Actual source code: mpibaij.c

petsc-dev 2014-02-02
Report Typos and Errors
  2: #include <../src/mat/impls/baij/mpi/mpibaij.h>   /*I  "petscmat.h"  I*/
  3: #include <petscblaslapack.h>
  4: #include <petscsf.h>

  6: extern PetscErrorCode MatSetUpMultiply_MPIBAIJ(Mat);
  7: extern PetscErrorCode MatDisAssemble_MPIBAIJ(Mat);
  8: extern PetscErrorCode MatGetValues_SeqBAIJ(Mat,PetscInt,const PetscInt[],PetscInt,const PetscInt [],PetscScalar []);
  9: extern PetscErrorCode MatSetValues_SeqBAIJ(Mat,PetscInt,const PetscInt[],PetscInt,const PetscInt [],const PetscScalar [],InsertMode);
 10: extern PetscErrorCode MatSetValuesBlocked_SeqBAIJ(Mat,PetscInt,const PetscInt[],PetscInt,const PetscInt[],const PetscScalar[],InsertMode);
 11: extern PetscErrorCode MatGetRow_SeqBAIJ(Mat,PetscInt,PetscInt*,PetscInt*[],PetscScalar*[]);
 12: extern PetscErrorCode MatRestoreRow_SeqBAIJ(Mat,PetscInt,PetscInt*,PetscInt*[],PetscScalar*[]);
 13: extern PetscErrorCode MatZeroRows_SeqBAIJ(Mat,PetscInt,const PetscInt[],PetscScalar,Vec,Vec);

 17: PetscErrorCode MatGetRowMaxAbs_MPIBAIJ(Mat A,Vec v,PetscInt idx[])
 18: {
 19:   Mat_MPIBAIJ    *a = (Mat_MPIBAIJ*)A->data;
 21:   PetscInt       i,*idxb = 0;
 22:   PetscScalar    *va,*vb;
 23:   Vec            vtmp;

 26:   MatGetRowMaxAbs(a->A,v,idx);
 27:   VecGetArray(v,&va);
 28:   if (idx) {
 29:     for (i=0; i<A->rmap->n; i++) {
 30:       if (PetscAbsScalar(va[i])) idx[i] += A->cmap->rstart;
 31:     }
 32:   }

 34:   VecCreateSeq(PETSC_COMM_SELF,A->rmap->n,&vtmp);
 35:   if (idx) {PetscMalloc1(A->rmap->n,&idxb);}
 36:   MatGetRowMaxAbs(a->B,vtmp,idxb);
 37:   VecGetArray(vtmp,&vb);

 39:   for (i=0; i<A->rmap->n; i++) {
 40:     if (PetscAbsScalar(va[i]) < PetscAbsScalar(vb[i])) {
 41:       va[i] = vb[i];
 42:       if (idx) idx[i] = A->cmap->bs*a->garray[idxb[i]/A->cmap->bs] + (idxb[i] % A->cmap->bs);
 43:     }
 44:   }

 46:   VecRestoreArray(v,&va);
 47:   VecRestoreArray(vtmp,&vb);
 48:   PetscFree(idxb);
 49:   VecDestroy(&vtmp);
 50:   return(0);
 51: }

 55: PetscErrorCode  MatStoreValues_MPIBAIJ(Mat mat)
 56: {
 57:   Mat_MPIBAIJ    *aij = (Mat_MPIBAIJ*)mat->data;

 61:   MatStoreValues(aij->A);
 62:   MatStoreValues(aij->B);
 63:   return(0);
 64: }

 68: PetscErrorCode  MatRetrieveValues_MPIBAIJ(Mat mat)
 69: {
 70:   Mat_MPIBAIJ    *aij = (Mat_MPIBAIJ*)mat->data;

 74:   MatRetrieveValues(aij->A);
 75:   MatRetrieveValues(aij->B);
 76:   return(0);
 77: }

 79: /*
 80:      Local utility routine that creates a mapping from the global column
 81:    number to the local number in the off-diagonal part of the local
 82:    storage of the matrix.  This is done in a non scalable way since the
 83:    length of colmap equals the global matrix length.
 84: */
 87: PetscErrorCode MatCreateColmap_MPIBAIJ_Private(Mat mat)
 88: {
 89:   Mat_MPIBAIJ    *baij = (Mat_MPIBAIJ*)mat->data;
 90:   Mat_SeqBAIJ    *B    = (Mat_SeqBAIJ*)baij->B->data;
 92:   PetscInt       nbs = B->nbs,i,bs=mat->rmap->bs;

 95: #if defined(PETSC_USE_CTABLE)
 96:   PetscTableCreate(baij->nbs,baij->Nbs+1,&baij->colmap);
 97:   for (i=0; i<nbs; i++) {
 98:     PetscTableAdd(baij->colmap,baij->garray[i]+1,i*bs+1,INSERT_VALUES);
 99:   }
100: #else
101:   PetscMalloc1((baij->Nbs+1),&baij->colmap);
102:   PetscLogObjectMemory((PetscObject)mat,baij->Nbs*sizeof(PetscInt));
103:   PetscMemzero(baij->colmap,baij->Nbs*sizeof(PetscInt));
104:   for (i=0; i<nbs; i++) baij->colmap[baij->garray[i]] = i*bs+1;
105: #endif
106:   return(0);
107: }

109: #define  MatSetValues_SeqBAIJ_A_Private(row,col,value,addv) \
110:   { \
111:  \
112:     brow = row/bs;  \
113:     rp   = aj + ai[brow]; ap = aa + bs2*ai[brow]; \
114:     rmax = aimax[brow]; nrow = ailen[brow]; \
115:     bcol = col/bs; \
116:     ridx = row % bs; cidx = col % bs; \
117:     low  = 0; high = nrow; \
118:     while (high-low > 3) { \
119:       t = (low+high)/2; \
120:       if (rp[t] > bcol) high = t; \
121:       else              low  = t; \
122:     } \
123:     for (_i=low; _i<high; _i++) { \
124:       if (rp[_i] > bcol) break; \
125:       if (rp[_i] == bcol) { \
126:         bap = ap +  bs2*_i + bs*cidx + ridx; \
127:         if (addv == ADD_VALUES) *bap += value;  \
128:         else                    *bap  = value;  \
129:         goto a_noinsert; \
130:       } \
131:     } \
132:     if (a->nonew == 1) goto a_noinsert; \
133:     if (a->nonew == -1) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero (%D, %D) into matrix", row, col); \
134:     MatSeqXAIJReallocateAIJ(A,a->mbs,bs2,nrow,brow,bcol,rmax,aa,ai,aj,rp,ap,aimax,a->nonew,MatScalar); \
135:     N = nrow++ - 1;  \
136:     /* shift up all the later entries in this row */ \
137:     for (ii=N; ii>=_i; ii--) { \
138:       rp[ii+1] = rp[ii]; \
139:       PetscMemcpy(ap+bs2*(ii+1),ap+bs2*(ii),bs2*sizeof(MatScalar)); \
140:     } \
141:     if (N>=_i) { PetscMemzero(ap+bs2*_i,bs2*sizeof(MatScalar)); }  \
142:     rp[_i]                      = bcol;  \
143:     ap[bs2*_i + bs*cidx + ridx] = value;  \
144: a_noinsert:; \
145:     ailen[brow] = nrow; \
146:   }

148: #define  MatSetValues_SeqBAIJ_B_Private(row,col,value,addv) \
149:   { \
150:     brow = row/bs;  \
151:     rp   = bj + bi[brow]; ap = ba + bs2*bi[brow]; \
152:     rmax = bimax[brow]; nrow = bilen[brow]; \
153:     bcol = col/bs; \
154:     ridx = row % bs; cidx = col % bs; \
155:     low  = 0; high = nrow; \
156:     while (high-low > 3) { \
157:       t = (low+high)/2; \
158:       if (rp[t] > bcol) high = t; \
159:       else              low  = t; \
160:     } \
161:     for (_i=low; _i<high; _i++) { \
162:       if (rp[_i] > bcol) break; \
163:       if (rp[_i] == bcol) { \
164:         bap = ap +  bs2*_i + bs*cidx + ridx; \
165:         if (addv == ADD_VALUES) *bap += value;  \
166:         else                    *bap  = value;  \
167:         goto b_noinsert; \
168:       } \
169:     } \
170:     if (b->nonew == 1) goto b_noinsert; \
171:     if (b->nonew == -1) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero (%D, %D) into matrix", row, col); \
172:     MatSeqXAIJReallocateAIJ(B,b->mbs,bs2,nrow,brow,bcol,rmax,ba,bi,bj,rp,ap,bimax,b->nonew,MatScalar); \
173:     N = nrow++ - 1;  \
174:     /* shift up all the later entries in this row */ \
175:     for (ii=N; ii>=_i; ii--) { \
176:       rp[ii+1] = rp[ii]; \
177:       PetscMemcpy(ap+bs2*(ii+1),ap+bs2*(ii),bs2*sizeof(MatScalar)); \
178:     } \
179:     if (N>=_i) { PetscMemzero(ap+bs2*_i,bs2*sizeof(MatScalar));}  \
180:     rp[_i]                      = bcol;  \
181:     ap[bs2*_i + bs*cidx + ridx] = value;  \
182: b_noinsert:; \
183:     bilen[brow] = nrow; \
184:   }

188: PetscErrorCode MatSetValues_MPIBAIJ(Mat mat,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],const PetscScalar v[],InsertMode addv)
189: {
190:   Mat_MPIBAIJ    *baij = (Mat_MPIBAIJ*)mat->data;
191:   MatScalar      value;
192:   PetscBool      roworiented = baij->roworiented;
194:   PetscInt       i,j,row,col;
195:   PetscInt       rstart_orig=mat->rmap->rstart;
196:   PetscInt       rend_orig  =mat->rmap->rend,cstart_orig=mat->cmap->rstart;
197:   PetscInt       cend_orig  =mat->cmap->rend,bs=mat->rmap->bs;

199:   /* Some Variables required in the macro */
200:   Mat         A     = baij->A;
201:   Mat_SeqBAIJ *a    = (Mat_SeqBAIJ*)(A)->data;
202:   PetscInt    *aimax=a->imax,*ai=a->i,*ailen=a->ilen,*aj=a->j;
203:   MatScalar   *aa   =a->a;

205:   Mat         B     = baij->B;
206:   Mat_SeqBAIJ *b    = (Mat_SeqBAIJ*)(B)->data;
207:   PetscInt    *bimax=b->imax,*bi=b->i,*bilen=b->ilen,*bj=b->j;
208:   MatScalar   *ba   =b->a;

210:   PetscInt  *rp,ii,nrow,_i,rmax,N,brow,bcol;
211:   PetscInt  low,high,t,ridx,cidx,bs2=a->bs2;
212:   MatScalar *ap,*bap;

216:   for (i=0; i<m; i++) {
217:     if (im[i] < 0) continue;
218: #if defined(PETSC_USE_DEBUG)
219:     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);
220: #endif
221:     if (im[i] >= rstart_orig && im[i] < rend_orig) {
222:       row = im[i] - rstart_orig;
223:       for (j=0; j<n; j++) {
224:         if (in[j] >= cstart_orig && in[j] < cend_orig) {
225:           col = in[j] - cstart_orig;
226:           if (roworiented) value = v[i*n+j];
227:           else             value = v[i+j*m];
228:           MatSetValues_SeqBAIJ_A_Private(row,col,value,addv);
229:           /* MatSetValues_SeqBAIJ(baij->A,1,&row,1,&col,&value,addv); */
230:         } else if (in[j] < 0) continue;
231: #if defined(PETSC_USE_DEBUG)
232:         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);
233: #endif
234:         else {
235:           if (mat->was_assembled) {
236:             if (!baij->colmap) {
237:               MatCreateColmap_MPIBAIJ_Private(mat);
238:             }
239: #if defined(PETSC_USE_CTABLE)
240:             PetscTableFind(baij->colmap,in[j]/bs + 1,&col);
241:             col  = col - 1;
242: #else
243:             col = baij->colmap[in[j]/bs] - 1;
244: #endif
245:             if (col < 0 && !((Mat_SeqBAIJ*)(baij->B->data))->nonew) {
246:               MatDisAssemble_MPIBAIJ(mat);
247:               col  =  in[j];
248:               /* Reinitialize the variables required by MatSetValues_SeqBAIJ_B_Private() */
249:               B    = baij->B;
250:               b    = (Mat_SeqBAIJ*)(B)->data;
251:               bimax=b->imax;bi=b->i;bilen=b->ilen;bj=b->j;
252:               ba   =b->a;
253:             } else if (col < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero (%D, %D) into matrix", im[i], in[j]);
254:             else col += in[j]%bs;
255:           } else col = in[j];
256:           if (roworiented) value = v[i*n+j];
257:           else             value = v[i+j*m];
258:           MatSetValues_SeqBAIJ_B_Private(row,col,value,addv);
259:           /* MatSetValues_SeqBAIJ(baij->B,1,&row,1,&col,&value,addv); */
260:         }
261:       }
262:     } else {
263:       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]);
264:       if (!baij->donotstash) {
265:         mat->assembled = PETSC_FALSE;
266:         if (roworiented) {
267:           MatStashValuesRow_Private(&mat->stash,im[i],n,in,v+i*n,PETSC_FALSE);
268:         } else {
269:           MatStashValuesCol_Private(&mat->stash,im[i],n,in,v+i,m,PETSC_FALSE);
270:         }
271:       }
272:     }
273:   }
274:   return(0);
275: }

279: PetscErrorCode MatSetValuesBlocked_MPIBAIJ(Mat mat,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],const PetscScalar v[],InsertMode addv)
280: {
281:   Mat_MPIBAIJ       *baij = (Mat_MPIBAIJ*)mat->data;
282:   const PetscScalar *value;
283:   MatScalar         *barray     = baij->barray;
284:   PetscBool         roworiented = baij->roworiented;
285:   PetscErrorCode    ierr;
286:   PetscInt          i,j,ii,jj,row,col,rstart=baij->rstartbs;
287:   PetscInt          rend=baij->rendbs,cstart=baij->cstartbs,stepval;
288:   PetscInt          cend=baij->cendbs,bs=mat->rmap->bs,bs2=baij->bs2;

291:   if (!barray) {
292:     PetscMalloc1(bs2,&barray);
293:     baij->barray = barray;
294:   }

296:   if (roworiented) stepval = (n-1)*bs;
297:   else stepval = (m-1)*bs;

299:   for (i=0; i<m; i++) {
300:     if (im[i] < 0) continue;
301: #if defined(PETSC_USE_DEBUG)
302:     if (im[i] >= baij->Mbs) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Row too large, row %D max %D",im[i],baij->Mbs-1);
303: #endif
304:     if (im[i] >= rstart && im[i] < rend) {
305:       row = im[i] - rstart;
306:       for (j=0; j<n; j++) {
307:         /* If NumCol = 1 then a copy is not required */
308:         if ((roworiented) && (n == 1)) {
309:           barray = (MatScalar*)v + i*bs2;
310:         } else if ((!roworiented) && (m == 1)) {
311:           barray = (MatScalar*)v + j*bs2;
312:         } else { /* Here a copy is required */
313:           if (roworiented) {
314:             value = v + (i*(stepval+bs) + j)*bs;
315:           } else {
316:             value = v + (j*(stepval+bs) + i)*bs;
317:           }
318:           for (ii=0; ii<bs; ii++,value+=bs+stepval) {
319:             for (jj=0; jj<bs; jj++) barray[jj] = value[jj];
320:             barray += bs;
321:           }
322:           barray -= bs2;
323:         }

325:         if (in[j] >= cstart && in[j] < cend) {
326:           col  = in[j] - cstart;
327:           MatSetValuesBlocked_SeqBAIJ(baij->A,1,&row,1,&col,barray,addv);
328:         } else if (in[j] < 0) continue;
329: #if defined(PETSC_USE_DEBUG)
330:         else if (in[j] >= baij->Nbs) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Column too large, col %D max %D",in[j],baij->Nbs-1);
331: #endif
332:         else {
333:           if (mat->was_assembled) {
334:             if (!baij->colmap) {
335:               MatCreateColmap_MPIBAIJ_Private(mat);
336:             }

338: #if defined(PETSC_USE_DEBUG)
339: #if defined(PETSC_USE_CTABLE)
340:             { PetscInt data;
341:               PetscTableFind(baij->colmap,in[j]+1,&data);
342:               if ((data - 1) % bs) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Incorrect colmap");
343:             }
344: #else
345:             if ((baij->colmap[in[j]] - 1) % bs) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Incorrect colmap");
346: #endif
347: #endif
348: #if defined(PETSC_USE_CTABLE)
349:             PetscTableFind(baij->colmap,in[j]+1,&col);
350:             col  = (col - 1)/bs;
351: #else
352:             col = (baij->colmap[in[j]] - 1)/bs;
353: #endif
354:             if (col < 0 && !((Mat_SeqBAIJ*)(baij->B->data))->nonew) {
355:               MatDisAssemble_MPIBAIJ(mat);
356:               col  =  in[j];
357:             } else if (col < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero (%D, %D) into matrix", bs*im[i], bs*in[j]);
358:           } else col = in[j];
359:           MatSetValuesBlocked_SeqBAIJ(baij->B,1,&row,1,&col,barray,addv);
360:         }
361:       }
362:     } else {
363:       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]);
364:       if (!baij->donotstash) {
365:         if (roworiented) {
366:           MatStashValuesRowBlocked_Private(&mat->bstash,im[i],n,in,v,m,n,i);
367:         } else {
368:           MatStashValuesColBlocked_Private(&mat->bstash,im[i],n,in,v,m,n,i);
369:         }
370:       }
371:     }
372:   }
373:   return(0);
374: }

376: #define HASH_KEY 0.6180339887
377: #define HASH(size,key,tmp) (tmp = (key)*HASH_KEY,(PetscInt)((size)*(tmp-(PetscInt)tmp)))
378: /* #define HASH(size,key) ((PetscInt)((size)*fmod(((key)*HASH_KEY),1))) */
379: /* #define HASH(size,key,tmp) ((PetscInt)((size)*fmod(((key)*HASH_KEY),1))) */
382: PetscErrorCode MatSetValues_MPIBAIJ_HT(Mat mat,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],const PetscScalar v[],InsertMode addv)
383: {
384:   Mat_MPIBAIJ    *baij       = (Mat_MPIBAIJ*)mat->data;
385:   PetscBool      roworiented = baij->roworiented;
387:   PetscInt       i,j,row,col;
388:   PetscInt       rstart_orig=mat->rmap->rstart;
389:   PetscInt       rend_orig  =mat->rmap->rend,Nbs=baij->Nbs;
390:   PetscInt       h1,key,size=baij->ht_size,bs=mat->rmap->bs,*HT=baij->ht,idx;
391:   PetscReal      tmp;
392:   MatScalar      **HD = baij->hd,value;
393: #if defined(PETSC_USE_DEBUG)
394:   PetscInt total_ct=baij->ht_total_ct,insert_ct=baij->ht_insert_ct;
395: #endif

399:   for (i=0; i<m; i++) {
400: #if defined(PETSC_USE_DEBUG)
401:     if (im[i] < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative row");
402:     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);
403: #endif
404:     row = im[i];
405:     if (row >= rstart_orig && row < rend_orig) {
406:       for (j=0; j<n; j++) {
407:         col = in[j];
408:         if (roworiented) value = v[i*n+j];
409:         else             value = v[i+j*m];
410:         /* Look up PetscInto the Hash Table */
411:         key = (row/bs)*Nbs+(col/bs)+1;
412:         h1  = HASH(size,key,tmp);


415:         idx = h1;
416: #if defined(PETSC_USE_DEBUG)
417:         insert_ct++;
418:         total_ct++;
419:         if (HT[idx] != key) {
420:           for (idx=h1; (idx<size) && (HT[idx]!=key); idx++,total_ct++) ;
421:           if (idx == size) {
422:             for (idx=0; (idx<h1) && (HT[idx]!=key); idx++,total_ct++) ;
423:             if (idx == h1) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"(%D,%D) has no entry in the hash table", row, col);
424:           }
425:         }
426: #else
427:         if (HT[idx] != key) {
428:           for (idx=h1; (idx<size) && (HT[idx]!=key); idx++) ;
429:           if (idx == size) {
430:             for (idx=0; (idx<h1) && (HT[idx]!=key); idx++) ;
431:             if (idx == h1) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"(%D,%D) has no entry in the hash table", row, col);
432:           }
433:         }
434: #endif
435:         /* A HASH table entry is found, so insert the values at the correct address */
436:         if (addv == ADD_VALUES) *(HD[idx]+ (col % bs)*bs + (row % bs)) += value;
437:         else                    *(HD[idx]+ (col % bs)*bs + (row % bs))  = value;
438:       }
439:     } else if (!baij->donotstash) {
440:       if (roworiented) {
441:         MatStashValuesRow_Private(&mat->stash,im[i],n,in,v+i*n,PETSC_FALSE);
442:       } else {
443:         MatStashValuesCol_Private(&mat->stash,im[i],n,in,v+i,m,PETSC_FALSE);
444:       }
445:     }
446:   }
447: #if defined(PETSC_USE_DEBUG)
448:   baij->ht_total_ct  = total_ct;
449:   baij->ht_insert_ct = insert_ct;
450: #endif
451:   return(0);
452: }

456: PetscErrorCode MatSetValuesBlocked_MPIBAIJ_HT(Mat mat,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],const PetscScalar v[],InsertMode addv)
457: {
458:   Mat_MPIBAIJ       *baij       = (Mat_MPIBAIJ*)mat->data;
459:   PetscBool         roworiented = baij->roworiented;
460:   PetscErrorCode    ierr;
461:   PetscInt          i,j,ii,jj,row,col;
462:   PetscInt          rstart=baij->rstartbs;
463:   PetscInt          rend  =mat->rmap->rend,stepval,bs=mat->rmap->bs,bs2=baij->bs2,nbs2=n*bs2;
464:   PetscInt          h1,key,size=baij->ht_size,idx,*HT=baij->ht,Nbs=baij->Nbs;
465:   PetscReal         tmp;
466:   MatScalar         **HD = baij->hd,*baij_a;
467:   const PetscScalar *v_t,*value;
468: #if defined(PETSC_USE_DEBUG)
469:   PetscInt total_ct=baij->ht_total_ct,insert_ct=baij->ht_insert_ct;
470: #endif

473:   if (roworiented) stepval = (n-1)*bs;
474:   else stepval = (m-1)*bs;

476:   for (i=0; i<m; i++) {
477: #if defined(PETSC_USE_DEBUG)
478:     if (im[i] < 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative row: %D",im[i]);
479:     if (im[i] >= baij->Mbs) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Row too large: row %D max %D",im[i],baij->Mbs-1);
480: #endif
481:     row = im[i];
482:     v_t = v + i*nbs2;
483:     if (row >= rstart && row < rend) {
484:       for (j=0; j<n; j++) {
485:         col = in[j];

487:         /* Look up into the Hash Table */
488:         key = row*Nbs+col+1;
489:         h1  = HASH(size,key,tmp);

491:         idx = h1;
492: #if defined(PETSC_USE_DEBUG)
493:         total_ct++;
494:         insert_ct++;
495:         if (HT[idx] != key) {
496:           for (idx=h1; (idx<size) && (HT[idx]!=key); idx++,total_ct++) ;
497:           if (idx == size) {
498:             for (idx=0; (idx<h1) && (HT[idx]!=key); idx++,total_ct++) ;
499:             if (idx == h1) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"(%D,%D) has no entry in the hash table", row, col);
500:           }
501:         }
502: #else
503:         if (HT[idx] != key) {
504:           for (idx=h1; (idx<size) && (HT[idx]!=key); idx++) ;
505:           if (idx == size) {
506:             for (idx=0; (idx<h1) && (HT[idx]!=key); idx++) ;
507:             if (idx == h1) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"(%D,%D) has no entry in the hash table", row, col);
508:           }
509:         }
510: #endif
511:         baij_a = HD[idx];
512:         if (roworiented) {
513:           /*value = v + i*(stepval+bs)*bs + j*bs;*/
514:           /* value = v + (i*(stepval+bs)+j)*bs; */
515:           value = v_t;
516:           v_t  += bs;
517:           if (addv == ADD_VALUES) {
518:             for (ii=0; ii<bs; ii++,value+=stepval) {
519:               for (jj=ii; jj<bs2; jj+=bs) {
520:                 baij_a[jj] += *value++;
521:               }
522:             }
523:           } else {
524:             for (ii=0; ii<bs; ii++,value+=stepval) {
525:               for (jj=ii; jj<bs2; jj+=bs) {
526:                 baij_a[jj] = *value++;
527:               }
528:             }
529:           }
530:         } else {
531:           value = v + j*(stepval+bs)*bs + i*bs;
532:           if (addv == ADD_VALUES) {
533:             for (ii=0; ii<bs; ii++,value+=stepval,baij_a+=bs) {
534:               for (jj=0; jj<bs; jj++) {
535:                 baij_a[jj] += *value++;
536:               }
537:             }
538:           } else {
539:             for (ii=0; ii<bs; ii++,value+=stepval,baij_a+=bs) {
540:               for (jj=0; jj<bs; jj++) {
541:                 baij_a[jj] = *value++;
542:               }
543:             }
544:           }
545:         }
546:       }
547:     } else {
548:       if (!baij->donotstash) {
549:         if (roworiented) {
550:           MatStashValuesRowBlocked_Private(&mat->bstash,im[i],n,in,v,m,n,i);
551:         } else {
552:           MatStashValuesColBlocked_Private(&mat->bstash,im[i],n,in,v,m,n,i);
553:         }
554:       }
555:     }
556:   }
557: #if defined(PETSC_USE_DEBUG)
558:   baij->ht_total_ct  = total_ct;
559:   baij->ht_insert_ct = insert_ct;
560: #endif
561:   return(0);
562: }

566: PetscErrorCode MatGetValues_MPIBAIJ(Mat mat,PetscInt m,const PetscInt idxm[],PetscInt n,const PetscInt idxn[],PetscScalar v[])
567: {
568:   Mat_MPIBAIJ    *baij = (Mat_MPIBAIJ*)mat->data;
570:   PetscInt       bs       = mat->rmap->bs,i,j,bsrstart = mat->rmap->rstart,bsrend = mat->rmap->rend;
571:   PetscInt       bscstart = mat->cmap->rstart,bscend = mat->cmap->rend,row,col,data;

574:   for (i=0; i<m; i++) {
575:     if (idxm[i] < 0) continue; /* SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative row: %D",idxm[i]);*/
576:     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);
577:     if (idxm[i] >= bsrstart && idxm[i] < bsrend) {
578:       row = idxm[i] - bsrstart;
579:       for (j=0; j<n; j++) {
580:         if (idxn[j] < 0) continue; /* SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative column: %D",idxn[j]); */
581:         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);
582:         if (idxn[j] >= bscstart && idxn[j] < bscend) {
583:           col  = idxn[j] - bscstart;
584:           MatGetValues_SeqBAIJ(baij->A,1,&row,1,&col,v+i*n+j);
585:         } else {
586:           if (!baij->colmap) {
587:             MatCreateColmap_MPIBAIJ_Private(mat);
588:           }
589: #if defined(PETSC_USE_CTABLE)
590:           PetscTableFind(baij->colmap,idxn[j]/bs+1,&data);
591:           data--;
592: #else
593:           data = baij->colmap[idxn[j]/bs]-1;
594: #endif
595:           if ((data < 0) || (baij->garray[data/bs] != idxn[j]/bs)) *(v+i*n+j) = 0.0;
596:           else {
597:             col  = data + idxn[j]%bs;
598:             MatGetValues_SeqBAIJ(baij->B,1,&row,1,&col,v+i*n+j);
599:           }
600:         }
601:       }
602:     } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Only local values currently supported");
603:   }
604:   return(0);
605: }

609: PetscErrorCode MatNorm_MPIBAIJ(Mat mat,NormType type,PetscReal *nrm)
610: {
611:   Mat_MPIBAIJ    *baij = (Mat_MPIBAIJ*)mat->data;
612:   Mat_SeqBAIJ    *amat = (Mat_SeqBAIJ*)baij->A->data,*bmat = (Mat_SeqBAIJ*)baij->B->data;
614:   PetscInt       i,j,bs2=baij->bs2,bs=baij->A->rmap->bs,nz,row,col;
615:   PetscReal      sum = 0.0;
616:   MatScalar      *v;

619:   if (baij->size == 1) {
620:      MatNorm(baij->A,type,nrm);
621:   } else {
622:     if (type == NORM_FROBENIUS) {
623:       v  = amat->a;
624:       nz = amat->nz*bs2;
625:       for (i=0; i<nz; i++) {
626:         sum += PetscRealPart(PetscConj(*v)*(*v)); v++;
627:       }
628:       v  = bmat->a;
629:       nz = bmat->nz*bs2;
630:       for (i=0; i<nz; i++) {
631:         sum += PetscRealPart(PetscConj(*v)*(*v)); v++;
632:       }
633:       MPI_Allreduce(&sum,nrm,1,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)mat));
634:       *nrm = PetscSqrtReal(*nrm);
635:     } else if (type == NORM_1) { /* max column sum */
636:       PetscReal *tmp,*tmp2;
637:       PetscInt  *jj,*garray=baij->garray,cstart=baij->rstartbs;
638:       PetscMalloc2(mat->cmap->N,&tmp,mat->cmap->N,&tmp2);
639:       PetscMemzero(tmp,mat->cmap->N*sizeof(PetscReal));
640:       v    = amat->a; jj = amat->j;
641:       for (i=0; i<amat->nz; i++) {
642:         for (j=0; j<bs; j++) {
643:           col = bs*(cstart + *jj) + j; /* column index */
644:           for (row=0; row<bs; row++) {
645:             tmp[col] += PetscAbsScalar(*v);  v++;
646:           }
647:         }
648:         jj++;
649:       }
650:       v = bmat->a; jj = bmat->j;
651:       for (i=0; i<bmat->nz; i++) {
652:         for (j=0; j<bs; j++) {
653:           col = bs*garray[*jj] + j;
654:           for (row=0; row<bs; row++) {
655:             tmp[col] += PetscAbsScalar(*v); v++;
656:           }
657:         }
658:         jj++;
659:       }
660:       MPI_Allreduce(tmp,tmp2,mat->cmap->N,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)mat));
661:       *nrm = 0.0;
662:       for (j=0; j<mat->cmap->N; j++) {
663:         if (tmp2[j] > *nrm) *nrm = tmp2[j];
664:       }
665:       PetscFree2(tmp,tmp2);
666:     } else if (type == NORM_INFINITY) { /* max row sum */
667:       PetscReal *sums;
668:       PetscMalloc1(bs,&sums);
669:       sum  = 0.0;
670:       for (j=0; j<amat->mbs; j++) {
671:         for (row=0; row<bs; row++) sums[row] = 0.0;
672:         v  = amat->a + bs2*amat->i[j];
673:         nz = amat->i[j+1]-amat->i[j];
674:         for (i=0; i<nz; i++) {
675:           for (col=0; col<bs; col++) {
676:             for (row=0; row<bs; row++) {
677:               sums[row] += PetscAbsScalar(*v); v++;
678:             }
679:           }
680:         }
681:         v  = bmat->a + bs2*bmat->i[j];
682:         nz = bmat->i[j+1]-bmat->i[j];
683:         for (i=0; i<nz; i++) {
684:           for (col=0; col<bs; col++) {
685:             for (row=0; row<bs; row++) {
686:               sums[row] += PetscAbsScalar(*v); v++;
687:             }
688:           }
689:         }
690:         for (row=0; row<bs; row++) {
691:           if (sums[row] > sum) sum = sums[row];
692:         }
693:       }
694:       MPI_Allreduce(&sum,nrm,1,MPIU_REAL,MPIU_MAX,PetscObjectComm((PetscObject)mat));
695:       PetscFree(sums);
696:     } else SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"No support for this norm yet");
697:   }
698:   return(0);
699: }

701: /*
702:   Creates the hash table, and sets the table
703:   This table is created only once.
704:   If new entried need to be added to the matrix
705:   then the hash table has to be destroyed and
706:   recreated.
707: */
710: PetscErrorCode MatCreateHashTable_MPIBAIJ_Private(Mat mat,PetscReal factor)
711: {
712:   Mat_MPIBAIJ    *baij = (Mat_MPIBAIJ*)mat->data;
713:   Mat            A     = baij->A,B=baij->B;
714:   Mat_SeqBAIJ    *a    = (Mat_SeqBAIJ*)A->data,*b=(Mat_SeqBAIJ*)B->data;
715:   PetscInt       i,j,k,nz=a->nz+b->nz,h1,*ai=a->i,*aj=a->j,*bi=b->i,*bj=b->j;
717:   PetscInt       ht_size,bs2=baij->bs2,rstart=baij->rstartbs;
718:   PetscInt       cstart=baij->cstartbs,*garray=baij->garray,row,col,Nbs=baij->Nbs;
719:   PetscInt       *HT,key;
720:   MatScalar      **HD;
721:   PetscReal      tmp;
722: #if defined(PETSC_USE_INFO)
723:   PetscInt ct=0,max=0;
724: #endif

727:   if (baij->ht) return(0);

729:   baij->ht_size = (PetscInt)(factor*nz);
730:   ht_size       = baij->ht_size;

732:   /* Allocate Memory for Hash Table */
733:   PetscCalloc2(ht_size,&baij->hd,ht_size,&baij->ht);
734:   HD   = baij->hd;
735:   HT   = baij->ht;

737:   /* Loop Over A */
738:   for (i=0; i<a->mbs; i++) {
739:     for (j=ai[i]; j<ai[i+1]; j++) {
740:       row = i+rstart;
741:       col = aj[j]+cstart;

743:       key = row*Nbs + col + 1;
744:       h1  = HASH(ht_size,key,tmp);
745:       for (k=0; k<ht_size; k++) {
746:         if (!HT[(h1+k)%ht_size]) {
747:           HT[(h1+k)%ht_size] = key;
748:           HD[(h1+k)%ht_size] = a->a + j*bs2;
749:           break;
750: #if defined(PETSC_USE_INFO)
751:         } else {
752:           ct++;
753: #endif
754:         }
755:       }
756: #if defined(PETSC_USE_INFO)
757:       if (k> max) max = k;
758: #endif
759:     }
760:   }
761:   /* Loop Over B */
762:   for (i=0; i<b->mbs; i++) {
763:     for (j=bi[i]; j<bi[i+1]; j++) {
764:       row = i+rstart;
765:       col = garray[bj[j]];
766:       key = row*Nbs + col + 1;
767:       h1  = HASH(ht_size,key,tmp);
768:       for (k=0; k<ht_size; k++) {
769:         if (!HT[(h1+k)%ht_size]) {
770:           HT[(h1+k)%ht_size] = key;
771:           HD[(h1+k)%ht_size] = b->a + j*bs2;
772:           break;
773: #if defined(PETSC_USE_INFO)
774:         } else {
775:           ct++;
776: #endif
777:         }
778:       }
779: #if defined(PETSC_USE_INFO)
780:       if (k> max) max = k;
781: #endif
782:     }
783:   }

785:   /* Print Summary */
786: #if defined(PETSC_USE_INFO)
787:   for (i=0,j=0; i<ht_size; i++) {
788:     if (HT[i]) j++;
789:   }
790:   PetscInfo2(mat,"Average Search = %5.2f,max search = %D\n",(!j)? 0.0:((PetscReal)(ct+j))/j,max);
791: #endif
792:   return(0);
793: }

797: PetscErrorCode MatAssemblyBegin_MPIBAIJ(Mat mat,MatAssemblyType mode)
798: {
799:   Mat_MPIBAIJ    *baij = (Mat_MPIBAIJ*)mat->data;
801:   PetscInt       nstash,reallocs;
802:   InsertMode     addv;

805:   if (baij->donotstash || mat->nooffprocentries) return(0);

807:   /* make sure all processors are either in INSERTMODE or ADDMODE */
808:   MPI_Allreduce((PetscEnum*)&mat->insertmode,(PetscEnum*)&addv,1,MPIU_ENUM,MPI_BOR,PetscObjectComm((PetscObject)mat));
809:   if (addv == (ADD_VALUES|INSERT_VALUES)) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_ARG_WRONGSTATE,"Some processors inserted others added");
810:   mat->insertmode = addv; /* in case this processor had no cache */

812:   MatStashScatterBegin_Private(mat,&mat->stash,mat->rmap->range);
813:   MatStashScatterBegin_Private(mat,&mat->bstash,baij->rangebs);
814:   MatStashGetInfo_Private(&mat->stash,&nstash,&reallocs);
815:   PetscInfo2(mat,"Stash has %D entries,uses %D mallocs.\n",nstash,reallocs);
816:   MatStashGetInfo_Private(&mat->bstash,&nstash,&reallocs);
817:   PetscInfo2(mat,"Block-Stash has %D entries, uses %D mallocs.\n",nstash,reallocs);
818:   return(0);
819: }

823: PetscErrorCode MatAssemblyEnd_MPIBAIJ(Mat mat,MatAssemblyType mode)
824: {
825:   Mat_MPIBAIJ    *baij=(Mat_MPIBAIJ*)mat->data;
826:   Mat_SeqBAIJ    *a   =(Mat_SeqBAIJ*)baij->A->data;
828:   PetscInt       i,j,rstart,ncols,flg,bs2=baij->bs2;
829:   PetscInt       *row,*col;
830:   PetscBool      r1,r2,r3,other_disassembled;
831:   MatScalar      *val;
832:   InsertMode     addv = mat->insertmode;
833:   PetscMPIInt    n;

836:   /* do not use 'b=(Mat_SeqBAIJ*)baij->B->data' as B can be reset in disassembly */
837:   if (!baij->donotstash && !mat->nooffprocentries) {
838:     while (1) {
839:       MatStashScatterGetMesg_Private(&mat->stash,&n,&row,&col,&val,&flg);
840:       if (!flg) break;

842:       for (i=0; i<n;) {
843:         /* Now identify the consecutive vals belonging to the same row */
844:         for (j=i,rstart=row[j]; j<n; j++) {
845:           if (row[j] != rstart) break;
846:         }
847:         if (j < n) ncols = j-i;
848:         else       ncols = n-i;
849:         /* Now assemble all these values with a single function call */
850:         MatSetValues_MPIBAIJ(mat,1,row+i,ncols,col+i,val+i,addv);
851:         i    = j;
852:       }
853:     }
854:     MatStashScatterEnd_Private(&mat->stash);
855:     /* Now process the block-stash. Since the values are stashed column-oriented,
856:        set the roworiented flag to column oriented, and after MatSetValues()
857:        restore the original flags */
858:     r1 = baij->roworiented;
859:     r2 = a->roworiented;
860:     r3 = ((Mat_SeqBAIJ*)baij->B->data)->roworiented;

862:     baij->roworiented = PETSC_FALSE;
863:     a->roworiented    = PETSC_FALSE;

865:     (((Mat_SeqBAIJ*)baij->B->data))->roworiented = PETSC_FALSE; /* b->roworiented */
866:     while (1) {
867:       MatStashScatterGetMesg_Private(&mat->bstash,&n,&row,&col,&val,&flg);
868:       if (!flg) break;

870:       for (i=0; i<n;) {
871:         /* Now identify the consecutive vals belonging to the same row */
872:         for (j=i,rstart=row[j]; j<n; j++) {
873:           if (row[j] != rstart) break;
874:         }
875:         if (j < n) ncols = j-i;
876:         else       ncols = n-i;
877:         MatSetValuesBlocked_MPIBAIJ(mat,1,row+i,ncols,col+i,val+i*bs2,addv);
878:         i    = j;
879:       }
880:     }
881:     MatStashScatterEnd_Private(&mat->bstash);

883:     baij->roworiented = r1;
884:     a->roworiented    = r2;

886:     ((Mat_SeqBAIJ*)baij->B->data)->roworiented = r3; /* b->roworiented */
887:   }

889:   MatAssemblyBegin(baij->A,mode);
890:   MatAssemblyEnd(baij->A,mode);

892:   /* determine if any processor has disassembled, if so we must
893:      also disassemble ourselfs, in order that we may reassemble. */
894:   /*
895:      if nonzero structure of submatrix B cannot change then we know that
896:      no processor disassembled thus we can skip this stuff
897:   */
898:   if (!((Mat_SeqBAIJ*)baij->B->data)->nonew) {
899:     MPI_Allreduce(&mat->was_assembled,&other_disassembled,1,MPIU_BOOL,MPI_PROD,PetscObjectComm((PetscObject)mat));
900:     if (mat->was_assembled && !other_disassembled) {
901:       MatDisAssemble_MPIBAIJ(mat);
902:     }
903:   }

905:   if (!mat->was_assembled && mode == MAT_FINAL_ASSEMBLY) {
906:     MatSetUpMultiply_MPIBAIJ(mat);
907:   }
908:   MatAssemblyBegin(baij->B,mode);
909:   MatAssemblyEnd(baij->B,mode);

911: #if defined(PETSC_USE_INFO)
912:   if (baij->ht && mode== MAT_FINAL_ASSEMBLY) {
913:     PetscInfo1(mat,"Average Hash Table Search in MatSetValues = %5.2f\n",((PetscReal)baij->ht_total_ct)/baij->ht_insert_ct);

915:     baij->ht_total_ct  = 0;
916:     baij->ht_insert_ct = 0;
917:   }
918: #endif
919:   if (baij->ht_flag && !baij->ht && mode == MAT_FINAL_ASSEMBLY) {
920:     MatCreateHashTable_MPIBAIJ_Private(mat,baij->ht_fact);

922:     mat->ops->setvalues        = MatSetValues_MPIBAIJ_HT;
923:     mat->ops->setvaluesblocked = MatSetValuesBlocked_MPIBAIJ_HT;
924:   }

926:   PetscFree2(baij->rowvalues,baij->rowindices);

928:   baij->rowvalues = 0;
929:   return(0);
930: }

932: #include <petscdraw.h>
935: static PetscErrorCode MatView_MPIBAIJ_ASCIIorDraworSocket(Mat mat,PetscViewer viewer)
936: {
937:   Mat_MPIBAIJ       *baij = (Mat_MPIBAIJ*)mat->data;
938:   PetscErrorCode    ierr;
939:   PetscMPIInt       size = baij->size,rank = baij->rank;
940:   PetscInt          bs   = mat->rmap->bs;
941:   PetscBool         iascii,isdraw;
942:   PetscViewer       sviewer;
943:   PetscViewerFormat format;

946:   PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii);
947:   PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERDRAW,&isdraw);
948:   if (iascii) {
949:     PetscViewerGetFormat(viewer,&format);
950:     if (format == PETSC_VIEWER_ASCII_INFO_DETAIL) {
951:       MatInfo info;
952:       MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);
953:       MatGetInfo(mat,MAT_LOCAL,&info);
954:       PetscViewerASCIISynchronizedAllow(viewer,PETSC_TRUE);
955:       PetscViewerASCIISynchronizedPrintf(viewer,"[%d] Local rows %D nz %D nz alloced %D bs %D mem %D\n",
956:                                                 rank,mat->rmap->n,(PetscInt)info.nz_used,(PetscInt)info.nz_allocated,mat->rmap->bs,(PetscInt)info.memory);
957:       MatGetInfo(baij->A,MAT_LOCAL,&info);
958:       PetscViewerASCIISynchronizedPrintf(viewer,"[%d] on-diagonal part: nz %D \n",rank,(PetscInt)info.nz_used);
959:       MatGetInfo(baij->B,MAT_LOCAL,&info);
960:       PetscViewerASCIISynchronizedPrintf(viewer,"[%d] off-diagonal part: nz %D \n",rank,(PetscInt)info.nz_used);
961:       PetscViewerFlush(viewer);
962:       PetscViewerASCIISynchronizedAllow(viewer,PETSC_FALSE);
963:       PetscViewerASCIIPrintf(viewer,"Information on VecScatter used in matrix-vector product: \n");
964:       VecScatterView(baij->Mvctx,viewer);
965:       return(0);
966:     } else if (format == PETSC_VIEWER_ASCII_INFO) {
967:       PetscViewerASCIIPrintf(viewer,"  block size is %D\n",bs);
968:       return(0);
969:     } else if (format == PETSC_VIEWER_ASCII_FACTOR_INFO) {
970:       return(0);
971:     }
972:   }

974:   if (isdraw) {
975:     PetscDraw draw;
976:     PetscBool isnull;
977:     PetscViewerDrawGetDraw(viewer,0,&draw);
978:     PetscDrawIsNull(draw,&isnull); if (isnull) return(0);
979:   }

981:   if (size == 1) {
982:     PetscObjectSetName((PetscObject)baij->A,((PetscObject)mat)->name);
983:     MatView(baij->A,viewer);
984:   } else {
985:     /* assemble the entire matrix onto first processor. */
986:     Mat         A;
987:     Mat_SeqBAIJ *Aloc;
988:     PetscInt    M = mat->rmap->N,N = mat->cmap->N,*ai,*aj,col,i,j,k,*rvals,mbs = baij->mbs;
989:     MatScalar   *a;

991:     /* Here we are creating a temporary matrix, so will assume MPIBAIJ is acceptable */
992:     /* Perhaps this should be the type of mat? */
993:     MatCreate(PetscObjectComm((PetscObject)mat),&A);
994:     if (!rank) {
995:       MatSetSizes(A,M,N,M,N);
996:     } else {
997:       MatSetSizes(A,0,0,M,N);
998:     }
999:     MatSetType(A,MATMPIBAIJ);
1000:     MatMPIBAIJSetPreallocation(A,mat->rmap->bs,0,NULL,0,NULL);
1001:     MatSetOption(A,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);
1002:     PetscLogObjectParent((PetscObject)mat,(PetscObject)A);

1004:     /* copy over the A part */
1005:     Aloc = (Mat_SeqBAIJ*)baij->A->data;
1006:     ai   = Aloc->i; aj = Aloc->j; a = Aloc->a;
1007:     PetscMalloc1(bs,&rvals);

1009:     for (i=0; i<mbs; i++) {
1010:       rvals[0] = bs*(baij->rstartbs + i);
1011:       for (j=1; j<bs; j++) rvals[j] = rvals[j-1] + 1;
1012:       for (j=ai[i]; j<ai[i+1]; j++) {
1013:         col = (baij->cstartbs+aj[j])*bs;
1014:         for (k=0; k<bs; k++) {
1015:           MatSetValues_MPIBAIJ(A,bs,rvals,1,&col,a,INSERT_VALUES);
1016:           col++; a += bs;
1017:         }
1018:       }
1019:     }
1020:     /* copy over the B part */
1021:     Aloc = (Mat_SeqBAIJ*)baij->B->data;
1022:     ai   = Aloc->i; aj = Aloc->j; a = Aloc->a;
1023:     for (i=0; i<mbs; i++) {
1024:       rvals[0] = bs*(baij->rstartbs + i);
1025:       for (j=1; j<bs; j++) rvals[j] = rvals[j-1] + 1;
1026:       for (j=ai[i]; j<ai[i+1]; j++) {
1027:         col = baij->garray[aj[j]]*bs;
1028:         for (k=0; k<bs; k++) {
1029:           MatSetValues_MPIBAIJ(A,bs,rvals,1,&col,a,INSERT_VALUES);
1030:           col++; a += bs;
1031:         }
1032:       }
1033:     }
1034:     PetscFree(rvals);
1035:     MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);
1036:     MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);
1037:     /*
1038:        Everyone has to call to draw the matrix since the graphics waits are
1039:        synchronized across all processors that share the PetscDraw object
1040:     */
1041:     PetscViewerGetSingleton(viewer,&sviewer);
1042:     if (!rank) {
1043:       PetscObjectSetName((PetscObject)((Mat_MPIBAIJ*)(A->data))->A,((PetscObject)mat)->name);
1044:       /* Set the type name to MATMPIBAIJ so that the correct type can be printed out by PetscObjectPrintClassNamePrefixType() in MatView_SeqBAIJ_ASCII()*/
1045:       PetscStrcpy(((PetscObject)((Mat_MPIBAIJ*)(A->data))->A)->type_name,MATMPIBAIJ);
1046:       MatView(((Mat_MPIBAIJ*)(A->data))->A,sviewer);
1047:     }
1048:     PetscViewerRestoreSingleton(viewer,&sviewer);
1049:     MatDestroy(&A);
1050:   }
1051:   return(0);
1052: }

1056: static PetscErrorCode MatView_MPIBAIJ_Binary(Mat mat,PetscViewer viewer)
1057: {
1058:   Mat_MPIBAIJ    *a = (Mat_MPIBAIJ*)mat->data;
1059:   Mat_SeqBAIJ    *A = (Mat_SeqBAIJ*)a->A->data;
1060:   Mat_SeqBAIJ    *B = (Mat_SeqBAIJ*)a->B->data;
1062:   PetscInt       i,*row_lens,*crow_lens,bs = mat->rmap->bs,j,k,bs2=a->bs2,header[4],nz,rlen;
1063:   PetscInt       *range=0,nzmax,*column_indices,cnt,col,*garray = a->garray,cstart = mat->cmap->rstart/bs,len,pcnt,l,ll;
1064:   int            fd;
1065:   PetscScalar    *column_values;
1066:   FILE           *file;
1067:   PetscMPIInt    rank,size,tag = ((PetscObject)viewer)->tag;
1068:   PetscInt       message_count,flowcontrolcount;

1071:   MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);
1072:   MPI_Comm_size(PetscObjectComm((PetscObject)mat),&size);
1073:   nz   = bs2*(A->nz + B->nz);
1074:   rlen = mat->rmap->n;
1075:   if (!rank) {
1076:     header[0] = MAT_FILE_CLASSID;
1077:     header[1] = mat->rmap->N;
1078:     header[2] = mat->cmap->N;

1080:     MPI_Reduce(&nz,&header[3],1,MPIU_INT,MPI_SUM,0,PetscObjectComm((PetscObject)mat));
1081:     PetscViewerBinaryGetDescriptor(viewer,&fd);
1082:     PetscBinaryWrite(fd,header,4,PETSC_INT,PETSC_TRUE);
1083:     /* get largest number of rows any processor has */
1084:     range = mat->rmap->range;
1085:     for (i=1; i<size; i++) {
1086:       rlen = PetscMax(rlen,range[i+1] - range[i]);
1087:     }
1088:   } else {
1089:     MPI_Reduce(&nz,0,1,MPIU_INT,MPI_SUM,0,PetscObjectComm((PetscObject)mat));
1090:   }

1092:   PetscMalloc1((rlen/bs),&crow_lens);
1093:   /* compute lengths of each row  */
1094:   for (i=0; i<a->mbs; i++) {
1095:     crow_lens[i] = A->i[i+1] - A->i[i] + B->i[i+1] - B->i[i];
1096:   }
1097:   /* store the row lengths to the file */
1098:   PetscViewerFlowControlStart(viewer,&message_count,&flowcontrolcount);
1099:   if (!rank) {
1100:     MPI_Status status;
1101:     PetscMalloc1(rlen,&row_lens);
1102:     rlen = (range[1] - range[0])/bs;
1103:     for (i=0; i<rlen; i++) {
1104:       for (j=0; j<bs; j++) {
1105:         row_lens[i*bs+j] = bs*crow_lens[i];
1106:       }
1107:     }
1108:     PetscBinaryWrite(fd,row_lens,bs*rlen,PETSC_INT,PETSC_TRUE);
1109:     for (i=1; i<size; i++) {
1110:       rlen = (range[i+1] - range[i])/bs;
1111:       PetscViewerFlowControlStepMaster(viewer,i,&message_count,flowcontrolcount);
1112:       MPI_Recv(crow_lens,rlen,MPIU_INT,i,tag,PetscObjectComm((PetscObject)mat),&status);
1113:       for (k=0; k<rlen; k++) {
1114:         for (j=0; j<bs; j++) {
1115:           row_lens[k*bs+j] = bs*crow_lens[k];
1116:         }
1117:       }
1118:       PetscBinaryWrite(fd,row_lens,bs*rlen,PETSC_INT,PETSC_TRUE);
1119:     }
1120:     PetscViewerFlowControlEndMaster(viewer,&message_count);
1121:     PetscFree(row_lens);
1122:   } else {
1123:     PetscViewerFlowControlStepWorker(viewer,rank,&message_count);
1124:     MPI_Send(crow_lens,mat->rmap->n/bs,MPIU_INT,0,tag,PetscObjectComm((PetscObject)mat));
1125:     PetscViewerFlowControlEndWorker(viewer,&message_count);
1126:   }
1127:   PetscFree(crow_lens);

1129:   /* load up the local column indices. Include for all rows not just one for each block row since process 0 does not have the
1130:      information needed to make it for each row from a block row. This does require more communication but still not more than
1131:      the communication needed for the nonzero values  */
1132:   nzmax = nz; /*  space a largest processor needs */
1133:   MPI_Reduce(&nz,&nzmax,1,MPIU_INT,MPI_MAX,0,PetscObjectComm((PetscObject)mat));
1134:   PetscMalloc1(nzmax,&column_indices);
1135:   cnt   = 0;
1136:   for (i=0; i<a->mbs; i++) {
1137:     pcnt = cnt;
1138:     for (j=B->i[i]; j<B->i[i+1]; j++) {
1139:       if ((col = garray[B->j[j]]) > cstart) break;
1140:       for (l=0; l<bs; l++) {
1141:         column_indices[cnt++] = bs*col+l;
1142:       }
1143:     }
1144:     for (k=A->i[i]; k<A->i[i+1]; k++) {
1145:       for (l=0; l<bs; l++) {
1146:         column_indices[cnt++] = bs*(A->j[k] + cstart)+l;
1147:       }
1148:     }
1149:     for (; j<B->i[i+1]; j++) {
1150:       for (l=0; l<bs; l++) {
1151:         column_indices[cnt++] = bs*garray[B->j[j]]+l;
1152:       }
1153:     }
1154:     len = cnt - pcnt;
1155:     for (k=1; k<bs; k++) {
1156:       PetscMemcpy(&column_indices[cnt],&column_indices[pcnt],len*sizeof(PetscInt));
1157:       cnt += len;
1158:     }
1159:   }
1160:   if (cnt != nz) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_LIB,"Internal PETSc error: cnt = %D nz = %D",cnt,nz);

1162:   /* store the columns to the file */
1163:   PetscViewerFlowControlStart(viewer,&message_count,&flowcontrolcount);
1164:   if (!rank) {
1165:     MPI_Status status;
1166:     PetscBinaryWrite(fd,column_indices,nz,PETSC_INT,PETSC_TRUE);
1167:     for (i=1; i<size; i++) {
1168:       PetscViewerFlowControlStepMaster(viewer,i,&message_count,flowcontrolcount);
1169:       MPI_Recv(&cnt,1,MPIU_INT,i,tag,PetscObjectComm((PetscObject)mat),&status);
1170:       MPI_Recv(column_indices,cnt,MPIU_INT,i,tag,PetscObjectComm((PetscObject)mat),&status);
1171:       PetscBinaryWrite(fd,column_indices,cnt,PETSC_INT,PETSC_TRUE);
1172:     }
1173:     PetscViewerFlowControlEndMaster(viewer,&message_count);
1174:   } else {
1175:     PetscViewerFlowControlStepWorker(viewer,rank,&message_count);
1176:     MPI_Send(&cnt,1,MPIU_INT,0,tag,PetscObjectComm((PetscObject)mat));
1177:     MPI_Send(column_indices,cnt,MPIU_INT,0,tag,PetscObjectComm((PetscObject)mat));
1178:     PetscViewerFlowControlEndWorker(viewer,&message_count);
1179:   }
1180:   PetscFree(column_indices);

1182:   /* load up the numerical values */
1183:   PetscMalloc1(nzmax,&column_values);
1184:   cnt  = 0;
1185:   for (i=0; i<a->mbs; i++) {
1186:     rlen = bs*(B->i[i+1] - B->i[i] + A->i[i+1] - A->i[i]);
1187:     for (j=B->i[i]; j<B->i[i+1]; j++) {
1188:       if (garray[B->j[j]] > cstart) break;
1189:       for (l=0; l<bs; l++) {
1190:         for (ll=0; ll<bs; ll++) {
1191:           column_values[cnt + l*rlen + ll] = B->a[bs2*j+l+bs*ll];
1192:         }
1193:       }
1194:       cnt += bs;
1195:     }
1196:     for (k=A->i[i]; k<A->i[i+1]; k++) {
1197:       for (l=0; l<bs; l++) {
1198:         for (ll=0; ll<bs; ll++) {
1199:           column_values[cnt + l*rlen + ll] = A->a[bs2*k+l+bs*ll];
1200:         }
1201:       }
1202:       cnt += bs;
1203:     }
1204:     for (; j<B->i[i+1]; j++) {
1205:       for (l=0; l<bs; l++) {
1206:         for (ll=0; ll<bs; ll++) {
1207:           column_values[cnt + l*rlen + ll] = B->a[bs2*j+l+bs*ll];
1208:         }
1209:       }
1210:       cnt += bs;
1211:     }
1212:     cnt += (bs-1)*rlen;
1213:   }
1214:   if (cnt != nz) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Internal PETSc error: cnt = %D nz = %D",cnt,nz);

1216:   /* store the column values to the file */
1217:   PetscViewerFlowControlStart(viewer,&message_count,&flowcontrolcount);
1218:   if (!rank) {
1219:     MPI_Status status;
1220:     PetscBinaryWrite(fd,column_values,nz,PETSC_SCALAR,PETSC_TRUE);
1221:     for (i=1; i<size; i++) {
1222:       PetscViewerFlowControlStepMaster(viewer,i,&message_count,flowcontrolcount);
1223:       MPI_Recv(&cnt,1,MPIU_INT,i,tag,PetscObjectComm((PetscObject)mat),&status);
1224:       MPI_Recv(column_values,cnt,MPIU_SCALAR,i,tag,PetscObjectComm((PetscObject)mat),&status);
1225:       PetscBinaryWrite(fd,column_values,cnt,PETSC_SCALAR,PETSC_TRUE);
1226:     }
1227:     PetscViewerFlowControlEndMaster(viewer,&message_count);
1228:   } else {
1229:     PetscViewerFlowControlStepWorker(viewer,rank,&message_count);
1230:     MPI_Send(&nz,1,MPIU_INT,0,tag,PetscObjectComm((PetscObject)mat));
1231:     MPI_Send(column_values,nz,MPIU_SCALAR,0,tag,PetscObjectComm((PetscObject)mat));
1232:     PetscViewerFlowControlEndWorker(viewer,&message_count);
1233:   }
1234:   PetscFree(column_values);

1236:   PetscViewerBinaryGetInfoPointer(viewer,&file);
1237:   if (file) {
1238:     fprintf(file,"-matload_block_size %d\n",(int)mat->rmap->bs);
1239:   }
1240:   return(0);
1241: }

1245: PetscErrorCode MatView_MPIBAIJ(Mat mat,PetscViewer viewer)
1246: {
1248:   PetscBool      iascii,isdraw,issocket,isbinary;

1251:   PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii);
1252:   PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERDRAW,&isdraw);
1253:   PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERSOCKET,&issocket);
1254:   PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERBINARY,&isbinary);
1255:   if (iascii || isdraw || issocket) {
1256:     MatView_MPIBAIJ_ASCIIorDraworSocket(mat,viewer);
1257:   } else if (isbinary) {
1258:     MatView_MPIBAIJ_Binary(mat,viewer);
1259:   }
1260:   return(0);
1261: }

1265: PetscErrorCode MatDestroy_MPIBAIJ(Mat mat)
1266: {
1267:   Mat_MPIBAIJ    *baij = (Mat_MPIBAIJ*)mat->data;

1271: #if defined(PETSC_USE_LOG)
1272:   PetscLogObjectState((PetscObject)mat,"Rows=%D,Cols=%D",mat->rmap->N,mat->cmap->N);
1273: #endif
1274:   MatStashDestroy_Private(&mat->stash);
1275:   MatStashDestroy_Private(&mat->bstash);
1276:   MatDestroy(&baij->A);
1277:   MatDestroy(&baij->B);
1278: #if defined(PETSC_USE_CTABLE)
1279:   PetscTableDestroy(&baij->colmap);
1280: #else
1281:   PetscFree(baij->colmap);
1282: #endif
1283:   PetscFree(baij->garray);
1284:   VecDestroy(&baij->lvec);
1285:   VecScatterDestroy(&baij->Mvctx);
1286:   PetscFree2(baij->rowvalues,baij->rowindices);
1287:   PetscFree(baij->barray);
1288:   PetscFree2(baij->hd,baij->ht);
1289:   PetscFree(baij->rangebs);
1290:   PetscFree(mat->data);

1292:   PetscObjectChangeTypeName((PetscObject)mat,0);
1293:   PetscObjectComposeFunction((PetscObject)mat,"MatStoreValues_C",NULL);
1294:   PetscObjectComposeFunction((PetscObject)mat,"MatRetrieveValues_C",NULL);
1295:   PetscObjectComposeFunction((PetscObject)mat,"MatGetDiagonalBlock_C",NULL);
1296:   PetscObjectComposeFunction((PetscObject)mat,"MatMPIBAIJSetPreallocation_C",NULL);
1297:   PetscObjectComposeFunction((PetscObject)mat,"MatMPIBAIJSetPreallocationCSR_C",NULL);
1298:   PetscObjectComposeFunction((PetscObject)mat,"MatDiagonalScaleLocal_C",NULL);
1299:   PetscObjectComposeFunction((PetscObject)mat,"MatSetHashTableFactor_C",NULL);
1300:   PetscObjectComposeFunction((PetscObject)mat,"MatConvert_mpibaij_mpisbaij_C",NULL);
1301:   PetscObjectComposeFunction((PetscObject)mat,"MatConvert_mpibaij_mpibstrm_C",NULL);
1302:   return(0);
1303: }

1307: PetscErrorCode MatMult_MPIBAIJ(Mat A,Vec xx,Vec yy)
1308: {
1309:   Mat_MPIBAIJ    *a = (Mat_MPIBAIJ*)A->data;
1311:   PetscInt       nt;

1314:   VecGetLocalSize(xx,&nt);
1315:   if (nt != A->cmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Incompatible partition of A and xx");
1316:   VecGetLocalSize(yy,&nt);
1317:   if (nt != A->rmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Incompatible parition of A and yy");
1318:   VecScatterBegin(a->Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);
1319:   (*a->A->ops->mult)(a->A,xx,yy);
1320:   VecScatterEnd(a->Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);
1321:   (*a->B->ops->multadd)(a->B,a->lvec,yy,yy);
1322:   return(0);
1323: }

1327: PetscErrorCode MatMultAdd_MPIBAIJ(Mat A,Vec xx,Vec yy,Vec zz)
1328: {
1329:   Mat_MPIBAIJ    *a = (Mat_MPIBAIJ*)A->data;

1333:   VecScatterBegin(a->Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);
1334:   (*a->A->ops->multadd)(a->A,xx,yy,zz);
1335:   VecScatterEnd(a->Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);
1336:   (*a->B->ops->multadd)(a->B,a->lvec,zz,zz);
1337:   return(0);
1338: }

1342: PetscErrorCode MatMultTranspose_MPIBAIJ(Mat A,Vec xx,Vec yy)
1343: {
1344:   Mat_MPIBAIJ    *a = (Mat_MPIBAIJ*)A->data;
1346:   PetscBool      merged;

1349:   VecScatterGetMerged(a->Mvctx,&merged);
1350:   /* do nondiagonal part */
1351:   (*a->B->ops->multtranspose)(a->B,xx,a->lvec);
1352:   if (!merged) {
1353:     /* send it on its way */
1354:     VecScatterBegin(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);
1355:     /* do local part */
1356:     (*a->A->ops->multtranspose)(a->A,xx,yy);
1357:     /* receive remote parts: note this assumes the values are not actually */
1358:     /* inserted in yy until the next line */
1359:     VecScatterEnd(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);
1360:   } else {
1361:     /* do local part */
1362:     (*a->A->ops->multtranspose)(a->A,xx,yy);
1363:     /* send it on its way */
1364:     VecScatterBegin(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);
1365:     /* values actually were received in the Begin() but we need to call this nop */
1366:     VecScatterEnd(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);
1367:   }
1368:   return(0);
1369: }

1373: PetscErrorCode MatMultTransposeAdd_MPIBAIJ(Mat A,Vec xx,Vec yy,Vec zz)
1374: {
1375:   Mat_MPIBAIJ    *a = (Mat_MPIBAIJ*)A->data;

1379:   /* do nondiagonal part */
1380:   (*a->B->ops->multtranspose)(a->B,xx,a->lvec);
1381:   /* send it on its way */
1382:   VecScatterBegin(a->Mvctx,a->lvec,zz,ADD_VALUES,SCATTER_REVERSE);
1383:   /* do local part */
1384:   (*a->A->ops->multtransposeadd)(a->A,xx,yy,zz);
1385:   /* receive remote parts: note this assumes the values are not actually */
1386:   /* inserted in yy until the next line, which is true for my implementation*/
1387:   /* but is not perhaps always true. */
1388:   VecScatterEnd(a->Mvctx,a->lvec,zz,ADD_VALUES,SCATTER_REVERSE);
1389:   return(0);
1390: }

1392: /*
1393:   This only works correctly for square matrices where the subblock A->A is the
1394:    diagonal block
1395: */
1398: PetscErrorCode MatGetDiagonal_MPIBAIJ(Mat A,Vec v)
1399: {
1400:   Mat_MPIBAIJ    *a = (Mat_MPIBAIJ*)A->data;

1404:   if (A->rmap->N != A->cmap->N) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Supports only square matrix where A->A is diag block");
1405:   MatGetDiagonal(a->A,v);
1406:   return(0);
1407: }

1411: PetscErrorCode MatScale_MPIBAIJ(Mat A,PetscScalar aa)
1412: {
1413:   Mat_MPIBAIJ    *a = (Mat_MPIBAIJ*)A->data;

1417:   MatScale(a->A,aa);
1418:   MatScale(a->B,aa);
1419:   return(0);
1420: }

1424: PetscErrorCode MatGetRow_MPIBAIJ(Mat matin,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v)
1425: {
1426:   Mat_MPIBAIJ    *mat = (Mat_MPIBAIJ*)matin->data;
1427:   PetscScalar    *vworkA,*vworkB,**pvA,**pvB,*v_p;
1429:   PetscInt       bs = matin->rmap->bs,bs2 = mat->bs2,i,*cworkA,*cworkB,**pcA,**pcB;
1430:   PetscInt       nztot,nzA,nzB,lrow,brstart = matin->rmap->rstart,brend = matin->rmap->rend;
1431:   PetscInt       *cmap,*idx_p,cstart = mat->cstartbs;

1434:   if (row < brstart || row >= brend) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Only local rows");
1435:   if (mat->getrowactive) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Already active");
1436:   mat->getrowactive = PETSC_TRUE;

1438:   if (!mat->rowvalues && (idx || v)) {
1439:     /*
1440:         allocate enough space to hold information from the longest row.
1441:     */
1442:     Mat_SeqBAIJ *Aa = (Mat_SeqBAIJ*)mat->A->data,*Ba = (Mat_SeqBAIJ*)mat->B->data;
1443:     PetscInt    max = 1,mbs = mat->mbs,tmp;
1444:     for (i=0; i<mbs; i++) {
1445:       tmp = Aa->i[i+1] - Aa->i[i] + Ba->i[i+1] - Ba->i[i];
1446:       if (max < tmp) max = tmp;
1447:     }
1448:     PetscMalloc2(max*bs2,&mat->rowvalues,max*bs2,&mat->rowindices);
1449:   }
1450:   lrow = row - brstart;

1452:   pvA = &vworkA; pcA = &cworkA; pvB = &vworkB; pcB = &cworkB;
1453:   if (!v)   {pvA = 0; pvB = 0;}
1454:   if (!idx) {pcA = 0; if (!v) pcB = 0;}
1455:   (*mat->A->ops->getrow)(mat->A,lrow,&nzA,pcA,pvA);
1456:   (*mat->B->ops->getrow)(mat->B,lrow,&nzB,pcB,pvB);
1457:   nztot = nzA + nzB;

1459:   cmap = mat->garray;
1460:   if (v  || idx) {
1461:     if (nztot) {
1462:       /* Sort by increasing column numbers, assuming A and B already sorted */
1463:       PetscInt imark = -1;
1464:       if (v) {
1465:         *v = v_p = mat->rowvalues;
1466:         for (i=0; i<nzB; i++) {
1467:           if (cmap[cworkB[i]/bs] < cstart) v_p[i] = vworkB[i];
1468:           else break;
1469:         }
1470:         imark = i;
1471:         for (i=0; i<nzA; i++)     v_p[imark+i] = vworkA[i];
1472:         for (i=imark; i<nzB; i++) v_p[nzA+i]   = vworkB[i];
1473:       }
1474:       if (idx) {
1475:         *idx = idx_p = mat->rowindices;
1476:         if (imark > -1) {
1477:           for (i=0; i<imark; i++) {
1478:             idx_p[i] = cmap[cworkB[i]/bs]*bs + cworkB[i]%bs;
1479:           }
1480:         } else {
1481:           for (i=0; i<nzB; i++) {
1482:             if (cmap[cworkB[i]/bs] < cstart) idx_p[i] = cmap[cworkB[i]/bs]*bs + cworkB[i]%bs;
1483:             else break;
1484:           }
1485:           imark = i;
1486:         }
1487:         for (i=0; i<nzA; i++)     idx_p[imark+i] = cstart*bs + cworkA[i];
1488:         for (i=imark; i<nzB; i++) idx_p[nzA+i]   = cmap[cworkB[i]/bs]*bs + cworkB[i]%bs ;
1489:       }
1490:     } else {
1491:       if (idx) *idx = 0;
1492:       if (v)   *v   = 0;
1493:     }
1494:   }
1495:   *nz  = nztot;
1496:   (*mat->A->ops->restorerow)(mat->A,lrow,&nzA,pcA,pvA);
1497:   (*mat->B->ops->restorerow)(mat->B,lrow,&nzB,pcB,pvB);
1498:   return(0);
1499: }

1503: PetscErrorCode MatRestoreRow_MPIBAIJ(Mat mat,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v)
1504: {
1505:   Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)mat->data;

1508:   if (!baij->getrowactive) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"MatGetRow not called");
1509:   baij->getrowactive = PETSC_FALSE;
1510:   return(0);
1511: }

1515: PetscErrorCode MatZeroEntries_MPIBAIJ(Mat A)
1516: {
1517:   Mat_MPIBAIJ    *l = (Mat_MPIBAIJ*)A->data;

1521:   MatZeroEntries(l->A);
1522:   MatZeroEntries(l->B);
1523:   return(0);
1524: }

1528: PetscErrorCode MatGetInfo_MPIBAIJ(Mat matin,MatInfoType flag,MatInfo *info)
1529: {
1530:   Mat_MPIBAIJ    *a = (Mat_MPIBAIJ*)matin->data;
1531:   Mat            A  = a->A,B = a->B;
1533:   PetscReal      isend[5],irecv[5];

1536:   info->block_size = (PetscReal)matin->rmap->bs;

1538:   MatGetInfo(A,MAT_LOCAL,info);

1540:   isend[0] = info->nz_used; isend[1] = info->nz_allocated; isend[2] = info->nz_unneeded;
1541:   isend[3] = info->memory;  isend[4] = info->mallocs;

1543:   MatGetInfo(B,MAT_LOCAL,info);

1545:   isend[0] += info->nz_used; isend[1] += info->nz_allocated; isend[2] += info->nz_unneeded;
1546:   isend[3] += info->memory;  isend[4] += info->mallocs;

1548:   if (flag == MAT_LOCAL) {
1549:     info->nz_used      = isend[0];
1550:     info->nz_allocated = isend[1];
1551:     info->nz_unneeded  = isend[2];
1552:     info->memory       = isend[3];
1553:     info->mallocs      = isend[4];
1554:   } else if (flag == MAT_GLOBAL_MAX) {
1555:     MPI_Allreduce(isend,irecv,5,MPIU_REAL,MPIU_MAX,PetscObjectComm((PetscObject)matin));

1557:     info->nz_used      = irecv[0];
1558:     info->nz_allocated = irecv[1];
1559:     info->nz_unneeded  = irecv[2];
1560:     info->memory       = irecv[3];
1561:     info->mallocs      = irecv[4];
1562:   } else if (flag == MAT_GLOBAL_SUM) {
1563:     MPI_Allreduce(isend,irecv,5,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)matin));

1565:     info->nz_used      = irecv[0];
1566:     info->nz_allocated = irecv[1];
1567:     info->nz_unneeded  = irecv[2];
1568:     info->memory       = irecv[3];
1569:     info->mallocs      = irecv[4];
1570:   } else SETERRQ1(PetscObjectComm((PetscObject)matin),PETSC_ERR_ARG_WRONG,"Unknown MatInfoType argument %d",(int)flag);
1571:   info->fill_ratio_given  = 0; /* no parallel LU/ILU/Cholesky */
1572:   info->fill_ratio_needed = 0;
1573:   info->factor_mallocs    = 0;
1574:   return(0);
1575: }

1579: PetscErrorCode MatSetOption_MPIBAIJ(Mat A,MatOption op,PetscBool flg)
1580: {
1581:   Mat_MPIBAIJ    *a = (Mat_MPIBAIJ*)A->data;

1585:   switch (op) {
1586:   case MAT_NEW_NONZERO_LOCATIONS:
1587:   case MAT_NEW_NONZERO_ALLOCATION_ERR:
1588:   case MAT_UNUSED_NONZERO_LOCATION_ERR:
1589:   case MAT_KEEP_NONZERO_PATTERN:
1590:   case MAT_NEW_NONZERO_LOCATION_ERR:
1591:     MatSetOption(a->A,op,flg);
1592:     MatSetOption(a->B,op,flg);
1593:     break;
1594:   case MAT_ROW_ORIENTED:
1595:     a->roworiented = flg;

1597:     MatSetOption(a->A,op,flg);
1598:     MatSetOption(a->B,op,flg);
1599:     break;
1600:   case MAT_NEW_DIAGONALS:
1601:     PetscInfo1(A,"Option %s ignored\n",MatOptions[op]);
1602:     break;
1603:   case MAT_IGNORE_OFF_PROC_ENTRIES:
1604:     a->donotstash = flg;
1605:     break;
1606:   case MAT_USE_HASH_TABLE:
1607:     a->ht_flag = flg;
1608:     break;
1609:   case MAT_SYMMETRIC:
1610:   case MAT_STRUCTURALLY_SYMMETRIC:
1611:   case MAT_HERMITIAN:
1612:   case MAT_SYMMETRY_ETERNAL:
1613:     MatSetOption(a->A,op,flg);
1614:     break;
1615:   default:
1616:     SETERRQ1(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"unknown option %d",op);
1617:   }
1618:   return(0);
1619: }

1623: PetscErrorCode MatTranspose_MPIBAIJ(Mat A,MatReuse reuse,Mat *matout)
1624: {
1625:   Mat_MPIBAIJ    *baij = (Mat_MPIBAIJ*)A->data;
1626:   Mat_SeqBAIJ    *Aloc;
1627:   Mat            B;
1629:   PetscInt       M =A->rmap->N,N=A->cmap->N,*ai,*aj,i,*rvals,j,k,col;
1630:   PetscInt       bs=A->rmap->bs,mbs=baij->mbs;
1631:   MatScalar      *a;

1634:   if (reuse == MAT_REUSE_MATRIX && A == *matout && M != N) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_SIZ,"Square matrix only for in-place");
1635:   if (reuse == MAT_INITIAL_MATRIX || *matout == A) {
1636:     MatCreate(PetscObjectComm((PetscObject)A),&B);
1637:     MatSetSizes(B,A->cmap->n,A->rmap->n,N,M);
1638:     MatSetType(B,((PetscObject)A)->type_name);
1639:     /* Do not know preallocation information, but must set block size */
1640:     MatMPIBAIJSetPreallocation(B,A->rmap->bs,PETSC_DECIDE,NULL,PETSC_DECIDE,NULL);
1641:   } else {
1642:     B = *matout;
1643:   }

1645:   /* copy over the A part */
1646:   Aloc = (Mat_SeqBAIJ*)baij->A->data;
1647:   ai   = Aloc->i; aj = Aloc->j; a = Aloc->a;
1648:   PetscMalloc1(bs,&rvals);

1650:   for (i=0; i<mbs; i++) {
1651:     rvals[0] = bs*(baij->rstartbs + i);
1652:     for (j=1; j<bs; j++) rvals[j] = rvals[j-1] + 1;
1653:     for (j=ai[i]; j<ai[i+1]; j++) {
1654:       col = (baij->cstartbs+aj[j])*bs;
1655:       for (k=0; k<bs; k++) {
1656:         MatSetValues_MPIBAIJ(B,1,&col,bs,rvals,a,INSERT_VALUES);

1658:         col++; a += bs;
1659:       }
1660:     }
1661:   }
1662:   /* copy over the B part */
1663:   Aloc = (Mat_SeqBAIJ*)baij->B->data;
1664:   ai   = Aloc->i; aj = Aloc->j; a = Aloc->a;
1665:   for (i=0; i<mbs; i++) {
1666:     rvals[0] = bs*(baij->rstartbs + i);
1667:     for (j=1; j<bs; j++) rvals[j] = rvals[j-1] + 1;
1668:     for (j=ai[i]; j<ai[i+1]; j++) {
1669:       col = baij->garray[aj[j]]*bs;
1670:       for (k=0; k<bs; k++) {
1671:         MatSetValues_MPIBAIJ(B,1,&col,bs,rvals,a,INSERT_VALUES);
1672:         col++;
1673:         a += bs;
1674:       }
1675:     }
1676:   }
1677:   PetscFree(rvals);
1678:   MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);
1679:   MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);

1681:   if (reuse == MAT_INITIAL_MATRIX || *matout != A) *matout = B;
1682:   else {
1683:     MatHeaderMerge(A,B);
1684:   }
1685:   return(0);
1686: }

1690: PetscErrorCode MatDiagonalScale_MPIBAIJ(Mat mat,Vec ll,Vec rr)
1691: {
1692:   Mat_MPIBAIJ    *baij = (Mat_MPIBAIJ*)mat->data;
1693:   Mat            a     = baij->A,b = baij->B;
1695:   PetscInt       s1,s2,s3;

1698:   MatGetLocalSize(mat,&s2,&s3);
1699:   if (rr) {
1700:     VecGetLocalSize(rr,&s1);
1701:     if (s1!=s3) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"right vector non-conforming local size");
1702:     /* Overlap communication with computation. */
1703:     VecScatterBegin(baij->Mvctx,rr,baij->lvec,INSERT_VALUES,SCATTER_FORWARD);
1704:   }
1705:   if (ll) {
1706:     VecGetLocalSize(ll,&s1);
1707:     if (s1!=s2) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"left vector non-conforming local size");
1708:     (*b->ops->diagonalscale)(b,ll,NULL);
1709:   }
1710:   /* scale  the diagonal block */
1711:   (*a->ops->diagonalscale)(a,ll,rr);

1713:   if (rr) {
1714:     /* Do a scatter end and then right scale the off-diagonal block */
1715:     VecScatterEnd(baij->Mvctx,rr,baij->lvec,INSERT_VALUES,SCATTER_FORWARD);
1716:     (*b->ops->diagonalscale)(b,NULL,baij->lvec);
1717:   }
1718:   return(0);
1719: }

1723: PetscErrorCode MatZeroRows_MPIBAIJ(Mat A,PetscInt N,const PetscInt rows[],PetscScalar diag,Vec x,Vec b)
1724: {
1725:   Mat_MPIBAIJ   *l      = (Mat_MPIBAIJ *) A->data;
1726:   PetscInt      *owners = A->rmap->range;
1727:   PetscInt       n      = A->rmap->n;
1728:   PetscMPIInt    size   = l->size;
1729:   PetscSF        sf;
1730:   PetscInt      *lrows;
1731:   PetscSFNode   *rrows;
1732:   PetscInt       lastidx = -1, r, p = 0, len = 0;

1736:   /* Create SF where leaves are input rows and roots are owned rows */
1737:   PetscMalloc1(n, &lrows);
1738:   for (r = 0; r < n; ++r) lrows[r] = -1;
1739:   PetscMalloc1(N, &rrows);
1740:   for (r = 0; r < N; ++r) {
1741:     const PetscInt idx   = rows[r];
1742:     PetscBool      found = PETSC_FALSE;
1743:     /* Trick for efficient searching for sorted rows */
1744:     if (lastidx > idx) p = 0;
1745:     lastidx = idx;
1746:     for (; p < size; ++p) {
1747:       if (idx >= owners[p] && idx < owners[p+1]) {
1748:         rrows[r].rank  = p;
1749:         rrows[r].index = rows[r] - owners[p];
1750:         found = PETSC_TRUE;
1751:         break;
1752:       }
1753:     }
1754:     if (!found) SETERRQ1(PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Row %d not found in matrix distribution", idx);
1755:   }
1756:   PetscSFCreate(PetscObjectComm((PetscObject) A), &sf);
1757:   PetscSFSetGraph(sf, n, N, NULL, PETSC_OWN_POINTER, rrows, PETSC_OWN_POINTER);
1758:   /* Collect flags for rows to be zeroed */
1759:   PetscSFReduceBegin(sf, MPIU_INT, (PetscInt *) rows, lrows, MPI_LOR);
1760:   PetscSFReduceEnd(sf, MPIU_INT, (PetscInt *) rows, lrows, MPI_LOR);
1761:   PetscSFDestroy(&sf);
1762:   /* Compress and put in row numbers */
1763:   for (r = 0; r < n; ++r) if (lrows[r] >= 0) lrows[len++] = r;
1764:   /* fix right hand side if needed */
1765:   if (x && b) {
1766:     const PetscScalar *xx;
1767:     PetscScalar       *bb;

1769:     VecGetArrayRead(x,&xx);
1770:     VecGetArray(b,&bb);
1771:     for (r = 0; r < len; ++r) bb[lrows[r]] = diag*xx[lrows[r]];
1772:     VecRestoreArrayRead(x,&xx);
1773:     VecRestoreArray(b,&bb);
1774:   }

1776:   /* actually zap the local rows */
1777:   /*
1778:         Zero the required rows. If the "diagonal block" of the matrix
1779:      is square and the user wishes to set the diagonal we use separate
1780:      code so that MatSetValues() is not called for each diagonal allocating
1781:      new memory, thus calling lots of mallocs and slowing things down.

1783:   */
1784:   /* must zero l->B before l->A because the (diag) case below may put values into l->B*/
1785:   MatZeroRows_SeqBAIJ(l->B,len,lrows,0.0,0,0);
1786:   if ((diag != 0.0) && (l->A->rmap->N == l->A->cmap->N)) {
1787:     MatZeroRows_SeqBAIJ(l->A,len,lrows,diag,0,0);
1788:   } else if (diag != 0.0) {
1789:     MatZeroRows_SeqBAIJ(l->A,len,lrows,0.0,0,0);
1790:     if (((Mat_SeqBAIJ*)l->A->data)->nonew) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"MatZeroRows() on rectangular matrices cannot be used with the Mat options \n\
1791:        MAT_NEW_NONZERO_LOCATIONS,MAT_NEW_NONZERO_LOCATION_ERR,MAT_NEW_NONZERO_ALLOCATION_ERR");
1792:     for (r = 0; r < len; ++r) {
1793:       const PetscInt row = lrows[r] + A->rmap->rstart;
1794:       MatSetValues(A,1,&row,1,&row,&diag,INSERT_VALUES);
1795:     }
1796:     MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);
1797:     MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);
1798:   } else {
1799:     MatZeroRows_SeqBAIJ(l->A,len,lrows,0.0,0,0);
1800:   }
1801:   PetscFree(lrows);
1802:   return(0);
1803: }

1807: PetscErrorCode MatZeroRowsColumns_MPIBAIJ(Mat A,PetscInt N,const PetscInt rows[],PetscScalar diag,Vec x,Vec b)
1808: {
1809:   Mat_MPIBAIJ       *l = (Mat_MPIBAIJ*)A->data;
1810:   PetscErrorCode    ierr;
1811:   PetscMPIInt       size = l->size,n = A->rmap->n,lastidx = -1;
1812:   PetscInt          i,j,k,r,p = 0,len = 0,row,col,count;
1813:   PetscInt          *lrows,*owners = A->rmap->range;
1814:   PetscSFNode       *rrows;
1815:   PetscSF           sf;
1816:   const PetscScalar *xx;
1817:   PetscScalar       *bb,*mask;
1818:   Vec               xmask,lmask;
1819:   Mat_SeqBAIJ       *baij = (Mat_SeqBAIJ*)l->B->data;
1820:   PetscInt           bs = A->rmap->bs, bs2 = baij->bs2;
1821:   PetscScalar       *aa;
1822: #if defined(PETSC_DEBUG)
1823:   PetscBool found = PETSC_FALSE;
1824: #endif

1827:   /* Create SF where leaves are input rows and roots are owned rows */
1828:   PetscMalloc1(n, &lrows);
1829:   for (r = 0; r < n; ++r) lrows[r] = -1;
1830:   PetscMalloc1(N, &rrows);
1831:   for (r = 0; r < N; ++r) {
1832:     const PetscInt idx   = rows[r];
1833:     PetscBool      found = PETSC_FALSE;
1834:     /* Trick for efficient searching for sorted rows */
1835:     if (lastidx > idx) p = 0;
1836:     lastidx = idx;
1837:     for (; p < size; ++p) {
1838:       if (idx >= owners[p] && idx < owners[p+1]) {
1839:         rrows[r].rank  = p;
1840:         rrows[r].index = rows[r] - owners[p];
1841:         found = PETSC_TRUE;
1842:         break;
1843:       }
1844:     }
1845:     if (!found) SETERRQ1(PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Row %d not found in matrix distribution", idx);
1846:   }
1847:   PetscSFCreate(PetscObjectComm((PetscObject) A), &sf);
1848:   PetscSFSetGraph(sf, n, N, NULL, PETSC_OWN_POINTER, rrows, PETSC_OWN_POINTER);
1849:   /* Collect flags for rows to be zeroed */
1850:   PetscSFReduceBegin(sf, MPIU_INT, (PetscInt *) rows, lrows, MPI_LOR);
1851:   PetscSFReduceEnd(sf, MPIU_INT, (PetscInt *) rows, lrows, MPI_LOR);
1852:   PetscSFDestroy(&sf);
1853:   /* Compress and put in row numbers */
1854:   for (r = 0; r < n; ++r) if (lrows[r] >= 0) lrows[len++] = r;
1855:   /* zero diagonal part of matrix */
1856:   MatZeroRowsColumns(l->A,len,lrows,diag,x,b);
1857:   /* handle off diagonal part of matrix */
1858:   MatGetVecs(A,&xmask,NULL);
1859:   VecDuplicate(l->lvec,&lmask);
1860:   VecGetArray(xmask,&bb);
1861:   for (i=0; i<len; i++) bb[lrows[i]] = 1;
1862:   VecRestoreArray(xmask,&bb);
1863:   VecScatterBegin(l->Mvctx,xmask,lmask,ADD_VALUES,SCATTER_FORWARD);
1864:   VecScatterEnd(l->Mvctx,xmask,lmask,ADD_VALUES,SCATTER_FORWARD);
1865:   VecDestroy(&xmask);
1866:   if (x) {
1867:     VecScatterBegin(l->Mvctx,x,l->lvec,INSERT_VALUES,SCATTER_FORWARD);
1868:     VecScatterEnd(l->Mvctx,x,l->lvec,INSERT_VALUES,SCATTER_FORWARD);
1869:     VecGetArrayRead(l->lvec,&xx);
1870:     VecGetArray(b,&bb);
1871:   }
1872:   VecGetArray(lmask,&mask);
1873:   /* remove zeroed rows of off diagonal matrix */
1874:   for (i = 0; i < len; ++i) {
1875:     row   = lrows[i];
1876:     count = (baij->i[row/bs +1] - baij->i[row/bs])*bs;
1877:     aa    = ((MatScalar*)(baij->a)) + baij->i[row/bs]*bs2 + (row%bs);
1878:     for (k = 0; k < count; ++k) {
1879:       aa[0] = 0.0;
1880:       aa   += bs;
1881:     }
1882:   }
1883:   /* loop over all elements of off process part of matrix zeroing removed columns*/
1884:   for (i = 0; i < l->B->rmap->N; ++i) {
1885:     row = i/bs;
1886:     for (j = baij->i[row]; j < baij->i[row+1]; ++j) {
1887:       for (k = 0; k < bs; ++k) {
1888:         col = bs*baij->j[j] + k;
1889:         if (PetscAbsScalar(mask[col])) {
1890:           aa = ((MatScalar*)(baij->a)) + j*bs2 + (i%bs) + bs*k;
1891:           if (b) bb[i] -= aa[0]*xx[col];
1892:           aa[0] = 0.0;
1893:         }
1894:       }
1895:     }
1896:   }
1897:   if (x) {
1898:     VecRestoreArray(b,&bb);
1899:     VecRestoreArrayRead(l->lvec,&xx);
1900:   }
1901:   VecRestoreArray(lmask,&mask);
1902:   VecDestroy(&lmask);
1903:   PetscFree(lrows);
1904:   return(0);
1905: }

1909: PetscErrorCode MatSetUnfactored_MPIBAIJ(Mat A)
1910: {
1911:   Mat_MPIBAIJ    *a = (Mat_MPIBAIJ*)A->data;

1915:   MatSetUnfactored(a->A);
1916:   return(0);
1917: }

1919: static PetscErrorCode MatDuplicate_MPIBAIJ(Mat,MatDuplicateOption,Mat*);

1923: PetscErrorCode MatEqual_MPIBAIJ(Mat A,Mat B,PetscBool  *flag)
1924: {
1925:   Mat_MPIBAIJ    *matB = (Mat_MPIBAIJ*)B->data,*matA = (Mat_MPIBAIJ*)A->data;
1926:   Mat            a,b,c,d;
1927:   PetscBool      flg;

1931:   a = matA->A; b = matA->B;
1932:   c = matB->A; d = matB->B;

1934:   MatEqual(a,c,&flg);
1935:   if (flg) {
1936:     MatEqual(b,d,&flg);
1937:   }
1938:   MPI_Allreduce(&flg,flag,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)A));
1939:   return(0);
1940: }

1944: PetscErrorCode MatCopy_MPIBAIJ(Mat A,Mat B,MatStructure str)
1945: {
1947:   Mat_MPIBAIJ    *a = (Mat_MPIBAIJ*)A->data;
1948:   Mat_MPIBAIJ    *b = (Mat_MPIBAIJ*)B->data;

1951:   /* If the two matrices don't have the same copy implementation, they aren't compatible for fast copy. */
1952:   if ((str != SAME_NONZERO_PATTERN) || (A->ops->copy != B->ops->copy)) {
1953:     MatCopy_Basic(A,B,str);
1954:   } else {
1955:     MatCopy(a->A,b->A,str);
1956:     MatCopy(a->B,b->B,str);
1957:   }
1958:   return(0);
1959: }

1963: PetscErrorCode MatSetUp_MPIBAIJ(Mat A)
1964: {

1968:    MatMPIBAIJSetPreallocation(A,A->rmap->bs,PETSC_DEFAULT,0,PETSC_DEFAULT,0);
1969:   return(0);
1970: }

1974: PetscErrorCode MatAXPY_MPIBAIJ(Mat Y,PetscScalar a,Mat X,MatStructure str)
1975: {
1977:   Mat_MPIBAIJ    *xx=(Mat_MPIBAIJ*)X->data,*yy=(Mat_MPIBAIJ*)Y->data;
1978:   PetscBLASInt   bnz,one=1;
1979:   Mat_SeqBAIJ    *x,*y;

1982:   if (str == SAME_NONZERO_PATTERN) {
1983:     PetscScalar alpha = a;
1984:     x    = (Mat_SeqBAIJ*)xx->A->data;
1985:     y    = (Mat_SeqBAIJ*)yy->A->data;
1986:     PetscBLASIntCast(x->nz,&bnz);
1987:     PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&bnz,&alpha,x->a,&one,y->a,&one));
1988:     x    = (Mat_SeqBAIJ*)xx->B->data;
1989:     y    = (Mat_SeqBAIJ*)yy->B->data;
1990:     PetscBLASIntCast(x->nz,&bnz);
1991:     PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&bnz,&alpha,x->a,&one,y->a,&one));
1992:   } else {
1993:     MatAXPY_Basic(Y,a,X,str);
1994:   }
1995:   return(0);
1996: }

2000: PetscErrorCode MatRealPart_MPIBAIJ(Mat A)
2001: {
2002:   Mat_MPIBAIJ    *a = (Mat_MPIBAIJ*)A->data;

2006:   MatRealPart(a->A);
2007:   MatRealPart(a->B);
2008:   return(0);
2009: }

2013: PetscErrorCode MatImaginaryPart_MPIBAIJ(Mat A)
2014: {
2015:   Mat_MPIBAIJ    *a = (Mat_MPIBAIJ*)A->data;

2019:   MatImaginaryPart(a->A);
2020:   MatImaginaryPart(a->B);
2021:   return(0);
2022: }

2026: PetscErrorCode MatGetSubMatrix_MPIBAIJ(Mat mat,IS isrow,IS iscol,MatReuse call,Mat *newmat)
2027: {
2029:   IS             iscol_local;
2030:   PetscInt       csize;

2033:   ISGetLocalSize(iscol,&csize);
2034:   if (call == MAT_REUSE_MATRIX) {
2035:     PetscObjectQuery((PetscObject)*newmat,"ISAllGather",(PetscObject*)&iscol_local);
2036:     if (!iscol_local) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Submatrix passed in was not used before, cannot reuse");
2037:   } else {
2038:     ISAllGather(iscol,&iscol_local);
2039:   }
2040:   MatGetSubMatrix_MPIBAIJ_Private(mat,isrow,iscol_local,csize,call,newmat);
2041:   if (call == MAT_INITIAL_MATRIX) {
2042:     PetscObjectCompose((PetscObject)*newmat,"ISAllGather",(PetscObject)iscol_local);
2043:     ISDestroy(&iscol_local);
2044:   }
2045:   return(0);
2046: }
2047: extern PetscErrorCode MatGetSubMatrices_MPIBAIJ_local(Mat,PetscInt,const IS[],const IS[],MatReuse,PetscBool*,PetscBool*,Mat*);
2050: /*
2051:   Not great since it makes two copies of the submatrix, first an SeqBAIJ
2052:   in local and then by concatenating the local matrices the end result.
2053:   Writing it directly would be much like MatGetSubMatrices_MPIBAIJ()
2054: */
2055: PetscErrorCode MatGetSubMatrix_MPIBAIJ_Private(Mat mat,IS isrow,IS iscol,PetscInt csize,MatReuse call,Mat *newmat)
2056: {
2058:   PetscMPIInt    rank,size;
2059:   PetscInt       i,m,n,rstart,row,rend,nz,*cwork,j,bs;
2060:   PetscInt       *ii,*jj,nlocal,*dlens,*olens,dlen,olen,jend,mglobal,ncol,nrow;
2061:   Mat            M,Mreuse;
2062:   MatScalar      *vwork,*aa;
2063:   MPI_Comm       comm;
2064:   IS             isrow_new, iscol_new;
2065:   PetscBool      idflag,allrows, allcols;
2066:   Mat_SeqBAIJ    *aij;

2069:   PetscObjectGetComm((PetscObject)mat,&comm);
2070:   MPI_Comm_rank(comm,&rank);
2071:   MPI_Comm_size(comm,&size);
2072:   /* The compression and expansion should be avoided. Doesn't point
2073:      out errors, might change the indices, hence buggey */
2074:   ISCompressIndicesGeneral(mat->rmap->N,mat->rmap->n,mat->rmap->bs,1,&isrow,&isrow_new);
2075:   ISCompressIndicesGeneral(mat->cmap->N,mat->cmap->n,mat->cmap->bs,1,&iscol,&iscol_new);

2077:   /* Check for special case: each processor gets entire matrix columns */
2078:   ISIdentity(iscol,&idflag);
2079:   ISGetLocalSize(iscol,&ncol);
2080:   if (idflag && ncol == mat->cmap->N) allcols = PETSC_TRUE;
2081:   else allcols = PETSC_FALSE;

2083:   ISIdentity(isrow,&idflag);
2084:   ISGetLocalSize(isrow,&nrow);
2085:   if (idflag && nrow == mat->rmap->N) allrows = PETSC_TRUE;
2086:   else allrows = PETSC_FALSE;

2088:   if (call ==  MAT_REUSE_MATRIX) {
2089:     PetscObjectQuery((PetscObject)*newmat,"SubMatrix",(PetscObject*)&Mreuse);
2090:     if (!Mreuse) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Submatrix passed in was not used before, cannot reuse");
2091:     MatGetSubMatrices_MPIBAIJ_local(mat,1,&isrow_new,&iscol_new,MAT_REUSE_MATRIX,&allrows,&allcols,&Mreuse);
2092:   } else {
2093:     MatGetSubMatrices_MPIBAIJ_local(mat,1,&isrow_new,&iscol_new,MAT_INITIAL_MATRIX,&allrows,&allcols,&Mreuse);
2094:   }
2095:   ISDestroy(&isrow_new);
2096:   ISDestroy(&iscol_new);
2097:   /*
2098:       m - number of local rows
2099:       n - number of columns (same on all processors)
2100:       rstart - first row in new global matrix generated
2101:   */
2102:   MatGetBlockSize(mat,&bs);
2103:   MatGetSize(Mreuse,&m,&n);
2104:   m    = m/bs;
2105:   n    = n/bs;

2107:   if (call == MAT_INITIAL_MATRIX) {
2108:     aij = (Mat_SeqBAIJ*)(Mreuse)->data;
2109:     ii  = aij->i;
2110:     jj  = aij->j;

2112:     /*
2113:         Determine the number of non-zeros in the diagonal and off-diagonal
2114:         portions of the matrix in order to do correct preallocation
2115:     */

2117:     /* first get start and end of "diagonal" columns */
2118:     if (csize == PETSC_DECIDE) {
2119:       ISGetSize(isrow,&mglobal);
2120:       if (mglobal == n*bs) { /* square matrix */
2121:         nlocal = m;
2122:       } else {
2123:         nlocal = n/size + ((n % size) > rank);
2124:       }
2125:     } else {
2126:       nlocal = csize/bs;
2127:     }
2128:     MPI_Scan(&nlocal,&rend,1,MPIU_INT,MPI_SUM,comm);
2129:     rstart = rend - nlocal;
2130:     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);

2132:     /* next, compute all the lengths */
2133:     PetscMalloc2(m+1,&dlens,m+1,&olens);
2134:     for (i=0; i<m; i++) {
2135:       jend = ii[i+1] - ii[i];
2136:       olen = 0;
2137:       dlen = 0;
2138:       for (j=0; j<jend; j++) {
2139:         if (*jj < rstart || *jj >= rend) olen++;
2140:         else dlen++;
2141:         jj++;
2142:       }
2143:       olens[i] = olen;
2144:       dlens[i] = dlen;
2145:     }
2146:     MatCreate(comm,&M);
2147:     MatSetSizes(M,bs*m,bs*nlocal,PETSC_DECIDE,bs*n);
2148:     MatSetType(M,((PetscObject)mat)->type_name);
2149:     MatMPIBAIJSetPreallocation(M,bs,0,dlens,0,olens);
2150:     PetscFree2(dlens,olens);
2151:   } else {
2152:     PetscInt ml,nl;

2154:     M    = *newmat;
2155:     MatGetLocalSize(M,&ml,&nl);
2156:     if (ml != m) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Previous matrix must be same size/layout as request");
2157:     MatZeroEntries(M);
2158:     /*
2159:          The next two lines are needed so we may call MatSetValues_MPIAIJ() below directly,
2160:        rather than the slower MatSetValues().
2161:     */
2162:     M->was_assembled = PETSC_TRUE;
2163:     M->assembled     = PETSC_FALSE;
2164:   }
2165:   MatSetOption(M,MAT_ROW_ORIENTED,PETSC_FALSE);
2166:   MatGetOwnershipRange(M,&rstart,&rend);
2167:   aij  = (Mat_SeqBAIJ*)(Mreuse)->data;
2168:   ii   = aij->i;
2169:   jj   = aij->j;
2170:   aa   = aij->a;
2171:   for (i=0; i<m; i++) {
2172:     row   = rstart/bs + i;
2173:     nz    = ii[i+1] - ii[i];
2174:     cwork = jj;     jj += nz;
2175:     vwork = aa;     aa += nz*bs*bs;
2176:     MatSetValuesBlocked_MPIBAIJ(M,1,&row,nz,cwork,vwork,INSERT_VALUES);
2177:   }

2179:   MatAssemblyBegin(M,MAT_FINAL_ASSEMBLY);
2180:   MatAssemblyEnd(M,MAT_FINAL_ASSEMBLY);
2181:   *newmat = M;

2183:   /* save submatrix used in processor for next request */
2184:   if (call ==  MAT_INITIAL_MATRIX) {
2185:     PetscObjectCompose((PetscObject)M,"SubMatrix",(PetscObject)Mreuse);
2186:     PetscObjectDereference((PetscObject)Mreuse);
2187:   }
2188:   return(0);
2189: }

2193: PetscErrorCode MatPermute_MPIBAIJ(Mat A,IS rowp,IS colp,Mat *B)
2194: {
2195:   MPI_Comm       comm,pcomm;
2196:   PetscInt       clocal_size,nrows;
2197:   const PetscInt *rows;
2198:   PetscMPIInt    size;
2199:   IS             crowp,lcolp;

2203:   PetscObjectGetComm((PetscObject)A,&comm);
2204:   /* make a collective version of 'rowp' */
2205:   PetscObjectGetComm((PetscObject)rowp,&pcomm);
2206:   if (pcomm==comm) {
2207:     crowp = rowp;
2208:   } else {
2209:     ISGetSize(rowp,&nrows);
2210:     ISGetIndices(rowp,&rows);
2211:     ISCreateGeneral(comm,nrows,rows,PETSC_COPY_VALUES,&crowp);
2212:     ISRestoreIndices(rowp,&rows);
2213:   }
2214:   ISSetPermutation(crowp);
2215:   /* make a local version of 'colp' */
2216:   PetscObjectGetComm((PetscObject)colp,&pcomm);
2217:   MPI_Comm_size(pcomm,&size);
2218:   if (size==1) {
2219:     lcolp = colp;
2220:   } else {
2221:     ISAllGather(colp,&lcolp);
2222:   }
2223:   ISSetPermutation(lcolp);
2224:   /* now we just get the submatrix */
2225:   MatGetLocalSize(A,PETSC_NULL,&clocal_size);
2226:   MatGetSubMatrix_MPIBAIJ_Private(A,crowp,lcolp,clocal_size,MAT_INITIAL_MATRIX,B);
2227:   /* clean up */
2228:   if (pcomm!=comm) {
2229:     ISDestroy(&crowp);
2230:   }
2231:   if (size>1) {
2232:     ISDestroy(&lcolp);
2233:   }
2234:   return(0);
2235: }

2239: PetscErrorCode  MatGetGhosts_MPIBAIJ(Mat mat,PetscInt *nghosts,const PetscInt *ghosts[])
2240: {
2241:   Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*) mat->data;
2242:   Mat_SeqBAIJ *B    = (Mat_SeqBAIJ*)baij->B->data;

2245:   if (nghosts) *nghosts = B->nbs;
2246:   if (ghosts) *ghosts = baij->garray;
2247:   return(0);
2248: }

2252: PetscErrorCode MatGetSeqNonzeroStructure_MPIBAIJ(Mat A,Mat *newmat)
2253: {
2254:   Mat            B;
2255:   Mat_MPIBAIJ    *a  = (Mat_MPIBAIJ*)A->data;
2256:   Mat_SeqBAIJ    *ad = (Mat_SeqBAIJ*)a->A->data,*bd = (Mat_SeqBAIJ*)a->B->data;
2257:   Mat_SeqAIJ     *b;
2259:   PetscMPIInt    size,rank,*recvcounts = 0,*displs = 0;
2260:   PetscInt       sendcount,i,*rstarts = A->rmap->range,n,cnt,j,bs = A->rmap->bs;
2261:   PetscInt       m,*garray = a->garray,*lens,*jsendbuf,*a_jsendbuf,*b_jsendbuf;

2264:   MPI_Comm_size(PetscObjectComm((PetscObject)A),&size);
2265:   MPI_Comm_rank(PetscObjectComm((PetscObject)A),&rank);

2267:   /* ----------------------------------------------------------------
2268:      Tell every processor the number of nonzeros per row
2269:   */
2270:   PetscMalloc1((A->rmap->N/bs),&lens);
2271:   for (i=A->rmap->rstart/bs; i<A->rmap->rend/bs; i++) {
2272:     lens[i] = ad->i[i-A->rmap->rstart/bs+1] - ad->i[i-A->rmap->rstart/bs] + bd->i[i-A->rmap->rstart/bs+1] - bd->i[i-A->rmap->rstart/bs];
2273:   }
2274:   sendcount = A->rmap->rend/bs - A->rmap->rstart/bs;
2275:   PetscMalloc1(2*size,&recvcounts);
2276:   displs    = recvcounts + size;
2277:   for (i=0; i<size; i++) {
2278:     recvcounts[i] = A->rmap->range[i+1]/bs - A->rmap->range[i]/bs;
2279:     displs[i]     = A->rmap->range[i]/bs;
2280:   }
2281: #if defined(PETSC_HAVE_MPI_IN_PLACE)
2282:   MPI_Allgatherv(MPI_IN_PLACE,0,MPI_DATATYPE_NULL,lens,recvcounts,displs,MPIU_INT,PetscObjectComm((PetscObject)A));
2283: #else
2284:   MPI_Allgatherv(lens+A->rmap->rstart/bs,sendcount,MPIU_INT,lens,recvcounts,displs,MPIU_INT,PetscObjectComm((PetscObject)A));
2285: #endif
2286:   /* ---------------------------------------------------------------
2287:      Create the sequential matrix of the same type as the local block diagonal
2288:   */
2289:   MatCreate(PETSC_COMM_SELF,&B);
2290:   MatSetSizes(B,A->rmap->N/bs,A->cmap->N/bs,PETSC_DETERMINE,PETSC_DETERMINE);
2291:   MatSetType(B,MATSEQAIJ);
2292:   MatSeqAIJSetPreallocation(B,0,lens);
2293:   b    = (Mat_SeqAIJ*)B->data;

2295:   /*--------------------------------------------------------------------
2296:     Copy my part of matrix column indices over
2297:   */
2298:   sendcount  = ad->nz + bd->nz;
2299:   jsendbuf   = b->j + b->i[rstarts[rank]/bs];
2300:   a_jsendbuf = ad->j;
2301:   b_jsendbuf = bd->j;
2302:   n          = A->rmap->rend/bs - A->rmap->rstart/bs;
2303:   cnt        = 0;
2304:   for (i=0; i<n; i++) {

2306:     /* put in lower diagonal portion */
2307:     m = bd->i[i+1] - bd->i[i];
2308:     while (m > 0) {
2309:       /* is it above diagonal (in bd (compressed) numbering) */
2310:       if (garray[*b_jsendbuf] > A->rmap->rstart/bs + i) break;
2311:       jsendbuf[cnt++] = garray[*b_jsendbuf++];
2312:       m--;
2313:     }

2315:     /* put in diagonal portion */
2316:     for (j=ad->i[i]; j<ad->i[i+1]; j++) {
2317:       jsendbuf[cnt++] = A->rmap->rstart/bs + *a_jsendbuf++;
2318:     }

2320:     /* put in upper diagonal portion */
2321:     while (m-- > 0) {
2322:       jsendbuf[cnt++] = garray[*b_jsendbuf++];
2323:     }
2324:   }
2325:   if (cnt != sendcount) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Corrupted PETSc matrix: nz given %D actual nz %D",sendcount,cnt);

2327:   /*--------------------------------------------------------------------
2328:     Gather all column indices to all processors
2329:   */
2330:   for (i=0; i<size; i++) {
2331:     recvcounts[i] = 0;
2332:     for (j=A->rmap->range[i]/bs; j<A->rmap->range[i+1]/bs; j++) {
2333:       recvcounts[i] += lens[j];
2334:     }
2335:   }
2336:   displs[0] = 0;
2337:   for (i=1; i<size; i++) {
2338:     displs[i] = displs[i-1] + recvcounts[i-1];
2339:   }
2340: #if defined(PETSC_HAVE_MPI_IN_PLACE)
2341:   MPI_Allgatherv(MPI_IN_PLACE,0,MPI_DATATYPE_NULL,b->j,recvcounts,displs,MPIU_INT,PetscObjectComm((PetscObject)A));
2342: #else
2343:   MPI_Allgatherv(jsendbuf,sendcount,MPIU_INT,b->j,recvcounts,displs,MPIU_INT,PetscObjectComm((PetscObject)A));
2344: #endif
2345:   /*--------------------------------------------------------------------
2346:     Assemble the matrix into useable form (note numerical values not yet set)
2347:   */
2348:   /* set the b->ilen (length of each row) values */
2349:   PetscMemcpy(b->ilen,lens,(A->rmap->N/bs)*sizeof(PetscInt));
2350:   /* set the b->i indices */
2351:   b->i[0] = 0;
2352:   for (i=1; i<=A->rmap->N/bs; i++) {
2353:     b->i[i] = b->i[i-1] + lens[i-1];
2354:   }
2355:   PetscFree(lens);
2356:   MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);
2357:   MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);
2358:   PetscFree(recvcounts);

2360:   if (A->symmetric) {
2361:     MatSetOption(B,MAT_SYMMETRIC,PETSC_TRUE);
2362:   } else if (A->hermitian) {
2363:     MatSetOption(B,MAT_HERMITIAN,PETSC_TRUE);
2364:   } else if (A->structurally_symmetric) {
2365:     MatSetOption(B,MAT_STRUCTURALLY_SYMMETRIC,PETSC_TRUE);
2366:   }
2367:   *newmat = B;
2368:   return(0);
2369: }

2373: PetscErrorCode MatSOR_MPIBAIJ(Mat matin,Vec bb,PetscReal omega,MatSORType flag,PetscReal fshift,PetscInt its,PetscInt lits,Vec xx)
2374: {
2375:   Mat_MPIBAIJ    *mat = (Mat_MPIBAIJ*)matin->data;
2377:   Vec            bb1 = 0;

2380:   if (flag == SOR_APPLY_UPPER) {
2381:     (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);
2382:     return(0);
2383:   }

2385:   if (its > 1 || ~flag & SOR_ZERO_INITIAL_GUESS) {
2386:     VecDuplicate(bb,&bb1);
2387:   }

2389:   if ((flag & SOR_LOCAL_SYMMETRIC_SWEEP) == SOR_LOCAL_SYMMETRIC_SWEEP) {
2390:     if (flag & SOR_ZERO_INITIAL_GUESS) {
2391:       (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);
2392:       its--;
2393:     }

2395:     while (its--) {
2396:       VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);
2397:       VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);

2399:       /* update rhs: bb1 = bb - B*x */
2400:       VecScale(mat->lvec,-1.0);
2401:       (*mat->B->ops->multadd)(mat->B,mat->lvec,bb,bb1);

2403:       /* local sweep */
2404:       (*mat->A->ops->sor)(mat->A,bb1,omega,SOR_SYMMETRIC_SWEEP,fshift,lits,1,xx);
2405:     }
2406:   } else if (flag & SOR_LOCAL_FORWARD_SWEEP) {
2407:     if (flag & SOR_ZERO_INITIAL_GUESS) {
2408:       (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);
2409:       its--;
2410:     }
2411:     while (its--) {
2412:       VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);
2413:       VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);

2415:       /* update rhs: bb1 = bb - B*x */
2416:       VecScale(mat->lvec,-1.0);
2417:       (*mat->B->ops->multadd)(mat->B,mat->lvec,bb,bb1);

2419:       /* local sweep */
2420:       (*mat->A->ops->sor)(mat->A,bb1,omega,SOR_FORWARD_SWEEP,fshift,lits,1,xx);
2421:     }
2422:   } else if (flag & SOR_LOCAL_BACKWARD_SWEEP) {
2423:     if (flag & SOR_ZERO_INITIAL_GUESS) {
2424:       (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);
2425:       its--;
2426:     }
2427:     while (its--) {
2428:       VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);
2429:       VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);

2431:       /* update rhs: bb1 = bb - B*x */
2432:       VecScale(mat->lvec,-1.0);
2433:       (*mat->B->ops->multadd)(mat->B,mat->lvec,bb,bb1);

2435:       /* local sweep */
2436:       (*mat->A->ops->sor)(mat->A,bb1,omega,SOR_BACKWARD_SWEEP,fshift,lits,1,xx);
2437:     }
2438:   } else SETERRQ(PetscObjectComm((PetscObject)matin),PETSC_ERR_SUP,"Parallel version of SOR requested not supported");

2440:   VecDestroy(&bb1);
2441:   return(0);
2442: }

2446: PetscErrorCode MatGetColumnNorms_MPIBAIJ(Mat A,NormType type,PetscReal *norms)
2447: {
2449:   Mat_MPIBAIJ    *aij = (Mat_MPIBAIJ*)A->data;
2450:   PetscInt       N,i,*garray = aij->garray;
2451:   PetscInt       ib,jb,bs = A->rmap->bs;
2452:   Mat_SeqBAIJ    *a_aij = (Mat_SeqBAIJ*) aij->A->data;
2453:   MatScalar      *a_val = a_aij->a;
2454:   Mat_SeqBAIJ    *b_aij = (Mat_SeqBAIJ*) aij->B->data;
2455:   MatScalar      *b_val = b_aij->a;
2456:   PetscReal      *work;

2459:   MatGetSize(A,NULL,&N);
2460:   PetscCalloc1(N,&work);
2461:   if (type == NORM_2) {
2462:     for (i=a_aij->i[0]; i<a_aij->i[aij->A->rmap->n/bs]; i++) {
2463:       for (jb=0; jb<bs; jb++) {
2464:         for (ib=0; ib<bs; ib++) {
2465:           work[A->cmap->rstart + a_aij->j[i] * bs + jb] += PetscAbsScalar(*a_val * *a_val);
2466:           a_val++;
2467:         }
2468:       }
2469:     }
2470:     for (i=b_aij->i[0]; i<b_aij->i[aij->B->rmap->n/bs]; i++) {
2471:       for (jb=0; jb<bs; jb++) {
2472:         for (ib=0; ib<bs; ib++) {
2473:           work[garray[b_aij->j[i]] * bs + jb] += PetscAbsScalar(*b_val * *b_val);
2474:           b_val++;
2475:         }
2476:       }
2477:     }
2478:   } else if (type == NORM_1) {
2479:     for (i=a_aij->i[0]; i<a_aij->i[aij->A->rmap->n/bs]; i++) {
2480:       for (jb=0; jb<bs; jb++) {
2481:         for (ib=0; ib<bs; ib++) {
2482:           work[A->cmap->rstart + a_aij->j[i] * bs + jb] += PetscAbsScalar(*a_val);
2483:           a_val++;
2484:         }
2485:       }
2486:     }
2487:     for (i=b_aij->i[0]; i<b_aij->i[aij->B->rmap->n/bs]; i++) {
2488:       for (jb=0; jb<bs; jb++) {
2489:        for (ib=0; ib<bs; ib++) {
2490:           work[garray[b_aij->j[i]] * bs + jb] += PetscAbsScalar(*b_val);
2491:           b_val++;
2492:         }
2493:       }
2494:     }
2495:   } else if (type == NORM_INFINITY) {
2496:     for (i=a_aij->i[0]; i<a_aij->i[aij->A->rmap->n/bs]; i++) {
2497:       for (jb=0; jb<bs; jb++) {
2498:         for (ib=0; ib<bs; ib++) {
2499:           int col = A->cmap->rstart + a_aij->j[i] * bs + jb;
2500:           work[col] = PetscMax(PetscAbsScalar(*a_val), work[col]);
2501:           a_val++;
2502:         }
2503:       }
2504:     }
2505:     for (i=b_aij->i[0]; i<b_aij->i[aij->B->rmap->n/bs]; i++) {
2506:       for (jb=0; jb<bs; jb++) {
2507:         for (ib=0; ib<bs; ib++) {
2508:           int col = garray[b_aij->j[i]] * bs + jb;
2509:           work[col] = PetscMax(PetscAbsScalar(*b_val), work[col]);
2510:           b_val++;
2511:         }
2512:       }
2513:     }
2514:   } else SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_WRONG,"Unknown NormType");
2515:   if (type == NORM_INFINITY) {
2516:     MPI_Allreduce(work,norms,N,MPIU_REAL,MPIU_MAX,PetscObjectComm((PetscObject)A));
2517:   } else {
2518:     MPI_Allreduce(work,norms,N,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)A));
2519:   }
2520:   PetscFree(work);
2521:   if (type == NORM_2) {
2522:     for (i=0; i<N; i++) norms[i] = PetscSqrtReal(norms[i]);
2523:   }
2524:   return(0);
2525: }

2529: PetscErrorCode  MatInvertBlockDiagonal_MPIBAIJ(Mat A,const PetscScalar **values)
2530: {
2531:   Mat_MPIBAIJ    *a = (Mat_MPIBAIJ*) A->data;

2535:   MatInvertBlockDiagonal(a->A,values);
2536:   return(0);
2537: }


2540: /* -------------------------------------------------------------------*/
2541: static struct _MatOps MatOps_Values = {MatSetValues_MPIBAIJ,
2542:                                        MatGetRow_MPIBAIJ,
2543:                                        MatRestoreRow_MPIBAIJ,
2544:                                        MatMult_MPIBAIJ,
2545:                                 /* 4*/ MatMultAdd_MPIBAIJ,
2546:                                        MatMultTranspose_MPIBAIJ,
2547:                                        MatMultTransposeAdd_MPIBAIJ,
2548:                                        0,
2549:                                        0,
2550:                                        0,
2551:                                 /*10*/ 0,
2552:                                        0,
2553:                                        0,
2554:                                        MatSOR_MPIBAIJ,
2555:                                        MatTranspose_MPIBAIJ,
2556:                                 /*15*/ MatGetInfo_MPIBAIJ,
2557:                                        MatEqual_MPIBAIJ,
2558:                                        MatGetDiagonal_MPIBAIJ,
2559:                                        MatDiagonalScale_MPIBAIJ,
2560:                                        MatNorm_MPIBAIJ,
2561:                                 /*20*/ MatAssemblyBegin_MPIBAIJ,
2562:                                        MatAssemblyEnd_MPIBAIJ,
2563:                                        MatSetOption_MPIBAIJ,
2564:                                        MatZeroEntries_MPIBAIJ,
2565:                                 /*24*/ MatZeroRows_MPIBAIJ,
2566:                                        0,
2567:                                        0,
2568:                                        0,
2569:                                        0,
2570:                                 /*29*/ MatSetUp_MPIBAIJ,
2571:                                        0,
2572:                                        0,
2573:                                        0,
2574:                                        0,
2575:                                 /*34*/ MatDuplicate_MPIBAIJ,
2576:                                        0,
2577:                                        0,
2578:                                        0,
2579:                                        0,
2580:                                 /*39*/ MatAXPY_MPIBAIJ,
2581:                                        MatGetSubMatrices_MPIBAIJ,
2582:                                        MatIncreaseOverlap_MPIBAIJ,
2583:                                        MatGetValues_MPIBAIJ,
2584:                                        MatCopy_MPIBAIJ,
2585:                                 /*44*/ 0,
2586:                                        MatScale_MPIBAIJ,
2587:                                        0,
2588:                                        0,
2589:                                        MatZeroRowsColumns_MPIBAIJ,
2590:                                 /*49*/ 0,
2591:                                        0,
2592:                                        0,
2593:                                        0,
2594:                                        0,
2595:                                 /*54*/ MatFDColoringCreate_MPIXAIJ,
2596:                                        0,
2597:                                        MatSetUnfactored_MPIBAIJ,
2598:                                        MatPermute_MPIBAIJ,
2599:                                        MatSetValuesBlocked_MPIBAIJ,
2600:                                 /*59*/ MatGetSubMatrix_MPIBAIJ,
2601:                                        MatDestroy_MPIBAIJ,
2602:                                        MatView_MPIBAIJ,
2603:                                        0,
2604:                                        0,
2605:                                 /*64*/ 0,
2606:                                        0,
2607:                                        0,
2608:                                        0,
2609:                                        0,
2610:                                 /*69*/ MatGetRowMaxAbs_MPIBAIJ,
2611:                                        0,
2612:                                        0,
2613:                                        0,
2614:                                        0,
2615:                                 /*74*/ 0,
2616:                                        MatFDColoringApply_BAIJ,
2617:                                        0,
2618:                                        0,
2619:                                        0,
2620:                                 /*79*/ 0,
2621:                                        0,
2622:                                        0,
2623:                                        0,
2624:                                        MatLoad_MPIBAIJ,
2625:                                 /*84*/ 0,
2626:                                        0,
2627:                                        0,
2628:                                        0,
2629:                                        0,
2630:                                 /*89*/ 0,
2631:                                        0,
2632:                                        0,
2633:                                        0,
2634:                                        0,
2635:                                 /*94*/ 0,
2636:                                        0,
2637:                                        0,
2638:                                        0,
2639:                                        0,
2640:                                 /*99*/ 0,
2641:                                        0,
2642:                                        0,
2643:                                        0,
2644:                                        0,
2645:                                 /*104*/0,
2646:                                        MatRealPart_MPIBAIJ,
2647:                                        MatImaginaryPart_MPIBAIJ,
2648:                                        0,
2649:                                        0,
2650:                                 /*109*/0,
2651:                                        0,
2652:                                        0,
2653:                                        0,
2654:                                        0,
2655:                                 /*114*/MatGetSeqNonzeroStructure_MPIBAIJ,
2656:                                        0,
2657:                                        MatGetGhosts_MPIBAIJ,
2658:                                        0,
2659:                                        0,
2660:                                 /*119*/0,
2661:                                        0,
2662:                                        0,
2663:                                        0,
2664:                                        MatGetMultiProcBlock_MPIBAIJ,
2665:                                 /*124*/0,
2666:                                        MatGetColumnNorms_MPIBAIJ,
2667:                                        MatInvertBlockDiagonal_MPIBAIJ,
2668:                                        0,
2669:                                        0,
2670:                                /*129*/ 0,
2671:                                        0,
2672:                                        0,
2673:                                        0,
2674:                                        0,
2675:                                /*134*/ 0,
2676:                                        0,
2677:                                        0,
2678:                                        0,
2679:                                        0,
2680:                                /*139*/ 0,
2681:                                        0,
2682:                                        0,
2683:                                        MatFDColoringSetUp_MPIXAIJ
2684: };

2688: PetscErrorCode  MatGetDiagonalBlock_MPIBAIJ(Mat A,Mat *a)
2689: {
2691:   *a = ((Mat_MPIBAIJ*)A->data)->A;
2692:   return(0);
2693: }

2695: PETSC_EXTERN PetscErrorCode MatConvert_MPIBAIJ_MPISBAIJ(Mat, MatType,MatReuse,Mat*);

2699: PetscErrorCode MatMPIBAIJSetPreallocationCSR_MPIBAIJ(Mat B,PetscInt bs,const PetscInt ii[],const PetscInt jj[],const PetscScalar V[])
2700: {
2701:   PetscInt       m,rstart,cstart,cend;
2702:   PetscInt       i,j,d,nz,nz_max=0,*d_nnz=0,*o_nnz=0;
2703:   const PetscInt *JJ    =0;
2704:   PetscScalar    *values=0;
2705:   PetscBool      roworiented = ((Mat_MPIBAIJ*)B->data)->roworiented;

2709:   PetscLayoutSetBlockSize(B->rmap,bs);
2710:   PetscLayoutSetBlockSize(B->cmap,bs);
2711:   PetscLayoutSetUp(B->rmap);
2712:   PetscLayoutSetUp(B->cmap);
2713:   PetscLayoutGetBlockSize(B->rmap,&bs);
2714:   m      = B->rmap->n/bs;
2715:   rstart = B->rmap->rstart/bs;
2716:   cstart = B->cmap->rstart/bs;
2717:   cend   = B->cmap->rend/bs;

2719:   if (ii[0]) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"ii[0] must be 0 but it is %D",ii[0]);
2720:   PetscMalloc2(m,&d_nnz,m,&o_nnz);
2721:   for (i=0; i<m; i++) {
2722:     nz = ii[i+1] - ii[i];
2723:     if (nz < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Local row %D has a negative number of columns %D",i,nz);
2724:     nz_max = PetscMax(nz_max,nz);
2725:     JJ     = jj + ii[i];
2726:     for (j=0; j<nz; j++) {
2727:       if (*JJ >= cstart) break;
2728:       JJ++;
2729:     }
2730:     d = 0;
2731:     for (; j<nz; j++) {
2732:       if (*JJ++ >= cend) break;
2733:       d++;
2734:     }
2735:     d_nnz[i] = d;
2736:     o_nnz[i] = nz - d;
2737:   }
2738:   MatMPIBAIJSetPreallocation(B,bs,0,d_nnz,0,o_nnz);
2739:   PetscFree2(d_nnz,o_nnz);

2741:   values = (PetscScalar*)V;
2742:   if (!values) {
2743:     PetscMalloc1(bs*bs*nz_max,&values);
2744:     PetscMemzero(values,bs*bs*nz_max*sizeof(PetscScalar));
2745:   }
2746:   for (i=0; i<m; i++) {
2747:     PetscInt          row    = i + rstart;
2748:     PetscInt          ncols  = ii[i+1] - ii[i];
2749:     const PetscInt    *icols = jj + ii[i];
2750:     if (!roworiented) {         /* block ordering matches the non-nested layout of MatSetValues so we can insert entire rows */
2751:       const PetscScalar *svals = values + (V ? (bs*bs*ii[i]) : 0);
2752:       MatSetValuesBlocked_MPIBAIJ(B,1,&row,ncols,icols,svals,INSERT_VALUES);
2753:     } else {                    /* block ordering does not match so we can only insert one block at a time. */
2754:       PetscInt j;
2755:       for (j=0; j<ncols; j++) {
2756:         const PetscScalar *svals = values + (V ? (bs*bs*(ii[i]+j)) : 0);
2757:         MatSetValuesBlocked_MPIBAIJ(B,1,&row,1,&icols[j],svals,INSERT_VALUES);
2758:       }
2759:     }
2760:   }

2762:   if (!V) { PetscFree(values); }
2763:   MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);
2764:   MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);
2765:   MatSetOption(B,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE);
2766:   return(0);
2767: }

2771: /*@C
2772:    MatMPIBAIJSetPreallocationCSR - Allocates memory for a sparse parallel matrix in BAIJ format
2773:    (the default parallel PETSc format).

2775:    Collective on MPI_Comm

2777:    Input Parameters:
2778: +  A - the matrix
2779: .  bs - the block size
2780: .  i - the indices into j for the start of each local row (starts with zero)
2781: .  j - the column indices for each local row (starts with zero) these must be sorted for each row
2782: -  v - optional values in the matrix

2784:    Level: developer

2786:    Notes: The order of the entries in values is specified by the MatOption MAT_ROW_ORIENTED.  For example, C programs
2787:    may want to use the default MAT_ROW_ORIENTED=PETSC_TRUE and use an array v[nnz][bs][bs] where the second index is
2788:    over rows within a block and the last index is over columns within a block row.  Fortran programs will likely set
2789:    MAT_ROW_ORIENTED=PETSC_FALSE and use a Fortran array v(bs,bs,nnz) in which the first index is over rows within a
2790:    block column and the second index is over columns within a block.

2792: .keywords: matrix, aij, compressed row, sparse, parallel

2794: .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIBAIJSetPreallocation(), MatCreateAIJ(), MPIAIJ, MatCreateMPIBAIJWithArrays(), MPIBAIJ
2795: @*/
2796: PetscErrorCode  MatMPIBAIJSetPreallocationCSR(Mat B,PetscInt bs,const PetscInt i[],const PetscInt j[], const PetscScalar v[])
2797: {

2804:   PetscTryMethod(B,"MatMPIBAIJSetPreallocationCSR_C",(Mat,PetscInt,const PetscInt[],const PetscInt[],const PetscScalar[]),(B,bs,i,j,v));
2805:   return(0);
2806: }

2810: PetscErrorCode  MatMPIBAIJSetPreallocation_MPIBAIJ(Mat B,PetscInt bs,PetscInt d_nz,const PetscInt *d_nnz,PetscInt o_nz,const PetscInt *o_nnz)
2811: {
2812:   Mat_MPIBAIJ    *b;
2814:   PetscInt       i;

2817:   PetscLayoutSetBlockSize(B->rmap,bs);
2818:   PetscLayoutSetBlockSize(B->cmap,bs);
2819:   PetscLayoutSetUp(B->rmap);
2820:   PetscLayoutSetUp(B->cmap);
2821:   PetscLayoutGetBlockSize(B->rmap,&bs);

2823:   if (d_nnz) {
2824:     for (i=0; i<B->rmap->n/bs; i++) {
2825:       if (d_nnz[i] < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"d_nnz cannot be less than -1: local row %D value %D",i,d_nnz[i]);
2826:     }
2827:   }
2828:   if (o_nnz) {
2829:     for (i=0; i<B->rmap->n/bs; i++) {
2830:       if (o_nnz[i] < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"o_nnz cannot be less than -1: local row %D value %D",i,o_nnz[i]);
2831:     }
2832:   }

2834:   b      = (Mat_MPIBAIJ*)B->data;
2835:   b->bs2 = bs*bs;
2836:   b->mbs = B->rmap->n/bs;
2837:   b->nbs = B->cmap->n/bs;
2838:   b->Mbs = B->rmap->N/bs;
2839:   b->Nbs = B->cmap->N/bs;

2841:   for (i=0; i<=b->size; i++) {
2842:     b->rangebs[i] = B->rmap->range[i]/bs;
2843:   }
2844:   b->rstartbs = B->rmap->rstart/bs;
2845:   b->rendbs   = B->rmap->rend/bs;
2846:   b->cstartbs = B->cmap->rstart/bs;
2847:   b->cendbs   = B->cmap->rend/bs;

2849:   if (!B->preallocated) {
2850:     MatCreate(PETSC_COMM_SELF,&b->A);
2851:     MatSetSizes(b->A,B->rmap->n,B->cmap->n,B->rmap->n,B->cmap->n);
2852:     MatSetType(b->A,MATSEQBAIJ);
2853:     PetscLogObjectParent((PetscObject)B,(PetscObject)b->A);
2854:     MatCreate(PETSC_COMM_SELF,&b->B);
2855:     MatSetSizes(b->B,B->rmap->n,B->cmap->N,B->rmap->n,B->cmap->N);
2856:     MatSetType(b->B,MATSEQBAIJ);
2857:     PetscLogObjectParent((PetscObject)B,(PetscObject)b->B);
2858:     MatStashCreate_Private(PetscObjectComm((PetscObject)B),bs,&B->bstash);
2859:   }

2861:   MatSeqBAIJSetPreallocation(b->A,bs,d_nz,d_nnz);
2862:   MatSeqBAIJSetPreallocation(b->B,bs,o_nz,o_nnz);
2863:   B->preallocated = PETSC_TRUE;
2864:   return(0);
2865: }

2867: extern PetscErrorCode  MatDiagonalScaleLocal_MPIBAIJ(Mat,Vec);
2868: extern PetscErrorCode  MatSetHashTableFactor_MPIBAIJ(Mat,PetscReal);

2872: PETSC_EXTERN PetscErrorCode MatConvert_MPIBAIJ_MPIAdj(Mat B, MatType newtype,MatReuse reuse,Mat *adj)
2873: {
2874:   Mat_MPIBAIJ    *b = (Mat_MPIBAIJ*)B->data;
2876:   Mat_SeqBAIJ    *d  = (Mat_SeqBAIJ*) b->A->data,*o = (Mat_SeqBAIJ*) b->B->data;
2877:   PetscInt       M   = B->rmap->n/B->rmap->bs,i,*ii,*jj,cnt,j,k,rstart = B->rmap->rstart/B->rmap->bs;
2878:   const PetscInt *id = d->i, *jd = d->j, *io = o->i, *jo = o->j, *garray = b->garray;

2881:   PetscMalloc1((M+1),&ii);
2882:   ii[0] = 0;
2883:   for (i=0; i<M; i++) {
2884:     if ((id[i+1] - id[i]) < 0) SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Indices wrong %D %D %D",i,id[i],id[i+1]);
2885:     if ((io[i+1] - io[i]) < 0) SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Indices wrong %D %D %D",i,io[i],io[i+1]);
2886:     ii[i+1] = ii[i] + id[i+1] - id[i] + io[i+1] - io[i];
2887:     /* remove one from count of matrix has diagonal */
2888:     for (j=id[i]; j<id[i+1]; j++) {
2889:       if (jd[j] == i) {ii[i+1]--;break;}
2890:     }
2891:   }
2892:   PetscMalloc1(ii[M],&jj);
2893:   cnt  = 0;
2894:   for (i=0; i<M; i++) {
2895:     for (j=io[i]; j<io[i+1]; j++) {
2896:       if (garray[jo[j]] > rstart) break;
2897:       jj[cnt++] = garray[jo[j]];
2898:     }
2899:     for (k=id[i]; k<id[i+1]; k++) {
2900:       if (jd[k] != i) {
2901:         jj[cnt++] = rstart + jd[k];
2902:       }
2903:     }
2904:     for (; j<io[i+1]; j++) {
2905:       jj[cnt++] = garray[jo[j]];
2906:     }
2907:   }
2908:   MatCreateMPIAdj(PetscObjectComm((PetscObject)B),M,B->cmap->N/B->rmap->bs,ii,jj,NULL,adj);
2909:   return(0);
2910: }

2912: #include <../src/mat/impls/aij/mpi/mpiaij.h>

2914: PETSC_EXTERN PetscErrorCode MatConvert_SeqBAIJ_SeqAIJ(Mat,MatType,MatReuse,Mat*);

2918: PETSC_EXTERN PetscErrorCode MatConvert_MPIBAIJ_MPIAIJ(Mat A,MatType newtype,MatReuse reuse,Mat *newmat)
2919: {
2921:   Mat_MPIBAIJ    *a = (Mat_MPIBAIJ*)A->data;
2922:   Mat            B;
2923:   Mat_MPIAIJ     *b;

2926:   if (!A->assembled) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"Matrix must be assembled");

2928:   MatCreate(PetscObjectComm((PetscObject)A),&B);
2929:   MatSetSizes(B,A->rmap->n,A->cmap->n,A->rmap->N,A->cmap->N);
2930:   MatSetType(B,MATMPIAIJ);
2931:   MatSeqAIJSetPreallocation(B,0,NULL);
2932:   MatMPIAIJSetPreallocation(B,0,NULL,0,NULL);
2933:   b    = (Mat_MPIAIJ*) B->data;

2935:   MatDestroy(&b->A);
2936:   MatDestroy(&b->B);
2937:   MatDisAssemble_MPIBAIJ(A);
2938:   MatConvert_SeqBAIJ_SeqAIJ(a->A, MATSEQAIJ, MAT_INITIAL_MATRIX, &b->A);
2939:   MatConvert_SeqBAIJ_SeqAIJ(a->B, MATSEQAIJ, MAT_INITIAL_MATRIX, &b->B);
2940:   MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);
2941:   MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);
2942:   MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);
2943:   MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);
2944:   if (reuse == MAT_REUSE_MATRIX) {
2945:     MatHeaderReplace(A,B);
2946:   } else {
2947:    *newmat = B;
2948:   }
2949:   return(0);
2950: }

2952: #if defined(PETSC_HAVE_MUMPS)
2953: PETSC_EXTERN PetscErrorCode MatGetFactor_baij_mumps(Mat,MatFactorType,Mat*);
2954: #endif

2956: /*MC
2957:    MATMPIBAIJ - MATMPIBAIJ = "mpibaij" - A matrix type to be used for distributed block sparse matrices.

2959:    Options Database Keys:
2960: + -mat_type mpibaij - sets the matrix type to "mpibaij" during a call to MatSetFromOptions()
2961: . -mat_block_size <bs> - set the blocksize used to store the matrix
2962: - -mat_use_hash_table <fact>

2964:   Level: beginner

2966: .seealso: MatCreateMPIBAIJ
2967: M*/

2969: PETSC_EXTERN PetscErrorCode MatConvert_MPIBAIJ_MPIBSTRM(Mat,MatType,MatReuse,Mat*);

2973: PETSC_EXTERN PetscErrorCode MatCreate_MPIBAIJ(Mat B)
2974: {
2975:   Mat_MPIBAIJ    *b;
2977:   PetscBool      flg;

2980:   PetscNewLog(B,&b);
2981:   B->data = (void*)b;

2983:   PetscMemcpy(B->ops,&MatOps_Values,sizeof(struct _MatOps));
2984:   B->assembled = PETSC_FALSE;

2986:   B->insertmode = NOT_SET_VALUES;
2987:   MPI_Comm_rank(PetscObjectComm((PetscObject)B),&b->rank);
2988:   MPI_Comm_size(PetscObjectComm((PetscObject)B),&b->size);

2990:   /* build local table of row and column ownerships */
2991:   PetscMalloc1((b->size+1),&b->rangebs);

2993:   /* build cache for off array entries formed */
2994:   MatStashCreate_Private(PetscObjectComm((PetscObject)B),1,&B->stash);

2996:   b->donotstash  = PETSC_FALSE;
2997:   b->colmap      = NULL;
2998:   b->garray      = NULL;
2999:   b->roworiented = PETSC_TRUE;

3001:   /* stuff used in block assembly */
3002:   b->barray = 0;

3004:   /* stuff used for matrix vector multiply */
3005:   b->lvec  = 0;
3006:   b->Mvctx = 0;

3008:   /* stuff for MatGetRow() */
3009:   b->rowindices   = 0;
3010:   b->rowvalues    = 0;
3011:   b->getrowactive = PETSC_FALSE;

3013:   /* hash table stuff */
3014:   b->ht           = 0;
3015:   b->hd           = 0;
3016:   b->ht_size      = 0;
3017:   b->ht_flag      = PETSC_FALSE;
3018:   b->ht_fact      = 0;
3019:   b->ht_total_ct  = 0;
3020:   b->ht_insert_ct = 0;

3022:   /* stuff for MatGetSubMatrices_MPIBAIJ_local() */
3023:   b->ijonly = PETSC_FALSE;

3025:   PetscOptionsBegin(PetscObjectComm((PetscObject)B),NULL,"Options for loading MPIBAIJ matrix 1","Mat");
3026:   PetscOptionsBool("-mat_use_hash_table","Use hash table to save memory in constructing matrix","MatSetOption",PETSC_FALSE,&flg,NULL);
3027:   if (flg) {
3028:     PetscReal fact = 1.39;
3029:     MatSetOption(B,MAT_USE_HASH_TABLE,PETSC_TRUE);
3030:     PetscOptionsReal("-mat_use_hash_table","Use hash table factor","MatMPIBAIJSetHashTableFactor",fact,&fact,NULL);
3031:     if (fact <= 1.0) fact = 1.39;
3032:     MatMPIBAIJSetHashTableFactor(B,fact);
3033:     PetscInfo1(B,"Hash table Factor used %5.2f\n",fact);
3034:   }
3035:   PetscOptionsEnd();

3037: #if defined(PETSC_HAVE_MUMPS)
3038:   PetscObjectComposeFunction((PetscObject)B,"MatGetFactor_mumps_C",MatGetFactor_baij_mumps);
3039: #endif
3040:   PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpibaij_mpiadj_C",MatConvert_MPIBAIJ_MPIAdj);
3041:   PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpibaij_mpiaij_C",MatConvert_MPIBAIJ_MPIAIJ);
3042:   PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpibaij_mpisbaij_C",MatConvert_MPIBAIJ_MPISBAIJ);
3043:   PetscObjectComposeFunction((PetscObject)B,"MatStoreValues_C",MatStoreValues_MPIBAIJ);
3044:   PetscObjectComposeFunction((PetscObject)B,"MatRetrieveValues_C",MatRetrieveValues_MPIBAIJ);
3045:   PetscObjectComposeFunction((PetscObject)B,"MatGetDiagonalBlock_C",MatGetDiagonalBlock_MPIBAIJ);
3046:   PetscObjectComposeFunction((PetscObject)B,"MatMPIBAIJSetPreallocation_C",MatMPIBAIJSetPreallocation_MPIBAIJ);
3047:   PetscObjectComposeFunction((PetscObject)B,"MatMPIBAIJSetPreallocationCSR_C",MatMPIBAIJSetPreallocationCSR_MPIBAIJ);
3048:   PetscObjectComposeFunction((PetscObject)B,"MatDiagonalScaleLocal_C",MatDiagonalScaleLocal_MPIBAIJ);
3049:   PetscObjectComposeFunction((PetscObject)B,"MatSetHashTableFactor_C",MatSetHashTableFactor_MPIBAIJ);
3050:   PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpibaij_mpibstrm_C",MatConvert_MPIBAIJ_MPIBSTRM);
3051:   PetscObjectChangeTypeName((PetscObject)B,MATMPIBAIJ);
3052:   return(0);
3053: }

3055: /*MC
3056:    MATBAIJ - MATBAIJ = "baij" - A matrix type to be used for block sparse matrices.

3058:    This matrix type is identical to MATSEQBAIJ when constructed with a single process communicator,
3059:    and MATMPIBAIJ otherwise.

3061:    Options Database Keys:
3062: . -mat_type baij - sets the matrix type to "baij" during a call to MatSetFromOptions()

3064:   Level: beginner

3066: .seealso: MatCreateBAIJ(),MATSEQBAIJ,MATMPIBAIJ, MatMPIBAIJSetPreallocation(), MatMPIBAIJSetPreallocationCSR()
3067: M*/

3071: /*@C
3072:    MatMPIBAIJSetPreallocation - Allocates memory for a sparse parallel matrix in block AIJ format
3073:    (block compressed row).  For good matrix assembly performance
3074:    the user should preallocate the matrix storage by setting the parameters
3075:    d_nz (or d_nnz) and o_nz (or o_nnz).  By setting these parameters accurately,
3076:    performance can be increased by more than a factor of 50.

3078:    Collective on Mat

3080:    Input Parameters:
3081: +  A - the matrix
3082: .  bs   - size of block
3083: .  d_nz  - number of block nonzeros per block row in diagonal portion of local
3084:            submatrix  (same for all local rows)
3085: .  d_nnz - array containing the number of block nonzeros in the various block rows
3086:            of the in diagonal portion of the local (possibly different for each block
3087:            row) or NULL.  If you plan to factor the matrix you must leave room for the diagonal entry and
3088:            set it even if it is zero.
3089: .  o_nz  - number of block nonzeros per block row in the off-diagonal portion of local
3090:            submatrix (same for all local rows).
3091: -  o_nnz - array containing the number of nonzeros in the various block rows of the
3092:            off-diagonal portion of the local submatrix (possibly different for
3093:            each block row) or NULL.

3095:    If the *_nnz parameter is given then the *_nz parameter is ignored

3097:    Options Database Keys:
3098: +   -mat_block_size - size of the blocks to use
3099: -   -mat_use_hash_table <fact>

3101:    Notes:
3102:    If PETSC_DECIDE or  PETSC_DETERMINE is used for a particular argument on one processor
3103:    than it must be used on all processors that share the object for that argument.

3105:    Storage Information:
3106:    For a square global matrix we define each processor's diagonal portion
3107:    to be its local rows and the corresponding columns (a square submatrix);
3108:    each processor's off-diagonal portion encompasses the remainder of the
3109:    local matrix (a rectangular submatrix).

3111:    The user can specify preallocated storage for the diagonal part of
3112:    the local submatrix with either d_nz or d_nnz (not both).  Set
3113:    d_nz=PETSC_DEFAULT and d_nnz=NULL for PETSc to control dynamic
3114:    memory allocation.  Likewise, specify preallocated storage for the
3115:    off-diagonal part of the local submatrix with o_nz or o_nnz (not both).

3117:    Consider a processor that owns rows 3, 4 and 5 of a parallel matrix. In
3118:    the figure below we depict these three local rows and all columns (0-11).

3120: .vb
3121:            0 1 2 3 4 5 6 7 8 9 10 11
3122:           --------------------------
3123:    row 3  |o o o d d d o o o o  o  o
3124:    row 4  |o o o d d d o o o o  o  o
3125:    row 5  |o o o d d d o o o o  o  o
3126:           --------------------------
3127: .ve

3129:    Thus, any entries in the d locations are stored in the d (diagonal)
3130:    submatrix, and any entries in the o locations are stored in the
3131:    o (off-diagonal) submatrix.  Note that the d and the o submatrices are
3132:    stored simply in the MATSEQBAIJ format for compressed row storage.

3134:    Now d_nz should indicate the number of block nonzeros per row in the d matrix,
3135:    and o_nz should indicate the number of block nonzeros per row in the o matrix.
3136:    In general, for PDE problems in which most nonzeros are near the diagonal,
3137:    one expects d_nz >> o_nz.   For large problems you MUST preallocate memory
3138:    or you will get TERRIBLE performance; see the users' manual chapter on
3139:    matrices.

3141:    You can call MatGetInfo() to get information on how effective the preallocation was;
3142:    for example the fields mallocs,nz_allocated,nz_used,nz_unneeded;
3143:    You can also run with the option -info and look for messages with the string
3144:    malloc in them to see if additional memory allocation was needed.

3146:    Level: intermediate

3148: .keywords: matrix, block, aij, compressed row, sparse, parallel

3150: .seealso: MatCreate(), MatCreateSeqBAIJ(), MatSetValues(), MatCreateBAIJ(), MatMPIBAIJSetPreallocationCSR(), PetscSplitOwnership()
3151: @*/
3152: PetscErrorCode  MatMPIBAIJSetPreallocation(Mat B,PetscInt bs,PetscInt d_nz,const PetscInt d_nnz[],PetscInt o_nz,const PetscInt o_nnz[])
3153: {

3160:   PetscTryMethod(B,"MatMPIBAIJSetPreallocation_C",(Mat,PetscInt,PetscInt,const PetscInt[],PetscInt,const PetscInt[]),(B,bs,d_nz,d_nnz,o_nz,o_nnz));
3161:   return(0);
3162: }

3166: /*@C
3167:    MatCreateBAIJ - Creates a sparse parallel matrix in block AIJ format
3168:    (block compressed row).  For good matrix assembly performance
3169:    the user should preallocate the matrix storage by setting the parameters
3170:    d_nz (or d_nnz) and o_nz (or o_nnz).  By setting these parameters accurately,
3171:    performance can be increased by more than a factor of 50.

3173:    Collective on MPI_Comm

3175:    Input Parameters:
3176: +  comm - MPI communicator
3177: .  bs   - size of blockk
3178: .  m - number of local rows (or PETSC_DECIDE to have calculated if M is given)
3179:            This value should be the same as the local size used in creating the
3180:            y vector for the matrix-vector product y = Ax.
3181: .  n - number of local columns (or PETSC_DECIDE to have calculated if N is given)
3182:            This value should be the same as the local size used in creating the
3183:            x vector for the matrix-vector product y = Ax.
3184: .  M - number of global rows (or PETSC_DETERMINE to have calculated if m is given)
3185: .  N - number of global columns (or PETSC_DETERMINE to have calculated if n is given)
3186: .  d_nz  - number of nonzero blocks per block row in diagonal portion of local
3187:            submatrix  (same for all local rows)
3188: .  d_nnz - array containing the number of nonzero blocks in the various block rows
3189:            of the in diagonal portion of the local (possibly different for each block
3190:            row) or NULL.  If you plan to factor the matrix you must leave room for the diagonal entry
3191:            and set it even if it is zero.
3192: .  o_nz  - number of nonzero blocks per block row in the off-diagonal portion of local
3193:            submatrix (same for all local rows).
3194: -  o_nnz - array containing the number of nonzero blocks in the various block rows of the
3195:            off-diagonal portion of the local submatrix (possibly different for
3196:            each block row) or NULL.

3198:    Output Parameter:
3199: .  A - the matrix

3201:    Options Database Keys:
3202: +   -mat_block_size - size of the blocks to use
3203: -   -mat_use_hash_table <fact>

3205:    It is recommended that one use the MatCreate(), MatSetType() and/or MatSetFromOptions(),
3206:    MatXXXXSetPreallocation() paradgm instead of this routine directly.
3207:    [MatXXXXSetPreallocation() is, for example, MatSeqAIJSetPreallocation]

3209:    Notes:
3210:    If the *_nnz parameter is given then the *_nz parameter is ignored

3212:    A nonzero block is any block that as 1 or more nonzeros in it

3214:    The user MUST specify either the local or global matrix dimensions
3215:    (possibly both).

3217:    If PETSC_DECIDE or  PETSC_DETERMINE is used for a particular argument on one processor
3218:    than it must be used on all processors that share the object for that argument.

3220:    Storage Information:
3221:    For a square global matrix we define each processor's diagonal portion
3222:    to be its local rows and the corresponding columns (a square submatrix);
3223:    each processor's off-diagonal portion encompasses the remainder of the
3224:    local matrix (a rectangular submatrix).

3226:    The user can specify preallocated storage for the diagonal part of
3227:    the local submatrix with either d_nz or d_nnz (not both).  Set
3228:    d_nz=PETSC_DEFAULT and d_nnz=NULL for PETSc to control dynamic
3229:    memory allocation.  Likewise, specify preallocated storage for the
3230:    off-diagonal part of the local submatrix with o_nz or o_nnz (not both).

3232:    Consider a processor that owns rows 3, 4 and 5 of a parallel matrix. In
3233:    the figure below we depict these three local rows and all columns (0-11).

3235: .vb
3236:            0 1 2 3 4 5 6 7 8 9 10 11
3237:           --------------------------
3238:    row 3  |o o o d d d o o o o  o  o
3239:    row 4  |o o o d d d o o o o  o  o
3240:    row 5  |o o o d d d o o o o  o  o
3241:           --------------------------
3242: .ve

3244:    Thus, any entries in the d locations are stored in the d (diagonal)
3245:    submatrix, and any entries in the o locations are stored in the
3246:    o (off-diagonal) submatrix.  Note that the d and the o submatrices are
3247:    stored simply in the MATSEQBAIJ format for compressed row storage.

3249:    Now d_nz should indicate the number of block nonzeros per row in the d matrix,
3250:    and o_nz should indicate the number of block nonzeros per row in the o matrix.
3251:    In general, for PDE problems in which most nonzeros are near the diagonal,
3252:    one expects d_nz >> o_nz.   For large problems you MUST preallocate memory
3253:    or you will get TERRIBLE performance; see the users' manual chapter on
3254:    matrices.

3256:    Level: intermediate

3258: .keywords: matrix, block, aij, compressed row, sparse, parallel

3260: .seealso: MatCreate(), MatCreateSeqBAIJ(), MatSetValues(), MatCreateBAIJ(), MatMPIBAIJSetPreallocation(), MatMPIBAIJSetPreallocationCSR()
3261: @*/
3262: PetscErrorCode  MatCreateBAIJ(MPI_Comm comm,PetscInt bs,PetscInt m,PetscInt n,PetscInt M,PetscInt N,PetscInt d_nz,const PetscInt d_nnz[],PetscInt o_nz,const PetscInt o_nnz[],Mat *A)
3263: {
3265:   PetscMPIInt    size;

3268:   MatCreate(comm,A);
3269:   MatSetSizes(*A,m,n,M,N);
3270:   MPI_Comm_size(comm,&size);
3271:   if (size > 1) {
3272:     MatSetType(*A,MATMPIBAIJ);
3273:     MatMPIBAIJSetPreallocation(*A,bs,d_nz,d_nnz,o_nz,o_nnz);
3274:   } else {
3275:     MatSetType(*A,MATSEQBAIJ);
3276:     MatSeqBAIJSetPreallocation(*A,bs,d_nz,d_nnz);
3277:   }
3278:   return(0);
3279: }

3283: static PetscErrorCode MatDuplicate_MPIBAIJ(Mat matin,MatDuplicateOption cpvalues,Mat *newmat)
3284: {
3285:   Mat            mat;
3286:   Mat_MPIBAIJ    *a,*oldmat = (Mat_MPIBAIJ*)matin->data;
3288:   PetscInt       len=0;

3291:   *newmat = 0;
3292:   MatCreate(PetscObjectComm((PetscObject)matin),&mat);
3293:   MatSetSizes(mat,matin->rmap->n,matin->cmap->n,matin->rmap->N,matin->cmap->N);
3294:   MatSetType(mat,((PetscObject)matin)->type_name);
3295:   PetscMemcpy(mat->ops,matin->ops,sizeof(struct _MatOps));

3297:   mat->factortype   = matin->factortype;
3298:   mat->preallocated = PETSC_TRUE;
3299:   mat->assembled    = PETSC_TRUE;
3300:   mat->insertmode   = NOT_SET_VALUES;

3302:   a             = (Mat_MPIBAIJ*)mat->data;
3303:   mat->rmap->bs = matin->rmap->bs;
3304:   a->bs2        = oldmat->bs2;
3305:   a->mbs        = oldmat->mbs;
3306:   a->nbs        = oldmat->nbs;
3307:   a->Mbs        = oldmat->Mbs;
3308:   a->Nbs        = oldmat->Nbs;

3310:   PetscLayoutReference(matin->rmap,&mat->rmap);
3311:   PetscLayoutReference(matin->cmap,&mat->cmap);

3313:   a->size         = oldmat->size;
3314:   a->rank         = oldmat->rank;
3315:   a->donotstash   = oldmat->donotstash;
3316:   a->roworiented  = oldmat->roworiented;
3317:   a->rowindices   = 0;
3318:   a->rowvalues    = 0;
3319:   a->getrowactive = PETSC_FALSE;
3320:   a->barray       = 0;
3321:   a->rstartbs     = oldmat->rstartbs;
3322:   a->rendbs       = oldmat->rendbs;
3323:   a->cstartbs     = oldmat->cstartbs;
3324:   a->cendbs       = oldmat->cendbs;

3326:   /* hash table stuff */
3327:   a->ht           = 0;
3328:   a->hd           = 0;
3329:   a->ht_size      = 0;
3330:   a->ht_flag      = oldmat->ht_flag;
3331:   a->ht_fact      = oldmat->ht_fact;
3332:   a->ht_total_ct  = 0;
3333:   a->ht_insert_ct = 0;

3335:   PetscMemcpy(a->rangebs,oldmat->rangebs,(a->size+1)*sizeof(PetscInt));
3336:   if (oldmat->colmap) {
3337: #if defined(PETSC_USE_CTABLE)
3338:     PetscTableCreateCopy(oldmat->colmap,&a->colmap);
3339: #else
3340:     PetscMalloc1((a->Nbs),&a->colmap);
3341:     PetscLogObjectMemory((PetscObject)mat,(a->Nbs)*sizeof(PetscInt));
3342:     PetscMemcpy(a->colmap,oldmat->colmap,(a->Nbs)*sizeof(PetscInt));
3343: #endif
3344:   } else a->colmap = 0;

3346:   if (oldmat->garray && (len = ((Mat_SeqBAIJ*)(oldmat->B->data))->nbs)) {
3347:     PetscMalloc1(len,&a->garray);
3348:     PetscLogObjectMemory((PetscObject)mat,len*sizeof(PetscInt));
3349:     PetscMemcpy(a->garray,oldmat->garray,len*sizeof(PetscInt));
3350:   } else a->garray = 0;

3352:   MatStashCreate_Private(PetscObjectComm((PetscObject)matin),matin->rmap->bs,&mat->bstash);
3353:   VecDuplicate(oldmat->lvec,&a->lvec);
3354:   PetscLogObjectParent((PetscObject)mat,(PetscObject)a->lvec);
3355:   VecScatterCopy(oldmat->Mvctx,&a->Mvctx);
3356:   PetscLogObjectParent((PetscObject)mat,(PetscObject)a->Mvctx);

3358:   MatDuplicate(oldmat->A,cpvalues,&a->A);
3359:   PetscLogObjectParent((PetscObject)mat,(PetscObject)a->A);
3360:   MatDuplicate(oldmat->B,cpvalues,&a->B);
3361:   PetscLogObjectParent((PetscObject)mat,(PetscObject)a->B);
3362:   PetscFunctionListDuplicate(((PetscObject)matin)->qlist,&((PetscObject)mat)->qlist);
3363:   *newmat = mat;
3364:   return(0);
3365: }

3369: PetscErrorCode MatLoad_MPIBAIJ(Mat newmat,PetscViewer viewer)
3370: {
3372:   int            fd;
3373:   PetscInt       i,nz,j,rstart,rend;
3374:   PetscScalar    *vals,*buf;
3375:   MPI_Comm       comm;
3376:   MPI_Status     status;
3377:   PetscMPIInt    rank,size,maxnz;
3378:   PetscInt       header[4],*rowlengths = 0,M,N,m,*rowners,*cols;
3379:   PetscInt       *locrowlens = NULL,*procsnz = NULL,*browners = NULL;
3380:   PetscInt       jj,*mycols,*ibuf,bs=1,Mbs,mbs,extra_rows,mmax;
3381:   PetscMPIInt    tag    = ((PetscObject)viewer)->tag;
3382:   PetscInt       *dlens = NULL,*odlens = NULL,*mask = NULL,*masked1 = NULL,*masked2 = NULL,rowcount,odcount;
3383:   PetscInt       dcount,kmax,k,nzcount,tmp,mend,sizesset=1,grows,gcols;

3386:   PetscObjectGetComm((PetscObject)viewer,&comm);
3387:   PetscOptionsBegin(comm,NULL,"Options for loading MPIBAIJ matrix 2","Mat");
3388:   PetscOptionsInt("-matload_block_size","Set the blocksize used to store the matrix","MatLoad",bs,&bs,NULL);
3389:   PetscOptionsEnd();

3391:   MPI_Comm_size(comm,&size);
3392:   MPI_Comm_rank(comm,&rank);
3393:   if (!rank) {
3394:     PetscViewerBinaryGetDescriptor(viewer,&fd);
3395:     PetscBinaryRead(fd,(char*)header,4,PETSC_INT);
3396:     if (header[0] != MAT_FILE_CLASSID) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED,"not matrix object");
3397:   }

3399:   if (newmat->rmap->n < 0 && newmat->rmap->N < 0 && newmat->cmap->n < 0 && newmat->cmap->N < 0) sizesset = 0;

3401:   MPI_Bcast(header+1,3,MPIU_INT,0,comm);
3402:   M    = header[1]; N = header[2];

3404:   /* If global rows/cols are set to PETSC_DECIDE, set it to the sizes given in the file */
3405:   if (sizesset && newmat->rmap->N < 0) newmat->rmap->N = M;
3406:   if (sizesset && newmat->cmap->N < 0) newmat->cmap->N = N;

3408:   /* If global sizes are set, check if they are consistent with that given in the file */
3409:   if (sizesset) {
3410:     MatGetSize(newmat,&grows,&gcols);
3411:   }
3412:   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);
3413:   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);

3415:   if (M != N) SETERRQ(PetscObjectComm((PetscObject)viewer),PETSC_ERR_SUP,"Can only do square matrices");

3417:   /*
3418:      This code adds extra rows to make sure the number of rows is
3419:      divisible by the blocksize
3420:   */
3421:   Mbs        = M/bs;
3422:   extra_rows = bs - M + bs*Mbs;
3423:   if (extra_rows == bs) extra_rows = 0;
3424:   else                  Mbs++;
3425:   if (extra_rows && !rank) {
3426:     PetscInfo(viewer,"Padding loaded matrix to match blocksize\n");
3427:   }

3429:   /* determine ownership of all rows */
3430:   if (newmat->rmap->n < 0) { /* PETSC_DECIDE */
3431:     mbs = Mbs/size + ((Mbs % size) > rank);
3432:     m   = mbs*bs;
3433:   } else { /* User set */
3434:     m   = newmat->rmap->n;
3435:     mbs = m/bs;
3436:   }
3437:   PetscMalloc2(size+1,&rowners,size+1,&browners);
3438:   MPI_Allgather(&mbs,1,MPIU_INT,rowners+1,1,MPIU_INT,comm);

3440:   /* process 0 needs enough room for process with most rows */
3441:   if (!rank) {
3442:     mmax = rowners[1];
3443:     for (i=2; i<=size; i++) {
3444:       mmax = PetscMax(mmax,rowners[i]);
3445:     }
3446:     mmax*=bs;
3447:   } else mmax = -1;             /* unused, but compiler warns anyway */

3449:   rowners[0] = 0;
3450:   for (i=2; i<=size; i++) rowners[i] += rowners[i-1];
3451:   for (i=0; i<=size; i++) browners[i] = rowners[i]*bs;
3452:   rstart = rowners[rank];
3453:   rend   = rowners[rank+1];

3455:   /* distribute row lengths to all processors */
3456:   PetscMalloc1(m,&locrowlens);
3457:   if (!rank) {
3458:     mend = m;
3459:     if (size == 1) mend = mend - extra_rows;
3460:     PetscBinaryRead(fd,locrowlens,mend,PETSC_INT);
3461:     for (j=mend; j<m; j++) locrowlens[j] = 1;
3462:     PetscMalloc1(mmax,&rowlengths);
3463:     PetscCalloc1(size,&procsnz);
3464:     for (j=0; j<m; j++) {
3465:       procsnz[0] += locrowlens[j];
3466:     }
3467:     for (i=1; i<size; i++) {
3468:       mend = browners[i+1] - browners[i];
3469:       if (i == size-1) mend = mend - extra_rows;
3470:       PetscBinaryRead(fd,rowlengths,mend,PETSC_INT);
3471:       for (j=mend; j<browners[i+1] - browners[i]; j++) rowlengths[j] = 1;
3472:       /* calculate the number of nonzeros on each processor */
3473:       for (j=0; j<browners[i+1]-browners[i]; j++) {
3474:         procsnz[i] += rowlengths[j];
3475:       }
3476:       MPI_Send(rowlengths,browners[i+1]-browners[i],MPIU_INT,i,tag,comm);
3477:     }
3478:     PetscFree(rowlengths);
3479:   } else {
3480:     MPI_Recv(locrowlens,m,MPIU_INT,0,tag,comm,&status);
3481:   }

3483:   if (!rank) {
3484:     /* determine max buffer needed and allocate it */
3485:     maxnz = procsnz[0];
3486:     for (i=1; i<size; i++) {
3487:       maxnz = PetscMax(maxnz,procsnz[i]);
3488:     }
3489:     PetscMalloc1(maxnz,&cols);

3491:     /* read in my part of the matrix column indices  */
3492:     nz     = procsnz[0];
3493:     PetscMalloc1((nz+1),&ibuf);
3494:     mycols = ibuf;
3495:     if (size == 1) nz -= extra_rows;
3496:     PetscBinaryRead(fd,mycols,nz,PETSC_INT);
3497:     if (size == 1) {
3498:       for (i=0; i< extra_rows; i++) mycols[nz+i] = M+i;
3499:     }

3501:     /* read in every ones (except the last) and ship off */
3502:     for (i=1; i<size-1; i++) {
3503:       nz   = procsnz[i];
3504:       PetscBinaryRead(fd,cols,nz,PETSC_INT);
3505:       MPI_Send(cols,nz,MPIU_INT,i,tag,comm);
3506:     }
3507:     /* read in the stuff for the last proc */
3508:     if (size != 1) {
3509:       nz   = procsnz[size-1] - extra_rows;  /* the extra rows are not on the disk */
3510:       PetscBinaryRead(fd,cols,nz,PETSC_INT);
3511:       for (i=0; i<extra_rows; i++) cols[nz+i] = M+i;
3512:       MPI_Send(cols,nz+extra_rows,MPIU_INT,size-1,tag,comm);
3513:     }
3514:     PetscFree(cols);
3515:   } else {
3516:     /* determine buffer space needed for message */
3517:     nz = 0;
3518:     for (i=0; i<m; i++) {
3519:       nz += locrowlens[i];
3520:     }
3521:     PetscMalloc1((nz+1),&ibuf);
3522:     mycols = ibuf;
3523:     /* receive message of column indices*/
3524:     MPI_Recv(mycols,nz,MPIU_INT,0,tag,comm,&status);
3525:     MPI_Get_count(&status,MPIU_INT,&maxnz);
3526:     if (maxnz != nz) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED,"something is wrong with file");
3527:   }

3529:   /* loop over local rows, determining number of off diagonal entries */
3530:   PetscMalloc2(rend-rstart,&dlens,rend-rstart,&odlens);
3531:   PetscCalloc3(Mbs,&mask,Mbs,&masked1,Mbs,&masked2);
3532:   rowcount = 0; nzcount = 0;
3533:   for (i=0; i<mbs; i++) {
3534:     dcount  = 0;
3535:     odcount = 0;
3536:     for (j=0; j<bs; j++) {
3537:       kmax = locrowlens[rowcount];
3538:       for (k=0; k<kmax; k++) {
3539:         tmp = mycols[nzcount++]/bs;
3540:         if (!mask[tmp]) {
3541:           mask[tmp] = 1;
3542:           if (tmp < rstart || tmp >= rend) masked2[odcount++] = tmp;
3543:           else masked1[dcount++] = tmp;
3544:         }
3545:       }
3546:       rowcount++;
3547:     }

3549:     dlens[i]  = dcount;
3550:     odlens[i] = odcount;

3552:     /* zero out the mask elements we set */
3553:     for (j=0; j<dcount; j++) mask[masked1[j]] = 0;
3554:     for (j=0; j<odcount; j++) mask[masked2[j]] = 0;
3555:   }


3558:   if (!sizesset) {
3559:     MatSetSizes(newmat,m,m,M+extra_rows,N+extra_rows);
3560:   }
3561:   MatMPIBAIJSetPreallocation(newmat,bs,0,dlens,0,odlens);

3563:   if (!rank) {
3564:     PetscMalloc1((maxnz+1),&buf);
3565:     /* read in my part of the matrix numerical values  */
3566:     nz     = procsnz[0];
3567:     vals   = buf;
3568:     mycols = ibuf;
3569:     if (size == 1) nz -= extra_rows;
3570:     PetscBinaryRead(fd,vals,nz,PETSC_SCALAR);
3571:     if (size == 1) {
3572:       for (i=0; i< extra_rows; i++) vals[nz+i] = 1.0;
3573:     }

3575:     /* insert into matrix */
3576:     jj = rstart*bs;
3577:     for (i=0; i<m; i++) {
3578:       MatSetValues_MPIBAIJ(newmat,1,&jj,locrowlens[i],mycols,vals,INSERT_VALUES);
3579:       mycols += locrowlens[i];
3580:       vals   += locrowlens[i];
3581:       jj++;
3582:     }
3583:     /* read in other processors (except the last one) and ship out */
3584:     for (i=1; i<size-1; i++) {
3585:       nz   = procsnz[i];
3586:       vals = buf;
3587:       PetscBinaryRead(fd,vals,nz,PETSC_SCALAR);
3588:       MPIULong_Send(vals,nz,MPIU_SCALAR,i,((PetscObject)newmat)->tag,comm);
3589:     }
3590:     /* the last proc */
3591:     if (size != 1) {
3592:       nz   = procsnz[i] - extra_rows;
3593:       vals = buf;
3594:       PetscBinaryRead(fd,vals,nz,PETSC_SCALAR);
3595:       for (i=0; i<extra_rows; i++) vals[nz+i] = 1.0;
3596:       MPIULong_Send(vals,nz+extra_rows,MPIU_SCALAR,size-1,((PetscObject)newmat)->tag,comm);
3597:     }
3598:     PetscFree(procsnz);
3599:   } else {
3600:     /* receive numeric values */
3601:     PetscMalloc1((nz+1),&buf);

3603:     /* receive message of values*/
3604:     vals   = buf;
3605:     mycols = ibuf;
3606:     MPIULong_Recv(vals,nz,MPIU_SCALAR,0,((PetscObject)newmat)->tag,comm);

3608:     /* insert into matrix */
3609:     jj = rstart*bs;
3610:     for (i=0; i<m; i++) {
3611:       MatSetValues_MPIBAIJ(newmat,1,&jj,locrowlens[i],mycols,vals,INSERT_VALUES);
3612:       mycols += locrowlens[i];
3613:       vals   += locrowlens[i];
3614:       jj++;
3615:     }
3616:   }
3617:   PetscFree(locrowlens);
3618:   PetscFree(buf);
3619:   PetscFree(ibuf);
3620:   PetscFree2(rowners,browners);
3621:   PetscFree2(dlens,odlens);
3622:   PetscFree3(mask,masked1,masked2);
3623:   MatAssemblyBegin(newmat,MAT_FINAL_ASSEMBLY);
3624:   MatAssemblyEnd(newmat,MAT_FINAL_ASSEMBLY);
3625:   return(0);
3626: }

3630: /*@
3631:    MatMPIBAIJSetHashTableFactor - Sets the factor required to compute the size of the HashTable.

3633:    Input Parameters:
3634: .  mat  - the matrix
3635: .  fact - factor

3637:    Not Collective, each process can use a different factor

3639:    Level: advanced

3641:   Notes:
3642:    This can also be set by the command line option: -mat_use_hash_table <fact>

3644: .keywords: matrix, hashtable, factor, HT

3646: .seealso: MatSetOption()
3647: @*/
3648: PetscErrorCode  MatMPIBAIJSetHashTableFactor(Mat mat,PetscReal fact)
3649: {

3653:   PetscTryMethod(mat,"MatSetHashTableFactor_C",(Mat,PetscReal),(mat,fact));
3654:   return(0);
3655: }

3659: PetscErrorCode  MatSetHashTableFactor_MPIBAIJ(Mat mat,PetscReal fact)
3660: {
3661:   Mat_MPIBAIJ *baij;

3664:   baij          = (Mat_MPIBAIJ*)mat->data;
3665:   baij->ht_fact = fact;
3666:   return(0);
3667: }

3671: PetscErrorCode  MatMPIBAIJGetSeqBAIJ(Mat A,Mat *Ad,Mat *Ao,const PetscInt *colmap[])
3672: {
3673:   Mat_MPIBAIJ *a = (Mat_MPIBAIJ*)A->data;

3676:   *Ad     = a->A;
3677:   *Ao     = a->B;
3678:   *colmap = a->garray;
3679:   return(0);
3680: }

3682: /*
3683:     Special version for direct calls from Fortran (to eliminate two function call overheads
3684: */
3685: #if defined(PETSC_HAVE_FORTRAN_CAPS)
3686: #define matmpibaijsetvaluesblocked_ MATMPIBAIJSETVALUESBLOCKED
3687: #elif !defined(PETSC_HAVE_FORTRAN_UNDERSCORE)
3688: #define matmpibaijsetvaluesblocked_ matmpibaijsetvaluesblocked
3689: #endif

3693: /*@C
3694:   MatMPIBAIJSetValuesBlocked - Direct Fortran call to replace call to MatSetValuesBlocked()

3696:   Collective on Mat

3698:   Input Parameters:
3699: + mat - the matrix
3700: . min - number of input rows
3701: . im - input rows
3702: . nin - number of input columns
3703: . in - input columns
3704: . v - numerical values input
3705: - addvin - INSERT_VALUES or ADD_VALUES

3707:   Notes: This has a complete copy of MatSetValuesBlocked_MPIBAIJ() which is terrible code un-reuse.

3709:   Level: advanced

3711: .seealso:   MatSetValuesBlocked()
3712: @*/
3713: PetscErrorCode matmpibaijsetvaluesblocked_(Mat *matin,PetscInt *min,const PetscInt im[],PetscInt *nin,const PetscInt in[],const MatScalar v[],InsertMode *addvin)
3714: {
3715:   /* convert input arguments to C version */
3716:   Mat        mat  = *matin;
3717:   PetscInt   m    = *min, n = *nin;
3718:   InsertMode addv = *addvin;

3720:   Mat_MPIBAIJ     *baij = (Mat_MPIBAIJ*)mat->data;
3721:   const MatScalar *value;
3722:   MatScalar       *barray     = baij->barray;
3723:   PetscBool       roworiented = baij->roworiented;
3724:   PetscErrorCode  ierr;
3725:   PetscInt        i,j,ii,jj,row,col,rstart=baij->rstartbs;
3726:   PetscInt        rend=baij->rendbs,cstart=baij->cstartbs,stepval;
3727:   PetscInt        cend=baij->cendbs,bs=mat->rmap->bs,bs2=baij->bs2;

3730:   /* tasks normally handled by MatSetValuesBlocked() */
3731:   if (mat->insertmode == NOT_SET_VALUES) mat->insertmode = addv;
3732: #if defined(PETSC_USE_DEBUG)
3733:   else if (mat->insertmode != addv) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Cannot mix add values and insert values");
3734:   if (mat->factortype) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Not for factored matrix");
3735: #endif
3736:   if (mat->assembled) {
3737:     mat->was_assembled = PETSC_TRUE;
3738:     mat->assembled     = PETSC_FALSE;
3739:   }
3740:   PetscLogEventBegin(MAT_SetValues,mat,0,0,0);


3743:   if (!barray) {
3744:     PetscMalloc1(bs2,&barray);
3745:     baij->barray = barray;
3746:   }

3748:   if (roworiented) stepval = (n-1)*bs;
3749:   else stepval = (m-1)*bs;

3751:   for (i=0; i<m; i++) {
3752:     if (im[i] < 0) continue;
3753: #if defined(PETSC_USE_DEBUG)
3754:     if (im[i] >= baij->Mbs) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Row too large, row %D max %D",im[i],baij->Mbs-1);
3755: #endif
3756:     if (im[i] >= rstart && im[i] < rend) {
3757:       row = im[i] - rstart;
3758:       for (j=0; j<n; j++) {
3759:         /* If NumCol = 1 then a copy is not required */
3760:         if ((roworiented) && (n == 1)) {
3761:           barray = (MatScalar*)v + i*bs2;
3762:         } else if ((!roworiented) && (m == 1)) {
3763:           barray = (MatScalar*)v + j*bs2;
3764:         } else { /* Here a copy is required */
3765:           if (roworiented) {
3766:             value = v + i*(stepval+bs)*bs + j*bs;
3767:           } else {
3768:             value = v + j*(stepval+bs)*bs + i*bs;
3769:           }
3770:           for (ii=0; ii<bs; ii++,value+=stepval) {
3771:             for (jj=0; jj<bs; jj++) {
3772:               *barray++ = *value++;
3773:             }
3774:           }
3775:           barray -=bs2;
3776:         }

3778:         if (in[j] >= cstart && in[j] < cend) {
3779:           col  = in[j] - cstart;
3780:           MatSetValuesBlocked_SeqBAIJ(baij->A,1,&row,1,&col,barray,addv);
3781:         } else if (in[j] < 0) continue;
3782: #if defined(PETSC_USE_DEBUG)
3783:         else if (in[j] >= baij->Nbs) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Column too large, col %D max %D",in[j],baij->Nbs-1);
3784: #endif
3785:         else {
3786:           if (mat->was_assembled) {
3787:             if (!baij->colmap) {
3788:               MatCreateColmap_MPIBAIJ_Private(mat);
3789:             }

3791: #if defined(PETSC_USE_DEBUG)
3792: #if defined(PETSC_USE_CTABLE)
3793:             { PetscInt data;
3794:               PetscTableFind(baij->colmap,in[j]+1,&data);
3795:               if ((data - 1) % bs) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Incorrect colmap");
3796:             }
3797: #else
3798:             if ((baij->colmap[in[j]] - 1) % bs) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Incorrect colmap");
3799: #endif
3800: #endif
3801: #if defined(PETSC_USE_CTABLE)
3802:             PetscTableFind(baij->colmap,in[j]+1,&col);
3803:             col  = (col - 1)/bs;
3804: #else
3805:             col = (baij->colmap[in[j]] - 1)/bs;
3806: #endif
3807:             if (col < 0 && !((Mat_SeqBAIJ*)(baij->A->data))->nonew) {
3808:               MatDisAssemble_MPIBAIJ(mat);
3809:               col  =  in[j];
3810:             }
3811:           } else col = in[j];
3812:           MatSetValuesBlocked_SeqBAIJ(baij->B,1,&row,1,&col,barray,addv);
3813:         }
3814:       }
3815:     } else {
3816:       if (!baij->donotstash) {
3817:         if (roworiented) {
3818:           MatStashValuesRowBlocked_Private(&mat->bstash,im[i],n,in,v,m,n,i);
3819:         } else {
3820:           MatStashValuesColBlocked_Private(&mat->bstash,im[i],n,in,v,m,n,i);
3821:         }
3822:       }
3823:     }
3824:   }

3826:   /* task normally handled by MatSetValuesBlocked() */
3827:   PetscLogEventEnd(MAT_SetValues,mat,0,0,0);
3828:   return(0);
3829: }

3833: /*@
3834:      MatCreateMPIBAIJWithArrays - creates a MPI BAIJ matrix using arrays that contain in standard
3835:          CSR format the local rows.

3837:    Collective on MPI_Comm

3839:    Input Parameters:
3840: +  comm - MPI communicator
3841: .  bs - the block size, only a block size of 1 is supported
3842: .  m - number of local rows (Cannot be PETSC_DECIDE)
3843: .  n - This value should be the same as the local size used in creating the
3844:        x vector for the matrix-vector product y = Ax. (or PETSC_DECIDE to have
3845:        calculated if N is given) For square matrices n is almost always m.
3846: .  M - number of global rows (or PETSC_DETERMINE to have calculated if m is given)
3847: .  N - number of global columns (or PETSC_DETERMINE to have calculated if n is given)
3848: .   i - row indices
3849: .   j - column indices
3850: -   a - matrix values

3852:    Output Parameter:
3853: .   mat - the matrix

3855:    Level: intermediate

3857:    Notes:
3858:        The i, j, and a arrays ARE copied by this routine into the internal format used by PETSc;
3859:      thus you CANNOT change the matrix entries by changing the values of a[] after you have
3860:      called this routine. Use MatCreateMPIAIJWithSplitArrays() to avoid needing to copy the arrays.

3862:      The order of the entries in values is the same as the block compressed sparse row storage format; that is, it is
3863:      the same as a three dimensional array in Fortran values(bs,bs,nnz) that contains the first column of the first
3864:      block, followed by the second column of the first block etc etc.  That is, the blocks are contiguous in memory
3865:      with column-major ordering within blocks.

3867:        The i and j indices are 0 based, and i indices are indices corresponding to the local j array.

3869: .keywords: matrix, aij, compressed row, sparse, parallel

3871: .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatMPIAIJSetPreallocationCSR(),
3872:           MPIAIJ, MatCreateAIJ(), MatCreateMPIAIJWithSplitArrays()
3873: @*/
3874: PetscErrorCode  MatCreateMPIBAIJWithArrays(MPI_Comm comm,PetscInt bs,PetscInt m,PetscInt n,PetscInt M,PetscInt N,const PetscInt i[],const PetscInt j[],const PetscScalar a[],Mat *mat)
3875: {

3879:   if (i[0]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"i (row indices) must start with 0");
3880:   if (m < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"local number of rows (m) cannot be PETSC_DECIDE, or negative");
3881:   MatCreate(comm,mat);
3882:   MatSetSizes(*mat,m,n,M,N);
3883:   MatSetType(*mat,MATMPISBAIJ);
3884:   MatSetOption(*mat,MAT_ROW_ORIENTED,PETSC_FALSE);
3885:   MatMPIBAIJSetPreallocationCSR(*mat,bs,i,j,a);
3886:   MatSetOption(*mat,MAT_ROW_ORIENTED,PETSC_TRUE);
3887:   return(0);
3888: }