PetscSolverGlobalMatrix.cc 35.6 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
    MatCreateAIJ(mpiCommGlobal, nRankRows, nRankRows, 
74 75 76 77
		 nOverallRows, nOverallRows,
		 0, d_nnz, 0, o_nnz, &matIntInt);

    MatSetOption(matIntInt, MAT_NEW_NONZERO_ALLOCATION_ERR, PETSC_FALSE);
78

Thomas Witkowski's avatar
Thomas Witkowski committed
79 80 81 82 83 84
#if (DEBUG != 0)
    MSG("Fill petsc matrix 1 needed %.5f seconds\n", MPI::Wtime() - wtime);
#endif

#if (DEBUG != 0)
    int a, b;
85 86 87
    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
88 89 90 91 92 93
      ("Wrong matrix ownership range!\n");
#endif


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

94
    int nComponents = mat->getNumRows();
Thomas Witkowski's avatar
Thomas Witkowski committed
95 96 97
    for (int i = 0; i < nComponents; i++)
      for (int j = 0; j < nComponents; j++)
	if ((*mat)[i][j])
98
	  setDofMatrix((*mat)[i][j], i, j);
Thomas Witkowski's avatar
Thomas Witkowski committed
99 100 101 102 103

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

104 105
    MatAssemblyBegin(matIntInt, MAT_FINAL_ASSEMBLY);
    MatAssemblyEnd(matIntInt, MAT_FINAL_ASSEMBLY);
Thomas Witkowski's avatar
Thomas Witkowski committed
106

107 108 109

    // === Remove Dirichlet BC DOFs. ===

110
    //    removeDirichletBcDofs(mat);
111 112
    

113
    // === Init PETSc solver. ===
114

115 116 117 118 119 120 121 122
    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);
123

124 125
    createFieldSplit(pcInterior);

126 127
    // Do not delete the solution vector, use it for the initial guess.
    if (!zeroStartVector)
128
      KSPSetInitialGuessNonzero(kspInterior, PETSC_TRUE);
129

130 131 132 133
    MSG("Fill petsc matrix needed %.5f seconds\n", MPI::Wtime() - wtime);
  }


134 135 136 137 138 139 140 141 142 143 144 145 146
  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 {
147 148 149 150
      MatCreateAIJ(mpiCommLocal, 
		   nRowsRankInterior, nRowsRankInterior,
		   nRowsOverallInterior, nRowsOverallInterior,
		   60, PETSC_NULL, 60, PETSC_NULL, &matIntInt);
151 152 153 154 155 156
    }

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

157 158 159 160
      MatCreateAIJ(mpiCommGlobal,
		   nRowsRankCoarse, nRowsRankCoarse,
		   nRowsOverallCoarse, nRowsOverallCoarse,
		   60, PETSC_NULL, 60, PETSC_NULL, &matCoarseCoarse);
161
      
162 163 164 165
      MatCreateAIJ(mpiCommGlobal,
		   nRowsRankCoarse, nRowsRankInterior,
		   nRowsOverallCoarse, nGlobalOverallInterior,
		   60, PETSC_NULL, 60, PETSC_NULL, &matCoarseInt);
166
      
167 168 169 170
      MatCreateAIJ(mpiCommGlobal,
		   nRowsRankInterior, nRowsRankCoarse,
		   nGlobalOverallInterior, nRowsOverallCoarse,
		   60, PETSC_NULL, 60, PETSC_NULL, &matIntCoarse);
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
    }

    // === 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);
    }


310 311
    // === Remove Dirichlet BC DOFs. ===

312
    //    removeDirichletBcDofs(mat);
313

314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330
    // === 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);  
  }


331 332 333
  void PetscSolverGlobalMatrix::fillPetscRhs(SystemVector *vec)
  {
    FUNCNAME("PetscSolverGlobalMatrix::fillPetscRhs()");
Thomas Witkowski's avatar
Thomas Witkowski committed
334

Thomas Witkowski's avatar
Thomas Witkowski committed
335 336 337 338
    VecCreateMPI(mpiCommGlobal, 
		 interiorMap->getRankDofs(), 
		 nGlobalOverallInterior,
		 &rhsInterior);
339

Thomas Witkowski's avatar
Thomas Witkowski committed
340 341 342 343 344 345
    if (coarseSpaceMap) 
      VecCreateMPI(mpiCommGlobal, 
		   coarseSpaceMap->getRankDofs(), 
		   coarseSpaceMap->getOverallDofs(),
		   &rhsCoarseSpace);
    
346
    TEST_EXIT_DBG(vec)("No DOF vector defined!\n");
347
    TEST_EXIT_DBG(interiorMap)("No parallel DOF map defined!\n");
Thomas Witkowski's avatar
Thomas Witkowski committed
348
    
349

Thomas Witkowski's avatar
Thomas Witkowski committed
350
    // === Transfer values from DOF vector to the PETSc vector. === 
Thomas Witkowski's avatar
Thomas Witkowski committed
351
    if (coarseSpaceMap) {
Thomas Witkowski's avatar
Thomas Witkowski committed
352 353
      for (int i = 0; i < vec->getSize(); i++)
	setDofVector(rhsInterior, rhsCoarseSpace, vec->getDOFVector(i), i);
Thomas Witkowski's avatar
Thomas Witkowski committed
354 355 356 357
    } else {
      for (int i = 0; i < vec->getSize(); i++)
	setDofVector(rhsInterior, vec->getDOFVector(i), i);
    }
Thomas Witkowski's avatar
Thomas Witkowski committed
358

359 360
    VecAssemblyBegin(rhsInterior);
    VecAssemblyEnd(rhsInterior);
361

Thomas Witkowski's avatar
Thomas Witkowski committed
362 363 364 365 366
    if (coarseSpaceMap) {
      VecAssemblyBegin(rhsCoarseSpace);
      VecAssemblyEnd(rhsCoarseSpace);
    }

367 368 369

    // === Remove Dirichlet BC DOFs. ===
    //    removeDirichletBcDofs(vec);
Thomas Witkowski's avatar
Thomas Witkowski committed
370
 
Thomas Witkowski's avatar
Thomas Witkowski committed
371 372 373
  }


374 375
  void PetscSolverGlobalMatrix::solvePetscMatrix(SystemVector &vec, 
						 AdaptInfo *adaptInfo)
Thomas Witkowski's avatar
Thomas Witkowski committed
376 377 378 379 380 381 382 383 384 385
  {
    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++)
386
	setDofVector(petscSolVec, vec.getDOFVector(i), i, true);
Thomas Witkowski's avatar
Thomas Witkowski committed
387 388 389 390 391
      
      VecAssemblyBegin(petscSolVec);
      VecAssemblyEnd(petscSolVec);
    }

392

Thomas Witkowski's avatar
Thomas Witkowski committed
393 394
    MatNullSpace matNullspace;
    Vec nullspaceBasis;
395

Thomas Witkowski's avatar
Thomas Witkowski committed
396 397
    if (nullspace.size() > 0 || hasConstantNullspace) {
      TEST_EXIT_DBG(nullspace.size() <= 1)("Not yet implemented!\n");
398

Thomas Witkowski's avatar
Thomas Witkowski committed
399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416
      if (nullspace.size() > 0) {
	VecDuplicate(petscSolVec, &nullspaceBasis);
	for (int i = 0; i < nComponents; i++)
	  setDofVector(nullspaceBasis, nullspace[0]->getDOFVector(i), i, true);
	
	VecAssemblyBegin(nullspaceBasis);
	VecAssemblyEnd(nullspaceBasis);
	
	MatNullSpaceCreate(mpiCommGlobal, (hasConstantNullspace ? PETSC_TRUE : PETSC_FALSE), 
			   1, &nullspaceBasis, &matNullspace);

	MatMult(matIntInt, nullspaceBasis, petscSolVec);
	PetscReal n;
	VecNorm(petscSolVec, NORM_2, &n);
	MSG("NORM IS: %e\n", n);
      } else {
	MatNullSpaceCreate(mpiCommGlobal, PETSC_TRUE, 0, PETSC_NULL, &matNullspace);
      }
417

Thomas Witkowski's avatar
Thomas Witkowski committed
418
      KSPSetNullSpace(kspInterior, matNullspace);
419

Thomas Witkowski's avatar
Thomas Witkowski committed
420
      // === Remove null space, if requested. ===
421

Thomas Witkowski's avatar
Thomas Witkowski committed
422 423 424 425 426 427 428 429
      if (removeRhsNullspace) {
	TEST_EXIT_DBG(coarseSpaceMap == NULL)("Not supported!\n");
	
	MatNullSpaceRemove(matNullspace, rhsInterior, PETSC_NULL);
      }
    } else {
      TEST_EXIT(removeRhsNullspace == false)
	("No nullspace provided that should be removed from rhs!\n");
430 431
    }

Thomas Witkowski's avatar
Thomas Witkowski committed
432

Thomas Witkowski's avatar
Thomas Witkowski committed
433
    // PETSc.
434
    solve(rhsInterior, petscSolVec);
Thomas Witkowski's avatar
Thomas Witkowski committed
435

436 437

    if (nullspace.size() > 0) {
Thomas Witkowski's avatar
Thomas Witkowski committed
438
      MatNullSpaceDestroy(&matNullspace);
439 440 441 442
      VecDestroy(&nullspaceBasis);
    }


Thomas Witkowski's avatar
Thomas Witkowski committed
443 444
    // === Transfere values from PETSc's solution vectors to the DOF vectors. ===
    PetscScalar *vecPointer;
445
    VecGetArray(petscSolVec, &vecPointer);    
Thomas Witkowski's avatar
Thomas Witkowski committed
446

447
    int c = 0;
Thomas Witkowski's avatar
Thomas Witkowski committed
448
    for (int i = 0; i < nComponents; i++) {
449
      DOFVector<double> &dv = *(vec.getDOFVector(i));
450

451
      DofMap& d = (*interiorMap)[dv.getFeSpace()].getMap();
452 453 454
      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
455 456 457 458 459 460 461 462
    }

    VecRestoreArray(petscSolVec, &vecPointer);


    // === Synchronize DOFs at common DOFs, i.e., DOFs that correspond to ===
    // === more than one partition.                                       ===
    meshDistributor->synchVector(vec);
463 464 465
  }


466 467 468 469
  void PetscSolverGlobalMatrix::solveGlobal(Vec &rhs, Vec &sol)
  {
    FUNCNAME("PetscSolverGlobalMatrix::solveGlobal()");

470 471 472
    double wtime = MPI::Wtime();
    double t0 = 0.0, t1 = 0.0;

473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491
    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);

492 493 494
    t0 = MPI::Wtime() - wtime;

    wtime = MPI::Wtime();
495
    KSPSolve(kspInterior, tmp, tmp);
496 497 498
    t1 = MPI::Wtime() - wtime;

    wtime = MPI::Wtime();
499 500 501 502 503 504 505 506 507 508 509

    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);
510 511
    t0 += MPI::Wtime() - wtime;

512
    //    MSG("TIMEING: %.5f %.5f\n", t0, t1);
513 514 515
  }


516 517 518 519
  void PetscSolverGlobalMatrix::destroyMatrixData()
  {
    FUNCNAME("PetscSolverGlobalMatrix::destroyMatrixData()");

520 521
    MatDestroy(&matIntInt);
    KSPDestroy(&kspInterior);
Thomas Witkowski's avatar
Thomas Witkowski committed
522

523 524 525 526 527 528 529
    if (coarseSpaceMap) {
      MatDestroy(&matCoarseCoarse);
      MatDestroy(&matCoarseInt);
      MatDestroy(&matIntCoarse);
    }

    if (petscSolVec != PETSC_NULL) {
Thomas Witkowski's avatar
Thomas Witkowski committed
530
      VecDestroy(&petscSolVec);
531 532
      petscSolVec = PETSC_NULL;
    }
Thomas Witkowski's avatar
Thomas Witkowski committed
533 534 535
  }


536 537 538 539
  void PetscSolverGlobalMatrix::destroyVectorData()
  {
    FUNCNAME("PetscSolverGlobalMatrix::destroyVectorData()");

540
    VecDestroy(&rhsInterior);
541 542 543

    if (coarseSpaceMap)
      VecDestroy(&rhsCoarseSpace);
544 545 546
  }


547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582
  void PetscSolverGlobalMatrix::createFieldSplit(PC pc)
  {
    FUNCNAME("PetscSolverGlobalMatrix::createFieldSplit()");

    vector<string> isNames;
    Parameters::get("parallel->solver->is blocks", isNames);

    int nBlocks = isNames.size();
    if (nBlocks == 0)
      return;

    for (int i = 0; i < nBlocks; i++) {
      MSG("Create for block %s\n", isNames[i].c_str());

      vector<int> blockComponents;
      Parameters::get("parallel->solver->is block " + lexical_cast<string>(i),
		      blockComponents);
      int nComponents = static_cast<int>(blockComponents.size());

      TEST_EXIT(nComponents > 0)("No is block for block %d defined!\n", i);

      // Check if blocks are continous
      for (int j = 0; j < nComponents; j++) {
	TEST_EXIT(blockComponents[j] == blockComponents[0] + j)
	  ("Does not yet support not continous IS blocks! Block %s\n", 
	   isNames[i].c_str());
      }

      IS is;
      interiorMap->createIndexSet(is, blockComponents[0], nComponents);
      PCFieldSplitSetIS(pc, isNames[i].c_str(), is);
      ISDestroy(&is);
    }
  }


583 584
  void PetscSolverGlobalMatrix::setDofMatrix(DOFMatrix* mat,
					     int nRowMat, int nColMat)
Thomas Witkowski's avatar
Thomas Witkowski committed
585 586 587 588 589 590
  {
    FUNCNAME("PetscSolverGlobalMatrix::setDofMatrix()");

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

    using mtl::tag::row; using mtl::tag::nz; using mtl::begin; using mtl::end;
591
    namespace traits = mtl::traits;
Thomas Witkowski's avatar
Thomas Witkowski committed
592 593 594 595 596 597 598 599 600 601 602 603 604 605 606
    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;

607 608 609
    // Get periodic mapping object
    PeriodicMap &perMap = meshDistributor->getPeriodicMap();

Thomas Witkowski's avatar
Thomas Witkowski committed
610 611 612 613 614 615
    // === 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) {

616 617 618
      const FiniteElemSpace *rowFe = mat->getRowFeSpace();
      const FiniteElemSpace *colFe = mat->getColFeSpace();

Thomas Witkowski's avatar
Thomas Witkowski committed
619
      // Global index of the current row DOF.
620
      int globalRowDof = (*interiorMap)[rowFe][*cursor].global;
621

Thomas Witkowski's avatar
Thomas Witkowski committed
622
      // Test if the current row DOF is a periodic DOF.
623
      bool periodicRow = perMap.isPeriodic(rowFe, globalRowDof);
624

Thomas Witkowski's avatar
Thomas Witkowski committed
625 626 627
      if (!periodicRow) {
	// === Row DOF index is not periodic. ===

628
	// Get PETSc's mat row index.
629
	int rowIndex = interiorMap->getMatIndex(nRowMat, globalRowDof);
Thomas Witkowski's avatar
Thomas Witkowski committed
630 631 632 633 634 635 636 637

	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.
638
	  int globalColDof = (*interiorMap)[colFe][col(*icursor)].global;
Thomas Witkowski's avatar
Thomas Witkowski committed
639
	  // Test if the current col dof is a periodic dof.
640
	  bool periodicCol = perMap.isPeriodic(colFe, globalColDof);
641
	  // Get PETSc's mat col index.
642
	  int colIndex = interiorMap->getMatIndex(nColMat, globalColDof);
Thomas Witkowski's avatar
Thomas Witkowski committed
643 644 645 646 647 648 649 650 651 652 653 654 655 656

	  // 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;
657
	    std::set<int>& perColAsc = perMap.getAssociations(colFe, globalColDof);
Thomas Witkowski's avatar
Thomas Witkowski committed
658 659
	    for (std::set<int>::iterator it = perColAsc.begin(); 
		 it != perColAsc.end(); ++it)
660
	      if (meshDistributor->getElementObjectDb().isValidPeriodicType(*it))
Thomas Witkowski's avatar
Thomas Witkowski committed
661 662 663 664 665 666 667 668 669 670 671 672 673 674
		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);
675

Thomas Witkowski's avatar
Thomas Witkowski committed
676 677 678 679 680 681
	    // 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++) {
682
 		TEST_EXIT_DBG(perMap.isPeriodic(colFe, *it, newCols[i]))
Thomas Witkowski's avatar
Thomas Witkowski committed
683 684 685
 		  ("Wrong periodic DOF associations at boundary %d with DOF %d!\n",
		   *it, newCols[i]);

686
		newCols.push_back(perMap.map(colFe, *it, newCols[i]));
Thomas Witkowski's avatar
Thomas Witkowski committed
687 688 689 690
	      }
	    }

	    for (unsigned int i = 0; i < newCols.size(); i++) {
691
	      cols.push_back(interiorMap->getMatIndex(nColMat, newCols[i]));
Thomas Witkowski's avatar
Thomas Witkowski committed
692 693 694 695 696
	      values.push_back(scaledValue);	      
	    }
	  }
	}

697
	MatSetValues(matIntInt, 1, &rowIndex, cols.size(), 
Thomas Witkowski's avatar
Thomas Witkowski committed
698 699 700 701 702 703 704 705 706 707 708 709 710 711 712
		     &(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.
713
	  int globalColDof = (*interiorMap)[colFe][col(*icursor)].global;
Thomas Witkowski's avatar
Thomas Witkowski committed
714 715 716 717 718 719 720 721 722 723

	  // 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;

724 725
	  if (perMap.isPeriodic(colFe, globalColDof)) {
	    std::set<int>& perColAsc = perMap.getAssociations(colFe, globalColDof);
Thomas Witkowski's avatar
Thomas Witkowski committed
726 727
	    for (std::set<int>::iterator it = perColAsc.begin(); 
		 it != perColAsc.end(); ++it)
728
	      if (meshDistributor->getElementObjectDb().isValidPeriodicType(*it))
Thomas Witkowski's avatar
Thomas Witkowski committed
729 730 731
		perAsc.insert(*it);
	  }

732
	  std::set<int>& perRowAsc = perMap.getAssociations(rowFe, globalRowDof);
Thomas Witkowski's avatar
Thomas Witkowski committed
733 734
	  for (std::set<int>::iterator it = perRowAsc.begin(); 
	       it != perRowAsc.end(); ++it)
735
	    if (meshDistributor->getElementObjectDb().isValidPeriodicType(*it))
Thomas Witkowski's avatar
Thomas Witkowski committed
736 737 738 739 740 741 742 743 744 745 746 747 748 749 750
	      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));

751 752
	  // Then, traverse the periodic associations of the row and column
	  // indices and create the corresponding entries.
Thomas Witkowski's avatar
Thomas Witkowski committed
753 754 755 756
	  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;
757

758 759
	      if (perMap.isPeriodic(rowFe, *it, entry[i].first))
		perRowDof = perMap.map(rowFe, *it, entry[i].first);
Thomas Witkowski's avatar
Thomas Witkowski committed
760 761 762 763
	      else
		perRowDof = entry[i].first;

	      int perColDof;
764 765
	      if (perMap.isPeriodic(colFe, *it, entry[i].second))
		perColDof = perMap.map(colFe, *it, entry[i].second);
Thomas Witkowski's avatar
Thomas Witkowski committed
766 767 768 769 770 771 772 773 774 775 776
	      else
		perColDof = entry[i].second;	      	      
	      

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


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

777
	  for (unsigned int i = 0; i < entry.size(); i++) {
778 779
	    int rowIdx = interiorMap->getMatIndex(nRowMat, entry[i].first);
	    int colIdx = interiorMap->getMatIndex(nColMat, entry[i].second);
Thomas Witkowski's avatar
Thomas Witkowski committed
780

781 782
	    colsMap[rowIdx].push_back(colIdx);
	    valsMap[rowIdx].push_back(scaledValue);
Thomas Witkowski's avatar
Thomas Witkowski committed
783 784 785 786 787 788 789 790 791 792 793 794
	  }
	}


	// === 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;
795

796
	  MatSetValues(matIntInt, 1, &rowIndex, rowIt->second.size(),
Thomas Witkowski's avatar
Thomas Witkowski committed
797 798 799 800 801 802 803
		       &(rowIt->second[0]), &(valsMap[rowIt->first][0]), ADD_VALUES);
	}
      }
    }
  }


Thomas Witkowski's avatar
Thomas Witkowski committed
804 805
  void PetscSolverGlobalMatrix::setDofVector(Vec vecInterior, 
					     Vec vecCoarse,
806
					     DOFVector<double>* vec, 
807
					     int nRowVec, 
808
					     bool rankOnly)
Thomas Witkowski's avatar
Thomas Witkowski committed
809 810 811
  {
    FUNCNAME("PetscSolverGlobalMatrix::setDofVector()");

812
    const FiniteElemSpace *feSpace = vec->getFeSpace();
813
    PeriodicMap &perMap = meshDistributor->getPeriodicMap();
814

Thomas Witkowski's avatar
Thomas Witkowski committed
815 816 817
    // Traverse all used DOFs in the dof vector.
    DOFVector<double>::Iterator dofIt(vec, USED_DOFS);
    for (dofIt.reset(); !dofIt.end(); ++dofIt) {
818
      if (rankOnly && !(*interiorMap)[feSpace].isRankDof(dofIt.getDOFIndex()))
Thomas Witkowski's avatar
Thomas Witkowski committed
819 820
	continue;

821
      if (isCoarseSpace(feSpace, dofIt.getDOFIndex())) {
822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840
	TEST_EXIT_DBG(vecCoarse != PETSC_NULL)("Should not happen!\n");

	int index = coarseSpaceMap->getMatIndex(nRowVec, dofIt.getDOFIndex());
	VecSetValue(vecCoarse, index, *dofIt, ADD_VALUES);
      } else {
	// Calculate global row index of the DOF.
	DegreeOfFreedom globalRowDof = 
	  (*interiorMap)[feSpace][dofIt.getDOFIndex()].global;
	
	// Get PETSc's mat index of the row DOF.
	int index = 0;
	if (interiorMap->isMatIndexFromGlobal())
	  index = 
	    interiorMap->getMatIndex(nRowVec, globalRowDof) + rStartInterior;
	else
	  index =
	    interiorMap->getMatIndex(nRowVec, dofIt.getDOFIndex()) + rStartInterior;
	
	if (perMap.isPeriodic(feSpace, globalRowDof)) {
Thomas Witkowski's avatar
Thomas Witkowski committed
841 842 843
	  std::set<int>& perAsc = perMap.getAssociations(feSpace, globalRowDof);
	  double value = *dofIt / (perAsc.size() + 1.0);
	  VecSetValue(vecInterior, index, value, ADD_VALUES);
844

Thomas Witkowski's avatar
Thomas Witkowski committed
845 846 847 848 849
	  for (std::set<int>::iterator perIt = perAsc.begin(); 
	       perIt != perAsc.end(); ++perIt) {
	    int mappedDof = perMap.map(feSpace, *perIt, globalRowDof);
	    int mappedIndex = interiorMap->getMatIndex(nRowVec, mappedDof);
	    VecSetValue(vecInterior, mappedIndex, value, ADD_VALUES);
850 851 852 853
	  }	  
	} else {	  
	  // The DOF index is not periodic.
	  
Thomas Witkowski's avatar
Thomas Witkowski committed
854 855
	  VecSetValue(vecInterior, index, *dofIt, ADD_VALUES);
	}
Thomas Witkowski's avatar
Thomas Witkowski committed
856 857 858 859 860
      }
    }
  }


861 862 863 864
  void PetscSolverGlobalMatrix::removeDirichletBcDofs(Matrix<DOFMatrix*> *mat)
  {
    FUNCNAME("PetscSolverGlobalMatrix::removeDirichletBcDofs()");

Thomas Witkowski's avatar
Thomas Witkowski committed
865
#if 0
866 867 868 869
    vector<int> dofsInterior, dofsCoarse;

    int nComponents = mat->getNumRows();
    for (int i = 0; i < nComponents; i++) {
Thomas Witkowski's avatar
Thomas Witkowski committed
870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889
      for (int j = 0; j < nComponents; j++) {
	if ((*mat)[i][j]) {
	  const FiniteElemSpace *feSpace = (*mat)[i][j]->getRowFeSpace();
	    
	  std::set<DegreeOfFreedom> &dirichletDofs = *((*mat)[i][j]->getApplyDBCs());

	  MSG("DIRICHLET DOFS: %d %d -> %d\n", i, j, dirichletDofs.size());
	  
	  for (std::set<DegreeOfFreedom>::iterator it = dirichletDofs.begin();
	       it != dirichletDofs.end(); ++it) {
	    if (isCoarseSpace(feSpace, *it)) {
	      if ((*coarseSpaceMap)[feSpace].isRankDof(*it)) {
		int globalDof = (*coarseSpaceMap)[feSpace][*it].global;
		dofsCoarse.push_back(coarseSpaceMap->getMatIndex(i, globalDof));
	      }
	    } else {
	      if ((*interiorMap)[feSpace].isRankDof(*it)) {
		int globalDof = (*interiorMap)[feSpace][*it].global;
		dofsInterior.push_back(interiorMap->getMatIndex(i, globalDof));
	      }
890
	    }
891
	  }
Thomas Witkowski's avatar
Thomas Witkowski committed
892 893
	} else {
	  MSG("NO MAT DIAG in %d\n", i);
894 895 896 897 898 899 900 901 902 903
	}
      }
    }

    MatZeroRows(matIntInt, dofsInterior.size(), &(dofsInterior[0]), 1.0, 
		PETSC_NULL, PETSC_NULL);

    if (coarseSpaceMap != NULL)
      MatZeroRows(matCoarseCoarse, dofsCoarse.size(), &(dofsCoarse[0]), 1.0, 
		  PETSC_NULL, PETSC_NULL);
Thomas Witkowski's avatar
Thomas Witkowski committed
904
#endif
905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921
  }


  void PetscSolverGlobalMatrix::removeDirichletBcDofs(SystemVector *vec)
  {
    FUNCNAME("PetscSolverGlobalMatrix::removeDirichletBcDofs()");

    int nComponents = vec->getSize();
    for (int i = 0; i < nComponents; i++) {
      const FiniteElemSpace *feSpace = vec->getDOFVector(i)->getFeSpace();

      map<DegreeOfFreedom, double> &dirichletValues = 
	vec->getDOFVector(i)->getDirichletValues();

      for (map<DegreeOfFreedom, double>::iterator it = dirichletValues.begin();
	   it != dirichletValues.end(); ++it) {
	if (isCoarseSpace(feSpace, it->first)) {
922 923 924
	  if ((*coarseSpaceMap)[feSpace].isRankDof(it->first)) {
	    int globalDof = (*coarseSpaceMap)[feSpace][it->first].global;
	    VecSetValue(rhsCoarseSpace, coarseSpaceMap->getMatIndex(i, globalDof), 
925
			it->second, INSERT_VALUES);
926
	  }
927 928
	} else {
	  if ((*interiorMap)[feSpace].isRankDof(it->first)) {
929 930
	    int globalDof = (*interiorMap)[feSpace][it->first].global;
	    VecSetValue(rhsInterior, interiorMap->getMatIndex(i, globalDof),
931 932 933 934 935
			it->second, INSERT_VALUES);	    
	  }
	}
      }
    }
936 937 938 939 940 941 942 943

    VecAssemblyBegin(rhsInterior);
    VecAssemblyEnd(rhsInterior);

    if (coarseSpaceMap) {
      VecAssemblyBegin(rhsCoarseSpace);
      VecAssemblyEnd(rhsCoarseSpace);
    }
944 945 946
  }


Thomas Witkowski's avatar
Thomas Witkowski committed
947 948 949 950 951 952 953
  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");

954
    vector<const FiniteElemSpace*> feSpaces = getFeSpaces(mat);
955 956
    int nRankRows = interiorMap->getRankDofs();
    int rankStartIndex = interiorMap->getStartDofs();
957

Thomas Witkowski's avatar
Thomas Witkowski committed
958 959 960 961 962 963 964 965 966 967 968 969 970
    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;

971 972 973 974
    // 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
975 976 977 978
    map<int, MatrixNnzEntry> sendMatrixEntry;

    // Maps to each DOF that must be send to another rank the rank number of the
    // receiving rank.
979
    map<pair<DegreeOfFreedom, int>, int> sendDofToRank;
Thomas Witkowski's avatar
Thomas Witkowski committed
980

981

982 983 984
    // 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++) {
985
      for (DofComm::Iterator it(meshDistributor->getDofComm().getRecvDofs(), feSpaces[i]);
986 987 988 989 990 991
	   !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
992 993
    }

994
    // Create list of ranks from which we receive data from.
Thomas Witkowski's avatar
Thomas Witkowski committed
995
    std::set<int> recvFromRank;
996
    for (unsigned int i = 0; i < feSpaces.size(); i++) 
997
      for (DofComm::Iterator it(meshDistributor->getDofComm().getSendDofs(), feSpaces[i]);
998 999 1000 1001 1002
	   !it.end(); it.nextRank())
	recvFromRank.insert(it.getRank());


    // === Traverse matrices to create nnz data. ===
Thomas Witkowski's avatar
Thomas Witkowski committed
1003

1004
    int nComponents = mat->getNumRows();
Thomas Witkowski's avatar
Thomas Witkowski committed
1005 1006 1007 1008 1009
    for (int i = 0; i < nComponents; i++) {
      for (int j = 0; j < nComponents; j++) {
 	if (!(*mat)[i][j])
	  continue;

1010 1011 1012 1013 1014
	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
1015 1016 1017 1018 1019 1020 1021 1022 1023 1024
	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) {
1025
	  int globalRowDof = (*interiorMap)[feSpaces[i]][*cursor].global;
1026

1027
	  // The corresponding global matrix row index of the current row DOF.
1028 1029
	  int petscRowIdx = interiorMap->getMatIndex(i, globalRowDof);
	  if ((*interiorMap)[feSpaces[i]].isRankDof(*cursor)) {
1030
    	    
1031 1032
	    // === 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
1033 1034
	    
	    // This is the local row index of the local PETSc matrix.
1035
	    int localPetscRowIdx = petscRowIdx - rankStartIndex;
Thomas Witkowski's avatar
Thomas Witkowski committed
1036 1037
	    
	    TEST_EXIT_DBG(localPetscRowIdx >= 0 && localPetscRowIdx < nRankRows)
1038
	      ("Should not happen! \n Debug info: DOF = %d   globalRowIndx = %d   petscRowIdx = %d   localPetscRowIdx = %d   rStart = %d   compontens = %d from %d   nRankRows = %d\n",
1039
	       *cursor,
1040
	       (*interiorMap)[feSpaces[i]][*cursor].global,
Thomas Witkowski's avatar
Thomas Witkowski committed
1041 1042
	       petscRowIdx, 
	       localPetscRowIdx, 
1043
	       rankStartIndex,
1044
	       i,
Thomas Witkowski's avatar
Thomas Witkowski committed
1045 1046 1047 1048 1049 1050 1051
	       nComponents, 
	       nRankRows);
	    
	    
	    // Traverse all non zero entries in this row.
	    for (icursor_type icursor = begin<nz>(cursor), 
		   icend = end<nz>(cursor); icursor != icend; ++icursor) {
1052 1053
	      int globalColDof = (*interiorMap)[feSpaces[j]][col(*icursor)].global;
	      int petscColIdx = interiorMap->getMatIndex(j, globalColDof);
Thomas Witkowski's avatar
Thomas Witkowski committed
1054 1055 1056
	      
	      if (value(*icursor) != 0.0 || petscRowIdx == petscColIdx) {
		// The row DOF is a rank DOF, if also the column is a rank DOF, 
1057 1058 1059 1060
		// 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
1061 1062 1063 1064 1065 1066
		  d_nnz[localPetscRowIdx]++;
		else
		  o_nnz[localPetscRowIdx]++;
	      }    
	    }
	  } else {
1067 1068 1069 1070 1071
	    // === 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
1072 1073
	    
	    // Send all non zero entries to the member of the row DOF.
1074
	    int sendToRank = sendDofToRank[make_pair(*cursor, i)];
Thomas Witkowski's avatar
Thomas Witkowski committed
1075 1076 1077 1078
	    
	    for (icursor_type icursor = begin<nz>(cursor), 
		   icend = end<nz>(cursor); icursor != icend; ++icursor) {
	      if (value(*icursor) != 0.0) {
1079
		int globalColDof = 
1080 1081
		  (*interiorMap)[feSpaces[j]][col(*icursor)].global;
		int petscColIdx = interiorMap->getMatIndex(j, globalColDof);
Thomas Witkowski's avatar
Thomas Witkowski committed
1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094
		
		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. ===

1095
    StdMpi<MatrixNnzEntry> stdMpi(mpiCommGlobal, true);
Thomas Witkowski's avatar
Thomas Witkowski committed
1096 1097 1098 1099 1100 1101 1102
    stdMpi.send(sendMatrixEntry);
    for (std::set<int>::iterator it = recvFromRank.begin(); 
	 it != recvFromRank.end(); ++it)
      stdMpi.recv(*it);
    stdMpi.startCommunication();


1103 1104
    // === 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
1105 1106 1107 1108 1109 1110 1111 1112

    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;

1113
	  int localRowIdx = r - rankStartIndex;
Thomas Witkowski's avatar
Thomas Witkowski committed
1114 1115 1116 1117 1118

	  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);
	  
1119
	  if (c < rankStartIndex || c >= rankStartIndex + nRankRows)
Thomas Witkowski's avatar
Thomas Witkowski committed
1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138
	    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);
  }

}