1 2 3 #include <../src/mat/impls/aij/mpi/mpiaij.h> /*I "petscmat.h" I*/ 4 #include <petsc/private/vecimpl.h> 5 #include <petsc/private/isimpl.h> 6 #include <petscblaslapack.h> 7 #include <petscsf.h> 8 9 /*MC 10 MATAIJ - MATAIJ = "aij" - A matrix type to be used for sparse matrices. 11 12 This matrix type is identical to MATSEQAIJ when constructed with a single process communicator, 13 and MATMPIAIJ otherwise. As a result, for single process communicators, 14 MatSeqAIJSetPreallocation is supported, and similarly MatMPIAIJSetPreallocation is supported 15 for communicators controlling multiple processes. It is recommended that you call both of 16 the above preallocation routines for simplicity. 17 18 Options Database Keys: 19 . -mat_type aij - sets the matrix type to "aij" during a call to MatSetFromOptions() 20 21 Developer Notes: Subclasses include MATAIJCUSP, MATAIJCUSPARSE, MATAIJPERM, MATAIJMKL, MATAIJCRL, and also automatically switches over to use inodes when 22 enough exist. 23 24 Level: beginner 25 26 .seealso: MatCreateAIJ(), MatCreateSeqAIJ(), MATSEQAIJ, MATMPIAIJ 27 M*/ 28 29 /*MC 30 MATAIJCRL - MATAIJCRL = "aijcrl" - A matrix type to be used for sparse matrices. 31 32 This matrix type is identical to MATSEQAIJCRL when constructed with a single process communicator, 33 and MATMPIAIJCRL otherwise. As a result, for single process communicators, 34 MatSeqAIJSetPreallocation() is supported, and similarly MatMPIAIJSetPreallocation() is supported 35 for communicators controlling multiple processes. It is recommended that you call both of 36 the above preallocation routines for simplicity. 37 38 Options Database Keys: 39 . -mat_type aijcrl - sets the matrix type to "aijcrl" during a call to MatSetFromOptions() 40 41 Level: beginner 42 43 .seealso: MatCreateMPIAIJCRL,MATSEQAIJCRL,MATMPIAIJCRL, MATSEQAIJCRL, MATMPIAIJCRL 44 M*/ 45 46 PetscErrorCode MatSetBlockSizes_MPIAIJ(Mat M, PetscInt rbs, PetscInt cbs) 47 { 48 PetscErrorCode ierr; 49 Mat_MPIAIJ *mat = (Mat_MPIAIJ*)M->data; 50 51 PetscFunctionBegin; 52 if (mat->A) { 53 ierr = MatSetBlockSizes(mat->A,rbs,cbs);CHKERRQ(ierr); 54 ierr = MatSetBlockSizes(mat->B,rbs,1);CHKERRQ(ierr); 55 } 56 PetscFunctionReturn(0); 57 } 58 59 PetscErrorCode MatFindNonzeroRows_MPIAIJ(Mat M,IS *keptrows) 60 { 61 PetscErrorCode ierr; 62 Mat_MPIAIJ *mat = (Mat_MPIAIJ*)M->data; 63 Mat_SeqAIJ *a = (Mat_SeqAIJ*)mat->A->data; 64 Mat_SeqAIJ *b = (Mat_SeqAIJ*)mat->B->data; 65 const PetscInt *ia,*ib; 66 const MatScalar *aa,*bb; 67 PetscInt na,nb,i,j,*rows,cnt=0,n0rows; 68 PetscInt m = M->rmap->n,rstart = M->rmap->rstart; 69 70 PetscFunctionBegin; 71 *keptrows = 0; 72 ia = a->i; 73 ib = b->i; 74 for (i=0; i<m; i++) { 75 na = ia[i+1] - ia[i]; 76 nb = ib[i+1] - ib[i]; 77 if (!na && !nb) { 78 cnt++; 79 goto ok1; 80 } 81 aa = a->a + ia[i]; 82 for (j=0; j<na; j++) { 83 if (aa[j] != 0.0) goto ok1; 84 } 85 bb = b->a + ib[i]; 86 for (j=0; j <nb; j++) { 87 if (bb[j] != 0.0) goto ok1; 88 } 89 cnt++; 90 ok1:; 91 } 92 ierr = MPIU_Allreduce(&cnt,&n0rows,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)M));CHKERRQ(ierr); 93 if (!n0rows) PetscFunctionReturn(0); 94 ierr = PetscMalloc1(M->rmap->n-cnt,&rows);CHKERRQ(ierr); 95 cnt = 0; 96 for (i=0; i<m; i++) { 97 na = ia[i+1] - ia[i]; 98 nb = ib[i+1] - ib[i]; 99 if (!na && !nb) continue; 100 aa = a->a + ia[i]; 101 for (j=0; j<na;j++) { 102 if (aa[j] != 0.0) { 103 rows[cnt++] = rstart + i; 104 goto ok2; 105 } 106 } 107 bb = b->a + ib[i]; 108 for (j=0; j<nb; j++) { 109 if (bb[j] != 0.0) { 110 rows[cnt++] = rstart + i; 111 goto ok2; 112 } 113 } 114 ok2:; 115 } 116 ierr = ISCreateGeneral(PetscObjectComm((PetscObject)M),cnt,rows,PETSC_OWN_POINTER,keptrows);CHKERRQ(ierr); 117 PetscFunctionReturn(0); 118 } 119 120 PetscErrorCode MatDiagonalSet_MPIAIJ(Mat Y,Vec D,InsertMode is) 121 { 122 PetscErrorCode ierr; 123 Mat_MPIAIJ *aij = (Mat_MPIAIJ*) Y->data; 124 125 PetscFunctionBegin; 126 if (Y->assembled && Y->rmap->rstart == Y->cmap->rstart && Y->rmap->rend == Y->cmap->rend) { 127 ierr = MatDiagonalSet(aij->A,D,is);CHKERRQ(ierr); 128 } else { 129 ierr = MatDiagonalSet_Default(Y,D,is);CHKERRQ(ierr); 130 } 131 PetscFunctionReturn(0); 132 } 133 134 PetscErrorCode MatFindZeroDiagonals_MPIAIJ(Mat M,IS *zrows) 135 { 136 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)M->data; 137 PetscErrorCode ierr; 138 PetscInt i,rstart,nrows,*rows; 139 140 PetscFunctionBegin; 141 *zrows = NULL; 142 ierr = MatFindZeroDiagonals_SeqAIJ_Private(aij->A,&nrows,&rows);CHKERRQ(ierr); 143 ierr = MatGetOwnershipRange(M,&rstart,NULL);CHKERRQ(ierr); 144 for (i=0; i<nrows; i++) rows[i] += rstart; 145 ierr = ISCreateGeneral(PetscObjectComm((PetscObject)M),nrows,rows,PETSC_OWN_POINTER,zrows);CHKERRQ(ierr); 146 PetscFunctionReturn(0); 147 } 148 149 PetscErrorCode MatGetColumnNorms_MPIAIJ(Mat A,NormType type,PetscReal *norms) 150 { 151 PetscErrorCode ierr; 152 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)A->data; 153 PetscInt i,n,*garray = aij->garray; 154 Mat_SeqAIJ *a_aij = (Mat_SeqAIJ*) aij->A->data; 155 Mat_SeqAIJ *b_aij = (Mat_SeqAIJ*) aij->B->data; 156 PetscReal *work; 157 158 PetscFunctionBegin; 159 ierr = MatGetSize(A,NULL,&n);CHKERRQ(ierr); 160 ierr = PetscCalloc1(n,&work);CHKERRQ(ierr); 161 if (type == NORM_2) { 162 for (i=0; i<a_aij->i[aij->A->rmap->n]; i++) { 163 work[A->cmap->rstart + a_aij->j[i]] += PetscAbsScalar(a_aij->a[i]*a_aij->a[i]); 164 } 165 for (i=0; i<b_aij->i[aij->B->rmap->n]; i++) { 166 work[garray[b_aij->j[i]]] += PetscAbsScalar(b_aij->a[i]*b_aij->a[i]); 167 } 168 } else if (type == NORM_1) { 169 for (i=0; i<a_aij->i[aij->A->rmap->n]; i++) { 170 work[A->cmap->rstart + a_aij->j[i]] += PetscAbsScalar(a_aij->a[i]); 171 } 172 for (i=0; i<b_aij->i[aij->B->rmap->n]; i++) { 173 work[garray[b_aij->j[i]]] += PetscAbsScalar(b_aij->a[i]); 174 } 175 } else if (type == NORM_INFINITY) { 176 for (i=0; i<a_aij->i[aij->A->rmap->n]; i++) { 177 work[A->cmap->rstart + a_aij->j[i]] = PetscMax(PetscAbsScalar(a_aij->a[i]), work[A->cmap->rstart + a_aij->j[i]]); 178 } 179 for (i=0; i<b_aij->i[aij->B->rmap->n]; i++) { 180 work[garray[b_aij->j[i]]] = PetscMax(PetscAbsScalar(b_aij->a[i]),work[garray[b_aij->j[i]]]); 181 } 182 183 } else SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_WRONG,"Unknown NormType"); 184 if (type == NORM_INFINITY) { 185 ierr = MPIU_Allreduce(work,norms,n,MPIU_REAL,MPIU_MAX,PetscObjectComm((PetscObject)A));CHKERRQ(ierr); 186 } else { 187 ierr = MPIU_Allreduce(work,norms,n,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)A));CHKERRQ(ierr); 188 } 189 ierr = PetscFree(work);CHKERRQ(ierr); 190 if (type == NORM_2) { 191 for (i=0; i<n; i++) norms[i] = PetscSqrtReal(norms[i]); 192 } 193 PetscFunctionReturn(0); 194 } 195 196 PetscErrorCode MatFindOffBlockDiagonalEntries_MPIAIJ(Mat A,IS *is) 197 { 198 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 199 IS sis,gis; 200 PetscErrorCode ierr; 201 const PetscInt *isis,*igis; 202 PetscInt n,*iis,nsis,ngis,rstart,i; 203 204 PetscFunctionBegin; 205 ierr = MatFindOffBlockDiagonalEntries(a->A,&sis);CHKERRQ(ierr); 206 ierr = MatFindNonzeroRows(a->B,&gis);CHKERRQ(ierr); 207 ierr = ISGetSize(gis,&ngis);CHKERRQ(ierr); 208 ierr = ISGetSize(sis,&nsis);CHKERRQ(ierr); 209 ierr = ISGetIndices(sis,&isis);CHKERRQ(ierr); 210 ierr = ISGetIndices(gis,&igis);CHKERRQ(ierr); 211 212 ierr = PetscMalloc1(ngis+nsis,&iis);CHKERRQ(ierr); 213 ierr = PetscMemcpy(iis,igis,ngis*sizeof(PetscInt));CHKERRQ(ierr); 214 ierr = PetscMemcpy(iis+ngis,isis,nsis*sizeof(PetscInt));CHKERRQ(ierr); 215 n = ngis + nsis; 216 ierr = PetscSortRemoveDupsInt(&n,iis);CHKERRQ(ierr); 217 ierr = MatGetOwnershipRange(A,&rstart,NULL);CHKERRQ(ierr); 218 for (i=0; i<n; i++) iis[i] += rstart; 219 ierr = ISCreateGeneral(PetscObjectComm((PetscObject)A),n,iis,PETSC_OWN_POINTER,is);CHKERRQ(ierr); 220 221 ierr = ISRestoreIndices(sis,&isis);CHKERRQ(ierr); 222 ierr = ISRestoreIndices(gis,&igis);CHKERRQ(ierr); 223 ierr = ISDestroy(&sis);CHKERRQ(ierr); 224 ierr = ISDestroy(&gis);CHKERRQ(ierr); 225 PetscFunctionReturn(0); 226 } 227 228 /* 229 Distributes a SeqAIJ matrix across a set of processes. Code stolen from 230 MatLoad_MPIAIJ(). Horrible lack of reuse. Should be a routine for each matrix type. 231 232 Only for square matrices 233 234 Used by a preconditioner, hence PETSC_EXTERN 235 */ 236 PETSC_EXTERN PetscErrorCode MatDistribute_MPIAIJ(MPI_Comm comm,Mat gmat,PetscInt m,MatReuse reuse,Mat *inmat) 237 { 238 PetscMPIInt rank,size; 239 PetscInt *rowners,*dlens,*olens,i,rstart,rend,j,jj,nz = 0,*gmataj,cnt,row,*ld,bses[2]; 240 PetscErrorCode ierr; 241 Mat mat; 242 Mat_SeqAIJ *gmata; 243 PetscMPIInt tag; 244 MPI_Status status; 245 PetscBool aij; 246 MatScalar *gmataa,*ao,*ad,*gmataarestore=0; 247 248 PetscFunctionBegin; 249 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 250 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 251 if (!rank) { 252 ierr = PetscObjectTypeCompare((PetscObject)gmat,MATSEQAIJ,&aij);CHKERRQ(ierr); 253 if (!aij) SETERRQ1(PetscObjectComm((PetscObject)gmat),PETSC_ERR_SUP,"Currently no support for input matrix of type %s\n",((PetscObject)gmat)->type_name); 254 } 255 if (reuse == MAT_INITIAL_MATRIX) { 256 ierr = MatCreate(comm,&mat);CHKERRQ(ierr); 257 ierr = MatSetSizes(mat,m,m,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 258 ierr = MatGetBlockSizes(gmat,&bses[0],&bses[1]);CHKERRQ(ierr); 259 ierr = MPI_Bcast(bses,2,MPIU_INT,0,comm);CHKERRQ(ierr); 260 ierr = MatSetBlockSizes(mat,bses[0],bses[1]);CHKERRQ(ierr); 261 ierr = MatSetType(mat,MATAIJ);CHKERRQ(ierr); 262 ierr = PetscMalloc1(size+1,&rowners);CHKERRQ(ierr); 263 ierr = PetscMalloc2(m,&dlens,m,&olens);CHKERRQ(ierr); 264 ierr = MPI_Allgather(&m,1,MPIU_INT,rowners+1,1,MPIU_INT,comm);CHKERRQ(ierr); 265 266 rowners[0] = 0; 267 for (i=2; i<=size; i++) rowners[i] += rowners[i-1]; 268 rstart = rowners[rank]; 269 rend = rowners[rank+1]; 270 ierr = PetscObjectGetNewTag((PetscObject)mat,&tag);CHKERRQ(ierr); 271 if (!rank) { 272 gmata = (Mat_SeqAIJ*) gmat->data; 273 /* send row lengths to all processors */ 274 for (i=0; i<m; i++) dlens[i] = gmata->ilen[i]; 275 for (i=1; i<size; i++) { 276 ierr = MPI_Send(gmata->ilen + rowners[i],rowners[i+1]-rowners[i],MPIU_INT,i,tag,comm);CHKERRQ(ierr); 277 } 278 /* determine number diagonal and off-diagonal counts */ 279 ierr = PetscMemzero(olens,m*sizeof(PetscInt));CHKERRQ(ierr); 280 ierr = PetscCalloc1(m,&ld);CHKERRQ(ierr); 281 jj = 0; 282 for (i=0; i<m; i++) { 283 for (j=0; j<dlens[i]; j++) { 284 if (gmata->j[jj] < rstart) ld[i]++; 285 if (gmata->j[jj] < rstart || gmata->j[jj] >= rend) olens[i]++; 286 jj++; 287 } 288 } 289 /* send column indices to other processes */ 290 for (i=1; i<size; i++) { 291 nz = gmata->i[rowners[i+1]]-gmata->i[rowners[i]]; 292 ierr = MPI_Send(&nz,1,MPIU_INT,i,tag,comm);CHKERRQ(ierr); 293 ierr = MPI_Send(gmata->j + gmata->i[rowners[i]],nz,MPIU_INT,i,tag,comm);CHKERRQ(ierr); 294 } 295 296 /* send numerical values to other processes */ 297 for (i=1; i<size; i++) { 298 nz = gmata->i[rowners[i+1]]-gmata->i[rowners[i]]; 299 ierr = MPI_Send(gmata->a + gmata->i[rowners[i]],nz,MPIU_SCALAR,i,tag,comm);CHKERRQ(ierr); 300 } 301 gmataa = gmata->a; 302 gmataj = gmata->j; 303 304 } else { 305 /* receive row lengths */ 306 ierr = MPI_Recv(dlens,m,MPIU_INT,0,tag,comm,&status);CHKERRQ(ierr); 307 /* receive column indices */ 308 ierr = MPI_Recv(&nz,1,MPIU_INT,0,tag,comm,&status);CHKERRQ(ierr); 309 ierr = PetscMalloc2(nz,&gmataa,nz,&gmataj);CHKERRQ(ierr); 310 ierr = MPI_Recv(gmataj,nz,MPIU_INT,0,tag,comm,&status);CHKERRQ(ierr); 311 /* determine number diagonal and off-diagonal counts */ 312 ierr = PetscMemzero(olens,m*sizeof(PetscInt));CHKERRQ(ierr); 313 ierr = PetscCalloc1(m,&ld);CHKERRQ(ierr); 314 jj = 0; 315 for (i=0; i<m; i++) { 316 for (j=0; j<dlens[i]; j++) { 317 if (gmataj[jj] < rstart) ld[i]++; 318 if (gmataj[jj] < rstart || gmataj[jj] >= rend) olens[i]++; 319 jj++; 320 } 321 } 322 /* receive numerical values */ 323 ierr = PetscMemzero(gmataa,nz*sizeof(PetscScalar));CHKERRQ(ierr); 324 ierr = MPI_Recv(gmataa,nz,MPIU_SCALAR,0,tag,comm,&status);CHKERRQ(ierr); 325 } 326 /* set preallocation */ 327 for (i=0; i<m; i++) { 328 dlens[i] -= olens[i]; 329 } 330 ierr = MatSeqAIJSetPreallocation(mat,0,dlens);CHKERRQ(ierr); 331 ierr = MatMPIAIJSetPreallocation(mat,0,dlens,0,olens);CHKERRQ(ierr); 332 333 for (i=0; i<m; i++) { 334 dlens[i] += olens[i]; 335 } 336 cnt = 0; 337 for (i=0; i<m; i++) { 338 row = rstart + i; 339 ierr = MatSetValues(mat,1,&row,dlens[i],gmataj+cnt,gmataa+cnt,INSERT_VALUES);CHKERRQ(ierr); 340 cnt += dlens[i]; 341 } 342 if (rank) { 343 ierr = PetscFree2(gmataa,gmataj);CHKERRQ(ierr); 344 } 345 ierr = PetscFree2(dlens,olens);CHKERRQ(ierr); 346 ierr = PetscFree(rowners);CHKERRQ(ierr); 347 348 ((Mat_MPIAIJ*)(mat->data))->ld = ld; 349 350 *inmat = mat; 351 } else { /* column indices are already set; only need to move over numerical values from process 0 */ 352 Mat_SeqAIJ *Ad = (Mat_SeqAIJ*)((Mat_MPIAIJ*)((*inmat)->data))->A->data; 353 Mat_SeqAIJ *Ao = (Mat_SeqAIJ*)((Mat_MPIAIJ*)((*inmat)->data))->B->data; 354 mat = *inmat; 355 ierr = PetscObjectGetNewTag((PetscObject)mat,&tag);CHKERRQ(ierr); 356 if (!rank) { 357 /* send numerical values to other processes */ 358 gmata = (Mat_SeqAIJ*) gmat->data; 359 ierr = MatGetOwnershipRanges(mat,(const PetscInt**)&rowners);CHKERRQ(ierr); 360 gmataa = gmata->a; 361 for (i=1; i<size; i++) { 362 nz = gmata->i[rowners[i+1]]-gmata->i[rowners[i]]; 363 ierr = MPI_Send(gmataa + gmata->i[rowners[i]],nz,MPIU_SCALAR,i,tag,comm);CHKERRQ(ierr); 364 } 365 nz = gmata->i[rowners[1]]-gmata->i[rowners[0]]; 366 } else { 367 /* receive numerical values from process 0*/ 368 nz = Ad->nz + Ao->nz; 369 ierr = PetscMalloc1(nz,&gmataa);CHKERRQ(ierr); gmataarestore = gmataa; 370 ierr = MPI_Recv(gmataa,nz,MPIU_SCALAR,0,tag,comm,&status);CHKERRQ(ierr); 371 } 372 /* transfer numerical values into the diagonal A and off diagonal B parts of mat */ 373 ld = ((Mat_MPIAIJ*)(mat->data))->ld; 374 ad = Ad->a; 375 ao = Ao->a; 376 if (mat->rmap->n) { 377 i = 0; 378 nz = ld[i]; ierr = PetscMemcpy(ao,gmataa,nz*sizeof(PetscScalar));CHKERRQ(ierr); ao += nz; gmataa += nz; 379 nz = Ad->i[i+1] - Ad->i[i]; ierr = PetscMemcpy(ad,gmataa,nz*sizeof(PetscScalar));CHKERRQ(ierr); ad += nz; gmataa += nz; 380 } 381 for (i=1; i<mat->rmap->n; i++) { 382 nz = Ao->i[i] - Ao->i[i-1] - ld[i-1] + ld[i]; ierr = PetscMemcpy(ao,gmataa,nz*sizeof(PetscScalar));CHKERRQ(ierr); ao += nz; gmataa += nz; 383 nz = Ad->i[i+1] - Ad->i[i]; ierr = PetscMemcpy(ad,gmataa,nz*sizeof(PetscScalar));CHKERRQ(ierr); ad += nz; gmataa += nz; 384 } 385 i--; 386 if (mat->rmap->n) { 387 nz = Ao->i[i+1] - Ao->i[i] - ld[i]; ierr = PetscMemcpy(ao,gmataa,nz*sizeof(PetscScalar));CHKERRQ(ierr); 388 } 389 if (rank) { 390 ierr = PetscFree(gmataarestore);CHKERRQ(ierr); 391 } 392 } 393 ierr = MatAssemblyBegin(mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 394 ierr = MatAssemblyEnd(mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 395 PetscFunctionReturn(0); 396 } 397 398 /* 399 Local utility routine that creates a mapping from the global column 400 number to the local number in the off-diagonal part of the local 401 storage of the matrix. When PETSC_USE_CTABLE is used this is scalable at 402 a slightly higher hash table cost; without it it is not scalable (each processor 403 has an order N integer array but is fast to acess. 404 */ 405 PetscErrorCode MatCreateColmap_MPIAIJ_Private(Mat mat) 406 { 407 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 408 PetscErrorCode ierr; 409 PetscInt n = aij->B->cmap->n,i; 410 411 PetscFunctionBegin; 412 if (!aij->garray) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"MPIAIJ Matrix was assembled but is missing garray"); 413 #if defined(PETSC_USE_CTABLE) 414 ierr = PetscTableCreate(n,mat->cmap->N+1,&aij->colmap);CHKERRQ(ierr); 415 for (i=0; i<n; i++) { 416 ierr = PetscTableAdd(aij->colmap,aij->garray[i]+1,i+1,INSERT_VALUES);CHKERRQ(ierr); 417 } 418 #else 419 ierr = PetscCalloc1(mat->cmap->N+1,&aij->colmap);CHKERRQ(ierr); 420 ierr = PetscLogObjectMemory((PetscObject)mat,(mat->cmap->N+1)*sizeof(PetscInt));CHKERRQ(ierr); 421 for (i=0; i<n; i++) aij->colmap[aij->garray[i]] = i+1; 422 #endif 423 PetscFunctionReturn(0); 424 } 425 426 #define MatSetValues_SeqAIJ_A_Private(row,col,value,addv,orow,ocol) \ 427 { \ 428 if (col <= lastcol1) low1 = 0; \ 429 else high1 = nrow1; \ 430 lastcol1 = col;\ 431 while (high1-low1 > 5) { \ 432 t = (low1+high1)/2; \ 433 if (rp1[t] > col) high1 = t; \ 434 else low1 = t; \ 435 } \ 436 for (_i=low1; _i<high1; _i++) { \ 437 if (rp1[_i] > col) break; \ 438 if (rp1[_i] == col) { \ 439 if (addv == ADD_VALUES) ap1[_i] += value; \ 440 else ap1[_i] = value; \ 441 goto a_noinsert; \ 442 } \ 443 } \ 444 if (value == 0.0 && ignorezeroentries && row != col) {low1 = 0; high1 = nrow1;goto a_noinsert;} \ 445 if (nonew == 1) {low1 = 0; high1 = nrow1; goto a_noinsert;} \ 446 if (nonew == -1) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero at global row/column (%D, %D) into matrix", orow, ocol); \ 447 MatSeqXAIJReallocateAIJ(A,am,1,nrow1,row,col,rmax1,aa,ai,aj,rp1,ap1,aimax,nonew,MatScalar); \ 448 N = nrow1++ - 1; a->nz++; high1++; \ 449 /* shift up all the later entries in this row */ \ 450 for (ii=N; ii>=_i; ii--) { \ 451 rp1[ii+1] = rp1[ii]; \ 452 ap1[ii+1] = ap1[ii]; \ 453 } \ 454 rp1[_i] = col; \ 455 ap1[_i] = value; \ 456 A->nonzerostate++;\ 457 a_noinsert: ; \ 458 ailen[row] = nrow1; \ 459 } 460 461 #define MatSetValues_SeqAIJ_B_Private(row,col,value,addv,orow,ocol) \ 462 { \ 463 if (col <= lastcol2) low2 = 0; \ 464 else high2 = nrow2; \ 465 lastcol2 = col; \ 466 while (high2-low2 > 5) { \ 467 t = (low2+high2)/2; \ 468 if (rp2[t] > col) high2 = t; \ 469 else low2 = t; \ 470 } \ 471 for (_i=low2; _i<high2; _i++) { \ 472 if (rp2[_i] > col) break; \ 473 if (rp2[_i] == col) { \ 474 if (addv == ADD_VALUES) ap2[_i] += value; \ 475 else ap2[_i] = value; \ 476 goto b_noinsert; \ 477 } \ 478 } \ 479 if (value == 0.0 && ignorezeroentries) {low2 = 0; high2 = nrow2; goto b_noinsert;} \ 480 if (nonew == 1) {low2 = 0; high2 = nrow2; goto b_noinsert;} \ 481 if (nonew == -1) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero at global row/column (%D, %D) into matrix", orow, ocol); \ 482 MatSeqXAIJReallocateAIJ(B,bm,1,nrow2,row,col,rmax2,ba,bi,bj,rp2,ap2,bimax,nonew,MatScalar); \ 483 N = nrow2++ - 1; b->nz++; high2++; \ 484 /* shift up all the later entries in this row */ \ 485 for (ii=N; ii>=_i; ii--) { \ 486 rp2[ii+1] = rp2[ii]; \ 487 ap2[ii+1] = ap2[ii]; \ 488 } \ 489 rp2[_i] = col; \ 490 ap2[_i] = value; \ 491 B->nonzerostate++; \ 492 b_noinsert: ; \ 493 bilen[row] = nrow2; \ 494 } 495 496 PetscErrorCode MatSetValuesRow_MPIAIJ(Mat A,PetscInt row,const PetscScalar v[]) 497 { 498 Mat_MPIAIJ *mat = (Mat_MPIAIJ*)A->data; 499 Mat_SeqAIJ *a = (Mat_SeqAIJ*)mat->A->data,*b = (Mat_SeqAIJ*)mat->B->data; 500 PetscErrorCode ierr; 501 PetscInt l,*garray = mat->garray,diag; 502 503 PetscFunctionBegin; 504 /* code only works for square matrices A */ 505 506 /* find size of row to the left of the diagonal part */ 507 ierr = MatGetOwnershipRange(A,&diag,0);CHKERRQ(ierr); 508 row = row - diag; 509 for (l=0; l<b->i[row+1]-b->i[row]; l++) { 510 if (garray[b->j[b->i[row]+l]] > diag) break; 511 } 512 ierr = PetscMemcpy(b->a+b->i[row],v,l*sizeof(PetscScalar));CHKERRQ(ierr); 513 514 /* diagonal part */ 515 ierr = PetscMemcpy(a->a+a->i[row],v+l,(a->i[row+1]-a->i[row])*sizeof(PetscScalar));CHKERRQ(ierr); 516 517 /* right of diagonal part */ 518 ierr = PetscMemcpy(b->a+b->i[row]+l,v+l+a->i[row+1]-a->i[row],(b->i[row+1]-b->i[row]-l)*sizeof(PetscScalar));CHKERRQ(ierr); 519 PetscFunctionReturn(0); 520 } 521 522 PetscErrorCode MatSetValues_MPIAIJ(Mat mat,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],const PetscScalar v[],InsertMode addv) 523 { 524 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 525 PetscScalar value; 526 PetscErrorCode ierr; 527 PetscInt i,j,rstart = mat->rmap->rstart,rend = mat->rmap->rend; 528 PetscInt cstart = mat->cmap->rstart,cend = mat->cmap->rend,row,col; 529 PetscBool roworiented = aij->roworiented; 530 531 /* Some Variables required in the macro */ 532 Mat A = aij->A; 533 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 534 PetscInt *aimax = a->imax,*ai = a->i,*ailen = a->ilen,*aj = a->j; 535 MatScalar *aa = a->a; 536 PetscBool ignorezeroentries = a->ignorezeroentries; 537 Mat B = aij->B; 538 Mat_SeqAIJ *b = (Mat_SeqAIJ*)B->data; 539 PetscInt *bimax = b->imax,*bi = b->i,*bilen = b->ilen,*bj = b->j,bm = aij->B->rmap->n,am = aij->A->rmap->n; 540 MatScalar *ba = b->a; 541 542 PetscInt *rp1,*rp2,ii,nrow1,nrow2,_i,rmax1,rmax2,N,low1,high1,low2,high2,t,lastcol1,lastcol2; 543 PetscInt nonew; 544 MatScalar *ap1,*ap2; 545 546 PetscFunctionBegin; 547 for (i=0; i<m; i++) { 548 if (im[i] < 0) continue; 549 #if defined(PETSC_USE_DEBUG) 550 if (im[i] >= mat->rmap->N) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Row too large: row %D max %D",im[i],mat->rmap->N-1); 551 #endif 552 if (im[i] >= rstart && im[i] < rend) { 553 row = im[i] - rstart; 554 lastcol1 = -1; 555 rp1 = aj + ai[row]; 556 ap1 = aa + ai[row]; 557 rmax1 = aimax[row]; 558 nrow1 = ailen[row]; 559 low1 = 0; 560 high1 = nrow1; 561 lastcol2 = -1; 562 rp2 = bj + bi[row]; 563 ap2 = ba + bi[row]; 564 rmax2 = bimax[row]; 565 nrow2 = bilen[row]; 566 low2 = 0; 567 high2 = nrow2; 568 569 for (j=0; j<n; j++) { 570 if (roworiented) value = v[i*n+j]; 571 else value = v[i+j*m]; 572 if (in[j] >= cstart && in[j] < cend) { 573 col = in[j] - cstart; 574 nonew = a->nonew; 575 if (ignorezeroentries && value == 0.0 && (addv == ADD_VALUES) && row != col) continue; 576 MatSetValues_SeqAIJ_A_Private(row,col,value,addv,im[i],in[j]); 577 } else if (in[j] < 0) continue; 578 #if defined(PETSC_USE_DEBUG) 579 else if (in[j] >= mat->cmap->N) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Column too large: col %D max %D",in[j],mat->cmap->N-1); 580 #endif 581 else { 582 if (mat->was_assembled) { 583 if (!aij->colmap) { 584 ierr = MatCreateColmap_MPIAIJ_Private(mat);CHKERRQ(ierr); 585 } 586 #if defined(PETSC_USE_CTABLE) 587 ierr = PetscTableFind(aij->colmap,in[j]+1,&col);CHKERRQ(ierr); 588 col--; 589 #else 590 col = aij->colmap[in[j]] - 1; 591 #endif 592 if (col < 0 && !((Mat_SeqAIJ*)(aij->B->data))->nonew) { 593 ierr = MatDisAssemble_MPIAIJ(mat);CHKERRQ(ierr); 594 col = in[j]; 595 /* Reinitialize the variables required by MatSetValues_SeqAIJ_B_Private() */ 596 B = aij->B; 597 b = (Mat_SeqAIJ*)B->data; 598 bimax = b->imax; bi = b->i; bilen = b->ilen; bj = b->j; ba = b->a; 599 rp2 = bj + bi[row]; 600 ap2 = ba + bi[row]; 601 rmax2 = bimax[row]; 602 nrow2 = bilen[row]; 603 low2 = 0; 604 high2 = nrow2; 605 bm = aij->B->rmap->n; 606 ba = b->a; 607 } else if (col < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero at global row/column (%D, %D) into matrix", im[i], in[j]); 608 } else col = in[j]; 609 nonew = b->nonew; 610 MatSetValues_SeqAIJ_B_Private(row,col,value,addv,im[i],in[j]); 611 } 612 } 613 } else { 614 if (mat->nooffprocentries) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Setting off process row %D even though MatSetOption(,MAT_NO_OFF_PROC_ENTRIES,PETSC_TRUE) was set",im[i]); 615 if (!aij->donotstash) { 616 mat->assembled = PETSC_FALSE; 617 if (roworiented) { 618 ierr = MatStashValuesRow_Private(&mat->stash,im[i],n,in,v+i*n,(PetscBool)(ignorezeroentries && (addv == ADD_VALUES)));CHKERRQ(ierr); 619 } else { 620 ierr = MatStashValuesCol_Private(&mat->stash,im[i],n,in,v+i,m,(PetscBool)(ignorezeroentries && (addv == ADD_VALUES)));CHKERRQ(ierr); 621 } 622 } 623 } 624 } 625 PetscFunctionReturn(0); 626 } 627 628 PetscErrorCode MatGetValues_MPIAIJ(Mat mat,PetscInt m,const PetscInt idxm[],PetscInt n,const PetscInt idxn[],PetscScalar v[]) 629 { 630 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 631 PetscErrorCode ierr; 632 PetscInt i,j,rstart = mat->rmap->rstart,rend = mat->rmap->rend; 633 PetscInt cstart = mat->cmap->rstart,cend = mat->cmap->rend,row,col; 634 635 PetscFunctionBegin; 636 for (i=0; i<m; i++) { 637 if (idxm[i] < 0) continue; /* SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative row: %D",idxm[i]);*/ 638 if (idxm[i] >= mat->rmap->N) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Row too large: row %D max %D",idxm[i],mat->rmap->N-1); 639 if (idxm[i] >= rstart && idxm[i] < rend) { 640 row = idxm[i] - rstart; 641 for (j=0; j<n; j++) { 642 if (idxn[j] < 0) continue; /* SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative column: %D",idxn[j]); */ 643 if (idxn[j] >= mat->cmap->N) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Column too large: col %D max %D",idxn[j],mat->cmap->N-1); 644 if (idxn[j] >= cstart && idxn[j] < cend) { 645 col = idxn[j] - cstart; 646 ierr = MatGetValues(aij->A,1,&row,1,&col,v+i*n+j);CHKERRQ(ierr); 647 } else { 648 if (!aij->colmap) { 649 ierr = MatCreateColmap_MPIAIJ_Private(mat);CHKERRQ(ierr); 650 } 651 #if defined(PETSC_USE_CTABLE) 652 ierr = PetscTableFind(aij->colmap,idxn[j]+1,&col);CHKERRQ(ierr); 653 col--; 654 #else 655 col = aij->colmap[idxn[j]] - 1; 656 #endif 657 if ((col < 0) || (aij->garray[col] != idxn[j])) *(v+i*n+j) = 0.0; 658 else { 659 ierr = MatGetValues(aij->B,1,&row,1,&col,v+i*n+j);CHKERRQ(ierr); 660 } 661 } 662 } 663 } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Only local values currently supported"); 664 } 665 PetscFunctionReturn(0); 666 } 667 668 extern PetscErrorCode MatMultDiagonalBlock_MPIAIJ(Mat,Vec,Vec); 669 670 PetscErrorCode MatAssemblyBegin_MPIAIJ(Mat mat,MatAssemblyType mode) 671 { 672 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 673 PetscErrorCode ierr; 674 PetscInt nstash,reallocs; 675 676 PetscFunctionBegin; 677 if (aij->donotstash || mat->nooffprocentries) PetscFunctionReturn(0); 678 679 ierr = MatStashScatterBegin_Private(mat,&mat->stash,mat->rmap->range);CHKERRQ(ierr); 680 ierr = MatStashGetInfo_Private(&mat->stash,&nstash,&reallocs);CHKERRQ(ierr); 681 ierr = PetscInfo2(aij->A,"Stash has %D entries, uses %D mallocs.\n",nstash,reallocs);CHKERRQ(ierr); 682 PetscFunctionReturn(0); 683 } 684 685 PetscErrorCode MatAssemblyEnd_MPIAIJ(Mat mat,MatAssemblyType mode) 686 { 687 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 688 Mat_SeqAIJ *a = (Mat_SeqAIJ*)aij->A->data; 689 PetscErrorCode ierr; 690 PetscMPIInt n; 691 PetscInt i,j,rstart,ncols,flg; 692 PetscInt *row,*col; 693 PetscBool other_disassembled; 694 PetscScalar *val; 695 696 /* do not use 'b = (Mat_SeqAIJ*)aij->B->data' as B can be reset in disassembly */ 697 698 PetscFunctionBegin; 699 if (!aij->donotstash && !mat->nooffprocentries) { 700 while (1) { 701 ierr = MatStashScatterGetMesg_Private(&mat->stash,&n,&row,&col,&val,&flg);CHKERRQ(ierr); 702 if (!flg) break; 703 704 for (i=0; i<n; ) { 705 /* Now identify the consecutive vals belonging to the same row */ 706 for (j=i,rstart=row[j]; j<n; j++) { 707 if (row[j] != rstart) break; 708 } 709 if (j < n) ncols = j-i; 710 else ncols = n-i; 711 /* Now assemble all these values with a single function call */ 712 ierr = MatSetValues_MPIAIJ(mat,1,row+i,ncols,col+i,val+i,mat->insertmode);CHKERRQ(ierr); 713 714 i = j; 715 } 716 } 717 ierr = MatStashScatterEnd_Private(&mat->stash);CHKERRQ(ierr); 718 } 719 ierr = MatAssemblyBegin(aij->A,mode);CHKERRQ(ierr); 720 ierr = MatAssemblyEnd(aij->A,mode);CHKERRQ(ierr); 721 722 /* determine if any processor has disassembled, if so we must 723 also disassemble ourselfs, in order that we may reassemble. */ 724 /* 725 if nonzero structure of submatrix B cannot change then we know that 726 no processor disassembled thus we can skip this stuff 727 */ 728 if (!((Mat_SeqAIJ*)aij->B->data)->nonew) { 729 ierr = MPIU_Allreduce(&mat->was_assembled,&other_disassembled,1,MPIU_BOOL,MPI_PROD,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 730 if (mat->was_assembled && !other_disassembled) { 731 ierr = MatDisAssemble_MPIAIJ(mat);CHKERRQ(ierr); 732 } 733 } 734 if (!mat->was_assembled && mode == MAT_FINAL_ASSEMBLY) { 735 ierr = MatSetUpMultiply_MPIAIJ(mat);CHKERRQ(ierr); 736 } 737 ierr = MatSetOption(aij->B,MAT_USE_INODES,PETSC_FALSE);CHKERRQ(ierr); 738 ierr = MatAssemblyBegin(aij->B,mode);CHKERRQ(ierr); 739 ierr = MatAssemblyEnd(aij->B,mode);CHKERRQ(ierr); 740 741 ierr = PetscFree2(aij->rowvalues,aij->rowindices);CHKERRQ(ierr); 742 743 aij->rowvalues = 0; 744 745 ierr = VecDestroy(&aij->diag);CHKERRQ(ierr); 746 if (a->inode.size) mat->ops->multdiagonalblock = MatMultDiagonalBlock_MPIAIJ; 747 748 /* if no new nonzero locations are allowed in matrix then only set the matrix state the first time through */ 749 if ((!mat->was_assembled && mode == MAT_FINAL_ASSEMBLY) || !((Mat_SeqAIJ*)(aij->A->data))->nonew) { 750 PetscObjectState state = aij->A->nonzerostate + aij->B->nonzerostate; 751 ierr = MPIU_Allreduce(&state,&mat->nonzerostate,1,MPIU_INT64,MPI_SUM,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 752 } 753 PetscFunctionReturn(0); 754 } 755 756 PetscErrorCode MatZeroEntries_MPIAIJ(Mat A) 757 { 758 Mat_MPIAIJ *l = (Mat_MPIAIJ*)A->data; 759 PetscErrorCode ierr; 760 761 PetscFunctionBegin; 762 ierr = MatZeroEntries(l->A);CHKERRQ(ierr); 763 ierr = MatZeroEntries(l->B);CHKERRQ(ierr); 764 PetscFunctionReturn(0); 765 } 766 767 PetscErrorCode MatZeroRows_MPIAIJ(Mat A,PetscInt N,const PetscInt rows[],PetscScalar diag,Vec x,Vec b) 768 { 769 Mat_MPIAIJ *mat = (Mat_MPIAIJ *) A->data; 770 PetscInt *lrows; 771 PetscInt r, len; 772 PetscErrorCode ierr; 773 774 PetscFunctionBegin; 775 /* get locally owned rows */ 776 ierr = MatZeroRowsMapLocal_Private(A,N,rows,&len,&lrows);CHKERRQ(ierr); 777 /* fix right hand side if needed */ 778 if (x && b) { 779 const PetscScalar *xx; 780 PetscScalar *bb; 781 782 ierr = VecGetArrayRead(x, &xx);CHKERRQ(ierr); 783 ierr = VecGetArray(b, &bb);CHKERRQ(ierr); 784 for (r = 0; r < len; ++r) bb[lrows[r]] = diag*xx[lrows[r]]; 785 ierr = VecRestoreArrayRead(x, &xx);CHKERRQ(ierr); 786 ierr = VecRestoreArray(b, &bb);CHKERRQ(ierr); 787 } 788 /* Must zero l->B before l->A because the (diag) case below may put values into l->B*/ 789 ierr = MatZeroRows(mat->B, len, lrows, 0.0, NULL, NULL);CHKERRQ(ierr); 790 if (A->congruentlayouts == -1) { /* first time we compare rows and cols layouts */ 791 PetscBool cong; 792 ierr = PetscLayoutCompare(A->rmap,A->cmap,&cong);CHKERRQ(ierr); 793 if (cong) A->congruentlayouts = 1; 794 else A->congruentlayouts = 0; 795 } 796 if ((diag != 0.0) && A->congruentlayouts) { 797 ierr = MatZeroRows(mat->A, len, lrows, diag, NULL, NULL);CHKERRQ(ierr); 798 } else if (diag != 0.0) { 799 ierr = MatZeroRows(mat->A, len, lrows, 0.0, NULL, NULL);CHKERRQ(ierr); 800 if (((Mat_SeqAIJ *) mat->A->data)->nonew) SETERRQ(PETSC_COMM_SELF, PETSC_ERR_SUP, "MatZeroRows() on rectangular matrices cannot be used with the Mat options\nMAT_NEW_NONZERO_LOCATIONS,MAT_NEW_NONZERO_LOCATION_ERR,MAT_NEW_NONZERO_ALLOCATION_ERR"); 801 for (r = 0; r < len; ++r) { 802 const PetscInt row = lrows[r] + A->rmap->rstart; 803 ierr = MatSetValues(A, 1, &row, 1, &row, &diag, INSERT_VALUES);CHKERRQ(ierr); 804 } 805 ierr = MatAssemblyBegin(A, MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 806 ierr = MatAssemblyEnd(A, MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 807 } else { 808 ierr = MatZeroRows(mat->A, len, lrows, 0.0, NULL, NULL);CHKERRQ(ierr); 809 } 810 ierr = PetscFree(lrows);CHKERRQ(ierr); 811 812 /* only change matrix nonzero state if pattern was allowed to be changed */ 813 if (!((Mat_SeqAIJ*)(mat->A->data))->keepnonzeropattern) { 814 PetscObjectState state = mat->A->nonzerostate + mat->B->nonzerostate; 815 ierr = MPIU_Allreduce(&state,&A->nonzerostate,1,MPIU_INT64,MPI_SUM,PetscObjectComm((PetscObject)A));CHKERRQ(ierr); 816 } 817 PetscFunctionReturn(0); 818 } 819 820 PetscErrorCode MatZeroRowsColumns_MPIAIJ(Mat A,PetscInt N,const PetscInt rows[],PetscScalar diag,Vec x,Vec b) 821 { 822 Mat_MPIAIJ *l = (Mat_MPIAIJ*)A->data; 823 PetscErrorCode ierr; 824 PetscMPIInt n = A->rmap->n; 825 PetscInt i,j,r,m,p = 0,len = 0; 826 PetscInt *lrows,*owners = A->rmap->range; 827 PetscSFNode *rrows; 828 PetscSF sf; 829 const PetscScalar *xx; 830 PetscScalar *bb,*mask; 831 Vec xmask,lmask; 832 Mat_SeqAIJ *aij = (Mat_SeqAIJ*)l->B->data; 833 const PetscInt *aj, *ii,*ridx; 834 PetscScalar *aa; 835 836 PetscFunctionBegin; 837 /* Create SF where leaves are input rows and roots are owned rows */ 838 ierr = PetscMalloc1(n, &lrows);CHKERRQ(ierr); 839 for (r = 0; r < n; ++r) lrows[r] = -1; 840 ierr = PetscMalloc1(N, &rrows);CHKERRQ(ierr); 841 for (r = 0; r < N; ++r) { 842 const PetscInt idx = rows[r]; 843 if (idx < 0 || A->rmap->N <= idx) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Row %D out of range [0,%D)",idx,A->rmap->N); 844 if (idx < owners[p] || owners[p+1] <= idx) { /* short-circuit the search if the last p owns this row too */ 845 ierr = PetscLayoutFindOwner(A->rmap,idx,&p);CHKERRQ(ierr); 846 } 847 rrows[r].rank = p; 848 rrows[r].index = rows[r] - owners[p]; 849 } 850 ierr = PetscSFCreate(PetscObjectComm((PetscObject) A), &sf);CHKERRQ(ierr); 851 ierr = PetscSFSetGraph(sf, n, N, NULL, PETSC_OWN_POINTER, rrows, PETSC_OWN_POINTER);CHKERRQ(ierr); 852 /* Collect flags for rows to be zeroed */ 853 ierr = PetscSFReduceBegin(sf, MPIU_INT, (PetscInt *) rows, lrows, MPI_LOR);CHKERRQ(ierr); 854 ierr = PetscSFReduceEnd(sf, MPIU_INT, (PetscInt *) rows, lrows, MPI_LOR);CHKERRQ(ierr); 855 ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 856 /* Compress and put in row numbers */ 857 for (r = 0; r < n; ++r) if (lrows[r] >= 0) lrows[len++] = r; 858 /* zero diagonal part of matrix */ 859 ierr = MatZeroRowsColumns(l->A,len,lrows,diag,x,b);CHKERRQ(ierr); 860 /* handle off diagonal part of matrix */ 861 ierr = MatCreateVecs(A,&xmask,NULL);CHKERRQ(ierr); 862 ierr = VecDuplicate(l->lvec,&lmask);CHKERRQ(ierr); 863 ierr = VecGetArray(xmask,&bb);CHKERRQ(ierr); 864 for (i=0; i<len; i++) bb[lrows[i]] = 1; 865 ierr = VecRestoreArray(xmask,&bb);CHKERRQ(ierr); 866 ierr = VecScatterBegin(l->Mvctx,xmask,lmask,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 867 ierr = VecScatterEnd(l->Mvctx,xmask,lmask,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 868 ierr = VecDestroy(&xmask);CHKERRQ(ierr); 869 if (x) { 870 ierr = VecScatterBegin(l->Mvctx,x,l->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 871 ierr = VecScatterEnd(l->Mvctx,x,l->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 872 ierr = VecGetArrayRead(l->lvec,&xx);CHKERRQ(ierr); 873 ierr = VecGetArray(b,&bb);CHKERRQ(ierr); 874 } 875 ierr = VecGetArray(lmask,&mask);CHKERRQ(ierr); 876 /* remove zeroed rows of off diagonal matrix */ 877 ii = aij->i; 878 for (i=0; i<len; i++) { 879 ierr = PetscMemzero(aij->a + ii[lrows[i]],(ii[lrows[i]+1] - ii[lrows[i]])*sizeof(PetscScalar));CHKERRQ(ierr); 880 } 881 /* loop over all elements of off process part of matrix zeroing removed columns*/ 882 if (aij->compressedrow.use) { 883 m = aij->compressedrow.nrows; 884 ii = aij->compressedrow.i; 885 ridx = aij->compressedrow.rindex; 886 for (i=0; i<m; i++) { 887 n = ii[i+1] - ii[i]; 888 aj = aij->j + ii[i]; 889 aa = aij->a + ii[i]; 890 891 for (j=0; j<n; j++) { 892 if (PetscAbsScalar(mask[*aj])) { 893 if (b) bb[*ridx] -= *aa*xx[*aj]; 894 *aa = 0.0; 895 } 896 aa++; 897 aj++; 898 } 899 ridx++; 900 } 901 } else { /* do not use compressed row format */ 902 m = l->B->rmap->n; 903 for (i=0; i<m; i++) { 904 n = ii[i+1] - ii[i]; 905 aj = aij->j + ii[i]; 906 aa = aij->a + ii[i]; 907 for (j=0; j<n; j++) { 908 if (PetscAbsScalar(mask[*aj])) { 909 if (b) bb[i] -= *aa*xx[*aj]; 910 *aa = 0.0; 911 } 912 aa++; 913 aj++; 914 } 915 } 916 } 917 if (x) { 918 ierr = VecRestoreArray(b,&bb);CHKERRQ(ierr); 919 ierr = VecRestoreArrayRead(l->lvec,&xx);CHKERRQ(ierr); 920 } 921 ierr = VecRestoreArray(lmask,&mask);CHKERRQ(ierr); 922 ierr = VecDestroy(&lmask);CHKERRQ(ierr); 923 ierr = PetscFree(lrows);CHKERRQ(ierr); 924 925 /* only change matrix nonzero state if pattern was allowed to be changed */ 926 if (!((Mat_SeqAIJ*)(l->A->data))->keepnonzeropattern) { 927 PetscObjectState state = l->A->nonzerostate + l->B->nonzerostate; 928 ierr = MPIU_Allreduce(&state,&A->nonzerostate,1,MPIU_INT64,MPI_SUM,PetscObjectComm((PetscObject)A));CHKERRQ(ierr); 929 } 930 PetscFunctionReturn(0); 931 } 932 933 PetscErrorCode MatMult_MPIAIJ(Mat A,Vec xx,Vec yy) 934 { 935 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 936 PetscErrorCode ierr; 937 PetscInt nt; 938 939 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 MatIsSymmetric_MPIAIJ(Mat A,PetscReal tol,PetscBool *f) 1042 { 1043 PetscErrorCode ierr; 1044 1045 PetscFunctionBegin; 1046 ierr = MatIsTranspose_MPIAIJ(A,A,tol,f);CHKERRQ(ierr); 1047 PetscFunctionReturn(0); 1048 } 1049 1050 PetscErrorCode MatMultTransposeAdd_MPIAIJ(Mat A,Vec xx,Vec yy,Vec zz) 1051 { 1052 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 1053 PetscErrorCode ierr; 1054 1055 PetscFunctionBegin; 1056 /* do nondiagonal part */ 1057 ierr = (*a->B->ops->multtranspose)(a->B,xx,a->lvec);CHKERRQ(ierr); 1058 /* send it on its way */ 1059 ierr = VecScatterBegin(a->Mvctx,a->lvec,zz,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1060 /* do local part */ 1061 ierr = (*a->A->ops->multtransposeadd)(a->A,xx,yy,zz);CHKERRQ(ierr); 1062 /* receive remote parts */ 1063 ierr = VecScatterEnd(a->Mvctx,a->lvec,zz,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1064 PetscFunctionReturn(0); 1065 } 1066 1067 /* 1068 This only works correctly for square matrices where the subblock A->A is the 1069 diagonal block 1070 */ 1071 PetscErrorCode MatGetDiagonal_MPIAIJ(Mat A,Vec v) 1072 { 1073 PetscErrorCode ierr; 1074 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 1075 1076 PetscFunctionBegin; 1077 if (A->rmap->N != A->cmap->N) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"Supports only square matrix where A->A is diag block"); 1078 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"); 1079 ierr = MatGetDiagonal(a->A,v);CHKERRQ(ierr); 1080 PetscFunctionReturn(0); 1081 } 1082 1083 PetscErrorCode MatScale_MPIAIJ(Mat A,PetscScalar aa) 1084 { 1085 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 1086 PetscErrorCode ierr; 1087 1088 PetscFunctionBegin; 1089 ierr = MatScale(a->A,aa);CHKERRQ(ierr); 1090 ierr = MatScale(a->B,aa);CHKERRQ(ierr); 1091 PetscFunctionReturn(0); 1092 } 1093 1094 PetscErrorCode MatDestroy_MPIAIJ(Mat mat) 1095 { 1096 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 1097 PetscErrorCode ierr; 1098 1099 PetscFunctionBegin; 1100 #if defined(PETSC_USE_LOG) 1101 PetscLogObjectState((PetscObject)mat,"Rows=%D, Cols=%D",mat->rmap->N,mat->cmap->N); 1102 #endif 1103 ierr = MatStashDestroy_Private(&mat->stash);CHKERRQ(ierr); 1104 ierr = VecDestroy(&aij->diag);CHKERRQ(ierr); 1105 ierr = MatDestroy(&aij->A);CHKERRQ(ierr); 1106 ierr = MatDestroy(&aij->B);CHKERRQ(ierr); 1107 #if defined(PETSC_USE_CTABLE) 1108 ierr = PetscTableDestroy(&aij->colmap);CHKERRQ(ierr); 1109 #else 1110 ierr = PetscFree(aij->colmap);CHKERRQ(ierr); 1111 #endif 1112 ierr = PetscFree(aij->garray);CHKERRQ(ierr); 1113 ierr = VecDestroy(&aij->lvec);CHKERRQ(ierr); 1114 ierr = VecScatterDestroy(&aij->Mvctx);CHKERRQ(ierr); 1115 if (aij->Mvctx_mpi1) {ierr = VecScatterDestroy(&aij->Mvctx_mpi1);CHKERRQ(ierr);} 1116 ierr = PetscFree2(aij->rowvalues,aij->rowindices);CHKERRQ(ierr); 1117 ierr = PetscFree(aij->ld);CHKERRQ(ierr); 1118 ierr = PetscFree(mat->data);CHKERRQ(ierr); 1119 1120 ierr = PetscObjectChangeTypeName((PetscObject)mat,0);CHKERRQ(ierr); 1121 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatStoreValues_C",NULL);CHKERRQ(ierr); 1122 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatRetrieveValues_C",NULL);CHKERRQ(ierr); 1123 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatIsTranspose_C",NULL);CHKERRQ(ierr); 1124 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatMPIAIJSetPreallocation_C",NULL);CHKERRQ(ierr); 1125 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatResetPreallocation_C",NULL);CHKERRQ(ierr); 1126 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatMPIAIJSetPreallocationCSR_C",NULL);CHKERRQ(ierr); 1127 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatDiagonalScaleLocal_C",NULL);CHKERRQ(ierr); 1128 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatConvert_mpiaij_mpisbaij_C",NULL);CHKERRQ(ierr); 1129 #if defined(PETSC_HAVE_ELEMENTAL) 1130 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatConvert_mpiaij_elemental_C",NULL);CHKERRQ(ierr); 1131 #endif 1132 #if defined(PETSC_HAVE_HYPRE) 1133 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatConvert_mpiaij_hypre_C",NULL);CHKERRQ(ierr); 1134 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatMatMatMult_transpose_mpiaij_mpiaij_C",NULL);CHKERRQ(ierr); 1135 #endif 1136 PetscFunctionReturn(0); 1137 } 1138 1139 PetscErrorCode MatView_MPIAIJ_Binary(Mat mat,PetscViewer viewer) 1140 { 1141 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 1142 Mat_SeqAIJ *A = (Mat_SeqAIJ*)aij->A->data; 1143 Mat_SeqAIJ *B = (Mat_SeqAIJ*)aij->B->data; 1144 PetscErrorCode ierr; 1145 PetscMPIInt rank,size,tag = ((PetscObject)viewer)->tag; 1146 int fd; 1147 PetscInt nz,header[4],*row_lengths,*range=0,rlen,i; 1148 PetscInt nzmax,*column_indices,j,k,col,*garray = aij->garray,cnt,cstart = mat->cmap->rstart,rnz = 0; 1149 PetscScalar *column_values; 1150 PetscInt message_count,flowcontrolcount; 1151 FILE *file; 1152 1153 PetscFunctionBegin; 1154 ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr); 1155 ierr = MPI_Comm_size(PetscObjectComm((PetscObject)mat),&size);CHKERRQ(ierr); 1156 nz = A->nz + B->nz; 1157 ierr = PetscViewerBinaryGetDescriptor(viewer,&fd);CHKERRQ(ierr); 1158 if (!rank) { 1159 header[0] = MAT_FILE_CLASSID; 1160 header[1] = mat->rmap->N; 1161 header[2] = mat->cmap->N; 1162 1163 ierr = MPI_Reduce(&nz,&header[3],1,MPIU_INT,MPI_SUM,0,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1164 ierr = PetscBinaryWrite(fd,header,4,PETSC_INT,PETSC_TRUE);CHKERRQ(ierr); 1165 /* get largest number of rows any processor has */ 1166 rlen = mat->rmap->n; 1167 range = mat->rmap->range; 1168 for (i=1; i<size; i++) rlen = PetscMax(rlen,range[i+1] - range[i]); 1169 } else { 1170 ierr = MPI_Reduce(&nz,0,1,MPIU_INT,MPI_SUM,0,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1171 rlen = mat->rmap->n; 1172 } 1173 1174 /* load up the local row counts */ 1175 ierr = PetscMalloc1(rlen+1,&row_lengths);CHKERRQ(ierr); 1176 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]; 1177 1178 /* store the row lengths to the file */ 1179 ierr = PetscViewerFlowControlStart(viewer,&message_count,&flowcontrolcount);CHKERRQ(ierr); 1180 if (!rank) { 1181 ierr = PetscBinaryWrite(fd,row_lengths,mat->rmap->n,PETSC_INT,PETSC_TRUE);CHKERRQ(ierr); 1182 for (i=1; i<size; i++) { 1183 ierr = PetscViewerFlowControlStepMaster(viewer,i,&message_count,flowcontrolcount);CHKERRQ(ierr); 1184 rlen = range[i+1] - range[i]; 1185 ierr = MPIULong_Recv(row_lengths,rlen,MPIU_INT,i,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1186 ierr = PetscBinaryWrite(fd,row_lengths,rlen,PETSC_INT,PETSC_TRUE);CHKERRQ(ierr); 1187 } 1188 ierr = PetscViewerFlowControlEndMaster(viewer,&message_count);CHKERRQ(ierr); 1189 } else { 1190 ierr = PetscViewerFlowControlStepWorker(viewer,rank,&message_count);CHKERRQ(ierr); 1191 ierr = MPIULong_Send(row_lengths,mat->rmap->n,MPIU_INT,0,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1192 ierr = PetscViewerFlowControlEndWorker(viewer,&message_count);CHKERRQ(ierr); 1193 } 1194 ierr = PetscFree(row_lengths);CHKERRQ(ierr); 1195 1196 /* load up the local column indices */ 1197 nzmax = nz; /* th processor needs space a largest processor needs */ 1198 ierr = MPI_Reduce(&nz,&nzmax,1,MPIU_INT,MPI_MAX,0,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1199 ierr = PetscMalloc1(nzmax+1,&column_indices);CHKERRQ(ierr); 1200 cnt = 0; 1201 for (i=0; i<mat->rmap->n; i++) { 1202 for (j=B->i[i]; j<B->i[i+1]; j++) { 1203 if ((col = garray[B->j[j]]) > cstart) break; 1204 column_indices[cnt++] = col; 1205 } 1206 for (k=A->i[i]; k<A->i[i+1]; k++) column_indices[cnt++] = A->j[k] + cstart; 1207 for (; j<B->i[i+1]; j++) column_indices[cnt++] = garray[B->j[j]]; 1208 } 1209 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); 1210 1211 /* store the column indices to the file */ 1212 ierr = PetscViewerFlowControlStart(viewer,&message_count,&flowcontrolcount);CHKERRQ(ierr); 1213 if (!rank) { 1214 MPI_Status status; 1215 ierr = PetscBinaryWrite(fd,column_indices,nz,PETSC_INT,PETSC_TRUE);CHKERRQ(ierr); 1216 for (i=1; i<size; i++) { 1217 ierr = PetscViewerFlowControlStepMaster(viewer,i,&message_count,flowcontrolcount);CHKERRQ(ierr); 1218 ierr = MPI_Recv(&rnz,1,MPIU_INT,i,tag,PetscObjectComm((PetscObject)mat),&status);CHKERRQ(ierr); 1219 if (rnz > nzmax) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_LIB,"Internal PETSc error: nz = %D nzmax = %D",nz,nzmax); 1220 ierr = MPIULong_Recv(column_indices,rnz,MPIU_INT,i,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1221 ierr = PetscBinaryWrite(fd,column_indices,rnz,PETSC_INT,PETSC_TRUE);CHKERRQ(ierr); 1222 } 1223 ierr = PetscViewerFlowControlEndMaster(viewer,&message_count);CHKERRQ(ierr); 1224 } else { 1225 ierr = PetscViewerFlowControlStepWorker(viewer,rank,&message_count);CHKERRQ(ierr); 1226 ierr = MPI_Send(&nz,1,MPIU_INT,0,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1227 ierr = MPIULong_Send(column_indices,nz,MPIU_INT,0,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1228 ierr = PetscViewerFlowControlEndWorker(viewer,&message_count);CHKERRQ(ierr); 1229 } 1230 ierr = PetscFree(column_indices);CHKERRQ(ierr); 1231 1232 /* load up the local column values */ 1233 ierr = PetscMalloc1(nzmax+1,&column_values);CHKERRQ(ierr); 1234 cnt = 0; 1235 for (i=0; i<mat->rmap->n; i++) { 1236 for (j=B->i[i]; j<B->i[i+1]; j++) { 1237 if (garray[B->j[j]] > cstart) break; 1238 column_values[cnt++] = B->a[j]; 1239 } 1240 for (k=A->i[i]; k<A->i[i+1]; k++) column_values[cnt++] = A->a[k]; 1241 for (; j<B->i[i+1]; j++) column_values[cnt++] = B->a[j]; 1242 } 1243 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); 1244 1245 /* store the column values to the file */ 1246 ierr = PetscViewerFlowControlStart(viewer,&message_count,&flowcontrolcount);CHKERRQ(ierr); 1247 if (!rank) { 1248 MPI_Status status; 1249 ierr = PetscBinaryWrite(fd,column_values,nz,PETSC_SCALAR,PETSC_TRUE);CHKERRQ(ierr); 1250 for (i=1; i<size; i++) { 1251 ierr = PetscViewerFlowControlStepMaster(viewer,i,&message_count,flowcontrolcount);CHKERRQ(ierr); 1252 ierr = MPI_Recv(&rnz,1,MPIU_INT,i,tag,PetscObjectComm((PetscObject)mat),&status);CHKERRQ(ierr); 1253 if (rnz > nzmax) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_LIB,"Internal PETSc error: nz = %D nzmax = %D",nz,nzmax); 1254 ierr = MPIULong_Recv(column_values,rnz,MPIU_SCALAR,i,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1255 ierr = PetscBinaryWrite(fd,column_values,rnz,PETSC_SCALAR,PETSC_TRUE);CHKERRQ(ierr); 1256 } 1257 ierr = PetscViewerFlowControlEndMaster(viewer,&message_count);CHKERRQ(ierr); 1258 } else { 1259 ierr = PetscViewerFlowControlStepWorker(viewer,rank,&message_count);CHKERRQ(ierr); 1260 ierr = MPI_Send(&nz,1,MPIU_INT,0,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1261 ierr = MPIULong_Send(column_values,nz,MPIU_SCALAR,0,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1262 ierr = PetscViewerFlowControlEndWorker(viewer,&message_count);CHKERRQ(ierr); 1263 } 1264 ierr = PetscFree(column_values);CHKERRQ(ierr); 1265 1266 ierr = PetscViewerBinaryGetInfoPointer(viewer,&file);CHKERRQ(ierr); 1267 if (file) fprintf(file,"-matload_block_size %d\n",(int)PetscAbs(mat->rmap->bs)); 1268 PetscFunctionReturn(0); 1269 } 1270 1271 #include <petscdraw.h> 1272 PetscErrorCode MatView_MPIAIJ_ASCIIorDraworSocket(Mat mat,PetscViewer viewer) 1273 { 1274 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 1275 PetscErrorCode ierr; 1276 PetscMPIInt rank = aij->rank,size = aij->size; 1277 PetscBool isdraw,iascii,isbinary; 1278 PetscViewer sviewer; 1279 PetscViewerFormat format; 1280 1281 PetscFunctionBegin; 1282 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERDRAW,&isdraw);CHKERRQ(ierr); 1283 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii);CHKERRQ(ierr); 1284 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERBINARY,&isbinary);CHKERRQ(ierr); 1285 if (iascii) { 1286 ierr = PetscViewerGetFormat(viewer,&format);CHKERRQ(ierr); 1287 if (format == PETSC_VIEWER_ASCII_INFO_DETAIL) { 1288 MatInfo info; 1289 PetscBool inodes; 1290 1291 ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr); 1292 ierr = MatGetInfo(mat,MAT_LOCAL,&info);CHKERRQ(ierr); 1293 ierr = MatInodeGetInodeSizes(aij->A,NULL,(PetscInt**)&inodes,NULL);CHKERRQ(ierr); 1294 ierr = PetscViewerASCIIPushSynchronized(viewer);CHKERRQ(ierr); 1295 if (!inodes) { 1296 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"[%d] Local rows %D nz %D nz alloced %D mem %D, not using I-node routines\n", 1297 rank,mat->rmap->n,(PetscInt)info.nz_used,(PetscInt)info.nz_allocated,(PetscInt)info.memory);CHKERRQ(ierr); 1298 } else { 1299 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"[%d] Local rows %D nz %D nz alloced %D mem %D, using I-node routines\n", 1300 rank,mat->rmap->n,(PetscInt)info.nz_used,(PetscInt)info.nz_allocated,(PetscInt)info.memory);CHKERRQ(ierr); 1301 } 1302 ierr = MatGetInfo(aij->A,MAT_LOCAL,&info);CHKERRQ(ierr); 1303 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"[%d] on-diagonal part: nz %D \n",rank,(PetscInt)info.nz_used);CHKERRQ(ierr); 1304 ierr = MatGetInfo(aij->B,MAT_LOCAL,&info);CHKERRQ(ierr); 1305 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"[%d] off-diagonal part: nz %D \n",rank,(PetscInt)info.nz_used);CHKERRQ(ierr); 1306 ierr = PetscViewerFlush(viewer);CHKERRQ(ierr); 1307 ierr = PetscViewerASCIIPopSynchronized(viewer);CHKERRQ(ierr); 1308 ierr = PetscViewerASCIIPrintf(viewer,"Information on VecScatter used in matrix-vector product: \n");CHKERRQ(ierr); 1309 ierr = VecScatterView(aij->Mvctx,viewer);CHKERRQ(ierr); 1310 PetscFunctionReturn(0); 1311 } else if (format == PETSC_VIEWER_ASCII_INFO) { 1312 PetscInt inodecount,inodelimit,*inodes; 1313 ierr = MatInodeGetInodeSizes(aij->A,&inodecount,&inodes,&inodelimit);CHKERRQ(ierr); 1314 if (inodes) { 1315 ierr = PetscViewerASCIIPrintf(viewer,"using I-node (on process 0) routines: found %D nodes, limit used is %D\n",inodecount,inodelimit);CHKERRQ(ierr); 1316 } else { 1317 ierr = PetscViewerASCIIPrintf(viewer,"not using I-node (on process 0) routines\n");CHKERRQ(ierr); 1318 } 1319 PetscFunctionReturn(0); 1320 } else if (format == PETSC_VIEWER_ASCII_FACTOR_INFO) { 1321 PetscFunctionReturn(0); 1322 } 1323 } else if (isbinary) { 1324 if (size == 1) { 1325 ierr = PetscObjectSetName((PetscObject)aij->A,((PetscObject)mat)->name);CHKERRQ(ierr); 1326 ierr = MatView(aij->A,viewer);CHKERRQ(ierr); 1327 } else { 1328 ierr = MatView_MPIAIJ_Binary(mat,viewer);CHKERRQ(ierr); 1329 } 1330 PetscFunctionReturn(0); 1331 } else if (isdraw) { 1332 PetscDraw draw; 1333 PetscBool isnull; 1334 ierr = PetscViewerDrawGetDraw(viewer,0,&draw);CHKERRQ(ierr); 1335 ierr = PetscDrawIsNull(draw,&isnull);CHKERRQ(ierr); 1336 if (isnull) PetscFunctionReturn(0); 1337 } 1338 1339 { 1340 /* assemble the entire matrix onto first processor. */ 1341 Mat A; 1342 Mat_SeqAIJ *Aloc; 1343 PetscInt M = mat->rmap->N,N = mat->cmap->N,m,*ai,*aj,row,*cols,i,*ct; 1344 MatScalar *a; 1345 1346 ierr = MatCreate(PetscObjectComm((PetscObject)mat),&A);CHKERRQ(ierr); 1347 if (!rank) { 1348 ierr = MatSetSizes(A,M,N,M,N);CHKERRQ(ierr); 1349 } else { 1350 ierr = MatSetSizes(A,0,0,M,N);CHKERRQ(ierr); 1351 } 1352 /* This is just a temporary matrix, so explicitly using MATMPIAIJ is probably best */ 1353 ierr = MatSetType(A,MATMPIAIJ);CHKERRQ(ierr); 1354 ierr = MatMPIAIJSetPreallocation(A,0,NULL,0,NULL);CHKERRQ(ierr); 1355 ierr = MatSetOption(A,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 1356 ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)A);CHKERRQ(ierr); 1357 1358 /* copy over the A part */ 1359 Aloc = (Mat_SeqAIJ*)aij->A->data; 1360 m = aij->A->rmap->n; ai = Aloc->i; aj = Aloc->j; a = Aloc->a; 1361 row = mat->rmap->rstart; 1362 for (i=0; i<ai[m]; i++) aj[i] += mat->cmap->rstart; 1363 for (i=0; i<m; i++) { 1364 ierr = MatSetValues(A,1,&row,ai[i+1]-ai[i],aj,a,INSERT_VALUES);CHKERRQ(ierr); 1365 row++; 1366 a += ai[i+1]-ai[i]; aj += ai[i+1]-ai[i]; 1367 } 1368 aj = Aloc->j; 1369 for (i=0; i<ai[m]; i++) aj[i] -= mat->cmap->rstart; 1370 1371 /* copy over the B part */ 1372 Aloc = (Mat_SeqAIJ*)aij->B->data; 1373 m = aij->B->rmap->n; ai = Aloc->i; aj = Aloc->j; a = Aloc->a; 1374 row = mat->rmap->rstart; 1375 ierr = PetscMalloc1(ai[m]+1,&cols);CHKERRQ(ierr); 1376 ct = cols; 1377 for (i=0; i<ai[m]; i++) cols[i] = aij->garray[aj[i]]; 1378 for (i=0; i<m; i++) { 1379 ierr = MatSetValues(A,1,&row,ai[i+1]-ai[i],cols,a,INSERT_VALUES);CHKERRQ(ierr); 1380 row++; 1381 a += ai[i+1]-ai[i]; cols += ai[i+1]-ai[i]; 1382 } 1383 ierr = PetscFree(ct);CHKERRQ(ierr); 1384 ierr = MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1385 ierr = MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1386 /* 1387 Everyone has to call to draw the matrix since the graphics waits are 1388 synchronized across all processors that share the PetscDraw object 1389 */ 1390 ierr = PetscViewerGetSubViewer(viewer,PETSC_COMM_SELF,&sviewer);CHKERRQ(ierr); 1391 if (!rank) { 1392 ierr = PetscObjectSetName((PetscObject)((Mat_MPIAIJ*)(A->data))->A,((PetscObject)mat)->name);CHKERRQ(ierr); 1393 ierr = MatView_SeqAIJ(((Mat_MPIAIJ*)(A->data))->A,sviewer);CHKERRQ(ierr); 1394 } 1395 ierr = PetscViewerRestoreSubViewer(viewer,PETSC_COMM_SELF,&sviewer);CHKERRQ(ierr); 1396 ierr = PetscViewerFlush(viewer);CHKERRQ(ierr); 1397 ierr = MatDestroy(&A);CHKERRQ(ierr); 1398 } 1399 PetscFunctionReturn(0); 1400 } 1401 1402 PetscErrorCode MatView_MPIAIJ(Mat mat,PetscViewer viewer) 1403 { 1404 PetscErrorCode ierr; 1405 PetscBool iascii,isdraw,issocket,isbinary; 1406 1407 PetscFunctionBegin; 1408 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii);CHKERRQ(ierr); 1409 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERDRAW,&isdraw);CHKERRQ(ierr); 1410 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERBINARY,&isbinary);CHKERRQ(ierr); 1411 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERSOCKET,&issocket);CHKERRQ(ierr); 1412 if (iascii || isdraw || isbinary || issocket) { 1413 ierr = MatView_MPIAIJ_ASCIIorDraworSocket(mat,viewer);CHKERRQ(ierr); 1414 } 1415 PetscFunctionReturn(0); 1416 } 1417 1418 PetscErrorCode MatSOR_MPIAIJ(Mat matin,Vec bb,PetscReal omega,MatSORType flag,PetscReal fshift,PetscInt its,PetscInt lits,Vec xx) 1419 { 1420 Mat_MPIAIJ *mat = (Mat_MPIAIJ*)matin->data; 1421 PetscErrorCode ierr; 1422 Vec bb1 = 0; 1423 PetscBool hasop; 1424 1425 PetscFunctionBegin; 1426 if (flag == SOR_APPLY_UPPER) { 1427 ierr = (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);CHKERRQ(ierr); 1428 PetscFunctionReturn(0); 1429 } 1430 1431 if (its > 1 || ~flag & SOR_ZERO_INITIAL_GUESS || flag & SOR_EISENSTAT) { 1432 ierr = VecDuplicate(bb,&bb1);CHKERRQ(ierr); 1433 } 1434 1435 if ((flag & SOR_LOCAL_SYMMETRIC_SWEEP) == SOR_LOCAL_SYMMETRIC_SWEEP) { 1436 if (flag & SOR_ZERO_INITIAL_GUESS) { 1437 ierr = (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);CHKERRQ(ierr); 1438 its--; 1439 } 1440 1441 while (its--) { 1442 ierr = VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1443 ierr = VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1444 1445 /* update rhs: bb1 = bb - B*x */ 1446 ierr = VecScale(mat->lvec,-1.0);CHKERRQ(ierr); 1447 ierr = (*mat->B->ops->multadd)(mat->B,mat->lvec,bb,bb1);CHKERRQ(ierr); 1448 1449 /* local sweep */ 1450 ierr = (*mat->A->ops->sor)(mat->A,bb1,omega,SOR_SYMMETRIC_SWEEP,fshift,lits,1,xx);CHKERRQ(ierr); 1451 } 1452 } else if (flag & SOR_LOCAL_FORWARD_SWEEP) { 1453 if (flag & SOR_ZERO_INITIAL_GUESS) { 1454 ierr = (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);CHKERRQ(ierr); 1455 its--; 1456 } 1457 while (its--) { 1458 ierr = VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1459 ierr = VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1460 1461 /* update rhs: bb1 = bb - B*x */ 1462 ierr = VecScale(mat->lvec,-1.0);CHKERRQ(ierr); 1463 ierr = (*mat->B->ops->multadd)(mat->B,mat->lvec,bb,bb1);CHKERRQ(ierr); 1464 1465 /* local sweep */ 1466 ierr = (*mat->A->ops->sor)(mat->A,bb1,omega,SOR_FORWARD_SWEEP,fshift,lits,1,xx);CHKERRQ(ierr); 1467 } 1468 } else if (flag & SOR_LOCAL_BACKWARD_SWEEP) { 1469 if (flag & SOR_ZERO_INITIAL_GUESS) { 1470 ierr = (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);CHKERRQ(ierr); 1471 its--; 1472 } 1473 while (its--) { 1474 ierr = VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1475 ierr = VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1476 1477 /* update rhs: bb1 = bb - B*x */ 1478 ierr = VecScale(mat->lvec,-1.0);CHKERRQ(ierr); 1479 ierr = (*mat->B->ops->multadd)(mat->B,mat->lvec,bb,bb1);CHKERRQ(ierr); 1480 1481 /* local sweep */ 1482 ierr = (*mat->A->ops->sor)(mat->A,bb1,omega,SOR_BACKWARD_SWEEP,fshift,lits,1,xx);CHKERRQ(ierr); 1483 } 1484 } else if (flag & SOR_EISENSTAT) { 1485 Vec xx1; 1486 1487 ierr = VecDuplicate(bb,&xx1);CHKERRQ(ierr); 1488 ierr = (*mat->A->ops->sor)(mat->A,bb,omega,(MatSORType)(SOR_ZERO_INITIAL_GUESS | SOR_LOCAL_BACKWARD_SWEEP),fshift,lits,1,xx);CHKERRQ(ierr); 1489 1490 ierr = VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1491 ierr = VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1492 if (!mat->diag) { 1493 ierr = MatCreateVecs(matin,&mat->diag,NULL);CHKERRQ(ierr); 1494 ierr = MatGetDiagonal(matin,mat->diag);CHKERRQ(ierr); 1495 } 1496 ierr = MatHasOperation(matin,MATOP_MULT_DIAGONAL_BLOCK,&hasop);CHKERRQ(ierr); 1497 if (hasop) { 1498 ierr = MatMultDiagonalBlock(matin,xx,bb1);CHKERRQ(ierr); 1499 } else { 1500 ierr = VecPointwiseMult(bb1,mat->diag,xx);CHKERRQ(ierr); 1501 } 1502 ierr = VecAYPX(bb1,(omega-2.0)/omega,bb);CHKERRQ(ierr); 1503 1504 ierr = MatMultAdd(mat->B,mat->lvec,bb1,bb1);CHKERRQ(ierr); 1505 1506 /* local sweep */ 1507 ierr = (*mat->A->ops->sor)(mat->A,bb1,omega,(MatSORType)(SOR_ZERO_INITIAL_GUESS | SOR_LOCAL_FORWARD_SWEEP),fshift,lits,1,xx1);CHKERRQ(ierr); 1508 ierr = VecAXPY(xx,1.0,xx1);CHKERRQ(ierr); 1509 ierr = VecDestroy(&xx1);CHKERRQ(ierr); 1510 } else SETERRQ(PetscObjectComm((PetscObject)matin),PETSC_ERR_SUP,"Parallel SOR not supported"); 1511 1512 ierr = VecDestroy(&bb1);CHKERRQ(ierr); 1513 1514 matin->factorerrortype = mat->A->factorerrortype; 1515 PetscFunctionReturn(0); 1516 } 1517 1518 PetscErrorCode MatPermute_MPIAIJ(Mat A,IS rowp,IS colp,Mat *B) 1519 { 1520 Mat aA,aB,Aperm; 1521 const PetscInt *rwant,*cwant,*gcols,*ai,*bi,*aj,*bj; 1522 PetscScalar *aa,*ba; 1523 PetscInt i,j,m,n,ng,anz,bnz,*dnnz,*onnz,*tdnnz,*tonnz,*rdest,*cdest,*work,*gcdest; 1524 PetscSF rowsf,sf; 1525 IS parcolp = NULL; 1526 PetscBool done; 1527 PetscErrorCode ierr; 1528 1529 PetscFunctionBegin; 1530 ierr = MatGetLocalSize(A,&m,&n);CHKERRQ(ierr); 1531 ierr = ISGetIndices(rowp,&rwant);CHKERRQ(ierr); 1532 ierr = ISGetIndices(colp,&cwant);CHKERRQ(ierr); 1533 ierr = PetscMalloc3(PetscMax(m,n),&work,m,&rdest,n,&cdest);CHKERRQ(ierr); 1534 1535 /* Invert row permutation to find out where my rows should go */ 1536 ierr = PetscSFCreate(PetscObjectComm((PetscObject)A),&rowsf);CHKERRQ(ierr); 1537 ierr = PetscSFSetGraphLayout(rowsf,A->rmap,A->rmap->n,NULL,PETSC_OWN_POINTER,rwant);CHKERRQ(ierr); 1538 ierr = PetscSFSetFromOptions(rowsf);CHKERRQ(ierr); 1539 for (i=0; i<m; i++) work[i] = A->rmap->rstart + i; 1540 ierr = PetscSFReduceBegin(rowsf,MPIU_INT,work,rdest,MPIU_REPLACE);CHKERRQ(ierr); 1541 ierr = PetscSFReduceEnd(rowsf,MPIU_INT,work,rdest,MPIU_REPLACE);CHKERRQ(ierr); 1542 1543 /* Invert column permutation to find out where my columns should go */ 1544 ierr = PetscSFCreate(PetscObjectComm((PetscObject)A),&sf);CHKERRQ(ierr); 1545 ierr = PetscSFSetGraphLayout(sf,A->cmap,A->cmap->n,NULL,PETSC_OWN_POINTER,cwant);CHKERRQ(ierr); 1546 ierr = PetscSFSetFromOptions(sf);CHKERRQ(ierr); 1547 for (i=0; i<n; i++) work[i] = A->cmap->rstart + i; 1548 ierr = PetscSFReduceBegin(sf,MPIU_INT,work,cdest,MPIU_REPLACE);CHKERRQ(ierr); 1549 ierr = PetscSFReduceEnd(sf,MPIU_INT,work,cdest,MPIU_REPLACE);CHKERRQ(ierr); 1550 ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 1551 1552 ierr = ISRestoreIndices(rowp,&rwant);CHKERRQ(ierr); 1553 ierr = ISRestoreIndices(colp,&cwant);CHKERRQ(ierr); 1554 ierr = MatMPIAIJGetSeqAIJ(A,&aA,&aB,&gcols);CHKERRQ(ierr); 1555 1556 /* Find out where my gcols should go */ 1557 ierr = MatGetSize(aB,NULL,&ng);CHKERRQ(ierr); 1558 ierr = PetscMalloc1(ng,&gcdest);CHKERRQ(ierr); 1559 ierr = PetscSFCreate(PetscObjectComm((PetscObject)A),&sf);CHKERRQ(ierr); 1560 ierr = PetscSFSetGraphLayout(sf,A->cmap,ng,NULL,PETSC_OWN_POINTER,gcols);CHKERRQ(ierr); 1561 ierr = PetscSFSetFromOptions(sf);CHKERRQ(ierr); 1562 ierr = PetscSFBcastBegin(sf,MPIU_INT,cdest,gcdest);CHKERRQ(ierr); 1563 ierr = PetscSFBcastEnd(sf,MPIU_INT,cdest,gcdest);CHKERRQ(ierr); 1564 ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 1565 1566 ierr = PetscCalloc4(m,&dnnz,m,&onnz,m,&tdnnz,m,&tonnz);CHKERRQ(ierr); 1567 ierr = MatGetRowIJ(aA,0,PETSC_FALSE,PETSC_FALSE,&anz,&ai,&aj,&done);CHKERRQ(ierr); 1568 ierr = MatGetRowIJ(aB,0,PETSC_FALSE,PETSC_FALSE,&bnz,&bi,&bj,&done);CHKERRQ(ierr); 1569 for (i=0; i<m; i++) { 1570 PetscInt row = rdest[i],rowner; 1571 ierr = PetscLayoutFindOwner(A->rmap,row,&rowner);CHKERRQ(ierr); 1572 for (j=ai[i]; j<ai[i+1]; j++) { 1573 PetscInt cowner,col = cdest[aj[j]]; 1574 ierr = PetscLayoutFindOwner(A->cmap,col,&cowner);CHKERRQ(ierr); /* Could build an index for the columns to eliminate this search */ 1575 if (rowner == cowner) dnnz[i]++; 1576 else onnz[i]++; 1577 } 1578 for (j=bi[i]; j<bi[i+1]; j++) { 1579 PetscInt cowner,col = gcdest[bj[j]]; 1580 ierr = PetscLayoutFindOwner(A->cmap,col,&cowner);CHKERRQ(ierr); 1581 if (rowner == cowner) dnnz[i]++; 1582 else onnz[i]++; 1583 } 1584 } 1585 ierr = PetscSFBcastBegin(rowsf,MPIU_INT,dnnz,tdnnz);CHKERRQ(ierr); 1586 ierr = PetscSFBcastEnd(rowsf,MPIU_INT,dnnz,tdnnz);CHKERRQ(ierr); 1587 ierr = PetscSFBcastBegin(rowsf,MPIU_INT,onnz,tonnz);CHKERRQ(ierr); 1588 ierr = PetscSFBcastEnd(rowsf,MPIU_INT,onnz,tonnz);CHKERRQ(ierr); 1589 ierr = PetscSFDestroy(&rowsf);CHKERRQ(ierr); 1590 1591 ierr = MatCreateAIJ(PetscObjectComm((PetscObject)A),A->rmap->n,A->cmap->n,A->rmap->N,A->cmap->N,0,tdnnz,0,tonnz,&Aperm);CHKERRQ(ierr); 1592 ierr = MatSeqAIJGetArray(aA,&aa);CHKERRQ(ierr); 1593 ierr = MatSeqAIJGetArray(aB,&ba);CHKERRQ(ierr); 1594 for (i=0; i<m; i++) { 1595 PetscInt *acols = dnnz,*bcols = onnz; /* Repurpose now-unneeded arrays */ 1596 PetscInt j0,rowlen; 1597 rowlen = ai[i+1] - ai[i]; 1598 for (j0=j=0; j<rowlen; j0=j) { /* rowlen could be larger than number of rows m, so sum in batches */ 1599 for ( ; j<PetscMin(rowlen,j0+m); j++) acols[j-j0] = cdest[aj[ai[i]+j]]; 1600 ierr = MatSetValues(Aperm,1,&rdest[i],j-j0,acols,aa+ai[i]+j0,INSERT_VALUES);CHKERRQ(ierr); 1601 } 1602 rowlen = bi[i+1] - bi[i]; 1603 for (j0=j=0; j<rowlen; j0=j) { 1604 for ( ; j<PetscMin(rowlen,j0+m); j++) bcols[j-j0] = gcdest[bj[bi[i]+j]]; 1605 ierr = MatSetValues(Aperm,1,&rdest[i],j-j0,bcols,ba+bi[i]+j0,INSERT_VALUES);CHKERRQ(ierr); 1606 } 1607 } 1608 ierr = MatAssemblyBegin(Aperm,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1609 ierr = MatAssemblyEnd(Aperm,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1610 ierr = MatRestoreRowIJ(aA,0,PETSC_FALSE,PETSC_FALSE,&anz,&ai,&aj,&done);CHKERRQ(ierr); 1611 ierr = MatRestoreRowIJ(aB,0,PETSC_FALSE,PETSC_FALSE,&bnz,&bi,&bj,&done);CHKERRQ(ierr); 1612 ierr = MatSeqAIJRestoreArray(aA,&aa);CHKERRQ(ierr); 1613 ierr = MatSeqAIJRestoreArray(aB,&ba);CHKERRQ(ierr); 1614 ierr = PetscFree4(dnnz,onnz,tdnnz,tonnz);CHKERRQ(ierr); 1615 ierr = PetscFree3(work,rdest,cdest);CHKERRQ(ierr); 1616 ierr = PetscFree(gcdest);CHKERRQ(ierr); 1617 if (parcolp) {ierr = ISDestroy(&colp);CHKERRQ(ierr);} 1618 *B = Aperm; 1619 PetscFunctionReturn(0); 1620 } 1621 1622 PetscErrorCode MatGetGhosts_MPIAIJ(Mat mat,PetscInt *nghosts,const PetscInt *ghosts[]) 1623 { 1624 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 1625 PetscErrorCode ierr; 1626 1627 PetscFunctionBegin; 1628 ierr = MatGetSize(aij->B,NULL,nghosts);CHKERRQ(ierr); 1629 if (ghosts) *ghosts = aij->garray; 1630 PetscFunctionReturn(0); 1631 } 1632 1633 PetscErrorCode MatGetInfo_MPIAIJ(Mat matin,MatInfoType flag,MatInfo *info) 1634 { 1635 Mat_MPIAIJ *mat = (Mat_MPIAIJ*)matin->data; 1636 Mat A = mat->A,B = mat->B; 1637 PetscErrorCode ierr; 1638 PetscReal isend[5],irecv[5]; 1639 1640 PetscFunctionBegin; 1641 info->block_size = 1.0; 1642 ierr = MatGetInfo(A,MAT_LOCAL,info);CHKERRQ(ierr); 1643 1644 isend[0] = info->nz_used; isend[1] = info->nz_allocated; isend[2] = info->nz_unneeded; 1645 isend[3] = info->memory; isend[4] = info->mallocs; 1646 1647 ierr = MatGetInfo(B,MAT_LOCAL,info);CHKERRQ(ierr); 1648 1649 isend[0] += info->nz_used; isend[1] += info->nz_allocated; isend[2] += info->nz_unneeded; 1650 isend[3] += info->memory; isend[4] += info->mallocs; 1651 if (flag == MAT_LOCAL) { 1652 info->nz_used = isend[0]; 1653 info->nz_allocated = isend[1]; 1654 info->nz_unneeded = isend[2]; 1655 info->memory = isend[3]; 1656 info->mallocs = isend[4]; 1657 } else if (flag == MAT_GLOBAL_MAX) { 1658 ierr = MPIU_Allreduce(isend,irecv,5,MPIU_REAL,MPIU_MAX,PetscObjectComm((PetscObject)matin));CHKERRQ(ierr); 1659 1660 info->nz_used = irecv[0]; 1661 info->nz_allocated = irecv[1]; 1662 info->nz_unneeded = irecv[2]; 1663 info->memory = irecv[3]; 1664 info->mallocs = irecv[4]; 1665 } else if (flag == MAT_GLOBAL_SUM) { 1666 ierr = MPIU_Allreduce(isend,irecv,5,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)matin));CHKERRQ(ierr); 1667 1668 info->nz_used = irecv[0]; 1669 info->nz_allocated = irecv[1]; 1670 info->nz_unneeded = irecv[2]; 1671 info->memory = irecv[3]; 1672 info->mallocs = irecv[4]; 1673 } 1674 info->fill_ratio_given = 0; /* no parallel LU/ILU/Cholesky */ 1675 info->fill_ratio_needed = 0; 1676 info->factor_mallocs = 0; 1677 PetscFunctionReturn(0); 1678 } 1679 1680 PetscErrorCode MatSetOption_MPIAIJ(Mat A,MatOption op,PetscBool flg) 1681 { 1682 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 1683 PetscErrorCode ierr; 1684 1685 PetscFunctionBegin; 1686 switch (op) { 1687 case MAT_NEW_NONZERO_LOCATIONS: 1688 case MAT_NEW_NONZERO_ALLOCATION_ERR: 1689 case MAT_UNUSED_NONZERO_LOCATION_ERR: 1690 case MAT_KEEP_NONZERO_PATTERN: 1691 case MAT_NEW_NONZERO_LOCATION_ERR: 1692 case MAT_USE_INODES: 1693 case MAT_IGNORE_ZERO_ENTRIES: 1694 MatCheckPreallocated(A,1); 1695 ierr = MatSetOption(a->A,op,flg);CHKERRQ(ierr); 1696 ierr = MatSetOption(a->B,op,flg);CHKERRQ(ierr); 1697 break; 1698 case MAT_ROW_ORIENTED: 1699 MatCheckPreallocated(A,1); 1700 a->roworiented = flg; 1701 1702 ierr = MatSetOption(a->A,op,flg);CHKERRQ(ierr); 1703 ierr = MatSetOption(a->B,op,flg);CHKERRQ(ierr); 1704 break; 1705 case MAT_NEW_DIAGONALS: 1706 ierr = PetscInfo1(A,"Option %s ignored\n",MatOptions[op]);CHKERRQ(ierr); 1707 break; 1708 case MAT_IGNORE_OFF_PROC_ENTRIES: 1709 a->donotstash = flg; 1710 break; 1711 case MAT_SPD: 1712 A->spd_set = PETSC_TRUE; 1713 A->spd = flg; 1714 if (flg) { 1715 A->symmetric = PETSC_TRUE; 1716 A->structurally_symmetric = PETSC_TRUE; 1717 A->symmetric_set = PETSC_TRUE; 1718 A->structurally_symmetric_set = PETSC_TRUE; 1719 } 1720 break; 1721 case MAT_SYMMETRIC: 1722 MatCheckPreallocated(A,1); 1723 ierr = MatSetOption(a->A,op,flg);CHKERRQ(ierr); 1724 break; 1725 case MAT_STRUCTURALLY_SYMMETRIC: 1726 MatCheckPreallocated(A,1); 1727 ierr = MatSetOption(a->A,op,flg);CHKERRQ(ierr); 1728 break; 1729 case MAT_HERMITIAN: 1730 MatCheckPreallocated(A,1); 1731 ierr = MatSetOption(a->A,op,flg);CHKERRQ(ierr); 1732 break; 1733 case MAT_SYMMETRY_ETERNAL: 1734 MatCheckPreallocated(A,1); 1735 ierr = MatSetOption(a->A,op,flg);CHKERRQ(ierr); 1736 break; 1737 case MAT_SUBMAT_SINGLEIS: 1738 A->submat_singleis = flg; 1739 break; 1740 case MAT_STRUCTURE_ONLY: 1741 /* The option is handled directly by MatSetOption() */ 1742 break; 1743 default: 1744 SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"unknown option %d",op); 1745 } 1746 PetscFunctionReturn(0); 1747 } 1748 1749 PetscErrorCode MatGetRow_MPIAIJ(Mat matin,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v) 1750 { 1751 Mat_MPIAIJ *mat = (Mat_MPIAIJ*)matin->data; 1752 PetscScalar *vworkA,*vworkB,**pvA,**pvB,*v_p; 1753 PetscErrorCode ierr; 1754 PetscInt i,*cworkA,*cworkB,**pcA,**pcB,cstart = matin->cmap->rstart; 1755 PetscInt nztot,nzA,nzB,lrow,rstart = matin->rmap->rstart,rend = matin->rmap->rend; 1756 PetscInt *cmap,*idx_p; 1757 1758 PetscFunctionBegin; 1759 if (mat->getrowactive) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Already active"); 1760 mat->getrowactive = PETSC_TRUE; 1761 1762 if (!mat->rowvalues && (idx || v)) { 1763 /* 1764 allocate enough space to hold information from the longest row. 1765 */ 1766 Mat_SeqAIJ *Aa = (Mat_SeqAIJ*)mat->A->data,*Ba = (Mat_SeqAIJ*)mat->B->data; 1767 PetscInt max = 1,tmp; 1768 for (i=0; i<matin->rmap->n; i++) { 1769 tmp = Aa->i[i+1] - Aa->i[i] + Ba->i[i+1] - Ba->i[i]; 1770 if (max < tmp) max = tmp; 1771 } 1772 ierr = PetscMalloc2(max,&mat->rowvalues,max,&mat->rowindices);CHKERRQ(ierr); 1773 } 1774 1775 if (row < rstart || row >= rend) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Only local rows"); 1776 lrow = row - rstart; 1777 1778 pvA = &vworkA; pcA = &cworkA; pvB = &vworkB; pcB = &cworkB; 1779 if (!v) {pvA = 0; pvB = 0;} 1780 if (!idx) {pcA = 0; if (!v) pcB = 0;} 1781 ierr = (*mat->A->ops->getrow)(mat->A,lrow,&nzA,pcA,pvA);CHKERRQ(ierr); 1782 ierr = (*mat->B->ops->getrow)(mat->B,lrow,&nzB,pcB,pvB);CHKERRQ(ierr); 1783 nztot = nzA + nzB; 1784 1785 cmap = mat->garray; 1786 if (v || idx) { 1787 if (nztot) { 1788 /* Sort by increasing column numbers, assuming A and B already sorted */ 1789 PetscInt imark = -1; 1790 if (v) { 1791 *v = v_p = mat->rowvalues; 1792 for (i=0; i<nzB; i++) { 1793 if (cmap[cworkB[i]] < cstart) v_p[i] = vworkB[i]; 1794 else break; 1795 } 1796 imark = i; 1797 for (i=0; i<nzA; i++) v_p[imark+i] = vworkA[i]; 1798 for (i=imark; i<nzB; i++) v_p[nzA+i] = vworkB[i]; 1799 } 1800 if (idx) { 1801 *idx = idx_p = mat->rowindices; 1802 if (imark > -1) { 1803 for (i=0; i<imark; i++) { 1804 idx_p[i] = cmap[cworkB[i]]; 1805 } 1806 } else { 1807 for (i=0; i<nzB; i++) { 1808 if (cmap[cworkB[i]] < cstart) idx_p[i] = cmap[cworkB[i]]; 1809 else break; 1810 } 1811 imark = i; 1812 } 1813 for (i=0; i<nzA; i++) idx_p[imark+i] = cstart + cworkA[i]; 1814 for (i=imark; i<nzB; i++) idx_p[nzA+i] = cmap[cworkB[i]]; 1815 } 1816 } else { 1817 if (idx) *idx = 0; 1818 if (v) *v = 0; 1819 } 1820 } 1821 *nz = nztot; 1822 ierr = (*mat->A->ops->restorerow)(mat->A,lrow,&nzA,pcA,pvA);CHKERRQ(ierr); 1823 ierr = (*mat->B->ops->restorerow)(mat->B,lrow,&nzB,pcB,pvB);CHKERRQ(ierr); 1824 PetscFunctionReturn(0); 1825 } 1826 1827 PetscErrorCode MatRestoreRow_MPIAIJ(Mat mat,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v) 1828 { 1829 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 1830 1831 PetscFunctionBegin; 1832 if (!aij->getrowactive) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"MatGetRow() must be called first"); 1833 aij->getrowactive = PETSC_FALSE; 1834 PetscFunctionReturn(0); 1835 } 1836 1837 PetscErrorCode MatNorm_MPIAIJ(Mat mat,NormType type,PetscReal *norm) 1838 { 1839 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 1840 Mat_SeqAIJ *amat = (Mat_SeqAIJ*)aij->A->data,*bmat = (Mat_SeqAIJ*)aij->B->data; 1841 PetscErrorCode ierr; 1842 PetscInt i,j,cstart = mat->cmap->rstart; 1843 PetscReal sum = 0.0; 1844 MatScalar *v; 1845 1846 PetscFunctionBegin; 1847 if (aij->size == 1) { 1848 ierr = MatNorm(aij->A,type,norm);CHKERRQ(ierr); 1849 } else { 1850 if (type == NORM_FROBENIUS) { 1851 v = amat->a; 1852 for (i=0; i<amat->nz; i++) { 1853 sum += PetscRealPart(PetscConj(*v)*(*v)); v++; 1854 } 1855 v = bmat->a; 1856 for (i=0; i<bmat->nz; i++) { 1857 sum += PetscRealPart(PetscConj(*v)*(*v)); v++; 1858 } 1859 ierr = MPIU_Allreduce(&sum,norm,1,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1860 *norm = PetscSqrtReal(*norm); 1861 ierr = PetscLogFlops(2*amat->nz+2*bmat->nz);CHKERRQ(ierr); 1862 } else if (type == NORM_1) { /* max column norm */ 1863 PetscReal *tmp,*tmp2; 1864 PetscInt *jj,*garray = aij->garray; 1865 ierr = PetscCalloc1(mat->cmap->N+1,&tmp);CHKERRQ(ierr); 1866 ierr = PetscMalloc1(mat->cmap->N+1,&tmp2);CHKERRQ(ierr); 1867 *norm = 0.0; 1868 v = amat->a; jj = amat->j; 1869 for (j=0; j<amat->nz; j++) { 1870 tmp[cstart + *jj++] += PetscAbsScalar(*v); v++; 1871 } 1872 v = bmat->a; jj = bmat->j; 1873 for (j=0; j<bmat->nz; j++) { 1874 tmp[garray[*jj++]] += PetscAbsScalar(*v); v++; 1875 } 1876 ierr = MPIU_Allreduce(tmp,tmp2,mat->cmap->N,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1877 for (j=0; j<mat->cmap->N; j++) { 1878 if (tmp2[j] > *norm) *norm = tmp2[j]; 1879 } 1880 ierr = PetscFree(tmp);CHKERRQ(ierr); 1881 ierr = PetscFree(tmp2);CHKERRQ(ierr); 1882 ierr = PetscLogFlops(PetscMax(amat->nz+bmat->nz-1,0));CHKERRQ(ierr); 1883 } else if (type == NORM_INFINITY) { /* max row norm */ 1884 PetscReal ntemp = 0.0; 1885 for (j=0; j<aij->A->rmap->n; j++) { 1886 v = amat->a + amat->i[j]; 1887 sum = 0.0; 1888 for (i=0; i<amat->i[j+1]-amat->i[j]; i++) { 1889 sum += PetscAbsScalar(*v); v++; 1890 } 1891 v = bmat->a + bmat->i[j]; 1892 for (i=0; i<bmat->i[j+1]-bmat->i[j]; i++) { 1893 sum += PetscAbsScalar(*v); v++; 1894 } 1895 if (sum > ntemp) ntemp = sum; 1896 } 1897 ierr = MPIU_Allreduce(&ntemp,norm,1,MPIU_REAL,MPIU_MAX,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1898 ierr = PetscLogFlops(PetscMax(amat->nz+bmat->nz-1,0));CHKERRQ(ierr); 1899 } else SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"No support for two norm"); 1900 } 1901 PetscFunctionReturn(0); 1902 } 1903 1904 PetscErrorCode MatTranspose_MPIAIJ(Mat A,MatReuse reuse,Mat *matout) 1905 { 1906 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 1907 Mat_SeqAIJ *Aloc=(Mat_SeqAIJ*)a->A->data,*Bloc=(Mat_SeqAIJ*)a->B->data; 1908 PetscErrorCode ierr; 1909 PetscInt M = A->rmap->N,N = A->cmap->N,ma,na,mb,nb,*ai,*aj,*bi,*bj,row,*cols,*cols_tmp,i; 1910 PetscInt cstart = A->cmap->rstart,ncol; 1911 Mat B; 1912 MatScalar *array; 1913 1914 PetscFunctionBegin; 1915 if (reuse == MAT_INPLACE_MATRIX && M != N) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_SIZ,"Square matrix only for in-place"); 1916 1917 ma = A->rmap->n; na = A->cmap->n; mb = a->B->rmap->n; nb = a->B->cmap->n; 1918 ai = Aloc->i; aj = Aloc->j; 1919 bi = Bloc->i; bj = Bloc->j; 1920 if (reuse == MAT_INITIAL_MATRIX || *matout == A) { 1921 PetscInt *d_nnz,*g_nnz,*o_nnz; 1922 PetscSFNode *oloc; 1923 PETSC_UNUSED PetscSF sf; 1924 1925 ierr = PetscMalloc4(na,&d_nnz,na,&o_nnz,nb,&g_nnz,nb,&oloc);CHKERRQ(ierr); 1926 /* compute d_nnz for preallocation */ 1927 ierr = PetscMemzero(d_nnz,na*sizeof(PetscInt));CHKERRQ(ierr); 1928 for (i=0; i<ai[ma]; i++) { 1929 d_nnz[aj[i]]++; 1930 aj[i] += cstart; /* global col index to be used by MatSetValues() */ 1931 } 1932 /* compute local off-diagonal contributions */ 1933 ierr = PetscMemzero(g_nnz,nb*sizeof(PetscInt));CHKERRQ(ierr); 1934 for (i=0; i<bi[ma]; i++) g_nnz[bj[i]]++; 1935 /* map those to global */ 1936 ierr = PetscSFCreate(PetscObjectComm((PetscObject)A),&sf);CHKERRQ(ierr); 1937 ierr = PetscSFSetGraphLayout(sf,A->cmap,nb,NULL,PETSC_USE_POINTER,a->garray);CHKERRQ(ierr); 1938 ierr = PetscSFSetFromOptions(sf);CHKERRQ(ierr); 1939 ierr = PetscMemzero(o_nnz,na*sizeof(PetscInt));CHKERRQ(ierr); 1940 ierr = PetscSFReduceBegin(sf,MPIU_INT,g_nnz,o_nnz,MPIU_SUM);CHKERRQ(ierr); 1941 ierr = PetscSFReduceEnd(sf,MPIU_INT,g_nnz,o_nnz,MPIU_SUM);CHKERRQ(ierr); 1942 ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 1943 1944 ierr = MatCreate(PetscObjectComm((PetscObject)A),&B);CHKERRQ(ierr); 1945 ierr = MatSetSizes(B,A->cmap->n,A->rmap->n,N,M);CHKERRQ(ierr); 1946 ierr = MatSetBlockSizes(B,PetscAbs(A->cmap->bs),PetscAbs(A->rmap->bs));CHKERRQ(ierr); 1947 ierr = MatSetType(B,((PetscObject)A)->type_name);CHKERRQ(ierr); 1948 ierr = MatMPIAIJSetPreallocation(B,0,d_nnz,0,o_nnz);CHKERRQ(ierr); 1949 ierr = PetscFree4(d_nnz,o_nnz,g_nnz,oloc);CHKERRQ(ierr); 1950 } else { 1951 B = *matout; 1952 ierr = MatSetOption(B,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_TRUE);CHKERRQ(ierr); 1953 for (i=0; i<ai[ma]; i++) aj[i] += cstart; /* global col index to be used by MatSetValues() */ 1954 } 1955 1956 /* copy over the A part */ 1957 array = Aloc->a; 1958 row = A->rmap->rstart; 1959 for (i=0; i<ma; i++) { 1960 ncol = ai[i+1]-ai[i]; 1961 ierr = MatSetValues(B,ncol,aj,1,&row,array,INSERT_VALUES);CHKERRQ(ierr); 1962 row++; 1963 array += ncol; aj += ncol; 1964 } 1965 aj = Aloc->j; 1966 for (i=0; i<ai[ma]; i++) aj[i] -= cstart; /* resume local col index */ 1967 1968 /* copy over the B part */ 1969 ierr = PetscCalloc1(bi[mb],&cols);CHKERRQ(ierr); 1970 array = Bloc->a; 1971 row = A->rmap->rstart; 1972 for (i=0; i<bi[mb]; i++) cols[i] = a->garray[bj[i]]; 1973 cols_tmp = cols; 1974 for (i=0; i<mb; i++) { 1975 ncol = bi[i+1]-bi[i]; 1976 ierr = MatSetValues(B,ncol,cols_tmp,1,&row,array,INSERT_VALUES);CHKERRQ(ierr); 1977 row++; 1978 array += ncol; cols_tmp += ncol; 1979 } 1980 ierr = PetscFree(cols);CHKERRQ(ierr); 1981 1982 ierr = MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1983 ierr = MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1984 if (reuse == MAT_INITIAL_MATRIX || reuse == MAT_REUSE_MATRIX) { 1985 *matout = B; 1986 } else { 1987 ierr = MatHeaderMerge(A,&B);CHKERRQ(ierr); 1988 } 1989 PetscFunctionReturn(0); 1990 } 1991 1992 PetscErrorCode MatDiagonalScale_MPIAIJ(Mat mat,Vec ll,Vec rr) 1993 { 1994 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 1995 Mat a = aij->A,b = aij->B; 1996 PetscErrorCode ierr; 1997 PetscInt s1,s2,s3; 1998 1999 PetscFunctionBegin; 2000 ierr = MatGetLocalSize(mat,&s2,&s3);CHKERRQ(ierr); 2001 if (rr) { 2002 ierr = VecGetLocalSize(rr,&s1);CHKERRQ(ierr); 2003 if (s1!=s3) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"right vector non-conforming local size"); 2004 /* Overlap communication with computation. */ 2005 ierr = VecScatterBegin(aij->Mvctx,rr,aij->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2006 } 2007 if (ll) { 2008 ierr = VecGetLocalSize(ll,&s1);CHKERRQ(ierr); 2009 if (s1!=s2) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"left vector non-conforming local size"); 2010 ierr = (*b->ops->diagonalscale)(b,ll,0);CHKERRQ(ierr); 2011 } 2012 /* scale the diagonal block */ 2013 ierr = (*a->ops->diagonalscale)(a,ll,rr);CHKERRQ(ierr); 2014 2015 if (rr) { 2016 /* Do a scatter end and then right scale the off-diagonal block */ 2017 ierr = VecScatterEnd(aij->Mvctx,rr,aij->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2018 ierr = (*b->ops->diagonalscale)(b,0,aij->lvec);CHKERRQ(ierr); 2019 } 2020 PetscFunctionReturn(0); 2021 } 2022 2023 PetscErrorCode MatSetUnfactored_MPIAIJ(Mat A) 2024 { 2025 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 2026 PetscErrorCode ierr; 2027 2028 PetscFunctionBegin; 2029 ierr = MatSetUnfactored(a->A);CHKERRQ(ierr); 2030 PetscFunctionReturn(0); 2031 } 2032 2033 PetscErrorCode MatEqual_MPIAIJ(Mat A,Mat B,PetscBool *flag) 2034 { 2035 Mat_MPIAIJ *matB = (Mat_MPIAIJ*)B->data,*matA = (Mat_MPIAIJ*)A->data; 2036 Mat a,b,c,d; 2037 PetscBool flg; 2038 PetscErrorCode ierr; 2039 2040 PetscFunctionBegin; 2041 a = matA->A; b = matA->B; 2042 c = matB->A; d = matB->B; 2043 2044 ierr = MatEqual(a,c,&flg);CHKERRQ(ierr); 2045 if (flg) { 2046 ierr = MatEqual(b,d,&flg);CHKERRQ(ierr); 2047 } 2048 ierr = MPIU_Allreduce(&flg,flag,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)A));CHKERRQ(ierr); 2049 PetscFunctionReturn(0); 2050 } 2051 2052 PetscErrorCode MatCopy_MPIAIJ(Mat A,Mat B,MatStructure str) 2053 { 2054 PetscErrorCode ierr; 2055 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 2056 Mat_MPIAIJ *b = (Mat_MPIAIJ*)B->data; 2057 2058 PetscFunctionBegin; 2059 /* If the two matrices don't have the same copy implementation, they aren't compatible for fast copy. */ 2060 if ((str != SAME_NONZERO_PATTERN) || (A->ops->copy != B->ops->copy)) { 2061 /* because of the column compression in the off-processor part of the matrix a->B, 2062 the number of columns in a->B and b->B may be different, hence we cannot call 2063 the MatCopy() directly on the two parts. If need be, we can provide a more 2064 efficient copy than the MatCopy_Basic() by first uncompressing the a->B matrices 2065 then copying the submatrices */ 2066 ierr = MatCopy_Basic(A,B,str);CHKERRQ(ierr); 2067 } else { 2068 ierr = MatCopy(a->A,b->A,str);CHKERRQ(ierr); 2069 ierr = MatCopy(a->B,b->B,str);CHKERRQ(ierr); 2070 } 2071 ierr = PetscObjectStateIncrease((PetscObject)B);CHKERRQ(ierr); 2072 PetscFunctionReturn(0); 2073 } 2074 2075 PetscErrorCode MatSetUp_MPIAIJ(Mat A) 2076 { 2077 PetscErrorCode ierr; 2078 2079 PetscFunctionBegin; 2080 ierr = MatMPIAIJSetPreallocation(A,PETSC_DEFAULT,0,PETSC_DEFAULT,0);CHKERRQ(ierr); 2081 PetscFunctionReturn(0); 2082 } 2083 2084 /* 2085 Computes the number of nonzeros per row needed for preallocation when X and Y 2086 have different nonzero structure. 2087 */ 2088 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) 2089 { 2090 PetscInt i,j,k,nzx,nzy; 2091 2092 PetscFunctionBegin; 2093 /* Set the number of nonzeros in the new matrix */ 2094 for (i=0; i<m; i++) { 2095 const PetscInt *xjj = xj+xi[i],*yjj = yj+yi[i]; 2096 nzx = xi[i+1] - xi[i]; 2097 nzy = yi[i+1] - yi[i]; 2098 nnz[i] = 0; 2099 for (j=0,k=0; j<nzx; j++) { /* Point in X */ 2100 for (; k<nzy && yltog[yjj[k]]<xltog[xjj[j]]; k++) nnz[i]++; /* Catch up to X */ 2101 if (k<nzy && yltog[yjj[k]]==xltog[xjj[j]]) k++; /* Skip duplicate */ 2102 nnz[i]++; 2103 } 2104 for (; k<nzy; k++) nnz[i]++; 2105 } 2106 PetscFunctionReturn(0); 2107 } 2108 2109 /* This is the same as MatAXPYGetPreallocation_SeqAIJ, except that the local-to-global map is provided */ 2110 static PetscErrorCode MatAXPYGetPreallocation_MPIAIJ(Mat Y,const PetscInt *yltog,Mat X,const PetscInt *xltog,PetscInt *nnz) 2111 { 2112 PetscErrorCode ierr; 2113 PetscInt m = Y->rmap->N; 2114 Mat_SeqAIJ *x = (Mat_SeqAIJ*)X->data; 2115 Mat_SeqAIJ *y = (Mat_SeqAIJ*)Y->data; 2116 2117 PetscFunctionBegin; 2118 ierr = MatAXPYGetPreallocation_MPIX_private(m,x->i,x->j,xltog,y->i,y->j,yltog,nnz);CHKERRQ(ierr); 2119 PetscFunctionReturn(0); 2120 } 2121 2122 PetscErrorCode MatAXPY_MPIAIJ(Mat Y,PetscScalar a,Mat X,MatStructure str) 2123 { 2124 PetscErrorCode ierr; 2125 Mat_MPIAIJ *xx = (Mat_MPIAIJ*)X->data,*yy = (Mat_MPIAIJ*)Y->data; 2126 PetscBLASInt bnz,one=1; 2127 Mat_SeqAIJ *x,*y; 2128 2129 PetscFunctionBegin; 2130 if (str == SAME_NONZERO_PATTERN) { 2131 PetscScalar alpha = a; 2132 x = (Mat_SeqAIJ*)xx->A->data; 2133 ierr = PetscBLASIntCast(x->nz,&bnz);CHKERRQ(ierr); 2134 y = (Mat_SeqAIJ*)yy->A->data; 2135 PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&bnz,&alpha,x->a,&one,y->a,&one)); 2136 x = (Mat_SeqAIJ*)xx->B->data; 2137 y = (Mat_SeqAIJ*)yy->B->data; 2138 ierr = PetscBLASIntCast(x->nz,&bnz);CHKERRQ(ierr); 2139 PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&bnz,&alpha,x->a,&one,y->a,&one)); 2140 ierr = PetscObjectStateIncrease((PetscObject)Y);CHKERRQ(ierr); 2141 } else if (str == SUBSET_NONZERO_PATTERN) { /* nonzeros of X is a subset of Y's */ 2142 ierr = MatAXPY_Basic(Y,a,X,str);CHKERRQ(ierr); 2143 } else { 2144 Mat B; 2145 PetscInt *nnz_d,*nnz_o; 2146 ierr = PetscMalloc1(yy->A->rmap->N,&nnz_d);CHKERRQ(ierr); 2147 ierr = PetscMalloc1(yy->B->rmap->N,&nnz_o);CHKERRQ(ierr); 2148 ierr = MatCreate(PetscObjectComm((PetscObject)Y),&B);CHKERRQ(ierr); 2149 ierr = PetscObjectSetName((PetscObject)B,((PetscObject)Y)->name);CHKERRQ(ierr); 2150 ierr = MatSetSizes(B,Y->rmap->n,Y->cmap->n,Y->rmap->N,Y->cmap->N);CHKERRQ(ierr); 2151 ierr = MatSetBlockSizesFromMats(B,Y,Y);CHKERRQ(ierr); 2152 ierr = MatSetType(B,MATMPIAIJ);CHKERRQ(ierr); 2153 ierr = MatAXPYGetPreallocation_SeqAIJ(yy->A,xx->A,nnz_d);CHKERRQ(ierr); 2154 ierr = MatAXPYGetPreallocation_MPIAIJ(yy->B,yy->garray,xx->B,xx->garray,nnz_o);CHKERRQ(ierr); 2155 ierr = MatMPIAIJSetPreallocation(B,0,nnz_d,0,nnz_o);CHKERRQ(ierr); 2156 ierr = MatAXPY_BasicWithPreallocation(B,Y,a,X,str);CHKERRQ(ierr); 2157 ierr = MatHeaderReplace(Y,&B);CHKERRQ(ierr); 2158 ierr = PetscFree(nnz_d);CHKERRQ(ierr); 2159 ierr = PetscFree(nnz_o);CHKERRQ(ierr); 2160 } 2161 PetscFunctionReturn(0); 2162 } 2163 2164 extern PetscErrorCode MatConjugate_SeqAIJ(Mat); 2165 2166 PetscErrorCode MatConjugate_MPIAIJ(Mat mat) 2167 { 2168 #if defined(PETSC_USE_COMPLEX) 2169 PetscErrorCode ierr; 2170 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 2171 2172 PetscFunctionBegin; 2173 ierr = MatConjugate_SeqAIJ(aij->A);CHKERRQ(ierr); 2174 ierr = MatConjugate_SeqAIJ(aij->B);CHKERRQ(ierr); 2175 #else 2176 PetscFunctionBegin; 2177 #endif 2178 PetscFunctionReturn(0); 2179 } 2180 2181 PetscErrorCode MatRealPart_MPIAIJ(Mat A) 2182 { 2183 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 2184 PetscErrorCode ierr; 2185 2186 PetscFunctionBegin; 2187 ierr = MatRealPart(a->A);CHKERRQ(ierr); 2188 ierr = MatRealPart(a->B);CHKERRQ(ierr); 2189 PetscFunctionReturn(0); 2190 } 2191 2192 PetscErrorCode MatImaginaryPart_MPIAIJ(Mat A) 2193 { 2194 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 2195 PetscErrorCode ierr; 2196 2197 PetscFunctionBegin; 2198 ierr = MatImaginaryPart(a->A);CHKERRQ(ierr); 2199 ierr = MatImaginaryPart(a->B);CHKERRQ(ierr); 2200 PetscFunctionReturn(0); 2201 } 2202 2203 PetscErrorCode MatGetRowMaxAbs_MPIAIJ(Mat A, Vec v, PetscInt idx[]) 2204 { 2205 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 2206 PetscErrorCode ierr; 2207 PetscInt i,*idxb = 0; 2208 PetscScalar *va,*vb; 2209 Vec vtmp; 2210 2211 PetscFunctionBegin; 2212 ierr = MatGetRowMaxAbs(a->A,v,idx);CHKERRQ(ierr); 2213 ierr = VecGetArray(v,&va);CHKERRQ(ierr); 2214 if (idx) { 2215 for (i=0; i<A->rmap->n; i++) { 2216 if (PetscAbsScalar(va[i])) idx[i] += A->cmap->rstart; 2217 } 2218 } 2219 2220 ierr = VecCreateSeq(PETSC_COMM_SELF,A->rmap->n,&vtmp);CHKERRQ(ierr); 2221 if (idx) { 2222 ierr = PetscMalloc1(A->rmap->n,&idxb);CHKERRQ(ierr); 2223 } 2224 ierr = MatGetRowMaxAbs(a->B,vtmp,idxb);CHKERRQ(ierr); 2225 ierr = VecGetArray(vtmp,&vb);CHKERRQ(ierr); 2226 2227 for (i=0; i<A->rmap->n; i++) { 2228 if (PetscAbsScalar(va[i]) < PetscAbsScalar(vb[i])) { 2229 va[i] = vb[i]; 2230 if (idx) idx[i] = a->garray[idxb[i]]; 2231 } 2232 } 2233 2234 ierr = VecRestoreArray(v,&va);CHKERRQ(ierr); 2235 ierr = VecRestoreArray(vtmp,&vb);CHKERRQ(ierr); 2236 ierr = PetscFree(idxb);CHKERRQ(ierr); 2237 ierr = VecDestroy(&vtmp);CHKERRQ(ierr); 2238 PetscFunctionReturn(0); 2239 } 2240 2241 PetscErrorCode MatGetRowMinAbs_MPIAIJ(Mat A, Vec v, PetscInt idx[]) 2242 { 2243 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 2244 PetscErrorCode ierr; 2245 PetscInt i,*idxb = 0; 2246 PetscScalar *va,*vb; 2247 Vec vtmp; 2248 2249 PetscFunctionBegin; 2250 ierr = MatGetRowMinAbs(a->A,v,idx);CHKERRQ(ierr); 2251 ierr = VecGetArray(v,&va);CHKERRQ(ierr); 2252 if (idx) { 2253 for (i=0; i<A->cmap->n; i++) { 2254 if (PetscAbsScalar(va[i])) idx[i] += A->cmap->rstart; 2255 } 2256 } 2257 2258 ierr = VecCreateSeq(PETSC_COMM_SELF,A->rmap->n,&vtmp);CHKERRQ(ierr); 2259 if (idx) { 2260 ierr = PetscMalloc1(A->rmap->n,&idxb);CHKERRQ(ierr); 2261 } 2262 ierr = MatGetRowMinAbs(a->B,vtmp,idxb);CHKERRQ(ierr); 2263 ierr = VecGetArray(vtmp,&vb);CHKERRQ(ierr); 2264 2265 for (i=0; i<A->rmap->n; i++) { 2266 if (PetscAbsScalar(va[i]) > PetscAbsScalar(vb[i])) { 2267 va[i] = vb[i]; 2268 if (idx) idx[i] = a->garray[idxb[i]]; 2269 } 2270 } 2271 2272 ierr = VecRestoreArray(v,&va);CHKERRQ(ierr); 2273 ierr = VecRestoreArray(vtmp,&vb);CHKERRQ(ierr); 2274 ierr = PetscFree(idxb);CHKERRQ(ierr); 2275 ierr = VecDestroy(&vtmp);CHKERRQ(ierr); 2276 PetscFunctionReturn(0); 2277 } 2278 2279 PetscErrorCode MatGetRowMin_MPIAIJ(Mat A, Vec v, PetscInt idx[]) 2280 { 2281 Mat_MPIAIJ *mat = (Mat_MPIAIJ*) A->data; 2282 PetscInt n = A->rmap->n; 2283 PetscInt cstart = A->cmap->rstart; 2284 PetscInt *cmap = mat->garray; 2285 PetscInt *diagIdx, *offdiagIdx; 2286 Vec diagV, offdiagV; 2287 PetscScalar *a, *diagA, *offdiagA; 2288 PetscInt r; 2289 PetscErrorCode ierr; 2290 2291 PetscFunctionBegin; 2292 ierr = PetscMalloc2(n,&diagIdx,n,&offdiagIdx);CHKERRQ(ierr); 2293 ierr = VecCreateSeq(PetscObjectComm((PetscObject)A), n, &diagV);CHKERRQ(ierr); 2294 ierr = VecCreateSeq(PetscObjectComm((PetscObject)A), n, &offdiagV);CHKERRQ(ierr); 2295 ierr = MatGetRowMin(mat->A, diagV, diagIdx);CHKERRQ(ierr); 2296 ierr = MatGetRowMin(mat->B, offdiagV, offdiagIdx);CHKERRQ(ierr); 2297 ierr = VecGetArray(v, &a);CHKERRQ(ierr); 2298 ierr = VecGetArray(diagV, &diagA);CHKERRQ(ierr); 2299 ierr = VecGetArray(offdiagV, &offdiagA);CHKERRQ(ierr); 2300 for (r = 0; r < n; ++r) { 2301 if (PetscAbsScalar(diagA[r]) <= PetscAbsScalar(offdiagA[r])) { 2302 a[r] = diagA[r]; 2303 idx[r] = cstart + diagIdx[r]; 2304 } else { 2305 a[r] = offdiagA[r]; 2306 idx[r] = cmap[offdiagIdx[r]]; 2307 } 2308 } 2309 ierr = VecRestoreArray(v, &a);CHKERRQ(ierr); 2310 ierr = VecRestoreArray(diagV, &diagA);CHKERRQ(ierr); 2311 ierr = VecRestoreArray(offdiagV, &offdiagA);CHKERRQ(ierr); 2312 ierr = VecDestroy(&diagV);CHKERRQ(ierr); 2313 ierr = VecDestroy(&offdiagV);CHKERRQ(ierr); 2314 ierr = PetscFree2(diagIdx, offdiagIdx);CHKERRQ(ierr); 2315 PetscFunctionReturn(0); 2316 } 2317 2318 PetscErrorCode MatGetRowMax_MPIAIJ(Mat A, Vec v, PetscInt idx[]) 2319 { 2320 Mat_MPIAIJ *mat = (Mat_MPIAIJ*) A->data; 2321 PetscInt n = A->rmap->n; 2322 PetscInt cstart = A->cmap->rstart; 2323 PetscInt *cmap = mat->garray; 2324 PetscInt *diagIdx, *offdiagIdx; 2325 Vec diagV, offdiagV; 2326 PetscScalar *a, *diagA, *offdiagA; 2327 PetscInt r; 2328 PetscErrorCode ierr; 2329 2330 PetscFunctionBegin; 2331 ierr = PetscMalloc2(n,&diagIdx,n,&offdiagIdx);CHKERRQ(ierr); 2332 ierr = VecCreateSeq(PETSC_COMM_SELF, n, &diagV);CHKERRQ(ierr); 2333 ierr = VecCreateSeq(PETSC_COMM_SELF, n, &offdiagV);CHKERRQ(ierr); 2334 ierr = MatGetRowMax(mat->A, diagV, diagIdx);CHKERRQ(ierr); 2335 ierr = MatGetRowMax(mat->B, offdiagV, offdiagIdx);CHKERRQ(ierr); 2336 ierr = VecGetArray(v, &a);CHKERRQ(ierr); 2337 ierr = VecGetArray(diagV, &diagA);CHKERRQ(ierr); 2338 ierr = VecGetArray(offdiagV, &offdiagA);CHKERRQ(ierr); 2339 for (r = 0; r < n; ++r) { 2340 if (PetscAbsScalar(diagA[r]) >= PetscAbsScalar(offdiagA[r])) { 2341 a[r] = diagA[r]; 2342 idx[r] = cstart + diagIdx[r]; 2343 } else { 2344 a[r] = offdiagA[r]; 2345 idx[r] = cmap[offdiagIdx[r]]; 2346 } 2347 } 2348 ierr = VecRestoreArray(v, &a);CHKERRQ(ierr); 2349 ierr = VecRestoreArray(diagV, &diagA);CHKERRQ(ierr); 2350 ierr = VecRestoreArray(offdiagV, &offdiagA);CHKERRQ(ierr); 2351 ierr = VecDestroy(&diagV);CHKERRQ(ierr); 2352 ierr = VecDestroy(&offdiagV);CHKERRQ(ierr); 2353 ierr = PetscFree2(diagIdx, offdiagIdx);CHKERRQ(ierr); 2354 PetscFunctionReturn(0); 2355 } 2356 2357 PetscErrorCode MatGetSeqNonzeroStructure_MPIAIJ(Mat mat,Mat *newmat) 2358 { 2359 PetscErrorCode ierr; 2360 Mat *dummy; 2361 2362 PetscFunctionBegin; 2363 ierr = MatCreateSubMatrix_MPIAIJ_All(mat,MAT_DO_NOT_GET_VALUES,MAT_INITIAL_MATRIX,&dummy);CHKERRQ(ierr); 2364 *newmat = *dummy; 2365 ierr = PetscFree(dummy);CHKERRQ(ierr); 2366 PetscFunctionReturn(0); 2367 } 2368 2369 PetscErrorCode MatInvertBlockDiagonal_MPIAIJ(Mat A,const PetscScalar **values) 2370 { 2371 Mat_MPIAIJ *a = (Mat_MPIAIJ*) A->data; 2372 PetscErrorCode ierr; 2373 2374 PetscFunctionBegin; 2375 ierr = MatInvertBlockDiagonal(a->A,values);CHKERRQ(ierr); 2376 A->factorerrortype = a->A->factorerrortype; 2377 PetscFunctionReturn(0); 2378 } 2379 2380 static PetscErrorCode MatSetRandom_MPIAIJ(Mat x,PetscRandom rctx) 2381 { 2382 PetscErrorCode ierr; 2383 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)x->data; 2384 2385 PetscFunctionBegin; 2386 ierr = MatSetRandom(aij->A,rctx);CHKERRQ(ierr); 2387 ierr = MatSetRandom(aij->B,rctx);CHKERRQ(ierr); 2388 ierr = MatAssemblyBegin(x,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2389 ierr = MatAssemblyEnd(x,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2390 PetscFunctionReturn(0); 2391 } 2392 2393 PetscErrorCode MatMPIAIJSetUseScalableIncreaseOverlap_MPIAIJ(Mat A,PetscBool sc) 2394 { 2395 PetscFunctionBegin; 2396 if (sc) A->ops->increaseoverlap = MatIncreaseOverlap_MPIAIJ_Scalable; 2397 else A->ops->increaseoverlap = MatIncreaseOverlap_MPIAIJ; 2398 PetscFunctionReturn(0); 2399 } 2400 2401 /*@ 2402 MatMPIAIJSetUseScalableIncreaseOverlap - Determine if the matrix uses a scalable algorithm to compute the overlap 2403 2404 Collective on Mat 2405 2406 Input Parameters: 2407 + A - the matrix 2408 - sc - PETSC_TRUE indicates use the scalable algorithm (default is not to use the scalable algorithm) 2409 2410 Level: advanced 2411 2412 @*/ 2413 PetscErrorCode MatMPIAIJSetUseScalableIncreaseOverlap(Mat A,PetscBool sc) 2414 { 2415 PetscErrorCode ierr; 2416 2417 PetscFunctionBegin; 2418 ierr = PetscTryMethod(A,"MatMPIAIJSetUseScalableIncreaseOverlap_C",(Mat,PetscBool),(A,sc));CHKERRQ(ierr); 2419 PetscFunctionReturn(0); 2420 } 2421 2422 PetscErrorCode MatSetFromOptions_MPIAIJ(PetscOptionItems *PetscOptionsObject,Mat A) 2423 { 2424 PetscErrorCode ierr; 2425 PetscBool sc = PETSC_FALSE,flg; 2426 2427 PetscFunctionBegin; 2428 ierr = PetscOptionsHead(PetscOptionsObject,"MPIAIJ options");CHKERRQ(ierr); 2429 ierr = PetscObjectOptionsBegin((PetscObject)A); 2430 if (A->ops->increaseoverlap == MatIncreaseOverlap_MPIAIJ_Scalable) sc = PETSC_TRUE; 2431 ierr = PetscOptionsBool("-mat_increase_overlap_scalable","Use a scalable algorithm to compute the overlap","MatIncreaseOverlap",sc,&sc,&flg);CHKERRQ(ierr); 2432 if (flg) { 2433 ierr = MatMPIAIJSetUseScalableIncreaseOverlap(A,sc);CHKERRQ(ierr); 2434 } 2435 ierr = PetscOptionsEnd();CHKERRQ(ierr); 2436 PetscFunctionReturn(0); 2437 } 2438 2439 PetscErrorCode MatShift_MPIAIJ(Mat Y,PetscScalar a) 2440 { 2441 PetscErrorCode ierr; 2442 Mat_MPIAIJ *maij = (Mat_MPIAIJ*)Y->data; 2443 Mat_SeqAIJ *aij = (Mat_SeqAIJ*)maij->A->data; 2444 2445 PetscFunctionBegin; 2446 if (!Y->preallocated) { 2447 ierr = MatMPIAIJSetPreallocation(Y,1,NULL,0,NULL);CHKERRQ(ierr); 2448 } else if (!aij->nz) { 2449 PetscInt nonew = aij->nonew; 2450 ierr = MatSeqAIJSetPreallocation(maij->A,1,NULL);CHKERRQ(ierr); 2451 aij->nonew = nonew; 2452 } 2453 ierr = MatShift_Basic(Y,a);CHKERRQ(ierr); 2454 PetscFunctionReturn(0); 2455 } 2456 2457 PetscErrorCode MatMissingDiagonal_MPIAIJ(Mat A,PetscBool *missing,PetscInt *d) 2458 { 2459 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 2460 PetscErrorCode ierr; 2461 2462 PetscFunctionBegin; 2463 if (A->rmap->n != A->cmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Only works for square matrices"); 2464 ierr = MatMissingDiagonal(a->A,missing,d);CHKERRQ(ierr); 2465 if (d) { 2466 PetscInt rstart; 2467 ierr = MatGetOwnershipRange(A,&rstart,NULL);CHKERRQ(ierr); 2468 *d += rstart; 2469 2470 } 2471 PetscFunctionReturn(0); 2472 } 2473 2474 2475 /* -------------------------------------------------------------------*/ 2476 static struct _MatOps MatOps_Values = {MatSetValues_MPIAIJ, 2477 MatGetRow_MPIAIJ, 2478 MatRestoreRow_MPIAIJ, 2479 MatMult_MPIAIJ, 2480 /* 4*/ MatMultAdd_MPIAIJ, 2481 MatMultTranspose_MPIAIJ, 2482 MatMultTransposeAdd_MPIAIJ, 2483 0, 2484 0, 2485 0, 2486 /*10*/ 0, 2487 0, 2488 0, 2489 MatSOR_MPIAIJ, 2490 MatTranspose_MPIAIJ, 2491 /*15*/ MatGetInfo_MPIAIJ, 2492 MatEqual_MPIAIJ, 2493 MatGetDiagonal_MPIAIJ, 2494 MatDiagonalScale_MPIAIJ, 2495 MatNorm_MPIAIJ, 2496 /*20*/ MatAssemblyBegin_MPIAIJ, 2497 MatAssemblyEnd_MPIAIJ, 2498 MatSetOption_MPIAIJ, 2499 MatZeroEntries_MPIAIJ, 2500 /*24*/ MatZeroRows_MPIAIJ, 2501 0, 2502 0, 2503 0, 2504 0, 2505 /*29*/ MatSetUp_MPIAIJ, 2506 0, 2507 0, 2508 MatGetDiagonalBlock_MPIAIJ, 2509 0, 2510 /*34*/ MatDuplicate_MPIAIJ, 2511 0, 2512 0, 2513 0, 2514 0, 2515 /*39*/ MatAXPY_MPIAIJ, 2516 MatCreateSubMatrices_MPIAIJ, 2517 MatIncreaseOverlap_MPIAIJ, 2518 MatGetValues_MPIAIJ, 2519 MatCopy_MPIAIJ, 2520 /*44*/ MatGetRowMax_MPIAIJ, 2521 MatScale_MPIAIJ, 2522 MatShift_MPIAIJ, 2523 MatDiagonalSet_MPIAIJ, 2524 MatZeroRowsColumns_MPIAIJ, 2525 /*49*/ MatSetRandom_MPIAIJ, 2526 0, 2527 0, 2528 0, 2529 0, 2530 /*54*/ MatFDColoringCreate_MPIXAIJ, 2531 0, 2532 MatSetUnfactored_MPIAIJ, 2533 MatPermute_MPIAIJ, 2534 0, 2535 /*59*/ MatCreateSubMatrix_MPIAIJ, 2536 MatDestroy_MPIAIJ, 2537 MatView_MPIAIJ, 2538 0, 2539 MatMatMatMult_MPIAIJ_MPIAIJ_MPIAIJ, 2540 /*64*/ MatMatMatMultSymbolic_MPIAIJ_MPIAIJ_MPIAIJ, 2541 MatMatMatMultNumeric_MPIAIJ_MPIAIJ_MPIAIJ, 2542 0, 2543 0, 2544 0, 2545 /*69*/ MatGetRowMaxAbs_MPIAIJ, 2546 MatGetRowMinAbs_MPIAIJ, 2547 0, 2548 0, 2549 0, 2550 0, 2551 /*75*/ MatFDColoringApply_AIJ, 2552 MatSetFromOptions_MPIAIJ, 2553 0, 2554 0, 2555 MatFindZeroDiagonals_MPIAIJ, 2556 /*80*/ 0, 2557 0, 2558 0, 2559 /*83*/ MatLoad_MPIAIJ, 2560 MatIsSymmetric_MPIAIJ, 2561 0, 2562 0, 2563 0, 2564 0, 2565 /*89*/ MatMatMult_MPIAIJ_MPIAIJ, 2566 MatMatMultSymbolic_MPIAIJ_MPIAIJ, 2567 MatMatMultNumeric_MPIAIJ_MPIAIJ, 2568 MatPtAP_MPIAIJ_MPIAIJ, 2569 MatPtAPSymbolic_MPIAIJ_MPIAIJ, 2570 /*94*/ MatPtAPNumeric_MPIAIJ_MPIAIJ, 2571 0, 2572 0, 2573 0, 2574 0, 2575 /*99*/ 0, 2576 0, 2577 0, 2578 MatConjugate_MPIAIJ, 2579 0, 2580 /*104*/MatSetValuesRow_MPIAIJ, 2581 MatRealPart_MPIAIJ, 2582 MatImaginaryPart_MPIAIJ, 2583 0, 2584 0, 2585 /*109*/0, 2586 0, 2587 MatGetRowMin_MPIAIJ, 2588 0, 2589 MatMissingDiagonal_MPIAIJ, 2590 /*114*/MatGetSeqNonzeroStructure_MPIAIJ, 2591 0, 2592 MatGetGhosts_MPIAIJ, 2593 0, 2594 0, 2595 /*119*/0, 2596 0, 2597 0, 2598 0, 2599 MatGetMultiProcBlock_MPIAIJ, 2600 /*124*/MatFindNonzeroRows_MPIAIJ, 2601 MatGetColumnNorms_MPIAIJ, 2602 MatInvertBlockDiagonal_MPIAIJ, 2603 0, 2604 MatCreateSubMatricesMPI_MPIAIJ, 2605 /*129*/0, 2606 MatTransposeMatMult_MPIAIJ_MPIAIJ, 2607 MatTransposeMatMultSymbolic_MPIAIJ_MPIAIJ, 2608 MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ, 2609 0, 2610 /*134*/0, 2611 0, 2612 MatRARt_MPIAIJ_MPIAIJ, 2613 0, 2614 0, 2615 /*139*/MatSetBlockSizes_MPIAIJ, 2616 0, 2617 0, 2618 MatFDColoringSetUp_MPIXAIJ, 2619 MatFindOffBlockDiagonalEntries_MPIAIJ, 2620 /*144*/MatCreateMPIMatConcatenateSeqMat_MPIAIJ 2621 }; 2622 2623 /* ----------------------------------------------------------------------------------------*/ 2624 2625 PetscErrorCode MatStoreValues_MPIAIJ(Mat mat) 2626 { 2627 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 2628 PetscErrorCode ierr; 2629 2630 PetscFunctionBegin; 2631 ierr = MatStoreValues(aij->A);CHKERRQ(ierr); 2632 ierr = MatStoreValues(aij->B);CHKERRQ(ierr); 2633 PetscFunctionReturn(0); 2634 } 2635 2636 PetscErrorCode MatRetrieveValues_MPIAIJ(Mat mat) 2637 { 2638 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 2639 PetscErrorCode ierr; 2640 2641 PetscFunctionBegin; 2642 ierr = MatRetrieveValues(aij->A);CHKERRQ(ierr); 2643 ierr = MatRetrieveValues(aij->B);CHKERRQ(ierr); 2644 PetscFunctionReturn(0); 2645 } 2646 2647 PetscErrorCode MatMPIAIJSetPreallocation_MPIAIJ(Mat B,PetscInt d_nz,const PetscInt d_nnz[],PetscInt o_nz,const PetscInt o_nnz[]) 2648 { 2649 Mat_MPIAIJ *b; 2650 PetscErrorCode ierr; 2651 2652 PetscFunctionBegin; 2653 ierr = PetscLayoutSetUp(B->rmap);CHKERRQ(ierr); 2654 ierr = PetscLayoutSetUp(B->cmap);CHKERRQ(ierr); 2655 b = (Mat_MPIAIJ*)B->data; 2656 2657 #if defined(PETSC_USE_CTABLE) 2658 ierr = PetscTableDestroy(&b->colmap);CHKERRQ(ierr); 2659 #else 2660 ierr = PetscFree(b->colmap);CHKERRQ(ierr); 2661 #endif 2662 ierr = PetscFree(b->garray);CHKERRQ(ierr); 2663 ierr = VecDestroy(&b->lvec);CHKERRQ(ierr); 2664 ierr = VecScatterDestroy(&b->Mvctx);CHKERRQ(ierr); 2665 2666 /* Because the B will have been resized we simply destroy it and create a new one each time */ 2667 ierr = MatDestroy(&b->B);CHKERRQ(ierr); 2668 ierr = MatCreate(PETSC_COMM_SELF,&b->B);CHKERRQ(ierr); 2669 ierr = MatSetSizes(b->B,B->rmap->n,B->cmap->N,B->rmap->n,B->cmap->N);CHKERRQ(ierr); 2670 ierr = MatSetBlockSizesFromMats(b->B,B,B);CHKERRQ(ierr); 2671 ierr = MatSetType(b->B,MATSEQAIJ);CHKERRQ(ierr); 2672 ierr = PetscLogObjectParent((PetscObject)B,(PetscObject)b->B);CHKERRQ(ierr); 2673 2674 if (!B->preallocated) { 2675 ierr = MatCreate(PETSC_COMM_SELF,&b->A);CHKERRQ(ierr); 2676 ierr = MatSetSizes(b->A,B->rmap->n,B->cmap->n,B->rmap->n,B->cmap->n);CHKERRQ(ierr); 2677 ierr = MatSetBlockSizesFromMats(b->A,B,B);CHKERRQ(ierr); 2678 ierr = MatSetType(b->A,MATSEQAIJ);CHKERRQ(ierr); 2679 ierr = PetscLogObjectParent((PetscObject)B,(PetscObject)b->A);CHKERRQ(ierr); 2680 } 2681 2682 ierr = MatSeqAIJSetPreallocation(b->A,d_nz,d_nnz);CHKERRQ(ierr); 2683 ierr = MatSeqAIJSetPreallocation(b->B,o_nz,o_nnz);CHKERRQ(ierr); 2684 B->preallocated = PETSC_TRUE; 2685 B->was_assembled = PETSC_FALSE; 2686 B->assembled = PETSC_FALSE;; 2687 PetscFunctionReturn(0); 2688 } 2689 2690 PetscErrorCode MatResetPreallocation_MPIAIJ(Mat B) 2691 { 2692 Mat_MPIAIJ *b; 2693 PetscErrorCode ierr; 2694 2695 PetscFunctionBegin; 2696 PetscValidHeaderSpecific(B,MAT_CLASSID,1); 2697 ierr = PetscLayoutSetUp(B->rmap);CHKERRQ(ierr); 2698 ierr = PetscLayoutSetUp(B->cmap);CHKERRQ(ierr); 2699 b = (Mat_MPIAIJ*)B->data; 2700 2701 #if defined(PETSC_USE_CTABLE) 2702 ierr = PetscTableDestroy(&b->colmap);CHKERRQ(ierr); 2703 #else 2704 ierr = PetscFree(b->colmap);CHKERRQ(ierr); 2705 #endif 2706 ierr = PetscFree(b->garray);CHKERRQ(ierr); 2707 ierr = VecDestroy(&b->lvec);CHKERRQ(ierr); 2708 ierr = VecScatterDestroy(&b->Mvctx);CHKERRQ(ierr); 2709 2710 ierr = MatResetPreallocation(b->A);CHKERRQ(ierr); 2711 ierr = MatResetPreallocation(b->B);CHKERRQ(ierr); 2712 B->preallocated = PETSC_TRUE; 2713 B->was_assembled = PETSC_FALSE; 2714 B->assembled = PETSC_FALSE; 2715 PetscFunctionReturn(0); 2716 } 2717 2718 PetscErrorCode MatDuplicate_MPIAIJ(Mat matin,MatDuplicateOption cpvalues,Mat *newmat) 2719 { 2720 Mat mat; 2721 Mat_MPIAIJ *a,*oldmat = (Mat_MPIAIJ*)matin->data; 2722 PetscErrorCode ierr; 2723 2724 PetscFunctionBegin; 2725 *newmat = 0; 2726 ierr = MatCreate(PetscObjectComm((PetscObject)matin),&mat);CHKERRQ(ierr); 2727 ierr = MatSetSizes(mat,matin->rmap->n,matin->cmap->n,matin->rmap->N,matin->cmap->N);CHKERRQ(ierr); 2728 ierr = MatSetBlockSizesFromMats(mat,matin,matin);CHKERRQ(ierr); 2729 ierr = MatSetType(mat,((PetscObject)matin)->type_name);CHKERRQ(ierr); 2730 ierr = PetscMemcpy(mat->ops,matin->ops,sizeof(struct _MatOps));CHKERRQ(ierr); 2731 a = (Mat_MPIAIJ*)mat->data; 2732 2733 mat->factortype = matin->factortype; 2734 mat->assembled = PETSC_TRUE; 2735 mat->insertmode = NOT_SET_VALUES; 2736 mat->preallocated = PETSC_TRUE; 2737 2738 a->size = oldmat->size; 2739 a->rank = oldmat->rank; 2740 a->donotstash = oldmat->donotstash; 2741 a->roworiented = oldmat->roworiented; 2742 a->rowindices = 0; 2743 a->rowvalues = 0; 2744 a->getrowactive = PETSC_FALSE; 2745 2746 ierr = PetscLayoutReference(matin->rmap,&mat->rmap);CHKERRQ(ierr); 2747 ierr = PetscLayoutReference(matin->cmap,&mat->cmap);CHKERRQ(ierr); 2748 2749 if (oldmat->colmap) { 2750 #if defined(PETSC_USE_CTABLE) 2751 ierr = PetscTableCreateCopy(oldmat->colmap,&a->colmap);CHKERRQ(ierr); 2752 #else 2753 ierr = PetscMalloc1(mat->cmap->N,&a->colmap);CHKERRQ(ierr); 2754 ierr = PetscLogObjectMemory((PetscObject)mat,(mat->cmap->N)*sizeof(PetscInt));CHKERRQ(ierr); 2755 ierr = PetscMemcpy(a->colmap,oldmat->colmap,(mat->cmap->N)*sizeof(PetscInt));CHKERRQ(ierr); 2756 #endif 2757 } else a->colmap = 0; 2758 if (oldmat->garray) { 2759 PetscInt len; 2760 len = oldmat->B->cmap->n; 2761 ierr = PetscMalloc1(len+1,&a->garray);CHKERRQ(ierr); 2762 ierr = PetscLogObjectMemory((PetscObject)mat,len*sizeof(PetscInt));CHKERRQ(ierr); 2763 if (len) { ierr = PetscMemcpy(a->garray,oldmat->garray,len*sizeof(PetscInt));CHKERRQ(ierr); } 2764 } else a->garray = 0; 2765 2766 ierr = VecDuplicate(oldmat->lvec,&a->lvec);CHKERRQ(ierr); 2767 ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)a->lvec);CHKERRQ(ierr); 2768 ierr = VecScatterCopy(oldmat->Mvctx,&a->Mvctx);CHKERRQ(ierr); 2769 ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)a->Mvctx);CHKERRQ(ierr); 2770 ierr = MatDuplicate(oldmat->A,cpvalues,&a->A);CHKERRQ(ierr); 2771 ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)a->A);CHKERRQ(ierr); 2772 ierr = MatDuplicate(oldmat->B,cpvalues,&a->B);CHKERRQ(ierr); 2773 ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)a->B);CHKERRQ(ierr); 2774 ierr = PetscFunctionListDuplicate(((PetscObject)matin)->qlist,&((PetscObject)mat)->qlist);CHKERRQ(ierr); 2775 *newmat = mat; 2776 PetscFunctionReturn(0); 2777 } 2778 2779 PetscErrorCode MatLoad_MPIAIJ(Mat newMat, PetscViewer viewer) 2780 { 2781 PetscScalar *vals,*svals; 2782 MPI_Comm comm; 2783 PetscErrorCode ierr; 2784 PetscMPIInt rank,size,tag = ((PetscObject)viewer)->tag; 2785 PetscInt i,nz,j,rstart,rend,mmax,maxnz = 0; 2786 PetscInt header[4],*rowlengths = 0,M,N,m,*cols; 2787 PetscInt *ourlens = NULL,*procsnz = NULL,*offlens = NULL,jj,*mycols,*smycols; 2788 PetscInt cend,cstart,n,*rowners; 2789 int fd; 2790 PetscInt bs = newMat->rmap->bs; 2791 2792 PetscFunctionBegin; 2793 /* force binary viewer to load .info file if it has not yet done so */ 2794 ierr = PetscViewerSetUp(viewer);CHKERRQ(ierr); 2795 ierr = PetscObjectGetComm((PetscObject)viewer,&comm);CHKERRQ(ierr); 2796 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 2797 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 2798 ierr = PetscViewerBinaryGetDescriptor(viewer,&fd);CHKERRQ(ierr); 2799 if (!rank) { 2800 ierr = PetscBinaryRead(fd,(char*)header,4,PETSC_INT);CHKERRQ(ierr); 2801 if (header[0] != MAT_FILE_CLASSID) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED,"not matrix object"); 2802 if (header[3] < 0) SETERRQ(PetscObjectComm((PetscObject)newMat),PETSC_ERR_FILE_UNEXPECTED,"Matrix stored in special format on disk,cannot load as MATMPIAIJ"); 2803 } 2804 2805 ierr = PetscOptionsBegin(comm,NULL,"Options for loading MATMPIAIJ matrix","Mat");CHKERRQ(ierr); 2806 ierr = PetscOptionsInt("-matload_block_size","Set the blocksize used to store the matrix","MatLoad",bs,&bs,NULL);CHKERRQ(ierr); 2807 ierr = PetscOptionsEnd();CHKERRQ(ierr); 2808 if (bs < 0) bs = 1; 2809 2810 ierr = MPI_Bcast(header+1,3,MPIU_INT,0,comm);CHKERRQ(ierr); 2811 M = header[1]; N = header[2]; 2812 2813 /* If global sizes are set, check if they are consistent with that given in the file */ 2814 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); 2815 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); 2816 2817 /* determine ownership of all (block) rows */ 2818 if (M%bs) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED, "Inconsistent # of rows (%d) and block size (%d)",M,bs); 2819 if (newMat->rmap->n < 0) m = bs*((M/bs)/size + (((M/bs) % size) > rank)); /* PETSC_DECIDE */ 2820 else m = newMat->rmap->n; /* Set by user */ 2821 2822 ierr = PetscMalloc1(size+1,&rowners);CHKERRQ(ierr); 2823 ierr = MPI_Allgather(&m,1,MPIU_INT,rowners+1,1,MPIU_INT,comm);CHKERRQ(ierr); 2824 2825 /* First process needs enough room for process with most rows */ 2826 if (!rank) { 2827 mmax = rowners[1]; 2828 for (i=2; i<=size; i++) { 2829 mmax = PetscMax(mmax, rowners[i]); 2830 } 2831 } else mmax = -1; /* unused, but compilers complain */ 2832 2833 rowners[0] = 0; 2834 for (i=2; i<=size; i++) { 2835 rowners[i] += rowners[i-1]; 2836 } 2837 rstart = rowners[rank]; 2838 rend = rowners[rank+1]; 2839 2840 /* distribute row lengths to all processors */ 2841 ierr = PetscMalloc2(m,&ourlens,m,&offlens);CHKERRQ(ierr); 2842 if (!rank) { 2843 ierr = PetscBinaryRead(fd,ourlens,m,PETSC_INT);CHKERRQ(ierr); 2844 ierr = PetscMalloc1(mmax,&rowlengths);CHKERRQ(ierr); 2845 ierr = PetscCalloc1(size,&procsnz);CHKERRQ(ierr); 2846 for (j=0; j<m; j++) { 2847 procsnz[0] += ourlens[j]; 2848 } 2849 for (i=1; i<size; i++) { 2850 ierr = PetscBinaryRead(fd,rowlengths,rowners[i+1]-rowners[i],PETSC_INT);CHKERRQ(ierr); 2851 /* calculate the number of nonzeros on each processor */ 2852 for (j=0; j<rowners[i+1]-rowners[i]; j++) { 2853 procsnz[i] += rowlengths[j]; 2854 } 2855 ierr = MPIULong_Send(rowlengths,rowners[i+1]-rowners[i],MPIU_INT,i,tag,comm);CHKERRQ(ierr); 2856 } 2857 ierr = PetscFree(rowlengths);CHKERRQ(ierr); 2858 } else { 2859 ierr = MPIULong_Recv(ourlens,m,MPIU_INT,0,tag,comm);CHKERRQ(ierr); 2860 } 2861 2862 if (!rank) { 2863 /* determine max buffer needed and allocate it */ 2864 maxnz = 0; 2865 for (i=0; i<size; i++) { 2866 maxnz = PetscMax(maxnz,procsnz[i]); 2867 } 2868 ierr = PetscMalloc1(maxnz,&cols);CHKERRQ(ierr); 2869 2870 /* read in my part of the matrix column indices */ 2871 nz = procsnz[0]; 2872 ierr = PetscMalloc1(nz,&mycols);CHKERRQ(ierr); 2873 ierr = PetscBinaryRead(fd,mycols,nz,PETSC_INT);CHKERRQ(ierr); 2874 2875 /* read in every one elses and ship off */ 2876 for (i=1; i<size; i++) { 2877 nz = procsnz[i]; 2878 ierr = PetscBinaryRead(fd,cols,nz,PETSC_INT);CHKERRQ(ierr); 2879 ierr = MPIULong_Send(cols,nz,MPIU_INT,i,tag,comm);CHKERRQ(ierr); 2880 } 2881 ierr = PetscFree(cols);CHKERRQ(ierr); 2882 } else { 2883 /* determine buffer space needed for message */ 2884 nz = 0; 2885 for (i=0; i<m; i++) { 2886 nz += ourlens[i]; 2887 } 2888 ierr = PetscMalloc1(nz,&mycols);CHKERRQ(ierr); 2889 2890 /* receive message of column indices*/ 2891 ierr = MPIULong_Recv(mycols,nz,MPIU_INT,0,tag,comm);CHKERRQ(ierr); 2892 } 2893 2894 /* determine column ownership if matrix is not square */ 2895 if (N != M) { 2896 if (newMat->cmap->n < 0) n = N/size + ((N % size) > rank); 2897 else n = newMat->cmap->n; 2898 ierr = MPI_Scan(&n,&cend,1,MPIU_INT,MPI_SUM,comm);CHKERRQ(ierr); 2899 cstart = cend - n; 2900 } else { 2901 cstart = rstart; 2902 cend = rend; 2903 n = cend - cstart; 2904 } 2905 2906 /* loop over local rows, determining number of off diagonal entries */ 2907 ierr = PetscMemzero(offlens,m*sizeof(PetscInt));CHKERRQ(ierr); 2908 jj = 0; 2909 for (i=0; i<m; i++) { 2910 for (j=0; j<ourlens[i]; j++) { 2911 if (mycols[jj] < cstart || mycols[jj] >= cend) offlens[i]++; 2912 jj++; 2913 } 2914 } 2915 2916 for (i=0; i<m; i++) { 2917 ourlens[i] -= offlens[i]; 2918 } 2919 ierr = MatSetSizes(newMat,m,n,M,N);CHKERRQ(ierr); 2920 2921 if (bs > 1) {ierr = MatSetBlockSize(newMat,bs);CHKERRQ(ierr);} 2922 2923 ierr = MatMPIAIJSetPreallocation(newMat,0,ourlens,0,offlens);CHKERRQ(ierr); 2924 2925 for (i=0; i<m; i++) { 2926 ourlens[i] += offlens[i]; 2927 } 2928 2929 if (!rank) { 2930 ierr = PetscMalloc1(maxnz+1,&vals);CHKERRQ(ierr); 2931 2932 /* read in my part of the matrix numerical values */ 2933 nz = procsnz[0]; 2934 ierr = PetscBinaryRead(fd,vals,nz,PETSC_SCALAR);CHKERRQ(ierr); 2935 2936 /* insert into matrix */ 2937 jj = rstart; 2938 smycols = mycols; 2939 svals = vals; 2940 for (i=0; i<m; i++) { 2941 ierr = MatSetValues_MPIAIJ(newMat,1,&jj,ourlens[i],smycols,svals,INSERT_VALUES);CHKERRQ(ierr); 2942 smycols += ourlens[i]; 2943 svals += ourlens[i]; 2944 jj++; 2945 } 2946 2947 /* read in other processors and ship out */ 2948 for (i=1; i<size; i++) { 2949 nz = procsnz[i]; 2950 ierr = PetscBinaryRead(fd,vals,nz,PETSC_SCALAR);CHKERRQ(ierr); 2951 ierr = MPIULong_Send(vals,nz,MPIU_SCALAR,i,((PetscObject)newMat)->tag,comm);CHKERRQ(ierr); 2952 } 2953 ierr = PetscFree(procsnz);CHKERRQ(ierr); 2954 } else { 2955 /* receive numeric values */ 2956 ierr = PetscMalloc1(nz+1,&vals);CHKERRQ(ierr); 2957 2958 /* receive message of values*/ 2959 ierr = MPIULong_Recv(vals,nz,MPIU_SCALAR,0,((PetscObject)newMat)->tag,comm);CHKERRQ(ierr); 2960 2961 /* insert into matrix */ 2962 jj = rstart; 2963 smycols = mycols; 2964 svals = vals; 2965 for (i=0; i<m; i++) { 2966 ierr = MatSetValues_MPIAIJ(newMat,1,&jj,ourlens[i],smycols,svals,INSERT_VALUES);CHKERRQ(ierr); 2967 smycols += ourlens[i]; 2968 svals += ourlens[i]; 2969 jj++; 2970 } 2971 } 2972 ierr = PetscFree2(ourlens,offlens);CHKERRQ(ierr); 2973 ierr = PetscFree(vals);CHKERRQ(ierr); 2974 ierr = PetscFree(mycols);CHKERRQ(ierr); 2975 ierr = PetscFree(rowners);CHKERRQ(ierr); 2976 ierr = MatAssemblyBegin(newMat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2977 ierr = MatAssemblyEnd(newMat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2978 PetscFunctionReturn(0); 2979 } 2980 2981 /* Not scalable because of ISAllGather() unless getting all columns. */ 2982 PetscErrorCode ISGetSeqIS_Private(Mat mat,IS iscol,IS *isseq) 2983 { 2984 PetscErrorCode ierr; 2985 IS iscol_local; 2986 PetscBool isstride; 2987 PetscMPIInt lisstride=0,gisstride; 2988 2989 PetscFunctionBegin; 2990 /* check if we are grabbing all columns*/ 2991 ierr = PetscObjectTypeCompare((PetscObject)iscol,ISSTRIDE,&isstride);CHKERRQ(ierr); 2992 2993 if (isstride) { 2994 PetscInt start,len,mstart,mlen; 2995 ierr = ISStrideGetInfo(iscol,&start,NULL);CHKERRQ(ierr); 2996 ierr = ISGetLocalSize(iscol,&len);CHKERRQ(ierr); 2997 ierr = MatGetOwnershipRangeColumn(mat,&mstart,&mlen);CHKERRQ(ierr); 2998 if (mstart == start && mlen-mstart == len) lisstride = 1; 2999 } 3000 3001 ierr = MPIU_Allreduce(&lisstride,&gisstride,1,MPI_INT,MPI_MIN,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 3002 if (gisstride) { 3003 PetscInt N; 3004 ierr = MatGetSize(mat,NULL,&N);CHKERRQ(ierr); 3005 ierr = ISCreateStride(PetscObjectComm((PetscObject)mat),N,0,1,&iscol_local);CHKERRQ(ierr); 3006 ierr = ISSetIdentity(iscol_local);CHKERRQ(ierr); 3007 ierr = PetscInfo(mat,"Optimizing for obtaining all columns of the matrix; skipping ISAllGather()\n");CHKERRQ(ierr); 3008 } else { 3009 PetscInt cbs; 3010 ierr = ISGetBlockSize(iscol,&cbs);CHKERRQ(ierr); 3011 ierr = ISAllGather(iscol,&iscol_local);CHKERRQ(ierr); 3012 ierr = ISSetBlockSize(iscol_local,cbs);CHKERRQ(ierr); 3013 } 3014 3015 *isseq = iscol_local; 3016 PetscFunctionReturn(0); 3017 } 3018 3019 /* 3020 Used by MatCreateSubMatrix_MPIAIJ_SameRowColDist() to avoid ISAllGather() and global size of iscol_local 3021 (see MatCreateSubMatrix_MPIAIJ_nonscalable) 3022 3023 Input Parameters: 3024 mat - matrix 3025 isrow - parallel row index set; its local indices are a subset of local columns of mat, 3026 i.e., mat->rstart <= isrow[i] < mat->rend 3027 iscol - parallel column index set; its local indices are a subset of local columns of mat, 3028 i.e., mat->cstart <= iscol[i] < mat->cend 3029 Output Parameter: 3030 isrow_d,iscol_d - sequential row and column index sets for retrieving mat->A 3031 iscol_o - sequential column index set for retrieving mat->B 3032 garray - column map; garray[i] indicates global location of iscol_o[i] in iscol 3033 */ 3034 PetscErrorCode ISGetSeqIS_SameColDist_Private(Mat mat,IS isrow,IS iscol,IS *isrow_d,IS *iscol_d,IS *iscol_o,const PetscInt *garray[]) 3035 { 3036 PetscErrorCode ierr; 3037 Vec x,cmap; 3038 const PetscInt *is_idx; 3039 PetscScalar *xarray,*cmaparray; 3040 PetscInt ncols,isstart,*idx,m,rstart,*cmap1,count; 3041 Mat_MPIAIJ *a=(Mat_MPIAIJ*)mat->data; 3042 Mat B=a->B; 3043 Vec lvec=a->lvec,lcmap; 3044 PetscInt i,cstart,cend,Bn=B->cmap->N; 3045 MPI_Comm comm; 3046 VecScatter Mvctx=a->Mvctx; 3047 3048 PetscFunctionBegin; 3049 ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 3050 ierr = ISGetLocalSize(iscol,&ncols);CHKERRQ(ierr); 3051 3052 /* (1) iscol is a sub-column vector of mat, pad it with '-1.' to form a full vector x */ 3053 ierr = MatCreateVecs(mat,&x,NULL);CHKERRQ(ierr); 3054 ierr = VecSet(x,-1.0);CHKERRQ(ierr); 3055 ierr = VecDuplicate(x,&cmap);CHKERRQ(ierr); 3056 ierr = VecSet(cmap,-1.0);CHKERRQ(ierr); 3057 3058 /* Get start indices */ 3059 ierr = MPI_Scan(&ncols,&isstart,1,MPIU_INT,MPI_SUM,comm);CHKERRQ(ierr); 3060 isstart -= ncols; 3061 ierr = MatGetOwnershipRangeColumn(mat,&cstart,&cend);CHKERRQ(ierr); 3062 3063 ierr = ISGetIndices(iscol,&is_idx);CHKERRQ(ierr); 3064 ierr = VecGetArray(x,&xarray);CHKERRQ(ierr); 3065 ierr = VecGetArray(cmap,&cmaparray);CHKERRQ(ierr); 3066 ierr = PetscMalloc1(ncols,&idx);CHKERRQ(ierr); 3067 for (i=0; i<ncols; i++) { 3068 xarray[is_idx[i]-cstart] = (PetscScalar)is_idx[i]; 3069 cmaparray[is_idx[i]-cstart] = i + isstart; /* global index of iscol[i] */ 3070 idx[i] = is_idx[i]-cstart; /* local index of iscol[i] */ 3071 } 3072 ierr = VecRestoreArray(x,&xarray);CHKERRQ(ierr); 3073 ierr = VecRestoreArray(cmap,&cmaparray);CHKERRQ(ierr); 3074 ierr = ISRestoreIndices(iscol,&is_idx);CHKERRQ(ierr); 3075 3076 /* Get iscol_d */ 3077 ierr = ISCreateGeneral(PETSC_COMM_SELF,ncols,idx,PETSC_OWN_POINTER,iscol_d);CHKERRQ(ierr); 3078 ierr = ISGetBlockSize(iscol,&i);CHKERRQ(ierr); 3079 ierr = ISSetBlockSize(*iscol_d,i);CHKERRQ(ierr); 3080 3081 /* Get isrow_d */ 3082 ierr = ISGetLocalSize(isrow,&m);CHKERRQ(ierr); 3083 rstart = mat->rmap->rstart; 3084 ierr = PetscMalloc1(m,&idx);CHKERRQ(ierr); 3085 ierr = ISGetIndices(isrow,&is_idx);CHKERRQ(ierr); 3086 for (i=0; i<m; i++) idx[i] = is_idx[i]-rstart; 3087 ierr = ISRestoreIndices(isrow,&is_idx);CHKERRQ(ierr); 3088 3089 ierr = ISCreateGeneral(PETSC_COMM_SELF,m,idx,PETSC_OWN_POINTER,isrow_d);CHKERRQ(ierr); 3090 ierr = ISGetBlockSize(isrow,&i);CHKERRQ(ierr); 3091 ierr = ISSetBlockSize(*isrow_d,i);CHKERRQ(ierr); 3092 3093 /* (2) Scatter x and cmap using aij->Mvctx to get their off-process portions (see MatMult_MPIAIJ) */ 3094 ierr = VecScatterBegin(Mvctx,x,lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3095 ierr = VecScatterEnd(Mvctx,x,lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3096 3097 ierr = VecDuplicate(lvec,&lcmap);CHKERRQ(ierr); 3098 3099 ierr = VecScatterBegin(Mvctx,cmap,lcmap,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3100 ierr = VecScatterEnd(Mvctx,cmap,lcmap,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3101 3102 /* (3) create sequential iscol_o (a subset of iscol) and isgarray */ 3103 /* off-process column indices */ 3104 count = 0; 3105 ierr = PetscMalloc1(Bn,&idx);CHKERRQ(ierr); 3106 ierr = PetscMalloc1(Bn,&cmap1);CHKERRQ(ierr); 3107 3108 ierr = VecGetArray(lvec,&xarray);CHKERRQ(ierr); 3109 ierr = VecGetArray(lcmap,&cmaparray);CHKERRQ(ierr); 3110 for (i=0; i<Bn; i++) { 3111 if (PetscRealPart(xarray[i]) > -1.0) { 3112 idx[count] = i; /* local column index in off-diagonal part B */ 3113 cmap1[count] = (PetscInt)PetscRealPart(cmaparray[i]); /* column index in submat */ 3114 count++; 3115 } 3116 } 3117 ierr = VecRestoreArray(lvec,&xarray);CHKERRQ(ierr); 3118 ierr = VecRestoreArray(lcmap,&cmaparray);CHKERRQ(ierr); 3119 3120 ierr = ISCreateGeneral(PETSC_COMM_SELF,count,idx,PETSC_COPY_VALUES,iscol_o);CHKERRQ(ierr); 3121 /* cannot ensure iscol_o has same blocksize as iscol! */ 3122 3123 ierr = PetscFree(idx);CHKERRQ(ierr); 3124 *garray = cmap1; 3125 3126 ierr = VecDestroy(&x);CHKERRQ(ierr); 3127 ierr = VecDestroy(&cmap);CHKERRQ(ierr); 3128 ierr = VecDestroy(&lcmap);CHKERRQ(ierr); 3129 PetscFunctionReturn(0); 3130 } 3131 3132 /* isrow and iscol have same processor distribution as mat, output *submat is a submatrix of local mat */ 3133 PetscErrorCode MatCreateSubMatrix_MPIAIJ_SameRowColDist(Mat mat,IS isrow,IS iscol,MatReuse call,Mat *submat) 3134 { 3135 PetscErrorCode ierr; 3136 Mat_MPIAIJ *a = (Mat_MPIAIJ*)mat->data,*asub; 3137 Mat M = NULL; 3138 MPI_Comm comm; 3139 IS iscol_d,isrow_d,iscol_o; 3140 Mat Asub = NULL,Bsub = NULL; 3141 PetscInt n; 3142 3143 PetscFunctionBegin; 3144 ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 3145 3146 if (call == MAT_REUSE_MATRIX) { 3147 /* Retrieve isrow_d, iscol_d and iscol_o from submat */ 3148 ierr = PetscObjectQuery((PetscObject)*submat,"isrow_d",(PetscObject*)&isrow_d);CHKERRQ(ierr); 3149 if (!isrow_d) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"isrow_d passed in was not used before, cannot reuse"); 3150 3151 ierr = PetscObjectQuery((PetscObject)*submat,"iscol_d",(PetscObject*)&iscol_d);CHKERRQ(ierr); 3152 if (!iscol_d) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"iscol_d passed in was not used before, cannot reuse"); 3153 3154 ierr = PetscObjectQuery((PetscObject)*submat,"iscol_o",(PetscObject*)&iscol_o);CHKERRQ(ierr); 3155 if (!iscol_o) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"iscol_o passed in was not used before, cannot reuse"); 3156 3157 /* Update diagonal and off-diagonal portions of submat */ 3158 asub = (Mat_MPIAIJ*)(*submat)->data; 3159 ierr = MatCreateSubMatrix_SeqAIJ(a->A,isrow_d,iscol_d,PETSC_DECIDE,MAT_REUSE_MATRIX,&asub->A);CHKERRQ(ierr); 3160 ierr = ISGetLocalSize(iscol_o,&n);CHKERRQ(ierr); 3161 if (n) { 3162 ierr = MatCreateSubMatrix_SeqAIJ(a->B,isrow_d,iscol_o,PETSC_DECIDE,MAT_REUSE_MATRIX,&asub->B);CHKERRQ(ierr); 3163 } 3164 ierr = MatAssemblyBegin(*submat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3165 ierr = MatAssemblyEnd(*submat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3166 3167 } else { /* call == MAT_INITIAL_MATRIX) */ 3168 const PetscInt *garray; 3169 PetscInt BsubN; 3170 3171 /* Create isrow_d, iscol_d, iscol_o and isgarray (replace isgarray with array?) */ 3172 ierr = ISGetSeqIS_SameColDist_Private(mat,isrow,iscol,&isrow_d,&iscol_d,&iscol_o,&garray);CHKERRQ(ierr); 3173 3174 /* Create local submatrices Asub and Bsub */ 3175 ierr = MatCreateSubMatrix_SeqAIJ(a->A,isrow_d,iscol_d,PETSC_DECIDE,MAT_INITIAL_MATRIX,&Asub);CHKERRQ(ierr); 3176 ierr = MatCreateSubMatrix_SeqAIJ(a->B,isrow_d,iscol_o,PETSC_DECIDE,MAT_INITIAL_MATRIX,&Bsub);CHKERRQ(ierr); 3177 3178 /* Create submatrix M */ 3179 ierr = MatCreateMPIAIJWithSeqAIJ(comm,Asub,Bsub,garray,&M);CHKERRQ(ierr); 3180 3181 /* If Bsub has empty columns, compress iscol_o such that it will retrieve condensed Bsub from a->B during reuse */ 3182 asub = (Mat_MPIAIJ*)M->data; 3183 3184 ierr = ISGetLocalSize(iscol_o,&BsubN);CHKERRQ(ierr); 3185 n = asub->B->cmap->N; 3186 if (BsubN > n) { 3187 /* This case can be tested using ~petsc/src/tao/bound/examples/tutorials/runplate2_3 */ 3188 const PetscInt *idx; 3189 PetscInt i,j,*idx_new,*subgarray = asub->garray; 3190 ierr = PetscInfo2(M,"submatrix Bn %D != BsubN %D, update iscol_o\n",n,BsubN);CHKERRQ(ierr); 3191 3192 ierr = PetscMalloc1(n,&idx_new);CHKERRQ(ierr); 3193 j = 0; 3194 ierr = ISGetIndices(iscol_o,&idx);CHKERRQ(ierr); 3195 for (i=0; i<n; i++) { 3196 if (j >= BsubN) break; 3197 while (subgarray[i] > garray[j]) j++; 3198 3199 if (subgarray[i] == garray[j]) { 3200 idx_new[i] = idx[j++]; 3201 } else SETERRQ4(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"subgarray[%D]=%D cannot < garray[%D]=%D",i,subgarray[i],j,garray[j]); 3202 } 3203 ierr = ISRestoreIndices(iscol_o,&idx);CHKERRQ(ierr); 3204 3205 ierr = ISDestroy(&iscol_o);CHKERRQ(ierr); 3206 ierr = ISCreateGeneral(PETSC_COMM_SELF,n,idx_new,PETSC_OWN_POINTER,&iscol_o);CHKERRQ(ierr); 3207 3208 } else if (BsubN < n) { 3209 SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Columns of Bsub cannot be smaller than B's",BsubN,asub->B->cmap->N); 3210 } 3211 3212 ierr = PetscFree(garray);CHKERRQ(ierr); 3213 *submat = M; 3214 3215 /* Save isrow_d, iscol_d and iscol_o used in processor for next request */ 3216 ierr = PetscObjectCompose((PetscObject)M,"isrow_d",(PetscObject)isrow_d);CHKERRQ(ierr); 3217 ierr = ISDestroy(&isrow_d);CHKERRQ(ierr); 3218 3219 ierr = PetscObjectCompose((PetscObject)M,"iscol_d",(PetscObject)iscol_d);CHKERRQ(ierr); 3220 ierr = ISDestroy(&iscol_d);CHKERRQ(ierr); 3221 3222 ierr = PetscObjectCompose((PetscObject)M,"iscol_o",(PetscObject)iscol_o);CHKERRQ(ierr); 3223 ierr = ISDestroy(&iscol_o);CHKERRQ(ierr); 3224 } 3225 PetscFunctionReturn(0); 3226 } 3227 3228 PetscErrorCode MatCreateSubMatrix_MPIAIJ(Mat mat,IS isrow,IS iscol,MatReuse call,Mat *newmat) 3229 { 3230 PetscErrorCode ierr; 3231 IS iscol_local=NULL,isrow_d; 3232 PetscInt csize; 3233 PetscInt n,i,j,start,end; 3234 PetscBool sameRowDist=PETSC_FALSE,sameDist[2],tsameDist[2]; 3235 MPI_Comm comm; 3236 3237 PetscFunctionBegin; 3238 /* If isrow has same processor distribution as mat, 3239 call MatCreateSubMatrix_MPIAIJ_SameRowDist() to avoid using a hash table with global size of iscol */ 3240 if (call == MAT_REUSE_MATRIX) { 3241 ierr = PetscObjectQuery((PetscObject)*newmat,"isrow_d",(PetscObject*)&isrow_d);CHKERRQ(ierr); 3242 if (isrow_d) { 3243 sameRowDist = PETSC_TRUE; 3244 tsameDist[1] = PETSC_TRUE; /* sameColDist */ 3245 } else { 3246 ierr = PetscObjectQuery((PetscObject)*newmat,"SubIScol",(PetscObject*)&iscol_local);CHKERRQ(ierr); 3247 if (iscol_local) { 3248 sameRowDist = PETSC_TRUE; 3249 tsameDist[1] = PETSC_FALSE; /* !sameColDist */ 3250 } 3251 } 3252 } else { 3253 /* Check if isrow has same processor distribution as mat */ 3254 sameDist[0] = PETSC_FALSE; 3255 ierr = ISGetLocalSize(isrow,&n);CHKERRQ(ierr); 3256 if (!n) { 3257 sameDist[0] = PETSC_TRUE; 3258 } else { 3259 ierr = ISGetMinMax(isrow,&i,&j);CHKERRQ(ierr); 3260 ierr = MatGetOwnershipRange(mat,&start,&end);CHKERRQ(ierr); 3261 if (i >= start && j < end) { 3262 sameDist[0] = PETSC_TRUE; 3263 } 3264 } 3265 3266 /* Check if iscol has same processor distribution as mat */ 3267 sameDist[1] = PETSC_FALSE; 3268 ierr = ISGetLocalSize(iscol,&n);CHKERRQ(ierr); 3269 if (!n) { 3270 sameDist[1] = PETSC_TRUE; 3271 } else { 3272 ierr = ISGetMinMax(iscol,&i,&j);CHKERRQ(ierr); 3273 ierr = MatGetOwnershipRangeColumn(mat,&start,&end);CHKERRQ(ierr); 3274 if (i >= start && j < end) sameDist[1] = PETSC_TRUE; 3275 } 3276 3277 ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 3278 ierr = MPIU_Allreduce(&sameDist,&tsameDist,2,MPIU_BOOL,MPI_LAND,comm);CHKERRQ(ierr); 3279 sameRowDist = tsameDist[0]; 3280 } 3281 3282 if (sameRowDist) { 3283 if (tsameDist[1]) { /* sameRowDist & sameColDist */ 3284 /* isrow and iscol have same processor distribution as mat */ 3285 ierr = MatCreateSubMatrix_MPIAIJ_SameRowColDist(mat,isrow,iscol,call,newmat);CHKERRQ(ierr); 3286 PetscFunctionReturn(0); 3287 } else { /* sameRowDist */ 3288 /* isrow has same processor distribution as mat */ 3289 if (call == MAT_INITIAL_MATRIX) { 3290 PetscBool sorted; 3291 ierr = ISGetSeqIS_Private(mat,iscol,&iscol_local);CHKERRQ(ierr); 3292 ierr = ISGetLocalSize(iscol_local,&n);CHKERRQ(ierr); /* local size of iscol_local = global columns of newmat */ 3293 ierr = ISGetSize(iscol,&i);CHKERRQ(ierr); 3294 if (n != i) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"n %d != size of iscol %d",n,i); 3295 3296 ierr = ISSorted(iscol_local,&sorted);CHKERRQ(ierr); 3297 if (sorted) { 3298 /* MatCreateSubMatrix_MPIAIJ_SameRowDist() requires iscol_local be sorted; it can have duplicate indices */ 3299 ierr = MatCreateSubMatrix_MPIAIJ_SameRowDist(mat,isrow,iscol,iscol_local,MAT_INITIAL_MATRIX,newmat);CHKERRQ(ierr); 3300 PetscFunctionReturn(0); 3301 } 3302 } else { /* call == MAT_REUSE_MATRIX */ 3303 IS iscol_sub; 3304 ierr = PetscObjectQuery((PetscObject)*newmat,"SubIScol",(PetscObject*)&iscol_sub);CHKERRQ(ierr); 3305 if (iscol_sub) { 3306 ierr = MatCreateSubMatrix_MPIAIJ_SameRowDist(mat,isrow,iscol,NULL,call,newmat);CHKERRQ(ierr); 3307 PetscFunctionReturn(0); 3308 } 3309 } 3310 } 3311 } 3312 3313 /* General case: iscol -> iscol_local which has global size of iscol */ 3314 if (call == MAT_REUSE_MATRIX) { 3315 ierr = PetscObjectQuery((PetscObject)*newmat,"ISAllGather",(PetscObject*)&iscol_local);CHKERRQ(ierr); 3316 if (!iscol_local) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Submatrix passed in was not used before, cannot reuse"); 3317 } else { 3318 if (!iscol_local) { 3319 ierr = ISGetSeqIS_Private(mat,iscol,&iscol_local);CHKERRQ(ierr); 3320 } 3321 } 3322 3323 ierr = ISGetLocalSize(iscol,&csize);CHKERRQ(ierr); 3324 ierr = MatCreateSubMatrix_MPIAIJ_nonscalable(mat,isrow,iscol_local,csize,call,newmat);CHKERRQ(ierr); 3325 3326 if (call == MAT_INITIAL_MATRIX) { 3327 ierr = PetscObjectCompose((PetscObject)*newmat,"ISAllGather",(PetscObject)iscol_local);CHKERRQ(ierr); 3328 ierr = ISDestroy(&iscol_local);CHKERRQ(ierr); 3329 } 3330 PetscFunctionReturn(0); 3331 } 3332 3333 /*@C 3334 MatCreateMPIAIJWithSeqAIJ - creates a MPIAIJ matrix using SeqAIJ matrices that contain the "diagonal" 3335 and "off-diagonal" part of the matrix in CSR format. 3336 3337 Collective on MPI_Comm 3338 3339 Input Parameters: 3340 + comm - MPI communicator 3341 . A - "diagonal" portion of matrix 3342 . B - "off-diagonal" portion of matrix, may have empty columns, will be destroyed by this routine 3343 - garray - global index of B columns 3344 3345 Output Parameter: 3346 . mat - the matrix, with input A as its local diagonal matrix 3347 Level: advanced 3348 3349 Notes: 3350 See MatCreateAIJ() for the definition of "diagonal" and "off-diagonal" portion of the matrix. 3351 A becomes part of output mat, B is destroyed by this routine. The user cannot use A and B anymore. 3352 3353 .seealso: MatCreateMPIAIJWithSplitArrays() 3354 @*/ 3355 PetscErrorCode MatCreateMPIAIJWithSeqAIJ(MPI_Comm comm,Mat A,Mat B,const PetscInt garray[],Mat *mat) 3356 { 3357 PetscErrorCode ierr; 3358 Mat_MPIAIJ *maij; 3359 Mat_SeqAIJ *b=(Mat_SeqAIJ*)B->data,*bnew; 3360 PetscInt *oi=b->i,*oj=b->j,i,nz,col; 3361 PetscScalar *oa=b->a; 3362 Mat Bnew; 3363 PetscInt m,n,N; 3364 3365 PetscFunctionBegin; 3366 ierr = MatCreate(comm,mat);CHKERRQ(ierr); 3367 ierr = MatGetSize(A,&m,&n);CHKERRQ(ierr); 3368 if (m != B->rmap->N) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Am %D != Bm %D",m,B->rmap->N); 3369 if (A->rmap->bs != B->rmap->bs) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"A row bs %D != B row bs %D",A->rmap->bs,B->rmap->bs); 3370 /* remove check below; When B is created using iscol_o from ISGetSeqIS_SameColDist_Private(), its bs may not be same as A */ 3371 /* if (A->cmap->bs != B->cmap->bs) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"A column bs %D != B column bs %D",A->cmap->bs,B->cmap->bs); */ 3372 3373 /* Get global columns of mat */ 3374 ierr = MPIU_Allreduce(&n,&N,1,MPIU_INT,MPI_SUM,comm);CHKERRQ(ierr); 3375 3376 ierr = MatSetSizes(*mat,m,n,PETSC_DECIDE,N);CHKERRQ(ierr); 3377 ierr = MatSetType(*mat,MATMPIAIJ);CHKERRQ(ierr); 3378 ierr = MatSetBlockSizes(*mat,A->rmap->bs,A->cmap->bs);CHKERRQ(ierr); 3379 maij = (Mat_MPIAIJ*)(*mat)->data; 3380 3381 (*mat)->preallocated = PETSC_TRUE; 3382 3383 ierr = PetscLayoutSetUp((*mat)->rmap);CHKERRQ(ierr); 3384 ierr = PetscLayoutSetUp((*mat)->cmap);CHKERRQ(ierr); 3385 3386 /* Set A as diagonal portion of *mat */ 3387 maij->A = A; 3388 3389 nz = oi[m]; 3390 for (i=0; i<nz; i++) { 3391 col = oj[i]; 3392 oj[i] = garray[col]; 3393 } 3394 3395 /* Set Bnew as off-diagonal portion of *mat */ 3396 ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,m,N,oi,oj,oa,&Bnew);CHKERRQ(ierr); 3397 bnew = (Mat_SeqAIJ*)Bnew->data; 3398 bnew->maxnz = b->maxnz; /* allocated nonzeros of B */ 3399 maij->B = Bnew; 3400 3401 if (B->rmap->N != Bnew->rmap->N) SETERRQ2(PETSC_COMM_SELF,0,"BN %d != BnewN %d",B->rmap->N,Bnew->rmap->N); 3402 3403 b->singlemalloc = PETSC_FALSE; /* B arrays are shared by Bnew */ 3404 b->free_a = PETSC_FALSE; 3405 b->free_ij = PETSC_FALSE; 3406 ierr = MatDestroy(&B);CHKERRQ(ierr); 3407 3408 bnew->singlemalloc = PETSC_TRUE; /* arrays will be freed by MatDestroy(&Bnew) */ 3409 bnew->free_a = PETSC_TRUE; 3410 bnew->free_ij = PETSC_TRUE; 3411 3412 /* condense columns of maij->B */ 3413 ierr = MatSetOption(*mat,MAT_NO_OFF_PROC_ENTRIES,PETSC_TRUE);CHKERRQ(ierr); 3414 ierr = MatAssemblyBegin(*mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3415 ierr = MatAssemblyEnd(*mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3416 ierr = MatSetOption(*mat,MAT_NO_OFF_PROC_ENTRIES,PETSC_FALSE);CHKERRQ(ierr); 3417 ierr = MatSetOption(*mat,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE);CHKERRQ(ierr); 3418 PetscFunctionReturn(0); 3419 } 3420 3421 extern PetscErrorCode MatCreateSubMatrices_MPIAIJ_SingleIS_Local(Mat,PetscInt,const IS[],const IS[],MatReuse,PetscBool,Mat*); 3422 3423 PetscErrorCode MatCreateSubMatrix_MPIAIJ_SameRowDist(Mat mat,IS isrow,IS iscol,IS iscol_local,MatReuse call,Mat *newmat) 3424 { 3425 PetscErrorCode ierr; 3426 PetscInt i,m,n,rstart,row,rend,nz,j,bs,cbs; 3427 PetscInt *ii,*jj,nlocal,*dlens,*olens,dlen,olen,jend,mglobal; 3428 Mat_MPIAIJ *a=(Mat_MPIAIJ*)mat->data; 3429 Mat M,Msub,B=a->B; 3430 MatScalar *aa; 3431 Mat_SeqAIJ *aij; 3432 PetscInt *garray = a->garray,*colsub,Ncols; 3433 PetscInt count,Bn=B->cmap->N,cstart=mat->cmap->rstart,cend=mat->cmap->rend; 3434 IS iscol_sub,iscmap; 3435 const PetscInt *is_idx,*cmap; 3436 PetscBool allcolumns=PETSC_FALSE; 3437 MPI_Comm comm; 3438 3439 PetscFunctionBegin; 3440 ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 3441 3442 if (call == MAT_REUSE_MATRIX) { 3443 ierr = PetscObjectQuery((PetscObject)*newmat,"SubIScol",(PetscObject*)&iscol_sub);CHKERRQ(ierr); 3444 if (!iscol_sub) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"SubIScol passed in was not used before, cannot reuse"); 3445 ierr = ISGetLocalSize(iscol_sub,&count);CHKERRQ(ierr); 3446 3447 ierr = PetscObjectQuery((PetscObject)*newmat,"Subcmap",(PetscObject*)&iscmap);CHKERRQ(ierr); 3448 if (!iscmap) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Subcmap passed in was not used before, cannot reuse"); 3449 3450 ierr = PetscObjectQuery((PetscObject)*newmat,"SubMatrix",(PetscObject*)&Msub);CHKERRQ(ierr); 3451 if (!Msub) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Submatrix passed in was not used before, cannot reuse"); 3452 3453 ierr = MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat,1,&isrow,&iscol_sub,MAT_REUSE_MATRIX,PETSC_FALSE,&Msub);CHKERRQ(ierr); 3454 3455 } else { /* call == MAT_INITIAL_MATRIX) */ 3456 PetscBool flg; 3457 3458 ierr = ISGetLocalSize(iscol,&n);CHKERRQ(ierr); 3459 ierr = ISGetSize(iscol,&Ncols);CHKERRQ(ierr); 3460 3461 /* (1) iscol -> nonscalable iscol_local */ 3462 /* Check for special case: each processor gets entire matrix columns */ 3463 ierr = ISIdentity(iscol_local,&flg);CHKERRQ(ierr); 3464 if (flg && n == mat->cmap->N) allcolumns = PETSC_TRUE; 3465 if (allcolumns) { 3466 iscol_sub = iscol_local; 3467 ierr = PetscObjectReference((PetscObject)iscol_local);CHKERRQ(ierr); 3468 ierr = ISCreateStride(PETSC_COMM_SELF,n,0,1,&iscmap);CHKERRQ(ierr); 3469 3470 } else { 3471 /* (2) iscol_local -> iscol_sub and iscmap. Implementation below requires iscol_local be sorted, it can have duplicate indices */ 3472 PetscInt *idx,*cmap1,k; 3473 ierr = PetscMalloc1(Ncols,&idx);CHKERRQ(ierr); 3474 ierr = PetscMalloc1(Ncols,&cmap1);CHKERRQ(ierr); 3475 ierr = ISGetIndices(iscol_local,&is_idx);CHKERRQ(ierr); 3476 count = 0; 3477 k = 0; 3478 for (i=0; i<Ncols; i++) { 3479 j = is_idx[i]; 3480 if (j >= cstart && j < cend) { 3481 /* diagonal part of mat */ 3482 idx[count] = j; 3483 cmap1[count++] = i; /* column index in submat */ 3484 } else if (Bn) { 3485 /* off-diagonal part of mat */ 3486 if (j == garray[k]) { 3487 idx[count] = j; 3488 cmap1[count++] = i; /* column index in submat */ 3489 } else if (j > garray[k]) { 3490 while (j > garray[k] && k < Bn-1) k++; 3491 if (j == garray[k]) { 3492 idx[count] = j; 3493 cmap1[count++] = i; /* column index in submat */ 3494 } 3495 } 3496 } 3497 } 3498 ierr = ISRestoreIndices(iscol_local,&is_idx);CHKERRQ(ierr); 3499 3500 ierr = ISCreateGeneral(PETSC_COMM_SELF,count,idx,PETSC_OWN_POINTER,&iscol_sub);CHKERRQ(ierr); 3501 ierr = ISGetBlockSize(iscol,&cbs);CHKERRQ(ierr); 3502 ierr = ISSetBlockSize(iscol_sub,cbs);CHKERRQ(ierr); 3503 3504 ierr = ISCreateGeneral(PetscObjectComm((PetscObject)iscol_local),count,cmap1,PETSC_OWN_POINTER,&iscmap);CHKERRQ(ierr); 3505 } 3506 3507 /* (3) Create sequential Msub */ 3508 ierr = MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat,1,&isrow,&iscol_sub,MAT_INITIAL_MATRIX,allcolumns,&Msub);CHKERRQ(ierr); 3509 } 3510 3511 ierr = ISGetLocalSize(iscol_sub,&count);CHKERRQ(ierr); 3512 aij = (Mat_SeqAIJ*)(Msub)->data; 3513 ii = aij->i; 3514 ierr = ISGetIndices(iscmap,&cmap);CHKERRQ(ierr); 3515 3516 /* 3517 m - number of local rows 3518 Ncols - number of columns (same on all processors) 3519 rstart - first row in new global matrix generated 3520 */ 3521 ierr = MatGetSize(Msub,&m,NULL);CHKERRQ(ierr); 3522 3523 if (call == MAT_INITIAL_MATRIX) { 3524 /* (4) Create parallel newmat */ 3525 PetscMPIInt rank,size; 3526 PetscInt csize; 3527 3528 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 3529 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 3530 3531 /* 3532 Determine the number of non-zeros in the diagonal and off-diagonal 3533 portions of the matrix in order to do correct preallocation 3534 */ 3535 3536 /* first get start and end of "diagonal" columns */ 3537 ierr = ISGetLocalSize(iscol,&csize);CHKERRQ(ierr); 3538 if (csize == PETSC_DECIDE) { 3539 ierr = ISGetSize(isrow,&mglobal);CHKERRQ(ierr); 3540 if (mglobal == Ncols) { /* square matrix */ 3541 nlocal = m; 3542 } else { 3543 nlocal = Ncols/size + ((Ncols % size) > rank); 3544 } 3545 } else { 3546 nlocal = csize; 3547 } 3548 ierr = MPI_Scan(&nlocal,&rend,1,MPIU_INT,MPI_SUM,comm);CHKERRQ(ierr); 3549 rstart = rend - nlocal; 3550 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); 3551 3552 /* next, compute all the lengths */ 3553 jj = aij->j; 3554 ierr = PetscMalloc1(2*m+1,&dlens);CHKERRQ(ierr); 3555 olens = dlens + m; 3556 for (i=0; i<m; i++) { 3557 jend = ii[i+1] - ii[i]; 3558 olen = 0; 3559 dlen = 0; 3560 for (j=0; j<jend; j++) { 3561 if (cmap[*jj] < rstart || cmap[*jj] >= rend) olen++; 3562 else dlen++; 3563 jj++; 3564 } 3565 olens[i] = olen; 3566 dlens[i] = dlen; 3567 } 3568 3569 ierr = ISGetBlockSize(isrow,&bs);CHKERRQ(ierr); 3570 ierr = ISGetBlockSize(iscol,&cbs);CHKERRQ(ierr); 3571 3572 ierr = MatCreate(comm,&M);CHKERRQ(ierr); 3573 ierr = MatSetSizes(M,m,nlocal,PETSC_DECIDE,Ncols);CHKERRQ(ierr); 3574 ierr = MatSetBlockSizes(M,bs,cbs);CHKERRQ(ierr); 3575 ierr = MatSetType(M,((PetscObject)mat)->type_name);CHKERRQ(ierr); 3576 ierr = MatMPIAIJSetPreallocation(M,0,dlens,0,olens);CHKERRQ(ierr); 3577 ierr = PetscFree(dlens);CHKERRQ(ierr); 3578 3579 } else { /* call == MAT_REUSE_MATRIX */ 3580 M = *newmat; 3581 ierr = MatGetLocalSize(M,&i,NULL);CHKERRQ(ierr); 3582 if (i != m) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Previous matrix must be same size/layout as request"); 3583 ierr = MatZeroEntries(M);CHKERRQ(ierr); 3584 /* 3585 The next two lines are needed so we may call MatSetValues_MPIAIJ() below directly, 3586 rather than the slower MatSetValues(). 3587 */ 3588 M->was_assembled = PETSC_TRUE; 3589 M->assembled = PETSC_FALSE; 3590 } 3591 3592 /* (5) Set values of Msub to *newmat */ 3593 ierr = PetscMalloc1(count,&colsub);CHKERRQ(ierr); 3594 ierr = MatGetOwnershipRange(M,&rstart,NULL);CHKERRQ(ierr); 3595 3596 jj = aij->j; 3597 aa = aij->a; 3598 for (i=0; i<m; i++) { 3599 row = rstart + i; 3600 nz = ii[i+1] - ii[i]; 3601 for (j=0; j<nz; j++) colsub[j] = cmap[jj[j]]; 3602 ierr = MatSetValues_MPIAIJ(M,1,&row,nz,colsub,aa,INSERT_VALUES);CHKERRQ(ierr); 3603 jj += nz; aa += nz; 3604 } 3605 ierr = ISRestoreIndices(iscmap,&cmap);CHKERRQ(ierr); 3606 3607 ierr = MatAssemblyBegin(M,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3608 ierr = MatAssemblyEnd(M,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3609 3610 ierr = PetscFree(colsub);CHKERRQ(ierr); 3611 3612 /* save Msub, iscol_sub and iscmap used in processor for next request */ 3613 if (call == MAT_INITIAL_MATRIX) { 3614 *newmat = M; 3615 ierr = PetscObjectCompose((PetscObject)(*newmat),"SubMatrix",(PetscObject)Msub);CHKERRQ(ierr); 3616 ierr = MatDestroy(&Msub);CHKERRQ(ierr); 3617 3618 ierr = PetscObjectCompose((PetscObject)(*newmat),"SubIScol",(PetscObject)iscol_sub);CHKERRQ(ierr); 3619 ierr = ISDestroy(&iscol_sub);CHKERRQ(ierr); 3620 3621 ierr = PetscObjectCompose((PetscObject)(*newmat),"Subcmap",(PetscObject)iscmap);CHKERRQ(ierr); 3622 ierr = ISDestroy(&iscmap);CHKERRQ(ierr); 3623 3624 if (iscol_local) { 3625 ierr = PetscObjectCompose((PetscObject)(*newmat),"ISAllGather",(PetscObject)iscol_local);CHKERRQ(ierr); 3626 ierr = ISDestroy(&iscol_local);CHKERRQ(ierr); 3627 } 3628 } 3629 PetscFunctionReturn(0); 3630 } 3631 3632 /* 3633 Not great since it makes two copies of the submatrix, first an SeqAIJ 3634 in local and then by concatenating the local matrices the end result. 3635 Writing it directly would be much like MatCreateSubMatrices_MPIAIJ() 3636 3637 Note: This requires a sequential iscol with all indices. 3638 */ 3639 PetscErrorCode MatCreateSubMatrix_MPIAIJ_nonscalable(Mat mat,IS isrow,IS iscol,PetscInt csize,MatReuse call,Mat *newmat) 3640 { 3641 PetscErrorCode ierr; 3642 PetscMPIInt rank,size; 3643 PetscInt i,m,n,rstart,row,rend,nz,*cwork,j,bs,cbs; 3644 PetscInt *ii,*jj,nlocal,*dlens,*olens,dlen,olen,jend,mglobal; 3645 Mat M,Mreuse; 3646 MatScalar *aa,*vwork; 3647 MPI_Comm comm; 3648 Mat_SeqAIJ *aij; 3649 PetscBool colflag,allcolumns=PETSC_FALSE; 3650 3651 PetscFunctionBegin; 3652 ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 3653 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 3654 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 3655 3656 /* Check for special case: each processor gets entire matrix columns */ 3657 ierr = ISIdentity(iscol,&colflag);CHKERRQ(ierr); 3658 ierr = ISGetLocalSize(iscol,&n);CHKERRQ(ierr); 3659 if (colflag && n == mat->cmap->N) allcolumns = PETSC_TRUE; 3660 3661 if (call == MAT_REUSE_MATRIX) { 3662 ierr = PetscObjectQuery((PetscObject)*newmat,"SubMatrix",(PetscObject*)&Mreuse);CHKERRQ(ierr); 3663 if (!Mreuse) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Submatrix passed in was not used before, cannot reuse"); 3664 ierr = MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat,1,&isrow,&iscol,MAT_REUSE_MATRIX,allcolumns,&Mreuse);CHKERRQ(ierr); 3665 } else { 3666 ierr = MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat,1,&isrow,&iscol,MAT_INITIAL_MATRIX,allcolumns,&Mreuse);CHKERRQ(ierr); 3667 } 3668 3669 /* 3670 m - number of local rows 3671 n - number of columns (same on all processors) 3672 rstart - first row in new global matrix generated 3673 */ 3674 ierr = MatGetSize(Mreuse,&m,&n);CHKERRQ(ierr); 3675 ierr = MatGetBlockSizes(Mreuse,&bs,&cbs);CHKERRQ(ierr); 3676 if (call == MAT_INITIAL_MATRIX) { 3677 aij = (Mat_SeqAIJ*)(Mreuse)->data; 3678 ii = aij->i; 3679 jj = aij->j; 3680 3681 /* 3682 Determine the number of non-zeros in the diagonal and off-diagonal 3683 portions of the matrix in order to do correct preallocation 3684 */ 3685 3686 /* first get start and end of "diagonal" columns */ 3687 if (csize == PETSC_DECIDE) { 3688 ierr = ISGetSize(isrow,&mglobal);CHKERRQ(ierr); 3689 if (mglobal == n) { /* square matrix */ 3690 nlocal = m; 3691 } else { 3692 nlocal = n/size + ((n % size) > rank); 3693 } 3694 } else { 3695 nlocal = csize; 3696 } 3697 ierr = MPI_Scan(&nlocal,&rend,1,MPIU_INT,MPI_SUM,comm);CHKERRQ(ierr); 3698 rstart = rend - nlocal; 3699 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); 3700 3701 /* next, compute all the lengths */ 3702 ierr = PetscMalloc1(2*m+1,&dlens);CHKERRQ(ierr); 3703 olens = dlens + m; 3704 for (i=0; i<m; i++) { 3705 jend = ii[i+1] - ii[i]; 3706 olen = 0; 3707 dlen = 0; 3708 for (j=0; j<jend; j++) { 3709 if (*jj < rstart || *jj >= rend) olen++; 3710 else dlen++; 3711 jj++; 3712 } 3713 olens[i] = olen; 3714 dlens[i] = dlen; 3715 } 3716 ierr = MatCreate(comm,&M);CHKERRQ(ierr); 3717 ierr = MatSetSizes(M,m,nlocal,PETSC_DECIDE,n);CHKERRQ(ierr); 3718 ierr = MatSetBlockSizes(M,bs,cbs);CHKERRQ(ierr); 3719 ierr = MatSetType(M,((PetscObject)mat)->type_name);CHKERRQ(ierr); 3720 ierr = MatMPIAIJSetPreallocation(M,0,dlens,0,olens);CHKERRQ(ierr); 3721 ierr = PetscFree(dlens);CHKERRQ(ierr); 3722 } else { 3723 PetscInt ml,nl; 3724 3725 M = *newmat; 3726 ierr = MatGetLocalSize(M,&ml,&nl);CHKERRQ(ierr); 3727 if (ml != m) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Previous matrix must be same size/layout as request"); 3728 ierr = MatZeroEntries(M);CHKERRQ(ierr); 3729 /* 3730 The next two lines are needed so we may call MatSetValues_MPIAIJ() below directly, 3731 rather than the slower MatSetValues(). 3732 */ 3733 M->was_assembled = PETSC_TRUE; 3734 M->assembled = PETSC_FALSE; 3735 } 3736 ierr = MatGetOwnershipRange(M,&rstart,&rend);CHKERRQ(ierr); 3737 aij = (Mat_SeqAIJ*)(Mreuse)->data; 3738 ii = aij->i; 3739 jj = aij->j; 3740 aa = aij->a; 3741 for (i=0; i<m; i++) { 3742 row = rstart + i; 3743 nz = ii[i+1] - ii[i]; 3744 cwork = jj; jj += nz; 3745 vwork = aa; aa += nz; 3746 ierr = MatSetValues_MPIAIJ(M,1,&row,nz,cwork,vwork,INSERT_VALUES);CHKERRQ(ierr); 3747 } 3748 3749 ierr = MatAssemblyBegin(M,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3750 ierr = MatAssemblyEnd(M,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3751 *newmat = M; 3752 3753 /* save submatrix used in processor for next request */ 3754 if (call == MAT_INITIAL_MATRIX) { 3755 ierr = PetscObjectCompose((PetscObject)M,"SubMatrix",(PetscObject)Mreuse);CHKERRQ(ierr); 3756 ierr = MatDestroy(&Mreuse);CHKERRQ(ierr); 3757 } 3758 PetscFunctionReturn(0); 3759 } 3760 3761 PetscErrorCode MatMPIAIJSetPreallocationCSR_MPIAIJ(Mat B,const PetscInt Ii[],const PetscInt J[],const PetscScalar v[]) 3762 { 3763 PetscInt m,cstart, cend,j,nnz,i,d; 3764 PetscInt *d_nnz,*o_nnz,nnz_max = 0,rstart,ii; 3765 const PetscInt *JJ; 3766 PetscScalar *values; 3767 PetscErrorCode ierr; 3768 PetscBool nooffprocentries; 3769 3770 PetscFunctionBegin; 3771 if (Ii[0]) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Ii[0] must be 0 it is %D",Ii[0]); 3772 3773 ierr = PetscLayoutSetUp(B->rmap);CHKERRQ(ierr); 3774 ierr = PetscLayoutSetUp(B->cmap);CHKERRQ(ierr); 3775 m = B->rmap->n; 3776 cstart = B->cmap->rstart; 3777 cend = B->cmap->rend; 3778 rstart = B->rmap->rstart; 3779 3780 ierr = PetscMalloc2(m,&d_nnz,m,&o_nnz);CHKERRQ(ierr); 3781 3782 #if defined(PETSC_USE_DEBUGGING) 3783 for (i=0; i<m; i++) { 3784 nnz = Ii[i+1]- Ii[i]; 3785 JJ = J + Ii[i]; 3786 if (nnz < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Local row %D has a negative %D number of columns",i,nnz); 3787 if (nnz && (JJ[0] < 0)) SETERRRQ1(PETSC_ERR_ARG_WRONGSTATE,"Row %D starts with negative column index",i,j); 3788 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); 3789 } 3790 #endif 3791 3792 for (i=0; i<m; i++) { 3793 nnz = Ii[i+1]- Ii[i]; 3794 JJ = J + Ii[i]; 3795 nnz_max = PetscMax(nnz_max,nnz); 3796 d = 0; 3797 for (j=0; j<nnz; j++) { 3798 if (cstart <= JJ[j] && JJ[j] < cend) d++; 3799 } 3800 d_nnz[i] = d; 3801 o_nnz[i] = nnz - d; 3802 } 3803 ierr = MatMPIAIJSetPreallocation(B,0,d_nnz,0,o_nnz);CHKERRQ(ierr); 3804 ierr = PetscFree2(d_nnz,o_nnz);CHKERRQ(ierr); 3805 3806 if (v) values = (PetscScalar*)v; 3807 else { 3808 ierr = PetscCalloc1(nnz_max+1,&values);CHKERRQ(ierr); 3809 } 3810 3811 for (i=0; i<m; i++) { 3812 ii = i + rstart; 3813 nnz = Ii[i+1]- Ii[i]; 3814 ierr = MatSetValues_MPIAIJ(B,1,&ii,nnz,J+Ii[i],values+(v ? Ii[i] : 0),INSERT_VALUES);CHKERRQ(ierr); 3815 } 3816 nooffprocentries = B->nooffprocentries; 3817 B->nooffprocentries = PETSC_TRUE; 3818 ierr = MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3819 ierr = MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3820 B->nooffprocentries = nooffprocentries; 3821 3822 if (!v) { 3823 ierr = PetscFree(values);CHKERRQ(ierr); 3824 } 3825 ierr = MatSetOption(B,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE);CHKERRQ(ierr); 3826 PetscFunctionReturn(0); 3827 } 3828 3829 /*@ 3830 MatMPIAIJSetPreallocationCSR - Allocates memory for a sparse parallel matrix in AIJ format 3831 (the default parallel PETSc format). 3832 3833 Collective on MPI_Comm 3834 3835 Input Parameters: 3836 + B - the matrix 3837 . i - the indices into j for the start of each local row (starts with zero) 3838 . j - the column indices for each local row (starts with zero) 3839 - v - optional values in the matrix 3840 3841 Level: developer 3842 3843 Notes: 3844 The i, j, and a arrays ARE copied by this routine into the internal format used by PETSc; 3845 thus you CANNOT change the matrix entries by changing the values of a[] after you have 3846 called this routine. Use MatCreateMPIAIJWithSplitArrays() to avoid needing to copy the arrays. 3847 3848 The i and j indices are 0 based, and i indices are indices corresponding to the local j array. 3849 3850 The format which is used for the sparse matrix input, is equivalent to a 3851 row-major ordering.. i.e for the following matrix, the input data expected is 3852 as shown 3853 3854 $ 1 0 0 3855 $ 2 0 3 P0 3856 $ ------- 3857 $ 4 5 6 P1 3858 $ 3859 $ Process0 [P0]: rows_owned=[0,1] 3860 $ i = {0,1,3} [size = nrow+1 = 2+1] 3861 $ j = {0,0,2} [size = 3] 3862 $ v = {1,2,3} [size = 3] 3863 $ 3864 $ Process1 [P1]: rows_owned=[2] 3865 $ i = {0,3} [size = nrow+1 = 1+1] 3866 $ j = {0,1,2} [size = 3] 3867 $ v = {4,5,6} [size = 3] 3868 3869 .keywords: matrix, aij, compressed row, sparse, parallel 3870 3871 .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatCreateAIJ(), MATMPIAIJ, 3872 MatCreateSeqAIJWithArrays(), MatCreateMPIAIJWithSplitArrays() 3873 @*/ 3874 PetscErrorCode MatMPIAIJSetPreallocationCSR(Mat B,const PetscInt i[],const PetscInt j[], const PetscScalar v[]) 3875 { 3876 PetscErrorCode ierr; 3877 3878 PetscFunctionBegin; 3879 ierr = PetscTryMethod(B,"MatMPIAIJSetPreallocationCSR_C",(Mat,const PetscInt[],const PetscInt[],const PetscScalar[]),(B,i,j,v));CHKERRQ(ierr); 3880 PetscFunctionReturn(0); 3881 } 3882 3883 /*@C 3884 MatMPIAIJSetPreallocation - Preallocates memory for a sparse parallel matrix in AIJ format 3885 (the default parallel PETSc format). For good matrix assembly performance 3886 the user should preallocate the matrix storage by setting the parameters 3887 d_nz (or d_nnz) and o_nz (or o_nnz). By setting these parameters accurately, 3888 performance can be increased by more than a factor of 50. 3889 3890 Collective on MPI_Comm 3891 3892 Input Parameters: 3893 + B - the matrix 3894 . d_nz - number of nonzeros per row in DIAGONAL portion of local submatrix 3895 (same value is used for all local rows) 3896 . d_nnz - array containing the number of nonzeros in the various rows of the 3897 DIAGONAL portion of the local submatrix (possibly different for each row) 3898 or NULL (PETSC_NULL_INTEGER in Fortran), if d_nz is used to specify the nonzero structure. 3899 The size of this array is equal to the number of local rows, i.e 'm'. 3900 For matrices that will be factored, you must leave room for (and set) 3901 the diagonal entry even if it is zero. 3902 . o_nz - number of nonzeros per row in the OFF-DIAGONAL portion of local 3903 submatrix (same value is used for all local rows). 3904 - o_nnz - array containing the number of nonzeros in the various rows of the 3905 OFF-DIAGONAL portion of the local submatrix (possibly different for 3906 each row) or NULL (PETSC_NULL_INTEGER in Fortran), if o_nz is used to specify the nonzero 3907 structure. The size of this array is equal to the number 3908 of local rows, i.e 'm'. 3909 3910 If the *_nnz parameter is given then the *_nz parameter is ignored 3911 3912 The AIJ format (also called the Yale sparse matrix format or 3913 compressed row storage (CSR)), is fully compatible with standard Fortran 77 3914 storage. The stored row and column indices begin with zero. 3915 See Users-Manual: ch_mat for details. 3916 3917 The parallel matrix is partitioned such that the first m0 rows belong to 3918 process 0, the next m1 rows belong to process 1, the next m2 rows belong 3919 to process 2 etc.. where m0,m1,m2... are the input parameter 'm'. 3920 3921 The DIAGONAL portion of the local submatrix of a processor can be defined 3922 as the submatrix which is obtained by extraction the part corresponding to 3923 the rows r1-r2 and columns c1-c2 of the global matrix, where r1 is the 3924 first row that belongs to the processor, r2 is the last row belonging to 3925 the this processor, and c1-c2 is range of indices of the local part of a 3926 vector suitable for applying the matrix to. This is an mxn matrix. In the 3927 common case of a square matrix, the row and column ranges are the same and 3928 the DIAGONAL part is also square. The remaining portion of the local 3929 submatrix (mxN) constitute the OFF-DIAGONAL portion. 3930 3931 If o_nnz, d_nnz are specified, then o_nz, and d_nz are ignored. 3932 3933 You can call MatGetInfo() to get information on how effective the preallocation was; 3934 for example the fields mallocs,nz_allocated,nz_used,nz_unneeded; 3935 You can also run with the option -info and look for messages with the string 3936 malloc in them to see if additional memory allocation was needed. 3937 3938 Example usage: 3939 3940 Consider the following 8x8 matrix with 34 non-zero values, that is 3941 assembled across 3 processors. Lets assume that proc0 owns 3 rows, 3942 proc1 owns 3 rows, proc2 owns 2 rows. This division can be shown 3943 as follows: 3944 3945 .vb 3946 1 2 0 | 0 3 0 | 0 4 3947 Proc0 0 5 6 | 7 0 0 | 8 0 3948 9 0 10 | 11 0 0 | 12 0 3949 ------------------------------------- 3950 13 0 14 | 15 16 17 | 0 0 3951 Proc1 0 18 0 | 19 20 21 | 0 0 3952 0 0 0 | 22 23 0 | 24 0 3953 ------------------------------------- 3954 Proc2 25 26 27 | 0 0 28 | 29 0 3955 30 0 0 | 31 32 33 | 0 34 3956 .ve 3957 3958 This can be represented as a collection of submatrices as: 3959 3960 .vb 3961 A B C 3962 D E F 3963 G H I 3964 .ve 3965 3966 Where the submatrices A,B,C are owned by proc0, D,E,F are 3967 owned by proc1, G,H,I are owned by proc2. 3968 3969 The 'm' parameters for proc0,proc1,proc2 are 3,3,2 respectively. 3970 The 'n' parameters for proc0,proc1,proc2 are 3,3,2 respectively. 3971 The 'M','N' parameters are 8,8, and have the same values on all procs. 3972 3973 The DIAGONAL submatrices corresponding to proc0,proc1,proc2 are 3974 submatrices [A], [E], [I] respectively. The OFF-DIAGONAL submatrices 3975 corresponding to proc0,proc1,proc2 are [BC], [DF], [GH] respectively. 3976 Internally, each processor stores the DIAGONAL part, and the OFF-DIAGONAL 3977 part as SeqAIJ matrices. for eg: proc1 will store [E] as a SeqAIJ 3978 matrix, ans [DF] as another SeqAIJ matrix. 3979 3980 When d_nz, o_nz parameters are specified, d_nz storage elements are 3981 allocated for every row of the local diagonal submatrix, and o_nz 3982 storage locations are allocated for every row of the OFF-DIAGONAL submat. 3983 One way to choose d_nz and o_nz is to use the max nonzerors per local 3984 rows for each of the local DIAGONAL, and the OFF-DIAGONAL submatrices. 3985 In this case, the values of d_nz,o_nz are: 3986 .vb 3987 proc0 : dnz = 2, o_nz = 2 3988 proc1 : dnz = 3, o_nz = 2 3989 proc2 : dnz = 1, o_nz = 4 3990 .ve 3991 We are allocating m*(d_nz+o_nz) storage locations for every proc. This 3992 translates to 3*(2+2)=12 for proc0, 3*(3+2)=15 for proc1, 2*(1+4)=10 3993 for proc3. i.e we are using 12+15+10=37 storage locations to store 3994 34 values. 3995 3996 When d_nnz, o_nnz parameters are specified, the storage is specified 3997 for every row, coresponding to both DIAGONAL and OFF-DIAGONAL submatrices. 3998 In the above case the values for d_nnz,o_nnz are: 3999 .vb 4000 proc0: d_nnz = [2,2,2] and o_nnz = [2,2,2] 4001 proc1: d_nnz = [3,3,2] and o_nnz = [2,1,1] 4002 proc2: d_nnz = [1,1] and o_nnz = [4,4] 4003 .ve 4004 Here the space allocated is sum of all the above values i.e 34, and 4005 hence pre-allocation is perfect. 4006 4007 Level: intermediate 4008 4009 .keywords: matrix, aij, compressed row, sparse, parallel 4010 4011 .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatCreateAIJ(), MatMPIAIJSetPreallocationCSR(), 4012 MATMPIAIJ, MatGetInfo(), PetscSplitOwnership() 4013 @*/ 4014 PetscErrorCode MatMPIAIJSetPreallocation(Mat B,PetscInt d_nz,const PetscInt d_nnz[],PetscInt o_nz,const PetscInt o_nnz[]) 4015 { 4016 PetscErrorCode ierr; 4017 4018 PetscFunctionBegin; 4019 PetscValidHeaderSpecific(B,MAT_CLASSID,1); 4020 PetscValidType(B,1); 4021 ierr = PetscTryMethod(B,"MatMPIAIJSetPreallocation_C",(Mat,PetscInt,const PetscInt[],PetscInt,const PetscInt[]),(B,d_nz,d_nnz,o_nz,o_nnz));CHKERRQ(ierr); 4022 PetscFunctionReturn(0); 4023 } 4024 4025 /*@ 4026 MatCreateMPIAIJWithArrays - creates a MPI AIJ matrix using arrays that contain in standard 4027 CSR format the local rows. 4028 4029 Collective on MPI_Comm 4030 4031 Input Parameters: 4032 + comm - MPI communicator 4033 . m - number of local rows (Cannot be PETSC_DECIDE) 4034 . n - This value should be the same as the local size used in creating the 4035 x vector for the matrix-vector product y = Ax. (or PETSC_DECIDE to have 4036 calculated if N is given) For square matrices n is almost always m. 4037 . M - number of global rows (or PETSC_DETERMINE to have calculated if m is given) 4038 . N - number of global columns (or PETSC_DETERMINE to have calculated if n is given) 4039 . i - row indices 4040 . j - column indices 4041 - a - matrix values 4042 4043 Output Parameter: 4044 . mat - the matrix 4045 4046 Level: intermediate 4047 4048 Notes: 4049 The i, j, and a arrays ARE copied by this routine into the internal format used by PETSc; 4050 thus you CANNOT change the matrix entries by changing the values of a[] after you have 4051 called this routine. Use MatCreateMPIAIJWithSplitArrays() to avoid needing to copy the arrays. 4052 4053 The i and j indices are 0 based, and i indices are indices corresponding to the local j array. 4054 4055 The format which is used for the sparse matrix input, is equivalent to a 4056 row-major ordering.. i.e for the following matrix, the input data expected is 4057 as shown 4058 4059 $ 1 0 0 4060 $ 2 0 3 P0 4061 $ ------- 4062 $ 4 5 6 P1 4063 $ 4064 $ Process0 [P0]: rows_owned=[0,1] 4065 $ i = {0,1,3} [size = nrow+1 = 2+1] 4066 $ j = {0,0,2} [size = 3] 4067 $ v = {1,2,3} [size = 3] 4068 $ 4069 $ Process1 [P1]: rows_owned=[2] 4070 $ i = {0,3} [size = nrow+1 = 1+1] 4071 $ j = {0,1,2} [size = 3] 4072 $ v = {4,5,6} [size = 3] 4073 4074 .keywords: matrix, aij, compressed row, sparse, parallel 4075 4076 .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatMPIAIJSetPreallocationCSR(), 4077 MATMPIAIJ, MatCreateAIJ(), MatCreateMPIAIJWithSplitArrays() 4078 @*/ 4079 PetscErrorCode MatCreateMPIAIJWithArrays(MPI_Comm comm,PetscInt m,PetscInt n,PetscInt M,PetscInt N,const PetscInt i[],const PetscInt j[],const PetscScalar a[],Mat *mat) 4080 { 4081 PetscErrorCode ierr; 4082 4083 PetscFunctionBegin; 4084 if (i[0]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"i (row indices) must start with 0"); 4085 if (m < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"local number of rows (m) cannot be PETSC_DECIDE, or negative"); 4086 ierr = MatCreate(comm,mat);CHKERRQ(ierr); 4087 ierr = MatSetSizes(*mat,m,n,M,N);CHKERRQ(ierr); 4088 /* ierr = MatSetBlockSizes(M,bs,cbs);CHKERRQ(ierr); */ 4089 ierr = MatSetType(*mat,MATMPIAIJ);CHKERRQ(ierr); 4090 ierr = MatMPIAIJSetPreallocationCSR(*mat,i,j,a);CHKERRQ(ierr); 4091 PetscFunctionReturn(0); 4092 } 4093 4094 /*@C 4095 MatCreateAIJ - Creates a sparse parallel matrix in AIJ format 4096 (the default parallel PETSc format). For good matrix assembly performance 4097 the user should preallocate the matrix storage by setting the parameters 4098 d_nz (or d_nnz) and o_nz (or o_nnz). By setting these parameters accurately, 4099 performance can be increased by more than a factor of 50. 4100 4101 Collective on MPI_Comm 4102 4103 Input Parameters: 4104 + comm - MPI communicator 4105 . m - number of local rows (or PETSC_DECIDE to have calculated if M is given) 4106 This value should be the same as the local size used in creating the 4107 y vector for the matrix-vector product y = Ax. 4108 . n - This value should be the same as the local size used in creating the 4109 x vector for the matrix-vector product y = Ax. (or PETSC_DECIDE to have 4110 calculated if N is given) For square matrices n is almost always m. 4111 . M - number of global rows (or PETSC_DETERMINE to have calculated if m is given) 4112 . N - number of global columns (or PETSC_DETERMINE to have calculated if n is given) 4113 . d_nz - number of nonzeros per row in DIAGONAL portion of local submatrix 4114 (same value is used for all local rows) 4115 . d_nnz - array containing the number of nonzeros in the various rows of the 4116 DIAGONAL portion of the local submatrix (possibly different for each row) 4117 or NULL, if d_nz is used to specify the nonzero structure. 4118 The size of this array is equal to the number of local rows, i.e 'm'. 4119 . o_nz - number of nonzeros per row in the OFF-DIAGONAL portion of local 4120 submatrix (same value is used for all local rows). 4121 - o_nnz - array containing the number of nonzeros in the various rows of the 4122 OFF-DIAGONAL portion of the local submatrix (possibly different for 4123 each row) or NULL, if o_nz is used to specify the nonzero 4124 structure. The size of this array is equal to the number 4125 of local rows, i.e 'm'. 4126 4127 Output Parameter: 4128 . A - the matrix 4129 4130 It is recommended that one use the MatCreate(), MatSetType() and/or MatSetFromOptions(), 4131 MatXXXXSetPreallocation() paradgm instead of this routine directly. 4132 [MatXXXXSetPreallocation() is, for example, MatSeqAIJSetPreallocation] 4133 4134 Notes: 4135 If the *_nnz parameter is given then the *_nz parameter is ignored 4136 4137 m,n,M,N parameters specify the size of the matrix, and its partitioning across 4138 processors, while d_nz,d_nnz,o_nz,o_nnz parameters specify the approximate 4139 storage requirements for this matrix. 4140 4141 If PETSC_DECIDE or PETSC_DETERMINE is used for a particular argument on one 4142 processor than it must be used on all processors that share the object for 4143 that argument. 4144 4145 The user MUST specify either the local or global matrix dimensions 4146 (possibly both). 4147 4148 The parallel matrix is partitioned across processors such that the 4149 first m0 rows belong to process 0, the next m1 rows belong to 4150 process 1, the next m2 rows belong to process 2 etc.. where 4151 m0,m1,m2,.. are the input parameter 'm'. i.e each processor stores 4152 values corresponding to [m x N] submatrix. 4153 4154 The columns are logically partitioned with the n0 columns belonging 4155 to 0th partition, the next n1 columns belonging to the next 4156 partition etc.. where n0,n1,n2... are the input parameter 'n'. 4157 4158 The DIAGONAL portion of the local submatrix on any given processor 4159 is the submatrix corresponding to the rows and columns m,n 4160 corresponding to the given processor. i.e diagonal matrix on 4161 process 0 is [m0 x n0], diagonal matrix on process 1 is [m1 x n1] 4162 etc. The remaining portion of the local submatrix [m x (N-n)] 4163 constitute the OFF-DIAGONAL portion. The example below better 4164 illustrates this concept. 4165 4166 For a square global matrix we define each processor's diagonal portion 4167 to be its local rows and the corresponding columns (a square submatrix); 4168 each processor's off-diagonal portion encompasses the remainder of the 4169 local matrix (a rectangular submatrix). 4170 4171 If o_nnz, d_nnz are specified, then o_nz, and d_nz are ignored. 4172 4173 When calling this routine with a single process communicator, a matrix of 4174 type SEQAIJ is returned. If a matrix of type MPIAIJ is desired for this 4175 type of communicator, use the construction mechanism 4176 .vb 4177 MatCreate(...,&A); MatSetType(A,MATMPIAIJ); MatSetSizes(A, m,n,M,N); MatMPIAIJSetPreallocation(A,...); 4178 .ve 4179 4180 $ MatCreate(...,&A); 4181 $ MatSetType(A,MATMPIAIJ); 4182 $ MatSetSizes(A, m,n,M,N); 4183 $ MatMPIAIJSetPreallocation(A,...); 4184 4185 By default, this format uses inodes (identical nodes) when possible. 4186 We search for consecutive rows with the same nonzero structure, thereby 4187 reusing matrix information to achieve increased efficiency. 4188 4189 Options Database Keys: 4190 + -mat_no_inode - Do not use inodes 4191 . -mat_inode_limit <limit> - Sets inode limit (max limit=5) 4192 - -mat_aij_oneindex - Internally use indexing starting at 1 4193 rather than 0. Note that when calling MatSetValues(), 4194 the user still MUST index entries starting at 0! 4195 4196 4197 Example usage: 4198 4199 Consider the following 8x8 matrix with 34 non-zero values, that is 4200 assembled across 3 processors. Lets assume that proc0 owns 3 rows, 4201 proc1 owns 3 rows, proc2 owns 2 rows. This division can be shown 4202 as follows 4203 4204 .vb 4205 1 2 0 | 0 3 0 | 0 4 4206 Proc0 0 5 6 | 7 0 0 | 8 0 4207 9 0 10 | 11 0 0 | 12 0 4208 ------------------------------------- 4209 13 0 14 | 15 16 17 | 0 0 4210 Proc1 0 18 0 | 19 20 21 | 0 0 4211 0 0 0 | 22 23 0 | 24 0 4212 ------------------------------------- 4213 Proc2 25 26 27 | 0 0 28 | 29 0 4214 30 0 0 | 31 32 33 | 0 34 4215 .ve 4216 4217 This can be represented as a collection of submatrices as 4218 4219 .vb 4220 A B C 4221 D E F 4222 G H I 4223 .ve 4224 4225 Where the submatrices A,B,C are owned by proc0, D,E,F are 4226 owned by proc1, G,H,I are owned by proc2. 4227 4228 The 'm' parameters for proc0,proc1,proc2 are 3,3,2 respectively. 4229 The 'n' parameters for proc0,proc1,proc2 are 3,3,2 respectively. 4230 The 'M','N' parameters are 8,8, and have the same values on all procs. 4231 4232 The DIAGONAL submatrices corresponding to proc0,proc1,proc2 are 4233 submatrices [A], [E], [I] respectively. The OFF-DIAGONAL submatrices 4234 corresponding to proc0,proc1,proc2 are [BC], [DF], [GH] respectively. 4235 Internally, each processor stores the DIAGONAL part, and the OFF-DIAGONAL 4236 part as SeqAIJ matrices. for eg: proc1 will store [E] as a SeqAIJ 4237 matrix, ans [DF] as another SeqAIJ matrix. 4238 4239 When d_nz, o_nz parameters are specified, d_nz storage elements are 4240 allocated for every row of the local diagonal submatrix, and o_nz 4241 storage locations are allocated for every row of the OFF-DIAGONAL submat. 4242 One way to choose d_nz and o_nz is to use the max nonzerors per local 4243 rows for each of the local DIAGONAL, and the OFF-DIAGONAL submatrices. 4244 In this case, the values of d_nz,o_nz are 4245 .vb 4246 proc0 : dnz = 2, o_nz = 2 4247 proc1 : dnz = 3, o_nz = 2 4248 proc2 : dnz = 1, o_nz = 4 4249 .ve 4250 We are allocating m*(d_nz+o_nz) storage locations for every proc. This 4251 translates to 3*(2+2)=12 for proc0, 3*(3+2)=15 for proc1, 2*(1+4)=10 4252 for proc3. i.e we are using 12+15+10=37 storage locations to store 4253 34 values. 4254 4255 When d_nnz, o_nnz parameters are specified, the storage is specified 4256 for every row, coresponding to both DIAGONAL and OFF-DIAGONAL submatrices. 4257 In the above case the values for d_nnz,o_nnz are 4258 .vb 4259 proc0: d_nnz = [2,2,2] and o_nnz = [2,2,2] 4260 proc1: d_nnz = [3,3,2] and o_nnz = [2,1,1] 4261 proc2: d_nnz = [1,1] and o_nnz = [4,4] 4262 .ve 4263 Here the space allocated is sum of all the above values i.e 34, and 4264 hence pre-allocation is perfect. 4265 4266 Level: intermediate 4267 4268 .keywords: matrix, aij, compressed row, sparse, parallel 4269 4270 .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatMPIAIJSetPreallocationCSR(), 4271 MATMPIAIJ, MatCreateMPIAIJWithArrays() 4272 @*/ 4273 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) 4274 { 4275 PetscErrorCode ierr; 4276 PetscMPIInt size; 4277 4278 PetscFunctionBegin; 4279 ierr = MatCreate(comm,A);CHKERRQ(ierr); 4280 ierr = MatSetSizes(*A,m,n,M,N);CHKERRQ(ierr); 4281 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 4282 if (size > 1) { 4283 ierr = MatSetType(*A,MATMPIAIJ);CHKERRQ(ierr); 4284 ierr = MatMPIAIJSetPreallocation(*A,d_nz,d_nnz,o_nz,o_nnz);CHKERRQ(ierr); 4285 } else { 4286 ierr = MatSetType(*A,MATSEQAIJ);CHKERRQ(ierr); 4287 ierr = MatSeqAIJSetPreallocation(*A,d_nz,d_nnz);CHKERRQ(ierr); 4288 } 4289 PetscFunctionReturn(0); 4290 } 4291 4292 PetscErrorCode MatMPIAIJGetSeqAIJ(Mat A,Mat *Ad,Mat *Ao,const PetscInt *colmap[]) 4293 { 4294 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 4295 PetscBool flg; 4296 PetscErrorCode ierr; 4297 4298 PetscFunctionBegin; 4299 ierr = PetscObjectTypeCompare((PetscObject)A,MATMPIAIJ,&flg);CHKERRQ(ierr); 4300 if (!flg) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"This function requires a MATMPIAIJ matrix as input"); 4301 if (Ad) *Ad = a->A; 4302 if (Ao) *Ao = a->B; 4303 if (colmap) *colmap = a->garray; 4304 PetscFunctionReturn(0); 4305 } 4306 4307 PetscErrorCode MatCreateMPIMatConcatenateSeqMat_MPIAIJ(MPI_Comm comm,Mat inmat,PetscInt n,MatReuse scall,Mat *outmat) 4308 { 4309 PetscErrorCode ierr; 4310 PetscInt m,N,i,rstart,nnz,Ii; 4311 PetscInt *indx; 4312 PetscScalar *values; 4313 4314 PetscFunctionBegin; 4315 ierr = MatGetSize(inmat,&m,&N);CHKERRQ(ierr); 4316 if (scall == MAT_INITIAL_MATRIX) { /* symbolic phase */ 4317 PetscInt *dnz,*onz,sum,bs,cbs; 4318 4319 if (n == PETSC_DECIDE) { 4320 ierr = PetscSplitOwnership(comm,&n,&N);CHKERRQ(ierr); 4321 } 4322 /* Check sum(n) = N */ 4323 ierr = MPIU_Allreduce(&n,&sum,1,MPIU_INT,MPI_SUM,comm);CHKERRQ(ierr); 4324 if (sum != N) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_INCOMP,"Sum of local columns %D != global columns %D",sum,N); 4325 4326 ierr = MPI_Scan(&m, &rstart,1,MPIU_INT,MPI_SUM,comm);CHKERRQ(ierr); 4327 rstart -= m; 4328 4329 ierr = MatPreallocateInitialize(comm,m,n,dnz,onz);CHKERRQ(ierr); 4330 for (i=0; i<m; i++) { 4331 ierr = MatGetRow_SeqAIJ(inmat,i,&nnz,&indx,NULL);CHKERRQ(ierr); 4332 ierr = MatPreallocateSet(i+rstart,nnz,indx,dnz,onz);CHKERRQ(ierr); 4333 ierr = MatRestoreRow_SeqAIJ(inmat,i,&nnz,&indx,NULL);CHKERRQ(ierr); 4334 } 4335 4336 ierr = MatCreate(comm,outmat);CHKERRQ(ierr); 4337 ierr = MatSetSizes(*outmat,m,n,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 4338 ierr = MatGetBlockSizes(inmat,&bs,&cbs);CHKERRQ(ierr); 4339 ierr = MatSetBlockSizes(*outmat,bs,cbs);CHKERRQ(ierr); 4340 ierr = MatSetType(*outmat,MATAIJ);CHKERRQ(ierr); 4341 ierr = MatSeqAIJSetPreallocation(*outmat,0,dnz);CHKERRQ(ierr); 4342 ierr = MatMPIAIJSetPreallocation(*outmat,0,dnz,0,onz);CHKERRQ(ierr); 4343 ierr = MatPreallocateFinalize(dnz,onz);CHKERRQ(ierr); 4344 } 4345 4346 /* numeric phase */ 4347 ierr = MatGetOwnershipRange(*outmat,&rstart,NULL);CHKERRQ(ierr); 4348 for (i=0; i<m; i++) { 4349 ierr = MatGetRow_SeqAIJ(inmat,i,&nnz,&indx,&values);CHKERRQ(ierr); 4350 Ii = i + rstart; 4351 ierr = MatSetValues(*outmat,1,&Ii,nnz,indx,values,INSERT_VALUES);CHKERRQ(ierr); 4352 ierr = MatRestoreRow_SeqAIJ(inmat,i,&nnz,&indx,&values);CHKERRQ(ierr); 4353 } 4354 ierr = MatAssemblyBegin(*outmat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4355 ierr = MatAssemblyEnd(*outmat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4356 PetscFunctionReturn(0); 4357 } 4358 4359 PetscErrorCode MatFileSplit(Mat A,char *outfile) 4360 { 4361 PetscErrorCode ierr; 4362 PetscMPIInt rank; 4363 PetscInt m,N,i,rstart,nnz; 4364 size_t len; 4365 const PetscInt *indx; 4366 PetscViewer out; 4367 char *name; 4368 Mat B; 4369 const PetscScalar *values; 4370 4371 PetscFunctionBegin; 4372 ierr = MatGetLocalSize(A,&m,0);CHKERRQ(ierr); 4373 ierr = MatGetSize(A,0,&N);CHKERRQ(ierr); 4374 /* Should this be the type of the diagonal block of A? */ 4375 ierr = MatCreate(PETSC_COMM_SELF,&B);CHKERRQ(ierr); 4376 ierr = MatSetSizes(B,m,N,m,N);CHKERRQ(ierr); 4377 ierr = MatSetBlockSizesFromMats(B,A,A);CHKERRQ(ierr); 4378 ierr = MatSetType(B,MATSEQAIJ);CHKERRQ(ierr); 4379 ierr = MatSeqAIJSetPreallocation(B,0,NULL);CHKERRQ(ierr); 4380 ierr = MatGetOwnershipRange(A,&rstart,0);CHKERRQ(ierr); 4381 for (i=0; i<m; i++) { 4382 ierr = MatGetRow(A,i+rstart,&nnz,&indx,&values);CHKERRQ(ierr); 4383 ierr = MatSetValues(B,1,&i,nnz,indx,values,INSERT_VALUES);CHKERRQ(ierr); 4384 ierr = MatRestoreRow(A,i+rstart,&nnz,&indx,&values);CHKERRQ(ierr); 4385 } 4386 ierr = MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4387 ierr = MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4388 4389 ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)A),&rank);CHKERRQ(ierr); 4390 ierr = PetscStrlen(outfile,&len);CHKERRQ(ierr); 4391 ierr = PetscMalloc1(len+5,&name);CHKERRQ(ierr); 4392 sprintf(name,"%s.%d",outfile,rank); 4393 ierr = PetscViewerBinaryOpen(PETSC_COMM_SELF,name,FILE_MODE_APPEND,&out);CHKERRQ(ierr); 4394 ierr = PetscFree(name);CHKERRQ(ierr); 4395 ierr = MatView(B,out);CHKERRQ(ierr); 4396 ierr = PetscViewerDestroy(&out);CHKERRQ(ierr); 4397 ierr = MatDestroy(&B);CHKERRQ(ierr); 4398 PetscFunctionReturn(0); 4399 } 4400 4401 PetscErrorCode MatDestroy_MPIAIJ_SeqsToMPI(Mat A) 4402 { 4403 PetscErrorCode ierr; 4404 Mat_Merge_SeqsToMPI *merge; 4405 PetscContainer container; 4406 4407 PetscFunctionBegin; 4408 ierr = PetscObjectQuery((PetscObject)A,"MatMergeSeqsToMPI",(PetscObject*)&container);CHKERRQ(ierr); 4409 if (container) { 4410 ierr = PetscContainerGetPointer(container,(void**)&merge);CHKERRQ(ierr); 4411 ierr = PetscFree(merge->id_r);CHKERRQ(ierr); 4412 ierr = PetscFree(merge->len_s);CHKERRQ(ierr); 4413 ierr = PetscFree(merge->len_r);CHKERRQ(ierr); 4414 ierr = PetscFree(merge->bi);CHKERRQ(ierr); 4415 ierr = PetscFree(merge->bj);CHKERRQ(ierr); 4416 ierr = PetscFree(merge->buf_ri[0]);CHKERRQ(ierr); 4417 ierr = PetscFree(merge->buf_ri);CHKERRQ(ierr); 4418 ierr = PetscFree(merge->buf_rj[0]);CHKERRQ(ierr); 4419 ierr = PetscFree(merge->buf_rj);CHKERRQ(ierr); 4420 ierr = PetscFree(merge->coi);CHKERRQ(ierr); 4421 ierr = PetscFree(merge->coj);CHKERRQ(ierr); 4422 ierr = PetscFree(merge->owners_co);CHKERRQ(ierr); 4423 ierr = PetscLayoutDestroy(&merge->rowmap);CHKERRQ(ierr); 4424 ierr = PetscFree(merge);CHKERRQ(ierr); 4425 ierr = PetscObjectCompose((PetscObject)A,"MatMergeSeqsToMPI",0);CHKERRQ(ierr); 4426 } 4427 ierr = MatDestroy_MPIAIJ(A);CHKERRQ(ierr); 4428 PetscFunctionReturn(0); 4429 } 4430 4431 #include <../src/mat/utils/freespace.h> 4432 #include <petscbt.h> 4433 4434 PetscErrorCode MatCreateMPIAIJSumSeqAIJNumeric(Mat seqmat,Mat mpimat) 4435 { 4436 PetscErrorCode ierr; 4437 MPI_Comm comm; 4438 Mat_SeqAIJ *a =(Mat_SeqAIJ*)seqmat->data; 4439 PetscMPIInt size,rank,taga,*len_s; 4440 PetscInt N=mpimat->cmap->N,i,j,*owners,*ai=a->i,*aj; 4441 PetscInt proc,m; 4442 PetscInt **buf_ri,**buf_rj; 4443 PetscInt k,anzi,*bj_i,*bi,*bj,arow,bnzi,nextaj; 4444 PetscInt nrows,**buf_ri_k,**nextrow,**nextai; 4445 MPI_Request *s_waits,*r_waits; 4446 MPI_Status *status; 4447 MatScalar *aa=a->a; 4448 MatScalar **abuf_r,*ba_i; 4449 Mat_Merge_SeqsToMPI *merge; 4450 PetscContainer container; 4451 4452 PetscFunctionBegin; 4453 ierr = PetscObjectGetComm((PetscObject)mpimat,&comm);CHKERRQ(ierr); 4454 ierr = PetscLogEventBegin(MAT_Seqstompinum,seqmat,0,0,0);CHKERRQ(ierr); 4455 4456 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 4457 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 4458 4459 ierr = PetscObjectQuery((PetscObject)mpimat,"MatMergeSeqsToMPI",(PetscObject*)&container);CHKERRQ(ierr); 4460 ierr = PetscContainerGetPointer(container,(void**)&merge);CHKERRQ(ierr); 4461 4462 bi = merge->bi; 4463 bj = merge->bj; 4464 buf_ri = merge->buf_ri; 4465 buf_rj = merge->buf_rj; 4466 4467 ierr = PetscMalloc1(size,&status);CHKERRQ(ierr); 4468 owners = merge->rowmap->range; 4469 len_s = merge->len_s; 4470 4471 /* send and recv matrix values */ 4472 /*-----------------------------*/ 4473 ierr = PetscObjectGetNewTag((PetscObject)mpimat,&taga);CHKERRQ(ierr); 4474 ierr = PetscPostIrecvScalar(comm,taga,merge->nrecv,merge->id_r,merge->len_r,&abuf_r,&r_waits);CHKERRQ(ierr); 4475 4476 ierr = PetscMalloc1(merge->nsend+1,&s_waits);CHKERRQ(ierr); 4477 for (proc=0,k=0; proc<size; proc++) { 4478 if (!len_s[proc]) continue; 4479 i = owners[proc]; 4480 ierr = MPI_Isend(aa+ai[i],len_s[proc],MPIU_MATSCALAR,proc,taga,comm,s_waits+k);CHKERRQ(ierr); 4481 k++; 4482 } 4483 4484 if (merge->nrecv) {ierr = MPI_Waitall(merge->nrecv,r_waits,status);CHKERRQ(ierr);} 4485 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,s_waits,status);CHKERRQ(ierr);} 4486 ierr = PetscFree(status);CHKERRQ(ierr); 4487 4488 ierr = PetscFree(s_waits);CHKERRQ(ierr); 4489 ierr = PetscFree(r_waits);CHKERRQ(ierr); 4490 4491 /* insert mat values of mpimat */ 4492 /*----------------------------*/ 4493 ierr = PetscMalloc1(N,&ba_i);CHKERRQ(ierr); 4494 ierr = PetscMalloc3(merge->nrecv,&buf_ri_k,merge->nrecv,&nextrow,merge->nrecv,&nextai);CHKERRQ(ierr); 4495 4496 for (k=0; k<merge->nrecv; k++) { 4497 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 4498 nrows = *(buf_ri_k[k]); 4499 nextrow[k] = buf_ri_k[k]+1; /* next row number of k-th recved i-structure */ 4500 nextai[k] = buf_ri_k[k] + (nrows + 1); /* poins to the next i-structure of k-th recved i-structure */ 4501 } 4502 4503 /* set values of ba */ 4504 m = merge->rowmap->n; 4505 for (i=0; i<m; i++) { 4506 arow = owners[rank] + i; 4507 bj_i = bj+bi[i]; /* col indices of the i-th row of mpimat */ 4508 bnzi = bi[i+1] - bi[i]; 4509 ierr = PetscMemzero(ba_i,bnzi*sizeof(PetscScalar));CHKERRQ(ierr); 4510 4511 /* add local non-zero vals of this proc's seqmat into ba */ 4512 anzi = ai[arow+1] - ai[arow]; 4513 aj = a->j + ai[arow]; 4514 aa = a->a + ai[arow]; 4515 nextaj = 0; 4516 for (j=0; nextaj<anzi; j++) { 4517 if (*(bj_i + j) == aj[nextaj]) { /* bcol == acol */ 4518 ba_i[j] += aa[nextaj++]; 4519 } 4520 } 4521 4522 /* add received vals into ba */ 4523 for (k=0; k<merge->nrecv; k++) { /* k-th received message */ 4524 /* i-th row */ 4525 if (i == *nextrow[k]) { 4526 anzi = *(nextai[k]+1) - *nextai[k]; 4527 aj = buf_rj[k] + *(nextai[k]); 4528 aa = abuf_r[k] + *(nextai[k]); 4529 nextaj = 0; 4530 for (j=0; nextaj<anzi; j++) { 4531 if (*(bj_i + j) == aj[nextaj]) { /* bcol == acol */ 4532 ba_i[j] += aa[nextaj++]; 4533 } 4534 } 4535 nextrow[k]++; nextai[k]++; 4536 } 4537 } 4538 ierr = MatSetValues(mpimat,1,&arow,bnzi,bj_i,ba_i,INSERT_VALUES);CHKERRQ(ierr); 4539 } 4540 ierr = MatAssemblyBegin(mpimat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4541 ierr = MatAssemblyEnd(mpimat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4542 4543 ierr = PetscFree(abuf_r[0]);CHKERRQ(ierr); 4544 ierr = PetscFree(abuf_r);CHKERRQ(ierr); 4545 ierr = PetscFree(ba_i);CHKERRQ(ierr); 4546 ierr = PetscFree3(buf_ri_k,nextrow,nextai);CHKERRQ(ierr); 4547 ierr = PetscLogEventEnd(MAT_Seqstompinum,seqmat,0,0,0);CHKERRQ(ierr); 4548 PetscFunctionReturn(0); 4549 } 4550 4551 PetscErrorCode MatCreateMPIAIJSumSeqAIJSymbolic(MPI_Comm comm,Mat seqmat,PetscInt m,PetscInt n,Mat *mpimat) 4552 { 4553 PetscErrorCode ierr; 4554 Mat B_mpi; 4555 Mat_SeqAIJ *a=(Mat_SeqAIJ*)seqmat->data; 4556 PetscMPIInt size,rank,tagi,tagj,*len_s,*len_si,*len_ri; 4557 PetscInt **buf_rj,**buf_ri,**buf_ri_k; 4558 PetscInt M=seqmat->rmap->n,N=seqmat->cmap->n,i,*owners,*ai=a->i,*aj=a->j; 4559 PetscInt len,proc,*dnz,*onz,bs,cbs; 4560 PetscInt k,anzi,*bi,*bj,*lnk,nlnk,arow,bnzi,nspacedouble=0; 4561 PetscInt nrows,*buf_s,*buf_si,*buf_si_i,**nextrow,**nextai; 4562 MPI_Request *si_waits,*sj_waits,*ri_waits,*rj_waits; 4563 MPI_Status *status; 4564 PetscFreeSpaceList free_space=NULL,current_space=NULL; 4565 PetscBT lnkbt; 4566 Mat_Merge_SeqsToMPI *merge; 4567 PetscContainer container; 4568 4569 PetscFunctionBegin; 4570 ierr = PetscLogEventBegin(MAT_Seqstompisym,seqmat,0,0,0);CHKERRQ(ierr); 4571 4572 /* make sure it is a PETSc comm */ 4573 ierr = PetscCommDuplicate(comm,&comm,NULL);CHKERRQ(ierr); 4574 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 4575 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 4576 4577 ierr = PetscNew(&merge);CHKERRQ(ierr); 4578 ierr = PetscMalloc1(size,&status);CHKERRQ(ierr); 4579 4580 /* determine row ownership */ 4581 /*---------------------------------------------------------*/ 4582 ierr = PetscLayoutCreate(comm,&merge->rowmap);CHKERRQ(ierr); 4583 ierr = PetscLayoutSetLocalSize(merge->rowmap,m);CHKERRQ(ierr); 4584 ierr = PetscLayoutSetSize(merge->rowmap,M);CHKERRQ(ierr); 4585 ierr = PetscLayoutSetBlockSize(merge->rowmap,1);CHKERRQ(ierr); 4586 ierr = PetscLayoutSetUp(merge->rowmap);CHKERRQ(ierr); 4587 ierr = PetscMalloc1(size,&len_si);CHKERRQ(ierr); 4588 ierr = PetscMalloc1(size,&merge->len_s);CHKERRQ(ierr); 4589 4590 m = merge->rowmap->n; 4591 owners = merge->rowmap->range; 4592 4593 /* determine the number of messages to send, their lengths */ 4594 /*---------------------------------------------------------*/ 4595 len_s = merge->len_s; 4596 4597 len = 0; /* length of buf_si[] */ 4598 merge->nsend = 0; 4599 for (proc=0; proc<size; proc++) { 4600 len_si[proc] = 0; 4601 if (proc == rank) { 4602 len_s[proc] = 0; 4603 } else { 4604 len_si[proc] = owners[proc+1] - owners[proc] + 1; 4605 len_s[proc] = ai[owners[proc+1]] - ai[owners[proc]]; /* num of rows to be sent to [proc] */ 4606 } 4607 if (len_s[proc]) { 4608 merge->nsend++; 4609 nrows = 0; 4610 for (i=owners[proc]; i<owners[proc+1]; i++) { 4611 if (ai[i+1] > ai[i]) nrows++; 4612 } 4613 len_si[proc] = 2*(nrows+1); 4614 len += len_si[proc]; 4615 } 4616 } 4617 4618 /* determine the number and length of messages to receive for ij-structure */ 4619 /*-------------------------------------------------------------------------*/ 4620 ierr = PetscGatherNumberOfMessages(comm,NULL,len_s,&merge->nrecv);CHKERRQ(ierr); 4621 ierr = PetscGatherMessageLengths2(comm,merge->nsend,merge->nrecv,len_s,len_si,&merge->id_r,&merge->len_r,&len_ri);CHKERRQ(ierr); 4622 4623 /* post the Irecv of j-structure */ 4624 /*-------------------------------*/ 4625 ierr = PetscCommGetNewTag(comm,&tagj);CHKERRQ(ierr); 4626 ierr = PetscPostIrecvInt(comm,tagj,merge->nrecv,merge->id_r,merge->len_r,&buf_rj,&rj_waits);CHKERRQ(ierr); 4627 4628 /* post the Isend of j-structure */ 4629 /*--------------------------------*/ 4630 ierr = PetscMalloc2(merge->nsend,&si_waits,merge->nsend,&sj_waits);CHKERRQ(ierr); 4631 4632 for (proc=0, k=0; proc<size; proc++) { 4633 if (!len_s[proc]) continue; 4634 i = owners[proc]; 4635 ierr = MPI_Isend(aj+ai[i],len_s[proc],MPIU_INT,proc,tagj,comm,sj_waits+k);CHKERRQ(ierr); 4636 k++; 4637 } 4638 4639 /* receives and sends of j-structure are complete */ 4640 /*------------------------------------------------*/ 4641 if (merge->nrecv) {ierr = MPI_Waitall(merge->nrecv,rj_waits,status);CHKERRQ(ierr);} 4642 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,sj_waits,status);CHKERRQ(ierr);} 4643 4644 /* send and recv i-structure */ 4645 /*---------------------------*/ 4646 ierr = PetscCommGetNewTag(comm,&tagi);CHKERRQ(ierr); 4647 ierr = PetscPostIrecvInt(comm,tagi,merge->nrecv,merge->id_r,len_ri,&buf_ri,&ri_waits);CHKERRQ(ierr); 4648 4649 ierr = PetscMalloc1(len+1,&buf_s);CHKERRQ(ierr); 4650 buf_si = buf_s; /* points to the beginning of k-th msg to be sent */ 4651 for (proc=0,k=0; proc<size; proc++) { 4652 if (!len_s[proc]) continue; 4653 /* form outgoing message for i-structure: 4654 buf_si[0]: nrows to be sent 4655 [1:nrows]: row index (global) 4656 [nrows+1:2*nrows+1]: i-structure index 4657 */ 4658 /*-------------------------------------------*/ 4659 nrows = len_si[proc]/2 - 1; 4660 buf_si_i = buf_si + nrows+1; 4661 buf_si[0] = nrows; 4662 buf_si_i[0] = 0; 4663 nrows = 0; 4664 for (i=owners[proc]; i<owners[proc+1]; i++) { 4665 anzi = ai[i+1] - ai[i]; 4666 if (anzi) { 4667 buf_si_i[nrows+1] = buf_si_i[nrows] + anzi; /* i-structure */ 4668 buf_si[nrows+1] = i-owners[proc]; /* local row index */ 4669 nrows++; 4670 } 4671 } 4672 ierr = MPI_Isend(buf_si,len_si[proc],MPIU_INT,proc,tagi,comm,si_waits+k);CHKERRQ(ierr); 4673 k++; 4674 buf_si += len_si[proc]; 4675 } 4676 4677 if (merge->nrecv) {ierr = MPI_Waitall(merge->nrecv,ri_waits,status);CHKERRQ(ierr);} 4678 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,si_waits,status);CHKERRQ(ierr);} 4679 4680 ierr = PetscInfo2(seqmat,"nsend: %D, nrecv: %D\n",merge->nsend,merge->nrecv);CHKERRQ(ierr); 4681 for (i=0; i<merge->nrecv; i++) { 4682 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); 4683 } 4684 4685 ierr = PetscFree(len_si);CHKERRQ(ierr); 4686 ierr = PetscFree(len_ri);CHKERRQ(ierr); 4687 ierr = PetscFree(rj_waits);CHKERRQ(ierr); 4688 ierr = PetscFree2(si_waits,sj_waits);CHKERRQ(ierr); 4689 ierr = PetscFree(ri_waits);CHKERRQ(ierr); 4690 ierr = PetscFree(buf_s);CHKERRQ(ierr); 4691 ierr = PetscFree(status);CHKERRQ(ierr); 4692 4693 /* compute a local seq matrix in each processor */ 4694 /*----------------------------------------------*/ 4695 /* allocate bi array and free space for accumulating nonzero column info */ 4696 ierr = PetscMalloc1(m+1,&bi);CHKERRQ(ierr); 4697 bi[0] = 0; 4698 4699 /* create and initialize a linked list */ 4700 nlnk = N+1; 4701 ierr = PetscLLCreate(N,N,nlnk,lnk,lnkbt);CHKERRQ(ierr); 4702 4703 /* initial FreeSpace size is 2*(num of local nnz(seqmat)) */ 4704 len = ai[owners[rank+1]] - ai[owners[rank]]; 4705 ierr = PetscFreeSpaceGet(PetscIntMultTruncate(2,len)+1,&free_space);CHKERRQ(ierr); 4706 4707 current_space = free_space; 4708 4709 /* determine symbolic info for each local row */ 4710 ierr = PetscMalloc3(merge->nrecv,&buf_ri_k,merge->nrecv,&nextrow,merge->nrecv,&nextai);CHKERRQ(ierr); 4711 4712 for (k=0; k<merge->nrecv; k++) { 4713 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 4714 nrows = *buf_ri_k[k]; 4715 nextrow[k] = buf_ri_k[k] + 1; /* next row number of k-th recved i-structure */ 4716 nextai[k] = buf_ri_k[k] + (nrows + 1); /* poins to the next i-structure of k-th recved i-structure */ 4717 } 4718 4719 ierr = MatPreallocateInitialize(comm,m,n,dnz,onz);CHKERRQ(ierr); 4720 len = 0; 4721 for (i=0; i<m; i++) { 4722 bnzi = 0; 4723 /* add local non-zero cols of this proc's seqmat into lnk */ 4724 arow = owners[rank] + i; 4725 anzi = ai[arow+1] - ai[arow]; 4726 aj = a->j + ai[arow]; 4727 ierr = PetscLLAddSorted(anzi,aj,N,nlnk,lnk,lnkbt);CHKERRQ(ierr); 4728 bnzi += nlnk; 4729 /* add received col data into lnk */ 4730 for (k=0; k<merge->nrecv; k++) { /* k-th received message */ 4731 if (i == *nextrow[k]) { /* i-th row */ 4732 anzi = *(nextai[k]+1) - *nextai[k]; 4733 aj = buf_rj[k] + *nextai[k]; 4734 ierr = PetscLLAddSorted(anzi,aj,N,nlnk,lnk,lnkbt);CHKERRQ(ierr); 4735 bnzi += nlnk; 4736 nextrow[k]++; nextai[k]++; 4737 } 4738 } 4739 if (len < bnzi) len = bnzi; /* =max(bnzi) */ 4740 4741 /* if free space is not available, make more free space */ 4742 if (current_space->local_remaining<bnzi) { 4743 ierr = PetscFreeSpaceGet(PetscIntSumTruncate(bnzi,current_space->total_array_size),¤t_space);CHKERRQ(ierr); 4744 nspacedouble++; 4745 } 4746 /* copy data into free space, then initialize lnk */ 4747 ierr = PetscLLClean(N,N,bnzi,lnk,current_space->array,lnkbt);CHKERRQ(ierr); 4748 ierr = MatPreallocateSet(i+owners[rank],bnzi,current_space->array,dnz,onz);CHKERRQ(ierr); 4749 4750 current_space->array += bnzi; 4751 current_space->local_used += bnzi; 4752 current_space->local_remaining -= bnzi; 4753 4754 bi[i+1] = bi[i] + bnzi; 4755 } 4756 4757 ierr = PetscFree3(buf_ri_k,nextrow,nextai);CHKERRQ(ierr); 4758 4759 ierr = PetscMalloc1(bi[m]+1,&bj);CHKERRQ(ierr); 4760 ierr = PetscFreeSpaceContiguous(&free_space,bj);CHKERRQ(ierr); 4761 ierr = PetscLLDestroy(lnk,lnkbt);CHKERRQ(ierr); 4762 4763 /* create symbolic parallel matrix B_mpi */ 4764 /*---------------------------------------*/ 4765 ierr = MatGetBlockSizes(seqmat,&bs,&cbs);CHKERRQ(ierr); 4766 ierr = MatCreate(comm,&B_mpi);CHKERRQ(ierr); 4767 if (n==PETSC_DECIDE) { 4768 ierr = MatSetSizes(B_mpi,m,n,PETSC_DETERMINE,N);CHKERRQ(ierr); 4769 } else { 4770 ierr = MatSetSizes(B_mpi,m,n,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 4771 } 4772 ierr = MatSetBlockSizes(B_mpi,bs,cbs);CHKERRQ(ierr); 4773 ierr = MatSetType(B_mpi,MATMPIAIJ);CHKERRQ(ierr); 4774 ierr = MatMPIAIJSetPreallocation(B_mpi,0,dnz,0,onz);CHKERRQ(ierr); 4775 ierr = MatPreallocateFinalize(dnz,onz);CHKERRQ(ierr); 4776 ierr = MatSetOption(B_mpi,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 4777 4778 /* B_mpi is not ready for use - assembly will be done by MatCreateMPIAIJSumSeqAIJNumeric() */ 4779 B_mpi->assembled = PETSC_FALSE; 4780 B_mpi->ops->destroy = MatDestroy_MPIAIJ_SeqsToMPI; 4781 merge->bi = bi; 4782 merge->bj = bj; 4783 merge->buf_ri = buf_ri; 4784 merge->buf_rj = buf_rj; 4785 merge->coi = NULL; 4786 merge->coj = NULL; 4787 merge->owners_co = NULL; 4788 4789 ierr = PetscCommDestroy(&comm);CHKERRQ(ierr); 4790 4791 /* attach the supporting struct to B_mpi for reuse */ 4792 ierr = PetscContainerCreate(PETSC_COMM_SELF,&container);CHKERRQ(ierr); 4793 ierr = PetscContainerSetPointer(container,merge);CHKERRQ(ierr); 4794 ierr = PetscObjectCompose((PetscObject)B_mpi,"MatMergeSeqsToMPI",(PetscObject)container);CHKERRQ(ierr); 4795 ierr = PetscContainerDestroy(&container);CHKERRQ(ierr); 4796 *mpimat = B_mpi; 4797 4798 ierr = PetscLogEventEnd(MAT_Seqstompisym,seqmat,0,0,0);CHKERRQ(ierr); 4799 PetscFunctionReturn(0); 4800 } 4801 4802 /*@C 4803 MatCreateMPIAIJSumSeqAIJ - Creates a MATMPIAIJ matrix by adding sequential 4804 matrices from each processor 4805 4806 Collective on MPI_Comm 4807 4808 Input Parameters: 4809 + comm - the communicators the parallel matrix will live on 4810 . seqmat - the input sequential matrices 4811 . m - number of local rows (or PETSC_DECIDE) 4812 . n - number of local columns (or PETSC_DECIDE) 4813 - scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX 4814 4815 Output Parameter: 4816 . mpimat - the parallel matrix generated 4817 4818 Level: advanced 4819 4820 Notes: 4821 The dimensions of the sequential matrix in each processor MUST be the same. 4822 The input seqmat is included into the container "Mat_Merge_SeqsToMPI", and will be 4823 destroyed when mpimat is destroyed. Call PetscObjectQuery() to access seqmat. 4824 @*/ 4825 PetscErrorCode MatCreateMPIAIJSumSeqAIJ(MPI_Comm comm,Mat seqmat,PetscInt m,PetscInt n,MatReuse scall,Mat *mpimat) 4826 { 4827 PetscErrorCode ierr; 4828 PetscMPIInt size; 4829 4830 PetscFunctionBegin; 4831 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 4832 if (size == 1) { 4833 ierr = PetscLogEventBegin(MAT_Seqstompi,seqmat,0,0,0);CHKERRQ(ierr); 4834 if (scall == MAT_INITIAL_MATRIX) { 4835 ierr = MatDuplicate(seqmat,MAT_COPY_VALUES,mpimat);CHKERRQ(ierr); 4836 } else { 4837 ierr = MatCopy(seqmat,*mpimat,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 4838 } 4839 ierr = PetscLogEventEnd(MAT_Seqstompi,seqmat,0,0,0);CHKERRQ(ierr); 4840 PetscFunctionReturn(0); 4841 } 4842 ierr = PetscLogEventBegin(MAT_Seqstompi,seqmat,0,0,0);CHKERRQ(ierr); 4843 if (scall == MAT_INITIAL_MATRIX) { 4844 ierr = MatCreateMPIAIJSumSeqAIJSymbolic(comm,seqmat,m,n,mpimat);CHKERRQ(ierr); 4845 } 4846 ierr = MatCreateMPIAIJSumSeqAIJNumeric(seqmat,*mpimat);CHKERRQ(ierr); 4847 ierr = PetscLogEventEnd(MAT_Seqstompi,seqmat,0,0,0);CHKERRQ(ierr); 4848 PetscFunctionReturn(0); 4849 } 4850 4851 /*@ 4852 MatMPIAIJGetLocalMat - Creates a SeqAIJ from a MATMPIAIJ matrix by taking all its local rows and putting them into a sequential matrix with 4853 mlocal rows and n columns. Where mlocal is the row count obtained with MatGetLocalSize() and n is the global column count obtained 4854 with MatGetSize() 4855 4856 Not Collective 4857 4858 Input Parameters: 4859 + A - the matrix 4860 . scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX 4861 4862 Output Parameter: 4863 . A_loc - the local sequential matrix generated 4864 4865 Level: developer 4866 4867 .seealso: MatGetOwnerShipRange(), MatMPIAIJGetLocalMatCondensed() 4868 4869 @*/ 4870 PetscErrorCode MatMPIAIJGetLocalMat(Mat A,MatReuse scall,Mat *A_loc) 4871 { 4872 PetscErrorCode ierr; 4873 Mat_MPIAIJ *mpimat=(Mat_MPIAIJ*)A->data; 4874 Mat_SeqAIJ *mat,*a,*b; 4875 PetscInt *ai,*aj,*bi,*bj,*cmap=mpimat->garray; 4876 MatScalar *aa,*ba,*cam; 4877 PetscScalar *ca; 4878 PetscInt am=A->rmap->n,i,j,k,cstart=A->cmap->rstart; 4879 PetscInt *ci,*cj,col,ncols_d,ncols_o,jo; 4880 PetscBool match; 4881 MPI_Comm comm; 4882 PetscMPIInt size; 4883 4884 PetscFunctionBegin; 4885 ierr = PetscObjectTypeCompare((PetscObject)A,MATMPIAIJ,&match);CHKERRQ(ierr); 4886 if (!match) SETERRQ(PetscObjectComm((PetscObject)A), PETSC_ERR_SUP,"Requires MATMPIAIJ matrix as input"); 4887 ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 4888 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 4889 if (size == 1 && scall == MAT_REUSE_MATRIX) PetscFunctionReturn(0); 4890 4891 ierr = PetscLogEventBegin(MAT_Getlocalmat,A,0,0,0);CHKERRQ(ierr); 4892 a = (Mat_SeqAIJ*)(mpimat->A)->data; 4893 b = (Mat_SeqAIJ*)(mpimat->B)->data; 4894 ai = a->i; aj = a->j; bi = b->i; bj = b->j; 4895 aa = a->a; ba = b->a; 4896 if (scall == MAT_INITIAL_MATRIX) { 4897 if (size == 1) { 4898 ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,am,A->cmap->N,ai,aj,aa,A_loc);CHKERRQ(ierr); 4899 PetscFunctionReturn(0); 4900 } 4901 4902 ierr = PetscMalloc1(1+am,&ci);CHKERRQ(ierr); 4903 ci[0] = 0; 4904 for (i=0; i<am; i++) { 4905 ci[i+1] = ci[i] + (ai[i+1] - ai[i]) + (bi[i+1] - bi[i]); 4906 } 4907 ierr = PetscMalloc1(1+ci[am],&cj);CHKERRQ(ierr); 4908 ierr = PetscMalloc1(1+ci[am],&ca);CHKERRQ(ierr); 4909 k = 0; 4910 for (i=0; i<am; i++) { 4911 ncols_o = bi[i+1] - bi[i]; 4912 ncols_d = ai[i+1] - ai[i]; 4913 /* off-diagonal portion of A */ 4914 for (jo=0; jo<ncols_o; jo++) { 4915 col = cmap[*bj]; 4916 if (col >= cstart) break; 4917 cj[k] = col; bj++; 4918 ca[k++] = *ba++; 4919 } 4920 /* diagonal portion of A */ 4921 for (j=0; j<ncols_d; j++) { 4922 cj[k] = cstart + *aj++; 4923 ca[k++] = *aa++; 4924 } 4925 /* off-diagonal portion of A */ 4926 for (j=jo; j<ncols_o; j++) { 4927 cj[k] = cmap[*bj++]; 4928 ca[k++] = *ba++; 4929 } 4930 } 4931 /* put together the new matrix */ 4932 ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,am,A->cmap->N,ci,cj,ca,A_loc);CHKERRQ(ierr); 4933 /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */ 4934 /* Since these are PETSc arrays, change flags to free them as necessary. */ 4935 mat = (Mat_SeqAIJ*)(*A_loc)->data; 4936 mat->free_a = PETSC_TRUE; 4937 mat->free_ij = PETSC_TRUE; 4938 mat->nonew = 0; 4939 } else if (scall == MAT_REUSE_MATRIX) { 4940 mat=(Mat_SeqAIJ*)(*A_loc)->data; 4941 ci = mat->i; cj = mat->j; cam = mat->a; 4942 for (i=0; i<am; i++) { 4943 /* off-diagonal portion of A */ 4944 ncols_o = bi[i+1] - bi[i]; 4945 for (jo=0; jo<ncols_o; jo++) { 4946 col = cmap[*bj]; 4947 if (col >= cstart) break; 4948 *cam++ = *ba++; bj++; 4949 } 4950 /* diagonal portion of A */ 4951 ncols_d = ai[i+1] - ai[i]; 4952 for (j=0; j<ncols_d; j++) *cam++ = *aa++; 4953 /* off-diagonal portion of A */ 4954 for (j=jo; j<ncols_o; j++) { 4955 *cam++ = *ba++; bj++; 4956 } 4957 } 4958 } else SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Invalid MatReuse %d",(int)scall); 4959 ierr = PetscLogEventEnd(MAT_Getlocalmat,A,0,0,0);CHKERRQ(ierr); 4960 PetscFunctionReturn(0); 4961 } 4962 4963 /*@C 4964 MatMPIAIJGetLocalMatCondensed - Creates a SeqAIJ matrix from an MATMPIAIJ matrix by taking all its local rows and NON-ZERO columns 4965 4966 Not Collective 4967 4968 Input Parameters: 4969 + A - the matrix 4970 . scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX 4971 - row, col - index sets of rows and columns to extract (or NULL) 4972 4973 Output Parameter: 4974 . A_loc - the local sequential matrix generated 4975 4976 Level: developer 4977 4978 .seealso: MatGetOwnershipRange(), MatMPIAIJGetLocalMat() 4979 4980 @*/ 4981 PetscErrorCode MatMPIAIJGetLocalMatCondensed(Mat A,MatReuse scall,IS *row,IS *col,Mat *A_loc) 4982 { 4983 Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data; 4984 PetscErrorCode ierr; 4985 PetscInt i,start,end,ncols,nzA,nzB,*cmap,imark,*idx; 4986 IS isrowa,iscola; 4987 Mat *aloc; 4988 PetscBool match; 4989 4990 PetscFunctionBegin; 4991 ierr = PetscObjectTypeCompare((PetscObject)A,MATMPIAIJ,&match);CHKERRQ(ierr); 4992 if (!match) SETERRQ(PetscObjectComm((PetscObject)A), PETSC_ERR_SUP,"Requires MATMPIAIJ matrix as input"); 4993 ierr = PetscLogEventBegin(MAT_Getlocalmatcondensed,A,0,0,0);CHKERRQ(ierr); 4994 if (!row) { 4995 start = A->rmap->rstart; end = A->rmap->rend; 4996 ierr = ISCreateStride(PETSC_COMM_SELF,end-start,start,1,&isrowa);CHKERRQ(ierr); 4997 } else { 4998 isrowa = *row; 4999 } 5000 if (!col) { 5001 start = A->cmap->rstart; 5002 cmap = a->garray; 5003 nzA = a->A->cmap->n; 5004 nzB = a->B->cmap->n; 5005 ierr = PetscMalloc1(nzA+nzB, &idx);CHKERRQ(ierr); 5006 ncols = 0; 5007 for (i=0; i<nzB; i++) { 5008 if (cmap[i] < start) idx[ncols++] = cmap[i]; 5009 else break; 5010 } 5011 imark = i; 5012 for (i=0; i<nzA; i++) idx[ncols++] = start + i; 5013 for (i=imark; i<nzB; i++) idx[ncols++] = cmap[i]; 5014 ierr = ISCreateGeneral(PETSC_COMM_SELF,ncols,idx,PETSC_OWN_POINTER,&iscola);CHKERRQ(ierr); 5015 } else { 5016 iscola = *col; 5017 } 5018 if (scall != MAT_INITIAL_MATRIX) { 5019 ierr = PetscMalloc1(1,&aloc);CHKERRQ(ierr); 5020 aloc[0] = *A_loc; 5021 } 5022 ierr = MatCreateSubMatrices(A,1,&isrowa,&iscola,scall,&aloc);CHKERRQ(ierr); 5023 *A_loc = aloc[0]; 5024 ierr = PetscFree(aloc);CHKERRQ(ierr); 5025 if (!row) { 5026 ierr = ISDestroy(&isrowa);CHKERRQ(ierr); 5027 } 5028 if (!col) { 5029 ierr = ISDestroy(&iscola);CHKERRQ(ierr); 5030 } 5031 ierr = PetscLogEventEnd(MAT_Getlocalmatcondensed,A,0,0,0);CHKERRQ(ierr); 5032 PetscFunctionReturn(0); 5033 } 5034 5035 /*@C 5036 MatGetBrowsOfAcols - Creates a SeqAIJ matrix by taking rows of B that equal to nonzero columns of local A 5037 5038 Collective on Mat 5039 5040 Input Parameters: 5041 + A,B - the matrices in mpiaij format 5042 . scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX 5043 - rowb, colb - index sets of rows and columns of B to extract (or NULL) 5044 5045 Output Parameter: 5046 + rowb, colb - index sets of rows and columns of B to extract 5047 - B_seq - the sequential matrix generated 5048 5049 Level: developer 5050 5051 @*/ 5052 PetscErrorCode MatGetBrowsOfAcols(Mat A,Mat B,MatReuse scall,IS *rowb,IS *colb,Mat *B_seq) 5053 { 5054 Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data; 5055 PetscErrorCode ierr; 5056 PetscInt *idx,i,start,ncols,nzA,nzB,*cmap,imark; 5057 IS isrowb,iscolb; 5058 Mat *bseq=NULL; 5059 5060 PetscFunctionBegin; 5061 if (A->cmap->rstart != B->rmap->rstart || A->cmap->rend != B->rmap->rend) { 5062 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); 5063 } 5064 ierr = PetscLogEventBegin(MAT_GetBrowsOfAcols,A,B,0,0);CHKERRQ(ierr); 5065 5066 if (scall == MAT_INITIAL_MATRIX) { 5067 start = A->cmap->rstart; 5068 cmap = a->garray; 5069 nzA = a->A->cmap->n; 5070 nzB = a->B->cmap->n; 5071 ierr = PetscMalloc1(nzA+nzB, &idx);CHKERRQ(ierr); 5072 ncols = 0; 5073 for (i=0; i<nzB; i++) { /* row < local row index */ 5074 if (cmap[i] < start) idx[ncols++] = cmap[i]; 5075 else break; 5076 } 5077 imark = i; 5078 for (i=0; i<nzA; i++) idx[ncols++] = start + i; /* local rows */ 5079 for (i=imark; i<nzB; i++) idx[ncols++] = cmap[i]; /* row > local row index */ 5080 ierr = ISCreateGeneral(PETSC_COMM_SELF,ncols,idx,PETSC_OWN_POINTER,&isrowb);CHKERRQ(ierr); 5081 ierr = ISCreateStride(PETSC_COMM_SELF,B->cmap->N,0,1,&iscolb);CHKERRQ(ierr); 5082 } else { 5083 if (!rowb || !colb) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"IS rowb and colb must be provided for MAT_REUSE_MATRIX"); 5084 isrowb = *rowb; iscolb = *colb; 5085 ierr = PetscMalloc1(1,&bseq);CHKERRQ(ierr); 5086 bseq[0] = *B_seq; 5087 } 5088 ierr = MatCreateSubMatrices(B,1,&isrowb,&iscolb,scall,&bseq);CHKERRQ(ierr); 5089 *B_seq = bseq[0]; 5090 ierr = PetscFree(bseq);CHKERRQ(ierr); 5091 if (!rowb) { 5092 ierr = ISDestroy(&isrowb);CHKERRQ(ierr); 5093 } else { 5094 *rowb = isrowb; 5095 } 5096 if (!colb) { 5097 ierr = ISDestroy(&iscolb);CHKERRQ(ierr); 5098 } else { 5099 *colb = iscolb; 5100 } 5101 ierr = PetscLogEventEnd(MAT_GetBrowsOfAcols,A,B,0,0);CHKERRQ(ierr); 5102 PetscFunctionReturn(0); 5103 } 5104 5105 /* 5106 MatGetBrowsOfAoCols_MPIAIJ - Creates a SeqAIJ matrix by taking rows of B that equal to nonzero columns 5107 of the OFF-DIAGONAL portion of local A 5108 5109 Collective on Mat 5110 5111 Input Parameters: 5112 + A,B - the matrices in mpiaij format 5113 - scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX 5114 5115 Output Parameter: 5116 + startsj_s - starting point in B's sending j-arrays, saved for MAT_REUSE (or NULL) 5117 . startsj_r - starting point in B's receiving j-arrays, saved for MAT_REUSE (or NULL) 5118 . bufa_ptr - array for sending matrix values, saved for MAT_REUSE (or NULL) 5119 - B_oth - the sequential matrix generated with size aBn=a->B->cmap->n by B->cmap->N 5120 5121 Level: developer 5122 5123 */ 5124 PetscErrorCode MatGetBrowsOfAoCols_MPIAIJ(Mat A,Mat B,MatReuse scall,PetscInt **startsj_s,PetscInt **startsj_r,MatScalar **bufa_ptr,Mat *B_oth) 5125 { 5126 VecScatter_MPI_General *gen_to,*gen_from; 5127 PetscErrorCode ierr; 5128 Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data; 5129 Mat_SeqAIJ *b_oth; 5130 VecScatter ctx; 5131 MPI_Comm comm; 5132 PetscMPIInt *rprocs,*sprocs,tag,rank; 5133 PetscInt *rowlen,*bufj,*bufJ,ncols,aBn=a->B->cmap->n,row,*b_othi,*b_othj; 5134 PetscInt *rvalues,*svalues; 5135 MatScalar *b_otha,*bufa,*bufA; 5136 PetscInt i,j,k,l,ll,nrecvs,nsends,nrows,*srow,*rstarts,*rstartsj = 0,*sstarts,*sstartsj,len; 5137 MPI_Request *rwaits = NULL,*swaits = NULL; 5138 MPI_Status *sstatus,rstatus; 5139 PetscMPIInt jj,size; 5140 PetscInt *cols,sbs,rbs; 5141 PetscScalar *vals; 5142 5143 PetscFunctionBegin; 5144 ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 5145 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 5146 5147 if (A->cmap->rstart != B->rmap->rstart || A->cmap->rend != B->rmap->rend) { 5148 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); 5149 } 5150 ierr = PetscLogEventBegin(MAT_GetBrowsOfAocols,A,B,0,0);CHKERRQ(ierr); 5151 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 5152 5153 if (size == 1) { 5154 startsj_s = NULL; 5155 bufa_ptr = NULL; 5156 *B_oth = NULL; 5157 PetscFunctionReturn(0); 5158 } 5159 5160 if (!a->Mvctx_mpi1) { /* create a->Mvctx_mpi1 to be used for Mat-Mat ops */ 5161 a->Mvctx_mpi1_flg = PETSC_TRUE; 5162 ierr = MatSetUpMultiply_MPIAIJ(A);CHKERRQ(ierr); 5163 } 5164 ctx = a->Mvctx_mpi1; 5165 tag = ((PetscObject)ctx)->tag; 5166 5167 gen_to = (VecScatter_MPI_General*)ctx->todata; 5168 gen_from = (VecScatter_MPI_General*)ctx->fromdata; 5169 nrecvs = gen_from->n; 5170 nsends = gen_to->n; 5171 5172 ierr = PetscMalloc2(nrecvs,&rwaits,nsends,&swaits);CHKERRQ(ierr); 5173 srow = gen_to->indices; /* local row index to be sent */ 5174 sstarts = gen_to->starts; 5175 sprocs = gen_to->procs; 5176 sstatus = gen_to->sstatus; 5177 sbs = gen_to->bs; 5178 rstarts = gen_from->starts; 5179 rprocs = gen_from->procs; 5180 rbs = gen_from->bs; 5181 5182 if (!startsj_s || !bufa_ptr) scall = MAT_INITIAL_MATRIX; 5183 if (scall == MAT_INITIAL_MATRIX) { 5184 /* i-array */ 5185 /*---------*/ 5186 /* post receives */ 5187 ierr = PetscMalloc1(rbs*(rstarts[nrecvs] - rstarts[0]),&rvalues);CHKERRQ(ierr); 5188 for (i=0; i<nrecvs; i++) { 5189 rowlen = rvalues + rstarts[i]*rbs; 5190 nrows = (rstarts[i+1]-rstarts[i])*rbs; /* num of indices to be received */ 5191 ierr = MPI_Irecv(rowlen,nrows,MPIU_INT,rprocs[i],tag,comm,rwaits+i);CHKERRQ(ierr); 5192 } 5193 5194 /* pack the outgoing message */ 5195 ierr = PetscMalloc2(nsends+1,&sstartsj,nrecvs+1,&rstartsj);CHKERRQ(ierr); 5196 5197 sstartsj[0] = 0; 5198 rstartsj[0] = 0; 5199 len = 0; /* total length of j or a array to be sent */ 5200 k = 0; 5201 ierr = PetscMalloc1(sbs*(sstarts[nsends] - sstarts[0]),&svalues);CHKERRQ(ierr); 5202 for (i=0; i<nsends; i++) { 5203 rowlen = svalues + sstarts[i]*sbs; 5204 nrows = sstarts[i+1]-sstarts[i]; /* num of block rows */ 5205 for (j=0; j<nrows; j++) { 5206 row = srow[k] + B->rmap->range[rank]; /* global row idx */ 5207 for (l=0; l<sbs; l++) { 5208 ierr = MatGetRow_MPIAIJ(B,row+l,&ncols,NULL,NULL);CHKERRQ(ierr); /* rowlength */ 5209 5210 rowlen[j*sbs+l] = ncols; 5211 5212 len += ncols; 5213 ierr = MatRestoreRow_MPIAIJ(B,row+l,&ncols,NULL,NULL);CHKERRQ(ierr); 5214 } 5215 k++; 5216 } 5217 ierr = MPI_Isend(rowlen,nrows*sbs,MPIU_INT,sprocs[i],tag,comm,swaits+i);CHKERRQ(ierr); 5218 5219 sstartsj[i+1] = len; /* starting point of (i+1)-th outgoing msg in bufj and bufa */ 5220 } 5221 /* recvs and sends of i-array are completed */ 5222 i = nrecvs; 5223 while (i--) { 5224 ierr = MPI_Waitany(nrecvs,rwaits,&jj,&rstatus);CHKERRQ(ierr); 5225 } 5226 if (nsends) {ierr = MPI_Waitall(nsends,swaits,sstatus);CHKERRQ(ierr);} 5227 ierr = PetscFree(svalues);CHKERRQ(ierr); 5228 5229 /* allocate buffers for sending j and a arrays */ 5230 ierr = PetscMalloc1(len+1,&bufj);CHKERRQ(ierr); 5231 ierr = PetscMalloc1(len+1,&bufa);CHKERRQ(ierr); 5232 5233 /* create i-array of B_oth */ 5234 ierr = PetscMalloc1(aBn+2,&b_othi);CHKERRQ(ierr); 5235 5236 b_othi[0] = 0; 5237 len = 0; /* total length of j or a array to be received */ 5238 k = 0; 5239 for (i=0; i<nrecvs; i++) { 5240 rowlen = rvalues + rstarts[i]*rbs; 5241 nrows = rbs*(rstarts[i+1]-rstarts[i]); /* num of rows to be received */ 5242 for (j=0; j<nrows; j++) { 5243 b_othi[k+1] = b_othi[k] + rowlen[j]; 5244 ierr = PetscIntSumError(rowlen[j],len,&len);CHKERRQ(ierr); 5245 k++; 5246 } 5247 rstartsj[i+1] = len; /* starting point of (i+1)-th incoming msg in bufj and bufa */ 5248 } 5249 ierr = PetscFree(rvalues);CHKERRQ(ierr); 5250 5251 /* allocate space for j and a arrrays of B_oth */ 5252 ierr = PetscMalloc1(b_othi[aBn]+1,&b_othj);CHKERRQ(ierr); 5253 ierr = PetscMalloc1(b_othi[aBn]+1,&b_otha);CHKERRQ(ierr); 5254 5255 /* j-array */ 5256 /*---------*/ 5257 /* post receives of j-array */ 5258 for (i=0; i<nrecvs; i++) { 5259 nrows = rstartsj[i+1]-rstartsj[i]; /* length of the msg received */ 5260 ierr = MPI_Irecv(b_othj+rstartsj[i],nrows,MPIU_INT,rprocs[i],tag,comm,rwaits+i);CHKERRQ(ierr); 5261 } 5262 5263 /* pack the outgoing message j-array */ 5264 k = 0; 5265 for (i=0; i<nsends; i++) { 5266 nrows = sstarts[i+1]-sstarts[i]; /* num of block rows */ 5267 bufJ = bufj+sstartsj[i]; 5268 for (j=0; j<nrows; j++) { 5269 row = srow[k++] + B->rmap->range[rank]; /* global row idx */ 5270 for (ll=0; ll<sbs; ll++) { 5271 ierr = MatGetRow_MPIAIJ(B,row+ll,&ncols,&cols,NULL);CHKERRQ(ierr); 5272 for (l=0; l<ncols; l++) { 5273 *bufJ++ = cols[l]; 5274 } 5275 ierr = MatRestoreRow_MPIAIJ(B,row+ll,&ncols,&cols,NULL);CHKERRQ(ierr); 5276 } 5277 } 5278 ierr = MPI_Isend(bufj+sstartsj[i],sstartsj[i+1]-sstartsj[i],MPIU_INT,sprocs[i],tag,comm,swaits+i);CHKERRQ(ierr); 5279 } 5280 5281 /* recvs and sends of j-array are completed */ 5282 i = nrecvs; 5283 while (i--) { 5284 ierr = MPI_Waitany(nrecvs,rwaits,&jj,&rstatus);CHKERRQ(ierr); 5285 } 5286 if (nsends) {ierr = MPI_Waitall(nsends,swaits,sstatus);CHKERRQ(ierr);} 5287 } else if (scall == MAT_REUSE_MATRIX) { 5288 sstartsj = *startsj_s; 5289 rstartsj = *startsj_r; 5290 bufa = *bufa_ptr; 5291 b_oth = (Mat_SeqAIJ*)(*B_oth)->data; 5292 b_otha = b_oth->a; 5293 } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE, "Matrix P does not posses an object container"); 5294 5295 /* a-array */ 5296 /*---------*/ 5297 /* post receives of a-array */ 5298 for (i=0; i<nrecvs; i++) { 5299 nrows = rstartsj[i+1]-rstartsj[i]; /* length of the msg received */ 5300 ierr = MPI_Irecv(b_otha+rstartsj[i],nrows,MPIU_SCALAR,rprocs[i],tag,comm,rwaits+i);CHKERRQ(ierr); 5301 } 5302 5303 /* pack the outgoing message a-array */ 5304 k = 0; 5305 for (i=0; i<nsends; i++) { 5306 nrows = sstarts[i+1]-sstarts[i]; /* num of block rows */ 5307 bufA = bufa+sstartsj[i]; 5308 for (j=0; j<nrows; j++) { 5309 row = srow[k++] + B->rmap->range[rank]; /* global row idx */ 5310 for (ll=0; ll<sbs; ll++) { 5311 ierr = MatGetRow_MPIAIJ(B,row+ll,&ncols,NULL,&vals);CHKERRQ(ierr); 5312 for (l=0; l<ncols; l++) { 5313 *bufA++ = vals[l]; 5314 } 5315 ierr = MatRestoreRow_MPIAIJ(B,row+ll,&ncols,NULL,&vals);CHKERRQ(ierr); 5316 } 5317 } 5318 ierr = MPI_Isend(bufa+sstartsj[i],sstartsj[i+1]-sstartsj[i],MPIU_SCALAR,sprocs[i],tag,comm,swaits+i);CHKERRQ(ierr); 5319 } 5320 /* recvs and sends of a-array are completed */ 5321 i = nrecvs; 5322 while (i--) { 5323 ierr = MPI_Waitany(nrecvs,rwaits,&jj,&rstatus);CHKERRQ(ierr); 5324 } 5325 if (nsends) {ierr = MPI_Waitall(nsends,swaits,sstatus);CHKERRQ(ierr);} 5326 ierr = PetscFree2(rwaits,swaits);CHKERRQ(ierr); 5327 5328 if (scall == MAT_INITIAL_MATRIX) { 5329 /* put together the new matrix */ 5330 ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,aBn,B->cmap->N,b_othi,b_othj,b_otha,B_oth);CHKERRQ(ierr); 5331 5332 /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */ 5333 /* Since these are PETSc arrays, change flags to free them as necessary. */ 5334 b_oth = (Mat_SeqAIJ*)(*B_oth)->data; 5335 b_oth->free_a = PETSC_TRUE; 5336 b_oth->free_ij = PETSC_TRUE; 5337 b_oth->nonew = 0; 5338 5339 ierr = PetscFree(bufj);CHKERRQ(ierr); 5340 if (!startsj_s || !bufa_ptr) { 5341 ierr = PetscFree2(sstartsj,rstartsj);CHKERRQ(ierr); 5342 ierr = PetscFree(bufa_ptr);CHKERRQ(ierr); 5343 } else { 5344 *startsj_s = sstartsj; 5345 *startsj_r = rstartsj; 5346 *bufa_ptr = bufa; 5347 } 5348 } 5349 ierr = PetscLogEventEnd(MAT_GetBrowsOfAocols,A,B,0,0);CHKERRQ(ierr); 5350 PetscFunctionReturn(0); 5351 } 5352 5353 /*@C 5354 MatGetCommunicationStructs - Provides access to the communication structures used in matrix-vector multiplication. 5355 5356 Not Collective 5357 5358 Input Parameters: 5359 . A - The matrix in mpiaij format 5360 5361 Output Parameter: 5362 + lvec - The local vector holding off-process values from the argument to a matrix-vector product 5363 . colmap - A map from global column index to local index into lvec 5364 - multScatter - A scatter from the argument of a matrix-vector product to lvec 5365 5366 Level: developer 5367 5368 @*/ 5369 #if defined(PETSC_USE_CTABLE) 5370 PetscErrorCode MatGetCommunicationStructs(Mat A, Vec *lvec, PetscTable *colmap, VecScatter *multScatter) 5371 #else 5372 PetscErrorCode MatGetCommunicationStructs(Mat A, Vec *lvec, PetscInt *colmap[], VecScatter *multScatter) 5373 #endif 5374 { 5375 Mat_MPIAIJ *a; 5376 5377 PetscFunctionBegin; 5378 PetscValidHeaderSpecific(A, MAT_CLASSID, 1); 5379 PetscValidPointer(lvec, 2); 5380 PetscValidPointer(colmap, 3); 5381 PetscValidPointer(multScatter, 4); 5382 a = (Mat_MPIAIJ*) A->data; 5383 if (lvec) *lvec = a->lvec; 5384 if (colmap) *colmap = a->colmap; 5385 if (multScatter) *multScatter = a->Mvctx; 5386 PetscFunctionReturn(0); 5387 } 5388 5389 PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJCRL(Mat,MatType,MatReuse,Mat*); 5390 PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJPERM(Mat,MatType,MatReuse,Mat*); 5391 #if defined(PETSC_HAVE_MKL_SPARSE) 5392 PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJMKL(Mat,MatType,MatReuse,Mat*); 5393 #endif 5394 PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPISBAIJ(Mat,MatType,MatReuse,Mat*); 5395 #if defined(PETSC_HAVE_ELEMENTAL) 5396 PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_Elemental(Mat,MatType,MatReuse,Mat*); 5397 #endif 5398 #if defined(PETSC_HAVE_HYPRE) 5399 PETSC_INTERN PetscErrorCode MatConvert_AIJ_HYPRE(Mat,MatType,MatReuse,Mat*); 5400 PETSC_INTERN PetscErrorCode MatMatMatMult_Transpose_AIJ_AIJ(Mat,Mat,Mat,MatReuse,PetscReal,Mat*); 5401 #endif 5402 PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_IS(Mat,MatType,MatReuse,Mat*); 5403 PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPISELL(Mat,MatType,MatReuse,Mat*); 5404 5405 /* 5406 Computes (B'*A')' since computing B*A directly is untenable 5407 5408 n p p 5409 ( ) ( ) ( ) 5410 m ( A ) * n ( B ) = m ( C ) 5411 ( ) ( ) ( ) 5412 5413 */ 5414 PetscErrorCode MatMatMultNumeric_MPIDense_MPIAIJ(Mat A,Mat B,Mat C) 5415 { 5416 PetscErrorCode ierr; 5417 Mat At,Bt,Ct; 5418 5419 PetscFunctionBegin; 5420 ierr = MatTranspose(A,MAT_INITIAL_MATRIX,&At);CHKERRQ(ierr); 5421 ierr = MatTranspose(B,MAT_INITIAL_MATRIX,&Bt);CHKERRQ(ierr); 5422 ierr = MatMatMult(Bt,At,MAT_INITIAL_MATRIX,1.0,&Ct);CHKERRQ(ierr); 5423 ierr = MatDestroy(&At);CHKERRQ(ierr); 5424 ierr = MatDestroy(&Bt);CHKERRQ(ierr); 5425 ierr = MatTranspose(Ct,MAT_REUSE_MATRIX,&C);CHKERRQ(ierr); 5426 ierr = MatDestroy(&Ct);CHKERRQ(ierr); 5427 PetscFunctionReturn(0); 5428 } 5429 5430 PetscErrorCode MatMatMultSymbolic_MPIDense_MPIAIJ(Mat A,Mat B,PetscReal fill,Mat *C) 5431 { 5432 PetscErrorCode ierr; 5433 PetscInt m=A->rmap->n,n=B->cmap->n; 5434 Mat Cmat; 5435 5436 PetscFunctionBegin; 5437 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); 5438 ierr = MatCreate(PetscObjectComm((PetscObject)A),&Cmat);CHKERRQ(ierr); 5439 ierr = MatSetSizes(Cmat,m,n,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 5440 ierr = MatSetBlockSizesFromMats(Cmat,A,B);CHKERRQ(ierr); 5441 ierr = MatSetType(Cmat,MATMPIDENSE);CHKERRQ(ierr); 5442 ierr = MatMPIDenseSetPreallocation(Cmat,NULL);CHKERRQ(ierr); 5443 ierr = MatAssemblyBegin(Cmat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5444 ierr = MatAssemblyEnd(Cmat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5445 5446 Cmat->ops->matmultnumeric = MatMatMultNumeric_MPIDense_MPIAIJ; 5447 5448 *C = Cmat; 5449 PetscFunctionReturn(0); 5450 } 5451 5452 /* ----------------------------------------------------------------*/ 5453 PETSC_INTERN PetscErrorCode MatMatMult_MPIDense_MPIAIJ(Mat A,Mat B,MatReuse scall,PetscReal fill,Mat *C) 5454 { 5455 PetscErrorCode ierr; 5456 5457 PetscFunctionBegin; 5458 if (scall == MAT_INITIAL_MATRIX) { 5459 ierr = PetscLogEventBegin(MAT_MatMultSymbolic,A,B,0,0);CHKERRQ(ierr); 5460 ierr = MatMatMultSymbolic_MPIDense_MPIAIJ(A,B,fill,C);CHKERRQ(ierr); 5461 ierr = PetscLogEventEnd(MAT_MatMultSymbolic,A,B,0,0);CHKERRQ(ierr); 5462 } 5463 ierr = PetscLogEventBegin(MAT_MatMultNumeric,A,B,0,0);CHKERRQ(ierr); 5464 ierr = MatMatMultNumeric_MPIDense_MPIAIJ(A,B,*C);CHKERRQ(ierr); 5465 ierr = PetscLogEventEnd(MAT_MatMultNumeric,A,B,0,0);CHKERRQ(ierr); 5466 PetscFunctionReturn(0); 5467 } 5468 5469 /*MC 5470 MATMPIAIJ - MATMPIAIJ = "mpiaij" - A matrix type to be used for parallel sparse matrices. 5471 5472 Options Database Keys: 5473 . -mat_type mpiaij - sets the matrix type to "mpiaij" during a call to MatSetFromOptions() 5474 5475 Level: beginner 5476 5477 .seealso: MatCreateAIJ() 5478 M*/ 5479 5480 PETSC_EXTERN PetscErrorCode MatCreate_MPIAIJ(Mat B) 5481 { 5482 Mat_MPIAIJ *b; 5483 PetscErrorCode ierr; 5484 PetscMPIInt size; 5485 5486 PetscFunctionBegin; 5487 ierr = MPI_Comm_size(PetscObjectComm((PetscObject)B),&size);CHKERRQ(ierr); 5488 5489 ierr = PetscNewLog(B,&b);CHKERRQ(ierr); 5490 B->data = (void*)b; 5491 ierr = PetscMemcpy(B->ops,&MatOps_Values,sizeof(struct _MatOps));CHKERRQ(ierr); 5492 B->assembled = PETSC_FALSE; 5493 B->insertmode = NOT_SET_VALUES; 5494 b->size = size; 5495 5496 ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)B),&b->rank);CHKERRQ(ierr); 5497 5498 /* build cache for off array entries formed */ 5499 ierr = MatStashCreate_Private(PetscObjectComm((PetscObject)B),1,&B->stash);CHKERRQ(ierr); 5500 5501 b->donotstash = PETSC_FALSE; 5502 b->colmap = 0; 5503 b->garray = 0; 5504 b->roworiented = PETSC_TRUE; 5505 5506 /* stuff used for matrix vector multiply */ 5507 b->lvec = NULL; 5508 b->Mvctx = NULL; 5509 5510 /* stuff for MatGetRow() */ 5511 b->rowindices = 0; 5512 b->rowvalues = 0; 5513 b->getrowactive = PETSC_FALSE; 5514 5515 /* flexible pointer used in CUSP/CUSPARSE classes */ 5516 b->spptr = NULL; 5517 5518 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMPIAIJSetUseScalableIncreaseOverlap_C",MatMPIAIJSetUseScalableIncreaseOverlap_MPIAIJ);CHKERRQ(ierr); 5519 ierr = PetscObjectComposeFunction((PetscObject)B,"MatStoreValues_C",MatStoreValues_MPIAIJ);CHKERRQ(ierr); 5520 ierr = PetscObjectComposeFunction((PetscObject)B,"MatRetrieveValues_C",MatRetrieveValues_MPIAIJ);CHKERRQ(ierr); 5521 ierr = PetscObjectComposeFunction((PetscObject)B,"MatIsTranspose_C",MatIsTranspose_MPIAIJ);CHKERRQ(ierr); 5522 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMPIAIJSetPreallocation_C",MatMPIAIJSetPreallocation_MPIAIJ);CHKERRQ(ierr); 5523 ierr = PetscObjectComposeFunction((PetscObject)B,"MatResetPreallocation_C",MatResetPreallocation_MPIAIJ);CHKERRQ(ierr); 5524 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMPIAIJSetPreallocationCSR_C",MatMPIAIJSetPreallocationCSR_MPIAIJ);CHKERRQ(ierr); 5525 ierr = PetscObjectComposeFunction((PetscObject)B,"MatDiagonalScaleLocal_C",MatDiagonalScaleLocal_MPIAIJ);CHKERRQ(ierr); 5526 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_mpiaijperm_C",MatConvert_MPIAIJ_MPIAIJPERM);CHKERRQ(ierr); 5527 #if defined(PETSC_HAVE_MKL_SPARSE) 5528 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_mpiaijmkl_C",MatConvert_MPIAIJ_MPIAIJMKL);CHKERRQ(ierr); 5529 #endif 5530 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_mpiaijcrl_C",MatConvert_MPIAIJ_MPIAIJCRL);CHKERRQ(ierr); 5531 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_mpisbaij_C",MatConvert_MPIAIJ_MPISBAIJ);CHKERRQ(ierr); 5532 #if defined(PETSC_HAVE_ELEMENTAL) 5533 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_elemental_C",MatConvert_MPIAIJ_Elemental);CHKERRQ(ierr); 5534 #endif 5535 #if defined(PETSC_HAVE_HYPRE) 5536 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_hypre_C",MatConvert_AIJ_HYPRE);CHKERRQ(ierr); 5537 #endif 5538 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_is_C",MatConvert_MPIAIJ_IS);CHKERRQ(ierr); 5539 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_mpisell_C",MatConvert_MPIAIJ_MPISELL);CHKERRQ(ierr); 5540 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMatMult_mpidense_mpiaij_C",MatMatMult_MPIDense_MPIAIJ);CHKERRQ(ierr); 5541 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMatMultSymbolic_mpidense_mpiaij_C",MatMatMultSymbolic_MPIDense_MPIAIJ);CHKERRQ(ierr); 5542 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMatMultNumeric_mpidense_mpiaij_C",MatMatMultNumeric_MPIDense_MPIAIJ);CHKERRQ(ierr); 5543 #if defined(PETSC_HAVE_HYPRE) 5544 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMatMatMult_transpose_mpiaij_mpiaij_C",MatMatMatMult_Transpose_AIJ_AIJ);CHKERRQ(ierr); 5545 #endif 5546 ierr = PetscObjectChangeTypeName((PetscObject)B,MATMPIAIJ);CHKERRQ(ierr); 5547 PetscFunctionReturn(0); 5548 } 5549 5550 /*@C 5551 MatCreateMPIAIJWithSplitArrays - creates a MPI AIJ matrix using arrays that contain the "diagonal" 5552 and "off-diagonal" part of the matrix in CSR format. 5553 5554 Collective on MPI_Comm 5555 5556 Input Parameters: 5557 + comm - MPI communicator 5558 . m - number of local rows (Cannot be PETSC_DECIDE) 5559 . n - This value should be the same as the local size used in creating the 5560 x vector for the matrix-vector product y = Ax. (or PETSC_DECIDE to have 5561 calculated if N is given) For square matrices n is almost always m. 5562 . M - number of global rows (or PETSC_DETERMINE to have calculated if m is given) 5563 . N - number of global columns (or PETSC_DETERMINE to have calculated if n is given) 5564 . i - row indices for "diagonal" portion of matrix 5565 . j - column indices 5566 . a - matrix values 5567 . oi - row indices for "off-diagonal" portion of matrix 5568 . oj - column indices 5569 - oa - matrix values 5570 5571 Output Parameter: 5572 . mat - the matrix 5573 5574 Level: advanced 5575 5576 Notes: 5577 The i, j, and a arrays ARE NOT copied by this routine into the internal format used by PETSc. The user 5578 must free the arrays once the matrix has been destroyed and not before. 5579 5580 The i and j indices are 0 based 5581 5582 See MatCreateAIJ() for the definition of "diagonal" and "off-diagonal" portion of the matrix 5583 5584 This sets local rows and cannot be used to set off-processor values. 5585 5586 Use of this routine is discouraged because it is inflexible and cumbersome to use. It is extremely rare that a 5587 legacy application natively assembles into exactly this split format. The code to do so is nontrivial and does 5588 not easily support in-place reassembly. It is recommended to use MatSetValues() (or a variant thereof) because 5589 the resulting assembly is easier to implement, will work with any matrix format, and the user does not have to 5590 keep track of the underlying array. Use MatSetOption(A,MAT_NO_OFF_PROC_ENTRIES,PETSC_TRUE) to disable all 5591 communication if it is known that only local entries will be set. 5592 5593 .keywords: matrix, aij, compressed row, sparse, parallel 5594 5595 .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatMPIAIJSetPreallocationCSR(), 5596 MATMPIAIJ, MatCreateAIJ(), MatCreateMPIAIJWithArrays() 5597 @*/ 5598 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) 5599 { 5600 PetscErrorCode ierr; 5601 Mat_MPIAIJ *maij; 5602 5603 PetscFunctionBegin; 5604 if (m < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"local number of rows (m) cannot be PETSC_DECIDE, or negative"); 5605 if (i[0]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"i (row indices) must start with 0"); 5606 if (oi[0]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"oi (row indices) must start with 0"); 5607 ierr = MatCreate(comm,mat);CHKERRQ(ierr); 5608 ierr = MatSetSizes(*mat,m,n,M,N);CHKERRQ(ierr); 5609 ierr = MatSetType(*mat,MATMPIAIJ);CHKERRQ(ierr); 5610 maij = (Mat_MPIAIJ*) (*mat)->data; 5611 5612 (*mat)->preallocated = PETSC_TRUE; 5613 5614 ierr = PetscLayoutSetUp((*mat)->rmap);CHKERRQ(ierr); 5615 ierr = PetscLayoutSetUp((*mat)->cmap);CHKERRQ(ierr); 5616 5617 ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,m,n,i,j,a,&maij->A);CHKERRQ(ierr); 5618 ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,m,(*mat)->cmap->N,oi,oj,oa,&maij->B);CHKERRQ(ierr); 5619 5620 ierr = MatAssemblyBegin(maij->A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5621 ierr = MatAssemblyEnd(maij->A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5622 ierr = MatAssemblyBegin(maij->B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5623 ierr = MatAssemblyEnd(maij->B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5624 5625 ierr = MatSetOption(*mat,MAT_NO_OFF_PROC_ENTRIES,PETSC_TRUE);CHKERRQ(ierr); 5626 ierr = MatAssemblyBegin(*mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5627 ierr = MatAssemblyEnd(*mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5628 ierr = MatSetOption(*mat,MAT_NO_OFF_PROC_ENTRIES,PETSC_FALSE);CHKERRQ(ierr); 5629 ierr = MatSetOption(*mat,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE);CHKERRQ(ierr); 5630 PetscFunctionReturn(0); 5631 } 5632 5633 /* 5634 Special version for direct calls from Fortran 5635 */ 5636 #include <petsc/private/fortranimpl.h> 5637 5638 /* Change these macros so can be used in void function */ 5639 #undef CHKERRQ 5640 #define CHKERRQ(ierr) CHKERRABORT(PETSC_COMM_WORLD,ierr) 5641 #undef SETERRQ2 5642 #define SETERRQ2(comm,ierr,b,c,d) CHKERRABORT(comm,ierr) 5643 #undef SETERRQ3 5644 #define SETERRQ3(comm,ierr,b,c,d,e) CHKERRABORT(comm,ierr) 5645 #undef SETERRQ 5646 #define SETERRQ(c,ierr,b) CHKERRABORT(c,ierr) 5647 5648 #if defined(PETSC_HAVE_FORTRAN_CAPS) 5649 #define matsetvaluesmpiaij_ MATSETVALUESMPIAIJ 5650 #elif !defined(PETSC_HAVE_FORTRAN_UNDERSCORE) 5651 #define matsetvaluesmpiaij_ matsetvaluesmpiaij 5652 #else 5653 #endif 5654 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) 5655 { 5656 Mat mat = *mmat; 5657 PetscInt m = *mm, n = *mn; 5658 InsertMode addv = *maddv; 5659 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 5660 PetscScalar value; 5661 PetscErrorCode ierr; 5662 5663 MatCheckPreallocated(mat,1); 5664 if (mat->insertmode == NOT_SET_VALUES) mat->insertmode = addv; 5665 5666 #if defined(PETSC_USE_DEBUG) 5667 else if (mat->insertmode != addv) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Cannot mix add values and insert values"); 5668 #endif 5669 { 5670 PetscInt i,j,rstart = mat->rmap->rstart,rend = mat->rmap->rend; 5671 PetscInt cstart = mat->cmap->rstart,cend = mat->cmap->rend,row,col; 5672 PetscBool roworiented = aij->roworiented; 5673 5674 /* Some Variables required in the macro */ 5675 Mat A = aij->A; 5676 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 5677 PetscInt *aimax = a->imax,*ai = a->i,*ailen = a->ilen,*aj = a->j; 5678 MatScalar *aa = a->a; 5679 PetscBool ignorezeroentries = (((a->ignorezeroentries)&&(addv==ADD_VALUES)) ? PETSC_TRUE : PETSC_FALSE); 5680 Mat B = aij->B; 5681 Mat_SeqAIJ *b = (Mat_SeqAIJ*)B->data; 5682 PetscInt *bimax = b->imax,*bi = b->i,*bilen = b->ilen,*bj = b->j,bm = aij->B->rmap->n,am = aij->A->rmap->n; 5683 MatScalar *ba = b->a; 5684 5685 PetscInt *rp1,*rp2,ii,nrow1,nrow2,_i,rmax1,rmax2,N,low1,high1,low2,high2,t,lastcol1,lastcol2; 5686 PetscInt nonew = a->nonew; 5687 MatScalar *ap1,*ap2; 5688 5689 PetscFunctionBegin; 5690 for (i=0; i<m; i++) { 5691 if (im[i] < 0) continue; 5692 #if defined(PETSC_USE_DEBUG) 5693 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); 5694 #endif 5695 if (im[i] >= rstart && im[i] < rend) { 5696 row = im[i] - rstart; 5697 lastcol1 = -1; 5698 rp1 = aj + ai[row]; 5699 ap1 = aa + ai[row]; 5700 rmax1 = aimax[row]; 5701 nrow1 = ailen[row]; 5702 low1 = 0; 5703 high1 = nrow1; 5704 lastcol2 = -1; 5705 rp2 = bj + bi[row]; 5706 ap2 = ba + bi[row]; 5707 rmax2 = bimax[row]; 5708 nrow2 = bilen[row]; 5709 low2 = 0; 5710 high2 = nrow2; 5711 5712 for (j=0; j<n; j++) { 5713 if (roworiented) value = v[i*n+j]; 5714 else value = v[i+j*m]; 5715 if (in[j] >= cstart && in[j] < cend) { 5716 col = in[j] - cstart; 5717 if (ignorezeroentries && value == 0.0 && (addv == ADD_VALUES) && row != col) continue; 5718 MatSetValues_SeqAIJ_A_Private(row,col,value,addv,im[i],in[j]); 5719 } else if (in[j] < 0) continue; 5720 #if defined(PETSC_USE_DEBUG) 5721 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); 5722 #endif 5723 else { 5724 if (mat->was_assembled) { 5725 if (!aij->colmap) { 5726 ierr = MatCreateColmap_MPIAIJ_Private(mat);CHKERRQ(ierr); 5727 } 5728 #if defined(PETSC_USE_CTABLE) 5729 ierr = PetscTableFind(aij->colmap,in[j]+1,&col);CHKERRQ(ierr); 5730 col--; 5731 #else 5732 col = aij->colmap[in[j]] - 1; 5733 #endif 5734 if (ignorezeroentries && value == 0.0 && (addv == ADD_VALUES) && row != col) continue; 5735 if (col < 0 && !((Mat_SeqAIJ*)(aij->A->data))->nonew) { 5736 ierr = MatDisAssemble_MPIAIJ(mat);CHKERRQ(ierr); 5737 col = in[j]; 5738 /* Reinitialize the variables required by MatSetValues_SeqAIJ_B_Private() */ 5739 B = aij->B; 5740 b = (Mat_SeqAIJ*)B->data; 5741 bimax = b->imax; bi = b->i; bilen = b->ilen; bj = b->j; 5742 rp2 = bj + bi[row]; 5743 ap2 = ba + bi[row]; 5744 rmax2 = bimax[row]; 5745 nrow2 = bilen[row]; 5746 low2 = 0; 5747 high2 = nrow2; 5748 bm = aij->B->rmap->n; 5749 ba = b->a; 5750 } 5751 } else col = in[j]; 5752 MatSetValues_SeqAIJ_B_Private(row,col,value,addv,im[i],in[j]); 5753 } 5754 } 5755 } else if (!aij->donotstash) { 5756 if (roworiented) { 5757 ierr = MatStashValuesRow_Private(&mat->stash,im[i],n,in,v+i*n,(PetscBool)(ignorezeroentries && (addv == ADD_VALUES)));CHKERRQ(ierr); 5758 } else { 5759 ierr = MatStashValuesCol_Private(&mat->stash,im[i],n,in,v+i,m,(PetscBool)(ignorezeroentries && (addv == ADD_VALUES)));CHKERRQ(ierr); 5760 } 5761 } 5762 } 5763 } 5764 PetscFunctionReturnVoid(); 5765 } 5766 5767