1 2 #include <../src/mat/impls/aij/mpi/mpiaij.h> /*I "petscmat.h" I*/ 3 #include <petsc-private/vecimpl.h> 4 #include <petsc-private/isimpl.h> 5 #include <petscblaslapack.h> 6 #include <petscsf.h> 7 8 /*MC 9 MATAIJ - MATAIJ = "aij" - A matrix type to be used for sparse matrices. 10 11 This matrix type is identical to MATSEQAIJ when constructed with a single process communicator, 12 and MATMPIAIJ otherwise. As a result, for single process communicators, 13 MatSeqAIJSetPreallocation is supported, and similarly MatMPIAIJSetPreallocation is supported 14 for communicators controlling multiple processes. It is recommended that you call both of 15 the above preallocation routines for simplicity. 16 17 Options Database Keys: 18 . -mat_type aij - sets the matrix type to "aij" during a call to MatSetFromOptions() 19 20 Developer Notes: Subclasses include MATAIJCUSP, MATAIJCUSPARSE, MATAIJPERM, MATAIJCRL, and also automatically switches over to use inodes when 21 enough exist. 22 23 Level: beginner 24 25 .seealso: MatCreateAIJ(), MatCreateSeqAIJ(), MATSEQAIJ,MATMPIAIJ 26 M*/ 27 28 /*MC 29 MATAIJCRL - MATAIJCRL = "aijcrl" - A matrix type to be used for sparse matrices. 30 31 This matrix type is identical to MATSEQAIJCRL when constructed with a single process communicator, 32 and MATMPIAIJCRL otherwise. As a result, for single process communicators, 33 MatSeqAIJSetPreallocation() is supported, and similarly MatMPIAIJSetPreallocation() is supported 34 for communicators controlling multiple processes. It is recommended that you call both of 35 the above preallocation routines for simplicity. 36 37 Options Database Keys: 38 . -mat_type aijcrl - sets the matrix type to "aijcrl" during a call to MatSetFromOptions() 39 40 Level: beginner 41 42 .seealso: MatCreateMPIAIJCRL,MATSEQAIJCRL,MATMPIAIJCRL, MATSEQAIJCRL, MATMPIAIJCRL 43 M*/ 44 45 #undef __FUNCT__ 46 #define __FUNCT__ "MatFindNonzeroRows_MPIAIJ" 47 PetscErrorCode MatFindNonzeroRows_MPIAIJ(Mat M,IS *keptrows) 48 { 49 PetscErrorCode ierr; 50 Mat_MPIAIJ *mat = (Mat_MPIAIJ*)M->data; 51 Mat_SeqAIJ *a = (Mat_SeqAIJ*)mat->A->data; 52 Mat_SeqAIJ *b = (Mat_SeqAIJ*)mat->B->data; 53 const PetscInt *ia,*ib; 54 const MatScalar *aa,*bb; 55 PetscInt na,nb,i,j,*rows,cnt=0,n0rows; 56 PetscInt m = M->rmap->n,rstart = M->rmap->rstart; 57 58 PetscFunctionBegin; 59 *keptrows = 0; 60 ia = a->i; 61 ib = b->i; 62 for (i=0; i<m; i++) { 63 na = ia[i+1] - ia[i]; 64 nb = ib[i+1] - ib[i]; 65 if (!na && !nb) { 66 cnt++; 67 goto ok1; 68 } 69 aa = a->a + ia[i]; 70 for (j=0; j<na; j++) { 71 if (aa[j] != 0.0) goto ok1; 72 } 73 bb = b->a + ib[i]; 74 for (j=0; j <nb; j++) { 75 if (bb[j] != 0.0) goto ok1; 76 } 77 cnt++; 78 ok1:; 79 } 80 ierr = MPI_Allreduce(&cnt,&n0rows,1,MPIU_INT,MPIU_SUM,PetscObjectComm((PetscObject)M));CHKERRQ(ierr); 81 if (!n0rows) PetscFunctionReturn(0); 82 ierr = PetscMalloc1(M->rmap->n-cnt,&rows);CHKERRQ(ierr); 83 cnt = 0; 84 for (i=0; i<m; i++) { 85 na = ia[i+1] - ia[i]; 86 nb = ib[i+1] - ib[i]; 87 if (!na && !nb) continue; 88 aa = a->a + ia[i]; 89 for (j=0; j<na;j++) { 90 if (aa[j] != 0.0) { 91 rows[cnt++] = rstart + i; 92 goto ok2; 93 } 94 } 95 bb = b->a + ib[i]; 96 for (j=0; j<nb; j++) { 97 if (bb[j] != 0.0) { 98 rows[cnt++] = rstart + i; 99 goto ok2; 100 } 101 } 102 ok2:; 103 } 104 ierr = ISCreateGeneral(PetscObjectComm((PetscObject)M),cnt,rows,PETSC_OWN_POINTER,keptrows);CHKERRQ(ierr); 105 PetscFunctionReturn(0); 106 } 107 108 #undef __FUNCT__ 109 #define __FUNCT__ "MatDiagonalSet_MPIAIJ" 110 PetscErrorCode MatDiagonalSet_MPIAIJ(Mat Y,Vec D,InsertMode is) 111 { 112 PetscErrorCode ierr; 113 Mat_MPIAIJ *aij = (Mat_MPIAIJ*) Y->data; 114 115 PetscFunctionBegin; 116 if (Y->assembled && Y->rmap->rstart == Y->cmap->rstart && Y->rmap->rend == Y->cmap->rend) { 117 ierr = MatDiagonalSet(aij->A,D,is);CHKERRQ(ierr); 118 } else { 119 ierr = MatDiagonalSet_Default(Y,D,is);CHKERRQ(ierr); 120 } 121 PetscFunctionReturn(0); 122 } 123 124 125 #undef __FUNCT__ 126 #define __FUNCT__ "MatFindZeroDiagonals_MPIAIJ" 127 PetscErrorCode MatFindZeroDiagonals_MPIAIJ(Mat M,IS *zrows) 128 { 129 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)M->data; 130 PetscErrorCode ierr; 131 PetscInt i,rstart,nrows,*rows; 132 133 PetscFunctionBegin; 134 *zrows = NULL; 135 ierr = MatFindZeroDiagonals_SeqAIJ_Private(aij->A,&nrows,&rows);CHKERRQ(ierr); 136 ierr = MatGetOwnershipRange(M,&rstart,NULL);CHKERRQ(ierr); 137 for (i=0; i<nrows; i++) rows[i] += rstart; 138 ierr = ISCreateGeneral(PetscObjectComm((PetscObject)M),nrows,rows,PETSC_OWN_POINTER,zrows);CHKERRQ(ierr); 139 PetscFunctionReturn(0); 140 } 141 142 #undef __FUNCT__ 143 #define __FUNCT__ "MatGetColumnNorms_MPIAIJ" 144 PetscErrorCode MatGetColumnNorms_MPIAIJ(Mat A,NormType type,PetscReal *norms) 145 { 146 PetscErrorCode ierr; 147 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)A->data; 148 PetscInt i,n,*garray = aij->garray; 149 Mat_SeqAIJ *a_aij = (Mat_SeqAIJ*) aij->A->data; 150 Mat_SeqAIJ *b_aij = (Mat_SeqAIJ*) aij->B->data; 151 PetscReal *work; 152 153 PetscFunctionBegin; 154 ierr = MatGetSize(A,NULL,&n);CHKERRQ(ierr); 155 ierr = PetscCalloc1(n,&work);CHKERRQ(ierr); 156 if (type == NORM_2) { 157 for (i=0; i<a_aij->i[aij->A->rmap->n]; i++) { 158 work[A->cmap->rstart + a_aij->j[i]] += PetscAbsScalar(a_aij->a[i]*a_aij->a[i]); 159 } 160 for (i=0; i<b_aij->i[aij->B->rmap->n]; i++) { 161 work[garray[b_aij->j[i]]] += PetscAbsScalar(b_aij->a[i]*b_aij->a[i]); 162 } 163 } else if (type == NORM_1) { 164 for (i=0; i<a_aij->i[aij->A->rmap->n]; i++) { 165 work[A->cmap->rstart + a_aij->j[i]] += PetscAbsScalar(a_aij->a[i]); 166 } 167 for (i=0; i<b_aij->i[aij->B->rmap->n]; i++) { 168 work[garray[b_aij->j[i]]] += PetscAbsScalar(b_aij->a[i]); 169 } 170 } else if (type == NORM_INFINITY) { 171 for (i=0; i<a_aij->i[aij->A->rmap->n]; i++) { 172 work[A->cmap->rstart + a_aij->j[i]] = PetscMax(PetscAbsScalar(a_aij->a[i]), work[A->cmap->rstart + a_aij->j[i]]); 173 } 174 for (i=0; i<b_aij->i[aij->B->rmap->n]; i++) { 175 work[garray[b_aij->j[i]]] = PetscMax(PetscAbsScalar(b_aij->a[i]),work[garray[b_aij->j[i]]]); 176 } 177 178 } else SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_WRONG,"Unknown NormType"); 179 if (type == NORM_INFINITY) { 180 ierr = MPI_Allreduce(work,norms,n,MPIU_REAL,MPIU_MAX,PetscObjectComm((PetscObject)A));CHKERRQ(ierr); 181 } else { 182 ierr = MPI_Allreduce(work,norms,n,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)A));CHKERRQ(ierr); 183 } 184 ierr = PetscFree(work);CHKERRQ(ierr); 185 if (type == NORM_2) { 186 for (i=0; i<n; i++) norms[i] = PetscSqrtReal(norms[i]); 187 } 188 PetscFunctionReturn(0); 189 } 190 191 #undef __FUNCT__ 192 #define __FUNCT__ "MatFindOffBlockDiagonalEntries_MPIAIJ" 193 PetscErrorCode MatFindOffBlockDiagonalEntries_MPIAIJ(Mat A,IS *is) 194 { 195 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 196 IS sis,gis; 197 PetscErrorCode ierr; 198 const PetscInt *isis,*igis; 199 PetscInt n,*iis,nsis,ngis,rstart,i; 200 201 PetscFunctionBegin; 202 ierr = MatFindOffBlockDiagonalEntries(a->A,&sis);CHKERRQ(ierr); 203 ierr = MatFindNonzeroRows(a->B,&gis);CHKERRQ(ierr); 204 ierr = ISGetSize(gis,&ngis);CHKERRQ(ierr); 205 ierr = ISGetSize(sis,&nsis);CHKERRQ(ierr); 206 ierr = ISGetIndices(sis,&isis);CHKERRQ(ierr); 207 ierr = ISGetIndices(gis,&igis);CHKERRQ(ierr); 208 209 ierr = PetscMalloc1(ngis+nsis,&iis);CHKERRQ(ierr); 210 ierr = PetscMemcpy(iis,igis,ngis*sizeof(PetscInt));CHKERRQ(ierr); 211 ierr = PetscMemcpy(iis+ngis,isis,nsis*sizeof(PetscInt));CHKERRQ(ierr); 212 n = ngis + nsis; 213 ierr = PetscSortRemoveDupsInt(&n,iis);CHKERRQ(ierr); 214 ierr = MatGetOwnershipRange(A,&rstart,NULL);CHKERRQ(ierr); 215 for (i=0; i<n; i++) iis[i] += rstart; 216 ierr = ISCreateGeneral(PetscObjectComm((PetscObject)A),n,iis,PETSC_OWN_POINTER,is);CHKERRQ(ierr); 217 218 ierr = ISRestoreIndices(sis,&isis);CHKERRQ(ierr); 219 ierr = ISRestoreIndices(gis,&igis);CHKERRQ(ierr); 220 ierr = ISDestroy(&sis);CHKERRQ(ierr); 221 ierr = ISDestroy(&gis);CHKERRQ(ierr); 222 PetscFunctionReturn(0); 223 } 224 225 #undef __FUNCT__ 226 #define __FUNCT__ "MatDistribute_MPIAIJ" 227 /* 228 Distributes a SeqAIJ matrix across a set of processes. Code stolen from 229 MatLoad_MPIAIJ(). Horrible lack of reuse. Should be a routine for each matrix type. 230 231 Only for square matrices 232 233 Used by a preconditioner, hence PETSC_EXTERN 234 */ 235 PETSC_EXTERN PetscErrorCode MatDistribute_MPIAIJ(MPI_Comm comm,Mat gmat,PetscInt m,MatReuse reuse,Mat *inmat) 236 { 237 PetscMPIInt rank,size; 238 PetscInt *rowners,*dlens,*olens,i,rstart,rend,j,jj,nz = 0,*gmataj,cnt,row,*ld,bses[2]; 239 PetscErrorCode ierr; 240 Mat mat; 241 Mat_SeqAIJ *gmata; 242 PetscMPIInt tag; 243 MPI_Status status; 244 PetscBool aij; 245 MatScalar *gmataa,*ao,*ad,*gmataarestore=0; 246 247 PetscFunctionBegin; 248 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 249 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 250 if (!rank) { 251 ierr = PetscObjectTypeCompare((PetscObject)gmat,MATSEQAIJ,&aij);CHKERRQ(ierr); 252 if (!aij) SETERRQ1(PetscObjectComm((PetscObject)gmat),PETSC_ERR_SUP,"Currently no support for input matrix of type %s\n",((PetscObject)gmat)->type_name); 253 } 254 if (reuse == MAT_INITIAL_MATRIX) { 255 ierr = MatCreate(comm,&mat);CHKERRQ(ierr); 256 ierr = MatSetSizes(mat,m,m,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 257 ierr = MatGetBlockSizes(gmat,&bses[0],&bses[1]);CHKERRQ(ierr); 258 ierr = MPI_Bcast(bses,2,MPIU_INT,0,comm);CHKERRQ(ierr); 259 ierr = MatSetBlockSizes(mat,bses[0],bses[1]);CHKERRQ(ierr); 260 ierr = MatSetType(mat,MATAIJ);CHKERRQ(ierr); 261 ierr = PetscMalloc1(size+1,&rowners);CHKERRQ(ierr); 262 ierr = PetscMalloc2(m,&dlens,m,&olens);CHKERRQ(ierr); 263 ierr = MPI_Allgather(&m,1,MPIU_INT,rowners+1,1,MPIU_INT,comm);CHKERRQ(ierr); 264 265 rowners[0] = 0; 266 for (i=2; i<=size; i++) rowners[i] += rowners[i-1]; 267 rstart = rowners[rank]; 268 rend = rowners[rank+1]; 269 ierr = PetscObjectGetNewTag((PetscObject)mat,&tag);CHKERRQ(ierr); 270 if (!rank) { 271 gmata = (Mat_SeqAIJ*) gmat->data; 272 /* send row lengths to all processors */ 273 for (i=0; i<m; i++) dlens[i] = gmata->ilen[i]; 274 for (i=1; i<size; i++) { 275 ierr = MPI_Send(gmata->ilen + rowners[i],rowners[i+1]-rowners[i],MPIU_INT,i,tag,comm);CHKERRQ(ierr); 276 } 277 /* determine number diagonal and off-diagonal counts */ 278 ierr = PetscMemzero(olens,m*sizeof(PetscInt));CHKERRQ(ierr); 279 ierr = PetscCalloc1(m,&ld);CHKERRQ(ierr); 280 jj = 0; 281 for (i=0; i<m; i++) { 282 for (j=0; j<dlens[i]; j++) { 283 if (gmata->j[jj] < rstart) ld[i]++; 284 if (gmata->j[jj] < rstart || gmata->j[jj] >= rend) olens[i]++; 285 jj++; 286 } 287 } 288 /* send column indices to other processes */ 289 for (i=1; i<size; i++) { 290 nz = gmata->i[rowners[i+1]]-gmata->i[rowners[i]]; 291 ierr = MPI_Send(&nz,1,MPIU_INT,i,tag,comm);CHKERRQ(ierr); 292 ierr = MPI_Send(gmata->j + gmata->i[rowners[i]],nz,MPIU_INT,i,tag,comm);CHKERRQ(ierr); 293 } 294 295 /* send numerical values to other processes */ 296 for (i=1; i<size; i++) { 297 nz = gmata->i[rowners[i+1]]-gmata->i[rowners[i]]; 298 ierr = MPI_Send(gmata->a + gmata->i[rowners[i]],nz,MPIU_SCALAR,i,tag,comm);CHKERRQ(ierr); 299 } 300 gmataa = gmata->a; 301 gmataj = gmata->j; 302 303 } else { 304 /* receive row lengths */ 305 ierr = MPI_Recv(dlens,m,MPIU_INT,0,tag,comm,&status);CHKERRQ(ierr); 306 /* receive column indices */ 307 ierr = MPI_Recv(&nz,1,MPIU_INT,0,tag,comm,&status);CHKERRQ(ierr); 308 ierr = PetscMalloc2(nz,&gmataa,nz,&gmataj);CHKERRQ(ierr); 309 ierr = MPI_Recv(gmataj,nz,MPIU_INT,0,tag,comm,&status);CHKERRQ(ierr); 310 /* determine number diagonal and off-diagonal counts */ 311 ierr = PetscMemzero(olens,m*sizeof(PetscInt));CHKERRQ(ierr); 312 ierr = PetscCalloc1(m,&ld);CHKERRQ(ierr); 313 jj = 0; 314 for (i=0; i<m; i++) { 315 for (j=0; j<dlens[i]; j++) { 316 if (gmataj[jj] < rstart) ld[i]++; 317 if (gmataj[jj] < rstart || gmataj[jj] >= rend) olens[i]++; 318 jj++; 319 } 320 } 321 /* receive numerical values */ 322 ierr = PetscMemzero(gmataa,nz*sizeof(PetscScalar));CHKERRQ(ierr); 323 ierr = MPI_Recv(gmataa,nz,MPIU_SCALAR,0,tag,comm,&status);CHKERRQ(ierr); 324 } 325 /* set preallocation */ 326 for (i=0; i<m; i++) { 327 dlens[i] -= olens[i]; 328 } 329 ierr = MatSeqAIJSetPreallocation(mat,0,dlens);CHKERRQ(ierr); 330 ierr = MatMPIAIJSetPreallocation(mat,0,dlens,0,olens);CHKERRQ(ierr); 331 332 for (i=0; i<m; i++) { 333 dlens[i] += olens[i]; 334 } 335 cnt = 0; 336 for (i=0; i<m; i++) { 337 row = rstart + i; 338 ierr = MatSetValues(mat,1,&row,dlens[i],gmataj+cnt,gmataa+cnt,INSERT_VALUES);CHKERRQ(ierr); 339 cnt += dlens[i]; 340 } 341 if (rank) { 342 ierr = PetscFree2(gmataa,gmataj);CHKERRQ(ierr); 343 } 344 ierr = PetscFree2(dlens,olens);CHKERRQ(ierr); 345 ierr = PetscFree(rowners);CHKERRQ(ierr); 346 347 ((Mat_MPIAIJ*)(mat->data))->ld = ld; 348 349 *inmat = mat; 350 } else { /* column indices are already set; only need to move over numerical values from process 0 */ 351 Mat_SeqAIJ *Ad = (Mat_SeqAIJ*)((Mat_MPIAIJ*)((*inmat)->data))->A->data; 352 Mat_SeqAIJ *Ao = (Mat_SeqAIJ*)((Mat_MPIAIJ*)((*inmat)->data))->B->data; 353 mat = *inmat; 354 ierr = PetscObjectGetNewTag((PetscObject)mat,&tag);CHKERRQ(ierr); 355 if (!rank) { 356 /* send numerical values to other processes */ 357 gmata = (Mat_SeqAIJ*) gmat->data; 358 ierr = MatGetOwnershipRanges(mat,(const PetscInt**)&rowners);CHKERRQ(ierr); 359 gmataa = gmata->a; 360 for (i=1; i<size; i++) { 361 nz = gmata->i[rowners[i+1]]-gmata->i[rowners[i]]; 362 ierr = MPI_Send(gmataa + gmata->i[rowners[i]],nz,MPIU_SCALAR,i,tag,comm);CHKERRQ(ierr); 363 } 364 nz = gmata->i[rowners[1]]-gmata->i[rowners[0]]; 365 } else { 366 /* receive numerical values from process 0*/ 367 nz = Ad->nz + Ao->nz; 368 ierr = PetscMalloc1(nz,&gmataa);CHKERRQ(ierr); gmataarestore = gmataa; 369 ierr = MPI_Recv(gmataa,nz,MPIU_SCALAR,0,tag,comm,&status);CHKERRQ(ierr); 370 } 371 /* transfer numerical values into the diagonal A and off diagonal B parts of mat */ 372 ld = ((Mat_MPIAIJ*)(mat->data))->ld; 373 ad = Ad->a; 374 ao = Ao->a; 375 if (mat->rmap->n) { 376 i = 0; 377 nz = ld[i]; ierr = PetscMemcpy(ao,gmataa,nz*sizeof(PetscScalar));CHKERRQ(ierr); ao += nz; gmataa += nz; 378 nz = Ad->i[i+1] - Ad->i[i]; ierr = PetscMemcpy(ad,gmataa,nz*sizeof(PetscScalar));CHKERRQ(ierr); ad += nz; gmataa += nz; 379 } 380 for (i=1; i<mat->rmap->n; i++) { 381 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; 382 nz = Ad->i[i+1] - Ad->i[i]; ierr = PetscMemcpy(ad,gmataa,nz*sizeof(PetscScalar));CHKERRQ(ierr); ad += nz; gmataa += nz; 383 } 384 i--; 385 if (mat->rmap->n) { 386 nz = Ao->i[i+1] - Ao->i[i] - ld[i]; ierr = PetscMemcpy(ao,gmataa,nz*sizeof(PetscScalar));CHKERRQ(ierr); 387 } 388 if (rank) { 389 ierr = PetscFree(gmataarestore);CHKERRQ(ierr); 390 } 391 } 392 ierr = MatAssemblyBegin(mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 393 ierr = MatAssemblyEnd(mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 394 PetscFunctionReturn(0); 395 } 396 397 /* 398 Local utility routine that creates a mapping from the global column 399 number to the local number in the off-diagonal part of the local 400 storage of the matrix. When PETSC_USE_CTABLE is used this is scalable at 401 a slightly higher hash table cost; without it it is not scalable (each processor 402 has an order N integer array but is fast to acess. 403 */ 404 #undef __FUNCT__ 405 #define __FUNCT__ "MatCreateColmap_MPIAIJ_Private" 406 PetscErrorCode MatCreateColmap_MPIAIJ_Private(Mat mat) 407 { 408 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 409 PetscErrorCode ierr; 410 PetscInt n = aij->B->cmap->n,i; 411 412 PetscFunctionBegin; 413 if (!aij->garray) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"MPIAIJ Matrix was assembled but is missing garray"); 414 #if defined(PETSC_USE_CTABLE) 415 ierr = PetscTableCreate(n,mat->cmap->N+1,&aij->colmap);CHKERRQ(ierr); 416 for (i=0; i<n; i++) { 417 ierr = PetscTableAdd(aij->colmap,aij->garray[i]+1,i+1,INSERT_VALUES);CHKERRQ(ierr); 418 } 419 #else 420 ierr = PetscCalloc1(mat->cmap->N+1,&aij->colmap);CHKERRQ(ierr); 421 ierr = PetscLogObjectMemory((PetscObject)mat,(mat->cmap->N+1)*sizeof(PetscInt));CHKERRQ(ierr); 422 for (i=0; i<n; i++) aij->colmap[aij->garray[i]] = i+1; 423 #endif 424 PetscFunctionReturn(0); 425 } 426 427 #define MatSetValues_SeqAIJ_A_Private(row,col,value,addv,orow,ocol) \ 428 { \ 429 if (col <= lastcol1) low1 = 0; \ 430 else high1 = nrow1; \ 431 lastcol1 = col;\ 432 while (high1-low1 > 5) { \ 433 t = (low1+high1)/2; \ 434 if (rp1[t] > col) high1 = t; \ 435 else low1 = t; \ 436 } \ 437 for (_i=low1; _i<high1; _i++) { \ 438 if (rp1[_i] > col) break; \ 439 if (rp1[_i] == col) { \ 440 if (addv == ADD_VALUES) ap1[_i] += value; \ 441 else ap1[_i] = value; \ 442 goto a_noinsert; \ 443 } \ 444 } \ 445 if (value == 0.0 && ignorezeroentries) {low1 = 0; high1 = nrow1;goto a_noinsert;} \ 446 if (nonew == 1) {low1 = 0; high1 = nrow1; goto a_noinsert;} \ 447 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); \ 448 MatSeqXAIJReallocateAIJ(A,am,1,nrow1,row,col,rmax1,aa,ai,aj,rp1,ap1,aimax,nonew,MatScalar); \ 449 N = nrow1++ - 1; a->nz++; high1++; \ 450 /* shift up all the later entries in this row */ \ 451 for (ii=N; ii>=_i; ii--) { \ 452 rp1[ii+1] = rp1[ii]; \ 453 ap1[ii+1] = ap1[ii]; \ 454 } \ 455 rp1[_i] = col; \ 456 ap1[_i] = value; \ 457 A->nonzerostate++;\ 458 a_noinsert: ; \ 459 ailen[row] = nrow1; \ 460 } 461 462 463 #define MatSetValues_SeqAIJ_B_Private(row,col,value,addv,orow,ocol) \ 464 { \ 465 if (col <= lastcol2) low2 = 0; \ 466 else high2 = nrow2; \ 467 lastcol2 = col; \ 468 while (high2-low2 > 5) { \ 469 t = (low2+high2)/2; \ 470 if (rp2[t] > col) high2 = t; \ 471 else low2 = t; \ 472 } \ 473 for (_i=low2; _i<high2; _i++) { \ 474 if (rp2[_i] > col) break; \ 475 if (rp2[_i] == col) { \ 476 if (addv == ADD_VALUES) ap2[_i] += value; \ 477 else ap2[_i] = value; \ 478 goto b_noinsert; \ 479 } \ 480 } \ 481 if (value == 0.0 && ignorezeroentries) {low2 = 0; high2 = nrow2; goto b_noinsert;} \ 482 if (nonew == 1) {low2 = 0; high2 = nrow2; goto b_noinsert;} \ 483 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); \ 484 MatSeqXAIJReallocateAIJ(B,bm,1,nrow2,row,col,rmax2,ba,bi,bj,rp2,ap2,bimax,nonew,MatScalar); \ 485 N = nrow2++ - 1; b->nz++; high2++; \ 486 /* shift up all the later entries in this row */ \ 487 for (ii=N; ii>=_i; ii--) { \ 488 rp2[ii+1] = rp2[ii]; \ 489 ap2[ii+1] = ap2[ii]; \ 490 } \ 491 rp2[_i] = col; \ 492 ap2[_i] = value; \ 493 B->nonzerostate++; \ 494 b_noinsert: ; \ 495 bilen[row] = nrow2; \ 496 } 497 498 #undef __FUNCT__ 499 #define __FUNCT__ "MatSetValuesRow_MPIAIJ" 500 PetscErrorCode MatSetValuesRow_MPIAIJ(Mat A,PetscInt row,const PetscScalar v[]) 501 { 502 Mat_MPIAIJ *mat = (Mat_MPIAIJ*)A->data; 503 Mat_SeqAIJ *a = (Mat_SeqAIJ*)mat->A->data,*b = (Mat_SeqAIJ*)mat->B->data; 504 PetscErrorCode ierr; 505 PetscInt l,*garray = mat->garray,diag; 506 507 PetscFunctionBegin; 508 /* code only works for square matrices A */ 509 510 /* find size of row to the left of the diagonal part */ 511 ierr = MatGetOwnershipRange(A,&diag,0);CHKERRQ(ierr); 512 row = row - diag; 513 for (l=0; l<b->i[row+1]-b->i[row]; l++) { 514 if (garray[b->j[b->i[row]+l]] > diag) break; 515 } 516 ierr = PetscMemcpy(b->a+b->i[row],v,l*sizeof(PetscScalar));CHKERRQ(ierr); 517 518 /* diagonal part */ 519 ierr = PetscMemcpy(a->a+a->i[row],v+l,(a->i[row+1]-a->i[row])*sizeof(PetscScalar));CHKERRQ(ierr); 520 521 /* right of diagonal part */ 522 ierr = PetscMemcpy(b->a+b->i[row]+l,v+l+a->i[row+1]-a->i[row],(b->i[row+1]-b->i[row]-l)*sizeof(PetscScalar));CHKERRQ(ierr); 523 PetscFunctionReturn(0); 524 } 525 526 #undef __FUNCT__ 527 #define __FUNCT__ "MatSetValues_MPIAIJ" 528 PetscErrorCode MatSetValues_MPIAIJ(Mat mat,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],const PetscScalar v[],InsertMode addv) 529 { 530 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 531 PetscScalar value; 532 PetscErrorCode ierr; 533 PetscInt i,j,rstart = mat->rmap->rstart,rend = mat->rmap->rend; 534 PetscInt cstart = mat->cmap->rstart,cend = mat->cmap->rend,row,col; 535 PetscBool roworiented = aij->roworiented; 536 537 /* Some Variables required in the macro */ 538 Mat A = aij->A; 539 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 540 PetscInt *aimax = a->imax,*ai = a->i,*ailen = a->ilen,*aj = a->j; 541 MatScalar *aa = a->a; 542 PetscBool ignorezeroentries = a->ignorezeroentries; 543 Mat B = aij->B; 544 Mat_SeqAIJ *b = (Mat_SeqAIJ*)B->data; 545 PetscInt *bimax = b->imax,*bi = b->i,*bilen = b->ilen,*bj = b->j,bm = aij->B->rmap->n,am = aij->A->rmap->n; 546 MatScalar *ba = b->a; 547 548 PetscInt *rp1,*rp2,ii,nrow1,nrow2,_i,rmax1,rmax2,N,low1,high1,low2,high2,t,lastcol1,lastcol2; 549 PetscInt nonew; 550 MatScalar *ap1,*ap2; 551 552 PetscFunctionBegin; 553 for (i=0; i<m; i++) { 554 if (im[i] < 0) continue; 555 #if defined(PETSC_USE_DEBUG) 556 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); 557 #endif 558 if (im[i] >= rstart && im[i] < rend) { 559 row = im[i] - rstart; 560 lastcol1 = -1; 561 rp1 = aj + ai[row]; 562 ap1 = aa + ai[row]; 563 rmax1 = aimax[row]; 564 nrow1 = ailen[row]; 565 low1 = 0; 566 high1 = nrow1; 567 lastcol2 = -1; 568 rp2 = bj + bi[row]; 569 ap2 = ba + bi[row]; 570 rmax2 = bimax[row]; 571 nrow2 = bilen[row]; 572 low2 = 0; 573 high2 = nrow2; 574 575 for (j=0; j<n; j++) { 576 if (roworiented) value = v[i*n+j]; 577 else value = v[i+j*m]; 578 if (ignorezeroentries && value == 0.0 && (addv == ADD_VALUES)) continue; 579 if (in[j] >= cstart && in[j] < cend) { 580 col = in[j] - cstart; 581 nonew = a->nonew; 582 MatSetValues_SeqAIJ_A_Private(row,col,value,addv,im[i],in[j]); 583 } else if (in[j] < 0) continue; 584 #if defined(PETSC_USE_DEBUG) 585 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); 586 #endif 587 else { 588 if (mat->was_assembled) { 589 if (!aij->colmap) { 590 ierr = MatCreateColmap_MPIAIJ_Private(mat);CHKERRQ(ierr); 591 } 592 #if defined(PETSC_USE_CTABLE) 593 ierr = PetscTableFind(aij->colmap,in[j]+1,&col);CHKERRQ(ierr); 594 col--; 595 #else 596 col = aij->colmap[in[j]] - 1; 597 #endif 598 if (col < 0 && !((Mat_SeqAIJ*)(aij->B->data))->nonew) { 599 ierr = MatDisAssemble_MPIAIJ(mat);CHKERRQ(ierr); 600 col = in[j]; 601 /* Reinitialize the variables required by MatSetValues_SeqAIJ_B_Private() */ 602 B = aij->B; 603 b = (Mat_SeqAIJ*)B->data; 604 bimax = b->imax; bi = b->i; bilen = b->ilen; bj = b->j; ba = b->a; 605 rp2 = bj + bi[row]; 606 ap2 = ba + bi[row]; 607 rmax2 = bimax[row]; 608 nrow2 = bilen[row]; 609 low2 = 0; 610 high2 = nrow2; 611 bm = aij->B->rmap->n; 612 ba = b->a; 613 } 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]); 614 } else col = in[j]; 615 nonew = b->nonew; 616 MatSetValues_SeqAIJ_B_Private(row,col,value,addv,im[i],in[j]); 617 } 618 } 619 } else { 620 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]); 621 if (!aij->donotstash) { 622 mat->assembled = PETSC_FALSE; 623 if (roworiented) { 624 ierr = MatStashValuesRow_Private(&mat->stash,im[i],n,in,v+i*n,(PetscBool)(ignorezeroentries && (addv == ADD_VALUES)));CHKERRQ(ierr); 625 } else { 626 ierr = MatStashValuesCol_Private(&mat->stash,im[i],n,in,v+i,m,(PetscBool)(ignorezeroentries && (addv == ADD_VALUES)));CHKERRQ(ierr); 627 } 628 } 629 } 630 } 631 PetscFunctionReturn(0); 632 } 633 634 #undef __FUNCT__ 635 #define __FUNCT__ "MatGetValues_MPIAIJ" 636 PetscErrorCode MatGetValues_MPIAIJ(Mat mat,PetscInt m,const PetscInt idxm[],PetscInt n,const PetscInt idxn[],PetscScalar v[]) 637 { 638 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 639 PetscErrorCode ierr; 640 PetscInt i,j,rstart = mat->rmap->rstart,rend = mat->rmap->rend; 641 PetscInt cstart = mat->cmap->rstart,cend = mat->cmap->rend,row,col; 642 643 PetscFunctionBegin; 644 for (i=0; i<m; i++) { 645 if (idxm[i] < 0) continue; /* SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative row: %D",idxm[i]);*/ 646 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); 647 if (idxm[i] >= rstart && idxm[i] < rend) { 648 row = idxm[i] - rstart; 649 for (j=0; j<n; j++) { 650 if (idxn[j] < 0) continue; /* SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative column: %D",idxn[j]); */ 651 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); 652 if (idxn[j] >= cstart && idxn[j] < cend) { 653 col = idxn[j] - cstart; 654 ierr = MatGetValues(aij->A,1,&row,1,&col,v+i*n+j);CHKERRQ(ierr); 655 } else { 656 if (!aij->colmap) { 657 ierr = MatCreateColmap_MPIAIJ_Private(mat);CHKERRQ(ierr); 658 } 659 #if defined(PETSC_USE_CTABLE) 660 ierr = PetscTableFind(aij->colmap,idxn[j]+1,&col);CHKERRQ(ierr); 661 col--; 662 #else 663 col = aij->colmap[idxn[j]] - 1; 664 #endif 665 if ((col < 0) || (aij->garray[col] != idxn[j])) *(v+i*n+j) = 0.0; 666 else { 667 ierr = MatGetValues(aij->B,1,&row,1,&col,v+i*n+j);CHKERRQ(ierr); 668 } 669 } 670 } 671 } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Only local values currently supported"); 672 } 673 PetscFunctionReturn(0); 674 } 675 676 extern PetscErrorCode MatMultDiagonalBlock_MPIAIJ(Mat,Vec,Vec); 677 678 #undef __FUNCT__ 679 #define __FUNCT__ "MatAssemblyBegin_MPIAIJ" 680 PetscErrorCode MatAssemblyBegin_MPIAIJ(Mat mat,MatAssemblyType mode) 681 { 682 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 683 PetscErrorCode ierr; 684 PetscInt nstash,reallocs; 685 InsertMode addv; 686 687 PetscFunctionBegin; 688 if (aij->donotstash || mat->nooffprocentries) PetscFunctionReturn(0); 689 690 /* make sure all processors are either in INSERTMODE or ADDMODE */ 691 ierr = MPI_Allreduce((PetscEnum*)&mat->insertmode,(PetscEnum*)&addv,1,MPIU_ENUM,MPI_BOR,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 692 if (addv == (ADD_VALUES|INSERT_VALUES)) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_ARG_WRONGSTATE,"Some processors inserted others added"); 693 mat->insertmode = addv; /* in case this processor had no cache */ 694 695 ierr = MatStashScatterBegin_Private(mat,&mat->stash,mat->rmap->range);CHKERRQ(ierr); 696 ierr = MatStashGetInfo_Private(&mat->stash,&nstash,&reallocs);CHKERRQ(ierr); 697 ierr = PetscInfo2(aij->A,"Stash has %D entries, uses %D mallocs.\n",nstash,reallocs);CHKERRQ(ierr); 698 PetscFunctionReturn(0); 699 } 700 701 #undef __FUNCT__ 702 #define __FUNCT__ "MatAssemblyEnd_MPIAIJ" 703 PetscErrorCode MatAssemblyEnd_MPIAIJ(Mat mat,MatAssemblyType mode) 704 { 705 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 706 Mat_SeqAIJ *a = (Mat_SeqAIJ*)aij->A->data; 707 PetscErrorCode ierr; 708 PetscMPIInt n; 709 PetscInt i,j,rstart,ncols,flg; 710 PetscInt *row,*col; 711 PetscBool other_disassembled; 712 PetscScalar *val; 713 InsertMode addv = mat->insertmode; 714 715 /* do not use 'b = (Mat_SeqAIJ*)aij->B->data' as B can be reset in disassembly */ 716 717 PetscFunctionBegin; 718 if (!aij->donotstash && !mat->nooffprocentries) { 719 while (1) { 720 ierr = MatStashScatterGetMesg_Private(&mat->stash,&n,&row,&col,&val,&flg);CHKERRQ(ierr); 721 if (!flg) break; 722 723 for (i=0; i<n; ) { 724 /* Now identify the consecutive vals belonging to the same row */ 725 for (j=i,rstart=row[j]; j<n; j++) { 726 if (row[j] != rstart) break; 727 } 728 if (j < n) ncols = j-i; 729 else ncols = n-i; 730 /* Now assemble all these values with a single function call */ 731 ierr = MatSetValues_MPIAIJ(mat,1,row+i,ncols,col+i,val+i,addv);CHKERRQ(ierr); 732 733 i = j; 734 } 735 } 736 ierr = MatStashScatterEnd_Private(&mat->stash);CHKERRQ(ierr); 737 } 738 ierr = MatAssemblyBegin(aij->A,mode);CHKERRQ(ierr); 739 ierr = MatAssemblyEnd(aij->A,mode);CHKERRQ(ierr); 740 741 /* determine if any processor has disassembled, if so we must 742 also disassemble ourselfs, in order that we may reassemble. */ 743 /* 744 if nonzero structure of submatrix B cannot change then we know that 745 no processor disassembled thus we can skip this stuff 746 */ 747 if (!((Mat_SeqAIJ*)aij->B->data)->nonew) { 748 ierr = MPI_Allreduce(&mat->was_assembled,&other_disassembled,1,MPIU_BOOL,MPI_PROD,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 749 if (mat->was_assembled && !other_disassembled) { 750 ierr = MatDisAssemble_MPIAIJ(mat);CHKERRQ(ierr); 751 } 752 } 753 if (!mat->was_assembled && mode == MAT_FINAL_ASSEMBLY) { 754 ierr = MatSetUpMultiply_MPIAIJ(mat);CHKERRQ(ierr); 755 } 756 ierr = MatSetOption(aij->B,MAT_USE_INODES,PETSC_FALSE);CHKERRQ(ierr); 757 ierr = MatAssemblyBegin(aij->B,mode);CHKERRQ(ierr); 758 ierr = MatAssemblyEnd(aij->B,mode);CHKERRQ(ierr); 759 760 ierr = PetscFree2(aij->rowvalues,aij->rowindices);CHKERRQ(ierr); 761 762 aij->rowvalues = 0; 763 764 ierr = VecDestroy(&aij->diag);CHKERRQ(ierr); 765 if (a->inode.size) mat->ops->multdiagonalblock = MatMultDiagonalBlock_MPIAIJ; 766 767 /* if no new nonzero locations are allowed in matrix then only set the matrix state the first time through */ 768 if ((!mat->was_assembled && mode == MAT_FINAL_ASSEMBLY) || !((Mat_SeqAIJ*)(aij->A->data))->nonew) { 769 PetscObjectState state = aij->A->nonzerostate + aij->B->nonzerostate; 770 ierr = MPI_Allreduce(&state,&mat->nonzerostate,1,MPIU_INT64,MPI_SUM,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 771 } 772 PetscFunctionReturn(0); 773 } 774 775 #undef __FUNCT__ 776 #define __FUNCT__ "MatZeroEntries_MPIAIJ" 777 PetscErrorCode MatZeroEntries_MPIAIJ(Mat A) 778 { 779 Mat_MPIAIJ *l = (Mat_MPIAIJ*)A->data; 780 PetscErrorCode ierr; 781 782 PetscFunctionBegin; 783 ierr = MatZeroEntries(l->A);CHKERRQ(ierr); 784 ierr = MatZeroEntries(l->B);CHKERRQ(ierr); 785 PetscFunctionReturn(0); 786 } 787 788 #undef __FUNCT__ 789 #define __FUNCT__ "MatZeroRows_MPIAIJ" 790 PetscErrorCode MatZeroRows_MPIAIJ(Mat A,PetscInt N,const PetscInt rows[],PetscScalar diag,Vec x,Vec b) 791 { 792 Mat_MPIAIJ *mat = (Mat_MPIAIJ *) A->data; 793 PetscInt *owners = A->rmap->range; 794 PetscInt n = A->rmap->n; 795 PetscSF sf; 796 PetscInt *lrows; 797 PetscSFNode *rrows; 798 PetscInt r, p = 0, len = 0; 799 PetscErrorCode ierr; 800 801 PetscFunctionBegin; 802 /* Create SF where leaves are input rows and roots are owned rows */ 803 ierr = PetscMalloc1(n, &lrows);CHKERRQ(ierr); 804 for (r = 0; r < n; ++r) lrows[r] = -1; 805 if (!A->nooffproczerorows) {ierr = PetscMalloc1(N, &rrows);CHKERRQ(ierr);} 806 for (r = 0; r < N; ++r) { 807 const PetscInt idx = rows[r]; 808 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); 809 if (idx < owners[p] || owners[p+1] <= idx) { /* short-circuit the search if the last p owns this row too */ 810 ierr = PetscLayoutFindOwner(A->rmap,idx,&p);CHKERRQ(ierr); 811 } 812 if (A->nooffproczerorows) { 813 if (p != mat->rank) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"MAT_NO_OFF_PROC_ZERO_ROWS set, but row %D is not owned by rank %d",idx,mat->rank); 814 lrows[len++] = idx - owners[p]; 815 } else { 816 rrows[r].rank = p; 817 rrows[r].index = rows[r] - owners[p]; 818 } 819 } 820 if (!A->nooffproczerorows) { 821 ierr = PetscSFCreate(PetscObjectComm((PetscObject) A), &sf);CHKERRQ(ierr); 822 ierr = PetscSFSetGraph(sf, n, N, NULL, PETSC_OWN_POINTER, rrows, PETSC_OWN_POINTER);CHKERRQ(ierr); 823 /* Collect flags for rows to be zeroed */ 824 ierr = PetscSFReduceBegin(sf, MPIU_INT, (PetscInt*)rows, lrows, MPI_LOR);CHKERRQ(ierr); 825 ierr = PetscSFReduceEnd(sf, MPIU_INT, (PetscInt*)rows, lrows, MPI_LOR);CHKERRQ(ierr); 826 ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 827 /* Compress and put in row numbers */ 828 for (r = 0; r < n; ++r) if (lrows[r] >= 0) lrows[len++] = r; 829 } 830 /* fix right hand side if needed */ 831 if (x && b) { 832 const PetscScalar *xx; 833 PetscScalar *bb; 834 835 ierr = VecGetArrayRead(x, &xx);CHKERRQ(ierr); 836 ierr = VecGetArray(b, &bb);CHKERRQ(ierr); 837 for (r = 0; r < len; ++r) bb[lrows[r]] = diag*xx[lrows[r]]; 838 ierr = VecRestoreArrayRead(x, &xx);CHKERRQ(ierr); 839 ierr = VecRestoreArray(b, &bb);CHKERRQ(ierr); 840 } 841 /* Must zero l->B before l->A because the (diag) case below may put values into l->B*/ 842 ierr = MatZeroRows(mat->B, len, lrows, 0.0, NULL, NULL);CHKERRQ(ierr); 843 if ((diag != 0.0) && (mat->A->rmap->N == mat->A->cmap->N)) { 844 ierr = MatZeroRows(mat->A, len, lrows, diag, NULL, NULL);CHKERRQ(ierr); 845 } else if (diag != 0.0) { 846 ierr = MatZeroRows(mat->A, len, lrows, 0.0, NULL, NULL);CHKERRQ(ierr); 847 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"); 848 for (r = 0; r < len; ++r) { 849 const PetscInt row = lrows[r] + A->rmap->rstart; 850 ierr = MatSetValues(A, 1, &row, 1, &row, &diag, INSERT_VALUES);CHKERRQ(ierr); 851 } 852 ierr = MatAssemblyBegin(A, MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 853 ierr = MatAssemblyEnd(A, MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 854 } else { 855 ierr = MatZeroRows(mat->A, len, lrows, 0.0, NULL, NULL);CHKERRQ(ierr); 856 } 857 ierr = PetscFree(lrows);CHKERRQ(ierr); 858 859 /* only change matrix nonzero state if pattern was allowed to be changed */ 860 if (!((Mat_SeqAIJ*)(mat->A->data))->keepnonzeropattern) { 861 PetscObjectState state = mat->A->nonzerostate + mat->B->nonzerostate; 862 ierr = MPI_Allreduce(&state,&A->nonzerostate,1,MPIU_INT64,MPI_SUM,PetscObjectComm((PetscObject)A));CHKERRQ(ierr); 863 } 864 PetscFunctionReturn(0); 865 } 866 867 #undef __FUNCT__ 868 #define __FUNCT__ "MatZeroRowsColumns_MPIAIJ" 869 PetscErrorCode MatZeroRowsColumns_MPIAIJ(Mat A,PetscInt N,const PetscInt rows[],PetscScalar diag,Vec x,Vec b) 870 { 871 Mat_MPIAIJ *l = (Mat_MPIAIJ*)A->data; 872 PetscErrorCode ierr; 873 PetscMPIInt n = A->rmap->n; 874 PetscInt i,j,r,m,p = 0,len = 0; 875 PetscInt *lrows,*owners = A->rmap->range; 876 PetscSFNode *rrows; 877 PetscSF sf; 878 const PetscScalar *xx; 879 PetscScalar *bb,*mask; 880 Vec xmask,lmask; 881 Mat_SeqAIJ *aij = (Mat_SeqAIJ*)l->B->data; 882 const PetscInt *aj, *ii,*ridx; 883 PetscScalar *aa; 884 885 PetscFunctionBegin; 886 /* Create SF where leaves are input rows and roots are owned rows */ 887 ierr = PetscMalloc1(n, &lrows);CHKERRQ(ierr); 888 for (r = 0; r < n; ++r) lrows[r] = -1; 889 ierr = PetscMalloc1(N, &rrows);CHKERRQ(ierr); 890 for (r = 0; r < N; ++r) { 891 const PetscInt idx = rows[r]; 892 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); 893 if (idx < owners[p] || owners[p+1] <= idx) { /* short-circuit the search if the last p owns this row too */ 894 ierr = PetscLayoutFindOwner(A->rmap,idx,&p);CHKERRQ(ierr); 895 } 896 rrows[r].rank = p; 897 rrows[r].index = rows[r] - owners[p]; 898 } 899 ierr = PetscSFCreate(PetscObjectComm((PetscObject) A), &sf);CHKERRQ(ierr); 900 ierr = PetscSFSetGraph(sf, n, N, NULL, PETSC_OWN_POINTER, rrows, PETSC_OWN_POINTER);CHKERRQ(ierr); 901 /* Collect flags for rows to be zeroed */ 902 ierr = PetscSFReduceBegin(sf, MPIU_INT, (PetscInt *) rows, lrows, MPI_LOR);CHKERRQ(ierr); 903 ierr = PetscSFReduceEnd(sf, MPIU_INT, (PetscInt *) rows, lrows, MPI_LOR);CHKERRQ(ierr); 904 ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 905 /* Compress and put in row numbers */ 906 for (r = 0; r < n; ++r) if (lrows[r] >= 0) lrows[len++] = r; 907 /* zero diagonal part of matrix */ 908 ierr = MatZeroRowsColumns(l->A,len,lrows,diag,x,b);CHKERRQ(ierr); 909 /* handle off diagonal part of matrix */ 910 ierr = MatCreateVecs(A,&xmask,NULL);CHKERRQ(ierr); 911 ierr = VecDuplicate(l->lvec,&lmask);CHKERRQ(ierr); 912 ierr = VecGetArray(xmask,&bb);CHKERRQ(ierr); 913 for (i=0; i<len; i++) bb[lrows[i]] = 1; 914 ierr = VecRestoreArray(xmask,&bb);CHKERRQ(ierr); 915 ierr = VecScatterBegin(l->Mvctx,xmask,lmask,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 916 ierr = VecScatterEnd(l->Mvctx,xmask,lmask,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 917 ierr = VecDestroy(&xmask);CHKERRQ(ierr); 918 if (x) { 919 ierr = VecScatterBegin(l->Mvctx,x,l->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 920 ierr = VecScatterEnd(l->Mvctx,x,l->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 921 ierr = VecGetArrayRead(l->lvec,&xx);CHKERRQ(ierr); 922 ierr = VecGetArray(b,&bb);CHKERRQ(ierr); 923 } 924 ierr = VecGetArray(lmask,&mask);CHKERRQ(ierr); 925 /* remove zeroed rows of off diagonal matrix */ 926 ii = aij->i; 927 for (i=0; i<len; i++) { 928 ierr = PetscMemzero(aij->a + ii[lrows[i]],(ii[lrows[i]+1] - ii[lrows[i]])*sizeof(PetscScalar));CHKERRQ(ierr); 929 } 930 /* loop over all elements of off process part of matrix zeroing removed columns*/ 931 if (aij->compressedrow.use) { 932 m = aij->compressedrow.nrows; 933 ii = aij->compressedrow.i; 934 ridx = aij->compressedrow.rindex; 935 for (i=0; i<m; i++) { 936 n = ii[i+1] - ii[i]; 937 aj = aij->j + ii[i]; 938 aa = aij->a + ii[i]; 939 940 for (j=0; j<n; j++) { 941 if (PetscAbsScalar(mask[*aj])) { 942 if (b) bb[*ridx] -= *aa*xx[*aj]; 943 *aa = 0.0; 944 } 945 aa++; 946 aj++; 947 } 948 ridx++; 949 } 950 } else { /* do not use compressed row format */ 951 m = l->B->rmap->n; 952 for (i=0; i<m; i++) { 953 n = ii[i+1] - ii[i]; 954 aj = aij->j + ii[i]; 955 aa = aij->a + ii[i]; 956 for (j=0; j<n; j++) { 957 if (PetscAbsScalar(mask[*aj])) { 958 if (b) bb[i] -= *aa*xx[*aj]; 959 *aa = 0.0; 960 } 961 aa++; 962 aj++; 963 } 964 } 965 } 966 if (x) { 967 ierr = VecRestoreArray(b,&bb);CHKERRQ(ierr); 968 ierr = VecRestoreArrayRead(l->lvec,&xx);CHKERRQ(ierr); 969 } 970 ierr = VecRestoreArray(lmask,&mask);CHKERRQ(ierr); 971 ierr = VecDestroy(&lmask);CHKERRQ(ierr); 972 ierr = PetscFree(lrows);CHKERRQ(ierr); 973 974 /* only change matrix nonzero state if pattern was allowed to be changed */ 975 if (!((Mat_SeqAIJ*)(l->A->data))->keepnonzeropattern) { 976 PetscObjectState state = l->A->nonzerostate + l->B->nonzerostate; 977 ierr = MPI_Allreduce(&state,&A->nonzerostate,1,MPIU_INT64,MPI_SUM,PetscObjectComm((PetscObject)A));CHKERRQ(ierr); 978 } 979 PetscFunctionReturn(0); 980 } 981 982 #undef __FUNCT__ 983 #define __FUNCT__ "MatMult_MPIAIJ" 984 PetscErrorCode MatMult_MPIAIJ(Mat A,Vec xx,Vec yy) 985 { 986 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 987 PetscErrorCode ierr; 988 PetscInt nt; 989 990 PetscFunctionBegin; 991 ierr = VecGetLocalSize(xx,&nt);CHKERRQ(ierr); 992 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); 993 ierr = VecScatterBegin(a->Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 994 ierr = (*a->A->ops->mult)(a->A,xx,yy);CHKERRQ(ierr); 995 ierr = VecScatterEnd(a->Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 996 ierr = (*a->B->ops->multadd)(a->B,a->lvec,yy,yy);CHKERRQ(ierr); 997 PetscFunctionReturn(0); 998 } 999 1000 #undef __FUNCT__ 1001 #define __FUNCT__ "MatMultDiagonalBlock_MPIAIJ" 1002 PetscErrorCode MatMultDiagonalBlock_MPIAIJ(Mat A,Vec bb,Vec xx) 1003 { 1004 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 1005 PetscErrorCode ierr; 1006 1007 PetscFunctionBegin; 1008 ierr = MatMultDiagonalBlock(a->A,bb,xx);CHKERRQ(ierr); 1009 PetscFunctionReturn(0); 1010 } 1011 1012 #undef __FUNCT__ 1013 #define __FUNCT__ "MatMultAdd_MPIAIJ" 1014 PetscErrorCode MatMultAdd_MPIAIJ(Mat A,Vec xx,Vec yy,Vec zz) 1015 { 1016 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 1017 PetscErrorCode ierr; 1018 1019 PetscFunctionBegin; 1020 ierr = VecScatterBegin(a->Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1021 ierr = (*a->A->ops->multadd)(a->A,xx,yy,zz);CHKERRQ(ierr); 1022 ierr = VecScatterEnd(a->Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1023 ierr = (*a->B->ops->multadd)(a->B,a->lvec,zz,zz);CHKERRQ(ierr); 1024 PetscFunctionReturn(0); 1025 } 1026 1027 #undef __FUNCT__ 1028 #define __FUNCT__ "MatMultTranspose_MPIAIJ" 1029 PetscErrorCode MatMultTranspose_MPIAIJ(Mat A,Vec xx,Vec yy) 1030 { 1031 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 1032 PetscErrorCode ierr; 1033 PetscBool merged; 1034 1035 PetscFunctionBegin; 1036 ierr = VecScatterGetMerged(a->Mvctx,&merged);CHKERRQ(ierr); 1037 /* do nondiagonal part */ 1038 ierr = (*a->B->ops->multtranspose)(a->B,xx,a->lvec);CHKERRQ(ierr); 1039 if (!merged) { 1040 /* send it on its way */ 1041 ierr = VecScatterBegin(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1042 /* do local part */ 1043 ierr = (*a->A->ops->multtranspose)(a->A,xx,yy);CHKERRQ(ierr); 1044 /* receive remote parts: note this assumes the values are not actually */ 1045 /* added in yy until the next line, */ 1046 ierr = VecScatterEnd(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1047 } else { 1048 /* do local part */ 1049 ierr = (*a->A->ops->multtranspose)(a->A,xx,yy);CHKERRQ(ierr); 1050 /* send it on its way */ 1051 ierr = VecScatterBegin(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1052 /* values actually were received in the Begin() but we need to call this nop */ 1053 ierr = VecScatterEnd(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1054 } 1055 PetscFunctionReturn(0); 1056 } 1057 1058 #undef __FUNCT__ 1059 #define __FUNCT__ "MatIsTranspose_MPIAIJ" 1060 PetscErrorCode MatIsTranspose_MPIAIJ(Mat Amat,Mat Bmat,PetscReal tol,PetscBool *f) 1061 { 1062 MPI_Comm comm; 1063 Mat_MPIAIJ *Aij = (Mat_MPIAIJ*) Amat->data, *Bij; 1064 Mat Adia = Aij->A, Bdia, Aoff,Boff,*Aoffs,*Boffs; 1065 IS Me,Notme; 1066 PetscErrorCode ierr; 1067 PetscInt M,N,first,last,*notme,i; 1068 PetscMPIInt size; 1069 1070 PetscFunctionBegin; 1071 /* Easy test: symmetric diagonal block */ 1072 Bij = (Mat_MPIAIJ*) Bmat->data; Bdia = Bij->A; 1073 ierr = MatIsTranspose(Adia,Bdia,tol,f);CHKERRQ(ierr); 1074 if (!*f) PetscFunctionReturn(0); 1075 ierr = PetscObjectGetComm((PetscObject)Amat,&comm);CHKERRQ(ierr); 1076 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 1077 if (size == 1) PetscFunctionReturn(0); 1078 1079 /* Hard test: off-diagonal block. This takes a MatGetSubMatrix. */ 1080 ierr = MatGetSize(Amat,&M,&N);CHKERRQ(ierr); 1081 ierr = MatGetOwnershipRange(Amat,&first,&last);CHKERRQ(ierr); 1082 ierr = PetscMalloc1(N-last+first,¬me);CHKERRQ(ierr); 1083 for (i=0; i<first; i++) notme[i] = i; 1084 for (i=last; i<M; i++) notme[i-last+first] = i; 1085 ierr = ISCreateGeneral(MPI_COMM_SELF,N-last+first,notme,PETSC_COPY_VALUES,&Notme);CHKERRQ(ierr); 1086 ierr = ISCreateStride(MPI_COMM_SELF,last-first,first,1,&Me);CHKERRQ(ierr); 1087 ierr = MatGetSubMatrices(Amat,1,&Me,&Notme,MAT_INITIAL_MATRIX,&Aoffs);CHKERRQ(ierr); 1088 Aoff = Aoffs[0]; 1089 ierr = MatGetSubMatrices(Bmat,1,&Notme,&Me,MAT_INITIAL_MATRIX,&Boffs);CHKERRQ(ierr); 1090 Boff = Boffs[0]; 1091 ierr = MatIsTranspose(Aoff,Boff,tol,f);CHKERRQ(ierr); 1092 ierr = MatDestroyMatrices(1,&Aoffs);CHKERRQ(ierr); 1093 ierr = MatDestroyMatrices(1,&Boffs);CHKERRQ(ierr); 1094 ierr = ISDestroy(&Me);CHKERRQ(ierr); 1095 ierr = ISDestroy(&Notme);CHKERRQ(ierr); 1096 ierr = PetscFree(notme);CHKERRQ(ierr); 1097 PetscFunctionReturn(0); 1098 } 1099 1100 #undef __FUNCT__ 1101 #define __FUNCT__ "MatMultTransposeAdd_MPIAIJ" 1102 PetscErrorCode MatMultTransposeAdd_MPIAIJ(Mat A,Vec xx,Vec yy,Vec zz) 1103 { 1104 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 1105 PetscErrorCode ierr; 1106 1107 PetscFunctionBegin; 1108 /* do nondiagonal part */ 1109 ierr = (*a->B->ops->multtranspose)(a->B,xx,a->lvec);CHKERRQ(ierr); 1110 /* send it on its way */ 1111 ierr = VecScatterBegin(a->Mvctx,a->lvec,zz,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1112 /* do local part */ 1113 ierr = (*a->A->ops->multtransposeadd)(a->A,xx,yy,zz);CHKERRQ(ierr); 1114 /* receive remote parts */ 1115 ierr = VecScatterEnd(a->Mvctx,a->lvec,zz,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1116 PetscFunctionReturn(0); 1117 } 1118 1119 /* 1120 This only works correctly for square matrices where the subblock A->A is the 1121 diagonal block 1122 */ 1123 #undef __FUNCT__ 1124 #define __FUNCT__ "MatGetDiagonal_MPIAIJ" 1125 PetscErrorCode MatGetDiagonal_MPIAIJ(Mat A,Vec v) 1126 { 1127 PetscErrorCode ierr; 1128 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 1129 1130 PetscFunctionBegin; 1131 if (A->rmap->N != A->cmap->N) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"Supports only square matrix where A->A is diag block"); 1132 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"); 1133 ierr = MatGetDiagonal(a->A,v);CHKERRQ(ierr); 1134 PetscFunctionReturn(0); 1135 } 1136 1137 #undef __FUNCT__ 1138 #define __FUNCT__ "MatScale_MPIAIJ" 1139 PetscErrorCode MatScale_MPIAIJ(Mat A,PetscScalar aa) 1140 { 1141 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 1142 PetscErrorCode ierr; 1143 1144 PetscFunctionBegin; 1145 ierr = MatScale(a->A,aa);CHKERRQ(ierr); 1146 ierr = MatScale(a->B,aa);CHKERRQ(ierr); 1147 PetscFunctionReturn(0); 1148 } 1149 1150 #undef __FUNCT__ 1151 #define __FUNCT__ "MatDestroy_MPIAIJ" 1152 PetscErrorCode MatDestroy_MPIAIJ(Mat mat) 1153 { 1154 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 1155 PetscErrorCode ierr; 1156 1157 PetscFunctionBegin; 1158 #if defined(PETSC_USE_LOG) 1159 PetscLogObjectState((PetscObject)mat,"Rows=%D, Cols=%D",mat->rmap->N,mat->cmap->N); 1160 #endif 1161 ierr = MatStashDestroy_Private(&mat->stash);CHKERRQ(ierr); 1162 ierr = VecDestroy(&aij->diag);CHKERRQ(ierr); 1163 ierr = MatDestroy(&aij->A);CHKERRQ(ierr); 1164 ierr = MatDestroy(&aij->B);CHKERRQ(ierr); 1165 #if defined(PETSC_USE_CTABLE) 1166 ierr = PetscTableDestroy(&aij->colmap);CHKERRQ(ierr); 1167 #else 1168 ierr = PetscFree(aij->colmap);CHKERRQ(ierr); 1169 #endif 1170 ierr = PetscFree(aij->garray);CHKERRQ(ierr); 1171 ierr = VecDestroy(&aij->lvec);CHKERRQ(ierr); 1172 ierr = VecScatterDestroy(&aij->Mvctx);CHKERRQ(ierr); 1173 ierr = PetscFree2(aij->rowvalues,aij->rowindices);CHKERRQ(ierr); 1174 ierr = PetscFree(aij->ld);CHKERRQ(ierr); 1175 ierr = PetscFree(mat->data);CHKERRQ(ierr); 1176 1177 ierr = PetscObjectChangeTypeName((PetscObject)mat,0);CHKERRQ(ierr); 1178 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatStoreValues_C",NULL);CHKERRQ(ierr); 1179 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatRetrieveValues_C",NULL);CHKERRQ(ierr); 1180 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatGetDiagonalBlock_C",NULL);CHKERRQ(ierr); 1181 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatIsTranspose_C",NULL);CHKERRQ(ierr); 1182 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatMPIAIJSetPreallocation_C",NULL);CHKERRQ(ierr); 1183 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatMPIAIJSetPreallocationCSR_C",NULL);CHKERRQ(ierr); 1184 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatDiagonalScaleLocal_C",NULL);CHKERRQ(ierr); 1185 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatConvert_mpiaij_mpisbaij_C",NULL);CHKERRQ(ierr); 1186 #if defined(PETSC_HAVE_ELEMENTAL) 1187 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatConvert_mpiaij_elemental_C",NULL);CHKERRQ(ierr); 1188 #endif 1189 PetscFunctionReturn(0); 1190 } 1191 1192 #undef __FUNCT__ 1193 #define __FUNCT__ "MatView_MPIAIJ_Binary" 1194 PetscErrorCode MatView_MPIAIJ_Binary(Mat mat,PetscViewer viewer) 1195 { 1196 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 1197 Mat_SeqAIJ *A = (Mat_SeqAIJ*)aij->A->data; 1198 Mat_SeqAIJ *B = (Mat_SeqAIJ*)aij->B->data; 1199 PetscErrorCode ierr; 1200 PetscMPIInt rank,size,tag = ((PetscObject)viewer)->tag; 1201 int fd; 1202 PetscInt nz,header[4],*row_lengths,*range=0,rlen,i; 1203 PetscInt nzmax,*column_indices,j,k,col,*garray = aij->garray,cnt,cstart = mat->cmap->rstart,rnz = 0; 1204 PetscScalar *column_values; 1205 PetscInt message_count,flowcontrolcount; 1206 FILE *file; 1207 1208 PetscFunctionBegin; 1209 ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr); 1210 ierr = MPI_Comm_size(PetscObjectComm((PetscObject)mat),&size);CHKERRQ(ierr); 1211 nz = A->nz + B->nz; 1212 ierr = PetscViewerBinaryGetDescriptor(viewer,&fd);CHKERRQ(ierr); 1213 if (!rank) { 1214 header[0] = MAT_FILE_CLASSID; 1215 header[1] = mat->rmap->N; 1216 header[2] = mat->cmap->N; 1217 1218 ierr = MPI_Reduce(&nz,&header[3],1,MPIU_INT,MPI_SUM,0,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1219 ierr = PetscBinaryWrite(fd,header,4,PETSC_INT,PETSC_TRUE);CHKERRQ(ierr); 1220 /* get largest number of rows any processor has */ 1221 rlen = mat->rmap->n; 1222 range = mat->rmap->range; 1223 for (i=1; i<size; i++) rlen = PetscMax(rlen,range[i+1] - range[i]); 1224 } else { 1225 ierr = MPI_Reduce(&nz,0,1,MPIU_INT,MPI_SUM,0,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1226 rlen = mat->rmap->n; 1227 } 1228 1229 /* load up the local row counts */ 1230 ierr = PetscMalloc1(rlen+1,&row_lengths);CHKERRQ(ierr); 1231 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]; 1232 1233 /* store the row lengths to the file */ 1234 ierr = PetscViewerFlowControlStart(viewer,&message_count,&flowcontrolcount);CHKERRQ(ierr); 1235 if (!rank) { 1236 ierr = PetscBinaryWrite(fd,row_lengths,mat->rmap->n,PETSC_INT,PETSC_TRUE);CHKERRQ(ierr); 1237 for (i=1; i<size; i++) { 1238 ierr = PetscViewerFlowControlStepMaster(viewer,i,&message_count,flowcontrolcount);CHKERRQ(ierr); 1239 rlen = range[i+1] - range[i]; 1240 ierr = MPIULong_Recv(row_lengths,rlen,MPIU_INT,i,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1241 ierr = PetscBinaryWrite(fd,row_lengths,rlen,PETSC_INT,PETSC_TRUE);CHKERRQ(ierr); 1242 } 1243 ierr = PetscViewerFlowControlEndMaster(viewer,&message_count);CHKERRQ(ierr); 1244 } else { 1245 ierr = PetscViewerFlowControlStepWorker(viewer,rank,&message_count);CHKERRQ(ierr); 1246 ierr = MPIULong_Send(row_lengths,mat->rmap->n,MPIU_INT,0,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1247 ierr = PetscViewerFlowControlEndWorker(viewer,&message_count);CHKERRQ(ierr); 1248 } 1249 ierr = PetscFree(row_lengths);CHKERRQ(ierr); 1250 1251 /* load up the local column indices */ 1252 nzmax = nz; /* th processor needs space a largest processor needs */ 1253 ierr = MPI_Reduce(&nz,&nzmax,1,MPIU_INT,MPI_MAX,0,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1254 ierr = PetscMalloc1(nzmax+1,&column_indices);CHKERRQ(ierr); 1255 cnt = 0; 1256 for (i=0; i<mat->rmap->n; i++) { 1257 for (j=B->i[i]; j<B->i[i+1]; j++) { 1258 if ((col = garray[B->j[j]]) > cstart) break; 1259 column_indices[cnt++] = col; 1260 } 1261 for (k=A->i[i]; k<A->i[i+1]; k++) column_indices[cnt++] = A->j[k] + cstart; 1262 for (; j<B->i[i+1]; j++) column_indices[cnt++] = garray[B->j[j]]; 1263 } 1264 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); 1265 1266 /* store the column indices to the file */ 1267 ierr = PetscViewerFlowControlStart(viewer,&message_count,&flowcontrolcount);CHKERRQ(ierr); 1268 if (!rank) { 1269 MPI_Status status; 1270 ierr = PetscBinaryWrite(fd,column_indices,nz,PETSC_INT,PETSC_TRUE);CHKERRQ(ierr); 1271 for (i=1; i<size; i++) { 1272 ierr = PetscViewerFlowControlStepMaster(viewer,i,&message_count,flowcontrolcount);CHKERRQ(ierr); 1273 ierr = MPI_Recv(&rnz,1,MPIU_INT,i,tag,PetscObjectComm((PetscObject)mat),&status);CHKERRQ(ierr); 1274 if (rnz > nzmax) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_LIB,"Internal PETSc error: nz = %D nzmax = %D",nz,nzmax); 1275 ierr = MPIULong_Recv(column_indices,rnz,MPIU_INT,i,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1276 ierr = PetscBinaryWrite(fd,column_indices,rnz,PETSC_INT,PETSC_TRUE);CHKERRQ(ierr); 1277 } 1278 ierr = PetscViewerFlowControlEndMaster(viewer,&message_count);CHKERRQ(ierr); 1279 } else { 1280 ierr = PetscViewerFlowControlStepWorker(viewer,rank,&message_count);CHKERRQ(ierr); 1281 ierr = MPI_Send(&nz,1,MPIU_INT,0,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1282 ierr = MPIULong_Send(column_indices,nz,MPIU_INT,0,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1283 ierr = PetscViewerFlowControlEndWorker(viewer,&message_count);CHKERRQ(ierr); 1284 } 1285 ierr = PetscFree(column_indices);CHKERRQ(ierr); 1286 1287 /* load up the local column values */ 1288 ierr = PetscMalloc1(nzmax+1,&column_values);CHKERRQ(ierr); 1289 cnt = 0; 1290 for (i=0; i<mat->rmap->n; i++) { 1291 for (j=B->i[i]; j<B->i[i+1]; j++) { 1292 if (garray[B->j[j]] > cstart) break; 1293 column_values[cnt++] = B->a[j]; 1294 } 1295 for (k=A->i[i]; k<A->i[i+1]; k++) column_values[cnt++] = A->a[k]; 1296 for (; j<B->i[i+1]; j++) column_values[cnt++] = B->a[j]; 1297 } 1298 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); 1299 1300 /* store the column values to the file */ 1301 ierr = PetscViewerFlowControlStart(viewer,&message_count,&flowcontrolcount);CHKERRQ(ierr); 1302 if (!rank) { 1303 MPI_Status status; 1304 ierr = PetscBinaryWrite(fd,column_values,nz,PETSC_SCALAR,PETSC_TRUE);CHKERRQ(ierr); 1305 for (i=1; i<size; i++) { 1306 ierr = PetscViewerFlowControlStepMaster(viewer,i,&message_count,flowcontrolcount);CHKERRQ(ierr); 1307 ierr = MPI_Recv(&rnz,1,MPIU_INT,i,tag,PetscObjectComm((PetscObject)mat),&status);CHKERRQ(ierr); 1308 if (rnz > nzmax) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_LIB,"Internal PETSc error: nz = %D nzmax = %D",nz,nzmax); 1309 ierr = MPIULong_Recv(column_values,rnz,MPIU_SCALAR,i,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1310 ierr = PetscBinaryWrite(fd,column_values,rnz,PETSC_SCALAR,PETSC_TRUE);CHKERRQ(ierr); 1311 } 1312 ierr = PetscViewerFlowControlEndMaster(viewer,&message_count);CHKERRQ(ierr); 1313 } else { 1314 ierr = PetscViewerFlowControlStepWorker(viewer,rank,&message_count);CHKERRQ(ierr); 1315 ierr = MPI_Send(&nz,1,MPIU_INT,0,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1316 ierr = MPIULong_Send(column_values,nz,MPIU_SCALAR,0,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1317 ierr = PetscViewerFlowControlEndWorker(viewer,&message_count);CHKERRQ(ierr); 1318 } 1319 ierr = PetscFree(column_values);CHKERRQ(ierr); 1320 1321 ierr = PetscViewerBinaryGetInfoPointer(viewer,&file);CHKERRQ(ierr); 1322 if (file) fprintf(file,"-matload_block_size %d\n",(int)PetscAbs(mat->rmap->bs)); 1323 PetscFunctionReturn(0); 1324 } 1325 1326 #include <petscdraw.h> 1327 #undef __FUNCT__ 1328 #define __FUNCT__ "MatView_MPIAIJ_ASCIIorDraworSocket" 1329 PetscErrorCode MatView_MPIAIJ_ASCIIorDraworSocket(Mat mat,PetscViewer viewer) 1330 { 1331 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 1332 PetscErrorCode ierr; 1333 PetscMPIInt rank = aij->rank,size = aij->size; 1334 PetscBool isdraw,iascii,isbinary; 1335 PetscViewer sviewer; 1336 PetscViewerFormat format; 1337 1338 PetscFunctionBegin; 1339 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERDRAW,&isdraw);CHKERRQ(ierr); 1340 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii);CHKERRQ(ierr); 1341 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERBINARY,&isbinary);CHKERRQ(ierr); 1342 if (iascii) { 1343 ierr = PetscViewerGetFormat(viewer,&format);CHKERRQ(ierr); 1344 if (format == PETSC_VIEWER_ASCII_INFO_DETAIL) { 1345 MatInfo info; 1346 PetscBool inodes; 1347 1348 ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr); 1349 ierr = MatGetInfo(mat,MAT_LOCAL,&info);CHKERRQ(ierr); 1350 ierr = MatInodeGetInodeSizes(aij->A,NULL,(PetscInt**)&inodes,NULL);CHKERRQ(ierr); 1351 ierr = PetscViewerASCIISynchronizedAllow(viewer,PETSC_TRUE);CHKERRQ(ierr); 1352 if (!inodes) { 1353 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"[%d] Local rows %D nz %D nz alloced %D mem %D, not using I-node routines\n", 1354 rank,mat->rmap->n,(PetscInt)info.nz_used,(PetscInt)info.nz_allocated,(PetscInt)info.memory);CHKERRQ(ierr); 1355 } else { 1356 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"[%d] Local rows %D nz %D nz alloced %D mem %D, using I-node routines\n", 1357 rank,mat->rmap->n,(PetscInt)info.nz_used,(PetscInt)info.nz_allocated,(PetscInt)info.memory);CHKERRQ(ierr); 1358 } 1359 ierr = MatGetInfo(aij->A,MAT_LOCAL,&info);CHKERRQ(ierr); 1360 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"[%d] on-diagonal part: nz %D \n",rank,(PetscInt)info.nz_used);CHKERRQ(ierr); 1361 ierr = MatGetInfo(aij->B,MAT_LOCAL,&info);CHKERRQ(ierr); 1362 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"[%d] off-diagonal part: nz %D \n",rank,(PetscInt)info.nz_used);CHKERRQ(ierr); 1363 ierr = PetscViewerFlush(viewer);CHKERRQ(ierr); 1364 ierr = PetscViewerASCIISynchronizedAllow(viewer,PETSC_FALSE);CHKERRQ(ierr); 1365 ierr = PetscViewerASCIIPrintf(viewer,"Information on VecScatter used in matrix-vector product: \n");CHKERRQ(ierr); 1366 ierr = VecScatterView(aij->Mvctx,viewer);CHKERRQ(ierr); 1367 PetscFunctionReturn(0); 1368 } else if (format == PETSC_VIEWER_ASCII_INFO) { 1369 PetscInt inodecount,inodelimit,*inodes; 1370 ierr = MatInodeGetInodeSizes(aij->A,&inodecount,&inodes,&inodelimit);CHKERRQ(ierr); 1371 if (inodes) { 1372 ierr = PetscViewerASCIIPrintf(viewer,"using I-node (on process 0) routines: found %D nodes, limit used is %D\n",inodecount,inodelimit);CHKERRQ(ierr); 1373 } else { 1374 ierr = PetscViewerASCIIPrintf(viewer,"not using I-node (on process 0) routines\n");CHKERRQ(ierr); 1375 } 1376 PetscFunctionReturn(0); 1377 } else if (format == PETSC_VIEWER_ASCII_FACTOR_INFO) { 1378 PetscFunctionReturn(0); 1379 } 1380 } else if (isbinary) { 1381 if (size == 1) { 1382 ierr = PetscObjectSetName((PetscObject)aij->A,((PetscObject)mat)->name);CHKERRQ(ierr); 1383 ierr = MatView(aij->A,viewer);CHKERRQ(ierr); 1384 } else { 1385 ierr = MatView_MPIAIJ_Binary(mat,viewer);CHKERRQ(ierr); 1386 } 1387 PetscFunctionReturn(0); 1388 } else if (isdraw) { 1389 PetscDraw draw; 1390 PetscBool isnull; 1391 ierr = PetscViewerDrawGetDraw(viewer,0,&draw);CHKERRQ(ierr); 1392 ierr = PetscDrawIsNull(draw,&isnull);CHKERRQ(ierr); if (isnull) PetscFunctionReturn(0); 1393 } 1394 1395 { 1396 /* assemble the entire matrix onto first processor. */ 1397 Mat A; 1398 Mat_SeqAIJ *Aloc; 1399 PetscInt M = mat->rmap->N,N = mat->cmap->N,m,*ai,*aj,row,*cols,i,*ct; 1400 MatScalar *a; 1401 1402 ierr = MatCreate(PetscObjectComm((PetscObject)mat),&A);CHKERRQ(ierr); 1403 if (!rank) { 1404 ierr = MatSetSizes(A,M,N,M,N);CHKERRQ(ierr); 1405 } else { 1406 ierr = MatSetSizes(A,0,0,M,N);CHKERRQ(ierr); 1407 } 1408 /* This is just a temporary matrix, so explicitly using MATMPIAIJ is probably best */ 1409 ierr = MatSetType(A,MATMPIAIJ);CHKERRQ(ierr); 1410 ierr = MatMPIAIJSetPreallocation(A,0,NULL,0,NULL);CHKERRQ(ierr); 1411 ierr = MatSetOption(A,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 1412 ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)A);CHKERRQ(ierr); 1413 1414 /* copy over the A part */ 1415 Aloc = (Mat_SeqAIJ*)aij->A->data; 1416 m = aij->A->rmap->n; ai = Aloc->i; aj = Aloc->j; a = Aloc->a; 1417 row = mat->rmap->rstart; 1418 for (i=0; i<ai[m]; i++) aj[i] += mat->cmap->rstart; 1419 for (i=0; i<m; i++) { 1420 ierr = MatSetValues(A,1,&row,ai[i+1]-ai[i],aj,a,INSERT_VALUES);CHKERRQ(ierr); 1421 row++; 1422 a += ai[i+1]-ai[i]; aj += ai[i+1]-ai[i]; 1423 } 1424 aj = Aloc->j; 1425 for (i=0; i<ai[m]; i++) aj[i] -= mat->cmap->rstart; 1426 1427 /* copy over the B part */ 1428 Aloc = (Mat_SeqAIJ*)aij->B->data; 1429 m = aij->B->rmap->n; ai = Aloc->i; aj = Aloc->j; a = Aloc->a; 1430 row = mat->rmap->rstart; 1431 ierr = PetscMalloc1(ai[m]+1,&cols);CHKERRQ(ierr); 1432 ct = cols; 1433 for (i=0; i<ai[m]; i++) cols[i] = aij->garray[aj[i]]; 1434 for (i=0; i<m; i++) { 1435 ierr = MatSetValues(A,1,&row,ai[i+1]-ai[i],cols,a,INSERT_VALUES);CHKERRQ(ierr); 1436 row++; 1437 a += ai[i+1]-ai[i]; cols += ai[i+1]-ai[i]; 1438 } 1439 ierr = PetscFree(ct);CHKERRQ(ierr); 1440 ierr = MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1441 ierr = MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1442 /* 1443 Everyone has to call to draw the matrix since the graphics waits are 1444 synchronized across all processors that share the PetscDraw object 1445 */ 1446 ierr = PetscViewerGetSingleton(viewer,&sviewer);CHKERRQ(ierr); 1447 if (!rank) { 1448 ierr = PetscObjectSetName((PetscObject)((Mat_MPIAIJ*)(A->data))->A,((PetscObject)mat)->name);CHKERRQ(ierr); 1449 ierr = MatView_SeqAIJ(((Mat_MPIAIJ*)(A->data))->A,sviewer);CHKERRQ(ierr); 1450 } 1451 ierr = PetscViewerRestoreSingleton(viewer,&sviewer);CHKERRQ(ierr); 1452 ierr = MatDestroy(&A);CHKERRQ(ierr); 1453 } 1454 PetscFunctionReturn(0); 1455 } 1456 1457 #undef __FUNCT__ 1458 #define __FUNCT__ "MatView_MPIAIJ" 1459 PetscErrorCode MatView_MPIAIJ(Mat mat,PetscViewer viewer) 1460 { 1461 PetscErrorCode ierr; 1462 PetscBool iascii,isdraw,issocket,isbinary; 1463 1464 PetscFunctionBegin; 1465 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii);CHKERRQ(ierr); 1466 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERDRAW,&isdraw);CHKERRQ(ierr); 1467 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERBINARY,&isbinary);CHKERRQ(ierr); 1468 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERSOCKET,&issocket);CHKERRQ(ierr); 1469 if (iascii || isdraw || isbinary || issocket) { 1470 ierr = MatView_MPIAIJ_ASCIIorDraworSocket(mat,viewer);CHKERRQ(ierr); 1471 } 1472 PetscFunctionReturn(0); 1473 } 1474 1475 #undef __FUNCT__ 1476 #define __FUNCT__ "MatSOR_MPIAIJ" 1477 PetscErrorCode MatSOR_MPIAIJ(Mat matin,Vec bb,PetscReal omega,MatSORType flag,PetscReal fshift,PetscInt its,PetscInt lits,Vec xx) 1478 { 1479 Mat_MPIAIJ *mat = (Mat_MPIAIJ*)matin->data; 1480 PetscErrorCode ierr; 1481 Vec bb1 = 0; 1482 PetscBool hasop; 1483 1484 PetscFunctionBegin; 1485 if (flag == SOR_APPLY_UPPER) { 1486 ierr = (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);CHKERRQ(ierr); 1487 PetscFunctionReturn(0); 1488 } 1489 1490 if (its > 1 || ~flag & SOR_ZERO_INITIAL_GUESS || flag & SOR_EISENSTAT) { 1491 ierr = VecDuplicate(bb,&bb1);CHKERRQ(ierr); 1492 } 1493 1494 if ((flag & SOR_LOCAL_SYMMETRIC_SWEEP) == SOR_LOCAL_SYMMETRIC_SWEEP) { 1495 if (flag & SOR_ZERO_INITIAL_GUESS) { 1496 ierr = (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);CHKERRQ(ierr); 1497 its--; 1498 } 1499 1500 while (its--) { 1501 ierr = VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1502 ierr = VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1503 1504 /* update rhs: bb1 = bb - B*x */ 1505 ierr = VecScale(mat->lvec,-1.0);CHKERRQ(ierr); 1506 ierr = (*mat->B->ops->multadd)(mat->B,mat->lvec,bb,bb1);CHKERRQ(ierr); 1507 1508 /* local sweep */ 1509 ierr = (*mat->A->ops->sor)(mat->A,bb1,omega,SOR_SYMMETRIC_SWEEP,fshift,lits,1,xx);CHKERRQ(ierr); 1510 } 1511 } else if (flag & SOR_LOCAL_FORWARD_SWEEP) { 1512 if (flag & SOR_ZERO_INITIAL_GUESS) { 1513 ierr = (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);CHKERRQ(ierr); 1514 its--; 1515 } 1516 while (its--) { 1517 ierr = VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1518 ierr = VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1519 1520 /* update rhs: bb1 = bb - B*x */ 1521 ierr = VecScale(mat->lvec,-1.0);CHKERRQ(ierr); 1522 ierr = (*mat->B->ops->multadd)(mat->B,mat->lvec,bb,bb1);CHKERRQ(ierr); 1523 1524 /* local sweep */ 1525 ierr = (*mat->A->ops->sor)(mat->A,bb1,omega,SOR_FORWARD_SWEEP,fshift,lits,1,xx);CHKERRQ(ierr); 1526 } 1527 } else if (flag & SOR_LOCAL_BACKWARD_SWEEP) { 1528 if (flag & SOR_ZERO_INITIAL_GUESS) { 1529 ierr = (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);CHKERRQ(ierr); 1530 its--; 1531 } 1532 while (its--) { 1533 ierr = VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1534 ierr = VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1535 1536 /* update rhs: bb1 = bb - B*x */ 1537 ierr = VecScale(mat->lvec,-1.0);CHKERRQ(ierr); 1538 ierr = (*mat->B->ops->multadd)(mat->B,mat->lvec,bb,bb1);CHKERRQ(ierr); 1539 1540 /* local sweep */ 1541 ierr = (*mat->A->ops->sor)(mat->A,bb1,omega,SOR_BACKWARD_SWEEP,fshift,lits,1,xx);CHKERRQ(ierr); 1542 } 1543 } else if (flag & SOR_EISENSTAT) { 1544 Vec xx1; 1545 1546 ierr = VecDuplicate(bb,&xx1);CHKERRQ(ierr); 1547 ierr = (*mat->A->ops->sor)(mat->A,bb,omega,(MatSORType)(SOR_ZERO_INITIAL_GUESS | SOR_LOCAL_BACKWARD_SWEEP),fshift,lits,1,xx);CHKERRQ(ierr); 1548 1549 ierr = VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1550 ierr = VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1551 if (!mat->diag) { 1552 ierr = MatCreateVecs(matin,&mat->diag,NULL);CHKERRQ(ierr); 1553 ierr = MatGetDiagonal(matin,mat->diag);CHKERRQ(ierr); 1554 } 1555 ierr = MatHasOperation(matin,MATOP_MULT_DIAGONAL_BLOCK,&hasop);CHKERRQ(ierr); 1556 if (hasop) { 1557 ierr = MatMultDiagonalBlock(matin,xx,bb1);CHKERRQ(ierr); 1558 } else { 1559 ierr = VecPointwiseMult(bb1,mat->diag,xx);CHKERRQ(ierr); 1560 } 1561 ierr = VecAYPX(bb1,(omega-2.0)/omega,bb);CHKERRQ(ierr); 1562 1563 ierr = MatMultAdd(mat->B,mat->lvec,bb1,bb1);CHKERRQ(ierr); 1564 1565 /* local sweep */ 1566 ierr = (*mat->A->ops->sor)(mat->A,bb1,omega,(MatSORType)(SOR_ZERO_INITIAL_GUESS | SOR_LOCAL_FORWARD_SWEEP),fshift,lits,1,xx1);CHKERRQ(ierr); 1567 ierr = VecAXPY(xx,1.0,xx1);CHKERRQ(ierr); 1568 ierr = VecDestroy(&xx1);CHKERRQ(ierr); 1569 } else SETERRQ(PetscObjectComm((PetscObject)matin),PETSC_ERR_SUP,"Parallel SOR not supported"); 1570 1571 ierr = VecDestroy(&bb1);CHKERRQ(ierr); 1572 PetscFunctionReturn(0); 1573 } 1574 1575 #undef __FUNCT__ 1576 #define __FUNCT__ "MatPermute_MPIAIJ" 1577 PetscErrorCode MatPermute_MPIAIJ(Mat A,IS rowp,IS colp,Mat *B) 1578 { 1579 Mat aA,aB,Aperm; 1580 const PetscInt *rwant,*cwant,*gcols,*ai,*bi,*aj,*bj; 1581 PetscScalar *aa,*ba; 1582 PetscInt i,j,m,n,ng,anz,bnz,*dnnz,*onnz,*tdnnz,*tonnz,*rdest,*cdest,*work,*gcdest; 1583 PetscSF rowsf,sf; 1584 IS parcolp = NULL; 1585 PetscBool done; 1586 PetscErrorCode ierr; 1587 1588 PetscFunctionBegin; 1589 ierr = MatGetLocalSize(A,&m,&n);CHKERRQ(ierr); 1590 ierr = ISGetIndices(rowp,&rwant);CHKERRQ(ierr); 1591 ierr = ISGetIndices(colp,&cwant);CHKERRQ(ierr); 1592 ierr = PetscMalloc3(PetscMax(m,n),&work,m,&rdest,n,&cdest);CHKERRQ(ierr); 1593 1594 /* Invert row permutation to find out where my rows should go */ 1595 ierr = PetscSFCreate(PetscObjectComm((PetscObject)A),&rowsf);CHKERRQ(ierr); 1596 ierr = PetscSFSetGraphLayout(rowsf,A->rmap,A->rmap->n,NULL,PETSC_OWN_POINTER,rwant);CHKERRQ(ierr); 1597 ierr = PetscSFSetFromOptions(rowsf);CHKERRQ(ierr); 1598 for (i=0; i<m; i++) work[i] = A->rmap->rstart + i; 1599 ierr = PetscSFReduceBegin(rowsf,MPIU_INT,work,rdest,MPIU_REPLACE);CHKERRQ(ierr); 1600 ierr = PetscSFReduceEnd(rowsf,MPIU_INT,work,rdest,MPIU_REPLACE);CHKERRQ(ierr); 1601 1602 /* Invert column permutation to find out where my columns should go */ 1603 ierr = PetscSFCreate(PetscObjectComm((PetscObject)A),&sf);CHKERRQ(ierr); 1604 ierr = PetscSFSetGraphLayout(sf,A->cmap,A->cmap->n,NULL,PETSC_OWN_POINTER,cwant);CHKERRQ(ierr); 1605 ierr = PetscSFSetFromOptions(sf);CHKERRQ(ierr); 1606 for (i=0; i<n; i++) work[i] = A->cmap->rstart + i; 1607 ierr = PetscSFReduceBegin(sf,MPIU_INT,work,cdest,MPIU_REPLACE);CHKERRQ(ierr); 1608 ierr = PetscSFReduceEnd(sf,MPIU_INT,work,cdest,MPIU_REPLACE);CHKERRQ(ierr); 1609 ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 1610 1611 ierr = ISRestoreIndices(rowp,&rwant);CHKERRQ(ierr); 1612 ierr = ISRestoreIndices(colp,&cwant);CHKERRQ(ierr); 1613 ierr = MatMPIAIJGetSeqAIJ(A,&aA,&aB,&gcols);CHKERRQ(ierr); 1614 1615 /* Find out where my gcols should go */ 1616 ierr = MatGetSize(aB,NULL,&ng);CHKERRQ(ierr); 1617 ierr = PetscMalloc1(ng,&gcdest);CHKERRQ(ierr); 1618 ierr = PetscSFCreate(PetscObjectComm((PetscObject)A),&sf);CHKERRQ(ierr); 1619 ierr = PetscSFSetGraphLayout(sf,A->cmap,ng,NULL,PETSC_OWN_POINTER,gcols);CHKERRQ(ierr); 1620 ierr = PetscSFSetFromOptions(sf);CHKERRQ(ierr); 1621 ierr = PetscSFBcastBegin(sf,MPIU_INT,cdest,gcdest);CHKERRQ(ierr); 1622 ierr = PetscSFBcastEnd(sf,MPIU_INT,cdest,gcdest);CHKERRQ(ierr); 1623 ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 1624 1625 ierr = PetscCalloc4(m,&dnnz,m,&onnz,m,&tdnnz,m,&tonnz);CHKERRQ(ierr); 1626 ierr = MatGetRowIJ(aA,0,PETSC_FALSE,PETSC_FALSE,&anz,&ai,&aj,&done);CHKERRQ(ierr); 1627 ierr = MatGetRowIJ(aB,0,PETSC_FALSE,PETSC_FALSE,&bnz,&bi,&bj,&done);CHKERRQ(ierr); 1628 for (i=0; i<m; i++) { 1629 PetscInt row = rdest[i],rowner; 1630 ierr = PetscLayoutFindOwner(A->rmap,row,&rowner);CHKERRQ(ierr); 1631 for (j=ai[i]; j<ai[i+1]; j++) { 1632 PetscInt cowner,col = cdest[aj[j]]; 1633 ierr = PetscLayoutFindOwner(A->cmap,col,&cowner);CHKERRQ(ierr); /* Could build an index for the columns to eliminate this search */ 1634 if (rowner == cowner) dnnz[i]++; 1635 else onnz[i]++; 1636 } 1637 for (j=bi[i]; j<bi[i+1]; j++) { 1638 PetscInt cowner,col = gcdest[bj[j]]; 1639 ierr = PetscLayoutFindOwner(A->cmap,col,&cowner);CHKERRQ(ierr); 1640 if (rowner == cowner) dnnz[i]++; 1641 else onnz[i]++; 1642 } 1643 } 1644 ierr = PetscSFBcastBegin(rowsf,MPIU_INT,dnnz,tdnnz);CHKERRQ(ierr); 1645 ierr = PetscSFBcastEnd(rowsf,MPIU_INT,dnnz,tdnnz);CHKERRQ(ierr); 1646 ierr = PetscSFBcastBegin(rowsf,MPIU_INT,onnz,tonnz);CHKERRQ(ierr); 1647 ierr = PetscSFBcastEnd(rowsf,MPIU_INT,onnz,tonnz);CHKERRQ(ierr); 1648 ierr = PetscSFDestroy(&rowsf);CHKERRQ(ierr); 1649 1650 ierr = MatCreateAIJ(PetscObjectComm((PetscObject)A),A->rmap->n,A->cmap->n,A->rmap->N,A->cmap->N,0,tdnnz,0,tonnz,&Aperm);CHKERRQ(ierr); 1651 ierr = MatSeqAIJGetArray(aA,&aa);CHKERRQ(ierr); 1652 ierr = MatSeqAIJGetArray(aB,&ba);CHKERRQ(ierr); 1653 for (i=0; i<m; i++) { 1654 PetscInt *acols = dnnz,*bcols = onnz; /* Repurpose now-unneeded arrays */ 1655 PetscInt j0,rowlen; 1656 rowlen = ai[i+1] - ai[i]; 1657 for (j0=j=0; j<rowlen; j0=j) { /* rowlen could be larger than number of rows m, so sum in batches */ 1658 for ( ; j<PetscMin(rowlen,j0+m); j++) acols[j-j0] = cdest[aj[ai[i]+j]]; 1659 ierr = MatSetValues(Aperm,1,&rdest[i],j-j0,acols,aa+ai[i]+j0,INSERT_VALUES);CHKERRQ(ierr); 1660 } 1661 rowlen = bi[i+1] - bi[i]; 1662 for (j0=j=0; j<rowlen; j0=j) { 1663 for ( ; j<PetscMin(rowlen,j0+m); j++) bcols[j-j0] = gcdest[bj[bi[i]+j]]; 1664 ierr = MatSetValues(Aperm,1,&rdest[i],j-j0,bcols,ba+bi[i]+j0,INSERT_VALUES);CHKERRQ(ierr); 1665 } 1666 } 1667 ierr = MatAssemblyBegin(Aperm,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1668 ierr = MatAssemblyEnd(Aperm,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1669 ierr = MatRestoreRowIJ(aA,0,PETSC_FALSE,PETSC_FALSE,&anz,&ai,&aj,&done);CHKERRQ(ierr); 1670 ierr = MatRestoreRowIJ(aB,0,PETSC_FALSE,PETSC_FALSE,&bnz,&bi,&bj,&done);CHKERRQ(ierr); 1671 ierr = MatSeqAIJRestoreArray(aA,&aa);CHKERRQ(ierr); 1672 ierr = MatSeqAIJRestoreArray(aB,&ba);CHKERRQ(ierr); 1673 ierr = PetscFree4(dnnz,onnz,tdnnz,tonnz);CHKERRQ(ierr); 1674 ierr = PetscFree3(work,rdest,cdest);CHKERRQ(ierr); 1675 ierr = PetscFree(gcdest);CHKERRQ(ierr); 1676 if (parcolp) {ierr = ISDestroy(&colp);CHKERRQ(ierr);} 1677 *B = Aperm; 1678 PetscFunctionReturn(0); 1679 } 1680 1681 #undef __FUNCT__ 1682 #define __FUNCT__ "MatGetInfo_MPIAIJ" 1683 PetscErrorCode MatGetInfo_MPIAIJ(Mat matin,MatInfoType flag,MatInfo *info) 1684 { 1685 Mat_MPIAIJ *mat = (Mat_MPIAIJ*)matin->data; 1686 Mat A = mat->A,B = mat->B; 1687 PetscErrorCode ierr; 1688 PetscReal isend[5],irecv[5]; 1689 1690 PetscFunctionBegin; 1691 info->block_size = 1.0; 1692 ierr = MatGetInfo(A,MAT_LOCAL,info);CHKERRQ(ierr); 1693 1694 isend[0] = info->nz_used; isend[1] = info->nz_allocated; isend[2] = info->nz_unneeded; 1695 isend[3] = info->memory; isend[4] = info->mallocs; 1696 1697 ierr = MatGetInfo(B,MAT_LOCAL,info);CHKERRQ(ierr); 1698 1699 isend[0] += info->nz_used; isend[1] += info->nz_allocated; isend[2] += info->nz_unneeded; 1700 isend[3] += info->memory; isend[4] += info->mallocs; 1701 if (flag == MAT_LOCAL) { 1702 info->nz_used = isend[0]; 1703 info->nz_allocated = isend[1]; 1704 info->nz_unneeded = isend[2]; 1705 info->memory = isend[3]; 1706 info->mallocs = isend[4]; 1707 } else if (flag == MAT_GLOBAL_MAX) { 1708 ierr = MPI_Allreduce(isend,irecv,5,MPIU_REAL,MPIU_MAX,PetscObjectComm((PetscObject)matin));CHKERRQ(ierr); 1709 1710 info->nz_used = irecv[0]; 1711 info->nz_allocated = irecv[1]; 1712 info->nz_unneeded = irecv[2]; 1713 info->memory = irecv[3]; 1714 info->mallocs = irecv[4]; 1715 } else if (flag == MAT_GLOBAL_SUM) { 1716 ierr = MPI_Allreduce(isend,irecv,5,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)matin));CHKERRQ(ierr); 1717 1718 info->nz_used = irecv[0]; 1719 info->nz_allocated = irecv[1]; 1720 info->nz_unneeded = irecv[2]; 1721 info->memory = irecv[3]; 1722 info->mallocs = irecv[4]; 1723 } 1724 info->fill_ratio_given = 0; /* no parallel LU/ILU/Cholesky */ 1725 info->fill_ratio_needed = 0; 1726 info->factor_mallocs = 0; 1727 PetscFunctionReturn(0); 1728 } 1729 1730 #undef __FUNCT__ 1731 #define __FUNCT__ "MatSetOption_MPIAIJ" 1732 PetscErrorCode MatSetOption_MPIAIJ(Mat A,MatOption op,PetscBool flg) 1733 { 1734 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 1735 PetscErrorCode ierr; 1736 1737 PetscFunctionBegin; 1738 switch (op) { 1739 case MAT_NEW_NONZERO_LOCATIONS: 1740 case MAT_NEW_NONZERO_ALLOCATION_ERR: 1741 case MAT_UNUSED_NONZERO_LOCATION_ERR: 1742 case MAT_KEEP_NONZERO_PATTERN: 1743 case MAT_NEW_NONZERO_LOCATION_ERR: 1744 case MAT_USE_INODES: 1745 case MAT_IGNORE_ZERO_ENTRIES: 1746 MatCheckPreallocated(A,1); 1747 ierr = MatSetOption(a->A,op,flg);CHKERRQ(ierr); 1748 ierr = MatSetOption(a->B,op,flg);CHKERRQ(ierr); 1749 break; 1750 case MAT_ROW_ORIENTED: 1751 a->roworiented = flg; 1752 1753 ierr = MatSetOption(a->A,op,flg);CHKERRQ(ierr); 1754 ierr = MatSetOption(a->B,op,flg);CHKERRQ(ierr); 1755 break; 1756 case MAT_NEW_DIAGONALS: 1757 ierr = PetscInfo1(A,"Option %s ignored\n",MatOptions[op]);CHKERRQ(ierr); 1758 break; 1759 case MAT_IGNORE_OFF_PROC_ENTRIES: 1760 a->donotstash = flg; 1761 break; 1762 case MAT_SPD: 1763 A->spd_set = PETSC_TRUE; 1764 A->spd = flg; 1765 if (flg) { 1766 A->symmetric = PETSC_TRUE; 1767 A->structurally_symmetric = PETSC_TRUE; 1768 A->symmetric_set = PETSC_TRUE; 1769 A->structurally_symmetric_set = PETSC_TRUE; 1770 } 1771 break; 1772 case MAT_SYMMETRIC: 1773 ierr = MatSetOption(a->A,op,flg);CHKERRQ(ierr); 1774 break; 1775 case MAT_STRUCTURALLY_SYMMETRIC: 1776 ierr = MatSetOption(a->A,op,flg);CHKERRQ(ierr); 1777 break; 1778 case MAT_HERMITIAN: 1779 ierr = MatSetOption(a->A,op,flg);CHKERRQ(ierr); 1780 break; 1781 case MAT_SYMMETRY_ETERNAL: 1782 ierr = MatSetOption(a->A,op,flg);CHKERRQ(ierr); 1783 break; 1784 default: 1785 SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"unknown option %d",op); 1786 } 1787 PetscFunctionReturn(0); 1788 } 1789 1790 #undef __FUNCT__ 1791 #define __FUNCT__ "MatGetRow_MPIAIJ" 1792 PetscErrorCode MatGetRow_MPIAIJ(Mat matin,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v) 1793 { 1794 Mat_MPIAIJ *mat = (Mat_MPIAIJ*)matin->data; 1795 PetscScalar *vworkA,*vworkB,**pvA,**pvB,*v_p; 1796 PetscErrorCode ierr; 1797 PetscInt i,*cworkA,*cworkB,**pcA,**pcB,cstart = matin->cmap->rstart; 1798 PetscInt nztot,nzA,nzB,lrow,rstart = matin->rmap->rstart,rend = matin->rmap->rend; 1799 PetscInt *cmap,*idx_p; 1800 1801 PetscFunctionBegin; 1802 if (mat->getrowactive) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Already active"); 1803 mat->getrowactive = PETSC_TRUE; 1804 1805 if (!mat->rowvalues && (idx || v)) { 1806 /* 1807 allocate enough space to hold information from the longest row. 1808 */ 1809 Mat_SeqAIJ *Aa = (Mat_SeqAIJ*)mat->A->data,*Ba = (Mat_SeqAIJ*)mat->B->data; 1810 PetscInt max = 1,tmp; 1811 for (i=0; i<matin->rmap->n; i++) { 1812 tmp = Aa->i[i+1] - Aa->i[i] + Ba->i[i+1] - Ba->i[i]; 1813 if (max < tmp) max = tmp; 1814 } 1815 ierr = PetscMalloc2(max,&mat->rowvalues,max,&mat->rowindices);CHKERRQ(ierr); 1816 } 1817 1818 if (row < rstart || row >= rend) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Only local rows"); 1819 lrow = row - rstart; 1820 1821 pvA = &vworkA; pcA = &cworkA; pvB = &vworkB; pcB = &cworkB; 1822 if (!v) {pvA = 0; pvB = 0;} 1823 if (!idx) {pcA = 0; if (!v) pcB = 0;} 1824 ierr = (*mat->A->ops->getrow)(mat->A,lrow,&nzA,pcA,pvA);CHKERRQ(ierr); 1825 ierr = (*mat->B->ops->getrow)(mat->B,lrow,&nzB,pcB,pvB);CHKERRQ(ierr); 1826 nztot = nzA + nzB; 1827 1828 cmap = mat->garray; 1829 if (v || idx) { 1830 if (nztot) { 1831 /* Sort by increasing column numbers, assuming A and B already sorted */ 1832 PetscInt imark = -1; 1833 if (v) { 1834 *v = v_p = mat->rowvalues; 1835 for (i=0; i<nzB; i++) { 1836 if (cmap[cworkB[i]] < cstart) v_p[i] = vworkB[i]; 1837 else break; 1838 } 1839 imark = i; 1840 for (i=0; i<nzA; i++) v_p[imark+i] = vworkA[i]; 1841 for (i=imark; i<nzB; i++) v_p[nzA+i] = vworkB[i]; 1842 } 1843 if (idx) { 1844 *idx = idx_p = mat->rowindices; 1845 if (imark > -1) { 1846 for (i=0; i<imark; i++) { 1847 idx_p[i] = cmap[cworkB[i]]; 1848 } 1849 } else { 1850 for (i=0; i<nzB; i++) { 1851 if (cmap[cworkB[i]] < cstart) idx_p[i] = cmap[cworkB[i]]; 1852 else break; 1853 } 1854 imark = i; 1855 } 1856 for (i=0; i<nzA; i++) idx_p[imark+i] = cstart + cworkA[i]; 1857 for (i=imark; i<nzB; i++) idx_p[nzA+i] = cmap[cworkB[i]]; 1858 } 1859 } else { 1860 if (idx) *idx = 0; 1861 if (v) *v = 0; 1862 } 1863 } 1864 *nz = nztot; 1865 ierr = (*mat->A->ops->restorerow)(mat->A,lrow,&nzA,pcA,pvA);CHKERRQ(ierr); 1866 ierr = (*mat->B->ops->restorerow)(mat->B,lrow,&nzB,pcB,pvB);CHKERRQ(ierr); 1867 PetscFunctionReturn(0); 1868 } 1869 1870 #undef __FUNCT__ 1871 #define __FUNCT__ "MatRestoreRow_MPIAIJ" 1872 PetscErrorCode MatRestoreRow_MPIAIJ(Mat mat,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v) 1873 { 1874 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 1875 1876 PetscFunctionBegin; 1877 if (!aij->getrowactive) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"MatGetRow() must be called first"); 1878 aij->getrowactive = PETSC_FALSE; 1879 PetscFunctionReturn(0); 1880 } 1881 1882 #undef __FUNCT__ 1883 #define __FUNCT__ "MatNorm_MPIAIJ" 1884 PetscErrorCode MatNorm_MPIAIJ(Mat mat,NormType type,PetscReal *norm) 1885 { 1886 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 1887 Mat_SeqAIJ *amat = (Mat_SeqAIJ*)aij->A->data,*bmat = (Mat_SeqAIJ*)aij->B->data; 1888 PetscErrorCode ierr; 1889 PetscInt i,j,cstart = mat->cmap->rstart; 1890 PetscReal sum = 0.0; 1891 MatScalar *v; 1892 1893 PetscFunctionBegin; 1894 if (aij->size == 1) { 1895 ierr = MatNorm(aij->A,type,norm);CHKERRQ(ierr); 1896 } else { 1897 if (type == NORM_FROBENIUS) { 1898 v = amat->a; 1899 for (i=0; i<amat->nz; i++) { 1900 sum += PetscRealPart(PetscConj(*v)*(*v)); v++; 1901 } 1902 v = bmat->a; 1903 for (i=0; i<bmat->nz; i++) { 1904 sum += PetscRealPart(PetscConj(*v)*(*v)); v++; 1905 } 1906 ierr = MPI_Allreduce(&sum,norm,1,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1907 *norm = PetscSqrtReal(*norm); 1908 } else if (type == NORM_1) { /* max column norm */ 1909 PetscReal *tmp,*tmp2; 1910 PetscInt *jj,*garray = aij->garray; 1911 ierr = PetscCalloc1(mat->cmap->N+1,&tmp);CHKERRQ(ierr); 1912 ierr = PetscMalloc1(mat->cmap->N+1,&tmp2);CHKERRQ(ierr); 1913 *norm = 0.0; 1914 v = amat->a; jj = amat->j; 1915 for (j=0; j<amat->nz; j++) { 1916 tmp[cstart + *jj++] += PetscAbsScalar(*v); v++; 1917 } 1918 v = bmat->a; jj = bmat->j; 1919 for (j=0; j<bmat->nz; j++) { 1920 tmp[garray[*jj++]] += PetscAbsScalar(*v); v++; 1921 } 1922 ierr = MPI_Allreduce(tmp,tmp2,mat->cmap->N,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1923 for (j=0; j<mat->cmap->N; j++) { 1924 if (tmp2[j] > *norm) *norm = tmp2[j]; 1925 } 1926 ierr = PetscFree(tmp);CHKERRQ(ierr); 1927 ierr = PetscFree(tmp2);CHKERRQ(ierr); 1928 } else if (type == NORM_INFINITY) { /* max row norm */ 1929 PetscReal ntemp = 0.0; 1930 for (j=0; j<aij->A->rmap->n; j++) { 1931 v = amat->a + amat->i[j]; 1932 sum = 0.0; 1933 for (i=0; i<amat->i[j+1]-amat->i[j]; i++) { 1934 sum += PetscAbsScalar(*v); v++; 1935 } 1936 v = bmat->a + bmat->i[j]; 1937 for (i=0; i<bmat->i[j+1]-bmat->i[j]; i++) { 1938 sum += PetscAbsScalar(*v); v++; 1939 } 1940 if (sum > ntemp) ntemp = sum; 1941 } 1942 ierr = MPI_Allreduce(&ntemp,norm,1,MPIU_REAL,MPIU_MAX,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1943 } else SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"No support for two norm"); 1944 } 1945 PetscFunctionReturn(0); 1946 } 1947 1948 #undef __FUNCT__ 1949 #define __FUNCT__ "MatTranspose_MPIAIJ" 1950 PetscErrorCode MatTranspose_MPIAIJ(Mat A,MatReuse reuse,Mat *matout) 1951 { 1952 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 1953 Mat_SeqAIJ *Aloc=(Mat_SeqAIJ*)a->A->data,*Bloc=(Mat_SeqAIJ*)a->B->data; 1954 PetscErrorCode ierr; 1955 PetscInt M = A->rmap->N,N = A->cmap->N,ma,na,mb,nb,*ai,*aj,*bi,*bj,row,*cols,*cols_tmp,i; 1956 PetscInt cstart = A->cmap->rstart,ncol; 1957 Mat B; 1958 MatScalar *array; 1959 1960 PetscFunctionBegin; 1961 if (reuse == MAT_REUSE_MATRIX && A == *matout && M != N) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_SIZ,"Square matrix only for in-place"); 1962 1963 ma = A->rmap->n; na = A->cmap->n; mb = a->B->rmap->n; nb = a->B->cmap->n; 1964 ai = Aloc->i; aj = Aloc->j; 1965 bi = Bloc->i; bj = Bloc->j; 1966 if (reuse == MAT_INITIAL_MATRIX || *matout == A) { 1967 PetscInt *d_nnz,*g_nnz,*o_nnz; 1968 PetscSFNode *oloc; 1969 PETSC_UNUSED PetscSF sf; 1970 1971 ierr = PetscMalloc4(na,&d_nnz,na,&o_nnz,nb,&g_nnz,nb,&oloc);CHKERRQ(ierr); 1972 /* compute d_nnz for preallocation */ 1973 ierr = PetscMemzero(d_nnz,na*sizeof(PetscInt));CHKERRQ(ierr); 1974 for (i=0; i<ai[ma]; i++) { 1975 d_nnz[aj[i]]++; 1976 aj[i] += cstart; /* global col index to be used by MatSetValues() */ 1977 } 1978 /* compute local off-diagonal contributions */ 1979 ierr = PetscMemzero(g_nnz,nb*sizeof(PetscInt));CHKERRQ(ierr); 1980 for (i=0; i<bi[ma]; i++) g_nnz[bj[i]]++; 1981 /* map those to global */ 1982 ierr = PetscSFCreate(PetscObjectComm((PetscObject)A),&sf);CHKERRQ(ierr); 1983 ierr = PetscSFSetGraphLayout(sf,A->cmap,nb,NULL,PETSC_USE_POINTER,a->garray);CHKERRQ(ierr); 1984 ierr = PetscSFSetFromOptions(sf);CHKERRQ(ierr); 1985 ierr = PetscMemzero(o_nnz,na*sizeof(PetscInt));CHKERRQ(ierr); 1986 ierr = PetscSFReduceBegin(sf,MPIU_INT,g_nnz,o_nnz,MPIU_SUM);CHKERRQ(ierr); 1987 ierr = PetscSFReduceEnd(sf,MPIU_INT,g_nnz,o_nnz,MPIU_SUM);CHKERRQ(ierr); 1988 ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 1989 1990 ierr = MatCreate(PetscObjectComm((PetscObject)A),&B);CHKERRQ(ierr); 1991 ierr = MatSetSizes(B,A->cmap->n,A->rmap->n,N,M);CHKERRQ(ierr); 1992 ierr = MatSetBlockSizes(B,PetscAbs(A->cmap->bs),PetscAbs(A->rmap->bs));CHKERRQ(ierr); 1993 ierr = MatSetType(B,((PetscObject)A)->type_name);CHKERRQ(ierr); 1994 ierr = MatMPIAIJSetPreallocation(B,0,d_nnz,0,o_nnz);CHKERRQ(ierr); 1995 ierr = PetscFree4(d_nnz,o_nnz,g_nnz,oloc);CHKERRQ(ierr); 1996 } else { 1997 B = *matout; 1998 ierr = MatSetOption(B,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_TRUE);CHKERRQ(ierr); 1999 for (i=0; i<ai[ma]; i++) aj[i] += cstart; /* global col index to be used by MatSetValues() */ 2000 } 2001 2002 /* copy over the A part */ 2003 array = Aloc->a; 2004 row = A->rmap->rstart; 2005 for (i=0; i<ma; i++) { 2006 ncol = ai[i+1]-ai[i]; 2007 ierr = MatSetValues(B,ncol,aj,1,&row,array,INSERT_VALUES);CHKERRQ(ierr); 2008 row++; 2009 array += ncol; aj += ncol; 2010 } 2011 aj = Aloc->j; 2012 for (i=0; i<ai[ma]; i++) aj[i] -= cstart; /* resume local col index */ 2013 2014 /* copy over the B part */ 2015 ierr = PetscCalloc1(bi[mb],&cols);CHKERRQ(ierr); 2016 array = Bloc->a; 2017 row = A->rmap->rstart; 2018 for (i=0; i<bi[mb]; i++) cols[i] = a->garray[bj[i]]; 2019 cols_tmp = cols; 2020 for (i=0; i<mb; i++) { 2021 ncol = bi[i+1]-bi[i]; 2022 ierr = MatSetValues(B,ncol,cols_tmp,1,&row,array,INSERT_VALUES);CHKERRQ(ierr); 2023 row++; 2024 array += ncol; cols_tmp += ncol; 2025 } 2026 ierr = PetscFree(cols);CHKERRQ(ierr); 2027 2028 ierr = MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2029 ierr = MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2030 if (reuse == MAT_INITIAL_MATRIX || *matout != A) { 2031 *matout = B; 2032 } else { 2033 ierr = MatHeaderMerge(A,B);CHKERRQ(ierr); 2034 } 2035 PetscFunctionReturn(0); 2036 } 2037 2038 #undef __FUNCT__ 2039 #define __FUNCT__ "MatDiagonalScale_MPIAIJ" 2040 PetscErrorCode MatDiagonalScale_MPIAIJ(Mat mat,Vec ll,Vec rr) 2041 { 2042 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 2043 Mat a = aij->A,b = aij->B; 2044 PetscErrorCode ierr; 2045 PetscInt s1,s2,s3; 2046 2047 PetscFunctionBegin; 2048 ierr = MatGetLocalSize(mat,&s2,&s3);CHKERRQ(ierr); 2049 if (rr) { 2050 ierr = VecGetLocalSize(rr,&s1);CHKERRQ(ierr); 2051 if (s1!=s3) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"right vector non-conforming local size"); 2052 /* Overlap communication with computation. */ 2053 ierr = VecScatterBegin(aij->Mvctx,rr,aij->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2054 } 2055 if (ll) { 2056 ierr = VecGetLocalSize(ll,&s1);CHKERRQ(ierr); 2057 if (s1!=s2) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"left vector non-conforming local size"); 2058 ierr = (*b->ops->diagonalscale)(b,ll,0);CHKERRQ(ierr); 2059 } 2060 /* scale the diagonal block */ 2061 ierr = (*a->ops->diagonalscale)(a,ll,rr);CHKERRQ(ierr); 2062 2063 if (rr) { 2064 /* Do a scatter end and then right scale the off-diagonal block */ 2065 ierr = VecScatterEnd(aij->Mvctx,rr,aij->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2066 ierr = (*b->ops->diagonalscale)(b,0,aij->lvec);CHKERRQ(ierr); 2067 } 2068 PetscFunctionReturn(0); 2069 } 2070 2071 #undef __FUNCT__ 2072 #define __FUNCT__ "MatSetUnfactored_MPIAIJ" 2073 PetscErrorCode MatSetUnfactored_MPIAIJ(Mat A) 2074 { 2075 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 2076 PetscErrorCode ierr; 2077 2078 PetscFunctionBegin; 2079 ierr = MatSetUnfactored(a->A);CHKERRQ(ierr); 2080 PetscFunctionReturn(0); 2081 } 2082 2083 #undef __FUNCT__ 2084 #define __FUNCT__ "MatEqual_MPIAIJ" 2085 PetscErrorCode MatEqual_MPIAIJ(Mat A,Mat B,PetscBool *flag) 2086 { 2087 Mat_MPIAIJ *matB = (Mat_MPIAIJ*)B->data,*matA = (Mat_MPIAIJ*)A->data; 2088 Mat a,b,c,d; 2089 PetscBool flg; 2090 PetscErrorCode ierr; 2091 2092 PetscFunctionBegin; 2093 a = matA->A; b = matA->B; 2094 c = matB->A; d = matB->B; 2095 2096 ierr = MatEqual(a,c,&flg);CHKERRQ(ierr); 2097 if (flg) { 2098 ierr = MatEqual(b,d,&flg);CHKERRQ(ierr); 2099 } 2100 ierr = MPI_Allreduce(&flg,flag,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)A));CHKERRQ(ierr); 2101 PetscFunctionReturn(0); 2102 } 2103 2104 #undef __FUNCT__ 2105 #define __FUNCT__ "MatCopy_MPIAIJ" 2106 PetscErrorCode MatCopy_MPIAIJ(Mat A,Mat B,MatStructure str) 2107 { 2108 PetscErrorCode ierr; 2109 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 2110 Mat_MPIAIJ *b = (Mat_MPIAIJ*)B->data; 2111 2112 PetscFunctionBegin; 2113 /* If the two matrices don't have the same copy implementation, they aren't compatible for fast copy. */ 2114 if ((str != SAME_NONZERO_PATTERN) || (A->ops->copy != B->ops->copy)) { 2115 /* because of the column compression in the off-processor part of the matrix a->B, 2116 the number of columns in a->B and b->B may be different, hence we cannot call 2117 the MatCopy() directly on the two parts. If need be, we can provide a more 2118 efficient copy than the MatCopy_Basic() by first uncompressing the a->B matrices 2119 then copying the submatrices */ 2120 ierr = MatCopy_Basic(A,B,str);CHKERRQ(ierr); 2121 } else { 2122 ierr = MatCopy(a->A,b->A,str);CHKERRQ(ierr); 2123 ierr = MatCopy(a->B,b->B,str);CHKERRQ(ierr); 2124 } 2125 PetscFunctionReturn(0); 2126 } 2127 2128 #undef __FUNCT__ 2129 #define __FUNCT__ "MatSetUp_MPIAIJ" 2130 PetscErrorCode MatSetUp_MPIAIJ(Mat A) 2131 { 2132 PetscErrorCode ierr; 2133 2134 PetscFunctionBegin; 2135 ierr = MatMPIAIJSetPreallocation(A,PETSC_DEFAULT,0,PETSC_DEFAULT,0);CHKERRQ(ierr); 2136 PetscFunctionReturn(0); 2137 } 2138 2139 /* 2140 Computes the number of nonzeros per row needed for preallocation when X and Y 2141 have different nonzero structure. 2142 */ 2143 #undef __FUNCT__ 2144 #define __FUNCT__ "MatAXPYGetPreallocation_MPIX_private" 2145 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) 2146 { 2147 PetscInt i,j,k,nzx,nzy; 2148 2149 PetscFunctionBegin; 2150 /* Set the number of nonzeros in the new matrix */ 2151 for (i=0; i<m; i++) { 2152 const PetscInt *xjj = xj+xi[i],*yjj = yj+yi[i]; 2153 nzx = xi[i+1] - xi[i]; 2154 nzy = yi[i+1] - yi[i]; 2155 nnz[i] = 0; 2156 for (j=0,k=0; j<nzx; j++) { /* Point in X */ 2157 for (; k<nzy && yltog[yjj[k]]<xltog[xjj[j]]; k++) nnz[i]++; /* Catch up to X */ 2158 if (k<nzy && yltog[yjj[k]]==xltog[xjj[j]]) k++; /* Skip duplicate */ 2159 nnz[i]++; 2160 } 2161 for (; k<nzy; k++) nnz[i]++; 2162 } 2163 PetscFunctionReturn(0); 2164 } 2165 2166 /* This is the same as MatAXPYGetPreallocation_SeqAIJ, except that the local-to-global map is provided */ 2167 #undef __FUNCT__ 2168 #define __FUNCT__ "MatAXPYGetPreallocation_MPIAIJ" 2169 static PetscErrorCode MatAXPYGetPreallocation_MPIAIJ(Mat Y,const PetscInt *yltog,Mat X,const PetscInt *xltog,PetscInt *nnz) 2170 { 2171 PetscErrorCode ierr; 2172 PetscInt m = Y->rmap->N; 2173 Mat_SeqAIJ *x = (Mat_SeqAIJ*)X->data; 2174 Mat_SeqAIJ *y = (Mat_SeqAIJ*)Y->data; 2175 2176 PetscFunctionBegin; 2177 ierr = MatAXPYGetPreallocation_MPIX_private(m,x->i,x->j,xltog,y->i,y->j,yltog,nnz);CHKERRQ(ierr); 2178 PetscFunctionReturn(0); 2179 } 2180 2181 #undef __FUNCT__ 2182 #define __FUNCT__ "MatAXPY_MPIAIJ" 2183 PetscErrorCode MatAXPY_MPIAIJ(Mat Y,PetscScalar a,Mat X,MatStructure str) 2184 { 2185 PetscErrorCode ierr; 2186 Mat_MPIAIJ *xx = (Mat_MPIAIJ*)X->data,*yy = (Mat_MPIAIJ*)Y->data; 2187 PetscBLASInt bnz,one=1; 2188 Mat_SeqAIJ *x,*y; 2189 2190 PetscFunctionBegin; 2191 if (str == SAME_NONZERO_PATTERN) { 2192 PetscScalar alpha = a; 2193 x = (Mat_SeqAIJ*)xx->A->data; 2194 ierr = PetscBLASIntCast(x->nz,&bnz);CHKERRQ(ierr); 2195 y = (Mat_SeqAIJ*)yy->A->data; 2196 PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&bnz,&alpha,x->a,&one,y->a,&one)); 2197 x = (Mat_SeqAIJ*)xx->B->data; 2198 y = (Mat_SeqAIJ*)yy->B->data; 2199 ierr = PetscBLASIntCast(x->nz,&bnz);CHKERRQ(ierr); 2200 PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&bnz,&alpha,x->a,&one,y->a,&one)); 2201 ierr = PetscObjectStateIncrease((PetscObject)Y);CHKERRQ(ierr); 2202 } else if (str == SUBSET_NONZERO_PATTERN) { /* nonzeros of X is a subset of Y's */ 2203 ierr = MatAXPY_Basic(Y,a,X,str);CHKERRQ(ierr); 2204 } else { 2205 Mat B; 2206 PetscInt *nnz_d,*nnz_o; 2207 ierr = PetscMalloc1(yy->A->rmap->N,&nnz_d);CHKERRQ(ierr); 2208 ierr = PetscMalloc1(yy->B->rmap->N,&nnz_o);CHKERRQ(ierr); 2209 ierr = MatCreate(PetscObjectComm((PetscObject)Y),&B);CHKERRQ(ierr); 2210 ierr = PetscObjectSetName((PetscObject)B,((PetscObject)Y)->name);CHKERRQ(ierr); 2211 ierr = MatSetSizes(B,Y->rmap->n,Y->cmap->n,Y->rmap->N,Y->cmap->N);CHKERRQ(ierr); 2212 ierr = MatSetBlockSizesFromMats(B,Y,Y);CHKERRQ(ierr); 2213 ierr = MatSetType(B,MATMPIAIJ);CHKERRQ(ierr); 2214 ierr = MatAXPYGetPreallocation_SeqAIJ(yy->A,xx->A,nnz_d);CHKERRQ(ierr); 2215 ierr = MatAXPYGetPreallocation_MPIAIJ(yy->B,yy->garray,xx->B,xx->garray,nnz_o);CHKERRQ(ierr); 2216 ierr = MatMPIAIJSetPreallocation(B,0,nnz_d,0,nnz_o);CHKERRQ(ierr); 2217 ierr = MatAXPY_BasicWithPreallocation(B,Y,a,X,str);CHKERRQ(ierr); 2218 ierr = MatHeaderReplace(Y,B);CHKERRQ(ierr); 2219 ierr = PetscFree(nnz_d);CHKERRQ(ierr); 2220 ierr = PetscFree(nnz_o);CHKERRQ(ierr); 2221 } 2222 PetscFunctionReturn(0); 2223 } 2224 2225 extern PetscErrorCode MatConjugate_SeqAIJ(Mat); 2226 2227 #undef __FUNCT__ 2228 #define __FUNCT__ "MatConjugate_MPIAIJ" 2229 PetscErrorCode MatConjugate_MPIAIJ(Mat mat) 2230 { 2231 #if defined(PETSC_USE_COMPLEX) 2232 PetscErrorCode ierr; 2233 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 2234 2235 PetscFunctionBegin; 2236 ierr = MatConjugate_SeqAIJ(aij->A);CHKERRQ(ierr); 2237 ierr = MatConjugate_SeqAIJ(aij->B);CHKERRQ(ierr); 2238 #else 2239 PetscFunctionBegin; 2240 #endif 2241 PetscFunctionReturn(0); 2242 } 2243 2244 #undef __FUNCT__ 2245 #define __FUNCT__ "MatRealPart_MPIAIJ" 2246 PetscErrorCode MatRealPart_MPIAIJ(Mat A) 2247 { 2248 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 2249 PetscErrorCode ierr; 2250 2251 PetscFunctionBegin; 2252 ierr = MatRealPart(a->A);CHKERRQ(ierr); 2253 ierr = MatRealPart(a->B);CHKERRQ(ierr); 2254 PetscFunctionReturn(0); 2255 } 2256 2257 #undef __FUNCT__ 2258 #define __FUNCT__ "MatImaginaryPart_MPIAIJ" 2259 PetscErrorCode MatImaginaryPart_MPIAIJ(Mat A) 2260 { 2261 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 2262 PetscErrorCode ierr; 2263 2264 PetscFunctionBegin; 2265 ierr = MatImaginaryPart(a->A);CHKERRQ(ierr); 2266 ierr = MatImaginaryPart(a->B);CHKERRQ(ierr); 2267 PetscFunctionReturn(0); 2268 } 2269 2270 #if defined(PETSC_HAVE_PBGL) 2271 2272 #include <boost/parallel/mpi/bsp_process_group.hpp> 2273 #include <boost/graph/distributed/ilu_default_graph.hpp> 2274 #include <boost/graph/distributed/ilu_0_block.hpp> 2275 #include <boost/graph/distributed/ilu_preconditioner.hpp> 2276 #include <boost/graph/distributed/petsc/interface.hpp> 2277 #include <boost/multi_array.hpp> 2278 #include <boost/parallel/distributed_property_map->hpp> 2279 2280 #undef __FUNCT__ 2281 #define __FUNCT__ "MatILUFactorSymbolic_MPIAIJ" 2282 /* 2283 This uses the parallel ILU factorization of Peter Gottschling <pgottsch@osl.iu.edu> 2284 */ 2285 PetscErrorCode MatILUFactorSymbolic_MPIAIJ(Mat fact,Mat A, IS isrow, IS iscol, const MatFactorInfo *info) 2286 { 2287 namespace petsc = boost::distributed::petsc; 2288 2289 namespace graph_dist = boost::graph::distributed; 2290 using boost::graph::distributed::ilu_default::process_group_type; 2291 using boost::graph::ilu_permuted; 2292 2293 PetscBool row_identity, col_identity; 2294 PetscContainer c; 2295 PetscInt m, n, M, N; 2296 PetscErrorCode ierr; 2297 2298 PetscFunctionBegin; 2299 if (info->levels != 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Only levels = 0 supported for parallel ilu"); 2300 ierr = ISIdentity(isrow, &row_identity);CHKERRQ(ierr); 2301 ierr = ISIdentity(iscol, &col_identity);CHKERRQ(ierr); 2302 if (!row_identity || !col_identity) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Row and column permutations must be identity for parallel ILU"); 2303 2304 process_group_type pg; 2305 typedef graph_dist::ilu_default::ilu_level_graph_type lgraph_type; 2306 lgraph_type *lgraph_p = new lgraph_type(petsc::num_global_vertices(A), pg, petsc::matrix_distribution(A, pg)); 2307 lgraph_type& level_graph = *lgraph_p; 2308 graph_dist::ilu_default::graph_type& graph(level_graph.graph); 2309 2310 petsc::read_matrix(A, graph, get(boost::edge_weight, graph)); 2311 ilu_permuted(level_graph); 2312 2313 /* put together the new matrix */ 2314 ierr = MatCreate(PetscObjectComm((PetscObject)A), fact);CHKERRQ(ierr); 2315 ierr = MatGetLocalSize(A, &m, &n);CHKERRQ(ierr); 2316 ierr = MatGetSize(A, &M, &N);CHKERRQ(ierr); 2317 ierr = MatSetSizes(fact, m, n, M, N);CHKERRQ(ierr); 2318 ierr = MatSetBlockSizesFromMats(fact,A,A);CHKERRQ(ierr); 2319 ierr = MatSetType(fact, ((PetscObject)A)->type_name);CHKERRQ(ierr); 2320 ierr = MatAssemblyBegin(fact, MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2321 ierr = MatAssemblyEnd(fact, MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2322 2323 ierr = PetscContainerCreate(PetscObjectComm((PetscObject)A), &c); 2324 ierr = PetscContainerSetPointer(c, lgraph_p); 2325 ierr = PetscObjectCompose((PetscObject) (fact), "graph", (PetscObject) c); 2326 ierr = PetscContainerDestroy(&c); 2327 PetscFunctionReturn(0); 2328 } 2329 2330 #undef __FUNCT__ 2331 #define __FUNCT__ "MatLUFactorNumeric_MPIAIJ" 2332 PetscErrorCode MatLUFactorNumeric_MPIAIJ(Mat B,Mat A, const MatFactorInfo *info) 2333 { 2334 PetscFunctionBegin; 2335 PetscFunctionReturn(0); 2336 } 2337 2338 #undef __FUNCT__ 2339 #define __FUNCT__ "MatSolve_MPIAIJ" 2340 /* 2341 This uses the parallel ILU factorization of Peter Gottschling <pgottsch@osl.iu.edu> 2342 */ 2343 PetscErrorCode MatSolve_MPIAIJ(Mat A, Vec b, Vec x) 2344 { 2345 namespace graph_dist = boost::graph::distributed; 2346 2347 typedef graph_dist::ilu_default::ilu_level_graph_type lgraph_type; 2348 lgraph_type *lgraph_p; 2349 PetscContainer c; 2350 PetscErrorCode ierr; 2351 2352 PetscFunctionBegin; 2353 ierr = PetscObjectQuery((PetscObject) A, "graph", (PetscObject*) &c);CHKERRQ(ierr); 2354 ierr = PetscContainerGetPointer(c, (void**) &lgraph_p);CHKERRQ(ierr); 2355 ierr = VecCopy(b, x);CHKERRQ(ierr); 2356 2357 PetscScalar *array_x; 2358 ierr = VecGetArray(x, &array_x);CHKERRQ(ierr); 2359 PetscInt sx; 2360 ierr = VecGetSize(x, &sx);CHKERRQ(ierr); 2361 2362 PetscScalar *array_b; 2363 ierr = VecGetArray(b, &array_b);CHKERRQ(ierr); 2364 PetscInt sb; 2365 ierr = VecGetSize(b, &sb);CHKERRQ(ierr); 2366 2367 lgraph_type& level_graph = *lgraph_p; 2368 graph_dist::ilu_default::graph_type& graph(level_graph.graph); 2369 2370 typedef boost::multi_array_ref<PetscScalar, 1> array_ref_type; 2371 array_ref_type ref_b(array_b, boost::extents[num_vertices(graph)]); 2372 array_ref_type ref_x(array_x, boost::extents[num_vertices(graph)]); 2373 2374 typedef boost::iterator_property_map<array_ref_type::iterator, 2375 boost::property_map<graph_dist::ilu_default::graph_type, boost::vertex_index_t>::type> gvector_type; 2376 gvector_type vector_b(ref_b.begin(), get(boost::vertex_index, graph)); 2377 gvector_type vector_x(ref_x.begin(), get(boost::vertex_index, graph)); 2378 2379 ilu_set_solve(*lgraph_p, vector_b, vector_x); 2380 PetscFunctionReturn(0); 2381 } 2382 #endif 2383 2384 #undef __FUNCT__ 2385 #define __FUNCT__ "MatGetRowMaxAbs_MPIAIJ" 2386 PetscErrorCode MatGetRowMaxAbs_MPIAIJ(Mat A, Vec v, PetscInt idx[]) 2387 { 2388 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 2389 PetscErrorCode ierr; 2390 PetscInt i,*idxb = 0; 2391 PetscScalar *va,*vb; 2392 Vec vtmp; 2393 2394 PetscFunctionBegin; 2395 ierr = MatGetRowMaxAbs(a->A,v,idx);CHKERRQ(ierr); 2396 ierr = VecGetArray(v,&va);CHKERRQ(ierr); 2397 if (idx) { 2398 for (i=0; i<A->rmap->n; i++) { 2399 if (PetscAbsScalar(va[i])) idx[i] += A->cmap->rstart; 2400 } 2401 } 2402 2403 ierr = VecCreateSeq(PETSC_COMM_SELF,A->rmap->n,&vtmp);CHKERRQ(ierr); 2404 if (idx) { 2405 ierr = PetscMalloc1(A->rmap->n,&idxb);CHKERRQ(ierr); 2406 } 2407 ierr = MatGetRowMaxAbs(a->B,vtmp,idxb);CHKERRQ(ierr); 2408 ierr = VecGetArray(vtmp,&vb);CHKERRQ(ierr); 2409 2410 for (i=0; i<A->rmap->n; i++) { 2411 if (PetscAbsScalar(va[i]) < PetscAbsScalar(vb[i])) { 2412 va[i] = vb[i]; 2413 if (idx) idx[i] = a->garray[idxb[i]]; 2414 } 2415 } 2416 2417 ierr = VecRestoreArray(v,&va);CHKERRQ(ierr); 2418 ierr = VecRestoreArray(vtmp,&vb);CHKERRQ(ierr); 2419 ierr = PetscFree(idxb);CHKERRQ(ierr); 2420 ierr = VecDestroy(&vtmp);CHKERRQ(ierr); 2421 PetscFunctionReturn(0); 2422 } 2423 2424 #undef __FUNCT__ 2425 #define __FUNCT__ "MatGetRowMinAbs_MPIAIJ" 2426 PetscErrorCode MatGetRowMinAbs_MPIAIJ(Mat A, Vec v, PetscInt idx[]) 2427 { 2428 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 2429 PetscErrorCode ierr; 2430 PetscInt i,*idxb = 0; 2431 PetscScalar *va,*vb; 2432 Vec vtmp; 2433 2434 PetscFunctionBegin; 2435 ierr = MatGetRowMinAbs(a->A,v,idx);CHKERRQ(ierr); 2436 ierr = VecGetArray(v,&va);CHKERRQ(ierr); 2437 if (idx) { 2438 for (i=0; i<A->cmap->n; i++) { 2439 if (PetscAbsScalar(va[i])) idx[i] += A->cmap->rstart; 2440 } 2441 } 2442 2443 ierr = VecCreateSeq(PETSC_COMM_SELF,A->rmap->n,&vtmp);CHKERRQ(ierr); 2444 if (idx) { 2445 ierr = PetscMalloc1(A->rmap->n,&idxb);CHKERRQ(ierr); 2446 } 2447 ierr = MatGetRowMinAbs(a->B,vtmp,idxb);CHKERRQ(ierr); 2448 ierr = VecGetArray(vtmp,&vb);CHKERRQ(ierr); 2449 2450 for (i=0; i<A->rmap->n; i++) { 2451 if (PetscAbsScalar(va[i]) > PetscAbsScalar(vb[i])) { 2452 va[i] = vb[i]; 2453 if (idx) idx[i] = a->garray[idxb[i]]; 2454 } 2455 } 2456 2457 ierr = VecRestoreArray(v,&va);CHKERRQ(ierr); 2458 ierr = VecRestoreArray(vtmp,&vb);CHKERRQ(ierr); 2459 ierr = PetscFree(idxb);CHKERRQ(ierr); 2460 ierr = VecDestroy(&vtmp);CHKERRQ(ierr); 2461 PetscFunctionReturn(0); 2462 } 2463 2464 #undef __FUNCT__ 2465 #define __FUNCT__ "MatGetRowMin_MPIAIJ" 2466 PetscErrorCode MatGetRowMin_MPIAIJ(Mat A, Vec v, PetscInt idx[]) 2467 { 2468 Mat_MPIAIJ *mat = (Mat_MPIAIJ*) A->data; 2469 PetscInt n = A->rmap->n; 2470 PetscInt cstart = A->cmap->rstart; 2471 PetscInt *cmap = mat->garray; 2472 PetscInt *diagIdx, *offdiagIdx; 2473 Vec diagV, offdiagV; 2474 PetscScalar *a, *diagA, *offdiagA; 2475 PetscInt r; 2476 PetscErrorCode ierr; 2477 2478 PetscFunctionBegin; 2479 ierr = PetscMalloc2(n,&diagIdx,n,&offdiagIdx);CHKERRQ(ierr); 2480 ierr = VecCreateSeq(PetscObjectComm((PetscObject)A), n, &diagV);CHKERRQ(ierr); 2481 ierr = VecCreateSeq(PetscObjectComm((PetscObject)A), n, &offdiagV);CHKERRQ(ierr); 2482 ierr = MatGetRowMin(mat->A, diagV, diagIdx);CHKERRQ(ierr); 2483 ierr = MatGetRowMin(mat->B, offdiagV, offdiagIdx);CHKERRQ(ierr); 2484 ierr = VecGetArray(v, &a);CHKERRQ(ierr); 2485 ierr = VecGetArray(diagV, &diagA);CHKERRQ(ierr); 2486 ierr = VecGetArray(offdiagV, &offdiagA);CHKERRQ(ierr); 2487 for (r = 0; r < n; ++r) { 2488 if (PetscAbsScalar(diagA[r]) <= PetscAbsScalar(offdiagA[r])) { 2489 a[r] = diagA[r]; 2490 idx[r] = cstart + diagIdx[r]; 2491 } else { 2492 a[r] = offdiagA[r]; 2493 idx[r] = cmap[offdiagIdx[r]]; 2494 } 2495 } 2496 ierr = VecRestoreArray(v, &a);CHKERRQ(ierr); 2497 ierr = VecRestoreArray(diagV, &diagA);CHKERRQ(ierr); 2498 ierr = VecRestoreArray(offdiagV, &offdiagA);CHKERRQ(ierr); 2499 ierr = VecDestroy(&diagV);CHKERRQ(ierr); 2500 ierr = VecDestroy(&offdiagV);CHKERRQ(ierr); 2501 ierr = PetscFree2(diagIdx, offdiagIdx);CHKERRQ(ierr); 2502 PetscFunctionReturn(0); 2503 } 2504 2505 #undef __FUNCT__ 2506 #define __FUNCT__ "MatGetRowMax_MPIAIJ" 2507 PetscErrorCode MatGetRowMax_MPIAIJ(Mat A, Vec v, PetscInt idx[]) 2508 { 2509 Mat_MPIAIJ *mat = (Mat_MPIAIJ*) A->data; 2510 PetscInt n = A->rmap->n; 2511 PetscInt cstart = A->cmap->rstart; 2512 PetscInt *cmap = mat->garray; 2513 PetscInt *diagIdx, *offdiagIdx; 2514 Vec diagV, offdiagV; 2515 PetscScalar *a, *diagA, *offdiagA; 2516 PetscInt r; 2517 PetscErrorCode ierr; 2518 2519 PetscFunctionBegin; 2520 ierr = PetscMalloc2(n,&diagIdx,n,&offdiagIdx);CHKERRQ(ierr); 2521 ierr = VecCreateSeq(PETSC_COMM_SELF, n, &diagV);CHKERRQ(ierr); 2522 ierr = VecCreateSeq(PETSC_COMM_SELF, n, &offdiagV);CHKERRQ(ierr); 2523 ierr = MatGetRowMax(mat->A, diagV, diagIdx);CHKERRQ(ierr); 2524 ierr = MatGetRowMax(mat->B, offdiagV, offdiagIdx);CHKERRQ(ierr); 2525 ierr = VecGetArray(v, &a);CHKERRQ(ierr); 2526 ierr = VecGetArray(diagV, &diagA);CHKERRQ(ierr); 2527 ierr = VecGetArray(offdiagV, &offdiagA);CHKERRQ(ierr); 2528 for (r = 0; r < n; ++r) { 2529 if (PetscAbsScalar(diagA[r]) >= PetscAbsScalar(offdiagA[r])) { 2530 a[r] = diagA[r]; 2531 idx[r] = cstart + diagIdx[r]; 2532 } else { 2533 a[r] = offdiagA[r]; 2534 idx[r] = cmap[offdiagIdx[r]]; 2535 } 2536 } 2537 ierr = VecRestoreArray(v, &a);CHKERRQ(ierr); 2538 ierr = VecRestoreArray(diagV, &diagA);CHKERRQ(ierr); 2539 ierr = VecRestoreArray(offdiagV, &offdiagA);CHKERRQ(ierr); 2540 ierr = VecDestroy(&diagV);CHKERRQ(ierr); 2541 ierr = VecDestroy(&offdiagV);CHKERRQ(ierr); 2542 ierr = PetscFree2(diagIdx, offdiagIdx);CHKERRQ(ierr); 2543 PetscFunctionReturn(0); 2544 } 2545 2546 #undef __FUNCT__ 2547 #define __FUNCT__ "MatGetSeqNonzeroStructure_MPIAIJ" 2548 PetscErrorCode MatGetSeqNonzeroStructure_MPIAIJ(Mat mat,Mat *newmat) 2549 { 2550 PetscErrorCode ierr; 2551 Mat *dummy; 2552 2553 PetscFunctionBegin; 2554 ierr = MatGetSubMatrix_MPIAIJ_All(mat,MAT_DO_NOT_GET_VALUES,MAT_INITIAL_MATRIX,&dummy);CHKERRQ(ierr); 2555 *newmat = *dummy; 2556 ierr = PetscFree(dummy);CHKERRQ(ierr); 2557 PetscFunctionReturn(0); 2558 } 2559 2560 #undef __FUNCT__ 2561 #define __FUNCT__ "MatInvertBlockDiagonal_MPIAIJ" 2562 PetscErrorCode MatInvertBlockDiagonal_MPIAIJ(Mat A,const PetscScalar **values) 2563 { 2564 Mat_MPIAIJ *a = (Mat_MPIAIJ*) A->data; 2565 PetscErrorCode ierr; 2566 2567 PetscFunctionBegin; 2568 ierr = MatInvertBlockDiagonal(a->A,values);CHKERRQ(ierr); 2569 PetscFunctionReturn(0); 2570 } 2571 2572 #undef __FUNCT__ 2573 #define __FUNCT__ "MatSetRandom_MPIAIJ" 2574 static PetscErrorCode MatSetRandom_MPIAIJ(Mat x,PetscRandom rctx) 2575 { 2576 PetscErrorCode ierr; 2577 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)x->data; 2578 2579 PetscFunctionBegin; 2580 ierr = MatSetRandom(aij->A,rctx);CHKERRQ(ierr); 2581 ierr = MatSetRandom(aij->B,rctx);CHKERRQ(ierr); 2582 ierr = MatAssemblyBegin(x,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2583 ierr = MatAssemblyEnd(x,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2584 PetscFunctionReturn(0); 2585 } 2586 2587 #undef __FUNCT__ 2588 #define __FUNCT__ "MatShift_MPIAIJ" 2589 PetscErrorCode MatShift_MPIAIJ(Mat Y,PetscScalar a) 2590 { 2591 PetscErrorCode ierr; 2592 Mat_MPIAIJ *maij = (Mat_MPIAIJ*)Y->data; 2593 Mat_SeqAIJ *aij = (Mat_SeqAIJ*)maij->A->data,*bij = (Mat_SeqAIJ*)maij->B->data; 2594 2595 PetscFunctionBegin; 2596 if (!aij->nz && !bij->nz) { 2597 ierr = MatMPIAIJSetPreallocation(Y,1,NULL,0,NULL);CHKERRQ(ierr); 2598 } 2599 ierr = MatShift_Basic(Y,a);CHKERRQ(ierr); 2600 PetscFunctionReturn(0); 2601 } 2602 2603 /* -------------------------------------------------------------------*/ 2604 static struct _MatOps MatOps_Values = {MatSetValues_MPIAIJ, 2605 MatGetRow_MPIAIJ, 2606 MatRestoreRow_MPIAIJ, 2607 MatMult_MPIAIJ, 2608 /* 4*/ MatMultAdd_MPIAIJ, 2609 MatMultTranspose_MPIAIJ, 2610 MatMultTransposeAdd_MPIAIJ, 2611 #if defined(PETSC_HAVE_PBGL) 2612 MatSolve_MPIAIJ, 2613 #else 2614 0, 2615 #endif 2616 0, 2617 0, 2618 /*10*/ 0, 2619 0, 2620 0, 2621 MatSOR_MPIAIJ, 2622 MatTranspose_MPIAIJ, 2623 /*15*/ MatGetInfo_MPIAIJ, 2624 MatEqual_MPIAIJ, 2625 MatGetDiagonal_MPIAIJ, 2626 MatDiagonalScale_MPIAIJ, 2627 MatNorm_MPIAIJ, 2628 /*20*/ MatAssemblyBegin_MPIAIJ, 2629 MatAssemblyEnd_MPIAIJ, 2630 MatSetOption_MPIAIJ, 2631 MatZeroEntries_MPIAIJ, 2632 /*24*/ MatZeroRows_MPIAIJ, 2633 0, 2634 #if defined(PETSC_HAVE_PBGL) 2635 0, 2636 #else 2637 0, 2638 #endif 2639 0, 2640 0, 2641 /*29*/ MatSetUp_MPIAIJ, 2642 #if defined(PETSC_HAVE_PBGL) 2643 0, 2644 #else 2645 0, 2646 #endif 2647 0, 2648 0, 2649 0, 2650 /*34*/ MatDuplicate_MPIAIJ, 2651 0, 2652 0, 2653 0, 2654 0, 2655 /*39*/ MatAXPY_MPIAIJ, 2656 MatGetSubMatrices_MPIAIJ, 2657 MatIncreaseOverlap_MPIAIJ, 2658 MatGetValues_MPIAIJ, 2659 MatCopy_MPIAIJ, 2660 /*44*/ MatGetRowMax_MPIAIJ, 2661 MatScale_MPIAIJ, 2662 MatShift_MPIAIJ, 2663 MatDiagonalSet_MPIAIJ, 2664 MatZeroRowsColumns_MPIAIJ, 2665 /*49*/ MatSetRandom_MPIAIJ, 2666 0, 2667 0, 2668 0, 2669 0, 2670 /*54*/ MatFDColoringCreate_MPIXAIJ, 2671 0, 2672 MatSetUnfactored_MPIAIJ, 2673 MatPermute_MPIAIJ, 2674 0, 2675 /*59*/ MatGetSubMatrix_MPIAIJ, 2676 MatDestroy_MPIAIJ, 2677 MatView_MPIAIJ, 2678 0, 2679 MatMatMatMult_MPIAIJ_MPIAIJ_MPIAIJ, 2680 /*64*/ MatMatMatMultSymbolic_MPIAIJ_MPIAIJ_MPIAIJ, 2681 MatMatMatMultNumeric_MPIAIJ_MPIAIJ_MPIAIJ, 2682 0, 2683 0, 2684 0, 2685 /*69*/ MatGetRowMaxAbs_MPIAIJ, 2686 MatGetRowMinAbs_MPIAIJ, 2687 0, 2688 MatSetColoring_MPIAIJ, 2689 0, 2690 MatSetValuesAdifor_MPIAIJ, 2691 /*75*/ MatFDColoringApply_AIJ, 2692 0, 2693 0, 2694 0, 2695 MatFindZeroDiagonals_MPIAIJ, 2696 /*80*/ 0, 2697 0, 2698 0, 2699 /*83*/ MatLoad_MPIAIJ, 2700 0, 2701 0, 2702 0, 2703 0, 2704 0, 2705 /*89*/ MatMatMult_MPIAIJ_MPIAIJ, 2706 MatMatMultSymbolic_MPIAIJ_MPIAIJ, 2707 MatMatMultNumeric_MPIAIJ_MPIAIJ, 2708 MatPtAP_MPIAIJ_MPIAIJ, 2709 MatPtAPSymbolic_MPIAIJ_MPIAIJ, 2710 /*94*/ MatPtAPNumeric_MPIAIJ_MPIAIJ, 2711 0, 2712 0, 2713 0, 2714 0, 2715 /*99*/ 0, 2716 0, 2717 0, 2718 MatConjugate_MPIAIJ, 2719 0, 2720 /*104*/MatSetValuesRow_MPIAIJ, 2721 MatRealPart_MPIAIJ, 2722 MatImaginaryPart_MPIAIJ, 2723 0, 2724 0, 2725 /*109*/0, 2726 0, 2727 MatGetRowMin_MPIAIJ, 2728 0, 2729 0, 2730 /*114*/MatGetSeqNonzeroStructure_MPIAIJ, 2731 0, 2732 0, 2733 0, 2734 0, 2735 /*119*/0, 2736 0, 2737 0, 2738 0, 2739 MatGetMultiProcBlock_MPIAIJ, 2740 /*124*/MatFindNonzeroRows_MPIAIJ, 2741 MatGetColumnNorms_MPIAIJ, 2742 MatInvertBlockDiagonal_MPIAIJ, 2743 0, 2744 MatGetSubMatricesParallel_MPIAIJ, 2745 /*129*/0, 2746 MatTransposeMatMult_MPIAIJ_MPIAIJ, 2747 MatTransposeMatMultSymbolic_MPIAIJ_MPIAIJ, 2748 MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ, 2749 0, 2750 /*134*/0, 2751 0, 2752 0, 2753 0, 2754 0, 2755 /*139*/0, 2756 0, 2757 0, 2758 MatFDColoringSetUp_MPIXAIJ, 2759 MatFindOffBlockDiagonalEntries_MPIAIJ, 2760 /*144*/MatCreateMPIMatConcatenateSeqMat_MPIAIJ 2761 }; 2762 2763 /* ----------------------------------------------------------------------------------------*/ 2764 2765 #undef __FUNCT__ 2766 #define __FUNCT__ "MatStoreValues_MPIAIJ" 2767 PetscErrorCode MatStoreValues_MPIAIJ(Mat mat) 2768 { 2769 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 2770 PetscErrorCode ierr; 2771 2772 PetscFunctionBegin; 2773 ierr = MatStoreValues(aij->A);CHKERRQ(ierr); 2774 ierr = MatStoreValues(aij->B);CHKERRQ(ierr); 2775 PetscFunctionReturn(0); 2776 } 2777 2778 #undef __FUNCT__ 2779 #define __FUNCT__ "MatRetrieveValues_MPIAIJ" 2780 PetscErrorCode MatRetrieveValues_MPIAIJ(Mat mat) 2781 { 2782 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 2783 PetscErrorCode ierr; 2784 2785 PetscFunctionBegin; 2786 ierr = MatRetrieveValues(aij->A);CHKERRQ(ierr); 2787 ierr = MatRetrieveValues(aij->B);CHKERRQ(ierr); 2788 PetscFunctionReturn(0); 2789 } 2790 2791 #undef __FUNCT__ 2792 #define __FUNCT__ "MatMPIAIJSetPreallocation_MPIAIJ" 2793 PetscErrorCode MatMPIAIJSetPreallocation_MPIAIJ(Mat B,PetscInt d_nz,const PetscInt d_nnz[],PetscInt o_nz,const PetscInt o_nnz[]) 2794 { 2795 Mat_MPIAIJ *b; 2796 PetscErrorCode ierr; 2797 2798 PetscFunctionBegin; 2799 ierr = PetscLayoutSetUp(B->rmap);CHKERRQ(ierr); 2800 ierr = PetscLayoutSetUp(B->cmap);CHKERRQ(ierr); 2801 b = (Mat_MPIAIJ*)B->data; 2802 2803 if (!B->preallocated) { 2804 /* Explicitly create 2 MATSEQAIJ matrices. */ 2805 ierr = MatCreate(PETSC_COMM_SELF,&b->A);CHKERRQ(ierr); 2806 ierr = MatSetSizes(b->A,B->rmap->n,B->cmap->n,B->rmap->n,B->cmap->n);CHKERRQ(ierr); 2807 ierr = MatSetBlockSizesFromMats(b->A,B,B);CHKERRQ(ierr); 2808 ierr = MatSetType(b->A,MATSEQAIJ);CHKERRQ(ierr); 2809 ierr = PetscLogObjectParent((PetscObject)B,(PetscObject)b->A);CHKERRQ(ierr); 2810 ierr = MatCreate(PETSC_COMM_SELF,&b->B);CHKERRQ(ierr); 2811 ierr = MatSetSizes(b->B,B->rmap->n,B->cmap->N,B->rmap->n,B->cmap->N);CHKERRQ(ierr); 2812 ierr = MatSetBlockSizesFromMats(b->B,B,B);CHKERRQ(ierr); 2813 ierr = MatSetType(b->B,MATSEQAIJ);CHKERRQ(ierr); 2814 ierr = PetscLogObjectParent((PetscObject)B,(PetscObject)b->B);CHKERRQ(ierr); 2815 } 2816 2817 ierr = MatSeqAIJSetPreallocation(b->A,d_nz,d_nnz);CHKERRQ(ierr); 2818 ierr = MatSeqAIJSetPreallocation(b->B,o_nz,o_nnz);CHKERRQ(ierr); 2819 B->preallocated = PETSC_TRUE; 2820 PetscFunctionReturn(0); 2821 } 2822 2823 #undef __FUNCT__ 2824 #define __FUNCT__ "MatDuplicate_MPIAIJ" 2825 PetscErrorCode MatDuplicate_MPIAIJ(Mat matin,MatDuplicateOption cpvalues,Mat *newmat) 2826 { 2827 Mat mat; 2828 Mat_MPIAIJ *a,*oldmat = (Mat_MPIAIJ*)matin->data; 2829 PetscErrorCode ierr; 2830 2831 PetscFunctionBegin; 2832 *newmat = 0; 2833 ierr = MatCreate(PetscObjectComm((PetscObject)matin),&mat);CHKERRQ(ierr); 2834 ierr = MatSetSizes(mat,matin->rmap->n,matin->cmap->n,matin->rmap->N,matin->cmap->N);CHKERRQ(ierr); 2835 ierr = MatSetBlockSizesFromMats(mat,matin,matin);CHKERRQ(ierr); 2836 ierr = MatSetType(mat,((PetscObject)matin)->type_name);CHKERRQ(ierr); 2837 ierr = PetscMemcpy(mat->ops,matin->ops,sizeof(struct _MatOps));CHKERRQ(ierr); 2838 a = (Mat_MPIAIJ*)mat->data; 2839 2840 mat->factortype = matin->factortype; 2841 mat->assembled = PETSC_TRUE; 2842 mat->insertmode = NOT_SET_VALUES; 2843 mat->preallocated = PETSC_TRUE; 2844 2845 a->size = oldmat->size; 2846 a->rank = oldmat->rank; 2847 a->donotstash = oldmat->donotstash; 2848 a->roworiented = oldmat->roworiented; 2849 a->rowindices = 0; 2850 a->rowvalues = 0; 2851 a->getrowactive = PETSC_FALSE; 2852 2853 ierr = PetscLayoutReference(matin->rmap,&mat->rmap);CHKERRQ(ierr); 2854 ierr = PetscLayoutReference(matin->cmap,&mat->cmap);CHKERRQ(ierr); 2855 2856 if (oldmat->colmap) { 2857 #if defined(PETSC_USE_CTABLE) 2858 ierr = PetscTableCreateCopy(oldmat->colmap,&a->colmap);CHKERRQ(ierr); 2859 #else 2860 ierr = PetscMalloc1(mat->cmap->N,&a->colmap);CHKERRQ(ierr); 2861 ierr = PetscLogObjectMemory((PetscObject)mat,(mat->cmap->N)*sizeof(PetscInt));CHKERRQ(ierr); 2862 ierr = PetscMemcpy(a->colmap,oldmat->colmap,(mat->cmap->N)*sizeof(PetscInt));CHKERRQ(ierr); 2863 #endif 2864 } else a->colmap = 0; 2865 if (oldmat->garray) { 2866 PetscInt len; 2867 len = oldmat->B->cmap->n; 2868 ierr = PetscMalloc1(len+1,&a->garray);CHKERRQ(ierr); 2869 ierr = PetscLogObjectMemory((PetscObject)mat,len*sizeof(PetscInt));CHKERRQ(ierr); 2870 if (len) { ierr = PetscMemcpy(a->garray,oldmat->garray,len*sizeof(PetscInt));CHKERRQ(ierr); } 2871 } else a->garray = 0; 2872 2873 ierr = VecDuplicate(oldmat->lvec,&a->lvec);CHKERRQ(ierr); 2874 ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)a->lvec);CHKERRQ(ierr); 2875 ierr = VecScatterCopy(oldmat->Mvctx,&a->Mvctx);CHKERRQ(ierr); 2876 ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)a->Mvctx);CHKERRQ(ierr); 2877 ierr = MatDuplicate(oldmat->A,cpvalues,&a->A);CHKERRQ(ierr); 2878 ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)a->A);CHKERRQ(ierr); 2879 ierr = MatDuplicate(oldmat->B,cpvalues,&a->B);CHKERRQ(ierr); 2880 ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)a->B);CHKERRQ(ierr); 2881 ierr = PetscFunctionListDuplicate(((PetscObject)matin)->qlist,&((PetscObject)mat)->qlist);CHKERRQ(ierr); 2882 *newmat = mat; 2883 PetscFunctionReturn(0); 2884 } 2885 2886 2887 2888 #undef __FUNCT__ 2889 #define __FUNCT__ "MatLoad_MPIAIJ" 2890 PetscErrorCode MatLoad_MPIAIJ(Mat newMat, PetscViewer viewer) 2891 { 2892 PetscScalar *vals,*svals; 2893 MPI_Comm comm; 2894 PetscErrorCode ierr; 2895 PetscMPIInt rank,size,tag = ((PetscObject)viewer)->tag; 2896 PetscInt i,nz,j,rstart,rend,mmax,maxnz = 0,grows,gcols; 2897 PetscInt header[4],*rowlengths = 0,M,N,m,*cols; 2898 PetscInt *ourlens = NULL,*procsnz = NULL,*offlens = NULL,jj,*mycols,*smycols; 2899 PetscInt cend,cstart,n,*rowners,sizesset=1; 2900 int fd; 2901 PetscInt bs = newMat->rmap->bs; 2902 2903 PetscFunctionBegin; 2904 /* force binary viewer to load .info file if it has not yet done so */ 2905 ierr = PetscViewerSetUp(viewer);CHKERRQ(ierr); 2906 ierr = PetscObjectGetComm((PetscObject)viewer,&comm);CHKERRQ(ierr); 2907 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 2908 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 2909 ierr = PetscViewerBinaryGetDescriptor(viewer,&fd);CHKERRQ(ierr); 2910 if (!rank) { 2911 ierr = PetscBinaryRead(fd,(char*)header,4,PETSC_INT);CHKERRQ(ierr); 2912 if (header[0] != MAT_FILE_CLASSID) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED,"not matrix object"); 2913 } 2914 2915 ierr = PetscOptionsBegin(comm,NULL,"Options for loading MPIAIJ matrix","Mat");CHKERRQ(ierr); 2916 ierr = PetscOptionsInt("-matload_block_size","Set the blocksize used to store the matrix","MatLoad",bs,&bs,NULL);CHKERRQ(ierr); 2917 ierr = PetscOptionsEnd();CHKERRQ(ierr); 2918 if (bs < 0) bs = 1; 2919 2920 if (newMat->rmap->n < 0 && newMat->rmap->N < 0 && newMat->cmap->n < 0 && newMat->cmap->N < 0) sizesset = 0; 2921 2922 ierr = MPI_Bcast(header+1,3,MPIU_INT,0,comm);CHKERRQ(ierr); 2923 M = header[1]; N = header[2]; 2924 /* If global rows/cols are set to PETSC_DECIDE, set it to the sizes given in the file */ 2925 if (sizesset && newMat->rmap->N < 0) newMat->rmap->N = M; 2926 if (sizesset && newMat->cmap->N < 0) newMat->cmap->N = N; 2927 2928 /* If global sizes are set, check if they are consistent with that given in the file */ 2929 if (sizesset) { 2930 ierr = MatGetSize(newMat,&grows,&gcols);CHKERRQ(ierr); 2931 } 2932 if (sizesset && newMat->rmap->N != grows) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED, "Inconsistent # of rows:Matrix in file has (%d) and input matrix has (%d)",M,grows); 2933 if (sizesset && newMat->cmap->N != gcols) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED, "Inconsistent # of cols:Matrix in file has (%d) and input matrix has (%d)",N,gcols); 2934 2935 /* determine ownership of all (block) rows */ 2936 if (M%bs) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED, "Inconsistent # of rows (%d) and block size (%d)",M,bs); 2937 if (newMat->rmap->n < 0) m = bs*((M/bs)/size + (((M/bs) % size) > rank)); /* PETSC_DECIDE */ 2938 else m = newMat->rmap->n; /* Set by user */ 2939 2940 ierr = PetscMalloc1(size+1,&rowners);CHKERRQ(ierr); 2941 ierr = MPI_Allgather(&m,1,MPIU_INT,rowners+1,1,MPIU_INT,comm);CHKERRQ(ierr); 2942 2943 /* First process needs enough room for process with most rows */ 2944 if (!rank) { 2945 mmax = rowners[1]; 2946 for (i=2; i<=size; i++) { 2947 mmax = PetscMax(mmax, rowners[i]); 2948 } 2949 } else mmax = -1; /* unused, but compilers complain */ 2950 2951 rowners[0] = 0; 2952 for (i=2; i<=size; i++) { 2953 rowners[i] += rowners[i-1]; 2954 } 2955 rstart = rowners[rank]; 2956 rend = rowners[rank+1]; 2957 2958 /* distribute row lengths to all processors */ 2959 ierr = PetscMalloc2(m,&ourlens,m,&offlens);CHKERRQ(ierr); 2960 if (!rank) { 2961 ierr = PetscBinaryRead(fd,ourlens,m,PETSC_INT);CHKERRQ(ierr); 2962 ierr = PetscMalloc1(mmax,&rowlengths);CHKERRQ(ierr); 2963 ierr = PetscCalloc1(size,&procsnz);CHKERRQ(ierr); 2964 for (j=0; j<m; j++) { 2965 procsnz[0] += ourlens[j]; 2966 } 2967 for (i=1; i<size; i++) { 2968 ierr = PetscBinaryRead(fd,rowlengths,rowners[i+1]-rowners[i],PETSC_INT);CHKERRQ(ierr); 2969 /* calculate the number of nonzeros on each processor */ 2970 for (j=0; j<rowners[i+1]-rowners[i]; j++) { 2971 procsnz[i] += rowlengths[j]; 2972 } 2973 ierr = MPIULong_Send(rowlengths,rowners[i+1]-rowners[i],MPIU_INT,i,tag,comm);CHKERRQ(ierr); 2974 } 2975 ierr = PetscFree(rowlengths);CHKERRQ(ierr); 2976 } else { 2977 ierr = MPIULong_Recv(ourlens,m,MPIU_INT,0,tag,comm);CHKERRQ(ierr); 2978 } 2979 2980 if (!rank) { 2981 /* determine max buffer needed and allocate it */ 2982 maxnz = 0; 2983 for (i=0; i<size; i++) { 2984 maxnz = PetscMax(maxnz,procsnz[i]); 2985 } 2986 ierr = PetscMalloc1(maxnz,&cols);CHKERRQ(ierr); 2987 2988 /* read in my part of the matrix column indices */ 2989 nz = procsnz[0]; 2990 ierr = PetscMalloc1(nz,&mycols);CHKERRQ(ierr); 2991 ierr = PetscBinaryRead(fd,mycols,nz,PETSC_INT);CHKERRQ(ierr); 2992 2993 /* read in every one elses and ship off */ 2994 for (i=1; i<size; i++) { 2995 nz = procsnz[i]; 2996 ierr = PetscBinaryRead(fd,cols,nz,PETSC_INT);CHKERRQ(ierr); 2997 ierr = MPIULong_Send(cols,nz,MPIU_INT,i,tag,comm);CHKERRQ(ierr); 2998 } 2999 ierr = PetscFree(cols);CHKERRQ(ierr); 3000 } else { 3001 /* determine buffer space needed for message */ 3002 nz = 0; 3003 for (i=0; i<m; i++) { 3004 nz += ourlens[i]; 3005 } 3006 ierr = PetscMalloc1(nz,&mycols);CHKERRQ(ierr); 3007 3008 /* receive message of column indices*/ 3009 ierr = MPIULong_Recv(mycols,nz,MPIU_INT,0,tag,comm);CHKERRQ(ierr); 3010 } 3011 3012 /* determine column ownership if matrix is not square */ 3013 if (N != M) { 3014 if (newMat->cmap->n < 0) n = N/size + ((N % size) > rank); 3015 else n = newMat->cmap->n; 3016 ierr = MPI_Scan(&n,&cend,1,MPIU_INT,MPI_SUM,comm);CHKERRQ(ierr); 3017 cstart = cend - n; 3018 } else { 3019 cstart = rstart; 3020 cend = rend; 3021 n = cend - cstart; 3022 } 3023 3024 /* loop over local rows, determining number of off diagonal entries */ 3025 ierr = PetscMemzero(offlens,m*sizeof(PetscInt));CHKERRQ(ierr); 3026 jj = 0; 3027 for (i=0; i<m; i++) { 3028 for (j=0; j<ourlens[i]; j++) { 3029 if (mycols[jj] < cstart || mycols[jj] >= cend) offlens[i]++; 3030 jj++; 3031 } 3032 } 3033 3034 for (i=0; i<m; i++) { 3035 ourlens[i] -= offlens[i]; 3036 } 3037 if (!sizesset) { 3038 ierr = MatSetSizes(newMat,m,n,M,N);CHKERRQ(ierr); 3039 } 3040 3041 if (bs > 1) {ierr = MatSetBlockSize(newMat,bs);CHKERRQ(ierr);} 3042 3043 ierr = MatMPIAIJSetPreallocation(newMat,0,ourlens,0,offlens);CHKERRQ(ierr); 3044 3045 for (i=0; i<m; i++) { 3046 ourlens[i] += offlens[i]; 3047 } 3048 3049 if (!rank) { 3050 ierr = PetscMalloc1(maxnz+1,&vals);CHKERRQ(ierr); 3051 3052 /* read in my part of the matrix numerical values */ 3053 nz = procsnz[0]; 3054 ierr = PetscBinaryRead(fd,vals,nz,PETSC_SCALAR);CHKERRQ(ierr); 3055 3056 /* insert into matrix */ 3057 jj = rstart; 3058 smycols = mycols; 3059 svals = vals; 3060 for (i=0; i<m; i++) { 3061 ierr = MatSetValues_MPIAIJ(newMat,1,&jj,ourlens[i],smycols,svals,INSERT_VALUES);CHKERRQ(ierr); 3062 smycols += ourlens[i]; 3063 svals += ourlens[i]; 3064 jj++; 3065 } 3066 3067 /* read in other processors and ship out */ 3068 for (i=1; i<size; i++) { 3069 nz = procsnz[i]; 3070 ierr = PetscBinaryRead(fd,vals,nz,PETSC_SCALAR);CHKERRQ(ierr); 3071 ierr = MPIULong_Send(vals,nz,MPIU_SCALAR,i,((PetscObject)newMat)->tag,comm);CHKERRQ(ierr); 3072 } 3073 ierr = PetscFree(procsnz);CHKERRQ(ierr); 3074 } else { 3075 /* receive numeric values */ 3076 ierr = PetscMalloc1(nz+1,&vals);CHKERRQ(ierr); 3077 3078 /* receive message of values*/ 3079 ierr = MPIULong_Recv(vals,nz,MPIU_SCALAR,0,((PetscObject)newMat)->tag,comm);CHKERRQ(ierr); 3080 3081 /* insert into matrix */ 3082 jj = rstart; 3083 smycols = mycols; 3084 svals = vals; 3085 for (i=0; i<m; i++) { 3086 ierr = MatSetValues_MPIAIJ(newMat,1,&jj,ourlens[i],smycols,svals,INSERT_VALUES);CHKERRQ(ierr); 3087 smycols += ourlens[i]; 3088 svals += ourlens[i]; 3089 jj++; 3090 } 3091 } 3092 ierr = PetscFree2(ourlens,offlens);CHKERRQ(ierr); 3093 ierr = PetscFree(vals);CHKERRQ(ierr); 3094 ierr = PetscFree(mycols);CHKERRQ(ierr); 3095 ierr = PetscFree(rowners);CHKERRQ(ierr); 3096 ierr = MatAssemblyBegin(newMat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3097 ierr = MatAssemblyEnd(newMat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3098 PetscFunctionReturn(0); 3099 } 3100 3101 #undef __FUNCT__ 3102 #define __FUNCT__ "MatGetSubMatrix_MPIAIJ" 3103 PetscErrorCode MatGetSubMatrix_MPIAIJ(Mat mat,IS isrow,IS iscol,MatReuse call,Mat *newmat) 3104 { 3105 PetscErrorCode ierr; 3106 IS iscol_local; 3107 PetscInt csize; 3108 3109 PetscFunctionBegin; 3110 ierr = ISGetLocalSize(iscol,&csize);CHKERRQ(ierr); 3111 if (call == MAT_REUSE_MATRIX) { 3112 ierr = PetscObjectQuery((PetscObject)*newmat,"ISAllGather",(PetscObject*)&iscol_local);CHKERRQ(ierr); 3113 if (!iscol_local) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Submatrix passed in was not used before, cannot reuse"); 3114 } else { 3115 PetscInt cbs; 3116 ierr = ISGetBlockSize(iscol,&cbs);CHKERRQ(ierr); 3117 ierr = ISAllGather(iscol,&iscol_local);CHKERRQ(ierr); 3118 ierr = ISSetBlockSize(iscol_local,cbs);CHKERRQ(ierr); 3119 } 3120 ierr = MatGetSubMatrix_MPIAIJ_Private(mat,isrow,iscol_local,csize,call,newmat);CHKERRQ(ierr); 3121 if (call == MAT_INITIAL_MATRIX) { 3122 ierr = PetscObjectCompose((PetscObject)*newmat,"ISAllGather",(PetscObject)iscol_local);CHKERRQ(ierr); 3123 ierr = ISDestroy(&iscol_local);CHKERRQ(ierr); 3124 } 3125 PetscFunctionReturn(0); 3126 } 3127 3128 extern PetscErrorCode MatGetSubMatrices_MPIAIJ_Local(Mat,PetscInt,const IS[],const IS[],MatReuse,PetscBool*,Mat*); 3129 #undef __FUNCT__ 3130 #define __FUNCT__ "MatGetSubMatrix_MPIAIJ_Private" 3131 /* 3132 Not great since it makes two copies of the submatrix, first an SeqAIJ 3133 in local and then by concatenating the local matrices the end result. 3134 Writing it directly would be much like MatGetSubMatrices_MPIAIJ() 3135 3136 Note: This requires a sequential iscol with all indices. 3137 */ 3138 PetscErrorCode MatGetSubMatrix_MPIAIJ_Private(Mat mat,IS isrow,IS iscol,PetscInt csize,MatReuse call,Mat *newmat) 3139 { 3140 PetscErrorCode ierr; 3141 PetscMPIInt rank,size; 3142 PetscInt i,m,n,rstart,row,rend,nz,*cwork,j,bs,cbs; 3143 PetscInt *ii,*jj,nlocal,*dlens,*olens,dlen,olen,jend,mglobal,ncol; 3144 PetscBool allcolumns, colflag; 3145 Mat M,Mreuse; 3146 MatScalar *vwork,*aa; 3147 MPI_Comm comm; 3148 Mat_SeqAIJ *aij; 3149 3150 PetscFunctionBegin; 3151 ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 3152 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 3153 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 3154 3155 ierr = ISIdentity(iscol,&colflag);CHKERRQ(ierr); 3156 ierr = ISGetLocalSize(iscol,&ncol);CHKERRQ(ierr); 3157 if (colflag && ncol == mat->cmap->N) { 3158 allcolumns = PETSC_TRUE; 3159 } else { 3160 allcolumns = PETSC_FALSE; 3161 } 3162 if (call == MAT_REUSE_MATRIX) { 3163 ierr = PetscObjectQuery((PetscObject)*newmat,"SubMatrix",(PetscObject*)&Mreuse);CHKERRQ(ierr); 3164 if (!Mreuse) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Submatrix passed in was not used before, cannot reuse"); 3165 ierr = MatGetSubMatrices_MPIAIJ_Local(mat,1,&isrow,&iscol,MAT_REUSE_MATRIX,&allcolumns,&Mreuse);CHKERRQ(ierr); 3166 } else { 3167 ierr = MatGetSubMatrices_MPIAIJ_Local(mat,1,&isrow,&iscol,MAT_INITIAL_MATRIX,&allcolumns,&Mreuse);CHKERRQ(ierr); 3168 } 3169 3170 /* 3171 m - number of local rows 3172 n - number of columns (same on all processors) 3173 rstart - first row in new global matrix generated 3174 */ 3175 ierr = MatGetSize(Mreuse,&m,&n);CHKERRQ(ierr); 3176 ierr = MatGetBlockSizes(Mreuse,&bs,&cbs);CHKERRQ(ierr); 3177 if (call == MAT_INITIAL_MATRIX) { 3178 aij = (Mat_SeqAIJ*)(Mreuse)->data; 3179 ii = aij->i; 3180 jj = aij->j; 3181 3182 /* 3183 Determine the number of non-zeros in the diagonal and off-diagonal 3184 portions of the matrix in order to do correct preallocation 3185 */ 3186 3187 /* first get start and end of "diagonal" columns */ 3188 if (csize == PETSC_DECIDE) { 3189 ierr = ISGetSize(isrow,&mglobal);CHKERRQ(ierr); 3190 if (mglobal == n) { /* square matrix */ 3191 nlocal = m; 3192 } else { 3193 nlocal = n/size + ((n % size) > rank); 3194 } 3195 } else { 3196 nlocal = csize; 3197 } 3198 ierr = MPI_Scan(&nlocal,&rend,1,MPIU_INT,MPI_SUM,comm);CHKERRQ(ierr); 3199 rstart = rend - nlocal; 3200 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); 3201 3202 /* next, compute all the lengths */ 3203 ierr = PetscMalloc1(2*m+1,&dlens);CHKERRQ(ierr); 3204 olens = dlens + m; 3205 for (i=0; i<m; i++) { 3206 jend = ii[i+1] - ii[i]; 3207 olen = 0; 3208 dlen = 0; 3209 for (j=0; j<jend; j++) { 3210 if (*jj < rstart || *jj >= rend) olen++; 3211 else dlen++; 3212 jj++; 3213 } 3214 olens[i] = olen; 3215 dlens[i] = dlen; 3216 } 3217 ierr = MatCreate(comm,&M);CHKERRQ(ierr); 3218 ierr = MatSetSizes(M,m,nlocal,PETSC_DECIDE,n);CHKERRQ(ierr); 3219 ierr = MatSetBlockSizes(M,bs,cbs);CHKERRQ(ierr); 3220 ierr = MatSetType(M,((PetscObject)mat)->type_name);CHKERRQ(ierr); 3221 ierr = MatMPIAIJSetPreallocation(M,0,dlens,0,olens);CHKERRQ(ierr); 3222 ierr = PetscFree(dlens);CHKERRQ(ierr); 3223 } else { 3224 PetscInt ml,nl; 3225 3226 M = *newmat; 3227 ierr = MatGetLocalSize(M,&ml,&nl);CHKERRQ(ierr); 3228 if (ml != m) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Previous matrix must be same size/layout as request"); 3229 ierr = MatZeroEntries(M);CHKERRQ(ierr); 3230 /* 3231 The next two lines are needed so we may call MatSetValues_MPIAIJ() below directly, 3232 rather than the slower MatSetValues(). 3233 */ 3234 M->was_assembled = PETSC_TRUE; 3235 M->assembled = PETSC_FALSE; 3236 } 3237 ierr = MatGetOwnershipRange(M,&rstart,&rend);CHKERRQ(ierr); 3238 aij = (Mat_SeqAIJ*)(Mreuse)->data; 3239 ii = aij->i; 3240 jj = aij->j; 3241 aa = aij->a; 3242 for (i=0; i<m; i++) { 3243 row = rstart + i; 3244 nz = ii[i+1] - ii[i]; 3245 cwork = jj; jj += nz; 3246 vwork = aa; aa += nz; 3247 ierr = MatSetValues_MPIAIJ(M,1,&row,nz,cwork,vwork,INSERT_VALUES);CHKERRQ(ierr); 3248 } 3249 3250 ierr = MatAssemblyBegin(M,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3251 ierr = MatAssemblyEnd(M,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3252 *newmat = M; 3253 3254 /* save submatrix used in processor for next request */ 3255 if (call == MAT_INITIAL_MATRIX) { 3256 ierr = PetscObjectCompose((PetscObject)M,"SubMatrix",(PetscObject)Mreuse);CHKERRQ(ierr); 3257 ierr = MatDestroy(&Mreuse);CHKERRQ(ierr); 3258 } 3259 PetscFunctionReturn(0); 3260 } 3261 3262 #undef __FUNCT__ 3263 #define __FUNCT__ "MatMPIAIJSetPreallocationCSR_MPIAIJ" 3264 PetscErrorCode MatMPIAIJSetPreallocationCSR_MPIAIJ(Mat B,const PetscInt Ii[],const PetscInt J[],const PetscScalar v[]) 3265 { 3266 PetscInt m,cstart, cend,j,nnz,i,d; 3267 PetscInt *d_nnz,*o_nnz,nnz_max = 0,rstart,ii; 3268 const PetscInt *JJ; 3269 PetscScalar *values; 3270 PetscErrorCode ierr; 3271 3272 PetscFunctionBegin; 3273 if (Ii[0]) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Ii[0] must be 0 it is %D",Ii[0]); 3274 3275 ierr = PetscLayoutSetUp(B->rmap);CHKERRQ(ierr); 3276 ierr = PetscLayoutSetUp(B->cmap);CHKERRQ(ierr); 3277 m = B->rmap->n; 3278 cstart = B->cmap->rstart; 3279 cend = B->cmap->rend; 3280 rstart = B->rmap->rstart; 3281 3282 ierr = PetscMalloc2(m,&d_nnz,m,&o_nnz);CHKERRQ(ierr); 3283 3284 #if defined(PETSC_USE_DEBUGGING) 3285 for (i=0; i<m; i++) { 3286 nnz = Ii[i+1]- Ii[i]; 3287 JJ = J + Ii[i]; 3288 if (nnz < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Local row %D has a negative %D number of columns",i,nnz); 3289 if (nnz && (JJ[0] < 0)) SETERRRQ1(PETSC_ERR_ARG_WRONGSTATE,"Row %D starts with negative column index",i,j); 3290 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); 3291 } 3292 #endif 3293 3294 for (i=0; i<m; i++) { 3295 nnz = Ii[i+1]- Ii[i]; 3296 JJ = J + Ii[i]; 3297 nnz_max = PetscMax(nnz_max,nnz); 3298 d = 0; 3299 for (j=0; j<nnz; j++) { 3300 if (cstart <= JJ[j] && JJ[j] < cend) d++; 3301 } 3302 d_nnz[i] = d; 3303 o_nnz[i] = nnz - d; 3304 } 3305 ierr = MatMPIAIJSetPreallocation(B,0,d_nnz,0,o_nnz);CHKERRQ(ierr); 3306 ierr = PetscFree2(d_nnz,o_nnz);CHKERRQ(ierr); 3307 3308 if (v) values = (PetscScalar*)v; 3309 else { 3310 ierr = PetscCalloc1(nnz_max+1,&values);CHKERRQ(ierr); 3311 } 3312 3313 for (i=0; i<m; i++) { 3314 ii = i + rstart; 3315 nnz = Ii[i+1]- Ii[i]; 3316 ierr = MatSetValues_MPIAIJ(B,1,&ii,nnz,J+Ii[i],values+(v ? Ii[i] : 0),INSERT_VALUES);CHKERRQ(ierr); 3317 } 3318 ierr = MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3319 ierr = MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3320 3321 if (!v) { 3322 ierr = PetscFree(values);CHKERRQ(ierr); 3323 } 3324 ierr = MatSetOption(B,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE);CHKERRQ(ierr); 3325 PetscFunctionReturn(0); 3326 } 3327 3328 #undef __FUNCT__ 3329 #define __FUNCT__ "MatMPIAIJSetPreallocationCSR" 3330 /*@ 3331 MatMPIAIJSetPreallocationCSR - Allocates memory for a sparse parallel matrix in AIJ format 3332 (the default parallel PETSc format). 3333 3334 Collective on MPI_Comm 3335 3336 Input Parameters: 3337 + B - the matrix 3338 . i - the indices into j for the start of each local row (starts with zero) 3339 . j - the column indices for each local row (starts with zero) 3340 - v - optional values in the matrix 3341 3342 Level: developer 3343 3344 Notes: 3345 The i, j, and a arrays ARE copied by this routine into the internal format used by PETSc; 3346 thus you CANNOT change the matrix entries by changing the values of a[] after you have 3347 called this routine. Use MatCreateMPIAIJWithSplitArrays() to avoid needing to copy the arrays. 3348 3349 The i and j indices are 0 based, and i indices are indices corresponding to the local j array. 3350 3351 The format which is used for the sparse matrix input, is equivalent to a 3352 row-major ordering.. i.e for the following matrix, the input data expected is 3353 as shown: 3354 3355 1 0 0 3356 2 0 3 P0 3357 ------- 3358 4 5 6 P1 3359 3360 Process0 [P0]: rows_owned=[0,1] 3361 i = {0,1,3} [size = nrow+1 = 2+1] 3362 j = {0,0,2} [size = nz = 6] 3363 v = {1,2,3} [size = nz = 6] 3364 3365 Process1 [P1]: rows_owned=[2] 3366 i = {0,3} [size = nrow+1 = 1+1] 3367 j = {0,1,2} [size = nz = 6] 3368 v = {4,5,6} [size = nz = 6] 3369 3370 .keywords: matrix, aij, compressed row, sparse, parallel 3371 3372 .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatCreateAIJ(), MPIAIJ, 3373 MatCreateSeqAIJWithArrays(), MatCreateMPIAIJWithSplitArrays() 3374 @*/ 3375 PetscErrorCode MatMPIAIJSetPreallocationCSR(Mat B,const PetscInt i[],const PetscInt j[], const PetscScalar v[]) 3376 { 3377 PetscErrorCode ierr; 3378 3379 PetscFunctionBegin; 3380 ierr = PetscTryMethod(B,"MatMPIAIJSetPreallocationCSR_C",(Mat,const PetscInt[],const PetscInt[],const PetscScalar[]),(B,i,j,v));CHKERRQ(ierr); 3381 PetscFunctionReturn(0); 3382 } 3383 3384 #undef __FUNCT__ 3385 #define __FUNCT__ "MatMPIAIJSetPreallocation" 3386 /*@C 3387 MatMPIAIJSetPreallocation - Preallocates memory for a sparse parallel matrix in AIJ format 3388 (the default parallel PETSc format). For good matrix assembly performance 3389 the user should preallocate the matrix storage by setting the parameters 3390 d_nz (or d_nnz) and o_nz (or o_nnz). By setting these parameters accurately, 3391 performance can be increased by more than a factor of 50. 3392 3393 Collective on MPI_Comm 3394 3395 Input Parameters: 3396 + B - the matrix 3397 . d_nz - number of nonzeros per row in DIAGONAL portion of local submatrix 3398 (same value is used for all local rows) 3399 . d_nnz - array containing the number of nonzeros in the various rows of the 3400 DIAGONAL portion of the local submatrix (possibly different for each row) 3401 or NULL, if d_nz is used to specify the nonzero structure. 3402 The size of this array is equal to the number of local rows, i.e 'm'. 3403 For matrices that will be factored, you must leave room for (and set) 3404 the diagonal entry even if it is zero. 3405 . o_nz - number of nonzeros per row in the OFF-DIAGONAL portion of local 3406 submatrix (same value is used for all local rows). 3407 - o_nnz - array containing the number of nonzeros in the various rows of the 3408 OFF-DIAGONAL portion of the local submatrix (possibly different for 3409 each row) or NULL, if o_nz is used to specify the nonzero 3410 structure. The size of this array is equal to the number 3411 of local rows, i.e 'm'. 3412 3413 If the *_nnz parameter is given then the *_nz parameter is ignored 3414 3415 The AIJ format (also called the Yale sparse matrix format or 3416 compressed row storage (CSR)), is fully compatible with standard Fortran 77 3417 storage. The stored row and column indices begin with zero. 3418 See Users-Manual: ch_mat for details. 3419 3420 The parallel matrix is partitioned such that the first m0 rows belong to 3421 process 0, the next m1 rows belong to process 1, the next m2 rows belong 3422 to process 2 etc.. where m0,m1,m2... are the input parameter 'm'. 3423 3424 The DIAGONAL portion of the local submatrix of a processor can be defined 3425 as the submatrix which is obtained by extraction the part corresponding to 3426 the rows r1-r2 and columns c1-c2 of the global matrix, where r1 is the 3427 first row that belongs to the processor, r2 is the last row belonging to 3428 the this processor, and c1-c2 is range of indices of the local part of a 3429 vector suitable for applying the matrix to. This is an mxn matrix. In the 3430 common case of a square matrix, the row and column ranges are the same and 3431 the DIAGONAL part is also square. The remaining portion of the local 3432 submatrix (mxN) constitute the OFF-DIAGONAL portion. 3433 3434 If o_nnz, d_nnz are specified, then o_nz, and d_nz are ignored. 3435 3436 You can call MatGetInfo() to get information on how effective the preallocation was; 3437 for example the fields mallocs,nz_allocated,nz_used,nz_unneeded; 3438 You can also run with the option -info and look for messages with the string 3439 malloc in them to see if additional memory allocation was needed. 3440 3441 Example usage: 3442 3443 Consider the following 8x8 matrix with 34 non-zero values, that is 3444 assembled across 3 processors. Lets assume that proc0 owns 3 rows, 3445 proc1 owns 3 rows, proc2 owns 2 rows. This division can be shown 3446 as follows: 3447 3448 .vb 3449 1 2 0 | 0 3 0 | 0 4 3450 Proc0 0 5 6 | 7 0 0 | 8 0 3451 9 0 10 | 11 0 0 | 12 0 3452 ------------------------------------- 3453 13 0 14 | 15 16 17 | 0 0 3454 Proc1 0 18 0 | 19 20 21 | 0 0 3455 0 0 0 | 22 23 0 | 24 0 3456 ------------------------------------- 3457 Proc2 25 26 27 | 0 0 28 | 29 0 3458 30 0 0 | 31 32 33 | 0 34 3459 .ve 3460 3461 This can be represented as a collection of submatrices as: 3462 3463 .vb 3464 A B C 3465 D E F 3466 G H I 3467 .ve 3468 3469 Where the submatrices A,B,C are owned by proc0, D,E,F are 3470 owned by proc1, G,H,I are owned by proc2. 3471 3472 The 'm' parameters for proc0,proc1,proc2 are 3,3,2 respectively. 3473 The 'n' parameters for proc0,proc1,proc2 are 3,3,2 respectively. 3474 The 'M','N' parameters are 8,8, and have the same values on all procs. 3475 3476 The DIAGONAL submatrices corresponding to proc0,proc1,proc2 are 3477 submatrices [A], [E], [I] respectively. The OFF-DIAGONAL submatrices 3478 corresponding to proc0,proc1,proc2 are [BC], [DF], [GH] respectively. 3479 Internally, each processor stores the DIAGONAL part, and the OFF-DIAGONAL 3480 part as SeqAIJ matrices. for eg: proc1 will store [E] as a SeqAIJ 3481 matrix, ans [DF] as another SeqAIJ matrix. 3482 3483 When d_nz, o_nz parameters are specified, d_nz storage elements are 3484 allocated for every row of the local diagonal submatrix, and o_nz 3485 storage locations are allocated for every row of the OFF-DIAGONAL submat. 3486 One way to choose d_nz and o_nz is to use the max nonzerors per local 3487 rows for each of the local DIAGONAL, and the OFF-DIAGONAL submatrices. 3488 In this case, the values of d_nz,o_nz are: 3489 .vb 3490 proc0 : dnz = 2, o_nz = 2 3491 proc1 : dnz = 3, o_nz = 2 3492 proc2 : dnz = 1, o_nz = 4 3493 .ve 3494 We are allocating m*(d_nz+o_nz) storage locations for every proc. This 3495 translates to 3*(2+2)=12 for proc0, 3*(3+2)=15 for proc1, 2*(1+4)=10 3496 for proc3. i.e we are using 12+15+10=37 storage locations to store 3497 34 values. 3498 3499 When d_nnz, o_nnz parameters are specified, the storage is specified 3500 for every row, coresponding to both DIAGONAL and OFF-DIAGONAL submatrices. 3501 In the above case the values for d_nnz,o_nnz are: 3502 .vb 3503 proc0: d_nnz = [2,2,2] and o_nnz = [2,2,2] 3504 proc1: d_nnz = [3,3,2] and o_nnz = [2,1,1] 3505 proc2: d_nnz = [1,1] and o_nnz = [4,4] 3506 .ve 3507 Here the space allocated is sum of all the above values i.e 34, and 3508 hence pre-allocation is perfect. 3509 3510 Level: intermediate 3511 3512 .keywords: matrix, aij, compressed row, sparse, parallel 3513 3514 .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatCreateAIJ(), MatMPIAIJSetPreallocationCSR(), 3515 MPIAIJ, MatGetInfo(), PetscSplitOwnership() 3516 @*/ 3517 PetscErrorCode MatMPIAIJSetPreallocation(Mat B,PetscInt d_nz,const PetscInt d_nnz[],PetscInt o_nz,const PetscInt o_nnz[]) 3518 { 3519 PetscErrorCode ierr; 3520 3521 PetscFunctionBegin; 3522 PetscValidHeaderSpecific(B,MAT_CLASSID,1); 3523 PetscValidType(B,1); 3524 ierr = PetscTryMethod(B,"MatMPIAIJSetPreallocation_C",(Mat,PetscInt,const PetscInt[],PetscInt,const PetscInt[]),(B,d_nz,d_nnz,o_nz,o_nnz));CHKERRQ(ierr); 3525 PetscFunctionReturn(0); 3526 } 3527 3528 #undef __FUNCT__ 3529 #define __FUNCT__ "MatCreateMPIAIJWithArrays" 3530 /*@ 3531 MatCreateMPIAIJWithArrays - creates a MPI AIJ matrix using arrays that contain in standard 3532 CSR format the local rows. 3533 3534 Collective on MPI_Comm 3535 3536 Input Parameters: 3537 + comm - MPI communicator 3538 . m - number of local rows (Cannot be PETSC_DECIDE) 3539 . n - This value should be the same as the local size used in creating the 3540 x vector for the matrix-vector product y = Ax. (or PETSC_DECIDE to have 3541 calculated if N is given) For square matrices n is almost always m. 3542 . M - number of global rows (or PETSC_DETERMINE to have calculated if m is given) 3543 . N - number of global columns (or PETSC_DETERMINE to have calculated if n is given) 3544 . i - row indices 3545 . j - column indices 3546 - a - matrix values 3547 3548 Output Parameter: 3549 . mat - the matrix 3550 3551 Level: intermediate 3552 3553 Notes: 3554 The i, j, and a arrays ARE copied by this routine into the internal format used by PETSc; 3555 thus you CANNOT change the matrix entries by changing the values of a[] after you have 3556 called this routine. Use MatCreateMPIAIJWithSplitArrays() to avoid needing to copy the arrays. 3557 3558 The i and j indices are 0 based, and i indices are indices corresponding to the local j array. 3559 3560 The format which is used for the sparse matrix input, is equivalent to a 3561 row-major ordering.. i.e for the following matrix, the input data expected is 3562 as shown: 3563 3564 1 0 0 3565 2 0 3 P0 3566 ------- 3567 4 5 6 P1 3568 3569 Process0 [P0]: rows_owned=[0,1] 3570 i = {0,1,3} [size = nrow+1 = 2+1] 3571 j = {0,0,2} [size = nz = 6] 3572 v = {1,2,3} [size = nz = 6] 3573 3574 Process1 [P1]: rows_owned=[2] 3575 i = {0,3} [size = nrow+1 = 1+1] 3576 j = {0,1,2} [size = nz = 6] 3577 v = {4,5,6} [size = nz = 6] 3578 3579 .keywords: matrix, aij, compressed row, sparse, parallel 3580 3581 .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatMPIAIJSetPreallocationCSR(), 3582 MPIAIJ, MatCreateAIJ(), MatCreateMPIAIJWithSplitArrays() 3583 @*/ 3584 PetscErrorCode MatCreateMPIAIJWithArrays(MPI_Comm comm,PetscInt m,PetscInt n,PetscInt M,PetscInt N,const PetscInt i[],const PetscInt j[],const PetscScalar a[],Mat *mat) 3585 { 3586 PetscErrorCode ierr; 3587 3588 PetscFunctionBegin; 3589 if (i[0]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"i (row indices) must start with 0"); 3590 if (m < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"local number of rows (m) cannot be PETSC_DECIDE, or negative"); 3591 ierr = MatCreate(comm,mat);CHKERRQ(ierr); 3592 ierr = MatSetSizes(*mat,m,n,M,N);CHKERRQ(ierr); 3593 /* ierr = MatSetBlockSizes(M,bs,cbs);CHKERRQ(ierr); */ 3594 ierr = MatSetType(*mat,MATMPIAIJ);CHKERRQ(ierr); 3595 ierr = MatMPIAIJSetPreallocationCSR(*mat,i,j,a);CHKERRQ(ierr); 3596 PetscFunctionReturn(0); 3597 } 3598 3599 #undef __FUNCT__ 3600 #define __FUNCT__ "MatCreateAIJ" 3601 /*@C 3602 MatCreateAIJ - Creates a sparse parallel matrix in AIJ format 3603 (the default parallel PETSc format). For good matrix assembly performance 3604 the user should preallocate the matrix storage by setting the parameters 3605 d_nz (or d_nnz) and o_nz (or o_nnz). By setting these parameters accurately, 3606 performance can be increased by more than a factor of 50. 3607 3608 Collective on MPI_Comm 3609 3610 Input Parameters: 3611 + comm - MPI communicator 3612 . m - number of local rows (or PETSC_DECIDE to have calculated if M is given) 3613 This value should be the same as the local size used in creating the 3614 y vector for the matrix-vector product y = Ax. 3615 . n - This value should be the same as the local size used in creating the 3616 x vector for the matrix-vector product y = Ax. (or PETSC_DECIDE to have 3617 calculated if N is given) For square matrices n is almost always m. 3618 . M - number of global rows (or PETSC_DETERMINE to have calculated if m is given) 3619 . N - number of global columns (or PETSC_DETERMINE to have calculated if n is given) 3620 . d_nz - number of nonzeros per row in DIAGONAL portion of local submatrix 3621 (same value is used for all local rows) 3622 . d_nnz - array containing the number of nonzeros in the various rows of the 3623 DIAGONAL portion of the local submatrix (possibly different for each row) 3624 or NULL, if d_nz is used to specify the nonzero structure. 3625 The size of this array is equal to the number of local rows, i.e 'm'. 3626 . o_nz - number of nonzeros per row in the OFF-DIAGONAL portion of local 3627 submatrix (same value is used for all local rows). 3628 - o_nnz - array containing the number of nonzeros in the various rows of the 3629 OFF-DIAGONAL portion of the local submatrix (possibly different for 3630 each row) or NULL, if o_nz is used to specify the nonzero 3631 structure. The size of this array is equal to the number 3632 of local rows, i.e 'm'. 3633 3634 Output Parameter: 3635 . A - the matrix 3636 3637 It is recommended that one use the MatCreate(), MatSetType() and/or MatSetFromOptions(), 3638 MatXXXXSetPreallocation() paradgm instead of this routine directly. 3639 [MatXXXXSetPreallocation() is, for example, MatSeqAIJSetPreallocation] 3640 3641 Notes: 3642 If the *_nnz parameter is given then the *_nz parameter is ignored 3643 3644 m,n,M,N parameters specify the size of the matrix, and its partitioning across 3645 processors, while d_nz,d_nnz,o_nz,o_nnz parameters specify the approximate 3646 storage requirements for this matrix. 3647 3648 If PETSC_DECIDE or PETSC_DETERMINE is used for a particular argument on one 3649 processor than it must be used on all processors that share the object for 3650 that argument. 3651 3652 The user MUST specify either the local or global matrix dimensions 3653 (possibly both). 3654 3655 The parallel matrix is partitioned across processors such that the 3656 first m0 rows belong to process 0, the next m1 rows belong to 3657 process 1, the next m2 rows belong to process 2 etc.. where 3658 m0,m1,m2,.. are the input parameter 'm'. i.e each processor stores 3659 values corresponding to [m x N] submatrix. 3660 3661 The columns are logically partitioned with the n0 columns belonging 3662 to 0th partition, the next n1 columns belonging to the next 3663 partition etc.. where n0,n1,n2... are the input parameter 'n'. 3664 3665 The DIAGONAL portion of the local submatrix on any given processor 3666 is the submatrix corresponding to the rows and columns m,n 3667 corresponding to the given processor. i.e diagonal matrix on 3668 process 0 is [m0 x n0], diagonal matrix on process 1 is [m1 x n1] 3669 etc. The remaining portion of the local submatrix [m x (N-n)] 3670 constitute the OFF-DIAGONAL portion. The example below better 3671 illustrates this concept. 3672 3673 For a square global matrix we define each processor's diagonal portion 3674 to be its local rows and the corresponding columns (a square submatrix); 3675 each processor's off-diagonal portion encompasses the remainder of the 3676 local matrix (a rectangular submatrix). 3677 3678 If o_nnz, d_nnz are specified, then o_nz, and d_nz are ignored. 3679 3680 When calling this routine with a single process communicator, a matrix of 3681 type SEQAIJ is returned. If a matrix of type MPIAIJ is desired for this 3682 type of communicator, use the construction mechanism: 3683 MatCreate(...,&A); MatSetType(A,MATMPIAIJ); MatSetSizes(A, m,n,M,N); MatMPIAIJSetPreallocation(A,...); 3684 3685 By default, this format uses inodes (identical nodes) when possible. 3686 We search for consecutive rows with the same nonzero structure, thereby 3687 reusing matrix information to achieve increased efficiency. 3688 3689 Options Database Keys: 3690 + -mat_no_inode - Do not use inodes 3691 . -mat_inode_limit <limit> - Sets inode limit (max limit=5) 3692 - -mat_aij_oneindex - Internally use indexing starting at 1 3693 rather than 0. Note that when calling MatSetValues(), 3694 the user still MUST index entries starting at 0! 3695 3696 3697 Example usage: 3698 3699 Consider the following 8x8 matrix with 34 non-zero values, that is 3700 assembled across 3 processors. Lets assume that proc0 owns 3 rows, 3701 proc1 owns 3 rows, proc2 owns 2 rows. This division can be shown 3702 as follows: 3703 3704 .vb 3705 1 2 0 | 0 3 0 | 0 4 3706 Proc0 0 5 6 | 7 0 0 | 8 0 3707 9 0 10 | 11 0 0 | 12 0 3708 ------------------------------------- 3709 13 0 14 | 15 16 17 | 0 0 3710 Proc1 0 18 0 | 19 20 21 | 0 0 3711 0 0 0 | 22 23 0 | 24 0 3712 ------------------------------------- 3713 Proc2 25 26 27 | 0 0 28 | 29 0 3714 30 0 0 | 31 32 33 | 0 34 3715 .ve 3716 3717 This can be represented as a collection of submatrices as: 3718 3719 .vb 3720 A B C 3721 D E F 3722 G H I 3723 .ve 3724 3725 Where the submatrices A,B,C are owned by proc0, D,E,F are 3726 owned by proc1, G,H,I are owned by proc2. 3727 3728 The 'm' parameters for proc0,proc1,proc2 are 3,3,2 respectively. 3729 The 'n' parameters for proc0,proc1,proc2 are 3,3,2 respectively. 3730 The 'M','N' parameters are 8,8, and have the same values on all procs. 3731 3732 The DIAGONAL submatrices corresponding to proc0,proc1,proc2 are 3733 submatrices [A], [E], [I] respectively. The OFF-DIAGONAL submatrices 3734 corresponding to proc0,proc1,proc2 are [BC], [DF], [GH] respectively. 3735 Internally, each processor stores the DIAGONAL part, and the OFF-DIAGONAL 3736 part as SeqAIJ matrices. for eg: proc1 will store [E] as a SeqAIJ 3737 matrix, ans [DF] as another SeqAIJ matrix. 3738 3739 When d_nz, o_nz parameters are specified, d_nz storage elements are 3740 allocated for every row of the local diagonal submatrix, and o_nz 3741 storage locations are allocated for every row of the OFF-DIAGONAL submat. 3742 One way to choose d_nz and o_nz is to use the max nonzerors per local 3743 rows for each of the local DIAGONAL, and the OFF-DIAGONAL submatrices. 3744 In this case, the values of d_nz,o_nz are: 3745 .vb 3746 proc0 : dnz = 2, o_nz = 2 3747 proc1 : dnz = 3, o_nz = 2 3748 proc2 : dnz = 1, o_nz = 4 3749 .ve 3750 We are allocating m*(d_nz+o_nz) storage locations for every proc. This 3751 translates to 3*(2+2)=12 for proc0, 3*(3+2)=15 for proc1, 2*(1+4)=10 3752 for proc3. i.e we are using 12+15+10=37 storage locations to store 3753 34 values. 3754 3755 When d_nnz, o_nnz parameters are specified, the storage is specified 3756 for every row, coresponding to both DIAGONAL and OFF-DIAGONAL submatrices. 3757 In the above case the values for d_nnz,o_nnz are: 3758 .vb 3759 proc0: d_nnz = [2,2,2] and o_nnz = [2,2,2] 3760 proc1: d_nnz = [3,3,2] and o_nnz = [2,1,1] 3761 proc2: d_nnz = [1,1] and o_nnz = [4,4] 3762 .ve 3763 Here the space allocated is sum of all the above values i.e 34, and 3764 hence pre-allocation is perfect. 3765 3766 Level: intermediate 3767 3768 .keywords: matrix, aij, compressed row, sparse, parallel 3769 3770 .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatMPIAIJSetPreallocationCSR(), 3771 MPIAIJ, MatCreateMPIAIJWithArrays() 3772 @*/ 3773 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) 3774 { 3775 PetscErrorCode ierr; 3776 PetscMPIInt size; 3777 3778 PetscFunctionBegin; 3779 ierr = MatCreate(comm,A);CHKERRQ(ierr); 3780 ierr = MatSetSizes(*A,m,n,M,N);CHKERRQ(ierr); 3781 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 3782 if (size > 1) { 3783 ierr = MatSetType(*A,MATMPIAIJ);CHKERRQ(ierr); 3784 ierr = MatMPIAIJSetPreallocation(*A,d_nz,d_nnz,o_nz,o_nnz);CHKERRQ(ierr); 3785 } else { 3786 ierr = MatSetType(*A,MATSEQAIJ);CHKERRQ(ierr); 3787 ierr = MatSeqAIJSetPreallocation(*A,d_nz,d_nnz);CHKERRQ(ierr); 3788 } 3789 PetscFunctionReturn(0); 3790 } 3791 3792 #undef __FUNCT__ 3793 #define __FUNCT__ "MatMPIAIJGetSeqAIJ" 3794 PetscErrorCode MatMPIAIJGetSeqAIJ(Mat A,Mat *Ad,Mat *Ao,const PetscInt *colmap[]) 3795 { 3796 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 3797 3798 PetscFunctionBegin; 3799 if (Ad) *Ad = a->A; 3800 if (Ao) *Ao = a->B; 3801 if (colmap) *colmap = a->garray; 3802 PetscFunctionReturn(0); 3803 } 3804 3805 #undef __FUNCT__ 3806 #define __FUNCT__ "MatSetColoring_MPIAIJ" 3807 PetscErrorCode MatSetColoring_MPIAIJ(Mat A,ISColoring coloring) 3808 { 3809 PetscErrorCode ierr; 3810 PetscInt i; 3811 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 3812 3813 PetscFunctionBegin; 3814 if (coloring->ctype == IS_COLORING_GLOBAL) { 3815 ISColoringValue *allcolors,*colors; 3816 ISColoring ocoloring; 3817 3818 /* set coloring for diagonal portion */ 3819 ierr = MatSetColoring_SeqAIJ(a->A,coloring);CHKERRQ(ierr); 3820 3821 /* set coloring for off-diagonal portion */ 3822 ierr = ISAllGatherColors(PetscObjectComm((PetscObject)A),coloring->n,coloring->colors,NULL,&allcolors);CHKERRQ(ierr); 3823 ierr = PetscMalloc1(a->B->cmap->n+1,&colors);CHKERRQ(ierr); 3824 for (i=0; i<a->B->cmap->n; i++) { 3825 colors[i] = allcolors[a->garray[i]]; 3826 } 3827 ierr = PetscFree(allcolors);CHKERRQ(ierr); 3828 ierr = ISColoringCreate(MPI_COMM_SELF,coloring->n,a->B->cmap->n,colors,PETSC_OWN_POINTER,&ocoloring);CHKERRQ(ierr); 3829 ierr = MatSetColoring_SeqAIJ(a->B,ocoloring);CHKERRQ(ierr); 3830 ierr = ISColoringDestroy(&ocoloring);CHKERRQ(ierr); 3831 } else if (coloring->ctype == IS_COLORING_GHOSTED) { 3832 ISColoringValue *colors; 3833 PetscInt *larray; 3834 ISColoring ocoloring; 3835 3836 /* set coloring for diagonal portion */ 3837 ierr = PetscMalloc1(a->A->cmap->n+1,&larray);CHKERRQ(ierr); 3838 for (i=0; i<a->A->cmap->n; i++) { 3839 larray[i] = i + A->cmap->rstart; 3840 } 3841 ierr = ISGlobalToLocalMappingApply(A->cmap->mapping,IS_GTOLM_MASK,a->A->cmap->n,larray,NULL,larray);CHKERRQ(ierr); 3842 ierr = PetscMalloc1(a->A->cmap->n+1,&colors);CHKERRQ(ierr); 3843 for (i=0; i<a->A->cmap->n; i++) { 3844 colors[i] = coloring->colors[larray[i]]; 3845 } 3846 ierr = PetscFree(larray);CHKERRQ(ierr); 3847 ierr = ISColoringCreate(PETSC_COMM_SELF,coloring->n,a->A->cmap->n,colors,PETSC_OWN_POINTER,&ocoloring);CHKERRQ(ierr); 3848 ierr = MatSetColoring_SeqAIJ(a->A,ocoloring);CHKERRQ(ierr); 3849 ierr = ISColoringDestroy(&ocoloring);CHKERRQ(ierr); 3850 3851 /* set coloring for off-diagonal portion */ 3852 ierr = PetscMalloc1(a->B->cmap->n+1,&larray);CHKERRQ(ierr); 3853 ierr = ISGlobalToLocalMappingApply(A->cmap->mapping,IS_GTOLM_MASK,a->B->cmap->n,a->garray,NULL,larray);CHKERRQ(ierr); 3854 ierr = PetscMalloc1(a->B->cmap->n+1,&colors);CHKERRQ(ierr); 3855 for (i=0; i<a->B->cmap->n; i++) { 3856 colors[i] = coloring->colors[larray[i]]; 3857 } 3858 ierr = PetscFree(larray);CHKERRQ(ierr); 3859 ierr = ISColoringCreate(MPI_COMM_SELF,coloring->n,a->B->cmap->n,colors,PETSC_OWN_POINTER,&ocoloring);CHKERRQ(ierr); 3860 ierr = MatSetColoring_SeqAIJ(a->B,ocoloring);CHKERRQ(ierr); 3861 ierr = ISColoringDestroy(&ocoloring);CHKERRQ(ierr); 3862 } else SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"No support ISColoringType %d",(int)coloring->ctype); 3863 PetscFunctionReturn(0); 3864 } 3865 3866 #undef __FUNCT__ 3867 #define __FUNCT__ "MatSetValuesAdifor_MPIAIJ" 3868 PetscErrorCode MatSetValuesAdifor_MPIAIJ(Mat A,PetscInt nl,void *advalues) 3869 { 3870 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 3871 PetscErrorCode ierr; 3872 3873 PetscFunctionBegin; 3874 ierr = MatSetValuesAdifor_SeqAIJ(a->A,nl,advalues);CHKERRQ(ierr); 3875 ierr = MatSetValuesAdifor_SeqAIJ(a->B,nl,advalues);CHKERRQ(ierr); 3876 PetscFunctionReturn(0); 3877 } 3878 3879 #undef __FUNCT__ 3880 #define __FUNCT__ "MatCreateMPIMatConcatenateSeqMat_MPIAIJ" 3881 PetscErrorCode MatCreateMPIMatConcatenateSeqMat_MPIAIJ(MPI_Comm comm,Mat inmat,PetscInt n,MatReuse scall,Mat *outmat) 3882 { 3883 PetscErrorCode ierr; 3884 PetscInt m,N,i,rstart,nnz,Ii; 3885 PetscInt *indx; 3886 PetscScalar *values; 3887 3888 PetscFunctionBegin; 3889 ierr = MatGetSize(inmat,&m,&N);CHKERRQ(ierr); 3890 if (scall == MAT_INITIAL_MATRIX) { /* symbolic phase */ 3891 PetscInt *dnz,*onz,sum,bs,cbs; 3892 3893 if (n == PETSC_DECIDE) { 3894 ierr = PetscSplitOwnership(comm,&n,&N);CHKERRQ(ierr); 3895 } 3896 /* Check sum(n) = N */ 3897 ierr = MPI_Allreduce(&n,&sum,1,MPIU_INT,MPI_SUM,comm);CHKERRQ(ierr); 3898 if (sum != N) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_INCOMP,"Sum of local columns != global columns %d",N); 3899 3900 ierr = MPI_Scan(&m, &rstart,1,MPIU_INT,MPI_SUM,comm);CHKERRQ(ierr); 3901 rstart -= m; 3902 3903 ierr = MatPreallocateInitialize(comm,m,n,dnz,onz);CHKERRQ(ierr); 3904 for (i=0; i<m; i++) { 3905 ierr = MatGetRow_SeqAIJ(inmat,i,&nnz,&indx,NULL);CHKERRQ(ierr); 3906 ierr = MatPreallocateSet(i+rstart,nnz,indx,dnz,onz);CHKERRQ(ierr); 3907 ierr = MatRestoreRow_SeqAIJ(inmat,i,&nnz,&indx,NULL);CHKERRQ(ierr); 3908 } 3909 3910 ierr = MatCreate(comm,outmat);CHKERRQ(ierr); 3911 ierr = MatSetSizes(*outmat,m,n,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 3912 ierr = MatGetBlockSizes(inmat,&bs,&cbs);CHKERRQ(ierr); 3913 ierr = MatSetBlockSizes(*outmat,bs,cbs);CHKERRQ(ierr); 3914 ierr = MatSetType(*outmat,MATMPIAIJ);CHKERRQ(ierr); 3915 ierr = MatMPIAIJSetPreallocation(*outmat,0,dnz,0,onz);CHKERRQ(ierr); 3916 ierr = MatPreallocateFinalize(dnz,onz);CHKERRQ(ierr); 3917 } 3918 3919 /* numeric phase */ 3920 ierr = MatGetOwnershipRange(*outmat,&rstart,NULL);CHKERRQ(ierr); 3921 for (i=0; i<m; i++) { 3922 ierr = MatGetRow_SeqAIJ(inmat,i,&nnz,&indx,&values);CHKERRQ(ierr); 3923 Ii = i + rstart; 3924 ierr = MatSetValues(*outmat,1,&Ii,nnz,indx,values,INSERT_VALUES);CHKERRQ(ierr); 3925 ierr = MatRestoreRow_SeqAIJ(inmat,i,&nnz,&indx,&values);CHKERRQ(ierr); 3926 } 3927 ierr = MatAssemblyBegin(*outmat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3928 ierr = MatAssemblyEnd(*outmat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3929 PetscFunctionReturn(0); 3930 } 3931 3932 #undef __FUNCT__ 3933 #define __FUNCT__ "MatFileSplit" 3934 PetscErrorCode MatFileSplit(Mat A,char *outfile) 3935 { 3936 PetscErrorCode ierr; 3937 PetscMPIInt rank; 3938 PetscInt m,N,i,rstart,nnz; 3939 size_t len; 3940 const PetscInt *indx; 3941 PetscViewer out; 3942 char *name; 3943 Mat B; 3944 const PetscScalar *values; 3945 3946 PetscFunctionBegin; 3947 ierr = MatGetLocalSize(A,&m,0);CHKERRQ(ierr); 3948 ierr = MatGetSize(A,0,&N);CHKERRQ(ierr); 3949 /* Should this be the type of the diagonal block of A? */ 3950 ierr = MatCreate(PETSC_COMM_SELF,&B);CHKERRQ(ierr); 3951 ierr = MatSetSizes(B,m,N,m,N);CHKERRQ(ierr); 3952 ierr = MatSetBlockSizesFromMats(B,A,A);CHKERRQ(ierr); 3953 ierr = MatSetType(B,MATSEQAIJ);CHKERRQ(ierr); 3954 ierr = MatSeqAIJSetPreallocation(B,0,NULL);CHKERRQ(ierr); 3955 ierr = MatGetOwnershipRange(A,&rstart,0);CHKERRQ(ierr); 3956 for (i=0; i<m; i++) { 3957 ierr = MatGetRow(A,i+rstart,&nnz,&indx,&values);CHKERRQ(ierr); 3958 ierr = MatSetValues(B,1,&i,nnz,indx,values,INSERT_VALUES);CHKERRQ(ierr); 3959 ierr = MatRestoreRow(A,i+rstart,&nnz,&indx,&values);CHKERRQ(ierr); 3960 } 3961 ierr = MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3962 ierr = MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3963 3964 ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)A),&rank);CHKERRQ(ierr); 3965 ierr = PetscStrlen(outfile,&len);CHKERRQ(ierr); 3966 ierr = PetscMalloc1(len+5,&name);CHKERRQ(ierr); 3967 sprintf(name,"%s.%d",outfile,rank); 3968 ierr = PetscViewerBinaryOpen(PETSC_COMM_SELF,name,FILE_MODE_APPEND,&out);CHKERRQ(ierr); 3969 ierr = PetscFree(name);CHKERRQ(ierr); 3970 ierr = MatView(B,out);CHKERRQ(ierr); 3971 ierr = PetscViewerDestroy(&out);CHKERRQ(ierr); 3972 ierr = MatDestroy(&B);CHKERRQ(ierr); 3973 PetscFunctionReturn(0); 3974 } 3975 3976 extern PetscErrorCode MatDestroy_MPIAIJ(Mat); 3977 #undef __FUNCT__ 3978 #define __FUNCT__ "MatDestroy_MPIAIJ_SeqsToMPI" 3979 PetscErrorCode MatDestroy_MPIAIJ_SeqsToMPI(Mat A) 3980 { 3981 PetscErrorCode ierr; 3982 Mat_Merge_SeqsToMPI *merge; 3983 PetscContainer container; 3984 3985 PetscFunctionBegin; 3986 ierr = PetscObjectQuery((PetscObject)A,"MatMergeSeqsToMPI",(PetscObject*)&container);CHKERRQ(ierr); 3987 if (container) { 3988 ierr = PetscContainerGetPointer(container,(void**)&merge);CHKERRQ(ierr); 3989 ierr = PetscFree(merge->id_r);CHKERRQ(ierr); 3990 ierr = PetscFree(merge->len_s);CHKERRQ(ierr); 3991 ierr = PetscFree(merge->len_r);CHKERRQ(ierr); 3992 ierr = PetscFree(merge->bi);CHKERRQ(ierr); 3993 ierr = PetscFree(merge->bj);CHKERRQ(ierr); 3994 ierr = PetscFree(merge->buf_ri[0]);CHKERRQ(ierr); 3995 ierr = PetscFree(merge->buf_ri);CHKERRQ(ierr); 3996 ierr = PetscFree(merge->buf_rj[0]);CHKERRQ(ierr); 3997 ierr = PetscFree(merge->buf_rj);CHKERRQ(ierr); 3998 ierr = PetscFree(merge->coi);CHKERRQ(ierr); 3999 ierr = PetscFree(merge->coj);CHKERRQ(ierr); 4000 ierr = PetscFree(merge->owners_co);CHKERRQ(ierr); 4001 ierr = PetscLayoutDestroy(&merge->rowmap);CHKERRQ(ierr); 4002 ierr = PetscFree(merge);CHKERRQ(ierr); 4003 ierr = PetscObjectCompose((PetscObject)A,"MatMergeSeqsToMPI",0);CHKERRQ(ierr); 4004 } 4005 ierr = MatDestroy_MPIAIJ(A);CHKERRQ(ierr); 4006 PetscFunctionReturn(0); 4007 } 4008 4009 #include <../src/mat/utils/freespace.h> 4010 #include <petscbt.h> 4011 4012 #undef __FUNCT__ 4013 #define __FUNCT__ "MatCreateMPIAIJSumSeqAIJNumeric" 4014 PetscErrorCode MatCreateMPIAIJSumSeqAIJNumeric(Mat seqmat,Mat mpimat) 4015 { 4016 PetscErrorCode ierr; 4017 MPI_Comm comm; 4018 Mat_SeqAIJ *a =(Mat_SeqAIJ*)seqmat->data; 4019 PetscMPIInt size,rank,taga,*len_s; 4020 PetscInt N=mpimat->cmap->N,i,j,*owners,*ai=a->i,*aj; 4021 PetscInt proc,m; 4022 PetscInt **buf_ri,**buf_rj; 4023 PetscInt k,anzi,*bj_i,*bi,*bj,arow,bnzi,nextaj; 4024 PetscInt nrows,**buf_ri_k,**nextrow,**nextai; 4025 MPI_Request *s_waits,*r_waits; 4026 MPI_Status *status; 4027 MatScalar *aa=a->a; 4028 MatScalar **abuf_r,*ba_i; 4029 Mat_Merge_SeqsToMPI *merge; 4030 PetscContainer container; 4031 4032 PetscFunctionBegin; 4033 ierr = PetscObjectGetComm((PetscObject)mpimat,&comm);CHKERRQ(ierr); 4034 ierr = PetscLogEventBegin(MAT_Seqstompinum,seqmat,0,0,0);CHKERRQ(ierr); 4035 4036 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 4037 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 4038 4039 ierr = PetscObjectQuery((PetscObject)mpimat,"MatMergeSeqsToMPI",(PetscObject*)&container);CHKERRQ(ierr); 4040 ierr = PetscContainerGetPointer(container,(void**)&merge);CHKERRQ(ierr); 4041 4042 bi = merge->bi; 4043 bj = merge->bj; 4044 buf_ri = merge->buf_ri; 4045 buf_rj = merge->buf_rj; 4046 4047 ierr = PetscMalloc1(size,&status);CHKERRQ(ierr); 4048 owners = merge->rowmap->range; 4049 len_s = merge->len_s; 4050 4051 /* send and recv matrix values */ 4052 /*-----------------------------*/ 4053 ierr = PetscObjectGetNewTag((PetscObject)mpimat,&taga);CHKERRQ(ierr); 4054 ierr = PetscPostIrecvScalar(comm,taga,merge->nrecv,merge->id_r,merge->len_r,&abuf_r,&r_waits);CHKERRQ(ierr); 4055 4056 ierr = PetscMalloc1(merge->nsend+1,&s_waits);CHKERRQ(ierr); 4057 for (proc=0,k=0; proc<size; proc++) { 4058 if (!len_s[proc]) continue; 4059 i = owners[proc]; 4060 ierr = MPI_Isend(aa+ai[i],len_s[proc],MPIU_MATSCALAR,proc,taga,comm,s_waits+k);CHKERRQ(ierr); 4061 k++; 4062 } 4063 4064 if (merge->nrecv) {ierr = MPI_Waitall(merge->nrecv,r_waits,status);CHKERRQ(ierr);} 4065 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,s_waits,status);CHKERRQ(ierr);} 4066 ierr = PetscFree(status);CHKERRQ(ierr); 4067 4068 ierr = PetscFree(s_waits);CHKERRQ(ierr); 4069 ierr = PetscFree(r_waits);CHKERRQ(ierr); 4070 4071 /* insert mat values of mpimat */ 4072 /*----------------------------*/ 4073 ierr = PetscMalloc1(N,&ba_i);CHKERRQ(ierr); 4074 ierr = PetscMalloc3(merge->nrecv,&buf_ri_k,merge->nrecv,&nextrow,merge->nrecv,&nextai);CHKERRQ(ierr); 4075 4076 for (k=0; k<merge->nrecv; k++) { 4077 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 4078 nrows = *(buf_ri_k[k]); 4079 nextrow[k] = buf_ri_k[k]+1; /* next row number of k-th recved i-structure */ 4080 nextai[k] = buf_ri_k[k] + (nrows + 1); /* poins to the next i-structure of k-th recved i-structure */ 4081 } 4082 4083 /* set values of ba */ 4084 m = merge->rowmap->n; 4085 for (i=0; i<m; i++) { 4086 arow = owners[rank] + i; 4087 bj_i = bj+bi[i]; /* col indices of the i-th row of mpimat */ 4088 bnzi = bi[i+1] - bi[i]; 4089 ierr = PetscMemzero(ba_i,bnzi*sizeof(PetscScalar));CHKERRQ(ierr); 4090 4091 /* add local non-zero vals of this proc's seqmat into ba */ 4092 anzi = ai[arow+1] - ai[arow]; 4093 aj = a->j + ai[arow]; 4094 aa = a->a + ai[arow]; 4095 nextaj = 0; 4096 for (j=0; nextaj<anzi; j++) { 4097 if (*(bj_i + j) == aj[nextaj]) { /* bcol == acol */ 4098 ba_i[j] += aa[nextaj++]; 4099 } 4100 } 4101 4102 /* add received vals into ba */ 4103 for (k=0; k<merge->nrecv; k++) { /* k-th received message */ 4104 /* i-th row */ 4105 if (i == *nextrow[k]) { 4106 anzi = *(nextai[k]+1) - *nextai[k]; 4107 aj = buf_rj[k] + *(nextai[k]); 4108 aa = abuf_r[k] + *(nextai[k]); 4109 nextaj = 0; 4110 for (j=0; nextaj<anzi; j++) { 4111 if (*(bj_i + j) == aj[nextaj]) { /* bcol == acol */ 4112 ba_i[j] += aa[nextaj++]; 4113 } 4114 } 4115 nextrow[k]++; nextai[k]++; 4116 } 4117 } 4118 ierr = MatSetValues(mpimat,1,&arow,bnzi,bj_i,ba_i,INSERT_VALUES);CHKERRQ(ierr); 4119 } 4120 ierr = MatAssemblyBegin(mpimat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4121 ierr = MatAssemblyEnd(mpimat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4122 4123 ierr = PetscFree(abuf_r[0]);CHKERRQ(ierr); 4124 ierr = PetscFree(abuf_r);CHKERRQ(ierr); 4125 ierr = PetscFree(ba_i);CHKERRQ(ierr); 4126 ierr = PetscFree3(buf_ri_k,nextrow,nextai);CHKERRQ(ierr); 4127 ierr = PetscLogEventEnd(MAT_Seqstompinum,seqmat,0,0,0);CHKERRQ(ierr); 4128 PetscFunctionReturn(0); 4129 } 4130 4131 extern PetscErrorCode MatDestroy_MPIAIJ_SeqsToMPI(Mat); 4132 4133 #undef __FUNCT__ 4134 #define __FUNCT__ "MatCreateMPIAIJSumSeqAIJSymbolic" 4135 PetscErrorCode MatCreateMPIAIJSumSeqAIJSymbolic(MPI_Comm comm,Mat seqmat,PetscInt m,PetscInt n,Mat *mpimat) 4136 { 4137 PetscErrorCode ierr; 4138 Mat B_mpi; 4139 Mat_SeqAIJ *a=(Mat_SeqAIJ*)seqmat->data; 4140 PetscMPIInt size,rank,tagi,tagj,*len_s,*len_si,*len_ri; 4141 PetscInt **buf_rj,**buf_ri,**buf_ri_k; 4142 PetscInt M=seqmat->rmap->n,N=seqmat->cmap->n,i,*owners,*ai=a->i,*aj=a->j; 4143 PetscInt len,proc,*dnz,*onz,bs,cbs; 4144 PetscInt k,anzi,*bi,*bj,*lnk,nlnk,arow,bnzi,nspacedouble=0; 4145 PetscInt nrows,*buf_s,*buf_si,*buf_si_i,**nextrow,**nextai; 4146 MPI_Request *si_waits,*sj_waits,*ri_waits,*rj_waits; 4147 MPI_Status *status; 4148 PetscFreeSpaceList free_space=NULL,current_space=NULL; 4149 PetscBT lnkbt; 4150 Mat_Merge_SeqsToMPI *merge; 4151 PetscContainer container; 4152 4153 PetscFunctionBegin; 4154 ierr = PetscLogEventBegin(MAT_Seqstompisym,seqmat,0,0,0);CHKERRQ(ierr); 4155 4156 /* make sure it is a PETSc comm */ 4157 ierr = PetscCommDuplicate(comm,&comm,NULL);CHKERRQ(ierr); 4158 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 4159 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 4160 4161 ierr = PetscNew(&merge);CHKERRQ(ierr); 4162 ierr = PetscMalloc1(size,&status);CHKERRQ(ierr); 4163 4164 /* determine row ownership */ 4165 /*---------------------------------------------------------*/ 4166 ierr = PetscLayoutCreate(comm,&merge->rowmap);CHKERRQ(ierr); 4167 ierr = PetscLayoutSetLocalSize(merge->rowmap,m);CHKERRQ(ierr); 4168 ierr = PetscLayoutSetSize(merge->rowmap,M);CHKERRQ(ierr); 4169 ierr = PetscLayoutSetBlockSize(merge->rowmap,1);CHKERRQ(ierr); 4170 ierr = PetscLayoutSetUp(merge->rowmap);CHKERRQ(ierr); 4171 ierr = PetscMalloc1(size,&len_si);CHKERRQ(ierr); 4172 ierr = PetscMalloc1(size,&merge->len_s);CHKERRQ(ierr); 4173 4174 m = merge->rowmap->n; 4175 owners = merge->rowmap->range; 4176 4177 /* determine the number of messages to send, their lengths */ 4178 /*---------------------------------------------------------*/ 4179 len_s = merge->len_s; 4180 4181 len = 0; /* length of buf_si[] */ 4182 merge->nsend = 0; 4183 for (proc=0; proc<size; proc++) { 4184 len_si[proc] = 0; 4185 if (proc == rank) { 4186 len_s[proc] = 0; 4187 } else { 4188 len_si[proc] = owners[proc+1] - owners[proc] + 1; 4189 len_s[proc] = ai[owners[proc+1]] - ai[owners[proc]]; /* num of rows to be sent to [proc] */ 4190 } 4191 if (len_s[proc]) { 4192 merge->nsend++; 4193 nrows = 0; 4194 for (i=owners[proc]; i<owners[proc+1]; i++) { 4195 if (ai[i+1] > ai[i]) nrows++; 4196 } 4197 len_si[proc] = 2*(nrows+1); 4198 len += len_si[proc]; 4199 } 4200 } 4201 4202 /* determine the number and length of messages to receive for ij-structure */ 4203 /*-------------------------------------------------------------------------*/ 4204 ierr = PetscGatherNumberOfMessages(comm,NULL,len_s,&merge->nrecv);CHKERRQ(ierr); 4205 ierr = PetscGatherMessageLengths2(comm,merge->nsend,merge->nrecv,len_s,len_si,&merge->id_r,&merge->len_r,&len_ri);CHKERRQ(ierr); 4206 4207 /* post the Irecv of j-structure */ 4208 /*-------------------------------*/ 4209 ierr = PetscCommGetNewTag(comm,&tagj);CHKERRQ(ierr); 4210 ierr = PetscPostIrecvInt(comm,tagj,merge->nrecv,merge->id_r,merge->len_r,&buf_rj,&rj_waits);CHKERRQ(ierr); 4211 4212 /* post the Isend of j-structure */ 4213 /*--------------------------------*/ 4214 ierr = PetscMalloc2(merge->nsend,&si_waits,merge->nsend,&sj_waits);CHKERRQ(ierr); 4215 4216 for (proc=0, k=0; proc<size; proc++) { 4217 if (!len_s[proc]) continue; 4218 i = owners[proc]; 4219 ierr = MPI_Isend(aj+ai[i],len_s[proc],MPIU_INT,proc,tagj,comm,sj_waits+k);CHKERRQ(ierr); 4220 k++; 4221 } 4222 4223 /* receives and sends of j-structure are complete */ 4224 /*------------------------------------------------*/ 4225 if (merge->nrecv) {ierr = MPI_Waitall(merge->nrecv,rj_waits,status);CHKERRQ(ierr);} 4226 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,sj_waits,status);CHKERRQ(ierr);} 4227 4228 /* send and recv i-structure */ 4229 /*---------------------------*/ 4230 ierr = PetscCommGetNewTag(comm,&tagi);CHKERRQ(ierr); 4231 ierr = PetscPostIrecvInt(comm,tagi,merge->nrecv,merge->id_r,len_ri,&buf_ri,&ri_waits);CHKERRQ(ierr); 4232 4233 ierr = PetscMalloc1(len+1,&buf_s);CHKERRQ(ierr); 4234 buf_si = buf_s; /* points to the beginning of k-th msg to be sent */ 4235 for (proc=0,k=0; proc<size; proc++) { 4236 if (!len_s[proc]) continue; 4237 /* form outgoing message for i-structure: 4238 buf_si[0]: nrows to be sent 4239 [1:nrows]: row index (global) 4240 [nrows+1:2*nrows+1]: i-structure index 4241 */ 4242 /*-------------------------------------------*/ 4243 nrows = len_si[proc]/2 - 1; 4244 buf_si_i = buf_si + nrows+1; 4245 buf_si[0] = nrows; 4246 buf_si_i[0] = 0; 4247 nrows = 0; 4248 for (i=owners[proc]; i<owners[proc+1]; i++) { 4249 anzi = ai[i+1] - ai[i]; 4250 if (anzi) { 4251 buf_si_i[nrows+1] = buf_si_i[nrows] + anzi; /* i-structure */ 4252 buf_si[nrows+1] = i-owners[proc]; /* local row index */ 4253 nrows++; 4254 } 4255 } 4256 ierr = MPI_Isend(buf_si,len_si[proc],MPIU_INT,proc,tagi,comm,si_waits+k);CHKERRQ(ierr); 4257 k++; 4258 buf_si += len_si[proc]; 4259 } 4260 4261 if (merge->nrecv) {ierr = MPI_Waitall(merge->nrecv,ri_waits,status);CHKERRQ(ierr);} 4262 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,si_waits,status);CHKERRQ(ierr);} 4263 4264 ierr = PetscInfo2(seqmat,"nsend: %D, nrecv: %D\n",merge->nsend,merge->nrecv);CHKERRQ(ierr); 4265 for (i=0; i<merge->nrecv; i++) { 4266 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); 4267 } 4268 4269 ierr = PetscFree(len_si);CHKERRQ(ierr); 4270 ierr = PetscFree(len_ri);CHKERRQ(ierr); 4271 ierr = PetscFree(rj_waits);CHKERRQ(ierr); 4272 ierr = PetscFree2(si_waits,sj_waits);CHKERRQ(ierr); 4273 ierr = PetscFree(ri_waits);CHKERRQ(ierr); 4274 ierr = PetscFree(buf_s);CHKERRQ(ierr); 4275 ierr = PetscFree(status);CHKERRQ(ierr); 4276 4277 /* compute a local seq matrix in each processor */ 4278 /*----------------------------------------------*/ 4279 /* allocate bi array and free space for accumulating nonzero column info */ 4280 ierr = PetscMalloc1(m+1,&bi);CHKERRQ(ierr); 4281 bi[0] = 0; 4282 4283 /* create and initialize a linked list */ 4284 nlnk = N+1; 4285 ierr = PetscLLCreate(N,N,nlnk,lnk,lnkbt);CHKERRQ(ierr); 4286 4287 /* initial FreeSpace size is 2*(num of local nnz(seqmat)) */ 4288 len = ai[owners[rank+1]] - ai[owners[rank]]; 4289 ierr = PetscFreeSpaceGet((PetscInt)(2*len+1),&free_space);CHKERRQ(ierr); 4290 4291 current_space = free_space; 4292 4293 /* determine symbolic info for each local row */ 4294 ierr = PetscMalloc3(merge->nrecv,&buf_ri_k,merge->nrecv,&nextrow,merge->nrecv,&nextai);CHKERRQ(ierr); 4295 4296 for (k=0; k<merge->nrecv; k++) { 4297 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 4298 nrows = *buf_ri_k[k]; 4299 nextrow[k] = buf_ri_k[k] + 1; /* next row number of k-th recved i-structure */ 4300 nextai[k] = buf_ri_k[k] + (nrows + 1); /* poins to the next i-structure of k-th recved i-structure */ 4301 } 4302 4303 ierr = MatPreallocateInitialize(comm,m,n,dnz,onz);CHKERRQ(ierr); 4304 len = 0; 4305 for (i=0; i<m; i++) { 4306 bnzi = 0; 4307 /* add local non-zero cols of this proc's seqmat into lnk */ 4308 arow = owners[rank] + i; 4309 anzi = ai[arow+1] - ai[arow]; 4310 aj = a->j + ai[arow]; 4311 ierr = PetscLLAddSorted(anzi,aj,N,nlnk,lnk,lnkbt);CHKERRQ(ierr); 4312 bnzi += nlnk; 4313 /* add received col data into lnk */ 4314 for (k=0; k<merge->nrecv; k++) { /* k-th received message */ 4315 if (i == *nextrow[k]) { /* i-th row */ 4316 anzi = *(nextai[k]+1) - *nextai[k]; 4317 aj = buf_rj[k] + *nextai[k]; 4318 ierr = PetscLLAddSorted(anzi,aj,N,nlnk,lnk,lnkbt);CHKERRQ(ierr); 4319 bnzi += nlnk; 4320 nextrow[k]++; nextai[k]++; 4321 } 4322 } 4323 if (len < bnzi) len = bnzi; /* =max(bnzi) */ 4324 4325 /* if free space is not available, make more free space */ 4326 if (current_space->local_remaining<bnzi) { 4327 ierr = PetscFreeSpaceGet(bnzi+current_space->total_array_size,¤t_space);CHKERRQ(ierr); 4328 nspacedouble++; 4329 } 4330 /* copy data into free space, then initialize lnk */ 4331 ierr = PetscLLClean(N,N,bnzi,lnk,current_space->array,lnkbt);CHKERRQ(ierr); 4332 ierr = MatPreallocateSet(i+owners[rank],bnzi,current_space->array,dnz,onz);CHKERRQ(ierr); 4333 4334 current_space->array += bnzi; 4335 current_space->local_used += bnzi; 4336 current_space->local_remaining -= bnzi; 4337 4338 bi[i+1] = bi[i] + bnzi; 4339 } 4340 4341 ierr = PetscFree3(buf_ri_k,nextrow,nextai);CHKERRQ(ierr); 4342 4343 ierr = PetscMalloc1(bi[m]+1,&bj);CHKERRQ(ierr); 4344 ierr = PetscFreeSpaceContiguous(&free_space,bj);CHKERRQ(ierr); 4345 ierr = PetscLLDestroy(lnk,lnkbt);CHKERRQ(ierr); 4346 4347 /* create symbolic parallel matrix B_mpi */ 4348 /*---------------------------------------*/ 4349 ierr = MatGetBlockSizes(seqmat,&bs,&cbs);CHKERRQ(ierr); 4350 ierr = MatCreate(comm,&B_mpi);CHKERRQ(ierr); 4351 if (n==PETSC_DECIDE) { 4352 ierr = MatSetSizes(B_mpi,m,n,PETSC_DETERMINE,N);CHKERRQ(ierr); 4353 } else { 4354 ierr = MatSetSizes(B_mpi,m,n,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 4355 } 4356 ierr = MatSetBlockSizes(B_mpi,bs,cbs);CHKERRQ(ierr); 4357 ierr = MatSetType(B_mpi,MATMPIAIJ);CHKERRQ(ierr); 4358 ierr = MatMPIAIJSetPreallocation(B_mpi,0,dnz,0,onz);CHKERRQ(ierr); 4359 ierr = MatPreallocateFinalize(dnz,onz);CHKERRQ(ierr); 4360 ierr = MatSetOption(B_mpi,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 4361 4362 /* B_mpi is not ready for use - assembly will be done by MatCreateMPIAIJSumSeqAIJNumeric() */ 4363 B_mpi->assembled = PETSC_FALSE; 4364 B_mpi->ops->destroy = MatDestroy_MPIAIJ_SeqsToMPI; 4365 merge->bi = bi; 4366 merge->bj = bj; 4367 merge->buf_ri = buf_ri; 4368 merge->buf_rj = buf_rj; 4369 merge->coi = NULL; 4370 merge->coj = NULL; 4371 merge->owners_co = NULL; 4372 4373 ierr = PetscCommDestroy(&comm);CHKERRQ(ierr); 4374 4375 /* attach the supporting struct to B_mpi for reuse */ 4376 ierr = PetscContainerCreate(PETSC_COMM_SELF,&container);CHKERRQ(ierr); 4377 ierr = PetscContainerSetPointer(container,merge);CHKERRQ(ierr); 4378 ierr = PetscObjectCompose((PetscObject)B_mpi,"MatMergeSeqsToMPI",(PetscObject)container);CHKERRQ(ierr); 4379 ierr = PetscContainerDestroy(&container);CHKERRQ(ierr); 4380 *mpimat = B_mpi; 4381 4382 ierr = PetscLogEventEnd(MAT_Seqstompisym,seqmat,0,0,0);CHKERRQ(ierr); 4383 PetscFunctionReturn(0); 4384 } 4385 4386 #undef __FUNCT__ 4387 #define __FUNCT__ "MatCreateMPIAIJSumSeqAIJ" 4388 /*@C 4389 MatCreateMPIAIJSumSeqAIJ - Creates a MPIAIJ matrix by adding sequential 4390 matrices from each processor 4391 4392 Collective on MPI_Comm 4393 4394 Input Parameters: 4395 + comm - the communicators the parallel matrix will live on 4396 . seqmat - the input sequential matrices 4397 . m - number of local rows (or PETSC_DECIDE) 4398 . n - number of local columns (or PETSC_DECIDE) 4399 - scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX 4400 4401 Output Parameter: 4402 . mpimat - the parallel matrix generated 4403 4404 Level: advanced 4405 4406 Notes: 4407 The dimensions of the sequential matrix in each processor MUST be the same. 4408 The input seqmat is included into the container "Mat_Merge_SeqsToMPI", and will be 4409 destroyed when mpimat is destroyed. Call PetscObjectQuery() to access seqmat. 4410 @*/ 4411 PetscErrorCode MatCreateMPIAIJSumSeqAIJ(MPI_Comm comm,Mat seqmat,PetscInt m,PetscInt n,MatReuse scall,Mat *mpimat) 4412 { 4413 PetscErrorCode ierr; 4414 PetscMPIInt size; 4415 4416 PetscFunctionBegin; 4417 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 4418 if (size == 1) { 4419 ierr = PetscLogEventBegin(MAT_Seqstompi,seqmat,0,0,0);CHKERRQ(ierr); 4420 if (scall == MAT_INITIAL_MATRIX) { 4421 ierr = MatDuplicate(seqmat,MAT_COPY_VALUES,mpimat);CHKERRQ(ierr); 4422 } else { 4423 ierr = MatCopy(seqmat,*mpimat,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 4424 } 4425 ierr = PetscLogEventEnd(MAT_Seqstompi,seqmat,0,0,0);CHKERRQ(ierr); 4426 PetscFunctionReturn(0); 4427 } 4428 ierr = PetscLogEventBegin(MAT_Seqstompi,seqmat,0,0,0);CHKERRQ(ierr); 4429 if (scall == MAT_INITIAL_MATRIX) { 4430 ierr = MatCreateMPIAIJSumSeqAIJSymbolic(comm,seqmat,m,n,mpimat);CHKERRQ(ierr); 4431 } 4432 ierr = MatCreateMPIAIJSumSeqAIJNumeric(seqmat,*mpimat);CHKERRQ(ierr); 4433 ierr = PetscLogEventEnd(MAT_Seqstompi,seqmat,0,0,0);CHKERRQ(ierr); 4434 PetscFunctionReturn(0); 4435 } 4436 4437 #undef __FUNCT__ 4438 #define __FUNCT__ "MatMPIAIJGetLocalMat" 4439 /*@ 4440 MatMPIAIJGetLocalMat - Creates a SeqAIJ from a MPIAIJ matrix by taking all its local rows and putting them into a sequential vector with 4441 mlocal rows and n columns. Where mlocal is the row count obtained with MatGetLocalSize() and n is the global column count obtained 4442 with MatGetSize() 4443 4444 Not Collective 4445 4446 Input Parameters: 4447 + A - the matrix 4448 . scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX 4449 4450 Output Parameter: 4451 . A_loc - the local sequential matrix generated 4452 4453 Level: developer 4454 4455 .seealso: MatGetOwnerShipRange(), MatMPIAIJGetLocalMatCondensed() 4456 4457 @*/ 4458 PetscErrorCode MatMPIAIJGetLocalMat(Mat A,MatReuse scall,Mat *A_loc) 4459 { 4460 PetscErrorCode ierr; 4461 Mat_MPIAIJ *mpimat=(Mat_MPIAIJ*)A->data; 4462 Mat_SeqAIJ *mat,*a,*b; 4463 PetscInt *ai,*aj,*bi,*bj,*cmap=mpimat->garray; 4464 MatScalar *aa,*ba,*cam; 4465 PetscScalar *ca; 4466 PetscInt am=A->rmap->n,i,j,k,cstart=A->cmap->rstart; 4467 PetscInt *ci,*cj,col,ncols_d,ncols_o,jo; 4468 PetscBool match; 4469 MPI_Comm comm; 4470 PetscMPIInt size; 4471 4472 PetscFunctionBegin; 4473 ierr = PetscObjectTypeCompare((PetscObject)A,MATMPIAIJ,&match);CHKERRQ(ierr); 4474 if (!match) SETERRQ(PetscObjectComm((PetscObject)A), PETSC_ERR_SUP,"Requires MPIAIJ matrix as input"); 4475 ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 4476 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 4477 if (size == 1 && scall == MAT_REUSE_MATRIX) PetscFunctionReturn(0); 4478 4479 ierr = PetscLogEventBegin(MAT_Getlocalmat,A,0,0,0);CHKERRQ(ierr); 4480 a = (Mat_SeqAIJ*)(mpimat->A)->data; 4481 b = (Mat_SeqAIJ*)(mpimat->B)->data; 4482 ai = a->i; aj = a->j; bi = b->i; bj = b->j; 4483 aa = a->a; ba = b->a; 4484 if (scall == MAT_INITIAL_MATRIX) { 4485 if (size == 1) { 4486 ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,am,A->cmap->N,ai,aj,aa,A_loc);CHKERRQ(ierr); 4487 PetscFunctionReturn(0); 4488 } 4489 4490 ierr = PetscMalloc1(1+am,&ci);CHKERRQ(ierr); 4491 ci[0] = 0; 4492 for (i=0; i<am; i++) { 4493 ci[i+1] = ci[i] + (ai[i+1] - ai[i]) + (bi[i+1] - bi[i]); 4494 } 4495 ierr = PetscMalloc1(1+ci[am],&cj);CHKERRQ(ierr); 4496 ierr = PetscMalloc1(1+ci[am],&ca);CHKERRQ(ierr); 4497 k = 0; 4498 for (i=0; i<am; i++) { 4499 ncols_o = bi[i+1] - bi[i]; 4500 ncols_d = ai[i+1] - ai[i]; 4501 /* off-diagonal portion of A */ 4502 for (jo=0; jo<ncols_o; jo++) { 4503 col = cmap[*bj]; 4504 if (col >= cstart) break; 4505 cj[k] = col; bj++; 4506 ca[k++] = *ba++; 4507 } 4508 /* diagonal portion of A */ 4509 for (j=0; j<ncols_d; j++) { 4510 cj[k] = cstart + *aj++; 4511 ca[k++] = *aa++; 4512 } 4513 /* off-diagonal portion of A */ 4514 for (j=jo; j<ncols_o; j++) { 4515 cj[k] = cmap[*bj++]; 4516 ca[k++] = *ba++; 4517 } 4518 } 4519 /* put together the new matrix */ 4520 ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,am,A->cmap->N,ci,cj,ca,A_loc);CHKERRQ(ierr); 4521 /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */ 4522 /* Since these are PETSc arrays, change flags to free them as necessary. */ 4523 mat = (Mat_SeqAIJ*)(*A_loc)->data; 4524 mat->free_a = PETSC_TRUE; 4525 mat->free_ij = PETSC_TRUE; 4526 mat->nonew = 0; 4527 } else if (scall == MAT_REUSE_MATRIX) { 4528 mat=(Mat_SeqAIJ*)(*A_loc)->data; 4529 ci = mat->i; cj = mat->j; cam = mat->a; 4530 for (i=0; i<am; i++) { 4531 /* off-diagonal portion of A */ 4532 ncols_o = bi[i+1] - bi[i]; 4533 for (jo=0; jo<ncols_o; jo++) { 4534 col = cmap[*bj]; 4535 if (col >= cstart) break; 4536 *cam++ = *ba++; bj++; 4537 } 4538 /* diagonal portion of A */ 4539 ncols_d = ai[i+1] - ai[i]; 4540 for (j=0; j<ncols_d; j++) *cam++ = *aa++; 4541 /* off-diagonal portion of A */ 4542 for (j=jo; j<ncols_o; j++) { 4543 *cam++ = *ba++; bj++; 4544 } 4545 } 4546 } else SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Invalid MatReuse %d",(int)scall); 4547 ierr = PetscLogEventEnd(MAT_Getlocalmat,A,0,0,0);CHKERRQ(ierr); 4548 PetscFunctionReturn(0); 4549 } 4550 4551 #undef __FUNCT__ 4552 #define __FUNCT__ "MatMPIAIJGetLocalMatCondensed" 4553 /*@C 4554 MatMPIAIJGetLocalMatCondensed - Creates a SeqAIJ matrix from an MPIAIJ matrix by taking all its local rows and NON-ZERO columns 4555 4556 Not Collective 4557 4558 Input Parameters: 4559 + A - the matrix 4560 . scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX 4561 - row, col - index sets of rows and columns to extract (or NULL) 4562 4563 Output Parameter: 4564 . A_loc - the local sequential matrix generated 4565 4566 Level: developer 4567 4568 .seealso: MatGetOwnershipRange(), MatMPIAIJGetLocalMat() 4569 4570 @*/ 4571 PetscErrorCode MatMPIAIJGetLocalMatCondensed(Mat A,MatReuse scall,IS *row,IS *col,Mat *A_loc) 4572 { 4573 Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data; 4574 PetscErrorCode ierr; 4575 PetscInt i,start,end,ncols,nzA,nzB,*cmap,imark,*idx; 4576 IS isrowa,iscola; 4577 Mat *aloc; 4578 PetscBool match; 4579 4580 PetscFunctionBegin; 4581 ierr = PetscObjectTypeCompare((PetscObject)A,MATMPIAIJ,&match);CHKERRQ(ierr); 4582 if (!match) SETERRQ(PetscObjectComm((PetscObject)A), PETSC_ERR_SUP,"Requires MPIAIJ matrix as input"); 4583 ierr = PetscLogEventBegin(MAT_Getlocalmatcondensed,A,0,0,0);CHKERRQ(ierr); 4584 if (!row) { 4585 start = A->rmap->rstart; end = A->rmap->rend; 4586 ierr = ISCreateStride(PETSC_COMM_SELF,end-start,start,1,&isrowa);CHKERRQ(ierr); 4587 } else { 4588 isrowa = *row; 4589 } 4590 if (!col) { 4591 start = A->cmap->rstart; 4592 cmap = a->garray; 4593 nzA = a->A->cmap->n; 4594 nzB = a->B->cmap->n; 4595 ierr = PetscMalloc1(nzA+nzB, &idx);CHKERRQ(ierr); 4596 ncols = 0; 4597 for (i=0; i<nzB; i++) { 4598 if (cmap[i] < start) idx[ncols++] = cmap[i]; 4599 else break; 4600 } 4601 imark = i; 4602 for (i=0; i<nzA; i++) idx[ncols++] = start + i; 4603 for (i=imark; i<nzB; i++) idx[ncols++] = cmap[i]; 4604 ierr = ISCreateGeneral(PETSC_COMM_SELF,ncols,idx,PETSC_OWN_POINTER,&iscola);CHKERRQ(ierr); 4605 } else { 4606 iscola = *col; 4607 } 4608 if (scall != MAT_INITIAL_MATRIX) { 4609 ierr = PetscMalloc1(1,&aloc);CHKERRQ(ierr); 4610 aloc[0] = *A_loc; 4611 } 4612 ierr = MatGetSubMatrices(A,1,&isrowa,&iscola,scall,&aloc);CHKERRQ(ierr); 4613 *A_loc = aloc[0]; 4614 ierr = PetscFree(aloc);CHKERRQ(ierr); 4615 if (!row) { 4616 ierr = ISDestroy(&isrowa);CHKERRQ(ierr); 4617 } 4618 if (!col) { 4619 ierr = ISDestroy(&iscola);CHKERRQ(ierr); 4620 } 4621 ierr = PetscLogEventEnd(MAT_Getlocalmatcondensed,A,0,0,0);CHKERRQ(ierr); 4622 PetscFunctionReturn(0); 4623 } 4624 4625 #undef __FUNCT__ 4626 #define __FUNCT__ "MatGetBrowsOfAcols" 4627 /*@C 4628 MatGetBrowsOfAcols - Creates a SeqAIJ matrix by taking rows of B that equal to nonzero columns of local A 4629 4630 Collective on Mat 4631 4632 Input Parameters: 4633 + A,B - the matrices in mpiaij format 4634 . scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX 4635 - rowb, colb - index sets of rows and columns of B to extract (or NULL) 4636 4637 Output Parameter: 4638 + rowb, colb - index sets of rows and columns of B to extract 4639 - B_seq - the sequential matrix generated 4640 4641 Level: developer 4642 4643 @*/ 4644 PetscErrorCode MatGetBrowsOfAcols(Mat A,Mat B,MatReuse scall,IS *rowb,IS *colb,Mat *B_seq) 4645 { 4646 Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data; 4647 PetscErrorCode ierr; 4648 PetscInt *idx,i,start,ncols,nzA,nzB,*cmap,imark; 4649 IS isrowb,iscolb; 4650 Mat *bseq=NULL; 4651 4652 PetscFunctionBegin; 4653 if (A->cmap->rstart != B->rmap->rstart || A->cmap->rend != B->rmap->rend) { 4654 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); 4655 } 4656 ierr = PetscLogEventBegin(MAT_GetBrowsOfAcols,A,B,0,0);CHKERRQ(ierr); 4657 4658 if (scall == MAT_INITIAL_MATRIX) { 4659 start = A->cmap->rstart; 4660 cmap = a->garray; 4661 nzA = a->A->cmap->n; 4662 nzB = a->B->cmap->n; 4663 ierr = PetscMalloc1(nzA+nzB, &idx);CHKERRQ(ierr); 4664 ncols = 0; 4665 for (i=0; i<nzB; i++) { /* row < local row index */ 4666 if (cmap[i] < start) idx[ncols++] = cmap[i]; 4667 else break; 4668 } 4669 imark = i; 4670 for (i=0; i<nzA; i++) idx[ncols++] = start + i; /* local rows */ 4671 for (i=imark; i<nzB; i++) idx[ncols++] = cmap[i]; /* row > local row index */ 4672 ierr = ISCreateGeneral(PETSC_COMM_SELF,ncols,idx,PETSC_OWN_POINTER,&isrowb);CHKERRQ(ierr); 4673 ierr = ISCreateStride(PETSC_COMM_SELF,B->cmap->N,0,1,&iscolb);CHKERRQ(ierr); 4674 } else { 4675 if (!rowb || !colb) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"IS rowb and colb must be provided for MAT_REUSE_MATRIX"); 4676 isrowb = *rowb; iscolb = *colb; 4677 ierr = PetscMalloc1(1,&bseq);CHKERRQ(ierr); 4678 bseq[0] = *B_seq; 4679 } 4680 ierr = MatGetSubMatrices(B,1,&isrowb,&iscolb,scall,&bseq);CHKERRQ(ierr); 4681 *B_seq = bseq[0]; 4682 ierr = PetscFree(bseq);CHKERRQ(ierr); 4683 if (!rowb) { 4684 ierr = ISDestroy(&isrowb);CHKERRQ(ierr); 4685 } else { 4686 *rowb = isrowb; 4687 } 4688 if (!colb) { 4689 ierr = ISDestroy(&iscolb);CHKERRQ(ierr); 4690 } else { 4691 *colb = iscolb; 4692 } 4693 ierr = PetscLogEventEnd(MAT_GetBrowsOfAcols,A,B,0,0);CHKERRQ(ierr); 4694 PetscFunctionReturn(0); 4695 } 4696 4697 #undef __FUNCT__ 4698 #define __FUNCT__ "MatGetBrowsOfAoCols_MPIAIJ" 4699 /* 4700 MatGetBrowsOfAoCols_MPIAIJ - Creates a SeqAIJ matrix by taking rows of B that equal to nonzero columns 4701 of the OFF-DIAGONAL portion of local A 4702 4703 Collective on Mat 4704 4705 Input Parameters: 4706 + A,B - the matrices in mpiaij format 4707 - scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX 4708 4709 Output Parameter: 4710 + startsj_s - starting point in B's sending j-arrays, saved for MAT_REUSE (or NULL) 4711 . startsj_r - starting point in B's receiving j-arrays, saved for MAT_REUSE (or NULL) 4712 . bufa_ptr - array for sending matrix values, saved for MAT_REUSE (or NULL) 4713 - B_oth - the sequential matrix generated with size aBn=a->B->cmap->n by B->cmap->N 4714 4715 Level: developer 4716 4717 */ 4718 PetscErrorCode MatGetBrowsOfAoCols_MPIAIJ(Mat A,Mat B,MatReuse scall,PetscInt **startsj_s,PetscInt **startsj_r,MatScalar **bufa_ptr,Mat *B_oth) 4719 { 4720 VecScatter_MPI_General *gen_to,*gen_from; 4721 PetscErrorCode ierr; 4722 Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data; 4723 Mat_SeqAIJ *b_oth; 4724 VecScatter ctx =a->Mvctx; 4725 MPI_Comm comm; 4726 PetscMPIInt *rprocs,*sprocs,tag=((PetscObject)ctx)->tag,rank; 4727 PetscInt *rowlen,*bufj,*bufJ,ncols,aBn=a->B->cmap->n,row,*b_othi,*b_othj; 4728 PetscScalar *rvalues,*svalues; 4729 MatScalar *b_otha,*bufa,*bufA; 4730 PetscInt i,j,k,l,ll,nrecvs,nsends,nrows,*srow,*rstarts,*rstartsj = 0,*sstarts,*sstartsj,len; 4731 MPI_Request *rwaits = NULL,*swaits = NULL; 4732 MPI_Status *sstatus,rstatus; 4733 PetscMPIInt jj,size; 4734 PetscInt *cols,sbs,rbs; 4735 PetscScalar *vals; 4736 4737 PetscFunctionBegin; 4738 ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 4739 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 4740 4741 if (A->cmap->rstart != B->rmap->rstart || A->cmap->rend != B->rmap->rend) { 4742 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); 4743 } 4744 ierr = PetscLogEventBegin(MAT_GetBrowsOfAocols,A,B,0,0);CHKERRQ(ierr); 4745 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 4746 4747 gen_to = (VecScatter_MPI_General*)ctx->todata; 4748 gen_from = (VecScatter_MPI_General*)ctx->fromdata; 4749 rvalues = gen_from->values; /* holds the length of receiving row */ 4750 svalues = gen_to->values; /* holds the length of sending row */ 4751 nrecvs = gen_from->n; 4752 nsends = gen_to->n; 4753 4754 ierr = PetscMalloc2(nrecvs,&rwaits,nsends,&swaits);CHKERRQ(ierr); 4755 srow = gen_to->indices; /* local row index to be sent */ 4756 sstarts = gen_to->starts; 4757 sprocs = gen_to->procs; 4758 sstatus = gen_to->sstatus; 4759 sbs = gen_to->bs; 4760 rstarts = gen_from->starts; 4761 rprocs = gen_from->procs; 4762 rbs = gen_from->bs; 4763 4764 if (!startsj_s || !bufa_ptr) scall = MAT_INITIAL_MATRIX; 4765 if (scall == MAT_INITIAL_MATRIX) { 4766 /* i-array */ 4767 /*---------*/ 4768 /* post receives */ 4769 for (i=0; i<nrecvs; i++) { 4770 rowlen = (PetscInt*)rvalues + rstarts[i]*rbs; 4771 nrows = (rstarts[i+1]-rstarts[i])*rbs; /* num of indices to be received */ 4772 ierr = MPI_Irecv(rowlen,nrows,MPIU_INT,rprocs[i],tag,comm,rwaits+i);CHKERRQ(ierr); 4773 } 4774 4775 /* pack the outgoing message */ 4776 ierr = PetscMalloc2(nsends+1,&sstartsj,nrecvs+1,&rstartsj);CHKERRQ(ierr); 4777 4778 sstartsj[0] = 0; 4779 rstartsj[0] = 0; 4780 len = 0; /* total length of j or a array to be sent */ 4781 k = 0; 4782 for (i=0; i<nsends; i++) { 4783 rowlen = (PetscInt*)svalues + sstarts[i]*sbs; 4784 nrows = sstarts[i+1]-sstarts[i]; /* num of block rows */ 4785 for (j=0; j<nrows; j++) { 4786 row = srow[k] + B->rmap->range[rank]; /* global row idx */ 4787 for (l=0; l<sbs; l++) { 4788 ierr = MatGetRow_MPIAIJ(B,row+l,&ncols,NULL,NULL);CHKERRQ(ierr); /* rowlength */ 4789 4790 rowlen[j*sbs+l] = ncols; 4791 4792 len += ncols; 4793 ierr = MatRestoreRow_MPIAIJ(B,row+l,&ncols,NULL,NULL);CHKERRQ(ierr); 4794 } 4795 k++; 4796 } 4797 ierr = MPI_Isend(rowlen,nrows*sbs,MPIU_INT,sprocs[i],tag,comm,swaits+i);CHKERRQ(ierr); 4798 4799 sstartsj[i+1] = len; /* starting point of (i+1)-th outgoing msg in bufj and bufa */ 4800 } 4801 /* recvs and sends of i-array are completed */ 4802 i = nrecvs; 4803 while (i--) { 4804 ierr = MPI_Waitany(nrecvs,rwaits,&jj,&rstatus);CHKERRQ(ierr); 4805 } 4806 if (nsends) {ierr = MPI_Waitall(nsends,swaits,sstatus);CHKERRQ(ierr);} 4807 4808 /* allocate buffers for sending j and a arrays */ 4809 ierr = PetscMalloc1(len+1,&bufj);CHKERRQ(ierr); 4810 ierr = PetscMalloc1(len+1,&bufa);CHKERRQ(ierr); 4811 4812 /* create i-array of B_oth */ 4813 ierr = PetscMalloc1(aBn+2,&b_othi);CHKERRQ(ierr); 4814 4815 b_othi[0] = 0; 4816 len = 0; /* total length of j or a array to be received */ 4817 k = 0; 4818 for (i=0; i<nrecvs; i++) { 4819 rowlen = (PetscInt*)rvalues + rstarts[i]*rbs; 4820 nrows = rbs*(rstarts[i+1]-rstarts[i]); /* num of rows to be recieved */ 4821 for (j=0; j<nrows; j++) { 4822 b_othi[k+1] = b_othi[k] + rowlen[j]; 4823 len += rowlen[j]; k++; 4824 } 4825 rstartsj[i+1] = len; /* starting point of (i+1)-th incoming msg in bufj and bufa */ 4826 } 4827 4828 /* allocate space for j and a arrrays of B_oth */ 4829 ierr = PetscMalloc1(b_othi[aBn]+1,&b_othj);CHKERRQ(ierr); 4830 ierr = PetscMalloc1(b_othi[aBn]+1,&b_otha);CHKERRQ(ierr); 4831 4832 /* j-array */ 4833 /*---------*/ 4834 /* post receives of j-array */ 4835 for (i=0; i<nrecvs; i++) { 4836 nrows = rstartsj[i+1]-rstartsj[i]; /* length of the msg received */ 4837 ierr = MPI_Irecv(b_othj+rstartsj[i],nrows,MPIU_INT,rprocs[i],tag,comm,rwaits+i);CHKERRQ(ierr); 4838 } 4839 4840 /* pack the outgoing message j-array */ 4841 k = 0; 4842 for (i=0; i<nsends; i++) { 4843 nrows = sstarts[i+1]-sstarts[i]; /* num of block rows */ 4844 bufJ = bufj+sstartsj[i]; 4845 for (j=0; j<nrows; j++) { 4846 row = srow[k++] + B->rmap->range[rank]; /* global row idx */ 4847 for (ll=0; ll<sbs; ll++) { 4848 ierr = MatGetRow_MPIAIJ(B,row+ll,&ncols,&cols,NULL);CHKERRQ(ierr); 4849 for (l=0; l<ncols; l++) { 4850 *bufJ++ = cols[l]; 4851 } 4852 ierr = MatRestoreRow_MPIAIJ(B,row+ll,&ncols,&cols,NULL);CHKERRQ(ierr); 4853 } 4854 } 4855 ierr = MPI_Isend(bufj+sstartsj[i],sstartsj[i+1]-sstartsj[i],MPIU_INT,sprocs[i],tag,comm,swaits+i);CHKERRQ(ierr); 4856 } 4857 4858 /* recvs and sends of j-array are completed */ 4859 i = nrecvs; 4860 while (i--) { 4861 ierr = MPI_Waitany(nrecvs,rwaits,&jj,&rstatus);CHKERRQ(ierr); 4862 } 4863 if (nsends) {ierr = MPI_Waitall(nsends,swaits,sstatus);CHKERRQ(ierr);} 4864 } else if (scall == MAT_REUSE_MATRIX) { 4865 sstartsj = *startsj_s; 4866 rstartsj = *startsj_r; 4867 bufa = *bufa_ptr; 4868 b_oth = (Mat_SeqAIJ*)(*B_oth)->data; 4869 b_otha = b_oth->a; 4870 } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE, "Matrix P does not posses an object container"); 4871 4872 /* a-array */ 4873 /*---------*/ 4874 /* post receives of a-array */ 4875 for (i=0; i<nrecvs; i++) { 4876 nrows = rstartsj[i+1]-rstartsj[i]; /* length of the msg received */ 4877 ierr = MPI_Irecv(b_otha+rstartsj[i],nrows,MPIU_SCALAR,rprocs[i],tag,comm,rwaits+i);CHKERRQ(ierr); 4878 } 4879 4880 /* pack the outgoing message a-array */ 4881 k = 0; 4882 for (i=0; i<nsends; i++) { 4883 nrows = sstarts[i+1]-sstarts[i]; /* num of block rows */ 4884 bufA = bufa+sstartsj[i]; 4885 for (j=0; j<nrows; j++) { 4886 row = srow[k++] + B->rmap->range[rank]; /* global row idx */ 4887 for (ll=0; ll<sbs; ll++) { 4888 ierr = MatGetRow_MPIAIJ(B,row+ll,&ncols,NULL,&vals);CHKERRQ(ierr); 4889 for (l=0; l<ncols; l++) { 4890 *bufA++ = vals[l]; 4891 } 4892 ierr = MatRestoreRow_MPIAIJ(B,row+ll,&ncols,NULL,&vals);CHKERRQ(ierr); 4893 } 4894 } 4895 ierr = MPI_Isend(bufa+sstartsj[i],sstartsj[i+1]-sstartsj[i],MPIU_SCALAR,sprocs[i],tag,comm,swaits+i);CHKERRQ(ierr); 4896 } 4897 /* recvs and sends of a-array are completed */ 4898 i = nrecvs; 4899 while (i--) { 4900 ierr = MPI_Waitany(nrecvs,rwaits,&jj,&rstatus);CHKERRQ(ierr); 4901 } 4902 if (nsends) {ierr = MPI_Waitall(nsends,swaits,sstatus);CHKERRQ(ierr);} 4903 ierr = PetscFree2(rwaits,swaits);CHKERRQ(ierr); 4904 4905 if (scall == MAT_INITIAL_MATRIX) { 4906 /* put together the new matrix */ 4907 ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,aBn,B->cmap->N,b_othi,b_othj,b_otha,B_oth);CHKERRQ(ierr); 4908 4909 /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */ 4910 /* Since these are PETSc arrays, change flags to free them as necessary. */ 4911 b_oth = (Mat_SeqAIJ*)(*B_oth)->data; 4912 b_oth->free_a = PETSC_TRUE; 4913 b_oth->free_ij = PETSC_TRUE; 4914 b_oth->nonew = 0; 4915 4916 ierr = PetscFree(bufj);CHKERRQ(ierr); 4917 if (!startsj_s || !bufa_ptr) { 4918 ierr = PetscFree2(sstartsj,rstartsj);CHKERRQ(ierr); 4919 ierr = PetscFree(bufa_ptr);CHKERRQ(ierr); 4920 } else { 4921 *startsj_s = sstartsj; 4922 *startsj_r = rstartsj; 4923 *bufa_ptr = bufa; 4924 } 4925 } 4926 ierr = PetscLogEventEnd(MAT_GetBrowsOfAocols,A,B,0,0);CHKERRQ(ierr); 4927 PetscFunctionReturn(0); 4928 } 4929 4930 #undef __FUNCT__ 4931 #define __FUNCT__ "MatGetCommunicationStructs" 4932 /*@C 4933 MatGetCommunicationStructs - Provides access to the communication structures used in matrix-vector multiplication. 4934 4935 Not Collective 4936 4937 Input Parameters: 4938 . A - The matrix in mpiaij format 4939 4940 Output Parameter: 4941 + lvec - The local vector holding off-process values from the argument to a matrix-vector product 4942 . colmap - A map from global column index to local index into lvec 4943 - multScatter - A scatter from the argument of a matrix-vector product to lvec 4944 4945 Level: developer 4946 4947 @*/ 4948 #if defined(PETSC_USE_CTABLE) 4949 PetscErrorCode MatGetCommunicationStructs(Mat A, Vec *lvec, PetscTable *colmap, VecScatter *multScatter) 4950 #else 4951 PetscErrorCode MatGetCommunicationStructs(Mat A, Vec *lvec, PetscInt *colmap[], VecScatter *multScatter) 4952 #endif 4953 { 4954 Mat_MPIAIJ *a; 4955 4956 PetscFunctionBegin; 4957 PetscValidHeaderSpecific(A, MAT_CLASSID, 1); 4958 PetscValidPointer(lvec, 2); 4959 PetscValidPointer(colmap, 3); 4960 PetscValidPointer(multScatter, 4); 4961 a = (Mat_MPIAIJ*) A->data; 4962 if (lvec) *lvec = a->lvec; 4963 if (colmap) *colmap = a->colmap; 4964 if (multScatter) *multScatter = a->Mvctx; 4965 PetscFunctionReturn(0); 4966 } 4967 4968 PETSC_EXTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJCRL(Mat,MatType,MatReuse,Mat*); 4969 PETSC_EXTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJPERM(Mat,MatType,MatReuse,Mat*); 4970 PETSC_EXTERN PetscErrorCode MatConvert_MPIAIJ_MPISBAIJ(Mat,MatType,MatReuse,Mat*); 4971 #if defined(PETSC_HAVE_ELEMENTAL) 4972 PETSC_EXTERN PetscErrorCode MatConvert_MPIAIJ_Elemental(Mat,MatType,MatReuse,Mat*); 4973 #endif 4974 4975 #undef __FUNCT__ 4976 #define __FUNCT__ "MatMatMultNumeric_MPIDense_MPIAIJ" 4977 /* 4978 Computes (B'*A')' since computing B*A directly is untenable 4979 4980 n p p 4981 ( ) ( ) ( ) 4982 m ( A ) * n ( B ) = m ( C ) 4983 ( ) ( ) ( ) 4984 4985 */ 4986 PetscErrorCode MatMatMultNumeric_MPIDense_MPIAIJ(Mat A,Mat B,Mat C) 4987 { 4988 PetscErrorCode ierr; 4989 Mat At,Bt,Ct; 4990 4991 PetscFunctionBegin; 4992 ierr = MatTranspose(A,MAT_INITIAL_MATRIX,&At);CHKERRQ(ierr); 4993 ierr = MatTranspose(B,MAT_INITIAL_MATRIX,&Bt);CHKERRQ(ierr); 4994 ierr = MatMatMult(Bt,At,MAT_INITIAL_MATRIX,1.0,&Ct);CHKERRQ(ierr); 4995 ierr = MatDestroy(&At);CHKERRQ(ierr); 4996 ierr = MatDestroy(&Bt);CHKERRQ(ierr); 4997 ierr = MatTranspose(Ct,MAT_REUSE_MATRIX,&C);CHKERRQ(ierr); 4998 ierr = MatDestroy(&Ct);CHKERRQ(ierr); 4999 PetscFunctionReturn(0); 5000 } 5001 5002 #undef __FUNCT__ 5003 #define __FUNCT__ "MatMatMultSymbolic_MPIDense_MPIAIJ" 5004 PetscErrorCode MatMatMultSymbolic_MPIDense_MPIAIJ(Mat A,Mat B,PetscReal fill,Mat *C) 5005 { 5006 PetscErrorCode ierr; 5007 PetscInt m=A->rmap->n,n=B->cmap->n; 5008 Mat Cmat; 5009 5010 PetscFunctionBegin; 5011 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); 5012 ierr = MatCreate(PetscObjectComm((PetscObject)A),&Cmat);CHKERRQ(ierr); 5013 ierr = MatSetSizes(Cmat,m,n,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 5014 ierr = MatSetBlockSizesFromMats(Cmat,A,B);CHKERRQ(ierr); 5015 ierr = MatSetType(Cmat,MATMPIDENSE);CHKERRQ(ierr); 5016 ierr = MatMPIDenseSetPreallocation(Cmat,NULL);CHKERRQ(ierr); 5017 ierr = MatAssemblyBegin(Cmat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5018 ierr = MatAssemblyEnd(Cmat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5019 5020 Cmat->ops->matmultnumeric = MatMatMultNumeric_MPIDense_MPIAIJ; 5021 5022 *C = Cmat; 5023 PetscFunctionReturn(0); 5024 } 5025 5026 /* ----------------------------------------------------------------*/ 5027 #undef __FUNCT__ 5028 #define __FUNCT__ "MatMatMult_MPIDense_MPIAIJ" 5029 PetscErrorCode MatMatMult_MPIDense_MPIAIJ(Mat A,Mat B,MatReuse scall,PetscReal fill,Mat *C) 5030 { 5031 PetscErrorCode ierr; 5032 5033 PetscFunctionBegin; 5034 if (scall == MAT_INITIAL_MATRIX) { 5035 ierr = PetscLogEventBegin(MAT_MatMultSymbolic,A,B,0,0);CHKERRQ(ierr); 5036 ierr = MatMatMultSymbolic_MPIDense_MPIAIJ(A,B,fill,C);CHKERRQ(ierr); 5037 ierr = PetscLogEventEnd(MAT_MatMultSymbolic,A,B,0,0);CHKERRQ(ierr); 5038 } 5039 ierr = PetscLogEventBegin(MAT_MatMultNumeric,A,B,0,0);CHKERRQ(ierr); 5040 ierr = MatMatMultNumeric_MPIDense_MPIAIJ(A,B,*C);CHKERRQ(ierr); 5041 ierr = PetscLogEventEnd(MAT_MatMultNumeric,A,B,0,0);CHKERRQ(ierr); 5042 PetscFunctionReturn(0); 5043 } 5044 5045 /*MC 5046 MATMPIAIJ - MATMPIAIJ = "mpiaij" - A matrix type to be used for parallel sparse matrices. 5047 5048 Options Database Keys: 5049 . -mat_type mpiaij - sets the matrix type to "mpiaij" during a call to MatSetFromOptions() 5050 5051 Level: beginner 5052 5053 .seealso: MatCreateAIJ() 5054 M*/ 5055 5056 #undef __FUNCT__ 5057 #define __FUNCT__ "MatCreate_MPIAIJ" 5058 PETSC_EXTERN PetscErrorCode MatCreate_MPIAIJ(Mat B) 5059 { 5060 Mat_MPIAIJ *b; 5061 PetscErrorCode ierr; 5062 PetscMPIInt size; 5063 5064 PetscFunctionBegin; 5065 ierr = MPI_Comm_size(PetscObjectComm((PetscObject)B),&size);CHKERRQ(ierr); 5066 5067 ierr = PetscNewLog(B,&b);CHKERRQ(ierr); 5068 B->data = (void*)b; 5069 ierr = PetscMemcpy(B->ops,&MatOps_Values,sizeof(struct _MatOps));CHKERRQ(ierr); 5070 B->assembled = PETSC_FALSE; 5071 B->insertmode = NOT_SET_VALUES; 5072 b->size = size; 5073 5074 ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)B),&b->rank);CHKERRQ(ierr); 5075 5076 /* build cache for off array entries formed */ 5077 ierr = MatStashCreate_Private(PetscObjectComm((PetscObject)B),1,&B->stash);CHKERRQ(ierr); 5078 5079 b->donotstash = PETSC_FALSE; 5080 b->colmap = 0; 5081 b->garray = 0; 5082 b->roworiented = PETSC_TRUE; 5083 5084 /* stuff used for matrix vector multiply */ 5085 b->lvec = NULL; 5086 b->Mvctx = NULL; 5087 5088 /* stuff for MatGetRow() */ 5089 b->rowindices = 0; 5090 b->rowvalues = 0; 5091 b->getrowactive = PETSC_FALSE; 5092 5093 /* flexible pointer used in CUSP/CUSPARSE classes */ 5094 b->spptr = NULL; 5095 5096 ierr = PetscObjectComposeFunction((PetscObject)B,"MatStoreValues_C",MatStoreValues_MPIAIJ);CHKERRQ(ierr); 5097 ierr = PetscObjectComposeFunction((PetscObject)B,"MatRetrieveValues_C",MatRetrieveValues_MPIAIJ);CHKERRQ(ierr); 5098 ierr = PetscObjectComposeFunction((PetscObject)B,"MatGetDiagonalBlock_C",MatGetDiagonalBlock_MPIAIJ);CHKERRQ(ierr); 5099 ierr = PetscObjectComposeFunction((PetscObject)B,"MatIsTranspose_C",MatIsTranspose_MPIAIJ);CHKERRQ(ierr); 5100 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMPIAIJSetPreallocation_C",MatMPIAIJSetPreallocation_MPIAIJ);CHKERRQ(ierr); 5101 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMPIAIJSetPreallocationCSR_C",MatMPIAIJSetPreallocationCSR_MPIAIJ);CHKERRQ(ierr); 5102 ierr = PetscObjectComposeFunction((PetscObject)B,"MatDiagonalScaleLocal_C",MatDiagonalScaleLocal_MPIAIJ);CHKERRQ(ierr); 5103 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_mpiaijperm_C",MatConvert_MPIAIJ_MPIAIJPERM);CHKERRQ(ierr); 5104 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_mpiaijcrl_C",MatConvert_MPIAIJ_MPIAIJCRL);CHKERRQ(ierr); 5105 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_mpisbaij_C",MatConvert_MPIAIJ_MPISBAIJ);CHKERRQ(ierr); 5106 #if defined(PETSC_HAVE_ELEMENTAL) 5107 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_elemental_C",MatConvert_MPIAIJ_Elemental);CHKERRQ(ierr); 5108 #endif 5109 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMatMult_mpidense_mpiaij_C",MatMatMult_MPIDense_MPIAIJ);CHKERRQ(ierr); 5110 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMatMultSymbolic_mpidense_mpiaij_C",MatMatMultSymbolic_MPIDense_MPIAIJ);CHKERRQ(ierr); 5111 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMatMultNumeric_mpidense_mpiaij_C",MatMatMultNumeric_MPIDense_MPIAIJ);CHKERRQ(ierr); 5112 ierr = PetscObjectChangeTypeName((PetscObject)B,MATMPIAIJ);CHKERRQ(ierr); 5113 PetscFunctionReturn(0); 5114 } 5115 5116 #undef __FUNCT__ 5117 #define __FUNCT__ "MatCreateMPIAIJWithSplitArrays" 5118 /*@C 5119 MatCreateMPIAIJWithSplitArrays - creates a MPI AIJ matrix using arrays that contain the "diagonal" 5120 and "off-diagonal" part of the matrix in CSR format. 5121 5122 Collective on MPI_Comm 5123 5124 Input Parameters: 5125 + comm - MPI communicator 5126 . m - number of local rows (Cannot be PETSC_DECIDE) 5127 . n - This value should be the same as the local size used in creating the 5128 x vector for the matrix-vector product y = Ax. (or PETSC_DECIDE to have 5129 calculated if N is given) For square matrices n is almost always m. 5130 . M - number of global rows (or PETSC_DETERMINE to have calculated if m is given) 5131 . N - number of global columns (or PETSC_DETERMINE to have calculated if n is given) 5132 . i - row indices for "diagonal" portion of matrix 5133 . j - column indices 5134 . a - matrix values 5135 . oi - row indices for "off-diagonal" portion of matrix 5136 . oj - column indices 5137 - oa - matrix values 5138 5139 Output Parameter: 5140 . mat - the matrix 5141 5142 Level: advanced 5143 5144 Notes: 5145 The i, j, and a arrays ARE NOT copied by this routine into the internal format used by PETSc. The user 5146 must free the arrays once the matrix has been destroyed and not before. 5147 5148 The i and j indices are 0 based 5149 5150 See MatCreateAIJ() for the definition of "diagonal" and "off-diagonal" portion of the matrix 5151 5152 This sets local rows and cannot be used to set off-processor values. 5153 5154 Use of this routine is discouraged because it is inflexible and cumbersome to use. It is extremely rare that a 5155 legacy application natively assembles into exactly this split format. The code to do so is nontrivial and does 5156 not easily support in-place reassembly. It is recommended to use MatSetValues() (or a variant thereof) because 5157 the resulting assembly is easier to implement, will work with any matrix format, and the user does not have to 5158 keep track of the underlying array. Use MatSetOption(A,MAT_IGNORE_OFF_PROC_ENTRIES,PETSC_TRUE) to disable all 5159 communication if it is known that only local entries will be set. 5160 5161 .keywords: matrix, aij, compressed row, sparse, parallel 5162 5163 .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatMPIAIJSetPreallocationCSR(), 5164 MPIAIJ, MatCreateAIJ(), MatCreateMPIAIJWithArrays() 5165 @*/ 5166 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) 5167 { 5168 PetscErrorCode ierr; 5169 Mat_MPIAIJ *maij; 5170 5171 PetscFunctionBegin; 5172 if (m < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"local number of rows (m) cannot be PETSC_DECIDE, or negative"); 5173 if (i[0]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"i (row indices) must start with 0"); 5174 if (oi[0]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"oi (row indices) must start with 0"); 5175 ierr = MatCreate(comm,mat);CHKERRQ(ierr); 5176 ierr = MatSetSizes(*mat,m,n,M,N);CHKERRQ(ierr); 5177 ierr = MatSetType(*mat,MATMPIAIJ);CHKERRQ(ierr); 5178 maij = (Mat_MPIAIJ*) (*mat)->data; 5179 5180 (*mat)->preallocated = PETSC_TRUE; 5181 5182 ierr = PetscLayoutSetUp((*mat)->rmap);CHKERRQ(ierr); 5183 ierr = PetscLayoutSetUp((*mat)->cmap);CHKERRQ(ierr); 5184 5185 ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,m,n,i,j,a,&maij->A);CHKERRQ(ierr); 5186 ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,m,(*mat)->cmap->N,oi,oj,oa,&maij->B);CHKERRQ(ierr); 5187 5188 ierr = MatAssemblyBegin(maij->A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5189 ierr = MatAssemblyEnd(maij->A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5190 ierr = MatAssemblyBegin(maij->B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5191 ierr = MatAssemblyEnd(maij->B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5192 5193 ierr = MatAssemblyBegin(*mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5194 ierr = MatAssemblyEnd(*mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5195 ierr = MatSetOption(*mat,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE);CHKERRQ(ierr); 5196 PetscFunctionReturn(0); 5197 } 5198 5199 /* 5200 Special version for direct calls from Fortran 5201 */ 5202 #include <petsc-private/fortranimpl.h> 5203 5204 #if defined(PETSC_HAVE_FORTRAN_CAPS) 5205 #define matsetvaluesmpiaij_ MATSETVALUESMPIAIJ 5206 #elif !defined(PETSC_HAVE_FORTRAN_UNDERSCORE) 5207 #define matsetvaluesmpiaij_ matsetvaluesmpiaij 5208 #endif 5209 5210 /* Change these macros so can be used in void function */ 5211 #undef CHKERRQ 5212 #define CHKERRQ(ierr) CHKERRABORT(PETSC_COMM_WORLD,ierr) 5213 #undef SETERRQ2 5214 #define SETERRQ2(comm,ierr,b,c,d) CHKERRABORT(comm,ierr) 5215 #undef SETERRQ3 5216 #define SETERRQ3(comm,ierr,b,c,d,e) CHKERRABORT(comm,ierr) 5217 #undef SETERRQ 5218 #define SETERRQ(c,ierr,b) CHKERRABORT(c,ierr) 5219 5220 #undef __FUNCT__ 5221 #define __FUNCT__ "matsetvaluesmpiaij_" 5222 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) 5223 { 5224 Mat mat = *mmat; 5225 PetscInt m = *mm, n = *mn; 5226 InsertMode addv = *maddv; 5227 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 5228 PetscScalar value; 5229 PetscErrorCode ierr; 5230 5231 MatCheckPreallocated(mat,1); 5232 if (mat->insertmode == NOT_SET_VALUES) mat->insertmode = addv; 5233 5234 #if defined(PETSC_USE_DEBUG) 5235 else if (mat->insertmode != addv) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Cannot mix add values and insert values"); 5236 #endif 5237 { 5238 PetscInt i,j,rstart = mat->rmap->rstart,rend = mat->rmap->rend; 5239 PetscInt cstart = mat->cmap->rstart,cend = mat->cmap->rend,row,col; 5240 PetscBool roworiented = aij->roworiented; 5241 5242 /* Some Variables required in the macro */ 5243 Mat A = aij->A; 5244 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 5245 PetscInt *aimax = a->imax,*ai = a->i,*ailen = a->ilen,*aj = a->j; 5246 MatScalar *aa = a->a; 5247 PetscBool ignorezeroentries = (((a->ignorezeroentries)&&(addv==ADD_VALUES)) ? PETSC_TRUE : PETSC_FALSE); 5248 Mat B = aij->B; 5249 Mat_SeqAIJ *b = (Mat_SeqAIJ*)B->data; 5250 PetscInt *bimax = b->imax,*bi = b->i,*bilen = b->ilen,*bj = b->j,bm = aij->B->rmap->n,am = aij->A->rmap->n; 5251 MatScalar *ba = b->a; 5252 5253 PetscInt *rp1,*rp2,ii,nrow1,nrow2,_i,rmax1,rmax2,N,low1,high1,low2,high2,t,lastcol1,lastcol2; 5254 PetscInt nonew = a->nonew; 5255 MatScalar *ap1,*ap2; 5256 5257 PetscFunctionBegin; 5258 for (i=0; i<m; i++) { 5259 if (im[i] < 0) continue; 5260 #if defined(PETSC_USE_DEBUG) 5261 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); 5262 #endif 5263 if (im[i] >= rstart && im[i] < rend) { 5264 row = im[i] - rstart; 5265 lastcol1 = -1; 5266 rp1 = aj + ai[row]; 5267 ap1 = aa + ai[row]; 5268 rmax1 = aimax[row]; 5269 nrow1 = ailen[row]; 5270 low1 = 0; 5271 high1 = nrow1; 5272 lastcol2 = -1; 5273 rp2 = bj + bi[row]; 5274 ap2 = ba + bi[row]; 5275 rmax2 = bimax[row]; 5276 nrow2 = bilen[row]; 5277 low2 = 0; 5278 high2 = nrow2; 5279 5280 for (j=0; j<n; j++) { 5281 if (roworiented) value = v[i*n+j]; 5282 else value = v[i+j*m]; 5283 if (ignorezeroentries && value == 0.0 && (addv == ADD_VALUES)) continue; 5284 if (in[j] >= cstart && in[j] < cend) { 5285 col = in[j] - cstart; 5286 MatSetValues_SeqAIJ_A_Private(row,col,value,addv,im[i],in[j]); 5287 } else if (in[j] < 0) continue; 5288 #if defined(PETSC_USE_DEBUG) 5289 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); 5290 #endif 5291 else { 5292 if (mat->was_assembled) { 5293 if (!aij->colmap) { 5294 ierr = MatCreateColmap_MPIAIJ_Private(mat);CHKERRQ(ierr); 5295 } 5296 #if defined(PETSC_USE_CTABLE) 5297 ierr = PetscTableFind(aij->colmap,in[j]+1,&col);CHKERRQ(ierr); 5298 col--; 5299 #else 5300 col = aij->colmap[in[j]] - 1; 5301 #endif 5302 if (col < 0 && !((Mat_SeqAIJ*)(aij->A->data))->nonew) { 5303 ierr = MatDisAssemble_MPIAIJ(mat);CHKERRQ(ierr); 5304 col = in[j]; 5305 /* Reinitialize the variables required by MatSetValues_SeqAIJ_B_Private() */ 5306 B = aij->B; 5307 b = (Mat_SeqAIJ*)B->data; 5308 bimax = b->imax; bi = b->i; bilen = b->ilen; bj = b->j; 5309 rp2 = bj + bi[row]; 5310 ap2 = ba + bi[row]; 5311 rmax2 = bimax[row]; 5312 nrow2 = bilen[row]; 5313 low2 = 0; 5314 high2 = nrow2; 5315 bm = aij->B->rmap->n; 5316 ba = b->a; 5317 } 5318 } else col = in[j]; 5319 MatSetValues_SeqAIJ_B_Private(row,col,value,addv,im[i],in[j]); 5320 } 5321 } 5322 } else if (!aij->donotstash) { 5323 if (roworiented) { 5324 ierr = MatStashValuesRow_Private(&mat->stash,im[i],n,in,v+i*n,(PetscBool)(ignorezeroentries && (addv == ADD_VALUES)));CHKERRQ(ierr); 5325 } else { 5326 ierr = MatStashValuesCol_Private(&mat->stash,im[i],n,in,v+i,m,(PetscBool)(ignorezeroentries && (addv == ADD_VALUES)));CHKERRQ(ierr); 5327 } 5328 } 5329 } 5330 } 5331 PetscFunctionReturnVoid(); 5332 } 5333 5334