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: 22 Subclasses include MATAIJCUSP, MATAIJCUSPARSE, MATAIJPERM, MATAIJSELL, MATAIJMKL, MATAIJCRL, and also automatically switches over to use inodes when 23 enough exist. 24 25 Level: beginner 26 27 .seealso: MatCreateAIJ(), MatCreateSeqAIJ(), MATSEQAIJ, MATMPIAIJ 28 M*/ 29 30 /*MC 31 MATAIJCRL - MATAIJCRL = "aijcrl" - A matrix type to be used for sparse matrices. 32 33 This matrix type is identical to MATSEQAIJCRL when constructed with a single process communicator, 34 and MATMPIAIJCRL otherwise. As a result, for single process communicators, 35 MatSeqAIJSetPreallocation() is supported, and similarly MatMPIAIJSetPreallocation() is supported 36 for communicators controlling multiple processes. It is recommended that you call both of 37 the above preallocation routines for simplicity. 38 39 Options Database Keys: 40 . -mat_type aijcrl - sets the matrix type to "aijcrl" during a call to MatSetFromOptions() 41 42 Level: beginner 43 44 .seealso: MatCreateMPIAIJCRL,MATSEQAIJCRL,MATMPIAIJCRL, MATSEQAIJCRL, MATMPIAIJCRL 45 M*/ 46 47 PetscErrorCode MatSetBlockSizes_MPIAIJ(Mat M, PetscInt rbs, PetscInt cbs) 48 { 49 PetscErrorCode ierr; 50 Mat_MPIAIJ *mat = (Mat_MPIAIJ*)M->data; 51 52 PetscFunctionBegin; 53 if (mat->A) { 54 ierr = MatSetBlockSizes(mat->A,rbs,cbs);CHKERRQ(ierr); 55 ierr = MatSetBlockSizes(mat->B,rbs,1);CHKERRQ(ierr); 56 } 57 PetscFunctionReturn(0); 58 } 59 60 PetscErrorCode MatFindNonzeroRows_MPIAIJ(Mat M,IS *keptrows) 61 { 62 PetscErrorCode ierr; 63 Mat_MPIAIJ *mat = (Mat_MPIAIJ*)M->data; 64 Mat_SeqAIJ *a = (Mat_SeqAIJ*)mat->A->data; 65 Mat_SeqAIJ *b = (Mat_SeqAIJ*)mat->B->data; 66 const PetscInt *ia,*ib; 67 const MatScalar *aa,*bb; 68 PetscInt na,nb,i,j,*rows,cnt=0,n0rows; 69 PetscInt m = M->rmap->n,rstart = M->rmap->rstart; 70 71 PetscFunctionBegin; 72 *keptrows = 0; 73 ia = a->i; 74 ib = b->i; 75 for (i=0; i<m; i++) { 76 na = ia[i+1] - ia[i]; 77 nb = ib[i+1] - ib[i]; 78 if (!na && !nb) { 79 cnt++; 80 goto ok1; 81 } 82 aa = a->a + ia[i]; 83 for (j=0; j<na; j++) { 84 if (aa[j] != 0.0) goto ok1; 85 } 86 bb = b->a + ib[i]; 87 for (j=0; j <nb; j++) { 88 if (bb[j] != 0.0) goto ok1; 89 } 90 cnt++; 91 ok1:; 92 } 93 ierr = MPIU_Allreduce(&cnt,&n0rows,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)M));CHKERRQ(ierr); 94 if (!n0rows) PetscFunctionReturn(0); 95 ierr = PetscMalloc1(M->rmap->n-cnt,&rows);CHKERRQ(ierr); 96 cnt = 0; 97 for (i=0; i<m; i++) { 98 na = ia[i+1] - ia[i]; 99 nb = ib[i+1] - ib[i]; 100 if (!na && !nb) continue; 101 aa = a->a + ia[i]; 102 for (j=0; j<na;j++) { 103 if (aa[j] != 0.0) { 104 rows[cnt++] = rstart + i; 105 goto ok2; 106 } 107 } 108 bb = b->a + ib[i]; 109 for (j=0; j<nb; j++) { 110 if (bb[j] != 0.0) { 111 rows[cnt++] = rstart + i; 112 goto ok2; 113 } 114 } 115 ok2:; 116 } 117 ierr = ISCreateGeneral(PetscObjectComm((PetscObject)M),cnt,rows,PETSC_OWN_POINTER,keptrows);CHKERRQ(ierr); 118 PetscFunctionReturn(0); 119 } 120 121 PetscErrorCode MatDiagonalSet_MPIAIJ(Mat Y,Vec D,InsertMode is) 122 { 123 PetscErrorCode ierr; 124 Mat_MPIAIJ *aij = (Mat_MPIAIJ*) Y->data; 125 PetscBool cong; 126 127 PetscFunctionBegin; 128 ierr = MatHasCongruentLayouts(Y,&cong);CHKERRQ(ierr); 129 if (Y->assembled && cong) { 130 ierr = MatDiagonalSet(aij->A,D,is);CHKERRQ(ierr); 131 } else { 132 ierr = MatDiagonalSet_Default(Y,D,is);CHKERRQ(ierr); 133 } 134 PetscFunctionReturn(0); 135 } 136 137 PetscErrorCode MatFindZeroDiagonals_MPIAIJ(Mat M,IS *zrows) 138 { 139 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)M->data; 140 PetscErrorCode ierr; 141 PetscInt i,rstart,nrows,*rows; 142 143 PetscFunctionBegin; 144 *zrows = NULL; 145 ierr = MatFindZeroDiagonals_SeqAIJ_Private(aij->A,&nrows,&rows);CHKERRQ(ierr); 146 ierr = MatGetOwnershipRange(M,&rstart,NULL);CHKERRQ(ierr); 147 for (i=0; i<nrows; i++) rows[i] += rstart; 148 ierr = ISCreateGeneral(PetscObjectComm((PetscObject)M),nrows,rows,PETSC_OWN_POINTER,zrows);CHKERRQ(ierr); 149 PetscFunctionReturn(0); 150 } 151 152 PetscErrorCode MatGetColumnNorms_MPIAIJ(Mat A,NormType type,PetscReal *norms) 153 { 154 PetscErrorCode ierr; 155 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)A->data; 156 PetscInt i,n,*garray = aij->garray; 157 Mat_SeqAIJ *a_aij = (Mat_SeqAIJ*) aij->A->data; 158 Mat_SeqAIJ *b_aij = (Mat_SeqAIJ*) aij->B->data; 159 PetscReal *work; 160 161 PetscFunctionBegin; 162 ierr = MatGetSize(A,NULL,&n);CHKERRQ(ierr); 163 ierr = PetscCalloc1(n,&work);CHKERRQ(ierr); 164 if (type == NORM_2) { 165 for (i=0; i<a_aij->i[aij->A->rmap->n]; i++) { 166 work[A->cmap->rstart + a_aij->j[i]] += PetscAbsScalar(a_aij->a[i]*a_aij->a[i]); 167 } 168 for (i=0; i<b_aij->i[aij->B->rmap->n]; i++) { 169 work[garray[b_aij->j[i]]] += PetscAbsScalar(b_aij->a[i]*b_aij->a[i]); 170 } 171 } else if (type == NORM_1) { 172 for (i=0; i<a_aij->i[aij->A->rmap->n]; i++) { 173 work[A->cmap->rstart + a_aij->j[i]] += PetscAbsScalar(a_aij->a[i]); 174 } 175 for (i=0; i<b_aij->i[aij->B->rmap->n]; i++) { 176 work[garray[b_aij->j[i]]] += PetscAbsScalar(b_aij->a[i]); 177 } 178 } else if (type == NORM_INFINITY) { 179 for (i=0; i<a_aij->i[aij->A->rmap->n]; i++) { 180 work[A->cmap->rstart + a_aij->j[i]] = PetscMax(PetscAbsScalar(a_aij->a[i]), work[A->cmap->rstart + a_aij->j[i]]); 181 } 182 for (i=0; i<b_aij->i[aij->B->rmap->n]; i++) { 183 work[garray[b_aij->j[i]]] = PetscMax(PetscAbsScalar(b_aij->a[i]),work[garray[b_aij->j[i]]]); 184 } 185 186 } else SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_WRONG,"Unknown NormType"); 187 if (type == NORM_INFINITY) { 188 ierr = MPIU_Allreduce(work,norms,n,MPIU_REAL,MPIU_MAX,PetscObjectComm((PetscObject)A));CHKERRQ(ierr); 189 } else { 190 ierr = MPIU_Allreduce(work,norms,n,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)A));CHKERRQ(ierr); 191 } 192 ierr = PetscFree(work);CHKERRQ(ierr); 193 if (type == NORM_2) { 194 for (i=0; i<n; i++) norms[i] = PetscSqrtReal(norms[i]); 195 } 196 PetscFunctionReturn(0); 197 } 198 199 PetscErrorCode MatFindOffBlockDiagonalEntries_MPIAIJ(Mat A,IS *is) 200 { 201 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 202 IS sis,gis; 203 PetscErrorCode ierr; 204 const PetscInt *isis,*igis; 205 PetscInt n,*iis,nsis,ngis,rstart,i; 206 207 PetscFunctionBegin; 208 ierr = MatFindOffBlockDiagonalEntries(a->A,&sis);CHKERRQ(ierr); 209 ierr = MatFindNonzeroRows(a->B,&gis);CHKERRQ(ierr); 210 ierr = ISGetSize(gis,&ngis);CHKERRQ(ierr); 211 ierr = ISGetSize(sis,&nsis);CHKERRQ(ierr); 212 ierr = ISGetIndices(sis,&isis);CHKERRQ(ierr); 213 ierr = ISGetIndices(gis,&igis);CHKERRQ(ierr); 214 215 ierr = PetscMalloc1(ngis+nsis,&iis);CHKERRQ(ierr); 216 ierr = PetscMemcpy(iis,igis,ngis*sizeof(PetscInt));CHKERRQ(ierr); 217 ierr = PetscMemcpy(iis+ngis,isis,nsis*sizeof(PetscInt));CHKERRQ(ierr); 218 n = ngis + nsis; 219 ierr = PetscSortRemoveDupsInt(&n,iis);CHKERRQ(ierr); 220 ierr = MatGetOwnershipRange(A,&rstart,NULL);CHKERRQ(ierr); 221 for (i=0; i<n; i++) iis[i] += rstart; 222 ierr = ISCreateGeneral(PetscObjectComm((PetscObject)A),n,iis,PETSC_OWN_POINTER,is);CHKERRQ(ierr); 223 224 ierr = ISRestoreIndices(sis,&isis);CHKERRQ(ierr); 225 ierr = ISRestoreIndices(gis,&igis);CHKERRQ(ierr); 226 ierr = ISDestroy(&sis);CHKERRQ(ierr); 227 ierr = ISDestroy(&gis);CHKERRQ(ierr); 228 PetscFunctionReturn(0); 229 } 230 231 /* 232 Distributes a SeqAIJ matrix across a set of processes. Code stolen from 233 MatLoad_MPIAIJ(). Horrible lack of reuse. Should be a routine for each matrix type. 234 235 Only for square matrices 236 237 Used by a preconditioner, hence PETSC_EXTERN 238 */ 239 PETSC_EXTERN PetscErrorCode MatDistribute_MPIAIJ(MPI_Comm comm,Mat gmat,PetscInt m,MatReuse reuse,Mat *inmat) 240 { 241 PetscMPIInt rank,size; 242 PetscInt *rowners,*dlens,*olens,i,rstart,rend,j,jj,nz = 0,*gmataj,cnt,row,*ld,bses[2]; 243 PetscErrorCode ierr; 244 Mat mat; 245 Mat_SeqAIJ *gmata; 246 PetscMPIInt tag; 247 MPI_Status status; 248 PetscBool aij; 249 MatScalar *gmataa,*ao,*ad,*gmataarestore=0; 250 251 PetscFunctionBegin; 252 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 253 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 254 if (!rank) { 255 ierr = PetscObjectTypeCompare((PetscObject)gmat,MATSEQAIJ,&aij);CHKERRQ(ierr); 256 if (!aij) SETERRQ1(PetscObjectComm((PetscObject)gmat),PETSC_ERR_SUP,"Currently no support for input matrix of type %s\n",((PetscObject)gmat)->type_name); 257 } 258 if (reuse == MAT_INITIAL_MATRIX) { 259 ierr = MatCreate(comm,&mat);CHKERRQ(ierr); 260 ierr = MatSetSizes(mat,m,m,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 261 ierr = MatGetBlockSizes(gmat,&bses[0],&bses[1]);CHKERRQ(ierr); 262 ierr = MPI_Bcast(bses,2,MPIU_INT,0,comm);CHKERRQ(ierr); 263 ierr = MatSetBlockSizes(mat,bses[0],bses[1]);CHKERRQ(ierr); 264 ierr = MatSetType(mat,MATAIJ);CHKERRQ(ierr); 265 ierr = PetscMalloc1(size+1,&rowners);CHKERRQ(ierr); 266 ierr = PetscMalloc2(m,&dlens,m,&olens);CHKERRQ(ierr); 267 ierr = MPI_Allgather(&m,1,MPIU_INT,rowners+1,1,MPIU_INT,comm);CHKERRQ(ierr); 268 269 rowners[0] = 0; 270 for (i=2; i<=size; i++) rowners[i] += rowners[i-1]; 271 rstart = rowners[rank]; 272 rend = rowners[rank+1]; 273 ierr = PetscObjectGetNewTag((PetscObject)mat,&tag);CHKERRQ(ierr); 274 if (!rank) { 275 gmata = (Mat_SeqAIJ*) gmat->data; 276 /* send row lengths to all processors */ 277 for (i=0; i<m; i++) dlens[i] = gmata->ilen[i]; 278 for (i=1; i<size; i++) { 279 ierr = MPI_Send(gmata->ilen + rowners[i],rowners[i+1]-rowners[i],MPIU_INT,i,tag,comm);CHKERRQ(ierr); 280 } 281 /* determine number diagonal and off-diagonal counts */ 282 ierr = PetscMemzero(olens,m*sizeof(PetscInt));CHKERRQ(ierr); 283 ierr = PetscCalloc1(m,&ld);CHKERRQ(ierr); 284 jj = 0; 285 for (i=0; i<m; i++) { 286 for (j=0; j<dlens[i]; j++) { 287 if (gmata->j[jj] < rstart) ld[i]++; 288 if (gmata->j[jj] < rstart || gmata->j[jj] >= rend) olens[i]++; 289 jj++; 290 } 291 } 292 /* send column indices to other processes */ 293 for (i=1; i<size; i++) { 294 nz = gmata->i[rowners[i+1]]-gmata->i[rowners[i]]; 295 ierr = MPI_Send(&nz,1,MPIU_INT,i,tag,comm);CHKERRQ(ierr); 296 ierr = MPI_Send(gmata->j + gmata->i[rowners[i]],nz,MPIU_INT,i,tag,comm);CHKERRQ(ierr); 297 } 298 299 /* send numerical values to other processes */ 300 for (i=1; i<size; i++) { 301 nz = gmata->i[rowners[i+1]]-gmata->i[rowners[i]]; 302 ierr = MPI_Send(gmata->a + gmata->i[rowners[i]],nz,MPIU_SCALAR,i,tag,comm);CHKERRQ(ierr); 303 } 304 gmataa = gmata->a; 305 gmataj = gmata->j; 306 307 } else { 308 /* receive row lengths */ 309 ierr = MPI_Recv(dlens,m,MPIU_INT,0,tag,comm,&status);CHKERRQ(ierr); 310 /* receive column indices */ 311 ierr = MPI_Recv(&nz,1,MPIU_INT,0,tag,comm,&status);CHKERRQ(ierr); 312 ierr = PetscMalloc2(nz,&gmataa,nz,&gmataj);CHKERRQ(ierr); 313 ierr = MPI_Recv(gmataj,nz,MPIU_INT,0,tag,comm,&status);CHKERRQ(ierr); 314 /* determine number diagonal and off-diagonal counts */ 315 ierr = PetscMemzero(olens,m*sizeof(PetscInt));CHKERRQ(ierr); 316 ierr = PetscCalloc1(m,&ld);CHKERRQ(ierr); 317 jj = 0; 318 for (i=0; i<m; i++) { 319 for (j=0; j<dlens[i]; j++) { 320 if (gmataj[jj] < rstart) ld[i]++; 321 if (gmataj[jj] < rstart || gmataj[jj] >= rend) olens[i]++; 322 jj++; 323 } 324 } 325 /* receive numerical values */ 326 ierr = PetscMemzero(gmataa,nz*sizeof(PetscScalar));CHKERRQ(ierr); 327 ierr = MPI_Recv(gmataa,nz,MPIU_SCALAR,0,tag,comm,&status);CHKERRQ(ierr); 328 } 329 /* set preallocation */ 330 for (i=0; i<m; i++) { 331 dlens[i] -= olens[i]; 332 } 333 ierr = MatSeqAIJSetPreallocation(mat,0,dlens);CHKERRQ(ierr); 334 ierr = MatMPIAIJSetPreallocation(mat,0,dlens,0,olens);CHKERRQ(ierr); 335 336 for (i=0; i<m; i++) { 337 dlens[i] += olens[i]; 338 } 339 cnt = 0; 340 for (i=0; i<m; i++) { 341 row = rstart + i; 342 ierr = MatSetValues(mat,1,&row,dlens[i],gmataj+cnt,gmataa+cnt,INSERT_VALUES);CHKERRQ(ierr); 343 cnt += dlens[i]; 344 } 345 if (rank) { 346 ierr = PetscFree2(gmataa,gmataj);CHKERRQ(ierr); 347 } 348 ierr = PetscFree2(dlens,olens);CHKERRQ(ierr); 349 ierr = PetscFree(rowners);CHKERRQ(ierr); 350 351 ((Mat_MPIAIJ*)(mat->data))->ld = ld; 352 353 *inmat = mat; 354 } else { /* column indices are already set; only need to move over numerical values from process 0 */ 355 Mat_SeqAIJ *Ad = (Mat_SeqAIJ*)((Mat_MPIAIJ*)((*inmat)->data))->A->data; 356 Mat_SeqAIJ *Ao = (Mat_SeqAIJ*)((Mat_MPIAIJ*)((*inmat)->data))->B->data; 357 mat = *inmat; 358 ierr = PetscObjectGetNewTag((PetscObject)mat,&tag);CHKERRQ(ierr); 359 if (!rank) { 360 /* send numerical values to other processes */ 361 gmata = (Mat_SeqAIJ*) gmat->data; 362 ierr = MatGetOwnershipRanges(mat,(const PetscInt**)&rowners);CHKERRQ(ierr); 363 gmataa = gmata->a; 364 for (i=1; i<size; i++) { 365 nz = gmata->i[rowners[i+1]]-gmata->i[rowners[i]]; 366 ierr = MPI_Send(gmataa + gmata->i[rowners[i]],nz,MPIU_SCALAR,i,tag,comm);CHKERRQ(ierr); 367 } 368 nz = gmata->i[rowners[1]]-gmata->i[rowners[0]]; 369 } else { 370 /* receive numerical values from process 0*/ 371 nz = Ad->nz + Ao->nz; 372 ierr = PetscMalloc1(nz,&gmataa);CHKERRQ(ierr); gmataarestore = gmataa; 373 ierr = MPI_Recv(gmataa,nz,MPIU_SCALAR,0,tag,comm,&status);CHKERRQ(ierr); 374 } 375 /* transfer numerical values into the diagonal A and off diagonal B parts of mat */ 376 ld = ((Mat_MPIAIJ*)(mat->data))->ld; 377 ad = Ad->a; 378 ao = Ao->a; 379 if (mat->rmap->n) { 380 i = 0; 381 nz = ld[i]; ierr = PetscMemcpy(ao,gmataa,nz*sizeof(PetscScalar));CHKERRQ(ierr); ao += nz; gmataa += nz; 382 nz = Ad->i[i+1] - Ad->i[i]; ierr = PetscMemcpy(ad,gmataa,nz*sizeof(PetscScalar));CHKERRQ(ierr); ad += nz; gmataa += nz; 383 } 384 for (i=1; i<mat->rmap->n; i++) { 385 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; 386 nz = Ad->i[i+1] - Ad->i[i]; ierr = PetscMemcpy(ad,gmataa,nz*sizeof(PetscScalar));CHKERRQ(ierr); ad += nz; gmataa += nz; 387 } 388 i--; 389 if (mat->rmap->n) { 390 nz = Ao->i[i+1] - Ao->i[i] - ld[i]; ierr = PetscMemcpy(ao,gmataa,nz*sizeof(PetscScalar));CHKERRQ(ierr); 391 } 392 if (rank) { 393 ierr = PetscFree(gmataarestore);CHKERRQ(ierr); 394 } 395 } 396 ierr = MatAssemblyBegin(mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 397 ierr = MatAssemblyEnd(mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 398 PetscFunctionReturn(0); 399 } 400 401 /* 402 Local utility routine that creates a mapping from the global column 403 number to the local number in the off-diagonal part of the local 404 storage of the matrix. When PETSC_USE_CTABLE is used this is scalable at 405 a slightly higher hash table cost; without it it is not scalable (each processor 406 has an order N integer array but is fast to acess. 407 */ 408 PetscErrorCode MatCreateColmap_MPIAIJ_Private(Mat mat) 409 { 410 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 411 PetscErrorCode ierr; 412 PetscInt n = aij->B->cmap->n,i; 413 414 PetscFunctionBegin; 415 if (!aij->garray) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"MPIAIJ Matrix was assembled but is missing garray"); 416 #if defined(PETSC_USE_CTABLE) 417 ierr = PetscTableCreate(n,mat->cmap->N+1,&aij->colmap);CHKERRQ(ierr); 418 for (i=0; i<n; i++) { 419 ierr = PetscTableAdd(aij->colmap,aij->garray[i]+1,i+1,INSERT_VALUES);CHKERRQ(ierr); 420 } 421 #else 422 ierr = PetscCalloc1(mat->cmap->N+1,&aij->colmap);CHKERRQ(ierr); 423 ierr = PetscLogObjectMemory((PetscObject)mat,(mat->cmap->N+1)*sizeof(PetscInt));CHKERRQ(ierr); 424 for (i=0; i<n; i++) aij->colmap[aij->garray[i]] = i+1; 425 #endif 426 PetscFunctionReturn(0); 427 } 428 429 #define MatSetValues_SeqAIJ_A_Private(row,col,value,addv,orow,ocol) \ 430 { \ 431 if (col <= lastcol1) low1 = 0; \ 432 else high1 = nrow1; \ 433 lastcol1 = col;\ 434 while (high1-low1 > 5) { \ 435 t = (low1+high1)/2; \ 436 if (rp1[t] > col) high1 = t; \ 437 else low1 = t; \ 438 } \ 439 for (_i=low1; _i<high1; _i++) { \ 440 if (rp1[_i] > col) break; \ 441 if (rp1[_i] == col) { \ 442 if (addv == ADD_VALUES) ap1[_i] += value; \ 443 else ap1[_i] = value; \ 444 goto a_noinsert; \ 445 } \ 446 } \ 447 if (value == 0.0 && ignorezeroentries && row != col) {low1 = 0; high1 = nrow1;goto a_noinsert;} \ 448 if (nonew == 1) {low1 = 0; high1 = nrow1; goto a_noinsert;} \ 449 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); \ 450 MatSeqXAIJReallocateAIJ(A,am,1,nrow1,row,col,rmax1,aa,ai,aj,rp1,ap1,aimax,nonew,MatScalar); \ 451 N = nrow1++ - 1; a->nz++; high1++; \ 452 /* shift up all the later entries in this row */ \ 453 for (ii=N; ii>=_i; ii--) { \ 454 rp1[ii+1] = rp1[ii]; \ 455 ap1[ii+1] = ap1[ii]; \ 456 } \ 457 rp1[_i] = col; \ 458 ap1[_i] = value; \ 459 A->nonzerostate++;\ 460 a_noinsert: ; \ 461 ailen[row] = nrow1; \ 462 } 463 464 #define MatSetValues_SeqAIJ_B_Private(row,col,value,addv,orow,ocol) \ 465 { \ 466 if (col <= lastcol2) low2 = 0; \ 467 else high2 = nrow2; \ 468 lastcol2 = col; \ 469 while (high2-low2 > 5) { \ 470 t = (low2+high2)/2; \ 471 if (rp2[t] > col) high2 = t; \ 472 else low2 = t; \ 473 } \ 474 for (_i=low2; _i<high2; _i++) { \ 475 if (rp2[_i] > col) break; \ 476 if (rp2[_i] == col) { \ 477 if (addv == ADD_VALUES) ap2[_i] += value; \ 478 else ap2[_i] = value; \ 479 goto b_noinsert; \ 480 } \ 481 } \ 482 if (value == 0.0 && ignorezeroentries) {low2 = 0; high2 = nrow2; goto b_noinsert;} \ 483 if (nonew == 1) {low2 = 0; high2 = nrow2; goto b_noinsert;} \ 484 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); \ 485 MatSeqXAIJReallocateAIJ(B,bm,1,nrow2,row,col,rmax2,ba,bi,bj,rp2,ap2,bimax,nonew,MatScalar); \ 486 N = nrow2++ - 1; b->nz++; high2++; \ 487 /* shift up all the later entries in this row */ \ 488 for (ii=N; ii>=_i; ii--) { \ 489 rp2[ii+1] = rp2[ii]; \ 490 ap2[ii+1] = ap2[ii]; \ 491 } \ 492 rp2[_i] = col; \ 493 ap2[_i] = value; \ 494 B->nonzerostate++; \ 495 b_noinsert: ; \ 496 bilen[row] = nrow2; \ 497 } 498 499 PetscErrorCode MatSetValuesRow_MPIAIJ(Mat A,PetscInt row,const PetscScalar v[]) 500 { 501 Mat_MPIAIJ *mat = (Mat_MPIAIJ*)A->data; 502 Mat_SeqAIJ *a = (Mat_SeqAIJ*)mat->A->data,*b = (Mat_SeqAIJ*)mat->B->data; 503 PetscErrorCode ierr; 504 PetscInt l,*garray = mat->garray,diag; 505 506 PetscFunctionBegin; 507 /* code only works for square matrices A */ 508 509 /* find size of row to the left of the diagonal part */ 510 ierr = MatGetOwnershipRange(A,&diag,0);CHKERRQ(ierr); 511 row = row - diag; 512 for (l=0; l<b->i[row+1]-b->i[row]; l++) { 513 if (garray[b->j[b->i[row]+l]] > diag) break; 514 } 515 ierr = PetscMemcpy(b->a+b->i[row],v,l*sizeof(PetscScalar));CHKERRQ(ierr); 516 517 /* diagonal part */ 518 ierr = PetscMemcpy(a->a+a->i[row],v+l,(a->i[row+1]-a->i[row])*sizeof(PetscScalar));CHKERRQ(ierr); 519 520 /* right of diagonal part */ 521 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); 522 PetscFunctionReturn(0); 523 } 524 525 PetscErrorCode MatSetValues_MPIAIJ(Mat mat,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],const PetscScalar v[],InsertMode addv) 526 { 527 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 528 PetscScalar value; 529 PetscErrorCode ierr; 530 PetscInt i,j,rstart = mat->rmap->rstart,rend = mat->rmap->rend; 531 PetscInt cstart = mat->cmap->rstart,cend = mat->cmap->rend,row,col; 532 PetscBool roworiented = aij->roworiented; 533 534 /* Some Variables required in the macro */ 535 Mat A = aij->A; 536 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 537 PetscInt *aimax = a->imax,*ai = a->i,*ailen = a->ilen,*aj = a->j; 538 MatScalar *aa = a->a; 539 PetscBool ignorezeroentries = a->ignorezeroentries; 540 Mat B = aij->B; 541 Mat_SeqAIJ *b = (Mat_SeqAIJ*)B->data; 542 PetscInt *bimax = b->imax,*bi = b->i,*bilen = b->ilen,*bj = b->j,bm = aij->B->rmap->n,am = aij->A->rmap->n; 543 MatScalar *ba = b->a; 544 545 PetscInt *rp1,*rp2,ii,nrow1,nrow2,_i,rmax1,rmax2,N,low1,high1,low2,high2,t,lastcol1,lastcol2; 546 PetscInt nonew; 547 MatScalar *ap1,*ap2; 548 549 PetscFunctionBegin; 550 for (i=0; i<m; i++) { 551 if (im[i] < 0) continue; 552 #if defined(PETSC_USE_DEBUG) 553 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); 554 #endif 555 if (im[i] >= rstart && im[i] < rend) { 556 row = im[i] - rstart; 557 lastcol1 = -1; 558 rp1 = aj + ai[row]; 559 ap1 = aa + ai[row]; 560 rmax1 = aimax[row]; 561 nrow1 = ailen[row]; 562 low1 = 0; 563 high1 = nrow1; 564 lastcol2 = -1; 565 rp2 = bj + bi[row]; 566 ap2 = ba + bi[row]; 567 rmax2 = bimax[row]; 568 nrow2 = bilen[row]; 569 low2 = 0; 570 high2 = nrow2; 571 572 for (j=0; j<n; j++) { 573 if (roworiented) value = v[i*n+j]; 574 else value = v[i+j*m]; 575 if (in[j] >= cstart && in[j] < cend) { 576 col = in[j] - cstart; 577 nonew = a->nonew; 578 if (ignorezeroentries && value == 0.0 && (addv == ADD_VALUES) && row != col) continue; 579 MatSetValues_SeqAIJ_A_Private(row,col,value,addv,im[i],in[j]); 580 } else if (in[j] < 0) continue; 581 #if defined(PETSC_USE_DEBUG) 582 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); 583 #endif 584 else { 585 if (mat->was_assembled) { 586 if (!aij->colmap) { 587 ierr = MatCreateColmap_MPIAIJ_Private(mat);CHKERRQ(ierr); 588 } 589 #if defined(PETSC_USE_CTABLE) 590 ierr = PetscTableFind(aij->colmap,in[j]+1,&col);CHKERRQ(ierr); 591 col--; 592 #else 593 col = aij->colmap[in[j]] - 1; 594 #endif 595 if (col < 0 && !((Mat_SeqAIJ*)(aij->B->data))->nonew) { 596 ierr = MatDisAssemble_MPIAIJ(mat);CHKERRQ(ierr); 597 col = in[j]; 598 /* Reinitialize the variables required by MatSetValues_SeqAIJ_B_Private() */ 599 B = aij->B; 600 b = (Mat_SeqAIJ*)B->data; 601 bimax = b->imax; bi = b->i; bilen = b->ilen; bj = b->j; ba = b->a; 602 rp2 = bj + bi[row]; 603 ap2 = ba + bi[row]; 604 rmax2 = bimax[row]; 605 nrow2 = bilen[row]; 606 low2 = 0; 607 high2 = nrow2; 608 bm = aij->B->rmap->n; 609 ba = b->a; 610 } else if (col < 0) { 611 if (1 == ((Mat_SeqAIJ*)(aij->B->data))->nonew) { 612 ierr = PetscInfo3(mat,"Skipping of insertion of new nonzero location in off-diagonal portion of matrix %g(%D,%D)\n",(double)PetscRealPart(value),im[i],in[j]);CHKERRQ(ierr); 613 } else SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero at global row/column (%D, %D) into matrix", im[i], in[j]); 614 } 615 } else col = in[j]; 616 nonew = b->nonew; 617 MatSetValues_SeqAIJ_B_Private(row,col,value,addv,im[i],in[j]); 618 } 619 } 620 } else { 621 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]); 622 if (!aij->donotstash) { 623 mat->assembled = PETSC_FALSE; 624 if (roworiented) { 625 ierr = MatStashValuesRow_Private(&mat->stash,im[i],n,in,v+i*n,(PetscBool)(ignorezeroentries && (addv == ADD_VALUES)));CHKERRQ(ierr); 626 } else { 627 ierr = MatStashValuesCol_Private(&mat->stash,im[i],n,in,v+i,m,(PetscBool)(ignorezeroentries && (addv == ADD_VALUES)));CHKERRQ(ierr); 628 } 629 } 630 } 631 } 632 PetscFunctionReturn(0); 633 } 634 635 /* 636 This function sets the j and ilen arrays (of the diagonal and off-diagonal part) of an MPIAIJ-matrix. 637 The values in mat_i have to be sorted and the values in mat_j have to be sorted for each row (CSR-like). 638 No off-processor parts off the matrix are allowed here and mat->was_assembled has to be PETSC_FALSE. 639 */ 640 PetscErrorCode MatSetValues_MPIAIJ_CopyFromCSRFormat_Symbolic(Mat mat,const PetscInt mat_j[],const PetscInt mat_i[]) 641 { 642 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 643 Mat A = aij->A; /* diagonal part of the matrix */ 644 Mat B = aij->B; /* offdiagonal part of the matrix */ 645 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 646 Mat_SeqAIJ *b = (Mat_SeqAIJ*)B->data; 647 PetscInt cstart = mat->cmap->rstart,cend = mat->cmap->rend,col; 648 PetscInt *ailen = a->ilen,*aj = a->j; 649 PetscInt *bilen = b->ilen,*bj = b->j; 650 PetscInt am = aij->A->rmap->n,j; 651 PetscInt diag_so_far = 0,dnz; 652 PetscInt offd_so_far = 0,onz; 653 654 PetscFunctionBegin; 655 /* Iterate over all rows of the matrix */ 656 for (j=0; j<am; j++) { 657 dnz = onz = 0; 658 /* Iterate over all non-zero columns of the current row */ 659 for (col=mat_i[j]; col<mat_i[j+1]; col++) { 660 /* If column is in the diagonal */ 661 if (mat_j[col] >= cstart && mat_j[col] < cend) { 662 aj[diag_so_far++] = mat_j[col] - cstart; 663 dnz++; 664 } else { /* off-diagonal entries */ 665 bj[offd_so_far++] = mat_j[col]; 666 onz++; 667 } 668 } 669 ailen[j] = dnz; 670 bilen[j] = onz; 671 } 672 PetscFunctionReturn(0); 673 } 674 675 /* 676 This function sets the local j, a and ilen arrays (of the diagonal and off-diagonal part) of an MPIAIJ-matrix. 677 The values in mat_i have to be sorted and the values in mat_j have to be sorted for each row (CSR-like). 678 No off-processor parts off the matrix are allowed here, they are set at a later point by MatSetValues_MPIAIJ. 679 Also, mat->was_assembled has to be false, otherwise the statement aj[rowstart_diag+dnz_row] = mat_j[col] - cstart; 680 would not be true and the more complex MatSetValues_MPIAIJ has to be used. 681 */ 682 PetscErrorCode MatSetValues_MPIAIJ_CopyFromCSRFormat(Mat mat,const PetscInt mat_j[],const PetscInt mat_i[],const PetscScalar mat_a[]) 683 { 684 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 685 Mat A = aij->A; /* diagonal part of the matrix */ 686 Mat B = aij->B; /* offdiagonal part of the matrix */ 687 Mat_SeqAIJ *aijd =(Mat_SeqAIJ*)(aij->A)->data,*aijo=(Mat_SeqAIJ*)(aij->B)->data; 688 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 689 Mat_SeqAIJ *b = (Mat_SeqAIJ*)B->data; 690 PetscInt cstart = mat->cmap->rstart,cend = mat->cmap->rend; 691 PetscInt *ailen = a->ilen,*aj = a->j; 692 PetscInt *bilen = b->ilen,*bj = b->j; 693 PetscInt am = aij->A->rmap->n,j; 694 PetscInt *full_diag_i=aijd->i,*full_offd_i=aijo->i; /* These variables can also include non-local elements, which are set at a later point. */ 695 PetscInt col,dnz_row,onz_row,rowstart_diag,rowstart_offd; 696 PetscScalar *aa = a->a,*ba = b->a; 697 698 PetscFunctionBegin; 699 /* Iterate over all rows of the matrix */ 700 for (j=0; j<am; j++) { 701 dnz_row = onz_row = 0; 702 rowstart_offd = full_offd_i[j]; 703 rowstart_diag = full_diag_i[j]; 704 /* Iterate over all non-zero columns of the current row */ 705 for (col=mat_i[j]; col<mat_i[j+1]; col++) { 706 /* If column is in the diagonal */ 707 if (mat_j[col] >= cstart && mat_j[col] < cend) { 708 aj[rowstart_diag+dnz_row] = mat_j[col] - cstart; 709 aa[rowstart_diag+dnz_row] = mat_a[col]; 710 dnz_row++; 711 } else { /* off-diagonal entries */ 712 bj[rowstart_offd+onz_row] = mat_j[col]; 713 ba[rowstart_offd+onz_row] = mat_a[col]; 714 onz_row++; 715 } 716 } 717 ailen[j] = dnz_row; 718 bilen[j] = onz_row; 719 } 720 PetscFunctionReturn(0); 721 } 722 723 PetscErrorCode MatGetValues_MPIAIJ(Mat mat,PetscInt m,const PetscInt idxm[],PetscInt n,const PetscInt idxn[],PetscScalar v[]) 724 { 725 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 726 PetscErrorCode ierr; 727 PetscInt i,j,rstart = mat->rmap->rstart,rend = mat->rmap->rend; 728 PetscInt cstart = mat->cmap->rstart,cend = mat->cmap->rend,row,col; 729 730 PetscFunctionBegin; 731 for (i=0; i<m; i++) { 732 if (idxm[i] < 0) continue; /* SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative row: %D",idxm[i]);*/ 733 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); 734 if (idxm[i] >= rstart && idxm[i] < rend) { 735 row = idxm[i] - rstart; 736 for (j=0; j<n; j++) { 737 if (idxn[j] < 0) continue; /* SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative column: %D",idxn[j]); */ 738 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); 739 if (idxn[j] >= cstart && idxn[j] < cend) { 740 col = idxn[j] - cstart; 741 ierr = MatGetValues(aij->A,1,&row,1,&col,v+i*n+j);CHKERRQ(ierr); 742 } else { 743 if (!aij->colmap) { 744 ierr = MatCreateColmap_MPIAIJ_Private(mat);CHKERRQ(ierr); 745 } 746 #if defined(PETSC_USE_CTABLE) 747 ierr = PetscTableFind(aij->colmap,idxn[j]+1,&col);CHKERRQ(ierr); 748 col--; 749 #else 750 col = aij->colmap[idxn[j]] - 1; 751 #endif 752 if ((col < 0) || (aij->garray[col] != idxn[j])) *(v+i*n+j) = 0.0; 753 else { 754 ierr = MatGetValues(aij->B,1,&row,1,&col,v+i*n+j);CHKERRQ(ierr); 755 } 756 } 757 } 758 } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Only local values currently supported"); 759 } 760 PetscFunctionReturn(0); 761 } 762 763 extern PetscErrorCode MatMultDiagonalBlock_MPIAIJ(Mat,Vec,Vec); 764 765 PetscErrorCode MatAssemblyBegin_MPIAIJ(Mat mat,MatAssemblyType mode) 766 { 767 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 768 PetscErrorCode ierr; 769 PetscInt nstash,reallocs; 770 771 PetscFunctionBegin; 772 if (aij->donotstash || mat->nooffprocentries) PetscFunctionReturn(0); 773 774 ierr = MatStashScatterBegin_Private(mat,&mat->stash,mat->rmap->range);CHKERRQ(ierr); 775 ierr = MatStashGetInfo_Private(&mat->stash,&nstash,&reallocs);CHKERRQ(ierr); 776 ierr = PetscInfo2(aij->A,"Stash has %D entries, uses %D mallocs.\n",nstash,reallocs);CHKERRQ(ierr); 777 PetscFunctionReturn(0); 778 } 779 780 PetscErrorCode MatAssemblyEnd_MPIAIJ(Mat mat,MatAssemblyType mode) 781 { 782 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 783 Mat_SeqAIJ *a = (Mat_SeqAIJ*)aij->A->data; 784 PetscErrorCode ierr; 785 PetscMPIInt n; 786 PetscInt i,j,rstart,ncols,flg; 787 PetscInt *row,*col; 788 PetscBool other_disassembled; 789 PetscScalar *val; 790 791 /* do not use 'b = (Mat_SeqAIJ*)aij->B->data' as B can be reset in disassembly */ 792 793 PetscFunctionBegin; 794 if (!aij->donotstash && !mat->nooffprocentries) { 795 while (1) { 796 ierr = MatStashScatterGetMesg_Private(&mat->stash,&n,&row,&col,&val,&flg);CHKERRQ(ierr); 797 if (!flg) break; 798 799 for (i=0; i<n; ) { 800 /* Now identify the consecutive vals belonging to the same row */ 801 for (j=i,rstart=row[j]; j<n; j++) { 802 if (row[j] != rstart) break; 803 } 804 if (j < n) ncols = j-i; 805 else ncols = n-i; 806 /* Now assemble all these values with a single function call */ 807 ierr = MatSetValues_MPIAIJ(mat,1,row+i,ncols,col+i,val+i,mat->insertmode);CHKERRQ(ierr); 808 809 i = j; 810 } 811 } 812 ierr = MatStashScatterEnd_Private(&mat->stash);CHKERRQ(ierr); 813 } 814 ierr = MatAssemblyBegin(aij->A,mode);CHKERRQ(ierr); 815 ierr = MatAssemblyEnd(aij->A,mode);CHKERRQ(ierr); 816 817 /* determine if any processor has disassembled, if so we must 818 also disassemble ourselfs, in order that we may reassemble. */ 819 /* 820 if nonzero structure of submatrix B cannot change then we know that 821 no processor disassembled thus we can skip this stuff 822 */ 823 if (!((Mat_SeqAIJ*)aij->B->data)->nonew) { 824 ierr = MPIU_Allreduce(&mat->was_assembled,&other_disassembled,1,MPIU_BOOL,MPI_PROD,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 825 if (mat->was_assembled && !other_disassembled) { 826 ierr = MatDisAssemble_MPIAIJ(mat);CHKERRQ(ierr); 827 } 828 } 829 if (!mat->was_assembled && mode == MAT_FINAL_ASSEMBLY) { 830 ierr = MatSetUpMultiply_MPIAIJ(mat);CHKERRQ(ierr); 831 } 832 ierr = MatSetOption(aij->B,MAT_USE_INODES,PETSC_FALSE);CHKERRQ(ierr); 833 ierr = MatAssemblyBegin(aij->B,mode);CHKERRQ(ierr); 834 ierr = MatAssemblyEnd(aij->B,mode);CHKERRQ(ierr); 835 836 ierr = PetscFree2(aij->rowvalues,aij->rowindices);CHKERRQ(ierr); 837 838 aij->rowvalues = 0; 839 840 ierr = VecDestroy(&aij->diag);CHKERRQ(ierr); 841 if (a->inode.size) mat->ops->multdiagonalblock = MatMultDiagonalBlock_MPIAIJ; 842 843 /* if no new nonzero locations are allowed in matrix then only set the matrix state the first time through */ 844 if ((!mat->was_assembled && mode == MAT_FINAL_ASSEMBLY) || !((Mat_SeqAIJ*)(aij->A->data))->nonew) { 845 PetscObjectState state = aij->A->nonzerostate + aij->B->nonzerostate; 846 ierr = MPIU_Allreduce(&state,&mat->nonzerostate,1,MPIU_INT64,MPI_SUM,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 847 } 848 PetscFunctionReturn(0); 849 } 850 851 PetscErrorCode MatZeroEntries_MPIAIJ(Mat A) 852 { 853 Mat_MPIAIJ *l = (Mat_MPIAIJ*)A->data; 854 PetscErrorCode ierr; 855 856 PetscFunctionBegin; 857 ierr = MatZeroEntries(l->A);CHKERRQ(ierr); 858 ierr = MatZeroEntries(l->B);CHKERRQ(ierr); 859 PetscFunctionReturn(0); 860 } 861 862 PetscErrorCode MatZeroRows_MPIAIJ(Mat A,PetscInt N,const PetscInt rows[],PetscScalar diag,Vec x,Vec b) 863 { 864 Mat_MPIAIJ *mat = (Mat_MPIAIJ *) A->data; 865 PetscInt *lrows; 866 PetscInt r, len; 867 PetscBool cong; 868 PetscErrorCode ierr; 869 870 PetscFunctionBegin; 871 /* get locally owned rows */ 872 ierr = MatZeroRowsMapLocal_Private(A,N,rows,&len,&lrows);CHKERRQ(ierr); 873 /* fix right hand side if needed */ 874 if (x && b) { 875 const PetscScalar *xx; 876 PetscScalar *bb; 877 878 ierr = VecGetArrayRead(x, &xx);CHKERRQ(ierr); 879 ierr = VecGetArray(b, &bb);CHKERRQ(ierr); 880 for (r = 0; r < len; ++r) bb[lrows[r]] = diag*xx[lrows[r]]; 881 ierr = VecRestoreArrayRead(x, &xx);CHKERRQ(ierr); 882 ierr = VecRestoreArray(b, &bb);CHKERRQ(ierr); 883 } 884 /* Must zero l->B before l->A because the (diag) case below may put values into l->B*/ 885 ierr = MatZeroRows(mat->B, len, lrows, 0.0, NULL, NULL);CHKERRQ(ierr); 886 ierr = MatHasCongruentLayouts(A,&cong);CHKERRQ(ierr); 887 if ((diag != 0.0) && cong) { 888 ierr = MatZeroRows(mat->A, len, lrows, diag, NULL, NULL);CHKERRQ(ierr); 889 } else if (diag != 0.0) { 890 ierr = MatZeroRows(mat->A, len, lrows, 0.0, NULL, NULL);CHKERRQ(ierr); 891 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"); 892 for (r = 0; r < len; ++r) { 893 const PetscInt row = lrows[r] + A->rmap->rstart; 894 ierr = MatSetValues(A, 1, &row, 1, &row, &diag, INSERT_VALUES);CHKERRQ(ierr); 895 } 896 ierr = MatAssemblyBegin(A, MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 897 ierr = MatAssemblyEnd(A, MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 898 } else { 899 ierr = MatZeroRows(mat->A, len, lrows, 0.0, NULL, NULL);CHKERRQ(ierr); 900 } 901 ierr = PetscFree(lrows);CHKERRQ(ierr); 902 903 /* only change matrix nonzero state if pattern was allowed to be changed */ 904 if (!((Mat_SeqAIJ*)(mat->A->data))->keepnonzeropattern) { 905 PetscObjectState state = mat->A->nonzerostate + mat->B->nonzerostate; 906 ierr = MPIU_Allreduce(&state,&A->nonzerostate,1,MPIU_INT64,MPI_SUM,PetscObjectComm((PetscObject)A));CHKERRQ(ierr); 907 } 908 PetscFunctionReturn(0); 909 } 910 911 PetscErrorCode MatZeroRowsColumns_MPIAIJ(Mat A,PetscInt N,const PetscInt rows[],PetscScalar diag,Vec x,Vec b) 912 { 913 Mat_MPIAIJ *l = (Mat_MPIAIJ*)A->data; 914 PetscErrorCode ierr; 915 PetscMPIInt n = A->rmap->n; 916 PetscInt i,j,r,m,p = 0,len = 0; 917 PetscInt *lrows,*owners = A->rmap->range; 918 PetscSFNode *rrows; 919 PetscSF sf; 920 const PetscScalar *xx; 921 PetscScalar *bb,*mask; 922 Vec xmask,lmask; 923 Mat_SeqAIJ *aij = (Mat_SeqAIJ*)l->B->data; 924 const PetscInt *aj, *ii,*ridx; 925 PetscScalar *aa; 926 927 PetscFunctionBegin; 928 /* Create SF where leaves are input rows and roots are owned rows */ 929 ierr = PetscMalloc1(n, &lrows);CHKERRQ(ierr); 930 for (r = 0; r < n; ++r) lrows[r] = -1; 931 ierr = PetscMalloc1(N, &rrows);CHKERRQ(ierr); 932 for (r = 0; r < N; ++r) { 933 const PetscInt idx = rows[r]; 934 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); 935 if (idx < owners[p] || owners[p+1] <= idx) { /* short-circuit the search if the last p owns this row too */ 936 ierr = PetscLayoutFindOwner(A->rmap,idx,&p);CHKERRQ(ierr); 937 } 938 rrows[r].rank = p; 939 rrows[r].index = rows[r] - owners[p]; 940 } 941 ierr = PetscSFCreate(PetscObjectComm((PetscObject) A), &sf);CHKERRQ(ierr); 942 ierr = PetscSFSetGraph(sf, n, N, NULL, PETSC_OWN_POINTER, rrows, PETSC_OWN_POINTER);CHKERRQ(ierr); 943 /* Collect flags for rows to be zeroed */ 944 ierr = PetscSFReduceBegin(sf, MPIU_INT, (PetscInt *) rows, lrows, MPI_LOR);CHKERRQ(ierr); 945 ierr = PetscSFReduceEnd(sf, MPIU_INT, (PetscInt *) rows, lrows, MPI_LOR);CHKERRQ(ierr); 946 ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 947 /* Compress and put in row numbers */ 948 for (r = 0; r < n; ++r) if (lrows[r] >= 0) lrows[len++] = r; 949 /* zero diagonal part of matrix */ 950 ierr = MatZeroRowsColumns(l->A,len,lrows,diag,x,b);CHKERRQ(ierr); 951 /* handle off diagonal part of matrix */ 952 ierr = MatCreateVecs(A,&xmask,NULL);CHKERRQ(ierr); 953 ierr = VecDuplicate(l->lvec,&lmask);CHKERRQ(ierr); 954 ierr = VecGetArray(xmask,&bb);CHKERRQ(ierr); 955 for (i=0; i<len; i++) bb[lrows[i]] = 1; 956 ierr = VecRestoreArray(xmask,&bb);CHKERRQ(ierr); 957 ierr = VecScatterBegin(l->Mvctx,xmask,lmask,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 958 ierr = VecScatterEnd(l->Mvctx,xmask,lmask,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 959 ierr = VecDestroy(&xmask);CHKERRQ(ierr); 960 if (x) { 961 ierr = VecScatterBegin(l->Mvctx,x,l->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 962 ierr = VecScatterEnd(l->Mvctx,x,l->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 963 ierr = VecGetArrayRead(l->lvec,&xx);CHKERRQ(ierr); 964 ierr = VecGetArray(b,&bb);CHKERRQ(ierr); 965 } 966 ierr = VecGetArray(lmask,&mask);CHKERRQ(ierr); 967 /* remove zeroed rows of off diagonal matrix */ 968 ii = aij->i; 969 for (i=0; i<len; i++) { 970 ierr = PetscMemzero(aij->a + ii[lrows[i]],(ii[lrows[i]+1] - ii[lrows[i]])*sizeof(PetscScalar));CHKERRQ(ierr); 971 } 972 /* loop over all elements of off process part of matrix zeroing removed columns*/ 973 if (aij->compressedrow.use) { 974 m = aij->compressedrow.nrows; 975 ii = aij->compressedrow.i; 976 ridx = aij->compressedrow.rindex; 977 for (i=0; i<m; i++) { 978 n = ii[i+1] - ii[i]; 979 aj = aij->j + ii[i]; 980 aa = aij->a + ii[i]; 981 982 for (j=0; j<n; j++) { 983 if (PetscAbsScalar(mask[*aj])) { 984 if (b) bb[*ridx] -= *aa*xx[*aj]; 985 *aa = 0.0; 986 } 987 aa++; 988 aj++; 989 } 990 ridx++; 991 } 992 } else { /* do not use compressed row format */ 993 m = l->B->rmap->n; 994 for (i=0; i<m; i++) { 995 n = ii[i+1] - ii[i]; 996 aj = aij->j + ii[i]; 997 aa = aij->a + ii[i]; 998 for (j=0; j<n; j++) { 999 if (PetscAbsScalar(mask[*aj])) { 1000 if (b) bb[i] -= *aa*xx[*aj]; 1001 *aa = 0.0; 1002 } 1003 aa++; 1004 aj++; 1005 } 1006 } 1007 } 1008 if (x) { 1009 ierr = VecRestoreArray(b,&bb);CHKERRQ(ierr); 1010 ierr = VecRestoreArrayRead(l->lvec,&xx);CHKERRQ(ierr); 1011 } 1012 ierr = VecRestoreArray(lmask,&mask);CHKERRQ(ierr); 1013 ierr = VecDestroy(&lmask);CHKERRQ(ierr); 1014 ierr = PetscFree(lrows);CHKERRQ(ierr); 1015 1016 /* only change matrix nonzero state if pattern was allowed to be changed */ 1017 if (!((Mat_SeqAIJ*)(l->A->data))->keepnonzeropattern) { 1018 PetscObjectState state = l->A->nonzerostate + l->B->nonzerostate; 1019 ierr = MPIU_Allreduce(&state,&A->nonzerostate,1,MPIU_INT64,MPI_SUM,PetscObjectComm((PetscObject)A));CHKERRQ(ierr); 1020 } 1021 PetscFunctionReturn(0); 1022 } 1023 1024 PetscErrorCode MatMult_MPIAIJ(Mat A,Vec xx,Vec yy) 1025 { 1026 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 1027 PetscErrorCode ierr; 1028 PetscInt nt; 1029 VecScatter Mvctx = a->Mvctx; 1030 1031 PetscFunctionBegin; 1032 ierr = VecGetLocalSize(xx,&nt);CHKERRQ(ierr); 1033 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); 1034 1035 ierr = VecScatterBegin(Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1036 ierr = (*a->A->ops->mult)(a->A,xx,yy);CHKERRQ(ierr); 1037 ierr = VecScatterEnd(Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1038 ierr = (*a->B->ops->multadd)(a->B,a->lvec,yy,yy);CHKERRQ(ierr); 1039 PetscFunctionReturn(0); 1040 } 1041 1042 PetscErrorCode MatMultDiagonalBlock_MPIAIJ(Mat A,Vec bb,Vec xx) 1043 { 1044 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 1045 PetscErrorCode ierr; 1046 1047 PetscFunctionBegin; 1048 ierr = MatMultDiagonalBlock(a->A,bb,xx);CHKERRQ(ierr); 1049 PetscFunctionReturn(0); 1050 } 1051 1052 PetscErrorCode MatMultAdd_MPIAIJ(Mat A,Vec xx,Vec yy,Vec zz) 1053 { 1054 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 1055 PetscErrorCode ierr; 1056 VecScatter Mvctx = a->Mvctx; 1057 1058 PetscFunctionBegin; 1059 if (a->Mvctx_mpi1_flg) Mvctx = a->Mvctx_mpi1; 1060 ierr = VecScatterBegin(Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1061 ierr = (*a->A->ops->multadd)(a->A,xx,yy,zz);CHKERRQ(ierr); 1062 ierr = VecScatterEnd(Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1063 ierr = (*a->B->ops->multadd)(a->B,a->lvec,zz,zz);CHKERRQ(ierr); 1064 PetscFunctionReturn(0); 1065 } 1066 1067 PetscErrorCode MatMultTranspose_MPIAIJ(Mat A,Vec xx,Vec yy) 1068 { 1069 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 1070 PetscErrorCode ierr; 1071 PetscBool merged; 1072 1073 PetscFunctionBegin; 1074 ierr = VecScatterGetMerged(a->Mvctx,&merged);CHKERRQ(ierr); 1075 /* do nondiagonal part */ 1076 ierr = (*a->B->ops->multtranspose)(a->B,xx,a->lvec);CHKERRQ(ierr); 1077 if (!merged) { 1078 /* send it on its way */ 1079 ierr = VecScatterBegin(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1080 /* do local part */ 1081 ierr = (*a->A->ops->multtranspose)(a->A,xx,yy);CHKERRQ(ierr); 1082 /* receive remote parts: note this assumes the values are not actually */ 1083 /* added in yy until the next line, */ 1084 ierr = VecScatterEnd(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1085 } else { 1086 /* do local part */ 1087 ierr = (*a->A->ops->multtranspose)(a->A,xx,yy);CHKERRQ(ierr); 1088 /* send it on its way */ 1089 ierr = VecScatterBegin(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1090 /* values actually were received in the Begin() but we need to call this nop */ 1091 ierr = VecScatterEnd(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1092 } 1093 PetscFunctionReturn(0); 1094 } 1095 1096 PetscErrorCode MatIsTranspose_MPIAIJ(Mat Amat,Mat Bmat,PetscReal tol,PetscBool *f) 1097 { 1098 MPI_Comm comm; 1099 Mat_MPIAIJ *Aij = (Mat_MPIAIJ*) Amat->data, *Bij; 1100 Mat Adia = Aij->A, Bdia, Aoff,Boff,*Aoffs,*Boffs; 1101 IS Me,Notme; 1102 PetscErrorCode ierr; 1103 PetscInt M,N,first,last,*notme,i; 1104 PetscBool lf; 1105 PetscMPIInt size; 1106 1107 PetscFunctionBegin; 1108 /* Easy test: symmetric diagonal block */ 1109 Bij = (Mat_MPIAIJ*) Bmat->data; Bdia = Bij->A; 1110 ierr = MatIsTranspose(Adia,Bdia,tol,&lf);CHKERRQ(ierr); 1111 ierr = MPIU_Allreduce(&lf,f,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)Amat));CHKERRQ(ierr); 1112 if (!*f) PetscFunctionReturn(0); 1113 ierr = PetscObjectGetComm((PetscObject)Amat,&comm);CHKERRQ(ierr); 1114 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 1115 if (size == 1) PetscFunctionReturn(0); 1116 1117 /* Hard test: off-diagonal block. This takes a MatCreateSubMatrix. */ 1118 ierr = MatGetSize(Amat,&M,&N);CHKERRQ(ierr); 1119 ierr = MatGetOwnershipRange(Amat,&first,&last);CHKERRQ(ierr); 1120 ierr = PetscMalloc1(N-last+first,¬me);CHKERRQ(ierr); 1121 for (i=0; i<first; i++) notme[i] = i; 1122 for (i=last; i<M; i++) notme[i-last+first] = i; 1123 ierr = ISCreateGeneral(MPI_COMM_SELF,N-last+first,notme,PETSC_COPY_VALUES,&Notme);CHKERRQ(ierr); 1124 ierr = ISCreateStride(MPI_COMM_SELF,last-first,first,1,&Me);CHKERRQ(ierr); 1125 ierr = MatCreateSubMatrices(Amat,1,&Me,&Notme,MAT_INITIAL_MATRIX,&Aoffs);CHKERRQ(ierr); 1126 Aoff = Aoffs[0]; 1127 ierr = MatCreateSubMatrices(Bmat,1,&Notme,&Me,MAT_INITIAL_MATRIX,&Boffs);CHKERRQ(ierr); 1128 Boff = Boffs[0]; 1129 ierr = MatIsTranspose(Aoff,Boff,tol,f);CHKERRQ(ierr); 1130 ierr = MatDestroyMatrices(1,&Aoffs);CHKERRQ(ierr); 1131 ierr = MatDestroyMatrices(1,&Boffs);CHKERRQ(ierr); 1132 ierr = ISDestroy(&Me);CHKERRQ(ierr); 1133 ierr = ISDestroy(&Notme);CHKERRQ(ierr); 1134 ierr = PetscFree(notme);CHKERRQ(ierr); 1135 PetscFunctionReturn(0); 1136 } 1137 1138 PetscErrorCode MatIsSymmetric_MPIAIJ(Mat A,PetscReal tol,PetscBool *f) 1139 { 1140 PetscErrorCode ierr; 1141 1142 PetscFunctionBegin; 1143 ierr = MatIsTranspose_MPIAIJ(A,A,tol,f);CHKERRQ(ierr); 1144 PetscFunctionReturn(0); 1145 } 1146 1147 PetscErrorCode MatMultTransposeAdd_MPIAIJ(Mat A,Vec xx,Vec yy,Vec zz) 1148 { 1149 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 1150 PetscErrorCode ierr; 1151 1152 PetscFunctionBegin; 1153 /* do nondiagonal part */ 1154 ierr = (*a->B->ops->multtranspose)(a->B,xx,a->lvec);CHKERRQ(ierr); 1155 /* send it on its way */ 1156 ierr = VecScatterBegin(a->Mvctx,a->lvec,zz,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1157 /* do local part */ 1158 ierr = (*a->A->ops->multtransposeadd)(a->A,xx,yy,zz);CHKERRQ(ierr); 1159 /* receive remote parts */ 1160 ierr = VecScatterEnd(a->Mvctx,a->lvec,zz,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1161 PetscFunctionReturn(0); 1162 } 1163 1164 /* 1165 This only works correctly for square matrices where the subblock A->A is the 1166 diagonal block 1167 */ 1168 PetscErrorCode MatGetDiagonal_MPIAIJ(Mat A,Vec v) 1169 { 1170 PetscErrorCode ierr; 1171 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 1172 1173 PetscFunctionBegin; 1174 if (A->rmap->N != A->cmap->N) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"Supports only square matrix where A->A is diag block"); 1175 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"); 1176 ierr = MatGetDiagonal(a->A,v);CHKERRQ(ierr); 1177 PetscFunctionReturn(0); 1178 } 1179 1180 PetscErrorCode MatScale_MPIAIJ(Mat A,PetscScalar aa) 1181 { 1182 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 1183 PetscErrorCode ierr; 1184 1185 PetscFunctionBegin; 1186 ierr = MatScale(a->A,aa);CHKERRQ(ierr); 1187 ierr = MatScale(a->B,aa);CHKERRQ(ierr); 1188 PetscFunctionReturn(0); 1189 } 1190 1191 PetscErrorCode MatDestroy_MPIAIJ(Mat mat) 1192 { 1193 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 1194 PetscErrorCode ierr; 1195 1196 PetscFunctionBegin; 1197 #if defined(PETSC_USE_LOG) 1198 PetscLogObjectState((PetscObject)mat,"Rows=%D, Cols=%D",mat->rmap->N,mat->cmap->N); 1199 #endif 1200 ierr = MatStashDestroy_Private(&mat->stash);CHKERRQ(ierr); 1201 ierr = VecDestroy(&aij->diag);CHKERRQ(ierr); 1202 ierr = MatDestroy(&aij->A);CHKERRQ(ierr); 1203 ierr = MatDestroy(&aij->B);CHKERRQ(ierr); 1204 #if defined(PETSC_USE_CTABLE) 1205 ierr = PetscTableDestroy(&aij->colmap);CHKERRQ(ierr); 1206 #else 1207 ierr = PetscFree(aij->colmap);CHKERRQ(ierr); 1208 #endif 1209 ierr = PetscFree(aij->garray);CHKERRQ(ierr); 1210 ierr = VecDestroy(&aij->lvec);CHKERRQ(ierr); 1211 ierr = VecScatterDestroy(&aij->Mvctx);CHKERRQ(ierr); 1212 if (aij->Mvctx_mpi1) {ierr = VecScatterDestroy(&aij->Mvctx_mpi1);CHKERRQ(ierr);} 1213 ierr = PetscFree2(aij->rowvalues,aij->rowindices);CHKERRQ(ierr); 1214 ierr = PetscFree(aij->ld);CHKERRQ(ierr); 1215 ierr = PetscFree(mat->data);CHKERRQ(ierr); 1216 1217 ierr = PetscObjectChangeTypeName((PetscObject)mat,0);CHKERRQ(ierr); 1218 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatStoreValues_C",NULL);CHKERRQ(ierr); 1219 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatRetrieveValues_C",NULL);CHKERRQ(ierr); 1220 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatIsTranspose_C",NULL);CHKERRQ(ierr); 1221 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatMPIAIJSetPreallocation_C",NULL);CHKERRQ(ierr); 1222 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatResetPreallocation_C",NULL);CHKERRQ(ierr); 1223 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatMPIAIJSetPreallocationCSR_C",NULL);CHKERRQ(ierr); 1224 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatDiagonalScaleLocal_C",NULL);CHKERRQ(ierr); 1225 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatConvert_mpiaij_mpisbaij_C",NULL);CHKERRQ(ierr); 1226 #if defined(PETSC_HAVE_ELEMENTAL) 1227 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatConvert_mpiaij_elemental_C",NULL);CHKERRQ(ierr); 1228 #endif 1229 #if defined(PETSC_HAVE_HYPRE) 1230 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatConvert_mpiaij_hypre_C",NULL);CHKERRQ(ierr); 1231 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatMatMatMult_transpose_mpiaij_mpiaij_C",NULL);CHKERRQ(ierr); 1232 #endif 1233 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatConvert_mpiaij_is_C",NULL);CHKERRQ(ierr); 1234 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatPtAP_is_mpiaij_C",NULL);CHKERRQ(ierr); 1235 PetscFunctionReturn(0); 1236 } 1237 1238 PetscErrorCode MatView_MPIAIJ_Binary(Mat mat,PetscViewer viewer) 1239 { 1240 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 1241 Mat_SeqAIJ *A = (Mat_SeqAIJ*)aij->A->data; 1242 Mat_SeqAIJ *B = (Mat_SeqAIJ*)aij->B->data; 1243 PetscErrorCode ierr; 1244 PetscMPIInt rank,size,tag = ((PetscObject)viewer)->tag; 1245 int fd; 1246 PetscInt nz,header[4],*row_lengths,*range=0,rlen,i; 1247 PetscInt nzmax,*column_indices,j,k,col,*garray = aij->garray,cnt,cstart = mat->cmap->rstart,rnz = 0; 1248 PetscScalar *column_values; 1249 PetscInt message_count,flowcontrolcount; 1250 FILE *file; 1251 1252 PetscFunctionBegin; 1253 ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr); 1254 ierr = MPI_Comm_size(PetscObjectComm((PetscObject)mat),&size);CHKERRQ(ierr); 1255 nz = A->nz + B->nz; 1256 ierr = PetscViewerBinaryGetDescriptor(viewer,&fd);CHKERRQ(ierr); 1257 if (!rank) { 1258 header[0] = MAT_FILE_CLASSID; 1259 header[1] = mat->rmap->N; 1260 header[2] = mat->cmap->N; 1261 1262 ierr = MPI_Reduce(&nz,&header[3],1,MPIU_INT,MPI_SUM,0,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1263 ierr = PetscBinaryWrite(fd,header,4,PETSC_INT,PETSC_TRUE);CHKERRQ(ierr); 1264 /* get largest number of rows any processor has */ 1265 rlen = mat->rmap->n; 1266 range = mat->rmap->range; 1267 for (i=1; i<size; i++) rlen = PetscMax(rlen,range[i+1] - range[i]); 1268 } else { 1269 ierr = MPI_Reduce(&nz,0,1,MPIU_INT,MPI_SUM,0,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1270 rlen = mat->rmap->n; 1271 } 1272 1273 /* load up the local row counts */ 1274 ierr = PetscMalloc1(rlen+1,&row_lengths);CHKERRQ(ierr); 1275 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]; 1276 1277 /* store the row lengths to the file */ 1278 ierr = PetscViewerFlowControlStart(viewer,&message_count,&flowcontrolcount);CHKERRQ(ierr); 1279 if (!rank) { 1280 ierr = PetscBinaryWrite(fd,row_lengths,mat->rmap->n,PETSC_INT,PETSC_TRUE);CHKERRQ(ierr); 1281 for (i=1; i<size; i++) { 1282 ierr = PetscViewerFlowControlStepMaster(viewer,i,&message_count,flowcontrolcount);CHKERRQ(ierr); 1283 rlen = range[i+1] - range[i]; 1284 ierr = MPIULong_Recv(row_lengths,rlen,MPIU_INT,i,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1285 ierr = PetscBinaryWrite(fd,row_lengths,rlen,PETSC_INT,PETSC_TRUE);CHKERRQ(ierr); 1286 } 1287 ierr = PetscViewerFlowControlEndMaster(viewer,&message_count);CHKERRQ(ierr); 1288 } else { 1289 ierr = PetscViewerFlowControlStepWorker(viewer,rank,&message_count);CHKERRQ(ierr); 1290 ierr = MPIULong_Send(row_lengths,mat->rmap->n,MPIU_INT,0,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1291 ierr = PetscViewerFlowControlEndWorker(viewer,&message_count);CHKERRQ(ierr); 1292 } 1293 ierr = PetscFree(row_lengths);CHKERRQ(ierr); 1294 1295 /* load up the local column indices */ 1296 nzmax = nz; /* th processor needs space a largest processor needs */ 1297 ierr = MPI_Reduce(&nz,&nzmax,1,MPIU_INT,MPI_MAX,0,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1298 ierr = PetscMalloc1(nzmax+1,&column_indices);CHKERRQ(ierr); 1299 cnt = 0; 1300 for (i=0; i<mat->rmap->n; i++) { 1301 for (j=B->i[i]; j<B->i[i+1]; j++) { 1302 if ((col = garray[B->j[j]]) > cstart) break; 1303 column_indices[cnt++] = col; 1304 } 1305 for (k=A->i[i]; k<A->i[i+1]; k++) column_indices[cnt++] = A->j[k] + cstart; 1306 for (; j<B->i[i+1]; j++) column_indices[cnt++] = garray[B->j[j]]; 1307 } 1308 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); 1309 1310 /* store the column indices to the file */ 1311 ierr = PetscViewerFlowControlStart(viewer,&message_count,&flowcontrolcount);CHKERRQ(ierr); 1312 if (!rank) { 1313 MPI_Status status; 1314 ierr = PetscBinaryWrite(fd,column_indices,nz,PETSC_INT,PETSC_TRUE);CHKERRQ(ierr); 1315 for (i=1; i<size; i++) { 1316 ierr = PetscViewerFlowControlStepMaster(viewer,i,&message_count,flowcontrolcount);CHKERRQ(ierr); 1317 ierr = MPI_Recv(&rnz,1,MPIU_INT,i,tag,PetscObjectComm((PetscObject)mat),&status);CHKERRQ(ierr); 1318 if (rnz > nzmax) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_LIB,"Internal PETSc error: nz = %D nzmax = %D",nz,nzmax); 1319 ierr = MPIULong_Recv(column_indices,rnz,MPIU_INT,i,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1320 ierr = PetscBinaryWrite(fd,column_indices,rnz,PETSC_INT,PETSC_TRUE);CHKERRQ(ierr); 1321 } 1322 ierr = PetscViewerFlowControlEndMaster(viewer,&message_count);CHKERRQ(ierr); 1323 } else { 1324 ierr = PetscViewerFlowControlStepWorker(viewer,rank,&message_count);CHKERRQ(ierr); 1325 ierr = MPI_Send(&nz,1,MPIU_INT,0,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1326 ierr = MPIULong_Send(column_indices,nz,MPIU_INT,0,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1327 ierr = PetscViewerFlowControlEndWorker(viewer,&message_count);CHKERRQ(ierr); 1328 } 1329 ierr = PetscFree(column_indices);CHKERRQ(ierr); 1330 1331 /* load up the local column values */ 1332 ierr = PetscMalloc1(nzmax+1,&column_values);CHKERRQ(ierr); 1333 cnt = 0; 1334 for (i=0; i<mat->rmap->n; i++) { 1335 for (j=B->i[i]; j<B->i[i+1]; j++) { 1336 if (garray[B->j[j]] > cstart) break; 1337 column_values[cnt++] = B->a[j]; 1338 } 1339 for (k=A->i[i]; k<A->i[i+1]; k++) column_values[cnt++] = A->a[k]; 1340 for (; j<B->i[i+1]; j++) column_values[cnt++] = B->a[j]; 1341 } 1342 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); 1343 1344 /* store the column values to the file */ 1345 ierr = PetscViewerFlowControlStart(viewer,&message_count,&flowcontrolcount);CHKERRQ(ierr); 1346 if (!rank) { 1347 MPI_Status status; 1348 ierr = PetscBinaryWrite(fd,column_values,nz,PETSC_SCALAR,PETSC_TRUE);CHKERRQ(ierr); 1349 for (i=1; i<size; i++) { 1350 ierr = PetscViewerFlowControlStepMaster(viewer,i,&message_count,flowcontrolcount);CHKERRQ(ierr); 1351 ierr = MPI_Recv(&rnz,1,MPIU_INT,i,tag,PetscObjectComm((PetscObject)mat),&status);CHKERRQ(ierr); 1352 if (rnz > nzmax) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_LIB,"Internal PETSc error: nz = %D nzmax = %D",nz,nzmax); 1353 ierr = MPIULong_Recv(column_values,rnz,MPIU_SCALAR,i,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1354 ierr = PetscBinaryWrite(fd,column_values,rnz,PETSC_SCALAR,PETSC_TRUE);CHKERRQ(ierr); 1355 } 1356 ierr = PetscViewerFlowControlEndMaster(viewer,&message_count);CHKERRQ(ierr); 1357 } else { 1358 ierr = PetscViewerFlowControlStepWorker(viewer,rank,&message_count);CHKERRQ(ierr); 1359 ierr = MPI_Send(&nz,1,MPIU_INT,0,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1360 ierr = MPIULong_Send(column_values,nz,MPIU_SCALAR,0,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1361 ierr = PetscViewerFlowControlEndWorker(viewer,&message_count);CHKERRQ(ierr); 1362 } 1363 ierr = PetscFree(column_values);CHKERRQ(ierr); 1364 1365 ierr = PetscViewerBinaryGetInfoPointer(viewer,&file);CHKERRQ(ierr); 1366 if (file) fprintf(file,"-matload_block_size %d\n",(int)PetscAbs(mat->rmap->bs)); 1367 PetscFunctionReturn(0); 1368 } 1369 1370 #include <petscdraw.h> 1371 PetscErrorCode MatView_MPIAIJ_ASCIIorDraworSocket(Mat mat,PetscViewer viewer) 1372 { 1373 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 1374 PetscErrorCode ierr; 1375 PetscMPIInt rank = aij->rank,size = aij->size; 1376 PetscBool isdraw,iascii,isbinary; 1377 PetscViewer sviewer; 1378 PetscViewerFormat format; 1379 1380 PetscFunctionBegin; 1381 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERDRAW,&isdraw);CHKERRQ(ierr); 1382 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii);CHKERRQ(ierr); 1383 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERBINARY,&isbinary);CHKERRQ(ierr); 1384 if (iascii) { 1385 ierr = PetscViewerGetFormat(viewer,&format);CHKERRQ(ierr); 1386 if (format == PETSC_VIEWER_LOAD_BALANCE) { 1387 PetscInt i,nmax = 0,nmin = PETSC_MAX_INT,navg = 0,*nz,nzlocal = ((Mat_SeqAIJ*) (aij->A->data))->nz + ((Mat_SeqAIJ*) (aij->B->data))->nz; 1388 ierr = PetscMalloc1(size,&nz);CHKERRQ(ierr); 1389 ierr = MPI_Allgather(&nzlocal,1,MPIU_INT,nz,1,MPIU_INT,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1390 for (i=0; i<(PetscInt)size; i++) { 1391 nmax = PetscMax(nmax,nz[i]); 1392 nmin = PetscMin(nmin,nz[i]); 1393 navg += nz[i]; 1394 } 1395 ierr = PetscFree(nz);CHKERRQ(ierr); 1396 navg = navg/size; 1397 ierr = PetscViewerASCIIPrintf(viewer,"Load Balance - Nonzeros: Min %D avg %D max %D\n",nmin,navg,nmax);CHKERRQ(ierr); 1398 PetscFunctionReturn(0); 1399 } 1400 ierr = PetscViewerGetFormat(viewer,&format);CHKERRQ(ierr); 1401 if (format == PETSC_VIEWER_ASCII_INFO_DETAIL) { 1402 MatInfo info; 1403 PetscBool inodes; 1404 1405 ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr); 1406 ierr = MatGetInfo(mat,MAT_LOCAL,&info);CHKERRQ(ierr); 1407 ierr = MatInodeGetInodeSizes(aij->A,NULL,(PetscInt**)&inodes,NULL);CHKERRQ(ierr); 1408 ierr = PetscViewerASCIIPushSynchronized(viewer);CHKERRQ(ierr); 1409 if (!inodes) { 1410 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"[%d] Local rows %D nz %D nz alloced %D mem %g, not using I-node routines\n", 1411 rank,mat->rmap->n,(PetscInt)info.nz_used,(PetscInt)info.nz_allocated,(double)info.memory);CHKERRQ(ierr); 1412 } else { 1413 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"[%d] Local rows %D nz %D nz alloced %D mem %g, using I-node routines\n", 1414 rank,mat->rmap->n,(PetscInt)info.nz_used,(PetscInt)info.nz_allocated,(double)info.memory);CHKERRQ(ierr); 1415 } 1416 ierr = MatGetInfo(aij->A,MAT_LOCAL,&info);CHKERRQ(ierr); 1417 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"[%d] on-diagonal part: nz %D \n",rank,(PetscInt)info.nz_used);CHKERRQ(ierr); 1418 ierr = MatGetInfo(aij->B,MAT_LOCAL,&info);CHKERRQ(ierr); 1419 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"[%d] off-diagonal part: nz %D \n",rank,(PetscInt)info.nz_used);CHKERRQ(ierr); 1420 ierr = PetscViewerFlush(viewer);CHKERRQ(ierr); 1421 ierr = PetscViewerASCIIPopSynchronized(viewer);CHKERRQ(ierr); 1422 ierr = PetscViewerASCIIPrintf(viewer,"Information on VecScatter used in matrix-vector product: \n");CHKERRQ(ierr); 1423 ierr = VecScatterView(aij->Mvctx,viewer);CHKERRQ(ierr); 1424 PetscFunctionReturn(0); 1425 } else if (format == PETSC_VIEWER_ASCII_INFO) { 1426 PetscInt inodecount,inodelimit,*inodes; 1427 ierr = MatInodeGetInodeSizes(aij->A,&inodecount,&inodes,&inodelimit);CHKERRQ(ierr); 1428 if (inodes) { 1429 ierr = PetscViewerASCIIPrintf(viewer,"using I-node (on process 0) routines: found %D nodes, limit used is %D\n",inodecount,inodelimit);CHKERRQ(ierr); 1430 } else { 1431 ierr = PetscViewerASCIIPrintf(viewer,"not using I-node (on process 0) routines\n");CHKERRQ(ierr); 1432 } 1433 PetscFunctionReturn(0); 1434 } else if (format == PETSC_VIEWER_ASCII_FACTOR_INFO) { 1435 PetscFunctionReturn(0); 1436 } 1437 } else if (isbinary) { 1438 if (size == 1) { 1439 ierr = PetscObjectSetName((PetscObject)aij->A,((PetscObject)mat)->name);CHKERRQ(ierr); 1440 ierr = MatView(aij->A,viewer);CHKERRQ(ierr); 1441 } else { 1442 ierr = MatView_MPIAIJ_Binary(mat,viewer);CHKERRQ(ierr); 1443 } 1444 PetscFunctionReturn(0); 1445 } else if (isdraw) { 1446 PetscDraw draw; 1447 PetscBool isnull; 1448 ierr = PetscViewerDrawGetDraw(viewer,0,&draw);CHKERRQ(ierr); 1449 ierr = PetscDrawIsNull(draw,&isnull);CHKERRQ(ierr); 1450 if (isnull) PetscFunctionReturn(0); 1451 } 1452 1453 { 1454 /* assemble the entire matrix onto first processor. */ 1455 Mat A; 1456 Mat_SeqAIJ *Aloc; 1457 PetscInt M = mat->rmap->N,N = mat->cmap->N,m,*ai,*aj,row,*cols,i,*ct; 1458 MatScalar *a; 1459 1460 ierr = MatCreate(PetscObjectComm((PetscObject)mat),&A);CHKERRQ(ierr); 1461 if (!rank) { 1462 ierr = MatSetSizes(A,M,N,M,N);CHKERRQ(ierr); 1463 } else { 1464 ierr = MatSetSizes(A,0,0,M,N);CHKERRQ(ierr); 1465 } 1466 /* This is just a temporary matrix, so explicitly using MATMPIAIJ is probably best */ 1467 ierr = MatSetType(A,MATMPIAIJ);CHKERRQ(ierr); 1468 ierr = MatMPIAIJSetPreallocation(A,0,NULL,0,NULL);CHKERRQ(ierr); 1469 ierr = MatSetOption(A,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 1470 ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)A);CHKERRQ(ierr); 1471 1472 /* copy over the A part */ 1473 Aloc = (Mat_SeqAIJ*)aij->A->data; 1474 m = aij->A->rmap->n; ai = Aloc->i; aj = Aloc->j; a = Aloc->a; 1475 row = mat->rmap->rstart; 1476 for (i=0; i<ai[m]; i++) aj[i] += mat->cmap->rstart; 1477 for (i=0; i<m; i++) { 1478 ierr = MatSetValues(A,1,&row,ai[i+1]-ai[i],aj,a,INSERT_VALUES);CHKERRQ(ierr); 1479 row++; 1480 a += ai[i+1]-ai[i]; aj += ai[i+1]-ai[i]; 1481 } 1482 aj = Aloc->j; 1483 for (i=0; i<ai[m]; i++) aj[i] -= mat->cmap->rstart; 1484 1485 /* copy over the B part */ 1486 Aloc = (Mat_SeqAIJ*)aij->B->data; 1487 m = aij->B->rmap->n; ai = Aloc->i; aj = Aloc->j; a = Aloc->a; 1488 row = mat->rmap->rstart; 1489 ierr = PetscMalloc1(ai[m]+1,&cols);CHKERRQ(ierr); 1490 ct = cols; 1491 for (i=0; i<ai[m]; i++) cols[i] = aij->garray[aj[i]]; 1492 for (i=0; i<m; i++) { 1493 ierr = MatSetValues(A,1,&row,ai[i+1]-ai[i],cols,a,INSERT_VALUES);CHKERRQ(ierr); 1494 row++; 1495 a += ai[i+1]-ai[i]; cols += ai[i+1]-ai[i]; 1496 } 1497 ierr = PetscFree(ct);CHKERRQ(ierr); 1498 ierr = MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1499 ierr = MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1500 /* 1501 Everyone has to call to draw the matrix since the graphics waits are 1502 synchronized across all processors that share the PetscDraw object 1503 */ 1504 ierr = PetscViewerGetSubViewer(viewer,PETSC_COMM_SELF,&sviewer);CHKERRQ(ierr); 1505 if (!rank) { 1506 ierr = PetscObjectSetName((PetscObject)((Mat_MPIAIJ*)(A->data))->A,((PetscObject)mat)->name);CHKERRQ(ierr); 1507 ierr = MatView_SeqAIJ(((Mat_MPIAIJ*)(A->data))->A,sviewer);CHKERRQ(ierr); 1508 } 1509 ierr = PetscViewerRestoreSubViewer(viewer,PETSC_COMM_SELF,&sviewer);CHKERRQ(ierr); 1510 ierr = PetscViewerFlush(viewer);CHKERRQ(ierr); 1511 ierr = MatDestroy(&A);CHKERRQ(ierr); 1512 } 1513 PetscFunctionReturn(0); 1514 } 1515 1516 PetscErrorCode MatView_MPIAIJ(Mat mat,PetscViewer viewer) 1517 { 1518 PetscErrorCode ierr; 1519 PetscBool iascii,isdraw,issocket,isbinary; 1520 1521 PetscFunctionBegin; 1522 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii);CHKERRQ(ierr); 1523 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERDRAW,&isdraw);CHKERRQ(ierr); 1524 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERBINARY,&isbinary);CHKERRQ(ierr); 1525 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERSOCKET,&issocket);CHKERRQ(ierr); 1526 if (iascii || isdraw || isbinary || issocket) { 1527 ierr = MatView_MPIAIJ_ASCIIorDraworSocket(mat,viewer);CHKERRQ(ierr); 1528 } 1529 PetscFunctionReturn(0); 1530 } 1531 1532 PetscErrorCode MatSOR_MPIAIJ(Mat matin,Vec bb,PetscReal omega,MatSORType flag,PetscReal fshift,PetscInt its,PetscInt lits,Vec xx) 1533 { 1534 Mat_MPIAIJ *mat = (Mat_MPIAIJ*)matin->data; 1535 PetscErrorCode ierr; 1536 Vec bb1 = 0; 1537 PetscBool hasop; 1538 1539 PetscFunctionBegin; 1540 if (flag == SOR_APPLY_UPPER) { 1541 ierr = (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);CHKERRQ(ierr); 1542 PetscFunctionReturn(0); 1543 } 1544 1545 if (its > 1 || ~flag & SOR_ZERO_INITIAL_GUESS || flag & SOR_EISENSTAT) { 1546 ierr = VecDuplicate(bb,&bb1);CHKERRQ(ierr); 1547 } 1548 1549 if ((flag & SOR_LOCAL_SYMMETRIC_SWEEP) == SOR_LOCAL_SYMMETRIC_SWEEP) { 1550 if (flag & SOR_ZERO_INITIAL_GUESS) { 1551 ierr = (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);CHKERRQ(ierr); 1552 its--; 1553 } 1554 1555 while (its--) { 1556 ierr = VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1557 ierr = VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1558 1559 /* update rhs: bb1 = bb - B*x */ 1560 ierr = VecScale(mat->lvec,-1.0);CHKERRQ(ierr); 1561 ierr = (*mat->B->ops->multadd)(mat->B,mat->lvec,bb,bb1);CHKERRQ(ierr); 1562 1563 /* local sweep */ 1564 ierr = (*mat->A->ops->sor)(mat->A,bb1,omega,SOR_SYMMETRIC_SWEEP,fshift,lits,1,xx);CHKERRQ(ierr); 1565 } 1566 } else if (flag & SOR_LOCAL_FORWARD_SWEEP) { 1567 if (flag & SOR_ZERO_INITIAL_GUESS) { 1568 ierr = (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);CHKERRQ(ierr); 1569 its--; 1570 } 1571 while (its--) { 1572 ierr = VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1573 ierr = VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1574 1575 /* update rhs: bb1 = bb - B*x */ 1576 ierr = VecScale(mat->lvec,-1.0);CHKERRQ(ierr); 1577 ierr = (*mat->B->ops->multadd)(mat->B,mat->lvec,bb,bb1);CHKERRQ(ierr); 1578 1579 /* local sweep */ 1580 ierr = (*mat->A->ops->sor)(mat->A,bb1,omega,SOR_FORWARD_SWEEP,fshift,lits,1,xx);CHKERRQ(ierr); 1581 } 1582 } else if (flag & SOR_LOCAL_BACKWARD_SWEEP) { 1583 if (flag & SOR_ZERO_INITIAL_GUESS) { 1584 ierr = (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);CHKERRQ(ierr); 1585 its--; 1586 } 1587 while (its--) { 1588 ierr = VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1589 ierr = VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1590 1591 /* update rhs: bb1 = bb - B*x */ 1592 ierr = VecScale(mat->lvec,-1.0);CHKERRQ(ierr); 1593 ierr = (*mat->B->ops->multadd)(mat->B,mat->lvec,bb,bb1);CHKERRQ(ierr); 1594 1595 /* local sweep */ 1596 ierr = (*mat->A->ops->sor)(mat->A,bb1,omega,SOR_BACKWARD_SWEEP,fshift,lits,1,xx);CHKERRQ(ierr); 1597 } 1598 } else if (flag & SOR_EISENSTAT) { 1599 Vec xx1; 1600 1601 ierr = VecDuplicate(bb,&xx1);CHKERRQ(ierr); 1602 ierr = (*mat->A->ops->sor)(mat->A,bb,omega,(MatSORType)(SOR_ZERO_INITIAL_GUESS | SOR_LOCAL_BACKWARD_SWEEP),fshift,lits,1,xx);CHKERRQ(ierr); 1603 1604 ierr = VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1605 ierr = VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1606 if (!mat->diag) { 1607 ierr = MatCreateVecs(matin,&mat->diag,NULL);CHKERRQ(ierr); 1608 ierr = MatGetDiagonal(matin,mat->diag);CHKERRQ(ierr); 1609 } 1610 ierr = MatHasOperation(matin,MATOP_MULT_DIAGONAL_BLOCK,&hasop);CHKERRQ(ierr); 1611 if (hasop) { 1612 ierr = MatMultDiagonalBlock(matin,xx,bb1);CHKERRQ(ierr); 1613 } else { 1614 ierr = VecPointwiseMult(bb1,mat->diag,xx);CHKERRQ(ierr); 1615 } 1616 ierr = VecAYPX(bb1,(omega-2.0)/omega,bb);CHKERRQ(ierr); 1617 1618 ierr = MatMultAdd(mat->B,mat->lvec,bb1,bb1);CHKERRQ(ierr); 1619 1620 /* local sweep */ 1621 ierr = (*mat->A->ops->sor)(mat->A,bb1,omega,(MatSORType)(SOR_ZERO_INITIAL_GUESS | SOR_LOCAL_FORWARD_SWEEP),fshift,lits,1,xx1);CHKERRQ(ierr); 1622 ierr = VecAXPY(xx,1.0,xx1);CHKERRQ(ierr); 1623 ierr = VecDestroy(&xx1);CHKERRQ(ierr); 1624 } else SETERRQ(PetscObjectComm((PetscObject)matin),PETSC_ERR_SUP,"Parallel SOR not supported"); 1625 1626 ierr = VecDestroy(&bb1);CHKERRQ(ierr); 1627 1628 matin->factorerrortype = mat->A->factorerrortype; 1629 PetscFunctionReturn(0); 1630 } 1631 1632 PetscErrorCode MatPermute_MPIAIJ(Mat A,IS rowp,IS colp,Mat *B) 1633 { 1634 Mat aA,aB,Aperm; 1635 const PetscInt *rwant,*cwant,*gcols,*ai,*bi,*aj,*bj; 1636 PetscScalar *aa,*ba; 1637 PetscInt i,j,m,n,ng,anz,bnz,*dnnz,*onnz,*tdnnz,*tonnz,*rdest,*cdest,*work,*gcdest; 1638 PetscSF rowsf,sf; 1639 IS parcolp = NULL; 1640 PetscBool done; 1641 PetscErrorCode ierr; 1642 1643 PetscFunctionBegin; 1644 ierr = MatGetLocalSize(A,&m,&n);CHKERRQ(ierr); 1645 ierr = ISGetIndices(rowp,&rwant);CHKERRQ(ierr); 1646 ierr = ISGetIndices(colp,&cwant);CHKERRQ(ierr); 1647 ierr = PetscMalloc3(PetscMax(m,n),&work,m,&rdest,n,&cdest);CHKERRQ(ierr); 1648 1649 /* Invert row permutation to find out where my rows should go */ 1650 ierr = PetscSFCreate(PetscObjectComm((PetscObject)A),&rowsf);CHKERRQ(ierr); 1651 ierr = PetscSFSetGraphLayout(rowsf,A->rmap,A->rmap->n,NULL,PETSC_OWN_POINTER,rwant);CHKERRQ(ierr); 1652 ierr = PetscSFSetFromOptions(rowsf);CHKERRQ(ierr); 1653 for (i=0; i<m; i++) work[i] = A->rmap->rstart + i; 1654 ierr = PetscSFReduceBegin(rowsf,MPIU_INT,work,rdest,MPIU_REPLACE);CHKERRQ(ierr); 1655 ierr = PetscSFReduceEnd(rowsf,MPIU_INT,work,rdest,MPIU_REPLACE);CHKERRQ(ierr); 1656 1657 /* Invert column permutation to find out where my columns should go */ 1658 ierr = PetscSFCreate(PetscObjectComm((PetscObject)A),&sf);CHKERRQ(ierr); 1659 ierr = PetscSFSetGraphLayout(sf,A->cmap,A->cmap->n,NULL,PETSC_OWN_POINTER,cwant);CHKERRQ(ierr); 1660 ierr = PetscSFSetFromOptions(sf);CHKERRQ(ierr); 1661 for (i=0; i<n; i++) work[i] = A->cmap->rstart + i; 1662 ierr = PetscSFReduceBegin(sf,MPIU_INT,work,cdest,MPIU_REPLACE);CHKERRQ(ierr); 1663 ierr = PetscSFReduceEnd(sf,MPIU_INT,work,cdest,MPIU_REPLACE);CHKERRQ(ierr); 1664 ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 1665 1666 ierr = ISRestoreIndices(rowp,&rwant);CHKERRQ(ierr); 1667 ierr = ISRestoreIndices(colp,&cwant);CHKERRQ(ierr); 1668 ierr = MatMPIAIJGetSeqAIJ(A,&aA,&aB,&gcols);CHKERRQ(ierr); 1669 1670 /* Find out where my gcols should go */ 1671 ierr = MatGetSize(aB,NULL,&ng);CHKERRQ(ierr); 1672 ierr = PetscMalloc1(ng,&gcdest);CHKERRQ(ierr); 1673 ierr = PetscSFCreate(PetscObjectComm((PetscObject)A),&sf);CHKERRQ(ierr); 1674 ierr = PetscSFSetGraphLayout(sf,A->cmap,ng,NULL,PETSC_OWN_POINTER,gcols);CHKERRQ(ierr); 1675 ierr = PetscSFSetFromOptions(sf);CHKERRQ(ierr); 1676 ierr = PetscSFBcastBegin(sf,MPIU_INT,cdest,gcdest);CHKERRQ(ierr); 1677 ierr = PetscSFBcastEnd(sf,MPIU_INT,cdest,gcdest);CHKERRQ(ierr); 1678 ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 1679 1680 ierr = PetscCalloc4(m,&dnnz,m,&onnz,m,&tdnnz,m,&tonnz);CHKERRQ(ierr); 1681 ierr = MatGetRowIJ(aA,0,PETSC_FALSE,PETSC_FALSE,&anz,&ai,&aj,&done);CHKERRQ(ierr); 1682 ierr = MatGetRowIJ(aB,0,PETSC_FALSE,PETSC_FALSE,&bnz,&bi,&bj,&done);CHKERRQ(ierr); 1683 for (i=0; i<m; i++) { 1684 PetscInt row = rdest[i],rowner; 1685 ierr = PetscLayoutFindOwner(A->rmap,row,&rowner);CHKERRQ(ierr); 1686 for (j=ai[i]; j<ai[i+1]; j++) { 1687 PetscInt cowner,col = cdest[aj[j]]; 1688 ierr = PetscLayoutFindOwner(A->cmap,col,&cowner);CHKERRQ(ierr); /* Could build an index for the columns to eliminate this search */ 1689 if (rowner == cowner) dnnz[i]++; 1690 else onnz[i]++; 1691 } 1692 for (j=bi[i]; j<bi[i+1]; j++) { 1693 PetscInt cowner,col = gcdest[bj[j]]; 1694 ierr = PetscLayoutFindOwner(A->cmap,col,&cowner);CHKERRQ(ierr); 1695 if (rowner == cowner) dnnz[i]++; 1696 else onnz[i]++; 1697 } 1698 } 1699 ierr = PetscSFBcastBegin(rowsf,MPIU_INT,dnnz,tdnnz);CHKERRQ(ierr); 1700 ierr = PetscSFBcastEnd(rowsf,MPIU_INT,dnnz,tdnnz);CHKERRQ(ierr); 1701 ierr = PetscSFBcastBegin(rowsf,MPIU_INT,onnz,tonnz);CHKERRQ(ierr); 1702 ierr = PetscSFBcastEnd(rowsf,MPIU_INT,onnz,tonnz);CHKERRQ(ierr); 1703 ierr = PetscSFDestroy(&rowsf);CHKERRQ(ierr); 1704 1705 ierr = MatCreateAIJ(PetscObjectComm((PetscObject)A),A->rmap->n,A->cmap->n,A->rmap->N,A->cmap->N,0,tdnnz,0,tonnz,&Aperm);CHKERRQ(ierr); 1706 ierr = MatSeqAIJGetArray(aA,&aa);CHKERRQ(ierr); 1707 ierr = MatSeqAIJGetArray(aB,&ba);CHKERRQ(ierr); 1708 for (i=0; i<m; i++) { 1709 PetscInt *acols = dnnz,*bcols = onnz; /* Repurpose now-unneeded arrays */ 1710 PetscInt j0,rowlen; 1711 rowlen = ai[i+1] - ai[i]; 1712 for (j0=j=0; j<rowlen; j0=j) { /* rowlen could be larger than number of rows m, so sum in batches */ 1713 for ( ; j<PetscMin(rowlen,j0+m); j++) acols[j-j0] = cdest[aj[ai[i]+j]]; 1714 ierr = MatSetValues(Aperm,1,&rdest[i],j-j0,acols,aa+ai[i]+j0,INSERT_VALUES);CHKERRQ(ierr); 1715 } 1716 rowlen = bi[i+1] - bi[i]; 1717 for (j0=j=0; j<rowlen; j0=j) { 1718 for ( ; j<PetscMin(rowlen,j0+m); j++) bcols[j-j0] = gcdest[bj[bi[i]+j]]; 1719 ierr = MatSetValues(Aperm,1,&rdest[i],j-j0,bcols,ba+bi[i]+j0,INSERT_VALUES);CHKERRQ(ierr); 1720 } 1721 } 1722 ierr = MatAssemblyBegin(Aperm,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1723 ierr = MatAssemblyEnd(Aperm,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1724 ierr = MatRestoreRowIJ(aA,0,PETSC_FALSE,PETSC_FALSE,&anz,&ai,&aj,&done);CHKERRQ(ierr); 1725 ierr = MatRestoreRowIJ(aB,0,PETSC_FALSE,PETSC_FALSE,&bnz,&bi,&bj,&done);CHKERRQ(ierr); 1726 ierr = MatSeqAIJRestoreArray(aA,&aa);CHKERRQ(ierr); 1727 ierr = MatSeqAIJRestoreArray(aB,&ba);CHKERRQ(ierr); 1728 ierr = PetscFree4(dnnz,onnz,tdnnz,tonnz);CHKERRQ(ierr); 1729 ierr = PetscFree3(work,rdest,cdest);CHKERRQ(ierr); 1730 ierr = PetscFree(gcdest);CHKERRQ(ierr); 1731 if (parcolp) {ierr = ISDestroy(&colp);CHKERRQ(ierr);} 1732 *B = Aperm; 1733 PetscFunctionReturn(0); 1734 } 1735 1736 PetscErrorCode MatGetGhosts_MPIAIJ(Mat mat,PetscInt *nghosts,const PetscInt *ghosts[]) 1737 { 1738 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 1739 PetscErrorCode ierr; 1740 1741 PetscFunctionBegin; 1742 ierr = MatGetSize(aij->B,NULL,nghosts);CHKERRQ(ierr); 1743 if (ghosts) *ghosts = aij->garray; 1744 PetscFunctionReturn(0); 1745 } 1746 1747 PetscErrorCode MatGetInfo_MPIAIJ(Mat matin,MatInfoType flag,MatInfo *info) 1748 { 1749 Mat_MPIAIJ *mat = (Mat_MPIAIJ*)matin->data; 1750 Mat A = mat->A,B = mat->B; 1751 PetscErrorCode ierr; 1752 PetscReal isend[5],irecv[5]; 1753 1754 PetscFunctionBegin; 1755 info->block_size = 1.0; 1756 ierr = MatGetInfo(A,MAT_LOCAL,info);CHKERRQ(ierr); 1757 1758 isend[0] = info->nz_used; isend[1] = info->nz_allocated; isend[2] = info->nz_unneeded; 1759 isend[3] = info->memory; isend[4] = info->mallocs; 1760 1761 ierr = MatGetInfo(B,MAT_LOCAL,info);CHKERRQ(ierr); 1762 1763 isend[0] += info->nz_used; isend[1] += info->nz_allocated; isend[2] += info->nz_unneeded; 1764 isend[3] += info->memory; isend[4] += info->mallocs; 1765 if (flag == MAT_LOCAL) { 1766 info->nz_used = isend[0]; 1767 info->nz_allocated = isend[1]; 1768 info->nz_unneeded = isend[2]; 1769 info->memory = isend[3]; 1770 info->mallocs = isend[4]; 1771 } else if (flag == MAT_GLOBAL_MAX) { 1772 ierr = MPIU_Allreduce(isend,irecv,5,MPIU_REAL,MPIU_MAX,PetscObjectComm((PetscObject)matin));CHKERRQ(ierr); 1773 1774 info->nz_used = irecv[0]; 1775 info->nz_allocated = irecv[1]; 1776 info->nz_unneeded = irecv[2]; 1777 info->memory = irecv[3]; 1778 info->mallocs = irecv[4]; 1779 } else if (flag == MAT_GLOBAL_SUM) { 1780 ierr = MPIU_Allreduce(isend,irecv,5,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)matin));CHKERRQ(ierr); 1781 1782 info->nz_used = irecv[0]; 1783 info->nz_allocated = irecv[1]; 1784 info->nz_unneeded = irecv[2]; 1785 info->memory = irecv[3]; 1786 info->mallocs = irecv[4]; 1787 } 1788 info->fill_ratio_given = 0; /* no parallel LU/ILU/Cholesky */ 1789 info->fill_ratio_needed = 0; 1790 info->factor_mallocs = 0; 1791 PetscFunctionReturn(0); 1792 } 1793 1794 PetscErrorCode MatSetOption_MPIAIJ(Mat A,MatOption op,PetscBool flg) 1795 { 1796 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 1797 PetscErrorCode ierr; 1798 1799 PetscFunctionBegin; 1800 switch (op) { 1801 case MAT_NEW_NONZERO_LOCATIONS: 1802 case MAT_NEW_NONZERO_ALLOCATION_ERR: 1803 case MAT_UNUSED_NONZERO_LOCATION_ERR: 1804 case MAT_KEEP_NONZERO_PATTERN: 1805 case MAT_NEW_NONZERO_LOCATION_ERR: 1806 case MAT_USE_INODES: 1807 case MAT_IGNORE_ZERO_ENTRIES: 1808 MatCheckPreallocated(A,1); 1809 ierr = MatSetOption(a->A,op,flg);CHKERRQ(ierr); 1810 ierr = MatSetOption(a->B,op,flg);CHKERRQ(ierr); 1811 break; 1812 case MAT_ROW_ORIENTED: 1813 MatCheckPreallocated(A,1); 1814 a->roworiented = flg; 1815 1816 ierr = MatSetOption(a->A,op,flg);CHKERRQ(ierr); 1817 ierr = MatSetOption(a->B,op,flg);CHKERRQ(ierr); 1818 break; 1819 case MAT_NEW_DIAGONALS: 1820 ierr = PetscInfo1(A,"Option %s ignored\n",MatOptions[op]);CHKERRQ(ierr); 1821 break; 1822 case MAT_IGNORE_OFF_PROC_ENTRIES: 1823 a->donotstash = flg; 1824 break; 1825 /* Symmetry flags are handled directly by MatSetOption() and they don't affect preallocation */ 1826 case MAT_SPD: 1827 case MAT_SYMMETRIC: 1828 case MAT_STRUCTURALLY_SYMMETRIC: 1829 case MAT_HERMITIAN: 1830 case MAT_SYMMETRY_ETERNAL: 1831 break; 1832 case MAT_SUBMAT_SINGLEIS: 1833 A->submat_singleis = flg; 1834 break; 1835 case MAT_STRUCTURE_ONLY: 1836 /* The option is handled directly by MatSetOption() */ 1837 break; 1838 default: 1839 SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"unknown option %d",op); 1840 } 1841 PetscFunctionReturn(0); 1842 } 1843 1844 PetscErrorCode MatGetRow_MPIAIJ(Mat matin,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v) 1845 { 1846 Mat_MPIAIJ *mat = (Mat_MPIAIJ*)matin->data; 1847 PetscScalar *vworkA,*vworkB,**pvA,**pvB,*v_p; 1848 PetscErrorCode ierr; 1849 PetscInt i,*cworkA,*cworkB,**pcA,**pcB,cstart = matin->cmap->rstart; 1850 PetscInt nztot,nzA,nzB,lrow,rstart = matin->rmap->rstart,rend = matin->rmap->rend; 1851 PetscInt *cmap,*idx_p; 1852 1853 PetscFunctionBegin; 1854 if (mat->getrowactive) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Already active"); 1855 mat->getrowactive = PETSC_TRUE; 1856 1857 if (!mat->rowvalues && (idx || v)) { 1858 /* 1859 allocate enough space to hold information from the longest row. 1860 */ 1861 Mat_SeqAIJ *Aa = (Mat_SeqAIJ*)mat->A->data,*Ba = (Mat_SeqAIJ*)mat->B->data; 1862 PetscInt max = 1,tmp; 1863 for (i=0; i<matin->rmap->n; i++) { 1864 tmp = Aa->i[i+1] - Aa->i[i] + Ba->i[i+1] - Ba->i[i]; 1865 if (max < tmp) max = tmp; 1866 } 1867 ierr = PetscMalloc2(max,&mat->rowvalues,max,&mat->rowindices);CHKERRQ(ierr); 1868 } 1869 1870 if (row < rstart || row >= rend) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Only local rows"); 1871 lrow = row - rstart; 1872 1873 pvA = &vworkA; pcA = &cworkA; pvB = &vworkB; pcB = &cworkB; 1874 if (!v) {pvA = 0; pvB = 0;} 1875 if (!idx) {pcA = 0; if (!v) pcB = 0;} 1876 ierr = (*mat->A->ops->getrow)(mat->A,lrow,&nzA,pcA,pvA);CHKERRQ(ierr); 1877 ierr = (*mat->B->ops->getrow)(mat->B,lrow,&nzB,pcB,pvB);CHKERRQ(ierr); 1878 nztot = nzA + nzB; 1879 1880 cmap = mat->garray; 1881 if (v || idx) { 1882 if (nztot) { 1883 /* Sort by increasing column numbers, assuming A and B already sorted */ 1884 PetscInt imark = -1; 1885 if (v) { 1886 *v = v_p = mat->rowvalues; 1887 for (i=0; i<nzB; i++) { 1888 if (cmap[cworkB[i]] < cstart) v_p[i] = vworkB[i]; 1889 else break; 1890 } 1891 imark = i; 1892 for (i=0; i<nzA; i++) v_p[imark+i] = vworkA[i]; 1893 for (i=imark; i<nzB; i++) v_p[nzA+i] = vworkB[i]; 1894 } 1895 if (idx) { 1896 *idx = idx_p = mat->rowindices; 1897 if (imark > -1) { 1898 for (i=0; i<imark; i++) { 1899 idx_p[i] = cmap[cworkB[i]]; 1900 } 1901 } else { 1902 for (i=0; i<nzB; i++) { 1903 if (cmap[cworkB[i]] < cstart) idx_p[i] = cmap[cworkB[i]]; 1904 else break; 1905 } 1906 imark = i; 1907 } 1908 for (i=0; i<nzA; i++) idx_p[imark+i] = cstart + cworkA[i]; 1909 for (i=imark; i<nzB; i++) idx_p[nzA+i] = cmap[cworkB[i]]; 1910 } 1911 } else { 1912 if (idx) *idx = 0; 1913 if (v) *v = 0; 1914 } 1915 } 1916 *nz = nztot; 1917 ierr = (*mat->A->ops->restorerow)(mat->A,lrow,&nzA,pcA,pvA);CHKERRQ(ierr); 1918 ierr = (*mat->B->ops->restorerow)(mat->B,lrow,&nzB,pcB,pvB);CHKERRQ(ierr); 1919 PetscFunctionReturn(0); 1920 } 1921 1922 PetscErrorCode MatRestoreRow_MPIAIJ(Mat mat,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v) 1923 { 1924 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 1925 1926 PetscFunctionBegin; 1927 if (!aij->getrowactive) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"MatGetRow() must be called first"); 1928 aij->getrowactive = PETSC_FALSE; 1929 PetscFunctionReturn(0); 1930 } 1931 1932 PetscErrorCode MatNorm_MPIAIJ(Mat mat,NormType type,PetscReal *norm) 1933 { 1934 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 1935 Mat_SeqAIJ *amat = (Mat_SeqAIJ*)aij->A->data,*bmat = (Mat_SeqAIJ*)aij->B->data; 1936 PetscErrorCode ierr; 1937 PetscInt i,j,cstart = mat->cmap->rstart; 1938 PetscReal sum = 0.0; 1939 MatScalar *v; 1940 1941 PetscFunctionBegin; 1942 if (aij->size == 1) { 1943 ierr = MatNorm(aij->A,type,norm);CHKERRQ(ierr); 1944 } else { 1945 if (type == NORM_FROBENIUS) { 1946 v = amat->a; 1947 for (i=0; i<amat->nz; i++) { 1948 sum += PetscRealPart(PetscConj(*v)*(*v)); v++; 1949 } 1950 v = bmat->a; 1951 for (i=0; i<bmat->nz; i++) { 1952 sum += PetscRealPart(PetscConj(*v)*(*v)); v++; 1953 } 1954 ierr = MPIU_Allreduce(&sum,norm,1,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1955 *norm = PetscSqrtReal(*norm); 1956 ierr = PetscLogFlops(2*amat->nz+2*bmat->nz);CHKERRQ(ierr); 1957 } else if (type == NORM_1) { /* max column norm */ 1958 PetscReal *tmp,*tmp2; 1959 PetscInt *jj,*garray = aij->garray; 1960 ierr = PetscCalloc1(mat->cmap->N+1,&tmp);CHKERRQ(ierr); 1961 ierr = PetscMalloc1(mat->cmap->N+1,&tmp2);CHKERRQ(ierr); 1962 *norm = 0.0; 1963 v = amat->a; jj = amat->j; 1964 for (j=0; j<amat->nz; j++) { 1965 tmp[cstart + *jj++] += PetscAbsScalar(*v); v++; 1966 } 1967 v = bmat->a; jj = bmat->j; 1968 for (j=0; j<bmat->nz; j++) { 1969 tmp[garray[*jj++]] += PetscAbsScalar(*v); v++; 1970 } 1971 ierr = MPIU_Allreduce(tmp,tmp2,mat->cmap->N,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1972 for (j=0; j<mat->cmap->N; j++) { 1973 if (tmp2[j] > *norm) *norm = tmp2[j]; 1974 } 1975 ierr = PetscFree(tmp);CHKERRQ(ierr); 1976 ierr = PetscFree(tmp2);CHKERRQ(ierr); 1977 ierr = PetscLogFlops(PetscMax(amat->nz+bmat->nz-1,0));CHKERRQ(ierr); 1978 } else if (type == NORM_INFINITY) { /* max row norm */ 1979 PetscReal ntemp = 0.0; 1980 for (j=0; j<aij->A->rmap->n; j++) { 1981 v = amat->a + amat->i[j]; 1982 sum = 0.0; 1983 for (i=0; i<amat->i[j+1]-amat->i[j]; i++) { 1984 sum += PetscAbsScalar(*v); v++; 1985 } 1986 v = bmat->a + bmat->i[j]; 1987 for (i=0; i<bmat->i[j+1]-bmat->i[j]; i++) { 1988 sum += PetscAbsScalar(*v); v++; 1989 } 1990 if (sum > ntemp) ntemp = sum; 1991 } 1992 ierr = MPIU_Allreduce(&ntemp,norm,1,MPIU_REAL,MPIU_MAX,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1993 ierr = PetscLogFlops(PetscMax(amat->nz+bmat->nz-1,0));CHKERRQ(ierr); 1994 } else SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"No support for two norm"); 1995 } 1996 PetscFunctionReturn(0); 1997 } 1998 1999 PetscErrorCode MatTranspose_MPIAIJ(Mat A,MatReuse reuse,Mat *matout) 2000 { 2001 Mat_MPIAIJ *a =(Mat_MPIAIJ*)A->data,*b; 2002 Mat_SeqAIJ *Aloc =(Mat_SeqAIJ*)a->A->data,*Bloc=(Mat_SeqAIJ*)a->B->data,*sub_B_diag; 2003 PetscInt M = A->rmap->N,N=A->cmap->N,ma,na,mb,nb,*ai,*aj,*bi,*bj,row,*cols,*cols_tmp,*B_diag_ilen,*B_diag_i,i,ncol,A_diag_ncol; 2004 PetscErrorCode ierr; 2005 Mat B,A_diag,*B_diag; 2006 MatScalar *array; 2007 2008 PetscFunctionBegin; 2009 ma = A->rmap->n; na = A->cmap->n; mb = a->B->rmap->n; nb = a->B->cmap->n; 2010 ai = Aloc->i; aj = Aloc->j; 2011 bi = Bloc->i; bj = Bloc->j; 2012 if (reuse == MAT_INITIAL_MATRIX || *matout == A) { 2013 PetscInt *d_nnz,*g_nnz,*o_nnz; 2014 PetscSFNode *oloc; 2015 PETSC_UNUSED PetscSF sf; 2016 2017 ierr = PetscMalloc4(na,&d_nnz,na,&o_nnz,nb,&g_nnz,nb,&oloc);CHKERRQ(ierr); 2018 /* compute d_nnz for preallocation */ 2019 ierr = PetscMemzero(d_nnz,na*sizeof(PetscInt));CHKERRQ(ierr); 2020 for (i=0; i<ai[ma]; i++) { 2021 d_nnz[aj[i]]++; 2022 } 2023 /* compute local off-diagonal contributions */ 2024 ierr = PetscMemzero(g_nnz,nb*sizeof(PetscInt));CHKERRQ(ierr); 2025 for (i=0; i<bi[ma]; i++) g_nnz[bj[i]]++; 2026 /* map those to global */ 2027 ierr = PetscSFCreate(PetscObjectComm((PetscObject)A),&sf);CHKERRQ(ierr); 2028 ierr = PetscSFSetGraphLayout(sf,A->cmap,nb,NULL,PETSC_USE_POINTER,a->garray);CHKERRQ(ierr); 2029 ierr = PetscSFSetFromOptions(sf);CHKERRQ(ierr); 2030 ierr = PetscMemzero(o_nnz,na*sizeof(PetscInt));CHKERRQ(ierr); 2031 ierr = PetscSFReduceBegin(sf,MPIU_INT,g_nnz,o_nnz,MPIU_SUM);CHKERRQ(ierr); 2032 ierr = PetscSFReduceEnd(sf,MPIU_INT,g_nnz,o_nnz,MPIU_SUM);CHKERRQ(ierr); 2033 ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 2034 2035 ierr = MatCreate(PetscObjectComm((PetscObject)A),&B);CHKERRQ(ierr); 2036 ierr = MatSetSizes(B,A->cmap->n,A->rmap->n,N,M);CHKERRQ(ierr); 2037 ierr = MatSetBlockSizes(B,PetscAbs(A->cmap->bs),PetscAbs(A->rmap->bs));CHKERRQ(ierr); 2038 ierr = MatSetType(B,((PetscObject)A)->type_name);CHKERRQ(ierr); 2039 ierr = MatMPIAIJSetPreallocation(B,0,d_nnz,0,o_nnz);CHKERRQ(ierr); 2040 ierr = PetscFree4(d_nnz,o_nnz,g_nnz,oloc);CHKERRQ(ierr); 2041 } else { 2042 B = *matout; 2043 ierr = MatSetOption(B,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_TRUE);CHKERRQ(ierr); 2044 } 2045 2046 b = (Mat_MPIAIJ*)B->data; 2047 A_diag = a->A; 2048 B_diag = &b->A; 2049 sub_B_diag = (Mat_SeqAIJ*)(*B_diag)->data; 2050 A_diag_ncol = A_diag->cmap->N; 2051 B_diag_ilen = sub_B_diag->ilen; 2052 B_diag_i = sub_B_diag->i; 2053 2054 /* Set ilen for diagonal of B */ 2055 for (i=0; i<A_diag_ncol; i++) { 2056 B_diag_ilen[i] = B_diag_i[i+1] - B_diag_i[i]; 2057 } 2058 2059 /* Transpose the diagonal part of the matrix. In contrast to the offdiagonal part, this can be done 2060 very quickly (=without using MatSetValues), because all writes are local. */ 2061 ierr = MatTranspose(A_diag,MAT_REUSE_MATRIX,B_diag);CHKERRQ(ierr); 2062 2063 /* copy over the B part */ 2064 ierr = PetscCalloc1(bi[mb],&cols);CHKERRQ(ierr); 2065 array = Bloc->a; 2066 row = A->rmap->rstart; 2067 for (i=0; i<bi[mb]; i++) cols[i] = a->garray[bj[i]]; 2068 cols_tmp = cols; 2069 for (i=0; i<mb; i++) { 2070 ncol = bi[i+1]-bi[i]; 2071 ierr = MatSetValues(B,ncol,cols_tmp,1,&row,array,INSERT_VALUES);CHKERRQ(ierr); 2072 row++; 2073 array += ncol; cols_tmp += ncol; 2074 } 2075 ierr = PetscFree(cols);CHKERRQ(ierr); 2076 2077 ierr = MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2078 ierr = MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2079 if (reuse == MAT_INITIAL_MATRIX || reuse == MAT_REUSE_MATRIX) { 2080 *matout = B; 2081 } else { 2082 ierr = MatHeaderMerge(A,&B);CHKERRQ(ierr); 2083 } 2084 PetscFunctionReturn(0); 2085 } 2086 2087 PetscErrorCode MatDiagonalScale_MPIAIJ(Mat mat,Vec ll,Vec rr) 2088 { 2089 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 2090 Mat a = aij->A,b = aij->B; 2091 PetscErrorCode ierr; 2092 PetscInt s1,s2,s3; 2093 2094 PetscFunctionBegin; 2095 ierr = MatGetLocalSize(mat,&s2,&s3);CHKERRQ(ierr); 2096 if (rr) { 2097 ierr = VecGetLocalSize(rr,&s1);CHKERRQ(ierr); 2098 if (s1!=s3) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"right vector non-conforming local size"); 2099 /* Overlap communication with computation. */ 2100 ierr = VecScatterBegin(aij->Mvctx,rr,aij->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2101 } 2102 if (ll) { 2103 ierr = VecGetLocalSize(ll,&s1);CHKERRQ(ierr); 2104 if (s1!=s2) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"left vector non-conforming local size"); 2105 ierr = (*b->ops->diagonalscale)(b,ll,0);CHKERRQ(ierr); 2106 } 2107 /* scale the diagonal block */ 2108 ierr = (*a->ops->diagonalscale)(a,ll,rr);CHKERRQ(ierr); 2109 2110 if (rr) { 2111 /* Do a scatter end and then right scale the off-diagonal block */ 2112 ierr = VecScatterEnd(aij->Mvctx,rr,aij->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2113 ierr = (*b->ops->diagonalscale)(b,0,aij->lvec);CHKERRQ(ierr); 2114 } 2115 PetscFunctionReturn(0); 2116 } 2117 2118 PetscErrorCode MatSetUnfactored_MPIAIJ(Mat A) 2119 { 2120 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 2121 PetscErrorCode ierr; 2122 2123 PetscFunctionBegin; 2124 ierr = MatSetUnfactored(a->A);CHKERRQ(ierr); 2125 PetscFunctionReturn(0); 2126 } 2127 2128 PetscErrorCode MatEqual_MPIAIJ(Mat A,Mat B,PetscBool *flag) 2129 { 2130 Mat_MPIAIJ *matB = (Mat_MPIAIJ*)B->data,*matA = (Mat_MPIAIJ*)A->data; 2131 Mat a,b,c,d; 2132 PetscBool flg; 2133 PetscErrorCode ierr; 2134 2135 PetscFunctionBegin; 2136 a = matA->A; b = matA->B; 2137 c = matB->A; d = matB->B; 2138 2139 ierr = MatEqual(a,c,&flg);CHKERRQ(ierr); 2140 if (flg) { 2141 ierr = MatEqual(b,d,&flg);CHKERRQ(ierr); 2142 } 2143 ierr = MPIU_Allreduce(&flg,flag,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)A));CHKERRQ(ierr); 2144 PetscFunctionReturn(0); 2145 } 2146 2147 PetscErrorCode MatCopy_MPIAIJ(Mat A,Mat B,MatStructure str) 2148 { 2149 PetscErrorCode ierr; 2150 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 2151 Mat_MPIAIJ *b = (Mat_MPIAIJ*)B->data; 2152 2153 PetscFunctionBegin; 2154 /* If the two matrices don't have the same copy implementation, they aren't compatible for fast copy. */ 2155 if ((str != SAME_NONZERO_PATTERN) || (A->ops->copy != B->ops->copy)) { 2156 /* because of the column compression in the off-processor part of the matrix a->B, 2157 the number of columns in a->B and b->B may be different, hence we cannot call 2158 the MatCopy() directly on the two parts. If need be, we can provide a more 2159 efficient copy than the MatCopy_Basic() by first uncompressing the a->B matrices 2160 then copying the submatrices */ 2161 ierr = MatCopy_Basic(A,B,str);CHKERRQ(ierr); 2162 } else { 2163 ierr = MatCopy(a->A,b->A,str);CHKERRQ(ierr); 2164 ierr = MatCopy(a->B,b->B,str);CHKERRQ(ierr); 2165 } 2166 ierr = PetscObjectStateIncrease((PetscObject)B);CHKERRQ(ierr); 2167 PetscFunctionReturn(0); 2168 } 2169 2170 PetscErrorCode MatSetUp_MPIAIJ(Mat A) 2171 { 2172 PetscErrorCode ierr; 2173 2174 PetscFunctionBegin; 2175 ierr = MatMPIAIJSetPreallocation(A,PETSC_DEFAULT,0,PETSC_DEFAULT,0);CHKERRQ(ierr); 2176 PetscFunctionReturn(0); 2177 } 2178 2179 /* 2180 Computes the number of nonzeros per row needed for preallocation when X and Y 2181 have different nonzero structure. 2182 */ 2183 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) 2184 { 2185 PetscInt i,j,k,nzx,nzy; 2186 2187 PetscFunctionBegin; 2188 /* Set the number of nonzeros in the new matrix */ 2189 for (i=0; i<m; i++) { 2190 const PetscInt *xjj = xj+xi[i],*yjj = yj+yi[i]; 2191 nzx = xi[i+1] - xi[i]; 2192 nzy = yi[i+1] - yi[i]; 2193 nnz[i] = 0; 2194 for (j=0,k=0; j<nzx; j++) { /* Point in X */ 2195 for (; k<nzy && yltog[yjj[k]]<xltog[xjj[j]]; k++) nnz[i]++; /* Catch up to X */ 2196 if (k<nzy && yltog[yjj[k]]==xltog[xjj[j]]) k++; /* Skip duplicate */ 2197 nnz[i]++; 2198 } 2199 for (; k<nzy; k++) nnz[i]++; 2200 } 2201 PetscFunctionReturn(0); 2202 } 2203 2204 /* This is the same as MatAXPYGetPreallocation_SeqAIJ, except that the local-to-global map is provided */ 2205 static PetscErrorCode MatAXPYGetPreallocation_MPIAIJ(Mat Y,const PetscInt *yltog,Mat X,const PetscInt *xltog,PetscInt *nnz) 2206 { 2207 PetscErrorCode ierr; 2208 PetscInt m = Y->rmap->N; 2209 Mat_SeqAIJ *x = (Mat_SeqAIJ*)X->data; 2210 Mat_SeqAIJ *y = (Mat_SeqAIJ*)Y->data; 2211 2212 PetscFunctionBegin; 2213 ierr = MatAXPYGetPreallocation_MPIX_private(m,x->i,x->j,xltog,y->i,y->j,yltog,nnz);CHKERRQ(ierr); 2214 PetscFunctionReturn(0); 2215 } 2216 2217 PetscErrorCode MatAXPY_MPIAIJ(Mat Y,PetscScalar a,Mat X,MatStructure str) 2218 { 2219 PetscErrorCode ierr; 2220 Mat_MPIAIJ *xx = (Mat_MPIAIJ*)X->data,*yy = (Mat_MPIAIJ*)Y->data; 2221 PetscBLASInt bnz,one=1; 2222 Mat_SeqAIJ *x,*y; 2223 2224 PetscFunctionBegin; 2225 if (str == SAME_NONZERO_PATTERN) { 2226 PetscScalar alpha = a; 2227 x = (Mat_SeqAIJ*)xx->A->data; 2228 ierr = PetscBLASIntCast(x->nz,&bnz);CHKERRQ(ierr); 2229 y = (Mat_SeqAIJ*)yy->A->data; 2230 PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&bnz,&alpha,x->a,&one,y->a,&one)); 2231 x = (Mat_SeqAIJ*)xx->B->data; 2232 y = (Mat_SeqAIJ*)yy->B->data; 2233 ierr = PetscBLASIntCast(x->nz,&bnz);CHKERRQ(ierr); 2234 PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&bnz,&alpha,x->a,&one,y->a,&one)); 2235 ierr = PetscObjectStateIncrease((PetscObject)Y);CHKERRQ(ierr); 2236 } else if (str == SUBSET_NONZERO_PATTERN) { /* nonzeros of X is a subset of Y's */ 2237 ierr = MatAXPY_Basic(Y,a,X,str);CHKERRQ(ierr); 2238 } else { 2239 Mat B; 2240 PetscInt *nnz_d,*nnz_o; 2241 ierr = PetscMalloc1(yy->A->rmap->N,&nnz_d);CHKERRQ(ierr); 2242 ierr = PetscMalloc1(yy->B->rmap->N,&nnz_o);CHKERRQ(ierr); 2243 ierr = MatCreate(PetscObjectComm((PetscObject)Y),&B);CHKERRQ(ierr); 2244 ierr = PetscObjectSetName((PetscObject)B,((PetscObject)Y)->name);CHKERRQ(ierr); 2245 ierr = MatSetSizes(B,Y->rmap->n,Y->cmap->n,Y->rmap->N,Y->cmap->N);CHKERRQ(ierr); 2246 ierr = MatSetBlockSizesFromMats(B,Y,Y);CHKERRQ(ierr); 2247 ierr = MatSetType(B,MATMPIAIJ);CHKERRQ(ierr); 2248 ierr = MatAXPYGetPreallocation_SeqAIJ(yy->A,xx->A,nnz_d);CHKERRQ(ierr); 2249 ierr = MatAXPYGetPreallocation_MPIAIJ(yy->B,yy->garray,xx->B,xx->garray,nnz_o);CHKERRQ(ierr); 2250 ierr = MatMPIAIJSetPreallocation(B,0,nnz_d,0,nnz_o);CHKERRQ(ierr); 2251 ierr = MatAXPY_BasicWithPreallocation(B,Y,a,X,str);CHKERRQ(ierr); 2252 ierr = MatHeaderReplace(Y,&B);CHKERRQ(ierr); 2253 ierr = PetscFree(nnz_d);CHKERRQ(ierr); 2254 ierr = PetscFree(nnz_o);CHKERRQ(ierr); 2255 } 2256 PetscFunctionReturn(0); 2257 } 2258 2259 extern PetscErrorCode MatConjugate_SeqAIJ(Mat); 2260 2261 PetscErrorCode MatConjugate_MPIAIJ(Mat mat) 2262 { 2263 #if defined(PETSC_USE_COMPLEX) 2264 PetscErrorCode ierr; 2265 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 2266 2267 PetscFunctionBegin; 2268 ierr = MatConjugate_SeqAIJ(aij->A);CHKERRQ(ierr); 2269 ierr = MatConjugate_SeqAIJ(aij->B);CHKERRQ(ierr); 2270 #else 2271 PetscFunctionBegin; 2272 #endif 2273 PetscFunctionReturn(0); 2274 } 2275 2276 PetscErrorCode MatRealPart_MPIAIJ(Mat A) 2277 { 2278 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 2279 PetscErrorCode ierr; 2280 2281 PetscFunctionBegin; 2282 ierr = MatRealPart(a->A);CHKERRQ(ierr); 2283 ierr = MatRealPart(a->B);CHKERRQ(ierr); 2284 PetscFunctionReturn(0); 2285 } 2286 2287 PetscErrorCode MatImaginaryPart_MPIAIJ(Mat A) 2288 { 2289 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 2290 PetscErrorCode ierr; 2291 2292 PetscFunctionBegin; 2293 ierr = MatImaginaryPart(a->A);CHKERRQ(ierr); 2294 ierr = MatImaginaryPart(a->B);CHKERRQ(ierr); 2295 PetscFunctionReturn(0); 2296 } 2297 2298 PetscErrorCode MatGetRowMaxAbs_MPIAIJ(Mat A, Vec v, PetscInt idx[]) 2299 { 2300 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 2301 PetscErrorCode ierr; 2302 PetscInt i,*idxb = 0; 2303 PetscScalar *va,*vb; 2304 Vec vtmp; 2305 2306 PetscFunctionBegin; 2307 ierr = MatGetRowMaxAbs(a->A,v,idx);CHKERRQ(ierr); 2308 ierr = VecGetArray(v,&va);CHKERRQ(ierr); 2309 if (idx) { 2310 for (i=0; i<A->rmap->n; i++) { 2311 if (PetscAbsScalar(va[i])) idx[i] += A->cmap->rstart; 2312 } 2313 } 2314 2315 ierr = VecCreateSeq(PETSC_COMM_SELF,A->rmap->n,&vtmp);CHKERRQ(ierr); 2316 if (idx) { 2317 ierr = PetscMalloc1(A->rmap->n,&idxb);CHKERRQ(ierr); 2318 } 2319 ierr = MatGetRowMaxAbs(a->B,vtmp,idxb);CHKERRQ(ierr); 2320 ierr = VecGetArray(vtmp,&vb);CHKERRQ(ierr); 2321 2322 for (i=0; i<A->rmap->n; i++) { 2323 if (PetscAbsScalar(va[i]) < PetscAbsScalar(vb[i])) { 2324 va[i] = vb[i]; 2325 if (idx) idx[i] = a->garray[idxb[i]]; 2326 } 2327 } 2328 2329 ierr = VecRestoreArray(v,&va);CHKERRQ(ierr); 2330 ierr = VecRestoreArray(vtmp,&vb);CHKERRQ(ierr); 2331 ierr = PetscFree(idxb);CHKERRQ(ierr); 2332 ierr = VecDestroy(&vtmp);CHKERRQ(ierr); 2333 PetscFunctionReturn(0); 2334 } 2335 2336 PetscErrorCode MatGetRowMinAbs_MPIAIJ(Mat A, Vec v, PetscInt idx[]) 2337 { 2338 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 2339 PetscErrorCode ierr; 2340 PetscInt i,*idxb = 0; 2341 PetscScalar *va,*vb; 2342 Vec vtmp; 2343 2344 PetscFunctionBegin; 2345 ierr = MatGetRowMinAbs(a->A,v,idx);CHKERRQ(ierr); 2346 ierr = VecGetArray(v,&va);CHKERRQ(ierr); 2347 if (idx) { 2348 for (i=0; i<A->cmap->n; i++) { 2349 if (PetscAbsScalar(va[i])) idx[i] += A->cmap->rstart; 2350 } 2351 } 2352 2353 ierr = VecCreateSeq(PETSC_COMM_SELF,A->rmap->n,&vtmp);CHKERRQ(ierr); 2354 if (idx) { 2355 ierr = PetscMalloc1(A->rmap->n,&idxb);CHKERRQ(ierr); 2356 } 2357 ierr = MatGetRowMinAbs(a->B,vtmp,idxb);CHKERRQ(ierr); 2358 ierr = VecGetArray(vtmp,&vb);CHKERRQ(ierr); 2359 2360 for (i=0; i<A->rmap->n; i++) { 2361 if (PetscAbsScalar(va[i]) > PetscAbsScalar(vb[i])) { 2362 va[i] = vb[i]; 2363 if (idx) idx[i] = a->garray[idxb[i]]; 2364 } 2365 } 2366 2367 ierr = VecRestoreArray(v,&va);CHKERRQ(ierr); 2368 ierr = VecRestoreArray(vtmp,&vb);CHKERRQ(ierr); 2369 ierr = PetscFree(idxb);CHKERRQ(ierr); 2370 ierr = VecDestroy(&vtmp);CHKERRQ(ierr); 2371 PetscFunctionReturn(0); 2372 } 2373 2374 PetscErrorCode MatGetRowMin_MPIAIJ(Mat A, Vec v, PetscInt idx[]) 2375 { 2376 Mat_MPIAIJ *mat = (Mat_MPIAIJ*) A->data; 2377 PetscInt n = A->rmap->n; 2378 PetscInt cstart = A->cmap->rstart; 2379 PetscInt *cmap = mat->garray; 2380 PetscInt *diagIdx, *offdiagIdx; 2381 Vec diagV, offdiagV; 2382 PetscScalar *a, *diagA, *offdiagA; 2383 PetscInt r; 2384 PetscErrorCode ierr; 2385 2386 PetscFunctionBegin; 2387 ierr = PetscMalloc2(n,&diagIdx,n,&offdiagIdx);CHKERRQ(ierr); 2388 ierr = VecCreateSeq(PetscObjectComm((PetscObject)A), n, &diagV);CHKERRQ(ierr); 2389 ierr = VecCreateSeq(PetscObjectComm((PetscObject)A), n, &offdiagV);CHKERRQ(ierr); 2390 ierr = MatGetRowMin(mat->A, diagV, diagIdx);CHKERRQ(ierr); 2391 ierr = MatGetRowMin(mat->B, offdiagV, offdiagIdx);CHKERRQ(ierr); 2392 ierr = VecGetArray(v, &a);CHKERRQ(ierr); 2393 ierr = VecGetArray(diagV, &diagA);CHKERRQ(ierr); 2394 ierr = VecGetArray(offdiagV, &offdiagA);CHKERRQ(ierr); 2395 for (r = 0; r < n; ++r) { 2396 if (PetscAbsScalar(diagA[r]) <= PetscAbsScalar(offdiagA[r])) { 2397 a[r] = diagA[r]; 2398 idx[r] = cstart + diagIdx[r]; 2399 } else { 2400 a[r] = offdiagA[r]; 2401 idx[r] = cmap[offdiagIdx[r]]; 2402 } 2403 } 2404 ierr = VecRestoreArray(v, &a);CHKERRQ(ierr); 2405 ierr = VecRestoreArray(diagV, &diagA);CHKERRQ(ierr); 2406 ierr = VecRestoreArray(offdiagV, &offdiagA);CHKERRQ(ierr); 2407 ierr = VecDestroy(&diagV);CHKERRQ(ierr); 2408 ierr = VecDestroy(&offdiagV);CHKERRQ(ierr); 2409 ierr = PetscFree2(diagIdx, offdiagIdx);CHKERRQ(ierr); 2410 PetscFunctionReturn(0); 2411 } 2412 2413 PetscErrorCode MatGetRowMax_MPIAIJ(Mat A, Vec v, PetscInt idx[]) 2414 { 2415 Mat_MPIAIJ *mat = (Mat_MPIAIJ*) A->data; 2416 PetscInt n = A->rmap->n; 2417 PetscInt cstart = A->cmap->rstart; 2418 PetscInt *cmap = mat->garray; 2419 PetscInt *diagIdx, *offdiagIdx; 2420 Vec diagV, offdiagV; 2421 PetscScalar *a, *diagA, *offdiagA; 2422 PetscInt r; 2423 PetscErrorCode ierr; 2424 2425 PetscFunctionBegin; 2426 ierr = PetscMalloc2(n,&diagIdx,n,&offdiagIdx);CHKERRQ(ierr); 2427 ierr = VecCreateSeq(PETSC_COMM_SELF, n, &diagV);CHKERRQ(ierr); 2428 ierr = VecCreateSeq(PETSC_COMM_SELF, n, &offdiagV);CHKERRQ(ierr); 2429 ierr = MatGetRowMax(mat->A, diagV, diagIdx);CHKERRQ(ierr); 2430 ierr = MatGetRowMax(mat->B, offdiagV, offdiagIdx);CHKERRQ(ierr); 2431 ierr = VecGetArray(v, &a);CHKERRQ(ierr); 2432 ierr = VecGetArray(diagV, &diagA);CHKERRQ(ierr); 2433 ierr = VecGetArray(offdiagV, &offdiagA);CHKERRQ(ierr); 2434 for (r = 0; r < n; ++r) { 2435 if (PetscAbsScalar(diagA[r]) >= PetscAbsScalar(offdiagA[r])) { 2436 a[r] = diagA[r]; 2437 idx[r] = cstart + diagIdx[r]; 2438 } else { 2439 a[r] = offdiagA[r]; 2440 idx[r] = cmap[offdiagIdx[r]]; 2441 } 2442 } 2443 ierr = VecRestoreArray(v, &a);CHKERRQ(ierr); 2444 ierr = VecRestoreArray(diagV, &diagA);CHKERRQ(ierr); 2445 ierr = VecRestoreArray(offdiagV, &offdiagA);CHKERRQ(ierr); 2446 ierr = VecDestroy(&diagV);CHKERRQ(ierr); 2447 ierr = VecDestroy(&offdiagV);CHKERRQ(ierr); 2448 ierr = PetscFree2(diagIdx, offdiagIdx);CHKERRQ(ierr); 2449 PetscFunctionReturn(0); 2450 } 2451 2452 PetscErrorCode MatGetSeqNonzeroStructure_MPIAIJ(Mat mat,Mat *newmat) 2453 { 2454 PetscErrorCode ierr; 2455 Mat *dummy; 2456 2457 PetscFunctionBegin; 2458 ierr = MatCreateSubMatrix_MPIAIJ_All(mat,MAT_DO_NOT_GET_VALUES,MAT_INITIAL_MATRIX,&dummy);CHKERRQ(ierr); 2459 *newmat = *dummy; 2460 ierr = PetscFree(dummy);CHKERRQ(ierr); 2461 PetscFunctionReturn(0); 2462 } 2463 2464 PetscErrorCode MatInvertBlockDiagonal_MPIAIJ(Mat A,const PetscScalar **values) 2465 { 2466 Mat_MPIAIJ *a = (Mat_MPIAIJ*) A->data; 2467 PetscErrorCode ierr; 2468 2469 PetscFunctionBegin; 2470 ierr = MatInvertBlockDiagonal(a->A,values);CHKERRQ(ierr); 2471 A->factorerrortype = a->A->factorerrortype; 2472 PetscFunctionReturn(0); 2473 } 2474 2475 static PetscErrorCode MatSetRandom_MPIAIJ(Mat x,PetscRandom rctx) 2476 { 2477 PetscErrorCode ierr; 2478 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)x->data; 2479 2480 PetscFunctionBegin; 2481 ierr = MatSetRandom(aij->A,rctx);CHKERRQ(ierr); 2482 ierr = MatSetRandom(aij->B,rctx);CHKERRQ(ierr); 2483 ierr = MatAssemblyBegin(x,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2484 ierr = MatAssemblyEnd(x,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2485 PetscFunctionReturn(0); 2486 } 2487 2488 PetscErrorCode MatMPIAIJSetUseScalableIncreaseOverlap_MPIAIJ(Mat A,PetscBool sc) 2489 { 2490 PetscFunctionBegin; 2491 if (sc) A->ops->increaseoverlap = MatIncreaseOverlap_MPIAIJ_Scalable; 2492 else A->ops->increaseoverlap = MatIncreaseOverlap_MPIAIJ; 2493 PetscFunctionReturn(0); 2494 } 2495 2496 /*@ 2497 MatMPIAIJSetUseScalableIncreaseOverlap - Determine if the matrix uses a scalable algorithm to compute the overlap 2498 2499 Collective on Mat 2500 2501 Input Parameters: 2502 + A - the matrix 2503 - sc - PETSC_TRUE indicates use the scalable algorithm (default is not to use the scalable algorithm) 2504 2505 Level: advanced 2506 2507 @*/ 2508 PetscErrorCode MatMPIAIJSetUseScalableIncreaseOverlap(Mat A,PetscBool sc) 2509 { 2510 PetscErrorCode ierr; 2511 2512 PetscFunctionBegin; 2513 ierr = PetscTryMethod(A,"MatMPIAIJSetUseScalableIncreaseOverlap_C",(Mat,PetscBool),(A,sc));CHKERRQ(ierr); 2514 PetscFunctionReturn(0); 2515 } 2516 2517 PetscErrorCode MatSetFromOptions_MPIAIJ(PetscOptionItems *PetscOptionsObject,Mat A) 2518 { 2519 PetscErrorCode ierr; 2520 PetscBool sc = PETSC_FALSE,flg; 2521 2522 PetscFunctionBegin; 2523 ierr = PetscOptionsHead(PetscOptionsObject,"MPIAIJ options");CHKERRQ(ierr); 2524 if (A->ops->increaseoverlap == MatIncreaseOverlap_MPIAIJ_Scalable) sc = PETSC_TRUE; 2525 ierr = PetscOptionsBool("-mat_increase_overlap_scalable","Use a scalable algorithm to compute the overlap","MatIncreaseOverlap",sc,&sc,&flg);CHKERRQ(ierr); 2526 if (flg) { 2527 ierr = MatMPIAIJSetUseScalableIncreaseOverlap(A,sc);CHKERRQ(ierr); 2528 } 2529 ierr = PetscOptionsTail();CHKERRQ(ierr); 2530 PetscFunctionReturn(0); 2531 } 2532 2533 PetscErrorCode MatShift_MPIAIJ(Mat Y,PetscScalar a) 2534 { 2535 PetscErrorCode ierr; 2536 Mat_MPIAIJ *maij = (Mat_MPIAIJ*)Y->data; 2537 Mat_SeqAIJ *aij = (Mat_SeqAIJ*)maij->A->data; 2538 2539 PetscFunctionBegin; 2540 if (!Y->preallocated) { 2541 ierr = MatMPIAIJSetPreallocation(Y,1,NULL,0,NULL);CHKERRQ(ierr); 2542 } else if (!aij->nz) { 2543 PetscInt nonew = aij->nonew; 2544 ierr = MatSeqAIJSetPreallocation(maij->A,1,NULL);CHKERRQ(ierr); 2545 aij->nonew = nonew; 2546 } 2547 ierr = MatShift_Basic(Y,a);CHKERRQ(ierr); 2548 PetscFunctionReturn(0); 2549 } 2550 2551 PetscErrorCode MatMissingDiagonal_MPIAIJ(Mat A,PetscBool *missing,PetscInt *d) 2552 { 2553 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 2554 PetscErrorCode ierr; 2555 2556 PetscFunctionBegin; 2557 if (A->rmap->n != A->cmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Only works for square matrices"); 2558 ierr = MatMissingDiagonal(a->A,missing,d);CHKERRQ(ierr); 2559 if (d) { 2560 PetscInt rstart; 2561 ierr = MatGetOwnershipRange(A,&rstart,NULL);CHKERRQ(ierr); 2562 *d += rstart; 2563 2564 } 2565 PetscFunctionReturn(0); 2566 } 2567 2568 PetscErrorCode MatInvertVariableBlockDiagonal_MPIAIJ(Mat A,PetscInt nblocks,const PetscInt *bsizes,PetscScalar *diag) 2569 { 2570 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 2571 PetscErrorCode ierr; 2572 2573 PetscFunctionBegin; 2574 ierr = MatInvertVariableBlockDiagonal(a->A,nblocks,bsizes,diag);CHKERRQ(ierr); 2575 PetscFunctionReturn(0); 2576 } 2577 2578 /* -------------------------------------------------------------------*/ 2579 static struct _MatOps MatOps_Values = {MatSetValues_MPIAIJ, 2580 MatGetRow_MPIAIJ, 2581 MatRestoreRow_MPIAIJ, 2582 MatMult_MPIAIJ, 2583 /* 4*/ MatMultAdd_MPIAIJ, 2584 MatMultTranspose_MPIAIJ, 2585 MatMultTransposeAdd_MPIAIJ, 2586 0, 2587 0, 2588 0, 2589 /*10*/ 0, 2590 0, 2591 0, 2592 MatSOR_MPIAIJ, 2593 MatTranspose_MPIAIJ, 2594 /*15*/ MatGetInfo_MPIAIJ, 2595 MatEqual_MPIAIJ, 2596 MatGetDiagonal_MPIAIJ, 2597 MatDiagonalScale_MPIAIJ, 2598 MatNorm_MPIAIJ, 2599 /*20*/ MatAssemblyBegin_MPIAIJ, 2600 MatAssemblyEnd_MPIAIJ, 2601 MatSetOption_MPIAIJ, 2602 MatZeroEntries_MPIAIJ, 2603 /*24*/ MatZeroRows_MPIAIJ, 2604 0, 2605 0, 2606 0, 2607 0, 2608 /*29*/ MatSetUp_MPIAIJ, 2609 0, 2610 0, 2611 MatGetDiagonalBlock_MPIAIJ, 2612 0, 2613 /*34*/ MatDuplicate_MPIAIJ, 2614 0, 2615 0, 2616 0, 2617 0, 2618 /*39*/ MatAXPY_MPIAIJ, 2619 MatCreateSubMatrices_MPIAIJ, 2620 MatIncreaseOverlap_MPIAIJ, 2621 MatGetValues_MPIAIJ, 2622 MatCopy_MPIAIJ, 2623 /*44*/ MatGetRowMax_MPIAIJ, 2624 MatScale_MPIAIJ, 2625 MatShift_MPIAIJ, 2626 MatDiagonalSet_MPIAIJ, 2627 MatZeroRowsColumns_MPIAIJ, 2628 /*49*/ MatSetRandom_MPIAIJ, 2629 0, 2630 0, 2631 0, 2632 0, 2633 /*54*/ MatFDColoringCreate_MPIXAIJ, 2634 0, 2635 MatSetUnfactored_MPIAIJ, 2636 MatPermute_MPIAIJ, 2637 0, 2638 /*59*/ MatCreateSubMatrix_MPIAIJ, 2639 MatDestroy_MPIAIJ, 2640 MatView_MPIAIJ, 2641 0, 2642 MatMatMatMult_MPIAIJ_MPIAIJ_MPIAIJ, 2643 /*64*/ MatMatMatMultSymbolic_MPIAIJ_MPIAIJ_MPIAIJ, 2644 MatMatMatMultNumeric_MPIAIJ_MPIAIJ_MPIAIJ, 2645 0, 2646 0, 2647 0, 2648 /*69*/ MatGetRowMaxAbs_MPIAIJ, 2649 MatGetRowMinAbs_MPIAIJ, 2650 0, 2651 0, 2652 0, 2653 0, 2654 /*75*/ MatFDColoringApply_AIJ, 2655 MatSetFromOptions_MPIAIJ, 2656 0, 2657 0, 2658 MatFindZeroDiagonals_MPIAIJ, 2659 /*80*/ 0, 2660 0, 2661 0, 2662 /*83*/ MatLoad_MPIAIJ, 2663 MatIsSymmetric_MPIAIJ, 2664 0, 2665 0, 2666 0, 2667 0, 2668 /*89*/ MatMatMult_MPIAIJ_MPIAIJ, 2669 MatMatMultSymbolic_MPIAIJ_MPIAIJ, 2670 MatMatMultNumeric_MPIAIJ_MPIAIJ, 2671 MatPtAP_MPIAIJ_MPIAIJ, 2672 MatPtAPSymbolic_MPIAIJ_MPIAIJ, 2673 /*94*/ MatPtAPNumeric_MPIAIJ_MPIAIJ, 2674 0, 2675 0, 2676 0, 2677 0, 2678 /*99*/ 0, 2679 0, 2680 0, 2681 MatConjugate_MPIAIJ, 2682 0, 2683 /*104*/MatSetValuesRow_MPIAIJ, 2684 MatRealPart_MPIAIJ, 2685 MatImaginaryPart_MPIAIJ, 2686 0, 2687 0, 2688 /*109*/0, 2689 0, 2690 MatGetRowMin_MPIAIJ, 2691 0, 2692 MatMissingDiagonal_MPIAIJ, 2693 /*114*/MatGetSeqNonzeroStructure_MPIAIJ, 2694 0, 2695 MatGetGhosts_MPIAIJ, 2696 0, 2697 0, 2698 /*119*/0, 2699 0, 2700 0, 2701 0, 2702 MatGetMultiProcBlock_MPIAIJ, 2703 /*124*/MatFindNonzeroRows_MPIAIJ, 2704 MatGetColumnNorms_MPIAIJ, 2705 MatInvertBlockDiagonal_MPIAIJ, 2706 MatInvertVariableBlockDiagonal_MPIAIJ, 2707 MatCreateSubMatricesMPI_MPIAIJ, 2708 /*129*/0, 2709 MatTransposeMatMult_MPIAIJ_MPIAIJ, 2710 MatTransposeMatMultSymbolic_MPIAIJ_MPIAIJ, 2711 MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ, 2712 0, 2713 /*134*/0, 2714 0, 2715 MatRARt_MPIAIJ_MPIAIJ, 2716 0, 2717 0, 2718 /*139*/MatSetBlockSizes_MPIAIJ, 2719 0, 2720 0, 2721 MatFDColoringSetUp_MPIXAIJ, 2722 MatFindOffBlockDiagonalEntries_MPIAIJ, 2723 /*144*/MatCreateMPIMatConcatenateSeqMat_MPIAIJ 2724 }; 2725 2726 /* ----------------------------------------------------------------------------------------*/ 2727 2728 PetscErrorCode MatStoreValues_MPIAIJ(Mat mat) 2729 { 2730 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 2731 PetscErrorCode ierr; 2732 2733 PetscFunctionBegin; 2734 ierr = MatStoreValues(aij->A);CHKERRQ(ierr); 2735 ierr = MatStoreValues(aij->B);CHKERRQ(ierr); 2736 PetscFunctionReturn(0); 2737 } 2738 2739 PetscErrorCode MatRetrieveValues_MPIAIJ(Mat mat) 2740 { 2741 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 2742 PetscErrorCode ierr; 2743 2744 PetscFunctionBegin; 2745 ierr = MatRetrieveValues(aij->A);CHKERRQ(ierr); 2746 ierr = MatRetrieveValues(aij->B);CHKERRQ(ierr); 2747 PetscFunctionReturn(0); 2748 } 2749 2750 PetscErrorCode MatMPIAIJSetPreallocation_MPIAIJ(Mat B,PetscInt d_nz,const PetscInt d_nnz[],PetscInt o_nz,const PetscInt o_nnz[]) 2751 { 2752 Mat_MPIAIJ *b; 2753 PetscErrorCode ierr; 2754 2755 PetscFunctionBegin; 2756 ierr = PetscLayoutSetUp(B->rmap);CHKERRQ(ierr); 2757 ierr = PetscLayoutSetUp(B->cmap);CHKERRQ(ierr); 2758 b = (Mat_MPIAIJ*)B->data; 2759 2760 #if defined(PETSC_USE_CTABLE) 2761 ierr = PetscTableDestroy(&b->colmap);CHKERRQ(ierr); 2762 #else 2763 ierr = PetscFree(b->colmap);CHKERRQ(ierr); 2764 #endif 2765 ierr = PetscFree(b->garray);CHKERRQ(ierr); 2766 ierr = VecDestroy(&b->lvec);CHKERRQ(ierr); 2767 ierr = VecScatterDestroy(&b->Mvctx);CHKERRQ(ierr); 2768 2769 /* Because the B will have been resized we simply destroy it and create a new one each time */ 2770 ierr = MatDestroy(&b->B);CHKERRQ(ierr); 2771 ierr = MatCreate(PETSC_COMM_SELF,&b->B);CHKERRQ(ierr); 2772 ierr = MatSetSizes(b->B,B->rmap->n,B->cmap->N,B->rmap->n,B->cmap->N);CHKERRQ(ierr); 2773 ierr = MatSetBlockSizesFromMats(b->B,B,B);CHKERRQ(ierr); 2774 ierr = MatSetType(b->B,MATSEQAIJ);CHKERRQ(ierr); 2775 ierr = PetscLogObjectParent((PetscObject)B,(PetscObject)b->B);CHKERRQ(ierr); 2776 2777 if (!B->preallocated) { 2778 ierr = MatCreate(PETSC_COMM_SELF,&b->A);CHKERRQ(ierr); 2779 ierr = MatSetSizes(b->A,B->rmap->n,B->cmap->n,B->rmap->n,B->cmap->n);CHKERRQ(ierr); 2780 ierr = MatSetBlockSizesFromMats(b->A,B,B);CHKERRQ(ierr); 2781 ierr = MatSetType(b->A,MATSEQAIJ);CHKERRQ(ierr); 2782 ierr = PetscLogObjectParent((PetscObject)B,(PetscObject)b->A);CHKERRQ(ierr); 2783 } 2784 2785 ierr = MatSeqAIJSetPreallocation(b->A,d_nz,d_nnz);CHKERRQ(ierr); 2786 ierr = MatSeqAIJSetPreallocation(b->B,o_nz,o_nnz);CHKERRQ(ierr); 2787 B->preallocated = PETSC_TRUE; 2788 B->was_assembled = PETSC_FALSE; 2789 B->assembled = PETSC_FALSE;; 2790 PetscFunctionReturn(0); 2791 } 2792 2793 PetscErrorCode MatResetPreallocation_MPIAIJ(Mat B) 2794 { 2795 Mat_MPIAIJ *b; 2796 PetscErrorCode ierr; 2797 2798 PetscFunctionBegin; 2799 PetscValidHeaderSpecific(B,MAT_CLASSID,1); 2800 ierr = PetscLayoutSetUp(B->rmap);CHKERRQ(ierr); 2801 ierr = PetscLayoutSetUp(B->cmap);CHKERRQ(ierr); 2802 b = (Mat_MPIAIJ*)B->data; 2803 2804 #if defined(PETSC_USE_CTABLE) 2805 ierr = PetscTableDestroy(&b->colmap);CHKERRQ(ierr); 2806 #else 2807 ierr = PetscFree(b->colmap);CHKERRQ(ierr); 2808 #endif 2809 ierr = PetscFree(b->garray);CHKERRQ(ierr); 2810 ierr = VecDestroy(&b->lvec);CHKERRQ(ierr); 2811 ierr = VecScatterDestroy(&b->Mvctx);CHKERRQ(ierr); 2812 2813 ierr = MatResetPreallocation(b->A);CHKERRQ(ierr); 2814 ierr = MatResetPreallocation(b->B);CHKERRQ(ierr); 2815 B->preallocated = PETSC_TRUE; 2816 B->was_assembled = PETSC_FALSE; 2817 B->assembled = PETSC_FALSE; 2818 PetscFunctionReturn(0); 2819 } 2820 2821 PetscErrorCode MatDuplicate_MPIAIJ(Mat matin,MatDuplicateOption cpvalues,Mat *newmat) 2822 { 2823 Mat mat; 2824 Mat_MPIAIJ *a,*oldmat = (Mat_MPIAIJ*)matin->data; 2825 PetscErrorCode ierr; 2826 2827 PetscFunctionBegin; 2828 *newmat = 0; 2829 ierr = MatCreate(PetscObjectComm((PetscObject)matin),&mat);CHKERRQ(ierr); 2830 ierr = MatSetSizes(mat,matin->rmap->n,matin->cmap->n,matin->rmap->N,matin->cmap->N);CHKERRQ(ierr); 2831 ierr = MatSetBlockSizesFromMats(mat,matin,matin);CHKERRQ(ierr); 2832 ierr = MatSetType(mat,((PetscObject)matin)->type_name);CHKERRQ(ierr); 2833 ierr = PetscMemcpy(mat->ops,matin->ops,sizeof(struct _MatOps));CHKERRQ(ierr); 2834 a = (Mat_MPIAIJ*)mat->data; 2835 2836 mat->factortype = matin->factortype; 2837 mat->assembled = PETSC_TRUE; 2838 mat->insertmode = NOT_SET_VALUES; 2839 mat->preallocated = PETSC_TRUE; 2840 2841 a->size = oldmat->size; 2842 a->rank = oldmat->rank; 2843 a->donotstash = oldmat->donotstash; 2844 a->roworiented = oldmat->roworiented; 2845 a->rowindices = 0; 2846 a->rowvalues = 0; 2847 a->getrowactive = PETSC_FALSE; 2848 2849 ierr = PetscLayoutReference(matin->rmap,&mat->rmap);CHKERRQ(ierr); 2850 ierr = PetscLayoutReference(matin->cmap,&mat->cmap);CHKERRQ(ierr); 2851 2852 if (oldmat->colmap) { 2853 #if defined(PETSC_USE_CTABLE) 2854 ierr = PetscTableCreateCopy(oldmat->colmap,&a->colmap);CHKERRQ(ierr); 2855 #else 2856 ierr = PetscMalloc1(mat->cmap->N,&a->colmap);CHKERRQ(ierr); 2857 ierr = PetscLogObjectMemory((PetscObject)mat,(mat->cmap->N)*sizeof(PetscInt));CHKERRQ(ierr); 2858 ierr = PetscMemcpy(a->colmap,oldmat->colmap,(mat->cmap->N)*sizeof(PetscInt));CHKERRQ(ierr); 2859 #endif 2860 } else a->colmap = 0; 2861 if (oldmat->garray) { 2862 PetscInt len; 2863 len = oldmat->B->cmap->n; 2864 ierr = PetscMalloc1(len+1,&a->garray);CHKERRQ(ierr); 2865 ierr = PetscLogObjectMemory((PetscObject)mat,len*sizeof(PetscInt));CHKERRQ(ierr); 2866 if (len) { ierr = PetscMemcpy(a->garray,oldmat->garray,len*sizeof(PetscInt));CHKERRQ(ierr); } 2867 } else a->garray = 0; 2868 2869 ierr = VecDuplicate(oldmat->lvec,&a->lvec);CHKERRQ(ierr); 2870 ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)a->lvec);CHKERRQ(ierr); 2871 ierr = VecScatterCopy(oldmat->Mvctx,&a->Mvctx);CHKERRQ(ierr); 2872 ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)a->Mvctx);CHKERRQ(ierr); 2873 2874 if (oldmat->Mvctx_mpi1) { 2875 ierr = VecScatterCopy(oldmat->Mvctx_mpi1,&a->Mvctx_mpi1);CHKERRQ(ierr); 2876 ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)a->Mvctx_mpi1);CHKERRQ(ierr); 2877 } 2878 2879 ierr = MatDuplicate(oldmat->A,cpvalues,&a->A);CHKERRQ(ierr); 2880 ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)a->A);CHKERRQ(ierr); 2881 ierr = MatDuplicate(oldmat->B,cpvalues,&a->B);CHKERRQ(ierr); 2882 ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)a->B);CHKERRQ(ierr); 2883 ierr = PetscFunctionListDuplicate(((PetscObject)matin)->qlist,&((PetscObject)mat)->qlist);CHKERRQ(ierr); 2884 *newmat = mat; 2885 PetscFunctionReturn(0); 2886 } 2887 2888 PetscErrorCode MatLoad_MPIAIJ(Mat newMat, PetscViewer viewer) 2889 { 2890 PetscBool isbinary, ishdf5; 2891 PetscErrorCode ierr; 2892 2893 PetscFunctionBegin; 2894 PetscValidHeaderSpecific(newMat,MAT_CLASSID,1); 2895 PetscValidHeaderSpecific(viewer,PETSC_VIEWER_CLASSID,2); 2896 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERBINARY,&isbinary);CHKERRQ(ierr); 2897 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERHDF5, &ishdf5);CHKERRQ(ierr); 2898 if (isbinary) { 2899 ierr = MatLoad_MPIAIJ_Binary(newMat,viewer);CHKERRQ(ierr); 2900 } else if (ishdf5) { 2901 #if defined(PETSC_HAVE_HDF5) 2902 ierr = MatLoad_AIJ_HDF5(newMat,viewer);CHKERRQ(ierr); 2903 #else 2904 SETERRQ(PetscObjectComm((PetscObject)newMat),PETSC_ERR_SUP,"HDF5 not supported in this build.\nPlease reconfigure using --download-hdf5"); 2905 #endif 2906 } else { 2907 SETERRQ2(PetscObjectComm((PetscObject)newMat),PETSC_ERR_SUP,"Viewer type %s not yet supported for reading %s matrices",((PetscObject)viewer)->type_name,((PetscObject)newMat)->type_name); 2908 } 2909 PetscFunctionReturn(0); 2910 } 2911 2912 PetscErrorCode MatLoad_MPIAIJ_Binary(Mat newMat, PetscViewer viewer) 2913 { 2914 PetscScalar *vals,*svals; 2915 MPI_Comm comm; 2916 PetscErrorCode ierr; 2917 PetscMPIInt rank,size,tag = ((PetscObject)viewer)->tag; 2918 PetscInt i,nz,j,rstart,rend,mmax,maxnz = 0; 2919 PetscInt header[4],*rowlengths = 0,M,N,m,*cols; 2920 PetscInt *ourlens = NULL,*procsnz = NULL,*offlens = NULL,jj,*mycols,*smycols; 2921 PetscInt cend,cstart,n,*rowners; 2922 int fd; 2923 PetscInt bs = newMat->rmap->bs; 2924 2925 PetscFunctionBegin; 2926 /* force binary viewer to load .info file if it has not yet done so */ 2927 ierr = PetscViewerSetUp(viewer);CHKERRQ(ierr); 2928 ierr = PetscObjectGetComm((PetscObject)viewer,&comm);CHKERRQ(ierr); 2929 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 2930 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 2931 ierr = PetscViewerBinaryGetDescriptor(viewer,&fd);CHKERRQ(ierr); 2932 if (!rank) { 2933 ierr = PetscBinaryRead(fd,(char*)header,4,PETSC_INT);CHKERRQ(ierr); 2934 if (header[0] != MAT_FILE_CLASSID) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED,"not matrix object"); 2935 if (header[3] < 0) SETERRQ(PetscObjectComm((PetscObject)newMat),PETSC_ERR_FILE_UNEXPECTED,"Matrix stored in special format on disk,cannot load as MATMPIAIJ"); 2936 } 2937 2938 ierr = PetscOptionsBegin(comm,NULL,"Options for loading MATMPIAIJ matrix","Mat");CHKERRQ(ierr); 2939 ierr = PetscOptionsInt("-matload_block_size","Set the blocksize used to store the matrix","MatLoad",bs,&bs,NULL);CHKERRQ(ierr); 2940 ierr = PetscOptionsEnd();CHKERRQ(ierr); 2941 if (bs < 0) bs = 1; 2942 2943 ierr = MPI_Bcast(header+1,3,MPIU_INT,0,comm);CHKERRQ(ierr); 2944 M = header[1]; N = header[2]; 2945 2946 /* If global sizes are set, check if they are consistent with that given in the file */ 2947 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); 2948 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); 2949 2950 /* determine ownership of all (block) rows */ 2951 if (M%bs) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED, "Inconsistent # of rows (%d) and block size (%d)",M,bs); 2952 if (newMat->rmap->n < 0) m = bs*((M/bs)/size + (((M/bs) % size) > rank)); /* PETSC_DECIDE */ 2953 else m = newMat->rmap->n; /* Set by user */ 2954 2955 ierr = PetscMalloc1(size+1,&rowners);CHKERRQ(ierr); 2956 ierr = MPI_Allgather(&m,1,MPIU_INT,rowners+1,1,MPIU_INT,comm);CHKERRQ(ierr); 2957 2958 /* First process needs enough room for process with most rows */ 2959 if (!rank) { 2960 mmax = rowners[1]; 2961 for (i=2; i<=size; i++) { 2962 mmax = PetscMax(mmax, rowners[i]); 2963 } 2964 } else mmax = -1; /* unused, but compilers complain */ 2965 2966 rowners[0] = 0; 2967 for (i=2; i<=size; i++) { 2968 rowners[i] += rowners[i-1]; 2969 } 2970 rstart = rowners[rank]; 2971 rend = rowners[rank+1]; 2972 2973 /* distribute row lengths to all processors */ 2974 ierr = PetscMalloc2(m,&ourlens,m,&offlens);CHKERRQ(ierr); 2975 if (!rank) { 2976 ierr = PetscBinaryRead(fd,ourlens,m,PETSC_INT);CHKERRQ(ierr); 2977 ierr = PetscMalloc1(mmax,&rowlengths);CHKERRQ(ierr); 2978 ierr = PetscCalloc1(size,&procsnz);CHKERRQ(ierr); 2979 for (j=0; j<m; j++) { 2980 procsnz[0] += ourlens[j]; 2981 } 2982 for (i=1; i<size; i++) { 2983 ierr = PetscBinaryRead(fd,rowlengths,rowners[i+1]-rowners[i],PETSC_INT);CHKERRQ(ierr); 2984 /* calculate the number of nonzeros on each processor */ 2985 for (j=0; j<rowners[i+1]-rowners[i]; j++) { 2986 procsnz[i] += rowlengths[j]; 2987 } 2988 ierr = MPIULong_Send(rowlengths,rowners[i+1]-rowners[i],MPIU_INT,i,tag,comm);CHKERRQ(ierr); 2989 } 2990 ierr = PetscFree(rowlengths);CHKERRQ(ierr); 2991 } else { 2992 ierr = MPIULong_Recv(ourlens,m,MPIU_INT,0,tag,comm);CHKERRQ(ierr); 2993 } 2994 2995 if (!rank) { 2996 /* determine max buffer needed and allocate it */ 2997 maxnz = 0; 2998 for (i=0; i<size; i++) { 2999 maxnz = PetscMax(maxnz,procsnz[i]); 3000 } 3001 ierr = PetscMalloc1(maxnz,&cols);CHKERRQ(ierr); 3002 3003 /* read in my part of the matrix column indices */ 3004 nz = procsnz[0]; 3005 ierr = PetscMalloc1(nz,&mycols);CHKERRQ(ierr); 3006 ierr = PetscBinaryRead(fd,mycols,nz,PETSC_INT);CHKERRQ(ierr); 3007 3008 /* read in every one elses and ship off */ 3009 for (i=1; i<size; i++) { 3010 nz = procsnz[i]; 3011 ierr = PetscBinaryRead(fd,cols,nz,PETSC_INT);CHKERRQ(ierr); 3012 ierr = MPIULong_Send(cols,nz,MPIU_INT,i,tag,comm);CHKERRQ(ierr); 3013 } 3014 ierr = PetscFree(cols);CHKERRQ(ierr); 3015 } else { 3016 /* determine buffer space needed for message */ 3017 nz = 0; 3018 for (i=0; i<m; i++) { 3019 nz += ourlens[i]; 3020 } 3021 ierr = PetscMalloc1(nz,&mycols);CHKERRQ(ierr); 3022 3023 /* receive message of column indices*/ 3024 ierr = MPIULong_Recv(mycols,nz,MPIU_INT,0,tag,comm);CHKERRQ(ierr); 3025 } 3026 3027 /* determine column ownership if matrix is not square */ 3028 if (N != M) { 3029 if (newMat->cmap->n < 0) n = N/size + ((N % size) > rank); 3030 else n = newMat->cmap->n; 3031 ierr = MPI_Scan(&n,&cend,1,MPIU_INT,MPI_SUM,comm);CHKERRQ(ierr); 3032 cstart = cend - n; 3033 } else { 3034 cstart = rstart; 3035 cend = rend; 3036 n = cend - cstart; 3037 } 3038 3039 /* loop over local rows, determining number of off diagonal entries */ 3040 ierr = PetscMemzero(offlens,m*sizeof(PetscInt));CHKERRQ(ierr); 3041 jj = 0; 3042 for (i=0; i<m; i++) { 3043 for (j=0; j<ourlens[i]; j++) { 3044 if (mycols[jj] < cstart || mycols[jj] >= cend) offlens[i]++; 3045 jj++; 3046 } 3047 } 3048 3049 for (i=0; i<m; i++) { 3050 ourlens[i] -= offlens[i]; 3051 } 3052 ierr = MatSetSizes(newMat,m,n,M,N);CHKERRQ(ierr); 3053 3054 if (bs > 1) {ierr = MatSetBlockSize(newMat,bs);CHKERRQ(ierr);} 3055 3056 ierr = MatMPIAIJSetPreallocation(newMat,0,ourlens,0,offlens);CHKERRQ(ierr); 3057 3058 for (i=0; i<m; i++) { 3059 ourlens[i] += offlens[i]; 3060 } 3061 3062 if (!rank) { 3063 ierr = PetscMalloc1(maxnz+1,&vals);CHKERRQ(ierr); 3064 3065 /* read in my part of the matrix numerical values */ 3066 nz = procsnz[0]; 3067 ierr = PetscBinaryRead(fd,vals,nz,PETSC_SCALAR);CHKERRQ(ierr); 3068 3069 /* insert into matrix */ 3070 jj = rstart; 3071 smycols = mycols; 3072 svals = vals; 3073 for (i=0; i<m; i++) { 3074 ierr = MatSetValues_MPIAIJ(newMat,1,&jj,ourlens[i],smycols,svals,INSERT_VALUES);CHKERRQ(ierr); 3075 smycols += ourlens[i]; 3076 svals += ourlens[i]; 3077 jj++; 3078 } 3079 3080 /* read in other processors and ship out */ 3081 for (i=1; i<size; i++) { 3082 nz = procsnz[i]; 3083 ierr = PetscBinaryRead(fd,vals,nz,PETSC_SCALAR);CHKERRQ(ierr); 3084 ierr = MPIULong_Send(vals,nz,MPIU_SCALAR,i,((PetscObject)newMat)->tag,comm);CHKERRQ(ierr); 3085 } 3086 ierr = PetscFree(procsnz);CHKERRQ(ierr); 3087 } else { 3088 /* receive numeric values */ 3089 ierr = PetscMalloc1(nz+1,&vals);CHKERRQ(ierr); 3090 3091 /* receive message of values*/ 3092 ierr = MPIULong_Recv(vals,nz,MPIU_SCALAR,0,((PetscObject)newMat)->tag,comm);CHKERRQ(ierr); 3093 3094 /* insert into matrix */ 3095 jj = rstart; 3096 smycols = mycols; 3097 svals = vals; 3098 for (i=0; i<m; i++) { 3099 ierr = MatSetValues_MPIAIJ(newMat,1,&jj,ourlens[i],smycols,svals,INSERT_VALUES);CHKERRQ(ierr); 3100 smycols += ourlens[i]; 3101 svals += ourlens[i]; 3102 jj++; 3103 } 3104 } 3105 ierr = PetscFree2(ourlens,offlens);CHKERRQ(ierr); 3106 ierr = PetscFree(vals);CHKERRQ(ierr); 3107 ierr = PetscFree(mycols);CHKERRQ(ierr); 3108 ierr = PetscFree(rowners);CHKERRQ(ierr); 3109 ierr = MatAssemblyBegin(newMat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3110 ierr = MatAssemblyEnd(newMat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3111 PetscFunctionReturn(0); 3112 } 3113 3114 /* Not scalable because of ISAllGather() unless getting all columns. */ 3115 PetscErrorCode ISGetSeqIS_Private(Mat mat,IS iscol,IS *isseq) 3116 { 3117 PetscErrorCode ierr; 3118 IS iscol_local; 3119 PetscBool isstride; 3120 PetscMPIInt lisstride=0,gisstride; 3121 3122 PetscFunctionBegin; 3123 /* check if we are grabbing all columns*/ 3124 ierr = PetscObjectTypeCompare((PetscObject)iscol,ISSTRIDE,&isstride);CHKERRQ(ierr); 3125 3126 if (isstride) { 3127 PetscInt start,len,mstart,mlen; 3128 ierr = ISStrideGetInfo(iscol,&start,NULL);CHKERRQ(ierr); 3129 ierr = ISGetLocalSize(iscol,&len);CHKERRQ(ierr); 3130 ierr = MatGetOwnershipRangeColumn(mat,&mstart,&mlen);CHKERRQ(ierr); 3131 if (mstart == start && mlen-mstart == len) lisstride = 1; 3132 } 3133 3134 ierr = MPIU_Allreduce(&lisstride,&gisstride,1,MPI_INT,MPI_MIN,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 3135 if (gisstride) { 3136 PetscInt N; 3137 ierr = MatGetSize(mat,NULL,&N);CHKERRQ(ierr); 3138 ierr = ISCreateStride(PetscObjectComm((PetscObject)mat),N,0,1,&iscol_local);CHKERRQ(ierr); 3139 ierr = ISSetIdentity(iscol_local);CHKERRQ(ierr); 3140 ierr = PetscInfo(mat,"Optimizing for obtaining all columns of the matrix; skipping ISAllGather()\n");CHKERRQ(ierr); 3141 } else { 3142 PetscInt cbs; 3143 ierr = ISGetBlockSize(iscol,&cbs);CHKERRQ(ierr); 3144 ierr = ISAllGather(iscol,&iscol_local);CHKERRQ(ierr); 3145 ierr = ISSetBlockSize(iscol_local,cbs);CHKERRQ(ierr); 3146 } 3147 3148 *isseq = iscol_local; 3149 PetscFunctionReturn(0); 3150 } 3151 3152 /* 3153 Used by MatCreateSubMatrix_MPIAIJ_SameRowColDist() to avoid ISAllGather() and global size of iscol_local 3154 (see MatCreateSubMatrix_MPIAIJ_nonscalable) 3155 3156 Input Parameters: 3157 mat - matrix 3158 isrow - parallel row index set; its local indices are a subset of local columns of mat, 3159 i.e., mat->rstart <= isrow[i] < mat->rend 3160 iscol - parallel column index set; its local indices are a subset of local columns of mat, 3161 i.e., mat->cstart <= iscol[i] < mat->cend 3162 Output Parameter: 3163 isrow_d,iscol_d - sequential row and column index sets for retrieving mat->A 3164 iscol_o - sequential column index set for retrieving mat->B 3165 garray - column map; garray[i] indicates global location of iscol_o[i] in iscol 3166 */ 3167 PetscErrorCode ISGetSeqIS_SameColDist_Private(Mat mat,IS isrow,IS iscol,IS *isrow_d,IS *iscol_d,IS *iscol_o,const PetscInt *garray[]) 3168 { 3169 PetscErrorCode ierr; 3170 Vec x,cmap; 3171 const PetscInt *is_idx; 3172 PetscScalar *xarray,*cmaparray; 3173 PetscInt ncols,isstart,*idx,m,rstart,*cmap1,count; 3174 Mat_MPIAIJ *a=(Mat_MPIAIJ*)mat->data; 3175 Mat B=a->B; 3176 Vec lvec=a->lvec,lcmap; 3177 PetscInt i,cstart,cend,Bn=B->cmap->N; 3178 MPI_Comm comm; 3179 VecScatter Mvctx=a->Mvctx; 3180 3181 PetscFunctionBegin; 3182 ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 3183 ierr = ISGetLocalSize(iscol,&ncols);CHKERRQ(ierr); 3184 3185 /* (1) iscol is a sub-column vector of mat, pad it with '-1.' to form a full vector x */ 3186 ierr = MatCreateVecs(mat,&x,NULL);CHKERRQ(ierr); 3187 ierr = VecSet(x,-1.0);CHKERRQ(ierr); 3188 ierr = VecDuplicate(x,&cmap);CHKERRQ(ierr); 3189 ierr = VecSet(cmap,-1.0);CHKERRQ(ierr); 3190 3191 /* Get start indices */ 3192 ierr = MPI_Scan(&ncols,&isstart,1,MPIU_INT,MPI_SUM,comm);CHKERRQ(ierr); 3193 isstart -= ncols; 3194 ierr = MatGetOwnershipRangeColumn(mat,&cstart,&cend);CHKERRQ(ierr); 3195 3196 ierr = ISGetIndices(iscol,&is_idx);CHKERRQ(ierr); 3197 ierr = VecGetArray(x,&xarray);CHKERRQ(ierr); 3198 ierr = VecGetArray(cmap,&cmaparray);CHKERRQ(ierr); 3199 ierr = PetscMalloc1(ncols,&idx);CHKERRQ(ierr); 3200 for (i=0; i<ncols; i++) { 3201 xarray[is_idx[i]-cstart] = (PetscScalar)is_idx[i]; 3202 cmaparray[is_idx[i]-cstart] = i + isstart; /* global index of iscol[i] */ 3203 idx[i] = is_idx[i]-cstart; /* local index of iscol[i] */ 3204 } 3205 ierr = VecRestoreArray(x,&xarray);CHKERRQ(ierr); 3206 ierr = VecRestoreArray(cmap,&cmaparray);CHKERRQ(ierr); 3207 ierr = ISRestoreIndices(iscol,&is_idx);CHKERRQ(ierr); 3208 3209 /* Get iscol_d */ 3210 ierr = ISCreateGeneral(PETSC_COMM_SELF,ncols,idx,PETSC_OWN_POINTER,iscol_d);CHKERRQ(ierr); 3211 ierr = ISGetBlockSize(iscol,&i);CHKERRQ(ierr); 3212 ierr = ISSetBlockSize(*iscol_d,i);CHKERRQ(ierr); 3213 3214 /* Get isrow_d */ 3215 ierr = ISGetLocalSize(isrow,&m);CHKERRQ(ierr); 3216 rstart = mat->rmap->rstart; 3217 ierr = PetscMalloc1(m,&idx);CHKERRQ(ierr); 3218 ierr = ISGetIndices(isrow,&is_idx);CHKERRQ(ierr); 3219 for (i=0; i<m; i++) idx[i] = is_idx[i]-rstart; 3220 ierr = ISRestoreIndices(isrow,&is_idx);CHKERRQ(ierr); 3221 3222 ierr = ISCreateGeneral(PETSC_COMM_SELF,m,idx,PETSC_OWN_POINTER,isrow_d);CHKERRQ(ierr); 3223 ierr = ISGetBlockSize(isrow,&i);CHKERRQ(ierr); 3224 ierr = ISSetBlockSize(*isrow_d,i);CHKERRQ(ierr); 3225 3226 /* (2) Scatter x and cmap using aij->Mvctx to get their off-process portions (see MatMult_MPIAIJ) */ 3227 ierr = VecScatterBegin(Mvctx,x,lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3228 ierr = VecScatterEnd(Mvctx,x,lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3229 3230 ierr = VecDuplicate(lvec,&lcmap);CHKERRQ(ierr); 3231 3232 ierr = VecScatterBegin(Mvctx,cmap,lcmap,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3233 ierr = VecScatterEnd(Mvctx,cmap,lcmap,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3234 3235 /* (3) create sequential iscol_o (a subset of iscol) and isgarray */ 3236 /* off-process column indices */ 3237 count = 0; 3238 ierr = PetscMalloc1(Bn,&idx);CHKERRQ(ierr); 3239 ierr = PetscMalloc1(Bn,&cmap1);CHKERRQ(ierr); 3240 3241 ierr = VecGetArray(lvec,&xarray);CHKERRQ(ierr); 3242 ierr = VecGetArray(lcmap,&cmaparray);CHKERRQ(ierr); 3243 for (i=0; i<Bn; i++) { 3244 if (PetscRealPart(xarray[i]) > -1.0) { 3245 idx[count] = i; /* local column index in off-diagonal part B */ 3246 cmap1[count] = (PetscInt)PetscRealPart(cmaparray[i]); /* column index in submat */ 3247 count++; 3248 } 3249 } 3250 ierr = VecRestoreArray(lvec,&xarray);CHKERRQ(ierr); 3251 ierr = VecRestoreArray(lcmap,&cmaparray);CHKERRQ(ierr); 3252 3253 ierr = ISCreateGeneral(PETSC_COMM_SELF,count,idx,PETSC_COPY_VALUES,iscol_o);CHKERRQ(ierr); 3254 /* cannot ensure iscol_o has same blocksize as iscol! */ 3255 3256 ierr = PetscFree(idx);CHKERRQ(ierr); 3257 *garray = cmap1; 3258 3259 ierr = VecDestroy(&x);CHKERRQ(ierr); 3260 ierr = VecDestroy(&cmap);CHKERRQ(ierr); 3261 ierr = VecDestroy(&lcmap);CHKERRQ(ierr); 3262 PetscFunctionReturn(0); 3263 } 3264 3265 /* isrow and iscol have same processor distribution as mat, output *submat is a submatrix of local mat */ 3266 PetscErrorCode MatCreateSubMatrix_MPIAIJ_SameRowColDist(Mat mat,IS isrow,IS iscol,MatReuse call,Mat *submat) 3267 { 3268 PetscErrorCode ierr; 3269 Mat_MPIAIJ *a = (Mat_MPIAIJ*)mat->data,*asub; 3270 Mat M = NULL; 3271 MPI_Comm comm; 3272 IS iscol_d,isrow_d,iscol_o; 3273 Mat Asub = NULL,Bsub = NULL; 3274 PetscInt n; 3275 3276 PetscFunctionBegin; 3277 ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 3278 3279 if (call == MAT_REUSE_MATRIX) { 3280 /* Retrieve isrow_d, iscol_d and iscol_o from submat */ 3281 ierr = PetscObjectQuery((PetscObject)*submat,"isrow_d",(PetscObject*)&isrow_d);CHKERRQ(ierr); 3282 if (!isrow_d) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"isrow_d passed in was not used before, cannot reuse"); 3283 3284 ierr = PetscObjectQuery((PetscObject)*submat,"iscol_d",(PetscObject*)&iscol_d);CHKERRQ(ierr); 3285 if (!iscol_d) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"iscol_d passed in was not used before, cannot reuse"); 3286 3287 ierr = PetscObjectQuery((PetscObject)*submat,"iscol_o",(PetscObject*)&iscol_o);CHKERRQ(ierr); 3288 if (!iscol_o) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"iscol_o passed in was not used before, cannot reuse"); 3289 3290 /* Update diagonal and off-diagonal portions of submat */ 3291 asub = (Mat_MPIAIJ*)(*submat)->data; 3292 ierr = MatCreateSubMatrix_SeqAIJ(a->A,isrow_d,iscol_d,PETSC_DECIDE,MAT_REUSE_MATRIX,&asub->A);CHKERRQ(ierr); 3293 ierr = ISGetLocalSize(iscol_o,&n);CHKERRQ(ierr); 3294 if (n) { 3295 ierr = MatCreateSubMatrix_SeqAIJ(a->B,isrow_d,iscol_o,PETSC_DECIDE,MAT_REUSE_MATRIX,&asub->B);CHKERRQ(ierr); 3296 } 3297 ierr = MatAssemblyBegin(*submat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3298 ierr = MatAssemblyEnd(*submat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3299 3300 } else { /* call == MAT_INITIAL_MATRIX) */ 3301 const PetscInt *garray; 3302 PetscInt BsubN; 3303 3304 /* Create isrow_d, iscol_d, iscol_o and isgarray (replace isgarray with array?) */ 3305 ierr = ISGetSeqIS_SameColDist_Private(mat,isrow,iscol,&isrow_d,&iscol_d,&iscol_o,&garray);CHKERRQ(ierr); 3306 3307 /* Create local submatrices Asub and Bsub */ 3308 ierr = MatCreateSubMatrix_SeqAIJ(a->A,isrow_d,iscol_d,PETSC_DECIDE,MAT_INITIAL_MATRIX,&Asub);CHKERRQ(ierr); 3309 ierr = MatCreateSubMatrix_SeqAIJ(a->B,isrow_d,iscol_o,PETSC_DECIDE,MAT_INITIAL_MATRIX,&Bsub);CHKERRQ(ierr); 3310 3311 /* Create submatrix M */ 3312 ierr = MatCreateMPIAIJWithSeqAIJ(comm,Asub,Bsub,garray,&M);CHKERRQ(ierr); 3313 3314 /* If Bsub has empty columns, compress iscol_o such that it will retrieve condensed Bsub from a->B during reuse */ 3315 asub = (Mat_MPIAIJ*)M->data; 3316 3317 ierr = ISGetLocalSize(iscol_o,&BsubN);CHKERRQ(ierr); 3318 n = asub->B->cmap->N; 3319 if (BsubN > n) { 3320 /* This case can be tested using ~petsc/src/tao/bound/examples/tutorials/runplate2_3 */ 3321 const PetscInt *idx; 3322 PetscInt i,j,*idx_new,*subgarray = asub->garray; 3323 ierr = PetscInfo2(M,"submatrix Bn %D != BsubN %D, update iscol_o\n",n,BsubN);CHKERRQ(ierr); 3324 3325 ierr = PetscMalloc1(n,&idx_new);CHKERRQ(ierr); 3326 j = 0; 3327 ierr = ISGetIndices(iscol_o,&idx);CHKERRQ(ierr); 3328 for (i=0; i<n; i++) { 3329 if (j >= BsubN) break; 3330 while (subgarray[i] > garray[j]) j++; 3331 3332 if (subgarray[i] == garray[j]) { 3333 idx_new[i] = idx[j++]; 3334 } else SETERRQ4(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"subgarray[%D]=%D cannot < garray[%D]=%D",i,subgarray[i],j,garray[j]); 3335 } 3336 ierr = ISRestoreIndices(iscol_o,&idx);CHKERRQ(ierr); 3337 3338 ierr = ISDestroy(&iscol_o);CHKERRQ(ierr); 3339 ierr = ISCreateGeneral(PETSC_COMM_SELF,n,idx_new,PETSC_OWN_POINTER,&iscol_o);CHKERRQ(ierr); 3340 3341 } else if (BsubN < n) { 3342 SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Columns of Bsub cannot be smaller than B's",BsubN,asub->B->cmap->N); 3343 } 3344 3345 ierr = PetscFree(garray);CHKERRQ(ierr); 3346 *submat = M; 3347 3348 /* Save isrow_d, iscol_d and iscol_o used in processor for next request */ 3349 ierr = PetscObjectCompose((PetscObject)M,"isrow_d",(PetscObject)isrow_d);CHKERRQ(ierr); 3350 ierr = ISDestroy(&isrow_d);CHKERRQ(ierr); 3351 3352 ierr = PetscObjectCompose((PetscObject)M,"iscol_d",(PetscObject)iscol_d);CHKERRQ(ierr); 3353 ierr = ISDestroy(&iscol_d);CHKERRQ(ierr); 3354 3355 ierr = PetscObjectCompose((PetscObject)M,"iscol_o",(PetscObject)iscol_o);CHKERRQ(ierr); 3356 ierr = ISDestroy(&iscol_o);CHKERRQ(ierr); 3357 } 3358 PetscFunctionReturn(0); 3359 } 3360 3361 PetscErrorCode MatCreateSubMatrix_MPIAIJ(Mat mat,IS isrow,IS iscol,MatReuse call,Mat *newmat) 3362 { 3363 PetscErrorCode ierr; 3364 IS iscol_local=NULL,isrow_d; 3365 PetscInt csize; 3366 PetscInt n,i,j,start,end; 3367 PetscBool sameRowDist=PETSC_FALSE,sameDist[2],tsameDist[2]; 3368 MPI_Comm comm; 3369 3370 PetscFunctionBegin; 3371 /* If isrow has same processor distribution as mat, 3372 call MatCreateSubMatrix_MPIAIJ_SameRowDist() to avoid using a hash table with global size of iscol */ 3373 if (call == MAT_REUSE_MATRIX) { 3374 ierr = PetscObjectQuery((PetscObject)*newmat,"isrow_d",(PetscObject*)&isrow_d);CHKERRQ(ierr); 3375 if (isrow_d) { 3376 sameRowDist = PETSC_TRUE; 3377 tsameDist[1] = PETSC_TRUE; /* sameColDist */ 3378 } else { 3379 ierr = PetscObjectQuery((PetscObject)*newmat,"SubIScol",(PetscObject*)&iscol_local);CHKERRQ(ierr); 3380 if (iscol_local) { 3381 sameRowDist = PETSC_TRUE; 3382 tsameDist[1] = PETSC_FALSE; /* !sameColDist */ 3383 } 3384 } 3385 } else { 3386 /* Check if isrow has same processor distribution as mat */ 3387 sameDist[0] = PETSC_FALSE; 3388 ierr = ISGetLocalSize(isrow,&n);CHKERRQ(ierr); 3389 if (!n) { 3390 sameDist[0] = PETSC_TRUE; 3391 } else { 3392 ierr = ISGetMinMax(isrow,&i,&j);CHKERRQ(ierr); 3393 ierr = MatGetOwnershipRange(mat,&start,&end);CHKERRQ(ierr); 3394 if (i >= start && j < end) { 3395 sameDist[0] = PETSC_TRUE; 3396 } 3397 } 3398 3399 /* Check if iscol has same processor distribution as mat */ 3400 sameDist[1] = PETSC_FALSE; 3401 ierr = ISGetLocalSize(iscol,&n);CHKERRQ(ierr); 3402 if (!n) { 3403 sameDist[1] = PETSC_TRUE; 3404 } else { 3405 ierr = ISGetMinMax(iscol,&i,&j);CHKERRQ(ierr); 3406 ierr = MatGetOwnershipRangeColumn(mat,&start,&end);CHKERRQ(ierr); 3407 if (i >= start && j < end) sameDist[1] = PETSC_TRUE; 3408 } 3409 3410 ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 3411 ierr = MPIU_Allreduce(&sameDist,&tsameDist,2,MPIU_BOOL,MPI_LAND,comm);CHKERRQ(ierr); 3412 sameRowDist = tsameDist[0]; 3413 } 3414 3415 if (sameRowDist) { 3416 if (tsameDist[1]) { /* sameRowDist & sameColDist */ 3417 /* isrow and iscol have same processor distribution as mat */ 3418 ierr = MatCreateSubMatrix_MPIAIJ_SameRowColDist(mat,isrow,iscol,call,newmat);CHKERRQ(ierr); 3419 PetscFunctionReturn(0); 3420 } else { /* sameRowDist */ 3421 /* isrow has same processor distribution as mat */ 3422 if (call == MAT_INITIAL_MATRIX) { 3423 PetscBool sorted; 3424 ierr = ISGetSeqIS_Private(mat,iscol,&iscol_local);CHKERRQ(ierr); 3425 ierr = ISGetLocalSize(iscol_local,&n);CHKERRQ(ierr); /* local size of iscol_local = global columns of newmat */ 3426 ierr = ISGetSize(iscol,&i);CHKERRQ(ierr); 3427 if (n != i) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"n %d != size of iscol %d",n,i); 3428 3429 ierr = ISSorted(iscol_local,&sorted);CHKERRQ(ierr); 3430 if (sorted) { 3431 /* MatCreateSubMatrix_MPIAIJ_SameRowDist() requires iscol_local be sorted; it can have duplicate indices */ 3432 ierr = MatCreateSubMatrix_MPIAIJ_SameRowDist(mat,isrow,iscol,iscol_local,MAT_INITIAL_MATRIX,newmat);CHKERRQ(ierr); 3433 PetscFunctionReturn(0); 3434 } 3435 } else { /* call == MAT_REUSE_MATRIX */ 3436 IS iscol_sub; 3437 ierr = PetscObjectQuery((PetscObject)*newmat,"SubIScol",(PetscObject*)&iscol_sub);CHKERRQ(ierr); 3438 if (iscol_sub) { 3439 ierr = MatCreateSubMatrix_MPIAIJ_SameRowDist(mat,isrow,iscol,NULL,call,newmat);CHKERRQ(ierr); 3440 PetscFunctionReturn(0); 3441 } 3442 } 3443 } 3444 } 3445 3446 /* General case: iscol -> iscol_local which has global size of iscol */ 3447 if (call == MAT_REUSE_MATRIX) { 3448 ierr = PetscObjectQuery((PetscObject)*newmat,"ISAllGather",(PetscObject*)&iscol_local);CHKERRQ(ierr); 3449 if (!iscol_local) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Submatrix passed in was not used before, cannot reuse"); 3450 } else { 3451 if (!iscol_local) { 3452 ierr = ISGetSeqIS_Private(mat,iscol,&iscol_local);CHKERRQ(ierr); 3453 } 3454 } 3455 3456 ierr = ISGetLocalSize(iscol,&csize);CHKERRQ(ierr); 3457 ierr = MatCreateSubMatrix_MPIAIJ_nonscalable(mat,isrow,iscol_local,csize,call,newmat);CHKERRQ(ierr); 3458 3459 if (call == MAT_INITIAL_MATRIX) { 3460 ierr = PetscObjectCompose((PetscObject)*newmat,"ISAllGather",(PetscObject)iscol_local);CHKERRQ(ierr); 3461 ierr = ISDestroy(&iscol_local);CHKERRQ(ierr); 3462 } 3463 PetscFunctionReturn(0); 3464 } 3465 3466 /*@C 3467 MatCreateMPIAIJWithSeqAIJ - creates a MPIAIJ matrix using SeqAIJ matrices that contain the "diagonal" 3468 and "off-diagonal" part of the matrix in CSR format. 3469 3470 Collective on MPI_Comm 3471 3472 Input Parameters: 3473 + comm - MPI communicator 3474 . A - "diagonal" portion of matrix 3475 . B - "off-diagonal" portion of matrix, may have empty columns, will be destroyed by this routine 3476 - garray - global index of B columns 3477 3478 Output Parameter: 3479 . mat - the matrix, with input A as its local diagonal matrix 3480 Level: advanced 3481 3482 Notes: 3483 See MatCreateAIJ() for the definition of "diagonal" and "off-diagonal" portion of the matrix. 3484 A becomes part of output mat, B is destroyed by this routine. The user cannot use A and B anymore. 3485 3486 .seealso: MatCreateMPIAIJWithSplitArrays() 3487 @*/ 3488 PetscErrorCode MatCreateMPIAIJWithSeqAIJ(MPI_Comm comm,Mat A,Mat B,const PetscInt garray[],Mat *mat) 3489 { 3490 PetscErrorCode ierr; 3491 Mat_MPIAIJ *maij; 3492 Mat_SeqAIJ *b=(Mat_SeqAIJ*)B->data,*bnew; 3493 PetscInt *oi=b->i,*oj=b->j,i,nz,col; 3494 PetscScalar *oa=b->a; 3495 Mat Bnew; 3496 PetscInt m,n,N; 3497 3498 PetscFunctionBegin; 3499 ierr = MatCreate(comm,mat);CHKERRQ(ierr); 3500 ierr = MatGetSize(A,&m,&n);CHKERRQ(ierr); 3501 if (m != B->rmap->N) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Am %D != Bm %D",m,B->rmap->N); 3502 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); 3503 /* remove check below; When B is created using iscol_o from ISGetSeqIS_SameColDist_Private(), its bs may not be same as A */ 3504 /* 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); */ 3505 3506 /* Get global columns of mat */ 3507 ierr = MPIU_Allreduce(&n,&N,1,MPIU_INT,MPI_SUM,comm);CHKERRQ(ierr); 3508 3509 ierr = MatSetSizes(*mat,m,n,PETSC_DECIDE,N);CHKERRQ(ierr); 3510 ierr = MatSetType(*mat,MATMPIAIJ);CHKERRQ(ierr); 3511 ierr = MatSetBlockSizes(*mat,A->rmap->bs,A->cmap->bs);CHKERRQ(ierr); 3512 maij = (Mat_MPIAIJ*)(*mat)->data; 3513 3514 (*mat)->preallocated = PETSC_TRUE; 3515 3516 ierr = PetscLayoutSetUp((*mat)->rmap);CHKERRQ(ierr); 3517 ierr = PetscLayoutSetUp((*mat)->cmap);CHKERRQ(ierr); 3518 3519 /* Set A as diagonal portion of *mat */ 3520 maij->A = A; 3521 3522 nz = oi[m]; 3523 for (i=0; i<nz; i++) { 3524 col = oj[i]; 3525 oj[i] = garray[col]; 3526 } 3527 3528 /* Set Bnew as off-diagonal portion of *mat */ 3529 ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,m,N,oi,oj,oa,&Bnew);CHKERRQ(ierr); 3530 bnew = (Mat_SeqAIJ*)Bnew->data; 3531 bnew->maxnz = b->maxnz; /* allocated nonzeros of B */ 3532 maij->B = Bnew; 3533 3534 if (B->rmap->N != Bnew->rmap->N) SETERRQ2(PETSC_COMM_SELF,0,"BN %d != BnewN %d",B->rmap->N,Bnew->rmap->N); 3535 3536 b->singlemalloc = PETSC_FALSE; /* B arrays are shared by Bnew */ 3537 b->free_a = PETSC_FALSE; 3538 b->free_ij = PETSC_FALSE; 3539 ierr = MatDestroy(&B);CHKERRQ(ierr); 3540 3541 bnew->singlemalloc = PETSC_TRUE; /* arrays will be freed by MatDestroy(&Bnew) */ 3542 bnew->free_a = PETSC_TRUE; 3543 bnew->free_ij = PETSC_TRUE; 3544 3545 /* condense columns of maij->B */ 3546 ierr = MatSetOption(*mat,MAT_NO_OFF_PROC_ENTRIES,PETSC_TRUE);CHKERRQ(ierr); 3547 ierr = MatAssemblyBegin(*mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3548 ierr = MatAssemblyEnd(*mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3549 ierr = MatSetOption(*mat,MAT_NO_OFF_PROC_ENTRIES,PETSC_FALSE);CHKERRQ(ierr); 3550 ierr = MatSetOption(*mat,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE);CHKERRQ(ierr); 3551 PetscFunctionReturn(0); 3552 } 3553 3554 extern PetscErrorCode MatCreateSubMatrices_MPIAIJ_SingleIS_Local(Mat,PetscInt,const IS[],const IS[],MatReuse,PetscBool,Mat*); 3555 3556 PetscErrorCode MatCreateSubMatrix_MPIAIJ_SameRowDist(Mat mat,IS isrow,IS iscol,IS iscol_local,MatReuse call,Mat *newmat) 3557 { 3558 PetscErrorCode ierr; 3559 PetscInt i,m,n,rstart,row,rend,nz,j,bs,cbs; 3560 PetscInt *ii,*jj,nlocal,*dlens,*olens,dlen,olen,jend,mglobal; 3561 Mat_MPIAIJ *a=(Mat_MPIAIJ*)mat->data; 3562 Mat M,Msub,B=a->B; 3563 MatScalar *aa; 3564 Mat_SeqAIJ *aij; 3565 PetscInt *garray = a->garray,*colsub,Ncols; 3566 PetscInt count,Bn=B->cmap->N,cstart=mat->cmap->rstart,cend=mat->cmap->rend; 3567 IS iscol_sub,iscmap; 3568 const PetscInt *is_idx,*cmap; 3569 PetscBool allcolumns=PETSC_FALSE; 3570 MPI_Comm comm; 3571 3572 PetscFunctionBegin; 3573 ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 3574 3575 if (call == MAT_REUSE_MATRIX) { 3576 ierr = PetscObjectQuery((PetscObject)*newmat,"SubIScol",(PetscObject*)&iscol_sub);CHKERRQ(ierr); 3577 if (!iscol_sub) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"SubIScol passed in was not used before, cannot reuse"); 3578 ierr = ISGetLocalSize(iscol_sub,&count);CHKERRQ(ierr); 3579 3580 ierr = PetscObjectQuery((PetscObject)*newmat,"Subcmap",(PetscObject*)&iscmap);CHKERRQ(ierr); 3581 if (!iscmap) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Subcmap passed in was not used before, cannot reuse"); 3582 3583 ierr = PetscObjectQuery((PetscObject)*newmat,"SubMatrix",(PetscObject*)&Msub);CHKERRQ(ierr); 3584 if (!Msub) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Submatrix passed in was not used before, cannot reuse"); 3585 3586 ierr = MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat,1,&isrow,&iscol_sub,MAT_REUSE_MATRIX,PETSC_FALSE,&Msub);CHKERRQ(ierr); 3587 3588 } else { /* call == MAT_INITIAL_MATRIX) */ 3589 PetscBool flg; 3590 3591 ierr = ISGetLocalSize(iscol,&n);CHKERRQ(ierr); 3592 ierr = ISGetSize(iscol,&Ncols);CHKERRQ(ierr); 3593 3594 /* (1) iscol -> nonscalable iscol_local */ 3595 /* Check for special case: each processor gets entire matrix columns */ 3596 ierr = ISIdentity(iscol_local,&flg);CHKERRQ(ierr); 3597 if (flg && n == mat->cmap->N) allcolumns = PETSC_TRUE; 3598 if (allcolumns) { 3599 iscol_sub = iscol_local; 3600 ierr = PetscObjectReference((PetscObject)iscol_local);CHKERRQ(ierr); 3601 ierr = ISCreateStride(PETSC_COMM_SELF,n,0,1,&iscmap);CHKERRQ(ierr); 3602 3603 } else { 3604 /* (2) iscol_local -> iscol_sub and iscmap. Implementation below requires iscol_local be sorted, it can have duplicate indices */ 3605 PetscInt *idx,*cmap1,k; 3606 ierr = PetscMalloc1(Ncols,&idx);CHKERRQ(ierr); 3607 ierr = PetscMalloc1(Ncols,&cmap1);CHKERRQ(ierr); 3608 ierr = ISGetIndices(iscol_local,&is_idx);CHKERRQ(ierr); 3609 count = 0; 3610 k = 0; 3611 for (i=0; i<Ncols; i++) { 3612 j = is_idx[i]; 3613 if (j >= cstart && j < cend) { 3614 /* diagonal part of mat */ 3615 idx[count] = j; 3616 cmap1[count++] = i; /* column index in submat */ 3617 } else if (Bn) { 3618 /* off-diagonal part of mat */ 3619 if (j == garray[k]) { 3620 idx[count] = j; 3621 cmap1[count++] = i; /* column index in submat */ 3622 } else if (j > garray[k]) { 3623 while (j > garray[k] && k < Bn-1) k++; 3624 if (j == garray[k]) { 3625 idx[count] = j; 3626 cmap1[count++] = i; /* column index in submat */ 3627 } 3628 } 3629 } 3630 } 3631 ierr = ISRestoreIndices(iscol_local,&is_idx);CHKERRQ(ierr); 3632 3633 ierr = ISCreateGeneral(PETSC_COMM_SELF,count,idx,PETSC_OWN_POINTER,&iscol_sub);CHKERRQ(ierr); 3634 ierr = ISGetBlockSize(iscol,&cbs);CHKERRQ(ierr); 3635 ierr = ISSetBlockSize(iscol_sub,cbs);CHKERRQ(ierr); 3636 3637 ierr = ISCreateGeneral(PetscObjectComm((PetscObject)iscol_local),count,cmap1,PETSC_OWN_POINTER,&iscmap);CHKERRQ(ierr); 3638 } 3639 3640 /* (3) Create sequential Msub */ 3641 ierr = MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat,1,&isrow,&iscol_sub,MAT_INITIAL_MATRIX,allcolumns,&Msub);CHKERRQ(ierr); 3642 } 3643 3644 ierr = ISGetLocalSize(iscol_sub,&count);CHKERRQ(ierr); 3645 aij = (Mat_SeqAIJ*)(Msub)->data; 3646 ii = aij->i; 3647 ierr = ISGetIndices(iscmap,&cmap);CHKERRQ(ierr); 3648 3649 /* 3650 m - number of local rows 3651 Ncols - number of columns (same on all processors) 3652 rstart - first row in new global matrix generated 3653 */ 3654 ierr = MatGetSize(Msub,&m,NULL);CHKERRQ(ierr); 3655 3656 if (call == MAT_INITIAL_MATRIX) { 3657 /* (4) Create parallel newmat */ 3658 PetscMPIInt rank,size; 3659 PetscInt csize; 3660 3661 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 3662 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 3663 3664 /* 3665 Determine the number of non-zeros in the diagonal and off-diagonal 3666 portions of the matrix in order to do correct preallocation 3667 */ 3668 3669 /* first get start and end of "diagonal" columns */ 3670 ierr = ISGetLocalSize(iscol,&csize);CHKERRQ(ierr); 3671 if (csize == PETSC_DECIDE) { 3672 ierr = ISGetSize(isrow,&mglobal);CHKERRQ(ierr); 3673 if (mglobal == Ncols) { /* square matrix */ 3674 nlocal = m; 3675 } else { 3676 nlocal = Ncols/size + ((Ncols % size) > rank); 3677 } 3678 } else { 3679 nlocal = csize; 3680 } 3681 ierr = MPI_Scan(&nlocal,&rend,1,MPIU_INT,MPI_SUM,comm);CHKERRQ(ierr); 3682 rstart = rend - nlocal; 3683 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); 3684 3685 /* next, compute all the lengths */ 3686 jj = aij->j; 3687 ierr = PetscMalloc1(2*m+1,&dlens);CHKERRQ(ierr); 3688 olens = dlens + m; 3689 for (i=0; i<m; i++) { 3690 jend = ii[i+1] - ii[i]; 3691 olen = 0; 3692 dlen = 0; 3693 for (j=0; j<jend; j++) { 3694 if (cmap[*jj] < rstart || cmap[*jj] >= rend) olen++; 3695 else dlen++; 3696 jj++; 3697 } 3698 olens[i] = olen; 3699 dlens[i] = dlen; 3700 } 3701 3702 ierr = ISGetBlockSize(isrow,&bs);CHKERRQ(ierr); 3703 ierr = ISGetBlockSize(iscol,&cbs);CHKERRQ(ierr); 3704 3705 ierr = MatCreate(comm,&M);CHKERRQ(ierr); 3706 ierr = MatSetSizes(M,m,nlocal,PETSC_DECIDE,Ncols);CHKERRQ(ierr); 3707 ierr = MatSetBlockSizes(M,bs,cbs);CHKERRQ(ierr); 3708 ierr = MatSetType(M,((PetscObject)mat)->type_name);CHKERRQ(ierr); 3709 ierr = MatMPIAIJSetPreallocation(M,0,dlens,0,olens);CHKERRQ(ierr); 3710 ierr = PetscFree(dlens);CHKERRQ(ierr); 3711 3712 } else { /* call == MAT_REUSE_MATRIX */ 3713 M = *newmat; 3714 ierr = MatGetLocalSize(M,&i,NULL);CHKERRQ(ierr); 3715 if (i != m) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Previous matrix must be same size/layout as request"); 3716 ierr = MatZeroEntries(M);CHKERRQ(ierr); 3717 /* 3718 The next two lines are needed so we may call MatSetValues_MPIAIJ() below directly, 3719 rather than the slower MatSetValues(). 3720 */ 3721 M->was_assembled = PETSC_TRUE; 3722 M->assembled = PETSC_FALSE; 3723 } 3724 3725 /* (5) Set values of Msub to *newmat */ 3726 ierr = PetscMalloc1(count,&colsub);CHKERRQ(ierr); 3727 ierr = MatGetOwnershipRange(M,&rstart,NULL);CHKERRQ(ierr); 3728 3729 jj = aij->j; 3730 aa = aij->a; 3731 for (i=0; i<m; i++) { 3732 row = rstart + i; 3733 nz = ii[i+1] - ii[i]; 3734 for (j=0; j<nz; j++) colsub[j] = cmap[jj[j]]; 3735 ierr = MatSetValues_MPIAIJ(M,1,&row,nz,colsub,aa,INSERT_VALUES);CHKERRQ(ierr); 3736 jj += nz; aa += nz; 3737 } 3738 ierr = ISRestoreIndices(iscmap,&cmap);CHKERRQ(ierr); 3739 3740 ierr = MatAssemblyBegin(M,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3741 ierr = MatAssemblyEnd(M,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3742 3743 ierr = PetscFree(colsub);CHKERRQ(ierr); 3744 3745 /* save Msub, iscol_sub and iscmap used in processor for next request */ 3746 if (call == MAT_INITIAL_MATRIX) { 3747 *newmat = M; 3748 ierr = PetscObjectCompose((PetscObject)(*newmat),"SubMatrix",(PetscObject)Msub);CHKERRQ(ierr); 3749 ierr = MatDestroy(&Msub);CHKERRQ(ierr); 3750 3751 ierr = PetscObjectCompose((PetscObject)(*newmat),"SubIScol",(PetscObject)iscol_sub);CHKERRQ(ierr); 3752 ierr = ISDestroy(&iscol_sub);CHKERRQ(ierr); 3753 3754 ierr = PetscObjectCompose((PetscObject)(*newmat),"Subcmap",(PetscObject)iscmap);CHKERRQ(ierr); 3755 ierr = ISDestroy(&iscmap);CHKERRQ(ierr); 3756 3757 if (iscol_local) { 3758 ierr = PetscObjectCompose((PetscObject)(*newmat),"ISAllGather",(PetscObject)iscol_local);CHKERRQ(ierr); 3759 ierr = ISDestroy(&iscol_local);CHKERRQ(ierr); 3760 } 3761 } 3762 PetscFunctionReturn(0); 3763 } 3764 3765 /* 3766 Not great since it makes two copies of the submatrix, first an SeqAIJ 3767 in local and then by concatenating the local matrices the end result. 3768 Writing it directly would be much like MatCreateSubMatrices_MPIAIJ() 3769 3770 Note: This requires a sequential iscol with all indices. 3771 */ 3772 PetscErrorCode MatCreateSubMatrix_MPIAIJ_nonscalable(Mat mat,IS isrow,IS iscol,PetscInt csize,MatReuse call,Mat *newmat) 3773 { 3774 PetscErrorCode ierr; 3775 PetscMPIInt rank,size; 3776 PetscInt i,m,n,rstart,row,rend,nz,*cwork,j,bs,cbs; 3777 PetscInt *ii,*jj,nlocal,*dlens,*olens,dlen,olen,jend,mglobal; 3778 Mat M,Mreuse; 3779 MatScalar *aa,*vwork; 3780 MPI_Comm comm; 3781 Mat_SeqAIJ *aij; 3782 PetscBool colflag,allcolumns=PETSC_FALSE; 3783 3784 PetscFunctionBegin; 3785 ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 3786 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 3787 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 3788 3789 /* Check for special case: each processor gets entire matrix columns */ 3790 ierr = ISIdentity(iscol,&colflag);CHKERRQ(ierr); 3791 ierr = ISGetLocalSize(iscol,&n);CHKERRQ(ierr); 3792 if (colflag && n == mat->cmap->N) allcolumns = PETSC_TRUE; 3793 3794 if (call == MAT_REUSE_MATRIX) { 3795 ierr = PetscObjectQuery((PetscObject)*newmat,"SubMatrix",(PetscObject*)&Mreuse);CHKERRQ(ierr); 3796 if (!Mreuse) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Submatrix passed in was not used before, cannot reuse"); 3797 ierr = MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat,1,&isrow,&iscol,MAT_REUSE_MATRIX,allcolumns,&Mreuse);CHKERRQ(ierr); 3798 } else { 3799 ierr = MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat,1,&isrow,&iscol,MAT_INITIAL_MATRIX,allcolumns,&Mreuse);CHKERRQ(ierr); 3800 } 3801 3802 /* 3803 m - number of local rows 3804 n - number of columns (same on all processors) 3805 rstart - first row in new global matrix generated 3806 */ 3807 ierr = MatGetSize(Mreuse,&m,&n);CHKERRQ(ierr); 3808 ierr = MatGetBlockSizes(Mreuse,&bs,&cbs);CHKERRQ(ierr); 3809 if (call == MAT_INITIAL_MATRIX) { 3810 aij = (Mat_SeqAIJ*)(Mreuse)->data; 3811 ii = aij->i; 3812 jj = aij->j; 3813 3814 /* 3815 Determine the number of non-zeros in the diagonal and off-diagonal 3816 portions of the matrix in order to do correct preallocation 3817 */ 3818 3819 /* first get start and end of "diagonal" columns */ 3820 if (csize == PETSC_DECIDE) { 3821 ierr = ISGetSize(isrow,&mglobal);CHKERRQ(ierr); 3822 if (mglobal == n) { /* square matrix */ 3823 nlocal = m; 3824 } else { 3825 nlocal = n/size + ((n % size) > rank); 3826 } 3827 } else { 3828 nlocal = csize; 3829 } 3830 ierr = MPI_Scan(&nlocal,&rend,1,MPIU_INT,MPI_SUM,comm);CHKERRQ(ierr); 3831 rstart = rend - nlocal; 3832 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); 3833 3834 /* next, compute all the lengths */ 3835 ierr = PetscMalloc1(2*m+1,&dlens);CHKERRQ(ierr); 3836 olens = dlens + m; 3837 for (i=0; i<m; i++) { 3838 jend = ii[i+1] - ii[i]; 3839 olen = 0; 3840 dlen = 0; 3841 for (j=0; j<jend; j++) { 3842 if (*jj < rstart || *jj >= rend) olen++; 3843 else dlen++; 3844 jj++; 3845 } 3846 olens[i] = olen; 3847 dlens[i] = dlen; 3848 } 3849 ierr = MatCreate(comm,&M);CHKERRQ(ierr); 3850 ierr = MatSetSizes(M,m,nlocal,PETSC_DECIDE,n);CHKERRQ(ierr); 3851 ierr = MatSetBlockSizes(M,bs,cbs);CHKERRQ(ierr); 3852 ierr = MatSetType(M,((PetscObject)mat)->type_name);CHKERRQ(ierr); 3853 ierr = MatMPIAIJSetPreallocation(M,0,dlens,0,olens);CHKERRQ(ierr); 3854 ierr = PetscFree(dlens);CHKERRQ(ierr); 3855 } else { 3856 PetscInt ml,nl; 3857 3858 M = *newmat; 3859 ierr = MatGetLocalSize(M,&ml,&nl);CHKERRQ(ierr); 3860 if (ml != m) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Previous matrix must be same size/layout as request"); 3861 ierr = MatZeroEntries(M);CHKERRQ(ierr); 3862 /* 3863 The next two lines are needed so we may call MatSetValues_MPIAIJ() below directly, 3864 rather than the slower MatSetValues(). 3865 */ 3866 M->was_assembled = PETSC_TRUE; 3867 M->assembled = PETSC_FALSE; 3868 } 3869 ierr = MatGetOwnershipRange(M,&rstart,&rend);CHKERRQ(ierr); 3870 aij = (Mat_SeqAIJ*)(Mreuse)->data; 3871 ii = aij->i; 3872 jj = aij->j; 3873 aa = aij->a; 3874 for (i=0; i<m; i++) { 3875 row = rstart + i; 3876 nz = ii[i+1] - ii[i]; 3877 cwork = jj; jj += nz; 3878 vwork = aa; aa += nz; 3879 ierr = MatSetValues_MPIAIJ(M,1,&row,nz,cwork,vwork,INSERT_VALUES);CHKERRQ(ierr); 3880 } 3881 3882 ierr = MatAssemblyBegin(M,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3883 ierr = MatAssemblyEnd(M,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3884 *newmat = M; 3885 3886 /* save submatrix used in processor for next request */ 3887 if (call == MAT_INITIAL_MATRIX) { 3888 ierr = PetscObjectCompose((PetscObject)M,"SubMatrix",(PetscObject)Mreuse);CHKERRQ(ierr); 3889 ierr = MatDestroy(&Mreuse);CHKERRQ(ierr); 3890 } 3891 PetscFunctionReturn(0); 3892 } 3893 3894 PetscErrorCode MatMPIAIJSetPreallocationCSR_MPIAIJ(Mat B,const PetscInt Ii[],const PetscInt J[],const PetscScalar v[]) 3895 { 3896 PetscInt m,cstart, cend,j,nnz,i,d; 3897 PetscInt *d_nnz,*o_nnz,nnz_max = 0,rstart,ii; 3898 const PetscInt *JJ; 3899 PetscScalar *values; 3900 PetscErrorCode ierr; 3901 PetscBool nooffprocentries; 3902 3903 PetscFunctionBegin; 3904 if (Ii && Ii[0]) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Ii[0] must be 0 it is %D",Ii[0]); 3905 3906 ierr = PetscLayoutSetUp(B->rmap);CHKERRQ(ierr); 3907 ierr = PetscLayoutSetUp(B->cmap);CHKERRQ(ierr); 3908 m = B->rmap->n; 3909 cstart = B->cmap->rstart; 3910 cend = B->cmap->rend; 3911 rstart = B->rmap->rstart; 3912 3913 ierr = PetscCalloc2(m,&d_nnz,m,&o_nnz);CHKERRQ(ierr); 3914 3915 #if defined(PETSC_USE_DEBUG) 3916 for (i=0; i<m && Ii; i++) { 3917 nnz = Ii[i+1]- Ii[i]; 3918 JJ = J + Ii[i]; 3919 if (nnz < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Local row %D has a negative %D number of columns",i,nnz); 3920 if (nnz && (JJ[0] < 0)) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Row %D starts with negative column index",i,JJ[0]); 3921 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); 3922 } 3923 #endif 3924 3925 for (i=0; i<m && Ii; i++) { 3926 nnz = Ii[i+1]- Ii[i]; 3927 JJ = J + Ii[i]; 3928 nnz_max = PetscMax(nnz_max,nnz); 3929 d = 0; 3930 for (j=0; j<nnz; j++) { 3931 if (cstart <= JJ[j] && JJ[j] < cend) d++; 3932 } 3933 d_nnz[i] = d; 3934 o_nnz[i] = nnz - d; 3935 } 3936 ierr = MatMPIAIJSetPreallocation(B,0,d_nnz,0,o_nnz);CHKERRQ(ierr); 3937 ierr = PetscFree2(d_nnz,o_nnz);CHKERRQ(ierr); 3938 3939 if (v) values = (PetscScalar*)v; 3940 else { 3941 ierr = PetscCalloc1(nnz_max+1,&values);CHKERRQ(ierr); 3942 } 3943 3944 for (i=0; i<m && Ii; i++) { 3945 ii = i + rstart; 3946 nnz = Ii[i+1]- Ii[i]; 3947 ierr = MatSetValues_MPIAIJ(B,1,&ii,nnz,J+Ii[i],values+(v ? Ii[i] : 0),INSERT_VALUES);CHKERRQ(ierr); 3948 } 3949 nooffprocentries = B->nooffprocentries; 3950 B->nooffprocentries = PETSC_TRUE; 3951 ierr = MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3952 ierr = MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3953 B->nooffprocentries = nooffprocentries; 3954 3955 if (!v) { 3956 ierr = PetscFree(values);CHKERRQ(ierr); 3957 } 3958 ierr = MatSetOption(B,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE);CHKERRQ(ierr); 3959 PetscFunctionReturn(0); 3960 } 3961 3962 /*@ 3963 MatMPIAIJSetPreallocationCSR - Allocates memory for a sparse parallel matrix in AIJ format 3964 (the default parallel PETSc format). 3965 3966 Collective on MPI_Comm 3967 3968 Input Parameters: 3969 + B - the matrix 3970 . i - the indices into j for the start of each local row (starts with zero) 3971 . j - the column indices for each local row (starts with zero) 3972 - v - optional values in the matrix 3973 3974 Level: developer 3975 3976 Notes: 3977 The i, j, and v arrays ARE copied by this routine into the internal format used by PETSc; 3978 thus you CANNOT change the matrix entries by changing the values of v[] after you have 3979 called this routine. Use MatCreateMPIAIJWithSplitArrays() to avoid needing to copy the arrays. 3980 3981 The i and j indices are 0 based, and i indices are indices corresponding to the local j array. 3982 3983 The format which is used for the sparse matrix input, is equivalent to a 3984 row-major ordering.. i.e for the following matrix, the input data expected is 3985 as shown 3986 3987 $ 1 0 0 3988 $ 2 0 3 P0 3989 $ ------- 3990 $ 4 5 6 P1 3991 $ 3992 $ Process0 [P0]: rows_owned=[0,1] 3993 $ i = {0,1,3} [size = nrow+1 = 2+1] 3994 $ j = {0,0,2} [size = 3] 3995 $ v = {1,2,3} [size = 3] 3996 $ 3997 $ Process1 [P1]: rows_owned=[2] 3998 $ i = {0,3} [size = nrow+1 = 1+1] 3999 $ j = {0,1,2} [size = 3] 4000 $ v = {4,5,6} [size = 3] 4001 4002 .keywords: matrix, aij, compressed row, sparse, parallel 4003 4004 .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatCreateAIJ(), MATMPIAIJ, 4005 MatCreateSeqAIJWithArrays(), MatCreateMPIAIJWithSplitArrays() 4006 @*/ 4007 PetscErrorCode MatMPIAIJSetPreallocationCSR(Mat B,const PetscInt i[],const PetscInt j[], const PetscScalar v[]) 4008 { 4009 PetscErrorCode ierr; 4010 4011 PetscFunctionBegin; 4012 ierr = PetscTryMethod(B,"MatMPIAIJSetPreallocationCSR_C",(Mat,const PetscInt[],const PetscInt[],const PetscScalar[]),(B,i,j,v));CHKERRQ(ierr); 4013 PetscFunctionReturn(0); 4014 } 4015 4016 /*@C 4017 MatMPIAIJSetPreallocation - Preallocates memory for a sparse parallel matrix in AIJ format 4018 (the default parallel PETSc format). For good matrix assembly performance 4019 the user should preallocate the matrix storage by setting the parameters 4020 d_nz (or d_nnz) and o_nz (or o_nnz). By setting these parameters accurately, 4021 performance can be increased by more than a factor of 50. 4022 4023 Collective on MPI_Comm 4024 4025 Input Parameters: 4026 + B - the matrix 4027 . d_nz - number of nonzeros per row in DIAGONAL portion of local submatrix 4028 (same value is used for all local rows) 4029 . d_nnz - array containing the number of nonzeros in the various rows of the 4030 DIAGONAL portion of the local submatrix (possibly different for each row) 4031 or NULL (PETSC_NULL_INTEGER in Fortran), if d_nz is used to specify the nonzero structure. 4032 The size of this array is equal to the number of local rows, i.e 'm'. 4033 For matrices that will be factored, you must leave room for (and set) 4034 the diagonal entry even if it is zero. 4035 . o_nz - number of nonzeros per row in the OFF-DIAGONAL portion of local 4036 submatrix (same value is used for all local rows). 4037 - o_nnz - array containing the number of nonzeros in the various rows of the 4038 OFF-DIAGONAL portion of the local submatrix (possibly different for 4039 each row) or NULL (PETSC_NULL_INTEGER in Fortran), if o_nz is used to specify the nonzero 4040 structure. The size of this array is equal to the number 4041 of local rows, i.e 'm'. 4042 4043 If the *_nnz parameter is given then the *_nz parameter is ignored 4044 4045 The AIJ format (also called the Yale sparse matrix format or 4046 compressed row storage (CSR)), is fully compatible with standard Fortran 77 4047 storage. The stored row and column indices begin with zero. 4048 See Users-Manual: ch_mat for details. 4049 4050 The parallel matrix is partitioned such that the first m0 rows belong to 4051 process 0, the next m1 rows belong to process 1, the next m2 rows belong 4052 to process 2 etc.. where m0,m1,m2... are the input parameter 'm'. 4053 4054 The DIAGONAL portion of the local submatrix of a processor can be defined 4055 as the submatrix which is obtained by extraction the part corresponding to 4056 the rows r1-r2 and columns c1-c2 of the global matrix, where r1 is the 4057 first row that belongs to the processor, r2 is the last row belonging to 4058 the this processor, and c1-c2 is range of indices of the local part of a 4059 vector suitable for applying the matrix to. This is an mxn matrix. In the 4060 common case of a square matrix, the row and column ranges are the same and 4061 the DIAGONAL part is also square. The remaining portion of the local 4062 submatrix (mxN) constitute the OFF-DIAGONAL portion. 4063 4064 If o_nnz, d_nnz are specified, then o_nz, and d_nz are ignored. 4065 4066 You can call MatGetInfo() to get information on how effective the preallocation was; 4067 for example the fields mallocs,nz_allocated,nz_used,nz_unneeded; 4068 You can also run with the option -info and look for messages with the string 4069 malloc in them to see if additional memory allocation was needed. 4070 4071 Example usage: 4072 4073 Consider the following 8x8 matrix with 34 non-zero values, that is 4074 assembled across 3 processors. Lets assume that proc0 owns 3 rows, 4075 proc1 owns 3 rows, proc2 owns 2 rows. This division can be shown 4076 as follows: 4077 4078 .vb 4079 1 2 0 | 0 3 0 | 0 4 4080 Proc0 0 5 6 | 7 0 0 | 8 0 4081 9 0 10 | 11 0 0 | 12 0 4082 ------------------------------------- 4083 13 0 14 | 15 16 17 | 0 0 4084 Proc1 0 18 0 | 19 20 21 | 0 0 4085 0 0 0 | 22 23 0 | 24 0 4086 ------------------------------------- 4087 Proc2 25 26 27 | 0 0 28 | 29 0 4088 30 0 0 | 31 32 33 | 0 34 4089 .ve 4090 4091 This can be represented as a collection of submatrices as: 4092 4093 .vb 4094 A B C 4095 D E F 4096 G H I 4097 .ve 4098 4099 Where the submatrices A,B,C are owned by proc0, D,E,F are 4100 owned by proc1, G,H,I are owned by proc2. 4101 4102 The 'm' parameters for proc0,proc1,proc2 are 3,3,2 respectively. 4103 The 'n' parameters for proc0,proc1,proc2 are 3,3,2 respectively. 4104 The 'M','N' parameters are 8,8, and have the same values on all procs. 4105 4106 The DIAGONAL submatrices corresponding to proc0,proc1,proc2 are 4107 submatrices [A], [E], [I] respectively. The OFF-DIAGONAL submatrices 4108 corresponding to proc0,proc1,proc2 are [BC], [DF], [GH] respectively. 4109 Internally, each processor stores the DIAGONAL part, and the OFF-DIAGONAL 4110 part as SeqAIJ matrices. for eg: proc1 will store [E] as a SeqAIJ 4111 matrix, ans [DF] as another SeqAIJ matrix. 4112 4113 When d_nz, o_nz parameters are specified, d_nz storage elements are 4114 allocated for every row of the local diagonal submatrix, and o_nz 4115 storage locations are allocated for every row of the OFF-DIAGONAL submat. 4116 One way to choose d_nz and o_nz is to use the max nonzerors per local 4117 rows for each of the local DIAGONAL, and the OFF-DIAGONAL submatrices. 4118 In this case, the values of d_nz,o_nz are: 4119 .vb 4120 proc0 : dnz = 2, o_nz = 2 4121 proc1 : dnz = 3, o_nz = 2 4122 proc2 : dnz = 1, o_nz = 4 4123 .ve 4124 We are allocating m*(d_nz+o_nz) storage locations for every proc. This 4125 translates to 3*(2+2)=12 for proc0, 3*(3+2)=15 for proc1, 2*(1+4)=10 4126 for proc3. i.e we are using 12+15+10=37 storage locations to store 4127 34 values. 4128 4129 When d_nnz, o_nnz parameters are specified, the storage is specified 4130 for every row, coresponding to both DIAGONAL and OFF-DIAGONAL submatrices. 4131 In the above case the values for d_nnz,o_nnz are: 4132 .vb 4133 proc0: d_nnz = [2,2,2] and o_nnz = [2,2,2] 4134 proc1: d_nnz = [3,3,2] and o_nnz = [2,1,1] 4135 proc2: d_nnz = [1,1] and o_nnz = [4,4] 4136 .ve 4137 Here the space allocated is sum of all the above values i.e 34, and 4138 hence pre-allocation is perfect. 4139 4140 Level: intermediate 4141 4142 .keywords: matrix, aij, compressed row, sparse, parallel 4143 4144 .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatCreateAIJ(), MatMPIAIJSetPreallocationCSR(), 4145 MATMPIAIJ, MatGetInfo(), PetscSplitOwnership() 4146 @*/ 4147 PetscErrorCode MatMPIAIJSetPreallocation(Mat B,PetscInt d_nz,const PetscInt d_nnz[],PetscInt o_nz,const PetscInt o_nnz[]) 4148 { 4149 PetscErrorCode ierr; 4150 4151 PetscFunctionBegin; 4152 PetscValidHeaderSpecific(B,MAT_CLASSID,1); 4153 PetscValidType(B,1); 4154 ierr = PetscTryMethod(B,"MatMPIAIJSetPreallocation_C",(Mat,PetscInt,const PetscInt[],PetscInt,const PetscInt[]),(B,d_nz,d_nnz,o_nz,o_nnz));CHKERRQ(ierr); 4155 PetscFunctionReturn(0); 4156 } 4157 4158 /*@ 4159 MatCreateMPIAIJWithArrays - creates a MPI AIJ matrix using arrays that contain in standard 4160 CSR format the local rows. 4161 4162 Collective on MPI_Comm 4163 4164 Input Parameters: 4165 + comm - MPI communicator 4166 . m - number of local rows (Cannot be PETSC_DECIDE) 4167 . n - This value should be the same as the local size used in creating the 4168 x vector for the matrix-vector product y = Ax. (or PETSC_DECIDE to have 4169 calculated if N is given) For square matrices n is almost always m. 4170 . M - number of global rows (or PETSC_DETERMINE to have calculated if m is given) 4171 . N - number of global columns (or PETSC_DETERMINE to have calculated if n is given) 4172 . i - row indices; that is i[0] = 0, i[row] = i[row-1] + number of elements in that row of the matrix 4173 . j - column indices 4174 - a - matrix values 4175 4176 Output Parameter: 4177 . mat - the matrix 4178 4179 Level: intermediate 4180 4181 Notes: 4182 The i, j, and a arrays ARE copied by this routine into the internal format used by PETSc; 4183 thus you CANNOT change the matrix entries by changing the values of a[] after you have 4184 called this routine. Use MatCreateMPIAIJWithSplitArrays() to avoid needing to copy the arrays. 4185 4186 The i and j indices are 0 based, and i indices are indices corresponding to the local j array. 4187 4188 The format which is used for the sparse matrix input, is equivalent to a 4189 row-major ordering.. i.e for the following matrix, the input data expected is 4190 as shown 4191 4192 $ 1 0 0 4193 $ 2 0 3 P0 4194 $ ------- 4195 $ 4 5 6 P1 4196 $ 4197 $ Process0 [P0]: rows_owned=[0,1] 4198 $ i = {0,1,3} [size = nrow+1 = 2+1] 4199 $ j = {0,0,2} [size = 3] 4200 $ v = {1,2,3} [size = 3] 4201 $ 4202 $ Process1 [P1]: rows_owned=[2] 4203 $ i = {0,3} [size = nrow+1 = 1+1] 4204 $ j = {0,1,2} [size = 3] 4205 $ v = {4,5,6} [size = 3] 4206 4207 .keywords: matrix, aij, compressed row, sparse, parallel 4208 4209 .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatMPIAIJSetPreallocationCSR(), 4210 MATMPIAIJ, MatCreateAIJ(), MatCreateMPIAIJWithSplitArrays() 4211 @*/ 4212 PetscErrorCode MatCreateMPIAIJWithArrays(MPI_Comm comm,PetscInt m,PetscInt n,PetscInt M,PetscInt N,const PetscInt i[],const PetscInt j[],const PetscScalar a[],Mat *mat) 4213 { 4214 PetscErrorCode ierr; 4215 4216 PetscFunctionBegin; 4217 if (i && i[0]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"i (row indices) must start with 0"); 4218 if (m < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"local number of rows (m) cannot be PETSC_DECIDE, or negative"); 4219 ierr = MatCreate(comm,mat);CHKERRQ(ierr); 4220 ierr = MatSetSizes(*mat,m,n,M,N);CHKERRQ(ierr); 4221 /* ierr = MatSetBlockSizes(M,bs,cbs);CHKERRQ(ierr); */ 4222 ierr = MatSetType(*mat,MATMPIAIJ);CHKERRQ(ierr); 4223 ierr = MatMPIAIJSetPreallocationCSR(*mat,i,j,a);CHKERRQ(ierr); 4224 PetscFunctionReturn(0); 4225 } 4226 4227 /*@C 4228 MatCreateAIJ - Creates a sparse parallel matrix in AIJ format 4229 (the default parallel PETSc format). For good matrix assembly performance 4230 the user should preallocate the matrix storage by setting the parameters 4231 d_nz (or d_nnz) and o_nz (or o_nnz). By setting these parameters accurately, 4232 performance can be increased by more than a factor of 50. 4233 4234 Collective on MPI_Comm 4235 4236 Input Parameters: 4237 + comm - MPI communicator 4238 . m - number of local rows (or PETSC_DECIDE to have calculated if M is given) 4239 This value should be the same as the local size used in creating the 4240 y vector for the matrix-vector product y = Ax. 4241 . n - This value should be the same as the local size used in creating the 4242 x vector for the matrix-vector product y = Ax. (or PETSC_DECIDE to have 4243 calculated if N is given) For square matrices n is almost always m. 4244 . M - number of global rows (or PETSC_DETERMINE to have calculated if m is given) 4245 . N - number of global columns (or PETSC_DETERMINE to have calculated if n is given) 4246 . d_nz - number of nonzeros per row in DIAGONAL portion of local submatrix 4247 (same value is used for all local rows) 4248 . d_nnz - array containing the number of nonzeros in the various rows of the 4249 DIAGONAL portion of the local submatrix (possibly different for each row) 4250 or NULL, if d_nz is used to specify the nonzero structure. 4251 The size of this array is equal to the number of local rows, i.e 'm'. 4252 . o_nz - number of nonzeros per row in the OFF-DIAGONAL portion of local 4253 submatrix (same value is used for all local rows). 4254 - o_nnz - array containing the number of nonzeros in the various rows of the 4255 OFF-DIAGONAL portion of the local submatrix (possibly different for 4256 each row) or NULL, if o_nz is used to specify the nonzero 4257 structure. The size of this array is equal to the number 4258 of local rows, i.e 'm'. 4259 4260 Output Parameter: 4261 . A - the matrix 4262 4263 It is recommended that one use the MatCreate(), MatSetType() and/or MatSetFromOptions(), 4264 MatXXXXSetPreallocation() paradgm instead of this routine directly. 4265 [MatXXXXSetPreallocation() is, for example, MatSeqAIJSetPreallocation] 4266 4267 Notes: 4268 If the *_nnz parameter is given then the *_nz parameter is ignored 4269 4270 m,n,M,N parameters specify the size of the matrix, and its partitioning across 4271 processors, while d_nz,d_nnz,o_nz,o_nnz parameters specify the approximate 4272 storage requirements for this matrix. 4273 4274 If PETSC_DECIDE or PETSC_DETERMINE is used for a particular argument on one 4275 processor than it must be used on all processors that share the object for 4276 that argument. 4277 4278 The user MUST specify either the local or global matrix dimensions 4279 (possibly both). 4280 4281 The parallel matrix is partitioned across processors such that the 4282 first m0 rows belong to process 0, the next m1 rows belong to 4283 process 1, the next m2 rows belong to process 2 etc.. where 4284 m0,m1,m2,.. are the input parameter 'm'. i.e each processor stores 4285 values corresponding to [m x N] submatrix. 4286 4287 The columns are logically partitioned with the n0 columns belonging 4288 to 0th partition, the next n1 columns belonging to the next 4289 partition etc.. where n0,n1,n2... are the input parameter 'n'. 4290 4291 The DIAGONAL portion of the local submatrix on any given processor 4292 is the submatrix corresponding to the rows and columns m,n 4293 corresponding to the given processor. i.e diagonal matrix on 4294 process 0 is [m0 x n0], diagonal matrix on process 1 is [m1 x n1] 4295 etc. The remaining portion of the local submatrix [m x (N-n)] 4296 constitute the OFF-DIAGONAL portion. The example below better 4297 illustrates this concept. 4298 4299 For a square global matrix we define each processor's diagonal portion 4300 to be its local rows and the corresponding columns (a square submatrix); 4301 each processor's off-diagonal portion encompasses the remainder of the 4302 local matrix (a rectangular submatrix). 4303 4304 If o_nnz, d_nnz are specified, then o_nz, and d_nz are ignored. 4305 4306 When calling this routine with a single process communicator, a matrix of 4307 type SEQAIJ is returned. If a matrix of type MPIAIJ is desired for this 4308 type of communicator, use the construction mechanism 4309 .vb 4310 MatCreate(...,&A); MatSetType(A,MATMPIAIJ); MatSetSizes(A, m,n,M,N); MatMPIAIJSetPreallocation(A,...); 4311 .ve 4312 4313 $ MatCreate(...,&A); 4314 $ MatSetType(A,MATMPIAIJ); 4315 $ MatSetSizes(A, m,n,M,N); 4316 $ MatMPIAIJSetPreallocation(A,...); 4317 4318 By default, this format uses inodes (identical nodes) when possible. 4319 We search for consecutive rows with the same nonzero structure, thereby 4320 reusing matrix information to achieve increased efficiency. 4321 4322 Options Database Keys: 4323 + -mat_no_inode - Do not use inodes 4324 - -mat_inode_limit <limit> - Sets inode limit (max limit=5) 4325 4326 4327 4328 Example usage: 4329 4330 Consider the following 8x8 matrix with 34 non-zero values, that is 4331 assembled across 3 processors. Lets assume that proc0 owns 3 rows, 4332 proc1 owns 3 rows, proc2 owns 2 rows. This division can be shown 4333 as follows 4334 4335 .vb 4336 1 2 0 | 0 3 0 | 0 4 4337 Proc0 0 5 6 | 7 0 0 | 8 0 4338 9 0 10 | 11 0 0 | 12 0 4339 ------------------------------------- 4340 13 0 14 | 15 16 17 | 0 0 4341 Proc1 0 18 0 | 19 20 21 | 0 0 4342 0 0 0 | 22 23 0 | 24 0 4343 ------------------------------------- 4344 Proc2 25 26 27 | 0 0 28 | 29 0 4345 30 0 0 | 31 32 33 | 0 34 4346 .ve 4347 4348 This can be represented as a collection of submatrices as 4349 4350 .vb 4351 A B C 4352 D E F 4353 G H I 4354 .ve 4355 4356 Where the submatrices A,B,C are owned by proc0, D,E,F are 4357 owned by proc1, G,H,I are owned by proc2. 4358 4359 The 'm' parameters for proc0,proc1,proc2 are 3,3,2 respectively. 4360 The 'n' parameters for proc0,proc1,proc2 are 3,3,2 respectively. 4361 The 'M','N' parameters are 8,8, and have the same values on all procs. 4362 4363 The DIAGONAL submatrices corresponding to proc0,proc1,proc2 are 4364 submatrices [A], [E], [I] respectively. The OFF-DIAGONAL submatrices 4365 corresponding to proc0,proc1,proc2 are [BC], [DF], [GH] respectively. 4366 Internally, each processor stores the DIAGONAL part, and the OFF-DIAGONAL 4367 part as SeqAIJ matrices. for eg: proc1 will store [E] as a SeqAIJ 4368 matrix, ans [DF] as another SeqAIJ matrix. 4369 4370 When d_nz, o_nz parameters are specified, d_nz storage elements are 4371 allocated for every row of the local diagonal submatrix, and o_nz 4372 storage locations are allocated for every row of the OFF-DIAGONAL submat. 4373 One way to choose d_nz and o_nz is to use the max nonzerors per local 4374 rows for each of the local DIAGONAL, and the OFF-DIAGONAL submatrices. 4375 In this case, the values of d_nz,o_nz are 4376 .vb 4377 proc0 : dnz = 2, o_nz = 2 4378 proc1 : dnz = 3, o_nz = 2 4379 proc2 : dnz = 1, o_nz = 4 4380 .ve 4381 We are allocating m*(d_nz+o_nz) storage locations for every proc. This 4382 translates to 3*(2+2)=12 for proc0, 3*(3+2)=15 for proc1, 2*(1+4)=10 4383 for proc3. i.e we are using 12+15+10=37 storage locations to store 4384 34 values. 4385 4386 When d_nnz, o_nnz parameters are specified, the storage is specified 4387 for every row, coresponding to both DIAGONAL and OFF-DIAGONAL submatrices. 4388 In the above case the values for d_nnz,o_nnz are 4389 .vb 4390 proc0: d_nnz = [2,2,2] and o_nnz = [2,2,2] 4391 proc1: d_nnz = [3,3,2] and o_nnz = [2,1,1] 4392 proc2: d_nnz = [1,1] and o_nnz = [4,4] 4393 .ve 4394 Here the space allocated is sum of all the above values i.e 34, and 4395 hence pre-allocation is perfect. 4396 4397 Level: intermediate 4398 4399 .keywords: matrix, aij, compressed row, sparse, parallel 4400 4401 .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatMPIAIJSetPreallocationCSR(), 4402 MATMPIAIJ, MatCreateMPIAIJWithArrays() 4403 @*/ 4404 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) 4405 { 4406 PetscErrorCode ierr; 4407 PetscMPIInt size; 4408 4409 PetscFunctionBegin; 4410 ierr = MatCreate(comm,A);CHKERRQ(ierr); 4411 ierr = MatSetSizes(*A,m,n,M,N);CHKERRQ(ierr); 4412 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 4413 if (size > 1) { 4414 ierr = MatSetType(*A,MATMPIAIJ);CHKERRQ(ierr); 4415 ierr = MatMPIAIJSetPreallocation(*A,d_nz,d_nnz,o_nz,o_nnz);CHKERRQ(ierr); 4416 } else { 4417 ierr = MatSetType(*A,MATSEQAIJ);CHKERRQ(ierr); 4418 ierr = MatSeqAIJSetPreallocation(*A,d_nz,d_nnz);CHKERRQ(ierr); 4419 } 4420 PetscFunctionReturn(0); 4421 } 4422 4423 PetscErrorCode MatMPIAIJGetSeqAIJ(Mat A,Mat *Ad,Mat *Ao,const PetscInt *colmap[]) 4424 { 4425 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 4426 PetscBool flg; 4427 PetscErrorCode ierr; 4428 4429 PetscFunctionBegin; 4430 ierr = PetscStrbeginswith(((PetscObject)A)->type_name,MATMPIAIJ,&flg);CHKERRQ(ierr); 4431 if (!flg) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"This function requires a MATMPIAIJ matrix as input"); 4432 if (Ad) *Ad = a->A; 4433 if (Ao) *Ao = a->B; 4434 if (colmap) *colmap = a->garray; 4435 PetscFunctionReturn(0); 4436 } 4437 4438 PetscErrorCode MatCreateMPIMatConcatenateSeqMat_MPIAIJ(MPI_Comm comm,Mat inmat,PetscInt n,MatReuse scall,Mat *outmat) 4439 { 4440 PetscErrorCode ierr; 4441 PetscInt m,N,i,rstart,nnz,Ii; 4442 PetscInt *indx; 4443 PetscScalar *values; 4444 4445 PetscFunctionBegin; 4446 ierr = MatGetSize(inmat,&m,&N);CHKERRQ(ierr); 4447 if (scall == MAT_INITIAL_MATRIX) { /* symbolic phase */ 4448 PetscInt *dnz,*onz,sum,bs,cbs; 4449 4450 if (n == PETSC_DECIDE) { 4451 ierr = PetscSplitOwnership(comm,&n,&N);CHKERRQ(ierr); 4452 } 4453 /* Check sum(n) = N */ 4454 ierr = MPIU_Allreduce(&n,&sum,1,MPIU_INT,MPI_SUM,comm);CHKERRQ(ierr); 4455 if (sum != N) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_INCOMP,"Sum of local columns %D != global columns %D",sum,N); 4456 4457 ierr = MPI_Scan(&m, &rstart,1,MPIU_INT,MPI_SUM,comm);CHKERRQ(ierr); 4458 rstart -= m; 4459 4460 ierr = MatPreallocateInitialize(comm,m,n,dnz,onz);CHKERRQ(ierr); 4461 for (i=0; i<m; i++) { 4462 ierr = MatGetRow_SeqAIJ(inmat,i,&nnz,&indx,NULL);CHKERRQ(ierr); 4463 ierr = MatPreallocateSet(i+rstart,nnz,indx,dnz,onz);CHKERRQ(ierr); 4464 ierr = MatRestoreRow_SeqAIJ(inmat,i,&nnz,&indx,NULL);CHKERRQ(ierr); 4465 } 4466 4467 ierr = MatCreate(comm,outmat);CHKERRQ(ierr); 4468 ierr = MatSetSizes(*outmat,m,n,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 4469 ierr = MatGetBlockSizes(inmat,&bs,&cbs);CHKERRQ(ierr); 4470 ierr = MatSetBlockSizes(*outmat,bs,cbs);CHKERRQ(ierr); 4471 ierr = MatSetType(*outmat,MATAIJ);CHKERRQ(ierr); 4472 ierr = MatSeqAIJSetPreallocation(*outmat,0,dnz);CHKERRQ(ierr); 4473 ierr = MatMPIAIJSetPreallocation(*outmat,0,dnz,0,onz);CHKERRQ(ierr); 4474 ierr = MatPreallocateFinalize(dnz,onz);CHKERRQ(ierr); 4475 } 4476 4477 /* numeric phase */ 4478 ierr = MatGetOwnershipRange(*outmat,&rstart,NULL);CHKERRQ(ierr); 4479 for (i=0; i<m; i++) { 4480 ierr = MatGetRow_SeqAIJ(inmat,i,&nnz,&indx,&values);CHKERRQ(ierr); 4481 Ii = i + rstart; 4482 ierr = MatSetValues(*outmat,1,&Ii,nnz,indx,values,INSERT_VALUES);CHKERRQ(ierr); 4483 ierr = MatRestoreRow_SeqAIJ(inmat,i,&nnz,&indx,&values);CHKERRQ(ierr); 4484 } 4485 ierr = MatAssemblyBegin(*outmat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4486 ierr = MatAssemblyEnd(*outmat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4487 PetscFunctionReturn(0); 4488 } 4489 4490 PetscErrorCode MatFileSplit(Mat A,char *outfile) 4491 { 4492 PetscErrorCode ierr; 4493 PetscMPIInt rank; 4494 PetscInt m,N,i,rstart,nnz; 4495 size_t len; 4496 const PetscInt *indx; 4497 PetscViewer out; 4498 char *name; 4499 Mat B; 4500 const PetscScalar *values; 4501 4502 PetscFunctionBegin; 4503 ierr = MatGetLocalSize(A,&m,0);CHKERRQ(ierr); 4504 ierr = MatGetSize(A,0,&N);CHKERRQ(ierr); 4505 /* Should this be the type of the diagonal block of A? */ 4506 ierr = MatCreate(PETSC_COMM_SELF,&B);CHKERRQ(ierr); 4507 ierr = MatSetSizes(B,m,N,m,N);CHKERRQ(ierr); 4508 ierr = MatSetBlockSizesFromMats(B,A,A);CHKERRQ(ierr); 4509 ierr = MatSetType(B,MATSEQAIJ);CHKERRQ(ierr); 4510 ierr = MatSeqAIJSetPreallocation(B,0,NULL);CHKERRQ(ierr); 4511 ierr = MatGetOwnershipRange(A,&rstart,0);CHKERRQ(ierr); 4512 for (i=0; i<m; i++) { 4513 ierr = MatGetRow(A,i+rstart,&nnz,&indx,&values);CHKERRQ(ierr); 4514 ierr = MatSetValues(B,1,&i,nnz,indx,values,INSERT_VALUES);CHKERRQ(ierr); 4515 ierr = MatRestoreRow(A,i+rstart,&nnz,&indx,&values);CHKERRQ(ierr); 4516 } 4517 ierr = MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4518 ierr = MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4519 4520 ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)A),&rank);CHKERRQ(ierr); 4521 ierr = PetscStrlen(outfile,&len);CHKERRQ(ierr); 4522 ierr = PetscMalloc1(len+5,&name);CHKERRQ(ierr); 4523 sprintf(name,"%s.%d",outfile,rank); 4524 ierr = PetscViewerBinaryOpen(PETSC_COMM_SELF,name,FILE_MODE_APPEND,&out);CHKERRQ(ierr); 4525 ierr = PetscFree(name);CHKERRQ(ierr); 4526 ierr = MatView(B,out);CHKERRQ(ierr); 4527 ierr = PetscViewerDestroy(&out);CHKERRQ(ierr); 4528 ierr = MatDestroy(&B);CHKERRQ(ierr); 4529 PetscFunctionReturn(0); 4530 } 4531 4532 PetscErrorCode MatDestroy_MPIAIJ_SeqsToMPI(Mat A) 4533 { 4534 PetscErrorCode ierr; 4535 Mat_Merge_SeqsToMPI *merge; 4536 PetscContainer container; 4537 4538 PetscFunctionBegin; 4539 ierr = PetscObjectQuery((PetscObject)A,"MatMergeSeqsToMPI",(PetscObject*)&container);CHKERRQ(ierr); 4540 if (container) { 4541 ierr = PetscContainerGetPointer(container,(void**)&merge);CHKERRQ(ierr); 4542 ierr = PetscFree(merge->id_r);CHKERRQ(ierr); 4543 ierr = PetscFree(merge->len_s);CHKERRQ(ierr); 4544 ierr = PetscFree(merge->len_r);CHKERRQ(ierr); 4545 ierr = PetscFree(merge->bi);CHKERRQ(ierr); 4546 ierr = PetscFree(merge->bj);CHKERRQ(ierr); 4547 ierr = PetscFree(merge->buf_ri[0]);CHKERRQ(ierr); 4548 ierr = PetscFree(merge->buf_ri);CHKERRQ(ierr); 4549 ierr = PetscFree(merge->buf_rj[0]);CHKERRQ(ierr); 4550 ierr = PetscFree(merge->buf_rj);CHKERRQ(ierr); 4551 ierr = PetscFree(merge->coi);CHKERRQ(ierr); 4552 ierr = PetscFree(merge->coj);CHKERRQ(ierr); 4553 ierr = PetscFree(merge->owners_co);CHKERRQ(ierr); 4554 ierr = PetscLayoutDestroy(&merge->rowmap);CHKERRQ(ierr); 4555 ierr = PetscFree(merge);CHKERRQ(ierr); 4556 ierr = PetscObjectCompose((PetscObject)A,"MatMergeSeqsToMPI",0);CHKERRQ(ierr); 4557 } 4558 ierr = MatDestroy_MPIAIJ(A);CHKERRQ(ierr); 4559 PetscFunctionReturn(0); 4560 } 4561 4562 #include <../src/mat/utils/freespace.h> 4563 #include <petscbt.h> 4564 4565 PetscErrorCode MatCreateMPIAIJSumSeqAIJNumeric(Mat seqmat,Mat mpimat) 4566 { 4567 PetscErrorCode ierr; 4568 MPI_Comm comm; 4569 Mat_SeqAIJ *a =(Mat_SeqAIJ*)seqmat->data; 4570 PetscMPIInt size,rank,taga,*len_s; 4571 PetscInt N=mpimat->cmap->N,i,j,*owners,*ai=a->i,*aj; 4572 PetscInt proc,m; 4573 PetscInt **buf_ri,**buf_rj; 4574 PetscInt k,anzi,*bj_i,*bi,*bj,arow,bnzi,nextaj; 4575 PetscInt nrows,**buf_ri_k,**nextrow,**nextai; 4576 MPI_Request *s_waits,*r_waits; 4577 MPI_Status *status; 4578 MatScalar *aa=a->a; 4579 MatScalar **abuf_r,*ba_i; 4580 Mat_Merge_SeqsToMPI *merge; 4581 PetscContainer container; 4582 4583 PetscFunctionBegin; 4584 ierr = PetscObjectGetComm((PetscObject)mpimat,&comm);CHKERRQ(ierr); 4585 ierr = PetscLogEventBegin(MAT_Seqstompinum,seqmat,0,0,0);CHKERRQ(ierr); 4586 4587 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 4588 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 4589 4590 ierr = PetscObjectQuery((PetscObject)mpimat,"MatMergeSeqsToMPI",(PetscObject*)&container);CHKERRQ(ierr); 4591 ierr = PetscContainerGetPointer(container,(void**)&merge);CHKERRQ(ierr); 4592 4593 bi = merge->bi; 4594 bj = merge->bj; 4595 buf_ri = merge->buf_ri; 4596 buf_rj = merge->buf_rj; 4597 4598 ierr = PetscMalloc1(size,&status);CHKERRQ(ierr); 4599 owners = merge->rowmap->range; 4600 len_s = merge->len_s; 4601 4602 /* send and recv matrix values */ 4603 /*-----------------------------*/ 4604 ierr = PetscObjectGetNewTag((PetscObject)mpimat,&taga);CHKERRQ(ierr); 4605 ierr = PetscPostIrecvScalar(comm,taga,merge->nrecv,merge->id_r,merge->len_r,&abuf_r,&r_waits);CHKERRQ(ierr); 4606 4607 ierr = PetscMalloc1(merge->nsend+1,&s_waits);CHKERRQ(ierr); 4608 for (proc=0,k=0; proc<size; proc++) { 4609 if (!len_s[proc]) continue; 4610 i = owners[proc]; 4611 ierr = MPI_Isend(aa+ai[i],len_s[proc],MPIU_MATSCALAR,proc,taga,comm,s_waits+k);CHKERRQ(ierr); 4612 k++; 4613 } 4614 4615 if (merge->nrecv) {ierr = MPI_Waitall(merge->nrecv,r_waits,status);CHKERRQ(ierr);} 4616 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,s_waits,status);CHKERRQ(ierr);} 4617 ierr = PetscFree(status);CHKERRQ(ierr); 4618 4619 ierr = PetscFree(s_waits);CHKERRQ(ierr); 4620 ierr = PetscFree(r_waits);CHKERRQ(ierr); 4621 4622 /* insert mat values of mpimat */ 4623 /*----------------------------*/ 4624 ierr = PetscMalloc1(N,&ba_i);CHKERRQ(ierr); 4625 ierr = PetscMalloc3(merge->nrecv,&buf_ri_k,merge->nrecv,&nextrow,merge->nrecv,&nextai);CHKERRQ(ierr); 4626 4627 for (k=0; k<merge->nrecv; k++) { 4628 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 4629 nrows = *(buf_ri_k[k]); 4630 nextrow[k] = buf_ri_k[k]+1; /* next row number of k-th recved i-structure */ 4631 nextai[k] = buf_ri_k[k] + (nrows + 1); /* poins to the next i-structure of k-th recved i-structure */ 4632 } 4633 4634 /* set values of ba */ 4635 m = merge->rowmap->n; 4636 for (i=0; i<m; i++) { 4637 arow = owners[rank] + i; 4638 bj_i = bj+bi[i]; /* col indices of the i-th row of mpimat */ 4639 bnzi = bi[i+1] - bi[i]; 4640 ierr = PetscMemzero(ba_i,bnzi*sizeof(PetscScalar));CHKERRQ(ierr); 4641 4642 /* add local non-zero vals of this proc's seqmat into ba */ 4643 anzi = ai[arow+1] - ai[arow]; 4644 aj = a->j + ai[arow]; 4645 aa = a->a + ai[arow]; 4646 nextaj = 0; 4647 for (j=0; nextaj<anzi; j++) { 4648 if (*(bj_i + j) == aj[nextaj]) { /* bcol == acol */ 4649 ba_i[j] += aa[nextaj++]; 4650 } 4651 } 4652 4653 /* add received vals into ba */ 4654 for (k=0; k<merge->nrecv; k++) { /* k-th received message */ 4655 /* i-th row */ 4656 if (i == *nextrow[k]) { 4657 anzi = *(nextai[k]+1) - *nextai[k]; 4658 aj = buf_rj[k] + *(nextai[k]); 4659 aa = abuf_r[k] + *(nextai[k]); 4660 nextaj = 0; 4661 for (j=0; nextaj<anzi; j++) { 4662 if (*(bj_i + j) == aj[nextaj]) { /* bcol == acol */ 4663 ba_i[j] += aa[nextaj++]; 4664 } 4665 } 4666 nextrow[k]++; nextai[k]++; 4667 } 4668 } 4669 ierr = MatSetValues(mpimat,1,&arow,bnzi,bj_i,ba_i,INSERT_VALUES);CHKERRQ(ierr); 4670 } 4671 ierr = MatAssemblyBegin(mpimat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4672 ierr = MatAssemblyEnd(mpimat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4673 4674 ierr = PetscFree(abuf_r[0]);CHKERRQ(ierr); 4675 ierr = PetscFree(abuf_r);CHKERRQ(ierr); 4676 ierr = PetscFree(ba_i);CHKERRQ(ierr); 4677 ierr = PetscFree3(buf_ri_k,nextrow,nextai);CHKERRQ(ierr); 4678 ierr = PetscLogEventEnd(MAT_Seqstompinum,seqmat,0,0,0);CHKERRQ(ierr); 4679 PetscFunctionReturn(0); 4680 } 4681 4682 PetscErrorCode MatCreateMPIAIJSumSeqAIJSymbolic(MPI_Comm comm,Mat seqmat,PetscInt m,PetscInt n,Mat *mpimat) 4683 { 4684 PetscErrorCode ierr; 4685 Mat B_mpi; 4686 Mat_SeqAIJ *a=(Mat_SeqAIJ*)seqmat->data; 4687 PetscMPIInt size,rank,tagi,tagj,*len_s,*len_si,*len_ri; 4688 PetscInt **buf_rj,**buf_ri,**buf_ri_k; 4689 PetscInt M=seqmat->rmap->n,N=seqmat->cmap->n,i,*owners,*ai=a->i,*aj=a->j; 4690 PetscInt len,proc,*dnz,*onz,bs,cbs; 4691 PetscInt k,anzi,*bi,*bj,*lnk,nlnk,arow,bnzi,nspacedouble=0; 4692 PetscInt nrows,*buf_s,*buf_si,*buf_si_i,**nextrow,**nextai; 4693 MPI_Request *si_waits,*sj_waits,*ri_waits,*rj_waits; 4694 MPI_Status *status; 4695 PetscFreeSpaceList free_space=NULL,current_space=NULL; 4696 PetscBT lnkbt; 4697 Mat_Merge_SeqsToMPI *merge; 4698 PetscContainer container; 4699 4700 PetscFunctionBegin; 4701 ierr = PetscLogEventBegin(MAT_Seqstompisym,seqmat,0,0,0);CHKERRQ(ierr); 4702 4703 /* make sure it is a PETSc comm */ 4704 ierr = PetscCommDuplicate(comm,&comm,NULL);CHKERRQ(ierr); 4705 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 4706 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 4707 4708 ierr = PetscNew(&merge);CHKERRQ(ierr); 4709 ierr = PetscMalloc1(size,&status);CHKERRQ(ierr); 4710 4711 /* determine row ownership */ 4712 /*---------------------------------------------------------*/ 4713 ierr = PetscLayoutCreate(comm,&merge->rowmap);CHKERRQ(ierr); 4714 ierr = PetscLayoutSetLocalSize(merge->rowmap,m);CHKERRQ(ierr); 4715 ierr = PetscLayoutSetSize(merge->rowmap,M);CHKERRQ(ierr); 4716 ierr = PetscLayoutSetBlockSize(merge->rowmap,1);CHKERRQ(ierr); 4717 ierr = PetscLayoutSetUp(merge->rowmap);CHKERRQ(ierr); 4718 ierr = PetscMalloc1(size,&len_si);CHKERRQ(ierr); 4719 ierr = PetscMalloc1(size,&merge->len_s);CHKERRQ(ierr); 4720 4721 m = merge->rowmap->n; 4722 owners = merge->rowmap->range; 4723 4724 /* determine the number of messages to send, their lengths */ 4725 /*---------------------------------------------------------*/ 4726 len_s = merge->len_s; 4727 4728 len = 0; /* length of buf_si[] */ 4729 merge->nsend = 0; 4730 for (proc=0; proc<size; proc++) { 4731 len_si[proc] = 0; 4732 if (proc == rank) { 4733 len_s[proc] = 0; 4734 } else { 4735 len_si[proc] = owners[proc+1] - owners[proc] + 1; 4736 len_s[proc] = ai[owners[proc+1]] - ai[owners[proc]]; /* num of rows to be sent to [proc] */ 4737 } 4738 if (len_s[proc]) { 4739 merge->nsend++; 4740 nrows = 0; 4741 for (i=owners[proc]; i<owners[proc+1]; i++) { 4742 if (ai[i+1] > ai[i]) nrows++; 4743 } 4744 len_si[proc] = 2*(nrows+1); 4745 len += len_si[proc]; 4746 } 4747 } 4748 4749 /* determine the number and length of messages to receive for ij-structure */ 4750 /*-------------------------------------------------------------------------*/ 4751 ierr = PetscGatherNumberOfMessages(comm,NULL,len_s,&merge->nrecv);CHKERRQ(ierr); 4752 ierr = PetscGatherMessageLengths2(comm,merge->nsend,merge->nrecv,len_s,len_si,&merge->id_r,&merge->len_r,&len_ri);CHKERRQ(ierr); 4753 4754 /* post the Irecv of j-structure */ 4755 /*-------------------------------*/ 4756 ierr = PetscCommGetNewTag(comm,&tagj);CHKERRQ(ierr); 4757 ierr = PetscPostIrecvInt(comm,tagj,merge->nrecv,merge->id_r,merge->len_r,&buf_rj,&rj_waits);CHKERRQ(ierr); 4758 4759 /* post the Isend of j-structure */ 4760 /*--------------------------------*/ 4761 ierr = PetscMalloc2(merge->nsend,&si_waits,merge->nsend,&sj_waits);CHKERRQ(ierr); 4762 4763 for (proc=0, k=0; proc<size; proc++) { 4764 if (!len_s[proc]) continue; 4765 i = owners[proc]; 4766 ierr = MPI_Isend(aj+ai[i],len_s[proc],MPIU_INT,proc,tagj,comm,sj_waits+k);CHKERRQ(ierr); 4767 k++; 4768 } 4769 4770 /* receives and sends of j-structure are complete */ 4771 /*------------------------------------------------*/ 4772 if (merge->nrecv) {ierr = MPI_Waitall(merge->nrecv,rj_waits,status);CHKERRQ(ierr);} 4773 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,sj_waits,status);CHKERRQ(ierr);} 4774 4775 /* send and recv i-structure */ 4776 /*---------------------------*/ 4777 ierr = PetscCommGetNewTag(comm,&tagi);CHKERRQ(ierr); 4778 ierr = PetscPostIrecvInt(comm,tagi,merge->nrecv,merge->id_r,len_ri,&buf_ri,&ri_waits);CHKERRQ(ierr); 4779 4780 ierr = PetscMalloc1(len+1,&buf_s);CHKERRQ(ierr); 4781 buf_si = buf_s; /* points to the beginning of k-th msg to be sent */ 4782 for (proc=0,k=0; proc<size; proc++) { 4783 if (!len_s[proc]) continue; 4784 /* form outgoing message for i-structure: 4785 buf_si[0]: nrows to be sent 4786 [1:nrows]: row index (global) 4787 [nrows+1:2*nrows+1]: i-structure index 4788 */ 4789 /*-------------------------------------------*/ 4790 nrows = len_si[proc]/2 - 1; 4791 buf_si_i = buf_si + nrows+1; 4792 buf_si[0] = nrows; 4793 buf_si_i[0] = 0; 4794 nrows = 0; 4795 for (i=owners[proc]; i<owners[proc+1]; i++) { 4796 anzi = ai[i+1] - ai[i]; 4797 if (anzi) { 4798 buf_si_i[nrows+1] = buf_si_i[nrows] + anzi; /* i-structure */ 4799 buf_si[nrows+1] = i-owners[proc]; /* local row index */ 4800 nrows++; 4801 } 4802 } 4803 ierr = MPI_Isend(buf_si,len_si[proc],MPIU_INT,proc,tagi,comm,si_waits+k);CHKERRQ(ierr); 4804 k++; 4805 buf_si += len_si[proc]; 4806 } 4807 4808 if (merge->nrecv) {ierr = MPI_Waitall(merge->nrecv,ri_waits,status);CHKERRQ(ierr);} 4809 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,si_waits,status);CHKERRQ(ierr);} 4810 4811 ierr = PetscInfo2(seqmat,"nsend: %D, nrecv: %D\n",merge->nsend,merge->nrecv);CHKERRQ(ierr); 4812 for (i=0; i<merge->nrecv; i++) { 4813 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); 4814 } 4815 4816 ierr = PetscFree(len_si);CHKERRQ(ierr); 4817 ierr = PetscFree(len_ri);CHKERRQ(ierr); 4818 ierr = PetscFree(rj_waits);CHKERRQ(ierr); 4819 ierr = PetscFree2(si_waits,sj_waits);CHKERRQ(ierr); 4820 ierr = PetscFree(ri_waits);CHKERRQ(ierr); 4821 ierr = PetscFree(buf_s);CHKERRQ(ierr); 4822 ierr = PetscFree(status);CHKERRQ(ierr); 4823 4824 /* compute a local seq matrix in each processor */ 4825 /*----------------------------------------------*/ 4826 /* allocate bi array and free space for accumulating nonzero column info */ 4827 ierr = PetscMalloc1(m+1,&bi);CHKERRQ(ierr); 4828 bi[0] = 0; 4829 4830 /* create and initialize a linked list */ 4831 nlnk = N+1; 4832 ierr = PetscLLCreate(N,N,nlnk,lnk,lnkbt);CHKERRQ(ierr); 4833 4834 /* initial FreeSpace size is 2*(num of local nnz(seqmat)) */ 4835 len = ai[owners[rank+1]] - ai[owners[rank]]; 4836 ierr = PetscFreeSpaceGet(PetscIntMultTruncate(2,len)+1,&free_space);CHKERRQ(ierr); 4837 4838 current_space = free_space; 4839 4840 /* determine symbolic info for each local row */ 4841 ierr = PetscMalloc3(merge->nrecv,&buf_ri_k,merge->nrecv,&nextrow,merge->nrecv,&nextai);CHKERRQ(ierr); 4842 4843 for (k=0; k<merge->nrecv; k++) { 4844 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 4845 nrows = *buf_ri_k[k]; 4846 nextrow[k] = buf_ri_k[k] + 1; /* next row number of k-th recved i-structure */ 4847 nextai[k] = buf_ri_k[k] + (nrows + 1); /* poins to the next i-structure of k-th recved i-structure */ 4848 } 4849 4850 ierr = MatPreallocateInitialize(comm,m,n,dnz,onz);CHKERRQ(ierr); 4851 len = 0; 4852 for (i=0; i<m; i++) { 4853 bnzi = 0; 4854 /* add local non-zero cols of this proc's seqmat into lnk */ 4855 arow = owners[rank] + i; 4856 anzi = ai[arow+1] - ai[arow]; 4857 aj = a->j + ai[arow]; 4858 ierr = PetscLLAddSorted(anzi,aj,N,nlnk,lnk,lnkbt);CHKERRQ(ierr); 4859 bnzi += nlnk; 4860 /* add received col data into lnk */ 4861 for (k=0; k<merge->nrecv; k++) { /* k-th received message */ 4862 if (i == *nextrow[k]) { /* i-th row */ 4863 anzi = *(nextai[k]+1) - *nextai[k]; 4864 aj = buf_rj[k] + *nextai[k]; 4865 ierr = PetscLLAddSorted(anzi,aj,N,nlnk,lnk,lnkbt);CHKERRQ(ierr); 4866 bnzi += nlnk; 4867 nextrow[k]++; nextai[k]++; 4868 } 4869 } 4870 if (len < bnzi) len = bnzi; /* =max(bnzi) */ 4871 4872 /* if free space is not available, make more free space */ 4873 if (current_space->local_remaining<bnzi) { 4874 ierr = PetscFreeSpaceGet(PetscIntSumTruncate(bnzi,current_space->total_array_size),¤t_space);CHKERRQ(ierr); 4875 nspacedouble++; 4876 } 4877 /* copy data into free space, then initialize lnk */ 4878 ierr = PetscLLClean(N,N,bnzi,lnk,current_space->array,lnkbt);CHKERRQ(ierr); 4879 ierr = MatPreallocateSet(i+owners[rank],bnzi,current_space->array,dnz,onz);CHKERRQ(ierr); 4880 4881 current_space->array += bnzi; 4882 current_space->local_used += bnzi; 4883 current_space->local_remaining -= bnzi; 4884 4885 bi[i+1] = bi[i] + bnzi; 4886 } 4887 4888 ierr = PetscFree3(buf_ri_k,nextrow,nextai);CHKERRQ(ierr); 4889 4890 ierr = PetscMalloc1(bi[m]+1,&bj);CHKERRQ(ierr); 4891 ierr = PetscFreeSpaceContiguous(&free_space,bj);CHKERRQ(ierr); 4892 ierr = PetscLLDestroy(lnk,lnkbt);CHKERRQ(ierr); 4893 4894 /* create symbolic parallel matrix B_mpi */ 4895 /*---------------------------------------*/ 4896 ierr = MatGetBlockSizes(seqmat,&bs,&cbs);CHKERRQ(ierr); 4897 ierr = MatCreate(comm,&B_mpi);CHKERRQ(ierr); 4898 if (n==PETSC_DECIDE) { 4899 ierr = MatSetSizes(B_mpi,m,n,PETSC_DETERMINE,N);CHKERRQ(ierr); 4900 } else { 4901 ierr = MatSetSizes(B_mpi,m,n,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 4902 } 4903 ierr = MatSetBlockSizes(B_mpi,bs,cbs);CHKERRQ(ierr); 4904 ierr = MatSetType(B_mpi,MATMPIAIJ);CHKERRQ(ierr); 4905 ierr = MatMPIAIJSetPreallocation(B_mpi,0,dnz,0,onz);CHKERRQ(ierr); 4906 ierr = MatPreallocateFinalize(dnz,onz);CHKERRQ(ierr); 4907 ierr = MatSetOption(B_mpi,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 4908 4909 /* B_mpi is not ready for use - assembly will be done by MatCreateMPIAIJSumSeqAIJNumeric() */ 4910 B_mpi->assembled = PETSC_FALSE; 4911 B_mpi->ops->destroy = MatDestroy_MPIAIJ_SeqsToMPI; 4912 merge->bi = bi; 4913 merge->bj = bj; 4914 merge->buf_ri = buf_ri; 4915 merge->buf_rj = buf_rj; 4916 merge->coi = NULL; 4917 merge->coj = NULL; 4918 merge->owners_co = NULL; 4919 4920 ierr = PetscCommDestroy(&comm);CHKERRQ(ierr); 4921 4922 /* attach the supporting struct to B_mpi for reuse */ 4923 ierr = PetscContainerCreate(PETSC_COMM_SELF,&container);CHKERRQ(ierr); 4924 ierr = PetscContainerSetPointer(container,merge);CHKERRQ(ierr); 4925 ierr = PetscObjectCompose((PetscObject)B_mpi,"MatMergeSeqsToMPI",(PetscObject)container);CHKERRQ(ierr); 4926 ierr = PetscContainerDestroy(&container);CHKERRQ(ierr); 4927 *mpimat = B_mpi; 4928 4929 ierr = PetscLogEventEnd(MAT_Seqstompisym,seqmat,0,0,0);CHKERRQ(ierr); 4930 PetscFunctionReturn(0); 4931 } 4932 4933 /*@C 4934 MatCreateMPIAIJSumSeqAIJ - Creates a MATMPIAIJ matrix by adding sequential 4935 matrices from each processor 4936 4937 Collective on MPI_Comm 4938 4939 Input Parameters: 4940 + comm - the communicators the parallel matrix will live on 4941 . seqmat - the input sequential matrices 4942 . m - number of local rows (or PETSC_DECIDE) 4943 . n - number of local columns (or PETSC_DECIDE) 4944 - scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX 4945 4946 Output Parameter: 4947 . mpimat - the parallel matrix generated 4948 4949 Level: advanced 4950 4951 Notes: 4952 The dimensions of the sequential matrix in each processor MUST be the same. 4953 The input seqmat is included into the container "Mat_Merge_SeqsToMPI", and will be 4954 destroyed when mpimat is destroyed. Call PetscObjectQuery() to access seqmat. 4955 @*/ 4956 PetscErrorCode MatCreateMPIAIJSumSeqAIJ(MPI_Comm comm,Mat seqmat,PetscInt m,PetscInt n,MatReuse scall,Mat *mpimat) 4957 { 4958 PetscErrorCode ierr; 4959 PetscMPIInt size; 4960 4961 PetscFunctionBegin; 4962 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 4963 if (size == 1) { 4964 ierr = PetscLogEventBegin(MAT_Seqstompi,seqmat,0,0,0);CHKERRQ(ierr); 4965 if (scall == MAT_INITIAL_MATRIX) { 4966 ierr = MatDuplicate(seqmat,MAT_COPY_VALUES,mpimat);CHKERRQ(ierr); 4967 } else { 4968 ierr = MatCopy(seqmat,*mpimat,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 4969 } 4970 ierr = PetscLogEventEnd(MAT_Seqstompi,seqmat,0,0,0);CHKERRQ(ierr); 4971 PetscFunctionReturn(0); 4972 } 4973 ierr = PetscLogEventBegin(MAT_Seqstompi,seqmat,0,0,0);CHKERRQ(ierr); 4974 if (scall == MAT_INITIAL_MATRIX) { 4975 ierr = MatCreateMPIAIJSumSeqAIJSymbolic(comm,seqmat,m,n,mpimat);CHKERRQ(ierr); 4976 } 4977 ierr = MatCreateMPIAIJSumSeqAIJNumeric(seqmat,*mpimat);CHKERRQ(ierr); 4978 ierr = PetscLogEventEnd(MAT_Seqstompi,seqmat,0,0,0);CHKERRQ(ierr); 4979 PetscFunctionReturn(0); 4980 } 4981 4982 /*@ 4983 MatMPIAIJGetLocalMat - Creates a SeqAIJ from a MATMPIAIJ matrix by taking all its local rows and putting them into a sequential matrix with 4984 mlocal rows and n columns. Where mlocal is the row count obtained with MatGetLocalSize() and n is the global column count obtained 4985 with MatGetSize() 4986 4987 Not Collective 4988 4989 Input Parameters: 4990 + A - the matrix 4991 . scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX 4992 4993 Output Parameter: 4994 . A_loc - the local sequential matrix generated 4995 4996 Level: developer 4997 4998 .seealso: MatGetOwnershipRange(), MatMPIAIJGetLocalMatCondensed() 4999 5000 @*/ 5001 PetscErrorCode MatMPIAIJGetLocalMat(Mat A,MatReuse scall,Mat *A_loc) 5002 { 5003 PetscErrorCode ierr; 5004 Mat_MPIAIJ *mpimat=(Mat_MPIAIJ*)A->data; 5005 Mat_SeqAIJ *mat,*a,*b; 5006 PetscInt *ai,*aj,*bi,*bj,*cmap=mpimat->garray; 5007 MatScalar *aa,*ba,*cam; 5008 PetscScalar *ca; 5009 PetscInt am=A->rmap->n,i,j,k,cstart=A->cmap->rstart; 5010 PetscInt *ci,*cj,col,ncols_d,ncols_o,jo; 5011 PetscBool match; 5012 MPI_Comm comm; 5013 PetscMPIInt size; 5014 5015 PetscFunctionBegin; 5016 ierr = PetscStrbeginswith(((PetscObject)A)->type_name,MATMPIAIJ,&match);CHKERRQ(ierr); 5017 if (!match) SETERRQ(PetscObjectComm((PetscObject)A), PETSC_ERR_SUP,"Requires MATMPIAIJ matrix as input"); 5018 ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 5019 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 5020 if (size == 1 && scall == MAT_REUSE_MATRIX) PetscFunctionReturn(0); 5021 5022 ierr = PetscLogEventBegin(MAT_Getlocalmat,A,0,0,0);CHKERRQ(ierr); 5023 a = (Mat_SeqAIJ*)(mpimat->A)->data; 5024 b = (Mat_SeqAIJ*)(mpimat->B)->data; 5025 ai = a->i; aj = a->j; bi = b->i; bj = b->j; 5026 aa = a->a; ba = b->a; 5027 if (scall == MAT_INITIAL_MATRIX) { 5028 if (size == 1) { 5029 ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,am,A->cmap->N,ai,aj,aa,A_loc);CHKERRQ(ierr); 5030 PetscFunctionReturn(0); 5031 } 5032 5033 ierr = PetscMalloc1(1+am,&ci);CHKERRQ(ierr); 5034 ci[0] = 0; 5035 for (i=0; i<am; i++) { 5036 ci[i+1] = ci[i] + (ai[i+1] - ai[i]) + (bi[i+1] - bi[i]); 5037 } 5038 ierr = PetscMalloc1(1+ci[am],&cj);CHKERRQ(ierr); 5039 ierr = PetscMalloc1(1+ci[am],&ca);CHKERRQ(ierr); 5040 k = 0; 5041 for (i=0; i<am; i++) { 5042 ncols_o = bi[i+1] - bi[i]; 5043 ncols_d = ai[i+1] - ai[i]; 5044 /* off-diagonal portion of A */ 5045 for (jo=0; jo<ncols_o; jo++) { 5046 col = cmap[*bj]; 5047 if (col >= cstart) break; 5048 cj[k] = col; bj++; 5049 ca[k++] = *ba++; 5050 } 5051 /* diagonal portion of A */ 5052 for (j=0; j<ncols_d; j++) { 5053 cj[k] = cstart + *aj++; 5054 ca[k++] = *aa++; 5055 } 5056 /* off-diagonal portion of A */ 5057 for (j=jo; j<ncols_o; j++) { 5058 cj[k] = cmap[*bj++]; 5059 ca[k++] = *ba++; 5060 } 5061 } 5062 /* put together the new matrix */ 5063 ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,am,A->cmap->N,ci,cj,ca,A_loc);CHKERRQ(ierr); 5064 /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */ 5065 /* Since these are PETSc arrays, change flags to free them as necessary. */ 5066 mat = (Mat_SeqAIJ*)(*A_loc)->data; 5067 mat->free_a = PETSC_TRUE; 5068 mat->free_ij = PETSC_TRUE; 5069 mat->nonew = 0; 5070 } else if (scall == MAT_REUSE_MATRIX) { 5071 mat=(Mat_SeqAIJ*)(*A_loc)->data; 5072 ci = mat->i; cj = mat->j; cam = mat->a; 5073 for (i=0; i<am; i++) { 5074 /* off-diagonal portion of A */ 5075 ncols_o = bi[i+1] - bi[i]; 5076 for (jo=0; jo<ncols_o; jo++) { 5077 col = cmap[*bj]; 5078 if (col >= cstart) break; 5079 *cam++ = *ba++; bj++; 5080 } 5081 /* diagonal portion of A */ 5082 ncols_d = ai[i+1] - ai[i]; 5083 for (j=0; j<ncols_d; j++) *cam++ = *aa++; 5084 /* off-diagonal portion of A */ 5085 for (j=jo; j<ncols_o; j++) { 5086 *cam++ = *ba++; bj++; 5087 } 5088 } 5089 } else SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Invalid MatReuse %d",(int)scall); 5090 ierr = PetscLogEventEnd(MAT_Getlocalmat,A,0,0,0);CHKERRQ(ierr); 5091 PetscFunctionReturn(0); 5092 } 5093 5094 /*@C 5095 MatMPIAIJGetLocalMatCondensed - Creates a SeqAIJ matrix from an MATMPIAIJ matrix by taking all its local rows and NON-ZERO columns 5096 5097 Not Collective 5098 5099 Input Parameters: 5100 + A - the matrix 5101 . scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX 5102 - row, col - index sets of rows and columns to extract (or NULL) 5103 5104 Output Parameter: 5105 . A_loc - the local sequential matrix generated 5106 5107 Level: developer 5108 5109 .seealso: MatGetOwnershipRange(), MatMPIAIJGetLocalMat() 5110 5111 @*/ 5112 PetscErrorCode MatMPIAIJGetLocalMatCondensed(Mat A,MatReuse scall,IS *row,IS *col,Mat *A_loc) 5113 { 5114 Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data; 5115 PetscErrorCode ierr; 5116 PetscInt i,start,end,ncols,nzA,nzB,*cmap,imark,*idx; 5117 IS isrowa,iscola; 5118 Mat *aloc; 5119 PetscBool match; 5120 5121 PetscFunctionBegin; 5122 ierr = PetscObjectTypeCompare((PetscObject)A,MATMPIAIJ,&match);CHKERRQ(ierr); 5123 if (!match) SETERRQ(PetscObjectComm((PetscObject)A), PETSC_ERR_SUP,"Requires MATMPIAIJ matrix as input"); 5124 ierr = PetscLogEventBegin(MAT_Getlocalmatcondensed,A,0,0,0);CHKERRQ(ierr); 5125 if (!row) { 5126 start = A->rmap->rstart; end = A->rmap->rend; 5127 ierr = ISCreateStride(PETSC_COMM_SELF,end-start,start,1,&isrowa);CHKERRQ(ierr); 5128 } else { 5129 isrowa = *row; 5130 } 5131 if (!col) { 5132 start = A->cmap->rstart; 5133 cmap = a->garray; 5134 nzA = a->A->cmap->n; 5135 nzB = a->B->cmap->n; 5136 ierr = PetscMalloc1(nzA+nzB, &idx);CHKERRQ(ierr); 5137 ncols = 0; 5138 for (i=0; i<nzB; i++) { 5139 if (cmap[i] < start) idx[ncols++] = cmap[i]; 5140 else break; 5141 } 5142 imark = i; 5143 for (i=0; i<nzA; i++) idx[ncols++] = start + i; 5144 for (i=imark; i<nzB; i++) idx[ncols++] = cmap[i]; 5145 ierr = ISCreateGeneral(PETSC_COMM_SELF,ncols,idx,PETSC_OWN_POINTER,&iscola);CHKERRQ(ierr); 5146 } else { 5147 iscola = *col; 5148 } 5149 if (scall != MAT_INITIAL_MATRIX) { 5150 ierr = PetscMalloc1(1,&aloc);CHKERRQ(ierr); 5151 aloc[0] = *A_loc; 5152 } 5153 ierr = MatCreateSubMatrices(A,1,&isrowa,&iscola,scall,&aloc);CHKERRQ(ierr); 5154 if (!col) { /* attach global id of condensed columns */ 5155 ierr = PetscObjectCompose((PetscObject)aloc[0],"_petsc_GetLocalMatCondensed_iscol",(PetscObject)iscola);CHKERRQ(ierr); 5156 } 5157 *A_loc = aloc[0]; 5158 ierr = PetscFree(aloc);CHKERRQ(ierr); 5159 if (!row) { 5160 ierr = ISDestroy(&isrowa);CHKERRQ(ierr); 5161 } 5162 if (!col) { 5163 ierr = ISDestroy(&iscola);CHKERRQ(ierr); 5164 } 5165 ierr = PetscLogEventEnd(MAT_Getlocalmatcondensed,A,0,0,0);CHKERRQ(ierr); 5166 PetscFunctionReturn(0); 5167 } 5168 5169 /*@C 5170 MatGetBrowsOfAcols - Creates a SeqAIJ matrix by taking rows of B that equal to nonzero columns of local A 5171 5172 Collective on Mat 5173 5174 Input Parameters: 5175 + A,B - the matrices in mpiaij format 5176 . scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX 5177 - rowb, colb - index sets of rows and columns of B to extract (or NULL) 5178 5179 Output Parameter: 5180 + rowb, colb - index sets of rows and columns of B to extract 5181 - B_seq - the sequential matrix generated 5182 5183 Level: developer 5184 5185 @*/ 5186 PetscErrorCode MatGetBrowsOfAcols(Mat A,Mat B,MatReuse scall,IS *rowb,IS *colb,Mat *B_seq) 5187 { 5188 Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data; 5189 PetscErrorCode ierr; 5190 PetscInt *idx,i,start,ncols,nzA,nzB,*cmap,imark; 5191 IS isrowb,iscolb; 5192 Mat *bseq=NULL; 5193 5194 PetscFunctionBegin; 5195 if (A->cmap->rstart != B->rmap->rstart || A->cmap->rend != B->rmap->rend) { 5196 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); 5197 } 5198 ierr = PetscLogEventBegin(MAT_GetBrowsOfAcols,A,B,0,0);CHKERRQ(ierr); 5199 5200 if (scall == MAT_INITIAL_MATRIX) { 5201 start = A->cmap->rstart; 5202 cmap = a->garray; 5203 nzA = a->A->cmap->n; 5204 nzB = a->B->cmap->n; 5205 ierr = PetscMalloc1(nzA+nzB, &idx);CHKERRQ(ierr); 5206 ncols = 0; 5207 for (i=0; i<nzB; i++) { /* row < local row index */ 5208 if (cmap[i] < start) idx[ncols++] = cmap[i]; 5209 else break; 5210 } 5211 imark = i; 5212 for (i=0; i<nzA; i++) idx[ncols++] = start + i; /* local rows */ 5213 for (i=imark; i<nzB; i++) idx[ncols++] = cmap[i]; /* row > local row index */ 5214 ierr = ISCreateGeneral(PETSC_COMM_SELF,ncols,idx,PETSC_OWN_POINTER,&isrowb);CHKERRQ(ierr); 5215 ierr = ISCreateStride(PETSC_COMM_SELF,B->cmap->N,0,1,&iscolb);CHKERRQ(ierr); 5216 } else { 5217 if (!rowb || !colb) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"IS rowb and colb must be provided for MAT_REUSE_MATRIX"); 5218 isrowb = *rowb; iscolb = *colb; 5219 ierr = PetscMalloc1(1,&bseq);CHKERRQ(ierr); 5220 bseq[0] = *B_seq; 5221 } 5222 ierr = MatCreateSubMatrices(B,1,&isrowb,&iscolb,scall,&bseq);CHKERRQ(ierr); 5223 *B_seq = bseq[0]; 5224 ierr = PetscFree(bseq);CHKERRQ(ierr); 5225 if (!rowb) { 5226 ierr = ISDestroy(&isrowb);CHKERRQ(ierr); 5227 } else { 5228 *rowb = isrowb; 5229 } 5230 if (!colb) { 5231 ierr = ISDestroy(&iscolb);CHKERRQ(ierr); 5232 } else { 5233 *colb = iscolb; 5234 } 5235 ierr = PetscLogEventEnd(MAT_GetBrowsOfAcols,A,B,0,0);CHKERRQ(ierr); 5236 PetscFunctionReturn(0); 5237 } 5238 5239 #include <petsc/private/vecscatterimpl.h> 5240 /* 5241 MatGetBrowsOfAoCols_MPIAIJ - Creates a SeqAIJ matrix by taking rows of B that equal to nonzero columns 5242 of the OFF-DIAGONAL portion of local A 5243 5244 Collective on Mat 5245 5246 Input Parameters: 5247 + A,B - the matrices in mpiaij format 5248 - scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX 5249 5250 Output Parameter: 5251 + startsj_s - starting point in B's sending j-arrays, saved for MAT_REUSE (or NULL) 5252 . startsj_r - starting point in B's receiving j-arrays, saved for MAT_REUSE (or NULL) 5253 . bufa_ptr - array for sending matrix values, saved for MAT_REUSE (or NULL) 5254 - B_oth - the sequential matrix generated with size aBn=a->B->cmap->n by B->cmap->N 5255 5256 Developer Notes: This directly accesses information inside the VecScatter associated with the matrix-vector product 5257 for this matrix. This is not desirable.. 5258 5259 Level: developer 5260 5261 */ 5262 PetscErrorCode MatGetBrowsOfAoCols_MPIAIJ(Mat A,Mat B,MatReuse scall,PetscInt **startsj_s,PetscInt **startsj_r,MatScalar **bufa_ptr,Mat *B_oth) 5263 { 5264 VecScatter_MPI_General *gen_to,*gen_from; 5265 PetscErrorCode ierr; 5266 Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data; 5267 Mat_SeqAIJ *b_oth; 5268 VecScatter ctx; 5269 MPI_Comm comm; 5270 PetscMPIInt *rprocs,*sprocs,tag,rank; 5271 PetscInt *rowlen,*bufj,*bufJ,ncols,aBn=a->B->cmap->n,row,*b_othi,*b_othj; 5272 PetscInt *rvalues,*svalues,*cols,sbs,rbs; 5273 PetscScalar *b_otha,*bufa,*bufA,*vals; 5274 PetscInt i,j,k,l,ll,nrecvs,nsends,nrows,*srow,*rstarts,*rstartsj = 0,*sstarts,*sstartsj,len; 5275 MPI_Request *rwaits = NULL,*swaits = NULL; 5276 MPI_Status *sstatus,rstatus; 5277 PetscMPIInt jj,size; 5278 VecScatterType type; 5279 PetscBool mpi1; 5280 5281 PetscFunctionBegin; 5282 ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 5283 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 5284 5285 if (A->cmap->rstart != B->rmap->rstart || A->cmap->rend != B->rmap->rend) { 5286 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); 5287 } 5288 ierr = PetscLogEventBegin(MAT_GetBrowsOfAocols,A,B,0,0);CHKERRQ(ierr); 5289 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 5290 5291 if (size == 1) { 5292 startsj_s = NULL; 5293 bufa_ptr = NULL; 5294 *B_oth = NULL; 5295 PetscFunctionReturn(0); 5296 } 5297 5298 ctx = a->Mvctx; 5299 ierr = VecScatterGetType(ctx,&type);CHKERRQ(ierr); 5300 ierr = PetscStrcmp(type,"mpi1",&mpi1);CHKERRQ(ierr); 5301 if (!mpi1) { 5302 /* a->Mvctx is not type MPI1 which is not implemented for Mat-Mat ops, 5303 thus create a->Mvctx_mpi1 */ 5304 if (!a->Mvctx_mpi1) { 5305 a->Mvctx_mpi1_flg = PETSC_TRUE; 5306 ierr = MatSetUpMultiply_MPIAIJ(A);CHKERRQ(ierr); 5307 } 5308 ctx = a->Mvctx_mpi1; 5309 } 5310 tag = ((PetscObject)ctx)->tag; 5311 5312 gen_to = (VecScatter_MPI_General*)ctx->todata; 5313 gen_from = (VecScatter_MPI_General*)ctx->fromdata; 5314 nrecvs = gen_from->n; 5315 nsends = gen_to->n; 5316 5317 ierr = PetscMalloc2(nrecvs,&rwaits,nsends,&swaits);CHKERRQ(ierr); 5318 srow = gen_to->indices; /* local row index to be sent */ 5319 sstarts = gen_to->starts; 5320 sprocs = gen_to->procs; 5321 sstatus = gen_to->sstatus; 5322 sbs = gen_to->bs; 5323 rstarts = gen_from->starts; 5324 rprocs = gen_from->procs; 5325 rbs = gen_from->bs; 5326 5327 if (!startsj_s || !bufa_ptr) scall = MAT_INITIAL_MATRIX; 5328 if (scall == MAT_INITIAL_MATRIX) { 5329 /* i-array */ 5330 /*---------*/ 5331 /* post receives */ 5332 ierr = PetscMalloc1(rbs*(rstarts[nrecvs] - rstarts[0]),&rvalues);CHKERRQ(ierr); 5333 for (i=0; i<nrecvs; i++) { 5334 rowlen = rvalues + rstarts[i]*rbs; 5335 nrows = (rstarts[i+1]-rstarts[i])*rbs; /* num of indices to be received */ 5336 ierr = MPI_Irecv(rowlen,nrows,MPIU_INT,rprocs[i],tag,comm,rwaits+i);CHKERRQ(ierr); 5337 } 5338 5339 /* pack the outgoing message */ 5340 ierr = PetscMalloc2(nsends+1,&sstartsj,nrecvs+1,&rstartsj);CHKERRQ(ierr); 5341 5342 sstartsj[0] = 0; 5343 rstartsj[0] = 0; 5344 len = 0; /* total length of j or a array to be sent */ 5345 k = 0; 5346 ierr = PetscMalloc1(sbs*(sstarts[nsends] - sstarts[0]),&svalues);CHKERRQ(ierr); 5347 for (i=0; i<nsends; i++) { 5348 rowlen = svalues + sstarts[i]*sbs; 5349 nrows = sstarts[i+1]-sstarts[i]; /* num of block rows */ 5350 for (j=0; j<nrows; j++) { 5351 row = srow[k] + B->rmap->range[rank]; /* global row idx */ 5352 for (l=0; l<sbs; l++) { 5353 ierr = MatGetRow_MPIAIJ(B,row+l,&ncols,NULL,NULL);CHKERRQ(ierr); /* rowlength */ 5354 5355 rowlen[j*sbs+l] = ncols; 5356 5357 len += ncols; 5358 ierr = MatRestoreRow_MPIAIJ(B,row+l,&ncols,NULL,NULL);CHKERRQ(ierr); 5359 } 5360 k++; 5361 } 5362 ierr = MPI_Isend(rowlen,nrows*sbs,MPIU_INT,sprocs[i],tag,comm,swaits+i);CHKERRQ(ierr); 5363 5364 sstartsj[i+1] = len; /* starting point of (i+1)-th outgoing msg in bufj and bufa */ 5365 } 5366 /* recvs and sends of i-array are completed */ 5367 i = nrecvs; 5368 while (i--) { 5369 ierr = MPI_Waitany(nrecvs,rwaits,&jj,&rstatus);CHKERRQ(ierr); 5370 } 5371 if (nsends) {ierr = MPI_Waitall(nsends,swaits,sstatus);CHKERRQ(ierr);} 5372 ierr = PetscFree(svalues);CHKERRQ(ierr); 5373 5374 /* allocate buffers for sending j and a arrays */ 5375 ierr = PetscMalloc1(len+1,&bufj);CHKERRQ(ierr); 5376 ierr = PetscMalloc1(len+1,&bufa);CHKERRQ(ierr); 5377 5378 /* create i-array of B_oth */ 5379 ierr = PetscMalloc1(aBn+2,&b_othi);CHKERRQ(ierr); 5380 5381 b_othi[0] = 0; 5382 len = 0; /* total length of j or a array to be received */ 5383 k = 0; 5384 for (i=0; i<nrecvs; i++) { 5385 rowlen = rvalues + rstarts[i]*rbs; 5386 nrows = rbs*(rstarts[i+1]-rstarts[i]); /* num of rows to be received */ 5387 for (j=0; j<nrows; j++) { 5388 b_othi[k+1] = b_othi[k] + rowlen[j]; 5389 ierr = PetscIntSumError(rowlen[j],len,&len);CHKERRQ(ierr); 5390 k++; 5391 } 5392 rstartsj[i+1] = len; /* starting point of (i+1)-th incoming msg in bufj and bufa */ 5393 } 5394 ierr = PetscFree(rvalues);CHKERRQ(ierr); 5395 5396 /* allocate space for j and a arrrays of B_oth */ 5397 ierr = PetscMalloc1(b_othi[aBn]+1,&b_othj);CHKERRQ(ierr); 5398 ierr = PetscMalloc1(b_othi[aBn]+1,&b_otha);CHKERRQ(ierr); 5399 5400 /* j-array */ 5401 /*---------*/ 5402 /* post receives of j-array */ 5403 for (i=0; i<nrecvs; i++) { 5404 nrows = rstartsj[i+1]-rstartsj[i]; /* length of the msg received */ 5405 ierr = MPI_Irecv(b_othj+rstartsj[i],nrows,MPIU_INT,rprocs[i],tag,comm,rwaits+i);CHKERRQ(ierr); 5406 } 5407 5408 /* pack the outgoing message j-array */ 5409 k = 0; 5410 for (i=0; i<nsends; i++) { 5411 nrows = sstarts[i+1]-sstarts[i]; /* num of block rows */ 5412 bufJ = bufj+sstartsj[i]; 5413 for (j=0; j<nrows; j++) { 5414 row = srow[k++] + B->rmap->range[rank]; /* global row idx */ 5415 for (ll=0; ll<sbs; ll++) { 5416 ierr = MatGetRow_MPIAIJ(B,row+ll,&ncols,&cols,NULL);CHKERRQ(ierr); 5417 for (l=0; l<ncols; l++) { 5418 *bufJ++ = cols[l]; 5419 } 5420 ierr = MatRestoreRow_MPIAIJ(B,row+ll,&ncols,&cols,NULL);CHKERRQ(ierr); 5421 } 5422 } 5423 ierr = MPI_Isend(bufj+sstartsj[i],sstartsj[i+1]-sstartsj[i],MPIU_INT,sprocs[i],tag,comm,swaits+i);CHKERRQ(ierr); 5424 } 5425 5426 /* recvs and sends of j-array are completed */ 5427 i = nrecvs; 5428 while (i--) { 5429 ierr = MPI_Waitany(nrecvs,rwaits,&jj,&rstatus);CHKERRQ(ierr); 5430 } 5431 if (nsends) {ierr = MPI_Waitall(nsends,swaits,sstatus);CHKERRQ(ierr);} 5432 } else if (scall == MAT_REUSE_MATRIX) { 5433 sstartsj = *startsj_s; 5434 rstartsj = *startsj_r; 5435 bufa = *bufa_ptr; 5436 b_oth = (Mat_SeqAIJ*)(*B_oth)->data; 5437 b_otha = b_oth->a; 5438 } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE, "Matrix P does not posses an object container"); 5439 5440 /* a-array */ 5441 /*---------*/ 5442 /* post receives of a-array */ 5443 for (i=0; i<nrecvs; i++) { 5444 nrows = rstartsj[i+1]-rstartsj[i]; /* length of the msg received */ 5445 ierr = MPI_Irecv(b_otha+rstartsj[i],nrows,MPIU_SCALAR,rprocs[i],tag,comm,rwaits+i);CHKERRQ(ierr); 5446 } 5447 5448 /* pack the outgoing message a-array */ 5449 k = 0; 5450 for (i=0; i<nsends; i++) { 5451 nrows = sstarts[i+1]-sstarts[i]; /* num of block rows */ 5452 bufA = bufa+sstartsj[i]; 5453 for (j=0; j<nrows; j++) { 5454 row = srow[k++] + B->rmap->range[rank]; /* global row idx */ 5455 for (ll=0; ll<sbs; ll++) { 5456 ierr = MatGetRow_MPIAIJ(B,row+ll,&ncols,NULL,&vals);CHKERRQ(ierr); 5457 for (l=0; l<ncols; l++) { 5458 *bufA++ = vals[l]; 5459 } 5460 ierr = MatRestoreRow_MPIAIJ(B,row+ll,&ncols,NULL,&vals);CHKERRQ(ierr); 5461 } 5462 } 5463 ierr = MPI_Isend(bufa+sstartsj[i],sstartsj[i+1]-sstartsj[i],MPIU_SCALAR,sprocs[i],tag,comm,swaits+i);CHKERRQ(ierr); 5464 } 5465 /* recvs and sends of a-array are completed */ 5466 i = nrecvs; 5467 while (i--) { 5468 ierr = MPI_Waitany(nrecvs,rwaits,&jj,&rstatus);CHKERRQ(ierr); 5469 } 5470 if (nsends) {ierr = MPI_Waitall(nsends,swaits,sstatus);CHKERRQ(ierr);} 5471 ierr = PetscFree2(rwaits,swaits);CHKERRQ(ierr); 5472 5473 if (scall == MAT_INITIAL_MATRIX) { 5474 /* put together the new matrix */ 5475 ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,aBn,B->cmap->N,b_othi,b_othj,b_otha,B_oth);CHKERRQ(ierr); 5476 5477 /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */ 5478 /* Since these are PETSc arrays, change flags to free them as necessary. */ 5479 b_oth = (Mat_SeqAIJ*)(*B_oth)->data; 5480 b_oth->free_a = PETSC_TRUE; 5481 b_oth->free_ij = PETSC_TRUE; 5482 b_oth->nonew = 0; 5483 5484 ierr = PetscFree(bufj);CHKERRQ(ierr); 5485 if (!startsj_s || !bufa_ptr) { 5486 ierr = PetscFree2(sstartsj,rstartsj);CHKERRQ(ierr); 5487 ierr = PetscFree(bufa_ptr);CHKERRQ(ierr); 5488 } else { 5489 *startsj_s = sstartsj; 5490 *startsj_r = rstartsj; 5491 *bufa_ptr = bufa; 5492 } 5493 } 5494 ierr = PetscLogEventEnd(MAT_GetBrowsOfAocols,A,B,0,0);CHKERRQ(ierr); 5495 PetscFunctionReturn(0); 5496 } 5497 5498 /*@C 5499 MatGetCommunicationStructs - Provides access to the communication structures used in matrix-vector multiplication. 5500 5501 Not Collective 5502 5503 Input Parameters: 5504 . A - The matrix in mpiaij format 5505 5506 Output Parameter: 5507 + lvec - The local vector holding off-process values from the argument to a matrix-vector product 5508 . colmap - A map from global column index to local index into lvec 5509 - multScatter - A scatter from the argument of a matrix-vector product to lvec 5510 5511 Level: developer 5512 5513 @*/ 5514 #if defined(PETSC_USE_CTABLE) 5515 PetscErrorCode MatGetCommunicationStructs(Mat A, Vec *lvec, PetscTable *colmap, VecScatter *multScatter) 5516 #else 5517 PetscErrorCode MatGetCommunicationStructs(Mat A, Vec *lvec, PetscInt *colmap[], VecScatter *multScatter) 5518 #endif 5519 { 5520 Mat_MPIAIJ *a; 5521 5522 PetscFunctionBegin; 5523 PetscValidHeaderSpecific(A, MAT_CLASSID, 1); 5524 PetscValidPointer(lvec, 2); 5525 PetscValidPointer(colmap, 3); 5526 PetscValidPointer(multScatter, 4); 5527 a = (Mat_MPIAIJ*) A->data; 5528 if (lvec) *lvec = a->lvec; 5529 if (colmap) *colmap = a->colmap; 5530 if (multScatter) *multScatter = a->Mvctx; 5531 PetscFunctionReturn(0); 5532 } 5533 5534 PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJCRL(Mat,MatType,MatReuse,Mat*); 5535 PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJPERM(Mat,MatType,MatReuse,Mat*); 5536 PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJSELL(Mat,MatType,MatReuse,Mat*); 5537 #if defined(PETSC_HAVE_MKL_SPARSE) 5538 PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJMKL(Mat,MatType,MatReuse,Mat*); 5539 #endif 5540 PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPISBAIJ(Mat,MatType,MatReuse,Mat*); 5541 #if defined(PETSC_HAVE_ELEMENTAL) 5542 PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_Elemental(Mat,MatType,MatReuse,Mat*); 5543 #endif 5544 #if defined(PETSC_HAVE_HYPRE) 5545 PETSC_INTERN PetscErrorCode MatConvert_AIJ_HYPRE(Mat,MatType,MatReuse,Mat*); 5546 PETSC_INTERN PetscErrorCode MatMatMatMult_Transpose_AIJ_AIJ(Mat,Mat,Mat,MatReuse,PetscReal,Mat*); 5547 #endif 5548 PETSC_INTERN PetscErrorCode MatConvert_XAIJ_IS(Mat,MatType,MatReuse,Mat*); 5549 PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPISELL(Mat,MatType,MatReuse,Mat*); 5550 PETSC_INTERN PetscErrorCode MatPtAP_IS_XAIJ(Mat,Mat,MatReuse,PetscReal,Mat*); 5551 5552 /* 5553 Computes (B'*A')' since computing B*A directly is untenable 5554 5555 n p p 5556 ( ) ( ) ( ) 5557 m ( A ) * n ( B ) = m ( C ) 5558 ( ) ( ) ( ) 5559 5560 */ 5561 PetscErrorCode MatMatMultNumeric_MPIDense_MPIAIJ(Mat A,Mat B,Mat C) 5562 { 5563 PetscErrorCode ierr; 5564 Mat At,Bt,Ct; 5565 5566 PetscFunctionBegin; 5567 ierr = MatTranspose(A,MAT_INITIAL_MATRIX,&At);CHKERRQ(ierr); 5568 ierr = MatTranspose(B,MAT_INITIAL_MATRIX,&Bt);CHKERRQ(ierr); 5569 ierr = MatMatMult(Bt,At,MAT_INITIAL_MATRIX,1.0,&Ct);CHKERRQ(ierr); 5570 ierr = MatDestroy(&At);CHKERRQ(ierr); 5571 ierr = MatDestroy(&Bt);CHKERRQ(ierr); 5572 ierr = MatTranspose(Ct,MAT_REUSE_MATRIX,&C);CHKERRQ(ierr); 5573 ierr = MatDestroy(&Ct);CHKERRQ(ierr); 5574 PetscFunctionReturn(0); 5575 } 5576 5577 PetscErrorCode MatMatMultSymbolic_MPIDense_MPIAIJ(Mat A,Mat B,PetscReal fill,Mat *C) 5578 { 5579 PetscErrorCode ierr; 5580 PetscInt m=A->rmap->n,n=B->cmap->n; 5581 Mat Cmat; 5582 5583 PetscFunctionBegin; 5584 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); 5585 ierr = MatCreate(PetscObjectComm((PetscObject)A),&Cmat);CHKERRQ(ierr); 5586 ierr = MatSetSizes(Cmat,m,n,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 5587 ierr = MatSetBlockSizesFromMats(Cmat,A,B);CHKERRQ(ierr); 5588 ierr = MatSetType(Cmat,MATMPIDENSE);CHKERRQ(ierr); 5589 ierr = MatMPIDenseSetPreallocation(Cmat,NULL);CHKERRQ(ierr); 5590 ierr = MatAssemblyBegin(Cmat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5591 ierr = MatAssemblyEnd(Cmat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5592 5593 Cmat->ops->matmultnumeric = MatMatMultNumeric_MPIDense_MPIAIJ; 5594 5595 *C = Cmat; 5596 PetscFunctionReturn(0); 5597 } 5598 5599 /* ----------------------------------------------------------------*/ 5600 PETSC_INTERN PetscErrorCode MatMatMult_MPIDense_MPIAIJ(Mat A,Mat B,MatReuse scall,PetscReal fill,Mat *C) 5601 { 5602 PetscErrorCode ierr; 5603 5604 PetscFunctionBegin; 5605 if (scall == MAT_INITIAL_MATRIX) { 5606 ierr = PetscLogEventBegin(MAT_MatMultSymbolic,A,B,0,0);CHKERRQ(ierr); 5607 ierr = MatMatMultSymbolic_MPIDense_MPIAIJ(A,B,fill,C);CHKERRQ(ierr); 5608 ierr = PetscLogEventEnd(MAT_MatMultSymbolic,A,B,0,0);CHKERRQ(ierr); 5609 } 5610 ierr = PetscLogEventBegin(MAT_MatMultNumeric,A,B,0,0);CHKERRQ(ierr); 5611 ierr = MatMatMultNumeric_MPIDense_MPIAIJ(A,B,*C);CHKERRQ(ierr); 5612 ierr = PetscLogEventEnd(MAT_MatMultNumeric,A,B,0,0);CHKERRQ(ierr); 5613 PetscFunctionReturn(0); 5614 } 5615 5616 /*MC 5617 MATMPIAIJ - MATMPIAIJ = "mpiaij" - A matrix type to be used for parallel sparse matrices. 5618 5619 Options Database Keys: 5620 . -mat_type mpiaij - sets the matrix type to "mpiaij" during a call to MatSetFromOptions() 5621 5622 Level: beginner 5623 5624 .seealso: MatCreateAIJ() 5625 M*/ 5626 5627 PETSC_EXTERN PetscErrorCode MatCreate_MPIAIJ(Mat B) 5628 { 5629 Mat_MPIAIJ *b; 5630 PetscErrorCode ierr; 5631 PetscMPIInt size; 5632 5633 PetscFunctionBegin; 5634 ierr = MPI_Comm_size(PetscObjectComm((PetscObject)B),&size);CHKERRQ(ierr); 5635 5636 ierr = PetscNewLog(B,&b);CHKERRQ(ierr); 5637 B->data = (void*)b; 5638 ierr = PetscMemcpy(B->ops,&MatOps_Values,sizeof(struct _MatOps));CHKERRQ(ierr); 5639 B->assembled = PETSC_FALSE; 5640 B->insertmode = NOT_SET_VALUES; 5641 b->size = size; 5642 5643 ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)B),&b->rank);CHKERRQ(ierr); 5644 5645 /* build cache for off array entries formed */ 5646 ierr = MatStashCreate_Private(PetscObjectComm((PetscObject)B),1,&B->stash);CHKERRQ(ierr); 5647 5648 b->donotstash = PETSC_FALSE; 5649 b->colmap = 0; 5650 b->garray = 0; 5651 b->roworiented = PETSC_TRUE; 5652 5653 /* stuff used for matrix vector multiply */ 5654 b->lvec = NULL; 5655 b->Mvctx = NULL; 5656 5657 /* stuff for MatGetRow() */ 5658 b->rowindices = 0; 5659 b->rowvalues = 0; 5660 b->getrowactive = PETSC_FALSE; 5661 5662 /* flexible pointer used in CUSP/CUSPARSE classes */ 5663 b->spptr = NULL; 5664 5665 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMPIAIJSetUseScalableIncreaseOverlap_C",MatMPIAIJSetUseScalableIncreaseOverlap_MPIAIJ);CHKERRQ(ierr); 5666 ierr = PetscObjectComposeFunction((PetscObject)B,"MatStoreValues_C",MatStoreValues_MPIAIJ);CHKERRQ(ierr); 5667 ierr = PetscObjectComposeFunction((PetscObject)B,"MatRetrieveValues_C",MatRetrieveValues_MPIAIJ);CHKERRQ(ierr); 5668 ierr = PetscObjectComposeFunction((PetscObject)B,"MatIsTranspose_C",MatIsTranspose_MPIAIJ);CHKERRQ(ierr); 5669 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMPIAIJSetPreallocation_C",MatMPIAIJSetPreallocation_MPIAIJ);CHKERRQ(ierr); 5670 ierr = PetscObjectComposeFunction((PetscObject)B,"MatResetPreallocation_C",MatResetPreallocation_MPIAIJ);CHKERRQ(ierr); 5671 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMPIAIJSetPreallocationCSR_C",MatMPIAIJSetPreallocationCSR_MPIAIJ);CHKERRQ(ierr); 5672 ierr = PetscObjectComposeFunction((PetscObject)B,"MatDiagonalScaleLocal_C",MatDiagonalScaleLocal_MPIAIJ);CHKERRQ(ierr); 5673 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_mpiaijperm_C",MatConvert_MPIAIJ_MPIAIJPERM);CHKERRQ(ierr); 5674 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_mpiaijsell_C",MatConvert_MPIAIJ_MPIAIJSELL);CHKERRQ(ierr); 5675 #if defined(PETSC_HAVE_MKL_SPARSE) 5676 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_mpiaijmkl_C",MatConvert_MPIAIJ_MPIAIJMKL);CHKERRQ(ierr); 5677 #endif 5678 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_mpiaijcrl_C",MatConvert_MPIAIJ_MPIAIJCRL);CHKERRQ(ierr); 5679 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_mpisbaij_C",MatConvert_MPIAIJ_MPISBAIJ);CHKERRQ(ierr); 5680 #if defined(PETSC_HAVE_ELEMENTAL) 5681 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_elemental_C",MatConvert_MPIAIJ_Elemental);CHKERRQ(ierr); 5682 #endif 5683 #if defined(PETSC_HAVE_HYPRE) 5684 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_hypre_C",MatConvert_AIJ_HYPRE);CHKERRQ(ierr); 5685 #endif 5686 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_is_C",MatConvert_XAIJ_IS);CHKERRQ(ierr); 5687 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_mpisell_C",MatConvert_MPIAIJ_MPISELL);CHKERRQ(ierr); 5688 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMatMult_mpidense_mpiaij_C",MatMatMult_MPIDense_MPIAIJ);CHKERRQ(ierr); 5689 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMatMultSymbolic_mpidense_mpiaij_C",MatMatMultSymbolic_MPIDense_MPIAIJ);CHKERRQ(ierr); 5690 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMatMultNumeric_mpidense_mpiaij_C",MatMatMultNumeric_MPIDense_MPIAIJ);CHKERRQ(ierr); 5691 #if defined(PETSC_HAVE_HYPRE) 5692 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMatMatMult_transpose_mpiaij_mpiaij_C",MatMatMatMult_Transpose_AIJ_AIJ);CHKERRQ(ierr); 5693 #endif 5694 ierr = PetscObjectComposeFunction((PetscObject)B,"MatPtAP_is_mpiaij_C",MatPtAP_IS_XAIJ);CHKERRQ(ierr); 5695 ierr = PetscObjectChangeTypeName((PetscObject)B,MATMPIAIJ);CHKERRQ(ierr); 5696 PetscFunctionReturn(0); 5697 } 5698 5699 /*@C 5700 MatCreateMPIAIJWithSplitArrays - creates a MPI AIJ matrix using arrays that contain the "diagonal" 5701 and "off-diagonal" part of the matrix in CSR format. 5702 5703 Collective on MPI_Comm 5704 5705 Input Parameters: 5706 + comm - MPI communicator 5707 . m - number of local rows (Cannot be PETSC_DECIDE) 5708 . n - This value should be the same as the local size used in creating the 5709 x vector for the matrix-vector product y = Ax. (or PETSC_DECIDE to have 5710 calculated if N is given) For square matrices n is almost always m. 5711 . M - number of global rows (or PETSC_DETERMINE to have calculated if m is given) 5712 . N - number of global columns (or PETSC_DETERMINE to have calculated if n is given) 5713 . i - row indices for "diagonal" portion of matrix; that is i[0] = 0, i[row] = i[row-1] + number of elements in that row of the matrix 5714 . j - column indices 5715 . a - matrix values 5716 . oi - row indices for "off-diagonal" portion of matrix; that is oi[0] = 0, oi[row] = oi[row-1] + number of elements in that row of the matrix 5717 . oj - column indices 5718 - oa - matrix values 5719 5720 Output Parameter: 5721 . mat - the matrix 5722 5723 Level: advanced 5724 5725 Notes: 5726 The i, j, and a arrays ARE NOT copied by this routine into the internal format used by PETSc. The user 5727 must free the arrays once the matrix has been destroyed and not before. 5728 5729 The i and j indices are 0 based 5730 5731 See MatCreateAIJ() for the definition of "diagonal" and "off-diagonal" portion of the matrix 5732 5733 This sets local rows and cannot be used to set off-processor values. 5734 5735 Use of this routine is discouraged because it is inflexible and cumbersome to use. It is extremely rare that a 5736 legacy application natively assembles into exactly this split format. The code to do so is nontrivial and does 5737 not easily support in-place reassembly. It is recommended to use MatSetValues() (or a variant thereof) because 5738 the resulting assembly is easier to implement, will work with any matrix format, and the user does not have to 5739 keep track of the underlying array. Use MatSetOption(A,MAT_NO_OFF_PROC_ENTRIES,PETSC_TRUE) to disable all 5740 communication if it is known that only local entries will be set. 5741 5742 .keywords: matrix, aij, compressed row, sparse, parallel 5743 5744 .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatMPIAIJSetPreallocationCSR(), 5745 MATMPIAIJ, MatCreateAIJ(), MatCreateMPIAIJWithArrays() 5746 @*/ 5747 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) 5748 { 5749 PetscErrorCode ierr; 5750 Mat_MPIAIJ *maij; 5751 5752 PetscFunctionBegin; 5753 if (m < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"local number of rows (m) cannot be PETSC_DECIDE, or negative"); 5754 if (i[0]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"i (row indices) must start with 0"); 5755 if (oi[0]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"oi (row indices) must start with 0"); 5756 ierr = MatCreate(comm,mat);CHKERRQ(ierr); 5757 ierr = MatSetSizes(*mat,m,n,M,N);CHKERRQ(ierr); 5758 ierr = MatSetType(*mat,MATMPIAIJ);CHKERRQ(ierr); 5759 maij = (Mat_MPIAIJ*) (*mat)->data; 5760 5761 (*mat)->preallocated = PETSC_TRUE; 5762 5763 ierr = PetscLayoutSetUp((*mat)->rmap);CHKERRQ(ierr); 5764 ierr = PetscLayoutSetUp((*mat)->cmap);CHKERRQ(ierr); 5765 5766 ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,m,n,i,j,a,&maij->A);CHKERRQ(ierr); 5767 ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,m,(*mat)->cmap->N,oi,oj,oa,&maij->B);CHKERRQ(ierr); 5768 5769 ierr = MatAssemblyBegin(maij->A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5770 ierr = MatAssemblyEnd(maij->A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5771 ierr = MatAssemblyBegin(maij->B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5772 ierr = MatAssemblyEnd(maij->B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5773 5774 ierr = MatSetOption(*mat,MAT_NO_OFF_PROC_ENTRIES,PETSC_TRUE);CHKERRQ(ierr); 5775 ierr = MatAssemblyBegin(*mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5776 ierr = MatAssemblyEnd(*mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5777 ierr = MatSetOption(*mat,MAT_NO_OFF_PROC_ENTRIES,PETSC_FALSE);CHKERRQ(ierr); 5778 ierr = MatSetOption(*mat,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE);CHKERRQ(ierr); 5779 PetscFunctionReturn(0); 5780 } 5781 5782 /* 5783 Special version for direct calls from Fortran 5784 */ 5785 #include <petsc/private/fortranimpl.h> 5786 5787 /* Change these macros so can be used in void function */ 5788 #undef CHKERRQ 5789 #define CHKERRQ(ierr) CHKERRABORT(PETSC_COMM_WORLD,ierr) 5790 #undef SETERRQ2 5791 #define SETERRQ2(comm,ierr,b,c,d) CHKERRABORT(comm,ierr) 5792 #undef SETERRQ3 5793 #define SETERRQ3(comm,ierr,b,c,d,e) CHKERRABORT(comm,ierr) 5794 #undef SETERRQ 5795 #define SETERRQ(c,ierr,b) CHKERRABORT(c,ierr) 5796 5797 #if defined(PETSC_HAVE_FORTRAN_CAPS) 5798 #define matsetvaluesmpiaij_ MATSETVALUESMPIAIJ 5799 #elif !defined(PETSC_HAVE_FORTRAN_UNDERSCORE) 5800 #define matsetvaluesmpiaij_ matsetvaluesmpiaij 5801 #else 5802 #endif 5803 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) 5804 { 5805 Mat mat = *mmat; 5806 PetscInt m = *mm, n = *mn; 5807 InsertMode addv = *maddv; 5808 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 5809 PetscScalar value; 5810 PetscErrorCode ierr; 5811 5812 MatCheckPreallocated(mat,1); 5813 if (mat->insertmode == NOT_SET_VALUES) mat->insertmode = addv; 5814 5815 #if defined(PETSC_USE_DEBUG) 5816 else if (mat->insertmode != addv) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Cannot mix add values and insert values"); 5817 #endif 5818 { 5819 PetscInt i,j,rstart = mat->rmap->rstart,rend = mat->rmap->rend; 5820 PetscInt cstart = mat->cmap->rstart,cend = mat->cmap->rend,row,col; 5821 PetscBool roworiented = aij->roworiented; 5822 5823 /* Some Variables required in the macro */ 5824 Mat A = aij->A; 5825 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 5826 PetscInt *aimax = a->imax,*ai = a->i,*ailen = a->ilen,*aj = a->j; 5827 MatScalar *aa = a->a; 5828 PetscBool ignorezeroentries = (((a->ignorezeroentries)&&(addv==ADD_VALUES)) ? PETSC_TRUE : PETSC_FALSE); 5829 Mat B = aij->B; 5830 Mat_SeqAIJ *b = (Mat_SeqAIJ*)B->data; 5831 PetscInt *bimax = b->imax,*bi = b->i,*bilen = b->ilen,*bj = b->j,bm = aij->B->rmap->n,am = aij->A->rmap->n; 5832 MatScalar *ba = b->a; 5833 5834 PetscInt *rp1,*rp2,ii,nrow1,nrow2,_i,rmax1,rmax2,N,low1,high1,low2,high2,t,lastcol1,lastcol2; 5835 PetscInt nonew = a->nonew; 5836 MatScalar *ap1,*ap2; 5837 5838 PetscFunctionBegin; 5839 for (i=0; i<m; i++) { 5840 if (im[i] < 0) continue; 5841 #if defined(PETSC_USE_DEBUG) 5842 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); 5843 #endif 5844 if (im[i] >= rstart && im[i] < rend) { 5845 row = im[i] - rstart; 5846 lastcol1 = -1; 5847 rp1 = aj + ai[row]; 5848 ap1 = aa + ai[row]; 5849 rmax1 = aimax[row]; 5850 nrow1 = ailen[row]; 5851 low1 = 0; 5852 high1 = nrow1; 5853 lastcol2 = -1; 5854 rp2 = bj + bi[row]; 5855 ap2 = ba + bi[row]; 5856 rmax2 = bimax[row]; 5857 nrow2 = bilen[row]; 5858 low2 = 0; 5859 high2 = nrow2; 5860 5861 for (j=0; j<n; j++) { 5862 if (roworiented) value = v[i*n+j]; 5863 else value = v[i+j*m]; 5864 if (in[j] >= cstart && in[j] < cend) { 5865 col = in[j] - cstart; 5866 if (ignorezeroentries && value == 0.0 && (addv == ADD_VALUES) && row != col) continue; 5867 MatSetValues_SeqAIJ_A_Private(row,col,value,addv,im[i],in[j]); 5868 } else if (in[j] < 0) continue; 5869 #if defined(PETSC_USE_DEBUG) 5870 /* extra brace on SETERRQ2() is required for --with-errorchecking=0 - due to the next 'else' clause */ 5871 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);} 5872 #endif 5873 else { 5874 if (mat->was_assembled) { 5875 if (!aij->colmap) { 5876 ierr = MatCreateColmap_MPIAIJ_Private(mat);CHKERRQ(ierr); 5877 } 5878 #if defined(PETSC_USE_CTABLE) 5879 ierr = PetscTableFind(aij->colmap,in[j]+1,&col);CHKERRQ(ierr); 5880 col--; 5881 #else 5882 col = aij->colmap[in[j]] - 1; 5883 #endif 5884 if (ignorezeroentries && value == 0.0 && (addv == ADD_VALUES) && row != col) continue; 5885 if (col < 0 && !((Mat_SeqAIJ*)(aij->A->data))->nonew) { 5886 ierr = MatDisAssemble_MPIAIJ(mat);CHKERRQ(ierr); 5887 col = in[j]; 5888 /* Reinitialize the variables required by MatSetValues_SeqAIJ_B_Private() */ 5889 B = aij->B; 5890 b = (Mat_SeqAIJ*)B->data; 5891 bimax = b->imax; bi = b->i; bilen = b->ilen; bj = b->j; 5892 rp2 = bj + bi[row]; 5893 ap2 = ba + bi[row]; 5894 rmax2 = bimax[row]; 5895 nrow2 = bilen[row]; 5896 low2 = 0; 5897 high2 = nrow2; 5898 bm = aij->B->rmap->n; 5899 ba = b->a; 5900 } 5901 } else col = in[j]; 5902 MatSetValues_SeqAIJ_B_Private(row,col,value,addv,im[i],in[j]); 5903 } 5904 } 5905 } else if (!aij->donotstash) { 5906 if (roworiented) { 5907 ierr = MatStashValuesRow_Private(&mat->stash,im[i],n,in,v+i*n,(PetscBool)(ignorezeroentries && (addv == ADD_VALUES)));CHKERRQ(ierr); 5908 } else { 5909 ierr = MatStashValuesCol_Private(&mat->stash,im[i],n,in,v+i,m,(PetscBool)(ignorezeroentries && (addv == ADD_VALUES)));CHKERRQ(ierr); 5910 } 5911 } 5912 } 5913 } 5914 PetscFunctionReturnVoid(); 5915 } 5916