PetscSolverGlobalMatrix.cc 25.4 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(dofMap)("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
    vector<const FiniteElemSpace*> feSpaces = getFeSpaces(mat);
30
31
32
33
    dofMap->update(feSpaces);

    int nRankRows = dofMap->getRankDofs();
    int nOverallRows = dofMap->getOverallDofs();
Thomas Witkowski's avatar
Thomas Witkowski committed
34

35
    // === Create PETSc vector (solution and a temporary vector). ===
Thomas Witkowski's avatar
Thomas Witkowski committed
36

37
38
    VecCreateMPI(mpiComm, nRankRows, nOverallRows, &petscSolVec);
    VecCreateMPI(mpiComm, nRankRows, nOverallRows, &petscTmpVec);
Thomas Witkowski's avatar
Thomas Witkowski committed
39

40
41
42
43
44
45
46
47
    int testddd = 1;
    Parameters::get("block size", testddd);
    if (testddd > 1) {
      VecSetBlockSize(petscSolVec, testddd);
      VecSetBlockSize(petscTmpVec, testddd);
    }


Thomas Witkowski's avatar
Thomas Witkowski committed
48
    int recvAllValues = 0;
49
50
    int sendValue = 
      static_cast<int>(meshDistributor->getLastMeshChangeIndex() != lastMeshNnz);
Thomas Witkowski's avatar
Thomas Witkowski committed
51
    mpiComm.Allreduce(&sendValue, &recvAllValues, 1, MPI_INT, MPI_SUM);
Thomas Witkowski's avatar
Thomas Witkowski committed
52

53
54
55
    recvAllValues = 1;

    if (!d_nnz || recvAllValues != 0 || alwaysCreateNnzStructure) {
56
57
58
59
60

      // Global DOF mapping must only be recomputed, if the NNZ structure has
      // changed (thus, also the mesh has changed).
      createGlobalDofMapping(feSpaces);

Thomas Witkowski's avatar
Thomas Witkowski committed
61
62
63
64
65
66
67
68
69
70
71
72
73
74
      if (d_nnz) {
	delete [] d_nnz;
	d_nnz = NULL;
	delete [] o_nnz;
	o_nnz = NULL;
      }

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


    // === Create PETSc matrix with the computed nnz data structure. ===

75
    MatCreateMPIAIJ(mpiComm, nRankRows, nRankRows, 
76
 		    nOverallRows, nOverallRows,
77
		    0, d_nnz, 0, o_nnz, &petscMatrix);
78

79
80
81
82
83
    if (testddd > 1) {
      MatSetBlockSize(petscMatrix, testddd);
      MSG("MAT SET BLOCK SIZE: %d\n", testddd);
    }

Thomas Witkowski's avatar
Thomas Witkowski committed
84
85
86
87
88
89
90
#if (DEBUG != 0)
    MSG("Fill petsc matrix 1 needed %.5f seconds\n", MPI::Wtime() - wtime);
#endif

#if (DEBUG != 0)
    int a, b;
    MatGetOwnershipRange(petscMatrix, &a, &b);
91
92
    TEST_EXIT(a == dofMap->getStartDofs())("Wrong matrix ownership range!\n");
    TEST_EXIT(b == dofMap->getStartDofs() + nRankRows)
Thomas Witkowski's avatar
Thomas Witkowski committed
93
94
95
96
97
98
      ("Wrong matrix ownership range!\n");
#endif


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

99
    int nComponents = mat->getNumRows();
Thomas Witkowski's avatar
Thomas Witkowski committed
100
101
102
    for (int i = 0; i < nComponents; i++)
      for (int j = 0; j < nComponents; j++)
	if ((*mat)[i][j])
103
	  setDofMatrix((*mat)[i][j], i, j);
Thomas Witkowski's avatar
Thomas Witkowski committed
104
105
106
107
108
109
110
111

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

    MatAssemblyBegin(petscMatrix, MAT_FINAL_ASSEMBLY);
    MatAssemblyEnd(petscMatrix, MAT_FINAL_ASSEMBLY);

112
    // === Init PETSc solver. ===
113
    KSPCreate(mpiComm, &solver);
114
115
116
117
    KSPGetPC(solver, &pc);
    KSPSetOperators(solver, petscMatrix, petscMatrix, SAME_NONZERO_PATTERN); 
    KSPSetTolerances(solver, 0.0, 1e-8, PETSC_DEFAULT, PETSC_DEFAULT);
    KSPSetType(solver, KSPBCGS);
118
    KSPSetOptionsPrefix(solver, kspPrefix.c_str());
119
120
121
122
123
124
125
    KSPSetFromOptions(solver);
    PCSetFromOptions(pc);

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

126
127
128
129
130
131
132
    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
133

134
    TEST_EXIT_DBG(vec)("No DOF vector defined!\n");
135
    TEST_EXIT_DBG(dofMap)("No parallel DOF map defined!\n");
136

137
    vector<const FiniteElemSpace*> feSpaces = getFeSpaces(vec);
138
139
    int nRankRows = dofMap->getRankDofs();
    int nOverallRows = dofMap->getOverallDofs();
140

141
    VecCreateMPI(mpiComm, nRankRows, nOverallRows, &petscRhsVec);
Thomas Witkowski's avatar
Thomas Witkowski committed
142

143
144
145
146
147
148
    int testddd = 1;
    Parameters::get("block size", testddd);
    if (testddd > 1)
      VecSetBlockSize(petscRhsVec, testddd);


Thomas Witkowski's avatar
Thomas Witkowski committed
149
    // === Transfer values from DOF vector to the PETSc vector. === 
150
    for (int i = 0; i < vec->getSize(); i++)
151
      setDofVector(petscRhsVec, vec->getDOFVector(i), i);
Thomas Witkowski's avatar
Thomas Witkowski committed
152
153
154

    VecAssemblyBegin(petscRhsVec);
    VecAssemblyEnd(petscRhsVec);
155
156

    if (removeRhsNullSpace) {
157
      MSG("Remove constant null space from the RHS!\n");
158
      MatNullSpace sp;
159
      MatNullSpaceCreate(mpiComm, PETSC_TRUE, 0, PETSC_NULL, &sp);
160
161
162
      MatNullSpaceRemove(sp, petscRhsVec, PETSC_NULL);
      MatNullSpaceDestroy(&sp);
    }
Thomas Witkowski's avatar
Thomas Witkowski committed
163
164
165
  }


166
167
  void PetscSolverGlobalMatrix::solvePetscMatrix(SystemVector &vec, 
						 AdaptInfo *adaptInfo)
Thomas Witkowski's avatar
Thomas Witkowski committed
168
169
170
171
172
173
174
175
176
177
  {
    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++)
178
	setDofVector(petscSolVec, vec.getDOFVector(i), i, true);
Thomas Witkowski's avatar
Thomas Witkowski committed
179
180
181
182
183
184
      
      VecAssemblyBegin(petscSolVec);
      VecAssemblyEnd(petscSolVec);
    }

    // PETSc.
185
186
187
188
189
    PetscErrorCode solverError = KSPSolve(solver, petscRhsVec, petscSolVec);
    if (solverError != 0) {
      AMDiS::finalize();
      exit(-1);
    }
Thomas Witkowski's avatar
Thomas Witkowski committed
190
191
192
193
194

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

195
    int c = 0;
Thomas Witkowski's avatar
Thomas Witkowski committed
196
    for (int i = 0; i < nComponents; i++) {
197
      DOFVector<double> &dv = *(vec.getDOFVector(i));
198
199
200
201
      DofMap& d = (*dofMap)[dv.getFeSpace()].getMap();
      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
202
203
204
205
206
207
208
209
210
211
212
213
    }

    VecRestoreArray(petscSolVec, &vecPointer);


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


    // Print iteration counter and residual norm of the solution.
    printSolutionInfo(adaptInfo);
214
215
216
217
218
219
220
221
222
  }


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

    MatDestroy(&petscMatrix);
    KSPDestroy(&solver);
223
224
    VecDestroy(&petscSolVec);
    VecDestroy(&petscTmpVec);
Thomas Witkowski's avatar
Thomas Witkowski committed
225
226
227
  }


228
229
230
231
232
233
234
235
  void PetscSolverGlobalMatrix::destroyVectorData()
  {
    FUNCNAME("PetscSolverGlobalMatrix::destroyVectorData()");

    VecDestroy(&petscRhsVec);
  }


236
237
  void PetscSolverGlobalMatrix::setDofMatrix(DOFMatrix* mat,
					     int nRowMat, int nColMat)
Thomas Witkowski's avatar
Thomas Witkowski committed
238
239
240
241
242
243
  {
    FUNCNAME("PetscSolverGlobalMatrix::setDofMatrix()");

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

    using mtl::tag::row; using mtl::tag::nz; using mtl::begin; using mtl::end;
244
    namespace traits = mtl::traits;
Thomas Witkowski's avatar
Thomas Witkowski committed
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
    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;

260
261
262
    // Get periodic mapping object
    PeriodicMap &perMap = meshDistributor->getPeriodicMap();

Thomas Witkowski's avatar
Thomas Witkowski committed
263
264
265
266
267
268
    // === 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) {

269
270
271
      const FiniteElemSpace *rowFe = mat->getRowFeSpace();
      const FiniteElemSpace *colFe = mat->getColFeSpace();

Thomas Witkowski's avatar
Thomas Witkowski committed
272
      // Global index of the current row DOF.
273
274
      int globalRowDof = (*dofMap)[rowFe][*cursor].global;

Thomas Witkowski's avatar
Thomas Witkowski committed
275
      // Test if the current row DOF is a periodic DOF.
276
      bool periodicRow = perMap.isPeriodic(rowFe, globalRowDof);
277

Thomas Witkowski's avatar
Thomas Witkowski committed
278
279
280
      if (!periodicRow) {
	// === Row DOF index is not periodic. ===

281
	// Get PETSc's mat row index.
282
	int rowIndex = dofToMatIndex.get(nRowMat, globalRowDof);
Thomas Witkowski's avatar
Thomas Witkowski committed
283
284
285
286
287
288
289
290

	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.
291
	  int globalColDof = (*dofMap)[colFe][col(*icursor)].global;
Thomas Witkowski's avatar
Thomas Witkowski committed
292
	  // Test if the current col dof is a periodic dof.
293
	  bool periodicCol = perMap.isPeriodic(colFe, globalColDof);
294
	  // Get PETSc's mat col index.
295
	  int colIndex = dofToMatIndex.get(nColMat, globalColDof);
Thomas Witkowski's avatar
Thomas Witkowski committed
296
297
298
299
300
301
302
303
304
305
306
307
308
309

	  // 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;
310
	    std::set<int>& perColAsc = perMap.getAssociations(colFe, globalColDof);
Thomas Witkowski's avatar
Thomas Witkowski committed
311
312
	    for (std::set<int>::iterator it = perColAsc.begin(); 
		 it != perColAsc.end(); ++it)
313
	      if (meshDistributor->getElementObjectDb().isValidPeriodicType(*it))
Thomas Witkowski's avatar
Thomas Witkowski committed
314
315
316
317
318
319
320
321
322
323
324
325
326
327
		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);
328

Thomas Witkowski's avatar
Thomas Witkowski committed
329
330
331
332
333
334
	    // 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++) {
335
 		TEST_EXIT_DBG(perMap.isPeriodic(colFe, *it, newCols[i]))
Thomas Witkowski's avatar
Thomas Witkowski committed
336
337
338
 		  ("Wrong periodic DOF associations at boundary %d with DOF %d!\n",
		   *it, newCols[i]);

339
		newCols.push_back(perMap.map(colFe, *it, newCols[i]));
Thomas Witkowski's avatar
Thomas Witkowski committed
340
341
342
343
	      }
	    }

	    for (unsigned int i = 0; i < newCols.size(); i++) {
344
	      cols.push_back(dofToMatIndex.get(nColMat, newCols[i]));
Thomas Witkowski's avatar
Thomas Witkowski committed
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
	      values.push_back(scaledValue);	      
	    }
	  }
	}

	MatSetValues(petscMatrix, 1, &rowIndex, cols.size(), 
		     &(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.
366
	  int globalColDof = (*dofMap)[colFe][col(*icursor)].global;
Thomas Witkowski's avatar
Thomas Witkowski committed
367
368
369
370
371
372
373
374
375
376

	  // 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;

377
378
	  if (perMap.isPeriodic(colFe, globalColDof)) {
	    std::set<int>& perColAsc = perMap.getAssociations(colFe, globalColDof);
Thomas Witkowski's avatar
Thomas Witkowski committed
379
380
	    for (std::set<int>::iterator it = perColAsc.begin(); 
		 it != perColAsc.end(); ++it)
381
	      if (meshDistributor->getElementObjectDb().isValidPeriodicType(*it))
Thomas Witkowski's avatar
Thomas Witkowski committed
382
383
384
		perAsc.insert(*it);
	  }

385
	  std::set<int>& perRowAsc = perMap.getAssociations(rowFe, globalRowDof);
Thomas Witkowski's avatar
Thomas Witkowski committed
386
387
	  for (std::set<int>::iterator it = perRowAsc.begin(); 
	       it != perRowAsc.end(); ++it)
388
	    if (meshDistributor->getElementObjectDb().isValidPeriodicType(*it))
Thomas Witkowski's avatar
Thomas Witkowski committed
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
	      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));

404
405
	  // Then, traverse the periodic associations of the row and column
	  // indices and create the corresponding entries.
Thomas Witkowski's avatar
Thomas Witkowski committed
406
407
408
409
	  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;
410

411
412
	      if (perMap.isPeriodic(rowFe, *it, entry[i].first))
		perRowDof = perMap.map(rowFe, *it, entry[i].first);
Thomas Witkowski's avatar
Thomas Witkowski committed
413
414
415
416
	      else
		perRowDof = entry[i].first;

	      int perColDof;
417
418
	      if (perMap.isPeriodic(colFe, *it, entry[i].second))
		perColDof = perMap.map(colFe, *it, entry[i].second);
Thomas Witkowski's avatar
Thomas Witkowski committed
419
420
421
422
423
424
425
426
427
428
429
	      else
		perColDof = entry[i].second;	      	      
	      

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


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

430
	  for (unsigned int i = 0; i < entry.size(); i++) {
431
432
	    int rowIdx = dofToMatIndex.get(nRowMat, entry[i].first);
	    int colIdx = dofToMatIndex.get(nColMat, entry[i].second);
Thomas Witkowski's avatar
Thomas Witkowski committed
433

434
435
	    colsMap[rowIdx].push_back(colIdx);
	    valsMap[rowIdx].push_back(scaledValue);
Thomas Witkowski's avatar
Thomas Witkowski committed
436
437
438
439
440
441
442
443
444
445
446
447
	  }
	}


	// === 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;
448

Thomas Witkowski's avatar
Thomas Witkowski committed
449
450
451
452
453
454
455
456
	  MatSetValues(petscMatrix, 1, &rowIndex, rowIt->second.size(),
		       &(rowIt->second[0]), &(valsMap[rowIt->first][0]), ADD_VALUES);
	}
      }
    }
  }


457
458
  void PetscSolverGlobalMatrix::setDofVector(Vec& petscVec, 
					     DOFVector<double>* vec, 
459
					     int nRowVec, 
460
					     bool rankOnly)
Thomas Witkowski's avatar
Thomas Witkowski committed
461
462
463
  {
    FUNCNAME("PetscSolverGlobalMatrix::setDofVector()");

464
    const FiniteElemSpace *feSpace = vec->getFeSpace();
465
    PeriodicMap &perMap = meshDistributor->getPeriodicMap();
466

Thomas Witkowski's avatar
Thomas Witkowski committed
467
468
469
    // Traverse all used DOFs in the dof vector.
    DOFVector<double>::Iterator dofIt(vec, USED_DOFS);
    for (dofIt.reset(); !dofIt.end(); ++dofIt) {
470
      if (rankOnly && !(*dofMap)[feSpace].isRankDof(dofIt.getDOFIndex()))
Thomas Witkowski's avatar
Thomas Witkowski committed
471
472
473
474
	continue;

      // Calculate global row index of the DOF.
      DegreeOfFreedom globalRowDof = 
475
	(*dofMap)[feSpace][dofIt.getDOFIndex()].global;
476
      // Get PETSc's mat index of the row DOF.
477
      int index = dofToMatIndex.get(nRowVec, globalRowDof);
Thomas Witkowski's avatar
Thomas Witkowski committed
478

479
480
      if (perMap.isPeriodic(feSpace, globalRowDof)) {
	std::set<int>& perAsc = perMap.getAssociations(feSpace, globalRowDof);
Thomas Witkowski's avatar
Thomas Witkowski committed
481
482
483
	double value = *dofIt / (perAsc.size() + 1.0);
	VecSetValues(petscVec, 1, &index, &value, ADD_VALUES);

484
485
486
	for (std::set<int>::iterator perIt = perAsc.begin(); 
	     perIt != perAsc.end(); ++perIt) {
	  int mappedDof = perMap.map(feSpace, *perIt, globalRowDof);
487
	  int mappedIndex = dofToMatIndex.get(nRowVec, mappedDof);
Thomas Witkowski's avatar
Thomas Witkowski committed
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
	  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");

506
    vector<const FiniteElemSpace*> feSpaces = getFeSpaces(mat);
507
508
509
    int nRankRows = dofMap->getRankDofs();
    int rankStartIndex = dofMap->getStartDofs();

Thomas Witkowski's avatar
Thomas Witkowski committed
510
511
512
513
514
515
516
517
518
519
520
521
522
    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;

523
524
525
526
    // 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
527
528
529
530
    map<int, MatrixNnzEntry> sendMatrixEntry;

    // Maps to each DOF that must be send to another rank the rank number of the
    // receiving rank.
531
    map<pair<DegreeOfFreedom, int>, int> sendDofToRank;
Thomas Witkowski's avatar
Thomas Witkowski committed
532

533

534
535
536
537
538
539
540
541
542
543
    // 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++) {
      for (DofComm::Iterator it(meshDistributor->getRecvDofs(), feSpaces[i]);
	   !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
544
545
    }

546
    // Create list of ranks from which we receive data from.
Thomas Witkowski's avatar
Thomas Witkowski committed
547
    std::set<int> recvFromRank;
548
549
550
551
552
553
554
    for (unsigned int i = 0; i < feSpaces.size(); i++) 
      for (DofComm::Iterator it(meshDistributor->getSendDofs(), feSpaces[i]);
	   !it.end(); it.nextRank())
	recvFromRank.insert(it.getRank());


    // === Traverse matrices to create nnz data. ===
Thomas Witkowski's avatar
Thomas Witkowski committed
555

556
    int nComponents = mat->getNumRows();
Thomas Witkowski's avatar
Thomas Witkowski committed
557
558
559
560
561
    for (int i = 0; i < nComponents; i++) {
      for (int j = 0; j < nComponents; j++) {
 	if (!(*mat)[i][j])
	  continue;

562
563
564
565
566
	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
567
568
569
570
571
572
573
574
575
576
	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) {
577
	  int globalRowDof = (*dofMap)[feSpaces[i]][*cursor].global;
578

579
	  // The corresponding global matrix row index of the current row DOF.
580
	  int petscRowIdx = dofToMatIndex.get(i, globalRowDof);
581
582
	  if ((*dofMap)[feSpaces[i]].isRankDof(*cursor)) {
    	    
583
584
	    // === 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
585
586
	    
	    // This is the local row index of the local PETSc matrix.
587
	    int localPetscRowIdx = petscRowIdx - rankStartIndex;
Thomas Witkowski's avatar
Thomas Witkowski committed
588
589
590
	    
	    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",
591
592
	       *cursor,
	       (*dofMap)[feSpaces[i]][*cursor].global,
Thomas Witkowski's avatar
Thomas Witkowski committed
593
594
	       petscRowIdx, 
	       localPetscRowIdx, 
595
	       rankStartIndex,
Thomas Witkowski's avatar
Thomas Witkowski committed
596
597
598
599
600
601
602
	       nComponents, 
	       nRankRows);
	    
	    
	    // Traverse all non zero entries in this row.
	    for (icursor_type icursor = begin<nz>(cursor), 
		   icend = end<nz>(cursor); icursor != icend; ++icursor) {
603
	      int globalColDof = (*dofMap)[feSpaces[j]][col(*icursor)].global;
604
	      int petscColIdx = dofToMatIndex.get(j, globalColDof);
Thomas Witkowski's avatar
Thomas Witkowski committed
605
606
607
	      
	      if (value(*icursor) != 0.0 || petscRowIdx == petscColIdx) {
		// The row DOF is a rank DOF, if also the column is a rank DOF, 
608
609
610
611
		// 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
612
613
614
615
616
617
		  d_nnz[localPetscRowIdx]++;
		else
		  o_nnz[localPetscRowIdx]++;
	      }    
	    }
	  } else {
618
619
620
621
622
	    // === 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
623
624
	    
	    // Send all non zero entries to the member of the row DOF.
625
	    int sendToRank = sendDofToRank[make_pair(*cursor, i)];
Thomas Witkowski's avatar
Thomas Witkowski committed
626
627
628
629
	    
	    for (icursor_type icursor = begin<nz>(cursor), 
		   icend = end<nz>(cursor); icursor != icend; ++icursor) {
	      if (value(*icursor) != 0.0) {
630
631
		int globalColDof = 
		  (*dofMap)[feSpaces[j]][col(*icursor)].global;
632
		int petscColIdx = dofToMatIndex.get(j, globalColDof);
Thomas Witkowski's avatar
Thomas Witkowski committed
633
634
635
636
637
638
639
640
641
642
643
644
645
		
		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. ===

646
    StdMpi<MatrixNnzEntry> stdMpi(mpiComm, true);
Thomas Witkowski's avatar
Thomas Witkowski committed
647
648
649
650
651
652
653
    stdMpi.send(sendMatrixEntry);
    for (std::set<int>::iterator it = recvFromRank.begin(); 
	 it != recvFromRank.end(); ++it)
      stdMpi.recv(*it);
    stdMpi.startCommunication();


654
655
    // === 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
656
657
658
659
660
661
662
663

    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;

664
	  int localRowIdx = r - rankStartIndex;
Thomas Witkowski's avatar
Thomas Witkowski committed
665
666
667
668
669

	  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);
	  
670
	  if (c < rankStartIndex || c >= rankStartIndex + nRankRows)
Thomas Witkowski's avatar
Thomas Witkowski committed
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
	    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);
  }

689
690
691
692
693

  void PetscSolverGlobalMatrix::createGlobalDofMapping(vector<const FiniteElemSpace*> &feSpaces)
  {
    FUNCNAME("PetscSolverGlobalMatrix::createGlobalDofMapping()");

694
695
    int offset = dofMap->getStartDofs();
    Mesh *mesh = feSpaces[0]->getMesh();
696

697
    dofToMatIndex.clear();
698
  
699
700
701
702
703
704
705
    for (unsigned int i = 0; i < feSpaces.size(); i++) {

      // === Create indices for all DOFs in rank' domain. ===
      std::set<const DegreeOfFreedom*> rankDofSet;
      mesh->getAllDofs(feSpaces[i], rankDofSet);
      for (std::set<const DegreeOfFreedom*>::iterator it = rankDofSet.begin();
	   it != rankDofSet.end(); ++it)
706
707
      if ((*dofMap)[feSpaces[i]].isRankDof(**it)) {
	int globalIndex = (*dofMap)[feSpaces[i]][**it].global;
708

709
710
	int globalMatIndex =  
	  globalIndex - (*dofMap)[feSpaces[i]].rStartDofs + offset;
711

712
	  dofToMatIndex.add(i, globalIndex, globalMatIndex);
713
714
	}

715
      // === Communicate interior boundary DOFs between domains. ===
716

717
      StdMpi<vector<int> > stdMpi(mpiComm);
718
719
720
721
722
723
    
      for (DofComm::Iterator it(meshDistributor->getSendDofs(), feSpaces[i]);
	   !it.end(); it.nextRank()) {
	vector<DegreeOfFreedom> sendGlobalDofs;

	for (; !it.endDofIter(); it.nextDof()) {
724
	  int globalIndex = (*dofMap)[feSpaces[i]][it.getDofIndex()].global;
725
	  int globalMatIndex = dofToMatIndex.get(i, globalIndex);
726
	  sendGlobalDofs.push_back(globalMatIndex);
727
728
729
730
731
732
733
734
735
736
737
738
739
740
	}

	stdMpi.send(it.getRank(), sendGlobalDofs);
      }

      for (DofComm::Iterator it(meshDistributor->getRecvDofs(), feSpaces[i]);
	   !it.end(); it.nextRank())
	stdMpi.recv(it.getRank());

      stdMpi.startCommunication();

      for (DofComm::Iterator it(meshDistributor->getRecvDofs(), feSpaces[i]);
	   !it.end(); it.nextRank())
	for (; !it.endDofIter(); it.nextDof()) {
741
	  int globalIndex = (*dofMap)[feSpaces[i]][it.getDofIndex()].global;
742
743
	  int globalMatIndex = 
	    stdMpi.getRecvData(it.getRank())[it.getDofCounter()];
744

745
	  dofToMatIndex.add(i, globalIndex, globalMatIndex);
746
747
748
749
750
751
752
753
754
755
756
757
	}


      // === Communicate DOFs on periodic boundaries. ===

      stdMpi.clear();

      for (DofComm::Iterator it(meshDistributor->getPeriodicDofs(), feSpaces[i]);
	   !it.end(); it.nextRank()) {
	vector<DegreeOfFreedom> sendGlobalDofs;
	
	for (; !it.endDofIter(); it.nextDof()) {
758
	  int ind0 = (*dofMap)[feSpaces[i]][it.getDofIndex()].global;
759
	  int ind1 = dofToMatIndex.get(i, ind0);
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776

	  sendGlobalDofs.push_back(ind0);
	  sendGlobalDofs.push_back(ind1);
	}

	stdMpi.send(it.getRank(), sendGlobalDofs);
	stdMpi.recv(it.getRank());
      }

      stdMpi.startCommunication();

      for (DofComm::Iterator it(meshDistributor->getPeriodicDofs(), feSpaces[i]);
	   !it.end(); it.nextRank())
	for (; !it.endDofIter(); it.nextDof()) {
	  int ind0 = stdMpi.getRecvData(it.getRank())[it.getDofCounter() * 2];
	  int ind1 = stdMpi.getRecvData(it.getRank())[it.getDofCounter() * 2 + 1];

777
	  dofToMatIndex.add(i, ind0, ind1);
778
779
780
	}
      
      // === Update offset. ===
781
      offset += (*dofMap)[feSpaces[i]].nRankDofs;
782
783
    }
  }
Thomas Witkowski's avatar
Thomas Witkowski committed
784
}