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) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero at global row/column (%D, %D) into matrix", im[i], in[j]); 608 } else col = in[j]; 609 nonew = b->nonew; 610 MatSetValues_SeqAIJ_B_Private(row,col,value,addv,im[i],in[j]); 611 } 612 } 613 } else { 614 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]); 615 if (!aij->donotstash) { 616 mat->assembled = PETSC_FALSE; 617 if (roworiented) { 618 ierr = MatStashValuesRow_Private(&mat->stash,im[i],n,in,v+i*n,(PetscBool)(ignorezeroentries && (addv == ADD_VALUES)));CHKERRQ(ierr); 619 } else { 620 ierr = MatStashValuesCol_Private(&mat->stash,im[i],n,in,v+i,m,(PetscBool)(ignorezeroentries && (addv == ADD_VALUES)));CHKERRQ(ierr); 621 } 622 } 623 } 624 } 625 PetscFunctionReturn(0); 626 } 627 628 PetscErrorCode MatGetValues_MPIAIJ(Mat mat,PetscInt m,const PetscInt idxm[],PetscInt n,const PetscInt idxn[],PetscScalar v[]) 629 { 630 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 631 PetscErrorCode ierr; 632 PetscInt i,j,rstart = mat->rmap->rstart,rend = mat->rmap->rend; 633 PetscInt cstart = mat->cmap->rstart,cend = mat->cmap->rend,row,col; 634 635 PetscFunctionBegin; 636 for (i=0; i<m; i++) { 637 if (idxm[i] < 0) continue; /* SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative row: %D",idxm[i]);*/ 638 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); 639 if (idxm[i] >= rstart && idxm[i] < rend) { 640 row = idxm[i] - rstart; 641 for (j=0; j<n; j++) { 642 if (idxn[j] < 0) continue; /* SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative column: %D",idxn[j]); */ 643 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); 644 if (idxn[j] >= cstart && idxn[j] < cend) { 645 col = idxn[j] - cstart; 646 ierr = MatGetValues(aij->A,1,&row,1,&col,v+i*n+j);CHKERRQ(ierr); 647 } else { 648 if (!aij->colmap) { 649 ierr = MatCreateColmap_MPIAIJ_Private(mat);CHKERRQ(ierr); 650 } 651 #if defined(PETSC_USE_CTABLE) 652 ierr = PetscTableFind(aij->colmap,idxn[j]+1,&col);CHKERRQ(ierr); 653 col--; 654 #else 655 col = aij->colmap[idxn[j]] - 1; 656 #endif 657 if ((col < 0) || (aij->garray[col] != idxn[j])) *(v+i*n+j) = 0.0; 658 else { 659 ierr = MatGetValues(aij->B,1,&row,1,&col,v+i*n+j);CHKERRQ(ierr); 660 } 661 } 662 } 663 } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Only local values currently supported"); 664 } 665 PetscFunctionReturn(0); 666 } 667 668 extern PetscErrorCode MatMultDiagonalBlock_MPIAIJ(Mat,Vec,Vec); 669 670 PetscErrorCode MatAssemblyBegin_MPIAIJ(Mat mat,MatAssemblyType mode) 671 { 672 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 673 PetscErrorCode ierr; 674 PetscInt nstash,reallocs; 675 676 PetscFunctionBegin; 677 if (aij->donotstash || mat->nooffprocentries) PetscFunctionReturn(0); 678 679 ierr = MatStashScatterBegin_Private(mat,&mat->stash,mat->rmap->range);CHKERRQ(ierr); 680 ierr = MatStashGetInfo_Private(&mat->stash,&nstash,&reallocs);CHKERRQ(ierr); 681 ierr = PetscInfo2(aij->A,"Stash has %D entries, uses %D mallocs.\n",nstash,reallocs);CHKERRQ(ierr); 682 PetscFunctionReturn(0); 683 } 684 685 PetscErrorCode MatAssemblyEnd_MPIAIJ(Mat mat,MatAssemblyType mode) 686 { 687 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 688 Mat_SeqAIJ *a = (Mat_SeqAIJ*)aij->A->data; 689 PetscErrorCode ierr; 690 PetscMPIInt n; 691 PetscInt i,j,rstart,ncols,flg; 692 PetscInt *row,*col; 693 PetscBool other_disassembled; 694 PetscScalar *val; 695 696 /* do not use 'b = (Mat_SeqAIJ*)aij->B->data' as B can be reset in disassembly */ 697 698 PetscFunctionBegin; 699 if (!aij->donotstash && !mat->nooffprocentries) { 700 while (1) { 701 ierr = MatStashScatterGetMesg_Private(&mat->stash,&n,&row,&col,&val,&flg);CHKERRQ(ierr); 702 if (!flg) break; 703 704 for (i=0; i<n; ) { 705 /* Now identify the consecutive vals belonging to the same row */ 706 for (j=i,rstart=row[j]; j<n; j++) { 707 if (row[j] != rstart) break; 708 } 709 if (j < n) ncols = j-i; 710 else ncols = n-i; 711 /* Now assemble all these values with a single function call */ 712 ierr = MatSetValues_MPIAIJ(mat,1,row+i,ncols,col+i,val+i,mat->insertmode);CHKERRQ(ierr); 713 714 i = j; 715 } 716 } 717 ierr = MatStashScatterEnd_Private(&mat->stash);CHKERRQ(ierr); 718 } 719 ierr = MatAssemblyBegin(aij->A,mode);CHKERRQ(ierr); 720 ierr = MatAssemblyEnd(aij->A,mode);CHKERRQ(ierr); 721 722 /* determine if any processor has disassembled, if so we must 723 also disassemble ourselfs, in order that we may reassemble. */ 724 /* 725 if nonzero structure of submatrix B cannot change then we know that 726 no processor disassembled thus we can skip this stuff 727 */ 728 if (!((Mat_SeqAIJ*)aij->B->data)->nonew) { 729 ierr = MPIU_Allreduce(&mat->was_assembled,&other_disassembled,1,MPIU_BOOL,MPI_PROD,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 730 if (mat->was_assembled && !other_disassembled) { 731 ierr = MatDisAssemble_MPIAIJ(mat);CHKERRQ(ierr); 732 } 733 } 734 if (!mat->was_assembled && mode == MAT_FINAL_ASSEMBLY) { 735 ierr = MatSetUpMultiply_MPIAIJ(mat);CHKERRQ(ierr); 736 } 737 ierr = MatSetOption(aij->B,MAT_USE_INODES,PETSC_FALSE);CHKERRQ(ierr); 738 ierr = MatAssemblyBegin(aij->B,mode);CHKERRQ(ierr); 739 ierr = MatAssemblyEnd(aij->B,mode);CHKERRQ(ierr); 740 741 ierr = PetscFree2(aij->rowvalues,aij->rowindices);CHKERRQ(ierr); 742 743 aij->rowvalues = 0; 744 745 ierr = VecDestroy(&aij->diag);CHKERRQ(ierr); 746 if (a->inode.size) mat->ops->multdiagonalblock = MatMultDiagonalBlock_MPIAIJ; 747 748 /* if no new nonzero locations are allowed in matrix then only set the matrix state the first time through */ 749 if ((!mat->was_assembled && mode == MAT_FINAL_ASSEMBLY) || !((Mat_SeqAIJ*)(aij->A->data))->nonew) { 750 PetscObjectState state = aij->A->nonzerostate + aij->B->nonzerostate; 751 ierr = MPIU_Allreduce(&state,&mat->nonzerostate,1,MPIU_INT64,MPI_SUM,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 752 } 753 PetscFunctionReturn(0); 754 } 755 756 PetscErrorCode MatZeroEntries_MPIAIJ(Mat A) 757 { 758 Mat_MPIAIJ *l = (Mat_MPIAIJ*)A->data; 759 PetscErrorCode ierr; 760 761 PetscFunctionBegin; 762 ierr = MatZeroEntries(l->A);CHKERRQ(ierr); 763 ierr = MatZeroEntries(l->B);CHKERRQ(ierr); 764 PetscFunctionReturn(0); 765 } 766 767 PetscErrorCode MatZeroRows_MPIAIJ(Mat A,PetscInt N,const PetscInt rows[],PetscScalar diag,Vec x,Vec b) 768 { 769 Mat_MPIAIJ *mat = (Mat_MPIAIJ *) A->data; 770 PetscInt *lrows; 771 PetscInt r, len; 772 PetscErrorCode ierr; 773 774 PetscFunctionBegin; 775 /* get locally owned rows */ 776 ierr = MatZeroRowsMapLocal_Private(A,N,rows,&len,&lrows);CHKERRQ(ierr); 777 /* fix right hand side if needed */ 778 if (x && b) { 779 const PetscScalar *xx; 780 PetscScalar *bb; 781 782 ierr = VecGetArrayRead(x, &xx);CHKERRQ(ierr); 783 ierr = VecGetArray(b, &bb);CHKERRQ(ierr); 784 for (r = 0; r < len; ++r) bb[lrows[r]] = diag*xx[lrows[r]]; 785 ierr = VecRestoreArrayRead(x, &xx);CHKERRQ(ierr); 786 ierr = VecRestoreArray(b, &bb);CHKERRQ(ierr); 787 } 788 /* Must zero l->B before l->A because the (diag) case below may put values into l->B*/ 789 ierr = MatZeroRows(mat->B, len, lrows, 0.0, NULL, NULL);CHKERRQ(ierr); 790 if (A->congruentlayouts == -1) { /* first time we compare rows and cols layouts */ 791 PetscBool cong; 792 ierr = PetscLayoutCompare(A->rmap,A->cmap,&cong);CHKERRQ(ierr); 793 if (cong) A->congruentlayouts = 1; 794 else A->congruentlayouts = 0; 795 } 796 if ((diag != 0.0) && A->congruentlayouts) { 797 ierr = MatZeroRows(mat->A, len, lrows, diag, NULL, NULL);CHKERRQ(ierr); 798 } else if (diag != 0.0) { 799 ierr = MatZeroRows(mat->A, len, lrows, 0.0, NULL, NULL);CHKERRQ(ierr); 800 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"); 801 for (r = 0; r < len; ++r) { 802 const PetscInt row = lrows[r] + A->rmap->rstart; 803 ierr = MatSetValues(A, 1, &row, 1, &row, &diag, INSERT_VALUES);CHKERRQ(ierr); 804 } 805 ierr = MatAssemblyBegin(A, MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 806 ierr = MatAssemblyEnd(A, MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 807 } else { 808 ierr = MatZeroRows(mat->A, len, lrows, 0.0, NULL, NULL);CHKERRQ(ierr); 809 } 810 ierr = PetscFree(lrows);CHKERRQ(ierr); 811 812 /* only change matrix nonzero state if pattern was allowed to be changed */ 813 if (!((Mat_SeqAIJ*)(mat->A->data))->keepnonzeropattern) { 814 PetscObjectState state = mat->A->nonzerostate + mat->B->nonzerostate; 815 ierr = MPIU_Allreduce(&state,&A->nonzerostate,1,MPIU_INT64,MPI_SUM,PetscObjectComm((PetscObject)A));CHKERRQ(ierr); 816 } 817 PetscFunctionReturn(0); 818 } 819 820 PetscErrorCode MatZeroRowsColumns_MPIAIJ(Mat A,PetscInt N,const PetscInt rows[],PetscScalar diag,Vec x,Vec b) 821 { 822 Mat_MPIAIJ *l = (Mat_MPIAIJ*)A->data; 823 PetscErrorCode ierr; 824 PetscMPIInt n = A->rmap->n; 825 PetscInt i,j,r,m,p = 0,len = 0; 826 PetscInt *lrows,*owners = A->rmap->range; 827 PetscSFNode *rrows; 828 PetscSF sf; 829 const PetscScalar *xx; 830 PetscScalar *bb,*mask; 831 Vec xmask,lmask; 832 Mat_SeqAIJ *aij = (Mat_SeqAIJ*)l->B->data; 833 const PetscInt *aj, *ii,*ridx; 834 PetscScalar *aa; 835 836 PetscFunctionBegin; 837 /* Create SF where leaves are input rows and roots are owned rows */ 838 ierr = PetscMalloc1(n, &lrows);CHKERRQ(ierr); 839 for (r = 0; r < n; ++r) lrows[r] = -1; 840 ierr = PetscMalloc1(N, &rrows);CHKERRQ(ierr); 841 for (r = 0; r < N; ++r) { 842 const PetscInt idx = rows[r]; 843 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); 844 if (idx < owners[p] || owners[p+1] <= idx) { /* short-circuit the search if the last p owns this row too */ 845 ierr = PetscLayoutFindOwner(A->rmap,idx,&p);CHKERRQ(ierr); 846 } 847 rrows[r].rank = p; 848 rrows[r].index = rows[r] - owners[p]; 849 } 850 ierr = PetscSFCreate(PetscObjectComm((PetscObject) A), &sf);CHKERRQ(ierr); 851 ierr = PetscSFSetGraph(sf, n, N, NULL, PETSC_OWN_POINTER, rrows, PETSC_OWN_POINTER);CHKERRQ(ierr); 852 /* Collect flags for rows to be zeroed */ 853 ierr = PetscSFReduceBegin(sf, MPIU_INT, (PetscInt *) rows, lrows, MPI_LOR);CHKERRQ(ierr); 854 ierr = PetscSFReduceEnd(sf, MPIU_INT, (PetscInt *) rows, lrows, MPI_LOR);CHKERRQ(ierr); 855 ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 856 /* Compress and put in row numbers */ 857 for (r = 0; r < n; ++r) if (lrows[r] >= 0) lrows[len++] = r; 858 /* zero diagonal part of matrix */ 859 ierr = MatZeroRowsColumns(l->A,len,lrows,diag,x,b);CHKERRQ(ierr); 860 /* handle off diagonal part of matrix */ 861 ierr = MatCreateVecs(A,&xmask,NULL);CHKERRQ(ierr); 862 ierr = VecDuplicate(l->lvec,&lmask);CHKERRQ(ierr); 863 ierr = VecGetArray(xmask,&bb);CHKERRQ(ierr); 864 for (i=0; i<len; i++) bb[lrows[i]] = 1; 865 ierr = VecRestoreArray(xmask,&bb);CHKERRQ(ierr); 866 ierr = VecScatterBegin(l->Mvctx,xmask,lmask,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 867 ierr = VecScatterEnd(l->Mvctx,xmask,lmask,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 868 ierr = VecDestroy(&xmask);CHKERRQ(ierr); 869 if (x) { 870 ierr = VecScatterBegin(l->Mvctx,x,l->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 871 ierr = VecScatterEnd(l->Mvctx,x,l->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 872 ierr = VecGetArrayRead(l->lvec,&xx);CHKERRQ(ierr); 873 ierr = VecGetArray(b,&bb);CHKERRQ(ierr); 874 } 875 ierr = VecGetArray(lmask,&mask);CHKERRQ(ierr); 876 /* remove zeroed rows of off diagonal matrix */ 877 ii = aij->i; 878 for (i=0; i<len; i++) { 879 ierr = PetscMemzero(aij->a + ii[lrows[i]],(ii[lrows[i]+1] - ii[lrows[i]])*sizeof(PetscScalar));CHKERRQ(ierr); 880 } 881 /* loop over all elements of off process part of matrix zeroing removed columns*/ 882 if (aij->compressedrow.use) { 883 m = aij->compressedrow.nrows; 884 ii = aij->compressedrow.i; 885 ridx = aij->compressedrow.rindex; 886 for (i=0; i<m; i++) { 887 n = ii[i+1] - ii[i]; 888 aj = aij->j + ii[i]; 889 aa = aij->a + ii[i]; 890 891 for (j=0; j<n; j++) { 892 if (PetscAbsScalar(mask[*aj])) { 893 if (b) bb[*ridx] -= *aa*xx[*aj]; 894 *aa = 0.0; 895 } 896 aa++; 897 aj++; 898 } 899 ridx++; 900 } 901 } else { /* do not use compressed row format */ 902 m = l->B->rmap->n; 903 for (i=0; i<m; i++) { 904 n = ii[i+1] - ii[i]; 905 aj = aij->j + ii[i]; 906 aa = aij->a + ii[i]; 907 for (j=0; j<n; j++) { 908 if (PetscAbsScalar(mask[*aj])) { 909 if (b) bb[i] -= *aa*xx[*aj]; 910 *aa = 0.0; 911 } 912 aa++; 913 aj++; 914 } 915 } 916 } 917 if (x) { 918 ierr = VecRestoreArray(b,&bb);CHKERRQ(ierr); 919 ierr = VecRestoreArrayRead(l->lvec,&xx);CHKERRQ(ierr); 920 } 921 ierr = VecRestoreArray(lmask,&mask);CHKERRQ(ierr); 922 ierr = VecDestroy(&lmask);CHKERRQ(ierr); 923 ierr = PetscFree(lrows);CHKERRQ(ierr); 924 925 /* only change matrix nonzero state if pattern was allowed to be changed */ 926 if (!((Mat_SeqAIJ*)(l->A->data))->keepnonzeropattern) { 927 PetscObjectState state = l->A->nonzerostate + l->B->nonzerostate; 928 ierr = MPIU_Allreduce(&state,&A->nonzerostate,1,MPIU_INT64,MPI_SUM,PetscObjectComm((PetscObject)A));CHKERRQ(ierr); 929 } 930 PetscFunctionReturn(0); 931 } 932 933 PetscErrorCode MatMult_MPIAIJ(Mat A,Vec xx,Vec yy) 934 { 935 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 936 PetscErrorCode ierr; 937 PetscInt nt; 938 VecScatter Mvctx = a->Mvctx; 939 940 PetscFunctionBegin; 941 ierr = VecGetLocalSize(xx,&nt);CHKERRQ(ierr); 942 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); 943 944 ierr = VecScatterBegin(Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 945 ierr = (*a->A->ops->mult)(a->A,xx,yy);CHKERRQ(ierr); 946 ierr = VecScatterEnd(Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 947 ierr = (*a->B->ops->multadd)(a->B,a->lvec,yy,yy);CHKERRQ(ierr); 948 PetscFunctionReturn(0); 949 } 950 951 PetscErrorCode MatMultDiagonalBlock_MPIAIJ(Mat A,Vec bb,Vec xx) 952 { 953 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 954 PetscErrorCode ierr; 955 956 PetscFunctionBegin; 957 ierr = MatMultDiagonalBlock(a->A,bb,xx);CHKERRQ(ierr); 958 PetscFunctionReturn(0); 959 } 960 961 PetscErrorCode MatMultAdd_MPIAIJ(Mat A,Vec xx,Vec yy,Vec zz) 962 { 963 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 964 PetscErrorCode ierr; 965 VecScatter Mvctx = a->Mvctx; 966 967 PetscFunctionBegin; 968 if (a->Mvctx_mpi1_flg) Mvctx = a->Mvctx_mpi1; 969 ierr = VecScatterBegin(Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 970 ierr = (*a->A->ops->multadd)(a->A,xx,yy,zz);CHKERRQ(ierr); 971 ierr = VecScatterEnd(Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 972 ierr = (*a->B->ops->multadd)(a->B,a->lvec,zz,zz);CHKERRQ(ierr); 973 PetscFunctionReturn(0); 974 } 975 976 PetscErrorCode MatMultTranspose_MPIAIJ(Mat A,Vec xx,Vec yy) 977 { 978 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 979 PetscErrorCode ierr; 980 PetscBool merged; 981 982 PetscFunctionBegin; 983 ierr = VecScatterGetMerged(a->Mvctx,&merged);CHKERRQ(ierr); 984 /* do nondiagonal part */ 985 ierr = (*a->B->ops->multtranspose)(a->B,xx,a->lvec);CHKERRQ(ierr); 986 if (!merged) { 987 /* send it on its way */ 988 ierr = VecScatterBegin(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 989 /* do local part */ 990 ierr = (*a->A->ops->multtranspose)(a->A,xx,yy);CHKERRQ(ierr); 991 /* receive remote parts: note this assumes the values are not actually */ 992 /* added in yy until the next line, */ 993 ierr = VecScatterEnd(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 994 } else { 995 /* do local part */ 996 ierr = (*a->A->ops->multtranspose)(a->A,xx,yy);CHKERRQ(ierr); 997 /* send it on its way */ 998 ierr = VecScatterBegin(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 999 /* values actually were received in the Begin() but we need to call this nop */ 1000 ierr = VecScatterEnd(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1001 } 1002 PetscFunctionReturn(0); 1003 } 1004 1005 PetscErrorCode MatIsTranspose_MPIAIJ(Mat Amat,Mat Bmat,PetscReal tol,PetscBool *f) 1006 { 1007 MPI_Comm comm; 1008 Mat_MPIAIJ *Aij = (Mat_MPIAIJ*) Amat->data, *Bij; 1009 Mat Adia = Aij->A, Bdia, Aoff,Boff,*Aoffs,*Boffs; 1010 IS Me,Notme; 1011 PetscErrorCode ierr; 1012 PetscInt M,N,first,last,*notme,i; 1013 PetscMPIInt size; 1014 1015 PetscFunctionBegin; 1016 /* Easy test: symmetric diagonal block */ 1017 Bij = (Mat_MPIAIJ*) Bmat->data; Bdia = Bij->A; 1018 ierr = MatIsTranspose(Adia,Bdia,tol,f);CHKERRQ(ierr); 1019 if (!*f) PetscFunctionReturn(0); 1020 ierr = PetscObjectGetComm((PetscObject)Amat,&comm);CHKERRQ(ierr); 1021 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 1022 if (size == 1) PetscFunctionReturn(0); 1023 1024 /* Hard test: off-diagonal block. This takes a MatCreateSubMatrix. */ 1025 ierr = MatGetSize(Amat,&M,&N);CHKERRQ(ierr); 1026 ierr = MatGetOwnershipRange(Amat,&first,&last);CHKERRQ(ierr); 1027 ierr = PetscMalloc1(N-last+first,¬me);CHKERRQ(ierr); 1028 for (i=0; i<first; i++) notme[i] = i; 1029 for (i=last; i<M; i++) notme[i-last+first] = i; 1030 ierr = ISCreateGeneral(MPI_COMM_SELF,N-last+first,notme,PETSC_COPY_VALUES,&Notme);CHKERRQ(ierr); 1031 ierr = ISCreateStride(MPI_COMM_SELF,last-first,first,1,&Me);CHKERRQ(ierr); 1032 ierr = MatCreateSubMatrices(Amat,1,&Me,&Notme,MAT_INITIAL_MATRIX,&Aoffs);CHKERRQ(ierr); 1033 Aoff = Aoffs[0]; 1034 ierr = MatCreateSubMatrices(Bmat,1,&Notme,&Me,MAT_INITIAL_MATRIX,&Boffs);CHKERRQ(ierr); 1035 Boff = Boffs[0]; 1036 ierr = MatIsTranspose(Aoff,Boff,tol,f);CHKERRQ(ierr); 1037 ierr = MatDestroyMatrices(1,&Aoffs);CHKERRQ(ierr); 1038 ierr = MatDestroyMatrices(1,&Boffs);CHKERRQ(ierr); 1039 ierr = ISDestroy(&Me);CHKERRQ(ierr); 1040 ierr = ISDestroy(&Notme);CHKERRQ(ierr); 1041 ierr = PetscFree(notme);CHKERRQ(ierr); 1042 PetscFunctionReturn(0); 1043 } 1044 1045 PetscErrorCode MatIsSymmetric_MPIAIJ(Mat A,PetscReal tol,PetscBool *f) 1046 { 1047 PetscErrorCode ierr; 1048 1049 PetscFunctionBegin; 1050 ierr = MatIsTranspose_MPIAIJ(A,A,tol,f);CHKERRQ(ierr); 1051 PetscFunctionReturn(0); 1052 } 1053 1054 PetscErrorCode MatMultTransposeAdd_MPIAIJ(Mat A,Vec xx,Vec yy,Vec zz) 1055 { 1056 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 1057 PetscErrorCode ierr; 1058 1059 PetscFunctionBegin; 1060 /* do nondiagonal part */ 1061 ierr = (*a->B->ops->multtranspose)(a->B,xx,a->lvec);CHKERRQ(ierr); 1062 /* send it on its way */ 1063 ierr = VecScatterBegin(a->Mvctx,a->lvec,zz,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1064 /* do local part */ 1065 ierr = (*a->A->ops->multtransposeadd)(a->A,xx,yy,zz);CHKERRQ(ierr); 1066 /* receive remote parts */ 1067 ierr = VecScatterEnd(a->Mvctx,a->lvec,zz,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1068 PetscFunctionReturn(0); 1069 } 1070 1071 /* 1072 This only works correctly for square matrices where the subblock A->A is the 1073 diagonal block 1074 */ 1075 PetscErrorCode MatGetDiagonal_MPIAIJ(Mat A,Vec v) 1076 { 1077 PetscErrorCode ierr; 1078 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 1079 1080 PetscFunctionBegin; 1081 if (A->rmap->N != A->cmap->N) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"Supports only square matrix where A->A is diag block"); 1082 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"); 1083 ierr = MatGetDiagonal(a->A,v);CHKERRQ(ierr); 1084 PetscFunctionReturn(0); 1085 } 1086 1087 PetscErrorCode MatScale_MPIAIJ(Mat A,PetscScalar aa) 1088 { 1089 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 1090 PetscErrorCode ierr; 1091 1092 PetscFunctionBegin; 1093 ierr = MatScale(a->A,aa);CHKERRQ(ierr); 1094 ierr = MatScale(a->B,aa);CHKERRQ(ierr); 1095 PetscFunctionReturn(0); 1096 } 1097 1098 PetscErrorCode MatDestroy_MPIAIJ(Mat mat) 1099 { 1100 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 1101 PetscErrorCode ierr; 1102 1103 PetscFunctionBegin; 1104 #if defined(PETSC_USE_LOG) 1105 PetscLogObjectState((PetscObject)mat,"Rows=%D, Cols=%D",mat->rmap->N,mat->cmap->N); 1106 #endif 1107 ierr = MatStashDestroy_Private(&mat->stash);CHKERRQ(ierr); 1108 ierr = VecDestroy(&aij->diag);CHKERRQ(ierr); 1109 ierr = MatDestroy(&aij->A);CHKERRQ(ierr); 1110 ierr = MatDestroy(&aij->B);CHKERRQ(ierr); 1111 #if defined(PETSC_USE_CTABLE) 1112 ierr = PetscTableDestroy(&aij->colmap);CHKERRQ(ierr); 1113 #else 1114 ierr = PetscFree(aij->colmap);CHKERRQ(ierr); 1115 #endif 1116 ierr = PetscFree(aij->garray);CHKERRQ(ierr); 1117 ierr = VecDestroy(&aij->lvec);CHKERRQ(ierr); 1118 ierr = VecScatterDestroy(&aij->Mvctx);CHKERRQ(ierr); 1119 if (aij->Mvctx_mpi1) {ierr = VecScatterDestroy(&aij->Mvctx_mpi1);CHKERRQ(ierr);} 1120 ierr = PetscFree2(aij->rowvalues,aij->rowindices);CHKERRQ(ierr); 1121 ierr = PetscFree(aij->ld);CHKERRQ(ierr); 1122 ierr = PetscFree(mat->data);CHKERRQ(ierr); 1123 1124 ierr = PetscObjectChangeTypeName((PetscObject)mat,0);CHKERRQ(ierr); 1125 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatStoreValues_C",NULL);CHKERRQ(ierr); 1126 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatRetrieveValues_C",NULL);CHKERRQ(ierr); 1127 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatIsTranspose_C",NULL);CHKERRQ(ierr); 1128 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatMPIAIJSetPreallocation_C",NULL);CHKERRQ(ierr); 1129 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatResetPreallocation_C",NULL);CHKERRQ(ierr); 1130 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatMPIAIJSetPreallocationCSR_C",NULL);CHKERRQ(ierr); 1131 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatDiagonalScaleLocal_C",NULL);CHKERRQ(ierr); 1132 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatConvert_mpiaij_mpisbaij_C",NULL);CHKERRQ(ierr); 1133 #if defined(PETSC_HAVE_ELEMENTAL) 1134 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatConvert_mpiaij_elemental_C",NULL);CHKERRQ(ierr); 1135 #endif 1136 #if defined(PETSC_HAVE_HYPRE) 1137 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatConvert_mpiaij_hypre_C",NULL);CHKERRQ(ierr); 1138 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatMatMatMult_transpose_mpiaij_mpiaij_C",NULL);CHKERRQ(ierr); 1139 #endif 1140 PetscFunctionReturn(0); 1141 } 1142 1143 PetscErrorCode MatView_MPIAIJ_Binary(Mat mat,PetscViewer viewer) 1144 { 1145 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 1146 Mat_SeqAIJ *A = (Mat_SeqAIJ*)aij->A->data; 1147 Mat_SeqAIJ *B = (Mat_SeqAIJ*)aij->B->data; 1148 PetscErrorCode ierr; 1149 PetscMPIInt rank,size,tag = ((PetscObject)viewer)->tag; 1150 int fd; 1151 PetscInt nz,header[4],*row_lengths,*range=0,rlen,i; 1152 PetscInt nzmax,*column_indices,j,k,col,*garray = aij->garray,cnt,cstart = mat->cmap->rstart,rnz = 0; 1153 PetscScalar *column_values; 1154 PetscInt message_count,flowcontrolcount; 1155 FILE *file; 1156 1157 PetscFunctionBegin; 1158 ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr); 1159 ierr = MPI_Comm_size(PetscObjectComm((PetscObject)mat),&size);CHKERRQ(ierr); 1160 nz = A->nz + B->nz; 1161 ierr = PetscViewerBinaryGetDescriptor(viewer,&fd);CHKERRQ(ierr); 1162 if (!rank) { 1163 header[0] = MAT_FILE_CLASSID; 1164 header[1] = mat->rmap->N; 1165 header[2] = mat->cmap->N; 1166 1167 ierr = MPI_Reduce(&nz,&header[3],1,MPIU_INT,MPI_SUM,0,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1168 ierr = PetscBinaryWrite(fd,header,4,PETSC_INT,PETSC_TRUE);CHKERRQ(ierr); 1169 /* get largest number of rows any processor has */ 1170 rlen = mat->rmap->n; 1171 range = mat->rmap->range; 1172 for (i=1; i<size; i++) rlen = PetscMax(rlen,range[i+1] - range[i]); 1173 } else { 1174 ierr = MPI_Reduce(&nz,0,1,MPIU_INT,MPI_SUM,0,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1175 rlen = mat->rmap->n; 1176 } 1177 1178 /* load up the local row counts */ 1179 ierr = PetscMalloc1(rlen+1,&row_lengths);CHKERRQ(ierr); 1180 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]; 1181 1182 /* store the row lengths to the file */ 1183 ierr = PetscViewerFlowControlStart(viewer,&message_count,&flowcontrolcount);CHKERRQ(ierr); 1184 if (!rank) { 1185 ierr = PetscBinaryWrite(fd,row_lengths,mat->rmap->n,PETSC_INT,PETSC_TRUE);CHKERRQ(ierr); 1186 for (i=1; i<size; i++) { 1187 ierr = PetscViewerFlowControlStepMaster(viewer,i,&message_count,flowcontrolcount);CHKERRQ(ierr); 1188 rlen = range[i+1] - range[i]; 1189 ierr = MPIULong_Recv(row_lengths,rlen,MPIU_INT,i,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1190 ierr = PetscBinaryWrite(fd,row_lengths,rlen,PETSC_INT,PETSC_TRUE);CHKERRQ(ierr); 1191 } 1192 ierr = PetscViewerFlowControlEndMaster(viewer,&message_count);CHKERRQ(ierr); 1193 } else { 1194 ierr = PetscViewerFlowControlStepWorker(viewer,rank,&message_count);CHKERRQ(ierr); 1195 ierr = MPIULong_Send(row_lengths,mat->rmap->n,MPIU_INT,0,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1196 ierr = PetscViewerFlowControlEndWorker(viewer,&message_count);CHKERRQ(ierr); 1197 } 1198 ierr = PetscFree(row_lengths);CHKERRQ(ierr); 1199 1200 /* load up the local column indices */ 1201 nzmax = nz; /* th processor needs space a largest processor needs */ 1202 ierr = MPI_Reduce(&nz,&nzmax,1,MPIU_INT,MPI_MAX,0,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1203 ierr = PetscMalloc1(nzmax+1,&column_indices);CHKERRQ(ierr); 1204 cnt = 0; 1205 for (i=0; i<mat->rmap->n; i++) { 1206 for (j=B->i[i]; j<B->i[i+1]; j++) { 1207 if ((col = garray[B->j[j]]) > cstart) break; 1208 column_indices[cnt++] = col; 1209 } 1210 for (k=A->i[i]; k<A->i[i+1]; k++) column_indices[cnt++] = A->j[k] + cstart; 1211 for (; j<B->i[i+1]; j++) column_indices[cnt++] = garray[B->j[j]]; 1212 } 1213 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); 1214 1215 /* store the column indices to the file */ 1216 ierr = PetscViewerFlowControlStart(viewer,&message_count,&flowcontrolcount);CHKERRQ(ierr); 1217 if (!rank) { 1218 MPI_Status status; 1219 ierr = PetscBinaryWrite(fd,column_indices,nz,PETSC_INT,PETSC_TRUE);CHKERRQ(ierr); 1220 for (i=1; i<size; i++) { 1221 ierr = PetscViewerFlowControlStepMaster(viewer,i,&message_count,flowcontrolcount);CHKERRQ(ierr); 1222 ierr = MPI_Recv(&rnz,1,MPIU_INT,i,tag,PetscObjectComm((PetscObject)mat),&status);CHKERRQ(ierr); 1223 if (rnz > nzmax) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_LIB,"Internal PETSc error: nz = %D nzmax = %D",nz,nzmax); 1224 ierr = MPIULong_Recv(column_indices,rnz,MPIU_INT,i,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1225 ierr = PetscBinaryWrite(fd,column_indices,rnz,PETSC_INT,PETSC_TRUE);CHKERRQ(ierr); 1226 } 1227 ierr = PetscViewerFlowControlEndMaster(viewer,&message_count);CHKERRQ(ierr); 1228 } else { 1229 ierr = PetscViewerFlowControlStepWorker(viewer,rank,&message_count);CHKERRQ(ierr); 1230 ierr = MPI_Send(&nz,1,MPIU_INT,0,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1231 ierr = MPIULong_Send(column_indices,nz,MPIU_INT,0,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1232 ierr = PetscViewerFlowControlEndWorker(viewer,&message_count);CHKERRQ(ierr); 1233 } 1234 ierr = PetscFree(column_indices);CHKERRQ(ierr); 1235 1236 /* load up the local column values */ 1237 ierr = PetscMalloc1(nzmax+1,&column_values);CHKERRQ(ierr); 1238 cnt = 0; 1239 for (i=0; i<mat->rmap->n; i++) { 1240 for (j=B->i[i]; j<B->i[i+1]; j++) { 1241 if (garray[B->j[j]] > cstart) break; 1242 column_values[cnt++] = B->a[j]; 1243 } 1244 for (k=A->i[i]; k<A->i[i+1]; k++) column_values[cnt++] = A->a[k]; 1245 for (; j<B->i[i+1]; j++) column_values[cnt++] = B->a[j]; 1246 } 1247 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); 1248 1249 /* store the column values to the file */ 1250 ierr = PetscViewerFlowControlStart(viewer,&message_count,&flowcontrolcount);CHKERRQ(ierr); 1251 if (!rank) { 1252 MPI_Status status; 1253 ierr = PetscBinaryWrite(fd,column_values,nz,PETSC_SCALAR,PETSC_TRUE);CHKERRQ(ierr); 1254 for (i=1; i<size; i++) { 1255 ierr = PetscViewerFlowControlStepMaster(viewer,i,&message_count,flowcontrolcount);CHKERRQ(ierr); 1256 ierr = MPI_Recv(&rnz,1,MPIU_INT,i,tag,PetscObjectComm((PetscObject)mat),&status);CHKERRQ(ierr); 1257 if (rnz > nzmax) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_LIB,"Internal PETSc error: nz = %D nzmax = %D",nz,nzmax); 1258 ierr = MPIULong_Recv(column_values,rnz,MPIU_SCALAR,i,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1259 ierr = PetscBinaryWrite(fd,column_values,rnz,PETSC_SCALAR,PETSC_TRUE);CHKERRQ(ierr); 1260 } 1261 ierr = PetscViewerFlowControlEndMaster(viewer,&message_count);CHKERRQ(ierr); 1262 } else { 1263 ierr = PetscViewerFlowControlStepWorker(viewer,rank,&message_count);CHKERRQ(ierr); 1264 ierr = MPI_Send(&nz,1,MPIU_INT,0,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1265 ierr = MPIULong_Send(column_values,nz,MPIU_SCALAR,0,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1266 ierr = PetscViewerFlowControlEndWorker(viewer,&message_count);CHKERRQ(ierr); 1267 } 1268 ierr = PetscFree(column_values);CHKERRQ(ierr); 1269 1270 ierr = PetscViewerBinaryGetInfoPointer(viewer,&file);CHKERRQ(ierr); 1271 if (file) fprintf(file,"-matload_block_size %d\n",(int)PetscAbs(mat->rmap->bs)); 1272 PetscFunctionReturn(0); 1273 } 1274 1275 #include <petscdraw.h> 1276 PetscErrorCode MatView_MPIAIJ_ASCIIorDraworSocket(Mat mat,PetscViewer viewer) 1277 { 1278 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 1279 PetscErrorCode ierr; 1280 PetscMPIInt rank = aij->rank,size = aij->size; 1281 PetscBool isdraw,iascii,isbinary; 1282 PetscViewer sviewer; 1283 PetscViewerFormat format; 1284 1285 PetscFunctionBegin; 1286 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERDRAW,&isdraw);CHKERRQ(ierr); 1287 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii);CHKERRQ(ierr); 1288 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERBINARY,&isbinary);CHKERRQ(ierr); 1289 if (iascii) { 1290 ierr = PetscViewerGetFormat(viewer,&format);CHKERRQ(ierr); 1291 if (format == PETSC_VIEWER_LOAD_BALANCE) { 1292 PetscInt i,nmax = 0,nmin = PETSC_MAX_INT,navg = 0,*nz,nzlocal = ((Mat_SeqAIJ*) (aij->A->data))->nz + ((Mat_SeqAIJ*) (aij->B->data))->nz; 1293 ierr = PetscMalloc1(size,&nz);CHKERRQ(ierr); 1294 ierr = MPI_Allgather(&nzlocal,1,MPIU_INT,nz,1,MPIU_INT,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1295 for (i=0; i<(PetscInt)size; i++) { 1296 nmax = PetscMax(nmax,nz[i]); 1297 nmin = PetscMin(nmin,nz[i]); 1298 navg += nz[i]; 1299 } 1300 ierr = PetscFree(nz);CHKERRQ(ierr); 1301 navg = navg/size; 1302 ierr = PetscViewerASCIIPrintf(viewer,"Load Balance - Nonzeros: Min %D avg %D max %D\n",nmin,navg,nmax);CHKERRQ(ierr); 1303 PetscFunctionReturn(0); 1304 } 1305 ierr = PetscViewerGetFormat(viewer,&format);CHKERRQ(ierr); 1306 if (format == PETSC_VIEWER_ASCII_INFO_DETAIL) { 1307 MatInfo info; 1308 PetscBool inodes; 1309 1310 ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr); 1311 ierr = MatGetInfo(mat,MAT_LOCAL,&info);CHKERRQ(ierr); 1312 ierr = MatInodeGetInodeSizes(aij->A,NULL,(PetscInt**)&inodes,NULL);CHKERRQ(ierr); 1313 ierr = PetscViewerASCIIPushSynchronized(viewer);CHKERRQ(ierr); 1314 if (!inodes) { 1315 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"[%d] Local rows %D nz %D nz alloced %D mem %g, not using I-node routines\n", 1316 rank,mat->rmap->n,(PetscInt)info.nz_used,(PetscInt)info.nz_allocated,(double)info.memory);CHKERRQ(ierr); 1317 } else { 1318 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"[%d] Local rows %D nz %D nz alloced %D mem %g, using I-node routines\n", 1319 rank,mat->rmap->n,(PetscInt)info.nz_used,(PetscInt)info.nz_allocated,(double)info.memory);CHKERRQ(ierr); 1320 } 1321 ierr = MatGetInfo(aij->A,MAT_LOCAL,&info);CHKERRQ(ierr); 1322 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"[%d] on-diagonal part: nz %D \n",rank,(PetscInt)info.nz_used);CHKERRQ(ierr); 1323 ierr = MatGetInfo(aij->B,MAT_LOCAL,&info);CHKERRQ(ierr); 1324 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"[%d] off-diagonal part: nz %D \n",rank,(PetscInt)info.nz_used);CHKERRQ(ierr); 1325 ierr = PetscViewerFlush(viewer);CHKERRQ(ierr); 1326 ierr = PetscViewerASCIIPopSynchronized(viewer);CHKERRQ(ierr); 1327 ierr = PetscViewerASCIIPrintf(viewer,"Information on VecScatter used in matrix-vector product: \n");CHKERRQ(ierr); 1328 ierr = VecScatterView(aij->Mvctx,viewer);CHKERRQ(ierr); 1329 PetscFunctionReturn(0); 1330 } else if (format == PETSC_VIEWER_ASCII_INFO) { 1331 PetscInt inodecount,inodelimit,*inodes; 1332 ierr = MatInodeGetInodeSizes(aij->A,&inodecount,&inodes,&inodelimit);CHKERRQ(ierr); 1333 if (inodes) { 1334 ierr = PetscViewerASCIIPrintf(viewer,"using I-node (on process 0) routines: found %D nodes, limit used is %D\n",inodecount,inodelimit);CHKERRQ(ierr); 1335 } else { 1336 ierr = PetscViewerASCIIPrintf(viewer,"not using I-node (on process 0) routines\n");CHKERRQ(ierr); 1337 } 1338 PetscFunctionReturn(0); 1339 } else if (format == PETSC_VIEWER_ASCII_FACTOR_INFO) { 1340 PetscFunctionReturn(0); 1341 } 1342 } else if (isbinary) { 1343 if (size == 1) { 1344 ierr = PetscObjectSetName((PetscObject)aij->A,((PetscObject)mat)->name);CHKERRQ(ierr); 1345 ierr = MatView(aij->A,viewer);CHKERRQ(ierr); 1346 } else { 1347 ierr = MatView_MPIAIJ_Binary(mat,viewer);CHKERRQ(ierr); 1348 } 1349 PetscFunctionReturn(0); 1350 } else if (isdraw) { 1351 PetscDraw draw; 1352 PetscBool isnull; 1353 ierr = PetscViewerDrawGetDraw(viewer,0,&draw);CHKERRQ(ierr); 1354 ierr = PetscDrawIsNull(draw,&isnull);CHKERRQ(ierr); 1355 if (isnull) PetscFunctionReturn(0); 1356 } 1357 1358 { 1359 /* assemble the entire matrix onto first processor. */ 1360 Mat A; 1361 Mat_SeqAIJ *Aloc; 1362 PetscInt M = mat->rmap->N,N = mat->cmap->N,m,*ai,*aj,row,*cols,i,*ct; 1363 MatScalar *a; 1364 1365 ierr = MatCreate(PetscObjectComm((PetscObject)mat),&A);CHKERRQ(ierr); 1366 if (!rank) { 1367 ierr = MatSetSizes(A,M,N,M,N);CHKERRQ(ierr); 1368 } else { 1369 ierr = MatSetSizes(A,0,0,M,N);CHKERRQ(ierr); 1370 } 1371 /* This is just a temporary matrix, so explicitly using MATMPIAIJ is probably best */ 1372 ierr = MatSetType(A,MATMPIAIJ);CHKERRQ(ierr); 1373 ierr = MatMPIAIJSetPreallocation(A,0,NULL,0,NULL);CHKERRQ(ierr); 1374 ierr = MatSetOption(A,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 1375 ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)A);CHKERRQ(ierr); 1376 1377 /* copy over the A part */ 1378 Aloc = (Mat_SeqAIJ*)aij->A->data; 1379 m = aij->A->rmap->n; ai = Aloc->i; aj = Aloc->j; a = Aloc->a; 1380 row = mat->rmap->rstart; 1381 for (i=0; i<ai[m]; i++) aj[i] += mat->cmap->rstart; 1382 for (i=0; i<m; i++) { 1383 ierr = MatSetValues(A,1,&row,ai[i+1]-ai[i],aj,a,INSERT_VALUES);CHKERRQ(ierr); 1384 row++; 1385 a += ai[i+1]-ai[i]; aj += ai[i+1]-ai[i]; 1386 } 1387 aj = Aloc->j; 1388 for (i=0; i<ai[m]; i++) aj[i] -= mat->cmap->rstart; 1389 1390 /* copy over the B part */ 1391 Aloc = (Mat_SeqAIJ*)aij->B->data; 1392 m = aij->B->rmap->n; ai = Aloc->i; aj = Aloc->j; a = Aloc->a; 1393 row = mat->rmap->rstart; 1394 ierr = PetscMalloc1(ai[m]+1,&cols);CHKERRQ(ierr); 1395 ct = cols; 1396 for (i=0; i<ai[m]; i++) cols[i] = aij->garray[aj[i]]; 1397 for (i=0; i<m; i++) { 1398 ierr = MatSetValues(A,1,&row,ai[i+1]-ai[i],cols,a,INSERT_VALUES);CHKERRQ(ierr); 1399 row++; 1400 a += ai[i+1]-ai[i]; cols += ai[i+1]-ai[i]; 1401 } 1402 ierr = PetscFree(ct);CHKERRQ(ierr); 1403 ierr = MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1404 ierr = MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1405 /* 1406 Everyone has to call to draw the matrix since the graphics waits are 1407 synchronized across all processors that share the PetscDraw object 1408 */ 1409 ierr = PetscViewerGetSubViewer(viewer,PETSC_COMM_SELF,&sviewer);CHKERRQ(ierr); 1410 if (!rank) { 1411 ierr = PetscObjectSetName((PetscObject)((Mat_MPIAIJ*)(A->data))->A,((PetscObject)mat)->name);CHKERRQ(ierr); 1412 ierr = MatView_SeqAIJ(((Mat_MPIAIJ*)(A->data))->A,sviewer);CHKERRQ(ierr); 1413 } 1414 ierr = PetscViewerRestoreSubViewer(viewer,PETSC_COMM_SELF,&sviewer);CHKERRQ(ierr); 1415 ierr = PetscViewerFlush(viewer);CHKERRQ(ierr); 1416 ierr = MatDestroy(&A);CHKERRQ(ierr); 1417 } 1418 PetscFunctionReturn(0); 1419 } 1420 1421 PetscErrorCode MatView_MPIAIJ(Mat mat,PetscViewer viewer) 1422 { 1423 PetscErrorCode ierr; 1424 PetscBool iascii,isdraw,issocket,isbinary; 1425 1426 PetscFunctionBegin; 1427 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii);CHKERRQ(ierr); 1428 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERDRAW,&isdraw);CHKERRQ(ierr); 1429 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERBINARY,&isbinary);CHKERRQ(ierr); 1430 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERSOCKET,&issocket);CHKERRQ(ierr); 1431 if (iascii || isdraw || isbinary || issocket) { 1432 ierr = MatView_MPIAIJ_ASCIIorDraworSocket(mat,viewer);CHKERRQ(ierr); 1433 } 1434 PetscFunctionReturn(0); 1435 } 1436 1437 PetscErrorCode MatSOR_MPIAIJ(Mat matin,Vec bb,PetscReal omega,MatSORType flag,PetscReal fshift,PetscInt its,PetscInt lits,Vec xx) 1438 { 1439 Mat_MPIAIJ *mat = (Mat_MPIAIJ*)matin->data; 1440 PetscErrorCode ierr; 1441 Vec bb1 = 0; 1442 PetscBool hasop; 1443 1444 PetscFunctionBegin; 1445 if (flag == SOR_APPLY_UPPER) { 1446 ierr = (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);CHKERRQ(ierr); 1447 PetscFunctionReturn(0); 1448 } 1449 1450 if (its > 1 || ~flag & SOR_ZERO_INITIAL_GUESS || flag & SOR_EISENSTAT) { 1451 ierr = VecDuplicate(bb,&bb1);CHKERRQ(ierr); 1452 } 1453 1454 if ((flag & SOR_LOCAL_SYMMETRIC_SWEEP) == SOR_LOCAL_SYMMETRIC_SWEEP) { 1455 if (flag & SOR_ZERO_INITIAL_GUESS) { 1456 ierr = (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);CHKERRQ(ierr); 1457 its--; 1458 } 1459 1460 while (its--) { 1461 ierr = VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1462 ierr = VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1463 1464 /* update rhs: bb1 = bb - B*x */ 1465 ierr = VecScale(mat->lvec,-1.0);CHKERRQ(ierr); 1466 ierr = (*mat->B->ops->multadd)(mat->B,mat->lvec,bb,bb1);CHKERRQ(ierr); 1467 1468 /* local sweep */ 1469 ierr = (*mat->A->ops->sor)(mat->A,bb1,omega,SOR_SYMMETRIC_SWEEP,fshift,lits,1,xx);CHKERRQ(ierr); 1470 } 1471 } else if (flag & SOR_LOCAL_FORWARD_SWEEP) { 1472 if (flag & SOR_ZERO_INITIAL_GUESS) { 1473 ierr = (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);CHKERRQ(ierr); 1474 its--; 1475 } 1476 while (its--) { 1477 ierr = VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1478 ierr = VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1479 1480 /* update rhs: bb1 = bb - B*x */ 1481 ierr = VecScale(mat->lvec,-1.0);CHKERRQ(ierr); 1482 ierr = (*mat->B->ops->multadd)(mat->B,mat->lvec,bb,bb1);CHKERRQ(ierr); 1483 1484 /* local sweep */ 1485 ierr = (*mat->A->ops->sor)(mat->A,bb1,omega,SOR_FORWARD_SWEEP,fshift,lits,1,xx);CHKERRQ(ierr); 1486 } 1487 } else if (flag & SOR_LOCAL_BACKWARD_SWEEP) { 1488 if (flag & SOR_ZERO_INITIAL_GUESS) { 1489 ierr = (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);CHKERRQ(ierr); 1490 its--; 1491 } 1492 while (its--) { 1493 ierr = VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1494 ierr = VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1495 1496 /* update rhs: bb1 = bb - B*x */ 1497 ierr = VecScale(mat->lvec,-1.0);CHKERRQ(ierr); 1498 ierr = (*mat->B->ops->multadd)(mat->B,mat->lvec,bb,bb1);CHKERRQ(ierr); 1499 1500 /* local sweep */ 1501 ierr = (*mat->A->ops->sor)(mat->A,bb1,omega,SOR_BACKWARD_SWEEP,fshift,lits,1,xx);CHKERRQ(ierr); 1502 } 1503 } else if (flag & SOR_EISENSTAT) { 1504 Vec xx1; 1505 1506 ierr = VecDuplicate(bb,&xx1);CHKERRQ(ierr); 1507 ierr = (*mat->A->ops->sor)(mat->A,bb,omega,(MatSORType)(SOR_ZERO_INITIAL_GUESS | SOR_LOCAL_BACKWARD_SWEEP),fshift,lits,1,xx);CHKERRQ(ierr); 1508 1509 ierr = VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1510 ierr = VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1511 if (!mat->diag) { 1512 ierr = MatCreateVecs(matin,&mat->diag,NULL);CHKERRQ(ierr); 1513 ierr = MatGetDiagonal(matin,mat->diag);CHKERRQ(ierr); 1514 } 1515 ierr = MatHasOperation(matin,MATOP_MULT_DIAGONAL_BLOCK,&hasop);CHKERRQ(ierr); 1516 if (hasop) { 1517 ierr = MatMultDiagonalBlock(matin,xx,bb1);CHKERRQ(ierr); 1518 } else { 1519 ierr = VecPointwiseMult(bb1,mat->diag,xx);CHKERRQ(ierr); 1520 } 1521 ierr = VecAYPX(bb1,(omega-2.0)/omega,bb);CHKERRQ(ierr); 1522 1523 ierr = MatMultAdd(mat->B,mat->lvec,bb1,bb1);CHKERRQ(ierr); 1524 1525 /* local sweep */ 1526 ierr = (*mat->A->ops->sor)(mat->A,bb1,omega,(MatSORType)(SOR_ZERO_INITIAL_GUESS | SOR_LOCAL_FORWARD_SWEEP),fshift,lits,1,xx1);CHKERRQ(ierr); 1527 ierr = VecAXPY(xx,1.0,xx1);CHKERRQ(ierr); 1528 ierr = VecDestroy(&xx1);CHKERRQ(ierr); 1529 } else SETERRQ(PetscObjectComm((PetscObject)matin),PETSC_ERR_SUP,"Parallel SOR not supported"); 1530 1531 ierr = VecDestroy(&bb1);CHKERRQ(ierr); 1532 1533 matin->factorerrortype = mat->A->factorerrortype; 1534 PetscFunctionReturn(0); 1535 } 1536 1537 PetscErrorCode MatPermute_MPIAIJ(Mat A,IS rowp,IS colp,Mat *B) 1538 { 1539 Mat aA,aB,Aperm; 1540 const PetscInt *rwant,*cwant,*gcols,*ai,*bi,*aj,*bj; 1541 PetscScalar *aa,*ba; 1542 PetscInt i,j,m,n,ng,anz,bnz,*dnnz,*onnz,*tdnnz,*tonnz,*rdest,*cdest,*work,*gcdest; 1543 PetscSF rowsf,sf; 1544 IS parcolp = NULL; 1545 PetscBool done; 1546 PetscErrorCode ierr; 1547 1548 PetscFunctionBegin; 1549 ierr = MatGetLocalSize(A,&m,&n);CHKERRQ(ierr); 1550 ierr = ISGetIndices(rowp,&rwant);CHKERRQ(ierr); 1551 ierr = ISGetIndices(colp,&cwant);CHKERRQ(ierr); 1552 ierr = PetscMalloc3(PetscMax(m,n),&work,m,&rdest,n,&cdest);CHKERRQ(ierr); 1553 1554 /* Invert row permutation to find out where my rows should go */ 1555 ierr = PetscSFCreate(PetscObjectComm((PetscObject)A),&rowsf);CHKERRQ(ierr); 1556 ierr = PetscSFSetGraphLayout(rowsf,A->rmap,A->rmap->n,NULL,PETSC_OWN_POINTER,rwant);CHKERRQ(ierr); 1557 ierr = PetscSFSetFromOptions(rowsf);CHKERRQ(ierr); 1558 for (i=0; i<m; i++) work[i] = A->rmap->rstart + i; 1559 ierr = PetscSFReduceBegin(rowsf,MPIU_INT,work,rdest,MPIU_REPLACE);CHKERRQ(ierr); 1560 ierr = PetscSFReduceEnd(rowsf,MPIU_INT,work,rdest,MPIU_REPLACE);CHKERRQ(ierr); 1561 1562 /* Invert column permutation to find out where my columns should go */ 1563 ierr = PetscSFCreate(PetscObjectComm((PetscObject)A),&sf);CHKERRQ(ierr); 1564 ierr = PetscSFSetGraphLayout(sf,A->cmap,A->cmap->n,NULL,PETSC_OWN_POINTER,cwant);CHKERRQ(ierr); 1565 ierr = PetscSFSetFromOptions(sf);CHKERRQ(ierr); 1566 for (i=0; i<n; i++) work[i] = A->cmap->rstart + i; 1567 ierr = PetscSFReduceBegin(sf,MPIU_INT,work,cdest,MPIU_REPLACE);CHKERRQ(ierr); 1568 ierr = PetscSFReduceEnd(sf,MPIU_INT,work,cdest,MPIU_REPLACE);CHKERRQ(ierr); 1569 ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 1570 1571 ierr = ISRestoreIndices(rowp,&rwant);CHKERRQ(ierr); 1572 ierr = ISRestoreIndices(colp,&cwant);CHKERRQ(ierr); 1573 ierr = MatMPIAIJGetSeqAIJ(A,&aA,&aB,&gcols);CHKERRQ(ierr); 1574 1575 /* Find out where my gcols should go */ 1576 ierr = MatGetSize(aB,NULL,&ng);CHKERRQ(ierr); 1577 ierr = PetscMalloc1(ng,&gcdest);CHKERRQ(ierr); 1578 ierr = PetscSFCreate(PetscObjectComm((PetscObject)A),&sf);CHKERRQ(ierr); 1579 ierr = PetscSFSetGraphLayout(sf,A->cmap,ng,NULL,PETSC_OWN_POINTER,gcols);CHKERRQ(ierr); 1580 ierr = PetscSFSetFromOptions(sf);CHKERRQ(ierr); 1581 ierr = PetscSFBcastBegin(sf,MPIU_INT,cdest,gcdest);CHKERRQ(ierr); 1582 ierr = PetscSFBcastEnd(sf,MPIU_INT,cdest,gcdest);CHKERRQ(ierr); 1583 ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 1584 1585 ierr = PetscCalloc4(m,&dnnz,m,&onnz,m,&tdnnz,m,&tonnz);CHKERRQ(ierr); 1586 ierr = MatGetRowIJ(aA,0,PETSC_FALSE,PETSC_FALSE,&anz,&ai,&aj,&done);CHKERRQ(ierr); 1587 ierr = MatGetRowIJ(aB,0,PETSC_FALSE,PETSC_FALSE,&bnz,&bi,&bj,&done);CHKERRQ(ierr); 1588 for (i=0; i<m; i++) { 1589 PetscInt row = rdest[i],rowner; 1590 ierr = PetscLayoutFindOwner(A->rmap,row,&rowner);CHKERRQ(ierr); 1591 for (j=ai[i]; j<ai[i+1]; j++) { 1592 PetscInt cowner,col = cdest[aj[j]]; 1593 ierr = PetscLayoutFindOwner(A->cmap,col,&cowner);CHKERRQ(ierr); /* Could build an index for the columns to eliminate this search */ 1594 if (rowner == cowner) dnnz[i]++; 1595 else onnz[i]++; 1596 } 1597 for (j=bi[i]; j<bi[i+1]; j++) { 1598 PetscInt cowner,col = gcdest[bj[j]]; 1599 ierr = PetscLayoutFindOwner(A->cmap,col,&cowner);CHKERRQ(ierr); 1600 if (rowner == cowner) dnnz[i]++; 1601 else onnz[i]++; 1602 } 1603 } 1604 ierr = PetscSFBcastBegin(rowsf,MPIU_INT,dnnz,tdnnz);CHKERRQ(ierr); 1605 ierr = PetscSFBcastEnd(rowsf,MPIU_INT,dnnz,tdnnz);CHKERRQ(ierr); 1606 ierr = PetscSFBcastBegin(rowsf,MPIU_INT,onnz,tonnz);CHKERRQ(ierr); 1607 ierr = PetscSFBcastEnd(rowsf,MPIU_INT,onnz,tonnz);CHKERRQ(ierr); 1608 ierr = PetscSFDestroy(&rowsf);CHKERRQ(ierr); 1609 1610 ierr = MatCreateAIJ(PetscObjectComm((PetscObject)A),A->rmap->n,A->cmap->n,A->rmap->N,A->cmap->N,0,tdnnz,0,tonnz,&Aperm);CHKERRQ(ierr); 1611 ierr = MatSeqAIJGetArray(aA,&aa);CHKERRQ(ierr); 1612 ierr = MatSeqAIJGetArray(aB,&ba);CHKERRQ(ierr); 1613 for (i=0; i<m; i++) { 1614 PetscInt *acols = dnnz,*bcols = onnz; /* Repurpose now-unneeded arrays */ 1615 PetscInt j0,rowlen; 1616 rowlen = ai[i+1] - ai[i]; 1617 for (j0=j=0; j<rowlen; j0=j) { /* rowlen could be larger than number of rows m, so sum in batches */ 1618 for ( ; j<PetscMin(rowlen,j0+m); j++) acols[j-j0] = cdest[aj[ai[i]+j]]; 1619 ierr = MatSetValues(Aperm,1,&rdest[i],j-j0,acols,aa+ai[i]+j0,INSERT_VALUES);CHKERRQ(ierr); 1620 } 1621 rowlen = bi[i+1] - bi[i]; 1622 for (j0=j=0; j<rowlen; j0=j) { 1623 for ( ; j<PetscMin(rowlen,j0+m); j++) bcols[j-j0] = gcdest[bj[bi[i]+j]]; 1624 ierr = MatSetValues(Aperm,1,&rdest[i],j-j0,bcols,ba+bi[i]+j0,INSERT_VALUES);CHKERRQ(ierr); 1625 } 1626 } 1627 ierr = MatAssemblyBegin(Aperm,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1628 ierr = MatAssemblyEnd(Aperm,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1629 ierr = MatRestoreRowIJ(aA,0,PETSC_FALSE,PETSC_FALSE,&anz,&ai,&aj,&done);CHKERRQ(ierr); 1630 ierr = MatRestoreRowIJ(aB,0,PETSC_FALSE,PETSC_FALSE,&bnz,&bi,&bj,&done);CHKERRQ(ierr); 1631 ierr = MatSeqAIJRestoreArray(aA,&aa);CHKERRQ(ierr); 1632 ierr = MatSeqAIJRestoreArray(aB,&ba);CHKERRQ(ierr); 1633 ierr = PetscFree4(dnnz,onnz,tdnnz,tonnz);CHKERRQ(ierr); 1634 ierr = PetscFree3(work,rdest,cdest);CHKERRQ(ierr); 1635 ierr = PetscFree(gcdest);CHKERRQ(ierr); 1636 if (parcolp) {ierr = ISDestroy(&colp);CHKERRQ(ierr);} 1637 *B = Aperm; 1638 PetscFunctionReturn(0); 1639 } 1640 1641 PetscErrorCode MatGetGhosts_MPIAIJ(Mat mat,PetscInt *nghosts,const PetscInt *ghosts[]) 1642 { 1643 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 1644 PetscErrorCode ierr; 1645 1646 PetscFunctionBegin; 1647 ierr = MatGetSize(aij->B,NULL,nghosts);CHKERRQ(ierr); 1648 if (ghosts) *ghosts = aij->garray; 1649 PetscFunctionReturn(0); 1650 } 1651 1652 PetscErrorCode MatGetInfo_MPIAIJ(Mat matin,MatInfoType flag,MatInfo *info) 1653 { 1654 Mat_MPIAIJ *mat = (Mat_MPIAIJ*)matin->data; 1655 Mat A = mat->A,B = mat->B; 1656 PetscErrorCode ierr; 1657 PetscReal isend[5],irecv[5]; 1658 1659 PetscFunctionBegin; 1660 info->block_size = 1.0; 1661 ierr = MatGetInfo(A,MAT_LOCAL,info);CHKERRQ(ierr); 1662 1663 isend[0] = info->nz_used; isend[1] = info->nz_allocated; isend[2] = info->nz_unneeded; 1664 isend[3] = info->memory; isend[4] = info->mallocs; 1665 1666 ierr = MatGetInfo(B,MAT_LOCAL,info);CHKERRQ(ierr); 1667 1668 isend[0] += info->nz_used; isend[1] += info->nz_allocated; isend[2] += info->nz_unneeded; 1669 isend[3] += info->memory; isend[4] += info->mallocs; 1670 if (flag == MAT_LOCAL) { 1671 info->nz_used = isend[0]; 1672 info->nz_allocated = isend[1]; 1673 info->nz_unneeded = isend[2]; 1674 info->memory = isend[3]; 1675 info->mallocs = isend[4]; 1676 } else if (flag == MAT_GLOBAL_MAX) { 1677 ierr = MPIU_Allreduce(isend,irecv,5,MPIU_REAL,MPIU_MAX,PetscObjectComm((PetscObject)matin));CHKERRQ(ierr); 1678 1679 info->nz_used = irecv[0]; 1680 info->nz_allocated = irecv[1]; 1681 info->nz_unneeded = irecv[2]; 1682 info->memory = irecv[3]; 1683 info->mallocs = irecv[4]; 1684 } else if (flag == MAT_GLOBAL_SUM) { 1685 ierr = MPIU_Allreduce(isend,irecv,5,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)matin));CHKERRQ(ierr); 1686 1687 info->nz_used = irecv[0]; 1688 info->nz_allocated = irecv[1]; 1689 info->nz_unneeded = irecv[2]; 1690 info->memory = irecv[3]; 1691 info->mallocs = irecv[4]; 1692 } 1693 info->fill_ratio_given = 0; /* no parallel LU/ILU/Cholesky */ 1694 info->fill_ratio_needed = 0; 1695 info->factor_mallocs = 0; 1696 PetscFunctionReturn(0); 1697 } 1698 1699 PetscErrorCode MatSetOption_MPIAIJ(Mat A,MatOption op,PetscBool flg) 1700 { 1701 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 1702 PetscErrorCode ierr; 1703 1704 PetscFunctionBegin; 1705 switch (op) { 1706 case MAT_NEW_NONZERO_LOCATIONS: 1707 case MAT_NEW_NONZERO_ALLOCATION_ERR: 1708 case MAT_UNUSED_NONZERO_LOCATION_ERR: 1709 case MAT_KEEP_NONZERO_PATTERN: 1710 case MAT_NEW_NONZERO_LOCATION_ERR: 1711 case MAT_USE_INODES: 1712 case MAT_IGNORE_ZERO_ENTRIES: 1713 MatCheckPreallocated(A,1); 1714 ierr = MatSetOption(a->A,op,flg);CHKERRQ(ierr); 1715 ierr = MatSetOption(a->B,op,flg);CHKERRQ(ierr); 1716 break; 1717 case MAT_ROW_ORIENTED: 1718 MatCheckPreallocated(A,1); 1719 a->roworiented = flg; 1720 1721 ierr = MatSetOption(a->A,op,flg);CHKERRQ(ierr); 1722 ierr = MatSetOption(a->B,op,flg);CHKERRQ(ierr); 1723 break; 1724 case MAT_NEW_DIAGONALS: 1725 ierr = PetscInfo1(A,"Option %s ignored\n",MatOptions[op]);CHKERRQ(ierr); 1726 break; 1727 case MAT_IGNORE_OFF_PROC_ENTRIES: 1728 a->donotstash = flg; 1729 break; 1730 case MAT_SPD: 1731 A->spd_set = PETSC_TRUE; 1732 A->spd = flg; 1733 if (flg) { 1734 A->symmetric = PETSC_TRUE; 1735 A->structurally_symmetric = PETSC_TRUE; 1736 A->symmetric_set = PETSC_TRUE; 1737 A->structurally_symmetric_set = PETSC_TRUE; 1738 } 1739 break; 1740 case MAT_SYMMETRIC: 1741 MatCheckPreallocated(A,1); 1742 ierr = MatSetOption(a->A,op,flg);CHKERRQ(ierr); 1743 break; 1744 case MAT_STRUCTURALLY_SYMMETRIC: 1745 MatCheckPreallocated(A,1); 1746 ierr = MatSetOption(a->A,op,flg);CHKERRQ(ierr); 1747 break; 1748 case MAT_HERMITIAN: 1749 MatCheckPreallocated(A,1); 1750 ierr = MatSetOption(a->A,op,flg);CHKERRQ(ierr); 1751 break; 1752 case MAT_SYMMETRY_ETERNAL: 1753 MatCheckPreallocated(A,1); 1754 ierr = MatSetOption(a->A,op,flg);CHKERRQ(ierr); 1755 break; 1756 case MAT_SUBMAT_SINGLEIS: 1757 A->submat_singleis = flg; 1758 break; 1759 case MAT_STRUCTURE_ONLY: 1760 /* The option is handled directly by MatSetOption() */ 1761 break; 1762 default: 1763 SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"unknown option %d",op); 1764 } 1765 PetscFunctionReturn(0); 1766 } 1767 1768 PetscErrorCode MatGetRow_MPIAIJ(Mat matin,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v) 1769 { 1770 Mat_MPIAIJ *mat = (Mat_MPIAIJ*)matin->data; 1771 PetscScalar *vworkA,*vworkB,**pvA,**pvB,*v_p; 1772 PetscErrorCode ierr; 1773 PetscInt i,*cworkA,*cworkB,**pcA,**pcB,cstart = matin->cmap->rstart; 1774 PetscInt nztot,nzA,nzB,lrow,rstart = matin->rmap->rstart,rend = matin->rmap->rend; 1775 PetscInt *cmap,*idx_p; 1776 1777 PetscFunctionBegin; 1778 if (mat->getrowactive) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Already active"); 1779 mat->getrowactive = PETSC_TRUE; 1780 1781 if (!mat->rowvalues && (idx || v)) { 1782 /* 1783 allocate enough space to hold information from the longest row. 1784 */ 1785 Mat_SeqAIJ *Aa = (Mat_SeqAIJ*)mat->A->data,*Ba = (Mat_SeqAIJ*)mat->B->data; 1786 PetscInt max = 1,tmp; 1787 for (i=0; i<matin->rmap->n; i++) { 1788 tmp = Aa->i[i+1] - Aa->i[i] + Ba->i[i+1] - Ba->i[i]; 1789 if (max < tmp) max = tmp; 1790 } 1791 ierr = PetscMalloc2(max,&mat->rowvalues,max,&mat->rowindices);CHKERRQ(ierr); 1792 } 1793 1794 if (row < rstart || row >= rend) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Only local rows"); 1795 lrow = row - rstart; 1796 1797 pvA = &vworkA; pcA = &cworkA; pvB = &vworkB; pcB = &cworkB; 1798 if (!v) {pvA = 0; pvB = 0;} 1799 if (!idx) {pcA = 0; if (!v) pcB = 0;} 1800 ierr = (*mat->A->ops->getrow)(mat->A,lrow,&nzA,pcA,pvA);CHKERRQ(ierr); 1801 ierr = (*mat->B->ops->getrow)(mat->B,lrow,&nzB,pcB,pvB);CHKERRQ(ierr); 1802 nztot = nzA + nzB; 1803 1804 cmap = mat->garray; 1805 if (v || idx) { 1806 if (nztot) { 1807 /* Sort by increasing column numbers, assuming A and B already sorted */ 1808 PetscInt imark = -1; 1809 if (v) { 1810 *v = v_p = mat->rowvalues; 1811 for (i=0; i<nzB; i++) { 1812 if (cmap[cworkB[i]] < cstart) v_p[i] = vworkB[i]; 1813 else break; 1814 } 1815 imark = i; 1816 for (i=0; i<nzA; i++) v_p[imark+i] = vworkA[i]; 1817 for (i=imark; i<nzB; i++) v_p[nzA+i] = vworkB[i]; 1818 } 1819 if (idx) { 1820 *idx = idx_p = mat->rowindices; 1821 if (imark > -1) { 1822 for (i=0; i<imark; i++) { 1823 idx_p[i] = cmap[cworkB[i]]; 1824 } 1825 } else { 1826 for (i=0; i<nzB; i++) { 1827 if (cmap[cworkB[i]] < cstart) idx_p[i] = cmap[cworkB[i]]; 1828 else break; 1829 } 1830 imark = i; 1831 } 1832 for (i=0; i<nzA; i++) idx_p[imark+i] = cstart + cworkA[i]; 1833 for (i=imark; i<nzB; i++) idx_p[nzA+i] = cmap[cworkB[i]]; 1834 } 1835 } else { 1836 if (idx) *idx = 0; 1837 if (v) *v = 0; 1838 } 1839 } 1840 *nz = nztot; 1841 ierr = (*mat->A->ops->restorerow)(mat->A,lrow,&nzA,pcA,pvA);CHKERRQ(ierr); 1842 ierr = (*mat->B->ops->restorerow)(mat->B,lrow,&nzB,pcB,pvB);CHKERRQ(ierr); 1843 PetscFunctionReturn(0); 1844 } 1845 1846 PetscErrorCode MatRestoreRow_MPIAIJ(Mat mat,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v) 1847 { 1848 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 1849 1850 PetscFunctionBegin; 1851 if (!aij->getrowactive) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"MatGetRow() must be called first"); 1852 aij->getrowactive = PETSC_FALSE; 1853 PetscFunctionReturn(0); 1854 } 1855 1856 PetscErrorCode MatNorm_MPIAIJ(Mat mat,NormType type,PetscReal *norm) 1857 { 1858 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 1859 Mat_SeqAIJ *amat = (Mat_SeqAIJ*)aij->A->data,*bmat = (Mat_SeqAIJ*)aij->B->data; 1860 PetscErrorCode ierr; 1861 PetscInt i,j,cstart = mat->cmap->rstart; 1862 PetscReal sum = 0.0; 1863 MatScalar *v; 1864 1865 PetscFunctionBegin; 1866 if (aij->size == 1) { 1867 ierr = MatNorm(aij->A,type,norm);CHKERRQ(ierr); 1868 } else { 1869 if (type == NORM_FROBENIUS) { 1870 v = amat->a; 1871 for (i=0; i<amat->nz; i++) { 1872 sum += PetscRealPart(PetscConj(*v)*(*v)); v++; 1873 } 1874 v = bmat->a; 1875 for (i=0; i<bmat->nz; i++) { 1876 sum += PetscRealPart(PetscConj(*v)*(*v)); v++; 1877 } 1878 ierr = MPIU_Allreduce(&sum,norm,1,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1879 *norm = PetscSqrtReal(*norm); 1880 ierr = PetscLogFlops(2*amat->nz+2*bmat->nz);CHKERRQ(ierr); 1881 } else if (type == NORM_1) { /* max column norm */ 1882 PetscReal *tmp,*tmp2; 1883 PetscInt *jj,*garray = aij->garray; 1884 ierr = PetscCalloc1(mat->cmap->N+1,&tmp);CHKERRQ(ierr); 1885 ierr = PetscMalloc1(mat->cmap->N+1,&tmp2);CHKERRQ(ierr); 1886 *norm = 0.0; 1887 v = amat->a; jj = amat->j; 1888 for (j=0; j<amat->nz; j++) { 1889 tmp[cstart + *jj++] += PetscAbsScalar(*v); v++; 1890 } 1891 v = bmat->a; jj = bmat->j; 1892 for (j=0; j<bmat->nz; j++) { 1893 tmp[garray[*jj++]] += PetscAbsScalar(*v); v++; 1894 } 1895 ierr = MPIU_Allreduce(tmp,tmp2,mat->cmap->N,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1896 for (j=0; j<mat->cmap->N; j++) { 1897 if (tmp2[j] > *norm) *norm = tmp2[j]; 1898 } 1899 ierr = PetscFree(tmp);CHKERRQ(ierr); 1900 ierr = PetscFree(tmp2);CHKERRQ(ierr); 1901 ierr = PetscLogFlops(PetscMax(amat->nz+bmat->nz-1,0));CHKERRQ(ierr); 1902 } else if (type == NORM_INFINITY) { /* max row norm */ 1903 PetscReal ntemp = 0.0; 1904 for (j=0; j<aij->A->rmap->n; j++) { 1905 v = amat->a + amat->i[j]; 1906 sum = 0.0; 1907 for (i=0; i<amat->i[j+1]-amat->i[j]; i++) { 1908 sum += PetscAbsScalar(*v); v++; 1909 } 1910 v = bmat->a + bmat->i[j]; 1911 for (i=0; i<bmat->i[j+1]-bmat->i[j]; i++) { 1912 sum += PetscAbsScalar(*v); v++; 1913 } 1914 if (sum > ntemp) ntemp = sum; 1915 } 1916 ierr = MPIU_Allreduce(&ntemp,norm,1,MPIU_REAL,MPIU_MAX,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1917 ierr = PetscLogFlops(PetscMax(amat->nz+bmat->nz-1,0));CHKERRQ(ierr); 1918 } else SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"No support for two norm"); 1919 } 1920 PetscFunctionReturn(0); 1921 } 1922 1923 PetscErrorCode MatTranspose_MPIAIJ(Mat A,MatReuse reuse,Mat *matout) 1924 { 1925 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 1926 Mat_SeqAIJ *Aloc=(Mat_SeqAIJ*)a->A->data,*Bloc=(Mat_SeqAIJ*)a->B->data; 1927 PetscErrorCode ierr; 1928 PetscInt M = A->rmap->N,N = A->cmap->N,ma,na,mb,nb,*ai,*aj,*bi,*bj,row,*cols,*cols_tmp,i; 1929 PetscInt cstart = A->cmap->rstart,ncol; 1930 Mat B; 1931 MatScalar *array; 1932 1933 PetscFunctionBegin; 1934 if (reuse == MAT_INPLACE_MATRIX && M != N) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_SIZ,"Square matrix only for in-place"); 1935 1936 ma = A->rmap->n; na = A->cmap->n; mb = a->B->rmap->n; nb = a->B->cmap->n; 1937 ai = Aloc->i; aj = Aloc->j; 1938 bi = Bloc->i; bj = Bloc->j; 1939 if (reuse == MAT_INITIAL_MATRIX || *matout == A) { 1940 PetscInt *d_nnz,*g_nnz,*o_nnz; 1941 PetscSFNode *oloc; 1942 PETSC_UNUSED PetscSF sf; 1943 1944 ierr = PetscMalloc4(na,&d_nnz,na,&o_nnz,nb,&g_nnz,nb,&oloc);CHKERRQ(ierr); 1945 /* compute d_nnz for preallocation */ 1946 ierr = PetscMemzero(d_nnz,na*sizeof(PetscInt));CHKERRQ(ierr); 1947 for (i=0; i<ai[ma]; i++) { 1948 d_nnz[aj[i]]++; 1949 aj[i] += cstart; /* global col index to be used by MatSetValues() */ 1950 } 1951 /* compute local off-diagonal contributions */ 1952 ierr = PetscMemzero(g_nnz,nb*sizeof(PetscInt));CHKERRQ(ierr); 1953 for (i=0; i<bi[ma]; i++) g_nnz[bj[i]]++; 1954 /* map those to global */ 1955 ierr = PetscSFCreate(PetscObjectComm((PetscObject)A),&sf);CHKERRQ(ierr); 1956 ierr = PetscSFSetGraphLayout(sf,A->cmap,nb,NULL,PETSC_USE_POINTER,a->garray);CHKERRQ(ierr); 1957 ierr = PetscSFSetFromOptions(sf);CHKERRQ(ierr); 1958 ierr = PetscMemzero(o_nnz,na*sizeof(PetscInt));CHKERRQ(ierr); 1959 ierr = PetscSFReduceBegin(sf,MPIU_INT,g_nnz,o_nnz,MPIU_SUM);CHKERRQ(ierr); 1960 ierr = PetscSFReduceEnd(sf,MPIU_INT,g_nnz,o_nnz,MPIU_SUM);CHKERRQ(ierr); 1961 ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 1962 1963 ierr = MatCreate(PetscObjectComm((PetscObject)A),&B);CHKERRQ(ierr); 1964 ierr = MatSetSizes(B,A->cmap->n,A->rmap->n,N,M);CHKERRQ(ierr); 1965 ierr = MatSetBlockSizes(B,PetscAbs(A->cmap->bs),PetscAbs(A->rmap->bs));CHKERRQ(ierr); 1966 ierr = MatSetType(B,((PetscObject)A)->type_name);CHKERRQ(ierr); 1967 ierr = MatMPIAIJSetPreallocation(B,0,d_nnz,0,o_nnz);CHKERRQ(ierr); 1968 ierr = PetscFree4(d_nnz,o_nnz,g_nnz,oloc);CHKERRQ(ierr); 1969 } else { 1970 B = *matout; 1971 ierr = MatSetOption(B,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_TRUE);CHKERRQ(ierr); 1972 for (i=0; i<ai[ma]; i++) aj[i] += cstart; /* global col index to be used by MatSetValues() */ 1973 } 1974 1975 /* copy over the A part */ 1976 array = Aloc->a; 1977 row = A->rmap->rstart; 1978 for (i=0; i<ma; i++) { 1979 ncol = ai[i+1]-ai[i]; 1980 ierr = MatSetValues(B,ncol,aj,1,&row,array,INSERT_VALUES);CHKERRQ(ierr); 1981 row++; 1982 array += ncol; aj += ncol; 1983 } 1984 aj = Aloc->j; 1985 for (i=0; i<ai[ma]; i++) aj[i] -= cstart; /* resume local col index */ 1986 1987 /* copy over the B part */ 1988 ierr = PetscCalloc1(bi[mb],&cols);CHKERRQ(ierr); 1989 array = Bloc->a; 1990 row = A->rmap->rstart; 1991 for (i=0; i<bi[mb]; i++) cols[i] = a->garray[bj[i]]; 1992 cols_tmp = cols; 1993 for (i=0; i<mb; i++) { 1994 ncol = bi[i+1]-bi[i]; 1995 ierr = MatSetValues(B,ncol,cols_tmp,1,&row,array,INSERT_VALUES);CHKERRQ(ierr); 1996 row++; 1997 array += ncol; cols_tmp += ncol; 1998 } 1999 ierr = PetscFree(cols);CHKERRQ(ierr); 2000 2001 ierr = MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2002 ierr = MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2003 if (reuse == MAT_INITIAL_MATRIX || reuse == MAT_REUSE_MATRIX) { 2004 *matout = B; 2005 } else { 2006 ierr = MatHeaderMerge(A,&B);CHKERRQ(ierr); 2007 } 2008 PetscFunctionReturn(0); 2009 } 2010 2011 PetscErrorCode MatDiagonalScale_MPIAIJ(Mat mat,Vec ll,Vec rr) 2012 { 2013 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 2014 Mat a = aij->A,b = aij->B; 2015 PetscErrorCode ierr; 2016 PetscInt s1,s2,s3; 2017 2018 PetscFunctionBegin; 2019 ierr = MatGetLocalSize(mat,&s2,&s3);CHKERRQ(ierr); 2020 if (rr) { 2021 ierr = VecGetLocalSize(rr,&s1);CHKERRQ(ierr); 2022 if (s1!=s3) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"right vector non-conforming local size"); 2023 /* Overlap communication with computation. */ 2024 ierr = VecScatterBegin(aij->Mvctx,rr,aij->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2025 } 2026 if (ll) { 2027 ierr = VecGetLocalSize(ll,&s1);CHKERRQ(ierr); 2028 if (s1!=s2) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"left vector non-conforming local size"); 2029 ierr = (*b->ops->diagonalscale)(b,ll,0);CHKERRQ(ierr); 2030 } 2031 /* scale the diagonal block */ 2032 ierr = (*a->ops->diagonalscale)(a,ll,rr);CHKERRQ(ierr); 2033 2034 if (rr) { 2035 /* Do a scatter end and then right scale the off-diagonal block */ 2036 ierr = VecScatterEnd(aij->Mvctx,rr,aij->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2037 ierr = (*b->ops->diagonalscale)(b,0,aij->lvec);CHKERRQ(ierr); 2038 } 2039 PetscFunctionReturn(0); 2040 } 2041 2042 PetscErrorCode MatSetUnfactored_MPIAIJ(Mat A) 2043 { 2044 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 2045 PetscErrorCode ierr; 2046 2047 PetscFunctionBegin; 2048 ierr = MatSetUnfactored(a->A);CHKERRQ(ierr); 2049 PetscFunctionReturn(0); 2050 } 2051 2052 PetscErrorCode MatEqual_MPIAIJ(Mat A,Mat B,PetscBool *flag) 2053 { 2054 Mat_MPIAIJ *matB = (Mat_MPIAIJ*)B->data,*matA = (Mat_MPIAIJ*)A->data; 2055 Mat a,b,c,d; 2056 PetscBool flg; 2057 PetscErrorCode ierr; 2058 2059 PetscFunctionBegin; 2060 a = matA->A; b = matA->B; 2061 c = matB->A; d = matB->B; 2062 2063 ierr = MatEqual(a,c,&flg);CHKERRQ(ierr); 2064 if (flg) { 2065 ierr = MatEqual(b,d,&flg);CHKERRQ(ierr); 2066 } 2067 ierr = MPIU_Allreduce(&flg,flag,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)A));CHKERRQ(ierr); 2068 PetscFunctionReturn(0); 2069 } 2070 2071 PetscErrorCode MatCopy_MPIAIJ(Mat A,Mat B,MatStructure str) 2072 { 2073 PetscErrorCode ierr; 2074 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 2075 Mat_MPIAIJ *b = (Mat_MPIAIJ*)B->data; 2076 2077 PetscFunctionBegin; 2078 /* If the two matrices don't have the same copy implementation, they aren't compatible for fast copy. */ 2079 if ((str != SAME_NONZERO_PATTERN) || (A->ops->copy != B->ops->copy)) { 2080 /* because of the column compression in the off-processor part of the matrix a->B, 2081 the number of columns in a->B and b->B may be different, hence we cannot call 2082 the MatCopy() directly on the two parts. If need be, we can provide a more 2083 efficient copy than the MatCopy_Basic() by first uncompressing the a->B matrices 2084 then copying the submatrices */ 2085 ierr = MatCopy_Basic(A,B,str);CHKERRQ(ierr); 2086 } else { 2087 ierr = MatCopy(a->A,b->A,str);CHKERRQ(ierr); 2088 ierr = MatCopy(a->B,b->B,str);CHKERRQ(ierr); 2089 } 2090 ierr = PetscObjectStateIncrease((PetscObject)B);CHKERRQ(ierr); 2091 PetscFunctionReturn(0); 2092 } 2093 2094 PetscErrorCode MatSetUp_MPIAIJ(Mat A) 2095 { 2096 PetscErrorCode ierr; 2097 2098 PetscFunctionBegin; 2099 ierr = MatMPIAIJSetPreallocation(A,PETSC_DEFAULT,0,PETSC_DEFAULT,0);CHKERRQ(ierr); 2100 PetscFunctionReturn(0); 2101 } 2102 2103 /* 2104 Computes the number of nonzeros per row needed for preallocation when X and Y 2105 have different nonzero structure. 2106 */ 2107 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) 2108 { 2109 PetscInt i,j,k,nzx,nzy; 2110 2111 PetscFunctionBegin; 2112 /* Set the number of nonzeros in the new matrix */ 2113 for (i=0; i<m; i++) { 2114 const PetscInt *xjj = xj+xi[i],*yjj = yj+yi[i]; 2115 nzx = xi[i+1] - xi[i]; 2116 nzy = yi[i+1] - yi[i]; 2117 nnz[i] = 0; 2118 for (j=0,k=0; j<nzx; j++) { /* Point in X */ 2119 for (; k<nzy && yltog[yjj[k]]<xltog[xjj[j]]; k++) nnz[i]++; /* Catch up to X */ 2120 if (k<nzy && yltog[yjj[k]]==xltog[xjj[j]]) k++; /* Skip duplicate */ 2121 nnz[i]++; 2122 } 2123 for (; k<nzy; k++) nnz[i]++; 2124 } 2125 PetscFunctionReturn(0); 2126 } 2127 2128 /* This is the same as MatAXPYGetPreallocation_SeqAIJ, except that the local-to-global map is provided */ 2129 static PetscErrorCode MatAXPYGetPreallocation_MPIAIJ(Mat Y,const PetscInt *yltog,Mat X,const PetscInt *xltog,PetscInt *nnz) 2130 { 2131 PetscErrorCode ierr; 2132 PetscInt m = Y->rmap->N; 2133 Mat_SeqAIJ *x = (Mat_SeqAIJ*)X->data; 2134 Mat_SeqAIJ *y = (Mat_SeqAIJ*)Y->data; 2135 2136 PetscFunctionBegin; 2137 ierr = MatAXPYGetPreallocation_MPIX_private(m,x->i,x->j,xltog,y->i,y->j,yltog,nnz);CHKERRQ(ierr); 2138 PetscFunctionReturn(0); 2139 } 2140 2141 PetscErrorCode MatAXPY_MPIAIJ(Mat Y,PetscScalar a,Mat X,MatStructure str) 2142 { 2143 PetscErrorCode ierr; 2144 Mat_MPIAIJ *xx = (Mat_MPIAIJ*)X->data,*yy = (Mat_MPIAIJ*)Y->data; 2145 PetscBLASInt bnz,one=1; 2146 Mat_SeqAIJ *x,*y; 2147 2148 PetscFunctionBegin; 2149 if (str == SAME_NONZERO_PATTERN) { 2150 PetscScalar alpha = a; 2151 x = (Mat_SeqAIJ*)xx->A->data; 2152 ierr = PetscBLASIntCast(x->nz,&bnz);CHKERRQ(ierr); 2153 y = (Mat_SeqAIJ*)yy->A->data; 2154 PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&bnz,&alpha,x->a,&one,y->a,&one)); 2155 x = (Mat_SeqAIJ*)xx->B->data; 2156 y = (Mat_SeqAIJ*)yy->B->data; 2157 ierr = PetscBLASIntCast(x->nz,&bnz);CHKERRQ(ierr); 2158 PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&bnz,&alpha,x->a,&one,y->a,&one)); 2159 ierr = PetscObjectStateIncrease((PetscObject)Y);CHKERRQ(ierr); 2160 } else if (str == SUBSET_NONZERO_PATTERN) { /* nonzeros of X is a subset of Y's */ 2161 ierr = MatAXPY_Basic(Y,a,X,str);CHKERRQ(ierr); 2162 } else { 2163 Mat B; 2164 PetscInt *nnz_d,*nnz_o; 2165 ierr = PetscMalloc1(yy->A->rmap->N,&nnz_d);CHKERRQ(ierr); 2166 ierr = PetscMalloc1(yy->B->rmap->N,&nnz_o);CHKERRQ(ierr); 2167 ierr = MatCreate(PetscObjectComm((PetscObject)Y),&B);CHKERRQ(ierr); 2168 ierr = PetscObjectSetName((PetscObject)B,((PetscObject)Y)->name);CHKERRQ(ierr); 2169 ierr = MatSetSizes(B,Y->rmap->n,Y->cmap->n,Y->rmap->N,Y->cmap->N);CHKERRQ(ierr); 2170 ierr = MatSetBlockSizesFromMats(B,Y,Y);CHKERRQ(ierr); 2171 ierr = MatSetType(B,MATMPIAIJ);CHKERRQ(ierr); 2172 ierr = MatAXPYGetPreallocation_SeqAIJ(yy->A,xx->A,nnz_d);CHKERRQ(ierr); 2173 ierr = MatAXPYGetPreallocation_MPIAIJ(yy->B,yy->garray,xx->B,xx->garray,nnz_o);CHKERRQ(ierr); 2174 ierr = MatMPIAIJSetPreallocation(B,0,nnz_d,0,nnz_o);CHKERRQ(ierr); 2175 ierr = MatAXPY_BasicWithPreallocation(B,Y,a,X,str);CHKERRQ(ierr); 2176 ierr = MatHeaderReplace(Y,&B);CHKERRQ(ierr); 2177 ierr = PetscFree(nnz_d);CHKERRQ(ierr); 2178 ierr = PetscFree(nnz_o);CHKERRQ(ierr); 2179 } 2180 PetscFunctionReturn(0); 2181 } 2182 2183 extern PetscErrorCode MatConjugate_SeqAIJ(Mat); 2184 2185 PetscErrorCode MatConjugate_MPIAIJ(Mat mat) 2186 { 2187 #if defined(PETSC_USE_COMPLEX) 2188 PetscErrorCode ierr; 2189 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 2190 2191 PetscFunctionBegin; 2192 ierr = MatConjugate_SeqAIJ(aij->A);CHKERRQ(ierr); 2193 ierr = MatConjugate_SeqAIJ(aij->B);CHKERRQ(ierr); 2194 #else 2195 PetscFunctionBegin; 2196 #endif 2197 PetscFunctionReturn(0); 2198 } 2199 2200 PetscErrorCode MatRealPart_MPIAIJ(Mat A) 2201 { 2202 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 2203 PetscErrorCode ierr; 2204 2205 PetscFunctionBegin; 2206 ierr = MatRealPart(a->A);CHKERRQ(ierr); 2207 ierr = MatRealPart(a->B);CHKERRQ(ierr); 2208 PetscFunctionReturn(0); 2209 } 2210 2211 PetscErrorCode MatImaginaryPart_MPIAIJ(Mat A) 2212 { 2213 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 2214 PetscErrorCode ierr; 2215 2216 PetscFunctionBegin; 2217 ierr = MatImaginaryPart(a->A);CHKERRQ(ierr); 2218 ierr = MatImaginaryPart(a->B);CHKERRQ(ierr); 2219 PetscFunctionReturn(0); 2220 } 2221 2222 PetscErrorCode MatGetRowMaxAbs_MPIAIJ(Mat A, Vec v, PetscInt idx[]) 2223 { 2224 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 2225 PetscErrorCode ierr; 2226 PetscInt i,*idxb = 0; 2227 PetscScalar *va,*vb; 2228 Vec vtmp; 2229 2230 PetscFunctionBegin; 2231 ierr = MatGetRowMaxAbs(a->A,v,idx);CHKERRQ(ierr); 2232 ierr = VecGetArray(v,&va);CHKERRQ(ierr); 2233 if (idx) { 2234 for (i=0; i<A->rmap->n; i++) { 2235 if (PetscAbsScalar(va[i])) idx[i] += A->cmap->rstart; 2236 } 2237 } 2238 2239 ierr = VecCreateSeq(PETSC_COMM_SELF,A->rmap->n,&vtmp);CHKERRQ(ierr); 2240 if (idx) { 2241 ierr = PetscMalloc1(A->rmap->n,&idxb);CHKERRQ(ierr); 2242 } 2243 ierr = MatGetRowMaxAbs(a->B,vtmp,idxb);CHKERRQ(ierr); 2244 ierr = VecGetArray(vtmp,&vb);CHKERRQ(ierr); 2245 2246 for (i=0; i<A->rmap->n; i++) { 2247 if (PetscAbsScalar(va[i]) < PetscAbsScalar(vb[i])) { 2248 va[i] = vb[i]; 2249 if (idx) idx[i] = a->garray[idxb[i]]; 2250 } 2251 } 2252 2253 ierr = VecRestoreArray(v,&va);CHKERRQ(ierr); 2254 ierr = VecRestoreArray(vtmp,&vb);CHKERRQ(ierr); 2255 ierr = PetscFree(idxb);CHKERRQ(ierr); 2256 ierr = VecDestroy(&vtmp);CHKERRQ(ierr); 2257 PetscFunctionReturn(0); 2258 } 2259 2260 PetscErrorCode MatGetRowMinAbs_MPIAIJ(Mat A, Vec v, PetscInt idx[]) 2261 { 2262 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 2263 PetscErrorCode ierr; 2264 PetscInt i,*idxb = 0; 2265 PetscScalar *va,*vb; 2266 Vec vtmp; 2267 2268 PetscFunctionBegin; 2269 ierr = MatGetRowMinAbs(a->A,v,idx);CHKERRQ(ierr); 2270 ierr = VecGetArray(v,&va);CHKERRQ(ierr); 2271 if (idx) { 2272 for (i=0; i<A->cmap->n; i++) { 2273 if (PetscAbsScalar(va[i])) idx[i] += A->cmap->rstart; 2274 } 2275 } 2276 2277 ierr = VecCreateSeq(PETSC_COMM_SELF,A->rmap->n,&vtmp);CHKERRQ(ierr); 2278 if (idx) { 2279 ierr = PetscMalloc1(A->rmap->n,&idxb);CHKERRQ(ierr); 2280 } 2281 ierr = MatGetRowMinAbs(a->B,vtmp,idxb);CHKERRQ(ierr); 2282 ierr = VecGetArray(vtmp,&vb);CHKERRQ(ierr); 2283 2284 for (i=0; i<A->rmap->n; i++) { 2285 if (PetscAbsScalar(va[i]) > PetscAbsScalar(vb[i])) { 2286 va[i] = vb[i]; 2287 if (idx) idx[i] = a->garray[idxb[i]]; 2288 } 2289 } 2290 2291 ierr = VecRestoreArray(v,&va);CHKERRQ(ierr); 2292 ierr = VecRestoreArray(vtmp,&vb);CHKERRQ(ierr); 2293 ierr = PetscFree(idxb);CHKERRQ(ierr); 2294 ierr = VecDestroy(&vtmp);CHKERRQ(ierr); 2295 PetscFunctionReturn(0); 2296 } 2297 2298 PetscErrorCode MatGetRowMin_MPIAIJ(Mat A, Vec v, PetscInt idx[]) 2299 { 2300 Mat_MPIAIJ *mat = (Mat_MPIAIJ*) A->data; 2301 PetscInt n = A->rmap->n; 2302 PetscInt cstart = A->cmap->rstart; 2303 PetscInt *cmap = mat->garray; 2304 PetscInt *diagIdx, *offdiagIdx; 2305 Vec diagV, offdiagV; 2306 PetscScalar *a, *diagA, *offdiagA; 2307 PetscInt r; 2308 PetscErrorCode ierr; 2309 2310 PetscFunctionBegin; 2311 ierr = PetscMalloc2(n,&diagIdx,n,&offdiagIdx);CHKERRQ(ierr); 2312 ierr = VecCreateSeq(PetscObjectComm((PetscObject)A), n, &diagV);CHKERRQ(ierr); 2313 ierr = VecCreateSeq(PetscObjectComm((PetscObject)A), n, &offdiagV);CHKERRQ(ierr); 2314 ierr = MatGetRowMin(mat->A, diagV, diagIdx);CHKERRQ(ierr); 2315 ierr = MatGetRowMin(mat->B, offdiagV, offdiagIdx);CHKERRQ(ierr); 2316 ierr = VecGetArray(v, &a);CHKERRQ(ierr); 2317 ierr = VecGetArray(diagV, &diagA);CHKERRQ(ierr); 2318 ierr = VecGetArray(offdiagV, &offdiagA);CHKERRQ(ierr); 2319 for (r = 0; r < n; ++r) { 2320 if (PetscAbsScalar(diagA[r]) <= PetscAbsScalar(offdiagA[r])) { 2321 a[r] = diagA[r]; 2322 idx[r] = cstart + diagIdx[r]; 2323 } else { 2324 a[r] = offdiagA[r]; 2325 idx[r] = cmap[offdiagIdx[r]]; 2326 } 2327 } 2328 ierr = VecRestoreArray(v, &a);CHKERRQ(ierr); 2329 ierr = VecRestoreArray(diagV, &diagA);CHKERRQ(ierr); 2330 ierr = VecRestoreArray(offdiagV, &offdiagA);CHKERRQ(ierr); 2331 ierr = VecDestroy(&diagV);CHKERRQ(ierr); 2332 ierr = VecDestroy(&offdiagV);CHKERRQ(ierr); 2333 ierr = PetscFree2(diagIdx, offdiagIdx);CHKERRQ(ierr); 2334 PetscFunctionReturn(0); 2335 } 2336 2337 PetscErrorCode MatGetRowMax_MPIAIJ(Mat A, Vec v, PetscInt idx[]) 2338 { 2339 Mat_MPIAIJ *mat = (Mat_MPIAIJ*) A->data; 2340 PetscInt n = A->rmap->n; 2341 PetscInt cstart = A->cmap->rstart; 2342 PetscInt *cmap = mat->garray; 2343 PetscInt *diagIdx, *offdiagIdx; 2344 Vec diagV, offdiagV; 2345 PetscScalar *a, *diagA, *offdiagA; 2346 PetscInt r; 2347 PetscErrorCode ierr; 2348 2349 PetscFunctionBegin; 2350 ierr = PetscMalloc2(n,&diagIdx,n,&offdiagIdx);CHKERRQ(ierr); 2351 ierr = VecCreateSeq(PETSC_COMM_SELF, n, &diagV);CHKERRQ(ierr); 2352 ierr = VecCreateSeq(PETSC_COMM_SELF, n, &offdiagV);CHKERRQ(ierr); 2353 ierr = MatGetRowMax(mat->A, diagV, diagIdx);CHKERRQ(ierr); 2354 ierr = MatGetRowMax(mat->B, offdiagV, offdiagIdx);CHKERRQ(ierr); 2355 ierr = VecGetArray(v, &a);CHKERRQ(ierr); 2356 ierr = VecGetArray(diagV, &diagA);CHKERRQ(ierr); 2357 ierr = VecGetArray(offdiagV, &offdiagA);CHKERRQ(ierr); 2358 for (r = 0; r < n; ++r) { 2359 if (PetscAbsScalar(diagA[r]) >= PetscAbsScalar(offdiagA[r])) { 2360 a[r] = diagA[r]; 2361 idx[r] = cstart + diagIdx[r]; 2362 } else { 2363 a[r] = offdiagA[r]; 2364 idx[r] = cmap[offdiagIdx[r]]; 2365 } 2366 } 2367 ierr = VecRestoreArray(v, &a);CHKERRQ(ierr); 2368 ierr = VecRestoreArray(diagV, &diagA);CHKERRQ(ierr); 2369 ierr = VecRestoreArray(offdiagV, &offdiagA);CHKERRQ(ierr); 2370 ierr = VecDestroy(&diagV);CHKERRQ(ierr); 2371 ierr = VecDestroy(&offdiagV);CHKERRQ(ierr); 2372 ierr = PetscFree2(diagIdx, offdiagIdx);CHKERRQ(ierr); 2373 PetscFunctionReturn(0); 2374 } 2375 2376 PetscErrorCode MatGetSeqNonzeroStructure_MPIAIJ(Mat mat,Mat *newmat) 2377 { 2378 PetscErrorCode ierr; 2379 Mat *dummy; 2380 2381 PetscFunctionBegin; 2382 ierr = MatCreateSubMatrix_MPIAIJ_All(mat,MAT_DO_NOT_GET_VALUES,MAT_INITIAL_MATRIX,&dummy);CHKERRQ(ierr); 2383 *newmat = *dummy; 2384 ierr = PetscFree(dummy);CHKERRQ(ierr); 2385 PetscFunctionReturn(0); 2386 } 2387 2388 PetscErrorCode MatInvertBlockDiagonal_MPIAIJ(Mat A,const PetscScalar **values) 2389 { 2390 Mat_MPIAIJ *a = (Mat_MPIAIJ*) A->data; 2391 PetscErrorCode ierr; 2392 2393 PetscFunctionBegin; 2394 ierr = MatInvertBlockDiagonal(a->A,values);CHKERRQ(ierr); 2395 A->factorerrortype = a->A->factorerrortype; 2396 PetscFunctionReturn(0); 2397 } 2398 2399 static PetscErrorCode MatSetRandom_MPIAIJ(Mat x,PetscRandom rctx) 2400 { 2401 PetscErrorCode ierr; 2402 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)x->data; 2403 2404 PetscFunctionBegin; 2405 ierr = MatSetRandom(aij->A,rctx);CHKERRQ(ierr); 2406 ierr = MatSetRandom(aij->B,rctx);CHKERRQ(ierr); 2407 ierr = MatAssemblyBegin(x,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2408 ierr = MatAssemblyEnd(x,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2409 PetscFunctionReturn(0); 2410 } 2411 2412 PetscErrorCode MatMPIAIJSetUseScalableIncreaseOverlap_MPIAIJ(Mat A,PetscBool sc) 2413 { 2414 PetscFunctionBegin; 2415 if (sc) A->ops->increaseoverlap = MatIncreaseOverlap_MPIAIJ_Scalable; 2416 else A->ops->increaseoverlap = MatIncreaseOverlap_MPIAIJ; 2417 PetscFunctionReturn(0); 2418 } 2419 2420 /*@ 2421 MatMPIAIJSetUseScalableIncreaseOverlap - Determine if the matrix uses a scalable algorithm to compute the overlap 2422 2423 Collective on Mat 2424 2425 Input Parameters: 2426 + A - the matrix 2427 - sc - PETSC_TRUE indicates use the scalable algorithm (default is not to use the scalable algorithm) 2428 2429 Level: advanced 2430 2431 @*/ 2432 PetscErrorCode MatMPIAIJSetUseScalableIncreaseOverlap(Mat A,PetscBool sc) 2433 { 2434 PetscErrorCode ierr; 2435 2436 PetscFunctionBegin; 2437 ierr = PetscTryMethod(A,"MatMPIAIJSetUseScalableIncreaseOverlap_C",(Mat,PetscBool),(A,sc));CHKERRQ(ierr); 2438 PetscFunctionReturn(0); 2439 } 2440 2441 PetscErrorCode MatSetFromOptions_MPIAIJ(PetscOptionItems *PetscOptionsObject,Mat A) 2442 { 2443 PetscErrorCode ierr; 2444 PetscBool sc = PETSC_FALSE,flg; 2445 2446 PetscFunctionBegin; 2447 ierr = PetscOptionsHead(PetscOptionsObject,"MPIAIJ options");CHKERRQ(ierr); 2448 ierr = PetscObjectOptionsBegin((PetscObject)A); 2449 if (A->ops->increaseoverlap == MatIncreaseOverlap_MPIAIJ_Scalable) sc = PETSC_TRUE; 2450 ierr = PetscOptionsBool("-mat_increase_overlap_scalable","Use a scalable algorithm to compute the overlap","MatIncreaseOverlap",sc,&sc,&flg);CHKERRQ(ierr); 2451 if (flg) { 2452 ierr = MatMPIAIJSetUseScalableIncreaseOverlap(A,sc);CHKERRQ(ierr); 2453 } 2454 ierr = PetscOptionsEnd();CHKERRQ(ierr); 2455 PetscFunctionReturn(0); 2456 } 2457 2458 PetscErrorCode MatShift_MPIAIJ(Mat Y,PetscScalar a) 2459 { 2460 PetscErrorCode ierr; 2461 Mat_MPIAIJ *maij = (Mat_MPIAIJ*)Y->data; 2462 Mat_SeqAIJ *aij = (Mat_SeqAIJ*)maij->A->data; 2463 2464 PetscFunctionBegin; 2465 if (!Y->preallocated) { 2466 ierr = MatMPIAIJSetPreallocation(Y,1,NULL,0,NULL);CHKERRQ(ierr); 2467 } else if (!aij->nz) { 2468 PetscInt nonew = aij->nonew; 2469 ierr = MatSeqAIJSetPreallocation(maij->A,1,NULL);CHKERRQ(ierr); 2470 aij->nonew = nonew; 2471 } 2472 ierr = MatShift_Basic(Y,a);CHKERRQ(ierr); 2473 PetscFunctionReturn(0); 2474 } 2475 2476 PetscErrorCode MatMissingDiagonal_MPIAIJ(Mat A,PetscBool *missing,PetscInt *d) 2477 { 2478 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 2479 PetscErrorCode ierr; 2480 2481 PetscFunctionBegin; 2482 if (A->rmap->n != A->cmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Only works for square matrices"); 2483 ierr = MatMissingDiagonal(a->A,missing,d);CHKERRQ(ierr); 2484 if (d) { 2485 PetscInt rstart; 2486 ierr = MatGetOwnershipRange(A,&rstart,NULL);CHKERRQ(ierr); 2487 *d += rstart; 2488 2489 } 2490 PetscFunctionReturn(0); 2491 } 2492 2493 2494 /* -------------------------------------------------------------------*/ 2495 static struct _MatOps MatOps_Values = {MatSetValues_MPIAIJ, 2496 MatGetRow_MPIAIJ, 2497 MatRestoreRow_MPIAIJ, 2498 MatMult_MPIAIJ, 2499 /* 4*/ MatMultAdd_MPIAIJ, 2500 MatMultTranspose_MPIAIJ, 2501 MatMultTransposeAdd_MPIAIJ, 2502 0, 2503 0, 2504 0, 2505 /*10*/ 0, 2506 0, 2507 0, 2508 MatSOR_MPIAIJ, 2509 MatTranspose_MPIAIJ, 2510 /*15*/ MatGetInfo_MPIAIJ, 2511 MatEqual_MPIAIJ, 2512 MatGetDiagonal_MPIAIJ, 2513 MatDiagonalScale_MPIAIJ, 2514 MatNorm_MPIAIJ, 2515 /*20*/ MatAssemblyBegin_MPIAIJ, 2516 MatAssemblyEnd_MPIAIJ, 2517 MatSetOption_MPIAIJ, 2518 MatZeroEntries_MPIAIJ, 2519 /*24*/ MatZeroRows_MPIAIJ, 2520 0, 2521 0, 2522 0, 2523 0, 2524 /*29*/ MatSetUp_MPIAIJ, 2525 0, 2526 0, 2527 MatGetDiagonalBlock_MPIAIJ, 2528 0, 2529 /*34*/ MatDuplicate_MPIAIJ, 2530 0, 2531 0, 2532 0, 2533 0, 2534 /*39*/ MatAXPY_MPIAIJ, 2535 MatCreateSubMatrices_MPIAIJ, 2536 MatIncreaseOverlap_MPIAIJ, 2537 MatGetValues_MPIAIJ, 2538 MatCopy_MPIAIJ, 2539 /*44*/ MatGetRowMax_MPIAIJ, 2540 MatScale_MPIAIJ, 2541 MatShift_MPIAIJ, 2542 MatDiagonalSet_MPIAIJ, 2543 MatZeroRowsColumns_MPIAIJ, 2544 /*49*/ MatSetRandom_MPIAIJ, 2545 0, 2546 0, 2547 0, 2548 0, 2549 /*54*/ MatFDColoringCreate_MPIXAIJ, 2550 0, 2551 MatSetUnfactored_MPIAIJ, 2552 MatPermute_MPIAIJ, 2553 0, 2554 /*59*/ MatCreateSubMatrix_MPIAIJ, 2555 MatDestroy_MPIAIJ, 2556 MatView_MPIAIJ, 2557 0, 2558 MatMatMatMult_MPIAIJ_MPIAIJ_MPIAIJ, 2559 /*64*/ MatMatMatMultSymbolic_MPIAIJ_MPIAIJ_MPIAIJ, 2560 MatMatMatMultNumeric_MPIAIJ_MPIAIJ_MPIAIJ, 2561 0, 2562 0, 2563 0, 2564 /*69*/ MatGetRowMaxAbs_MPIAIJ, 2565 MatGetRowMinAbs_MPIAIJ, 2566 0, 2567 0, 2568 0, 2569 0, 2570 /*75*/ MatFDColoringApply_AIJ, 2571 MatSetFromOptions_MPIAIJ, 2572 0, 2573 0, 2574 MatFindZeroDiagonals_MPIAIJ, 2575 /*80*/ 0, 2576 0, 2577 0, 2578 /*83*/ MatLoad_MPIAIJ, 2579 MatIsSymmetric_MPIAIJ, 2580 0, 2581 0, 2582 0, 2583 0, 2584 /*89*/ MatMatMult_MPIAIJ_MPIAIJ, 2585 MatMatMultSymbolic_MPIAIJ_MPIAIJ, 2586 MatMatMultNumeric_MPIAIJ_MPIAIJ, 2587 MatPtAP_MPIAIJ_MPIAIJ, 2588 MatPtAPSymbolic_MPIAIJ_MPIAIJ, 2589 /*94*/ MatPtAPNumeric_MPIAIJ_MPIAIJ, 2590 0, 2591 0, 2592 0, 2593 0, 2594 /*99*/ 0, 2595 0, 2596 0, 2597 MatConjugate_MPIAIJ, 2598 0, 2599 /*104*/MatSetValuesRow_MPIAIJ, 2600 MatRealPart_MPIAIJ, 2601 MatImaginaryPart_MPIAIJ, 2602 0, 2603 0, 2604 /*109*/0, 2605 0, 2606 MatGetRowMin_MPIAIJ, 2607 0, 2608 MatMissingDiagonal_MPIAIJ, 2609 /*114*/MatGetSeqNonzeroStructure_MPIAIJ, 2610 0, 2611 MatGetGhosts_MPIAIJ, 2612 0, 2613 0, 2614 /*119*/0, 2615 0, 2616 0, 2617 0, 2618 MatGetMultiProcBlock_MPIAIJ, 2619 /*124*/MatFindNonzeroRows_MPIAIJ, 2620 MatGetColumnNorms_MPIAIJ, 2621 MatInvertBlockDiagonal_MPIAIJ, 2622 0, 2623 MatCreateSubMatricesMPI_MPIAIJ, 2624 /*129*/0, 2625 MatTransposeMatMult_MPIAIJ_MPIAIJ, 2626 MatTransposeMatMultSymbolic_MPIAIJ_MPIAIJ, 2627 MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ, 2628 0, 2629 /*134*/0, 2630 0, 2631 MatRARt_MPIAIJ_MPIAIJ, 2632 0, 2633 0, 2634 /*139*/MatSetBlockSizes_MPIAIJ, 2635 0, 2636 0, 2637 MatFDColoringSetUp_MPIXAIJ, 2638 MatFindOffBlockDiagonalEntries_MPIAIJ, 2639 /*144*/MatCreateMPIMatConcatenateSeqMat_MPIAIJ 2640 }; 2641 2642 /* ----------------------------------------------------------------------------------------*/ 2643 2644 PetscErrorCode MatStoreValues_MPIAIJ(Mat mat) 2645 { 2646 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 2647 PetscErrorCode ierr; 2648 2649 PetscFunctionBegin; 2650 ierr = MatStoreValues(aij->A);CHKERRQ(ierr); 2651 ierr = MatStoreValues(aij->B);CHKERRQ(ierr); 2652 PetscFunctionReturn(0); 2653 } 2654 2655 PetscErrorCode MatRetrieveValues_MPIAIJ(Mat mat) 2656 { 2657 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 2658 PetscErrorCode ierr; 2659 2660 PetscFunctionBegin; 2661 ierr = MatRetrieveValues(aij->A);CHKERRQ(ierr); 2662 ierr = MatRetrieveValues(aij->B);CHKERRQ(ierr); 2663 PetscFunctionReturn(0); 2664 } 2665 2666 PetscErrorCode MatMPIAIJSetPreallocation_MPIAIJ(Mat B,PetscInt d_nz,const PetscInt d_nnz[],PetscInt o_nz,const PetscInt o_nnz[]) 2667 { 2668 Mat_MPIAIJ *b; 2669 PetscErrorCode ierr; 2670 2671 PetscFunctionBegin; 2672 ierr = PetscLayoutSetUp(B->rmap);CHKERRQ(ierr); 2673 ierr = PetscLayoutSetUp(B->cmap);CHKERRQ(ierr); 2674 b = (Mat_MPIAIJ*)B->data; 2675 2676 #if defined(PETSC_USE_CTABLE) 2677 ierr = PetscTableDestroy(&b->colmap);CHKERRQ(ierr); 2678 #else 2679 ierr = PetscFree(b->colmap);CHKERRQ(ierr); 2680 #endif 2681 ierr = PetscFree(b->garray);CHKERRQ(ierr); 2682 ierr = VecDestroy(&b->lvec);CHKERRQ(ierr); 2683 ierr = VecScatterDestroy(&b->Mvctx);CHKERRQ(ierr); 2684 2685 /* Because the B will have been resized we simply destroy it and create a new one each time */ 2686 ierr = MatDestroy(&b->B);CHKERRQ(ierr); 2687 ierr = MatCreate(PETSC_COMM_SELF,&b->B);CHKERRQ(ierr); 2688 ierr = MatSetSizes(b->B,B->rmap->n,B->cmap->N,B->rmap->n,B->cmap->N);CHKERRQ(ierr); 2689 ierr = MatSetBlockSizesFromMats(b->B,B,B);CHKERRQ(ierr); 2690 ierr = MatSetType(b->B,MATSEQAIJ);CHKERRQ(ierr); 2691 ierr = PetscLogObjectParent((PetscObject)B,(PetscObject)b->B);CHKERRQ(ierr); 2692 2693 if (!B->preallocated) { 2694 ierr = MatCreate(PETSC_COMM_SELF,&b->A);CHKERRQ(ierr); 2695 ierr = MatSetSizes(b->A,B->rmap->n,B->cmap->n,B->rmap->n,B->cmap->n);CHKERRQ(ierr); 2696 ierr = MatSetBlockSizesFromMats(b->A,B,B);CHKERRQ(ierr); 2697 ierr = MatSetType(b->A,MATSEQAIJ);CHKERRQ(ierr); 2698 ierr = PetscLogObjectParent((PetscObject)B,(PetscObject)b->A);CHKERRQ(ierr); 2699 } 2700 2701 ierr = MatSeqAIJSetPreallocation(b->A,d_nz,d_nnz);CHKERRQ(ierr); 2702 ierr = MatSeqAIJSetPreallocation(b->B,o_nz,o_nnz);CHKERRQ(ierr); 2703 B->preallocated = PETSC_TRUE; 2704 B->was_assembled = PETSC_FALSE; 2705 B->assembled = PETSC_FALSE;; 2706 PetscFunctionReturn(0); 2707 } 2708 2709 PetscErrorCode MatResetPreallocation_MPIAIJ(Mat B) 2710 { 2711 Mat_MPIAIJ *b; 2712 PetscErrorCode ierr; 2713 2714 PetscFunctionBegin; 2715 PetscValidHeaderSpecific(B,MAT_CLASSID,1); 2716 ierr = PetscLayoutSetUp(B->rmap);CHKERRQ(ierr); 2717 ierr = PetscLayoutSetUp(B->cmap);CHKERRQ(ierr); 2718 b = (Mat_MPIAIJ*)B->data; 2719 2720 #if defined(PETSC_USE_CTABLE) 2721 ierr = PetscTableDestroy(&b->colmap);CHKERRQ(ierr); 2722 #else 2723 ierr = PetscFree(b->colmap);CHKERRQ(ierr); 2724 #endif 2725 ierr = PetscFree(b->garray);CHKERRQ(ierr); 2726 ierr = VecDestroy(&b->lvec);CHKERRQ(ierr); 2727 ierr = VecScatterDestroy(&b->Mvctx);CHKERRQ(ierr); 2728 2729 ierr = MatResetPreallocation(b->A);CHKERRQ(ierr); 2730 ierr = MatResetPreallocation(b->B);CHKERRQ(ierr); 2731 B->preallocated = PETSC_TRUE; 2732 B->was_assembled = PETSC_FALSE; 2733 B->assembled = PETSC_FALSE; 2734 PetscFunctionReturn(0); 2735 } 2736 2737 PetscErrorCode MatDuplicate_MPIAIJ(Mat matin,MatDuplicateOption cpvalues,Mat *newmat) 2738 { 2739 Mat mat; 2740 Mat_MPIAIJ *a,*oldmat = (Mat_MPIAIJ*)matin->data; 2741 PetscErrorCode ierr; 2742 2743 PetscFunctionBegin; 2744 *newmat = 0; 2745 ierr = MatCreate(PetscObjectComm((PetscObject)matin),&mat);CHKERRQ(ierr); 2746 ierr = MatSetSizes(mat,matin->rmap->n,matin->cmap->n,matin->rmap->N,matin->cmap->N);CHKERRQ(ierr); 2747 ierr = MatSetBlockSizesFromMats(mat,matin,matin);CHKERRQ(ierr); 2748 ierr = MatSetType(mat,((PetscObject)matin)->type_name);CHKERRQ(ierr); 2749 ierr = PetscMemcpy(mat->ops,matin->ops,sizeof(struct _MatOps));CHKERRQ(ierr); 2750 a = (Mat_MPIAIJ*)mat->data; 2751 2752 mat->factortype = matin->factortype; 2753 mat->assembled = PETSC_TRUE; 2754 mat->insertmode = NOT_SET_VALUES; 2755 mat->preallocated = PETSC_TRUE; 2756 2757 a->size = oldmat->size; 2758 a->rank = oldmat->rank; 2759 a->donotstash = oldmat->donotstash; 2760 a->roworiented = oldmat->roworiented; 2761 a->rowindices = 0; 2762 a->rowvalues = 0; 2763 a->getrowactive = PETSC_FALSE; 2764 2765 ierr = PetscLayoutReference(matin->rmap,&mat->rmap);CHKERRQ(ierr); 2766 ierr = PetscLayoutReference(matin->cmap,&mat->cmap);CHKERRQ(ierr); 2767 2768 if (oldmat->colmap) { 2769 #if defined(PETSC_USE_CTABLE) 2770 ierr = PetscTableCreateCopy(oldmat->colmap,&a->colmap);CHKERRQ(ierr); 2771 #else 2772 ierr = PetscMalloc1(mat->cmap->N,&a->colmap);CHKERRQ(ierr); 2773 ierr = PetscLogObjectMemory((PetscObject)mat,(mat->cmap->N)*sizeof(PetscInt));CHKERRQ(ierr); 2774 ierr = PetscMemcpy(a->colmap,oldmat->colmap,(mat->cmap->N)*sizeof(PetscInt));CHKERRQ(ierr); 2775 #endif 2776 } else a->colmap = 0; 2777 if (oldmat->garray) { 2778 PetscInt len; 2779 len = oldmat->B->cmap->n; 2780 ierr = PetscMalloc1(len+1,&a->garray);CHKERRQ(ierr); 2781 ierr = PetscLogObjectMemory((PetscObject)mat,len*sizeof(PetscInt));CHKERRQ(ierr); 2782 if (len) { ierr = PetscMemcpy(a->garray,oldmat->garray,len*sizeof(PetscInt));CHKERRQ(ierr); } 2783 } else a->garray = 0; 2784 2785 ierr = VecDuplicate(oldmat->lvec,&a->lvec);CHKERRQ(ierr); 2786 ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)a->lvec);CHKERRQ(ierr); 2787 ierr = VecScatterCopy(oldmat->Mvctx,&a->Mvctx);CHKERRQ(ierr); 2788 ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)a->Mvctx);CHKERRQ(ierr); 2789 2790 if (oldmat->Mvctx_mpi1) { 2791 ierr = VecScatterCopy(oldmat->Mvctx_mpi1,&a->Mvctx_mpi1);CHKERRQ(ierr); 2792 ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)a->Mvctx_mpi1);CHKERRQ(ierr); 2793 } 2794 2795 ierr = MatDuplicate(oldmat->A,cpvalues,&a->A);CHKERRQ(ierr); 2796 ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)a->A);CHKERRQ(ierr); 2797 ierr = MatDuplicate(oldmat->B,cpvalues,&a->B);CHKERRQ(ierr); 2798 ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)a->B);CHKERRQ(ierr); 2799 ierr = PetscFunctionListDuplicate(((PetscObject)matin)->qlist,&((PetscObject)mat)->qlist);CHKERRQ(ierr); 2800 *newmat = mat; 2801 PetscFunctionReturn(0); 2802 } 2803 2804 PetscErrorCode MatLoad_MPIAIJ(Mat newMat, PetscViewer viewer) 2805 { 2806 PetscScalar *vals,*svals; 2807 MPI_Comm comm; 2808 PetscErrorCode ierr; 2809 PetscMPIInt rank,size,tag = ((PetscObject)viewer)->tag; 2810 PetscInt i,nz,j,rstart,rend,mmax,maxnz = 0; 2811 PetscInt header[4],*rowlengths = 0,M,N,m,*cols; 2812 PetscInt *ourlens = NULL,*procsnz = NULL,*offlens = NULL,jj,*mycols,*smycols; 2813 PetscInt cend,cstart,n,*rowners; 2814 int fd; 2815 PetscInt bs = newMat->rmap->bs; 2816 2817 PetscFunctionBegin; 2818 /* force binary viewer to load .info file if it has not yet done so */ 2819 ierr = PetscViewerSetUp(viewer);CHKERRQ(ierr); 2820 ierr = PetscObjectGetComm((PetscObject)viewer,&comm);CHKERRQ(ierr); 2821 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 2822 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 2823 ierr = PetscViewerBinaryGetDescriptor(viewer,&fd);CHKERRQ(ierr); 2824 if (!rank) { 2825 ierr = PetscBinaryRead(fd,(char*)header,4,PETSC_INT);CHKERRQ(ierr); 2826 if (header[0] != MAT_FILE_CLASSID) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED,"not matrix object"); 2827 if (header[3] < 0) SETERRQ(PetscObjectComm((PetscObject)newMat),PETSC_ERR_FILE_UNEXPECTED,"Matrix stored in special format on disk,cannot load as MATMPIAIJ"); 2828 } 2829 2830 ierr = PetscOptionsBegin(comm,NULL,"Options for loading MATMPIAIJ matrix","Mat");CHKERRQ(ierr); 2831 ierr = PetscOptionsInt("-matload_block_size","Set the blocksize used to store the matrix","MatLoad",bs,&bs,NULL);CHKERRQ(ierr); 2832 ierr = PetscOptionsEnd();CHKERRQ(ierr); 2833 if (bs < 0) bs = 1; 2834 2835 ierr = MPI_Bcast(header+1,3,MPIU_INT,0,comm);CHKERRQ(ierr); 2836 M = header[1]; N = header[2]; 2837 2838 /* If global sizes are set, check if they are consistent with that given in the file */ 2839 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); 2840 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); 2841 2842 /* determine ownership of all (block) rows */ 2843 if (M%bs) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED, "Inconsistent # of rows (%d) and block size (%d)",M,bs); 2844 if (newMat->rmap->n < 0) m = bs*((M/bs)/size + (((M/bs) % size) > rank)); /* PETSC_DECIDE */ 2845 else m = newMat->rmap->n; /* Set by user */ 2846 2847 ierr = PetscMalloc1(size+1,&rowners);CHKERRQ(ierr); 2848 ierr = MPI_Allgather(&m,1,MPIU_INT,rowners+1,1,MPIU_INT,comm);CHKERRQ(ierr); 2849 2850 /* First process needs enough room for process with most rows */ 2851 if (!rank) { 2852 mmax = rowners[1]; 2853 for (i=2; i<=size; i++) { 2854 mmax = PetscMax(mmax, rowners[i]); 2855 } 2856 } else mmax = -1; /* unused, but compilers complain */ 2857 2858 rowners[0] = 0; 2859 for (i=2; i<=size; i++) { 2860 rowners[i] += rowners[i-1]; 2861 } 2862 rstart = rowners[rank]; 2863 rend = rowners[rank+1]; 2864 2865 /* distribute row lengths to all processors */ 2866 ierr = PetscMalloc2(m,&ourlens,m,&offlens);CHKERRQ(ierr); 2867 if (!rank) { 2868 ierr = PetscBinaryRead(fd,ourlens,m,PETSC_INT);CHKERRQ(ierr); 2869 ierr = PetscMalloc1(mmax,&rowlengths);CHKERRQ(ierr); 2870 ierr = PetscCalloc1(size,&procsnz);CHKERRQ(ierr); 2871 for (j=0; j<m; j++) { 2872 procsnz[0] += ourlens[j]; 2873 } 2874 for (i=1; i<size; i++) { 2875 ierr = PetscBinaryRead(fd,rowlengths,rowners[i+1]-rowners[i],PETSC_INT);CHKERRQ(ierr); 2876 /* calculate the number of nonzeros on each processor */ 2877 for (j=0; j<rowners[i+1]-rowners[i]; j++) { 2878 procsnz[i] += rowlengths[j]; 2879 } 2880 ierr = MPIULong_Send(rowlengths,rowners[i+1]-rowners[i],MPIU_INT,i,tag,comm);CHKERRQ(ierr); 2881 } 2882 ierr = PetscFree(rowlengths);CHKERRQ(ierr); 2883 } else { 2884 ierr = MPIULong_Recv(ourlens,m,MPIU_INT,0,tag,comm);CHKERRQ(ierr); 2885 } 2886 2887 if (!rank) { 2888 /* determine max buffer needed and allocate it */ 2889 maxnz = 0; 2890 for (i=0; i<size; i++) { 2891 maxnz = PetscMax(maxnz,procsnz[i]); 2892 } 2893 ierr = PetscMalloc1(maxnz,&cols);CHKERRQ(ierr); 2894 2895 /* read in my part of the matrix column indices */ 2896 nz = procsnz[0]; 2897 ierr = PetscMalloc1(nz,&mycols);CHKERRQ(ierr); 2898 ierr = PetscBinaryRead(fd,mycols,nz,PETSC_INT);CHKERRQ(ierr); 2899 2900 /* read in every one elses and ship off */ 2901 for (i=1; i<size; i++) { 2902 nz = procsnz[i]; 2903 ierr = PetscBinaryRead(fd,cols,nz,PETSC_INT);CHKERRQ(ierr); 2904 ierr = MPIULong_Send(cols,nz,MPIU_INT,i,tag,comm);CHKERRQ(ierr); 2905 } 2906 ierr = PetscFree(cols);CHKERRQ(ierr); 2907 } else { 2908 /* determine buffer space needed for message */ 2909 nz = 0; 2910 for (i=0; i<m; i++) { 2911 nz += ourlens[i]; 2912 } 2913 ierr = PetscMalloc1(nz,&mycols);CHKERRQ(ierr); 2914 2915 /* receive message of column indices*/ 2916 ierr = MPIULong_Recv(mycols,nz,MPIU_INT,0,tag,comm);CHKERRQ(ierr); 2917 } 2918 2919 /* determine column ownership if matrix is not square */ 2920 if (N != M) { 2921 if (newMat->cmap->n < 0) n = N/size + ((N % size) > rank); 2922 else n = newMat->cmap->n; 2923 ierr = MPI_Scan(&n,&cend,1,MPIU_INT,MPI_SUM,comm);CHKERRQ(ierr); 2924 cstart = cend - n; 2925 } else { 2926 cstart = rstart; 2927 cend = rend; 2928 n = cend - cstart; 2929 } 2930 2931 /* loop over local rows, determining number of off diagonal entries */ 2932 ierr = PetscMemzero(offlens,m*sizeof(PetscInt));CHKERRQ(ierr); 2933 jj = 0; 2934 for (i=0; i<m; i++) { 2935 for (j=0; j<ourlens[i]; j++) { 2936 if (mycols[jj] < cstart || mycols[jj] >= cend) offlens[i]++; 2937 jj++; 2938 } 2939 } 2940 2941 for (i=0; i<m; i++) { 2942 ourlens[i] -= offlens[i]; 2943 } 2944 ierr = MatSetSizes(newMat,m,n,M,N);CHKERRQ(ierr); 2945 2946 if (bs > 1) {ierr = MatSetBlockSize(newMat,bs);CHKERRQ(ierr);} 2947 2948 ierr = MatMPIAIJSetPreallocation(newMat,0,ourlens,0,offlens);CHKERRQ(ierr); 2949 2950 for (i=0; i<m; i++) { 2951 ourlens[i] += offlens[i]; 2952 } 2953 2954 if (!rank) { 2955 ierr = PetscMalloc1(maxnz+1,&vals);CHKERRQ(ierr); 2956 2957 /* read in my part of the matrix numerical values */ 2958 nz = procsnz[0]; 2959 ierr = PetscBinaryRead(fd,vals,nz,PETSC_SCALAR);CHKERRQ(ierr); 2960 2961 /* insert into matrix */ 2962 jj = rstart; 2963 smycols = mycols; 2964 svals = vals; 2965 for (i=0; i<m; i++) { 2966 ierr = MatSetValues_MPIAIJ(newMat,1,&jj,ourlens[i],smycols,svals,INSERT_VALUES);CHKERRQ(ierr); 2967 smycols += ourlens[i]; 2968 svals += ourlens[i]; 2969 jj++; 2970 } 2971 2972 /* read in other processors and ship out */ 2973 for (i=1; i<size; i++) { 2974 nz = procsnz[i]; 2975 ierr = PetscBinaryRead(fd,vals,nz,PETSC_SCALAR);CHKERRQ(ierr); 2976 ierr = MPIULong_Send(vals,nz,MPIU_SCALAR,i,((PetscObject)newMat)->tag,comm);CHKERRQ(ierr); 2977 } 2978 ierr = PetscFree(procsnz);CHKERRQ(ierr); 2979 } else { 2980 /* receive numeric values */ 2981 ierr = PetscMalloc1(nz+1,&vals);CHKERRQ(ierr); 2982 2983 /* receive message of values*/ 2984 ierr = MPIULong_Recv(vals,nz,MPIU_SCALAR,0,((PetscObject)newMat)->tag,comm);CHKERRQ(ierr); 2985 2986 /* insert into matrix */ 2987 jj = rstart; 2988 smycols = mycols; 2989 svals = vals; 2990 for (i=0; i<m; i++) { 2991 ierr = MatSetValues_MPIAIJ(newMat,1,&jj,ourlens[i],smycols,svals,INSERT_VALUES);CHKERRQ(ierr); 2992 smycols += ourlens[i]; 2993 svals += ourlens[i]; 2994 jj++; 2995 } 2996 } 2997 ierr = PetscFree2(ourlens,offlens);CHKERRQ(ierr); 2998 ierr = PetscFree(vals);CHKERRQ(ierr); 2999 ierr = PetscFree(mycols);CHKERRQ(ierr); 3000 ierr = PetscFree(rowners);CHKERRQ(ierr); 3001 ierr = MatAssemblyBegin(newMat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3002 ierr = MatAssemblyEnd(newMat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3003 PetscFunctionReturn(0); 3004 } 3005 3006 /* Not scalable because of ISAllGather() unless getting all columns. */ 3007 PetscErrorCode ISGetSeqIS_Private(Mat mat,IS iscol,IS *isseq) 3008 { 3009 PetscErrorCode ierr; 3010 IS iscol_local; 3011 PetscBool isstride; 3012 PetscMPIInt lisstride=0,gisstride; 3013 3014 PetscFunctionBegin; 3015 /* check if we are grabbing all columns*/ 3016 ierr = PetscObjectTypeCompare((PetscObject)iscol,ISSTRIDE,&isstride);CHKERRQ(ierr); 3017 3018 if (isstride) { 3019 PetscInt start,len,mstart,mlen; 3020 ierr = ISStrideGetInfo(iscol,&start,NULL);CHKERRQ(ierr); 3021 ierr = ISGetLocalSize(iscol,&len);CHKERRQ(ierr); 3022 ierr = MatGetOwnershipRangeColumn(mat,&mstart,&mlen);CHKERRQ(ierr); 3023 if (mstart == start && mlen-mstart == len) lisstride = 1; 3024 } 3025 3026 ierr = MPIU_Allreduce(&lisstride,&gisstride,1,MPI_INT,MPI_MIN,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 3027 if (gisstride) { 3028 PetscInt N; 3029 ierr = MatGetSize(mat,NULL,&N);CHKERRQ(ierr); 3030 ierr = ISCreateStride(PetscObjectComm((PetscObject)mat),N,0,1,&iscol_local);CHKERRQ(ierr); 3031 ierr = ISSetIdentity(iscol_local);CHKERRQ(ierr); 3032 ierr = PetscInfo(mat,"Optimizing for obtaining all columns of the matrix; skipping ISAllGather()\n");CHKERRQ(ierr); 3033 } else { 3034 PetscInt cbs; 3035 ierr = ISGetBlockSize(iscol,&cbs);CHKERRQ(ierr); 3036 ierr = ISAllGather(iscol,&iscol_local);CHKERRQ(ierr); 3037 ierr = ISSetBlockSize(iscol_local,cbs);CHKERRQ(ierr); 3038 } 3039 3040 *isseq = iscol_local; 3041 PetscFunctionReturn(0); 3042 } 3043 3044 /* 3045 Used by MatCreateSubMatrix_MPIAIJ_SameRowColDist() to avoid ISAllGather() and global size of iscol_local 3046 (see MatCreateSubMatrix_MPIAIJ_nonscalable) 3047 3048 Input Parameters: 3049 mat - matrix 3050 isrow - parallel row index set; its local indices are a subset of local columns of mat, 3051 i.e., mat->rstart <= isrow[i] < mat->rend 3052 iscol - parallel column index set; its local indices are a subset of local columns of mat, 3053 i.e., mat->cstart <= iscol[i] < mat->cend 3054 Output Parameter: 3055 isrow_d,iscol_d - sequential row and column index sets for retrieving mat->A 3056 iscol_o - sequential column index set for retrieving mat->B 3057 garray - column map; garray[i] indicates global location of iscol_o[i] in iscol 3058 */ 3059 PetscErrorCode ISGetSeqIS_SameColDist_Private(Mat mat,IS isrow,IS iscol,IS *isrow_d,IS *iscol_d,IS *iscol_o,const PetscInt *garray[]) 3060 { 3061 PetscErrorCode ierr; 3062 Vec x,cmap; 3063 const PetscInt *is_idx; 3064 PetscScalar *xarray,*cmaparray; 3065 PetscInt ncols,isstart,*idx,m,rstart,*cmap1,count; 3066 Mat_MPIAIJ *a=(Mat_MPIAIJ*)mat->data; 3067 Mat B=a->B; 3068 Vec lvec=a->lvec,lcmap; 3069 PetscInt i,cstart,cend,Bn=B->cmap->N; 3070 MPI_Comm comm; 3071 VecScatter Mvctx=a->Mvctx; 3072 3073 PetscFunctionBegin; 3074 ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 3075 ierr = ISGetLocalSize(iscol,&ncols);CHKERRQ(ierr); 3076 3077 /* (1) iscol is a sub-column vector of mat, pad it with '-1.' to form a full vector x */ 3078 ierr = MatCreateVecs(mat,&x,NULL);CHKERRQ(ierr); 3079 ierr = VecSet(x,-1.0);CHKERRQ(ierr); 3080 ierr = VecDuplicate(x,&cmap);CHKERRQ(ierr); 3081 ierr = VecSet(cmap,-1.0);CHKERRQ(ierr); 3082 3083 /* Get start indices */ 3084 ierr = MPI_Scan(&ncols,&isstart,1,MPIU_INT,MPI_SUM,comm);CHKERRQ(ierr); 3085 isstart -= ncols; 3086 ierr = MatGetOwnershipRangeColumn(mat,&cstart,&cend);CHKERRQ(ierr); 3087 3088 ierr = ISGetIndices(iscol,&is_idx);CHKERRQ(ierr); 3089 ierr = VecGetArray(x,&xarray);CHKERRQ(ierr); 3090 ierr = VecGetArray(cmap,&cmaparray);CHKERRQ(ierr); 3091 ierr = PetscMalloc1(ncols,&idx);CHKERRQ(ierr); 3092 for (i=0; i<ncols; i++) { 3093 xarray[is_idx[i]-cstart] = (PetscScalar)is_idx[i]; 3094 cmaparray[is_idx[i]-cstart] = i + isstart; /* global index of iscol[i] */ 3095 idx[i] = is_idx[i]-cstart; /* local index of iscol[i] */ 3096 } 3097 ierr = VecRestoreArray(x,&xarray);CHKERRQ(ierr); 3098 ierr = VecRestoreArray(cmap,&cmaparray);CHKERRQ(ierr); 3099 ierr = ISRestoreIndices(iscol,&is_idx);CHKERRQ(ierr); 3100 3101 /* Get iscol_d */ 3102 ierr = ISCreateGeneral(PETSC_COMM_SELF,ncols,idx,PETSC_OWN_POINTER,iscol_d);CHKERRQ(ierr); 3103 ierr = ISGetBlockSize(iscol,&i);CHKERRQ(ierr); 3104 ierr = ISSetBlockSize(*iscol_d,i);CHKERRQ(ierr); 3105 3106 /* Get isrow_d */ 3107 ierr = ISGetLocalSize(isrow,&m);CHKERRQ(ierr); 3108 rstart = mat->rmap->rstart; 3109 ierr = PetscMalloc1(m,&idx);CHKERRQ(ierr); 3110 ierr = ISGetIndices(isrow,&is_idx);CHKERRQ(ierr); 3111 for (i=0; i<m; i++) idx[i] = is_idx[i]-rstart; 3112 ierr = ISRestoreIndices(isrow,&is_idx);CHKERRQ(ierr); 3113 3114 ierr = ISCreateGeneral(PETSC_COMM_SELF,m,idx,PETSC_OWN_POINTER,isrow_d);CHKERRQ(ierr); 3115 ierr = ISGetBlockSize(isrow,&i);CHKERRQ(ierr); 3116 ierr = ISSetBlockSize(*isrow_d,i);CHKERRQ(ierr); 3117 3118 /* (2) Scatter x and cmap using aij->Mvctx to get their off-process portions (see MatMult_MPIAIJ) */ 3119 ierr = VecScatterBegin(Mvctx,x,lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3120 ierr = VecScatterEnd(Mvctx,x,lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3121 3122 ierr = VecDuplicate(lvec,&lcmap);CHKERRQ(ierr); 3123 3124 ierr = VecScatterBegin(Mvctx,cmap,lcmap,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3125 ierr = VecScatterEnd(Mvctx,cmap,lcmap,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3126 3127 /* (3) create sequential iscol_o (a subset of iscol) and isgarray */ 3128 /* off-process column indices */ 3129 count = 0; 3130 ierr = PetscMalloc1(Bn,&idx);CHKERRQ(ierr); 3131 ierr = PetscMalloc1(Bn,&cmap1);CHKERRQ(ierr); 3132 3133 ierr = VecGetArray(lvec,&xarray);CHKERRQ(ierr); 3134 ierr = VecGetArray(lcmap,&cmaparray);CHKERRQ(ierr); 3135 for (i=0; i<Bn; i++) { 3136 if (PetscRealPart(xarray[i]) > -1.0) { 3137 idx[count] = i; /* local column index in off-diagonal part B */ 3138 cmap1[count] = (PetscInt)PetscRealPart(cmaparray[i]); /* column index in submat */ 3139 count++; 3140 } 3141 } 3142 ierr = VecRestoreArray(lvec,&xarray);CHKERRQ(ierr); 3143 ierr = VecRestoreArray(lcmap,&cmaparray);CHKERRQ(ierr); 3144 3145 ierr = ISCreateGeneral(PETSC_COMM_SELF,count,idx,PETSC_COPY_VALUES,iscol_o);CHKERRQ(ierr); 3146 /* cannot ensure iscol_o has same blocksize as iscol! */ 3147 3148 ierr = PetscFree(idx);CHKERRQ(ierr); 3149 *garray = cmap1; 3150 3151 ierr = VecDestroy(&x);CHKERRQ(ierr); 3152 ierr = VecDestroy(&cmap);CHKERRQ(ierr); 3153 ierr = VecDestroy(&lcmap);CHKERRQ(ierr); 3154 PetscFunctionReturn(0); 3155 } 3156 3157 /* isrow and iscol have same processor distribution as mat, output *submat is a submatrix of local mat */ 3158 PetscErrorCode MatCreateSubMatrix_MPIAIJ_SameRowColDist(Mat mat,IS isrow,IS iscol,MatReuse call,Mat *submat) 3159 { 3160 PetscErrorCode ierr; 3161 Mat_MPIAIJ *a = (Mat_MPIAIJ*)mat->data,*asub; 3162 Mat M = NULL; 3163 MPI_Comm comm; 3164 IS iscol_d,isrow_d,iscol_o; 3165 Mat Asub = NULL,Bsub = NULL; 3166 PetscInt n; 3167 3168 PetscFunctionBegin; 3169 ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 3170 3171 if (call == MAT_REUSE_MATRIX) { 3172 /* Retrieve isrow_d, iscol_d and iscol_o from submat */ 3173 ierr = PetscObjectQuery((PetscObject)*submat,"isrow_d",(PetscObject*)&isrow_d);CHKERRQ(ierr); 3174 if (!isrow_d) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"isrow_d passed in was not used before, cannot reuse"); 3175 3176 ierr = PetscObjectQuery((PetscObject)*submat,"iscol_d",(PetscObject*)&iscol_d);CHKERRQ(ierr); 3177 if (!iscol_d) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"iscol_d passed in was not used before, cannot reuse"); 3178 3179 ierr = PetscObjectQuery((PetscObject)*submat,"iscol_o",(PetscObject*)&iscol_o);CHKERRQ(ierr); 3180 if (!iscol_o) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"iscol_o passed in was not used before, cannot reuse"); 3181 3182 /* Update diagonal and off-diagonal portions of submat */ 3183 asub = (Mat_MPIAIJ*)(*submat)->data; 3184 ierr = MatCreateSubMatrix_SeqAIJ(a->A,isrow_d,iscol_d,PETSC_DECIDE,MAT_REUSE_MATRIX,&asub->A);CHKERRQ(ierr); 3185 ierr = ISGetLocalSize(iscol_o,&n);CHKERRQ(ierr); 3186 if (n) { 3187 ierr = MatCreateSubMatrix_SeqAIJ(a->B,isrow_d,iscol_o,PETSC_DECIDE,MAT_REUSE_MATRIX,&asub->B);CHKERRQ(ierr); 3188 } 3189 ierr = MatAssemblyBegin(*submat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3190 ierr = MatAssemblyEnd(*submat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3191 3192 } else { /* call == MAT_INITIAL_MATRIX) */ 3193 const PetscInt *garray; 3194 PetscInt BsubN; 3195 3196 /* Create isrow_d, iscol_d, iscol_o and isgarray (replace isgarray with array?) */ 3197 ierr = ISGetSeqIS_SameColDist_Private(mat,isrow,iscol,&isrow_d,&iscol_d,&iscol_o,&garray);CHKERRQ(ierr); 3198 3199 /* Create local submatrices Asub and Bsub */ 3200 ierr = MatCreateSubMatrix_SeqAIJ(a->A,isrow_d,iscol_d,PETSC_DECIDE,MAT_INITIAL_MATRIX,&Asub);CHKERRQ(ierr); 3201 ierr = MatCreateSubMatrix_SeqAIJ(a->B,isrow_d,iscol_o,PETSC_DECIDE,MAT_INITIAL_MATRIX,&Bsub);CHKERRQ(ierr); 3202 3203 /* Create submatrix M */ 3204 ierr = MatCreateMPIAIJWithSeqAIJ(comm,Asub,Bsub,garray,&M);CHKERRQ(ierr); 3205 3206 /* If Bsub has empty columns, compress iscol_o such that it will retrieve condensed Bsub from a->B during reuse */ 3207 asub = (Mat_MPIAIJ*)M->data; 3208 3209 ierr = ISGetLocalSize(iscol_o,&BsubN);CHKERRQ(ierr); 3210 n = asub->B->cmap->N; 3211 if (BsubN > n) { 3212 /* This case can be tested using ~petsc/src/tao/bound/examples/tutorials/runplate2_3 */ 3213 const PetscInt *idx; 3214 PetscInt i,j,*idx_new,*subgarray = asub->garray; 3215 ierr = PetscInfo2(M,"submatrix Bn %D != BsubN %D, update iscol_o\n",n,BsubN);CHKERRQ(ierr); 3216 3217 ierr = PetscMalloc1(n,&idx_new);CHKERRQ(ierr); 3218 j = 0; 3219 ierr = ISGetIndices(iscol_o,&idx);CHKERRQ(ierr); 3220 for (i=0; i<n; i++) { 3221 if (j >= BsubN) break; 3222 while (subgarray[i] > garray[j]) j++; 3223 3224 if (subgarray[i] == garray[j]) { 3225 idx_new[i] = idx[j++]; 3226 } else SETERRQ4(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"subgarray[%D]=%D cannot < garray[%D]=%D",i,subgarray[i],j,garray[j]); 3227 } 3228 ierr = ISRestoreIndices(iscol_o,&idx);CHKERRQ(ierr); 3229 3230 ierr = ISDestroy(&iscol_o);CHKERRQ(ierr); 3231 ierr = ISCreateGeneral(PETSC_COMM_SELF,n,idx_new,PETSC_OWN_POINTER,&iscol_o);CHKERRQ(ierr); 3232 3233 } else if (BsubN < n) { 3234 SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Columns of Bsub cannot be smaller than B's",BsubN,asub->B->cmap->N); 3235 } 3236 3237 ierr = PetscFree(garray);CHKERRQ(ierr); 3238 *submat = M; 3239 3240 /* Save isrow_d, iscol_d and iscol_o used in processor for next request */ 3241 ierr = PetscObjectCompose((PetscObject)M,"isrow_d",(PetscObject)isrow_d);CHKERRQ(ierr); 3242 ierr = ISDestroy(&isrow_d);CHKERRQ(ierr); 3243 3244 ierr = PetscObjectCompose((PetscObject)M,"iscol_d",(PetscObject)iscol_d);CHKERRQ(ierr); 3245 ierr = ISDestroy(&iscol_d);CHKERRQ(ierr); 3246 3247 ierr = PetscObjectCompose((PetscObject)M,"iscol_o",(PetscObject)iscol_o);CHKERRQ(ierr); 3248 ierr = ISDestroy(&iscol_o);CHKERRQ(ierr); 3249 } 3250 PetscFunctionReturn(0); 3251 } 3252 3253 PetscErrorCode MatCreateSubMatrix_MPIAIJ(Mat mat,IS isrow,IS iscol,MatReuse call,Mat *newmat) 3254 { 3255 PetscErrorCode ierr; 3256 IS iscol_local=NULL,isrow_d; 3257 PetscInt csize; 3258 PetscInt n,i,j,start,end; 3259 PetscBool sameRowDist=PETSC_FALSE,sameDist[2],tsameDist[2]; 3260 MPI_Comm comm; 3261 3262 PetscFunctionBegin; 3263 /* If isrow has same processor distribution as mat, 3264 call MatCreateSubMatrix_MPIAIJ_SameRowDist() to avoid using a hash table with global size of iscol */ 3265 if (call == MAT_REUSE_MATRIX) { 3266 ierr = PetscObjectQuery((PetscObject)*newmat,"isrow_d",(PetscObject*)&isrow_d);CHKERRQ(ierr); 3267 if (isrow_d) { 3268 sameRowDist = PETSC_TRUE; 3269 tsameDist[1] = PETSC_TRUE; /* sameColDist */ 3270 } else { 3271 ierr = PetscObjectQuery((PetscObject)*newmat,"SubIScol",(PetscObject*)&iscol_local);CHKERRQ(ierr); 3272 if (iscol_local) { 3273 sameRowDist = PETSC_TRUE; 3274 tsameDist[1] = PETSC_FALSE; /* !sameColDist */ 3275 } 3276 } 3277 } else { 3278 /* Check if isrow has same processor distribution as mat */ 3279 sameDist[0] = PETSC_FALSE; 3280 ierr = ISGetLocalSize(isrow,&n);CHKERRQ(ierr); 3281 if (!n) { 3282 sameDist[0] = PETSC_TRUE; 3283 } else { 3284 ierr = ISGetMinMax(isrow,&i,&j);CHKERRQ(ierr); 3285 ierr = MatGetOwnershipRange(mat,&start,&end);CHKERRQ(ierr); 3286 if (i >= start && j < end) { 3287 sameDist[0] = PETSC_TRUE; 3288 } 3289 } 3290 3291 /* Check if iscol has same processor distribution as mat */ 3292 sameDist[1] = PETSC_FALSE; 3293 ierr = ISGetLocalSize(iscol,&n);CHKERRQ(ierr); 3294 if (!n) { 3295 sameDist[1] = PETSC_TRUE; 3296 } else { 3297 ierr = ISGetMinMax(iscol,&i,&j);CHKERRQ(ierr); 3298 ierr = MatGetOwnershipRangeColumn(mat,&start,&end);CHKERRQ(ierr); 3299 if (i >= start && j < end) sameDist[1] = PETSC_TRUE; 3300 } 3301 3302 ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 3303 ierr = MPIU_Allreduce(&sameDist,&tsameDist,2,MPIU_BOOL,MPI_LAND,comm);CHKERRQ(ierr); 3304 sameRowDist = tsameDist[0]; 3305 } 3306 3307 if (sameRowDist) { 3308 if (tsameDist[1]) { /* sameRowDist & sameColDist */ 3309 /* isrow and iscol have same processor distribution as mat */ 3310 ierr = MatCreateSubMatrix_MPIAIJ_SameRowColDist(mat,isrow,iscol,call,newmat);CHKERRQ(ierr); 3311 PetscFunctionReturn(0); 3312 } else { /* sameRowDist */ 3313 /* isrow has same processor distribution as mat */ 3314 if (call == MAT_INITIAL_MATRIX) { 3315 PetscBool sorted; 3316 ierr = ISGetSeqIS_Private(mat,iscol,&iscol_local);CHKERRQ(ierr); 3317 ierr = ISGetLocalSize(iscol_local,&n);CHKERRQ(ierr); /* local size of iscol_local = global columns of newmat */ 3318 ierr = ISGetSize(iscol,&i);CHKERRQ(ierr); 3319 if (n != i) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"n %d != size of iscol %d",n,i); 3320 3321 ierr = ISSorted(iscol_local,&sorted);CHKERRQ(ierr); 3322 if (sorted) { 3323 /* MatCreateSubMatrix_MPIAIJ_SameRowDist() requires iscol_local be sorted; it can have duplicate indices */ 3324 ierr = MatCreateSubMatrix_MPIAIJ_SameRowDist(mat,isrow,iscol,iscol_local,MAT_INITIAL_MATRIX,newmat);CHKERRQ(ierr); 3325 PetscFunctionReturn(0); 3326 } 3327 } else { /* call == MAT_REUSE_MATRIX */ 3328 IS iscol_sub; 3329 ierr = PetscObjectQuery((PetscObject)*newmat,"SubIScol",(PetscObject*)&iscol_sub);CHKERRQ(ierr); 3330 if (iscol_sub) { 3331 ierr = MatCreateSubMatrix_MPIAIJ_SameRowDist(mat,isrow,iscol,NULL,call,newmat);CHKERRQ(ierr); 3332 PetscFunctionReturn(0); 3333 } 3334 } 3335 } 3336 } 3337 3338 /* General case: iscol -> iscol_local which has global size of iscol */ 3339 if (call == MAT_REUSE_MATRIX) { 3340 ierr = PetscObjectQuery((PetscObject)*newmat,"ISAllGather",(PetscObject*)&iscol_local);CHKERRQ(ierr); 3341 if (!iscol_local) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Submatrix passed in was not used before, cannot reuse"); 3342 } else { 3343 if (!iscol_local) { 3344 ierr = ISGetSeqIS_Private(mat,iscol,&iscol_local);CHKERRQ(ierr); 3345 } 3346 } 3347 3348 ierr = ISGetLocalSize(iscol,&csize);CHKERRQ(ierr); 3349 ierr = MatCreateSubMatrix_MPIAIJ_nonscalable(mat,isrow,iscol_local,csize,call,newmat);CHKERRQ(ierr); 3350 3351 if (call == MAT_INITIAL_MATRIX) { 3352 ierr = PetscObjectCompose((PetscObject)*newmat,"ISAllGather",(PetscObject)iscol_local);CHKERRQ(ierr); 3353 ierr = ISDestroy(&iscol_local);CHKERRQ(ierr); 3354 } 3355 PetscFunctionReturn(0); 3356 } 3357 3358 /*@C 3359 MatCreateMPIAIJWithSeqAIJ - creates a MPIAIJ matrix using SeqAIJ matrices that contain the "diagonal" 3360 and "off-diagonal" part of the matrix in CSR format. 3361 3362 Collective on MPI_Comm 3363 3364 Input Parameters: 3365 + comm - MPI communicator 3366 . A - "diagonal" portion of matrix 3367 . B - "off-diagonal" portion of matrix, may have empty columns, will be destroyed by this routine 3368 - garray - global index of B columns 3369 3370 Output Parameter: 3371 . mat - the matrix, with input A as its local diagonal matrix 3372 Level: advanced 3373 3374 Notes: 3375 See MatCreateAIJ() for the definition of "diagonal" and "off-diagonal" portion of the matrix. 3376 A becomes part of output mat, B is destroyed by this routine. The user cannot use A and B anymore. 3377 3378 .seealso: MatCreateMPIAIJWithSplitArrays() 3379 @*/ 3380 PetscErrorCode MatCreateMPIAIJWithSeqAIJ(MPI_Comm comm,Mat A,Mat B,const PetscInt garray[],Mat *mat) 3381 { 3382 PetscErrorCode ierr; 3383 Mat_MPIAIJ *maij; 3384 Mat_SeqAIJ *b=(Mat_SeqAIJ*)B->data,*bnew; 3385 PetscInt *oi=b->i,*oj=b->j,i,nz,col; 3386 PetscScalar *oa=b->a; 3387 Mat Bnew; 3388 PetscInt m,n,N; 3389 3390 PetscFunctionBegin; 3391 ierr = MatCreate(comm,mat);CHKERRQ(ierr); 3392 ierr = MatGetSize(A,&m,&n);CHKERRQ(ierr); 3393 if (m != B->rmap->N) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Am %D != Bm %D",m,B->rmap->N); 3394 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); 3395 /* remove check below; When B is created using iscol_o from ISGetSeqIS_SameColDist_Private(), its bs may not be same as A */ 3396 /* 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); */ 3397 3398 /* Get global columns of mat */ 3399 ierr = MPIU_Allreduce(&n,&N,1,MPIU_INT,MPI_SUM,comm);CHKERRQ(ierr); 3400 3401 ierr = MatSetSizes(*mat,m,n,PETSC_DECIDE,N);CHKERRQ(ierr); 3402 ierr = MatSetType(*mat,MATMPIAIJ);CHKERRQ(ierr); 3403 ierr = MatSetBlockSizes(*mat,A->rmap->bs,A->cmap->bs);CHKERRQ(ierr); 3404 maij = (Mat_MPIAIJ*)(*mat)->data; 3405 3406 (*mat)->preallocated = PETSC_TRUE; 3407 3408 ierr = PetscLayoutSetUp((*mat)->rmap);CHKERRQ(ierr); 3409 ierr = PetscLayoutSetUp((*mat)->cmap);CHKERRQ(ierr); 3410 3411 /* Set A as diagonal portion of *mat */ 3412 maij->A = A; 3413 3414 nz = oi[m]; 3415 for (i=0; i<nz; i++) { 3416 col = oj[i]; 3417 oj[i] = garray[col]; 3418 } 3419 3420 /* Set Bnew as off-diagonal portion of *mat */ 3421 ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,m,N,oi,oj,oa,&Bnew);CHKERRQ(ierr); 3422 bnew = (Mat_SeqAIJ*)Bnew->data; 3423 bnew->maxnz = b->maxnz; /* allocated nonzeros of B */ 3424 maij->B = Bnew; 3425 3426 if (B->rmap->N != Bnew->rmap->N) SETERRQ2(PETSC_COMM_SELF,0,"BN %d != BnewN %d",B->rmap->N,Bnew->rmap->N); 3427 3428 b->singlemalloc = PETSC_FALSE; /* B arrays are shared by Bnew */ 3429 b->free_a = PETSC_FALSE; 3430 b->free_ij = PETSC_FALSE; 3431 ierr = MatDestroy(&B);CHKERRQ(ierr); 3432 3433 bnew->singlemalloc = PETSC_TRUE; /* arrays will be freed by MatDestroy(&Bnew) */ 3434 bnew->free_a = PETSC_TRUE; 3435 bnew->free_ij = PETSC_TRUE; 3436 3437 /* condense columns of maij->B */ 3438 ierr = MatSetOption(*mat,MAT_NO_OFF_PROC_ENTRIES,PETSC_TRUE);CHKERRQ(ierr); 3439 ierr = MatAssemblyBegin(*mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3440 ierr = MatAssemblyEnd(*mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3441 ierr = MatSetOption(*mat,MAT_NO_OFF_PROC_ENTRIES,PETSC_FALSE);CHKERRQ(ierr); 3442 ierr = MatSetOption(*mat,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE);CHKERRQ(ierr); 3443 PetscFunctionReturn(0); 3444 } 3445 3446 extern PetscErrorCode MatCreateSubMatrices_MPIAIJ_SingleIS_Local(Mat,PetscInt,const IS[],const IS[],MatReuse,PetscBool,Mat*); 3447 3448 PetscErrorCode MatCreateSubMatrix_MPIAIJ_SameRowDist(Mat mat,IS isrow,IS iscol,IS iscol_local,MatReuse call,Mat *newmat) 3449 { 3450 PetscErrorCode ierr; 3451 PetscInt i,m,n,rstart,row,rend,nz,j,bs,cbs; 3452 PetscInt *ii,*jj,nlocal,*dlens,*olens,dlen,olen,jend,mglobal; 3453 Mat_MPIAIJ *a=(Mat_MPIAIJ*)mat->data; 3454 Mat M,Msub,B=a->B; 3455 MatScalar *aa; 3456 Mat_SeqAIJ *aij; 3457 PetscInt *garray = a->garray,*colsub,Ncols; 3458 PetscInt count,Bn=B->cmap->N,cstart=mat->cmap->rstart,cend=mat->cmap->rend; 3459 IS iscol_sub,iscmap; 3460 const PetscInt *is_idx,*cmap; 3461 PetscBool allcolumns=PETSC_FALSE; 3462 MPI_Comm comm; 3463 3464 PetscFunctionBegin; 3465 ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 3466 3467 if (call == MAT_REUSE_MATRIX) { 3468 ierr = PetscObjectQuery((PetscObject)*newmat,"SubIScol",(PetscObject*)&iscol_sub);CHKERRQ(ierr); 3469 if (!iscol_sub) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"SubIScol passed in was not used before, cannot reuse"); 3470 ierr = ISGetLocalSize(iscol_sub,&count);CHKERRQ(ierr); 3471 3472 ierr = PetscObjectQuery((PetscObject)*newmat,"Subcmap",(PetscObject*)&iscmap);CHKERRQ(ierr); 3473 if (!iscmap) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Subcmap passed in was not used before, cannot reuse"); 3474 3475 ierr = PetscObjectQuery((PetscObject)*newmat,"SubMatrix",(PetscObject*)&Msub);CHKERRQ(ierr); 3476 if (!Msub) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Submatrix passed in was not used before, cannot reuse"); 3477 3478 ierr = MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat,1,&isrow,&iscol_sub,MAT_REUSE_MATRIX,PETSC_FALSE,&Msub);CHKERRQ(ierr); 3479 3480 } else { /* call == MAT_INITIAL_MATRIX) */ 3481 PetscBool flg; 3482 3483 ierr = ISGetLocalSize(iscol,&n);CHKERRQ(ierr); 3484 ierr = ISGetSize(iscol,&Ncols);CHKERRQ(ierr); 3485 3486 /* (1) iscol -> nonscalable iscol_local */ 3487 /* Check for special case: each processor gets entire matrix columns */ 3488 ierr = ISIdentity(iscol_local,&flg);CHKERRQ(ierr); 3489 if (flg && n == mat->cmap->N) allcolumns = PETSC_TRUE; 3490 if (allcolumns) { 3491 iscol_sub = iscol_local; 3492 ierr = PetscObjectReference((PetscObject)iscol_local);CHKERRQ(ierr); 3493 ierr = ISCreateStride(PETSC_COMM_SELF,n,0,1,&iscmap);CHKERRQ(ierr); 3494 3495 } else { 3496 /* (2) iscol_local -> iscol_sub and iscmap. Implementation below requires iscol_local be sorted, it can have duplicate indices */ 3497 PetscInt *idx,*cmap1,k; 3498 ierr = PetscMalloc1(Ncols,&idx);CHKERRQ(ierr); 3499 ierr = PetscMalloc1(Ncols,&cmap1);CHKERRQ(ierr); 3500 ierr = ISGetIndices(iscol_local,&is_idx);CHKERRQ(ierr); 3501 count = 0; 3502 k = 0; 3503 for (i=0; i<Ncols; i++) { 3504 j = is_idx[i]; 3505 if (j >= cstart && j < cend) { 3506 /* diagonal part of mat */ 3507 idx[count] = j; 3508 cmap1[count++] = i; /* column index in submat */ 3509 } else if (Bn) { 3510 /* off-diagonal part of mat */ 3511 if (j == garray[k]) { 3512 idx[count] = j; 3513 cmap1[count++] = i; /* column index in submat */ 3514 } else if (j > garray[k]) { 3515 while (j > garray[k] && k < Bn-1) k++; 3516 if (j == garray[k]) { 3517 idx[count] = j; 3518 cmap1[count++] = i; /* column index in submat */ 3519 } 3520 } 3521 } 3522 } 3523 ierr = ISRestoreIndices(iscol_local,&is_idx);CHKERRQ(ierr); 3524 3525 ierr = ISCreateGeneral(PETSC_COMM_SELF,count,idx,PETSC_OWN_POINTER,&iscol_sub);CHKERRQ(ierr); 3526 ierr = ISGetBlockSize(iscol,&cbs);CHKERRQ(ierr); 3527 ierr = ISSetBlockSize(iscol_sub,cbs);CHKERRQ(ierr); 3528 3529 ierr = ISCreateGeneral(PetscObjectComm((PetscObject)iscol_local),count,cmap1,PETSC_OWN_POINTER,&iscmap);CHKERRQ(ierr); 3530 } 3531 3532 /* (3) Create sequential Msub */ 3533 ierr = MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat,1,&isrow,&iscol_sub,MAT_INITIAL_MATRIX,allcolumns,&Msub);CHKERRQ(ierr); 3534 } 3535 3536 ierr = ISGetLocalSize(iscol_sub,&count);CHKERRQ(ierr); 3537 aij = (Mat_SeqAIJ*)(Msub)->data; 3538 ii = aij->i; 3539 ierr = ISGetIndices(iscmap,&cmap);CHKERRQ(ierr); 3540 3541 /* 3542 m - number of local rows 3543 Ncols - number of columns (same on all processors) 3544 rstart - first row in new global matrix generated 3545 */ 3546 ierr = MatGetSize(Msub,&m,NULL);CHKERRQ(ierr); 3547 3548 if (call == MAT_INITIAL_MATRIX) { 3549 /* (4) Create parallel newmat */ 3550 PetscMPIInt rank,size; 3551 PetscInt csize; 3552 3553 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 3554 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 3555 3556 /* 3557 Determine the number of non-zeros in the diagonal and off-diagonal 3558 portions of the matrix in order to do correct preallocation 3559 */ 3560 3561 /* first get start and end of "diagonal" columns */ 3562 ierr = ISGetLocalSize(iscol,&csize);CHKERRQ(ierr); 3563 if (csize == PETSC_DECIDE) { 3564 ierr = ISGetSize(isrow,&mglobal);CHKERRQ(ierr); 3565 if (mglobal == Ncols) { /* square matrix */ 3566 nlocal = m; 3567 } else { 3568 nlocal = Ncols/size + ((Ncols % size) > rank); 3569 } 3570 } else { 3571 nlocal = csize; 3572 } 3573 ierr = MPI_Scan(&nlocal,&rend,1,MPIU_INT,MPI_SUM,comm);CHKERRQ(ierr); 3574 rstart = rend - nlocal; 3575 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); 3576 3577 /* next, compute all the lengths */ 3578 jj = aij->j; 3579 ierr = PetscMalloc1(2*m+1,&dlens);CHKERRQ(ierr); 3580 olens = dlens + m; 3581 for (i=0; i<m; i++) { 3582 jend = ii[i+1] - ii[i]; 3583 olen = 0; 3584 dlen = 0; 3585 for (j=0; j<jend; j++) { 3586 if (cmap[*jj] < rstart || cmap[*jj] >= rend) olen++; 3587 else dlen++; 3588 jj++; 3589 } 3590 olens[i] = olen; 3591 dlens[i] = dlen; 3592 } 3593 3594 ierr = ISGetBlockSize(isrow,&bs);CHKERRQ(ierr); 3595 ierr = ISGetBlockSize(iscol,&cbs);CHKERRQ(ierr); 3596 3597 ierr = MatCreate(comm,&M);CHKERRQ(ierr); 3598 ierr = MatSetSizes(M,m,nlocal,PETSC_DECIDE,Ncols);CHKERRQ(ierr); 3599 ierr = MatSetBlockSizes(M,bs,cbs);CHKERRQ(ierr); 3600 ierr = MatSetType(M,((PetscObject)mat)->type_name);CHKERRQ(ierr); 3601 ierr = MatMPIAIJSetPreallocation(M,0,dlens,0,olens);CHKERRQ(ierr); 3602 ierr = PetscFree(dlens);CHKERRQ(ierr); 3603 3604 } else { /* call == MAT_REUSE_MATRIX */ 3605 M = *newmat; 3606 ierr = MatGetLocalSize(M,&i,NULL);CHKERRQ(ierr); 3607 if (i != m) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Previous matrix must be same size/layout as request"); 3608 ierr = MatZeroEntries(M);CHKERRQ(ierr); 3609 /* 3610 The next two lines are needed so we may call MatSetValues_MPIAIJ() below directly, 3611 rather than the slower MatSetValues(). 3612 */ 3613 M->was_assembled = PETSC_TRUE; 3614 M->assembled = PETSC_FALSE; 3615 } 3616 3617 /* (5) Set values of Msub to *newmat */ 3618 ierr = PetscMalloc1(count,&colsub);CHKERRQ(ierr); 3619 ierr = MatGetOwnershipRange(M,&rstart,NULL);CHKERRQ(ierr); 3620 3621 jj = aij->j; 3622 aa = aij->a; 3623 for (i=0; i<m; i++) { 3624 row = rstart + i; 3625 nz = ii[i+1] - ii[i]; 3626 for (j=0; j<nz; j++) colsub[j] = cmap[jj[j]]; 3627 ierr = MatSetValues_MPIAIJ(M,1,&row,nz,colsub,aa,INSERT_VALUES);CHKERRQ(ierr); 3628 jj += nz; aa += nz; 3629 } 3630 ierr = ISRestoreIndices(iscmap,&cmap);CHKERRQ(ierr); 3631 3632 ierr = MatAssemblyBegin(M,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3633 ierr = MatAssemblyEnd(M,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3634 3635 ierr = PetscFree(colsub);CHKERRQ(ierr); 3636 3637 /* save Msub, iscol_sub and iscmap used in processor for next request */ 3638 if (call == MAT_INITIAL_MATRIX) { 3639 *newmat = M; 3640 ierr = PetscObjectCompose((PetscObject)(*newmat),"SubMatrix",(PetscObject)Msub);CHKERRQ(ierr); 3641 ierr = MatDestroy(&Msub);CHKERRQ(ierr); 3642 3643 ierr = PetscObjectCompose((PetscObject)(*newmat),"SubIScol",(PetscObject)iscol_sub);CHKERRQ(ierr); 3644 ierr = ISDestroy(&iscol_sub);CHKERRQ(ierr); 3645 3646 ierr = PetscObjectCompose((PetscObject)(*newmat),"Subcmap",(PetscObject)iscmap);CHKERRQ(ierr); 3647 ierr = ISDestroy(&iscmap);CHKERRQ(ierr); 3648 3649 if (iscol_local) { 3650 ierr = PetscObjectCompose((PetscObject)(*newmat),"ISAllGather",(PetscObject)iscol_local);CHKERRQ(ierr); 3651 ierr = ISDestroy(&iscol_local);CHKERRQ(ierr); 3652 } 3653 } 3654 PetscFunctionReturn(0); 3655 } 3656 3657 /* 3658 Not great since it makes two copies of the submatrix, first an SeqAIJ 3659 in local and then by concatenating the local matrices the end result. 3660 Writing it directly would be much like MatCreateSubMatrices_MPIAIJ() 3661 3662 Note: This requires a sequential iscol with all indices. 3663 */ 3664 PetscErrorCode MatCreateSubMatrix_MPIAIJ_nonscalable(Mat mat,IS isrow,IS iscol,PetscInt csize,MatReuse call,Mat *newmat) 3665 { 3666 PetscErrorCode ierr; 3667 PetscMPIInt rank,size; 3668 PetscInt i,m,n,rstart,row,rend,nz,*cwork,j,bs,cbs; 3669 PetscInt *ii,*jj,nlocal,*dlens,*olens,dlen,olen,jend,mglobal; 3670 Mat M,Mreuse; 3671 MatScalar *aa,*vwork; 3672 MPI_Comm comm; 3673 Mat_SeqAIJ *aij; 3674 PetscBool colflag,allcolumns=PETSC_FALSE; 3675 3676 PetscFunctionBegin; 3677 ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 3678 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 3679 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 3680 3681 /* Check for special case: each processor gets entire matrix columns */ 3682 ierr = ISIdentity(iscol,&colflag);CHKERRQ(ierr); 3683 ierr = ISGetLocalSize(iscol,&n);CHKERRQ(ierr); 3684 if (colflag && n == mat->cmap->N) allcolumns = PETSC_TRUE; 3685 3686 if (call == MAT_REUSE_MATRIX) { 3687 ierr = PetscObjectQuery((PetscObject)*newmat,"SubMatrix",(PetscObject*)&Mreuse);CHKERRQ(ierr); 3688 if (!Mreuse) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Submatrix passed in was not used before, cannot reuse"); 3689 ierr = MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat,1,&isrow,&iscol,MAT_REUSE_MATRIX,allcolumns,&Mreuse);CHKERRQ(ierr); 3690 } else { 3691 ierr = MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat,1,&isrow,&iscol,MAT_INITIAL_MATRIX,allcolumns,&Mreuse);CHKERRQ(ierr); 3692 } 3693 3694 /* 3695 m - number of local rows 3696 n - number of columns (same on all processors) 3697 rstart - first row in new global matrix generated 3698 */ 3699 ierr = MatGetSize(Mreuse,&m,&n);CHKERRQ(ierr); 3700 ierr = MatGetBlockSizes(Mreuse,&bs,&cbs);CHKERRQ(ierr); 3701 if (call == MAT_INITIAL_MATRIX) { 3702 aij = (Mat_SeqAIJ*)(Mreuse)->data; 3703 ii = aij->i; 3704 jj = aij->j; 3705 3706 /* 3707 Determine the number of non-zeros in the diagonal and off-diagonal 3708 portions of the matrix in order to do correct preallocation 3709 */ 3710 3711 /* first get start and end of "diagonal" columns */ 3712 if (csize == PETSC_DECIDE) { 3713 ierr = ISGetSize(isrow,&mglobal);CHKERRQ(ierr); 3714 if (mglobal == n) { /* square matrix */ 3715 nlocal = m; 3716 } else { 3717 nlocal = n/size + ((n % size) > rank); 3718 } 3719 } else { 3720 nlocal = csize; 3721 } 3722 ierr = MPI_Scan(&nlocal,&rend,1,MPIU_INT,MPI_SUM,comm);CHKERRQ(ierr); 3723 rstart = rend - nlocal; 3724 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); 3725 3726 /* next, compute all the lengths */ 3727 ierr = PetscMalloc1(2*m+1,&dlens);CHKERRQ(ierr); 3728 olens = dlens + m; 3729 for (i=0; i<m; i++) { 3730 jend = ii[i+1] - ii[i]; 3731 olen = 0; 3732 dlen = 0; 3733 for (j=0; j<jend; j++) { 3734 if (*jj < rstart || *jj >= rend) olen++; 3735 else dlen++; 3736 jj++; 3737 } 3738 olens[i] = olen; 3739 dlens[i] = dlen; 3740 } 3741 ierr = MatCreate(comm,&M);CHKERRQ(ierr); 3742 ierr = MatSetSizes(M,m,nlocal,PETSC_DECIDE,n);CHKERRQ(ierr); 3743 ierr = MatSetBlockSizes(M,bs,cbs);CHKERRQ(ierr); 3744 ierr = MatSetType(M,((PetscObject)mat)->type_name);CHKERRQ(ierr); 3745 ierr = MatMPIAIJSetPreallocation(M,0,dlens,0,olens);CHKERRQ(ierr); 3746 ierr = PetscFree(dlens);CHKERRQ(ierr); 3747 } else { 3748 PetscInt ml,nl; 3749 3750 M = *newmat; 3751 ierr = MatGetLocalSize(M,&ml,&nl);CHKERRQ(ierr); 3752 if (ml != m) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Previous matrix must be same size/layout as request"); 3753 ierr = MatZeroEntries(M);CHKERRQ(ierr); 3754 /* 3755 The next two lines are needed so we may call MatSetValues_MPIAIJ() below directly, 3756 rather than the slower MatSetValues(). 3757 */ 3758 M->was_assembled = PETSC_TRUE; 3759 M->assembled = PETSC_FALSE; 3760 } 3761 ierr = MatGetOwnershipRange(M,&rstart,&rend);CHKERRQ(ierr); 3762 aij = (Mat_SeqAIJ*)(Mreuse)->data; 3763 ii = aij->i; 3764 jj = aij->j; 3765 aa = aij->a; 3766 for (i=0; i<m; i++) { 3767 row = rstart + i; 3768 nz = ii[i+1] - ii[i]; 3769 cwork = jj; jj += nz; 3770 vwork = aa; aa += nz; 3771 ierr = MatSetValues_MPIAIJ(M,1,&row,nz,cwork,vwork,INSERT_VALUES);CHKERRQ(ierr); 3772 } 3773 3774 ierr = MatAssemblyBegin(M,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3775 ierr = MatAssemblyEnd(M,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3776 *newmat = M; 3777 3778 /* save submatrix used in processor for next request */ 3779 if (call == MAT_INITIAL_MATRIX) { 3780 ierr = PetscObjectCompose((PetscObject)M,"SubMatrix",(PetscObject)Mreuse);CHKERRQ(ierr); 3781 ierr = MatDestroy(&Mreuse);CHKERRQ(ierr); 3782 } 3783 PetscFunctionReturn(0); 3784 } 3785 3786 PetscErrorCode MatMPIAIJSetPreallocationCSR_MPIAIJ(Mat B,const PetscInt Ii[],const PetscInt J[],const PetscScalar v[]) 3787 { 3788 PetscInt m,cstart, cend,j,nnz,i,d; 3789 PetscInt *d_nnz,*o_nnz,nnz_max = 0,rstart,ii; 3790 const PetscInt *JJ; 3791 PetscScalar *values; 3792 PetscErrorCode ierr; 3793 PetscBool nooffprocentries; 3794 3795 PetscFunctionBegin; 3796 if (Ii && Ii[0]) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Ii[0] must be 0 it is %D",Ii[0]); 3797 3798 ierr = PetscLayoutSetUp(B->rmap);CHKERRQ(ierr); 3799 ierr = PetscLayoutSetUp(B->cmap);CHKERRQ(ierr); 3800 m = B->rmap->n; 3801 cstart = B->cmap->rstart; 3802 cend = B->cmap->rend; 3803 rstart = B->rmap->rstart; 3804 3805 ierr = PetscMalloc2(m,&d_nnz,m,&o_nnz);CHKERRQ(ierr); 3806 3807 #if defined(PETSC_USE_DEBUG) 3808 for (i=0; i<m; i++) { 3809 nnz = Ii[i+1]- Ii[i]; 3810 JJ = J + Ii[i]; 3811 if (nnz < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Local row %D has a negative %D number of columns",i,nnz); 3812 if (nnz && (JJ[0] < 0)) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Row %D starts with negative column index",i,JJ[0]); 3813 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); 3814 } 3815 #endif 3816 3817 for (i=0; i<m; i++) { 3818 nnz = Ii[i+1]- Ii[i]; 3819 JJ = J + Ii[i]; 3820 nnz_max = PetscMax(nnz_max,nnz); 3821 d = 0; 3822 for (j=0; j<nnz; j++) { 3823 if (cstart <= JJ[j] && JJ[j] < cend) d++; 3824 } 3825 d_nnz[i] = d; 3826 o_nnz[i] = nnz - d; 3827 } 3828 ierr = MatMPIAIJSetPreallocation(B,0,d_nnz,0,o_nnz);CHKERRQ(ierr); 3829 ierr = PetscFree2(d_nnz,o_nnz);CHKERRQ(ierr); 3830 3831 if (v) values = (PetscScalar*)v; 3832 else { 3833 ierr = PetscCalloc1(nnz_max+1,&values);CHKERRQ(ierr); 3834 } 3835 3836 for (i=0; i<m; i++) { 3837 ii = i + rstart; 3838 nnz = Ii[i+1]- Ii[i]; 3839 ierr = MatSetValues_MPIAIJ(B,1,&ii,nnz,J+Ii[i],values+(v ? Ii[i] : 0),INSERT_VALUES);CHKERRQ(ierr); 3840 } 3841 nooffprocentries = B->nooffprocentries; 3842 B->nooffprocentries = PETSC_TRUE; 3843 ierr = MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3844 ierr = MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3845 B->nooffprocentries = nooffprocentries; 3846 3847 if (!v) { 3848 ierr = PetscFree(values);CHKERRQ(ierr); 3849 } 3850 ierr = MatSetOption(B,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE);CHKERRQ(ierr); 3851 PetscFunctionReturn(0); 3852 } 3853 3854 /*@ 3855 MatMPIAIJSetPreallocationCSR - Allocates memory for a sparse parallel matrix in AIJ format 3856 (the default parallel PETSc format). 3857 3858 Collective on MPI_Comm 3859 3860 Input Parameters: 3861 + B - the matrix 3862 . i - the indices into j for the start of each local row (starts with zero) 3863 . j - the column indices for each local row (starts with zero) 3864 - v - optional values in the matrix 3865 3866 Level: developer 3867 3868 Notes: 3869 The i, j, and a arrays ARE copied by this routine into the internal format used by PETSc; 3870 thus you CANNOT change the matrix entries by changing the values of a[] after you have 3871 called this routine. Use MatCreateMPIAIJWithSplitArrays() to avoid needing to copy the arrays. 3872 3873 The i and j indices are 0 based, and i indices are indices corresponding to the local j array. 3874 3875 The format which is used for the sparse matrix input, is equivalent to a 3876 row-major ordering.. i.e for the following matrix, the input data expected is 3877 as shown 3878 3879 $ 1 0 0 3880 $ 2 0 3 P0 3881 $ ------- 3882 $ 4 5 6 P1 3883 $ 3884 $ Process0 [P0]: rows_owned=[0,1] 3885 $ i = {0,1,3} [size = nrow+1 = 2+1] 3886 $ j = {0,0,2} [size = 3] 3887 $ v = {1,2,3} [size = 3] 3888 $ 3889 $ Process1 [P1]: rows_owned=[2] 3890 $ i = {0,3} [size = nrow+1 = 1+1] 3891 $ j = {0,1,2} [size = 3] 3892 $ v = {4,5,6} [size = 3] 3893 3894 .keywords: matrix, aij, compressed row, sparse, parallel 3895 3896 .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatCreateAIJ(), MATMPIAIJ, 3897 MatCreateSeqAIJWithArrays(), MatCreateMPIAIJWithSplitArrays() 3898 @*/ 3899 PetscErrorCode MatMPIAIJSetPreallocationCSR(Mat B,const PetscInt i[],const PetscInt j[], const PetscScalar v[]) 3900 { 3901 PetscErrorCode ierr; 3902 3903 PetscFunctionBegin; 3904 ierr = PetscTryMethod(B,"MatMPIAIJSetPreallocationCSR_C",(Mat,const PetscInt[],const PetscInt[],const PetscScalar[]),(B,i,j,v));CHKERRQ(ierr); 3905 PetscFunctionReturn(0); 3906 } 3907 3908 /*@C 3909 MatMPIAIJSetPreallocation - Preallocates memory for a sparse parallel matrix in AIJ format 3910 (the default parallel PETSc format). For good matrix assembly performance 3911 the user should preallocate the matrix storage by setting the parameters 3912 d_nz (or d_nnz) and o_nz (or o_nnz). By setting these parameters accurately, 3913 performance can be increased by more than a factor of 50. 3914 3915 Collective on MPI_Comm 3916 3917 Input Parameters: 3918 + B - the matrix 3919 . d_nz - number of nonzeros per row in DIAGONAL portion of local submatrix 3920 (same value is used for all local rows) 3921 . d_nnz - array containing the number of nonzeros in the various rows of the 3922 DIAGONAL portion of the local submatrix (possibly different for each row) 3923 or NULL (PETSC_NULL_INTEGER in Fortran), if d_nz is used to specify the nonzero structure. 3924 The size of this array is equal to the number of local rows, i.e 'm'. 3925 For matrices that will be factored, you must leave room for (and set) 3926 the diagonal entry even if it is zero. 3927 . o_nz - number of nonzeros per row in the OFF-DIAGONAL portion of local 3928 submatrix (same value is used for all local rows). 3929 - o_nnz - array containing the number of nonzeros in the various rows of the 3930 OFF-DIAGONAL portion of the local submatrix (possibly different for 3931 each row) or NULL (PETSC_NULL_INTEGER in Fortran), if o_nz is used to specify the nonzero 3932 structure. The size of this array is equal to the number 3933 of local rows, i.e 'm'. 3934 3935 If the *_nnz parameter is given then the *_nz parameter is ignored 3936 3937 The AIJ format (also called the Yale sparse matrix format or 3938 compressed row storage (CSR)), is fully compatible with standard Fortran 77 3939 storage. The stored row and column indices begin with zero. 3940 See Users-Manual: ch_mat for details. 3941 3942 The parallel matrix is partitioned such that the first m0 rows belong to 3943 process 0, the next m1 rows belong to process 1, the next m2 rows belong 3944 to process 2 etc.. where m0,m1,m2... are the input parameter 'm'. 3945 3946 The DIAGONAL portion of the local submatrix of a processor can be defined 3947 as the submatrix which is obtained by extraction the part corresponding to 3948 the rows r1-r2 and columns c1-c2 of the global matrix, where r1 is the 3949 first row that belongs to the processor, r2 is the last row belonging to 3950 the this processor, and c1-c2 is range of indices of the local part of a 3951 vector suitable for applying the matrix to. This is an mxn matrix. In the 3952 common case of a square matrix, the row and column ranges are the same and 3953 the DIAGONAL part is also square. The remaining portion of the local 3954 submatrix (mxN) constitute the OFF-DIAGONAL portion. 3955 3956 If o_nnz, d_nnz are specified, then o_nz, and d_nz are ignored. 3957 3958 You can call MatGetInfo() to get information on how effective the preallocation was; 3959 for example the fields mallocs,nz_allocated,nz_used,nz_unneeded; 3960 You can also run with the option -info and look for messages with the string 3961 malloc in them to see if additional memory allocation was needed. 3962 3963 Example usage: 3964 3965 Consider the following 8x8 matrix with 34 non-zero values, that is 3966 assembled across 3 processors. Lets assume that proc0 owns 3 rows, 3967 proc1 owns 3 rows, proc2 owns 2 rows. This division can be shown 3968 as follows: 3969 3970 .vb 3971 1 2 0 | 0 3 0 | 0 4 3972 Proc0 0 5 6 | 7 0 0 | 8 0 3973 9 0 10 | 11 0 0 | 12 0 3974 ------------------------------------- 3975 13 0 14 | 15 16 17 | 0 0 3976 Proc1 0 18 0 | 19 20 21 | 0 0 3977 0 0 0 | 22 23 0 | 24 0 3978 ------------------------------------- 3979 Proc2 25 26 27 | 0 0 28 | 29 0 3980 30 0 0 | 31 32 33 | 0 34 3981 .ve 3982 3983 This can be represented as a collection of submatrices as: 3984 3985 .vb 3986 A B C 3987 D E F 3988 G H I 3989 .ve 3990 3991 Where the submatrices A,B,C are owned by proc0, D,E,F are 3992 owned by proc1, G,H,I are owned by proc2. 3993 3994 The 'm' parameters for proc0,proc1,proc2 are 3,3,2 respectively. 3995 The 'n' parameters for proc0,proc1,proc2 are 3,3,2 respectively. 3996 The 'M','N' parameters are 8,8, and have the same values on all procs. 3997 3998 The DIAGONAL submatrices corresponding to proc0,proc1,proc2 are 3999 submatrices [A], [E], [I] respectively. The OFF-DIAGONAL submatrices 4000 corresponding to proc0,proc1,proc2 are [BC], [DF], [GH] respectively. 4001 Internally, each processor stores the DIAGONAL part, and the OFF-DIAGONAL 4002 part as SeqAIJ matrices. for eg: proc1 will store [E] as a SeqAIJ 4003 matrix, ans [DF] as another SeqAIJ matrix. 4004 4005 When d_nz, o_nz parameters are specified, d_nz storage elements are 4006 allocated for every row of the local diagonal submatrix, and o_nz 4007 storage locations are allocated for every row of the OFF-DIAGONAL submat. 4008 One way to choose d_nz and o_nz is to use the max nonzerors per local 4009 rows for each of the local DIAGONAL, and the OFF-DIAGONAL submatrices. 4010 In this case, the values of d_nz,o_nz are: 4011 .vb 4012 proc0 : dnz = 2, o_nz = 2 4013 proc1 : dnz = 3, o_nz = 2 4014 proc2 : dnz = 1, o_nz = 4 4015 .ve 4016 We are allocating m*(d_nz+o_nz) storage locations for every proc. This 4017 translates to 3*(2+2)=12 for proc0, 3*(3+2)=15 for proc1, 2*(1+4)=10 4018 for proc3. i.e we are using 12+15+10=37 storage locations to store 4019 34 values. 4020 4021 When d_nnz, o_nnz parameters are specified, the storage is specified 4022 for every row, coresponding to both DIAGONAL and OFF-DIAGONAL submatrices. 4023 In the above case the values for d_nnz,o_nnz are: 4024 .vb 4025 proc0: d_nnz = [2,2,2] and o_nnz = [2,2,2] 4026 proc1: d_nnz = [3,3,2] and o_nnz = [2,1,1] 4027 proc2: d_nnz = [1,1] and o_nnz = [4,4] 4028 .ve 4029 Here the space allocated is sum of all the above values i.e 34, and 4030 hence pre-allocation is perfect. 4031 4032 Level: intermediate 4033 4034 .keywords: matrix, aij, compressed row, sparse, parallel 4035 4036 .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatCreateAIJ(), MatMPIAIJSetPreallocationCSR(), 4037 MATMPIAIJ, MatGetInfo(), PetscSplitOwnership() 4038 @*/ 4039 PetscErrorCode MatMPIAIJSetPreallocation(Mat B,PetscInt d_nz,const PetscInt d_nnz[],PetscInt o_nz,const PetscInt o_nnz[]) 4040 { 4041 PetscErrorCode ierr; 4042 4043 PetscFunctionBegin; 4044 PetscValidHeaderSpecific(B,MAT_CLASSID,1); 4045 PetscValidType(B,1); 4046 ierr = PetscTryMethod(B,"MatMPIAIJSetPreallocation_C",(Mat,PetscInt,const PetscInt[],PetscInt,const PetscInt[]),(B,d_nz,d_nnz,o_nz,o_nnz));CHKERRQ(ierr); 4047 PetscFunctionReturn(0); 4048 } 4049 4050 /*@ 4051 MatCreateMPIAIJWithArrays - creates a MPI AIJ matrix using arrays that contain in standard 4052 CSR format the local rows. 4053 4054 Collective on MPI_Comm 4055 4056 Input Parameters: 4057 + comm - MPI communicator 4058 . m - number of local rows (Cannot be PETSC_DECIDE) 4059 . n - This value should be the same as the local size used in creating the 4060 x vector for the matrix-vector product y = Ax. (or PETSC_DECIDE to have 4061 calculated if N is given) For square matrices n is almost always m. 4062 . M - number of global rows (or PETSC_DETERMINE to have calculated if m is given) 4063 . N - number of global columns (or PETSC_DETERMINE to have calculated if n is given) 4064 . i - row indices 4065 . j - column indices 4066 - a - matrix values 4067 4068 Output Parameter: 4069 . mat - the matrix 4070 4071 Level: intermediate 4072 4073 Notes: 4074 The i, j, and a arrays ARE copied by this routine into the internal format used by PETSc; 4075 thus you CANNOT change the matrix entries by changing the values of a[] after you have 4076 called this routine. Use MatCreateMPIAIJWithSplitArrays() to avoid needing to copy the arrays. 4077 4078 The i and j indices are 0 based, and i indices are indices corresponding to the local j array. 4079 4080 The format which is used for the sparse matrix input, is equivalent to a 4081 row-major ordering.. i.e for the following matrix, the input data expected is 4082 as shown 4083 4084 $ 1 0 0 4085 $ 2 0 3 P0 4086 $ ------- 4087 $ 4 5 6 P1 4088 $ 4089 $ Process0 [P0]: rows_owned=[0,1] 4090 $ i = {0,1,3} [size = nrow+1 = 2+1] 4091 $ j = {0,0,2} [size = 3] 4092 $ v = {1,2,3} [size = 3] 4093 $ 4094 $ Process1 [P1]: rows_owned=[2] 4095 $ i = {0,3} [size = nrow+1 = 1+1] 4096 $ j = {0,1,2} [size = 3] 4097 $ v = {4,5,6} [size = 3] 4098 4099 .keywords: matrix, aij, compressed row, sparse, parallel 4100 4101 .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatMPIAIJSetPreallocationCSR(), 4102 MATMPIAIJ, MatCreateAIJ(), MatCreateMPIAIJWithSplitArrays() 4103 @*/ 4104 PetscErrorCode MatCreateMPIAIJWithArrays(MPI_Comm comm,PetscInt m,PetscInt n,PetscInt M,PetscInt N,const PetscInt i[],const PetscInt j[],const PetscScalar a[],Mat *mat) 4105 { 4106 PetscErrorCode ierr; 4107 4108 PetscFunctionBegin; 4109 if (i && i[0]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"i (row indices) must start with 0"); 4110 if (m < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"local number of rows (m) cannot be PETSC_DECIDE, or negative"); 4111 ierr = MatCreate(comm,mat);CHKERRQ(ierr); 4112 ierr = MatSetSizes(*mat,m,n,M,N);CHKERRQ(ierr); 4113 /* ierr = MatSetBlockSizes(M,bs,cbs);CHKERRQ(ierr); */ 4114 ierr = MatSetType(*mat,MATMPIAIJ);CHKERRQ(ierr); 4115 ierr = MatMPIAIJSetPreallocationCSR(*mat,i,j,a);CHKERRQ(ierr); 4116 PetscFunctionReturn(0); 4117 } 4118 4119 /*@C 4120 MatCreateAIJ - Creates a sparse parallel matrix in AIJ format 4121 (the default parallel PETSc format). For good matrix assembly performance 4122 the user should preallocate the matrix storage by setting the parameters 4123 d_nz (or d_nnz) and o_nz (or o_nnz). By setting these parameters accurately, 4124 performance can be increased by more than a factor of 50. 4125 4126 Collective on MPI_Comm 4127 4128 Input Parameters: 4129 + comm - MPI communicator 4130 . m - number of local rows (or PETSC_DECIDE to have calculated if M is given) 4131 This value should be the same as the local size used in creating the 4132 y vector for the matrix-vector product y = Ax. 4133 . n - This value should be the same as the local size used in creating the 4134 x vector for the matrix-vector product y = Ax. (or PETSC_DECIDE to have 4135 calculated if N is given) For square matrices n is almost always m. 4136 . M - number of global rows (or PETSC_DETERMINE to have calculated if m is given) 4137 . N - number of global columns (or PETSC_DETERMINE to have calculated if n is given) 4138 . d_nz - number of nonzeros per row in DIAGONAL portion of local submatrix 4139 (same value is used for all local rows) 4140 . d_nnz - array containing the number of nonzeros in the various rows of the 4141 DIAGONAL portion of the local submatrix (possibly different for each row) 4142 or NULL, if d_nz is used to specify the nonzero structure. 4143 The size of this array is equal to the number of local rows, i.e 'm'. 4144 . o_nz - number of nonzeros per row in the OFF-DIAGONAL portion of local 4145 submatrix (same value is used for all local rows). 4146 - o_nnz - array containing the number of nonzeros in the various rows of the 4147 OFF-DIAGONAL portion of the local submatrix (possibly different for 4148 each row) or NULL, if o_nz is used to specify the nonzero 4149 structure. The size of this array is equal to the number 4150 of local rows, i.e 'm'. 4151 4152 Output Parameter: 4153 . A - the matrix 4154 4155 It is recommended that one use the MatCreate(), MatSetType() and/or MatSetFromOptions(), 4156 MatXXXXSetPreallocation() paradgm instead of this routine directly. 4157 [MatXXXXSetPreallocation() is, for example, MatSeqAIJSetPreallocation] 4158 4159 Notes: 4160 If the *_nnz parameter is given then the *_nz parameter is ignored 4161 4162 m,n,M,N parameters specify the size of the matrix, and its partitioning across 4163 processors, while d_nz,d_nnz,o_nz,o_nnz parameters specify the approximate 4164 storage requirements for this matrix. 4165 4166 If PETSC_DECIDE or PETSC_DETERMINE is used for a particular argument on one 4167 processor than it must be used on all processors that share the object for 4168 that argument. 4169 4170 The user MUST specify either the local or global matrix dimensions 4171 (possibly both). 4172 4173 The parallel matrix is partitioned across processors such that the 4174 first m0 rows belong to process 0, the next m1 rows belong to 4175 process 1, the next m2 rows belong to process 2 etc.. where 4176 m0,m1,m2,.. are the input parameter 'm'. i.e each processor stores 4177 values corresponding to [m x N] submatrix. 4178 4179 The columns are logically partitioned with the n0 columns belonging 4180 to 0th partition, the next n1 columns belonging to the next 4181 partition etc.. where n0,n1,n2... are the input parameter 'n'. 4182 4183 The DIAGONAL portion of the local submatrix on any given processor 4184 is the submatrix corresponding to the rows and columns m,n 4185 corresponding to the given processor. i.e diagonal matrix on 4186 process 0 is [m0 x n0], diagonal matrix on process 1 is [m1 x n1] 4187 etc. The remaining portion of the local submatrix [m x (N-n)] 4188 constitute the OFF-DIAGONAL portion. The example below better 4189 illustrates this concept. 4190 4191 For a square global matrix we define each processor's diagonal portion 4192 to be its local rows and the corresponding columns (a square submatrix); 4193 each processor's off-diagonal portion encompasses the remainder of the 4194 local matrix (a rectangular submatrix). 4195 4196 If o_nnz, d_nnz are specified, then o_nz, and d_nz are ignored. 4197 4198 When calling this routine with a single process communicator, a matrix of 4199 type SEQAIJ is returned. If a matrix of type MPIAIJ is desired for this 4200 type of communicator, use the construction mechanism 4201 .vb 4202 MatCreate(...,&A); MatSetType(A,MATMPIAIJ); MatSetSizes(A, m,n,M,N); MatMPIAIJSetPreallocation(A,...); 4203 .ve 4204 4205 $ MatCreate(...,&A); 4206 $ MatSetType(A,MATMPIAIJ); 4207 $ MatSetSizes(A, m,n,M,N); 4208 $ MatMPIAIJSetPreallocation(A,...); 4209 4210 By default, this format uses inodes (identical nodes) when possible. 4211 We search for consecutive rows with the same nonzero structure, thereby 4212 reusing matrix information to achieve increased efficiency. 4213 4214 Options Database Keys: 4215 + -mat_no_inode - Do not use inodes 4216 - -mat_inode_limit <limit> - Sets inode limit (max limit=5) 4217 4218 4219 4220 Example usage: 4221 4222 Consider the following 8x8 matrix with 34 non-zero values, that is 4223 assembled across 3 processors. Lets assume that proc0 owns 3 rows, 4224 proc1 owns 3 rows, proc2 owns 2 rows. This division can be shown 4225 as follows 4226 4227 .vb 4228 1 2 0 | 0 3 0 | 0 4 4229 Proc0 0 5 6 | 7 0 0 | 8 0 4230 9 0 10 | 11 0 0 | 12 0 4231 ------------------------------------- 4232 13 0 14 | 15 16 17 | 0 0 4233 Proc1 0 18 0 | 19 20 21 | 0 0 4234 0 0 0 | 22 23 0 | 24 0 4235 ------------------------------------- 4236 Proc2 25 26 27 | 0 0 28 | 29 0 4237 30 0 0 | 31 32 33 | 0 34 4238 .ve 4239 4240 This can be represented as a collection of submatrices as 4241 4242 .vb 4243 A B C 4244 D E F 4245 G H I 4246 .ve 4247 4248 Where the submatrices A,B,C are owned by proc0, D,E,F are 4249 owned by proc1, G,H,I are owned by proc2. 4250 4251 The 'm' parameters for proc0,proc1,proc2 are 3,3,2 respectively. 4252 The 'n' parameters for proc0,proc1,proc2 are 3,3,2 respectively. 4253 The 'M','N' parameters are 8,8, and have the same values on all procs. 4254 4255 The DIAGONAL submatrices corresponding to proc0,proc1,proc2 are 4256 submatrices [A], [E], [I] respectively. The OFF-DIAGONAL submatrices 4257 corresponding to proc0,proc1,proc2 are [BC], [DF], [GH] respectively. 4258 Internally, each processor stores the DIAGONAL part, and the OFF-DIAGONAL 4259 part as SeqAIJ matrices. for eg: proc1 will store [E] as a SeqAIJ 4260 matrix, ans [DF] as another SeqAIJ matrix. 4261 4262 When d_nz, o_nz parameters are specified, d_nz storage elements are 4263 allocated for every row of the local diagonal submatrix, and o_nz 4264 storage locations are allocated for every row of the OFF-DIAGONAL submat. 4265 One way to choose d_nz and o_nz is to use the max nonzerors per local 4266 rows for each of the local DIAGONAL, and the OFF-DIAGONAL submatrices. 4267 In this case, the values of d_nz,o_nz are 4268 .vb 4269 proc0 : dnz = 2, o_nz = 2 4270 proc1 : dnz = 3, o_nz = 2 4271 proc2 : dnz = 1, o_nz = 4 4272 .ve 4273 We are allocating m*(d_nz+o_nz) storage locations for every proc. This 4274 translates to 3*(2+2)=12 for proc0, 3*(3+2)=15 for proc1, 2*(1+4)=10 4275 for proc3. i.e we are using 12+15+10=37 storage locations to store 4276 34 values. 4277 4278 When d_nnz, o_nnz parameters are specified, the storage is specified 4279 for every row, coresponding to both DIAGONAL and OFF-DIAGONAL submatrices. 4280 In the above case the values for d_nnz,o_nnz are 4281 .vb 4282 proc0: d_nnz = [2,2,2] and o_nnz = [2,2,2] 4283 proc1: d_nnz = [3,3,2] and o_nnz = [2,1,1] 4284 proc2: d_nnz = [1,1] and o_nnz = [4,4] 4285 .ve 4286 Here the space allocated is sum of all the above values i.e 34, and 4287 hence pre-allocation is perfect. 4288 4289 Level: intermediate 4290 4291 .keywords: matrix, aij, compressed row, sparse, parallel 4292 4293 .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatMPIAIJSetPreallocationCSR(), 4294 MATMPIAIJ, MatCreateMPIAIJWithArrays() 4295 @*/ 4296 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) 4297 { 4298 PetscErrorCode ierr; 4299 PetscMPIInt size; 4300 4301 PetscFunctionBegin; 4302 ierr = MatCreate(comm,A);CHKERRQ(ierr); 4303 ierr = MatSetSizes(*A,m,n,M,N);CHKERRQ(ierr); 4304 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 4305 if (size > 1) { 4306 ierr = MatSetType(*A,MATMPIAIJ);CHKERRQ(ierr); 4307 ierr = MatMPIAIJSetPreallocation(*A,d_nz,d_nnz,o_nz,o_nnz);CHKERRQ(ierr); 4308 } else { 4309 ierr = MatSetType(*A,MATSEQAIJ);CHKERRQ(ierr); 4310 ierr = MatSeqAIJSetPreallocation(*A,d_nz,d_nnz);CHKERRQ(ierr); 4311 } 4312 PetscFunctionReturn(0); 4313 } 4314 4315 PetscErrorCode MatMPIAIJGetSeqAIJ(Mat A,Mat *Ad,Mat *Ao,const PetscInt *colmap[]) 4316 { 4317 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 4318 PetscBool flg; 4319 PetscErrorCode ierr; 4320 4321 PetscFunctionBegin; 4322 ierr = PetscObjectTypeCompare((PetscObject)A,MATMPIAIJ,&flg);CHKERRQ(ierr); 4323 if (!flg) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"This function requires a MATMPIAIJ matrix as input"); 4324 if (Ad) *Ad = a->A; 4325 if (Ao) *Ao = a->B; 4326 if (colmap) *colmap = a->garray; 4327 PetscFunctionReturn(0); 4328 } 4329 4330 PetscErrorCode MatCreateMPIMatConcatenateSeqMat_MPIAIJ(MPI_Comm comm,Mat inmat,PetscInt n,MatReuse scall,Mat *outmat) 4331 { 4332 PetscErrorCode ierr; 4333 PetscInt m,N,i,rstart,nnz,Ii; 4334 PetscInt *indx; 4335 PetscScalar *values; 4336 4337 PetscFunctionBegin; 4338 ierr = MatGetSize(inmat,&m,&N);CHKERRQ(ierr); 4339 if (scall == MAT_INITIAL_MATRIX) { /* symbolic phase */ 4340 PetscInt *dnz,*onz,sum,bs,cbs; 4341 4342 if (n == PETSC_DECIDE) { 4343 ierr = PetscSplitOwnership(comm,&n,&N);CHKERRQ(ierr); 4344 } 4345 /* Check sum(n) = N */ 4346 ierr = MPIU_Allreduce(&n,&sum,1,MPIU_INT,MPI_SUM,comm);CHKERRQ(ierr); 4347 if (sum != N) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_INCOMP,"Sum of local columns %D != global columns %D",sum,N); 4348 4349 ierr = MPI_Scan(&m, &rstart,1,MPIU_INT,MPI_SUM,comm);CHKERRQ(ierr); 4350 rstart -= m; 4351 4352 ierr = MatPreallocateInitialize(comm,m,n,dnz,onz);CHKERRQ(ierr); 4353 for (i=0; i<m; i++) { 4354 ierr = MatGetRow_SeqAIJ(inmat,i,&nnz,&indx,NULL);CHKERRQ(ierr); 4355 ierr = MatPreallocateSet(i+rstart,nnz,indx,dnz,onz);CHKERRQ(ierr); 4356 ierr = MatRestoreRow_SeqAIJ(inmat,i,&nnz,&indx,NULL);CHKERRQ(ierr); 4357 } 4358 4359 ierr = MatCreate(comm,outmat);CHKERRQ(ierr); 4360 ierr = MatSetSizes(*outmat,m,n,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 4361 ierr = MatGetBlockSizes(inmat,&bs,&cbs);CHKERRQ(ierr); 4362 ierr = MatSetBlockSizes(*outmat,bs,cbs);CHKERRQ(ierr); 4363 ierr = MatSetType(*outmat,MATAIJ);CHKERRQ(ierr); 4364 ierr = MatSeqAIJSetPreallocation(*outmat,0,dnz);CHKERRQ(ierr); 4365 ierr = MatMPIAIJSetPreallocation(*outmat,0,dnz,0,onz);CHKERRQ(ierr); 4366 ierr = MatPreallocateFinalize(dnz,onz);CHKERRQ(ierr); 4367 } 4368 4369 /* numeric phase */ 4370 ierr = MatGetOwnershipRange(*outmat,&rstart,NULL);CHKERRQ(ierr); 4371 for (i=0; i<m; i++) { 4372 ierr = MatGetRow_SeqAIJ(inmat,i,&nnz,&indx,&values);CHKERRQ(ierr); 4373 Ii = i + rstart; 4374 ierr = MatSetValues(*outmat,1,&Ii,nnz,indx,values,INSERT_VALUES);CHKERRQ(ierr); 4375 ierr = MatRestoreRow_SeqAIJ(inmat,i,&nnz,&indx,&values);CHKERRQ(ierr); 4376 } 4377 ierr = MatAssemblyBegin(*outmat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4378 ierr = MatAssemblyEnd(*outmat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4379 PetscFunctionReturn(0); 4380 } 4381 4382 PetscErrorCode MatFileSplit(Mat A,char *outfile) 4383 { 4384 PetscErrorCode ierr; 4385 PetscMPIInt rank; 4386 PetscInt m,N,i,rstart,nnz; 4387 size_t len; 4388 const PetscInt *indx; 4389 PetscViewer out; 4390 char *name; 4391 Mat B; 4392 const PetscScalar *values; 4393 4394 PetscFunctionBegin; 4395 ierr = MatGetLocalSize(A,&m,0);CHKERRQ(ierr); 4396 ierr = MatGetSize(A,0,&N);CHKERRQ(ierr); 4397 /* Should this be the type of the diagonal block of A? */ 4398 ierr = MatCreate(PETSC_COMM_SELF,&B);CHKERRQ(ierr); 4399 ierr = MatSetSizes(B,m,N,m,N);CHKERRQ(ierr); 4400 ierr = MatSetBlockSizesFromMats(B,A,A);CHKERRQ(ierr); 4401 ierr = MatSetType(B,MATSEQAIJ);CHKERRQ(ierr); 4402 ierr = MatSeqAIJSetPreallocation(B,0,NULL);CHKERRQ(ierr); 4403 ierr = MatGetOwnershipRange(A,&rstart,0);CHKERRQ(ierr); 4404 for (i=0; i<m; i++) { 4405 ierr = MatGetRow(A,i+rstart,&nnz,&indx,&values);CHKERRQ(ierr); 4406 ierr = MatSetValues(B,1,&i,nnz,indx,values,INSERT_VALUES);CHKERRQ(ierr); 4407 ierr = MatRestoreRow(A,i+rstart,&nnz,&indx,&values);CHKERRQ(ierr); 4408 } 4409 ierr = MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4410 ierr = MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4411 4412 ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)A),&rank);CHKERRQ(ierr); 4413 ierr = PetscStrlen(outfile,&len);CHKERRQ(ierr); 4414 ierr = PetscMalloc1(len+5,&name);CHKERRQ(ierr); 4415 sprintf(name,"%s.%d",outfile,rank); 4416 ierr = PetscViewerBinaryOpen(PETSC_COMM_SELF,name,FILE_MODE_APPEND,&out);CHKERRQ(ierr); 4417 ierr = PetscFree(name);CHKERRQ(ierr); 4418 ierr = MatView(B,out);CHKERRQ(ierr); 4419 ierr = PetscViewerDestroy(&out);CHKERRQ(ierr); 4420 ierr = MatDestroy(&B);CHKERRQ(ierr); 4421 PetscFunctionReturn(0); 4422 } 4423 4424 PetscErrorCode MatDestroy_MPIAIJ_SeqsToMPI(Mat A) 4425 { 4426 PetscErrorCode ierr; 4427 Mat_Merge_SeqsToMPI *merge; 4428 PetscContainer container; 4429 4430 PetscFunctionBegin; 4431 ierr = PetscObjectQuery((PetscObject)A,"MatMergeSeqsToMPI",(PetscObject*)&container);CHKERRQ(ierr); 4432 if (container) { 4433 ierr = PetscContainerGetPointer(container,(void**)&merge);CHKERRQ(ierr); 4434 ierr = PetscFree(merge->id_r);CHKERRQ(ierr); 4435 ierr = PetscFree(merge->len_s);CHKERRQ(ierr); 4436 ierr = PetscFree(merge->len_r);CHKERRQ(ierr); 4437 ierr = PetscFree(merge->bi);CHKERRQ(ierr); 4438 ierr = PetscFree(merge->bj);CHKERRQ(ierr); 4439 ierr = PetscFree(merge->buf_ri[0]);CHKERRQ(ierr); 4440 ierr = PetscFree(merge->buf_ri);CHKERRQ(ierr); 4441 ierr = PetscFree(merge->buf_rj[0]);CHKERRQ(ierr); 4442 ierr = PetscFree(merge->buf_rj);CHKERRQ(ierr); 4443 ierr = PetscFree(merge->coi);CHKERRQ(ierr); 4444 ierr = PetscFree(merge->coj);CHKERRQ(ierr); 4445 ierr = PetscFree(merge->owners_co);CHKERRQ(ierr); 4446 ierr = PetscLayoutDestroy(&merge->rowmap);CHKERRQ(ierr); 4447 ierr = PetscFree(merge);CHKERRQ(ierr); 4448 ierr = PetscObjectCompose((PetscObject)A,"MatMergeSeqsToMPI",0);CHKERRQ(ierr); 4449 } 4450 ierr = MatDestroy_MPIAIJ(A);CHKERRQ(ierr); 4451 PetscFunctionReturn(0); 4452 } 4453 4454 #include <../src/mat/utils/freespace.h> 4455 #include <petscbt.h> 4456 4457 PetscErrorCode MatCreateMPIAIJSumSeqAIJNumeric(Mat seqmat,Mat mpimat) 4458 { 4459 PetscErrorCode ierr; 4460 MPI_Comm comm; 4461 Mat_SeqAIJ *a =(Mat_SeqAIJ*)seqmat->data; 4462 PetscMPIInt size,rank,taga,*len_s; 4463 PetscInt N=mpimat->cmap->N,i,j,*owners,*ai=a->i,*aj; 4464 PetscInt proc,m; 4465 PetscInt **buf_ri,**buf_rj; 4466 PetscInt k,anzi,*bj_i,*bi,*bj,arow,bnzi,nextaj; 4467 PetscInt nrows,**buf_ri_k,**nextrow,**nextai; 4468 MPI_Request *s_waits,*r_waits; 4469 MPI_Status *status; 4470 MatScalar *aa=a->a; 4471 MatScalar **abuf_r,*ba_i; 4472 Mat_Merge_SeqsToMPI *merge; 4473 PetscContainer container; 4474 4475 PetscFunctionBegin; 4476 ierr = PetscObjectGetComm((PetscObject)mpimat,&comm);CHKERRQ(ierr); 4477 ierr = PetscLogEventBegin(MAT_Seqstompinum,seqmat,0,0,0);CHKERRQ(ierr); 4478 4479 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 4480 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 4481 4482 ierr = PetscObjectQuery((PetscObject)mpimat,"MatMergeSeqsToMPI",(PetscObject*)&container);CHKERRQ(ierr); 4483 ierr = PetscContainerGetPointer(container,(void**)&merge);CHKERRQ(ierr); 4484 4485 bi = merge->bi; 4486 bj = merge->bj; 4487 buf_ri = merge->buf_ri; 4488 buf_rj = merge->buf_rj; 4489 4490 ierr = PetscMalloc1(size,&status);CHKERRQ(ierr); 4491 owners = merge->rowmap->range; 4492 len_s = merge->len_s; 4493 4494 /* send and recv matrix values */ 4495 /*-----------------------------*/ 4496 ierr = PetscObjectGetNewTag((PetscObject)mpimat,&taga);CHKERRQ(ierr); 4497 ierr = PetscPostIrecvScalar(comm,taga,merge->nrecv,merge->id_r,merge->len_r,&abuf_r,&r_waits);CHKERRQ(ierr); 4498 4499 ierr = PetscMalloc1(merge->nsend+1,&s_waits);CHKERRQ(ierr); 4500 for (proc=0,k=0; proc<size; proc++) { 4501 if (!len_s[proc]) continue; 4502 i = owners[proc]; 4503 ierr = MPI_Isend(aa+ai[i],len_s[proc],MPIU_MATSCALAR,proc,taga,comm,s_waits+k);CHKERRQ(ierr); 4504 k++; 4505 } 4506 4507 if (merge->nrecv) {ierr = MPI_Waitall(merge->nrecv,r_waits,status);CHKERRQ(ierr);} 4508 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,s_waits,status);CHKERRQ(ierr);} 4509 ierr = PetscFree(status);CHKERRQ(ierr); 4510 4511 ierr = PetscFree(s_waits);CHKERRQ(ierr); 4512 ierr = PetscFree(r_waits);CHKERRQ(ierr); 4513 4514 /* insert mat values of mpimat */ 4515 /*----------------------------*/ 4516 ierr = PetscMalloc1(N,&ba_i);CHKERRQ(ierr); 4517 ierr = PetscMalloc3(merge->nrecv,&buf_ri_k,merge->nrecv,&nextrow,merge->nrecv,&nextai);CHKERRQ(ierr); 4518 4519 for (k=0; k<merge->nrecv; k++) { 4520 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 4521 nrows = *(buf_ri_k[k]); 4522 nextrow[k] = buf_ri_k[k]+1; /* next row number of k-th recved i-structure */ 4523 nextai[k] = buf_ri_k[k] + (nrows + 1); /* poins to the next i-structure of k-th recved i-structure */ 4524 } 4525 4526 /* set values of ba */ 4527 m = merge->rowmap->n; 4528 for (i=0; i<m; i++) { 4529 arow = owners[rank] + i; 4530 bj_i = bj+bi[i]; /* col indices of the i-th row of mpimat */ 4531 bnzi = bi[i+1] - bi[i]; 4532 ierr = PetscMemzero(ba_i,bnzi*sizeof(PetscScalar));CHKERRQ(ierr); 4533 4534 /* add local non-zero vals of this proc's seqmat into ba */ 4535 anzi = ai[arow+1] - ai[arow]; 4536 aj = a->j + ai[arow]; 4537 aa = a->a + ai[arow]; 4538 nextaj = 0; 4539 for (j=0; nextaj<anzi; j++) { 4540 if (*(bj_i + j) == aj[nextaj]) { /* bcol == acol */ 4541 ba_i[j] += aa[nextaj++]; 4542 } 4543 } 4544 4545 /* add received vals into ba */ 4546 for (k=0; k<merge->nrecv; k++) { /* k-th received message */ 4547 /* i-th row */ 4548 if (i == *nextrow[k]) { 4549 anzi = *(nextai[k]+1) - *nextai[k]; 4550 aj = buf_rj[k] + *(nextai[k]); 4551 aa = abuf_r[k] + *(nextai[k]); 4552 nextaj = 0; 4553 for (j=0; nextaj<anzi; j++) { 4554 if (*(bj_i + j) == aj[nextaj]) { /* bcol == acol */ 4555 ba_i[j] += aa[nextaj++]; 4556 } 4557 } 4558 nextrow[k]++; nextai[k]++; 4559 } 4560 } 4561 ierr = MatSetValues(mpimat,1,&arow,bnzi,bj_i,ba_i,INSERT_VALUES);CHKERRQ(ierr); 4562 } 4563 ierr = MatAssemblyBegin(mpimat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4564 ierr = MatAssemblyEnd(mpimat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4565 4566 ierr = PetscFree(abuf_r[0]);CHKERRQ(ierr); 4567 ierr = PetscFree(abuf_r);CHKERRQ(ierr); 4568 ierr = PetscFree(ba_i);CHKERRQ(ierr); 4569 ierr = PetscFree3(buf_ri_k,nextrow,nextai);CHKERRQ(ierr); 4570 ierr = PetscLogEventEnd(MAT_Seqstompinum,seqmat,0,0,0);CHKERRQ(ierr); 4571 PetscFunctionReturn(0); 4572 } 4573 4574 PetscErrorCode MatCreateMPIAIJSumSeqAIJSymbolic(MPI_Comm comm,Mat seqmat,PetscInt m,PetscInt n,Mat *mpimat) 4575 { 4576 PetscErrorCode ierr; 4577 Mat B_mpi; 4578 Mat_SeqAIJ *a=(Mat_SeqAIJ*)seqmat->data; 4579 PetscMPIInt size,rank,tagi,tagj,*len_s,*len_si,*len_ri; 4580 PetscInt **buf_rj,**buf_ri,**buf_ri_k; 4581 PetscInt M=seqmat->rmap->n,N=seqmat->cmap->n,i,*owners,*ai=a->i,*aj=a->j; 4582 PetscInt len,proc,*dnz,*onz,bs,cbs; 4583 PetscInt k,anzi,*bi,*bj,*lnk,nlnk,arow,bnzi,nspacedouble=0; 4584 PetscInt nrows,*buf_s,*buf_si,*buf_si_i,**nextrow,**nextai; 4585 MPI_Request *si_waits,*sj_waits,*ri_waits,*rj_waits; 4586 MPI_Status *status; 4587 PetscFreeSpaceList free_space=NULL,current_space=NULL; 4588 PetscBT lnkbt; 4589 Mat_Merge_SeqsToMPI *merge; 4590 PetscContainer container; 4591 4592 PetscFunctionBegin; 4593 ierr = PetscLogEventBegin(MAT_Seqstompisym,seqmat,0,0,0);CHKERRQ(ierr); 4594 4595 /* make sure it is a PETSc comm */ 4596 ierr = PetscCommDuplicate(comm,&comm,NULL);CHKERRQ(ierr); 4597 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 4598 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 4599 4600 ierr = PetscNew(&merge);CHKERRQ(ierr); 4601 ierr = PetscMalloc1(size,&status);CHKERRQ(ierr); 4602 4603 /* determine row ownership */ 4604 /*---------------------------------------------------------*/ 4605 ierr = PetscLayoutCreate(comm,&merge->rowmap);CHKERRQ(ierr); 4606 ierr = PetscLayoutSetLocalSize(merge->rowmap,m);CHKERRQ(ierr); 4607 ierr = PetscLayoutSetSize(merge->rowmap,M);CHKERRQ(ierr); 4608 ierr = PetscLayoutSetBlockSize(merge->rowmap,1);CHKERRQ(ierr); 4609 ierr = PetscLayoutSetUp(merge->rowmap);CHKERRQ(ierr); 4610 ierr = PetscMalloc1(size,&len_si);CHKERRQ(ierr); 4611 ierr = PetscMalloc1(size,&merge->len_s);CHKERRQ(ierr); 4612 4613 m = merge->rowmap->n; 4614 owners = merge->rowmap->range; 4615 4616 /* determine the number of messages to send, their lengths */ 4617 /*---------------------------------------------------------*/ 4618 len_s = merge->len_s; 4619 4620 len = 0; /* length of buf_si[] */ 4621 merge->nsend = 0; 4622 for (proc=0; proc<size; proc++) { 4623 len_si[proc] = 0; 4624 if (proc == rank) { 4625 len_s[proc] = 0; 4626 } else { 4627 len_si[proc] = owners[proc+1] - owners[proc] + 1; 4628 len_s[proc] = ai[owners[proc+1]] - ai[owners[proc]]; /* num of rows to be sent to [proc] */ 4629 } 4630 if (len_s[proc]) { 4631 merge->nsend++; 4632 nrows = 0; 4633 for (i=owners[proc]; i<owners[proc+1]; i++) { 4634 if (ai[i+1] > ai[i]) nrows++; 4635 } 4636 len_si[proc] = 2*(nrows+1); 4637 len += len_si[proc]; 4638 } 4639 } 4640 4641 /* determine the number and length of messages to receive for ij-structure */ 4642 /*-------------------------------------------------------------------------*/ 4643 ierr = PetscGatherNumberOfMessages(comm,NULL,len_s,&merge->nrecv);CHKERRQ(ierr); 4644 ierr = PetscGatherMessageLengths2(comm,merge->nsend,merge->nrecv,len_s,len_si,&merge->id_r,&merge->len_r,&len_ri);CHKERRQ(ierr); 4645 4646 /* post the Irecv of j-structure */ 4647 /*-------------------------------*/ 4648 ierr = PetscCommGetNewTag(comm,&tagj);CHKERRQ(ierr); 4649 ierr = PetscPostIrecvInt(comm,tagj,merge->nrecv,merge->id_r,merge->len_r,&buf_rj,&rj_waits);CHKERRQ(ierr); 4650 4651 /* post the Isend of j-structure */ 4652 /*--------------------------------*/ 4653 ierr = PetscMalloc2(merge->nsend,&si_waits,merge->nsend,&sj_waits);CHKERRQ(ierr); 4654 4655 for (proc=0, k=0; proc<size; proc++) { 4656 if (!len_s[proc]) continue; 4657 i = owners[proc]; 4658 ierr = MPI_Isend(aj+ai[i],len_s[proc],MPIU_INT,proc,tagj,comm,sj_waits+k);CHKERRQ(ierr); 4659 k++; 4660 } 4661 4662 /* receives and sends of j-structure are complete */ 4663 /*------------------------------------------------*/ 4664 if (merge->nrecv) {ierr = MPI_Waitall(merge->nrecv,rj_waits,status);CHKERRQ(ierr);} 4665 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,sj_waits,status);CHKERRQ(ierr);} 4666 4667 /* send and recv i-structure */ 4668 /*---------------------------*/ 4669 ierr = PetscCommGetNewTag(comm,&tagi);CHKERRQ(ierr); 4670 ierr = PetscPostIrecvInt(comm,tagi,merge->nrecv,merge->id_r,len_ri,&buf_ri,&ri_waits);CHKERRQ(ierr); 4671 4672 ierr = PetscMalloc1(len+1,&buf_s);CHKERRQ(ierr); 4673 buf_si = buf_s; /* points to the beginning of k-th msg to be sent */ 4674 for (proc=0,k=0; proc<size; proc++) { 4675 if (!len_s[proc]) continue; 4676 /* form outgoing message for i-structure: 4677 buf_si[0]: nrows to be sent 4678 [1:nrows]: row index (global) 4679 [nrows+1:2*nrows+1]: i-structure index 4680 */ 4681 /*-------------------------------------------*/ 4682 nrows = len_si[proc]/2 - 1; 4683 buf_si_i = buf_si + nrows+1; 4684 buf_si[0] = nrows; 4685 buf_si_i[0] = 0; 4686 nrows = 0; 4687 for (i=owners[proc]; i<owners[proc+1]; i++) { 4688 anzi = ai[i+1] - ai[i]; 4689 if (anzi) { 4690 buf_si_i[nrows+1] = buf_si_i[nrows] + anzi; /* i-structure */ 4691 buf_si[nrows+1] = i-owners[proc]; /* local row index */ 4692 nrows++; 4693 } 4694 } 4695 ierr = MPI_Isend(buf_si,len_si[proc],MPIU_INT,proc,tagi,comm,si_waits+k);CHKERRQ(ierr); 4696 k++; 4697 buf_si += len_si[proc]; 4698 } 4699 4700 if (merge->nrecv) {ierr = MPI_Waitall(merge->nrecv,ri_waits,status);CHKERRQ(ierr);} 4701 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,si_waits,status);CHKERRQ(ierr);} 4702 4703 ierr = PetscInfo2(seqmat,"nsend: %D, nrecv: %D\n",merge->nsend,merge->nrecv);CHKERRQ(ierr); 4704 for (i=0; i<merge->nrecv; i++) { 4705 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); 4706 } 4707 4708 ierr = PetscFree(len_si);CHKERRQ(ierr); 4709 ierr = PetscFree(len_ri);CHKERRQ(ierr); 4710 ierr = PetscFree(rj_waits);CHKERRQ(ierr); 4711 ierr = PetscFree2(si_waits,sj_waits);CHKERRQ(ierr); 4712 ierr = PetscFree(ri_waits);CHKERRQ(ierr); 4713 ierr = PetscFree(buf_s);CHKERRQ(ierr); 4714 ierr = PetscFree(status);CHKERRQ(ierr); 4715 4716 /* compute a local seq matrix in each processor */ 4717 /*----------------------------------------------*/ 4718 /* allocate bi array and free space for accumulating nonzero column info */ 4719 ierr = PetscMalloc1(m+1,&bi);CHKERRQ(ierr); 4720 bi[0] = 0; 4721 4722 /* create and initialize a linked list */ 4723 nlnk = N+1; 4724 ierr = PetscLLCreate(N,N,nlnk,lnk,lnkbt);CHKERRQ(ierr); 4725 4726 /* initial FreeSpace size is 2*(num of local nnz(seqmat)) */ 4727 len = ai[owners[rank+1]] - ai[owners[rank]]; 4728 ierr = PetscFreeSpaceGet(PetscIntMultTruncate(2,len)+1,&free_space);CHKERRQ(ierr); 4729 4730 current_space = free_space; 4731 4732 /* determine symbolic info for each local row */ 4733 ierr = PetscMalloc3(merge->nrecv,&buf_ri_k,merge->nrecv,&nextrow,merge->nrecv,&nextai);CHKERRQ(ierr); 4734 4735 for (k=0; k<merge->nrecv; k++) { 4736 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 4737 nrows = *buf_ri_k[k]; 4738 nextrow[k] = buf_ri_k[k] + 1; /* next row number of k-th recved i-structure */ 4739 nextai[k] = buf_ri_k[k] + (nrows + 1); /* poins to the next i-structure of k-th recved i-structure */ 4740 } 4741 4742 ierr = MatPreallocateInitialize(comm,m,n,dnz,onz);CHKERRQ(ierr); 4743 len = 0; 4744 for (i=0; i<m; i++) { 4745 bnzi = 0; 4746 /* add local non-zero cols of this proc's seqmat into lnk */ 4747 arow = owners[rank] + i; 4748 anzi = ai[arow+1] - ai[arow]; 4749 aj = a->j + ai[arow]; 4750 ierr = PetscLLAddSorted(anzi,aj,N,nlnk,lnk,lnkbt);CHKERRQ(ierr); 4751 bnzi += nlnk; 4752 /* add received col data into lnk */ 4753 for (k=0; k<merge->nrecv; k++) { /* k-th received message */ 4754 if (i == *nextrow[k]) { /* i-th row */ 4755 anzi = *(nextai[k]+1) - *nextai[k]; 4756 aj = buf_rj[k] + *nextai[k]; 4757 ierr = PetscLLAddSorted(anzi,aj,N,nlnk,lnk,lnkbt);CHKERRQ(ierr); 4758 bnzi += nlnk; 4759 nextrow[k]++; nextai[k]++; 4760 } 4761 } 4762 if (len < bnzi) len = bnzi; /* =max(bnzi) */ 4763 4764 /* if free space is not available, make more free space */ 4765 if (current_space->local_remaining<bnzi) { 4766 ierr = PetscFreeSpaceGet(PetscIntSumTruncate(bnzi,current_space->total_array_size),¤t_space);CHKERRQ(ierr); 4767 nspacedouble++; 4768 } 4769 /* copy data into free space, then initialize lnk */ 4770 ierr = PetscLLClean(N,N,bnzi,lnk,current_space->array,lnkbt);CHKERRQ(ierr); 4771 ierr = MatPreallocateSet(i+owners[rank],bnzi,current_space->array,dnz,onz);CHKERRQ(ierr); 4772 4773 current_space->array += bnzi; 4774 current_space->local_used += bnzi; 4775 current_space->local_remaining -= bnzi; 4776 4777 bi[i+1] = bi[i] + bnzi; 4778 } 4779 4780 ierr = PetscFree3(buf_ri_k,nextrow,nextai);CHKERRQ(ierr); 4781 4782 ierr = PetscMalloc1(bi[m]+1,&bj);CHKERRQ(ierr); 4783 ierr = PetscFreeSpaceContiguous(&free_space,bj);CHKERRQ(ierr); 4784 ierr = PetscLLDestroy(lnk,lnkbt);CHKERRQ(ierr); 4785 4786 /* create symbolic parallel matrix B_mpi */ 4787 /*---------------------------------------*/ 4788 ierr = MatGetBlockSizes(seqmat,&bs,&cbs);CHKERRQ(ierr); 4789 ierr = MatCreate(comm,&B_mpi);CHKERRQ(ierr); 4790 if (n==PETSC_DECIDE) { 4791 ierr = MatSetSizes(B_mpi,m,n,PETSC_DETERMINE,N);CHKERRQ(ierr); 4792 } else { 4793 ierr = MatSetSizes(B_mpi,m,n,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 4794 } 4795 ierr = MatSetBlockSizes(B_mpi,bs,cbs);CHKERRQ(ierr); 4796 ierr = MatSetType(B_mpi,MATMPIAIJ);CHKERRQ(ierr); 4797 ierr = MatMPIAIJSetPreallocation(B_mpi,0,dnz,0,onz);CHKERRQ(ierr); 4798 ierr = MatPreallocateFinalize(dnz,onz);CHKERRQ(ierr); 4799 ierr = MatSetOption(B_mpi,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 4800 4801 /* B_mpi is not ready for use - assembly will be done by MatCreateMPIAIJSumSeqAIJNumeric() */ 4802 B_mpi->assembled = PETSC_FALSE; 4803 B_mpi->ops->destroy = MatDestroy_MPIAIJ_SeqsToMPI; 4804 merge->bi = bi; 4805 merge->bj = bj; 4806 merge->buf_ri = buf_ri; 4807 merge->buf_rj = buf_rj; 4808 merge->coi = NULL; 4809 merge->coj = NULL; 4810 merge->owners_co = NULL; 4811 4812 ierr = PetscCommDestroy(&comm);CHKERRQ(ierr); 4813 4814 /* attach the supporting struct to B_mpi for reuse */ 4815 ierr = PetscContainerCreate(PETSC_COMM_SELF,&container);CHKERRQ(ierr); 4816 ierr = PetscContainerSetPointer(container,merge);CHKERRQ(ierr); 4817 ierr = PetscObjectCompose((PetscObject)B_mpi,"MatMergeSeqsToMPI",(PetscObject)container);CHKERRQ(ierr); 4818 ierr = PetscContainerDestroy(&container);CHKERRQ(ierr); 4819 *mpimat = B_mpi; 4820 4821 ierr = PetscLogEventEnd(MAT_Seqstompisym,seqmat,0,0,0);CHKERRQ(ierr); 4822 PetscFunctionReturn(0); 4823 } 4824 4825 /*@C 4826 MatCreateMPIAIJSumSeqAIJ - Creates a MATMPIAIJ matrix by adding sequential 4827 matrices from each processor 4828 4829 Collective on MPI_Comm 4830 4831 Input Parameters: 4832 + comm - the communicators the parallel matrix will live on 4833 . seqmat - the input sequential matrices 4834 . m - number of local rows (or PETSC_DECIDE) 4835 . n - number of local columns (or PETSC_DECIDE) 4836 - scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX 4837 4838 Output Parameter: 4839 . mpimat - the parallel matrix generated 4840 4841 Level: advanced 4842 4843 Notes: 4844 The dimensions of the sequential matrix in each processor MUST be the same. 4845 The input seqmat is included into the container "Mat_Merge_SeqsToMPI", and will be 4846 destroyed when mpimat is destroyed. Call PetscObjectQuery() to access seqmat. 4847 @*/ 4848 PetscErrorCode MatCreateMPIAIJSumSeqAIJ(MPI_Comm comm,Mat seqmat,PetscInt m,PetscInt n,MatReuse scall,Mat *mpimat) 4849 { 4850 PetscErrorCode ierr; 4851 PetscMPIInt size; 4852 4853 PetscFunctionBegin; 4854 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 4855 if (size == 1) { 4856 ierr = PetscLogEventBegin(MAT_Seqstompi,seqmat,0,0,0);CHKERRQ(ierr); 4857 if (scall == MAT_INITIAL_MATRIX) { 4858 ierr = MatDuplicate(seqmat,MAT_COPY_VALUES,mpimat);CHKERRQ(ierr); 4859 } else { 4860 ierr = MatCopy(seqmat,*mpimat,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 4861 } 4862 ierr = PetscLogEventEnd(MAT_Seqstompi,seqmat,0,0,0);CHKERRQ(ierr); 4863 PetscFunctionReturn(0); 4864 } 4865 ierr = PetscLogEventBegin(MAT_Seqstompi,seqmat,0,0,0);CHKERRQ(ierr); 4866 if (scall == MAT_INITIAL_MATRIX) { 4867 ierr = MatCreateMPIAIJSumSeqAIJSymbolic(comm,seqmat,m,n,mpimat);CHKERRQ(ierr); 4868 } 4869 ierr = MatCreateMPIAIJSumSeqAIJNumeric(seqmat,*mpimat);CHKERRQ(ierr); 4870 ierr = PetscLogEventEnd(MAT_Seqstompi,seqmat,0,0,0);CHKERRQ(ierr); 4871 PetscFunctionReturn(0); 4872 } 4873 4874 /*@ 4875 MatMPIAIJGetLocalMat - Creates a SeqAIJ from a MATMPIAIJ matrix by taking all its local rows and putting them into a sequential matrix with 4876 mlocal rows and n columns. Where mlocal is the row count obtained with MatGetLocalSize() and n is the global column count obtained 4877 with MatGetSize() 4878 4879 Not Collective 4880 4881 Input Parameters: 4882 + A - the matrix 4883 . scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX 4884 4885 Output Parameter: 4886 . A_loc - the local sequential matrix generated 4887 4888 Level: developer 4889 4890 .seealso: MatGetOwnerShipRange(), MatMPIAIJGetLocalMatCondensed() 4891 4892 @*/ 4893 PetscErrorCode MatMPIAIJGetLocalMat(Mat A,MatReuse scall,Mat *A_loc) 4894 { 4895 PetscErrorCode ierr; 4896 Mat_MPIAIJ *mpimat=(Mat_MPIAIJ*)A->data; 4897 Mat_SeqAIJ *mat,*a,*b; 4898 PetscInt *ai,*aj,*bi,*bj,*cmap=mpimat->garray; 4899 MatScalar *aa,*ba,*cam; 4900 PetscScalar *ca; 4901 PetscInt am=A->rmap->n,i,j,k,cstart=A->cmap->rstart; 4902 PetscInt *ci,*cj,col,ncols_d,ncols_o,jo; 4903 PetscBool match; 4904 MPI_Comm comm; 4905 PetscMPIInt size; 4906 4907 PetscFunctionBegin; 4908 ierr = PetscObjectTypeCompare((PetscObject)A,MATMPIAIJ,&match);CHKERRQ(ierr); 4909 if (!match) SETERRQ(PetscObjectComm((PetscObject)A), PETSC_ERR_SUP,"Requires MATMPIAIJ matrix as input"); 4910 ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 4911 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 4912 if (size == 1 && scall == MAT_REUSE_MATRIX) PetscFunctionReturn(0); 4913 4914 ierr = PetscLogEventBegin(MAT_Getlocalmat,A,0,0,0);CHKERRQ(ierr); 4915 a = (Mat_SeqAIJ*)(mpimat->A)->data; 4916 b = (Mat_SeqAIJ*)(mpimat->B)->data; 4917 ai = a->i; aj = a->j; bi = b->i; bj = b->j; 4918 aa = a->a; ba = b->a; 4919 if (scall == MAT_INITIAL_MATRIX) { 4920 if (size == 1) { 4921 ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,am,A->cmap->N,ai,aj,aa,A_loc);CHKERRQ(ierr); 4922 PetscFunctionReturn(0); 4923 } 4924 4925 ierr = PetscMalloc1(1+am,&ci);CHKERRQ(ierr); 4926 ci[0] = 0; 4927 for (i=0; i<am; i++) { 4928 ci[i+1] = ci[i] + (ai[i+1] - ai[i]) + (bi[i+1] - bi[i]); 4929 } 4930 ierr = PetscMalloc1(1+ci[am],&cj);CHKERRQ(ierr); 4931 ierr = PetscMalloc1(1+ci[am],&ca);CHKERRQ(ierr); 4932 k = 0; 4933 for (i=0; i<am; i++) { 4934 ncols_o = bi[i+1] - bi[i]; 4935 ncols_d = ai[i+1] - ai[i]; 4936 /* off-diagonal portion of A */ 4937 for (jo=0; jo<ncols_o; jo++) { 4938 col = cmap[*bj]; 4939 if (col >= cstart) break; 4940 cj[k] = col; bj++; 4941 ca[k++] = *ba++; 4942 } 4943 /* diagonal portion of A */ 4944 for (j=0; j<ncols_d; j++) { 4945 cj[k] = cstart + *aj++; 4946 ca[k++] = *aa++; 4947 } 4948 /* off-diagonal portion of A */ 4949 for (j=jo; j<ncols_o; j++) { 4950 cj[k] = cmap[*bj++]; 4951 ca[k++] = *ba++; 4952 } 4953 } 4954 /* put together the new matrix */ 4955 ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,am,A->cmap->N,ci,cj,ca,A_loc);CHKERRQ(ierr); 4956 /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */ 4957 /* Since these are PETSc arrays, change flags to free them as necessary. */ 4958 mat = (Mat_SeqAIJ*)(*A_loc)->data; 4959 mat->free_a = PETSC_TRUE; 4960 mat->free_ij = PETSC_TRUE; 4961 mat->nonew = 0; 4962 } else if (scall == MAT_REUSE_MATRIX) { 4963 mat=(Mat_SeqAIJ*)(*A_loc)->data; 4964 ci = mat->i; cj = mat->j; cam = mat->a; 4965 for (i=0; i<am; i++) { 4966 /* off-diagonal portion of A */ 4967 ncols_o = bi[i+1] - bi[i]; 4968 for (jo=0; jo<ncols_o; jo++) { 4969 col = cmap[*bj]; 4970 if (col >= cstart) break; 4971 *cam++ = *ba++; bj++; 4972 } 4973 /* diagonal portion of A */ 4974 ncols_d = ai[i+1] - ai[i]; 4975 for (j=0; j<ncols_d; j++) *cam++ = *aa++; 4976 /* off-diagonal portion of A */ 4977 for (j=jo; j<ncols_o; j++) { 4978 *cam++ = *ba++; bj++; 4979 } 4980 } 4981 } else SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Invalid MatReuse %d",(int)scall); 4982 ierr = PetscLogEventEnd(MAT_Getlocalmat,A,0,0,0);CHKERRQ(ierr); 4983 PetscFunctionReturn(0); 4984 } 4985 4986 /*@C 4987 MatMPIAIJGetLocalMatCondensed - Creates a SeqAIJ matrix from an MATMPIAIJ matrix by taking all its local rows and NON-ZERO columns 4988 4989 Not Collective 4990 4991 Input Parameters: 4992 + A - the matrix 4993 . scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX 4994 - row, col - index sets of rows and columns to extract (or NULL) 4995 4996 Output Parameter: 4997 . A_loc - the local sequential matrix generated 4998 4999 Level: developer 5000 5001 .seealso: MatGetOwnershipRange(), MatMPIAIJGetLocalMat() 5002 5003 @*/ 5004 PetscErrorCode MatMPIAIJGetLocalMatCondensed(Mat A,MatReuse scall,IS *row,IS *col,Mat *A_loc) 5005 { 5006 Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data; 5007 PetscErrorCode ierr; 5008 PetscInt i,start,end,ncols,nzA,nzB,*cmap,imark,*idx; 5009 IS isrowa,iscola; 5010 Mat *aloc; 5011 PetscBool match; 5012 5013 PetscFunctionBegin; 5014 ierr = PetscObjectTypeCompare((PetscObject)A,MATMPIAIJ,&match);CHKERRQ(ierr); 5015 if (!match) SETERRQ(PetscObjectComm((PetscObject)A), PETSC_ERR_SUP,"Requires MATMPIAIJ matrix as input"); 5016 ierr = PetscLogEventBegin(MAT_Getlocalmatcondensed,A,0,0,0);CHKERRQ(ierr); 5017 if (!row) { 5018 start = A->rmap->rstart; end = A->rmap->rend; 5019 ierr = ISCreateStride(PETSC_COMM_SELF,end-start,start,1,&isrowa);CHKERRQ(ierr); 5020 } else { 5021 isrowa = *row; 5022 } 5023 if (!col) { 5024 start = A->cmap->rstart; 5025 cmap = a->garray; 5026 nzA = a->A->cmap->n; 5027 nzB = a->B->cmap->n; 5028 ierr = PetscMalloc1(nzA+nzB, &idx);CHKERRQ(ierr); 5029 ncols = 0; 5030 for (i=0; i<nzB; i++) { 5031 if (cmap[i] < start) idx[ncols++] = cmap[i]; 5032 else break; 5033 } 5034 imark = i; 5035 for (i=0; i<nzA; i++) idx[ncols++] = start + i; 5036 for (i=imark; i<nzB; i++) idx[ncols++] = cmap[i]; 5037 ierr = ISCreateGeneral(PETSC_COMM_SELF,ncols,idx,PETSC_OWN_POINTER,&iscola);CHKERRQ(ierr); 5038 } else { 5039 iscola = *col; 5040 } 5041 if (scall != MAT_INITIAL_MATRIX) { 5042 ierr = PetscMalloc1(1,&aloc);CHKERRQ(ierr); 5043 aloc[0] = *A_loc; 5044 } 5045 ierr = MatCreateSubMatrices(A,1,&isrowa,&iscola,scall,&aloc);CHKERRQ(ierr); 5046 *A_loc = aloc[0]; 5047 ierr = PetscFree(aloc);CHKERRQ(ierr); 5048 if (!row) { 5049 ierr = ISDestroy(&isrowa);CHKERRQ(ierr); 5050 } 5051 if (!col) { 5052 ierr = ISDestroy(&iscola);CHKERRQ(ierr); 5053 } 5054 ierr = PetscLogEventEnd(MAT_Getlocalmatcondensed,A,0,0,0);CHKERRQ(ierr); 5055 PetscFunctionReturn(0); 5056 } 5057 5058 /*@C 5059 MatGetBrowsOfAcols - Creates a SeqAIJ matrix by taking rows of B that equal to nonzero columns of local A 5060 5061 Collective on Mat 5062 5063 Input Parameters: 5064 + A,B - the matrices in mpiaij format 5065 . scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX 5066 - rowb, colb - index sets of rows and columns of B to extract (or NULL) 5067 5068 Output Parameter: 5069 + rowb, colb - index sets of rows and columns of B to extract 5070 - B_seq - the sequential matrix generated 5071 5072 Level: developer 5073 5074 @*/ 5075 PetscErrorCode MatGetBrowsOfAcols(Mat A,Mat B,MatReuse scall,IS *rowb,IS *colb,Mat *B_seq) 5076 { 5077 Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data; 5078 PetscErrorCode ierr; 5079 PetscInt *idx,i,start,ncols,nzA,nzB,*cmap,imark; 5080 IS isrowb,iscolb; 5081 Mat *bseq=NULL; 5082 5083 PetscFunctionBegin; 5084 if (A->cmap->rstart != B->rmap->rstart || A->cmap->rend != B->rmap->rend) { 5085 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); 5086 } 5087 ierr = PetscLogEventBegin(MAT_GetBrowsOfAcols,A,B,0,0);CHKERRQ(ierr); 5088 5089 if (scall == MAT_INITIAL_MATRIX) { 5090 start = A->cmap->rstart; 5091 cmap = a->garray; 5092 nzA = a->A->cmap->n; 5093 nzB = a->B->cmap->n; 5094 ierr = PetscMalloc1(nzA+nzB, &idx);CHKERRQ(ierr); 5095 ncols = 0; 5096 for (i=0; i<nzB; i++) { /* row < local row index */ 5097 if (cmap[i] < start) idx[ncols++] = cmap[i]; 5098 else break; 5099 } 5100 imark = i; 5101 for (i=0; i<nzA; i++) idx[ncols++] = start + i; /* local rows */ 5102 for (i=imark; i<nzB; i++) idx[ncols++] = cmap[i]; /* row > local row index */ 5103 ierr = ISCreateGeneral(PETSC_COMM_SELF,ncols,idx,PETSC_OWN_POINTER,&isrowb);CHKERRQ(ierr); 5104 ierr = ISCreateStride(PETSC_COMM_SELF,B->cmap->N,0,1,&iscolb);CHKERRQ(ierr); 5105 } else { 5106 if (!rowb || !colb) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"IS rowb and colb must be provided for MAT_REUSE_MATRIX"); 5107 isrowb = *rowb; iscolb = *colb; 5108 ierr = PetscMalloc1(1,&bseq);CHKERRQ(ierr); 5109 bseq[0] = *B_seq; 5110 } 5111 ierr = MatCreateSubMatrices(B,1,&isrowb,&iscolb,scall,&bseq);CHKERRQ(ierr); 5112 *B_seq = bseq[0]; 5113 ierr = PetscFree(bseq);CHKERRQ(ierr); 5114 if (!rowb) { 5115 ierr = ISDestroy(&isrowb);CHKERRQ(ierr); 5116 } else { 5117 *rowb = isrowb; 5118 } 5119 if (!colb) { 5120 ierr = ISDestroy(&iscolb);CHKERRQ(ierr); 5121 } else { 5122 *colb = iscolb; 5123 } 5124 ierr = PetscLogEventEnd(MAT_GetBrowsOfAcols,A,B,0,0);CHKERRQ(ierr); 5125 PetscFunctionReturn(0); 5126 } 5127 5128 /* 5129 MatGetBrowsOfAoCols_MPIAIJ - Creates a SeqAIJ matrix by taking rows of B that equal to nonzero columns 5130 of the OFF-DIAGONAL portion of local A 5131 5132 Collective on Mat 5133 5134 Input Parameters: 5135 + A,B - the matrices in mpiaij format 5136 - scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX 5137 5138 Output Parameter: 5139 + startsj_s - starting point in B's sending j-arrays, saved for MAT_REUSE (or NULL) 5140 . startsj_r - starting point in B's receiving j-arrays, saved for MAT_REUSE (or NULL) 5141 . bufa_ptr - array for sending matrix values, saved for MAT_REUSE (or NULL) 5142 - B_oth - the sequential matrix generated with size aBn=a->B->cmap->n by B->cmap->N 5143 5144 Level: developer 5145 5146 */ 5147 PetscErrorCode MatGetBrowsOfAoCols_MPIAIJ(Mat A,Mat B,MatReuse scall,PetscInt **startsj_s,PetscInt **startsj_r,MatScalar **bufa_ptr,Mat *B_oth) 5148 { 5149 VecScatter_MPI_General *gen_to,*gen_from; 5150 PetscErrorCode ierr; 5151 Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data; 5152 Mat_SeqAIJ *b_oth; 5153 VecScatter ctx; 5154 MPI_Comm comm; 5155 PetscMPIInt *rprocs,*sprocs,tag,rank; 5156 PetscInt *rowlen,*bufj,*bufJ,ncols,aBn=a->B->cmap->n,row,*b_othi,*b_othj; 5157 PetscInt *rvalues,*svalues,*cols,sbs,rbs; 5158 PetscScalar *b_otha,*bufa,*bufA,*vals; 5159 PetscInt i,j,k,l,ll,nrecvs,nsends,nrows,*srow,*rstarts,*rstartsj = 0,*sstarts,*sstartsj,len; 5160 MPI_Request *rwaits = NULL,*swaits = NULL; 5161 MPI_Status *sstatus,rstatus; 5162 PetscMPIInt jj,size; 5163 VecScatterType type; 5164 PetscBool mpi1; 5165 5166 PetscFunctionBegin; 5167 ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 5168 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 5169 5170 if (A->cmap->rstart != B->rmap->rstart || A->cmap->rend != B->rmap->rend) { 5171 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); 5172 } 5173 ierr = PetscLogEventBegin(MAT_GetBrowsOfAocols,A,B,0,0);CHKERRQ(ierr); 5174 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 5175 5176 if (size == 1) { 5177 startsj_s = NULL; 5178 bufa_ptr = NULL; 5179 *B_oth = NULL; 5180 PetscFunctionReturn(0); 5181 } 5182 5183 ctx = a->Mvctx; 5184 ierr = VecScatterGetType(ctx,&type);CHKERRQ(ierr); 5185 ierr = PetscStrcmp(type,"mpi1",&mpi1);CHKERRQ(ierr); 5186 if (!mpi1) { 5187 /* a->Mvctx is not type MPI1 which is not implemented for Mat-Mat ops, 5188 thus create a->Mvctx_mpi1 */ 5189 if (!a->Mvctx_mpi1) { 5190 a->Mvctx_mpi1_flg = PETSC_TRUE; 5191 ierr = MatSetUpMultiply_MPIAIJ(A);CHKERRQ(ierr); 5192 } 5193 ctx = a->Mvctx_mpi1; 5194 } 5195 tag = ((PetscObject)ctx)->tag; 5196 5197 gen_to = (VecScatter_MPI_General*)ctx->todata; 5198 gen_from = (VecScatter_MPI_General*)ctx->fromdata; 5199 nrecvs = gen_from->n; 5200 nsends = gen_to->n; 5201 5202 ierr = PetscMalloc2(nrecvs,&rwaits,nsends,&swaits);CHKERRQ(ierr); 5203 srow = gen_to->indices; /* local row index to be sent */ 5204 sstarts = gen_to->starts; 5205 sprocs = gen_to->procs; 5206 sstatus = gen_to->sstatus; 5207 sbs = gen_to->bs; 5208 rstarts = gen_from->starts; 5209 rprocs = gen_from->procs; 5210 rbs = gen_from->bs; 5211 5212 if (!startsj_s || !bufa_ptr) scall = MAT_INITIAL_MATRIX; 5213 if (scall == MAT_INITIAL_MATRIX) { 5214 /* i-array */ 5215 /*---------*/ 5216 /* post receives */ 5217 ierr = PetscMalloc1(rbs*(rstarts[nrecvs] - rstarts[0]),&rvalues);CHKERRQ(ierr); 5218 for (i=0; i<nrecvs; i++) { 5219 rowlen = rvalues + rstarts[i]*rbs; 5220 nrows = (rstarts[i+1]-rstarts[i])*rbs; /* num of indices to be received */ 5221 ierr = MPI_Irecv(rowlen,nrows,MPIU_INT,rprocs[i],tag,comm,rwaits+i);CHKERRQ(ierr); 5222 } 5223 5224 /* pack the outgoing message */ 5225 ierr = PetscMalloc2(nsends+1,&sstartsj,nrecvs+1,&rstartsj);CHKERRQ(ierr); 5226 5227 sstartsj[0] = 0; 5228 rstartsj[0] = 0; 5229 len = 0; /* total length of j or a array to be sent */ 5230 k = 0; 5231 ierr = PetscMalloc1(sbs*(sstarts[nsends] - sstarts[0]),&svalues);CHKERRQ(ierr); 5232 for (i=0; i<nsends; i++) { 5233 rowlen = svalues + sstarts[i]*sbs; 5234 nrows = sstarts[i+1]-sstarts[i]; /* num of block rows */ 5235 for (j=0; j<nrows; j++) { 5236 row = srow[k] + B->rmap->range[rank]; /* global row idx */ 5237 for (l=0; l<sbs; l++) { 5238 ierr = MatGetRow_MPIAIJ(B,row+l,&ncols,NULL,NULL);CHKERRQ(ierr); /* rowlength */ 5239 5240 rowlen[j*sbs+l] = ncols; 5241 5242 len += ncols; 5243 ierr = MatRestoreRow_MPIAIJ(B,row+l,&ncols,NULL,NULL);CHKERRQ(ierr); 5244 } 5245 k++; 5246 } 5247 ierr = MPI_Isend(rowlen,nrows*sbs,MPIU_INT,sprocs[i],tag,comm,swaits+i);CHKERRQ(ierr); 5248 5249 sstartsj[i+1] = len; /* starting point of (i+1)-th outgoing msg in bufj and bufa */ 5250 } 5251 /* recvs and sends of i-array are completed */ 5252 i = nrecvs; 5253 while (i--) { 5254 ierr = MPI_Waitany(nrecvs,rwaits,&jj,&rstatus);CHKERRQ(ierr); 5255 } 5256 if (nsends) {ierr = MPI_Waitall(nsends,swaits,sstatus);CHKERRQ(ierr);} 5257 ierr = PetscFree(svalues);CHKERRQ(ierr); 5258 5259 /* allocate buffers for sending j and a arrays */ 5260 ierr = PetscMalloc1(len+1,&bufj);CHKERRQ(ierr); 5261 ierr = PetscMalloc1(len+1,&bufa);CHKERRQ(ierr); 5262 5263 /* create i-array of B_oth */ 5264 ierr = PetscMalloc1(aBn+2,&b_othi);CHKERRQ(ierr); 5265 5266 b_othi[0] = 0; 5267 len = 0; /* total length of j or a array to be received */ 5268 k = 0; 5269 for (i=0; i<nrecvs; i++) { 5270 rowlen = rvalues + rstarts[i]*rbs; 5271 nrows = rbs*(rstarts[i+1]-rstarts[i]); /* num of rows to be received */ 5272 for (j=0; j<nrows; j++) { 5273 b_othi[k+1] = b_othi[k] + rowlen[j]; 5274 ierr = PetscIntSumError(rowlen[j],len,&len);CHKERRQ(ierr); 5275 k++; 5276 } 5277 rstartsj[i+1] = len; /* starting point of (i+1)-th incoming msg in bufj and bufa */ 5278 } 5279 ierr = PetscFree(rvalues);CHKERRQ(ierr); 5280 5281 /* allocate space for j and a arrrays of B_oth */ 5282 ierr = PetscMalloc1(b_othi[aBn]+1,&b_othj);CHKERRQ(ierr); 5283 ierr = PetscMalloc1(b_othi[aBn]+1,&b_otha);CHKERRQ(ierr); 5284 5285 /* j-array */ 5286 /*---------*/ 5287 /* post receives of j-array */ 5288 for (i=0; i<nrecvs; i++) { 5289 nrows = rstartsj[i+1]-rstartsj[i]; /* length of the msg received */ 5290 ierr = MPI_Irecv(b_othj+rstartsj[i],nrows,MPIU_INT,rprocs[i],tag,comm,rwaits+i);CHKERRQ(ierr); 5291 } 5292 5293 /* pack the outgoing message j-array */ 5294 k = 0; 5295 for (i=0; i<nsends; i++) { 5296 nrows = sstarts[i+1]-sstarts[i]; /* num of block rows */ 5297 bufJ = bufj+sstartsj[i]; 5298 for (j=0; j<nrows; j++) { 5299 row = srow[k++] + B->rmap->range[rank]; /* global row idx */ 5300 for (ll=0; ll<sbs; ll++) { 5301 ierr = MatGetRow_MPIAIJ(B,row+ll,&ncols,&cols,NULL);CHKERRQ(ierr); 5302 for (l=0; l<ncols; l++) { 5303 *bufJ++ = cols[l]; 5304 } 5305 ierr = MatRestoreRow_MPIAIJ(B,row+ll,&ncols,&cols,NULL);CHKERRQ(ierr); 5306 } 5307 } 5308 ierr = MPI_Isend(bufj+sstartsj[i],sstartsj[i+1]-sstartsj[i],MPIU_INT,sprocs[i],tag,comm,swaits+i);CHKERRQ(ierr); 5309 } 5310 5311 /* recvs and sends of j-array are completed */ 5312 i = nrecvs; 5313 while (i--) { 5314 ierr = MPI_Waitany(nrecvs,rwaits,&jj,&rstatus);CHKERRQ(ierr); 5315 } 5316 if (nsends) {ierr = MPI_Waitall(nsends,swaits,sstatus);CHKERRQ(ierr);} 5317 } else if (scall == MAT_REUSE_MATRIX) { 5318 sstartsj = *startsj_s; 5319 rstartsj = *startsj_r; 5320 bufa = *bufa_ptr; 5321 b_oth = (Mat_SeqAIJ*)(*B_oth)->data; 5322 b_otha = b_oth->a; 5323 } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE, "Matrix P does not posses an object container"); 5324 5325 /* a-array */ 5326 /*---------*/ 5327 /* post receives of a-array */ 5328 for (i=0; i<nrecvs; i++) { 5329 nrows = rstartsj[i+1]-rstartsj[i]; /* length of the msg received */ 5330 ierr = MPI_Irecv(b_otha+rstartsj[i],nrows,MPIU_SCALAR,rprocs[i],tag,comm,rwaits+i);CHKERRQ(ierr); 5331 } 5332 5333 /* pack the outgoing message a-array */ 5334 k = 0; 5335 for (i=0; i<nsends; i++) { 5336 nrows = sstarts[i+1]-sstarts[i]; /* num of block rows */ 5337 bufA = bufa+sstartsj[i]; 5338 for (j=0; j<nrows; j++) { 5339 row = srow[k++] + B->rmap->range[rank]; /* global row idx */ 5340 for (ll=0; ll<sbs; ll++) { 5341 ierr = MatGetRow_MPIAIJ(B,row+ll,&ncols,NULL,&vals);CHKERRQ(ierr); 5342 for (l=0; l<ncols; l++) { 5343 *bufA++ = vals[l]; 5344 } 5345 ierr = MatRestoreRow_MPIAIJ(B,row+ll,&ncols,NULL,&vals);CHKERRQ(ierr); 5346 } 5347 } 5348 ierr = MPI_Isend(bufa+sstartsj[i],sstartsj[i+1]-sstartsj[i],MPIU_SCALAR,sprocs[i],tag,comm,swaits+i);CHKERRQ(ierr); 5349 } 5350 /* recvs and sends of a-array are completed */ 5351 i = nrecvs; 5352 while (i--) { 5353 ierr = MPI_Waitany(nrecvs,rwaits,&jj,&rstatus);CHKERRQ(ierr); 5354 } 5355 if (nsends) {ierr = MPI_Waitall(nsends,swaits,sstatus);CHKERRQ(ierr);} 5356 ierr = PetscFree2(rwaits,swaits);CHKERRQ(ierr); 5357 5358 if (scall == MAT_INITIAL_MATRIX) { 5359 /* put together the new matrix */ 5360 ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,aBn,B->cmap->N,b_othi,b_othj,b_otha,B_oth);CHKERRQ(ierr); 5361 5362 /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */ 5363 /* Since these are PETSc arrays, change flags to free them as necessary. */ 5364 b_oth = (Mat_SeqAIJ*)(*B_oth)->data; 5365 b_oth->free_a = PETSC_TRUE; 5366 b_oth->free_ij = PETSC_TRUE; 5367 b_oth->nonew = 0; 5368 5369 ierr = PetscFree(bufj);CHKERRQ(ierr); 5370 if (!startsj_s || !bufa_ptr) { 5371 ierr = PetscFree2(sstartsj,rstartsj);CHKERRQ(ierr); 5372 ierr = PetscFree(bufa_ptr);CHKERRQ(ierr); 5373 } else { 5374 *startsj_s = sstartsj; 5375 *startsj_r = rstartsj; 5376 *bufa_ptr = bufa; 5377 } 5378 } 5379 ierr = PetscLogEventEnd(MAT_GetBrowsOfAocols,A,B,0,0);CHKERRQ(ierr); 5380 PetscFunctionReturn(0); 5381 } 5382 5383 /*@C 5384 MatGetCommunicationStructs - Provides access to the communication structures used in matrix-vector multiplication. 5385 5386 Not Collective 5387 5388 Input Parameters: 5389 . A - The matrix in mpiaij format 5390 5391 Output Parameter: 5392 + lvec - The local vector holding off-process values from the argument to a matrix-vector product 5393 . colmap - A map from global column index to local index into lvec 5394 - multScatter - A scatter from the argument of a matrix-vector product to lvec 5395 5396 Level: developer 5397 5398 @*/ 5399 #if defined(PETSC_USE_CTABLE) 5400 PetscErrorCode MatGetCommunicationStructs(Mat A, Vec *lvec, PetscTable *colmap, VecScatter *multScatter) 5401 #else 5402 PetscErrorCode MatGetCommunicationStructs(Mat A, Vec *lvec, PetscInt *colmap[], VecScatter *multScatter) 5403 #endif 5404 { 5405 Mat_MPIAIJ *a; 5406 5407 PetscFunctionBegin; 5408 PetscValidHeaderSpecific(A, MAT_CLASSID, 1); 5409 PetscValidPointer(lvec, 2); 5410 PetscValidPointer(colmap, 3); 5411 PetscValidPointer(multScatter, 4); 5412 a = (Mat_MPIAIJ*) A->data; 5413 if (lvec) *lvec = a->lvec; 5414 if (colmap) *colmap = a->colmap; 5415 if (multScatter) *multScatter = a->Mvctx; 5416 PetscFunctionReturn(0); 5417 } 5418 5419 PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJCRL(Mat,MatType,MatReuse,Mat*); 5420 PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJPERM(Mat,MatType,MatReuse,Mat*); 5421 #if defined(PETSC_HAVE_MKL_SPARSE) 5422 PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJMKL(Mat,MatType,MatReuse,Mat*); 5423 #endif 5424 PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPISBAIJ(Mat,MatType,MatReuse,Mat*); 5425 #if defined(PETSC_HAVE_ELEMENTAL) 5426 PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_Elemental(Mat,MatType,MatReuse,Mat*); 5427 #endif 5428 #if defined(PETSC_HAVE_HYPRE) 5429 PETSC_INTERN PetscErrorCode MatConvert_AIJ_HYPRE(Mat,MatType,MatReuse,Mat*); 5430 PETSC_INTERN PetscErrorCode MatMatMatMult_Transpose_AIJ_AIJ(Mat,Mat,Mat,MatReuse,PetscReal,Mat*); 5431 #endif 5432 PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_IS(Mat,MatType,MatReuse,Mat*); 5433 PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPISELL(Mat,MatType,MatReuse,Mat*); 5434 5435 /* 5436 Computes (B'*A')' since computing B*A directly is untenable 5437 5438 n p p 5439 ( ) ( ) ( ) 5440 m ( A ) * n ( B ) = m ( C ) 5441 ( ) ( ) ( ) 5442 5443 */ 5444 PetscErrorCode MatMatMultNumeric_MPIDense_MPIAIJ(Mat A,Mat B,Mat C) 5445 { 5446 PetscErrorCode ierr; 5447 Mat At,Bt,Ct; 5448 5449 PetscFunctionBegin; 5450 ierr = MatTranspose(A,MAT_INITIAL_MATRIX,&At);CHKERRQ(ierr); 5451 ierr = MatTranspose(B,MAT_INITIAL_MATRIX,&Bt);CHKERRQ(ierr); 5452 ierr = MatMatMult(Bt,At,MAT_INITIAL_MATRIX,1.0,&Ct);CHKERRQ(ierr); 5453 ierr = MatDestroy(&At);CHKERRQ(ierr); 5454 ierr = MatDestroy(&Bt);CHKERRQ(ierr); 5455 ierr = MatTranspose(Ct,MAT_REUSE_MATRIX,&C);CHKERRQ(ierr); 5456 ierr = MatDestroy(&Ct);CHKERRQ(ierr); 5457 PetscFunctionReturn(0); 5458 } 5459 5460 PetscErrorCode MatMatMultSymbolic_MPIDense_MPIAIJ(Mat A,Mat B,PetscReal fill,Mat *C) 5461 { 5462 PetscErrorCode ierr; 5463 PetscInt m=A->rmap->n,n=B->cmap->n; 5464 Mat Cmat; 5465 5466 PetscFunctionBegin; 5467 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); 5468 ierr = MatCreate(PetscObjectComm((PetscObject)A),&Cmat);CHKERRQ(ierr); 5469 ierr = MatSetSizes(Cmat,m,n,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 5470 ierr = MatSetBlockSizesFromMats(Cmat,A,B);CHKERRQ(ierr); 5471 ierr = MatSetType(Cmat,MATMPIDENSE);CHKERRQ(ierr); 5472 ierr = MatMPIDenseSetPreallocation(Cmat,NULL);CHKERRQ(ierr); 5473 ierr = MatAssemblyBegin(Cmat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5474 ierr = MatAssemblyEnd(Cmat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5475 5476 Cmat->ops->matmultnumeric = MatMatMultNumeric_MPIDense_MPIAIJ; 5477 5478 *C = Cmat; 5479 PetscFunctionReturn(0); 5480 } 5481 5482 /* ----------------------------------------------------------------*/ 5483 PETSC_INTERN PetscErrorCode MatMatMult_MPIDense_MPIAIJ(Mat A,Mat B,MatReuse scall,PetscReal fill,Mat *C) 5484 { 5485 PetscErrorCode ierr; 5486 5487 PetscFunctionBegin; 5488 if (scall == MAT_INITIAL_MATRIX) { 5489 ierr = PetscLogEventBegin(MAT_MatMultSymbolic,A,B,0,0);CHKERRQ(ierr); 5490 ierr = MatMatMultSymbolic_MPIDense_MPIAIJ(A,B,fill,C);CHKERRQ(ierr); 5491 ierr = PetscLogEventEnd(MAT_MatMultSymbolic,A,B,0,0);CHKERRQ(ierr); 5492 } 5493 ierr = PetscLogEventBegin(MAT_MatMultNumeric,A,B,0,0);CHKERRQ(ierr); 5494 ierr = MatMatMultNumeric_MPIDense_MPIAIJ(A,B,*C);CHKERRQ(ierr); 5495 ierr = PetscLogEventEnd(MAT_MatMultNumeric,A,B,0,0);CHKERRQ(ierr); 5496 PetscFunctionReturn(0); 5497 } 5498 5499 /*MC 5500 MATMPIAIJ - MATMPIAIJ = "mpiaij" - A matrix type to be used for parallel sparse matrices. 5501 5502 Options Database Keys: 5503 . -mat_type mpiaij - sets the matrix type to "mpiaij" during a call to MatSetFromOptions() 5504 5505 Level: beginner 5506 5507 .seealso: MatCreateAIJ() 5508 M*/ 5509 5510 PETSC_EXTERN PetscErrorCode MatCreate_MPIAIJ(Mat B) 5511 { 5512 Mat_MPIAIJ *b; 5513 PetscErrorCode ierr; 5514 PetscMPIInt size; 5515 5516 PetscFunctionBegin; 5517 ierr = MPI_Comm_size(PetscObjectComm((PetscObject)B),&size);CHKERRQ(ierr); 5518 5519 ierr = PetscNewLog(B,&b);CHKERRQ(ierr); 5520 B->data = (void*)b; 5521 ierr = PetscMemcpy(B->ops,&MatOps_Values,sizeof(struct _MatOps));CHKERRQ(ierr); 5522 B->assembled = PETSC_FALSE; 5523 B->insertmode = NOT_SET_VALUES; 5524 b->size = size; 5525 5526 ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)B),&b->rank);CHKERRQ(ierr); 5527 5528 /* build cache for off array entries formed */ 5529 ierr = MatStashCreate_Private(PetscObjectComm((PetscObject)B),1,&B->stash);CHKERRQ(ierr); 5530 5531 b->donotstash = PETSC_FALSE; 5532 b->colmap = 0; 5533 b->garray = 0; 5534 b->roworiented = PETSC_TRUE; 5535 5536 /* stuff used for matrix vector multiply */ 5537 b->lvec = NULL; 5538 b->Mvctx = NULL; 5539 5540 /* stuff for MatGetRow() */ 5541 b->rowindices = 0; 5542 b->rowvalues = 0; 5543 b->getrowactive = PETSC_FALSE; 5544 5545 /* flexible pointer used in CUSP/CUSPARSE classes */ 5546 b->spptr = NULL; 5547 5548 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMPIAIJSetUseScalableIncreaseOverlap_C",MatMPIAIJSetUseScalableIncreaseOverlap_MPIAIJ);CHKERRQ(ierr); 5549 ierr = PetscObjectComposeFunction((PetscObject)B,"MatStoreValues_C",MatStoreValues_MPIAIJ);CHKERRQ(ierr); 5550 ierr = PetscObjectComposeFunction((PetscObject)B,"MatRetrieveValues_C",MatRetrieveValues_MPIAIJ);CHKERRQ(ierr); 5551 ierr = PetscObjectComposeFunction((PetscObject)B,"MatIsTranspose_C",MatIsTranspose_MPIAIJ);CHKERRQ(ierr); 5552 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMPIAIJSetPreallocation_C",MatMPIAIJSetPreallocation_MPIAIJ);CHKERRQ(ierr); 5553 ierr = PetscObjectComposeFunction((PetscObject)B,"MatResetPreallocation_C",MatResetPreallocation_MPIAIJ);CHKERRQ(ierr); 5554 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMPIAIJSetPreallocationCSR_C",MatMPIAIJSetPreallocationCSR_MPIAIJ);CHKERRQ(ierr); 5555 ierr = PetscObjectComposeFunction((PetscObject)B,"MatDiagonalScaleLocal_C",MatDiagonalScaleLocal_MPIAIJ);CHKERRQ(ierr); 5556 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_mpiaijperm_C",MatConvert_MPIAIJ_MPIAIJPERM);CHKERRQ(ierr); 5557 #if defined(PETSC_HAVE_MKL_SPARSE) 5558 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_mpiaijmkl_C",MatConvert_MPIAIJ_MPIAIJMKL);CHKERRQ(ierr); 5559 #endif 5560 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_mpiaijcrl_C",MatConvert_MPIAIJ_MPIAIJCRL);CHKERRQ(ierr); 5561 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_mpisbaij_C",MatConvert_MPIAIJ_MPISBAIJ);CHKERRQ(ierr); 5562 #if defined(PETSC_HAVE_ELEMENTAL) 5563 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_elemental_C",MatConvert_MPIAIJ_Elemental);CHKERRQ(ierr); 5564 #endif 5565 #if defined(PETSC_HAVE_HYPRE) 5566 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_hypre_C",MatConvert_AIJ_HYPRE);CHKERRQ(ierr); 5567 #endif 5568 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_is_C",MatConvert_MPIAIJ_IS);CHKERRQ(ierr); 5569 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_mpisell_C",MatConvert_MPIAIJ_MPISELL);CHKERRQ(ierr); 5570 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMatMult_mpidense_mpiaij_C",MatMatMult_MPIDense_MPIAIJ);CHKERRQ(ierr); 5571 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMatMultSymbolic_mpidense_mpiaij_C",MatMatMultSymbolic_MPIDense_MPIAIJ);CHKERRQ(ierr); 5572 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMatMultNumeric_mpidense_mpiaij_C",MatMatMultNumeric_MPIDense_MPIAIJ);CHKERRQ(ierr); 5573 #if defined(PETSC_HAVE_HYPRE) 5574 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMatMatMult_transpose_mpiaij_mpiaij_C",MatMatMatMult_Transpose_AIJ_AIJ);CHKERRQ(ierr); 5575 #endif 5576 ierr = PetscObjectChangeTypeName((PetscObject)B,MATMPIAIJ);CHKERRQ(ierr); 5577 PetscFunctionReturn(0); 5578 } 5579 5580 /*@C 5581 MatCreateMPIAIJWithSplitArrays - creates a MPI AIJ matrix using arrays that contain the "diagonal" 5582 and "off-diagonal" part of the matrix in CSR format. 5583 5584 Collective on MPI_Comm 5585 5586 Input Parameters: 5587 + comm - MPI communicator 5588 . m - number of local rows (Cannot be PETSC_DECIDE) 5589 . n - This value should be the same as the local size used in creating the 5590 x vector for the matrix-vector product y = Ax. (or PETSC_DECIDE to have 5591 calculated if N is given) For square matrices n is almost always m. 5592 . M - number of global rows (or PETSC_DETERMINE to have calculated if m is given) 5593 . N - number of global columns (or PETSC_DETERMINE to have calculated if n is given) 5594 . i - row indices for "diagonal" portion of matrix 5595 . j - column indices 5596 . a - matrix values 5597 . oi - row indices for "off-diagonal" portion of matrix 5598 . oj - column indices 5599 - oa - matrix values 5600 5601 Output Parameter: 5602 . mat - the matrix 5603 5604 Level: advanced 5605 5606 Notes: 5607 The i, j, and a arrays ARE NOT copied by this routine into the internal format used by PETSc. The user 5608 must free the arrays once the matrix has been destroyed and not before. 5609 5610 The i and j indices are 0 based 5611 5612 See MatCreateAIJ() for the definition of "diagonal" and "off-diagonal" portion of the matrix 5613 5614 This sets local rows and cannot be used to set off-processor values. 5615 5616 Use of this routine is discouraged because it is inflexible and cumbersome to use. It is extremely rare that a 5617 legacy application natively assembles into exactly this split format. The code to do so is nontrivial and does 5618 not easily support in-place reassembly. It is recommended to use MatSetValues() (or a variant thereof) because 5619 the resulting assembly is easier to implement, will work with any matrix format, and the user does not have to 5620 keep track of the underlying array. Use MatSetOption(A,MAT_NO_OFF_PROC_ENTRIES,PETSC_TRUE) to disable all 5621 communication if it is known that only local entries will be set. 5622 5623 .keywords: matrix, aij, compressed row, sparse, parallel 5624 5625 .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatMPIAIJSetPreallocationCSR(), 5626 MATMPIAIJ, MatCreateAIJ(), MatCreateMPIAIJWithArrays() 5627 @*/ 5628 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) 5629 { 5630 PetscErrorCode ierr; 5631 Mat_MPIAIJ *maij; 5632 5633 PetscFunctionBegin; 5634 if (m < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"local number of rows (m) cannot be PETSC_DECIDE, or negative"); 5635 if (i[0]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"i (row indices) must start with 0"); 5636 if (oi[0]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"oi (row indices) must start with 0"); 5637 ierr = MatCreate(comm,mat);CHKERRQ(ierr); 5638 ierr = MatSetSizes(*mat,m,n,M,N);CHKERRQ(ierr); 5639 ierr = MatSetType(*mat,MATMPIAIJ);CHKERRQ(ierr); 5640 maij = (Mat_MPIAIJ*) (*mat)->data; 5641 5642 (*mat)->preallocated = PETSC_TRUE; 5643 5644 ierr = PetscLayoutSetUp((*mat)->rmap);CHKERRQ(ierr); 5645 ierr = PetscLayoutSetUp((*mat)->cmap);CHKERRQ(ierr); 5646 5647 ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,m,n,i,j,a,&maij->A);CHKERRQ(ierr); 5648 ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,m,(*mat)->cmap->N,oi,oj,oa,&maij->B);CHKERRQ(ierr); 5649 5650 ierr = MatAssemblyBegin(maij->A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5651 ierr = MatAssemblyEnd(maij->A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5652 ierr = MatAssemblyBegin(maij->B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5653 ierr = MatAssemblyEnd(maij->B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5654 5655 ierr = MatSetOption(*mat,MAT_NO_OFF_PROC_ENTRIES,PETSC_TRUE);CHKERRQ(ierr); 5656 ierr = MatAssemblyBegin(*mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5657 ierr = MatAssemblyEnd(*mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5658 ierr = MatSetOption(*mat,MAT_NO_OFF_PROC_ENTRIES,PETSC_FALSE);CHKERRQ(ierr); 5659 ierr = MatSetOption(*mat,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE);CHKERRQ(ierr); 5660 PetscFunctionReturn(0); 5661 } 5662 5663 /* 5664 Special version for direct calls from Fortran 5665 */ 5666 #include <petsc/private/fortranimpl.h> 5667 5668 /* Change these macros so can be used in void function */ 5669 #undef CHKERRQ 5670 #define CHKERRQ(ierr) CHKERRABORT(PETSC_COMM_WORLD,ierr) 5671 #undef SETERRQ2 5672 #define SETERRQ2(comm,ierr,b,c,d) CHKERRABORT(comm,ierr) 5673 #undef SETERRQ3 5674 #define SETERRQ3(comm,ierr,b,c,d,e) CHKERRABORT(comm,ierr) 5675 #undef SETERRQ 5676 #define SETERRQ(c,ierr,b) CHKERRABORT(c,ierr) 5677 5678 #if defined(PETSC_HAVE_FORTRAN_CAPS) 5679 #define matsetvaluesmpiaij_ MATSETVALUESMPIAIJ 5680 #elif !defined(PETSC_HAVE_FORTRAN_UNDERSCORE) 5681 #define matsetvaluesmpiaij_ matsetvaluesmpiaij 5682 #else 5683 #endif 5684 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) 5685 { 5686 Mat mat = *mmat; 5687 PetscInt m = *mm, n = *mn; 5688 InsertMode addv = *maddv; 5689 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 5690 PetscScalar value; 5691 PetscErrorCode ierr; 5692 5693 MatCheckPreallocated(mat,1); 5694 if (mat->insertmode == NOT_SET_VALUES) mat->insertmode = addv; 5695 5696 #if defined(PETSC_USE_DEBUG) 5697 else if (mat->insertmode != addv) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Cannot mix add values and insert values"); 5698 #endif 5699 { 5700 PetscInt i,j,rstart = mat->rmap->rstart,rend = mat->rmap->rend; 5701 PetscInt cstart = mat->cmap->rstart,cend = mat->cmap->rend,row,col; 5702 PetscBool roworiented = aij->roworiented; 5703 5704 /* Some Variables required in the macro */ 5705 Mat A = aij->A; 5706 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 5707 PetscInt *aimax = a->imax,*ai = a->i,*ailen = a->ilen,*aj = a->j; 5708 MatScalar *aa = a->a; 5709 PetscBool ignorezeroentries = (((a->ignorezeroentries)&&(addv==ADD_VALUES)) ? PETSC_TRUE : PETSC_FALSE); 5710 Mat B = aij->B; 5711 Mat_SeqAIJ *b = (Mat_SeqAIJ*)B->data; 5712 PetscInt *bimax = b->imax,*bi = b->i,*bilen = b->ilen,*bj = b->j,bm = aij->B->rmap->n,am = aij->A->rmap->n; 5713 MatScalar *ba = b->a; 5714 5715 PetscInt *rp1,*rp2,ii,nrow1,nrow2,_i,rmax1,rmax2,N,low1,high1,low2,high2,t,lastcol1,lastcol2; 5716 PetscInt nonew = a->nonew; 5717 MatScalar *ap1,*ap2; 5718 5719 PetscFunctionBegin; 5720 for (i=0; i<m; i++) { 5721 if (im[i] < 0) continue; 5722 #if defined(PETSC_USE_DEBUG) 5723 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); 5724 #endif 5725 if (im[i] >= rstart && im[i] < rend) { 5726 row = im[i] - rstart; 5727 lastcol1 = -1; 5728 rp1 = aj + ai[row]; 5729 ap1 = aa + ai[row]; 5730 rmax1 = aimax[row]; 5731 nrow1 = ailen[row]; 5732 low1 = 0; 5733 high1 = nrow1; 5734 lastcol2 = -1; 5735 rp2 = bj + bi[row]; 5736 ap2 = ba + bi[row]; 5737 rmax2 = bimax[row]; 5738 nrow2 = bilen[row]; 5739 low2 = 0; 5740 high2 = nrow2; 5741 5742 for (j=0; j<n; j++) { 5743 if (roworiented) value = v[i*n+j]; 5744 else value = v[i+j*m]; 5745 if (in[j] >= cstart && in[j] < cend) { 5746 col = in[j] - cstart; 5747 if (ignorezeroentries && value == 0.0 && (addv == ADD_VALUES) && row != col) continue; 5748 MatSetValues_SeqAIJ_A_Private(row,col,value,addv,im[i],in[j]); 5749 } else if (in[j] < 0) continue; 5750 #if defined(PETSC_USE_DEBUG) 5751 /* extra brace on SETERRQ2() is required for --with-errorchecking=0 - due to the next 'else' clause */ 5752 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);} 5753 #endif 5754 else { 5755 if (mat->was_assembled) { 5756 if (!aij->colmap) { 5757 ierr = MatCreateColmap_MPIAIJ_Private(mat);CHKERRQ(ierr); 5758 } 5759 #if defined(PETSC_USE_CTABLE) 5760 ierr = PetscTableFind(aij->colmap,in[j]+1,&col);CHKERRQ(ierr); 5761 col--; 5762 #else 5763 col = aij->colmap[in[j]] - 1; 5764 #endif 5765 if (ignorezeroentries && value == 0.0 && (addv == ADD_VALUES) && row != col) continue; 5766 if (col < 0 && !((Mat_SeqAIJ*)(aij->A->data))->nonew) { 5767 ierr = MatDisAssemble_MPIAIJ(mat);CHKERRQ(ierr); 5768 col = in[j]; 5769 /* Reinitialize the variables required by MatSetValues_SeqAIJ_B_Private() */ 5770 B = aij->B; 5771 b = (Mat_SeqAIJ*)B->data; 5772 bimax = b->imax; bi = b->i; bilen = b->ilen; bj = b->j; 5773 rp2 = bj + bi[row]; 5774 ap2 = ba + bi[row]; 5775 rmax2 = bimax[row]; 5776 nrow2 = bilen[row]; 5777 low2 = 0; 5778 high2 = nrow2; 5779 bm = aij->B->rmap->n; 5780 ba = b->a; 5781 } 5782 } else col = in[j]; 5783 MatSetValues_SeqAIJ_B_Private(row,col,value,addv,im[i],in[j]); 5784 } 5785 } 5786 } else if (!aij->donotstash) { 5787 if (roworiented) { 5788 ierr = MatStashValuesRow_Private(&mat->stash,im[i],n,in,v+i*n,(PetscBool)(ignorezeroentries && (addv == ADD_VALUES)));CHKERRQ(ierr); 5789 } else { 5790 ierr = MatStashValuesCol_Private(&mat->stash,im[i],n,in,v+i,m,(PetscBool)(ignorezeroentries && (addv == ADD_VALUES)));CHKERRQ(ierr); 5791 } 5792 } 5793 } 5794 } 5795 PetscFunctionReturnVoid(); 5796 } 5797 5798