Actual source code: mpiaij.c

petsc-3.3-p2 2012-07-13
  2: #include <../src/mat/impls/aij/mpi/mpiaij.h>   /*I "petscmat.h" I*/
  3: #include <petscblaslapack.h>

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

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

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

 17:   Developer Notes: Subclasses include MATAIJCUSP, MATAIJPERM, MATAIJCRL, and also automatically switches over to use inodes when 
 18:    enough exist.

 20:   Level: beginner

 22: .seealso: MatCreateAIJ(), MatCreateSeqAIJ(), MATSEQAIJ,MATMPIAIJ
 23: M*/

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

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

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

 37:   Level: beginner

 39: .seealso: MatCreateMPIAIJCRL,MATSEQAIJCRL,MATMPIAIJCRL, MATSEQAIJCRL, MATMPIAIJCRL
 40: M*/

 44: PetscErrorCode MatFindNonZeroRows_MPIAIJ(Mat M,IS *keptrows)
 45: {
 46:   PetscErrorCode  ierr;
 47:   Mat_MPIAIJ      *mat = (Mat_MPIAIJ*)M->data;
 48:   Mat_SeqAIJ      *a = (Mat_SeqAIJ*)mat->A->data;
 49:   Mat_SeqAIJ      *b = (Mat_SeqAIJ*)mat->B->data;
 50:   const PetscInt  *ia,*ib;
 51:   const MatScalar *aa,*bb;
 52:   PetscInt        na,nb,i,j,*rows,cnt=0,n0rows;
 53:   PetscInt        m = M->rmap->n,rstart = M->rmap->rstart;

 56:   *keptrows = 0;
 57:   ia = a->i;
 58:   ib = b->i;
 59:   for (i=0; i<m; i++) {
 60:     na = ia[i+1] - ia[i];
 61:     nb = ib[i+1] - ib[i];
 62:     if (!na && !nb) {
 63:       cnt++;
 64:       goto ok1;
 65:     }
 66:     aa = a->a + ia[i];
 67:     for (j=0; j<na; j++) {
 68:       if (aa[j] != 0.0) goto ok1;
 69:     }
 70:     bb = b->a + ib[i];
 71:     for (j=0; j <nb; j++) {
 72:       if (bb[j] != 0.0) goto ok1;
 73:     }
 74:     cnt++;
 75:     ok1:;
 76:   }
 77:   MPI_Allreduce(&cnt,&n0rows,1,MPIU_INT,MPI_SUM,((PetscObject)M)->comm);
 78:   if (!n0rows) return(0);
 79:   PetscMalloc((M->rmap->n-cnt)*sizeof(PetscInt),&rows);
 80:   cnt = 0;
 81:   for (i=0; i<m; i++) {
 82:     na = ia[i+1] - ia[i];
 83:     nb = ib[i+1] - ib[i];
 84:     if (!na && !nb) continue;
 85:     aa = a->a + ia[i];
 86:     for(j=0; j<na;j++) {
 87:       if (aa[j] != 0.0) {
 88:         rows[cnt++] = rstart + i;
 89:         goto ok2;
 90:       }
 91:     }
 92:     bb = b->a + ib[i];
 93:     for (j=0; j<nb; j++) {
 94:       if (bb[j] != 0.0) {
 95:         rows[cnt++] = rstart + i;
 96:         goto ok2;
 97:       }
 98:     }
 99:     ok2:;
100:   }
101:   ISCreateGeneral(((PetscObject)M)->comm,cnt,rows,PETSC_OWN_POINTER,keptrows);
102:   return(0);
103: }

107: PetscErrorCode MatGetColumnNorms_MPIAIJ(Mat A,NormType type,PetscReal *norms)
108: {
110:   Mat_MPIAIJ     *aij = (Mat_MPIAIJ*)A->data;
111:   PetscInt       i,n,*garray = aij->garray;
112:   Mat_SeqAIJ     *a_aij = (Mat_SeqAIJ*) aij->A->data;
113:   Mat_SeqAIJ     *b_aij = (Mat_SeqAIJ*) aij->B->data;
114:   PetscReal      *work;

117:   MatGetSize(A,PETSC_NULL,&n);
118:   PetscMalloc(n*sizeof(PetscReal),&work);
119:   PetscMemzero(work,n*sizeof(PetscReal));
120:   if (type == NORM_2) {
121:     for (i=0; i<a_aij->i[aij->A->rmap->n]; i++) {
122:       work[A->cmap->rstart + a_aij->j[i]] += PetscAbsScalar(a_aij->a[i]*a_aij->a[i]);
123:     }
124:     for (i=0; i<b_aij->i[aij->B->rmap->n]; i++) {
125:       work[garray[b_aij->j[i]]] += PetscAbsScalar(b_aij->a[i]*b_aij->a[i]);
126:     }
127:   } else if (type == NORM_1) {
128:     for (i=0; i<a_aij->i[aij->A->rmap->n]; i++) {
129:       work[A->cmap->rstart + a_aij->j[i]] += PetscAbsScalar(a_aij->a[i]);
130:     }
131:     for (i=0; i<b_aij->i[aij->B->rmap->n]; i++) {
132:       work[garray[b_aij->j[i]]] += PetscAbsScalar(b_aij->a[i]);
133:     }
134:   } else if (type == NORM_INFINITY) {
135:     for (i=0; i<a_aij->i[aij->A->rmap->n]; i++) {
136:       work[A->cmap->rstart + a_aij->j[i]] = PetscMax(PetscAbsScalar(a_aij->a[i]), work[A->cmap->rstart + a_aij->j[i]]);
137:     }
138:     for (i=0; i<b_aij->i[aij->B->rmap->n]; i++) {
139:       work[garray[b_aij->j[i]]] = PetscMax(PetscAbsScalar(b_aij->a[i]),work[garray[b_aij->j[i]]]);
140:     }

142:   } else SETERRQ(((PetscObject)A)->comm,PETSC_ERR_ARG_WRONG,"Unknown NormType");
143:   if (type == NORM_INFINITY) {
144:     MPI_Allreduce(work,norms,n,MPIU_REAL,MPIU_MAX,A->hdr.comm);
145:   } else {
146:     MPI_Allreduce(work,norms,n,MPIU_REAL,MPIU_SUM,A->hdr.comm);
147:   }
148:   PetscFree(work);
149:   if (type == NORM_2) {
150:     for (i=0; i<n; i++) norms[i] = PetscSqrtReal(norms[i]);
151:   }
152:   return(0);
153: }

157: /*
158:     Distributes a SeqAIJ matrix across a set of processes. Code stolen from
159:     MatLoad_MPIAIJ(). Horrible lack of reuse. Should be a routine for each matrix type.

161:     Only for square matrices
162: */
163: PetscErrorCode MatDistribute_MPIAIJ(MPI_Comm comm,Mat gmat,PetscInt m,MatReuse reuse,Mat *inmat)
164: {
165:   PetscMPIInt    rank,size;
166:   PetscInt       *rowners,*dlens,*olens,i,rstart,rend,j,jj,nz,*gmataj,cnt,row,*ld;
168:   Mat            mat;
169:   Mat_SeqAIJ     *gmata;
170:   PetscMPIInt    tag;
171:   MPI_Status     status;
172:   PetscBool      aij;
173:   MatScalar      *gmataa,*ao,*ad,*gmataarestore=0;

176:   CHKMEMQ;
177:   MPI_Comm_rank(comm,&rank);
178:   MPI_Comm_size(comm,&size);
179:   if (!rank) {
180:     PetscObjectTypeCompare((PetscObject)gmat,MATSEQAIJ,&aij);
181:     if (!aij) SETERRQ1(((PetscObject)gmat)->comm,PETSC_ERR_SUP,"Currently no support for input matrix of type %s\n",((PetscObject)gmat)->type_name);
182:   }
183:   if (reuse == MAT_INITIAL_MATRIX) {
184:     MatCreate(comm,&mat);
185:     MatSetSizes(mat,m,m,PETSC_DETERMINE,PETSC_DETERMINE);
186:     MatSetBlockSizes(mat,gmat->rmap->bs,gmat->cmap->bs);
187:     MatSetType(mat,MATAIJ);
188:     PetscMalloc((size+1)*sizeof(PetscInt),&rowners);
189:     PetscMalloc2(m,PetscInt,&dlens,m,PetscInt,&olens);
190:     MPI_Allgather(&m,1,MPIU_INT,rowners+1,1,MPIU_INT,comm);
191:     rowners[0] = 0;
192:     for (i=2; i<=size; i++) {
193:       rowners[i] += rowners[i-1];
194:     }
195:     rstart = rowners[rank];
196:     rend   = rowners[rank+1];
197:     PetscObjectGetNewTag((PetscObject)mat,&tag);
198:     if (!rank) {
199:       gmata = (Mat_SeqAIJ*) gmat->data;
200:       /* send row lengths to all processors */
201:       for (i=0; i<m; i++) dlens[i] = gmata->ilen[i];
202:       for (i=1; i<size; i++) {
203:         MPI_Send(gmata->ilen + rowners[i],rowners[i+1]-rowners[i],MPIU_INT,i,tag,comm);
204:       }
205:       /* determine number diagonal and off-diagonal counts */
206:       PetscMemzero(olens,m*sizeof(PetscInt));
207:       PetscMalloc(m*sizeof(PetscInt),&ld);
208:       PetscMemzero(ld,m*sizeof(PetscInt));
209:       jj = 0;
210:       for (i=0; i<m; i++) {
211:         for (j=0; j<dlens[i]; j++) {
212:           if (gmata->j[jj] < rstart) ld[i]++;
213:           if (gmata->j[jj] < rstart || gmata->j[jj] >= rend) olens[i]++;
214:           jj++;
215:         }
216:       }
217:       /* send column indices to other processes */
218:       for (i=1; i<size; i++) {
219:         nz   = gmata->i[rowners[i+1]]-gmata->i[rowners[i]];
220:         MPI_Send(&nz,1,MPIU_INT,i,tag,comm);
221:         MPI_Send(gmata->j + gmata->i[rowners[i]],nz,MPIU_INT,i,tag,comm);
222:       }

224:       /* send numerical values to other processes */
225:       for (i=1; i<size; i++) {
226:         nz   = gmata->i[rowners[i+1]]-gmata->i[rowners[i]];
227:         MPI_Send(gmata->a + gmata->i[rowners[i]],nz,MPIU_SCALAR,i,tag,comm);
228:       }
229:       gmataa = gmata->a;
230:       gmataj = gmata->j;

232:     } else {
233:       /* receive row lengths */
234:       MPI_Recv(dlens,m,MPIU_INT,0,tag,comm,&status);
235:       /* receive column indices */
236:       MPI_Recv(&nz,1,MPIU_INT,0,tag,comm,&status);
237:       PetscMalloc2(nz,PetscScalar,&gmataa,nz,PetscInt,&gmataj);
238:       MPI_Recv(gmataj,nz,MPIU_INT,0,tag,comm,&status);
239:       /* determine number diagonal and off-diagonal counts */
240:       PetscMemzero(olens,m*sizeof(PetscInt));
241:       PetscMalloc(m*sizeof(PetscInt),&ld);
242:       PetscMemzero(ld,m*sizeof(PetscInt));
243:       jj = 0;
244:       for (i=0; i<m; i++) {
245:         for (j=0; j<dlens[i]; j++) {
246:           if (gmataj[jj] < rstart) ld[i]++;
247:           if (gmataj[jj] < rstart || gmataj[jj] >= rend) olens[i]++;
248:           jj++;
249:         }
250:       }
251:       /* receive numerical values */
252:       PetscMemzero(gmataa,nz*sizeof(PetscScalar));
253:       MPI_Recv(gmataa,nz,MPIU_SCALAR,0,tag,comm,&status);
254:     }
255:     /* set preallocation */
256:     for (i=0; i<m; i++) {
257:       dlens[i] -= olens[i];
258:     }
259:     MatSeqAIJSetPreallocation(mat,0,dlens);
260:     MatMPIAIJSetPreallocation(mat,0,dlens,0,olens);
261: 
262:     for (i=0; i<m; i++) {
263:       dlens[i] += olens[i];
264:     }
265:     cnt  = 0;
266:     for (i=0; i<m; i++) {
267:       row  = rstart + i;
268:       MatSetValues(mat,1,&row,dlens[i],gmataj+cnt,gmataa+cnt,INSERT_VALUES);
269:       cnt += dlens[i];
270:     }
271:     if (rank) {
272:       PetscFree2(gmataa,gmataj);
273:     }
274:     PetscFree2(dlens,olens);
275:     PetscFree(rowners);
276:     ((Mat_MPIAIJ*)(mat->data))->ld = ld;
277:     *inmat = mat;
278:   } else {   /* column indices are already set; only need to move over numerical values from process 0 */
279:     Mat_SeqAIJ *Ad = (Mat_SeqAIJ*)((Mat_MPIAIJ*)((*inmat)->data))->A->data;
280:     Mat_SeqAIJ *Ao = (Mat_SeqAIJ*)((Mat_MPIAIJ*)((*inmat)->data))->B->data;
281:     mat   = *inmat;
282:     PetscObjectGetNewTag((PetscObject)mat,&tag);
283:     if (!rank) {
284:       /* send numerical values to other processes */
285:       gmata = (Mat_SeqAIJ*) gmat->data;
286:       MatGetOwnershipRanges(mat,(const PetscInt**)&rowners);
287:       gmataa = gmata->a;
288:       for (i=1; i<size; i++) {
289:         nz   = gmata->i[rowners[i+1]]-gmata->i[rowners[i]];
290:         MPI_Send(gmataa + gmata->i[rowners[i]],nz,MPIU_SCALAR,i,tag,comm);
291:       }
292:       nz   = gmata->i[rowners[1]]-gmata->i[rowners[0]];
293:     } else {
294:       /* receive numerical values from process 0*/
295:       nz   = Ad->nz + Ao->nz;
296:       PetscMalloc(nz*sizeof(PetscScalar),&gmataa); gmataarestore = gmataa;
297:       MPI_Recv(gmataa,nz,MPIU_SCALAR,0,tag,comm,&status);
298:     }
299:     /* transfer numerical values into the diagonal A and off diagonal B parts of mat */
300:     ld = ((Mat_MPIAIJ*)(mat->data))->ld;
301:     ad = Ad->a;
302:     ao = Ao->a;
303:     if (mat->rmap->n) {
304:       i  = 0;
305:       nz = ld[i];                                   PetscMemcpy(ao,gmataa,nz*sizeof(PetscScalar)); ao += nz; gmataa += nz;
306:       nz = Ad->i[i+1] - Ad->i[i];                   PetscMemcpy(ad,gmataa,nz*sizeof(PetscScalar)); ad += nz; gmataa += nz;
307:     }
308:     for (i=1; i<mat->rmap->n; i++) {
309:       nz = Ao->i[i] - Ao->i[i-1] - ld[i-1] + ld[i]; PetscMemcpy(ao,gmataa,nz*sizeof(PetscScalar)); ao += nz; gmataa += nz;
310:       nz = Ad->i[i+1] - Ad->i[i];                   PetscMemcpy(ad,gmataa,nz*sizeof(PetscScalar)); ad += nz; gmataa += nz;
311:     }
312:     i--;
313:     if (mat->rmap->n) {
314:       nz = Ao->i[i+1] - Ao->i[i] - ld[i];           PetscMemcpy(ao,gmataa,nz*sizeof(PetscScalar)); ao += nz; gmataa += nz;
315:     }
316:     if (rank) {
317:       PetscFree(gmataarestore);
318:     }
319:   }
320:   MatAssemblyBegin(mat,MAT_FINAL_ASSEMBLY);
321:   MatAssemblyEnd(mat,MAT_FINAL_ASSEMBLY);
322:   CHKMEMQ;
323:   return(0);
324: }

326: /* 
327:   Local utility routine that creates a mapping from the global column 
328: number to the local number in the off-diagonal part of the local 
329: storage of the matrix.  When PETSC_USE_CTABLE is used this is scalable at 
330: a slightly higher hash table cost; without it it is not scalable (each processor
331: has an order N integer array but is fast to acess.
332: */
335: PetscErrorCode MatCreateColmap_MPIAIJ_Private(Mat mat)
336: {
337:   Mat_MPIAIJ     *aij = (Mat_MPIAIJ*)mat->data;
339:   PetscInt       n = aij->B->cmap->n,i;

342:   if (!aij->garray) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"MPIAIJ Matrix was assembled but is missing garray");
343: #if defined (PETSC_USE_CTABLE)
344:   PetscTableCreate(n,mat->cmap->N+1,&aij->colmap);
345:   for (i=0; i<n; i++){
346:     PetscTableAdd(aij->colmap,aij->garray[i]+1,i+1,INSERT_VALUES);
347:   }
348: #else
349:   PetscMalloc((mat->cmap->N+1)*sizeof(PetscInt),&aij->colmap);
350:   PetscLogObjectMemory(mat,mat->cmap->N*sizeof(PetscInt));
351:   PetscMemzero(aij->colmap,mat->cmap->N*sizeof(PetscInt));
352:   for (i=0; i<n; i++) aij->colmap[aij->garray[i]] = i+1;
353: #endif
354:   return(0);
355: }

357: #define MatSetValues_SeqAIJ_A_Private(row,col,value,addv) \
358: { \
359:     if (col <= lastcol1) low1 = 0; else high1 = nrow1; \
360:     lastcol1 = col;\
361:     while (high1-low1 > 5) { \
362:       t = (low1+high1)/2; \
363:       if (rp1[t] > col) high1 = t; \
364:       else             low1  = t; \
365:     } \
366:       for (_i=low1; _i<high1; _i++) { \
367:         if (rp1[_i] > col) break; \
368:         if (rp1[_i] == col) { \
369:           if (addv == ADD_VALUES) ap1[_i] += value;   \
370:           else                    ap1[_i] = value; \
371:           goto a_noinsert; \
372:         } \
373:       }  \
374:       if (value == 0.0 && ignorezeroentries) {low1 = 0; high1 = nrow1;goto a_noinsert;} \
375:       if (nonew == 1) {low1 = 0; high1 = nrow1; goto a_noinsert;}                \
376:       if (nonew == -1) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero (%D, %D) into matrix", row, col); \
377:       MatSeqXAIJReallocateAIJ(A,am,1,nrow1,row,col,rmax1,aa,ai,aj,rp1,ap1,aimax,nonew,MatScalar); \
378:       N = nrow1++ - 1; a->nz++; high1++; \
379:       /* shift up all the later entries in this row */ \
380:       for (ii=N; ii>=_i; ii--) { \
381:         rp1[ii+1] = rp1[ii]; \
382:         ap1[ii+1] = ap1[ii]; \
383:       } \
384:       rp1[_i] = col;  \
385:       ap1[_i] = value;  \
386:       a_noinsert: ; \
387:       ailen[row] = nrow1; \
388: } 


391: #define MatSetValues_SeqAIJ_B_Private(row,col,value,addv) \
392: { \
393:     if (col <= lastcol2) low2 = 0; else high2 = nrow2; \
394:     lastcol2 = col;\
395:     while (high2-low2 > 5) { \
396:       t = (low2+high2)/2; \
397:       if (rp2[t] > col) high2 = t; \
398:       else             low2  = t; \
399:     } \
400:     for (_i=low2; _i<high2; _i++) {                \
401:       if (rp2[_i] > col) break;                        \
402:       if (rp2[_i] == col) {                              \
403:         if (addv == ADD_VALUES) ap2[_i] += value;     \
404:         else                    ap2[_i] = value;      \
405:         goto b_noinsert;                              \
406:       }                                                      \
407:     }                                                              \
408:     if (value == 0.0 && ignorezeroentries) {low2 = 0; high2 = nrow2; goto b_noinsert;} \
409:     if (nonew == 1) {low2 = 0; high2 = nrow2; goto b_noinsert;}                \
410:     if (nonew == -1) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero (%D, %D) into matrix", row, col); \
411:     MatSeqXAIJReallocateAIJ(B,bm,1,nrow2,row,col,rmax2,ba,bi,bj,rp2,ap2,bimax,nonew,MatScalar); \
412:     N = nrow2++ - 1; b->nz++; high2++;                                        \
413:     /* shift up all the later entries in this row */                        \
414:     for (ii=N; ii>=_i; ii--) {                                                \
415:       rp2[ii+1] = rp2[ii];                                                \
416:       ap2[ii+1] = ap2[ii];                                                \
417:     }                                                                        \
418:     rp2[_i] = col;                                                        \
419:     ap2[_i] = value;                                                        \
420:     b_noinsert: ;                                                                \
421:     bilen[row] = nrow2;                                                        \
422: }

426: PetscErrorCode MatSetValuesRow_MPIAIJ(Mat A,PetscInt row,const PetscScalar v[])
427: {
428:   Mat_MPIAIJ     *mat = (Mat_MPIAIJ*)A->data;
429:   Mat_SeqAIJ     *a = (Mat_SeqAIJ*)mat->A->data,*b = (Mat_SeqAIJ*)mat->B->data;
431:   PetscInt       l,*garray = mat->garray,diag;

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

436:   /* find size of row to the left of the diagonal part */
437:   MatGetOwnershipRange(A,&diag,0);
438:   row  = row - diag;
439:   for (l=0; l<b->i[row+1]-b->i[row]; l++) {
440:     if (garray[b->j[b->i[row]+l]] > diag) break;
441:   }
442:   PetscMemcpy(b->a+b->i[row],v,l*sizeof(PetscScalar));

444:   /* diagonal part */
445:   PetscMemcpy(a->a+a->i[row],v+l,(a->i[row+1]-a->i[row])*sizeof(PetscScalar));

447:   /* right of diagonal part */
448:   PetscMemcpy(b->a+b->i[row]+l,v+l+a->i[row+1]-a->i[row],(b->i[row+1]-b->i[row]-l)*sizeof(PetscScalar));
449:   return(0);
450: }

454: PetscErrorCode MatSetValues_MPIAIJ(Mat mat,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],const PetscScalar v[],InsertMode addv)
455: {
456:   Mat_MPIAIJ     *aij = (Mat_MPIAIJ*)mat->data;
457:   PetscScalar    value;
459:   PetscInt       i,j,rstart = mat->rmap->rstart,rend = mat->rmap->rend;
460:   PetscInt       cstart = mat->cmap->rstart,cend = mat->cmap->rend,row,col;
461:   PetscBool      roworiented = aij->roworiented;

463:   /* Some Variables required in the macro */
464:   Mat            A = aij->A;
465:   Mat_SeqAIJ     *a = (Mat_SeqAIJ*)A->data;
466:   PetscInt       *aimax = a->imax,*ai = a->i,*ailen = a->ilen,*aj = a->j;
467:   MatScalar      *aa = a->a;
468:   PetscBool      ignorezeroentries = a->ignorezeroentries;
469:   Mat            B = aij->B;
470:   Mat_SeqAIJ     *b = (Mat_SeqAIJ*)B->data;
471:   PetscInt       *bimax = b->imax,*bi = b->i,*bilen = b->ilen,*bj = b->j,bm = aij->B->rmap->n,am = aij->A->rmap->n;
472:   MatScalar      *ba = b->a;

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

480:   for (i=0; i<m; i++) {
481:     if (im[i] < 0) continue;
482: #if defined(PETSC_USE_DEBUG)
483:     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);
484: #endif
485:     if (im[i] >= rstart && im[i] < rend) {
486:       row      = im[i] - rstart;
487:       lastcol1 = -1;
488:       rp1      = aj + ai[row];
489:       ap1      = aa + ai[row];
490:       rmax1    = aimax[row];
491:       nrow1    = ailen[row];
492:       low1     = 0;
493:       high1    = nrow1;
494:       lastcol2 = -1;
495:       rp2      = bj + bi[row];
496:       ap2      = ba + bi[row];
497:       rmax2    = bimax[row];
498:       nrow2    = bilen[row];
499:       low2     = 0;
500:       high2    = nrow2;

502:       for (j=0; j<n; j++) {
503:         if (v) {if (roworiented) value = v[i*n+j]; else value = v[i+j*m];} else value = 0.0;
504:         if (ignorezeroentries && value == 0.0 && (addv == ADD_VALUES)) continue;
505:         if (in[j] >= cstart && in[j] < cend){
506:           col = in[j] - cstart;
507:           nonew = a->nonew;
508:           MatSetValues_SeqAIJ_A_Private(row,col,value,addv);
509:         } else if (in[j] < 0) continue;
510: #if defined(PETSC_USE_DEBUG)
511:         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);
512: #endif
513:         else {
514:           if (mat->was_assembled) {
515:             if (!aij->colmap) {
516:               MatCreateColmap_MPIAIJ_Private(mat);
517:             }
518: #if defined (PETSC_USE_CTABLE)
519:             PetscTableFind(aij->colmap,in[j]+1,&col);
520:             col--;
521: #else
522:             col = aij->colmap[in[j]] - 1;
523: #endif
524:             if (col < 0 && !((Mat_SeqAIJ*)(aij->A->data))->nonew) {
525:               MatDisAssemble_MPIAIJ(mat);
526:               col =  in[j];
527:               /* Reinitialize the variables required by MatSetValues_SeqAIJ_B_Private() */
528:               B = aij->B;
529:               b = (Mat_SeqAIJ*)B->data;
530:               bimax = b->imax; bi = b->i; bilen = b->ilen; bj = b->j; ba = b->a;
531:               rp2      = bj + bi[row];
532:               ap2      = ba + bi[row];
533:               rmax2    = bimax[row];
534:               nrow2    = bilen[row];
535:               low2     = 0;
536:               high2    = nrow2;
537:               bm       = aij->B->rmap->n;
538:               ba = b->a;
539:             }
540:           } else col = in[j];
541:           nonew = b->nonew;
542:           MatSetValues_SeqAIJ_B_Private(row,col,value,addv);
543:         }
544:       }
545:     } else {
546:       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]);
547:       if (!aij->donotstash) {
548:         mat->assembled = PETSC_FALSE;
549:         if (roworiented) {
550:           MatStashValuesRow_Private(&mat->stash,im[i],n,in,v+i*n,(PetscBool)(ignorezeroentries && (addv == ADD_VALUES)));
551:         } else {
552:           MatStashValuesCol_Private(&mat->stash,im[i],n,in,v+i,m,(PetscBool)(ignorezeroentries && (addv == ADD_VALUES)));
553:         }
554:       }
555:     }
556:   }
557:   return(0);
558: }

562: PetscErrorCode MatGetValues_MPIAIJ(Mat mat,PetscInt m,const PetscInt idxm[],PetscInt n,const PetscInt idxn[],PetscScalar v[])
563: {
564:   Mat_MPIAIJ     *aij = (Mat_MPIAIJ*)mat->data;
566:   PetscInt       i,j,rstart = mat->rmap->rstart,rend = mat->rmap->rend;
567:   PetscInt       cstart = mat->cmap->rstart,cend = mat->cmap->rend,row,col;

570:   for (i=0; i<m; i++) {
571:     if (idxm[i] < 0) continue; /* SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative row: %D",idxm[i]);*/
572:     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);
573:     if (idxm[i] >= rstart && idxm[i] < rend) {
574:       row = idxm[i] - rstart;
575:       for (j=0; j<n; j++) {
576:         if (idxn[j] < 0) continue; /* SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative column: %D",idxn[j]); */
577:         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);
578:         if (idxn[j] >= cstart && idxn[j] < cend){
579:           col = idxn[j] - cstart;
580:           MatGetValues(aij->A,1,&row,1,&col,v+i*n+j);
581:         } else {
582:           if (!aij->colmap) {
583:             MatCreateColmap_MPIAIJ_Private(mat);
584:           }
585: #if defined (PETSC_USE_CTABLE)
586:           PetscTableFind(aij->colmap,idxn[j]+1,&col);
587:           col --;
588: #else
589:           col = aij->colmap[idxn[j]] - 1;
590: #endif
591:           if ((col < 0) || (aij->garray[col] != idxn[j])) *(v+i*n+j) = 0.0;
592:           else {
593:             MatGetValues(aij->B,1,&row,1,&col,v+i*n+j);
594:           }
595:         }
596:       }
597:     } else {
598:       SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Only local values currently supported");
599:     }
600:   }
601:   return(0);
602: }

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

608: PetscErrorCode MatAssemblyBegin_MPIAIJ(Mat mat,MatAssemblyType mode)
609: {
610:   Mat_MPIAIJ     *aij = (Mat_MPIAIJ*)mat->data;
612:   PetscInt       nstash,reallocs;
613:   InsertMode     addv;

616:   if (aij->donotstash || mat->nooffprocentries) {
617:     return(0);
618:   }

620:   /* make sure all processors are either in INSERTMODE or ADDMODE */
621:   MPI_Allreduce(&mat->insertmode,&addv,1,MPI_INT,MPI_BOR,((PetscObject)mat)->comm);
622:   if (addv == (ADD_VALUES|INSERT_VALUES)) SETERRQ(((PetscObject)mat)->comm,PETSC_ERR_ARG_WRONGSTATE,"Some processors inserted others added");
623:   mat->insertmode = addv; /* in case this processor had no cache */

625:   MatStashScatterBegin_Private(mat,&mat->stash,mat->rmap->range);
626:   MatStashGetInfo_Private(&mat->stash,&nstash,&reallocs);
627:   PetscInfo2(aij->A,"Stash has %D entries, uses %D mallocs.\n",nstash,reallocs);
628:   return(0);
629: }

633: PetscErrorCode MatAssemblyEnd_MPIAIJ(Mat mat,MatAssemblyType mode)
634: {
635:   Mat_MPIAIJ     *aij = (Mat_MPIAIJ*)mat->data;
636:   Mat_SeqAIJ     *a=(Mat_SeqAIJ *)aij->A->data;
638:   PetscMPIInt    n;
639:   PetscInt       i,j,rstart,ncols,flg;
640:   PetscInt       *row,*col;
641:   PetscBool      other_disassembled;
642:   PetscScalar    *val;
643:   InsertMode     addv = mat->insertmode;

645:   /* do not use 'b = (Mat_SeqAIJ *)aij->B->data' as B can be reset in disassembly */
647:   if (!aij->donotstash && !mat->nooffprocentries) {
648:     while (1) {
649:       MatStashScatterGetMesg_Private(&mat->stash,&n,&row,&col,&val,&flg);
650:       if (!flg) break;

652:       for (i=0; i<n;) {
653:         /* Now identify the consecutive vals belonging to the same row */
654:         for (j=i,rstart=row[j]; j<n; j++) { if (row[j] != rstart) break; }
655:         if (j < n) ncols = j-i;
656:         else       ncols = n-i;
657:         /* Now assemble all these values with a single function call */
658:         MatSetValues_MPIAIJ(mat,1,row+i,ncols,col+i,val+i,addv);
659:         i = j;
660:       }
661:     }
662:     MatStashScatterEnd_Private(&mat->stash);
663:   }
664:   MatAssemblyBegin(aij->A,mode);
665:   MatAssemblyEnd(aij->A,mode);

667:   /* determine if any processor has disassembled, if so we must 
668:      also disassemble ourselfs, in order that we may reassemble. */
669:   /*
670:      if nonzero structure of submatrix B cannot change then we know that
671:      no processor disassembled thus we can skip this stuff
672:   */
673:   if (!((Mat_SeqAIJ*)aij->B->data)->nonew)  {
674:     MPI_Allreduce(&mat->was_assembled,&other_disassembled,1,MPI_INT,MPI_PROD,((PetscObject)mat)->comm);
675:     if (mat->was_assembled && !other_disassembled) {
676:       MatDisAssemble_MPIAIJ(mat);
677:     }
678:   }
679:   if (!mat->was_assembled && mode == MAT_FINAL_ASSEMBLY) {
680:     MatSetUpMultiply_MPIAIJ(mat);
681:   }
682:   MatSetOption(aij->B,MAT_USE_INODES,PETSC_FALSE);
683:   MatSetOption(aij->B,MAT_CHECK_COMPRESSED_ROW,PETSC_FALSE);
684:   MatAssemblyBegin(aij->B,mode);
685:   MatAssemblyEnd(aij->B,mode);

687:   PetscFree2(aij->rowvalues,aij->rowindices);
688:   aij->rowvalues = 0;

690:   /* used by MatAXPY() */
691:   a->xtoy = 0; ((Mat_SeqAIJ *)aij->B->data)->xtoy = 0;  /* b->xtoy = 0 */
692:   a->XtoY = 0; ((Mat_SeqAIJ *)aij->B->data)->XtoY = 0;  /* b->XtoY = 0 */

694:   VecDestroy(&aij->diag);
695:   if (a->inode.size) mat->ops->multdiagonalblock = MatMultDiagonalBlock_MPIAIJ;
696:   return(0);
697: }

701: PetscErrorCode MatZeroEntries_MPIAIJ(Mat A)
702: {
703:   Mat_MPIAIJ     *l = (Mat_MPIAIJ*)A->data;

707:   MatZeroEntries(l->A);
708:   MatZeroEntries(l->B);
709:   return(0);
710: }

714: PetscErrorCode MatZeroRows_MPIAIJ(Mat A,PetscInt N,const PetscInt rows[],PetscScalar diag,Vec x,Vec b)
715: {
716:   Mat_MPIAIJ        *l = (Mat_MPIAIJ*)A->data;
717:   PetscErrorCode    ierr;
718:   PetscMPIInt       size = l->size,imdex,n,rank = l->rank,tag = ((PetscObject)A)->tag,lastidx = -1;
719:   PetscInt          i,*owners = A->rmap->range;
720:   PetscInt          *nprocs,j,idx,nsends,row;
721:   PetscInt          nmax,*svalues,*starts,*owner,nrecvs;
722:   PetscInt          *rvalues,count,base,slen,*source;
723:   PetscInt          *lens,*lrows,*values,rstart=A->rmap->rstart;
724:   MPI_Comm          comm = ((PetscObject)A)->comm;
725:   MPI_Request       *send_waits,*recv_waits;
726:   MPI_Status        recv_status,*send_status;
727:   const PetscScalar *xx;
728:   PetscScalar       *bb;
729: #if defined(PETSC_DEBUG)
730:   PetscBool      found = PETSC_FALSE;
731: #endif

734:   /*  first count number of contributors to each processor */
735:   PetscMalloc(2*size*sizeof(PetscInt),&nprocs);
736:   PetscMemzero(nprocs,2*size*sizeof(PetscInt));
737:   PetscMalloc((N+1)*sizeof(PetscInt),&owner); /* see note*/
738:   j = 0;
739:   for (i=0; i<N; i++) {
740:     if (lastidx > (idx = rows[i])) j = 0;
741:     lastidx = idx;
742:     for (; j<size; j++) {
743:       if (idx >= owners[j] && idx < owners[j+1]) {
744:         nprocs[2*j]++;
745:         nprocs[2*j+1] = 1;
746:         owner[i] = j;
747: #if defined(PETSC_DEBUG)
748:         found = PETSC_TRUE;
749: #endif
750:         break;
751:       }
752:     }
753: #if defined(PETSC_DEBUG)
754:     if (!found) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Index out of range");
755:     found = PETSC_FALSE;
756: #endif
757:   }
758:   nsends = 0;  for (i=0; i<size; i++) { nsends += nprocs[2*i+1];}

760:   if (A->nooffproczerorows) {
761:     if (nsends > 1) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"You called MatSetOption(,MAT_NO_OFF_PROC_ZERO_ROWS,PETSC_TRUE) but set an off process zero row");
762:     nrecvs = nsends;
763:     nmax   = N;
764:   } else {
765:     /* inform other processors of number of messages and max length*/
766:     PetscMaxSum(comm,nprocs,&nmax,&nrecvs);
767:   }

769:   /* post receives:   */
770:   PetscMalloc((nrecvs+1)*(nmax+1)*sizeof(PetscInt),&rvalues);
771:   PetscMalloc((nrecvs+1)*sizeof(MPI_Request),&recv_waits);
772:   for (i=0; i<nrecvs; i++) {
773:     MPI_Irecv(rvalues+nmax*i,nmax,MPIU_INT,MPI_ANY_SOURCE,tag,comm,recv_waits+i);
774:   }

776:   /* do sends:
777:       1) starts[i] gives the starting index in svalues for stuff going to 
778:          the ith processor
779:   */
780:   PetscMalloc((N+1)*sizeof(PetscInt),&svalues);
781:   PetscMalloc((nsends+1)*sizeof(MPI_Request),&send_waits);
782:   PetscMalloc((size+1)*sizeof(PetscInt),&starts);
783:   starts[0] = 0;
784:   for (i=1; i<size; i++) { starts[i] = starts[i-1] + nprocs[2*i-2];}
785:   for (i=0; i<N; i++) {
786:     svalues[starts[owner[i]]++] = rows[i];
787:   }

789:   starts[0] = 0;
790:   for (i=1; i<size+1; i++) { starts[i] = starts[i-1] + nprocs[2*i-2];}
791:   count = 0;
792:   for (i=0; i<size; i++) {
793:     if (nprocs[2*i+1]) {
794:       MPI_Isend(svalues+starts[i],nprocs[2*i],MPIU_INT,i,tag,comm,send_waits+count++);
795:     }
796:   }
797:   PetscFree(starts);

799:   base = owners[rank];

801:   /*  wait on receives */
802:   PetscMalloc2(nrecvs,PetscInt,&lens,nrecvs,PetscInt,&source);
803:   count  = nrecvs; slen = 0;
804:   while (count) {
805:     MPI_Waitany(nrecvs,recv_waits,&imdex,&recv_status);
806:     /* unpack receives into our local space */
807:     MPI_Get_count(&recv_status,MPIU_INT,&n);
808:     source[imdex]  = recv_status.MPI_SOURCE;
809:     lens[imdex]    = n;
810:     slen          += n;
811:     count--;
812:   }
813:   PetscFree(recv_waits);
814: 
815:   /* move the data into the send scatter */
816:   PetscMalloc((slen+1)*sizeof(PetscInt),&lrows);
817:   count = 0;
818:   for (i=0; i<nrecvs; i++) {
819:     values = rvalues + i*nmax;
820:     for (j=0; j<lens[i]; j++) {
821:       lrows[count++] = values[j] - base;
822:     }
823:   }
824:   PetscFree(rvalues);
825:   PetscFree2(lens,source);
826:   PetscFree(owner);
827:   PetscFree(nprocs);
828: 
829:   /* fix right hand side if needed */
830:   if (x && b) {
831:     VecGetArrayRead(x,&xx);
832:     VecGetArray(b,&bb);
833:     for (i=0; i<slen; i++) {
834:       bb[lrows[i]] = diag*xx[lrows[i]];
835:     }
836:     VecRestoreArrayRead(x,&xx);
837:     VecRestoreArray(b,&bb);
838:   }
839:   /*
840:         Zero the required rows. If the "diagonal block" of the matrix
841:      is square and the user wishes to set the diagonal we use separate
842:      code so that MatSetValues() is not called for each diagonal allocating
843:      new memory, thus calling lots of mallocs and slowing things down.

845:   */
846:   /* must zero l->B before l->A because the (diag) case below may put values into l->B*/
847:   MatZeroRows(l->B,slen,lrows,0.0,0,0);
848:   if ((diag != 0.0) && (l->A->rmap->N == l->A->cmap->N)) {
849:     MatZeroRows(l->A,slen,lrows,diag,0,0);
850:   } else if (diag != 0.0) {
851:     MatZeroRows(l->A,slen,lrows,0.0,0,0);
852:     if (((Mat_SeqAIJ*)l->A->data)->nonew) {
853:       SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"MatZeroRows() on rectangular matrices cannot be used with the Mat options\n\
854: MAT_NEW_NONZERO_LOCATIONS,MAT_NEW_NONZERO_LOCATION_ERR,MAT_NEW_NONZERO_ALLOCATION_ERR");
855:     }
856:     for (i = 0; i < slen; i++) {
857:       row  = lrows[i] + rstart;
858:       MatSetValues(A,1,&row,1,&row,&diag,INSERT_VALUES);
859:     }
860:     MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);
861:     MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);
862:   } else {
863:     MatZeroRows(l->A,slen,lrows,0.0,0,0);
864:   }
865:   PetscFree(lrows);

867:   /* wait on sends */
868:   if (nsends) {
869:     PetscMalloc(nsends*sizeof(MPI_Status),&send_status);
870:     MPI_Waitall(nsends,send_waits,send_status);
871:     PetscFree(send_status);
872:   }
873:   PetscFree(send_waits);
874:   PetscFree(svalues);
875:   return(0);
876: }

880: PetscErrorCode MatZeroRowsColumns_MPIAIJ(Mat A,PetscInt N,const PetscInt rows[],PetscScalar diag,Vec x,Vec b)
881: {
882:   Mat_MPIAIJ        *l = (Mat_MPIAIJ*)A->data;
883:   PetscErrorCode    ierr;
884:   PetscMPIInt       size = l->size,imdex,n,rank = l->rank,tag = ((PetscObject)A)->tag,lastidx = -1;
885:   PetscInt          i,*owners = A->rmap->range;
886:   PetscInt          *nprocs,j,idx,nsends;
887:   PetscInt          nmax,*svalues,*starts,*owner,nrecvs;
888:   PetscInt          *rvalues,count,base,slen,*source;
889:   PetscInt          *lens,*lrows,*values,m;
890:   MPI_Comm          comm = ((PetscObject)A)->comm;
891:   MPI_Request       *send_waits,*recv_waits;
892:   MPI_Status        recv_status,*send_status;
893:   const PetscScalar *xx;
894:   PetscScalar       *bb,*mask;
895:   Vec               xmask,lmask;
896:   Mat_SeqAIJ        *aij = (Mat_SeqAIJ*)l->B->data;
897:   const PetscInt    *aj, *ii,*ridx;
898:   PetscScalar       *aa;
899: #if defined(PETSC_DEBUG)
900:   PetscBool         found = PETSC_FALSE;
901: #endif

904:   /*  first count number of contributors to each processor */
905:   PetscMalloc(2*size*sizeof(PetscInt),&nprocs);
906:   PetscMemzero(nprocs,2*size*sizeof(PetscInt));
907:   PetscMalloc((N+1)*sizeof(PetscInt),&owner); /* see note*/
908:   j = 0;
909:   for (i=0; i<N; i++) {
910:     if (lastidx > (idx = rows[i])) j = 0;
911:     lastidx = idx;
912:     for (; j<size; j++) {
913:       if (idx >= owners[j] && idx < owners[j+1]) {
914:         nprocs[2*j]++;
915:         nprocs[2*j+1] = 1;
916:         owner[i] = j;
917: #if defined(PETSC_DEBUG)
918:         found = PETSC_TRUE;
919: #endif
920:         break;
921:       }
922:     }
923: #if defined(PETSC_DEBUG)
924:     if (!found) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Index out of range");
925:     found = PETSC_FALSE;
926: #endif
927:   }
928:   nsends = 0;  for (i=0; i<size; i++) { nsends += nprocs[2*i+1];}

930:   /* inform other processors of number of messages and max length*/
931:   PetscMaxSum(comm,nprocs,&nmax,&nrecvs);

933:   /* post receives:   */
934:   PetscMalloc((nrecvs+1)*(nmax+1)*sizeof(PetscInt),&rvalues);
935:   PetscMalloc((nrecvs+1)*sizeof(MPI_Request),&recv_waits);
936:   for (i=0; i<nrecvs; i++) {
937:     MPI_Irecv(rvalues+nmax*i,nmax,MPIU_INT,MPI_ANY_SOURCE,tag,comm,recv_waits+i);
938:   }

940:   /* do sends:
941:       1) starts[i] gives the starting index in svalues for stuff going to 
942:          the ith processor
943:   */
944:   PetscMalloc((N+1)*sizeof(PetscInt),&svalues);
945:   PetscMalloc((nsends+1)*sizeof(MPI_Request),&send_waits);
946:   PetscMalloc((size+1)*sizeof(PetscInt),&starts);
947:   starts[0] = 0;
948:   for (i=1; i<size; i++) { starts[i] = starts[i-1] + nprocs[2*i-2];}
949:   for (i=0; i<N; i++) {
950:     svalues[starts[owner[i]]++] = rows[i];
951:   }

953:   starts[0] = 0;
954:   for (i=1; i<size+1; i++) { starts[i] = starts[i-1] + nprocs[2*i-2];}
955:   count = 0;
956:   for (i=0; i<size; i++) {
957:     if (nprocs[2*i+1]) {
958:       MPI_Isend(svalues+starts[i],nprocs[2*i],MPIU_INT,i,tag,comm,send_waits+count++);
959:     }
960:   }
961:   PetscFree(starts);

963:   base = owners[rank];

965:   /*  wait on receives */
966:   PetscMalloc2(nrecvs,PetscInt,&lens,nrecvs,PetscInt,&source);
967:   count  = nrecvs; slen = 0;
968:   while (count) {
969:     MPI_Waitany(nrecvs,recv_waits,&imdex,&recv_status);
970:     /* unpack receives into our local space */
971:     MPI_Get_count(&recv_status,MPIU_INT,&n);
972:     source[imdex]  = recv_status.MPI_SOURCE;
973:     lens[imdex]    = n;
974:     slen          += n;
975:     count--;
976:   }
977:   PetscFree(recv_waits);
978: 
979:   /* move the data into the send scatter */
980:   PetscMalloc((slen+1)*sizeof(PetscInt),&lrows);
981:   count = 0;
982:   for (i=0; i<nrecvs; i++) {
983:     values = rvalues + i*nmax;
984:     for (j=0; j<lens[i]; j++) {
985:       lrows[count++] = values[j] - base;
986:     }
987:   }
988:   PetscFree(rvalues);
989:   PetscFree2(lens,source);
990:   PetscFree(owner);
991:   PetscFree(nprocs);
992:   /* lrows are the local rows to be zeroed, slen is the number of local rows */

994:   /* zero diagonal part of matrix */
995:   MatZeroRowsColumns(l->A,slen,lrows,diag,x,b);
996: 
997:   /* handle off diagonal part of matrix */
998:   MatGetVecs(A,&xmask,PETSC_NULL);
999:   VecDuplicate(l->lvec,&lmask);
1000:   VecGetArray(xmask,&bb);
1001:   for (i=0; i<slen; i++) {
1002:     bb[lrows[i]] = 1;
1003:   }
1004:   VecRestoreArray(xmask,&bb);
1005:   VecScatterBegin(l->Mvctx,xmask,lmask,ADD_VALUES,SCATTER_FORWARD);
1006:   VecScatterEnd(l->Mvctx,xmask,lmask,ADD_VALUES,SCATTER_FORWARD);
1007:   VecDestroy(&xmask);
1008:   if (x) {
1009:     VecScatterBegin(l->Mvctx,x,l->lvec,ADD_VALUES,SCATTER_FORWARD);
1010:     VecScatterEnd(l->Mvctx,x,l->lvec,ADD_VALUES,SCATTER_FORWARD);
1011:     VecGetArrayRead(l->lvec,&xx);
1012:     VecGetArray(b,&bb);
1013:   }
1014:   VecGetArray(lmask,&mask);

1016:   /* remove zeroed rows of off diagonal matrix */
1017:   ii = aij->i;
1018:   for (i=0; i<slen; i++) {
1019:     PetscMemzero(aij->a + ii[lrows[i]],(ii[lrows[i]+1] - ii[lrows[i]])*sizeof(PetscScalar));
1020:   }

1022:   /* loop over all elements of off process part of matrix zeroing removed columns*/
1023:   if (aij->compressedrow.use){
1024:     m    = aij->compressedrow.nrows;
1025:     ii   = aij->compressedrow.i;
1026:     ridx = aij->compressedrow.rindex;
1027:     for (i=0; i<m; i++){
1028:       n   = ii[i+1] - ii[i];
1029:       aj  = aij->j + ii[i];
1030:       aa  = aij->a + ii[i];

1032:       for (j=0; j<n; j++) {
1033:         if (PetscAbsScalar(mask[*aj])) {
1034:           if (b) bb[*ridx] -= *aa*xx[*aj];
1035:           *aa        = 0.0;
1036:         }
1037:         aa++;
1038:         aj++;
1039:       }
1040:       ridx++;
1041:     }
1042:   } else { /* do not use compressed row format */
1043:     m = l->B->rmap->n;
1044:     for (i=0; i<m; i++) {
1045:       n   = ii[i+1] - ii[i];
1046:       aj  = aij->j + ii[i];
1047:       aa  = aij->a + ii[i];
1048:       for (j=0; j<n; j++) {
1049:         if (PetscAbsScalar(mask[*aj])) {
1050:           if (b) bb[i] -= *aa*xx[*aj];
1051:           *aa    = 0.0;
1052:         }
1053:         aa++;
1054:         aj++;
1055:       }
1056:     }
1057:   }
1058:   if (x) {
1059:     VecRestoreArray(b,&bb);
1060:     VecRestoreArrayRead(l->lvec,&xx);
1061:   }
1062:   VecRestoreArray(lmask,&mask);
1063:   VecDestroy(&lmask);
1064:   PetscFree(lrows);

1066:   /* wait on sends */
1067:   if (nsends) {
1068:     PetscMalloc(nsends*sizeof(MPI_Status),&send_status);
1069:     MPI_Waitall(nsends,send_waits,send_status);
1070:     PetscFree(send_status);
1071:   }
1072:   PetscFree(send_waits);
1073:   PetscFree(svalues);

1075:   return(0);
1076: }

1080: PetscErrorCode MatMult_MPIAIJ(Mat A,Vec xx,Vec yy)
1081: {
1082:   Mat_MPIAIJ     *a = (Mat_MPIAIJ*)A->data;
1084:   PetscInt       nt;

1087:   VecGetLocalSize(xx,&nt);
1088:   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);
1089:   VecScatterBegin(a->Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);
1090:   (*a->A->ops->mult)(a->A,xx,yy);
1091:   VecScatterEnd(a->Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);
1092:   (*a->B->ops->multadd)(a->B,a->lvec,yy,yy);
1093:   return(0);
1094: }

1098: PetscErrorCode MatMultDiagonalBlock_MPIAIJ(Mat A,Vec bb,Vec xx)
1099: {
1100:   Mat_MPIAIJ     *a = (Mat_MPIAIJ*)A->data;

1104:   MatMultDiagonalBlock(a->A,bb,xx);
1105:   return(0);
1106: }

1110: PetscErrorCode MatMultAdd_MPIAIJ(Mat A,Vec xx,Vec yy,Vec zz)
1111: {
1112:   Mat_MPIAIJ     *a = (Mat_MPIAIJ*)A->data;

1116:   VecScatterBegin(a->Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);
1117:   (*a->A->ops->multadd)(a->A,xx,yy,zz);
1118:   VecScatterEnd(a->Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);
1119:   (*a->B->ops->multadd)(a->B,a->lvec,zz,zz);
1120:   return(0);
1121: }

1125: PetscErrorCode MatMultTranspose_MPIAIJ(Mat A,Vec xx,Vec yy)
1126: {
1127:   Mat_MPIAIJ     *a = (Mat_MPIAIJ*)A->data;
1129:   PetscBool      merged;

1132:   VecScatterGetMerged(a->Mvctx,&merged);
1133:   /* do nondiagonal part */
1134:   (*a->B->ops->multtranspose)(a->B,xx,a->lvec);
1135:   if (!merged) {
1136:     /* send it on its way */
1137:     VecScatterBegin(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);
1138:     /* do local part */
1139:     (*a->A->ops->multtranspose)(a->A,xx,yy);
1140:     /* receive remote parts: note this assumes the values are not actually */
1141:     /* added in yy until the next line, */
1142:     VecScatterEnd(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);
1143:   } else {
1144:     /* do local part */
1145:     (*a->A->ops->multtranspose)(a->A,xx,yy);
1146:     /* send it on its way */
1147:     VecScatterBegin(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);
1148:     /* values actually were received in the Begin() but we need to call this nop */
1149:     VecScatterEnd(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);
1150:   }
1151:   return(0);
1152: }

1154: EXTERN_C_BEGIN
1157: PetscErrorCode  MatIsTranspose_MPIAIJ(Mat Amat,Mat Bmat,PetscReal tol,PetscBool  *f)
1158: {
1159:   MPI_Comm       comm;
1160:   Mat_MPIAIJ     *Aij = (Mat_MPIAIJ *) Amat->data, *Bij;
1161:   Mat            Adia = Aij->A, Bdia, Aoff,Boff,*Aoffs,*Boffs;
1162:   IS             Me,Notme;
1164:   PetscInt       M,N,first,last,*notme,i;
1165:   PetscMPIInt    size;


1169:   /* Easy test: symmetric diagonal block */
1170:   Bij = (Mat_MPIAIJ *) Bmat->data; Bdia = Bij->A;
1171:   MatIsTranspose(Adia,Bdia,tol,f);
1172:   if (!*f) return(0);
1173:   PetscObjectGetComm((PetscObject)Amat,&comm);
1174:   MPI_Comm_size(comm,&size);
1175:   if (size == 1) return(0);

1177:   /* Hard test: off-diagonal block. This takes a MatGetSubMatrix. */
1178:   MatGetSize(Amat,&M,&N);
1179:   MatGetOwnershipRange(Amat,&first,&last);
1180:   PetscMalloc((N-last+first)*sizeof(PetscInt),&notme);
1181:   for (i=0; i<first; i++) notme[i] = i;
1182:   for (i=last; i<M; i++) notme[i-last+first] = i;
1183:   ISCreateGeneral(MPI_COMM_SELF,N-last+first,notme,PETSC_COPY_VALUES,&Notme);
1184:   ISCreateStride(MPI_COMM_SELF,last-first,first,1,&Me);
1185:   MatGetSubMatrices(Amat,1,&Me,&Notme,MAT_INITIAL_MATRIX,&Aoffs);
1186:   Aoff = Aoffs[0];
1187:   MatGetSubMatrices(Bmat,1,&Notme,&Me,MAT_INITIAL_MATRIX,&Boffs);
1188:   Boff = Boffs[0];
1189:   MatIsTranspose(Aoff,Boff,tol,f);
1190:   MatDestroyMatrices(1,&Aoffs);
1191:   MatDestroyMatrices(1,&Boffs);
1192:   ISDestroy(&Me);
1193:   ISDestroy(&Notme);
1194:   PetscFree(notme);
1195:    return(0);
1196: }
1197: EXTERN_C_END

1201: PetscErrorCode MatMultTransposeAdd_MPIAIJ(Mat A,Vec xx,Vec yy,Vec zz)
1202: {
1203:   Mat_MPIAIJ     *a = (Mat_MPIAIJ*)A->data;

1207:   /* do nondiagonal part */
1208:   (*a->B->ops->multtranspose)(a->B,xx,a->lvec);
1209:   /* send it on its way */
1210:   VecScatterBegin(a->Mvctx,a->lvec,zz,ADD_VALUES,SCATTER_REVERSE);
1211:   /* do local part */
1212:   (*a->A->ops->multtransposeadd)(a->A,xx,yy,zz);
1213:   /* receive remote parts */
1214:   VecScatterEnd(a->Mvctx,a->lvec,zz,ADD_VALUES,SCATTER_REVERSE);
1215:   return(0);
1216: }

1218: /*
1219:   This only works correctly for square matrices where the subblock A->A is the 
1220:    diagonal block
1221: */
1224: PetscErrorCode MatGetDiagonal_MPIAIJ(Mat A,Vec v)
1225: {
1227:   Mat_MPIAIJ     *a = (Mat_MPIAIJ*)A->data;

1230:   if (A->rmap->N != A->cmap->N) SETERRQ(((PetscObject)A)->comm,PETSC_ERR_SUP,"Supports only square matrix where A->A is diag block");
1231:   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");
1232:   MatGetDiagonal(a->A,v);
1233:   return(0);
1234: }

1238: PetscErrorCode MatScale_MPIAIJ(Mat A,PetscScalar aa)
1239: {
1240:   Mat_MPIAIJ     *a = (Mat_MPIAIJ*)A->data;

1244:   MatScale(a->A,aa);
1245:   MatScale(a->B,aa);
1246:   return(0);
1247: }

1251: PetscErrorCode MatDestroy_MPIAIJ(Mat mat)
1252: {
1253:   Mat_MPIAIJ     *aij = (Mat_MPIAIJ*)mat->data;

1257: #if defined(PETSC_USE_LOG)
1258:   PetscLogObjectState((PetscObject)mat,"Rows=%D, Cols=%D",mat->rmap->N,mat->cmap->N);
1259: #endif
1260:   MatStashDestroy_Private(&mat->stash);
1261:   VecDestroy(&aij->diag);
1262:   MatDestroy(&aij->A);
1263:   MatDestroy(&aij->B);
1264: #if defined (PETSC_USE_CTABLE)
1265:   PetscTableDestroy(&aij->colmap);
1266: #else
1267:   PetscFree(aij->colmap);
1268: #endif
1269:   PetscFree(aij->garray);
1270:   VecDestroy(&aij->lvec);
1271:   VecScatterDestroy(&aij->Mvctx);
1272:   PetscFree2(aij->rowvalues,aij->rowindices);
1273:   PetscFree(aij->ld);
1274:   PetscFree(mat->data);

1276:   PetscObjectChangeTypeName((PetscObject)mat,0);
1277:   PetscObjectComposeFunction((PetscObject)mat,"MatStoreValues_C","",PETSC_NULL);
1278:   PetscObjectComposeFunction((PetscObject)mat,"MatRetrieveValues_C","",PETSC_NULL);
1279:   PetscObjectComposeFunction((PetscObject)mat,"MatGetDiagonalBlock_C","",PETSC_NULL);
1280:   PetscObjectComposeFunction((PetscObject)mat,"MatIsTranspose_C","",PETSC_NULL);
1281:   PetscObjectComposeFunction((PetscObject)mat,"MatMPIAIJSetPreallocation_C","",PETSC_NULL);
1282:   PetscObjectComposeFunction((PetscObject)mat,"MatMPIAIJSetPreallocationCSR_C","",PETSC_NULL);
1283:   PetscObjectComposeFunction((PetscObject)mat,"MatDiagonalScaleLocal_C","",PETSC_NULL);
1284:   PetscObjectComposeFunction((PetscObject)mat,"MatConvert_mpiaij_mpisbaij_C","",PETSC_NULL);
1285:   return(0);
1286: }

1290: PetscErrorCode MatView_MPIAIJ_Binary(Mat mat,PetscViewer viewer)
1291: {
1292:   Mat_MPIAIJ        *aij = (Mat_MPIAIJ*)mat->data;
1293:   Mat_SeqAIJ*       A = (Mat_SeqAIJ*)aij->A->data;
1294:   Mat_SeqAIJ*       B = (Mat_SeqAIJ*)aij->B->data;
1295:   PetscErrorCode    ierr;
1296:   PetscMPIInt       rank,size,tag = ((PetscObject)viewer)->tag;
1297:   int               fd;
1298:   PetscInt          nz,header[4],*row_lengths,*range=0,rlen,i;
1299:   PetscInt          nzmax,*column_indices,j,k,col,*garray = aij->garray,cnt,cstart = mat->cmap->rstart,rnz;
1300:   PetscScalar       *column_values;
1301:   PetscInt          message_count,flowcontrolcount;

1304:   MPI_Comm_rank(((PetscObject)mat)->comm,&rank);
1305:   MPI_Comm_size(((PetscObject)mat)->comm,&size);
1306:   nz   = A->nz + B->nz;
1307:   if (!rank) {
1308:     header[0] = MAT_FILE_CLASSID;
1309:     header[1] = mat->rmap->N;
1310:     header[2] = mat->cmap->N;
1311:     MPI_Reduce(&nz,&header[3],1,MPIU_INT,MPI_SUM,0,((PetscObject)mat)->comm);
1312:     PetscViewerBinaryGetDescriptor(viewer,&fd);
1313:     PetscBinaryWrite(fd,header,4,PETSC_INT,PETSC_TRUE);
1314:     /* get largest number of rows any processor has */
1315:     rlen = mat->rmap->n;
1316:     range = mat->rmap->range;
1317:     for (i=1; i<size; i++) {
1318:       rlen = PetscMax(rlen,range[i+1] - range[i]);
1319:     }
1320:   } else {
1321:     MPI_Reduce(&nz,0,1,MPIU_INT,MPI_SUM,0,((PetscObject)mat)->comm);
1322:     rlen = mat->rmap->n;
1323:   }

1325:   /* load up the local row counts */
1326:   PetscMalloc((rlen+1)*sizeof(PetscInt),&row_lengths);
1327:   for (i=0; i<mat->rmap->n; i++) {
1328:     row_lengths[i] = A->i[i+1] - A->i[i] + B->i[i+1] - B->i[i];
1329:   }

1331:   /* store the row lengths to the file */
1332:   PetscViewerFlowControlStart(viewer,&message_count,&flowcontrolcount);
1333:   if (!rank) {
1334:     PetscBinaryWrite(fd,row_lengths,mat->rmap->n,PETSC_INT,PETSC_TRUE);
1335:     for (i=1; i<size; i++) {
1336:       PetscViewerFlowControlStepMaster(viewer,i,message_count,flowcontrolcount);
1337:       rlen = range[i+1] - range[i];
1338:       MPIULong_Recv(row_lengths,rlen,MPIU_INT,i,tag,((PetscObject)mat)->comm);
1339:       PetscBinaryWrite(fd,row_lengths,rlen,PETSC_INT,PETSC_TRUE);
1340:     }
1341:     PetscViewerFlowControlEndMaster(viewer,message_count);
1342:   } else {
1343:     PetscViewerFlowControlStepWorker(viewer,rank,message_count);
1344:     MPIULong_Send(row_lengths,mat->rmap->n,MPIU_INT,0,tag,((PetscObject)mat)->comm);
1345:     PetscViewerFlowControlEndWorker(viewer,message_count);
1346:   }
1347:   PetscFree(row_lengths);

1349:   /* load up the local column indices */
1350:   nzmax = nz; /* )th processor needs space a largest processor needs */
1351:   MPI_Reduce(&nz,&nzmax,1,MPIU_INT,MPI_MAX,0,((PetscObject)mat)->comm);
1352:   PetscMalloc((nzmax+1)*sizeof(PetscInt),&column_indices);
1353:   cnt  = 0;
1354:   for (i=0; i<mat->rmap->n; i++) {
1355:     for (j=B->i[i]; j<B->i[i+1]; j++) {
1356:       if ( (col = garray[B->j[j]]) > cstart) break;
1357:       column_indices[cnt++] = col;
1358:     }
1359:     for (k=A->i[i]; k<A->i[i+1]; k++) {
1360:       column_indices[cnt++] = A->j[k] + cstart;
1361:     }
1362:     for (; j<B->i[i+1]; j++) {
1363:       column_indices[cnt++] = garray[B->j[j]];
1364:     }
1365:   }
1366:   if (cnt != A->nz + B->nz) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_LIB,"Internal PETSc error: cnt = %D nz = %D",cnt,A->nz+B->nz);

1368:   /* store the column indices to the file */
1369:    PetscViewerFlowControlStart(viewer,&message_count,&flowcontrolcount);
1370:   if (!rank) {
1371:     MPI_Status status;
1372:     PetscBinaryWrite(fd,column_indices,nz,PETSC_INT,PETSC_TRUE);
1373:     for (i=1; i<size; i++) {
1374:       PetscViewerFlowControlStepMaster(viewer,i,message_count,flowcontrolcount);
1375:       MPI_Recv(&rnz,1,MPIU_INT,i,tag,((PetscObject)mat)->comm,&status);
1376:       if (rnz > nzmax) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_LIB,"Internal PETSc error: nz = %D nzmax = %D",nz,nzmax);
1377:       MPIULong_Recv(column_indices,rnz,MPIU_INT,i,tag,((PetscObject)mat)->comm);
1378:       PetscBinaryWrite(fd,column_indices,rnz,PETSC_INT,PETSC_TRUE);
1379:     }
1380:      PetscViewerFlowControlEndMaster(viewer,message_count);
1381:   } else {
1382:     PetscViewerFlowControlStepWorker(viewer,rank,message_count);
1383:     MPI_Send(&nz,1,MPIU_INT,0,tag,((PetscObject)mat)->comm);
1384:     MPIULong_Send(column_indices,nz,MPIU_INT,0,tag,((PetscObject)mat)->comm);
1385:     PetscViewerFlowControlEndWorker(viewer,message_count);
1386:   }
1387:   PetscFree(column_indices);

1389:   /* load up the local column values */
1390:   PetscMalloc((nzmax+1)*sizeof(PetscScalar),&column_values);
1391:   cnt  = 0;
1392:   for (i=0; i<mat->rmap->n; i++) {
1393:     for (j=B->i[i]; j<B->i[i+1]; j++) {
1394:       if ( garray[B->j[j]] > cstart) break;
1395:       column_values[cnt++] = B->a[j];
1396:     }
1397:     for (k=A->i[i]; k<A->i[i+1]; k++) {
1398:       column_values[cnt++] = A->a[k];
1399:     }
1400:     for (; j<B->i[i+1]; j++) {
1401:       column_values[cnt++] = B->a[j];
1402:     }
1403:   }
1404:   if (cnt != A->nz + B->nz) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Internal PETSc error: cnt = %D nz = %D",cnt,A->nz+B->nz);

1406:   /* store the column values to the file */
1407:    PetscViewerFlowControlStart(viewer,&message_count,&flowcontrolcount);
1408:   if (!rank) {
1409:     MPI_Status status;
1410:     PetscBinaryWrite(fd,column_values,nz,PETSC_SCALAR,PETSC_TRUE);
1411:     for (i=1; i<size; i++) {
1412:        PetscViewerFlowControlStepMaster(viewer,i,message_count,flowcontrolcount);
1413:       MPI_Recv(&rnz,1,MPIU_INT,i,tag,((PetscObject)mat)->comm,&status);
1414:       if (rnz > nzmax) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_LIB,"Internal PETSc error: nz = %D nzmax = %D",nz,nzmax);
1415:       MPIULong_Recv(column_values,rnz,MPIU_SCALAR,i,tag,((PetscObject)mat)->comm);
1416:       PetscBinaryWrite(fd,column_values,rnz,PETSC_SCALAR,PETSC_TRUE);
1417:     }
1418:     PetscViewerFlowControlEndMaster(viewer,message_count);
1419:   } else {
1420:     PetscViewerFlowControlStepWorker(viewer,rank,message_count);
1421:     MPI_Send(&nz,1,MPIU_INT,0,tag,((PetscObject)mat)->comm);
1422:     MPIULong_Send(column_values,nz,MPIU_SCALAR,0,tag,((PetscObject)mat)->comm);
1423:     PetscViewerFlowControlEndWorker(viewer,message_count);
1424:   }
1425:   PetscFree(column_values);
1426:   return(0);
1427: }

1431: PetscErrorCode MatView_MPIAIJ_ASCIIorDraworSocket(Mat mat,PetscViewer viewer)
1432: {
1433:   Mat_MPIAIJ        *aij = (Mat_MPIAIJ*)mat->data;
1434:   PetscErrorCode    ierr;
1435:   PetscMPIInt       rank = aij->rank,size = aij->size;
1436:   PetscBool         isdraw,iascii,isbinary;
1437:   PetscViewer       sviewer;
1438:   PetscViewerFormat format;

1441:   PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERDRAW,&isdraw);
1442:   PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii);
1443:   PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERBINARY,&isbinary);
1444:   if (iascii) {
1445:     PetscViewerGetFormat(viewer,&format);
1446:     if (format == PETSC_VIEWER_ASCII_INFO_DETAIL) {
1447:       MatInfo    info;
1448:       PetscBool  inodes;

1450:       MPI_Comm_rank(((PetscObject)mat)->comm,&rank);
1451:       MatGetInfo(mat,MAT_LOCAL,&info);
1452:       MatInodeGetInodeSizes(aij->A,PETSC_NULL,(PetscInt **)&inodes,PETSC_NULL);
1453:       PetscViewerASCIISynchronizedAllow(viewer,PETSC_TRUE);
1454:       if (!inodes) {
1455:         PetscViewerASCIISynchronizedPrintf(viewer,"[%d] Local rows %D nz %D nz alloced %D mem %D, not using I-node routines\n",
1456:                                               rank,mat->rmap->n,(PetscInt)info.nz_used,(PetscInt)info.nz_allocated,(PetscInt)info.memory);
1457:       } else {
1458:         PetscViewerASCIISynchronizedPrintf(viewer,"[%d] Local rows %D nz %D nz alloced %D mem %D, using I-node routines\n",
1459:                     rank,mat->rmap->n,(PetscInt)info.nz_used,(PetscInt)info.nz_allocated,(PetscInt)info.memory);
1460:       }
1461:       MatGetInfo(aij->A,MAT_LOCAL,&info);
1462:       PetscViewerASCIISynchronizedPrintf(viewer,"[%d] on-diagonal part: nz %D \n",rank,(PetscInt)info.nz_used);
1463:       MatGetInfo(aij->B,MAT_LOCAL,&info);
1464:       PetscViewerASCIISynchronizedPrintf(viewer,"[%d] off-diagonal part: nz %D \n",rank,(PetscInt)info.nz_used);
1465:       PetscViewerFlush(viewer);
1466:       PetscViewerASCIISynchronizedAllow(viewer,PETSC_FALSE);
1467:       PetscViewerASCIIPrintf(viewer,"Information on VecScatter used in matrix-vector product: \n");
1468:       VecScatterView(aij->Mvctx,viewer);
1469:       return(0);
1470:     } else if (format == PETSC_VIEWER_ASCII_INFO) {
1471:       PetscInt   inodecount,inodelimit,*inodes;
1472:       MatInodeGetInodeSizes(aij->A,&inodecount,&inodes,&inodelimit);
1473:       if (inodes) {
1474:         PetscViewerASCIIPrintf(viewer,"using I-node (on process 0) routines: found %D nodes, limit used is %D\n",inodecount,inodelimit);
1475:       } else {
1476:         PetscViewerASCIIPrintf(viewer,"not using I-node (on process 0) routines\n");
1477:       }
1478:       return(0);
1479:     } else if (format == PETSC_VIEWER_ASCII_FACTOR_INFO) {
1480:       return(0);
1481:     }
1482:   } else if (isbinary) {
1483:     if (size == 1) {
1484:       PetscObjectSetName((PetscObject)aij->A,((PetscObject)mat)->name);
1485:       MatView(aij->A,viewer);
1486:     } else {
1487:       MatView_MPIAIJ_Binary(mat,viewer);
1488:     }
1489:     return(0);
1490:   } else if (isdraw) {
1491:     PetscDraw  draw;
1492:     PetscBool  isnull;
1493:     PetscViewerDrawGetDraw(viewer,0,&draw);
1494:     PetscDrawIsNull(draw,&isnull); if (isnull) return(0);
1495:   }

1497:   if (size == 1) {
1498:     PetscObjectSetName((PetscObject)aij->A,((PetscObject)mat)->name);
1499:     MatView(aij->A,viewer);
1500:   } else {
1501:     /* assemble the entire matrix onto first processor. */
1502:     Mat         A;
1503:     Mat_SeqAIJ  *Aloc;
1504:     PetscInt    M = mat->rmap->N,N = mat->cmap->N,m,*ai,*aj,row,*cols,i,*ct;
1505:     MatScalar   *a;

1507:     if (mat->rmap->N > 1024) {
1508:       PetscBool  flg = PETSC_FALSE;

1510:       PetscOptionsGetBool(((PetscObject) mat)->prefix, "-mat_ascii_output_large", &flg,PETSC_NULL);
1511:       if (!flg) {
1512:         SETERRQ(((PetscObject)mat)->comm,PETSC_ERR_ARG_OUTOFRANGE,"ASCII matrix output not allowed for matrices with more than 1024 rows, use binary format instead.\nYou can override this restriction using -mat_ascii_output_large.");
1513:       }
1514:     }

1516:     MatCreate(((PetscObject)mat)->comm,&A);
1517:     if (!rank) {
1518:       MatSetSizes(A,M,N,M,N);
1519:     } else {
1520:       MatSetSizes(A,0,0,M,N);
1521:     }
1522:     /* This is just a temporary matrix, so explicitly using MATMPIAIJ is probably best */
1523:     MatSetType(A,MATMPIAIJ);
1524:     MatMPIAIJSetPreallocation(A,0,PETSC_NULL,0,PETSC_NULL);
1525:     MatSetOption(A,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);
1526:     PetscLogObjectParent(mat,A);

1528:     /* copy over the A part */
1529:     Aloc = (Mat_SeqAIJ*)aij->A->data;
1530:     m = aij->A->rmap->n; ai = Aloc->i; aj = Aloc->j; a = Aloc->a;
1531:     row = mat->rmap->rstart;
1532:     for (i=0; i<ai[m]; i++) {aj[i] += mat->cmap->rstart ;}
1533:     for (i=0; i<m; i++) {
1534:       MatSetValues(A,1,&row,ai[i+1]-ai[i],aj,a,INSERT_VALUES);
1535:       row++; a += ai[i+1]-ai[i]; aj += ai[i+1]-ai[i];
1536:     }
1537:     aj = Aloc->j;
1538:     for (i=0; i<ai[m]; i++) {aj[i] -= mat->cmap->rstart;}

1540:     /* copy over the B part */
1541:     Aloc = (Mat_SeqAIJ*)aij->B->data;
1542:     m    = aij->B->rmap->n;  ai = Aloc->i; aj = Aloc->j; a = Aloc->a;
1543:     row  = mat->rmap->rstart;
1544:     PetscMalloc((ai[m]+1)*sizeof(PetscInt),&cols);
1545:     ct   = cols;
1546:     for (i=0; i<ai[m]; i++) {cols[i] = aij->garray[aj[i]];}
1547:     for (i=0; i<m; i++) {
1548:       MatSetValues(A,1,&row,ai[i+1]-ai[i],cols,a,INSERT_VALUES);
1549:       row++; a += ai[i+1]-ai[i]; cols += ai[i+1]-ai[i];
1550:     }
1551:     PetscFree(ct);
1552:     MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);
1553:     MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);
1554:     /* 
1555:        Everyone has to call to draw the matrix since the graphics waits are
1556:        synchronized across all processors that share the PetscDraw object
1557:     */
1558:     PetscViewerGetSingleton(viewer,&sviewer);
1559:     if (!rank) {
1560:       PetscObjectSetName((PetscObject)((Mat_MPIAIJ*)(A->data))->A,((PetscObject)mat)->name);
1561:       /* Set the type name to MATMPIAIJ so that the correct type can be printed out by PetscObjectPrintClassNamePrefixType() in MatView_SeqAIJ_ASCII()*/
1562:       PetscStrcpy(((PetscObject)((Mat_MPIAIJ*)(A->data))->A)->type_name,MATMPIAIJ);
1563:       MatView(((Mat_MPIAIJ*)(A->data))->A,sviewer);
1564:     }
1565:     PetscViewerRestoreSingleton(viewer,&sviewer);
1566:     MatDestroy(&A);
1567:   }
1568:   return(0);
1569: }

1573: PetscErrorCode MatView_MPIAIJ(Mat mat,PetscViewer viewer)
1574: {
1576:   PetscBool      iascii,isdraw,issocket,isbinary;
1577: 
1579:   PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii);
1580:   PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERDRAW,&isdraw);
1581:   PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERBINARY,&isbinary);
1582:   PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERSOCKET,&issocket);
1583:   if (iascii || isdraw || isbinary || issocket) {
1584:     MatView_MPIAIJ_ASCIIorDraworSocket(mat,viewer);
1585:   } else {
1586:     SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"Viewer type %s not supported by MPIAIJ matrices",((PetscObject)viewer)->type_name);
1587:   }
1588:   return(0);
1589: }

1593: PetscErrorCode MatSOR_MPIAIJ(Mat matin,Vec bb,PetscReal omega,MatSORType flag,PetscReal fshift,PetscInt its,PetscInt lits,Vec xx)
1594: {
1595:   Mat_MPIAIJ     *mat = (Mat_MPIAIJ*)matin->data;
1597:   Vec            bb1 = 0;
1598:   PetscBool      hasop;

1601:   if (its > 1 || ~flag & SOR_ZERO_INITIAL_GUESS || flag & SOR_EISENSTAT) {
1602:     VecDuplicate(bb,&bb1);
1603:   }

1605:   if (flag == SOR_APPLY_UPPER) {
1606:     (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);
1607:     return(0);
1608:   }

1610:   if ((flag & SOR_LOCAL_SYMMETRIC_SWEEP) == SOR_LOCAL_SYMMETRIC_SWEEP){
1611:     if (flag & SOR_ZERO_INITIAL_GUESS) {
1612:       (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);
1613:       its--;
1614:     }
1615: 
1616:     while (its--) {
1617:       VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);
1618:       VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);

1620:       /* update rhs: bb1 = bb - B*x */
1621:       VecScale(mat->lvec,-1.0);
1622:       (*mat->B->ops->multadd)(mat->B,mat->lvec,bb,bb1);

1624:       /* local sweep */
1625:       (*mat->A->ops->sor)(mat->A,bb1,omega,SOR_SYMMETRIC_SWEEP,fshift,lits,1,xx);
1626:     }
1627:   } else if (flag & SOR_LOCAL_FORWARD_SWEEP){
1628:     if (flag & SOR_ZERO_INITIAL_GUESS) {
1629:       (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);
1630:       its--;
1631:     }
1632:     while (its--) {
1633:       VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);
1634:       VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);

1636:       /* update rhs: bb1 = bb - B*x */
1637:       VecScale(mat->lvec,-1.0);
1638:       (*mat->B->ops->multadd)(mat->B,mat->lvec,bb,bb1);

1640:       /* local sweep */
1641:       (*mat->A->ops->sor)(mat->A,bb1,omega,SOR_FORWARD_SWEEP,fshift,lits,1,xx);
1642:     }
1643:   } else if (flag & SOR_LOCAL_BACKWARD_SWEEP){
1644:     if (flag & SOR_ZERO_INITIAL_GUESS) {
1645:       (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);
1646:       its--;
1647:     }
1648:     while (its--) {
1649:       VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);
1650:       VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);

1652:       /* update rhs: bb1 = bb - B*x */
1653:       VecScale(mat->lvec,-1.0);
1654:       (*mat->B->ops->multadd)(mat->B,mat->lvec,bb,bb1);

1656:       /* local sweep */
1657:       (*mat->A->ops->sor)(mat->A,bb1,omega,SOR_BACKWARD_SWEEP,fshift,lits,1,xx);
1658:     }
1659:   }  else if (flag & SOR_EISENSTAT) {
1660:     Vec         xx1;

1662:     VecDuplicate(bb,&xx1);
1663:     (*mat->A->ops->sor)(mat->A,bb,omega,(MatSORType)(SOR_ZERO_INITIAL_GUESS | SOR_LOCAL_BACKWARD_SWEEP),fshift,lits,1,xx);

1665:     VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);
1666:     VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);
1667:     if (!mat->diag) {
1668:       MatGetVecs(matin,&mat->diag,PETSC_NULL);
1669:       MatGetDiagonal(matin,mat->diag);
1670:     }
1671:     MatHasOperation(matin,MATOP_MULT_DIAGONAL_BLOCK,&hasop);
1672:     if (hasop) {
1673:       MatMultDiagonalBlock(matin,xx,bb1);
1674:     } else {
1675:       VecPointwiseMult(bb1,mat->diag,xx);
1676:     }
1677:     VecAYPX(bb1,(omega-2.0)/omega,bb);

1679:     MatMultAdd(mat->B,mat->lvec,bb1,bb1);

1681:     /* local sweep */
1682:     (*mat->A->ops->sor)(mat->A,bb1,omega,(MatSORType)(SOR_ZERO_INITIAL_GUESS | SOR_LOCAL_FORWARD_SWEEP),fshift,lits,1,xx1);
1683:     VecAXPY(xx,1.0,xx1);
1684:     VecDestroy(&xx1);
1685:   } else SETERRQ(((PetscObject)matin)->comm,PETSC_ERR_SUP,"Parallel SOR not supported");

1687:   VecDestroy(&bb1);
1688:   return(0);
1689: }

1693: PetscErrorCode MatPermute_MPIAIJ(Mat A,IS rowp,IS colp,Mat *B)
1694: {
1695:   MPI_Comm       comm;
1696:   PetscInt       first,local_rowsize,local_colsize;
1697:   const PetscInt *rows;
1698:   IS             crowp,growp,irowp,lrowp,lcolp,icolp;

1702:   PetscObjectGetComm((PetscObject)A,&comm);
1703:   /* make a collective version of 'rowp', this is to be tolerant of users who pass serial index sets */
1704:   ISOnComm(rowp,comm,PETSC_USE_POINTER,&crowp);
1705:   /* collect the global row permutation and invert it */
1706:   ISAllGather(crowp,&growp);
1707:   ISSetPermutation(growp);
1708:   ISDestroy(&crowp);
1709:   ISInvertPermutation(growp,PETSC_DECIDE,&irowp);
1710:   ISDestroy(&growp);
1711:   /* get the local target indices */
1712:   MatGetOwnershipRange(A,&first,PETSC_NULL);
1713:   MatGetLocalSize(A,&local_rowsize,&local_colsize);
1714:   ISGetIndices(irowp,&rows);
1715:   ISCreateGeneral(PETSC_COMM_SELF,local_rowsize,rows+first,PETSC_COPY_VALUES,&lrowp);
1716:   ISRestoreIndices(irowp,&rows);
1717:   ISDestroy(&irowp);
1718:   /* the column permutation is so much easier;
1719:      make a local version of 'colp' and invert it */
1720:   ISOnComm(colp,PETSC_COMM_SELF,PETSC_USE_POINTER,&lcolp);
1721:   ISSetPermutation(lcolp);
1722:   ISInvertPermutation(lcolp,PETSC_DECIDE,&icolp);
1723:   ISDestroy(&lcolp);
1724:   /* now we just get the submatrix */
1725:   MatGetSubMatrix_MPIAIJ_Private(A,lrowp,icolp,local_colsize,MAT_INITIAL_MATRIX,B);
1726:   /* clean up */
1727:   ISDestroy(&lrowp);
1728:   ISDestroy(&icolp);
1729:   return(0);
1730: }

1734: PetscErrorCode MatGetInfo_MPIAIJ(Mat matin,MatInfoType flag,MatInfo *info)
1735: {
1736:   Mat_MPIAIJ     *mat = (Mat_MPIAIJ*)matin->data;
1737:   Mat            A = mat->A,B = mat->B;
1739:   PetscReal      isend[5],irecv[5];

1742:   info->block_size     = 1.0;
1743:   MatGetInfo(A,MAT_LOCAL,info);
1744:   isend[0] = info->nz_used; isend[1] = info->nz_allocated; isend[2] = info->nz_unneeded;
1745:   isend[3] = info->memory;  isend[4] = info->mallocs;
1746:   MatGetInfo(B,MAT_LOCAL,info);
1747:   isend[0] += info->nz_used; isend[1] += info->nz_allocated; isend[2] += info->nz_unneeded;
1748:   isend[3] += info->memory;  isend[4] += info->mallocs;
1749:   if (flag == MAT_LOCAL) {
1750:     info->nz_used      = isend[0];
1751:     info->nz_allocated = isend[1];
1752:     info->nz_unneeded  = isend[2];
1753:     info->memory       = isend[3];
1754:     info->mallocs      = isend[4];
1755:   } else if (flag == MAT_GLOBAL_MAX) {
1756:     MPI_Allreduce(isend,irecv,5,MPIU_REAL,MPIU_MAX,((PetscObject)matin)->comm);
1757:     info->nz_used      = irecv[0];
1758:     info->nz_allocated = irecv[1];
1759:     info->nz_unneeded  = irecv[2];
1760:     info->memory       = irecv[3];
1761:     info->mallocs      = irecv[4];
1762:   } else if (flag == MAT_GLOBAL_SUM) {
1763:     MPI_Allreduce(isend,irecv,5,MPIU_REAL,MPIU_SUM,((PetscObject)matin)->comm);
1764:     info->nz_used      = irecv[0];
1765:     info->nz_allocated = irecv[1];
1766:     info->nz_unneeded  = irecv[2];
1767:     info->memory       = irecv[3];
1768:     info->mallocs      = irecv[4];
1769:   }
1770:   info->fill_ratio_given  = 0; /* no parallel LU/ILU/Cholesky */
1771:   info->fill_ratio_needed = 0;
1772:   info->factor_mallocs    = 0;

1774:   return(0);
1775: }

1779: PetscErrorCode MatSetOption_MPIAIJ(Mat A,MatOption op,PetscBool  flg)
1780: {
1781:   Mat_MPIAIJ     *a = (Mat_MPIAIJ*)A->data;

1785:   switch (op) {
1786:   case MAT_NEW_NONZERO_LOCATIONS:
1787:   case MAT_NEW_NONZERO_ALLOCATION_ERR:
1788:   case MAT_UNUSED_NONZERO_LOCATION_ERR:
1789:   case MAT_KEEP_NONZERO_PATTERN:
1790:   case MAT_NEW_NONZERO_LOCATION_ERR:
1791:   case MAT_USE_INODES:
1792:   case MAT_IGNORE_ZERO_ENTRIES:
1793:     MatCheckPreallocated(A,1);
1794:     MatSetOption(a->A,op,flg);
1795:     MatSetOption(a->B,op,flg);
1796:     break;
1797:   case MAT_ROW_ORIENTED:
1798:     a->roworiented = flg;
1799:     MatSetOption(a->A,op,flg);
1800:     MatSetOption(a->B,op,flg);
1801:     break;
1802:   case MAT_NEW_DIAGONALS:
1803:     PetscInfo1(A,"Option %s ignored\n",MatOptions[op]);
1804:     break;
1805:   case MAT_IGNORE_OFF_PROC_ENTRIES:
1806:     a->donotstash = flg;
1807:     break;
1808:   case MAT_SPD:
1809:     A->spd_set                         = PETSC_TRUE;
1810:     A->spd                             = flg;
1811:     if (flg) {
1812:       A->symmetric                     = PETSC_TRUE;
1813:       A->structurally_symmetric        = PETSC_TRUE;
1814:       A->symmetric_set                 = PETSC_TRUE;
1815:       A->structurally_symmetric_set    = PETSC_TRUE;
1816:     }
1817:     break;
1818:   case MAT_SYMMETRIC:
1819:     MatSetOption(a->A,op,flg);
1820:     break;
1821:   case MAT_STRUCTURALLY_SYMMETRIC:
1822:     MatSetOption(a->A,op,flg);
1823:     break;
1824:   case MAT_HERMITIAN:
1825:     MatSetOption(a->A,op,flg);
1826:     break;
1827:   case MAT_SYMMETRY_ETERNAL:
1828:     MatSetOption(a->A,op,flg);
1829:     break;
1830:   default:
1831:     SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"unknown option %d",op);
1832:   }
1833:   return(0);
1834: }

1838: PetscErrorCode MatGetRow_MPIAIJ(Mat matin,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v)
1839: {
1840:   Mat_MPIAIJ     *mat = (Mat_MPIAIJ*)matin->data;
1841:   PetscScalar    *vworkA,*vworkB,**pvA,**pvB,*v_p;
1843:   PetscInt       i,*cworkA,*cworkB,**pcA,**pcB,cstart = matin->cmap->rstart;
1844:   PetscInt       nztot,nzA,nzB,lrow,rstart = matin->rmap->rstart,rend = matin->rmap->rend;
1845:   PetscInt       *cmap,*idx_p;

1848:   if (mat->getrowactive) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Already active");
1849:   mat->getrowactive = PETSC_TRUE;

1851:   if (!mat->rowvalues && (idx || v)) {
1852:     /*
1853:         allocate enough space to hold information from the longest row.
1854:     */
1855:     Mat_SeqAIJ *Aa = (Mat_SeqAIJ*)mat->A->data,*Ba = (Mat_SeqAIJ*)mat->B->data;
1856:     PetscInt   max = 1,tmp;
1857:     for (i=0; i<matin->rmap->n; i++) {
1858:       tmp = Aa->i[i+1] - Aa->i[i] + Ba->i[i+1] - Ba->i[i];
1859:       if (max < tmp) { max = tmp; }
1860:     }
1861:     PetscMalloc2(max,PetscScalar,&mat->rowvalues,max,PetscInt,&mat->rowindices);
1862:   }

1864:   if (row < rstart || row >= rend) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Only local rows");
1865:   lrow = row - rstart;

1867:   pvA = &vworkA; pcA = &cworkA; pvB = &vworkB; pcB = &cworkB;
1868:   if (!v)   {pvA = 0; pvB = 0;}
1869:   if (!idx) {pcA = 0; if (!v) pcB = 0;}
1870:   (*mat->A->ops->getrow)(mat->A,lrow,&nzA,pcA,pvA);
1871:   (*mat->B->ops->getrow)(mat->B,lrow,&nzB,pcB,pvB);
1872:   nztot = nzA + nzB;

1874:   cmap  = mat->garray;
1875:   if (v  || idx) {
1876:     if (nztot) {
1877:       /* Sort by increasing column numbers, assuming A and B already sorted */
1878:       PetscInt imark = -1;
1879:       if (v) {
1880:         *v = v_p = mat->rowvalues;
1881:         for (i=0; i<nzB; i++) {
1882:           if (cmap[cworkB[i]] < cstart)   v_p[i] = vworkB[i];
1883:           else break;
1884:         }
1885:         imark = i;
1886:         for (i=0; i<nzA; i++)     v_p[imark+i] = vworkA[i];
1887:         for (i=imark; i<nzB; i++) v_p[nzA+i]   = vworkB[i];
1888:       }
1889:       if (idx) {
1890:         *idx = idx_p = mat->rowindices;
1891:         if (imark > -1) {
1892:           for (i=0; i<imark; i++) {
1893:             idx_p[i] = cmap[cworkB[i]];
1894:           }
1895:         } else {
1896:           for (i=0; i<nzB; i++) {
1897:             if (cmap[cworkB[i]] < cstart)   idx_p[i] = cmap[cworkB[i]];
1898:             else break;
1899:           }
1900:           imark = i;
1901:         }
1902:         for (i=0; i<nzA; i++)     idx_p[imark+i] = cstart + cworkA[i];
1903:         for (i=imark; i<nzB; i++) idx_p[nzA+i]   = cmap[cworkB[i]];
1904:       }
1905:     } else {
1906:       if (idx) *idx = 0;
1907:       if (v)   *v   = 0;
1908:     }
1909:   }
1910:   *nz = nztot;
1911:   (*mat->A->ops->restorerow)(mat->A,lrow,&nzA,pcA,pvA);
1912:   (*mat->B->ops->restorerow)(mat->B,lrow,&nzB,pcB,pvB);
1913:   return(0);
1914: }

1918: PetscErrorCode MatRestoreRow_MPIAIJ(Mat mat,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v)
1919: {
1920:   Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;

1923:   if (!aij->getrowactive) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"MatGetRow() must be called first");
1924:   aij->getrowactive = PETSC_FALSE;
1925:   return(0);
1926: }

1930: PetscErrorCode MatNorm_MPIAIJ(Mat mat,NormType type,PetscReal *norm)
1931: {
1932:   Mat_MPIAIJ     *aij = (Mat_MPIAIJ*)mat->data;
1933:   Mat_SeqAIJ     *amat = (Mat_SeqAIJ*)aij->A->data,*bmat = (Mat_SeqAIJ*)aij->B->data;
1935:   PetscInt       i,j,cstart = mat->cmap->rstart;
1936:   PetscReal      sum = 0.0;
1937:   MatScalar      *v;

1940:   if (aij->size == 1) {
1941:      MatNorm(aij->A,type,norm);
1942:   } else {
1943:     if (type == NORM_FROBENIUS) {
1944:       v = amat->a;
1945:       for (i=0; i<amat->nz; i++) {
1946: #if defined(PETSC_USE_COMPLEX)
1947:         sum += PetscRealPart(PetscConj(*v)*(*v)); v++;
1948: #else
1949:         sum += (*v)*(*v); v++;
1950: #endif
1951:       }
1952:       v = bmat->a;
1953:       for (i=0; i<bmat->nz; i++) {
1954: #if defined(PETSC_USE_COMPLEX)
1955:         sum += PetscRealPart(PetscConj(*v)*(*v)); v++;
1956: #else
1957:         sum += (*v)*(*v); v++;
1958: #endif
1959:       }
1960:       MPI_Allreduce(&sum,norm,1,MPIU_REAL,MPIU_SUM,((PetscObject)mat)->comm);
1961:       *norm = PetscSqrtReal(*norm);
1962:     } else if (type == NORM_1) { /* max column norm */
1963:       PetscReal *tmp,*tmp2;
1964:       PetscInt  *jj,*garray = aij->garray;
1965:       PetscMalloc((mat->cmap->N+1)*sizeof(PetscReal),&tmp);
1966:       PetscMalloc((mat->cmap->N+1)*sizeof(PetscReal),&tmp2);
1967:       PetscMemzero(tmp,mat->cmap->N*sizeof(PetscReal));
1968:       *norm = 0.0;
1969:       v = amat->a; jj = amat->j;
1970:       for (j=0; j<amat->nz; j++) {
1971:         tmp[cstart + *jj++ ] += PetscAbsScalar(*v);  v++;
1972:       }
1973:       v = bmat->a; jj = bmat->j;
1974:       for (j=0; j<bmat->nz; j++) {
1975:         tmp[garray[*jj++]] += PetscAbsScalar(*v); v++;
1976:       }
1977:       MPI_Allreduce(tmp,tmp2,mat->cmap->N,MPIU_REAL,MPIU_SUM,((PetscObject)mat)->comm);
1978:       for (j=0; j<mat->cmap->N; j++) {
1979:         if (tmp2[j] > *norm) *norm = tmp2[j];
1980:       }
1981:       PetscFree(tmp);
1982:       PetscFree(tmp2);
1983:     } else if (type == NORM_INFINITY) { /* max row norm */
1984:       PetscReal ntemp = 0.0;
1985:       for (j=0; j<aij->A->rmap->n; j++) {
1986:         v = amat->a + amat->i[j];
1987:         sum = 0.0;
1988:         for (i=0; i<amat->i[j+1]-amat->i[j]; i++) {
1989:           sum += PetscAbsScalar(*v); v++;
1990:         }
1991:         v = bmat->a + bmat->i[j];
1992:         for (i=0; i<bmat->i[j+1]-bmat->i[j]; i++) {
1993:           sum += PetscAbsScalar(*v); v++;
1994:         }
1995:         if (sum > ntemp) ntemp = sum;
1996:       }
1997:       MPI_Allreduce(&ntemp,norm,1,MPIU_REAL,MPIU_MAX,((PetscObject)mat)->comm);
1998:     } else {
1999:       SETERRQ(((PetscObject)mat)->comm,PETSC_ERR_SUP,"No support for two norm");
2000:     }
2001:   }
2002:   return(0);
2003: }

2007: PetscErrorCode MatTranspose_MPIAIJ(Mat A,MatReuse reuse,Mat *matout)
2008: {
2009:   Mat_MPIAIJ     *a = (Mat_MPIAIJ*)A->data;
2010:   Mat_SeqAIJ     *Aloc=(Mat_SeqAIJ*)a->A->data,*Bloc=(Mat_SeqAIJ*)a->B->data;
2012:   PetscInt       M = A->rmap->N,N = A->cmap->N,ma,na,mb,*ai,*aj,*bi,*bj,row,*cols,*cols_tmp,i,*d_nnz;
2013:   PetscInt       cstart=A->cmap->rstart,ncol;
2014:   Mat            B;
2015:   MatScalar      *array;

2018:   if (reuse == MAT_REUSE_MATRIX && A == *matout && M != N) SETERRQ(((PetscObject)A)->comm,PETSC_ERR_ARG_SIZ,"Square matrix only for in-place");

2020:   ma = A->rmap->n; na = A->cmap->n; mb = a->B->rmap->n;
2021:   ai = Aloc->i; aj = Aloc->j;
2022:   bi = Bloc->i; bj = Bloc->j;
2023:   if (reuse == MAT_INITIAL_MATRIX || *matout == A) {
2024:     /* compute d_nnz for preallocation; o_nnz is approximated by d_nnz to avoid communication */
2025:     PetscMalloc((1+na)*sizeof(PetscInt),&d_nnz);
2026:     PetscMemzero(d_nnz,(1+na)*sizeof(PetscInt));
2027:     for (i=0; i<ai[ma]; i++){
2028:       d_nnz[aj[i]] ++;
2029:       aj[i] += cstart; /* global col index to be used by MatSetValues() */
2030:     }

2032:     MatCreate(((PetscObject)A)->comm,&B);
2033:     MatSetSizes(B,A->cmap->n,A->rmap->n,N,M);
2034:     MatSetBlockSizes(B,A->cmap->bs,A->rmap->bs);
2035:     MatSetType(B,((PetscObject)A)->type_name);
2036:     MatMPIAIJSetPreallocation(B,0,d_nnz,0,d_nnz);
2037:     MatSetOption(B,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);
2038:     PetscFree(d_nnz);
2039:   } else {
2040:     B = *matout;
2041:     MatSetOption(B,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_TRUE);
2042:     for (i=0; i<ai[ma]; i++){
2043:       aj[i] += cstart; /* global col index to be used by MatSetValues() */
2044:     }
2045:   }

2047:   /* copy over the A part */
2048:   array = Aloc->a;
2049:   row = A->rmap->rstart;
2050:   for (i=0; i<ma; i++) {
2051:     ncol = ai[i+1]-ai[i];
2052:     MatSetValues(B,ncol,aj,1,&row,array,INSERT_VALUES);
2053:     row++; array += ncol; aj += ncol;
2054:   }
2055:   aj = Aloc->j;
2056:   for (i=0; i<ai[ma]; i++) aj[i] -= cstart; /* resume local col index */

2058:   /* copy over the B part */
2059:   PetscMalloc(bi[mb]*sizeof(PetscInt),&cols);
2060:   PetscMemzero(cols,bi[mb]*sizeof(PetscInt));
2061:   array = Bloc->a;
2062:   row = A->rmap->rstart;
2063:   for (i=0; i<bi[mb]; i++) {cols[i] = a->garray[bj[i]];}
2064:   cols_tmp = cols;
2065:   for (i=0; i<mb; i++) {
2066:     ncol = bi[i+1]-bi[i];
2067:     MatSetValues(B,ncol,cols_tmp,1,&row,array,INSERT_VALUES);
2068:     row++; array += ncol; cols_tmp += ncol;
2069:   }
2070:   PetscFree(cols);
2071: 
2072:   MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);
2073:   MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);
2074:   if (reuse == MAT_INITIAL_MATRIX || *matout != A) {
2075:     *matout = B;
2076:   } else {
2077:     MatHeaderMerge(A,B);
2078:   }
2079:   return(0);
2080: }

2084: PetscErrorCode MatDiagonalScale_MPIAIJ(Mat mat,Vec ll,Vec rr)
2085: {
2086:   Mat_MPIAIJ     *aij = (Mat_MPIAIJ*)mat->data;
2087:   Mat            a = aij->A,b = aij->B;
2089:   PetscInt       s1,s2,s3;

2092:   MatGetLocalSize(mat,&s2,&s3);
2093:   if (rr) {
2094:     VecGetLocalSize(rr,&s1);
2095:     if (s1!=s3) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"right vector non-conforming local size");
2096:     /* Overlap communication with computation. */
2097:     VecScatterBegin(aij->Mvctx,rr,aij->lvec,INSERT_VALUES,SCATTER_FORWARD);
2098:   }
2099:   if (ll) {
2100:     VecGetLocalSize(ll,&s1);
2101:     if (s1!=s2) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"left vector non-conforming local size");
2102:     (*b->ops->diagonalscale)(b,ll,0);
2103:   }
2104:   /* scale  the diagonal block */
2105:   (*a->ops->diagonalscale)(a,ll,rr);

2107:   if (rr) {
2108:     /* Do a scatter end and then right scale the off-diagonal block */
2109:     VecScatterEnd(aij->Mvctx,rr,aij->lvec,INSERT_VALUES,SCATTER_FORWARD);
2110:     (*b->ops->diagonalscale)(b,0,aij->lvec);
2111:   }
2112: 
2113:   return(0);
2114: }

2118: PetscErrorCode MatSetUnfactored_MPIAIJ(Mat A)
2119: {
2120:   Mat_MPIAIJ     *a   = (Mat_MPIAIJ*)A->data;

2124:   MatSetUnfactored(a->A);
2125:   return(0);
2126: }

2130: PetscErrorCode MatEqual_MPIAIJ(Mat A,Mat B,PetscBool  *flag)
2131: {
2132:   Mat_MPIAIJ     *matB = (Mat_MPIAIJ*)B->data,*matA = (Mat_MPIAIJ*)A->data;
2133:   Mat            a,b,c,d;
2134:   PetscBool      flg;

2138:   a = matA->A; b = matA->B;
2139:   c = matB->A; d = matB->B;

2141:   MatEqual(a,c,&flg);
2142:   if (flg) {
2143:     MatEqual(b,d,&flg);
2144:   }
2145:   MPI_Allreduce(&flg,flag,1,MPI_INT,MPI_LAND,((PetscObject)A)->comm);
2146:   return(0);
2147: }

2151: PetscErrorCode MatCopy_MPIAIJ(Mat A,Mat B,MatStructure str)
2152: {
2154:   Mat_MPIAIJ     *a = (Mat_MPIAIJ *)A->data;
2155:   Mat_MPIAIJ     *b = (Mat_MPIAIJ *)B->data;

2158:   /* If the two matrices don't have the same copy implementation, they aren't compatible for fast copy. */
2159:   if ((str != SAME_NONZERO_PATTERN) || (A->ops->copy != B->ops->copy)) {
2160:     /* because of the column compression in the off-processor part of the matrix a->B,
2161:        the number of columns in a->B and b->B may be different, hence we cannot call
2162:        the MatCopy() directly on the two parts. If need be, we can provide a more 
2163:        efficient copy than the MatCopy_Basic() by first uncompressing the a->B matrices
2164:        then copying the submatrices */
2165:     MatCopy_Basic(A,B,str);
2166:   } else {
2167:     MatCopy(a->A,b->A,str);
2168:     MatCopy(a->B,b->B,str);
2169:   }
2170:   return(0);
2171: }

2175: PetscErrorCode MatSetUp_MPIAIJ(Mat A)
2176: {

2180:    MatMPIAIJSetPreallocation(A,PETSC_DEFAULT,0,PETSC_DEFAULT,0);
2181:   return(0);
2182: }

2186: /* This is the same as MatAXPYGetPreallocation_SeqAIJ, except that the local-to-global map is provided */
2187: static PetscErrorCode MatAXPYGetPreallocation_MPIAIJ(Mat Y,const PetscInt *yltog,Mat X,const PetscInt *xltog,PetscInt* nnz)
2188: {
2189:   PetscInt          i,m=Y->rmap->N;
2190:   Mat_SeqAIJ        *x = (Mat_SeqAIJ*)X->data;
2191:   Mat_SeqAIJ        *y = (Mat_SeqAIJ*)Y->data;
2192:   const PetscInt    *xi = x->i,*yi = y->i;

2195:   /* Set the number of nonzeros in the new matrix */
2196:   for(i=0; i<m; i++) {
2197:     PetscInt j,k,nzx = xi[i+1] - xi[i],nzy = yi[i+1] - yi[i];
2198:     const PetscInt *xj = x->j+xi[i],*yj = y->j+yi[i];
2199:     nnz[i] = 0;
2200:     for (j=0,k=0; j<nzx; j++) {                   /* Point in X */
2201:       for (; k<nzy && yltog[yj[k]]<xltog[xj[j]]; k++) nnz[i]++; /* Catch up to X */
2202:       if (k<nzy && yltog[yj[k]]==xltog[xj[j]]) k++;             /* Skip duplicate */
2203:       nnz[i]++;
2204:     }
2205:     for (; k<nzy; k++) nnz[i]++;
2206:   }
2207:   return(0);
2208: }

2212: PetscErrorCode MatAXPY_MPIAIJ(Mat Y,PetscScalar a,Mat X,MatStructure str)
2213: {
2215:   PetscInt       i;
2216:   Mat_MPIAIJ     *xx = (Mat_MPIAIJ *)X->data,*yy = (Mat_MPIAIJ *)Y->data;
2217:   PetscBLASInt   bnz,one=1;
2218:   Mat_SeqAIJ     *x,*y;

2221:   if (str == SAME_NONZERO_PATTERN) {
2222:     PetscScalar alpha = a;
2223:     x = (Mat_SeqAIJ *)xx->A->data;
2224:     y = (Mat_SeqAIJ *)yy->A->data;
2225:     bnz = PetscBLASIntCast(x->nz);
2226:     BLASaxpy_(&bnz,&alpha,x->a,&one,y->a,&one);
2227:     x = (Mat_SeqAIJ *)xx->B->data;
2228:     y = (Mat_SeqAIJ *)yy->B->data;
2229:     bnz = PetscBLASIntCast(x->nz);
2230:     BLASaxpy_(&bnz,&alpha,x->a,&one,y->a,&one);
2231:   } else if (str == SUBSET_NONZERO_PATTERN) {
2232:     MatAXPY_SeqAIJ(yy->A,a,xx->A,str);

2234:     x = (Mat_SeqAIJ *)xx->B->data;
2235:     y = (Mat_SeqAIJ *)yy->B->data;
2236:     if (y->xtoy && y->XtoY != xx->B) {
2237:       PetscFree(y->xtoy);
2238:       MatDestroy(&y->XtoY);
2239:     }
2240:     if (!y->xtoy) { /* get xtoy */
2241:       MatAXPYGetxtoy_Private(xx->B->rmap->n,x->i,x->j,xx->garray,y->i,y->j,yy->garray,&y->xtoy);
2242:       y->XtoY = xx->B;
2243:       PetscObjectReference((PetscObject)xx->B);
2244:     }
2245:     for (i=0; i<x->nz; i++) y->a[y->xtoy[i]] += a*(x->a[i]);
2246:   } else {
2247:     Mat B;
2248:     PetscInt *nnz_d,*nnz_o;
2249:     PetscMalloc(yy->A->rmap->N*sizeof(PetscInt),&nnz_d);
2250:     PetscMalloc(yy->B->rmap->N*sizeof(PetscInt),&nnz_o);
2251:     MatCreate(((PetscObject)Y)->comm,&B);
2252:     PetscObjectSetName((PetscObject)B,((PetscObject)Y)->name);
2253:     MatSetSizes(B,Y->rmap->n,Y->cmap->n,Y->rmap->N,Y->cmap->N);
2254:     MatSetBlockSizes(B,Y->rmap->bs,Y->cmap->bs);
2255:     MatSetType(B,MATMPIAIJ);
2256:     MatAXPYGetPreallocation_SeqAIJ(yy->A,xx->A,nnz_d);
2257:     MatAXPYGetPreallocation_MPIAIJ(yy->B,yy->garray,xx->B,xx->garray,nnz_o);
2258:     MatMPIAIJSetPreallocation(B,0,nnz_d,0,nnz_o);
2259:     MatAXPY_BasicWithPreallocation(B,Y,a,X,str);
2260:     MatHeaderReplace(Y,B);
2261:     PetscFree(nnz_d);
2262:     PetscFree(nnz_o);
2263:   }
2264:   return(0);
2265: }

2267: extern PetscErrorCode  MatConjugate_SeqAIJ(Mat);

2271: PetscErrorCode  MatConjugate_MPIAIJ(Mat mat)
2272: {
2273: #if defined(PETSC_USE_COMPLEX)
2275:   Mat_MPIAIJ     *aij = (Mat_MPIAIJ *)mat->data;

2278:   MatConjugate_SeqAIJ(aij->A);
2279:   MatConjugate_SeqAIJ(aij->B);
2280: #else
2282: #endif
2283:   return(0);
2284: }

2288: PetscErrorCode MatRealPart_MPIAIJ(Mat A)
2289: {
2290:   Mat_MPIAIJ   *a = (Mat_MPIAIJ*)A->data;

2294:   MatRealPart(a->A);
2295:   MatRealPart(a->B);
2296:   return(0);
2297: }

2301: PetscErrorCode MatImaginaryPart_MPIAIJ(Mat A)
2302: {
2303:   Mat_MPIAIJ   *a = (Mat_MPIAIJ*)A->data;

2307:   MatImaginaryPart(a->A);
2308:   MatImaginaryPart(a->B);
2309:   return(0);
2310: }

2312: #ifdef PETSC_HAVE_PBGL

2314: #include <boost/parallel/mpi/bsp_process_group.hpp>
2315: #include <boost/graph/distributed/ilu_default_graph.hpp>
2316: #include <boost/graph/distributed/ilu_0_block.hpp>
2317: #include <boost/graph/distributed/ilu_preconditioner.hpp>
2318: #include <boost/graph/distributed/petsc/interface.hpp>
2319: #include <boost/multi_array.hpp>
2320: #include <boost/parallel/distributed_property_map->hpp>

2324: /*
2325:   This uses the parallel ILU factorization of Peter Gottschling <pgottsch@osl.iu.edu>
2326: */
2327: PetscErrorCode MatILUFactorSymbolic_MPIAIJ(Mat fact,Mat A, IS isrow, IS iscol, const MatFactorInfo *info)
2328: {
2329:   namespace petsc = boost::distributed::petsc;
2330: 
2331:   namespace graph_dist = boost::graph::distributed;
2332:   using boost::graph::distributed::ilu_default::process_group_type;
2333:   using boost::graph::ilu_permuted;

2335:   PetscBool       row_identity, col_identity;
2336:   PetscContainer  c;
2337:   PetscInt        m, n, M, N;
2338:   PetscErrorCode  ierr;

2341:   if (info->levels != 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Only levels = 0 supported for parallel ilu");
2342:   ISIdentity(isrow, &row_identity);
2343:   ISIdentity(iscol, &col_identity);
2344:   if (!row_identity || !col_identity) {
2345:     SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Row and column permutations must be identity for parallel ILU");
2346:   }

2348:   process_group_type pg;
2349:   typedef graph_dist::ilu_default::ilu_level_graph_type  lgraph_type;
2350:   lgraph_type*   lgraph_p = new lgraph_type(petsc::num_global_vertices(A), pg, petsc::matrix_distribution(A, pg));
2351:   lgraph_type&   level_graph = *lgraph_p;
2352:   graph_dist::ilu_default::graph_type&            graph(level_graph.graph);

2354:   petsc::read_matrix(A, graph, get(boost::edge_weight, graph));
2355:   ilu_permuted(level_graph);

2357:   /* put together the new matrix */
2358:   MatCreate(((PetscObject)A)->comm, fact);
2359:   MatGetLocalSize(A, &m, &n);
2360:   MatGetSize(A, &M, &N);
2361:   MatSetSizes(fact, m, n, M, N);
2362:   MatSetBlockSizes(fact,A->rmap->bs,A->cmap->bs);
2363:   MatSetType(fact, ((PetscObject)A)->type_name);
2364:   MatAssemblyBegin(fact, MAT_FINAL_ASSEMBLY);
2365:   MatAssemblyEnd(fact, MAT_FINAL_ASSEMBLY);

2367:   PetscContainerCreate(((PetscObject)A)->comm, &c);
2368:   PetscContainerSetPointer(c, lgraph_p);
2369:   PetscObjectCompose((PetscObject) (fact), "graph", (PetscObject) c);
2370:   PetscContainerDestroy(&c);
2371:   return(0);
2372: }

2376: PetscErrorCode MatLUFactorNumeric_MPIAIJ(Mat B,Mat A, const MatFactorInfo *info)
2377: {
2379:   return(0);
2380: }

2384: /*
2385:   This uses the parallel ILU factorization of Peter Gottschling <pgottsch@osl.iu.edu>
2386: */
2387: PetscErrorCode MatSolve_MPIAIJ(Mat A, Vec b, Vec x)
2388: {
2389:   namespace graph_dist = boost::graph::distributed;

2391:   typedef graph_dist::ilu_default::ilu_level_graph_type  lgraph_type;
2392:   lgraph_type*   lgraph_p;
2393:   PetscContainer c;

2397:   PetscObjectQuery((PetscObject) A, "graph", (PetscObject *) &c);
2398:   PetscContainerGetPointer(c, (void **) &lgraph_p);
2399:   VecCopy(b, x);

2401:   PetscScalar* array_x;
2402:   VecGetArray(x, &array_x);
2403:   PetscInt sx;
2404:   VecGetSize(x, &sx);
2405: 
2406:   PetscScalar* array_b;
2407:   VecGetArray(b, &array_b);
2408:   PetscInt sb;
2409:   VecGetSize(b, &sb);

2411:   lgraph_type&   level_graph = *lgraph_p;
2412:   graph_dist::ilu_default::graph_type&            graph(level_graph.graph);

2414:   typedef boost::multi_array_ref<PetscScalar, 1> array_ref_type;
2415:   array_ref_type                                 ref_b(array_b, boost::extents[num_vertices(graph)]),
2416:                                                  ref_x(array_x, boost::extents[num_vertices(graph)]);

2418:   typedef boost::iterator_property_map<array_ref_type::iterator,
2419:                                 boost::property_map<graph_dist::ilu_default::graph_type, boost::vertex_index_t>::type>  gvector_type;
2420:   gvector_type                                   vector_b(ref_b.begin(), get(boost::vertex_index, graph)),
2421:                                                  vector_x(ref_x.begin(), get(boost::vertex_index, graph));
2422: 
2423:   ilu_set_solve(*lgraph_p, vector_b, vector_x);

2425:   return(0);
2426: }
2427: #endif

2429: typedef struct { /* used by MatGetRedundantMatrix() for reusing matredundant */
2430:   PetscInt       nzlocal,nsends,nrecvs;
2431:   PetscMPIInt    *send_rank,*recv_rank;
2432:   PetscInt       *sbuf_nz,*rbuf_nz,*sbuf_j,**rbuf_j;
2433:   PetscScalar    *sbuf_a,**rbuf_a;
2434:   PetscErrorCode (*Destroy)(Mat);
2435: } Mat_Redundant;

2439: PetscErrorCode PetscContainerDestroy_MatRedundant(void *ptr)
2440: {
2441:   PetscErrorCode       ierr;
2442:   Mat_Redundant        *redund=(Mat_Redundant*)ptr;
2443:   PetscInt             i;

2446:   PetscFree2(redund->send_rank,redund->recv_rank);
2447:   PetscFree(redund->sbuf_j);
2448:   PetscFree(redund->sbuf_a);
2449:   for (i=0; i<redund->nrecvs; i++){
2450:     PetscFree(redund->rbuf_j[i]);
2451:     PetscFree(redund->rbuf_a[i]);
2452:   }
2453:   PetscFree4(redund->sbuf_nz,redund->rbuf_nz,redund->rbuf_j,redund->rbuf_a);
2454:   PetscFree(redund);
2455:   return(0);
2456: }

2460: PetscErrorCode MatDestroy_MatRedundant(Mat A)
2461: {
2462:   PetscErrorCode  ierr;
2463:   PetscContainer  container;
2464:   Mat_Redundant   *redund=PETSC_NULL;

2467:   PetscObjectQuery((PetscObject)A,"Mat_Redundant",(PetscObject *)&container);
2468:   if (!container) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Container does not exit");
2469:   PetscContainerGetPointer(container,(void **)&redund);
2470:   A->ops->destroy = redund->Destroy;
2471:   PetscObjectCompose((PetscObject)A,"Mat_Redundant",0);
2472:   if (A->ops->destroy) {
2473:     (*A->ops->destroy)(A);
2474:   }
2475:   return(0);
2476: }

2480: PetscErrorCode MatGetRedundantMatrix_MPIAIJ(Mat mat,PetscInt nsubcomm,MPI_Comm subcomm,PetscInt mlocal_sub,MatReuse reuse,Mat *matredundant)
2481: {
2482:   PetscMPIInt    rank,size;
2483:   MPI_Comm       comm=((PetscObject)mat)->comm;
2485:   PetscInt       nsends=0,nrecvs=0,i,rownz_max=0;
2486:   PetscMPIInt    *send_rank=PETSC_NULL,*recv_rank=PETSC_NULL;
2487:   PetscInt       *rowrange=mat->rmap->range;
2488:   Mat_MPIAIJ     *aij = (Mat_MPIAIJ*)mat->data;
2489:   Mat            A=aij->A,B=aij->B,C=*matredundant;
2490:   Mat_SeqAIJ     *a=(Mat_SeqAIJ*)A->data,*b=(Mat_SeqAIJ*)B->data;
2491:   PetscScalar    *sbuf_a;
2492:   PetscInt       nzlocal=a->nz+b->nz;
2493:   PetscInt       j,cstart=mat->cmap->rstart,cend=mat->cmap->rend,row,nzA,nzB,ncols,*cworkA,*cworkB;
2494:   PetscInt       rstart=mat->rmap->rstart,rend=mat->rmap->rend,*bmap=aij->garray,M,N;
2495:   PetscInt       *cols,ctmp,lwrite,*rptr,l,*sbuf_j;
2496:   MatScalar      *aworkA,*aworkB;
2497:   PetscScalar    *vals;
2498:   PetscMPIInt    tag1,tag2,tag3,imdex;
2499:   MPI_Request    *s_waits1=PETSC_NULL,*s_waits2=PETSC_NULL,*s_waits3=PETSC_NULL,
2500:                  *r_waits1=PETSC_NULL,*r_waits2=PETSC_NULL,*r_waits3=PETSC_NULL;
2501:   MPI_Status     recv_status,*send_status;
2502:   PetscInt       *sbuf_nz=PETSC_NULL,*rbuf_nz=PETSC_NULL,count;
2503:   PetscInt       **rbuf_j=PETSC_NULL;
2504:   PetscScalar    **rbuf_a=PETSC_NULL;
2505:   Mat_Redundant  *redund=PETSC_NULL;
2506:   PetscContainer container;

2509:   MPI_Comm_rank(comm,&rank);
2510:   MPI_Comm_size(comm,&size);

2512:   if (reuse == MAT_REUSE_MATRIX) {
2513:     MatGetSize(C,&M,&N);
2514:     if (M != N || M != mat->rmap->N) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Cannot reuse matrix. Wrong global size");
2515:     MatGetLocalSize(C,&M,&N);
2516:     if (M != N || M != mlocal_sub) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Cannot reuse matrix. Wrong local size");
2517:     PetscObjectQuery((PetscObject)C,"Mat_Redundant",(PetscObject *)&container);
2518:     if (!container) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Container does not exit");
2519:     PetscContainerGetPointer(container,(void **)&redund);
2520:     if (nzlocal != redund->nzlocal) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Cannot reuse matrix. Wrong nzlocal");

2522:     nsends    = redund->nsends;
2523:     nrecvs    = redund->nrecvs;
2524:     send_rank = redund->send_rank;
2525:     recv_rank = redund->recv_rank;
2526:     sbuf_nz   = redund->sbuf_nz;
2527:     rbuf_nz   = redund->rbuf_nz;
2528:     sbuf_j    = redund->sbuf_j;
2529:     sbuf_a    = redund->sbuf_a;
2530:     rbuf_j    = redund->rbuf_j;
2531:     rbuf_a    = redund->rbuf_a;
2532:   }

2534:   if (reuse == MAT_INITIAL_MATRIX){
2535:     PetscMPIInt  subrank,subsize;
2536:     PetscInt     nleftover,np_subcomm;
2537:     /* get the destination processors' id send_rank, nsends and nrecvs */
2538:     MPI_Comm_rank(subcomm,&subrank);
2539:     MPI_Comm_size(subcomm,&subsize);
2540:     PetscMalloc2(size,PetscMPIInt,&send_rank,size,PetscMPIInt,&recv_rank);
2541:     np_subcomm = size/nsubcomm;
2542:     nleftover  = size - nsubcomm*np_subcomm;
2543:     nsends = 0; nrecvs = 0;
2544:     for (i=0; i<size; i++){ /* i=rank*/
2545:       if (subrank == i/nsubcomm && rank != i){ /* my_subrank == other's subrank */
2546:         send_rank[nsends] = i; nsends++;
2547:         recv_rank[nrecvs++] = i;
2548:       }
2549:     }
2550:     if (rank >= size - nleftover){/* this proc is a leftover processor */
2551:       i = size-nleftover-1;
2552:       j = 0;
2553:       while (j < nsubcomm - nleftover){
2554:         send_rank[nsends++] = i;
2555:         i--; j++;
2556:       }
2557:     }

2559:     if (nleftover && subsize == size/nsubcomm && subrank==subsize-1){ /* this proc recvs from leftover processors */
2560:       for (i=0; i<nleftover; i++){
2561:         recv_rank[nrecvs++] = size-nleftover+i;
2562:       }
2563:     }

2565:     /* allocate sbuf_j, sbuf_a */
2566:     i = nzlocal + rowrange[rank+1] - rowrange[rank] + 2;
2567:     PetscMalloc(i*sizeof(PetscInt),&sbuf_j);
2568:     PetscMalloc((nzlocal+1)*sizeof(PetscScalar),&sbuf_a);
2569:   } /* endof if (reuse == MAT_INITIAL_MATRIX) */

2571:   /* copy mat's local entries into the buffers */
2572:   if (reuse == MAT_INITIAL_MATRIX){
2573:     rownz_max = 0;
2574:     rptr = sbuf_j;
2575:     cols = sbuf_j + rend-rstart + 1;
2576:     vals = sbuf_a;
2577:     rptr[0] = 0;
2578:     for (i=0; i<rend-rstart; i++){
2579:       row = i + rstart;
2580:       nzA    = a->i[i+1] - a->i[i]; nzB = b->i[i+1] - b->i[i];
2581:       ncols  = nzA + nzB;
2582:       cworkA = a->j + a->i[i]; cworkB = b->j + b->i[i];
2583:       aworkA = a->a + a->i[i]; aworkB = b->a + b->i[i];
2584:       /* load the column indices for this row into cols */
2585:       lwrite = 0;
2586:       for (l=0; l<nzB; l++) {
2587:         if ((ctmp = bmap[cworkB[l]]) < cstart){
2588:           vals[lwrite]   = aworkB[l];
2589:           cols[lwrite++] = ctmp;
2590:         }
2591:       }
2592:       for (l=0; l<nzA; l++){
2593:         vals[lwrite]   = aworkA[l];
2594:         cols[lwrite++] = cstart + cworkA[l];
2595:       }
2596:       for (l=0; l<nzB; l++) {
2597:         if ((ctmp = bmap[cworkB[l]]) >= cend){
2598:           vals[lwrite]   = aworkB[l];
2599:           cols[lwrite++] = ctmp;
2600:         }
2601:       }
2602:       vals += ncols;
2603:       cols += ncols;
2604:       rptr[i+1] = rptr[i] + ncols;
2605:       if (rownz_max < ncols) rownz_max = ncols;
2606:     }
2607:     if (rptr[rend-rstart] != a->nz + b->nz) SETERRQ4(PETSC_COMM_SELF,PETSC_ERR_PLIB, "rptr[%d] %d != %d + %d",rend-rstart,rptr[rend-rstart+1],a->nz,b->nz);
2608:   } else { /* only copy matrix values into sbuf_a */
2609:     rptr = sbuf_j;
2610:     vals = sbuf_a;
2611:     rptr[0] = 0;
2612:     for (i=0; i<rend-rstart; i++){
2613:       row = i + rstart;
2614:       nzA    = a->i[i+1] - a->i[i]; nzB = b->i[i+1] - b->i[i];
2615:       ncols  = nzA + nzB;
2616:       cworkA = a->j + a->i[i]; cworkB = b->j + b->i[i];
2617:       aworkA = a->a + a->i[i]; aworkB = b->a + b->i[i];
2618:       lwrite = 0;
2619:       for (l=0; l<nzB; l++) {
2620:         if ((ctmp = bmap[cworkB[l]]) < cstart) vals[lwrite++] = aworkB[l];
2621:       }
2622:       for (l=0; l<nzA; l++) vals[lwrite++] = aworkA[l];
2623:       for (l=0; l<nzB; l++) {
2624:         if ((ctmp = bmap[cworkB[l]]) >= cend) vals[lwrite++] = aworkB[l];
2625:       }
2626:       vals += ncols;
2627:       rptr[i+1] = rptr[i] + ncols;
2628:     }
2629:   } /* endof if (reuse == MAT_INITIAL_MATRIX) */

2631:   /* send nzlocal to others, and recv other's nzlocal */
2632:   /*--------------------------------------------------*/
2633:   if (reuse == MAT_INITIAL_MATRIX){
2634:     PetscMalloc2(3*(nsends + nrecvs)+1,MPI_Request,&s_waits3,nsends+1,MPI_Status,&send_status);
2635:     s_waits2 = s_waits3 + nsends;
2636:     s_waits1 = s_waits2 + nsends;
2637:     r_waits1 = s_waits1 + nsends;
2638:     r_waits2 = r_waits1 + nrecvs;
2639:     r_waits3 = r_waits2 + nrecvs;
2640:   } else {
2641:     PetscMalloc2(nsends + nrecvs +1,MPI_Request,&s_waits3,nsends+1,MPI_Status,&send_status);
2642:     r_waits3 = s_waits3 + nsends;
2643:   }

2645:   PetscObjectGetNewTag((PetscObject)mat,&tag3);
2646:   if (reuse == MAT_INITIAL_MATRIX){
2647:     /* get new tags to keep the communication clean */
2648:     PetscObjectGetNewTag((PetscObject)mat,&tag1);
2649:     PetscObjectGetNewTag((PetscObject)mat,&tag2);
2650:     PetscMalloc4(nsends,PetscInt,&sbuf_nz,nrecvs,PetscInt,&rbuf_nz,nrecvs,PetscInt*,&rbuf_j,nrecvs,PetscScalar*,&rbuf_a);

2652:     /* post receives of other's nzlocal */
2653:     for (i=0; i<nrecvs; i++){
2654:       MPI_Irecv(rbuf_nz+i,1,MPIU_INT,MPI_ANY_SOURCE,tag1,comm,r_waits1+i);
2655:     }
2656:     /* send nzlocal to others */
2657:     for (i=0; i<nsends; i++){
2658:       sbuf_nz[i] = nzlocal;
2659:       MPI_Isend(sbuf_nz+i,1,MPIU_INT,send_rank[i],tag1,comm,s_waits1+i);
2660:     }
2661:     /* wait on receives of nzlocal; allocate space for rbuf_j, rbuf_a */
2662:     count = nrecvs;
2663:     while (count) {
2664:       MPI_Waitany(nrecvs,r_waits1,&imdex,&recv_status);
2665:       recv_rank[imdex] = recv_status.MPI_SOURCE;
2666:       /* allocate rbuf_a and rbuf_j; then post receives of rbuf_j */
2667:       PetscMalloc((rbuf_nz[imdex]+1)*sizeof(PetscScalar),&rbuf_a[imdex]);

2669:       i = rowrange[recv_status.MPI_SOURCE+1] - rowrange[recv_status.MPI_SOURCE]; /* number of expected mat->i */
2670:       rbuf_nz[imdex] += i + 2;
2671:       PetscMalloc(rbuf_nz[imdex]*sizeof(PetscInt),&rbuf_j[imdex]);
2672:       MPI_Irecv(rbuf_j[imdex],rbuf_nz[imdex],MPIU_INT,recv_status.MPI_SOURCE,tag2,comm,r_waits2+imdex);
2673:       count--;
2674:     }
2675:     /* wait on sends of nzlocal */
2676:     if (nsends) {MPI_Waitall(nsends,s_waits1,send_status);}
2677:     /* send mat->i,j to others, and recv from other's */
2678:     /*------------------------------------------------*/
2679:     for (i=0; i<nsends; i++){
2680:       j = nzlocal + rowrange[rank+1] - rowrange[rank] + 1;
2681:       MPI_Isend(sbuf_j,j,MPIU_INT,send_rank[i],tag2,comm,s_waits2+i);
2682:     }
2683:     /* wait on receives of mat->i,j */
2684:     /*------------------------------*/
2685:     count = nrecvs;
2686:     while (count) {
2687:       MPI_Waitany(nrecvs,r_waits2,&imdex,&recv_status);
2688:       if (recv_rank[imdex] != recv_status.MPI_SOURCE) SETERRQ2(PETSC_COMM_SELF,1, "recv_rank %d != MPI_SOURCE %d",recv_rank[imdex],recv_status.MPI_SOURCE);
2689:       count--;
2690:     }
2691:     /* wait on sends of mat->i,j */
2692:     /*---------------------------*/
2693:     if (nsends) {
2694:       MPI_Waitall(nsends,s_waits2,send_status);
2695:     }
2696:   } /* endof if (reuse == MAT_INITIAL_MATRIX) */

2698:   /* post receives, send and receive mat->a */
2699:   /*----------------------------------------*/
2700:   for (imdex=0; imdex<nrecvs; imdex++) {
2701:     MPI_Irecv(rbuf_a[imdex],rbuf_nz[imdex],MPIU_SCALAR,recv_rank[imdex],tag3,comm,r_waits3+imdex);
2702:   }
2703:   for (i=0; i<nsends; i++){
2704:     MPI_Isend(sbuf_a,nzlocal,MPIU_SCALAR,send_rank[i],tag3,comm,s_waits3+i);
2705:   }
2706:   count = nrecvs;
2707:   while (count) {
2708:     MPI_Waitany(nrecvs,r_waits3,&imdex,&recv_status);
2709:     if (recv_rank[imdex] != recv_status.MPI_SOURCE) SETERRQ2(PETSC_COMM_SELF,1, "recv_rank %d != MPI_SOURCE %d",recv_rank[imdex],recv_status.MPI_SOURCE);
2710:     count--;
2711:   }
2712:   if (nsends) {
2713:     MPI_Waitall(nsends,s_waits3,send_status);
2714:   }

2716:   PetscFree2(s_waits3,send_status);

2718:   /* create redundant matrix */
2719:   /*-------------------------*/
2720:   if (reuse == MAT_INITIAL_MATRIX){
2721:     /* compute rownz_max for preallocation */
2722:     for (imdex=0; imdex<nrecvs; imdex++){
2723:       j = rowrange[recv_rank[imdex]+1] - rowrange[recv_rank[imdex]];
2724:       rptr = rbuf_j[imdex];
2725:       for (i=0; i<j; i++){
2726:         ncols = rptr[i+1] - rptr[i];
2727:         if (rownz_max < ncols) rownz_max = ncols;
2728:       }
2729:     }

2731:     MatCreate(subcomm,&C);
2732:     MatSetSizes(C,mlocal_sub,mlocal_sub,PETSC_DECIDE,PETSC_DECIDE);
2733:     MatSetBlockSizes(C,mat->rmap->bs,mat->cmap->bs);
2734:     MatSetFromOptions(C);
2735:     MatSeqAIJSetPreallocation(C,rownz_max,PETSC_NULL);
2736:     MatMPIAIJSetPreallocation(C,rownz_max,PETSC_NULL,rownz_max,PETSC_NULL);
2737:   } else {
2738:     C = *matredundant;
2739:   }

2741:   /* insert local matrix entries */
2742:   rptr = sbuf_j;
2743:   cols = sbuf_j + rend-rstart + 1;
2744:   vals = sbuf_a;
2745:   for (i=0; i<rend-rstart; i++){
2746:     row   = i + rstart;
2747:     ncols = rptr[i+1] - rptr[i];
2748:     MatSetValues(C,1,&row,ncols,cols,vals,INSERT_VALUES);
2749:     vals += ncols;
2750:     cols += ncols;
2751:   }
2752:   /* insert received matrix entries */
2753:   for (imdex=0; imdex<nrecvs; imdex++){
2754:     rstart = rowrange[recv_rank[imdex]];
2755:     rend   = rowrange[recv_rank[imdex]+1];
2756:     rptr = rbuf_j[imdex];
2757:     cols = rbuf_j[imdex] + rend-rstart + 1;
2758:     vals = rbuf_a[imdex];
2759:     for (i=0; i<rend-rstart; i++){
2760:       row   = i + rstart;
2761:       ncols = rptr[i+1] - rptr[i];
2762:       MatSetValues(C,1,&row,ncols,cols,vals,INSERT_VALUES);
2763:       vals += ncols;
2764:       cols += ncols;
2765:     }
2766:   }
2767:   MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);
2768:   MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);
2769:   MatGetSize(C,&M,&N);
2770:   if (M != mat->rmap->N || N != mat->cmap->N) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_INCOMP,"redundant mat size %d != input mat size %d",M,mat->rmap->N);
2771:   if (reuse == MAT_INITIAL_MATRIX) {
2772:     PetscContainer container;
2773:     *matredundant = C;
2774:     /* create a supporting struct and attach it to C for reuse */
2775:     PetscNewLog(C,Mat_Redundant,&redund);
2776:     PetscContainerCreate(PETSC_COMM_SELF,&container);
2777:     PetscContainerSetPointer(container,redund);
2778:     PetscContainerSetUserDestroy(container,PetscContainerDestroy_MatRedundant);
2779:     PetscObjectCompose((PetscObject)C,"Mat_Redundant",(PetscObject)container);
2780:     PetscContainerDestroy(&container);

2782:     redund->nzlocal = nzlocal;
2783:     redund->nsends  = nsends;
2784:     redund->nrecvs  = nrecvs;
2785:     redund->send_rank = send_rank;
2786:     redund->recv_rank = recv_rank;
2787:     redund->sbuf_nz = sbuf_nz;
2788:     redund->rbuf_nz = rbuf_nz;
2789:     redund->sbuf_j  = sbuf_j;
2790:     redund->sbuf_a  = sbuf_a;
2791:     redund->rbuf_j  = rbuf_j;
2792:     redund->rbuf_a  = rbuf_a;

2794:     redund->Destroy = C->ops->destroy;
2795:     C->ops->destroy = MatDestroy_MatRedundant;
2796:   }
2797:   return(0);
2798: }

2802: PetscErrorCode MatGetRowMaxAbs_MPIAIJ(Mat A, Vec v, PetscInt idx[])
2803: {
2804:   Mat_MPIAIJ     *a = (Mat_MPIAIJ*)A->data;
2806:   PetscInt       i,*idxb = 0;
2807:   PetscScalar    *va,*vb;
2808:   Vec            vtmp;

2811:   MatGetRowMaxAbs(a->A,v,idx);
2812:   VecGetArray(v,&va);
2813:   if (idx) {
2814:     for (i=0; i<A->rmap->n; i++) {
2815:       if (PetscAbsScalar(va[i])) idx[i] += A->cmap->rstart;
2816:     }
2817:   }

2819:   VecCreateSeq(PETSC_COMM_SELF,A->rmap->n,&vtmp);
2820:   if (idx) {
2821:     PetscMalloc(A->rmap->n*sizeof(PetscInt),&idxb);
2822:   }
2823:   MatGetRowMaxAbs(a->B,vtmp,idxb);
2824:   VecGetArray(vtmp,&vb);

2826:   for (i=0; i<A->rmap->n; i++){
2827:     if (PetscAbsScalar(va[i]) < PetscAbsScalar(vb[i])) {
2828:       va[i] = vb[i];
2829:       if (idx) idx[i] = a->garray[idxb[i]];
2830:     }
2831:   }

2833:   VecRestoreArray(v,&va);
2834:   VecRestoreArray(vtmp,&vb);
2835:   PetscFree(idxb);
2836:   VecDestroy(&vtmp);
2837:   return(0);
2838: }

2842: PetscErrorCode MatGetRowMinAbs_MPIAIJ(Mat A, Vec v, PetscInt idx[])
2843: {
2844:   Mat_MPIAIJ     *a = (Mat_MPIAIJ*)A->data;
2846:   PetscInt       i,*idxb = 0;
2847:   PetscScalar    *va,*vb;
2848:   Vec            vtmp;

2851:   MatGetRowMinAbs(a->A,v,idx);
2852:   VecGetArray(v,&va);
2853:   if (idx) {
2854:     for (i=0; i<A->cmap->n; i++) {
2855:       if (PetscAbsScalar(va[i])) idx[i] += A->cmap->rstart;
2856:     }
2857:   }

2859:   VecCreateSeq(PETSC_COMM_SELF,A->rmap->n,&vtmp);
2860:   if (idx) {
2861:     PetscMalloc(A->rmap->n*sizeof(PetscInt),&idxb);
2862:   }
2863:   MatGetRowMinAbs(a->B,vtmp,idxb);
2864:   VecGetArray(vtmp,&vb);

2866:   for (i=0; i<A->rmap->n; i++){
2867:     if (PetscAbsScalar(va[i]) > PetscAbsScalar(vb[i])) {
2868:       va[i] = vb[i];
2869:       if (idx) idx[i] = a->garray[idxb[i]];
2870:     }
2871:   }

2873:   VecRestoreArray(v,&va);
2874:   VecRestoreArray(vtmp,&vb);
2875:   PetscFree(idxb);
2876:   VecDestroy(&vtmp);
2877:   return(0);
2878: }

2882: PetscErrorCode MatGetRowMin_MPIAIJ(Mat A, Vec v, PetscInt idx[])
2883: {
2884:   Mat_MPIAIJ    *mat    = (Mat_MPIAIJ *) A->data;
2885:   PetscInt       n      = A->rmap->n;
2886:   PetscInt       cstart = A->cmap->rstart;
2887:   PetscInt      *cmap   = mat->garray;
2888:   PetscInt      *diagIdx, *offdiagIdx;
2889:   Vec            diagV, offdiagV;
2890:   PetscScalar   *a, *diagA, *offdiagA;
2891:   PetscInt       r;

2895:   PetscMalloc2(n,PetscInt,&diagIdx,n,PetscInt,&offdiagIdx);
2896:   VecCreateSeq(((PetscObject)A)->comm, n, &diagV);
2897:   VecCreateSeq(((PetscObject)A)->comm, n, &offdiagV);
2898:   MatGetRowMin(mat->A, diagV,    diagIdx);
2899:   MatGetRowMin(mat->B, offdiagV, offdiagIdx);
2900:   VecGetArray(v,        &a);
2901:   VecGetArray(diagV,    &diagA);
2902:   VecGetArray(offdiagV, &offdiagA);
2903:   for(r = 0; r < n; ++r) {
2904:     if (PetscAbsScalar(diagA[r]) <= PetscAbsScalar(offdiagA[r])) {
2905:       a[r]   = diagA[r];
2906:       idx[r] = cstart + diagIdx[r];
2907:     } else {
2908:       a[r]   = offdiagA[r];
2909:       idx[r] = cmap[offdiagIdx[r]];
2910:     }
2911:   }
2912:   VecRestoreArray(v,        &a);
2913:   VecRestoreArray(diagV,    &diagA);
2914:   VecRestoreArray(offdiagV, &offdiagA);
2915:   VecDestroy(&diagV);
2916:   VecDestroy(&offdiagV);
2917:   PetscFree2(diagIdx, offdiagIdx);
2918:   return(0);
2919: }

2923: PetscErrorCode MatGetRowMax_MPIAIJ(Mat A, Vec v, PetscInt idx[])
2924: {
2925:   Mat_MPIAIJ    *mat    = (Mat_MPIAIJ *) A->data;
2926:   PetscInt       n      = A->rmap->n;
2927:   PetscInt       cstart = A->cmap->rstart;
2928:   PetscInt      *cmap   = mat->garray;
2929:   PetscInt      *diagIdx, *offdiagIdx;
2930:   Vec            diagV, offdiagV;
2931:   PetscScalar   *a, *diagA, *offdiagA;
2932:   PetscInt       r;

2936:   PetscMalloc2(n,PetscInt,&diagIdx,n,PetscInt,&offdiagIdx);
2937:   VecCreateSeq(((PetscObject)A)->comm, n, &diagV);
2938:   VecCreateSeq(((PetscObject)A)->comm, n, &offdiagV);
2939:   MatGetRowMax(mat->A, diagV,    diagIdx);
2940:   MatGetRowMax(mat->B, offdiagV, offdiagIdx);
2941:   VecGetArray(v,        &a);
2942:   VecGetArray(diagV,    &diagA);
2943:   VecGetArray(offdiagV, &offdiagA);
2944:   for(r = 0; r < n; ++r) {
2945:     if (PetscAbsScalar(diagA[r]) >= PetscAbsScalar(offdiagA[r])) {
2946:       a[r]   = diagA[r];
2947:       idx[r] = cstart + diagIdx[r];
2948:     } else {
2949:       a[r]   = offdiagA[r];
2950:       idx[r] = cmap[offdiagIdx[r]];
2951:     }
2952:   }
2953:   VecRestoreArray(v,        &a);
2954:   VecRestoreArray(diagV,    &diagA);
2955:   VecRestoreArray(offdiagV, &offdiagA);
2956:   VecDestroy(&diagV);
2957:   VecDestroy(&offdiagV);
2958:   PetscFree2(diagIdx, offdiagIdx);
2959:   return(0);
2960: }

2964: PetscErrorCode MatGetSeqNonzeroStructure_MPIAIJ(Mat mat,Mat *newmat)
2965: {
2967:   Mat            *dummy;

2970:   MatGetSubMatrix_MPIAIJ_All(mat,MAT_DO_NOT_GET_VALUES,MAT_INITIAL_MATRIX,&dummy);
2971:   *newmat = *dummy;
2972:   PetscFree(dummy);
2973:   return(0);
2974: }

2976: extern PetscErrorCode  MatFDColoringApply_AIJ(Mat,MatFDColoring,Vec,MatStructure*,void*);

2980: PetscErrorCode  MatInvertBlockDiagonal_MPIAIJ(Mat A,const PetscScalar **values)
2981: {
2982:   Mat_MPIAIJ    *a = (Mat_MPIAIJ*) A->data;

2986:   MatInvertBlockDiagonal(a->A,values);
2987:   return(0);
2988: }


2991: /* -------------------------------------------------------------------*/
2992: static struct _MatOps MatOps_Values = {MatSetValues_MPIAIJ,
2993:        MatGetRow_MPIAIJ,
2994:        MatRestoreRow_MPIAIJ,
2995:        MatMult_MPIAIJ,
2996: /* 4*/ MatMultAdd_MPIAIJ,
2997:        MatMultTranspose_MPIAIJ,
2998:        MatMultTransposeAdd_MPIAIJ,
2999: #ifdef PETSC_HAVE_PBGL
3000:        MatSolve_MPIAIJ,
3001: #else
3002:        0,
3003: #endif
3004:        0,
3005:        0,
3006: /*10*/ 0,
3007:        0,
3008:        0,
3009:        MatSOR_MPIAIJ,
3010:        MatTranspose_MPIAIJ,
3011: /*15*/ MatGetInfo_MPIAIJ,
3012:        MatEqual_MPIAIJ,
3013:        MatGetDiagonal_MPIAIJ,
3014:        MatDiagonalScale_MPIAIJ,
3015:        MatNorm_MPIAIJ,
3016: /*20*/ MatAssemblyBegin_MPIAIJ,
3017:        MatAssemblyEnd_MPIAIJ,
3018:        MatSetOption_MPIAIJ,
3019:        MatZeroEntries_MPIAIJ,
3020: /*24*/ MatZeroRows_MPIAIJ,
3021:        0,
3022: #ifdef PETSC_HAVE_PBGL
3023:        0,
3024: #else
3025:        0,
3026: #endif
3027:        0,
3028:        0,
3029: /*29*/ MatSetUp_MPIAIJ,
3030: #ifdef PETSC_HAVE_PBGL
3031:        0,
3032: #else
3033:        0,
3034: #endif
3035:        0,
3036:        0,
3037:        0,
3038: /*34*/ MatDuplicate_MPIAIJ,
3039:        0,
3040:        0,
3041:        0,
3042:        0,
3043: /*39*/ MatAXPY_MPIAIJ,
3044:        MatGetSubMatrices_MPIAIJ,
3045:        MatIncreaseOverlap_MPIAIJ,
3046:        MatGetValues_MPIAIJ,
3047:        MatCopy_MPIAIJ,
3048: /*44*/ MatGetRowMax_MPIAIJ,
3049:        MatScale_MPIAIJ,
3050:        0,
3051:        0,
3052:        MatZeroRowsColumns_MPIAIJ,
3053: /*49*/ 0,
3054:        0,
3055:        0,
3056:        0,
3057:        0,
3058: /*54*/ MatFDColoringCreate_MPIAIJ,
3059:        0,
3060:        MatSetUnfactored_MPIAIJ,
3061:        0, /* MatPermute_MPIAIJ, impl currently broken */
3062:        0,
3063: /*59*/ MatGetSubMatrix_MPIAIJ,
3064:        MatDestroy_MPIAIJ,
3065:        MatView_MPIAIJ,
3066:        0,
3067:        0,
3068: /*64*/ 0,
3069:        0,
3070:        0,
3071:        0,
3072:        0,
3073: /*69*/ MatGetRowMaxAbs_MPIAIJ,
3074:        MatGetRowMinAbs_MPIAIJ,
3075:        0,
3076:        MatSetColoring_MPIAIJ,
3077: #if defined(PETSC_HAVE_ADIC)
3078:        MatSetValuesAdic_MPIAIJ,
3079: #else
3080:        0,
3081: #endif
3082:        MatSetValuesAdifor_MPIAIJ,
3083: /*75*/ MatFDColoringApply_AIJ,
3084:        0,
3085:        0,
3086:        0,
3087:        0,
3088: /*80*/ 0,
3089:        0,
3090:        0,
3091: /*83*/ MatLoad_MPIAIJ,
3092:        0,
3093:        0,
3094:        0,
3095:        0,
3096:        0,
3097: /*89*/ MatMatMult_MPIAIJ_MPIAIJ,
3098:        MatMatMultSymbolic_MPIAIJ_MPIAIJ,
3099:        MatMatMultNumeric_MPIAIJ_MPIAIJ,
3100:        MatPtAP_Basic,
3101:        MatPtAPSymbolic_MPIAIJ,
3102: /*94*/ MatPtAPNumeric_MPIAIJ,
3103:        0,
3104:        0,
3105:        0,
3106:        0,
3107: /*99*/ 0,
3108:        MatPtAPSymbolic_MPIAIJ_MPIAIJ,
3109:        MatPtAPNumeric_MPIAIJ_MPIAIJ,
3110:        MatConjugate_MPIAIJ,
3111:        0,
3112: /*104*/MatSetValuesRow_MPIAIJ,
3113:        MatRealPart_MPIAIJ,
3114:        MatImaginaryPart_MPIAIJ,
3115:        0,
3116:        0,
3117: /*109*/0,
3118:        MatGetRedundantMatrix_MPIAIJ,
3119:        MatGetRowMin_MPIAIJ,
3120:        0,
3121:        0,
3122: /*114*/MatGetSeqNonzeroStructure_MPIAIJ,
3123:        0,
3124:        0,
3125:        0,
3126:        0,
3127: /*119*/0,
3128:        0,
3129:        0,
3130:        0,
3131:        MatGetMultiProcBlock_MPIAIJ,
3132: /*124*/MatFindNonZeroRows_MPIAIJ,
3133:        MatGetColumnNorms_MPIAIJ,
3134:        MatInvertBlockDiagonal_MPIAIJ,
3135:        0,
3136:        MatGetSubMatricesParallel_MPIAIJ,
3137: /*129*/0,
3138:        MatTransposeMatMult_MPIAIJ_MPIAIJ,
3139:        MatTransposeMatMultSymbolic_MPIAIJ_MPIAIJ,
3140:        MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ,
3141:        0,
3142: /*134*/0,
3143:        0,
3144:        0,
3145:        0,
3146:        0
3147: };

3149: /* ----------------------------------------------------------------------------------------*/

3151: EXTERN_C_BEGIN
3154: PetscErrorCode  MatStoreValues_MPIAIJ(Mat mat)
3155: {
3156:   Mat_MPIAIJ     *aij = (Mat_MPIAIJ *)mat->data;

3160:   MatStoreValues(aij->A);
3161:   MatStoreValues(aij->B);
3162:   return(0);
3163: }
3164: EXTERN_C_END

3166: EXTERN_C_BEGIN
3169: PetscErrorCode  MatRetrieveValues_MPIAIJ(Mat mat)
3170: {
3171:   Mat_MPIAIJ     *aij = (Mat_MPIAIJ *)mat->data;

3175:   MatRetrieveValues(aij->A);
3176:   MatRetrieveValues(aij->B);
3177:   return(0);
3178: }
3179: EXTERN_C_END

3181: EXTERN_C_BEGIN
3184: PetscErrorCode  MatMPIAIJSetPreallocation_MPIAIJ(Mat B,PetscInt d_nz,const PetscInt d_nnz[],PetscInt o_nz,const PetscInt o_nnz[])
3185: {
3186:   Mat_MPIAIJ     *b;
3188:   PetscInt       i;
3189:   PetscBool      d_realalloc = PETSC_FALSE,o_realalloc = PETSC_FALSE;

3192:   if (d_nz >= 0 || d_nnz) d_realalloc = PETSC_TRUE;
3193:   if (o_nz >= 0 || o_nnz) o_realalloc = PETSC_TRUE;
3194:   if (d_nz == PETSC_DEFAULT || d_nz == PETSC_DECIDE) d_nz = 5;
3195:   if (o_nz == PETSC_DEFAULT || o_nz == PETSC_DECIDE) o_nz = 2;
3196:   if (d_nz < 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"d_nz cannot be less than 0: value %D",d_nz);
3197:   if (o_nz < 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"o_nz cannot be less than 0: value %D",o_nz);

3199:   PetscLayoutSetUp(B->rmap);
3200:   PetscLayoutSetUp(B->cmap);
3201:   if (d_nnz) {
3202:     for (i=0; i<B->rmap->n; i++) {
3203:       if (d_nnz[i] < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"d_nnz cannot be less than 0: local row %D value %D",i,d_nnz[i]);
3204:     }
3205:   }
3206:   if (o_nnz) {
3207:     for (i=0; i<B->rmap->n; i++) {
3208:       if (o_nnz[i] < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"o_nnz cannot be less than 0: local row %D value %D",i,o_nnz[i]);
3209:     }
3210:   }
3211:   b = (Mat_MPIAIJ*)B->data;

3213:   if (!B->preallocated) {
3214:     /* Explicitly create 2 MATSEQAIJ matrices. */
3215:     MatCreate(PETSC_COMM_SELF,&b->A);
3216:     MatSetSizes(b->A,B->rmap->n,B->cmap->n,B->rmap->n,B->cmap->n);
3217:     MatSetBlockSizes(b->A,B->rmap->bs,B->cmap->bs);
3218:     MatSetType(b->A,MATSEQAIJ);
3219:     PetscLogObjectParent(B,b->A);
3220:     MatCreate(PETSC_COMM_SELF,&b->B);
3221:     MatSetSizes(b->B,B->rmap->n,B->cmap->N,B->rmap->n,B->cmap->N);
3222:     MatSetBlockSizes(b->B,B->rmap->bs,B->cmap->bs);
3223:     MatSetType(b->B,MATSEQAIJ);
3224:     PetscLogObjectParent(B,b->B);
3225:   }

3227:   MatSeqAIJSetPreallocation(b->A,d_nz,d_nnz);
3228:   MatSeqAIJSetPreallocation(b->B,o_nz,o_nnz);
3229:   /* Do not error if the user did not give real preallocation information. Ugly because this would overwrite a previous user call to MatSetOption(). */
3230:   if (!d_realalloc) {MatSetOption(b->A,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);}
3231:   if (!o_realalloc) {MatSetOption(b->B,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);}
3232:   B->preallocated = PETSC_TRUE;
3233:   return(0);
3234: }
3235: EXTERN_C_END

3239: PetscErrorCode MatDuplicate_MPIAIJ(Mat matin,MatDuplicateOption cpvalues,Mat *newmat)
3240: {
3241:   Mat            mat;
3242:   Mat_MPIAIJ     *a,*oldmat = (Mat_MPIAIJ*)matin->data;

3246:   *newmat       = 0;
3247:   MatCreate(((PetscObject)matin)->comm,&mat);
3248:   MatSetSizes(mat,matin->rmap->n,matin->cmap->n,matin->rmap->N,matin->cmap->N);
3249:   MatSetBlockSizes(mat,matin->rmap->bs,matin->cmap->bs);
3250:   MatSetType(mat,((PetscObject)matin)->type_name);
3251:   PetscMemcpy(mat->ops,matin->ops,sizeof(struct _MatOps));
3252:   a    = (Mat_MPIAIJ*)mat->data;
3253: 
3254:   mat->factortype    = matin->factortype;
3255:   mat->rmap->bs      = matin->rmap->bs;
3256:   mat->cmap->bs      = matin->cmap->bs;
3257:   mat->assembled    = PETSC_TRUE;
3258:   mat->insertmode   = NOT_SET_VALUES;
3259:   mat->preallocated = PETSC_TRUE;

3261:   a->size           = oldmat->size;
3262:   a->rank           = oldmat->rank;
3263:   a->donotstash     = oldmat->donotstash;
3264:   a->roworiented    = oldmat->roworiented;
3265:   a->rowindices     = 0;
3266:   a->rowvalues      = 0;
3267:   a->getrowactive   = PETSC_FALSE;

3269:   PetscLayoutReference(matin->rmap,&mat->rmap);
3270:   PetscLayoutReference(matin->cmap,&mat->cmap);

3272:   if (oldmat->colmap) {
3273: #if defined (PETSC_USE_CTABLE)
3274:     PetscTableCreateCopy(oldmat->colmap,&a->colmap);
3275: #else
3276:     PetscMalloc((mat->cmap->N)*sizeof(PetscInt),&a->colmap);
3277:     PetscLogObjectMemory(mat,(mat->cmap->N)*sizeof(PetscInt));
3278:     PetscMemcpy(a->colmap,oldmat->colmap,(mat->cmap->N)*sizeof(PetscInt));
3279: #endif
3280:   } else a->colmap = 0;
3281:   if (oldmat->garray) {
3282:     PetscInt len;
3283:     len  = oldmat->B->cmap->n;
3284:     PetscMalloc((len+1)*sizeof(PetscInt),&a->garray);
3285:     PetscLogObjectMemory(mat,len*sizeof(PetscInt));
3286:     if (len) { PetscMemcpy(a->garray,oldmat->garray,len*sizeof(PetscInt)); }
3287:   } else a->garray = 0;
3288: 
3289:   VecDuplicate(oldmat->lvec,&a->lvec);
3290:   PetscLogObjectParent(mat,a->lvec);
3291:   VecScatterCopy(oldmat->Mvctx,&a->Mvctx);
3292:   PetscLogObjectParent(mat,a->Mvctx);
3293:   MatDuplicate(oldmat->A,cpvalues,&a->A);
3294:   PetscLogObjectParent(mat,a->A);
3295:   MatDuplicate(oldmat->B,cpvalues,&a->B);
3296:   PetscLogObjectParent(mat,a->B);
3297:   PetscFListDuplicate(((PetscObject)matin)->qlist,&((PetscObject)mat)->qlist);
3298:   *newmat = mat;
3299:   return(0);
3300: }



3306: PetscErrorCode MatLoad_MPIAIJ(Mat newMat, PetscViewer viewer)
3307: {
3308:   PetscScalar    *vals,*svals;
3309:   MPI_Comm       comm = ((PetscObject)viewer)->comm;
3311:   PetscMPIInt    rank,size,tag = ((PetscObject)viewer)->tag;
3312:   PetscInt       i,nz,j,rstart,rend,mmax,maxnz = 0,grows,gcols;
3313:   PetscInt       header[4],*rowlengths = 0,M,N,m,*cols;
3314:   PetscInt       *ourlens = PETSC_NULL,*procsnz = PETSC_NULL,*offlens = PETSC_NULL,jj,*mycols,*smycols;
3315:   PetscInt       cend,cstart,n,*rowners,sizesset=1;
3316:   int            fd;

3319:   MPI_Comm_size(comm,&size);
3320:   MPI_Comm_rank(comm,&rank);
3321:   if (!rank) {
3322:     PetscViewerBinaryGetDescriptor(viewer,&fd);
3323:     PetscBinaryRead(fd,(char *)header,4,PETSC_INT);
3324:     if (header[0] != MAT_FILE_CLASSID) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED,"not matrix object");
3325:   }

3327:   if (newMat->rmap->n < 0 && newMat->rmap->N < 0 && newMat->cmap->n < 0 && newMat->cmap->N < 0) sizesset = 0;

3329:   MPI_Bcast(header+1,3,MPIU_INT,0,comm);
3330:   M = header[1]; N = header[2];
3331:   /* If global rows/cols are set to PETSC_DECIDE, set it to the sizes given in the file */
3332:   if (sizesset && newMat->rmap->N < 0) newMat->rmap->N = M;
3333:   if (sizesset && newMat->cmap->N < 0) newMat->cmap->N = N;
3334: 
3335:   /* If global sizes are set, check if they are consistent with that given in the file */
3336:   if (sizesset) {
3337:     MatGetSize(newMat,&grows,&gcols);
3338:   }
3339:   if (sizesset && newMat->rmap->N != grows) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED, "Inconsistent # of rows:Matrix in file has (%d) and input matrix has (%d)",M,grows);
3340:   if (sizesset && newMat->cmap->N != gcols) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED, "Inconsistent # of cols:Matrix in file has (%d) and input matrix has (%d)",N,gcols);

3342:   /* determine ownership of all rows */
3343:   if (newMat->rmap->n < 0 ) m    = M/size + ((M % size) > rank); /* PETSC_DECIDE */
3344:   else m = newMat->rmap->n; /* Set by user */
3345: 
3346:   PetscMalloc((size+1)*sizeof(PetscInt),&rowners);
3347:   MPI_Allgather(&m,1,MPIU_INT,rowners+1,1,MPIU_INT,comm);

3349:   /* First process needs enough room for process with most rows */
3350:   if (!rank) {
3351:     mmax       = rowners[1];
3352:     for (i=2; i<size; i++) {
3353:       mmax = PetscMax(mmax,rowners[i]);
3354:     }
3355:   } else mmax = m;

3357:   rowners[0] = 0;
3358:   for (i=2; i<=size; i++) {
3359:     rowners[i] += rowners[i-1];
3360:   }
3361:   rstart = rowners[rank];
3362:   rend   = rowners[rank+1];

3364:   /* distribute row lengths to all processors */
3365:   PetscMalloc2(mmax,PetscInt,&ourlens,mmax,PetscInt,&offlens);
3366:   if (!rank) {
3367:     PetscBinaryRead(fd,ourlens,m,PETSC_INT);
3368:     PetscMalloc(m*sizeof(PetscInt),&rowlengths);
3369:     PetscMalloc(size*sizeof(PetscInt),&procsnz);
3370:     PetscMemzero(procsnz,size*sizeof(PetscInt));
3371:     for (j=0; j<m; j++) {
3372:       procsnz[0] += ourlens[j];
3373:     }
3374:     for (i=1; i<size; i++) {
3375:       PetscBinaryRead(fd,rowlengths,rowners[i+1]-rowners[i],PETSC_INT);
3376:       /* calculate the number of nonzeros on each processor */
3377:       for (j=0; j<rowners[i+1]-rowners[i]; j++) {
3378:         procsnz[i] += rowlengths[j];
3379:       }
3380:       MPIULong_Send(rowlengths,rowners[i+1]-rowners[i],MPIU_INT,i,tag,comm);
3381:     }
3382:     PetscFree(rowlengths);
3383:   } else {
3384:     MPIULong_Recv(ourlens,m,MPIU_INT,0,tag,comm);
3385:   }

3387:   if (!rank) {
3388:     /* determine max buffer needed and allocate it */
3389:     maxnz = 0;
3390:     for (i=0; i<size; i++) {
3391:       maxnz = PetscMax(maxnz,procsnz[i]);
3392:     }
3393:     PetscMalloc(maxnz*sizeof(PetscInt),&cols);

3395:     /* read in my part of the matrix column indices  */
3396:     nz   = procsnz[0];
3397:     PetscMalloc(nz*sizeof(PetscInt),&mycols);
3398:     PetscBinaryRead(fd,mycols,nz,PETSC_INT);

3400:     /* read in every one elses and ship off */
3401:     for (i=1; i<size; i++) {
3402:       nz     = procsnz[i];
3403:       PetscBinaryRead(fd,cols,nz,PETSC_INT);
3404:       MPIULong_Send(cols,nz,MPIU_INT,i,tag,comm);
3405:     }
3406:     PetscFree(cols);
3407:   } else {
3408:     /* determine buffer space needed for message */
3409:     nz = 0;
3410:     for (i=0; i<m; i++) {
3411:       nz += ourlens[i];
3412:     }
3413:     PetscMalloc(nz*sizeof(PetscInt),&mycols);

3415:     /* receive message of column indices*/
3416:     MPIULong_Recv(mycols,nz,MPIU_INT,0,tag,comm);
3417:   }

3419:   /* determine column ownership if matrix is not square */
3420:   if (N != M) {
3421:     if (newMat->cmap->n < 0) n      = N/size + ((N % size) > rank);
3422:     else n = newMat->cmap->n;
3423:     MPI_Scan(&n,&cend,1,MPIU_INT,MPI_SUM,comm);
3424:     cstart = cend - n;
3425:   } else {
3426:     cstart = rstart;
3427:     cend   = rend;
3428:     n      = cend - cstart;
3429:   }

3431:   /* loop over local rows, determining number of off diagonal entries */
3432:   PetscMemzero(offlens,m*sizeof(PetscInt));
3433:   jj = 0;
3434:   for (i=0; i<m; i++) {
3435:     for (j=0; j<ourlens[i]; j++) {
3436:       if (mycols[jj] < cstart || mycols[jj] >= cend) offlens[i]++;
3437:       jj++;
3438:     }
3439:   }

3441:   for (i=0; i<m; i++) {
3442:     ourlens[i] -= offlens[i];
3443:   }
3444:   if (!sizesset) {
3445:     MatSetSizes(newMat,m,n,M,N);
3446:   }
3447:   MatMPIAIJSetPreallocation(newMat,0,ourlens,0,offlens);

3449:   for (i=0; i<m; i++) {
3450:     ourlens[i] += offlens[i];
3451:   }

3453:   if (!rank) {
3454:     PetscMalloc((maxnz+1)*sizeof(PetscScalar),&vals);

3456:     /* read in my part of the matrix numerical values  */
3457:     nz   = procsnz[0];
3458:     PetscBinaryRead(fd,vals,nz,PETSC_SCALAR);
3459: 
3460:     /* insert into matrix */
3461:     jj      = rstart;
3462:     smycols = mycols;
3463:     svals   = vals;
3464:     for (i=0; i<m; i++) {
3465:       MatSetValues_MPIAIJ(newMat,1,&jj,ourlens[i],smycols,svals,INSERT_VALUES);
3466:       smycols += ourlens[i];
3467:       svals   += ourlens[i];
3468:       jj++;
3469:     }

3471:     /* read in other processors and ship out */
3472:     for (i=1; i<size; i++) {
3473:       nz     = procsnz[i];
3474:       PetscBinaryRead(fd,vals,nz,PETSC_SCALAR);
3475:       MPIULong_Send(vals,nz,MPIU_SCALAR,i,((PetscObject)newMat)->tag,comm);
3476:     }
3477:     PetscFree(procsnz);
3478:   } else {
3479:     /* receive numeric values */
3480:     PetscMalloc((nz+1)*sizeof(PetscScalar),&vals);

3482:     /* receive message of values*/
3483:     MPIULong_Recv(vals,nz,MPIU_SCALAR,0,((PetscObject)newMat)->tag,comm);

3485:     /* insert into matrix */
3486:     jj      = rstart;
3487:     smycols = mycols;
3488:     svals   = vals;
3489:     for (i=0; i<m; i++) {
3490:       MatSetValues_MPIAIJ(newMat,1,&jj,ourlens[i],smycols,svals,INSERT_VALUES);
3491:       smycols += ourlens[i];
3492:       svals   += ourlens[i];
3493:       jj++;
3494:     }
3495:   }
3496:   PetscFree2(ourlens,offlens);
3497:   PetscFree(vals);
3498:   PetscFree(mycols);
3499:   PetscFree(rowners);

3501:   MatAssemblyBegin(newMat,MAT_FINAL_ASSEMBLY);
3502:   MatAssemblyEnd(newMat,MAT_FINAL_ASSEMBLY);
3503:   return(0);
3504: }

3508: PetscErrorCode MatGetSubMatrix_MPIAIJ(Mat mat,IS isrow,IS iscol,MatReuse call,Mat *newmat)
3509: {
3511:   IS             iscol_local;
3512:   PetscInt       csize;

3515:   ISGetLocalSize(iscol,&csize);
3516:   if (call == MAT_REUSE_MATRIX) {
3517:     PetscObjectQuery((PetscObject)*newmat,"ISAllGather",(PetscObject*)&iscol_local);
3518:     if (!iscol_local) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Submatrix passed in was not used before, cannot reuse");
3519:   } else {
3520:     PetscInt cbs;
3521:     ISGetBlockSize(iscol,&cbs);
3522:     ISAllGather(iscol,&iscol_local);
3523:     ISSetBlockSize(iscol_local,cbs);
3524:   }
3525:   MatGetSubMatrix_MPIAIJ_Private(mat,isrow,iscol_local,csize,call,newmat);
3526:   if (call == MAT_INITIAL_MATRIX) {
3527:     PetscObjectCompose((PetscObject)*newmat,"ISAllGather",(PetscObject)iscol_local);
3528:     ISDestroy(&iscol_local);
3529:   }
3530:   return(0);
3531: }

3533: extern PetscErrorCode MatGetSubMatrices_MPIAIJ_Local(Mat,PetscInt,const IS[],const IS[],MatReuse,PetscBool*,Mat*);
3536: /*
3537:     Not great since it makes two copies of the submatrix, first an SeqAIJ 
3538:   in local and then by concatenating the local matrices the end result.
3539:   Writing it directly would be much like MatGetSubMatrices_MPIAIJ()

3541:   Note: This requires a sequential iscol with all indices.
3542: */
3543: PetscErrorCode MatGetSubMatrix_MPIAIJ_Private(Mat mat,IS isrow,IS iscol,PetscInt csize,MatReuse call,Mat *newmat)
3544: {
3546:   PetscMPIInt    rank,size;
3547:   PetscInt       i,m,n,rstart,row,rend,nz,*cwork,j,bs,cbs;
3548:   PetscInt       *ii,*jj,nlocal,*dlens,*olens,dlen,olen,jend,mglobal,ncol;
3549:   PetscBool      allcolumns, colflag;
3550:   Mat            M,Mreuse;
3551:   MatScalar      *vwork,*aa;
3552:   MPI_Comm       comm = ((PetscObject)mat)->comm;
3553:   Mat_SeqAIJ     *aij;


3557:   MPI_Comm_rank(comm,&rank);
3558:   MPI_Comm_size(comm,&size);

3560:   ISIdentity(iscol,&colflag);
3561:   ISGetLocalSize(iscol,&ncol);
3562:   if (colflag && ncol == mat->cmap->N){
3563:     allcolumns = PETSC_TRUE;
3564:   } else {
3565:     allcolumns = PETSC_FALSE;
3566:   }
3567:   if (call ==  MAT_REUSE_MATRIX) {
3568:     PetscObjectQuery((PetscObject)*newmat,"SubMatrix",(PetscObject *)&Mreuse);
3569:     if (!Mreuse) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Submatrix passed in was not used before, cannot reuse");
3570:     MatGetSubMatrices_MPIAIJ_Local(mat,1,&isrow,&iscol,MAT_REUSE_MATRIX,&allcolumns,&Mreuse);
3571:   } else {
3572:     MatGetSubMatrices_MPIAIJ_Local(mat,1,&isrow,&iscol,MAT_INITIAL_MATRIX,&allcolumns,&Mreuse);
3573:   }

3575:   /* 
3576:       m - number of local rows
3577:       n - number of columns (same on all processors)
3578:       rstart - first row in new global matrix generated
3579:   */
3580:   MatGetSize(Mreuse,&m,&n);
3581:   MatGetBlockSizes(Mreuse,&bs,&cbs);
3582:   if (call == MAT_INITIAL_MATRIX) {
3583:     aij = (Mat_SeqAIJ*)(Mreuse)->data;
3584:     ii  = aij->i;
3585:     jj  = aij->j;

3587:     /*
3588:         Determine the number of non-zeros in the diagonal and off-diagonal 
3589:         portions of the matrix in order to do correct preallocation
3590:     */

3592:     /* first get start and end of "diagonal" columns */
3593:     if (csize == PETSC_DECIDE) {
3594:       ISGetSize(isrow,&mglobal);
3595:       if (mglobal == n) { /* square matrix */
3596:         nlocal = m;
3597:       } else {
3598:         nlocal = n/size + ((n % size) > rank);
3599:       }
3600:     } else {
3601:       nlocal = csize;
3602:     }
3603:     MPI_Scan(&nlocal,&rend,1,MPIU_INT,MPI_SUM,comm);
3604:     rstart = rend - nlocal;
3605:     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);

3607:     /* next, compute all the lengths */
3608:     PetscMalloc((2*m+1)*sizeof(PetscInt),&dlens);
3609:     olens = dlens + m;
3610:     for (i=0; i<m; i++) {
3611:       jend = ii[i+1] - ii[i];
3612:       olen = 0;
3613:       dlen = 0;
3614:       for (j=0; j<jend; j++) {
3615:         if (*jj < rstart || *jj >= rend) olen++;
3616:         else dlen++;
3617:         jj++;
3618:       }
3619:       olens[i] = olen;
3620:       dlens[i] = dlen;
3621:     }
3622:     MatCreate(comm,&M);
3623:     MatSetSizes(M,m,nlocal,PETSC_DECIDE,n);
3624:     MatSetBlockSizes(M,bs,cbs);
3625:     MatSetType(M,((PetscObject)mat)->type_name);
3626:     MatMPIAIJSetPreallocation(M,0,dlens,0,olens);
3627:     PetscFree(dlens);
3628:   } else {
3629:     PetscInt ml,nl;

3631:     M = *newmat;
3632:     MatGetLocalSize(M,&ml,&nl);
3633:     if (ml != m) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Previous matrix must be same size/layout as request");
3634:     MatZeroEntries(M);
3635:     /*
3636:          The next two lines are needed so we may call MatSetValues_MPIAIJ() below directly,
3637:        rather than the slower MatSetValues().
3638:     */
3639:     M->was_assembled = PETSC_TRUE;
3640:     M->assembled     = PETSC_FALSE;
3641:   }
3642:   MatGetOwnershipRange(M,&rstart,&rend);
3643:   aij = (Mat_SeqAIJ*)(Mreuse)->data;
3644:   ii  = aij->i;
3645:   jj  = aij->j;
3646:   aa  = aij->a;
3647:   for (i=0; i<m; i++) {
3648:     row   = rstart + i;
3649:     nz    = ii[i+1] - ii[i];
3650:     cwork = jj;     jj += nz;
3651:     vwork = aa;     aa += nz;
3652:     MatSetValues_MPIAIJ(M,1,&row,nz,cwork,vwork,INSERT_VALUES);
3653:   }

3655:   MatAssemblyBegin(M,MAT_FINAL_ASSEMBLY);
3656:   MatAssemblyEnd(M,MAT_FINAL_ASSEMBLY);
3657:   *newmat = M;

3659:   /* save submatrix used in processor for next request */
3660:   if (call ==  MAT_INITIAL_MATRIX) {
3661:     PetscObjectCompose((PetscObject)M,"SubMatrix",(PetscObject)Mreuse);
3662:     MatDestroy(&Mreuse);
3663:   }

3665:   return(0);
3666: }

3668: EXTERN_C_BEGIN
3671: PetscErrorCode  MatMPIAIJSetPreallocationCSR_MPIAIJ(Mat B,const PetscInt Ii[],const PetscInt J[],const PetscScalar v[])
3672: {
3673:   PetscInt       m,cstart, cend,j,nnz,i,d;
3674:   PetscInt       *d_nnz,*o_nnz,nnz_max = 0,rstart,ii;
3675:   const PetscInt *JJ;
3676:   PetscScalar    *values;

3680:   if (Ii[0]) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Ii[0] must be 0 it is %D",Ii[0]);

3682:   PetscLayoutSetUp(B->rmap);
3683:   PetscLayoutSetUp(B->cmap);
3684:   m      = B->rmap->n;
3685:   cstart = B->cmap->rstart;
3686:   cend   = B->cmap->rend;
3687:   rstart = B->rmap->rstart;

3689:   PetscMalloc2(m,PetscInt,&d_nnz,m,PetscInt,&o_nnz);

3691: #if defined(PETSC_USE_DEBUGGING)
3692:   for (i=0; i<m; i++) {
3693:     nnz     = Ii[i+1]- Ii[i];
3694:     JJ      = J + Ii[i];
3695:     if (nnz < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Local row %D has a negative %D number of columns",i,nnz);
3696:     if (nnz && (JJ[0] < 0)) SETERRRQ1(PETSC_ERR_ARG_WRONGSTATE,"Row %D starts with negative column index",i,j);
3697:     if (nnz && (JJ[nnz-1] >= B->cmap->N) SETERRRQ3(PETSC_ERR_ARG_WRONGSTATE,"Row %D ends with too large a column index %D (max allowed %D)",i,JJ[nnz-1],B->cmap->N);
3698:   }
3699: #endif

3701:   for (i=0; i<m; i++) {
3702:     nnz     = Ii[i+1]- Ii[i];
3703:     JJ      = J + Ii[i];
3704:     nnz_max = PetscMax(nnz_max,nnz);
3705:     d       = 0;
3706:     for (j=0; j<nnz; j++) {
3707:       if (cstart <= JJ[j] && JJ[j] < cend) d++;
3708:     }
3709:     d_nnz[i] = d;
3710:     o_nnz[i] = nnz - d;
3711:   }
3712:   MatMPIAIJSetPreallocation(B,0,d_nnz,0,o_nnz);
3713:   PetscFree2(d_nnz,o_nnz);

3715:   if (v) values = (PetscScalar*)v;
3716:   else {
3717:     PetscMalloc((nnz_max+1)*sizeof(PetscScalar),&values);
3718:     PetscMemzero(values,nnz_max*sizeof(PetscScalar));
3719:   }

3721:   for (i=0; i<m; i++) {
3722:     ii   = i + rstart;
3723:     nnz  = Ii[i+1]- Ii[i];
3724:     MatSetValues_MPIAIJ(B,1,&ii,nnz,J+Ii[i],values+(v ? Ii[i] : 0),INSERT_VALUES);
3725:   }
3726:   MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);
3727:   MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);

3729:   if (!v) {
3730:     PetscFree(values);
3731:   }
3732:   MatSetOption(B,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE);
3733:   return(0);
3734: }
3735: EXTERN_C_END

3739: /*@
3740:    MatMPIAIJSetPreallocationCSR - Allocates memory for a sparse parallel matrix in AIJ format
3741:    (the default parallel PETSc format).  

3743:    Collective on MPI_Comm

3745:    Input Parameters:
3746: +  B - the matrix 
3747: .  i - the indices into j for the start of each local row (starts with zero)
3748: .  j - the column indices for each local row (starts with zero)
3749: -  v - optional values in the matrix

3751:    Level: developer

3753:    Notes:
3754:        The i, j, and a arrays ARE copied by this routine into the internal format used by PETSc;
3755:      thus you CANNOT change the matrix entries by changing the values of a[] after you have 
3756:      called this routine. Use MatCreateMPIAIJWithSplitArrays() to avoid needing to copy the arrays.

3758:        The i and j indices are 0 based, and i indices are indices corresponding to the local j array.

3760:        The format which is used for the sparse matrix input, is equivalent to a
3761:     row-major ordering.. i.e for the following matrix, the input data expected is
3762:     as shown:

3764:         1 0 0
3765:         2 0 3     P0
3766:        -------
3767:         4 5 6     P1

3769:      Process0 [P0]: rows_owned=[0,1]
3770:         i =  {0,1,3}  [size = nrow+1  = 2+1]
3771:         j =  {0,0,2}  [size = nz = 6]
3772:         v =  {1,2,3}  [size = nz = 6]

3774:      Process1 [P1]: rows_owned=[2]
3775:         i =  {0,3}    [size = nrow+1  = 1+1]
3776:         j =  {0,1,2}  [size = nz = 6]
3777:         v =  {4,5,6}  [size = nz = 6]

3779: .keywords: matrix, aij, compressed row, sparse, parallel

3781: .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatCreateAIJ(), MPIAIJ,
3782:           MatCreateSeqAIJWithArrays(), MatCreateMPIAIJWithSplitArrays()
3783: @*/
3784: PetscErrorCode  MatMPIAIJSetPreallocationCSR(Mat B,const PetscInt i[],const PetscInt j[], const PetscScalar v[])
3785: {

3789:   PetscTryMethod(B,"MatMPIAIJSetPreallocationCSR_C",(Mat,const PetscInt[],const PetscInt[],const PetscScalar[]),(B,i,j,v));
3790:   return(0);
3791: }

3795: /*@C
3796:    MatMPIAIJSetPreallocation - Preallocates memory for a sparse parallel matrix in AIJ format
3797:    (the default parallel PETSc format).  For good matrix assembly performance
3798:    the user should preallocate the matrix storage by setting the parameters 
3799:    d_nz (or d_nnz) and o_nz (or o_nnz).  By setting these parameters accurately,
3800:    performance can be increased by more than a factor of 50.

3802:    Collective on MPI_Comm

3804:    Input Parameters:
3805: +  A - the matrix 
3806: .  d_nz  - number of nonzeros per row in DIAGONAL portion of local submatrix
3807:            (same value is used for all local rows)
3808: .  d_nnz - array containing the number of nonzeros in the various rows of the 
3809:            DIAGONAL portion of the local submatrix (possibly different for each row)
3810:            or PETSC_NULL, if d_nz is used to specify the nonzero structure. 
3811:            The size of this array is equal to the number of local rows, i.e 'm'. 
3812:            For matrices that will be factored, you must leave room for (and set)
3813:            the diagonal entry even if it is zero.
3814: .  o_nz  - number of nonzeros per row in the OFF-DIAGONAL portion of local
3815:            submatrix (same value is used for all local rows).
3816: -  o_nnz - array containing the number of nonzeros in the various rows of the
3817:            OFF-DIAGONAL portion of the local submatrix (possibly different for
3818:            each row) or PETSC_NULL, if o_nz is used to specify the nonzero 
3819:            structure. The size of this array is equal to the number 
3820:            of local rows, i.e 'm'. 

3822:    If the *_nnz parameter is given then the *_nz parameter is ignored

3824:    The AIJ format (also called the Yale sparse matrix format or
3825:    compressed row storage (CSR)), is fully compatible with standard Fortran 77
3826:    storage.  The stored row and column indices begin with zero. 
3827:    See the <A href="../../docs/manual.pdf#nameddest=ch_mat">Mat chapter of the users manual</A> for details.

3829:    The parallel matrix is partitioned such that the first m0 rows belong to 
3830:    process 0, the next m1 rows belong to process 1, the next m2 rows belong 
3831:    to process 2 etc.. where m0,m1,m2... are the input parameter 'm'.

3833:    The DIAGONAL portion of the local submatrix of a processor can be defined
3834:    as the submatrix which is obtained by extraction the part corresponding to
3835:    the rows r1-r2 and columns c1-c2 of the global matrix, where r1 is the
3836:    first row that belongs to the processor, r2 is the last row belonging to
3837:    the this processor, and c1-c2 is range of indices of the local part of a
3838:    vector suitable for applying the matrix to.  This is an mxn matrix.  In the
3839:    common case of a square matrix, the row and column ranges are the same and
3840:    the DIAGONAL part is also square. The remaining portion of the local
3841:    submatrix (mxN) constitute the OFF-DIAGONAL portion.

3843:    If o_nnz, d_nnz are specified, then o_nz, and d_nz are ignored.

3845:    You can call MatGetInfo() to get information on how effective the preallocation was;
3846:    for example the fields mallocs,nz_allocated,nz_used,nz_unneeded;
3847:    You can also run with the option -info and look for messages with the string 
3848:    malloc in them to see if additional memory allocation was needed.

3850:    Example usage:
3851:   
3852:    Consider the following 8x8 matrix with 34 non-zero values, that is 
3853:    assembled across 3 processors. Lets assume that proc0 owns 3 rows,
3854:    proc1 owns 3 rows, proc2 owns 2 rows. This division can be shown 
3855:    as follows:

3857: .vb
3858:             1  2  0  |  0  3  0  |  0  4
3859:     Proc0   0  5  6  |  7  0  0  |  8  0
3860:             9  0 10  | 11  0  0  | 12  0
3861:     -------------------------------------
3862:            13  0 14  | 15 16 17  |  0  0
3863:     Proc1   0 18  0  | 19 20 21  |  0  0 
3864:             0  0  0  | 22 23  0  | 24  0
3865:     -------------------------------------
3866:     Proc2  25 26 27  |  0  0 28  | 29  0
3867:            30  0  0  | 31 32 33  |  0 34
3868: .ve

3870:    This can be represented as a collection of submatrices as:

3872: .vb
3873:       A B C
3874:       D E F
3875:       G H I
3876: .ve

3878:    Where the submatrices A,B,C are owned by proc0, D,E,F are
3879:    owned by proc1, G,H,I are owned by proc2.

3881:    The 'm' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
3882:    The 'n' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
3883:    The 'M','N' parameters are 8,8, and have the same values on all procs.

3885:    The DIAGONAL submatrices corresponding to proc0,proc1,proc2 are
3886:    submatrices [A], [E], [I] respectively. The OFF-DIAGONAL submatrices
3887:    corresponding to proc0,proc1,proc2 are [BC], [DF], [GH] respectively.
3888:    Internally, each processor stores the DIAGONAL part, and the OFF-DIAGONAL
3889:    part as SeqAIJ matrices. for eg: proc1 will store [E] as a SeqAIJ
3890:    matrix, ans [DF] as another SeqAIJ matrix.

3892:    When d_nz, o_nz parameters are specified, d_nz storage elements are
3893:    allocated for every row of the local diagonal submatrix, and o_nz
3894:    storage locations are allocated for every row of the OFF-DIAGONAL submat.
3895:    One way to choose d_nz and o_nz is to use the max nonzerors per local 
3896:    rows for each of the local DIAGONAL, and the OFF-DIAGONAL submatrices. 
3897:    In this case, the values of d_nz,o_nz are:
3898: .vb
3899:      proc0 : dnz = 2, o_nz = 2
3900:      proc1 : dnz = 3, o_nz = 2
3901:      proc2 : dnz = 1, o_nz = 4
3902: .ve
3903:    We are allocating m*(d_nz+o_nz) storage locations for every proc. This
3904:    translates to 3*(2+2)=12 for proc0, 3*(3+2)=15 for proc1, 2*(1+4)=10
3905:    for proc3. i.e we are using 12+15+10=37 storage locations to store 
3906:    34 values.

3908:    When d_nnz, o_nnz parameters are specified, the storage is specified
3909:    for every row, coresponding to both DIAGONAL and OFF-DIAGONAL submatrices.
3910:    In the above case the values for d_nnz,o_nnz are:
3911: .vb
3912:      proc0: d_nnz = [2,2,2] and o_nnz = [2,2,2]
3913:      proc1: d_nnz = [3,3,2] and o_nnz = [2,1,1]
3914:      proc2: d_nnz = [1,1]   and o_nnz = [4,4]
3915: .ve
3916:    Here the space allocated is sum of all the above values i.e 34, and
3917:    hence pre-allocation is perfect.

3919:    Level: intermediate

3921: .keywords: matrix, aij, compressed row, sparse, parallel

3923: .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatCreateAIJ(), MatMPIAIJSetPreallocationCSR(),
3924:           MPIAIJ, MatGetInfo()
3925: @*/
3926: PetscErrorCode  MatMPIAIJSetPreallocation(Mat B,PetscInt d_nz,const PetscInt d_nnz[],PetscInt o_nz,const PetscInt o_nnz[])
3927: {

3933:   PetscTryMethod(B,"MatMPIAIJSetPreallocation_C",(Mat,PetscInt,const PetscInt[],PetscInt,const PetscInt[]),(B,d_nz,d_nnz,o_nz,o_nnz));
3934:   return(0);
3935: }

3939: /*@
3940:      MatCreateMPIAIJWithArrays - creates a MPI AIJ matrix using arrays that contain in standard
3941:          CSR format the local rows.

3943:    Collective on MPI_Comm

3945:    Input Parameters:
3946: +  comm - MPI communicator
3947: .  m - number of local rows (Cannot be PETSC_DECIDE)
3948: .  n - This value should be the same as the local size used in creating the 
3949:        x vector for the matrix-vector product y = Ax. (or PETSC_DECIDE to have
3950:        calculated if N is given) For square matrices n is almost always m.
3951: .  M - number of global rows (or PETSC_DETERMINE to have calculated if m is given)
3952: .  N - number of global columns (or PETSC_DETERMINE to have calculated if n is given)
3953: .   i - row indices
3954: .   j - column indices
3955: -   a - matrix values

3957:    Output Parameter:
3958: .   mat - the matrix

3960:    Level: intermediate

3962:    Notes:
3963:        The i, j, and a arrays ARE copied by this routine into the internal format used by PETSc;
3964:      thus you CANNOT change the matrix entries by changing the values of a[] after you have 
3965:      called this routine. Use MatCreateMPIAIJWithSplitArrays() to avoid needing to copy the arrays.

3967:        The i and j indices are 0 based, and i indices are indices corresponding to the local j array.

3969:        The format which is used for the sparse matrix input, is equivalent to a
3970:     row-major ordering.. i.e for the following matrix, the input data expected is
3971:     as shown:

3973:         1 0 0
3974:         2 0 3     P0
3975:        -------
3976:         4 5 6     P1

3978:      Process0 [P0]: rows_owned=[0,1]
3979:         i =  {0,1,3}  [size = nrow+1  = 2+1]
3980:         j =  {0,0,2}  [size = nz = 6]
3981:         v =  {1,2,3}  [size = nz = 6]

3983:      Process1 [P1]: rows_owned=[2]
3984:         i =  {0,3}    [size = nrow+1  = 1+1]
3985:         j =  {0,1,2}  [size = nz = 6]
3986:         v =  {4,5,6}  [size = nz = 6]

3988: .keywords: matrix, aij, compressed row, sparse, parallel

3990: .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatMPIAIJSetPreallocationCSR(),
3991:           MPIAIJ, MatCreateAIJ(), MatCreateMPIAIJWithSplitArrays()
3992: @*/
3993: PetscErrorCode  MatCreateMPIAIJWithArrays(MPI_Comm comm,PetscInt m,PetscInt n,PetscInt M,PetscInt N,const PetscInt i[],const PetscInt j[],const PetscScalar a[],Mat *mat)
3994: {

3998:   if (i[0]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"i (row indices) must start with 0");
3999:   if (m < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"local number of rows (m) cannot be PETSC_DECIDE, or negative");
4000:   MatCreate(comm,mat);
4001:   MatSetSizes(*mat,m,n,M,N);
4002:   /* MatSetBlockSizes(M,bs,cbs);  */
4003:   MatSetType(*mat,MATMPIAIJ);
4004:   MatMPIAIJSetPreallocationCSR(*mat,i,j,a);
4005:   return(0);
4006: }

4010: /*@C
4011:    MatCreateAIJ - Creates a sparse parallel matrix in AIJ format
4012:    (the default parallel PETSc format).  For good matrix assembly performance
4013:    the user should preallocate the matrix storage by setting the parameters 
4014:    d_nz (or d_nnz) and o_nz (or o_nnz).  By setting these parameters accurately,
4015:    performance can be increased by more than a factor of 50.

4017:    Collective on MPI_Comm

4019:    Input Parameters:
4020: +  comm - MPI communicator
4021: .  m - number of local rows (or PETSC_DECIDE to have calculated if M is given)
4022:            This value should be the same as the local size used in creating the 
4023:            y vector for the matrix-vector product y = Ax.
4024: .  n - This value should be the same as the local size used in creating the 
4025:        x vector for the matrix-vector product y = Ax. (or PETSC_DECIDE to have
4026:        calculated if N is given) For square matrices n is almost always m.
4027: .  M - number of global rows (or PETSC_DETERMINE to have calculated if m is given)
4028: .  N - number of global columns (or PETSC_DETERMINE to have calculated if n is given)
4029: .  d_nz  - number of nonzeros per row in DIAGONAL portion of local submatrix
4030:            (same value is used for all local rows)
4031: .  d_nnz - array containing the number of nonzeros in the various rows of the 
4032:            DIAGONAL portion of the local submatrix (possibly different for each row)
4033:            or PETSC_NULL, if d_nz is used to specify the nonzero structure. 
4034:            The size of this array is equal to the number of local rows, i.e 'm'. 
4035: .  o_nz  - number of nonzeros per row in the OFF-DIAGONAL portion of local
4036:            submatrix (same value is used for all local rows).
4037: -  o_nnz - array containing the number of nonzeros in the various rows of the
4038:            OFF-DIAGONAL portion of the local submatrix (possibly different for
4039:            each row) or PETSC_NULL, if o_nz is used to specify the nonzero 
4040:            structure. The size of this array is equal to the number 
4041:            of local rows, i.e 'm'. 

4043:    Output Parameter:
4044: .  A - the matrix 

4046:    It is recommended that one use the MatCreate(), MatSetType() and/or MatSetFromOptions(),
4047:    MatXXXXSetPreallocation() paradgm instead of this routine directly. 
4048:    [MatXXXXSetPreallocation() is, for example, MatSeqAIJSetPreallocation]

4050:    Notes:
4051:    If the *_nnz parameter is given then the *_nz parameter is ignored

4053:    m,n,M,N parameters specify the size of the matrix, and its partitioning across
4054:    processors, while d_nz,d_nnz,o_nz,o_nnz parameters specify the approximate
4055:    storage requirements for this matrix.

4057:    If PETSC_DECIDE or  PETSC_DETERMINE is used for a particular argument on one 
4058:    processor than it must be used on all processors that share the object for 
4059:    that argument.

4061:    The user MUST specify either the local or global matrix dimensions
4062:    (possibly both).

4064:    The parallel matrix is partitioned across processors such that the
4065:    first m0 rows belong to process 0, the next m1 rows belong to
4066:    process 1, the next m2 rows belong to process 2 etc.. where
4067:    m0,m1,m2,.. are the input parameter 'm'. i.e each processor stores
4068:    values corresponding to [m x N] submatrix.

4070:    The columns are logically partitioned with the n0 columns belonging
4071:    to 0th partition, the next n1 columns belonging to the next
4072:    partition etc.. where n0,n1,n2... are the the input parameter 'n'.

4074:    The DIAGONAL portion of the local submatrix on any given processor
4075:    is the submatrix corresponding to the rows and columns m,n
4076:    corresponding to the given processor. i.e diagonal matrix on
4077:    process 0 is [m0 x n0], diagonal matrix on process 1 is [m1 x n1]
4078:    etc. The remaining portion of the local submatrix [m x (N-n)]
4079:    constitute the OFF-DIAGONAL portion. The example below better
4080:    illustrates this concept.

4082:    For a square global matrix we define each processor's diagonal portion 
4083:    to be its local rows and the corresponding columns (a square submatrix);  
4084:    each processor's off-diagonal portion encompasses the remainder of the
4085:    local matrix (a rectangular submatrix). 

4087:    If o_nnz, d_nnz are specified, then o_nz, and d_nz are ignored.

4089:    When calling this routine with a single process communicator, a matrix of
4090:    type SEQAIJ is returned.  If a matrix of type MPIAIJ is desired for this
4091:    type of communicator, use the construction mechanism:
4092:      MatCreate(...,&A); MatSetType(A,MATMPIAIJ); MatSetSizes(A, m,n,M,N); MatMPIAIJSetPreallocation(A,...);
4093:  
4094:    By default, this format uses inodes (identical nodes) when possible.
4095:    We search for consecutive rows with the same nonzero structure, thereby
4096:    reusing matrix information to achieve increased efficiency.

4098:    Options Database Keys:
4099: +  -mat_no_inode  - Do not use inodes
4100: .  -mat_inode_limit <limit> - Sets inode limit (max limit=5)
4101: -  -mat_aij_oneindex - Internally use indexing starting at 1
4102:         rather than 0.  Note that when calling MatSetValues(),
4103:         the user still MUST index entries starting at 0!


4106:    Example usage:
4107:   
4108:    Consider the following 8x8 matrix with 34 non-zero values, that is 
4109:    assembled across 3 processors. Lets assume that proc0 owns 3 rows,
4110:    proc1 owns 3 rows, proc2 owns 2 rows. This division can be shown 
4111:    as follows:

4113: .vb
4114:             1  2  0  |  0  3  0  |  0  4
4115:     Proc0   0  5  6  |  7  0  0  |  8  0
4116:             9  0 10  | 11  0  0  | 12  0
4117:     -------------------------------------
4118:            13  0 14  | 15 16 17  |  0  0
4119:     Proc1   0 18  0  | 19 20 21  |  0  0 
4120:             0  0  0  | 22 23  0  | 24  0
4121:     -------------------------------------
4122:     Proc2  25 26 27  |  0  0 28  | 29  0
4123:            30  0  0  | 31 32 33  |  0 34
4124: .ve

4126:    This can be represented as a collection of submatrices as:

4128: .vb
4129:       A B C
4130:       D E F
4131:       G H I
4132: .ve

4134:    Where the submatrices A,B,C are owned by proc0, D,E,F are
4135:    owned by proc1, G,H,I are owned by proc2.

4137:    The 'm' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
4138:    The 'n' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
4139:    The 'M','N' parameters are 8,8, and have the same values on all procs.

4141:    The DIAGONAL submatrices corresponding to proc0,proc1,proc2 are
4142:    submatrices [A], [E], [I] respectively. The OFF-DIAGONAL submatrices
4143:    corresponding to proc0,proc1,proc2 are [BC], [DF], [GH] respectively.
4144:    Internally, each processor stores the DIAGONAL part, and the OFF-DIAGONAL
4145:    part as SeqAIJ matrices. for eg: proc1 will store [E] as a SeqAIJ
4146:    matrix, ans [DF] as another SeqAIJ matrix.

4148:    When d_nz, o_nz parameters are specified, d_nz storage elements are
4149:    allocated for every row of the local diagonal submatrix, and o_nz
4150:    storage locations are allocated for every row of the OFF-DIAGONAL submat.
4151:    One way to choose d_nz and o_nz is to use the max nonzerors per local 
4152:    rows for each of the local DIAGONAL, and the OFF-DIAGONAL submatrices. 
4153:    In this case, the values of d_nz,o_nz are:
4154: .vb
4155:      proc0 : dnz = 2, o_nz = 2
4156:      proc1 : dnz = 3, o_nz = 2
4157:      proc2 : dnz = 1, o_nz = 4
4158: .ve
4159:    We are allocating m*(d_nz+o_nz) storage locations for every proc. This
4160:    translates to 3*(2+2)=12 for proc0, 3*(3+2)=15 for proc1, 2*(1+4)=10
4161:    for proc3. i.e we are using 12+15+10=37 storage locations to store 
4162:    34 values.

4164:    When d_nnz, o_nnz parameters are specified, the storage is specified
4165:    for every row, coresponding to both DIAGONAL and OFF-DIAGONAL submatrices.
4166:    In the above case the values for d_nnz,o_nnz are:
4167: .vb
4168:      proc0: d_nnz = [2,2,2] and o_nnz = [2,2,2]
4169:      proc1: d_nnz = [3,3,2] and o_nnz = [2,1,1]
4170:      proc2: d_nnz = [1,1]   and o_nnz = [4,4]
4171: .ve
4172:    Here the space allocated is sum of all the above values i.e 34, and
4173:    hence pre-allocation is perfect.

4175:    Level: intermediate

4177: .keywords: matrix, aij, compressed row, sparse, parallel

4179: .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatMPIAIJSetPreallocationCSR(),
4180:           MPIAIJ, MatCreateMPIAIJWithArrays()
4181: @*/
4182: 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)
4183: {
4185:   PetscMPIInt    size;

4188:   MatCreate(comm,A);
4189:   MatSetSizes(*A,m,n,M,N);
4190:   MPI_Comm_size(comm,&size);
4191:   if (size > 1) {
4192:     MatSetType(*A,MATMPIAIJ);
4193:     MatMPIAIJSetPreallocation(*A,d_nz,d_nnz,o_nz,o_nnz);
4194:   } else {
4195:     MatSetType(*A,MATSEQAIJ);
4196:     MatSeqAIJSetPreallocation(*A,d_nz,d_nnz);
4197:   }
4198:   return(0);
4199: }

4203: PetscErrorCode  MatMPIAIJGetSeqAIJ(Mat A,Mat *Ad,Mat *Ao,PetscInt *colmap[])
4204: {
4205:   Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;

4208:   *Ad     = a->A;
4209:   *Ao     = a->B;
4210:   *colmap = a->garray;
4211:   return(0);
4212: }

4216: PetscErrorCode MatSetColoring_MPIAIJ(Mat A,ISColoring coloring)
4217: {
4219:   PetscInt       i;
4220:   Mat_MPIAIJ     *a = (Mat_MPIAIJ*)A->data;

4223:   if (coloring->ctype == IS_COLORING_GLOBAL) {
4224:     ISColoringValue *allcolors,*colors;
4225:     ISColoring      ocoloring;

4227:     /* set coloring for diagonal portion */
4228:     MatSetColoring_SeqAIJ(a->A,coloring);

4230:     /* set coloring for off-diagonal portion */
4231:     ISAllGatherColors(((PetscObject)A)->comm,coloring->n,coloring->colors,PETSC_NULL,&allcolors);
4232:     PetscMalloc((a->B->cmap->n+1)*sizeof(ISColoringValue),&colors);
4233:     for (i=0; i<a->B->cmap->n; i++) {
4234:       colors[i] = allcolors[a->garray[i]];
4235:     }
4236:     PetscFree(allcolors);
4237:     ISColoringCreate(MPI_COMM_SELF,coloring->n,a->B->cmap->n,colors,&ocoloring);
4238:     MatSetColoring_SeqAIJ(a->B,ocoloring);
4239:     ISColoringDestroy(&ocoloring);
4240:   } else if (coloring->ctype == IS_COLORING_GHOSTED) {
4241:     ISColoringValue *colors;
4242:     PetscInt        *larray;
4243:     ISColoring      ocoloring;

4245:     /* set coloring for diagonal portion */
4246:     PetscMalloc((a->A->cmap->n+1)*sizeof(PetscInt),&larray);
4247:     for (i=0; i<a->A->cmap->n; i++) {
4248:       larray[i] = i + A->cmap->rstart;
4249:     }
4250:     ISGlobalToLocalMappingApply(A->cmap->mapping,IS_GTOLM_MASK,a->A->cmap->n,larray,PETSC_NULL,larray);
4251:     PetscMalloc((a->A->cmap->n+1)*sizeof(ISColoringValue),&colors);
4252:     for (i=0; i<a->A->cmap->n; i++) {
4253:       colors[i] = coloring->colors[larray[i]];
4254:     }
4255:     PetscFree(larray);
4256:     ISColoringCreate(PETSC_COMM_SELF,coloring->n,a->A->cmap->n,colors,&ocoloring);
4257:     MatSetColoring_SeqAIJ(a->A,ocoloring);
4258:     ISColoringDestroy(&ocoloring);

4260:     /* set coloring for off-diagonal portion */
4261:     PetscMalloc((a->B->cmap->n+1)*sizeof(PetscInt),&larray);
4262:     ISGlobalToLocalMappingApply(A->cmap->mapping,IS_GTOLM_MASK,a->B->cmap->n,a->garray,PETSC_NULL,larray);
4263:     PetscMalloc((a->B->cmap->n+1)*sizeof(ISColoringValue),&colors);
4264:     for (i=0; i<a->B->cmap->n; i++) {
4265:       colors[i] = coloring->colors[larray[i]];
4266:     }
4267:     PetscFree(larray);
4268:     ISColoringCreate(MPI_COMM_SELF,coloring->n,a->B->cmap->n,colors,&ocoloring);
4269:     MatSetColoring_SeqAIJ(a->B,ocoloring);
4270:     ISColoringDestroy(&ocoloring);
4271:   } else SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"No support ISColoringType %d",(int)coloring->ctype);

4273:   return(0);
4274: }

4276: #if defined(PETSC_HAVE_ADIC)
4279: PetscErrorCode MatSetValuesAdic_MPIAIJ(Mat A,void *advalues)
4280: {
4281:   Mat_MPIAIJ     *a = (Mat_MPIAIJ*)A->data;

4285:   MatSetValuesAdic_SeqAIJ(a->A,advalues);
4286:   MatSetValuesAdic_SeqAIJ(a->B,advalues);
4287:   return(0);
4288: }
4289: #endif

4293: PetscErrorCode MatSetValuesAdifor_MPIAIJ(Mat A,PetscInt nl,void *advalues)
4294: {
4295:   Mat_MPIAIJ     *a = (Mat_MPIAIJ*)A->data;

4299:   MatSetValuesAdifor_SeqAIJ(a->A,nl,advalues);
4300:   MatSetValuesAdifor_SeqAIJ(a->B,nl,advalues);
4301:   return(0);
4302: }

4306: PetscErrorCode  MatCreateMPIAIJConcatenateSeqAIJSymbolic(MPI_Comm comm,Mat inmat,PetscInt n,Mat *outmat)
4307: {
4309:   PetscInt       m,N,i,rstart,nnz,*dnz,*onz,sum,bs,cbs;
4310:   PetscInt       *indx;

4313:   /* This routine will ONLY return MPIAIJ type matrix */
4314:   MatGetSize(inmat,&m,&N);
4315:   MatGetBlockSizes(inmat,&bs,&cbs);
4316:   if (n == PETSC_DECIDE){
4317:     PetscSplitOwnership(comm,&n,&N);
4318:   }
4319:   /* Check sum(n) = N */
4320:   MPI_Allreduce(&n,&sum,1,MPIU_INT,MPI_SUM,comm);
4321:   if (sum != N) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_INCOMP,"Sum of local columns != global columns %d",N);
4322: 
4323:   MPI_Scan(&m, &rstart,1,MPIU_INT,MPI_SUM,comm);
4324:   rstart -= m;

4326:   MatPreallocateInitialize(comm,m,n,dnz,onz);
4327:   for (i=0;i<m;i++) {
4328:     MatGetRow_SeqAIJ(inmat,i,&nnz,&indx,PETSC_NULL);
4329:     MatPreallocateSet(i+rstart,nnz,indx,dnz,onz);
4330:     MatRestoreRow_SeqAIJ(inmat,i,&nnz,&indx,PETSC_NULL);
4331:   }
4332: 
4333:   MatCreate(comm,outmat);
4334:   MatSetSizes(*outmat,m,n,PETSC_DETERMINE,PETSC_DETERMINE);
4335:   MatSetBlockSizes(*outmat,bs,cbs);
4336:   MatSetType(*outmat,MATMPIAIJ);
4337:   MatMPIAIJSetPreallocation(*outmat,0,dnz,0,onz);
4338:   MatPreallocateFinalize(dnz,onz);
4339:   return(0);
4340: }

4344: PetscErrorCode  MatCreateMPIAIJConcatenateSeqAIJNumeric(MPI_Comm comm,Mat inmat,PetscInt n,Mat outmat)
4345: {
4347:   PetscInt       m,N,i,rstart,nnz,Ii;
4348:   PetscInt       *indx;
4349:   PetscScalar    *values;

4352:   MatGetSize(inmat,&m,&N);
4353:   MatGetOwnershipRange(outmat,&rstart,PETSC_NULL);
4354:   for (i=0;i<m;i++) {
4355:     MatGetRow_SeqAIJ(inmat,i,&nnz,&indx,&values);
4356:     Ii    = i + rstart;
4357:     MatSetValues_MPIAIJ(outmat,1,&Ii,nnz,indx,values,INSERT_VALUES);
4358:     MatRestoreRow_SeqAIJ(inmat,i,&nnz,&indx,&values);
4359:   }
4360:   MatAssemblyBegin(outmat,MAT_FINAL_ASSEMBLY);
4361:   MatAssemblyEnd(outmat,MAT_FINAL_ASSEMBLY);
4362:   return(0);
4363: }

4367: /*@
4368:       MatCreateMPIAIJConcatenateSeqAIJ - Creates a single large PETSc matrix by concatenating sequential
4369:                  matrices from each processor

4371:     Collective on MPI_Comm

4373:    Input Parameters:
4374: +    comm - the communicators the parallel matrix will live on
4375: .    inmat - the input sequential matrices
4376: .    n - number of local columns (or PETSC_DECIDE)
4377: -    scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX

4379:    Output Parameter:
4380: .    outmat - the parallel matrix generated

4382:     Level: advanced

4384:    Notes: The number of columns of the matrix in EACH processor MUST be the same.

4386: @*/
4387: PetscErrorCode  MatCreateMPIAIJConcatenateSeqAIJ(MPI_Comm comm,Mat inmat,PetscInt n,MatReuse scall,Mat *outmat)
4388: {

4392:   PetscLogEventBegin(MAT_Merge,inmat,0,0,0);
4393:   if (scall == MAT_INITIAL_MATRIX){
4394:     MatCreateMPIAIJConcatenateSeqAIJSymbolic(comm,inmat,n,outmat);
4395:   }
4396:   MatCreateMPIAIJConcatenateSeqAIJNumeric(comm,inmat,n,*outmat);
4397:   PetscLogEventEnd(MAT_Merge,inmat,0,0,0);
4398:   return(0);
4399: }

4403: PetscErrorCode MatFileSplit(Mat A,char *outfile)
4404: {
4405:   PetscErrorCode    ierr;
4406:   PetscMPIInt       rank;
4407:   PetscInt          m,N,i,rstart,nnz;
4408:   size_t            len;
4409:   const PetscInt    *indx;
4410:   PetscViewer       out;
4411:   char              *name;
4412:   Mat               B;
4413:   const PetscScalar *values;

4416:   MatGetLocalSize(A,&m,0);
4417:   MatGetSize(A,0,&N);
4418:   /* Should this be the type of the diagonal block of A? */
4419:   MatCreate(PETSC_COMM_SELF,&B);
4420:   MatSetSizes(B,m,N,m,N);
4421:   MatSetBlockSizes(B,A->rmap->bs,A->cmap->bs);
4422:   MatSetType(B,MATSEQAIJ);
4423:   MatSeqAIJSetPreallocation(B,0,PETSC_NULL);
4424:   MatGetOwnershipRange(A,&rstart,0);
4425:   for (i=0;i<m;i++) {
4426:     MatGetRow(A,i+rstart,&nnz,&indx,&values);
4427:     MatSetValues(B,1,&i,nnz,indx,values,INSERT_VALUES);
4428:     MatRestoreRow(A,i+rstart,&nnz,&indx,&values);
4429:   }
4430:   MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);
4431:   MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);

4433:   MPI_Comm_rank(((PetscObject)A)->comm,&rank);
4434:   PetscStrlen(outfile,&len);
4435:   PetscMalloc((len+5)*sizeof(char),&name);
4436:   sprintf(name,"%s.%d",outfile,rank);
4437:   PetscViewerBinaryOpen(PETSC_COMM_SELF,name,FILE_MODE_APPEND,&out);
4438:   PetscFree(name);
4439:   MatView(B,out);
4440:   PetscViewerDestroy(&out);
4441:   MatDestroy(&B);
4442:   return(0);
4443: }

4445: extern PetscErrorCode MatDestroy_MPIAIJ(Mat);
4448: PetscErrorCode  MatDestroy_MPIAIJ_SeqsToMPI(Mat A)
4449: {
4450:   PetscErrorCode       ierr;
4451:   Mat_Merge_SeqsToMPI  *merge;
4452:   PetscContainer       container;

4455:   PetscObjectQuery((PetscObject)A,"MatMergeSeqsToMPI",(PetscObject *)&container);
4456:   if (container) {
4457:     PetscContainerGetPointer(container,(void **)&merge);
4458:     PetscFree(merge->id_r);
4459:     PetscFree(merge->len_s);
4460:     PetscFree(merge->len_r);
4461:     PetscFree(merge->bi);
4462:     PetscFree(merge->bj);
4463:     PetscFree(merge->buf_ri[0]);
4464:     PetscFree(merge->buf_ri);
4465:     PetscFree(merge->buf_rj[0]);
4466:     PetscFree(merge->buf_rj);
4467:     PetscFree(merge->coi);
4468:     PetscFree(merge->coj);
4469:     PetscFree(merge->owners_co);
4470:     PetscLayoutDestroy(&merge->rowmap);
4471:     PetscFree(merge);
4472:     PetscObjectCompose((PetscObject)A,"MatMergeSeqsToMPI",0);
4473:   }
4474:   MatDestroy_MPIAIJ(A);
4475:   return(0);
4476: }

4478: #include <../src/mat/utils/freespace.h>
4479: #include <petscbt.h>

4483: PetscErrorCode  MatCreateMPIAIJSumSeqAIJNumeric(Mat seqmat,Mat mpimat)
4484: {
4485:   PetscErrorCode       ierr;
4486:   MPI_Comm             comm=((PetscObject)mpimat)->comm;
4487:   Mat_SeqAIJ           *a=(Mat_SeqAIJ*)seqmat->data;
4488:   PetscMPIInt          size,rank,taga,*len_s;
4489:   PetscInt             N=mpimat->cmap->N,i,j,*owners,*ai=a->i,*aj=a->j;
4490:   PetscInt             proc,m;
4491:   PetscInt             **buf_ri,**buf_rj;
4492:   PetscInt             k,anzi,*bj_i,*bi,*bj,arow,bnzi,nextaj;
4493:   PetscInt             nrows,**buf_ri_k,**nextrow,**nextai;
4494:   MPI_Request          *s_waits,*r_waits;
4495:   MPI_Status           *status;
4496:   MatScalar            *aa=a->a;
4497:   MatScalar            **abuf_r,*ba_i;
4498:   Mat_Merge_SeqsToMPI  *merge;
4499:   PetscContainer       container;

4502:   PetscLogEventBegin(MAT_Seqstompinum,seqmat,0,0,0);

4504:   MPI_Comm_size(comm,&size);
4505:   MPI_Comm_rank(comm,&rank);

4507:   PetscObjectQuery((PetscObject)mpimat,"MatMergeSeqsToMPI",(PetscObject *)&container);
4508:   PetscContainerGetPointer(container,(void **)&merge);

4510:   bi     = merge->bi;
4511:   bj     = merge->bj;
4512:   buf_ri = merge->buf_ri;
4513:   buf_rj = merge->buf_rj;

4515:   PetscMalloc(size*sizeof(MPI_Status),&status);
4516:   owners = merge->rowmap->range;
4517:   len_s  = merge->len_s;

4519:   /* send and recv matrix values */
4520:   /*-----------------------------*/
4521:   PetscObjectGetNewTag((PetscObject)mpimat,&taga);
4522:   PetscPostIrecvScalar(comm,taga,merge->nrecv,merge->id_r,merge->len_r,&abuf_r,&r_waits);

4524:   PetscMalloc((merge->nsend+1)*sizeof(MPI_Request),&s_waits);
4525:   for (proc=0,k=0; proc<size; proc++){
4526:     if (!len_s[proc]) continue;
4527:     i = owners[proc];
4528:     MPI_Isend(aa+ai[i],len_s[proc],MPIU_MATSCALAR,proc,taga,comm,s_waits+k);
4529:     k++;
4530:   }

4532:   if (merge->nrecv) {MPI_Waitall(merge->nrecv,r_waits,status);}
4533:   if (merge->nsend) {MPI_Waitall(merge->nsend,s_waits,status);}
4534:   PetscFree(status);

4536:   PetscFree(s_waits);
4537:   PetscFree(r_waits);

4539:   /* insert mat values of mpimat */
4540:   /*----------------------------*/
4541:   PetscMalloc(N*sizeof(PetscScalar),&ba_i);
4542:   PetscMalloc3(merge->nrecv,PetscInt*,&buf_ri_k,merge->nrecv,PetscInt*,&nextrow,merge->nrecv,PetscInt*,&nextai);

4544:   for (k=0; k<merge->nrecv; k++){
4545:     buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */
4546:     nrows = *(buf_ri_k[k]);
4547:     nextrow[k]  = buf_ri_k[k]+1;  /* next row number of k-th recved i-structure */
4548:     nextai[k]   = buf_ri_k[k] + (nrows + 1);/* poins to the next i-structure of k-th recved i-structure  */
4549:   }

4551:   /* set values of ba */
4552:   m = merge->rowmap->n;
4553:   for (i=0; i<m; i++) {
4554:     arow = owners[rank] + i;
4555:     bj_i = bj+bi[i];  /* col indices of the i-th row of mpimat */
4556:     bnzi = bi[i+1] - bi[i];
4557:     PetscMemzero(ba_i,bnzi*sizeof(PetscScalar));

4559:     /* add local non-zero vals of this proc's seqmat into ba */
4560:     anzi = ai[arow+1] - ai[arow];
4561:     aj   = a->j + ai[arow];
4562:     aa   = a->a + ai[arow];
4563:     nextaj = 0;
4564:     for (j=0; nextaj<anzi; j++){
4565:       if (*(bj_i + j) == aj[nextaj]){ /* bcol == acol */
4566:         ba_i[j] += aa[nextaj++];
4567:       }
4568:     }

4570:     /* add received vals into ba */
4571:     for (k=0; k<merge->nrecv; k++){ /* k-th received message */
4572:       /* i-th row */
4573:       if (i == *nextrow[k]) {
4574:         anzi = *(nextai[k]+1) - *nextai[k];
4575:         aj   = buf_rj[k] + *(nextai[k]);
4576:         aa   = abuf_r[k] + *(nextai[k]);
4577:         nextaj = 0;
4578:         for (j=0; nextaj<anzi; j++){
4579:           if (*(bj_i + j) == aj[nextaj]){ /* bcol == acol */
4580:             ba_i[j] += aa[nextaj++];
4581:           }
4582:         }
4583:         nextrow[k]++; nextai[k]++;
4584:       }
4585:     }
4586:     MatSetValues(mpimat,1,&arow,bnzi,bj_i,ba_i,INSERT_VALUES);
4587:   }
4588:   MatAssemblyBegin(mpimat,MAT_FINAL_ASSEMBLY);
4589:   MatAssemblyEnd(mpimat,MAT_FINAL_ASSEMBLY);

4591:   PetscFree(abuf_r[0]);
4592:   PetscFree(abuf_r);
4593:   PetscFree(ba_i);
4594:   PetscFree3(buf_ri_k,nextrow,nextai);
4595:   PetscLogEventEnd(MAT_Seqstompinum,seqmat,0,0,0);
4596:   return(0);
4597: }

4599: extern PetscErrorCode  MatDestroy_MPIAIJ_SeqsToMPI(Mat);

4603: PetscErrorCode  MatCreateMPIAIJSumSeqAIJSymbolic(MPI_Comm comm,Mat seqmat,PetscInt m,PetscInt n,Mat *mpimat)
4604: {
4605:   PetscErrorCode       ierr;
4606:   Mat                  B_mpi;
4607:   Mat_SeqAIJ           *a=(Mat_SeqAIJ*)seqmat->data;
4608:   PetscMPIInt          size,rank,tagi,tagj,*len_s,*len_si,*len_ri;
4609:   PetscInt             **buf_rj,**buf_ri,**buf_ri_k;
4610:   PetscInt             M=seqmat->rmap->n,N=seqmat->cmap->n,i,*owners,*ai=a->i,*aj=a->j;
4611:   PetscInt             len,proc,*dnz,*onz,bs,cbs;
4612:   PetscInt             k,anzi,*bi,*bj,*lnk,nlnk,arow,bnzi,nspacedouble=0;
4613:   PetscInt             nrows,*buf_s,*buf_si,*buf_si_i,**nextrow,**nextai;
4614:   MPI_Request          *si_waits,*sj_waits,*ri_waits,*rj_waits;
4615:   MPI_Status           *status;
4616:   PetscFreeSpaceList   free_space=PETSC_NULL,current_space=PETSC_NULL;
4617:   PetscBT              lnkbt;
4618:   Mat_Merge_SeqsToMPI  *merge;
4619:   PetscContainer       container;

4622:   PetscLogEventBegin(MAT_Seqstompisym,seqmat,0,0,0);

4624:   /* make sure it is a PETSc comm */
4625:   PetscCommDuplicate(comm,&comm,PETSC_NULL);
4626:   MPI_Comm_size(comm,&size);
4627:   MPI_Comm_rank(comm,&rank);

4629:   PetscNew(Mat_Merge_SeqsToMPI,&merge);
4630:   PetscMalloc(size*sizeof(MPI_Status),&status);

4632:   /* determine row ownership */
4633:   /*---------------------------------------------------------*/
4634:   PetscLayoutCreate(comm,&merge->rowmap);
4635:   PetscLayoutSetLocalSize(merge->rowmap,m);
4636:   PetscLayoutSetSize(merge->rowmap,M);
4637:   PetscLayoutSetBlockSize(merge->rowmap,1);
4638:   PetscLayoutSetUp(merge->rowmap);
4639:   PetscMalloc(size*sizeof(PetscMPIInt),&len_si);
4640:   PetscMalloc(size*sizeof(PetscMPIInt),&merge->len_s);

4642:   m      = merge->rowmap->n;
4643:   M      = merge->rowmap->N;
4644:   owners = merge->rowmap->range;

4646:   /* determine the number of messages to send, their lengths */
4647:   /*---------------------------------------------------------*/
4648:   len_s  = merge->len_s;

4650:   len = 0;  /* length of buf_si[] */
4651:   merge->nsend = 0;
4652:   for (proc=0; proc<size; proc++){
4653:     len_si[proc] = 0;
4654:     if (proc == rank){
4655:       len_s[proc] = 0;
4656:     } else {
4657:       len_si[proc] = owners[proc+1] - owners[proc] + 1;
4658:       len_s[proc] = ai[owners[proc+1]] - ai[owners[proc]]; /* num of rows to be sent to [proc] */
4659:     }
4660:     if (len_s[proc]) {
4661:       merge->nsend++;
4662:       nrows = 0;
4663:       for (i=owners[proc]; i<owners[proc+1]; i++){
4664:         if (ai[i+1] > ai[i]) nrows++;
4665:       }
4666:       len_si[proc] = 2*(nrows+1);
4667:       len += len_si[proc];
4668:     }
4669:   }

4671:   /* determine the number and length of messages to receive for ij-structure */
4672:   /*-------------------------------------------------------------------------*/
4673:   PetscGatherNumberOfMessages(comm,PETSC_NULL,len_s,&merge->nrecv);
4674:   PetscGatherMessageLengths2(comm,merge->nsend,merge->nrecv,len_s,len_si,&merge->id_r,&merge->len_r,&len_ri);

4676:   /* post the Irecv of j-structure */
4677:   /*-------------------------------*/
4678:   PetscCommGetNewTag(comm,&tagj);
4679:   PetscPostIrecvInt(comm,tagj,merge->nrecv,merge->id_r,merge->len_r,&buf_rj,&rj_waits);

4681:   /* post the Isend of j-structure */
4682:   /*--------------------------------*/
4683:   PetscMalloc2(merge->nsend,MPI_Request,&si_waits,merge->nsend,MPI_Request,&sj_waits);

4685:   for (proc=0, k=0; proc<size; proc++){
4686:     if (!len_s[proc]) continue;
4687:     i = owners[proc];
4688:     MPI_Isend(aj+ai[i],len_s[proc],MPIU_INT,proc,tagj,comm,sj_waits+k);
4689:     k++;
4690:   }

4692:   /* receives and sends of j-structure are complete */
4693:   /*------------------------------------------------*/
4694:   if (merge->nrecv) {MPI_Waitall(merge->nrecv,rj_waits,status);}
4695:   if (merge->nsend) {MPI_Waitall(merge->nsend,sj_waits,status);}

4697:   /* send and recv i-structure */
4698:   /*---------------------------*/
4699:   PetscCommGetNewTag(comm,&tagi);
4700:   PetscPostIrecvInt(comm,tagi,merge->nrecv,merge->id_r,len_ri,&buf_ri,&ri_waits);

4702:   PetscMalloc((len+1)*sizeof(PetscInt),&buf_s);
4703:   buf_si = buf_s;  /* points to the beginning of k-th msg to be sent */
4704:   for (proc=0,k=0; proc<size; proc++){
4705:     if (!len_s[proc]) continue;
4706:     /* form outgoing message for i-structure:
4707:          buf_si[0]:                 nrows to be sent
4708:                [1:nrows]:           row index (global)
4709:                [nrows+1:2*nrows+1]: i-structure index
4710:     */
4711:     /*-------------------------------------------*/
4712:     nrows = len_si[proc]/2 - 1;
4713:     buf_si_i    = buf_si + nrows+1;
4714:     buf_si[0]   = nrows;
4715:     buf_si_i[0] = 0;
4716:     nrows = 0;
4717:     for (i=owners[proc]; i<owners[proc+1]; i++){
4718:       anzi = ai[i+1] - ai[i];
4719:       if (anzi) {
4720:         buf_si_i[nrows+1] = buf_si_i[nrows] + anzi; /* i-structure */
4721:         buf_si[nrows+1] = i-owners[proc]; /* local row index */
4722:         nrows++;
4723:       }
4724:     }
4725:     MPI_Isend(buf_si,len_si[proc],MPIU_INT,proc,tagi,comm,si_waits+k);
4726:     k++;
4727:     buf_si += len_si[proc];
4728:   }

4730:   if (merge->nrecv) {MPI_Waitall(merge->nrecv,ri_waits,status);}
4731:   if (merge->nsend) {MPI_Waitall(merge->nsend,si_waits,status);}

4733:   PetscInfo2(seqmat,"nsend: %D, nrecv: %D\n",merge->nsend,merge->nrecv);
4734:   for (i=0; i<merge->nrecv; i++){
4735:     PetscInfo3(seqmat,"recv len_ri=%D, len_rj=%D from [%D]\n",len_ri[i],merge->len_r[i],merge->id_r[i]);
4736:   }

4738:   PetscFree(len_si);
4739:   PetscFree(len_ri);
4740:   PetscFree(rj_waits);
4741:   PetscFree2(si_waits,sj_waits);
4742:   PetscFree(ri_waits);
4743:   PetscFree(buf_s);
4744:   PetscFree(status);

4746:   /* compute a local seq matrix in each processor */
4747:   /*----------------------------------------------*/
4748:   /* allocate bi array and free space for accumulating nonzero column info */
4749:   PetscMalloc((m+1)*sizeof(PetscInt),&bi);
4750:   bi[0] = 0;

4752:   /* create and initialize a linked list */
4753:   nlnk = N+1;
4754:   PetscLLCreate(N,N,nlnk,lnk,lnkbt);

4756:   /* initial FreeSpace size is 2*(num of local nnz(seqmat)) */
4757:   len = 0;
4758:   len  = ai[owners[rank+1]] - ai[owners[rank]];
4759:   PetscFreeSpaceGet((PetscInt)(2*len+1),&free_space);
4760:   current_space = free_space;

4762:   /* determine symbolic info for each local row */
4763:   PetscMalloc3(merge->nrecv,PetscInt*,&buf_ri_k,merge->nrecv,PetscInt*,&nextrow,merge->nrecv,PetscInt*,&nextai);

4765:   for (k=0; k<merge->nrecv; k++){
4766:     buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */
4767:     nrows = *buf_ri_k[k];
4768:     nextrow[k]  = buf_ri_k[k] + 1;  /* next row number of k-th recved i-structure */
4769:     nextai[k]   = buf_ri_k[k] + (nrows + 1);/* poins to the next i-structure of k-th recved i-structure  */
4770:   }

4772:   MatPreallocateInitialize(comm,m,n,dnz,onz);
4773:   len = 0;
4774:   for (i=0;i<m;i++) {
4775:     bnzi   = 0;
4776:     /* add local non-zero cols of this proc's seqmat into lnk */
4777:     arow   = owners[rank] + i;
4778:     anzi   = ai[arow+1] - ai[arow];
4779:     aj     = a->j + ai[arow];
4780:     PetscLLAddSorted(anzi,aj,N,nlnk,lnk,lnkbt);
4781:     bnzi += nlnk;
4782:     /* add received col data into lnk */
4783:     for (k=0; k<merge->nrecv; k++){ /* k-th received message */
4784:       if (i == *nextrow[k]) { /* i-th row */
4785:         anzi = *(nextai[k]+1) - *nextai[k];
4786:         aj   = buf_rj[k] + *nextai[k];
4787:         PetscLLAddSorted(anzi,aj,N,nlnk,lnk,lnkbt);
4788:         bnzi += nlnk;
4789:         nextrow[k]++; nextai[k]++;
4790:       }
4791:     }
4792:     if (len < bnzi) len = bnzi;  /* =max(bnzi) */

4794:     /* if free space is not available, make more free space */
4795:     if (current_space->local_remaining<bnzi) {
4796:       PetscFreeSpaceGet(bnzi+current_space->total_array_size,&current_space);
4797:       nspacedouble++;
4798:     }
4799:     /* copy data into free space, then initialize lnk */
4800:     PetscLLClean(N,N,bnzi,lnk,current_space->array,lnkbt);
4801:     MatPreallocateSet(i+owners[rank],bnzi,current_space->array,dnz,onz);

4803:     current_space->array           += bnzi;
4804:     current_space->local_used      += bnzi;
4805:     current_space->local_remaining -= bnzi;

4807:     bi[i+1] = bi[i] + bnzi;
4808:   }

4810:   PetscFree3(buf_ri_k,nextrow,nextai);

4812:   PetscMalloc((bi[m]+1)*sizeof(PetscInt),&bj);
4813:   PetscFreeSpaceContiguous(&free_space,bj);
4814:   PetscLLDestroy(lnk,lnkbt);

4816:   /* create symbolic parallel matrix B_mpi */
4817:   /*---------------------------------------*/
4818:     MatGetBlockSizes(seqmat,&bs,&cbs);
4819:   MatCreate(comm,&B_mpi);
4820:   if (n==PETSC_DECIDE) {
4821:     MatSetSizes(B_mpi,m,n,PETSC_DETERMINE,N);
4822:   } else {
4823:     MatSetSizes(B_mpi,m,n,PETSC_DETERMINE,PETSC_DETERMINE);
4824:   }
4825:   MatSetBlockSizes(B_mpi,bs,cbs);
4826:   MatSetType(B_mpi,MATMPIAIJ);
4827:   MatMPIAIJSetPreallocation(B_mpi,0,dnz,0,onz);
4828:   MatPreallocateFinalize(dnz,onz);
4829:   MatSetOption(B_mpi,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);

4831:   /* B_mpi is not ready for use - assembly will be done by MatCreateMPIAIJSumSeqAIJNumeric() */
4832:   B_mpi->assembled     = PETSC_FALSE;
4833:   B_mpi->ops->destroy  = MatDestroy_MPIAIJ_SeqsToMPI;
4834:   merge->bi            = bi;
4835:   merge->bj            = bj;
4836:   merge->buf_ri        = buf_ri;
4837:   merge->buf_rj        = buf_rj;
4838:   merge->coi           = PETSC_NULL;
4839:   merge->coj           = PETSC_NULL;
4840:   merge->owners_co     = PETSC_NULL;

4842:   PetscCommDestroy(&comm);

4844:   /* attach the supporting struct to B_mpi for reuse */
4845:   PetscContainerCreate(PETSC_COMM_SELF,&container);
4846:   PetscContainerSetPointer(container,merge);
4847:   PetscObjectCompose((PetscObject)B_mpi,"MatMergeSeqsToMPI",(PetscObject)container);
4848:   PetscContainerDestroy(&container);
4849:   *mpimat = B_mpi;

4851:   PetscLogEventEnd(MAT_Seqstompisym,seqmat,0,0,0);
4852:   return(0);
4853: }

4857: /*@C
4858:       MatCreateMPIAIJSumSeqAIJ - Creates a MPIAIJ matrix by adding sequential
4859:                  matrices from each processor

4861:     Collective on MPI_Comm

4863:    Input Parameters:
4864: +    comm - the communicators the parallel matrix will live on
4865: .    seqmat - the input sequential matrices
4866: .    m - number of local rows (or PETSC_DECIDE)
4867: .    n - number of local columns (or PETSC_DECIDE)
4868: -    scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX

4870:    Output Parameter:
4871: .    mpimat - the parallel matrix generated

4873:     Level: advanced

4875:    Notes:
4876:      The dimensions of the sequential matrix in each processor MUST be the same.
4877:      The input seqmat is included into the container "Mat_Merge_SeqsToMPI", and will be
4878:      destroyed when mpimat is destroyed. Call PetscObjectQuery() to access seqmat.
4879: @*/
4880: PetscErrorCode  MatCreateMPIAIJSumSeqAIJ(MPI_Comm comm,Mat seqmat,PetscInt m,PetscInt n,MatReuse scall,Mat *mpimat)
4881: {
4882:   PetscErrorCode   ierr;
4883:   PetscMPIInt     size;

4886:   MPI_Comm_size(comm,&size);
4887:   if (size == 1){
4888:      PetscLogEventBegin(MAT_Seqstompi,seqmat,0,0,0);
4889:      if (scall == MAT_INITIAL_MATRIX){
4890:        MatDuplicate(seqmat,MAT_COPY_VALUES,mpimat);
4891:      } else {
4892:        MatCopy(seqmat,*mpimat,SAME_NONZERO_PATTERN);
4893:      }
4894:      PetscLogEventEnd(MAT_Seqstompi,seqmat,0,0,0);
4895:      return(0);
4896:   }
4897:   PetscLogEventBegin(MAT_Seqstompi,seqmat,0,0,0);
4898:   if (scall == MAT_INITIAL_MATRIX){
4899:     MatCreateMPIAIJSumSeqAIJSymbolic(comm,seqmat,m,n,mpimat);
4900:   }
4901:   MatCreateMPIAIJSumSeqAIJNumeric(seqmat,*mpimat);
4902:   PetscLogEventEnd(MAT_Seqstompi,seqmat,0,0,0);
4903:   return(0);
4904: }

4908: /*@
4909:      MatMPIAIJGetLocalMat - Creates a SeqAIJ from a MPIAIJ matrix by taking all its local rows and putting them into a sequential vector with
4910:           mlocal rows and n columns. Where mlocal is the row count obtained with MatGetLocalSize() and n is the global column count obtained
4911:           with MatGetSize()

4913:     Not Collective

4915:    Input Parameters:
4916: +    A - the matrix 
4917: .    scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX 

4919:    Output Parameter:
4920: .    A_loc - the local sequential matrix generated

4922:     Level: developer

4924: .seealso: MatGetOwnerShipRange(), MatMPIAIJGetLocalMatCondensed()

4926: @*/
4927: PetscErrorCode  MatMPIAIJGetLocalMat(Mat A,MatReuse scall,Mat *A_loc)
4928: {
4929:   PetscErrorCode  ierr;
4930:   Mat_MPIAIJ      *mpimat=(Mat_MPIAIJ*)A->data;
4931:   Mat_SeqAIJ      *mat,*a=(Mat_SeqAIJ*)(mpimat->A)->data,*b=(Mat_SeqAIJ*)(mpimat->B)->data;
4932:   PetscInt        *ai=a->i,*aj=a->j,*bi=b->i,*bj=b->j,*cmap=mpimat->garray;
4933:   MatScalar       *aa=a->a,*ba=b->a,*cam;
4934:   PetscScalar     *ca;
4935:   PetscInt        am=A->rmap->n,i,j,k,cstart=A->cmap->rstart;
4936:   PetscInt        *ci,*cj,col,ncols_d,ncols_o,jo;
4937:   PetscBool       match;

4940:   PetscObjectTypeCompare((PetscObject)A,MATMPIAIJ,&match);
4941:   if (!match) SETERRQ(((PetscObject)A)->comm, PETSC_ERR_SUP,"Requires MPIAIJ matrix as input");
4942:   PetscLogEventBegin(MAT_Getlocalmat,A,0,0,0);
4943:   if (scall == MAT_INITIAL_MATRIX){
4944:     PetscMalloc((1+am)*sizeof(PetscInt),&ci);
4945:     ci[0] = 0;
4946:     for (i=0; i<am; i++){
4947:       ci[i+1] = ci[i] + (ai[i+1] - ai[i]) + (bi[i+1] - bi[i]);
4948:     }
4949:     PetscMalloc((1+ci[am])*sizeof(PetscInt),&cj);
4950:     PetscMalloc((1+ci[am])*sizeof(PetscScalar),&ca);
4951:     k = 0;
4952:     for (i=0; i<am; i++) {
4953:       ncols_o = bi[i+1] - bi[i];
4954:       ncols_d = ai[i+1] - ai[i];
4955:       /* off-diagonal portion of A */
4956:       for (jo=0; jo<ncols_o; jo++) {
4957:         col = cmap[*bj];
4958:         if (col >= cstart) break;
4959:         cj[k]   = col; bj++;
4960:         ca[k++] = *ba++;
4961:       }
4962:       /* diagonal portion of A */
4963:       for (j=0; j<ncols_d; j++) {
4964:         cj[k]   = cstart + *aj++;
4965:         ca[k++] = *aa++;
4966:       }
4967:       /* off-diagonal portion of A */
4968:       for (j=jo; j<ncols_o; j++) {
4969:         cj[k]   = cmap[*bj++];
4970:         ca[k++] = *ba++;
4971:       }
4972:     }
4973:     /* put together the new matrix */
4974:     MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,am,A->cmap->N,ci,cj,ca,A_loc);
4975:     /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */
4976:     /* Since these are PETSc arrays, change flags to free them as necessary. */
4977:     mat          = (Mat_SeqAIJ*)(*A_loc)->data;
4978:     mat->free_a  = PETSC_TRUE;
4979:     mat->free_ij = PETSC_TRUE;
4980:     mat->nonew   = 0;
4981:   } else if (scall == MAT_REUSE_MATRIX){
4982:     mat=(Mat_SeqAIJ*)(*A_loc)->data;
4983:     ci = mat->i; cj = mat->j; cam = mat->a;
4984:     for (i=0; i<am; i++) {
4985:       /* off-diagonal portion of A */
4986:       ncols_o = bi[i+1] - bi[i];
4987:       for (jo=0; jo<ncols_o; jo++) {
4988:         col = cmap[*bj];
4989:         if (col >= cstart) break;
4990:         *cam++ = *ba++; bj++;
4991:       }
4992:       /* diagonal portion of A */
4993:       ncols_d = ai[i+1] - ai[i];
4994:       for (j=0; j<ncols_d; j++) *cam++ = *aa++;
4995:       /* off-diagonal portion of A */
4996:       for (j=jo; j<ncols_o; j++) {
4997:         *cam++ = *ba++; bj++;
4998:       }
4999:     }
5000:   } else SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Invalid MatReuse %d",(int)scall);
5001:   PetscLogEventEnd(MAT_Getlocalmat,A,0,0,0);
5002:   return(0);
5003: }

5007: /*@C
5008:      MatMPIAIJGetLocalMatCondensed - Creates a SeqAIJ matrix from an MPIAIJ matrix by taking all its local rows and NON-ZERO columns

5010:     Not Collective

5012:    Input Parameters:
5013: +    A - the matrix 
5014: .    scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX
5015: -    row, col - index sets of rows and columns to extract (or PETSC_NULL)  

5017:    Output Parameter:
5018: .    A_loc - the local sequential matrix generated

5020:     Level: developer

5022: .seealso: MatGetOwnershipRange(), MatMPIAIJGetLocalMat()

5024: @*/
5025: PetscErrorCode  MatMPIAIJGetLocalMatCondensed(Mat A,MatReuse scall,IS *row,IS *col,Mat *A_loc)
5026: {
5027:   Mat_MPIAIJ        *a=(Mat_MPIAIJ*)A->data;
5028:   PetscErrorCode    ierr;
5029:   PetscInt          i,start,end,ncols,nzA,nzB,*cmap,imark,*idx;
5030:   IS                isrowa,iscola;
5031:   Mat               *aloc;
5032:   PetscBool       match;

5035:   PetscObjectTypeCompare((PetscObject)A,MATMPIAIJ,&match);
5036:   if (!match) SETERRQ(((PetscObject)A)->comm, PETSC_ERR_SUP,"Requires MPIAIJ matrix as input");
5037:   PetscLogEventBegin(MAT_Getlocalmatcondensed,A,0,0,0);
5038:   if (!row){
5039:     start = A->rmap->rstart; end = A->rmap->rend;
5040:     ISCreateStride(PETSC_COMM_SELF,end-start,start,1,&isrowa);
5041:   } else {
5042:     isrowa = *row;
5043:   }
5044:   if (!col){
5045:     start = A->cmap->rstart;
5046:     cmap  = a->garray;
5047:     nzA   = a->A->cmap->n;
5048:     nzB   = a->B->cmap->n;
5049:     PetscMalloc((nzA+nzB)*sizeof(PetscInt), &idx);
5050:     ncols = 0;
5051:     for (i=0; i<nzB; i++) {
5052:       if (cmap[i] < start) idx[ncols++] = cmap[i];
5053:       else break;
5054:     }
5055:     imark = i;
5056:     for (i=0; i<nzA; i++) idx[ncols++] = start + i;
5057:     for (i=imark; i<nzB; i++) idx[ncols++] = cmap[i];
5058:     ISCreateGeneral(PETSC_COMM_SELF,ncols,idx,PETSC_OWN_POINTER,&iscola);
5059:   } else {
5060:     iscola = *col;
5061:   }
5062:   if (scall != MAT_INITIAL_MATRIX){
5063:     PetscMalloc(sizeof(Mat),&aloc);
5064:     aloc[0] = *A_loc;
5065:   }
5066:   MatGetSubMatrices(A,1,&isrowa,&iscola,scall,&aloc);
5067:   *A_loc = aloc[0];
5068:   PetscFree(aloc);
5069:   if (!row){
5070:     ISDestroy(&isrowa);
5071:   }
5072:   if (!col){
5073:     ISDestroy(&iscola);
5074:   }
5075:   PetscLogEventEnd(MAT_Getlocalmatcondensed,A,0,0,0);
5076:   return(0);
5077: }

5081: /*@C
5082:     MatGetBrowsOfAcols - Creates a SeqAIJ matrix by taking rows of B that equal to nonzero columns of local A 

5084:     Collective on Mat

5086:    Input Parameters:
5087: +    A,B - the matrices in mpiaij format
5088: .    scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX
5089: -    rowb, colb - index sets of rows and columns of B to extract (or PETSC_NULL)   

5091:    Output Parameter:
5092: +    rowb, colb - index sets of rows and columns of B to extract 
5093: -    B_seq - the sequential matrix generated

5095:     Level: developer

5097: @*/
5098: PetscErrorCode  MatGetBrowsOfAcols(Mat A,Mat B,MatReuse scall,IS *rowb,IS *colb,Mat *B_seq)
5099: {
5100:   Mat_MPIAIJ        *a=(Mat_MPIAIJ*)A->data;
5101:   PetscErrorCode    ierr;
5102:   PetscInt          *idx,i,start,ncols,nzA,nzB,*cmap,imark;
5103:   IS                isrowb,iscolb;
5104:   Mat               *bseq=PETSC_NULL;
5105: 
5107:   if (A->cmap->rstart != B->rmap->rstart || A->cmap->rend != B->rmap->rend){
5108:     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);
5109:   }
5110:   PetscLogEventBegin(MAT_GetBrowsOfAcols,A,B,0,0);
5111: 
5112:   if (scall == MAT_INITIAL_MATRIX){
5113:     start = A->cmap->rstart;
5114:     cmap  = a->garray;
5115:     nzA   = a->A->cmap->n;
5116:     nzB   = a->B->cmap->n;
5117:     PetscMalloc((nzA+nzB)*sizeof(PetscInt), &idx);
5118:     ncols = 0;
5119:     for (i=0; i<nzB; i++) {  /* row < local row index */
5120:       if (cmap[i] < start) idx[ncols++] = cmap[i];
5121:       else break;
5122:     }
5123:     imark = i;
5124:     for (i=0; i<nzA; i++) idx[ncols++] = start + i;  /* local rows */
5125:     for (i=imark; i<nzB; i++) idx[ncols++] = cmap[i]; /* row > local row index */
5126:     ISCreateGeneral(PETSC_COMM_SELF,ncols,idx,PETSC_OWN_POINTER,&isrowb);
5127:     ISCreateStride(PETSC_COMM_SELF,B->cmap->N,0,1,&iscolb);
5128:   } else {
5129:     if (!rowb || !colb) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"IS rowb and colb must be provided for MAT_REUSE_MATRIX");
5130:     isrowb = *rowb; iscolb = *colb;
5131:     PetscMalloc(sizeof(Mat),&bseq);
5132:     bseq[0] = *B_seq;
5133:   }
5134:   MatGetSubMatrices(B,1,&isrowb,&iscolb,scall,&bseq);
5135:   *B_seq = bseq[0];
5136:   PetscFree(bseq);
5137:   if (!rowb){
5138:     ISDestroy(&isrowb);
5139:   } else {
5140:     *rowb = isrowb;
5141:   }
5142:   if (!colb){
5143:     ISDestroy(&iscolb);
5144:   } else {
5145:     *colb = iscolb;
5146:   }
5147:   PetscLogEventEnd(MAT_GetBrowsOfAcols,A,B,0,0);
5148:   return(0);
5149: }

5153: /*
5154:     MatGetBrowsOfAoCols_MPIAIJ - Creates a SeqAIJ matrix by taking rows of B that equal to nonzero columns 
5155:     of the OFF-DIAGONAL portion of local A

5157:     Collective on Mat

5159:    Input Parameters:
5160: +    A,B - the matrices in mpiaij format
5161: -    scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX

5163:    Output Parameter:
5164: +    startsj_s - starting point in B's sending j-arrays, saved for MAT_REUSE (or PETSC_NULL) 
5165: .    startsj_r - starting point in B's receiving j-arrays, saved for MAT_REUSE (or PETSC_NULL)
5166: .    bufa_ptr - array for sending matrix values, saved for MAT_REUSE (or PETSC_NULL) 
5167: -    B_oth - the sequential matrix generated with size aBn=a->B->cmap->n by B->cmap->N

5169:     Level: developer

5171: */
5172: PetscErrorCode  MatGetBrowsOfAoCols_MPIAIJ(Mat A,Mat B,MatReuse scall,PetscInt **startsj_s,PetscInt **startsj_r,MatScalar **bufa_ptr,Mat *B_oth)
5173: {
5174:   VecScatter_MPI_General *gen_to,*gen_from;
5175:   PetscErrorCode         ierr;
5176:   Mat_MPIAIJ             *a=(Mat_MPIAIJ*)A->data;
5177:   Mat_SeqAIJ             *b_oth;
5178:   VecScatter             ctx=a->Mvctx;
5179:   MPI_Comm               comm=((PetscObject)ctx)->comm;
5180:   PetscMPIInt            *rprocs,*sprocs,tag=((PetscObject)ctx)->tag,rank;
5181:   PetscInt               *rowlen,*bufj,*bufJ,ncols,aBn=a->B->cmap->n,row,*b_othi,*b_othj;
5182:   PetscScalar            *rvalues,*svalues;
5183:   MatScalar              *b_otha,*bufa,*bufA;
5184:   PetscInt               i,j,k,l,ll,nrecvs,nsends,nrows,*srow,*rstarts,*rstartsj = 0,*sstarts,*sstartsj,len;
5185:   MPI_Request            *rwaits = PETSC_NULL,*swaits = PETSC_NULL;
5186:   MPI_Status             *sstatus,rstatus;
5187:   PetscMPIInt            jj;
5188:   PetscInt               *cols,sbs,rbs;
5189:   PetscScalar            *vals;

5192:   if (A->cmap->rstart != B->rmap->rstart || A->cmap->rend != B->rmap->rend){
5193:     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);
5194:   }
5195:   PetscLogEventBegin(MAT_GetBrowsOfAocols,A,B,0,0);
5196:   MPI_Comm_rank(comm,&rank);

5198:   gen_to   = (VecScatter_MPI_General*)ctx->todata;
5199:   gen_from = (VecScatter_MPI_General*)ctx->fromdata;
5200:   rvalues  = gen_from->values; /* holds the length of receiving row */
5201:   svalues  = gen_to->values;   /* holds the length of sending row */
5202:   nrecvs   = gen_from->n;
5203:   nsends   = gen_to->n;

5205:   PetscMalloc2(nrecvs,MPI_Request,&rwaits,nsends,MPI_Request,&swaits);
5206:   srow     = gen_to->indices;   /* local row index to be sent */
5207:   sstarts  = gen_to->starts;
5208:   sprocs   = gen_to->procs;
5209:   sstatus  = gen_to->sstatus;
5210:   sbs      = gen_to->bs;
5211:   rstarts  = gen_from->starts;
5212:   rprocs   = gen_from->procs;
5213:   rbs      = gen_from->bs;

5215:   if (!startsj_s || !bufa_ptr) scall = MAT_INITIAL_MATRIX;
5216:   if (scall == MAT_INITIAL_MATRIX){
5217:     /* i-array */
5218:     /*---------*/
5219:     /*  post receives */
5220:     for (i=0; i<nrecvs; i++){
5221:       rowlen = (PetscInt*)rvalues + rstarts[i]*rbs;
5222:       nrows = (rstarts[i+1]-rstarts[i])*rbs; /* num of indices to be received */
5223:       MPI_Irecv(rowlen,nrows,MPIU_INT,rprocs[i],tag,comm,rwaits+i);
5224:     }

5226:     /* pack the outgoing message */
5227:     PetscMalloc2(nsends+1,PetscInt,&sstartsj,nrecvs+1,PetscInt,&rstartsj);
5228:     sstartsj[0] = 0;  rstartsj[0] = 0;
5229:     len = 0; /* total length of j or a array to be sent */
5230:     k = 0;
5231:     for (i=0; i<nsends; i++){
5232:       rowlen = (PetscInt*)svalues + sstarts[i]*sbs;
5233:       nrows = sstarts[i+1]-sstarts[i]; /* num of block rows */
5234:       for (j=0; j<nrows; j++) {
5235:         row = srow[k] + B->rmap->range[rank]; /* global row idx */
5236:         for (l=0; l<sbs; l++){
5237:           MatGetRow_MPIAIJ(B,row+l,&ncols,PETSC_NULL,PETSC_NULL); /* rowlength */
5238:           rowlen[j*sbs+l] = ncols;
5239:           len += ncols;
5240:           MatRestoreRow_MPIAIJ(B,row+l,&ncols,PETSC_NULL,PETSC_NULL);
5241:         }
5242:         k++;
5243:       }
5244:       MPI_Isend(rowlen,nrows*sbs,MPIU_INT,sprocs[i],tag,comm,swaits+i);
5245:       sstartsj[i+1] = len;  /* starting point of (i+1)-th outgoing msg in bufj and bufa */
5246:     }
5247:     /* recvs and sends of i-array are completed */
5248:     i = nrecvs;
5249:     while (i--) {
5250:       MPI_Waitany(nrecvs,rwaits,&jj,&rstatus);
5251:     }
5252:     if (nsends) {MPI_Waitall(nsends,swaits,sstatus);}

5254:     /* allocate buffers for sending j and a arrays */
5255:     PetscMalloc((len+1)*sizeof(PetscInt),&bufj);
5256:     PetscMalloc((len+1)*sizeof(PetscScalar),&bufa);

5258:     /* create i-array of B_oth */
5259:     PetscMalloc((aBn+2)*sizeof(PetscInt),&b_othi);
5260:     b_othi[0] = 0;
5261:     len = 0; /* total length of j or a array to be received */
5262:     k = 0;
5263:     for (i=0; i<nrecvs; i++){
5264:       rowlen = (PetscInt*)rvalues + rstarts[i]*rbs;
5265:       nrows = rbs*(rstarts[i+1]-rstarts[i]); /* num of rows to be recieved */
5266:       for (j=0; j<nrows; j++) {
5267:         b_othi[k+1] = b_othi[k] + rowlen[j];
5268:         len += rowlen[j]; k++;
5269:       }
5270:       rstartsj[i+1] = len; /* starting point of (i+1)-th incoming msg in bufj and bufa */
5271:     }

5273:     /* allocate space for j and a arrrays of B_oth */
5274:     PetscMalloc((b_othi[aBn]+1)*sizeof(PetscInt),&b_othj);
5275:     PetscMalloc((b_othi[aBn]+1)*sizeof(MatScalar),&b_otha);

5277:     /* j-array */
5278:     /*---------*/
5279:     /*  post receives of j-array */
5280:     for (i=0; i<nrecvs; i++){
5281:       nrows = rstartsj[i+1]-rstartsj[i]; /* length of the msg received */
5282:       MPI_Irecv(b_othj+rstartsj[i],nrows,MPIU_INT,rprocs[i],tag,comm,rwaits+i);
5283:     }

5285:     /* pack the outgoing message j-array */
5286:     k = 0;
5287:     for (i=0; i<nsends; i++){
5288:       nrows = sstarts[i+1]-sstarts[i]; /* num of block rows */
5289:       bufJ = bufj+sstartsj[i];
5290:       for (j=0; j<nrows; j++) {
5291:         row  = srow[k++] + B->rmap->range[rank]; /* global row idx */
5292:         for (ll=0; ll<sbs; ll++){
5293:           MatGetRow_MPIAIJ(B,row+ll,&ncols,&cols,PETSC_NULL);
5294:           for (l=0; l<ncols; l++){
5295:             *bufJ++ = cols[l];
5296:           }
5297:           MatRestoreRow_MPIAIJ(B,row+ll,&ncols,&cols,PETSC_NULL);
5298:         }
5299:       }
5300:       MPI_Isend(bufj+sstartsj[i],sstartsj[i+1]-sstartsj[i],MPIU_INT,sprocs[i],tag,comm,swaits+i);
5301:     }

5303:     /* recvs and sends of j-array are completed */
5304:     i = nrecvs;
5305:     while (i--) {
5306:       MPI_Waitany(nrecvs,rwaits,&jj,&rstatus);
5307:     }
5308:     if (nsends) {MPI_Waitall(nsends,swaits,sstatus);}
5309:   } else if (scall == MAT_REUSE_MATRIX){
5310:     sstartsj = *startsj_s;
5311:     rstartsj = *startsj_r;
5312:     bufa     = *bufa_ptr;
5313:     b_oth    = (Mat_SeqAIJ*)(*B_oth)->data;
5314:     b_otha   = b_oth->a;
5315:   } else {
5316:     SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE, "Matrix P does not posses an object container");
5317:   }

5319:   /* a-array */
5320:   /*---------*/
5321:   /*  post receives of a-array */
5322:   for (i=0; i<nrecvs; i++){
5323:     nrows = rstartsj[i+1]-rstartsj[i]; /* length of the msg received */
5324:     MPI_Irecv(b_otha+rstartsj[i],nrows,MPIU_SCALAR,rprocs[i],tag,comm,rwaits+i);
5325:   }

5327:   /* pack the outgoing message a-array */
5328:   k = 0;
5329:   for (i=0; i<nsends; i++){
5330:     nrows = sstarts[i+1]-sstarts[i]; /* num of block rows */
5331:     bufA = bufa+sstartsj[i];
5332:     for (j=0; j<nrows; j++) {
5333:       row  = srow[k++] + B->rmap->range[rank]; /* global row idx */
5334:       for (ll=0; ll<sbs; ll++){
5335:         MatGetRow_MPIAIJ(B,row+ll,&ncols,PETSC_NULL,&vals);
5336:         for (l=0; l<ncols; l++){
5337:           *bufA++ = vals[l];
5338:         }
5339:         MatRestoreRow_MPIAIJ(B,row+ll,&ncols,PETSC_NULL,&vals);
5340:       }
5341:     }
5342:     MPI_Isend(bufa+sstartsj[i],sstartsj[i+1]-sstartsj[i],MPIU_SCALAR,sprocs[i],tag,comm,swaits+i);
5343:   }
5344:   /* recvs and sends of a-array are completed */
5345:   i = nrecvs;
5346:   while (i--) {
5347:     MPI_Waitany(nrecvs,rwaits,&jj,&rstatus);
5348:   }
5349:   if (nsends) {MPI_Waitall(nsends,swaits,sstatus);}
5350:   PetscFree2(rwaits,swaits);

5352:   if (scall == MAT_INITIAL_MATRIX){
5353:     /* put together the new matrix */
5354:     MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,aBn,B->cmap->N,b_othi,b_othj,b_otha,B_oth);

5356:     /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */
5357:     /* Since these are PETSc arrays, change flags to free them as necessary. */
5358:     b_oth          = (Mat_SeqAIJ *)(*B_oth)->data;
5359:     b_oth->free_a  = PETSC_TRUE;
5360:     b_oth->free_ij = PETSC_TRUE;
5361:     b_oth->nonew   = 0;

5363:     PetscFree(bufj);
5364:     if (!startsj_s || !bufa_ptr){
5365:       PetscFree2(sstartsj,rstartsj);
5366:       PetscFree(bufa_ptr);
5367:     } else {
5368:       *startsj_s = sstartsj;
5369:       *startsj_r = rstartsj;
5370:       *bufa_ptr  = bufa;
5371:     }
5372:   }
5373:   PetscLogEventEnd(MAT_GetBrowsOfAocols,A,B,0,0);
5374:   return(0);
5375: }

5379: /*@C
5380:   MatGetCommunicationStructs - Provides access to the communication structures used in matrix-vector multiplication.

5382:   Not Collective

5384:   Input Parameters:
5385: . A - The matrix in mpiaij format

5387:   Output Parameter:
5388: + lvec - The local vector holding off-process values from the argument to a matrix-vector product
5389: . colmap - A map from global column index to local index into lvec
5390: - multScatter - A scatter from the argument of a matrix-vector product to lvec

5392:   Level: developer

5394: @*/
5395: #if defined (PETSC_USE_CTABLE)
5396: PetscErrorCode  MatGetCommunicationStructs(Mat A, Vec *lvec, PetscTable *colmap, VecScatter *multScatter)
5397: #else
5398: PetscErrorCode  MatGetCommunicationStructs(Mat A, Vec *lvec, PetscInt *colmap[], VecScatter *multScatter)
5399: #endif
5400: {
5401:   Mat_MPIAIJ *a;

5408:   a = (Mat_MPIAIJ *) A->data;
5409:   if (lvec) *lvec = a->lvec;
5410:   if (colmap) *colmap = a->colmap;
5411:   if (multScatter) *multScatter = a->Mvctx;
5412:   return(0);
5413: }

5415: EXTERN_C_BEGIN
5416: extern PetscErrorCode  MatConvert_MPIAIJ_MPIAIJCRL(Mat,const MatType,MatReuse,Mat*);
5417: extern PetscErrorCode  MatConvert_MPIAIJ_MPIAIJPERM(Mat,const MatType,MatReuse,Mat*);
5418: extern PetscErrorCode  MatConvert_MPIAIJ_MPISBAIJ(Mat,const MatType,MatReuse,Mat*);
5419: EXTERN_C_END

5423: /*
5424:     Computes (B'*A')' since computing B*A directly is untenable

5426:                n                       p                          p
5427:         (              )       (              )         (                  )
5428:       m (      A       )  *  n (       B      )   =   m (         C        )
5429:         (              )       (              )         (                  )

5431: */
5432: PetscErrorCode MatMatMultNumeric_MPIDense_MPIAIJ(Mat A,Mat B,Mat C)
5433: {
5434:   PetscErrorCode     ierr;
5435:   Mat                At,Bt,Ct;

5438:   MatTranspose(A,MAT_INITIAL_MATRIX,&At);
5439:   MatTranspose(B,MAT_INITIAL_MATRIX,&Bt);
5440:   MatMatMult(Bt,At,MAT_INITIAL_MATRIX,1.0,&Ct);
5441:   MatDestroy(&At);
5442:   MatDestroy(&Bt);
5443:   MatTranspose(Ct,MAT_REUSE_MATRIX,&C);
5444:   MatDestroy(&Ct);
5445:   return(0);
5446: }

5450: PetscErrorCode MatMatMultSymbolic_MPIDense_MPIAIJ(Mat A,Mat B,PetscReal fill,Mat *C)
5451: {
5453:   PetscInt       m=A->rmap->n,n=B->cmap->n;
5454:   Mat            Cmat;

5457:   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);
5458:   MatCreate(((PetscObject)A)->comm,&Cmat);
5459:   MatSetSizes(Cmat,m,n,PETSC_DETERMINE,PETSC_DETERMINE);
5460:   MatSetBlockSizes(Cmat,A->rmap->bs,B->cmap->bs);
5461:   MatSetType(Cmat,MATMPIDENSE);
5462:   MatMPIDenseSetPreallocation(Cmat,PETSC_NULL);
5463:   MatAssemblyBegin(Cmat,MAT_FINAL_ASSEMBLY);
5464:   MatAssemblyEnd(Cmat,MAT_FINAL_ASSEMBLY);
5465:   *C   = Cmat;
5466:   (*C)->ops->matmult = MatMatMult_MPIDense_MPIAIJ;
5467:   return(0);
5468: }

5470: /* ----------------------------------------------------------------*/
5473: PetscErrorCode MatMatMult_MPIDense_MPIAIJ(Mat A,Mat B,MatReuse scall,PetscReal fill,Mat *C)
5474: {

5478:   if (scall == MAT_INITIAL_MATRIX){
5479:     MatMatMultSymbolic_MPIDense_MPIAIJ(A,B,fill,C);
5480:   }
5481:   MatMatMultNumeric_MPIDense_MPIAIJ(A,B,*C);
5482:   return(0);
5483: }

5485: EXTERN_C_BEGIN
5486: #if defined(PETSC_HAVE_MUMPS)
5487: extern PetscErrorCode MatGetFactor_aij_mumps(Mat,MatFactorType,Mat*);
5488: #endif
5489: #if defined(PETSC_HAVE_PASTIX)
5490: extern PetscErrorCode MatGetFactor_mpiaij_pastix(Mat,MatFactorType,Mat*);
5491: #endif
5492: #if defined(PETSC_HAVE_SUPERLU_DIST)
5493: extern PetscErrorCode MatGetFactor_mpiaij_superlu_dist(Mat,MatFactorType,Mat*);
5494: #endif
5495: #if defined(PETSC_HAVE_SPOOLES)
5496: extern PetscErrorCode MatGetFactor_mpiaij_spooles(Mat,MatFactorType,Mat*);
5497: #endif
5498: EXTERN_C_END

5500: /*MC
5501:    MATMPIAIJ - MATMPIAIJ = "mpiaij" - A matrix type to be used for parallel sparse matrices.

5503:    Options Database Keys:
5504: . -mat_type mpiaij - sets the matrix type to "mpiaij" during a call to MatSetFromOptions()

5506:   Level: beginner

5508: .seealso: MatCreateAIJ()
5509: M*/

5511: EXTERN_C_BEGIN
5514: PetscErrorCode  MatCreate_MPIAIJ(Mat B)
5515: {
5516:   Mat_MPIAIJ     *b;
5518:   PetscMPIInt    size;

5521:   MPI_Comm_size(((PetscObject)B)->comm,&size);

5523:   PetscNewLog(B,Mat_MPIAIJ,&b);
5524:   B->data         = (void*)b;
5525:   PetscMemcpy(B->ops,&MatOps_Values,sizeof(struct _MatOps));
5526:   B->assembled    = PETSC_FALSE;

5528:   B->insertmode   = NOT_SET_VALUES;
5529:   b->size         = size;
5530:   MPI_Comm_rank(((PetscObject)B)->comm,&b->rank);

5532:   /* build cache for off array entries formed */
5533:   MatStashCreate_Private(((PetscObject)B)->comm,1,&B->stash);
5534:   b->donotstash  = PETSC_FALSE;
5535:   b->colmap      = 0;
5536:   b->garray      = 0;
5537:   b->roworiented = PETSC_TRUE;

5539:   /* stuff used for matrix vector multiply */
5540:   b->lvec      = PETSC_NULL;
5541:   b->Mvctx     = PETSC_NULL;

5543:   /* stuff for MatGetRow() */
5544:   b->rowindices   = 0;
5545:   b->rowvalues    = 0;
5546:   b->getrowactive = PETSC_FALSE;

5548: #if defined(PETSC_HAVE_SPOOLES)
5549:   PetscObjectComposeFunctionDynamic((PetscObject)B,"MatGetFactor_spooles_C",
5550:                                      "MatGetFactor_mpiaij_spooles",
5551:                                      MatGetFactor_mpiaij_spooles);
5552: #endif
5553: #if defined(PETSC_HAVE_MUMPS)
5554:   PetscObjectComposeFunctionDynamic((PetscObject)B,"MatGetFactor_mumps_C",
5555:                                      "MatGetFactor_aij_mumps",
5556:                                      MatGetFactor_aij_mumps);
5557: #endif
5558: #if defined(PETSC_HAVE_PASTIX)
5559:   PetscObjectComposeFunctionDynamic((PetscObject)B,"MatGetFactor_pastix_C",
5560:                                            "MatGetFactor_mpiaij_pastix",
5561:                                            MatGetFactor_mpiaij_pastix);
5562: #endif
5563: #if defined(PETSC_HAVE_SUPERLU_DIST)
5564:   PetscObjectComposeFunctionDynamic((PetscObject)B,"MatGetFactor_superlu_dist_C",
5565:                                      "MatGetFactor_mpiaij_superlu_dist",
5566:                                      MatGetFactor_mpiaij_superlu_dist);
5567: #endif
5568:   PetscObjectComposeFunctionDynamic((PetscObject)B,"MatStoreValues_C",
5569:                                      "MatStoreValues_MPIAIJ",
5570:                                      MatStoreValues_MPIAIJ);
5571:   PetscObjectComposeFunctionDynamic((PetscObject)B,"MatRetrieveValues_C",
5572:                                      "MatRetrieveValues_MPIAIJ",
5573:                                      MatRetrieveValues_MPIAIJ);
5574:   PetscObjectComposeFunctionDynamic((PetscObject)B,"MatGetDiagonalBlock_C",
5575:                                      "MatGetDiagonalBlock_MPIAIJ",
5576:                                      MatGetDiagonalBlock_MPIAIJ);
5577:   PetscObjectComposeFunctionDynamic((PetscObject)B,"MatIsTranspose_C",
5578:                                      "MatIsTranspose_MPIAIJ",
5579:                                      MatIsTranspose_MPIAIJ);
5580:   PetscObjectComposeFunctionDynamic((PetscObject)B,"MatMPIAIJSetPreallocation_C",
5581:                                      "MatMPIAIJSetPreallocation_MPIAIJ",
5582:                                      MatMPIAIJSetPreallocation_MPIAIJ);
5583:   PetscObjectComposeFunctionDynamic((PetscObject)B,"MatMPIAIJSetPreallocationCSR_C",
5584:                                      "MatMPIAIJSetPreallocationCSR_MPIAIJ",
5585:                                      MatMPIAIJSetPreallocationCSR_MPIAIJ);
5586:   PetscObjectComposeFunctionDynamic((PetscObject)B,"MatDiagonalScaleLocal_C",
5587:                                      "MatDiagonalScaleLocal_MPIAIJ",
5588:                                      MatDiagonalScaleLocal_MPIAIJ);
5589:   PetscObjectComposeFunctionDynamic((PetscObject)B,"MatConvert_mpiaij_mpiaijperm_C",
5590:                                      "MatConvert_MPIAIJ_MPIAIJPERM",
5591:                                       MatConvert_MPIAIJ_MPIAIJPERM);
5592:   PetscObjectComposeFunctionDynamic((PetscObject)B,"MatConvert_mpiaij_mpiaijcrl_C",
5593:                                      "MatConvert_MPIAIJ_MPIAIJCRL",
5594:                                       MatConvert_MPIAIJ_MPIAIJCRL);
5595:   PetscObjectComposeFunctionDynamic((PetscObject)B,"MatConvert_mpiaij_mpisbaij_C",
5596:                                      "MatConvert_MPIAIJ_MPISBAIJ",
5597:                                       MatConvert_MPIAIJ_MPISBAIJ);
5598:   PetscObjectComposeFunctionDynamic((PetscObject)B,"MatMatMult_mpidense_mpiaij_C",
5599:                                      "MatMatMult_MPIDense_MPIAIJ",
5600:                                       MatMatMult_MPIDense_MPIAIJ);
5601:   PetscObjectComposeFunctionDynamic((PetscObject)B,"MatMatMultSymbolic_mpidense_mpiaij_C",
5602:                                      "MatMatMultSymbolic_MPIDense_MPIAIJ",
5603:                                      MatMatMultSymbolic_MPIDense_MPIAIJ);
5604:   PetscObjectComposeFunctionDynamic((PetscObject)B,"MatMatMultNumeric_mpidense_mpiaij_C",
5605:                                      "MatMatMultNumeric_MPIDense_MPIAIJ",
5606:                                       MatMatMultNumeric_MPIDense_MPIAIJ);
5607:   PetscObjectChangeTypeName((PetscObject)B,MATMPIAIJ);
5608:   return(0);
5609: }
5610: EXTERN_C_END

5614: /*@
5615:      MatCreateMPIAIJWithSplitArrays - creates a MPI AIJ matrix using arrays that contain the "diagonal"
5616:          and "off-diagonal" part of the matrix in CSR format.

5618:    Collective on MPI_Comm

5620:    Input Parameters:
5621: +  comm - MPI communicator
5622: .  m - number of local rows (Cannot be PETSC_DECIDE)
5623: .  n - This value should be the same as the local size used in creating the 
5624:        x vector for the matrix-vector product y = Ax. (or PETSC_DECIDE to have
5625:        calculated if N is given) For square matrices n is almost always m.
5626: .  M - number of global rows (or PETSC_DETERMINE to have calculated if m is given)
5627: .  N - number of global columns (or PETSC_DETERMINE to have calculated if n is given)
5628: .   i - row indices for "diagonal" portion of matrix
5629: .   j - column indices
5630: .   a - matrix values
5631: .   oi - row indices for "off-diagonal" portion of matrix
5632: .   oj - column indices
5633: -   oa - matrix values

5635:    Output Parameter:
5636: .   mat - the matrix

5638:    Level: advanced

5640:    Notes:
5641:        The i, j, and a arrays ARE NOT copied by this routine into the internal format used by PETSc. The user
5642:        must free the arrays once the matrix has been destroyed and not before.

5644:        The i and j indices are 0 based
5645:  
5646:        See MatCreateAIJ() for the definition of "diagonal" and "off-diagonal" portion of the matrix

5648:        This sets local rows and cannot be used to set off-processor values. 

5650:        You cannot later use MatSetValues() to change values in this matrix.

5652: .keywords: matrix, aij, compressed row, sparse, parallel

5654: .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatMPIAIJSetPreallocationCSR(),
5655:           MPIAIJ, MatCreateAIJ(), MatCreateMPIAIJWithArrays()
5656: @*/
5657: PetscErrorCode  MatCreateMPIAIJWithSplitArrays(MPI_Comm comm,PetscInt m,PetscInt n,PetscInt M,PetscInt N,PetscInt i[],PetscInt j[],PetscScalar a[],
5658:                                                                 PetscInt oi[], PetscInt oj[],PetscScalar oa[],Mat *mat)
5659: {
5661:   Mat_MPIAIJ     *maij;

5664:   if (m < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"local number of rows (m) cannot be PETSC_DECIDE, or negative");
5665:   if (i[0]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"i (row indices) must start with 0");
5666:   if (oi[0]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"oi (row indices) must start with 0");
5667:   MatCreate(comm,mat);
5668:   MatSetSizes(*mat,m,n,M,N);
5669:   MatSetType(*mat,MATMPIAIJ);
5670:   maij = (Mat_MPIAIJ*) (*mat)->data;
5671:   maij->donotstash     = PETSC_TRUE;
5672:   (*mat)->preallocated = PETSC_TRUE;

5674:   PetscLayoutSetUp((*mat)->rmap);
5675:   PetscLayoutSetUp((*mat)->cmap);

5677:   MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,m,n,i,j,a,&maij->A);
5678:   MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,m,(*mat)->cmap->N,oi,oj,oa,&maij->B);

5680:   MatAssemblyBegin(maij->A,MAT_FINAL_ASSEMBLY);
5681:   MatAssemblyEnd(maij->A,MAT_FINAL_ASSEMBLY);
5682:   MatAssemblyBegin(maij->B,MAT_FINAL_ASSEMBLY);
5683:   MatAssemblyEnd(maij->B,MAT_FINAL_ASSEMBLY);

5685:   MatAssemblyBegin(*mat,MAT_FINAL_ASSEMBLY);
5686:   MatAssemblyEnd(*mat,MAT_FINAL_ASSEMBLY);
5687:   return(0);
5688: }

5690: /*
5691:     Special version for direct calls from Fortran 
5692: */
5693: #include <petsc-private/fortranimpl.h>

5695: #if defined(PETSC_HAVE_FORTRAN_CAPS)
5696: #define matsetvaluesmpiaij_ MATSETVALUESMPIAIJ
5697: #elif !defined(PETSC_HAVE_FORTRAN_UNDERSCORE)
5698: #define matsetvaluesmpiaij_ matsetvaluesmpiaij
5699: #endif

5701: /* Change these macros so can be used in void function */
5702: #undef CHKERRQ
5703: #define CHKERRQ(ierr) CHKERRABORT(PETSC_COMM_WORLD,ierr) 
5704: #undef SETERRQ2
5705: #define SETERRQ2(comm,ierr,b,c,d) CHKERRABORT(comm,ierr) 
5706: #undef SETERRQ3
5707: #define SETERRQ3(comm,ierr,b,c,d,e) CHKERRABORT(comm,ierr)
5708: #undef SETERRQ
5709: #define SETERRQ(c,ierr,b) CHKERRABORT(c,ierr) 

5711: EXTERN_C_BEGIN
5714: void PETSC_STDCALL matsetvaluesmpiaij_(Mat *mmat,PetscInt *mm,const PetscInt im[],PetscInt *mn,const PetscInt in[],const PetscScalar v[],InsertMode *maddv,PetscErrorCode *_ierr)
5715: {
5716:   Mat             mat = *mmat;
5717:   PetscInt        m = *mm, n = *mn;
5718:   InsertMode      addv = *maddv;
5719:   Mat_MPIAIJ      *aij = (Mat_MPIAIJ*)mat->data;
5720:   PetscScalar     value;
5721:   PetscErrorCode  ierr;

5723:   MatCheckPreallocated(mat,1);
5724:   if (mat->insertmode == NOT_SET_VALUES) {
5725:     mat->insertmode = addv;
5726:   }
5727: #if defined(PETSC_USE_DEBUG)
5728:   else if (mat->insertmode != addv) {
5729:     SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Cannot mix add values and insert values");
5730:   }
5731: #endif
5732:   {
5733:   PetscInt        i,j,rstart = mat->rmap->rstart,rend = mat->rmap->rend;
5734:   PetscInt        cstart = mat->cmap->rstart,cend = mat->cmap->rend,row,col;
5735:   PetscBool       roworiented = aij->roworiented;

5737:   /* Some Variables required in the macro */
5738:   Mat             A = aij->A;
5739:   Mat_SeqAIJ      *a = (Mat_SeqAIJ*)A->data;
5740:   PetscInt        *aimax = a->imax,*ai = a->i,*ailen = a->ilen,*aj = a->j;
5741:   MatScalar       *aa = a->a;
5742:   PetscBool       ignorezeroentries = (((a->ignorezeroentries)&&(addv==ADD_VALUES))?PETSC_TRUE:PETSC_FALSE);
5743:   Mat             B = aij->B;
5744:   Mat_SeqAIJ      *b = (Mat_SeqAIJ*)B->data;
5745:   PetscInt        *bimax = b->imax,*bi = b->i,*bilen = b->ilen,*bj = b->j,bm = aij->B->rmap->n,am = aij->A->rmap->n;
5746:   MatScalar       *ba = b->a;

5748:   PetscInt        *rp1,*rp2,ii,nrow1,nrow2,_i,rmax1,rmax2,N,low1,high1,low2,high2,t,lastcol1,lastcol2;
5749:   PetscInt        nonew = a->nonew;
5750:   MatScalar       *ap1,*ap2;

5753:   for (i=0; i<m; i++) {
5754:     if (im[i] < 0) continue;
5755: #if defined(PETSC_USE_DEBUG)
5756:     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);
5757: #endif
5758:     if (im[i] >= rstart && im[i] < rend) {
5759:       row      = im[i] - rstart;
5760:       lastcol1 = -1;
5761:       rp1      = aj + ai[row];
5762:       ap1      = aa + ai[row];
5763:       rmax1    = aimax[row];
5764:       nrow1    = ailen[row];
5765:       low1     = 0;
5766:       high1    = nrow1;
5767:       lastcol2 = -1;
5768:       rp2      = bj + bi[row];
5769:       ap2      = ba + bi[row];
5770:       rmax2    = bimax[row];
5771:       nrow2    = bilen[row];
5772:       low2     = 0;
5773:       high2    = nrow2;

5775:       for (j=0; j<n; j++) {
5776:         if (roworiented) value = v[i*n+j]; else value = v[i+j*m];
5777:         if (ignorezeroentries && value == 0.0 && (addv == ADD_VALUES)) continue;
5778:         if (in[j] >= cstart && in[j] < cend){
5779:           col = in[j] - cstart;
5780:           MatSetValues_SeqAIJ_A_Private(row,col,value,addv);
5781:         } else if (in[j] < 0) continue;
5782: #if defined(PETSC_USE_DEBUG)
5783:         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);
5784: #endif
5785:         else {
5786:           if (mat->was_assembled) {
5787:             if (!aij->colmap) {
5788:               MatCreateColmap_MPIAIJ_Private(mat);
5789:             }
5790: #if defined (PETSC_USE_CTABLE)
5791:             PetscTableFind(aij->colmap,in[j]+1,&col);
5792:             col--;
5793: #else
5794:             col = aij->colmap[in[j]] - 1;
5795: #endif
5796:             if (col < 0 && !((Mat_SeqAIJ*)(aij->A->data))->nonew) {
5797:               MatDisAssemble_MPIAIJ(mat);
5798:               col =  in[j];
5799:               /* Reinitialize the variables required by MatSetValues_SeqAIJ_B_Private() */
5800:               B = aij->B;
5801:               b = (Mat_SeqAIJ*)B->data;
5802:               bimax = b->imax; bi = b->i; bilen = b->ilen; bj = b->j;
5803:               rp2      = bj + bi[row];
5804:               ap2      = ba + bi[row];
5805:               rmax2    = bimax[row];
5806:               nrow2    = bilen[row];
5807:               low2     = 0;
5808:               high2    = nrow2;
5809:               bm       = aij->B->rmap->n;
5810:               ba = b->a;
5811:             }
5812:           } else col = in[j];
5813:           MatSetValues_SeqAIJ_B_Private(row,col,value,addv);
5814:         }
5815:       }
5816:     } else {
5817:       if (!aij->donotstash) {
5818:         if (roworiented) {
5819:           MatStashValuesRow_Private(&mat->stash,im[i],n,in,v+i*n,(PetscBool)(ignorezeroentries && (addv == ADD_VALUES)));
5820:         } else {
5821:           MatStashValuesCol_Private(&mat->stash,im[i],n,in,v+i,m,(PetscBool)(ignorezeroentries && (addv == ADD_VALUES)));
5822:         }
5823:       }
5824:     }
5825:   }}
5826:   PetscFunctionReturnVoid();
5827: }
5828: EXTERN_C_END