1 2 3 #include <../src/mat/impls/aij/mpi/mpiaij.h> /*I "petscmat.h" I*/ 4 #include <petsc/private/vecimpl.h> 5 #include <petsc/private/isimpl.h> 6 #include <petscblaslapack.h> 7 #include <petscsf.h> 8 9 /*MC 10 MATAIJ - MATAIJ = "aij" - A matrix type to be used for sparse matrices. 11 12 This matrix type is identical to MATSEQAIJ when constructed with a single process communicator, 13 and MATMPIAIJ otherwise. As a result, for single process communicators, 14 MatSeqAIJSetPreallocation is supported, and similarly MatMPIAIJSetPreallocation is supported 15 for communicators controlling multiple processes. It is recommended that you call both of 16 the above preallocation routines for simplicity. 17 18 Options Database Keys: 19 . -mat_type aij - sets the matrix type to "aij" during a call to MatSetFromOptions() 20 21 Developer Notes: Subclasses include MATAIJCUSP, MATAIJCUSPARSE, MATAIJPERM, MATAIJCRL, and also automatically switches over to use inodes when 22 enough exist. 23 24 Level: beginner 25 26 .seealso: MatCreateAIJ(), MatCreateSeqAIJ(), MATSEQAIJ, MATMPIAIJ 27 M*/ 28 29 /*MC 30 MATAIJCRL - MATAIJCRL = "aijcrl" - A matrix type to be used for sparse matrices. 31 32 This matrix type is identical to MATSEQAIJCRL when constructed with a single process communicator, 33 and MATMPIAIJCRL otherwise. As a result, for single process communicators, 34 MatSeqAIJSetPreallocation() is supported, and similarly MatMPIAIJSetPreallocation() is supported 35 for communicators controlling multiple processes. It is recommended that you call both of 36 the above preallocation routines for simplicity. 37 38 Options Database Keys: 39 . -mat_type aijcrl - sets the matrix type to "aijcrl" during a call to MatSetFromOptions() 40 41 Level: beginner 42 43 .seealso: MatCreateMPIAIJCRL,MATSEQAIJCRL,MATMPIAIJCRL, MATSEQAIJCRL, MATMPIAIJCRL 44 M*/ 45 46 PetscErrorCode MatSetBlockSizes_MPIAIJ(Mat M, PetscInt rbs, PetscInt cbs) 47 { 48 PetscErrorCode ierr; 49 Mat_MPIAIJ *mat = (Mat_MPIAIJ*)M->data; 50 51 PetscFunctionBegin; 52 if (mat->A) { 53 ierr = MatSetBlockSizes(mat->A,rbs,cbs);CHKERRQ(ierr); 54 ierr = MatSetBlockSizes(mat->B,rbs,1);CHKERRQ(ierr); 55 } 56 PetscFunctionReturn(0); 57 } 58 59 PetscErrorCode MatFindNonzeroRows_MPIAIJ(Mat M,IS *keptrows) 60 { 61 PetscErrorCode ierr; 62 Mat_MPIAIJ *mat = (Mat_MPIAIJ*)M->data; 63 Mat_SeqAIJ *a = (Mat_SeqAIJ*)mat->A->data; 64 Mat_SeqAIJ *b = (Mat_SeqAIJ*)mat->B->data; 65 const PetscInt *ia,*ib; 66 const MatScalar *aa,*bb; 67 PetscInt na,nb,i,j,*rows,cnt=0,n0rows; 68 PetscInt m = M->rmap->n,rstart = M->rmap->rstart; 69 70 PetscFunctionBegin; 71 *keptrows = 0; 72 ia = a->i; 73 ib = b->i; 74 for (i=0; i<m; i++) { 75 na = ia[i+1] - ia[i]; 76 nb = ib[i+1] - ib[i]; 77 if (!na && !nb) { 78 cnt++; 79 goto ok1; 80 } 81 aa = a->a + ia[i]; 82 for (j=0; j<na; j++) { 83 if (aa[j] != 0.0) goto ok1; 84 } 85 bb = b->a + ib[i]; 86 for (j=0; j <nb; j++) { 87 if (bb[j] != 0.0) goto ok1; 88 } 89 cnt++; 90 ok1:; 91 } 92 ierr = MPIU_Allreduce(&cnt,&n0rows,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)M));CHKERRQ(ierr); 93 if (!n0rows) PetscFunctionReturn(0); 94 ierr = PetscMalloc1(M->rmap->n-cnt,&rows);CHKERRQ(ierr); 95 cnt = 0; 96 for (i=0; i<m; i++) { 97 na = ia[i+1] - ia[i]; 98 nb = ib[i+1] - ib[i]; 99 if (!na && !nb) continue; 100 aa = a->a + ia[i]; 101 for (j=0; j<na;j++) { 102 if (aa[j] != 0.0) { 103 rows[cnt++] = rstart + i; 104 goto ok2; 105 } 106 } 107 bb = b->a + ib[i]; 108 for (j=0; j<nb; j++) { 109 if (bb[j] != 0.0) { 110 rows[cnt++] = rstart + i; 111 goto ok2; 112 } 113 } 114 ok2:; 115 } 116 ierr = ISCreateGeneral(PetscObjectComm((PetscObject)M),cnt,rows,PETSC_OWN_POINTER,keptrows);CHKERRQ(ierr); 117 PetscFunctionReturn(0); 118 } 119 120 PetscErrorCode MatDiagonalSet_MPIAIJ(Mat Y,Vec D,InsertMode is) 121 { 122 PetscErrorCode ierr; 123 Mat_MPIAIJ *aij = (Mat_MPIAIJ*) Y->data; 124 125 PetscFunctionBegin; 126 if (Y->assembled && Y->rmap->rstart == Y->cmap->rstart && Y->rmap->rend == Y->cmap->rend) { 127 ierr = MatDiagonalSet(aij->A,D,is);CHKERRQ(ierr); 128 } else { 129 ierr = MatDiagonalSet_Default(Y,D,is);CHKERRQ(ierr); 130 } 131 PetscFunctionReturn(0); 132 } 133 134 PetscErrorCode MatFindZeroDiagonals_MPIAIJ(Mat M,IS *zrows) 135 { 136 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)M->data; 137 PetscErrorCode ierr; 138 PetscInt i,rstart,nrows,*rows; 139 140 PetscFunctionBegin; 141 *zrows = NULL; 142 ierr = MatFindZeroDiagonals_SeqAIJ_Private(aij->A,&nrows,&rows);CHKERRQ(ierr); 143 ierr = MatGetOwnershipRange(M,&rstart,NULL);CHKERRQ(ierr); 144 for (i=0; i<nrows; i++) rows[i] += rstart; 145 ierr = ISCreateGeneral(PetscObjectComm((PetscObject)M),nrows,rows,PETSC_OWN_POINTER,zrows);CHKERRQ(ierr); 146 PetscFunctionReturn(0); 147 } 148 149 PetscErrorCode MatGetColumnNorms_MPIAIJ(Mat A,NormType type,PetscReal *norms) 150 { 151 PetscErrorCode ierr; 152 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)A->data; 153 PetscInt i,n,*garray = aij->garray; 154 Mat_SeqAIJ *a_aij = (Mat_SeqAIJ*) aij->A->data; 155 Mat_SeqAIJ *b_aij = (Mat_SeqAIJ*) aij->B->data; 156 PetscReal *work; 157 158 PetscFunctionBegin; 159 ierr = MatGetSize(A,NULL,&n);CHKERRQ(ierr); 160 ierr = PetscCalloc1(n,&work);CHKERRQ(ierr); 161 if (type == NORM_2) { 162 for (i=0; i<a_aij->i[aij->A->rmap->n]; i++) { 163 work[A->cmap->rstart + a_aij->j[i]] += PetscAbsScalar(a_aij->a[i]*a_aij->a[i]); 164 } 165 for (i=0; i<b_aij->i[aij->B->rmap->n]; i++) { 166 work[garray[b_aij->j[i]]] += PetscAbsScalar(b_aij->a[i]*b_aij->a[i]); 167 } 168 } else if (type == NORM_1) { 169 for (i=0; i<a_aij->i[aij->A->rmap->n]; i++) { 170 work[A->cmap->rstart + a_aij->j[i]] += PetscAbsScalar(a_aij->a[i]); 171 } 172 for (i=0; i<b_aij->i[aij->B->rmap->n]; i++) { 173 work[garray[b_aij->j[i]]] += PetscAbsScalar(b_aij->a[i]); 174 } 175 } else if (type == NORM_INFINITY) { 176 for (i=0; i<a_aij->i[aij->A->rmap->n]; i++) { 177 work[A->cmap->rstart + a_aij->j[i]] = PetscMax(PetscAbsScalar(a_aij->a[i]), work[A->cmap->rstart + a_aij->j[i]]); 178 } 179 for (i=0; i<b_aij->i[aij->B->rmap->n]; i++) { 180 work[garray[b_aij->j[i]]] = PetscMax(PetscAbsScalar(b_aij->a[i]),work[garray[b_aij->j[i]]]); 181 } 182 183 } else SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_WRONG,"Unknown NormType"); 184 if (type == NORM_INFINITY) { 185 ierr = MPIU_Allreduce(work,norms,n,MPIU_REAL,MPIU_MAX,PetscObjectComm((PetscObject)A));CHKERRQ(ierr); 186 } else { 187 ierr = MPIU_Allreduce(work,norms,n,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)A));CHKERRQ(ierr); 188 } 189 ierr = PetscFree(work);CHKERRQ(ierr); 190 if (type == NORM_2) { 191 for (i=0; i<n; i++) norms[i] = PetscSqrtReal(norms[i]); 192 } 193 PetscFunctionReturn(0); 194 } 195 196 PetscErrorCode MatFindOffBlockDiagonalEntries_MPIAIJ(Mat A,IS *is) 197 { 198 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 199 IS sis,gis; 200 PetscErrorCode ierr; 201 const PetscInt *isis,*igis; 202 PetscInt n,*iis,nsis,ngis,rstart,i; 203 204 PetscFunctionBegin; 205 ierr = MatFindOffBlockDiagonalEntries(a->A,&sis);CHKERRQ(ierr); 206 ierr = MatFindNonzeroRows(a->B,&gis);CHKERRQ(ierr); 207 ierr = ISGetSize(gis,&ngis);CHKERRQ(ierr); 208 ierr = ISGetSize(sis,&nsis);CHKERRQ(ierr); 209 ierr = ISGetIndices(sis,&isis);CHKERRQ(ierr); 210 ierr = ISGetIndices(gis,&igis);CHKERRQ(ierr); 211 212 ierr = PetscMalloc1(ngis+nsis,&iis);CHKERRQ(ierr); 213 ierr = PetscMemcpy(iis,igis,ngis*sizeof(PetscInt));CHKERRQ(ierr); 214 ierr = PetscMemcpy(iis+ngis,isis,nsis*sizeof(PetscInt));CHKERRQ(ierr); 215 n = ngis + nsis; 216 ierr = PetscSortRemoveDupsInt(&n,iis);CHKERRQ(ierr); 217 ierr = MatGetOwnershipRange(A,&rstart,NULL);CHKERRQ(ierr); 218 for (i=0; i<n; i++) iis[i] += rstart; 219 ierr = ISCreateGeneral(PetscObjectComm((PetscObject)A),n,iis,PETSC_OWN_POINTER,is);CHKERRQ(ierr); 220 221 ierr = ISRestoreIndices(sis,&isis);CHKERRQ(ierr); 222 ierr = ISRestoreIndices(gis,&igis);CHKERRQ(ierr); 223 ierr = ISDestroy(&sis);CHKERRQ(ierr); 224 ierr = ISDestroy(&gis);CHKERRQ(ierr); 225 PetscFunctionReturn(0); 226 } 227 228 /* 229 Distributes a SeqAIJ matrix across a set of processes. Code stolen from 230 MatLoad_MPIAIJ(). Horrible lack of reuse. Should be a routine for each matrix type. 231 232 Only for square matrices 233 234 Used by a preconditioner, hence PETSC_EXTERN 235 */ 236 PETSC_EXTERN PetscErrorCode MatDistribute_MPIAIJ(MPI_Comm comm,Mat gmat,PetscInt m,MatReuse reuse,Mat *inmat) 237 { 238 PetscMPIInt rank,size; 239 PetscInt *rowners,*dlens,*olens,i,rstart,rend,j,jj,nz = 0,*gmataj,cnt,row,*ld,bses[2]; 240 PetscErrorCode ierr; 241 Mat mat; 242 Mat_SeqAIJ *gmata; 243 PetscMPIInt tag; 244 MPI_Status status; 245 PetscBool aij; 246 MatScalar *gmataa,*ao,*ad,*gmataarestore=0; 247 248 PetscFunctionBegin; 249 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 250 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 251 if (!rank) { 252 ierr = PetscObjectTypeCompare((PetscObject)gmat,MATSEQAIJ,&aij);CHKERRQ(ierr); 253 if (!aij) SETERRQ1(PetscObjectComm((PetscObject)gmat),PETSC_ERR_SUP,"Currently no support for input matrix of type %s\n",((PetscObject)gmat)->type_name); 254 } 255 if (reuse == MAT_INITIAL_MATRIX) { 256 ierr = MatCreate(comm,&mat);CHKERRQ(ierr); 257 ierr = MatSetSizes(mat,m,m,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 258 ierr = MatGetBlockSizes(gmat,&bses[0],&bses[1]);CHKERRQ(ierr); 259 ierr = MPI_Bcast(bses,2,MPIU_INT,0,comm);CHKERRQ(ierr); 260 ierr = MatSetBlockSizes(mat,bses[0],bses[1]);CHKERRQ(ierr); 261 ierr = MatSetType(mat,MATAIJ);CHKERRQ(ierr); 262 ierr = PetscMalloc1(size+1,&rowners);CHKERRQ(ierr); 263 ierr = PetscMalloc2(m,&dlens,m,&olens);CHKERRQ(ierr); 264 ierr = MPI_Allgather(&m,1,MPIU_INT,rowners+1,1,MPIU_INT,comm);CHKERRQ(ierr); 265 266 rowners[0] = 0; 267 for (i=2; i<=size; i++) rowners[i] += rowners[i-1]; 268 rstart = rowners[rank]; 269 rend = rowners[rank+1]; 270 ierr = PetscObjectGetNewTag((PetscObject)mat,&tag);CHKERRQ(ierr); 271 if (!rank) { 272 gmata = (Mat_SeqAIJ*) gmat->data; 273 /* send row lengths to all processors */ 274 for (i=0; i<m; i++) dlens[i] = gmata->ilen[i]; 275 for (i=1; i<size; i++) { 276 ierr = MPI_Send(gmata->ilen + rowners[i],rowners[i+1]-rowners[i],MPIU_INT,i,tag,comm);CHKERRQ(ierr); 277 } 278 /* determine number diagonal and off-diagonal counts */ 279 ierr = PetscMemzero(olens,m*sizeof(PetscInt));CHKERRQ(ierr); 280 ierr = PetscCalloc1(m,&ld);CHKERRQ(ierr); 281 jj = 0; 282 for (i=0; i<m; i++) { 283 for (j=0; j<dlens[i]; j++) { 284 if (gmata->j[jj] < rstart) ld[i]++; 285 if (gmata->j[jj] < rstart || gmata->j[jj] >= rend) olens[i]++; 286 jj++; 287 } 288 } 289 /* send column indices to other processes */ 290 for (i=1; i<size; i++) { 291 nz = gmata->i[rowners[i+1]]-gmata->i[rowners[i]]; 292 ierr = MPI_Send(&nz,1,MPIU_INT,i,tag,comm);CHKERRQ(ierr); 293 ierr = MPI_Send(gmata->j + gmata->i[rowners[i]],nz,MPIU_INT,i,tag,comm);CHKERRQ(ierr); 294 } 295 296 /* send numerical values to other processes */ 297 for (i=1; i<size; i++) { 298 nz = gmata->i[rowners[i+1]]-gmata->i[rowners[i]]; 299 ierr = MPI_Send(gmata->a + gmata->i[rowners[i]],nz,MPIU_SCALAR,i,tag,comm);CHKERRQ(ierr); 300 } 301 gmataa = gmata->a; 302 gmataj = gmata->j; 303 304 } else { 305 /* receive row lengths */ 306 ierr = MPI_Recv(dlens,m,MPIU_INT,0,tag,comm,&status);CHKERRQ(ierr); 307 /* receive column indices */ 308 ierr = MPI_Recv(&nz,1,MPIU_INT,0,tag,comm,&status);CHKERRQ(ierr); 309 ierr = PetscMalloc2(nz,&gmataa,nz,&gmataj);CHKERRQ(ierr); 310 ierr = MPI_Recv(gmataj,nz,MPIU_INT,0,tag,comm,&status);CHKERRQ(ierr); 311 /* determine number diagonal and off-diagonal counts */ 312 ierr = PetscMemzero(olens,m*sizeof(PetscInt));CHKERRQ(ierr); 313 ierr = PetscCalloc1(m,&ld);CHKERRQ(ierr); 314 jj = 0; 315 for (i=0; i<m; i++) { 316 for (j=0; j<dlens[i]; j++) { 317 if (gmataj[jj] < rstart) ld[i]++; 318 if (gmataj[jj] < rstart || gmataj[jj] >= rend) olens[i]++; 319 jj++; 320 } 321 } 322 /* receive numerical values */ 323 ierr = PetscMemzero(gmataa,nz*sizeof(PetscScalar));CHKERRQ(ierr); 324 ierr = MPI_Recv(gmataa,nz,MPIU_SCALAR,0,tag,comm,&status);CHKERRQ(ierr); 325 } 326 /* set preallocation */ 327 for (i=0; i<m; i++) { 328 dlens[i] -= olens[i]; 329 } 330 ierr = MatSeqAIJSetPreallocation(mat,0,dlens);CHKERRQ(ierr); 331 ierr = MatMPIAIJSetPreallocation(mat,0,dlens,0,olens);CHKERRQ(ierr); 332 333 for (i=0; i<m; i++) { 334 dlens[i] += olens[i]; 335 } 336 cnt = 0; 337 for (i=0; i<m; i++) { 338 row = rstart + i; 339 ierr = MatSetValues(mat,1,&row,dlens[i],gmataj+cnt,gmataa+cnt,INSERT_VALUES);CHKERRQ(ierr); 340 cnt += dlens[i]; 341 } 342 if (rank) { 343 ierr = PetscFree2(gmataa,gmataj);CHKERRQ(ierr); 344 } 345 ierr = PetscFree2(dlens,olens);CHKERRQ(ierr); 346 ierr = PetscFree(rowners);CHKERRQ(ierr); 347 348 ((Mat_MPIAIJ*)(mat->data))->ld = ld; 349 350 *inmat = mat; 351 } else { /* column indices are already set; only need to move over numerical values from process 0 */ 352 Mat_SeqAIJ *Ad = (Mat_SeqAIJ*)((Mat_MPIAIJ*)((*inmat)->data))->A->data; 353 Mat_SeqAIJ *Ao = (Mat_SeqAIJ*)((Mat_MPIAIJ*)((*inmat)->data))->B->data; 354 mat = *inmat; 355 ierr = PetscObjectGetNewTag((PetscObject)mat,&tag);CHKERRQ(ierr); 356 if (!rank) { 357 /* send numerical values to other processes */ 358 gmata = (Mat_SeqAIJ*) gmat->data; 359 ierr = MatGetOwnershipRanges(mat,(const PetscInt**)&rowners);CHKERRQ(ierr); 360 gmataa = gmata->a; 361 for (i=1; i<size; i++) { 362 nz = gmata->i[rowners[i+1]]-gmata->i[rowners[i]]; 363 ierr = MPI_Send(gmataa + gmata->i[rowners[i]],nz,MPIU_SCALAR,i,tag,comm);CHKERRQ(ierr); 364 } 365 nz = gmata->i[rowners[1]]-gmata->i[rowners[0]]; 366 } else { 367 /* receive numerical values from process 0*/ 368 nz = Ad->nz + Ao->nz; 369 ierr = PetscMalloc1(nz,&gmataa);CHKERRQ(ierr); gmataarestore = gmataa; 370 ierr = MPI_Recv(gmataa,nz,MPIU_SCALAR,0,tag,comm,&status);CHKERRQ(ierr); 371 } 372 /* transfer numerical values into the diagonal A and off diagonal B parts of mat */ 373 ld = ((Mat_MPIAIJ*)(mat->data))->ld; 374 ad = Ad->a; 375 ao = Ao->a; 376 if (mat->rmap->n) { 377 i = 0; 378 nz = ld[i]; ierr = PetscMemcpy(ao,gmataa,nz*sizeof(PetscScalar));CHKERRQ(ierr); ao += nz; gmataa += nz; 379 nz = Ad->i[i+1] - Ad->i[i]; ierr = PetscMemcpy(ad,gmataa,nz*sizeof(PetscScalar));CHKERRQ(ierr); ad += nz; gmataa += nz; 380 } 381 for (i=1; i<mat->rmap->n; i++) { 382 nz = Ao->i[i] - Ao->i[i-1] - ld[i-1] + ld[i]; ierr = PetscMemcpy(ao,gmataa,nz*sizeof(PetscScalar));CHKERRQ(ierr); ao += nz; gmataa += nz; 383 nz = Ad->i[i+1] - Ad->i[i]; ierr = PetscMemcpy(ad,gmataa,nz*sizeof(PetscScalar));CHKERRQ(ierr); ad += nz; gmataa += nz; 384 } 385 i--; 386 if (mat->rmap->n) { 387 nz = Ao->i[i+1] - Ao->i[i] - ld[i]; ierr = PetscMemcpy(ao,gmataa,nz*sizeof(PetscScalar));CHKERRQ(ierr); 388 } 389 if (rank) { 390 ierr = PetscFree(gmataarestore);CHKERRQ(ierr); 391 } 392 } 393 ierr = MatAssemblyBegin(mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 394 ierr = MatAssemblyEnd(mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 395 PetscFunctionReturn(0); 396 } 397 398 /* 399 Local utility routine that creates a mapping from the global column 400 number to the local number in the off-diagonal part of the local 401 storage of the matrix. When PETSC_USE_CTABLE is used this is scalable at 402 a slightly higher hash table cost; without it it is not scalable (each processor 403 has an order N integer array but is fast to acess. 404 */ 405 PetscErrorCode MatCreateColmap_MPIAIJ_Private(Mat mat) 406 { 407 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 408 PetscErrorCode ierr; 409 PetscInt n = aij->B->cmap->n,i; 410 411 PetscFunctionBegin; 412 if (!aij->garray) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"MPIAIJ Matrix was assembled but is missing garray"); 413 #if defined(PETSC_USE_CTABLE) 414 ierr = PetscTableCreate(n,mat->cmap->N+1,&aij->colmap);CHKERRQ(ierr); 415 for (i=0; i<n; i++) { 416 ierr = PetscTableAdd(aij->colmap,aij->garray[i]+1,i+1,INSERT_VALUES);CHKERRQ(ierr); 417 } 418 #else 419 ierr = PetscCalloc1(mat->cmap->N+1,&aij->colmap);CHKERRQ(ierr); 420 ierr = PetscLogObjectMemory((PetscObject)mat,(mat->cmap->N+1)*sizeof(PetscInt));CHKERRQ(ierr); 421 for (i=0; i<n; i++) aij->colmap[aij->garray[i]] = i+1; 422 #endif 423 PetscFunctionReturn(0); 424 } 425 426 #define MatSetValues_SeqAIJ_A_Private(row,col,value,addv,orow,ocol) \ 427 { \ 428 if (col <= lastcol1) low1 = 0; \ 429 else high1 = nrow1; \ 430 lastcol1 = col;\ 431 while (high1-low1 > 5) { \ 432 t = (low1+high1)/2; \ 433 if (rp1[t] > col) high1 = t; \ 434 else low1 = t; \ 435 } \ 436 for (_i=low1; _i<high1; _i++) { \ 437 if (rp1[_i] > col) break; \ 438 if (rp1[_i] == col) { \ 439 if (addv == ADD_VALUES) ap1[_i] += value; \ 440 else ap1[_i] = value; \ 441 goto a_noinsert; \ 442 } \ 443 } \ 444 if (value == 0.0 && ignorezeroentries && row != col) {low1 = 0; high1 = nrow1;goto a_noinsert;} \ 445 if (nonew == 1) {low1 = 0; high1 = nrow1; goto a_noinsert;} \ 446 if (nonew == -1) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero at global row/column (%D, %D) into matrix", orow, ocol); \ 447 MatSeqXAIJReallocateAIJ(A,am,1,nrow1,row,col,rmax1,aa,ai,aj,rp1,ap1,aimax,nonew,MatScalar); \ 448 N = nrow1++ - 1; a->nz++; high1++; \ 449 /* shift up all the later entries in this row */ \ 450 for (ii=N; ii>=_i; ii--) { \ 451 rp1[ii+1] = rp1[ii]; \ 452 ap1[ii+1] = ap1[ii]; \ 453 } \ 454 rp1[_i] = col; \ 455 ap1[_i] = value; \ 456 A->nonzerostate++;\ 457 a_noinsert: ; \ 458 ailen[row] = nrow1; \ 459 } 460 461 #define MatSetValues_SeqAIJ_B_Private(row,col,value,addv,orow,ocol) \ 462 { \ 463 if (col <= lastcol2) low2 = 0; \ 464 else high2 = nrow2; \ 465 lastcol2 = col; \ 466 while (high2-low2 > 5) { \ 467 t = (low2+high2)/2; \ 468 if (rp2[t] > col) high2 = t; \ 469 else low2 = t; \ 470 } \ 471 for (_i=low2; _i<high2; _i++) { \ 472 if (rp2[_i] > col) break; \ 473 if (rp2[_i] == col) { \ 474 if (addv == ADD_VALUES) ap2[_i] += value; \ 475 else ap2[_i] = value; \ 476 goto b_noinsert; \ 477 } \ 478 } \ 479 if (value == 0.0 && ignorezeroentries) {low2 = 0; high2 = nrow2; goto b_noinsert;} \ 480 if (nonew == 1) {low2 = 0; high2 = nrow2; goto b_noinsert;} \ 481 if (nonew == -1) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero at global row/column (%D, %D) into matrix", orow, ocol); \ 482 MatSeqXAIJReallocateAIJ(B,bm,1,nrow2,row,col,rmax2,ba,bi,bj,rp2,ap2,bimax,nonew,MatScalar); \ 483 N = nrow2++ - 1; b->nz++; high2++; \ 484 /* shift up all the later entries in this row */ \ 485 for (ii=N; ii>=_i; ii--) { \ 486 rp2[ii+1] = rp2[ii]; \ 487 ap2[ii+1] = ap2[ii]; \ 488 } \ 489 rp2[_i] = col; \ 490 ap2[_i] = value; \ 491 B->nonzerostate++; \ 492 b_noinsert: ; \ 493 bilen[row] = nrow2; \ 494 } 495 496 PetscErrorCode MatSetValuesRow_MPIAIJ(Mat A,PetscInt row,const PetscScalar v[]) 497 { 498 Mat_MPIAIJ *mat = (Mat_MPIAIJ*)A->data; 499 Mat_SeqAIJ *a = (Mat_SeqAIJ*)mat->A->data,*b = (Mat_SeqAIJ*)mat->B->data; 500 PetscErrorCode ierr; 501 PetscInt l,*garray = mat->garray,diag; 502 503 PetscFunctionBegin; 504 /* code only works for square matrices A */ 505 506 /* find size of row to the left of the diagonal part */ 507 ierr = MatGetOwnershipRange(A,&diag,0);CHKERRQ(ierr); 508 row = row - diag; 509 for (l=0; l<b->i[row+1]-b->i[row]; l++) { 510 if (garray[b->j[b->i[row]+l]] > diag) break; 511 } 512 ierr = PetscMemcpy(b->a+b->i[row],v,l*sizeof(PetscScalar));CHKERRQ(ierr); 513 514 /* diagonal part */ 515 ierr = PetscMemcpy(a->a+a->i[row],v+l,(a->i[row+1]-a->i[row])*sizeof(PetscScalar));CHKERRQ(ierr); 516 517 /* right of diagonal part */ 518 ierr = PetscMemcpy(b->a+b->i[row]+l,v+l+a->i[row+1]-a->i[row],(b->i[row+1]-b->i[row]-l)*sizeof(PetscScalar));CHKERRQ(ierr); 519 PetscFunctionReturn(0); 520 } 521 522 PetscErrorCode MatSetValues_MPIAIJ(Mat mat,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],const PetscScalar v[],InsertMode addv) 523 { 524 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 525 PetscScalar value; 526 PetscErrorCode ierr; 527 PetscInt i,j,rstart = mat->rmap->rstart,rend = mat->rmap->rend; 528 PetscInt cstart = mat->cmap->rstart,cend = mat->cmap->rend,row,col; 529 PetscBool roworiented = aij->roworiented; 530 531 /* Some Variables required in the macro */ 532 Mat A = aij->A; 533 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 534 PetscInt *aimax = a->imax,*ai = a->i,*ailen = a->ilen,*aj = a->j; 535 MatScalar *aa = a->a; 536 PetscBool ignorezeroentries = a->ignorezeroentries; 537 Mat B = aij->B; 538 Mat_SeqAIJ *b = (Mat_SeqAIJ*)B->data; 539 PetscInt *bimax = b->imax,*bi = b->i,*bilen = b->ilen,*bj = b->j,bm = aij->B->rmap->n,am = aij->A->rmap->n; 540 MatScalar *ba = b->a; 541 542 PetscInt *rp1,*rp2,ii,nrow1,nrow2,_i,rmax1,rmax2,N,low1,high1,low2,high2,t,lastcol1,lastcol2; 543 PetscInt nonew; 544 MatScalar *ap1,*ap2; 545 546 PetscFunctionBegin; 547 for (i=0; i<m; i++) { 548 if (im[i] < 0) continue; 549 #if defined(PETSC_USE_DEBUG) 550 if (im[i] >= mat->rmap->N) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Row too large: row %D max %D",im[i],mat->rmap->N-1); 551 #endif 552 if (im[i] >= rstart && im[i] < rend) { 553 row = im[i] - rstart; 554 lastcol1 = -1; 555 rp1 = aj + ai[row]; 556 ap1 = aa + ai[row]; 557 rmax1 = aimax[row]; 558 nrow1 = ailen[row]; 559 low1 = 0; 560 high1 = nrow1; 561 lastcol2 = -1; 562 rp2 = bj + bi[row]; 563 ap2 = ba + bi[row]; 564 rmax2 = bimax[row]; 565 nrow2 = bilen[row]; 566 low2 = 0; 567 high2 = nrow2; 568 569 for (j=0; j<n; j++) { 570 if (roworiented) value = v[i*n+j]; 571 else value = v[i+j*m]; 572 if (in[j] >= cstart && in[j] < cend) { 573 col = in[j] - cstart; 574 nonew = a->nonew; 575 if (ignorezeroentries && value == 0.0 && (addv == ADD_VALUES) && row != col) continue; 576 MatSetValues_SeqAIJ_A_Private(row,col,value,addv,im[i],in[j]); 577 } else if (in[j] < 0) continue; 578 #if defined(PETSC_USE_DEBUG) 579 else if (in[j] >= mat->cmap->N) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Column too large: col %D max %D",in[j],mat->cmap->N-1); 580 #endif 581 else { 582 if (mat->was_assembled) { 583 if (!aij->colmap) { 584 ierr = MatCreateColmap_MPIAIJ_Private(mat);CHKERRQ(ierr); 585 } 586 #if defined(PETSC_USE_CTABLE) 587 ierr = PetscTableFind(aij->colmap,in[j]+1,&col);CHKERRQ(ierr); 588 col--; 589 #else 590 col = aij->colmap[in[j]] - 1; 591 #endif 592 if (col < 0 && !((Mat_SeqAIJ*)(aij->B->data))->nonew) { 593 ierr = MatDisAssemble_MPIAIJ(mat);CHKERRQ(ierr); 594 col = in[j]; 595 /* Reinitialize the variables required by MatSetValues_SeqAIJ_B_Private() */ 596 B = aij->B; 597 b = (Mat_SeqAIJ*)B->data; 598 bimax = b->imax; bi = b->i; bilen = b->ilen; bj = b->j; ba = b->a; 599 rp2 = bj + bi[row]; 600 ap2 = ba + bi[row]; 601 rmax2 = bimax[row]; 602 nrow2 = bilen[row]; 603 low2 = 0; 604 high2 = nrow2; 605 bm = aij->B->rmap->n; 606 ba = b->a; 607 } else if (col < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero at global row/column (%D, %D) into matrix", im[i], in[j]); 608 } else col = in[j]; 609 nonew = b->nonew; 610 MatSetValues_SeqAIJ_B_Private(row,col,value,addv,im[i],in[j]); 611 } 612 } 613 } else { 614 if (mat->nooffprocentries) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Setting off process row %D even though MatSetOption(,MAT_NO_OFF_PROC_ENTRIES,PETSC_TRUE) was set",im[i]); 615 if (!aij->donotstash) { 616 mat->assembled = PETSC_FALSE; 617 if (roworiented) { 618 ierr = MatStashValuesRow_Private(&mat->stash,im[i],n,in,v+i*n,(PetscBool)(ignorezeroentries && (addv == ADD_VALUES)));CHKERRQ(ierr); 619 } else { 620 ierr = MatStashValuesCol_Private(&mat->stash,im[i],n,in,v+i,m,(PetscBool)(ignorezeroentries && (addv == ADD_VALUES)));CHKERRQ(ierr); 621 } 622 } 623 } 624 } 625 PetscFunctionReturn(0); 626 } 627 628 PetscErrorCode MatGetValues_MPIAIJ(Mat mat,PetscInt m,const PetscInt idxm[],PetscInt n,const PetscInt idxn[],PetscScalar v[]) 629 { 630 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 631 PetscErrorCode ierr; 632 PetscInt i,j,rstart = mat->rmap->rstart,rend = mat->rmap->rend; 633 PetscInt cstart = mat->cmap->rstart,cend = mat->cmap->rend,row,col; 634 635 PetscFunctionBegin; 636 for (i=0; i<m; i++) { 637 if (idxm[i] < 0) continue; /* SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative row: %D",idxm[i]);*/ 638 if (idxm[i] >= mat->rmap->N) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Row too large: row %D max %D",idxm[i],mat->rmap->N-1); 639 if (idxm[i] >= rstart && idxm[i] < rend) { 640 row = idxm[i] - rstart; 641 for (j=0; j<n; j++) { 642 if (idxn[j] < 0) continue; /* SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative column: %D",idxn[j]); */ 643 if (idxn[j] >= mat->cmap->N) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Column too large: col %D max %D",idxn[j],mat->cmap->N-1); 644 if (idxn[j] >= cstart && idxn[j] < cend) { 645 col = idxn[j] - cstart; 646 ierr = MatGetValues(aij->A,1,&row,1,&col,v+i*n+j);CHKERRQ(ierr); 647 } else { 648 if (!aij->colmap) { 649 ierr = MatCreateColmap_MPIAIJ_Private(mat);CHKERRQ(ierr); 650 } 651 #if defined(PETSC_USE_CTABLE) 652 ierr = PetscTableFind(aij->colmap,idxn[j]+1,&col);CHKERRQ(ierr); 653 col--; 654 #else 655 col = aij->colmap[idxn[j]] - 1; 656 #endif 657 if ((col < 0) || (aij->garray[col] != idxn[j])) *(v+i*n+j) = 0.0; 658 else { 659 ierr = MatGetValues(aij->B,1,&row,1,&col,v+i*n+j);CHKERRQ(ierr); 660 } 661 } 662 } 663 } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Only local values currently supported"); 664 } 665 PetscFunctionReturn(0); 666 } 667 668 extern PetscErrorCode MatMultDiagonalBlock_MPIAIJ(Mat,Vec,Vec); 669 670 PetscErrorCode MatAssemblyBegin_MPIAIJ(Mat mat,MatAssemblyType mode) 671 { 672 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 673 PetscErrorCode ierr; 674 PetscInt nstash,reallocs; 675 676 PetscFunctionBegin; 677 if (aij->donotstash || mat->nooffprocentries) PetscFunctionReturn(0); 678 679 ierr = MatStashScatterBegin_Private(mat,&mat->stash,mat->rmap->range);CHKERRQ(ierr); 680 ierr = MatStashGetInfo_Private(&mat->stash,&nstash,&reallocs);CHKERRQ(ierr); 681 ierr = PetscInfo2(aij->A,"Stash has %D entries, uses %D mallocs.\n",nstash,reallocs);CHKERRQ(ierr); 682 PetscFunctionReturn(0); 683 } 684 685 PetscErrorCode MatAssemblyEnd_MPIAIJ(Mat mat,MatAssemblyType mode) 686 { 687 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 688 Mat_SeqAIJ *a = (Mat_SeqAIJ*)aij->A->data; 689 PetscErrorCode ierr; 690 PetscMPIInt n; 691 PetscInt i,j,rstart,ncols,flg; 692 PetscInt *row,*col; 693 PetscBool other_disassembled; 694 PetscScalar *val; 695 696 /* do not use 'b = (Mat_SeqAIJ*)aij->B->data' as B can be reset in disassembly */ 697 698 PetscFunctionBegin; 699 if (!aij->donotstash && !mat->nooffprocentries) { 700 while (1) { 701 ierr = MatStashScatterGetMesg_Private(&mat->stash,&n,&row,&col,&val,&flg);CHKERRQ(ierr); 702 if (!flg) break; 703 704 for (i=0; i<n; ) { 705 /* Now identify the consecutive vals belonging to the same row */ 706 for (j=i,rstart=row[j]; j<n; j++) { 707 if (row[j] != rstart) break; 708 } 709 if (j < n) ncols = j-i; 710 else ncols = n-i; 711 /* Now assemble all these values with a single function call */ 712 ierr = MatSetValues_MPIAIJ(mat,1,row+i,ncols,col+i,val+i,mat->insertmode);CHKERRQ(ierr); 713 714 i = j; 715 } 716 } 717 ierr = MatStashScatterEnd_Private(&mat->stash);CHKERRQ(ierr); 718 } 719 ierr = MatAssemblyBegin(aij->A,mode);CHKERRQ(ierr); 720 ierr = MatAssemblyEnd(aij->A,mode);CHKERRQ(ierr); 721 722 /* determine if any processor has disassembled, if so we must 723 also disassemble ourselfs, in order that we may reassemble. */ 724 /* 725 if nonzero structure of submatrix B cannot change then we know that 726 no processor disassembled thus we can skip this stuff 727 */ 728 if (!((Mat_SeqAIJ*)aij->B->data)->nonew) { 729 ierr = MPIU_Allreduce(&mat->was_assembled,&other_disassembled,1,MPIU_BOOL,MPI_PROD,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 730 if (mat->was_assembled && !other_disassembled) { 731 ierr = MatDisAssemble_MPIAIJ(mat);CHKERRQ(ierr); 732 } 733 } 734 if (!mat->was_assembled && mode == MAT_FINAL_ASSEMBLY) { 735 ierr = MatSetUpMultiply_MPIAIJ(mat);CHKERRQ(ierr); 736 } 737 ierr = MatSetOption(aij->B,MAT_USE_INODES,PETSC_FALSE);CHKERRQ(ierr); 738 ierr = MatAssemblyBegin(aij->B,mode);CHKERRQ(ierr); 739 ierr = MatAssemblyEnd(aij->B,mode);CHKERRQ(ierr); 740 741 ierr = PetscFree2(aij->rowvalues,aij->rowindices);CHKERRQ(ierr); 742 743 aij->rowvalues = 0; 744 745 ierr = VecDestroy(&aij->diag);CHKERRQ(ierr); 746 if (a->inode.size) mat->ops->multdiagonalblock = MatMultDiagonalBlock_MPIAIJ; 747 748 /* if no new nonzero locations are allowed in matrix then only set the matrix state the first time through */ 749 if ((!mat->was_assembled && mode == MAT_FINAL_ASSEMBLY) || !((Mat_SeqAIJ*)(aij->A->data))->nonew) { 750 PetscObjectState state = aij->A->nonzerostate + aij->B->nonzerostate; 751 ierr = MPIU_Allreduce(&state,&mat->nonzerostate,1,MPIU_INT64,MPI_SUM,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 752 } 753 PetscFunctionReturn(0); 754 } 755 756 PetscErrorCode MatZeroEntries_MPIAIJ(Mat A) 757 { 758 Mat_MPIAIJ *l = (Mat_MPIAIJ*)A->data; 759 PetscErrorCode ierr; 760 761 PetscFunctionBegin; 762 ierr = MatZeroEntries(l->A);CHKERRQ(ierr); 763 ierr = MatZeroEntries(l->B);CHKERRQ(ierr); 764 PetscFunctionReturn(0); 765 } 766 767 PetscErrorCode MatZeroRows_MPIAIJ(Mat A,PetscInt N,const PetscInt rows[],PetscScalar diag,Vec x,Vec b) 768 { 769 Mat_MPIAIJ *mat = (Mat_MPIAIJ *) A->data; 770 PetscInt *lrows; 771 PetscInt r, len; 772 PetscErrorCode ierr; 773 774 PetscFunctionBegin; 775 /* get locally owned rows */ 776 ierr = MatZeroRowsMapLocal_Private(A,N,rows,&len,&lrows);CHKERRQ(ierr); 777 /* fix right hand side if needed */ 778 if (x && b) { 779 const PetscScalar *xx; 780 PetscScalar *bb; 781 782 ierr = VecGetArrayRead(x, &xx);CHKERRQ(ierr); 783 ierr = VecGetArray(b, &bb);CHKERRQ(ierr); 784 for (r = 0; r < len; ++r) bb[lrows[r]] = diag*xx[lrows[r]]; 785 ierr = VecRestoreArrayRead(x, &xx);CHKERRQ(ierr); 786 ierr = VecRestoreArray(b, &bb);CHKERRQ(ierr); 787 } 788 /* Must zero l->B before l->A because the (diag) case below may put values into l->B*/ 789 ierr = MatZeroRows(mat->B, len, lrows, 0.0, NULL, NULL);CHKERRQ(ierr); 790 if (A->congruentlayouts == -1) { /* first time we compare rows and cols layouts */ 791 PetscBool cong; 792 ierr = PetscLayoutCompare(A->rmap,A->cmap,&cong);CHKERRQ(ierr); 793 if (cong) A->congruentlayouts = 1; 794 else A->congruentlayouts = 0; 795 } 796 if ((diag != 0.0) && A->congruentlayouts) { 797 ierr = MatZeroRows(mat->A, len, lrows, diag, NULL, NULL);CHKERRQ(ierr); 798 } else if (diag != 0.0) { 799 ierr = MatZeroRows(mat->A, len, lrows, 0.0, NULL, NULL);CHKERRQ(ierr); 800 if (((Mat_SeqAIJ *) mat->A->data)->nonew) SETERRQ(PETSC_COMM_SELF, PETSC_ERR_SUP, "MatZeroRows() on rectangular matrices cannot be used with the Mat options\nMAT_NEW_NONZERO_LOCATIONS,MAT_NEW_NONZERO_LOCATION_ERR,MAT_NEW_NONZERO_ALLOCATION_ERR"); 801 for (r = 0; r < len; ++r) { 802 const PetscInt row = lrows[r] + A->rmap->rstart; 803 ierr = MatSetValues(A, 1, &row, 1, &row, &diag, INSERT_VALUES);CHKERRQ(ierr); 804 } 805 ierr = MatAssemblyBegin(A, MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 806 ierr = MatAssemblyEnd(A, MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 807 } else { 808 ierr = MatZeroRows(mat->A, len, lrows, 0.0, NULL, NULL);CHKERRQ(ierr); 809 } 810 ierr = PetscFree(lrows);CHKERRQ(ierr); 811 812 /* only change matrix nonzero state if pattern was allowed to be changed */ 813 if (!((Mat_SeqAIJ*)(mat->A->data))->keepnonzeropattern) { 814 PetscObjectState state = mat->A->nonzerostate + mat->B->nonzerostate; 815 ierr = MPIU_Allreduce(&state,&A->nonzerostate,1,MPIU_INT64,MPI_SUM,PetscObjectComm((PetscObject)A));CHKERRQ(ierr); 816 } 817 PetscFunctionReturn(0); 818 } 819 820 PetscErrorCode MatZeroRowsColumns_MPIAIJ(Mat A,PetscInt N,const PetscInt rows[],PetscScalar diag,Vec x,Vec b) 821 { 822 Mat_MPIAIJ *l = (Mat_MPIAIJ*)A->data; 823 PetscErrorCode ierr; 824 PetscMPIInt n = A->rmap->n; 825 PetscInt i,j,r,m,p = 0,len = 0; 826 PetscInt *lrows,*owners = A->rmap->range; 827 PetscSFNode *rrows; 828 PetscSF sf; 829 const PetscScalar *xx; 830 PetscScalar *bb,*mask; 831 Vec xmask,lmask; 832 Mat_SeqAIJ *aij = (Mat_SeqAIJ*)l->B->data; 833 const PetscInt *aj, *ii,*ridx; 834 PetscScalar *aa; 835 836 PetscFunctionBegin; 837 /* Create SF where leaves are input rows and roots are owned rows */ 838 ierr = PetscMalloc1(n, &lrows);CHKERRQ(ierr); 839 for (r = 0; r < n; ++r) lrows[r] = -1; 840 ierr = PetscMalloc1(N, &rrows);CHKERRQ(ierr); 841 for (r = 0; r < N; ++r) { 842 const PetscInt idx = rows[r]; 843 if (idx < 0 || A->rmap->N <= idx) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Row %D out of range [0,%D)",idx,A->rmap->N); 844 if (idx < owners[p] || owners[p+1] <= idx) { /* short-circuit the search if the last p owns this row too */ 845 ierr = PetscLayoutFindOwner(A->rmap,idx,&p);CHKERRQ(ierr); 846 } 847 rrows[r].rank = p; 848 rrows[r].index = rows[r] - owners[p]; 849 } 850 ierr = PetscSFCreate(PetscObjectComm((PetscObject) A), &sf);CHKERRQ(ierr); 851 ierr = PetscSFSetGraph(sf, n, N, NULL, PETSC_OWN_POINTER, rrows, PETSC_OWN_POINTER);CHKERRQ(ierr); 852 /* Collect flags for rows to be zeroed */ 853 ierr = PetscSFReduceBegin(sf, MPIU_INT, (PetscInt *) rows, lrows, MPI_LOR);CHKERRQ(ierr); 854 ierr = PetscSFReduceEnd(sf, MPIU_INT, (PetscInt *) rows, lrows, MPI_LOR);CHKERRQ(ierr); 855 ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 856 /* Compress and put in row numbers */ 857 for (r = 0; r < n; ++r) if (lrows[r] >= 0) lrows[len++] = r; 858 /* zero diagonal part of matrix */ 859 ierr = MatZeroRowsColumns(l->A,len,lrows,diag,x,b);CHKERRQ(ierr); 860 /* handle off diagonal part of matrix */ 861 ierr = MatCreateVecs(A,&xmask,NULL);CHKERRQ(ierr); 862 ierr = VecDuplicate(l->lvec,&lmask);CHKERRQ(ierr); 863 ierr = VecGetArray(xmask,&bb);CHKERRQ(ierr); 864 for (i=0; i<len; i++) bb[lrows[i]] = 1; 865 ierr = VecRestoreArray(xmask,&bb);CHKERRQ(ierr); 866 ierr = VecScatterBegin(l->Mvctx,xmask,lmask,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 867 ierr = VecScatterEnd(l->Mvctx,xmask,lmask,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 868 ierr = VecDestroy(&xmask);CHKERRQ(ierr); 869 if (x) { 870 ierr = VecScatterBegin(l->Mvctx,x,l->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 871 ierr = VecScatterEnd(l->Mvctx,x,l->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 872 ierr = VecGetArrayRead(l->lvec,&xx);CHKERRQ(ierr); 873 ierr = VecGetArray(b,&bb);CHKERRQ(ierr); 874 } 875 ierr = VecGetArray(lmask,&mask);CHKERRQ(ierr); 876 /* remove zeroed rows of off diagonal matrix */ 877 ii = aij->i; 878 for (i=0; i<len; i++) { 879 ierr = PetscMemzero(aij->a + ii[lrows[i]],(ii[lrows[i]+1] - ii[lrows[i]])*sizeof(PetscScalar));CHKERRQ(ierr); 880 } 881 /* loop over all elements of off process part of matrix zeroing removed columns*/ 882 if (aij->compressedrow.use) { 883 m = aij->compressedrow.nrows; 884 ii = aij->compressedrow.i; 885 ridx = aij->compressedrow.rindex; 886 for (i=0; i<m; i++) { 887 n = ii[i+1] - ii[i]; 888 aj = aij->j + ii[i]; 889 aa = aij->a + ii[i]; 890 891 for (j=0; j<n; j++) { 892 if (PetscAbsScalar(mask[*aj])) { 893 if (b) bb[*ridx] -= *aa*xx[*aj]; 894 *aa = 0.0; 895 } 896 aa++; 897 aj++; 898 } 899 ridx++; 900 } 901 } else { /* do not use compressed row format */ 902 m = l->B->rmap->n; 903 for (i=0; i<m; i++) { 904 n = ii[i+1] - ii[i]; 905 aj = aij->j + ii[i]; 906 aa = aij->a + ii[i]; 907 for (j=0; j<n; j++) { 908 if (PetscAbsScalar(mask[*aj])) { 909 if (b) bb[i] -= *aa*xx[*aj]; 910 *aa = 0.0; 911 } 912 aa++; 913 aj++; 914 } 915 } 916 } 917 if (x) { 918 ierr = VecRestoreArray(b,&bb);CHKERRQ(ierr); 919 ierr = VecRestoreArrayRead(l->lvec,&xx);CHKERRQ(ierr); 920 } 921 ierr = VecRestoreArray(lmask,&mask);CHKERRQ(ierr); 922 ierr = VecDestroy(&lmask);CHKERRQ(ierr); 923 ierr = PetscFree(lrows);CHKERRQ(ierr); 924 925 /* only change matrix nonzero state if pattern was allowed to be changed */ 926 if (!((Mat_SeqAIJ*)(l->A->data))->keepnonzeropattern) { 927 PetscObjectState state = l->A->nonzerostate + l->B->nonzerostate; 928 ierr = MPIU_Allreduce(&state,&A->nonzerostate,1,MPIU_INT64,MPI_SUM,PetscObjectComm((PetscObject)A));CHKERRQ(ierr); 929 } 930 PetscFunctionReturn(0); 931 } 932 933 PetscErrorCode MatMult_MPIAIJ(Mat A,Vec xx,Vec yy) 934 { 935 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 936 PetscErrorCode ierr; 937 PetscInt nt; 938 939 PetscFunctionBegin; 940 ierr = VecGetLocalSize(xx,&nt);CHKERRQ(ierr); 941 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); 942 ierr = VecScatterBegin(a->Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 943 ierr = (*a->A->ops->mult)(a->A,xx,yy);CHKERRQ(ierr); 944 ierr = VecScatterEnd(a->Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 945 ierr = (*a->B->ops->multadd)(a->B,a->lvec,yy,yy);CHKERRQ(ierr); 946 PetscFunctionReturn(0); 947 } 948 949 PetscErrorCode MatMultDiagonalBlock_MPIAIJ(Mat A,Vec bb,Vec xx) 950 { 951 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 952 PetscErrorCode ierr; 953 954 PetscFunctionBegin; 955 ierr = MatMultDiagonalBlock(a->A,bb,xx);CHKERRQ(ierr); 956 PetscFunctionReturn(0); 957 } 958 959 PetscErrorCode MatMultAdd_MPIAIJ(Mat A,Vec xx,Vec yy,Vec zz) 960 { 961 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 962 PetscErrorCode ierr; 963 964 PetscFunctionBegin; 965 ierr = VecScatterBegin(a->Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 966 ierr = (*a->A->ops->multadd)(a->A,xx,yy,zz);CHKERRQ(ierr); 967 ierr = VecScatterEnd(a->Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 968 ierr = (*a->B->ops->multadd)(a->B,a->lvec,zz,zz);CHKERRQ(ierr); 969 PetscFunctionReturn(0); 970 } 971 972 PetscErrorCode MatMultTranspose_MPIAIJ(Mat A,Vec xx,Vec yy) 973 { 974 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 975 PetscErrorCode ierr; 976 PetscBool merged; 977 978 PetscFunctionBegin; 979 ierr = VecScatterGetMerged(a->Mvctx,&merged);CHKERRQ(ierr); 980 /* do nondiagonal part */ 981 ierr = (*a->B->ops->multtranspose)(a->B,xx,a->lvec);CHKERRQ(ierr); 982 if (!merged) { 983 /* send it on its way */ 984 ierr = VecScatterBegin(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 985 /* do local part */ 986 ierr = (*a->A->ops->multtranspose)(a->A,xx,yy);CHKERRQ(ierr); 987 /* receive remote parts: note this assumes the values are not actually */ 988 /* added in yy until the next line, */ 989 ierr = VecScatterEnd(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 990 } else { 991 /* do local part */ 992 ierr = (*a->A->ops->multtranspose)(a->A,xx,yy);CHKERRQ(ierr); 993 /* send it on its way */ 994 ierr = VecScatterBegin(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 995 /* values actually were received in the Begin() but we need to call this nop */ 996 ierr = VecScatterEnd(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 997 } 998 PetscFunctionReturn(0); 999 } 1000 1001 PetscErrorCode MatIsTranspose_MPIAIJ(Mat Amat,Mat Bmat,PetscReal tol,PetscBool *f) 1002 { 1003 MPI_Comm comm; 1004 Mat_MPIAIJ *Aij = (Mat_MPIAIJ*) Amat->data, *Bij; 1005 Mat Adia = Aij->A, Bdia, Aoff,Boff,*Aoffs,*Boffs; 1006 IS Me,Notme; 1007 PetscErrorCode ierr; 1008 PetscInt M,N,first,last,*notme,i; 1009 PetscMPIInt size; 1010 1011 PetscFunctionBegin; 1012 /* Easy test: symmetric diagonal block */ 1013 Bij = (Mat_MPIAIJ*) Bmat->data; Bdia = Bij->A; 1014 ierr = MatIsTranspose(Adia,Bdia,tol,f);CHKERRQ(ierr); 1015 if (!*f) PetscFunctionReturn(0); 1016 ierr = PetscObjectGetComm((PetscObject)Amat,&comm);CHKERRQ(ierr); 1017 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 1018 if (size == 1) PetscFunctionReturn(0); 1019 1020 /* Hard test: off-diagonal block. This takes a MatCreateSubMatrix. */ 1021 ierr = MatGetSize(Amat,&M,&N);CHKERRQ(ierr); 1022 ierr = MatGetOwnershipRange(Amat,&first,&last);CHKERRQ(ierr); 1023 ierr = PetscMalloc1(N-last+first,¬me);CHKERRQ(ierr); 1024 for (i=0; i<first; i++) notme[i] = i; 1025 for (i=last; i<M; i++) notme[i-last+first] = i; 1026 ierr = ISCreateGeneral(MPI_COMM_SELF,N-last+first,notme,PETSC_COPY_VALUES,&Notme);CHKERRQ(ierr); 1027 ierr = ISCreateStride(MPI_COMM_SELF,last-first,first,1,&Me);CHKERRQ(ierr); 1028 ierr = MatCreateSubMatrices(Amat,1,&Me,&Notme,MAT_INITIAL_MATRIX,&Aoffs);CHKERRQ(ierr); 1029 Aoff = Aoffs[0]; 1030 ierr = MatCreateSubMatrices(Bmat,1,&Notme,&Me,MAT_INITIAL_MATRIX,&Boffs);CHKERRQ(ierr); 1031 Boff = Boffs[0]; 1032 ierr = MatIsTranspose(Aoff,Boff,tol,f);CHKERRQ(ierr); 1033 ierr = MatDestroyMatrices(1,&Aoffs);CHKERRQ(ierr); 1034 ierr = MatDestroyMatrices(1,&Boffs);CHKERRQ(ierr); 1035 ierr = ISDestroy(&Me);CHKERRQ(ierr); 1036 ierr = ISDestroy(&Notme);CHKERRQ(ierr); 1037 ierr = PetscFree(notme);CHKERRQ(ierr); 1038 PetscFunctionReturn(0); 1039 } 1040 1041 PetscErrorCode MatMultTransposeAdd_MPIAIJ(Mat A,Vec xx,Vec yy,Vec zz) 1042 { 1043 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 1044 PetscErrorCode ierr; 1045 1046 PetscFunctionBegin; 1047 /* do nondiagonal part */ 1048 ierr = (*a->B->ops->multtranspose)(a->B,xx,a->lvec);CHKERRQ(ierr); 1049 /* send it on its way */ 1050 ierr = VecScatterBegin(a->Mvctx,a->lvec,zz,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1051 /* do local part */ 1052 ierr = (*a->A->ops->multtransposeadd)(a->A,xx,yy,zz);CHKERRQ(ierr); 1053 /* receive remote parts */ 1054 ierr = VecScatterEnd(a->Mvctx,a->lvec,zz,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1055 PetscFunctionReturn(0); 1056 } 1057 1058 /* 1059 This only works correctly for square matrices where the subblock A->A is the 1060 diagonal block 1061 */ 1062 PetscErrorCode MatGetDiagonal_MPIAIJ(Mat A,Vec v) 1063 { 1064 PetscErrorCode ierr; 1065 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 1066 1067 PetscFunctionBegin; 1068 if (A->rmap->N != A->cmap->N) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"Supports only square matrix where A->A is diag block"); 1069 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"); 1070 ierr = MatGetDiagonal(a->A,v);CHKERRQ(ierr); 1071 PetscFunctionReturn(0); 1072 } 1073 1074 PetscErrorCode MatScale_MPIAIJ(Mat A,PetscScalar aa) 1075 { 1076 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 1077 PetscErrorCode ierr; 1078 1079 PetscFunctionBegin; 1080 ierr = MatScale(a->A,aa);CHKERRQ(ierr); 1081 ierr = MatScale(a->B,aa);CHKERRQ(ierr); 1082 PetscFunctionReturn(0); 1083 } 1084 1085 PetscErrorCode MatDestroy_MPIAIJ(Mat mat) 1086 { 1087 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 1088 PetscErrorCode ierr; 1089 1090 PetscFunctionBegin; 1091 #if defined(PETSC_USE_LOG) 1092 PetscLogObjectState((PetscObject)mat,"Rows=%D, Cols=%D",mat->rmap->N,mat->cmap->N); 1093 #endif 1094 ierr = MatStashDestroy_Private(&mat->stash);CHKERRQ(ierr); 1095 ierr = VecDestroy(&aij->diag);CHKERRQ(ierr); 1096 ierr = MatDestroy(&aij->A);CHKERRQ(ierr); 1097 ierr = MatDestroy(&aij->B);CHKERRQ(ierr); 1098 #if defined(PETSC_USE_CTABLE) 1099 ierr = PetscTableDestroy(&aij->colmap);CHKERRQ(ierr); 1100 #else 1101 ierr = PetscFree(aij->colmap);CHKERRQ(ierr); 1102 #endif 1103 ierr = PetscFree(aij->garray);CHKERRQ(ierr); 1104 ierr = VecDestroy(&aij->lvec);CHKERRQ(ierr); 1105 ierr = VecScatterDestroy(&aij->Mvctx);CHKERRQ(ierr); 1106 ierr = PetscFree2(aij->rowvalues,aij->rowindices);CHKERRQ(ierr); 1107 ierr = PetscFree(aij->ld);CHKERRQ(ierr); 1108 ierr = PetscFree(mat->data);CHKERRQ(ierr); 1109 1110 ierr = PetscObjectChangeTypeName((PetscObject)mat,0);CHKERRQ(ierr); 1111 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatStoreValues_C",NULL);CHKERRQ(ierr); 1112 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatRetrieveValues_C",NULL);CHKERRQ(ierr); 1113 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatIsTranspose_C",NULL);CHKERRQ(ierr); 1114 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatMPIAIJSetPreallocation_C",NULL);CHKERRQ(ierr); 1115 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatMPIAIJSetPreallocationCSR_C",NULL);CHKERRQ(ierr); 1116 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatDiagonalScaleLocal_C",NULL);CHKERRQ(ierr); 1117 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatConvert_mpiaij_mpisbaij_C",NULL);CHKERRQ(ierr); 1118 #if defined(PETSC_HAVE_ELEMENTAL) 1119 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatConvert_mpiaij_elemental_C",NULL);CHKERRQ(ierr); 1120 #endif 1121 #if defined(PETSC_HAVE_HYPRE) 1122 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatConvert_mpiaij_hypre_C",NULL);CHKERRQ(ierr); 1123 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatMatMatMult_transpose_mpiaij_mpiaij_C",NULL);CHKERRQ(ierr); 1124 #endif 1125 PetscFunctionReturn(0); 1126 } 1127 1128 PetscErrorCode MatView_MPIAIJ_Binary(Mat mat,PetscViewer viewer) 1129 { 1130 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 1131 Mat_SeqAIJ *A = (Mat_SeqAIJ*)aij->A->data; 1132 Mat_SeqAIJ *B = (Mat_SeqAIJ*)aij->B->data; 1133 PetscErrorCode ierr; 1134 PetscMPIInt rank,size,tag = ((PetscObject)viewer)->tag; 1135 int fd; 1136 PetscInt nz,header[4],*row_lengths,*range=0,rlen,i; 1137 PetscInt nzmax,*column_indices,j,k,col,*garray = aij->garray,cnt,cstart = mat->cmap->rstart,rnz = 0; 1138 PetscScalar *column_values; 1139 PetscInt message_count,flowcontrolcount; 1140 FILE *file; 1141 1142 PetscFunctionBegin; 1143 ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr); 1144 ierr = MPI_Comm_size(PetscObjectComm((PetscObject)mat),&size);CHKERRQ(ierr); 1145 nz = A->nz + B->nz; 1146 ierr = PetscViewerBinaryGetDescriptor(viewer,&fd);CHKERRQ(ierr); 1147 if (!rank) { 1148 header[0] = MAT_FILE_CLASSID; 1149 header[1] = mat->rmap->N; 1150 header[2] = mat->cmap->N; 1151 1152 ierr = MPI_Reduce(&nz,&header[3],1,MPIU_INT,MPI_SUM,0,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1153 ierr = PetscBinaryWrite(fd,header,4,PETSC_INT,PETSC_TRUE);CHKERRQ(ierr); 1154 /* get largest number of rows any processor has */ 1155 rlen = mat->rmap->n; 1156 range = mat->rmap->range; 1157 for (i=1; i<size; i++) rlen = PetscMax(rlen,range[i+1] - range[i]); 1158 } else { 1159 ierr = MPI_Reduce(&nz,0,1,MPIU_INT,MPI_SUM,0,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1160 rlen = mat->rmap->n; 1161 } 1162 1163 /* load up the local row counts */ 1164 ierr = PetscMalloc1(rlen+1,&row_lengths);CHKERRQ(ierr); 1165 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]; 1166 1167 /* store the row lengths to the file */ 1168 ierr = PetscViewerFlowControlStart(viewer,&message_count,&flowcontrolcount);CHKERRQ(ierr); 1169 if (!rank) { 1170 ierr = PetscBinaryWrite(fd,row_lengths,mat->rmap->n,PETSC_INT,PETSC_TRUE);CHKERRQ(ierr); 1171 for (i=1; i<size; i++) { 1172 ierr = PetscViewerFlowControlStepMaster(viewer,i,&message_count,flowcontrolcount);CHKERRQ(ierr); 1173 rlen = range[i+1] - range[i]; 1174 ierr = MPIULong_Recv(row_lengths,rlen,MPIU_INT,i,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1175 ierr = PetscBinaryWrite(fd,row_lengths,rlen,PETSC_INT,PETSC_TRUE);CHKERRQ(ierr); 1176 } 1177 ierr = PetscViewerFlowControlEndMaster(viewer,&message_count);CHKERRQ(ierr); 1178 } else { 1179 ierr = PetscViewerFlowControlStepWorker(viewer,rank,&message_count);CHKERRQ(ierr); 1180 ierr = MPIULong_Send(row_lengths,mat->rmap->n,MPIU_INT,0,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1181 ierr = PetscViewerFlowControlEndWorker(viewer,&message_count);CHKERRQ(ierr); 1182 } 1183 ierr = PetscFree(row_lengths);CHKERRQ(ierr); 1184 1185 /* load up the local column indices */ 1186 nzmax = nz; /* th processor needs space a largest processor needs */ 1187 ierr = MPI_Reduce(&nz,&nzmax,1,MPIU_INT,MPI_MAX,0,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1188 ierr = PetscMalloc1(nzmax+1,&column_indices);CHKERRQ(ierr); 1189 cnt = 0; 1190 for (i=0; i<mat->rmap->n; i++) { 1191 for (j=B->i[i]; j<B->i[i+1]; j++) { 1192 if ((col = garray[B->j[j]]) > cstart) break; 1193 column_indices[cnt++] = col; 1194 } 1195 for (k=A->i[i]; k<A->i[i+1]; k++) column_indices[cnt++] = A->j[k] + cstart; 1196 for (; j<B->i[i+1]; j++) column_indices[cnt++] = garray[B->j[j]]; 1197 } 1198 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); 1199 1200 /* store the column indices to the file */ 1201 ierr = PetscViewerFlowControlStart(viewer,&message_count,&flowcontrolcount);CHKERRQ(ierr); 1202 if (!rank) { 1203 MPI_Status status; 1204 ierr = PetscBinaryWrite(fd,column_indices,nz,PETSC_INT,PETSC_TRUE);CHKERRQ(ierr); 1205 for (i=1; i<size; i++) { 1206 ierr = PetscViewerFlowControlStepMaster(viewer,i,&message_count,flowcontrolcount);CHKERRQ(ierr); 1207 ierr = MPI_Recv(&rnz,1,MPIU_INT,i,tag,PetscObjectComm((PetscObject)mat),&status);CHKERRQ(ierr); 1208 if (rnz > nzmax) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_LIB,"Internal PETSc error: nz = %D nzmax = %D",nz,nzmax); 1209 ierr = MPIULong_Recv(column_indices,rnz,MPIU_INT,i,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1210 ierr = PetscBinaryWrite(fd,column_indices,rnz,PETSC_INT,PETSC_TRUE);CHKERRQ(ierr); 1211 } 1212 ierr = PetscViewerFlowControlEndMaster(viewer,&message_count);CHKERRQ(ierr); 1213 } else { 1214 ierr = PetscViewerFlowControlStepWorker(viewer,rank,&message_count);CHKERRQ(ierr); 1215 ierr = MPI_Send(&nz,1,MPIU_INT,0,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1216 ierr = MPIULong_Send(column_indices,nz,MPIU_INT,0,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1217 ierr = PetscViewerFlowControlEndWorker(viewer,&message_count);CHKERRQ(ierr); 1218 } 1219 ierr = PetscFree(column_indices);CHKERRQ(ierr); 1220 1221 /* load up the local column values */ 1222 ierr = PetscMalloc1(nzmax+1,&column_values);CHKERRQ(ierr); 1223 cnt = 0; 1224 for (i=0; i<mat->rmap->n; i++) { 1225 for (j=B->i[i]; j<B->i[i+1]; j++) { 1226 if (garray[B->j[j]] > cstart) break; 1227 column_values[cnt++] = B->a[j]; 1228 } 1229 for (k=A->i[i]; k<A->i[i+1]; k++) column_values[cnt++] = A->a[k]; 1230 for (; j<B->i[i+1]; j++) column_values[cnt++] = B->a[j]; 1231 } 1232 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); 1233 1234 /* store the column values to the file */ 1235 ierr = PetscViewerFlowControlStart(viewer,&message_count,&flowcontrolcount);CHKERRQ(ierr); 1236 if (!rank) { 1237 MPI_Status status; 1238 ierr = PetscBinaryWrite(fd,column_values,nz,PETSC_SCALAR,PETSC_TRUE);CHKERRQ(ierr); 1239 for (i=1; i<size; i++) { 1240 ierr = PetscViewerFlowControlStepMaster(viewer,i,&message_count,flowcontrolcount);CHKERRQ(ierr); 1241 ierr = MPI_Recv(&rnz,1,MPIU_INT,i,tag,PetscObjectComm((PetscObject)mat),&status);CHKERRQ(ierr); 1242 if (rnz > nzmax) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_LIB,"Internal PETSc error: nz = %D nzmax = %D",nz,nzmax); 1243 ierr = MPIULong_Recv(column_values,rnz,MPIU_SCALAR,i,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1244 ierr = PetscBinaryWrite(fd,column_values,rnz,PETSC_SCALAR,PETSC_TRUE);CHKERRQ(ierr); 1245 } 1246 ierr = PetscViewerFlowControlEndMaster(viewer,&message_count);CHKERRQ(ierr); 1247 } else { 1248 ierr = PetscViewerFlowControlStepWorker(viewer,rank,&message_count);CHKERRQ(ierr); 1249 ierr = MPI_Send(&nz,1,MPIU_INT,0,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1250 ierr = MPIULong_Send(column_values,nz,MPIU_SCALAR,0,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1251 ierr = PetscViewerFlowControlEndWorker(viewer,&message_count);CHKERRQ(ierr); 1252 } 1253 ierr = PetscFree(column_values);CHKERRQ(ierr); 1254 1255 ierr = PetscViewerBinaryGetInfoPointer(viewer,&file);CHKERRQ(ierr); 1256 if (file) fprintf(file,"-matload_block_size %d\n",(int)PetscAbs(mat->rmap->bs)); 1257 PetscFunctionReturn(0); 1258 } 1259 1260 #include <petscdraw.h> 1261 PetscErrorCode MatView_MPIAIJ_ASCIIorDraworSocket(Mat mat,PetscViewer viewer) 1262 { 1263 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 1264 PetscErrorCode ierr; 1265 PetscMPIInt rank = aij->rank,size = aij->size; 1266 PetscBool isdraw,iascii,isbinary; 1267 PetscViewer sviewer; 1268 PetscViewerFormat format; 1269 1270 PetscFunctionBegin; 1271 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERDRAW,&isdraw);CHKERRQ(ierr); 1272 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii);CHKERRQ(ierr); 1273 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERBINARY,&isbinary);CHKERRQ(ierr); 1274 if (iascii) { 1275 ierr = PetscViewerGetFormat(viewer,&format);CHKERRQ(ierr); 1276 if (format == PETSC_VIEWER_ASCII_INFO_DETAIL) { 1277 MatInfo info; 1278 PetscBool inodes; 1279 1280 ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr); 1281 ierr = MatGetInfo(mat,MAT_LOCAL,&info);CHKERRQ(ierr); 1282 ierr = MatInodeGetInodeSizes(aij->A,NULL,(PetscInt**)&inodes,NULL);CHKERRQ(ierr); 1283 ierr = PetscViewerASCIIPushSynchronized(viewer);CHKERRQ(ierr); 1284 if (!inodes) { 1285 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"[%d] Local rows %D nz %D nz alloced %D mem %D, not using I-node routines\n", 1286 rank,mat->rmap->n,(PetscInt)info.nz_used,(PetscInt)info.nz_allocated,(PetscInt)info.memory);CHKERRQ(ierr); 1287 } else { 1288 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"[%d] Local rows %D nz %D nz alloced %D mem %D, using I-node routines\n", 1289 rank,mat->rmap->n,(PetscInt)info.nz_used,(PetscInt)info.nz_allocated,(PetscInt)info.memory);CHKERRQ(ierr); 1290 } 1291 ierr = MatGetInfo(aij->A,MAT_LOCAL,&info);CHKERRQ(ierr); 1292 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"[%d] on-diagonal part: nz %D \n",rank,(PetscInt)info.nz_used);CHKERRQ(ierr); 1293 ierr = MatGetInfo(aij->B,MAT_LOCAL,&info);CHKERRQ(ierr); 1294 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"[%d] off-diagonal part: nz %D \n",rank,(PetscInt)info.nz_used);CHKERRQ(ierr); 1295 ierr = PetscViewerFlush(viewer);CHKERRQ(ierr); 1296 ierr = PetscViewerASCIIPopSynchronized(viewer);CHKERRQ(ierr); 1297 ierr = PetscViewerASCIIPrintf(viewer,"Information on VecScatter used in matrix-vector product: \n");CHKERRQ(ierr); 1298 ierr = VecScatterView(aij->Mvctx,viewer);CHKERRQ(ierr); 1299 PetscFunctionReturn(0); 1300 } else if (format == PETSC_VIEWER_ASCII_INFO) { 1301 PetscInt inodecount,inodelimit,*inodes; 1302 ierr = MatInodeGetInodeSizes(aij->A,&inodecount,&inodes,&inodelimit);CHKERRQ(ierr); 1303 if (inodes) { 1304 ierr = PetscViewerASCIIPrintf(viewer,"using I-node (on process 0) routines: found %D nodes, limit used is %D\n",inodecount,inodelimit);CHKERRQ(ierr); 1305 } else { 1306 ierr = PetscViewerASCIIPrintf(viewer,"not using I-node (on process 0) routines\n");CHKERRQ(ierr); 1307 } 1308 PetscFunctionReturn(0); 1309 } else if (format == PETSC_VIEWER_ASCII_FACTOR_INFO) { 1310 PetscFunctionReturn(0); 1311 } 1312 } else if (isbinary) { 1313 if (size == 1) { 1314 ierr = PetscObjectSetName((PetscObject)aij->A,((PetscObject)mat)->name);CHKERRQ(ierr); 1315 ierr = MatView(aij->A,viewer);CHKERRQ(ierr); 1316 } else { 1317 ierr = MatView_MPIAIJ_Binary(mat,viewer);CHKERRQ(ierr); 1318 } 1319 PetscFunctionReturn(0); 1320 } else if (isdraw) { 1321 PetscDraw draw; 1322 PetscBool isnull; 1323 ierr = PetscViewerDrawGetDraw(viewer,0,&draw);CHKERRQ(ierr); 1324 ierr = PetscDrawIsNull(draw,&isnull);CHKERRQ(ierr); 1325 if (isnull) PetscFunctionReturn(0); 1326 } 1327 1328 { 1329 /* assemble the entire matrix onto first processor. */ 1330 Mat A; 1331 Mat_SeqAIJ *Aloc; 1332 PetscInt M = mat->rmap->N,N = mat->cmap->N,m,*ai,*aj,row,*cols,i,*ct; 1333 MatScalar *a; 1334 1335 ierr = MatCreate(PetscObjectComm((PetscObject)mat),&A);CHKERRQ(ierr); 1336 if (!rank) { 1337 ierr = MatSetSizes(A,M,N,M,N);CHKERRQ(ierr); 1338 } else { 1339 ierr = MatSetSizes(A,0,0,M,N);CHKERRQ(ierr); 1340 } 1341 /* This is just a temporary matrix, so explicitly using MATMPIAIJ is probably best */ 1342 ierr = MatSetType(A,MATMPIAIJ);CHKERRQ(ierr); 1343 ierr = MatMPIAIJSetPreallocation(A,0,NULL,0,NULL);CHKERRQ(ierr); 1344 ierr = MatSetOption(A,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 1345 ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)A);CHKERRQ(ierr); 1346 1347 /* copy over the A part */ 1348 Aloc = (Mat_SeqAIJ*)aij->A->data; 1349 m = aij->A->rmap->n; ai = Aloc->i; aj = Aloc->j; a = Aloc->a; 1350 row = mat->rmap->rstart; 1351 for (i=0; i<ai[m]; i++) aj[i] += mat->cmap->rstart; 1352 for (i=0; i<m; i++) { 1353 ierr = MatSetValues(A,1,&row,ai[i+1]-ai[i],aj,a,INSERT_VALUES);CHKERRQ(ierr); 1354 row++; 1355 a += ai[i+1]-ai[i]; aj += ai[i+1]-ai[i]; 1356 } 1357 aj = Aloc->j; 1358 for (i=0; i<ai[m]; i++) aj[i] -= mat->cmap->rstart; 1359 1360 /* copy over the B part */ 1361 Aloc = (Mat_SeqAIJ*)aij->B->data; 1362 m = aij->B->rmap->n; ai = Aloc->i; aj = Aloc->j; a = Aloc->a; 1363 row = mat->rmap->rstart; 1364 ierr = PetscMalloc1(ai[m]+1,&cols);CHKERRQ(ierr); 1365 ct = cols; 1366 for (i=0; i<ai[m]; i++) cols[i] = aij->garray[aj[i]]; 1367 for (i=0; i<m; i++) { 1368 ierr = MatSetValues(A,1,&row,ai[i+1]-ai[i],cols,a,INSERT_VALUES);CHKERRQ(ierr); 1369 row++; 1370 a += ai[i+1]-ai[i]; cols += ai[i+1]-ai[i]; 1371 } 1372 ierr = PetscFree(ct);CHKERRQ(ierr); 1373 ierr = MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1374 ierr = MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1375 /* 1376 Everyone has to call to draw the matrix since the graphics waits are 1377 synchronized across all processors that share the PetscDraw object 1378 */ 1379 ierr = PetscViewerGetSubViewer(viewer,PETSC_COMM_SELF,&sviewer);CHKERRQ(ierr); 1380 if (!rank) { 1381 ierr = PetscObjectSetName((PetscObject)((Mat_MPIAIJ*)(A->data))->A,((PetscObject)mat)->name);CHKERRQ(ierr); 1382 ierr = MatView_SeqAIJ(((Mat_MPIAIJ*)(A->data))->A,sviewer);CHKERRQ(ierr); 1383 } 1384 ierr = PetscViewerRestoreSubViewer(viewer,PETSC_COMM_SELF,&sviewer);CHKERRQ(ierr); 1385 ierr = PetscViewerFlush(viewer);CHKERRQ(ierr); 1386 ierr = MatDestroy(&A);CHKERRQ(ierr); 1387 } 1388 PetscFunctionReturn(0); 1389 } 1390 1391 PetscErrorCode MatView_MPIAIJ(Mat mat,PetscViewer viewer) 1392 { 1393 PetscErrorCode ierr; 1394 PetscBool iascii,isdraw,issocket,isbinary; 1395 1396 PetscFunctionBegin; 1397 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii);CHKERRQ(ierr); 1398 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERDRAW,&isdraw);CHKERRQ(ierr); 1399 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERBINARY,&isbinary);CHKERRQ(ierr); 1400 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERSOCKET,&issocket);CHKERRQ(ierr); 1401 if (iascii || isdraw || isbinary || issocket) { 1402 ierr = MatView_MPIAIJ_ASCIIorDraworSocket(mat,viewer);CHKERRQ(ierr); 1403 } 1404 PetscFunctionReturn(0); 1405 } 1406 1407 PetscErrorCode MatSOR_MPIAIJ(Mat matin,Vec bb,PetscReal omega,MatSORType flag,PetscReal fshift,PetscInt its,PetscInt lits,Vec xx) 1408 { 1409 Mat_MPIAIJ *mat = (Mat_MPIAIJ*)matin->data; 1410 PetscErrorCode ierr; 1411 Vec bb1 = 0; 1412 PetscBool hasop; 1413 1414 PetscFunctionBegin; 1415 if (flag == SOR_APPLY_UPPER) { 1416 ierr = (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);CHKERRQ(ierr); 1417 PetscFunctionReturn(0); 1418 } 1419 1420 if (its > 1 || ~flag & SOR_ZERO_INITIAL_GUESS || flag & SOR_EISENSTAT) { 1421 ierr = VecDuplicate(bb,&bb1);CHKERRQ(ierr); 1422 } 1423 1424 if ((flag & SOR_LOCAL_SYMMETRIC_SWEEP) == SOR_LOCAL_SYMMETRIC_SWEEP) { 1425 if (flag & SOR_ZERO_INITIAL_GUESS) { 1426 ierr = (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);CHKERRQ(ierr); 1427 its--; 1428 } 1429 1430 while (its--) { 1431 ierr = VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1432 ierr = VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1433 1434 /* update rhs: bb1 = bb - B*x */ 1435 ierr = VecScale(mat->lvec,-1.0);CHKERRQ(ierr); 1436 ierr = (*mat->B->ops->multadd)(mat->B,mat->lvec,bb,bb1);CHKERRQ(ierr); 1437 1438 /* local sweep */ 1439 ierr = (*mat->A->ops->sor)(mat->A,bb1,omega,SOR_SYMMETRIC_SWEEP,fshift,lits,1,xx);CHKERRQ(ierr); 1440 } 1441 } else if (flag & SOR_LOCAL_FORWARD_SWEEP) { 1442 if (flag & SOR_ZERO_INITIAL_GUESS) { 1443 ierr = (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);CHKERRQ(ierr); 1444 its--; 1445 } 1446 while (its--) { 1447 ierr = VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1448 ierr = VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1449 1450 /* update rhs: bb1 = bb - B*x */ 1451 ierr = VecScale(mat->lvec,-1.0);CHKERRQ(ierr); 1452 ierr = (*mat->B->ops->multadd)(mat->B,mat->lvec,bb,bb1);CHKERRQ(ierr); 1453 1454 /* local sweep */ 1455 ierr = (*mat->A->ops->sor)(mat->A,bb1,omega,SOR_FORWARD_SWEEP,fshift,lits,1,xx);CHKERRQ(ierr); 1456 } 1457 } else if (flag & SOR_LOCAL_BACKWARD_SWEEP) { 1458 if (flag & SOR_ZERO_INITIAL_GUESS) { 1459 ierr = (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);CHKERRQ(ierr); 1460 its--; 1461 } 1462 while (its--) { 1463 ierr = VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1464 ierr = VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1465 1466 /* update rhs: bb1 = bb - B*x */ 1467 ierr = VecScale(mat->lvec,-1.0);CHKERRQ(ierr); 1468 ierr = (*mat->B->ops->multadd)(mat->B,mat->lvec,bb,bb1);CHKERRQ(ierr); 1469 1470 /* local sweep */ 1471 ierr = (*mat->A->ops->sor)(mat->A,bb1,omega,SOR_BACKWARD_SWEEP,fshift,lits,1,xx);CHKERRQ(ierr); 1472 } 1473 } else if (flag & SOR_EISENSTAT) { 1474 Vec xx1; 1475 1476 ierr = VecDuplicate(bb,&xx1);CHKERRQ(ierr); 1477 ierr = (*mat->A->ops->sor)(mat->A,bb,omega,(MatSORType)(SOR_ZERO_INITIAL_GUESS | SOR_LOCAL_BACKWARD_SWEEP),fshift,lits,1,xx);CHKERRQ(ierr); 1478 1479 ierr = VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1480 ierr = VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1481 if (!mat->diag) { 1482 ierr = MatCreateVecs(matin,&mat->diag,NULL);CHKERRQ(ierr); 1483 ierr = MatGetDiagonal(matin,mat->diag);CHKERRQ(ierr); 1484 } 1485 ierr = MatHasOperation(matin,MATOP_MULT_DIAGONAL_BLOCK,&hasop);CHKERRQ(ierr); 1486 if (hasop) { 1487 ierr = MatMultDiagonalBlock(matin,xx,bb1);CHKERRQ(ierr); 1488 } else { 1489 ierr = VecPointwiseMult(bb1,mat->diag,xx);CHKERRQ(ierr); 1490 } 1491 ierr = VecAYPX(bb1,(omega-2.0)/omega,bb);CHKERRQ(ierr); 1492 1493 ierr = MatMultAdd(mat->B,mat->lvec,bb1,bb1);CHKERRQ(ierr); 1494 1495 /* local sweep */ 1496 ierr = (*mat->A->ops->sor)(mat->A,bb1,omega,(MatSORType)(SOR_ZERO_INITIAL_GUESS | SOR_LOCAL_FORWARD_SWEEP),fshift,lits,1,xx1);CHKERRQ(ierr); 1497 ierr = VecAXPY(xx,1.0,xx1);CHKERRQ(ierr); 1498 ierr = VecDestroy(&xx1);CHKERRQ(ierr); 1499 } else SETERRQ(PetscObjectComm((PetscObject)matin),PETSC_ERR_SUP,"Parallel SOR not supported"); 1500 1501 ierr = VecDestroy(&bb1);CHKERRQ(ierr); 1502 1503 matin->factorerrortype = mat->A->factorerrortype; 1504 PetscFunctionReturn(0); 1505 } 1506 1507 PetscErrorCode MatPermute_MPIAIJ(Mat A,IS rowp,IS colp,Mat *B) 1508 { 1509 Mat aA,aB,Aperm; 1510 const PetscInt *rwant,*cwant,*gcols,*ai,*bi,*aj,*bj; 1511 PetscScalar *aa,*ba; 1512 PetscInt i,j,m,n,ng,anz,bnz,*dnnz,*onnz,*tdnnz,*tonnz,*rdest,*cdest,*work,*gcdest; 1513 PetscSF rowsf,sf; 1514 IS parcolp = NULL; 1515 PetscBool done; 1516 PetscErrorCode ierr; 1517 1518 PetscFunctionBegin; 1519 ierr = MatGetLocalSize(A,&m,&n);CHKERRQ(ierr); 1520 ierr = ISGetIndices(rowp,&rwant);CHKERRQ(ierr); 1521 ierr = ISGetIndices(colp,&cwant);CHKERRQ(ierr); 1522 ierr = PetscMalloc3(PetscMax(m,n),&work,m,&rdest,n,&cdest);CHKERRQ(ierr); 1523 1524 /* Invert row permutation to find out where my rows should go */ 1525 ierr = PetscSFCreate(PetscObjectComm((PetscObject)A),&rowsf);CHKERRQ(ierr); 1526 ierr = PetscSFSetGraphLayout(rowsf,A->rmap,A->rmap->n,NULL,PETSC_OWN_POINTER,rwant);CHKERRQ(ierr); 1527 ierr = PetscSFSetFromOptions(rowsf);CHKERRQ(ierr); 1528 for (i=0; i<m; i++) work[i] = A->rmap->rstart + i; 1529 ierr = PetscSFReduceBegin(rowsf,MPIU_INT,work,rdest,MPIU_REPLACE);CHKERRQ(ierr); 1530 ierr = PetscSFReduceEnd(rowsf,MPIU_INT,work,rdest,MPIU_REPLACE);CHKERRQ(ierr); 1531 1532 /* Invert column permutation to find out where my columns should go */ 1533 ierr = PetscSFCreate(PetscObjectComm((PetscObject)A),&sf);CHKERRQ(ierr); 1534 ierr = PetscSFSetGraphLayout(sf,A->cmap,A->cmap->n,NULL,PETSC_OWN_POINTER,cwant);CHKERRQ(ierr); 1535 ierr = PetscSFSetFromOptions(sf);CHKERRQ(ierr); 1536 for (i=0; i<n; i++) work[i] = A->cmap->rstart + i; 1537 ierr = PetscSFReduceBegin(sf,MPIU_INT,work,cdest,MPIU_REPLACE);CHKERRQ(ierr); 1538 ierr = PetscSFReduceEnd(sf,MPIU_INT,work,cdest,MPIU_REPLACE);CHKERRQ(ierr); 1539 ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 1540 1541 ierr = ISRestoreIndices(rowp,&rwant);CHKERRQ(ierr); 1542 ierr = ISRestoreIndices(colp,&cwant);CHKERRQ(ierr); 1543 ierr = MatMPIAIJGetSeqAIJ(A,&aA,&aB,&gcols);CHKERRQ(ierr); 1544 1545 /* Find out where my gcols should go */ 1546 ierr = MatGetSize(aB,NULL,&ng);CHKERRQ(ierr); 1547 ierr = PetscMalloc1(ng,&gcdest);CHKERRQ(ierr); 1548 ierr = PetscSFCreate(PetscObjectComm((PetscObject)A),&sf);CHKERRQ(ierr); 1549 ierr = PetscSFSetGraphLayout(sf,A->cmap,ng,NULL,PETSC_OWN_POINTER,gcols);CHKERRQ(ierr); 1550 ierr = PetscSFSetFromOptions(sf);CHKERRQ(ierr); 1551 ierr = PetscSFBcastBegin(sf,MPIU_INT,cdest,gcdest);CHKERRQ(ierr); 1552 ierr = PetscSFBcastEnd(sf,MPIU_INT,cdest,gcdest);CHKERRQ(ierr); 1553 ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 1554 1555 ierr = PetscCalloc4(m,&dnnz,m,&onnz,m,&tdnnz,m,&tonnz);CHKERRQ(ierr); 1556 ierr = MatGetRowIJ(aA,0,PETSC_FALSE,PETSC_FALSE,&anz,&ai,&aj,&done);CHKERRQ(ierr); 1557 ierr = MatGetRowIJ(aB,0,PETSC_FALSE,PETSC_FALSE,&bnz,&bi,&bj,&done);CHKERRQ(ierr); 1558 for (i=0; i<m; i++) { 1559 PetscInt row = rdest[i],rowner; 1560 ierr = PetscLayoutFindOwner(A->rmap,row,&rowner);CHKERRQ(ierr); 1561 for (j=ai[i]; j<ai[i+1]; j++) { 1562 PetscInt cowner,col = cdest[aj[j]]; 1563 ierr = PetscLayoutFindOwner(A->cmap,col,&cowner);CHKERRQ(ierr); /* Could build an index for the columns to eliminate this search */ 1564 if (rowner == cowner) dnnz[i]++; 1565 else onnz[i]++; 1566 } 1567 for (j=bi[i]; j<bi[i+1]; j++) { 1568 PetscInt cowner,col = gcdest[bj[j]]; 1569 ierr = PetscLayoutFindOwner(A->cmap,col,&cowner);CHKERRQ(ierr); 1570 if (rowner == cowner) dnnz[i]++; 1571 else onnz[i]++; 1572 } 1573 } 1574 ierr = PetscSFBcastBegin(rowsf,MPIU_INT,dnnz,tdnnz);CHKERRQ(ierr); 1575 ierr = PetscSFBcastEnd(rowsf,MPIU_INT,dnnz,tdnnz);CHKERRQ(ierr); 1576 ierr = PetscSFBcastBegin(rowsf,MPIU_INT,onnz,tonnz);CHKERRQ(ierr); 1577 ierr = PetscSFBcastEnd(rowsf,MPIU_INT,onnz,tonnz);CHKERRQ(ierr); 1578 ierr = PetscSFDestroy(&rowsf);CHKERRQ(ierr); 1579 1580 ierr = MatCreateAIJ(PetscObjectComm((PetscObject)A),A->rmap->n,A->cmap->n,A->rmap->N,A->cmap->N,0,tdnnz,0,tonnz,&Aperm);CHKERRQ(ierr); 1581 ierr = MatSeqAIJGetArray(aA,&aa);CHKERRQ(ierr); 1582 ierr = MatSeqAIJGetArray(aB,&ba);CHKERRQ(ierr); 1583 for (i=0; i<m; i++) { 1584 PetscInt *acols = dnnz,*bcols = onnz; /* Repurpose now-unneeded arrays */ 1585 PetscInt j0,rowlen; 1586 rowlen = ai[i+1] - ai[i]; 1587 for (j0=j=0; j<rowlen; j0=j) { /* rowlen could be larger than number of rows m, so sum in batches */ 1588 for ( ; j<PetscMin(rowlen,j0+m); j++) acols[j-j0] = cdest[aj[ai[i]+j]]; 1589 ierr = MatSetValues(Aperm,1,&rdest[i],j-j0,acols,aa+ai[i]+j0,INSERT_VALUES);CHKERRQ(ierr); 1590 } 1591 rowlen = bi[i+1] - bi[i]; 1592 for (j0=j=0; j<rowlen; j0=j) { 1593 for ( ; j<PetscMin(rowlen,j0+m); j++) bcols[j-j0] = gcdest[bj[bi[i]+j]]; 1594 ierr = MatSetValues(Aperm,1,&rdest[i],j-j0,bcols,ba+bi[i]+j0,INSERT_VALUES);CHKERRQ(ierr); 1595 } 1596 } 1597 ierr = MatAssemblyBegin(Aperm,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1598 ierr = MatAssemblyEnd(Aperm,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1599 ierr = MatRestoreRowIJ(aA,0,PETSC_FALSE,PETSC_FALSE,&anz,&ai,&aj,&done);CHKERRQ(ierr); 1600 ierr = MatRestoreRowIJ(aB,0,PETSC_FALSE,PETSC_FALSE,&bnz,&bi,&bj,&done);CHKERRQ(ierr); 1601 ierr = MatSeqAIJRestoreArray(aA,&aa);CHKERRQ(ierr); 1602 ierr = MatSeqAIJRestoreArray(aB,&ba);CHKERRQ(ierr); 1603 ierr = PetscFree4(dnnz,onnz,tdnnz,tonnz);CHKERRQ(ierr); 1604 ierr = PetscFree3(work,rdest,cdest);CHKERRQ(ierr); 1605 ierr = PetscFree(gcdest);CHKERRQ(ierr); 1606 if (parcolp) {ierr = ISDestroy(&colp);CHKERRQ(ierr);} 1607 *B = Aperm; 1608 PetscFunctionReturn(0); 1609 } 1610 1611 PetscErrorCode MatGetGhosts_MPIAIJ(Mat mat,PetscInt *nghosts,const PetscInt *ghosts[]) 1612 { 1613 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 1614 PetscErrorCode ierr; 1615 1616 PetscFunctionBegin; 1617 ierr = MatGetSize(aij->B,NULL,nghosts);CHKERRQ(ierr); 1618 if (ghosts) *ghosts = aij->garray; 1619 PetscFunctionReturn(0); 1620 } 1621 1622 PetscErrorCode MatGetInfo_MPIAIJ(Mat matin,MatInfoType flag,MatInfo *info) 1623 { 1624 Mat_MPIAIJ *mat = (Mat_MPIAIJ*)matin->data; 1625 Mat A = mat->A,B = mat->B; 1626 PetscErrorCode ierr; 1627 PetscReal isend[5],irecv[5]; 1628 1629 PetscFunctionBegin; 1630 info->block_size = 1.0; 1631 ierr = MatGetInfo(A,MAT_LOCAL,info);CHKERRQ(ierr); 1632 1633 isend[0] = info->nz_used; isend[1] = info->nz_allocated; isend[2] = info->nz_unneeded; 1634 isend[3] = info->memory; isend[4] = info->mallocs; 1635 1636 ierr = MatGetInfo(B,MAT_LOCAL,info);CHKERRQ(ierr); 1637 1638 isend[0] += info->nz_used; isend[1] += info->nz_allocated; isend[2] += info->nz_unneeded; 1639 isend[3] += info->memory; isend[4] += info->mallocs; 1640 if (flag == MAT_LOCAL) { 1641 info->nz_used = isend[0]; 1642 info->nz_allocated = isend[1]; 1643 info->nz_unneeded = isend[2]; 1644 info->memory = isend[3]; 1645 info->mallocs = isend[4]; 1646 } else if (flag == MAT_GLOBAL_MAX) { 1647 ierr = MPIU_Allreduce(isend,irecv,5,MPIU_REAL,MPIU_MAX,PetscObjectComm((PetscObject)matin));CHKERRQ(ierr); 1648 1649 info->nz_used = irecv[0]; 1650 info->nz_allocated = irecv[1]; 1651 info->nz_unneeded = irecv[2]; 1652 info->memory = irecv[3]; 1653 info->mallocs = irecv[4]; 1654 } else if (flag == MAT_GLOBAL_SUM) { 1655 ierr = MPIU_Allreduce(isend,irecv,5,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)matin));CHKERRQ(ierr); 1656 1657 info->nz_used = irecv[0]; 1658 info->nz_allocated = irecv[1]; 1659 info->nz_unneeded = irecv[2]; 1660 info->memory = irecv[3]; 1661 info->mallocs = irecv[4]; 1662 } 1663 info->fill_ratio_given = 0; /* no parallel LU/ILU/Cholesky */ 1664 info->fill_ratio_needed = 0; 1665 info->factor_mallocs = 0; 1666 PetscFunctionReturn(0); 1667 } 1668 1669 PetscErrorCode MatSetOption_MPIAIJ(Mat A,MatOption op,PetscBool flg) 1670 { 1671 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 1672 PetscErrorCode ierr; 1673 1674 PetscFunctionBegin; 1675 switch (op) { 1676 case MAT_NEW_NONZERO_LOCATIONS: 1677 case MAT_NEW_NONZERO_ALLOCATION_ERR: 1678 case MAT_UNUSED_NONZERO_LOCATION_ERR: 1679 case MAT_KEEP_NONZERO_PATTERN: 1680 case MAT_NEW_NONZERO_LOCATION_ERR: 1681 case MAT_USE_INODES: 1682 case MAT_IGNORE_ZERO_ENTRIES: 1683 MatCheckPreallocated(A,1); 1684 ierr = MatSetOption(a->A,op,flg);CHKERRQ(ierr); 1685 ierr = MatSetOption(a->B,op,flg);CHKERRQ(ierr); 1686 break; 1687 case MAT_ROW_ORIENTED: 1688 MatCheckPreallocated(A,1); 1689 a->roworiented = flg; 1690 1691 ierr = MatSetOption(a->A,op,flg);CHKERRQ(ierr); 1692 ierr = MatSetOption(a->B,op,flg);CHKERRQ(ierr); 1693 break; 1694 case MAT_NEW_DIAGONALS: 1695 ierr = PetscInfo1(A,"Option %s ignored\n",MatOptions[op]);CHKERRQ(ierr); 1696 break; 1697 case MAT_IGNORE_OFF_PROC_ENTRIES: 1698 a->donotstash = flg; 1699 break; 1700 case MAT_SPD: 1701 A->spd_set = PETSC_TRUE; 1702 A->spd = flg; 1703 if (flg) { 1704 A->symmetric = PETSC_TRUE; 1705 A->structurally_symmetric = PETSC_TRUE; 1706 A->symmetric_set = PETSC_TRUE; 1707 A->structurally_symmetric_set = PETSC_TRUE; 1708 } 1709 break; 1710 case MAT_SYMMETRIC: 1711 MatCheckPreallocated(A,1); 1712 ierr = MatSetOption(a->A,op,flg);CHKERRQ(ierr); 1713 break; 1714 case MAT_STRUCTURALLY_SYMMETRIC: 1715 MatCheckPreallocated(A,1); 1716 ierr = MatSetOption(a->A,op,flg);CHKERRQ(ierr); 1717 break; 1718 case MAT_HERMITIAN: 1719 MatCheckPreallocated(A,1); 1720 ierr = MatSetOption(a->A,op,flg);CHKERRQ(ierr); 1721 break; 1722 case MAT_SYMMETRY_ETERNAL: 1723 MatCheckPreallocated(A,1); 1724 ierr = MatSetOption(a->A,op,flg);CHKERRQ(ierr); 1725 break; 1726 case MAT_SUBMAT_SINGLEIS: 1727 A->submat_singleis = flg; 1728 break; 1729 default: 1730 SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"unknown option %d",op); 1731 } 1732 PetscFunctionReturn(0); 1733 } 1734 1735 PetscErrorCode MatGetRow_MPIAIJ(Mat matin,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v) 1736 { 1737 Mat_MPIAIJ *mat = (Mat_MPIAIJ*)matin->data; 1738 PetscScalar *vworkA,*vworkB,**pvA,**pvB,*v_p; 1739 PetscErrorCode ierr; 1740 PetscInt i,*cworkA,*cworkB,**pcA,**pcB,cstart = matin->cmap->rstart; 1741 PetscInt nztot,nzA,nzB,lrow,rstart = matin->rmap->rstart,rend = matin->rmap->rend; 1742 PetscInt *cmap,*idx_p; 1743 1744 PetscFunctionBegin; 1745 if (mat->getrowactive) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Already active"); 1746 mat->getrowactive = PETSC_TRUE; 1747 1748 if (!mat->rowvalues && (idx || v)) { 1749 /* 1750 allocate enough space to hold information from the longest row. 1751 */ 1752 Mat_SeqAIJ *Aa = (Mat_SeqAIJ*)mat->A->data,*Ba = (Mat_SeqAIJ*)mat->B->data; 1753 PetscInt max = 1,tmp; 1754 for (i=0; i<matin->rmap->n; i++) { 1755 tmp = Aa->i[i+1] - Aa->i[i] + Ba->i[i+1] - Ba->i[i]; 1756 if (max < tmp) max = tmp; 1757 } 1758 ierr = PetscMalloc2(max,&mat->rowvalues,max,&mat->rowindices);CHKERRQ(ierr); 1759 } 1760 1761 if (row < rstart || row >= rend) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Only local rows"); 1762 lrow = row - rstart; 1763 1764 pvA = &vworkA; pcA = &cworkA; pvB = &vworkB; pcB = &cworkB; 1765 if (!v) {pvA = 0; pvB = 0;} 1766 if (!idx) {pcA = 0; if (!v) pcB = 0;} 1767 ierr = (*mat->A->ops->getrow)(mat->A,lrow,&nzA,pcA,pvA);CHKERRQ(ierr); 1768 ierr = (*mat->B->ops->getrow)(mat->B,lrow,&nzB,pcB,pvB);CHKERRQ(ierr); 1769 nztot = nzA + nzB; 1770 1771 cmap = mat->garray; 1772 if (v || idx) { 1773 if (nztot) { 1774 /* Sort by increasing column numbers, assuming A and B already sorted */ 1775 PetscInt imark = -1; 1776 if (v) { 1777 *v = v_p = mat->rowvalues; 1778 for (i=0; i<nzB; i++) { 1779 if (cmap[cworkB[i]] < cstart) v_p[i] = vworkB[i]; 1780 else break; 1781 } 1782 imark = i; 1783 for (i=0; i<nzA; i++) v_p[imark+i] = vworkA[i]; 1784 for (i=imark; i<nzB; i++) v_p[nzA+i] = vworkB[i]; 1785 } 1786 if (idx) { 1787 *idx = idx_p = mat->rowindices; 1788 if (imark > -1) { 1789 for (i=0; i<imark; i++) { 1790 idx_p[i] = cmap[cworkB[i]]; 1791 } 1792 } else { 1793 for (i=0; i<nzB; i++) { 1794 if (cmap[cworkB[i]] < cstart) idx_p[i] = cmap[cworkB[i]]; 1795 else break; 1796 } 1797 imark = i; 1798 } 1799 for (i=0; i<nzA; i++) idx_p[imark+i] = cstart + cworkA[i]; 1800 for (i=imark; i<nzB; i++) idx_p[nzA+i] = cmap[cworkB[i]]; 1801 } 1802 } else { 1803 if (idx) *idx = 0; 1804 if (v) *v = 0; 1805 } 1806 } 1807 *nz = nztot; 1808 ierr = (*mat->A->ops->restorerow)(mat->A,lrow,&nzA,pcA,pvA);CHKERRQ(ierr); 1809 ierr = (*mat->B->ops->restorerow)(mat->B,lrow,&nzB,pcB,pvB);CHKERRQ(ierr); 1810 PetscFunctionReturn(0); 1811 } 1812 1813 PetscErrorCode MatRestoreRow_MPIAIJ(Mat mat,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v) 1814 { 1815 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 1816 1817 PetscFunctionBegin; 1818 if (!aij->getrowactive) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"MatGetRow() must be called first"); 1819 aij->getrowactive = PETSC_FALSE; 1820 PetscFunctionReturn(0); 1821 } 1822 1823 PetscErrorCode MatNorm_MPIAIJ(Mat mat,NormType type,PetscReal *norm) 1824 { 1825 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 1826 Mat_SeqAIJ *amat = (Mat_SeqAIJ*)aij->A->data,*bmat = (Mat_SeqAIJ*)aij->B->data; 1827 PetscErrorCode ierr; 1828 PetscInt i,j,cstart = mat->cmap->rstart; 1829 PetscReal sum = 0.0; 1830 MatScalar *v; 1831 1832 PetscFunctionBegin; 1833 if (aij->size == 1) { 1834 ierr = MatNorm(aij->A,type,norm);CHKERRQ(ierr); 1835 } else { 1836 if (type == NORM_FROBENIUS) { 1837 v = amat->a; 1838 for (i=0; i<amat->nz; i++) { 1839 sum += PetscRealPart(PetscConj(*v)*(*v)); v++; 1840 } 1841 v = bmat->a; 1842 for (i=0; i<bmat->nz; i++) { 1843 sum += PetscRealPart(PetscConj(*v)*(*v)); v++; 1844 } 1845 ierr = MPIU_Allreduce(&sum,norm,1,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1846 *norm = PetscSqrtReal(*norm); 1847 ierr = PetscLogFlops(2*amat->nz+2*bmat->nz);CHKERRQ(ierr); 1848 } else if (type == NORM_1) { /* max column norm */ 1849 PetscReal *tmp,*tmp2; 1850 PetscInt *jj,*garray = aij->garray; 1851 ierr = PetscCalloc1(mat->cmap->N+1,&tmp);CHKERRQ(ierr); 1852 ierr = PetscMalloc1(mat->cmap->N+1,&tmp2);CHKERRQ(ierr); 1853 *norm = 0.0; 1854 v = amat->a; jj = amat->j; 1855 for (j=0; j<amat->nz; j++) { 1856 tmp[cstart + *jj++] += PetscAbsScalar(*v); v++; 1857 } 1858 v = bmat->a; jj = bmat->j; 1859 for (j=0; j<bmat->nz; j++) { 1860 tmp[garray[*jj++]] += PetscAbsScalar(*v); v++; 1861 } 1862 ierr = MPIU_Allreduce(tmp,tmp2,mat->cmap->N,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1863 for (j=0; j<mat->cmap->N; j++) { 1864 if (tmp2[j] > *norm) *norm = tmp2[j]; 1865 } 1866 ierr = PetscFree(tmp);CHKERRQ(ierr); 1867 ierr = PetscFree(tmp2);CHKERRQ(ierr); 1868 ierr = PetscLogFlops(PetscMax(amat->nz+bmat->nz-1,0));CHKERRQ(ierr); 1869 } else if (type == NORM_INFINITY) { /* max row norm */ 1870 PetscReal ntemp = 0.0; 1871 for (j=0; j<aij->A->rmap->n; j++) { 1872 v = amat->a + amat->i[j]; 1873 sum = 0.0; 1874 for (i=0; i<amat->i[j+1]-amat->i[j]; i++) { 1875 sum += PetscAbsScalar(*v); v++; 1876 } 1877 v = bmat->a + bmat->i[j]; 1878 for (i=0; i<bmat->i[j+1]-bmat->i[j]; i++) { 1879 sum += PetscAbsScalar(*v); v++; 1880 } 1881 if (sum > ntemp) ntemp = sum; 1882 } 1883 ierr = MPIU_Allreduce(&ntemp,norm,1,MPIU_REAL,MPIU_MAX,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1884 ierr = PetscLogFlops(PetscMax(amat->nz+bmat->nz-1,0));CHKERRQ(ierr); 1885 } else SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"No support for two norm"); 1886 } 1887 PetscFunctionReturn(0); 1888 } 1889 1890 PetscErrorCode MatTranspose_MPIAIJ(Mat A,MatReuse reuse,Mat *matout) 1891 { 1892 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 1893 Mat_SeqAIJ *Aloc=(Mat_SeqAIJ*)a->A->data,*Bloc=(Mat_SeqAIJ*)a->B->data; 1894 PetscErrorCode ierr; 1895 PetscInt M = A->rmap->N,N = A->cmap->N,ma,na,mb,nb,*ai,*aj,*bi,*bj,row,*cols,*cols_tmp,i; 1896 PetscInt cstart = A->cmap->rstart,ncol; 1897 Mat B; 1898 MatScalar *array; 1899 1900 PetscFunctionBegin; 1901 if (reuse == MAT_INPLACE_MATRIX && M != N) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_SIZ,"Square matrix only for in-place"); 1902 1903 ma = A->rmap->n; na = A->cmap->n; mb = a->B->rmap->n; nb = a->B->cmap->n; 1904 ai = Aloc->i; aj = Aloc->j; 1905 bi = Bloc->i; bj = Bloc->j; 1906 if (reuse == MAT_INITIAL_MATRIX || *matout == A) { 1907 PetscInt *d_nnz,*g_nnz,*o_nnz; 1908 PetscSFNode *oloc; 1909 PETSC_UNUSED PetscSF sf; 1910 1911 ierr = PetscMalloc4(na,&d_nnz,na,&o_nnz,nb,&g_nnz,nb,&oloc);CHKERRQ(ierr); 1912 /* compute d_nnz for preallocation */ 1913 ierr = PetscMemzero(d_nnz,na*sizeof(PetscInt));CHKERRQ(ierr); 1914 for (i=0; i<ai[ma]; i++) { 1915 d_nnz[aj[i]]++; 1916 aj[i] += cstart; /* global col index to be used by MatSetValues() */ 1917 } 1918 /* compute local off-diagonal contributions */ 1919 ierr = PetscMemzero(g_nnz,nb*sizeof(PetscInt));CHKERRQ(ierr); 1920 for (i=0; i<bi[ma]; i++) g_nnz[bj[i]]++; 1921 /* map those to global */ 1922 ierr = PetscSFCreate(PetscObjectComm((PetscObject)A),&sf);CHKERRQ(ierr); 1923 ierr = PetscSFSetGraphLayout(sf,A->cmap,nb,NULL,PETSC_USE_POINTER,a->garray);CHKERRQ(ierr); 1924 ierr = PetscSFSetFromOptions(sf);CHKERRQ(ierr); 1925 ierr = PetscMemzero(o_nnz,na*sizeof(PetscInt));CHKERRQ(ierr); 1926 ierr = PetscSFReduceBegin(sf,MPIU_INT,g_nnz,o_nnz,MPIU_SUM);CHKERRQ(ierr); 1927 ierr = PetscSFReduceEnd(sf,MPIU_INT,g_nnz,o_nnz,MPIU_SUM);CHKERRQ(ierr); 1928 ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 1929 1930 ierr = MatCreate(PetscObjectComm((PetscObject)A),&B);CHKERRQ(ierr); 1931 ierr = MatSetSizes(B,A->cmap->n,A->rmap->n,N,M);CHKERRQ(ierr); 1932 ierr = MatSetBlockSizes(B,PetscAbs(A->cmap->bs),PetscAbs(A->rmap->bs));CHKERRQ(ierr); 1933 ierr = MatSetType(B,((PetscObject)A)->type_name);CHKERRQ(ierr); 1934 ierr = MatMPIAIJSetPreallocation(B,0,d_nnz,0,o_nnz);CHKERRQ(ierr); 1935 ierr = PetscFree4(d_nnz,o_nnz,g_nnz,oloc);CHKERRQ(ierr); 1936 } else { 1937 B = *matout; 1938 ierr = MatSetOption(B,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_TRUE);CHKERRQ(ierr); 1939 for (i=0; i<ai[ma]; i++) aj[i] += cstart; /* global col index to be used by MatSetValues() */ 1940 } 1941 1942 /* copy over the A part */ 1943 array = Aloc->a; 1944 row = A->rmap->rstart; 1945 for (i=0; i<ma; i++) { 1946 ncol = ai[i+1]-ai[i]; 1947 ierr = MatSetValues(B,ncol,aj,1,&row,array,INSERT_VALUES);CHKERRQ(ierr); 1948 row++; 1949 array += ncol; aj += ncol; 1950 } 1951 aj = Aloc->j; 1952 for (i=0; i<ai[ma]; i++) aj[i] -= cstart; /* resume local col index */ 1953 1954 /* copy over the B part */ 1955 ierr = PetscCalloc1(bi[mb],&cols);CHKERRQ(ierr); 1956 array = Bloc->a; 1957 row = A->rmap->rstart; 1958 for (i=0; i<bi[mb]; i++) cols[i] = a->garray[bj[i]]; 1959 cols_tmp = cols; 1960 for (i=0; i<mb; i++) { 1961 ncol = bi[i+1]-bi[i]; 1962 ierr = MatSetValues(B,ncol,cols_tmp,1,&row,array,INSERT_VALUES);CHKERRQ(ierr); 1963 row++; 1964 array += ncol; cols_tmp += ncol; 1965 } 1966 ierr = PetscFree(cols);CHKERRQ(ierr); 1967 1968 ierr = MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1969 ierr = MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1970 if (reuse == MAT_INITIAL_MATRIX || reuse == MAT_REUSE_MATRIX) { 1971 *matout = B; 1972 } else { 1973 ierr = MatHeaderMerge(A,&B);CHKERRQ(ierr); 1974 } 1975 PetscFunctionReturn(0); 1976 } 1977 1978 PetscErrorCode MatDiagonalScale_MPIAIJ(Mat mat,Vec ll,Vec rr) 1979 { 1980 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 1981 Mat a = aij->A,b = aij->B; 1982 PetscErrorCode ierr; 1983 PetscInt s1,s2,s3; 1984 1985 PetscFunctionBegin; 1986 ierr = MatGetLocalSize(mat,&s2,&s3);CHKERRQ(ierr); 1987 if (rr) { 1988 ierr = VecGetLocalSize(rr,&s1);CHKERRQ(ierr); 1989 if (s1!=s3) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"right vector non-conforming local size"); 1990 /* Overlap communication with computation. */ 1991 ierr = VecScatterBegin(aij->Mvctx,rr,aij->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1992 } 1993 if (ll) { 1994 ierr = VecGetLocalSize(ll,&s1);CHKERRQ(ierr); 1995 if (s1!=s2) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"left vector non-conforming local size"); 1996 ierr = (*b->ops->diagonalscale)(b,ll,0);CHKERRQ(ierr); 1997 } 1998 /* scale the diagonal block */ 1999 ierr = (*a->ops->diagonalscale)(a,ll,rr);CHKERRQ(ierr); 2000 2001 if (rr) { 2002 /* Do a scatter end and then right scale the off-diagonal block */ 2003 ierr = VecScatterEnd(aij->Mvctx,rr,aij->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2004 ierr = (*b->ops->diagonalscale)(b,0,aij->lvec);CHKERRQ(ierr); 2005 } 2006 PetscFunctionReturn(0); 2007 } 2008 2009 PetscErrorCode MatSetUnfactored_MPIAIJ(Mat A) 2010 { 2011 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 2012 PetscErrorCode ierr; 2013 2014 PetscFunctionBegin; 2015 ierr = MatSetUnfactored(a->A);CHKERRQ(ierr); 2016 PetscFunctionReturn(0); 2017 } 2018 2019 PetscErrorCode MatEqual_MPIAIJ(Mat A,Mat B,PetscBool *flag) 2020 { 2021 Mat_MPIAIJ *matB = (Mat_MPIAIJ*)B->data,*matA = (Mat_MPIAIJ*)A->data; 2022 Mat a,b,c,d; 2023 PetscBool flg; 2024 PetscErrorCode ierr; 2025 2026 PetscFunctionBegin; 2027 a = matA->A; b = matA->B; 2028 c = matB->A; d = matB->B; 2029 2030 ierr = MatEqual(a,c,&flg);CHKERRQ(ierr); 2031 if (flg) { 2032 ierr = MatEqual(b,d,&flg);CHKERRQ(ierr); 2033 } 2034 ierr = MPIU_Allreduce(&flg,flag,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)A));CHKERRQ(ierr); 2035 PetscFunctionReturn(0); 2036 } 2037 2038 PetscErrorCode MatCopy_MPIAIJ(Mat A,Mat B,MatStructure str) 2039 { 2040 PetscErrorCode ierr; 2041 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 2042 Mat_MPIAIJ *b = (Mat_MPIAIJ*)B->data; 2043 2044 PetscFunctionBegin; 2045 /* If the two matrices don't have the same copy implementation, they aren't compatible for fast copy. */ 2046 if ((str != SAME_NONZERO_PATTERN) || (A->ops->copy != B->ops->copy)) { 2047 /* because of the column compression in the off-processor part of the matrix a->B, 2048 the number of columns in a->B and b->B may be different, hence we cannot call 2049 the MatCopy() directly on the two parts. If need be, we can provide a more 2050 efficient copy than the MatCopy_Basic() by first uncompressing the a->B matrices 2051 then copying the submatrices */ 2052 ierr = MatCopy_Basic(A,B,str);CHKERRQ(ierr); 2053 } else { 2054 ierr = MatCopy(a->A,b->A,str);CHKERRQ(ierr); 2055 ierr = MatCopy(a->B,b->B,str);CHKERRQ(ierr); 2056 } 2057 PetscFunctionReturn(0); 2058 } 2059 2060 PetscErrorCode MatSetUp_MPIAIJ(Mat A) 2061 { 2062 PetscErrorCode ierr; 2063 2064 PetscFunctionBegin; 2065 ierr = MatMPIAIJSetPreallocation(A,PETSC_DEFAULT,0,PETSC_DEFAULT,0);CHKERRQ(ierr); 2066 PetscFunctionReturn(0); 2067 } 2068 2069 /* 2070 Computes the number of nonzeros per row needed for preallocation when X and Y 2071 have different nonzero structure. 2072 */ 2073 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) 2074 { 2075 PetscInt i,j,k,nzx,nzy; 2076 2077 PetscFunctionBegin; 2078 /* Set the number of nonzeros in the new matrix */ 2079 for (i=0; i<m; i++) { 2080 const PetscInt *xjj = xj+xi[i],*yjj = yj+yi[i]; 2081 nzx = xi[i+1] - xi[i]; 2082 nzy = yi[i+1] - yi[i]; 2083 nnz[i] = 0; 2084 for (j=0,k=0; j<nzx; j++) { /* Point in X */ 2085 for (; k<nzy && yltog[yjj[k]]<xltog[xjj[j]]; k++) nnz[i]++; /* Catch up to X */ 2086 if (k<nzy && yltog[yjj[k]]==xltog[xjj[j]]) k++; /* Skip duplicate */ 2087 nnz[i]++; 2088 } 2089 for (; k<nzy; k++) nnz[i]++; 2090 } 2091 PetscFunctionReturn(0); 2092 } 2093 2094 /* This is the same as MatAXPYGetPreallocation_SeqAIJ, except that the local-to-global map is provided */ 2095 static PetscErrorCode MatAXPYGetPreallocation_MPIAIJ(Mat Y,const PetscInt *yltog,Mat X,const PetscInt *xltog,PetscInt *nnz) 2096 { 2097 PetscErrorCode ierr; 2098 PetscInt m = Y->rmap->N; 2099 Mat_SeqAIJ *x = (Mat_SeqAIJ*)X->data; 2100 Mat_SeqAIJ *y = (Mat_SeqAIJ*)Y->data; 2101 2102 PetscFunctionBegin; 2103 ierr = MatAXPYGetPreallocation_MPIX_private(m,x->i,x->j,xltog,y->i,y->j,yltog,nnz);CHKERRQ(ierr); 2104 PetscFunctionReturn(0); 2105 } 2106 2107 PetscErrorCode MatAXPY_MPIAIJ(Mat Y,PetscScalar a,Mat X,MatStructure str) 2108 { 2109 PetscErrorCode ierr; 2110 Mat_MPIAIJ *xx = (Mat_MPIAIJ*)X->data,*yy = (Mat_MPIAIJ*)Y->data; 2111 PetscBLASInt bnz,one=1; 2112 Mat_SeqAIJ *x,*y; 2113 2114 PetscFunctionBegin; 2115 if (str == SAME_NONZERO_PATTERN) { 2116 PetscScalar alpha = a; 2117 x = (Mat_SeqAIJ*)xx->A->data; 2118 ierr = PetscBLASIntCast(x->nz,&bnz);CHKERRQ(ierr); 2119 y = (Mat_SeqAIJ*)yy->A->data; 2120 PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&bnz,&alpha,x->a,&one,y->a,&one)); 2121 x = (Mat_SeqAIJ*)xx->B->data; 2122 y = (Mat_SeqAIJ*)yy->B->data; 2123 ierr = PetscBLASIntCast(x->nz,&bnz);CHKERRQ(ierr); 2124 PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&bnz,&alpha,x->a,&one,y->a,&one)); 2125 ierr = PetscObjectStateIncrease((PetscObject)Y);CHKERRQ(ierr); 2126 } else if (str == SUBSET_NONZERO_PATTERN) { /* nonzeros of X is a subset of Y's */ 2127 ierr = MatAXPY_Basic(Y,a,X,str);CHKERRQ(ierr); 2128 } else { 2129 Mat B; 2130 PetscInt *nnz_d,*nnz_o; 2131 ierr = PetscMalloc1(yy->A->rmap->N,&nnz_d);CHKERRQ(ierr); 2132 ierr = PetscMalloc1(yy->B->rmap->N,&nnz_o);CHKERRQ(ierr); 2133 ierr = MatCreate(PetscObjectComm((PetscObject)Y),&B);CHKERRQ(ierr); 2134 ierr = PetscObjectSetName((PetscObject)B,((PetscObject)Y)->name);CHKERRQ(ierr); 2135 ierr = MatSetSizes(B,Y->rmap->n,Y->cmap->n,Y->rmap->N,Y->cmap->N);CHKERRQ(ierr); 2136 ierr = MatSetBlockSizesFromMats(B,Y,Y);CHKERRQ(ierr); 2137 ierr = MatSetType(B,MATMPIAIJ);CHKERRQ(ierr); 2138 ierr = MatAXPYGetPreallocation_SeqAIJ(yy->A,xx->A,nnz_d);CHKERRQ(ierr); 2139 ierr = MatAXPYGetPreallocation_MPIAIJ(yy->B,yy->garray,xx->B,xx->garray,nnz_o);CHKERRQ(ierr); 2140 ierr = MatMPIAIJSetPreallocation(B,0,nnz_d,0,nnz_o);CHKERRQ(ierr); 2141 ierr = MatAXPY_BasicWithPreallocation(B,Y,a,X,str);CHKERRQ(ierr); 2142 ierr = MatHeaderReplace(Y,&B);CHKERRQ(ierr); 2143 ierr = PetscFree(nnz_d);CHKERRQ(ierr); 2144 ierr = PetscFree(nnz_o);CHKERRQ(ierr); 2145 } 2146 PetscFunctionReturn(0); 2147 } 2148 2149 extern PetscErrorCode MatConjugate_SeqAIJ(Mat); 2150 2151 PetscErrorCode MatConjugate_MPIAIJ(Mat mat) 2152 { 2153 #if defined(PETSC_USE_COMPLEX) 2154 PetscErrorCode ierr; 2155 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 2156 2157 PetscFunctionBegin; 2158 ierr = MatConjugate_SeqAIJ(aij->A);CHKERRQ(ierr); 2159 ierr = MatConjugate_SeqAIJ(aij->B);CHKERRQ(ierr); 2160 #else 2161 PetscFunctionBegin; 2162 #endif 2163 PetscFunctionReturn(0); 2164 } 2165 2166 PetscErrorCode MatRealPart_MPIAIJ(Mat A) 2167 { 2168 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 2169 PetscErrorCode ierr; 2170 2171 PetscFunctionBegin; 2172 ierr = MatRealPart(a->A);CHKERRQ(ierr); 2173 ierr = MatRealPart(a->B);CHKERRQ(ierr); 2174 PetscFunctionReturn(0); 2175 } 2176 2177 PetscErrorCode MatImaginaryPart_MPIAIJ(Mat A) 2178 { 2179 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 2180 PetscErrorCode ierr; 2181 2182 PetscFunctionBegin; 2183 ierr = MatImaginaryPart(a->A);CHKERRQ(ierr); 2184 ierr = MatImaginaryPart(a->B);CHKERRQ(ierr); 2185 PetscFunctionReturn(0); 2186 } 2187 2188 PetscErrorCode MatGetRowMaxAbs_MPIAIJ(Mat A, Vec v, PetscInt idx[]) 2189 { 2190 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 2191 PetscErrorCode ierr; 2192 PetscInt i,*idxb = 0; 2193 PetscScalar *va,*vb; 2194 Vec vtmp; 2195 2196 PetscFunctionBegin; 2197 ierr = MatGetRowMaxAbs(a->A,v,idx);CHKERRQ(ierr); 2198 ierr = VecGetArray(v,&va);CHKERRQ(ierr); 2199 if (idx) { 2200 for (i=0; i<A->rmap->n; i++) { 2201 if (PetscAbsScalar(va[i])) idx[i] += A->cmap->rstart; 2202 } 2203 } 2204 2205 ierr = VecCreateSeq(PETSC_COMM_SELF,A->rmap->n,&vtmp);CHKERRQ(ierr); 2206 if (idx) { 2207 ierr = PetscMalloc1(A->rmap->n,&idxb);CHKERRQ(ierr); 2208 } 2209 ierr = MatGetRowMaxAbs(a->B,vtmp,idxb);CHKERRQ(ierr); 2210 ierr = VecGetArray(vtmp,&vb);CHKERRQ(ierr); 2211 2212 for (i=0; i<A->rmap->n; i++) { 2213 if (PetscAbsScalar(va[i]) < PetscAbsScalar(vb[i])) { 2214 va[i] = vb[i]; 2215 if (idx) idx[i] = a->garray[idxb[i]]; 2216 } 2217 } 2218 2219 ierr = VecRestoreArray(v,&va);CHKERRQ(ierr); 2220 ierr = VecRestoreArray(vtmp,&vb);CHKERRQ(ierr); 2221 ierr = PetscFree(idxb);CHKERRQ(ierr); 2222 ierr = VecDestroy(&vtmp);CHKERRQ(ierr); 2223 PetscFunctionReturn(0); 2224 } 2225 2226 PetscErrorCode MatGetRowMinAbs_MPIAIJ(Mat A, Vec v, PetscInt idx[]) 2227 { 2228 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 2229 PetscErrorCode ierr; 2230 PetscInt i,*idxb = 0; 2231 PetscScalar *va,*vb; 2232 Vec vtmp; 2233 2234 PetscFunctionBegin; 2235 ierr = MatGetRowMinAbs(a->A,v,idx);CHKERRQ(ierr); 2236 ierr = VecGetArray(v,&va);CHKERRQ(ierr); 2237 if (idx) { 2238 for (i=0; i<A->cmap->n; i++) { 2239 if (PetscAbsScalar(va[i])) idx[i] += A->cmap->rstart; 2240 } 2241 } 2242 2243 ierr = VecCreateSeq(PETSC_COMM_SELF,A->rmap->n,&vtmp);CHKERRQ(ierr); 2244 if (idx) { 2245 ierr = PetscMalloc1(A->rmap->n,&idxb);CHKERRQ(ierr); 2246 } 2247 ierr = MatGetRowMinAbs(a->B,vtmp,idxb);CHKERRQ(ierr); 2248 ierr = VecGetArray(vtmp,&vb);CHKERRQ(ierr); 2249 2250 for (i=0; i<A->rmap->n; i++) { 2251 if (PetscAbsScalar(va[i]) > PetscAbsScalar(vb[i])) { 2252 va[i] = vb[i]; 2253 if (idx) idx[i] = a->garray[idxb[i]]; 2254 } 2255 } 2256 2257 ierr = VecRestoreArray(v,&va);CHKERRQ(ierr); 2258 ierr = VecRestoreArray(vtmp,&vb);CHKERRQ(ierr); 2259 ierr = PetscFree(idxb);CHKERRQ(ierr); 2260 ierr = VecDestroy(&vtmp);CHKERRQ(ierr); 2261 PetscFunctionReturn(0); 2262 } 2263 2264 PetscErrorCode MatGetRowMin_MPIAIJ(Mat A, Vec v, PetscInt idx[]) 2265 { 2266 Mat_MPIAIJ *mat = (Mat_MPIAIJ*) A->data; 2267 PetscInt n = A->rmap->n; 2268 PetscInt cstart = A->cmap->rstart; 2269 PetscInt *cmap = mat->garray; 2270 PetscInt *diagIdx, *offdiagIdx; 2271 Vec diagV, offdiagV; 2272 PetscScalar *a, *diagA, *offdiagA; 2273 PetscInt r; 2274 PetscErrorCode ierr; 2275 2276 PetscFunctionBegin; 2277 ierr = PetscMalloc2(n,&diagIdx,n,&offdiagIdx);CHKERRQ(ierr); 2278 ierr = VecCreateSeq(PetscObjectComm((PetscObject)A), n, &diagV);CHKERRQ(ierr); 2279 ierr = VecCreateSeq(PetscObjectComm((PetscObject)A), n, &offdiagV);CHKERRQ(ierr); 2280 ierr = MatGetRowMin(mat->A, diagV, diagIdx);CHKERRQ(ierr); 2281 ierr = MatGetRowMin(mat->B, offdiagV, offdiagIdx);CHKERRQ(ierr); 2282 ierr = VecGetArray(v, &a);CHKERRQ(ierr); 2283 ierr = VecGetArray(diagV, &diagA);CHKERRQ(ierr); 2284 ierr = VecGetArray(offdiagV, &offdiagA);CHKERRQ(ierr); 2285 for (r = 0; r < n; ++r) { 2286 if (PetscAbsScalar(diagA[r]) <= PetscAbsScalar(offdiagA[r])) { 2287 a[r] = diagA[r]; 2288 idx[r] = cstart + diagIdx[r]; 2289 } else { 2290 a[r] = offdiagA[r]; 2291 idx[r] = cmap[offdiagIdx[r]]; 2292 } 2293 } 2294 ierr = VecRestoreArray(v, &a);CHKERRQ(ierr); 2295 ierr = VecRestoreArray(diagV, &diagA);CHKERRQ(ierr); 2296 ierr = VecRestoreArray(offdiagV, &offdiagA);CHKERRQ(ierr); 2297 ierr = VecDestroy(&diagV);CHKERRQ(ierr); 2298 ierr = VecDestroy(&offdiagV);CHKERRQ(ierr); 2299 ierr = PetscFree2(diagIdx, offdiagIdx);CHKERRQ(ierr); 2300 PetscFunctionReturn(0); 2301 } 2302 2303 PetscErrorCode MatGetRowMax_MPIAIJ(Mat A, Vec v, PetscInt idx[]) 2304 { 2305 Mat_MPIAIJ *mat = (Mat_MPIAIJ*) A->data; 2306 PetscInt n = A->rmap->n; 2307 PetscInt cstart = A->cmap->rstart; 2308 PetscInt *cmap = mat->garray; 2309 PetscInt *diagIdx, *offdiagIdx; 2310 Vec diagV, offdiagV; 2311 PetscScalar *a, *diagA, *offdiagA; 2312 PetscInt r; 2313 PetscErrorCode ierr; 2314 2315 PetscFunctionBegin; 2316 ierr = PetscMalloc2(n,&diagIdx,n,&offdiagIdx);CHKERRQ(ierr); 2317 ierr = VecCreateSeq(PETSC_COMM_SELF, n, &diagV);CHKERRQ(ierr); 2318 ierr = VecCreateSeq(PETSC_COMM_SELF, n, &offdiagV);CHKERRQ(ierr); 2319 ierr = MatGetRowMax(mat->A, diagV, diagIdx);CHKERRQ(ierr); 2320 ierr = MatGetRowMax(mat->B, offdiagV, offdiagIdx);CHKERRQ(ierr); 2321 ierr = VecGetArray(v, &a);CHKERRQ(ierr); 2322 ierr = VecGetArray(diagV, &diagA);CHKERRQ(ierr); 2323 ierr = VecGetArray(offdiagV, &offdiagA);CHKERRQ(ierr); 2324 for (r = 0; r < n; ++r) { 2325 if (PetscAbsScalar(diagA[r]) >= PetscAbsScalar(offdiagA[r])) { 2326 a[r] = diagA[r]; 2327 idx[r] = cstart + diagIdx[r]; 2328 } else { 2329 a[r] = offdiagA[r]; 2330 idx[r] = cmap[offdiagIdx[r]]; 2331 } 2332 } 2333 ierr = VecRestoreArray(v, &a);CHKERRQ(ierr); 2334 ierr = VecRestoreArray(diagV, &diagA);CHKERRQ(ierr); 2335 ierr = VecRestoreArray(offdiagV, &offdiagA);CHKERRQ(ierr); 2336 ierr = VecDestroy(&diagV);CHKERRQ(ierr); 2337 ierr = VecDestroy(&offdiagV);CHKERRQ(ierr); 2338 ierr = PetscFree2(diagIdx, offdiagIdx);CHKERRQ(ierr); 2339 PetscFunctionReturn(0); 2340 } 2341 2342 PetscErrorCode MatGetSeqNonzeroStructure_MPIAIJ(Mat mat,Mat *newmat) 2343 { 2344 PetscErrorCode ierr; 2345 Mat *dummy; 2346 2347 PetscFunctionBegin; 2348 ierr = MatCreateSubMatrix_MPIAIJ_All(mat,MAT_DO_NOT_GET_VALUES,MAT_INITIAL_MATRIX,&dummy);CHKERRQ(ierr); 2349 *newmat = *dummy; 2350 ierr = PetscFree(dummy);CHKERRQ(ierr); 2351 PetscFunctionReturn(0); 2352 } 2353 2354 PetscErrorCode MatInvertBlockDiagonal_MPIAIJ(Mat A,const PetscScalar **values) 2355 { 2356 Mat_MPIAIJ *a = (Mat_MPIAIJ*) A->data; 2357 PetscErrorCode ierr; 2358 2359 PetscFunctionBegin; 2360 ierr = MatInvertBlockDiagonal(a->A,values);CHKERRQ(ierr); 2361 A->factorerrortype = a->A->factorerrortype; 2362 PetscFunctionReturn(0); 2363 } 2364 2365 static PetscErrorCode MatSetRandom_MPIAIJ(Mat x,PetscRandom rctx) 2366 { 2367 PetscErrorCode ierr; 2368 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)x->data; 2369 2370 PetscFunctionBegin; 2371 ierr = MatSetRandom(aij->A,rctx);CHKERRQ(ierr); 2372 ierr = MatSetRandom(aij->B,rctx);CHKERRQ(ierr); 2373 ierr = MatAssemblyBegin(x,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2374 ierr = MatAssemblyEnd(x,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2375 PetscFunctionReturn(0); 2376 } 2377 2378 PetscErrorCode MatMPIAIJSetUseScalableIncreaseOverlap_MPIAIJ(Mat A,PetscBool sc) 2379 { 2380 PetscFunctionBegin; 2381 if (sc) A->ops->increaseoverlap = MatIncreaseOverlap_MPIAIJ_Scalable; 2382 else A->ops->increaseoverlap = MatIncreaseOverlap_MPIAIJ; 2383 PetscFunctionReturn(0); 2384 } 2385 2386 /*@ 2387 MatMPIAIJSetUseScalableIncreaseOverlap - Determine if the matrix uses a scalable algorithm to compute the overlap 2388 2389 Collective on Mat 2390 2391 Input Parameters: 2392 + A - the matrix 2393 - sc - PETSC_TRUE indicates use the scalable algorithm (default is not to use the scalable algorithm) 2394 2395 Level: advanced 2396 2397 @*/ 2398 PetscErrorCode MatMPIAIJSetUseScalableIncreaseOverlap(Mat A,PetscBool sc) 2399 { 2400 PetscErrorCode ierr; 2401 2402 PetscFunctionBegin; 2403 ierr = PetscTryMethod(A,"MatMPIAIJSetUseScalableIncreaseOverlap_C",(Mat,PetscBool),(A,sc));CHKERRQ(ierr); 2404 PetscFunctionReturn(0); 2405 } 2406 2407 PetscErrorCode MatSetFromOptions_MPIAIJ(PetscOptionItems *PetscOptionsObject,Mat A) 2408 { 2409 PetscErrorCode ierr; 2410 PetscBool sc = PETSC_FALSE,flg; 2411 2412 PetscFunctionBegin; 2413 ierr = PetscOptionsHead(PetscOptionsObject,"MPIAIJ options");CHKERRQ(ierr); 2414 ierr = PetscObjectOptionsBegin((PetscObject)A); 2415 if (A->ops->increaseoverlap == MatIncreaseOverlap_MPIAIJ_Scalable) sc = PETSC_TRUE; 2416 ierr = PetscOptionsBool("-mat_increase_overlap_scalable","Use a scalable algorithm to compute the overlap","MatIncreaseOverlap",sc,&sc,&flg);CHKERRQ(ierr); 2417 if (flg) { 2418 ierr = MatMPIAIJSetUseScalableIncreaseOverlap(A,sc);CHKERRQ(ierr); 2419 } 2420 ierr = PetscOptionsEnd();CHKERRQ(ierr); 2421 PetscFunctionReturn(0); 2422 } 2423 2424 PetscErrorCode MatShift_MPIAIJ(Mat Y,PetscScalar a) 2425 { 2426 PetscErrorCode ierr; 2427 Mat_MPIAIJ *maij = (Mat_MPIAIJ*)Y->data; 2428 Mat_SeqAIJ *aij = (Mat_SeqAIJ*)maij->A->data; 2429 2430 PetscFunctionBegin; 2431 if (!Y->preallocated) { 2432 ierr = MatMPIAIJSetPreallocation(Y,1,NULL,0,NULL);CHKERRQ(ierr); 2433 } else if (!aij->nz) { 2434 PetscInt nonew = aij->nonew; 2435 ierr = MatSeqAIJSetPreallocation(maij->A,1,NULL);CHKERRQ(ierr); 2436 aij->nonew = nonew; 2437 } 2438 ierr = MatShift_Basic(Y,a);CHKERRQ(ierr); 2439 PetscFunctionReturn(0); 2440 } 2441 2442 PetscErrorCode MatMissingDiagonal_MPIAIJ(Mat A,PetscBool *missing,PetscInt *d) 2443 { 2444 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 2445 PetscErrorCode ierr; 2446 2447 PetscFunctionBegin; 2448 if (A->rmap->n != A->cmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Only works for square matrices"); 2449 ierr = MatMissingDiagonal(a->A,missing,d);CHKERRQ(ierr); 2450 if (d) { 2451 PetscInt rstart; 2452 ierr = MatGetOwnershipRange(A,&rstart,NULL);CHKERRQ(ierr); 2453 *d += rstart; 2454 2455 } 2456 PetscFunctionReturn(0); 2457 } 2458 2459 2460 /* -------------------------------------------------------------------*/ 2461 static struct _MatOps MatOps_Values = {MatSetValues_MPIAIJ, 2462 MatGetRow_MPIAIJ, 2463 MatRestoreRow_MPIAIJ, 2464 MatMult_MPIAIJ, 2465 /* 4*/ MatMultAdd_MPIAIJ, 2466 MatMultTranspose_MPIAIJ, 2467 MatMultTransposeAdd_MPIAIJ, 2468 0, 2469 0, 2470 0, 2471 /*10*/ 0, 2472 0, 2473 0, 2474 MatSOR_MPIAIJ, 2475 MatTranspose_MPIAIJ, 2476 /*15*/ MatGetInfo_MPIAIJ, 2477 MatEqual_MPIAIJ, 2478 MatGetDiagonal_MPIAIJ, 2479 MatDiagonalScale_MPIAIJ, 2480 MatNorm_MPIAIJ, 2481 /*20*/ MatAssemblyBegin_MPIAIJ, 2482 MatAssemblyEnd_MPIAIJ, 2483 MatSetOption_MPIAIJ, 2484 MatZeroEntries_MPIAIJ, 2485 /*24*/ MatZeroRows_MPIAIJ, 2486 0, 2487 0, 2488 0, 2489 0, 2490 /*29*/ MatSetUp_MPIAIJ, 2491 0, 2492 0, 2493 MatGetDiagonalBlock_MPIAIJ, 2494 0, 2495 /*34*/ MatDuplicate_MPIAIJ, 2496 0, 2497 0, 2498 0, 2499 0, 2500 /*39*/ MatAXPY_MPIAIJ, 2501 MatCreateSubMatrices_MPIAIJ, 2502 MatIncreaseOverlap_MPIAIJ, 2503 MatGetValues_MPIAIJ, 2504 MatCopy_MPIAIJ, 2505 /*44*/ MatGetRowMax_MPIAIJ, 2506 MatScale_MPIAIJ, 2507 MatShift_MPIAIJ, 2508 MatDiagonalSet_MPIAIJ, 2509 MatZeroRowsColumns_MPIAIJ, 2510 /*49*/ MatSetRandom_MPIAIJ, 2511 0, 2512 0, 2513 0, 2514 0, 2515 /*54*/ MatFDColoringCreate_MPIXAIJ, 2516 0, 2517 MatSetUnfactored_MPIAIJ, 2518 MatPermute_MPIAIJ, 2519 0, 2520 /*59*/ MatCreateSubMatrix_MPIAIJ, 2521 MatDestroy_MPIAIJ, 2522 MatView_MPIAIJ, 2523 0, 2524 MatMatMatMult_MPIAIJ_MPIAIJ_MPIAIJ, 2525 /*64*/ MatMatMatMultSymbolic_MPIAIJ_MPIAIJ_MPIAIJ, 2526 MatMatMatMultNumeric_MPIAIJ_MPIAIJ_MPIAIJ, 2527 0, 2528 0, 2529 0, 2530 /*69*/ MatGetRowMaxAbs_MPIAIJ, 2531 MatGetRowMinAbs_MPIAIJ, 2532 0, 2533 0, 2534 0, 2535 0, 2536 /*75*/ MatFDColoringApply_AIJ, 2537 MatSetFromOptions_MPIAIJ, 2538 0, 2539 0, 2540 MatFindZeroDiagonals_MPIAIJ, 2541 /*80*/ 0, 2542 0, 2543 0, 2544 /*83*/ MatLoad_MPIAIJ, 2545 0, 2546 0, 2547 0, 2548 0, 2549 0, 2550 /*89*/ MatMatMult_MPIAIJ_MPIAIJ, 2551 MatMatMultSymbolic_MPIAIJ_MPIAIJ, 2552 MatMatMultNumeric_MPIAIJ_MPIAIJ, 2553 MatPtAP_MPIAIJ_MPIAIJ, 2554 MatPtAPSymbolic_MPIAIJ_MPIAIJ, 2555 /*94*/ MatPtAPNumeric_MPIAIJ_MPIAIJ, 2556 0, 2557 0, 2558 0, 2559 0, 2560 /*99*/ 0, 2561 0, 2562 0, 2563 MatConjugate_MPIAIJ, 2564 0, 2565 /*104*/MatSetValuesRow_MPIAIJ, 2566 MatRealPart_MPIAIJ, 2567 MatImaginaryPart_MPIAIJ, 2568 0, 2569 0, 2570 /*109*/0, 2571 0, 2572 MatGetRowMin_MPIAIJ, 2573 0, 2574 MatMissingDiagonal_MPIAIJ, 2575 /*114*/MatGetSeqNonzeroStructure_MPIAIJ, 2576 0, 2577 MatGetGhosts_MPIAIJ, 2578 0, 2579 0, 2580 /*119*/0, 2581 0, 2582 0, 2583 0, 2584 MatGetMultiProcBlock_MPIAIJ, 2585 /*124*/MatFindNonzeroRows_MPIAIJ, 2586 MatGetColumnNorms_MPIAIJ, 2587 MatInvertBlockDiagonal_MPIAIJ, 2588 0, 2589 MatCreateSubMatricesMPI_MPIAIJ, 2590 /*129*/0, 2591 MatTransposeMatMult_MPIAIJ_MPIAIJ, 2592 MatTransposeMatMultSymbolic_MPIAIJ_MPIAIJ, 2593 MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ, 2594 0, 2595 /*134*/0, 2596 0, 2597 0, 2598 0, 2599 0, 2600 /*139*/MatSetBlockSizes_MPIAIJ, 2601 0, 2602 0, 2603 MatFDColoringSetUp_MPIXAIJ, 2604 MatFindOffBlockDiagonalEntries_MPIAIJ, 2605 /*144*/MatCreateMPIMatConcatenateSeqMat_MPIAIJ 2606 }; 2607 2608 /* ----------------------------------------------------------------------------------------*/ 2609 2610 PetscErrorCode MatStoreValues_MPIAIJ(Mat mat) 2611 { 2612 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 2613 PetscErrorCode ierr; 2614 2615 PetscFunctionBegin; 2616 ierr = MatStoreValues(aij->A);CHKERRQ(ierr); 2617 ierr = MatStoreValues(aij->B);CHKERRQ(ierr); 2618 PetscFunctionReturn(0); 2619 } 2620 2621 PetscErrorCode MatRetrieveValues_MPIAIJ(Mat mat) 2622 { 2623 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 2624 PetscErrorCode ierr; 2625 2626 PetscFunctionBegin; 2627 ierr = MatRetrieveValues(aij->A);CHKERRQ(ierr); 2628 ierr = MatRetrieveValues(aij->B);CHKERRQ(ierr); 2629 PetscFunctionReturn(0); 2630 } 2631 2632 PetscErrorCode MatMPIAIJSetPreallocation_MPIAIJ(Mat B,PetscInt d_nz,const PetscInt d_nnz[],PetscInt o_nz,const PetscInt o_nnz[]) 2633 { 2634 Mat_MPIAIJ *b; 2635 PetscErrorCode ierr; 2636 2637 PetscFunctionBegin; 2638 ierr = PetscLayoutSetUp(B->rmap);CHKERRQ(ierr); 2639 ierr = PetscLayoutSetUp(B->cmap);CHKERRQ(ierr); 2640 b = (Mat_MPIAIJ*)B->data; 2641 2642 #if defined(PETSC_USE_CTABLE) 2643 ierr = PetscTableDestroy(&b->colmap);CHKERRQ(ierr); 2644 #else 2645 ierr = PetscFree(b->colmap);CHKERRQ(ierr); 2646 #endif 2647 ierr = PetscFree(b->garray);CHKERRQ(ierr); 2648 ierr = VecDestroy(&b->lvec);CHKERRQ(ierr); 2649 ierr = VecScatterDestroy(&b->Mvctx);CHKERRQ(ierr); 2650 2651 /* Because the B will have been resized we simply destroy it and create a new one each time */ 2652 ierr = MatDestroy(&b->B);CHKERRQ(ierr); 2653 ierr = MatCreate(PETSC_COMM_SELF,&b->B);CHKERRQ(ierr); 2654 ierr = MatSetSizes(b->B,B->rmap->n,B->cmap->N,B->rmap->n,B->cmap->N);CHKERRQ(ierr); 2655 ierr = MatSetBlockSizesFromMats(b->B,B,B);CHKERRQ(ierr); 2656 ierr = MatSetType(b->B,MATSEQAIJ);CHKERRQ(ierr); 2657 ierr = PetscLogObjectParent((PetscObject)B,(PetscObject)b->B);CHKERRQ(ierr); 2658 2659 if (!B->preallocated) { 2660 ierr = MatCreate(PETSC_COMM_SELF,&b->A);CHKERRQ(ierr); 2661 ierr = MatSetSizes(b->A,B->rmap->n,B->cmap->n,B->rmap->n,B->cmap->n);CHKERRQ(ierr); 2662 ierr = MatSetBlockSizesFromMats(b->A,B,B);CHKERRQ(ierr); 2663 ierr = MatSetType(b->A,MATSEQAIJ);CHKERRQ(ierr); 2664 ierr = PetscLogObjectParent((PetscObject)B,(PetscObject)b->A);CHKERRQ(ierr); 2665 } 2666 2667 ierr = MatSeqAIJSetPreallocation(b->A,d_nz,d_nnz);CHKERRQ(ierr); 2668 ierr = MatSeqAIJSetPreallocation(b->B,o_nz,o_nnz);CHKERRQ(ierr); 2669 B->preallocated = PETSC_TRUE; 2670 B->was_assembled = PETSC_FALSE; 2671 B->assembled = PETSC_FALSE;; 2672 PetscFunctionReturn(0); 2673 } 2674 2675 PetscErrorCode MatDuplicate_MPIAIJ(Mat matin,MatDuplicateOption cpvalues,Mat *newmat) 2676 { 2677 Mat mat; 2678 Mat_MPIAIJ *a,*oldmat = (Mat_MPIAIJ*)matin->data; 2679 PetscErrorCode ierr; 2680 2681 PetscFunctionBegin; 2682 *newmat = 0; 2683 ierr = MatCreate(PetscObjectComm((PetscObject)matin),&mat);CHKERRQ(ierr); 2684 ierr = MatSetSizes(mat,matin->rmap->n,matin->cmap->n,matin->rmap->N,matin->cmap->N);CHKERRQ(ierr); 2685 ierr = MatSetBlockSizesFromMats(mat,matin,matin);CHKERRQ(ierr); 2686 ierr = MatSetType(mat,((PetscObject)matin)->type_name);CHKERRQ(ierr); 2687 ierr = PetscMemcpy(mat->ops,matin->ops,sizeof(struct _MatOps));CHKERRQ(ierr); 2688 a = (Mat_MPIAIJ*)mat->data; 2689 2690 mat->factortype = matin->factortype; 2691 mat->assembled = PETSC_TRUE; 2692 mat->insertmode = NOT_SET_VALUES; 2693 mat->preallocated = PETSC_TRUE; 2694 2695 a->size = oldmat->size; 2696 a->rank = oldmat->rank; 2697 a->donotstash = oldmat->donotstash; 2698 a->roworiented = oldmat->roworiented; 2699 a->rowindices = 0; 2700 a->rowvalues = 0; 2701 a->getrowactive = PETSC_FALSE; 2702 2703 ierr = PetscLayoutReference(matin->rmap,&mat->rmap);CHKERRQ(ierr); 2704 ierr = PetscLayoutReference(matin->cmap,&mat->cmap);CHKERRQ(ierr); 2705 2706 if (oldmat->colmap) { 2707 #if defined(PETSC_USE_CTABLE) 2708 ierr = PetscTableCreateCopy(oldmat->colmap,&a->colmap);CHKERRQ(ierr); 2709 #else 2710 ierr = PetscMalloc1(mat->cmap->N,&a->colmap);CHKERRQ(ierr); 2711 ierr = PetscLogObjectMemory((PetscObject)mat,(mat->cmap->N)*sizeof(PetscInt));CHKERRQ(ierr); 2712 ierr = PetscMemcpy(a->colmap,oldmat->colmap,(mat->cmap->N)*sizeof(PetscInt));CHKERRQ(ierr); 2713 #endif 2714 } else a->colmap = 0; 2715 if (oldmat->garray) { 2716 PetscInt len; 2717 len = oldmat->B->cmap->n; 2718 ierr = PetscMalloc1(len+1,&a->garray);CHKERRQ(ierr); 2719 ierr = PetscLogObjectMemory((PetscObject)mat,len*sizeof(PetscInt));CHKERRQ(ierr); 2720 if (len) { ierr = PetscMemcpy(a->garray,oldmat->garray,len*sizeof(PetscInt));CHKERRQ(ierr); } 2721 } else a->garray = 0; 2722 2723 ierr = VecDuplicate(oldmat->lvec,&a->lvec);CHKERRQ(ierr); 2724 ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)a->lvec);CHKERRQ(ierr); 2725 ierr = VecScatterCopy(oldmat->Mvctx,&a->Mvctx);CHKERRQ(ierr); 2726 ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)a->Mvctx);CHKERRQ(ierr); 2727 ierr = MatDuplicate(oldmat->A,cpvalues,&a->A);CHKERRQ(ierr); 2728 ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)a->A);CHKERRQ(ierr); 2729 ierr = MatDuplicate(oldmat->B,cpvalues,&a->B);CHKERRQ(ierr); 2730 ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)a->B);CHKERRQ(ierr); 2731 ierr = PetscFunctionListDuplicate(((PetscObject)matin)->qlist,&((PetscObject)mat)->qlist);CHKERRQ(ierr); 2732 *newmat = mat; 2733 PetscFunctionReturn(0); 2734 } 2735 2736 PetscErrorCode MatLoad_MPIAIJ(Mat newMat, PetscViewer viewer) 2737 { 2738 PetscScalar *vals,*svals; 2739 MPI_Comm comm; 2740 PetscErrorCode ierr; 2741 PetscMPIInt rank,size,tag = ((PetscObject)viewer)->tag; 2742 PetscInt i,nz,j,rstart,rend,mmax,maxnz = 0; 2743 PetscInt header[4],*rowlengths = 0,M,N,m,*cols; 2744 PetscInt *ourlens = NULL,*procsnz = NULL,*offlens = NULL,jj,*mycols,*smycols; 2745 PetscInt cend,cstart,n,*rowners; 2746 int fd; 2747 PetscInt bs = newMat->rmap->bs; 2748 2749 PetscFunctionBegin; 2750 /* force binary viewer to load .info file if it has not yet done so */ 2751 ierr = PetscViewerSetUp(viewer);CHKERRQ(ierr); 2752 ierr = PetscObjectGetComm((PetscObject)viewer,&comm);CHKERRQ(ierr); 2753 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 2754 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 2755 ierr = PetscViewerBinaryGetDescriptor(viewer,&fd);CHKERRQ(ierr); 2756 if (!rank) { 2757 ierr = PetscBinaryRead(fd,(char*)header,4,PETSC_INT);CHKERRQ(ierr); 2758 if (header[0] != MAT_FILE_CLASSID) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED,"not matrix object"); 2759 if (header[3] < 0) SETERRQ(PetscObjectComm((PetscObject)newMat),PETSC_ERR_FILE_UNEXPECTED,"Matrix stored in special format on disk,cannot load as MATMPIAIJ"); 2760 } 2761 2762 ierr = PetscOptionsBegin(comm,NULL,"Options for loading MATMPIAIJ matrix","Mat");CHKERRQ(ierr); 2763 ierr = PetscOptionsInt("-matload_block_size","Set the blocksize used to store the matrix","MatLoad",bs,&bs,NULL);CHKERRQ(ierr); 2764 ierr = PetscOptionsEnd();CHKERRQ(ierr); 2765 if (bs < 0) bs = 1; 2766 2767 ierr = MPI_Bcast(header+1,3,MPIU_INT,0,comm);CHKERRQ(ierr); 2768 M = header[1]; N = header[2]; 2769 2770 /* If global sizes are set, check if they are consistent with that given in the file */ 2771 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); 2772 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); 2773 2774 /* determine ownership of all (block) rows */ 2775 if (M%bs) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED, "Inconsistent # of rows (%d) and block size (%d)",M,bs); 2776 if (newMat->rmap->n < 0) m = bs*((M/bs)/size + (((M/bs) % size) > rank)); /* PETSC_DECIDE */ 2777 else m = newMat->rmap->n; /* Set by user */ 2778 2779 ierr = PetscMalloc1(size+1,&rowners);CHKERRQ(ierr); 2780 ierr = MPI_Allgather(&m,1,MPIU_INT,rowners+1,1,MPIU_INT,comm);CHKERRQ(ierr); 2781 2782 /* First process needs enough room for process with most rows */ 2783 if (!rank) { 2784 mmax = rowners[1]; 2785 for (i=2; i<=size; i++) { 2786 mmax = PetscMax(mmax, rowners[i]); 2787 } 2788 } else mmax = -1; /* unused, but compilers complain */ 2789 2790 rowners[0] = 0; 2791 for (i=2; i<=size; i++) { 2792 rowners[i] += rowners[i-1]; 2793 } 2794 rstart = rowners[rank]; 2795 rend = rowners[rank+1]; 2796 2797 /* distribute row lengths to all processors */ 2798 ierr = PetscMalloc2(m,&ourlens,m,&offlens);CHKERRQ(ierr); 2799 if (!rank) { 2800 ierr = PetscBinaryRead(fd,ourlens,m,PETSC_INT);CHKERRQ(ierr); 2801 ierr = PetscMalloc1(mmax,&rowlengths);CHKERRQ(ierr); 2802 ierr = PetscCalloc1(size,&procsnz);CHKERRQ(ierr); 2803 for (j=0; j<m; j++) { 2804 procsnz[0] += ourlens[j]; 2805 } 2806 for (i=1; i<size; i++) { 2807 ierr = PetscBinaryRead(fd,rowlengths,rowners[i+1]-rowners[i],PETSC_INT);CHKERRQ(ierr); 2808 /* calculate the number of nonzeros on each processor */ 2809 for (j=0; j<rowners[i+1]-rowners[i]; j++) { 2810 procsnz[i] += rowlengths[j]; 2811 } 2812 ierr = MPIULong_Send(rowlengths,rowners[i+1]-rowners[i],MPIU_INT,i,tag,comm);CHKERRQ(ierr); 2813 } 2814 ierr = PetscFree(rowlengths);CHKERRQ(ierr); 2815 } else { 2816 ierr = MPIULong_Recv(ourlens,m,MPIU_INT,0,tag,comm);CHKERRQ(ierr); 2817 } 2818 2819 if (!rank) { 2820 /* determine max buffer needed and allocate it */ 2821 maxnz = 0; 2822 for (i=0; i<size; i++) { 2823 maxnz = PetscMax(maxnz,procsnz[i]); 2824 } 2825 ierr = PetscMalloc1(maxnz,&cols);CHKERRQ(ierr); 2826 2827 /* read in my part of the matrix column indices */ 2828 nz = procsnz[0]; 2829 ierr = PetscMalloc1(nz,&mycols);CHKERRQ(ierr); 2830 ierr = PetscBinaryRead(fd,mycols,nz,PETSC_INT);CHKERRQ(ierr); 2831 2832 /* read in every one elses and ship off */ 2833 for (i=1; i<size; i++) { 2834 nz = procsnz[i]; 2835 ierr = PetscBinaryRead(fd,cols,nz,PETSC_INT);CHKERRQ(ierr); 2836 ierr = MPIULong_Send(cols,nz,MPIU_INT,i,tag,comm);CHKERRQ(ierr); 2837 } 2838 ierr = PetscFree(cols);CHKERRQ(ierr); 2839 } else { 2840 /* determine buffer space needed for message */ 2841 nz = 0; 2842 for (i=0; i<m; i++) { 2843 nz += ourlens[i]; 2844 } 2845 ierr = PetscMalloc1(nz,&mycols);CHKERRQ(ierr); 2846 2847 /* receive message of column indices*/ 2848 ierr = MPIULong_Recv(mycols,nz,MPIU_INT,0,tag,comm);CHKERRQ(ierr); 2849 } 2850 2851 /* determine column ownership if matrix is not square */ 2852 if (N != M) { 2853 if (newMat->cmap->n < 0) n = N/size + ((N % size) > rank); 2854 else n = newMat->cmap->n; 2855 ierr = MPI_Scan(&n,&cend,1,MPIU_INT,MPI_SUM,comm);CHKERRQ(ierr); 2856 cstart = cend - n; 2857 } else { 2858 cstart = rstart; 2859 cend = rend; 2860 n = cend - cstart; 2861 } 2862 2863 /* loop over local rows, determining number of off diagonal entries */ 2864 ierr = PetscMemzero(offlens,m*sizeof(PetscInt));CHKERRQ(ierr); 2865 jj = 0; 2866 for (i=0; i<m; i++) { 2867 for (j=0; j<ourlens[i]; j++) { 2868 if (mycols[jj] < cstart || mycols[jj] >= cend) offlens[i]++; 2869 jj++; 2870 } 2871 } 2872 2873 for (i=0; i<m; i++) { 2874 ourlens[i] -= offlens[i]; 2875 } 2876 ierr = MatSetSizes(newMat,m,n,M,N);CHKERRQ(ierr); 2877 2878 if (bs > 1) {ierr = MatSetBlockSize(newMat,bs);CHKERRQ(ierr);} 2879 2880 ierr = MatMPIAIJSetPreallocation(newMat,0,ourlens,0,offlens);CHKERRQ(ierr); 2881 2882 for (i=0; i<m; i++) { 2883 ourlens[i] += offlens[i]; 2884 } 2885 2886 if (!rank) { 2887 ierr = PetscMalloc1(maxnz+1,&vals);CHKERRQ(ierr); 2888 2889 /* read in my part of the matrix numerical values */ 2890 nz = procsnz[0]; 2891 ierr = PetscBinaryRead(fd,vals,nz,PETSC_SCALAR);CHKERRQ(ierr); 2892 2893 /* insert into matrix */ 2894 jj = rstart; 2895 smycols = mycols; 2896 svals = vals; 2897 for (i=0; i<m; i++) { 2898 ierr = MatSetValues_MPIAIJ(newMat,1,&jj,ourlens[i],smycols,svals,INSERT_VALUES);CHKERRQ(ierr); 2899 smycols += ourlens[i]; 2900 svals += ourlens[i]; 2901 jj++; 2902 } 2903 2904 /* read in other processors and ship out */ 2905 for (i=1; i<size; i++) { 2906 nz = procsnz[i]; 2907 ierr = PetscBinaryRead(fd,vals,nz,PETSC_SCALAR);CHKERRQ(ierr); 2908 ierr = MPIULong_Send(vals,nz,MPIU_SCALAR,i,((PetscObject)newMat)->tag,comm);CHKERRQ(ierr); 2909 } 2910 ierr = PetscFree(procsnz);CHKERRQ(ierr); 2911 } else { 2912 /* receive numeric values */ 2913 ierr = PetscMalloc1(nz+1,&vals);CHKERRQ(ierr); 2914 2915 /* receive message of values*/ 2916 ierr = MPIULong_Recv(vals,nz,MPIU_SCALAR,0,((PetscObject)newMat)->tag,comm);CHKERRQ(ierr); 2917 2918 /* insert into matrix */ 2919 jj = rstart; 2920 smycols = mycols; 2921 svals = vals; 2922 for (i=0; i<m; i++) { 2923 ierr = MatSetValues_MPIAIJ(newMat,1,&jj,ourlens[i],smycols,svals,INSERT_VALUES);CHKERRQ(ierr); 2924 smycols += ourlens[i]; 2925 svals += ourlens[i]; 2926 jj++; 2927 } 2928 } 2929 ierr = PetscFree2(ourlens,offlens);CHKERRQ(ierr); 2930 ierr = PetscFree(vals);CHKERRQ(ierr); 2931 ierr = PetscFree(mycols);CHKERRQ(ierr); 2932 ierr = PetscFree(rowners);CHKERRQ(ierr); 2933 ierr = MatAssemblyBegin(newMat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2934 ierr = MatAssemblyEnd(newMat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2935 PetscFunctionReturn(0); 2936 } 2937 2938 /* Not scalable because of ISAllGather() unless getting all columns. */ 2939 PetscErrorCode ISGetSeqIS_Private(Mat mat,IS iscol,IS *isseq) 2940 { 2941 PetscErrorCode ierr; 2942 IS iscol_local; 2943 PetscBool isstride; 2944 PetscMPIInt lisstride=0,gisstride; 2945 2946 PetscFunctionBegin; 2947 /* check if we are grabbing all columns*/ 2948 ierr = PetscObjectTypeCompare((PetscObject)iscol,ISSTRIDE,&isstride);CHKERRQ(ierr); 2949 2950 if (isstride) { 2951 PetscInt start,len,mstart,mlen; 2952 ierr = ISStrideGetInfo(iscol,&start,NULL);CHKERRQ(ierr); 2953 ierr = ISGetLocalSize(iscol,&len);CHKERRQ(ierr); 2954 ierr = MatGetOwnershipRangeColumn(mat,&mstart,&mlen);CHKERRQ(ierr); 2955 if (mstart == start && mlen-mstart == len) lisstride = 1; 2956 } 2957 2958 ierr = MPIU_Allreduce(&lisstride,&gisstride,1,MPI_INT,MPI_MIN,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 2959 if (gisstride) { 2960 PetscInt N; 2961 ierr = MatGetSize(mat,NULL,&N);CHKERRQ(ierr); 2962 ierr = ISCreateStride(PetscObjectComm((PetscObject)mat),N,0,1,&iscol_local);CHKERRQ(ierr); 2963 ierr = ISSetIdentity(iscol_local);CHKERRQ(ierr); 2964 ierr = PetscInfo(mat,"Optimizing for obtaining all columns of the matrix; skipping ISAllGather()\n");CHKERRQ(ierr); 2965 } else { 2966 PetscInt cbs; 2967 ierr = ISGetBlockSize(iscol,&cbs);CHKERRQ(ierr); 2968 ierr = ISAllGather(iscol,&iscol_local);CHKERRQ(ierr); 2969 ierr = ISSetBlockSize(iscol_local,cbs);CHKERRQ(ierr); 2970 } 2971 2972 *isseq = iscol_local; 2973 PetscFunctionReturn(0); 2974 } 2975 2976 /* 2977 Used for avoiding ISAllGather() and global size of iscol_local (see MatCreateSubMatrix_MPIAIJ_nonscalable) 2978 2979 Input Parameters: 2980 mat - matrix 2981 iscol - parallel column index set; its local indices are a subset of local columns of mat, 2982 i.e., mat->cstart <= iscol[i] < mat->cend 2983 Output Parameter: 2984 iscol_sub - sequential column index set, a subset of iscol 2985 iscmap - column map; iscmap[i] indicates global location of iscol_sub[i] in iscol 2986 */ 2987 PetscErrorCode ISGetSeqIS_SameColDist_Private(Mat mat,IS isrow,IS iscol,IS *isrow_d,IS *iscol_d,IS *iscol_o,IS *isgarray) 2988 { 2989 PetscErrorCode ierr; 2990 Vec x,cmap; 2991 const PetscInt *is_idx; 2992 PetscScalar *xarray,*cmaparray; 2993 PetscInt ncols,isstart,*idx,*camp,count,m,rstart; 2994 Mat_MPIAIJ *a=(Mat_MPIAIJ*)mat->data; 2995 Mat B=a->B; 2996 Vec lvec=a->lvec,lcmap; 2997 PetscInt i,cstart,cend,Bn=B->cmap->N; 2998 MPI_Comm comm; 2999 3000 PetscFunctionBegin; 3001 ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 3002 PetscMPIInt rank; 3003 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 3004 ierr = ISGetLocalSize(iscol,&ncols);CHKERRQ(ierr); 3005 3006 /* (1) iscol is a sub-column vector of mat, pad it with '-1.' to form a full vector x */ 3007 ierr = MatCreateVecs(mat,&x,NULL);CHKERRQ(ierr); 3008 ierr = VecDuplicate(x,&cmap);CHKERRQ(ierr); 3009 ierr = VecSet(x,-1.0);CHKERRQ(ierr); 3010 3011 /* get start indices */ 3012 ierr = MPI_Scan(&ncols,&isstart,1,MPIU_INT,MPI_SUM,comm);CHKERRQ(ierr); 3013 isstart -= ncols; 3014 ierr = MatGetOwnershipRangeColumn(mat,&cstart,&cend);CHKERRQ(ierr); 3015 3016 ierr = ISGetIndices(iscol,&is_idx);CHKERRQ(ierr); 3017 ierr = VecGetArray(x,&xarray);CHKERRQ(ierr); 3018 ierr = VecGetArray(cmap,&cmaparray);CHKERRQ(ierr); 3019 ierr = PetscMalloc1(ncols,&idx);CHKERRQ(ierr); 3020 for (i=0; i<ncols; i++) { 3021 xarray[is_idx[i]-cstart] = (PetscScalar)is_idx[i]; 3022 cmaparray[is_idx[i]-cstart] = i + isstart; /* global index of iscol[i] */ 3023 3024 idx[i] = is_idx[i]-cstart; 3025 } 3026 ierr = VecRestoreArray(x,&xarray);CHKERRQ(ierr); 3027 ierr = VecRestoreArray(cmap,&cmaparray);CHKERRQ(ierr); 3028 ierr = ISRestoreIndices(iscol,&is_idx);CHKERRQ(ierr); 3029 3030 ierr = ISCreateGeneral(PETSC_COMM_SELF,ncols,idx,PETSC_COPY_VALUES,iscol_d);CHKERRQ(ierr); 3031 ierr = ISGetBlockSize(iscol,&i);CHKERRQ(ierr); 3032 ierr = ISSetBlockSize(*iscol_d,i);CHKERRQ(ierr); 3033 ierr = PetscFree(idx);CHKERRQ(ierr); 3034 3035 ierr = ISGetLocalSize(isrow,&m);CHKERRQ(ierr); 3036 rstart = mat->rmap->rstart; 3037 ierr = PetscMalloc1(m,&idx);CHKERRQ(ierr); 3038 ierr = ISGetIndices(isrow,&is_idx);CHKERRQ(ierr); 3039 for (i=0; i<m; i++) { 3040 idx[i] = is_idx[i]-rstart; 3041 } 3042 ierr = ISRestoreIndices(isrow,&is_idx);CHKERRQ(ierr); 3043 3044 ierr = ISCreateGeneral(PETSC_COMM_SELF,m,idx,PETSC_COPY_VALUES,isrow_d);CHKERRQ(ierr); 3045 ierr = ISGetBlockSize(isrow,&i);CHKERRQ(ierr); 3046 ierr = ISSetBlockSize(*isrow_d,i);CHKERRQ(ierr); 3047 ierr = PetscFree(idx);CHKERRQ(ierr); 3048 3049 /* (2) scatter x and cmap using aij->Mvctx to get their off-process portions (see MatMult_MPIAIJ) */ 3050 ierr = VecScatterBegin(a->Mvctx,x,lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3051 3052 ierr = VecDuplicate(lvec,&lcmap);CHKERRQ(ierr); 3053 3054 ierr = VecScatterEnd(a->Mvctx,x,lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3055 ierr = VecScatterBegin(a->Mvctx,cmap,lcmap,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3056 3057 count = ncols + Bn; 3058 ierr = PetscMalloc2(count,&idx,count,&camp);CHKERRQ(ierr); 3059 3060 ierr = VecScatterEnd(a->Mvctx,cmap,lcmap,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3061 3062 /* (3) create scalable iscol_sub (a subset of iscol) and iscmap */ 3063 /* local column indices */ 3064 ierr = ISGetIndices(iscol,&is_idx);CHKERRQ(ierr); 3065 for (i=0; i<ncols; i++) { 3066 idx[i] = is_idx[i]; 3067 camp[i] = i + isstart; /* global index of iscol = column index in submat */ 3068 } 3069 ierr = ISRestoreIndices(iscol,&is_idx);CHKERRQ(ierr); 3070 count = ncols; 3071 3072 /* off-process column indices */ 3073 PetscInt *idx1,*cmap1,count1=0; 3074 ierr = PetscMalloc2(Bn,&idx1,Bn,&cmap1);CHKERRQ(ierr); 3075 3076 ierr = VecGetArray(lvec,&xarray);CHKERRQ(ierr); 3077 ierr = VecGetArray(lcmap,&cmaparray);CHKERRQ(ierr); 3078 for (i=0; i<Bn; i++) { 3079 if (PetscRealPart(xarray[i]) > -1.0) { 3080 idx1[count1] = i; /* local column index in off-diagonal part B */ 3081 cmap1[count1++] = (PetscInt)PetscRealPart(cmaparray[i]); /* column index in submat */ 3082 3083 idx[count] = (PetscInt)PetscRealPart(xarray[i]); 3084 camp[count++] = (PetscInt)PetscRealPart(cmaparray[i]); /* column index in submat */ 3085 } 3086 } 3087 ierr = VecRestoreArray(lvec,&xarray);CHKERRQ(ierr); 3088 ierr = VecRestoreArray(lcmap,&cmaparray);CHKERRQ(ierr); 3089 3090 ierr = PetscSortIntWithArray(count,camp,idx);CHKERRQ(ierr); 3091 3092 ierr = ISCreateGeneral(PETSC_COMM_SELF,count1,idx1,PETSC_COPY_VALUES,iscol_o);CHKERRQ(ierr); 3093 ierr = ISCreateGeneral(PETSC_COMM_SELF,count1,cmap1,PETSC_COPY_VALUES,isgarray);CHKERRQ(ierr); 3094 ierr = ISGetBlockSize(iscol,&i);CHKERRQ(ierr); 3095 ierr = ISSetBlockSize(*iscol_o,i);CHKERRQ(ierr); 3096 ierr = PetscFree2(idx1,cmap1);CHKERRQ(ierr); 3097 3098 ierr = PetscFree2(idx,camp);CHKERRQ(ierr); 3099 ierr = VecDestroy(&x);CHKERRQ(ierr); 3100 ierr = VecDestroy(&cmap);CHKERRQ(ierr); 3101 ierr = VecDestroy(&lcmap);CHKERRQ(ierr); 3102 PetscFunctionReturn(0); 3103 } 3104 3105 PetscErrorCode MatCreateSubMatrix_MPIAIJ(Mat mat,IS isrow,IS iscol,MatReuse call,Mat *newmat) 3106 { 3107 PetscErrorCode ierr; 3108 IS iscol_local,isrow_d; 3109 PetscInt csize; 3110 PetscInt n,i,j,start,end; 3111 PetscBool sameRowDist=PETSC_FALSE,sameDist[2],tsameDist[2]; 3112 MPI_Comm comm; 3113 3114 PetscFunctionBegin; 3115 /* If isrow has same processor distribution as mat, 3116 call MatCreateSubMatrix_MPIAIJ_SameRowDist() to avoid using a hash table with global size of iscol */ 3117 if (call == MAT_REUSE_MATRIX) { 3118 ierr = PetscObjectQuery((PetscObject)*newmat,"isrow_d",(PetscObject*)&isrow_d);CHKERRQ(ierr); 3119 if (isrow_d) { 3120 sameRowDist = PETSC_TRUE; 3121 tsameDist[1] = PETSC_TRUE; /* sameColDist */ 3122 } else { 3123 ierr = PetscObjectQuery((PetscObject)*newmat,"SubIScol",(PetscObject*)&iscol_local);CHKERRQ(ierr); 3124 if (iscol_local) { 3125 sameRowDist = PETSC_TRUE; 3126 tsameDist[1] = PETSC_FALSE; /* !sameColDist */ 3127 } 3128 } 3129 } else { 3130 /* check if isrow has same processor distribution as mat */ 3131 sameDist[0] = PETSC_FALSE; 3132 ierr = ISGetLocalSize(isrow,&n);CHKERRQ(ierr); 3133 if (!n) { 3134 sameDist[0] = PETSC_TRUE; 3135 } else { 3136 ierr = ISGetMinMax(isrow,&i,&j);CHKERRQ(ierr); 3137 ierr = MatGetOwnershipRange(mat,&start,&end);CHKERRQ(ierr); 3138 if (i >= start && j < end) { 3139 sameDist[0] = PETSC_TRUE; 3140 } 3141 } 3142 3143 /* check if iscol has same processor distribution as mat */ 3144 sameDist[1] = PETSC_FALSE; 3145 ierr = ISGetLocalSize(iscol,&n);CHKERRQ(ierr); 3146 if (!n) { 3147 sameDist[1] = PETSC_TRUE; 3148 } else { 3149 ierr = ISGetMinMax(iscol,&i,&j);CHKERRQ(ierr); 3150 ierr = MatGetOwnershipRangeColumn(mat,&start,&end);CHKERRQ(ierr); 3151 if (i >= start && j < end) sameDist[1] = PETSC_TRUE; 3152 } 3153 3154 ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 3155 ierr = MPIU_Allreduce(&sameDist,&tsameDist,2,MPIU_BOOL,MPI_LAND,comm);CHKERRQ(ierr); 3156 sameRowDist = tsameDist[0]; 3157 } 3158 3159 if (sameRowDist) { 3160 ierr = MatCreateSubMatrix_MPIAIJ_SameRowDist(mat,isrow,iscol,call,tsameDist[1],newmat);CHKERRQ(ierr); 3161 PetscFunctionReturn(0); 3162 } 3163 3164 /* General case: iscol -> iscol_local which has global size of iscol */ 3165 if (call == MAT_REUSE_MATRIX) { 3166 ierr = PetscObjectQuery((PetscObject)*newmat,"ISAllGather",(PetscObject*)&iscol_local);CHKERRQ(ierr); 3167 if (!iscol_local) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Submatrix passed in was not used before, cannot reuse"); 3168 } else { 3169 ierr = ISGetSeqIS_Private(mat,iscol,&iscol_local);CHKERRQ(ierr); 3170 } 3171 3172 ierr = ISGetLocalSize(iscol,&csize);CHKERRQ(ierr); 3173 ierr = MatCreateSubMatrix_MPIAIJ_nonscalable(mat,isrow,iscol_local,csize,call,newmat);CHKERRQ(ierr); 3174 3175 if (call == MAT_INITIAL_MATRIX) { 3176 ierr = PetscObjectCompose((PetscObject)*newmat,"ISAllGather",(PetscObject)iscol_local);CHKERRQ(ierr); 3177 ierr = ISDestroy(&iscol_local);CHKERRQ(ierr); 3178 } 3179 PetscFunctionReturn(0); 3180 } 3181 3182 /*@C 3183 MatCreateMPIAIJWithSeqAIJ - creates a MPIAIJ matrix using SeqAIJ matrices that contain the "diagonal" 3184 and "off-diagonal" part of the matrix in CSR format. 3185 3186 Collective on MPI_Comm 3187 3188 Input Parameters: 3189 + comm - MPI communicator 3190 . A - "diagonal" portion of matrix 3191 . B - "off-diagonal" portion of matrix, destroyed by this routine 3192 - garray - global index of B columns 3193 3194 Output Parameter: 3195 . mat - the matrix, with input A as its local diagonal matrix 3196 Level: advanced 3197 3198 Notes: 3199 See MatCreateAIJ() for the definition of "diagonal" and "off-diagonal" portion of the matrix. 3200 A becomes part of output mat, B is destroyed by this routine. The user cannot use A and B anymore. 3201 3202 .seealso: MatCreateMPIAIJWithSplitArrays() 3203 @*/ 3204 PetscErrorCode MatCreateMPIAIJWithSeqAIJ(MPI_Comm comm,Mat A,Mat B,const PetscInt garray[],Mat *mat) 3205 { 3206 PetscErrorCode ierr; 3207 Mat_MPIAIJ *maij; 3208 Mat_SeqAIJ *b=(Mat_SeqAIJ*)B->data; 3209 PetscInt *oi=b->i,*oj=b->j,i,nz,col; 3210 PetscScalar *oa=b->a; 3211 Mat Btmp; 3212 PetscInt m,n,N; 3213 3214 PetscFunctionBegin; 3215 ierr = MatCreate(comm,mat);CHKERRQ(ierr); 3216 ierr = MatGetSize(A,&m,&n);CHKERRQ(ierr); 3217 if (m != B->rmap->N) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Am %d != Bm %d",m,B->rmap->N); 3218 3219 /* get global columns of mat */ 3220 ierr = MPIU_Allreduce(&n,&N,1,MPI_INT,MPI_SUM,comm);CHKERRQ(ierr); 3221 3222 ierr = MatSetSizes(*mat,m,n,PETSC_DECIDE,N);CHKERRQ(ierr); 3223 ierr = MatSetType(*mat,MATMPIAIJ);CHKERRQ(ierr); 3224 maij = (Mat_MPIAIJ*)(*mat)->data; 3225 3226 (*mat)->preallocated = PETSC_TRUE; 3227 3228 ierr = PetscLayoutSetUp((*mat)->rmap);CHKERRQ(ierr); 3229 ierr = PetscLayoutSetUp((*mat)->cmap);CHKERRQ(ierr); 3230 3231 maij->A = A; 3232 3233 nz = oi[m]; 3234 for (i=0; i<nz; i++) { 3235 col = oj[i]; 3236 oj[i] = garray[col]; 3237 } 3238 3239 ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,m,N,oi,oj,oa,&Btmp);CHKERRQ(ierr); 3240 maij->B = Btmp; 3241 3242 if (B->rmap->N != Btmp->rmap->N) SETERRQ2(PETSC_COMM_SELF,0,"BN %d != BtmpN %d",B->rmap->N,Btmp->rmap->N); 3243 3244 b->singlemalloc = PETSC_FALSE; /* these arrays are shared by Btmp */ 3245 b->free_a = PETSC_FALSE; 3246 b->free_ij = PETSC_FALSE; 3247 ierr = MatDestroy(&B);CHKERRQ(ierr); 3248 3249 b = (Mat_SeqAIJ*)Btmp->data; 3250 b->singlemalloc = PETSC_TRUE; /* arrays will be freed by MatDestroy(&Btmp) */ 3251 b->free_a = PETSC_TRUE; 3252 b->free_ij = PETSC_TRUE; 3253 3254 /* condense columns of maij->B */ 3255 ierr = MatSetOption(*mat,MAT_NO_OFF_PROC_ENTRIES,PETSC_TRUE);CHKERRQ(ierr); 3256 ierr = MatAssemblyBegin(*mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3257 ierr = MatAssemblyEnd(*mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3258 ierr = MatSetOption(*mat,MAT_NO_OFF_PROC_ENTRIES,PETSC_FALSE);CHKERRQ(ierr); 3259 ierr = MatSetOption(*mat,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE);CHKERRQ(ierr); 3260 PetscFunctionReturn(0); 3261 } 3262 3263 extern PetscErrorCode MatCreateSubMatrices_MPIAIJ_SingleIS_Local(Mat,PetscInt,const IS[],const IS[],MatReuse,PetscBool,Mat*); 3264 3265 PetscErrorCode MatCreateSubMatrix_MPIAIJ_SameRowDist(Mat mat,IS isrow,IS iscol,MatReuse call,PetscBool sameColDist,Mat *newmat) 3266 { 3267 PetscErrorCode ierr; 3268 PetscInt i,m,n,rstart,row,rend,nz,j,bs,cbs; 3269 PetscInt *ii,*jj,nlocal,*dlens,*olens,dlen,olen,jend,mglobal; 3270 Mat_MPIAIJ *a=(Mat_MPIAIJ*)mat->data; 3271 Mat M,Msub,B=a->B,Mnew=NULL; 3272 MatScalar *aa; 3273 Mat_SeqAIJ *aij; 3274 PetscInt *garray = a->garray,*colsub,Ncols; 3275 PetscInt count,Bn=B->cmap->N,cstart=mat->cmap->rstart,cend=mat->cmap->rend; 3276 IS iscol_sub,iscmap; 3277 const PetscInt *is_idx,*cmap; 3278 PetscBool allcolumns=PETSC_FALSE; 3279 IS iscol_local=NULL; 3280 MPI_Comm comm; 3281 IS iscol_d,isrow_d,iscol_o,isgarray; 3282 Mat Asub=NULL,Bsub=NULL; 3283 3284 PetscFunctionBegin; 3285 ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 3286 PetscMPIInt rank; 3287 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 3288 3289 if (call == MAT_REUSE_MATRIX) { 3290 if (sameColDist) { 3291 Mat_MPIAIJ *matsub=(Mat_MPIAIJ*)(*newmat)->data; 3292 3293 ierr = PetscObjectQuery((PetscObject)*newmat,"isrow_d",(PetscObject*)&isrow_d);CHKERRQ(ierr); 3294 if (!isrow_d) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"isrow_d passed in was not used before, cannot reuse"); 3295 3296 ierr = PetscObjectQuery((PetscObject)*newmat,"iscol_d",(PetscObject*)&iscol_d);CHKERRQ(ierr); 3297 if (!iscol_d) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"iscol_d passed in was not used before, cannot reuse"); 3298 3299 ierr = PetscObjectQuery((PetscObject)*newmat,"iscol_o",(PetscObject*)&iscol_o);CHKERRQ(ierr); 3300 if (!iscol_o) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"iscol_o passed in was not used before, cannot reuse"); 3301 3302 ierr = MatCreateSubMatrix_SeqAIJ(a->A,isrow_d,iscol_d,PETSC_DECIDE,call,&matsub->A);CHKERRQ(ierr); 3303 ierr = MatCreateSubMatrix_SeqAIJ(a->B,isrow_d,iscol_o,PETSC_DECIDE,call,&matsub->B);CHKERRQ(ierr); 3304 ierr = MatAssemblyBegin(*newmat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3305 ierr = MatAssemblyEnd(*newmat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3306 PetscFunctionReturn(0); 3307 3308 } else { 3309 SETERRQ(PETSC_COMM_SELF,0,"not done yet"); 3310 //=============================== 3311 ierr = PetscObjectQuery((PetscObject)*newmat,"SubIScol",(PetscObject*)&iscol_sub);CHKERRQ(ierr); 3312 if (!iscol_sub) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"SubIScol passed in was not used before, cannot reuse"); 3313 ierr = ISGetLocalSize(iscol_sub,&count);CHKERRQ(ierr); 3314 3315 ierr = PetscObjectQuery((PetscObject)*newmat,"Subcmap",(PetscObject*)&iscmap);CHKERRQ(ierr); 3316 if (!iscmap) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Subcmap passed in was not used before, cannot reuse"); 3317 3318 ierr = PetscObjectQuery((PetscObject)*newmat,"SubMatrix",(PetscObject*)&Msub);CHKERRQ(ierr); 3319 if (!Msub) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Submatrix passed in was not used before, cannot reuse"); 3320 3321 ierr = MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat,1,&isrow,&iscol_sub,MAT_REUSE_MATRIX,PETSC_FALSE,&Msub);CHKERRQ(ierr); 3322 } 3323 3324 } else { /* call == MAT_INITIAL_MATRIX) */ 3325 ierr = ISGetLocalSize(iscol,&n);CHKERRQ(ierr); 3326 ierr = ISGetSize(iscol,&Ncols);CHKERRQ(ierr); 3327 3328 if (sameColDist) { 3329 /* iscol has same processor distribution as mat, *newmat is a submatrix of local mat */ 3330 ierr = ISGetSeqIS_SameColDist_Private(mat,isrow,iscol,&isrow_d,&iscol_d,&iscol_o,&isgarray);CHKERRQ(ierr); 3331 3332 ierr = MatCreateSubMatrix_SeqAIJ(a->A,isrow_d,iscol_d,PETSC_DECIDE,call,&Asub);CHKERRQ(ierr); 3333 ierr = MatCreateSubMatrix_SeqAIJ(a->B,isrow_d,iscol_o,PETSC_DECIDE,call,&Bsub);CHKERRQ(ierr); 3334 #if 0 3335 ierr = ISGetBlockSize(isrow,&bs);CHKERRQ(ierr); 3336 ierr = ISGetBlockSize(iscol,&cbs);CHKERRQ(ierr); 3337 ierr = MatSetBlockSizes(Asub,bs,cbs);CHKERRQ(ierr); 3338 printf("[%d] bs %d cbs %d Asub rbs %d, cbs %d\n",rank,bs,cbs,Asub->rmap->bs,Asub->cmap->bs); 3339 #endif 3340 const PetscInt *garray1; 3341 ierr = ISGetIndices(isgarray,&garray1);CHKERRQ(ierr); 3342 ierr = MatCreateMPIAIJWithSeqAIJ(comm,Asub,Bsub,garray1,&Mnew);CHKERRQ(ierr); 3343 ierr = ISRestoreIndices(isgarray,&garray1);CHKERRQ(ierr); 3344 //ierr = MatView(Mnew,PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr); 3345 3346 ierr = ISDestroy(&isgarray);CHKERRQ(ierr); 3347 3348 /* save isrow_d, iscol_d and iscol_o used in processor for next request */ 3349 ierr = PetscObjectCompose((PetscObject)Mnew,"isrow_d",(PetscObject)isrow_d);CHKERRQ(ierr); 3350 ierr = ISDestroy(&isrow_d);CHKERRQ(ierr); 3351 3352 ierr = PetscObjectCompose((PetscObject)Mnew,"iscol_d",(PetscObject)iscol_d);CHKERRQ(ierr); 3353 ierr = ISDestroy(&iscol_d);CHKERRQ(ierr); 3354 3355 ierr = PetscObjectCompose((PetscObject)Mnew,"iscol_o",(PetscObject)iscol_o);CHKERRQ(ierr); 3356 ierr = ISDestroy(&iscol_o);CHKERRQ(ierr); 3357 3358 *newmat = Mnew; 3359 PetscFunctionReturn(0); 3360 3361 } else { /* !sameColDist: iscol -> nonscalable iscol_local, then get scalable iscol_sub and iscmap */ 3362 PetscBool flg; 3363 3364 /* (1) Create scalable iscol_sub and iscmap */ 3365 /* iscol -> nonscalable iscol_local */ 3366 ierr = ISGetSeqIS_Private(mat,iscol,&iscol_local);CHKERRQ(ierr); 3367 ierr = ISGetLocalSize(iscol_local,&n);CHKERRQ(ierr); /* local size of iscol_local = global columns of newmat */ 3368 if (n != Ncols) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"n %d != Ncols %d",n,Ncols); 3369 3370 /* Check for special case: each processor gets entire matrix columns */ 3371 ierr = ISIdentity(iscol_local,&flg);CHKERRQ(ierr); 3372 if (flg && n == mat->cmap->N) allcolumns = PETSC_TRUE; 3373 if (allcolumns) { 3374 iscol_sub = iscol_local; 3375 ierr = PetscObjectReference((PetscObject)iscol_local);CHKERRQ(ierr); 3376 ierr = ISCreateStride(PETSC_COMM_SELF,n,0,1,&iscmap);CHKERRQ(ierr); 3377 3378 } else { /* iscol_local -> iscol_sub and iscmap */ 3379 PetscInt *idx,*cmap1,k,cbs; 3380 3381 /* implementation below requires iscol_local be sorted, it can have duplicate indices */ 3382 ierr = ISSorted(iscol_local,&flg);CHKERRQ(ierr); 3383 if (!flg) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"unsorted iscol_local is not implemented yet"); 3384 3385 ierr = PetscMalloc2(Ncols,&idx,Ncols,&cmap1);CHKERRQ(ierr); 3386 ierr = ISGetIndices(iscol_local,&is_idx);CHKERRQ(ierr); 3387 count = 0; 3388 k = 0; 3389 for (i=0; i<Ncols; i++) { 3390 j = is_idx[i]; 3391 if (j >= cstart && j < cend) { 3392 /* diagonal part of mat */ 3393 idx[count] = j; 3394 cmap1[count++] = i; /* column index in submat */ 3395 } else if (Bn) { 3396 /* off-diagonal part of mat */ 3397 if (j == garray[k]) { 3398 idx[count] = j; 3399 cmap1[count++] = i; /* column index in submat */ 3400 } else if (j > garray[k]) { 3401 while (j > garray[k] && k < Bn-1) k++; 3402 if (j == garray[k]) { 3403 idx[count] = j; 3404 cmap1[count++] = i; /* column index in submat */ 3405 } 3406 } 3407 } 3408 } 3409 ierr = ISRestoreIndices(iscol_local,&is_idx);CHKERRQ(ierr); 3410 3411 ierr = ISCreateGeneral(PETSC_COMM_SELF,count,idx,PETSC_COPY_VALUES,&iscol_sub);CHKERRQ(ierr); 3412 ierr = ISGetBlockSize(iscol,&cbs);CHKERRQ(ierr); 3413 ierr = ISSetBlockSize(iscol_sub,cbs);CHKERRQ(ierr); 3414 3415 ierr = ISCreateGeneral(PetscObjectComm((PetscObject)iscol_local),count,cmap1,PETSC_COPY_VALUES,&iscmap);CHKERRQ(ierr); 3416 ierr = PetscFree2(idx,cmap1);CHKERRQ(ierr); 3417 } 3418 3419 /* (2) Create sequential Msub */ 3420 ierr = MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat,1,&isrow,&iscol_sub,MAT_INITIAL_MATRIX,allcolumns,&Msub);CHKERRQ(ierr); 3421 } 3422 } 3423 3424 ierr = ISGetLocalSize(iscol_sub,&count);CHKERRQ(ierr); 3425 aij = (Mat_SeqAIJ*)(Msub)->data; 3426 ii = aij->i; 3427 ierr = ISGetIndices(iscmap,&cmap);CHKERRQ(ierr); 3428 3429 /* 3430 m - number of local rows 3431 Ncols - number of columns (same on all processors) 3432 rstart - first row in new global matrix generated 3433 */ 3434 ierr = MatGetSize(Msub,&m,NULL);CHKERRQ(ierr); 3435 3436 if (call == MAT_INITIAL_MATRIX && !sameColDist) { 3437 /* (3) Create parallel newmat */ 3438 PetscMPIInt rank,size; 3439 PetscInt csize; 3440 3441 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 3442 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 3443 3444 /* 3445 Determine the number of non-zeros in the diagonal and off-diagonal 3446 portions of the matrix in order to do correct preallocation 3447 */ 3448 3449 /* first get start and end of "diagonal" columns */ 3450 ierr = ISGetLocalSize(iscol,&csize);CHKERRQ(ierr); 3451 if (csize == PETSC_DECIDE) { 3452 ierr = ISGetSize(isrow,&mglobal);CHKERRQ(ierr); 3453 if (mglobal == Ncols) { /* square matrix */ 3454 nlocal = m; 3455 } else { 3456 nlocal = Ncols/size + ((Ncols % size) > rank); 3457 } 3458 } else { 3459 nlocal = csize; 3460 } 3461 ierr = MPI_Scan(&nlocal,&rend,1,MPIU_INT,MPI_SUM,comm);CHKERRQ(ierr); 3462 rstart = rend - nlocal; 3463 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); 3464 3465 /* next, compute all the lengths */ 3466 jj = aij->j; 3467 ierr = PetscMalloc1(2*m+1,&dlens);CHKERRQ(ierr); 3468 olens = dlens + m; 3469 for (i=0; i<m; i++) { 3470 jend = ii[i+1] - ii[i]; 3471 olen = 0; 3472 dlen = 0; 3473 for (j=0; j<jend; j++) { 3474 if (cmap[*jj] < rstart || cmap[*jj] >= rend) olen++; 3475 else dlen++; 3476 jj++; 3477 } 3478 olens[i] = olen; 3479 dlens[i] = dlen; 3480 } 3481 ierr = MatGetBlockSizes(Msub,&bs,&cbs);CHKERRQ(ierr); 3482 3483 ierr = MatCreate(comm,&M);CHKERRQ(ierr); 3484 ierr = MatSetSizes(M,m,nlocal,PETSC_DECIDE,Ncols);CHKERRQ(ierr); 3485 ierr = MatSetBlockSizes(M,bs,cbs);CHKERRQ(ierr); 3486 ierr = MatSetType(M,((PetscObject)mat)->type_name);CHKERRQ(ierr); 3487 ierr = MatMPIAIJSetPreallocation(M,0,dlens,0,olens);CHKERRQ(ierr); 3488 ierr = PetscFree(dlens);CHKERRQ(ierr); 3489 3490 } else { /* call == MAT_REUSE_MATRIX */ 3491 M = *newmat; 3492 ierr = MatGetLocalSize(M,&i,NULL);CHKERRQ(ierr); 3493 if (i != m) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Previous matrix must be same size/layout as request"); 3494 ierr = MatZeroEntries(M);CHKERRQ(ierr); 3495 /* 3496 The next two lines are needed so we may call MatSetValues_MPIAIJ() below directly, 3497 rather than the slower MatSetValues(). 3498 */ 3499 M->was_assembled = PETSC_TRUE; 3500 M->assembled = PETSC_FALSE; 3501 } 3502 3503 /* (4) Set values of Msub to *newmat */ 3504 ierr = PetscMalloc1(count,&colsub);CHKERRQ(ierr); 3505 ierr = MatGetOwnershipRange(M,&rstart,NULL);CHKERRQ(ierr); 3506 3507 jj = aij->j; 3508 aa = aij->a; 3509 for (i=0; i<m; i++) { 3510 row = rstart + i; 3511 nz = ii[i+1] - ii[i]; 3512 for (j=0; j<nz; j++) colsub[j] = cmap[jj[j]]; 3513 ierr = MatSetValues_MPIAIJ(M,1,&row,nz,colsub,aa,INSERT_VALUES);CHKERRQ(ierr); 3514 jj += nz; aa += nz; 3515 } 3516 ierr = ISRestoreIndices(iscmap,&cmap);CHKERRQ(ierr); 3517 3518 ierr = MatAssemblyBegin(M,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3519 ierr = MatAssemblyEnd(M,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3520 3521 ierr = PetscFree(colsub);CHKERRQ(ierr); 3522 3523 /* save Msub, iscol_sub and iscmap used in processor for next request */ 3524 if (call == MAT_INITIAL_MATRIX) { 3525 if (!sameColDist) *newmat = M; 3526 ierr = PetscObjectCompose((PetscObject)(*newmat),"SubMatrix",(PetscObject)Msub);CHKERRQ(ierr); 3527 ierr = MatDestroy(&Msub);CHKERRQ(ierr); 3528 3529 ierr = PetscObjectCompose((PetscObject)(*newmat),"SubIScol",(PetscObject)iscol_sub);CHKERRQ(ierr); 3530 ierr = ISDestroy(&iscol_sub);CHKERRQ(ierr); 3531 3532 ierr = PetscObjectCompose((PetscObject)(*newmat),"Subcmap",(PetscObject)iscmap);CHKERRQ(ierr); 3533 ierr = ISDestroy(&iscmap);CHKERRQ(ierr); 3534 3535 if (iscol_local) { 3536 ierr = PetscObjectCompose((PetscObject)(*newmat),"ISAllGather",(PetscObject)iscol_local);CHKERRQ(ierr); 3537 ierr = ISDestroy(&iscol_local);CHKERRQ(ierr); 3538 } 3539 3540 if (*newmat != M) {ierr = MatDestroy(&M);CHKERRQ(ierr);} 3541 } 3542 PetscFunctionReturn(0); 3543 } 3544 3545 /* 3546 Not great since it makes two copies of the submatrix, first an SeqAIJ 3547 in local and then by concatenating the local matrices the end result. 3548 Writing it directly would be much like MatCreateSubMatrices_MPIAIJ() 3549 3550 Note: This requires a sequential iscol with all indices. 3551 */ 3552 PetscErrorCode MatCreateSubMatrix_MPIAIJ_nonscalable(Mat mat,IS isrow,IS iscol,PetscInt csize,MatReuse call,Mat *newmat) 3553 { 3554 PetscErrorCode ierr; 3555 PetscMPIInt rank,size; 3556 PetscInt i,m,n,rstart,row,rend,nz,*cwork,j,bs,cbs; 3557 PetscInt *ii,*jj,nlocal,*dlens,*olens,dlen,olen,jend,mglobal; 3558 Mat M,Mreuse; 3559 MatScalar *aa,*vwork; 3560 MPI_Comm comm; 3561 Mat_SeqAIJ *aij; 3562 PetscBool colflag,allcolumns=PETSC_FALSE; 3563 3564 PetscFunctionBegin; 3565 ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 3566 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 3567 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 3568 3569 /* Check for special case: each processor gets entire matrix columns */ 3570 ierr = ISIdentity(iscol,&colflag);CHKERRQ(ierr); 3571 ierr = ISGetLocalSize(iscol,&n);CHKERRQ(ierr); 3572 if (colflag && n == mat->cmap->N) allcolumns = PETSC_TRUE; 3573 3574 if (call == MAT_REUSE_MATRIX) { 3575 ierr = PetscObjectQuery((PetscObject)*newmat,"SubMatrix",(PetscObject*)&Mreuse);CHKERRQ(ierr); 3576 if (!Mreuse) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Submatrix passed in was not used before, cannot reuse"); 3577 ierr = MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat,1,&isrow,&iscol,MAT_REUSE_MATRIX,allcolumns,&Mreuse);CHKERRQ(ierr); 3578 } else { 3579 ierr = MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat,1,&isrow,&iscol,MAT_INITIAL_MATRIX,allcolumns,&Mreuse);CHKERRQ(ierr); 3580 } 3581 3582 /* 3583 m - number of local rows 3584 n - number of columns (same on all processors) 3585 rstart - first row in new global matrix generated 3586 */ 3587 ierr = MatGetSize(Mreuse,&m,&n);CHKERRQ(ierr); 3588 ierr = MatGetBlockSizes(Mreuse,&bs,&cbs);CHKERRQ(ierr); 3589 if (call == MAT_INITIAL_MATRIX) { 3590 aij = (Mat_SeqAIJ*)(Mreuse)->data; 3591 ii = aij->i; 3592 jj = aij->j; 3593 3594 /* 3595 Determine the number of non-zeros in the diagonal and off-diagonal 3596 portions of the matrix in order to do correct preallocation 3597 */ 3598 3599 /* first get start and end of "diagonal" columns */ 3600 if (csize == PETSC_DECIDE) { 3601 ierr = ISGetSize(isrow,&mglobal);CHKERRQ(ierr); 3602 if (mglobal == n) { /* square matrix */ 3603 nlocal = m; 3604 } else { 3605 nlocal = n/size + ((n % size) > rank); 3606 } 3607 } else { 3608 nlocal = csize; 3609 } 3610 ierr = MPI_Scan(&nlocal,&rend,1,MPIU_INT,MPI_SUM,comm);CHKERRQ(ierr); 3611 rstart = rend - nlocal; 3612 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); 3613 3614 /* next, compute all the lengths */ 3615 ierr = PetscMalloc1(2*m+1,&dlens);CHKERRQ(ierr); 3616 olens = dlens + m; 3617 for (i=0; i<m; i++) { 3618 jend = ii[i+1] - ii[i]; 3619 olen = 0; 3620 dlen = 0; 3621 for (j=0; j<jend; j++) { 3622 if (*jj < rstart || *jj >= rend) olen++; 3623 else dlen++; 3624 jj++; 3625 } 3626 olens[i] = olen; 3627 dlens[i] = dlen; 3628 } 3629 ierr = MatCreate(comm,&M);CHKERRQ(ierr); 3630 ierr = MatSetSizes(M,m,nlocal,PETSC_DECIDE,n);CHKERRQ(ierr); 3631 ierr = MatSetBlockSizes(M,bs,cbs);CHKERRQ(ierr); 3632 ierr = MatSetType(M,((PetscObject)mat)->type_name);CHKERRQ(ierr); 3633 ierr = MatMPIAIJSetPreallocation(M,0,dlens,0,olens);CHKERRQ(ierr); 3634 ierr = PetscFree(dlens);CHKERRQ(ierr); 3635 } else { 3636 PetscInt ml,nl; 3637 3638 M = *newmat; 3639 ierr = MatGetLocalSize(M,&ml,&nl);CHKERRQ(ierr); 3640 if (ml != m) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Previous matrix must be same size/layout as request"); 3641 ierr = MatZeroEntries(M);CHKERRQ(ierr); 3642 /* 3643 The next two lines are needed so we may call MatSetValues_MPIAIJ() below directly, 3644 rather than the slower MatSetValues(). 3645 */ 3646 M->was_assembled = PETSC_TRUE; 3647 M->assembled = PETSC_FALSE; 3648 } 3649 ierr = MatGetOwnershipRange(M,&rstart,&rend);CHKERRQ(ierr); 3650 aij = (Mat_SeqAIJ*)(Mreuse)->data; 3651 ii = aij->i; 3652 jj = aij->j; 3653 aa = aij->a; 3654 for (i=0; i<m; i++) { 3655 row = rstart + i; 3656 nz = ii[i+1] - ii[i]; 3657 cwork = jj; jj += nz; 3658 vwork = aa; aa += nz; 3659 ierr = MatSetValues_MPIAIJ(M,1,&row,nz,cwork,vwork,INSERT_VALUES);CHKERRQ(ierr); 3660 } 3661 3662 ierr = MatAssemblyBegin(M,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3663 ierr = MatAssemblyEnd(M,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3664 *newmat = M; 3665 3666 /* save submatrix used in processor for next request */ 3667 if (call == MAT_INITIAL_MATRIX) { 3668 ierr = PetscObjectCompose((PetscObject)M,"SubMatrix",(PetscObject)Mreuse);CHKERRQ(ierr); 3669 ierr = MatDestroy(&Mreuse);CHKERRQ(ierr); 3670 } 3671 PetscFunctionReturn(0); 3672 } 3673 3674 PetscErrorCode MatMPIAIJSetPreallocationCSR_MPIAIJ(Mat B,const PetscInt Ii[],const PetscInt J[],const PetscScalar v[]) 3675 { 3676 PetscInt m,cstart, cend,j,nnz,i,d; 3677 PetscInt *d_nnz,*o_nnz,nnz_max = 0,rstart,ii; 3678 const PetscInt *JJ; 3679 PetscScalar *values; 3680 PetscErrorCode ierr; 3681 PetscBool nooffprocentries; 3682 3683 PetscFunctionBegin; 3684 if (Ii[0]) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Ii[0] must be 0 it is %D",Ii[0]); 3685 3686 ierr = PetscLayoutSetUp(B->rmap);CHKERRQ(ierr); 3687 ierr = PetscLayoutSetUp(B->cmap);CHKERRQ(ierr); 3688 m = B->rmap->n; 3689 cstart = B->cmap->rstart; 3690 cend = B->cmap->rend; 3691 rstart = B->rmap->rstart; 3692 3693 ierr = PetscMalloc2(m,&d_nnz,m,&o_nnz);CHKERRQ(ierr); 3694 3695 #if defined(PETSC_USE_DEBUGGING) 3696 for (i=0; i<m; i++) { 3697 nnz = Ii[i+1]- Ii[i]; 3698 JJ = J + Ii[i]; 3699 if (nnz < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Local row %D has a negative %D number of columns",i,nnz); 3700 if (nnz && (JJ[0] < 0)) SETERRRQ1(PETSC_ERR_ARG_WRONGSTATE,"Row %D starts with negative column index",i,j); 3701 if (nnz && (JJ[nnz-1] >= B->cmap->N) SETERRRQ3(PETSC_ERR_ARG_WRONGSTATE,"Row %D ends with too large a column index %D (max allowed %D)",i,JJ[nnz-1],B->cmap->N); 3702 } 3703 #endif 3704 3705 for (i=0; i<m; i++) { 3706 nnz = Ii[i+1]- Ii[i]; 3707 JJ = J + Ii[i]; 3708 nnz_max = PetscMax(nnz_max,nnz); 3709 d = 0; 3710 for (j=0; j<nnz; j++) { 3711 if (cstart <= JJ[j] && JJ[j] < cend) d++; 3712 } 3713 d_nnz[i] = d; 3714 o_nnz[i] = nnz - d; 3715 } 3716 ierr = MatMPIAIJSetPreallocation(B,0,d_nnz,0,o_nnz);CHKERRQ(ierr); 3717 ierr = PetscFree2(d_nnz,o_nnz);CHKERRQ(ierr); 3718 3719 if (v) values = (PetscScalar*)v; 3720 else { 3721 ierr = PetscCalloc1(nnz_max+1,&values);CHKERRQ(ierr); 3722 } 3723 3724 for (i=0; i<m; i++) { 3725 ii = i + rstart; 3726 nnz = Ii[i+1]- Ii[i]; 3727 ierr = MatSetValues_MPIAIJ(B,1,&ii,nnz,J+Ii[i],values+(v ? Ii[i] : 0),INSERT_VALUES);CHKERRQ(ierr); 3728 } 3729 nooffprocentries = B->nooffprocentries; 3730 B->nooffprocentries = PETSC_TRUE; 3731 ierr = MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3732 ierr = MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3733 B->nooffprocentries = nooffprocentries; 3734 3735 if (!v) { 3736 ierr = PetscFree(values);CHKERRQ(ierr); 3737 } 3738 ierr = MatSetOption(B,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE);CHKERRQ(ierr); 3739 PetscFunctionReturn(0); 3740 } 3741 3742 /*@ 3743 MatMPIAIJSetPreallocationCSR - Allocates memory for a sparse parallel matrix in AIJ format 3744 (the default parallel PETSc format). 3745 3746 Collective on MPI_Comm 3747 3748 Input Parameters: 3749 + B - the matrix 3750 . i - the indices into j for the start of each local row (starts with zero) 3751 . j - the column indices for each local row (starts with zero) 3752 - v - optional values in the matrix 3753 3754 Level: developer 3755 3756 Notes: 3757 The i, j, and a arrays ARE copied by this routine into the internal format used by PETSc; 3758 thus you CANNOT change the matrix entries by changing the values of a[] after you have 3759 called this routine. Use MatCreateMPIAIJWithSplitArrays() to avoid needing to copy the arrays. 3760 3761 The i and j indices are 0 based, and i indices are indices corresponding to the local j array. 3762 3763 The format which is used for the sparse matrix input, is equivalent to a 3764 row-major ordering.. i.e for the following matrix, the input data expected is 3765 as shown 3766 3767 $ 1 0 0 3768 $ 2 0 3 P0 3769 $ ------- 3770 $ 4 5 6 P1 3771 $ 3772 $ Process0 [P0]: rows_owned=[0,1] 3773 $ i = {0,1,3} [size = nrow+1 = 2+1] 3774 $ j = {0,0,2} [size = 3] 3775 $ v = {1,2,3} [size = 3] 3776 $ 3777 $ Process1 [P1]: rows_owned=[2] 3778 $ i = {0,3} [size = nrow+1 = 1+1] 3779 $ j = {0,1,2} [size = 3] 3780 $ v = {4,5,6} [size = 3] 3781 3782 .keywords: matrix, aij, compressed row, sparse, parallel 3783 3784 .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatCreateAIJ(), MATMPIAIJ, 3785 MatCreateSeqAIJWithArrays(), MatCreateMPIAIJWithSplitArrays() 3786 @*/ 3787 PetscErrorCode MatMPIAIJSetPreallocationCSR(Mat B,const PetscInt i[],const PetscInt j[], const PetscScalar v[]) 3788 { 3789 PetscErrorCode ierr; 3790 3791 PetscFunctionBegin; 3792 ierr = PetscTryMethod(B,"MatMPIAIJSetPreallocationCSR_C",(Mat,const PetscInt[],const PetscInt[],const PetscScalar[]),(B,i,j,v));CHKERRQ(ierr); 3793 PetscFunctionReturn(0); 3794 } 3795 3796 /*@C 3797 MatMPIAIJSetPreallocation - Preallocates memory for a sparse parallel matrix in AIJ format 3798 (the default parallel PETSc format). For good matrix assembly performance 3799 the user should preallocate the matrix storage by setting the parameters 3800 d_nz (or d_nnz) and o_nz (or o_nnz). By setting these parameters accurately, 3801 performance can be increased by more than a factor of 50. 3802 3803 Collective on MPI_Comm 3804 3805 Input Parameters: 3806 + B - the matrix 3807 . d_nz - number of nonzeros per row in DIAGONAL portion of local submatrix 3808 (same value is used for all local rows) 3809 . d_nnz - array containing the number of nonzeros in the various rows of the 3810 DIAGONAL portion of the local submatrix (possibly different for each row) 3811 or NULL (PETSC_NULL_INTEGER in Fortran), if d_nz is used to specify the nonzero structure. 3812 The size of this array is equal to the number of local rows, i.e 'm'. 3813 For matrices that will be factored, you must leave room for (and set) 3814 the diagonal entry even if it is zero. 3815 . o_nz - number of nonzeros per row in the OFF-DIAGONAL portion of local 3816 submatrix (same value is used for all local rows). 3817 - o_nnz - array containing the number of nonzeros in the various rows of the 3818 OFF-DIAGONAL portion of the local submatrix (possibly different for 3819 each row) or NULL (PETSC_NULL_INTEGER in Fortran), if o_nz is used to specify the nonzero 3820 structure. The size of this array is equal to the number 3821 of local rows, i.e 'm'. 3822 3823 If the *_nnz parameter is given then the *_nz parameter is ignored 3824 3825 The AIJ format (also called the Yale sparse matrix format or 3826 compressed row storage (CSR)), is fully compatible with standard Fortran 77 3827 storage. The stored row and column indices begin with zero. 3828 See Users-Manual: ch_mat for details. 3829 3830 The parallel matrix is partitioned such that the first m0 rows belong to 3831 process 0, the next m1 rows belong to process 1, the next m2 rows belong 3832 to process 2 etc.. where m0,m1,m2... are the input parameter 'm'. 3833 3834 The DIAGONAL portion of the local submatrix of a processor can be defined 3835 as the submatrix which is obtained by extraction the part corresponding to 3836 the rows r1-r2 and columns c1-c2 of the global matrix, where r1 is the 3837 first row that belongs to the processor, r2 is the last row belonging to 3838 the this processor, and c1-c2 is range of indices of the local part of a 3839 vector suitable for applying the matrix to. This is an mxn matrix. In the 3840 common case of a square matrix, the row and column ranges are the same and 3841 the DIAGONAL part is also square. The remaining portion of the local 3842 submatrix (mxN) constitute the OFF-DIAGONAL portion. 3843 3844 If o_nnz, d_nnz are specified, then o_nz, and d_nz are ignored. 3845 3846 You can call MatGetInfo() to get information on how effective the preallocation was; 3847 for example the fields mallocs,nz_allocated,nz_used,nz_unneeded; 3848 You can also run with the option -info and look for messages with the string 3849 malloc in them to see if additional memory allocation was needed. 3850 3851 Example usage: 3852 3853 Consider the following 8x8 matrix with 34 non-zero values, that is 3854 assembled across 3 processors. Lets assume that proc0 owns 3 rows, 3855 proc1 owns 3 rows, proc2 owns 2 rows. This division can be shown 3856 as follows: 3857 3858 .vb 3859 1 2 0 | 0 3 0 | 0 4 3860 Proc0 0 5 6 | 7 0 0 | 8 0 3861 9 0 10 | 11 0 0 | 12 0 3862 ------------------------------------- 3863 13 0 14 | 15 16 17 | 0 0 3864 Proc1 0 18 0 | 19 20 21 | 0 0 3865 0 0 0 | 22 23 0 | 24 0 3866 ------------------------------------- 3867 Proc2 25 26 27 | 0 0 28 | 29 0 3868 30 0 0 | 31 32 33 | 0 34 3869 .ve 3870 3871 This can be represented as a collection of submatrices as: 3872 3873 .vb 3874 A B C 3875 D E F 3876 G H I 3877 .ve 3878 3879 Where the submatrices A,B,C are owned by proc0, D,E,F are 3880 owned by proc1, G,H,I are owned by proc2. 3881 3882 The 'm' parameters for proc0,proc1,proc2 are 3,3,2 respectively. 3883 The 'n' parameters for proc0,proc1,proc2 are 3,3,2 respectively. 3884 The 'M','N' parameters are 8,8, and have the same values on all procs. 3885 3886 The DIAGONAL submatrices corresponding to proc0,proc1,proc2 are 3887 submatrices [A], [E], [I] respectively. The OFF-DIAGONAL submatrices 3888 corresponding to proc0,proc1,proc2 are [BC], [DF], [GH] respectively. 3889 Internally, each processor stores the DIAGONAL part, and the OFF-DIAGONAL 3890 part as SeqAIJ matrices. for eg: proc1 will store [E] as a SeqAIJ 3891 matrix, ans [DF] as another SeqAIJ matrix. 3892 3893 When d_nz, o_nz parameters are specified, d_nz storage elements are 3894 allocated for every row of the local diagonal submatrix, and o_nz 3895 storage locations are allocated for every row of the OFF-DIAGONAL submat. 3896 One way to choose d_nz and o_nz is to use the max nonzerors per local 3897 rows for each of the local DIAGONAL, and the OFF-DIAGONAL submatrices. 3898 In this case, the values of d_nz,o_nz are: 3899 .vb 3900 proc0 : dnz = 2, o_nz = 2 3901 proc1 : dnz = 3, o_nz = 2 3902 proc2 : dnz = 1, o_nz = 4 3903 .ve 3904 We are allocating m*(d_nz+o_nz) storage locations for every proc. This 3905 translates to 3*(2+2)=12 for proc0, 3*(3+2)=15 for proc1, 2*(1+4)=10 3906 for proc3. i.e we are using 12+15+10=37 storage locations to store 3907 34 values. 3908 3909 When d_nnz, o_nnz parameters are specified, the storage is specified 3910 for every row, coresponding to both DIAGONAL and OFF-DIAGONAL submatrices. 3911 In the above case the values for d_nnz,o_nnz are: 3912 .vb 3913 proc0: d_nnz = [2,2,2] and o_nnz = [2,2,2] 3914 proc1: d_nnz = [3,3,2] and o_nnz = [2,1,1] 3915 proc2: d_nnz = [1,1] and o_nnz = [4,4] 3916 .ve 3917 Here the space allocated is sum of all the above values i.e 34, and 3918 hence pre-allocation is perfect. 3919 3920 Level: intermediate 3921 3922 .keywords: matrix, aij, compressed row, sparse, parallel 3923 3924 .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatCreateAIJ(), MatMPIAIJSetPreallocationCSR(), 3925 MATMPIAIJ, MatGetInfo(), PetscSplitOwnership() 3926 @*/ 3927 PetscErrorCode MatMPIAIJSetPreallocation(Mat B,PetscInt d_nz,const PetscInt d_nnz[],PetscInt o_nz,const PetscInt o_nnz[]) 3928 { 3929 PetscErrorCode ierr; 3930 3931 PetscFunctionBegin; 3932 PetscValidHeaderSpecific(B,MAT_CLASSID,1); 3933 PetscValidType(B,1); 3934 ierr = PetscTryMethod(B,"MatMPIAIJSetPreallocation_C",(Mat,PetscInt,const PetscInt[],PetscInt,const PetscInt[]),(B,d_nz,d_nnz,o_nz,o_nnz));CHKERRQ(ierr); 3935 PetscFunctionReturn(0); 3936 } 3937 3938 /*@ 3939 MatCreateMPIAIJWithArrays - creates a MPI AIJ matrix using arrays that contain in standard 3940 CSR format the local rows. 3941 3942 Collective on MPI_Comm 3943 3944 Input Parameters: 3945 + comm - MPI communicator 3946 . m - number of local rows (Cannot be PETSC_DECIDE) 3947 . n - This value should be the same as the local size used in creating the 3948 x vector for the matrix-vector product y = Ax. (or PETSC_DECIDE to have 3949 calculated if N is given) For square matrices n is almost always m. 3950 . M - number of global rows (or PETSC_DETERMINE to have calculated if m is given) 3951 . N - number of global columns (or PETSC_DETERMINE to have calculated if n is given) 3952 . i - row indices 3953 . j - column indices 3954 - a - matrix values 3955 3956 Output Parameter: 3957 . mat - the matrix 3958 3959 Level: intermediate 3960 3961 Notes: 3962 The i, j, and a arrays ARE copied by this routine into the internal format used by PETSc; 3963 thus you CANNOT change the matrix entries by changing the values of a[] after you have 3964 called this routine. Use MatCreateMPIAIJWithSplitArrays() to avoid needing to copy the arrays. 3965 3966 The i and j indices are 0 based, and i indices are indices corresponding to the local j array. 3967 3968 The format which is used for the sparse matrix input, is equivalent to a 3969 row-major ordering.. i.e for the following matrix, the input data expected is 3970 as shown 3971 3972 $ 1 0 0 3973 $ 2 0 3 P0 3974 $ ------- 3975 $ 4 5 6 P1 3976 $ 3977 $ Process0 [P0]: rows_owned=[0,1] 3978 $ i = {0,1,3} [size = nrow+1 = 2+1] 3979 $ j = {0,0,2} [size = 3] 3980 $ v = {1,2,3} [size = 3] 3981 $ 3982 $ Process1 [P1]: rows_owned=[2] 3983 $ i = {0,3} [size = nrow+1 = 1+1] 3984 $ j = {0,1,2} [size = 3] 3985 $ v = {4,5,6} [size = 3] 3986 3987 .keywords: matrix, aij, compressed row, sparse, parallel 3988 3989 .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatMPIAIJSetPreallocationCSR(), 3990 MATMPIAIJ, MatCreateAIJ(), MatCreateMPIAIJWithSplitArrays() 3991 @*/ 3992 PetscErrorCode MatCreateMPIAIJWithArrays(MPI_Comm comm,PetscInt m,PetscInt n,PetscInt M,PetscInt N,const PetscInt i[],const PetscInt j[],const PetscScalar a[],Mat *mat) 3993 { 3994 PetscErrorCode ierr; 3995 3996 PetscFunctionBegin; 3997 if (i[0]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"i (row indices) must start with 0"); 3998 if (m < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"local number of rows (m) cannot be PETSC_DECIDE, or negative"); 3999 ierr = MatCreate(comm,mat);CHKERRQ(ierr); 4000 ierr = MatSetSizes(*mat,m,n,M,N);CHKERRQ(ierr); 4001 /* ierr = MatSetBlockSizes(M,bs,cbs);CHKERRQ(ierr); */ 4002 ierr = MatSetType(*mat,MATMPIAIJ);CHKERRQ(ierr); 4003 ierr = MatMPIAIJSetPreallocationCSR(*mat,i,j,a);CHKERRQ(ierr); 4004 PetscFunctionReturn(0); 4005 } 4006 4007 /*@C 4008 MatCreateAIJ - Creates a sparse parallel matrix in AIJ format 4009 (the default parallel PETSc format). For good matrix assembly performance 4010 the user should preallocate the matrix storage by setting the parameters 4011 d_nz (or d_nnz) and o_nz (or o_nnz). By setting these parameters accurately, 4012 performance can be increased by more than a factor of 50. 4013 4014 Collective on MPI_Comm 4015 4016 Input Parameters: 4017 + comm - MPI communicator 4018 . m - number of local rows (or PETSC_DECIDE to have calculated if M is given) 4019 This value should be the same as the local size used in creating the 4020 y vector for the matrix-vector product y = Ax. 4021 . n - This value should be the same as the local size used in creating the 4022 x vector for the matrix-vector product y = Ax. (or PETSC_DECIDE to have 4023 calculated if N is given) For square matrices n is almost always m. 4024 . M - number of global rows (or PETSC_DETERMINE to have calculated if m is given) 4025 . N - number of global columns (or PETSC_DETERMINE to have calculated if n is given) 4026 . d_nz - number of nonzeros per row in DIAGONAL portion of local submatrix 4027 (same value is used for all local rows) 4028 . d_nnz - array containing the number of nonzeros in the various rows of the 4029 DIAGONAL portion of the local submatrix (possibly different for each row) 4030 or NULL, if d_nz is used to specify the nonzero structure. 4031 The size of this array is equal to the number of local rows, i.e 'm'. 4032 . o_nz - number of nonzeros per row in the OFF-DIAGONAL portion of local 4033 submatrix (same value is used for all local rows). 4034 - o_nnz - array containing the number of nonzeros in the various rows of the 4035 OFF-DIAGONAL portion of the local submatrix (possibly different for 4036 each row) or NULL, if o_nz is used to specify the nonzero 4037 structure. The size of this array is equal to the number 4038 of local rows, i.e 'm'. 4039 4040 Output Parameter: 4041 . A - the matrix 4042 4043 It is recommended that one use the MatCreate(), MatSetType() and/or MatSetFromOptions(), 4044 MatXXXXSetPreallocation() paradgm instead of this routine directly. 4045 [MatXXXXSetPreallocation() is, for example, MatSeqAIJSetPreallocation] 4046 4047 Notes: 4048 If the *_nnz parameter is given then the *_nz parameter is ignored 4049 4050 m,n,M,N parameters specify the size of the matrix, and its partitioning across 4051 processors, while d_nz,d_nnz,o_nz,o_nnz parameters specify the approximate 4052 storage requirements for this matrix. 4053 4054 If PETSC_DECIDE or PETSC_DETERMINE is used for a particular argument on one 4055 processor than it must be used on all processors that share the object for 4056 that argument. 4057 4058 The user MUST specify either the local or global matrix dimensions 4059 (possibly both). 4060 4061 The parallel matrix is partitioned across processors such that the 4062 first m0 rows belong to process 0, the next m1 rows belong to 4063 process 1, the next m2 rows belong to process 2 etc.. where 4064 m0,m1,m2,.. are the input parameter 'm'. i.e each processor stores 4065 values corresponding to [m x N] submatrix. 4066 4067 The columns are logically partitioned with the n0 columns belonging 4068 to 0th partition, the next n1 columns belonging to the next 4069 partition etc.. where n0,n1,n2... are the input parameter 'n'. 4070 4071 The DIAGONAL portion of the local submatrix on any given processor 4072 is the submatrix corresponding to the rows and columns m,n 4073 corresponding to the given processor. i.e diagonal matrix on 4074 process 0 is [m0 x n0], diagonal matrix on process 1 is [m1 x n1] 4075 etc. The remaining portion of the local submatrix [m x (N-n)] 4076 constitute the OFF-DIAGONAL portion. The example below better 4077 illustrates this concept. 4078 4079 For a square global matrix we define each processor's diagonal portion 4080 to be its local rows and the corresponding columns (a square submatrix); 4081 each processor's off-diagonal portion encompasses the remainder of the 4082 local matrix (a rectangular submatrix). 4083 4084 If o_nnz, d_nnz are specified, then o_nz, and d_nz are ignored. 4085 4086 When calling this routine with a single process communicator, a matrix of 4087 type SEQAIJ is returned. If a matrix of type MATMPIAIJ is desired for this 4088 type of communicator, use the construction mechanism: 4089 MatCreate(...,&A); MatSetType(A,MATMPIAIJ); MatSetSizes(A, m,n,M,N); MatMPIAIJSetPreallocation(A,...); 4090 4091 By default, this format uses inodes (identical nodes) when possible. 4092 We search for consecutive rows with the same nonzero structure, thereby 4093 reusing matrix information to achieve increased efficiency. 4094 4095 Options Database Keys: 4096 + -mat_no_inode - Do not use inodes 4097 . -mat_inode_limit <limit> - Sets inode limit (max limit=5) 4098 - -mat_aij_oneindex - Internally use indexing starting at 1 4099 rather than 0. Note that when calling MatSetValues(), 4100 the user still MUST index entries starting at 0! 4101 4102 4103 Example usage: 4104 4105 Consider the following 8x8 matrix with 34 non-zero values, that is 4106 assembled across 3 processors. Lets assume that proc0 owns 3 rows, 4107 proc1 owns 3 rows, proc2 owns 2 rows. This division can be shown 4108 as follows: 4109 4110 .vb 4111 1 2 0 | 0 3 0 | 0 4 4112 Proc0 0 5 6 | 7 0 0 | 8 0 4113 9 0 10 | 11 0 0 | 12 0 4114 ------------------------------------- 4115 13 0 14 | 15 16 17 | 0 0 4116 Proc1 0 18 0 | 19 20 21 | 0 0 4117 0 0 0 | 22 23 0 | 24 0 4118 ------------------------------------- 4119 Proc2 25 26 27 | 0 0 28 | 29 0 4120 30 0 0 | 31 32 33 | 0 34 4121 .ve 4122 4123 This can be represented as a collection of submatrices as: 4124 4125 .vb 4126 A B C 4127 D E F 4128 G H I 4129 .ve 4130 4131 Where the submatrices A,B,C are owned by proc0, D,E,F are 4132 owned by proc1, G,H,I are owned by proc2. 4133 4134 The 'm' parameters for proc0,proc1,proc2 are 3,3,2 respectively. 4135 The 'n' parameters for proc0,proc1,proc2 are 3,3,2 respectively. 4136 The 'M','N' parameters are 8,8, and have the same values on all procs. 4137 4138 The DIAGONAL submatrices corresponding to proc0,proc1,proc2 are 4139 submatrices [A], [E], [I] respectively. The OFF-DIAGONAL submatrices 4140 corresponding to proc0,proc1,proc2 are [BC], [DF], [GH] respectively. 4141 Internally, each processor stores the DIAGONAL part, and the OFF-DIAGONAL 4142 part as SeqAIJ matrices. for eg: proc1 will store [E] as a SeqAIJ 4143 matrix, ans [DF] as another SeqAIJ matrix. 4144 4145 When d_nz, o_nz parameters are specified, d_nz storage elements are 4146 allocated for every row of the local diagonal submatrix, and o_nz 4147 storage locations are allocated for every row of the OFF-DIAGONAL submat. 4148 One way to choose d_nz and o_nz is to use the max nonzerors per local 4149 rows for each of the local DIAGONAL, and the OFF-DIAGONAL submatrices. 4150 In this case, the values of d_nz,o_nz are: 4151 .vb 4152 proc0 : dnz = 2, o_nz = 2 4153 proc1 : dnz = 3, o_nz = 2 4154 proc2 : dnz = 1, o_nz = 4 4155 .ve 4156 We are allocating m*(d_nz+o_nz) storage locations for every proc. This 4157 translates to 3*(2+2)=12 for proc0, 3*(3+2)=15 for proc1, 2*(1+4)=10 4158 for proc3. i.e we are using 12+15+10=37 storage locations to store 4159 34 values. 4160 4161 When d_nnz, o_nnz parameters are specified, the storage is specified 4162 for every row, coresponding to both DIAGONAL and OFF-DIAGONAL submatrices. 4163 In the above case the values for d_nnz,o_nnz are: 4164 .vb 4165 proc0: d_nnz = [2,2,2] and o_nnz = [2,2,2] 4166 proc1: d_nnz = [3,3,2] and o_nnz = [2,1,1] 4167 proc2: d_nnz = [1,1] and o_nnz = [4,4] 4168 .ve 4169 Here the space allocated is sum of all the above values i.e 34, and 4170 hence pre-allocation is perfect. 4171 4172 Level: intermediate 4173 4174 .keywords: matrix, aij, compressed row, sparse, parallel 4175 4176 .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatMPIAIJSetPreallocationCSR(), 4177 MATMPIAIJ, MatCreateMPIAIJWithArrays() 4178 @*/ 4179 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) 4180 { 4181 PetscErrorCode ierr; 4182 PetscMPIInt size; 4183 4184 PetscFunctionBegin; 4185 ierr = MatCreate(comm,A);CHKERRQ(ierr); 4186 ierr = MatSetSizes(*A,m,n,M,N);CHKERRQ(ierr); 4187 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 4188 if (size > 1) { 4189 ierr = MatSetType(*A,MATMPIAIJ);CHKERRQ(ierr); 4190 ierr = MatMPIAIJSetPreallocation(*A,d_nz,d_nnz,o_nz,o_nnz);CHKERRQ(ierr); 4191 } else { 4192 ierr = MatSetType(*A,MATSEQAIJ);CHKERRQ(ierr); 4193 ierr = MatSeqAIJSetPreallocation(*A,d_nz,d_nnz);CHKERRQ(ierr); 4194 } 4195 PetscFunctionReturn(0); 4196 } 4197 4198 PetscErrorCode MatMPIAIJGetSeqAIJ(Mat A,Mat *Ad,Mat *Ao,const PetscInt *colmap[]) 4199 { 4200 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 4201 PetscBool flg; 4202 PetscErrorCode ierr; 4203 4204 PetscFunctionBegin; 4205 ierr = PetscObjectTypeCompare((PetscObject)A,MATMPIAIJ,&flg);CHKERRQ(ierr); 4206 if (!flg) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"This function requires a MATMPIAIJ matrix as input"); 4207 if (Ad) *Ad = a->A; 4208 if (Ao) *Ao = a->B; 4209 if (colmap) *colmap = a->garray; 4210 PetscFunctionReturn(0); 4211 } 4212 4213 PetscErrorCode MatCreateMPIMatConcatenateSeqMat_MPIAIJ(MPI_Comm comm,Mat inmat,PetscInt n,MatReuse scall,Mat *outmat) 4214 { 4215 PetscErrorCode ierr; 4216 PetscInt m,N,i,rstart,nnz,Ii; 4217 PetscInt *indx; 4218 PetscScalar *values; 4219 4220 PetscFunctionBegin; 4221 ierr = MatGetSize(inmat,&m,&N);CHKERRQ(ierr); 4222 if (scall == MAT_INITIAL_MATRIX) { /* symbolic phase */ 4223 PetscInt *dnz,*onz,sum,bs,cbs; 4224 4225 if (n == PETSC_DECIDE) { 4226 ierr = PetscSplitOwnership(comm,&n,&N);CHKERRQ(ierr); 4227 } 4228 /* Check sum(n) = N */ 4229 ierr = MPIU_Allreduce(&n,&sum,1,MPIU_INT,MPI_SUM,comm);CHKERRQ(ierr); 4230 if (sum != N) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_INCOMP,"Sum of local columns != global columns %d",N); 4231 4232 ierr = MPI_Scan(&m, &rstart,1,MPIU_INT,MPI_SUM,comm);CHKERRQ(ierr); 4233 rstart -= m; 4234 4235 ierr = MatPreallocateInitialize(comm,m,n,dnz,onz);CHKERRQ(ierr); 4236 for (i=0; i<m; i++) { 4237 ierr = MatGetRow_SeqAIJ(inmat,i,&nnz,&indx,NULL);CHKERRQ(ierr); 4238 ierr = MatPreallocateSet(i+rstart,nnz,indx,dnz,onz);CHKERRQ(ierr); 4239 ierr = MatRestoreRow_SeqAIJ(inmat,i,&nnz,&indx,NULL);CHKERRQ(ierr); 4240 } 4241 4242 ierr = MatCreate(comm,outmat);CHKERRQ(ierr); 4243 ierr = MatSetSizes(*outmat,m,n,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 4244 ierr = MatGetBlockSizes(inmat,&bs,&cbs);CHKERRQ(ierr); 4245 ierr = MatSetBlockSizes(*outmat,bs,cbs);CHKERRQ(ierr); 4246 ierr = MatSetType(*outmat,MATAIJ);CHKERRQ(ierr); 4247 ierr = MatSeqAIJSetPreallocation(*outmat,0,dnz);CHKERRQ(ierr); 4248 ierr = MatMPIAIJSetPreallocation(*outmat,0,dnz,0,onz);CHKERRQ(ierr); 4249 ierr = MatPreallocateFinalize(dnz,onz);CHKERRQ(ierr); 4250 } 4251 4252 /* numeric phase */ 4253 ierr = MatGetOwnershipRange(*outmat,&rstart,NULL);CHKERRQ(ierr); 4254 for (i=0; i<m; i++) { 4255 ierr = MatGetRow_SeqAIJ(inmat,i,&nnz,&indx,&values);CHKERRQ(ierr); 4256 Ii = i + rstart; 4257 ierr = MatSetValues(*outmat,1,&Ii,nnz,indx,values,INSERT_VALUES);CHKERRQ(ierr); 4258 ierr = MatRestoreRow_SeqAIJ(inmat,i,&nnz,&indx,&values);CHKERRQ(ierr); 4259 } 4260 ierr = MatAssemblyBegin(*outmat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4261 ierr = MatAssemblyEnd(*outmat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4262 PetscFunctionReturn(0); 4263 } 4264 4265 PetscErrorCode MatFileSplit(Mat A,char *outfile) 4266 { 4267 PetscErrorCode ierr; 4268 PetscMPIInt rank; 4269 PetscInt m,N,i,rstart,nnz; 4270 size_t len; 4271 const PetscInt *indx; 4272 PetscViewer out; 4273 char *name; 4274 Mat B; 4275 const PetscScalar *values; 4276 4277 PetscFunctionBegin; 4278 ierr = MatGetLocalSize(A,&m,0);CHKERRQ(ierr); 4279 ierr = MatGetSize(A,0,&N);CHKERRQ(ierr); 4280 /* Should this be the type of the diagonal block of A? */ 4281 ierr = MatCreate(PETSC_COMM_SELF,&B);CHKERRQ(ierr); 4282 ierr = MatSetSizes(B,m,N,m,N);CHKERRQ(ierr); 4283 ierr = MatSetBlockSizesFromMats(B,A,A);CHKERRQ(ierr); 4284 ierr = MatSetType(B,MATSEQAIJ);CHKERRQ(ierr); 4285 ierr = MatSeqAIJSetPreallocation(B,0,NULL);CHKERRQ(ierr); 4286 ierr = MatGetOwnershipRange(A,&rstart,0);CHKERRQ(ierr); 4287 for (i=0; i<m; i++) { 4288 ierr = MatGetRow(A,i+rstart,&nnz,&indx,&values);CHKERRQ(ierr); 4289 ierr = MatSetValues(B,1,&i,nnz,indx,values,INSERT_VALUES);CHKERRQ(ierr); 4290 ierr = MatRestoreRow(A,i+rstart,&nnz,&indx,&values);CHKERRQ(ierr); 4291 } 4292 ierr = MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4293 ierr = MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4294 4295 ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)A),&rank);CHKERRQ(ierr); 4296 ierr = PetscStrlen(outfile,&len);CHKERRQ(ierr); 4297 ierr = PetscMalloc1(len+5,&name);CHKERRQ(ierr); 4298 sprintf(name,"%s.%d",outfile,rank); 4299 ierr = PetscViewerBinaryOpen(PETSC_COMM_SELF,name,FILE_MODE_APPEND,&out);CHKERRQ(ierr); 4300 ierr = PetscFree(name);CHKERRQ(ierr); 4301 ierr = MatView(B,out);CHKERRQ(ierr); 4302 ierr = PetscViewerDestroy(&out);CHKERRQ(ierr); 4303 ierr = MatDestroy(&B);CHKERRQ(ierr); 4304 PetscFunctionReturn(0); 4305 } 4306 4307 PetscErrorCode MatDestroy_MPIAIJ_SeqsToMPI(Mat A) 4308 { 4309 PetscErrorCode ierr; 4310 Mat_Merge_SeqsToMPI *merge; 4311 PetscContainer container; 4312 4313 PetscFunctionBegin; 4314 ierr = PetscObjectQuery((PetscObject)A,"MatMergeSeqsToMPI",(PetscObject*)&container);CHKERRQ(ierr); 4315 if (container) { 4316 ierr = PetscContainerGetPointer(container,(void**)&merge);CHKERRQ(ierr); 4317 ierr = PetscFree(merge->id_r);CHKERRQ(ierr); 4318 ierr = PetscFree(merge->len_s);CHKERRQ(ierr); 4319 ierr = PetscFree(merge->len_r);CHKERRQ(ierr); 4320 ierr = PetscFree(merge->bi);CHKERRQ(ierr); 4321 ierr = PetscFree(merge->bj);CHKERRQ(ierr); 4322 ierr = PetscFree(merge->buf_ri[0]);CHKERRQ(ierr); 4323 ierr = PetscFree(merge->buf_ri);CHKERRQ(ierr); 4324 ierr = PetscFree(merge->buf_rj[0]);CHKERRQ(ierr); 4325 ierr = PetscFree(merge->buf_rj);CHKERRQ(ierr); 4326 ierr = PetscFree(merge->coi);CHKERRQ(ierr); 4327 ierr = PetscFree(merge->coj);CHKERRQ(ierr); 4328 ierr = PetscFree(merge->owners_co);CHKERRQ(ierr); 4329 ierr = PetscLayoutDestroy(&merge->rowmap);CHKERRQ(ierr); 4330 ierr = PetscFree(merge);CHKERRQ(ierr); 4331 ierr = PetscObjectCompose((PetscObject)A,"MatMergeSeqsToMPI",0);CHKERRQ(ierr); 4332 } 4333 ierr = MatDestroy_MPIAIJ(A);CHKERRQ(ierr); 4334 PetscFunctionReturn(0); 4335 } 4336 4337 #include <../src/mat/utils/freespace.h> 4338 #include <petscbt.h> 4339 4340 PetscErrorCode MatCreateMPIAIJSumSeqAIJNumeric(Mat seqmat,Mat mpimat) 4341 { 4342 PetscErrorCode ierr; 4343 MPI_Comm comm; 4344 Mat_SeqAIJ *a =(Mat_SeqAIJ*)seqmat->data; 4345 PetscMPIInt size,rank,taga,*len_s; 4346 PetscInt N=mpimat->cmap->N,i,j,*owners,*ai=a->i,*aj; 4347 PetscInt proc,m; 4348 PetscInt **buf_ri,**buf_rj; 4349 PetscInt k,anzi,*bj_i,*bi,*bj,arow,bnzi,nextaj; 4350 PetscInt nrows,**buf_ri_k,**nextrow,**nextai; 4351 MPI_Request *s_waits,*r_waits; 4352 MPI_Status *status; 4353 MatScalar *aa=a->a; 4354 MatScalar **abuf_r,*ba_i; 4355 Mat_Merge_SeqsToMPI *merge; 4356 PetscContainer container; 4357 4358 PetscFunctionBegin; 4359 ierr = PetscObjectGetComm((PetscObject)mpimat,&comm);CHKERRQ(ierr); 4360 ierr = PetscLogEventBegin(MAT_Seqstompinum,seqmat,0,0,0);CHKERRQ(ierr); 4361 4362 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 4363 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 4364 4365 ierr = PetscObjectQuery((PetscObject)mpimat,"MatMergeSeqsToMPI",(PetscObject*)&container);CHKERRQ(ierr); 4366 ierr = PetscContainerGetPointer(container,(void**)&merge);CHKERRQ(ierr); 4367 4368 bi = merge->bi; 4369 bj = merge->bj; 4370 buf_ri = merge->buf_ri; 4371 buf_rj = merge->buf_rj; 4372 4373 ierr = PetscMalloc1(size,&status);CHKERRQ(ierr); 4374 owners = merge->rowmap->range; 4375 len_s = merge->len_s; 4376 4377 /* send and recv matrix values */ 4378 /*-----------------------------*/ 4379 ierr = PetscObjectGetNewTag((PetscObject)mpimat,&taga);CHKERRQ(ierr); 4380 ierr = PetscPostIrecvScalar(comm,taga,merge->nrecv,merge->id_r,merge->len_r,&abuf_r,&r_waits);CHKERRQ(ierr); 4381 4382 ierr = PetscMalloc1(merge->nsend+1,&s_waits);CHKERRQ(ierr); 4383 for (proc=0,k=0; proc<size; proc++) { 4384 if (!len_s[proc]) continue; 4385 i = owners[proc]; 4386 ierr = MPI_Isend(aa+ai[i],len_s[proc],MPIU_MATSCALAR,proc,taga,comm,s_waits+k);CHKERRQ(ierr); 4387 k++; 4388 } 4389 4390 if (merge->nrecv) {ierr = MPI_Waitall(merge->nrecv,r_waits,status);CHKERRQ(ierr);} 4391 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,s_waits,status);CHKERRQ(ierr);} 4392 ierr = PetscFree(status);CHKERRQ(ierr); 4393 4394 ierr = PetscFree(s_waits);CHKERRQ(ierr); 4395 ierr = PetscFree(r_waits);CHKERRQ(ierr); 4396 4397 /* insert mat values of mpimat */ 4398 /*----------------------------*/ 4399 ierr = PetscMalloc1(N,&ba_i);CHKERRQ(ierr); 4400 ierr = PetscMalloc3(merge->nrecv,&buf_ri_k,merge->nrecv,&nextrow,merge->nrecv,&nextai);CHKERRQ(ierr); 4401 4402 for (k=0; k<merge->nrecv; k++) { 4403 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 4404 nrows = *(buf_ri_k[k]); 4405 nextrow[k] = buf_ri_k[k]+1; /* next row number of k-th recved i-structure */ 4406 nextai[k] = buf_ri_k[k] + (nrows + 1); /* poins to the next i-structure of k-th recved i-structure */ 4407 } 4408 4409 /* set values of ba */ 4410 m = merge->rowmap->n; 4411 for (i=0; i<m; i++) { 4412 arow = owners[rank] + i; 4413 bj_i = bj+bi[i]; /* col indices of the i-th row of mpimat */ 4414 bnzi = bi[i+1] - bi[i]; 4415 ierr = PetscMemzero(ba_i,bnzi*sizeof(PetscScalar));CHKERRQ(ierr); 4416 4417 /* add local non-zero vals of this proc's seqmat into ba */ 4418 anzi = ai[arow+1] - ai[arow]; 4419 aj = a->j + ai[arow]; 4420 aa = a->a + ai[arow]; 4421 nextaj = 0; 4422 for (j=0; nextaj<anzi; j++) { 4423 if (*(bj_i + j) == aj[nextaj]) { /* bcol == acol */ 4424 ba_i[j] += aa[nextaj++]; 4425 } 4426 } 4427 4428 /* add received vals into ba */ 4429 for (k=0; k<merge->nrecv; k++) { /* k-th received message */ 4430 /* i-th row */ 4431 if (i == *nextrow[k]) { 4432 anzi = *(nextai[k]+1) - *nextai[k]; 4433 aj = buf_rj[k] + *(nextai[k]); 4434 aa = abuf_r[k] + *(nextai[k]); 4435 nextaj = 0; 4436 for (j=0; nextaj<anzi; j++) { 4437 if (*(bj_i + j) == aj[nextaj]) { /* bcol == acol */ 4438 ba_i[j] += aa[nextaj++]; 4439 } 4440 } 4441 nextrow[k]++; nextai[k]++; 4442 } 4443 } 4444 ierr = MatSetValues(mpimat,1,&arow,bnzi,bj_i,ba_i,INSERT_VALUES);CHKERRQ(ierr); 4445 } 4446 ierr = MatAssemblyBegin(mpimat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4447 ierr = MatAssemblyEnd(mpimat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4448 4449 ierr = PetscFree(abuf_r[0]);CHKERRQ(ierr); 4450 ierr = PetscFree(abuf_r);CHKERRQ(ierr); 4451 ierr = PetscFree(ba_i);CHKERRQ(ierr); 4452 ierr = PetscFree3(buf_ri_k,nextrow,nextai);CHKERRQ(ierr); 4453 ierr = PetscLogEventEnd(MAT_Seqstompinum,seqmat,0,0,0);CHKERRQ(ierr); 4454 PetscFunctionReturn(0); 4455 } 4456 4457 PetscErrorCode MatCreateMPIAIJSumSeqAIJSymbolic(MPI_Comm comm,Mat seqmat,PetscInt m,PetscInt n,Mat *mpimat) 4458 { 4459 PetscErrorCode ierr; 4460 Mat B_mpi; 4461 Mat_SeqAIJ *a=(Mat_SeqAIJ*)seqmat->data; 4462 PetscMPIInt size,rank,tagi,tagj,*len_s,*len_si,*len_ri; 4463 PetscInt **buf_rj,**buf_ri,**buf_ri_k; 4464 PetscInt M=seqmat->rmap->n,N=seqmat->cmap->n,i,*owners,*ai=a->i,*aj=a->j; 4465 PetscInt len,proc,*dnz,*onz,bs,cbs; 4466 PetscInt k,anzi,*bi,*bj,*lnk,nlnk,arow,bnzi,nspacedouble=0; 4467 PetscInt nrows,*buf_s,*buf_si,*buf_si_i,**nextrow,**nextai; 4468 MPI_Request *si_waits,*sj_waits,*ri_waits,*rj_waits; 4469 MPI_Status *status; 4470 PetscFreeSpaceList free_space=NULL,current_space=NULL; 4471 PetscBT lnkbt; 4472 Mat_Merge_SeqsToMPI *merge; 4473 PetscContainer container; 4474 4475 PetscFunctionBegin; 4476 ierr = PetscLogEventBegin(MAT_Seqstompisym,seqmat,0,0,0);CHKERRQ(ierr); 4477 4478 /* make sure it is a PETSc comm */ 4479 ierr = PetscCommDuplicate(comm,&comm,NULL);CHKERRQ(ierr); 4480 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 4481 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 4482 4483 ierr = PetscNew(&merge);CHKERRQ(ierr); 4484 ierr = PetscMalloc1(size,&status);CHKERRQ(ierr); 4485 4486 /* determine row ownership */ 4487 /*---------------------------------------------------------*/ 4488 ierr = PetscLayoutCreate(comm,&merge->rowmap);CHKERRQ(ierr); 4489 ierr = PetscLayoutSetLocalSize(merge->rowmap,m);CHKERRQ(ierr); 4490 ierr = PetscLayoutSetSize(merge->rowmap,M);CHKERRQ(ierr); 4491 ierr = PetscLayoutSetBlockSize(merge->rowmap,1);CHKERRQ(ierr); 4492 ierr = PetscLayoutSetUp(merge->rowmap);CHKERRQ(ierr); 4493 ierr = PetscMalloc1(size,&len_si);CHKERRQ(ierr); 4494 ierr = PetscMalloc1(size,&merge->len_s);CHKERRQ(ierr); 4495 4496 m = merge->rowmap->n; 4497 owners = merge->rowmap->range; 4498 4499 /* determine the number of messages to send, their lengths */ 4500 /*---------------------------------------------------------*/ 4501 len_s = merge->len_s; 4502 4503 len = 0; /* length of buf_si[] */ 4504 merge->nsend = 0; 4505 for (proc=0; proc<size; proc++) { 4506 len_si[proc] = 0; 4507 if (proc == rank) { 4508 len_s[proc] = 0; 4509 } else { 4510 len_si[proc] = owners[proc+1] - owners[proc] + 1; 4511 len_s[proc] = ai[owners[proc+1]] - ai[owners[proc]]; /* num of rows to be sent to [proc] */ 4512 } 4513 if (len_s[proc]) { 4514 merge->nsend++; 4515 nrows = 0; 4516 for (i=owners[proc]; i<owners[proc+1]; i++) { 4517 if (ai[i+1] > ai[i]) nrows++; 4518 } 4519 len_si[proc] = 2*(nrows+1); 4520 len += len_si[proc]; 4521 } 4522 } 4523 4524 /* determine the number and length of messages to receive for ij-structure */ 4525 /*-------------------------------------------------------------------------*/ 4526 ierr = PetscGatherNumberOfMessages(comm,NULL,len_s,&merge->nrecv);CHKERRQ(ierr); 4527 ierr = PetscGatherMessageLengths2(comm,merge->nsend,merge->nrecv,len_s,len_si,&merge->id_r,&merge->len_r,&len_ri);CHKERRQ(ierr); 4528 4529 /* post the Irecv of j-structure */ 4530 /*-------------------------------*/ 4531 ierr = PetscCommGetNewTag(comm,&tagj);CHKERRQ(ierr); 4532 ierr = PetscPostIrecvInt(comm,tagj,merge->nrecv,merge->id_r,merge->len_r,&buf_rj,&rj_waits);CHKERRQ(ierr); 4533 4534 /* post the Isend of j-structure */ 4535 /*--------------------------------*/ 4536 ierr = PetscMalloc2(merge->nsend,&si_waits,merge->nsend,&sj_waits);CHKERRQ(ierr); 4537 4538 for (proc=0, k=0; proc<size; proc++) { 4539 if (!len_s[proc]) continue; 4540 i = owners[proc]; 4541 ierr = MPI_Isend(aj+ai[i],len_s[proc],MPIU_INT,proc,tagj,comm,sj_waits+k);CHKERRQ(ierr); 4542 k++; 4543 } 4544 4545 /* receives and sends of j-structure are complete */ 4546 /*------------------------------------------------*/ 4547 if (merge->nrecv) {ierr = MPI_Waitall(merge->nrecv,rj_waits,status);CHKERRQ(ierr);} 4548 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,sj_waits,status);CHKERRQ(ierr);} 4549 4550 /* send and recv i-structure */ 4551 /*---------------------------*/ 4552 ierr = PetscCommGetNewTag(comm,&tagi);CHKERRQ(ierr); 4553 ierr = PetscPostIrecvInt(comm,tagi,merge->nrecv,merge->id_r,len_ri,&buf_ri,&ri_waits);CHKERRQ(ierr); 4554 4555 ierr = PetscMalloc1(len+1,&buf_s);CHKERRQ(ierr); 4556 buf_si = buf_s; /* points to the beginning of k-th msg to be sent */ 4557 for (proc=0,k=0; proc<size; proc++) { 4558 if (!len_s[proc]) continue; 4559 /* form outgoing message for i-structure: 4560 buf_si[0]: nrows to be sent 4561 [1:nrows]: row index (global) 4562 [nrows+1:2*nrows+1]: i-structure index 4563 */ 4564 /*-------------------------------------------*/ 4565 nrows = len_si[proc]/2 - 1; 4566 buf_si_i = buf_si + nrows+1; 4567 buf_si[0] = nrows; 4568 buf_si_i[0] = 0; 4569 nrows = 0; 4570 for (i=owners[proc]; i<owners[proc+1]; i++) { 4571 anzi = ai[i+1] - ai[i]; 4572 if (anzi) { 4573 buf_si_i[nrows+1] = buf_si_i[nrows] + anzi; /* i-structure */ 4574 buf_si[nrows+1] = i-owners[proc]; /* local row index */ 4575 nrows++; 4576 } 4577 } 4578 ierr = MPI_Isend(buf_si,len_si[proc],MPIU_INT,proc,tagi,comm,si_waits+k);CHKERRQ(ierr); 4579 k++; 4580 buf_si += len_si[proc]; 4581 } 4582 4583 if (merge->nrecv) {ierr = MPI_Waitall(merge->nrecv,ri_waits,status);CHKERRQ(ierr);} 4584 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,si_waits,status);CHKERRQ(ierr);} 4585 4586 ierr = PetscInfo2(seqmat,"nsend: %D, nrecv: %D\n",merge->nsend,merge->nrecv);CHKERRQ(ierr); 4587 for (i=0; i<merge->nrecv; i++) { 4588 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); 4589 } 4590 4591 ierr = PetscFree(len_si);CHKERRQ(ierr); 4592 ierr = PetscFree(len_ri);CHKERRQ(ierr); 4593 ierr = PetscFree(rj_waits);CHKERRQ(ierr); 4594 ierr = PetscFree2(si_waits,sj_waits);CHKERRQ(ierr); 4595 ierr = PetscFree(ri_waits);CHKERRQ(ierr); 4596 ierr = PetscFree(buf_s);CHKERRQ(ierr); 4597 ierr = PetscFree(status);CHKERRQ(ierr); 4598 4599 /* compute a local seq matrix in each processor */ 4600 /*----------------------------------------------*/ 4601 /* allocate bi array and free space for accumulating nonzero column info */ 4602 ierr = PetscMalloc1(m+1,&bi);CHKERRQ(ierr); 4603 bi[0] = 0; 4604 4605 /* create and initialize a linked list */ 4606 nlnk = N+1; 4607 ierr = PetscLLCreate(N,N,nlnk,lnk,lnkbt);CHKERRQ(ierr); 4608 4609 /* initial FreeSpace size is 2*(num of local nnz(seqmat)) */ 4610 len = ai[owners[rank+1]] - ai[owners[rank]]; 4611 ierr = PetscFreeSpaceGet(PetscIntMultTruncate(2,len)+1,&free_space);CHKERRQ(ierr); 4612 4613 current_space = free_space; 4614 4615 /* determine symbolic info for each local row */ 4616 ierr = PetscMalloc3(merge->nrecv,&buf_ri_k,merge->nrecv,&nextrow,merge->nrecv,&nextai);CHKERRQ(ierr); 4617 4618 for (k=0; k<merge->nrecv; k++) { 4619 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 4620 nrows = *buf_ri_k[k]; 4621 nextrow[k] = buf_ri_k[k] + 1; /* next row number of k-th recved i-structure */ 4622 nextai[k] = buf_ri_k[k] + (nrows + 1); /* poins to the next i-structure of k-th recved i-structure */ 4623 } 4624 4625 ierr = MatPreallocateInitialize(comm,m,n,dnz,onz);CHKERRQ(ierr); 4626 len = 0; 4627 for (i=0; i<m; i++) { 4628 bnzi = 0; 4629 /* add local non-zero cols of this proc's seqmat into lnk */ 4630 arow = owners[rank] + i; 4631 anzi = ai[arow+1] - ai[arow]; 4632 aj = a->j + ai[arow]; 4633 ierr = PetscLLAddSorted(anzi,aj,N,nlnk,lnk,lnkbt);CHKERRQ(ierr); 4634 bnzi += nlnk; 4635 /* add received col data into lnk */ 4636 for (k=0; k<merge->nrecv; k++) { /* k-th received message */ 4637 if (i == *nextrow[k]) { /* i-th row */ 4638 anzi = *(nextai[k]+1) - *nextai[k]; 4639 aj = buf_rj[k] + *nextai[k]; 4640 ierr = PetscLLAddSorted(anzi,aj,N,nlnk,lnk,lnkbt);CHKERRQ(ierr); 4641 bnzi += nlnk; 4642 nextrow[k]++; nextai[k]++; 4643 } 4644 } 4645 if (len < bnzi) len = bnzi; /* =max(bnzi) */ 4646 4647 /* if free space is not available, make more free space */ 4648 if (current_space->local_remaining<bnzi) { 4649 ierr = PetscFreeSpaceGet(PetscIntSumTruncate(bnzi,current_space->total_array_size),¤t_space);CHKERRQ(ierr); 4650 nspacedouble++; 4651 } 4652 /* copy data into free space, then initialize lnk */ 4653 ierr = PetscLLClean(N,N,bnzi,lnk,current_space->array,lnkbt);CHKERRQ(ierr); 4654 ierr = MatPreallocateSet(i+owners[rank],bnzi,current_space->array,dnz,onz);CHKERRQ(ierr); 4655 4656 current_space->array += bnzi; 4657 current_space->local_used += bnzi; 4658 current_space->local_remaining -= bnzi; 4659 4660 bi[i+1] = bi[i] + bnzi; 4661 } 4662 4663 ierr = PetscFree3(buf_ri_k,nextrow,nextai);CHKERRQ(ierr); 4664 4665 ierr = PetscMalloc1(bi[m]+1,&bj);CHKERRQ(ierr); 4666 ierr = PetscFreeSpaceContiguous(&free_space,bj);CHKERRQ(ierr); 4667 ierr = PetscLLDestroy(lnk,lnkbt);CHKERRQ(ierr); 4668 4669 /* create symbolic parallel matrix B_mpi */ 4670 /*---------------------------------------*/ 4671 ierr = MatGetBlockSizes(seqmat,&bs,&cbs);CHKERRQ(ierr); 4672 ierr = MatCreate(comm,&B_mpi);CHKERRQ(ierr); 4673 if (n==PETSC_DECIDE) { 4674 ierr = MatSetSizes(B_mpi,m,n,PETSC_DETERMINE,N);CHKERRQ(ierr); 4675 } else { 4676 ierr = MatSetSizes(B_mpi,m,n,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 4677 } 4678 ierr = MatSetBlockSizes(B_mpi,bs,cbs);CHKERRQ(ierr); 4679 ierr = MatSetType(B_mpi,MATMPIAIJ);CHKERRQ(ierr); 4680 ierr = MatMPIAIJSetPreallocation(B_mpi,0,dnz,0,onz);CHKERRQ(ierr); 4681 ierr = MatPreallocateFinalize(dnz,onz);CHKERRQ(ierr); 4682 ierr = MatSetOption(B_mpi,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 4683 4684 /* B_mpi is not ready for use - assembly will be done by MatCreateMPIAIJSumSeqAIJNumeric() */ 4685 B_mpi->assembled = PETSC_FALSE; 4686 B_mpi->ops->destroy = MatDestroy_MPIAIJ_SeqsToMPI; 4687 merge->bi = bi; 4688 merge->bj = bj; 4689 merge->buf_ri = buf_ri; 4690 merge->buf_rj = buf_rj; 4691 merge->coi = NULL; 4692 merge->coj = NULL; 4693 merge->owners_co = NULL; 4694 4695 ierr = PetscCommDestroy(&comm);CHKERRQ(ierr); 4696 4697 /* attach the supporting struct to B_mpi for reuse */ 4698 ierr = PetscContainerCreate(PETSC_COMM_SELF,&container);CHKERRQ(ierr); 4699 ierr = PetscContainerSetPointer(container,merge);CHKERRQ(ierr); 4700 ierr = PetscObjectCompose((PetscObject)B_mpi,"MatMergeSeqsToMPI",(PetscObject)container);CHKERRQ(ierr); 4701 ierr = PetscContainerDestroy(&container);CHKERRQ(ierr); 4702 *mpimat = B_mpi; 4703 4704 ierr = PetscLogEventEnd(MAT_Seqstompisym,seqmat,0,0,0);CHKERRQ(ierr); 4705 PetscFunctionReturn(0); 4706 } 4707 4708 /*@C 4709 MatCreateMPIAIJSumSeqAIJ - Creates a MATMPIAIJ matrix by adding sequential 4710 matrices from each processor 4711 4712 Collective on MPI_Comm 4713 4714 Input Parameters: 4715 + comm - the communicators the parallel matrix will live on 4716 . seqmat - the input sequential matrices 4717 . m - number of local rows (or PETSC_DECIDE) 4718 . n - number of local columns (or PETSC_DECIDE) 4719 - scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX 4720 4721 Output Parameter: 4722 . mpimat - the parallel matrix generated 4723 4724 Level: advanced 4725 4726 Notes: 4727 The dimensions of the sequential matrix in each processor MUST be the same. 4728 The input seqmat is included into the container "Mat_Merge_SeqsToMPI", and will be 4729 destroyed when mpimat is destroyed. Call PetscObjectQuery() to access seqmat. 4730 @*/ 4731 PetscErrorCode MatCreateMPIAIJSumSeqAIJ(MPI_Comm comm,Mat seqmat,PetscInt m,PetscInt n,MatReuse scall,Mat *mpimat) 4732 { 4733 PetscErrorCode ierr; 4734 PetscMPIInt size; 4735 4736 PetscFunctionBegin; 4737 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 4738 if (size == 1) { 4739 ierr = PetscLogEventBegin(MAT_Seqstompi,seqmat,0,0,0);CHKERRQ(ierr); 4740 if (scall == MAT_INITIAL_MATRIX) { 4741 ierr = MatDuplicate(seqmat,MAT_COPY_VALUES,mpimat);CHKERRQ(ierr); 4742 } else { 4743 ierr = MatCopy(seqmat,*mpimat,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 4744 } 4745 ierr = PetscLogEventEnd(MAT_Seqstompi,seqmat,0,0,0);CHKERRQ(ierr); 4746 PetscFunctionReturn(0); 4747 } 4748 ierr = PetscLogEventBegin(MAT_Seqstompi,seqmat,0,0,0);CHKERRQ(ierr); 4749 if (scall == MAT_INITIAL_MATRIX) { 4750 ierr = MatCreateMPIAIJSumSeqAIJSymbolic(comm,seqmat,m,n,mpimat);CHKERRQ(ierr); 4751 } 4752 ierr = MatCreateMPIAIJSumSeqAIJNumeric(seqmat,*mpimat);CHKERRQ(ierr); 4753 ierr = PetscLogEventEnd(MAT_Seqstompi,seqmat,0,0,0);CHKERRQ(ierr); 4754 PetscFunctionReturn(0); 4755 } 4756 4757 /*@ 4758 MatMPIAIJGetLocalMat - Creates a SeqAIJ from a MATMPIAIJ matrix by taking all its local rows and putting them into a sequential vector with 4759 mlocal rows and n columns. Where mlocal is the row count obtained with MatGetLocalSize() and n is the global column count obtained 4760 with MatGetSize() 4761 4762 Not Collective 4763 4764 Input Parameters: 4765 + A - the matrix 4766 . scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX 4767 4768 Output Parameter: 4769 . A_loc - the local sequential matrix generated 4770 4771 Level: developer 4772 4773 .seealso: MatGetOwnerShipRange(), MatMPIAIJGetLocalMatCondensed() 4774 4775 @*/ 4776 PetscErrorCode MatMPIAIJGetLocalMat(Mat A,MatReuse scall,Mat *A_loc) 4777 { 4778 PetscErrorCode ierr; 4779 Mat_MPIAIJ *mpimat=(Mat_MPIAIJ*)A->data; 4780 Mat_SeqAIJ *mat,*a,*b; 4781 PetscInt *ai,*aj,*bi,*bj,*cmap=mpimat->garray; 4782 MatScalar *aa,*ba,*cam; 4783 PetscScalar *ca; 4784 PetscInt am=A->rmap->n,i,j,k,cstart=A->cmap->rstart; 4785 PetscInt *ci,*cj,col,ncols_d,ncols_o,jo; 4786 PetscBool match; 4787 MPI_Comm comm; 4788 PetscMPIInt size; 4789 4790 PetscFunctionBegin; 4791 ierr = PetscObjectTypeCompare((PetscObject)A,MATMPIAIJ,&match);CHKERRQ(ierr); 4792 if (!match) SETERRQ(PetscObjectComm((PetscObject)A), PETSC_ERR_SUP,"Requires MATMPIAIJ matrix as input"); 4793 ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 4794 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 4795 if (size == 1 && scall == MAT_REUSE_MATRIX) PetscFunctionReturn(0); 4796 4797 ierr = PetscLogEventBegin(MAT_Getlocalmat,A,0,0,0);CHKERRQ(ierr); 4798 a = (Mat_SeqAIJ*)(mpimat->A)->data; 4799 b = (Mat_SeqAIJ*)(mpimat->B)->data; 4800 ai = a->i; aj = a->j; bi = b->i; bj = b->j; 4801 aa = a->a; ba = b->a; 4802 if (scall == MAT_INITIAL_MATRIX) { 4803 if (size == 1) { 4804 ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,am,A->cmap->N,ai,aj,aa,A_loc);CHKERRQ(ierr); 4805 PetscFunctionReturn(0); 4806 } 4807 4808 ierr = PetscMalloc1(1+am,&ci);CHKERRQ(ierr); 4809 ci[0] = 0; 4810 for (i=0; i<am; i++) { 4811 ci[i+1] = ci[i] + (ai[i+1] - ai[i]) + (bi[i+1] - bi[i]); 4812 } 4813 ierr = PetscMalloc1(1+ci[am],&cj);CHKERRQ(ierr); 4814 ierr = PetscMalloc1(1+ci[am],&ca);CHKERRQ(ierr); 4815 k = 0; 4816 for (i=0; i<am; i++) { 4817 ncols_o = bi[i+1] - bi[i]; 4818 ncols_d = ai[i+1] - ai[i]; 4819 /* off-diagonal portion of A */ 4820 for (jo=0; jo<ncols_o; jo++) { 4821 col = cmap[*bj]; 4822 if (col >= cstart) break; 4823 cj[k] = col; bj++; 4824 ca[k++] = *ba++; 4825 } 4826 /* diagonal portion of A */ 4827 for (j=0; j<ncols_d; j++) { 4828 cj[k] = cstart + *aj++; 4829 ca[k++] = *aa++; 4830 } 4831 /* off-diagonal portion of A */ 4832 for (j=jo; j<ncols_o; j++) { 4833 cj[k] = cmap[*bj++]; 4834 ca[k++] = *ba++; 4835 } 4836 } 4837 /* put together the new matrix */ 4838 ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,am,A->cmap->N,ci,cj,ca,A_loc);CHKERRQ(ierr); 4839 /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */ 4840 /* Since these are PETSc arrays, change flags to free them as necessary. */ 4841 mat = (Mat_SeqAIJ*)(*A_loc)->data; 4842 mat->free_a = PETSC_TRUE; 4843 mat->free_ij = PETSC_TRUE; 4844 mat->nonew = 0; 4845 } else if (scall == MAT_REUSE_MATRIX) { 4846 mat=(Mat_SeqAIJ*)(*A_loc)->data; 4847 ci = mat->i; cj = mat->j; cam = mat->a; 4848 for (i=0; i<am; i++) { 4849 /* off-diagonal portion of A */ 4850 ncols_o = bi[i+1] - bi[i]; 4851 for (jo=0; jo<ncols_o; jo++) { 4852 col = cmap[*bj]; 4853 if (col >= cstart) break; 4854 *cam++ = *ba++; bj++; 4855 } 4856 /* diagonal portion of A */ 4857 ncols_d = ai[i+1] - ai[i]; 4858 for (j=0; j<ncols_d; j++) *cam++ = *aa++; 4859 /* off-diagonal portion of A */ 4860 for (j=jo; j<ncols_o; j++) { 4861 *cam++ = *ba++; bj++; 4862 } 4863 } 4864 } else SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Invalid MatReuse %d",(int)scall); 4865 ierr = PetscLogEventEnd(MAT_Getlocalmat,A,0,0,0);CHKERRQ(ierr); 4866 PetscFunctionReturn(0); 4867 } 4868 4869 /*@C 4870 MatMPIAIJGetLocalMatCondensed - Creates a SeqAIJ matrix from an MATMPIAIJ matrix by taking all its local rows and NON-ZERO columns 4871 4872 Not Collective 4873 4874 Input Parameters: 4875 + A - the matrix 4876 . scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX 4877 - row, col - index sets of rows and columns to extract (or NULL) 4878 4879 Output Parameter: 4880 . A_loc - the local sequential matrix generated 4881 4882 Level: developer 4883 4884 .seealso: MatGetOwnershipRange(), MatMPIAIJGetLocalMat() 4885 4886 @*/ 4887 PetscErrorCode MatMPIAIJGetLocalMatCondensed(Mat A,MatReuse scall,IS *row,IS *col,Mat *A_loc) 4888 { 4889 Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data; 4890 PetscErrorCode ierr; 4891 PetscInt i,start,end,ncols,nzA,nzB,*cmap,imark,*idx; 4892 IS isrowa,iscola; 4893 Mat *aloc; 4894 PetscBool match; 4895 4896 PetscFunctionBegin; 4897 ierr = PetscObjectTypeCompare((PetscObject)A,MATMPIAIJ,&match);CHKERRQ(ierr); 4898 if (!match) SETERRQ(PetscObjectComm((PetscObject)A), PETSC_ERR_SUP,"Requires MATMPIAIJ matrix as input"); 4899 ierr = PetscLogEventBegin(MAT_Getlocalmatcondensed,A,0,0,0);CHKERRQ(ierr); 4900 if (!row) { 4901 start = A->rmap->rstart; end = A->rmap->rend; 4902 ierr = ISCreateStride(PETSC_COMM_SELF,end-start,start,1,&isrowa);CHKERRQ(ierr); 4903 } else { 4904 isrowa = *row; 4905 } 4906 if (!col) { 4907 start = A->cmap->rstart; 4908 cmap = a->garray; 4909 nzA = a->A->cmap->n; 4910 nzB = a->B->cmap->n; 4911 ierr = PetscMalloc1(nzA+nzB, &idx);CHKERRQ(ierr); 4912 ncols = 0; 4913 for (i=0; i<nzB; i++) { 4914 if (cmap[i] < start) idx[ncols++] = cmap[i]; 4915 else break; 4916 } 4917 imark = i; 4918 for (i=0; i<nzA; i++) idx[ncols++] = start + i; 4919 for (i=imark; i<nzB; i++) idx[ncols++] = cmap[i]; 4920 ierr = ISCreateGeneral(PETSC_COMM_SELF,ncols,idx,PETSC_OWN_POINTER,&iscola);CHKERRQ(ierr); 4921 } else { 4922 iscola = *col; 4923 } 4924 if (scall != MAT_INITIAL_MATRIX) { 4925 ierr = PetscMalloc1(1,&aloc);CHKERRQ(ierr); 4926 aloc[0] = *A_loc; 4927 } 4928 ierr = MatCreateSubMatrices(A,1,&isrowa,&iscola,scall,&aloc);CHKERRQ(ierr); 4929 *A_loc = aloc[0]; 4930 ierr = PetscFree(aloc);CHKERRQ(ierr); 4931 if (!row) { 4932 ierr = ISDestroy(&isrowa);CHKERRQ(ierr); 4933 } 4934 if (!col) { 4935 ierr = ISDestroy(&iscola);CHKERRQ(ierr); 4936 } 4937 ierr = PetscLogEventEnd(MAT_Getlocalmatcondensed,A,0,0,0);CHKERRQ(ierr); 4938 PetscFunctionReturn(0); 4939 } 4940 4941 /*@C 4942 MatGetBrowsOfAcols - Creates a SeqAIJ matrix by taking rows of B that equal to nonzero columns of local A 4943 4944 Collective on Mat 4945 4946 Input Parameters: 4947 + A,B - the matrices in mpiaij format 4948 . scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX 4949 - rowb, colb - index sets of rows and columns of B to extract (or NULL) 4950 4951 Output Parameter: 4952 + rowb, colb - index sets of rows and columns of B to extract 4953 - B_seq - the sequential matrix generated 4954 4955 Level: developer 4956 4957 @*/ 4958 PetscErrorCode MatGetBrowsOfAcols(Mat A,Mat B,MatReuse scall,IS *rowb,IS *colb,Mat *B_seq) 4959 { 4960 Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data; 4961 PetscErrorCode ierr; 4962 PetscInt *idx,i,start,ncols,nzA,nzB,*cmap,imark; 4963 IS isrowb,iscolb; 4964 Mat *bseq=NULL; 4965 4966 PetscFunctionBegin; 4967 if (A->cmap->rstart != B->rmap->rstart || A->cmap->rend != B->rmap->rend) { 4968 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); 4969 } 4970 ierr = PetscLogEventBegin(MAT_GetBrowsOfAcols,A,B,0,0);CHKERRQ(ierr); 4971 4972 if (scall == MAT_INITIAL_MATRIX) { 4973 start = A->cmap->rstart; 4974 cmap = a->garray; 4975 nzA = a->A->cmap->n; 4976 nzB = a->B->cmap->n; 4977 ierr = PetscMalloc1(nzA+nzB, &idx);CHKERRQ(ierr); 4978 ncols = 0; 4979 for (i=0; i<nzB; i++) { /* row < local row index */ 4980 if (cmap[i] < start) idx[ncols++] = cmap[i]; 4981 else break; 4982 } 4983 imark = i; 4984 for (i=0; i<nzA; i++) idx[ncols++] = start + i; /* local rows */ 4985 for (i=imark; i<nzB; i++) idx[ncols++] = cmap[i]; /* row > local row index */ 4986 ierr = ISCreateGeneral(PETSC_COMM_SELF,ncols,idx,PETSC_OWN_POINTER,&isrowb);CHKERRQ(ierr); 4987 ierr = ISCreateStride(PETSC_COMM_SELF,B->cmap->N,0,1,&iscolb);CHKERRQ(ierr); 4988 } else { 4989 if (!rowb || !colb) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"IS rowb and colb must be provided for MAT_REUSE_MATRIX"); 4990 isrowb = *rowb; iscolb = *colb; 4991 ierr = PetscMalloc1(1,&bseq);CHKERRQ(ierr); 4992 bseq[0] = *B_seq; 4993 } 4994 ierr = MatCreateSubMatrices(B,1,&isrowb,&iscolb,scall,&bseq);CHKERRQ(ierr); 4995 *B_seq = bseq[0]; 4996 ierr = PetscFree(bseq);CHKERRQ(ierr); 4997 if (!rowb) { 4998 ierr = ISDestroy(&isrowb);CHKERRQ(ierr); 4999 } else { 5000 *rowb = isrowb; 5001 } 5002 if (!colb) { 5003 ierr = ISDestroy(&iscolb);CHKERRQ(ierr); 5004 } else { 5005 *colb = iscolb; 5006 } 5007 ierr = PetscLogEventEnd(MAT_GetBrowsOfAcols,A,B,0,0);CHKERRQ(ierr); 5008 PetscFunctionReturn(0); 5009 } 5010 5011 /* 5012 MatGetBrowsOfAoCols_MPIAIJ - Creates a SeqAIJ matrix by taking rows of B that equal to nonzero columns 5013 of the OFF-DIAGONAL portion of local A 5014 5015 Collective on Mat 5016 5017 Input Parameters: 5018 + A,B - the matrices in mpiaij format 5019 - scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX 5020 5021 Output Parameter: 5022 + startsj_s - starting point in B's sending j-arrays, saved for MAT_REUSE (or NULL) 5023 . startsj_r - starting point in B's receiving j-arrays, saved for MAT_REUSE (or NULL) 5024 . bufa_ptr - array for sending matrix values, saved for MAT_REUSE (or NULL) 5025 - B_oth - the sequential matrix generated with size aBn=a->B->cmap->n by B->cmap->N 5026 5027 Level: developer 5028 5029 */ 5030 PetscErrorCode MatGetBrowsOfAoCols_MPIAIJ(Mat A,Mat B,MatReuse scall,PetscInt **startsj_s,PetscInt **startsj_r,MatScalar **bufa_ptr,Mat *B_oth) 5031 { 5032 VecScatter_MPI_General *gen_to,*gen_from; 5033 PetscErrorCode ierr; 5034 Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data; 5035 Mat_SeqAIJ *b_oth; 5036 VecScatter ctx =a->Mvctx; 5037 MPI_Comm comm; 5038 PetscMPIInt *rprocs,*sprocs,tag=((PetscObject)ctx)->tag,rank; 5039 PetscInt *rowlen,*bufj,*bufJ,ncols,aBn=a->B->cmap->n,row,*b_othi,*b_othj; 5040 PetscInt *rvalues,*svalues; 5041 MatScalar *b_otha,*bufa,*bufA; 5042 PetscInt i,j,k,l,ll,nrecvs,nsends,nrows,*srow,*rstarts,*rstartsj = 0,*sstarts,*sstartsj,len; 5043 MPI_Request *rwaits = NULL,*swaits = NULL; 5044 MPI_Status *sstatus,rstatus; 5045 PetscMPIInt jj,size; 5046 PetscInt *cols,sbs,rbs; 5047 PetscScalar *vals; 5048 5049 PetscFunctionBegin; 5050 ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 5051 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 5052 5053 if (A->cmap->rstart != B->rmap->rstart || A->cmap->rend != B->rmap->rend) { 5054 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); 5055 } 5056 ierr = PetscLogEventBegin(MAT_GetBrowsOfAocols,A,B,0,0);CHKERRQ(ierr); 5057 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 5058 5059 if (size == 1) { 5060 startsj_s = NULL; 5061 bufa_ptr = NULL; 5062 *B_oth = NULL; 5063 PetscFunctionReturn(0); 5064 } 5065 5066 gen_to = (VecScatter_MPI_General*)ctx->todata; 5067 gen_from = (VecScatter_MPI_General*)ctx->fromdata; 5068 nrecvs = gen_from->n; 5069 nsends = gen_to->n; 5070 5071 ierr = PetscMalloc2(nrecvs,&rwaits,nsends,&swaits);CHKERRQ(ierr); 5072 srow = gen_to->indices; /* local row index to be sent */ 5073 sstarts = gen_to->starts; 5074 sprocs = gen_to->procs; 5075 sstatus = gen_to->sstatus; 5076 sbs = gen_to->bs; 5077 rstarts = gen_from->starts; 5078 rprocs = gen_from->procs; 5079 rbs = gen_from->bs; 5080 5081 if (!startsj_s || !bufa_ptr) scall = MAT_INITIAL_MATRIX; 5082 if (scall == MAT_INITIAL_MATRIX) { 5083 /* i-array */ 5084 /*---------*/ 5085 /* post receives */ 5086 ierr = PetscMalloc1(rbs*(rstarts[nrecvs] - rstarts[0]),&rvalues);CHKERRQ(ierr); 5087 for (i=0; i<nrecvs; i++) { 5088 rowlen = rvalues + rstarts[i]*rbs; 5089 nrows = (rstarts[i+1]-rstarts[i])*rbs; /* num of indices to be received */ 5090 ierr = MPI_Irecv(rowlen,nrows,MPIU_INT,rprocs[i],tag,comm,rwaits+i);CHKERRQ(ierr); 5091 } 5092 5093 /* pack the outgoing message */ 5094 ierr = PetscMalloc2(nsends+1,&sstartsj,nrecvs+1,&rstartsj);CHKERRQ(ierr); 5095 5096 sstartsj[0] = 0; 5097 rstartsj[0] = 0; 5098 len = 0; /* total length of j or a array to be sent */ 5099 k = 0; 5100 ierr = PetscMalloc1(sbs*(sstarts[nsends] - sstarts[0]),&svalues);CHKERRQ(ierr); 5101 for (i=0; i<nsends; i++) { 5102 rowlen = svalues + sstarts[i]*sbs; 5103 nrows = sstarts[i+1]-sstarts[i]; /* num of block rows */ 5104 for (j=0; j<nrows; j++) { 5105 row = srow[k] + B->rmap->range[rank]; /* global row idx */ 5106 for (l=0; l<sbs; l++) { 5107 ierr = MatGetRow_MPIAIJ(B,row+l,&ncols,NULL,NULL);CHKERRQ(ierr); /* rowlength */ 5108 5109 rowlen[j*sbs+l] = ncols; 5110 5111 len += ncols; 5112 ierr = MatRestoreRow_MPIAIJ(B,row+l,&ncols,NULL,NULL);CHKERRQ(ierr); 5113 } 5114 k++; 5115 } 5116 ierr = MPI_Isend(rowlen,nrows*sbs,MPIU_INT,sprocs[i],tag,comm,swaits+i);CHKERRQ(ierr); 5117 5118 sstartsj[i+1] = len; /* starting point of (i+1)-th outgoing msg in bufj and bufa */ 5119 } 5120 /* recvs and sends of i-array are completed */ 5121 i = nrecvs; 5122 while (i--) { 5123 ierr = MPI_Waitany(nrecvs,rwaits,&jj,&rstatus);CHKERRQ(ierr); 5124 } 5125 if (nsends) {ierr = MPI_Waitall(nsends,swaits,sstatus);CHKERRQ(ierr);} 5126 ierr = PetscFree(svalues);CHKERRQ(ierr); 5127 5128 /* allocate buffers for sending j and a arrays */ 5129 ierr = PetscMalloc1(len+1,&bufj);CHKERRQ(ierr); 5130 ierr = PetscMalloc1(len+1,&bufa);CHKERRQ(ierr); 5131 5132 /* create i-array of B_oth */ 5133 ierr = PetscMalloc1(aBn+2,&b_othi);CHKERRQ(ierr); 5134 5135 b_othi[0] = 0; 5136 len = 0; /* total length of j or a array to be received */ 5137 k = 0; 5138 for (i=0; i<nrecvs; i++) { 5139 rowlen = rvalues + rstarts[i]*rbs; 5140 nrows = rbs*(rstarts[i+1]-rstarts[i]); /* num of rows to be received */ 5141 for (j=0; j<nrows; j++) { 5142 b_othi[k+1] = b_othi[k] + rowlen[j]; 5143 ierr = PetscIntSumError(rowlen[j],len,&len);CHKERRQ(ierr); 5144 k++; 5145 } 5146 rstartsj[i+1] = len; /* starting point of (i+1)-th incoming msg in bufj and bufa */ 5147 } 5148 ierr = PetscFree(rvalues);CHKERRQ(ierr); 5149 5150 /* allocate space for j and a arrrays of B_oth */ 5151 ierr = PetscMalloc1(b_othi[aBn]+1,&b_othj);CHKERRQ(ierr); 5152 ierr = PetscMalloc1(b_othi[aBn]+1,&b_otha);CHKERRQ(ierr); 5153 5154 /* j-array */ 5155 /*---------*/ 5156 /* post receives of j-array */ 5157 for (i=0; i<nrecvs; i++) { 5158 nrows = rstartsj[i+1]-rstartsj[i]; /* length of the msg received */ 5159 ierr = MPI_Irecv(b_othj+rstartsj[i],nrows,MPIU_INT,rprocs[i],tag,comm,rwaits+i);CHKERRQ(ierr); 5160 } 5161 5162 /* pack the outgoing message j-array */ 5163 k = 0; 5164 for (i=0; i<nsends; i++) { 5165 nrows = sstarts[i+1]-sstarts[i]; /* num of block rows */ 5166 bufJ = bufj+sstartsj[i]; 5167 for (j=0; j<nrows; j++) { 5168 row = srow[k++] + B->rmap->range[rank]; /* global row idx */ 5169 for (ll=0; ll<sbs; ll++) { 5170 ierr = MatGetRow_MPIAIJ(B,row+ll,&ncols,&cols,NULL);CHKERRQ(ierr); 5171 for (l=0; l<ncols; l++) { 5172 *bufJ++ = cols[l]; 5173 } 5174 ierr = MatRestoreRow_MPIAIJ(B,row+ll,&ncols,&cols,NULL);CHKERRQ(ierr); 5175 } 5176 } 5177 ierr = MPI_Isend(bufj+sstartsj[i],sstartsj[i+1]-sstartsj[i],MPIU_INT,sprocs[i],tag,comm,swaits+i);CHKERRQ(ierr); 5178 } 5179 5180 /* recvs and sends of j-array are completed */ 5181 i = nrecvs; 5182 while (i--) { 5183 ierr = MPI_Waitany(nrecvs,rwaits,&jj,&rstatus);CHKERRQ(ierr); 5184 } 5185 if (nsends) {ierr = MPI_Waitall(nsends,swaits,sstatus);CHKERRQ(ierr);} 5186 } else if (scall == MAT_REUSE_MATRIX) { 5187 sstartsj = *startsj_s; 5188 rstartsj = *startsj_r; 5189 bufa = *bufa_ptr; 5190 b_oth = (Mat_SeqAIJ*)(*B_oth)->data; 5191 b_otha = b_oth->a; 5192 } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE, "Matrix P does not posses an object container"); 5193 5194 /* a-array */ 5195 /*---------*/ 5196 /* post receives of a-array */ 5197 for (i=0; i<nrecvs; i++) { 5198 nrows = rstartsj[i+1]-rstartsj[i]; /* length of the msg received */ 5199 ierr = MPI_Irecv(b_otha+rstartsj[i],nrows,MPIU_SCALAR,rprocs[i],tag,comm,rwaits+i);CHKERRQ(ierr); 5200 } 5201 5202 /* pack the outgoing message a-array */ 5203 k = 0; 5204 for (i=0; i<nsends; i++) { 5205 nrows = sstarts[i+1]-sstarts[i]; /* num of block rows */ 5206 bufA = bufa+sstartsj[i]; 5207 for (j=0; j<nrows; j++) { 5208 row = srow[k++] + B->rmap->range[rank]; /* global row idx */ 5209 for (ll=0; ll<sbs; ll++) { 5210 ierr = MatGetRow_MPIAIJ(B,row+ll,&ncols,NULL,&vals);CHKERRQ(ierr); 5211 for (l=0; l<ncols; l++) { 5212 *bufA++ = vals[l]; 5213 } 5214 ierr = MatRestoreRow_MPIAIJ(B,row+ll,&ncols,NULL,&vals);CHKERRQ(ierr); 5215 } 5216 } 5217 ierr = MPI_Isend(bufa+sstartsj[i],sstartsj[i+1]-sstartsj[i],MPIU_SCALAR,sprocs[i],tag,comm,swaits+i);CHKERRQ(ierr); 5218 } 5219 /* recvs and sends of a-array are completed */ 5220 i = nrecvs; 5221 while (i--) { 5222 ierr = MPI_Waitany(nrecvs,rwaits,&jj,&rstatus);CHKERRQ(ierr); 5223 } 5224 if (nsends) {ierr = MPI_Waitall(nsends,swaits,sstatus);CHKERRQ(ierr);} 5225 ierr = PetscFree2(rwaits,swaits);CHKERRQ(ierr); 5226 5227 if (scall == MAT_INITIAL_MATRIX) { 5228 /* put together the new matrix */ 5229 ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,aBn,B->cmap->N,b_othi,b_othj,b_otha,B_oth);CHKERRQ(ierr); 5230 5231 /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */ 5232 /* Since these are PETSc arrays, change flags to free them as necessary. */ 5233 b_oth = (Mat_SeqAIJ*)(*B_oth)->data; 5234 b_oth->free_a = PETSC_TRUE; 5235 b_oth->free_ij = PETSC_TRUE; 5236 b_oth->nonew = 0; 5237 5238 ierr = PetscFree(bufj);CHKERRQ(ierr); 5239 if (!startsj_s || !bufa_ptr) { 5240 ierr = PetscFree2(sstartsj,rstartsj);CHKERRQ(ierr); 5241 ierr = PetscFree(bufa_ptr);CHKERRQ(ierr); 5242 } else { 5243 *startsj_s = sstartsj; 5244 *startsj_r = rstartsj; 5245 *bufa_ptr = bufa; 5246 } 5247 } 5248 ierr = PetscLogEventEnd(MAT_GetBrowsOfAocols,A,B,0,0);CHKERRQ(ierr); 5249 PetscFunctionReturn(0); 5250 } 5251 5252 /*@C 5253 MatGetCommunicationStructs - Provides access to the communication structures used in matrix-vector multiplication. 5254 5255 Not Collective 5256 5257 Input Parameters: 5258 . A - The matrix in mpiaij format 5259 5260 Output Parameter: 5261 + lvec - The local vector holding off-process values from the argument to a matrix-vector product 5262 . colmap - A map from global column index to local index into lvec 5263 - multScatter - A scatter from the argument of a matrix-vector product to lvec 5264 5265 Level: developer 5266 5267 @*/ 5268 #if defined(PETSC_USE_CTABLE) 5269 PetscErrorCode MatGetCommunicationStructs(Mat A, Vec *lvec, PetscTable *colmap, VecScatter *multScatter) 5270 #else 5271 PetscErrorCode MatGetCommunicationStructs(Mat A, Vec *lvec, PetscInt *colmap[], VecScatter *multScatter) 5272 #endif 5273 { 5274 Mat_MPIAIJ *a; 5275 5276 PetscFunctionBegin; 5277 PetscValidHeaderSpecific(A, MAT_CLASSID, 1); 5278 PetscValidPointer(lvec, 2); 5279 PetscValidPointer(colmap, 3); 5280 PetscValidPointer(multScatter, 4); 5281 a = (Mat_MPIAIJ*) A->data; 5282 if (lvec) *lvec = a->lvec; 5283 if (colmap) *colmap = a->colmap; 5284 if (multScatter) *multScatter = a->Mvctx; 5285 PetscFunctionReturn(0); 5286 } 5287 5288 PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJCRL(Mat,MatType,MatReuse,Mat*); 5289 PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJPERM(Mat,MatType,MatReuse,Mat*); 5290 PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPISBAIJ(Mat,MatType,MatReuse,Mat*); 5291 #if defined(PETSC_HAVE_ELEMENTAL) 5292 PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_Elemental(Mat,MatType,MatReuse,Mat*); 5293 #endif 5294 #if defined(PETSC_HAVE_HYPRE) 5295 PETSC_INTERN PetscErrorCode MatConvert_AIJ_HYPRE(Mat,MatType,MatReuse,Mat*); 5296 PETSC_INTERN PetscErrorCode MatMatMatMult_Transpose_AIJ_AIJ(Mat,Mat,Mat,MatReuse,PetscReal,Mat*); 5297 #endif 5298 PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_IS(Mat,MatType,MatReuse,Mat*); 5299 5300 /* 5301 Computes (B'*A')' since computing B*A directly is untenable 5302 5303 n p p 5304 ( ) ( ) ( ) 5305 m ( A ) * n ( B ) = m ( C ) 5306 ( ) ( ) ( ) 5307 5308 */ 5309 PetscErrorCode MatMatMultNumeric_MPIDense_MPIAIJ(Mat A,Mat B,Mat C) 5310 { 5311 PetscErrorCode ierr; 5312 Mat At,Bt,Ct; 5313 5314 PetscFunctionBegin; 5315 ierr = MatTranspose(A,MAT_INITIAL_MATRIX,&At);CHKERRQ(ierr); 5316 ierr = MatTranspose(B,MAT_INITIAL_MATRIX,&Bt);CHKERRQ(ierr); 5317 ierr = MatMatMult(Bt,At,MAT_INITIAL_MATRIX,1.0,&Ct);CHKERRQ(ierr); 5318 ierr = MatDestroy(&At);CHKERRQ(ierr); 5319 ierr = MatDestroy(&Bt);CHKERRQ(ierr); 5320 ierr = MatTranspose(Ct,MAT_REUSE_MATRIX,&C);CHKERRQ(ierr); 5321 ierr = MatDestroy(&Ct);CHKERRQ(ierr); 5322 PetscFunctionReturn(0); 5323 } 5324 5325 PetscErrorCode MatMatMultSymbolic_MPIDense_MPIAIJ(Mat A,Mat B,PetscReal fill,Mat *C) 5326 { 5327 PetscErrorCode ierr; 5328 PetscInt m=A->rmap->n,n=B->cmap->n; 5329 Mat Cmat; 5330 5331 PetscFunctionBegin; 5332 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); 5333 ierr = MatCreate(PetscObjectComm((PetscObject)A),&Cmat);CHKERRQ(ierr); 5334 ierr = MatSetSizes(Cmat,m,n,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 5335 ierr = MatSetBlockSizesFromMats(Cmat,A,B);CHKERRQ(ierr); 5336 ierr = MatSetType(Cmat,MATMPIDENSE);CHKERRQ(ierr); 5337 ierr = MatMPIDenseSetPreallocation(Cmat,NULL);CHKERRQ(ierr); 5338 ierr = MatAssemblyBegin(Cmat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5339 ierr = MatAssemblyEnd(Cmat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5340 5341 Cmat->ops->matmultnumeric = MatMatMultNumeric_MPIDense_MPIAIJ; 5342 5343 *C = Cmat; 5344 PetscFunctionReturn(0); 5345 } 5346 5347 /* ----------------------------------------------------------------*/ 5348 PETSC_INTERN PetscErrorCode MatMatMult_MPIDense_MPIAIJ(Mat A,Mat B,MatReuse scall,PetscReal fill,Mat *C) 5349 { 5350 PetscErrorCode ierr; 5351 5352 PetscFunctionBegin; 5353 if (scall == MAT_INITIAL_MATRIX) { 5354 ierr = PetscLogEventBegin(MAT_MatMultSymbolic,A,B,0,0);CHKERRQ(ierr); 5355 ierr = MatMatMultSymbolic_MPIDense_MPIAIJ(A,B,fill,C);CHKERRQ(ierr); 5356 ierr = PetscLogEventEnd(MAT_MatMultSymbolic,A,B,0,0);CHKERRQ(ierr); 5357 } 5358 ierr = PetscLogEventBegin(MAT_MatMultNumeric,A,B,0,0);CHKERRQ(ierr); 5359 ierr = MatMatMultNumeric_MPIDense_MPIAIJ(A,B,*C);CHKERRQ(ierr); 5360 ierr = PetscLogEventEnd(MAT_MatMultNumeric,A,B,0,0);CHKERRQ(ierr); 5361 PetscFunctionReturn(0); 5362 } 5363 5364 /*MC 5365 MATMPIAIJ - MATMPIAIJ = "mpiaij" - A matrix type to be used for parallel sparse matrices. 5366 5367 Options Database Keys: 5368 . -mat_type mpiaij - sets the matrix type to "mpiaij" during a call to MatSetFromOptions() 5369 5370 Level: beginner 5371 5372 .seealso: MatCreateAIJ() 5373 M*/ 5374 5375 PETSC_EXTERN PetscErrorCode MatCreate_MPIAIJ(Mat B) 5376 { 5377 Mat_MPIAIJ *b; 5378 PetscErrorCode ierr; 5379 PetscMPIInt size; 5380 5381 PetscFunctionBegin; 5382 ierr = MPI_Comm_size(PetscObjectComm((PetscObject)B),&size);CHKERRQ(ierr); 5383 5384 ierr = PetscNewLog(B,&b);CHKERRQ(ierr); 5385 B->data = (void*)b; 5386 ierr = PetscMemcpy(B->ops,&MatOps_Values,sizeof(struct _MatOps));CHKERRQ(ierr); 5387 B->assembled = PETSC_FALSE; 5388 B->insertmode = NOT_SET_VALUES; 5389 b->size = size; 5390 5391 ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)B),&b->rank);CHKERRQ(ierr); 5392 5393 /* build cache for off array entries formed */ 5394 ierr = MatStashCreate_Private(PetscObjectComm((PetscObject)B),1,&B->stash);CHKERRQ(ierr); 5395 5396 b->donotstash = PETSC_FALSE; 5397 b->colmap = 0; 5398 b->garray = 0; 5399 b->roworiented = PETSC_TRUE; 5400 5401 /* stuff used for matrix vector multiply */ 5402 b->lvec = NULL; 5403 b->Mvctx = NULL; 5404 5405 /* stuff for MatGetRow() */ 5406 b->rowindices = 0; 5407 b->rowvalues = 0; 5408 b->getrowactive = PETSC_FALSE; 5409 5410 /* flexible pointer used in CUSP/CUSPARSE classes */ 5411 b->spptr = NULL; 5412 5413 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMPIAIJSetUseScalableIncreaseOverlap_C",MatMPIAIJSetUseScalableIncreaseOverlap_MPIAIJ);CHKERRQ(ierr); 5414 ierr = PetscObjectComposeFunction((PetscObject)B,"MatStoreValues_C",MatStoreValues_MPIAIJ);CHKERRQ(ierr); 5415 ierr = PetscObjectComposeFunction((PetscObject)B,"MatRetrieveValues_C",MatRetrieveValues_MPIAIJ);CHKERRQ(ierr); 5416 ierr = PetscObjectComposeFunction((PetscObject)B,"MatIsTranspose_C",MatIsTranspose_MPIAIJ);CHKERRQ(ierr); 5417 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMPIAIJSetPreallocation_C",MatMPIAIJSetPreallocation_MPIAIJ);CHKERRQ(ierr); 5418 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMPIAIJSetPreallocationCSR_C",MatMPIAIJSetPreallocationCSR_MPIAIJ);CHKERRQ(ierr); 5419 ierr = PetscObjectComposeFunction((PetscObject)B,"MatDiagonalScaleLocal_C",MatDiagonalScaleLocal_MPIAIJ);CHKERRQ(ierr); 5420 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_mpiaijperm_C",MatConvert_MPIAIJ_MPIAIJPERM);CHKERRQ(ierr); 5421 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_mpiaijcrl_C",MatConvert_MPIAIJ_MPIAIJCRL);CHKERRQ(ierr); 5422 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_mpisbaij_C",MatConvert_MPIAIJ_MPISBAIJ);CHKERRQ(ierr); 5423 #if defined(PETSC_HAVE_ELEMENTAL) 5424 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_elemental_C",MatConvert_MPIAIJ_Elemental);CHKERRQ(ierr); 5425 #endif 5426 #if defined(PETSC_HAVE_HYPRE) 5427 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_hypre_C",MatConvert_AIJ_HYPRE);CHKERRQ(ierr); 5428 #endif 5429 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_is_C",MatConvert_MPIAIJ_IS);CHKERRQ(ierr); 5430 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMatMult_mpidense_mpiaij_C",MatMatMult_MPIDense_MPIAIJ);CHKERRQ(ierr); 5431 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMatMultSymbolic_mpidense_mpiaij_C",MatMatMultSymbolic_MPIDense_MPIAIJ);CHKERRQ(ierr); 5432 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMatMultNumeric_mpidense_mpiaij_C",MatMatMultNumeric_MPIDense_MPIAIJ);CHKERRQ(ierr); 5433 #if defined(PETSC_HAVE_HYPRE) 5434 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMatMatMult_transpose_mpiaij_mpiaij_C",MatMatMatMult_Transpose_AIJ_AIJ);CHKERRQ(ierr); 5435 #endif 5436 ierr = PetscObjectChangeTypeName((PetscObject)B,MATMPIAIJ);CHKERRQ(ierr); 5437 PetscFunctionReturn(0); 5438 } 5439 5440 /*@C 5441 MatCreateMPIAIJWithSplitArrays - creates a MPI AIJ matrix using arrays that contain the "diagonal" 5442 and "off-diagonal" part of the matrix in CSR format. 5443 5444 Collective on MPI_Comm 5445 5446 Input Parameters: 5447 + comm - MPI communicator 5448 . m - number of local rows (Cannot be PETSC_DECIDE) 5449 . n - This value should be the same as the local size used in creating the 5450 x vector for the matrix-vector product y = Ax. (or PETSC_DECIDE to have 5451 calculated if N is given) For square matrices n is almost always m. 5452 . M - number of global rows (or PETSC_DETERMINE to have calculated if m is given) 5453 . N - number of global columns (or PETSC_DETERMINE to have calculated if n is given) 5454 . i - row indices for "diagonal" portion of matrix 5455 . j - column indices 5456 . a - matrix values 5457 . oi - row indices for "off-diagonal" portion of matrix 5458 . oj - column indices 5459 - oa - matrix values 5460 5461 Output Parameter: 5462 . mat - the matrix 5463 5464 Level: advanced 5465 5466 Notes: 5467 The i, j, and a arrays ARE NOT copied by this routine into the internal format used by PETSc. The user 5468 must free the arrays once the matrix has been destroyed and not before. 5469 5470 The i and j indices are 0 based 5471 5472 See MatCreateAIJ() for the definition of "diagonal" and "off-diagonal" portion of the matrix 5473 5474 This sets local rows and cannot be used to set off-processor values. 5475 5476 Use of this routine is discouraged because it is inflexible and cumbersome to use. It is extremely rare that a 5477 legacy application natively assembles into exactly this split format. The code to do so is nontrivial and does 5478 not easily support in-place reassembly. It is recommended to use MatSetValues() (or a variant thereof) because 5479 the resulting assembly is easier to implement, will work with any matrix format, and the user does not have to 5480 keep track of the underlying array. Use MatSetOption(A,MAT_NO_OFF_PROC_ENTRIES,PETSC_TRUE) to disable all 5481 communication if it is known that only local entries will be set. 5482 5483 .keywords: matrix, aij, compressed row, sparse, parallel 5484 5485 .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatMPIAIJSetPreallocationCSR(), 5486 MATMPIAIJ, MatCreateAIJ(), MatCreateMPIAIJWithArrays() 5487 @*/ 5488 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) 5489 { 5490 PetscErrorCode ierr; 5491 Mat_MPIAIJ *maij; 5492 5493 PetscFunctionBegin; 5494 if (m < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"local number of rows (m) cannot be PETSC_DECIDE, or negative"); 5495 if (i[0]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"i (row indices) must start with 0"); 5496 if (oi[0]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"oi (row indices) must start with 0"); 5497 ierr = MatCreate(comm,mat);CHKERRQ(ierr); 5498 ierr = MatSetSizes(*mat,m,n,M,N);CHKERRQ(ierr); 5499 ierr = MatSetType(*mat,MATMPIAIJ);CHKERRQ(ierr); 5500 maij = (Mat_MPIAIJ*) (*mat)->data; 5501 5502 (*mat)->preallocated = PETSC_TRUE; 5503 5504 ierr = PetscLayoutSetUp((*mat)->rmap);CHKERRQ(ierr); 5505 ierr = PetscLayoutSetUp((*mat)->cmap);CHKERRQ(ierr); 5506 5507 ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,m,n,i,j,a,&maij->A);CHKERRQ(ierr); 5508 ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,m,(*mat)->cmap->N,oi,oj,oa,&maij->B);CHKERRQ(ierr); 5509 5510 ierr = MatAssemblyBegin(maij->A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5511 ierr = MatAssemblyEnd(maij->A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5512 ierr = MatAssemblyBegin(maij->B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5513 ierr = MatAssemblyEnd(maij->B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5514 5515 ierr = MatSetOption(*mat,MAT_NO_OFF_PROC_ENTRIES,PETSC_TRUE);CHKERRQ(ierr); 5516 ierr = MatAssemblyBegin(*mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5517 ierr = MatAssemblyEnd(*mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5518 ierr = MatSetOption(*mat,MAT_NO_OFF_PROC_ENTRIES,PETSC_FALSE);CHKERRQ(ierr); 5519 ierr = MatSetOption(*mat,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE);CHKERRQ(ierr); 5520 PetscFunctionReturn(0); 5521 } 5522 5523 /* 5524 Special version for direct calls from Fortran 5525 */ 5526 #include <petsc/private/fortranimpl.h> 5527 5528 /* Change these macros so can be used in void function */ 5529 #undef CHKERRQ 5530 #define CHKERRQ(ierr) CHKERRABORT(PETSC_COMM_WORLD,ierr) 5531 #undef SETERRQ2 5532 #define SETERRQ2(comm,ierr,b,c,d) CHKERRABORT(comm,ierr) 5533 #undef SETERRQ3 5534 #define SETERRQ3(comm,ierr,b,c,d,e) CHKERRABORT(comm,ierr) 5535 #undef SETERRQ 5536 #define SETERRQ(c,ierr,b) CHKERRABORT(c,ierr) 5537 5538 #if defined(PETSC_HAVE_FORTRAN_CAPS) 5539 #define matsetvaluesmpiaij_ MATSETVALUESMPIAIJ 5540 #elif !defined(PETSC_HAVE_FORTRAN_UNDERSCORE) 5541 #define matsetvaluesmpiaij_ matsetvaluesmpiaij 5542 #else 5543 #endif 5544 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) 5545 { 5546 Mat mat = *mmat; 5547 PetscInt m = *mm, n = *mn; 5548 InsertMode addv = *maddv; 5549 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 5550 PetscScalar value; 5551 PetscErrorCode ierr; 5552 5553 MatCheckPreallocated(mat,1); 5554 if (mat->insertmode == NOT_SET_VALUES) mat->insertmode = addv; 5555 5556 #if defined(PETSC_USE_DEBUG) 5557 else if (mat->insertmode != addv) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Cannot mix add values and insert values"); 5558 #endif 5559 { 5560 PetscInt i,j,rstart = mat->rmap->rstart,rend = mat->rmap->rend; 5561 PetscInt cstart = mat->cmap->rstart,cend = mat->cmap->rend,row,col; 5562 PetscBool roworiented = aij->roworiented; 5563 5564 /* Some Variables required in the macro */ 5565 Mat A = aij->A; 5566 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 5567 PetscInt *aimax = a->imax,*ai = a->i,*ailen = a->ilen,*aj = a->j; 5568 MatScalar *aa = a->a; 5569 PetscBool ignorezeroentries = (((a->ignorezeroentries)&&(addv==ADD_VALUES)) ? PETSC_TRUE : PETSC_FALSE); 5570 Mat B = aij->B; 5571 Mat_SeqAIJ *b = (Mat_SeqAIJ*)B->data; 5572 PetscInt *bimax = b->imax,*bi = b->i,*bilen = b->ilen,*bj = b->j,bm = aij->B->rmap->n,am = aij->A->rmap->n; 5573 MatScalar *ba = b->a; 5574 5575 PetscInt *rp1,*rp2,ii,nrow1,nrow2,_i,rmax1,rmax2,N,low1,high1,low2,high2,t,lastcol1,lastcol2; 5576 PetscInt nonew = a->nonew; 5577 MatScalar *ap1,*ap2; 5578 5579 PetscFunctionBegin; 5580 for (i=0; i<m; i++) { 5581 if (im[i] < 0) continue; 5582 #if defined(PETSC_USE_DEBUG) 5583 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); 5584 #endif 5585 if (im[i] >= rstart && im[i] < rend) { 5586 row = im[i] - rstart; 5587 lastcol1 = -1; 5588 rp1 = aj + ai[row]; 5589 ap1 = aa + ai[row]; 5590 rmax1 = aimax[row]; 5591 nrow1 = ailen[row]; 5592 low1 = 0; 5593 high1 = nrow1; 5594 lastcol2 = -1; 5595 rp2 = bj + bi[row]; 5596 ap2 = ba + bi[row]; 5597 rmax2 = bimax[row]; 5598 nrow2 = bilen[row]; 5599 low2 = 0; 5600 high2 = nrow2; 5601 5602 for (j=0; j<n; j++) { 5603 if (roworiented) value = v[i*n+j]; 5604 else value = v[i+j*m]; 5605 if (in[j] >= cstart && in[j] < cend) { 5606 col = in[j] - cstart; 5607 if (ignorezeroentries && value == 0.0 && (addv == ADD_VALUES) && row != col) continue; 5608 MatSetValues_SeqAIJ_A_Private(row,col,value,addv,im[i],in[j]); 5609 } else if (in[j] < 0) continue; 5610 #if defined(PETSC_USE_DEBUG) 5611 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); 5612 #endif 5613 else { 5614 if (mat->was_assembled) { 5615 if (!aij->colmap) { 5616 ierr = MatCreateColmap_MPIAIJ_Private(mat);CHKERRQ(ierr); 5617 } 5618 #if defined(PETSC_USE_CTABLE) 5619 ierr = PetscTableFind(aij->colmap,in[j]+1,&col);CHKERRQ(ierr); 5620 col--; 5621 #else 5622 col = aij->colmap[in[j]] - 1; 5623 #endif 5624 if (ignorezeroentries && value == 0.0 && (addv == ADD_VALUES) && row != col) continue; 5625 if (col < 0 && !((Mat_SeqAIJ*)(aij->A->data))->nonew) { 5626 ierr = MatDisAssemble_MPIAIJ(mat);CHKERRQ(ierr); 5627 col = in[j]; 5628 /* Reinitialize the variables required by MatSetValues_SeqAIJ_B_Private() */ 5629 B = aij->B; 5630 b = (Mat_SeqAIJ*)B->data; 5631 bimax = b->imax; bi = b->i; bilen = b->ilen; bj = b->j; 5632 rp2 = bj + bi[row]; 5633 ap2 = ba + bi[row]; 5634 rmax2 = bimax[row]; 5635 nrow2 = bilen[row]; 5636 low2 = 0; 5637 high2 = nrow2; 5638 bm = aij->B->rmap->n; 5639 ba = b->a; 5640 } 5641 } else col = in[j]; 5642 MatSetValues_SeqAIJ_B_Private(row,col,value,addv,im[i],in[j]); 5643 } 5644 } 5645 } else if (!aij->donotstash) { 5646 if (roworiented) { 5647 ierr = MatStashValuesRow_Private(&mat->stash,im[i],n,in,v+i*n,(PetscBool)(ignorezeroentries && (addv == ADD_VALUES)));CHKERRQ(ierr); 5648 } else { 5649 ierr = MatStashValuesCol_Private(&mat->stash,im[i],n,in,v+i,m,(PetscBool)(ignorezeroentries && (addv == ADD_VALUES)));CHKERRQ(ierr); 5650 } 5651 } 5652 } 5653 } 5654 PetscFunctionReturnVoid(); 5655 } 5656 5657