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, MATAIJMKL, 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 PetscErrorCode MatFindZeroDiagonals_MPIAIJ(Mat M,IS *zrows) 135 { 136 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)M->data; 137 PetscErrorCode ierr; 138 PetscInt i,rstart,nrows,*rows; 139 140 PetscFunctionBegin; 141 *zrows = NULL; 142 ierr = MatFindZeroDiagonals_SeqAIJ_Private(aij->A,&nrows,&rows);CHKERRQ(ierr); 143 ierr = MatGetOwnershipRange(M,&rstart,NULL);CHKERRQ(ierr); 144 for (i=0; i<nrows; i++) rows[i] += rstart; 145 ierr = ISCreateGeneral(PetscObjectComm((PetscObject)M),nrows,rows,PETSC_OWN_POINTER,zrows);CHKERRQ(ierr); 146 PetscFunctionReturn(0); 147 } 148 149 PetscErrorCode MatGetColumnNorms_MPIAIJ(Mat A,NormType type,PetscReal *norms) 150 { 151 PetscErrorCode ierr; 152 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)A->data; 153 PetscInt i,n,*garray = aij->garray; 154 Mat_SeqAIJ *a_aij = (Mat_SeqAIJ*) aij->A->data; 155 Mat_SeqAIJ *b_aij = (Mat_SeqAIJ*) aij->B->data; 156 PetscReal *work; 157 158 PetscFunctionBegin; 159 ierr = MatGetSize(A,NULL,&n);CHKERRQ(ierr); 160 ierr = PetscCalloc1(n,&work);CHKERRQ(ierr); 161 if (type == NORM_2) { 162 for (i=0; i<a_aij->i[aij->A->rmap->n]; i++) { 163 work[A->cmap->rstart + a_aij->j[i]] += PetscAbsScalar(a_aij->a[i]*a_aij->a[i]); 164 } 165 for (i=0; i<b_aij->i[aij->B->rmap->n]; i++) { 166 work[garray[b_aij->j[i]]] += PetscAbsScalar(b_aij->a[i]*b_aij->a[i]); 167 } 168 } else if (type == NORM_1) { 169 for (i=0; i<a_aij->i[aij->A->rmap->n]; i++) { 170 work[A->cmap->rstart + a_aij->j[i]] += PetscAbsScalar(a_aij->a[i]); 171 } 172 for (i=0; i<b_aij->i[aij->B->rmap->n]; i++) { 173 work[garray[b_aij->j[i]]] += PetscAbsScalar(b_aij->a[i]); 174 } 175 } else if (type == NORM_INFINITY) { 176 for (i=0; i<a_aij->i[aij->A->rmap->n]; i++) { 177 work[A->cmap->rstart + a_aij->j[i]] = PetscMax(PetscAbsScalar(a_aij->a[i]), work[A->cmap->rstart + a_aij->j[i]]); 178 } 179 for (i=0; i<b_aij->i[aij->B->rmap->n]; i++) { 180 work[garray[b_aij->j[i]]] = PetscMax(PetscAbsScalar(b_aij->a[i]),work[garray[b_aij->j[i]]]); 181 } 182 183 } else SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_WRONG,"Unknown NormType"); 184 if (type == NORM_INFINITY) { 185 ierr = MPIU_Allreduce(work,norms,n,MPIU_REAL,MPIU_MAX,PetscObjectComm((PetscObject)A));CHKERRQ(ierr); 186 } else { 187 ierr = MPIU_Allreduce(work,norms,n,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)A));CHKERRQ(ierr); 188 } 189 ierr = PetscFree(work);CHKERRQ(ierr); 190 if (type == NORM_2) { 191 for (i=0; i<n; i++) norms[i] = PetscSqrtReal(norms[i]); 192 } 193 PetscFunctionReturn(0); 194 } 195 196 PetscErrorCode MatFindOffBlockDiagonalEntries_MPIAIJ(Mat A,IS *is) 197 { 198 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 199 IS sis,gis; 200 PetscErrorCode ierr; 201 const PetscInt *isis,*igis; 202 PetscInt n,*iis,nsis,ngis,rstart,i; 203 204 PetscFunctionBegin; 205 ierr = MatFindOffBlockDiagonalEntries(a->A,&sis);CHKERRQ(ierr); 206 ierr = MatFindNonzeroRows(a->B,&gis);CHKERRQ(ierr); 207 ierr = ISGetSize(gis,&ngis);CHKERRQ(ierr); 208 ierr = ISGetSize(sis,&nsis);CHKERRQ(ierr); 209 ierr = ISGetIndices(sis,&isis);CHKERRQ(ierr); 210 ierr = ISGetIndices(gis,&igis);CHKERRQ(ierr); 211 212 ierr = PetscMalloc1(ngis+nsis,&iis);CHKERRQ(ierr); 213 ierr = PetscMemcpy(iis,igis,ngis*sizeof(PetscInt));CHKERRQ(ierr); 214 ierr = PetscMemcpy(iis+ngis,isis,nsis*sizeof(PetscInt));CHKERRQ(ierr); 215 n = ngis + nsis; 216 ierr = PetscSortRemoveDupsInt(&n,iis);CHKERRQ(ierr); 217 ierr = MatGetOwnershipRange(A,&rstart,NULL);CHKERRQ(ierr); 218 for (i=0; i<n; i++) iis[i] += rstart; 219 ierr = ISCreateGeneral(PetscObjectComm((PetscObject)A),n,iis,PETSC_OWN_POINTER,is);CHKERRQ(ierr); 220 221 ierr = ISRestoreIndices(sis,&isis);CHKERRQ(ierr); 222 ierr = ISRestoreIndices(gis,&igis);CHKERRQ(ierr); 223 ierr = ISDestroy(&sis);CHKERRQ(ierr); 224 ierr = ISDestroy(&gis);CHKERRQ(ierr); 225 PetscFunctionReturn(0); 226 } 227 228 /* 229 Distributes a SeqAIJ matrix across a set of processes. Code stolen from 230 MatLoad_MPIAIJ(). Horrible lack of reuse. Should be a routine for each matrix type. 231 232 Only for square matrices 233 234 Used by a preconditioner, hence PETSC_EXTERN 235 */ 236 PETSC_EXTERN PetscErrorCode MatDistribute_MPIAIJ(MPI_Comm comm,Mat gmat,PetscInt m,MatReuse reuse,Mat *inmat) 237 { 238 PetscMPIInt rank,size; 239 PetscInt *rowners,*dlens,*olens,i,rstart,rend,j,jj,nz = 0,*gmataj,cnt,row,*ld,bses[2]; 240 PetscErrorCode ierr; 241 Mat mat; 242 Mat_SeqAIJ *gmata; 243 PetscMPIInt tag; 244 MPI_Status status; 245 PetscBool aij; 246 MatScalar *gmataa,*ao,*ad,*gmataarestore=0; 247 248 PetscFunctionBegin; 249 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 250 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 251 if (!rank) { 252 ierr = PetscObjectTypeCompare((PetscObject)gmat,MATSEQAIJ,&aij);CHKERRQ(ierr); 253 if (!aij) SETERRQ1(PetscObjectComm((PetscObject)gmat),PETSC_ERR_SUP,"Currently no support for input matrix of type %s\n",((PetscObject)gmat)->type_name); 254 } 255 if (reuse == MAT_INITIAL_MATRIX) { 256 ierr = MatCreate(comm,&mat);CHKERRQ(ierr); 257 ierr = MatSetSizes(mat,m,m,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 258 ierr = MatGetBlockSizes(gmat,&bses[0],&bses[1]);CHKERRQ(ierr); 259 ierr = MPI_Bcast(bses,2,MPIU_INT,0,comm);CHKERRQ(ierr); 260 ierr = MatSetBlockSizes(mat,bses[0],bses[1]);CHKERRQ(ierr); 261 ierr = MatSetType(mat,MATAIJ);CHKERRQ(ierr); 262 ierr = PetscMalloc1(size+1,&rowners);CHKERRQ(ierr); 263 ierr = PetscMalloc2(m,&dlens,m,&olens);CHKERRQ(ierr); 264 ierr = MPI_Allgather(&m,1,MPIU_INT,rowners+1,1,MPIU_INT,comm);CHKERRQ(ierr); 265 266 rowners[0] = 0; 267 for (i=2; i<=size; i++) rowners[i] += rowners[i-1]; 268 rstart = rowners[rank]; 269 rend = rowners[rank+1]; 270 ierr = PetscObjectGetNewTag((PetscObject)mat,&tag);CHKERRQ(ierr); 271 if (!rank) { 272 gmata = (Mat_SeqAIJ*) gmat->data; 273 /* send row lengths to all processors */ 274 for (i=0; i<m; i++) dlens[i] = gmata->ilen[i]; 275 for (i=1; i<size; i++) { 276 ierr = MPI_Send(gmata->ilen + rowners[i],rowners[i+1]-rowners[i],MPIU_INT,i,tag,comm);CHKERRQ(ierr); 277 } 278 /* determine number diagonal and off-diagonal counts */ 279 ierr = PetscMemzero(olens,m*sizeof(PetscInt));CHKERRQ(ierr); 280 ierr = PetscCalloc1(m,&ld);CHKERRQ(ierr); 281 jj = 0; 282 for (i=0; i<m; i++) { 283 for (j=0; j<dlens[i]; j++) { 284 if (gmata->j[jj] < rstart) ld[i]++; 285 if (gmata->j[jj] < rstart || gmata->j[jj] >= rend) olens[i]++; 286 jj++; 287 } 288 } 289 /* send column indices to other processes */ 290 for (i=1; i<size; i++) { 291 nz = gmata->i[rowners[i+1]]-gmata->i[rowners[i]]; 292 ierr = MPI_Send(&nz,1,MPIU_INT,i,tag,comm);CHKERRQ(ierr); 293 ierr = MPI_Send(gmata->j + gmata->i[rowners[i]],nz,MPIU_INT,i,tag,comm);CHKERRQ(ierr); 294 } 295 296 /* send numerical values to other processes */ 297 for (i=1; i<size; i++) { 298 nz = gmata->i[rowners[i+1]]-gmata->i[rowners[i]]; 299 ierr = MPI_Send(gmata->a + gmata->i[rowners[i]],nz,MPIU_SCALAR,i,tag,comm);CHKERRQ(ierr); 300 } 301 gmataa = gmata->a; 302 gmataj = gmata->j; 303 304 } else { 305 /* receive row lengths */ 306 ierr = MPI_Recv(dlens,m,MPIU_INT,0,tag,comm,&status);CHKERRQ(ierr); 307 /* receive column indices */ 308 ierr = MPI_Recv(&nz,1,MPIU_INT,0,tag,comm,&status);CHKERRQ(ierr); 309 ierr = PetscMalloc2(nz,&gmataa,nz,&gmataj);CHKERRQ(ierr); 310 ierr = MPI_Recv(gmataj,nz,MPIU_INT,0,tag,comm,&status);CHKERRQ(ierr); 311 /* determine number diagonal and off-diagonal counts */ 312 ierr = PetscMemzero(olens,m*sizeof(PetscInt));CHKERRQ(ierr); 313 ierr = PetscCalloc1(m,&ld);CHKERRQ(ierr); 314 jj = 0; 315 for (i=0; i<m; i++) { 316 for (j=0; j<dlens[i]; j++) { 317 if (gmataj[jj] < rstart) ld[i]++; 318 if (gmataj[jj] < rstart || gmataj[jj] >= rend) olens[i]++; 319 jj++; 320 } 321 } 322 /* receive numerical values */ 323 ierr = PetscMemzero(gmataa,nz*sizeof(PetscScalar));CHKERRQ(ierr); 324 ierr = MPI_Recv(gmataa,nz,MPIU_SCALAR,0,tag,comm,&status);CHKERRQ(ierr); 325 } 326 /* set preallocation */ 327 for (i=0; i<m; i++) { 328 dlens[i] -= olens[i]; 329 } 330 ierr = MatSeqAIJSetPreallocation(mat,0,dlens);CHKERRQ(ierr); 331 ierr = MatMPIAIJSetPreallocation(mat,0,dlens,0,olens);CHKERRQ(ierr); 332 333 for (i=0; i<m; i++) { 334 dlens[i] += olens[i]; 335 } 336 cnt = 0; 337 for (i=0; i<m; i++) { 338 row = rstart + i; 339 ierr = MatSetValues(mat,1,&row,dlens[i],gmataj+cnt,gmataa+cnt,INSERT_VALUES);CHKERRQ(ierr); 340 cnt += dlens[i]; 341 } 342 if (rank) { 343 ierr = PetscFree2(gmataa,gmataj);CHKERRQ(ierr); 344 } 345 ierr = PetscFree2(dlens,olens);CHKERRQ(ierr); 346 ierr = PetscFree(rowners);CHKERRQ(ierr); 347 348 ((Mat_MPIAIJ*)(mat->data))->ld = ld; 349 350 *inmat = mat; 351 } else { /* column indices are already set; only need to move over numerical values from process 0 */ 352 Mat_SeqAIJ *Ad = (Mat_SeqAIJ*)((Mat_MPIAIJ*)((*inmat)->data))->A->data; 353 Mat_SeqAIJ *Ao = (Mat_SeqAIJ*)((Mat_MPIAIJ*)((*inmat)->data))->B->data; 354 mat = *inmat; 355 ierr = PetscObjectGetNewTag((PetscObject)mat,&tag);CHKERRQ(ierr); 356 if (!rank) { 357 /* send numerical values to other processes */ 358 gmata = (Mat_SeqAIJ*) gmat->data; 359 ierr = MatGetOwnershipRanges(mat,(const PetscInt**)&rowners);CHKERRQ(ierr); 360 gmataa = gmata->a; 361 for (i=1; i<size; i++) { 362 nz = gmata->i[rowners[i+1]]-gmata->i[rowners[i]]; 363 ierr = MPI_Send(gmataa + gmata->i[rowners[i]],nz,MPIU_SCALAR,i,tag,comm);CHKERRQ(ierr); 364 } 365 nz = gmata->i[rowners[1]]-gmata->i[rowners[0]]; 366 } else { 367 /* receive numerical values from process 0*/ 368 nz = Ad->nz + Ao->nz; 369 ierr = PetscMalloc1(nz,&gmataa);CHKERRQ(ierr); gmataarestore = gmataa; 370 ierr = MPI_Recv(gmataa,nz,MPIU_SCALAR,0,tag,comm,&status);CHKERRQ(ierr); 371 } 372 /* transfer numerical values into the diagonal A and off diagonal B parts of mat */ 373 ld = ((Mat_MPIAIJ*)(mat->data))->ld; 374 ad = Ad->a; 375 ao = Ao->a; 376 if (mat->rmap->n) { 377 i = 0; 378 nz = ld[i]; ierr = PetscMemcpy(ao,gmataa,nz*sizeof(PetscScalar));CHKERRQ(ierr); ao += nz; gmataa += nz; 379 nz = Ad->i[i+1] - Ad->i[i]; ierr = PetscMemcpy(ad,gmataa,nz*sizeof(PetscScalar));CHKERRQ(ierr); ad += nz; gmataa += nz; 380 } 381 for (i=1; i<mat->rmap->n; i++) { 382 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; 383 nz = Ad->i[i+1] - Ad->i[i]; ierr = PetscMemcpy(ad,gmataa,nz*sizeof(PetscScalar));CHKERRQ(ierr); ad += nz; gmataa += nz; 384 } 385 i--; 386 if (mat->rmap->n) { 387 nz = Ao->i[i+1] - Ao->i[i] - ld[i]; ierr = PetscMemcpy(ao,gmataa,nz*sizeof(PetscScalar));CHKERRQ(ierr); 388 } 389 if (rank) { 390 ierr = PetscFree(gmataarestore);CHKERRQ(ierr); 391 } 392 } 393 ierr = MatAssemblyBegin(mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 394 ierr = MatAssemblyEnd(mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 395 PetscFunctionReturn(0); 396 } 397 398 /* 399 Local utility routine that creates a mapping from the global column 400 number to the local number in the off-diagonal part of the local 401 storage of the matrix. When PETSC_USE_CTABLE is used this is scalable at 402 a slightly higher hash table cost; without it it is not scalable (each processor 403 has an order N integer array but is fast to acess. 404 */ 405 PetscErrorCode MatCreateColmap_MPIAIJ_Private(Mat mat) 406 { 407 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 408 PetscErrorCode ierr; 409 PetscInt n = aij->B->cmap->n,i; 410 411 PetscFunctionBegin; 412 if (!aij->garray) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"MPIAIJ Matrix was assembled but is missing garray"); 413 #if defined(PETSC_USE_CTABLE) 414 ierr = PetscTableCreate(n,mat->cmap->N+1,&aij->colmap);CHKERRQ(ierr); 415 for (i=0; i<n; i++) { 416 ierr = PetscTableAdd(aij->colmap,aij->garray[i]+1,i+1,INSERT_VALUES);CHKERRQ(ierr); 417 } 418 #else 419 ierr = PetscCalloc1(mat->cmap->N+1,&aij->colmap);CHKERRQ(ierr); 420 ierr = PetscLogObjectMemory((PetscObject)mat,(mat->cmap->N+1)*sizeof(PetscInt));CHKERRQ(ierr); 421 for (i=0; i<n; i++) aij->colmap[aij->garray[i]] = i+1; 422 #endif 423 PetscFunctionReturn(0); 424 } 425 426 #define MatSetValues_SeqAIJ_A_Private(row,col,value,addv,orow,ocol) \ 427 { \ 428 if (col <= lastcol1) low1 = 0; \ 429 else high1 = nrow1; \ 430 lastcol1 = col;\ 431 while (high1-low1 > 5) { \ 432 t = (low1+high1)/2; \ 433 if (rp1[t] > col) high1 = t; \ 434 else low1 = t; \ 435 } \ 436 for (_i=low1; _i<high1; _i++) { \ 437 if (rp1[_i] > col) break; \ 438 if (rp1[_i] == col) { \ 439 if (addv == ADD_VALUES) ap1[_i] += value; \ 440 else ap1[_i] = value; \ 441 goto a_noinsert; \ 442 } \ 443 } \ 444 if (value == 0.0 && ignorezeroentries && row != col) {low1 = 0; high1 = nrow1;goto a_noinsert;} \ 445 if (nonew == 1) {low1 = 0; high1 = nrow1; goto a_noinsert;} \ 446 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); \ 447 MatSeqXAIJReallocateAIJ(A,am,1,nrow1,row,col,rmax1,aa,ai,aj,rp1,ap1,aimax,nonew,MatScalar); \ 448 N = nrow1++ - 1; a->nz++; high1++; \ 449 /* shift up all the later entries in this row */ \ 450 for (ii=N; ii>=_i; ii--) { \ 451 rp1[ii+1] = rp1[ii]; \ 452 ap1[ii+1] = ap1[ii]; \ 453 } \ 454 rp1[_i] = col; \ 455 ap1[_i] = value; \ 456 A->nonzerostate++;\ 457 a_noinsert: ; \ 458 ailen[row] = nrow1; \ 459 } 460 461 #define MatSetValues_SeqAIJ_B_Private(row,col,value,addv,orow,ocol) \ 462 { \ 463 if (col <= lastcol2) low2 = 0; \ 464 else high2 = nrow2; \ 465 lastcol2 = col; \ 466 while (high2-low2 > 5) { \ 467 t = (low2+high2)/2; \ 468 if (rp2[t] > col) high2 = t; \ 469 else low2 = t; \ 470 } \ 471 for (_i=low2; _i<high2; _i++) { \ 472 if (rp2[_i] > col) break; \ 473 if (rp2[_i] == col) { \ 474 if (addv == ADD_VALUES) ap2[_i] += value; \ 475 else ap2[_i] = value; \ 476 goto b_noinsert; \ 477 } \ 478 } \ 479 if (value == 0.0 && ignorezeroentries) {low2 = 0; high2 = nrow2; goto b_noinsert;} \ 480 if (nonew == 1) {low2 = 0; high2 = nrow2; goto b_noinsert;} \ 481 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); \ 482 MatSeqXAIJReallocateAIJ(B,bm,1,nrow2,row,col,rmax2,ba,bi,bj,rp2,ap2,bimax,nonew,MatScalar); \ 483 N = nrow2++ - 1; b->nz++; high2++; \ 484 /* shift up all the later entries in this row */ \ 485 for (ii=N; ii>=_i; ii--) { \ 486 rp2[ii+1] = rp2[ii]; \ 487 ap2[ii+1] = ap2[ii]; \ 488 } \ 489 rp2[_i] = col; \ 490 ap2[_i] = value; \ 491 B->nonzerostate++; \ 492 b_noinsert: ; \ 493 bilen[row] = nrow2; \ 494 } 495 496 PetscErrorCode MatSetValuesRow_MPIAIJ(Mat A,PetscInt row,const PetscScalar v[]) 497 { 498 Mat_MPIAIJ *mat = (Mat_MPIAIJ*)A->data; 499 Mat_SeqAIJ *a = (Mat_SeqAIJ*)mat->A->data,*b = (Mat_SeqAIJ*)mat->B->data; 500 PetscErrorCode ierr; 501 PetscInt l,*garray = mat->garray,diag; 502 503 PetscFunctionBegin; 504 /* code only works for square matrices A */ 505 506 /* find size of row to the left of the diagonal part */ 507 ierr = MatGetOwnershipRange(A,&diag,0);CHKERRQ(ierr); 508 row = row - diag; 509 for (l=0; l<b->i[row+1]-b->i[row]; l++) { 510 if (garray[b->j[b->i[row]+l]] > diag) break; 511 } 512 ierr = PetscMemcpy(b->a+b->i[row],v,l*sizeof(PetscScalar));CHKERRQ(ierr); 513 514 /* diagonal part */ 515 ierr = PetscMemcpy(a->a+a->i[row],v+l,(a->i[row+1]-a->i[row])*sizeof(PetscScalar));CHKERRQ(ierr); 516 517 /* right of diagonal part */ 518 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); 519 PetscFunctionReturn(0); 520 } 521 522 PetscErrorCode MatSetValues_MPIAIJ(Mat mat,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],const PetscScalar v[],InsertMode addv) 523 { 524 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 525 PetscScalar value; 526 PetscErrorCode ierr; 527 PetscInt i,j,rstart = mat->rmap->rstart,rend = mat->rmap->rend; 528 PetscInt cstart = mat->cmap->rstart,cend = mat->cmap->rend,row,col; 529 PetscBool roworiented = aij->roworiented; 530 531 /* Some Variables required in the macro */ 532 Mat A = aij->A; 533 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 534 PetscInt *aimax = a->imax,*ai = a->i,*ailen = a->ilen,*aj = a->j; 535 MatScalar *aa = a->a; 536 PetscBool ignorezeroentries = a->ignorezeroentries; 537 Mat B = aij->B; 538 Mat_SeqAIJ *b = (Mat_SeqAIJ*)B->data; 539 PetscInt *bimax = b->imax,*bi = b->i,*bilen = b->ilen,*bj = b->j,bm = aij->B->rmap->n,am = aij->A->rmap->n; 540 MatScalar *ba = b->a; 541 542 PetscInt *rp1,*rp2,ii,nrow1,nrow2,_i,rmax1,rmax2,N,low1,high1,low2,high2,t,lastcol1,lastcol2; 543 PetscInt nonew; 544 MatScalar *ap1,*ap2; 545 546 PetscFunctionBegin; 547 for (i=0; i<m; i++) { 548 if (im[i] < 0) continue; 549 #if defined(PETSC_USE_DEBUG) 550 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); 551 #endif 552 if (im[i] >= rstart && im[i] < rend) { 553 row = im[i] - rstart; 554 lastcol1 = -1; 555 rp1 = aj + ai[row]; 556 ap1 = aa + ai[row]; 557 rmax1 = aimax[row]; 558 nrow1 = ailen[row]; 559 low1 = 0; 560 high1 = nrow1; 561 lastcol2 = -1; 562 rp2 = bj + bi[row]; 563 ap2 = ba + bi[row]; 564 rmax2 = bimax[row]; 565 nrow2 = bilen[row]; 566 low2 = 0; 567 high2 = nrow2; 568 569 for (j=0; j<n; j++) { 570 if (roworiented) value = v[i*n+j]; 571 else value = v[i+j*m]; 572 if (in[j] >= cstart && in[j] < cend) { 573 col = in[j] - cstart; 574 nonew = a->nonew; 575 if (ignorezeroentries && value == 0.0 && (addv == ADD_VALUES) && row != col) continue; 576 MatSetValues_SeqAIJ_A_Private(row,col,value,addv,im[i],in[j]); 577 } else if (in[j] < 0) continue; 578 #if defined(PETSC_USE_DEBUG) 579 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); 580 #endif 581 else { 582 if (mat->was_assembled) { 583 if (!aij->colmap) { 584 ierr = MatCreateColmap_MPIAIJ_Private(mat);CHKERRQ(ierr); 585 } 586 #if defined(PETSC_USE_CTABLE) 587 ierr = PetscTableFind(aij->colmap,in[j]+1,&col);CHKERRQ(ierr); 588 col--; 589 #else 590 col = aij->colmap[in[j]] - 1; 591 #endif 592 if (col < 0 && !((Mat_SeqAIJ*)(aij->B->data))->nonew) { 593 ierr = MatDisAssemble_MPIAIJ(mat);CHKERRQ(ierr); 594 col = in[j]; 595 /* Reinitialize the variables required by MatSetValues_SeqAIJ_B_Private() */ 596 B = aij->B; 597 b = (Mat_SeqAIJ*)B->data; 598 bimax = b->imax; bi = b->i; bilen = b->ilen; bj = b->j; ba = b->a; 599 rp2 = bj + bi[row]; 600 ap2 = ba + bi[row]; 601 rmax2 = bimax[row]; 602 nrow2 = bilen[row]; 603 low2 = 0; 604 high2 = nrow2; 605 bm = aij->B->rmap->n; 606 ba = b->a; 607 } else if (col < 0) { 608 if (1 == ((Mat_SeqAIJ*)(aij->B->data))->nonew) { 609 ierr = PetscInfo3(mat,"Skipping of insertion of new nonzero location in off-diagonal portion of matrix %g(%D,%D)\n",(double)PetscRealPart(value),im[i],in[j]);CHKERRQ(ierr); 610 } else SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero at global row/column (%D, %D) into matrix", im[i], in[j]); 611 } 612 } else col = in[j]; 613 nonew = b->nonew; 614 MatSetValues_SeqAIJ_B_Private(row,col,value,addv,im[i],in[j]); 615 } 616 } 617 } else { 618 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]); 619 if (!aij->donotstash) { 620 mat->assembled = PETSC_FALSE; 621 if (roworiented) { 622 ierr = MatStashValuesRow_Private(&mat->stash,im[i],n,in,v+i*n,(PetscBool)(ignorezeroentries && (addv == ADD_VALUES)));CHKERRQ(ierr); 623 } else { 624 ierr = MatStashValuesCol_Private(&mat->stash,im[i],n,in,v+i,m,(PetscBool)(ignorezeroentries && (addv == ADD_VALUES)));CHKERRQ(ierr); 625 } 626 } 627 } 628 } 629 PetscFunctionReturn(0); 630 } 631 632 PetscErrorCode MatGetValues_MPIAIJ(Mat mat,PetscInt m,const PetscInt idxm[],PetscInt n,const PetscInt idxn[],PetscScalar v[]) 633 { 634 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 635 PetscErrorCode ierr; 636 PetscInt i,j,rstart = mat->rmap->rstart,rend = mat->rmap->rend; 637 PetscInt cstart = mat->cmap->rstart,cend = mat->cmap->rend,row,col; 638 639 PetscFunctionBegin; 640 for (i=0; i<m; i++) { 641 if (idxm[i] < 0) continue; /* SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative row: %D",idxm[i]);*/ 642 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); 643 if (idxm[i] >= rstart && idxm[i] < rend) { 644 row = idxm[i] - rstart; 645 for (j=0; j<n; j++) { 646 if (idxn[j] < 0) continue; /* SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative column: %D",idxn[j]); */ 647 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); 648 if (idxn[j] >= cstart && idxn[j] < cend) { 649 col = idxn[j] - cstart; 650 ierr = MatGetValues(aij->A,1,&row,1,&col,v+i*n+j);CHKERRQ(ierr); 651 } else { 652 if (!aij->colmap) { 653 ierr = MatCreateColmap_MPIAIJ_Private(mat);CHKERRQ(ierr); 654 } 655 #if defined(PETSC_USE_CTABLE) 656 ierr = PetscTableFind(aij->colmap,idxn[j]+1,&col);CHKERRQ(ierr); 657 col--; 658 #else 659 col = aij->colmap[idxn[j]] - 1; 660 #endif 661 if ((col < 0) || (aij->garray[col] != idxn[j])) *(v+i*n+j) = 0.0; 662 else { 663 ierr = MatGetValues(aij->B,1,&row,1,&col,v+i*n+j);CHKERRQ(ierr); 664 } 665 } 666 } 667 } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Only local values currently supported"); 668 } 669 PetscFunctionReturn(0); 670 } 671 672 extern PetscErrorCode MatMultDiagonalBlock_MPIAIJ(Mat,Vec,Vec); 673 674 PetscErrorCode MatAssemblyBegin_MPIAIJ(Mat mat,MatAssemblyType mode) 675 { 676 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 677 PetscErrorCode ierr; 678 PetscInt nstash,reallocs; 679 680 PetscFunctionBegin; 681 if (aij->donotstash || mat->nooffprocentries) PetscFunctionReturn(0); 682 683 ierr = MatStashScatterBegin_Private(mat,&mat->stash,mat->rmap->range);CHKERRQ(ierr); 684 ierr = MatStashGetInfo_Private(&mat->stash,&nstash,&reallocs);CHKERRQ(ierr); 685 ierr = PetscInfo2(aij->A,"Stash has %D entries, uses %D mallocs.\n",nstash,reallocs);CHKERRQ(ierr); 686 PetscFunctionReturn(0); 687 } 688 689 PetscErrorCode MatAssemblyEnd_MPIAIJ(Mat mat,MatAssemblyType mode) 690 { 691 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 692 Mat_SeqAIJ *a = (Mat_SeqAIJ*)aij->A->data; 693 PetscErrorCode ierr; 694 PetscMPIInt n; 695 PetscInt i,j,rstart,ncols,flg; 696 PetscInt *row,*col; 697 PetscBool other_disassembled; 698 PetscScalar *val; 699 700 /* do not use 'b = (Mat_SeqAIJ*)aij->B->data' as B can be reset in disassembly */ 701 702 PetscFunctionBegin; 703 if (!aij->donotstash && !mat->nooffprocentries) { 704 while (1) { 705 ierr = MatStashScatterGetMesg_Private(&mat->stash,&n,&row,&col,&val,&flg);CHKERRQ(ierr); 706 if (!flg) break; 707 708 for (i=0; i<n; ) { 709 /* Now identify the consecutive vals belonging to the same row */ 710 for (j=i,rstart=row[j]; j<n; j++) { 711 if (row[j] != rstart) break; 712 } 713 if (j < n) ncols = j-i; 714 else ncols = n-i; 715 /* Now assemble all these values with a single function call */ 716 ierr = MatSetValues_MPIAIJ(mat,1,row+i,ncols,col+i,val+i,mat->insertmode);CHKERRQ(ierr); 717 718 i = j; 719 } 720 } 721 ierr = MatStashScatterEnd_Private(&mat->stash);CHKERRQ(ierr); 722 } 723 ierr = MatAssemblyBegin(aij->A,mode);CHKERRQ(ierr); 724 ierr = MatAssemblyEnd(aij->A,mode);CHKERRQ(ierr); 725 726 /* determine if any processor has disassembled, if so we must 727 also disassemble ourselfs, in order that we may reassemble. */ 728 /* 729 if nonzero structure of submatrix B cannot change then we know that 730 no processor disassembled thus we can skip this stuff 731 */ 732 if (!((Mat_SeqAIJ*)aij->B->data)->nonew) { 733 ierr = MPIU_Allreduce(&mat->was_assembled,&other_disassembled,1,MPIU_BOOL,MPI_PROD,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 734 if (mat->was_assembled && !other_disassembled) { 735 ierr = MatDisAssemble_MPIAIJ(mat);CHKERRQ(ierr); 736 } 737 } 738 if (!mat->was_assembled && mode == MAT_FINAL_ASSEMBLY) { 739 ierr = MatSetUpMultiply_MPIAIJ(mat);CHKERRQ(ierr); 740 } 741 ierr = MatSetOption(aij->B,MAT_USE_INODES,PETSC_FALSE);CHKERRQ(ierr); 742 ierr = MatAssemblyBegin(aij->B,mode);CHKERRQ(ierr); 743 ierr = MatAssemblyEnd(aij->B,mode);CHKERRQ(ierr); 744 745 ierr = PetscFree2(aij->rowvalues,aij->rowindices);CHKERRQ(ierr); 746 747 aij->rowvalues = 0; 748 749 ierr = VecDestroy(&aij->diag);CHKERRQ(ierr); 750 if (a->inode.size) mat->ops->multdiagonalblock = MatMultDiagonalBlock_MPIAIJ; 751 752 /* if no new nonzero locations are allowed in matrix then only set the matrix state the first time through */ 753 if ((!mat->was_assembled && mode == MAT_FINAL_ASSEMBLY) || !((Mat_SeqAIJ*)(aij->A->data))->nonew) { 754 PetscObjectState state = aij->A->nonzerostate + aij->B->nonzerostate; 755 ierr = MPIU_Allreduce(&state,&mat->nonzerostate,1,MPIU_INT64,MPI_SUM,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 756 } 757 PetscFunctionReturn(0); 758 } 759 760 PetscErrorCode MatZeroEntries_MPIAIJ(Mat A) 761 { 762 Mat_MPIAIJ *l = (Mat_MPIAIJ*)A->data; 763 PetscErrorCode ierr; 764 765 PetscFunctionBegin; 766 ierr = MatZeroEntries(l->A);CHKERRQ(ierr); 767 ierr = MatZeroEntries(l->B);CHKERRQ(ierr); 768 PetscFunctionReturn(0); 769 } 770 771 PetscErrorCode MatZeroRows_MPIAIJ(Mat A,PetscInt N,const PetscInt rows[],PetscScalar diag,Vec x,Vec b) 772 { 773 Mat_MPIAIJ *mat = (Mat_MPIAIJ *) A->data; 774 PetscInt *lrows; 775 PetscInt r, len; 776 PetscErrorCode ierr; 777 778 PetscFunctionBegin; 779 /* get locally owned rows */ 780 ierr = MatZeroRowsMapLocal_Private(A,N,rows,&len,&lrows);CHKERRQ(ierr); 781 /* fix right hand side if needed */ 782 if (x && b) { 783 const PetscScalar *xx; 784 PetscScalar *bb; 785 786 ierr = VecGetArrayRead(x, &xx);CHKERRQ(ierr); 787 ierr = VecGetArray(b, &bb);CHKERRQ(ierr); 788 for (r = 0; r < len; ++r) bb[lrows[r]] = diag*xx[lrows[r]]; 789 ierr = VecRestoreArrayRead(x, &xx);CHKERRQ(ierr); 790 ierr = VecRestoreArray(b, &bb);CHKERRQ(ierr); 791 } 792 /* Must zero l->B before l->A because the (diag) case below may put values into l->B*/ 793 ierr = MatZeroRows(mat->B, len, lrows, 0.0, NULL, NULL);CHKERRQ(ierr); 794 if (A->congruentlayouts == -1) { /* first time we compare rows and cols layouts */ 795 PetscBool cong; 796 ierr = PetscLayoutCompare(A->rmap,A->cmap,&cong);CHKERRQ(ierr); 797 if (cong) A->congruentlayouts = 1; 798 else A->congruentlayouts = 0; 799 } 800 if ((diag != 0.0) && A->congruentlayouts) { 801 ierr = MatZeroRows(mat->A, len, lrows, diag, NULL, NULL);CHKERRQ(ierr); 802 } else if (diag != 0.0) { 803 ierr = MatZeroRows(mat->A, len, lrows, 0.0, NULL, NULL);CHKERRQ(ierr); 804 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"); 805 for (r = 0; r < len; ++r) { 806 const PetscInt row = lrows[r] + A->rmap->rstart; 807 ierr = MatSetValues(A, 1, &row, 1, &row, &diag, INSERT_VALUES);CHKERRQ(ierr); 808 } 809 ierr = MatAssemblyBegin(A, MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 810 ierr = MatAssemblyEnd(A, MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 811 } else { 812 ierr = MatZeroRows(mat->A, len, lrows, 0.0, NULL, NULL);CHKERRQ(ierr); 813 } 814 ierr = PetscFree(lrows);CHKERRQ(ierr); 815 816 /* only change matrix nonzero state if pattern was allowed to be changed */ 817 if (!((Mat_SeqAIJ*)(mat->A->data))->keepnonzeropattern) { 818 PetscObjectState state = mat->A->nonzerostate + mat->B->nonzerostate; 819 ierr = MPIU_Allreduce(&state,&A->nonzerostate,1,MPIU_INT64,MPI_SUM,PetscObjectComm((PetscObject)A));CHKERRQ(ierr); 820 } 821 PetscFunctionReturn(0); 822 } 823 824 PetscErrorCode MatZeroRowsColumns_MPIAIJ(Mat A,PetscInt N,const PetscInt rows[],PetscScalar diag,Vec x,Vec b) 825 { 826 Mat_MPIAIJ *l = (Mat_MPIAIJ*)A->data; 827 PetscErrorCode ierr; 828 PetscMPIInt n = A->rmap->n; 829 PetscInt i,j,r,m,p = 0,len = 0; 830 PetscInt *lrows,*owners = A->rmap->range; 831 PetscSFNode *rrows; 832 PetscSF sf; 833 const PetscScalar *xx; 834 PetscScalar *bb,*mask; 835 Vec xmask,lmask; 836 Mat_SeqAIJ *aij = (Mat_SeqAIJ*)l->B->data; 837 const PetscInt *aj, *ii,*ridx; 838 PetscScalar *aa; 839 840 PetscFunctionBegin; 841 /* Create SF where leaves are input rows and roots are owned rows */ 842 ierr = PetscMalloc1(n, &lrows);CHKERRQ(ierr); 843 for (r = 0; r < n; ++r) lrows[r] = -1; 844 ierr = PetscMalloc1(N, &rrows);CHKERRQ(ierr); 845 for (r = 0; r < N; ++r) { 846 const PetscInt idx = rows[r]; 847 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); 848 if (idx < owners[p] || owners[p+1] <= idx) { /* short-circuit the search if the last p owns this row too */ 849 ierr = PetscLayoutFindOwner(A->rmap,idx,&p);CHKERRQ(ierr); 850 } 851 rrows[r].rank = p; 852 rrows[r].index = rows[r] - owners[p]; 853 } 854 ierr = PetscSFCreate(PetscObjectComm((PetscObject) A), &sf);CHKERRQ(ierr); 855 ierr = PetscSFSetGraph(sf, n, N, NULL, PETSC_OWN_POINTER, rrows, PETSC_OWN_POINTER);CHKERRQ(ierr); 856 /* Collect flags for rows to be zeroed */ 857 ierr = PetscSFReduceBegin(sf, MPIU_INT, (PetscInt *) rows, lrows, MPI_LOR);CHKERRQ(ierr); 858 ierr = PetscSFReduceEnd(sf, MPIU_INT, (PetscInt *) rows, lrows, MPI_LOR);CHKERRQ(ierr); 859 ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 860 /* Compress and put in row numbers */ 861 for (r = 0; r < n; ++r) if (lrows[r] >= 0) lrows[len++] = r; 862 /* zero diagonal part of matrix */ 863 ierr = MatZeroRowsColumns(l->A,len,lrows,diag,x,b);CHKERRQ(ierr); 864 /* handle off diagonal part of matrix */ 865 ierr = MatCreateVecs(A,&xmask,NULL);CHKERRQ(ierr); 866 ierr = VecDuplicate(l->lvec,&lmask);CHKERRQ(ierr); 867 ierr = VecGetArray(xmask,&bb);CHKERRQ(ierr); 868 for (i=0; i<len; i++) bb[lrows[i]] = 1; 869 ierr = VecRestoreArray(xmask,&bb);CHKERRQ(ierr); 870 ierr = VecScatterBegin(l->Mvctx,xmask,lmask,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 871 ierr = VecScatterEnd(l->Mvctx,xmask,lmask,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 872 ierr = VecDestroy(&xmask);CHKERRQ(ierr); 873 if (x) { 874 ierr = VecScatterBegin(l->Mvctx,x,l->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 875 ierr = VecScatterEnd(l->Mvctx,x,l->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 876 ierr = VecGetArrayRead(l->lvec,&xx);CHKERRQ(ierr); 877 ierr = VecGetArray(b,&bb);CHKERRQ(ierr); 878 } 879 ierr = VecGetArray(lmask,&mask);CHKERRQ(ierr); 880 /* remove zeroed rows of off diagonal matrix */ 881 ii = aij->i; 882 for (i=0; i<len; i++) { 883 ierr = PetscMemzero(aij->a + ii[lrows[i]],(ii[lrows[i]+1] - ii[lrows[i]])*sizeof(PetscScalar));CHKERRQ(ierr); 884 } 885 /* loop over all elements of off process part of matrix zeroing removed columns*/ 886 if (aij->compressedrow.use) { 887 m = aij->compressedrow.nrows; 888 ii = aij->compressedrow.i; 889 ridx = aij->compressedrow.rindex; 890 for (i=0; i<m; i++) { 891 n = ii[i+1] - ii[i]; 892 aj = aij->j + ii[i]; 893 aa = aij->a + ii[i]; 894 895 for (j=0; j<n; j++) { 896 if (PetscAbsScalar(mask[*aj])) { 897 if (b) bb[*ridx] -= *aa*xx[*aj]; 898 *aa = 0.0; 899 } 900 aa++; 901 aj++; 902 } 903 ridx++; 904 } 905 } else { /* do not use compressed row format */ 906 m = l->B->rmap->n; 907 for (i=0; i<m; i++) { 908 n = ii[i+1] - ii[i]; 909 aj = aij->j + ii[i]; 910 aa = aij->a + ii[i]; 911 for (j=0; j<n; j++) { 912 if (PetscAbsScalar(mask[*aj])) { 913 if (b) bb[i] -= *aa*xx[*aj]; 914 *aa = 0.0; 915 } 916 aa++; 917 aj++; 918 } 919 } 920 } 921 if (x) { 922 ierr = VecRestoreArray(b,&bb);CHKERRQ(ierr); 923 ierr = VecRestoreArrayRead(l->lvec,&xx);CHKERRQ(ierr); 924 } 925 ierr = VecRestoreArray(lmask,&mask);CHKERRQ(ierr); 926 ierr = VecDestroy(&lmask);CHKERRQ(ierr); 927 ierr = PetscFree(lrows);CHKERRQ(ierr); 928 929 /* only change matrix nonzero state if pattern was allowed to be changed */ 930 if (!((Mat_SeqAIJ*)(l->A->data))->keepnonzeropattern) { 931 PetscObjectState state = l->A->nonzerostate + l->B->nonzerostate; 932 ierr = MPIU_Allreduce(&state,&A->nonzerostate,1,MPIU_INT64,MPI_SUM,PetscObjectComm((PetscObject)A));CHKERRQ(ierr); 933 } 934 PetscFunctionReturn(0); 935 } 936 937 PetscErrorCode MatMult_MPIAIJ(Mat A,Vec xx,Vec yy) 938 { 939 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 940 PetscErrorCode ierr; 941 PetscInt nt; 942 VecScatter Mvctx = a->Mvctx; 943 944 PetscFunctionBegin; 945 ierr = VecGetLocalSize(xx,&nt);CHKERRQ(ierr); 946 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); 947 948 ierr = VecScatterBegin(Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 949 ierr = (*a->A->ops->mult)(a->A,xx,yy);CHKERRQ(ierr); 950 ierr = VecScatterEnd(Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 951 ierr = (*a->B->ops->multadd)(a->B,a->lvec,yy,yy);CHKERRQ(ierr); 952 PetscFunctionReturn(0); 953 } 954 955 PetscErrorCode MatMultDiagonalBlock_MPIAIJ(Mat A,Vec bb,Vec xx) 956 { 957 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 958 PetscErrorCode ierr; 959 960 PetscFunctionBegin; 961 ierr = MatMultDiagonalBlock(a->A,bb,xx);CHKERRQ(ierr); 962 PetscFunctionReturn(0); 963 } 964 965 PetscErrorCode MatMultAdd_MPIAIJ(Mat A,Vec xx,Vec yy,Vec zz) 966 { 967 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 968 PetscErrorCode ierr; 969 VecScatter Mvctx = a->Mvctx; 970 971 PetscFunctionBegin; 972 if (a->Mvctx_mpi1_flg) Mvctx = a->Mvctx_mpi1; 973 ierr = VecScatterBegin(Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 974 ierr = (*a->A->ops->multadd)(a->A,xx,yy,zz);CHKERRQ(ierr); 975 ierr = VecScatterEnd(Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 976 ierr = (*a->B->ops->multadd)(a->B,a->lvec,zz,zz);CHKERRQ(ierr); 977 PetscFunctionReturn(0); 978 } 979 980 PetscErrorCode MatMultTranspose_MPIAIJ(Mat A,Vec xx,Vec yy) 981 { 982 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 983 PetscErrorCode ierr; 984 PetscBool merged; 985 986 PetscFunctionBegin; 987 ierr = VecScatterGetMerged(a->Mvctx,&merged);CHKERRQ(ierr); 988 /* do nondiagonal part */ 989 ierr = (*a->B->ops->multtranspose)(a->B,xx,a->lvec);CHKERRQ(ierr); 990 if (!merged) { 991 /* send it on its way */ 992 ierr = VecScatterBegin(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 993 /* do local part */ 994 ierr = (*a->A->ops->multtranspose)(a->A,xx,yy);CHKERRQ(ierr); 995 /* receive remote parts: note this assumes the values are not actually */ 996 /* added in yy until the next line, */ 997 ierr = VecScatterEnd(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 998 } else { 999 /* do local part */ 1000 ierr = (*a->A->ops->multtranspose)(a->A,xx,yy);CHKERRQ(ierr); 1001 /* send it on its way */ 1002 ierr = VecScatterBegin(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1003 /* values actually were received in the Begin() but we need to call this nop */ 1004 ierr = VecScatterEnd(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1005 } 1006 PetscFunctionReturn(0); 1007 } 1008 1009 PetscErrorCode MatIsTranspose_MPIAIJ(Mat Amat,Mat Bmat,PetscReal tol,PetscBool *f) 1010 { 1011 MPI_Comm comm; 1012 Mat_MPIAIJ *Aij = (Mat_MPIAIJ*) Amat->data, *Bij; 1013 Mat Adia = Aij->A, Bdia, Aoff,Boff,*Aoffs,*Boffs; 1014 IS Me,Notme; 1015 PetscErrorCode ierr; 1016 PetscInt M,N,first,last,*notme,i; 1017 PetscMPIInt size; 1018 1019 PetscFunctionBegin; 1020 /* Easy test: symmetric diagonal block */ 1021 Bij = (Mat_MPIAIJ*) Bmat->data; Bdia = Bij->A; 1022 ierr = MatIsTranspose(Adia,Bdia,tol,f);CHKERRQ(ierr); 1023 if (!*f) PetscFunctionReturn(0); 1024 ierr = PetscObjectGetComm((PetscObject)Amat,&comm);CHKERRQ(ierr); 1025 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 1026 if (size == 1) PetscFunctionReturn(0); 1027 1028 /* Hard test: off-diagonal block. This takes a MatCreateSubMatrix. */ 1029 ierr = MatGetSize(Amat,&M,&N);CHKERRQ(ierr); 1030 ierr = MatGetOwnershipRange(Amat,&first,&last);CHKERRQ(ierr); 1031 ierr = PetscMalloc1(N-last+first,¬me);CHKERRQ(ierr); 1032 for (i=0; i<first; i++) notme[i] = i; 1033 for (i=last; i<M; i++) notme[i-last+first] = i; 1034 ierr = ISCreateGeneral(MPI_COMM_SELF,N-last+first,notme,PETSC_COPY_VALUES,&Notme);CHKERRQ(ierr); 1035 ierr = ISCreateStride(MPI_COMM_SELF,last-first,first,1,&Me);CHKERRQ(ierr); 1036 ierr = MatCreateSubMatrices(Amat,1,&Me,&Notme,MAT_INITIAL_MATRIX,&Aoffs);CHKERRQ(ierr); 1037 Aoff = Aoffs[0]; 1038 ierr = MatCreateSubMatrices(Bmat,1,&Notme,&Me,MAT_INITIAL_MATRIX,&Boffs);CHKERRQ(ierr); 1039 Boff = Boffs[0]; 1040 ierr = MatIsTranspose(Aoff,Boff,tol,f);CHKERRQ(ierr); 1041 ierr = MatDestroyMatrices(1,&Aoffs);CHKERRQ(ierr); 1042 ierr = MatDestroyMatrices(1,&Boffs);CHKERRQ(ierr); 1043 ierr = ISDestroy(&Me);CHKERRQ(ierr); 1044 ierr = ISDestroy(&Notme);CHKERRQ(ierr); 1045 ierr = PetscFree(notme);CHKERRQ(ierr); 1046 PetscFunctionReturn(0); 1047 } 1048 1049 PetscErrorCode MatIsSymmetric_MPIAIJ(Mat A,PetscReal tol,PetscBool *f) 1050 { 1051 PetscErrorCode ierr; 1052 1053 PetscFunctionBegin; 1054 ierr = MatIsTranspose_MPIAIJ(A,A,tol,f);CHKERRQ(ierr); 1055 PetscFunctionReturn(0); 1056 } 1057 1058 PetscErrorCode MatMultTransposeAdd_MPIAIJ(Mat A,Vec xx,Vec yy,Vec zz) 1059 { 1060 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 1061 PetscErrorCode ierr; 1062 1063 PetscFunctionBegin; 1064 /* do nondiagonal part */ 1065 ierr = (*a->B->ops->multtranspose)(a->B,xx,a->lvec);CHKERRQ(ierr); 1066 /* send it on its way */ 1067 ierr = VecScatterBegin(a->Mvctx,a->lvec,zz,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1068 /* do local part */ 1069 ierr = (*a->A->ops->multtransposeadd)(a->A,xx,yy,zz);CHKERRQ(ierr); 1070 /* receive remote parts */ 1071 ierr = VecScatterEnd(a->Mvctx,a->lvec,zz,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1072 PetscFunctionReturn(0); 1073 } 1074 1075 /* 1076 This only works correctly for square matrices where the subblock A->A is the 1077 diagonal block 1078 */ 1079 PetscErrorCode MatGetDiagonal_MPIAIJ(Mat A,Vec v) 1080 { 1081 PetscErrorCode ierr; 1082 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 1083 1084 PetscFunctionBegin; 1085 if (A->rmap->N != A->cmap->N) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"Supports only square matrix where A->A is diag block"); 1086 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"); 1087 ierr = MatGetDiagonal(a->A,v);CHKERRQ(ierr); 1088 PetscFunctionReturn(0); 1089 } 1090 1091 PetscErrorCode MatScale_MPIAIJ(Mat A,PetscScalar aa) 1092 { 1093 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 1094 PetscErrorCode ierr; 1095 1096 PetscFunctionBegin; 1097 ierr = MatScale(a->A,aa);CHKERRQ(ierr); 1098 ierr = MatScale(a->B,aa);CHKERRQ(ierr); 1099 PetscFunctionReturn(0); 1100 } 1101 1102 PetscErrorCode MatDestroy_MPIAIJ(Mat mat) 1103 { 1104 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 1105 PetscErrorCode ierr; 1106 1107 PetscFunctionBegin; 1108 #if defined(PETSC_USE_LOG) 1109 PetscLogObjectState((PetscObject)mat,"Rows=%D, Cols=%D",mat->rmap->N,mat->cmap->N); 1110 #endif 1111 ierr = MatStashDestroy_Private(&mat->stash);CHKERRQ(ierr); 1112 ierr = VecDestroy(&aij->diag);CHKERRQ(ierr); 1113 ierr = MatDestroy(&aij->A);CHKERRQ(ierr); 1114 ierr = MatDestroy(&aij->B);CHKERRQ(ierr); 1115 #if defined(PETSC_USE_CTABLE) 1116 ierr = PetscTableDestroy(&aij->colmap);CHKERRQ(ierr); 1117 #else 1118 ierr = PetscFree(aij->colmap);CHKERRQ(ierr); 1119 #endif 1120 ierr = PetscFree(aij->garray);CHKERRQ(ierr); 1121 ierr = VecDestroy(&aij->lvec);CHKERRQ(ierr); 1122 ierr = VecScatterDestroy(&aij->Mvctx);CHKERRQ(ierr); 1123 if (aij->Mvctx_mpi1) {ierr = VecScatterDestroy(&aij->Mvctx_mpi1);CHKERRQ(ierr);} 1124 ierr = PetscFree2(aij->rowvalues,aij->rowindices);CHKERRQ(ierr); 1125 ierr = PetscFree(aij->ld);CHKERRQ(ierr); 1126 ierr = PetscFree(mat->data);CHKERRQ(ierr); 1127 1128 ierr = PetscObjectChangeTypeName((PetscObject)mat,0);CHKERRQ(ierr); 1129 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatStoreValues_C",NULL);CHKERRQ(ierr); 1130 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatRetrieveValues_C",NULL);CHKERRQ(ierr); 1131 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatIsTranspose_C",NULL);CHKERRQ(ierr); 1132 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatMPIAIJSetPreallocation_C",NULL);CHKERRQ(ierr); 1133 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatResetPreallocation_C",NULL);CHKERRQ(ierr); 1134 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatMPIAIJSetPreallocationCSR_C",NULL);CHKERRQ(ierr); 1135 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatDiagonalScaleLocal_C",NULL);CHKERRQ(ierr); 1136 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatConvert_mpiaij_mpisbaij_C",NULL);CHKERRQ(ierr); 1137 #if defined(PETSC_HAVE_ELEMENTAL) 1138 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatConvert_mpiaij_elemental_C",NULL);CHKERRQ(ierr); 1139 #endif 1140 #if defined(PETSC_HAVE_HYPRE) 1141 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatConvert_mpiaij_hypre_C",NULL);CHKERRQ(ierr); 1142 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatMatMatMult_transpose_mpiaij_mpiaij_C",NULL);CHKERRQ(ierr); 1143 #endif 1144 PetscFunctionReturn(0); 1145 } 1146 1147 PetscErrorCode MatView_MPIAIJ_Binary(Mat mat,PetscViewer viewer) 1148 { 1149 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 1150 Mat_SeqAIJ *A = (Mat_SeqAIJ*)aij->A->data; 1151 Mat_SeqAIJ *B = (Mat_SeqAIJ*)aij->B->data; 1152 PetscErrorCode ierr; 1153 PetscMPIInt rank,size,tag = ((PetscObject)viewer)->tag; 1154 int fd; 1155 PetscInt nz,header[4],*row_lengths,*range=0,rlen,i; 1156 PetscInt nzmax,*column_indices,j,k,col,*garray = aij->garray,cnt,cstart = mat->cmap->rstart,rnz = 0; 1157 PetscScalar *column_values; 1158 PetscInt message_count,flowcontrolcount; 1159 FILE *file; 1160 1161 PetscFunctionBegin; 1162 ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr); 1163 ierr = MPI_Comm_size(PetscObjectComm((PetscObject)mat),&size);CHKERRQ(ierr); 1164 nz = A->nz + B->nz; 1165 ierr = PetscViewerBinaryGetDescriptor(viewer,&fd);CHKERRQ(ierr); 1166 if (!rank) { 1167 header[0] = MAT_FILE_CLASSID; 1168 header[1] = mat->rmap->N; 1169 header[2] = mat->cmap->N; 1170 1171 ierr = MPI_Reduce(&nz,&header[3],1,MPIU_INT,MPI_SUM,0,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1172 ierr = PetscBinaryWrite(fd,header,4,PETSC_INT,PETSC_TRUE);CHKERRQ(ierr); 1173 /* get largest number of rows any processor has */ 1174 rlen = mat->rmap->n; 1175 range = mat->rmap->range; 1176 for (i=1; i<size; i++) rlen = PetscMax(rlen,range[i+1] - range[i]); 1177 } else { 1178 ierr = MPI_Reduce(&nz,0,1,MPIU_INT,MPI_SUM,0,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1179 rlen = mat->rmap->n; 1180 } 1181 1182 /* load up the local row counts */ 1183 ierr = PetscMalloc1(rlen+1,&row_lengths);CHKERRQ(ierr); 1184 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]; 1185 1186 /* store the row lengths to the file */ 1187 ierr = PetscViewerFlowControlStart(viewer,&message_count,&flowcontrolcount);CHKERRQ(ierr); 1188 if (!rank) { 1189 ierr = PetscBinaryWrite(fd,row_lengths,mat->rmap->n,PETSC_INT,PETSC_TRUE);CHKERRQ(ierr); 1190 for (i=1; i<size; i++) { 1191 ierr = PetscViewerFlowControlStepMaster(viewer,i,&message_count,flowcontrolcount);CHKERRQ(ierr); 1192 rlen = range[i+1] - range[i]; 1193 ierr = MPIULong_Recv(row_lengths,rlen,MPIU_INT,i,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1194 ierr = PetscBinaryWrite(fd,row_lengths,rlen,PETSC_INT,PETSC_TRUE);CHKERRQ(ierr); 1195 } 1196 ierr = PetscViewerFlowControlEndMaster(viewer,&message_count);CHKERRQ(ierr); 1197 } else { 1198 ierr = PetscViewerFlowControlStepWorker(viewer,rank,&message_count);CHKERRQ(ierr); 1199 ierr = MPIULong_Send(row_lengths,mat->rmap->n,MPIU_INT,0,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1200 ierr = PetscViewerFlowControlEndWorker(viewer,&message_count);CHKERRQ(ierr); 1201 } 1202 ierr = PetscFree(row_lengths);CHKERRQ(ierr); 1203 1204 /* load up the local column indices */ 1205 nzmax = nz; /* th processor needs space a largest processor needs */ 1206 ierr = MPI_Reduce(&nz,&nzmax,1,MPIU_INT,MPI_MAX,0,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1207 ierr = PetscMalloc1(nzmax+1,&column_indices);CHKERRQ(ierr); 1208 cnt = 0; 1209 for (i=0; i<mat->rmap->n; i++) { 1210 for (j=B->i[i]; j<B->i[i+1]; j++) { 1211 if ((col = garray[B->j[j]]) > cstart) break; 1212 column_indices[cnt++] = col; 1213 } 1214 for (k=A->i[i]; k<A->i[i+1]; k++) column_indices[cnt++] = A->j[k] + cstart; 1215 for (; j<B->i[i+1]; j++) column_indices[cnt++] = garray[B->j[j]]; 1216 } 1217 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); 1218 1219 /* store the column indices to the file */ 1220 ierr = PetscViewerFlowControlStart(viewer,&message_count,&flowcontrolcount);CHKERRQ(ierr); 1221 if (!rank) { 1222 MPI_Status status; 1223 ierr = PetscBinaryWrite(fd,column_indices,nz,PETSC_INT,PETSC_TRUE);CHKERRQ(ierr); 1224 for (i=1; i<size; i++) { 1225 ierr = PetscViewerFlowControlStepMaster(viewer,i,&message_count,flowcontrolcount);CHKERRQ(ierr); 1226 ierr = MPI_Recv(&rnz,1,MPIU_INT,i,tag,PetscObjectComm((PetscObject)mat),&status);CHKERRQ(ierr); 1227 if (rnz > nzmax) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_LIB,"Internal PETSc error: nz = %D nzmax = %D",nz,nzmax); 1228 ierr = MPIULong_Recv(column_indices,rnz,MPIU_INT,i,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1229 ierr = PetscBinaryWrite(fd,column_indices,rnz,PETSC_INT,PETSC_TRUE);CHKERRQ(ierr); 1230 } 1231 ierr = PetscViewerFlowControlEndMaster(viewer,&message_count);CHKERRQ(ierr); 1232 } else { 1233 ierr = PetscViewerFlowControlStepWorker(viewer,rank,&message_count);CHKERRQ(ierr); 1234 ierr = MPI_Send(&nz,1,MPIU_INT,0,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1235 ierr = MPIULong_Send(column_indices,nz,MPIU_INT,0,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1236 ierr = PetscViewerFlowControlEndWorker(viewer,&message_count);CHKERRQ(ierr); 1237 } 1238 ierr = PetscFree(column_indices);CHKERRQ(ierr); 1239 1240 /* load up the local column values */ 1241 ierr = PetscMalloc1(nzmax+1,&column_values);CHKERRQ(ierr); 1242 cnt = 0; 1243 for (i=0; i<mat->rmap->n; i++) { 1244 for (j=B->i[i]; j<B->i[i+1]; j++) { 1245 if (garray[B->j[j]] > cstart) break; 1246 column_values[cnt++] = B->a[j]; 1247 } 1248 for (k=A->i[i]; k<A->i[i+1]; k++) column_values[cnt++] = A->a[k]; 1249 for (; j<B->i[i+1]; j++) column_values[cnt++] = B->a[j]; 1250 } 1251 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); 1252 1253 /* store the column values to the file */ 1254 ierr = PetscViewerFlowControlStart(viewer,&message_count,&flowcontrolcount);CHKERRQ(ierr); 1255 if (!rank) { 1256 MPI_Status status; 1257 ierr = PetscBinaryWrite(fd,column_values,nz,PETSC_SCALAR,PETSC_TRUE);CHKERRQ(ierr); 1258 for (i=1; i<size; i++) { 1259 ierr = PetscViewerFlowControlStepMaster(viewer,i,&message_count,flowcontrolcount);CHKERRQ(ierr); 1260 ierr = MPI_Recv(&rnz,1,MPIU_INT,i,tag,PetscObjectComm((PetscObject)mat),&status);CHKERRQ(ierr); 1261 if (rnz > nzmax) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_LIB,"Internal PETSc error: nz = %D nzmax = %D",nz,nzmax); 1262 ierr = MPIULong_Recv(column_values,rnz,MPIU_SCALAR,i,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1263 ierr = PetscBinaryWrite(fd,column_values,rnz,PETSC_SCALAR,PETSC_TRUE);CHKERRQ(ierr); 1264 } 1265 ierr = PetscViewerFlowControlEndMaster(viewer,&message_count);CHKERRQ(ierr); 1266 } else { 1267 ierr = PetscViewerFlowControlStepWorker(viewer,rank,&message_count);CHKERRQ(ierr); 1268 ierr = MPI_Send(&nz,1,MPIU_INT,0,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1269 ierr = MPIULong_Send(column_values,nz,MPIU_SCALAR,0,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1270 ierr = PetscViewerFlowControlEndWorker(viewer,&message_count);CHKERRQ(ierr); 1271 } 1272 ierr = PetscFree(column_values);CHKERRQ(ierr); 1273 1274 ierr = PetscViewerBinaryGetInfoPointer(viewer,&file);CHKERRQ(ierr); 1275 if (file) fprintf(file,"-matload_block_size %d\n",(int)PetscAbs(mat->rmap->bs)); 1276 PetscFunctionReturn(0); 1277 } 1278 1279 #include <petscdraw.h> 1280 PetscErrorCode MatView_MPIAIJ_ASCIIorDraworSocket(Mat mat,PetscViewer viewer) 1281 { 1282 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 1283 PetscErrorCode ierr; 1284 PetscMPIInt rank = aij->rank,size = aij->size; 1285 PetscBool isdraw,iascii,isbinary; 1286 PetscViewer sviewer; 1287 PetscViewerFormat format; 1288 1289 PetscFunctionBegin; 1290 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERDRAW,&isdraw);CHKERRQ(ierr); 1291 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii);CHKERRQ(ierr); 1292 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERBINARY,&isbinary);CHKERRQ(ierr); 1293 if (iascii) { 1294 ierr = PetscViewerGetFormat(viewer,&format);CHKERRQ(ierr); 1295 if (format == PETSC_VIEWER_LOAD_BALANCE) { 1296 PetscInt i,nmax = 0,nmin = PETSC_MAX_INT,navg = 0,*nz,nzlocal = ((Mat_SeqAIJ*) (aij->A->data))->nz + ((Mat_SeqAIJ*) (aij->B->data))->nz; 1297 ierr = PetscMalloc1(size,&nz);CHKERRQ(ierr); 1298 ierr = MPI_Allgather(&nzlocal,1,MPIU_INT,nz,1,MPIU_INT,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1299 for (i=0; i<(PetscInt)size; i++) { 1300 nmax = PetscMax(nmax,nz[i]); 1301 nmin = PetscMin(nmin,nz[i]); 1302 navg += nz[i]; 1303 } 1304 ierr = PetscFree(nz);CHKERRQ(ierr); 1305 navg = navg/size; 1306 ierr = PetscViewerASCIIPrintf(viewer,"Load Balance - Nonzeros: Min %D avg %D max %D\n",nmin,navg,nmax);CHKERRQ(ierr); 1307 PetscFunctionReturn(0); 1308 } 1309 ierr = PetscViewerGetFormat(viewer,&format);CHKERRQ(ierr); 1310 if (format == PETSC_VIEWER_ASCII_INFO_DETAIL) { 1311 MatInfo info; 1312 PetscBool inodes; 1313 1314 ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr); 1315 ierr = MatGetInfo(mat,MAT_LOCAL,&info);CHKERRQ(ierr); 1316 ierr = MatInodeGetInodeSizes(aij->A,NULL,(PetscInt**)&inodes,NULL);CHKERRQ(ierr); 1317 ierr = PetscViewerASCIIPushSynchronized(viewer);CHKERRQ(ierr); 1318 if (!inodes) { 1319 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"[%d] Local rows %D nz %D nz alloced %D mem %g, not using I-node routines\n", 1320 rank,mat->rmap->n,(PetscInt)info.nz_used,(PetscInt)info.nz_allocated,(double)info.memory);CHKERRQ(ierr); 1321 } else { 1322 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"[%d] Local rows %D nz %D nz alloced %D mem %g, using I-node routines\n", 1323 rank,mat->rmap->n,(PetscInt)info.nz_used,(PetscInt)info.nz_allocated,(double)info.memory);CHKERRQ(ierr); 1324 } 1325 ierr = MatGetInfo(aij->A,MAT_LOCAL,&info);CHKERRQ(ierr); 1326 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"[%d] on-diagonal part: nz %D \n",rank,(PetscInt)info.nz_used);CHKERRQ(ierr); 1327 ierr = MatGetInfo(aij->B,MAT_LOCAL,&info);CHKERRQ(ierr); 1328 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"[%d] off-diagonal part: nz %D \n",rank,(PetscInt)info.nz_used);CHKERRQ(ierr); 1329 ierr = PetscViewerFlush(viewer);CHKERRQ(ierr); 1330 ierr = PetscViewerASCIIPopSynchronized(viewer);CHKERRQ(ierr); 1331 ierr = PetscViewerASCIIPrintf(viewer,"Information on VecScatter used in matrix-vector product: \n");CHKERRQ(ierr); 1332 ierr = VecScatterView(aij->Mvctx,viewer);CHKERRQ(ierr); 1333 PetscFunctionReturn(0); 1334 } else if (format == PETSC_VIEWER_ASCII_INFO) { 1335 PetscInt inodecount,inodelimit,*inodes; 1336 ierr = MatInodeGetInodeSizes(aij->A,&inodecount,&inodes,&inodelimit);CHKERRQ(ierr); 1337 if (inodes) { 1338 ierr = PetscViewerASCIIPrintf(viewer,"using I-node (on process 0) routines: found %D nodes, limit used is %D\n",inodecount,inodelimit);CHKERRQ(ierr); 1339 } else { 1340 ierr = PetscViewerASCIIPrintf(viewer,"not using I-node (on process 0) routines\n");CHKERRQ(ierr); 1341 } 1342 PetscFunctionReturn(0); 1343 } else if (format == PETSC_VIEWER_ASCII_FACTOR_INFO) { 1344 PetscFunctionReturn(0); 1345 } 1346 } else if (isbinary) { 1347 if (size == 1) { 1348 ierr = PetscObjectSetName((PetscObject)aij->A,((PetscObject)mat)->name);CHKERRQ(ierr); 1349 ierr = MatView(aij->A,viewer);CHKERRQ(ierr); 1350 } else { 1351 ierr = MatView_MPIAIJ_Binary(mat,viewer);CHKERRQ(ierr); 1352 } 1353 PetscFunctionReturn(0); 1354 } else if (isdraw) { 1355 PetscDraw draw; 1356 PetscBool isnull; 1357 ierr = PetscViewerDrawGetDraw(viewer,0,&draw);CHKERRQ(ierr); 1358 ierr = PetscDrawIsNull(draw,&isnull);CHKERRQ(ierr); 1359 if (isnull) PetscFunctionReturn(0); 1360 } 1361 1362 { 1363 /* assemble the entire matrix onto first processor. */ 1364 Mat A; 1365 Mat_SeqAIJ *Aloc; 1366 PetscInt M = mat->rmap->N,N = mat->cmap->N,m,*ai,*aj,row,*cols,i,*ct; 1367 MatScalar *a; 1368 1369 ierr = MatCreate(PetscObjectComm((PetscObject)mat),&A);CHKERRQ(ierr); 1370 if (!rank) { 1371 ierr = MatSetSizes(A,M,N,M,N);CHKERRQ(ierr); 1372 } else { 1373 ierr = MatSetSizes(A,0,0,M,N);CHKERRQ(ierr); 1374 } 1375 /* This is just a temporary matrix, so explicitly using MATMPIAIJ is probably best */ 1376 ierr = MatSetType(A,MATMPIAIJ);CHKERRQ(ierr); 1377 ierr = MatMPIAIJSetPreallocation(A,0,NULL,0,NULL);CHKERRQ(ierr); 1378 ierr = MatSetOption(A,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 1379 ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)A);CHKERRQ(ierr); 1380 1381 /* copy over the A part */ 1382 Aloc = (Mat_SeqAIJ*)aij->A->data; 1383 m = aij->A->rmap->n; ai = Aloc->i; aj = Aloc->j; a = Aloc->a; 1384 row = mat->rmap->rstart; 1385 for (i=0; i<ai[m]; i++) aj[i] += mat->cmap->rstart; 1386 for (i=0; i<m; i++) { 1387 ierr = MatSetValues(A,1,&row,ai[i+1]-ai[i],aj,a,INSERT_VALUES);CHKERRQ(ierr); 1388 row++; 1389 a += ai[i+1]-ai[i]; aj += ai[i+1]-ai[i]; 1390 } 1391 aj = Aloc->j; 1392 for (i=0; i<ai[m]; i++) aj[i] -= mat->cmap->rstart; 1393 1394 /* copy over the B part */ 1395 Aloc = (Mat_SeqAIJ*)aij->B->data; 1396 m = aij->B->rmap->n; ai = Aloc->i; aj = Aloc->j; a = Aloc->a; 1397 row = mat->rmap->rstart; 1398 ierr = PetscMalloc1(ai[m]+1,&cols);CHKERRQ(ierr); 1399 ct = cols; 1400 for (i=0; i<ai[m]; i++) cols[i] = aij->garray[aj[i]]; 1401 for (i=0; i<m; i++) { 1402 ierr = MatSetValues(A,1,&row,ai[i+1]-ai[i],cols,a,INSERT_VALUES);CHKERRQ(ierr); 1403 row++; 1404 a += ai[i+1]-ai[i]; cols += ai[i+1]-ai[i]; 1405 } 1406 ierr = PetscFree(ct);CHKERRQ(ierr); 1407 ierr = MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1408 ierr = MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1409 /* 1410 Everyone has to call to draw the matrix since the graphics waits are 1411 synchronized across all processors that share the PetscDraw object 1412 */ 1413 ierr = PetscViewerGetSubViewer(viewer,PETSC_COMM_SELF,&sviewer);CHKERRQ(ierr); 1414 if (!rank) { 1415 ierr = PetscObjectSetName((PetscObject)((Mat_MPIAIJ*)(A->data))->A,((PetscObject)mat)->name);CHKERRQ(ierr); 1416 ierr = MatView_SeqAIJ(((Mat_MPIAIJ*)(A->data))->A,sviewer);CHKERRQ(ierr); 1417 } 1418 ierr = PetscViewerRestoreSubViewer(viewer,PETSC_COMM_SELF,&sviewer);CHKERRQ(ierr); 1419 ierr = PetscViewerFlush(viewer);CHKERRQ(ierr); 1420 ierr = MatDestroy(&A);CHKERRQ(ierr); 1421 } 1422 PetscFunctionReturn(0); 1423 } 1424 1425 PetscErrorCode MatView_MPIAIJ(Mat mat,PetscViewer viewer) 1426 { 1427 PetscErrorCode ierr; 1428 PetscBool iascii,isdraw,issocket,isbinary; 1429 1430 PetscFunctionBegin; 1431 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii);CHKERRQ(ierr); 1432 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERDRAW,&isdraw);CHKERRQ(ierr); 1433 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERBINARY,&isbinary);CHKERRQ(ierr); 1434 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERSOCKET,&issocket);CHKERRQ(ierr); 1435 if (iascii || isdraw || isbinary || issocket) { 1436 ierr = MatView_MPIAIJ_ASCIIorDraworSocket(mat,viewer);CHKERRQ(ierr); 1437 } 1438 PetscFunctionReturn(0); 1439 } 1440 1441 PetscErrorCode MatSOR_MPIAIJ(Mat matin,Vec bb,PetscReal omega,MatSORType flag,PetscReal fshift,PetscInt its,PetscInt lits,Vec xx) 1442 { 1443 Mat_MPIAIJ *mat = (Mat_MPIAIJ*)matin->data; 1444 PetscErrorCode ierr; 1445 Vec bb1 = 0; 1446 PetscBool hasop; 1447 1448 PetscFunctionBegin; 1449 if (flag == SOR_APPLY_UPPER) { 1450 ierr = (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);CHKERRQ(ierr); 1451 PetscFunctionReturn(0); 1452 } 1453 1454 if (its > 1 || ~flag & SOR_ZERO_INITIAL_GUESS || flag & SOR_EISENSTAT) { 1455 ierr = VecDuplicate(bb,&bb1);CHKERRQ(ierr); 1456 } 1457 1458 if ((flag & SOR_LOCAL_SYMMETRIC_SWEEP) == SOR_LOCAL_SYMMETRIC_SWEEP) { 1459 if (flag & SOR_ZERO_INITIAL_GUESS) { 1460 ierr = (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);CHKERRQ(ierr); 1461 its--; 1462 } 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_SYMMETRIC_SWEEP,fshift,lits,1,xx);CHKERRQ(ierr); 1474 } 1475 } else if (flag & SOR_LOCAL_FORWARD_SWEEP) { 1476 if (flag & SOR_ZERO_INITIAL_GUESS) { 1477 ierr = (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);CHKERRQ(ierr); 1478 its--; 1479 } 1480 while (its--) { 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 1484 /* update rhs: bb1 = bb - B*x */ 1485 ierr = VecScale(mat->lvec,-1.0);CHKERRQ(ierr); 1486 ierr = (*mat->B->ops->multadd)(mat->B,mat->lvec,bb,bb1);CHKERRQ(ierr); 1487 1488 /* local sweep */ 1489 ierr = (*mat->A->ops->sor)(mat->A,bb1,omega,SOR_FORWARD_SWEEP,fshift,lits,1,xx);CHKERRQ(ierr); 1490 } 1491 } else if (flag & SOR_LOCAL_BACKWARD_SWEEP) { 1492 if (flag & SOR_ZERO_INITIAL_GUESS) { 1493 ierr = (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);CHKERRQ(ierr); 1494 its--; 1495 } 1496 while (its--) { 1497 ierr = VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1498 ierr = VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1499 1500 /* update rhs: bb1 = bb - B*x */ 1501 ierr = VecScale(mat->lvec,-1.0);CHKERRQ(ierr); 1502 ierr = (*mat->B->ops->multadd)(mat->B,mat->lvec,bb,bb1);CHKERRQ(ierr); 1503 1504 /* local sweep */ 1505 ierr = (*mat->A->ops->sor)(mat->A,bb1,omega,SOR_BACKWARD_SWEEP,fshift,lits,1,xx);CHKERRQ(ierr); 1506 } 1507 } else if (flag & SOR_EISENSTAT) { 1508 Vec xx1; 1509 1510 ierr = VecDuplicate(bb,&xx1);CHKERRQ(ierr); 1511 ierr = (*mat->A->ops->sor)(mat->A,bb,omega,(MatSORType)(SOR_ZERO_INITIAL_GUESS | SOR_LOCAL_BACKWARD_SWEEP),fshift,lits,1,xx);CHKERRQ(ierr); 1512 1513 ierr = VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1514 ierr = VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1515 if (!mat->diag) { 1516 ierr = MatCreateVecs(matin,&mat->diag,NULL);CHKERRQ(ierr); 1517 ierr = MatGetDiagonal(matin,mat->diag);CHKERRQ(ierr); 1518 } 1519 ierr = MatHasOperation(matin,MATOP_MULT_DIAGONAL_BLOCK,&hasop);CHKERRQ(ierr); 1520 if (hasop) { 1521 ierr = MatMultDiagonalBlock(matin,xx,bb1);CHKERRQ(ierr); 1522 } else { 1523 ierr = VecPointwiseMult(bb1,mat->diag,xx);CHKERRQ(ierr); 1524 } 1525 ierr = VecAYPX(bb1,(omega-2.0)/omega,bb);CHKERRQ(ierr); 1526 1527 ierr = MatMultAdd(mat->B,mat->lvec,bb1,bb1);CHKERRQ(ierr); 1528 1529 /* local sweep */ 1530 ierr = (*mat->A->ops->sor)(mat->A,bb1,omega,(MatSORType)(SOR_ZERO_INITIAL_GUESS | SOR_LOCAL_FORWARD_SWEEP),fshift,lits,1,xx1);CHKERRQ(ierr); 1531 ierr = VecAXPY(xx,1.0,xx1);CHKERRQ(ierr); 1532 ierr = VecDestroy(&xx1);CHKERRQ(ierr); 1533 } else SETERRQ(PetscObjectComm((PetscObject)matin),PETSC_ERR_SUP,"Parallel SOR not supported"); 1534 1535 ierr = VecDestroy(&bb1);CHKERRQ(ierr); 1536 1537 matin->factorerrortype = mat->A->factorerrortype; 1538 PetscFunctionReturn(0); 1539 } 1540 1541 PetscErrorCode MatPermute_MPIAIJ(Mat A,IS rowp,IS colp,Mat *B) 1542 { 1543 Mat aA,aB,Aperm; 1544 const PetscInt *rwant,*cwant,*gcols,*ai,*bi,*aj,*bj; 1545 PetscScalar *aa,*ba; 1546 PetscInt i,j,m,n,ng,anz,bnz,*dnnz,*onnz,*tdnnz,*tonnz,*rdest,*cdest,*work,*gcdest; 1547 PetscSF rowsf,sf; 1548 IS parcolp = NULL; 1549 PetscBool done; 1550 PetscErrorCode ierr; 1551 1552 PetscFunctionBegin; 1553 ierr = MatGetLocalSize(A,&m,&n);CHKERRQ(ierr); 1554 ierr = ISGetIndices(rowp,&rwant);CHKERRQ(ierr); 1555 ierr = ISGetIndices(colp,&cwant);CHKERRQ(ierr); 1556 ierr = PetscMalloc3(PetscMax(m,n),&work,m,&rdest,n,&cdest);CHKERRQ(ierr); 1557 1558 /* Invert row permutation to find out where my rows should go */ 1559 ierr = PetscSFCreate(PetscObjectComm((PetscObject)A),&rowsf);CHKERRQ(ierr); 1560 ierr = PetscSFSetGraphLayout(rowsf,A->rmap,A->rmap->n,NULL,PETSC_OWN_POINTER,rwant);CHKERRQ(ierr); 1561 ierr = PetscSFSetFromOptions(rowsf);CHKERRQ(ierr); 1562 for (i=0; i<m; i++) work[i] = A->rmap->rstart + i; 1563 ierr = PetscSFReduceBegin(rowsf,MPIU_INT,work,rdest,MPIU_REPLACE);CHKERRQ(ierr); 1564 ierr = PetscSFReduceEnd(rowsf,MPIU_INT,work,rdest,MPIU_REPLACE);CHKERRQ(ierr); 1565 1566 /* Invert column permutation to find out where my columns should go */ 1567 ierr = PetscSFCreate(PetscObjectComm((PetscObject)A),&sf);CHKERRQ(ierr); 1568 ierr = PetscSFSetGraphLayout(sf,A->cmap,A->cmap->n,NULL,PETSC_OWN_POINTER,cwant);CHKERRQ(ierr); 1569 ierr = PetscSFSetFromOptions(sf);CHKERRQ(ierr); 1570 for (i=0; i<n; i++) work[i] = A->cmap->rstart + i; 1571 ierr = PetscSFReduceBegin(sf,MPIU_INT,work,cdest,MPIU_REPLACE);CHKERRQ(ierr); 1572 ierr = PetscSFReduceEnd(sf,MPIU_INT,work,cdest,MPIU_REPLACE);CHKERRQ(ierr); 1573 ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 1574 1575 ierr = ISRestoreIndices(rowp,&rwant);CHKERRQ(ierr); 1576 ierr = ISRestoreIndices(colp,&cwant);CHKERRQ(ierr); 1577 ierr = MatMPIAIJGetSeqAIJ(A,&aA,&aB,&gcols);CHKERRQ(ierr); 1578 1579 /* Find out where my gcols should go */ 1580 ierr = MatGetSize(aB,NULL,&ng);CHKERRQ(ierr); 1581 ierr = PetscMalloc1(ng,&gcdest);CHKERRQ(ierr); 1582 ierr = PetscSFCreate(PetscObjectComm((PetscObject)A),&sf);CHKERRQ(ierr); 1583 ierr = PetscSFSetGraphLayout(sf,A->cmap,ng,NULL,PETSC_OWN_POINTER,gcols);CHKERRQ(ierr); 1584 ierr = PetscSFSetFromOptions(sf);CHKERRQ(ierr); 1585 ierr = PetscSFBcastBegin(sf,MPIU_INT,cdest,gcdest);CHKERRQ(ierr); 1586 ierr = PetscSFBcastEnd(sf,MPIU_INT,cdest,gcdest);CHKERRQ(ierr); 1587 ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 1588 1589 ierr = PetscCalloc4(m,&dnnz,m,&onnz,m,&tdnnz,m,&tonnz);CHKERRQ(ierr); 1590 ierr = MatGetRowIJ(aA,0,PETSC_FALSE,PETSC_FALSE,&anz,&ai,&aj,&done);CHKERRQ(ierr); 1591 ierr = MatGetRowIJ(aB,0,PETSC_FALSE,PETSC_FALSE,&bnz,&bi,&bj,&done);CHKERRQ(ierr); 1592 for (i=0; i<m; i++) { 1593 PetscInt row = rdest[i],rowner; 1594 ierr = PetscLayoutFindOwner(A->rmap,row,&rowner);CHKERRQ(ierr); 1595 for (j=ai[i]; j<ai[i+1]; j++) { 1596 PetscInt cowner,col = cdest[aj[j]]; 1597 ierr = PetscLayoutFindOwner(A->cmap,col,&cowner);CHKERRQ(ierr); /* Could build an index for the columns to eliminate this search */ 1598 if (rowner == cowner) dnnz[i]++; 1599 else onnz[i]++; 1600 } 1601 for (j=bi[i]; j<bi[i+1]; j++) { 1602 PetscInt cowner,col = gcdest[bj[j]]; 1603 ierr = PetscLayoutFindOwner(A->cmap,col,&cowner);CHKERRQ(ierr); 1604 if (rowner == cowner) dnnz[i]++; 1605 else onnz[i]++; 1606 } 1607 } 1608 ierr = PetscSFBcastBegin(rowsf,MPIU_INT,dnnz,tdnnz);CHKERRQ(ierr); 1609 ierr = PetscSFBcastEnd(rowsf,MPIU_INT,dnnz,tdnnz);CHKERRQ(ierr); 1610 ierr = PetscSFBcastBegin(rowsf,MPIU_INT,onnz,tonnz);CHKERRQ(ierr); 1611 ierr = PetscSFBcastEnd(rowsf,MPIU_INT,onnz,tonnz);CHKERRQ(ierr); 1612 ierr = PetscSFDestroy(&rowsf);CHKERRQ(ierr); 1613 1614 ierr = MatCreateAIJ(PetscObjectComm((PetscObject)A),A->rmap->n,A->cmap->n,A->rmap->N,A->cmap->N,0,tdnnz,0,tonnz,&Aperm);CHKERRQ(ierr); 1615 ierr = MatSeqAIJGetArray(aA,&aa);CHKERRQ(ierr); 1616 ierr = MatSeqAIJGetArray(aB,&ba);CHKERRQ(ierr); 1617 for (i=0; i<m; i++) { 1618 PetscInt *acols = dnnz,*bcols = onnz; /* Repurpose now-unneeded arrays */ 1619 PetscInt j0,rowlen; 1620 rowlen = ai[i+1] - ai[i]; 1621 for (j0=j=0; j<rowlen; j0=j) { /* rowlen could be larger than number of rows m, so sum in batches */ 1622 for ( ; j<PetscMin(rowlen,j0+m); j++) acols[j-j0] = cdest[aj[ai[i]+j]]; 1623 ierr = MatSetValues(Aperm,1,&rdest[i],j-j0,acols,aa+ai[i]+j0,INSERT_VALUES);CHKERRQ(ierr); 1624 } 1625 rowlen = bi[i+1] - bi[i]; 1626 for (j0=j=0; j<rowlen; j0=j) { 1627 for ( ; j<PetscMin(rowlen,j0+m); j++) bcols[j-j0] = gcdest[bj[bi[i]+j]]; 1628 ierr = MatSetValues(Aperm,1,&rdest[i],j-j0,bcols,ba+bi[i]+j0,INSERT_VALUES);CHKERRQ(ierr); 1629 } 1630 } 1631 ierr = MatAssemblyBegin(Aperm,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1632 ierr = MatAssemblyEnd(Aperm,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1633 ierr = MatRestoreRowIJ(aA,0,PETSC_FALSE,PETSC_FALSE,&anz,&ai,&aj,&done);CHKERRQ(ierr); 1634 ierr = MatRestoreRowIJ(aB,0,PETSC_FALSE,PETSC_FALSE,&bnz,&bi,&bj,&done);CHKERRQ(ierr); 1635 ierr = MatSeqAIJRestoreArray(aA,&aa);CHKERRQ(ierr); 1636 ierr = MatSeqAIJRestoreArray(aB,&ba);CHKERRQ(ierr); 1637 ierr = PetscFree4(dnnz,onnz,tdnnz,tonnz);CHKERRQ(ierr); 1638 ierr = PetscFree3(work,rdest,cdest);CHKERRQ(ierr); 1639 ierr = PetscFree(gcdest);CHKERRQ(ierr); 1640 if (parcolp) {ierr = ISDestroy(&colp);CHKERRQ(ierr);} 1641 *B = Aperm; 1642 PetscFunctionReturn(0); 1643 } 1644 1645 PetscErrorCode MatGetGhosts_MPIAIJ(Mat mat,PetscInt *nghosts,const PetscInt *ghosts[]) 1646 { 1647 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 1648 PetscErrorCode ierr; 1649 1650 PetscFunctionBegin; 1651 ierr = MatGetSize(aij->B,NULL,nghosts);CHKERRQ(ierr); 1652 if (ghosts) *ghosts = aij->garray; 1653 PetscFunctionReturn(0); 1654 } 1655 1656 PetscErrorCode MatGetInfo_MPIAIJ(Mat matin,MatInfoType flag,MatInfo *info) 1657 { 1658 Mat_MPIAIJ *mat = (Mat_MPIAIJ*)matin->data; 1659 Mat A = mat->A,B = mat->B; 1660 PetscErrorCode ierr; 1661 PetscReal isend[5],irecv[5]; 1662 1663 PetscFunctionBegin; 1664 info->block_size = 1.0; 1665 ierr = MatGetInfo(A,MAT_LOCAL,info);CHKERRQ(ierr); 1666 1667 isend[0] = info->nz_used; isend[1] = info->nz_allocated; isend[2] = info->nz_unneeded; 1668 isend[3] = info->memory; isend[4] = info->mallocs; 1669 1670 ierr = MatGetInfo(B,MAT_LOCAL,info);CHKERRQ(ierr); 1671 1672 isend[0] += info->nz_used; isend[1] += info->nz_allocated; isend[2] += info->nz_unneeded; 1673 isend[3] += info->memory; isend[4] += info->mallocs; 1674 if (flag == MAT_LOCAL) { 1675 info->nz_used = isend[0]; 1676 info->nz_allocated = isend[1]; 1677 info->nz_unneeded = isend[2]; 1678 info->memory = isend[3]; 1679 info->mallocs = isend[4]; 1680 } else if (flag == MAT_GLOBAL_MAX) { 1681 ierr = MPIU_Allreduce(isend,irecv,5,MPIU_REAL,MPIU_MAX,PetscObjectComm((PetscObject)matin));CHKERRQ(ierr); 1682 1683 info->nz_used = irecv[0]; 1684 info->nz_allocated = irecv[1]; 1685 info->nz_unneeded = irecv[2]; 1686 info->memory = irecv[3]; 1687 info->mallocs = irecv[4]; 1688 } else if (flag == MAT_GLOBAL_SUM) { 1689 ierr = MPIU_Allreduce(isend,irecv,5,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)matin));CHKERRQ(ierr); 1690 1691 info->nz_used = irecv[0]; 1692 info->nz_allocated = irecv[1]; 1693 info->nz_unneeded = irecv[2]; 1694 info->memory = irecv[3]; 1695 info->mallocs = irecv[4]; 1696 } 1697 info->fill_ratio_given = 0; /* no parallel LU/ILU/Cholesky */ 1698 info->fill_ratio_needed = 0; 1699 info->factor_mallocs = 0; 1700 PetscFunctionReturn(0); 1701 } 1702 1703 PetscErrorCode MatSetOption_MPIAIJ(Mat A,MatOption op,PetscBool flg) 1704 { 1705 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 1706 PetscErrorCode ierr; 1707 1708 PetscFunctionBegin; 1709 switch (op) { 1710 case MAT_NEW_NONZERO_LOCATIONS: 1711 case MAT_NEW_NONZERO_ALLOCATION_ERR: 1712 case MAT_UNUSED_NONZERO_LOCATION_ERR: 1713 case MAT_KEEP_NONZERO_PATTERN: 1714 case MAT_NEW_NONZERO_LOCATION_ERR: 1715 case MAT_USE_INODES: 1716 case MAT_IGNORE_ZERO_ENTRIES: 1717 MatCheckPreallocated(A,1); 1718 ierr = MatSetOption(a->A,op,flg);CHKERRQ(ierr); 1719 ierr = MatSetOption(a->B,op,flg);CHKERRQ(ierr); 1720 break; 1721 case MAT_ROW_ORIENTED: 1722 MatCheckPreallocated(A,1); 1723 a->roworiented = flg; 1724 1725 ierr = MatSetOption(a->A,op,flg);CHKERRQ(ierr); 1726 ierr = MatSetOption(a->B,op,flg);CHKERRQ(ierr); 1727 break; 1728 case MAT_NEW_DIAGONALS: 1729 ierr = PetscInfo1(A,"Option %s ignored\n",MatOptions[op]);CHKERRQ(ierr); 1730 break; 1731 case MAT_IGNORE_OFF_PROC_ENTRIES: 1732 a->donotstash = flg; 1733 break; 1734 case MAT_SPD: 1735 A->spd_set = PETSC_TRUE; 1736 A->spd = flg; 1737 if (flg) { 1738 A->symmetric = PETSC_TRUE; 1739 A->structurally_symmetric = PETSC_TRUE; 1740 A->symmetric_set = PETSC_TRUE; 1741 A->structurally_symmetric_set = PETSC_TRUE; 1742 } 1743 break; 1744 case MAT_SYMMETRIC: 1745 MatCheckPreallocated(A,1); 1746 ierr = MatSetOption(a->A,op,flg);CHKERRQ(ierr); 1747 break; 1748 case MAT_STRUCTURALLY_SYMMETRIC: 1749 MatCheckPreallocated(A,1); 1750 ierr = MatSetOption(a->A,op,flg);CHKERRQ(ierr); 1751 break; 1752 case MAT_HERMITIAN: 1753 MatCheckPreallocated(A,1); 1754 ierr = MatSetOption(a->A,op,flg);CHKERRQ(ierr); 1755 break; 1756 case MAT_SYMMETRY_ETERNAL: 1757 MatCheckPreallocated(A,1); 1758 ierr = MatSetOption(a->A,op,flg);CHKERRQ(ierr); 1759 break; 1760 case MAT_SUBMAT_SINGLEIS: 1761 A->submat_singleis = flg; 1762 break; 1763 case MAT_STRUCTURE_ONLY: 1764 /* The option is handled directly by MatSetOption() */ 1765 break; 1766 default: 1767 SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"unknown option %d",op); 1768 } 1769 PetscFunctionReturn(0); 1770 } 1771 1772 PetscErrorCode MatGetRow_MPIAIJ(Mat matin,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v) 1773 { 1774 Mat_MPIAIJ *mat = (Mat_MPIAIJ*)matin->data; 1775 PetscScalar *vworkA,*vworkB,**pvA,**pvB,*v_p; 1776 PetscErrorCode ierr; 1777 PetscInt i,*cworkA,*cworkB,**pcA,**pcB,cstart = matin->cmap->rstart; 1778 PetscInt nztot,nzA,nzB,lrow,rstart = matin->rmap->rstart,rend = matin->rmap->rend; 1779 PetscInt *cmap,*idx_p; 1780 1781 PetscFunctionBegin; 1782 if (mat->getrowactive) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Already active"); 1783 mat->getrowactive = PETSC_TRUE; 1784 1785 if (!mat->rowvalues && (idx || v)) { 1786 /* 1787 allocate enough space to hold information from the longest row. 1788 */ 1789 Mat_SeqAIJ *Aa = (Mat_SeqAIJ*)mat->A->data,*Ba = (Mat_SeqAIJ*)mat->B->data; 1790 PetscInt max = 1,tmp; 1791 for (i=0; i<matin->rmap->n; i++) { 1792 tmp = Aa->i[i+1] - Aa->i[i] + Ba->i[i+1] - Ba->i[i]; 1793 if (max < tmp) max = tmp; 1794 } 1795 ierr = PetscMalloc2(max,&mat->rowvalues,max,&mat->rowindices);CHKERRQ(ierr); 1796 } 1797 1798 if (row < rstart || row >= rend) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Only local rows"); 1799 lrow = row - rstart; 1800 1801 pvA = &vworkA; pcA = &cworkA; pvB = &vworkB; pcB = &cworkB; 1802 if (!v) {pvA = 0; pvB = 0;} 1803 if (!idx) {pcA = 0; if (!v) pcB = 0;} 1804 ierr = (*mat->A->ops->getrow)(mat->A,lrow,&nzA,pcA,pvA);CHKERRQ(ierr); 1805 ierr = (*mat->B->ops->getrow)(mat->B,lrow,&nzB,pcB,pvB);CHKERRQ(ierr); 1806 nztot = nzA + nzB; 1807 1808 cmap = mat->garray; 1809 if (v || idx) { 1810 if (nztot) { 1811 /* Sort by increasing column numbers, assuming A and B already sorted */ 1812 PetscInt imark = -1; 1813 if (v) { 1814 *v = v_p = mat->rowvalues; 1815 for (i=0; i<nzB; i++) { 1816 if (cmap[cworkB[i]] < cstart) v_p[i] = vworkB[i]; 1817 else break; 1818 } 1819 imark = i; 1820 for (i=0; i<nzA; i++) v_p[imark+i] = vworkA[i]; 1821 for (i=imark; i<nzB; i++) v_p[nzA+i] = vworkB[i]; 1822 } 1823 if (idx) { 1824 *idx = idx_p = mat->rowindices; 1825 if (imark > -1) { 1826 for (i=0; i<imark; i++) { 1827 idx_p[i] = cmap[cworkB[i]]; 1828 } 1829 } else { 1830 for (i=0; i<nzB; i++) { 1831 if (cmap[cworkB[i]] < cstart) idx_p[i] = cmap[cworkB[i]]; 1832 else break; 1833 } 1834 imark = i; 1835 } 1836 for (i=0; i<nzA; i++) idx_p[imark+i] = cstart + cworkA[i]; 1837 for (i=imark; i<nzB; i++) idx_p[nzA+i] = cmap[cworkB[i]]; 1838 } 1839 } else { 1840 if (idx) *idx = 0; 1841 if (v) *v = 0; 1842 } 1843 } 1844 *nz = nztot; 1845 ierr = (*mat->A->ops->restorerow)(mat->A,lrow,&nzA,pcA,pvA);CHKERRQ(ierr); 1846 ierr = (*mat->B->ops->restorerow)(mat->B,lrow,&nzB,pcB,pvB);CHKERRQ(ierr); 1847 PetscFunctionReturn(0); 1848 } 1849 1850 PetscErrorCode MatRestoreRow_MPIAIJ(Mat mat,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v) 1851 { 1852 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 1853 1854 PetscFunctionBegin; 1855 if (!aij->getrowactive) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"MatGetRow() must be called first"); 1856 aij->getrowactive = PETSC_FALSE; 1857 PetscFunctionReturn(0); 1858 } 1859 1860 PetscErrorCode MatNorm_MPIAIJ(Mat mat,NormType type,PetscReal *norm) 1861 { 1862 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 1863 Mat_SeqAIJ *amat = (Mat_SeqAIJ*)aij->A->data,*bmat = (Mat_SeqAIJ*)aij->B->data; 1864 PetscErrorCode ierr; 1865 PetscInt i,j,cstart = mat->cmap->rstart; 1866 PetscReal sum = 0.0; 1867 MatScalar *v; 1868 1869 PetscFunctionBegin; 1870 if (aij->size == 1) { 1871 ierr = MatNorm(aij->A,type,norm);CHKERRQ(ierr); 1872 } else { 1873 if (type == NORM_FROBENIUS) { 1874 v = amat->a; 1875 for (i=0; i<amat->nz; i++) { 1876 sum += PetscRealPart(PetscConj(*v)*(*v)); v++; 1877 } 1878 v = bmat->a; 1879 for (i=0; i<bmat->nz; i++) { 1880 sum += PetscRealPart(PetscConj(*v)*(*v)); v++; 1881 } 1882 ierr = MPIU_Allreduce(&sum,norm,1,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1883 *norm = PetscSqrtReal(*norm); 1884 ierr = PetscLogFlops(2*amat->nz+2*bmat->nz);CHKERRQ(ierr); 1885 } else if (type == NORM_1) { /* max column norm */ 1886 PetscReal *tmp,*tmp2; 1887 PetscInt *jj,*garray = aij->garray; 1888 ierr = PetscCalloc1(mat->cmap->N+1,&tmp);CHKERRQ(ierr); 1889 ierr = PetscMalloc1(mat->cmap->N+1,&tmp2);CHKERRQ(ierr); 1890 *norm = 0.0; 1891 v = amat->a; jj = amat->j; 1892 for (j=0; j<amat->nz; j++) { 1893 tmp[cstart + *jj++] += PetscAbsScalar(*v); v++; 1894 } 1895 v = bmat->a; jj = bmat->j; 1896 for (j=0; j<bmat->nz; j++) { 1897 tmp[garray[*jj++]] += PetscAbsScalar(*v); v++; 1898 } 1899 ierr = MPIU_Allreduce(tmp,tmp2,mat->cmap->N,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1900 for (j=0; j<mat->cmap->N; j++) { 1901 if (tmp2[j] > *norm) *norm = tmp2[j]; 1902 } 1903 ierr = PetscFree(tmp);CHKERRQ(ierr); 1904 ierr = PetscFree(tmp2);CHKERRQ(ierr); 1905 ierr = PetscLogFlops(PetscMax(amat->nz+bmat->nz-1,0));CHKERRQ(ierr); 1906 } else if (type == NORM_INFINITY) { /* max row norm */ 1907 PetscReal ntemp = 0.0; 1908 for (j=0; j<aij->A->rmap->n; j++) { 1909 v = amat->a + amat->i[j]; 1910 sum = 0.0; 1911 for (i=0; i<amat->i[j+1]-amat->i[j]; i++) { 1912 sum += PetscAbsScalar(*v); v++; 1913 } 1914 v = bmat->a + bmat->i[j]; 1915 for (i=0; i<bmat->i[j+1]-bmat->i[j]; i++) { 1916 sum += PetscAbsScalar(*v); v++; 1917 } 1918 if (sum > ntemp) ntemp = sum; 1919 } 1920 ierr = MPIU_Allreduce(&ntemp,norm,1,MPIU_REAL,MPIU_MAX,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1921 ierr = PetscLogFlops(PetscMax(amat->nz+bmat->nz-1,0));CHKERRQ(ierr); 1922 } else SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"No support for two norm"); 1923 } 1924 PetscFunctionReturn(0); 1925 } 1926 1927 PetscErrorCode MatTranspose_MPIAIJ(Mat A,MatReuse reuse,Mat *matout) 1928 { 1929 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 1930 Mat_SeqAIJ *Aloc=(Mat_SeqAIJ*)a->A->data,*Bloc=(Mat_SeqAIJ*)a->B->data; 1931 PetscErrorCode ierr; 1932 PetscInt M = A->rmap->N,N = A->cmap->N,ma,na,mb,nb,*ai,*aj,*bi,*bj,row,*cols,*cols_tmp,i; 1933 PetscInt cstart = A->cmap->rstart,ncol; 1934 Mat B; 1935 MatScalar *array; 1936 1937 PetscFunctionBegin; 1938 if (reuse == MAT_INPLACE_MATRIX && M != N) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_SIZ,"Square matrix only for in-place"); 1939 1940 ma = A->rmap->n; na = A->cmap->n; mb = a->B->rmap->n; nb = a->B->cmap->n; 1941 ai = Aloc->i; aj = Aloc->j; 1942 bi = Bloc->i; bj = Bloc->j; 1943 if (reuse == MAT_INITIAL_MATRIX || *matout == A) { 1944 PetscInt *d_nnz,*g_nnz,*o_nnz; 1945 PetscSFNode *oloc; 1946 PETSC_UNUSED PetscSF sf; 1947 1948 ierr = PetscMalloc4(na,&d_nnz,na,&o_nnz,nb,&g_nnz,nb,&oloc);CHKERRQ(ierr); 1949 /* compute d_nnz for preallocation */ 1950 ierr = PetscMemzero(d_nnz,na*sizeof(PetscInt));CHKERRQ(ierr); 1951 for (i=0; i<ai[ma]; i++) { 1952 d_nnz[aj[i]]++; 1953 aj[i] += cstart; /* global col index to be used by MatSetValues() */ 1954 } 1955 /* compute local off-diagonal contributions */ 1956 ierr = PetscMemzero(g_nnz,nb*sizeof(PetscInt));CHKERRQ(ierr); 1957 for (i=0; i<bi[ma]; i++) g_nnz[bj[i]]++; 1958 /* map those to global */ 1959 ierr = PetscSFCreate(PetscObjectComm((PetscObject)A),&sf);CHKERRQ(ierr); 1960 ierr = PetscSFSetGraphLayout(sf,A->cmap,nb,NULL,PETSC_USE_POINTER,a->garray);CHKERRQ(ierr); 1961 ierr = PetscSFSetFromOptions(sf);CHKERRQ(ierr); 1962 ierr = PetscMemzero(o_nnz,na*sizeof(PetscInt));CHKERRQ(ierr); 1963 ierr = PetscSFReduceBegin(sf,MPIU_INT,g_nnz,o_nnz,MPIU_SUM);CHKERRQ(ierr); 1964 ierr = PetscSFReduceEnd(sf,MPIU_INT,g_nnz,o_nnz,MPIU_SUM);CHKERRQ(ierr); 1965 ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 1966 1967 ierr = MatCreate(PetscObjectComm((PetscObject)A),&B);CHKERRQ(ierr); 1968 ierr = MatSetSizes(B,A->cmap->n,A->rmap->n,N,M);CHKERRQ(ierr); 1969 ierr = MatSetBlockSizes(B,PetscAbs(A->cmap->bs),PetscAbs(A->rmap->bs));CHKERRQ(ierr); 1970 ierr = MatSetType(B,((PetscObject)A)->type_name);CHKERRQ(ierr); 1971 ierr = MatMPIAIJSetPreallocation(B,0,d_nnz,0,o_nnz);CHKERRQ(ierr); 1972 ierr = PetscFree4(d_nnz,o_nnz,g_nnz,oloc);CHKERRQ(ierr); 1973 } else { 1974 B = *matout; 1975 ierr = MatSetOption(B,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_TRUE);CHKERRQ(ierr); 1976 for (i=0; i<ai[ma]; i++) aj[i] += cstart; /* global col index to be used by MatSetValues() */ 1977 } 1978 1979 /* copy over the A part */ 1980 array = Aloc->a; 1981 row = A->rmap->rstart; 1982 for (i=0; i<ma; i++) { 1983 ncol = ai[i+1]-ai[i]; 1984 ierr = MatSetValues(B,ncol,aj,1,&row,array,INSERT_VALUES);CHKERRQ(ierr); 1985 row++; 1986 array += ncol; aj += ncol; 1987 } 1988 aj = Aloc->j; 1989 for (i=0; i<ai[ma]; i++) aj[i] -= cstart; /* resume local col index */ 1990 1991 /* copy over the B part */ 1992 ierr = PetscCalloc1(bi[mb],&cols);CHKERRQ(ierr); 1993 array = Bloc->a; 1994 row = A->rmap->rstart; 1995 for (i=0; i<bi[mb]; i++) cols[i] = a->garray[bj[i]]; 1996 cols_tmp = cols; 1997 for (i=0; i<mb; i++) { 1998 ncol = bi[i+1]-bi[i]; 1999 ierr = MatSetValues(B,ncol,cols_tmp,1,&row,array,INSERT_VALUES);CHKERRQ(ierr); 2000 row++; 2001 array += ncol; cols_tmp += ncol; 2002 } 2003 ierr = PetscFree(cols);CHKERRQ(ierr); 2004 2005 ierr = MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2006 ierr = MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2007 if (reuse == MAT_INITIAL_MATRIX || reuse == MAT_REUSE_MATRIX) { 2008 *matout = B; 2009 } else { 2010 ierr = MatHeaderMerge(A,&B);CHKERRQ(ierr); 2011 } 2012 PetscFunctionReturn(0); 2013 } 2014 2015 PetscErrorCode MatDiagonalScale_MPIAIJ(Mat mat,Vec ll,Vec rr) 2016 { 2017 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 2018 Mat a = aij->A,b = aij->B; 2019 PetscErrorCode ierr; 2020 PetscInt s1,s2,s3; 2021 2022 PetscFunctionBegin; 2023 ierr = MatGetLocalSize(mat,&s2,&s3);CHKERRQ(ierr); 2024 if (rr) { 2025 ierr = VecGetLocalSize(rr,&s1);CHKERRQ(ierr); 2026 if (s1!=s3) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"right vector non-conforming local size"); 2027 /* Overlap communication with computation. */ 2028 ierr = VecScatterBegin(aij->Mvctx,rr,aij->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2029 } 2030 if (ll) { 2031 ierr = VecGetLocalSize(ll,&s1);CHKERRQ(ierr); 2032 if (s1!=s2) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"left vector non-conforming local size"); 2033 ierr = (*b->ops->diagonalscale)(b,ll,0);CHKERRQ(ierr); 2034 } 2035 /* scale the diagonal block */ 2036 ierr = (*a->ops->diagonalscale)(a,ll,rr);CHKERRQ(ierr); 2037 2038 if (rr) { 2039 /* Do a scatter end and then right scale the off-diagonal block */ 2040 ierr = VecScatterEnd(aij->Mvctx,rr,aij->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2041 ierr = (*b->ops->diagonalscale)(b,0,aij->lvec);CHKERRQ(ierr); 2042 } 2043 PetscFunctionReturn(0); 2044 } 2045 2046 PetscErrorCode MatSetUnfactored_MPIAIJ(Mat A) 2047 { 2048 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 2049 PetscErrorCode ierr; 2050 2051 PetscFunctionBegin; 2052 ierr = MatSetUnfactored(a->A);CHKERRQ(ierr); 2053 PetscFunctionReturn(0); 2054 } 2055 2056 PetscErrorCode MatEqual_MPIAIJ(Mat A,Mat B,PetscBool *flag) 2057 { 2058 Mat_MPIAIJ *matB = (Mat_MPIAIJ*)B->data,*matA = (Mat_MPIAIJ*)A->data; 2059 Mat a,b,c,d; 2060 PetscBool flg; 2061 PetscErrorCode ierr; 2062 2063 PetscFunctionBegin; 2064 a = matA->A; b = matA->B; 2065 c = matB->A; d = matB->B; 2066 2067 ierr = MatEqual(a,c,&flg);CHKERRQ(ierr); 2068 if (flg) { 2069 ierr = MatEqual(b,d,&flg);CHKERRQ(ierr); 2070 } 2071 ierr = MPIU_Allreduce(&flg,flag,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)A));CHKERRQ(ierr); 2072 PetscFunctionReturn(0); 2073 } 2074 2075 PetscErrorCode MatCopy_MPIAIJ(Mat A,Mat B,MatStructure str) 2076 { 2077 PetscErrorCode ierr; 2078 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 2079 Mat_MPIAIJ *b = (Mat_MPIAIJ*)B->data; 2080 2081 PetscFunctionBegin; 2082 /* If the two matrices don't have the same copy implementation, they aren't compatible for fast copy. */ 2083 if ((str != SAME_NONZERO_PATTERN) || (A->ops->copy != B->ops->copy)) { 2084 /* because of the column compression in the off-processor part of the matrix a->B, 2085 the number of columns in a->B and b->B may be different, hence we cannot call 2086 the MatCopy() directly on the two parts. If need be, we can provide a more 2087 efficient copy than the MatCopy_Basic() by first uncompressing the a->B matrices 2088 then copying the submatrices */ 2089 ierr = MatCopy_Basic(A,B,str);CHKERRQ(ierr); 2090 } else { 2091 ierr = MatCopy(a->A,b->A,str);CHKERRQ(ierr); 2092 ierr = MatCopy(a->B,b->B,str);CHKERRQ(ierr); 2093 } 2094 ierr = PetscObjectStateIncrease((PetscObject)B);CHKERRQ(ierr); 2095 PetscFunctionReturn(0); 2096 } 2097 2098 PetscErrorCode MatSetUp_MPIAIJ(Mat A) 2099 { 2100 PetscErrorCode ierr; 2101 2102 PetscFunctionBegin; 2103 ierr = MatMPIAIJSetPreallocation(A,PETSC_DEFAULT,0,PETSC_DEFAULT,0);CHKERRQ(ierr); 2104 PetscFunctionReturn(0); 2105 } 2106 2107 /* 2108 Computes the number of nonzeros per row needed for preallocation when X and Y 2109 have different nonzero structure. 2110 */ 2111 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) 2112 { 2113 PetscInt i,j,k,nzx,nzy; 2114 2115 PetscFunctionBegin; 2116 /* Set the number of nonzeros in the new matrix */ 2117 for (i=0; i<m; i++) { 2118 const PetscInt *xjj = xj+xi[i],*yjj = yj+yi[i]; 2119 nzx = xi[i+1] - xi[i]; 2120 nzy = yi[i+1] - yi[i]; 2121 nnz[i] = 0; 2122 for (j=0,k=0; j<nzx; j++) { /* Point in X */ 2123 for (; k<nzy && yltog[yjj[k]]<xltog[xjj[j]]; k++) nnz[i]++; /* Catch up to X */ 2124 if (k<nzy && yltog[yjj[k]]==xltog[xjj[j]]) k++; /* Skip duplicate */ 2125 nnz[i]++; 2126 } 2127 for (; k<nzy; k++) nnz[i]++; 2128 } 2129 PetscFunctionReturn(0); 2130 } 2131 2132 /* This is the same as MatAXPYGetPreallocation_SeqAIJ, except that the local-to-global map is provided */ 2133 static PetscErrorCode MatAXPYGetPreallocation_MPIAIJ(Mat Y,const PetscInt *yltog,Mat X,const PetscInt *xltog,PetscInt *nnz) 2134 { 2135 PetscErrorCode ierr; 2136 PetscInt m = Y->rmap->N; 2137 Mat_SeqAIJ *x = (Mat_SeqAIJ*)X->data; 2138 Mat_SeqAIJ *y = (Mat_SeqAIJ*)Y->data; 2139 2140 PetscFunctionBegin; 2141 ierr = MatAXPYGetPreallocation_MPIX_private(m,x->i,x->j,xltog,y->i,y->j,yltog,nnz);CHKERRQ(ierr); 2142 PetscFunctionReturn(0); 2143 } 2144 2145 PetscErrorCode MatAXPY_MPIAIJ(Mat Y,PetscScalar a,Mat X,MatStructure str) 2146 { 2147 PetscErrorCode ierr; 2148 Mat_MPIAIJ *xx = (Mat_MPIAIJ*)X->data,*yy = (Mat_MPIAIJ*)Y->data; 2149 PetscBLASInt bnz,one=1; 2150 Mat_SeqAIJ *x,*y; 2151 2152 PetscFunctionBegin; 2153 if (str == SAME_NONZERO_PATTERN) { 2154 PetscScalar alpha = a; 2155 x = (Mat_SeqAIJ*)xx->A->data; 2156 ierr = PetscBLASIntCast(x->nz,&bnz);CHKERRQ(ierr); 2157 y = (Mat_SeqAIJ*)yy->A->data; 2158 PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&bnz,&alpha,x->a,&one,y->a,&one)); 2159 x = (Mat_SeqAIJ*)xx->B->data; 2160 y = (Mat_SeqAIJ*)yy->B->data; 2161 ierr = PetscBLASIntCast(x->nz,&bnz);CHKERRQ(ierr); 2162 PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&bnz,&alpha,x->a,&one,y->a,&one)); 2163 ierr = PetscObjectStateIncrease((PetscObject)Y);CHKERRQ(ierr); 2164 } else if (str == SUBSET_NONZERO_PATTERN) { /* nonzeros of X is a subset of Y's */ 2165 ierr = MatAXPY_Basic(Y,a,X,str);CHKERRQ(ierr); 2166 } else { 2167 Mat B; 2168 PetscInt *nnz_d,*nnz_o; 2169 ierr = PetscMalloc1(yy->A->rmap->N,&nnz_d);CHKERRQ(ierr); 2170 ierr = PetscMalloc1(yy->B->rmap->N,&nnz_o);CHKERRQ(ierr); 2171 ierr = MatCreate(PetscObjectComm((PetscObject)Y),&B);CHKERRQ(ierr); 2172 ierr = PetscObjectSetName((PetscObject)B,((PetscObject)Y)->name);CHKERRQ(ierr); 2173 ierr = MatSetSizes(B,Y->rmap->n,Y->cmap->n,Y->rmap->N,Y->cmap->N);CHKERRQ(ierr); 2174 ierr = MatSetBlockSizesFromMats(B,Y,Y);CHKERRQ(ierr); 2175 ierr = MatSetType(B,MATMPIAIJ);CHKERRQ(ierr); 2176 ierr = MatAXPYGetPreallocation_SeqAIJ(yy->A,xx->A,nnz_d);CHKERRQ(ierr); 2177 ierr = MatAXPYGetPreallocation_MPIAIJ(yy->B,yy->garray,xx->B,xx->garray,nnz_o);CHKERRQ(ierr); 2178 ierr = MatMPIAIJSetPreallocation(B,0,nnz_d,0,nnz_o);CHKERRQ(ierr); 2179 ierr = MatAXPY_BasicWithPreallocation(B,Y,a,X,str);CHKERRQ(ierr); 2180 ierr = MatHeaderReplace(Y,&B);CHKERRQ(ierr); 2181 ierr = PetscFree(nnz_d);CHKERRQ(ierr); 2182 ierr = PetscFree(nnz_o);CHKERRQ(ierr); 2183 } 2184 PetscFunctionReturn(0); 2185 } 2186 2187 extern PetscErrorCode MatConjugate_SeqAIJ(Mat); 2188 2189 PetscErrorCode MatConjugate_MPIAIJ(Mat mat) 2190 { 2191 #if defined(PETSC_USE_COMPLEX) 2192 PetscErrorCode ierr; 2193 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 2194 2195 PetscFunctionBegin; 2196 ierr = MatConjugate_SeqAIJ(aij->A);CHKERRQ(ierr); 2197 ierr = MatConjugate_SeqAIJ(aij->B);CHKERRQ(ierr); 2198 #else 2199 PetscFunctionBegin; 2200 #endif 2201 PetscFunctionReturn(0); 2202 } 2203 2204 PetscErrorCode MatRealPart_MPIAIJ(Mat A) 2205 { 2206 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 2207 PetscErrorCode ierr; 2208 2209 PetscFunctionBegin; 2210 ierr = MatRealPart(a->A);CHKERRQ(ierr); 2211 ierr = MatRealPart(a->B);CHKERRQ(ierr); 2212 PetscFunctionReturn(0); 2213 } 2214 2215 PetscErrorCode MatImaginaryPart_MPIAIJ(Mat A) 2216 { 2217 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 2218 PetscErrorCode ierr; 2219 2220 PetscFunctionBegin; 2221 ierr = MatImaginaryPart(a->A);CHKERRQ(ierr); 2222 ierr = MatImaginaryPart(a->B);CHKERRQ(ierr); 2223 PetscFunctionReturn(0); 2224 } 2225 2226 PetscErrorCode MatGetRowMaxAbs_MPIAIJ(Mat A, Vec v, PetscInt idx[]) 2227 { 2228 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 2229 PetscErrorCode ierr; 2230 PetscInt i,*idxb = 0; 2231 PetscScalar *va,*vb; 2232 Vec vtmp; 2233 2234 PetscFunctionBegin; 2235 ierr = MatGetRowMaxAbs(a->A,v,idx);CHKERRQ(ierr); 2236 ierr = VecGetArray(v,&va);CHKERRQ(ierr); 2237 if (idx) { 2238 for (i=0; i<A->rmap->n; i++) { 2239 if (PetscAbsScalar(va[i])) idx[i] += A->cmap->rstart; 2240 } 2241 } 2242 2243 ierr = VecCreateSeq(PETSC_COMM_SELF,A->rmap->n,&vtmp);CHKERRQ(ierr); 2244 if (idx) { 2245 ierr = PetscMalloc1(A->rmap->n,&idxb);CHKERRQ(ierr); 2246 } 2247 ierr = MatGetRowMaxAbs(a->B,vtmp,idxb);CHKERRQ(ierr); 2248 ierr = VecGetArray(vtmp,&vb);CHKERRQ(ierr); 2249 2250 for (i=0; i<A->rmap->n; i++) { 2251 if (PetscAbsScalar(va[i]) < PetscAbsScalar(vb[i])) { 2252 va[i] = vb[i]; 2253 if (idx) idx[i] = a->garray[idxb[i]]; 2254 } 2255 } 2256 2257 ierr = VecRestoreArray(v,&va);CHKERRQ(ierr); 2258 ierr = VecRestoreArray(vtmp,&vb);CHKERRQ(ierr); 2259 ierr = PetscFree(idxb);CHKERRQ(ierr); 2260 ierr = VecDestroy(&vtmp);CHKERRQ(ierr); 2261 PetscFunctionReturn(0); 2262 } 2263 2264 PetscErrorCode MatGetRowMinAbs_MPIAIJ(Mat A, Vec v, PetscInt idx[]) 2265 { 2266 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 2267 PetscErrorCode ierr; 2268 PetscInt i,*idxb = 0; 2269 PetscScalar *va,*vb; 2270 Vec vtmp; 2271 2272 PetscFunctionBegin; 2273 ierr = MatGetRowMinAbs(a->A,v,idx);CHKERRQ(ierr); 2274 ierr = VecGetArray(v,&va);CHKERRQ(ierr); 2275 if (idx) { 2276 for (i=0; i<A->cmap->n; i++) { 2277 if (PetscAbsScalar(va[i])) idx[i] += A->cmap->rstart; 2278 } 2279 } 2280 2281 ierr = VecCreateSeq(PETSC_COMM_SELF,A->rmap->n,&vtmp);CHKERRQ(ierr); 2282 if (idx) { 2283 ierr = PetscMalloc1(A->rmap->n,&idxb);CHKERRQ(ierr); 2284 } 2285 ierr = MatGetRowMinAbs(a->B,vtmp,idxb);CHKERRQ(ierr); 2286 ierr = VecGetArray(vtmp,&vb);CHKERRQ(ierr); 2287 2288 for (i=0; i<A->rmap->n; i++) { 2289 if (PetscAbsScalar(va[i]) > PetscAbsScalar(vb[i])) { 2290 va[i] = vb[i]; 2291 if (idx) idx[i] = a->garray[idxb[i]]; 2292 } 2293 } 2294 2295 ierr = VecRestoreArray(v,&va);CHKERRQ(ierr); 2296 ierr = VecRestoreArray(vtmp,&vb);CHKERRQ(ierr); 2297 ierr = PetscFree(idxb);CHKERRQ(ierr); 2298 ierr = VecDestroy(&vtmp);CHKERRQ(ierr); 2299 PetscFunctionReturn(0); 2300 } 2301 2302 PetscErrorCode MatGetRowMin_MPIAIJ(Mat A, Vec v, PetscInt idx[]) 2303 { 2304 Mat_MPIAIJ *mat = (Mat_MPIAIJ*) A->data; 2305 PetscInt n = A->rmap->n; 2306 PetscInt cstart = A->cmap->rstart; 2307 PetscInt *cmap = mat->garray; 2308 PetscInt *diagIdx, *offdiagIdx; 2309 Vec diagV, offdiagV; 2310 PetscScalar *a, *diagA, *offdiagA; 2311 PetscInt r; 2312 PetscErrorCode ierr; 2313 2314 PetscFunctionBegin; 2315 ierr = PetscMalloc2(n,&diagIdx,n,&offdiagIdx);CHKERRQ(ierr); 2316 ierr = VecCreateSeq(PetscObjectComm((PetscObject)A), n, &diagV);CHKERRQ(ierr); 2317 ierr = VecCreateSeq(PetscObjectComm((PetscObject)A), n, &offdiagV);CHKERRQ(ierr); 2318 ierr = MatGetRowMin(mat->A, diagV, diagIdx);CHKERRQ(ierr); 2319 ierr = MatGetRowMin(mat->B, offdiagV, offdiagIdx);CHKERRQ(ierr); 2320 ierr = VecGetArray(v, &a);CHKERRQ(ierr); 2321 ierr = VecGetArray(diagV, &diagA);CHKERRQ(ierr); 2322 ierr = VecGetArray(offdiagV, &offdiagA);CHKERRQ(ierr); 2323 for (r = 0; r < n; ++r) { 2324 if (PetscAbsScalar(diagA[r]) <= PetscAbsScalar(offdiagA[r])) { 2325 a[r] = diagA[r]; 2326 idx[r] = cstart + diagIdx[r]; 2327 } else { 2328 a[r] = offdiagA[r]; 2329 idx[r] = cmap[offdiagIdx[r]]; 2330 } 2331 } 2332 ierr = VecRestoreArray(v, &a);CHKERRQ(ierr); 2333 ierr = VecRestoreArray(diagV, &diagA);CHKERRQ(ierr); 2334 ierr = VecRestoreArray(offdiagV, &offdiagA);CHKERRQ(ierr); 2335 ierr = VecDestroy(&diagV);CHKERRQ(ierr); 2336 ierr = VecDestroy(&offdiagV);CHKERRQ(ierr); 2337 ierr = PetscFree2(diagIdx, offdiagIdx);CHKERRQ(ierr); 2338 PetscFunctionReturn(0); 2339 } 2340 2341 PetscErrorCode MatGetRowMax_MPIAIJ(Mat A, Vec v, PetscInt idx[]) 2342 { 2343 Mat_MPIAIJ *mat = (Mat_MPIAIJ*) A->data; 2344 PetscInt n = A->rmap->n; 2345 PetscInt cstart = A->cmap->rstart; 2346 PetscInt *cmap = mat->garray; 2347 PetscInt *diagIdx, *offdiagIdx; 2348 Vec diagV, offdiagV; 2349 PetscScalar *a, *diagA, *offdiagA; 2350 PetscInt r; 2351 PetscErrorCode ierr; 2352 2353 PetscFunctionBegin; 2354 ierr = PetscMalloc2(n,&diagIdx,n,&offdiagIdx);CHKERRQ(ierr); 2355 ierr = VecCreateSeq(PETSC_COMM_SELF, n, &diagV);CHKERRQ(ierr); 2356 ierr = VecCreateSeq(PETSC_COMM_SELF, n, &offdiagV);CHKERRQ(ierr); 2357 ierr = MatGetRowMax(mat->A, diagV, diagIdx);CHKERRQ(ierr); 2358 ierr = MatGetRowMax(mat->B, offdiagV, offdiagIdx);CHKERRQ(ierr); 2359 ierr = VecGetArray(v, &a);CHKERRQ(ierr); 2360 ierr = VecGetArray(diagV, &diagA);CHKERRQ(ierr); 2361 ierr = VecGetArray(offdiagV, &offdiagA);CHKERRQ(ierr); 2362 for (r = 0; r < n; ++r) { 2363 if (PetscAbsScalar(diagA[r]) >= PetscAbsScalar(offdiagA[r])) { 2364 a[r] = diagA[r]; 2365 idx[r] = cstart + diagIdx[r]; 2366 } else { 2367 a[r] = offdiagA[r]; 2368 idx[r] = cmap[offdiagIdx[r]]; 2369 } 2370 } 2371 ierr = VecRestoreArray(v, &a);CHKERRQ(ierr); 2372 ierr = VecRestoreArray(diagV, &diagA);CHKERRQ(ierr); 2373 ierr = VecRestoreArray(offdiagV, &offdiagA);CHKERRQ(ierr); 2374 ierr = VecDestroy(&diagV);CHKERRQ(ierr); 2375 ierr = VecDestroy(&offdiagV);CHKERRQ(ierr); 2376 ierr = PetscFree2(diagIdx, offdiagIdx);CHKERRQ(ierr); 2377 PetscFunctionReturn(0); 2378 } 2379 2380 PetscErrorCode MatGetSeqNonzeroStructure_MPIAIJ(Mat mat,Mat *newmat) 2381 { 2382 PetscErrorCode ierr; 2383 Mat *dummy; 2384 2385 PetscFunctionBegin; 2386 ierr = MatCreateSubMatrix_MPIAIJ_All(mat,MAT_DO_NOT_GET_VALUES,MAT_INITIAL_MATRIX,&dummy);CHKERRQ(ierr); 2387 *newmat = *dummy; 2388 ierr = PetscFree(dummy);CHKERRQ(ierr); 2389 PetscFunctionReturn(0); 2390 } 2391 2392 PetscErrorCode MatInvertBlockDiagonal_MPIAIJ(Mat A,const PetscScalar **values) 2393 { 2394 Mat_MPIAIJ *a = (Mat_MPIAIJ*) A->data; 2395 PetscErrorCode ierr; 2396 2397 PetscFunctionBegin; 2398 ierr = MatInvertBlockDiagonal(a->A,values);CHKERRQ(ierr); 2399 A->factorerrortype = a->A->factorerrortype; 2400 PetscFunctionReturn(0); 2401 } 2402 2403 static PetscErrorCode MatSetRandom_MPIAIJ(Mat x,PetscRandom rctx) 2404 { 2405 PetscErrorCode ierr; 2406 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)x->data; 2407 2408 PetscFunctionBegin; 2409 ierr = MatSetRandom(aij->A,rctx);CHKERRQ(ierr); 2410 ierr = MatSetRandom(aij->B,rctx);CHKERRQ(ierr); 2411 ierr = MatAssemblyBegin(x,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2412 ierr = MatAssemblyEnd(x,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2413 PetscFunctionReturn(0); 2414 } 2415 2416 PetscErrorCode MatMPIAIJSetUseScalableIncreaseOverlap_MPIAIJ(Mat A,PetscBool sc) 2417 { 2418 PetscFunctionBegin; 2419 if (sc) A->ops->increaseoverlap = MatIncreaseOverlap_MPIAIJ_Scalable; 2420 else A->ops->increaseoverlap = MatIncreaseOverlap_MPIAIJ; 2421 PetscFunctionReturn(0); 2422 } 2423 2424 /*@ 2425 MatMPIAIJSetUseScalableIncreaseOverlap - Determine if the matrix uses a scalable algorithm to compute the overlap 2426 2427 Collective on Mat 2428 2429 Input Parameters: 2430 + A - the matrix 2431 - sc - PETSC_TRUE indicates use the scalable algorithm (default is not to use the scalable algorithm) 2432 2433 Level: advanced 2434 2435 @*/ 2436 PetscErrorCode MatMPIAIJSetUseScalableIncreaseOverlap(Mat A,PetscBool sc) 2437 { 2438 PetscErrorCode ierr; 2439 2440 PetscFunctionBegin; 2441 ierr = PetscTryMethod(A,"MatMPIAIJSetUseScalableIncreaseOverlap_C",(Mat,PetscBool),(A,sc));CHKERRQ(ierr); 2442 PetscFunctionReturn(0); 2443 } 2444 2445 PetscErrorCode MatSetFromOptions_MPIAIJ(PetscOptionItems *PetscOptionsObject,Mat A) 2446 { 2447 PetscErrorCode ierr; 2448 PetscBool sc = PETSC_FALSE,flg; 2449 2450 PetscFunctionBegin; 2451 ierr = PetscOptionsHead(PetscOptionsObject,"MPIAIJ options");CHKERRQ(ierr); 2452 ierr = PetscObjectOptionsBegin((PetscObject)A); 2453 if (A->ops->increaseoverlap == MatIncreaseOverlap_MPIAIJ_Scalable) sc = PETSC_TRUE; 2454 ierr = PetscOptionsBool("-mat_increase_overlap_scalable","Use a scalable algorithm to compute the overlap","MatIncreaseOverlap",sc,&sc,&flg);CHKERRQ(ierr); 2455 if (flg) { 2456 ierr = MatMPIAIJSetUseScalableIncreaseOverlap(A,sc);CHKERRQ(ierr); 2457 } 2458 ierr = PetscOptionsEnd();CHKERRQ(ierr); 2459 PetscFunctionReturn(0); 2460 } 2461 2462 PetscErrorCode MatShift_MPIAIJ(Mat Y,PetscScalar a) 2463 { 2464 PetscErrorCode ierr; 2465 Mat_MPIAIJ *maij = (Mat_MPIAIJ*)Y->data; 2466 Mat_SeqAIJ *aij = (Mat_SeqAIJ*)maij->A->data; 2467 2468 PetscFunctionBegin; 2469 if (!Y->preallocated) { 2470 ierr = MatMPIAIJSetPreallocation(Y,1,NULL,0,NULL);CHKERRQ(ierr); 2471 } else if (!aij->nz) { 2472 PetscInt nonew = aij->nonew; 2473 ierr = MatSeqAIJSetPreallocation(maij->A,1,NULL);CHKERRQ(ierr); 2474 aij->nonew = nonew; 2475 } 2476 ierr = MatShift_Basic(Y,a);CHKERRQ(ierr); 2477 PetscFunctionReturn(0); 2478 } 2479 2480 PetscErrorCode MatMissingDiagonal_MPIAIJ(Mat A,PetscBool *missing,PetscInt *d) 2481 { 2482 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 2483 PetscErrorCode ierr; 2484 2485 PetscFunctionBegin; 2486 if (A->rmap->n != A->cmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Only works for square matrices"); 2487 ierr = MatMissingDiagonal(a->A,missing,d);CHKERRQ(ierr); 2488 if (d) { 2489 PetscInt rstart; 2490 ierr = MatGetOwnershipRange(A,&rstart,NULL);CHKERRQ(ierr); 2491 *d += rstart; 2492 2493 } 2494 PetscFunctionReturn(0); 2495 } 2496 2497 2498 /* -------------------------------------------------------------------*/ 2499 static struct _MatOps MatOps_Values = {MatSetValues_MPIAIJ, 2500 MatGetRow_MPIAIJ, 2501 MatRestoreRow_MPIAIJ, 2502 MatMult_MPIAIJ, 2503 /* 4*/ MatMultAdd_MPIAIJ, 2504 MatMultTranspose_MPIAIJ, 2505 MatMultTransposeAdd_MPIAIJ, 2506 0, 2507 0, 2508 0, 2509 /*10*/ 0, 2510 0, 2511 0, 2512 MatSOR_MPIAIJ, 2513 MatTranspose_MPIAIJ, 2514 /*15*/ MatGetInfo_MPIAIJ, 2515 MatEqual_MPIAIJ, 2516 MatGetDiagonal_MPIAIJ, 2517 MatDiagonalScale_MPIAIJ, 2518 MatNorm_MPIAIJ, 2519 /*20*/ MatAssemblyBegin_MPIAIJ, 2520 MatAssemblyEnd_MPIAIJ, 2521 MatSetOption_MPIAIJ, 2522 MatZeroEntries_MPIAIJ, 2523 /*24*/ MatZeroRows_MPIAIJ, 2524 0, 2525 0, 2526 0, 2527 0, 2528 /*29*/ MatSetUp_MPIAIJ, 2529 0, 2530 0, 2531 MatGetDiagonalBlock_MPIAIJ, 2532 0, 2533 /*34*/ MatDuplicate_MPIAIJ, 2534 0, 2535 0, 2536 0, 2537 0, 2538 /*39*/ MatAXPY_MPIAIJ, 2539 MatCreateSubMatrices_MPIAIJ, 2540 MatIncreaseOverlap_MPIAIJ, 2541 MatGetValues_MPIAIJ, 2542 MatCopy_MPIAIJ, 2543 /*44*/ MatGetRowMax_MPIAIJ, 2544 MatScale_MPIAIJ, 2545 MatShift_MPIAIJ, 2546 MatDiagonalSet_MPIAIJ, 2547 MatZeroRowsColumns_MPIAIJ, 2548 /*49*/ MatSetRandom_MPIAIJ, 2549 0, 2550 0, 2551 0, 2552 0, 2553 /*54*/ MatFDColoringCreate_MPIXAIJ, 2554 0, 2555 MatSetUnfactored_MPIAIJ, 2556 MatPermute_MPIAIJ, 2557 0, 2558 /*59*/ MatCreateSubMatrix_MPIAIJ, 2559 MatDestroy_MPIAIJ, 2560 MatView_MPIAIJ, 2561 0, 2562 MatMatMatMult_MPIAIJ_MPIAIJ_MPIAIJ, 2563 /*64*/ MatMatMatMultSymbolic_MPIAIJ_MPIAIJ_MPIAIJ, 2564 MatMatMatMultNumeric_MPIAIJ_MPIAIJ_MPIAIJ, 2565 0, 2566 0, 2567 0, 2568 /*69*/ MatGetRowMaxAbs_MPIAIJ, 2569 MatGetRowMinAbs_MPIAIJ, 2570 0, 2571 0, 2572 0, 2573 0, 2574 /*75*/ MatFDColoringApply_AIJ, 2575 MatSetFromOptions_MPIAIJ, 2576 0, 2577 0, 2578 MatFindZeroDiagonals_MPIAIJ, 2579 /*80*/ 0, 2580 0, 2581 0, 2582 /*83*/ MatLoad_MPIAIJ, 2583 MatIsSymmetric_MPIAIJ, 2584 0, 2585 0, 2586 0, 2587 0, 2588 /*89*/ MatMatMult_MPIAIJ_MPIAIJ, 2589 MatMatMultSymbolic_MPIAIJ_MPIAIJ, 2590 MatMatMultNumeric_MPIAIJ_MPIAIJ, 2591 MatPtAP_MPIAIJ_MPIAIJ, 2592 MatPtAPSymbolic_MPIAIJ_MPIAIJ, 2593 /*94*/ MatPtAPNumeric_MPIAIJ_MPIAIJ, 2594 0, 2595 0, 2596 0, 2597 0, 2598 /*99*/ 0, 2599 0, 2600 0, 2601 MatConjugate_MPIAIJ, 2602 0, 2603 /*104*/MatSetValuesRow_MPIAIJ, 2604 MatRealPart_MPIAIJ, 2605 MatImaginaryPart_MPIAIJ, 2606 0, 2607 0, 2608 /*109*/0, 2609 0, 2610 MatGetRowMin_MPIAIJ, 2611 0, 2612 MatMissingDiagonal_MPIAIJ, 2613 /*114*/MatGetSeqNonzeroStructure_MPIAIJ, 2614 0, 2615 MatGetGhosts_MPIAIJ, 2616 0, 2617 0, 2618 /*119*/0, 2619 0, 2620 0, 2621 0, 2622 MatGetMultiProcBlock_MPIAIJ, 2623 /*124*/MatFindNonzeroRows_MPIAIJ, 2624 MatGetColumnNorms_MPIAIJ, 2625 MatInvertBlockDiagonal_MPIAIJ, 2626 0, 2627 MatCreateSubMatricesMPI_MPIAIJ, 2628 /*129*/0, 2629 MatTransposeMatMult_MPIAIJ_MPIAIJ, 2630 MatTransposeMatMultSymbolic_MPIAIJ_MPIAIJ, 2631 MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ, 2632 0, 2633 /*134*/0, 2634 0, 2635 MatRARt_MPIAIJ_MPIAIJ, 2636 0, 2637 0, 2638 /*139*/MatSetBlockSizes_MPIAIJ, 2639 0, 2640 0, 2641 MatFDColoringSetUp_MPIXAIJ, 2642 MatFindOffBlockDiagonalEntries_MPIAIJ, 2643 /*144*/MatCreateMPIMatConcatenateSeqMat_MPIAIJ 2644 }; 2645 2646 /* ----------------------------------------------------------------------------------------*/ 2647 2648 PetscErrorCode MatStoreValues_MPIAIJ(Mat mat) 2649 { 2650 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 2651 PetscErrorCode ierr; 2652 2653 PetscFunctionBegin; 2654 ierr = MatStoreValues(aij->A);CHKERRQ(ierr); 2655 ierr = MatStoreValues(aij->B);CHKERRQ(ierr); 2656 PetscFunctionReturn(0); 2657 } 2658 2659 PetscErrorCode MatRetrieveValues_MPIAIJ(Mat mat) 2660 { 2661 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 2662 PetscErrorCode ierr; 2663 2664 PetscFunctionBegin; 2665 ierr = MatRetrieveValues(aij->A);CHKERRQ(ierr); 2666 ierr = MatRetrieveValues(aij->B);CHKERRQ(ierr); 2667 PetscFunctionReturn(0); 2668 } 2669 2670 PetscErrorCode MatMPIAIJSetPreallocation_MPIAIJ(Mat B,PetscInt d_nz,const PetscInt d_nnz[],PetscInt o_nz,const PetscInt o_nnz[]) 2671 { 2672 Mat_MPIAIJ *b; 2673 PetscErrorCode ierr; 2674 2675 PetscFunctionBegin; 2676 ierr = PetscLayoutSetUp(B->rmap);CHKERRQ(ierr); 2677 ierr = PetscLayoutSetUp(B->cmap);CHKERRQ(ierr); 2678 b = (Mat_MPIAIJ*)B->data; 2679 2680 #if defined(PETSC_USE_CTABLE) 2681 ierr = PetscTableDestroy(&b->colmap);CHKERRQ(ierr); 2682 #else 2683 ierr = PetscFree(b->colmap);CHKERRQ(ierr); 2684 #endif 2685 ierr = PetscFree(b->garray);CHKERRQ(ierr); 2686 ierr = VecDestroy(&b->lvec);CHKERRQ(ierr); 2687 ierr = VecScatterDestroy(&b->Mvctx);CHKERRQ(ierr); 2688 2689 /* Because the B will have been resized we simply destroy it and create a new one each time */ 2690 ierr = MatDestroy(&b->B);CHKERRQ(ierr); 2691 ierr = MatCreate(PETSC_COMM_SELF,&b->B);CHKERRQ(ierr); 2692 ierr = MatSetSizes(b->B,B->rmap->n,B->cmap->N,B->rmap->n,B->cmap->N);CHKERRQ(ierr); 2693 ierr = MatSetBlockSizesFromMats(b->B,B,B);CHKERRQ(ierr); 2694 ierr = MatSetType(b->B,MATSEQAIJ);CHKERRQ(ierr); 2695 ierr = PetscLogObjectParent((PetscObject)B,(PetscObject)b->B);CHKERRQ(ierr); 2696 2697 if (!B->preallocated) { 2698 ierr = MatCreate(PETSC_COMM_SELF,&b->A);CHKERRQ(ierr); 2699 ierr = MatSetSizes(b->A,B->rmap->n,B->cmap->n,B->rmap->n,B->cmap->n);CHKERRQ(ierr); 2700 ierr = MatSetBlockSizesFromMats(b->A,B,B);CHKERRQ(ierr); 2701 ierr = MatSetType(b->A,MATSEQAIJ);CHKERRQ(ierr); 2702 ierr = PetscLogObjectParent((PetscObject)B,(PetscObject)b->A);CHKERRQ(ierr); 2703 } 2704 2705 ierr = MatSeqAIJSetPreallocation(b->A,d_nz,d_nnz);CHKERRQ(ierr); 2706 ierr = MatSeqAIJSetPreallocation(b->B,o_nz,o_nnz);CHKERRQ(ierr); 2707 B->preallocated = PETSC_TRUE; 2708 B->was_assembled = PETSC_FALSE; 2709 B->assembled = PETSC_FALSE;; 2710 PetscFunctionReturn(0); 2711 } 2712 2713 PetscErrorCode MatResetPreallocation_MPIAIJ(Mat B) 2714 { 2715 Mat_MPIAIJ *b; 2716 PetscErrorCode ierr; 2717 2718 PetscFunctionBegin; 2719 PetscValidHeaderSpecific(B,MAT_CLASSID,1); 2720 ierr = PetscLayoutSetUp(B->rmap);CHKERRQ(ierr); 2721 ierr = PetscLayoutSetUp(B->cmap);CHKERRQ(ierr); 2722 b = (Mat_MPIAIJ*)B->data; 2723 2724 #if defined(PETSC_USE_CTABLE) 2725 ierr = PetscTableDestroy(&b->colmap);CHKERRQ(ierr); 2726 #else 2727 ierr = PetscFree(b->colmap);CHKERRQ(ierr); 2728 #endif 2729 ierr = PetscFree(b->garray);CHKERRQ(ierr); 2730 ierr = VecDestroy(&b->lvec);CHKERRQ(ierr); 2731 ierr = VecScatterDestroy(&b->Mvctx);CHKERRQ(ierr); 2732 2733 ierr = MatResetPreallocation(b->A);CHKERRQ(ierr); 2734 ierr = MatResetPreallocation(b->B);CHKERRQ(ierr); 2735 B->preallocated = PETSC_TRUE; 2736 B->was_assembled = PETSC_FALSE; 2737 B->assembled = PETSC_FALSE; 2738 PetscFunctionReturn(0); 2739 } 2740 2741 PetscErrorCode MatDuplicate_MPIAIJ(Mat matin,MatDuplicateOption cpvalues,Mat *newmat) 2742 { 2743 Mat mat; 2744 Mat_MPIAIJ *a,*oldmat = (Mat_MPIAIJ*)matin->data; 2745 PetscErrorCode ierr; 2746 2747 PetscFunctionBegin; 2748 *newmat = 0; 2749 ierr = MatCreate(PetscObjectComm((PetscObject)matin),&mat);CHKERRQ(ierr); 2750 ierr = MatSetSizes(mat,matin->rmap->n,matin->cmap->n,matin->rmap->N,matin->cmap->N);CHKERRQ(ierr); 2751 ierr = MatSetBlockSizesFromMats(mat,matin,matin);CHKERRQ(ierr); 2752 ierr = MatSetType(mat,((PetscObject)matin)->type_name);CHKERRQ(ierr); 2753 ierr = PetscMemcpy(mat->ops,matin->ops,sizeof(struct _MatOps));CHKERRQ(ierr); 2754 a = (Mat_MPIAIJ*)mat->data; 2755 2756 mat->factortype = matin->factortype; 2757 mat->assembled = PETSC_TRUE; 2758 mat->insertmode = NOT_SET_VALUES; 2759 mat->preallocated = PETSC_TRUE; 2760 2761 a->size = oldmat->size; 2762 a->rank = oldmat->rank; 2763 a->donotstash = oldmat->donotstash; 2764 a->roworiented = oldmat->roworiented; 2765 a->rowindices = 0; 2766 a->rowvalues = 0; 2767 a->getrowactive = PETSC_FALSE; 2768 2769 ierr = PetscLayoutReference(matin->rmap,&mat->rmap);CHKERRQ(ierr); 2770 ierr = PetscLayoutReference(matin->cmap,&mat->cmap);CHKERRQ(ierr); 2771 2772 if (oldmat->colmap) { 2773 #if defined(PETSC_USE_CTABLE) 2774 ierr = PetscTableCreateCopy(oldmat->colmap,&a->colmap);CHKERRQ(ierr); 2775 #else 2776 ierr = PetscMalloc1(mat->cmap->N,&a->colmap);CHKERRQ(ierr); 2777 ierr = PetscLogObjectMemory((PetscObject)mat,(mat->cmap->N)*sizeof(PetscInt));CHKERRQ(ierr); 2778 ierr = PetscMemcpy(a->colmap,oldmat->colmap,(mat->cmap->N)*sizeof(PetscInt));CHKERRQ(ierr); 2779 #endif 2780 } else a->colmap = 0; 2781 if (oldmat->garray) { 2782 PetscInt len; 2783 len = oldmat->B->cmap->n; 2784 ierr = PetscMalloc1(len+1,&a->garray);CHKERRQ(ierr); 2785 ierr = PetscLogObjectMemory((PetscObject)mat,len*sizeof(PetscInt));CHKERRQ(ierr); 2786 if (len) { ierr = PetscMemcpy(a->garray,oldmat->garray,len*sizeof(PetscInt));CHKERRQ(ierr); } 2787 } else a->garray = 0; 2788 2789 ierr = VecDuplicate(oldmat->lvec,&a->lvec);CHKERRQ(ierr); 2790 ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)a->lvec);CHKERRQ(ierr); 2791 ierr = VecScatterCopy(oldmat->Mvctx,&a->Mvctx);CHKERRQ(ierr); 2792 ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)a->Mvctx);CHKERRQ(ierr); 2793 2794 if (oldmat->Mvctx_mpi1) { 2795 ierr = VecScatterCopy(oldmat->Mvctx_mpi1,&a->Mvctx_mpi1);CHKERRQ(ierr); 2796 ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)a->Mvctx_mpi1);CHKERRQ(ierr); 2797 } 2798 2799 ierr = MatDuplicate(oldmat->A,cpvalues,&a->A);CHKERRQ(ierr); 2800 ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)a->A);CHKERRQ(ierr); 2801 ierr = MatDuplicate(oldmat->B,cpvalues,&a->B);CHKERRQ(ierr); 2802 ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)a->B);CHKERRQ(ierr); 2803 ierr = PetscFunctionListDuplicate(((PetscObject)matin)->qlist,&((PetscObject)mat)->qlist);CHKERRQ(ierr); 2804 *newmat = mat; 2805 PetscFunctionReturn(0); 2806 } 2807 2808 PetscErrorCode MatLoad_MPIAIJ(Mat newMat, PetscViewer viewer) 2809 { 2810 PetscScalar *vals,*svals; 2811 MPI_Comm comm; 2812 PetscErrorCode ierr; 2813 PetscMPIInt rank,size,tag = ((PetscObject)viewer)->tag; 2814 PetscInt i,nz,j,rstart,rend,mmax,maxnz = 0; 2815 PetscInt header[4],*rowlengths = 0,M,N,m,*cols; 2816 PetscInt *ourlens = NULL,*procsnz = NULL,*offlens = NULL,jj,*mycols,*smycols; 2817 PetscInt cend,cstart,n,*rowners; 2818 int fd; 2819 PetscInt bs = newMat->rmap->bs; 2820 2821 PetscFunctionBegin; 2822 /* force binary viewer to load .info file if it has not yet done so */ 2823 ierr = PetscViewerSetUp(viewer);CHKERRQ(ierr); 2824 ierr = PetscObjectGetComm((PetscObject)viewer,&comm);CHKERRQ(ierr); 2825 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 2826 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 2827 ierr = PetscViewerBinaryGetDescriptor(viewer,&fd);CHKERRQ(ierr); 2828 if (!rank) { 2829 ierr = PetscBinaryRead(fd,(char*)header,4,PETSC_INT);CHKERRQ(ierr); 2830 if (header[0] != MAT_FILE_CLASSID) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED,"not matrix object"); 2831 if (header[3] < 0) SETERRQ(PetscObjectComm((PetscObject)newMat),PETSC_ERR_FILE_UNEXPECTED,"Matrix stored in special format on disk,cannot load as MATMPIAIJ"); 2832 } 2833 2834 ierr = PetscOptionsBegin(comm,NULL,"Options for loading MATMPIAIJ matrix","Mat");CHKERRQ(ierr); 2835 ierr = PetscOptionsInt("-matload_block_size","Set the blocksize used to store the matrix","MatLoad",bs,&bs,NULL);CHKERRQ(ierr); 2836 ierr = PetscOptionsEnd();CHKERRQ(ierr); 2837 if (bs < 0) bs = 1; 2838 2839 ierr = MPI_Bcast(header+1,3,MPIU_INT,0,comm);CHKERRQ(ierr); 2840 M = header[1]; N = header[2]; 2841 2842 /* If global sizes are set, check if they are consistent with that given in the file */ 2843 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); 2844 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); 2845 2846 /* determine ownership of all (block) rows */ 2847 if (M%bs) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED, "Inconsistent # of rows (%d) and block size (%d)",M,bs); 2848 if (newMat->rmap->n < 0) m = bs*((M/bs)/size + (((M/bs) % size) > rank)); /* PETSC_DECIDE */ 2849 else m = newMat->rmap->n; /* Set by user */ 2850 2851 ierr = PetscMalloc1(size+1,&rowners);CHKERRQ(ierr); 2852 ierr = MPI_Allgather(&m,1,MPIU_INT,rowners+1,1,MPIU_INT,comm);CHKERRQ(ierr); 2853 2854 /* First process needs enough room for process with most rows */ 2855 if (!rank) { 2856 mmax = rowners[1]; 2857 for (i=2; i<=size; i++) { 2858 mmax = PetscMax(mmax, rowners[i]); 2859 } 2860 } else mmax = -1; /* unused, but compilers complain */ 2861 2862 rowners[0] = 0; 2863 for (i=2; i<=size; i++) { 2864 rowners[i] += rowners[i-1]; 2865 } 2866 rstart = rowners[rank]; 2867 rend = rowners[rank+1]; 2868 2869 /* distribute row lengths to all processors */ 2870 ierr = PetscMalloc2(m,&ourlens,m,&offlens);CHKERRQ(ierr); 2871 if (!rank) { 2872 ierr = PetscBinaryRead(fd,ourlens,m,PETSC_INT);CHKERRQ(ierr); 2873 ierr = PetscMalloc1(mmax,&rowlengths);CHKERRQ(ierr); 2874 ierr = PetscCalloc1(size,&procsnz);CHKERRQ(ierr); 2875 for (j=0; j<m; j++) { 2876 procsnz[0] += ourlens[j]; 2877 } 2878 for (i=1; i<size; i++) { 2879 ierr = PetscBinaryRead(fd,rowlengths,rowners[i+1]-rowners[i],PETSC_INT);CHKERRQ(ierr); 2880 /* calculate the number of nonzeros on each processor */ 2881 for (j=0; j<rowners[i+1]-rowners[i]; j++) { 2882 procsnz[i] += rowlengths[j]; 2883 } 2884 ierr = MPIULong_Send(rowlengths,rowners[i+1]-rowners[i],MPIU_INT,i,tag,comm);CHKERRQ(ierr); 2885 } 2886 ierr = PetscFree(rowlengths);CHKERRQ(ierr); 2887 } else { 2888 ierr = MPIULong_Recv(ourlens,m,MPIU_INT,0,tag,comm);CHKERRQ(ierr); 2889 } 2890 2891 if (!rank) { 2892 /* determine max buffer needed and allocate it */ 2893 maxnz = 0; 2894 for (i=0; i<size; i++) { 2895 maxnz = PetscMax(maxnz,procsnz[i]); 2896 } 2897 ierr = PetscMalloc1(maxnz,&cols);CHKERRQ(ierr); 2898 2899 /* read in my part of the matrix column indices */ 2900 nz = procsnz[0]; 2901 ierr = PetscMalloc1(nz,&mycols);CHKERRQ(ierr); 2902 ierr = PetscBinaryRead(fd,mycols,nz,PETSC_INT);CHKERRQ(ierr); 2903 2904 /* read in every one elses and ship off */ 2905 for (i=1; i<size; i++) { 2906 nz = procsnz[i]; 2907 ierr = PetscBinaryRead(fd,cols,nz,PETSC_INT);CHKERRQ(ierr); 2908 ierr = MPIULong_Send(cols,nz,MPIU_INT,i,tag,comm);CHKERRQ(ierr); 2909 } 2910 ierr = PetscFree(cols);CHKERRQ(ierr); 2911 } else { 2912 /* determine buffer space needed for message */ 2913 nz = 0; 2914 for (i=0; i<m; i++) { 2915 nz += ourlens[i]; 2916 } 2917 ierr = PetscMalloc1(nz,&mycols);CHKERRQ(ierr); 2918 2919 /* receive message of column indices*/ 2920 ierr = MPIULong_Recv(mycols,nz,MPIU_INT,0,tag,comm);CHKERRQ(ierr); 2921 } 2922 2923 /* determine column ownership if matrix is not square */ 2924 if (N != M) { 2925 if (newMat->cmap->n < 0) n = N/size + ((N % size) > rank); 2926 else n = newMat->cmap->n; 2927 ierr = MPI_Scan(&n,&cend,1,MPIU_INT,MPI_SUM,comm);CHKERRQ(ierr); 2928 cstart = cend - n; 2929 } else { 2930 cstart = rstart; 2931 cend = rend; 2932 n = cend - cstart; 2933 } 2934 2935 /* loop over local rows, determining number of off diagonal entries */ 2936 ierr = PetscMemzero(offlens,m*sizeof(PetscInt));CHKERRQ(ierr); 2937 jj = 0; 2938 for (i=0; i<m; i++) { 2939 for (j=0; j<ourlens[i]; j++) { 2940 if (mycols[jj] < cstart || mycols[jj] >= cend) offlens[i]++; 2941 jj++; 2942 } 2943 } 2944 2945 for (i=0; i<m; i++) { 2946 ourlens[i] -= offlens[i]; 2947 } 2948 ierr = MatSetSizes(newMat,m,n,M,N);CHKERRQ(ierr); 2949 2950 if (bs > 1) {ierr = MatSetBlockSize(newMat,bs);CHKERRQ(ierr);} 2951 2952 ierr = MatMPIAIJSetPreallocation(newMat,0,ourlens,0,offlens);CHKERRQ(ierr); 2953 2954 for (i=0; i<m; i++) { 2955 ourlens[i] += offlens[i]; 2956 } 2957 2958 if (!rank) { 2959 ierr = PetscMalloc1(maxnz+1,&vals);CHKERRQ(ierr); 2960 2961 /* read in my part of the matrix numerical values */ 2962 nz = procsnz[0]; 2963 ierr = PetscBinaryRead(fd,vals,nz,PETSC_SCALAR);CHKERRQ(ierr); 2964 2965 /* insert into matrix */ 2966 jj = rstart; 2967 smycols = mycols; 2968 svals = vals; 2969 for (i=0; i<m; i++) { 2970 ierr = MatSetValues_MPIAIJ(newMat,1,&jj,ourlens[i],smycols,svals,INSERT_VALUES);CHKERRQ(ierr); 2971 smycols += ourlens[i]; 2972 svals += ourlens[i]; 2973 jj++; 2974 } 2975 2976 /* read in other processors and ship out */ 2977 for (i=1; i<size; i++) { 2978 nz = procsnz[i]; 2979 ierr = PetscBinaryRead(fd,vals,nz,PETSC_SCALAR);CHKERRQ(ierr); 2980 ierr = MPIULong_Send(vals,nz,MPIU_SCALAR,i,((PetscObject)newMat)->tag,comm);CHKERRQ(ierr); 2981 } 2982 ierr = PetscFree(procsnz);CHKERRQ(ierr); 2983 } else { 2984 /* receive numeric values */ 2985 ierr = PetscMalloc1(nz+1,&vals);CHKERRQ(ierr); 2986 2987 /* receive message of values*/ 2988 ierr = MPIULong_Recv(vals,nz,MPIU_SCALAR,0,((PetscObject)newMat)->tag,comm);CHKERRQ(ierr); 2989 2990 /* insert into matrix */ 2991 jj = rstart; 2992 smycols = mycols; 2993 svals = vals; 2994 for (i=0; i<m; i++) { 2995 ierr = MatSetValues_MPIAIJ(newMat,1,&jj,ourlens[i],smycols,svals,INSERT_VALUES);CHKERRQ(ierr); 2996 smycols += ourlens[i]; 2997 svals += ourlens[i]; 2998 jj++; 2999 } 3000 } 3001 ierr = PetscFree2(ourlens,offlens);CHKERRQ(ierr); 3002 ierr = PetscFree(vals);CHKERRQ(ierr); 3003 ierr = PetscFree(mycols);CHKERRQ(ierr); 3004 ierr = PetscFree(rowners);CHKERRQ(ierr); 3005 ierr = MatAssemblyBegin(newMat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3006 ierr = MatAssemblyEnd(newMat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3007 PetscFunctionReturn(0); 3008 } 3009 3010 /* Not scalable because of ISAllGather() unless getting all columns. */ 3011 PetscErrorCode ISGetSeqIS_Private(Mat mat,IS iscol,IS *isseq) 3012 { 3013 PetscErrorCode ierr; 3014 IS iscol_local; 3015 PetscBool isstride; 3016 PetscMPIInt lisstride=0,gisstride; 3017 3018 PetscFunctionBegin; 3019 /* check if we are grabbing all columns*/ 3020 ierr = PetscObjectTypeCompare((PetscObject)iscol,ISSTRIDE,&isstride);CHKERRQ(ierr); 3021 3022 if (isstride) { 3023 PetscInt start,len,mstart,mlen; 3024 ierr = ISStrideGetInfo(iscol,&start,NULL);CHKERRQ(ierr); 3025 ierr = ISGetLocalSize(iscol,&len);CHKERRQ(ierr); 3026 ierr = MatGetOwnershipRangeColumn(mat,&mstart,&mlen);CHKERRQ(ierr); 3027 if (mstart == start && mlen-mstart == len) lisstride = 1; 3028 } 3029 3030 ierr = MPIU_Allreduce(&lisstride,&gisstride,1,MPI_INT,MPI_MIN,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 3031 if (gisstride) { 3032 PetscInt N; 3033 ierr = MatGetSize(mat,NULL,&N);CHKERRQ(ierr); 3034 ierr = ISCreateStride(PetscObjectComm((PetscObject)mat),N,0,1,&iscol_local);CHKERRQ(ierr); 3035 ierr = ISSetIdentity(iscol_local);CHKERRQ(ierr); 3036 ierr = PetscInfo(mat,"Optimizing for obtaining all columns of the matrix; skipping ISAllGather()\n");CHKERRQ(ierr); 3037 } else { 3038 PetscInt cbs; 3039 ierr = ISGetBlockSize(iscol,&cbs);CHKERRQ(ierr); 3040 ierr = ISAllGather(iscol,&iscol_local);CHKERRQ(ierr); 3041 ierr = ISSetBlockSize(iscol_local,cbs);CHKERRQ(ierr); 3042 } 3043 3044 *isseq = iscol_local; 3045 PetscFunctionReturn(0); 3046 } 3047 3048 /* 3049 Used by MatCreateSubMatrix_MPIAIJ_SameRowColDist() to avoid ISAllGather() and global size of iscol_local 3050 (see MatCreateSubMatrix_MPIAIJ_nonscalable) 3051 3052 Input Parameters: 3053 mat - matrix 3054 isrow - parallel row index set; its local indices are a subset of local columns of mat, 3055 i.e., mat->rstart <= isrow[i] < mat->rend 3056 iscol - parallel column index set; its local indices are a subset of local columns of mat, 3057 i.e., mat->cstart <= iscol[i] < mat->cend 3058 Output Parameter: 3059 isrow_d,iscol_d - sequential row and column index sets for retrieving mat->A 3060 iscol_o - sequential column index set for retrieving mat->B 3061 garray - column map; garray[i] indicates global location of iscol_o[i] in iscol 3062 */ 3063 PetscErrorCode ISGetSeqIS_SameColDist_Private(Mat mat,IS isrow,IS iscol,IS *isrow_d,IS *iscol_d,IS *iscol_o,const PetscInt *garray[]) 3064 { 3065 PetscErrorCode ierr; 3066 Vec x,cmap; 3067 const PetscInt *is_idx; 3068 PetscScalar *xarray,*cmaparray; 3069 PetscInt ncols,isstart,*idx,m,rstart,*cmap1,count; 3070 Mat_MPIAIJ *a=(Mat_MPIAIJ*)mat->data; 3071 Mat B=a->B; 3072 Vec lvec=a->lvec,lcmap; 3073 PetscInt i,cstart,cend,Bn=B->cmap->N; 3074 MPI_Comm comm; 3075 VecScatter Mvctx=a->Mvctx; 3076 3077 PetscFunctionBegin; 3078 ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 3079 ierr = ISGetLocalSize(iscol,&ncols);CHKERRQ(ierr); 3080 3081 /* (1) iscol is a sub-column vector of mat, pad it with '-1.' to form a full vector x */ 3082 ierr = MatCreateVecs(mat,&x,NULL);CHKERRQ(ierr); 3083 ierr = VecSet(x,-1.0);CHKERRQ(ierr); 3084 ierr = VecDuplicate(x,&cmap);CHKERRQ(ierr); 3085 ierr = VecSet(cmap,-1.0);CHKERRQ(ierr); 3086 3087 /* Get start indices */ 3088 ierr = MPI_Scan(&ncols,&isstart,1,MPIU_INT,MPI_SUM,comm);CHKERRQ(ierr); 3089 isstart -= ncols; 3090 ierr = MatGetOwnershipRangeColumn(mat,&cstart,&cend);CHKERRQ(ierr); 3091 3092 ierr = ISGetIndices(iscol,&is_idx);CHKERRQ(ierr); 3093 ierr = VecGetArray(x,&xarray);CHKERRQ(ierr); 3094 ierr = VecGetArray(cmap,&cmaparray);CHKERRQ(ierr); 3095 ierr = PetscMalloc1(ncols,&idx);CHKERRQ(ierr); 3096 for (i=0; i<ncols; i++) { 3097 xarray[is_idx[i]-cstart] = (PetscScalar)is_idx[i]; 3098 cmaparray[is_idx[i]-cstart] = i + isstart; /* global index of iscol[i] */ 3099 idx[i] = is_idx[i]-cstart; /* local index of iscol[i] */ 3100 } 3101 ierr = VecRestoreArray(x,&xarray);CHKERRQ(ierr); 3102 ierr = VecRestoreArray(cmap,&cmaparray);CHKERRQ(ierr); 3103 ierr = ISRestoreIndices(iscol,&is_idx);CHKERRQ(ierr); 3104 3105 /* Get iscol_d */ 3106 ierr = ISCreateGeneral(PETSC_COMM_SELF,ncols,idx,PETSC_OWN_POINTER,iscol_d);CHKERRQ(ierr); 3107 ierr = ISGetBlockSize(iscol,&i);CHKERRQ(ierr); 3108 ierr = ISSetBlockSize(*iscol_d,i);CHKERRQ(ierr); 3109 3110 /* Get isrow_d */ 3111 ierr = ISGetLocalSize(isrow,&m);CHKERRQ(ierr); 3112 rstart = mat->rmap->rstart; 3113 ierr = PetscMalloc1(m,&idx);CHKERRQ(ierr); 3114 ierr = ISGetIndices(isrow,&is_idx);CHKERRQ(ierr); 3115 for (i=0; i<m; i++) idx[i] = is_idx[i]-rstart; 3116 ierr = ISRestoreIndices(isrow,&is_idx);CHKERRQ(ierr); 3117 3118 ierr = ISCreateGeneral(PETSC_COMM_SELF,m,idx,PETSC_OWN_POINTER,isrow_d);CHKERRQ(ierr); 3119 ierr = ISGetBlockSize(isrow,&i);CHKERRQ(ierr); 3120 ierr = ISSetBlockSize(*isrow_d,i);CHKERRQ(ierr); 3121 3122 /* (2) Scatter x and cmap using aij->Mvctx to get their off-process portions (see MatMult_MPIAIJ) */ 3123 ierr = VecScatterBegin(Mvctx,x,lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3124 ierr = VecScatterEnd(Mvctx,x,lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3125 3126 ierr = VecDuplicate(lvec,&lcmap);CHKERRQ(ierr); 3127 3128 ierr = VecScatterBegin(Mvctx,cmap,lcmap,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3129 ierr = VecScatterEnd(Mvctx,cmap,lcmap,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3130 3131 /* (3) create sequential iscol_o (a subset of iscol) and isgarray */ 3132 /* off-process column indices */ 3133 count = 0; 3134 ierr = PetscMalloc1(Bn,&idx);CHKERRQ(ierr); 3135 ierr = PetscMalloc1(Bn,&cmap1);CHKERRQ(ierr); 3136 3137 ierr = VecGetArray(lvec,&xarray);CHKERRQ(ierr); 3138 ierr = VecGetArray(lcmap,&cmaparray);CHKERRQ(ierr); 3139 for (i=0; i<Bn; i++) { 3140 if (PetscRealPart(xarray[i]) > -1.0) { 3141 idx[count] = i; /* local column index in off-diagonal part B */ 3142 cmap1[count] = (PetscInt)PetscRealPart(cmaparray[i]); /* column index in submat */ 3143 count++; 3144 } 3145 } 3146 ierr = VecRestoreArray(lvec,&xarray);CHKERRQ(ierr); 3147 ierr = VecRestoreArray(lcmap,&cmaparray);CHKERRQ(ierr); 3148 3149 ierr = ISCreateGeneral(PETSC_COMM_SELF,count,idx,PETSC_COPY_VALUES,iscol_o);CHKERRQ(ierr); 3150 /* cannot ensure iscol_o has same blocksize as iscol! */ 3151 3152 ierr = PetscFree(idx);CHKERRQ(ierr); 3153 *garray = cmap1; 3154 3155 ierr = VecDestroy(&x);CHKERRQ(ierr); 3156 ierr = VecDestroy(&cmap);CHKERRQ(ierr); 3157 ierr = VecDestroy(&lcmap);CHKERRQ(ierr); 3158 PetscFunctionReturn(0); 3159 } 3160 3161 /* isrow and iscol have same processor distribution as mat, output *submat is a submatrix of local mat */ 3162 PetscErrorCode MatCreateSubMatrix_MPIAIJ_SameRowColDist(Mat mat,IS isrow,IS iscol,MatReuse call,Mat *submat) 3163 { 3164 PetscErrorCode ierr; 3165 Mat_MPIAIJ *a = (Mat_MPIAIJ*)mat->data,*asub; 3166 Mat M = NULL; 3167 MPI_Comm comm; 3168 IS iscol_d,isrow_d,iscol_o; 3169 Mat Asub = NULL,Bsub = NULL; 3170 PetscInt n; 3171 3172 PetscFunctionBegin; 3173 ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 3174 3175 if (call == MAT_REUSE_MATRIX) { 3176 /* Retrieve isrow_d, iscol_d and iscol_o from submat */ 3177 ierr = PetscObjectQuery((PetscObject)*submat,"isrow_d",(PetscObject*)&isrow_d);CHKERRQ(ierr); 3178 if (!isrow_d) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"isrow_d passed in was not used before, cannot reuse"); 3179 3180 ierr = PetscObjectQuery((PetscObject)*submat,"iscol_d",(PetscObject*)&iscol_d);CHKERRQ(ierr); 3181 if (!iscol_d) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"iscol_d passed in was not used before, cannot reuse"); 3182 3183 ierr = PetscObjectQuery((PetscObject)*submat,"iscol_o",(PetscObject*)&iscol_o);CHKERRQ(ierr); 3184 if (!iscol_o) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"iscol_o passed in was not used before, cannot reuse"); 3185 3186 /* Update diagonal and off-diagonal portions of submat */ 3187 asub = (Mat_MPIAIJ*)(*submat)->data; 3188 ierr = MatCreateSubMatrix_SeqAIJ(a->A,isrow_d,iscol_d,PETSC_DECIDE,MAT_REUSE_MATRIX,&asub->A);CHKERRQ(ierr); 3189 ierr = ISGetLocalSize(iscol_o,&n);CHKERRQ(ierr); 3190 if (n) { 3191 ierr = MatCreateSubMatrix_SeqAIJ(a->B,isrow_d,iscol_o,PETSC_DECIDE,MAT_REUSE_MATRIX,&asub->B);CHKERRQ(ierr); 3192 } 3193 ierr = MatAssemblyBegin(*submat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3194 ierr = MatAssemblyEnd(*submat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3195 3196 } else { /* call == MAT_INITIAL_MATRIX) */ 3197 const PetscInt *garray; 3198 PetscInt BsubN; 3199 3200 /* Create isrow_d, iscol_d, iscol_o and isgarray (replace isgarray with array?) */ 3201 ierr = ISGetSeqIS_SameColDist_Private(mat,isrow,iscol,&isrow_d,&iscol_d,&iscol_o,&garray);CHKERRQ(ierr); 3202 3203 /* Create local submatrices Asub and Bsub */ 3204 ierr = MatCreateSubMatrix_SeqAIJ(a->A,isrow_d,iscol_d,PETSC_DECIDE,MAT_INITIAL_MATRIX,&Asub);CHKERRQ(ierr); 3205 ierr = MatCreateSubMatrix_SeqAIJ(a->B,isrow_d,iscol_o,PETSC_DECIDE,MAT_INITIAL_MATRIX,&Bsub);CHKERRQ(ierr); 3206 3207 /* Create submatrix M */ 3208 ierr = MatCreateMPIAIJWithSeqAIJ(comm,Asub,Bsub,garray,&M);CHKERRQ(ierr); 3209 3210 /* If Bsub has empty columns, compress iscol_o such that it will retrieve condensed Bsub from a->B during reuse */ 3211 asub = (Mat_MPIAIJ*)M->data; 3212 3213 ierr = ISGetLocalSize(iscol_o,&BsubN);CHKERRQ(ierr); 3214 n = asub->B->cmap->N; 3215 if (BsubN > n) { 3216 /* This case can be tested using ~petsc/src/tao/bound/examples/tutorials/runplate2_3 */ 3217 const PetscInt *idx; 3218 PetscInt i,j,*idx_new,*subgarray = asub->garray; 3219 ierr = PetscInfo2(M,"submatrix Bn %D != BsubN %D, update iscol_o\n",n,BsubN);CHKERRQ(ierr); 3220 3221 ierr = PetscMalloc1(n,&idx_new);CHKERRQ(ierr); 3222 j = 0; 3223 ierr = ISGetIndices(iscol_o,&idx);CHKERRQ(ierr); 3224 for (i=0; i<n; i++) { 3225 if (j >= BsubN) break; 3226 while (subgarray[i] > garray[j]) j++; 3227 3228 if (subgarray[i] == garray[j]) { 3229 idx_new[i] = idx[j++]; 3230 } else SETERRQ4(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"subgarray[%D]=%D cannot < garray[%D]=%D",i,subgarray[i],j,garray[j]); 3231 } 3232 ierr = ISRestoreIndices(iscol_o,&idx);CHKERRQ(ierr); 3233 3234 ierr = ISDestroy(&iscol_o);CHKERRQ(ierr); 3235 ierr = ISCreateGeneral(PETSC_COMM_SELF,n,idx_new,PETSC_OWN_POINTER,&iscol_o);CHKERRQ(ierr); 3236 3237 } else if (BsubN < n) { 3238 SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Columns of Bsub cannot be smaller than B's",BsubN,asub->B->cmap->N); 3239 } 3240 3241 ierr = PetscFree(garray);CHKERRQ(ierr); 3242 *submat = M; 3243 3244 /* Save isrow_d, iscol_d and iscol_o used in processor for next request */ 3245 ierr = PetscObjectCompose((PetscObject)M,"isrow_d",(PetscObject)isrow_d);CHKERRQ(ierr); 3246 ierr = ISDestroy(&isrow_d);CHKERRQ(ierr); 3247 3248 ierr = PetscObjectCompose((PetscObject)M,"iscol_d",(PetscObject)iscol_d);CHKERRQ(ierr); 3249 ierr = ISDestroy(&iscol_d);CHKERRQ(ierr); 3250 3251 ierr = PetscObjectCompose((PetscObject)M,"iscol_o",(PetscObject)iscol_o);CHKERRQ(ierr); 3252 ierr = ISDestroy(&iscol_o);CHKERRQ(ierr); 3253 } 3254 PetscFunctionReturn(0); 3255 } 3256 3257 PetscErrorCode MatCreateSubMatrix_MPIAIJ(Mat mat,IS isrow,IS iscol,MatReuse call,Mat *newmat) 3258 { 3259 PetscErrorCode ierr; 3260 IS iscol_local=NULL,isrow_d; 3261 PetscInt csize; 3262 PetscInt n,i,j,start,end; 3263 PetscBool sameRowDist=PETSC_FALSE,sameDist[2],tsameDist[2]; 3264 MPI_Comm comm; 3265 3266 PetscFunctionBegin; 3267 /* If isrow has same processor distribution as mat, 3268 call MatCreateSubMatrix_MPIAIJ_SameRowDist() to avoid using a hash table with global size of iscol */ 3269 if (call == MAT_REUSE_MATRIX) { 3270 ierr = PetscObjectQuery((PetscObject)*newmat,"isrow_d",(PetscObject*)&isrow_d);CHKERRQ(ierr); 3271 if (isrow_d) { 3272 sameRowDist = PETSC_TRUE; 3273 tsameDist[1] = PETSC_TRUE; /* sameColDist */ 3274 } else { 3275 ierr = PetscObjectQuery((PetscObject)*newmat,"SubIScol",(PetscObject*)&iscol_local);CHKERRQ(ierr); 3276 if (iscol_local) { 3277 sameRowDist = PETSC_TRUE; 3278 tsameDist[1] = PETSC_FALSE; /* !sameColDist */ 3279 } 3280 } 3281 } else { 3282 /* Check if isrow has same processor distribution as mat */ 3283 sameDist[0] = PETSC_FALSE; 3284 ierr = ISGetLocalSize(isrow,&n);CHKERRQ(ierr); 3285 if (!n) { 3286 sameDist[0] = PETSC_TRUE; 3287 } else { 3288 ierr = ISGetMinMax(isrow,&i,&j);CHKERRQ(ierr); 3289 ierr = MatGetOwnershipRange(mat,&start,&end);CHKERRQ(ierr); 3290 if (i >= start && j < end) { 3291 sameDist[0] = PETSC_TRUE; 3292 } 3293 } 3294 3295 /* Check if iscol has same processor distribution as mat */ 3296 sameDist[1] = PETSC_FALSE; 3297 ierr = ISGetLocalSize(iscol,&n);CHKERRQ(ierr); 3298 if (!n) { 3299 sameDist[1] = PETSC_TRUE; 3300 } else { 3301 ierr = ISGetMinMax(iscol,&i,&j);CHKERRQ(ierr); 3302 ierr = MatGetOwnershipRangeColumn(mat,&start,&end);CHKERRQ(ierr); 3303 if (i >= start && j < end) sameDist[1] = PETSC_TRUE; 3304 } 3305 3306 ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 3307 ierr = MPIU_Allreduce(&sameDist,&tsameDist,2,MPIU_BOOL,MPI_LAND,comm);CHKERRQ(ierr); 3308 sameRowDist = tsameDist[0]; 3309 } 3310 3311 if (sameRowDist) { 3312 if (tsameDist[1]) { /* sameRowDist & sameColDist */ 3313 /* isrow and iscol have same processor distribution as mat */ 3314 ierr = MatCreateSubMatrix_MPIAIJ_SameRowColDist(mat,isrow,iscol,call,newmat);CHKERRQ(ierr); 3315 PetscFunctionReturn(0); 3316 } else { /* sameRowDist */ 3317 /* isrow has same processor distribution as mat */ 3318 if (call == MAT_INITIAL_MATRIX) { 3319 PetscBool sorted; 3320 ierr = ISGetSeqIS_Private(mat,iscol,&iscol_local);CHKERRQ(ierr); 3321 ierr = ISGetLocalSize(iscol_local,&n);CHKERRQ(ierr); /* local size of iscol_local = global columns of newmat */ 3322 ierr = ISGetSize(iscol,&i);CHKERRQ(ierr); 3323 if (n != i) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"n %d != size of iscol %d",n,i); 3324 3325 ierr = ISSorted(iscol_local,&sorted);CHKERRQ(ierr); 3326 if (sorted) { 3327 /* MatCreateSubMatrix_MPIAIJ_SameRowDist() requires iscol_local be sorted; it can have duplicate indices */ 3328 ierr = MatCreateSubMatrix_MPIAIJ_SameRowDist(mat,isrow,iscol,iscol_local,MAT_INITIAL_MATRIX,newmat);CHKERRQ(ierr); 3329 PetscFunctionReturn(0); 3330 } 3331 } else { /* call == MAT_REUSE_MATRIX */ 3332 IS iscol_sub; 3333 ierr = PetscObjectQuery((PetscObject)*newmat,"SubIScol",(PetscObject*)&iscol_sub);CHKERRQ(ierr); 3334 if (iscol_sub) { 3335 ierr = MatCreateSubMatrix_MPIAIJ_SameRowDist(mat,isrow,iscol,NULL,call,newmat);CHKERRQ(ierr); 3336 PetscFunctionReturn(0); 3337 } 3338 } 3339 } 3340 } 3341 3342 /* General case: iscol -> iscol_local which has global size of iscol */ 3343 if (call == MAT_REUSE_MATRIX) { 3344 ierr = PetscObjectQuery((PetscObject)*newmat,"ISAllGather",(PetscObject*)&iscol_local);CHKERRQ(ierr); 3345 if (!iscol_local) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Submatrix passed in was not used before, cannot reuse"); 3346 } else { 3347 if (!iscol_local) { 3348 ierr = ISGetSeqIS_Private(mat,iscol,&iscol_local);CHKERRQ(ierr); 3349 } 3350 } 3351 3352 ierr = ISGetLocalSize(iscol,&csize);CHKERRQ(ierr); 3353 ierr = MatCreateSubMatrix_MPIAIJ_nonscalable(mat,isrow,iscol_local,csize,call,newmat);CHKERRQ(ierr); 3354 3355 if (call == MAT_INITIAL_MATRIX) { 3356 ierr = PetscObjectCompose((PetscObject)*newmat,"ISAllGather",(PetscObject)iscol_local);CHKERRQ(ierr); 3357 ierr = ISDestroy(&iscol_local);CHKERRQ(ierr); 3358 } 3359 PetscFunctionReturn(0); 3360 } 3361 3362 /*@C 3363 MatCreateMPIAIJWithSeqAIJ - creates a MPIAIJ matrix using SeqAIJ matrices that contain the "diagonal" 3364 and "off-diagonal" part of the matrix in CSR format. 3365 3366 Collective on MPI_Comm 3367 3368 Input Parameters: 3369 + comm - MPI communicator 3370 . A - "diagonal" portion of matrix 3371 . B - "off-diagonal" portion of matrix, may have empty columns, will be destroyed by this routine 3372 - garray - global index of B columns 3373 3374 Output Parameter: 3375 . mat - the matrix, with input A as its local diagonal matrix 3376 Level: advanced 3377 3378 Notes: 3379 See MatCreateAIJ() for the definition of "diagonal" and "off-diagonal" portion of the matrix. 3380 A becomes part of output mat, B is destroyed by this routine. The user cannot use A and B anymore. 3381 3382 .seealso: MatCreateMPIAIJWithSplitArrays() 3383 @*/ 3384 PetscErrorCode MatCreateMPIAIJWithSeqAIJ(MPI_Comm comm,Mat A,Mat B,const PetscInt garray[],Mat *mat) 3385 { 3386 PetscErrorCode ierr; 3387 Mat_MPIAIJ *maij; 3388 Mat_SeqAIJ *b=(Mat_SeqAIJ*)B->data,*bnew; 3389 PetscInt *oi=b->i,*oj=b->j,i,nz,col; 3390 PetscScalar *oa=b->a; 3391 Mat Bnew; 3392 PetscInt m,n,N; 3393 3394 PetscFunctionBegin; 3395 ierr = MatCreate(comm,mat);CHKERRQ(ierr); 3396 ierr = MatGetSize(A,&m,&n);CHKERRQ(ierr); 3397 if (m != B->rmap->N) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Am %D != Bm %D",m,B->rmap->N); 3398 if (A->rmap->bs != B->rmap->bs) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"A row bs %D != B row bs %D",A->rmap->bs,B->rmap->bs); 3399 /* remove check below; When B is created using iscol_o from ISGetSeqIS_SameColDist_Private(), its bs may not be same as A */ 3400 /* if (A->cmap->bs != B->cmap->bs) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"A column bs %D != B column bs %D",A->cmap->bs,B->cmap->bs); */ 3401 3402 /* Get global columns of mat */ 3403 ierr = MPIU_Allreduce(&n,&N,1,MPIU_INT,MPI_SUM,comm);CHKERRQ(ierr); 3404 3405 ierr = MatSetSizes(*mat,m,n,PETSC_DECIDE,N);CHKERRQ(ierr); 3406 ierr = MatSetType(*mat,MATMPIAIJ);CHKERRQ(ierr); 3407 ierr = MatSetBlockSizes(*mat,A->rmap->bs,A->cmap->bs);CHKERRQ(ierr); 3408 maij = (Mat_MPIAIJ*)(*mat)->data; 3409 3410 (*mat)->preallocated = PETSC_TRUE; 3411 3412 ierr = PetscLayoutSetUp((*mat)->rmap);CHKERRQ(ierr); 3413 ierr = PetscLayoutSetUp((*mat)->cmap);CHKERRQ(ierr); 3414 3415 /* Set A as diagonal portion of *mat */ 3416 maij->A = A; 3417 3418 nz = oi[m]; 3419 for (i=0; i<nz; i++) { 3420 col = oj[i]; 3421 oj[i] = garray[col]; 3422 } 3423 3424 /* Set Bnew as off-diagonal portion of *mat */ 3425 ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,m,N,oi,oj,oa,&Bnew);CHKERRQ(ierr); 3426 bnew = (Mat_SeqAIJ*)Bnew->data; 3427 bnew->maxnz = b->maxnz; /* allocated nonzeros of B */ 3428 maij->B = Bnew; 3429 3430 if (B->rmap->N != Bnew->rmap->N) SETERRQ2(PETSC_COMM_SELF,0,"BN %d != BnewN %d",B->rmap->N,Bnew->rmap->N); 3431 3432 b->singlemalloc = PETSC_FALSE; /* B arrays are shared by Bnew */ 3433 b->free_a = PETSC_FALSE; 3434 b->free_ij = PETSC_FALSE; 3435 ierr = MatDestroy(&B);CHKERRQ(ierr); 3436 3437 bnew->singlemalloc = PETSC_TRUE; /* arrays will be freed by MatDestroy(&Bnew) */ 3438 bnew->free_a = PETSC_TRUE; 3439 bnew->free_ij = PETSC_TRUE; 3440 3441 /* condense columns of maij->B */ 3442 ierr = MatSetOption(*mat,MAT_NO_OFF_PROC_ENTRIES,PETSC_TRUE);CHKERRQ(ierr); 3443 ierr = MatAssemblyBegin(*mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3444 ierr = MatAssemblyEnd(*mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3445 ierr = MatSetOption(*mat,MAT_NO_OFF_PROC_ENTRIES,PETSC_FALSE);CHKERRQ(ierr); 3446 ierr = MatSetOption(*mat,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE);CHKERRQ(ierr); 3447 PetscFunctionReturn(0); 3448 } 3449 3450 extern PetscErrorCode MatCreateSubMatrices_MPIAIJ_SingleIS_Local(Mat,PetscInt,const IS[],const IS[],MatReuse,PetscBool,Mat*); 3451 3452 PetscErrorCode MatCreateSubMatrix_MPIAIJ_SameRowDist(Mat mat,IS isrow,IS iscol,IS iscol_local,MatReuse call,Mat *newmat) 3453 { 3454 PetscErrorCode ierr; 3455 PetscInt i,m,n,rstart,row,rend,nz,j,bs,cbs; 3456 PetscInt *ii,*jj,nlocal,*dlens,*olens,dlen,olen,jend,mglobal; 3457 Mat_MPIAIJ *a=(Mat_MPIAIJ*)mat->data; 3458 Mat M,Msub,B=a->B; 3459 MatScalar *aa; 3460 Mat_SeqAIJ *aij; 3461 PetscInt *garray = a->garray,*colsub,Ncols; 3462 PetscInt count,Bn=B->cmap->N,cstart=mat->cmap->rstart,cend=mat->cmap->rend; 3463 IS iscol_sub,iscmap; 3464 const PetscInt *is_idx,*cmap; 3465 PetscBool allcolumns=PETSC_FALSE; 3466 MPI_Comm comm; 3467 3468 PetscFunctionBegin; 3469 ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 3470 3471 if (call == MAT_REUSE_MATRIX) { 3472 ierr = PetscObjectQuery((PetscObject)*newmat,"SubIScol",(PetscObject*)&iscol_sub);CHKERRQ(ierr); 3473 if (!iscol_sub) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"SubIScol passed in was not used before, cannot reuse"); 3474 ierr = ISGetLocalSize(iscol_sub,&count);CHKERRQ(ierr); 3475 3476 ierr = PetscObjectQuery((PetscObject)*newmat,"Subcmap",(PetscObject*)&iscmap);CHKERRQ(ierr); 3477 if (!iscmap) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Subcmap passed in was not used before, cannot reuse"); 3478 3479 ierr = PetscObjectQuery((PetscObject)*newmat,"SubMatrix",(PetscObject*)&Msub);CHKERRQ(ierr); 3480 if (!Msub) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Submatrix passed in was not used before, cannot reuse"); 3481 3482 ierr = MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat,1,&isrow,&iscol_sub,MAT_REUSE_MATRIX,PETSC_FALSE,&Msub);CHKERRQ(ierr); 3483 3484 } else { /* call == MAT_INITIAL_MATRIX) */ 3485 PetscBool flg; 3486 3487 ierr = ISGetLocalSize(iscol,&n);CHKERRQ(ierr); 3488 ierr = ISGetSize(iscol,&Ncols);CHKERRQ(ierr); 3489 3490 /* (1) iscol -> nonscalable iscol_local */ 3491 /* Check for special case: each processor gets entire matrix columns */ 3492 ierr = ISIdentity(iscol_local,&flg);CHKERRQ(ierr); 3493 if (flg && n == mat->cmap->N) allcolumns = PETSC_TRUE; 3494 if (allcolumns) { 3495 iscol_sub = iscol_local; 3496 ierr = PetscObjectReference((PetscObject)iscol_local);CHKERRQ(ierr); 3497 ierr = ISCreateStride(PETSC_COMM_SELF,n,0,1,&iscmap);CHKERRQ(ierr); 3498 3499 } else { 3500 /* (2) iscol_local -> iscol_sub and iscmap. Implementation below requires iscol_local be sorted, it can have duplicate indices */ 3501 PetscInt *idx,*cmap1,k; 3502 ierr = PetscMalloc1(Ncols,&idx);CHKERRQ(ierr); 3503 ierr = PetscMalloc1(Ncols,&cmap1);CHKERRQ(ierr); 3504 ierr = ISGetIndices(iscol_local,&is_idx);CHKERRQ(ierr); 3505 count = 0; 3506 k = 0; 3507 for (i=0; i<Ncols; i++) { 3508 j = is_idx[i]; 3509 if (j >= cstart && j < cend) { 3510 /* diagonal part of mat */ 3511 idx[count] = j; 3512 cmap1[count++] = i; /* column index in submat */ 3513 } else if (Bn) { 3514 /* off-diagonal part of mat */ 3515 if (j == garray[k]) { 3516 idx[count] = j; 3517 cmap1[count++] = i; /* column index in submat */ 3518 } else if (j > garray[k]) { 3519 while (j > garray[k] && k < Bn-1) k++; 3520 if (j == garray[k]) { 3521 idx[count] = j; 3522 cmap1[count++] = i; /* column index in submat */ 3523 } 3524 } 3525 } 3526 } 3527 ierr = ISRestoreIndices(iscol_local,&is_idx);CHKERRQ(ierr); 3528 3529 ierr = ISCreateGeneral(PETSC_COMM_SELF,count,idx,PETSC_OWN_POINTER,&iscol_sub);CHKERRQ(ierr); 3530 ierr = ISGetBlockSize(iscol,&cbs);CHKERRQ(ierr); 3531 ierr = ISSetBlockSize(iscol_sub,cbs);CHKERRQ(ierr); 3532 3533 ierr = ISCreateGeneral(PetscObjectComm((PetscObject)iscol_local),count,cmap1,PETSC_OWN_POINTER,&iscmap);CHKERRQ(ierr); 3534 } 3535 3536 /* (3) Create sequential Msub */ 3537 ierr = MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat,1,&isrow,&iscol_sub,MAT_INITIAL_MATRIX,allcolumns,&Msub);CHKERRQ(ierr); 3538 } 3539 3540 ierr = ISGetLocalSize(iscol_sub,&count);CHKERRQ(ierr); 3541 aij = (Mat_SeqAIJ*)(Msub)->data; 3542 ii = aij->i; 3543 ierr = ISGetIndices(iscmap,&cmap);CHKERRQ(ierr); 3544 3545 /* 3546 m - number of local rows 3547 Ncols - number of columns (same on all processors) 3548 rstart - first row in new global matrix generated 3549 */ 3550 ierr = MatGetSize(Msub,&m,NULL);CHKERRQ(ierr); 3551 3552 if (call == MAT_INITIAL_MATRIX) { 3553 /* (4) Create parallel newmat */ 3554 PetscMPIInt rank,size; 3555 PetscInt csize; 3556 3557 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 3558 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 3559 3560 /* 3561 Determine the number of non-zeros in the diagonal and off-diagonal 3562 portions of the matrix in order to do correct preallocation 3563 */ 3564 3565 /* first get start and end of "diagonal" columns */ 3566 ierr = ISGetLocalSize(iscol,&csize);CHKERRQ(ierr); 3567 if (csize == PETSC_DECIDE) { 3568 ierr = ISGetSize(isrow,&mglobal);CHKERRQ(ierr); 3569 if (mglobal == Ncols) { /* square matrix */ 3570 nlocal = m; 3571 } else { 3572 nlocal = Ncols/size + ((Ncols % size) > rank); 3573 } 3574 } else { 3575 nlocal = csize; 3576 } 3577 ierr = MPI_Scan(&nlocal,&rend,1,MPIU_INT,MPI_SUM,comm);CHKERRQ(ierr); 3578 rstart = rend - nlocal; 3579 if (rank == size - 1 && rend != Ncols) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Local column sizes %D do not add up to total number of columns %D",rend,Ncols); 3580 3581 /* next, compute all the lengths */ 3582 jj = aij->j; 3583 ierr = PetscMalloc1(2*m+1,&dlens);CHKERRQ(ierr); 3584 olens = dlens + m; 3585 for (i=0; i<m; i++) { 3586 jend = ii[i+1] - ii[i]; 3587 olen = 0; 3588 dlen = 0; 3589 for (j=0; j<jend; j++) { 3590 if (cmap[*jj] < rstart || cmap[*jj] >= rend) olen++; 3591 else dlen++; 3592 jj++; 3593 } 3594 olens[i] = olen; 3595 dlens[i] = dlen; 3596 } 3597 3598 ierr = ISGetBlockSize(isrow,&bs);CHKERRQ(ierr); 3599 ierr = ISGetBlockSize(iscol,&cbs);CHKERRQ(ierr); 3600 3601 ierr = MatCreate(comm,&M);CHKERRQ(ierr); 3602 ierr = MatSetSizes(M,m,nlocal,PETSC_DECIDE,Ncols);CHKERRQ(ierr); 3603 ierr = MatSetBlockSizes(M,bs,cbs);CHKERRQ(ierr); 3604 ierr = MatSetType(M,((PetscObject)mat)->type_name);CHKERRQ(ierr); 3605 ierr = MatMPIAIJSetPreallocation(M,0,dlens,0,olens);CHKERRQ(ierr); 3606 ierr = PetscFree(dlens);CHKERRQ(ierr); 3607 3608 } else { /* call == MAT_REUSE_MATRIX */ 3609 M = *newmat; 3610 ierr = MatGetLocalSize(M,&i,NULL);CHKERRQ(ierr); 3611 if (i != m) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Previous matrix must be same size/layout as request"); 3612 ierr = MatZeroEntries(M);CHKERRQ(ierr); 3613 /* 3614 The next two lines are needed so we may call MatSetValues_MPIAIJ() below directly, 3615 rather than the slower MatSetValues(). 3616 */ 3617 M->was_assembled = PETSC_TRUE; 3618 M->assembled = PETSC_FALSE; 3619 } 3620 3621 /* (5) Set values of Msub to *newmat */ 3622 ierr = PetscMalloc1(count,&colsub);CHKERRQ(ierr); 3623 ierr = MatGetOwnershipRange(M,&rstart,NULL);CHKERRQ(ierr); 3624 3625 jj = aij->j; 3626 aa = aij->a; 3627 for (i=0; i<m; i++) { 3628 row = rstart + i; 3629 nz = ii[i+1] - ii[i]; 3630 for (j=0; j<nz; j++) colsub[j] = cmap[jj[j]]; 3631 ierr = MatSetValues_MPIAIJ(M,1,&row,nz,colsub,aa,INSERT_VALUES);CHKERRQ(ierr); 3632 jj += nz; aa += nz; 3633 } 3634 ierr = ISRestoreIndices(iscmap,&cmap);CHKERRQ(ierr); 3635 3636 ierr = MatAssemblyBegin(M,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3637 ierr = MatAssemblyEnd(M,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3638 3639 ierr = PetscFree(colsub);CHKERRQ(ierr); 3640 3641 /* save Msub, iscol_sub and iscmap used in processor for next request */ 3642 if (call == MAT_INITIAL_MATRIX) { 3643 *newmat = M; 3644 ierr = PetscObjectCompose((PetscObject)(*newmat),"SubMatrix",(PetscObject)Msub);CHKERRQ(ierr); 3645 ierr = MatDestroy(&Msub);CHKERRQ(ierr); 3646 3647 ierr = PetscObjectCompose((PetscObject)(*newmat),"SubIScol",(PetscObject)iscol_sub);CHKERRQ(ierr); 3648 ierr = ISDestroy(&iscol_sub);CHKERRQ(ierr); 3649 3650 ierr = PetscObjectCompose((PetscObject)(*newmat),"Subcmap",(PetscObject)iscmap);CHKERRQ(ierr); 3651 ierr = ISDestroy(&iscmap);CHKERRQ(ierr); 3652 3653 if (iscol_local) { 3654 ierr = PetscObjectCompose((PetscObject)(*newmat),"ISAllGather",(PetscObject)iscol_local);CHKERRQ(ierr); 3655 ierr = ISDestroy(&iscol_local);CHKERRQ(ierr); 3656 } 3657 } 3658 PetscFunctionReturn(0); 3659 } 3660 3661 /* 3662 Not great since it makes two copies of the submatrix, first an SeqAIJ 3663 in local and then by concatenating the local matrices the end result. 3664 Writing it directly would be much like MatCreateSubMatrices_MPIAIJ() 3665 3666 Note: This requires a sequential iscol with all indices. 3667 */ 3668 PetscErrorCode MatCreateSubMatrix_MPIAIJ_nonscalable(Mat mat,IS isrow,IS iscol,PetscInt csize,MatReuse call,Mat *newmat) 3669 { 3670 PetscErrorCode ierr; 3671 PetscMPIInt rank,size; 3672 PetscInt i,m,n,rstart,row,rend,nz,*cwork,j,bs,cbs; 3673 PetscInt *ii,*jj,nlocal,*dlens,*olens,dlen,olen,jend,mglobal; 3674 Mat M,Mreuse; 3675 MatScalar *aa,*vwork; 3676 MPI_Comm comm; 3677 Mat_SeqAIJ *aij; 3678 PetscBool colflag,allcolumns=PETSC_FALSE; 3679 3680 PetscFunctionBegin; 3681 ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 3682 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 3683 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 3684 3685 /* Check for special case: each processor gets entire matrix columns */ 3686 ierr = ISIdentity(iscol,&colflag);CHKERRQ(ierr); 3687 ierr = ISGetLocalSize(iscol,&n);CHKERRQ(ierr); 3688 if (colflag && n == mat->cmap->N) allcolumns = PETSC_TRUE; 3689 3690 if (call == MAT_REUSE_MATRIX) { 3691 ierr = PetscObjectQuery((PetscObject)*newmat,"SubMatrix",(PetscObject*)&Mreuse);CHKERRQ(ierr); 3692 if (!Mreuse) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Submatrix passed in was not used before, cannot reuse"); 3693 ierr = MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat,1,&isrow,&iscol,MAT_REUSE_MATRIX,allcolumns,&Mreuse);CHKERRQ(ierr); 3694 } else { 3695 ierr = MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat,1,&isrow,&iscol,MAT_INITIAL_MATRIX,allcolumns,&Mreuse);CHKERRQ(ierr); 3696 } 3697 3698 /* 3699 m - number of local rows 3700 n - number of columns (same on all processors) 3701 rstart - first row in new global matrix generated 3702 */ 3703 ierr = MatGetSize(Mreuse,&m,&n);CHKERRQ(ierr); 3704 ierr = MatGetBlockSizes(Mreuse,&bs,&cbs);CHKERRQ(ierr); 3705 if (call == MAT_INITIAL_MATRIX) { 3706 aij = (Mat_SeqAIJ*)(Mreuse)->data; 3707 ii = aij->i; 3708 jj = aij->j; 3709 3710 /* 3711 Determine the number of non-zeros in the diagonal and off-diagonal 3712 portions of the matrix in order to do correct preallocation 3713 */ 3714 3715 /* first get start and end of "diagonal" columns */ 3716 if (csize == PETSC_DECIDE) { 3717 ierr = ISGetSize(isrow,&mglobal);CHKERRQ(ierr); 3718 if (mglobal == n) { /* square matrix */ 3719 nlocal = m; 3720 } else { 3721 nlocal = n/size + ((n % size) > rank); 3722 } 3723 } else { 3724 nlocal = csize; 3725 } 3726 ierr = MPI_Scan(&nlocal,&rend,1,MPIU_INT,MPI_SUM,comm);CHKERRQ(ierr); 3727 rstart = rend - nlocal; 3728 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); 3729 3730 /* next, compute all the lengths */ 3731 ierr = PetscMalloc1(2*m+1,&dlens);CHKERRQ(ierr); 3732 olens = dlens + m; 3733 for (i=0; i<m; i++) { 3734 jend = ii[i+1] - ii[i]; 3735 olen = 0; 3736 dlen = 0; 3737 for (j=0; j<jend; j++) { 3738 if (*jj < rstart || *jj >= rend) olen++; 3739 else dlen++; 3740 jj++; 3741 } 3742 olens[i] = olen; 3743 dlens[i] = dlen; 3744 } 3745 ierr = MatCreate(comm,&M);CHKERRQ(ierr); 3746 ierr = MatSetSizes(M,m,nlocal,PETSC_DECIDE,n);CHKERRQ(ierr); 3747 ierr = MatSetBlockSizes(M,bs,cbs);CHKERRQ(ierr); 3748 ierr = MatSetType(M,((PetscObject)mat)->type_name);CHKERRQ(ierr); 3749 ierr = MatMPIAIJSetPreallocation(M,0,dlens,0,olens);CHKERRQ(ierr); 3750 ierr = PetscFree(dlens);CHKERRQ(ierr); 3751 } else { 3752 PetscInt ml,nl; 3753 3754 M = *newmat; 3755 ierr = MatGetLocalSize(M,&ml,&nl);CHKERRQ(ierr); 3756 if (ml != m) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Previous matrix must be same size/layout as request"); 3757 ierr = MatZeroEntries(M);CHKERRQ(ierr); 3758 /* 3759 The next two lines are needed so we may call MatSetValues_MPIAIJ() below directly, 3760 rather than the slower MatSetValues(). 3761 */ 3762 M->was_assembled = PETSC_TRUE; 3763 M->assembled = PETSC_FALSE; 3764 } 3765 ierr = MatGetOwnershipRange(M,&rstart,&rend);CHKERRQ(ierr); 3766 aij = (Mat_SeqAIJ*)(Mreuse)->data; 3767 ii = aij->i; 3768 jj = aij->j; 3769 aa = aij->a; 3770 for (i=0; i<m; i++) { 3771 row = rstart + i; 3772 nz = ii[i+1] - ii[i]; 3773 cwork = jj; jj += nz; 3774 vwork = aa; aa += nz; 3775 ierr = MatSetValues_MPIAIJ(M,1,&row,nz,cwork,vwork,INSERT_VALUES);CHKERRQ(ierr); 3776 } 3777 3778 ierr = MatAssemblyBegin(M,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3779 ierr = MatAssemblyEnd(M,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3780 *newmat = M; 3781 3782 /* save submatrix used in processor for next request */ 3783 if (call == MAT_INITIAL_MATRIX) { 3784 ierr = PetscObjectCompose((PetscObject)M,"SubMatrix",(PetscObject)Mreuse);CHKERRQ(ierr); 3785 ierr = MatDestroy(&Mreuse);CHKERRQ(ierr); 3786 } 3787 PetscFunctionReturn(0); 3788 } 3789 3790 PetscErrorCode MatMPIAIJSetPreallocationCSR_MPIAIJ(Mat B,const PetscInt Ii[],const PetscInt J[],const PetscScalar v[]) 3791 { 3792 PetscInt m,cstart, cend,j,nnz,i,d; 3793 PetscInt *d_nnz,*o_nnz,nnz_max = 0,rstart,ii; 3794 const PetscInt *JJ; 3795 PetscScalar *values; 3796 PetscErrorCode ierr; 3797 PetscBool nooffprocentries; 3798 3799 PetscFunctionBegin; 3800 if (Ii && Ii[0]) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Ii[0] must be 0 it is %D",Ii[0]); 3801 3802 ierr = PetscLayoutSetUp(B->rmap);CHKERRQ(ierr); 3803 ierr = PetscLayoutSetUp(B->cmap);CHKERRQ(ierr); 3804 m = B->rmap->n; 3805 cstart = B->cmap->rstart; 3806 cend = B->cmap->rend; 3807 rstart = B->rmap->rstart; 3808 3809 ierr = PetscMalloc2(m,&d_nnz,m,&o_nnz);CHKERRQ(ierr); 3810 3811 #if defined(PETSC_USE_DEBUG) 3812 for (i=0; i<m; i++) { 3813 nnz = Ii[i+1]- Ii[i]; 3814 JJ = J + Ii[i]; 3815 if (nnz < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Local row %D has a negative %D number of columns",i,nnz); 3816 if (nnz && (JJ[0] < 0)) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Row %D starts with negative column index",i,JJ[0]); 3817 if (nnz && (JJ[nnz-1] >= B->cmap->N)) SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Row %D ends with too large a column index %D (max allowed %D)",i,JJ[nnz-1],B->cmap->N); 3818 } 3819 #endif 3820 3821 for (i=0; i<m; i++) { 3822 nnz = Ii[i+1]- Ii[i]; 3823 JJ = J + Ii[i]; 3824 nnz_max = PetscMax(nnz_max,nnz); 3825 d = 0; 3826 for (j=0; j<nnz; j++) { 3827 if (cstart <= JJ[j] && JJ[j] < cend) d++; 3828 } 3829 d_nnz[i] = d; 3830 o_nnz[i] = nnz - d; 3831 } 3832 ierr = MatMPIAIJSetPreallocation(B,0,d_nnz,0,o_nnz);CHKERRQ(ierr); 3833 ierr = PetscFree2(d_nnz,o_nnz);CHKERRQ(ierr); 3834 3835 if (v) values = (PetscScalar*)v; 3836 else { 3837 ierr = PetscCalloc1(nnz_max+1,&values);CHKERRQ(ierr); 3838 } 3839 3840 for (i=0; i<m; i++) { 3841 ii = i + rstart; 3842 nnz = Ii[i+1]- Ii[i]; 3843 ierr = MatSetValues_MPIAIJ(B,1,&ii,nnz,J+Ii[i],values+(v ? Ii[i] : 0),INSERT_VALUES);CHKERRQ(ierr); 3844 } 3845 nooffprocentries = B->nooffprocentries; 3846 B->nooffprocentries = PETSC_TRUE; 3847 ierr = MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3848 ierr = MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3849 B->nooffprocentries = nooffprocentries; 3850 3851 if (!v) { 3852 ierr = PetscFree(values);CHKERRQ(ierr); 3853 } 3854 ierr = MatSetOption(B,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE);CHKERRQ(ierr); 3855 PetscFunctionReturn(0); 3856 } 3857 3858 /*@ 3859 MatMPIAIJSetPreallocationCSR - Allocates memory for a sparse parallel matrix in AIJ format 3860 (the default parallel PETSc format). 3861 3862 Collective on MPI_Comm 3863 3864 Input Parameters: 3865 + B - the matrix 3866 . i - the indices into j for the start of each local row (starts with zero) 3867 . j - the column indices for each local row (starts with zero) 3868 - v - optional values in the matrix 3869 3870 Level: developer 3871 3872 Notes: 3873 The i, j, and v arrays ARE copied by this routine into the internal format used by PETSc; 3874 thus you CANNOT change the matrix entries by changing the values of v[] after you have 3875 called this routine. Use MatCreateMPIAIJWithSplitArrays() to avoid needing to copy the arrays. 3876 3877 The i and j indices are 0 based, and i indices are indices corresponding to the local j array. 3878 3879 The format which is used for the sparse matrix input, is equivalent to a 3880 row-major ordering.. i.e for the following matrix, the input data expected is 3881 as shown 3882 3883 $ 1 0 0 3884 $ 2 0 3 P0 3885 $ ------- 3886 $ 4 5 6 P1 3887 $ 3888 $ Process0 [P0]: rows_owned=[0,1] 3889 $ i = {0,1,3} [size = nrow+1 = 2+1] 3890 $ j = {0,0,2} [size = 3] 3891 $ v = {1,2,3} [size = 3] 3892 $ 3893 $ Process1 [P1]: rows_owned=[2] 3894 $ i = {0,3} [size = nrow+1 = 1+1] 3895 $ j = {0,1,2} [size = 3] 3896 $ v = {4,5,6} [size = 3] 3897 3898 .keywords: matrix, aij, compressed row, sparse, parallel 3899 3900 .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatCreateAIJ(), MATMPIAIJ, 3901 MatCreateSeqAIJWithArrays(), MatCreateMPIAIJWithSplitArrays() 3902 @*/ 3903 PetscErrorCode MatMPIAIJSetPreallocationCSR(Mat B,const PetscInt i[],const PetscInt j[], const PetscScalar v[]) 3904 { 3905 PetscErrorCode ierr; 3906 3907 PetscFunctionBegin; 3908 ierr = PetscTryMethod(B,"MatMPIAIJSetPreallocationCSR_C",(Mat,const PetscInt[],const PetscInt[],const PetscScalar[]),(B,i,j,v));CHKERRQ(ierr); 3909 PetscFunctionReturn(0); 3910 } 3911 3912 /*@C 3913 MatMPIAIJSetPreallocation - Preallocates memory for a sparse parallel matrix in AIJ format 3914 (the default parallel PETSc format). For good matrix assembly performance 3915 the user should preallocate the matrix storage by setting the parameters 3916 d_nz (or d_nnz) and o_nz (or o_nnz). By setting these parameters accurately, 3917 performance can be increased by more than a factor of 50. 3918 3919 Collective on MPI_Comm 3920 3921 Input Parameters: 3922 + B - the matrix 3923 . d_nz - number of nonzeros per row in DIAGONAL portion of local submatrix 3924 (same value is used for all local rows) 3925 . d_nnz - array containing the number of nonzeros in the various rows of the 3926 DIAGONAL portion of the local submatrix (possibly different for each row) 3927 or NULL (PETSC_NULL_INTEGER in Fortran), if d_nz is used to specify the nonzero structure. 3928 The size of this array is equal to the number of local rows, i.e 'm'. 3929 For matrices that will be factored, you must leave room for (and set) 3930 the diagonal entry even if it is zero. 3931 . o_nz - number of nonzeros per row in the OFF-DIAGONAL portion of local 3932 submatrix (same value is used for all local rows). 3933 - o_nnz - array containing the number of nonzeros in the various rows of the 3934 OFF-DIAGONAL portion of the local submatrix (possibly different for 3935 each row) or NULL (PETSC_NULL_INTEGER in Fortran), if o_nz is used to specify the nonzero 3936 structure. The size of this array is equal to the number 3937 of local rows, i.e 'm'. 3938 3939 If the *_nnz parameter is given then the *_nz parameter is ignored 3940 3941 The AIJ format (also called the Yale sparse matrix format or 3942 compressed row storage (CSR)), is fully compatible with standard Fortran 77 3943 storage. The stored row and column indices begin with zero. 3944 See Users-Manual: ch_mat for details. 3945 3946 The parallel matrix is partitioned such that the first m0 rows belong to 3947 process 0, the next m1 rows belong to process 1, the next m2 rows belong 3948 to process 2 etc.. where m0,m1,m2... are the input parameter 'm'. 3949 3950 The DIAGONAL portion of the local submatrix of a processor can be defined 3951 as the submatrix which is obtained by extraction the part corresponding to 3952 the rows r1-r2 and columns c1-c2 of the global matrix, where r1 is the 3953 first row that belongs to the processor, r2 is the last row belonging to 3954 the this processor, and c1-c2 is range of indices of the local part of a 3955 vector suitable for applying the matrix to. This is an mxn matrix. In the 3956 common case of a square matrix, the row and column ranges are the same and 3957 the DIAGONAL part is also square. The remaining portion of the local 3958 submatrix (mxN) constitute the OFF-DIAGONAL portion. 3959 3960 If o_nnz, d_nnz are specified, then o_nz, and d_nz are ignored. 3961 3962 You can call MatGetInfo() to get information on how effective the preallocation was; 3963 for example the fields mallocs,nz_allocated,nz_used,nz_unneeded; 3964 You can also run with the option -info and look for messages with the string 3965 malloc in them to see if additional memory allocation was needed. 3966 3967 Example usage: 3968 3969 Consider the following 8x8 matrix with 34 non-zero values, that is 3970 assembled across 3 processors. Lets assume that proc0 owns 3 rows, 3971 proc1 owns 3 rows, proc2 owns 2 rows. This division can be shown 3972 as follows: 3973 3974 .vb 3975 1 2 0 | 0 3 0 | 0 4 3976 Proc0 0 5 6 | 7 0 0 | 8 0 3977 9 0 10 | 11 0 0 | 12 0 3978 ------------------------------------- 3979 13 0 14 | 15 16 17 | 0 0 3980 Proc1 0 18 0 | 19 20 21 | 0 0 3981 0 0 0 | 22 23 0 | 24 0 3982 ------------------------------------- 3983 Proc2 25 26 27 | 0 0 28 | 29 0 3984 30 0 0 | 31 32 33 | 0 34 3985 .ve 3986 3987 This can be represented as a collection of submatrices as: 3988 3989 .vb 3990 A B C 3991 D E F 3992 G H I 3993 .ve 3994 3995 Where the submatrices A,B,C are owned by proc0, D,E,F are 3996 owned by proc1, G,H,I are owned by proc2. 3997 3998 The 'm' parameters for proc0,proc1,proc2 are 3,3,2 respectively. 3999 The 'n' parameters for proc0,proc1,proc2 are 3,3,2 respectively. 4000 The 'M','N' parameters are 8,8, and have the same values on all procs. 4001 4002 The DIAGONAL submatrices corresponding to proc0,proc1,proc2 are 4003 submatrices [A], [E], [I] respectively. The OFF-DIAGONAL submatrices 4004 corresponding to proc0,proc1,proc2 are [BC], [DF], [GH] respectively. 4005 Internally, each processor stores the DIAGONAL part, and the OFF-DIAGONAL 4006 part as SeqAIJ matrices. for eg: proc1 will store [E] as a SeqAIJ 4007 matrix, ans [DF] as another SeqAIJ matrix. 4008 4009 When d_nz, o_nz parameters are specified, d_nz storage elements are 4010 allocated for every row of the local diagonal submatrix, and o_nz 4011 storage locations are allocated for every row of the OFF-DIAGONAL submat. 4012 One way to choose d_nz and o_nz is to use the max nonzerors per local 4013 rows for each of the local DIAGONAL, and the OFF-DIAGONAL submatrices. 4014 In this case, the values of d_nz,o_nz are: 4015 .vb 4016 proc0 : dnz = 2, o_nz = 2 4017 proc1 : dnz = 3, o_nz = 2 4018 proc2 : dnz = 1, o_nz = 4 4019 .ve 4020 We are allocating m*(d_nz+o_nz) storage locations for every proc. This 4021 translates to 3*(2+2)=12 for proc0, 3*(3+2)=15 for proc1, 2*(1+4)=10 4022 for proc3. i.e we are using 12+15+10=37 storage locations to store 4023 34 values. 4024 4025 When d_nnz, o_nnz parameters are specified, the storage is specified 4026 for every row, coresponding to both DIAGONAL and OFF-DIAGONAL submatrices. 4027 In the above case the values for d_nnz,o_nnz are: 4028 .vb 4029 proc0: d_nnz = [2,2,2] and o_nnz = [2,2,2] 4030 proc1: d_nnz = [3,3,2] and o_nnz = [2,1,1] 4031 proc2: d_nnz = [1,1] and o_nnz = [4,4] 4032 .ve 4033 Here the space allocated is sum of all the above values i.e 34, and 4034 hence pre-allocation is perfect. 4035 4036 Level: intermediate 4037 4038 .keywords: matrix, aij, compressed row, sparse, parallel 4039 4040 .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatCreateAIJ(), MatMPIAIJSetPreallocationCSR(), 4041 MATMPIAIJ, MatGetInfo(), PetscSplitOwnership() 4042 @*/ 4043 PetscErrorCode MatMPIAIJSetPreallocation(Mat B,PetscInt d_nz,const PetscInt d_nnz[],PetscInt o_nz,const PetscInt o_nnz[]) 4044 { 4045 PetscErrorCode ierr; 4046 4047 PetscFunctionBegin; 4048 PetscValidHeaderSpecific(B,MAT_CLASSID,1); 4049 PetscValidType(B,1); 4050 ierr = PetscTryMethod(B,"MatMPIAIJSetPreallocation_C",(Mat,PetscInt,const PetscInt[],PetscInt,const PetscInt[]),(B,d_nz,d_nnz,o_nz,o_nnz));CHKERRQ(ierr); 4051 PetscFunctionReturn(0); 4052 } 4053 4054 /*@ 4055 MatCreateMPIAIJWithArrays - creates a MPI AIJ matrix using arrays that contain in standard 4056 CSR format the local rows. 4057 4058 Collective on MPI_Comm 4059 4060 Input Parameters: 4061 + comm - MPI communicator 4062 . m - number of local rows (Cannot be PETSC_DECIDE) 4063 . n - This value should be the same as the local size used in creating the 4064 x vector for the matrix-vector product y = Ax. (or PETSC_DECIDE to have 4065 calculated if N is given) For square matrices n is almost always m. 4066 . M - number of global rows (or PETSC_DETERMINE to have calculated if m is given) 4067 . N - number of global columns (or PETSC_DETERMINE to have calculated if n is given) 4068 . i - row indices 4069 . j - column indices 4070 - a - matrix values 4071 4072 Output Parameter: 4073 . mat - the matrix 4074 4075 Level: intermediate 4076 4077 Notes: 4078 The i, j, and a arrays ARE copied by this routine into the internal format used by PETSc; 4079 thus you CANNOT change the matrix entries by changing the values of a[] after you have 4080 called this routine. Use MatCreateMPIAIJWithSplitArrays() to avoid needing to copy the arrays. 4081 4082 The i and j indices are 0 based, and i indices are indices corresponding to the local j array. 4083 4084 The format which is used for the sparse matrix input, is equivalent to a 4085 row-major ordering.. i.e for the following matrix, the input data expected is 4086 as shown 4087 4088 $ 1 0 0 4089 $ 2 0 3 P0 4090 $ ------- 4091 $ 4 5 6 P1 4092 $ 4093 $ Process0 [P0]: rows_owned=[0,1] 4094 $ i = {0,1,3} [size = nrow+1 = 2+1] 4095 $ j = {0,0,2} [size = 3] 4096 $ v = {1,2,3} [size = 3] 4097 $ 4098 $ Process1 [P1]: rows_owned=[2] 4099 $ i = {0,3} [size = nrow+1 = 1+1] 4100 $ j = {0,1,2} [size = 3] 4101 $ v = {4,5,6} [size = 3] 4102 4103 .keywords: matrix, aij, compressed row, sparse, parallel 4104 4105 .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatMPIAIJSetPreallocationCSR(), 4106 MATMPIAIJ, MatCreateAIJ(), MatCreateMPIAIJWithSplitArrays() 4107 @*/ 4108 PetscErrorCode MatCreateMPIAIJWithArrays(MPI_Comm comm,PetscInt m,PetscInt n,PetscInt M,PetscInt N,const PetscInt i[],const PetscInt j[],const PetscScalar a[],Mat *mat) 4109 { 4110 PetscErrorCode ierr; 4111 4112 PetscFunctionBegin; 4113 if (i && i[0]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"i (row indices) must start with 0"); 4114 if (m < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"local number of rows (m) cannot be PETSC_DECIDE, or negative"); 4115 ierr = MatCreate(comm,mat);CHKERRQ(ierr); 4116 ierr = MatSetSizes(*mat,m,n,M,N);CHKERRQ(ierr); 4117 /* ierr = MatSetBlockSizes(M,bs,cbs);CHKERRQ(ierr); */ 4118 ierr = MatSetType(*mat,MATMPIAIJ);CHKERRQ(ierr); 4119 ierr = MatMPIAIJSetPreallocationCSR(*mat,i,j,a);CHKERRQ(ierr); 4120 PetscFunctionReturn(0); 4121 } 4122 4123 /*@C 4124 MatCreateAIJ - Creates a sparse parallel matrix in AIJ format 4125 (the default parallel PETSc format). For good matrix assembly performance 4126 the user should preallocate the matrix storage by setting the parameters 4127 d_nz (or d_nnz) and o_nz (or o_nnz). By setting these parameters accurately, 4128 performance can be increased by more than a factor of 50. 4129 4130 Collective on MPI_Comm 4131 4132 Input Parameters: 4133 + comm - MPI communicator 4134 . m - number of local rows (or PETSC_DECIDE to have calculated if M is given) 4135 This value should be the same as the local size used in creating the 4136 y vector for the matrix-vector product y = Ax. 4137 . n - This value should be the same as the local size used in creating the 4138 x vector for the matrix-vector product y = Ax. (or PETSC_DECIDE to have 4139 calculated if N is given) For square matrices n is almost always m. 4140 . M - number of global rows (or PETSC_DETERMINE to have calculated if m is given) 4141 . N - number of global columns (or PETSC_DETERMINE to have calculated if n is given) 4142 . d_nz - number of nonzeros per row in DIAGONAL portion of local submatrix 4143 (same value is used for all local rows) 4144 . d_nnz - array containing the number of nonzeros in the various rows of the 4145 DIAGONAL portion of the local submatrix (possibly different for each row) 4146 or NULL, if d_nz is used to specify the nonzero structure. 4147 The size of this array is equal to the number of local rows, i.e 'm'. 4148 . o_nz - number of nonzeros per row in the OFF-DIAGONAL portion of local 4149 submatrix (same value is used for all local rows). 4150 - o_nnz - array containing the number of nonzeros in the various rows of the 4151 OFF-DIAGONAL portion of the local submatrix (possibly different for 4152 each row) or NULL, if o_nz is used to specify the nonzero 4153 structure. The size of this array is equal to the number 4154 of local rows, i.e 'm'. 4155 4156 Output Parameter: 4157 . A - the matrix 4158 4159 It is recommended that one use the MatCreate(), MatSetType() and/or MatSetFromOptions(), 4160 MatXXXXSetPreallocation() paradgm instead of this routine directly. 4161 [MatXXXXSetPreallocation() is, for example, MatSeqAIJSetPreallocation] 4162 4163 Notes: 4164 If the *_nnz parameter is given then the *_nz parameter is ignored 4165 4166 m,n,M,N parameters specify the size of the matrix, and its partitioning across 4167 processors, while d_nz,d_nnz,o_nz,o_nnz parameters specify the approximate 4168 storage requirements for this matrix. 4169 4170 If PETSC_DECIDE or PETSC_DETERMINE is used for a particular argument on one 4171 processor than it must be used on all processors that share the object for 4172 that argument. 4173 4174 The user MUST specify either the local or global matrix dimensions 4175 (possibly both). 4176 4177 The parallel matrix is partitioned across processors such that the 4178 first m0 rows belong to process 0, the next m1 rows belong to 4179 process 1, the next m2 rows belong to process 2 etc.. where 4180 m0,m1,m2,.. are the input parameter 'm'. i.e each processor stores 4181 values corresponding to [m x N] submatrix. 4182 4183 The columns are logically partitioned with the n0 columns belonging 4184 to 0th partition, the next n1 columns belonging to the next 4185 partition etc.. where n0,n1,n2... are the input parameter 'n'. 4186 4187 The DIAGONAL portion of the local submatrix on any given processor 4188 is the submatrix corresponding to the rows and columns m,n 4189 corresponding to the given processor. i.e diagonal matrix on 4190 process 0 is [m0 x n0], diagonal matrix on process 1 is [m1 x n1] 4191 etc. The remaining portion of the local submatrix [m x (N-n)] 4192 constitute the OFF-DIAGONAL portion. The example below better 4193 illustrates this concept. 4194 4195 For a square global matrix we define each processor's diagonal portion 4196 to be its local rows and the corresponding columns (a square submatrix); 4197 each processor's off-diagonal portion encompasses the remainder of the 4198 local matrix (a rectangular submatrix). 4199 4200 If o_nnz, d_nnz are specified, then o_nz, and d_nz are ignored. 4201 4202 When calling this routine with a single process communicator, a matrix of 4203 type SEQAIJ is returned. If a matrix of type MPIAIJ is desired for this 4204 type of communicator, use the construction mechanism 4205 .vb 4206 MatCreate(...,&A); MatSetType(A,MATMPIAIJ); MatSetSizes(A, m,n,M,N); MatMPIAIJSetPreallocation(A,...); 4207 .ve 4208 4209 $ MatCreate(...,&A); 4210 $ MatSetType(A,MATMPIAIJ); 4211 $ MatSetSizes(A, m,n,M,N); 4212 $ MatMPIAIJSetPreallocation(A,...); 4213 4214 By default, this format uses inodes (identical nodes) when possible. 4215 We search for consecutive rows with the same nonzero structure, thereby 4216 reusing matrix information to achieve increased efficiency. 4217 4218 Options Database Keys: 4219 + -mat_no_inode - Do not use inodes 4220 - -mat_inode_limit <limit> - Sets inode limit (max limit=5) 4221 4222 4223 4224 Example usage: 4225 4226 Consider the following 8x8 matrix with 34 non-zero values, that is 4227 assembled across 3 processors. Lets assume that proc0 owns 3 rows, 4228 proc1 owns 3 rows, proc2 owns 2 rows. This division can be shown 4229 as follows 4230 4231 .vb 4232 1 2 0 | 0 3 0 | 0 4 4233 Proc0 0 5 6 | 7 0 0 | 8 0 4234 9 0 10 | 11 0 0 | 12 0 4235 ------------------------------------- 4236 13 0 14 | 15 16 17 | 0 0 4237 Proc1 0 18 0 | 19 20 21 | 0 0 4238 0 0 0 | 22 23 0 | 24 0 4239 ------------------------------------- 4240 Proc2 25 26 27 | 0 0 28 | 29 0 4241 30 0 0 | 31 32 33 | 0 34 4242 .ve 4243 4244 This can be represented as a collection of submatrices as 4245 4246 .vb 4247 A B C 4248 D E F 4249 G H I 4250 .ve 4251 4252 Where the submatrices A,B,C are owned by proc0, D,E,F are 4253 owned by proc1, G,H,I are owned by proc2. 4254 4255 The 'm' parameters for proc0,proc1,proc2 are 3,3,2 respectively. 4256 The 'n' parameters for proc0,proc1,proc2 are 3,3,2 respectively. 4257 The 'M','N' parameters are 8,8, and have the same values on all procs. 4258 4259 The DIAGONAL submatrices corresponding to proc0,proc1,proc2 are 4260 submatrices [A], [E], [I] respectively. The OFF-DIAGONAL submatrices 4261 corresponding to proc0,proc1,proc2 are [BC], [DF], [GH] respectively. 4262 Internally, each processor stores the DIAGONAL part, and the OFF-DIAGONAL 4263 part as SeqAIJ matrices. for eg: proc1 will store [E] as a SeqAIJ 4264 matrix, ans [DF] as another SeqAIJ matrix. 4265 4266 When d_nz, o_nz parameters are specified, d_nz storage elements are 4267 allocated for every row of the local diagonal submatrix, and o_nz 4268 storage locations are allocated for every row of the OFF-DIAGONAL submat. 4269 One way to choose d_nz and o_nz is to use the max nonzerors per local 4270 rows for each of the local DIAGONAL, and the OFF-DIAGONAL submatrices. 4271 In this case, the values of d_nz,o_nz are 4272 .vb 4273 proc0 : dnz = 2, o_nz = 2 4274 proc1 : dnz = 3, o_nz = 2 4275 proc2 : dnz = 1, o_nz = 4 4276 .ve 4277 We are allocating m*(d_nz+o_nz) storage locations for every proc. This 4278 translates to 3*(2+2)=12 for proc0, 3*(3+2)=15 for proc1, 2*(1+4)=10 4279 for proc3. i.e we are using 12+15+10=37 storage locations to store 4280 34 values. 4281 4282 When d_nnz, o_nnz parameters are specified, the storage is specified 4283 for every row, coresponding to both DIAGONAL and OFF-DIAGONAL submatrices. 4284 In the above case the values for d_nnz,o_nnz are 4285 .vb 4286 proc0: d_nnz = [2,2,2] and o_nnz = [2,2,2] 4287 proc1: d_nnz = [3,3,2] and o_nnz = [2,1,1] 4288 proc2: d_nnz = [1,1] and o_nnz = [4,4] 4289 .ve 4290 Here the space allocated is sum of all the above values i.e 34, and 4291 hence pre-allocation is perfect. 4292 4293 Level: intermediate 4294 4295 .keywords: matrix, aij, compressed row, sparse, parallel 4296 4297 .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatMPIAIJSetPreallocationCSR(), 4298 MATMPIAIJ, MatCreateMPIAIJWithArrays() 4299 @*/ 4300 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) 4301 { 4302 PetscErrorCode ierr; 4303 PetscMPIInt size; 4304 4305 PetscFunctionBegin; 4306 ierr = MatCreate(comm,A);CHKERRQ(ierr); 4307 ierr = MatSetSizes(*A,m,n,M,N);CHKERRQ(ierr); 4308 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 4309 if (size > 1) { 4310 ierr = MatSetType(*A,MATMPIAIJ);CHKERRQ(ierr); 4311 ierr = MatMPIAIJSetPreallocation(*A,d_nz,d_nnz,o_nz,o_nnz);CHKERRQ(ierr); 4312 } else { 4313 ierr = MatSetType(*A,MATSEQAIJ);CHKERRQ(ierr); 4314 ierr = MatSeqAIJSetPreallocation(*A,d_nz,d_nnz);CHKERRQ(ierr); 4315 } 4316 PetscFunctionReturn(0); 4317 } 4318 4319 PetscErrorCode MatMPIAIJGetSeqAIJ(Mat A,Mat *Ad,Mat *Ao,const PetscInt *colmap[]) 4320 { 4321 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 4322 PetscBool flg; 4323 PetscErrorCode ierr; 4324 4325 PetscFunctionBegin; 4326 ierr = PetscObjectTypeCompare((PetscObject)A,MATMPIAIJ,&flg);CHKERRQ(ierr); 4327 if (!flg) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"This function requires a MATMPIAIJ matrix as input"); 4328 if (Ad) *Ad = a->A; 4329 if (Ao) *Ao = a->B; 4330 if (colmap) *colmap = a->garray; 4331 PetscFunctionReturn(0); 4332 } 4333 4334 PetscErrorCode MatCreateMPIMatConcatenateSeqMat_MPIAIJ(MPI_Comm comm,Mat inmat,PetscInt n,MatReuse scall,Mat *outmat) 4335 { 4336 PetscErrorCode ierr; 4337 PetscInt m,N,i,rstart,nnz,Ii; 4338 PetscInt *indx; 4339 PetscScalar *values; 4340 4341 PetscFunctionBegin; 4342 ierr = MatGetSize(inmat,&m,&N);CHKERRQ(ierr); 4343 if (scall == MAT_INITIAL_MATRIX) { /* symbolic phase */ 4344 PetscInt *dnz,*onz,sum,bs,cbs; 4345 4346 if (n == PETSC_DECIDE) { 4347 ierr = PetscSplitOwnership(comm,&n,&N);CHKERRQ(ierr); 4348 } 4349 /* Check sum(n) = N */ 4350 ierr = MPIU_Allreduce(&n,&sum,1,MPIU_INT,MPI_SUM,comm);CHKERRQ(ierr); 4351 if (sum != N) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_INCOMP,"Sum of local columns %D != global columns %D",sum,N); 4352 4353 ierr = MPI_Scan(&m, &rstart,1,MPIU_INT,MPI_SUM,comm);CHKERRQ(ierr); 4354 rstart -= m; 4355 4356 ierr = MatPreallocateInitialize(comm,m,n,dnz,onz);CHKERRQ(ierr); 4357 for (i=0; i<m; i++) { 4358 ierr = MatGetRow_SeqAIJ(inmat,i,&nnz,&indx,NULL);CHKERRQ(ierr); 4359 ierr = MatPreallocateSet(i+rstart,nnz,indx,dnz,onz);CHKERRQ(ierr); 4360 ierr = MatRestoreRow_SeqAIJ(inmat,i,&nnz,&indx,NULL);CHKERRQ(ierr); 4361 } 4362 4363 ierr = MatCreate(comm,outmat);CHKERRQ(ierr); 4364 ierr = MatSetSizes(*outmat,m,n,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 4365 ierr = MatGetBlockSizes(inmat,&bs,&cbs);CHKERRQ(ierr); 4366 ierr = MatSetBlockSizes(*outmat,bs,cbs);CHKERRQ(ierr); 4367 ierr = MatSetType(*outmat,MATAIJ);CHKERRQ(ierr); 4368 ierr = MatSeqAIJSetPreallocation(*outmat,0,dnz);CHKERRQ(ierr); 4369 ierr = MatMPIAIJSetPreallocation(*outmat,0,dnz,0,onz);CHKERRQ(ierr); 4370 ierr = MatPreallocateFinalize(dnz,onz);CHKERRQ(ierr); 4371 } 4372 4373 /* numeric phase */ 4374 ierr = MatGetOwnershipRange(*outmat,&rstart,NULL);CHKERRQ(ierr); 4375 for (i=0; i<m; i++) { 4376 ierr = MatGetRow_SeqAIJ(inmat,i,&nnz,&indx,&values);CHKERRQ(ierr); 4377 Ii = i + rstart; 4378 ierr = MatSetValues(*outmat,1,&Ii,nnz,indx,values,INSERT_VALUES);CHKERRQ(ierr); 4379 ierr = MatRestoreRow_SeqAIJ(inmat,i,&nnz,&indx,&values);CHKERRQ(ierr); 4380 } 4381 ierr = MatAssemblyBegin(*outmat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4382 ierr = MatAssemblyEnd(*outmat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4383 PetscFunctionReturn(0); 4384 } 4385 4386 PetscErrorCode MatFileSplit(Mat A,char *outfile) 4387 { 4388 PetscErrorCode ierr; 4389 PetscMPIInt rank; 4390 PetscInt m,N,i,rstart,nnz; 4391 size_t len; 4392 const PetscInt *indx; 4393 PetscViewer out; 4394 char *name; 4395 Mat B; 4396 const PetscScalar *values; 4397 4398 PetscFunctionBegin; 4399 ierr = MatGetLocalSize(A,&m,0);CHKERRQ(ierr); 4400 ierr = MatGetSize(A,0,&N);CHKERRQ(ierr); 4401 /* Should this be the type of the diagonal block of A? */ 4402 ierr = MatCreate(PETSC_COMM_SELF,&B);CHKERRQ(ierr); 4403 ierr = MatSetSizes(B,m,N,m,N);CHKERRQ(ierr); 4404 ierr = MatSetBlockSizesFromMats(B,A,A);CHKERRQ(ierr); 4405 ierr = MatSetType(B,MATSEQAIJ);CHKERRQ(ierr); 4406 ierr = MatSeqAIJSetPreallocation(B,0,NULL);CHKERRQ(ierr); 4407 ierr = MatGetOwnershipRange(A,&rstart,0);CHKERRQ(ierr); 4408 for (i=0; i<m; i++) { 4409 ierr = MatGetRow(A,i+rstart,&nnz,&indx,&values);CHKERRQ(ierr); 4410 ierr = MatSetValues(B,1,&i,nnz,indx,values,INSERT_VALUES);CHKERRQ(ierr); 4411 ierr = MatRestoreRow(A,i+rstart,&nnz,&indx,&values);CHKERRQ(ierr); 4412 } 4413 ierr = MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4414 ierr = MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4415 4416 ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)A),&rank);CHKERRQ(ierr); 4417 ierr = PetscStrlen(outfile,&len);CHKERRQ(ierr); 4418 ierr = PetscMalloc1(len+5,&name);CHKERRQ(ierr); 4419 sprintf(name,"%s.%d",outfile,rank); 4420 ierr = PetscViewerBinaryOpen(PETSC_COMM_SELF,name,FILE_MODE_APPEND,&out);CHKERRQ(ierr); 4421 ierr = PetscFree(name);CHKERRQ(ierr); 4422 ierr = MatView(B,out);CHKERRQ(ierr); 4423 ierr = PetscViewerDestroy(&out);CHKERRQ(ierr); 4424 ierr = MatDestroy(&B);CHKERRQ(ierr); 4425 PetscFunctionReturn(0); 4426 } 4427 4428 PetscErrorCode MatDestroy_MPIAIJ_SeqsToMPI(Mat A) 4429 { 4430 PetscErrorCode ierr; 4431 Mat_Merge_SeqsToMPI *merge; 4432 PetscContainer container; 4433 4434 PetscFunctionBegin; 4435 ierr = PetscObjectQuery((PetscObject)A,"MatMergeSeqsToMPI",(PetscObject*)&container);CHKERRQ(ierr); 4436 if (container) { 4437 ierr = PetscContainerGetPointer(container,(void**)&merge);CHKERRQ(ierr); 4438 ierr = PetscFree(merge->id_r);CHKERRQ(ierr); 4439 ierr = PetscFree(merge->len_s);CHKERRQ(ierr); 4440 ierr = PetscFree(merge->len_r);CHKERRQ(ierr); 4441 ierr = PetscFree(merge->bi);CHKERRQ(ierr); 4442 ierr = PetscFree(merge->bj);CHKERRQ(ierr); 4443 ierr = PetscFree(merge->buf_ri[0]);CHKERRQ(ierr); 4444 ierr = PetscFree(merge->buf_ri);CHKERRQ(ierr); 4445 ierr = PetscFree(merge->buf_rj[0]);CHKERRQ(ierr); 4446 ierr = PetscFree(merge->buf_rj);CHKERRQ(ierr); 4447 ierr = PetscFree(merge->coi);CHKERRQ(ierr); 4448 ierr = PetscFree(merge->coj);CHKERRQ(ierr); 4449 ierr = PetscFree(merge->owners_co);CHKERRQ(ierr); 4450 ierr = PetscLayoutDestroy(&merge->rowmap);CHKERRQ(ierr); 4451 ierr = PetscFree(merge);CHKERRQ(ierr); 4452 ierr = PetscObjectCompose((PetscObject)A,"MatMergeSeqsToMPI",0);CHKERRQ(ierr); 4453 } 4454 ierr = MatDestroy_MPIAIJ(A);CHKERRQ(ierr); 4455 PetscFunctionReturn(0); 4456 } 4457 4458 #include <../src/mat/utils/freespace.h> 4459 #include <petscbt.h> 4460 4461 PetscErrorCode MatCreateMPIAIJSumSeqAIJNumeric(Mat seqmat,Mat mpimat) 4462 { 4463 PetscErrorCode ierr; 4464 MPI_Comm comm; 4465 Mat_SeqAIJ *a =(Mat_SeqAIJ*)seqmat->data; 4466 PetscMPIInt size,rank,taga,*len_s; 4467 PetscInt N=mpimat->cmap->N,i,j,*owners,*ai=a->i,*aj; 4468 PetscInt proc,m; 4469 PetscInt **buf_ri,**buf_rj; 4470 PetscInt k,anzi,*bj_i,*bi,*bj,arow,bnzi,nextaj; 4471 PetscInt nrows,**buf_ri_k,**nextrow,**nextai; 4472 MPI_Request *s_waits,*r_waits; 4473 MPI_Status *status; 4474 MatScalar *aa=a->a; 4475 MatScalar **abuf_r,*ba_i; 4476 Mat_Merge_SeqsToMPI *merge; 4477 PetscContainer container; 4478 4479 PetscFunctionBegin; 4480 ierr = PetscObjectGetComm((PetscObject)mpimat,&comm);CHKERRQ(ierr); 4481 ierr = PetscLogEventBegin(MAT_Seqstompinum,seqmat,0,0,0);CHKERRQ(ierr); 4482 4483 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 4484 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 4485 4486 ierr = PetscObjectQuery((PetscObject)mpimat,"MatMergeSeqsToMPI",(PetscObject*)&container);CHKERRQ(ierr); 4487 ierr = PetscContainerGetPointer(container,(void**)&merge);CHKERRQ(ierr); 4488 4489 bi = merge->bi; 4490 bj = merge->bj; 4491 buf_ri = merge->buf_ri; 4492 buf_rj = merge->buf_rj; 4493 4494 ierr = PetscMalloc1(size,&status);CHKERRQ(ierr); 4495 owners = merge->rowmap->range; 4496 len_s = merge->len_s; 4497 4498 /* send and recv matrix values */ 4499 /*-----------------------------*/ 4500 ierr = PetscObjectGetNewTag((PetscObject)mpimat,&taga);CHKERRQ(ierr); 4501 ierr = PetscPostIrecvScalar(comm,taga,merge->nrecv,merge->id_r,merge->len_r,&abuf_r,&r_waits);CHKERRQ(ierr); 4502 4503 ierr = PetscMalloc1(merge->nsend+1,&s_waits);CHKERRQ(ierr); 4504 for (proc=0,k=0; proc<size; proc++) { 4505 if (!len_s[proc]) continue; 4506 i = owners[proc]; 4507 ierr = MPI_Isend(aa+ai[i],len_s[proc],MPIU_MATSCALAR,proc,taga,comm,s_waits+k);CHKERRQ(ierr); 4508 k++; 4509 } 4510 4511 if (merge->nrecv) {ierr = MPI_Waitall(merge->nrecv,r_waits,status);CHKERRQ(ierr);} 4512 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,s_waits,status);CHKERRQ(ierr);} 4513 ierr = PetscFree(status);CHKERRQ(ierr); 4514 4515 ierr = PetscFree(s_waits);CHKERRQ(ierr); 4516 ierr = PetscFree(r_waits);CHKERRQ(ierr); 4517 4518 /* insert mat values of mpimat */ 4519 /*----------------------------*/ 4520 ierr = PetscMalloc1(N,&ba_i);CHKERRQ(ierr); 4521 ierr = PetscMalloc3(merge->nrecv,&buf_ri_k,merge->nrecv,&nextrow,merge->nrecv,&nextai);CHKERRQ(ierr); 4522 4523 for (k=0; k<merge->nrecv; k++) { 4524 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 4525 nrows = *(buf_ri_k[k]); 4526 nextrow[k] = buf_ri_k[k]+1; /* next row number of k-th recved i-structure */ 4527 nextai[k] = buf_ri_k[k] + (nrows + 1); /* poins to the next i-structure of k-th recved i-structure */ 4528 } 4529 4530 /* set values of ba */ 4531 m = merge->rowmap->n; 4532 for (i=0; i<m; i++) { 4533 arow = owners[rank] + i; 4534 bj_i = bj+bi[i]; /* col indices of the i-th row of mpimat */ 4535 bnzi = bi[i+1] - bi[i]; 4536 ierr = PetscMemzero(ba_i,bnzi*sizeof(PetscScalar));CHKERRQ(ierr); 4537 4538 /* add local non-zero vals of this proc's seqmat into ba */ 4539 anzi = ai[arow+1] - ai[arow]; 4540 aj = a->j + ai[arow]; 4541 aa = a->a + ai[arow]; 4542 nextaj = 0; 4543 for (j=0; nextaj<anzi; j++) { 4544 if (*(bj_i + j) == aj[nextaj]) { /* bcol == acol */ 4545 ba_i[j] += aa[nextaj++]; 4546 } 4547 } 4548 4549 /* add received vals into ba */ 4550 for (k=0; k<merge->nrecv; k++) { /* k-th received message */ 4551 /* i-th row */ 4552 if (i == *nextrow[k]) { 4553 anzi = *(nextai[k]+1) - *nextai[k]; 4554 aj = buf_rj[k] + *(nextai[k]); 4555 aa = abuf_r[k] + *(nextai[k]); 4556 nextaj = 0; 4557 for (j=0; nextaj<anzi; j++) { 4558 if (*(bj_i + j) == aj[nextaj]) { /* bcol == acol */ 4559 ba_i[j] += aa[nextaj++]; 4560 } 4561 } 4562 nextrow[k]++; nextai[k]++; 4563 } 4564 } 4565 ierr = MatSetValues(mpimat,1,&arow,bnzi,bj_i,ba_i,INSERT_VALUES);CHKERRQ(ierr); 4566 } 4567 ierr = MatAssemblyBegin(mpimat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4568 ierr = MatAssemblyEnd(mpimat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4569 4570 ierr = PetscFree(abuf_r[0]);CHKERRQ(ierr); 4571 ierr = PetscFree(abuf_r);CHKERRQ(ierr); 4572 ierr = PetscFree(ba_i);CHKERRQ(ierr); 4573 ierr = PetscFree3(buf_ri_k,nextrow,nextai);CHKERRQ(ierr); 4574 ierr = PetscLogEventEnd(MAT_Seqstompinum,seqmat,0,0,0);CHKERRQ(ierr); 4575 PetscFunctionReturn(0); 4576 } 4577 4578 PetscErrorCode MatCreateMPIAIJSumSeqAIJSymbolic(MPI_Comm comm,Mat seqmat,PetscInt m,PetscInt n,Mat *mpimat) 4579 { 4580 PetscErrorCode ierr; 4581 Mat B_mpi; 4582 Mat_SeqAIJ *a=(Mat_SeqAIJ*)seqmat->data; 4583 PetscMPIInt size,rank,tagi,tagj,*len_s,*len_si,*len_ri; 4584 PetscInt **buf_rj,**buf_ri,**buf_ri_k; 4585 PetscInt M=seqmat->rmap->n,N=seqmat->cmap->n,i,*owners,*ai=a->i,*aj=a->j; 4586 PetscInt len,proc,*dnz,*onz,bs,cbs; 4587 PetscInt k,anzi,*bi,*bj,*lnk,nlnk,arow,bnzi,nspacedouble=0; 4588 PetscInt nrows,*buf_s,*buf_si,*buf_si_i,**nextrow,**nextai; 4589 MPI_Request *si_waits,*sj_waits,*ri_waits,*rj_waits; 4590 MPI_Status *status; 4591 PetscFreeSpaceList free_space=NULL,current_space=NULL; 4592 PetscBT lnkbt; 4593 Mat_Merge_SeqsToMPI *merge; 4594 PetscContainer container; 4595 4596 PetscFunctionBegin; 4597 ierr = PetscLogEventBegin(MAT_Seqstompisym,seqmat,0,0,0);CHKERRQ(ierr); 4598 4599 /* make sure it is a PETSc comm */ 4600 ierr = PetscCommDuplicate(comm,&comm,NULL);CHKERRQ(ierr); 4601 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 4602 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 4603 4604 ierr = PetscNew(&merge);CHKERRQ(ierr); 4605 ierr = PetscMalloc1(size,&status);CHKERRQ(ierr); 4606 4607 /* determine row ownership */ 4608 /*---------------------------------------------------------*/ 4609 ierr = PetscLayoutCreate(comm,&merge->rowmap);CHKERRQ(ierr); 4610 ierr = PetscLayoutSetLocalSize(merge->rowmap,m);CHKERRQ(ierr); 4611 ierr = PetscLayoutSetSize(merge->rowmap,M);CHKERRQ(ierr); 4612 ierr = PetscLayoutSetBlockSize(merge->rowmap,1);CHKERRQ(ierr); 4613 ierr = PetscLayoutSetUp(merge->rowmap);CHKERRQ(ierr); 4614 ierr = PetscMalloc1(size,&len_si);CHKERRQ(ierr); 4615 ierr = PetscMalloc1(size,&merge->len_s);CHKERRQ(ierr); 4616 4617 m = merge->rowmap->n; 4618 owners = merge->rowmap->range; 4619 4620 /* determine the number of messages to send, their lengths */ 4621 /*---------------------------------------------------------*/ 4622 len_s = merge->len_s; 4623 4624 len = 0; /* length of buf_si[] */ 4625 merge->nsend = 0; 4626 for (proc=0; proc<size; proc++) { 4627 len_si[proc] = 0; 4628 if (proc == rank) { 4629 len_s[proc] = 0; 4630 } else { 4631 len_si[proc] = owners[proc+1] - owners[proc] + 1; 4632 len_s[proc] = ai[owners[proc+1]] - ai[owners[proc]]; /* num of rows to be sent to [proc] */ 4633 } 4634 if (len_s[proc]) { 4635 merge->nsend++; 4636 nrows = 0; 4637 for (i=owners[proc]; i<owners[proc+1]; i++) { 4638 if (ai[i+1] > ai[i]) nrows++; 4639 } 4640 len_si[proc] = 2*(nrows+1); 4641 len += len_si[proc]; 4642 } 4643 } 4644 4645 /* determine the number and length of messages to receive for ij-structure */ 4646 /*-------------------------------------------------------------------------*/ 4647 ierr = PetscGatherNumberOfMessages(comm,NULL,len_s,&merge->nrecv);CHKERRQ(ierr); 4648 ierr = PetscGatherMessageLengths2(comm,merge->nsend,merge->nrecv,len_s,len_si,&merge->id_r,&merge->len_r,&len_ri);CHKERRQ(ierr); 4649 4650 /* post the Irecv of j-structure */ 4651 /*-------------------------------*/ 4652 ierr = PetscCommGetNewTag(comm,&tagj);CHKERRQ(ierr); 4653 ierr = PetscPostIrecvInt(comm,tagj,merge->nrecv,merge->id_r,merge->len_r,&buf_rj,&rj_waits);CHKERRQ(ierr); 4654 4655 /* post the Isend of j-structure */ 4656 /*--------------------------------*/ 4657 ierr = PetscMalloc2(merge->nsend,&si_waits,merge->nsend,&sj_waits);CHKERRQ(ierr); 4658 4659 for (proc=0, k=0; proc<size; proc++) { 4660 if (!len_s[proc]) continue; 4661 i = owners[proc]; 4662 ierr = MPI_Isend(aj+ai[i],len_s[proc],MPIU_INT,proc,tagj,comm,sj_waits+k);CHKERRQ(ierr); 4663 k++; 4664 } 4665 4666 /* receives and sends of j-structure are complete */ 4667 /*------------------------------------------------*/ 4668 if (merge->nrecv) {ierr = MPI_Waitall(merge->nrecv,rj_waits,status);CHKERRQ(ierr);} 4669 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,sj_waits,status);CHKERRQ(ierr);} 4670 4671 /* send and recv i-structure */ 4672 /*---------------------------*/ 4673 ierr = PetscCommGetNewTag(comm,&tagi);CHKERRQ(ierr); 4674 ierr = PetscPostIrecvInt(comm,tagi,merge->nrecv,merge->id_r,len_ri,&buf_ri,&ri_waits);CHKERRQ(ierr); 4675 4676 ierr = PetscMalloc1(len+1,&buf_s);CHKERRQ(ierr); 4677 buf_si = buf_s; /* points to the beginning of k-th msg to be sent */ 4678 for (proc=0,k=0; proc<size; proc++) { 4679 if (!len_s[proc]) continue; 4680 /* form outgoing message for i-structure: 4681 buf_si[0]: nrows to be sent 4682 [1:nrows]: row index (global) 4683 [nrows+1:2*nrows+1]: i-structure index 4684 */ 4685 /*-------------------------------------------*/ 4686 nrows = len_si[proc]/2 - 1; 4687 buf_si_i = buf_si + nrows+1; 4688 buf_si[0] = nrows; 4689 buf_si_i[0] = 0; 4690 nrows = 0; 4691 for (i=owners[proc]; i<owners[proc+1]; i++) { 4692 anzi = ai[i+1] - ai[i]; 4693 if (anzi) { 4694 buf_si_i[nrows+1] = buf_si_i[nrows] + anzi; /* i-structure */ 4695 buf_si[nrows+1] = i-owners[proc]; /* local row index */ 4696 nrows++; 4697 } 4698 } 4699 ierr = MPI_Isend(buf_si,len_si[proc],MPIU_INT,proc,tagi,comm,si_waits+k);CHKERRQ(ierr); 4700 k++; 4701 buf_si += len_si[proc]; 4702 } 4703 4704 if (merge->nrecv) {ierr = MPI_Waitall(merge->nrecv,ri_waits,status);CHKERRQ(ierr);} 4705 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,si_waits,status);CHKERRQ(ierr);} 4706 4707 ierr = PetscInfo2(seqmat,"nsend: %D, nrecv: %D\n",merge->nsend,merge->nrecv);CHKERRQ(ierr); 4708 for (i=0; i<merge->nrecv; i++) { 4709 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); 4710 } 4711 4712 ierr = PetscFree(len_si);CHKERRQ(ierr); 4713 ierr = PetscFree(len_ri);CHKERRQ(ierr); 4714 ierr = PetscFree(rj_waits);CHKERRQ(ierr); 4715 ierr = PetscFree2(si_waits,sj_waits);CHKERRQ(ierr); 4716 ierr = PetscFree(ri_waits);CHKERRQ(ierr); 4717 ierr = PetscFree(buf_s);CHKERRQ(ierr); 4718 ierr = PetscFree(status);CHKERRQ(ierr); 4719 4720 /* compute a local seq matrix in each processor */ 4721 /*----------------------------------------------*/ 4722 /* allocate bi array and free space for accumulating nonzero column info */ 4723 ierr = PetscMalloc1(m+1,&bi);CHKERRQ(ierr); 4724 bi[0] = 0; 4725 4726 /* create and initialize a linked list */ 4727 nlnk = N+1; 4728 ierr = PetscLLCreate(N,N,nlnk,lnk,lnkbt);CHKERRQ(ierr); 4729 4730 /* initial FreeSpace size is 2*(num of local nnz(seqmat)) */ 4731 len = ai[owners[rank+1]] - ai[owners[rank]]; 4732 ierr = PetscFreeSpaceGet(PetscIntMultTruncate(2,len)+1,&free_space);CHKERRQ(ierr); 4733 4734 current_space = free_space; 4735 4736 /* determine symbolic info for each local row */ 4737 ierr = PetscMalloc3(merge->nrecv,&buf_ri_k,merge->nrecv,&nextrow,merge->nrecv,&nextai);CHKERRQ(ierr); 4738 4739 for (k=0; k<merge->nrecv; k++) { 4740 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 4741 nrows = *buf_ri_k[k]; 4742 nextrow[k] = buf_ri_k[k] + 1; /* next row number of k-th recved i-structure */ 4743 nextai[k] = buf_ri_k[k] + (nrows + 1); /* poins to the next i-structure of k-th recved i-structure */ 4744 } 4745 4746 ierr = MatPreallocateInitialize(comm,m,n,dnz,onz);CHKERRQ(ierr); 4747 len = 0; 4748 for (i=0; i<m; i++) { 4749 bnzi = 0; 4750 /* add local non-zero cols of this proc's seqmat into lnk */ 4751 arow = owners[rank] + i; 4752 anzi = ai[arow+1] - ai[arow]; 4753 aj = a->j + ai[arow]; 4754 ierr = PetscLLAddSorted(anzi,aj,N,nlnk,lnk,lnkbt);CHKERRQ(ierr); 4755 bnzi += nlnk; 4756 /* add received col data into lnk */ 4757 for (k=0; k<merge->nrecv; k++) { /* k-th received message */ 4758 if (i == *nextrow[k]) { /* i-th row */ 4759 anzi = *(nextai[k]+1) - *nextai[k]; 4760 aj = buf_rj[k] + *nextai[k]; 4761 ierr = PetscLLAddSorted(anzi,aj,N,nlnk,lnk,lnkbt);CHKERRQ(ierr); 4762 bnzi += nlnk; 4763 nextrow[k]++; nextai[k]++; 4764 } 4765 } 4766 if (len < bnzi) len = bnzi; /* =max(bnzi) */ 4767 4768 /* if free space is not available, make more free space */ 4769 if (current_space->local_remaining<bnzi) { 4770 ierr = PetscFreeSpaceGet(PetscIntSumTruncate(bnzi,current_space->total_array_size),¤t_space);CHKERRQ(ierr); 4771 nspacedouble++; 4772 } 4773 /* copy data into free space, then initialize lnk */ 4774 ierr = PetscLLClean(N,N,bnzi,lnk,current_space->array,lnkbt);CHKERRQ(ierr); 4775 ierr = MatPreallocateSet(i+owners[rank],bnzi,current_space->array,dnz,onz);CHKERRQ(ierr); 4776 4777 current_space->array += bnzi; 4778 current_space->local_used += bnzi; 4779 current_space->local_remaining -= bnzi; 4780 4781 bi[i+1] = bi[i] + bnzi; 4782 } 4783 4784 ierr = PetscFree3(buf_ri_k,nextrow,nextai);CHKERRQ(ierr); 4785 4786 ierr = PetscMalloc1(bi[m]+1,&bj);CHKERRQ(ierr); 4787 ierr = PetscFreeSpaceContiguous(&free_space,bj);CHKERRQ(ierr); 4788 ierr = PetscLLDestroy(lnk,lnkbt);CHKERRQ(ierr); 4789 4790 /* create symbolic parallel matrix B_mpi */ 4791 /*---------------------------------------*/ 4792 ierr = MatGetBlockSizes(seqmat,&bs,&cbs);CHKERRQ(ierr); 4793 ierr = MatCreate(comm,&B_mpi);CHKERRQ(ierr); 4794 if (n==PETSC_DECIDE) { 4795 ierr = MatSetSizes(B_mpi,m,n,PETSC_DETERMINE,N);CHKERRQ(ierr); 4796 } else { 4797 ierr = MatSetSizes(B_mpi,m,n,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 4798 } 4799 ierr = MatSetBlockSizes(B_mpi,bs,cbs);CHKERRQ(ierr); 4800 ierr = MatSetType(B_mpi,MATMPIAIJ);CHKERRQ(ierr); 4801 ierr = MatMPIAIJSetPreallocation(B_mpi,0,dnz,0,onz);CHKERRQ(ierr); 4802 ierr = MatPreallocateFinalize(dnz,onz);CHKERRQ(ierr); 4803 ierr = MatSetOption(B_mpi,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 4804 4805 /* B_mpi is not ready for use - assembly will be done by MatCreateMPIAIJSumSeqAIJNumeric() */ 4806 B_mpi->assembled = PETSC_FALSE; 4807 B_mpi->ops->destroy = MatDestroy_MPIAIJ_SeqsToMPI; 4808 merge->bi = bi; 4809 merge->bj = bj; 4810 merge->buf_ri = buf_ri; 4811 merge->buf_rj = buf_rj; 4812 merge->coi = NULL; 4813 merge->coj = NULL; 4814 merge->owners_co = NULL; 4815 4816 ierr = PetscCommDestroy(&comm);CHKERRQ(ierr); 4817 4818 /* attach the supporting struct to B_mpi for reuse */ 4819 ierr = PetscContainerCreate(PETSC_COMM_SELF,&container);CHKERRQ(ierr); 4820 ierr = PetscContainerSetPointer(container,merge);CHKERRQ(ierr); 4821 ierr = PetscObjectCompose((PetscObject)B_mpi,"MatMergeSeqsToMPI",(PetscObject)container);CHKERRQ(ierr); 4822 ierr = PetscContainerDestroy(&container);CHKERRQ(ierr); 4823 *mpimat = B_mpi; 4824 4825 ierr = PetscLogEventEnd(MAT_Seqstompisym,seqmat,0,0,0);CHKERRQ(ierr); 4826 PetscFunctionReturn(0); 4827 } 4828 4829 /*@C 4830 MatCreateMPIAIJSumSeqAIJ - Creates a MATMPIAIJ matrix by adding sequential 4831 matrices from each processor 4832 4833 Collective on MPI_Comm 4834 4835 Input Parameters: 4836 + comm - the communicators the parallel matrix will live on 4837 . seqmat - the input sequential matrices 4838 . m - number of local rows (or PETSC_DECIDE) 4839 . n - number of local columns (or PETSC_DECIDE) 4840 - scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX 4841 4842 Output Parameter: 4843 . mpimat - the parallel matrix generated 4844 4845 Level: advanced 4846 4847 Notes: 4848 The dimensions of the sequential matrix in each processor MUST be the same. 4849 The input seqmat is included into the container "Mat_Merge_SeqsToMPI", and will be 4850 destroyed when mpimat is destroyed. Call PetscObjectQuery() to access seqmat. 4851 @*/ 4852 PetscErrorCode MatCreateMPIAIJSumSeqAIJ(MPI_Comm comm,Mat seqmat,PetscInt m,PetscInt n,MatReuse scall,Mat *mpimat) 4853 { 4854 PetscErrorCode ierr; 4855 PetscMPIInt size; 4856 4857 PetscFunctionBegin; 4858 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 4859 if (size == 1) { 4860 ierr = PetscLogEventBegin(MAT_Seqstompi,seqmat,0,0,0);CHKERRQ(ierr); 4861 if (scall == MAT_INITIAL_MATRIX) { 4862 ierr = MatDuplicate(seqmat,MAT_COPY_VALUES,mpimat);CHKERRQ(ierr); 4863 } else { 4864 ierr = MatCopy(seqmat,*mpimat,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 4865 } 4866 ierr = PetscLogEventEnd(MAT_Seqstompi,seqmat,0,0,0);CHKERRQ(ierr); 4867 PetscFunctionReturn(0); 4868 } 4869 ierr = PetscLogEventBegin(MAT_Seqstompi,seqmat,0,0,0);CHKERRQ(ierr); 4870 if (scall == MAT_INITIAL_MATRIX) { 4871 ierr = MatCreateMPIAIJSumSeqAIJSymbolic(comm,seqmat,m,n,mpimat);CHKERRQ(ierr); 4872 } 4873 ierr = MatCreateMPIAIJSumSeqAIJNumeric(seqmat,*mpimat);CHKERRQ(ierr); 4874 ierr = PetscLogEventEnd(MAT_Seqstompi,seqmat,0,0,0);CHKERRQ(ierr); 4875 PetscFunctionReturn(0); 4876 } 4877 4878 /*@ 4879 MatMPIAIJGetLocalMat - Creates a SeqAIJ from a MATMPIAIJ matrix by taking all its local rows and putting them into a sequential matrix with 4880 mlocal rows and n columns. Where mlocal is the row count obtained with MatGetLocalSize() and n is the global column count obtained 4881 with MatGetSize() 4882 4883 Not Collective 4884 4885 Input Parameters: 4886 + A - the matrix 4887 . scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX 4888 4889 Output Parameter: 4890 . A_loc - the local sequential matrix generated 4891 4892 Level: developer 4893 4894 .seealso: MatGetOwnerShipRange(), MatMPIAIJGetLocalMatCondensed() 4895 4896 @*/ 4897 PetscErrorCode MatMPIAIJGetLocalMat(Mat A,MatReuse scall,Mat *A_loc) 4898 { 4899 PetscErrorCode ierr; 4900 Mat_MPIAIJ *mpimat=(Mat_MPIAIJ*)A->data; 4901 Mat_SeqAIJ *mat,*a,*b; 4902 PetscInt *ai,*aj,*bi,*bj,*cmap=mpimat->garray; 4903 MatScalar *aa,*ba,*cam; 4904 PetscScalar *ca; 4905 PetscInt am=A->rmap->n,i,j,k,cstart=A->cmap->rstart; 4906 PetscInt *ci,*cj,col,ncols_d,ncols_o,jo; 4907 PetscBool match; 4908 MPI_Comm comm; 4909 PetscMPIInt size; 4910 4911 PetscFunctionBegin; 4912 ierr = PetscObjectTypeCompare((PetscObject)A,MATMPIAIJ,&match);CHKERRQ(ierr); 4913 if (!match) SETERRQ(PetscObjectComm((PetscObject)A), PETSC_ERR_SUP,"Requires MATMPIAIJ matrix as input"); 4914 ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 4915 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 4916 if (size == 1 && scall == MAT_REUSE_MATRIX) PetscFunctionReturn(0); 4917 4918 ierr = PetscLogEventBegin(MAT_Getlocalmat,A,0,0,0);CHKERRQ(ierr); 4919 a = (Mat_SeqAIJ*)(mpimat->A)->data; 4920 b = (Mat_SeqAIJ*)(mpimat->B)->data; 4921 ai = a->i; aj = a->j; bi = b->i; bj = b->j; 4922 aa = a->a; ba = b->a; 4923 if (scall == MAT_INITIAL_MATRIX) { 4924 if (size == 1) { 4925 ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,am,A->cmap->N,ai,aj,aa,A_loc);CHKERRQ(ierr); 4926 PetscFunctionReturn(0); 4927 } 4928 4929 ierr = PetscMalloc1(1+am,&ci);CHKERRQ(ierr); 4930 ci[0] = 0; 4931 for (i=0; i<am; i++) { 4932 ci[i+1] = ci[i] + (ai[i+1] - ai[i]) + (bi[i+1] - bi[i]); 4933 } 4934 ierr = PetscMalloc1(1+ci[am],&cj);CHKERRQ(ierr); 4935 ierr = PetscMalloc1(1+ci[am],&ca);CHKERRQ(ierr); 4936 k = 0; 4937 for (i=0; i<am; i++) { 4938 ncols_o = bi[i+1] - bi[i]; 4939 ncols_d = ai[i+1] - ai[i]; 4940 /* off-diagonal portion of A */ 4941 for (jo=0; jo<ncols_o; jo++) { 4942 col = cmap[*bj]; 4943 if (col >= cstart) break; 4944 cj[k] = col; bj++; 4945 ca[k++] = *ba++; 4946 } 4947 /* diagonal portion of A */ 4948 for (j=0; j<ncols_d; j++) { 4949 cj[k] = cstart + *aj++; 4950 ca[k++] = *aa++; 4951 } 4952 /* off-diagonal portion of A */ 4953 for (j=jo; j<ncols_o; j++) { 4954 cj[k] = cmap[*bj++]; 4955 ca[k++] = *ba++; 4956 } 4957 } 4958 /* put together the new matrix */ 4959 ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,am,A->cmap->N,ci,cj,ca,A_loc);CHKERRQ(ierr); 4960 /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */ 4961 /* Since these are PETSc arrays, change flags to free them as necessary. */ 4962 mat = (Mat_SeqAIJ*)(*A_loc)->data; 4963 mat->free_a = PETSC_TRUE; 4964 mat->free_ij = PETSC_TRUE; 4965 mat->nonew = 0; 4966 } else if (scall == MAT_REUSE_MATRIX) { 4967 mat=(Mat_SeqAIJ*)(*A_loc)->data; 4968 ci = mat->i; cj = mat->j; cam = mat->a; 4969 for (i=0; i<am; i++) { 4970 /* off-diagonal portion of A */ 4971 ncols_o = bi[i+1] - bi[i]; 4972 for (jo=0; jo<ncols_o; jo++) { 4973 col = cmap[*bj]; 4974 if (col >= cstart) break; 4975 *cam++ = *ba++; bj++; 4976 } 4977 /* diagonal portion of A */ 4978 ncols_d = ai[i+1] - ai[i]; 4979 for (j=0; j<ncols_d; j++) *cam++ = *aa++; 4980 /* off-diagonal portion of A */ 4981 for (j=jo; j<ncols_o; j++) { 4982 *cam++ = *ba++; bj++; 4983 } 4984 } 4985 } else SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Invalid MatReuse %d",(int)scall); 4986 ierr = PetscLogEventEnd(MAT_Getlocalmat,A,0,0,0);CHKERRQ(ierr); 4987 PetscFunctionReturn(0); 4988 } 4989 4990 /*@C 4991 MatMPIAIJGetLocalMatCondensed - Creates a SeqAIJ matrix from an MATMPIAIJ matrix by taking all its local rows and NON-ZERO columns 4992 4993 Not Collective 4994 4995 Input Parameters: 4996 + A - the matrix 4997 . scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX 4998 - row, col - index sets of rows and columns to extract (or NULL) 4999 5000 Output Parameter: 5001 . A_loc - the local sequential matrix generated 5002 5003 Level: developer 5004 5005 .seealso: MatGetOwnershipRange(), MatMPIAIJGetLocalMat() 5006 5007 @*/ 5008 PetscErrorCode MatMPIAIJGetLocalMatCondensed(Mat A,MatReuse scall,IS *row,IS *col,Mat *A_loc) 5009 { 5010 Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data; 5011 PetscErrorCode ierr; 5012 PetscInt i,start,end,ncols,nzA,nzB,*cmap,imark,*idx; 5013 IS isrowa,iscola; 5014 Mat *aloc; 5015 PetscBool match; 5016 5017 PetscFunctionBegin; 5018 ierr = PetscObjectTypeCompare((PetscObject)A,MATMPIAIJ,&match);CHKERRQ(ierr); 5019 if (!match) SETERRQ(PetscObjectComm((PetscObject)A), PETSC_ERR_SUP,"Requires MATMPIAIJ matrix as input"); 5020 ierr = PetscLogEventBegin(MAT_Getlocalmatcondensed,A,0,0,0);CHKERRQ(ierr); 5021 if (!row) { 5022 start = A->rmap->rstart; end = A->rmap->rend; 5023 ierr = ISCreateStride(PETSC_COMM_SELF,end-start,start,1,&isrowa);CHKERRQ(ierr); 5024 } else { 5025 isrowa = *row; 5026 } 5027 if (!col) { 5028 start = A->cmap->rstart; 5029 cmap = a->garray; 5030 nzA = a->A->cmap->n; 5031 nzB = a->B->cmap->n; 5032 ierr = PetscMalloc1(nzA+nzB, &idx);CHKERRQ(ierr); 5033 ncols = 0; 5034 for (i=0; i<nzB; i++) { 5035 if (cmap[i] < start) idx[ncols++] = cmap[i]; 5036 else break; 5037 } 5038 imark = i; 5039 for (i=0; i<nzA; i++) idx[ncols++] = start + i; 5040 for (i=imark; i<nzB; i++) idx[ncols++] = cmap[i]; 5041 ierr = ISCreateGeneral(PETSC_COMM_SELF,ncols,idx,PETSC_OWN_POINTER,&iscola);CHKERRQ(ierr); 5042 } else { 5043 iscola = *col; 5044 } 5045 if (scall != MAT_INITIAL_MATRIX) { 5046 ierr = PetscMalloc1(1,&aloc);CHKERRQ(ierr); 5047 aloc[0] = *A_loc; 5048 } 5049 ierr = MatCreateSubMatrices(A,1,&isrowa,&iscola,scall,&aloc);CHKERRQ(ierr); 5050 *A_loc = aloc[0]; 5051 ierr = PetscFree(aloc);CHKERRQ(ierr); 5052 if (!row) { 5053 ierr = ISDestroy(&isrowa);CHKERRQ(ierr); 5054 } 5055 if (!col) { 5056 ierr = ISDestroy(&iscola);CHKERRQ(ierr); 5057 } 5058 ierr = PetscLogEventEnd(MAT_Getlocalmatcondensed,A,0,0,0);CHKERRQ(ierr); 5059 PetscFunctionReturn(0); 5060 } 5061 5062 /*@C 5063 MatGetBrowsOfAcols - Creates a SeqAIJ matrix by taking rows of B that equal to nonzero columns of local A 5064 5065 Collective on Mat 5066 5067 Input Parameters: 5068 + A,B - the matrices in mpiaij format 5069 . scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX 5070 - rowb, colb - index sets of rows and columns of B to extract (or NULL) 5071 5072 Output Parameter: 5073 + rowb, colb - index sets of rows and columns of B to extract 5074 - B_seq - the sequential matrix generated 5075 5076 Level: developer 5077 5078 @*/ 5079 PetscErrorCode MatGetBrowsOfAcols(Mat A,Mat B,MatReuse scall,IS *rowb,IS *colb,Mat *B_seq) 5080 { 5081 Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data; 5082 PetscErrorCode ierr; 5083 PetscInt *idx,i,start,ncols,nzA,nzB,*cmap,imark; 5084 IS isrowb,iscolb; 5085 Mat *bseq=NULL; 5086 5087 PetscFunctionBegin; 5088 if (A->cmap->rstart != B->rmap->rstart || A->cmap->rend != B->rmap->rend) { 5089 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); 5090 } 5091 ierr = PetscLogEventBegin(MAT_GetBrowsOfAcols,A,B,0,0);CHKERRQ(ierr); 5092 5093 if (scall == MAT_INITIAL_MATRIX) { 5094 start = A->cmap->rstart; 5095 cmap = a->garray; 5096 nzA = a->A->cmap->n; 5097 nzB = a->B->cmap->n; 5098 ierr = PetscMalloc1(nzA+nzB, &idx);CHKERRQ(ierr); 5099 ncols = 0; 5100 for (i=0; i<nzB; i++) { /* row < local row index */ 5101 if (cmap[i] < start) idx[ncols++] = cmap[i]; 5102 else break; 5103 } 5104 imark = i; 5105 for (i=0; i<nzA; i++) idx[ncols++] = start + i; /* local rows */ 5106 for (i=imark; i<nzB; i++) idx[ncols++] = cmap[i]; /* row > local row index */ 5107 ierr = ISCreateGeneral(PETSC_COMM_SELF,ncols,idx,PETSC_OWN_POINTER,&isrowb);CHKERRQ(ierr); 5108 ierr = ISCreateStride(PETSC_COMM_SELF,B->cmap->N,0,1,&iscolb);CHKERRQ(ierr); 5109 } else { 5110 if (!rowb || !colb) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"IS rowb and colb must be provided for MAT_REUSE_MATRIX"); 5111 isrowb = *rowb; iscolb = *colb; 5112 ierr = PetscMalloc1(1,&bseq);CHKERRQ(ierr); 5113 bseq[0] = *B_seq; 5114 } 5115 ierr = MatCreateSubMatrices(B,1,&isrowb,&iscolb,scall,&bseq);CHKERRQ(ierr); 5116 *B_seq = bseq[0]; 5117 ierr = PetscFree(bseq);CHKERRQ(ierr); 5118 if (!rowb) { 5119 ierr = ISDestroy(&isrowb);CHKERRQ(ierr); 5120 } else { 5121 *rowb = isrowb; 5122 } 5123 if (!colb) { 5124 ierr = ISDestroy(&iscolb);CHKERRQ(ierr); 5125 } else { 5126 *colb = iscolb; 5127 } 5128 ierr = PetscLogEventEnd(MAT_GetBrowsOfAcols,A,B,0,0);CHKERRQ(ierr); 5129 PetscFunctionReturn(0); 5130 } 5131 5132 /* 5133 MatGetBrowsOfAoCols_MPIAIJ - Creates a SeqAIJ matrix by taking rows of B that equal to nonzero columns 5134 of the OFF-DIAGONAL portion of local A 5135 5136 Collective on Mat 5137 5138 Input Parameters: 5139 + A,B - the matrices in mpiaij format 5140 - scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX 5141 5142 Output Parameter: 5143 + startsj_s - starting point in B's sending j-arrays, saved for MAT_REUSE (or NULL) 5144 . startsj_r - starting point in B's receiving j-arrays, saved for MAT_REUSE (or NULL) 5145 . bufa_ptr - array for sending matrix values, saved for MAT_REUSE (or NULL) 5146 - B_oth - the sequential matrix generated with size aBn=a->B->cmap->n by B->cmap->N 5147 5148 Level: developer 5149 5150 */ 5151 PetscErrorCode MatGetBrowsOfAoCols_MPIAIJ(Mat A,Mat B,MatReuse scall,PetscInt **startsj_s,PetscInt **startsj_r,MatScalar **bufa_ptr,Mat *B_oth) 5152 { 5153 VecScatter_MPI_General *gen_to,*gen_from; 5154 PetscErrorCode ierr; 5155 Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data; 5156 Mat_SeqAIJ *b_oth; 5157 VecScatter ctx; 5158 MPI_Comm comm; 5159 PetscMPIInt *rprocs,*sprocs,tag,rank; 5160 PetscInt *rowlen,*bufj,*bufJ,ncols,aBn=a->B->cmap->n,row,*b_othi,*b_othj; 5161 PetscInt *rvalues,*svalues,*cols,sbs,rbs; 5162 PetscScalar *b_otha,*bufa,*bufA,*vals; 5163 PetscInt i,j,k,l,ll,nrecvs,nsends,nrows,*srow,*rstarts,*rstartsj = 0,*sstarts,*sstartsj,len; 5164 MPI_Request *rwaits = NULL,*swaits = NULL; 5165 MPI_Status *sstatus,rstatus; 5166 PetscMPIInt jj,size; 5167 VecScatterType type; 5168 PetscBool mpi1; 5169 5170 PetscFunctionBegin; 5171 ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 5172 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 5173 5174 if (A->cmap->rstart != B->rmap->rstart || A->cmap->rend != B->rmap->rend) { 5175 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); 5176 } 5177 ierr = PetscLogEventBegin(MAT_GetBrowsOfAocols,A,B,0,0);CHKERRQ(ierr); 5178 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 5179 5180 if (size == 1) { 5181 startsj_s = NULL; 5182 bufa_ptr = NULL; 5183 *B_oth = NULL; 5184 PetscFunctionReturn(0); 5185 } 5186 5187 ctx = a->Mvctx; 5188 ierr = VecScatterGetType(ctx,&type);CHKERRQ(ierr); 5189 ierr = PetscStrcmp(type,"mpi1",&mpi1);CHKERRQ(ierr); 5190 if (!mpi1) { 5191 /* a->Mvctx is not type MPI1 which is not implemented for Mat-Mat ops, 5192 thus create a->Mvctx_mpi1 */ 5193 if (!a->Mvctx_mpi1) { 5194 a->Mvctx_mpi1_flg = PETSC_TRUE; 5195 ierr = MatSetUpMultiply_MPIAIJ(A);CHKERRQ(ierr); 5196 } 5197 ctx = a->Mvctx_mpi1; 5198 } 5199 tag = ((PetscObject)ctx)->tag; 5200 5201 gen_to = (VecScatter_MPI_General*)ctx->todata; 5202 gen_from = (VecScatter_MPI_General*)ctx->fromdata; 5203 nrecvs = gen_from->n; 5204 nsends = gen_to->n; 5205 5206 ierr = PetscMalloc2(nrecvs,&rwaits,nsends,&swaits);CHKERRQ(ierr); 5207 srow = gen_to->indices; /* local row index to be sent */ 5208 sstarts = gen_to->starts; 5209 sprocs = gen_to->procs; 5210 sstatus = gen_to->sstatus; 5211 sbs = gen_to->bs; 5212 rstarts = gen_from->starts; 5213 rprocs = gen_from->procs; 5214 rbs = gen_from->bs; 5215 5216 if (!startsj_s || !bufa_ptr) scall = MAT_INITIAL_MATRIX; 5217 if (scall == MAT_INITIAL_MATRIX) { 5218 /* i-array */ 5219 /*---------*/ 5220 /* post receives */ 5221 ierr = PetscMalloc1(rbs*(rstarts[nrecvs] - rstarts[0]),&rvalues);CHKERRQ(ierr); 5222 for (i=0; i<nrecvs; i++) { 5223 rowlen = rvalues + rstarts[i]*rbs; 5224 nrows = (rstarts[i+1]-rstarts[i])*rbs; /* num of indices to be received */ 5225 ierr = MPI_Irecv(rowlen,nrows,MPIU_INT,rprocs[i],tag,comm,rwaits+i);CHKERRQ(ierr); 5226 } 5227 5228 /* pack the outgoing message */ 5229 ierr = PetscMalloc2(nsends+1,&sstartsj,nrecvs+1,&rstartsj);CHKERRQ(ierr); 5230 5231 sstartsj[0] = 0; 5232 rstartsj[0] = 0; 5233 len = 0; /* total length of j or a array to be sent */ 5234 k = 0; 5235 ierr = PetscMalloc1(sbs*(sstarts[nsends] - sstarts[0]),&svalues);CHKERRQ(ierr); 5236 for (i=0; i<nsends; i++) { 5237 rowlen = svalues + sstarts[i]*sbs; 5238 nrows = sstarts[i+1]-sstarts[i]; /* num of block rows */ 5239 for (j=0; j<nrows; j++) { 5240 row = srow[k] + B->rmap->range[rank]; /* global row idx */ 5241 for (l=0; l<sbs; l++) { 5242 ierr = MatGetRow_MPIAIJ(B,row+l,&ncols,NULL,NULL);CHKERRQ(ierr); /* rowlength */ 5243 5244 rowlen[j*sbs+l] = ncols; 5245 5246 len += ncols; 5247 ierr = MatRestoreRow_MPIAIJ(B,row+l,&ncols,NULL,NULL);CHKERRQ(ierr); 5248 } 5249 k++; 5250 } 5251 ierr = MPI_Isend(rowlen,nrows*sbs,MPIU_INT,sprocs[i],tag,comm,swaits+i);CHKERRQ(ierr); 5252 5253 sstartsj[i+1] = len; /* starting point of (i+1)-th outgoing msg in bufj and bufa */ 5254 } 5255 /* recvs and sends of i-array are completed */ 5256 i = nrecvs; 5257 while (i--) { 5258 ierr = MPI_Waitany(nrecvs,rwaits,&jj,&rstatus);CHKERRQ(ierr); 5259 } 5260 if (nsends) {ierr = MPI_Waitall(nsends,swaits,sstatus);CHKERRQ(ierr);} 5261 ierr = PetscFree(svalues);CHKERRQ(ierr); 5262 5263 /* allocate buffers for sending j and a arrays */ 5264 ierr = PetscMalloc1(len+1,&bufj);CHKERRQ(ierr); 5265 ierr = PetscMalloc1(len+1,&bufa);CHKERRQ(ierr); 5266 5267 /* create i-array of B_oth */ 5268 ierr = PetscMalloc1(aBn+2,&b_othi);CHKERRQ(ierr); 5269 5270 b_othi[0] = 0; 5271 len = 0; /* total length of j or a array to be received */ 5272 k = 0; 5273 for (i=0; i<nrecvs; i++) { 5274 rowlen = rvalues + rstarts[i]*rbs; 5275 nrows = rbs*(rstarts[i+1]-rstarts[i]); /* num of rows to be received */ 5276 for (j=0; j<nrows; j++) { 5277 b_othi[k+1] = b_othi[k] + rowlen[j]; 5278 ierr = PetscIntSumError(rowlen[j],len,&len);CHKERRQ(ierr); 5279 k++; 5280 } 5281 rstartsj[i+1] = len; /* starting point of (i+1)-th incoming msg in bufj and bufa */ 5282 } 5283 ierr = PetscFree(rvalues);CHKERRQ(ierr); 5284 5285 /* allocate space for j and a arrrays of B_oth */ 5286 ierr = PetscMalloc1(b_othi[aBn]+1,&b_othj);CHKERRQ(ierr); 5287 ierr = PetscMalloc1(b_othi[aBn]+1,&b_otha);CHKERRQ(ierr); 5288 5289 /* j-array */ 5290 /*---------*/ 5291 /* post receives of j-array */ 5292 for (i=0; i<nrecvs; i++) { 5293 nrows = rstartsj[i+1]-rstartsj[i]; /* length of the msg received */ 5294 ierr = MPI_Irecv(b_othj+rstartsj[i],nrows,MPIU_INT,rprocs[i],tag,comm,rwaits+i);CHKERRQ(ierr); 5295 } 5296 5297 /* pack the outgoing message j-array */ 5298 k = 0; 5299 for (i=0; i<nsends; i++) { 5300 nrows = sstarts[i+1]-sstarts[i]; /* num of block rows */ 5301 bufJ = bufj+sstartsj[i]; 5302 for (j=0; j<nrows; j++) { 5303 row = srow[k++] + B->rmap->range[rank]; /* global row idx */ 5304 for (ll=0; ll<sbs; ll++) { 5305 ierr = MatGetRow_MPIAIJ(B,row+ll,&ncols,&cols,NULL);CHKERRQ(ierr); 5306 for (l=0; l<ncols; l++) { 5307 *bufJ++ = cols[l]; 5308 } 5309 ierr = MatRestoreRow_MPIAIJ(B,row+ll,&ncols,&cols,NULL);CHKERRQ(ierr); 5310 } 5311 } 5312 ierr = MPI_Isend(bufj+sstartsj[i],sstartsj[i+1]-sstartsj[i],MPIU_INT,sprocs[i],tag,comm,swaits+i);CHKERRQ(ierr); 5313 } 5314 5315 /* recvs and sends of j-array are completed */ 5316 i = nrecvs; 5317 while (i--) { 5318 ierr = MPI_Waitany(nrecvs,rwaits,&jj,&rstatus);CHKERRQ(ierr); 5319 } 5320 if (nsends) {ierr = MPI_Waitall(nsends,swaits,sstatus);CHKERRQ(ierr);} 5321 } else if (scall == MAT_REUSE_MATRIX) { 5322 sstartsj = *startsj_s; 5323 rstartsj = *startsj_r; 5324 bufa = *bufa_ptr; 5325 b_oth = (Mat_SeqAIJ*)(*B_oth)->data; 5326 b_otha = b_oth->a; 5327 } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE, "Matrix P does not posses an object container"); 5328 5329 /* a-array */ 5330 /*---------*/ 5331 /* post receives of a-array */ 5332 for (i=0; i<nrecvs; i++) { 5333 nrows = rstartsj[i+1]-rstartsj[i]; /* length of the msg received */ 5334 ierr = MPI_Irecv(b_otha+rstartsj[i],nrows,MPIU_SCALAR,rprocs[i],tag,comm,rwaits+i);CHKERRQ(ierr); 5335 } 5336 5337 /* pack the outgoing message a-array */ 5338 k = 0; 5339 for (i=0; i<nsends; i++) { 5340 nrows = sstarts[i+1]-sstarts[i]; /* num of block rows */ 5341 bufA = bufa+sstartsj[i]; 5342 for (j=0; j<nrows; j++) { 5343 row = srow[k++] + B->rmap->range[rank]; /* global row idx */ 5344 for (ll=0; ll<sbs; ll++) { 5345 ierr = MatGetRow_MPIAIJ(B,row+ll,&ncols,NULL,&vals);CHKERRQ(ierr); 5346 for (l=0; l<ncols; l++) { 5347 *bufA++ = vals[l]; 5348 } 5349 ierr = MatRestoreRow_MPIAIJ(B,row+ll,&ncols,NULL,&vals);CHKERRQ(ierr); 5350 } 5351 } 5352 ierr = MPI_Isend(bufa+sstartsj[i],sstartsj[i+1]-sstartsj[i],MPIU_SCALAR,sprocs[i],tag,comm,swaits+i);CHKERRQ(ierr); 5353 } 5354 /* recvs and sends of a-array are completed */ 5355 i = nrecvs; 5356 while (i--) { 5357 ierr = MPI_Waitany(nrecvs,rwaits,&jj,&rstatus);CHKERRQ(ierr); 5358 } 5359 if (nsends) {ierr = MPI_Waitall(nsends,swaits,sstatus);CHKERRQ(ierr);} 5360 ierr = PetscFree2(rwaits,swaits);CHKERRQ(ierr); 5361 5362 if (scall == MAT_INITIAL_MATRIX) { 5363 /* put together the new matrix */ 5364 ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,aBn,B->cmap->N,b_othi,b_othj,b_otha,B_oth);CHKERRQ(ierr); 5365 5366 /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */ 5367 /* Since these are PETSc arrays, change flags to free them as necessary. */ 5368 b_oth = (Mat_SeqAIJ*)(*B_oth)->data; 5369 b_oth->free_a = PETSC_TRUE; 5370 b_oth->free_ij = PETSC_TRUE; 5371 b_oth->nonew = 0; 5372 5373 ierr = PetscFree(bufj);CHKERRQ(ierr); 5374 if (!startsj_s || !bufa_ptr) { 5375 ierr = PetscFree2(sstartsj,rstartsj);CHKERRQ(ierr); 5376 ierr = PetscFree(bufa_ptr);CHKERRQ(ierr); 5377 } else { 5378 *startsj_s = sstartsj; 5379 *startsj_r = rstartsj; 5380 *bufa_ptr = bufa; 5381 } 5382 } 5383 ierr = PetscLogEventEnd(MAT_GetBrowsOfAocols,A,B,0,0);CHKERRQ(ierr); 5384 PetscFunctionReturn(0); 5385 } 5386 5387 /*@C 5388 MatGetCommunicationStructs - Provides access to the communication structures used in matrix-vector multiplication. 5389 5390 Not Collective 5391 5392 Input Parameters: 5393 . A - The matrix in mpiaij format 5394 5395 Output Parameter: 5396 + lvec - The local vector holding off-process values from the argument to a matrix-vector product 5397 . colmap - A map from global column index to local index into lvec 5398 - multScatter - A scatter from the argument of a matrix-vector product to lvec 5399 5400 Level: developer 5401 5402 @*/ 5403 #if defined(PETSC_USE_CTABLE) 5404 PetscErrorCode MatGetCommunicationStructs(Mat A, Vec *lvec, PetscTable *colmap, VecScatter *multScatter) 5405 #else 5406 PetscErrorCode MatGetCommunicationStructs(Mat A, Vec *lvec, PetscInt *colmap[], VecScatter *multScatter) 5407 #endif 5408 { 5409 Mat_MPIAIJ *a; 5410 5411 PetscFunctionBegin; 5412 PetscValidHeaderSpecific(A, MAT_CLASSID, 1); 5413 PetscValidPointer(lvec, 2); 5414 PetscValidPointer(colmap, 3); 5415 PetscValidPointer(multScatter, 4); 5416 a = (Mat_MPIAIJ*) A->data; 5417 if (lvec) *lvec = a->lvec; 5418 if (colmap) *colmap = a->colmap; 5419 if (multScatter) *multScatter = a->Mvctx; 5420 PetscFunctionReturn(0); 5421 } 5422 5423 PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJCRL(Mat,MatType,MatReuse,Mat*); 5424 PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJPERM(Mat,MatType,MatReuse,Mat*); 5425 #if defined(PETSC_HAVE_MKL_SPARSE) 5426 PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJMKL(Mat,MatType,MatReuse,Mat*); 5427 #endif 5428 PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPISBAIJ(Mat,MatType,MatReuse,Mat*); 5429 #if defined(PETSC_HAVE_ELEMENTAL) 5430 PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_Elemental(Mat,MatType,MatReuse,Mat*); 5431 #endif 5432 #if defined(PETSC_HAVE_HYPRE) 5433 PETSC_INTERN PetscErrorCode MatConvert_AIJ_HYPRE(Mat,MatType,MatReuse,Mat*); 5434 PETSC_INTERN PetscErrorCode MatMatMatMult_Transpose_AIJ_AIJ(Mat,Mat,Mat,MatReuse,PetscReal,Mat*); 5435 #endif 5436 PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_IS(Mat,MatType,MatReuse,Mat*); 5437 PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPISELL(Mat,MatType,MatReuse,Mat*); 5438 5439 /* 5440 Computes (B'*A')' since computing B*A directly is untenable 5441 5442 n p p 5443 ( ) ( ) ( ) 5444 m ( A ) * n ( B ) = m ( C ) 5445 ( ) ( ) ( ) 5446 5447 */ 5448 PetscErrorCode MatMatMultNumeric_MPIDense_MPIAIJ(Mat A,Mat B,Mat C) 5449 { 5450 PetscErrorCode ierr; 5451 Mat At,Bt,Ct; 5452 5453 PetscFunctionBegin; 5454 ierr = MatTranspose(A,MAT_INITIAL_MATRIX,&At);CHKERRQ(ierr); 5455 ierr = MatTranspose(B,MAT_INITIAL_MATRIX,&Bt);CHKERRQ(ierr); 5456 ierr = MatMatMult(Bt,At,MAT_INITIAL_MATRIX,1.0,&Ct);CHKERRQ(ierr); 5457 ierr = MatDestroy(&At);CHKERRQ(ierr); 5458 ierr = MatDestroy(&Bt);CHKERRQ(ierr); 5459 ierr = MatTranspose(Ct,MAT_REUSE_MATRIX,&C);CHKERRQ(ierr); 5460 ierr = MatDestroy(&Ct);CHKERRQ(ierr); 5461 PetscFunctionReturn(0); 5462 } 5463 5464 PetscErrorCode MatMatMultSymbolic_MPIDense_MPIAIJ(Mat A,Mat B,PetscReal fill,Mat *C) 5465 { 5466 PetscErrorCode ierr; 5467 PetscInt m=A->rmap->n,n=B->cmap->n; 5468 Mat Cmat; 5469 5470 PetscFunctionBegin; 5471 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); 5472 ierr = MatCreate(PetscObjectComm((PetscObject)A),&Cmat);CHKERRQ(ierr); 5473 ierr = MatSetSizes(Cmat,m,n,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 5474 ierr = MatSetBlockSizesFromMats(Cmat,A,B);CHKERRQ(ierr); 5475 ierr = MatSetType(Cmat,MATMPIDENSE);CHKERRQ(ierr); 5476 ierr = MatMPIDenseSetPreallocation(Cmat,NULL);CHKERRQ(ierr); 5477 ierr = MatAssemblyBegin(Cmat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5478 ierr = MatAssemblyEnd(Cmat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5479 5480 Cmat->ops->matmultnumeric = MatMatMultNumeric_MPIDense_MPIAIJ; 5481 5482 *C = Cmat; 5483 PetscFunctionReturn(0); 5484 } 5485 5486 /* ----------------------------------------------------------------*/ 5487 PETSC_INTERN PetscErrorCode MatMatMult_MPIDense_MPIAIJ(Mat A,Mat B,MatReuse scall,PetscReal fill,Mat *C) 5488 { 5489 PetscErrorCode ierr; 5490 5491 PetscFunctionBegin; 5492 if (scall == MAT_INITIAL_MATRIX) { 5493 ierr = PetscLogEventBegin(MAT_MatMultSymbolic,A,B,0,0);CHKERRQ(ierr); 5494 ierr = MatMatMultSymbolic_MPIDense_MPIAIJ(A,B,fill,C);CHKERRQ(ierr); 5495 ierr = PetscLogEventEnd(MAT_MatMultSymbolic,A,B,0,0);CHKERRQ(ierr); 5496 } 5497 ierr = PetscLogEventBegin(MAT_MatMultNumeric,A,B,0,0);CHKERRQ(ierr); 5498 ierr = MatMatMultNumeric_MPIDense_MPIAIJ(A,B,*C);CHKERRQ(ierr); 5499 ierr = PetscLogEventEnd(MAT_MatMultNumeric,A,B,0,0);CHKERRQ(ierr); 5500 PetscFunctionReturn(0); 5501 } 5502 5503 /*MC 5504 MATMPIAIJ - MATMPIAIJ = "mpiaij" - A matrix type to be used for parallel sparse matrices. 5505 5506 Options Database Keys: 5507 . -mat_type mpiaij - sets the matrix type to "mpiaij" during a call to MatSetFromOptions() 5508 5509 Level: beginner 5510 5511 .seealso: MatCreateAIJ() 5512 M*/ 5513 5514 PETSC_EXTERN PetscErrorCode MatCreate_MPIAIJ(Mat B) 5515 { 5516 Mat_MPIAIJ *b; 5517 PetscErrorCode ierr; 5518 PetscMPIInt size; 5519 5520 PetscFunctionBegin; 5521 ierr = MPI_Comm_size(PetscObjectComm((PetscObject)B),&size);CHKERRQ(ierr); 5522 5523 ierr = PetscNewLog(B,&b);CHKERRQ(ierr); 5524 B->data = (void*)b; 5525 ierr = PetscMemcpy(B->ops,&MatOps_Values,sizeof(struct _MatOps));CHKERRQ(ierr); 5526 B->assembled = PETSC_FALSE; 5527 B->insertmode = NOT_SET_VALUES; 5528 b->size = size; 5529 5530 ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)B),&b->rank);CHKERRQ(ierr); 5531 5532 /* build cache for off array entries formed */ 5533 ierr = MatStashCreate_Private(PetscObjectComm((PetscObject)B),1,&B->stash);CHKERRQ(ierr); 5534 5535 b->donotstash = PETSC_FALSE; 5536 b->colmap = 0; 5537 b->garray = 0; 5538 b->roworiented = PETSC_TRUE; 5539 5540 /* stuff used for matrix vector multiply */ 5541 b->lvec = NULL; 5542 b->Mvctx = NULL; 5543 5544 /* stuff for MatGetRow() */ 5545 b->rowindices = 0; 5546 b->rowvalues = 0; 5547 b->getrowactive = PETSC_FALSE; 5548 5549 /* flexible pointer used in CUSP/CUSPARSE classes */ 5550 b->spptr = NULL; 5551 5552 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMPIAIJSetUseScalableIncreaseOverlap_C",MatMPIAIJSetUseScalableIncreaseOverlap_MPIAIJ);CHKERRQ(ierr); 5553 ierr = PetscObjectComposeFunction((PetscObject)B,"MatStoreValues_C",MatStoreValues_MPIAIJ);CHKERRQ(ierr); 5554 ierr = PetscObjectComposeFunction((PetscObject)B,"MatRetrieveValues_C",MatRetrieveValues_MPIAIJ);CHKERRQ(ierr); 5555 ierr = PetscObjectComposeFunction((PetscObject)B,"MatIsTranspose_C",MatIsTranspose_MPIAIJ);CHKERRQ(ierr); 5556 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMPIAIJSetPreallocation_C",MatMPIAIJSetPreallocation_MPIAIJ);CHKERRQ(ierr); 5557 ierr = PetscObjectComposeFunction((PetscObject)B,"MatResetPreallocation_C",MatResetPreallocation_MPIAIJ);CHKERRQ(ierr); 5558 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMPIAIJSetPreallocationCSR_C",MatMPIAIJSetPreallocationCSR_MPIAIJ);CHKERRQ(ierr); 5559 ierr = PetscObjectComposeFunction((PetscObject)B,"MatDiagonalScaleLocal_C",MatDiagonalScaleLocal_MPIAIJ);CHKERRQ(ierr); 5560 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_mpiaijperm_C",MatConvert_MPIAIJ_MPIAIJPERM);CHKERRQ(ierr); 5561 #if defined(PETSC_HAVE_MKL_SPARSE) 5562 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_mpiaijmkl_C",MatConvert_MPIAIJ_MPIAIJMKL);CHKERRQ(ierr); 5563 #endif 5564 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_mpiaijcrl_C",MatConvert_MPIAIJ_MPIAIJCRL);CHKERRQ(ierr); 5565 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_mpisbaij_C",MatConvert_MPIAIJ_MPISBAIJ);CHKERRQ(ierr); 5566 #if defined(PETSC_HAVE_ELEMENTAL) 5567 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_elemental_C",MatConvert_MPIAIJ_Elemental);CHKERRQ(ierr); 5568 #endif 5569 #if defined(PETSC_HAVE_HYPRE) 5570 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_hypre_C",MatConvert_AIJ_HYPRE);CHKERRQ(ierr); 5571 #endif 5572 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_is_C",MatConvert_MPIAIJ_IS);CHKERRQ(ierr); 5573 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_mpisell_C",MatConvert_MPIAIJ_MPISELL);CHKERRQ(ierr); 5574 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMatMult_mpidense_mpiaij_C",MatMatMult_MPIDense_MPIAIJ);CHKERRQ(ierr); 5575 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMatMultSymbolic_mpidense_mpiaij_C",MatMatMultSymbolic_MPIDense_MPIAIJ);CHKERRQ(ierr); 5576 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMatMultNumeric_mpidense_mpiaij_C",MatMatMultNumeric_MPIDense_MPIAIJ);CHKERRQ(ierr); 5577 #if defined(PETSC_HAVE_HYPRE) 5578 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMatMatMult_transpose_mpiaij_mpiaij_C",MatMatMatMult_Transpose_AIJ_AIJ);CHKERRQ(ierr); 5579 #endif 5580 ierr = PetscObjectChangeTypeName((PetscObject)B,MATMPIAIJ);CHKERRQ(ierr); 5581 PetscFunctionReturn(0); 5582 } 5583 5584 /*@C 5585 MatCreateMPIAIJWithSplitArrays - creates a MPI AIJ matrix using arrays that contain the "diagonal" 5586 and "off-diagonal" part of the matrix in CSR format. 5587 5588 Collective on MPI_Comm 5589 5590 Input Parameters: 5591 + comm - MPI communicator 5592 . m - number of local rows (Cannot be PETSC_DECIDE) 5593 . n - This value should be the same as the local size used in creating the 5594 x vector for the matrix-vector product y = Ax. (or PETSC_DECIDE to have 5595 calculated if N is given) For square matrices n is almost always m. 5596 . M - number of global rows (or PETSC_DETERMINE to have calculated if m is given) 5597 . N - number of global columns (or PETSC_DETERMINE to have calculated if n is given) 5598 . i - row indices for "diagonal" portion of matrix 5599 . j - column indices 5600 . a - matrix values 5601 . oi - row indices for "off-diagonal" portion of matrix 5602 . oj - column indices 5603 - oa - matrix values 5604 5605 Output Parameter: 5606 . mat - the matrix 5607 5608 Level: advanced 5609 5610 Notes: 5611 The i, j, and a arrays ARE NOT copied by this routine into the internal format used by PETSc. The user 5612 must free the arrays once the matrix has been destroyed and not before. 5613 5614 The i and j indices are 0 based 5615 5616 See MatCreateAIJ() for the definition of "diagonal" and "off-diagonal" portion of the matrix 5617 5618 This sets local rows and cannot be used to set off-processor values. 5619 5620 Use of this routine is discouraged because it is inflexible and cumbersome to use. It is extremely rare that a 5621 legacy application natively assembles into exactly this split format. The code to do so is nontrivial and does 5622 not easily support in-place reassembly. It is recommended to use MatSetValues() (or a variant thereof) because 5623 the resulting assembly is easier to implement, will work with any matrix format, and the user does not have to 5624 keep track of the underlying array. Use MatSetOption(A,MAT_NO_OFF_PROC_ENTRIES,PETSC_TRUE) to disable all 5625 communication if it is known that only local entries will be set. 5626 5627 .keywords: matrix, aij, compressed row, sparse, parallel 5628 5629 .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatMPIAIJSetPreallocationCSR(), 5630 MATMPIAIJ, MatCreateAIJ(), MatCreateMPIAIJWithArrays() 5631 @*/ 5632 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) 5633 { 5634 PetscErrorCode ierr; 5635 Mat_MPIAIJ *maij; 5636 5637 PetscFunctionBegin; 5638 if (m < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"local number of rows (m) cannot be PETSC_DECIDE, or negative"); 5639 if (i[0]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"i (row indices) must start with 0"); 5640 if (oi[0]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"oi (row indices) must start with 0"); 5641 ierr = MatCreate(comm,mat);CHKERRQ(ierr); 5642 ierr = MatSetSizes(*mat,m,n,M,N);CHKERRQ(ierr); 5643 ierr = MatSetType(*mat,MATMPIAIJ);CHKERRQ(ierr); 5644 maij = (Mat_MPIAIJ*) (*mat)->data; 5645 5646 (*mat)->preallocated = PETSC_TRUE; 5647 5648 ierr = PetscLayoutSetUp((*mat)->rmap);CHKERRQ(ierr); 5649 ierr = PetscLayoutSetUp((*mat)->cmap);CHKERRQ(ierr); 5650 5651 ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,m,n,i,j,a,&maij->A);CHKERRQ(ierr); 5652 ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,m,(*mat)->cmap->N,oi,oj,oa,&maij->B);CHKERRQ(ierr); 5653 5654 ierr = MatAssemblyBegin(maij->A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5655 ierr = MatAssemblyEnd(maij->A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5656 ierr = MatAssemblyBegin(maij->B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5657 ierr = MatAssemblyEnd(maij->B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5658 5659 ierr = MatSetOption(*mat,MAT_NO_OFF_PROC_ENTRIES,PETSC_TRUE);CHKERRQ(ierr); 5660 ierr = MatAssemblyBegin(*mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5661 ierr = MatAssemblyEnd(*mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5662 ierr = MatSetOption(*mat,MAT_NO_OFF_PROC_ENTRIES,PETSC_FALSE);CHKERRQ(ierr); 5663 ierr = MatSetOption(*mat,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE);CHKERRQ(ierr); 5664 PetscFunctionReturn(0); 5665 } 5666 5667 /* 5668 Special version for direct calls from Fortran 5669 */ 5670 #include <petsc/private/fortranimpl.h> 5671 5672 /* Change these macros so can be used in void function */ 5673 #undef CHKERRQ 5674 #define CHKERRQ(ierr) CHKERRABORT(PETSC_COMM_WORLD,ierr) 5675 #undef SETERRQ2 5676 #define SETERRQ2(comm,ierr,b,c,d) CHKERRABORT(comm,ierr) 5677 #undef SETERRQ3 5678 #define SETERRQ3(comm,ierr,b,c,d,e) CHKERRABORT(comm,ierr) 5679 #undef SETERRQ 5680 #define SETERRQ(c,ierr,b) CHKERRABORT(c,ierr) 5681 5682 #if defined(PETSC_HAVE_FORTRAN_CAPS) 5683 #define matsetvaluesmpiaij_ MATSETVALUESMPIAIJ 5684 #elif !defined(PETSC_HAVE_FORTRAN_UNDERSCORE) 5685 #define matsetvaluesmpiaij_ matsetvaluesmpiaij 5686 #else 5687 #endif 5688 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) 5689 { 5690 Mat mat = *mmat; 5691 PetscInt m = *mm, n = *mn; 5692 InsertMode addv = *maddv; 5693 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 5694 PetscScalar value; 5695 PetscErrorCode ierr; 5696 5697 MatCheckPreallocated(mat,1); 5698 if (mat->insertmode == NOT_SET_VALUES) mat->insertmode = addv; 5699 5700 #if defined(PETSC_USE_DEBUG) 5701 else if (mat->insertmode != addv) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Cannot mix add values and insert values"); 5702 #endif 5703 { 5704 PetscInt i,j,rstart = mat->rmap->rstart,rend = mat->rmap->rend; 5705 PetscInt cstart = mat->cmap->rstart,cend = mat->cmap->rend,row,col; 5706 PetscBool roworiented = aij->roworiented; 5707 5708 /* Some Variables required in the macro */ 5709 Mat A = aij->A; 5710 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 5711 PetscInt *aimax = a->imax,*ai = a->i,*ailen = a->ilen,*aj = a->j; 5712 MatScalar *aa = a->a; 5713 PetscBool ignorezeroentries = (((a->ignorezeroentries)&&(addv==ADD_VALUES)) ? PETSC_TRUE : PETSC_FALSE); 5714 Mat B = aij->B; 5715 Mat_SeqAIJ *b = (Mat_SeqAIJ*)B->data; 5716 PetscInt *bimax = b->imax,*bi = b->i,*bilen = b->ilen,*bj = b->j,bm = aij->B->rmap->n,am = aij->A->rmap->n; 5717 MatScalar *ba = b->a; 5718 5719 PetscInt *rp1,*rp2,ii,nrow1,nrow2,_i,rmax1,rmax2,N,low1,high1,low2,high2,t,lastcol1,lastcol2; 5720 PetscInt nonew = a->nonew; 5721 MatScalar *ap1,*ap2; 5722 5723 PetscFunctionBegin; 5724 for (i=0; i<m; i++) { 5725 if (im[i] < 0) continue; 5726 #if defined(PETSC_USE_DEBUG) 5727 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); 5728 #endif 5729 if (im[i] >= rstart && im[i] < rend) { 5730 row = im[i] - rstart; 5731 lastcol1 = -1; 5732 rp1 = aj + ai[row]; 5733 ap1 = aa + ai[row]; 5734 rmax1 = aimax[row]; 5735 nrow1 = ailen[row]; 5736 low1 = 0; 5737 high1 = nrow1; 5738 lastcol2 = -1; 5739 rp2 = bj + bi[row]; 5740 ap2 = ba + bi[row]; 5741 rmax2 = bimax[row]; 5742 nrow2 = bilen[row]; 5743 low2 = 0; 5744 high2 = nrow2; 5745 5746 for (j=0; j<n; j++) { 5747 if (roworiented) value = v[i*n+j]; 5748 else value = v[i+j*m]; 5749 if (in[j] >= cstart && in[j] < cend) { 5750 col = in[j] - cstart; 5751 if (ignorezeroentries && value == 0.0 && (addv == ADD_VALUES) && row != col) continue; 5752 MatSetValues_SeqAIJ_A_Private(row,col,value,addv,im[i],in[j]); 5753 } else if (in[j] < 0) continue; 5754 #if defined(PETSC_USE_DEBUG) 5755 /* extra brace on SETERRQ2() is required for --with-errorchecking=0 - due to the next 'else' clause */ 5756 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);} 5757 #endif 5758 else { 5759 if (mat->was_assembled) { 5760 if (!aij->colmap) { 5761 ierr = MatCreateColmap_MPIAIJ_Private(mat);CHKERRQ(ierr); 5762 } 5763 #if defined(PETSC_USE_CTABLE) 5764 ierr = PetscTableFind(aij->colmap,in[j]+1,&col);CHKERRQ(ierr); 5765 col--; 5766 #else 5767 col = aij->colmap[in[j]] - 1; 5768 #endif 5769 if (ignorezeroentries && value == 0.0 && (addv == ADD_VALUES) && row != col) continue; 5770 if (col < 0 && !((Mat_SeqAIJ*)(aij->A->data))->nonew) { 5771 ierr = MatDisAssemble_MPIAIJ(mat);CHKERRQ(ierr); 5772 col = in[j]; 5773 /* Reinitialize the variables required by MatSetValues_SeqAIJ_B_Private() */ 5774 B = aij->B; 5775 b = (Mat_SeqAIJ*)B->data; 5776 bimax = b->imax; bi = b->i; bilen = b->ilen; bj = b->j; 5777 rp2 = bj + bi[row]; 5778 ap2 = ba + bi[row]; 5779 rmax2 = bimax[row]; 5780 nrow2 = bilen[row]; 5781 low2 = 0; 5782 high2 = nrow2; 5783 bm = aij->B->rmap->n; 5784 ba = b->a; 5785 } 5786 } else col = in[j]; 5787 MatSetValues_SeqAIJ_B_Private(row,col,value,addv,im[i],in[j]); 5788 } 5789 } 5790 } else if (!aij->donotstash) { 5791 if (roworiented) { 5792 ierr = MatStashValuesRow_Private(&mat->stash,im[i],n,in,v+i*n,(PetscBool)(ignorezeroentries && (addv == ADD_VALUES)));CHKERRQ(ierr); 5793 } else { 5794 ierr = MatStashValuesCol_Private(&mat->stash,im[i],n,in,v+i,m,(PetscBool)(ignorezeroentries && (addv == ADD_VALUES)));CHKERRQ(ierr); 5795 } 5796 } 5797 } 5798 } 5799 PetscFunctionReturnVoid(); 5800 } 5801 5802