Actual source code: bjacobi.c

petsc-3.6.1 2015-08-06
Report Typos and Errors
  2: /*
  3:    Defines a block Jacobi preconditioner.
  4: */
  5: #include <petsc/private/pcimpl.h>              /*I "petscpc.h" I*/
  6: #include <../src/ksp/pc/impls/bjacobi/bjacobi.h>

  8: static PetscErrorCode PCSetUp_BJacobi_Singleblock(PC,Mat,Mat);
  9: static PetscErrorCode PCSetUp_BJacobi_Multiblock(PC,Mat,Mat);
 10: static PetscErrorCode PCSetUp_BJacobi_Multiproc(PC);

 14: static PetscErrorCode PCSetUp_BJacobi(PC pc)
 15: {
 16:   PC_BJacobi     *jac = (PC_BJacobi*)pc->data;
 17:   Mat            mat  = pc->mat,pmat = pc->pmat;
 18:   PetscErrorCode ierr,(*f)(Mat,Mat*);
 19:   PetscInt       N,M,start,i,sum,end;
 20:   PetscInt       bs,i_start=-1,i_end=-1;
 21:   PetscMPIInt    rank,size;
 22:   const char     *pprefix,*mprefix;

 25:   MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);
 26:   MPI_Comm_size(PetscObjectComm((PetscObject)pc),&size);
 27:   MatGetLocalSize(pc->pmat,&M,&N);
 28:   MatGetBlockSize(pc->pmat,&bs);

 30:   if (jac->n > 0 && jac->n < size) {
 31:     PCSetUp_BJacobi_Multiproc(pc);
 32:     return(0);
 33:   }

 35:   /* --------------------------------------------------------------------------
 36:       Determines the number of blocks assigned to each processor
 37:   -----------------------------------------------------------------------------*/

 39:   /*   local block count  given */
 40:   if (jac->n_local > 0 && jac->n < 0) {
 41:     MPI_Allreduce(&jac->n_local,&jac->n,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)pc));
 42:     if (jac->l_lens) { /* check that user set these correctly */
 43:       sum = 0;
 44:       for (i=0; i<jac->n_local; i++) {
 45:         if (jac->l_lens[i]/bs*bs !=jac->l_lens[i]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Mat blocksize doesn't match block Jacobi layout");
 46:         sum += jac->l_lens[i];
 47:       }
 48:       if (sum != M) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Local lens set incorrectly");
 49:     } else {
 50:       PetscMalloc1(jac->n_local,&jac->l_lens);
 51:       for (i=0; i<jac->n_local; i++) jac->l_lens[i] = bs*((M/bs)/jac->n_local + (((M/bs) % jac->n_local) > i));
 52:     }
 53:   } else if (jac->n > 0 && jac->n_local < 0) { /* global block count given */
 54:     /* global blocks given: determine which ones are local */
 55:     if (jac->g_lens) {
 56:       /* check if the g_lens is has valid entries */
 57:       for (i=0; i<jac->n; i++) {
 58:         if (!jac->g_lens[i]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Zero block not allowed");
 59:         if (jac->g_lens[i]/bs*bs != jac->g_lens[i]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Mat blocksize doesn't match block Jacobi layout");
 60:       }
 61:       if (size == 1) {
 62:         jac->n_local = jac->n;
 63:         PetscMalloc1(jac->n_local,&jac->l_lens);
 64:         PetscMemcpy(jac->l_lens,jac->g_lens,jac->n_local*sizeof(PetscInt));
 65:         /* check that user set these correctly */
 66:         sum = 0;
 67:         for (i=0; i<jac->n_local; i++) sum += jac->l_lens[i];
 68:         if (sum != M) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Global lens set incorrectly");
 69:       } else {
 70:         MatGetOwnershipRange(pc->pmat,&start,&end);
 71:         /* loop over blocks determing first one owned by me */
 72:         sum = 0;
 73:         for (i=0; i<jac->n+1; i++) {
 74:           if (sum == start) { i_start = i; goto start_1;}
 75:           if (i < jac->n) sum += jac->g_lens[i];
 76:         }
 77:         SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Block sizes used in PCBJacobiSetTotalBlocks()\nare not compatible with parallel matrix layout");
 78: start_1:
 79:         for (i=i_start; i<jac->n+1; i++) {
 80:           if (sum == end) { i_end = i; goto end_1; }
 81:           if (i < jac->n) sum += jac->g_lens[i];
 82:         }
 83:         SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Block sizes used in PCBJacobiSetTotalBlocks()\nare not compatible with parallel matrix layout");
 84: end_1:
 85:         jac->n_local = i_end - i_start;
 86:         PetscMalloc1(jac->n_local,&jac->l_lens);
 87:         PetscMemcpy(jac->l_lens,jac->g_lens+i_start,jac->n_local*sizeof(PetscInt));
 88:       }
 89:     } else { /* no global blocks given, determine then using default layout */
 90:       jac->n_local = jac->n/size + ((jac->n % size) > rank);
 91:       PetscMalloc1(jac->n_local,&jac->l_lens);
 92:       for (i=0; i<jac->n_local; i++) {
 93:         jac->l_lens[i] = ((M/bs)/jac->n_local + (((M/bs) % jac->n_local) > i))*bs;
 94:         if (!jac->l_lens[i]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Too many blocks given");
 95:       }
 96:     }
 97:   } else if (jac->n < 0 && jac->n_local < 0) { /* no blocks given */
 98:     jac->n         = size;
 99:     jac->n_local   = 1;
100:     PetscMalloc1(1,&jac->l_lens);
101:     jac->l_lens[0] = M;
102:   } else { /* jac->n > 0 && jac->n_local > 0 */
103:     if (!jac->l_lens) {
104:       PetscMalloc1(jac->n_local,&jac->l_lens);
105:       for (i=0; i<jac->n_local; i++) jac->l_lens[i] = bs*((M/bs)/jac->n_local + (((M/bs) % jac->n_local) > i));
106:     }
107:   }
108:   if (jac->n_local < 1) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Number of blocks is less than number of processors");

110:   /* -------------------------
111:       Determines mat and pmat
112:   ---------------------------*/
113:   PetscObjectQueryFunction((PetscObject)pc->mat,"MatGetDiagonalBlock_C",&f);
114:   if (!f && size == 1) {
115:     mat  = pc->mat;
116:     pmat = pc->pmat;
117:   } else {
118:     if (pc->useAmat) {
119:       /* use block from Amat matrix, not Pmat for local MatMult() */
120:       MatGetDiagonalBlock(pc->mat,&mat);
121:       /* make submatrix have same prefix as entire matrix */
122:       PetscObjectGetOptionsPrefix((PetscObject)pc->mat,&mprefix);
123:       PetscObjectSetOptionsPrefix((PetscObject)mat,mprefix);
124:     }
125:     if (pc->pmat != pc->mat || !pc->useAmat) {
126:       MatGetDiagonalBlock(pc->pmat,&pmat);
127:       /* make submatrix have same prefix as entire matrix */
128:       PetscObjectGetOptionsPrefix((PetscObject)pc->pmat,&pprefix);
129:       PetscObjectSetOptionsPrefix((PetscObject)pmat,pprefix);
130:     } else pmat = mat;
131:   }

133:   /* ------
134:      Setup code depends on the number of blocks
135:   */
136:   if (jac->n_local == 1) {
137:     PCSetUp_BJacobi_Singleblock(pc,mat,pmat);
138:   } else {
139:     PCSetUp_BJacobi_Multiblock(pc,mat,pmat);
140:   }
141:   return(0);
142: }

144: /* Default destroy, if it has never been setup */
147: static PetscErrorCode PCDestroy_BJacobi(PC pc)
148: {
149:   PC_BJacobi     *jac = (PC_BJacobi*)pc->data;

153:   PetscFree(jac->g_lens);
154:   PetscFree(jac->l_lens);
155:   PetscFree(pc->data);
156:   return(0);
157: }


162: static PetscErrorCode PCSetFromOptions_BJacobi(PetscOptions *PetscOptionsObject,PC pc)
163: {
164:   PC_BJacobi     *jac = (PC_BJacobi*)pc->data;
166:   PetscInt       blocks,i;
167:   PetscBool      flg;

170:   PetscOptionsHead(PetscOptionsObject,"Block Jacobi options");
171:   PetscOptionsInt("-pc_bjacobi_blocks","Total number of blocks","PCBJacobiSetTotalBlocks",jac->n,&blocks,&flg);
172:   if (flg) {
173:     PCBJacobiSetTotalBlocks(pc,blocks,NULL);
174:   }
175:   if (jac->ksp) {
176:     /* The sub-KSP has already been set up (e.g., PCSetUp_BJacobi_Singleblock), but KSPSetFromOptions was not called
177:      * unless we had already been called. */
178:     for (i=0; i<jac->n_local; i++) {
179:       KSPSetFromOptions(jac->ksp[i]);
180:     }
181:   }
182:   PetscOptionsTail();
183:   return(0);
184: }

186: #include <petscdraw.h>
189: static PetscErrorCode PCView_BJacobi(PC pc,PetscViewer viewer)
190: {
191:   PC_BJacobi           *jac   = (PC_BJacobi*)pc->data;
192:   PC_BJacobi_Multiproc *mpjac = (PC_BJacobi_Multiproc*)jac->data;
193:   PetscErrorCode       ierr;
194:   PetscMPIInt          rank;
195:   PetscInt             i;
196:   PetscBool            iascii,isstring,isdraw;
197:   PetscViewer          sviewer;

200:   PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii);
201:   PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERSTRING,&isstring);
202:   PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERDRAW,&isdraw);
203:   if (iascii) {
204:     if (pc->useAmat) {
205:       PetscViewerASCIIPrintf(viewer,"  block Jacobi: using Amat local matrix, number of blocks = %D\n",jac->n);
206:     }
207:     PetscViewerASCIIPrintf(viewer,"  block Jacobi: number of blocks = %D\n",jac->n);
208:     MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);
209:     if (jac->same_local_solves) {
210:       PetscViewerASCIIPrintf(viewer,"  Local solve is same for all blocks, in the following KSP and PC objects:\n");
211:       if (jac->ksp && !jac->psubcomm) {
212:         PetscViewerGetSingleton(viewer,&sviewer);
213:         if (!rank) {
214:           PetscViewerASCIIPushTab(viewer);
215:           KSPView(jac->ksp[0],sviewer);
216:           PetscViewerASCIIPopTab(viewer);
217:         }
218:         PetscViewerRestoreSingleton(viewer,&sviewer);
219:       } else if (jac->psubcomm && !jac->psubcomm->color) {
220:         PetscViewerASCIIGetStdout(PetscSubcommChild(mpjac->psubcomm),&sviewer);
221:         PetscViewerASCIIPushTab(viewer);
222:         KSPView(*(jac->ksp),sviewer);
223:         PetscViewerASCIIPopTab(viewer);
224:       }
225:     } else {
226:       PetscInt n_global;
227:       MPI_Allreduce(&jac->n_local,&n_global,1,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)pc));
228:       PetscViewerASCIISynchronizedAllow(viewer,PETSC_TRUE);
229:       PetscViewerASCIIPrintf(viewer,"  Local solve info for each block is in the following KSP and PC objects:\n");
230:       PetscViewerASCIISynchronizedPrintf(viewer,"[%d] number of local blocks = %D, first local block number = %D\n",
231:                                                 rank,jac->n_local,jac->first_local);
232:       PetscViewerASCIIPushTab(viewer);
233:       PetscViewerGetSingleton(viewer,&sviewer);
234:       for (i=0; i<jac->n_local; i++) {
235:         PetscViewerASCIISynchronizedPrintf(viewer,"[%d] local block number %D\n",rank,i);
236:         KSPView(jac->ksp[i],sviewer);
237:         PetscViewerASCIISynchronizedPrintf(viewer,"- - - - - - - - - - - - - - - - - -\n");
238:       }
239:       PetscViewerRestoreSingleton(viewer,&sviewer);
240:       PetscViewerASCIIPopTab(viewer);
241:       PetscViewerFlush(viewer);
242:       PetscViewerASCIISynchronizedAllow(viewer,PETSC_FALSE);
243:     }
244:   } else if (isstring) {
245:     PetscViewerStringSPrintf(viewer," blks=%D",jac->n);
246:     PetscViewerGetSingleton(viewer,&sviewer);
247:     if (jac->ksp) {KSPView(jac->ksp[0],sviewer);}
248:     PetscViewerRestoreSingleton(viewer,&sviewer);
249:   } else if (isdraw) {
250:     PetscDraw draw;
251:     char      str[25];
252:     PetscReal x,y,bottom,h;

254:     PetscViewerDrawGetDraw(viewer,0,&draw);
255:     PetscDrawGetCurrentPoint(draw,&x,&y);
256:     PetscSNPrintf(str,25,"Number blocks %D",jac->n);
257:     PetscDrawStringBoxed(draw,x,y,PETSC_DRAW_RED,PETSC_DRAW_BLACK,str,NULL,&h);
258:     bottom = y - h;
259:     PetscDrawPushCurrentPoint(draw,x,bottom);
260:     /* warning the communicator on viewer is different then on ksp in parallel */
261:     if (jac->ksp) {KSPView(jac->ksp[0],viewer);}
262:     PetscDrawPopCurrentPoint(draw);
263:   }
264:   return(0);
265: }

267: /* -------------------------------------------------------------------------------------*/

271: static PetscErrorCode  PCBJacobiGetSubKSP_BJacobi(PC pc,PetscInt *n_local,PetscInt *first_local,KSP **ksp)
272: {
273:   PC_BJacobi *jac = (PC_BJacobi*)pc->data;;

276:   if (!pc->setupcalled) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_ARG_WRONGSTATE,"Must call KSPSetUp() or PCSetUp() first");

278:   if (n_local) *n_local = jac->n_local;
279:   if (first_local) *first_local = jac->first_local;
280:   *ksp                   = jac->ksp;
281:   jac->same_local_solves = PETSC_FALSE;        /* Assume that local solves are now different;
282:                                                   not necessarily true though!  This flag is
283:                                                   used only for PCView_BJacobi() */
284:   return(0);
285: }

289: static PetscErrorCode  PCBJacobiSetTotalBlocks_BJacobi(PC pc,PetscInt blocks,PetscInt *lens)
290: {
291:   PC_BJacobi     *jac = (PC_BJacobi*)pc->data;

295:   if (pc->setupcalled > 0 && jac->n!=blocks) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_ORDER,"Cannot alter number of blocks after PCSetUp()/KSPSetUp() has been called");
296:   jac->n = blocks;
297:   if (!lens) jac->g_lens = 0;
298:   else {
299:     PetscMalloc1(blocks,&jac->g_lens);
300:     PetscLogObjectMemory((PetscObject)pc,blocks*sizeof(PetscInt));
301:     PetscMemcpy(jac->g_lens,lens,blocks*sizeof(PetscInt));
302:   }
303:   return(0);
304: }

308: static PetscErrorCode  PCBJacobiGetTotalBlocks_BJacobi(PC pc, PetscInt *blocks, const PetscInt *lens[])
309: {
310:   PC_BJacobi *jac = (PC_BJacobi*) pc->data;

313:   *blocks = jac->n;
314:   if (lens) *lens = jac->g_lens;
315:   return(0);
316: }

320: static PetscErrorCode  PCBJacobiSetLocalBlocks_BJacobi(PC pc,PetscInt blocks,const PetscInt lens[])
321: {
322:   PC_BJacobi     *jac;

326:   jac = (PC_BJacobi*)pc->data;

328:   jac->n_local = blocks;
329:   if (!lens) jac->l_lens = 0;
330:   else {
331:     PetscMalloc1(blocks,&jac->l_lens);
332:     PetscLogObjectMemory((PetscObject)pc,blocks*sizeof(PetscInt));
333:     PetscMemcpy(jac->l_lens,lens,blocks*sizeof(PetscInt));
334:   }
335:   return(0);
336: }

340: static PetscErrorCode  PCBJacobiGetLocalBlocks_BJacobi(PC pc, PetscInt *blocks, const PetscInt *lens[])
341: {
342:   PC_BJacobi *jac = (PC_BJacobi*) pc->data;

345:   *blocks = jac->n_local;
346:   if (lens) *lens = jac->l_lens;
347:   return(0);
348: }

350: /* -------------------------------------------------------------------------------------*/

354: /*@C
355:    PCBJacobiGetSubKSP - Gets the local KSP contexts for all blocks on
356:    this processor.

358:    Note Collective

360:    Input Parameter:
361: .  pc - the preconditioner context

363:    Output Parameters:
364: +  n_local - the number of blocks on this processor, or NULL
365: .  first_local - the global number of the first block on this processor, or NULL
366: -  ksp - the array of KSP contexts

368:    Notes:
369:    After PCBJacobiGetSubKSP() the array of KSP contexts is not to be freed.

371:    Currently for some matrix implementations only 1 block per processor
372:    is supported.

374:    You must call KSPSetUp() or PCSetUp() before calling PCBJacobiGetSubKSP().

376:    Fortran Usage: You must pass in a KSP array that is large enough to contain all the local KSPs.
377:       You can call PCBJacobiGetSubKSP(pc,nlocal,firstlocal,NULL_OBJECT,ierr) to determine how large the
378:       KSP array must be.

380:    Level: advanced

382: .keywords:  block, Jacobi, get, sub, KSP, context

384: .seealso: PCBJacobiGetSubKSP()
385: @*/
386: PetscErrorCode  PCBJacobiGetSubKSP(PC pc,PetscInt *n_local,PetscInt *first_local,KSP *ksp[])
387: {

392:   PetscUseMethod(pc,"PCBJacobiGetSubKSP_C",(PC,PetscInt*,PetscInt*,KSP **),(pc,n_local,first_local,ksp));
393:   return(0);
394: }

398: /*@
399:    PCBJacobiSetTotalBlocks - Sets the global number of blocks for the block
400:    Jacobi preconditioner.

402:    Collective on PC

404:    Input Parameters:
405: +  pc - the preconditioner context
406: .  blocks - the number of blocks
407: -  lens - [optional] integer array containing the size of each block

409:    Options Database Key:
410: .  -pc_bjacobi_blocks <blocks> - Sets the number of global blocks

412:    Notes:
413:    Currently only a limited number of blocking configurations are supported.
414:    All processors sharing the PC must call this routine with the same data.

416:    Level: intermediate

418: .keywords:  set, number, Jacobi, global, total, blocks

420: .seealso: PCSetUseAmat(), PCBJacobiSetLocalBlocks()
421: @*/
422: PetscErrorCode  PCBJacobiSetTotalBlocks(PC pc,PetscInt blocks,const PetscInt lens[])
423: {

428:   if (blocks <= 0) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_ARG_OUTOFRANGE,"Must have positive blocks");
429:   PetscTryMethod(pc,"PCBJacobiSetTotalBlocks_C",(PC,PetscInt,const PetscInt[]),(pc,blocks,lens));
430:   return(0);
431: }

435: /*@C
436:    PCBJacobiGetTotalBlocks - Gets the global number of blocks for the block
437:    Jacobi preconditioner.

439:    Not Collective

441:    Input Parameter:
442: .  pc - the preconditioner context

444:    Output parameters:
445: +  blocks - the number of blocks
446: -  lens - integer array containing the size of each block

448:    Level: intermediate

450: .keywords:  get, number, Jacobi, global, total, blocks

452: .seealso: PCSetUseAmat(), PCBJacobiGetLocalBlocks()
453: @*/
454: PetscErrorCode  PCBJacobiGetTotalBlocks(PC pc, PetscInt *blocks, const PetscInt *lens[])
455: {

461:   PetscUseMethod(pc,"PCBJacobiGetTotalBlocks_C",(PC,PetscInt*, const PetscInt *[]),(pc,blocks,lens));
462:   return(0);
463: }

467: /*@
468:    PCBJacobiSetLocalBlocks - Sets the local number of blocks for the block
469:    Jacobi preconditioner.

471:    Not Collective

473:    Input Parameters:
474: +  pc - the preconditioner context
475: .  blocks - the number of blocks
476: -  lens - [optional] integer array containing size of each block

478:    Note:
479:    Currently only a limited number of blocking configurations are supported.

481:    Level: intermediate

483: .keywords: PC, set, number, Jacobi, local, blocks

485: .seealso: PCSetUseAmat(), PCBJacobiSetTotalBlocks()
486: @*/
487: PetscErrorCode  PCBJacobiSetLocalBlocks(PC pc,PetscInt blocks,const PetscInt lens[])
488: {

493:   if (blocks < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Must have nonegative blocks");
494:   PetscTryMethod(pc,"PCBJacobiSetLocalBlocks_C",(PC,PetscInt,const PetscInt []),(pc,blocks,lens));
495:   return(0);
496: }

500: /*@C
501:    PCBJacobiGetLocalBlocks - Gets the local number of blocks for the block
502:    Jacobi preconditioner.

504:    Not Collective

506:    Input Parameters:
507: +  pc - the preconditioner context
508: .  blocks - the number of blocks
509: -  lens - [optional] integer array containing size of each block

511:    Note:
512:    Currently only a limited number of blocking configurations are supported.

514:    Level: intermediate

516: .keywords: PC, get, number, Jacobi, local, blocks

518: .seealso: PCSetUseAmat(), PCBJacobiGetTotalBlocks()
519: @*/
520: PetscErrorCode  PCBJacobiGetLocalBlocks(PC pc, PetscInt *blocks, const PetscInt *lens[])
521: {

527:   PetscUseMethod(pc,"PCBJacobiGetLocalBlocks_C",(PC,PetscInt*, const PetscInt *[]),(pc,blocks,lens));
528:   return(0);
529: }

531: /* -----------------------------------------------------------------------------------*/

533: /*MC
534:    PCBJACOBI - Use block Jacobi preconditioning, each block is (approximately) solved with
535:            its own KSP object.

537:    Options Database Keys:
538: .  -pc_use_amat - use Amat to apply block of operator in inner Krylov method

540:    Notes: Each processor can have one or more blocks, but a block cannot be shared by more
541:      than one processor. Defaults to one block per processor.

543:      To set options on the solvers for each block append -sub_ to all the KSP, KSP, and PC
544:         options database keys. For example, -sub_pc_type ilu -sub_pc_factor_levels 1 -sub_ksp_type preonly

546:      To set the options on the solvers separate for each block call PCBJacobiGetSubKSP()
547:          and set the options directly on the resulting KSP object (you can access its PC
548:          KSPGetPC())

550:    Level: beginner

552:    Concepts: block Jacobi

554:    Developer Notes: This preconditioner does not currently work with CUDA/CUSP for a couple of reasons.
555:        (1) It creates seq vectors as work vectors that should be cusp
556:        (2) The use of VecPlaceArray() is not handled properly by CUSP (that is it will not know where
557:            the ownership of the vector is so may use wrong values) even if it did know the ownership
558:            it may induce extra copy ups and downs. Satish suggests a VecTransplantArray() to handle two
559:            vectors sharing the same pointer and handling the CUSP side as well instead of VecGetArray()/VecPlaceArray().


562: .seealso:  PCCreate(), PCSetType(), PCType (for list of available types), PC,
563:            PCASM, PCSetUseAmat(), PCGetUseAmat(), PCBJacobiGetSubKSP(), PCBJacobiSetTotalBlocks(),
564:            PCBJacobiSetLocalBlocks(), PCSetModifySubmatrices()
565: M*/

569: PETSC_EXTERN PetscErrorCode PCCreate_BJacobi(PC pc)
570: {
572:   PetscMPIInt    rank;
573:   PC_BJacobi     *jac;

576:   PetscNewLog(pc,&jac);
577:   MPI_Comm_rank(PetscObjectComm((PetscObject)pc),&rank);

579:   pc->ops->apply           = 0;
580:   pc->ops->applytranspose  = 0;
581:   pc->ops->setup           = PCSetUp_BJacobi;
582:   pc->ops->destroy         = PCDestroy_BJacobi;
583:   pc->ops->setfromoptions  = PCSetFromOptions_BJacobi;
584:   pc->ops->view            = PCView_BJacobi;
585:   pc->ops->applyrichardson = 0;

587:   pc->data               = (void*)jac;
588:   jac->n                 = -1;
589:   jac->n_local           = -1;
590:   jac->first_local       = rank;
591:   jac->ksp               = 0;
592:   jac->same_local_solves = PETSC_TRUE;
593:   jac->g_lens            = 0;
594:   jac->l_lens            = 0;
595:   jac->psubcomm          = 0;

597:   PetscObjectComposeFunction((PetscObject)pc,"PCBJacobiGetSubKSP_C",PCBJacobiGetSubKSP_BJacobi);
598:   PetscObjectComposeFunction((PetscObject)pc,"PCBJacobiSetTotalBlocks_C",PCBJacobiSetTotalBlocks_BJacobi);
599:   PetscObjectComposeFunction((PetscObject)pc,"PCBJacobiGetTotalBlocks_C",PCBJacobiGetTotalBlocks_BJacobi);
600:   PetscObjectComposeFunction((PetscObject)pc,"PCBJacobiSetLocalBlocks_C",PCBJacobiSetLocalBlocks_BJacobi);
601:   PetscObjectComposeFunction((PetscObject)pc,"PCBJacobiGetLocalBlocks_C",PCBJacobiGetLocalBlocks_BJacobi);
602:   return(0);
603: }

605: /* --------------------------------------------------------------------------------------------*/
606: /*
607:         These are for a single block per processor; works for AIJ, BAIJ; Seq and MPI
608: */
611: PetscErrorCode PCReset_BJacobi_Singleblock(PC pc)
612: {
613:   PC_BJacobi             *jac  = (PC_BJacobi*)pc->data;
614:   PC_BJacobi_Singleblock *bjac = (PC_BJacobi_Singleblock*)jac->data;
615:   PetscErrorCode         ierr;

618:   KSPReset(jac->ksp[0]);
619:   VecDestroy(&bjac->x);
620:   VecDestroy(&bjac->y);
621:   return(0);
622: }

626: PetscErrorCode PCDestroy_BJacobi_Singleblock(PC pc)
627: {
628:   PC_BJacobi             *jac  = (PC_BJacobi*)pc->data;
629:   PC_BJacobi_Singleblock *bjac = (PC_BJacobi_Singleblock*)jac->data;
630:   PetscErrorCode         ierr;

633:   PCReset_BJacobi_Singleblock(pc);
634:   KSPDestroy(&jac->ksp[0]);
635:   PetscFree(jac->ksp);
636:   PetscFree(jac->l_lens);
637:   PetscFree(jac->g_lens);
638:   PetscFree(bjac);
639:   PetscFree(pc->data);
640:   return(0);
641: }

645: PetscErrorCode PCSetUpOnBlocks_BJacobi_Singleblock(PC pc)
646: {
648:   PC_BJacobi     *jac = (PC_BJacobi*)pc->data;

651:   KSPSetUp(jac->ksp[0]);
652:   return(0);
653: }

657: PetscErrorCode PCApply_BJacobi_Singleblock(PC pc,Vec x,Vec y)
658: {
659:   PetscErrorCode         ierr;
660:   PC_BJacobi             *jac  = (PC_BJacobi*)pc->data;
661:   PC_BJacobi_Singleblock *bjac = (PC_BJacobi_Singleblock*)jac->data;

664:   VecGetLocalVectorRead(x, bjac->x);
665:   VecGetLocalVector(y, bjac->y);
666:  /* Since the inner KSP matrix may point directly to the diagonal block of an MPI matrix the inner
667:      matrix may change even if the outter KSP/PC has not updated the preconditioner, this will trigger a rebuild
668:      of the inner preconditioner automatically unless we pass down the outter preconditioners reuse flag.*/
669:   KSPSetReusePreconditioner(jac->ksp[0],pc->reusepreconditioner);
670:   KSPSolve(jac->ksp[0],bjac->x,bjac->y);
671:   VecRestoreLocalVectorRead(x, bjac->x);
672:   VecRestoreLocalVector(y, bjac->y);
673:   return(0);
674: }

678: PetscErrorCode PCApplySymmetricLeft_BJacobi_Singleblock(PC pc,Vec x,Vec y)
679: {
680:   PetscErrorCode         ierr;
681:   PC_BJacobi             *jac  = (PC_BJacobi*)pc->data;
682:   PC_BJacobi_Singleblock *bjac = (PC_BJacobi_Singleblock*)jac->data;
683:   PetscScalar            *y_array;
684:   const PetscScalar      *x_array;
685:   PC                     subpc;

688:   /*
689:       The VecPlaceArray() is to avoid having to copy the
690:     y vector into the bjac->x vector. The reason for
691:     the bjac->x vector is that we need a sequential vector
692:     for the sequential solve.
693:   */
694:   VecGetArrayRead(x,&x_array);
695:   VecGetArray(y,&y_array);
696:   VecPlaceArray(bjac->x,x_array);
697:   VecPlaceArray(bjac->y,y_array);
698:   /* apply the symmetric left portion of the inner PC operator */
699:   /* note this by-passes the inner KSP and its options completely */
700:   KSPGetPC(jac->ksp[0],&subpc);
701:   PCApplySymmetricLeft(subpc,bjac->x,bjac->y);
702:   VecResetArray(bjac->x);
703:   VecResetArray(bjac->y);
704:   VecRestoreArrayRead(x,&x_array);
705:   VecRestoreArray(y,&y_array);
706:   return(0);
707: }

711: PetscErrorCode PCApplySymmetricRight_BJacobi_Singleblock(PC pc,Vec x,Vec y)
712: {
713:   PetscErrorCode         ierr;
714:   PC_BJacobi             *jac  = (PC_BJacobi*)pc->data;
715:   PC_BJacobi_Singleblock *bjac = (PC_BJacobi_Singleblock*)jac->data;
716:   PetscScalar            *y_array;
717:   const PetscScalar      *x_array;
718:   PC                     subpc;

721:   /*
722:       The VecPlaceArray() is to avoid having to copy the
723:     y vector into the bjac->x vector. The reason for
724:     the bjac->x vector is that we need a sequential vector
725:     for the sequential solve.
726:   */
727:   VecGetArrayRead(x,&x_array);
728:   VecGetArray(y,&y_array);
729:   VecPlaceArray(bjac->x,x_array);
730:   VecPlaceArray(bjac->y,y_array);

732:   /* apply the symmetric right portion of the inner PC operator */
733:   /* note this by-passes the inner KSP and its options completely */

735:   KSPGetPC(jac->ksp[0],&subpc);
736:   PCApplySymmetricRight(subpc,bjac->x,bjac->y);

738:   VecRestoreArrayRead(x,&x_array);
739:   VecRestoreArray(y,&y_array);
740:   return(0);
741: }

745: PetscErrorCode PCApplyTranspose_BJacobi_Singleblock(PC pc,Vec x,Vec y)
746: {
747:   PetscErrorCode         ierr;
748:   PC_BJacobi             *jac  = (PC_BJacobi*)pc->data;
749:   PC_BJacobi_Singleblock *bjac = (PC_BJacobi_Singleblock*)jac->data;
750:   PetscScalar            *y_array;
751:   const PetscScalar      *x_array;

754:   /*
755:       The VecPlaceArray() is to avoid having to copy the
756:     y vector into the bjac->x vector. The reason for
757:     the bjac->x vector is that we need a sequential vector
758:     for the sequential solve.
759:   */
760:   VecGetArrayRead(x,&x_array);
761:   VecGetArray(y,&y_array);
762:   VecPlaceArray(bjac->x,x_array);
763:   VecPlaceArray(bjac->y,y_array);
764:   KSPSolveTranspose(jac->ksp[0],bjac->x,bjac->y);
765:   VecResetArray(bjac->x);
766:   VecResetArray(bjac->y);
767:   VecRestoreArrayRead(x,&x_array);
768:   VecRestoreArray(y,&y_array);
769:   return(0);
770: }

774: static PetscErrorCode PCSetUp_BJacobi_Singleblock(PC pc,Mat mat,Mat pmat)
775: {
776:   PC_BJacobi             *jac = (PC_BJacobi*)pc->data;
777:   PetscErrorCode         ierr;
778:   PetscInt               m;
779:   KSP                    ksp;
780:   PC_BJacobi_Singleblock *bjac;
781:   PetscBool              wasSetup = PETSC_TRUE;

784:   if (!pc->setupcalled) {
785:     const char *prefix;

787:     if (!jac->ksp) {
788:       wasSetup = PETSC_FALSE;

790:       KSPCreate(PETSC_COMM_SELF,&ksp);
791:       KSPSetErrorIfNotConverged(ksp,pc->erroriffailure);
792:       PetscObjectIncrementTabLevel((PetscObject)ksp,(PetscObject)pc,1);
793:       PetscLogObjectParent((PetscObject)pc,(PetscObject)ksp);
794:       KSPSetType(ksp,KSPPREONLY);
795:       PCGetOptionsPrefix(pc,&prefix);
796:       KSPSetOptionsPrefix(ksp,prefix);
797:       KSPAppendOptionsPrefix(ksp,"sub_");

799:       pc->ops->reset               = PCReset_BJacobi_Singleblock;
800:       pc->ops->destroy             = PCDestroy_BJacobi_Singleblock;
801:       pc->ops->apply               = PCApply_BJacobi_Singleblock;
802:       pc->ops->applysymmetricleft  = PCApplySymmetricLeft_BJacobi_Singleblock;
803:       pc->ops->applysymmetricright = PCApplySymmetricRight_BJacobi_Singleblock;
804:       pc->ops->applytranspose      = PCApplyTranspose_BJacobi_Singleblock;
805:       pc->ops->setuponblocks       = PCSetUpOnBlocks_BJacobi_Singleblock;

807:       PetscMalloc1(1,&jac->ksp);
808:       jac->ksp[0] = ksp;

810:       PetscNewLog(pc,&bjac);
811:       jac->data = (void*)bjac;
812:     } else {
813:       ksp  = jac->ksp[0];
814:       bjac = (PC_BJacobi_Singleblock*)jac->data;
815:     }

817:     /*
818:       The reason we need to generate these vectors is to serve
819:       as the right-hand side and solution vector for the solve on the
820:       block. We do not need to allocate space for the vectors since
821:       that is provided via VecPlaceArray() just before the call to
822:       KSPSolve() on the block.
823:     */
824:     MatGetSize(pmat,&m,&m);
825:     VecCreateSeqWithArray(PETSC_COMM_SELF,1,m,NULL,&bjac->x);
826:     VecCreateSeqWithArray(PETSC_COMM_SELF,1,m,NULL,&bjac->y);
827: #ifdef PETSC_HAVE_CUSP
828:     VecSetType(bjac->x,VECCUSP);
829:     VecSetType(bjac->y,VECCUSP);
830: #endif
831:     PetscLogObjectParent((PetscObject)pc,(PetscObject)bjac->x);
832:     PetscLogObjectParent((PetscObject)pc,(PetscObject)bjac->y);
833:   } else {
834:     ksp  = jac->ksp[0];
835:     bjac = (PC_BJacobi_Singleblock*)jac->data;
836:   }
837:   if (pc->useAmat) {
838:     KSPSetOperators(ksp,mat,pmat);
839:   } else {
840:     KSPSetOperators(ksp,pmat,pmat);
841:   }
842:   if (!wasSetup && pc->setfromoptionscalled) {
843:     /* If PCSetFromOptions_BJacobi is called later, KSPSetFromOptions will be called at that time. */
844:     KSPSetFromOptions(ksp);
845:   }
846:   return(0);
847: }

849: /* ---------------------------------------------------------------------------------------------*/
852: PetscErrorCode PCReset_BJacobi_Multiblock(PC pc)
853: {
854:   PC_BJacobi            *jac  = (PC_BJacobi*)pc->data;
855:   PC_BJacobi_Multiblock *bjac = (PC_BJacobi_Multiblock*)jac->data;
856:   PetscErrorCode        ierr;
857:   PetscInt              i;

860:   if (bjac && bjac->pmat) {
861:     MatDestroyMatrices(jac->n_local,&bjac->pmat);
862:     if (pc->useAmat) {
863:       MatDestroyMatrices(jac->n_local,&bjac->mat);
864:     }
865:   }

867:   for (i=0; i<jac->n_local; i++) {
868:     KSPReset(jac->ksp[i]);
869:     if (bjac && bjac->x) {
870:       VecDestroy(&bjac->x[i]);
871:       VecDestroy(&bjac->y[i]);
872:       ISDestroy(&bjac->is[i]);
873:     }
874:   }
875:   PetscFree(jac->l_lens);
876:   PetscFree(jac->g_lens);
877:   return(0);
878: }

882: PetscErrorCode PCDestroy_BJacobi_Multiblock(PC pc)
883: {
884:   PC_BJacobi            *jac  = (PC_BJacobi*)pc->data;
885:   PC_BJacobi_Multiblock *bjac = (PC_BJacobi_Multiblock*)jac->data;
886:   PetscErrorCode        ierr;
887:   PetscInt              i;

890:   PCReset_BJacobi_Multiblock(pc);
891:   if (bjac) {
892:     PetscFree2(bjac->x,bjac->y);
893:     PetscFree(bjac->starts);
894:     PetscFree(bjac->is);
895:   }
896:   PetscFree(jac->data);
897:   for (i=0; i<jac->n_local; i++) {
898:     KSPDestroy(&jac->ksp[i]);
899:   }
900:   PetscFree(jac->ksp);
901:   PetscFree(pc->data);
902:   return(0);
903: }

907: PetscErrorCode PCSetUpOnBlocks_BJacobi_Multiblock(PC pc)
908: {
909:   PC_BJacobi     *jac = (PC_BJacobi*)pc->data;
911:   PetscInt       i,n_local = jac->n_local;

914:   for (i=0; i<n_local; i++) {
915:     KSPSetUp(jac->ksp[i]);
916:   }
917:   return(0);
918: }

920: /*
921:       Preconditioner for block Jacobi
922: */
925: PetscErrorCode PCApply_BJacobi_Multiblock(PC pc,Vec x,Vec y)
926: {
927:   PC_BJacobi            *jac = (PC_BJacobi*)pc->data;
928:   PetscErrorCode        ierr;
929:   PetscInt              i,n_local = jac->n_local;
930:   PC_BJacobi_Multiblock *bjac = (PC_BJacobi_Multiblock*)jac->data;
931:   PetscScalar           *yin;
932:   const PetscScalar     *xin;

935:   VecGetArrayRead(x,&xin);
936:   VecGetArray(y,&yin);
937:   for (i=0; i<n_local; i++) {
938:     /*
939:        To avoid copying the subvector from x into a workspace we instead
940:        make the workspace vector array point to the subpart of the array of
941:        the global vector.
942:     */
943:     VecPlaceArray(bjac->x[i],xin+bjac->starts[i]);
944:     VecPlaceArray(bjac->y[i],yin+bjac->starts[i]);

946:     PetscLogEventBegin(PC_ApplyOnBlocks,jac->ksp[i],bjac->x[i],bjac->y[i],0);
947:     KSPSolve(jac->ksp[i],bjac->x[i],bjac->y[i]);
948:     PetscLogEventEnd(PC_ApplyOnBlocks,jac->ksp[i],bjac->x[i],bjac->y[i],0);

950:     VecResetArray(bjac->x[i]);
951:     VecResetArray(bjac->y[i]);
952:   }
953:   VecRestoreArrayRead(x,&xin);
954:   VecRestoreArray(y,&yin);
955:   return(0);
956: }

958: /*
959:       Preconditioner for block Jacobi
960: */
963: PetscErrorCode PCApplyTranspose_BJacobi_Multiblock(PC pc,Vec x,Vec y)
964: {
965:   PC_BJacobi            *jac = (PC_BJacobi*)pc->data;
966:   PetscErrorCode        ierr;
967:   PetscInt              i,n_local = jac->n_local;
968:   PC_BJacobi_Multiblock *bjac = (PC_BJacobi_Multiblock*)jac->data;
969:   PetscScalar           *yin;
970:   const PetscScalar     *xin;

973:   VecGetArrayRead(x,&xin);
974:   VecGetArray(y,&yin);
975:   for (i=0; i<n_local; i++) {
976:     /*
977:        To avoid copying the subvector from x into a workspace we instead
978:        make the workspace vector array point to the subpart of the array of
979:        the global vector.
980:     */
981:     VecPlaceArray(bjac->x[i],xin+bjac->starts[i]);
982:     VecPlaceArray(bjac->y[i],yin+bjac->starts[i]);

984:     PetscLogEventBegin(PC_ApplyTransposeOnBlocks,jac->ksp[i],bjac->x[i],bjac->y[i],0);
985:     KSPSolveTranspose(jac->ksp[i],bjac->x[i],bjac->y[i]);
986:     PetscLogEventEnd(PC_ApplyTransposeOnBlocks,jac->ksp[i],bjac->x[i],bjac->y[i],0);

988:     VecResetArray(bjac->x[i]);
989:     VecResetArray(bjac->y[i]);
990:   }
991:   VecRestoreArrayRead(x,&xin);
992:   VecRestoreArray(y,&yin);
993:   return(0);
994: }

998: static PetscErrorCode PCSetUp_BJacobi_Multiblock(PC pc,Mat mat,Mat pmat)
999: {
1000:   PC_BJacobi            *jac = (PC_BJacobi*)pc->data;
1001:   PetscErrorCode        ierr;
1002:   PetscInt              m,n_local,N,M,start,i;
1003:   const char            *prefix,*pprefix,*mprefix;
1004:   KSP                   ksp;
1005:   Vec                   x,y;
1006:   PC_BJacobi_Multiblock *bjac = (PC_BJacobi_Multiblock*)jac->data;
1007:   PC                    subpc;
1008:   IS                    is;
1009:   MatReuse              scall;

1012:   MatGetLocalSize(pc->pmat,&M,&N);

1014:   n_local = jac->n_local;

1016:   if (pc->useAmat) {
1017:     PetscBool same;
1018:     PetscObjectTypeCompare((PetscObject)mat,((PetscObject)pmat)->type_name,&same);
1019:     if (!same) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_ARG_INCOMP,"Matrices not of same type");
1020:   }

1022:   if (!pc->setupcalled) {
1023:     scall = MAT_INITIAL_MATRIX;

1025:     if (!jac->ksp) {
1026:       pc->ops->reset         = PCReset_BJacobi_Multiblock;
1027:       pc->ops->destroy       = PCDestroy_BJacobi_Multiblock;
1028:       pc->ops->apply         = PCApply_BJacobi_Multiblock;
1029:       pc->ops->applytranspose= PCApplyTranspose_BJacobi_Multiblock;
1030:       pc->ops->setuponblocks = PCSetUpOnBlocks_BJacobi_Multiblock;

1032:       PetscNewLog(pc,&bjac);
1033:       PetscMalloc1(n_local,&jac->ksp);
1034:       PetscLogObjectMemory((PetscObject)pc,sizeof(n_local*sizeof(KSP)));
1035:       PetscMalloc2(n_local,&bjac->x,n_local,&bjac->y);
1036:       PetscMalloc1(n_local,&bjac->starts);
1037:       PetscLogObjectMemory((PetscObject)pc,sizeof(n_local*sizeof(PetscScalar)));

1039:       jac->data = (void*)bjac;
1040:       PetscMalloc1(n_local,&bjac->is);
1041:       PetscLogObjectMemory((PetscObject)pc,sizeof(n_local*sizeof(IS)));

1043:       for (i=0; i<n_local; i++) {
1044:         KSPCreate(PETSC_COMM_SELF,&ksp);
1045:         KSPSetErrorIfNotConverged(ksp,pc->erroriffailure);
1046:         PetscObjectIncrementTabLevel((PetscObject)ksp,(PetscObject)pc,1);
1047:         PetscLogObjectParent((PetscObject)pc,(PetscObject)ksp);
1048:         KSPSetType(ksp,KSPPREONLY);
1049:         KSPGetPC(ksp,&subpc);
1050:         PCGetOptionsPrefix(pc,&prefix);
1051:         KSPSetOptionsPrefix(ksp,prefix);
1052:         KSPAppendOptionsPrefix(ksp,"sub_");

1054:         jac->ksp[i] = ksp;
1055:       }
1056:     } else {
1057:       bjac = (PC_BJacobi_Multiblock*)jac->data;
1058:     }

1060:     start = 0;
1061:     for (i=0; i<n_local; i++) {
1062:       m = jac->l_lens[i];
1063:       /*
1064:       The reason we need to generate these vectors is to serve
1065:       as the right-hand side and solution vector for the solve on the
1066:       block. We do not need to allocate space for the vectors since
1067:       that is provided via VecPlaceArray() just before the call to
1068:       KSPSolve() on the block.

1070:       */
1071:       VecCreateSeq(PETSC_COMM_SELF,m,&x);
1072:       VecCreateSeqWithArray(PETSC_COMM_SELF,1,m,NULL,&y);
1073: #ifdef PETSC_HAVE_CUSP
1074:       VecSetType(x,VECCUSP);
1075:       VecSetType(y,VECCUSP);
1076: #endif
1077:       PetscLogObjectParent((PetscObject)pc,(PetscObject)x);
1078:       PetscLogObjectParent((PetscObject)pc,(PetscObject)y);

1080:       bjac->x[i]      = x;
1081:       bjac->y[i]      = y;
1082:       bjac->starts[i] = start;

1084:       ISCreateStride(PETSC_COMM_SELF,m,start,1,&is);
1085:       bjac->is[i] = is;
1086:       PetscLogObjectParent((PetscObject)pc,(PetscObject)is);

1088:       start += m;
1089:     }
1090:   } else {
1091:     bjac = (PC_BJacobi_Multiblock*)jac->data;
1092:     /*
1093:        Destroy the blocks from the previous iteration
1094:     */
1095:     if (pc->flag == DIFFERENT_NONZERO_PATTERN) {
1096:       MatDestroyMatrices(n_local,&bjac->pmat);
1097:       if (pc->useAmat) {
1098:         MatDestroyMatrices(n_local,&bjac->mat);
1099:       }
1100:       scall = MAT_INITIAL_MATRIX;
1101:     } else scall = MAT_REUSE_MATRIX;
1102:   }

1104:   MatGetSubMatrices(pmat,n_local,bjac->is,bjac->is,scall,&bjac->pmat);
1105:   if (pc->useAmat) {
1106:     PetscObjectGetOptionsPrefix((PetscObject)mat,&mprefix);
1107:     MatGetSubMatrices(mat,n_local,bjac->is,bjac->is,scall,&bjac->mat);
1108:   }
1109:   /* Return control to the user so that the submatrices can be modified (e.g., to apply
1110:      different boundary conditions for the submatrices than for the global problem) */
1111:   PCModifySubMatrices(pc,n_local,bjac->is,bjac->is,bjac->pmat,pc->modifysubmatricesP);

1113:   PetscObjectGetOptionsPrefix((PetscObject)pmat,&pprefix);
1114:   for (i=0; i<n_local; i++) {
1115:     PetscLogObjectParent((PetscObject)pc,(PetscObject)bjac->pmat[i]);
1116:     PetscObjectSetOptionsPrefix((PetscObject)bjac->pmat[i],pprefix);
1117:     if (pc->useAmat) {
1118:       PetscLogObjectParent((PetscObject)pc,(PetscObject)bjac->mat[i]);
1119:       PetscObjectSetOptionsPrefix((PetscObject)bjac->mat[i],mprefix);
1120:       KSPSetOperators(jac->ksp[i],bjac->mat[i],bjac->pmat[i]);
1121:     } else {
1122:       KSPSetOperators(jac->ksp[i],bjac->pmat[i],bjac->pmat[i]);
1123:     }
1124:     if (pc->setfromoptionscalled) {
1125:       KSPSetFromOptions(jac->ksp[i]);
1126:     }
1127:   }
1128:   return(0);
1129: }

1131: /* ---------------------------------------------------------------------------------------------*/
1132: /*
1133:       These are for a single block with multiple processes;
1134: */
1137: static PetscErrorCode PCReset_BJacobi_Multiproc(PC pc)
1138: {
1139:   PC_BJacobi           *jac   = (PC_BJacobi*)pc->data;
1140:   PC_BJacobi_Multiproc *mpjac = (PC_BJacobi_Multiproc*)jac->data;
1141:   PetscErrorCode       ierr;

1144:   VecDestroy(&mpjac->ysub);
1145:   VecDestroy(&mpjac->xsub);
1146:   MatDestroy(&mpjac->submats);
1147:   if (jac->ksp) {KSPReset(jac->ksp[0]);}
1148:   return(0);
1149: }

1153: static PetscErrorCode PCDestroy_BJacobi_Multiproc(PC pc)
1154: {
1155:   PC_BJacobi           *jac   = (PC_BJacobi*)pc->data;
1156:   PC_BJacobi_Multiproc *mpjac = (PC_BJacobi_Multiproc*)jac->data;
1157:   PetscErrorCode       ierr;

1160:   PCReset_BJacobi_Multiproc(pc);
1161:   KSPDestroy(&jac->ksp[0]);
1162:   PetscFree(jac->ksp);
1163:   PetscSubcommDestroy(&mpjac->psubcomm);

1165:   PetscFree(mpjac);
1166:   PetscFree(pc->data);
1167:   return(0);
1168: }

1172: static PetscErrorCode PCApply_BJacobi_Multiproc(PC pc,Vec x,Vec y)
1173: {
1174:   PC_BJacobi           *jac   = (PC_BJacobi*)pc->data;
1175:   PC_BJacobi_Multiproc *mpjac = (PC_BJacobi_Multiproc*)jac->data;
1176:   PetscErrorCode       ierr;
1177:   PetscScalar          *yarray;
1178:   const PetscScalar    *xarray;

1181:   /* place x's and y's local arrays into xsub and ysub */
1182:   VecGetArrayRead(x,&xarray);
1183:   VecGetArray(y,&yarray);
1184:   VecPlaceArray(mpjac->xsub,xarray);
1185:   VecPlaceArray(mpjac->ysub,yarray);

1187:   /* apply preconditioner on each matrix block */
1188:   PetscLogEventBegin(PC_ApplyOnMproc,jac->ksp[0],mpjac->xsub,mpjac->ysub,0);
1189:   KSPSolve(jac->ksp[0],mpjac->xsub,mpjac->ysub);
1190:   PetscLogEventEnd(PC_ApplyOnMproc,jac->ksp[0],mpjac->xsub,mpjac->ysub,0);

1192:   VecResetArray(mpjac->xsub);
1193:   VecResetArray(mpjac->ysub);
1194:   VecRestoreArrayRead(x,&xarray);
1195:   VecRestoreArray(y,&yarray);
1196:   return(0);
1197: }

1199: #include <petsc/private/matimpl.h>
1202: static PetscErrorCode PCSetUp_BJacobi_Multiproc(PC pc)
1203: {
1204:   PC_BJacobi           *jac   = (PC_BJacobi*)pc->data;
1205:   PC_BJacobi_Multiproc *mpjac = (PC_BJacobi_Multiproc*)jac->data;
1206:   PetscErrorCode       ierr;
1207:   PetscInt             m,n;
1208:   MPI_Comm             comm,subcomm=0;
1209:   const char           *prefix;
1210:   PetscBool            wasSetup = PETSC_TRUE;

1213:   PetscObjectGetComm((PetscObject)pc,&comm);
1214:   if (jac->n_local > 1) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Only a single block in a subcommunicator is supported");
1215:   jac->n_local = 1; /* currently only a single block is supported for a subcommunicator */
1216:   if (!pc->setupcalled) {
1217:     wasSetup  = PETSC_FALSE;
1218:     PetscNewLog(pc,&mpjac);
1219:     jac->data = (void*)mpjac;

1221:     /* initialize datastructure mpjac */
1222:     if (!jac->psubcomm) {
1223:       /* Create default contiguous subcommunicatiors if user does not provide them */
1224:       PetscSubcommCreate(comm,&jac->psubcomm);
1225:       PetscSubcommSetNumber(jac->psubcomm,jac->n);
1226:       PetscSubcommSetType(jac->psubcomm,PETSC_SUBCOMM_CONTIGUOUS);
1227:       PetscLogObjectMemory((PetscObject)pc,sizeof(PetscSubcomm));
1228:     }
1229:     mpjac->psubcomm = jac->psubcomm;
1230:     subcomm         = PetscSubcommChild(mpjac->psubcomm);

1232:     /* Get matrix blocks of pmat */
1233:     if (!pc->pmat->ops->getmultiprocblock) SETERRQ(PetscObjectComm((PetscObject)pc->pmat),PETSC_ERR_SUP,"No support for the requested operation");
1234:     (*pc->pmat->ops->getmultiprocblock)(pc->pmat,subcomm,MAT_INITIAL_MATRIX,&mpjac->submats);

1236:     /* create a new PC that processors in each subcomm have copy of */
1237:     PetscMalloc1(1,&jac->ksp);
1238:     KSPCreate(subcomm,&jac->ksp[0]);
1239:     KSPSetErrorIfNotConverged(jac->ksp[0],pc->erroriffailure);
1240:     PetscObjectIncrementTabLevel((PetscObject)jac->ksp[0],(PetscObject)pc,1);
1241:     PetscLogObjectParent((PetscObject)pc,(PetscObject)jac->ksp[0]);
1242:     KSPSetOperators(jac->ksp[0],mpjac->submats,mpjac->submats);
1243:     KSPGetPC(jac->ksp[0],&mpjac->pc);

1245:     PCGetOptionsPrefix(pc,&prefix);
1246:     KSPSetOptionsPrefix(jac->ksp[0],prefix);
1247:     KSPAppendOptionsPrefix(jac->ksp[0],"sub_");
1248:     /*
1249:       PetscMPIInt rank,subsize,subrank;
1250:       MPI_Comm_rank(comm,&rank);
1251:       MPI_Comm_size(subcomm,&subsize);
1252:       MPI_Comm_rank(subcomm,&subrank);

1254:       MatGetLocalSize(mpjac->submats,&m,NULL);
1255:       MatGetSize(mpjac->submats,&n,NULL);
1256:       PetscSynchronizedPrintf(comm,"[%d], sub-size %d,sub-rank %d\n",rank,subsize,subrank);
1257:       PetscSynchronizedFlush(comm,PETSC_STDOUT);
1258:     */

1260:     /* create dummy vectors xsub and ysub */
1261:     MatGetLocalSize(mpjac->submats,&m,&n);
1262:     VecCreateMPIWithArray(subcomm,1,n,PETSC_DECIDE,NULL,&mpjac->xsub);
1263:     VecCreateMPIWithArray(subcomm,1,m,PETSC_DECIDE,NULL,&mpjac->ysub);
1264: #ifdef PETSC_HAVE_CUSP
1265:     VecSetType(mpjac->xsub,VECMPICUSP);
1266:     VecSetType(mpjac->ysub,VECMPICUSP);
1267: #endif
1268:     PetscLogObjectParent((PetscObject)pc,(PetscObject)mpjac->xsub);
1269:     PetscLogObjectParent((PetscObject)pc,(PetscObject)mpjac->ysub);

1271:     pc->ops->reset   = PCReset_BJacobi_Multiproc;
1272:     pc->ops->destroy = PCDestroy_BJacobi_Multiproc;
1273:     pc->ops->apply   = PCApply_BJacobi_Multiproc;
1274:   } else { /* pc->setupcalled */
1275:     subcomm = PetscSubcommChild(mpjac->psubcomm);
1276:     if (pc->flag == DIFFERENT_NONZERO_PATTERN) {
1277:       /* destroy old matrix blocks, then get new matrix blocks */
1278:       if (mpjac->submats) {MatDestroy(&mpjac->submats);}
1279:       (*pc->pmat->ops->getmultiprocblock)(pc->pmat,subcomm,MAT_INITIAL_MATRIX,&mpjac->submats);
1280:     } else {
1281:       (*pc->pmat->ops->getmultiprocblock)(pc->pmat,subcomm,MAT_REUSE_MATRIX,&mpjac->submats);
1282:     }
1283:     KSPSetOperators(jac->ksp[0],mpjac->submats,mpjac->submats);
1284:   }

1286:   if (!wasSetup && pc->setfromoptionscalled) {
1287:     KSPSetFromOptions(jac->ksp[0]);
1288:   }
1289:   return(0);
1290: }