PetscSolverGlobalMatrix.cc 22.4 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 24
  {
    FUNCNAME("PetscSolverGlobalMatrix::fillPetscMatrix()");

    TEST_EXIT_DBG(meshDistributor)("No mesh distributor object defined!\n");
25
    TEST_EXIT_DBG(interiorMap)("No parallel mapping object defined!\n");
Thomas Witkowski's avatar
Thomas Witkowski committed
26
    TEST_EXIT_DBG(mat)("No DOF matrix defined!\n");
27
    
Thomas Witkowski's avatar
Thomas Witkowski committed
28
    double wtime = MPI::Wtime();
29

30 31 32
    
    // === Check if mesh was changed and, in this case, recompute matrix ===
    // === nnz structure and matrix indices.                             ===
33

Thomas Witkowski's avatar
Thomas Witkowski committed
34
    int recvAllValues = 0;
35 36
    int sendValue = 
      static_cast<int>(meshDistributor->getLastMeshChangeIndex() != lastMeshNnz);
37
    mpiCommGlobal.Allreduce(&sendValue, &recvAllValues, 1, MPI_INT, MPI_SUM);
Thomas Witkowski's avatar
Thomas Witkowski committed
38

39
    if (!d_nnz || recvAllValues != 0 || alwaysCreateNnzStructure) {
40
      vector<const FiniteElemSpace*> feSpaces = getFeSpaces(mat);
41 42
      interiorMap->setComputeMatIndex(true, true);
      interiorMap->update(feSpaces);
43

Thomas Witkowski's avatar
Thomas Witkowski committed
44 45 46 47 48 49 50 51 52 53 54 55
      if (d_nnz) {
	delete [] d_nnz;
	d_nnz = NULL;
	delete [] o_nnz;
	o_nnz = NULL;
      }

      createPetscNnzStructure(mat);
      lastMeshNnz = meshDistributor->getLastMeshChangeIndex();
    }


56 57
    // === Create PETSc vector (solution and a temporary vector). ===

58 59
    int nRankRows = interiorMap->getRankDofs();
    int nOverallRows = interiorMap->getOverallDofs();
60

61
    VecCreateMPI(mpiCommGlobal, nRankRows, nOverallRows, &petscSolVec);
62 63 64

    int testddd = 1;
    Parameters::get("block size", testddd);
65
    if (testddd > 1)
66
      VecSetBlockSize(petscSolVec, testddd);
67

68 69


Thomas Witkowski's avatar
Thomas Witkowski committed
70 71
    // === Create PETSc matrix with the computed nnz data structure. ===

72
    MatCreateMPIAIJ(mpiCommGlobal, nRankRows, nRankRows, 
73
 		    nOverallRows, nOverallRows,
74
		    0, d_nnz, 0, o_nnz, &matIntInt);
75

76
    if (testddd > 1) {
77
      MatSetBlockSize(matIntInt, testddd);
78 79 80
      MSG("MAT SET BLOCK SIZE: %d\n", testddd);
    }

Thomas Witkowski's avatar
Thomas Witkowski committed
81 82 83 84 85 86
#if (DEBUG != 0)
    MSG("Fill petsc matrix 1 needed %.5f seconds\n", MPI::Wtime() - wtime);
#endif

#if (DEBUG != 0)
    int a, b;
87 88 89
    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
90 91 92 93 94 95
      ("Wrong matrix ownership range!\n");
#endif


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

96
    int nComponents = mat->getNumRows();
Thomas Witkowski's avatar
Thomas Witkowski committed
97 98 99
    for (int i = 0; i < nComponents; i++)
      for (int j = 0; j < nComponents; j++)
	if ((*mat)[i][j])
100
	  setDofMatrix((*mat)[i][j], i, j);
Thomas Witkowski's avatar
Thomas Witkowski committed
101 102 103 104 105

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

106 107
    MatAssemblyBegin(matIntInt, MAT_FINAL_ASSEMBLY);
    MatAssemblyEnd(matIntInt, MAT_FINAL_ASSEMBLY);
Thomas Witkowski's avatar
Thomas Witkowski committed
108

109
    // === Init PETSc solver. ===
110 111 112 113 114 115 116 117
    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);
118 119 120

    // Do not delete the solution vector, use it for the initial guess.
    if (!zeroStartVector)
121
      KSPSetInitialGuessNonzero(kspInterior, PETSC_TRUE);
122

123 124 125 126 127 128 129
    MSG("Fill petsc matrix needed %.5f seconds\n", MPI::Wtime() - wtime);
  }


  void PetscSolverGlobalMatrix::fillPetscRhs(SystemVector *vec)
  {
    FUNCNAME("PetscSolverGlobalMatrix::fillPetscRhs()");
Thomas Witkowski's avatar
Thomas Witkowski committed
130

131
    TEST_EXIT_DBG(vec)("No DOF vector defined!\n");
132
    TEST_EXIT_DBG(interiorMap)("No parallel DOF map defined!\n");
133

134 135
    int nRankRows = interiorMap->getRankDofs();
    int nOverallRows = interiorMap->getOverallDofs();
136

137
    VecCreateMPI(mpiCommGlobal, nRankRows, nOverallRows, &rhsInterior);
Thomas Witkowski's avatar
Thomas Witkowski committed
138

139 140 141
    int testddd = 1;
    Parameters::get("block size", testddd);
    if (testddd > 1)
142
      VecSetBlockSize(rhsInterior, testddd);
143 144


Thomas Witkowski's avatar
Thomas Witkowski committed
145
    // === Transfer values from DOF vector to the PETSc vector. === 
146
    for (int i = 0; i < vec->getSize(); i++)
147
      setDofVector(rhsInterior, vec->getDOFVector(i), i);
Thomas Witkowski's avatar
Thomas Witkowski committed
148

149 150
    VecAssemblyBegin(rhsInterior);
    VecAssemblyEnd(rhsInterior);
151 152

    if (removeRhsNullSpace) {
153
      MSG("Remove constant null space from the RHS!\n");
154
      MatNullSpace sp;
155 156
      MatNullSpaceCreate(mpiCommGlobal, PETSC_TRUE, 0, PETSC_NULL, &sp);
      MatNullSpaceRemove(sp, rhsInterior, PETSC_NULL);
157 158
      MatNullSpaceDestroy(&sp);
    }
Thomas Witkowski's avatar
Thomas Witkowski committed
159 160 161
  }


162 163
  void PetscSolverGlobalMatrix::solvePetscMatrix(SystemVector &vec, 
						 AdaptInfo *adaptInfo)
Thomas Witkowski's avatar
Thomas Witkowski committed
164 165 166 167 168 169 170 171 172 173
  {
    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++)
174
	setDofVector(petscSolVec, vec.getDOFVector(i), i, true);
Thomas Witkowski's avatar
Thomas Witkowski committed
175 176 177 178 179 180
      
      VecAssemblyBegin(petscSolVec);
      VecAssemblyEnd(petscSolVec);
    }

    // PETSc.
181
    solve(rhsInterior, petscSolVec);
Thomas Witkowski's avatar
Thomas Witkowski committed
182 183 184 185 186

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

187
    int c = 0;
Thomas Witkowski's avatar
Thomas Witkowski committed
188
    for (int i = 0; i < nComponents; i++) {
189
      DOFVector<double> &dv = *(vec.getDOFVector(i));
190
      DofMap& d = (*interiorMap)[dv.getFeSpace()].getMap();
191 192 193
      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
194 195 196 197 198 199 200 201
    }

    VecRestoreArray(petscSolVec, &vecPointer);


    // === Synchronize DOFs at common DOFs, i.e., DOFs that correspond to ===
    // === more than one partition.                                       ===
    meshDistributor->synchVector(vec);
202 203 204 205 206 207 208
  }


  void PetscSolverGlobalMatrix::destroyMatrixData()
  {
    FUNCNAME("PetscSolverGlobalMatrix::destroyMatrixData()");

209 210
    MatDestroy(&matIntInt);
    KSPDestroy(&kspInterior);
211
    VecDestroy(&petscSolVec);
Thomas Witkowski's avatar
Thomas Witkowski committed
212 213 214
  }


215 216 217 218
  void PetscSolverGlobalMatrix::destroyVectorData()
  {
    FUNCNAME("PetscSolverGlobalMatrix::destroyVectorData()");

219
    VecDestroy(&rhsInterior);
220 221 222
  }


223 224
  void PetscSolverGlobalMatrix::setDofMatrix(DOFMatrix* mat,
					     int nRowMat, int nColMat)
Thomas Witkowski's avatar
Thomas Witkowski committed
225 226 227 228 229 230
  {
    FUNCNAME("PetscSolverGlobalMatrix::setDofMatrix()");

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

    using mtl::tag::row; using mtl::tag::nz; using mtl::begin; using mtl::end;
231
    namespace traits = mtl::traits;
Thomas Witkowski's avatar
Thomas Witkowski committed
232 233 234 235 236 237 238 239 240 241 242 243 244 245 246
    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;

247 248 249
    // Get periodic mapping object
    PeriodicMap &perMap = meshDistributor->getPeriodicMap();

Thomas Witkowski's avatar
Thomas Witkowski committed
250 251 252 253 254 255
    // === 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) {

256 257 258
      const FiniteElemSpace *rowFe = mat->getRowFeSpace();
      const FiniteElemSpace *colFe = mat->getColFeSpace();

Thomas Witkowski's avatar
Thomas Witkowski committed
259
      // Global index of the current row DOF.
260
      int globalRowDof = (*interiorMap)[rowFe][*cursor].global;
261

Thomas Witkowski's avatar
Thomas Witkowski committed
262
      // Test if the current row DOF is a periodic DOF.
263
      bool periodicRow = perMap.isPeriodic(rowFe, globalRowDof);
264

Thomas Witkowski's avatar
Thomas Witkowski committed
265 266 267
      if (!periodicRow) {
	// === Row DOF index is not periodic. ===

268
	// Get PETSc's mat row index.
269
	int rowIndex = interiorMap->getMatIndex(nRowMat, globalRowDof);
Thomas Witkowski's avatar
Thomas Witkowski committed
270 271 272 273 274 275 276 277

	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.
278
	  int globalColDof = (*interiorMap)[colFe][col(*icursor)].global;
Thomas Witkowski's avatar
Thomas Witkowski committed
279
	  // Test if the current col dof is a periodic dof.
280
	  bool periodicCol = perMap.isPeriodic(colFe, globalColDof);
281
	  // Get PETSc's mat col index.
282
	  int colIndex = interiorMap->getMatIndex(nColMat, globalColDof);
Thomas Witkowski's avatar
Thomas Witkowski committed
283 284 285 286 287 288 289 290 291 292 293 294 295 296

	  // 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;
297
	    std::set<int>& perColAsc = perMap.getAssociations(colFe, globalColDof);
Thomas Witkowski's avatar
Thomas Witkowski committed
298 299
	    for (std::set<int>::iterator it = perColAsc.begin(); 
		 it != perColAsc.end(); ++it)
300
	      if (meshDistributor->getElementObjectDb().isValidPeriodicType(*it))
Thomas Witkowski's avatar
Thomas Witkowski committed
301 302 303 304 305 306 307 308 309 310 311 312 313 314
		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);
315

Thomas Witkowski's avatar
Thomas Witkowski committed
316 317 318 319 320 321
	    // 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++) {
322
 		TEST_EXIT_DBG(perMap.isPeriodic(colFe, *it, newCols[i]))
Thomas Witkowski's avatar
Thomas Witkowski committed
323 324 325
 		  ("Wrong periodic DOF associations at boundary %d with DOF %d!\n",
		   *it, newCols[i]);

326
		newCols.push_back(perMap.map(colFe, *it, newCols[i]));
Thomas Witkowski's avatar
Thomas Witkowski committed
327 328 329 330
	      }
	    }

	    for (unsigned int i = 0; i < newCols.size(); i++) {
331
	      cols.push_back(interiorMap->getMatIndex(nColMat, newCols[i]));
Thomas Witkowski's avatar
Thomas Witkowski committed
332 333 334 335 336
	      values.push_back(scaledValue);	      
	    }
	  }
	}

337
	MatSetValues(matIntInt, 1, &rowIndex, cols.size(), 
Thomas Witkowski's avatar
Thomas Witkowski committed
338 339 340 341 342 343 344 345 346 347 348 349 350 351 352
		     &(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.
353
	  int globalColDof = (*interiorMap)[colFe][col(*icursor)].global;
Thomas Witkowski's avatar
Thomas Witkowski committed
354 355 356 357 358 359 360 361 362 363

	  // 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;

364 365
	  if (perMap.isPeriodic(colFe, globalColDof)) {
	    std::set<int>& perColAsc = perMap.getAssociations(colFe, globalColDof);
Thomas Witkowski's avatar
Thomas Witkowski committed
366 367
	    for (std::set<int>::iterator it = perColAsc.begin(); 
		 it != perColAsc.end(); ++it)
368
	      if (meshDistributor->getElementObjectDb().isValidPeriodicType(*it))
Thomas Witkowski's avatar
Thomas Witkowski committed
369 370 371
		perAsc.insert(*it);
	  }

372
	  std::set<int>& perRowAsc = perMap.getAssociations(rowFe, globalRowDof);
Thomas Witkowski's avatar
Thomas Witkowski committed
373 374
	  for (std::set<int>::iterator it = perRowAsc.begin(); 
	       it != perRowAsc.end(); ++it)
375
	    if (meshDistributor->getElementObjectDb().isValidPeriodicType(*it))
Thomas Witkowski's avatar
Thomas Witkowski committed
376 377 378 379 380 381 382 383 384 385 386 387 388 389 390
	      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));

391 392
	  // Then, traverse the periodic associations of the row and column
	  // indices and create the corresponding entries.
Thomas Witkowski's avatar
Thomas Witkowski committed
393 394 395 396
	  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;
397

398 399
	      if (perMap.isPeriodic(rowFe, *it, entry[i].first))
		perRowDof = perMap.map(rowFe, *it, entry[i].first);
Thomas Witkowski's avatar
Thomas Witkowski committed
400 401 402 403
	      else
		perRowDof = entry[i].first;

	      int perColDof;
404 405
	      if (perMap.isPeriodic(colFe, *it, entry[i].second))
		perColDof = perMap.map(colFe, *it, entry[i].second);
Thomas Witkowski's avatar
Thomas Witkowski committed
406 407 408 409 410 411 412 413 414 415 416
	      else
		perColDof = entry[i].second;	      	      
	      

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


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

417
	  for (unsigned int i = 0; i < entry.size(); i++) {
418 419
	    int rowIdx = interiorMap->getMatIndex(nRowMat, entry[i].first);
	    int colIdx = interiorMap->getMatIndex(nColMat, entry[i].second);
Thomas Witkowski's avatar
Thomas Witkowski committed
420

421 422
	    colsMap[rowIdx].push_back(colIdx);
	    valsMap[rowIdx].push_back(scaledValue);
Thomas Witkowski's avatar
Thomas Witkowski committed
423 424 425 426 427 428 429 430 431 432 433 434
	  }
	}


	// === 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;
435

436
	  MatSetValues(matIntInt, 1, &rowIndex, rowIt->second.size(),
Thomas Witkowski's avatar
Thomas Witkowski committed
437 438 439 440 441 442 443
		       &(rowIt->second[0]), &(valsMap[rowIt->first][0]), ADD_VALUES);
	}
      }
    }
  }


444 445
  void PetscSolverGlobalMatrix::setDofVector(Vec& petscVec, 
					     DOFVector<double>* vec, 
446
					     int nRowVec, 
447
					     bool rankOnly)
Thomas Witkowski's avatar
Thomas Witkowski committed
448 449 450
  {
    FUNCNAME("PetscSolverGlobalMatrix::setDofVector()");

451
    const FiniteElemSpace *feSpace = vec->getFeSpace();
452
    PeriodicMap &perMap = meshDistributor->getPeriodicMap();
453

Thomas Witkowski's avatar
Thomas Witkowski committed
454 455 456
    // Traverse all used DOFs in the dof vector.
    DOFVector<double>::Iterator dofIt(vec, USED_DOFS);
    for (dofIt.reset(); !dofIt.end(); ++dofIt) {
457
      if (rankOnly && !(*interiorMap)[feSpace].isRankDof(dofIt.getDOFIndex()))
Thomas Witkowski's avatar
Thomas Witkowski committed
458 459 460 461
	continue;

      // Calculate global row index of the DOF.
      DegreeOfFreedom globalRowDof = 
462
	(*interiorMap)[feSpace][dofIt.getDOFIndex()].global;
463
      // Get PETSc's mat index of the row DOF.
464
      int index = interiorMap->getMatIndex(nRowVec, globalRowDof);
Thomas Witkowski's avatar
Thomas Witkowski committed
465

466 467
      if (perMap.isPeriodic(feSpace, globalRowDof)) {
	std::set<int>& perAsc = perMap.getAssociations(feSpace, globalRowDof);
Thomas Witkowski's avatar
Thomas Witkowski committed
468 469 470
	double value = *dofIt / (perAsc.size() + 1.0);
	VecSetValues(petscVec, 1, &index, &value, ADD_VALUES);

471 472 473
	for (std::set<int>::iterator perIt = perAsc.begin(); 
	     perIt != perAsc.end(); ++perIt) {
	  int mappedDof = perMap.map(feSpace, *perIt, globalRowDof);
474
	  int mappedIndex = interiorMap->getMatIndex(nRowVec, mappedDof);
Thomas Witkowski's avatar
Thomas Witkowski committed
475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492
	  VecSetValues(petscVec, 1, &mappedIndex, &value, ADD_VALUES);
	}
      } else {
	// The DOF index is not periodic.
	double value = *dofIt;
	VecSetValues(petscVec, 1, &index, &value, ADD_VALUES);
      }
    }
  }


  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");

493
    vector<const FiniteElemSpace*> feSpaces = getFeSpaces(mat);
494 495
    int nRankRows = interiorMap->getRankDofs();
    int rankStartIndex = interiorMap->getStartDofs();
496

Thomas Witkowski's avatar
Thomas Witkowski committed
497 498 499 500 501 502 503 504 505 506 507 508 509
    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;

510 511 512 513
    // 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
514 515 516 517
    map<int, MatrixNnzEntry> sendMatrixEntry;

    // Maps to each DOF that must be send to another rank the rank number of the
    // receiving rank.
518
    map<pair<DegreeOfFreedom, int>, int> sendDofToRank;
Thomas Witkowski's avatar
Thomas Witkowski committed
519

520

521 522 523
    // 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++) {
524
      for (DofComm::Iterator it(meshDistributor->getDofComm().getRecvDofs(), feSpaces[i]);
525 526 527 528 529 530
	   !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
531 532
    }

533
    // Create list of ranks from which we receive data from.
Thomas Witkowski's avatar
Thomas Witkowski committed
534
    std::set<int> recvFromRank;
535
    for (unsigned int i = 0; i < feSpaces.size(); i++) 
536
      for (DofComm::Iterator it(meshDistributor->getDofComm().getSendDofs(), feSpaces[i]);
537 538 539 540 541
	   !it.end(); it.nextRank())
	recvFromRank.insert(it.getRank());


    // === Traverse matrices to create nnz data. ===
Thomas Witkowski's avatar
Thomas Witkowski committed
542

543
    int nComponents = mat->getNumRows();
Thomas Witkowski's avatar
Thomas Witkowski committed
544 545 546 547 548
    for (int i = 0; i < nComponents; i++) {
      for (int j = 0; j < nComponents; j++) {
 	if (!(*mat)[i][j])
	  continue;

549 550 551 552 553
	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
554 555 556 557 558 559 560 561 562 563
	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) {
564
	  int globalRowDof = (*interiorMap)[feSpaces[i]][*cursor].global;
565

566
	  // The corresponding global matrix row index of the current row DOF.
567 568
	  int petscRowIdx = interiorMap->getMatIndex(i, globalRowDof);
	  if ((*interiorMap)[feSpaces[i]].isRankDof(*cursor)) {
569
    	    
570 571
	    // === 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
572 573
	    
	    // This is the local row index of the local PETSc matrix.
574
	    int localPetscRowIdx = petscRowIdx - rankStartIndex;
Thomas Witkowski's avatar
Thomas Witkowski committed
575 576 577
	    
	    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",
578
	       *cursor,
579
	       (*interiorMap)[feSpaces[i]][*cursor].global,
Thomas Witkowski's avatar
Thomas Witkowski committed
580 581
	       petscRowIdx, 
	       localPetscRowIdx, 
582
	       rankStartIndex,
Thomas Witkowski's avatar
Thomas Witkowski committed
583 584 585 586 587 588 589
	       nComponents, 
	       nRankRows);
	    
	    
	    // Traverse all non zero entries in this row.
	    for (icursor_type icursor = begin<nz>(cursor), 
		   icend = end<nz>(cursor); icursor != icend; ++icursor) {
590 591
	      int globalColDof = (*interiorMap)[feSpaces[j]][col(*icursor)].global;
	      int petscColIdx = interiorMap->getMatIndex(j, globalColDof);
Thomas Witkowski's avatar
Thomas Witkowski committed
592 593 594
	      
	      if (value(*icursor) != 0.0 || petscRowIdx == petscColIdx) {
		// The row DOF is a rank DOF, if also the column is a rank DOF, 
595 596 597 598
		// 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
599 600 601 602 603 604
		  d_nnz[localPetscRowIdx]++;
		else
		  o_nnz[localPetscRowIdx]++;
	      }    
	    }
	  } else {
605 606 607 608 609
	    // === 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
610 611
	    
	    // Send all non zero entries to the member of the row DOF.
612
	    int sendToRank = sendDofToRank[make_pair(*cursor, i)];
Thomas Witkowski's avatar
Thomas Witkowski committed
613 614 615 616
	    
	    for (icursor_type icursor = begin<nz>(cursor), 
		   icend = end<nz>(cursor); icursor != icend; ++icursor) {
	      if (value(*icursor) != 0.0) {
617
		int globalColDof = 
618 619
		  (*interiorMap)[feSpaces[j]][col(*icursor)].global;
		int petscColIdx = interiorMap->getMatIndex(j, globalColDof);
Thomas Witkowski's avatar
Thomas Witkowski committed
620 621 622 623 624 625 626 627 628 629 630 631 632
		
		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. ===

633
    StdMpi<MatrixNnzEntry> stdMpi(mpiCommGlobal, true);
Thomas Witkowski's avatar
Thomas Witkowski committed
634 635 636 637 638 639 640
    stdMpi.send(sendMatrixEntry);
    for (std::set<int>::iterator it = recvFromRank.begin(); 
	 it != recvFromRank.end(); ++it)
      stdMpi.recv(*it);
    stdMpi.startCommunication();


641 642
    // === 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
643 644 645 646 647 648 649 650

    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;

651
	  int localRowIdx = r - rankStartIndex;
Thomas Witkowski's avatar
Thomas Witkowski committed
652 653 654 655 656

	  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);
	  
657
	  if (c < rankStartIndex || c >= rankStartIndex + nRankRows)
Thomas Witkowski's avatar
Thomas Witkowski committed
658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676
	    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);
  }

}