
    VecGetArray(tmp, &tmpValues);
    VecGetArray(rhs, &rhsValues);

    for (int i = 0; i < interiorMap->getRankDofs(); i++)
      tmpValues[i] = rhsValues[i];

    VecRestoreArray(rhs, &rhsValues);
    VecRestoreArray(tmp, &tmpValues);

    KSPSolve(kspInterior, tmp, tmp);

    VecGetArray(tmp, &tmpValues);
    VecGetArray(sol, &rhsValues);

    for (int i = 0; i < interiorMap->getRankDofs(); i++) 
      rhsValues[i] = tmpValues[i];

    VecRestoreArray(sol, &rhsValues);
    VecRestoreArray(tmp, &tmpValues);

    VecDestroy(&tmp);
  }


465 466 467 468
  void PetscSolverGlobalMatrix::destroyMatrixData()
  {
    FUNCNAME("PetscSolverGlobalMatrix::destroyMatrixData()");

469 470
    MatDestroy(&matIntInt);
    KSPDestroy(&kspInterior);
Thomas Witkowski's avatar
Thomas Witkowski committed
471 472 473 474

    if (petscSolVec != PETSC_NULL) 
      VecDestroy(&petscSolVec);
    petscSolVec = PETSC_NULL;
Thomas Witkowski's avatar
Thomas Witkowski committed
475 476 477
  }


478 479 480 481
  void PetscSolverGlobalMatrix::destroyVectorData()
  {
    FUNCNAME("PetscSolverGlobalMatrix::destroyVectorData()");

482
    VecDestroy(&rhsInterior);
483 484 485
  }


486 487
  void PetscSolverGlobalMatrix::setDofMatrix(DOFMatrix* mat,
					     int nRowMat, int nColMat)
Thomas Witkowski's avatar
Thomas Witkowski committed
488 489 490 491 492 493
  {
    FUNCNAME("PetscSolverGlobalMatrix::setDofMatrix()");

    TEST_EXIT(mat)("No DOFMatrix!\n");

    using mtl::tag::row; using mtl::tag::nz; using mtl::begin; using mtl::end;
494
    namespace traits = mtl::traits;
Thomas Witkowski's avatar
Thomas Witkowski committed
495 496 497 498 499 500 501 502 503 504 505 506 507 508 509
    typedef DOFMatrix::base_matrix_type Matrix;

    traits::col<Matrix>::type col(mat->getBaseMatrix());
    traits::const_value<Matrix>::type value(mat->getBaseMatrix());

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

    vector<int> cols;
    vector<double> values;
    cols.reserve(300);
    values.reserve(300);
    
    vector<int> globalCols;

510 511 512
    // Get periodic mapping object
    PeriodicMap &perMap = meshDistributor->getPeriodicMap();

Thomas Witkowski's avatar
Thomas Witkowski committed
513 514 515 516 517 518
    // === Traverse all rows of the dof matrix and insert row wise the values ===
    // === to the PETSc matrix.                                               ===

    for (cursor_type cursor = begin<row>(mat->getBaseMatrix()), 
	   cend = end<row>(mat->getBaseMatrix()); cursor != cend; ++cursor) {

519 520 521
      const FiniteElemSpace *rowFe = mat->getRowFeSpace();
      const FiniteElemSpace *colFe = mat->getColFeSpace();

Thomas Witkowski's avatar
Thomas Witkowski committed
522
      // Global index of the current row DOF.
523
      int globalRowDof = (*interiorMap)[rowFe][*cursor].global;
524

Thomas Witkowski's avatar
Thomas Witkowski committed
525
      // Test if the current row DOF is a periodic DOF.
526
      bool periodicRow = perMap.isPeriodic(rowFe, globalRowDof);
527

Thomas Witkowski's avatar
Thomas Witkowski committed
528 529 530
      if (!periodicRow) {
	// === Row DOF index is not periodic. ===

531
	// Get PETSc's mat row index.
532
	int rowIndex = interiorMap->getMatIndex(nRowMat, globalRowDof);
Thomas Witkowski's avatar
Thomas Witkowski committed
533 534 535 536 537 538 539 540

	cols.clear();
	values.clear();

	for (icursor_type icursor = begin<nz>(cursor), icend = end<nz>(cursor); 
	     icursor != icend; ++icursor) {

	  // Global index of the current column index.
541
	  int globalColDof = (*interiorMap)[colFe][col(*icursor)].global;
Thomas Witkowski's avatar
Thomas Witkowski committed
542
	  // Test if the current col dof is a periodic dof.
543
	  bool periodicCol = perMap.isPeriodic(colFe, globalColDof);
544
	  // Get PETSc's mat col index.
545
	  int colIndex = interiorMap->getMatIndex(nColMat, globalColDof);
Thomas Witkowski's avatar
Thomas Witkowski committed
546 547 548 549 550 551 552 553 554 555 556 557 558 559

	  // Ignore all zero entries, expect it is a diagonal entry.
 	  if (value(*icursor) == 0.0 && rowIndex != colIndex)
 	    continue;

	  if (!periodicCol) {
	    // Calculate the exact position of the column index in the PETSc matrix.
	    cols.push_back(colIndex);
	    values.push_back(value(*icursor));
	  } else {
	    // === Row index is not periodic, but column index is. ===

	    // Create set of all periodic associations of the column index.
	    std::set<int> perAsc;
560
	    std::set<int>& perColAsc = perMap.getAssociations(colFe, globalColDof);
Thomas Witkowski's avatar
Thomas Witkowski committed
561 562
	    for (std::set<int>::iterator it = perColAsc.begin(); 
		 it != perColAsc.end(); ++it)
563
	      if (meshDistributor->getElementObjectDb().isValidPeriodicType(*it))
Thomas Witkowski's avatar
Thomas Witkowski committed
564 565 566 567 568 569 570 571 572 573 574 575 576 577
		perAsc.insert(*it);
    
	    // Scale value to the number of periodic associations of the column index.
	    double scaledValue = 
	      value(*icursor) * pow(0.5, static_cast<double>(perAsc.size()));

	    
	    // === Create set of all matrix column indices due to the periodic ===
	    // === associations of the column DOF index.                       ===

	    vector<int> newCols;

	    // First, add the original matrix index.
	    newCols.push_back(globalColDof);
578

Thomas Witkowski's avatar
Thomas Witkowski committed
579 580 581 582 583 584
	    // And add all periodic matrix indices.
	    for (std::set<int>::iterator it = perAsc.begin(); 
		 it != perAsc.end(); ++it) {
	      int nCols = static_cast<int>(newCols.size());

	      for (int i = 0; i < nCols; i++) {
585
 		TEST_EXIT_DBG(perMap.isPeriodic(colFe, *it, newCols[i]))
Thomas Witkowski's avatar
Thomas Witkowski committed
586 587 588
 		  ("Wrong periodic DOF associations at boundary %d with DOF %d!\n",
		   *it, newCols[i]);

589
		newCols.push_back(perMap.map(colFe, *it, newCols[i]));
Thomas Witkowski's avatar
Thomas Witkowski committed
590 591 592 593
	      }
	    }

	    for (unsigned int i = 0; i < newCols.size(); i++) {
594
	      cols.push_back(interiorMap->getMatIndex(nColMat, newCols[i]));
Thomas Witkowski's avatar
Thomas Witkowski committed
595 596 597 598 599
	      values.push_back(scaledValue);	      
	    }
	  }
	}

600
	MatSetValues(matIntInt, 1, &rowIndex, cols.size(), 
Thomas Witkowski's avatar
Thomas Witkowski committed
601 602 603 604 605 606 607 608 609 610 611 612 613 614 615
		     &(cols[0]), &(values[0]), ADD_VALUES);	
      } else {
	// === Row DOF index is periodic. ===

	// Because this row is periodic, we will have to add the entries of this 
	// matrix row to multiple rows. The following maps store to each row an
	// array of column indices and values of the entries that must be added to
	// the PETSc matrix.
	map<int, vector<int> > colsMap;
	map<int, vector<double> > valsMap;

	// Traverse all column entries.
	for (icursor_type icursor = begin<nz>(cursor), icend = end<nz>(cursor); 
	     icursor != icend; ++icursor) {
	  // Global index of the current column index.
616
	  int globalColDof = (*interiorMap)[colFe][col(*icursor)].global;
Thomas Witkowski's avatar
Thomas Witkowski committed
617 618 619 620 621 622 623 624 625 626

	  // Ignore all zero entries, expect it is a diagonal entry.
 	  if (value(*icursor) == 0.0 && globalRowDof != globalColDof)
 	    continue;

	  // === Add all periodic associations of both, the row and the column ===
	  // === indices to the set perAsc.                                    ===

	  std::set<int> perAsc;

627 628
	  if (perMap.isPeriodic(colFe, globalColDof)) {
	    std::set<int>& perColAsc = perMap.getAssociations(colFe, globalColDof);
Thomas Witkowski's avatar
Thomas Witkowski committed
629 630
	    for (std::set<int>::iterator it = perColAsc.begin(); 
		 it != perColAsc.end(); ++it)
631
	      if (meshDistributor->getElementObjectDb().isValidPeriodicType(*it))
Thomas Witkowski's avatar
Thomas Witkowski committed
632 633 634
		perAsc.insert(*it);
	  }

635
	  std::set<int>& perRowAsc = perMap.getAssociations(rowFe, globalRowDof);
Thomas Witkowski's avatar
Thomas Witkowski committed
636 637
	  for (std::set<int>::iterator it = perRowAsc.begin(); 
	       it != perRowAsc.end(); ++it)
638
	    if (meshDistributor->getElementObjectDb().isValidPeriodicType(*it))
Thomas Witkowski's avatar
Thomas Witkowski committed
639 640 641 642 643 644 645 646 647 648 649 650 651 652 653
	      perAsc.insert(*it);

	  // Scale the value with respect to the number of periodic associations.
	  double scaledValue = 
	    value(*icursor) * pow(0.5, static_cast<double>(perAsc.size()));


	  // === Create all matrix entries with respect to the periodic  ===
	  // === associations of the row and column indices.             ===

	  vector<pair<int, int> > entry;
	  
	  // First, add the original entry.
	  entry.push_back(make_pair(globalRowDof, globalColDof));

654 655
	  // Then, traverse the periodic associations of the row and column
	  // indices and create the corresponding entries.
Thomas Witkowski's avatar
Thomas Witkowski committed
656 657 658 659
	  for (std::set<int>::iterator it = perAsc.begin(); it != perAsc.end(); ++it) {
	    int nEntry = static_cast<int>(entry.size());
	    for (int i = 0; i < nEntry; i++) {
	      int perRowDof = 0;
660

661 662
	      if (perMap.isPeriodic(rowFe, *it, entry[i].first))
		perRowDof = perMap.map(rowFe, *it, entry[i].first);
Thomas Witkowski's avatar
Thomas Witkowski committed
663 664 665 666
	      else
		perRowDof = entry[i].first;

	      int perColDof;
667 668
	      if (perMap.isPeriodic(colFe, *it, entry[i].second))
		perColDof = perMap.map(colFe, *it, entry[i].second);
Thomas Witkowski's avatar
Thomas Witkowski committed
669 670 671 672 673 674 675 676 677 678 679
	      else
		perColDof = entry[i].second;	      	      
	      

	      entry.push_back(make_pair(perRowDof, perColDof));
	    }
	  }


	  // === Translate the matrix entries to PETSc's matrix.

680
	  for (unsigned int i = 0; i < entry.size(); i++) {
681 682
	    int rowIdx = interiorMap->getMatIndex(nRowMat, entry[i].first);
	    int colIdx = interiorMap->getMatIndex(nColMat, entry[i].second);
Thomas Witkowski's avatar
Thomas Witkowski committed
683

684 685
	    colsMap[rowIdx].push_back(colIdx);
	    valsMap[rowIdx].push_back(scaledValue);
Thomas Witkowski's avatar
Thomas Witkowski committed
686 687 688 689 690 691 692 693 694 695 696 697
	  }
	}


	// === Finally, add all periodic rows to the PETSc matrix. ===

	for (map<int, vector<int> >::iterator rowIt = colsMap.begin();
	     rowIt != colsMap.end(); ++rowIt) {
	  TEST_EXIT_DBG(rowIt->second.size() == valsMap[rowIt->first].size())
	    ("Should not happen!\n");

	  int rowIndex = rowIt->first;
698

699
	  MatSetValues(matIntInt, 1, &rowIndex, rowIt->second.size(),
Thomas Witkowski's avatar
Thomas Witkowski committed
700 701 702 703 704 705 706
		       &(rowIt->second[0]), &(valsMap[rowIt->first][0]), ADD_VALUES);
	}
      }
    }
  }


707 708
  void PetscSolverGlobalMatrix::setDofVector(Vec& petscVec, 
					     DOFVector<double>* vec, 
709
					     int nRowVec, 
710
					     bool rankOnly)
Thomas Witkowski's avatar
Thomas Witkowski committed
711 712 713
  {
    FUNCNAME("PetscSolverGlobalMatrix::setDofVector()");

714
    const FiniteElemSpace *feSpace = vec->getFeSpace();
715
    PeriodicMap &perMap = meshDistributor->getPeriodicMap();
716

Thomas Witkowski's avatar
Thomas Witkowski committed
717 718 719
    // Traverse all used DOFs in the dof vector.
    DOFVector<double>::Iterator dofIt(vec, USED_DOFS);
    for (dofIt.reset(); !dofIt.end(); ++dofIt) {
720
      if (rankOnly && !(*interiorMap)[feSpace].isRankDof(dofIt.getDOFIndex()))
Thomas Witkowski's avatar
Thomas Witkowski committed
721 722 723 724
	continue;

      // Calculate global row index of the DOF.
      DegreeOfFreedom globalRowDof = 
725
	(*interiorMap)[feSpace][dofIt.getDOFIndex()].global;
726
      // Get PETSc's mat index of the row DOF.
Thomas Witkowski's avatar
Thomas Witkowski committed
727 728 729 730 731
      int index = 0;
      if (interiorMap->isMatIndexFromGlobal())
	index = interiorMap->getMatIndex(nRowVec, globalRowDof);
      else
	index = interiorMap->getMatIndex(nRowVec, dofIt.getDOFIndex());
Thomas Witkowski's avatar
Thomas Witkowski committed
732

733 734
      if (perMap.isPeriodic(feSpace, globalRowDof)) {
	std::set<int>& perAsc = perMap.getAssociations(feSpace, globalRowDof);
Thomas Witkowski's avatar
Thomas Witkowski committed
735
	double value = *dofIt / (perAsc.size() + 1.0);
Thomas Witkowski's avatar
Thomas Witkowski committed
736
	VecSetValue(petscVec, index, value, ADD_VALUES);
Thomas Witkowski's avatar
Thomas Witkowski committed
737

738 739 740
	for (std::set<int>::iterator perIt = perAsc.begin(); 
	     perIt != perAsc.end(); ++perIt) {
	  int mappedDof = perMap.map(feSpace, *perIt, globalRowDof);
741
	  int mappedIndex = interiorMap->getMatIndex(nRowVec, mappedDof);
Thomas Witkowski's avatar
Thomas Witkowski committed
742
	  VecSetValue(petscVec, mappedIndex, value, ADD_VALUES);
Thomas Witkowski's avatar
Thomas Witkowski committed
743 744 745
	}
      } else {
	// The DOF index is not periodic.
Thomas Witkowski's avatar
Thomas Witkowski committed
746
	VecSetValue(petscVec, index, *dofIt, ADD_VALUES);
Thomas Witkowski's avatar
Thomas Witkowski committed
747 748 749 750 751 752 753 754 755 756 757 758
      }
    }
  }


  void PetscSolverGlobalMatrix::createPetscNnzStructure(Matrix<DOFMatrix*> *mat)
  {
    FUNCNAME("PetscSolverGlobalMatrix::createPetscNnzStructure()");

    TEST_EXIT_DBG(!d_nnz)("There is something wrong!\n");
    TEST_EXIT_DBG(!o_nnz)("There is something wrong!\n");

759
    vector<const FiniteElemSpace*> feSpaces = getFeSpaces(mat);
760 761
    int nRankRows = interiorMap->getRankDofs();
    int rankStartIndex = interiorMap->getStartDofs();
762

Thomas Witkowski's avatar
Thomas Witkowski committed
763 764 765 766 767 768 769 770 771 772 773 774 775
    d_nnz = new int[nRankRows];
    o_nnz = new int[nRankRows];
    for (int i = 0; i < nRankRows; i++) {
      d_nnz[i] = 0;
      o_nnz[i] = 0;
    }

    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 vector<pair<int, int> > MatrixNnzEntry;
    typedef map<int, DofContainer> RankToDofContainer;

776 777 778 779
    // Stores to each rank a list of nnz entries (i.e. pairs of row and column
    // index) that this rank will send to. These nnz entries will be assembled
    // on this rank, but because the row DOFs are not DOFs of this rank they 
    // will be send to the owner of the row DOFs.
Thomas Witkowski's avatar
Thomas Witkowski committed
780 781 782 783
    map<int, MatrixNnzEntry> sendMatrixEntry;

    // Maps to each DOF that must be send to another rank the rank number of the
    // receiving rank.
784
    map<pair<DegreeOfFreedom, int>, int> sendDofToRank;
Thomas Witkowski's avatar
Thomas Witkowski committed
785

786

787 788 789
    // First, create for all ranks, to which we send data to, MatrixNnzEntry 
    // object with 0 entries.
    for (unsigned int i = 0; i < feSpaces.size(); i++) {
790
      for (DofComm::Iterator it(meshDistributor->getDofComm().getRecvDofs(), feSpaces[i]);
791 792 793 794 795 796
	   !it.end(); it.nextRank()) {
	sendMatrixEntry[it.getRank()].resize(0);
	
	for (; !it.endDofIter(); it.nextDof())
	  sendDofToRank[make_pair(it.getDofIndex(), i)] = it.getRank();
      }
Thomas Witkowski's avatar
Thomas Witkowski committed
797 798
    }

799
    // Create list of ranks from which we receive data from.
Thomas Witkowski's avatar
Thomas Witkowski committed
800
    std::set<int> recvFromRank;
801
    for (unsigned int i = 0; i < feSpaces.size(); i++) 
802
      for (DofComm::Iterator it(meshDistributor->getDofComm().getSendDofs(), feSpaces[i]);
803 804 805 806 807
	   !it.end(); it.nextRank())
	recvFromRank.insert(it.getRank());


    // === Traverse matrices to create nnz data. ===
Thomas Witkowski's avatar
Thomas Witkowski committed
808

809
    int nComponents = mat->getNumRows();
Thomas Witkowski's avatar
Thomas Witkowski committed
810 811 812 813 814
    for (int i = 0; i < nComponents; i++) {
      for (int j = 0; j < nComponents; j++) {
 	if (!(*mat)[i][j])
	  continue;

815 816 817 818 819
	TEST_EXIT_DBG((*mat)[i][j]->getRowFeSpace() == feSpaces[i])
	  ("Should not happen!\n");
	TEST_EXIT_DBG((*mat)[i][j]->getColFeSpace() == feSpaces[j])
	  ("Should not happen!\n");

Thomas Witkowski's avatar
Thomas Witkowski committed
820 821 822 823 824 825 826 827 828 829
	Matrix bmat = (*mat)[i][j]->getBaseMatrix();

	traits::col<Matrix>::type col(bmat);
	traits::const_value<Matrix>::type value(bmat);
	  
	typedef traits::range_generator<row, Matrix>::type cursor_type;
	typedef traits::range_generator<nz, cursor_type>::type icursor_type;
	
	for (cursor_type cursor = begin<row>(bmat), 
	       cend = end<row>(bmat); cursor != cend; ++cursor) {
830
	  int globalRowDof = (*interiorMap)[feSpaces[i]][*cursor].global;
831

832
	  // The corresponding global matrix row index of the current row DOF.
833 834
	  int petscRowIdx = interiorMap->getMatIndex(i, globalRowDof);
	  if ((*interiorMap)[feSpaces[i]].isRankDof(*cursor)) {
835
    	    
836 837
	    // === The current row DOF is a rank DOF, so create the       ===
	    // === corresponding nnz values directly on rank's nnz data.  ===
Thomas Witkowski's avatar
Thomas Witkowski committed
838 839
	    
	    // This is the local row index of the local PETSc matrix.
840
	    int localPetscRowIdx = petscRowIdx - rankStartIndex;
Thomas Witkowski's avatar
Thomas Witkowski committed
841 842 843
	    
	    TEST_EXIT_DBG(localPetscRowIdx >= 0 && localPetscRowIdx < nRankRows)
	      ("Should not happen! \n Debug info: localRowIdx = %d   globalRowIndx = %d   petscRowIdx = %d   localPetscRowIdx = %d   rStart = %d   nCompontens = %d   nRankRows = %d\n",
844
	       *cursor,
845
	       (*interiorMap)[feSpaces[i]][*cursor].global,
Thomas Witkowski's avatar
Thomas Witkowski committed
846 847
	       petscRowIdx, 
	       localPetscRowIdx, 
848
	       rankStartIndex,
Thomas Witkowski's avatar
Thomas Witkowski committed
849 850 851 852 853 854 855
	       nComponents, 
	       nRankRows);
	    
	    
	    // Traverse all non zero entries in this row.
	    for (icursor_type icursor = begin<nz>(cursor), 
		   icend = end<nz>(cursor); icursor != icend; ++icursor) {
856 857
	      int globalColDof = (*interiorMap)[feSpaces[j]][col(*icursor)].global;
	      int petscColIdx = interiorMap->getMatIndex(j, globalColDof);
Thomas Witkowski's avatar
Thomas Witkowski committed
858 859 860
	      
	      if (value(*icursor) != 0.0 || petscRowIdx == petscColIdx) {
		// The row DOF is a rank DOF, if also the column is a rank DOF, 
861 862 863 864
		// increment the d_nnz values for this row, otherwise the 
		// o_nnz value.
		if (petscColIdx >= rankStartIndex && 
		    petscColIdx < rankStartIndex + nRankRows)
Thomas Witkowski's avatar
Thomas Witkowski committed
865 866 867 868 869 870
		  d_nnz[localPetscRowIdx]++;
		else
		  o_nnz[localPetscRowIdx]++;
	      }    
	    }
	  } else {
871 872 873 874 875
	    // === The current row DOF is not a rank DOF, i.e., its values   ===
	    // === are also created on this rank, but afterthere they will   ===
	    // === be send to another rank. So we need to send also the      ===
	    // === corresponding nnz structure of this row to the corres-    ===
	    // === ponding rank.                                             ===
Thomas Witkowski's avatar
Thomas Witkowski committed
876 877
	    
	    // Send all non zero entries to the member of the row DOF.
878
	    int sendToRank = sendDofToRank[make_pair(*cursor, i)];
Thomas Witkowski's avatar
Thomas Witkowski committed
879 880 881 882
	    
	    for (icursor_type icursor = begin<nz>(cursor), 
		   icend = end<nz>(cursor); icursor != icend; ++icursor) {
	      if (value(*icursor) != 0.0) {
883
		int globalColDof = 
884 885
		  (*interiorMap)[feSpaces[j]][col(*icursor)].global;
		int petscColIdx = interiorMap->getMatIndex(j, globalColDof);
Thomas Witkowski's avatar
Thomas Witkowski committed
886 887 888 889 890 891 892 893 894 895 896 897 898
		
		sendMatrixEntry[sendToRank].
		  push_back(make_pair(petscRowIdx, petscColIdx));
	      }
	    }
	    
	  } // if (isRankDof[*cursor]) ... else ...
	} // for each row in mat[i][j]
      } 
    }

    // === Send and recv the nnz row structure to/from other ranks. ===

899
    StdMpi<MatrixNnzEntry> stdMpi(mpiCommGlobal, true);
Thomas Witkowski's avatar
Thomas Witkowski committed
900 901 902 903 904 905 906
    stdMpi.send(sendMatrixEntry);
    for (std::set<int>::iterator it = recvFromRank.begin(); 
	 it != recvFromRank.end(); ++it)
      stdMpi.recv(*it);
    stdMpi.startCommunication();


907 908
    // === Evaluate the nnz structure this rank got from other ranks and add ===
    // === it to the PETSc nnz data structure.                               ===
Thomas Witkowski's avatar
Thomas Witkowski committed
909 910 911 912 913 914 915 916

    for (map<int, MatrixNnzEntry>::iterator it = stdMpi.getRecvData().begin();
	 it != stdMpi.getRecvData().end(); ++it) {
      if (it->second.size() > 0) {
	for (unsigned int i = 0; i < it->second.size(); i++) {
	  int r = it->second[i].first;
	  int c = it->second[i].second;

917
	  int localRowIdx = r - rankStartIndex;
Thomas Witkowski's avatar
Thomas Witkowski committed
918 919 920 921 922

	  TEST_EXIT_DBG(localRowIdx >= 0 && localRowIdx < nRankRows)
	    ("Got row index %d/%d (nRankRows = %d) from rank %d. Should not happen!\n",
	     r, localRowIdx, nRankRows, it->first);
	  
923
	  if (c < rankStartIndex || c >= rankStartIndex + nRankRows)
Thomas Witkowski's avatar
Thomas Witkowski committed
924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942
	    o_nnz[localRowIdx]++;
	  else
	    d_nnz[localRowIdx]++;
	}
      }
    }

    // The above algorithm for calculating the number of nnz per row over-
    // approximates the value, i.e., the number is always equal or larger to 
    // the real number of nnz values in the global parallel matrix. For small
    // matrices, the problem may arise, that the result is larger than the
    // number of elements in a row. This is fixed in the following.

    if (nRankRows < 100) 
      for (int i = 0; i < nRankRows; i++)
	d_nnz[i] = std::min(d_nnz[i], nRankRows);
  }

}