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