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