Actual source code: mpiaij.c

petsc-3.13.4 2020-08-01
Report Typos and Errors
  1:  #include <../src/mat/impls/aij/mpi/mpiaij.h>
  2:  #include <petsc/private/vecimpl.h>
  3:  #include <petsc/private/vecscatterimpl.h>
  4:  #include <petsc/private/isimpl.h>
  5:  #include <petscblaslapack.h>
  6:  #include <petscsf.h>
  7:  #include <petsc/private/hashmapi.h>

  9: /*MC
 10:    MATAIJ - MATAIJ = "aij" - A matrix type to be used for sparse matrices.

 12:    This matrix type is identical to MATSEQAIJ when constructed with a single process communicator,
 13:    and MATMPIAIJ otherwise.  As a result, for single process communicators,
 14:   MatSeqAIJSetPreallocation is supported, and similarly MatMPIAIJSetPreallocation() is supported
 15:   for communicators controlling multiple processes.  It is recommended that you call both of
 16:   the above preallocation routines for simplicity.

 18:    Options Database Keys:
 19: . -mat_type aij - sets the matrix type to "aij" during a call to MatSetFromOptions()

 21:   Developer Notes:
 22:     Subclasses include MATAIJCUSP, MATAIJCUSPARSE, MATAIJPERM, MATAIJSELL, MATAIJMKL, MATAIJCRL, and also automatically switches over to use inodes when
 23:    enough exist.

 25:   Level: beginner

 27: .seealso: MatCreateAIJ(), MatCreateSeqAIJ(), MATSEQAIJ, MATMPIAIJ
 28: M*/

 30: /*MC
 31:    MATAIJCRL - MATAIJCRL = "aijcrl" - A matrix type to be used for sparse matrices.

 33:    This matrix type is identical to MATSEQAIJCRL when constructed with a single process communicator,
 34:    and MATMPIAIJCRL otherwise.  As a result, for single process communicators,
 35:    MatSeqAIJSetPreallocation() is supported, and similarly MatMPIAIJSetPreallocation() is supported
 36:   for communicators controlling multiple processes.  It is recommended that you call both of
 37:   the above preallocation routines for simplicity.

 39:    Options Database Keys:
 40: . -mat_type aijcrl - sets the matrix type to "aijcrl" during a call to MatSetFromOptions()

 42:   Level: beginner

 44: .seealso: MatCreateMPIAIJCRL,MATSEQAIJCRL,MATMPIAIJCRL, MATSEQAIJCRL, MATMPIAIJCRL
 45: M*/

 47: static PetscErrorCode MatBindToCPU_MPIAIJ(Mat A,PetscBool flg)
 48: {
 49:   Mat_MPIAIJ     *a = (Mat_MPIAIJ*)A->data;

 53: #if defined(PETSC_HAVE_CUDA) || defined(PETSC_HAVE_VIENNACL)
 54:   A->boundtocpu = flg;
 55: #endif
 56:   if (a->A) {
 57:     MatBindToCPU(a->A,flg);
 58:   }
 59:   if (a->B) {
 60:     MatBindToCPU(a->B,flg);
 61:   }
 62:   return(0);
 63: }


 66: PetscErrorCode MatSetBlockSizes_MPIAIJ(Mat M, PetscInt rbs, PetscInt cbs)
 67: {
 69:   Mat_MPIAIJ     *mat = (Mat_MPIAIJ*)M->data;

 72:   if (mat->A) {
 73:     MatSetBlockSizes(mat->A,rbs,cbs);
 74:     MatSetBlockSizes(mat->B,rbs,1);
 75:   }
 76:   return(0);
 77: }

 79: PetscErrorCode MatFindNonzeroRows_MPIAIJ(Mat M,IS *keptrows)
 80: {
 81:   PetscErrorCode  ierr;
 82:   Mat_MPIAIJ      *mat = (Mat_MPIAIJ*)M->data;
 83:   Mat_SeqAIJ      *a   = (Mat_SeqAIJ*)mat->A->data;
 84:   Mat_SeqAIJ      *b   = (Mat_SeqAIJ*)mat->B->data;
 85:   const PetscInt  *ia,*ib;
 86:   const MatScalar *aa,*bb;
 87:   PetscInt        na,nb,i,j,*rows,cnt=0,n0rows;
 88:   PetscInt        m = M->rmap->n,rstart = M->rmap->rstart;

 91:   *keptrows = 0;
 92:   ia        = a->i;
 93:   ib        = b->i;
 94:   for (i=0; i<m; i++) {
 95:     na = ia[i+1] - ia[i];
 96:     nb = ib[i+1] - ib[i];
 97:     if (!na && !nb) {
 98:       cnt++;
 99:       goto ok1;
100:     }
101:     aa = a->a + ia[i];
102:     for (j=0; j<na; j++) {
103:       if (aa[j] != 0.0) goto ok1;
104:     }
105:     bb = b->a + ib[i];
106:     for (j=0; j <nb; j++) {
107:       if (bb[j] != 0.0) goto ok1;
108:     }
109:     cnt++;
110: ok1:;
111:   }
112:   MPIU_Allreduce(&cnt,&n0rows,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)M));
113:   if (!n0rows) return(0);
114:   PetscMalloc1(M->rmap->n-cnt,&rows);
115:   cnt  = 0;
116:   for (i=0; i<m; i++) {
117:     na = ia[i+1] - ia[i];
118:     nb = ib[i+1] - ib[i];
119:     if (!na && !nb) continue;
120:     aa = a->a + ia[i];
121:     for (j=0; j<na;j++) {
122:       if (aa[j] != 0.0) {
123:         rows[cnt++] = rstart + i;
124:         goto ok2;
125:       }
126:     }
127:     bb = b->a + ib[i];
128:     for (j=0; j<nb; j++) {
129:       if (bb[j] != 0.0) {
130:         rows[cnt++] = rstart + i;
131:         goto ok2;
132:       }
133:     }
134: ok2:;
135:   }
136:   ISCreateGeneral(PetscObjectComm((PetscObject)M),cnt,rows,PETSC_OWN_POINTER,keptrows);
137:   return(0);
138: }

140: PetscErrorCode  MatDiagonalSet_MPIAIJ(Mat Y,Vec D,InsertMode is)
141: {
142:   PetscErrorCode    ierr;
143:   Mat_MPIAIJ        *aij = (Mat_MPIAIJ*) Y->data;
144:   PetscBool         cong;

147:   MatHasCongruentLayouts(Y,&cong);
148:   if (Y->assembled && cong) {
149:     MatDiagonalSet(aij->A,D,is);
150:   } else {
151:     MatDiagonalSet_Default(Y,D,is);
152:   }
153:   return(0);
154: }

156: PetscErrorCode MatFindZeroDiagonals_MPIAIJ(Mat M,IS *zrows)
157: {
158:   Mat_MPIAIJ     *aij = (Mat_MPIAIJ*)M->data;
160:   PetscInt       i,rstart,nrows,*rows;

163:   *zrows = NULL;
164:   MatFindZeroDiagonals_SeqAIJ_Private(aij->A,&nrows,&rows);
165:   MatGetOwnershipRange(M,&rstart,NULL);
166:   for (i=0; i<nrows; i++) rows[i] += rstart;
167:   ISCreateGeneral(PetscObjectComm((PetscObject)M),nrows,rows,PETSC_OWN_POINTER,zrows);
168:   return(0);
169: }

171: PetscErrorCode MatGetColumnNorms_MPIAIJ(Mat A,NormType type,PetscReal *norms)
172: {
174:   Mat_MPIAIJ     *aij = (Mat_MPIAIJ*)A->data;
175:   PetscInt       i,n,*garray = aij->garray;
176:   Mat_SeqAIJ     *a_aij = (Mat_SeqAIJ*) aij->A->data;
177:   Mat_SeqAIJ     *b_aij = (Mat_SeqAIJ*) aij->B->data;
178:   PetscReal      *work;

181:   MatGetSize(A,NULL,&n);
182:   PetscCalloc1(n,&work);
183:   if (type == NORM_2) {
184:     for (i=0; i<a_aij->i[aij->A->rmap->n]; i++) {
185:       work[A->cmap->rstart + a_aij->j[i]] += PetscAbsScalar(a_aij->a[i]*a_aij->a[i]);
186:     }
187:     for (i=0; i<b_aij->i[aij->B->rmap->n]; i++) {
188:       work[garray[b_aij->j[i]]] += PetscAbsScalar(b_aij->a[i]*b_aij->a[i]);
189:     }
190:   } else if (type == NORM_1) {
191:     for (i=0; i<a_aij->i[aij->A->rmap->n]; i++) {
192:       work[A->cmap->rstart + a_aij->j[i]] += PetscAbsScalar(a_aij->a[i]);
193:     }
194:     for (i=0; i<b_aij->i[aij->B->rmap->n]; i++) {
195:       work[garray[b_aij->j[i]]] += PetscAbsScalar(b_aij->a[i]);
196:     }
197:   } else if (type == NORM_INFINITY) {
198:     for (i=0; i<a_aij->i[aij->A->rmap->n]; i++) {
199:       work[A->cmap->rstart + a_aij->j[i]] = PetscMax(PetscAbsScalar(a_aij->a[i]), work[A->cmap->rstart + a_aij->j[i]]);
200:     }
201:     for (i=0; i<b_aij->i[aij->B->rmap->n]; i++) {
202:       work[garray[b_aij->j[i]]] = PetscMax(PetscAbsScalar(b_aij->a[i]),work[garray[b_aij->j[i]]]);
203:     }

205:   } else SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_WRONG,"Unknown NormType");
206:   if (type == NORM_INFINITY) {
207:     MPIU_Allreduce(work,norms,n,MPIU_REAL,MPIU_MAX,PetscObjectComm((PetscObject)A));
208:   } else {
209:     MPIU_Allreduce(work,norms,n,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)A));
210:   }
211:   PetscFree(work);
212:   if (type == NORM_2) {
213:     for (i=0; i<n; i++) norms[i] = PetscSqrtReal(norms[i]);
214:   }
215:   return(0);
216: }

218: PetscErrorCode MatFindOffBlockDiagonalEntries_MPIAIJ(Mat A,IS *is)
219: {
220:   Mat_MPIAIJ      *a  = (Mat_MPIAIJ*)A->data;
221:   IS              sis,gis;
222:   PetscErrorCode  ierr;
223:   const PetscInt  *isis,*igis;
224:   PetscInt        n,*iis,nsis,ngis,rstart,i;

227:   MatFindOffBlockDiagonalEntries(a->A,&sis);
228:   MatFindNonzeroRows(a->B,&gis);
229:   ISGetSize(gis,&ngis);
230:   ISGetSize(sis,&nsis);
231:   ISGetIndices(sis,&isis);
232:   ISGetIndices(gis,&igis);

234:   PetscMalloc1(ngis+nsis,&iis);
235:   PetscArraycpy(iis,igis,ngis);
236:   PetscArraycpy(iis+ngis,isis,nsis);
237:   n    = ngis + nsis;
238:   PetscSortRemoveDupsInt(&n,iis);
239:   MatGetOwnershipRange(A,&rstart,NULL);
240:   for (i=0; i<n; i++) iis[i] += rstart;
241:   ISCreateGeneral(PetscObjectComm((PetscObject)A),n,iis,PETSC_OWN_POINTER,is);

243:   ISRestoreIndices(sis,&isis);
244:   ISRestoreIndices(gis,&igis);
245:   ISDestroy(&sis);
246:   ISDestroy(&gis);
247:   return(0);
248: }

250: /*
251:     Distributes a SeqAIJ matrix across a set of processes. Code stolen from
252:     MatLoad_MPIAIJ(). Horrible lack of reuse. Should be a routine for each matrix type.

254:     Only for square matrices

256:     Used by a preconditioner, hence PETSC_EXTERN
257: */
258: PETSC_EXTERN PetscErrorCode MatDistribute_MPIAIJ(MPI_Comm comm,Mat gmat,PetscInt m,MatReuse reuse,Mat *inmat)
259: {
260:   PetscMPIInt    rank,size;
261:   PetscInt       *rowners,*dlens,*olens,i,rstart,rend,j,jj,nz = 0,*gmataj,cnt,row,*ld,bses[2];
263:   Mat            mat;
264:   Mat_SeqAIJ     *gmata;
265:   PetscMPIInt    tag;
266:   MPI_Status     status;
267:   PetscBool      aij;
268:   MatScalar      *gmataa,*ao,*ad,*gmataarestore=0;

271:   MPI_Comm_rank(comm,&rank);
272:   MPI_Comm_size(comm,&size);
273:   if (!rank) {
274:     PetscObjectTypeCompare((PetscObject)gmat,MATSEQAIJ,&aij);
275:     if (!aij) SETERRQ1(PetscObjectComm((PetscObject)gmat),PETSC_ERR_SUP,"Currently no support for input matrix of type %s\n",((PetscObject)gmat)->type_name);
276:   }
277:   if (reuse == MAT_INITIAL_MATRIX) {
278:     MatCreate(comm,&mat);
279:     MatSetSizes(mat,m,m,PETSC_DETERMINE,PETSC_DETERMINE);
280:     MatGetBlockSizes(gmat,&bses[0],&bses[1]);
281:     MPI_Bcast(bses,2,MPIU_INT,0,comm);
282:     MatSetBlockSizes(mat,bses[0],bses[1]);
283:     MatSetType(mat,MATAIJ);
284:     PetscMalloc1(size+1,&rowners);
285:     PetscMalloc2(m,&dlens,m,&olens);
286:     MPI_Allgather(&m,1,MPIU_INT,rowners+1,1,MPIU_INT,comm);

288:     rowners[0] = 0;
289:     for (i=2; i<=size; i++) rowners[i] += rowners[i-1];
290:     rstart = rowners[rank];
291:     rend   = rowners[rank+1];
292:     PetscObjectGetNewTag((PetscObject)mat,&tag);
293:     if (!rank) {
294:       gmata = (Mat_SeqAIJ*) gmat->data;
295:       /* send row lengths to all processors */
296:       for (i=0; i<m; i++) dlens[i] = gmata->ilen[i];
297:       for (i=1; i<size; i++) {
298:         MPI_Send(gmata->ilen + rowners[i],rowners[i+1]-rowners[i],MPIU_INT,i,tag,comm);
299:       }
300:       /* determine number diagonal and off-diagonal counts */
301:       PetscArrayzero(olens,m);
302:       PetscCalloc1(m,&ld);
303:       jj   = 0;
304:       for (i=0; i<m; i++) {
305:         for (j=0; j<dlens[i]; j++) {
306:           if (gmata->j[jj] < rstart) ld[i]++;
307:           if (gmata->j[jj] < rstart || gmata->j[jj] >= rend) olens[i]++;
308:           jj++;
309:         }
310:       }
311:       /* send column indices to other processes */
312:       for (i=1; i<size; i++) {
313:         nz   = gmata->i[rowners[i+1]]-gmata->i[rowners[i]];
314:         MPI_Send(&nz,1,MPIU_INT,i,tag,comm);
315:         MPI_Send(gmata->j + gmata->i[rowners[i]],nz,MPIU_INT,i,tag,comm);
316:       }

318:       /* send numerical values to other processes */
319:       for (i=1; i<size; i++) {
320:         nz   = gmata->i[rowners[i+1]]-gmata->i[rowners[i]];
321:         MPI_Send(gmata->a + gmata->i[rowners[i]],nz,MPIU_SCALAR,i,tag,comm);
322:       }
323:       gmataa = gmata->a;
324:       gmataj = gmata->j;

326:     } else {
327:       /* receive row lengths */
328:       MPI_Recv(dlens,m,MPIU_INT,0,tag,comm,&status);
329:       /* receive column indices */
330:       MPI_Recv(&nz,1,MPIU_INT,0,tag,comm,&status);
331:       PetscMalloc2(nz,&gmataa,nz,&gmataj);
332:       MPI_Recv(gmataj,nz,MPIU_INT,0,tag,comm,&status);
333:       /* determine number diagonal and off-diagonal counts */
334:       PetscArrayzero(olens,m);
335:       PetscCalloc1(m,&ld);
336:       jj   = 0;
337:       for (i=0; i<m; i++) {
338:         for (j=0; j<dlens[i]; j++) {
339:           if (gmataj[jj] < rstart) ld[i]++;
340:           if (gmataj[jj] < rstart || gmataj[jj] >= rend) olens[i]++;
341:           jj++;
342:         }
343:       }
344:       /* receive numerical values */
345:       PetscArrayzero(gmataa,nz);
346:       MPI_Recv(gmataa,nz,MPIU_SCALAR,0,tag,comm,&status);
347:     }
348:     /* set preallocation */
349:     for (i=0; i<m; i++) {
350:       dlens[i] -= olens[i];
351:     }
352:     MatSeqAIJSetPreallocation(mat,0,dlens);
353:     MatMPIAIJSetPreallocation(mat,0,dlens,0,olens);

355:     for (i=0; i<m; i++) {
356:       dlens[i] += olens[i];
357:     }
358:     cnt = 0;
359:     for (i=0; i<m; i++) {
360:       row  = rstart + i;
361:       MatSetValues(mat,1,&row,dlens[i],gmataj+cnt,gmataa+cnt,INSERT_VALUES);
362:       cnt += dlens[i];
363:     }
364:     if (rank) {
365:       PetscFree2(gmataa,gmataj);
366:     }
367:     PetscFree2(dlens,olens);
368:     PetscFree(rowners);

370:     ((Mat_MPIAIJ*)(mat->data))->ld = ld;

372:     *inmat = mat;
373:   } else {   /* column indices are already set; only need to move over numerical values from process 0 */
374:     Mat_SeqAIJ *Ad = (Mat_SeqAIJ*)((Mat_MPIAIJ*)((*inmat)->data))->A->data;
375:     Mat_SeqAIJ *Ao = (Mat_SeqAIJ*)((Mat_MPIAIJ*)((*inmat)->data))->B->data;
376:     mat  = *inmat;
377:     PetscObjectGetNewTag((PetscObject)mat,&tag);
378:     if (!rank) {
379:       /* send numerical values to other processes */
380:       gmata  = (Mat_SeqAIJ*) gmat->data;
381:       MatGetOwnershipRanges(mat,(const PetscInt**)&rowners);
382:       gmataa = gmata->a;
383:       for (i=1; i<size; i++) {
384:         nz   = gmata->i[rowners[i+1]]-gmata->i[rowners[i]];
385:         MPI_Send(gmataa + gmata->i[rowners[i]],nz,MPIU_SCALAR,i,tag,comm);
386:       }
387:       nz = gmata->i[rowners[1]]-gmata->i[rowners[0]];
388:     } else {
389:       /* receive numerical values from process 0*/
390:       nz   = Ad->nz + Ao->nz;
391:       PetscMalloc1(nz,&gmataa); gmataarestore = gmataa;
392:       MPI_Recv(gmataa,nz,MPIU_SCALAR,0,tag,comm,&status);
393:     }
394:     /* transfer numerical values into the diagonal A and off diagonal B parts of mat */
395:     ld = ((Mat_MPIAIJ*)(mat->data))->ld;
396:     ad = Ad->a;
397:     ao = Ao->a;
398:     if (mat->rmap->n) {
399:       i  = 0;
400:       nz = ld[i];                                   PetscArraycpy(ao,gmataa,nz); ao += nz; gmataa += nz;
401:       nz = Ad->i[i+1] - Ad->i[i];                   PetscArraycpy(ad,gmataa,nz); ad += nz; gmataa += nz;
402:     }
403:     for (i=1; i<mat->rmap->n; i++) {
404:       nz = Ao->i[i] - Ao->i[i-1] - ld[i-1] + ld[i]; PetscArraycpy(ao,gmataa,nz); ao += nz; gmataa += nz;
405:       nz = Ad->i[i+1] - Ad->i[i];                   PetscArraycpy(ad,gmataa,nz); ad += nz; gmataa += nz;
406:     }
407:     i--;
408:     if (mat->rmap->n) {
409:       nz = Ao->i[i+1] - Ao->i[i] - ld[i];           PetscArraycpy(ao,gmataa,nz);
410:     }
411:     if (rank) {
412:       PetscFree(gmataarestore);
413:     }
414:   }
415:   MatAssemblyBegin(mat,MAT_FINAL_ASSEMBLY);
416:   MatAssemblyEnd(mat,MAT_FINAL_ASSEMBLY);
417:   return(0);
418: }

420: /*
421:   Local utility routine that creates a mapping from the global column
422: number to the local number in the off-diagonal part of the local
423: storage of the matrix.  When PETSC_USE_CTABLE is used this is scalable at
424: a slightly higher hash table cost; without it it is not scalable (each processor
425: has an order N integer array but is fast to acess.
426: */
427: PetscErrorCode MatCreateColmap_MPIAIJ_Private(Mat mat)
428: {
429:   Mat_MPIAIJ     *aij = (Mat_MPIAIJ*)mat->data;
431:   PetscInt       n = aij->B->cmap->n,i;

434:   if (!aij->garray) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"MPIAIJ Matrix was assembled but is missing garray");
435: #if defined(PETSC_USE_CTABLE)
436:   PetscTableCreate(n,mat->cmap->N+1,&aij->colmap);
437:   for (i=0; i<n; i++) {
438:     PetscTableAdd(aij->colmap,aij->garray[i]+1,i+1,INSERT_VALUES);
439:   }
440: #else
441:   PetscCalloc1(mat->cmap->N+1,&aij->colmap);
442:   PetscLogObjectMemory((PetscObject)mat,(mat->cmap->N+1)*sizeof(PetscInt));
443:   for (i=0; i<n; i++) aij->colmap[aij->garray[i]] = i+1;
444: #endif
445:   return(0);
446: }

448: #define MatSetValues_SeqAIJ_A_Private(row,col,value,addv,orow,ocol)     \
449: { \
450:     if (col <= lastcol1)  low1 = 0;     \
451:     else                 high1 = nrow1; \
452:     lastcol1 = col;\
453:     while (high1-low1 > 5) { \
454:       t = (low1+high1)/2; \
455:       if (rp1[t] > col) high1 = t; \
456:       else              low1  = t; \
457:     } \
458:       for (_i=low1; _i<high1; _i++) { \
459:         if (rp1[_i] > col) break; \
460:         if (rp1[_i] == col) { \
461:           if (addv == ADD_VALUES) { \
462:             ap1[_i] += value;   \
463:             /* Not sure LogFlops will slow dow the code or not */ \
464:             (void)PetscLogFlops(1.0);   \
465:            } \
466:           else                    ap1[_i] = value; \
467:           inserted = PETSC_TRUE; \
468:           goto a_noinsert; \
469:         } \
470:       }  \
471:       if (value == 0.0 && ignorezeroentries && row != col) {low1 = 0; high1 = nrow1;goto a_noinsert;} \
472:       if (nonew == 1) {low1 = 0; high1 = nrow1; goto a_noinsert;}                \
473:       if (nonew == -1) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero at global row/column (%D, %D) into matrix", orow, ocol); \
474:       MatSeqXAIJReallocateAIJ(A,am,1,nrow1,row,col,rmax1,aa,ai,aj,rp1,ap1,aimax,nonew,MatScalar); \
475:       N = nrow1++ - 1; a->nz++; high1++; \
476:       /* shift up all the later entries in this row */ \
477:       PetscArraymove(rp1+_i+1,rp1+_i,N-_i+1);\
478:       PetscArraymove(ap1+_i+1,ap1+_i,N-_i+1);\
479:       rp1[_i] = col;  \
480:       ap1[_i] = value;  \
481:       A->nonzerostate++;\
482:       a_noinsert: ; \
483:       ailen[row] = nrow1; \
484: }

486: #define MatSetValues_SeqAIJ_B_Private(row,col,value,addv,orow,ocol) \
487:   { \
488:     if (col <= lastcol2) low2 = 0;                        \
489:     else high2 = nrow2;                                   \
490:     lastcol2 = col;                                       \
491:     while (high2-low2 > 5) {                              \
492:       t = (low2+high2)/2;                                 \
493:       if (rp2[t] > col) high2 = t;                        \
494:       else             low2  = t;                         \
495:     }                                                     \
496:     for (_i=low2; _i<high2; _i++) {                       \
497:       if (rp2[_i] > col) break;                           \
498:       if (rp2[_i] == col) {                               \
499:         if (addv == ADD_VALUES) {                         \
500:           ap2[_i] += value;                               \
501:           (void)PetscLogFlops(1.0);                       \
502:         }                                                 \
503:         else                    ap2[_i] = value;          \
504:         inserted = PETSC_TRUE;                            \
505:         goto b_noinsert;                                  \
506:       }                                                   \
507:     }                                                     \
508:     if (value == 0.0 && ignorezeroentries) {low2 = 0; high2 = nrow2; goto b_noinsert;} \
509:     if (nonew == 1) {low2 = 0; high2 = nrow2; goto b_noinsert;}                        \
510:     if (nonew == -1) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero at global row/column (%D, %D) into matrix", orow, ocol); \
511:     MatSeqXAIJReallocateAIJ(B,bm,1,nrow2,row,col,rmax2,ba,bi,bj,rp2,ap2,bimax,nonew,MatScalar); \
512:     N = nrow2++ - 1; b->nz++; high2++;                    \
513:     /* shift up all the later entries in this row */      \
514:     PetscArraymove(rp2+_i+1,rp2+_i,N-_i+1);\
515:     PetscArraymove(ap2+_i+1,ap2+_i,N-_i+1);\
516:     rp2[_i] = col;                                        \
517:     ap2[_i] = value;                                      \
518:     B->nonzerostate++;                                    \
519:     b_noinsert: ;                                         \
520:     bilen[row] = nrow2;                                   \
521:   }

523: PetscErrorCode MatSetValuesRow_MPIAIJ(Mat A,PetscInt row,const PetscScalar v[])
524: {
525:   Mat_MPIAIJ     *mat = (Mat_MPIAIJ*)A->data;
526:   Mat_SeqAIJ     *a   = (Mat_SeqAIJ*)mat->A->data,*b = (Mat_SeqAIJ*)mat->B->data;
528:   PetscInt       l,*garray = mat->garray,diag;

531:   /* code only works for square matrices A */

533:   /* find size of row to the left of the diagonal part */
534:   MatGetOwnershipRange(A,&diag,0);
535:   row  = row - diag;
536:   for (l=0; l<b->i[row+1]-b->i[row]; l++) {
537:     if (garray[b->j[b->i[row]+l]] > diag) break;
538:   }
539:   PetscArraycpy(b->a+b->i[row],v,l);

541:   /* diagonal part */
542:   PetscArraycpy(a->a+a->i[row],v+l,(a->i[row+1]-a->i[row]));

544:   /* right of diagonal part */
545:   PetscArraycpy(b->a+b->i[row]+l,v+l+a->i[row+1]-a->i[row],b->i[row+1]-b->i[row]-l);
546: #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA)
547:   if (A->offloadmask != PETSC_OFFLOAD_UNALLOCATED && (l || (a->i[row+1]-a->i[row]) || (b->i[row+1]-b->i[row]-l))) A->offloadmask = PETSC_OFFLOAD_CPU;
548: #endif
549:   return(0);
550: }

552: PetscErrorCode MatSetValues_MPIAIJ(Mat mat,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],const PetscScalar v[],InsertMode addv)
553: {
554:   Mat_MPIAIJ     *aij = (Mat_MPIAIJ*)mat->data;
555:   PetscScalar    value = 0.0;
557:   PetscInt       i,j,rstart  = mat->rmap->rstart,rend = mat->rmap->rend;
558:   PetscInt       cstart      = mat->cmap->rstart,cend = mat->cmap->rend,row,col;
559:   PetscBool      roworiented = aij->roworiented;

561:   /* Some Variables required in the macro */
562:   Mat        A                    = aij->A;
563:   Mat_SeqAIJ *a                   = (Mat_SeqAIJ*)A->data;
564:   PetscInt   *aimax               = a->imax,*ai = a->i,*ailen = a->ilen,*aj = a->j;
565:   MatScalar  *aa                  = a->a;
566:   PetscBool  ignorezeroentries    = a->ignorezeroentries;
567:   Mat        B                    = aij->B;
568:   Mat_SeqAIJ *b                   = (Mat_SeqAIJ*)B->data;
569:   PetscInt   *bimax               = b->imax,*bi = b->i,*bilen = b->ilen,*bj = b->j,bm = aij->B->rmap->n,am = aij->A->rmap->n;
570:   MatScalar  *ba                  = b->a;
571:   /* This variable below is only for the PETSC_HAVE_VIENNACL or PETSC_HAVE_CUDA cases, but we define it in all cases because we
572:    * cannot use "#if defined" inside a macro. */
573:   PETSC_UNUSED PetscBool inserted = PETSC_FALSE;

575:   PetscInt  *rp1,*rp2,ii,nrow1,nrow2,_i,rmax1,rmax2,N,low1,high1,low2,high2,t,lastcol1,lastcol2;
576:   PetscInt  nonew;
577:   MatScalar *ap1,*ap2;

580:   for (i=0; i<m; i++) {
581:     if (im[i] < 0) continue;
582: #if defined(PETSC_USE_DEBUG)
583:     if (im[i] >= mat->rmap->N) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Row too large: row %D max %D",im[i],mat->rmap->N-1);
584: #endif
585:     if (im[i] >= rstart && im[i] < rend) {
586:       row      = im[i] - rstart;
587:       lastcol1 = -1;
588:       rp1      = aj + ai[row];
589:       ap1      = aa + ai[row];
590:       rmax1    = aimax[row];
591:       nrow1    = ailen[row];
592:       low1     = 0;
593:       high1    = nrow1;
594:       lastcol2 = -1;
595:       rp2      = bj + bi[row];
596:       ap2      = ba + bi[row];
597:       rmax2    = bimax[row];
598:       nrow2    = bilen[row];
599:       low2     = 0;
600:       high2    = nrow2;

602:       for (j=0; j<n; j++) {
603:         if (v)  value = roworiented ? v[i*n+j] : v[i+j*m];
604:         if (ignorezeroentries && value == 0.0 && (addv == ADD_VALUES) && im[i] != in[j]) continue;
605:         if (in[j] >= cstart && in[j] < cend) {
606:           col   = in[j] - cstart;
607:           nonew = a->nonew;
608:           MatSetValues_SeqAIJ_A_Private(row,col,value,addv,im[i],in[j]);
609: #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA)
610:           if (A->offloadmask != PETSC_OFFLOAD_UNALLOCATED && inserted) A->offloadmask = PETSC_OFFLOAD_CPU;
611: #endif
612:         } else if (in[j] < 0) continue;
613: #if defined(PETSC_USE_DEBUG)
614:         else if (in[j] >= mat->cmap->N) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Column too large: col %D max %D",in[j],mat->cmap->N-1);
615: #endif
616:         else {
617:           if (mat->was_assembled) {
618:             if (!aij->colmap) {
619:               MatCreateColmap_MPIAIJ_Private(mat);
620:             }
621: #if defined(PETSC_USE_CTABLE)
622:             PetscTableFind(aij->colmap,in[j]+1,&col);
623:             col--;
624: #else
625:             col = aij->colmap[in[j]] - 1;
626: #endif
627:             if (col < 0 && !((Mat_SeqAIJ*)(aij->B->data))->nonew) {
628:               MatDisAssemble_MPIAIJ(mat);
629:               col  =  in[j];
630:               /* Reinitialize the variables required by MatSetValues_SeqAIJ_B_Private() */
631:               B        = aij->B;
632:               b        = (Mat_SeqAIJ*)B->data;
633:               bimax    = b->imax; bi = b->i; bilen = b->ilen; bj = b->j; ba = b->a;
634:               rp2      = bj + bi[row];
635:               ap2      = ba + bi[row];
636:               rmax2    = bimax[row];
637:               nrow2    = bilen[row];
638:               low2     = 0;
639:               high2    = nrow2;
640:               bm       = aij->B->rmap->n;
641:               ba       = b->a;
642:               inserted = PETSC_FALSE;
643:             } else if (col < 0) {
644:               if (1 == ((Mat_SeqAIJ*)(aij->B->data))->nonew) {
645:                 PetscInfo3(mat,"Skipping of insertion of new nonzero location in off-diagonal portion of matrix %g(%D,%D)\n",(double)PetscRealPart(value),im[i],in[j]);
646:               } else SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero at global row/column (%D, %D) into matrix", im[i], in[j]);
647:             }
648:           } else col = in[j];
649:           nonew = b->nonew;
650:           MatSetValues_SeqAIJ_B_Private(row,col,value,addv,im[i],in[j]);
651: #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA)
652:           if (B->offloadmask != PETSC_OFFLOAD_UNALLOCATED && inserted) B->offloadmask = PETSC_OFFLOAD_CPU;
653: #endif
654:         }
655:       }
656:     } else {
657:       if (mat->nooffprocentries) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Setting off process row %D even though MatSetOption(,MAT_NO_OFF_PROC_ENTRIES,PETSC_TRUE) was set",im[i]);
658:       if (!aij->donotstash) {
659:         mat->assembled = PETSC_FALSE;
660:         if (roworiented) {
661:           MatStashValuesRow_Private(&mat->stash,im[i],n,in,v+i*n,(PetscBool)(ignorezeroentries && (addv == ADD_VALUES)));
662:         } else {
663:           MatStashValuesCol_Private(&mat->stash,im[i],n,in,v+i,m,(PetscBool)(ignorezeroentries && (addv == ADD_VALUES)));
664:         }
665:       }
666:     }
667:   }
668:   return(0);
669: }

671: /*
672:     This function sets the j and ilen arrays (of the diagonal and off-diagonal part) of an MPIAIJ-matrix.
673:     The values in mat_i have to be sorted and the values in mat_j have to be sorted for each row (CSR-like).
674:     No off-processor parts off the matrix are allowed here and mat->was_assembled has to be PETSC_FALSE.
675: */
676: PetscErrorCode MatSetValues_MPIAIJ_CopyFromCSRFormat_Symbolic(Mat mat,const PetscInt mat_j[],const PetscInt mat_i[])
677: {
678:   Mat_MPIAIJ     *aij        = (Mat_MPIAIJ*)mat->data;
679:   Mat            A           = aij->A; /* diagonal part of the matrix */
680:   Mat            B           = aij->B; /* offdiagonal part of the matrix */
681:   Mat_SeqAIJ     *a          = (Mat_SeqAIJ*)A->data;
682:   Mat_SeqAIJ     *b          = (Mat_SeqAIJ*)B->data;
683:   PetscInt       cstart      = mat->cmap->rstart,cend = mat->cmap->rend,col;
684:   PetscInt       *ailen      = a->ilen,*aj = a->j;
685:   PetscInt       *bilen      = b->ilen,*bj = b->j;
686:   PetscInt       am          = aij->A->rmap->n,j;
687:   PetscInt       diag_so_far = 0,dnz;
688:   PetscInt       offd_so_far = 0,onz;

691:   /* Iterate over all rows of the matrix */
692:   for (j=0; j<am; j++) {
693:     dnz = onz = 0;
694:     /*  Iterate over all non-zero columns of the current row */
695:     for (col=mat_i[j]; col<mat_i[j+1]; col++) {
696:       /* If column is in the diagonal */
697:       if (mat_j[col] >= cstart && mat_j[col] < cend) {
698:         aj[diag_so_far++] = mat_j[col] - cstart;
699:         dnz++;
700:       } else { /* off-diagonal entries */
701:         bj[offd_so_far++] = mat_j[col];
702:         onz++;
703:       }
704:     }
705:     ailen[j] = dnz;
706:     bilen[j] = onz;
707:   }
708:   return(0);
709: }

711: /*
712:     This function sets the local j, a and ilen arrays (of the diagonal and off-diagonal part) of an MPIAIJ-matrix.
713:     The values in mat_i have to be sorted and the values in mat_j have to be sorted for each row (CSR-like).
714:     No off-processor parts off the matrix are allowed here, they are set at a later point by MatSetValues_MPIAIJ.
715:     Also, mat->was_assembled has to be false, otherwise the statement aj[rowstart_diag+dnz_row] = mat_j[col] - cstart;
716:     would not be true and the more complex MatSetValues_MPIAIJ has to be used.
717: */
718: PetscErrorCode MatSetValues_MPIAIJ_CopyFromCSRFormat(Mat mat,const PetscInt mat_j[],const PetscInt mat_i[],const PetscScalar mat_a[])
719: {
720:   Mat_MPIAIJ     *aij   = (Mat_MPIAIJ*)mat->data;
721:   Mat            A      = aij->A; /* diagonal part of the matrix */
722:   Mat            B      = aij->B; /* offdiagonal part of the matrix */
723:   Mat_SeqAIJ     *aijd  =(Mat_SeqAIJ*)(aij->A)->data,*aijo=(Mat_SeqAIJ*)(aij->B)->data;
724:   Mat_SeqAIJ     *a     = (Mat_SeqAIJ*)A->data;
725:   Mat_SeqAIJ     *b     = (Mat_SeqAIJ*)B->data;
726:   PetscInt       cstart = mat->cmap->rstart,cend = mat->cmap->rend;
727:   PetscInt       *ailen = a->ilen,*aj = a->j;
728:   PetscInt       *bilen = b->ilen,*bj = b->j;
729:   PetscInt       am     = aij->A->rmap->n,j;
730:   PetscInt       *full_diag_i=aijd->i,*full_offd_i=aijo->i; /* These variables can also include non-local elements, which are set at a later point. */
731:   PetscInt       col,dnz_row,onz_row,rowstart_diag,rowstart_offd;
732:   PetscScalar    *aa = a->a,*ba = b->a;

735:   /* Iterate over all rows of the matrix */
736:   for (j=0; j<am; j++) {
737:     dnz_row = onz_row = 0;
738:     rowstart_offd = full_offd_i[j];
739:     rowstart_diag = full_diag_i[j];
740:     /*  Iterate over all non-zero columns of the current row */
741:     for (col=mat_i[j]; col<mat_i[j+1]; col++) {
742:       /* If column is in the diagonal */
743:       if (mat_j[col] >= cstart && mat_j[col] < cend) {
744:         aj[rowstart_diag+dnz_row] = mat_j[col] - cstart;
745:         aa[rowstart_diag+dnz_row] = mat_a[col];
746:         dnz_row++;
747:       } else { /* off-diagonal entries */
748:         bj[rowstart_offd+onz_row] = mat_j[col];
749:         ba[rowstart_offd+onz_row] = mat_a[col];
750:         onz_row++;
751:       }
752:     }
753:     ailen[j] = dnz_row;
754:     bilen[j] = onz_row;
755:   }
756:   return(0);
757: }

759: PetscErrorCode MatGetValues_MPIAIJ(Mat mat,PetscInt m,const PetscInt idxm[],PetscInt n,const PetscInt idxn[],PetscScalar v[])
760: {
761:   Mat_MPIAIJ     *aij = (Mat_MPIAIJ*)mat->data;
763:   PetscInt       i,j,rstart = mat->rmap->rstart,rend = mat->rmap->rend;
764:   PetscInt       cstart = mat->cmap->rstart,cend = mat->cmap->rend,row,col;

767:   for (i=0; i<m; i++) {
768:     if (idxm[i] < 0) continue; /* SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative row: %D",idxm[i]);*/
769:     if (idxm[i] >= mat->rmap->N) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Row too large: row %D max %D",idxm[i],mat->rmap->N-1);
770:     if (idxm[i] >= rstart && idxm[i] < rend) {
771:       row = idxm[i] - rstart;
772:       for (j=0; j<n; j++) {
773:         if (idxn[j] < 0) continue; /* SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative column: %D",idxn[j]); */
774:         if (idxn[j] >= mat->cmap->N) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Column too large: col %D max %D",idxn[j],mat->cmap->N-1);
775:         if (idxn[j] >= cstart && idxn[j] < cend) {
776:           col  = idxn[j] - cstart;
777:           MatGetValues(aij->A,1,&row,1,&col,v+i*n+j);
778:         } else {
779:           if (!aij->colmap) {
780:             MatCreateColmap_MPIAIJ_Private(mat);
781:           }
782: #if defined(PETSC_USE_CTABLE)
783:           PetscTableFind(aij->colmap,idxn[j]+1,&col);
784:           col--;
785: #else
786:           col = aij->colmap[idxn[j]] - 1;
787: #endif
788:           if ((col < 0) || (aij->garray[col] != idxn[j])) *(v+i*n+j) = 0.0;
789:           else {
790:             MatGetValues(aij->B,1,&row,1,&col,v+i*n+j);
791:           }
792:         }
793:       }
794:     } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Only local values currently supported");
795:   }
796:   return(0);
797: }

799: extern PetscErrorCode MatMultDiagonalBlock_MPIAIJ(Mat,Vec,Vec);

801: PetscErrorCode MatAssemblyBegin_MPIAIJ(Mat mat,MatAssemblyType mode)
802: {
803:   Mat_MPIAIJ     *aij = (Mat_MPIAIJ*)mat->data;
805:   PetscInt       nstash,reallocs;

808:   if (aij->donotstash || mat->nooffprocentries) return(0);

810:   MatStashScatterBegin_Private(mat,&mat->stash,mat->rmap->range);
811:   MatStashGetInfo_Private(&mat->stash,&nstash,&reallocs);
812:   PetscInfo2(aij->A,"Stash has %D entries, uses %D mallocs.\n",nstash,reallocs);
813:   return(0);
814: }

816: PetscErrorCode MatAssemblyEnd_MPIAIJ(Mat mat,MatAssemblyType mode)
817: {
818:   Mat_MPIAIJ     *aij = (Mat_MPIAIJ*)mat->data;
819:   Mat_SeqAIJ     *a   = (Mat_SeqAIJ*)aij->A->data;
821:   PetscMPIInt    n;
822:   PetscInt       i,j,rstart,ncols,flg;
823:   PetscInt       *row,*col;
824:   PetscBool      other_disassembled;
825:   PetscScalar    *val;

827:   /* do not use 'b = (Mat_SeqAIJ*)aij->B->data' as B can be reset in disassembly */

830:   if (!aij->donotstash && !mat->nooffprocentries) {
831:     while (1) {
832:       MatStashScatterGetMesg_Private(&mat->stash,&n,&row,&col,&val,&flg);
833:       if (!flg) break;

835:       for (i=0; i<n; ) {
836:         /* Now identify the consecutive vals belonging to the same row */
837:         for (j=i,rstart=row[j]; j<n; j++) {
838:           if (row[j] != rstart) break;
839:         }
840:         if (j < n) ncols = j-i;
841:         else       ncols = n-i;
842:         /* Now assemble all these values with a single function call */
843:         MatSetValues_MPIAIJ(mat,1,row+i,ncols,col+i,val+i,mat->insertmode);
844:         i    = j;
845:       }
846:     }
847:     MatStashScatterEnd_Private(&mat->stash);
848:   }
849: #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA)
850:   if (mat->offloadmask == PETSC_OFFLOAD_CPU) aij->A->offloadmask = PETSC_OFFLOAD_CPU;
851:   /* We call MatBindToCPU() on aij->A and aij->B here, because if MatBindToCPU_MPIAIJ() is called before assembly, it cannot bind these. */
852:   if (mat->boundtocpu) {
853:     MatBindToCPU(aij->A,PETSC_TRUE);
854:     MatBindToCPU(aij->B,PETSC_TRUE);
855:   }
856: #endif
857:   MatAssemblyBegin(aij->A,mode);
858:   MatAssemblyEnd(aij->A,mode);

860:   /* determine if any processor has disassembled, if so we must
861:      also disassemble ourself, in order that we may reassemble. */
862:   /*
863:      if nonzero structure of submatrix B cannot change then we know that
864:      no processor disassembled thus we can skip this stuff
865:   */
866:   if (!((Mat_SeqAIJ*)aij->B->data)->nonew) {
867:     MPIU_Allreduce(&mat->was_assembled,&other_disassembled,1,MPIU_BOOL,MPI_PROD,PetscObjectComm((PetscObject)mat));
868:     if (mat->was_assembled && !other_disassembled) {
869: #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA)
870:       aij->B->offloadmask = PETSC_OFFLOAD_BOTH; /* do not copy on the GPU when assembling inside MatDisAssemble_MPIAIJ */
871: #endif
872:       MatDisAssemble_MPIAIJ(mat);
873:     }
874:   }
875:   if (!mat->was_assembled && mode == MAT_FINAL_ASSEMBLY) {
876:     MatSetUpMultiply_MPIAIJ(mat);
877:   }
878:   MatSetOption(aij->B,MAT_USE_INODES,PETSC_FALSE);
879: #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA)
880:   if (mat->offloadmask == PETSC_OFFLOAD_CPU && aij->B->offloadmask != PETSC_OFFLOAD_UNALLOCATED) aij->B->offloadmask = PETSC_OFFLOAD_CPU;
881: #endif
882:   MatAssemblyBegin(aij->B,mode);
883:   MatAssemblyEnd(aij->B,mode);

885:   PetscFree2(aij->rowvalues,aij->rowindices);

887:   aij->rowvalues = 0;

889:   VecDestroy(&aij->diag);
890:   if (a->inode.size) mat->ops->multdiagonalblock = MatMultDiagonalBlock_MPIAIJ;

892:   /* if no new nonzero locations are allowed in matrix then only set the matrix state the first time through */
893:   if ((!mat->was_assembled && mode == MAT_FINAL_ASSEMBLY) || !((Mat_SeqAIJ*)(aij->A->data))->nonew) {
894:     PetscObjectState state = aij->A->nonzerostate + aij->B->nonzerostate;
895:     MPIU_Allreduce(&state,&mat->nonzerostate,1,MPIU_INT64,MPI_SUM,PetscObjectComm((PetscObject)mat));
896:   }
897: #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA)
898:   mat->offloadmask = PETSC_OFFLOAD_BOTH;
899: #endif
900:   return(0);
901: }

903: PetscErrorCode MatZeroEntries_MPIAIJ(Mat A)
904: {
905:   Mat_MPIAIJ     *l = (Mat_MPIAIJ*)A->data;

909:   MatZeroEntries(l->A);
910:   MatZeroEntries(l->B);
911:   return(0);
912: }

914: PetscErrorCode MatZeroRows_MPIAIJ(Mat A,PetscInt N,const PetscInt rows[],PetscScalar diag,Vec x,Vec b)
915: {
916:   Mat_MPIAIJ      *mat = (Mat_MPIAIJ *) A->data;
917:   PetscObjectState sA, sB;
918:   PetscInt        *lrows;
919:   PetscInt         r, len;
920:   PetscBool        cong, lch, gch;
921:   PetscErrorCode   ierr;

924:   /* get locally owned rows */
925:   MatZeroRowsMapLocal_Private(A,N,rows,&len,&lrows);
926:   MatHasCongruentLayouts(A,&cong);
927:   /* fix right hand side if needed */
928:   if (x && b) {
929:     const PetscScalar *xx;
930:     PetscScalar       *bb;

932:     if (!cong) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"Need matching row/col layout");
933:     VecGetArrayRead(x, &xx);
934:     VecGetArray(b, &bb);
935:     for (r = 0; r < len; ++r) bb[lrows[r]] = diag*xx[lrows[r]];
936:     VecRestoreArrayRead(x, &xx);
937:     VecRestoreArray(b, &bb);
938:   }

940:   sA = mat->A->nonzerostate;
941:   sB = mat->B->nonzerostate;

943:   if (diag != 0.0 && cong) {
944:     MatZeroRows(mat->A, len, lrows, diag, NULL, NULL);
945:     MatZeroRows(mat->B, len, lrows, 0.0, NULL, NULL);
946:   } else if (diag != 0.0) { /* non-square or non congruent layouts -> if keepnonzeropattern is false, we allow for new insertion */
947:     Mat_SeqAIJ *aijA = (Mat_SeqAIJ*)mat->A->data;
948:     Mat_SeqAIJ *aijB = (Mat_SeqAIJ*)mat->B->data;
949:     PetscInt   nnwA, nnwB;
950:     PetscBool  nnzA, nnzB;

952:     nnwA = aijA->nonew;
953:     nnwB = aijB->nonew;
954:     nnzA = aijA->keepnonzeropattern;
955:     nnzB = aijB->keepnonzeropattern;
956:     if (!nnzA) {
957:       PetscInfo(mat->A,"Requested to not keep the pattern and add a nonzero diagonal; may encounter reallocations on diagonal block.\n");
958:       aijA->nonew = 0;
959:     }
960:     if (!nnzB) {
961:       PetscInfo(mat->B,"Requested to not keep the pattern and add a nonzero diagonal; may encounter reallocations on off-diagonal block.\n");
962:       aijB->nonew = 0;
963:     }
964:     /* Must zero here before the next loop */
965:     MatZeroRows(mat->A, len, lrows, 0.0, NULL, NULL);
966:     MatZeroRows(mat->B, len, lrows, 0.0, NULL, NULL);
967:     for (r = 0; r < len; ++r) {
968:       const PetscInt row = lrows[r] + A->rmap->rstart;
969:       if (row >= A->cmap->N) continue;
970:       MatSetValues(A, 1, &row, 1, &row, &diag, INSERT_VALUES);
971:     }
972:     aijA->nonew = nnwA;
973:     aijB->nonew = nnwB;
974:   } else {
975:     MatZeroRows(mat->A, len, lrows, 0.0, NULL, NULL);
976:     MatZeroRows(mat->B, len, lrows, 0.0, NULL, NULL);
977:   }
978:   PetscFree(lrows);
979:   MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);
980:   MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);

982:   /* reduce nonzerostate */
983:   lch = (PetscBool)(sA != mat->A->nonzerostate || sB != mat->B->nonzerostate);
984:   MPIU_Allreduce(&lch,&gch,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)A));
985:   if (gch) A->nonzerostate++;
986:   return(0);
987: }

989: PetscErrorCode MatZeroRowsColumns_MPIAIJ(Mat A,PetscInt N,const PetscInt rows[],PetscScalar diag,Vec x,Vec b)
990: {
991:   Mat_MPIAIJ        *l = (Mat_MPIAIJ*)A->data;
992:   PetscErrorCode    ierr;
993:   PetscMPIInt       n = A->rmap->n;
994:   PetscInt          i,j,r,m,len = 0;
995:   PetscInt          *lrows,*owners = A->rmap->range;
996:   PetscMPIInt       p = 0;
997:   PetscSFNode       *rrows;
998:   PetscSF           sf;
999:   const PetscScalar *xx;
1000:   PetscScalar       *bb,*mask;
1001:   Vec               xmask,lmask;
1002:   Mat_SeqAIJ        *aij = (Mat_SeqAIJ*)l->B->data;
1003:   const PetscInt    *aj, *ii,*ridx;
1004:   PetscScalar       *aa;

1007:   /* Create SF where leaves are input rows and roots are owned rows */
1008:   PetscMalloc1(n, &lrows);
1009:   for (r = 0; r < n; ++r) lrows[r] = -1;
1010:   PetscMalloc1(N, &rrows);
1011:   for (r = 0; r < N; ++r) {
1012:     const PetscInt idx   = rows[r];
1013:     if (idx < 0 || A->rmap->N <= idx) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Row %D out of range [0,%D)",idx,A->rmap->N);
1014:     if (idx < owners[p] || owners[p+1] <= idx) { /* short-circuit the search if the last p owns this row too */
1015:       PetscLayoutFindOwner(A->rmap,idx,&p);
1016:     }
1017:     rrows[r].rank  = p;
1018:     rrows[r].index = rows[r] - owners[p];
1019:   }
1020:   PetscSFCreate(PetscObjectComm((PetscObject) A), &sf);
1021:   PetscSFSetGraph(sf, n, N, NULL, PETSC_OWN_POINTER, rrows, PETSC_OWN_POINTER);
1022:   /* Collect flags for rows to be zeroed */
1023:   PetscSFReduceBegin(sf, MPIU_INT, (PetscInt *) rows, lrows, MPI_LOR);
1024:   PetscSFReduceEnd(sf, MPIU_INT, (PetscInt *) rows, lrows, MPI_LOR);
1025:   PetscSFDestroy(&sf);
1026:   /* Compress and put in row numbers */
1027:   for (r = 0; r < n; ++r) if (lrows[r] >= 0) lrows[len++] = r;
1028:   /* zero diagonal part of matrix */
1029:   MatZeroRowsColumns(l->A,len,lrows,diag,x,b);
1030:   /* handle off diagonal part of matrix */
1031:   MatCreateVecs(A,&xmask,NULL);
1032:   VecDuplicate(l->lvec,&lmask);
1033:   VecGetArray(xmask,&bb);
1034:   for (i=0; i<len; i++) bb[lrows[i]] = 1;
1035:   VecRestoreArray(xmask,&bb);
1036:   VecScatterBegin(l->Mvctx,xmask,lmask,ADD_VALUES,SCATTER_FORWARD);
1037:   VecScatterEnd(l->Mvctx,xmask,lmask,ADD_VALUES,SCATTER_FORWARD);
1038:   VecDestroy(&xmask);
1039:   if (x && b) { /* this code is buggy when the row and column layout don't match */
1040:     PetscBool cong;

1042:     MatHasCongruentLayouts(A,&cong);
1043:     if (!cong) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"Need matching row/col layout");
1044:     VecScatterBegin(l->Mvctx,x,l->lvec,INSERT_VALUES,SCATTER_FORWARD);
1045:     VecScatterEnd(l->Mvctx,x,l->lvec,INSERT_VALUES,SCATTER_FORWARD);
1046:     VecGetArrayRead(l->lvec,&xx);
1047:     VecGetArray(b,&bb);
1048:   }
1049:   VecGetArray(lmask,&mask);
1050:   /* remove zeroed rows of off diagonal matrix */
1051:   ii = aij->i;
1052:   for (i=0; i<len; i++) {
1053:     PetscArrayzero(aij->a + ii[lrows[i]],ii[lrows[i]+1] - ii[lrows[i]]);
1054:   }
1055:   /* loop over all elements of off process part of matrix zeroing removed columns*/
1056:   if (aij->compressedrow.use) {
1057:     m    = aij->compressedrow.nrows;
1058:     ii   = aij->compressedrow.i;
1059:     ridx = aij->compressedrow.rindex;
1060:     for (i=0; i<m; i++) {
1061:       n  = ii[i+1] - ii[i];
1062:       aj = aij->j + ii[i];
1063:       aa = aij->a + ii[i];

1065:       for (j=0; j<n; j++) {
1066:         if (PetscAbsScalar(mask[*aj])) {
1067:           if (b) bb[*ridx] -= *aa*xx[*aj];
1068:           *aa = 0.0;
1069:         }
1070:         aa++;
1071:         aj++;
1072:       }
1073:       ridx++;
1074:     }
1075:   } else { /* do not use compressed row format */
1076:     m = l->B->rmap->n;
1077:     for (i=0; i<m; i++) {
1078:       n  = ii[i+1] - ii[i];
1079:       aj = aij->j + ii[i];
1080:       aa = aij->a + ii[i];
1081:       for (j=0; j<n; j++) {
1082:         if (PetscAbsScalar(mask[*aj])) {
1083:           if (b) bb[i] -= *aa*xx[*aj];
1084:           *aa = 0.0;
1085:         }
1086:         aa++;
1087:         aj++;
1088:       }
1089:     }
1090:   }
1091:   if (x && b) {
1092:     VecRestoreArray(b,&bb);
1093:     VecRestoreArrayRead(l->lvec,&xx);
1094:   }
1095:   VecRestoreArray(lmask,&mask);
1096:   VecDestroy(&lmask);
1097:   PetscFree(lrows);

1099:   /* only change matrix nonzero state if pattern was allowed to be changed */
1100:   if (!((Mat_SeqAIJ*)(l->A->data))->keepnonzeropattern) {
1101:     PetscObjectState state = l->A->nonzerostate + l->B->nonzerostate;
1102:     MPIU_Allreduce(&state,&A->nonzerostate,1,MPIU_INT64,MPI_SUM,PetscObjectComm((PetscObject)A));
1103:   }
1104:   return(0);
1105: }

1107: PetscErrorCode MatMult_MPIAIJ(Mat A,Vec xx,Vec yy)
1108: {
1109:   Mat_MPIAIJ     *a = (Mat_MPIAIJ*)A->data;
1111:   PetscInt       nt;
1112:   VecScatter     Mvctx = a->Mvctx;

1115:   VecGetLocalSize(xx,&nt);
1116:   if (nt != A->cmap->n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Incompatible partition of A (%D) and xx (%D)",A->cmap->n,nt);
1117:   VecScatterBegin(Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);
1118:   (*a->A->ops->mult)(a->A,xx,yy);
1119:   VecScatterEnd(Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);
1120:   (*a->B->ops->multadd)(a->B,a->lvec,yy,yy);
1121:   return(0);
1122: }

1124: PetscErrorCode MatMultDiagonalBlock_MPIAIJ(Mat A,Vec bb,Vec xx)
1125: {
1126:   Mat_MPIAIJ     *a = (Mat_MPIAIJ*)A->data;

1130:   MatMultDiagonalBlock(a->A,bb,xx);
1131:   return(0);
1132: }

1134: PetscErrorCode MatMultAdd_MPIAIJ(Mat A,Vec xx,Vec yy,Vec zz)
1135: {
1136:   Mat_MPIAIJ     *a = (Mat_MPIAIJ*)A->data;
1138:   VecScatter     Mvctx = a->Mvctx;

1141:   if (a->Mvctx_mpi1_flg) Mvctx = a->Mvctx_mpi1;
1142:   VecScatterBegin(Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);
1143:   (*a->A->ops->multadd)(a->A,xx,yy,zz);
1144:   VecScatterEnd(Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);
1145:   (*a->B->ops->multadd)(a->B,a->lvec,zz,zz);
1146:   return(0);
1147: }

1149: PetscErrorCode MatMultTranspose_MPIAIJ(Mat A,Vec xx,Vec yy)
1150: {
1151:   Mat_MPIAIJ     *a = (Mat_MPIAIJ*)A->data;

1155:   /* do nondiagonal part */
1156:   (*a->B->ops->multtranspose)(a->B,xx,a->lvec);
1157:   /* do local part */
1158:   (*a->A->ops->multtranspose)(a->A,xx,yy);
1159:   /* add partial results together */
1160:   VecScatterBegin(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);
1161:   VecScatterEnd(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);
1162:   return(0);
1163: }

1165: PetscErrorCode MatIsTranspose_MPIAIJ(Mat Amat,Mat Bmat,PetscReal tol,PetscBool  *f)
1166: {
1167:   MPI_Comm       comm;
1168:   Mat_MPIAIJ     *Aij = (Mat_MPIAIJ*) Amat->data, *Bij;
1169:   Mat            Adia = Aij->A, Bdia, Aoff,Boff,*Aoffs,*Boffs;
1170:   IS             Me,Notme;
1172:   PetscInt       M,N,first,last,*notme,i;
1173:   PetscBool      lf;
1174:   PetscMPIInt    size;

1177:   /* Easy test: symmetric diagonal block */
1178:   Bij  = (Mat_MPIAIJ*) Bmat->data; Bdia = Bij->A;
1179:   MatIsTranspose(Adia,Bdia,tol,&lf);
1180:   MPIU_Allreduce(&lf,f,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)Amat));
1181:   if (!*f) return(0);
1182:   PetscObjectGetComm((PetscObject)Amat,&comm);
1183:   MPI_Comm_size(comm,&size);
1184:   if (size == 1) return(0);

1186:   /* Hard test: off-diagonal block. This takes a MatCreateSubMatrix. */
1187:   MatGetSize(Amat,&M,&N);
1188:   MatGetOwnershipRange(Amat,&first,&last);
1189:   PetscMalloc1(N-last+first,&notme);
1190:   for (i=0; i<first; i++) notme[i] = i;
1191:   for (i=last; i<M; i++) notme[i-last+first] = i;
1192:   ISCreateGeneral(MPI_COMM_SELF,N-last+first,notme,PETSC_COPY_VALUES,&Notme);
1193:   ISCreateStride(MPI_COMM_SELF,last-first,first,1,&Me);
1194:   MatCreateSubMatrices(Amat,1,&Me,&Notme,MAT_INITIAL_MATRIX,&Aoffs);
1195:   Aoff = Aoffs[0];
1196:   MatCreateSubMatrices(Bmat,1,&Notme,&Me,MAT_INITIAL_MATRIX,&Boffs);
1197:   Boff = Boffs[0];
1198:   MatIsTranspose(Aoff,Boff,tol,f);
1199:   MatDestroyMatrices(1,&Aoffs);
1200:   MatDestroyMatrices(1,&Boffs);
1201:   ISDestroy(&Me);
1202:   ISDestroy(&Notme);
1203:   PetscFree(notme);
1204:   return(0);
1205: }

1207: PetscErrorCode MatIsSymmetric_MPIAIJ(Mat A,PetscReal tol,PetscBool  *f)
1208: {

1212:   MatIsTranspose_MPIAIJ(A,A,tol,f);
1213:   return(0);
1214: }

1216: PetscErrorCode MatMultTransposeAdd_MPIAIJ(Mat A,Vec xx,Vec yy,Vec zz)
1217: {
1218:   Mat_MPIAIJ     *a = (Mat_MPIAIJ*)A->data;

1222:   /* do nondiagonal part */
1223:   (*a->B->ops->multtranspose)(a->B,xx,a->lvec);
1224:   /* do local part */
1225:   (*a->A->ops->multtransposeadd)(a->A,xx,yy,zz);
1226:   /* add partial results together */
1227:   VecScatterBegin(a->Mvctx,a->lvec,zz,ADD_VALUES,SCATTER_REVERSE);
1228:   VecScatterEnd(a->Mvctx,a->lvec,zz,ADD_VALUES,SCATTER_REVERSE);
1229:   return(0);
1230: }

1232: /*
1233:   This only works correctly for square matrices where the subblock A->A is the
1234:    diagonal block
1235: */
1236: PetscErrorCode MatGetDiagonal_MPIAIJ(Mat A,Vec v)
1237: {
1239:   Mat_MPIAIJ     *a = (Mat_MPIAIJ*)A->data;

1242:   if (A->rmap->N != A->cmap->N) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"Supports only square matrix where A->A is diag block");
1243:   if (A->rmap->rstart != A->cmap->rstart || A->rmap->rend != A->cmap->rend) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"row partition must equal col partition");
1244:   MatGetDiagonal(a->A,v);
1245:   return(0);
1246: }

1248: PetscErrorCode MatScale_MPIAIJ(Mat A,PetscScalar aa)
1249: {
1250:   Mat_MPIAIJ     *a = (Mat_MPIAIJ*)A->data;

1254:   MatScale(a->A,aa);
1255:   MatScale(a->B,aa);
1256:   return(0);
1257: }

1259: PetscErrorCode MatDestroy_MPIAIJ(Mat mat)
1260: {
1261:   Mat_MPIAIJ     *aij = (Mat_MPIAIJ*)mat->data;

1265: #if defined(PETSC_USE_LOG)
1266:   PetscLogObjectState((PetscObject)mat,"Rows=%D, Cols=%D",mat->rmap->N,mat->cmap->N);
1267: #endif
1268:   MatStashDestroy_Private(&mat->stash);
1269:   VecDestroy(&aij->diag);
1270:   MatDestroy(&aij->A);
1271:   MatDestroy(&aij->B);
1272: #if defined(PETSC_USE_CTABLE)
1273:   PetscTableDestroy(&aij->colmap);
1274: #else
1275:   PetscFree(aij->colmap);
1276: #endif
1277:   PetscFree(aij->garray);
1278:   VecDestroy(&aij->lvec);
1279:   VecScatterDestroy(&aij->Mvctx);
1280:   if (aij->Mvctx_mpi1) {VecScatterDestroy(&aij->Mvctx_mpi1);}
1281:   PetscFree2(aij->rowvalues,aij->rowindices);
1282:   PetscFree(aij->ld);
1283:   PetscFree(mat->data);

1285:   PetscObjectChangeTypeName((PetscObject)mat,0);
1286:   PetscObjectComposeFunction((PetscObject)mat,"MatStoreValues_C",NULL);
1287:   PetscObjectComposeFunction((PetscObject)mat,"MatRetrieveValues_C",NULL);
1288:   PetscObjectComposeFunction((PetscObject)mat,"MatIsTranspose_C",NULL);
1289:   PetscObjectComposeFunction((PetscObject)mat,"MatMPIAIJSetPreallocation_C",NULL);
1290:   PetscObjectComposeFunction((PetscObject)mat,"MatResetPreallocation_C",NULL);
1291:   PetscObjectComposeFunction((PetscObject)mat,"MatMPIAIJSetPreallocationCSR_C",NULL);
1292:   PetscObjectComposeFunction((PetscObject)mat,"MatDiagonalScaleLocal_C",NULL);
1293:   PetscObjectComposeFunction((PetscObject)mat,"MatConvert_mpiaij_mpibaij_C",NULL);
1294:   PetscObjectComposeFunction((PetscObject)mat,"MatConvert_mpiaij_mpisbaij_C",NULL);
1295: #if defined(PETSC_HAVE_ELEMENTAL)
1296:   PetscObjectComposeFunction((PetscObject)mat,"MatConvert_mpiaij_elemental_C",NULL);
1297: #endif
1298: #if defined(PETSC_HAVE_HYPRE)
1299:   PetscObjectComposeFunction((PetscObject)mat,"MatConvert_mpiaij_hypre_C",NULL);
1300:   PetscObjectComposeFunction((PetscObject)mat,"MatProductSetFromOptions_transpose_mpiaij_mpiaij_C",NULL);
1301: #endif
1302:   PetscObjectComposeFunction((PetscObject)mat,"MatConvert_mpiaij_is_C",NULL);
1303:   PetscObjectComposeFunction((PetscObject)mat,"MatProductSetFromOptions_is_mpiaij_C",NULL);
1304:   PetscObjectComposeFunction((PetscObject)mat,"MatProductSetFromOptions_mpiaij_mpiaij_C",NULL);
1305:   return(0);
1306: }

1308: PetscErrorCode MatView_MPIAIJ_Binary(Mat mat,PetscViewer viewer)
1309: {
1310:   Mat_MPIAIJ        *aij = (Mat_MPIAIJ*)mat->data;
1311:   Mat_SeqAIJ        *A   = (Mat_SeqAIJ*)aij->A->data;
1312:   Mat_SeqAIJ        *B   = (Mat_SeqAIJ*)aij->B->data;
1313:   const PetscInt    *garray = aij->garray;
1314:   PetscInt          header[4],M,N,m,rs,cs,nz,cnt,i,ja,jb;
1315:   PetscInt          *rowlens;
1316:   PetscInt          *colidxs;
1317:   PetscScalar       *matvals;
1318:   PetscErrorCode    ierr;

1321:   PetscViewerSetUp(viewer);

1323:   M  = mat->rmap->N;
1324:   N  = mat->cmap->N;
1325:   m  = mat->rmap->n;
1326:   rs = mat->rmap->rstart;
1327:   cs = mat->cmap->rstart;
1328:   nz = A->nz + B->nz;

1330:   /* write matrix header */
1331:   header[0] = MAT_FILE_CLASSID;
1332:   header[1] = M; header[2] = N; header[3] = nz;
1333:   MPI_Reduce(&nz,&header[3],1,MPIU_INT,MPI_SUM,0,PetscObjectComm((PetscObject)mat));
1334:   PetscViewerBinaryWrite(viewer,header,4,PETSC_INT);

1336:   /* fill in and store row lengths  */
1337:   PetscMalloc1(m,&rowlens);
1338:   for (i=0; i<m; i++) rowlens[i] = A->i[i+1] - A->i[i] + B->i[i+1] - B->i[i];
1339:   PetscViewerBinaryWriteAll(viewer,rowlens,m,rs,M,PETSC_INT);
1340:   PetscFree(rowlens);

1342:   /* fill in and store column indices */
1343:   PetscMalloc1(nz,&colidxs);
1344:   for (cnt=0, i=0; i<m; i++) {
1345:     for (jb=B->i[i]; jb<B->i[i+1]; jb++) {
1346:       if (garray[B->j[jb]] > cs) break;
1347:       colidxs[cnt++] = garray[B->j[jb]];
1348:     }
1349:     for (ja=A->i[i]; ja<A->i[i+1]; ja++)
1350:       colidxs[cnt++] = A->j[ja] + cs;
1351:     for (; jb<B->i[i+1]; jb++)
1352:       colidxs[cnt++] = garray[B->j[jb]];
1353:   }
1354:   if (cnt != nz) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_LIB,"Internal PETSc error: cnt = %D nz = %D",cnt,nz);
1355:   PetscViewerBinaryWriteAll(viewer,colidxs,nz,PETSC_DETERMINE,PETSC_DETERMINE,PETSC_INT);
1356:   PetscFree(colidxs);

1358:   /* fill in and store nonzero values */
1359:   PetscMalloc1(nz,&matvals);
1360:   for (cnt=0, i=0; i<m; i++) {
1361:     for (jb=B->i[i]; jb<B->i[i+1]; jb++) {
1362:       if (garray[B->j[jb]] > cs) break;
1363:       matvals[cnt++] = B->a[jb];
1364:     }
1365:     for (ja=A->i[i]; ja<A->i[i+1]; ja++)
1366:       matvals[cnt++] = A->a[ja];
1367:     for (; jb<B->i[i+1]; jb++)
1368:       matvals[cnt++] = B->a[jb];
1369:   }
1370:   if (cnt != nz) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_LIB,"Internal PETSc error: cnt = %D nz = %D",cnt,nz);
1371:   PetscViewerBinaryWriteAll(viewer,matvals,nz,PETSC_DETERMINE,PETSC_DETERMINE,PETSC_SCALAR);
1372:   PetscFree(matvals);

1374:   /* write block size option to the viewer's .info file */
1375:   MatView_Binary_BlockSizes(mat,viewer);
1376:   return(0);
1377: }

1379:  #include <petscdraw.h>
1380: PetscErrorCode MatView_MPIAIJ_ASCIIorDraworSocket(Mat mat,PetscViewer viewer)
1381: {
1382:   Mat_MPIAIJ        *aij = (Mat_MPIAIJ*)mat->data;
1383:   PetscErrorCode    ierr;
1384:   PetscMPIInt       rank = aij->rank,size = aij->size;
1385:   PetscBool         isdraw,iascii,isbinary;
1386:   PetscViewer       sviewer;
1387:   PetscViewerFormat format;

1390:   PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERDRAW,&isdraw);
1391:   PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii);
1392:   PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERBINARY,&isbinary);
1393:   if (iascii) {
1394:     PetscViewerGetFormat(viewer,&format);
1395:     if (format == PETSC_VIEWER_LOAD_BALANCE) {
1396:       PetscInt i,nmax = 0,nmin = PETSC_MAX_INT,navg = 0,*nz,nzlocal = ((Mat_SeqAIJ*) (aij->A->data))->nz + ((Mat_SeqAIJ*) (aij->B->data))->nz;
1397:       PetscMalloc1(size,&nz);
1398:       MPI_Allgather(&nzlocal,1,MPIU_INT,nz,1,MPIU_INT,PetscObjectComm((PetscObject)mat));
1399:       for (i=0; i<(PetscInt)size; i++) {
1400:         nmax = PetscMax(nmax,nz[i]);
1401:         nmin = PetscMin(nmin,nz[i]);
1402:         navg += nz[i];
1403:       }
1404:       PetscFree(nz);
1405:       navg = navg/size;
1406:       PetscViewerASCIIPrintf(viewer,"Load Balance - Nonzeros: Min %D  avg %D  max %D\n",nmin,navg,nmax);
1407:       return(0);
1408:     }
1409:     PetscViewerGetFormat(viewer,&format);
1410:     if (format == PETSC_VIEWER_ASCII_INFO_DETAIL) {
1411:       MatInfo   info;
1412:       PetscBool inodes;

1414:       MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);
1415:       MatGetInfo(mat,MAT_LOCAL,&info);
1416:       MatInodeGetInodeSizes(aij->A,NULL,(PetscInt**)&inodes,NULL);
1417:       PetscViewerASCIIPushSynchronized(viewer);
1418:       if (!inodes) {
1419:         PetscViewerASCIISynchronizedPrintf(viewer,"[%d] Local rows %D nz %D nz alloced %D mem %g, not using I-node routines\n",
1420:                                                   rank,mat->rmap->n,(PetscInt)info.nz_used,(PetscInt)info.nz_allocated,(double)info.memory);
1421:       } else {
1422:         PetscViewerASCIISynchronizedPrintf(viewer,"[%d] Local rows %D nz %D nz alloced %D mem %g, using I-node routines\n",
1423:                                                   rank,mat->rmap->n,(PetscInt)info.nz_used,(PetscInt)info.nz_allocated,(double)info.memory);
1424:       }
1425:       MatGetInfo(aij->A,MAT_LOCAL,&info);
1426:       PetscViewerASCIISynchronizedPrintf(viewer,"[%d] on-diagonal part: nz %D \n",rank,(PetscInt)info.nz_used);
1427:       MatGetInfo(aij->B,MAT_LOCAL,&info);
1428:       PetscViewerASCIISynchronizedPrintf(viewer,"[%d] off-diagonal part: nz %D \n",rank,(PetscInt)info.nz_used);
1429:       PetscViewerFlush(viewer);
1430:       PetscViewerASCIIPopSynchronized(viewer);
1431:       PetscViewerASCIIPrintf(viewer,"Information on VecScatter used in matrix-vector product: \n");
1432:       VecScatterView(aij->Mvctx,viewer);
1433:       return(0);
1434:     } else if (format == PETSC_VIEWER_ASCII_INFO) {
1435:       PetscInt inodecount,inodelimit,*inodes;
1436:       MatInodeGetInodeSizes(aij->A,&inodecount,&inodes,&inodelimit);
1437:       if (inodes) {
1438:         PetscViewerASCIIPrintf(viewer,"using I-node (on process 0) routines: found %D nodes, limit used is %D\n",inodecount,inodelimit);
1439:       } else {
1440:         PetscViewerASCIIPrintf(viewer,"not using I-node (on process 0) routines\n");
1441:       }
1442:       return(0);
1443:     } else if (format == PETSC_VIEWER_ASCII_FACTOR_INFO) {
1444:       return(0);
1445:     }
1446:   } else if (isbinary) {
1447:     if (size == 1) {
1448:       PetscObjectSetName((PetscObject)aij->A,((PetscObject)mat)->name);
1449:       MatView(aij->A,viewer);
1450:     } else {
1451:       MatView_MPIAIJ_Binary(mat,viewer);
1452:     }
1453:     return(0);
1454:   } else if (iascii && size == 1) {
1455:     PetscObjectSetName((PetscObject)aij->A,((PetscObject)mat)->name);
1456:     MatView(aij->A,viewer);
1457:     return(0);
1458:   } else if (isdraw) {
1459:     PetscDraw draw;
1460:     PetscBool isnull;
1461:     PetscViewerDrawGetDraw(viewer,0,&draw);
1462:     PetscDrawIsNull(draw,&isnull);
1463:     if (isnull) return(0);
1464:   }

1466:   { /* assemble the entire matrix onto first processor */
1467:     Mat A = NULL, Av;
1468:     IS  isrow,iscol;

1470:     ISCreateStride(PetscObjectComm((PetscObject)mat),!rank ? mat->rmap->N : 0,0,1,&isrow);
1471:     ISCreateStride(PetscObjectComm((PetscObject)mat),!rank ? mat->cmap->N : 0,0,1,&iscol);
1472:     MatCreateSubMatrix(mat,isrow,iscol,MAT_INITIAL_MATRIX,&A);
1473:     MatMPIAIJGetSeqAIJ(A,&Av,NULL,NULL);
1474: /*  The commented code uses MatCreateSubMatrices instead */
1475: /*
1476:     Mat *AA, A = NULL, Av;
1477:     IS  isrow,iscol;

1479:     ISCreateStride(PetscObjectComm((PetscObject)mat),!rank ? mat->rmap->N : 0,0,1,&isrow);
1480:     ISCreateStride(PetscObjectComm((PetscObject)mat),!rank ? mat->cmap->N : 0,0,1,&iscol);
1481:     MatCreateSubMatrices(mat,1,&isrow,&iscol,MAT_INITIAL_MATRIX,&AA);
1482:     if (!rank) {
1483:        PetscObjectReference((PetscObject)AA[0]);
1484:        A    = AA[0];
1485:        Av   = AA[0];
1486:     }
1487:     MatDestroySubMatrices(1,&AA);
1488: */
1489:     ISDestroy(&iscol);
1490:     ISDestroy(&isrow);
1491:     /*
1492:        Everyone has to call to draw the matrix since the graphics waits are
1493:        synchronized across all processors that share the PetscDraw object
1494:     */
1495:     PetscViewerGetSubViewer(viewer,PETSC_COMM_SELF,&sviewer);
1496:     if (!rank) {
1497:       if (((PetscObject)mat)->name) {
1498:         PetscObjectSetName((PetscObject)Av,((PetscObject)mat)->name);
1499:       }
1500:       MatView_SeqAIJ(Av,sviewer);
1501:     }
1502:     PetscViewerRestoreSubViewer(viewer,PETSC_COMM_SELF,&sviewer);
1503:     PetscViewerFlush(viewer);
1504:     MatDestroy(&A);
1505:   }
1506:   return(0);
1507: }

1509: PetscErrorCode MatView_MPIAIJ(Mat mat,PetscViewer viewer)
1510: {
1512:   PetscBool      iascii,isdraw,issocket,isbinary;

1515:   PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii);
1516:   PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERDRAW,&isdraw);
1517:   PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERBINARY,&isbinary);
1518:   PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERSOCKET,&issocket);
1519:   if (iascii || isdraw || isbinary || issocket) {
1520:     MatView_MPIAIJ_ASCIIorDraworSocket(mat,viewer);
1521:   }
1522:   return(0);
1523: }

1525: PetscErrorCode MatSOR_MPIAIJ(Mat matin,Vec bb,PetscReal omega,MatSORType flag,PetscReal fshift,PetscInt its,PetscInt lits,Vec xx)
1526: {
1527:   Mat_MPIAIJ     *mat = (Mat_MPIAIJ*)matin->data;
1529:   Vec            bb1 = 0;
1530:   PetscBool      hasop;

1533:   if (flag == SOR_APPLY_UPPER) {
1534:     (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);
1535:     return(0);
1536:   }

1538:   if (its > 1 || ~flag & SOR_ZERO_INITIAL_GUESS || flag & SOR_EISENSTAT) {
1539:     VecDuplicate(bb,&bb1);
1540:   }

1542:   if ((flag & SOR_LOCAL_SYMMETRIC_SWEEP) == SOR_LOCAL_SYMMETRIC_SWEEP) {
1543:     if (flag & SOR_ZERO_INITIAL_GUESS) {
1544:       (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);
1545:       its--;
1546:     }

1548:     while (its--) {
1549:       VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);
1550:       VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);

1552:       /* update rhs: bb1 = bb - B*x */
1553:       VecScale(mat->lvec,-1.0);
1554:       (*mat->B->ops->multadd)(mat->B,mat->lvec,bb,bb1);

1556:       /* local sweep */
1557:       (*mat->A->ops->sor)(mat->A,bb1,omega,SOR_SYMMETRIC_SWEEP,fshift,lits,1,xx);
1558:     }
1559:   } else if (flag & SOR_LOCAL_FORWARD_SWEEP) {
1560:     if (flag & SOR_ZERO_INITIAL_GUESS) {
1561:       (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);
1562:       its--;
1563:     }
1564:     while (its--) {
1565:       VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);
1566:       VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);

1568:       /* update rhs: bb1 = bb - B*x */
1569:       VecScale(mat->lvec,-1.0);
1570:       (*mat->B->ops->multadd)(mat->B,mat->lvec,bb,bb1);

1572:       /* local sweep */
1573:       (*mat->A->ops->sor)(mat->A,bb1,omega,SOR_FORWARD_SWEEP,fshift,lits,1,xx);
1574:     }
1575:   } else if (flag & SOR_LOCAL_BACKWARD_SWEEP) {
1576:     if (flag & SOR_ZERO_INITIAL_GUESS) {
1577:       (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);
1578:       its--;
1579:     }
1580:     while (its--) {
1581:       VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);
1582:       VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);

1584:       /* update rhs: bb1 = bb - B*x */
1585:       VecScale(mat->lvec,-1.0);
1586:       (*mat->B->ops->multadd)(mat->B,mat->lvec,bb,bb1);

1588:       /* local sweep */
1589:       (*mat->A->ops->sor)(mat->A,bb1,omega,SOR_BACKWARD_SWEEP,fshift,lits,1,xx);
1590:     }
1591:   } else if (flag & SOR_EISENSTAT) {
1592:     Vec xx1;

1594:     VecDuplicate(bb,&xx1);
1595:     (*mat->A->ops->sor)(mat->A,bb,omega,(MatSORType)(SOR_ZERO_INITIAL_GUESS | SOR_LOCAL_BACKWARD_SWEEP),fshift,lits,1,xx);

1597:     VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);
1598:     VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);
1599:     if (!mat->diag) {
1600:       MatCreateVecs(matin,&mat->diag,NULL);
1601:       MatGetDiagonal(matin,mat->diag);
1602:     }
1603:     MatHasOperation(matin,MATOP_MULT_DIAGONAL_BLOCK,&hasop);
1604:     if (hasop) {
1605:       MatMultDiagonalBlock(matin,xx,bb1);
1606:     } else {
1607:       VecPointwiseMult(bb1,mat->diag,xx);
1608:     }
1609:     VecAYPX(bb1,(omega-2.0)/omega,bb);

1611:     MatMultAdd(mat->B,mat->lvec,bb1,bb1);

1613:     /* local sweep */
1614:     (*mat->A->ops->sor)(mat->A,bb1,omega,(MatSORType)(SOR_ZERO_INITIAL_GUESS | SOR_LOCAL_FORWARD_SWEEP),fshift,lits,1,xx1);
1615:     VecAXPY(xx,1.0,xx1);
1616:     VecDestroy(&xx1);
1617:   } else SETERRQ(PetscObjectComm((PetscObject)matin),PETSC_ERR_SUP,"Parallel SOR not supported");

1619:   VecDestroy(&bb1);

1621:   matin->factorerrortype = mat->A->factorerrortype;
1622:   return(0);
1623: }

1625: PetscErrorCode MatPermute_MPIAIJ(Mat A,IS rowp,IS colp,Mat *B)
1626: {
1627:   Mat            aA,aB,Aperm;
1628:   const PetscInt *rwant,*cwant,*gcols,*ai,*bi,*aj,*bj;
1629:   PetscScalar    *aa,*ba;
1630:   PetscInt       i,j,m,n,ng,anz,bnz,*dnnz,*onnz,*tdnnz,*tonnz,*rdest,*cdest,*work,*gcdest;
1631:   PetscSF        rowsf,sf;
1632:   IS             parcolp = NULL;
1633:   PetscBool      done;

1637:   MatGetLocalSize(A,&m,&n);
1638:   ISGetIndices(rowp,&rwant);
1639:   ISGetIndices(colp,&cwant);
1640:   PetscMalloc3(PetscMax(m,n),&work,m,&rdest,n,&cdest);

1642:   /* Invert row permutation to find out where my rows should go */
1643:   PetscSFCreate(PetscObjectComm((PetscObject)A),&rowsf);
1644:   PetscSFSetGraphLayout(rowsf,A->rmap,A->rmap->n,NULL,PETSC_OWN_POINTER,rwant);
1645:   PetscSFSetFromOptions(rowsf);
1646:   for (i=0; i<m; i++) work[i] = A->rmap->rstart + i;
1647:   PetscSFReduceBegin(rowsf,MPIU_INT,work,rdest,MPIU_REPLACE);
1648:   PetscSFReduceEnd(rowsf,MPIU_INT,work,rdest,MPIU_REPLACE);

1650:   /* Invert column permutation to find out where my columns should go */
1651:   PetscSFCreate(PetscObjectComm((PetscObject)A),&sf);
1652:   PetscSFSetGraphLayout(sf,A->cmap,A->cmap->n,NULL,PETSC_OWN_POINTER,cwant);
1653:   PetscSFSetFromOptions(sf);
1654:   for (i=0; i<n; i++) work[i] = A->cmap->rstart + i;
1655:   PetscSFReduceBegin(sf,MPIU_INT,work,cdest,MPIU_REPLACE);
1656:   PetscSFReduceEnd(sf,MPIU_INT,work,cdest,MPIU_REPLACE);
1657:   PetscSFDestroy(&sf);

1659:   ISRestoreIndices(rowp,&rwant);
1660:   ISRestoreIndices(colp,&cwant);
1661:   MatMPIAIJGetSeqAIJ(A,&aA,&aB,&gcols);

1663:   /* Find out where my gcols should go */
1664:   MatGetSize(aB,NULL,&ng);
1665:   PetscMalloc1(ng,&gcdest);
1666:   PetscSFCreate(PetscObjectComm((PetscObject)A),&sf);
1667:   PetscSFSetGraphLayout(sf,A->cmap,ng,NULL,PETSC_OWN_POINTER,gcols);
1668:   PetscSFSetFromOptions(sf);
1669:   PetscSFBcastBegin(sf,MPIU_INT,cdest,gcdest);
1670:   PetscSFBcastEnd(sf,MPIU_INT,cdest,gcdest);
1671:   PetscSFDestroy(&sf);

1673:   PetscCalloc4(m,&dnnz,m,&onnz,m,&tdnnz,m,&tonnz);
1674:   MatGetRowIJ(aA,0,PETSC_FALSE,PETSC_FALSE,&anz,&ai,&aj,&done);
1675:   MatGetRowIJ(aB,0,PETSC_FALSE,PETSC_FALSE,&bnz,&bi,&bj,&done);
1676:   for (i=0; i<m; i++) {
1677:     PetscInt    row = rdest[i];
1678:     PetscMPIInt rowner;
1679:     PetscLayoutFindOwner(A->rmap,row,&rowner);
1680:     for (j=ai[i]; j<ai[i+1]; j++) {
1681:       PetscInt    col = cdest[aj[j]];
1682:       PetscMPIInt cowner;
1683:       PetscLayoutFindOwner(A->cmap,col,&cowner); /* Could build an index for the columns to eliminate this search */
1684:       if (rowner == cowner) dnnz[i]++;
1685:       else onnz[i]++;
1686:     }
1687:     for (j=bi[i]; j<bi[i+1]; j++) {
1688:       PetscInt    col = gcdest[bj[j]];
1689:       PetscMPIInt cowner;
1690:       PetscLayoutFindOwner(A->cmap,col,&cowner);
1691:       if (rowner == cowner) dnnz[i]++;
1692:       else onnz[i]++;
1693:     }
1694:   }
1695:   PetscSFBcastBegin(rowsf,MPIU_INT,dnnz,tdnnz);
1696:   PetscSFBcastEnd(rowsf,MPIU_INT,dnnz,tdnnz);
1697:   PetscSFBcastBegin(rowsf,MPIU_INT,onnz,tonnz);
1698:   PetscSFBcastEnd(rowsf,MPIU_INT,onnz,tonnz);
1699:   PetscSFDestroy(&rowsf);

1701:   MatCreateAIJ(PetscObjectComm((PetscObject)A),A->rmap->n,A->cmap->n,A->rmap->N,A->cmap->N,0,tdnnz,0,tonnz,&Aperm);
1702:   MatSeqAIJGetArray(aA,&aa);
1703:   MatSeqAIJGetArray(aB,&ba);
1704:   for (i=0; i<m; i++) {
1705:     PetscInt *acols = dnnz,*bcols = onnz; /* Repurpose now-unneeded arrays */
1706:     PetscInt j0,rowlen;
1707:     rowlen = ai[i+1] - ai[i];
1708:     for (j0=j=0; j<rowlen; j0=j) { /* rowlen could be larger than number of rows m, so sum in batches */
1709:       for ( ; j<PetscMin(rowlen,j0+m); j++) acols[j-j0] = cdest[aj[ai[i]+j]];
1710:       MatSetValues(Aperm,1,&rdest[i],j-j0,acols,aa+ai[i]+j0,INSERT_VALUES);
1711:     }
1712:     rowlen = bi[i+1] - bi[i];
1713:     for (j0=j=0; j<rowlen; j0=j) {
1714:       for ( ; j<PetscMin(rowlen,j0+m); j++) bcols[j-j0] = gcdest[bj[bi[i]+j]];
1715:       MatSetValues(Aperm,1,&rdest[i],j-j0,bcols,ba+bi[i]+j0,INSERT_VALUES);
1716:     }
1717:   }
1718:   MatAssemblyBegin(Aperm,MAT_FINAL_ASSEMBLY);
1719:   MatAssemblyEnd(Aperm,MAT_FINAL_ASSEMBLY);
1720:   MatRestoreRowIJ(aA,0,PETSC_FALSE,PETSC_FALSE,&anz,&ai,&aj,&done);
1721:   MatRestoreRowIJ(aB,0,PETSC_FALSE,PETSC_FALSE,&bnz,&bi,&bj,&done);
1722:   MatSeqAIJRestoreArray(aA,&aa);
1723:   MatSeqAIJRestoreArray(aB,&ba);
1724:   PetscFree4(dnnz,onnz,tdnnz,tonnz);
1725:   PetscFree3(work,rdest,cdest);
1726:   PetscFree(gcdest);
1727:   if (parcolp) {ISDestroy(&colp);}
1728:   *B = Aperm;
1729:   return(0);
1730: }

1732: PetscErrorCode  MatGetGhosts_MPIAIJ(Mat mat,PetscInt *nghosts,const PetscInt *ghosts[])
1733: {
1734:   Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;

1738:   MatGetSize(aij->B,NULL,nghosts);
1739:   if (ghosts) *ghosts = aij->garray;
1740:   return(0);
1741: }

1743: PetscErrorCode MatGetInfo_MPIAIJ(Mat matin,MatInfoType flag,MatInfo *info)
1744: {
1745:   Mat_MPIAIJ     *mat = (Mat_MPIAIJ*)matin->data;
1746:   Mat            A    = mat->A,B = mat->B;
1748:   PetscLogDouble isend[5],irecv[5];

1751:   info->block_size = 1.0;
1752:   MatGetInfo(A,MAT_LOCAL,info);

1754:   isend[0] = info->nz_used; isend[1] = info->nz_allocated; isend[2] = info->nz_unneeded;
1755:   isend[3] = info->memory;  isend[4] = info->mallocs;

1757:   MatGetInfo(B,MAT_LOCAL,info);

1759:   isend[0] += info->nz_used; isend[1] += info->nz_allocated; isend[2] += info->nz_unneeded;
1760:   isend[3] += info->memory;  isend[4] += info->mallocs;
1761:   if (flag == MAT_LOCAL) {
1762:     info->nz_used      = isend[0];
1763:     info->nz_allocated = isend[1];
1764:     info->nz_unneeded  = isend[2];
1765:     info->memory       = isend[3];
1766:     info->mallocs      = isend[4];
1767:   } else if (flag == MAT_GLOBAL_MAX) {
1768:     MPIU_Allreduce(isend,irecv,5,MPIU_PETSCLOGDOUBLE,MPI_MAX,PetscObjectComm((PetscObject)matin));

1770:     info->nz_used      = irecv[0];
1771:     info->nz_allocated = irecv[1];
1772:     info->nz_unneeded  = irecv[2];
1773:     info->memory       = irecv[3];
1774:     info->mallocs      = irecv[4];
1775:   } else if (flag == MAT_GLOBAL_SUM) {
1776:     MPIU_Allreduce(isend,irecv,5,MPIU_PETSCLOGDOUBLE,MPI_SUM,PetscObjectComm((PetscObject)matin));

1778:     info->nz_used      = irecv[0];
1779:     info->nz_allocated = irecv[1];
1780:     info->nz_unneeded  = irecv[2];
1781:     info->memory       = irecv[3];
1782:     info->mallocs      = irecv[4];
1783:   }
1784:   info->fill_ratio_given  = 0; /* no parallel LU/ILU/Cholesky */
1785:   info->fill_ratio_needed = 0;
1786:   info->factor_mallocs    = 0;
1787:   return(0);
1788: }

1790: PetscErrorCode MatSetOption_MPIAIJ(Mat A,MatOption op,PetscBool flg)
1791: {
1792:   Mat_MPIAIJ     *a = (Mat_MPIAIJ*)A->data;

1796:   switch (op) {
1797:   case MAT_NEW_NONZERO_LOCATIONS:
1798:   case MAT_NEW_NONZERO_ALLOCATION_ERR:
1799:   case MAT_UNUSED_NONZERO_LOCATION_ERR:
1800:   case MAT_KEEP_NONZERO_PATTERN:
1801:   case MAT_NEW_NONZERO_LOCATION_ERR:
1802:   case MAT_USE_INODES:
1803:   case MAT_IGNORE_ZERO_ENTRIES:
1804:     MatCheckPreallocated(A,1);
1805:     MatSetOption(a->A,op,flg);
1806:     MatSetOption(a->B,op,flg);
1807:     break;
1808:   case MAT_ROW_ORIENTED:
1809:     MatCheckPreallocated(A,1);
1810:     a->roworiented = flg;

1812:     MatSetOption(a->A,op,flg);
1813:     MatSetOption(a->B,op,flg);
1814:     break;
1815:   case MAT_NEW_DIAGONALS:
1816:   case MAT_SORTED_FULL:
1817:     PetscInfo1(A,"Option %s ignored\n",MatOptions[op]);
1818:     break;
1819:   case MAT_IGNORE_OFF_PROC_ENTRIES:
1820:     a->donotstash = flg;
1821:     break;
1822:   /* Symmetry flags are handled directly by MatSetOption() and they don't affect preallocation */
1823:   case MAT_SPD:
1824:   case MAT_SYMMETRIC:
1825:   case MAT_STRUCTURALLY_SYMMETRIC:
1826:   case MAT_HERMITIAN:
1827:   case MAT_SYMMETRY_ETERNAL:
1828:     break;
1829:   case MAT_SUBMAT_SINGLEIS:
1830:     A->submat_singleis = flg;
1831:     break;
1832:   case MAT_STRUCTURE_ONLY:
1833:     /* The option is handled directly by MatSetOption() */
1834:     break;
1835:   default:
1836:     SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"unknown option %d",op);
1837:   }
1838:   return(0);
1839: }

1841: PetscErrorCode MatGetRow_MPIAIJ(Mat matin,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v)
1842: {
1843:   Mat_MPIAIJ     *mat = (Mat_MPIAIJ*)matin->data;
1844:   PetscScalar    *vworkA,*vworkB,**pvA,**pvB,*v_p;
1846:   PetscInt       i,*cworkA,*cworkB,**pcA,**pcB,cstart = matin->cmap->rstart;
1847:   PetscInt       nztot,nzA,nzB,lrow,rstart = matin->rmap->rstart,rend = matin->rmap->rend;
1848:   PetscInt       *cmap,*idx_p;

1851:   if (mat->getrowactive) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Already active");
1852:   mat->getrowactive = PETSC_TRUE;

1854:   if (!mat->rowvalues && (idx || v)) {
1855:     /*
1856:         allocate enough space to hold information from the longest row.
1857:     */
1858:     Mat_SeqAIJ *Aa = (Mat_SeqAIJ*)mat->A->data,*Ba = (Mat_SeqAIJ*)mat->B->data;
1859:     PetscInt   max = 1,tmp;
1860:     for (i=0; i<matin->rmap->n; i++) {
1861:       tmp = Aa->i[i+1] - Aa->i[i] + Ba->i[i+1] - Ba->i[i];
1862:       if (max < tmp) max = tmp;
1863:     }
1864:     PetscMalloc2(max,&mat->rowvalues,max,&mat->rowindices);
1865:   }

1867:   if (row < rstart || row >= rend) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Only local rows");
1868:   lrow = row - rstart;

1870:   pvA = &vworkA; pcA = &cworkA; pvB = &vworkB; pcB = &cworkB;
1871:   if (!v)   {pvA = 0; pvB = 0;}
1872:   if (!idx) {pcA = 0; if (!v) pcB = 0;}
1873:   (*mat->A->ops->getrow)(mat->A,lrow,&nzA,pcA,pvA);
1874:   (*mat->B->ops->getrow)(mat->B,lrow,&nzB,pcB,pvB);
1875:   nztot = nzA + nzB;

1877:   cmap = mat->garray;
1878:   if (v  || idx) {
1879:     if (nztot) {
1880:       /* Sort by increasing column numbers, assuming A and B already sorted */
1881:       PetscInt imark = -1;
1882:       if (v) {
1883:         *v = v_p = mat->rowvalues;
1884:         for (i=0; i<nzB; i++) {
1885:           if (cmap[cworkB[i]] < cstart) v_p[i] = vworkB[i];
1886:           else break;
1887:         }
1888:         imark = i;
1889:         for (i=0; i<nzA; i++)     v_p[imark+i] = vworkA[i];
1890:         for (i=imark; i<nzB; i++) v_p[nzA+i]   = vworkB[i];
1891:       }
1892:       if (idx) {
1893:         *idx = idx_p = mat->rowindices;
1894:         if (imark > -1) {
1895:           for (i=0; i<imark; i++) {
1896:             idx_p[i] = cmap[cworkB[i]];
1897:           }
1898:         } else {
1899:           for (i=0; i<nzB; i++) {
1900:             if (cmap[cworkB[i]] < cstart) idx_p[i] = cmap[cworkB[i]];
1901:             else break;
1902:           }
1903:           imark = i;
1904:         }
1905:         for (i=0; i<nzA; i++)     idx_p[imark+i] = cstart + cworkA[i];
1906:         for (i=imark; i<nzB; i++) idx_p[nzA+i]   = cmap[cworkB[i]];
1907:       }
1908:     } else {
1909:       if (idx) *idx = 0;
1910:       if (v)   *v   = 0;
1911:     }
1912:   }
1913:   *nz  = nztot;
1914:   (*mat->A->ops->restorerow)(mat->A,lrow,&nzA,pcA,pvA);
1915:   (*mat->B->ops->restorerow)(mat->B,lrow,&nzB,pcB,pvB);
1916:   return(0);
1917: }

1919: PetscErrorCode MatRestoreRow_MPIAIJ(Mat mat,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v)
1920: {
1921:   Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;

1924:   if (!aij->getrowactive) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"MatGetRow() must be called first");
1925:   aij->getrowactive = PETSC_FALSE;
1926:   return(0);
1927: }

1929: PetscErrorCode MatNorm_MPIAIJ(Mat mat,NormType type,PetscReal *norm)
1930: {
1931:   Mat_MPIAIJ     *aij  = (Mat_MPIAIJ*)mat->data;
1932:   Mat_SeqAIJ     *amat = (Mat_SeqAIJ*)aij->A->data,*bmat = (Mat_SeqAIJ*)aij->B->data;
1934:   PetscInt       i,j,cstart = mat->cmap->rstart;
1935:   PetscReal      sum = 0.0;
1936:   MatScalar      *v;

1939:   if (aij->size == 1) {
1940:      MatNorm(aij->A,type,norm);
1941:   } else {
1942:     if (type == NORM_FROBENIUS) {
1943:       v = amat->a;
1944:       for (i=0; i<amat->nz; i++) {
1945:         sum += PetscRealPart(PetscConj(*v)*(*v)); v++;
1946:       }
1947:       v = bmat->a;
1948:       for (i=0; i<bmat->nz; i++) {
1949:         sum += PetscRealPart(PetscConj(*v)*(*v)); v++;
1950:       }
1951:       MPIU_Allreduce(&sum,norm,1,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)mat));
1952:       *norm = PetscSqrtReal(*norm);
1953:       PetscLogFlops(2.0*amat->nz+2.0*bmat->nz);
1954:     } else if (type == NORM_1) { /* max column norm */
1955:       PetscReal *tmp,*tmp2;
1956:       PetscInt  *jj,*garray = aij->garray;
1957:       PetscCalloc1(mat->cmap->N+1,&tmp);
1958:       PetscMalloc1(mat->cmap->N+1,&tmp2);
1959:       *norm = 0.0;
1960:       v     = amat->a; jj = amat->j;
1961:       for (j=0; j<amat->nz; j++) {
1962:         tmp[cstart + *jj++] += PetscAbsScalar(*v);  v++;
1963:       }
1964:       v = bmat->a; jj = bmat->j;
1965:       for (j=0; j<bmat->nz; j++) {
1966:         tmp[garray[*jj++]] += PetscAbsScalar(*v); v++;
1967:       }
1968:       MPIU_Allreduce(tmp,tmp2,mat->cmap->N,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)mat));
1969:       for (j=0; j<mat->cmap->N; j++) {
1970:         if (tmp2[j] > *norm) *norm = tmp2[j];
1971:       }
1972:       PetscFree(tmp);
1973:       PetscFree(tmp2);
1974:       PetscLogFlops(PetscMax(amat->nz+bmat->nz-1,0));
1975:     } else if (type == NORM_INFINITY) { /* max row norm */
1976:       PetscReal ntemp = 0.0;
1977:       for (j=0; j<aij->A->rmap->n; j++) {
1978:         v   = amat->a + amat->i[j];
1979:         sum = 0.0;
1980:         for (i=0; i<amat->i[j+1]-amat->i[j]; i++) {
1981:           sum += PetscAbsScalar(*v); v++;
1982:         }
1983:         v = bmat->a + bmat->i[j];
1984:         for (i=0; i<bmat->i[j+1]-bmat->i[j]; i++) {
1985:           sum += PetscAbsScalar(*v); v++;
1986:         }
1987:         if (sum > ntemp) ntemp = sum;
1988:       }
1989:       MPIU_Allreduce(&ntemp,norm,1,MPIU_REAL,MPIU_MAX,PetscObjectComm((PetscObject)mat));
1990:       PetscLogFlops(PetscMax(amat->nz+bmat->nz-1,0));
1991:     } else SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"No support for two norm");
1992:   }
1993:   return(0);
1994: }

1996: PetscErrorCode MatTranspose_MPIAIJ(Mat A,MatReuse reuse,Mat *matout)
1997: {
1998:   Mat_MPIAIJ      *a    =(Mat_MPIAIJ*)A->data,*b;
1999:   Mat_SeqAIJ      *Aloc =(Mat_SeqAIJ*)a->A->data,*Bloc=(Mat_SeqAIJ*)a->B->data,*sub_B_diag;
2000:   PetscInt        M     = A->rmap->N,N=A->cmap->N,ma,na,mb,nb,row,*cols,*cols_tmp,*B_diag_ilen,i,ncol,A_diag_ncol;
2001:   const PetscInt  *ai,*aj,*bi,*bj,*B_diag_i;
2002:   PetscErrorCode  ierr;
2003:   Mat             B,A_diag,*B_diag;
2004:   const MatScalar *array;

2007:   ma = A->rmap->n; na = A->cmap->n; mb = a->B->rmap->n; nb = a->B->cmap->n;
2008:   ai = Aloc->i; aj = Aloc->j;
2009:   bi = Bloc->i; bj = Bloc->j;
2010:   if (reuse == MAT_INITIAL_MATRIX || *matout == A) {
2011:     PetscInt             *d_nnz,*g_nnz,*o_nnz;
2012:     PetscSFNode          *oloc;
2013:     PETSC_UNUSED PetscSF sf;

2015:     PetscMalloc4(na,&d_nnz,na,&o_nnz,nb,&g_nnz,nb,&oloc);
2016:     /* compute d_nnz for preallocation */
2017:     PetscArrayzero(d_nnz,na);
2018:     for (i=0; i<ai[ma]; i++) {
2019:       d_nnz[aj[i]]++;
2020:     }
2021:     /* compute local off-diagonal contributions */
2022:     PetscArrayzero(g_nnz,nb);
2023:     for (i=0; i<bi[ma]; i++) g_nnz[bj[i]]++;
2024:     /* map those to global */
2025:     PetscSFCreate(PetscObjectComm((PetscObject)A),&sf);
2026:     PetscSFSetGraphLayout(sf,A->cmap,nb,NULL,PETSC_USE_POINTER,a->garray);
2027:     PetscSFSetFromOptions(sf);
2028:     PetscArrayzero(o_nnz,na);
2029:     PetscSFReduceBegin(sf,MPIU_INT,g_nnz,o_nnz,MPIU_SUM);
2030:     PetscSFReduceEnd(sf,MPIU_INT,g_nnz,o_nnz,MPIU_SUM);
2031:     PetscSFDestroy(&sf);

2033:     MatCreate(PetscObjectComm((PetscObject)A),&B);
2034:     MatSetSizes(B,A->cmap->n,A->rmap->n,N,M);
2035:     MatSetBlockSizes(B,PetscAbs(A->cmap->bs),PetscAbs(A->rmap->bs));
2036:     MatSetType(B,((PetscObject)A)->type_name);
2037:     MatMPIAIJSetPreallocation(B,0,d_nnz,0,o_nnz);
2038:     PetscFree4(d_nnz,o_nnz,g_nnz,oloc);
2039:   } else {
2040:     B    = *matout;
2041:     MatSetOption(B,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_TRUE);
2042:   }

2044:   b           = (Mat_MPIAIJ*)B->data;
2045:   A_diag      = a->A;
2046:   B_diag      = &b->A;
2047:   sub_B_diag  = (Mat_SeqAIJ*)(*B_diag)->data;
2048:   A_diag_ncol = A_diag->cmap->N;
2049:   B_diag_ilen = sub_B_diag->ilen;
2050:   B_diag_i    = sub_B_diag->i;

2052:   /* Set ilen for diagonal of B */
2053:   for (i=0; i<A_diag_ncol; i++) {
2054:     B_diag_ilen[i] = B_diag_i[i+1] - B_diag_i[i];
2055:   }

2057:   /* Transpose the diagonal part of the matrix. In contrast to the offdiagonal part, this can be done
2058:   very quickly (=without using MatSetValues), because all writes are local. */
2059:   MatTranspose(A_diag,MAT_REUSE_MATRIX,B_diag);

2061:   /* copy over the B part */
2062:   PetscMalloc1(bi[mb],&cols);
2063:   array = Bloc->a;
2064:   row   = A->rmap->rstart;
2065:   for (i=0; i<bi[mb]; i++) cols[i] = a->garray[bj[i]];
2066:   cols_tmp = cols;
2067:   for (i=0; i<mb; i++) {
2068:     ncol = bi[i+1]-bi[i];
2069:     MatSetValues(B,ncol,cols_tmp,1,&row,array,INSERT_VALUES);
2070:     row++;
2071:     array += ncol; cols_tmp += ncol;
2072:   }
2073:   PetscFree(cols);

2075:   MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);
2076:   MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);
2077:   if (reuse == MAT_INITIAL_MATRIX || reuse == MAT_REUSE_MATRIX) {
2078:     *matout = B;
2079:   } else {
2080:     MatHeaderMerge(A,&B);
2081:   }
2082:   return(0);
2083: }

2085: PetscErrorCode MatDiagonalScale_MPIAIJ(Mat mat,Vec ll,Vec rr)
2086: {
2087:   Mat_MPIAIJ     *aij = (Mat_MPIAIJ*)mat->data;
2088:   Mat            a    = aij->A,b = aij->B;
2090:   PetscInt       s1,s2,s3;

2093:   MatGetLocalSize(mat,&s2,&s3);
2094:   if (rr) {
2095:     VecGetLocalSize(rr,&s1);
2096:     if (s1!=s3) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"right vector non-conforming local size");
2097:     /* Overlap communication with computation. */
2098:     VecScatterBegin(aij->Mvctx,rr,aij->lvec,INSERT_VALUES,SCATTER_FORWARD);
2099:   }
2100:   if (ll) {
2101:     VecGetLocalSize(ll,&s1);
2102:     if (s1!=s2) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"left vector non-conforming local size");
2103:     (*b->ops->diagonalscale)(b,ll,0);
2104:   }
2105:   /* scale  the diagonal block */
2106:   (*a->ops->diagonalscale)(a,ll,rr);

2108:   if (rr) {
2109:     /* Do a scatter end and then right scale the off-diagonal block */
2110:     VecScatterEnd(aij->Mvctx,rr,aij->lvec,INSERT_VALUES,SCATTER_FORWARD);
2111:     (*b->ops->diagonalscale)(b,0,aij->lvec);
2112:   }
2113:   return(0);
2114: }

2116: PetscErrorCode MatSetUnfactored_MPIAIJ(Mat A)
2117: {
2118:   Mat_MPIAIJ     *a = (Mat_MPIAIJ*)A->data;

2122:   MatSetUnfactored(a->A);
2123:   return(0);
2124: }

2126: PetscErrorCode MatEqual_MPIAIJ(Mat A,Mat B,PetscBool  *flag)
2127: {
2128:   Mat_MPIAIJ     *matB = (Mat_MPIAIJ*)B->data,*matA = (Mat_MPIAIJ*)A->data;
2129:   Mat            a,b,c,d;
2130:   PetscBool      flg;

2134:   a = matA->A; b = matA->B;
2135:   c = matB->A; d = matB->B;

2137:   MatEqual(a,c,&flg);
2138:   if (flg) {
2139:     MatEqual(b,d,&flg);
2140:   }
2141:   MPIU_Allreduce(&flg,flag,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)A));
2142:   return(0);
2143: }

2145: PetscErrorCode MatCopy_MPIAIJ(Mat A,Mat B,MatStructure str)
2146: {
2148:   Mat_MPIAIJ     *a = (Mat_MPIAIJ*)A->data;
2149:   Mat_MPIAIJ     *b = (Mat_MPIAIJ*)B->data;

2152:   /* If the two matrices don't have the same copy implementation, they aren't compatible for fast copy. */
2153:   if ((str != SAME_NONZERO_PATTERN) || (A->ops->copy != B->ops->copy)) {
2154:     /* because of the column compression in the off-processor part of the matrix a->B,
2155:        the number of columns in a->B and b->B may be different, hence we cannot call
2156:        the MatCopy() directly on the two parts. If need be, we can provide a more
2157:        efficient copy than the MatCopy_Basic() by first uncompressing the a->B matrices
2158:        then copying the submatrices */
2159:     MatCopy_Basic(A,B,str);
2160:   } else {
2161:     MatCopy(a->A,b->A,str);
2162:     MatCopy(a->B,b->B,str);
2163:   }
2164:   PetscObjectStateIncrease((PetscObject)B);
2165:   return(0);
2166: }

2168: PetscErrorCode MatSetUp_MPIAIJ(Mat A)
2169: {

2173:   MatMPIAIJSetPreallocation(A,PETSC_DEFAULT,0,PETSC_DEFAULT,0);
2174:   return(0);
2175: }

2177: /*
2178:    Computes the number of nonzeros per row needed for preallocation when X and Y
2179:    have different nonzero structure.
2180: */
2181: PetscErrorCode MatAXPYGetPreallocation_MPIX_private(PetscInt m,const PetscInt *xi,const PetscInt *xj,const PetscInt *xltog,const PetscInt *yi,const PetscInt *yj,const PetscInt *yltog,PetscInt *nnz)
2182: {
2183:   PetscInt       i,j,k,nzx,nzy;

2186:   /* Set the number of nonzeros in the new matrix */
2187:   for (i=0; i<m; i++) {
2188:     const PetscInt *xjj = xj+xi[i],*yjj = yj+yi[i];
2189:     nzx = xi[i+1] - xi[i];
2190:     nzy = yi[i+1] - yi[i];
2191:     nnz[i] = 0;
2192:     for (j=0,k=0; j<nzx; j++) {                   /* Point in X */
2193:       for (; k<nzy && yltog[yjj[k]]<xltog[xjj[j]]; k++) nnz[i]++; /* Catch up to X */
2194:       if (k<nzy && yltog[yjj[k]]==xltog[xjj[j]]) k++;             /* Skip duplicate */
2195:       nnz[i]++;
2196:     }
2197:     for (; k<nzy; k++) nnz[i]++;
2198:   }
2199:   return(0);
2200: }

2202: /* This is the same as MatAXPYGetPreallocation_SeqAIJ, except that the local-to-global map is provided */
2203: static PetscErrorCode MatAXPYGetPreallocation_MPIAIJ(Mat Y,const PetscInt *yltog,Mat X,const PetscInt *xltog,PetscInt *nnz)
2204: {
2206:   PetscInt       m = Y->rmap->N;
2207:   Mat_SeqAIJ     *x = (Mat_SeqAIJ*)X->data;
2208:   Mat_SeqAIJ     *y = (Mat_SeqAIJ*)Y->data;

2211:   MatAXPYGetPreallocation_MPIX_private(m,x->i,x->j,xltog,y->i,y->j,yltog,nnz);
2212:   return(0);
2213: }

2215: PetscErrorCode MatAXPY_MPIAIJ(Mat Y,PetscScalar a,Mat X,MatStructure str)
2216: {
2218:   Mat_MPIAIJ     *xx = (Mat_MPIAIJ*)X->data,*yy = (Mat_MPIAIJ*)Y->data;
2219:   PetscBLASInt   bnz,one=1;
2220:   Mat_SeqAIJ     *x,*y;

2223:   if (str == SAME_NONZERO_PATTERN) {
2224:     PetscScalar alpha = a;
2225:     x    = (Mat_SeqAIJ*)xx->A->data;
2226:     PetscBLASIntCast(x->nz,&bnz);
2227:     y    = (Mat_SeqAIJ*)yy->A->data;
2228:     PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&bnz,&alpha,x->a,&one,y->a,&one));
2229:     x    = (Mat_SeqAIJ*)xx->B->data;
2230:     y    = (Mat_SeqAIJ*)yy->B->data;
2231:     PetscBLASIntCast(x->nz,&bnz);
2232:     PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&bnz,&alpha,x->a,&one,y->a,&one));
2233:     PetscObjectStateIncrease((PetscObject)Y);
2234:     /* the MatAXPY_Basic* subroutines calls MatAssembly, so the matrix on the GPU
2235:        will be updated */
2236: #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA)
2237:     if (Y->offloadmask != PETSC_OFFLOAD_UNALLOCATED) {
2238:       Y->offloadmask = PETSC_OFFLOAD_CPU;
2239:     }
2240: #endif
2241:   } else if (str == SUBSET_NONZERO_PATTERN) { /* nonzeros of X is a subset of Y's */
2242:     MatAXPY_Basic(Y,a,X,str);
2243:   } else {
2244:     Mat      B;
2245:     PetscInt *nnz_d,*nnz_o;
2246:     PetscMalloc1(yy->A->rmap->N,&nnz_d);
2247:     PetscMalloc1(yy->B->rmap->N,&nnz_o);
2248:     MatCreate(PetscObjectComm((PetscObject)Y),&B);
2249:     PetscObjectSetName((PetscObject)B,((PetscObject)Y)->name);
2250:     MatSetSizes(B,Y->rmap->n,Y->cmap->n,Y->rmap->N,Y->cmap->N);
2251:     MatSetBlockSizesFromMats(B,Y,Y);
2252:     MatSetType(B,MATMPIAIJ);
2253:     MatAXPYGetPreallocation_SeqAIJ(yy->A,xx->A,nnz_d);
2254:     MatAXPYGetPreallocation_MPIAIJ(yy->B,yy->garray,xx->B,xx->garray,nnz_o);
2255:     MatMPIAIJSetPreallocation(B,0,nnz_d,0,nnz_o);
2256:     MatAXPY_BasicWithPreallocation(B,Y,a,X,str);
2257:     MatHeaderReplace(Y,&B);
2258:     PetscFree(nnz_d);
2259:     PetscFree(nnz_o);
2260:   }
2261:   return(0);
2262: }

2264: extern PetscErrorCode  MatConjugate_SeqAIJ(Mat);

2266: PetscErrorCode  MatConjugate_MPIAIJ(Mat mat)
2267: {
2268: #if defined(PETSC_USE_COMPLEX)
2270:   Mat_MPIAIJ     *aij = (Mat_MPIAIJ*)mat->data;

2273:   MatConjugate_SeqAIJ(aij->A);
2274:   MatConjugate_SeqAIJ(aij->B);
2275: #else
2277: #endif
2278:   return(0);
2279: }

2281: PetscErrorCode MatRealPart_MPIAIJ(Mat A)
2282: {
2283:   Mat_MPIAIJ     *a = (Mat_MPIAIJ*)A->data;

2287:   MatRealPart(a->A);
2288:   MatRealPart(a->B);
2289:   return(0);
2290: }

2292: PetscErrorCode MatImaginaryPart_MPIAIJ(Mat A)
2293: {
2294:   Mat_MPIAIJ     *a = (Mat_MPIAIJ*)A->data;

2298:   MatImaginaryPart(a->A);
2299:   MatImaginaryPart(a->B);
2300:   return(0);
2301: }

2303: PetscErrorCode MatGetRowMaxAbs_MPIAIJ(Mat A, Vec v, PetscInt idx[])
2304: {
2305:   Mat_MPIAIJ     *a = (Mat_MPIAIJ*)A->data;
2307:   PetscInt       i,*idxb = 0;
2308:   PetscScalar    *va,*vb;
2309:   Vec            vtmp;

2312:   MatGetRowMaxAbs(a->A,v,idx);
2313:   VecGetArray(v,&va);
2314:   if (idx) {
2315:     for (i=0; i<A->rmap->n; i++) {
2316:       if (PetscAbsScalar(va[i])) idx[i] += A->cmap->rstart;
2317:     }
2318:   }

2320:   VecCreateSeq(PETSC_COMM_SELF,A->rmap->n,&vtmp);
2321:   if (idx) {
2322:     PetscMalloc1(A->rmap->n,&idxb);
2323:   }
2324:   MatGetRowMaxAbs(a->B,vtmp,idxb);
2325:   VecGetArray(vtmp,&vb);

2327:   for (i=0; i<A->rmap->n; i++) {
2328:     if (PetscAbsScalar(va[i]) < PetscAbsScalar(vb[i])) {
2329:       va[i] = vb[i];
2330:       if (idx) idx[i] = a->garray[idxb[i]];
2331:     }
2332:   }

2334:   VecRestoreArray(v,&va);
2335:   VecRestoreArray(vtmp,&vb);
2336:   PetscFree(idxb);
2337:   VecDestroy(&vtmp);
2338:   return(0);
2339: }

2341: PetscErrorCode MatGetRowMinAbs_MPIAIJ(Mat A, Vec v, PetscInt idx[])
2342: {
2343:   Mat_MPIAIJ     *a = (Mat_MPIAIJ*)A->data;
2345:   PetscInt       i,*idxb = 0;
2346:   PetscScalar    *va,*vb;
2347:   Vec            vtmp;

2350:   MatGetRowMinAbs(a->A,v,idx);
2351:   VecGetArray(v,&va);
2352:   if (idx) {
2353:     for (i=0; i<A->cmap->n; i++) {
2354:       if (PetscAbsScalar(va[i])) idx[i] += A->cmap->rstart;
2355:     }
2356:   }

2358:   VecCreateSeq(PETSC_COMM_SELF,A->rmap->n,&vtmp);
2359:   if (idx) {
2360:     PetscMalloc1(A->rmap->n,&idxb);
2361:   }
2362:   MatGetRowMinAbs(a->B,vtmp,idxb);
2363:   VecGetArray(vtmp,&vb);

2365:   for (i=0; i<A->rmap->n; i++) {
2366:     if (PetscAbsScalar(va[i]) > PetscAbsScalar(vb[i])) {
2367:       va[i] = vb[i];
2368:       if (idx) idx[i] = a->garray[idxb[i]];
2369:     }
2370:   }

2372:   VecRestoreArray(v,&va);
2373:   VecRestoreArray(vtmp,&vb);
2374:   PetscFree(idxb);
2375:   VecDestroy(&vtmp);
2376:   return(0);
2377: }

2379: PetscErrorCode MatGetRowMin_MPIAIJ(Mat A, Vec v, PetscInt idx[])
2380: {
2381:   Mat_MPIAIJ     *mat   = (Mat_MPIAIJ*) A->data;
2382:   PetscInt       n      = A->rmap->n;
2383:   PetscInt       cstart = A->cmap->rstart;
2384:   PetscInt       *cmap  = mat->garray;
2385:   PetscInt       *diagIdx, *offdiagIdx;
2386:   Vec            diagV, offdiagV;
2387:   PetscScalar    *a, *diagA, *offdiagA;
2388:   PetscInt       r;

2392:   PetscMalloc2(n,&diagIdx,n,&offdiagIdx);
2393:   VecCreateSeq(PetscObjectComm((PetscObject)A), n, &diagV);
2394:   VecCreateSeq(PetscObjectComm((PetscObject)A), n, &offdiagV);
2395:   MatGetRowMin(mat->A, diagV,    diagIdx);
2396:   MatGetRowMin(mat->B, offdiagV, offdiagIdx);
2397:   VecGetArray(v,        &a);
2398:   VecGetArray(diagV,    &diagA);
2399:   VecGetArray(offdiagV, &offdiagA);
2400:   for (r = 0; r < n; ++r) {
2401:     if (PetscAbsScalar(diagA[r]) <= PetscAbsScalar(offdiagA[r])) {
2402:       a[r]   = diagA[r];
2403:       idx[r] = cstart + diagIdx[r];
2404:     } else {
2405:       a[r]   = offdiagA[r];
2406:       idx[r] = cmap[offdiagIdx[r]];
2407:     }
2408:   }
2409:   VecRestoreArray(v,        &a);
2410:   VecRestoreArray(diagV,    &diagA);
2411:   VecRestoreArray(offdiagV, &offdiagA);
2412:   VecDestroy(&diagV);
2413:   VecDestroy(&offdiagV);
2414:   PetscFree2(diagIdx, offdiagIdx);
2415:   return(0);
2416: }

2418: PetscErrorCode MatGetRowMax_MPIAIJ(Mat A, Vec v, PetscInt idx[])
2419: {
2420:   Mat_MPIAIJ     *mat   = (Mat_MPIAIJ*) A->data;
2421:   PetscInt       n      = A->rmap->n;
2422:   PetscInt       cstart = A->cmap->rstart;
2423:   PetscInt       *cmap  = mat->garray;
2424:   PetscInt       *diagIdx, *offdiagIdx;
2425:   Vec            diagV, offdiagV;
2426:   PetscScalar    *a, *diagA, *offdiagA;
2427:   PetscInt       r;

2431:   PetscMalloc2(n,&diagIdx,n,&offdiagIdx);
2432:   VecCreateSeq(PETSC_COMM_SELF, n, &diagV);
2433:   VecCreateSeq(PETSC_COMM_SELF, n, &offdiagV);
2434:   MatGetRowMax(mat->A, diagV,    diagIdx);
2435:   MatGetRowMax(mat->B, offdiagV, offdiagIdx);
2436:   VecGetArray(v,        &a);
2437:   VecGetArray(diagV,    &diagA);
2438:   VecGetArray(offdiagV, &offdiagA);
2439:   for (r = 0; r < n; ++r) {
2440:     if (PetscAbsScalar(diagA[r]) >= PetscAbsScalar(offdiagA[r])) {
2441:       a[r]   = diagA[r];
2442:       idx[r] = cstart + diagIdx[r];
2443:     } else {
2444:       a[r]   = offdiagA[r];
2445:       idx[r] = cmap[offdiagIdx[r]];
2446:     }
2447:   }
2448:   VecRestoreArray(v,        &a);
2449:   VecRestoreArray(diagV,    &diagA);
2450:   VecRestoreArray(offdiagV, &offdiagA);
2451:   VecDestroy(&diagV);
2452:   VecDestroy(&offdiagV);
2453:   PetscFree2(diagIdx, offdiagIdx);
2454:   return(0);
2455: }

2457: PetscErrorCode MatGetSeqNonzeroStructure_MPIAIJ(Mat mat,Mat *newmat)
2458: {
2460:   Mat            *dummy;

2463:   MatCreateSubMatrix_MPIAIJ_All(mat,MAT_DO_NOT_GET_VALUES,MAT_INITIAL_MATRIX,&dummy);
2464:   *newmat = *dummy;
2465:   PetscFree(dummy);
2466:   return(0);
2467: }

2469: PetscErrorCode  MatInvertBlockDiagonal_MPIAIJ(Mat A,const PetscScalar **values)
2470: {
2471:   Mat_MPIAIJ     *a = (Mat_MPIAIJ*) A->data;

2475:   MatInvertBlockDiagonal(a->A,values);
2476:   A->factorerrortype = a->A->factorerrortype;
2477:   return(0);
2478: }

2480: static PetscErrorCode  MatSetRandom_MPIAIJ(Mat x,PetscRandom rctx)
2481: {
2483:   Mat_MPIAIJ     *aij = (Mat_MPIAIJ*)x->data;

2486:   if (!x->assembled && !x->preallocated) SETERRQ(PetscObjectComm((PetscObject)x), PETSC_ERR_ARG_WRONGSTATE, "MatSetRandom on an unassembled and unpreallocated MATMPIAIJ is not allowed");
2487:   MatSetRandom(aij->A,rctx);
2488:   if (x->assembled) {
2489:     MatSetRandom(aij->B,rctx);
2490:   } else {
2491:     MatSetRandomSkipColumnRange_SeqAIJ_Private(aij->B,x->cmap->rstart,x->cmap->rend,rctx);
2492:   }
2493:   MatAssemblyBegin(x,MAT_FINAL_ASSEMBLY);
2494:   MatAssemblyEnd(x,MAT_FINAL_ASSEMBLY);
2495:   return(0);
2496: }

2498: PetscErrorCode MatMPIAIJSetUseScalableIncreaseOverlap_MPIAIJ(Mat A,PetscBool sc)
2499: {
2501:   if (sc) A->ops->increaseoverlap = MatIncreaseOverlap_MPIAIJ_Scalable;
2502:   else A->ops->increaseoverlap    = MatIncreaseOverlap_MPIAIJ;
2503:   return(0);
2504: }

2506: /*@
2507:    MatMPIAIJSetUseScalableIncreaseOverlap - Determine if the matrix uses a scalable algorithm to compute the overlap

2509:    Collective on Mat

2511:    Input Parameters:
2512: +    A - the matrix
2513: -    sc - PETSC_TRUE indicates use the scalable algorithm (default is not to use the scalable algorithm)

2515:  Level: advanced

2517: @*/
2518: PetscErrorCode MatMPIAIJSetUseScalableIncreaseOverlap(Mat A,PetscBool sc)
2519: {
2520:   PetscErrorCode       ierr;

2523:   PetscTryMethod(A,"MatMPIAIJSetUseScalableIncreaseOverlap_C",(Mat,PetscBool),(A,sc));
2524:   return(0);
2525: }

2527: PetscErrorCode MatSetFromOptions_MPIAIJ(PetscOptionItems *PetscOptionsObject,Mat A)
2528: {
2529:   PetscErrorCode       ierr;
2530:   PetscBool            sc = PETSC_FALSE,flg;

2533:   PetscOptionsHead(PetscOptionsObject,"MPIAIJ options");
2534:   if (A->ops->increaseoverlap == MatIncreaseOverlap_MPIAIJ_Scalable) sc = PETSC_TRUE;
2535:   PetscOptionsBool("-mat_increase_overlap_scalable","Use a scalable algorithm to compute the overlap","MatIncreaseOverlap",sc,&sc,&flg);
2536:   if (flg) {
2537:     MatMPIAIJSetUseScalableIncreaseOverlap(A,sc);
2538:   }
2539:   PetscOptionsTail();
2540:   return(0);
2541: }

2543: PetscErrorCode MatShift_MPIAIJ(Mat Y,PetscScalar a)
2544: {
2546:   Mat_MPIAIJ     *maij = (Mat_MPIAIJ*)Y->data;
2547:   Mat_SeqAIJ     *aij = (Mat_SeqAIJ*)maij->A->data;

2550:   if (!Y->preallocated) {
2551:     MatMPIAIJSetPreallocation(Y,1,NULL,0,NULL);
2552:   } else if (!aij->nz) {
2553:     PetscInt nonew = aij->nonew;
2554:     MatSeqAIJSetPreallocation(maij->A,1,NULL);
2555:     aij->nonew = nonew;
2556:   }
2557:   MatShift_Basic(Y,a);
2558:   return(0);
2559: }

2561: PetscErrorCode MatMissingDiagonal_MPIAIJ(Mat A,PetscBool  *missing,PetscInt *d)
2562: {
2563:   Mat_MPIAIJ     *a = (Mat_MPIAIJ*)A->data;

2567:   if (A->rmap->n != A->cmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Only works for square matrices");
2568:   MatMissingDiagonal(a->A,missing,d);
2569:   if (d) {
2570:     PetscInt rstart;
2571:     MatGetOwnershipRange(A,&rstart,NULL);
2572:     *d += rstart;

2574:   }
2575:   return(0);
2576: }

2578: PetscErrorCode MatInvertVariableBlockDiagonal_MPIAIJ(Mat A,PetscInt nblocks,const PetscInt *bsizes,PetscScalar *diag)
2579: {
2580:   Mat_MPIAIJ     *a = (Mat_MPIAIJ*)A->data;

2584:   MatInvertVariableBlockDiagonal(a->A,nblocks,bsizes,diag);
2585:   return(0);
2586: }

2588: /* -------------------------------------------------------------------*/
2589: static struct _MatOps MatOps_Values = {MatSetValues_MPIAIJ,
2590:                                        MatGetRow_MPIAIJ,
2591:                                        MatRestoreRow_MPIAIJ,
2592:                                        MatMult_MPIAIJ,
2593:                                 /* 4*/ MatMultAdd_MPIAIJ,
2594:                                        MatMultTranspose_MPIAIJ,
2595:                                        MatMultTransposeAdd_MPIAIJ,
2596:                                        0,
2597:                                        0,
2598:                                        0,
2599:                                 /*10*/ 0,
2600:                                        0,
2601:                                        0,
2602:                                        MatSOR_MPIAIJ,
2603:                                        MatTranspose_MPIAIJ,
2604:                                 /*15*/ MatGetInfo_MPIAIJ,
2605:                                        MatEqual_MPIAIJ,
2606:                                        MatGetDiagonal_MPIAIJ,
2607:                                        MatDiagonalScale_MPIAIJ,
2608:                                        MatNorm_MPIAIJ,
2609:                                 /*20*/ MatAssemblyBegin_MPIAIJ,
2610:                                        MatAssemblyEnd_MPIAIJ,
2611:                                        MatSetOption_MPIAIJ,
2612:                                        MatZeroEntries_MPIAIJ,
2613:                                 /*24*/ MatZeroRows_MPIAIJ,
2614:                                        0,
2615:                                        0,
2616:                                        0,
2617:                                        0,
2618:                                 /*29*/ MatSetUp_MPIAIJ,
2619:                                        0,
2620:                                        0,
2621:                                        MatGetDiagonalBlock_MPIAIJ,
2622:                                        0,
2623:                                 /*34*/ MatDuplicate_MPIAIJ,
2624:                                        0,
2625:                                        0,
2626:                                        0,
2627:                                        0,
2628:                                 /*39*/ MatAXPY_MPIAIJ,
2629:                                        MatCreateSubMatrices_MPIAIJ,
2630:                                        MatIncreaseOverlap_MPIAIJ,
2631:                                        MatGetValues_MPIAIJ,
2632:                                        MatCopy_MPIAIJ,
2633:                                 /*44*/ MatGetRowMax_MPIAIJ,
2634:                                        MatScale_MPIAIJ,
2635:                                        MatShift_MPIAIJ,
2636:                                        MatDiagonalSet_MPIAIJ,
2637:                                        MatZeroRowsColumns_MPIAIJ,
2638:                                 /*49*/ MatSetRandom_MPIAIJ,
2639:                                        0,
2640:                                        0,
2641:                                        0,
2642:                                        0,
2643:                                 /*54*/ MatFDColoringCreate_MPIXAIJ,
2644:                                        0,
2645:                                        MatSetUnfactored_MPIAIJ,
2646:                                        MatPermute_MPIAIJ,
2647:                                        0,
2648:                                 /*59*/ MatCreateSubMatrix_MPIAIJ,
2649:                                        MatDestroy_MPIAIJ,
2650:                                        MatView_MPIAIJ,
2651:                                        0,
2652:                                        0,
2653:                                 /*64*/ 0,
2654:                                        MatMatMatMultNumeric_MPIAIJ_MPIAIJ_MPIAIJ,
2655:                                        0,
2656:                                        0,
2657:                                        0,
2658:                                 /*69*/ MatGetRowMaxAbs_MPIAIJ,
2659:                                        MatGetRowMinAbs_MPIAIJ,
2660:                                        0,
2661:                                        0,
2662:                                        0,
2663:                                        0,
2664:                                 /*75*/ MatFDColoringApply_AIJ,
2665:                                        MatSetFromOptions_MPIAIJ,
2666:                                        0,
2667:                                        0,
2668:                                        MatFindZeroDiagonals_MPIAIJ,
2669:                                 /*80*/ 0,
2670:                                        0,
2671:                                        0,
2672:                                 /*83*/ MatLoad_MPIAIJ,
2673:                                        MatIsSymmetric_MPIAIJ,
2674:                                        0,
2675:                                        0,
2676:                                        0,
2677:                                        0,
2678:                                 /*89*/ 0,
2679:                                        0,
2680:                                        MatMatMultNumeric_MPIAIJ_MPIAIJ,
2681:                                        0,
2682:                                        0,
2683:                                 /*94*/ MatPtAPNumeric_MPIAIJ_MPIAIJ,
2684:                                        0,
2685:                                        0,
2686:                                        0,
2687:                                        MatBindToCPU_MPIAIJ,
2688:                                 /*99*/ MatProductSetFromOptions_MPIAIJ,
2689:                                        0,
2690:                                        0,
2691:                                        MatConjugate_MPIAIJ,
2692:                                        0,
2693:                                 /*104*/MatSetValuesRow_MPIAIJ,
2694:                                        MatRealPart_MPIAIJ,
2695:                                        MatImaginaryPart_MPIAIJ,
2696:                                        0,
2697:                                        0,
2698:                                 /*109*/0,
2699:                                        0,
2700:                                        MatGetRowMin_MPIAIJ,
2701:                                        0,
2702:                                        MatMissingDiagonal_MPIAIJ,
2703:                                 /*114*/MatGetSeqNonzeroStructure_MPIAIJ,
2704:                                        0,
2705:                                        MatGetGhosts_MPIAIJ,
2706:                                        0,
2707:                                        0,
2708:                                 /*119*/0,
2709:                                        0,
2710:                                        0,
2711:                                        0,
2712:                                        MatGetMultiProcBlock_MPIAIJ,
2713:                                 /*124*/MatFindNonzeroRows_MPIAIJ,
2714:                                        MatGetColumnNorms_MPIAIJ,
2715:                                        MatInvertBlockDiagonal_MPIAIJ,
2716:                                        MatInvertVariableBlockDiagonal_MPIAIJ,
2717:                                        MatCreateSubMatricesMPI_MPIAIJ,
2718:                                 /*129*/0,
2719:                                        0,
2720:                                        0,
2721:                                        MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ,
2722:                                        0,
2723:                                 /*134*/0,
2724:                                        0,
2725:                                        0,
2726:                                        0,
2727:                                        0,
2728:                                 /*139*/MatSetBlockSizes_MPIAIJ,
2729:                                        0,
2730:                                        0,
2731:                                        MatFDColoringSetUp_MPIXAIJ,
2732:                                        MatFindOffBlockDiagonalEntries_MPIAIJ,
2733:                                        MatCreateMPIMatConcatenateSeqMat_MPIAIJ,
2734:                                 /*145*/0,
2735:                                        0,
2736:                                        0
2737: };

2739: /* ----------------------------------------------------------------------------------------*/

2741: PetscErrorCode  MatStoreValues_MPIAIJ(Mat mat)
2742: {
2743:   Mat_MPIAIJ     *aij = (Mat_MPIAIJ*)mat->data;

2747:   MatStoreValues(aij->A);
2748:   MatStoreValues(aij->B);
2749:   return(0);
2750: }

2752: PetscErrorCode  MatRetrieveValues_MPIAIJ(Mat mat)
2753: {
2754:   Mat_MPIAIJ     *aij = (Mat_MPIAIJ*)mat->data;

2758:   MatRetrieveValues(aij->A);
2759:   MatRetrieveValues(aij->B);
2760:   return(0);
2761: }

2763: PetscErrorCode  MatMPIAIJSetPreallocation_MPIAIJ(Mat B,PetscInt d_nz,const PetscInt d_nnz[],PetscInt o_nz,const PetscInt o_nnz[])
2764: {
2765:   Mat_MPIAIJ     *b;
2767:   PetscMPIInt    size;

2770:   PetscLayoutSetUp(B->rmap);
2771:   PetscLayoutSetUp(B->cmap);
2772:   b = (Mat_MPIAIJ*)B->data;

2774: #if defined(PETSC_USE_CTABLE)
2775:   PetscTableDestroy(&b->colmap);
2776: #else
2777:   PetscFree(b->colmap);
2778: #endif
2779:   PetscFree(b->garray);
2780:   VecDestroy(&b->lvec);
2781:   VecScatterDestroy(&b->Mvctx);

2783:   /* Because the B will have been resized we simply destroy it and create a new one each time */
2784:   MPI_Comm_size(PetscObjectComm((PetscObject)B),&size);
2785:   MatDestroy(&b->B);
2786:   MatCreate(PETSC_COMM_SELF,&b->B);
2787:   MatSetSizes(b->B,B->rmap->n,size > 1 ? B->cmap->N : 0,B->rmap->n,size > 1 ? B->cmap->N : 0);
2788:   MatSetBlockSizesFromMats(b->B,B,B);
2789:   MatSetType(b->B,MATSEQAIJ);
2790:   PetscLogObjectParent((PetscObject)B,(PetscObject)b->B);

2792:   if (!B->preallocated) {
2793:     MatCreate(PETSC_COMM_SELF,&b->A);
2794:     MatSetSizes(b->A,B->rmap->n,B->cmap->n,B->rmap->n,B->cmap->n);
2795:     MatSetBlockSizesFromMats(b->A,B,B);
2796:     MatSetType(b->A,MATSEQAIJ);
2797:     PetscLogObjectParent((PetscObject)B,(PetscObject)b->A);
2798:   }

2800:   MatSeqAIJSetPreallocation(b->A,d_nz,d_nnz);
2801:   MatSeqAIJSetPreallocation(b->B,o_nz,o_nnz);
2802:   B->preallocated  = PETSC_TRUE;
2803:   B->was_assembled = PETSC_FALSE;
2804:   B->assembled     = PETSC_FALSE;
2805:   return(0);
2806: }

2808: PetscErrorCode MatResetPreallocation_MPIAIJ(Mat B)
2809: {
2810:   Mat_MPIAIJ     *b;

2815:   PetscLayoutSetUp(B->rmap);
2816:   PetscLayoutSetUp(B->cmap);
2817:   b = (Mat_MPIAIJ*)B->data;

2819: #if defined(PETSC_USE_CTABLE)
2820:   PetscTableDestroy(&b->colmap);
2821: #else
2822:   PetscFree(b->colmap);
2823: #endif
2824:   PetscFree(b->garray);
2825:   VecDestroy(&b->lvec);
2826:   VecScatterDestroy(&b->Mvctx);

2828:   MatResetPreallocation(b->A);
2829:   MatResetPreallocation(b->B);
2830:   B->preallocated  = PETSC_TRUE;
2831:   B->was_assembled = PETSC_FALSE;
2832:   B->assembled = PETSC_FALSE;
2833:   return(0);
2834: }

2836: PetscErrorCode MatDuplicate_MPIAIJ(Mat matin,MatDuplicateOption cpvalues,Mat *newmat)
2837: {
2838:   Mat            mat;
2839:   Mat_MPIAIJ     *a,*oldmat = (Mat_MPIAIJ*)matin->data;

2843:   *newmat = 0;
2844:   MatCreate(PetscObjectComm((PetscObject)matin),&mat);
2845:   MatSetSizes(mat,matin->rmap->n,matin->cmap->n,matin->rmap->N,matin->cmap->N);
2846:   MatSetBlockSizesFromMats(mat,matin,matin);
2847:   MatSetType(mat,((PetscObject)matin)->type_name);
2848:   a       = (Mat_MPIAIJ*)mat->data;

2850:   mat->factortype   = matin->factortype;
2851:   mat->assembled    = matin->assembled;
2852:   mat->insertmode   = NOT_SET_VALUES;
2853:   mat->preallocated = matin->preallocated;

2855:   a->size         = oldmat->size;
2856:   a->rank         = oldmat->rank;
2857:   a->donotstash   = oldmat->donotstash;
2858:   a->roworiented  = oldmat->roworiented;
2859:   a->rowindices   = NULL;
2860:   a->rowvalues    = NULL;
2861:   a->getrowactive = PETSC_FALSE;

2863:   PetscLayoutReference(matin->rmap,&mat->rmap);
2864:   PetscLayoutReference(matin->cmap,&mat->cmap);

2866:   if (oldmat->colmap) {
2867: #if defined(PETSC_USE_CTABLE)
2868:     PetscTableCreateCopy(oldmat->colmap,&a->colmap);
2869: #else
2870:     PetscMalloc1(mat->cmap->N,&a->colmap);
2871:     PetscLogObjectMemory((PetscObject)mat,(mat->cmap->N)*sizeof(PetscInt));
2872:     PetscArraycpy(a->colmap,oldmat->colmap,mat->cmap->N);
2873: #endif
2874:   } else a->colmap = NULL;
2875:   if (oldmat->garray) {
2876:     PetscInt len;
2877:     len  = oldmat->B->cmap->n;
2878:     PetscMalloc1(len+1,&a->garray);
2879:     PetscLogObjectMemory((PetscObject)mat,len*sizeof(PetscInt));
2880:     if (len) { PetscArraycpy(a->garray,oldmat->garray,len); }
2881:   } else a->garray = NULL;

2883:   /* It may happen MatDuplicate is called with a non-assembled matrix
2884:      In fact, MatDuplicate only requires the matrix to be preallocated
2885:      This may happen inside a DMCreateMatrix_Shell */
2886:   if (oldmat->lvec) {
2887:     VecDuplicate(oldmat->lvec,&a->lvec);
2888:     PetscLogObjectParent((PetscObject)mat,(PetscObject)a->lvec);
2889:   }
2890:   if (oldmat->Mvctx) {
2891:     VecScatterCopy(oldmat->Mvctx,&a->Mvctx);
2892:     PetscLogObjectParent((PetscObject)mat,(PetscObject)a->Mvctx);
2893:   }
2894:   if (oldmat->Mvctx_mpi1) {
2895:     VecScatterCopy(oldmat->Mvctx_mpi1,&a->Mvctx_mpi1);
2896:     PetscLogObjectParent((PetscObject)mat,(PetscObject)a->Mvctx_mpi1);
2897:   }

2899:   MatDuplicate(oldmat->A,cpvalues,&a->A);
2900:   PetscLogObjectParent((PetscObject)mat,(PetscObject)a->A);
2901:   MatDuplicate(oldmat->B,cpvalues,&a->B);
2902:   PetscLogObjectParent((PetscObject)mat,(PetscObject)a->B);
2903:   PetscFunctionListDuplicate(((PetscObject)matin)->qlist,&((PetscObject)mat)->qlist);
2904:   *newmat = mat;
2905:   return(0);
2906: }

2908: PetscErrorCode MatLoad_MPIAIJ(Mat newMat, PetscViewer viewer)
2909: {
2910:   PetscBool      isbinary, ishdf5;

2916:   /* force binary viewer to load .info file if it has not yet done so */
2917:   PetscViewerSetUp(viewer);
2918:   PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERBINARY,&isbinary);
2919:   PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERHDF5,  &ishdf5);
2920:   if (isbinary) {
2921:     MatLoad_MPIAIJ_Binary(newMat,viewer);
2922:   } else if (ishdf5) {
2923: #if defined(PETSC_HAVE_HDF5)
2924:     MatLoad_AIJ_HDF5(newMat,viewer);
2925: #else
2926:     SETERRQ(PetscObjectComm((PetscObject)newMat),PETSC_ERR_SUP,"HDF5 not supported in this build.\nPlease reconfigure using --download-hdf5");
2927: #endif
2928:   } else {
2929:     SETERRQ2(PetscObjectComm((PetscObject)newMat),PETSC_ERR_SUP,"Viewer type %s not yet supported for reading %s matrices",((PetscObject)viewer)->type_name,((PetscObject)newMat)->type_name);
2930:   }
2931:   return(0);
2932: }

2934: PetscErrorCode MatLoad_MPIAIJ_Binary(Mat mat, PetscViewer viewer)
2935: {
2936:   PetscInt       header[4],M,N,m,nz,rows,cols,sum,i;
2937:   PetscInt       *rowidxs,*colidxs;
2938:   PetscScalar    *matvals;

2942:   PetscViewerSetUp(viewer);

2944:   /* read in matrix header */
2945:   PetscViewerBinaryRead(viewer,header,4,NULL,PETSC_INT);
2946:   if (header[0] != MAT_FILE_CLASSID) SETERRQ(PetscObjectComm((PetscObject)viewer),PETSC_ERR_FILE_UNEXPECTED,"Not a matrix object in file");
2947:   M  = header[1]; N = header[2]; nz = header[3];
2948:   if (M < 0) SETERRQ1(PetscObjectComm((PetscObject)viewer),PETSC_ERR_FILE_UNEXPECTED,"Matrix row size (%D) in file is negative",M);
2949:   if (N < 0) SETERRQ1(PetscObjectComm((PetscObject)viewer),PETSC_ERR_FILE_UNEXPECTED,"Matrix column size (%D) in file is negative",N);
2950:   if (nz < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED,"Matrix stored in special format on disk, cannot load as MPIAIJ");

2952:   /* set block sizes from the viewer's .info file */
2953:   MatLoad_Binary_BlockSizes(mat,viewer);
2954:   /* set global sizes if not set already */
2955:   if (mat->rmap->N < 0) mat->rmap->N = M;
2956:   if (mat->cmap->N < 0) mat->cmap->N = N;
2957:   PetscLayoutSetUp(mat->rmap);
2958:   PetscLayoutSetUp(mat->cmap);

2960:   /* check if the matrix sizes are correct */
2961:   MatGetSize(mat,&rows,&cols);
2962:   if (M != rows || N != cols) SETERRQ4(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED, "Matrix in file of different sizes (%D, %D) than the input matrix (%D, %D)",M,N,rows,cols);

2964:   /* read in row lengths and build row indices */
2965:   MatGetLocalSize(mat,&m,NULL);
2966:   PetscMalloc1(m+1,&rowidxs);
2967:   PetscViewerBinaryReadAll(viewer,rowidxs+1,m,PETSC_DECIDE,M,PETSC_INT);
2968:   rowidxs[0] = 0; for (i=0; i<m; i++) rowidxs[i+1] += rowidxs[i];
2969:   MPIU_Allreduce(&rowidxs[m],&sum,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)viewer));
2970:   if (sum != nz) SETERRQ2(PetscObjectComm((PetscObject)viewer),PETSC_ERR_FILE_UNEXPECTED,"Inconsistent matrix data in file: nonzeros = %D, sum-row-lengths = %D\n",nz,sum);
2971:   /* read in column indices and matrix values */
2972:   PetscMalloc2(rowidxs[m],&colidxs,rowidxs[m],&matvals);
2973:   PetscViewerBinaryReadAll(viewer,colidxs,rowidxs[m],PETSC_DETERMINE,PETSC_DETERMINE,PETSC_INT);
2974:   PetscViewerBinaryReadAll(viewer,matvals,rowidxs[m],PETSC_DETERMINE,PETSC_DETERMINE,PETSC_SCALAR);
2975:   /* store matrix indices and values */
2976:   MatMPIAIJSetPreallocationCSR(mat,rowidxs,colidxs,matvals);
2977:   PetscFree(rowidxs);
2978:   PetscFree2(colidxs,matvals);
2979:   return(0);
2980: }

2982: /* Not scalable because of ISAllGather() unless getting all columns. */
2983: PetscErrorCode ISGetSeqIS_Private(Mat mat,IS iscol,IS *isseq)
2984: {
2986:   IS             iscol_local;
2987:   PetscBool      isstride;
2988:   PetscMPIInt    lisstride=0,gisstride;

2991:   /* check if we are grabbing all columns*/
2992:   PetscObjectTypeCompare((PetscObject)iscol,ISSTRIDE,&isstride);

2994:   if (isstride) {
2995:     PetscInt  start,len,mstart,mlen;
2996:     ISStrideGetInfo(iscol,&start,NULL);
2997:     ISGetLocalSize(iscol,&len);
2998:     MatGetOwnershipRangeColumn(mat,&mstart,&mlen);
2999:     if (mstart == start && mlen-mstart == len) lisstride = 1;
3000:   }

3002:   MPIU_Allreduce(&lisstride,&gisstride,1,MPI_INT,MPI_MIN,PetscObjectComm((PetscObject)mat));
3003:   if (gisstride) {
3004:     PetscInt N;
3005:     MatGetSize(mat,NULL,&N);
3006:     ISCreateStride(PETSC_COMM_SELF,N,0,1,&iscol_local);
3007:     ISSetIdentity(iscol_local);
3008:     PetscInfo(mat,"Optimizing for obtaining all columns of the matrix; skipping ISAllGather()\n");
3009:   } else {
3010:     PetscInt cbs;
3011:     ISGetBlockSize(iscol,&cbs);
3012:     ISAllGather(iscol,&iscol_local);
3013:     ISSetBlockSize(iscol_local,cbs);
3014:   }

3016:   *isseq = iscol_local;
3017:   return(0);
3018: }

3020: /*
3021:  Used by MatCreateSubMatrix_MPIAIJ_SameRowColDist() to avoid ISAllGather() and global size of iscol_local
3022:  (see MatCreateSubMatrix_MPIAIJ_nonscalable)

3024:  Input Parameters:
3025:    mat - matrix
3026:    isrow - parallel row index set; its local indices are a subset of local columns of mat,
3027:            i.e., mat->rstart <= isrow[i] < mat->rend
3028:    iscol - parallel column index set; its local indices are a subset of local columns of mat,
3029:            i.e., mat->cstart <= iscol[i] < mat->cend
3030:  Output Parameter:
3031:    isrow_d,iscol_d - sequential row and column index sets for retrieving mat->A
3032:    iscol_o - sequential column index set for retrieving mat->B
3033:    garray - column map; garray[i] indicates global location of iscol_o[i] in iscol
3034:  */
3035: PetscErrorCode ISGetSeqIS_SameColDist_Private(Mat mat,IS isrow,IS iscol,IS *isrow_d,IS *iscol_d,IS *iscol_o,const PetscInt *garray[])
3036: {
3038:   Vec            x,cmap;
3039:   const PetscInt *is_idx;
3040:   PetscScalar    *xarray,*cmaparray;
3041:   PetscInt       ncols,isstart,*idx,m,rstart,*cmap1,count;
3042:   Mat_MPIAIJ     *a=(Mat_MPIAIJ*)mat->data;
3043:   Mat            B=a->B;
3044:   Vec            lvec=a->lvec,lcmap;
3045:   PetscInt       i,cstart,cend,Bn=B->cmap->N;
3046:   MPI_Comm       comm;
3047:   VecScatter     Mvctx=a->Mvctx;

3050:   PetscObjectGetComm((PetscObject)mat,&comm);
3051:   ISGetLocalSize(iscol,&ncols);

3053:   /* (1) iscol is a sub-column vector of mat, pad it with '-1.' to form a full vector x */
3054:   MatCreateVecs(mat,&x,NULL);
3055:   VecSet(x,-1.0);
3056:   VecDuplicate(x,&cmap);
3057:   VecSet(cmap,-1.0);

3059:   /* Get start indices */
3060:   MPI_Scan(&ncols,&isstart,1,MPIU_INT,MPI_SUM,comm);
3061:   isstart -= ncols;
3062:   MatGetOwnershipRangeColumn(mat,&cstart,&cend);

3064:   ISGetIndices(iscol,&is_idx);
3065:   VecGetArray(x,&xarray);
3066:   VecGetArray(cmap,&cmaparray);
3067:   PetscMalloc1(ncols,&idx);
3068:   for (i=0; i<ncols; i++) {
3069:     xarray[is_idx[i]-cstart]    = (PetscScalar)is_idx[i];
3070:     cmaparray[is_idx[i]-cstart] = i + isstart;      /* global index of iscol[i] */
3071:     idx[i]                      = is_idx[i]-cstart; /* local index of iscol[i]  */
3072:   }
3073:   VecRestoreArray(x,&xarray);
3074:   VecRestoreArray(cmap,&cmaparray);
3075:   ISRestoreIndices(iscol,&is_idx);

3077:   /* Get iscol_d */
3078:   ISCreateGeneral(PETSC_COMM_SELF,ncols,idx,PETSC_OWN_POINTER,iscol_d);
3079:   ISGetBlockSize(iscol,&i);
3080:   ISSetBlockSize(*iscol_d,i);

3082:   /* Get isrow_d */
3083:   ISGetLocalSize(isrow,&m);
3084:   rstart = mat->rmap->rstart;
3085:   PetscMalloc1(m,&idx);
3086:   ISGetIndices(isrow,&is_idx);
3087:   for (i=0; i<m; i++) idx[i] = is_idx[i]-rstart;
3088:   ISRestoreIndices(isrow,&is_idx);

3090:   ISCreateGeneral(PETSC_COMM_SELF,m,idx,PETSC_OWN_POINTER,isrow_d);
3091:   ISGetBlockSize(isrow,&i);
3092:   ISSetBlockSize(*isrow_d,i);

3094:   /* (2) Scatter x and cmap using aij->Mvctx to get their off-process portions (see MatMult_MPIAIJ) */
3095:   VecScatterBegin(Mvctx,x,lvec,INSERT_VALUES,SCATTER_FORWARD);
3096:   VecScatterEnd(Mvctx,x,lvec,INSERT_VALUES,SCATTER_FORWARD);

3098:   VecDuplicate(lvec,&lcmap);

3100:   VecScatterBegin(Mvctx,cmap,lcmap,INSERT_VALUES,SCATTER_FORWARD);
3101:   VecScatterEnd(Mvctx,cmap,lcmap,INSERT_VALUES,SCATTER_FORWARD);

3103:   /* (3) create sequential iscol_o (a subset of iscol) and isgarray */
3104:   /* off-process column indices */
3105:   count = 0;
3106:   PetscMalloc1(Bn,&idx);
3107:   PetscMalloc1(Bn,&cmap1);

3109:   VecGetArray(lvec,&xarray);
3110:   VecGetArray(lcmap,&cmaparray);
3111:   for (i=0; i<Bn; i++) {
3112:     if (PetscRealPart(xarray[i]) > -1.0) {
3113:       idx[count]     = i;                   /* local column index in off-diagonal part B */
3114:       cmap1[count] = (PetscInt)PetscRealPart(cmaparray[i]);  /* column index in submat */
3115:       count++;
3116:     }
3117:   }
3118:   VecRestoreArray(lvec,&xarray);
3119:   VecRestoreArray(lcmap,&cmaparray);

3121:   ISCreateGeneral(PETSC_COMM_SELF,count,idx,PETSC_COPY_VALUES,iscol_o);
3122:   /* cannot ensure iscol_o has same blocksize as iscol! */

3124:   PetscFree(idx);
3125:   *garray = cmap1;

3127:   VecDestroy(&x);
3128:   VecDestroy(&cmap);
3129:   VecDestroy(&lcmap);
3130:   return(0);
3131: }

3133: /* isrow and iscol have same processor distribution as mat, output *submat is a submatrix of local mat */
3134: PetscErrorCode MatCreateSubMatrix_MPIAIJ_SameRowColDist(Mat mat,IS isrow,IS iscol,MatReuse call,Mat *submat)
3135: {
3137:   Mat_MPIAIJ     *a = (Mat_MPIAIJ*)mat->data,*asub;
3138:   Mat            M = NULL;
3139:   MPI_Comm       comm;
3140:   IS             iscol_d,isrow_d,iscol_o;
3141:   Mat            Asub = NULL,Bsub = NULL;
3142:   PetscInt       n;

3145:   PetscObjectGetComm((PetscObject)mat,&comm);

3147:   if (call == MAT_REUSE_MATRIX) {
3148:     /* Retrieve isrow_d, iscol_d and iscol_o from submat */
3149:     PetscObjectQuery((PetscObject)*submat,"isrow_d",(PetscObject*)&isrow_d);
3150:     if (!isrow_d) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"isrow_d passed in was not used before, cannot reuse");

3152:     PetscObjectQuery((PetscObject)*submat,"iscol_d",(PetscObject*)&iscol_d);
3153:     if (!iscol_d) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"iscol_d passed in was not used before, cannot reuse");

3155:     PetscObjectQuery((PetscObject)*submat,"iscol_o",(PetscObject*)&iscol_o);
3156:     if (!iscol_o) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"iscol_o passed in was not used before, cannot reuse");

3158:     /* Update diagonal and off-diagonal portions of submat */
3159:     asub = (Mat_MPIAIJ*)(*submat)->data;
3160:     MatCreateSubMatrix_SeqAIJ(a->A,isrow_d,iscol_d,PETSC_DECIDE,MAT_REUSE_MATRIX,&asub->A);
3161:     ISGetLocalSize(iscol_o,&n);
3162:     if (n) {
3163:       MatCreateSubMatrix_SeqAIJ(a->B,isrow_d,iscol_o,PETSC_DECIDE,MAT_REUSE_MATRIX,&asub->B);
3164:     }
3165:     MatAssemblyBegin(*submat,MAT_FINAL_ASSEMBLY);
3166:     MatAssemblyEnd(*submat,MAT_FINAL_ASSEMBLY);

3168:   } else { /* call == MAT_INITIAL_MATRIX) */
3169:     const PetscInt *garray;
3170:     PetscInt        BsubN;

3172:     /* Create isrow_d, iscol_d, iscol_o and isgarray (replace isgarray with array?) */
3173:     ISGetSeqIS_SameColDist_Private(mat,isrow,iscol,&isrow_d,&iscol_d,&iscol_o,&garray);

3175:     /* Create local submatrices Asub and Bsub */
3176:     MatCreateSubMatrix_SeqAIJ(a->A,isrow_d,iscol_d,PETSC_DECIDE,MAT_INITIAL_MATRIX,&Asub);
3177:     MatCreateSubMatrix_SeqAIJ(a->B,isrow_d,iscol_o,PETSC_DECIDE,MAT_INITIAL_MATRIX,&Bsub);

3179:     /* Create submatrix M */
3180:     MatCreateMPIAIJWithSeqAIJ(comm,Asub,Bsub,garray,&M);

3182:     /* If Bsub has empty columns, compress iscol_o such that it will retrieve condensed Bsub from a->B during reuse */
3183:     asub = (Mat_MPIAIJ*)M->data;

3185:     ISGetLocalSize(iscol_o,&BsubN);
3186:     n = asub->B->cmap->N;
3187:     if (BsubN > n) {
3188:       /* This case can be tested using ~petsc/src/tao/bound/tutorials/runplate2_3 */
3189:       const PetscInt *idx;
3190:       PetscInt       i,j,*idx_new,*subgarray = asub->garray;
3191:       PetscInfo2(M,"submatrix Bn %D != BsubN %D, update iscol_o\n",n,BsubN);

3193:       PetscMalloc1(n,&idx_new);
3194:       j = 0;
3195:       ISGetIndices(iscol_o,&idx);
3196:       for (i=0; i<n; i++) {
3197:         if (j >= BsubN) break;
3198:         while (subgarray[i] > garray[j]) j++;

3200:         if (subgarray[i] == garray[j]) {
3201:           idx_new[i] = idx[j++];
3202:         } else SETERRQ4(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"subgarray[%D]=%D cannot < garray[%D]=%D",i,subgarray[i],j,garray[j]);
3203:       }
3204:       ISRestoreIndices(iscol_o,&idx);

3206:       ISDestroy(&iscol_o);
3207:       ISCreateGeneral(PETSC_COMM_SELF,n,idx_new,PETSC_OWN_POINTER,&iscol_o);

3209:     } else if (BsubN < n) {
3210:       SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Columns of Bsub cannot be smaller than B's",BsubN,asub->B->cmap->N);
3211:     }

3213:     PetscFree(garray);
3214:     *submat = M;

3216:     /* Save isrow_d, iscol_d and iscol_o used in processor for next request */
3217:     PetscObjectCompose((PetscObject)M,"isrow_d",(PetscObject)isrow_d);
3218:     ISDestroy(&isrow_d);

3220:     PetscObjectCompose((PetscObject)M,"iscol_d",(PetscObject)iscol_d);
3221:     ISDestroy(&iscol_d);

3223:     PetscObjectCompose((PetscObject)M,"iscol_o",(PetscObject)iscol_o);
3224:     ISDestroy(&iscol_o);
3225:   }
3226:   return(0);
3227: }

3229: PetscErrorCode MatCreateSubMatrix_MPIAIJ(Mat mat,IS isrow,IS iscol,MatReuse call,Mat *newmat)
3230: {
3232:   IS             iscol_local=NULL,isrow_d;
3233:   PetscInt       csize;
3234:   PetscInt       n,i,j,start,end;
3235:   PetscBool      sameRowDist=PETSC_FALSE,sameDist[2],tsameDist[2];
3236:   MPI_Comm       comm;

3239:   /* If isrow has same processor distribution as mat,
3240:      call MatCreateSubMatrix_MPIAIJ_SameRowDist() to avoid using a hash table with global size of iscol */
3241:   if (call == MAT_REUSE_MATRIX) {
3242:     PetscObjectQuery((PetscObject)*newmat,"isrow_d",(PetscObject*)&isrow_d);
3243:     if (isrow_d) {
3244:       sameRowDist  = PETSC_TRUE;
3245:       tsameDist[1] = PETSC_TRUE; /* sameColDist */
3246:     } else {
3247:       PetscObjectQuery((PetscObject)*newmat,"SubIScol",(PetscObject*)&iscol_local);
3248:       if (iscol_local) {
3249:         sameRowDist  = PETSC_TRUE;
3250:         tsameDist[1] = PETSC_FALSE; /* !sameColDist */
3251:       }
3252:     }
3253:   } else {
3254:     /* Check if isrow has same processor distribution as mat */
3255:     sameDist[0] = PETSC_FALSE;
3256:     ISGetLocalSize(isrow,&n);
3257:     if (!n) {
3258:       sameDist[0] = PETSC_TRUE;
3259:     } else {
3260:       ISGetMinMax(isrow,&i,&j);
3261:       MatGetOwnershipRange(mat,&start,&end);
3262:       if (i >= start && j < end) {
3263:         sameDist[0] = PETSC_TRUE;
3264:       }
3265:     }

3267:     /* Check if iscol has same processor distribution as mat */
3268:     sameDist[1] = PETSC_FALSE;
3269:     ISGetLocalSize(iscol,&n);
3270:     if (!n) {
3271:       sameDist[1] = PETSC_TRUE;
3272:     } else {
3273:       ISGetMinMax(iscol,&i,&j);
3274:       MatGetOwnershipRangeColumn(mat,&start,&end);
3275:       if (i >= start && j < end) sameDist[1] = PETSC_TRUE;
3276:     }

3278:     PetscObjectGetComm((PetscObject)mat,&comm);
3279:     MPIU_Allreduce(&sameDist,&tsameDist,2,MPIU_BOOL,MPI_LAND,comm);
3280:     sameRowDist = tsameDist[0];
3281:   }

3283:   if (sameRowDist) {
3284:     if (tsameDist[1]) { /* sameRowDist & sameColDist */
3285:       /* isrow and iscol have same processor distribution as mat */
3286:       MatCreateSubMatrix_MPIAIJ_SameRowColDist(mat,isrow,iscol,call,newmat);
3287:       return(0);
3288:     } else { /* sameRowDist */
3289:       /* isrow has same processor distribution as mat */
3290:       if (call == MAT_INITIAL_MATRIX) {
3291:         PetscBool sorted;
3292:         ISGetSeqIS_Private(mat,iscol,&iscol_local);
3293:         ISGetLocalSize(iscol_local,&n); /* local size of iscol_local = global columns of newmat */
3294:         ISGetSize(iscol,&i);
3295:         if (n != i) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"n %d != size of iscol %d",n,i);

3297:         ISSorted(iscol_local,&sorted);
3298:         if (sorted) {
3299:           /* MatCreateSubMatrix_MPIAIJ_SameRowDist() requires iscol_local be sorted; it can have duplicate indices */
3300:           MatCreateSubMatrix_MPIAIJ_SameRowDist(mat,isrow,iscol,iscol_local,MAT_INITIAL_MATRIX,newmat);
3301:           return(0);
3302:         }
3303:       } else { /* call == MAT_REUSE_MATRIX */
3304:         IS    iscol_sub;
3305:         PetscObjectQuery((PetscObject)*newmat,"SubIScol",(PetscObject*)&iscol_sub);
3306:         if (iscol_sub) {
3307:           MatCreateSubMatrix_MPIAIJ_SameRowDist(mat,isrow,iscol,NULL,call,newmat);
3308:           return(0);
3309:         }
3310:       }
3311:     }
3312:   }

3314:   /* General case: iscol -> iscol_local which has global size of iscol */
3315:   if (call == MAT_REUSE_MATRIX) {
3316:     PetscObjectQuery((PetscObject)*newmat,"ISAllGather",(PetscObject*)&iscol_local);
3317:     if (!iscol_local) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Submatrix passed in was not used before, cannot reuse");
3318:   } else {
3319:     if (!iscol_local) {
3320:       ISGetSeqIS_Private(mat,iscol,&iscol_local);
3321:     }
3322:   }

3324:   ISGetLocalSize(iscol,&csize);
3325:   MatCreateSubMatrix_MPIAIJ_nonscalable(mat,isrow,iscol_local,csize,call,newmat);

3327:   if (call == MAT_INITIAL_MATRIX) {
3328:     PetscObjectCompose((PetscObject)*newmat,"ISAllGather",(PetscObject)iscol_local);
3329:     ISDestroy(&iscol_local);
3330:   }
3331:   return(0);
3332: }

3334: /*@C
3335:      MatCreateMPIAIJWithSeqAIJ - creates a MPIAIJ matrix using SeqAIJ matrices that contain the "diagonal"
3336:          and "off-diagonal" part of the matrix in CSR format.

3338:    Collective

3340:    Input Parameters:
3341: +  comm - MPI communicator
3342: .  A - "diagonal" portion of matrix
3343: .  B - "off-diagonal" portion of matrix, may have empty columns, will be destroyed by this routine
3344: -  garray - global index of B columns

3346:    Output Parameter:
3347: .   mat - the matrix, with input A as its local diagonal matrix
3348:    Level: advanced

3350:    Notes:
3351:        See MatCreateAIJ() for the definition of "diagonal" and "off-diagonal" portion of the matrix.
3352:        A becomes part of output mat, B is destroyed by this routine. The user cannot use A and B anymore.

3354: .seealso: MatCreateMPIAIJWithSplitArrays()
3355: @*/
3356: PetscErrorCode MatCreateMPIAIJWithSeqAIJ(MPI_Comm comm,Mat A,Mat B,const PetscInt garray[],Mat *mat)
3357: {
3359:   Mat_MPIAIJ     *maij;
3360:   Mat_SeqAIJ     *b=(Mat_SeqAIJ*)B->data,*bnew;
3361:   PetscInt       *oi=b->i,*oj=b->j,i,nz,col;
3362:   PetscScalar    *oa=b->a;
3363:   Mat            Bnew;
3364:   PetscInt       m,n,N;

3367:   MatCreate(comm,mat);
3368:   MatGetSize(A,&m,&n);
3369:   if (m != B->rmap->N) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Am %D != Bm %D",m,B->rmap->N);
3370:   if (A->rmap->bs != B->rmap->bs) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"A row bs %D != B row bs %D",A->rmap->bs,B->rmap->bs);
3371:   /* remove check below; When B is created using iscol_o from ISGetSeqIS_SameColDist_Private(), its bs may not be same as A */
3372:   /* if (A->cmap->bs != B->cmap->bs) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"A column bs %D != B column bs %D",A->cmap->bs,B->cmap->bs); */

3374:   /* Get global columns of mat */
3375:   MPIU_Allreduce(&n,&N,1,MPIU_INT,MPI_SUM,comm);

3377:   MatSetSizes(*mat,m,n,PETSC_DECIDE,N);
3378:   MatSetType(*mat,MATMPIAIJ);
3379:   MatSetBlockSizes(*mat,A->rmap->bs,A->cmap->bs);
3380:   maij = (Mat_MPIAIJ*)(*mat)->data;

3382:   (*mat)->preallocated = PETSC_TRUE;

3384:   PetscLayoutSetUp((*mat)->rmap);
3385:   PetscLayoutSetUp((*mat)->cmap);

3387:   /* Set A as diagonal portion of *mat */
3388:   maij->A = A;

3390:   nz = oi[m];
3391:   for (i=0; i<nz; i++) {
3392:     col   = oj[i];
3393:     oj[i] = garray[col];
3394:   }

3396:    /* Set Bnew as off-diagonal portion of *mat */
3397:   MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,m,N,oi,oj,oa,&Bnew);
3398:   bnew        = (Mat_SeqAIJ*)Bnew->data;
3399:   bnew->maxnz = b->maxnz; /* allocated nonzeros of B */
3400:   maij->B     = Bnew;

3402:   if (B->rmap->N != Bnew->rmap->N) SETERRQ2(PETSC_COMM_SELF,0,"BN %d != BnewN %d",B->rmap->N,Bnew->rmap->N);

3404:   b->singlemalloc = PETSC_FALSE; /* B arrays are shared by Bnew */
3405:   b->free_a       = PETSC_FALSE;
3406:   b->free_ij      = PETSC_FALSE;
3407:   MatDestroy(&B);

3409:   bnew->singlemalloc = PETSC_TRUE; /* arrays will be freed by MatDestroy(&Bnew) */
3410:   bnew->free_a       = PETSC_TRUE;
3411:   bnew->free_ij      = PETSC_TRUE;

3413:   /* condense columns of maij->B */
3414:   MatSetOption(*mat,MAT_NO_OFF_PROC_ENTRIES,PETSC_TRUE);
3415:   MatAssemblyBegin(*mat,MAT_FINAL_ASSEMBLY);
3416:   MatAssemblyEnd(*mat,MAT_FINAL_ASSEMBLY);
3417:   MatSetOption(*mat,MAT_NO_OFF_PROC_ENTRIES,PETSC_FALSE);
3418:   MatSetOption(*mat,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE);
3419:   return(0);
3420: }

3422: extern PetscErrorCode MatCreateSubMatrices_MPIAIJ_SingleIS_Local(Mat,PetscInt,const IS[],const IS[],MatReuse,PetscBool,Mat*);

3424: PetscErrorCode MatCreateSubMatrix_MPIAIJ_SameRowDist(Mat mat,IS isrow,IS iscol,IS iscol_local,MatReuse call,Mat *newmat)
3425: {
3427:   PetscInt       i,m,n,rstart,row,rend,nz,j,bs,cbs;
3428:   PetscInt       *ii,*jj,nlocal,*dlens,*olens,dlen,olen,jend,mglobal;
3429:   Mat_MPIAIJ     *a=(Mat_MPIAIJ*)mat->data;
3430:   Mat            M,Msub,B=a->B;
3431:   MatScalar      *aa;
3432:   Mat_SeqAIJ     *aij;
3433:   PetscInt       *garray = a->garray,*colsub,Ncols;
3434:   PetscInt       count,Bn=B->cmap->N,cstart=mat->cmap->rstart,cend=mat->cmap->rend;
3435:   IS             iscol_sub,iscmap;
3436:   const PetscInt *is_idx,*cmap;
3437:   PetscBool      allcolumns=PETSC_FALSE;
3438:   MPI_Comm       comm;

3441:   PetscObjectGetComm((PetscObject)mat,&comm);

3443:   if (call == MAT_REUSE_MATRIX) {
3444:     PetscObjectQuery((PetscObject)*newmat,"SubIScol",(PetscObject*)&iscol_sub);
3445:     if (!iscol_sub) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"SubIScol passed in was not used before, cannot reuse");
3446:     ISGetLocalSize(iscol_sub,&count);

3448:     PetscObjectQuery((PetscObject)*newmat,"Subcmap",(PetscObject*)&iscmap);
3449:     if (!iscmap) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Subcmap passed in was not used before, cannot reuse");

3451:     PetscObjectQuery((PetscObject)*newmat,"SubMatrix",(PetscObject*)&Msub);
3452:     if (!Msub) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Submatrix passed in was not used before, cannot reuse");

3454:     MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat,1,&isrow,&iscol_sub,MAT_REUSE_MATRIX,PETSC_FALSE,&Msub);

3456:   } else { /* call == MAT_INITIAL_MATRIX) */
3457:     PetscBool flg;

3459:     ISGetLocalSize(iscol,&n);
3460:     ISGetSize(iscol,&Ncols);

3462:     /* (1) iscol -> nonscalable iscol_local */
3463:     /* Check for special case: each processor gets entire matrix columns */
3464:     ISIdentity(iscol_local,&flg);
3465:     if (flg && n == mat->cmap->N) allcolumns = PETSC_TRUE;
3466:     MPIU_Allreduce(MPI_IN_PLACE,&allcolumns,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)mat));
3467:     if (allcolumns) {
3468:       iscol_sub = iscol_local;
3469:       PetscObjectReference((PetscObject)iscol_local);
3470:       ISCreateStride(PETSC_COMM_SELF,n,0,1,&iscmap);

3472:     } else {
3473:       /* (2) iscol_local -> iscol_sub and iscmap. Implementation below requires iscol_local be sorted, it can have duplicate indices */
3474:       PetscInt *idx,*cmap1,k;
3475:       PetscMalloc1(Ncols,&idx);
3476:       PetscMalloc1(Ncols,&cmap1);
3477:       ISGetIndices(iscol_local,&is_idx);
3478:       count = 0;
3479:       k     = 0;
3480:       for (i=0; i<Ncols; i++) {
3481:         j = is_idx[i];
3482:         if (j >= cstart && j < cend) {
3483:           /* diagonal part of mat */
3484:           idx[count]     = j;
3485:           cmap1[count++] = i; /* column index in submat */
3486:         } else if (Bn) {
3487:           /* off-diagonal part of mat */
3488:           if (j == garray[k]) {
3489:             idx[count]     = j;
3490:             cmap1[count++] = i;  /* column index in submat */
3491:           } else if (j > garray[k]) {
3492:             while (j > garray[k] && k < Bn-1) k++;
3493:             if (j == garray[k]) {
3494:               idx[count]     = j;
3495:               cmap1[count++] = i; /* column index in submat */
3496:             }
3497:           }
3498:         }
3499:       }
3500:       ISRestoreIndices(iscol_local,&is_idx);

3502:       ISCreateGeneral(PETSC_COMM_SELF,count,idx,PETSC_OWN_POINTER,&iscol_sub);
3503:       ISGetBlockSize(iscol,&cbs);
3504:       ISSetBlockSize(iscol_sub,cbs);

3506:       ISCreateGeneral(PetscObjectComm((PetscObject)iscol_local),count,cmap1,PETSC_OWN_POINTER,&iscmap);
3507:     }

3509:     /* (3) Create sequential Msub */
3510:     MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat,1,&isrow,&iscol_sub,MAT_INITIAL_MATRIX,allcolumns,&Msub);
3511:   }

3513:   ISGetLocalSize(iscol_sub,&count);
3514:   aij  = (Mat_SeqAIJ*)(Msub)->data;
3515:   ii   = aij->i;
3516:   ISGetIndices(iscmap,&cmap);

3518:   /*
3519:       m - number of local rows
3520:       Ncols - number of columns (same on all processors)
3521:       rstart - first row in new global matrix generated
3522:   */
3523:   MatGetSize(Msub,&m,NULL);

3525:   if (call == MAT_INITIAL_MATRIX) {
3526:     /* (4) Create parallel newmat */
3527:     PetscMPIInt    rank,size;
3528:     PetscInt       csize;

3530:     MPI_Comm_size(comm,&size);
3531:     MPI_Comm_rank(comm,&rank);

3533:     /*
3534:         Determine the number of non-zeros in the diagonal and off-diagonal
3535:         portions of the matrix in order to do correct preallocation
3536:     */

3538:     /* first get start and end of "diagonal" columns */
3539:     ISGetLocalSize(iscol,&csize);
3540:     if (csize == PETSC_DECIDE) {
3541:       ISGetSize(isrow,&mglobal);
3542:       if (mglobal == Ncols) { /* square matrix */
3543:         nlocal = m;
3544:       } else {
3545:         nlocal = Ncols/size + ((Ncols % size) > rank);
3546:       }
3547:     } else {
3548:       nlocal = csize;
3549:     }
3550:     MPI_Scan(&nlocal,&rend,1,MPIU_INT,MPI_SUM,comm);
3551:     rstart = rend - nlocal;
3552:     if (rank == size - 1 && rend != Ncols) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Local column sizes %D do not add up to total number of columns %D",rend,Ncols);

3554:     /* next, compute all the lengths */
3555:     jj    = aij->j;
3556:     PetscMalloc1(2*m+1,&dlens);
3557:     olens = dlens + m;
3558:     for (i=0; i<m; i++) {
3559:       jend = ii[i+1] - ii[i];
3560:       olen = 0;
3561:       dlen = 0;
3562:       for (j=0; j<jend; j++) {
3563:         if (cmap[*jj] < rstart || cmap[*jj] >= rend) olen++;
3564:         else dlen++;
3565:         jj++;
3566:       }
3567:       olens[i] = olen;
3568:       dlens[i] = dlen;
3569:     }

3571:     ISGetBlockSize(isrow,&bs);
3572:     ISGetBlockSize(iscol,&cbs);

3574:     MatCreate(comm,&M);
3575:     MatSetSizes(M,m,nlocal,PETSC_DECIDE,Ncols);
3576:     MatSetBlockSizes(M,bs,cbs);
3577:     MatSetType(M,((PetscObject)mat)->type_name);
3578:     MatMPIAIJSetPreallocation(M,0,dlens,0,olens);
3579:     PetscFree(dlens);

3581:   } else { /* call == MAT_REUSE_MATRIX */
3582:     M    = *newmat;
3583:     MatGetLocalSize(M,&i,NULL);
3584:     if (i != m) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Previous matrix must be same size/layout as request");
3585:     MatZeroEntries(M);
3586:     /*
3587:          The next two lines are needed so we may call MatSetValues_MPIAIJ() below directly,
3588:        rather than the slower MatSetValues().
3589:     */
3590:     M->was_assembled = PETSC_TRUE;
3591:     M->assembled     = PETSC_FALSE;
3592:   }

3594:   /* (5) Set values of Msub to *newmat */
3595:   PetscMalloc1(count,&colsub);
3596:   MatGetOwnershipRange(M,&rstart,NULL);

3598:   jj   = aij->j;
3599:   aa   = aij->a;
3600:   for (i=0; i<m; i++) {
3601:     row = rstart + i;
3602:     nz  = ii[i+1] - ii[i];
3603:     for (j=0; j<nz; j++) colsub[j] = cmap[jj[j]];
3604:     MatSetValues_MPIAIJ(M,1,&row,nz,colsub,aa,INSERT_VALUES);
3605:     jj += nz; aa += nz;
3606:   }
3607:   ISRestoreIndices(iscmap,&cmap);

3609:   MatAssemblyBegin(M,MAT_FINAL_ASSEMBLY);
3610:   MatAssemblyEnd(M,MAT_FINAL_ASSEMBLY);

3612:   PetscFree(colsub);

3614:   /* save Msub, iscol_sub and iscmap used in processor for next request */
3615:   if (call ==  MAT_INITIAL_MATRIX) {
3616:     *newmat = M;
3617:     PetscObjectCompose((PetscObject)(*newmat),"SubMatrix",(PetscObject)Msub);
3618:     MatDestroy(&Msub);

3620:     PetscObjectCompose((PetscObject)(*newmat),"SubIScol",(PetscObject)iscol_sub);
3621:     ISDestroy(&iscol_sub);

3623:     PetscObjectCompose((PetscObject)(*newmat),"Subcmap",(PetscObject)iscmap);
3624:     ISDestroy(&iscmap);

3626:     if (iscol_local) {
3627:       PetscObjectCompose((PetscObject)(*newmat),"ISAllGather",(PetscObject)iscol_local);
3628:       ISDestroy(&iscol_local);
3629:     }
3630:   }
3631:   return(0);
3632: }

3634: /*
3635:     Not great since it makes two copies of the submatrix, first an SeqAIJ
3636:   in local and then by concatenating the local matrices the end result.
3637:   Writing it directly would be much like MatCreateSubMatrices_MPIAIJ()

3639:   Note: This requires a sequential iscol with all indices.
3640: */
3641: PetscErrorCode MatCreateSubMatrix_MPIAIJ_nonscalable(Mat mat,IS isrow,IS iscol,PetscInt csize,MatReuse call,Mat *newmat)
3642: {
3644:   PetscMPIInt    rank,size;
3645:   PetscInt       i,m,n,rstart,row,rend,nz,*cwork,j,bs,cbs;
3646:   PetscInt       *ii,*jj,nlocal,*dlens,*olens,dlen,olen,jend,mglobal;
3647:   Mat            M,Mreuse;
3648:   MatScalar      *aa,*vwork;
3649:   MPI_Comm       comm;
3650:   Mat_SeqAIJ     *aij;
3651:   PetscBool      colflag,allcolumns=PETSC_FALSE;

3654:   PetscObjectGetComm((PetscObject)mat,&comm);
3655:   MPI_Comm_rank(comm,&rank);
3656:   MPI_Comm_size(comm,&size);

3658:   /* Check for special case: each processor gets entire matrix columns */
3659:   ISIdentity(iscol,&colflag);
3660:   ISGetLocalSize(iscol,&n);
3661:   if (colflag && n == mat->cmap->N) allcolumns = PETSC_TRUE;
3662:   MPIU_Allreduce(MPI_IN_PLACE,&allcolumns,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)mat));

3664:   if (call ==  MAT_REUSE_MATRIX) {
3665:     PetscObjectQuery((PetscObject)*newmat,"SubMatrix",(PetscObject*)&Mreuse);
3666:     if (!Mreuse) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Submatrix passed in was not used before, cannot reuse");
3667:     MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat,1,&isrow,&iscol,MAT_REUSE_MATRIX,allcolumns,&Mreuse);
3668:   } else {
3669:     MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat,1,&isrow,&iscol,MAT_INITIAL_MATRIX,allcolumns,&Mreuse);
3670:   }

3672:   /*
3673:       m - number of local rows
3674:       n - number of columns (same on all processors)
3675:       rstart - first row in new global matrix generated
3676:   */
3677:   MatGetSize(Mreuse,&m,&n);
3678:   MatGetBlockSizes(Mreuse,&bs,&cbs);
3679:   if (call == MAT_INITIAL_MATRIX) {
3680:     aij = (Mat_SeqAIJ*)(Mreuse)->data;
3681:     ii  = aij->i;
3682:     jj  = aij->j;

3684:     /*
3685:         Determine the number of non-zeros in the diagonal and off-diagonal
3686:         portions of the matrix in order to do correct preallocation
3687:     */

3689:     /* first get start and end of "diagonal" columns */
3690:     if (csize == PETSC_DECIDE) {
3691:       ISGetSize(isrow,&mglobal);
3692:       if (mglobal == n) { /* square matrix */
3693:         nlocal = m;
3694:       } else {
3695:         nlocal = n/size + ((n % size) > rank);
3696:       }
3697:     } else {
3698:       nlocal = csize;
3699:     }
3700:     MPI_Scan(&nlocal,&rend,1,MPIU_INT,MPI_SUM,comm);
3701:     rstart = rend - nlocal;
3702:     if (rank == size - 1 && rend != n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Local column sizes %D do not add up to total number of columns %D",rend,n);

3704:     /* next, compute all the lengths */
3705:     PetscMalloc1(2*m+1,&dlens);
3706:     olens = dlens + m;
3707:     for (i=0; i<m; i++) {
3708:       jend = ii[i+1] - ii[i];
3709:       olen = 0;
3710:       dlen = 0;
3711:       for (j=0; j<jend; j++) {
3712:         if (*jj < rstart || *jj >= rend) olen++;
3713:         else dlen++;
3714:         jj++;
3715:       }
3716:       olens[i] = olen;
3717:       dlens[i] = dlen;
3718:     }
3719:     MatCreate(comm,&M);
3720:     MatSetSizes(M,m,nlocal,PETSC_DECIDE,n);
3721:     MatSetBlockSizes(M,bs,cbs);
3722:     MatSetType(M,((PetscObject)mat)->type_name);
3723:     MatMPIAIJSetPreallocation(M,0,dlens,0,olens);
3724:     PetscFree(dlens);
3725:   } else {
3726:     PetscInt ml,nl;

3728:     M    = *newmat;
3729:     MatGetLocalSize(M,&ml,&nl);
3730:     if (ml != m) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Previous matrix must be same size/layout as request");
3731:     MatZeroEntries(M);
3732:     /*
3733:          The next two lines are needed so we may call MatSetValues_MPIAIJ() below directly,
3734:        rather than the slower MatSetValues().
3735:     */
3736:     M->was_assembled = PETSC_TRUE;
3737:     M->assembled     = PETSC_FALSE;
3738:   }
3739:   MatGetOwnershipRange(M,&rstart,&rend);
3740:   aij  = (Mat_SeqAIJ*)(Mreuse)->data;
3741:   ii   = aij->i;
3742:   jj   = aij->j;
3743:   aa   = aij->a;
3744:   for (i=0; i<m; i++) {
3745:     row   = rstart + i;
3746:     nz    = ii[i+1] - ii[i];
3747:     cwork = jj;     jj += nz;
3748:     vwork = aa;     aa += nz;
3749:     MatSetValues_MPIAIJ(M,1,&row,nz,cwork,vwork,INSERT_VALUES);
3750:   }

3752:   MatAssemblyBegin(M,MAT_FINAL_ASSEMBLY);
3753:   MatAssemblyEnd(M,MAT_FINAL_ASSEMBLY);
3754:   *newmat = M;

3756:   /* save submatrix used in processor for next request */
3757:   if (call ==  MAT_INITIAL_MATRIX) {
3758:     PetscObjectCompose((PetscObject)M,"SubMatrix",(PetscObject)Mreuse);
3759:     MatDestroy(&Mreuse);
3760:   }
3761:   return(0);
3762: }

3764: PetscErrorCode MatMPIAIJSetPreallocationCSR_MPIAIJ(Mat B,const PetscInt Ii[],const PetscInt J[],const PetscScalar v[])
3765: {
3766:   PetscInt       m,cstart, cend,j,nnz,i,d;
3767:   PetscInt       *d_nnz,*o_nnz,nnz_max = 0,rstart,ii;
3768:   const PetscInt *JJ;
3770:   PetscBool      nooffprocentries;

3773:   if (Ii[0]) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Ii[0] must be 0 it is %D",Ii[0]);

3775:   PetscLayoutSetUp(B->rmap);
3776:   PetscLayoutSetUp(B->cmap);
3777:   m      = B->rmap->n;
3778:   cstart = B->cmap->rstart;
3779:   cend   = B->cmap->rend;
3780:   rstart = B->rmap->rstart;

3782:   PetscCalloc2(m,&d_nnz,m,&o_nnz);

3784: #if defined(PETSC_USE_DEBUG)
3785:   for (i=0; i<m; i++) {
3786:     nnz = Ii[i+1]- Ii[i];
3787:     JJ  = J + Ii[i];
3788:     if (nnz < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Local row %D has a negative %D number of columns",i,nnz);
3789:     if (nnz && (JJ[0] < 0)) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Row %D starts with negative column index",i,JJ[0]);
3790:     if (nnz && (JJ[nnz-1] >= B->cmap->N)) SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Row %D ends with too large a column index %D (max allowed %D)",i,JJ[nnz-1],B->cmap->N);
3791:   }
3792: #endif

3794:   for (i=0; i<m; i++) {
3795:     nnz     = Ii[i+1]- Ii[i];
3796:     JJ      = J + Ii[i];
3797:     nnz_max = PetscMax(nnz_max,nnz);
3798:     d       = 0;
3799:     for (j=0; j<nnz; j++) {
3800:       if (cstart <= JJ[j] && JJ[j] < cend) d++;
3801:     }
3802:     d_nnz[i] = d;
3803:     o_nnz[i] = nnz - d;
3804:   }
3805:   MatMPIAIJSetPreallocation(B,0,d_nnz,0,o_nnz);
3806:   PetscFree2(d_nnz,o_nnz);

3808:   for (i=0; i<m; i++) {
3809:     ii   = i + rstart;
3810:     MatSetValues_MPIAIJ(B,1,&ii,Ii[i+1] - Ii[i],J+Ii[i], v ? v + Ii[i] : NULL,INSERT_VALUES);
3811:   }
3812:   nooffprocentries    = B->nooffprocentries;
3813:   B->nooffprocentries = PETSC_TRUE;
3814:   MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);
3815:   MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);
3816:   B->nooffprocentries = nooffprocentries;

3818:   MatSetOption(B,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE);
3819:   return(0);
3820: }

3822: /*@
3823:    MatMPIAIJSetPreallocationCSR - Allocates memory for a sparse parallel matrix in AIJ format
3824:    (the default parallel PETSc format).

3826:    Collective

3828:    Input Parameters:
3829: +  B - the matrix
3830: .  i - the indices into j for the start of each local row (starts with zero)
3831: .  j - the column indices for each local row (starts with zero)
3832: -  v - optional values in the matrix

3834:    Level: developer

3836:    Notes:
3837:        The i, j, and v arrays ARE copied by this routine into the internal format used by PETSc;
3838:      thus you CANNOT change the matrix entries by changing the values of v[] after you have
3839:      called this routine. Use MatCreateMPIAIJWithSplitArrays() to avoid needing to copy the arrays.

3841:        The i and j indices are 0 based, and i indices are indices corresponding to the local j array.

3843:        The format which is used for the sparse matrix input, is equivalent to a
3844:     row-major ordering.. i.e for the following matrix, the input data expected is
3845:     as shown

3847: $        1 0 0
3848: $        2 0 3     P0
3849: $       -------
3850: $        4 5 6     P1
3851: $
3852: $     Process0 [P0]: rows_owned=[0,1]
3853: $        i =  {0,1,3}  [size = nrow+1  = 2+1]
3854: $        j =  {0,0,2}  [size = 3]
3855: $        v =  {1,2,3}  [size = 3]
3856: $
3857: $     Process1 [P1]: rows_owned=[2]
3858: $        i =  {0,3}    [size = nrow+1  = 1+1]
3859: $        j =  {0,1,2}  [size = 3]
3860: $        v =  {4,5,6}  [size = 3]

3862: .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatCreateAIJ(), MATMPIAIJ,
3863:           MatCreateSeqAIJWithArrays(), MatCreateMPIAIJWithSplitArrays()
3864: @*/
3865: PetscErrorCode  MatMPIAIJSetPreallocationCSR(Mat B,const PetscInt i[],const PetscInt j[], const PetscScalar v[])
3866: {

3870:   PetscTryMethod(B,"MatMPIAIJSetPreallocationCSR_C",(Mat,const PetscInt[],const PetscInt[],const PetscScalar[]),(B,i,j,v));
3871:   return(0);
3872: }

3874: /*@C
3875:    MatMPIAIJSetPreallocation - Preallocates memory for a sparse parallel matrix in AIJ format
3876:    (the default parallel PETSc format).  For good matrix assembly performance
3877:    the user should preallocate the matrix storage by setting the parameters
3878:    d_nz (or d_nnz) and o_nz (or o_nnz).  By setting these parameters accurately,
3879:    performance can be increased by more than a factor of 50.

3881:    Collective

3883:    Input Parameters:
3884: +  B - the matrix
3885: .  d_nz  - number of nonzeros per row in DIAGONAL portion of local submatrix
3886:            (same value is used for all local rows)
3887: .  d_nnz - array containing the number of nonzeros in the various rows of the
3888:            DIAGONAL portion of the local submatrix (possibly different for each row)
3889:            or NULL (PETSC_NULL_INTEGER in Fortran), if d_nz is used to specify the nonzero structure.
3890:            The size of this array is equal to the number of local rows, i.e 'm'.
3891:            For matrices that will be factored, you must leave room for (and set)
3892:            the diagonal entry even if it is zero.
3893: .  o_nz  - number of nonzeros per row in the OFF-DIAGONAL portion of local
3894:            submatrix (same value is used for all local rows).
3895: -  o_nnz - array containing the number of nonzeros in the various rows of the
3896:            OFF-DIAGONAL portion of the local submatrix (possibly different for
3897:            each row) or NULL (PETSC_NULL_INTEGER in Fortran), if o_nz is used to specify the nonzero
3898:            structure. The size of this array is equal to the number
3899:            of local rows, i.e 'm'.

3901:    If the *_nnz parameter is given then the *_nz parameter is ignored

3903:    The AIJ format (also called the Yale sparse matrix format or
3904:    compressed row storage (CSR)), is fully compatible with standard Fortran 77
3905:    storage.  The stored row and column indices begin with zero.
3906:    See Users-Manual: ch_mat for details.

3908:    The parallel matrix is partitioned such that the first m0 rows belong to
3909:    process 0, the next m1 rows belong to process 1, the next m2 rows belong
3910:    to process 2 etc.. where m0,m1,m2... are the input parameter 'm'.

3912:    The DIAGONAL portion of the local submatrix of a processor can be defined
3913:    as the submatrix which is obtained by extraction the part corresponding to
3914:    the rows r1-r2 and columns c1-c2 of the global matrix, where r1 is the
3915:    first row that belongs to the processor, r2 is the last row belonging to
3916:    the this processor, and c1-c2 is range of indices of the local part of a
3917:    vector suitable for applying the matrix to.  This is an mxn matrix.  In the
3918:    common case of a square matrix, the row and column ranges are the same and
3919:    the DIAGONAL part is also square. The remaining portion of the local
3920:    submatrix (mxN) constitute the OFF-DIAGONAL portion.

3922:    If o_nnz, d_nnz are specified, then o_nz, and d_nz are ignored.

3924:    You can call MatGetInfo() to get information on how effective the preallocation was;
3925:    for example the fields mallocs,nz_allocated,nz_used,nz_unneeded;
3926:    You can also run with the option -info and look for messages with the string
3927:    malloc in them to see if additional memory allocation was needed.

3929:    Example usage:

3931:    Consider the following 8x8 matrix with 34 non-zero values, that is
3932:    assembled across 3 processors. Lets assume that proc0 owns 3 rows,
3933:    proc1 owns 3 rows, proc2 owns 2 rows. This division can be shown
3934:    as follows:

3936: .vb
3937:             1  2  0  |  0  3  0  |  0  4
3938:     Proc0   0  5  6  |  7  0  0  |  8  0
3939:             9  0 10  | 11  0  0  | 12  0
3940:     -------------------------------------
3941:            13  0 14  | 15 16 17  |  0  0
3942:     Proc1   0 18  0  | 19 20 21  |  0  0
3943:             0  0  0  | 22 23  0  | 24  0
3944:     -------------------------------------
3945:     Proc2  25 26 27  |  0  0 28  | 29  0
3946:            30  0  0  | 31 32 33  |  0 34
3947: .ve

3949:    This can be represented as a collection of submatrices as:

3951: .vb
3952:       A B C
3953:       D E F
3954:       G H I
3955: .ve

3957:    Where the submatrices A,B,C are owned by proc0, D,E,F are
3958:    owned by proc1, G,H,I are owned by proc2.

3960:    The 'm' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
3961:    The 'n' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
3962:    The 'M','N' parameters are 8,8, and have the same values on all procs.

3964:    The DIAGONAL submatrices corresponding to proc0,proc1,proc2 are
3965:    submatrices [A], [E], [I] respectively. The OFF-DIAGONAL submatrices
3966:    corresponding to proc0,proc1,proc2 are [BC], [DF], [GH] respectively.
3967:    Internally, each processor stores the DIAGONAL part, and the OFF-DIAGONAL
3968:    part as SeqAIJ matrices. for eg: proc1 will store [E] as a SeqAIJ
3969:    matrix, ans [DF] as another SeqAIJ matrix.

3971:    When d_nz, o_nz parameters are specified, d_nz storage elements are
3972:    allocated for every row of the local diagonal submatrix, and o_nz
3973:    storage locations are allocated for every row of the OFF-DIAGONAL submat.
3974:    One way to choose d_nz and o_nz is to use the max nonzerors per local
3975:    rows for each of the local DIAGONAL, and the OFF-DIAGONAL submatrices.
3976:    In this case, the values of d_nz,o_nz are:
3977: .vb
3978:      proc0 : dnz = 2, o_nz = 2
3979:      proc1 : dnz = 3, o_nz = 2
3980:      proc2 : dnz = 1, o_nz = 4
3981: .ve
3982:    We are allocating m*(d_nz+o_nz) storage locations for every proc. This
3983:    translates to 3*(2+2)=12 for proc0, 3*(3+2)=15 for proc1, 2*(1+4)=10
3984:    for proc3. i.e we are using 12+15+10=37 storage locations to store
3985:    34 values.

3987:    When d_nnz, o_nnz parameters are specified, the storage is specified
3988:    for every row, coresponding to both DIAGONAL and OFF-DIAGONAL submatrices.
3989:    In the above case the values for d_nnz,o_nnz are:
3990: .vb
3991:      proc0: d_nnz = [2,2,2] and o_nnz = [2,2,2]
3992:      proc1: d_nnz = [3,3,2] and o_nnz = [2,1,1]
3993:      proc2: d_nnz = [1,1]   and o_nnz = [4,4]
3994: .ve
3995:    Here the space allocated is sum of all the above values i.e 34, and
3996:    hence pre-allocation is perfect.

3998:    Level: intermediate

4000: .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatCreateAIJ(), MatMPIAIJSetPreallocationCSR(),
4001:           MATMPIAIJ, MatGetInfo(), PetscSplitOwnership()
4002: @*/
4003: PetscErrorCode MatMPIAIJSetPreallocation(Mat B,PetscInt d_nz,const PetscInt d_nnz[],PetscInt o_nz,const PetscInt o_nnz[])
4004: {

4010:   PetscTryMethod(B,"MatMPIAIJSetPreallocation_C",(Mat,PetscInt,const PetscInt[],PetscInt,const PetscInt[]),(B,d_nz,d_nnz,o_nz,o_nnz));
4011:   return(0);
4012: }

4014: /*@
4015:      MatCreateMPIAIJWithArrays - creates a MPI AIJ matrix using arrays that contain in standard
4016:          CSR format for the local rows.

4018:    Collective

4020:    Input Parameters:
4021: +  comm - MPI communicator
4022: .  m - number of local rows (Cannot be PETSC_DECIDE)
4023: .  n - This value should be the same as the local size used in creating the
4024:        x vector for the matrix-vector product y = Ax. (or PETSC_DECIDE to have
4025:        calculated if N is given) For square matrices n is almost always m.
4026: .  M - number of global rows (or PETSC_DETERMINE to have calculated if m is given)
4027: .  N - number of global columns (or PETSC_DETERMINE to have calculated if n is given)
4028: .   i - row indices; that is i[0] = 0, i[row] = i[row-1] + number of elements in that row of the matrix
4029: .   j - column indices
4030: -   a - matrix values

4032:    Output Parameter:
4033: .   mat - the matrix

4035:    Level: intermediate

4037:    Notes:
4038:        The i, j, and a arrays ARE copied by this routine into the internal format used by PETSc;
4039:      thus you CANNOT change the matrix entries by changing the values of a[] after you have
4040:      called this routine. Use MatCreateMPIAIJWithSplitArrays() to avoid needing to copy the arrays.

4042:        The i and j indices are 0 based, and i indices are indices corresponding to the local j array.

4044:        The format which is used for the sparse matrix input, is equivalent to a
4045:     row-major ordering.. i.e for the following matrix, the input data expected is
4046:     as shown

4048:        Once you have created the matrix you can update it with new numerical values using MatUpdateMPIAIJWithArrays

4050: $        1 0 0
4051: $        2 0 3     P0
4052: $       -------
4053: $        4 5 6     P1
4054: $
4055: $     Process0 [P0]: rows_owned=[0,1]
4056: $        i =  {0,1,3}  [size = nrow+1  = 2+1]
4057: $        j =  {0,0,2}  [size = 3]
4058: $        v =  {1,2,3}  [size = 3]
4059: $
4060: $     Process1 [P1]: rows_owned=[2]
4061: $        i =  {0,3}    [size = nrow+1  = 1+1]
4062: $        j =  {0,1,2}  [size = 3]
4063: $        v =  {4,5,6}  [size = 3]

4065: .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatMPIAIJSetPreallocationCSR(),
4066:           MATMPIAIJ, MatCreateAIJ(), MatCreateMPIAIJWithSplitArrays(), MatUpdateMPIAIJWithArrays()
4067: @*/
4068: PetscErrorCode MatCreateMPIAIJWithArrays(MPI_Comm comm,PetscInt m,PetscInt n,PetscInt M,PetscInt N,const PetscInt i[],const PetscInt j[],const PetscScalar a[],Mat *mat)
4069: {

4073:   if (i && i[0]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"i (row indices) must start with 0");
4074:   if (m < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"local number of rows (m) cannot be PETSC_DECIDE, or negative");
4075:   MatCreate(comm,mat);
4076:   MatSetSizes(*mat,m,n,M,N);
4077:   /* MatSetBlockSizes(M,bs,cbs); */
4078:   MatSetType(*mat,MATMPIAIJ);
4079:   MatMPIAIJSetPreallocationCSR(*mat,i,j,a);
4080:   return(0);
4081: }

4083: /*@
4084:      MatUpdateMPIAIJWithArrays - updates a MPI AIJ matrix using arrays that contain in standard
4085:          CSR format for the local rows. Only the numerical values are updated the other arrays must be identical

4087:    Collective

4089:    Input Parameters:
4090: +  mat - the matrix
4091: .  m - number of local rows (Cannot be PETSC_DECIDE)
4092: .  n - This value should be the same as the local size used in creating the
4093:        x vector for the matrix-vector product y = Ax. (or PETSC_DECIDE to have
4094:        calculated if N is given) For square matrices n is almost always m.
4095: .  M - number of global rows (or PETSC_DETERMINE to have calculated if m is given)
4096: .  N - number of global columns (or PETSC_DETERMINE to have calculated if n is given)
4097: .  Ii - row indices; that is Ii[0] = 0, Ii[row] = Ii[row-1] + number of elements in that row of the matrix
4098: .  J - column indices
4099: -  v - matrix values

4101:    Level: intermediate

4103: .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatMPIAIJSetPreallocationCSR(),
4104:           MATMPIAIJ, MatCreateAIJ(), MatCreateMPIAIJWithSplitArrays(), MatUpdateMPIAIJWithArrays()
4105: @*/
4106: PetscErrorCode MatUpdateMPIAIJWithArrays(Mat mat,PetscInt m,PetscInt n,PetscInt M,PetscInt N,const PetscInt Ii[],const PetscInt J[],const PetscScalar v[])
4107: {
4109:   PetscInt       cstart,nnz,i,j;
4110:   PetscInt       *ld;
4111:   PetscBool      nooffprocentries;
4112:   Mat_MPIAIJ     *Aij = (Mat_MPIAIJ*)mat->data;
4113:   Mat_SeqAIJ     *Ad  = (Mat_SeqAIJ*)Aij->A->data, *Ao  = (Mat_SeqAIJ*)Aij->B->data;
4114:   PetscScalar    *ad = Ad->a, *ao = Ao->a;
4115:   const PetscInt *Adi = Ad->i;
4116:   PetscInt       ldi,Iii,md;

4119:   if (Ii[0]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"i (row indices) must start with 0");
4120:   if (m < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"local number of rows (m) cannot be PETSC_DECIDE, or negative");
4121:   if (m != mat->rmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Local number of rows cannot change from call to MatUpdateMPIAIJWithArrays()");
4122:   if (n != mat->cmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Local number of columns cannot change from call to MatUpdateMPIAIJWithArrays()");

4124:   cstart = mat->cmap->rstart;
4125:   if (!Aij->ld) {
4126:     /* count number of entries below block diagonal */
4127:     PetscCalloc1(m,&ld);
4128:     Aij->ld = ld;
4129:     for (i=0; i<m; i++) {
4130:       nnz  = Ii[i+1]- Ii[i];
4131:       j     = 0;
4132:       while  (J[j] < cstart && j < nnz) {j++;}
4133:       J    += nnz;
4134:       ld[i] = j;
4135:     }
4136:   } else {
4137:     ld = Aij->ld;
4138:   }

4140:   for (i=0; i<m; i++) {
4141:     nnz  = Ii[i+1]- Ii[i];
4142:     Iii  = Ii[i];
4143:     ldi  = ld[i];
4144:     md   = Adi[i+1]-Adi[i];
4145:     PetscArraycpy(ao,v + Iii,ldi);
4146:     PetscArraycpy(ad,v + Iii + ldi,md);
4147:     PetscArraycpy(ao + ldi,v + Iii + ldi + md,nnz - ldi - md);
4148:     ad  += md;
4149:     ao  += nnz - md;
4150:   }
4151:   nooffprocentries      = mat->nooffprocentries;
4152:   mat->nooffprocentries = PETSC_TRUE;
4153:   PetscObjectStateIncrease((PetscObject)Aij->A);
4154:   PetscObjectStateIncrease((PetscObject)Aij->B);
4155:   PetscObjectStateIncrease((PetscObject)mat);
4156:   MatAssemblyBegin(mat,MAT_FINAL_ASSEMBLY);
4157:   MatAssemblyEnd(mat,MAT_FINAL_ASSEMBLY);
4158:   mat->nooffprocentries = nooffprocentries;
4159:   return(0);
4160: }

4162: /*@C
4163:    MatCreateAIJ - Creates a sparse parallel matrix in AIJ format
4164:    (the default parallel PETSc format).  For good matrix assembly performance
4165:    the user should preallocate the matrix storage by setting the parameters
4166:    d_nz (or d_nnz) and o_nz (or o_nnz).  By setting these parameters accurately,
4167:    performance can be increased by more than a factor of 50.

4169:    Collective

4171:    Input Parameters:
4172: +  comm - MPI communicator
4173: .  m - number of local rows (or PETSC_DECIDE to have calculated if M is given)
4174:            This value should be the same as the local size used in creating the
4175:            y vector for the matrix-vector product y = Ax.
4176: .  n - This value should be the same as the local size used in creating the
4177:        x vector for the matrix-vector product y = Ax. (or PETSC_DECIDE to have
4178:        calculated if N is given) For square matrices n is almost always m.
4179: .  M - number of global rows (or PETSC_DETERMINE to have calculated if m is given)
4180: .  N - number of global columns (or PETSC_DETERMINE to have calculated if n is given)
4181: .  d_nz  - number of nonzeros per row in DIAGONAL portion of local submatrix
4182:            (same value is used for all local rows)
4183: .  d_nnz - array containing the number of nonzeros in the various rows of the
4184:            DIAGONAL portion of the local submatrix (possibly different for each row)
4185:            or NULL, if d_nz is used to specify the nonzero structure.
4186:            The size of this array is equal to the number of local rows, i.e 'm'.
4187: .  o_nz  - number of nonzeros per row in the OFF-DIAGONAL portion of local
4188:            submatrix (same value is used for all local rows).
4189: -  o_nnz - array containing the number of nonzeros in the various rows of the
4190:            OFF-DIAGONAL portion of the local submatrix (possibly different for
4191:            each row) or NULL, if o_nz is used to specify the nonzero
4192:            structure. The size of this array is equal to the number
4193:            of local rows, i.e 'm'.

4195:    Output Parameter:
4196: .  A - the matrix

4198:    It is recommended that one use the MatCreate(), MatSetType() and/or MatSetFromOptions(),
4199:    MatXXXXSetPreallocation() paradigm instead of this routine directly.
4200:    [MatXXXXSetPreallocation() is, for example, MatSeqAIJSetPreallocation]

4202:    Notes:
4203:    If the *_nnz parameter is given then the *_nz parameter is ignored

4205:    m,n,M,N parameters specify the size of the matrix, and its partitioning across
4206:    processors, while d_nz,d_nnz,o_nz,o_nnz parameters specify the approximate
4207:    storage requirements for this matrix.

4209:    If PETSC_DECIDE or  PETSC_DETERMINE is used for a particular argument on one
4210:    processor than it must be used on all processors that share the object for
4211:    that argument.

4213:    The user MUST specify either the local or global matrix dimensions
4214:    (possibly both).

4216:    The parallel matrix is partitioned across processors such that the
4217:    first m0 rows belong to process 0, the next m1 rows belong to
4218:    process 1, the next m2 rows belong to process 2 etc.. where
4219:    m0,m1,m2,.. are the input parameter 'm'. i.e each processor stores
4220:    values corresponding to [m x N] submatrix.

4222:    The columns are logically partitioned with the n0 columns belonging
4223:    to 0th partition, the next n1 columns belonging to the next
4224:    partition etc.. where n0,n1,n2... are the input parameter 'n'.

4226:    The DIAGONAL portion of the local submatrix on any given processor
4227:    is the submatrix corresponding to the rows and columns m,n
4228:    corresponding to the given processor. i.e diagonal matrix on
4229:    process 0 is [m0 x n0], diagonal matrix on process 1 is [m1 x n1]
4230:    etc. The remaining portion of the local submatrix [m x (N-n)]
4231:    constitute the OFF-DIAGONAL portion. The example below better
4232:    illustrates this concept.

4234:    For a square global matrix we define each processor's diagonal portion
4235:    to be its local rows and the corresponding columns (a square submatrix);
4236:    each processor's off-diagonal portion encompasses the remainder of the
4237:    local matrix (a rectangular submatrix).

4239:    If o_nnz, d_nnz are specified, then o_nz, and d_nz are ignored.

4241:    When calling this routine with a single process communicator, a matrix of
4242:    type SEQAIJ is returned.  If a matrix of type MPIAIJ is desired for this
4243:    type of communicator, use the construction mechanism
4244: .vb
4245:      MatCreate(...,&A); MatSetType(A,MATMPIAIJ); MatSetSizes(A, m,n,M,N); MatMPIAIJSetPreallocation(A,...);
4246: .ve

4248: $     MatCreate(...,&A);
4249: $     MatSetType(A,MATMPIAIJ);
4250: $     MatSetSizes(A, m,n,M,N);
4251: $     MatMPIAIJSetPreallocation(A,...);

4253:    By default, this format uses inodes (identical nodes) when possible.
4254:    We search for consecutive rows with the same nonzero structure, thereby
4255:    reusing matrix information to achieve increased efficiency.

4257:    Options Database Keys:
4258: +  -mat_no_inode  - Do not use inodes
4259: -  -mat_inode_limit <limit> - Sets inode limit (max limit=5)



4263:    Example usage:

4265:    Consider the following 8x8 matrix with 34 non-zero values, that is
4266:    assembled across 3 processors. Lets assume that proc0 owns 3 rows,
4267:    proc1 owns 3 rows, proc2 owns 2 rows. This division can be shown
4268:    as follows

4270: .vb
4271:             1  2  0  |  0  3  0  |  0  4
4272:     Proc0   0  5  6  |  7  0  0  |  8  0
4273:             9  0 10  | 11  0  0  | 12  0
4274:     -------------------------------------
4275:            13  0 14  | 15 16 17  |  0  0
4276:     Proc1   0 18  0  | 19 20 21  |  0  0
4277:             0  0  0  | 22 23  0  | 24  0
4278:     -------------------------------------
4279:     Proc2  25 26 27  |  0  0 28  | 29  0
4280:            30  0  0  | 31 32 33  |  0 34
4281: .ve

4283:    This can be represented as a collection of submatrices as

4285: .vb
4286:       A B C
4287:       D E F
4288:       G H I
4289: .ve

4291:    Where the submatrices A,B,C are owned by proc0, D,E,F are
4292:    owned by proc1, G,H,I are owned by proc2.

4294:    The 'm' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
4295:    The 'n' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
4296:    The 'M','N' parameters are 8,8, and have the same values on all procs.

4298:    The DIAGONAL submatrices corresponding to proc0,proc1,proc2 are
4299:    submatrices [A], [E], [I] respectively. The OFF-DIAGONAL submatrices
4300:    corresponding to proc0,proc1,proc2 are [BC], [DF], [GH] respectively.
4301:    Internally, each processor stores the DIAGONAL part, and the OFF-DIAGONAL
4302:    part as SeqAIJ matrices. for eg: proc1 will store [E] as a SeqAIJ
4303:    matrix, ans [DF] as another SeqAIJ matrix.

4305:    When d_nz, o_nz parameters are specified, d_nz storage elements are
4306:    allocated for every row of the local diagonal submatrix, and o_nz
4307:    storage locations are allocated for every row of the OFF-DIAGONAL submat.
4308:    One way to choose d_nz and o_nz is to use the max nonzerors per local
4309:    rows for each of the local DIAGONAL, and the OFF-DIAGONAL submatrices.
4310:    In this case, the values of d_nz,o_nz are
4311: .vb
4312:      proc0 : dnz = 2, o_nz = 2
4313:      proc1 : dnz = 3, o_nz = 2
4314:      proc2 : dnz = 1, o_nz = 4
4315: .ve
4316:    We are allocating m*(d_nz+o_nz) storage locations for every proc. This
4317:    translates to 3*(2+2)=12 for proc0, 3*(3+2)=15 for proc1, 2*(1+4)=10
4318:    for proc3. i.e we are using 12+15+10=37 storage locations to store
4319:    34 values.

4321:    When d_nnz, o_nnz parameters are specified, the storage is specified
4322:    for every row, coresponding to both DIAGONAL and OFF-DIAGONAL submatrices.
4323:    In the above case the values for d_nnz,o_nnz are
4324: .vb
4325:      proc0: d_nnz = [2,2,2] and o_nnz = [2,2,2]
4326:      proc1: d_nnz = [3,3,2] and o_nnz = [2,1,1]
4327:      proc2: d_nnz = [1,1]   and o_nnz = [4,4]
4328: .ve
4329:    Here the space allocated is sum of all the above values i.e 34, and
4330:    hence pre-allocation is perfect.

4332:    Level: intermediate

4334: .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatMPIAIJSetPreallocationCSR(),
4335:           MATMPIAIJ, MatCreateMPIAIJWithArrays()
4336: @*/
4337: PetscErrorCode  MatCreateAIJ(MPI_Comm comm,PetscInt m,PetscInt n,PetscInt M,PetscInt N,PetscInt d_nz,const PetscInt d_nnz[],PetscInt o_nz,const PetscInt o_nnz[],Mat *A)
4338: {
4340:   PetscMPIInt    size;

4343:   MatCreate(comm,A);
4344:   MatSetSizes(*A,m,n,M,N);
4345:   MPI_Comm_size(comm,&size);
4346:   if (size > 1) {
4347:     MatSetType(*A,MATMPIAIJ);
4348:     MatMPIAIJSetPreallocation(*A,d_nz,d_nnz,o_nz,o_nnz);
4349:   } else {
4350:     MatSetType(*A,MATSEQAIJ);
4351:     MatSeqAIJSetPreallocation(*A,d_nz,d_nnz);
4352:   }
4353:   return(0);
4354: }

4356: /*@C
4357:   MatMPIAIJGetSeqAIJ - Returns the local piece of this distributed matrix
4358:   
4359:   Not collective
4360:   
4361:   Input Parameter:
4362: . A - The MPIAIJ matrix

4364:   Output Parameters:
4365: + Ad - The local diagonal block as a SeqAIJ matrix
4366: . Ao - The local off-diagonal block as a SeqAIJ matrix
4367: - colmap - An array mapping local column numbers of Ao to global column numbers of the parallel matrix

4369:   Note: The rows in Ad and Ao are in [0, Nr), where Nr is the number of local rows on this process. The columns
4370:   in Ad are in [0, Nc) where Nc is the number of local columns. The columns are Ao are in [0, Nco), where Nco is
4371:   the number of nonzero columns in the local off-diagonal piece of the matrix A. The array colmap maps these
4372:   local column numbers to global column numbers in the original matrix.

4374:   Level: intermediate

4376: .seealso: MatMPIAIJGetLocalMat(), MatMPIAIJGetLocalMatCondensed(), MatCreateAIJ(), MATMPIAIJ, MATSEQAIJ
4377: @*/
4378: PetscErrorCode MatMPIAIJGetSeqAIJ(Mat A,Mat *Ad,Mat *Ao,const PetscInt *colmap[])
4379: {
4380:   Mat_MPIAIJ     *a = (Mat_MPIAIJ*)A->data;
4381:   PetscBool      flg;

4385:   PetscStrbeginswith(((PetscObject)A)->type_name,MATMPIAIJ,&flg);
4386:   if (!flg) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"This function requires a MATMPIAIJ matrix as input");
4387:   if (Ad)     *Ad     = a->A;
4388:   if (Ao)     *Ao     = a->B;
4389:   if (colmap) *colmap = a->garray;
4390:   return(0);
4391: }

4393: PetscErrorCode MatCreateMPIMatConcatenateSeqMat_MPIAIJ(MPI_Comm comm,Mat inmat,PetscInt n,MatReuse scall,Mat *outmat)
4394: {
4396:   PetscInt       m,N,i,rstart,nnz,Ii;
4397:   PetscInt       *indx;
4398:   PetscScalar    *values;

4401:   MatGetSize(inmat,&m,&N);
4402:   if (scall == MAT_INITIAL_MATRIX) { /* symbolic phase */
4403:     PetscInt       *dnz,*onz,sum,bs,cbs;

4405:     if (n == PETSC_DECIDE) {
4406:       PetscSplitOwnership(comm,&n,&N);
4407:     }
4408:     /* Check sum(n) = N */
4409:     MPIU_Allreduce(&n,&sum,1,MPIU_INT,MPI_SUM,comm);
4410:     if (sum != N) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_INCOMP,"Sum of local columns %D != global columns %D",sum,N);

4412:     MPI_Scan(&m, &rstart,1,MPIU_INT,MPI_SUM,comm);
4413:     rstart -= m;

4415:     MatPreallocateInitialize(comm,m,n,dnz,onz);
4416:     for (i=0; i<m; i++) {
4417:       MatGetRow_SeqAIJ(inmat,i,&nnz,&indx,NULL);
4418:       MatPreallocateSet(i+rstart,nnz,indx,dnz,onz);
4419:       MatRestoreRow_SeqAIJ(inmat,i,&nnz,&indx,NULL);
4420:     }

4422:     MatCreate(comm,outmat);
4423:     MatSetSizes(*outmat,m,n,PETSC_DETERMINE,PETSC_DETERMINE);
4424:     MatGetBlockSizes(inmat,&bs,&cbs);
4425:     MatSetBlockSizes(*outmat,bs,cbs);
4426:     MatSetType(*outmat,MATAIJ);
4427:     MatSeqAIJSetPreallocation(*outmat,0,dnz);
4428:     MatMPIAIJSetPreallocation(*outmat,0,dnz,0,onz);
4429:     MatPreallocateFinalize(dnz,onz);
4430:   }

4432:   /* numeric phase */
4433:   MatGetOwnershipRange(*outmat,&rstart,NULL);
4434:   for (i=0; i<m; i++) {
4435:     MatGetRow_SeqAIJ(inmat,i,&nnz,&indx,&values);
4436:     Ii   = i + rstart;
4437:     MatSetValues(*outmat,1,&Ii,nnz,indx,values,INSERT_VALUES);
4438:     MatRestoreRow_SeqAIJ(inmat,i,&nnz,&indx,&values);
4439:   }
4440:   MatAssemblyBegin(*outmat,MAT_FINAL_ASSEMBLY);
4441:   MatAssemblyEnd(*outmat,MAT_FINAL_ASSEMBLY);
4442:   return(0);
4443: }

4445: PetscErrorCode MatFileSplit(Mat A,char *outfile)
4446: {
4447:   PetscErrorCode    ierr;
4448:   PetscMPIInt       rank;
4449:   PetscInt          m,N,i,rstart,nnz;
4450:   size_t            len;
4451:   const PetscInt    *indx;
4452:   PetscViewer       out;
4453:   char              *name;
4454:   Mat               B;
4455:   const PetscScalar *values;

4458:   MatGetLocalSize(A,&m,0);
4459:   MatGetSize(A,0,&N);
4460:   /* Should this be the type of the diagonal block of A? */
4461:   MatCreate(PETSC_COMM_SELF,&B);
4462:   MatSetSizes(B,m,N,m,N);
4463:   MatSetBlockSizesFromMats(B,A,A);
4464:   MatSetType(B,MATSEQAIJ);
4465:   MatSeqAIJSetPreallocation(B,0,NULL);
4466:   MatGetOwnershipRange(A,&rstart,0);
4467:   for (i=0; i<m; i++) {
4468:     MatGetRow(A,i+rstart,&nnz,&indx,&values);
4469:     MatSetValues(B,1,&i,nnz,indx,values,INSERT_VALUES);
4470:     MatRestoreRow(A,i+rstart,&nnz,&indx,&values);
4471:   }
4472:   MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);
4473:   MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);

4475:   MPI_Comm_rank(PetscObjectComm((PetscObject)A),&rank);
4476:   PetscStrlen(outfile,&len);
4477:   PetscMalloc1(len+5,&name);
4478:   sprintf(name,"%s.%d",outfile,rank);
4479:   PetscViewerBinaryOpen(PETSC_COMM_SELF,name,FILE_MODE_APPEND,&out);
4480:   PetscFree(name);
4481:   MatView(B,out);
4482:   PetscViewerDestroy(&out);
4483:   MatDestroy(&B);
4484:   return(0);
4485: }

4487: PetscErrorCode MatDestroy_MPIAIJ_SeqsToMPI(Mat A)
4488: {
4489:   PetscErrorCode      ierr;
4490:   Mat_Merge_SeqsToMPI *merge;
4491:   PetscContainer      container;

4494:   PetscObjectQuery((PetscObject)A,"MatMergeSeqsToMPI",(PetscObject*)&container);
4495:   if (container) {
4496:     PetscContainerGetPointer(container,(void**)&merge);
4497:     PetscFree(merge->id_r);
4498:     PetscFree(merge->len_s);
4499:     PetscFree(merge->len_r);
4500:     PetscFree(merge->bi);
4501:     PetscFree(merge->bj);
4502:     PetscFree(merge->buf_ri[0]);
4503:     PetscFree(merge->buf_ri);
4504:     PetscFree(merge->buf_rj[0]);
4505:     PetscFree(merge->buf_rj);
4506:     PetscFree(merge->coi);
4507:     PetscFree(merge->coj);
4508:     PetscFree(merge->owners_co);
4509:     PetscLayoutDestroy(&merge->rowmap);
4510:     PetscFree(merge);
4511:     PetscObjectCompose((PetscObject)A,"MatMergeSeqsToMPI",0);
4512:   }
4513:   MatDestroy_MPIAIJ(A);
4514:   return(0);
4515: }

4517:  #include <../src/mat/utils/freespace.h>
4518:  #include <petscbt.h>

4520: PetscErrorCode MatCreateMPIAIJSumSeqAIJNumeric(Mat seqmat,Mat mpimat)
4521: {
4522:   PetscErrorCode      ierr;
4523:   MPI_Comm            comm;
4524:   Mat_SeqAIJ          *a  =(Mat_SeqAIJ*)seqmat->data;
4525:   PetscMPIInt         size,rank,taga,*len_s;
4526:   PetscInt            N=mpimat->cmap->N,i,j,*owners,*ai=a->i,*aj;
4527:   PetscInt            proc,m;
4528:   PetscInt            **buf_ri,**buf_rj;
4529:   PetscInt            k,anzi,*bj_i,*bi,*bj,arow,bnzi,nextaj;
4530:   PetscInt            nrows,**buf_ri_k,**nextrow,**nextai;
4531:   MPI_Request         *s_waits,*r_waits;
4532:   MPI_Status          *status;
4533:   MatScalar           *aa=a->a;
4534:   MatScalar           **abuf_r,*ba_i;
4535:   Mat_Merge_SeqsToMPI *merge;
4536:   PetscContainer      container;

4539:   PetscObjectGetComm((PetscObject)mpimat,&comm);
4540:   PetscLogEventBegin(MAT_Seqstompinum,seqmat,0,0,0);

4542:   MPI_Comm_size(comm,&size);
4543:   MPI_Comm_rank(comm,&rank);

4545:   PetscObjectQuery((PetscObject)mpimat,"MatMergeSeqsToMPI",(PetscObject*)&container);
4546:   PetscContainerGetPointer(container,(void**)&merge);

4548:   bi     = merge->bi;
4549:   bj     = merge->bj;
4550:   buf_ri = merge->buf_ri;
4551:   buf_rj = merge->buf_rj;

4553:   PetscMalloc1(size,&status);
4554:   owners = merge->rowmap->range;
4555:   len_s  = merge->len_s;

4557:   /* send and recv matrix values */
4558:   /*-----------------------------*/
4559:   PetscObjectGetNewTag((PetscObject)mpimat,&taga);
4560:   PetscPostIrecvScalar(comm,taga,merge->nrecv,merge->id_r,merge->len_r,&abuf_r,&r_waits);

4562:   PetscMalloc1(merge->nsend+1,&s_waits);
4563:   for (proc=0,k=0; proc<size; proc++) {
4564:     if (!len_s[proc]) continue;
4565:     i    = owners[proc];
4566:     MPI_Isend(aa+ai[i],len_s[proc],MPIU_MATSCALAR,proc,taga,comm,s_waits+k);
4567:     k++;
4568:   }

4570:   if (merge->nrecv) {MPI_Waitall(merge->nrecv,r_waits,status);}
4571:   if (merge->nsend) {MPI_Waitall(merge->nsend,s_waits,status);}
4572:   PetscFree(status);

4574:   PetscFree(s_waits);
4575:   PetscFree(r_waits);

4577:   /* insert mat values of mpimat */
4578:   /*----------------------------*/
4579:   PetscMalloc1(N,&ba_i);
4580:   PetscMalloc3(merge->nrecv,&buf_ri_k,merge->nrecv,&nextrow,merge->nrecv,&nextai);

4582:   for (k=0; k<merge->nrecv; k++) {
4583:     buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */
4584:     nrows       = *(buf_ri_k[k]);
4585:     nextrow[k]  = buf_ri_k[k]+1;  /* next row number of k-th recved i-structure */
4586:     nextai[k]   = buf_ri_k[k] + (nrows + 1); /* poins to the next i-structure of k-th recved i-structure  */
4587:   }

4589:   /* set values of ba */
4590:   m = merge->rowmap->n;
4591:   for (i=0; i<m; i++) {
4592:     arow = owners[rank] + i;
4593:     bj_i = bj+bi[i];  /* col indices of the i-th row of mpimat */
4594:     bnzi = bi[i+1] - bi[i];
4595:     PetscArrayzero(ba_i,bnzi);

4597:     /* add local non-zero vals of this proc's seqmat into ba */
4598:     anzi   = ai[arow+1] - ai[arow];
4599:     aj     = a->j + ai[arow];
4600:     aa     = a->a + ai[arow];
4601:     nextaj = 0;
4602:     for (j=0; nextaj<anzi; j++) {
4603:       if (*(bj_i + j) == aj[nextaj]) { /* bcol == acol */
4604:         ba_i[j] += aa[nextaj++];
4605:       }
4606:     }

4608:     /* add received vals into ba */
4609:     for (k=0; k<merge->nrecv; k++) { /* k-th received message */
4610:       /* i-th row */
4611:       if (i == *nextrow[k]) {
4612:         anzi   = *(nextai[k]+1) - *nextai[k];
4613:         aj     = buf_rj[k] + *(nextai[k]);
4614:         aa     = abuf_r[k] + *(nextai[k]);
4615:         nextaj = 0;
4616:         for (j=0; nextaj<anzi; j++) {
4617:           if (*(bj_i + j) == aj[nextaj]) { /* bcol == acol */
4618:             ba_i[j] += aa[nextaj++];
4619:           }
4620:         }
4621:         nextrow[k]++; nextai[k]++;
4622:       }
4623:     }
4624:     MatSetValues(mpimat,1,&arow,bnzi,bj_i,ba_i,INSERT_VALUES);
4625:   }
4626:   MatAssemblyBegin(mpimat,MAT_FINAL_ASSEMBLY);
4627:   MatAssemblyEnd(mpimat,MAT_FINAL_ASSEMBLY);

4629:   PetscFree(abuf_r[0]);
4630:   PetscFree(abuf_r);
4631:   PetscFree(ba_i);
4632:   PetscFree3(buf_ri_k,nextrow,nextai);
4633:   PetscLogEventEnd(MAT_Seqstompinum,seqmat,0,0,0);
4634:   return(0);
4635: }

4637: PetscErrorCode  MatCreateMPIAIJSumSeqAIJSymbolic(MPI_Comm comm,Mat seqmat,PetscInt m,PetscInt n,Mat *mpimat)
4638: {
4639:   PetscErrorCode      ierr;
4640:   Mat                 B_mpi;
4641:   Mat_SeqAIJ          *a=(Mat_SeqAIJ*)seqmat->data;
4642:   PetscMPIInt         size,rank,tagi,tagj,*len_s,*len_si,*len_ri;
4643:   PetscInt            **buf_rj,**buf_ri,**buf_ri_k;
4644:   PetscInt            M=seqmat->rmap->n,N=seqmat->cmap->n,i,*owners,*ai=a->i,*aj=a->j;
4645:   PetscInt            len,proc,*dnz,*onz,bs,cbs;
4646:   PetscInt            k,anzi,*bi,*bj,*lnk,nlnk,arow,bnzi,nspacedouble=0;
4647:   PetscInt            nrows,*buf_s,*buf_si,*buf_si_i,**nextrow,**nextai;
4648:   MPI_Request         *si_waits,*sj_waits,*ri_waits,*rj_waits;
4649:   MPI_Status          *status;
4650:   PetscFreeSpaceList  free_space=NULL,current_space=NULL;
4651:   PetscBT             lnkbt;
4652:   Mat_Merge_SeqsToMPI *merge;
4653:   PetscContainer      container;

4656:   PetscLogEventBegin(MAT_Seqstompisym,seqmat,0,0,0);

4658:   /* make sure it is a PETSc comm */
4659:   PetscCommDuplicate(comm,&comm,NULL);
4660:   MPI_Comm_size(comm,&size);
4661:   MPI_Comm_rank(comm,&rank);

4663:   PetscNew(&merge);
4664:   PetscMalloc1(size,&status);

4666:   /* determine row ownership */
4667:   /*---------------------------------------------------------*/
4668:   PetscLayoutCreate(comm,&merge->rowmap);
4669:   PetscLayoutSetLocalSize(merge->rowmap,m);
4670:   PetscLayoutSetSize(merge->rowmap,M);
4671:   PetscLayoutSetBlockSize(merge->rowmap,1);
4672:   PetscLayoutSetUp(merge->rowmap);
4673:   PetscMalloc1(size,&len_si);
4674:   PetscMalloc1(size,&merge->len_s);

4676:   m      = merge->rowmap->n;
4677:   owners = merge->rowmap->range;

4679:   /* determine the number of messages to send, their lengths */
4680:   /*---------------------------------------------------------*/
4681:   len_s = merge->len_s;

4683:   len          = 0; /* length of buf_si[] */
4684:   merge->nsend = 0;
4685:   for (proc=0; proc<size; proc++) {
4686:     len_si[proc] = 0;
4687:     if (proc == rank) {
4688:       len_s[proc] = 0;
4689:     } else {
4690:       len_si[proc] = owners[proc+1] - owners[proc] + 1;
4691:       len_s[proc]  = ai[owners[proc+1]] - ai[owners[proc]]; /* num of rows to be sent to [proc] */
4692:     }
4693:     if (len_s[proc]) {
4694:       merge->nsend++;
4695:       nrows = 0;
4696:       for (i=owners[proc]; i<owners[proc+1]; i++) {
4697:         if (ai[i+1] > ai[i]) nrows++;
4698:       }
4699:       len_si[proc] = 2*(nrows+1);
4700:       len         += len_si[proc];
4701:     }
4702:   }

4704:   /* determine the number and length of messages to receive for ij-structure */
4705:   /*-------------------------------------------------------------------------*/
4706:   PetscGatherNumberOfMessages(comm,NULL,len_s,&merge->nrecv);
4707:   PetscGatherMessageLengths2(comm,merge->nsend,merge->nrecv,len_s,len_si,&merge->id_r,&merge->len_r,&len_ri);

4709:   /* post the Irecv of j-structure */
4710:   /*-------------------------------*/
4711:   PetscCommGetNewTag(comm,&tagj);
4712:   PetscPostIrecvInt(comm,tagj,merge->nrecv,merge->id_r,merge->len_r,&buf_rj,&rj_waits);

4714:   /* post the Isend of j-structure */
4715:   /*--------------------------------*/
4716:   PetscMalloc2(merge->nsend,&si_waits,merge->nsend,&sj_waits);

4718:   for (proc=0, k=0; proc<size; proc++) {
4719:     if (!len_s[proc]) continue;
4720:     i    = owners[proc];
4721:     MPI_Isend(aj+ai[i],len_s[proc],MPIU_INT,proc,tagj,comm,sj_waits+k);
4722:     k++;
4723:   }

4725:   /* receives and sends of j-structure are complete */
4726:   /*------------------------------------------------*/
4727:   if (merge->nrecv) {MPI_Waitall(merge->nrecv,rj_waits,status);}
4728:   if (merge->nsend) {MPI_Waitall(merge->nsend,sj_waits,status);}

4730:   /* send and recv i-structure */
4731:   /*---------------------------*/
4732:   PetscCommGetNewTag(comm,&tagi);
4733:   PetscPostIrecvInt(comm,tagi,merge->nrecv,merge->id_r,len_ri,&buf_ri,&ri_waits);

4735:   PetscMalloc1(len+1,&buf_s);
4736:   buf_si = buf_s;  /* points to the beginning of k-th msg to be sent */
4737:   for (proc=0,k=0; proc<size; proc++) {
4738:     if (!len_s[proc]) continue;
4739:     /* form outgoing message for i-structure:
4740:          buf_si[0]:                 nrows to be sent
4741:                [1:nrows]:           row index (global)
4742:                [nrows+1:2*nrows+1]: i-structure index
4743:     */
4744:     /*-------------------------------------------*/
4745:     nrows       = len_si[proc]/2 - 1;
4746:     buf_si_i    = buf_si + nrows+1;
4747:     buf_si[0]   = nrows;
4748:     buf_si_i[0] = 0;
4749:     nrows       = 0;
4750:     for (i=owners[proc]; i<owners[proc+1]; i++) {
4751:       anzi = ai[i+1] - ai[i];
4752:       if (anzi) {
4753:         buf_si_i[nrows+1] = buf_si_i[nrows] + anzi; /* i-structure */
4754:         buf_si[nrows+1]   = i-owners[proc]; /* local row index */
4755:         nrows++;
4756:       }
4757:     }
4758:     MPI_Isend(buf_si,len_si[proc],MPIU_INT,proc,tagi,comm,si_waits+k);
4759:     k++;
4760:     buf_si += len_si[proc];
4761:   }

4763:   if (merge->nrecv) {MPI_Waitall(merge->nrecv,ri_waits,status);}
4764:   if (merge->nsend) {MPI_Waitall(merge->nsend,si_waits,status);}

4766:   PetscInfo2(seqmat,"nsend: %D, nrecv: %D\n",merge->nsend,merge->nrecv);
4767:   for (i=0; i<merge->nrecv; i++) {
4768:     PetscInfo3(seqmat,"recv len_ri=%D, len_rj=%D from [%D]\n",len_ri[i],merge->len_r[i],merge->id_r[i]);
4769:   }

4771:   PetscFree(len_si);
4772:   PetscFree(len_ri);
4773:   PetscFree(rj_waits);
4774:   PetscFree2(si_waits,sj_waits);
4775:   PetscFree(ri_waits);
4776:   PetscFree(buf_s);
4777:   PetscFree(status);

4779:   /* compute a local seq matrix in each processor */
4780:   /*----------------------------------------------*/
4781:   /* allocate bi array and free space for accumulating nonzero column info */
4782:   PetscMalloc1(m+1,&bi);
4783:   bi[0] = 0;

4785:   /* create and initialize a linked list */
4786:   nlnk = N+1;
4787:   PetscLLCreate(N,N,nlnk,lnk,lnkbt);

4789:   /* initial FreeSpace size is 2*(num of local nnz(seqmat)) */
4790:   len  = ai[owners[rank+1]] - ai[owners[rank]];
4791:   PetscFreeSpaceGet(PetscIntMultTruncate(2,len)+1,&free_space);

4793:   current_space = free_space;

4795:   /* determine symbolic info for each local row */
4796:   PetscMalloc3(merge->nrecv,&buf_ri_k,merge->nrecv,&nextrow,merge->nrecv,&nextai);

4798:   for (k=0; k<merge->nrecv; k++) {
4799:     buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */
4800:     nrows       = *buf_ri_k[k];
4801:     nextrow[k]  = buf_ri_k[k] + 1;  /* next row number of k-th recved i-structure */
4802:     nextai[k]   = buf_ri_k[k] + (nrows + 1); /* poins to the next i-structure of k-th recved i-structure  */
4803:   }

4805:   MatPreallocateInitialize(comm,m,n,dnz,onz);
4806:   len  = 0;
4807:   for (i=0; i<m; i++) {
4808:     bnzi = 0;
4809:     /* add local non-zero cols of this proc's seqmat into lnk */
4810:     arow  = owners[rank] + i;
4811:     anzi  = ai[arow+1] - ai[arow];
4812:     aj    = a->j + ai[arow];
4813:     PetscLLAddSorted(anzi,aj,N,nlnk,lnk,lnkbt);
4814:     bnzi += nlnk;
4815:     /* add received col data into lnk */
4816:     for (k=0; k<merge->nrecv; k++) { /* k-th received message */
4817:       if (i == *nextrow[k]) { /* i-th row */
4818:         anzi  = *(nextai[k]+1) - *nextai[k];
4819:         aj    = buf_rj[k] + *nextai[k];
4820:         PetscLLAddSorted(anzi,aj,N,nlnk,lnk,lnkbt);
4821:         bnzi += nlnk;
4822:         nextrow[k]++; nextai[k]++;
4823:       }
4824:     }
4825:     if (len < bnzi) len = bnzi;  /* =max(bnzi) */

4827:     /* if free space is not available, make more free space */
4828:     if (current_space->local_remaining<bnzi) {
4829:       PetscFreeSpaceGet(PetscIntSumTruncate(bnzi,current_space->total_array_size),&current_space);
4830:       nspacedouble++;
4831:     }
4832:     /* copy data into free space, then initialize lnk */
4833:     PetscLLClean(N,N,bnzi,lnk,current_space->array,lnkbt);
4834:     MatPreallocateSet(i+owners[rank],bnzi,current_space->array,dnz,onz);

4836:     current_space->array           += bnzi;
4837:     current_space->local_used      += bnzi;
4838:     current_space->local_remaining -= bnzi;

4840:     bi[i+1] = bi[i] + bnzi;
4841:   }

4843:   PetscFree3(buf_ri_k,nextrow,nextai);

4845:   PetscMalloc1(bi[m]+1,&bj);
4846:   PetscFreeSpaceContiguous(&free_space,bj);
4847:   PetscLLDestroy(lnk,lnkbt);

4849:   /* create symbolic parallel matrix B_mpi */
4850:   /*---------------------------------------*/
4851:   MatGetBlockSizes(seqmat,&bs,&cbs);
4852:   MatCreate(comm,&B_mpi);
4853:   if (n==PETSC_DECIDE) {
4854:     MatSetSizes(B_mpi,m,n,PETSC_DETERMINE,N);
4855:   } else {
4856:     MatSetSizes(B_mpi,m,n,PETSC_DETERMINE,PETSC_DETERMINE);
4857:   }
4858:   MatSetBlockSizes(B_mpi,bs,cbs);
4859:   MatSetType(B_mpi,MATMPIAIJ);
4860:   MatMPIAIJSetPreallocation(B_mpi,0,dnz,0,onz);
4861:   MatPreallocateFinalize(dnz,onz);
4862:   MatSetOption(B_mpi,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);

4864:   /* B_mpi is not ready for use - assembly will be done by MatCreateMPIAIJSumSeqAIJNumeric() */
4865:   B_mpi->assembled    = PETSC_FALSE;
4866:   B_mpi->ops->destroy = MatDestroy_MPIAIJ_SeqsToMPI;
4867:   merge->bi           = bi;
4868:   merge->bj           = bj;
4869:   merge->buf_ri       = buf_ri;
4870:   merge->buf_rj       = buf_rj;
4871:   merge->coi          = NULL;
4872:   merge->coj          = NULL;
4873:   merge->owners_co    = NULL;

4875:   PetscCommDestroy(&comm);

4877:   /* attach the supporting struct to B_mpi for reuse */
4878:   PetscContainerCreate(PETSC_COMM_SELF,&container);
4879:   PetscContainerSetPointer(container,merge);
4880:   PetscObjectCompose((PetscObject)B_mpi,"MatMergeSeqsToMPI",(PetscObject)container);
4881:   PetscContainerDestroy(&container);
4882:   *mpimat = B_mpi;

4884:   PetscLogEventEnd(MAT_Seqstompisym,seqmat,0,0,0);
4885:   return(0);
4886: }

4888: /*@C
4889:       MatCreateMPIAIJSumSeqAIJ - Creates a MATMPIAIJ matrix by adding sequential
4890:                  matrices from each processor

4892:     Collective

4894:    Input Parameters:
4895: +    comm - the communicators the parallel matrix will live on
4896: .    seqmat - the input sequential matrices
4897: .    m - number of local rows (or PETSC_DECIDE)
4898: .    n - number of local columns (or PETSC_DECIDE)
4899: -    scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX

4901:    Output Parameter:
4902: .    mpimat - the parallel matrix generated

4904:     Level: advanced

4906:    Notes:
4907:      The dimensions of the sequential matrix in each processor MUST be the same.
4908:      The input seqmat is included into the container "Mat_Merge_SeqsToMPI", and will be
4909:      destroyed when mpimat is destroyed. Call PetscObjectQuery() to access seqmat.
4910: @*/
4911: PetscErrorCode MatCreateMPIAIJSumSeqAIJ(MPI_Comm comm,Mat seqmat,PetscInt m,PetscInt n,MatReuse scall,Mat *mpimat)
4912: {
4914:   PetscMPIInt    size;

4917:   MPI_Comm_size(comm,&size);
4918:   if (size == 1) {
4919:     PetscLogEventBegin(MAT_Seqstompi,seqmat,0,0,0);
4920:     if (scall == MAT_INITIAL_MATRIX) {
4921:       MatDuplicate(seqmat,MAT_COPY_VALUES,mpimat);
4922:     } else {
4923:       MatCopy(seqmat,*mpimat,SAME_NONZERO_PATTERN);
4924:     }
4925:     PetscLogEventEnd(MAT_Seqstompi,seqmat,0,0,0);
4926:     return(0);
4927:   }
4928:   PetscLogEventBegin(MAT_Seqstompi,seqmat,0,0,0);
4929:   if (scall == MAT_INITIAL_MATRIX) {
4930:     MatCreateMPIAIJSumSeqAIJSymbolic(comm,seqmat,m,n,mpimat);
4931:   }
4932:   MatCreateMPIAIJSumSeqAIJNumeric(seqmat,*mpimat);
4933:   PetscLogEventEnd(MAT_Seqstompi,seqmat,0,0,0);
4934:   return(0);
4935: }

4937: /*@
4938:      MatMPIAIJGetLocalMat - Creates a SeqAIJ from a MATMPIAIJ matrix by taking all its local rows and putting them into a sequential matrix with
4939:           mlocal rows and n columns. Where mlocal is the row count obtained with MatGetLocalSize() and n is the global column count obtained
4940:           with MatGetSize()

4942:     Not Collective

4944:    Input Parameters:
4945: +    A - the matrix
4946: -    scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX

4948:    Output Parameter:
4949: .    A_loc - the local sequential matrix generated

4951:     Level: developer

4953:    Notes:
4954:      When the communicator associated with A has size 1 and MAT_INITIAL_MATRIX is requested, the matrix returned is the diagonal part of A.
4955:      If MAT_REUSE_MATRIX is requested with comm size 1, MatCopy(Adiag,*A_loc,SAME_NONZERO_PATTERN) is called.
4956:      This means that one can preallocate the proper sequential matrix first and then call this routine with MAT_REUSE_MATRIX to safely
4957:      modify the values of the returned A_loc.

4959: .seealso: MatGetOwnershipRange(), MatMPIAIJGetLocalMatCondensed()

4961: @*/
4962: PetscErrorCode MatMPIAIJGetLocalMat(Mat A,MatReuse scall,Mat *A_loc)
4963: {
4965:   Mat_MPIAIJ     *mpimat=(Mat_MPIAIJ*)A->data;
4966:   Mat_SeqAIJ     *mat,*a,*b;
4967:   PetscInt       *ai,*aj,*bi,*bj,*cmap=mpimat->garray;
4968:   MatScalar      *aa,*ba,*cam;
4969:   PetscScalar    *ca;
4970:   PetscMPIInt    size;
4971:   PetscInt       am=A->rmap->n,i,j,k,cstart=A->cmap->rstart;
4972:   PetscInt       *ci,*cj,col,ncols_d,ncols_o,jo;
4973:   PetscBool      match;

4976:   PetscStrbeginswith(((PetscObject)A)->type_name,MATMPIAIJ,&match);
4977:   if (!match) SETERRQ(PetscObjectComm((PetscObject)A), PETSC_ERR_SUP,"Requires MATMPIAIJ matrix as input");
4978:   MPI_Comm_size(PetscObjectComm((PetscObject)A),&size);
4979:   if (size == 1) {
4980:     if (scall == MAT_INITIAL_MATRIX) {
4981:       PetscObjectReference((PetscObject)mpimat->A);
4982:       *A_loc = mpimat->A;
4983:     } else if (scall == MAT_REUSE_MATRIX) {
4984:       MatCopy(mpimat->A,*A_loc,SAME_NONZERO_PATTERN);
4985:     }
4986:     return(0);
4987:   }

4989:   PetscLogEventBegin(MAT_Getlocalmat,A,0,0,0);
4990:   a = (Mat_SeqAIJ*)(mpimat->A)->data;
4991:   b = (Mat_SeqAIJ*)(mpimat->B)->data;
4992:   ai = a->i; aj = a->j; bi = b->i; bj = b->j;
4993:   aa = a->a; ba = b->a;
4994:   if (scall == MAT_INITIAL_MATRIX) {
4995:     PetscMalloc1(1+am,&ci);
4996:     ci[0] = 0;
4997:     for (i=0; i<am; i++) {
4998:       ci[i+1] = ci[i] + (ai[i+1] - ai[i]) + (bi[i+1] - bi[i]);
4999:     }
5000:     PetscMalloc1(1+ci[am],&cj);
5001:     PetscMalloc1(1+ci[am],&ca);
5002:     k    = 0;
5003:     for (i=0; i<am; i++) {
5004:       ncols_o = bi[i+1] - bi[i];
5005:       ncols_d = ai[i+1] - ai[i];
5006:       /* off-diagonal portion of A */
5007:       for (jo=0; jo<ncols_o; jo++) {
5008:         col = cmap[*bj];
5009:         if (col >= cstart) break;
5010:         cj[k]   = col; bj++;
5011:         ca[k++] = *ba++;
5012:       }
5013:       /* diagonal portion of A */
5014:       for (j=0; j<ncols_d; j++) {
5015:         cj[k]   = cstart + *aj++;
5016:         ca[k++] = *aa++;
5017:       }
5018:       /* off-diagonal portion of A */
5019:       for (j=jo; j<ncols_o; j++) {
5020:         cj[k]   = cmap[*bj++];
5021:         ca[k++] = *ba++;
5022:       }
5023:     }
5024:     /* put together the new matrix */
5025:     MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,am,A->cmap->N,ci,cj,ca,A_loc);
5026:     /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */
5027:     /* Since these are PETSc arrays, change flags to free them as necessary. */
5028:     mat          = (Mat_SeqAIJ*)(*A_loc)->data;
5029:     mat->free_a  = PETSC_TRUE;
5030:     mat->free_ij = PETSC_TRUE;
5031:     mat->nonew   = 0;
5032:   } else if (scall == MAT_REUSE_MATRIX) {
5033:     mat=(Mat_SeqAIJ*)(*A_loc)->data;
5034:     ci = mat->i; cj = mat->j; cam = mat->a;
5035:     for (i=0; i<am; i++) {
5036:       /* off-diagonal portion of A */
5037:       ncols_o = bi[i+1] - bi[i];
5038:       for (jo=0; jo<ncols_o; jo++) {
5039:         col = cmap[*bj];
5040:         if (col >= cstart) break;
5041:         *cam++ = *ba++; bj++;
5042:       }
5043:       /* diagonal portion of A */
5044:       ncols_d = ai[i+1] - ai[i];
5045:       for (j=0; j<ncols_d; j++) *cam++ = *aa++;
5046:       /* off-diagonal portion of A */
5047:       for (j=jo; j<ncols_o; j++) {
5048:         *cam++ = *ba++; bj++;
5049:       }
5050:     }
5051:   } else SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Invalid MatReuse %d",(int)scall);
5052:   PetscLogEventEnd(MAT_Getlocalmat,A,0,0,0);
5053:   return(0);
5054: }

5056: /*@C
5057:      MatMPIAIJGetLocalMatCondensed - Creates a SeqAIJ matrix from an MATMPIAIJ matrix by taking all its local rows and NON-ZERO columns

5059:     Not Collective

5061:    Input Parameters:
5062: +    A - the matrix
5063: .    scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX
5064: -    row, col - index sets of rows and columns to extract (or NULL)

5066:    Output Parameter:
5067: .    A_loc - the local sequential matrix generated

5069:     Level: developer

5071: .seealso: MatGetOwnershipRange(), MatMPIAIJGetLocalMat()

5073: @*/
5074: PetscErrorCode MatMPIAIJGetLocalMatCondensed(Mat A,MatReuse scall,IS *row,IS *col,Mat *A_loc)
5075: {
5076:   Mat_MPIAIJ     *a=(Mat_MPIAIJ*)A->data;
5078:   PetscInt       i,start,end,ncols,nzA,nzB,*cmap,imark,*idx;
5079:   IS             isrowa,iscola;
5080:   Mat            *aloc;
5081:   PetscBool      match;

5084:   PetscObjectTypeCompare((PetscObject)A,MATMPIAIJ,&match);
5085:   if (!match) SETERRQ(PetscObjectComm((PetscObject)A), PETSC_ERR_SUP,"Requires MATMPIAIJ matrix as input");
5086:   PetscLogEventBegin(MAT_Getlocalmatcondensed,A,0,0,0);
5087:   if (!row) {
5088:     start = A->rmap->rstart; end = A->rmap->rend;
5089:     ISCreateStride(PETSC_COMM_SELF,end-start,start,1,&isrowa);
5090:   } else {
5091:     isrowa = *row;
5092:   }
5093:   if (!col) {
5094:     start = A->cmap->rstart;
5095:     cmap  = a->garray;
5096:     nzA   = a->A->cmap->n;
5097:     nzB   = a->B->cmap->n;
5098:     PetscMalloc1(nzA+nzB, &idx);
5099:     ncols = 0;
5100:     for (i=0; i<nzB; i++) {
5101:       if (cmap[i] < start) idx[ncols++] = cmap[i];
5102:       else break;
5103:     }
5104:     imark = i;
5105:     for (i=0; i<nzA; i++) idx[ncols++] = start + i;
5106:     for (i=imark; i<nzB; i++) idx[ncols++] = cmap[i];
5107:     ISCreateGeneral(PETSC_COMM_SELF,ncols,idx,PETSC_OWN_POINTER,&iscola);
5108:   } else {
5109:     iscola = *col;
5110:   }
5111:   if (scall != MAT_INITIAL_MATRIX) {
5112:     PetscMalloc1(1,&aloc);
5113:     aloc[0] = *A_loc;
5114:   }
5115:   MatCreateSubMatrices(A,1,&isrowa,&iscola,scall,&aloc);
5116:   if (!col) { /* attach global id of condensed columns */
5117:     PetscObjectCompose((PetscObject)aloc[0],"_petsc_GetLocalMatCondensed_iscol",(PetscObject)iscola);
5118:   }
5119:   *A_loc = aloc[0];
5120:   PetscFree(aloc);
5121:   if (!row) {
5122:     ISDestroy(&isrowa);
5123:   }
5124:   if (!col) {
5125:     ISDestroy(&iscola);
5126:   }
5127:   PetscLogEventEnd(MAT_Getlocalmatcondensed,A,0,0,0);
5128:   return(0);
5129: }

5131: /*
5132:  * Destroy a mat that may be composed with PetscSF communication objects.
5133:  * The SF objects were created in MatCreateSeqSubMatrixWithRows_Private.
5134:  * */
5135: PetscErrorCode MatDestroy_SeqAIJ_PetscSF(Mat mat)
5136: {
5137:   PetscSF          sf,osf;
5138:   IS               map;
5139:   PetscErrorCode   ierr;

5142:   PetscObjectQuery((PetscObject)mat,"diagsf",(PetscObject*)&sf);
5143:   PetscObjectQuery((PetscObject)mat,"offdiagsf",(PetscObject*)&osf);
5144:   PetscSFDestroy(&sf);
5145:   PetscSFDestroy(&osf);
5146:   PetscObjectQuery((PetscObject)mat,"aoffdiagtopothmapping",(PetscObject*)&map);
5147:   ISDestroy(&map);
5148:   MatDestroy_SeqAIJ(mat);
5149:   return(0);
5150: }

5152: /*
5153:  * Create a sequential AIJ matrix based on row indices. a whole column is extracted once a row is matched.
5154:  * Row could be local or remote.The routine is designed to be scalable in memory so that nothing is based
5155:  * on a global size.
5156:  * */
5157: PetscErrorCode MatCreateSeqSubMatrixWithRows_Private(Mat P,IS rows,Mat *P_oth)
5158: {
5159:   Mat_MPIAIJ               *p=(Mat_MPIAIJ*)P->data;
5160:   Mat_SeqAIJ               *pd=(Mat_SeqAIJ*)(p->A)->data,*po=(Mat_SeqAIJ*)(p->B)->data,*p_oth;
5161:   PetscInt                 plocalsize,nrows,*ilocal,*oilocal,i,lidx,*nrcols,*nlcols,ncol;
5162:   PetscMPIInt              owner;
5163:   PetscSFNode              *iremote,*oiremote;
5164:   const PetscInt           *lrowindices;
5165:   PetscErrorCode           ierr;
5166:   PetscSF                  sf,osf;
5167:   PetscInt                 pcstart,*roffsets,*loffsets,*pnnz,j;
5168:   PetscInt                 ontotalcols,dntotalcols,ntotalcols,nout;
5169:   MPI_Comm                 comm;
5170:   ISLocalToGlobalMapping   mapping;

5173:   PetscObjectGetComm((PetscObject)P,&comm);
5174:   /* plocalsize is the number of roots
5175:    * nrows is the number of leaves
5176:    * */
5177:   MatGetLocalSize(P,&plocalsize,NULL);
5178:   ISGetLocalSize(rows,&nrows);
5179:   PetscCalloc1(nrows,&iremote);
5180:   ISGetIndices(rows,&lrowindices);
5181:   for (i=0;i<nrows;i++) {
5182:     /* Find a remote index and an owner for a row
5183:      * The row could be local or remote
5184:      * */
5185:     owner = 0;
5186:     lidx  = 0;
5187:     PetscLayoutFindOwnerIndex(P->rmap,lrowindices[i],&owner,&lidx);
5188:     iremote[i].index = lidx;
5189:     iremote[i].rank  = owner;
5190:   }
5191:   /* Create SF to communicate how many nonzero columns for each row */
5192:   PetscSFCreate(comm,&sf);
5193:   /* SF will figure out the number of nonzero colunms for each row, and their
5194:    * offsets
5195:    * */
5196:   PetscSFSetGraph(sf,plocalsize,nrows,NULL,PETSC_OWN_POINTER,iremote,PETSC_OWN_POINTER);
5197:   PetscSFSetFromOptions(sf);
5198:   PetscSFSetUp(sf);

5200:   PetscCalloc1(2*(plocalsize+1),&roffsets);
5201:   PetscCalloc1(2*plocalsize,&nrcols);
5202:   PetscCalloc1(nrows,&pnnz);
5203:   roffsets[0] = 0;
5204:   roffsets[1] = 0;
5205:   for (i=0;i<plocalsize;i++) {
5206:     /* diag */
5207:     nrcols[i*2+0] = pd->i[i+1] - pd->i[i];
5208:     /* off diag */
5209:     nrcols[i*2+1] = po->i[i+1] - po->i[i];
5210:     /* compute offsets so that we relative location for each row */
5211:     roffsets[(i+1)*2+0] = roffsets[i*2+0] + nrcols[i*2+0];
5212:     roffsets[(i+1)*2+1] = roffsets[i*2+1] + nrcols[i*2+1];
5213:   }
5214:   PetscCalloc1(2*nrows,&nlcols);
5215:   PetscCalloc1(2*nrows,&loffsets);
5216:   /* 'r' means root, and 'l' means leaf */
5217:   PetscSFBcastBegin(sf,MPIU_2INT,nrcols,nlcols);
5218:   PetscSFBcastBegin(sf,MPIU_2INT,roffsets,loffsets);
5219:   PetscSFBcastEnd(sf,MPIU_2INT,nrcols,nlcols);
5220:   PetscSFBcastEnd(sf,MPIU_2INT,roffsets,loffsets);
5221:   PetscSFDestroy(&sf);
5222:   PetscFree(roffsets);
5223:   PetscFree(nrcols);
5224:   dntotalcols = 0;
5225:   ontotalcols = 0;
5226:   ncol = 0;
5227:   for (i=0;i<nrows;i++) {
5228:     pnnz[i] = nlcols[i*2+0] + nlcols[i*2+1];
5229:     ncol = PetscMax(pnnz[i],ncol);
5230:     /* diag */
5231:     dntotalcols += nlcols[i*2+0];
5232:     /* off diag */
5233:     ontotalcols += nlcols[i*2+1];
5234:   }
5235:   /* We do not need to figure the right number of columns
5236:    * since all the calculations will be done by going through the raw data
5237:    * */
5238:   MatCreateSeqAIJ(PETSC_COMM_SELF,nrows,ncol,0,pnnz,P_oth);
5239:   MatSetUp(*P_oth);
5240:   PetscFree(pnnz);
5241:   p_oth = (Mat_SeqAIJ*) (*P_oth)->data;
5242:   /* diag */
5243:   PetscCalloc1(dntotalcols,&iremote);
5244:   /* off diag */
5245:   PetscCalloc1(ontotalcols,&oiremote);
5246:   /* diag */
5247:   PetscCalloc1(dntotalcols,&ilocal);
5248:   /* off diag */
5249:   PetscCalloc1(ontotalcols,&oilocal);
5250:   dntotalcols = 0;
5251:   ontotalcols = 0;
5252:   ntotalcols  = 0;
5253:   for (i=0;i<nrows;i++) {
5254:     owner = 0;
5255:     PetscLayoutFindOwnerIndex(P->rmap,lrowindices[i],&owner,NULL);
5256:     /* Set iremote for diag matrix */
5257:     for (j=0;j<nlcols[i*2+0];j++) {
5258:       iremote[dntotalcols].index   = loffsets[i*2+0] + j;
5259:       iremote[dntotalcols].rank    = owner;
5260:       /* P_oth is seqAIJ so that ilocal need to point to the first part of memory */
5261:       ilocal[dntotalcols++]        = ntotalcols++;
5262:     }
5263:     /* off diag */
5264:     for (j=0;j<nlcols[i*2+1];j++) {
5265:       oiremote[ontotalcols].index   = loffsets[i*2+1] + j;
5266:       oiremote[ontotalcols].rank    = owner;
5267:       oilocal[ontotalcols++]        = ntotalcols++;
5268:     }
5269:   }
5270:   ISRestoreIndices(rows,&lrowindices);
5271:   PetscFree(loffsets);
5272:   PetscFree(nlcols);
5273:   PetscSFCreate(comm,&sf);
5274:   /* P serves as roots and P_oth is leaves
5275:    * Diag matrix
5276:    * */
5277:   PetscSFSetGraph(sf,pd->i[plocalsize],dntotalcols,ilocal,PETSC_OWN_POINTER,iremote,PETSC_OWN_POINTER);
5278:   PetscSFSetFromOptions(sf);
5279:   PetscSFSetUp(sf);

5281:   PetscSFCreate(comm,&osf);
5282:   /* Off diag */
5283:   PetscSFSetGraph(osf,po->i[plocalsize],ontotalcols,oilocal,PETSC_OWN_POINTER,oiremote,PETSC_OWN_POINTER);
5284:   PetscSFSetFromOptions(osf);
5285:   PetscSFSetUp(osf);
5286:   /* We operate on the matrix internal data for saving memory */
5287:   PetscSFBcastBegin(sf,MPIU_SCALAR,pd->a,p_oth->a);
5288:   PetscSFBcastBegin(osf,MPIU_SCALAR,po->a,p_oth->a);
5289:   MatGetOwnershipRangeColumn(P,&pcstart,NULL);
5290:   /* Convert to global indices for diag matrix */
5291:   for (i=0;i<pd->i[plocalsize];i++) pd->j[i] += pcstart;
5292:   PetscSFBcastBegin(sf,MPIU_INT,pd->j,p_oth->j);
5293:   /* We want P_oth store global indices */
5294:   ISLocalToGlobalMappingCreate(comm,1,p->B->cmap->n,p->garray,PETSC_COPY_VALUES,&mapping);
5295:   /* Use memory scalable approach */
5296:   ISLocalToGlobalMappingSetType(mapping,ISLOCALTOGLOBALMAPPINGHASH);
5297:   ISLocalToGlobalMappingApply(mapping,po->i[plocalsize],po->j,po->j);
5298:   PetscSFBcastBegin(osf,MPIU_INT,po->j,p_oth->j);
5299:   PetscSFBcastEnd(sf,MPIU_INT,pd->j,p_oth->j);
5300:   /* Convert back to local indices */
5301:   for (i=0;i<pd->i[plocalsize];i++) pd->j[i] -= pcstart;
5302:   PetscSFBcastEnd(osf,MPIU_INT,po->j,p_oth->j);
5303:   nout = 0;
5304:   ISGlobalToLocalMappingApply(mapping,IS_GTOLM_DROP,po->i[plocalsize],po->j,&nout,po->j);
5305:   if (nout != po->i[plocalsize]) SETERRQ2(comm,PETSC_ERR_ARG_INCOMP,"n %D does not equal to nout %D \n",po->i[plocalsize],nout);
5306:   ISLocalToGlobalMappingDestroy(&mapping);
5307:   /* Exchange values */
5308:   PetscSFBcastEnd(sf,MPIU_SCALAR,pd->a,p_oth->a);
5309:   PetscSFBcastEnd(osf,MPIU_SCALAR,po->a,p_oth->a);
5310:   /* Stop PETSc from shrinking memory */
5311:   for (i=0;i<nrows;i++) p_oth->ilen[i] = p_oth->imax[i];
5312:   MatAssemblyBegin(*P_oth,MAT_FINAL_ASSEMBLY);
5313:   MatAssemblyEnd(*P_oth,MAT_FINAL_ASSEMBLY);
5314:   /* Attach PetscSF objects to P_oth so that we can reuse it later */
5315:   PetscObjectCompose((PetscObject)*P_oth,"diagsf",(PetscObject)sf);
5316:   PetscObjectCompose((PetscObject)*P_oth,"offdiagsf",(PetscObject)osf);
5317:   /* ``New MatDestroy" takes care of PetscSF objects as well */
5318:   (*P_oth)->ops->destroy = MatDestroy_SeqAIJ_PetscSF;
5319:   return(0);
5320: }

5322: /*
5323:  * Creates a SeqAIJ matrix by taking rows of B that equal to nonzero columns of local A
5324:  * This supports MPIAIJ and MAIJ
5325:  * */
5326: PetscErrorCode MatGetBrowsOfAcols_MPIXAIJ(Mat A,Mat P,PetscInt dof,MatReuse reuse,Mat *P_oth)
5327: {
5328:   Mat_MPIAIJ            *a=(Mat_MPIAIJ*)A->data,*p=(Mat_MPIAIJ*)P->data;
5329:   Mat_SeqAIJ            *p_oth;
5330:   Mat_SeqAIJ            *pd=(Mat_SeqAIJ*)(p->A)->data,*po=(Mat_SeqAIJ*)(p->B)->data;
5331:   IS                    rows,map;
5332:   PetscHMapI            hamp;
5333:   PetscInt              i,htsize,*rowindices,off,*mapping,key,count;
5334:   MPI_Comm              comm;
5335:   PetscSF               sf,osf;
5336:   PetscBool             has;
5337:   PetscErrorCode        ierr;

5340:   PetscObjectGetComm((PetscObject)A,&comm);
5341:   PetscLogEventBegin(MAT_GetBrowsOfAocols,A,P,0,0);
5342:   /* If it is the first time, create an index set of off-diag nonzero columns of A,
5343:    *  and then create a submatrix (that often is an overlapping matrix)
5344:    * */
5345:   if (reuse==MAT_INITIAL_MATRIX) {
5346:     /* Use a hash table to figure out unique keys */
5347:     PetscHMapICreate(&hamp);
5348:     PetscHMapIResize(hamp,a->B->cmap->n);
5349:     PetscCalloc1(a->B->cmap->n,&mapping);
5350:     count = 0;
5351:     /* Assume that  a->g is sorted, otherwise the following does not make sense */
5352:     for (i=0;i<a->B->cmap->n;i++) {
5353:       key  = a->garray[i]/dof;
5354:       PetscHMapIHas(hamp,key,&has);
5355:       if (!has) {
5356:         mapping[i] = count;
5357:         PetscHMapISet(hamp,key,count++);
5358:       } else {
5359:         /* Current 'i' has the same value the previous step */
5360:         mapping[i] = count-1;
5361:       }
5362:     }
5363:     ISCreateGeneral(comm,a->B->cmap->n,mapping,PETSC_OWN_POINTER,&map);
5364:     PetscHMapIGetSize(hamp,&htsize);
5365:     if (htsize!=count) SETERRQ2(comm,PETSC_ERR_ARG_INCOMP," Size of hash map %D is inconsistent with count %D \n",htsize,count);
5366:     PetscCalloc1(htsize,&rowindices);
5367:     off = 0;
5368:     PetscHMapIGetKeys(hamp,&off,rowindices);
5369:     PetscHMapIDestroy(&hamp);
5370:     PetscSortInt(htsize,rowindices);
5371:     ISCreateGeneral(comm,htsize,rowindices,PETSC_OWN_POINTER,&rows);
5372:     /* In case, the matrix was already created but users want to recreate the matrix */
5373:     MatDestroy(P_oth);
5374:     MatCreateSeqSubMatrixWithRows_Private(P,rows,P_oth);
5375:     PetscObjectCompose((PetscObject)*P_oth,"aoffdiagtopothmapping",(PetscObject)map);
5376:     ISDestroy(&rows);
5377:   } else if (reuse==MAT_REUSE_MATRIX) {
5378:     /* If matrix was already created, we simply update values using SF objects
5379:      * that as attached to the matrix ealier.
5380:      *  */
5381:     PetscObjectQuery((PetscObject)*P_oth,"diagsf",(PetscObject*)&sf);
5382:     PetscObjectQuery((PetscObject)*P_oth,"offdiagsf",(PetscObject*)&osf);
5383:     if (!sf || !osf) {
5384:       SETERRQ(comm,PETSC_ERR_ARG_NULL,"Matrix is not initialized yet \n");
5385:     }
5386:     p_oth = (Mat_SeqAIJ*) (*P_oth)->data;
5387:     /* Update values in place */
5388:     PetscSFBcastBegin(sf,MPIU_SCALAR,pd->a,p_oth->a);
5389:     PetscSFBcastBegin(osf,MPIU_SCALAR,po->a,p_oth->a);
5390:     PetscSFBcastEnd(sf,MPIU_SCALAR,pd->a,p_oth->a);
5391:     PetscSFBcastEnd(osf,MPIU_SCALAR,po->a,p_oth->a);
5392:   } else {
5393:     SETERRQ(comm,PETSC_ERR_ARG_UNKNOWN_TYPE,"Unknown reuse type \n");
5394:   }
5395:   PetscLogEventEnd(MAT_GetBrowsOfAocols,A,P,0,0);
5396:   return(0);
5397: }

5399: /*@C
5400:     MatGetBrowsOfAcols - Creates a SeqAIJ matrix by taking rows of B that equal to nonzero columns of local A

5402:     Collective on Mat

5404:    Input Parameters:
5405: +    A,B - the matrices in mpiaij format
5406: .    scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX
5407: -    rowb, colb - index sets of rows and columns of B to extract (or NULL)

5409:    Output Parameter:
5410: +    rowb, colb - index sets of rows and columns of B to extract
5411: -    B_seq - the sequential matrix generated

5413:     Level: developer

5415: @*/
5416: PetscErrorCode MatGetBrowsOfAcols(Mat A,Mat B,MatReuse scall,IS *rowb,IS *colb,Mat *B_seq)
5417: {
5418:   Mat_MPIAIJ     *a=(Mat_MPIAIJ*)A->data;
5420:   PetscInt       *idx,i,start,ncols,nzA,nzB,*cmap,imark;
5421:   IS             isrowb,iscolb;
5422:   Mat            *bseq=NULL;

5425:   if (A->cmap->rstart != B->rmap->rstart || A->cmap->rend != B->rmap->rend) {
5426:     SETERRQ4(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Matrix local dimensions are incompatible, (%D, %D) != (%D,%D)",A->cmap->rstart,A->cmap->rend,B->rmap->rstart,B->rmap->rend);
5427:   }
5428:   PetscLogEventBegin(MAT_GetBrowsOfAcols,A,B,0,0);

5430:   if (scall == MAT_INITIAL_MATRIX) {
5431:     start = A->cmap->rstart;
5432:     cmap  = a->garray;
5433:     nzA   = a->A->cmap->n;
5434:     nzB   = a->B->cmap->n;
5435:     PetscMalloc1(nzA+nzB, &idx);
5436:     ncols = 0;
5437:     for (i=0; i<nzB; i++) {  /* row < local row index */
5438:       if (cmap[i] < start) idx[ncols++] = cmap[i];
5439:       else break;
5440:     }
5441:     imark = i;
5442:     for (i=0; i<nzA; i++) idx[ncols++] = start + i;  /* local rows */
5443:     for (i=imark; i<nzB; i++) idx[ncols++] = cmap[i]; /* row > local row index */
5444:     ISCreateGeneral(PETSC_COMM_SELF,ncols,idx,PETSC_OWN_POINTER,&isrowb);
5445:     ISCreateStride(PETSC_COMM_SELF,B->cmap->N,0,1,&iscolb);
5446:   } else {
5447:     if (!rowb || !colb) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"IS rowb and colb must be provided for MAT_REUSE_MATRIX");
5448:     isrowb  = *rowb; iscolb = *colb;
5449:     PetscMalloc1(1,&bseq);
5450:     bseq[0] = *B_seq;
5451:   }
5452:   MatCreateSubMatrices(B,1,&isrowb,&iscolb,scall,&bseq);
5453:   *B_seq = bseq[0];
5454:   PetscFree(bseq);
5455:   if (!rowb) {
5456:     ISDestroy(&isrowb);
5457:   } else {
5458:     *rowb = isrowb;
5459:   }
5460:   if (!colb) {
5461:     ISDestroy(&iscolb);
5462:   } else {
5463:     *colb = iscolb;
5464:   }
5465:   PetscLogEventEnd(MAT_GetBrowsOfAcols,A,B,0,0);
5466:   return(0);
5467: }

5469: /*
5470:     MatGetBrowsOfAoCols_MPIAIJ - Creates a SeqAIJ matrix by taking rows of B that equal to nonzero columns
5471:     of the OFF-DIAGONAL portion of local A

5473:     Collective on Mat

5475:    Input Parameters:
5476: +    A,B - the matrices in mpiaij format
5477: -    scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX

5479:    Output Parameter:
5480: +    startsj_s - starting point in B's sending j-arrays, saved for MAT_REUSE (or NULL)
5481: .    startsj_r - starting point in B's receiving j-arrays, saved for MAT_REUSE (or NULL)
5482: .    bufa_ptr - array for sending matrix values, saved for MAT_REUSE (or NULL)
5483: -    B_oth - the sequential matrix generated with size aBn=a->B->cmap->n by B->cmap->N

5485:     Developer Notes: This directly accesses information inside the VecScatter associated with the matrix-vector product
5486:      for this matrix. This is not desirable..

5488:     Level: developer

5490: */
5491: PetscErrorCode MatGetBrowsOfAoCols_MPIAIJ(Mat A,Mat B,MatReuse scall,PetscInt **startsj_s,PetscInt **startsj_r,MatScalar **bufa_ptr,Mat *B_oth)
5492: {
5493:   PetscErrorCode         ierr;
5494:   Mat_MPIAIJ             *a=(Mat_MPIAIJ*)A->data;
5495:   Mat_SeqAIJ             *b_oth;
5496:   VecScatter             ctx;
5497:   MPI_Comm               comm;
5498:   const PetscMPIInt      *rprocs,*sprocs;
5499:   const PetscInt         *srow,*rstarts,*sstarts;
5500:   PetscInt               *rowlen,*bufj,*bufJ,ncols = 0,aBn=a->B->cmap->n,row,*b_othi,*b_othj,*rvalues=NULL,*svalues=NULL,*cols,sbs,rbs;
5501:   PetscInt               i,j,k=0,l,ll,nrecvs,nsends,nrows,*rstartsj = 0,*sstartsj,len;
5502:   PetscScalar            *b_otha,*bufa,*bufA,*vals = NULL;
5503:   MPI_Request            *rwaits = NULL,*swaits = NULL;
5504:   MPI_Status             rstatus;
5505:   PetscMPIInt            jj,size,tag,rank,nsends_mpi,nrecvs_mpi;

5508:   PetscObjectGetComm((PetscObject)A,&comm);
5509:   MPI_Comm_size(comm,&size);

5511:   if (A->cmap->rstart != B->rmap->rstart || A->cmap->rend != B->rmap->rend) {
5512:     SETERRQ4(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Matrix local dimensions are incompatible, (%d, %d) != (%d,%d)",A->cmap->rstart,A->cmap->rend,B->rmap->rstart,B->rmap->rend);
5513:   }
5514:   PetscLogEventBegin(MAT_GetBrowsOfAocols,A,B,0,0);
5515:   MPI_Comm_rank(comm,&rank);

5517:   if (size == 1) {
5518:     startsj_s = NULL;
5519:     bufa_ptr  = NULL;
5520:     *B_oth    = NULL;
5521:     return(0);
5522:   }

5524:   ctx = a->Mvctx;
5525:   tag = ((PetscObject)ctx)->tag;

5527:   if (ctx->inuse) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE," Scatter ctx already in use");
5528:   VecScatterGetRemote_Private(ctx,PETSC_TRUE/*send*/,&nsends,&sstarts,&srow,&sprocs,&sbs);
5529:   /* rprocs[] must be ordered so that indices received from them are ordered in rvalues[], which is key to algorithms used in this subroutine */
5530:   VecScatterGetRemoteOrdered_Private(ctx,PETSC_FALSE/*recv*/,&nrecvs,&rstarts,NULL/*indices not needed*/,&rprocs,&rbs);
5531:   PetscMPIIntCast(nsends,&nsends_mpi);
5532:   PetscMPIIntCast(nrecvs,&nrecvs_mpi);
5533:   PetscMalloc2(nrecvs,&rwaits,nsends,&swaits);

5535:   if (!startsj_s || !bufa_ptr) scall = MAT_INITIAL_MATRIX;
5536:   if (scall == MAT_INITIAL_MATRIX) {
5537:     /* i-array */
5538:     /*---------*/
5539:     /*  post receives */
5540:     if (nrecvs) {PetscMalloc1(rbs*(rstarts[nrecvs] - rstarts[0]),&rvalues);} /* rstarts can be NULL when nrecvs=0 */
5541:     for (i=0; i<nrecvs; i++) {
5542:       rowlen = rvalues + rstarts[i]*rbs;
5543:       nrows  = (rstarts[i+1]-rstarts[i])*rbs; /* num of indices to be received */
5544:       MPI_Irecv(rowlen,nrows,MPIU_INT,rprocs[i],tag,comm,rwaits+i);
5545:     }

5547:     /* pack the outgoing message */
5548:     PetscMalloc2(nsends+1,&sstartsj,nrecvs+1,&rstartsj);

5550:     sstartsj[0] = 0;
5551:     rstartsj[0] = 0;
5552:     len         = 0; /* total length of j or a array to be sent */
5553:     if (nsends) {
5554:       k    = sstarts[0]; /* ATTENTION: sstarts[0] and rstarts[0] are not necessarily zero */
5555:       PetscMalloc1(sbs*(sstarts[nsends]-sstarts[0]),&svalues);
5556:     }
5557:     for (i=0; i<nsends; i++) {
5558:       rowlen = svalues + (sstarts[i]-sstarts[0])*sbs;
5559:       nrows  = sstarts[i+1]-sstarts[i]; /* num of block rows */
5560:       for (j=0; j<nrows; j++) {
5561:         row = srow[k] + B->rmap->range[rank]; /* global row idx */
5562:         for (l=0; l<sbs; l++) {
5563:           MatGetRow_MPIAIJ(B,row+l,&ncols,NULL,NULL); /* rowlength */

5565:           rowlen[j*sbs+l] = ncols;

5567:           len += ncols;
5568:           MatRestoreRow_MPIAIJ(B,row+l,&ncols,NULL,NULL);
5569:         }
5570:         k++;
5571:       }
5572:       MPI_Isend(rowlen,nrows*sbs,MPIU_INT,sprocs[i],tag,comm,swaits+i);

5574:       sstartsj[i+1] = len;  /* starting point of (i+1)-th outgoing msg in bufj and bufa */
5575:     }
5576:     /* recvs and sends of i-array are completed */
5577:     i = nrecvs;
5578:     while (i--) {
5579:       MPI_Waitany(nrecvs_mpi,rwaits,&jj,&rstatus);
5580:     }
5581:     if (nsends) {MPI_Waitall(nsends_mpi,swaits,MPI_STATUSES_IGNORE);}
5582:     PetscFree(svalues);

5584:     /* allocate buffers for sending j and a arrays */
5585:     PetscMalloc1(len+1,&bufj);
5586:     PetscMalloc1(len+1,&bufa);

5588:     /* create i-array of B_oth */
5589:     PetscMalloc1(aBn+2,&b_othi);

5591:     b_othi[0] = 0;
5592:     len       = 0; /* total length of j or a array to be received */
5593:     k         = 0;
5594:     for (i=0; i<nrecvs; i++) {
5595:       rowlen = rvalues + (rstarts[i]-rstarts[0])*rbs;
5596:       nrows  = (rstarts[i+1]-rstarts[i])*rbs; /* num of rows to be received */
5597:       for (j=0; j<nrows; j++) {
5598:         b_othi[k+1] = b_othi[k] + rowlen[j];
5599:         PetscIntSumError(rowlen[j],len,&len);
5600:         k++;
5601:       }
5602:       rstartsj[i+1] = len; /* starting point of (i+1)-th incoming msg in bufj and bufa */
5603:     }
5604:     PetscFree(rvalues);

5606:     /* allocate space for j and a arrrays of B_oth */
5607:     PetscMalloc1(b_othi[aBn]+1,&b_othj);
5608:     PetscMalloc1(b_othi[aBn]+1,&b_otha);

5610:     /* j-array */
5611:     /*---------*/
5612:     /*  post receives of j-array */
5613:     for (i=0; i<nrecvs; i++) {
5614:       nrows = rstartsj[i+1]-rstartsj[i]; /* length of the msg received */
5615:       MPI_Irecv(b_othj+rstartsj[i],nrows,MPIU_INT,rprocs[i],tag,comm,rwaits+i);
5616:     }

5618:     /* pack the outgoing message j-array */
5619:     if (nsends) k = sstarts[0];
5620:     for (i=0; i<nsends; i++) {
5621:       nrows = sstarts[i+1]-sstarts[i]; /* num of block rows */
5622:       bufJ  = bufj+sstartsj[i];
5623:       for (j=0; j<nrows; j++) {
5624:         row = srow[k++] + B->rmap->range[rank];  /* global row idx */
5625:         for (ll=0; ll<sbs; ll++) {
5626:           MatGetRow_MPIAIJ(B,row+ll,&ncols,&cols,NULL);
5627:           for (l=0; l<ncols; l++) {
5628:             *bufJ++ = cols[l];
5629:           }
5630:           MatRestoreRow_MPIAIJ(B,row+ll,&ncols,&cols,NULL);
5631:         }
5632:       }
5633:       MPI_Isend(bufj+sstartsj[i],sstartsj[i+1]-sstartsj[i],MPIU_INT,sprocs[i],tag,comm,swaits+i);
5634:     }

5636:     /* recvs and sends of j-array are completed */
5637:     i = nrecvs;
5638:     while (i--) {
5639:       MPI_Waitany(nrecvs_mpi,rwaits,&jj,&rstatus);
5640:     }
5641:     if (nsends) {MPI_Waitall(nsends_mpi,swaits,MPI_STATUSES_IGNORE);}
5642:   } else if (scall == MAT_REUSE_MATRIX) {
5643:     sstartsj = *startsj_s;
5644:     rstartsj = *startsj_r;
5645:     bufa     = *bufa_ptr;
5646:     b_oth    = (Mat_SeqAIJ*)(*B_oth)->data;
5647:     b_otha   = b_oth->a;
5648:   } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE, "Matrix P does not posses an object container");

5650:   /* a-array */
5651:   /*---------*/
5652:   /*  post receives of a-array */
5653:   for (i=0; i<nrecvs; i++) {
5654:     nrows = rstartsj[i+1]-rstartsj[i]; /* length of the msg received */
5655:     MPI_Irecv(b_otha+rstartsj[i],nrows,MPIU_SCALAR,rprocs[i],tag,comm,rwaits+i);
5656:   }

5658:   /* pack the outgoing message a-array */
5659:   if (nsends) k = sstarts[0];
5660:   for (i=0; i<nsends; i++) {
5661:     nrows = sstarts[i+1]-sstarts[i]; /* num of block rows */
5662:     bufA  = bufa+sstartsj[i];
5663:     for (j=0; j<nrows; j++) {
5664:       row = srow[k++] + B->rmap->range[rank];  /* global row idx */
5665:       for (ll=0; ll<sbs; ll++) {
5666:         MatGetRow_MPIAIJ(B,row+ll,&ncols,NULL,&vals);
5667:         for (l=0; l<ncols; l++) {
5668:           *bufA++ = vals[l];
5669:         }
5670:         MatRestoreRow_MPIAIJ(B,row+ll,&ncols,NULL,&vals);
5671:       }
5672:     }
5673:     MPI_Isend(bufa+sstartsj[i],sstartsj[i+1]-sstartsj[i],MPIU_SCALAR,sprocs[i],tag,comm,swaits+i);
5674:   }
5675:   /* recvs and sends of a-array are completed */
5676:   i = nrecvs;
5677:   while (i--) {
5678:     MPI_Waitany(nrecvs_mpi,rwaits,&jj,&rstatus);
5679:   }
5680:   if (nsends) {MPI_Waitall(nsends_mpi,swaits,MPI_STATUSES_IGNORE);}
5681:   PetscFree2(rwaits,swaits);

5683:   if (scall == MAT_INITIAL_MATRIX) {
5684:     /* put together the new matrix */
5685:     MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,aBn,B->cmap->N,b_othi,b_othj,b_otha,B_oth);

5687:     /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */
5688:     /* Since these are PETSc arrays, change flags to free them as necessary. */
5689:     b_oth          = (Mat_SeqAIJ*)(*B_oth)->data;
5690:     b_oth->free_a  = PETSC_TRUE;
5691:     b_oth->free_ij = PETSC_TRUE;
5692:     b_oth->nonew   = 0;

5694:     PetscFree(bufj);
5695:     if (!startsj_s || !bufa_ptr) {
5696:       PetscFree2(sstartsj,rstartsj);
5697:       PetscFree(bufa_ptr);
5698:     } else {
5699:       *startsj_s = sstartsj;
5700:       *startsj_r = rstartsj;
5701:       *bufa_ptr  = bufa;
5702:     }
5703:   }

5705:   VecScatterRestoreRemote_Private(ctx,PETSC_TRUE,&nsends,&sstarts,&srow,&sprocs,&sbs);
5706:   VecScatterRestoreRemoteOrdered_Private(ctx,PETSC_FALSE,&nrecvs,&rstarts,NULL,&rprocs,&rbs);
5707:   PetscLogEventEnd(MAT_GetBrowsOfAocols,A,B,0,0);
5708:   return(0);
5709: }

5711: /*@C
5712:   MatGetCommunicationStructs - Provides access to the communication structures used in matrix-vector multiplication.

5714:   Not Collective

5716:   Input Parameters:
5717: . A - The matrix in mpiaij format

5719:   Output Parameter:
5720: + lvec - The local vector holding off-process values from the argument to a matrix-vector product
5721: . colmap - A map from global column index to local index into lvec
5722: - multScatter - A scatter from the argument of a matrix-vector product to lvec

5724:   Level: developer

5726: @*/
5727: #if defined(PETSC_USE_CTABLE)
5728: PetscErrorCode MatGetCommunicationStructs(Mat A, Vec *lvec, PetscTable *colmap, VecScatter *multScatter)
5729: #else
5730: PetscErrorCode MatGetCommunicationStructs(Mat A, Vec *lvec, PetscInt *colmap[], VecScatter *multScatter)
5731: #endif
5732: {
5733:   Mat_MPIAIJ *a;

5740:   a = (Mat_MPIAIJ*) A->data;
5741:   if (lvec) *lvec = a->lvec;
5742:   if (colmap) *colmap = a->colmap;
5743:   if (multScatter) *multScatter = a->Mvctx;
5744:   return(0);
5745: }

5747: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJCRL(Mat,MatType,MatReuse,Mat*);
5748: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJPERM(Mat,MatType,MatReuse,Mat*);
5749: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJSELL(Mat,MatType,MatReuse,Mat*);
5750: #if defined(PETSC_HAVE_MKL_SPARSE)
5751: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJMKL(Mat,MatType,MatReuse,Mat*);
5752: #endif
5753: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIBAIJ(Mat,MatType,MatReuse,Mat*);
5754: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPISBAIJ(Mat,MatType,MatReuse,Mat*);
5755: #if defined(PETSC_HAVE_ELEMENTAL)
5756: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_Elemental(Mat,MatType,MatReuse,Mat*);
5757: #endif
5758: #if defined(PETSC_HAVE_HYPRE)
5759: PETSC_INTERN PetscErrorCode MatConvert_AIJ_HYPRE(Mat,MatType,MatReuse,Mat*);
5760: #endif
5761: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPISELL(Mat,MatType,MatReuse,Mat*);
5762: PETSC_INTERN PetscErrorCode MatConvert_XAIJ_IS(Mat,MatType,MatReuse,Mat*);
5763: PETSC_INTERN PetscErrorCode MatProductSetFromOptions_IS_XAIJ(Mat);

5765: /*
5766:     Computes (B'*A')' since computing B*A directly is untenable

5768:                n                       p                          p
5769:         (              )       (              )         (                  )
5770:       m (      A       )  *  n (       B      )   =   m (         C        )
5771:         (              )       (              )         (                  )

5773: */
5774: PetscErrorCode MatMatMultNumeric_MPIDense_MPIAIJ(Mat A,Mat B,Mat C)
5775: {
5777:   Mat            At,Bt,Ct;

5780:   MatTranspose(A,MAT_INITIAL_MATRIX,&At);
5781:   MatTranspose(B,MAT_INITIAL_MATRIX,&Bt);
5782:   MatMatMult(Bt,At,MAT_INITIAL_MATRIX,1.0,&Ct);
5783:   MatDestroy(&At);
5784:   MatDestroy(&Bt);
5785:   MatTranspose(Ct,MAT_REUSE_MATRIX,&C);
5786:   MatDestroy(&Ct);
5787:   return(0);
5788: }

5790: PetscErrorCode MatMatMultSymbolic_MPIDense_MPIAIJ(Mat A,Mat B,PetscReal fill,Mat C)
5791: {
5793:   PetscInt       m=A->rmap->n,n=B->cmap->n;

5796:   if (A->cmap->n != B->rmap->n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"A->cmap->n %d != B->rmap->n %d\n",A->cmap->n,B->rmap->n);
5797:   MatSetSizes(C,m,n,PETSC_DETERMINE,PETSC_DETERMINE);
5798:   MatSetBlockSizesFromMats(C,A,B);
5799:   MatSetType(C,MATMPIDENSE);
5800:   MatMPIDenseSetPreallocation(C,NULL);
5801:   MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);
5802:   MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);

5804:   C->ops->matmultnumeric = MatMatMultNumeric_MPIDense_MPIAIJ;
5805:   return(0);
5806: }

5808: /* ----------------------------------------------------------------*/
5809: static PetscErrorCode MatProductSetFromOptions_MPIDense_MPIAIJ_AB(Mat C)
5810: {
5811:   Mat_Product *product = C->product;
5812:   Mat         A = product->A,B=product->B;

5815:   if (A->cmap->rstart != B->rmap->rstart || A->cmap->rend != B->rmap->rend)
5816:     SETERRQ4(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Matrix local dimensions are incompatible, (%D, %D) != (%D,%D)",A->cmap->rstart,A->cmap->rend,B->rmap->rstart,B->rmap->rend);

5818:   C->ops->matmultsymbolic = MatMatMultSymbolic_MPIDense_MPIAIJ;
5819:   C->ops->productsymbolic = MatProductSymbolic_AB;
5820:   return(0);
5821: }

5823: PETSC_INTERN PetscErrorCode MatProductSetFromOptions_MPIDense_MPIAIJ(Mat C)
5824: {
5826:   Mat_Product    *product = C->product;

5829:   if (product->type == MATPRODUCT_AB) {
5830:     MatProductSetFromOptions_MPIDense_MPIAIJ_AB(C);
5831:   } else SETERRQ1(PetscObjectComm((PetscObject)C),PETSC_ERR_SUP,"MatProduct type %s is not supported for MPIDense and MPIAIJ matrices",MatProductTypes[product->type]);
5832:   return(0);
5833: }
5834: /* ----------------------------------------------------------------*/

5836: /*MC
5837:    MATMPIAIJ - MATMPIAIJ = "mpiaij" - A matrix type to be used for parallel sparse matrices.

5839:    Options Database Keys:
5840: . -mat_type mpiaij - sets the matrix type to "mpiaij" during a call to MatSetFromOptions()

5842:    Level: beginner

5844:    Notes:
5845:     MatSetValues() may be called for this matrix type with a NULL argument for the numerical values,
5846:     in this case the values associated with the rows and columns one passes in are set to zero
5847:     in the matrix

5849:     MatSetOptions(,MAT_STRUCTURE_ONLY,PETSC_TRUE) may be called for this matrix type. In this no
5850:     space is allocated for the nonzero entries and any entries passed with MatSetValues() are ignored

5852: .seealso: MatCreateAIJ()
5853: M*/

5855: PETSC_EXTERN PetscErrorCode MatCreate_MPIAIJ(Mat B)
5856: {
5857:   Mat_MPIAIJ     *b;
5859:   PetscMPIInt    size;

5862:   MPI_Comm_size(PetscObjectComm((PetscObject)B),&size);

5864:   PetscNewLog(B,&b);
5865:   B->data       = (void*)b;
5866:   PetscMemcpy(B->ops,&MatOps_Values,sizeof(struct _MatOps));
5867:   B->assembled  = PETSC_FALSE;
5868:   B->insertmode = NOT_SET_VALUES;
5869:   b->size       = size;

5871:   MPI_Comm_rank(PetscObjectComm((PetscObject)B),&b->rank);

5873:   /* build cache for off array entries formed */
5874:   MatStashCreate_Private(PetscObjectComm((PetscObject)B),1,&B->stash);

5876:   b->donotstash  = PETSC_FALSE;
5877:   b->colmap      = 0;
5878:   b->garray      = 0;
5879:   b->roworiented = PETSC_TRUE;

5881:   /* stuff used for matrix vector multiply */
5882:   b->lvec  = NULL;
5883:   b->Mvctx = NULL;

5885:   /* stuff for MatGetRow() */
5886:   b->rowindices   = 0;
5887:   b->rowvalues    = 0;
5888:   b->getrowactive = PETSC_FALSE;

5890:   /* flexible pointer used in CUSP/CUSPARSE classes */
5891:   b->spptr = NULL;

5893:   PetscObjectComposeFunction((PetscObject)B,"MatMPIAIJSetUseScalableIncreaseOverlap_C",MatMPIAIJSetUseScalableIncreaseOverlap_MPIAIJ);
5894:   PetscObjectComposeFunction((PetscObject)B,"MatStoreValues_C",MatStoreValues_MPIAIJ);
5895:   PetscObjectComposeFunction((PetscObject)B,"MatRetrieveValues_C",MatRetrieveValues_MPIAIJ);
5896:   PetscObjectComposeFunction((PetscObject)B,"MatIsTranspose_C",MatIsTranspose_MPIAIJ);
5897:   PetscObjectComposeFunction((PetscObject)B,"MatMPIAIJSetPreallocation_C",MatMPIAIJSetPreallocation_MPIAIJ);
5898:   PetscObjectComposeFunction((PetscObject)B,"MatResetPreallocation_C",MatResetPreallocation_MPIAIJ);
5899:   PetscObjectComposeFunction((PetscObject)B,"MatMPIAIJSetPreallocationCSR_C",MatMPIAIJSetPreallocationCSR_MPIAIJ);
5900:   PetscObjectComposeFunction((PetscObject)B,"MatDiagonalScaleLocal_C",MatDiagonalScaleLocal_MPIAIJ);
5901:   PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_mpiaijperm_C",MatConvert_MPIAIJ_MPIAIJPERM);
5902:   PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_mpiaijsell_C",MatConvert_MPIAIJ_MPIAIJSELL);
5903: #if defined(PETSC_HAVE_MKL_SPARSE)
5904:   PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_mpiaijmkl_C",MatConvert_MPIAIJ_MPIAIJMKL);
5905: #endif
5906:   PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_mpiaijcrl_C",MatConvert_MPIAIJ_MPIAIJCRL);
5907:   PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_mpibaij_C",MatConvert_MPIAIJ_MPIBAIJ);
5908:   PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_mpisbaij_C",MatConvert_MPIAIJ_MPISBAIJ);
5909: #if defined(PETSC_HAVE_ELEMENTAL)
5910:   PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_elemental_C",MatConvert_MPIAIJ_Elemental);
5911: #endif
5912:   PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_is_C",MatConvert_XAIJ_IS);
5913:   PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_mpisell_C",MatConvert_MPIAIJ_MPISELL);
5914:   PetscObjectComposeFunction((PetscObject)B,"MatMatMultSymbolic_mpidense_mpiaij_C",MatMatMultSymbolic_MPIDense_MPIAIJ);
5915:   PetscObjectComposeFunction((PetscObject)B,"MatMatMultNumeric_mpidense_mpiaij_C",MatMatMultNumeric_MPIDense_MPIAIJ);
5916: #if defined(PETSC_HAVE_HYPRE)
5917:   PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_hypre_C",MatConvert_AIJ_HYPRE);
5918:   PetscObjectComposeFunction((PetscObject)B,"MatProductSetFromOptions_transpose_mpiaij_mpiaij_C",MatProductSetFromOptions_Transpose_AIJ_AIJ);
5919: #endif
5920:   PetscObjectComposeFunction((PetscObject)B,"MatProductSetFromOptions_is_mpiaij_C",MatProductSetFromOptions_IS_XAIJ);
5921:   PetscObjectComposeFunction((PetscObject)B,"MatProductSetFromOptions_mpiaij_mpiaij_C",MatProductSetFromOptions_MPIAIJ);
5922:   PetscObjectChangeTypeName((PetscObject)B,MATMPIAIJ);
5923:   return(0);
5924: }

5926: /*@C
5927:      MatCreateMPIAIJWithSplitArrays - creates a MPI AIJ matrix using arrays that contain the "diagonal"
5928:          and "off-diagonal" part of the matrix in CSR format.

5930:    Collective

5932:    Input Parameters:
5933: +  comm - MPI communicator
5934: .  m - number of local rows (Cannot be PETSC_DECIDE)
5935: .  n - This value should be the same as the local size used in creating the
5936:        x vector for the matrix-vector product y = Ax. (or PETSC_DECIDE to have
5937:        calculated if N is given) For square matrices n is almost always m.
5938: .  M - number of global rows (or PETSC_DETERMINE to have calculated if m is given)
5939: .  N - number of global columns (or PETSC_DETERMINE to have calculated if n is given)
5940: .   i - row indices for "diagonal" portion of matrix; that is i[0] = 0, i[row] = i[row-1] + number of elements in that row of the matrix
5941: .   j - column indices
5942: .   a - matrix values
5943: .   oi - row indices for "off-diagonal" portion of matrix; that is oi[0] = 0, oi[row] = oi[row-1] + number of elements in that row of the matrix
5944: .   oj - column indices
5945: -   oa - matrix values

5947:    Output Parameter:
5948: .   mat - the matrix

5950:    Level: advanced

5952:    Notes:
5953:        The i, j, and a arrays ARE NOT copied by this routine into the internal format used by PETSc. The user
5954:        must free the arrays once the matrix has been destroyed and not before.

5956:        The i and j indices are 0 based

5958:        See MatCreateAIJ() for the definition of "diagonal" and "off-diagonal" portion of the matrix

5960:        This sets local rows and cannot be used to set off-processor values.

5962:        Use of this routine is discouraged because it is inflexible and cumbersome to use. It is extremely rare that a
5963:        legacy Section 1.5 Writing Application Codes with PETSc natively assembles into exactly this split format. The code to do so is nontrivial and does
5964:        not easily support in-place reassembly. It is recommended to use MatSetValues() (or a variant thereof) because
5965:        the resulting assembly is easier to implement, will work with any matrix format, and the user does not have to
5966:        keep track of the underlying array. Use MatSetOption(A,MAT_NO_OFF_PROC_ENTRIES,PETSC_TRUE) to disable all
5967:        communication if it is known that only local entries will be set.

5969: .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatMPIAIJSetPreallocationCSR(),
5970:           MATMPIAIJ, MatCreateAIJ(), MatCreateMPIAIJWithArrays()
5971: @*/
5972: PetscErrorCode MatCreateMPIAIJWithSplitArrays(MPI_Comm comm,PetscInt m,PetscInt n,PetscInt M,PetscInt N,PetscInt i[],PetscInt j[],PetscScalar a[],PetscInt oi[], PetscInt oj[],PetscScalar oa[],Mat *mat)
5973: {
5975:   Mat_MPIAIJ     *maij;

5978:   if (m < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"local number of rows (m) cannot be PETSC_DECIDE, or negative");
5979:   if (i[0]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"i (row indices) must start with 0");
5980:   if (oi[0]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"oi (row indices) must start with 0");
5981:   MatCreate(comm,mat);
5982:   MatSetSizes(*mat,m,n,M,N);
5983:   MatSetType(*mat,MATMPIAIJ);
5984:   maij = (Mat_MPIAIJ*) (*mat)->data;

5986:   (*mat)->preallocated = PETSC_TRUE;

5988:   PetscLayoutSetUp((*mat)->rmap);
5989:   PetscLayoutSetUp((*mat)->cmap);

5991:   MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,m,n,i,j,a,&maij->A);
5992:   MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,m,(*mat)->cmap->N,oi,oj,oa,&maij->B);

5994:   MatAssemblyBegin(maij->A,MAT_FINAL_ASSEMBLY);
5995:   MatAssemblyEnd(maij->A,MAT_FINAL_ASSEMBLY);
5996:   MatAssemblyBegin(maij->B,MAT_FINAL_ASSEMBLY);
5997:   MatAssemblyEnd(maij->B,MAT_FINAL_ASSEMBLY);

5999:   MatSetOption(*mat,MAT_NO_OFF_PROC_ENTRIES,PETSC_TRUE);
6000:   MatAssemblyBegin(*mat,MAT_FINAL_ASSEMBLY);
6001:   MatAssemblyEnd(*mat,MAT_FINAL_ASSEMBLY);
6002:   MatSetOption(*mat,MAT_NO_OFF_PROC_ENTRIES,PETSC_FALSE);
6003:   MatSetOption(*mat,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE);
6004:   return(0);
6005: }

6007: /*
6008:     Special version for direct calls from Fortran
6009: */
6010:  #include <petsc/private/fortranimpl.h>

6012: /* Change these macros so can be used in void function */
6013: #undef CHKERRQ
6014: #define CHKERRQ(ierr) CHKERRABORT(PETSC_COMM_WORLD,ierr)
6015: #undef SETERRQ2
6016: #define SETERRQ2(comm,ierr,b,c,d) CHKERRABORT(comm,ierr)
6017: #undef SETERRQ3
6018: #define SETERRQ3(comm,ierr,b,c,d,e) CHKERRABORT(comm,ierr)
6019: #undef SETERRQ
6020: #define SETERRQ(c,ierr,b) CHKERRABORT(c,ierr)

6022: #if defined(PETSC_HAVE_FORTRAN_CAPS)
6023: #define matsetvaluesmpiaij_ MATSETVALUESMPIAIJ
6024: #elif !defined(PETSC_HAVE_FORTRAN_UNDERSCORE)
6025: #define matsetvaluesmpiaij_ matsetvaluesmpiaij
6026: #else
6027: #endif
6028: PETSC_EXTERN void matsetvaluesmpiaij_(Mat *mmat,PetscInt *mm,const PetscInt im[],PetscInt *mn,const PetscInt in[],const PetscScalar v[],InsertMode *maddv,PetscErrorCode *_ierr)
6029: {
6030:   Mat            mat  = *mmat;
6031:   PetscInt       m    = *mm, n = *mn;
6032:   InsertMode     addv = *maddv;
6033:   Mat_MPIAIJ     *aij = (Mat_MPIAIJ*)mat->data;
6034:   PetscScalar    value;

6037:   MatCheckPreallocated(mat,1);
6038:   if (mat->insertmode == NOT_SET_VALUES) mat->insertmode = addv;

6040: #if defined(PETSC_USE_DEBUG)
6041:   else if (mat->insertmode != addv) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Cannot mix add values and insert values");
6042: #endif
6043:   {
6044:     PetscInt  i,j,rstart  = mat->rmap->rstart,rend = mat->rmap->rend;
6045:     PetscInt  cstart      = mat->cmap->rstart,cend = mat->cmap->rend,row,col;
6046:     PetscBool roworiented = aij->roworiented;

6048:     /* Some Variables required in the macro */
6049:     Mat        A                    = aij->A;
6050:     Mat_SeqAIJ *a                   = (Mat_SeqAIJ*)A->data;
6051:     PetscInt   *aimax               = a->imax,*ai = a->i,*ailen = a->ilen,*aj = a->j;
6052:     MatScalar  *aa                  = a->a;
6053:     PetscBool  ignorezeroentries    = (((a->ignorezeroentries)&&(addv==ADD_VALUES)) ? PETSC_TRUE : PETSC_FALSE);
6054:     Mat        B                    = aij->B;
6055:     Mat_SeqAIJ *b                   = (Mat_SeqAIJ*)B->data;
6056:     PetscInt   *bimax               = b->imax,*bi = b->i,*bilen = b->ilen,*bj = b->j,bm = aij->B->rmap->n,am = aij->A->rmap->n;
6057:     MatScalar  *ba                  = b->a;
6058:     /* This variable below is only for the PETSC_HAVE_VIENNACL or PETSC_HAVE_CUDA cases, but we define it in all cases because we
6059:      * cannot use "#if defined" inside a macro. */
6060:     PETSC_UNUSED PetscBool inserted = PETSC_FALSE;

6062:     PetscInt  *rp1,*rp2,ii,nrow1,nrow2,_i,rmax1,rmax2,N,low1,high1,low2,high2,t,lastcol1,lastcol2;
6063:     PetscInt  nonew = a->nonew;
6064:     MatScalar *ap1,*ap2;

6067:     for (i=0; i<m; i++) {
6068:       if (im[i] < 0) continue;
6069: #if defined(PETSC_USE_DEBUG)
6070:       if (im[i] >= mat->rmap->N) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Row too large: row %D max %D",im[i],mat->rmap->N-1);
6071: #endif
6072:       if (im[i] >= rstart && im[i] < rend) {
6073:         row      = im[i] - rstart;
6074:         lastcol1 = -1;
6075:         rp1      = aj + ai[row];
6076:         ap1      = aa + ai[row];
6077:         rmax1    = aimax[row];
6078:         nrow1    = ailen[row];
6079:         low1     = 0;
6080:         high1    = nrow1;
6081:         lastcol2 = -1;
6082:         rp2      = bj + bi[row];
6083:         ap2      = ba + bi[row];
6084:         rmax2    = bimax[row];
6085:         nrow2    = bilen[row];
6086:         low2     = 0;
6087:         high2    = nrow2;

6089:         for (j=0; j<n; j++) {
6090:           if (roworiented) value = v[i*n+j];
6091:           else value = v[i+j*m];
6092:           if (ignorezeroentries && value == 0.0 && (addv == ADD_VALUES) && im[i] != in[j]) continue;
6093:           if (in[j] >= cstart && in[j] < cend) {
6094:             col = in[j] - cstart;
6095:             MatSetValues_SeqAIJ_A_Private(row,col,value,addv,im[i],in[j]);
6096: #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA)
6097:             if (A->offloadmask != PETSC_OFFLOAD_UNALLOCATED && inserted) A->offloadmask = PETSC_OFFLOAD_CPU;
6098: #endif
6099:           } else if (in[j] < 0) continue;
6100: #if defined(PETSC_USE_DEBUG)
6101:           /* extra brace on SETERRQ2() is required for --with-errorchecking=0 - due to the next 'else' clause */
6102:           else if (in[j] >= mat->cmap->N) {SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Column too large: col %D max %D",in[j],mat->cmap->N-1);}
6103: #endif
6104:           else {
6105:             if (mat->was_assembled) {
6106:               if (!aij->colmap) {
6107:                 MatCreateColmap_MPIAIJ_Private(mat);
6108:               }
6109: #if defined(PETSC_USE_CTABLE)
6110:               PetscTableFind(aij->colmap,in[j]+1,&col);
6111:               col--;
6112: #else
6113:               col = aij->colmap[in[j]] - 1;
6114: #endif
6115:               if (col < 0 && !((Mat_SeqAIJ*)(aij->A->data))->nonew) {
6116:                 MatDisAssemble_MPIAIJ(mat);
6117:                 col  =  in[j];
6118:                 /* Reinitialize the variables required by MatSetValues_SeqAIJ_B_Private() */
6119:                 B        = aij->B;
6120:                 b        = (Mat_SeqAIJ*)B->data;
6121:                 bimax    = b->imax; bi = b->i; bilen = b->ilen; bj = b->j;
6122:                 rp2      = bj + bi[row];
6123:                 ap2      = ba + bi[row];
6124:                 rmax2    = bimax[row];
6125:                 nrow2    = bilen[row];
6126:                 low2     = 0;
6127:                 high2    = nrow2;
6128:                 bm       = aij->B->rmap->n;
6129:                 ba       = b->a;
6130:                 inserted = PETSC_FALSE;
6131:               }
6132:             } else col = in[j];
6133:             MatSetValues_SeqAIJ_B_Private(row,col,value,addv,im[i],in[j]);
6134: #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA)
6135:             if (B->offloadmask != PETSC_OFFLOAD_UNALLOCATED && inserted) B->offloadmask = PETSC_OFFLOAD_CPU;
6136: #endif
6137:           }
6138:         }
6139:       } else if (!aij->donotstash) {
6140:         if (roworiented) {
6141:           MatStashValuesRow_Private(&mat->stash,im[i],n,in,v+i*n,(PetscBool)(ignorezeroentries && (addv == ADD_VALUES)));
6142:         } else {
6143:           MatStashValuesCol_Private(&mat->stash,im[i],n,in,v+i,m,(PetscBool)(ignorezeroentries && (addv == ADD_VALUES)));
6144:         }
6145:       }
6146:     }
6147:   }
6148:   PetscFunctionReturnVoid();
6149: }