PetscSolverGlobalMatrix.cc 36.4 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
  {
    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
    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);
122

123
    initPreconditioner(pcInterior);
124

125 126
    // Do not delete the solution vector, use it for the initial guess.
    if (!zeroStartVector)
127
      KSPSetInitialGuessNonzero(kspInterior, PETSC_TRUE);
128

Thomas Witkowski's avatar
Thomas Witkowski committed
129 130 131
#if (DEBUG != 0)
    MSG("Fill petsc matrix 3 needed %.5f seconds\n", MPI::Wtime() - wtime);
#endif
132 133 134
  }


135 136 137 138 139 140 141 142
  void PetscSolverGlobalMatrix::fillPetscMatrix(DOFMatrix *mat)
  {
    Matrix<DOFMatrix*> m(1, 1);
    m[0][0] = mat;
    fillPetscMatrix(&m);
  }


143 144 145 146 147 148 149 150 151 152 153 154 155
  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 {
156 157 158 159
      MatCreateAIJ(mpiCommLocal, 
		   nRowsRankInterior, nRowsRankInterior,
		   nRowsOverallInterior, nRowsOverallInterior,
		   60, PETSC_NULL, 60, PETSC_NULL, &matIntInt);
160 161 162 163 164 165
    }

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

166 167 168 169
      MatCreateAIJ(mpiCommGlobal,
		   nRowsRankCoarse, nRowsRankCoarse,
		   nRowsOverallCoarse, nRowsOverallCoarse,
		   60, PETSC_NULL, 60, PETSC_NULL, &matCoarseCoarse);
170
      
171 172 173 174
      MatCreateAIJ(mpiCommGlobal,
		   nRowsRankCoarse, nRowsRankInterior,
		   nRowsOverallCoarse, nGlobalOverallInterior,
		   60, PETSC_NULL, 60, PETSC_NULL, &matCoarseInt);
175
      
176 177 178 179
      MatCreateAIJ(mpiCommGlobal,
		   nRowsRankInterior, nRowsRankCoarse,
		   nGlobalOverallInterior, nRowsOverallCoarse,
		   60, PETSC_NULL, 60, PETSC_NULL, &matIntCoarse);
180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318
    }

    // === 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);
    }


319 320
    // === Remove Dirichlet BC DOFs. ===

321
    //    removeDirichletBcDofs(mat);
322

323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338
    // === 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);
    KSPGetPC(kspInterior, &pcInterior);
    PCSetType(pcInterior, PCLU);
    if (subdomainLevel == 0)
      PCFactorSetMatSolverPackage(pcInterior, MATSOLVERUMFPACK);
    else
      PCFactorSetMatSolverPackage(pcInterior, MATSOLVERMUMPS);
    KSPSetFromOptions(kspInterior);  
  }


339 340 341
  void PetscSolverGlobalMatrix::fillPetscRhs(SystemVector *vec)
  {
    FUNCNAME("PetscSolverGlobalMatrix::fillPetscRhs()");
Thomas Witkowski's avatar
Thomas Witkowski committed
342

Thomas Witkowski's avatar
Thomas Witkowski committed
343 344 345 346
    VecCreateMPI(mpiCommGlobal, 
		 interiorMap->getRankDofs(), 
		 nGlobalOverallInterior,
		 &rhsInterior);
347

Thomas Witkowski's avatar
Thomas Witkowski committed
348 349 350 351 352 353
    if (coarseSpaceMap) 
      VecCreateMPI(mpiCommGlobal, 
		   coarseSpaceMap->getRankDofs(), 
		   coarseSpaceMap->getOverallDofs(),
		   &rhsCoarseSpace);
    
354
    TEST_EXIT_DBG(vec)("No DOF vector defined!\n");
355
    TEST_EXIT_DBG(interiorMap)("No parallel DOF map defined!\n");
Thomas Witkowski's avatar
Thomas Witkowski committed
356
    
357

Thomas Witkowski's avatar
Thomas Witkowski committed
358
    // === Transfer values from DOF vector to the PETSc vector. === 
Thomas Witkowski's avatar
Thomas Witkowski committed
359
    if (coarseSpaceMap) {
Thomas Witkowski's avatar
Thomas Witkowski committed
360 361
      for (int i = 0; i < vec->getSize(); i++)
	setDofVector(rhsInterior, rhsCoarseSpace, vec->getDOFVector(i), i);
Thomas Witkowski's avatar
Thomas Witkowski committed
362 363 364 365
    } else {
      for (int i = 0; i < vec->getSize(); i++)
	setDofVector(rhsInterior, vec->getDOFVector(i), i);
    }
Thomas Witkowski's avatar
Thomas Witkowski committed
366

367 368
    VecAssemblyBegin(rhsInterior);
    VecAssemblyEnd(rhsInterior);
369

Thomas Witkowski's avatar
Thomas Witkowski committed
370 371 372 373 374
    if (coarseSpaceMap) {
      VecAssemblyBegin(rhsCoarseSpace);
      VecAssemblyEnd(rhsCoarseSpace);
    }

375 376 377

    // === Remove Dirichlet BC DOFs. ===
    //    removeDirichletBcDofs(vec);
Thomas Witkowski's avatar
Thomas Witkowski committed
378
 
Thomas Witkowski's avatar
Thomas Witkowski committed
379 380 381
  }


382 383
  void PetscSolverGlobalMatrix::solvePetscMatrix(SystemVector &vec, 
						 AdaptInfo *adaptInfo)
Thomas Witkowski's avatar
Thomas Witkowski committed
384 385 386 387 388 389 390 391 392 393
  {
    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++)
394
	setDofVector(petscSolVec, vec.getDOFVector(i), i, true);
Thomas Witkowski's avatar
Thomas Witkowski committed
395 396 397 398 399
      
      VecAssemblyBegin(petscSolVec);
      VecAssemblyEnd(petscSolVec);
    }

400

Thomas Witkowski's avatar
Thomas Witkowski committed
401 402
    MatNullSpace matNullspace;
    Vec nullspaceBasis;
403

404 405 406
    if (nullspace.size() > 0 || 
	hasConstantNullspace ||
	constNullspaceComponent.size() > 0) {
Thomas Witkowski's avatar
Thomas Witkowski committed
407
      TEST_EXIT_DBG(nullspace.size() <= 1)("Not yet implemented!\n");
408

409 410 411 412 413 414 415 416 417 418
      if (constNullspaceComponent.size() > 0) {
	nullspace.clear();
	SystemVector *basisVec = new SystemVector(vec);
	basisVec->set(0.0);
	for (unsigned int i = 0; i < constNullspaceComponent.size(); i++)
	  basisVec->getDOFVector(constNullspaceComponent[i])->set(1.0);

	nullspace.push_back(basisVec);
      } 

Thomas Witkowski's avatar
Thomas Witkowski committed
419 420 421 422 423 424 425
      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);
426 427

	VecNormalize(nullspaceBasis, PETSC_NULL);
Thomas Witkowski's avatar
Thomas Witkowski committed
428 429 430 431 432 433 434 435 436 437 438
	
	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);
      }
439

440
      MatSetNullSpace(matIntInt, matNullspace);
Thomas Witkowski's avatar
Thomas Witkowski committed
441
      KSPSetNullSpace(kspInterior, matNullspace);
442

Thomas Witkowski's avatar
Thomas Witkowski committed
443
      // === Remove null space, if requested. ===
444

Thomas Witkowski's avatar
Thomas Witkowski committed
445 446 447 448 449 450 451 452
      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");
453 454
    }

Thomas Witkowski's avatar
Thomas Witkowski committed
455

Thomas Witkowski's avatar
Thomas Witkowski committed
456
    // PETSc.
457
    solve(rhsInterior, petscSolVec);
Thomas Witkowski's avatar
Thomas Witkowski committed
458

459 460

    if (nullspace.size() > 0) {
Thomas Witkowski's avatar
Thomas Witkowski committed
461
      MatNullSpaceDestroy(&matNullspace);
462 463 464 465
      VecDestroy(&nullspaceBasis);
    }


Thomas Witkowski's avatar
Thomas Witkowski committed
466 467
    // === Transfere values from PETSc's solution vectors to the DOF vectors. ===
    PetscScalar *vecPointer;
468
    VecGetArray(petscSolVec, &vecPointer);    
Thomas Witkowski's avatar
Thomas Witkowski committed
469

470
    int c = 0;
Thomas Witkowski's avatar
Thomas Witkowski committed
471
    for (int i = 0; i < nComponents; i++) {
472
      DOFVector<double> &dv = *(vec.getDOFVector(i));
473

474
      DofMap& d = (*interiorMap)[dv.getFeSpace()].getMap();
475 476 477
      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
478 479 480
    }

    VecRestoreArray(petscSolVec, &vecPointer);
481 482 483
  }


484 485 486 487
  void PetscSolverGlobalMatrix::solveGlobal(Vec &rhs, Vec &sol)
  {
    FUNCNAME("PetscSolverGlobalMatrix::solveGlobal()");

488 489 490
    double wtime = MPI::Wtime();
    double t0 = 0.0, t1 = 0.0;

491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509
    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);

510 511 512
    t0 = MPI::Wtime() - wtime;

    wtime = MPI::Wtime();
513
    KSPSolve(kspInterior, tmp, tmp);
514 515 516
    t1 = MPI::Wtime() - wtime;

    wtime = MPI::Wtime();
517 518 519 520 521 522 523 524 525 526 527

    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);
528 529
    t0 += MPI::Wtime() - wtime;

530
    //    MSG("TIMEING: %.5f %.5f\n", t0, t1);
531 532 533
  }


534 535 536 537
  void PetscSolverGlobalMatrix::destroyMatrixData()
  {
    FUNCNAME("PetscSolverGlobalMatrix::destroyMatrixData()");

538 539
    exitPreconditioner(pcInterior);

540 541
    MatDestroy(&matIntInt);
    KSPDestroy(&kspInterior);
Thomas Witkowski's avatar
Thomas Witkowski committed
542

543 544 545 546 547 548 549
    if (coarseSpaceMap) {
      MatDestroy(&matCoarseCoarse);
      MatDestroy(&matCoarseInt);
      MatDestroy(&matIntCoarse);
    }

    if (petscSolVec != PETSC_NULL) {
Thomas Witkowski's avatar
Thomas Witkowski committed
550
      VecDestroy(&petscSolVec);
551 552
      petscSolVec = PETSC_NULL;
    }
Thomas Witkowski's avatar
Thomas Witkowski committed
553 554 555
  }


556 557 558 559
  void PetscSolverGlobalMatrix::destroyVectorData()
  {
    FUNCNAME("PetscSolverGlobalMatrix::destroyVectorData()");

560
    VecDestroy(&rhsInterior);
561 562 563

    if (coarseSpaceMap)
      VecDestroy(&rhsCoarseSpace);
564 565 566
  }


567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602
  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);
    }
  }


603 604 605 606 607 608 609 610 611 612 613 614 615 616 617
  void PetscSolverGlobalMatrix::initPreconditioner(PC pc)
  {
    FUNCNAME("PetscSolverGlobalMatrix::initPreconditioner()");

    PCSetFromOptions(pc);
    createFieldSplit(pc);
  }

 
  void PetscSolverGlobalMatrix::exitPreconditioner(PC pc)
  {
    FUNCNAME("PetscSolverGlobalMatrix::exitPreconditioner()");
  }


618 619
  void PetscSolverGlobalMatrix::setDofMatrix(DOFMatrix* mat,
					     int nRowMat, int nColMat)
Thomas Witkowski's avatar
Thomas Witkowski committed
620 621 622 623 624 625
  {
    FUNCNAME("PetscSolverGlobalMatrix::setDofMatrix()");

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

    using mtl::tag::row; using mtl::tag::nz; using mtl::begin; using mtl::end;
626
    namespace traits = mtl::traits;
Thomas Witkowski's avatar
Thomas Witkowski committed
627 628 629 630 631 632 633 634 635 636 637 638 639 640 641
    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;

642 643 644
    // Get periodic mapping object
    PeriodicMap &perMap = meshDistributor->getPeriodicMap();

Thomas Witkowski's avatar
Thomas Witkowski committed
645 646 647 648 649 650
    const FiniteElemSpace *rowFe = mat->getRowFeSpace();
    const FiniteElemSpace *colFe = mat->getColFeSpace();
    DofMap& rowMap = (*interiorMap)[rowFe].getMap();
    DofMap& colMap = (*interiorMap)[colFe].getMap();

    // === Traverse all rows of the DOF matrix and insert row wise the values ===
Thomas Witkowski's avatar
Thomas Witkowski committed
651 652 653 654 655
    // === to the PETSc matrix.                                               ===

    for (cursor_type cursor = begin<row>(mat->getBaseMatrix()), 
	   cend = end<row>(mat->getBaseMatrix()); cursor != cend; ++cursor) {
      // Global index of the current row DOF.
Thomas Witkowski's avatar
Thomas Witkowski committed
656
      int globalRowDof = rowMap[*cursor].global;
657

Thomas Witkowski's avatar
Thomas Witkowski committed
658
      // Test if the current row DOF is a periodic DOF.
659
      bool periodicRow = perMap.isPeriodic(rowFe, globalRowDof);
660

Thomas Witkowski's avatar
Thomas Witkowski committed
661 662 663
      if (!periodicRow) {
	// === Row DOF index is not periodic. ===

664
	// Get PETSc's mat row index.
665
	int rowIndex = interiorMap->getMatIndex(nRowMat, globalRowDof);
Thomas Witkowski's avatar
Thomas Witkowski committed
666 667 668 669 670 671 672 673

	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.
Thomas Witkowski's avatar
Thomas Witkowski committed
674
	  int globalColDof = colMap[col(*icursor)].global;
Thomas Witkowski's avatar
Thomas Witkowski committed
675
	  // Test if the current col dof is a periodic dof.
676
	  bool periodicCol = perMap.isPeriodic(colFe, globalColDof);
677
	  // Get PETSc's mat col index.
678
	  int colIndex = interiorMap->getMatIndex(nColMat, globalColDof);
Thomas Witkowski's avatar
Thomas Witkowski committed
679 680 681 682 683 684 685

	  // 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.
Thomas Witkowski's avatar
Thomas Witkowski committed
686 687
 	    cols.push_back(colIndex);
 	    values.push_back(value(*icursor));
Thomas Witkowski's avatar
Thomas Witkowski committed
688 689 690 691 692
	  } else {
	    // === Row index is not periodic, but column index is. ===

	    // Create set of all periodic associations of the column index.
	    std::set<int> perAsc;
693
	    std::set<int>& perColAsc = perMap.getAssociations(colFe, globalColDof);
Thomas Witkowski's avatar
Thomas Witkowski committed
694 695
	    for (std::set<int>::iterator it = perColAsc.begin(); 
		 it != perColAsc.end(); ++it)
696
	      if (meshDistributor->getElementObjectDb().isValidPeriodicType(*it))
Thomas Witkowski's avatar
Thomas Witkowski committed
697 698 699 700 701 702 703 704 705 706 707 708 709 710
		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);
711

Thomas Witkowski's avatar
Thomas Witkowski committed
712 713 714 715 716 717
	    // 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++) {
718
 		TEST_EXIT_DBG(perMap.isPeriodic(colFe, *it, newCols[i]))
Thomas Witkowski's avatar
Thomas Witkowski committed
719 720 721
 		  ("Wrong periodic DOF associations at boundary %d with DOF %d!\n",
		   *it, newCols[i]);

722
		newCols.push_back(perMap.map(colFe, *it, newCols[i]));
Thomas Witkowski's avatar
Thomas Witkowski committed
723 724 725 726
	      }
	    }

	    for (unsigned int i = 0; i < newCols.size(); i++) {
727
	      cols.push_back(interiorMap->getMatIndex(nColMat, newCols[i]));
Thomas Witkowski's avatar
Thomas Witkowski committed
728 729 730 731 732
	      values.push_back(scaledValue);	      
	    }
	  }
	}

Thomas Witkowski's avatar
Thomas Witkowski committed
733 734
  	MatSetValues(matIntInt, 1, &rowIndex, cols.size(), 
  		     &(cols[0]), &(values[0]), ADD_VALUES);	
Thomas Witkowski's avatar
Thomas Witkowski committed
735 736 737 738 739 740 741 742 743 744 745 746 747 748
      } 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.
749
	  int globalColDof = (*interiorMap)[colFe][col(*icursor)].global;
Thomas Witkowski's avatar
Thomas Witkowski committed
750 751 752 753 754 755 756 757 758 759

	  // 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;

760 761
	  if (perMap.isPeriodic(colFe, globalColDof)) {
	    std::set<int>& perColAsc = perMap.getAssociations(colFe, globalColDof);
Thomas Witkowski's avatar
Thomas Witkowski committed
762 763
	    for (std::set<int>::iterator it = perColAsc.begin(); 
		 it != perColAsc.end(); ++it)
764
	      if (meshDistributor->getElementObjectDb().isValidPeriodicType(*it))
Thomas Witkowski's avatar
Thomas Witkowski committed
765 766 767
		perAsc.insert(*it);
	  }

768
	  std::set<int>& perRowAsc = perMap.getAssociations(rowFe, globalRowDof);
Thomas Witkowski's avatar
Thomas Witkowski committed
769 770
	  for (std::set<int>::iterator it = perRowAsc.begin(); 
	       it != perRowAsc.end(); ++it)
771
	    if (meshDistributor->getElementObjectDb().isValidPeriodicType(*it))
Thomas Witkowski's avatar
Thomas Witkowski committed
772 773 774 775 776 777 778 779 780 781 782 783 784 785 786
	      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));

787 788
	  // Then, traverse the periodic associations of the row and column
	  // indices and create the corresponding entries.
Thomas Witkowski's avatar
Thomas Witkowski committed
789 790 791 792
	  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;
793

794 795
	      if (perMap.isPeriodic(rowFe, *it, entry[i].first))
		perRowDof = perMap.map(rowFe, *it, entry[i].first);
Thomas Witkowski's avatar
Thomas Witkowski committed
796 797 798 799
	      else
		perRowDof = entry[i].first;

	      int perColDof;
800 801
	      if (perMap.isPeriodic(colFe, *it, entry[i].second))
		perColDof = perMap.map(colFe, *it, entry[i].second);
Thomas Witkowski's avatar
Thomas Witkowski committed
802 803 804 805 806 807 808 809 810 811
	      else
		perColDof = entry[i].second;	      	      
	      
	      entry.push_back(make_pair(perRowDof, perColDof));
	    }
	  }


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

812
	  for (unsigned int i = 0; i < entry.size(); i++) {
813 814
	    int rowIdx = interiorMap->getMatIndex(nRowMat, entry[i].first);
	    int colIdx = interiorMap->getMatIndex(nColMat, entry[i].second);
Thomas Witkowski's avatar
Thomas Witkowski committed
815

816 817
	    colsMap[rowIdx].push_back(colIdx);
	    valsMap[rowIdx].push_back(scaledValue);
Thomas Witkowski's avatar
Thomas Witkowski committed
818 819 820 821 822 823 824 825 826 827 828 829
	  }
	}


	// === 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;
830

831
	  MatSetValues(matIntInt, 1, &rowIndex, rowIt->second.size(),
Thomas Witkowski's avatar
Thomas Witkowski committed
832 833 834 835 836 837 838
		       &(rowIt->second[0]), &(valsMap[rowIt->first][0]), ADD_VALUES);
	}
      }
    }
  }


Thomas Witkowski's avatar
Thomas Witkowski committed
839 840
  void PetscSolverGlobalMatrix::setDofVector(Vec vecInterior, 
					     Vec vecCoarse,
841
					     DOFVector<double>* vec, 
842
					     int nRowVec, 
843
					     bool rankOnly)
Thomas Witkowski's avatar
Thomas Witkowski committed
844 845 846
  {
    FUNCNAME("PetscSolverGlobalMatrix::setDofVector()");

847
    const FiniteElemSpace *feSpace = vec->getFeSpace();
848
    PeriodicMap &perMap = meshDistributor->getPeriodicMap();
849

Thomas Witkowski's avatar
Thomas Witkowski committed
850 851 852
    // Traverse all used DOFs in the dof vector.
    DOFVector<double>::Iterator dofIt(vec, USED_DOFS);
    for (dofIt.reset(); !dofIt.end(); ++dofIt) {
853

854
      if (rankOnly && !(*interiorMap)[feSpace].isRankDof(dofIt.getDOFIndex()))
Thomas Witkowski's avatar
Thomas Witkowski committed
855 856
	continue;

857
      if (isCoarseSpace(feSpace, dofIt.getDOFIndex())) {
858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874
	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;
875

876
	if (perMap.isPeriodic(feSpace, globalRowDof)) {
Thomas Witkowski's avatar
Thomas Witkowski committed
877 878 879
	  std::set<int>& perAsc = perMap.getAssociations(feSpace, globalRowDof);
	  double value = *dofIt / (perAsc.size() + 1.0);
	  VecSetValue(vecInterior, index, value, ADD_VALUES);
880

Thomas Witkowski's avatar
Thomas Witkowski committed
881 882 883 884
	  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);
885

Thomas Witkowski's avatar
Thomas Witkowski committed
886
	    VecSetValue(vecInterior, mappedIndex, value, ADD_VALUES);
887 888 889 890
	  }	  
	} else {	  
	  // The DOF index is not periodic.
	  
Thomas Witkowski's avatar
Thomas Witkowski committed
891 892
	  VecSetValue(vecInterior, index, *dofIt, ADD_VALUES);
	}
Thomas Witkowski's avatar
Thomas Witkowski committed
893 894 895 896 897
      }
    }
  }


898 899 900 901
  void PetscSolverGlobalMatrix::removeDirichletBcDofs(Matrix<DOFMatrix*> *mat)
  {
    FUNCNAME("PetscSolverGlobalMatrix::removeDirichletBcDofs()");

Thomas Witkowski's avatar
Thomas Witkowski committed
902
#if 0
903 904 905 906
    vector<int> dofsInterior, dofsCoarse;

    int nComponents = mat->getNumRows();
    for (int i = 0; i < nComponents; i++) {
Thomas Witkowski's avatar
Thomas Witkowski committed
907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926
      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));
	      }
927
	    }
928
	  }
Thomas Witkowski's avatar
Thomas Witkowski committed
929 930
	} else {
	  MSG("NO MAT DIAG in %d\n", i);
931 932 933 934 935 936 937 938 939 940
	}
      }
    }

    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
941
#endif
942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958
  }


  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)) {
959 960 961
	  if ((*coarseSpaceMap)[feSpace].isRankDof(it->first)) {
	    int globalDof = (*coarseSpaceMap)[feSpace][it->first].global;
	    VecSetValue(rhsCoarseSpace, coarseSpaceMap->getMatIndex(i, globalDof), 
962
			it->second, INSERT_VALUES);
963
	  }
964 965
	} else {
	  if ((*interiorMap)[feSpace].isRankDof(it->first)) {
966 967
	    int globalDof = (*interiorMap)[feSpace][it->first].global;
	    VecSetValue(rhsInterior, interiorMap->getMatIndex(i, globalDof),
968 969 970 971 972
			it->second, INSERT_VALUES);	    
	  }
	}
      }
    }
973 974 975 976 977 978 979 980

    VecAssemblyBegin(rhsInterior);
    VecAssemblyEnd(rhsInterior);

    if (coarseSpaceMap) {
      VecAssemblyBegin(rhsCoarseSpace);
      VecAssemblyEnd(rhsCoarseSpace);
    }
981 982 983
  }


Thomas Witkowski's avatar
Thomas Witkowski committed
984 985 986 987 988 989 990
  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");

991
    vector<const FiniteElemSpace*> feSpaces = getFeSpaces(mat);
992 993
    int nRankRows = interiorMap->getRankDofs();
    int rankStartIndex = interiorMap->getStartDofs();
994

Thomas Witkowski's avatar
Thomas Witkowski committed
995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007
    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;

1008 1009 1010 1011
    // 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
1012 1013 1014 1015
    map<int, MatrixNnzEntry> sendMatrixEntry;

    // Maps to each DOF that must be send to another rank the rank number of the
    // receiving rank.
1016
    map<pair<DegreeOfFreedom, int>, int> sendDofToRank;
Thomas Witkowski's avatar
Thomas Witkowski committed
1017

1018

1019 1020 1021
    // 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++) {
1022
      for (DofComm::Iterator it(meshDistributor->getDofComm().getRecvDofs(), feSpaces[i]);
1023 1024 1025 1026 1027 1028
	   !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
1029 1030
    }

1031
    // Create list of ranks from which we receive data from.
Thomas Witkowski's avatar
Thomas Witkowski committed
1032
    std::set<int> recvFromRank;
1033
    for (unsigned int i = 0; i < feSpaces.size(); i++) 
1034
      for (DofComm::Iterator it(meshDistributor->getDofComm().getSendDofs(), feSpaces[i]);
1035 1036 1037 1038 1039
	   !it.end(); it.nextRank())
	recvFromRank.insert(it.getRank());


    // === Traverse matrices to create nnz data. ===
Thomas Witkowski's avatar
Thomas Witkowski committed
1040

1041
    int nComponents = mat->getNumRows();
Thomas Witkowski's avatar
Thomas Witkowski committed
1042 1043 1044 1045 1046
    for (int i = 0; i < nComponents; i++) {
      for (int j = 0; j < nComponents; j++) {
 	if (!(*mat)[i][j])
	  continue;

1047 1048 1049 1050 1051
	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
1052 1053 1054 1055 1056 1057 1058 1059 1060 1061
	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) {
1062
	  int globalRowDof = (*interiorMap)[feSpaces[i]][*cursor].global;
1063

1064
	  // The corresponding global matrix row index of the current row DOF.
1065 1066
	  int petscRowIdx = interiorMap->getMatIndex(i, globalRowDof);
	  if ((*interiorMap)[feSpaces[i]].isRankDof(*cursor)) {
1067
    	    
1068 1069
	    // === 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
1070 1071
	    
	    // This is the local row index of the local PETSc matrix.
1072
	    int localPetscRowIdx = petscRowIdx - rankStartIndex;
Thomas Witkowski's avatar
Thomas Witkowski committed
1073 1074
	    
	    TEST_EXIT_DBG(localPetscRowIdx >= 0 && localPetscRowIdx < nRankRows)
1075
	      ("Should not happen! \n Debug info: DOF = %d   globalRowIndx = %d   petscRowIdx = %d   localPetscRowIdx = %d   rStart = %d   compontens = %d from %d   nRankRows = %d\n",
1076
	       *cursor,
1077
	       (*interiorMap)[feSpaces[i]][*cursor].global,
Thomas Witkowski's avatar
Thomas Witkowski committed
1078 1079
	       petscRowIdx, 
	       localPetscRowIdx, 
1080
	       rankStartIndex,
1081
	       i,
Thomas Witkowski's avatar
Thomas Witkowski committed
1082 1083 1084 1085 1086 1087 1088
	       nComponents, 
	       nRankRows);
	    
	    
	    // Traverse all non zero entries in this row.
	    for (icursor_type icursor = begin<nz>(cursor), 
		   icend = end<nz>(cursor); icursor != icend; ++icursor) {
1089 1090
	      int globalColDof = (*interiorMap)[feSpaces[j]][col(*icursor)].global;
	      int petscColIdx = interiorMap->getMatIndex(j, globalColDof);
Thomas Witkowski's avatar
Thomas Witkowski committed
1091 1092 1093
	      
	      if (value(*icursor) != 0.0 || petscRowIdx == petscColIdx) {
		// The row DOF is a rank DOF, if also the column is a rank DOF, 
1094 1095 1096 1097
		// 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
1098 1099 1100 1101 1102 1103
		  d_nnz[localPetscRowIdx]++;
		else
		  o_nnz[localPetscRowIdx]++;
	      }    
	    }
	  } else {
1104 1105 1106 1107 1108
	    // === 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
1109 1110
	    
	    // Send all non zero entries to the member of the row DOF.
1111
	    int sendToRank = sendDofToRank[make_pair(*cursor, i)];
Thomas Witkowski's avatar
Thomas Witkowski committed
1112 1113 1114 1115
	    
	    for (icursor_type icursor = begin<nz>(cursor), 
		   icend = end<nz>(cursor); icursor != icend; ++icursor) {
	      if (value(*icursor) != 0.0) {
1116
		int globalColDof = 
1117 1118
		  (*interiorMap)[feSpaces[j]][col(*icursor)].global;
		int petscColIdx = interiorMap->getMatIndex(j, globalColDof);
Thomas Witkowski's avatar
Thomas Witkowski committed
1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131
		
		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. ===

1132
    StdMpi<MatrixNnzEntry> stdMpi(mpiCommGlobal, true);
Thomas Witkowski's avatar
Thomas Witkowski committed
1133 1134 1135 1136 1137 1138 1139
    stdMpi.send(sendMatrixEntry);
    for (std::set<int>::iterator it = recvFromRank.begin(); 
	 it != recvFromRank.end(); ++it)
      stdMpi.recv(*it);
    stdMpi.startCommunication();


1140 1141
    // === 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
1142 1143 1144 1145 1146 1147 1148 1149

    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;

1150
	  int localRowIdx = r - rankStartIndex;
Thomas Witkowski's avatar
Thomas Witkowski committed
1151 1152 1153 1154 1155

	  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);
	  
1156
	  if (c < rankStartIndex || c >= rankStartIndex + nRankRows)
Thomas Witkowski's avatar
Thomas Witkowski committed
1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175
	    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);
  }

}