Actual source code: mmbaij.c

  1: /*$Id: mmbaij.c,v 1.46 2001/09/25 00:31:36 balay Exp $*/

  3: /*
  4:    Support for the parallel BAIJ matrix vector multiply
  5: */
 6:  #include src/mat/impls/baij/mpi/mpibaij.h
 7:  #include src/vec/vecimpl.h

  9: EXTERN int MatSetValuesBlocked_SeqBAIJ(Mat,int,const int[],int,const int[],const PetscScalar[],InsertMode);

 13: int MatSetUpMultiply_MPIBAIJ(Mat mat)
 14: {
 15:   Mat_MPIBAIJ        *baij = (Mat_MPIBAIJ*)mat->data;
 16:   Mat_SeqBAIJ        *B = (Mat_SeqBAIJ*)(baij->B->data);
 17:   int                Nbs = baij->Nbs,i,j,*indices,*aj = B->j,ierr,ec = 0,*garray;
 18:   int                bs = baij->bs,*stmp;
 19:   IS                 from,to;
 20:   Vec                gvec;
 21: #if defined (PETSC_USE_CTABLE)
 22:   PetscTable         gid1_lid1;
 23:   PetscTablePosition tpos;
 24:   int                gid,lid;
 25: #endif  


 29: #if defined (PETSC_USE_CTABLE)
 30:   /* use a table - Mark Adams */
 31:   PetscTableCreate(B->mbs,&gid1_lid1);
 32:   for (i=0; i<B->mbs; i++) {
 33:     for (j=0; j<B->ilen[i]; j++) {
 34:       int data,gid1 = aj[B->i[i]+j] + 1;
 35:       PetscTableFind(gid1_lid1,gid1,&data) ;
 36:       if (!data) {
 37:         /* one based table */
 38:         PetscTableAdd(gid1_lid1,gid1,++ec);
 39:       }
 40:     }
 41:   }
 42:   /* form array of columns we need */
 43:   PetscMalloc((ec+1)*sizeof(int),&garray);
 44:   PetscTableGetHeadPosition(gid1_lid1,&tpos);
 45:   while (tpos) {
 46:     PetscTableGetNext(gid1_lid1,&tpos,&gid,&lid);
 47:     gid--; lid--;
 48:     garray[lid] = gid;
 49:   }
 50:   PetscSortInt(ec,garray);
 51:   /* qsort(garray, ec, sizeof(int), intcomparcarc); */
 52:   PetscTableRemoveAll(gid1_lid1);
 53:   for (i=0; i<ec; i++) {
 54:     PetscTableAdd(gid1_lid1,garray[i]+1,i+1);
 55:   }
 56:   /* compact out the extra columns in B */
 57:   for (i=0; i<B->mbs; i++) {
 58:     for (j=0; j<B->ilen[i]; j++) {
 59:       int gid1 = aj[B->i[i] + j] + 1;
 60:       PetscTableFind(gid1_lid1,gid1,&lid);
 61:       lid --;
 62:       aj[B->i[i]+j] = lid;
 63:     }
 64:   }
 65:   B->nbs     = ec;
 66:   baij->B->n = ec*B->bs;
 67:   PetscTableDelete(gid1_lid1);
 68:   /* Mark Adams */
 69: #else
 70:   /* Make an array as long as the number of columns */
 71:   /* mark those columns that are in baij->B */
 72:   PetscMalloc((Nbs+1)*sizeof(int),&indices);
 73:   PetscMemzero(indices,Nbs*sizeof(int));
 74:   for (i=0; i<B->mbs; i++) {
 75:     for (j=0; j<B->ilen[i]; j++) {
 76:       if (!indices[aj[B->i[i] + j]]) ec++;
 77:       indices[aj[B->i[i] + j]] = 1;
 78:     }
 79:   }

 81:   /* form array of columns we need */
 82:   PetscMalloc((ec+1)*sizeof(int),&garray);
 83:   ec = 0;
 84:   for (i=0; i<Nbs; i++) {
 85:     if (indices[i]) {
 86:       garray[ec++] = i;
 87:     }
 88:   }

 90:   /* make indices now point into garray */
 91:   for (i=0; i<ec; i++) {
 92:     indices[garray[i]] = i;
 93:   }

 95:   /* compact out the extra columns in B */
 96:   for (i=0; i<B->mbs; i++) {
 97:     for (j=0; j<B->ilen[i]; j++) {
 98:       aj[B->i[i] + j] = indices[aj[B->i[i] + j]];
 99:     }
100:   }
101:   B->nbs       = ec;
102:   baij->B->n   = ec*B->bs;
103:   PetscFree(indices);
104: #endif  

106:   /* create local vector that is used to scatter into */
107:   VecCreateSeq(PETSC_COMM_SELF,ec*bs,&baij->lvec);

109:   /* create two temporary index sets for building scatter-gather */
110:   for (i=0; i<ec; i++) {
111:     garray[i] = bs*garray[i];
112:   }
113:   ISCreateBlock(PETSC_COMM_SELF,bs,ec,garray,&from);
114:   for (i=0; i<ec; i++) {
115:     garray[i] = garray[i]/bs;
116:   }

118:   PetscMalloc((ec+1)*sizeof(int),&stmp);
119:   for (i=0; i<ec; i++) { stmp[i] = bs*i; }
120:   ISCreateBlock(PETSC_COMM_SELF,bs,ec,stmp,&to);
121:   PetscFree(stmp);

123:   /* create temporary global vector to generate scatter context */
124:   /* this is inefficient, but otherwise we must do either 
125:      1) save garray until the first actual scatter when the vector is known or
126:      2) have another way of generating a scatter context without a vector.*/
127:   VecCreateMPI(mat->comm,mat->n,mat->N,&gvec);

129:   /* gnerate the scatter context */
130:   VecScatterCreate(gvec,from,baij->lvec,to,&baij->Mvctx);

132:   /*
133:       Post the receives for the first matrix vector product. We sync-chronize after
134:     this on the chance that the user immediately calls MatMult() after assemblying 
135:     the matrix.
136:   */
137:   VecScatterPostRecvs(gvec,baij->lvec,INSERT_VALUES,SCATTER_FORWARD,baij->Mvctx);
138:   MPI_Barrier(mat->comm);

140:   PetscLogObjectParent(mat,baij->Mvctx);
141:   PetscLogObjectParent(mat,baij->lvec);
142:   PetscLogObjectParent(mat,from);
143:   PetscLogObjectParent(mat,to);
144:   baij->garray = garray;
145:   PetscLogObjectMemory(mat,(ec+1)*sizeof(int));
146:   ISDestroy(from);
147:   ISDestroy(to);
148:   VecDestroy(gvec);
149:   return(0);
150: }

152: /*
153:      Takes the local part of an already assembled MPIBAIJ matrix
154:    and disassembles it. This is to allow new nonzeros into the matrix
155:    that require more communication in the matrix vector multiply. 
156:    Thus certain data-structures must be rebuilt.

158:    Kind of slow! But that's what application programmers get when 
159:    they are sloppy.
160: */
163: int DisAssemble_MPIBAIJ(Mat A)
164: {
165:   Mat_MPIBAIJ  *baij = (Mat_MPIBAIJ*)A->data;
166:   Mat          B = baij->B,Bnew;
167:   Mat_SeqBAIJ  *Bbaij = (Mat_SeqBAIJ*)B->data;
168:   int          ierr,i,j,mbs=Bbaij->mbs,n = A->N,col,*garray=baij->garray;
169:   int          bs2 = baij->bs2,*nz,ec,m = A->m;
170:   MatScalar    *a = Bbaij->a;
171:   PetscScalar  *atmp;
172: #if defined(PETSC_USE_MAT_SINGLE)
173:   int          k;
174: #endif

177:   /* free stuff related to matrix-vec multiply */
178:   VecGetSize(baij->lvec,&ec); /* needed for PetscLogObjectMemory below */
179:   VecDestroy(baij->lvec); baij->lvec = 0;
180:   VecScatterDestroy(baij->Mvctx); baij->Mvctx = 0;
181:   if (baij->colmap) {
182: #if defined (PETSC_USE_CTABLE)
183:     PetscTableDelete(baij->colmap); baij->colmap = 0;
184: #else
185:     PetscFree(baij->colmap);
186:     baij->colmap = 0;
187:     PetscLogObjectMemory(A,-Bbaij->nbs*sizeof(int));
188: #endif
189:   }

191:   /* make sure that B is assembled so we can access its values */
192:   MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);
193:   MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);

195:   /* invent new B and copy stuff over */
196:   PetscMalloc(mbs*sizeof(int),&nz);
197:   for (i=0; i<mbs; i++) {
198:     nz[i] = Bbaij->i[i+1]-Bbaij->i[i];
199:   }
200:   MatCreateSeqBAIJ(PETSC_COMM_SELF,baij->bs,m,n,0,nz,&Bnew);
201:   MatSetOption(Bnew,MAT_COLUMN_ORIENTED);

203: #if defined(PETSC_USE_MAT_SINGLE)
204:   PetscMalloc(bs2*sizeof(PetscScalar),&atmp);
205: #endif
206:     for (i=0; i<mbs; i++) {
207:       for (j=Bbaij->i[i]; j<Bbaij->i[i+1]; j++) {
208:         col  = garray[Bbaij->j[j]];
209: #if defined(PETSC_USE_MAT_SINGLE)
210:         for (k=0; k<bs2; k++) atmp[k] = a[j*bs2+k];
211: #else
212:         atmp = a + j*bs2;
213: #endif
214:         MatSetValuesBlocked_SeqBAIJ(Bnew,1,&i,1,&col,atmp,B->insertmode);
215:       }
216:     }
217:   MatSetOption(Bnew,MAT_ROW_ORIENTED);

219: #if defined(PETSC_USE_MAT_SINGLE)
220:   PetscFree(atmp);
221: #endif

223:   PetscFree(nz);
224:   PetscFree(baij->garray);
225:   baij->garray = 0;
226:   PetscLogObjectMemory(A,-ec*sizeof(int));
227:   MatDestroy(B);
228:   PetscLogObjectParent(A,Bnew);
229:   baij->B = Bnew;
230:   A->was_assembled = PETSC_FALSE;
231:   return(0);
232: }

234: /*      ugly stuff added for Glenn someday we should fix this up */

236: static int *uglyrmapd = 0,*uglyrmapo = 0;  /* mapping from the local ordering to the "diagonal" and "off-diagonal"
237:                                       parts of the local matrix */
238: static Vec uglydd = 0,uglyoo = 0;   /* work vectors used to scale the two parts of the local matrix */


243: int MatMPIBAIJDiagonalScaleLocalSetUp(Mat inA,Vec scale)
244: {
245:   Mat_MPIBAIJ  *ina = (Mat_MPIBAIJ*) inA->data; /*access private part of matrix */
246:   Mat_SeqBAIJ  *A = (Mat_SeqBAIJ*)ina->A->data;
247:   Mat_SeqBAIJ  *B = (Mat_SeqBAIJ*)ina->B->data;
248:   int          ierr,bs = A->bs,i,n,nt,j,cstart,cend,no,*garray = ina->garray,*lindices;
249:   int          *r_rmapd,*r_rmapo;
250: 
252:   MatGetOwnershipRange(inA,&cstart,&cend);
253:   MatGetSize(ina->A,PETSC_NULL,&n);
254:   PetscMalloc((inA->bmapping->n+1)*sizeof(int),&r_rmapd);
255:   PetscMemzero(r_rmapd,inA->bmapping->n*sizeof(int));
256:   nt   = 0;
257:   for (i=0; i<inA->bmapping->n; i++) {
258:     if (inA->bmapping->indices[i]*bs >= cstart && inA->bmapping->indices[i]*bs < cend) {
259:       nt++;
260:       r_rmapd[i] = inA->bmapping->indices[i] + 1;
261:     }
262:   }
263:   if (nt*bs != n) SETERRQ2(1,"Hmm nt*bs %d n %d",nt*bs,n);
264:   PetscMalloc((n+1)*sizeof(int),&uglyrmapd);
265:   for (i=0; i<inA->bmapping->n; i++) {
266:     if (r_rmapd[i]){
267:       for (j=0; j<bs; j++) {
268:         uglyrmapd[(r_rmapd[i]-1)*bs+j-cstart] = i*bs + j;
269:       }
270:     }
271:   }
272:   PetscFree(r_rmapd);
273:   VecCreateSeq(PETSC_COMM_SELF,n,&uglydd);

275:   PetscMalloc((ina->Nbs+1)*sizeof(int),&lindices);
276:   PetscMemzero(lindices,ina->Nbs*sizeof(int));
277:   for (i=0; i<B->nbs; i++) {
278:     lindices[garray[i]] = i+1;
279:   }
280:   no   = inA->bmapping->n - nt;
281:   PetscMalloc((inA->bmapping->n+1)*sizeof(int),&r_rmapo);
282:   PetscMemzero(r_rmapo,inA->bmapping->n*sizeof(int));
283:   nt   = 0;
284:   for (i=0; i<inA->bmapping->n; i++) {
285:     if (lindices[inA->bmapping->indices[i]]) {
286:       nt++;
287:       r_rmapo[i] = lindices[inA->bmapping->indices[i]];
288:     }
289:   }
290:   if (nt > no) SETERRQ2(1,"Hmm nt %d no %d",nt,n);
291:   PetscFree(lindices);
292:   PetscMalloc((nt*bs+1)*sizeof(int),&uglyrmapo);
293:   for (i=0; i<inA->bmapping->n; i++) {
294:     if (r_rmapo[i]){
295:       for (j=0; j<bs; j++) {
296:         uglyrmapo[(r_rmapo[i]-1)*bs+j] = i*bs + j;
297:       }
298:     }
299:   }
300:   PetscFree(r_rmapo);
301:   VecCreateSeq(PETSC_COMM_SELF,nt*bs,&uglyoo);

303:   return(0);
304: }

308: int MatMPIBAIJDiagonalScaleLocal(Mat A,Vec scale)
309: {
310:   /* This routine should really be abandoned as it duplicates MatDiagonalScaleLocal */
311:   int ierr,(*f)(Mat,Vec);

314:   PetscObjectQueryFunction((PetscObject)A,"MatDiagonalScaleLocal_C",(void (**)(void))&f);
315:   if (f) {
316:     (*f)(A,scale);
317:   }
318:   return(0);
319: }

321: EXTERN_C_BEGIN
324: int MatDiagonalScaleLocal_MPIBAIJ(Mat A,Vec scale)
325: {
326:   Mat_MPIBAIJ  *a = (Mat_MPIBAIJ*) A->data; /*access private part of matrix */
327:   int          ierr,n,i;
328:   PetscScalar  *d,*o,*s;
329: 
331:   if (!uglyrmapd) {
332:     MatMPIBAIJDiagonalScaleLocalSetUp(A,scale);
333:   }

335:   VecGetArray(scale,&s);
336: 
337:   VecGetLocalSize(uglydd,&n);
338:   VecGetArray(uglydd,&d);
339:   for (i=0; i<n; i++) {
340:     d[i] = s[uglyrmapd[i]]; /* copy "diagonal" (true local) portion of scale into dd vector */
341:   }
342:   VecRestoreArray(uglydd,&d);
343:   /* column scale "diagonal" portion of local matrix */
344:   MatDiagonalScale(a->A,PETSC_NULL,uglydd);

346:   VecGetLocalSize(uglyoo,&n);
347:   VecGetArray(uglyoo,&o);
348:   for (i=0; i<n; i++) {
349:     o[i] = s[uglyrmapo[i]]; /* copy "off-diagonal" portion of scale into oo vector */
350:   }
351:   VecRestoreArray(scale,&s);
352:   VecRestoreArray(uglyoo,&o);
353:   /* column scale "off-diagonal" portion of local matrix */
354:   MatDiagonalScale(a->B,PETSC_NULL,uglyoo);

356:   return(0);
357: }
358: EXTERN_C_END