Actual source code: nn.c


  2: #include <../src/ksp/pc/impls/is/nn/nn.h>

  4: /* -------------------------------------------------------------------------- */
  5: /*
  6:    PCSetUp_NN - Prepares for the use of the NN preconditioner
  7:                     by setting data structures and options.

  9:    Input Parameter:
 10: .  pc - the preconditioner context

 12:    Application Interface Routine: PCSetUp()

 14:    Notes:
 15:    The interface routine PCSetUp() is not usually called directly by
 16:    the user, but instead is called by PCApply() if necessary.
 17: */
 18: static PetscErrorCode PCSetUp_NN(PC pc)
 19: {

 23:   if (!pc->setupcalled) {
 24:     /* Set up all the "iterative substructuring" common block */
 25:     PCISSetUp(pc,PETSC_TRUE,PETSC_TRUE);
 26:     /* Create the coarse matrix. */
 27:     PCNNCreateCoarseMatrix(pc);
 28:   }
 29:   return(0);
 30: }

 32: /* -------------------------------------------------------------------------- */
 33: /*
 34:    PCApply_NN - Applies the NN preconditioner to a vector.

 36:    Input Parameters:
 37: .  pc - the preconditioner context
 38: .  r - input vector (global)

 40:    Output Parameter:
 41: .  z - output vector (global)

 43:    Application Interface Routine: PCApply()
 44:  */
 45: static PetscErrorCode PCApply_NN(PC pc,Vec r,Vec z)
 46: {
 47:   PC_IS          *pcis = (PC_IS*)(pc->data);
 49:   PetscScalar    m_one = -1.0;
 50:   Vec            w     = pcis->vec1_global;

 53:   /*
 54:     Dirichlet solvers.
 55:     Solving $ B_I^{(i)}r_I^{(i)} $ at each processor.
 56:     Storing the local results at vec2_D
 57:   */
 58:   VecScatterBegin(pcis->global_to_D,r,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);
 59:   VecScatterEnd  (pcis->global_to_D,r,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);
 60:   KSPSolve(pcis->ksp_D,pcis->vec1_D,pcis->vec2_D);

 62:   /*
 63:     Computing $ r_B - \sum_j \tilde R_j^T A_{BI}^{(j)} (B_I^{(j)}r_I^{(j)}) $ .
 64:     Storing the result in the interface portion of the global vector w.
 65:   */
 66:   MatMult(pcis->A_BI,pcis->vec2_D,pcis->vec1_B);
 67:   VecScale(pcis->vec1_B,m_one);
 68:   VecCopy(r,w);
 69:   VecScatterBegin(pcis->global_to_B,pcis->vec1_B,w,ADD_VALUES,SCATTER_REVERSE);
 70:   VecScatterEnd  (pcis->global_to_B,pcis->vec1_B,w,ADD_VALUES,SCATTER_REVERSE);

 72:   /*
 73:     Apply the interface preconditioner
 74:   */
 75:   PCNNApplyInterfacePreconditioner(pc,w,z,pcis->work_N,pcis->vec1_B,pcis->vec2_B,pcis->vec3_B,pcis->vec1_D,
 76:                                           pcis->vec3_D,pcis->vec1_N,pcis->vec2_N);

 78:   /*
 79:     Computing $ t_I^{(i)} = A_{IB}^{(i)} \tilde R_i z_B $
 80:     The result is stored in vec1_D.
 81:   */
 82:   VecScatterBegin(pcis->global_to_B,z,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);
 83:   VecScatterEnd  (pcis->global_to_B,z,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);
 84:   MatMult(pcis->A_IB,pcis->vec1_B,pcis->vec1_D);

 86:   /*
 87:     Dirichlet solvers.
 88:     Computing $ B_I^{(i)}t_I^{(i)} $ and sticking into the global vector the blocks
 89:     $ B_I^{(i)}r_I^{(i)} - B_I^{(i)}t_I^{(i)} $.
 90:   */
 91:   VecScatterBegin(pcis->global_to_D,pcis->vec2_D,z,INSERT_VALUES,SCATTER_REVERSE);
 92:   VecScatterEnd  (pcis->global_to_D,pcis->vec2_D,z,INSERT_VALUES,SCATTER_REVERSE);
 93:   KSPSolve(pcis->ksp_D,pcis->vec1_D,pcis->vec2_D);
 94:   VecScale(pcis->vec2_D,m_one);
 95:   VecScatterBegin(pcis->global_to_D,pcis->vec2_D,z,ADD_VALUES,SCATTER_REVERSE);
 96:   VecScatterEnd  (pcis->global_to_D,pcis->vec2_D,z,ADD_VALUES,SCATTER_REVERSE);
 97:   return(0);
 98: }

100: /* -------------------------------------------------------------------------- */
101: /*
102:    PCDestroy_NN - Destroys the private context for the NN preconditioner
103:    that was created with PCCreate_NN().

105:    Input Parameter:
106: .  pc - the preconditioner context

108:    Application Interface Routine: PCDestroy()
109: */
110: static PetscErrorCode PCDestroy_NN(PC pc)
111: {
112:   PC_NN          *pcnn = (PC_NN*)pc->data;

116:   PCISDestroy(pc);

118:   MatDestroy(&pcnn->coarse_mat);
119:   VecDestroy(&pcnn->coarse_x);
120:   VecDestroy(&pcnn->coarse_b);
121:   KSPDestroy(&pcnn->ksp_coarse);
122:   if (pcnn->DZ_IN) {
123:     PetscFree(pcnn->DZ_IN[0]);
124:     PetscFree(pcnn->DZ_IN);
125:   }

127:   /*
128:       Free the private data structure that was hanging off the PC
129:   */
130:   PetscFree(pc->data);
131:   return(0);
132: }

134: /* -------------------------------------------------------------------------- */
135: /*MC
136:    PCNN - Balancing Neumann-Neumann for scalar elliptic PDEs.

138:    Options Database Keys:
139: +    -pc_nn_turn_off_first_balancing - do not balance the residual before solving the local Neumann problems
140:                                        (this skips the first coarse grid solve in the preconditioner)
141: .    -pc_nn_turn_off_second_balancing - do not balance the solution solving the local Neumann problems
142:                                        (this skips the second coarse grid solve in the preconditioner)
143: .    -pc_is_damp_fixed <fact> -
144: .    -pc_is_remove_nullspace_fixed -
145: .    -pc_is_set_damping_factor_floating <fact> -
146: .    -pc_is_not_damp_floating -
147: -    -pc_is_not_remove_nullspace_floating -

149:    Level: intermediate

151:    Notes:
152:     The matrix used with this preconditioner must be of type MATIS

154:           Unlike more 'conventional' Neumann-Neumann preconditioners this iterates over ALL the
155:           degrees of freedom, NOT just those on the interface (this allows the use of approximate solvers
156:           on the subdomains; though in our experience using approximate solvers is slower.).

158:           Options for the coarse grid preconditioner can be set with -nn_coarse_pc_xxx
159:           Options for the Dirichlet subproblem preconditioner can be set with -is_localD_pc_xxx
160:           Options for the Neumann subproblem preconditioner can be set with -is_localN_pc_xxx

162:    Contributed by Paulo Goldfeld

164: .seealso:  PCCreate(), PCSetType(), PCType (for list of available types), PC,  MATIS
165: M*/

167: PETSC_EXTERN PetscErrorCode PCCreate_NN(PC pc)
168: {
170:   PC_NN          *pcnn;

173:   /*
174:      Creates the private data structure for this preconditioner and
175:      attach it to the PC object.
176:   */
177:   PetscNewLog(pc,&pcnn);
178:   pc->data = (void*)pcnn;

180:   PCISCreate(pc);
181:   pcnn->coarse_mat = NULL;
182:   pcnn->coarse_x   = NULL;
183:   pcnn->coarse_b   = NULL;
184:   pcnn->ksp_coarse = NULL;
185:   pcnn->DZ_IN      = NULL;

187:   /*
188:       Set the pointers for the functions that are provided above.
189:       Now when the user-level routines (such as PCApply(), PCDestroy(), etc.)
190:       are called, they will automatically call these functions.  Note we
191:       choose not to provide a couple of these functions since they are
192:       not needed.
193:   */
194:   pc->ops->apply               = PCApply_NN;
195:   pc->ops->applytranspose      = NULL;
196:   pc->ops->setup               = PCSetUp_NN;
197:   pc->ops->destroy             = PCDestroy_NN;
198:   pc->ops->view                = NULL;
199:   pc->ops->applyrichardson     = NULL;
200:   pc->ops->applysymmetricleft  = NULL;
201:   pc->ops->applysymmetricright = NULL;
202:   return(0);
203: }

205: /* -------------------------------------------------------------------------- */
206: /*
207:    PCNNCreateCoarseMatrix -
208: */
209: PetscErrorCode PCNNCreateCoarseMatrix(PC pc)
210: {
211:   MPI_Request    *send_request, *recv_request;
213:   PetscInt       i, j, k;
214:   PetscScalar    *mat;     /* Sub-matrix with this subdomain's contribution to the coarse matrix             */
215:   PetscScalar    **DZ_OUT; /* proc[k].DZ_OUT[i][] = bit of vector to be sent from processor k to processor i */

217:   /* aliasing some names */
218:   PC_IS       *pcis     = (PC_IS*)(pc->data);
219:   PC_NN       *pcnn     = (PC_NN*)pc->data;
220:   PetscInt    n_neigh   = pcis->n_neigh;
221:   PetscInt    *neigh    = pcis->neigh;
222:   PetscInt    *n_shared = pcis->n_shared;
223:   PetscInt    **shared  = pcis->shared;
224:   PetscScalar **DZ_IN;   /* Must be initialized after memory allocation. */

227:   /* Allocate memory for mat (the +1 is to handle the case n_neigh equal to zero) */
228:   PetscMalloc1(n_neigh*n_neigh+1,&mat);

230:   /* Allocate memory for DZ */
231:   /* Notice that DZ_OUT[0] is allocated some space that is never used. */
232:   /* This is just in order to DZ_OUT and DZ_IN to have exactly the same form. */
233:   {
234:     PetscInt size_of_Z = 0;
235:     PetscMalloc ((n_neigh+1)*sizeof(PetscScalar*),&pcnn->DZ_IN);
236:     DZ_IN = pcnn->DZ_IN;
237:     PetscMalloc ((n_neigh+1)*sizeof(PetscScalar*),&DZ_OUT);
238:     for (i=0; i<n_neigh; i++) size_of_Z += n_shared[i];
239:     PetscMalloc ((size_of_Z+1)*sizeof(PetscScalar),&DZ_IN[0]);
240:     PetscMalloc ((size_of_Z+1)*sizeof(PetscScalar),&DZ_OUT[0]);
241:   }
242:   for (i=1; i<n_neigh; i++) {
243:     DZ_IN[i]  = DZ_IN [i-1] + n_shared[i-1];
244:     DZ_OUT[i] = DZ_OUT[i-1] + n_shared[i-1];
245:   }

247:   /* Set the values of DZ_OUT, in order to send this info to the neighbours */
248:   /* First, set the auxiliary array pcis->work_N. */
249:   PCISScatterArrayNToVecB(pcis->work_N,pcis->D,INSERT_VALUES,SCATTER_REVERSE,pc);
250:   for (i=1; i<n_neigh; i++) {
251:     for (j=0; j<n_shared[i]; j++) {
252:       DZ_OUT[i][j] = pcis->work_N[shared[i][j]];
253:     }
254:   }

256:   /* Non-blocking send/receive the common-interface chunks of scaled nullspaces */
257:   /* Notice that send_request[] and recv_request[] could have one less element. */
258:   /* We make them longer to have request[i] corresponding to neigh[i].          */
259:   {
260:     PetscMPIInt tag;
261:     PetscObjectGetNewTag((PetscObject)pc,&tag);
262:     PetscMalloc2(n_neigh+1,&send_request,n_neigh+1,&recv_request);
263:     for (i=1; i<n_neigh; i++) {
264:       MPI_Isend((void*)(DZ_OUT[i]),n_shared[i],MPIU_SCALAR,neigh[i],tag,PetscObjectComm((PetscObject)pc),&(send_request[i]));
265:       MPI_Irecv((void*)(DZ_IN [i]),n_shared[i],MPIU_SCALAR,neigh[i],tag,PetscObjectComm((PetscObject)pc),&(recv_request[i]));
266:     }
267:   }

269:   /* Set DZ_IN[0][] (recall that neigh[0]==rank, always) */
270:   for (j=0; j<n_shared[0]; j++) DZ_IN[0][j] = pcis->work_N[shared[0][j]];

272:   /* Start computing with local D*Z while communication goes on.    */
273:   /* Apply Schur complement. The result is "stored" in vec (more    */
274:   /* precisely, vec points to the result, stored in pc_nn->vec1_B)  */
275:   /* and also scattered to pcnn->work_N.                            */
276:   PCNNApplySchurToChunk(pc,n_shared[0],shared[0],DZ_IN[0],pcis->work_N,pcis->vec1_B,pcis->vec2_B,pcis->vec1_D,pcis->vec2_D);

278:   /* Compute the first column, while completing the receiving. */
279:   for (i=0; i<n_neigh; i++) {
280:     MPI_Status  stat;
281:     PetscMPIInt ind=0;
282:     if (i>0) {MPI_Waitany(n_neigh-1,recv_request+1,&ind,&stat); ind++;}
283:     mat[ind*n_neigh+0] = 0.0;
284:     for (k=0; k<n_shared[ind]; k++) mat[ind*n_neigh+0] += DZ_IN[ind][k] * pcis->work_N[shared[ind][k]];
285:   }

287:   /* Compute the remaining of the columns */
288:   for (j=1; j<n_neigh; j++) {
289:     PCNNApplySchurToChunk(pc,n_shared[j],shared[j],DZ_IN[j],pcis->work_N,pcis->vec1_B,pcis->vec2_B,pcis->vec1_D,pcis->vec2_D);
290:     for (i=0; i<n_neigh; i++) {
291:       mat[i*n_neigh+j] = 0.0;
292:       for (k=0; k<n_shared[i]; k++) mat[i*n_neigh+j] += DZ_IN[i][k] * pcis->work_N[shared[i][k]];
293:     }
294:   }

296:   /* Complete the sending. */
297:   if (n_neigh>1) {
298:     MPI_Status *stat;
299:     PetscMalloc1(n_neigh-1,&stat);
300:     if (n_neigh-1) {MPI_Waitall(n_neigh-1,&(send_request[1]),stat);}
301:     PetscFree(stat);
302:   }

304:   /* Free the memory for the MPI requests */
305:   PetscFree2(send_request,recv_request);

307:   /* Free the memory for DZ_OUT */
308:   if (DZ_OUT) {
309:     PetscFree(DZ_OUT[0]);
310:     PetscFree(DZ_OUT);
311:   }

313:   {
314:     PetscMPIInt size;
315:     MPI_Comm_size(PetscObjectComm((PetscObject)pc),&size);
316:     /* Create the global coarse vectors (rhs and solution). */
317:     VecCreateMPI(PetscObjectComm((PetscObject)pc),1,size,&(pcnn->coarse_b));
318:     VecDuplicate(pcnn->coarse_b,&(pcnn->coarse_x));
319:     /* Create and set the global coarse AIJ matrix. */
320:     MatCreate(PetscObjectComm((PetscObject)pc),&(pcnn->coarse_mat));
321:     MatSetSizes(pcnn->coarse_mat,1,1,size,size);
322:     MatSetType(pcnn->coarse_mat,MATAIJ);
323:     MatSeqAIJSetPreallocation(pcnn->coarse_mat,1,NULL);
324:     MatMPIAIJSetPreallocation(pcnn->coarse_mat,1,NULL,n_neigh,NULL);
325:     MatSetOption(pcnn->coarse_mat,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);
326:     MatSetOption(pcnn->coarse_mat,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);
327:     MatSetValues(pcnn->coarse_mat,n_neigh,neigh,n_neigh,neigh,mat,ADD_VALUES);
328:     MatAssemblyBegin(pcnn->coarse_mat,MAT_FINAL_ASSEMBLY);
329:     MatAssemblyEnd  (pcnn->coarse_mat,MAT_FINAL_ASSEMBLY);
330:   }

332:   {
333:     PetscMPIInt rank;
334:     PetscScalar one = 1.0;
335:     MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);
336:     /* "Zero out" rows of not-purely-Neumann subdomains */
337:     if (pcis->pure_neumann) {  /* does NOT zero the row; create an empty index set. The reason is that MatZeroRows() is collective. */
338:       MatZeroRows(pcnn->coarse_mat,0,NULL,one,NULL,NULL);
339:     } else { /* here it DOES zero the row, since it's not a floating subdomain. */
340:       PetscInt row = (PetscInt) rank;
341:       MatZeroRows(pcnn->coarse_mat,1,&row,one,NULL,NULL);
342:     }
343:   }

345:   /* Create the coarse linear solver context */
346:   {
347:     PC  pc_ctx, inner_pc;
348:     KSP inner_ksp;

350:     KSPCreate(PetscObjectComm((PetscObject)pc),&pcnn->ksp_coarse);
351:     PetscObjectIncrementTabLevel((PetscObject)pcnn->ksp_coarse,(PetscObject)pc,2);
352:     KSPSetOperators(pcnn->ksp_coarse,pcnn->coarse_mat,pcnn->coarse_mat);
353:     KSPGetPC(pcnn->ksp_coarse,&pc_ctx);
354:     PCSetType(pc_ctx,PCREDUNDANT);
355:     KSPSetType(pcnn->ksp_coarse,KSPPREONLY);
356:     PCRedundantGetKSP(pc_ctx,&inner_ksp);
357:     KSPGetPC(inner_ksp,&inner_pc);
358:     PCSetType(inner_pc,PCLU);
359:     KSPSetOptionsPrefix(pcnn->ksp_coarse,"nn_coarse_");
360:     KSPSetFromOptions(pcnn->ksp_coarse);
361:     /* the vectors in the following line are dummy arguments, just telling the KSP the vector size. Values are not used */
362:     KSPSetUp(pcnn->ksp_coarse);
363:   }

365:   /* Free the memory for mat */
366:   PetscFree(mat);

368:   /* for DEBUGGING, save the coarse matrix to a file. */
369:   {
370:     PetscBool flg = PETSC_FALSE;
371:     PetscOptionsGetBool(NULL,NULL,"-pc_nn_save_coarse_matrix",&flg,NULL);
372:     if (flg) {
373:       PetscViewer viewer;
374:       PetscViewerASCIIOpen(PETSC_COMM_WORLD,"coarse.m",&viewer);
375:       PetscViewerPushFormat(viewer,PETSC_VIEWER_ASCII_MATLAB);
376:       MatView(pcnn->coarse_mat,viewer);
377:       PetscViewerPopFormat(viewer);
378:       PetscViewerDestroy(&viewer);
379:     }
380:   }

382:   /*  Set the variable pcnn->factor_coarse_rhs. */
383:   pcnn->factor_coarse_rhs = (pcis->pure_neumann) ? 1.0 : 0.0;

385:   /* See historical note 02, at the bottom of this file. */
386:   return(0);
387: }

389: /* -------------------------------------------------------------------------- */
390: /*
391:    PCNNApplySchurToChunk -

393:    Input parameters:
394: .  pcnn
395: .  n - size of chunk
396: .  idx - indices of chunk
397: .  chunk - values

399:    Output parameters:
400: .  array_N - result of Schur complement applied to chunk, scattered to big array
401: .  vec1_B  - result of Schur complement applied to chunk
402: .  vec2_B  - garbage (used as work space)
403: .  vec1_D  - garbage (used as work space)
404: .  vec2_D  - garbage (used as work space)

406: */
407: PetscErrorCode PCNNApplySchurToChunk(PC pc, PetscInt n, PetscInt *idx, PetscScalar *chunk, PetscScalar *array_N, Vec vec1_B, Vec vec2_B, Vec vec1_D, Vec vec2_D)
408: {
410:   PetscInt       i;
411:   PC_IS          *pcis = (PC_IS*)(pc->data);

414:   PetscArrayzero(array_N, pcis->n);
415:   for (i=0; i<n; i++) array_N[idx[i]] = chunk[i];
416:   PCISScatterArrayNToVecB(array_N,vec2_B,INSERT_VALUES,SCATTER_FORWARD,pc);
417:   PCISApplySchur(pc,vec2_B,vec1_B,(Vec)0,vec1_D,vec2_D);
418:   PCISScatterArrayNToVecB(array_N,vec1_B,INSERT_VALUES,SCATTER_REVERSE,pc);
419:   return(0);
420: }

422: /* -------------------------------------------------------------------------- */
423: /*
424:    PCNNApplyInterfacePreconditioner - Apply the interface preconditioner, i.e.,
425:                                       the preconditioner for the Schur complement.

427:    Input parameter:
428: .  r - global vector of interior and interface nodes. The values on the interior nodes are NOT used.

430:    Output parameters:
431: .  z - global vector of interior and interface nodes. The values on the interface are the result of
432:        the application of the interface preconditioner to the interface part of r. The values on the
433:        interior nodes are garbage.
434: .  work_N - array of local nodes (interior and interface, including ghosts); returns garbage (used as work space)
435: .  vec1_B - vector of local interface nodes (including ghosts); returns garbage (used as work space)
436: .  vec2_B - vector of local interface nodes (including ghosts); returns garbage (used as work space)
437: .  vec3_B - vector of local interface nodes (including ghosts); returns garbage (used as work space)
438: .  vec1_D - vector of local interior nodes; returns garbage (used as work space)
439: .  vec2_D - vector of local interior nodes; returns garbage (used as work space)
440: .  vec1_N - vector of local nodes (interior and interface, including ghosts); returns garbage (used as work space)
441: .  vec2_N - vector of local nodes (interior and interface, including ghosts); returns garbage (used as work space)

443: */
444: PetscErrorCode PCNNApplyInterfacePreconditioner(PC pc, Vec r, Vec z, PetscScalar *work_N, Vec vec1_B, Vec vec2_B, Vec vec3_B, Vec vec1_D,Vec vec2_D, Vec vec1_N, Vec vec2_N)
445: {
447:   PC_IS          *pcis = (PC_IS*)(pc->data);

450:   /*
451:     First balancing step.
452:   */
453:   {
454:     PetscBool flg = PETSC_FALSE;
455:     PetscOptionsGetBool(NULL,NULL,"-pc_nn_turn_off_first_balancing",&flg,NULL);
456:     if (!flg) {
457:       PCNNBalancing(pc,r,(Vec)0,z,vec1_B,vec2_B,(Vec)0,vec1_D,vec2_D,work_N);
458:     } else {
459:       VecCopy(r,z);
460:     }
461:   }

463:   /*
464:     Extract the local interface part of z and scale it by D
465:   */
466:   VecScatterBegin(pcis->global_to_B,z,vec1_B,INSERT_VALUES,SCATTER_FORWARD);
467:   VecScatterEnd  (pcis->global_to_B,z,vec1_B,INSERT_VALUES,SCATTER_FORWARD);
468:   VecPointwiseMult(vec2_B,pcis->D,vec1_B);

470:   /* Neumann Solver */
471:   PCISApplyInvSchur(pc,vec2_B,vec1_B,vec1_N,vec2_N);

473:   /*
474:     Second balancing step.
475:   */
476:   {
477:     PetscBool flg = PETSC_FALSE;
478:     PetscOptionsGetBool(NULL,NULL,"-pc_turn_off_second_balancing",&flg,NULL);
479:     if (!flg) {
480:       PCNNBalancing(pc,r,vec1_B,z,vec2_B,vec3_B,(Vec)0,vec1_D,vec2_D,work_N);
481:     } else {
482:       VecPointwiseMult(vec2_B,pcis->D,vec1_B);
483:       VecSet(z,0.0);
484:       VecScatterBegin(pcis->global_to_B,vec2_B,z,ADD_VALUES,SCATTER_REVERSE);
485:       VecScatterEnd  (pcis->global_to_B,vec2_B,z,ADD_VALUES,SCATTER_REVERSE);
486:     }
487:   }
488:   return(0);
489: }

491: /* -------------------------------------------------------------------------- */
492: /*
493:    PCNNBalancing - Computes z, as given in equations (15) and (16) (if the
494:                    input argument u is provided), or s, as given in equations
495:                    (12) and (13), if the input argument u is a null vector.
496:                    Notice that the input argument u plays the role of u_i in
497:                    equation (14). The equation numbers refer to [Man93].

499:    Input Parameters:
500: .  pcnn - NN preconditioner context.
501: .  r - MPI vector of all nodes (interior and interface). It's preserved.
502: .  u - (Optional) sequential vector of local interface nodes. It's preserved UNLESS vec3_B is null.

504:    Output Parameters:
505: .  z - MPI vector of interior and interface nodes. Returns s or z (see description above).
506: .  vec1_B - Sequential vector of local interface nodes. Workspace.
507: .  vec2_B - Sequential vector of local interface nodes. Workspace.
508: .  vec3_B - (Optional) sequential vector of local interface nodes. Workspace.
509: .  vec1_D - Sequential vector of local interior nodes. Workspace.
510: .  vec2_D - Sequential vector of local interior nodes. Workspace.
511: .  work_N - Array of all local nodes (interior and interface). Workspace.

513: */
514: PetscErrorCode PCNNBalancing(PC pc, Vec r, Vec u, Vec z, Vec vec1_B, Vec vec2_B, Vec vec3_B,Vec vec1_D, Vec vec2_D, PetscScalar *work_N)
515: {
517:   PetscInt       k;
518:   PetscScalar    value;
519:   PetscScalar    *lambda;
520:   PC_NN          *pcnn = (PC_NN*)(pc->data);
521:   PC_IS          *pcis = (PC_IS*)(pc->data);

524:   PetscLogEventBegin(PC_ApplyCoarse,pc,0,0,0);
525:   if (u) {
526:     if (!vec3_B) vec3_B = u;
527:     VecPointwiseMult(vec1_B,pcis->D,u);
528:     VecSet(z,0.0);
529:     VecScatterBegin(pcis->global_to_B,vec1_B,z,ADD_VALUES,SCATTER_REVERSE);
530:     VecScatterEnd  (pcis->global_to_B,vec1_B,z,ADD_VALUES,SCATTER_REVERSE);
531:     VecScatterBegin(pcis->global_to_B,z,vec2_B,INSERT_VALUES,SCATTER_FORWARD);
532:     VecScatterEnd  (pcis->global_to_B,z,vec2_B,INSERT_VALUES,SCATTER_FORWARD);
533:     PCISApplySchur(pc,vec2_B,vec3_B,(Vec)0,vec1_D,vec2_D);
534:     VecScale(vec3_B,-1.0);
535:     VecCopy(r,z);
536:     VecScatterBegin(pcis->global_to_B,vec3_B,z,ADD_VALUES,SCATTER_REVERSE);
537:     VecScatterEnd  (pcis->global_to_B,vec3_B,z,ADD_VALUES,SCATTER_REVERSE);
538:   } else {
539:     VecCopy(r,z);
540:   }
541:   VecScatterBegin(pcis->global_to_B,z,vec2_B,INSERT_VALUES,SCATTER_FORWARD);
542:   VecScatterEnd  (pcis->global_to_B,z,vec2_B,INSERT_VALUES,SCATTER_FORWARD);
543:   PCISScatterArrayNToVecB(work_N,vec2_B,INSERT_VALUES,SCATTER_REVERSE,pc);
544:   for (k=0, value=0.0; k<pcis->n_shared[0]; k++) value += pcnn->DZ_IN[0][k] * work_N[pcis->shared[0][k]];
545:   value *= pcnn->factor_coarse_rhs;  /* This factor is set in CreateCoarseMatrix(). */
546:   {
547:     PetscMPIInt rank;
548:     MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);
549:     VecSetValue(pcnn->coarse_b,rank,value,INSERT_VALUES);
550:     /*
551:        Since we are only inserting local values (one value actually) we don't need to do the
552:        reduction that tells us there is no data that needs to be moved. Hence we comment out these
553:        VecAssemblyBegin(pcnn->coarse_b);
554:        VecAssemblyEnd  (pcnn->coarse_b);
555:     */
556:   }
557:   KSPSolve(pcnn->ksp_coarse,pcnn->coarse_b,pcnn->coarse_x);
558:   if (!u) { VecScale(pcnn->coarse_x,-1.0); }
559:   VecGetArray(pcnn->coarse_x,&lambda);
560:   for (k=0; k<pcis->n_shared[0]; k++) work_N[pcis->shared[0][k]] = *lambda * pcnn->DZ_IN[0][k];
561:   VecRestoreArray(pcnn->coarse_x,&lambda);
562:   PCISScatterArrayNToVecB(work_N,vec2_B,INSERT_VALUES,SCATTER_FORWARD,pc);
563:   VecSet(z,0.0);
564:   VecScatterBegin(pcis->global_to_B,vec2_B,z,ADD_VALUES,SCATTER_REVERSE);
565:   VecScatterEnd  (pcis->global_to_B,vec2_B,z,ADD_VALUES,SCATTER_REVERSE);
566:   if (!u) {
567:     VecScatterBegin(pcis->global_to_B,z,vec2_B,INSERT_VALUES,SCATTER_FORWARD);
568:     VecScatterEnd  (pcis->global_to_B,z,vec2_B,INSERT_VALUES,SCATTER_FORWARD);
569:     PCISApplySchur(pc,vec2_B,vec1_B,(Vec)0,vec1_D,vec2_D);
570:     VecCopy(r,z);
571:   }
572:   VecScatterBegin(pcis->global_to_B,vec1_B,z,ADD_VALUES,SCATTER_REVERSE);
573:   VecScatterEnd  (pcis->global_to_B,vec1_B,z,ADD_VALUES,SCATTER_REVERSE);
574:   PetscLogEventEnd(PC_ApplyCoarse,pc,0,0,0);
575:   return(0);
576: }




581: /*  -------   E N D   O F   T H E   C O D E   -------  */
582: /*                                                     */
583: /*  From now on, "footnotes" (or "historical notes").  */
584: /*                                                     */
585: /*  -------------------------------------------------  */



589: /* --------------------------------------------------------------------------
590:    Historical note 01
591:    -------------------------------------------------------------------------- */
592: /*
593:    We considered the possibility of an alternative D_i that would still
594:    provide a partition of unity (i.e., $ \sum_i  N_i D_i N_i^T = I $).
595:    The basic principle was still the pseudo-inverse of the counting
596:    function; the difference was that we would not count subdomains
597:    that do not contribute to the coarse space (i.e., not pure-Neumann
598:    subdomains).

600:    This turned out to be a bad idea:  we would solve trivial Neumann
601:    problems in the not pure-Neumann subdomains, since we would be scaling
602:    the balanced residual by zero.
603: */




608: /* --------------------------------------------------------------------------
609:    Historical note 02
610:    -------------------------------------------------------------------------- */
611: /*
612:    We tried an alternative coarse problem, that would eliminate exactly a
613:    constant error. Turned out not to improve the overall convergence.
614: */