1 2 3 #include <../src/mat/impls/aij/mpi/mpiaij.h> /*I "petscmat.h" I*/ 4 #include <petsc/private/vecimpl.h> 5 #include <petsc/private/isimpl.h> 6 #include <petscblaslapack.h> 7 #include <petscsf.h> 8 9 /*MC 10 MATAIJ - MATAIJ = "aij" - A matrix type to be used for sparse matrices. 11 12 This matrix type is identical to MATSEQAIJ when constructed with a single process communicator, 13 and MATMPIAIJ otherwise. As a result, for single process communicators, 14 MatSeqAIJSetPreallocation is supported, and similarly MatMPIAIJSetPreallocation is supported 15 for communicators controlling multiple processes. It is recommended that you call both of 16 the above preallocation routines for simplicity. 17 18 Options Database Keys: 19 . -mat_type aij - sets the matrix type to "aij" during a call to MatSetFromOptions() 20 21 Developer Notes: Subclasses include MATAIJCUSP, MATAIJCUSPARSE, MATAIJPERM, MATAIJCRL, and also automatically switches over to use inodes when 22 enough exist. 23 24 Level: beginner 25 26 .seealso: MatCreateAIJ(), MatCreateSeqAIJ(), MATSEQAIJ, MATMPIAIJ 27 M*/ 28 29 /*MC 30 MATAIJCRL - MATAIJCRL = "aijcrl" - A matrix type to be used for sparse matrices. 31 32 This matrix type is identical to MATSEQAIJCRL when constructed with a single process communicator, 33 and MATMPIAIJCRL otherwise. As a result, for single process communicators, 34 MatSeqAIJSetPreallocation() is supported, and similarly MatMPIAIJSetPreallocation() is supported 35 for communicators controlling multiple processes. It is recommended that you call both of 36 the above preallocation routines for simplicity. 37 38 Options Database Keys: 39 . -mat_type aijcrl - sets the matrix type to "aijcrl" during a call to MatSetFromOptions() 40 41 Level: beginner 42 43 .seealso: MatCreateMPIAIJCRL,MATSEQAIJCRL,MATMPIAIJCRL, MATSEQAIJCRL, MATMPIAIJCRL 44 M*/ 45 46 PetscErrorCode MatSetBlockSizes_MPIAIJ(Mat M, PetscInt rbs, PetscInt cbs) 47 { 48 PetscErrorCode ierr; 49 Mat_MPIAIJ *mat = (Mat_MPIAIJ*)M->data; 50 51 PetscFunctionBegin; 52 if (mat->A) { 53 ierr = MatSetBlockSizes(mat->A,rbs,cbs);CHKERRQ(ierr); 54 ierr = MatSetBlockSizes(mat->B,rbs,1);CHKERRQ(ierr); 55 } 56 PetscFunctionReturn(0); 57 } 58 59 PetscErrorCode MatFindNonzeroRows_MPIAIJ(Mat M,IS *keptrows) 60 { 61 PetscErrorCode ierr; 62 Mat_MPIAIJ *mat = (Mat_MPIAIJ*)M->data; 63 Mat_SeqAIJ *a = (Mat_SeqAIJ*)mat->A->data; 64 Mat_SeqAIJ *b = (Mat_SeqAIJ*)mat->B->data; 65 const PetscInt *ia,*ib; 66 const MatScalar *aa,*bb; 67 PetscInt na,nb,i,j,*rows,cnt=0,n0rows; 68 PetscInt m = M->rmap->n,rstart = M->rmap->rstart; 69 70 PetscFunctionBegin; 71 *keptrows = 0; 72 ia = a->i; 73 ib = b->i; 74 for (i=0; i<m; i++) { 75 na = ia[i+1] - ia[i]; 76 nb = ib[i+1] - ib[i]; 77 if (!na && !nb) { 78 cnt++; 79 goto ok1; 80 } 81 aa = a->a + ia[i]; 82 for (j=0; j<na; j++) { 83 if (aa[j] != 0.0) goto ok1; 84 } 85 bb = b->a + ib[i]; 86 for (j=0; j <nb; j++) { 87 if (bb[j] != 0.0) goto ok1; 88 } 89 cnt++; 90 ok1:; 91 } 92 ierr = MPIU_Allreduce(&cnt,&n0rows,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)M));CHKERRQ(ierr); 93 if (!n0rows) PetscFunctionReturn(0); 94 ierr = PetscMalloc1(M->rmap->n-cnt,&rows);CHKERRQ(ierr); 95 cnt = 0; 96 for (i=0; i<m; i++) { 97 na = ia[i+1] - ia[i]; 98 nb = ib[i+1] - ib[i]; 99 if (!na && !nb) continue; 100 aa = a->a + ia[i]; 101 for (j=0; j<na;j++) { 102 if (aa[j] != 0.0) { 103 rows[cnt++] = rstart + i; 104 goto ok2; 105 } 106 } 107 bb = b->a + ib[i]; 108 for (j=0; j<nb; j++) { 109 if (bb[j] != 0.0) { 110 rows[cnt++] = rstart + i; 111 goto ok2; 112 } 113 } 114 ok2:; 115 } 116 ierr = ISCreateGeneral(PetscObjectComm((PetscObject)M),cnt,rows,PETSC_OWN_POINTER,keptrows);CHKERRQ(ierr); 117 PetscFunctionReturn(0); 118 } 119 120 PetscErrorCode MatDiagonalSet_MPIAIJ(Mat Y,Vec D,InsertMode is) 121 { 122 PetscErrorCode ierr; 123 Mat_MPIAIJ *aij = (Mat_MPIAIJ*) Y->data; 124 125 PetscFunctionBegin; 126 if (Y->assembled && Y->rmap->rstart == Y->cmap->rstart && Y->rmap->rend == Y->cmap->rend) { 127 ierr = MatDiagonalSet(aij->A,D,is);CHKERRQ(ierr); 128 } else { 129 ierr = MatDiagonalSet_Default(Y,D,is);CHKERRQ(ierr); 130 } 131 PetscFunctionReturn(0); 132 } 133 134 135 PetscErrorCode MatFindZeroDiagonals_MPIAIJ(Mat M,IS *zrows) 136 { 137 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)M->data; 138 PetscErrorCode ierr; 139 PetscInt i,rstart,nrows,*rows; 140 141 PetscFunctionBegin; 142 *zrows = NULL; 143 ierr = MatFindZeroDiagonals_SeqAIJ_Private(aij->A,&nrows,&rows);CHKERRQ(ierr); 144 ierr = MatGetOwnershipRange(M,&rstart,NULL);CHKERRQ(ierr); 145 for (i=0; i<nrows; i++) rows[i] += rstart; 146 ierr = ISCreateGeneral(PetscObjectComm((PetscObject)M),nrows,rows,PETSC_OWN_POINTER,zrows);CHKERRQ(ierr); 147 PetscFunctionReturn(0); 148 } 149 150 PetscErrorCode MatGetColumnNorms_MPIAIJ(Mat A,NormType type,PetscReal *norms) 151 { 152 PetscErrorCode ierr; 153 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)A->data; 154 PetscInt i,n,*garray = aij->garray; 155 Mat_SeqAIJ *a_aij = (Mat_SeqAIJ*) aij->A->data; 156 Mat_SeqAIJ *b_aij = (Mat_SeqAIJ*) aij->B->data; 157 PetscReal *work; 158 159 PetscFunctionBegin; 160 ierr = MatGetSize(A,NULL,&n);CHKERRQ(ierr); 161 ierr = PetscCalloc1(n,&work);CHKERRQ(ierr); 162 if (type == NORM_2) { 163 for (i=0; i<a_aij->i[aij->A->rmap->n]; i++) { 164 work[A->cmap->rstart + a_aij->j[i]] += PetscAbsScalar(a_aij->a[i]*a_aij->a[i]); 165 } 166 for (i=0; i<b_aij->i[aij->B->rmap->n]; i++) { 167 work[garray[b_aij->j[i]]] += PetscAbsScalar(b_aij->a[i]*b_aij->a[i]); 168 } 169 } else if (type == NORM_1) { 170 for (i=0; i<a_aij->i[aij->A->rmap->n]; i++) { 171 work[A->cmap->rstart + a_aij->j[i]] += PetscAbsScalar(a_aij->a[i]); 172 } 173 for (i=0; i<b_aij->i[aij->B->rmap->n]; i++) { 174 work[garray[b_aij->j[i]]] += PetscAbsScalar(b_aij->a[i]); 175 } 176 } else if (type == NORM_INFINITY) { 177 for (i=0; i<a_aij->i[aij->A->rmap->n]; i++) { 178 work[A->cmap->rstart + a_aij->j[i]] = PetscMax(PetscAbsScalar(a_aij->a[i]), work[A->cmap->rstart + a_aij->j[i]]); 179 } 180 for (i=0; i<b_aij->i[aij->B->rmap->n]; i++) { 181 work[garray[b_aij->j[i]]] = PetscMax(PetscAbsScalar(b_aij->a[i]),work[garray[b_aij->j[i]]]); 182 } 183 184 } else SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_WRONG,"Unknown NormType"); 185 if (type == NORM_INFINITY) { 186 ierr = MPIU_Allreduce(work,norms,n,MPIU_REAL,MPIU_MAX,PetscObjectComm((PetscObject)A));CHKERRQ(ierr); 187 } else { 188 ierr = MPIU_Allreduce(work,norms,n,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)A));CHKERRQ(ierr); 189 } 190 ierr = PetscFree(work);CHKERRQ(ierr); 191 if (type == NORM_2) { 192 for (i=0; i<n; i++) norms[i] = PetscSqrtReal(norms[i]); 193 } 194 PetscFunctionReturn(0); 195 } 196 197 PetscErrorCode MatFindOffBlockDiagonalEntries_MPIAIJ(Mat A,IS *is) 198 { 199 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 200 IS sis,gis; 201 PetscErrorCode ierr; 202 const PetscInt *isis,*igis; 203 PetscInt n,*iis,nsis,ngis,rstart,i; 204 205 PetscFunctionBegin; 206 ierr = MatFindOffBlockDiagonalEntries(a->A,&sis);CHKERRQ(ierr); 207 ierr = MatFindNonzeroRows(a->B,&gis);CHKERRQ(ierr); 208 ierr = ISGetSize(gis,&ngis);CHKERRQ(ierr); 209 ierr = ISGetSize(sis,&nsis);CHKERRQ(ierr); 210 ierr = ISGetIndices(sis,&isis);CHKERRQ(ierr); 211 ierr = ISGetIndices(gis,&igis);CHKERRQ(ierr); 212 213 ierr = PetscMalloc1(ngis+nsis,&iis);CHKERRQ(ierr); 214 ierr = PetscMemcpy(iis,igis,ngis*sizeof(PetscInt));CHKERRQ(ierr); 215 ierr = PetscMemcpy(iis+ngis,isis,nsis*sizeof(PetscInt));CHKERRQ(ierr); 216 n = ngis + nsis; 217 ierr = PetscSortRemoveDupsInt(&n,iis);CHKERRQ(ierr); 218 ierr = MatGetOwnershipRange(A,&rstart,NULL);CHKERRQ(ierr); 219 for (i=0; i<n; i++) iis[i] += rstart; 220 ierr = ISCreateGeneral(PetscObjectComm((PetscObject)A),n,iis,PETSC_OWN_POINTER,is);CHKERRQ(ierr); 221 222 ierr = ISRestoreIndices(sis,&isis);CHKERRQ(ierr); 223 ierr = ISRestoreIndices(gis,&igis);CHKERRQ(ierr); 224 ierr = ISDestroy(&sis);CHKERRQ(ierr); 225 ierr = ISDestroy(&gis);CHKERRQ(ierr); 226 PetscFunctionReturn(0); 227 } 228 229 /* 230 Distributes a SeqAIJ matrix across a set of processes. Code stolen from 231 MatLoad_MPIAIJ(). Horrible lack of reuse. Should be a routine for each matrix type. 232 233 Only for square matrices 234 235 Used by a preconditioner, hence PETSC_EXTERN 236 */ 237 PETSC_EXTERN PetscErrorCode MatDistribute_MPIAIJ(MPI_Comm comm,Mat gmat,PetscInt m,MatReuse reuse,Mat *inmat) 238 { 239 PetscMPIInt rank,size; 240 PetscInt *rowners,*dlens,*olens,i,rstart,rend,j,jj,nz = 0,*gmataj,cnt,row,*ld,bses[2]; 241 PetscErrorCode ierr; 242 Mat mat; 243 Mat_SeqAIJ *gmata; 244 PetscMPIInt tag; 245 MPI_Status status; 246 PetscBool aij; 247 MatScalar *gmataa,*ao,*ad,*gmataarestore=0; 248 249 PetscFunctionBegin; 250 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 251 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 252 if (!rank) { 253 ierr = PetscObjectTypeCompare((PetscObject)gmat,MATSEQAIJ,&aij);CHKERRQ(ierr); 254 if (!aij) SETERRQ1(PetscObjectComm((PetscObject)gmat),PETSC_ERR_SUP,"Currently no support for input matrix of type %s\n",((PetscObject)gmat)->type_name); 255 } 256 if (reuse == MAT_INITIAL_MATRIX) { 257 ierr = MatCreate(comm,&mat);CHKERRQ(ierr); 258 ierr = MatSetSizes(mat,m,m,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 259 ierr = MatGetBlockSizes(gmat,&bses[0],&bses[1]);CHKERRQ(ierr); 260 ierr = MPI_Bcast(bses,2,MPIU_INT,0,comm);CHKERRQ(ierr); 261 ierr = MatSetBlockSizes(mat,bses[0],bses[1]);CHKERRQ(ierr); 262 ierr = MatSetType(mat,MATAIJ);CHKERRQ(ierr); 263 ierr = PetscMalloc1(size+1,&rowners);CHKERRQ(ierr); 264 ierr = PetscMalloc2(m,&dlens,m,&olens);CHKERRQ(ierr); 265 ierr = MPI_Allgather(&m,1,MPIU_INT,rowners+1,1,MPIU_INT,comm);CHKERRQ(ierr); 266 267 rowners[0] = 0; 268 for (i=2; i<=size; i++) rowners[i] += rowners[i-1]; 269 rstart = rowners[rank]; 270 rend = rowners[rank+1]; 271 ierr = PetscObjectGetNewTag((PetscObject)mat,&tag);CHKERRQ(ierr); 272 if (!rank) { 273 gmata = (Mat_SeqAIJ*) gmat->data; 274 /* send row lengths to all processors */ 275 for (i=0; i<m; i++) dlens[i] = gmata->ilen[i]; 276 for (i=1; i<size; i++) { 277 ierr = MPI_Send(gmata->ilen + rowners[i],rowners[i+1]-rowners[i],MPIU_INT,i,tag,comm);CHKERRQ(ierr); 278 } 279 /* determine number diagonal and off-diagonal counts */ 280 ierr = PetscMemzero(olens,m*sizeof(PetscInt));CHKERRQ(ierr); 281 ierr = PetscCalloc1(m,&ld);CHKERRQ(ierr); 282 jj = 0; 283 for (i=0; i<m; i++) { 284 for (j=0; j<dlens[i]; j++) { 285 if (gmata->j[jj] < rstart) ld[i]++; 286 if (gmata->j[jj] < rstart || gmata->j[jj] >= rend) olens[i]++; 287 jj++; 288 } 289 } 290 /* send column indices to other processes */ 291 for (i=1; i<size; i++) { 292 nz = gmata->i[rowners[i+1]]-gmata->i[rowners[i]]; 293 ierr = MPI_Send(&nz,1,MPIU_INT,i,tag,comm);CHKERRQ(ierr); 294 ierr = MPI_Send(gmata->j + gmata->i[rowners[i]],nz,MPIU_INT,i,tag,comm);CHKERRQ(ierr); 295 } 296 297 /* send numerical values to other processes */ 298 for (i=1; i<size; i++) { 299 nz = gmata->i[rowners[i+1]]-gmata->i[rowners[i]]; 300 ierr = MPI_Send(gmata->a + gmata->i[rowners[i]],nz,MPIU_SCALAR,i,tag,comm);CHKERRQ(ierr); 301 } 302 gmataa = gmata->a; 303 gmataj = gmata->j; 304 305 } else { 306 /* receive row lengths */ 307 ierr = MPI_Recv(dlens,m,MPIU_INT,0,tag,comm,&status);CHKERRQ(ierr); 308 /* receive column indices */ 309 ierr = MPI_Recv(&nz,1,MPIU_INT,0,tag,comm,&status);CHKERRQ(ierr); 310 ierr = PetscMalloc2(nz,&gmataa,nz,&gmataj);CHKERRQ(ierr); 311 ierr = MPI_Recv(gmataj,nz,MPIU_INT,0,tag,comm,&status);CHKERRQ(ierr); 312 /* determine number diagonal and off-diagonal counts */ 313 ierr = PetscMemzero(olens,m*sizeof(PetscInt));CHKERRQ(ierr); 314 ierr = PetscCalloc1(m,&ld);CHKERRQ(ierr); 315 jj = 0; 316 for (i=0; i<m; i++) { 317 for (j=0; j<dlens[i]; j++) { 318 if (gmataj[jj] < rstart) ld[i]++; 319 if (gmataj[jj] < rstart || gmataj[jj] >= rend) olens[i]++; 320 jj++; 321 } 322 } 323 /* receive numerical values */ 324 ierr = PetscMemzero(gmataa,nz*sizeof(PetscScalar));CHKERRQ(ierr); 325 ierr = MPI_Recv(gmataa,nz,MPIU_SCALAR,0,tag,comm,&status);CHKERRQ(ierr); 326 } 327 /* set preallocation */ 328 for (i=0; i<m; i++) { 329 dlens[i] -= olens[i]; 330 } 331 ierr = MatSeqAIJSetPreallocation(mat,0,dlens);CHKERRQ(ierr); 332 ierr = MatMPIAIJSetPreallocation(mat,0,dlens,0,olens);CHKERRQ(ierr); 333 334 for (i=0; i<m; i++) { 335 dlens[i] += olens[i]; 336 } 337 cnt = 0; 338 for (i=0; i<m; i++) { 339 row = rstart + i; 340 ierr = MatSetValues(mat,1,&row,dlens[i],gmataj+cnt,gmataa+cnt,INSERT_VALUES);CHKERRQ(ierr); 341 cnt += dlens[i]; 342 } 343 if (rank) { 344 ierr = PetscFree2(gmataa,gmataj);CHKERRQ(ierr); 345 } 346 ierr = PetscFree2(dlens,olens);CHKERRQ(ierr); 347 ierr = PetscFree(rowners);CHKERRQ(ierr); 348 349 ((Mat_MPIAIJ*)(mat->data))->ld = ld; 350 351 *inmat = mat; 352 } else { /* column indices are already set; only need to move over numerical values from process 0 */ 353 Mat_SeqAIJ *Ad = (Mat_SeqAIJ*)((Mat_MPIAIJ*)((*inmat)->data))->A->data; 354 Mat_SeqAIJ *Ao = (Mat_SeqAIJ*)((Mat_MPIAIJ*)((*inmat)->data))->B->data; 355 mat = *inmat; 356 ierr = PetscObjectGetNewTag((PetscObject)mat,&tag);CHKERRQ(ierr); 357 if (!rank) { 358 /* send numerical values to other processes */ 359 gmata = (Mat_SeqAIJ*) gmat->data; 360 ierr = MatGetOwnershipRanges(mat,(const PetscInt**)&rowners);CHKERRQ(ierr); 361 gmataa = gmata->a; 362 for (i=1; i<size; i++) { 363 nz = gmata->i[rowners[i+1]]-gmata->i[rowners[i]]; 364 ierr = MPI_Send(gmataa + gmata->i[rowners[i]],nz,MPIU_SCALAR,i,tag,comm);CHKERRQ(ierr); 365 } 366 nz = gmata->i[rowners[1]]-gmata->i[rowners[0]]; 367 } else { 368 /* receive numerical values from process 0*/ 369 nz = Ad->nz + Ao->nz; 370 ierr = PetscMalloc1(nz,&gmataa);CHKERRQ(ierr); gmataarestore = gmataa; 371 ierr = MPI_Recv(gmataa,nz,MPIU_SCALAR,0,tag,comm,&status);CHKERRQ(ierr); 372 } 373 /* transfer numerical values into the diagonal A and off diagonal B parts of mat */ 374 ld = ((Mat_MPIAIJ*)(mat->data))->ld; 375 ad = Ad->a; 376 ao = Ao->a; 377 if (mat->rmap->n) { 378 i = 0; 379 nz = ld[i]; ierr = PetscMemcpy(ao,gmataa,nz*sizeof(PetscScalar));CHKERRQ(ierr); ao += nz; gmataa += nz; 380 nz = Ad->i[i+1] - Ad->i[i]; ierr = PetscMemcpy(ad,gmataa,nz*sizeof(PetscScalar));CHKERRQ(ierr); ad += nz; gmataa += nz; 381 } 382 for (i=1; i<mat->rmap->n; i++) { 383 nz = Ao->i[i] - Ao->i[i-1] - ld[i-1] + ld[i]; ierr = PetscMemcpy(ao,gmataa,nz*sizeof(PetscScalar));CHKERRQ(ierr); ao += nz; gmataa += nz; 384 nz = Ad->i[i+1] - Ad->i[i]; ierr = PetscMemcpy(ad,gmataa,nz*sizeof(PetscScalar));CHKERRQ(ierr); ad += nz; gmataa += nz; 385 } 386 i--; 387 if (mat->rmap->n) { 388 nz = Ao->i[i+1] - Ao->i[i] - ld[i]; ierr = PetscMemcpy(ao,gmataa,nz*sizeof(PetscScalar));CHKERRQ(ierr); 389 } 390 if (rank) { 391 ierr = PetscFree(gmataarestore);CHKERRQ(ierr); 392 } 393 } 394 ierr = MatAssemblyBegin(mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 395 ierr = MatAssemblyEnd(mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 396 PetscFunctionReturn(0); 397 } 398 399 /* 400 Local utility routine that creates a mapping from the global column 401 number to the local number in the off-diagonal part of the local 402 storage of the matrix. When PETSC_USE_CTABLE is used this is scalable at 403 a slightly higher hash table cost; without it it is not scalable (each processor 404 has an order N integer array but is fast to acess. 405 */ 406 PetscErrorCode MatCreateColmap_MPIAIJ_Private(Mat mat) 407 { 408 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 409 PetscErrorCode ierr; 410 PetscInt n = aij->B->cmap->n,i; 411 412 PetscFunctionBegin; 413 if (!aij->garray) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"MPIAIJ Matrix was assembled but is missing garray"); 414 #if defined(PETSC_USE_CTABLE) 415 ierr = PetscTableCreate(n,mat->cmap->N+1,&aij->colmap);CHKERRQ(ierr); 416 for (i=0; i<n; i++) { 417 ierr = PetscTableAdd(aij->colmap,aij->garray[i]+1,i+1,INSERT_VALUES);CHKERRQ(ierr); 418 } 419 #else 420 ierr = PetscCalloc1(mat->cmap->N+1,&aij->colmap);CHKERRQ(ierr); 421 ierr = PetscLogObjectMemory((PetscObject)mat,(mat->cmap->N+1)*sizeof(PetscInt));CHKERRQ(ierr); 422 for (i=0; i<n; i++) aij->colmap[aij->garray[i]] = i+1; 423 #endif 424 PetscFunctionReturn(0); 425 } 426 427 #define MatSetValues_SeqAIJ_A_Private(row,col,value,addv,orow,ocol) \ 428 { \ 429 if (col <= lastcol1) low1 = 0; \ 430 else high1 = nrow1; \ 431 lastcol1 = col;\ 432 while (high1-low1 > 5) { \ 433 t = (low1+high1)/2; \ 434 if (rp1[t] > col) high1 = t; \ 435 else low1 = t; \ 436 } \ 437 for (_i=low1; _i<high1; _i++) { \ 438 if (rp1[_i] > col) break; \ 439 if (rp1[_i] == col) { \ 440 if (addv == ADD_VALUES) ap1[_i] += value; \ 441 else ap1[_i] = value; \ 442 goto a_noinsert; \ 443 } \ 444 } \ 445 if (value == 0.0 && ignorezeroentries && row != col) {low1 = 0; high1 = nrow1;goto a_noinsert;} \ 446 if (nonew == 1) {low1 = 0; high1 = nrow1; goto a_noinsert;} \ 447 if (nonew == -1) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero at global row/column (%D, %D) into matrix", orow, ocol); \ 448 MatSeqXAIJReallocateAIJ(A,am,1,nrow1,row,col,rmax1,aa,ai,aj,rp1,ap1,aimax,nonew,MatScalar); \ 449 N = nrow1++ - 1; a->nz++; high1++; \ 450 /* shift up all the later entries in this row */ \ 451 for (ii=N; ii>=_i; ii--) { \ 452 rp1[ii+1] = rp1[ii]; \ 453 ap1[ii+1] = ap1[ii]; \ 454 } \ 455 rp1[_i] = col; \ 456 ap1[_i] = value; \ 457 A->nonzerostate++;\ 458 a_noinsert: ; \ 459 ailen[row] = nrow1; \ 460 } 461 462 463 #define MatSetValues_SeqAIJ_B_Private(row,col,value,addv,orow,ocol) \ 464 { \ 465 if (col <= lastcol2) low2 = 0; \ 466 else high2 = nrow2; \ 467 lastcol2 = col; \ 468 while (high2-low2 > 5) { \ 469 t = (low2+high2)/2; \ 470 if (rp2[t] > col) high2 = t; \ 471 else low2 = t; \ 472 } \ 473 for (_i=low2; _i<high2; _i++) { \ 474 if (rp2[_i] > col) break; \ 475 if (rp2[_i] == col) { \ 476 if (addv == ADD_VALUES) ap2[_i] += value; \ 477 else ap2[_i] = value; \ 478 goto b_noinsert; \ 479 } \ 480 } \ 481 if (value == 0.0 && ignorezeroentries) {low2 = 0; high2 = nrow2; goto b_noinsert;} \ 482 if (nonew == 1) {low2 = 0; high2 = nrow2; goto b_noinsert;} \ 483 if (nonew == -1) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero at global row/column (%D, %D) into matrix", orow, ocol); \ 484 MatSeqXAIJReallocateAIJ(B,bm,1,nrow2,row,col,rmax2,ba,bi,bj,rp2,ap2,bimax,nonew,MatScalar); \ 485 N = nrow2++ - 1; b->nz++; high2++; \ 486 /* shift up all the later entries in this row */ \ 487 for (ii=N; ii>=_i; ii--) { \ 488 rp2[ii+1] = rp2[ii]; \ 489 ap2[ii+1] = ap2[ii]; \ 490 } \ 491 rp2[_i] = col; \ 492 ap2[_i] = value; \ 493 B->nonzerostate++; \ 494 b_noinsert: ; \ 495 bilen[row] = nrow2; \ 496 } 497 498 PetscErrorCode MatSetValuesRow_MPIAIJ(Mat A,PetscInt row,const PetscScalar v[]) 499 { 500 Mat_MPIAIJ *mat = (Mat_MPIAIJ*)A->data; 501 Mat_SeqAIJ *a = (Mat_SeqAIJ*)mat->A->data,*b = (Mat_SeqAIJ*)mat->B->data; 502 PetscErrorCode ierr; 503 PetscInt l,*garray = mat->garray,diag; 504 505 PetscFunctionBegin; 506 /* code only works for square matrices A */ 507 508 /* find size of row to the left of the diagonal part */ 509 ierr = MatGetOwnershipRange(A,&diag,0);CHKERRQ(ierr); 510 row = row - diag; 511 for (l=0; l<b->i[row+1]-b->i[row]; l++) { 512 if (garray[b->j[b->i[row]+l]] > diag) break; 513 } 514 ierr = PetscMemcpy(b->a+b->i[row],v,l*sizeof(PetscScalar));CHKERRQ(ierr); 515 516 /* diagonal part */ 517 ierr = PetscMemcpy(a->a+a->i[row],v+l,(a->i[row+1]-a->i[row])*sizeof(PetscScalar));CHKERRQ(ierr); 518 519 /* right of diagonal part */ 520 ierr = PetscMemcpy(b->a+b->i[row]+l,v+l+a->i[row+1]-a->i[row],(b->i[row+1]-b->i[row]-l)*sizeof(PetscScalar));CHKERRQ(ierr); 521 PetscFunctionReturn(0); 522 } 523 524 PetscErrorCode MatSetValues_MPIAIJ(Mat mat,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],const PetscScalar v[],InsertMode addv) 525 { 526 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 527 PetscScalar value; 528 PetscErrorCode ierr; 529 PetscInt i,j,rstart = mat->rmap->rstart,rend = mat->rmap->rend; 530 PetscInt cstart = mat->cmap->rstart,cend = mat->cmap->rend,row,col; 531 PetscBool roworiented = aij->roworiented; 532 533 /* Some Variables required in the macro */ 534 Mat A = aij->A; 535 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 536 PetscInt *aimax = a->imax,*ai = a->i,*ailen = a->ilen,*aj = a->j; 537 MatScalar *aa = a->a; 538 PetscBool ignorezeroentries = a->ignorezeroentries; 539 Mat B = aij->B; 540 Mat_SeqAIJ *b = (Mat_SeqAIJ*)B->data; 541 PetscInt *bimax = b->imax,*bi = b->i,*bilen = b->ilen,*bj = b->j,bm = aij->B->rmap->n,am = aij->A->rmap->n; 542 MatScalar *ba = b->a; 543 544 PetscInt *rp1,*rp2,ii,nrow1,nrow2,_i,rmax1,rmax2,N,low1,high1,low2,high2,t,lastcol1,lastcol2; 545 PetscInt nonew; 546 MatScalar *ap1,*ap2; 547 548 PetscFunctionBegin; 549 for (i=0; i<m; i++) { 550 if (im[i] < 0) continue; 551 #if defined(PETSC_USE_DEBUG) 552 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); 553 #endif 554 if (im[i] >= rstart && im[i] < rend) { 555 row = im[i] - rstart; 556 lastcol1 = -1; 557 rp1 = aj + ai[row]; 558 ap1 = aa + ai[row]; 559 rmax1 = aimax[row]; 560 nrow1 = ailen[row]; 561 low1 = 0; 562 high1 = nrow1; 563 lastcol2 = -1; 564 rp2 = bj + bi[row]; 565 ap2 = ba + bi[row]; 566 rmax2 = bimax[row]; 567 nrow2 = bilen[row]; 568 low2 = 0; 569 high2 = nrow2; 570 571 for (j=0; j<n; j++) { 572 if (roworiented) value = v[i*n+j]; 573 else value = v[i+j*m]; 574 if (in[j] >= cstart && in[j] < cend) { 575 col = in[j] - cstart; 576 nonew = a->nonew; 577 if (ignorezeroentries && value == 0.0 && (addv == ADD_VALUES) && row != col) continue; 578 MatSetValues_SeqAIJ_A_Private(row,col,value,addv,im[i],in[j]); 579 } else if (in[j] < 0) continue; 580 #if defined(PETSC_USE_DEBUG) 581 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); 582 #endif 583 else { 584 if (mat->was_assembled) { 585 if (!aij->colmap) { 586 ierr = MatCreateColmap_MPIAIJ_Private(mat);CHKERRQ(ierr); 587 } 588 #if defined(PETSC_USE_CTABLE) 589 ierr = PetscTableFind(aij->colmap,in[j]+1,&col);CHKERRQ(ierr); 590 col--; 591 #else 592 col = aij->colmap[in[j]] - 1; 593 #endif 594 if (col < 0 && !((Mat_SeqAIJ*)(aij->B->data))->nonew) { 595 ierr = MatDisAssemble_MPIAIJ(mat);CHKERRQ(ierr); 596 col = in[j]; 597 /* Reinitialize the variables required by MatSetValues_SeqAIJ_B_Private() */ 598 B = aij->B; 599 b = (Mat_SeqAIJ*)B->data; 600 bimax = b->imax; bi = b->i; bilen = b->ilen; bj = b->j; ba = b->a; 601 rp2 = bj + bi[row]; 602 ap2 = ba + bi[row]; 603 rmax2 = bimax[row]; 604 nrow2 = bilen[row]; 605 low2 = 0; 606 high2 = nrow2; 607 bm = aij->B->rmap->n; 608 ba = b->a; 609 } else if (col < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero at global row/column (%D, %D) into matrix", im[i], in[j]); 610 } else col = in[j]; 611 nonew = b->nonew; 612 MatSetValues_SeqAIJ_B_Private(row,col,value,addv,im[i],in[j]); 613 } 614 } 615 } else { 616 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]); 617 if (!aij->donotstash) { 618 mat->assembled = PETSC_FALSE; 619 if (roworiented) { 620 ierr = MatStashValuesRow_Private(&mat->stash,im[i],n,in,v+i*n,(PetscBool)(ignorezeroentries && (addv == ADD_VALUES)));CHKERRQ(ierr); 621 } else { 622 ierr = MatStashValuesCol_Private(&mat->stash,im[i],n,in,v+i,m,(PetscBool)(ignorezeroentries && (addv == ADD_VALUES)));CHKERRQ(ierr); 623 } 624 } 625 } 626 } 627 PetscFunctionReturn(0); 628 } 629 630 PetscErrorCode MatGetValues_MPIAIJ(Mat mat,PetscInt m,const PetscInt idxm[],PetscInt n,const PetscInt idxn[],PetscScalar v[]) 631 { 632 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 633 PetscErrorCode ierr; 634 PetscInt i,j,rstart = mat->rmap->rstart,rend = mat->rmap->rend; 635 PetscInt cstart = mat->cmap->rstart,cend = mat->cmap->rend,row,col; 636 637 PetscFunctionBegin; 638 for (i=0; i<m; i++) { 639 if (idxm[i] < 0) continue; /* SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative row: %D",idxm[i]);*/ 640 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); 641 if (idxm[i] >= rstart && idxm[i] < rend) { 642 row = idxm[i] - rstart; 643 for (j=0; j<n; j++) { 644 if (idxn[j] < 0) continue; /* SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative column: %D",idxn[j]); */ 645 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); 646 if (idxn[j] >= cstart && idxn[j] < cend) { 647 col = idxn[j] - cstart; 648 ierr = MatGetValues(aij->A,1,&row,1,&col,v+i*n+j);CHKERRQ(ierr); 649 } else { 650 if (!aij->colmap) { 651 ierr = MatCreateColmap_MPIAIJ_Private(mat);CHKERRQ(ierr); 652 } 653 #if defined(PETSC_USE_CTABLE) 654 ierr = PetscTableFind(aij->colmap,idxn[j]+1,&col);CHKERRQ(ierr); 655 col--; 656 #else 657 col = aij->colmap[idxn[j]] - 1; 658 #endif 659 if ((col < 0) || (aij->garray[col] != idxn[j])) *(v+i*n+j) = 0.0; 660 else { 661 ierr = MatGetValues(aij->B,1,&row,1,&col,v+i*n+j);CHKERRQ(ierr); 662 } 663 } 664 } 665 } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Only local values currently supported"); 666 } 667 PetscFunctionReturn(0); 668 } 669 670 extern PetscErrorCode MatMultDiagonalBlock_MPIAIJ(Mat,Vec,Vec); 671 672 PetscErrorCode MatAssemblyBegin_MPIAIJ(Mat mat,MatAssemblyType mode) 673 { 674 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 675 PetscErrorCode ierr; 676 PetscInt nstash,reallocs; 677 678 PetscFunctionBegin; 679 if (aij->donotstash || mat->nooffprocentries) PetscFunctionReturn(0); 680 681 ierr = MatStashScatterBegin_Private(mat,&mat->stash,mat->rmap->range);CHKERRQ(ierr); 682 ierr = MatStashGetInfo_Private(&mat->stash,&nstash,&reallocs);CHKERRQ(ierr); 683 ierr = PetscInfo2(aij->A,"Stash has %D entries, uses %D mallocs.\n",nstash,reallocs);CHKERRQ(ierr); 684 PetscFunctionReturn(0); 685 } 686 687 PetscErrorCode MatAssemblyEnd_MPIAIJ(Mat mat,MatAssemblyType mode) 688 { 689 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 690 Mat_SeqAIJ *a = (Mat_SeqAIJ*)aij->A->data; 691 PetscErrorCode ierr; 692 PetscMPIInt n; 693 PetscInt i,j,rstart,ncols,flg; 694 PetscInt *row,*col; 695 PetscBool other_disassembled; 696 PetscScalar *val; 697 698 /* do not use 'b = (Mat_SeqAIJ*)aij->B->data' as B can be reset in disassembly */ 699 700 PetscFunctionBegin; 701 if (!aij->donotstash && !mat->nooffprocentries) { 702 while (1) { 703 ierr = MatStashScatterGetMesg_Private(&mat->stash,&n,&row,&col,&val,&flg);CHKERRQ(ierr); 704 if (!flg) break; 705 706 for (i=0; i<n; ) { 707 /* Now identify the consecutive vals belonging to the same row */ 708 for (j=i,rstart=row[j]; j<n; j++) { 709 if (row[j] != rstart) break; 710 } 711 if (j < n) ncols = j-i; 712 else ncols = n-i; 713 /* Now assemble all these values with a single function call */ 714 ierr = MatSetValues_MPIAIJ(mat,1,row+i,ncols,col+i,val+i,mat->insertmode);CHKERRQ(ierr); 715 716 i = j; 717 } 718 } 719 ierr = MatStashScatterEnd_Private(&mat->stash);CHKERRQ(ierr); 720 } 721 ierr = MatAssemblyBegin(aij->A,mode);CHKERRQ(ierr); 722 ierr = MatAssemblyEnd(aij->A,mode);CHKERRQ(ierr); 723 724 /* determine if any processor has disassembled, if so we must 725 also disassemble ourselfs, in order that we may reassemble. */ 726 /* 727 if nonzero structure of submatrix B cannot change then we know that 728 no processor disassembled thus we can skip this stuff 729 */ 730 if (!((Mat_SeqAIJ*)aij->B->data)->nonew) { 731 ierr = MPIU_Allreduce(&mat->was_assembled,&other_disassembled,1,MPIU_BOOL,MPI_PROD,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 732 if (mat->was_assembled && !other_disassembled) { 733 ierr = MatDisAssemble_MPIAIJ(mat);CHKERRQ(ierr); 734 } 735 } 736 if (!mat->was_assembled && mode == MAT_FINAL_ASSEMBLY) { 737 ierr = MatSetUpMultiply_MPIAIJ(mat);CHKERRQ(ierr); 738 } 739 ierr = MatSetOption(aij->B,MAT_USE_INODES,PETSC_FALSE);CHKERRQ(ierr); 740 ierr = MatAssemblyBegin(aij->B,mode);CHKERRQ(ierr); 741 ierr = MatAssemblyEnd(aij->B,mode);CHKERRQ(ierr); 742 743 ierr = PetscFree2(aij->rowvalues,aij->rowindices);CHKERRQ(ierr); 744 745 aij->rowvalues = 0; 746 747 ierr = VecDestroy(&aij->diag);CHKERRQ(ierr); 748 if (a->inode.size) mat->ops->multdiagonalblock = MatMultDiagonalBlock_MPIAIJ; 749 750 /* if no new nonzero locations are allowed in matrix then only set the matrix state the first time through */ 751 if ((!mat->was_assembled && mode == MAT_FINAL_ASSEMBLY) || !((Mat_SeqAIJ*)(aij->A->data))->nonew) { 752 PetscObjectState state = aij->A->nonzerostate + aij->B->nonzerostate; 753 ierr = MPIU_Allreduce(&state,&mat->nonzerostate,1,MPIU_INT64,MPI_SUM,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 754 } 755 PetscFunctionReturn(0); 756 } 757 758 PetscErrorCode MatZeroEntries_MPIAIJ(Mat A) 759 { 760 Mat_MPIAIJ *l = (Mat_MPIAIJ*)A->data; 761 PetscErrorCode ierr; 762 763 PetscFunctionBegin; 764 ierr = MatZeroEntries(l->A);CHKERRQ(ierr); 765 ierr = MatZeroEntries(l->B);CHKERRQ(ierr); 766 PetscFunctionReturn(0); 767 } 768 769 PetscErrorCode MatZeroRows_MPIAIJ(Mat A,PetscInt N,const PetscInt rows[],PetscScalar diag,Vec x,Vec b) 770 { 771 Mat_MPIAIJ *mat = (Mat_MPIAIJ *) A->data; 772 PetscInt *lrows; 773 PetscInt r, len; 774 PetscErrorCode ierr; 775 776 PetscFunctionBegin; 777 /* get locally owned rows */ 778 ierr = MatZeroRowsMapLocal_Private(A,N,rows,&len,&lrows);CHKERRQ(ierr); 779 /* fix right hand side if needed */ 780 if (x && b) { 781 const PetscScalar *xx; 782 PetscScalar *bb; 783 784 ierr = VecGetArrayRead(x, &xx);CHKERRQ(ierr); 785 ierr = VecGetArray(b, &bb);CHKERRQ(ierr); 786 for (r = 0; r < len; ++r) bb[lrows[r]] = diag*xx[lrows[r]]; 787 ierr = VecRestoreArrayRead(x, &xx);CHKERRQ(ierr); 788 ierr = VecRestoreArray(b, &bb);CHKERRQ(ierr); 789 } 790 /* Must zero l->B before l->A because the (diag) case below may put values into l->B*/ 791 ierr = MatZeroRows(mat->B, len, lrows, 0.0, NULL, NULL);CHKERRQ(ierr); 792 if (A->congruentlayouts == -1) { /* first time we compare rows and cols layouts */ 793 PetscBool cong; 794 ierr = PetscLayoutCompare(A->rmap,A->cmap,&cong);CHKERRQ(ierr); 795 if (cong) A->congruentlayouts = 1; 796 else A->congruentlayouts = 0; 797 } 798 if ((diag != 0.0) && A->congruentlayouts) { 799 ierr = MatZeroRows(mat->A, len, lrows, diag, NULL, NULL);CHKERRQ(ierr); 800 } else if (diag != 0.0) { 801 ierr = MatZeroRows(mat->A, len, lrows, 0.0, NULL, NULL);CHKERRQ(ierr); 802 if (((Mat_SeqAIJ *) mat->A->data)->nonew) SETERRQ(PETSC_COMM_SELF, PETSC_ERR_SUP, "MatZeroRows() on rectangular matrices cannot be used with the Mat options\nMAT_NEW_NONZERO_LOCATIONS,MAT_NEW_NONZERO_LOCATION_ERR,MAT_NEW_NONZERO_ALLOCATION_ERR"); 803 for (r = 0; r < len; ++r) { 804 const PetscInt row = lrows[r] + A->rmap->rstart; 805 ierr = MatSetValues(A, 1, &row, 1, &row, &diag, INSERT_VALUES);CHKERRQ(ierr); 806 } 807 ierr = MatAssemblyBegin(A, MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 808 ierr = MatAssemblyEnd(A, MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 809 } else { 810 ierr = MatZeroRows(mat->A, len, lrows, 0.0, NULL, NULL);CHKERRQ(ierr); 811 } 812 ierr = PetscFree(lrows);CHKERRQ(ierr); 813 814 /* only change matrix nonzero state if pattern was allowed to be changed */ 815 if (!((Mat_SeqAIJ*)(mat->A->data))->keepnonzeropattern) { 816 PetscObjectState state = mat->A->nonzerostate + mat->B->nonzerostate; 817 ierr = MPIU_Allreduce(&state,&A->nonzerostate,1,MPIU_INT64,MPI_SUM,PetscObjectComm((PetscObject)A));CHKERRQ(ierr); 818 } 819 PetscFunctionReturn(0); 820 } 821 822 PetscErrorCode MatZeroRowsColumns_MPIAIJ(Mat A,PetscInt N,const PetscInt rows[],PetscScalar diag,Vec x,Vec b) 823 { 824 Mat_MPIAIJ *l = (Mat_MPIAIJ*)A->data; 825 PetscErrorCode ierr; 826 PetscMPIInt n = A->rmap->n; 827 PetscInt i,j,r,m,p = 0,len = 0; 828 PetscInt *lrows,*owners = A->rmap->range; 829 PetscSFNode *rrows; 830 PetscSF sf; 831 const PetscScalar *xx; 832 PetscScalar *bb,*mask; 833 Vec xmask,lmask; 834 Mat_SeqAIJ *aij = (Mat_SeqAIJ*)l->B->data; 835 const PetscInt *aj, *ii,*ridx; 836 PetscScalar *aa; 837 838 PetscFunctionBegin; 839 /* Create SF where leaves are input rows and roots are owned rows */ 840 ierr = PetscMalloc1(n, &lrows);CHKERRQ(ierr); 841 for (r = 0; r < n; ++r) lrows[r] = -1; 842 ierr = PetscMalloc1(N, &rrows);CHKERRQ(ierr); 843 for (r = 0; r < N; ++r) { 844 const PetscInt idx = rows[r]; 845 if (idx < 0 || A->rmap->N <= idx) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Row %D out of range [0,%D)",idx,A->rmap->N); 846 if (idx < owners[p] || owners[p+1] <= idx) { /* short-circuit the search if the last p owns this row too */ 847 ierr = PetscLayoutFindOwner(A->rmap,idx,&p);CHKERRQ(ierr); 848 } 849 rrows[r].rank = p; 850 rrows[r].index = rows[r] - owners[p]; 851 } 852 ierr = PetscSFCreate(PetscObjectComm((PetscObject) A), &sf);CHKERRQ(ierr); 853 ierr = PetscSFSetGraph(sf, n, N, NULL, PETSC_OWN_POINTER, rrows, PETSC_OWN_POINTER);CHKERRQ(ierr); 854 /* Collect flags for rows to be zeroed */ 855 ierr = PetscSFReduceBegin(sf, MPIU_INT, (PetscInt *) rows, lrows, MPI_LOR);CHKERRQ(ierr); 856 ierr = PetscSFReduceEnd(sf, MPIU_INT, (PetscInt *) rows, lrows, MPI_LOR);CHKERRQ(ierr); 857 ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 858 /* Compress and put in row numbers */ 859 for (r = 0; r < n; ++r) if (lrows[r] >= 0) lrows[len++] = r; 860 /* zero diagonal part of matrix */ 861 ierr = MatZeroRowsColumns(l->A,len,lrows,diag,x,b);CHKERRQ(ierr); 862 /* handle off diagonal part of matrix */ 863 ierr = MatCreateVecs(A,&xmask,NULL);CHKERRQ(ierr); 864 ierr = VecDuplicate(l->lvec,&lmask);CHKERRQ(ierr); 865 ierr = VecGetArray(xmask,&bb);CHKERRQ(ierr); 866 for (i=0; i<len; i++) bb[lrows[i]] = 1; 867 ierr = VecRestoreArray(xmask,&bb);CHKERRQ(ierr); 868 ierr = VecScatterBegin(l->Mvctx,xmask,lmask,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 869 ierr = VecScatterEnd(l->Mvctx,xmask,lmask,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 870 ierr = VecDestroy(&xmask);CHKERRQ(ierr); 871 if (x) { 872 ierr = VecScatterBegin(l->Mvctx,x,l->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 873 ierr = VecScatterEnd(l->Mvctx,x,l->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 874 ierr = VecGetArrayRead(l->lvec,&xx);CHKERRQ(ierr); 875 ierr = VecGetArray(b,&bb);CHKERRQ(ierr); 876 } 877 ierr = VecGetArray(lmask,&mask);CHKERRQ(ierr); 878 /* remove zeroed rows of off diagonal matrix */ 879 ii = aij->i; 880 for (i=0; i<len; i++) { 881 ierr = PetscMemzero(aij->a + ii[lrows[i]],(ii[lrows[i]+1] - ii[lrows[i]])*sizeof(PetscScalar));CHKERRQ(ierr); 882 } 883 /* loop over all elements of off process part of matrix zeroing removed columns*/ 884 if (aij->compressedrow.use) { 885 m = aij->compressedrow.nrows; 886 ii = aij->compressedrow.i; 887 ridx = aij->compressedrow.rindex; 888 for (i=0; i<m; i++) { 889 n = ii[i+1] - ii[i]; 890 aj = aij->j + ii[i]; 891 aa = aij->a + ii[i]; 892 893 for (j=0; j<n; j++) { 894 if (PetscAbsScalar(mask[*aj])) { 895 if (b) bb[*ridx] -= *aa*xx[*aj]; 896 *aa = 0.0; 897 } 898 aa++; 899 aj++; 900 } 901 ridx++; 902 } 903 } else { /* do not use compressed row format */ 904 m = l->B->rmap->n; 905 for (i=0; i<m; i++) { 906 n = ii[i+1] - ii[i]; 907 aj = aij->j + ii[i]; 908 aa = aij->a + ii[i]; 909 for (j=0; j<n; j++) { 910 if (PetscAbsScalar(mask[*aj])) { 911 if (b) bb[i] -= *aa*xx[*aj]; 912 *aa = 0.0; 913 } 914 aa++; 915 aj++; 916 } 917 } 918 } 919 if (x) { 920 ierr = VecRestoreArray(b,&bb);CHKERRQ(ierr); 921 ierr = VecRestoreArrayRead(l->lvec,&xx);CHKERRQ(ierr); 922 } 923 ierr = VecRestoreArray(lmask,&mask);CHKERRQ(ierr); 924 ierr = VecDestroy(&lmask);CHKERRQ(ierr); 925 ierr = PetscFree(lrows);CHKERRQ(ierr); 926 927 /* only change matrix nonzero state if pattern was allowed to be changed */ 928 if (!((Mat_SeqAIJ*)(l->A->data))->keepnonzeropattern) { 929 PetscObjectState state = l->A->nonzerostate + l->B->nonzerostate; 930 ierr = MPIU_Allreduce(&state,&A->nonzerostate,1,MPIU_INT64,MPI_SUM,PetscObjectComm((PetscObject)A));CHKERRQ(ierr); 931 } 932 PetscFunctionReturn(0); 933 } 934 935 PetscErrorCode MatMult_MPIAIJ(Mat A,Vec xx,Vec yy) 936 { 937 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 938 PetscErrorCode ierr; 939 PetscInt nt; 940 941 PetscFunctionBegin; 942 ierr = VecGetLocalSize(xx,&nt);CHKERRQ(ierr); 943 if (nt != A->cmap->n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Incompatible partition of A (%D) and xx (%D)",A->cmap->n,nt); 944 ierr = VecScatterBegin(a->Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 945 ierr = (*a->A->ops->mult)(a->A,xx,yy);CHKERRQ(ierr); 946 ierr = VecScatterEnd(a->Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 947 ierr = (*a->B->ops->multadd)(a->B,a->lvec,yy,yy);CHKERRQ(ierr); 948 PetscFunctionReturn(0); 949 } 950 951 PetscErrorCode MatMultDiagonalBlock_MPIAIJ(Mat A,Vec bb,Vec xx) 952 { 953 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 954 PetscErrorCode ierr; 955 956 PetscFunctionBegin; 957 ierr = MatMultDiagonalBlock(a->A,bb,xx);CHKERRQ(ierr); 958 PetscFunctionReturn(0); 959 } 960 961 PetscErrorCode MatMultAdd_MPIAIJ(Mat A,Vec xx,Vec yy,Vec zz) 962 { 963 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 964 PetscErrorCode ierr; 965 966 PetscFunctionBegin; 967 ierr = VecScatterBegin(a->Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 968 ierr = (*a->A->ops->multadd)(a->A,xx,yy,zz);CHKERRQ(ierr); 969 ierr = VecScatterEnd(a->Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 970 ierr = (*a->B->ops->multadd)(a->B,a->lvec,zz,zz);CHKERRQ(ierr); 971 PetscFunctionReturn(0); 972 } 973 974 PetscErrorCode MatMultTranspose_MPIAIJ(Mat A,Vec xx,Vec yy) 975 { 976 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 977 PetscErrorCode ierr; 978 PetscBool merged; 979 980 PetscFunctionBegin; 981 ierr = VecScatterGetMerged(a->Mvctx,&merged);CHKERRQ(ierr); 982 /* do nondiagonal part */ 983 ierr = (*a->B->ops->multtranspose)(a->B,xx,a->lvec);CHKERRQ(ierr); 984 if (!merged) { 985 /* send it on its way */ 986 ierr = VecScatterBegin(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 987 /* do local part */ 988 ierr = (*a->A->ops->multtranspose)(a->A,xx,yy);CHKERRQ(ierr); 989 /* receive remote parts: note this assumes the values are not actually */ 990 /* added in yy until the next line, */ 991 ierr = VecScatterEnd(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 992 } else { 993 /* do local part */ 994 ierr = (*a->A->ops->multtranspose)(a->A,xx,yy);CHKERRQ(ierr); 995 /* send it on its way */ 996 ierr = VecScatterBegin(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 997 /* values actually were received in the Begin() but we need to call this nop */ 998 ierr = VecScatterEnd(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 999 } 1000 PetscFunctionReturn(0); 1001 } 1002 1003 PetscErrorCode MatIsTranspose_MPIAIJ(Mat Amat,Mat Bmat,PetscReal tol,PetscBool *f) 1004 { 1005 MPI_Comm comm; 1006 Mat_MPIAIJ *Aij = (Mat_MPIAIJ*) Amat->data, *Bij; 1007 Mat Adia = Aij->A, Bdia, Aoff,Boff,*Aoffs,*Boffs; 1008 IS Me,Notme; 1009 PetscErrorCode ierr; 1010 PetscInt M,N,first,last,*notme,i; 1011 PetscMPIInt size; 1012 1013 PetscFunctionBegin; 1014 /* Easy test: symmetric diagonal block */ 1015 Bij = (Mat_MPIAIJ*) Bmat->data; Bdia = Bij->A; 1016 ierr = MatIsTranspose(Adia,Bdia,tol,f);CHKERRQ(ierr); 1017 if (!*f) PetscFunctionReturn(0); 1018 ierr = PetscObjectGetComm((PetscObject)Amat,&comm);CHKERRQ(ierr); 1019 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 1020 if (size == 1) PetscFunctionReturn(0); 1021 1022 /* Hard test: off-diagonal block. This takes a MatCreateSubMatrix. */ 1023 ierr = MatGetSize(Amat,&M,&N);CHKERRQ(ierr); 1024 ierr = MatGetOwnershipRange(Amat,&first,&last);CHKERRQ(ierr); 1025 ierr = PetscMalloc1(N-last+first,¬me);CHKERRQ(ierr); 1026 for (i=0; i<first; i++) notme[i] = i; 1027 for (i=last; i<M; i++) notme[i-last+first] = i; 1028 ierr = ISCreateGeneral(MPI_COMM_SELF,N-last+first,notme,PETSC_COPY_VALUES,&Notme);CHKERRQ(ierr); 1029 ierr = ISCreateStride(MPI_COMM_SELF,last-first,first,1,&Me);CHKERRQ(ierr); 1030 ierr = MatCreateSubMatrices(Amat,1,&Me,&Notme,MAT_INITIAL_MATRIX,&Aoffs);CHKERRQ(ierr); 1031 Aoff = Aoffs[0]; 1032 ierr = MatCreateSubMatrices(Bmat,1,&Notme,&Me,MAT_INITIAL_MATRIX,&Boffs);CHKERRQ(ierr); 1033 Boff = Boffs[0]; 1034 ierr = MatIsTranspose(Aoff,Boff,tol,f);CHKERRQ(ierr); 1035 ierr = MatDestroyMatrices(1,&Aoffs);CHKERRQ(ierr); 1036 ierr = MatDestroyMatrices(1,&Boffs);CHKERRQ(ierr); 1037 ierr = ISDestroy(&Me);CHKERRQ(ierr); 1038 ierr = ISDestroy(&Notme);CHKERRQ(ierr); 1039 ierr = PetscFree(notme);CHKERRQ(ierr); 1040 PetscFunctionReturn(0); 1041 } 1042 1043 PetscErrorCode MatMultTransposeAdd_MPIAIJ(Mat A,Vec xx,Vec yy,Vec zz) 1044 { 1045 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 1046 PetscErrorCode ierr; 1047 1048 PetscFunctionBegin; 1049 /* do nondiagonal part */ 1050 ierr = (*a->B->ops->multtranspose)(a->B,xx,a->lvec);CHKERRQ(ierr); 1051 /* send it on its way */ 1052 ierr = VecScatterBegin(a->Mvctx,a->lvec,zz,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1053 /* do local part */ 1054 ierr = (*a->A->ops->multtransposeadd)(a->A,xx,yy,zz);CHKERRQ(ierr); 1055 /* receive remote parts */ 1056 ierr = VecScatterEnd(a->Mvctx,a->lvec,zz,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1057 PetscFunctionReturn(0); 1058 } 1059 1060 /* 1061 This only works correctly for square matrices where the subblock A->A is the 1062 diagonal block 1063 */ 1064 PetscErrorCode MatGetDiagonal_MPIAIJ(Mat A,Vec v) 1065 { 1066 PetscErrorCode ierr; 1067 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 1068 1069 PetscFunctionBegin; 1070 if (A->rmap->N != A->cmap->N) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"Supports only square matrix where A->A is diag block"); 1071 if (A->rmap->rstart != A->cmap->rstart || A->rmap->rend != A->cmap->rend) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"row partition must equal col partition"); 1072 ierr = MatGetDiagonal(a->A,v);CHKERRQ(ierr); 1073 PetscFunctionReturn(0); 1074 } 1075 1076 PetscErrorCode MatScale_MPIAIJ(Mat A,PetscScalar aa) 1077 { 1078 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 1079 PetscErrorCode ierr; 1080 1081 PetscFunctionBegin; 1082 ierr = MatScale(a->A,aa);CHKERRQ(ierr); 1083 ierr = MatScale(a->B,aa);CHKERRQ(ierr); 1084 PetscFunctionReturn(0); 1085 } 1086 1087 PetscErrorCode MatDestroy_MPIAIJ(Mat mat) 1088 { 1089 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 1090 PetscErrorCode ierr; 1091 1092 PetscFunctionBegin; 1093 #if defined(PETSC_USE_LOG) 1094 PetscLogObjectState((PetscObject)mat,"Rows=%D, Cols=%D",mat->rmap->N,mat->cmap->N); 1095 #endif 1096 ierr = MatStashDestroy_Private(&mat->stash);CHKERRQ(ierr); 1097 ierr = VecDestroy(&aij->diag);CHKERRQ(ierr); 1098 ierr = MatDestroy(&aij->A);CHKERRQ(ierr); 1099 ierr = MatDestroy(&aij->B);CHKERRQ(ierr); 1100 #if defined(PETSC_USE_CTABLE) 1101 ierr = PetscTableDestroy(&aij->colmap);CHKERRQ(ierr); 1102 #else 1103 ierr = PetscFree(aij->colmap);CHKERRQ(ierr); 1104 #endif 1105 ierr = PetscFree(aij->garray);CHKERRQ(ierr); 1106 ierr = VecDestroy(&aij->lvec);CHKERRQ(ierr); 1107 ierr = VecScatterDestroy(&aij->Mvctx);CHKERRQ(ierr); 1108 ierr = PetscFree2(aij->rowvalues,aij->rowindices);CHKERRQ(ierr); 1109 ierr = PetscFree(aij->ld);CHKERRQ(ierr); 1110 ierr = PetscFree(mat->data);CHKERRQ(ierr); 1111 1112 ierr = PetscObjectChangeTypeName((PetscObject)mat,0);CHKERRQ(ierr); 1113 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatStoreValues_C",NULL);CHKERRQ(ierr); 1114 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatRetrieveValues_C",NULL);CHKERRQ(ierr); 1115 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatIsTranspose_C",NULL);CHKERRQ(ierr); 1116 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatMPIAIJSetPreallocation_C",NULL);CHKERRQ(ierr); 1117 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatMPIAIJSetPreallocationCSR_C",NULL);CHKERRQ(ierr); 1118 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatDiagonalScaleLocal_C",NULL);CHKERRQ(ierr); 1119 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatConvert_mpiaij_mpisbaij_C",NULL);CHKERRQ(ierr); 1120 #if defined(PETSC_HAVE_ELEMENTAL) 1121 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatConvert_mpiaij_elemental_C",NULL);CHKERRQ(ierr); 1122 #endif 1123 #if defined(PETSC_HAVE_HYPRE) 1124 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatConvert_mpiaij_hypre_C",NULL);CHKERRQ(ierr); 1125 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatMatMatMult_transpose_mpiaij_mpiaij_C",NULL);CHKERRQ(ierr); 1126 #endif 1127 PetscFunctionReturn(0); 1128 } 1129 1130 PetscErrorCode MatView_MPIAIJ_Binary(Mat mat,PetscViewer viewer) 1131 { 1132 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 1133 Mat_SeqAIJ *A = (Mat_SeqAIJ*)aij->A->data; 1134 Mat_SeqAIJ *B = (Mat_SeqAIJ*)aij->B->data; 1135 PetscErrorCode ierr; 1136 PetscMPIInt rank,size,tag = ((PetscObject)viewer)->tag; 1137 int fd; 1138 PetscInt nz,header[4],*row_lengths,*range=0,rlen,i; 1139 PetscInt nzmax,*column_indices,j,k,col,*garray = aij->garray,cnt,cstart = mat->cmap->rstart,rnz = 0; 1140 PetscScalar *column_values; 1141 PetscInt message_count,flowcontrolcount; 1142 FILE *file; 1143 1144 PetscFunctionBegin; 1145 ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr); 1146 ierr = MPI_Comm_size(PetscObjectComm((PetscObject)mat),&size);CHKERRQ(ierr); 1147 nz = A->nz + B->nz; 1148 ierr = PetscViewerBinaryGetDescriptor(viewer,&fd);CHKERRQ(ierr); 1149 if (!rank) { 1150 header[0] = MAT_FILE_CLASSID; 1151 header[1] = mat->rmap->N; 1152 header[2] = mat->cmap->N; 1153 1154 ierr = MPI_Reduce(&nz,&header[3],1,MPIU_INT,MPI_SUM,0,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1155 ierr = PetscBinaryWrite(fd,header,4,PETSC_INT,PETSC_TRUE);CHKERRQ(ierr); 1156 /* get largest number of rows any processor has */ 1157 rlen = mat->rmap->n; 1158 range = mat->rmap->range; 1159 for (i=1; i<size; i++) rlen = PetscMax(rlen,range[i+1] - range[i]); 1160 } else { 1161 ierr = MPI_Reduce(&nz,0,1,MPIU_INT,MPI_SUM,0,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1162 rlen = mat->rmap->n; 1163 } 1164 1165 /* load up the local row counts */ 1166 ierr = PetscMalloc1(rlen+1,&row_lengths);CHKERRQ(ierr); 1167 for (i=0; i<mat->rmap->n; i++) row_lengths[i] = A->i[i+1] - A->i[i] + B->i[i+1] - B->i[i]; 1168 1169 /* store the row lengths to the file */ 1170 ierr = PetscViewerFlowControlStart(viewer,&message_count,&flowcontrolcount);CHKERRQ(ierr); 1171 if (!rank) { 1172 ierr = PetscBinaryWrite(fd,row_lengths,mat->rmap->n,PETSC_INT,PETSC_TRUE);CHKERRQ(ierr); 1173 for (i=1; i<size; i++) { 1174 ierr = PetscViewerFlowControlStepMaster(viewer,i,&message_count,flowcontrolcount);CHKERRQ(ierr); 1175 rlen = range[i+1] - range[i]; 1176 ierr = MPIULong_Recv(row_lengths,rlen,MPIU_INT,i,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1177 ierr = PetscBinaryWrite(fd,row_lengths,rlen,PETSC_INT,PETSC_TRUE);CHKERRQ(ierr); 1178 } 1179 ierr = PetscViewerFlowControlEndMaster(viewer,&message_count);CHKERRQ(ierr); 1180 } else { 1181 ierr = PetscViewerFlowControlStepWorker(viewer,rank,&message_count);CHKERRQ(ierr); 1182 ierr = MPIULong_Send(row_lengths,mat->rmap->n,MPIU_INT,0,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1183 ierr = PetscViewerFlowControlEndWorker(viewer,&message_count);CHKERRQ(ierr); 1184 } 1185 ierr = PetscFree(row_lengths);CHKERRQ(ierr); 1186 1187 /* load up the local column indices */ 1188 nzmax = nz; /* th processor needs space a largest processor needs */ 1189 ierr = MPI_Reduce(&nz,&nzmax,1,MPIU_INT,MPI_MAX,0,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1190 ierr = PetscMalloc1(nzmax+1,&column_indices);CHKERRQ(ierr); 1191 cnt = 0; 1192 for (i=0; i<mat->rmap->n; i++) { 1193 for (j=B->i[i]; j<B->i[i+1]; j++) { 1194 if ((col = garray[B->j[j]]) > cstart) break; 1195 column_indices[cnt++] = col; 1196 } 1197 for (k=A->i[i]; k<A->i[i+1]; k++) column_indices[cnt++] = A->j[k] + cstart; 1198 for (; j<B->i[i+1]; j++) column_indices[cnt++] = garray[B->j[j]]; 1199 } 1200 if (cnt != A->nz + B->nz) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_LIB,"Internal PETSc error: cnt = %D nz = %D",cnt,A->nz+B->nz); 1201 1202 /* store the column indices to the file */ 1203 ierr = PetscViewerFlowControlStart(viewer,&message_count,&flowcontrolcount);CHKERRQ(ierr); 1204 if (!rank) { 1205 MPI_Status status; 1206 ierr = PetscBinaryWrite(fd,column_indices,nz,PETSC_INT,PETSC_TRUE);CHKERRQ(ierr); 1207 for (i=1; i<size; i++) { 1208 ierr = PetscViewerFlowControlStepMaster(viewer,i,&message_count,flowcontrolcount);CHKERRQ(ierr); 1209 ierr = MPI_Recv(&rnz,1,MPIU_INT,i,tag,PetscObjectComm((PetscObject)mat),&status);CHKERRQ(ierr); 1210 if (rnz > nzmax) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_LIB,"Internal PETSc error: nz = %D nzmax = %D",nz,nzmax); 1211 ierr = MPIULong_Recv(column_indices,rnz,MPIU_INT,i,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1212 ierr = PetscBinaryWrite(fd,column_indices,rnz,PETSC_INT,PETSC_TRUE);CHKERRQ(ierr); 1213 } 1214 ierr = PetscViewerFlowControlEndMaster(viewer,&message_count);CHKERRQ(ierr); 1215 } else { 1216 ierr = PetscViewerFlowControlStepWorker(viewer,rank,&message_count);CHKERRQ(ierr); 1217 ierr = MPI_Send(&nz,1,MPIU_INT,0,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1218 ierr = MPIULong_Send(column_indices,nz,MPIU_INT,0,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1219 ierr = PetscViewerFlowControlEndWorker(viewer,&message_count);CHKERRQ(ierr); 1220 } 1221 ierr = PetscFree(column_indices);CHKERRQ(ierr); 1222 1223 /* load up the local column values */ 1224 ierr = PetscMalloc1(nzmax+1,&column_values);CHKERRQ(ierr); 1225 cnt = 0; 1226 for (i=0; i<mat->rmap->n; i++) { 1227 for (j=B->i[i]; j<B->i[i+1]; j++) { 1228 if (garray[B->j[j]] > cstart) break; 1229 column_values[cnt++] = B->a[j]; 1230 } 1231 for (k=A->i[i]; k<A->i[i+1]; k++) column_values[cnt++] = A->a[k]; 1232 for (; j<B->i[i+1]; j++) column_values[cnt++] = B->a[j]; 1233 } 1234 if (cnt != A->nz + B->nz) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Internal PETSc error: cnt = %D nz = %D",cnt,A->nz+B->nz); 1235 1236 /* store the column values to the file */ 1237 ierr = PetscViewerFlowControlStart(viewer,&message_count,&flowcontrolcount);CHKERRQ(ierr); 1238 if (!rank) { 1239 MPI_Status status; 1240 ierr = PetscBinaryWrite(fd,column_values,nz,PETSC_SCALAR,PETSC_TRUE);CHKERRQ(ierr); 1241 for (i=1; i<size; i++) { 1242 ierr = PetscViewerFlowControlStepMaster(viewer,i,&message_count,flowcontrolcount);CHKERRQ(ierr); 1243 ierr = MPI_Recv(&rnz,1,MPIU_INT,i,tag,PetscObjectComm((PetscObject)mat),&status);CHKERRQ(ierr); 1244 if (rnz > nzmax) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_LIB,"Internal PETSc error: nz = %D nzmax = %D",nz,nzmax); 1245 ierr = MPIULong_Recv(column_values,rnz,MPIU_SCALAR,i,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1246 ierr = PetscBinaryWrite(fd,column_values,rnz,PETSC_SCALAR,PETSC_TRUE);CHKERRQ(ierr); 1247 } 1248 ierr = PetscViewerFlowControlEndMaster(viewer,&message_count);CHKERRQ(ierr); 1249 } else { 1250 ierr = PetscViewerFlowControlStepWorker(viewer,rank,&message_count);CHKERRQ(ierr); 1251 ierr = MPI_Send(&nz,1,MPIU_INT,0,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1252 ierr = MPIULong_Send(column_values,nz,MPIU_SCALAR,0,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1253 ierr = PetscViewerFlowControlEndWorker(viewer,&message_count);CHKERRQ(ierr); 1254 } 1255 ierr = PetscFree(column_values);CHKERRQ(ierr); 1256 1257 ierr = PetscViewerBinaryGetInfoPointer(viewer,&file);CHKERRQ(ierr); 1258 if (file) fprintf(file,"-matload_block_size %d\n",(int)PetscAbs(mat->rmap->bs)); 1259 PetscFunctionReturn(0); 1260 } 1261 1262 #include <petscdraw.h> 1263 PetscErrorCode MatView_MPIAIJ_ASCIIorDraworSocket(Mat mat,PetscViewer viewer) 1264 { 1265 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 1266 PetscErrorCode ierr; 1267 PetscMPIInt rank = aij->rank,size = aij->size; 1268 PetscBool isdraw,iascii,isbinary; 1269 PetscViewer sviewer; 1270 PetscViewerFormat format; 1271 1272 PetscFunctionBegin; 1273 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERDRAW,&isdraw);CHKERRQ(ierr); 1274 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii);CHKERRQ(ierr); 1275 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERBINARY,&isbinary);CHKERRQ(ierr); 1276 if (iascii) { 1277 ierr = PetscViewerGetFormat(viewer,&format);CHKERRQ(ierr); 1278 if (format == PETSC_VIEWER_ASCII_INFO_DETAIL) { 1279 MatInfo info; 1280 PetscBool inodes; 1281 1282 ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr); 1283 ierr = MatGetInfo(mat,MAT_LOCAL,&info);CHKERRQ(ierr); 1284 ierr = MatInodeGetInodeSizes(aij->A,NULL,(PetscInt**)&inodes,NULL);CHKERRQ(ierr); 1285 ierr = PetscViewerASCIIPushSynchronized(viewer);CHKERRQ(ierr); 1286 if (!inodes) { 1287 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"[%d] Local rows %D nz %D nz alloced %D mem %D, not using I-node routines\n", 1288 rank,mat->rmap->n,(PetscInt)info.nz_used,(PetscInt)info.nz_allocated,(PetscInt)info.memory);CHKERRQ(ierr); 1289 } else { 1290 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"[%d] Local rows %D nz %D nz alloced %D mem %D, using I-node routines\n", 1291 rank,mat->rmap->n,(PetscInt)info.nz_used,(PetscInt)info.nz_allocated,(PetscInt)info.memory);CHKERRQ(ierr); 1292 } 1293 ierr = MatGetInfo(aij->A,MAT_LOCAL,&info);CHKERRQ(ierr); 1294 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"[%d] on-diagonal part: nz %D \n",rank,(PetscInt)info.nz_used);CHKERRQ(ierr); 1295 ierr = MatGetInfo(aij->B,MAT_LOCAL,&info);CHKERRQ(ierr); 1296 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"[%d] off-diagonal part: nz %D \n",rank,(PetscInt)info.nz_used);CHKERRQ(ierr); 1297 ierr = PetscViewerFlush(viewer);CHKERRQ(ierr); 1298 ierr = PetscViewerASCIIPopSynchronized(viewer);CHKERRQ(ierr); 1299 ierr = PetscViewerASCIIPrintf(viewer,"Information on VecScatter used in matrix-vector product: \n");CHKERRQ(ierr); 1300 ierr = VecScatterView(aij->Mvctx,viewer);CHKERRQ(ierr); 1301 PetscFunctionReturn(0); 1302 } else if (format == PETSC_VIEWER_ASCII_INFO) { 1303 PetscInt inodecount,inodelimit,*inodes; 1304 ierr = MatInodeGetInodeSizes(aij->A,&inodecount,&inodes,&inodelimit);CHKERRQ(ierr); 1305 if (inodes) { 1306 ierr = PetscViewerASCIIPrintf(viewer,"using I-node (on process 0) routines: found %D nodes, limit used is %D\n",inodecount,inodelimit);CHKERRQ(ierr); 1307 } else { 1308 ierr = PetscViewerASCIIPrintf(viewer,"not using I-node (on process 0) routines\n");CHKERRQ(ierr); 1309 } 1310 PetscFunctionReturn(0); 1311 } else if (format == PETSC_VIEWER_ASCII_FACTOR_INFO) { 1312 PetscFunctionReturn(0); 1313 } 1314 } else if (isbinary) { 1315 if (size == 1) { 1316 ierr = PetscObjectSetName((PetscObject)aij->A,((PetscObject)mat)->name);CHKERRQ(ierr); 1317 ierr = MatView(aij->A,viewer);CHKERRQ(ierr); 1318 } else { 1319 ierr = MatView_MPIAIJ_Binary(mat,viewer);CHKERRQ(ierr); 1320 } 1321 PetscFunctionReturn(0); 1322 } else if (isdraw) { 1323 PetscDraw draw; 1324 PetscBool isnull; 1325 ierr = PetscViewerDrawGetDraw(viewer,0,&draw);CHKERRQ(ierr); 1326 ierr = PetscDrawIsNull(draw,&isnull);CHKERRQ(ierr); 1327 if (isnull) PetscFunctionReturn(0); 1328 } 1329 1330 { 1331 /* assemble the entire matrix onto first processor. */ 1332 Mat A; 1333 Mat_SeqAIJ *Aloc; 1334 PetscInt M = mat->rmap->N,N = mat->cmap->N,m,*ai,*aj,row,*cols,i,*ct; 1335 MatScalar *a; 1336 1337 ierr = MatCreate(PetscObjectComm((PetscObject)mat),&A);CHKERRQ(ierr); 1338 if (!rank) { 1339 ierr = MatSetSizes(A,M,N,M,N);CHKERRQ(ierr); 1340 } else { 1341 ierr = MatSetSizes(A,0,0,M,N);CHKERRQ(ierr); 1342 } 1343 /* This is just a temporary matrix, so explicitly using MATMPIAIJ is probably best */ 1344 ierr = MatSetType(A,MATMPIAIJ);CHKERRQ(ierr); 1345 ierr = MatMPIAIJSetPreallocation(A,0,NULL,0,NULL);CHKERRQ(ierr); 1346 ierr = MatSetOption(A,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 1347 ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)A);CHKERRQ(ierr); 1348 1349 /* copy over the A part */ 1350 Aloc = (Mat_SeqAIJ*)aij->A->data; 1351 m = aij->A->rmap->n; ai = Aloc->i; aj = Aloc->j; a = Aloc->a; 1352 row = mat->rmap->rstart; 1353 for (i=0; i<ai[m]; i++) aj[i] += mat->cmap->rstart; 1354 for (i=0; i<m; i++) { 1355 ierr = MatSetValues(A,1,&row,ai[i+1]-ai[i],aj,a,INSERT_VALUES);CHKERRQ(ierr); 1356 row++; 1357 a += ai[i+1]-ai[i]; aj += ai[i+1]-ai[i]; 1358 } 1359 aj = Aloc->j; 1360 for (i=0; i<ai[m]; i++) aj[i] -= mat->cmap->rstart; 1361 1362 /* copy over the B part */ 1363 Aloc = (Mat_SeqAIJ*)aij->B->data; 1364 m = aij->B->rmap->n; ai = Aloc->i; aj = Aloc->j; a = Aloc->a; 1365 row = mat->rmap->rstart; 1366 ierr = PetscMalloc1(ai[m]+1,&cols);CHKERRQ(ierr); 1367 ct = cols; 1368 for (i=0; i<ai[m]; i++) cols[i] = aij->garray[aj[i]]; 1369 for (i=0; i<m; i++) { 1370 ierr = MatSetValues(A,1,&row,ai[i+1]-ai[i],cols,a,INSERT_VALUES);CHKERRQ(ierr); 1371 row++; 1372 a += ai[i+1]-ai[i]; cols += ai[i+1]-ai[i]; 1373 } 1374 ierr = PetscFree(ct);CHKERRQ(ierr); 1375 ierr = MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1376 ierr = MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1377 /* 1378 Everyone has to call to draw the matrix since the graphics waits are 1379 synchronized across all processors that share the PetscDraw object 1380 */ 1381 ierr = PetscViewerGetSubViewer(viewer,PETSC_COMM_SELF,&sviewer);CHKERRQ(ierr); 1382 if (!rank) { 1383 ierr = PetscObjectSetName((PetscObject)((Mat_MPIAIJ*)(A->data))->A,((PetscObject)mat)->name);CHKERRQ(ierr); 1384 ierr = MatView_SeqAIJ(((Mat_MPIAIJ*)(A->data))->A,sviewer);CHKERRQ(ierr); 1385 } 1386 ierr = PetscViewerRestoreSubViewer(viewer,PETSC_COMM_SELF,&sviewer);CHKERRQ(ierr); 1387 ierr = PetscViewerFlush(viewer);CHKERRQ(ierr); 1388 ierr = MatDestroy(&A);CHKERRQ(ierr); 1389 } 1390 PetscFunctionReturn(0); 1391 } 1392 1393 PetscErrorCode MatView_MPIAIJ(Mat mat,PetscViewer viewer) 1394 { 1395 PetscErrorCode ierr; 1396 PetscBool iascii,isdraw,issocket,isbinary; 1397 1398 PetscFunctionBegin; 1399 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii);CHKERRQ(ierr); 1400 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERDRAW,&isdraw);CHKERRQ(ierr); 1401 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERBINARY,&isbinary);CHKERRQ(ierr); 1402 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERSOCKET,&issocket);CHKERRQ(ierr); 1403 if (iascii || isdraw || isbinary || issocket) { 1404 ierr = MatView_MPIAIJ_ASCIIorDraworSocket(mat,viewer);CHKERRQ(ierr); 1405 } 1406 PetscFunctionReturn(0); 1407 } 1408 1409 PetscErrorCode MatSOR_MPIAIJ(Mat matin,Vec bb,PetscReal omega,MatSORType flag,PetscReal fshift,PetscInt its,PetscInt lits,Vec xx) 1410 { 1411 Mat_MPIAIJ *mat = (Mat_MPIAIJ*)matin->data; 1412 PetscErrorCode ierr; 1413 Vec bb1 = 0; 1414 PetscBool hasop; 1415 1416 PetscFunctionBegin; 1417 if (flag == SOR_APPLY_UPPER) { 1418 ierr = (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);CHKERRQ(ierr); 1419 PetscFunctionReturn(0); 1420 } 1421 1422 if (its > 1 || ~flag & SOR_ZERO_INITIAL_GUESS || flag & SOR_EISENSTAT) { 1423 ierr = VecDuplicate(bb,&bb1);CHKERRQ(ierr); 1424 } 1425 1426 if ((flag & SOR_LOCAL_SYMMETRIC_SWEEP) == SOR_LOCAL_SYMMETRIC_SWEEP) { 1427 if (flag & SOR_ZERO_INITIAL_GUESS) { 1428 ierr = (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);CHKERRQ(ierr); 1429 its--; 1430 } 1431 1432 while (its--) { 1433 ierr = VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1434 ierr = VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1435 1436 /* update rhs: bb1 = bb - B*x */ 1437 ierr = VecScale(mat->lvec,-1.0);CHKERRQ(ierr); 1438 ierr = (*mat->B->ops->multadd)(mat->B,mat->lvec,bb,bb1);CHKERRQ(ierr); 1439 1440 /* local sweep */ 1441 ierr = (*mat->A->ops->sor)(mat->A,bb1,omega,SOR_SYMMETRIC_SWEEP,fshift,lits,1,xx);CHKERRQ(ierr); 1442 } 1443 } else if (flag & SOR_LOCAL_FORWARD_SWEEP) { 1444 if (flag & SOR_ZERO_INITIAL_GUESS) { 1445 ierr = (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);CHKERRQ(ierr); 1446 its--; 1447 } 1448 while (its--) { 1449 ierr = VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1450 ierr = VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1451 1452 /* update rhs: bb1 = bb - B*x */ 1453 ierr = VecScale(mat->lvec,-1.0);CHKERRQ(ierr); 1454 ierr = (*mat->B->ops->multadd)(mat->B,mat->lvec,bb,bb1);CHKERRQ(ierr); 1455 1456 /* local sweep */ 1457 ierr = (*mat->A->ops->sor)(mat->A,bb1,omega,SOR_FORWARD_SWEEP,fshift,lits,1,xx);CHKERRQ(ierr); 1458 } 1459 } else if (flag & SOR_LOCAL_BACKWARD_SWEEP) { 1460 if (flag & SOR_ZERO_INITIAL_GUESS) { 1461 ierr = (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);CHKERRQ(ierr); 1462 its--; 1463 } 1464 while (its--) { 1465 ierr = VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1466 ierr = VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1467 1468 /* update rhs: bb1 = bb - B*x */ 1469 ierr = VecScale(mat->lvec,-1.0);CHKERRQ(ierr); 1470 ierr = (*mat->B->ops->multadd)(mat->B,mat->lvec,bb,bb1);CHKERRQ(ierr); 1471 1472 /* local sweep */ 1473 ierr = (*mat->A->ops->sor)(mat->A,bb1,omega,SOR_BACKWARD_SWEEP,fshift,lits,1,xx);CHKERRQ(ierr); 1474 } 1475 } else if (flag & SOR_EISENSTAT) { 1476 Vec xx1; 1477 1478 ierr = VecDuplicate(bb,&xx1);CHKERRQ(ierr); 1479 ierr = (*mat->A->ops->sor)(mat->A,bb,omega,(MatSORType)(SOR_ZERO_INITIAL_GUESS | SOR_LOCAL_BACKWARD_SWEEP),fshift,lits,1,xx);CHKERRQ(ierr); 1480 1481 ierr = VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1482 ierr = VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1483 if (!mat->diag) { 1484 ierr = MatCreateVecs(matin,&mat->diag,NULL);CHKERRQ(ierr); 1485 ierr = MatGetDiagonal(matin,mat->diag);CHKERRQ(ierr); 1486 } 1487 ierr = MatHasOperation(matin,MATOP_MULT_DIAGONAL_BLOCK,&hasop);CHKERRQ(ierr); 1488 if (hasop) { 1489 ierr = MatMultDiagonalBlock(matin,xx,bb1);CHKERRQ(ierr); 1490 } else { 1491 ierr = VecPointwiseMult(bb1,mat->diag,xx);CHKERRQ(ierr); 1492 } 1493 ierr = VecAYPX(bb1,(omega-2.0)/omega,bb);CHKERRQ(ierr); 1494 1495 ierr = MatMultAdd(mat->B,mat->lvec,bb1,bb1);CHKERRQ(ierr); 1496 1497 /* local sweep */ 1498 ierr = (*mat->A->ops->sor)(mat->A,bb1,omega,(MatSORType)(SOR_ZERO_INITIAL_GUESS | SOR_LOCAL_FORWARD_SWEEP),fshift,lits,1,xx1);CHKERRQ(ierr); 1499 ierr = VecAXPY(xx,1.0,xx1);CHKERRQ(ierr); 1500 ierr = VecDestroy(&xx1);CHKERRQ(ierr); 1501 } else SETERRQ(PetscObjectComm((PetscObject)matin),PETSC_ERR_SUP,"Parallel SOR not supported"); 1502 1503 ierr = VecDestroy(&bb1);CHKERRQ(ierr); 1504 1505 matin->factorerrortype = mat->A->factorerrortype; 1506 PetscFunctionReturn(0); 1507 } 1508 1509 PetscErrorCode MatPermute_MPIAIJ(Mat A,IS rowp,IS colp,Mat *B) 1510 { 1511 Mat aA,aB,Aperm; 1512 const PetscInt *rwant,*cwant,*gcols,*ai,*bi,*aj,*bj; 1513 PetscScalar *aa,*ba; 1514 PetscInt i,j,m,n,ng,anz,bnz,*dnnz,*onnz,*tdnnz,*tonnz,*rdest,*cdest,*work,*gcdest; 1515 PetscSF rowsf,sf; 1516 IS parcolp = NULL; 1517 PetscBool done; 1518 PetscErrorCode ierr; 1519 1520 PetscFunctionBegin; 1521 ierr = MatGetLocalSize(A,&m,&n);CHKERRQ(ierr); 1522 ierr = ISGetIndices(rowp,&rwant);CHKERRQ(ierr); 1523 ierr = ISGetIndices(colp,&cwant);CHKERRQ(ierr); 1524 ierr = PetscMalloc3(PetscMax(m,n),&work,m,&rdest,n,&cdest);CHKERRQ(ierr); 1525 1526 /* Invert row permutation to find out where my rows should go */ 1527 ierr = PetscSFCreate(PetscObjectComm((PetscObject)A),&rowsf);CHKERRQ(ierr); 1528 ierr = PetscSFSetGraphLayout(rowsf,A->rmap,A->rmap->n,NULL,PETSC_OWN_POINTER,rwant);CHKERRQ(ierr); 1529 ierr = PetscSFSetFromOptions(rowsf);CHKERRQ(ierr); 1530 for (i=0; i<m; i++) work[i] = A->rmap->rstart + i; 1531 ierr = PetscSFReduceBegin(rowsf,MPIU_INT,work,rdest,MPIU_REPLACE);CHKERRQ(ierr); 1532 ierr = PetscSFReduceEnd(rowsf,MPIU_INT,work,rdest,MPIU_REPLACE);CHKERRQ(ierr); 1533 1534 /* Invert column permutation to find out where my columns should go */ 1535 ierr = PetscSFCreate(PetscObjectComm((PetscObject)A),&sf);CHKERRQ(ierr); 1536 ierr = PetscSFSetGraphLayout(sf,A->cmap,A->cmap->n,NULL,PETSC_OWN_POINTER,cwant);CHKERRQ(ierr); 1537 ierr = PetscSFSetFromOptions(sf);CHKERRQ(ierr); 1538 for (i=0; i<n; i++) work[i] = A->cmap->rstart + i; 1539 ierr = PetscSFReduceBegin(sf,MPIU_INT,work,cdest,MPIU_REPLACE);CHKERRQ(ierr); 1540 ierr = PetscSFReduceEnd(sf,MPIU_INT,work,cdest,MPIU_REPLACE);CHKERRQ(ierr); 1541 ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 1542 1543 ierr = ISRestoreIndices(rowp,&rwant);CHKERRQ(ierr); 1544 ierr = ISRestoreIndices(colp,&cwant);CHKERRQ(ierr); 1545 ierr = MatMPIAIJGetSeqAIJ(A,&aA,&aB,&gcols);CHKERRQ(ierr); 1546 1547 /* Find out where my gcols should go */ 1548 ierr = MatGetSize(aB,NULL,&ng);CHKERRQ(ierr); 1549 ierr = PetscMalloc1(ng,&gcdest);CHKERRQ(ierr); 1550 ierr = PetscSFCreate(PetscObjectComm((PetscObject)A),&sf);CHKERRQ(ierr); 1551 ierr = PetscSFSetGraphLayout(sf,A->cmap,ng,NULL,PETSC_OWN_POINTER,gcols);CHKERRQ(ierr); 1552 ierr = PetscSFSetFromOptions(sf);CHKERRQ(ierr); 1553 ierr = PetscSFBcastBegin(sf,MPIU_INT,cdest,gcdest);CHKERRQ(ierr); 1554 ierr = PetscSFBcastEnd(sf,MPIU_INT,cdest,gcdest);CHKERRQ(ierr); 1555 ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 1556 1557 ierr = PetscCalloc4(m,&dnnz,m,&onnz,m,&tdnnz,m,&tonnz);CHKERRQ(ierr); 1558 ierr = MatGetRowIJ(aA,0,PETSC_FALSE,PETSC_FALSE,&anz,&ai,&aj,&done);CHKERRQ(ierr); 1559 ierr = MatGetRowIJ(aB,0,PETSC_FALSE,PETSC_FALSE,&bnz,&bi,&bj,&done);CHKERRQ(ierr); 1560 for (i=0; i<m; i++) { 1561 PetscInt row = rdest[i],rowner; 1562 ierr = PetscLayoutFindOwner(A->rmap,row,&rowner);CHKERRQ(ierr); 1563 for (j=ai[i]; j<ai[i+1]; j++) { 1564 PetscInt cowner,col = cdest[aj[j]]; 1565 ierr = PetscLayoutFindOwner(A->cmap,col,&cowner);CHKERRQ(ierr); /* Could build an index for the columns to eliminate this search */ 1566 if (rowner == cowner) dnnz[i]++; 1567 else onnz[i]++; 1568 } 1569 for (j=bi[i]; j<bi[i+1]; j++) { 1570 PetscInt cowner,col = gcdest[bj[j]]; 1571 ierr = PetscLayoutFindOwner(A->cmap,col,&cowner);CHKERRQ(ierr); 1572 if (rowner == cowner) dnnz[i]++; 1573 else onnz[i]++; 1574 } 1575 } 1576 ierr = PetscSFBcastBegin(rowsf,MPIU_INT,dnnz,tdnnz);CHKERRQ(ierr); 1577 ierr = PetscSFBcastEnd(rowsf,MPIU_INT,dnnz,tdnnz);CHKERRQ(ierr); 1578 ierr = PetscSFBcastBegin(rowsf,MPIU_INT,onnz,tonnz);CHKERRQ(ierr); 1579 ierr = PetscSFBcastEnd(rowsf,MPIU_INT,onnz,tonnz);CHKERRQ(ierr); 1580 ierr = PetscSFDestroy(&rowsf);CHKERRQ(ierr); 1581 1582 ierr = MatCreateAIJ(PetscObjectComm((PetscObject)A),A->rmap->n,A->cmap->n,A->rmap->N,A->cmap->N,0,tdnnz,0,tonnz,&Aperm);CHKERRQ(ierr); 1583 ierr = MatSeqAIJGetArray(aA,&aa);CHKERRQ(ierr); 1584 ierr = MatSeqAIJGetArray(aB,&ba);CHKERRQ(ierr); 1585 for (i=0; i<m; i++) { 1586 PetscInt *acols = dnnz,*bcols = onnz; /* Repurpose now-unneeded arrays */ 1587 PetscInt j0,rowlen; 1588 rowlen = ai[i+1] - ai[i]; 1589 for (j0=j=0; j<rowlen; j0=j) { /* rowlen could be larger than number of rows m, so sum in batches */ 1590 for ( ; j<PetscMin(rowlen,j0+m); j++) acols[j-j0] = cdest[aj[ai[i]+j]]; 1591 ierr = MatSetValues(Aperm,1,&rdest[i],j-j0,acols,aa+ai[i]+j0,INSERT_VALUES);CHKERRQ(ierr); 1592 } 1593 rowlen = bi[i+1] - bi[i]; 1594 for (j0=j=0; j<rowlen; j0=j) { 1595 for ( ; j<PetscMin(rowlen,j0+m); j++) bcols[j-j0] = gcdest[bj[bi[i]+j]]; 1596 ierr = MatSetValues(Aperm,1,&rdest[i],j-j0,bcols,ba+bi[i]+j0,INSERT_VALUES);CHKERRQ(ierr); 1597 } 1598 } 1599 ierr = MatAssemblyBegin(Aperm,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1600 ierr = MatAssemblyEnd(Aperm,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1601 ierr = MatRestoreRowIJ(aA,0,PETSC_FALSE,PETSC_FALSE,&anz,&ai,&aj,&done);CHKERRQ(ierr); 1602 ierr = MatRestoreRowIJ(aB,0,PETSC_FALSE,PETSC_FALSE,&bnz,&bi,&bj,&done);CHKERRQ(ierr); 1603 ierr = MatSeqAIJRestoreArray(aA,&aa);CHKERRQ(ierr); 1604 ierr = MatSeqAIJRestoreArray(aB,&ba);CHKERRQ(ierr); 1605 ierr = PetscFree4(dnnz,onnz,tdnnz,tonnz);CHKERRQ(ierr); 1606 ierr = PetscFree3(work,rdest,cdest);CHKERRQ(ierr); 1607 ierr = PetscFree(gcdest);CHKERRQ(ierr); 1608 if (parcolp) {ierr = ISDestroy(&colp);CHKERRQ(ierr);} 1609 *B = Aperm; 1610 PetscFunctionReturn(0); 1611 } 1612 1613 PetscErrorCode MatGetGhosts_MPIAIJ(Mat mat,PetscInt *nghosts,const PetscInt *ghosts[]) 1614 { 1615 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 1616 PetscErrorCode ierr; 1617 1618 PetscFunctionBegin; 1619 ierr = MatGetSize(aij->B,NULL,nghosts);CHKERRQ(ierr); 1620 if (ghosts) *ghosts = aij->garray; 1621 PetscFunctionReturn(0); 1622 } 1623 1624 PetscErrorCode MatGetInfo_MPIAIJ(Mat matin,MatInfoType flag,MatInfo *info) 1625 { 1626 Mat_MPIAIJ *mat = (Mat_MPIAIJ*)matin->data; 1627 Mat A = mat->A,B = mat->B; 1628 PetscErrorCode ierr; 1629 PetscReal isend[5],irecv[5]; 1630 1631 PetscFunctionBegin; 1632 info->block_size = 1.0; 1633 ierr = MatGetInfo(A,MAT_LOCAL,info);CHKERRQ(ierr); 1634 1635 isend[0] = info->nz_used; isend[1] = info->nz_allocated; isend[2] = info->nz_unneeded; 1636 isend[3] = info->memory; isend[4] = info->mallocs; 1637 1638 ierr = MatGetInfo(B,MAT_LOCAL,info);CHKERRQ(ierr); 1639 1640 isend[0] += info->nz_used; isend[1] += info->nz_allocated; isend[2] += info->nz_unneeded; 1641 isend[3] += info->memory; isend[4] += info->mallocs; 1642 if (flag == MAT_LOCAL) { 1643 info->nz_used = isend[0]; 1644 info->nz_allocated = isend[1]; 1645 info->nz_unneeded = isend[2]; 1646 info->memory = isend[3]; 1647 info->mallocs = isend[4]; 1648 } else if (flag == MAT_GLOBAL_MAX) { 1649 ierr = MPIU_Allreduce(isend,irecv,5,MPIU_REAL,MPIU_MAX,PetscObjectComm((PetscObject)matin));CHKERRQ(ierr); 1650 1651 info->nz_used = irecv[0]; 1652 info->nz_allocated = irecv[1]; 1653 info->nz_unneeded = irecv[2]; 1654 info->memory = irecv[3]; 1655 info->mallocs = irecv[4]; 1656 } else if (flag == MAT_GLOBAL_SUM) { 1657 ierr = MPIU_Allreduce(isend,irecv,5,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)matin));CHKERRQ(ierr); 1658 1659 info->nz_used = irecv[0]; 1660 info->nz_allocated = irecv[1]; 1661 info->nz_unneeded = irecv[2]; 1662 info->memory = irecv[3]; 1663 info->mallocs = irecv[4]; 1664 } 1665 info->fill_ratio_given = 0; /* no parallel LU/ILU/Cholesky */ 1666 info->fill_ratio_needed = 0; 1667 info->factor_mallocs = 0; 1668 PetscFunctionReturn(0); 1669 } 1670 1671 PetscErrorCode MatSetOption_MPIAIJ(Mat A,MatOption op,PetscBool flg) 1672 { 1673 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 1674 PetscErrorCode ierr; 1675 1676 PetscFunctionBegin; 1677 switch (op) { 1678 case MAT_NEW_NONZERO_LOCATIONS: 1679 case MAT_NEW_NONZERO_ALLOCATION_ERR: 1680 case MAT_UNUSED_NONZERO_LOCATION_ERR: 1681 case MAT_KEEP_NONZERO_PATTERN: 1682 case MAT_NEW_NONZERO_LOCATION_ERR: 1683 case MAT_USE_INODES: 1684 case MAT_IGNORE_ZERO_ENTRIES: 1685 MatCheckPreallocated(A,1); 1686 ierr = MatSetOption(a->A,op,flg);CHKERRQ(ierr); 1687 ierr = MatSetOption(a->B,op,flg);CHKERRQ(ierr); 1688 break; 1689 case MAT_ROW_ORIENTED: 1690 MatCheckPreallocated(A,1); 1691 a->roworiented = flg; 1692 1693 ierr = MatSetOption(a->A,op,flg);CHKERRQ(ierr); 1694 ierr = MatSetOption(a->B,op,flg);CHKERRQ(ierr); 1695 break; 1696 case MAT_NEW_DIAGONALS: 1697 ierr = PetscInfo1(A,"Option %s ignored\n",MatOptions[op]);CHKERRQ(ierr); 1698 break; 1699 case MAT_IGNORE_OFF_PROC_ENTRIES: 1700 a->donotstash = flg; 1701 break; 1702 case MAT_SPD: 1703 A->spd_set = PETSC_TRUE; 1704 A->spd = flg; 1705 if (flg) { 1706 A->symmetric = PETSC_TRUE; 1707 A->structurally_symmetric = PETSC_TRUE; 1708 A->symmetric_set = PETSC_TRUE; 1709 A->structurally_symmetric_set = PETSC_TRUE; 1710 } 1711 break; 1712 case MAT_SYMMETRIC: 1713 MatCheckPreallocated(A,1); 1714 ierr = MatSetOption(a->A,op,flg);CHKERRQ(ierr); 1715 break; 1716 case MAT_STRUCTURALLY_SYMMETRIC: 1717 MatCheckPreallocated(A,1); 1718 ierr = MatSetOption(a->A,op,flg);CHKERRQ(ierr); 1719 break; 1720 case MAT_HERMITIAN: 1721 MatCheckPreallocated(A,1); 1722 ierr = MatSetOption(a->A,op,flg);CHKERRQ(ierr); 1723 break; 1724 case MAT_SYMMETRY_ETERNAL: 1725 MatCheckPreallocated(A,1); 1726 ierr = MatSetOption(a->A,op,flg);CHKERRQ(ierr); 1727 break; 1728 case MAT_SUBMAT_SINGLEIS: 1729 A->submat_singleis = flg; 1730 break; 1731 default: 1732 SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"unknown option %d",op); 1733 } 1734 PetscFunctionReturn(0); 1735 } 1736 1737 PetscErrorCode MatGetRow_MPIAIJ(Mat matin,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v) 1738 { 1739 Mat_MPIAIJ *mat = (Mat_MPIAIJ*)matin->data; 1740 PetscScalar *vworkA,*vworkB,**pvA,**pvB,*v_p; 1741 PetscErrorCode ierr; 1742 PetscInt i,*cworkA,*cworkB,**pcA,**pcB,cstart = matin->cmap->rstart; 1743 PetscInt nztot,nzA,nzB,lrow,rstart = matin->rmap->rstart,rend = matin->rmap->rend; 1744 PetscInt *cmap,*idx_p; 1745 1746 PetscFunctionBegin; 1747 if (mat->getrowactive) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Already active"); 1748 mat->getrowactive = PETSC_TRUE; 1749 1750 if (!mat->rowvalues && (idx || v)) { 1751 /* 1752 allocate enough space to hold information from the longest row. 1753 */ 1754 Mat_SeqAIJ *Aa = (Mat_SeqAIJ*)mat->A->data,*Ba = (Mat_SeqAIJ*)mat->B->data; 1755 PetscInt max = 1,tmp; 1756 for (i=0; i<matin->rmap->n; i++) { 1757 tmp = Aa->i[i+1] - Aa->i[i] + Ba->i[i+1] - Ba->i[i]; 1758 if (max < tmp) max = tmp; 1759 } 1760 ierr = PetscMalloc2(max,&mat->rowvalues,max,&mat->rowindices);CHKERRQ(ierr); 1761 } 1762 1763 if (row < rstart || row >= rend) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Only local rows"); 1764 lrow = row - rstart; 1765 1766 pvA = &vworkA; pcA = &cworkA; pvB = &vworkB; pcB = &cworkB; 1767 if (!v) {pvA = 0; pvB = 0;} 1768 if (!idx) {pcA = 0; if (!v) pcB = 0;} 1769 ierr = (*mat->A->ops->getrow)(mat->A,lrow,&nzA,pcA,pvA);CHKERRQ(ierr); 1770 ierr = (*mat->B->ops->getrow)(mat->B,lrow,&nzB,pcB,pvB);CHKERRQ(ierr); 1771 nztot = nzA + nzB; 1772 1773 cmap = mat->garray; 1774 if (v || idx) { 1775 if (nztot) { 1776 /* Sort by increasing column numbers, assuming A and B already sorted */ 1777 PetscInt imark = -1; 1778 if (v) { 1779 *v = v_p = mat->rowvalues; 1780 for (i=0; i<nzB; i++) { 1781 if (cmap[cworkB[i]] < cstart) v_p[i] = vworkB[i]; 1782 else break; 1783 } 1784 imark = i; 1785 for (i=0; i<nzA; i++) v_p[imark+i] = vworkA[i]; 1786 for (i=imark; i<nzB; i++) v_p[nzA+i] = vworkB[i]; 1787 } 1788 if (idx) { 1789 *idx = idx_p = mat->rowindices; 1790 if (imark > -1) { 1791 for (i=0; i<imark; i++) { 1792 idx_p[i] = cmap[cworkB[i]]; 1793 } 1794 } else { 1795 for (i=0; i<nzB; i++) { 1796 if (cmap[cworkB[i]] < cstart) idx_p[i] = cmap[cworkB[i]]; 1797 else break; 1798 } 1799 imark = i; 1800 } 1801 for (i=0; i<nzA; i++) idx_p[imark+i] = cstart + cworkA[i]; 1802 for (i=imark; i<nzB; i++) idx_p[nzA+i] = cmap[cworkB[i]]; 1803 } 1804 } else { 1805 if (idx) *idx = 0; 1806 if (v) *v = 0; 1807 } 1808 } 1809 *nz = nztot; 1810 ierr = (*mat->A->ops->restorerow)(mat->A,lrow,&nzA,pcA,pvA);CHKERRQ(ierr); 1811 ierr = (*mat->B->ops->restorerow)(mat->B,lrow,&nzB,pcB,pvB);CHKERRQ(ierr); 1812 PetscFunctionReturn(0); 1813 } 1814 1815 PetscErrorCode MatRestoreRow_MPIAIJ(Mat mat,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v) 1816 { 1817 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 1818 1819 PetscFunctionBegin; 1820 if (!aij->getrowactive) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"MatGetRow() must be called first"); 1821 aij->getrowactive = PETSC_FALSE; 1822 PetscFunctionReturn(0); 1823 } 1824 1825 PetscErrorCode MatNorm_MPIAIJ(Mat mat,NormType type,PetscReal *norm) 1826 { 1827 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 1828 Mat_SeqAIJ *amat = (Mat_SeqAIJ*)aij->A->data,*bmat = (Mat_SeqAIJ*)aij->B->data; 1829 PetscErrorCode ierr; 1830 PetscInt i,j,cstart = mat->cmap->rstart; 1831 PetscReal sum = 0.0; 1832 MatScalar *v; 1833 1834 PetscFunctionBegin; 1835 if (aij->size == 1) { 1836 ierr = MatNorm(aij->A,type,norm);CHKERRQ(ierr); 1837 } else { 1838 if (type == NORM_FROBENIUS) { 1839 v = amat->a; 1840 for (i=0; i<amat->nz; i++) { 1841 sum += PetscRealPart(PetscConj(*v)*(*v)); v++; 1842 } 1843 v = bmat->a; 1844 for (i=0; i<bmat->nz; i++) { 1845 sum += PetscRealPart(PetscConj(*v)*(*v)); v++; 1846 } 1847 ierr = MPIU_Allreduce(&sum,norm,1,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1848 *norm = PetscSqrtReal(*norm); 1849 ierr = PetscLogFlops(2*amat->nz+2*bmat->nz);CHKERRQ(ierr); 1850 } else if (type == NORM_1) { /* max column norm */ 1851 PetscReal *tmp,*tmp2; 1852 PetscInt *jj,*garray = aij->garray; 1853 ierr = PetscCalloc1(mat->cmap->N+1,&tmp);CHKERRQ(ierr); 1854 ierr = PetscMalloc1(mat->cmap->N+1,&tmp2);CHKERRQ(ierr); 1855 *norm = 0.0; 1856 v = amat->a; jj = amat->j; 1857 for (j=0; j<amat->nz; j++) { 1858 tmp[cstart + *jj++] += PetscAbsScalar(*v); v++; 1859 } 1860 v = bmat->a; jj = bmat->j; 1861 for (j=0; j<bmat->nz; j++) { 1862 tmp[garray[*jj++]] += PetscAbsScalar(*v); v++; 1863 } 1864 ierr = MPIU_Allreduce(tmp,tmp2,mat->cmap->N,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1865 for (j=0; j<mat->cmap->N; j++) { 1866 if (tmp2[j] > *norm) *norm = tmp2[j]; 1867 } 1868 ierr = PetscFree(tmp);CHKERRQ(ierr); 1869 ierr = PetscFree(tmp2);CHKERRQ(ierr); 1870 ierr = PetscLogFlops(PetscMax(amat->nz+bmat->nz-1,0));CHKERRQ(ierr); 1871 } else if (type == NORM_INFINITY) { /* max row norm */ 1872 PetscReal ntemp = 0.0; 1873 for (j=0; j<aij->A->rmap->n; j++) { 1874 v = amat->a + amat->i[j]; 1875 sum = 0.0; 1876 for (i=0; i<amat->i[j+1]-amat->i[j]; i++) { 1877 sum += PetscAbsScalar(*v); v++; 1878 } 1879 v = bmat->a + bmat->i[j]; 1880 for (i=0; i<bmat->i[j+1]-bmat->i[j]; i++) { 1881 sum += PetscAbsScalar(*v); v++; 1882 } 1883 if (sum > ntemp) ntemp = sum; 1884 } 1885 ierr = MPIU_Allreduce(&ntemp,norm,1,MPIU_REAL,MPIU_MAX,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1886 ierr = PetscLogFlops(PetscMax(amat->nz+bmat->nz-1,0));CHKERRQ(ierr); 1887 } else SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"No support for two norm"); 1888 } 1889 PetscFunctionReturn(0); 1890 } 1891 1892 PetscErrorCode MatTranspose_MPIAIJ(Mat A,MatReuse reuse,Mat *matout) 1893 { 1894 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 1895 Mat_SeqAIJ *Aloc=(Mat_SeqAIJ*)a->A->data,*Bloc=(Mat_SeqAIJ*)a->B->data; 1896 PetscErrorCode ierr; 1897 PetscInt M = A->rmap->N,N = A->cmap->N,ma,na,mb,nb,*ai,*aj,*bi,*bj,row,*cols,*cols_tmp,i; 1898 PetscInt cstart = A->cmap->rstart,ncol; 1899 Mat B; 1900 MatScalar *array; 1901 1902 PetscFunctionBegin; 1903 if (reuse == MAT_INPLACE_MATRIX && M != N) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_SIZ,"Square matrix only for in-place"); 1904 1905 ma = A->rmap->n; na = A->cmap->n; mb = a->B->rmap->n; nb = a->B->cmap->n; 1906 ai = Aloc->i; aj = Aloc->j; 1907 bi = Bloc->i; bj = Bloc->j; 1908 if (reuse == MAT_INITIAL_MATRIX || *matout == A) { 1909 PetscInt *d_nnz,*g_nnz,*o_nnz; 1910 PetscSFNode *oloc; 1911 PETSC_UNUSED PetscSF sf; 1912 1913 ierr = PetscMalloc4(na,&d_nnz,na,&o_nnz,nb,&g_nnz,nb,&oloc);CHKERRQ(ierr); 1914 /* compute d_nnz for preallocation */ 1915 ierr = PetscMemzero(d_nnz,na*sizeof(PetscInt));CHKERRQ(ierr); 1916 for (i=0; i<ai[ma]; i++) { 1917 d_nnz[aj[i]]++; 1918 aj[i] += cstart; /* global col index to be used by MatSetValues() */ 1919 } 1920 /* compute local off-diagonal contributions */ 1921 ierr = PetscMemzero(g_nnz,nb*sizeof(PetscInt));CHKERRQ(ierr); 1922 for (i=0; i<bi[ma]; i++) g_nnz[bj[i]]++; 1923 /* map those to global */ 1924 ierr = PetscSFCreate(PetscObjectComm((PetscObject)A),&sf);CHKERRQ(ierr); 1925 ierr = PetscSFSetGraphLayout(sf,A->cmap,nb,NULL,PETSC_USE_POINTER,a->garray);CHKERRQ(ierr); 1926 ierr = PetscSFSetFromOptions(sf);CHKERRQ(ierr); 1927 ierr = PetscMemzero(o_nnz,na*sizeof(PetscInt));CHKERRQ(ierr); 1928 ierr = PetscSFReduceBegin(sf,MPIU_INT,g_nnz,o_nnz,MPIU_SUM);CHKERRQ(ierr); 1929 ierr = PetscSFReduceEnd(sf,MPIU_INT,g_nnz,o_nnz,MPIU_SUM);CHKERRQ(ierr); 1930 ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 1931 1932 ierr = MatCreate(PetscObjectComm((PetscObject)A),&B);CHKERRQ(ierr); 1933 ierr = MatSetSizes(B,A->cmap->n,A->rmap->n,N,M);CHKERRQ(ierr); 1934 ierr = MatSetBlockSizes(B,PetscAbs(A->cmap->bs),PetscAbs(A->rmap->bs));CHKERRQ(ierr); 1935 ierr = MatSetType(B,((PetscObject)A)->type_name);CHKERRQ(ierr); 1936 ierr = MatMPIAIJSetPreallocation(B,0,d_nnz,0,o_nnz);CHKERRQ(ierr); 1937 ierr = PetscFree4(d_nnz,o_nnz,g_nnz,oloc);CHKERRQ(ierr); 1938 } else { 1939 B = *matout; 1940 ierr = MatSetOption(B,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_TRUE);CHKERRQ(ierr); 1941 for (i=0; i<ai[ma]; i++) aj[i] += cstart; /* global col index to be used by MatSetValues() */ 1942 } 1943 1944 /* copy over the A part */ 1945 array = Aloc->a; 1946 row = A->rmap->rstart; 1947 for (i=0; i<ma; i++) { 1948 ncol = ai[i+1]-ai[i]; 1949 ierr = MatSetValues(B,ncol,aj,1,&row,array,INSERT_VALUES);CHKERRQ(ierr); 1950 row++; 1951 array += ncol; aj += ncol; 1952 } 1953 aj = Aloc->j; 1954 for (i=0; i<ai[ma]; i++) aj[i] -= cstart; /* resume local col index */ 1955 1956 /* copy over the B part */ 1957 ierr = PetscCalloc1(bi[mb],&cols);CHKERRQ(ierr); 1958 array = Bloc->a; 1959 row = A->rmap->rstart; 1960 for (i=0; i<bi[mb]; i++) cols[i] = a->garray[bj[i]]; 1961 cols_tmp = cols; 1962 for (i=0; i<mb; i++) { 1963 ncol = bi[i+1]-bi[i]; 1964 ierr = MatSetValues(B,ncol,cols_tmp,1,&row,array,INSERT_VALUES);CHKERRQ(ierr); 1965 row++; 1966 array += ncol; cols_tmp += ncol; 1967 } 1968 ierr = PetscFree(cols);CHKERRQ(ierr); 1969 1970 ierr = MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1971 ierr = MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1972 if (reuse == MAT_INITIAL_MATRIX || reuse == MAT_REUSE_MATRIX) { 1973 *matout = B; 1974 } else { 1975 ierr = MatHeaderMerge(A,&B);CHKERRQ(ierr); 1976 } 1977 PetscFunctionReturn(0); 1978 } 1979 1980 PetscErrorCode MatDiagonalScale_MPIAIJ(Mat mat,Vec ll,Vec rr) 1981 { 1982 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 1983 Mat a = aij->A,b = aij->B; 1984 PetscErrorCode ierr; 1985 PetscInt s1,s2,s3; 1986 1987 PetscFunctionBegin; 1988 ierr = MatGetLocalSize(mat,&s2,&s3);CHKERRQ(ierr); 1989 if (rr) { 1990 ierr = VecGetLocalSize(rr,&s1);CHKERRQ(ierr); 1991 if (s1!=s3) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"right vector non-conforming local size"); 1992 /* Overlap communication with computation. */ 1993 ierr = VecScatterBegin(aij->Mvctx,rr,aij->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1994 } 1995 if (ll) { 1996 ierr = VecGetLocalSize(ll,&s1);CHKERRQ(ierr); 1997 if (s1!=s2) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"left vector non-conforming local size"); 1998 ierr = (*b->ops->diagonalscale)(b,ll,0);CHKERRQ(ierr); 1999 } 2000 /* scale the diagonal block */ 2001 ierr = (*a->ops->diagonalscale)(a,ll,rr);CHKERRQ(ierr); 2002 2003 if (rr) { 2004 /* Do a scatter end and then right scale the off-diagonal block */ 2005 ierr = VecScatterEnd(aij->Mvctx,rr,aij->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2006 ierr = (*b->ops->diagonalscale)(b,0,aij->lvec);CHKERRQ(ierr); 2007 } 2008 PetscFunctionReturn(0); 2009 } 2010 2011 PetscErrorCode MatSetUnfactored_MPIAIJ(Mat A) 2012 { 2013 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 2014 PetscErrorCode ierr; 2015 2016 PetscFunctionBegin; 2017 ierr = MatSetUnfactored(a->A);CHKERRQ(ierr); 2018 PetscFunctionReturn(0); 2019 } 2020 2021 PetscErrorCode MatEqual_MPIAIJ(Mat A,Mat B,PetscBool *flag) 2022 { 2023 Mat_MPIAIJ *matB = (Mat_MPIAIJ*)B->data,*matA = (Mat_MPIAIJ*)A->data; 2024 Mat a,b,c,d; 2025 PetscBool flg; 2026 PetscErrorCode ierr; 2027 2028 PetscFunctionBegin; 2029 a = matA->A; b = matA->B; 2030 c = matB->A; d = matB->B; 2031 2032 ierr = MatEqual(a,c,&flg);CHKERRQ(ierr); 2033 if (flg) { 2034 ierr = MatEqual(b,d,&flg);CHKERRQ(ierr); 2035 } 2036 ierr = MPIU_Allreduce(&flg,flag,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)A));CHKERRQ(ierr); 2037 PetscFunctionReturn(0); 2038 } 2039 2040 PetscErrorCode MatCopy_MPIAIJ(Mat A,Mat B,MatStructure str) 2041 { 2042 PetscErrorCode ierr; 2043 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 2044 Mat_MPIAIJ *b = (Mat_MPIAIJ*)B->data; 2045 2046 PetscFunctionBegin; 2047 /* If the two matrices don't have the same copy implementation, they aren't compatible for fast copy. */ 2048 if ((str != SAME_NONZERO_PATTERN) || (A->ops->copy != B->ops->copy)) { 2049 /* because of the column compression in the off-processor part of the matrix a->B, 2050 the number of columns in a->B and b->B may be different, hence we cannot call 2051 the MatCopy() directly on the two parts. If need be, we can provide a more 2052 efficient copy than the MatCopy_Basic() by first uncompressing the a->B matrices 2053 then copying the submatrices */ 2054 ierr = MatCopy_Basic(A,B,str);CHKERRQ(ierr); 2055 } else { 2056 ierr = MatCopy(a->A,b->A,str);CHKERRQ(ierr); 2057 ierr = MatCopy(a->B,b->B,str);CHKERRQ(ierr); 2058 } 2059 PetscFunctionReturn(0); 2060 } 2061 2062 PetscErrorCode MatSetUp_MPIAIJ(Mat A) 2063 { 2064 PetscErrorCode ierr; 2065 2066 PetscFunctionBegin; 2067 ierr = MatMPIAIJSetPreallocation(A,PETSC_DEFAULT,0,PETSC_DEFAULT,0);CHKERRQ(ierr); 2068 PetscFunctionReturn(0); 2069 } 2070 2071 /* 2072 Computes the number of nonzeros per row needed for preallocation when X and Y 2073 have different nonzero structure. 2074 */ 2075 PetscErrorCode MatAXPYGetPreallocation_MPIX_private(PetscInt m,const PetscInt *xi,const PetscInt *xj,const PetscInt *xltog,const PetscInt *yi,const PetscInt *yj,const PetscInt *yltog,PetscInt *nnz) 2076 { 2077 PetscInt i,j,k,nzx,nzy; 2078 2079 PetscFunctionBegin; 2080 /* Set the number of nonzeros in the new matrix */ 2081 for (i=0; i<m; i++) { 2082 const PetscInt *xjj = xj+xi[i],*yjj = yj+yi[i]; 2083 nzx = xi[i+1] - xi[i]; 2084 nzy = yi[i+1] - yi[i]; 2085 nnz[i] = 0; 2086 for (j=0,k=0; j<nzx; j++) { /* Point in X */ 2087 for (; k<nzy && yltog[yjj[k]]<xltog[xjj[j]]; k++) nnz[i]++; /* Catch up to X */ 2088 if (k<nzy && yltog[yjj[k]]==xltog[xjj[j]]) k++; /* Skip duplicate */ 2089 nnz[i]++; 2090 } 2091 for (; k<nzy; k++) nnz[i]++; 2092 } 2093 PetscFunctionReturn(0); 2094 } 2095 2096 /* This is the same as MatAXPYGetPreallocation_SeqAIJ, except that the local-to-global map is provided */ 2097 static PetscErrorCode MatAXPYGetPreallocation_MPIAIJ(Mat Y,const PetscInt *yltog,Mat X,const PetscInt *xltog,PetscInt *nnz) 2098 { 2099 PetscErrorCode ierr; 2100 PetscInt m = Y->rmap->N; 2101 Mat_SeqAIJ *x = (Mat_SeqAIJ*)X->data; 2102 Mat_SeqAIJ *y = (Mat_SeqAIJ*)Y->data; 2103 2104 PetscFunctionBegin; 2105 ierr = MatAXPYGetPreallocation_MPIX_private(m,x->i,x->j,xltog,y->i,y->j,yltog,nnz);CHKERRQ(ierr); 2106 PetscFunctionReturn(0); 2107 } 2108 2109 PetscErrorCode MatAXPY_MPIAIJ(Mat Y,PetscScalar a,Mat X,MatStructure str) 2110 { 2111 PetscErrorCode ierr; 2112 Mat_MPIAIJ *xx = (Mat_MPIAIJ*)X->data,*yy = (Mat_MPIAIJ*)Y->data; 2113 PetscBLASInt bnz,one=1; 2114 Mat_SeqAIJ *x,*y; 2115 2116 PetscFunctionBegin; 2117 if (str == SAME_NONZERO_PATTERN) { 2118 PetscScalar alpha = a; 2119 x = (Mat_SeqAIJ*)xx->A->data; 2120 ierr = PetscBLASIntCast(x->nz,&bnz);CHKERRQ(ierr); 2121 y = (Mat_SeqAIJ*)yy->A->data; 2122 PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&bnz,&alpha,x->a,&one,y->a,&one)); 2123 x = (Mat_SeqAIJ*)xx->B->data; 2124 y = (Mat_SeqAIJ*)yy->B->data; 2125 ierr = PetscBLASIntCast(x->nz,&bnz);CHKERRQ(ierr); 2126 PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&bnz,&alpha,x->a,&one,y->a,&one)); 2127 ierr = PetscObjectStateIncrease((PetscObject)Y);CHKERRQ(ierr); 2128 } else if (str == SUBSET_NONZERO_PATTERN) { /* nonzeros of X is a subset of Y's */ 2129 ierr = MatAXPY_Basic(Y,a,X,str);CHKERRQ(ierr); 2130 } else { 2131 Mat B; 2132 PetscInt *nnz_d,*nnz_o; 2133 ierr = PetscMalloc1(yy->A->rmap->N,&nnz_d);CHKERRQ(ierr); 2134 ierr = PetscMalloc1(yy->B->rmap->N,&nnz_o);CHKERRQ(ierr); 2135 ierr = MatCreate(PetscObjectComm((PetscObject)Y),&B);CHKERRQ(ierr); 2136 ierr = PetscObjectSetName((PetscObject)B,((PetscObject)Y)->name);CHKERRQ(ierr); 2137 ierr = MatSetSizes(B,Y->rmap->n,Y->cmap->n,Y->rmap->N,Y->cmap->N);CHKERRQ(ierr); 2138 ierr = MatSetBlockSizesFromMats(B,Y,Y);CHKERRQ(ierr); 2139 ierr = MatSetType(B,MATMPIAIJ);CHKERRQ(ierr); 2140 ierr = MatAXPYGetPreallocation_SeqAIJ(yy->A,xx->A,nnz_d);CHKERRQ(ierr); 2141 ierr = MatAXPYGetPreallocation_MPIAIJ(yy->B,yy->garray,xx->B,xx->garray,nnz_o);CHKERRQ(ierr); 2142 ierr = MatMPIAIJSetPreallocation(B,0,nnz_d,0,nnz_o);CHKERRQ(ierr); 2143 ierr = MatAXPY_BasicWithPreallocation(B,Y,a,X,str);CHKERRQ(ierr); 2144 ierr = MatHeaderReplace(Y,&B);CHKERRQ(ierr); 2145 ierr = PetscFree(nnz_d);CHKERRQ(ierr); 2146 ierr = PetscFree(nnz_o);CHKERRQ(ierr); 2147 } 2148 PetscFunctionReturn(0); 2149 } 2150 2151 extern PetscErrorCode MatConjugate_SeqAIJ(Mat); 2152 2153 PetscErrorCode MatConjugate_MPIAIJ(Mat mat) 2154 { 2155 #if defined(PETSC_USE_COMPLEX) 2156 PetscErrorCode ierr; 2157 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 2158 2159 PetscFunctionBegin; 2160 ierr = MatConjugate_SeqAIJ(aij->A);CHKERRQ(ierr); 2161 ierr = MatConjugate_SeqAIJ(aij->B);CHKERRQ(ierr); 2162 #else 2163 PetscFunctionBegin; 2164 #endif 2165 PetscFunctionReturn(0); 2166 } 2167 2168 PetscErrorCode MatRealPart_MPIAIJ(Mat A) 2169 { 2170 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 2171 PetscErrorCode ierr; 2172 2173 PetscFunctionBegin; 2174 ierr = MatRealPart(a->A);CHKERRQ(ierr); 2175 ierr = MatRealPart(a->B);CHKERRQ(ierr); 2176 PetscFunctionReturn(0); 2177 } 2178 2179 PetscErrorCode MatImaginaryPart_MPIAIJ(Mat A) 2180 { 2181 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 2182 PetscErrorCode ierr; 2183 2184 PetscFunctionBegin; 2185 ierr = MatImaginaryPart(a->A);CHKERRQ(ierr); 2186 ierr = MatImaginaryPart(a->B);CHKERRQ(ierr); 2187 PetscFunctionReturn(0); 2188 } 2189 2190 PetscErrorCode MatGetRowMaxAbs_MPIAIJ(Mat A, Vec v, PetscInt idx[]) 2191 { 2192 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 2193 PetscErrorCode ierr; 2194 PetscInt i,*idxb = 0; 2195 PetscScalar *va,*vb; 2196 Vec vtmp; 2197 2198 PetscFunctionBegin; 2199 ierr = MatGetRowMaxAbs(a->A,v,idx);CHKERRQ(ierr); 2200 ierr = VecGetArray(v,&va);CHKERRQ(ierr); 2201 if (idx) { 2202 for (i=0; i<A->rmap->n; i++) { 2203 if (PetscAbsScalar(va[i])) idx[i] += A->cmap->rstart; 2204 } 2205 } 2206 2207 ierr = VecCreateSeq(PETSC_COMM_SELF,A->rmap->n,&vtmp);CHKERRQ(ierr); 2208 if (idx) { 2209 ierr = PetscMalloc1(A->rmap->n,&idxb);CHKERRQ(ierr); 2210 } 2211 ierr = MatGetRowMaxAbs(a->B,vtmp,idxb);CHKERRQ(ierr); 2212 ierr = VecGetArray(vtmp,&vb);CHKERRQ(ierr); 2213 2214 for (i=0; i<A->rmap->n; i++) { 2215 if (PetscAbsScalar(va[i]) < PetscAbsScalar(vb[i])) { 2216 va[i] = vb[i]; 2217 if (idx) idx[i] = a->garray[idxb[i]]; 2218 } 2219 } 2220 2221 ierr = VecRestoreArray(v,&va);CHKERRQ(ierr); 2222 ierr = VecRestoreArray(vtmp,&vb);CHKERRQ(ierr); 2223 ierr = PetscFree(idxb);CHKERRQ(ierr); 2224 ierr = VecDestroy(&vtmp);CHKERRQ(ierr); 2225 PetscFunctionReturn(0); 2226 } 2227 2228 PetscErrorCode MatGetRowMinAbs_MPIAIJ(Mat A, Vec v, PetscInt idx[]) 2229 { 2230 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 2231 PetscErrorCode ierr; 2232 PetscInt i,*idxb = 0; 2233 PetscScalar *va,*vb; 2234 Vec vtmp; 2235 2236 PetscFunctionBegin; 2237 ierr = MatGetRowMinAbs(a->A,v,idx);CHKERRQ(ierr); 2238 ierr = VecGetArray(v,&va);CHKERRQ(ierr); 2239 if (idx) { 2240 for (i=0; i<A->cmap->n; i++) { 2241 if (PetscAbsScalar(va[i])) idx[i] += A->cmap->rstart; 2242 } 2243 } 2244 2245 ierr = VecCreateSeq(PETSC_COMM_SELF,A->rmap->n,&vtmp);CHKERRQ(ierr); 2246 if (idx) { 2247 ierr = PetscMalloc1(A->rmap->n,&idxb);CHKERRQ(ierr); 2248 } 2249 ierr = MatGetRowMinAbs(a->B,vtmp,idxb);CHKERRQ(ierr); 2250 ierr = VecGetArray(vtmp,&vb);CHKERRQ(ierr); 2251 2252 for (i=0; i<A->rmap->n; i++) { 2253 if (PetscAbsScalar(va[i]) > PetscAbsScalar(vb[i])) { 2254 va[i] = vb[i]; 2255 if (idx) idx[i] = a->garray[idxb[i]]; 2256 } 2257 } 2258 2259 ierr = VecRestoreArray(v,&va);CHKERRQ(ierr); 2260 ierr = VecRestoreArray(vtmp,&vb);CHKERRQ(ierr); 2261 ierr = PetscFree(idxb);CHKERRQ(ierr); 2262 ierr = VecDestroy(&vtmp);CHKERRQ(ierr); 2263 PetscFunctionReturn(0); 2264 } 2265 2266 PetscErrorCode MatGetRowMin_MPIAIJ(Mat A, Vec v, PetscInt idx[]) 2267 { 2268 Mat_MPIAIJ *mat = (Mat_MPIAIJ*) A->data; 2269 PetscInt n = A->rmap->n; 2270 PetscInt cstart = A->cmap->rstart; 2271 PetscInt *cmap = mat->garray; 2272 PetscInt *diagIdx, *offdiagIdx; 2273 Vec diagV, offdiagV; 2274 PetscScalar *a, *diagA, *offdiagA; 2275 PetscInt r; 2276 PetscErrorCode ierr; 2277 2278 PetscFunctionBegin; 2279 ierr = PetscMalloc2(n,&diagIdx,n,&offdiagIdx);CHKERRQ(ierr); 2280 ierr = VecCreateSeq(PetscObjectComm((PetscObject)A), n, &diagV);CHKERRQ(ierr); 2281 ierr = VecCreateSeq(PetscObjectComm((PetscObject)A), n, &offdiagV);CHKERRQ(ierr); 2282 ierr = MatGetRowMin(mat->A, diagV, diagIdx);CHKERRQ(ierr); 2283 ierr = MatGetRowMin(mat->B, offdiagV, offdiagIdx);CHKERRQ(ierr); 2284 ierr = VecGetArray(v, &a);CHKERRQ(ierr); 2285 ierr = VecGetArray(diagV, &diagA);CHKERRQ(ierr); 2286 ierr = VecGetArray(offdiagV, &offdiagA);CHKERRQ(ierr); 2287 for (r = 0; r < n; ++r) { 2288 if (PetscAbsScalar(diagA[r]) <= PetscAbsScalar(offdiagA[r])) { 2289 a[r] = diagA[r]; 2290 idx[r] = cstart + diagIdx[r]; 2291 } else { 2292 a[r] = offdiagA[r]; 2293 idx[r] = cmap[offdiagIdx[r]]; 2294 } 2295 } 2296 ierr = VecRestoreArray(v, &a);CHKERRQ(ierr); 2297 ierr = VecRestoreArray(diagV, &diagA);CHKERRQ(ierr); 2298 ierr = VecRestoreArray(offdiagV, &offdiagA);CHKERRQ(ierr); 2299 ierr = VecDestroy(&diagV);CHKERRQ(ierr); 2300 ierr = VecDestroy(&offdiagV);CHKERRQ(ierr); 2301 ierr = PetscFree2(diagIdx, offdiagIdx);CHKERRQ(ierr); 2302 PetscFunctionReturn(0); 2303 } 2304 2305 PetscErrorCode MatGetRowMax_MPIAIJ(Mat A, Vec v, PetscInt idx[]) 2306 { 2307 Mat_MPIAIJ *mat = (Mat_MPIAIJ*) A->data; 2308 PetscInt n = A->rmap->n; 2309 PetscInt cstart = A->cmap->rstart; 2310 PetscInt *cmap = mat->garray; 2311 PetscInt *diagIdx, *offdiagIdx; 2312 Vec diagV, offdiagV; 2313 PetscScalar *a, *diagA, *offdiagA; 2314 PetscInt r; 2315 PetscErrorCode ierr; 2316 2317 PetscFunctionBegin; 2318 ierr = PetscMalloc2(n,&diagIdx,n,&offdiagIdx);CHKERRQ(ierr); 2319 ierr = VecCreateSeq(PETSC_COMM_SELF, n, &diagV);CHKERRQ(ierr); 2320 ierr = VecCreateSeq(PETSC_COMM_SELF, n, &offdiagV);CHKERRQ(ierr); 2321 ierr = MatGetRowMax(mat->A, diagV, diagIdx);CHKERRQ(ierr); 2322 ierr = MatGetRowMax(mat->B, offdiagV, offdiagIdx);CHKERRQ(ierr); 2323 ierr = VecGetArray(v, &a);CHKERRQ(ierr); 2324 ierr = VecGetArray(diagV, &diagA);CHKERRQ(ierr); 2325 ierr = VecGetArray(offdiagV, &offdiagA);CHKERRQ(ierr); 2326 for (r = 0; r < n; ++r) { 2327 if (PetscAbsScalar(diagA[r]) >= PetscAbsScalar(offdiagA[r])) { 2328 a[r] = diagA[r]; 2329 idx[r] = cstart + diagIdx[r]; 2330 } else { 2331 a[r] = offdiagA[r]; 2332 idx[r] = cmap[offdiagIdx[r]]; 2333 } 2334 } 2335 ierr = VecRestoreArray(v, &a);CHKERRQ(ierr); 2336 ierr = VecRestoreArray(diagV, &diagA);CHKERRQ(ierr); 2337 ierr = VecRestoreArray(offdiagV, &offdiagA);CHKERRQ(ierr); 2338 ierr = VecDestroy(&diagV);CHKERRQ(ierr); 2339 ierr = VecDestroy(&offdiagV);CHKERRQ(ierr); 2340 ierr = PetscFree2(diagIdx, offdiagIdx);CHKERRQ(ierr); 2341 PetscFunctionReturn(0); 2342 } 2343 2344 PetscErrorCode MatGetSeqNonzeroStructure_MPIAIJ(Mat mat,Mat *newmat) 2345 { 2346 PetscErrorCode ierr; 2347 Mat *dummy; 2348 2349 PetscFunctionBegin; 2350 ierr = MatCreateSubMatrix_MPIAIJ_All(mat,MAT_DO_NOT_GET_VALUES,MAT_INITIAL_MATRIX,&dummy);CHKERRQ(ierr); 2351 *newmat = *dummy; 2352 ierr = PetscFree(dummy);CHKERRQ(ierr); 2353 PetscFunctionReturn(0); 2354 } 2355 2356 PetscErrorCode MatInvertBlockDiagonal_MPIAIJ(Mat A,const PetscScalar **values) 2357 { 2358 Mat_MPIAIJ *a = (Mat_MPIAIJ*) A->data; 2359 PetscErrorCode ierr; 2360 2361 PetscFunctionBegin; 2362 ierr = MatInvertBlockDiagonal(a->A,values);CHKERRQ(ierr); 2363 A->factorerrortype = a->A->factorerrortype; 2364 PetscFunctionReturn(0); 2365 } 2366 2367 static PetscErrorCode MatSetRandom_MPIAIJ(Mat x,PetscRandom rctx) 2368 { 2369 PetscErrorCode ierr; 2370 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)x->data; 2371 2372 PetscFunctionBegin; 2373 ierr = MatSetRandom(aij->A,rctx);CHKERRQ(ierr); 2374 ierr = MatSetRandom(aij->B,rctx);CHKERRQ(ierr); 2375 ierr = MatAssemblyBegin(x,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2376 ierr = MatAssemblyEnd(x,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2377 PetscFunctionReturn(0); 2378 } 2379 2380 PetscErrorCode MatMPIAIJSetUseScalableIncreaseOverlap_MPIAIJ(Mat A,PetscBool sc) 2381 { 2382 PetscFunctionBegin; 2383 if (sc) A->ops->increaseoverlap = MatIncreaseOverlap_MPIAIJ_Scalable; 2384 else A->ops->increaseoverlap = MatIncreaseOverlap_MPIAIJ; 2385 PetscFunctionReturn(0); 2386 } 2387 2388 /*@ 2389 MatMPIAIJSetUseScalableIncreaseOverlap - Determine if the matrix uses a scalable algorithm to compute the overlap 2390 2391 Collective on Mat 2392 2393 Input Parameters: 2394 + A - the matrix 2395 - sc - PETSC_TRUE indicates use the scalable algorithm (default is not to use the scalable algorithm) 2396 2397 Level: advanced 2398 2399 @*/ 2400 PetscErrorCode MatMPIAIJSetUseScalableIncreaseOverlap(Mat A,PetscBool sc) 2401 { 2402 PetscErrorCode ierr; 2403 2404 PetscFunctionBegin; 2405 ierr = PetscTryMethod(A,"MatMPIAIJSetUseScalableIncreaseOverlap_C",(Mat,PetscBool),(A,sc));CHKERRQ(ierr); 2406 PetscFunctionReturn(0); 2407 } 2408 2409 PetscErrorCode MatSetFromOptions_MPIAIJ(PetscOptionItems *PetscOptionsObject,Mat A) 2410 { 2411 PetscErrorCode ierr; 2412 PetscBool sc = PETSC_FALSE,flg; 2413 2414 PetscFunctionBegin; 2415 ierr = PetscOptionsHead(PetscOptionsObject,"MPIAIJ options");CHKERRQ(ierr); 2416 ierr = PetscObjectOptionsBegin((PetscObject)A); 2417 if (A->ops->increaseoverlap == MatIncreaseOverlap_MPIAIJ_Scalable) sc = PETSC_TRUE; 2418 ierr = PetscOptionsBool("-mat_increase_overlap_scalable","Use a scalable algorithm to compute the overlap","MatIncreaseOverlap",sc,&sc,&flg);CHKERRQ(ierr); 2419 if (flg) { 2420 ierr = MatMPIAIJSetUseScalableIncreaseOverlap(A,sc);CHKERRQ(ierr); 2421 } 2422 ierr = PetscOptionsEnd();CHKERRQ(ierr); 2423 PetscFunctionReturn(0); 2424 } 2425 2426 PetscErrorCode MatShift_MPIAIJ(Mat Y,PetscScalar a) 2427 { 2428 PetscErrorCode ierr; 2429 Mat_MPIAIJ *maij = (Mat_MPIAIJ*)Y->data; 2430 Mat_SeqAIJ *aij = (Mat_SeqAIJ*)maij->A->data; 2431 2432 PetscFunctionBegin; 2433 if (!Y->preallocated) { 2434 ierr = MatMPIAIJSetPreallocation(Y,1,NULL,0,NULL);CHKERRQ(ierr); 2435 } else if (!aij->nz) { 2436 PetscInt nonew = aij->nonew; 2437 ierr = MatSeqAIJSetPreallocation(maij->A,1,NULL);CHKERRQ(ierr); 2438 aij->nonew = nonew; 2439 } 2440 ierr = MatShift_Basic(Y,a);CHKERRQ(ierr); 2441 PetscFunctionReturn(0); 2442 } 2443 2444 PetscErrorCode MatMissingDiagonal_MPIAIJ(Mat A,PetscBool *missing,PetscInt *d) 2445 { 2446 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 2447 PetscErrorCode ierr; 2448 2449 PetscFunctionBegin; 2450 if (A->rmap->n != A->cmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Only works for square matrices"); 2451 ierr = MatMissingDiagonal(a->A,missing,d);CHKERRQ(ierr); 2452 if (d) { 2453 PetscInt rstart; 2454 ierr = MatGetOwnershipRange(A,&rstart,NULL);CHKERRQ(ierr); 2455 *d += rstart; 2456 2457 } 2458 PetscFunctionReturn(0); 2459 } 2460 2461 2462 /* -------------------------------------------------------------------*/ 2463 static struct _MatOps MatOps_Values = {MatSetValues_MPIAIJ, 2464 MatGetRow_MPIAIJ, 2465 MatRestoreRow_MPIAIJ, 2466 MatMult_MPIAIJ, 2467 /* 4*/ MatMultAdd_MPIAIJ, 2468 MatMultTranspose_MPIAIJ, 2469 MatMultTransposeAdd_MPIAIJ, 2470 0, 2471 0, 2472 0, 2473 /*10*/ 0, 2474 0, 2475 0, 2476 MatSOR_MPIAIJ, 2477 MatTranspose_MPIAIJ, 2478 /*15*/ MatGetInfo_MPIAIJ, 2479 MatEqual_MPIAIJ, 2480 MatGetDiagonal_MPIAIJ, 2481 MatDiagonalScale_MPIAIJ, 2482 MatNorm_MPIAIJ, 2483 /*20*/ MatAssemblyBegin_MPIAIJ, 2484 MatAssemblyEnd_MPIAIJ, 2485 MatSetOption_MPIAIJ, 2486 MatZeroEntries_MPIAIJ, 2487 /*24*/ MatZeroRows_MPIAIJ, 2488 0, 2489 0, 2490 0, 2491 0, 2492 /*29*/ MatSetUp_MPIAIJ, 2493 0, 2494 0, 2495 MatGetDiagonalBlock_MPIAIJ, 2496 0, 2497 /*34*/ MatDuplicate_MPIAIJ, 2498 0, 2499 0, 2500 0, 2501 0, 2502 /*39*/ MatAXPY_MPIAIJ, 2503 MatCreateSubMatrices_MPIAIJ, 2504 MatIncreaseOverlap_MPIAIJ, 2505 MatGetValues_MPIAIJ, 2506 MatCopy_MPIAIJ, 2507 /*44*/ MatGetRowMax_MPIAIJ, 2508 MatScale_MPIAIJ, 2509 MatShift_MPIAIJ, 2510 MatDiagonalSet_MPIAIJ, 2511 MatZeroRowsColumns_MPIAIJ, 2512 /*49*/ MatSetRandom_MPIAIJ, 2513 0, 2514 0, 2515 0, 2516 0, 2517 /*54*/ MatFDColoringCreate_MPIXAIJ, 2518 0, 2519 MatSetUnfactored_MPIAIJ, 2520 MatPermute_MPIAIJ, 2521 0, 2522 /*59*/ MatCreateSubMatrix_MPIAIJ, 2523 MatDestroy_MPIAIJ, 2524 MatView_MPIAIJ, 2525 0, 2526 MatMatMatMult_MPIAIJ_MPIAIJ_MPIAIJ, 2527 /*64*/ MatMatMatMultSymbolic_MPIAIJ_MPIAIJ_MPIAIJ, 2528 MatMatMatMultNumeric_MPIAIJ_MPIAIJ_MPIAIJ, 2529 0, 2530 0, 2531 0, 2532 /*69*/ MatGetRowMaxAbs_MPIAIJ, 2533 MatGetRowMinAbs_MPIAIJ, 2534 0, 2535 0, 2536 0, 2537 0, 2538 /*75*/ MatFDColoringApply_AIJ, 2539 MatSetFromOptions_MPIAIJ, 2540 0, 2541 0, 2542 MatFindZeroDiagonals_MPIAIJ, 2543 /*80*/ 0, 2544 0, 2545 0, 2546 /*83*/ MatLoad_MPIAIJ, 2547 0, 2548 0, 2549 0, 2550 0, 2551 0, 2552 /*89*/ MatMatMult_MPIAIJ_MPIAIJ, 2553 MatMatMultSymbolic_MPIAIJ_MPIAIJ, 2554 MatMatMultNumeric_MPIAIJ_MPIAIJ, 2555 MatPtAP_MPIAIJ_MPIAIJ, 2556 MatPtAPSymbolic_MPIAIJ_MPIAIJ, 2557 /*94*/ MatPtAPNumeric_MPIAIJ_MPIAIJ, 2558 0, 2559 0, 2560 0, 2561 0, 2562 /*99*/ 0, 2563 0, 2564 0, 2565 MatConjugate_MPIAIJ, 2566 0, 2567 /*104*/MatSetValuesRow_MPIAIJ, 2568 MatRealPart_MPIAIJ, 2569 MatImaginaryPart_MPIAIJ, 2570 0, 2571 0, 2572 /*109*/0, 2573 0, 2574 MatGetRowMin_MPIAIJ, 2575 0, 2576 MatMissingDiagonal_MPIAIJ, 2577 /*114*/MatGetSeqNonzeroStructure_MPIAIJ, 2578 0, 2579 MatGetGhosts_MPIAIJ, 2580 0, 2581 0, 2582 /*119*/0, 2583 0, 2584 0, 2585 0, 2586 MatGetMultiProcBlock_MPIAIJ, 2587 /*124*/MatFindNonzeroRows_MPIAIJ, 2588 MatGetColumnNorms_MPIAIJ, 2589 MatInvertBlockDiagonal_MPIAIJ, 2590 0, 2591 MatCreateSubMatricesMPI_MPIAIJ, 2592 /*129*/0, 2593 MatTransposeMatMult_MPIAIJ_MPIAIJ, 2594 MatTransposeMatMultSymbolic_MPIAIJ_MPIAIJ, 2595 MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ, 2596 0, 2597 /*134*/0, 2598 0, 2599 0, 2600 0, 2601 0, 2602 /*139*/MatSetBlockSizes_MPIAIJ, 2603 0, 2604 0, 2605 MatFDColoringSetUp_MPIXAIJ, 2606 MatFindOffBlockDiagonalEntries_MPIAIJ, 2607 /*144*/MatCreateMPIMatConcatenateSeqMat_MPIAIJ 2608 }; 2609 2610 /* ----------------------------------------------------------------------------------------*/ 2611 2612 PetscErrorCode MatStoreValues_MPIAIJ(Mat mat) 2613 { 2614 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 2615 PetscErrorCode ierr; 2616 2617 PetscFunctionBegin; 2618 ierr = MatStoreValues(aij->A);CHKERRQ(ierr); 2619 ierr = MatStoreValues(aij->B);CHKERRQ(ierr); 2620 PetscFunctionReturn(0); 2621 } 2622 2623 PetscErrorCode MatRetrieveValues_MPIAIJ(Mat mat) 2624 { 2625 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 2626 PetscErrorCode ierr; 2627 2628 PetscFunctionBegin; 2629 ierr = MatRetrieveValues(aij->A);CHKERRQ(ierr); 2630 ierr = MatRetrieveValues(aij->B);CHKERRQ(ierr); 2631 PetscFunctionReturn(0); 2632 } 2633 2634 PetscErrorCode MatMPIAIJSetPreallocation_MPIAIJ(Mat B,PetscInt d_nz,const PetscInt d_nnz[],PetscInt o_nz,const PetscInt o_nnz[]) 2635 { 2636 Mat_MPIAIJ *b; 2637 PetscErrorCode ierr; 2638 2639 PetscFunctionBegin; 2640 ierr = PetscLayoutSetUp(B->rmap);CHKERRQ(ierr); 2641 ierr = PetscLayoutSetUp(B->cmap);CHKERRQ(ierr); 2642 b = (Mat_MPIAIJ*)B->data; 2643 2644 #if defined(PETSC_USE_CTABLE) 2645 ierr = PetscTableDestroy(&b->colmap);CHKERRQ(ierr); 2646 #else 2647 ierr = PetscFree(b->colmap);CHKERRQ(ierr); 2648 #endif 2649 ierr = PetscFree(b->garray);CHKERRQ(ierr); 2650 ierr = VecDestroy(&b->lvec);CHKERRQ(ierr); 2651 ierr = VecScatterDestroy(&b->Mvctx);CHKERRQ(ierr); 2652 2653 /* Because the B will have been resized we simply destroy it and create a new one each time */ 2654 ierr = MatDestroy(&b->B);CHKERRQ(ierr); 2655 ierr = MatCreate(PETSC_COMM_SELF,&b->B);CHKERRQ(ierr); 2656 ierr = MatSetSizes(b->B,B->rmap->n,B->cmap->N,B->rmap->n,B->cmap->N);CHKERRQ(ierr); 2657 ierr = MatSetBlockSizesFromMats(b->B,B,B);CHKERRQ(ierr); 2658 ierr = MatSetType(b->B,MATSEQAIJ);CHKERRQ(ierr); 2659 ierr = PetscLogObjectParent((PetscObject)B,(PetscObject)b->B);CHKERRQ(ierr); 2660 2661 if (!B->preallocated) { 2662 ierr = MatCreate(PETSC_COMM_SELF,&b->A);CHKERRQ(ierr); 2663 ierr = MatSetSizes(b->A,B->rmap->n,B->cmap->n,B->rmap->n,B->cmap->n);CHKERRQ(ierr); 2664 ierr = MatSetBlockSizesFromMats(b->A,B,B);CHKERRQ(ierr); 2665 ierr = MatSetType(b->A,MATSEQAIJ);CHKERRQ(ierr); 2666 ierr = PetscLogObjectParent((PetscObject)B,(PetscObject)b->A);CHKERRQ(ierr); 2667 } 2668 2669 ierr = MatSeqAIJSetPreallocation(b->A,d_nz,d_nnz);CHKERRQ(ierr); 2670 ierr = MatSeqAIJSetPreallocation(b->B,o_nz,o_nnz);CHKERRQ(ierr); 2671 B->preallocated = PETSC_TRUE; 2672 B->was_assembled = PETSC_FALSE; 2673 B->assembled = PETSC_FALSE;; 2674 PetscFunctionReturn(0); 2675 } 2676 2677 PetscErrorCode MatDuplicate_MPIAIJ(Mat matin,MatDuplicateOption cpvalues,Mat *newmat) 2678 { 2679 Mat mat; 2680 Mat_MPIAIJ *a,*oldmat = (Mat_MPIAIJ*)matin->data; 2681 PetscErrorCode ierr; 2682 2683 PetscFunctionBegin; 2684 *newmat = 0; 2685 ierr = MatCreate(PetscObjectComm((PetscObject)matin),&mat);CHKERRQ(ierr); 2686 ierr = MatSetSizes(mat,matin->rmap->n,matin->cmap->n,matin->rmap->N,matin->cmap->N);CHKERRQ(ierr); 2687 ierr = MatSetBlockSizesFromMats(mat,matin,matin);CHKERRQ(ierr); 2688 ierr = MatSetType(mat,((PetscObject)matin)->type_name);CHKERRQ(ierr); 2689 ierr = PetscMemcpy(mat->ops,matin->ops,sizeof(struct _MatOps));CHKERRQ(ierr); 2690 a = (Mat_MPIAIJ*)mat->data; 2691 2692 mat->factortype = matin->factortype; 2693 mat->assembled = PETSC_TRUE; 2694 mat->insertmode = NOT_SET_VALUES; 2695 mat->preallocated = PETSC_TRUE; 2696 2697 a->size = oldmat->size; 2698 a->rank = oldmat->rank; 2699 a->donotstash = oldmat->donotstash; 2700 a->roworiented = oldmat->roworiented; 2701 a->rowindices = 0; 2702 a->rowvalues = 0; 2703 a->getrowactive = PETSC_FALSE; 2704 2705 ierr = PetscLayoutReference(matin->rmap,&mat->rmap);CHKERRQ(ierr); 2706 ierr = PetscLayoutReference(matin->cmap,&mat->cmap);CHKERRQ(ierr); 2707 2708 if (oldmat->colmap) { 2709 #if defined(PETSC_USE_CTABLE) 2710 ierr = PetscTableCreateCopy(oldmat->colmap,&a->colmap);CHKERRQ(ierr); 2711 #else 2712 ierr = PetscMalloc1(mat->cmap->N,&a->colmap);CHKERRQ(ierr); 2713 ierr = PetscLogObjectMemory((PetscObject)mat,(mat->cmap->N)*sizeof(PetscInt));CHKERRQ(ierr); 2714 ierr = PetscMemcpy(a->colmap,oldmat->colmap,(mat->cmap->N)*sizeof(PetscInt));CHKERRQ(ierr); 2715 #endif 2716 } else a->colmap = 0; 2717 if (oldmat->garray) { 2718 PetscInt len; 2719 len = oldmat->B->cmap->n; 2720 ierr = PetscMalloc1(len+1,&a->garray);CHKERRQ(ierr); 2721 ierr = PetscLogObjectMemory((PetscObject)mat,len*sizeof(PetscInt));CHKERRQ(ierr); 2722 if (len) { ierr = PetscMemcpy(a->garray,oldmat->garray,len*sizeof(PetscInt));CHKERRQ(ierr); } 2723 } else a->garray = 0; 2724 2725 ierr = VecDuplicate(oldmat->lvec,&a->lvec);CHKERRQ(ierr); 2726 ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)a->lvec);CHKERRQ(ierr); 2727 ierr = VecScatterCopy(oldmat->Mvctx,&a->Mvctx);CHKERRQ(ierr); 2728 ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)a->Mvctx);CHKERRQ(ierr); 2729 ierr = MatDuplicate(oldmat->A,cpvalues,&a->A);CHKERRQ(ierr); 2730 ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)a->A);CHKERRQ(ierr); 2731 ierr = MatDuplicate(oldmat->B,cpvalues,&a->B);CHKERRQ(ierr); 2732 ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)a->B);CHKERRQ(ierr); 2733 ierr = PetscFunctionListDuplicate(((PetscObject)matin)->qlist,&((PetscObject)mat)->qlist);CHKERRQ(ierr); 2734 *newmat = mat; 2735 PetscFunctionReturn(0); 2736 } 2737 2738 2739 2740 PetscErrorCode MatLoad_MPIAIJ(Mat newMat, PetscViewer viewer) 2741 { 2742 PetscScalar *vals,*svals; 2743 MPI_Comm comm; 2744 PetscErrorCode ierr; 2745 PetscMPIInt rank,size,tag = ((PetscObject)viewer)->tag; 2746 PetscInt i,nz,j,rstart,rend,mmax,maxnz = 0; 2747 PetscInt header[4],*rowlengths = 0,M,N,m,*cols; 2748 PetscInt *ourlens = NULL,*procsnz = NULL,*offlens = NULL,jj,*mycols,*smycols; 2749 PetscInt cend,cstart,n,*rowners; 2750 int fd; 2751 PetscInt bs = newMat->rmap->bs; 2752 2753 PetscFunctionBegin; 2754 /* force binary viewer to load .info file if it has not yet done so */ 2755 ierr = PetscViewerSetUp(viewer);CHKERRQ(ierr); 2756 ierr = PetscObjectGetComm((PetscObject)viewer,&comm);CHKERRQ(ierr); 2757 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 2758 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 2759 ierr = PetscViewerBinaryGetDescriptor(viewer,&fd);CHKERRQ(ierr); 2760 if (!rank) { 2761 ierr = PetscBinaryRead(fd,(char*)header,4,PETSC_INT);CHKERRQ(ierr); 2762 if (header[0] != MAT_FILE_CLASSID) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED,"not matrix object"); 2763 if (header[3] < 0) SETERRQ(PetscObjectComm((PetscObject)newMat),PETSC_ERR_FILE_UNEXPECTED,"Matrix stored in special format on disk,cannot load as MATMPIAIJ"); 2764 } 2765 2766 ierr = PetscOptionsBegin(comm,NULL,"Options for loading MATMPIAIJ matrix","Mat");CHKERRQ(ierr); 2767 ierr = PetscOptionsInt("-matload_block_size","Set the blocksize used to store the matrix","MatLoad",bs,&bs,NULL);CHKERRQ(ierr); 2768 ierr = PetscOptionsEnd();CHKERRQ(ierr); 2769 if (bs < 0) bs = 1; 2770 2771 ierr = MPI_Bcast(header+1,3,MPIU_INT,0,comm);CHKERRQ(ierr); 2772 M = header[1]; N = header[2]; 2773 2774 /* If global sizes are set, check if they are consistent with that given in the file */ 2775 if (newMat->rmap->N >= 0 && newMat->rmap->N != M) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED,"Inconsistent # of rows:Matrix in file has (%D) and input matrix has (%D)",newMat->rmap->N,M); 2776 if (newMat->cmap->N >=0 && newMat->cmap->N != N) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED,"Inconsistent # of cols:Matrix in file has (%D) and input matrix has (%D)",newMat->cmap->N,N); 2777 2778 /* determine ownership of all (block) rows */ 2779 if (M%bs) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED, "Inconsistent # of rows (%d) and block size (%d)",M,bs); 2780 if (newMat->rmap->n < 0) m = bs*((M/bs)/size + (((M/bs) % size) > rank)); /* PETSC_DECIDE */ 2781 else m = newMat->rmap->n; /* Set by user */ 2782 2783 ierr = PetscMalloc1(size+1,&rowners);CHKERRQ(ierr); 2784 ierr = MPI_Allgather(&m,1,MPIU_INT,rowners+1,1,MPIU_INT,comm);CHKERRQ(ierr); 2785 2786 /* First process needs enough room for process with most rows */ 2787 if (!rank) { 2788 mmax = rowners[1]; 2789 for (i=2; i<=size; i++) { 2790 mmax = PetscMax(mmax, rowners[i]); 2791 } 2792 } else mmax = -1; /* unused, but compilers complain */ 2793 2794 rowners[0] = 0; 2795 for (i=2; i<=size; i++) { 2796 rowners[i] += rowners[i-1]; 2797 } 2798 rstart = rowners[rank]; 2799 rend = rowners[rank+1]; 2800 2801 /* distribute row lengths to all processors */ 2802 ierr = PetscMalloc2(m,&ourlens,m,&offlens);CHKERRQ(ierr); 2803 if (!rank) { 2804 ierr = PetscBinaryRead(fd,ourlens,m,PETSC_INT);CHKERRQ(ierr); 2805 ierr = PetscMalloc1(mmax,&rowlengths);CHKERRQ(ierr); 2806 ierr = PetscCalloc1(size,&procsnz);CHKERRQ(ierr); 2807 for (j=0; j<m; j++) { 2808 procsnz[0] += ourlens[j]; 2809 } 2810 for (i=1; i<size; i++) { 2811 ierr = PetscBinaryRead(fd,rowlengths,rowners[i+1]-rowners[i],PETSC_INT);CHKERRQ(ierr); 2812 /* calculate the number of nonzeros on each processor */ 2813 for (j=0; j<rowners[i+1]-rowners[i]; j++) { 2814 procsnz[i] += rowlengths[j]; 2815 } 2816 ierr = MPIULong_Send(rowlengths,rowners[i+1]-rowners[i],MPIU_INT,i,tag,comm);CHKERRQ(ierr); 2817 } 2818 ierr = PetscFree(rowlengths);CHKERRQ(ierr); 2819 } else { 2820 ierr = MPIULong_Recv(ourlens,m,MPIU_INT,0,tag,comm);CHKERRQ(ierr); 2821 } 2822 2823 if (!rank) { 2824 /* determine max buffer needed and allocate it */ 2825 maxnz = 0; 2826 for (i=0; i<size; i++) { 2827 maxnz = PetscMax(maxnz,procsnz[i]); 2828 } 2829 ierr = PetscMalloc1(maxnz,&cols);CHKERRQ(ierr); 2830 2831 /* read in my part of the matrix column indices */ 2832 nz = procsnz[0]; 2833 ierr = PetscMalloc1(nz,&mycols);CHKERRQ(ierr); 2834 ierr = PetscBinaryRead(fd,mycols,nz,PETSC_INT);CHKERRQ(ierr); 2835 2836 /* read in every one elses and ship off */ 2837 for (i=1; i<size; i++) { 2838 nz = procsnz[i]; 2839 ierr = PetscBinaryRead(fd,cols,nz,PETSC_INT);CHKERRQ(ierr); 2840 ierr = MPIULong_Send(cols,nz,MPIU_INT,i,tag,comm);CHKERRQ(ierr); 2841 } 2842 ierr = PetscFree(cols);CHKERRQ(ierr); 2843 } else { 2844 /* determine buffer space needed for message */ 2845 nz = 0; 2846 for (i=0; i<m; i++) { 2847 nz += ourlens[i]; 2848 } 2849 ierr = PetscMalloc1(nz,&mycols);CHKERRQ(ierr); 2850 2851 /* receive message of column indices*/ 2852 ierr = MPIULong_Recv(mycols,nz,MPIU_INT,0,tag,comm);CHKERRQ(ierr); 2853 } 2854 2855 /* determine column ownership if matrix is not square */ 2856 if (N != M) { 2857 if (newMat->cmap->n < 0) n = N/size + ((N % size) > rank); 2858 else n = newMat->cmap->n; 2859 ierr = MPI_Scan(&n,&cend,1,MPIU_INT,MPI_SUM,comm);CHKERRQ(ierr); 2860 cstart = cend - n; 2861 } else { 2862 cstart = rstart; 2863 cend = rend; 2864 n = cend - cstart; 2865 } 2866 2867 /* loop over local rows, determining number of off diagonal entries */ 2868 ierr = PetscMemzero(offlens,m*sizeof(PetscInt));CHKERRQ(ierr); 2869 jj = 0; 2870 for (i=0; i<m; i++) { 2871 for (j=0; j<ourlens[i]; j++) { 2872 if (mycols[jj] < cstart || mycols[jj] >= cend) offlens[i]++; 2873 jj++; 2874 } 2875 } 2876 2877 for (i=0; i<m; i++) { 2878 ourlens[i] -= offlens[i]; 2879 } 2880 ierr = MatSetSizes(newMat,m,n,M,N);CHKERRQ(ierr); 2881 2882 if (bs > 1) {ierr = MatSetBlockSize(newMat,bs);CHKERRQ(ierr);} 2883 2884 ierr = MatMPIAIJSetPreallocation(newMat,0,ourlens,0,offlens);CHKERRQ(ierr); 2885 2886 for (i=0; i<m; i++) { 2887 ourlens[i] += offlens[i]; 2888 } 2889 2890 if (!rank) { 2891 ierr = PetscMalloc1(maxnz+1,&vals);CHKERRQ(ierr); 2892 2893 /* read in my part of the matrix numerical values */ 2894 nz = procsnz[0]; 2895 ierr = PetscBinaryRead(fd,vals,nz,PETSC_SCALAR);CHKERRQ(ierr); 2896 2897 /* insert into matrix */ 2898 jj = rstart; 2899 smycols = mycols; 2900 svals = vals; 2901 for (i=0; i<m; i++) { 2902 ierr = MatSetValues_MPIAIJ(newMat,1,&jj,ourlens[i],smycols,svals,INSERT_VALUES);CHKERRQ(ierr); 2903 smycols += ourlens[i]; 2904 svals += ourlens[i]; 2905 jj++; 2906 } 2907 2908 /* read in other processors and ship out */ 2909 for (i=1; i<size; i++) { 2910 nz = procsnz[i]; 2911 ierr = PetscBinaryRead(fd,vals,nz,PETSC_SCALAR);CHKERRQ(ierr); 2912 ierr = MPIULong_Send(vals,nz,MPIU_SCALAR,i,((PetscObject)newMat)->tag,comm);CHKERRQ(ierr); 2913 } 2914 ierr = PetscFree(procsnz);CHKERRQ(ierr); 2915 } else { 2916 /* receive numeric values */ 2917 ierr = PetscMalloc1(nz+1,&vals);CHKERRQ(ierr); 2918 2919 /* receive message of values*/ 2920 ierr = MPIULong_Recv(vals,nz,MPIU_SCALAR,0,((PetscObject)newMat)->tag,comm);CHKERRQ(ierr); 2921 2922 /* insert into matrix */ 2923 jj = rstart; 2924 smycols = mycols; 2925 svals = vals; 2926 for (i=0; i<m; i++) { 2927 ierr = MatSetValues_MPIAIJ(newMat,1,&jj,ourlens[i],smycols,svals,INSERT_VALUES);CHKERRQ(ierr); 2928 smycols += ourlens[i]; 2929 svals += ourlens[i]; 2930 jj++; 2931 } 2932 } 2933 ierr = PetscFree2(ourlens,offlens);CHKERRQ(ierr); 2934 ierr = PetscFree(vals);CHKERRQ(ierr); 2935 ierr = PetscFree(mycols);CHKERRQ(ierr); 2936 ierr = PetscFree(rowners);CHKERRQ(ierr); 2937 ierr = MatAssemblyBegin(newMat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2938 ierr = MatAssemblyEnd(newMat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2939 PetscFunctionReturn(0); 2940 } 2941 2942 /* TODO: Not scalable because of ISAllGather() unless getting all columns. */ 2943 PetscErrorCode MatCreateSubMatrix_MPIAIJ(Mat mat,IS isrow,IS iscol,MatReuse call,Mat *newmat) 2944 { 2945 PetscErrorCode ierr; 2946 IS iscol_local; 2947 PetscInt csize; 2948 2949 PetscFunctionBegin; 2950 ierr = ISGetLocalSize(iscol,&csize);CHKERRQ(ierr); 2951 if (call == MAT_REUSE_MATRIX) { 2952 ierr = PetscObjectQuery((PetscObject)*newmat,"ISAllGather",(PetscObject*)&iscol_local);CHKERRQ(ierr); 2953 if (!iscol_local) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Submatrix passed in was not used before, cannot reuse"); 2954 } else { 2955 /* check if we are grabbing all columns*/ 2956 PetscBool isstride; 2957 PetscMPIInt lisstride = 0,gisstride; 2958 ierr = PetscObjectTypeCompare((PetscObject)iscol,ISSTRIDE,&isstride);CHKERRQ(ierr); 2959 if (isstride) { 2960 PetscInt start,len,mstart,mlen; 2961 ierr = ISStrideGetInfo(iscol,&start,NULL);CHKERRQ(ierr); 2962 ierr = ISGetLocalSize(iscol,&len);CHKERRQ(ierr); 2963 ierr = MatGetOwnershipRangeColumn(mat,&mstart,&mlen);CHKERRQ(ierr); 2964 if (mstart == start && mlen-mstart == len) lisstride = 1; 2965 } 2966 ierr = MPIU_Allreduce(&lisstride,&gisstride,1,MPI_INT,MPI_MIN,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 2967 if (gisstride) { 2968 PetscInt N; 2969 ierr = MatGetSize(mat,NULL,&N);CHKERRQ(ierr); 2970 ierr = ISCreateStride(PetscObjectComm((PetscObject)mat),N,0,1,&iscol_local);CHKERRQ(ierr); 2971 ierr = ISSetIdentity(iscol_local);CHKERRQ(ierr); 2972 ierr = PetscInfo(mat,"Optimizing for obtaining all columns of the matrix; skipping ISAllGather()\n");CHKERRQ(ierr); 2973 } else { 2974 PetscInt cbs; 2975 ierr = ISGetBlockSize(iscol,&cbs);CHKERRQ(ierr); 2976 ierr = ISAllGather(iscol,&iscol_local);CHKERRQ(ierr); 2977 ierr = ISSetBlockSize(iscol_local,cbs);CHKERRQ(ierr); 2978 } 2979 } 2980 #if 0 2981 //==================== new ==================== 2982 PetscMPIInt rank; 2983 MPI_Comm comm; 2984 Mat_MPIAIJ *a = (Mat_MPIAIJ*)mat->data; 2985 Mat A = a->A,B = a->B; 2986 PetscInt i,j,nlocal,*garray = a->garray; 2987 PetscInt count = 0,An=A->cmap->N,Bn=B->cmap->N,cstart=mat->cmap->rstart,cend=mat->cmap->rend; 2988 const PetscInt *is_idx; 2989 2990 ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 2991 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 2992 2993 if (rank==-1) { 2994 ierr = ISView(iscol_local,0);CHKERRQ(ierr); 2995 } 2996 2997 ierr = ISGetLocalSize(iscol_local,&nlocal);CHKERRQ(ierr); 2998 //printf("[%d] cstart/end: %d - %d; nlocal %d\n",rank,cstart,cend,nlocal); 2999 ierr = ISGetIndices(iscol_local,&is_idx);CHKERRQ(ierr); 3000 3001 /* (1) create a reduced scalable iscol_loc */ 3002 IS iscol_loc; 3003 PetscInt idx[nlocal],cmap[nlocal]; 3004 count = 0; 3005 3006 /* A part */ 3007 j = cstart; 3008 for (i=0; i<nlocal; i++) { 3009 if (j >= cend) break; 3010 if (is_idx[i] == j) { 3011 idx[count] = j; 3012 cmap[count] = i; /* column index in submat */ 3013 count++; j++; 3014 } else if (is_idx[i] > j) { 3015 while (is_idx[i] > j && j < cend-1) j++; 3016 if (is_idx[i] == j) { 3017 idx[count] = j; 3018 cmap[count] = i; /* column index in submat */ 3019 count++; j++; 3020 } 3021 } 3022 } 3023 3024 /* B part */ 3025 j = 0; 3026 for (i=0; i<nlocal; i++) { 3027 if (j >= Bn) break; 3028 if (is_idx[i] == garray[j]) { 3029 idx[count] = garray[j]; 3030 cmap[count] = i; /* column index in submat */ 3031 count++; j++; 3032 } else if (is_idx[i] > garray[j]) { 3033 while (is_idx[i] > garray[j] && j < Bn-1) j++; 3034 if (is_idx[i] == garray[j]) { 3035 idx[count] = garray[j]; 3036 cmap[count] = i; /* column index in submat */ 3037 count++; j++; 3038 } 3039 } 3040 } 3041 //printf("[%d] nlocal %d, N %d, Bn %d + An %d = %d, count %d\n",rank,nlocal,mat->cmap->N,Bn,An,Bn+An,count); 3042 3043 ierr = ISCreateGeneral(PetscObjectComm((PetscObject)iscol_local),count,idx,PETSC_COPY_VALUES,&iscol_loc);CHKERRQ(ierr); 3044 //ierr = ISSort(iscol_loc);CHKERRQ(ierr); 3045 if (rank==-1) { 3046 ierr = ISView(iscol_loc,0);CHKERRQ(ierr); 3047 for (i=0; i<count; i++) printf("cmap[%d] = %d\n",i,cmap[i]); 3048 } 3049 3050 ierr = ISRestoreIndices(iscol_local,&is_idx);CHKERRQ(ierr); 3051 ierr = ISDestroy(&iscol_loc);CHKERRQ(ierr); 3052 //========================================================== 3053 #endif 3054 ierr = MatCreateSubMatrix_MPIAIJ_Private(mat,isrow,iscol_local,csize,call,newmat);CHKERRQ(ierr); 3055 if (call == MAT_INITIAL_MATRIX) { 3056 ierr = PetscObjectCompose((PetscObject)*newmat,"ISAllGather",(PetscObject)iscol_local);CHKERRQ(ierr); 3057 ierr = ISDestroy(&iscol_local);CHKERRQ(ierr); 3058 } 3059 PetscFunctionReturn(0); 3060 } 3061 3062 extern PetscErrorCode MatCreateSubMatrices_MPIAIJ_Local(Mat,PetscInt,const IS[],const IS[],MatReuse,Mat*); 3063 extern PetscErrorCode MatCreateSubMatrices_MPIAIJ_SingleIS_Local(Mat,PetscInt,const IS[],const IS[],MatReuse,PetscBool,Mat*); 3064 3065 PetscErrorCode MatCreateSubMatrix_MPIAIJ_Private_SameDist(Mat mat,IS isrow,IS iscol,PetscInt csize,MatReuse call,Mat *newmat) 3066 { 3067 PetscErrorCode ierr; 3068 PetscMPIInt rank,size; 3069 PetscInt i,m,n,rstart,row,rend,nz,*cwork,j,bs,cbs; 3070 PetscInt *ii,*jj,nlocal,*dlens,*olens,dlen,olen,jend,mglobal; 3071 Mat M,Mreuse; 3072 MatScalar *vwork,*aa; 3073 MPI_Comm comm; 3074 Mat_SeqAIJ *aij; 3075 PetscBool colflag,allcolumns=PETSC_FALSE; 3076 3077 PetscFunctionBegin; 3078 printf("MatCreateSubMatrix_MPIAIJ_Private_SameDist...\n"); 3079 3080 ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 3081 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 3082 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 3083 3084 //ierr = ISView(isrow,PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr); 3085 3086 /* Check for special case: each processor gets entire matrix columns */ 3087 ierr = ISIdentity(iscol,&colflag);CHKERRQ(ierr); 3088 ierr = ISGetLocalSize(iscol,&nlocal);CHKERRQ(ierr); 3089 if (colflag && nlocal == mat->cmap->N) allcolumns = PETSC_TRUE; 3090 3091 //==================== new ==================== 3092 Mat_MPIAIJ *a = (Mat_MPIAIJ*)mat->data; 3093 Mat A = a->A,B = a->B; 3094 PetscInt *garray = a->garray; 3095 PetscInt count = 0,An=A->cmap->N,Bn=B->cmap->N,cstart=mat->cmap->rstart,cend=mat->cmap->rend; 3096 const PetscInt *is_idx; 3097 IS iscol_local = iscol; 3098 PetscBool rowflag,newAlg = PETSC_FALSE; 3099 3100 #if 0 3101 ierr = ISSorted(iscol,&colflag);CHKERRQ(ierr); 3102 if (!colflag) printf("[%d] iscol !sorted\n",rank); 3103 3104 ierr = ISSorted(isrow,&rowflag);CHKERRQ(ierr); 3105 if (!rowflag) printf("[%d] isrow !sorted\n",rank); 3106 3107 if (colflag && rowflag) { 3108 ierr = MatGetOwnershipRange(mat,&rstart,&rend);CHKERRQ(ierr); 3109 ierr = MatGetOwnershipRangeColumn(mat,&cstart,&cend);CHKERRQ(ierr); 3110 3111 ierr = ISGetIndices(iscol_local,&is_idx);CHKERRQ(ierr); 3112 if (is_idx[0] >= cstart && is_idx[nlocal]<cend) newAlg=PETSC_TRUE; 3113 ierr = ISRestoreIndices(iscol_local,&is_idx);CHKERRQ(ierr); 3114 3115 ierr = ISGetIndices(isrow,&is_idx);CHKERRQ(ierr); 3116 if (is_idx[0] < rstart || is_idx[nlocal]>=rend) newAlg=PETSC_FALSE; 3117 ierr = ISRestoreIndices(isrow,&is_idx);CHKERRQ(ierr); 3118 } 3119 3120 printf("[%d] newAlg %d\n",rank,newAlg); 3121 #endif 3122 3123 if (rank==-1) { 3124 ierr = ISView(iscol_local,0);CHKERRQ(ierr); 3125 } 3126 3127 //printf("[%d] cstart/end: %d - %d; nlocal %d\n",rank,cstart,cend,nlocal); 3128 ierr = ISGetIndices(iscol_local,&is_idx);CHKERRQ(ierr); 3129 3130 /* (1) create a reduced scalable iscol_loc */ 3131 IS iscol_loc; 3132 PetscInt idx[nlocal],cmap[nlocal]; 3133 count = 0; 3134 3135 /* A part */ 3136 j = cstart; 3137 for (i=0; i<nlocal; i++) { 3138 if (j >= cend) break; 3139 if (is_idx[i] == j) { 3140 idx[count] = j; 3141 cmap[count] = i; /* column index in submat */ 3142 count++; j++; 3143 } else if (is_idx[i] > j) { 3144 while (is_idx[i] > j && j < cend-1) j++; 3145 if (is_idx[i] == j) { 3146 idx[count] = j; 3147 cmap[count] = i; /* column index in submat */ 3148 count++; j++; 3149 } 3150 } 3151 } 3152 3153 /* B part */ 3154 j = 0; 3155 for (i=0; i<nlocal; i++) { 3156 if (j >= Bn) break; 3157 if (is_idx[i] == garray[j]) { 3158 idx[count] = garray[j]; 3159 cmap[count] = i; /* column index in submat */ 3160 count++; j++; 3161 } else if (is_idx[i] > garray[j]) { 3162 while (is_idx[i] > garray[j] && j < Bn-1) j++; 3163 if (is_idx[i] == garray[j]) { 3164 idx[count] = garray[j]; 3165 cmap[count] = i; /* column index in submat */ 3166 count++; j++; 3167 } 3168 } 3169 } 3170 ierr = ISRestoreIndices(iscol_local,&is_idx);CHKERRQ(ierr); 3171 //printf("[%d] nlocal %d, N %d, Bn %d + An %d = %d, count %d\n",rank,nlocal,mat->cmap->N,Bn,An,Bn+An,count); 3172 3173 ierr = ISCreateGeneral(PetscObjectComm((PetscObject)iscol_local),count,idx,PETSC_COPY_VALUES,&iscol_loc);CHKERRQ(ierr); 3174 ierr = ISSort(iscol_loc);CHKERRQ(ierr); 3175 ierr = PetscSortInt(count,cmap);CHKERRQ(ierr); 3176 if (rank==-1) { 3177 ierr = ISView(iscol_loc,0);CHKERRQ(ierr); 3178 for (i=0; i<count; i++) printf("cmap[%d] = %d\n",i,cmap[i]); 3179 } 3180 3181 Mat Msub; 3182 ierr = MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat,1,&isrow,&iscol_loc,MAT_INITIAL_MATRIX,allcolumns,&Msub);CHKERRQ(ierr); 3183 if (rank==-1) { 3184 printf("[%d] Msub: %d %d \n",rank,Msub->rmap->N,Msub->cmap->N); 3185 ierr = MatView(Msub,0);CHKERRQ(ierr); 3186 } 3187 ierr = ISDestroy(&iscol_loc);CHKERRQ(ierr); 3188 //========================================================== 3189 3190 if (call == MAT_REUSE_MATRIX) { 3191 ierr = PetscObjectQuery((PetscObject)*newmat,"SubMatrix",(PetscObject*)&Mreuse);CHKERRQ(ierr); 3192 if (!Mreuse) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Submatrix passed in was not used before, cannot reuse"); 3193 ierr = MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat,1,&isrow,&iscol,MAT_REUSE_MATRIX,allcolumns,&Mreuse);CHKERRQ(ierr); 3194 } else { 3195 ierr = MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat,1,&isrow,&iscol,MAT_INITIAL_MATRIX,allcolumns,&Mreuse);CHKERRQ(ierr); 3196 } 3197 3198 /* 3199 m - number of local rows 3200 n - number of columns (same on all processors) 3201 rstart - first row in new global matrix generated 3202 */ 3203 ierr = MatGetSize(Mreuse,&m,&n);CHKERRQ(ierr); 3204 ierr = MatGetBlockSizes(Mreuse,&bs,&cbs);CHKERRQ(ierr); 3205 if (call == MAT_INITIAL_MATRIX) { 3206 aij = (Mat_SeqAIJ*)(Mreuse)->data; 3207 ii = aij->i; 3208 jj = aij->j; 3209 3210 /* 3211 Determine the number of non-zeros in the diagonal and off-diagonal 3212 portions of the matrix in order to do correct preallocation 3213 */ 3214 3215 /* first get start and end of "diagonal" columns */ 3216 if (csize == PETSC_DECIDE) { 3217 ierr = ISGetSize(isrow,&mglobal);CHKERRQ(ierr); 3218 if (mglobal == n) { /* square matrix */ 3219 nlocal = m; 3220 } else { 3221 nlocal = n/size + ((n % size) > rank); 3222 } 3223 } else { 3224 nlocal = csize; 3225 } 3226 ierr = MPI_Scan(&nlocal,&rend,1,MPIU_INT,MPI_SUM,comm);CHKERRQ(ierr); 3227 rstart = rend - nlocal; 3228 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); 3229 3230 /* next, compute all the lengths */ 3231 ierr = PetscMalloc1(2*m+1,&dlens);CHKERRQ(ierr); 3232 olens = dlens + m; 3233 for (i=0; i<m; i++) { 3234 jend = ii[i+1] - ii[i]; 3235 olen = 0; 3236 dlen = 0; 3237 for (j=0; j<jend; j++) { 3238 if (*jj < rstart || *jj >= rend) olen++; 3239 else dlen++; 3240 jj++; 3241 } 3242 olens[i] = olen; 3243 dlens[i] = dlen; 3244 } 3245 ierr = MatCreate(comm,&M);CHKERRQ(ierr); 3246 ierr = MatSetSizes(M,m,nlocal,PETSC_DECIDE,n);CHKERRQ(ierr); 3247 ierr = MatSetBlockSizes(M,bs,cbs);CHKERRQ(ierr); 3248 ierr = MatSetType(M,((PetscObject)mat)->type_name);CHKERRQ(ierr); 3249 ierr = MatMPIAIJSetPreallocation(M,0,dlens,0,olens);CHKERRQ(ierr); 3250 ierr = PetscFree(dlens);CHKERRQ(ierr); 3251 } else { 3252 PetscInt ml,nl; 3253 3254 M = *newmat; 3255 ierr = MatGetLocalSize(M,&ml,&nl);CHKERRQ(ierr); 3256 if (ml != m) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Previous matrix must be same size/layout as request"); 3257 ierr = MatZeroEntries(M);CHKERRQ(ierr); 3258 /* 3259 The next two lines are needed so we may call MatSetValues_MPIAIJ() below directly, 3260 rather than the slower MatSetValues(). 3261 */ 3262 M->was_assembled = PETSC_TRUE; 3263 M->assembled = PETSC_FALSE; 3264 } 3265 3266 //========================== 3267 Mat_SeqAIJ *aijsub = (Mat_SeqAIJ*)(Msub)->data; 3268 PetscInt *iisub = aijsub->i,*jjsub = aijsub->j,nzsub,colsub; 3269 PetscScalar *aasub = aijsub->a,vsub; 3270 //=========================== 3271 3272 ierr = MatGetOwnershipRange(M,&rstart,&rend);CHKERRQ(ierr); 3273 aij = (Mat_SeqAIJ*)(Mreuse)->data; 3274 ii = aij->i; 3275 jj = aij->j; 3276 aa = aij->a; 3277 for (i=0; i<m; i++) { 3278 row = rstart + i; 3279 nz = ii[i+1] - ii[i]; 3280 cwork = jj; jj += nz; 3281 vwork = aa; aa += nz; 3282 3283 nzsub = iisub[i+1] - iisub[i]; 3284 if (nzsub != nz) printf(" nzsub != nz\n"); 3285 3286 for (j=0; j<nz; j++) { 3287 colsub = cmap[jjsub[iisub[i]+j]]; 3288 vsub = aasub[iisub[i]+j]; 3289 if (colsub != cwork[j]) printf("%d != %d\n",colsub,cwork[j]); 3290 if (vwork[j] != vsub) printf(" %g != %g\n",vwork[j],vsub); 3291 ierr = MatSetValues_MPIAIJ(M,1,&row,1,&colsub,&vsub,INSERT_VALUES);CHKERRQ(ierr); 3292 } 3293 3294 //ierr = MatSetValues_MPIAIJ(M,1,&row,nz,cwork,vwork,INSERT_VALUES);CHKERRQ(ierr); 3295 } 3296 3297 ierr = MatAssemblyBegin(M,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3298 ierr = MatAssemblyEnd(M,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3299 *newmat = M; 3300 //ierr = MatView(M,PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr); 3301 3302 /* save submatrix used in processor for next request */ 3303 if (call == MAT_INITIAL_MATRIX) { 3304 ierr = PetscObjectCompose((PetscObject)M,"SubMatrix",(PetscObject)Mreuse);CHKERRQ(ierr); 3305 ierr = MatDestroy(&Mreuse);CHKERRQ(ierr); 3306 } 3307 3308 ierr = MatDestroy(&Msub);CHKERRQ(ierr); 3309 PetscFunctionReturn(0); 3310 } 3311 3312 /* 3313 Not great since it makes two copies of the submatrix, first an SeqAIJ 3314 in local and then by concatenating the local matrices the end result. 3315 Writing it directly would be much like MatCreateSubMatrices_MPIAIJ() 3316 3317 Note: This requires a sequential iscol with all indices. 3318 */ 3319 PetscErrorCode MatCreateSubMatrix_MPIAIJ_Private(Mat mat,IS isrow,IS iscol,PetscInt csize,MatReuse call,Mat *newmat) 3320 { 3321 PetscErrorCode ierr; 3322 PetscMPIInt rank,size; 3323 PetscInt i,m,n,rstart,row,rend,nz,*cwork,j,bs,cbs; 3324 PetscInt *ii,*jj,nlocal,*dlens,*olens,dlen,olen,jend,mglobal; 3325 Mat M,Mreuse; 3326 MatScalar *vwork,*aa; 3327 MPI_Comm comm; 3328 Mat_SeqAIJ *aij; 3329 PetscBool sameDist=PETSC_FALSE,tsameDist; 3330 3331 PetscFunctionBegin; 3332 ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 3333 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 3334 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 3335 3336 /* If isrow has same processor distribution as mat, then use a scalable routine */ 3337 ierr = ISGetLocalSize(isrow,&n);CHKERRQ(ierr); 3338 if (!n) { 3339 sameDist=PETSC_TRUE; 3340 } else { 3341 ierr = ISGetMinMax(isrow,&i,&j);CHKERRQ(ierr); 3342 ierr = MatGetOwnershipRange(mat,&rstart,&rend);CHKERRQ(ierr); 3343 if (i >= rstart && j < rend) sameDist=PETSC_TRUE; 3344 } 3345 ierr = MPIU_Allreduce(&sameDist,&tsameDist,1,MPIU_BOOL,MPI_LAND,comm);CHKERRQ(ierr); 3346 if (tsameDist) { 3347 if (!rank) printf(" will call _Private_SameDist() ...\n"); 3348 //ierr = MatCreateSubMatrix_MPIAIJ_Private_SameDist(mat,isrow,iscol,csize,call,newmat);CHKERRQ(ierr); 3349 //PetscFunctionReturn(0); 3350 } 3351 3352 if (call == MAT_REUSE_MATRIX) { 3353 ierr = PetscObjectQuery((PetscObject)*newmat,"SubMatrix",(PetscObject*)&Mreuse);CHKERRQ(ierr); 3354 if (!Mreuse) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Submatrix passed in was not used before, cannot reuse"); 3355 ierr = MatCreateSubMatrices_MPIAIJ_Local(mat,1,&isrow,&iscol,MAT_REUSE_MATRIX,&Mreuse);CHKERRQ(ierr); 3356 } else { 3357 ierr = MatCreateSubMatrices_MPIAIJ_Local(mat,1,&isrow,&iscol,MAT_INITIAL_MATRIX,&Mreuse);CHKERRQ(ierr); 3358 } 3359 3360 /* 3361 m - number of local rows 3362 n - number of columns (same on all processors) 3363 rstart - first row in new global matrix generated 3364 */ 3365 ierr = MatGetSize(Mreuse,&m,&n);CHKERRQ(ierr); 3366 ierr = MatGetBlockSizes(Mreuse,&bs,&cbs);CHKERRQ(ierr); 3367 if (call == MAT_INITIAL_MATRIX) { 3368 aij = (Mat_SeqAIJ*)(Mreuse)->data; 3369 ii = aij->i; 3370 jj = aij->j; 3371 3372 /* 3373 Determine the number of non-zeros in the diagonal and off-diagonal 3374 portions of the matrix in order to do correct preallocation 3375 */ 3376 3377 /* first get start and end of "diagonal" columns */ 3378 if (csize == PETSC_DECIDE) { 3379 ierr = ISGetSize(isrow,&mglobal);CHKERRQ(ierr); 3380 if (mglobal == n) { /* square matrix */ 3381 nlocal = m; 3382 } else { 3383 nlocal = n/size + ((n % size) > rank); 3384 } 3385 } else { 3386 nlocal = csize; 3387 } 3388 ierr = MPI_Scan(&nlocal,&rend,1,MPIU_INT,MPI_SUM,comm);CHKERRQ(ierr); 3389 rstart = rend - nlocal; 3390 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); 3391 3392 /* next, compute all the lengths */ 3393 ierr = PetscMalloc1(2*m+1,&dlens);CHKERRQ(ierr); 3394 olens = dlens + m; 3395 for (i=0; i<m; i++) { 3396 jend = ii[i+1] - ii[i]; 3397 olen = 0; 3398 dlen = 0; 3399 for (j=0; j<jend; j++) { 3400 if (*jj < rstart || *jj >= rend) olen++; 3401 else dlen++; 3402 jj++; 3403 } 3404 olens[i] = olen; 3405 dlens[i] = dlen; 3406 } 3407 ierr = MatCreate(comm,&M);CHKERRQ(ierr); 3408 ierr = MatSetSizes(M,m,nlocal,PETSC_DECIDE,n);CHKERRQ(ierr); 3409 ierr = MatSetBlockSizes(M,bs,cbs);CHKERRQ(ierr); 3410 ierr = MatSetType(M,((PetscObject)mat)->type_name);CHKERRQ(ierr); 3411 ierr = MatMPIAIJSetPreallocation(M,0,dlens,0,olens);CHKERRQ(ierr); 3412 ierr = PetscFree(dlens);CHKERRQ(ierr); 3413 } else { 3414 PetscInt ml,nl; 3415 3416 M = *newmat; 3417 ierr = MatGetLocalSize(M,&ml,&nl);CHKERRQ(ierr); 3418 if (ml != m) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Previous matrix must be same size/layout as request"); 3419 ierr = MatZeroEntries(M);CHKERRQ(ierr); 3420 /* 3421 The next two lines are needed so we may call MatSetValues_MPIAIJ() below directly, 3422 rather than the slower MatSetValues(). 3423 */ 3424 M->was_assembled = PETSC_TRUE; 3425 M->assembled = PETSC_FALSE; 3426 } 3427 ierr = MatGetOwnershipRange(M,&rstart,&rend);CHKERRQ(ierr); 3428 aij = (Mat_SeqAIJ*)(Mreuse)->data; 3429 ii = aij->i; 3430 jj = aij->j; 3431 aa = aij->a; 3432 for (i=0; i<m; i++) { 3433 row = rstart + i; 3434 nz = ii[i+1] - ii[i]; 3435 cwork = jj; jj += nz; 3436 vwork = aa; aa += nz; 3437 ierr = MatSetValues_MPIAIJ(M,1,&row,nz,cwork,vwork,INSERT_VALUES);CHKERRQ(ierr); 3438 } 3439 3440 ierr = MatAssemblyBegin(M,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3441 ierr = MatAssemblyEnd(M,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3442 *newmat = M; 3443 3444 /* save submatrix used in processor for next request */ 3445 if (call == MAT_INITIAL_MATRIX) { 3446 ierr = PetscObjectCompose((PetscObject)M,"SubMatrix",(PetscObject)Mreuse);CHKERRQ(ierr); 3447 ierr = MatDestroy(&Mreuse);CHKERRQ(ierr); 3448 } 3449 PetscFunctionReturn(0); 3450 } 3451 3452 PetscErrorCode MatMPIAIJSetPreallocationCSR_MPIAIJ(Mat B,const PetscInt Ii[],const PetscInt J[],const PetscScalar v[]) 3453 { 3454 PetscInt m,cstart, cend,j,nnz,i,d; 3455 PetscInt *d_nnz,*o_nnz,nnz_max = 0,rstart,ii; 3456 const PetscInt *JJ; 3457 PetscScalar *values; 3458 PetscErrorCode ierr; 3459 PetscBool nooffprocentries; 3460 3461 PetscFunctionBegin; 3462 if (Ii[0]) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Ii[0] must be 0 it is %D",Ii[0]); 3463 3464 ierr = PetscLayoutSetUp(B->rmap);CHKERRQ(ierr); 3465 ierr = PetscLayoutSetUp(B->cmap);CHKERRQ(ierr); 3466 m = B->rmap->n; 3467 cstart = B->cmap->rstart; 3468 cend = B->cmap->rend; 3469 rstart = B->rmap->rstart; 3470 3471 ierr = PetscMalloc2(m,&d_nnz,m,&o_nnz);CHKERRQ(ierr); 3472 3473 #if defined(PETSC_USE_DEBUGGING) 3474 for (i=0; i<m; i++) { 3475 nnz = Ii[i+1]- Ii[i]; 3476 JJ = J + Ii[i]; 3477 if (nnz < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Local row %D has a negative %D number of columns",i,nnz); 3478 if (nnz && (JJ[0] < 0)) SETERRRQ1(PETSC_ERR_ARG_WRONGSTATE,"Row %D starts with negative column index",i,j); 3479 if (nnz && (JJ[nnz-1] >= B->cmap->N) SETERRRQ3(PETSC_ERR_ARG_WRONGSTATE,"Row %D ends with too large a column index %D (max allowed %D)",i,JJ[nnz-1],B->cmap->N); 3480 } 3481 #endif 3482 3483 for (i=0; i<m; i++) { 3484 nnz = Ii[i+1]- Ii[i]; 3485 JJ = J + Ii[i]; 3486 nnz_max = PetscMax(nnz_max,nnz); 3487 d = 0; 3488 for (j=0; j<nnz; j++) { 3489 if (cstart <= JJ[j] && JJ[j] < cend) d++; 3490 } 3491 d_nnz[i] = d; 3492 o_nnz[i] = nnz - d; 3493 } 3494 ierr = MatMPIAIJSetPreallocation(B,0,d_nnz,0,o_nnz);CHKERRQ(ierr); 3495 ierr = PetscFree2(d_nnz,o_nnz);CHKERRQ(ierr); 3496 3497 if (v) values = (PetscScalar*)v; 3498 else { 3499 ierr = PetscCalloc1(nnz_max+1,&values);CHKERRQ(ierr); 3500 } 3501 3502 for (i=0; i<m; i++) { 3503 ii = i + rstart; 3504 nnz = Ii[i+1]- Ii[i]; 3505 ierr = MatSetValues_MPIAIJ(B,1,&ii,nnz,J+Ii[i],values+(v ? Ii[i] : 0),INSERT_VALUES);CHKERRQ(ierr); 3506 } 3507 nooffprocentries = B->nooffprocentries; 3508 B->nooffprocentries = PETSC_TRUE; 3509 ierr = MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3510 ierr = MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3511 B->nooffprocentries = nooffprocentries; 3512 3513 if (!v) { 3514 ierr = PetscFree(values);CHKERRQ(ierr); 3515 } 3516 ierr = MatSetOption(B,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE);CHKERRQ(ierr); 3517 PetscFunctionReturn(0); 3518 } 3519 3520 /*@ 3521 MatMPIAIJSetPreallocationCSR - Allocates memory for a sparse parallel matrix in AIJ format 3522 (the default parallel PETSc format). 3523 3524 Collective on MPI_Comm 3525 3526 Input Parameters: 3527 + B - the matrix 3528 . i - the indices into j for the start of each local row (starts with zero) 3529 . j - the column indices for each local row (starts with zero) 3530 - v - optional values in the matrix 3531 3532 Level: developer 3533 3534 Notes: 3535 The i, j, and a arrays ARE copied by this routine into the internal format used by PETSc; 3536 thus you CANNOT change the matrix entries by changing the values of a[] after you have 3537 called this routine. Use MatCreateMPIAIJWithSplitArrays() to avoid needing to copy the arrays. 3538 3539 The i and j indices are 0 based, and i indices are indices corresponding to the local j array. 3540 3541 The format which is used for the sparse matrix input, is equivalent to a 3542 row-major ordering.. i.e for the following matrix, the input data expected is 3543 as shown 3544 3545 $ 1 0 0 3546 $ 2 0 3 P0 3547 $ ------- 3548 $ 4 5 6 P1 3549 $ 3550 $ Process0 [P0]: rows_owned=[0,1] 3551 $ i = {0,1,3} [size = nrow+1 = 2+1] 3552 $ j = {0,0,2} [size = 3] 3553 $ v = {1,2,3} [size = 3] 3554 $ 3555 $ Process1 [P1]: rows_owned=[2] 3556 $ i = {0,3} [size = nrow+1 = 1+1] 3557 $ j = {0,1,2} [size = 3] 3558 $ v = {4,5,6} [size = 3] 3559 3560 .keywords: matrix, aij, compressed row, sparse, parallel 3561 3562 .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatCreateAIJ(), MATMPIAIJ, 3563 MatCreateSeqAIJWithArrays(), MatCreateMPIAIJWithSplitArrays() 3564 @*/ 3565 PetscErrorCode MatMPIAIJSetPreallocationCSR(Mat B,const PetscInt i[],const PetscInt j[], const PetscScalar v[]) 3566 { 3567 PetscErrorCode ierr; 3568 3569 PetscFunctionBegin; 3570 ierr = PetscTryMethod(B,"MatMPIAIJSetPreallocationCSR_C",(Mat,const PetscInt[],const PetscInt[],const PetscScalar[]),(B,i,j,v));CHKERRQ(ierr); 3571 PetscFunctionReturn(0); 3572 } 3573 3574 /*@C 3575 MatMPIAIJSetPreallocation - Preallocates memory for a sparse parallel matrix in AIJ format 3576 (the default parallel PETSc format). For good matrix assembly performance 3577 the user should preallocate the matrix storage by setting the parameters 3578 d_nz (or d_nnz) and o_nz (or o_nnz). By setting these parameters accurately, 3579 performance can be increased by more than a factor of 50. 3580 3581 Collective on MPI_Comm 3582 3583 Input Parameters: 3584 + B - the matrix 3585 . d_nz - number of nonzeros per row in DIAGONAL portion of local submatrix 3586 (same value is used for all local rows) 3587 . d_nnz - array containing the number of nonzeros in the various rows of the 3588 DIAGONAL portion of the local submatrix (possibly different for each row) 3589 or NULL (PETSC_NULL_INTEGER in Fortran), if d_nz is used to specify the nonzero structure. 3590 The size of this array is equal to the number of local rows, i.e 'm'. 3591 For matrices that will be factored, you must leave room for (and set) 3592 the diagonal entry even if it is zero. 3593 . o_nz - number of nonzeros per row in the OFF-DIAGONAL portion of local 3594 submatrix (same value is used for all local rows). 3595 - o_nnz - array containing the number of nonzeros in the various rows of the 3596 OFF-DIAGONAL portion of the local submatrix (possibly different for 3597 each row) or NULL (PETSC_NULL_INTEGER in Fortran), if o_nz is used to specify the nonzero 3598 structure. The size of this array is equal to the number 3599 of local rows, i.e 'm'. 3600 3601 If the *_nnz parameter is given then the *_nz parameter is ignored 3602 3603 The AIJ format (also called the Yale sparse matrix format or 3604 compressed row storage (CSR)), is fully compatible with standard Fortran 77 3605 storage. The stored row and column indices begin with zero. 3606 See Users-Manual: ch_mat for details. 3607 3608 The parallel matrix is partitioned such that the first m0 rows belong to 3609 process 0, the next m1 rows belong to process 1, the next m2 rows belong 3610 to process 2 etc.. where m0,m1,m2... are the input parameter 'm'. 3611 3612 The DIAGONAL portion of the local submatrix of a processor can be defined 3613 as the submatrix which is obtained by extraction the part corresponding to 3614 the rows r1-r2 and columns c1-c2 of the global matrix, where r1 is the 3615 first row that belongs to the processor, r2 is the last row belonging to 3616 the this processor, and c1-c2 is range of indices of the local part of a 3617 vector suitable for applying the matrix to. This is an mxn matrix. In the 3618 common case of a square matrix, the row and column ranges are the same and 3619 the DIAGONAL part is also square. The remaining portion of the local 3620 submatrix (mxN) constitute the OFF-DIAGONAL portion. 3621 3622 If o_nnz, d_nnz are specified, then o_nz, and d_nz are ignored. 3623 3624 You can call MatGetInfo() to get information on how effective the preallocation was; 3625 for example the fields mallocs,nz_allocated,nz_used,nz_unneeded; 3626 You can also run with the option -info and look for messages with the string 3627 malloc in them to see if additional memory allocation was needed. 3628 3629 Example usage: 3630 3631 Consider the following 8x8 matrix with 34 non-zero values, that is 3632 assembled across 3 processors. Lets assume that proc0 owns 3 rows, 3633 proc1 owns 3 rows, proc2 owns 2 rows. This division can be shown 3634 as follows: 3635 3636 .vb 3637 1 2 0 | 0 3 0 | 0 4 3638 Proc0 0 5 6 | 7 0 0 | 8 0 3639 9 0 10 | 11 0 0 | 12 0 3640 ------------------------------------- 3641 13 0 14 | 15 16 17 | 0 0 3642 Proc1 0 18 0 | 19 20 21 | 0 0 3643 0 0 0 | 22 23 0 | 24 0 3644 ------------------------------------- 3645 Proc2 25 26 27 | 0 0 28 | 29 0 3646 30 0 0 | 31 32 33 | 0 34 3647 .ve 3648 3649 This can be represented as a collection of submatrices as: 3650 3651 .vb 3652 A B C 3653 D E F 3654 G H I 3655 .ve 3656 3657 Where the submatrices A,B,C are owned by proc0, D,E,F are 3658 owned by proc1, G,H,I are owned by proc2. 3659 3660 The 'm' parameters for proc0,proc1,proc2 are 3,3,2 respectively. 3661 The 'n' parameters for proc0,proc1,proc2 are 3,3,2 respectively. 3662 The 'M','N' parameters are 8,8, and have the same values on all procs. 3663 3664 The DIAGONAL submatrices corresponding to proc0,proc1,proc2 are 3665 submatrices [A], [E], [I] respectively. The OFF-DIAGONAL submatrices 3666 corresponding to proc0,proc1,proc2 are [BC], [DF], [GH] respectively. 3667 Internally, each processor stores the DIAGONAL part, and the OFF-DIAGONAL 3668 part as SeqAIJ matrices. for eg: proc1 will store [E] as a SeqAIJ 3669 matrix, ans [DF] as another SeqAIJ matrix. 3670 3671 When d_nz, o_nz parameters are specified, d_nz storage elements are 3672 allocated for every row of the local diagonal submatrix, and o_nz 3673 storage locations are allocated for every row of the OFF-DIAGONAL submat. 3674 One way to choose d_nz and o_nz is to use the max nonzerors per local 3675 rows for each of the local DIAGONAL, and the OFF-DIAGONAL submatrices. 3676 In this case, the values of d_nz,o_nz are: 3677 .vb 3678 proc0 : dnz = 2, o_nz = 2 3679 proc1 : dnz = 3, o_nz = 2 3680 proc2 : dnz = 1, o_nz = 4 3681 .ve 3682 We are allocating m*(d_nz+o_nz) storage locations for every proc. This 3683 translates to 3*(2+2)=12 for proc0, 3*(3+2)=15 for proc1, 2*(1+4)=10 3684 for proc3. i.e we are using 12+15+10=37 storage locations to store 3685 34 values. 3686 3687 When d_nnz, o_nnz parameters are specified, the storage is specified 3688 for every row, coresponding to both DIAGONAL and OFF-DIAGONAL submatrices. 3689 In the above case the values for d_nnz,o_nnz are: 3690 .vb 3691 proc0: d_nnz = [2,2,2] and o_nnz = [2,2,2] 3692 proc1: d_nnz = [3,3,2] and o_nnz = [2,1,1] 3693 proc2: d_nnz = [1,1] and o_nnz = [4,4] 3694 .ve 3695 Here the space allocated is sum of all the above values i.e 34, and 3696 hence pre-allocation is perfect. 3697 3698 Level: intermediate 3699 3700 .keywords: matrix, aij, compressed row, sparse, parallel 3701 3702 .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatCreateAIJ(), MatMPIAIJSetPreallocationCSR(), 3703 MATMPIAIJ, MatGetInfo(), PetscSplitOwnership() 3704 @*/ 3705 PetscErrorCode MatMPIAIJSetPreallocation(Mat B,PetscInt d_nz,const PetscInt d_nnz[],PetscInt o_nz,const PetscInt o_nnz[]) 3706 { 3707 PetscErrorCode ierr; 3708 3709 PetscFunctionBegin; 3710 PetscValidHeaderSpecific(B,MAT_CLASSID,1); 3711 PetscValidType(B,1); 3712 ierr = PetscTryMethod(B,"MatMPIAIJSetPreallocation_C",(Mat,PetscInt,const PetscInt[],PetscInt,const PetscInt[]),(B,d_nz,d_nnz,o_nz,o_nnz));CHKERRQ(ierr); 3713 PetscFunctionReturn(0); 3714 } 3715 3716 /*@ 3717 MatCreateMPIAIJWithArrays - creates a MPI AIJ matrix using arrays that contain in standard 3718 CSR format the local rows. 3719 3720 Collective on MPI_Comm 3721 3722 Input Parameters: 3723 + comm - MPI communicator 3724 . m - number of local rows (Cannot be PETSC_DECIDE) 3725 . n - This value should be the same as the local size used in creating the 3726 x vector for the matrix-vector product y = Ax. (or PETSC_DECIDE to have 3727 calculated if N is given) For square matrices n is almost always m. 3728 . M - number of global rows (or PETSC_DETERMINE to have calculated if m is given) 3729 . N - number of global columns (or PETSC_DETERMINE to have calculated if n is given) 3730 . i - row indices 3731 . j - column indices 3732 - a - matrix values 3733 3734 Output Parameter: 3735 . mat - the matrix 3736 3737 Level: intermediate 3738 3739 Notes: 3740 The i, j, and a arrays ARE copied by this routine into the internal format used by PETSc; 3741 thus you CANNOT change the matrix entries by changing the values of a[] after you have 3742 called this routine. Use MatCreateMPIAIJWithSplitArrays() to avoid needing to copy the arrays. 3743 3744 The i and j indices are 0 based, and i indices are indices corresponding to the local j array. 3745 3746 The format which is used for the sparse matrix input, is equivalent to a 3747 row-major ordering.. i.e for the following matrix, the input data expected is 3748 as shown 3749 3750 $ 1 0 0 3751 $ 2 0 3 P0 3752 $ ------- 3753 $ 4 5 6 P1 3754 $ 3755 $ Process0 [P0]: rows_owned=[0,1] 3756 $ i = {0,1,3} [size = nrow+1 = 2+1] 3757 $ j = {0,0,2} [size = 3] 3758 $ v = {1,2,3} [size = 3] 3759 $ 3760 $ Process1 [P1]: rows_owned=[2] 3761 $ i = {0,3} [size = nrow+1 = 1+1] 3762 $ j = {0,1,2} [size = 3] 3763 $ v = {4,5,6} [size = 3] 3764 3765 .keywords: matrix, aij, compressed row, sparse, parallel 3766 3767 .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatMPIAIJSetPreallocationCSR(), 3768 MATMPIAIJ, MatCreateAIJ(), MatCreateMPIAIJWithSplitArrays() 3769 @*/ 3770 PetscErrorCode MatCreateMPIAIJWithArrays(MPI_Comm comm,PetscInt m,PetscInt n,PetscInt M,PetscInt N,const PetscInt i[],const PetscInt j[],const PetscScalar a[],Mat *mat) 3771 { 3772 PetscErrorCode ierr; 3773 3774 PetscFunctionBegin; 3775 if (i[0]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"i (row indices) must start with 0"); 3776 if (m < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"local number of rows (m) cannot be PETSC_DECIDE, or negative"); 3777 ierr = MatCreate(comm,mat);CHKERRQ(ierr); 3778 ierr = MatSetSizes(*mat,m,n,M,N);CHKERRQ(ierr); 3779 /* ierr = MatSetBlockSizes(M,bs,cbs);CHKERRQ(ierr); */ 3780 ierr = MatSetType(*mat,MATMPIAIJ);CHKERRQ(ierr); 3781 ierr = MatMPIAIJSetPreallocationCSR(*mat,i,j,a);CHKERRQ(ierr); 3782 PetscFunctionReturn(0); 3783 } 3784 3785 /*@C 3786 MatCreateAIJ - Creates a sparse parallel matrix in AIJ format 3787 (the default parallel PETSc format). For good matrix assembly performance 3788 the user should preallocate the matrix storage by setting the parameters 3789 d_nz (or d_nnz) and o_nz (or o_nnz). By setting these parameters accurately, 3790 performance can be increased by more than a factor of 50. 3791 3792 Collective on MPI_Comm 3793 3794 Input Parameters: 3795 + comm - MPI communicator 3796 . m - number of local rows (or PETSC_DECIDE to have calculated if M is given) 3797 This value should be the same as the local size used in creating the 3798 y vector for the matrix-vector product y = Ax. 3799 . n - This value should be the same as the local size used in creating the 3800 x vector for the matrix-vector product y = Ax. (or PETSC_DECIDE to have 3801 calculated if N is given) For square matrices n is almost always m. 3802 . M - number of global rows (or PETSC_DETERMINE to have calculated if m is given) 3803 . N - number of global columns (or PETSC_DETERMINE to have calculated if n is given) 3804 . d_nz - number of nonzeros per row in DIAGONAL portion of local submatrix 3805 (same value is used for all local rows) 3806 . d_nnz - array containing the number of nonzeros in the various rows of the 3807 DIAGONAL portion of the local submatrix (possibly different for each row) 3808 or NULL, if d_nz is used to specify the nonzero structure. 3809 The size of this array is equal to the number of local rows, i.e 'm'. 3810 . o_nz - number of nonzeros per row in the OFF-DIAGONAL portion of local 3811 submatrix (same value is used for all local rows). 3812 - o_nnz - array containing the number of nonzeros in the various rows of the 3813 OFF-DIAGONAL portion of the local submatrix (possibly different for 3814 each row) or NULL, if o_nz is used to specify the nonzero 3815 structure. The size of this array is equal to the number 3816 of local rows, i.e 'm'. 3817 3818 Output Parameter: 3819 . A - the matrix 3820 3821 It is recommended that one use the MatCreate(), MatSetType() and/or MatSetFromOptions(), 3822 MatXXXXSetPreallocation() paradgm instead of this routine directly. 3823 [MatXXXXSetPreallocation() is, for example, MatSeqAIJSetPreallocation] 3824 3825 Notes: 3826 If the *_nnz parameter is given then the *_nz parameter is ignored 3827 3828 m,n,M,N parameters specify the size of the matrix, and its partitioning across 3829 processors, while d_nz,d_nnz,o_nz,o_nnz parameters specify the approximate 3830 storage requirements for this matrix. 3831 3832 If PETSC_DECIDE or PETSC_DETERMINE is used for a particular argument on one 3833 processor than it must be used on all processors that share the object for 3834 that argument. 3835 3836 The user MUST specify either the local or global matrix dimensions 3837 (possibly both). 3838 3839 The parallel matrix is partitioned across processors such that the 3840 first m0 rows belong to process 0, the next m1 rows belong to 3841 process 1, the next m2 rows belong to process 2 etc.. where 3842 m0,m1,m2,.. are the input parameter 'm'. i.e each processor stores 3843 values corresponding to [m x N] submatrix. 3844 3845 The columns are logically partitioned with the n0 columns belonging 3846 to 0th partition, the next n1 columns belonging to the next 3847 partition etc.. where n0,n1,n2... are the input parameter 'n'. 3848 3849 The DIAGONAL portion of the local submatrix on any given processor 3850 is the submatrix corresponding to the rows and columns m,n 3851 corresponding to the given processor. i.e diagonal matrix on 3852 process 0 is [m0 x n0], diagonal matrix on process 1 is [m1 x n1] 3853 etc. The remaining portion of the local submatrix [m x (N-n)] 3854 constitute the OFF-DIAGONAL portion. The example below better 3855 illustrates this concept. 3856 3857 For a square global matrix we define each processor's diagonal portion 3858 to be its local rows and the corresponding columns (a square submatrix); 3859 each processor's off-diagonal portion encompasses the remainder of the 3860 local matrix (a rectangular submatrix). 3861 3862 If o_nnz, d_nnz are specified, then o_nz, and d_nz are ignored. 3863 3864 When calling this routine with a single process communicator, a matrix of 3865 type SEQAIJ is returned. If a matrix of type MATMPIAIJ is desired for this 3866 type of communicator, use the construction mechanism: 3867 MatCreate(...,&A); MatSetType(A,MATMPIAIJ); MatSetSizes(A, m,n,M,N); MatMPIAIJSetPreallocation(A,...); 3868 3869 By default, this format uses inodes (identical nodes) when possible. 3870 We search for consecutive rows with the same nonzero structure, thereby 3871 reusing matrix information to achieve increased efficiency. 3872 3873 Options Database Keys: 3874 + -mat_no_inode - Do not use inodes 3875 . -mat_inode_limit <limit> - Sets inode limit (max limit=5) 3876 - -mat_aij_oneindex - Internally use indexing starting at 1 3877 rather than 0. Note that when calling MatSetValues(), 3878 the user still MUST index entries starting at 0! 3879 3880 3881 Example usage: 3882 3883 Consider the following 8x8 matrix with 34 non-zero values, that is 3884 assembled across 3 processors. Lets assume that proc0 owns 3 rows, 3885 proc1 owns 3 rows, proc2 owns 2 rows. This division can be shown 3886 as follows: 3887 3888 .vb 3889 1 2 0 | 0 3 0 | 0 4 3890 Proc0 0 5 6 | 7 0 0 | 8 0 3891 9 0 10 | 11 0 0 | 12 0 3892 ------------------------------------- 3893 13 0 14 | 15 16 17 | 0 0 3894 Proc1 0 18 0 | 19 20 21 | 0 0 3895 0 0 0 | 22 23 0 | 24 0 3896 ------------------------------------- 3897 Proc2 25 26 27 | 0 0 28 | 29 0 3898 30 0 0 | 31 32 33 | 0 34 3899 .ve 3900 3901 This can be represented as a collection of submatrices as: 3902 3903 .vb 3904 A B C 3905 D E F 3906 G H I 3907 .ve 3908 3909 Where the submatrices A,B,C are owned by proc0, D,E,F are 3910 owned by proc1, G,H,I are owned by proc2. 3911 3912 The 'm' parameters for proc0,proc1,proc2 are 3,3,2 respectively. 3913 The 'n' parameters for proc0,proc1,proc2 are 3,3,2 respectively. 3914 The 'M','N' parameters are 8,8, and have the same values on all procs. 3915 3916 The DIAGONAL submatrices corresponding to proc0,proc1,proc2 are 3917 submatrices [A], [E], [I] respectively. The OFF-DIAGONAL submatrices 3918 corresponding to proc0,proc1,proc2 are [BC], [DF], [GH] respectively. 3919 Internally, each processor stores the DIAGONAL part, and the OFF-DIAGONAL 3920 part as SeqAIJ matrices. for eg: proc1 will store [E] as a SeqAIJ 3921 matrix, ans [DF] as another SeqAIJ matrix. 3922 3923 When d_nz, o_nz parameters are specified, d_nz storage elements are 3924 allocated for every row of the local diagonal submatrix, and o_nz 3925 storage locations are allocated for every row of the OFF-DIAGONAL submat. 3926 One way to choose d_nz and o_nz is to use the max nonzerors per local 3927 rows for each of the local DIAGONAL, and the OFF-DIAGONAL submatrices. 3928 In this case, the values of d_nz,o_nz are: 3929 .vb 3930 proc0 : dnz = 2, o_nz = 2 3931 proc1 : dnz = 3, o_nz = 2 3932 proc2 : dnz = 1, o_nz = 4 3933 .ve 3934 We are allocating m*(d_nz+o_nz) storage locations for every proc. This 3935 translates to 3*(2+2)=12 for proc0, 3*(3+2)=15 for proc1, 2*(1+4)=10 3936 for proc3. i.e we are using 12+15+10=37 storage locations to store 3937 34 values. 3938 3939 When d_nnz, o_nnz parameters are specified, the storage is specified 3940 for every row, coresponding to both DIAGONAL and OFF-DIAGONAL submatrices. 3941 In the above case the values for d_nnz,o_nnz are: 3942 .vb 3943 proc0: d_nnz = [2,2,2] and o_nnz = [2,2,2] 3944 proc1: d_nnz = [3,3,2] and o_nnz = [2,1,1] 3945 proc2: d_nnz = [1,1] and o_nnz = [4,4] 3946 .ve 3947 Here the space allocated is sum of all the above values i.e 34, and 3948 hence pre-allocation is perfect. 3949 3950 Level: intermediate 3951 3952 .keywords: matrix, aij, compressed row, sparse, parallel 3953 3954 .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatMPIAIJSetPreallocationCSR(), 3955 MATMPIAIJ, MatCreateMPIAIJWithArrays() 3956 @*/ 3957 PetscErrorCode MatCreateAIJ(MPI_Comm comm,PetscInt m,PetscInt n,PetscInt M,PetscInt N,PetscInt d_nz,const PetscInt d_nnz[],PetscInt o_nz,const PetscInt o_nnz[],Mat *A) 3958 { 3959 PetscErrorCode ierr; 3960 PetscMPIInt size; 3961 3962 PetscFunctionBegin; 3963 ierr = MatCreate(comm,A);CHKERRQ(ierr); 3964 ierr = MatSetSizes(*A,m,n,M,N);CHKERRQ(ierr); 3965 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 3966 if (size > 1) { 3967 ierr = MatSetType(*A,MATMPIAIJ);CHKERRQ(ierr); 3968 ierr = MatMPIAIJSetPreallocation(*A,d_nz,d_nnz,o_nz,o_nnz);CHKERRQ(ierr); 3969 } else { 3970 ierr = MatSetType(*A,MATSEQAIJ);CHKERRQ(ierr); 3971 ierr = MatSeqAIJSetPreallocation(*A,d_nz,d_nnz);CHKERRQ(ierr); 3972 } 3973 PetscFunctionReturn(0); 3974 } 3975 3976 PetscErrorCode MatMPIAIJGetSeqAIJ(Mat A,Mat *Ad,Mat *Ao,const PetscInt *colmap[]) 3977 { 3978 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 3979 PetscBool flg; 3980 PetscErrorCode ierr; 3981 3982 PetscFunctionBegin; 3983 ierr = PetscObjectTypeCompare((PetscObject)A,MATMPIAIJ,&flg);CHKERRQ(ierr); 3984 if (!flg) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"This function requires a MATMPIAIJ matrix as input"); 3985 if (Ad) *Ad = a->A; 3986 if (Ao) *Ao = a->B; 3987 if (colmap) *colmap = a->garray; 3988 PetscFunctionReturn(0); 3989 } 3990 3991 PetscErrorCode MatCreateMPIMatConcatenateSeqMat_MPIAIJ(MPI_Comm comm,Mat inmat,PetscInt n,MatReuse scall,Mat *outmat) 3992 { 3993 PetscErrorCode ierr; 3994 PetscInt m,N,i,rstart,nnz,Ii; 3995 PetscInt *indx; 3996 PetscScalar *values; 3997 3998 PetscFunctionBegin; 3999 ierr = MatGetSize(inmat,&m,&N);CHKERRQ(ierr); 4000 if (scall == MAT_INITIAL_MATRIX) { /* symbolic phase */ 4001 PetscInt *dnz,*onz,sum,bs,cbs; 4002 4003 if (n == PETSC_DECIDE) { 4004 ierr = PetscSplitOwnership(comm,&n,&N);CHKERRQ(ierr); 4005 } 4006 /* Check sum(n) = N */ 4007 ierr = MPIU_Allreduce(&n,&sum,1,MPIU_INT,MPI_SUM,comm);CHKERRQ(ierr); 4008 if (sum != N) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_INCOMP,"Sum of local columns != global columns %d",N); 4009 4010 ierr = MPI_Scan(&m, &rstart,1,MPIU_INT,MPI_SUM,comm);CHKERRQ(ierr); 4011 rstart -= m; 4012 4013 ierr = MatPreallocateInitialize(comm,m,n,dnz,onz);CHKERRQ(ierr); 4014 for (i=0; i<m; i++) { 4015 ierr = MatGetRow_SeqAIJ(inmat,i,&nnz,&indx,NULL);CHKERRQ(ierr); 4016 ierr = MatPreallocateSet(i+rstart,nnz,indx,dnz,onz);CHKERRQ(ierr); 4017 ierr = MatRestoreRow_SeqAIJ(inmat,i,&nnz,&indx,NULL);CHKERRQ(ierr); 4018 } 4019 4020 ierr = MatCreate(comm,outmat);CHKERRQ(ierr); 4021 ierr = MatSetSizes(*outmat,m,n,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 4022 ierr = MatGetBlockSizes(inmat,&bs,&cbs);CHKERRQ(ierr); 4023 ierr = MatSetBlockSizes(*outmat,bs,cbs);CHKERRQ(ierr); 4024 ierr = MatSetType(*outmat,MATAIJ);CHKERRQ(ierr); 4025 ierr = MatSeqAIJSetPreallocation(*outmat,0,dnz);CHKERRQ(ierr); 4026 ierr = MatMPIAIJSetPreallocation(*outmat,0,dnz,0,onz);CHKERRQ(ierr); 4027 ierr = MatPreallocateFinalize(dnz,onz);CHKERRQ(ierr); 4028 } 4029 4030 /* numeric phase */ 4031 ierr = MatGetOwnershipRange(*outmat,&rstart,NULL);CHKERRQ(ierr); 4032 for (i=0; i<m; i++) { 4033 ierr = MatGetRow_SeqAIJ(inmat,i,&nnz,&indx,&values);CHKERRQ(ierr); 4034 Ii = i + rstart; 4035 ierr = MatSetValues(*outmat,1,&Ii,nnz,indx,values,INSERT_VALUES);CHKERRQ(ierr); 4036 ierr = MatRestoreRow_SeqAIJ(inmat,i,&nnz,&indx,&values);CHKERRQ(ierr); 4037 } 4038 ierr = MatAssemblyBegin(*outmat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4039 ierr = MatAssemblyEnd(*outmat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4040 PetscFunctionReturn(0); 4041 } 4042 4043 PetscErrorCode MatFileSplit(Mat A,char *outfile) 4044 { 4045 PetscErrorCode ierr; 4046 PetscMPIInt rank; 4047 PetscInt m,N,i,rstart,nnz; 4048 size_t len; 4049 const PetscInt *indx; 4050 PetscViewer out; 4051 char *name; 4052 Mat B; 4053 const PetscScalar *values; 4054 4055 PetscFunctionBegin; 4056 ierr = MatGetLocalSize(A,&m,0);CHKERRQ(ierr); 4057 ierr = MatGetSize(A,0,&N);CHKERRQ(ierr); 4058 /* Should this be the type of the diagonal block of A? */ 4059 ierr = MatCreate(PETSC_COMM_SELF,&B);CHKERRQ(ierr); 4060 ierr = MatSetSizes(B,m,N,m,N);CHKERRQ(ierr); 4061 ierr = MatSetBlockSizesFromMats(B,A,A);CHKERRQ(ierr); 4062 ierr = MatSetType(B,MATSEQAIJ);CHKERRQ(ierr); 4063 ierr = MatSeqAIJSetPreallocation(B,0,NULL);CHKERRQ(ierr); 4064 ierr = MatGetOwnershipRange(A,&rstart,0);CHKERRQ(ierr); 4065 for (i=0; i<m; i++) { 4066 ierr = MatGetRow(A,i+rstart,&nnz,&indx,&values);CHKERRQ(ierr); 4067 ierr = MatSetValues(B,1,&i,nnz,indx,values,INSERT_VALUES);CHKERRQ(ierr); 4068 ierr = MatRestoreRow(A,i+rstart,&nnz,&indx,&values);CHKERRQ(ierr); 4069 } 4070 ierr = MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4071 ierr = MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4072 4073 ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)A),&rank);CHKERRQ(ierr); 4074 ierr = PetscStrlen(outfile,&len);CHKERRQ(ierr); 4075 ierr = PetscMalloc1(len+5,&name);CHKERRQ(ierr); 4076 sprintf(name,"%s.%d",outfile,rank); 4077 ierr = PetscViewerBinaryOpen(PETSC_COMM_SELF,name,FILE_MODE_APPEND,&out);CHKERRQ(ierr); 4078 ierr = PetscFree(name);CHKERRQ(ierr); 4079 ierr = MatView(B,out);CHKERRQ(ierr); 4080 ierr = PetscViewerDestroy(&out);CHKERRQ(ierr); 4081 ierr = MatDestroy(&B);CHKERRQ(ierr); 4082 PetscFunctionReturn(0); 4083 } 4084 4085 PetscErrorCode MatDestroy_MPIAIJ_SeqsToMPI(Mat A) 4086 { 4087 PetscErrorCode ierr; 4088 Mat_Merge_SeqsToMPI *merge; 4089 PetscContainer container; 4090 4091 PetscFunctionBegin; 4092 ierr = PetscObjectQuery((PetscObject)A,"MatMergeSeqsToMPI",(PetscObject*)&container);CHKERRQ(ierr); 4093 if (container) { 4094 ierr = PetscContainerGetPointer(container,(void**)&merge);CHKERRQ(ierr); 4095 ierr = PetscFree(merge->id_r);CHKERRQ(ierr); 4096 ierr = PetscFree(merge->len_s);CHKERRQ(ierr); 4097 ierr = PetscFree(merge->len_r);CHKERRQ(ierr); 4098 ierr = PetscFree(merge->bi);CHKERRQ(ierr); 4099 ierr = PetscFree(merge->bj);CHKERRQ(ierr); 4100 ierr = PetscFree(merge->buf_ri[0]);CHKERRQ(ierr); 4101 ierr = PetscFree(merge->buf_ri);CHKERRQ(ierr); 4102 ierr = PetscFree(merge->buf_rj[0]);CHKERRQ(ierr); 4103 ierr = PetscFree(merge->buf_rj);CHKERRQ(ierr); 4104 ierr = PetscFree(merge->coi);CHKERRQ(ierr); 4105 ierr = PetscFree(merge->coj);CHKERRQ(ierr); 4106 ierr = PetscFree(merge->owners_co);CHKERRQ(ierr); 4107 ierr = PetscLayoutDestroy(&merge->rowmap);CHKERRQ(ierr); 4108 ierr = PetscFree(merge);CHKERRQ(ierr); 4109 ierr = PetscObjectCompose((PetscObject)A,"MatMergeSeqsToMPI",0);CHKERRQ(ierr); 4110 } 4111 ierr = MatDestroy_MPIAIJ(A);CHKERRQ(ierr); 4112 PetscFunctionReturn(0); 4113 } 4114 4115 #include <../src/mat/utils/freespace.h> 4116 #include <petscbt.h> 4117 4118 PetscErrorCode MatCreateMPIAIJSumSeqAIJNumeric(Mat seqmat,Mat mpimat) 4119 { 4120 PetscErrorCode ierr; 4121 MPI_Comm comm; 4122 Mat_SeqAIJ *a =(Mat_SeqAIJ*)seqmat->data; 4123 PetscMPIInt size,rank,taga,*len_s; 4124 PetscInt N=mpimat->cmap->N,i,j,*owners,*ai=a->i,*aj; 4125 PetscInt proc,m; 4126 PetscInt **buf_ri,**buf_rj; 4127 PetscInt k,anzi,*bj_i,*bi,*bj,arow,bnzi,nextaj; 4128 PetscInt nrows,**buf_ri_k,**nextrow,**nextai; 4129 MPI_Request *s_waits,*r_waits; 4130 MPI_Status *status; 4131 MatScalar *aa=a->a; 4132 MatScalar **abuf_r,*ba_i; 4133 Mat_Merge_SeqsToMPI *merge; 4134 PetscContainer container; 4135 4136 PetscFunctionBegin; 4137 ierr = PetscObjectGetComm((PetscObject)mpimat,&comm);CHKERRQ(ierr); 4138 ierr = PetscLogEventBegin(MAT_Seqstompinum,seqmat,0,0,0);CHKERRQ(ierr); 4139 4140 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 4141 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 4142 4143 ierr = PetscObjectQuery((PetscObject)mpimat,"MatMergeSeqsToMPI",(PetscObject*)&container);CHKERRQ(ierr); 4144 ierr = PetscContainerGetPointer(container,(void**)&merge);CHKERRQ(ierr); 4145 4146 bi = merge->bi; 4147 bj = merge->bj; 4148 buf_ri = merge->buf_ri; 4149 buf_rj = merge->buf_rj; 4150 4151 ierr = PetscMalloc1(size,&status);CHKERRQ(ierr); 4152 owners = merge->rowmap->range; 4153 len_s = merge->len_s; 4154 4155 /* send and recv matrix values */ 4156 /*-----------------------------*/ 4157 ierr = PetscObjectGetNewTag((PetscObject)mpimat,&taga);CHKERRQ(ierr); 4158 ierr = PetscPostIrecvScalar(comm,taga,merge->nrecv,merge->id_r,merge->len_r,&abuf_r,&r_waits);CHKERRQ(ierr); 4159 4160 ierr = PetscMalloc1(merge->nsend+1,&s_waits);CHKERRQ(ierr); 4161 for (proc=0,k=0; proc<size; proc++) { 4162 if (!len_s[proc]) continue; 4163 i = owners[proc]; 4164 ierr = MPI_Isend(aa+ai[i],len_s[proc],MPIU_MATSCALAR,proc,taga,comm,s_waits+k);CHKERRQ(ierr); 4165 k++; 4166 } 4167 4168 if (merge->nrecv) {ierr = MPI_Waitall(merge->nrecv,r_waits,status);CHKERRQ(ierr);} 4169 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,s_waits,status);CHKERRQ(ierr);} 4170 ierr = PetscFree(status);CHKERRQ(ierr); 4171 4172 ierr = PetscFree(s_waits);CHKERRQ(ierr); 4173 ierr = PetscFree(r_waits);CHKERRQ(ierr); 4174 4175 /* insert mat values of mpimat */ 4176 /*----------------------------*/ 4177 ierr = PetscMalloc1(N,&ba_i);CHKERRQ(ierr); 4178 ierr = PetscMalloc3(merge->nrecv,&buf_ri_k,merge->nrecv,&nextrow,merge->nrecv,&nextai);CHKERRQ(ierr); 4179 4180 for (k=0; k<merge->nrecv; k++) { 4181 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 4182 nrows = *(buf_ri_k[k]); 4183 nextrow[k] = buf_ri_k[k]+1; /* next row number of k-th recved i-structure */ 4184 nextai[k] = buf_ri_k[k] + (nrows + 1); /* poins to the next i-structure of k-th recved i-structure */ 4185 } 4186 4187 /* set values of ba */ 4188 m = merge->rowmap->n; 4189 for (i=0; i<m; i++) { 4190 arow = owners[rank] + i; 4191 bj_i = bj+bi[i]; /* col indices of the i-th row of mpimat */ 4192 bnzi = bi[i+1] - bi[i]; 4193 ierr = PetscMemzero(ba_i,bnzi*sizeof(PetscScalar));CHKERRQ(ierr); 4194 4195 /* add local non-zero vals of this proc's seqmat into ba */ 4196 anzi = ai[arow+1] - ai[arow]; 4197 aj = a->j + ai[arow]; 4198 aa = a->a + ai[arow]; 4199 nextaj = 0; 4200 for (j=0; nextaj<anzi; j++) { 4201 if (*(bj_i + j) == aj[nextaj]) { /* bcol == acol */ 4202 ba_i[j] += aa[nextaj++]; 4203 } 4204 } 4205 4206 /* add received vals into ba */ 4207 for (k=0; k<merge->nrecv; k++) { /* k-th received message */ 4208 /* i-th row */ 4209 if (i == *nextrow[k]) { 4210 anzi = *(nextai[k]+1) - *nextai[k]; 4211 aj = buf_rj[k] + *(nextai[k]); 4212 aa = abuf_r[k] + *(nextai[k]); 4213 nextaj = 0; 4214 for (j=0; nextaj<anzi; j++) { 4215 if (*(bj_i + j) == aj[nextaj]) { /* bcol == acol */ 4216 ba_i[j] += aa[nextaj++]; 4217 } 4218 } 4219 nextrow[k]++; nextai[k]++; 4220 } 4221 } 4222 ierr = MatSetValues(mpimat,1,&arow,bnzi,bj_i,ba_i,INSERT_VALUES);CHKERRQ(ierr); 4223 } 4224 ierr = MatAssemblyBegin(mpimat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4225 ierr = MatAssemblyEnd(mpimat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4226 4227 ierr = PetscFree(abuf_r[0]);CHKERRQ(ierr); 4228 ierr = PetscFree(abuf_r);CHKERRQ(ierr); 4229 ierr = PetscFree(ba_i);CHKERRQ(ierr); 4230 ierr = PetscFree3(buf_ri_k,nextrow,nextai);CHKERRQ(ierr); 4231 ierr = PetscLogEventEnd(MAT_Seqstompinum,seqmat,0,0,0);CHKERRQ(ierr); 4232 PetscFunctionReturn(0); 4233 } 4234 4235 PetscErrorCode MatCreateMPIAIJSumSeqAIJSymbolic(MPI_Comm comm,Mat seqmat,PetscInt m,PetscInt n,Mat *mpimat) 4236 { 4237 PetscErrorCode ierr; 4238 Mat B_mpi; 4239 Mat_SeqAIJ *a=(Mat_SeqAIJ*)seqmat->data; 4240 PetscMPIInt size,rank,tagi,tagj,*len_s,*len_si,*len_ri; 4241 PetscInt **buf_rj,**buf_ri,**buf_ri_k; 4242 PetscInt M=seqmat->rmap->n,N=seqmat->cmap->n,i,*owners,*ai=a->i,*aj=a->j; 4243 PetscInt len,proc,*dnz,*onz,bs,cbs; 4244 PetscInt k,anzi,*bi,*bj,*lnk,nlnk,arow,bnzi,nspacedouble=0; 4245 PetscInt nrows,*buf_s,*buf_si,*buf_si_i,**nextrow,**nextai; 4246 MPI_Request *si_waits,*sj_waits,*ri_waits,*rj_waits; 4247 MPI_Status *status; 4248 PetscFreeSpaceList free_space=NULL,current_space=NULL; 4249 PetscBT lnkbt; 4250 Mat_Merge_SeqsToMPI *merge; 4251 PetscContainer container; 4252 4253 PetscFunctionBegin; 4254 ierr = PetscLogEventBegin(MAT_Seqstompisym,seqmat,0,0,0);CHKERRQ(ierr); 4255 4256 /* make sure it is a PETSc comm */ 4257 ierr = PetscCommDuplicate(comm,&comm,NULL);CHKERRQ(ierr); 4258 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 4259 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 4260 4261 ierr = PetscNew(&merge);CHKERRQ(ierr); 4262 ierr = PetscMalloc1(size,&status);CHKERRQ(ierr); 4263 4264 /* determine row ownership */ 4265 /*---------------------------------------------------------*/ 4266 ierr = PetscLayoutCreate(comm,&merge->rowmap);CHKERRQ(ierr); 4267 ierr = PetscLayoutSetLocalSize(merge->rowmap,m);CHKERRQ(ierr); 4268 ierr = PetscLayoutSetSize(merge->rowmap,M);CHKERRQ(ierr); 4269 ierr = PetscLayoutSetBlockSize(merge->rowmap,1);CHKERRQ(ierr); 4270 ierr = PetscLayoutSetUp(merge->rowmap);CHKERRQ(ierr); 4271 ierr = PetscMalloc1(size,&len_si);CHKERRQ(ierr); 4272 ierr = PetscMalloc1(size,&merge->len_s);CHKERRQ(ierr); 4273 4274 m = merge->rowmap->n; 4275 owners = merge->rowmap->range; 4276 4277 /* determine the number of messages to send, their lengths */ 4278 /*---------------------------------------------------------*/ 4279 len_s = merge->len_s; 4280 4281 len = 0; /* length of buf_si[] */ 4282 merge->nsend = 0; 4283 for (proc=0; proc<size; proc++) { 4284 len_si[proc] = 0; 4285 if (proc == rank) { 4286 len_s[proc] = 0; 4287 } else { 4288 len_si[proc] = owners[proc+1] - owners[proc] + 1; 4289 len_s[proc] = ai[owners[proc+1]] - ai[owners[proc]]; /* num of rows to be sent to [proc] */ 4290 } 4291 if (len_s[proc]) { 4292 merge->nsend++; 4293 nrows = 0; 4294 for (i=owners[proc]; i<owners[proc+1]; i++) { 4295 if (ai[i+1] > ai[i]) nrows++; 4296 } 4297 len_si[proc] = 2*(nrows+1); 4298 len += len_si[proc]; 4299 } 4300 } 4301 4302 /* determine the number and length of messages to receive for ij-structure */ 4303 /*-------------------------------------------------------------------------*/ 4304 ierr = PetscGatherNumberOfMessages(comm,NULL,len_s,&merge->nrecv);CHKERRQ(ierr); 4305 ierr = PetscGatherMessageLengths2(comm,merge->nsend,merge->nrecv,len_s,len_si,&merge->id_r,&merge->len_r,&len_ri);CHKERRQ(ierr); 4306 4307 /* post the Irecv of j-structure */ 4308 /*-------------------------------*/ 4309 ierr = PetscCommGetNewTag(comm,&tagj);CHKERRQ(ierr); 4310 ierr = PetscPostIrecvInt(comm,tagj,merge->nrecv,merge->id_r,merge->len_r,&buf_rj,&rj_waits);CHKERRQ(ierr); 4311 4312 /* post the Isend of j-structure */ 4313 /*--------------------------------*/ 4314 ierr = PetscMalloc2(merge->nsend,&si_waits,merge->nsend,&sj_waits);CHKERRQ(ierr); 4315 4316 for (proc=0, k=0; proc<size; proc++) { 4317 if (!len_s[proc]) continue; 4318 i = owners[proc]; 4319 ierr = MPI_Isend(aj+ai[i],len_s[proc],MPIU_INT,proc,tagj,comm,sj_waits+k);CHKERRQ(ierr); 4320 k++; 4321 } 4322 4323 /* receives and sends of j-structure are complete */ 4324 /*------------------------------------------------*/ 4325 if (merge->nrecv) {ierr = MPI_Waitall(merge->nrecv,rj_waits,status);CHKERRQ(ierr);} 4326 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,sj_waits,status);CHKERRQ(ierr);} 4327 4328 /* send and recv i-structure */ 4329 /*---------------------------*/ 4330 ierr = PetscCommGetNewTag(comm,&tagi);CHKERRQ(ierr); 4331 ierr = PetscPostIrecvInt(comm,tagi,merge->nrecv,merge->id_r,len_ri,&buf_ri,&ri_waits);CHKERRQ(ierr); 4332 4333 ierr = PetscMalloc1(len+1,&buf_s);CHKERRQ(ierr); 4334 buf_si = buf_s; /* points to the beginning of k-th msg to be sent */ 4335 for (proc=0,k=0; proc<size; proc++) { 4336 if (!len_s[proc]) continue; 4337 /* form outgoing message for i-structure: 4338 buf_si[0]: nrows to be sent 4339 [1:nrows]: row index (global) 4340 [nrows+1:2*nrows+1]: i-structure index 4341 */ 4342 /*-------------------------------------------*/ 4343 nrows = len_si[proc]/2 - 1; 4344 buf_si_i = buf_si + nrows+1; 4345 buf_si[0] = nrows; 4346 buf_si_i[0] = 0; 4347 nrows = 0; 4348 for (i=owners[proc]; i<owners[proc+1]; i++) { 4349 anzi = ai[i+1] - ai[i]; 4350 if (anzi) { 4351 buf_si_i[nrows+1] = buf_si_i[nrows] + anzi; /* i-structure */ 4352 buf_si[nrows+1] = i-owners[proc]; /* local row index */ 4353 nrows++; 4354 } 4355 } 4356 ierr = MPI_Isend(buf_si,len_si[proc],MPIU_INT,proc,tagi,comm,si_waits+k);CHKERRQ(ierr); 4357 k++; 4358 buf_si += len_si[proc]; 4359 } 4360 4361 if (merge->nrecv) {ierr = MPI_Waitall(merge->nrecv,ri_waits,status);CHKERRQ(ierr);} 4362 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,si_waits,status);CHKERRQ(ierr);} 4363 4364 ierr = PetscInfo2(seqmat,"nsend: %D, nrecv: %D\n",merge->nsend,merge->nrecv);CHKERRQ(ierr); 4365 for (i=0; i<merge->nrecv; i++) { 4366 ierr = PetscInfo3(seqmat,"recv len_ri=%D, len_rj=%D from [%D]\n",len_ri[i],merge->len_r[i],merge->id_r[i]);CHKERRQ(ierr); 4367 } 4368 4369 ierr = PetscFree(len_si);CHKERRQ(ierr); 4370 ierr = PetscFree(len_ri);CHKERRQ(ierr); 4371 ierr = PetscFree(rj_waits);CHKERRQ(ierr); 4372 ierr = PetscFree2(si_waits,sj_waits);CHKERRQ(ierr); 4373 ierr = PetscFree(ri_waits);CHKERRQ(ierr); 4374 ierr = PetscFree(buf_s);CHKERRQ(ierr); 4375 ierr = PetscFree(status);CHKERRQ(ierr); 4376 4377 /* compute a local seq matrix in each processor */ 4378 /*----------------------------------------------*/ 4379 /* allocate bi array and free space for accumulating nonzero column info */ 4380 ierr = PetscMalloc1(m+1,&bi);CHKERRQ(ierr); 4381 bi[0] = 0; 4382 4383 /* create and initialize a linked list */ 4384 nlnk = N+1; 4385 ierr = PetscLLCreate(N,N,nlnk,lnk,lnkbt);CHKERRQ(ierr); 4386 4387 /* initial FreeSpace size is 2*(num of local nnz(seqmat)) */ 4388 len = ai[owners[rank+1]] - ai[owners[rank]]; 4389 ierr = PetscFreeSpaceGet(PetscIntMultTruncate(2,len)+1,&free_space);CHKERRQ(ierr); 4390 4391 current_space = free_space; 4392 4393 /* determine symbolic info for each local row */ 4394 ierr = PetscMalloc3(merge->nrecv,&buf_ri_k,merge->nrecv,&nextrow,merge->nrecv,&nextai);CHKERRQ(ierr); 4395 4396 for (k=0; k<merge->nrecv; k++) { 4397 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 4398 nrows = *buf_ri_k[k]; 4399 nextrow[k] = buf_ri_k[k] + 1; /* next row number of k-th recved i-structure */ 4400 nextai[k] = buf_ri_k[k] + (nrows + 1); /* poins to the next i-structure of k-th recved i-structure */ 4401 } 4402 4403 ierr = MatPreallocateInitialize(comm,m,n,dnz,onz);CHKERRQ(ierr); 4404 len = 0; 4405 for (i=0; i<m; i++) { 4406 bnzi = 0; 4407 /* add local non-zero cols of this proc's seqmat into lnk */ 4408 arow = owners[rank] + i; 4409 anzi = ai[arow+1] - ai[arow]; 4410 aj = a->j + ai[arow]; 4411 ierr = PetscLLAddSorted(anzi,aj,N,nlnk,lnk,lnkbt);CHKERRQ(ierr); 4412 bnzi += nlnk; 4413 /* add received col data into lnk */ 4414 for (k=0; k<merge->nrecv; k++) { /* k-th received message */ 4415 if (i == *nextrow[k]) { /* i-th row */ 4416 anzi = *(nextai[k]+1) - *nextai[k]; 4417 aj = buf_rj[k] + *nextai[k]; 4418 ierr = PetscLLAddSorted(anzi,aj,N,nlnk,lnk,lnkbt);CHKERRQ(ierr); 4419 bnzi += nlnk; 4420 nextrow[k]++; nextai[k]++; 4421 } 4422 } 4423 if (len < bnzi) len = bnzi; /* =max(bnzi) */ 4424 4425 /* if free space is not available, make more free space */ 4426 if (current_space->local_remaining<bnzi) { 4427 ierr = PetscFreeSpaceGet(PetscIntSumTruncate(bnzi,current_space->total_array_size),¤t_space);CHKERRQ(ierr); 4428 nspacedouble++; 4429 } 4430 /* copy data into free space, then initialize lnk */ 4431 ierr = PetscLLClean(N,N,bnzi,lnk,current_space->array,lnkbt);CHKERRQ(ierr); 4432 ierr = MatPreallocateSet(i+owners[rank],bnzi,current_space->array,dnz,onz);CHKERRQ(ierr); 4433 4434 current_space->array += bnzi; 4435 current_space->local_used += bnzi; 4436 current_space->local_remaining -= bnzi; 4437 4438 bi[i+1] = bi[i] + bnzi; 4439 } 4440 4441 ierr = PetscFree3(buf_ri_k,nextrow,nextai);CHKERRQ(ierr); 4442 4443 ierr = PetscMalloc1(bi[m]+1,&bj);CHKERRQ(ierr); 4444 ierr = PetscFreeSpaceContiguous(&free_space,bj);CHKERRQ(ierr); 4445 ierr = PetscLLDestroy(lnk,lnkbt);CHKERRQ(ierr); 4446 4447 /* create symbolic parallel matrix B_mpi */ 4448 /*---------------------------------------*/ 4449 ierr = MatGetBlockSizes(seqmat,&bs,&cbs);CHKERRQ(ierr); 4450 ierr = MatCreate(comm,&B_mpi);CHKERRQ(ierr); 4451 if (n==PETSC_DECIDE) { 4452 ierr = MatSetSizes(B_mpi,m,n,PETSC_DETERMINE,N);CHKERRQ(ierr); 4453 } else { 4454 ierr = MatSetSizes(B_mpi,m,n,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 4455 } 4456 ierr = MatSetBlockSizes(B_mpi,bs,cbs);CHKERRQ(ierr); 4457 ierr = MatSetType(B_mpi,MATMPIAIJ);CHKERRQ(ierr); 4458 ierr = MatMPIAIJSetPreallocation(B_mpi,0,dnz,0,onz);CHKERRQ(ierr); 4459 ierr = MatPreallocateFinalize(dnz,onz);CHKERRQ(ierr); 4460 ierr = MatSetOption(B_mpi,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 4461 4462 /* B_mpi is not ready for use - assembly will be done by MatCreateMPIAIJSumSeqAIJNumeric() */ 4463 B_mpi->assembled = PETSC_FALSE; 4464 B_mpi->ops->destroy = MatDestroy_MPIAIJ_SeqsToMPI; 4465 merge->bi = bi; 4466 merge->bj = bj; 4467 merge->buf_ri = buf_ri; 4468 merge->buf_rj = buf_rj; 4469 merge->coi = NULL; 4470 merge->coj = NULL; 4471 merge->owners_co = NULL; 4472 4473 ierr = PetscCommDestroy(&comm);CHKERRQ(ierr); 4474 4475 /* attach the supporting struct to B_mpi for reuse */ 4476 ierr = PetscContainerCreate(PETSC_COMM_SELF,&container);CHKERRQ(ierr); 4477 ierr = PetscContainerSetPointer(container,merge);CHKERRQ(ierr); 4478 ierr = PetscObjectCompose((PetscObject)B_mpi,"MatMergeSeqsToMPI",(PetscObject)container);CHKERRQ(ierr); 4479 ierr = PetscContainerDestroy(&container);CHKERRQ(ierr); 4480 *mpimat = B_mpi; 4481 4482 ierr = PetscLogEventEnd(MAT_Seqstompisym,seqmat,0,0,0);CHKERRQ(ierr); 4483 PetscFunctionReturn(0); 4484 } 4485 4486 /*@C 4487 MatCreateMPIAIJSumSeqAIJ - Creates a MATMPIAIJ matrix by adding sequential 4488 matrices from each processor 4489 4490 Collective on MPI_Comm 4491 4492 Input Parameters: 4493 + comm - the communicators the parallel matrix will live on 4494 . seqmat - the input sequential matrices 4495 . m - number of local rows (or PETSC_DECIDE) 4496 . n - number of local columns (or PETSC_DECIDE) 4497 - scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX 4498 4499 Output Parameter: 4500 . mpimat - the parallel matrix generated 4501 4502 Level: advanced 4503 4504 Notes: 4505 The dimensions of the sequential matrix in each processor MUST be the same. 4506 The input seqmat is included into the container "Mat_Merge_SeqsToMPI", and will be 4507 destroyed when mpimat is destroyed. Call PetscObjectQuery() to access seqmat. 4508 @*/ 4509 PetscErrorCode MatCreateMPIAIJSumSeqAIJ(MPI_Comm comm,Mat seqmat,PetscInt m,PetscInt n,MatReuse scall,Mat *mpimat) 4510 { 4511 PetscErrorCode ierr; 4512 PetscMPIInt size; 4513 4514 PetscFunctionBegin; 4515 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 4516 if (size == 1) { 4517 ierr = PetscLogEventBegin(MAT_Seqstompi,seqmat,0,0,0);CHKERRQ(ierr); 4518 if (scall == MAT_INITIAL_MATRIX) { 4519 ierr = MatDuplicate(seqmat,MAT_COPY_VALUES,mpimat);CHKERRQ(ierr); 4520 } else { 4521 ierr = MatCopy(seqmat,*mpimat,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 4522 } 4523 ierr = PetscLogEventEnd(MAT_Seqstompi,seqmat,0,0,0);CHKERRQ(ierr); 4524 PetscFunctionReturn(0); 4525 } 4526 ierr = PetscLogEventBegin(MAT_Seqstompi,seqmat,0,0,0);CHKERRQ(ierr); 4527 if (scall == MAT_INITIAL_MATRIX) { 4528 ierr = MatCreateMPIAIJSumSeqAIJSymbolic(comm,seqmat,m,n,mpimat);CHKERRQ(ierr); 4529 } 4530 ierr = MatCreateMPIAIJSumSeqAIJNumeric(seqmat,*mpimat);CHKERRQ(ierr); 4531 ierr = PetscLogEventEnd(MAT_Seqstompi,seqmat,0,0,0);CHKERRQ(ierr); 4532 PetscFunctionReturn(0); 4533 } 4534 4535 /*@ 4536 MatMPIAIJGetLocalMat - Creates a SeqAIJ from a MATMPIAIJ matrix by taking all its local rows and putting them into a sequential vector with 4537 mlocal rows and n columns. Where mlocal is the row count obtained with MatGetLocalSize() and n is the global column count obtained 4538 with MatGetSize() 4539 4540 Not Collective 4541 4542 Input Parameters: 4543 + A - the matrix 4544 . scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX 4545 4546 Output Parameter: 4547 . A_loc - the local sequential matrix generated 4548 4549 Level: developer 4550 4551 .seealso: MatGetOwnerShipRange(), MatMPIAIJGetLocalMatCondensed() 4552 4553 @*/ 4554 PetscErrorCode MatMPIAIJGetLocalMat(Mat A,MatReuse scall,Mat *A_loc) 4555 { 4556 PetscErrorCode ierr; 4557 Mat_MPIAIJ *mpimat=(Mat_MPIAIJ*)A->data; 4558 Mat_SeqAIJ *mat,*a,*b; 4559 PetscInt *ai,*aj,*bi,*bj,*cmap=mpimat->garray; 4560 MatScalar *aa,*ba,*cam; 4561 PetscScalar *ca; 4562 PetscInt am=A->rmap->n,i,j,k,cstart=A->cmap->rstart; 4563 PetscInt *ci,*cj,col,ncols_d,ncols_o,jo; 4564 PetscBool match; 4565 MPI_Comm comm; 4566 PetscMPIInt size; 4567 4568 PetscFunctionBegin; 4569 ierr = PetscObjectTypeCompare((PetscObject)A,MATMPIAIJ,&match);CHKERRQ(ierr); 4570 if (!match) SETERRQ(PetscObjectComm((PetscObject)A), PETSC_ERR_SUP,"Requires MATMPIAIJ matrix as input"); 4571 ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 4572 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 4573 if (size == 1 && scall == MAT_REUSE_MATRIX) PetscFunctionReturn(0); 4574 4575 ierr = PetscLogEventBegin(MAT_Getlocalmat,A,0,0,0);CHKERRQ(ierr); 4576 a = (Mat_SeqAIJ*)(mpimat->A)->data; 4577 b = (Mat_SeqAIJ*)(mpimat->B)->data; 4578 ai = a->i; aj = a->j; bi = b->i; bj = b->j; 4579 aa = a->a; ba = b->a; 4580 if (scall == MAT_INITIAL_MATRIX) { 4581 if (size == 1) { 4582 ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,am,A->cmap->N,ai,aj,aa,A_loc);CHKERRQ(ierr); 4583 PetscFunctionReturn(0); 4584 } 4585 4586 ierr = PetscMalloc1(1+am,&ci);CHKERRQ(ierr); 4587 ci[0] = 0; 4588 for (i=0; i<am; i++) { 4589 ci[i+1] = ci[i] + (ai[i+1] - ai[i]) + (bi[i+1] - bi[i]); 4590 } 4591 ierr = PetscMalloc1(1+ci[am],&cj);CHKERRQ(ierr); 4592 ierr = PetscMalloc1(1+ci[am],&ca);CHKERRQ(ierr); 4593 k = 0; 4594 for (i=0; i<am; i++) { 4595 ncols_o = bi[i+1] - bi[i]; 4596 ncols_d = ai[i+1] - ai[i]; 4597 /* off-diagonal portion of A */ 4598 for (jo=0; jo<ncols_o; jo++) { 4599 col = cmap[*bj]; 4600 if (col >= cstart) break; 4601 cj[k] = col; bj++; 4602 ca[k++] = *ba++; 4603 } 4604 /* diagonal portion of A */ 4605 for (j=0; j<ncols_d; j++) { 4606 cj[k] = cstart + *aj++; 4607 ca[k++] = *aa++; 4608 } 4609 /* off-diagonal portion of A */ 4610 for (j=jo; j<ncols_o; j++) { 4611 cj[k] = cmap[*bj++]; 4612 ca[k++] = *ba++; 4613 } 4614 } 4615 /* put together the new matrix */ 4616 ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,am,A->cmap->N,ci,cj,ca,A_loc);CHKERRQ(ierr); 4617 /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */ 4618 /* Since these are PETSc arrays, change flags to free them as necessary. */ 4619 mat = (Mat_SeqAIJ*)(*A_loc)->data; 4620 mat->free_a = PETSC_TRUE; 4621 mat->free_ij = PETSC_TRUE; 4622 mat->nonew = 0; 4623 } else if (scall == MAT_REUSE_MATRIX) { 4624 mat=(Mat_SeqAIJ*)(*A_loc)->data; 4625 ci = mat->i; cj = mat->j; cam = mat->a; 4626 for (i=0; i<am; i++) { 4627 /* off-diagonal portion of A */ 4628 ncols_o = bi[i+1] - bi[i]; 4629 for (jo=0; jo<ncols_o; jo++) { 4630 col = cmap[*bj]; 4631 if (col >= cstart) break; 4632 *cam++ = *ba++; bj++; 4633 } 4634 /* diagonal portion of A */ 4635 ncols_d = ai[i+1] - ai[i]; 4636 for (j=0; j<ncols_d; j++) *cam++ = *aa++; 4637 /* off-diagonal portion of A */ 4638 for (j=jo; j<ncols_o; j++) { 4639 *cam++ = *ba++; bj++; 4640 } 4641 } 4642 } else SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Invalid MatReuse %d",(int)scall); 4643 ierr = PetscLogEventEnd(MAT_Getlocalmat,A,0,0,0);CHKERRQ(ierr); 4644 PetscFunctionReturn(0); 4645 } 4646 4647 /*@C 4648 MatMPIAIJGetLocalMatCondensed - Creates a SeqAIJ matrix from an MATMPIAIJ matrix by taking all its local rows and NON-ZERO columns 4649 4650 Not Collective 4651 4652 Input Parameters: 4653 + A - the matrix 4654 . scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX 4655 - row, col - index sets of rows and columns to extract (or NULL) 4656 4657 Output Parameter: 4658 . A_loc - the local sequential matrix generated 4659 4660 Level: developer 4661 4662 .seealso: MatGetOwnershipRange(), MatMPIAIJGetLocalMat() 4663 4664 @*/ 4665 PetscErrorCode MatMPIAIJGetLocalMatCondensed(Mat A,MatReuse scall,IS *row,IS *col,Mat *A_loc) 4666 { 4667 Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data; 4668 PetscErrorCode ierr; 4669 PetscInt i,start,end,ncols,nzA,nzB,*cmap,imark,*idx; 4670 IS isrowa,iscola; 4671 Mat *aloc; 4672 PetscBool match; 4673 4674 PetscFunctionBegin; 4675 ierr = PetscObjectTypeCompare((PetscObject)A,MATMPIAIJ,&match);CHKERRQ(ierr); 4676 if (!match) SETERRQ(PetscObjectComm((PetscObject)A), PETSC_ERR_SUP,"Requires MATMPIAIJ matrix as input"); 4677 ierr = PetscLogEventBegin(MAT_Getlocalmatcondensed,A,0,0,0);CHKERRQ(ierr); 4678 if (!row) { 4679 start = A->rmap->rstart; end = A->rmap->rend; 4680 ierr = ISCreateStride(PETSC_COMM_SELF,end-start,start,1,&isrowa);CHKERRQ(ierr); 4681 } else { 4682 isrowa = *row; 4683 } 4684 if (!col) { 4685 start = A->cmap->rstart; 4686 cmap = a->garray; 4687 nzA = a->A->cmap->n; 4688 nzB = a->B->cmap->n; 4689 ierr = PetscMalloc1(nzA+nzB, &idx);CHKERRQ(ierr); 4690 ncols = 0; 4691 for (i=0; i<nzB; i++) { 4692 if (cmap[i] < start) idx[ncols++] = cmap[i]; 4693 else break; 4694 } 4695 imark = i; 4696 for (i=0; i<nzA; i++) idx[ncols++] = start + i; 4697 for (i=imark; i<nzB; i++) idx[ncols++] = cmap[i]; 4698 ierr = ISCreateGeneral(PETSC_COMM_SELF,ncols,idx,PETSC_OWN_POINTER,&iscola);CHKERRQ(ierr); 4699 } else { 4700 iscola = *col; 4701 } 4702 if (scall != MAT_INITIAL_MATRIX) { 4703 ierr = PetscMalloc1(1,&aloc);CHKERRQ(ierr); 4704 aloc[0] = *A_loc; 4705 } 4706 ierr = MatCreateSubMatrices(A,1,&isrowa,&iscola,scall,&aloc);CHKERRQ(ierr); 4707 *A_loc = aloc[0]; 4708 ierr = PetscFree(aloc);CHKERRQ(ierr); 4709 if (!row) { 4710 ierr = ISDestroy(&isrowa);CHKERRQ(ierr); 4711 } 4712 if (!col) { 4713 ierr = ISDestroy(&iscola);CHKERRQ(ierr); 4714 } 4715 ierr = PetscLogEventEnd(MAT_Getlocalmatcondensed,A,0,0,0);CHKERRQ(ierr); 4716 PetscFunctionReturn(0); 4717 } 4718 4719 /*@C 4720 MatGetBrowsOfAcols - Creates a SeqAIJ matrix by taking rows of B that equal to nonzero columns of local A 4721 4722 Collective on Mat 4723 4724 Input Parameters: 4725 + A,B - the matrices in mpiaij format 4726 . scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX 4727 - rowb, colb - index sets of rows and columns of B to extract (or NULL) 4728 4729 Output Parameter: 4730 + rowb, colb - index sets of rows and columns of B to extract 4731 - B_seq - the sequential matrix generated 4732 4733 Level: developer 4734 4735 @*/ 4736 PetscErrorCode MatGetBrowsOfAcols(Mat A,Mat B,MatReuse scall,IS *rowb,IS *colb,Mat *B_seq) 4737 { 4738 Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data; 4739 PetscErrorCode ierr; 4740 PetscInt *idx,i,start,ncols,nzA,nzB,*cmap,imark; 4741 IS isrowb,iscolb; 4742 Mat *bseq=NULL; 4743 4744 PetscFunctionBegin; 4745 if (A->cmap->rstart != B->rmap->rstart || A->cmap->rend != B->rmap->rend) { 4746 SETERRQ4(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Matrix local dimensions are incompatible, (%D, %D) != (%D,%D)",A->cmap->rstart,A->cmap->rend,B->rmap->rstart,B->rmap->rend); 4747 } 4748 ierr = PetscLogEventBegin(MAT_GetBrowsOfAcols,A,B,0,0);CHKERRQ(ierr); 4749 4750 if (scall == MAT_INITIAL_MATRIX) { 4751 start = A->cmap->rstart; 4752 cmap = a->garray; 4753 nzA = a->A->cmap->n; 4754 nzB = a->B->cmap->n; 4755 ierr = PetscMalloc1(nzA+nzB, &idx);CHKERRQ(ierr); 4756 ncols = 0; 4757 for (i=0; i<nzB; i++) { /* row < local row index */ 4758 if (cmap[i] < start) idx[ncols++] = cmap[i]; 4759 else break; 4760 } 4761 imark = i; 4762 for (i=0; i<nzA; i++) idx[ncols++] = start + i; /* local rows */ 4763 for (i=imark; i<nzB; i++) idx[ncols++] = cmap[i]; /* row > local row index */ 4764 ierr = ISCreateGeneral(PETSC_COMM_SELF,ncols,idx,PETSC_OWN_POINTER,&isrowb);CHKERRQ(ierr); 4765 ierr = ISCreateStride(PETSC_COMM_SELF,B->cmap->N,0,1,&iscolb);CHKERRQ(ierr); 4766 } else { 4767 if (!rowb || !colb) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"IS rowb and colb must be provided for MAT_REUSE_MATRIX"); 4768 isrowb = *rowb; iscolb = *colb; 4769 ierr = PetscMalloc1(1,&bseq);CHKERRQ(ierr); 4770 bseq[0] = *B_seq; 4771 } 4772 ierr = MatCreateSubMatrices(B,1,&isrowb,&iscolb,scall,&bseq);CHKERRQ(ierr); 4773 *B_seq = bseq[0]; 4774 ierr = PetscFree(bseq);CHKERRQ(ierr); 4775 if (!rowb) { 4776 ierr = ISDestroy(&isrowb);CHKERRQ(ierr); 4777 } else { 4778 *rowb = isrowb; 4779 } 4780 if (!colb) { 4781 ierr = ISDestroy(&iscolb);CHKERRQ(ierr); 4782 } else { 4783 *colb = iscolb; 4784 } 4785 ierr = PetscLogEventEnd(MAT_GetBrowsOfAcols,A,B,0,0);CHKERRQ(ierr); 4786 PetscFunctionReturn(0); 4787 } 4788 4789 /* 4790 MatGetBrowsOfAoCols_MPIAIJ - Creates a SeqAIJ matrix by taking rows of B that equal to nonzero columns 4791 of the OFF-DIAGONAL portion of local A 4792 4793 Collective on Mat 4794 4795 Input Parameters: 4796 + A,B - the matrices in mpiaij format 4797 - scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX 4798 4799 Output Parameter: 4800 + startsj_s - starting point in B's sending j-arrays, saved for MAT_REUSE (or NULL) 4801 . startsj_r - starting point in B's receiving j-arrays, saved for MAT_REUSE (or NULL) 4802 . bufa_ptr - array for sending matrix values, saved for MAT_REUSE (or NULL) 4803 - B_oth - the sequential matrix generated with size aBn=a->B->cmap->n by B->cmap->N 4804 4805 Level: developer 4806 4807 */ 4808 PetscErrorCode MatGetBrowsOfAoCols_MPIAIJ(Mat A,Mat B,MatReuse scall,PetscInt **startsj_s,PetscInt **startsj_r,MatScalar **bufa_ptr,Mat *B_oth) 4809 { 4810 VecScatter_MPI_General *gen_to,*gen_from; 4811 PetscErrorCode ierr; 4812 Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data; 4813 Mat_SeqAIJ *b_oth; 4814 VecScatter ctx =a->Mvctx; 4815 MPI_Comm comm; 4816 PetscMPIInt *rprocs,*sprocs,tag=((PetscObject)ctx)->tag,rank; 4817 PetscInt *rowlen,*bufj,*bufJ,ncols,aBn=a->B->cmap->n,row,*b_othi,*b_othj; 4818 PetscInt *rvalues,*svalues; 4819 MatScalar *b_otha,*bufa,*bufA; 4820 PetscInt i,j,k,l,ll,nrecvs,nsends,nrows,*srow,*rstarts,*rstartsj = 0,*sstarts,*sstartsj,len; 4821 MPI_Request *rwaits = NULL,*swaits = NULL; 4822 MPI_Status *sstatus,rstatus; 4823 PetscMPIInt jj,size; 4824 PetscInt *cols,sbs,rbs; 4825 PetscScalar *vals; 4826 4827 PetscFunctionBegin; 4828 ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 4829 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 4830 4831 if (A->cmap->rstart != B->rmap->rstart || A->cmap->rend != B->rmap->rend) { 4832 SETERRQ4(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Matrix local dimensions are incompatible, (%d, %d) != (%d,%d)",A->cmap->rstart,A->cmap->rend,B->rmap->rstart,B->rmap->rend); 4833 } 4834 ierr = PetscLogEventBegin(MAT_GetBrowsOfAocols,A,B,0,0);CHKERRQ(ierr); 4835 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 4836 4837 if (size == 1) { 4838 startsj_s = NULL; 4839 bufa_ptr = NULL; 4840 *B_oth = NULL; 4841 PetscFunctionReturn(0); 4842 } 4843 4844 gen_to = (VecScatter_MPI_General*)ctx->todata; 4845 gen_from = (VecScatter_MPI_General*)ctx->fromdata; 4846 nrecvs = gen_from->n; 4847 nsends = gen_to->n; 4848 4849 ierr = PetscMalloc2(nrecvs,&rwaits,nsends,&swaits);CHKERRQ(ierr); 4850 srow = gen_to->indices; /* local row index to be sent */ 4851 sstarts = gen_to->starts; 4852 sprocs = gen_to->procs; 4853 sstatus = gen_to->sstatus; 4854 sbs = gen_to->bs; 4855 rstarts = gen_from->starts; 4856 rprocs = gen_from->procs; 4857 rbs = gen_from->bs; 4858 4859 if (!startsj_s || !bufa_ptr) scall = MAT_INITIAL_MATRIX; 4860 if (scall == MAT_INITIAL_MATRIX) { 4861 /* i-array */ 4862 /*---------*/ 4863 /* post receives */ 4864 ierr = PetscMalloc1(rbs*(rstarts[nrecvs] - rstarts[0]),&rvalues);CHKERRQ(ierr); 4865 for (i=0; i<nrecvs; i++) { 4866 rowlen = rvalues + rstarts[i]*rbs; 4867 nrows = (rstarts[i+1]-rstarts[i])*rbs; /* num of indices to be received */ 4868 ierr = MPI_Irecv(rowlen,nrows,MPIU_INT,rprocs[i],tag,comm,rwaits+i);CHKERRQ(ierr); 4869 } 4870 4871 /* pack the outgoing message */ 4872 ierr = PetscMalloc2(nsends+1,&sstartsj,nrecvs+1,&rstartsj);CHKERRQ(ierr); 4873 4874 sstartsj[0] = 0; 4875 rstartsj[0] = 0; 4876 len = 0; /* total length of j or a array to be sent */ 4877 k = 0; 4878 ierr = PetscMalloc1(sbs*(sstarts[nsends] - sstarts[0]),&svalues);CHKERRQ(ierr); 4879 for (i=0; i<nsends; i++) { 4880 rowlen = svalues + sstarts[i]*sbs; 4881 nrows = sstarts[i+1]-sstarts[i]; /* num of block rows */ 4882 for (j=0; j<nrows; j++) { 4883 row = srow[k] + B->rmap->range[rank]; /* global row idx */ 4884 for (l=0; l<sbs; l++) { 4885 ierr = MatGetRow_MPIAIJ(B,row+l,&ncols,NULL,NULL);CHKERRQ(ierr); /* rowlength */ 4886 4887 rowlen[j*sbs+l] = ncols; 4888 4889 len += ncols; 4890 ierr = MatRestoreRow_MPIAIJ(B,row+l,&ncols,NULL,NULL);CHKERRQ(ierr); 4891 } 4892 k++; 4893 } 4894 ierr = MPI_Isend(rowlen,nrows*sbs,MPIU_INT,sprocs[i],tag,comm,swaits+i);CHKERRQ(ierr); 4895 4896 sstartsj[i+1] = len; /* starting point of (i+1)-th outgoing msg in bufj and bufa */ 4897 } 4898 /* recvs and sends of i-array are completed */ 4899 i = nrecvs; 4900 while (i--) { 4901 ierr = MPI_Waitany(nrecvs,rwaits,&jj,&rstatus);CHKERRQ(ierr); 4902 } 4903 if (nsends) {ierr = MPI_Waitall(nsends,swaits,sstatus);CHKERRQ(ierr);} 4904 ierr = PetscFree(svalues);CHKERRQ(ierr); 4905 4906 /* allocate buffers for sending j and a arrays */ 4907 ierr = PetscMalloc1(len+1,&bufj);CHKERRQ(ierr); 4908 ierr = PetscMalloc1(len+1,&bufa);CHKERRQ(ierr); 4909 4910 /* create i-array of B_oth */ 4911 ierr = PetscMalloc1(aBn+2,&b_othi);CHKERRQ(ierr); 4912 4913 b_othi[0] = 0; 4914 len = 0; /* total length of j or a array to be received */ 4915 k = 0; 4916 for (i=0; i<nrecvs; i++) { 4917 rowlen = rvalues + rstarts[i]*rbs; 4918 nrows = rbs*(rstarts[i+1]-rstarts[i]); /* num of rows to be received */ 4919 for (j=0; j<nrows; j++) { 4920 b_othi[k+1] = b_othi[k] + rowlen[j]; 4921 ierr = PetscIntSumError(rowlen[j],len,&len);CHKERRQ(ierr); 4922 k++; 4923 } 4924 rstartsj[i+1] = len; /* starting point of (i+1)-th incoming msg in bufj and bufa */ 4925 } 4926 ierr = PetscFree(rvalues);CHKERRQ(ierr); 4927 4928 /* allocate space for j and a arrrays of B_oth */ 4929 ierr = PetscMalloc1(b_othi[aBn]+1,&b_othj);CHKERRQ(ierr); 4930 ierr = PetscMalloc1(b_othi[aBn]+1,&b_otha);CHKERRQ(ierr); 4931 4932 /* j-array */ 4933 /*---------*/ 4934 /* post receives of j-array */ 4935 for (i=0; i<nrecvs; i++) { 4936 nrows = rstartsj[i+1]-rstartsj[i]; /* length of the msg received */ 4937 ierr = MPI_Irecv(b_othj+rstartsj[i],nrows,MPIU_INT,rprocs[i],tag,comm,rwaits+i);CHKERRQ(ierr); 4938 } 4939 4940 /* pack the outgoing message j-array */ 4941 k = 0; 4942 for (i=0; i<nsends; i++) { 4943 nrows = sstarts[i+1]-sstarts[i]; /* num of block rows */ 4944 bufJ = bufj+sstartsj[i]; 4945 for (j=0; j<nrows; j++) { 4946 row = srow[k++] + B->rmap->range[rank]; /* global row idx */ 4947 for (ll=0; ll<sbs; ll++) { 4948 ierr = MatGetRow_MPIAIJ(B,row+ll,&ncols,&cols,NULL);CHKERRQ(ierr); 4949 for (l=0; l<ncols; l++) { 4950 *bufJ++ = cols[l]; 4951 } 4952 ierr = MatRestoreRow_MPIAIJ(B,row+ll,&ncols,&cols,NULL);CHKERRQ(ierr); 4953 } 4954 } 4955 ierr = MPI_Isend(bufj+sstartsj[i],sstartsj[i+1]-sstartsj[i],MPIU_INT,sprocs[i],tag,comm,swaits+i);CHKERRQ(ierr); 4956 } 4957 4958 /* recvs and sends of j-array are completed */ 4959 i = nrecvs; 4960 while (i--) { 4961 ierr = MPI_Waitany(nrecvs,rwaits,&jj,&rstatus);CHKERRQ(ierr); 4962 } 4963 if (nsends) {ierr = MPI_Waitall(nsends,swaits,sstatus);CHKERRQ(ierr);} 4964 } else if (scall == MAT_REUSE_MATRIX) { 4965 sstartsj = *startsj_s; 4966 rstartsj = *startsj_r; 4967 bufa = *bufa_ptr; 4968 b_oth = (Mat_SeqAIJ*)(*B_oth)->data; 4969 b_otha = b_oth->a; 4970 } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE, "Matrix P does not posses an object container"); 4971 4972 /* a-array */ 4973 /*---------*/ 4974 /* post receives of a-array */ 4975 for (i=0; i<nrecvs; i++) { 4976 nrows = rstartsj[i+1]-rstartsj[i]; /* length of the msg received */ 4977 ierr = MPI_Irecv(b_otha+rstartsj[i],nrows,MPIU_SCALAR,rprocs[i],tag,comm,rwaits+i);CHKERRQ(ierr); 4978 } 4979 4980 /* pack the outgoing message a-array */ 4981 k = 0; 4982 for (i=0; i<nsends; i++) { 4983 nrows = sstarts[i+1]-sstarts[i]; /* num of block rows */ 4984 bufA = bufa+sstartsj[i]; 4985 for (j=0; j<nrows; j++) { 4986 row = srow[k++] + B->rmap->range[rank]; /* global row idx */ 4987 for (ll=0; ll<sbs; ll++) { 4988 ierr = MatGetRow_MPIAIJ(B,row+ll,&ncols,NULL,&vals);CHKERRQ(ierr); 4989 for (l=0; l<ncols; l++) { 4990 *bufA++ = vals[l]; 4991 } 4992 ierr = MatRestoreRow_MPIAIJ(B,row+ll,&ncols,NULL,&vals);CHKERRQ(ierr); 4993 } 4994 } 4995 ierr = MPI_Isend(bufa+sstartsj[i],sstartsj[i+1]-sstartsj[i],MPIU_SCALAR,sprocs[i],tag,comm,swaits+i);CHKERRQ(ierr); 4996 } 4997 /* recvs and sends of a-array are completed */ 4998 i = nrecvs; 4999 while (i--) { 5000 ierr = MPI_Waitany(nrecvs,rwaits,&jj,&rstatus);CHKERRQ(ierr); 5001 } 5002 if (nsends) {ierr = MPI_Waitall(nsends,swaits,sstatus);CHKERRQ(ierr);} 5003 ierr = PetscFree2(rwaits,swaits);CHKERRQ(ierr); 5004 5005 if (scall == MAT_INITIAL_MATRIX) { 5006 /* put together the new matrix */ 5007 ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,aBn,B->cmap->N,b_othi,b_othj,b_otha,B_oth);CHKERRQ(ierr); 5008 5009 /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */ 5010 /* Since these are PETSc arrays, change flags to free them as necessary. */ 5011 b_oth = (Mat_SeqAIJ*)(*B_oth)->data; 5012 b_oth->free_a = PETSC_TRUE; 5013 b_oth->free_ij = PETSC_TRUE; 5014 b_oth->nonew = 0; 5015 5016 ierr = PetscFree(bufj);CHKERRQ(ierr); 5017 if (!startsj_s || !bufa_ptr) { 5018 ierr = PetscFree2(sstartsj,rstartsj);CHKERRQ(ierr); 5019 ierr = PetscFree(bufa_ptr);CHKERRQ(ierr); 5020 } else { 5021 *startsj_s = sstartsj; 5022 *startsj_r = rstartsj; 5023 *bufa_ptr = bufa; 5024 } 5025 } 5026 ierr = PetscLogEventEnd(MAT_GetBrowsOfAocols,A,B,0,0);CHKERRQ(ierr); 5027 PetscFunctionReturn(0); 5028 } 5029 5030 /*@C 5031 MatGetCommunicationStructs - Provides access to the communication structures used in matrix-vector multiplication. 5032 5033 Not Collective 5034 5035 Input Parameters: 5036 . A - The matrix in mpiaij format 5037 5038 Output Parameter: 5039 + lvec - The local vector holding off-process values from the argument to a matrix-vector product 5040 . colmap - A map from global column index to local index into lvec 5041 - multScatter - A scatter from the argument of a matrix-vector product to lvec 5042 5043 Level: developer 5044 5045 @*/ 5046 #if defined(PETSC_USE_CTABLE) 5047 PetscErrorCode MatGetCommunicationStructs(Mat A, Vec *lvec, PetscTable *colmap, VecScatter *multScatter) 5048 #else 5049 PetscErrorCode MatGetCommunicationStructs(Mat A, Vec *lvec, PetscInt *colmap[], VecScatter *multScatter) 5050 #endif 5051 { 5052 Mat_MPIAIJ *a; 5053 5054 PetscFunctionBegin; 5055 PetscValidHeaderSpecific(A, MAT_CLASSID, 1); 5056 PetscValidPointer(lvec, 2); 5057 PetscValidPointer(colmap, 3); 5058 PetscValidPointer(multScatter, 4); 5059 a = (Mat_MPIAIJ*) A->data; 5060 if (lvec) *lvec = a->lvec; 5061 if (colmap) *colmap = a->colmap; 5062 if (multScatter) *multScatter = a->Mvctx; 5063 PetscFunctionReturn(0); 5064 } 5065 5066 PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJCRL(Mat,MatType,MatReuse,Mat*); 5067 PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJPERM(Mat,MatType,MatReuse,Mat*); 5068 PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPISBAIJ(Mat,MatType,MatReuse,Mat*); 5069 #if defined(PETSC_HAVE_ELEMENTAL) 5070 PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_Elemental(Mat,MatType,MatReuse,Mat*); 5071 #endif 5072 #if defined(PETSC_HAVE_HYPRE) 5073 PETSC_INTERN PetscErrorCode MatConvert_AIJ_HYPRE(Mat,MatType,MatReuse,Mat*); 5074 PETSC_INTERN PetscErrorCode MatMatMatMult_Transpose_AIJ_AIJ(Mat,Mat,Mat,MatReuse,PetscReal,Mat*); 5075 #endif 5076 PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_IS(Mat,MatType,MatReuse,Mat*); 5077 5078 /* 5079 Computes (B'*A')' since computing B*A directly is untenable 5080 5081 n p p 5082 ( ) ( ) ( ) 5083 m ( A ) * n ( B ) = m ( C ) 5084 ( ) ( ) ( ) 5085 5086 */ 5087 PetscErrorCode MatMatMultNumeric_MPIDense_MPIAIJ(Mat A,Mat B,Mat C) 5088 { 5089 PetscErrorCode ierr; 5090 Mat At,Bt,Ct; 5091 5092 PetscFunctionBegin; 5093 ierr = MatTranspose(A,MAT_INITIAL_MATRIX,&At);CHKERRQ(ierr); 5094 ierr = MatTranspose(B,MAT_INITIAL_MATRIX,&Bt);CHKERRQ(ierr); 5095 ierr = MatMatMult(Bt,At,MAT_INITIAL_MATRIX,1.0,&Ct);CHKERRQ(ierr); 5096 ierr = MatDestroy(&At);CHKERRQ(ierr); 5097 ierr = MatDestroy(&Bt);CHKERRQ(ierr); 5098 ierr = MatTranspose(Ct,MAT_REUSE_MATRIX,&C);CHKERRQ(ierr); 5099 ierr = MatDestroy(&Ct);CHKERRQ(ierr); 5100 PetscFunctionReturn(0); 5101 } 5102 5103 PetscErrorCode MatMatMultSymbolic_MPIDense_MPIAIJ(Mat A,Mat B,PetscReal fill,Mat *C) 5104 { 5105 PetscErrorCode ierr; 5106 PetscInt m=A->rmap->n,n=B->cmap->n; 5107 Mat Cmat; 5108 5109 PetscFunctionBegin; 5110 if (A->cmap->n != B->rmap->n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"A->cmap->n %d != B->rmap->n %d\n",A->cmap->n,B->rmap->n); 5111 ierr = MatCreate(PetscObjectComm((PetscObject)A),&Cmat);CHKERRQ(ierr); 5112 ierr = MatSetSizes(Cmat,m,n,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 5113 ierr = MatSetBlockSizesFromMats(Cmat,A,B);CHKERRQ(ierr); 5114 ierr = MatSetType(Cmat,MATMPIDENSE);CHKERRQ(ierr); 5115 ierr = MatMPIDenseSetPreallocation(Cmat,NULL);CHKERRQ(ierr); 5116 ierr = MatAssemblyBegin(Cmat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5117 ierr = MatAssemblyEnd(Cmat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5118 5119 Cmat->ops->matmultnumeric = MatMatMultNumeric_MPIDense_MPIAIJ; 5120 5121 *C = Cmat; 5122 PetscFunctionReturn(0); 5123 } 5124 5125 /* ----------------------------------------------------------------*/ 5126 PETSC_INTERN PetscErrorCode MatMatMult_MPIDense_MPIAIJ(Mat A,Mat B,MatReuse scall,PetscReal fill,Mat *C) 5127 { 5128 PetscErrorCode ierr; 5129 5130 PetscFunctionBegin; 5131 if (scall == MAT_INITIAL_MATRIX) { 5132 ierr = PetscLogEventBegin(MAT_MatMultSymbolic,A,B,0,0);CHKERRQ(ierr); 5133 ierr = MatMatMultSymbolic_MPIDense_MPIAIJ(A,B,fill,C);CHKERRQ(ierr); 5134 ierr = PetscLogEventEnd(MAT_MatMultSymbolic,A,B,0,0);CHKERRQ(ierr); 5135 } 5136 ierr = PetscLogEventBegin(MAT_MatMultNumeric,A,B,0,0);CHKERRQ(ierr); 5137 ierr = MatMatMultNumeric_MPIDense_MPIAIJ(A,B,*C);CHKERRQ(ierr); 5138 ierr = PetscLogEventEnd(MAT_MatMultNumeric,A,B,0,0);CHKERRQ(ierr); 5139 PetscFunctionReturn(0); 5140 } 5141 5142 /*MC 5143 MATMPIAIJ - MATMPIAIJ = "mpiaij" - A matrix type to be used for parallel sparse matrices. 5144 5145 Options Database Keys: 5146 . -mat_type mpiaij - sets the matrix type to "mpiaij" during a call to MatSetFromOptions() 5147 5148 Level: beginner 5149 5150 .seealso: MatCreateAIJ() 5151 M*/ 5152 5153 PETSC_EXTERN PetscErrorCode MatCreate_MPIAIJ(Mat B) 5154 { 5155 Mat_MPIAIJ *b; 5156 PetscErrorCode ierr; 5157 PetscMPIInt size; 5158 5159 PetscFunctionBegin; 5160 ierr = MPI_Comm_size(PetscObjectComm((PetscObject)B),&size);CHKERRQ(ierr); 5161 5162 ierr = PetscNewLog(B,&b);CHKERRQ(ierr); 5163 B->data = (void*)b; 5164 ierr = PetscMemcpy(B->ops,&MatOps_Values,sizeof(struct _MatOps));CHKERRQ(ierr); 5165 B->assembled = PETSC_FALSE; 5166 B->insertmode = NOT_SET_VALUES; 5167 b->size = size; 5168 5169 ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)B),&b->rank);CHKERRQ(ierr); 5170 5171 /* build cache for off array entries formed */ 5172 ierr = MatStashCreate_Private(PetscObjectComm((PetscObject)B),1,&B->stash);CHKERRQ(ierr); 5173 5174 b->donotstash = PETSC_FALSE; 5175 b->colmap = 0; 5176 b->garray = 0; 5177 b->roworiented = PETSC_TRUE; 5178 5179 /* stuff used for matrix vector multiply */ 5180 b->lvec = NULL; 5181 b->Mvctx = NULL; 5182 5183 /* stuff for MatGetRow() */ 5184 b->rowindices = 0; 5185 b->rowvalues = 0; 5186 b->getrowactive = PETSC_FALSE; 5187 5188 /* flexible pointer used in CUSP/CUSPARSE classes */ 5189 b->spptr = NULL; 5190 5191 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMPIAIJSetUseScalableIncreaseOverlap_C",MatMPIAIJSetUseScalableIncreaseOverlap_MPIAIJ);CHKERRQ(ierr); 5192 ierr = PetscObjectComposeFunction((PetscObject)B,"MatStoreValues_C",MatStoreValues_MPIAIJ);CHKERRQ(ierr); 5193 ierr = PetscObjectComposeFunction((PetscObject)B,"MatRetrieveValues_C",MatRetrieveValues_MPIAIJ);CHKERRQ(ierr); 5194 ierr = PetscObjectComposeFunction((PetscObject)B,"MatIsTranspose_C",MatIsTranspose_MPIAIJ);CHKERRQ(ierr); 5195 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMPIAIJSetPreallocation_C",MatMPIAIJSetPreallocation_MPIAIJ);CHKERRQ(ierr); 5196 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMPIAIJSetPreallocationCSR_C",MatMPIAIJSetPreallocationCSR_MPIAIJ);CHKERRQ(ierr); 5197 ierr = PetscObjectComposeFunction((PetscObject)B,"MatDiagonalScaleLocal_C",MatDiagonalScaleLocal_MPIAIJ);CHKERRQ(ierr); 5198 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_mpiaijperm_C",MatConvert_MPIAIJ_MPIAIJPERM);CHKERRQ(ierr); 5199 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_mpiaijcrl_C",MatConvert_MPIAIJ_MPIAIJCRL);CHKERRQ(ierr); 5200 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_mpisbaij_C",MatConvert_MPIAIJ_MPISBAIJ);CHKERRQ(ierr); 5201 #if defined(PETSC_HAVE_ELEMENTAL) 5202 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_elemental_C",MatConvert_MPIAIJ_Elemental);CHKERRQ(ierr); 5203 #endif 5204 #if defined(PETSC_HAVE_HYPRE) 5205 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_hypre_C",MatConvert_AIJ_HYPRE);CHKERRQ(ierr); 5206 #endif 5207 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_is_C",MatConvert_MPIAIJ_IS);CHKERRQ(ierr); 5208 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMatMult_mpidense_mpiaij_C",MatMatMult_MPIDense_MPIAIJ);CHKERRQ(ierr); 5209 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMatMultSymbolic_mpidense_mpiaij_C",MatMatMultSymbolic_MPIDense_MPIAIJ);CHKERRQ(ierr); 5210 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMatMultNumeric_mpidense_mpiaij_C",MatMatMultNumeric_MPIDense_MPIAIJ);CHKERRQ(ierr); 5211 #if defined(PETSC_HAVE_HYPRE) 5212 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMatMatMult_transpose_mpiaij_mpiaij_C",MatMatMatMult_Transpose_AIJ_AIJ);CHKERRQ(ierr); 5213 #endif 5214 ierr = PetscObjectChangeTypeName((PetscObject)B,MATMPIAIJ);CHKERRQ(ierr); 5215 PetscFunctionReturn(0); 5216 } 5217 5218 /*@C 5219 MatCreateMPIAIJWithSplitArrays - creates a MPI AIJ matrix using arrays that contain the "diagonal" 5220 and "off-diagonal" part of the matrix in CSR format. 5221 5222 Collective on MPI_Comm 5223 5224 Input Parameters: 5225 + comm - MPI communicator 5226 . m - number of local rows (Cannot be PETSC_DECIDE) 5227 . n - This value should be the same as the local size used in creating the 5228 x vector for the matrix-vector product y = Ax. (or PETSC_DECIDE to have 5229 calculated if N is given) For square matrices n is almost always m. 5230 . M - number of global rows (or PETSC_DETERMINE to have calculated if m is given) 5231 . N - number of global columns (or PETSC_DETERMINE to have calculated if n is given) 5232 . i - row indices for "diagonal" portion of matrix 5233 . j - column indices 5234 . a - matrix values 5235 . oi - row indices for "off-diagonal" portion of matrix 5236 . oj - column indices 5237 - oa - matrix values 5238 5239 Output Parameter: 5240 . mat - the matrix 5241 5242 Level: advanced 5243 5244 Notes: 5245 The i, j, and a arrays ARE NOT copied by this routine into the internal format used by PETSc. The user 5246 must free the arrays once the matrix has been destroyed and not before. 5247 5248 The i and j indices are 0 based 5249 5250 See MatCreateAIJ() for the definition of "diagonal" and "off-diagonal" portion of the matrix 5251 5252 This sets local rows and cannot be used to set off-processor values. 5253 5254 Use of this routine is discouraged because it is inflexible and cumbersome to use. It is extremely rare that a 5255 legacy application natively assembles into exactly this split format. The code to do so is nontrivial and does 5256 not easily support in-place reassembly. It is recommended to use MatSetValues() (or a variant thereof) because 5257 the resulting assembly is easier to implement, will work with any matrix format, and the user does not have to 5258 keep track of the underlying array. Use MatSetOption(A,MAT_NO_OFF_PROC_ENTRIES,PETSC_TRUE) to disable all 5259 communication if it is known that only local entries will be set. 5260 5261 .keywords: matrix, aij, compressed row, sparse, parallel 5262 5263 .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatMPIAIJSetPreallocationCSR(), 5264 MATMPIAIJ, MatCreateAIJ(), MatCreateMPIAIJWithArrays() 5265 @*/ 5266 PetscErrorCode MatCreateMPIAIJWithSplitArrays(MPI_Comm comm,PetscInt m,PetscInt n,PetscInt M,PetscInt N,PetscInt i[],PetscInt j[],PetscScalar a[],PetscInt oi[], PetscInt oj[],PetscScalar oa[],Mat *mat) 5267 { 5268 PetscErrorCode ierr; 5269 Mat_MPIAIJ *maij; 5270 5271 PetscFunctionBegin; 5272 if (m < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"local number of rows (m) cannot be PETSC_DECIDE, or negative"); 5273 if (i[0]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"i (row indices) must start with 0"); 5274 if (oi[0]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"oi (row indices) must start with 0"); 5275 ierr = MatCreate(comm,mat);CHKERRQ(ierr); 5276 ierr = MatSetSizes(*mat,m,n,M,N);CHKERRQ(ierr); 5277 ierr = MatSetType(*mat,MATMPIAIJ);CHKERRQ(ierr); 5278 maij = (Mat_MPIAIJ*) (*mat)->data; 5279 5280 (*mat)->preallocated = PETSC_TRUE; 5281 5282 ierr = PetscLayoutSetUp((*mat)->rmap);CHKERRQ(ierr); 5283 ierr = PetscLayoutSetUp((*mat)->cmap);CHKERRQ(ierr); 5284 5285 ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,m,n,i,j,a,&maij->A);CHKERRQ(ierr); 5286 ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,m,(*mat)->cmap->N,oi,oj,oa,&maij->B);CHKERRQ(ierr); 5287 5288 ierr = MatAssemblyBegin(maij->A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5289 ierr = MatAssemblyEnd(maij->A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5290 ierr = MatAssemblyBegin(maij->B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5291 ierr = MatAssemblyEnd(maij->B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5292 5293 ierr = MatSetOption(*mat,MAT_NO_OFF_PROC_ENTRIES,PETSC_TRUE);CHKERRQ(ierr); 5294 ierr = MatAssemblyBegin(*mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5295 ierr = MatAssemblyEnd(*mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5296 ierr = MatSetOption(*mat,MAT_NO_OFF_PROC_ENTRIES,PETSC_FALSE);CHKERRQ(ierr); 5297 ierr = MatSetOption(*mat,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE);CHKERRQ(ierr); 5298 PetscFunctionReturn(0); 5299 } 5300 5301 /* 5302 Special version for direct calls from Fortran 5303 */ 5304 #include <petsc/private/fortranimpl.h> 5305 5306 /* Change these macros so can be used in void function */ 5307 #undef CHKERRQ 5308 #define CHKERRQ(ierr) CHKERRABORT(PETSC_COMM_WORLD,ierr) 5309 #undef SETERRQ2 5310 #define SETERRQ2(comm,ierr,b,c,d) CHKERRABORT(comm,ierr) 5311 #undef SETERRQ3 5312 #define SETERRQ3(comm,ierr,b,c,d,e) CHKERRABORT(comm,ierr) 5313 #undef SETERRQ 5314 #define SETERRQ(c,ierr,b) CHKERRABORT(c,ierr) 5315 5316 #if defined(PETSC_HAVE_FORTRAN_CAPS) 5317 #define matsetvaluesmpiaij_ MATSETVALUESMPIAIJ 5318 #elif !defined(PETSC_HAVE_FORTRAN_UNDERSCORE) 5319 #define matsetvaluesmpiaij_ matsetvaluesmpiaij 5320 #else 5321 #endif 5322 PETSC_EXTERN void PETSC_STDCALL matsetvaluesmpiaij_(Mat *mmat,PetscInt *mm,const PetscInt im[],PetscInt *mn,const PetscInt in[],const PetscScalar v[],InsertMode *maddv,PetscErrorCode *_ierr) 5323 { 5324 Mat mat = *mmat; 5325 PetscInt m = *mm, n = *mn; 5326 InsertMode addv = *maddv; 5327 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 5328 PetscScalar value; 5329 PetscErrorCode ierr; 5330 5331 MatCheckPreallocated(mat,1); 5332 if (mat->insertmode == NOT_SET_VALUES) mat->insertmode = addv; 5333 5334 #if defined(PETSC_USE_DEBUG) 5335 else if (mat->insertmode != addv) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Cannot mix add values and insert values"); 5336 #endif 5337 { 5338 PetscInt i,j,rstart = mat->rmap->rstart,rend = mat->rmap->rend; 5339 PetscInt cstart = mat->cmap->rstart,cend = mat->cmap->rend,row,col; 5340 PetscBool roworiented = aij->roworiented; 5341 5342 /* Some Variables required in the macro */ 5343 Mat A = aij->A; 5344 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 5345 PetscInt *aimax = a->imax,*ai = a->i,*ailen = a->ilen,*aj = a->j; 5346 MatScalar *aa = a->a; 5347 PetscBool ignorezeroentries = (((a->ignorezeroentries)&&(addv==ADD_VALUES)) ? PETSC_TRUE : PETSC_FALSE); 5348 Mat B = aij->B; 5349 Mat_SeqAIJ *b = (Mat_SeqAIJ*)B->data; 5350 PetscInt *bimax = b->imax,*bi = b->i,*bilen = b->ilen,*bj = b->j,bm = aij->B->rmap->n,am = aij->A->rmap->n; 5351 MatScalar *ba = b->a; 5352 5353 PetscInt *rp1,*rp2,ii,nrow1,nrow2,_i,rmax1,rmax2,N,low1,high1,low2,high2,t,lastcol1,lastcol2; 5354 PetscInt nonew = a->nonew; 5355 MatScalar *ap1,*ap2; 5356 5357 PetscFunctionBegin; 5358 for (i=0; i<m; i++) { 5359 if (im[i] < 0) continue; 5360 #if defined(PETSC_USE_DEBUG) 5361 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); 5362 #endif 5363 if (im[i] >= rstart && im[i] < rend) { 5364 row = im[i] - rstart; 5365 lastcol1 = -1; 5366 rp1 = aj + ai[row]; 5367 ap1 = aa + ai[row]; 5368 rmax1 = aimax[row]; 5369 nrow1 = ailen[row]; 5370 low1 = 0; 5371 high1 = nrow1; 5372 lastcol2 = -1; 5373 rp2 = bj + bi[row]; 5374 ap2 = ba + bi[row]; 5375 rmax2 = bimax[row]; 5376 nrow2 = bilen[row]; 5377 low2 = 0; 5378 high2 = nrow2; 5379 5380 for (j=0; j<n; j++) { 5381 if (roworiented) value = v[i*n+j]; 5382 else value = v[i+j*m]; 5383 if (in[j] >= cstart && in[j] < cend) { 5384 col = in[j] - cstart; 5385 if (ignorezeroentries && value == 0.0 && (addv == ADD_VALUES) && row != col) continue; 5386 MatSetValues_SeqAIJ_A_Private(row,col,value,addv,im[i],in[j]); 5387 } else if (in[j] < 0) continue; 5388 #if defined(PETSC_USE_DEBUG) 5389 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); 5390 #endif 5391 else { 5392 if (mat->was_assembled) { 5393 if (!aij->colmap) { 5394 ierr = MatCreateColmap_MPIAIJ_Private(mat);CHKERRQ(ierr); 5395 } 5396 #if defined(PETSC_USE_CTABLE) 5397 ierr = PetscTableFind(aij->colmap,in[j]+1,&col);CHKERRQ(ierr); 5398 col--; 5399 #else 5400 col = aij->colmap[in[j]] - 1; 5401 #endif 5402 if (ignorezeroentries && value == 0.0 && (addv == ADD_VALUES) && row != col) continue; 5403 if (col < 0 && !((Mat_SeqAIJ*)(aij->A->data))->nonew) { 5404 ierr = MatDisAssemble_MPIAIJ(mat);CHKERRQ(ierr); 5405 col = in[j]; 5406 /* Reinitialize the variables required by MatSetValues_SeqAIJ_B_Private() */ 5407 B = aij->B; 5408 b = (Mat_SeqAIJ*)B->data; 5409 bimax = b->imax; bi = b->i; bilen = b->ilen; bj = b->j; 5410 rp2 = bj + bi[row]; 5411 ap2 = ba + bi[row]; 5412 rmax2 = bimax[row]; 5413 nrow2 = bilen[row]; 5414 low2 = 0; 5415 high2 = nrow2; 5416 bm = aij->B->rmap->n; 5417 ba = b->a; 5418 } 5419 } else col = in[j]; 5420 MatSetValues_SeqAIJ_B_Private(row,col,value,addv,im[i],in[j]); 5421 } 5422 } 5423 } else if (!aij->donotstash) { 5424 if (roworiented) { 5425 ierr = MatStashValuesRow_Private(&mat->stash,im[i],n,in,v+i*n,(PetscBool)(ignorezeroentries && (addv == ADD_VALUES)));CHKERRQ(ierr); 5426 } else { 5427 ierr = MatStashValuesCol_Private(&mat->stash,im[i],n,in,v+i,m,(PetscBool)(ignorezeroentries && (addv == ADD_VALUES)));CHKERRQ(ierr); 5428 } 5429 } 5430 } 5431 } 5432 PetscFunctionReturnVoid(); 5433 } 5434 5435