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,MPI_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 = PetscViewerASCIIPushSynchronized(viewer);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 = PetscViewerASCIIPopSynchronized(viewer);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 = PetscViewerGetSubViewer(viewer,PETSC_COMM_SELF,&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 = PetscViewerRestoreSubViewer(viewer,PETSC_COMM_SELF,&sviewer);CHKERRQ(ierr); 1452 ierr = PetscViewerFlush(viewer);CHKERRQ(ierr); 1453 ierr = MatDestroy(&A);CHKERRQ(ierr); 1454 } 1455 PetscFunctionReturn(0); 1456 } 1457 1458 #undef __FUNCT__ 1459 #define __FUNCT__ "MatView_MPIAIJ" 1460 PetscErrorCode MatView_MPIAIJ(Mat mat,PetscViewer viewer) 1461 { 1462 PetscErrorCode ierr; 1463 PetscBool iascii,isdraw,issocket,isbinary; 1464 1465 PetscFunctionBegin; 1466 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii);CHKERRQ(ierr); 1467 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERDRAW,&isdraw);CHKERRQ(ierr); 1468 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERBINARY,&isbinary);CHKERRQ(ierr); 1469 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERSOCKET,&issocket);CHKERRQ(ierr); 1470 if (iascii || isdraw || isbinary || issocket) { 1471 ierr = MatView_MPIAIJ_ASCIIorDraworSocket(mat,viewer);CHKERRQ(ierr); 1472 } 1473 PetscFunctionReturn(0); 1474 } 1475 1476 #undef __FUNCT__ 1477 #define __FUNCT__ "MatSOR_MPIAIJ" 1478 PetscErrorCode MatSOR_MPIAIJ(Mat matin,Vec bb,PetscReal omega,MatSORType flag,PetscReal fshift,PetscInt its,PetscInt lits,Vec xx) 1479 { 1480 Mat_MPIAIJ *mat = (Mat_MPIAIJ*)matin->data; 1481 PetscErrorCode ierr; 1482 Vec bb1 = 0; 1483 PetscBool hasop; 1484 1485 PetscFunctionBegin; 1486 if (flag == SOR_APPLY_UPPER) { 1487 ierr = (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);CHKERRQ(ierr); 1488 PetscFunctionReturn(0); 1489 } 1490 1491 if (its > 1 || ~flag & SOR_ZERO_INITIAL_GUESS || flag & SOR_EISENSTAT) { 1492 ierr = VecDuplicate(bb,&bb1);CHKERRQ(ierr); 1493 } 1494 1495 if ((flag & SOR_LOCAL_SYMMETRIC_SWEEP) == SOR_LOCAL_SYMMETRIC_SWEEP) { 1496 if (flag & SOR_ZERO_INITIAL_GUESS) { 1497 ierr = (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);CHKERRQ(ierr); 1498 its--; 1499 } 1500 1501 while (its--) { 1502 ierr = VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1503 ierr = VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1504 1505 /* update rhs: bb1 = bb - B*x */ 1506 ierr = VecScale(mat->lvec,-1.0);CHKERRQ(ierr); 1507 ierr = (*mat->B->ops->multadd)(mat->B,mat->lvec,bb,bb1);CHKERRQ(ierr); 1508 1509 /* local sweep */ 1510 ierr = (*mat->A->ops->sor)(mat->A,bb1,omega,SOR_SYMMETRIC_SWEEP,fshift,lits,1,xx);CHKERRQ(ierr); 1511 } 1512 } else if (flag & SOR_LOCAL_FORWARD_SWEEP) { 1513 if (flag & SOR_ZERO_INITIAL_GUESS) { 1514 ierr = (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);CHKERRQ(ierr); 1515 its--; 1516 } 1517 while (its--) { 1518 ierr = VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1519 ierr = VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1520 1521 /* update rhs: bb1 = bb - B*x */ 1522 ierr = VecScale(mat->lvec,-1.0);CHKERRQ(ierr); 1523 ierr = (*mat->B->ops->multadd)(mat->B,mat->lvec,bb,bb1);CHKERRQ(ierr); 1524 1525 /* local sweep */ 1526 ierr = (*mat->A->ops->sor)(mat->A,bb1,omega,SOR_FORWARD_SWEEP,fshift,lits,1,xx);CHKERRQ(ierr); 1527 } 1528 } else if (flag & SOR_LOCAL_BACKWARD_SWEEP) { 1529 if (flag & SOR_ZERO_INITIAL_GUESS) { 1530 ierr = (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);CHKERRQ(ierr); 1531 its--; 1532 } 1533 while (its--) { 1534 ierr = VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1535 ierr = VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1536 1537 /* update rhs: bb1 = bb - B*x */ 1538 ierr = VecScale(mat->lvec,-1.0);CHKERRQ(ierr); 1539 ierr = (*mat->B->ops->multadd)(mat->B,mat->lvec,bb,bb1);CHKERRQ(ierr); 1540 1541 /* local sweep */ 1542 ierr = (*mat->A->ops->sor)(mat->A,bb1,omega,SOR_BACKWARD_SWEEP,fshift,lits,1,xx);CHKERRQ(ierr); 1543 } 1544 } else if (flag & SOR_EISENSTAT) { 1545 Vec xx1; 1546 1547 ierr = VecDuplicate(bb,&xx1);CHKERRQ(ierr); 1548 ierr = (*mat->A->ops->sor)(mat->A,bb,omega,(MatSORType)(SOR_ZERO_INITIAL_GUESS | SOR_LOCAL_BACKWARD_SWEEP),fshift,lits,1,xx);CHKERRQ(ierr); 1549 1550 ierr = VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1551 ierr = VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1552 if (!mat->diag) { 1553 ierr = MatCreateVecs(matin,&mat->diag,NULL);CHKERRQ(ierr); 1554 ierr = MatGetDiagonal(matin,mat->diag);CHKERRQ(ierr); 1555 } 1556 ierr = MatHasOperation(matin,MATOP_MULT_DIAGONAL_BLOCK,&hasop);CHKERRQ(ierr); 1557 if (hasop) { 1558 ierr = MatMultDiagonalBlock(matin,xx,bb1);CHKERRQ(ierr); 1559 } else { 1560 ierr = VecPointwiseMult(bb1,mat->diag,xx);CHKERRQ(ierr); 1561 } 1562 ierr = VecAYPX(bb1,(omega-2.0)/omega,bb);CHKERRQ(ierr); 1563 1564 ierr = MatMultAdd(mat->B,mat->lvec,bb1,bb1);CHKERRQ(ierr); 1565 1566 /* local sweep */ 1567 ierr = (*mat->A->ops->sor)(mat->A,bb1,omega,(MatSORType)(SOR_ZERO_INITIAL_GUESS | SOR_LOCAL_FORWARD_SWEEP),fshift,lits,1,xx1);CHKERRQ(ierr); 1568 ierr = VecAXPY(xx,1.0,xx1);CHKERRQ(ierr); 1569 ierr = VecDestroy(&xx1);CHKERRQ(ierr); 1570 } else SETERRQ(PetscObjectComm((PetscObject)matin),PETSC_ERR_SUP,"Parallel SOR not supported"); 1571 1572 ierr = VecDestroy(&bb1);CHKERRQ(ierr); 1573 PetscFunctionReturn(0); 1574 } 1575 1576 #undef __FUNCT__ 1577 #define __FUNCT__ "MatPermute_MPIAIJ" 1578 PetscErrorCode MatPermute_MPIAIJ(Mat A,IS rowp,IS colp,Mat *B) 1579 { 1580 Mat aA,aB,Aperm; 1581 const PetscInt *rwant,*cwant,*gcols,*ai,*bi,*aj,*bj; 1582 PetscScalar *aa,*ba; 1583 PetscInt i,j,m,n,ng,anz,bnz,*dnnz,*onnz,*tdnnz,*tonnz,*rdest,*cdest,*work,*gcdest; 1584 PetscSF rowsf,sf; 1585 IS parcolp = NULL; 1586 PetscBool done; 1587 PetscErrorCode ierr; 1588 1589 PetscFunctionBegin; 1590 ierr = MatGetLocalSize(A,&m,&n);CHKERRQ(ierr); 1591 ierr = ISGetIndices(rowp,&rwant);CHKERRQ(ierr); 1592 ierr = ISGetIndices(colp,&cwant);CHKERRQ(ierr); 1593 ierr = PetscMalloc3(PetscMax(m,n),&work,m,&rdest,n,&cdest);CHKERRQ(ierr); 1594 1595 /* Invert row permutation to find out where my rows should go */ 1596 ierr = PetscSFCreate(PetscObjectComm((PetscObject)A),&rowsf);CHKERRQ(ierr); 1597 ierr = PetscSFSetGraphLayout(rowsf,A->rmap,A->rmap->n,NULL,PETSC_OWN_POINTER,rwant);CHKERRQ(ierr); 1598 ierr = PetscSFSetFromOptions(rowsf);CHKERRQ(ierr); 1599 for (i=0; i<m; i++) work[i] = A->rmap->rstart + i; 1600 ierr = PetscSFReduceBegin(rowsf,MPIU_INT,work,rdest,MPIU_REPLACE);CHKERRQ(ierr); 1601 ierr = PetscSFReduceEnd(rowsf,MPIU_INT,work,rdest,MPIU_REPLACE);CHKERRQ(ierr); 1602 1603 /* Invert column permutation to find out where my columns should go */ 1604 ierr = PetscSFCreate(PetscObjectComm((PetscObject)A),&sf);CHKERRQ(ierr); 1605 ierr = PetscSFSetGraphLayout(sf,A->cmap,A->cmap->n,NULL,PETSC_OWN_POINTER,cwant);CHKERRQ(ierr); 1606 ierr = PetscSFSetFromOptions(sf);CHKERRQ(ierr); 1607 for (i=0; i<n; i++) work[i] = A->cmap->rstart + i; 1608 ierr = PetscSFReduceBegin(sf,MPIU_INT,work,cdest,MPIU_REPLACE);CHKERRQ(ierr); 1609 ierr = PetscSFReduceEnd(sf,MPIU_INT,work,cdest,MPIU_REPLACE);CHKERRQ(ierr); 1610 ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 1611 1612 ierr = ISRestoreIndices(rowp,&rwant);CHKERRQ(ierr); 1613 ierr = ISRestoreIndices(colp,&cwant);CHKERRQ(ierr); 1614 ierr = MatMPIAIJGetSeqAIJ(A,&aA,&aB,&gcols);CHKERRQ(ierr); 1615 1616 /* Find out where my gcols should go */ 1617 ierr = MatGetSize(aB,NULL,&ng);CHKERRQ(ierr); 1618 ierr = PetscMalloc1(ng,&gcdest);CHKERRQ(ierr); 1619 ierr = PetscSFCreate(PetscObjectComm((PetscObject)A),&sf);CHKERRQ(ierr); 1620 ierr = PetscSFSetGraphLayout(sf,A->cmap,ng,NULL,PETSC_OWN_POINTER,gcols);CHKERRQ(ierr); 1621 ierr = PetscSFSetFromOptions(sf);CHKERRQ(ierr); 1622 ierr = PetscSFBcastBegin(sf,MPIU_INT,cdest,gcdest);CHKERRQ(ierr); 1623 ierr = PetscSFBcastEnd(sf,MPIU_INT,cdest,gcdest);CHKERRQ(ierr); 1624 ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 1625 1626 ierr = PetscCalloc4(m,&dnnz,m,&onnz,m,&tdnnz,m,&tonnz);CHKERRQ(ierr); 1627 ierr = MatGetRowIJ(aA,0,PETSC_FALSE,PETSC_FALSE,&anz,&ai,&aj,&done);CHKERRQ(ierr); 1628 ierr = MatGetRowIJ(aB,0,PETSC_FALSE,PETSC_FALSE,&bnz,&bi,&bj,&done);CHKERRQ(ierr); 1629 for (i=0; i<m; i++) { 1630 PetscInt row = rdest[i],rowner; 1631 ierr = PetscLayoutFindOwner(A->rmap,row,&rowner);CHKERRQ(ierr); 1632 for (j=ai[i]; j<ai[i+1]; j++) { 1633 PetscInt cowner,col = cdest[aj[j]]; 1634 ierr = PetscLayoutFindOwner(A->cmap,col,&cowner);CHKERRQ(ierr); /* Could build an index for the columns to eliminate this search */ 1635 if (rowner == cowner) dnnz[i]++; 1636 else onnz[i]++; 1637 } 1638 for (j=bi[i]; j<bi[i+1]; j++) { 1639 PetscInt cowner,col = gcdest[bj[j]]; 1640 ierr = PetscLayoutFindOwner(A->cmap,col,&cowner);CHKERRQ(ierr); 1641 if (rowner == cowner) dnnz[i]++; 1642 else onnz[i]++; 1643 } 1644 } 1645 ierr = PetscSFBcastBegin(rowsf,MPIU_INT,dnnz,tdnnz);CHKERRQ(ierr); 1646 ierr = PetscSFBcastEnd(rowsf,MPIU_INT,dnnz,tdnnz);CHKERRQ(ierr); 1647 ierr = PetscSFBcastBegin(rowsf,MPIU_INT,onnz,tonnz);CHKERRQ(ierr); 1648 ierr = PetscSFBcastEnd(rowsf,MPIU_INT,onnz,tonnz);CHKERRQ(ierr); 1649 ierr = PetscSFDestroy(&rowsf);CHKERRQ(ierr); 1650 1651 ierr = MatCreateAIJ(PetscObjectComm((PetscObject)A),A->rmap->n,A->cmap->n,A->rmap->N,A->cmap->N,0,tdnnz,0,tonnz,&Aperm);CHKERRQ(ierr); 1652 ierr = MatSeqAIJGetArray(aA,&aa);CHKERRQ(ierr); 1653 ierr = MatSeqAIJGetArray(aB,&ba);CHKERRQ(ierr); 1654 for (i=0; i<m; i++) { 1655 PetscInt *acols = dnnz,*bcols = onnz; /* Repurpose now-unneeded arrays */ 1656 PetscInt j0,rowlen; 1657 rowlen = ai[i+1] - ai[i]; 1658 for (j0=j=0; j<rowlen; j0=j) { /* rowlen could be larger than number of rows m, so sum in batches */ 1659 for ( ; j<PetscMin(rowlen,j0+m); j++) acols[j-j0] = cdest[aj[ai[i]+j]]; 1660 ierr = MatSetValues(Aperm,1,&rdest[i],j-j0,acols,aa+ai[i]+j0,INSERT_VALUES);CHKERRQ(ierr); 1661 } 1662 rowlen = bi[i+1] - bi[i]; 1663 for (j0=j=0; j<rowlen; j0=j) { 1664 for ( ; j<PetscMin(rowlen,j0+m); j++) bcols[j-j0] = gcdest[bj[bi[i]+j]]; 1665 ierr = MatSetValues(Aperm,1,&rdest[i],j-j0,bcols,ba+bi[i]+j0,INSERT_VALUES);CHKERRQ(ierr); 1666 } 1667 } 1668 ierr = MatAssemblyBegin(Aperm,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1669 ierr = MatAssemblyEnd(Aperm,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1670 ierr = MatRestoreRowIJ(aA,0,PETSC_FALSE,PETSC_FALSE,&anz,&ai,&aj,&done);CHKERRQ(ierr); 1671 ierr = MatRestoreRowIJ(aB,0,PETSC_FALSE,PETSC_FALSE,&bnz,&bi,&bj,&done);CHKERRQ(ierr); 1672 ierr = MatSeqAIJRestoreArray(aA,&aa);CHKERRQ(ierr); 1673 ierr = MatSeqAIJRestoreArray(aB,&ba);CHKERRQ(ierr); 1674 ierr = PetscFree4(dnnz,onnz,tdnnz,tonnz);CHKERRQ(ierr); 1675 ierr = PetscFree3(work,rdest,cdest);CHKERRQ(ierr); 1676 ierr = PetscFree(gcdest);CHKERRQ(ierr); 1677 if (parcolp) {ierr = ISDestroy(&colp);CHKERRQ(ierr);} 1678 *B = Aperm; 1679 PetscFunctionReturn(0); 1680 } 1681 1682 #undef __FUNCT__ 1683 #define __FUNCT__ "MatGetGhosts_MPIAIJ" 1684 PetscErrorCode MatGetGhosts_MPIAIJ(Mat mat,PetscInt *nghosts,const PetscInt *ghosts[]) 1685 { 1686 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 1687 PetscErrorCode ierr; 1688 1689 PetscFunctionBegin; 1690 ierr = MatGetSize(aij->B,NULL,nghosts);CHKERRQ(ierr); 1691 if (ghosts) *ghosts = aij->garray; 1692 PetscFunctionReturn(0); 1693 } 1694 1695 #undef __FUNCT__ 1696 #define __FUNCT__ "MatGetInfo_MPIAIJ" 1697 PetscErrorCode MatGetInfo_MPIAIJ(Mat matin,MatInfoType flag,MatInfo *info) 1698 { 1699 Mat_MPIAIJ *mat = (Mat_MPIAIJ*)matin->data; 1700 Mat A = mat->A,B = mat->B; 1701 PetscErrorCode ierr; 1702 PetscReal isend[5],irecv[5]; 1703 1704 PetscFunctionBegin; 1705 info->block_size = 1.0; 1706 ierr = MatGetInfo(A,MAT_LOCAL,info);CHKERRQ(ierr); 1707 1708 isend[0] = info->nz_used; isend[1] = info->nz_allocated; isend[2] = info->nz_unneeded; 1709 isend[3] = info->memory; isend[4] = info->mallocs; 1710 1711 ierr = MatGetInfo(B,MAT_LOCAL,info);CHKERRQ(ierr); 1712 1713 isend[0] += info->nz_used; isend[1] += info->nz_allocated; isend[2] += info->nz_unneeded; 1714 isend[3] += info->memory; isend[4] += info->mallocs; 1715 if (flag == MAT_LOCAL) { 1716 info->nz_used = isend[0]; 1717 info->nz_allocated = isend[1]; 1718 info->nz_unneeded = isend[2]; 1719 info->memory = isend[3]; 1720 info->mallocs = isend[4]; 1721 } else if (flag == MAT_GLOBAL_MAX) { 1722 ierr = MPI_Allreduce(isend,irecv,5,MPIU_REAL,MPIU_MAX,PetscObjectComm((PetscObject)matin));CHKERRQ(ierr); 1723 1724 info->nz_used = irecv[0]; 1725 info->nz_allocated = irecv[1]; 1726 info->nz_unneeded = irecv[2]; 1727 info->memory = irecv[3]; 1728 info->mallocs = irecv[4]; 1729 } else if (flag == MAT_GLOBAL_SUM) { 1730 ierr = MPI_Allreduce(isend,irecv,5,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)matin));CHKERRQ(ierr); 1731 1732 info->nz_used = irecv[0]; 1733 info->nz_allocated = irecv[1]; 1734 info->nz_unneeded = irecv[2]; 1735 info->memory = irecv[3]; 1736 info->mallocs = irecv[4]; 1737 } 1738 info->fill_ratio_given = 0; /* no parallel LU/ILU/Cholesky */ 1739 info->fill_ratio_needed = 0; 1740 info->factor_mallocs = 0; 1741 PetscFunctionReturn(0); 1742 } 1743 1744 #undef __FUNCT__ 1745 #define __FUNCT__ "MatSetOption_MPIAIJ" 1746 PetscErrorCode MatSetOption_MPIAIJ(Mat A,MatOption op,PetscBool flg) 1747 { 1748 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 1749 PetscErrorCode ierr; 1750 1751 PetscFunctionBegin; 1752 switch (op) { 1753 case MAT_NEW_NONZERO_LOCATIONS: 1754 case MAT_NEW_NONZERO_ALLOCATION_ERR: 1755 case MAT_UNUSED_NONZERO_LOCATION_ERR: 1756 case MAT_KEEP_NONZERO_PATTERN: 1757 case MAT_NEW_NONZERO_LOCATION_ERR: 1758 case MAT_USE_INODES: 1759 case MAT_IGNORE_ZERO_ENTRIES: 1760 MatCheckPreallocated(A,1); 1761 ierr = MatSetOption(a->A,op,flg);CHKERRQ(ierr); 1762 ierr = MatSetOption(a->B,op,flg);CHKERRQ(ierr); 1763 break; 1764 case MAT_ROW_ORIENTED: 1765 a->roworiented = flg; 1766 1767 ierr = MatSetOption(a->A,op,flg);CHKERRQ(ierr); 1768 ierr = MatSetOption(a->B,op,flg);CHKERRQ(ierr); 1769 break; 1770 case MAT_NEW_DIAGONALS: 1771 ierr = PetscInfo1(A,"Option %s ignored\n",MatOptions[op]);CHKERRQ(ierr); 1772 break; 1773 case MAT_IGNORE_OFF_PROC_ENTRIES: 1774 a->donotstash = flg; 1775 break; 1776 case MAT_SPD: 1777 A->spd_set = PETSC_TRUE; 1778 A->spd = flg; 1779 if (flg) { 1780 A->symmetric = PETSC_TRUE; 1781 A->structurally_symmetric = PETSC_TRUE; 1782 A->symmetric_set = PETSC_TRUE; 1783 A->structurally_symmetric_set = PETSC_TRUE; 1784 } 1785 break; 1786 case MAT_SYMMETRIC: 1787 ierr = MatSetOption(a->A,op,flg);CHKERRQ(ierr); 1788 break; 1789 case MAT_STRUCTURALLY_SYMMETRIC: 1790 ierr = MatSetOption(a->A,op,flg);CHKERRQ(ierr); 1791 break; 1792 case MAT_HERMITIAN: 1793 ierr = MatSetOption(a->A,op,flg);CHKERRQ(ierr); 1794 break; 1795 case MAT_SYMMETRY_ETERNAL: 1796 ierr = MatSetOption(a->A,op,flg);CHKERRQ(ierr); 1797 break; 1798 default: 1799 SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"unknown option %d",op); 1800 } 1801 PetscFunctionReturn(0); 1802 } 1803 1804 #undef __FUNCT__ 1805 #define __FUNCT__ "MatGetRow_MPIAIJ" 1806 PetscErrorCode MatGetRow_MPIAIJ(Mat matin,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v) 1807 { 1808 Mat_MPIAIJ *mat = (Mat_MPIAIJ*)matin->data; 1809 PetscScalar *vworkA,*vworkB,**pvA,**pvB,*v_p; 1810 PetscErrorCode ierr; 1811 PetscInt i,*cworkA,*cworkB,**pcA,**pcB,cstart = matin->cmap->rstart; 1812 PetscInt nztot,nzA,nzB,lrow,rstart = matin->rmap->rstart,rend = matin->rmap->rend; 1813 PetscInt *cmap,*idx_p; 1814 1815 PetscFunctionBegin; 1816 if (mat->getrowactive) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Already active"); 1817 mat->getrowactive = PETSC_TRUE; 1818 1819 if (!mat->rowvalues && (idx || v)) { 1820 /* 1821 allocate enough space to hold information from the longest row. 1822 */ 1823 Mat_SeqAIJ *Aa = (Mat_SeqAIJ*)mat->A->data,*Ba = (Mat_SeqAIJ*)mat->B->data; 1824 PetscInt max = 1,tmp; 1825 for (i=0; i<matin->rmap->n; i++) { 1826 tmp = Aa->i[i+1] - Aa->i[i] + Ba->i[i+1] - Ba->i[i]; 1827 if (max < tmp) max = tmp; 1828 } 1829 ierr = PetscMalloc2(max,&mat->rowvalues,max,&mat->rowindices);CHKERRQ(ierr); 1830 } 1831 1832 if (row < rstart || row >= rend) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Only local rows"); 1833 lrow = row - rstart; 1834 1835 pvA = &vworkA; pcA = &cworkA; pvB = &vworkB; pcB = &cworkB; 1836 if (!v) {pvA = 0; pvB = 0;} 1837 if (!idx) {pcA = 0; if (!v) pcB = 0;} 1838 ierr = (*mat->A->ops->getrow)(mat->A,lrow,&nzA,pcA,pvA);CHKERRQ(ierr); 1839 ierr = (*mat->B->ops->getrow)(mat->B,lrow,&nzB,pcB,pvB);CHKERRQ(ierr); 1840 nztot = nzA + nzB; 1841 1842 cmap = mat->garray; 1843 if (v || idx) { 1844 if (nztot) { 1845 /* Sort by increasing column numbers, assuming A and B already sorted */ 1846 PetscInt imark = -1; 1847 if (v) { 1848 *v = v_p = mat->rowvalues; 1849 for (i=0; i<nzB; i++) { 1850 if (cmap[cworkB[i]] < cstart) v_p[i] = vworkB[i]; 1851 else break; 1852 } 1853 imark = i; 1854 for (i=0; i<nzA; i++) v_p[imark+i] = vworkA[i]; 1855 for (i=imark; i<nzB; i++) v_p[nzA+i] = vworkB[i]; 1856 } 1857 if (idx) { 1858 *idx = idx_p = mat->rowindices; 1859 if (imark > -1) { 1860 for (i=0; i<imark; i++) { 1861 idx_p[i] = cmap[cworkB[i]]; 1862 } 1863 } else { 1864 for (i=0; i<nzB; i++) { 1865 if (cmap[cworkB[i]] < cstart) idx_p[i] = cmap[cworkB[i]]; 1866 else break; 1867 } 1868 imark = i; 1869 } 1870 for (i=0; i<nzA; i++) idx_p[imark+i] = cstart + cworkA[i]; 1871 for (i=imark; i<nzB; i++) idx_p[nzA+i] = cmap[cworkB[i]]; 1872 } 1873 } else { 1874 if (idx) *idx = 0; 1875 if (v) *v = 0; 1876 } 1877 } 1878 *nz = nztot; 1879 ierr = (*mat->A->ops->restorerow)(mat->A,lrow,&nzA,pcA,pvA);CHKERRQ(ierr); 1880 ierr = (*mat->B->ops->restorerow)(mat->B,lrow,&nzB,pcB,pvB);CHKERRQ(ierr); 1881 PetscFunctionReturn(0); 1882 } 1883 1884 #undef __FUNCT__ 1885 #define __FUNCT__ "MatRestoreRow_MPIAIJ" 1886 PetscErrorCode MatRestoreRow_MPIAIJ(Mat mat,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v) 1887 { 1888 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 1889 1890 PetscFunctionBegin; 1891 if (!aij->getrowactive) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"MatGetRow() must be called first"); 1892 aij->getrowactive = PETSC_FALSE; 1893 PetscFunctionReturn(0); 1894 } 1895 1896 #undef __FUNCT__ 1897 #define __FUNCT__ "MatNorm_MPIAIJ" 1898 PetscErrorCode MatNorm_MPIAIJ(Mat mat,NormType type,PetscReal *norm) 1899 { 1900 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 1901 Mat_SeqAIJ *amat = (Mat_SeqAIJ*)aij->A->data,*bmat = (Mat_SeqAIJ*)aij->B->data; 1902 PetscErrorCode ierr; 1903 PetscInt i,j,cstart = mat->cmap->rstart; 1904 PetscReal sum = 0.0; 1905 MatScalar *v; 1906 1907 PetscFunctionBegin; 1908 if (aij->size == 1) { 1909 ierr = MatNorm(aij->A,type,norm);CHKERRQ(ierr); 1910 } else { 1911 if (type == NORM_FROBENIUS) { 1912 v = amat->a; 1913 for (i=0; i<amat->nz; i++) { 1914 sum += PetscRealPart(PetscConj(*v)*(*v)); v++; 1915 } 1916 v = bmat->a; 1917 for (i=0; i<bmat->nz; i++) { 1918 sum += PetscRealPart(PetscConj(*v)*(*v)); v++; 1919 } 1920 ierr = MPI_Allreduce(&sum,norm,1,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1921 *norm = PetscSqrtReal(*norm); 1922 } else if (type == NORM_1) { /* max column norm */ 1923 PetscReal *tmp,*tmp2; 1924 PetscInt *jj,*garray = aij->garray; 1925 ierr = PetscCalloc1(mat->cmap->N+1,&tmp);CHKERRQ(ierr); 1926 ierr = PetscMalloc1(mat->cmap->N+1,&tmp2);CHKERRQ(ierr); 1927 *norm = 0.0; 1928 v = amat->a; jj = amat->j; 1929 for (j=0; j<amat->nz; j++) { 1930 tmp[cstart + *jj++] += PetscAbsScalar(*v); v++; 1931 } 1932 v = bmat->a; jj = bmat->j; 1933 for (j=0; j<bmat->nz; j++) { 1934 tmp[garray[*jj++]] += PetscAbsScalar(*v); v++; 1935 } 1936 ierr = MPI_Allreduce(tmp,tmp2,mat->cmap->N,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1937 for (j=0; j<mat->cmap->N; j++) { 1938 if (tmp2[j] > *norm) *norm = tmp2[j]; 1939 } 1940 ierr = PetscFree(tmp);CHKERRQ(ierr); 1941 ierr = PetscFree(tmp2);CHKERRQ(ierr); 1942 } else if (type == NORM_INFINITY) { /* max row norm */ 1943 PetscReal ntemp = 0.0; 1944 for (j=0; j<aij->A->rmap->n; j++) { 1945 v = amat->a + amat->i[j]; 1946 sum = 0.0; 1947 for (i=0; i<amat->i[j+1]-amat->i[j]; i++) { 1948 sum += PetscAbsScalar(*v); v++; 1949 } 1950 v = bmat->a + bmat->i[j]; 1951 for (i=0; i<bmat->i[j+1]-bmat->i[j]; i++) { 1952 sum += PetscAbsScalar(*v); v++; 1953 } 1954 if (sum > ntemp) ntemp = sum; 1955 } 1956 ierr = MPI_Allreduce(&ntemp,norm,1,MPIU_REAL,MPIU_MAX,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1957 } else SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"No support for two norm"); 1958 } 1959 PetscFunctionReturn(0); 1960 } 1961 1962 #undef __FUNCT__ 1963 #define __FUNCT__ "MatTranspose_MPIAIJ" 1964 PetscErrorCode MatTranspose_MPIAIJ(Mat A,MatReuse reuse,Mat *matout) 1965 { 1966 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 1967 Mat_SeqAIJ *Aloc=(Mat_SeqAIJ*)a->A->data,*Bloc=(Mat_SeqAIJ*)a->B->data; 1968 PetscErrorCode ierr; 1969 PetscInt M = A->rmap->N,N = A->cmap->N,ma,na,mb,nb,*ai,*aj,*bi,*bj,row,*cols,*cols_tmp,i; 1970 PetscInt cstart = A->cmap->rstart,ncol; 1971 Mat B; 1972 MatScalar *array; 1973 1974 PetscFunctionBegin; 1975 if (reuse == MAT_REUSE_MATRIX && A == *matout && M != N) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_SIZ,"Square matrix only for in-place"); 1976 1977 ma = A->rmap->n; na = A->cmap->n; mb = a->B->rmap->n; nb = a->B->cmap->n; 1978 ai = Aloc->i; aj = Aloc->j; 1979 bi = Bloc->i; bj = Bloc->j; 1980 if (reuse == MAT_INITIAL_MATRIX || *matout == A) { 1981 PetscInt *d_nnz,*g_nnz,*o_nnz; 1982 PetscSFNode *oloc; 1983 PETSC_UNUSED PetscSF sf; 1984 1985 ierr = PetscMalloc4(na,&d_nnz,na,&o_nnz,nb,&g_nnz,nb,&oloc);CHKERRQ(ierr); 1986 /* compute d_nnz for preallocation */ 1987 ierr = PetscMemzero(d_nnz,na*sizeof(PetscInt));CHKERRQ(ierr); 1988 for (i=0; i<ai[ma]; i++) { 1989 d_nnz[aj[i]]++; 1990 aj[i] += cstart; /* global col index to be used by MatSetValues() */ 1991 } 1992 /* compute local off-diagonal contributions */ 1993 ierr = PetscMemzero(g_nnz,nb*sizeof(PetscInt));CHKERRQ(ierr); 1994 for (i=0; i<bi[ma]; i++) g_nnz[bj[i]]++; 1995 /* map those to global */ 1996 ierr = PetscSFCreate(PetscObjectComm((PetscObject)A),&sf);CHKERRQ(ierr); 1997 ierr = PetscSFSetGraphLayout(sf,A->cmap,nb,NULL,PETSC_USE_POINTER,a->garray);CHKERRQ(ierr); 1998 ierr = PetscSFSetFromOptions(sf);CHKERRQ(ierr); 1999 ierr = PetscMemzero(o_nnz,na*sizeof(PetscInt));CHKERRQ(ierr); 2000 ierr = PetscSFReduceBegin(sf,MPIU_INT,g_nnz,o_nnz,MPIU_SUM);CHKERRQ(ierr); 2001 ierr = PetscSFReduceEnd(sf,MPIU_INT,g_nnz,o_nnz,MPIU_SUM);CHKERRQ(ierr); 2002 ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 2003 2004 ierr = MatCreate(PetscObjectComm((PetscObject)A),&B);CHKERRQ(ierr); 2005 ierr = MatSetSizes(B,A->cmap->n,A->rmap->n,N,M);CHKERRQ(ierr); 2006 ierr = MatSetBlockSizes(B,PetscAbs(A->cmap->bs),PetscAbs(A->rmap->bs));CHKERRQ(ierr); 2007 ierr = MatSetType(B,((PetscObject)A)->type_name);CHKERRQ(ierr); 2008 ierr = MatMPIAIJSetPreallocation(B,0,d_nnz,0,o_nnz);CHKERRQ(ierr); 2009 ierr = PetscFree4(d_nnz,o_nnz,g_nnz,oloc);CHKERRQ(ierr); 2010 } else { 2011 B = *matout; 2012 ierr = MatSetOption(B,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_TRUE);CHKERRQ(ierr); 2013 for (i=0; i<ai[ma]; i++) aj[i] += cstart; /* global col index to be used by MatSetValues() */ 2014 } 2015 2016 /* copy over the A part */ 2017 array = Aloc->a; 2018 row = A->rmap->rstart; 2019 for (i=0; i<ma; i++) { 2020 ncol = ai[i+1]-ai[i]; 2021 ierr = MatSetValues(B,ncol,aj,1,&row,array,INSERT_VALUES);CHKERRQ(ierr); 2022 row++; 2023 array += ncol; aj += ncol; 2024 } 2025 aj = Aloc->j; 2026 for (i=0; i<ai[ma]; i++) aj[i] -= cstart; /* resume local col index */ 2027 2028 /* copy over the B part */ 2029 ierr = PetscCalloc1(bi[mb],&cols);CHKERRQ(ierr); 2030 array = Bloc->a; 2031 row = A->rmap->rstart; 2032 for (i=0; i<bi[mb]; i++) cols[i] = a->garray[bj[i]]; 2033 cols_tmp = cols; 2034 for (i=0; i<mb; i++) { 2035 ncol = bi[i+1]-bi[i]; 2036 ierr = MatSetValues(B,ncol,cols_tmp,1,&row,array,INSERT_VALUES);CHKERRQ(ierr); 2037 row++; 2038 array += ncol; cols_tmp += ncol; 2039 } 2040 ierr = PetscFree(cols);CHKERRQ(ierr); 2041 2042 ierr = MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2043 ierr = MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2044 if (reuse == MAT_INITIAL_MATRIX || *matout != A) { 2045 *matout = B; 2046 } else { 2047 ierr = MatHeaderMerge(A,B);CHKERRQ(ierr); 2048 } 2049 PetscFunctionReturn(0); 2050 } 2051 2052 #undef __FUNCT__ 2053 #define __FUNCT__ "MatDiagonalScale_MPIAIJ" 2054 PetscErrorCode MatDiagonalScale_MPIAIJ(Mat mat,Vec ll,Vec rr) 2055 { 2056 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 2057 Mat a = aij->A,b = aij->B; 2058 PetscErrorCode ierr; 2059 PetscInt s1,s2,s3; 2060 2061 PetscFunctionBegin; 2062 ierr = MatGetLocalSize(mat,&s2,&s3);CHKERRQ(ierr); 2063 if (rr) { 2064 ierr = VecGetLocalSize(rr,&s1);CHKERRQ(ierr); 2065 if (s1!=s3) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"right vector non-conforming local size"); 2066 /* Overlap communication with computation. */ 2067 ierr = VecScatterBegin(aij->Mvctx,rr,aij->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2068 } 2069 if (ll) { 2070 ierr = VecGetLocalSize(ll,&s1);CHKERRQ(ierr); 2071 if (s1!=s2) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"left vector non-conforming local size"); 2072 ierr = (*b->ops->diagonalscale)(b,ll,0);CHKERRQ(ierr); 2073 } 2074 /* scale the diagonal block */ 2075 ierr = (*a->ops->diagonalscale)(a,ll,rr);CHKERRQ(ierr); 2076 2077 if (rr) { 2078 /* Do a scatter end and then right scale the off-diagonal block */ 2079 ierr = VecScatterEnd(aij->Mvctx,rr,aij->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2080 ierr = (*b->ops->diagonalscale)(b,0,aij->lvec);CHKERRQ(ierr); 2081 } 2082 PetscFunctionReturn(0); 2083 } 2084 2085 #undef __FUNCT__ 2086 #define __FUNCT__ "MatSetUnfactored_MPIAIJ" 2087 PetscErrorCode MatSetUnfactored_MPIAIJ(Mat A) 2088 { 2089 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 2090 PetscErrorCode ierr; 2091 2092 PetscFunctionBegin; 2093 ierr = MatSetUnfactored(a->A);CHKERRQ(ierr); 2094 PetscFunctionReturn(0); 2095 } 2096 2097 #undef __FUNCT__ 2098 #define __FUNCT__ "MatEqual_MPIAIJ" 2099 PetscErrorCode MatEqual_MPIAIJ(Mat A,Mat B,PetscBool *flag) 2100 { 2101 Mat_MPIAIJ *matB = (Mat_MPIAIJ*)B->data,*matA = (Mat_MPIAIJ*)A->data; 2102 Mat a,b,c,d; 2103 PetscBool flg; 2104 PetscErrorCode ierr; 2105 2106 PetscFunctionBegin; 2107 a = matA->A; b = matA->B; 2108 c = matB->A; d = matB->B; 2109 2110 ierr = MatEqual(a,c,&flg);CHKERRQ(ierr); 2111 if (flg) { 2112 ierr = MatEqual(b,d,&flg);CHKERRQ(ierr); 2113 } 2114 ierr = MPI_Allreduce(&flg,flag,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)A));CHKERRQ(ierr); 2115 PetscFunctionReturn(0); 2116 } 2117 2118 #undef __FUNCT__ 2119 #define __FUNCT__ "MatCopy_MPIAIJ" 2120 PetscErrorCode MatCopy_MPIAIJ(Mat A,Mat B,MatStructure str) 2121 { 2122 PetscErrorCode ierr; 2123 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 2124 Mat_MPIAIJ *b = (Mat_MPIAIJ*)B->data; 2125 2126 PetscFunctionBegin; 2127 /* If the two matrices don't have the same copy implementation, they aren't compatible for fast copy. */ 2128 if ((str != SAME_NONZERO_PATTERN) || (A->ops->copy != B->ops->copy)) { 2129 /* because of the column compression in the off-processor part of the matrix a->B, 2130 the number of columns in a->B and b->B may be different, hence we cannot call 2131 the MatCopy() directly on the two parts. If need be, we can provide a more 2132 efficient copy than the MatCopy_Basic() by first uncompressing the a->B matrices 2133 then copying the submatrices */ 2134 ierr = MatCopy_Basic(A,B,str);CHKERRQ(ierr); 2135 } else { 2136 ierr = MatCopy(a->A,b->A,str);CHKERRQ(ierr); 2137 ierr = MatCopy(a->B,b->B,str);CHKERRQ(ierr); 2138 } 2139 PetscFunctionReturn(0); 2140 } 2141 2142 #undef __FUNCT__ 2143 #define __FUNCT__ "MatSetUp_MPIAIJ" 2144 PetscErrorCode MatSetUp_MPIAIJ(Mat A) 2145 { 2146 PetscErrorCode ierr; 2147 2148 PetscFunctionBegin; 2149 ierr = MatMPIAIJSetPreallocation(A,PETSC_DEFAULT,0,PETSC_DEFAULT,0);CHKERRQ(ierr); 2150 PetscFunctionReturn(0); 2151 } 2152 2153 /* 2154 Computes the number of nonzeros per row needed for preallocation when X and Y 2155 have different nonzero structure. 2156 */ 2157 #undef __FUNCT__ 2158 #define __FUNCT__ "MatAXPYGetPreallocation_MPIX_private" 2159 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) 2160 { 2161 PetscInt i,j,k,nzx,nzy; 2162 2163 PetscFunctionBegin; 2164 /* Set the number of nonzeros in the new matrix */ 2165 for (i=0; i<m; i++) { 2166 const PetscInt *xjj = xj+xi[i],*yjj = yj+yi[i]; 2167 nzx = xi[i+1] - xi[i]; 2168 nzy = yi[i+1] - yi[i]; 2169 nnz[i] = 0; 2170 for (j=0,k=0; j<nzx; j++) { /* Point in X */ 2171 for (; k<nzy && yltog[yjj[k]]<xltog[xjj[j]]; k++) nnz[i]++; /* Catch up to X */ 2172 if (k<nzy && yltog[yjj[k]]==xltog[xjj[j]]) k++; /* Skip duplicate */ 2173 nnz[i]++; 2174 } 2175 for (; k<nzy; k++) nnz[i]++; 2176 } 2177 PetscFunctionReturn(0); 2178 } 2179 2180 /* This is the same as MatAXPYGetPreallocation_SeqAIJ, except that the local-to-global map is provided */ 2181 #undef __FUNCT__ 2182 #define __FUNCT__ "MatAXPYGetPreallocation_MPIAIJ" 2183 static PetscErrorCode MatAXPYGetPreallocation_MPIAIJ(Mat Y,const PetscInt *yltog,Mat X,const PetscInt *xltog,PetscInt *nnz) 2184 { 2185 PetscErrorCode ierr; 2186 PetscInt m = Y->rmap->N; 2187 Mat_SeqAIJ *x = (Mat_SeqAIJ*)X->data; 2188 Mat_SeqAIJ *y = (Mat_SeqAIJ*)Y->data; 2189 2190 PetscFunctionBegin; 2191 ierr = MatAXPYGetPreallocation_MPIX_private(m,x->i,x->j,xltog,y->i,y->j,yltog,nnz);CHKERRQ(ierr); 2192 PetscFunctionReturn(0); 2193 } 2194 2195 #undef __FUNCT__ 2196 #define __FUNCT__ "MatAXPY_MPIAIJ" 2197 PetscErrorCode MatAXPY_MPIAIJ(Mat Y,PetscScalar a,Mat X,MatStructure str) 2198 { 2199 PetscErrorCode ierr; 2200 Mat_MPIAIJ *xx = (Mat_MPIAIJ*)X->data,*yy = (Mat_MPIAIJ*)Y->data; 2201 PetscBLASInt bnz,one=1; 2202 Mat_SeqAIJ *x,*y; 2203 2204 PetscFunctionBegin; 2205 if (str == SAME_NONZERO_PATTERN) { 2206 PetscScalar alpha = a; 2207 x = (Mat_SeqAIJ*)xx->A->data; 2208 ierr = PetscBLASIntCast(x->nz,&bnz);CHKERRQ(ierr); 2209 y = (Mat_SeqAIJ*)yy->A->data; 2210 PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&bnz,&alpha,x->a,&one,y->a,&one)); 2211 x = (Mat_SeqAIJ*)xx->B->data; 2212 y = (Mat_SeqAIJ*)yy->B->data; 2213 ierr = PetscBLASIntCast(x->nz,&bnz);CHKERRQ(ierr); 2214 PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&bnz,&alpha,x->a,&one,y->a,&one)); 2215 ierr = PetscObjectStateIncrease((PetscObject)Y);CHKERRQ(ierr); 2216 } else if (str == SUBSET_NONZERO_PATTERN) { /* nonzeros of X is a subset of Y's */ 2217 ierr = MatAXPY_Basic(Y,a,X,str);CHKERRQ(ierr); 2218 } else { 2219 Mat B; 2220 PetscInt *nnz_d,*nnz_o; 2221 ierr = PetscMalloc1(yy->A->rmap->N,&nnz_d);CHKERRQ(ierr); 2222 ierr = PetscMalloc1(yy->B->rmap->N,&nnz_o);CHKERRQ(ierr); 2223 ierr = MatCreate(PetscObjectComm((PetscObject)Y),&B);CHKERRQ(ierr); 2224 ierr = PetscObjectSetName((PetscObject)B,((PetscObject)Y)->name);CHKERRQ(ierr); 2225 ierr = MatSetSizes(B,Y->rmap->n,Y->cmap->n,Y->rmap->N,Y->cmap->N);CHKERRQ(ierr); 2226 ierr = MatSetBlockSizesFromMats(B,Y,Y);CHKERRQ(ierr); 2227 ierr = MatSetType(B,MATMPIAIJ);CHKERRQ(ierr); 2228 ierr = MatAXPYGetPreallocation_SeqAIJ(yy->A,xx->A,nnz_d);CHKERRQ(ierr); 2229 ierr = MatAXPYGetPreallocation_MPIAIJ(yy->B,yy->garray,xx->B,xx->garray,nnz_o);CHKERRQ(ierr); 2230 ierr = MatMPIAIJSetPreallocation(B,0,nnz_d,0,nnz_o);CHKERRQ(ierr); 2231 ierr = MatAXPY_BasicWithPreallocation(B,Y,a,X,str);CHKERRQ(ierr); 2232 ierr = MatHeaderReplace(Y,B);CHKERRQ(ierr); 2233 ierr = PetscFree(nnz_d);CHKERRQ(ierr); 2234 ierr = PetscFree(nnz_o);CHKERRQ(ierr); 2235 } 2236 PetscFunctionReturn(0); 2237 } 2238 2239 extern PetscErrorCode MatConjugate_SeqAIJ(Mat); 2240 2241 #undef __FUNCT__ 2242 #define __FUNCT__ "MatConjugate_MPIAIJ" 2243 PetscErrorCode MatConjugate_MPIAIJ(Mat mat) 2244 { 2245 #if defined(PETSC_USE_COMPLEX) 2246 PetscErrorCode ierr; 2247 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 2248 2249 PetscFunctionBegin; 2250 ierr = MatConjugate_SeqAIJ(aij->A);CHKERRQ(ierr); 2251 ierr = MatConjugate_SeqAIJ(aij->B);CHKERRQ(ierr); 2252 #else 2253 PetscFunctionBegin; 2254 #endif 2255 PetscFunctionReturn(0); 2256 } 2257 2258 #undef __FUNCT__ 2259 #define __FUNCT__ "MatRealPart_MPIAIJ" 2260 PetscErrorCode MatRealPart_MPIAIJ(Mat A) 2261 { 2262 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 2263 PetscErrorCode ierr; 2264 2265 PetscFunctionBegin; 2266 ierr = MatRealPart(a->A);CHKERRQ(ierr); 2267 ierr = MatRealPart(a->B);CHKERRQ(ierr); 2268 PetscFunctionReturn(0); 2269 } 2270 2271 #undef __FUNCT__ 2272 #define __FUNCT__ "MatImaginaryPart_MPIAIJ" 2273 PetscErrorCode MatImaginaryPart_MPIAIJ(Mat A) 2274 { 2275 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 2276 PetscErrorCode ierr; 2277 2278 PetscFunctionBegin; 2279 ierr = MatImaginaryPart(a->A);CHKERRQ(ierr); 2280 ierr = MatImaginaryPart(a->B);CHKERRQ(ierr); 2281 PetscFunctionReturn(0); 2282 } 2283 2284 #if defined(PETSC_HAVE_PBGL) 2285 2286 #include <boost/parallel/mpi/bsp_process_group.hpp> 2287 #include <boost/graph/distributed/ilu_default_graph.hpp> 2288 #include <boost/graph/distributed/ilu_0_block.hpp> 2289 #include <boost/graph/distributed/ilu_preconditioner.hpp> 2290 #include <boost/graph/distributed/petsc/interface.hpp> 2291 #include <boost/multi_array.hpp> 2292 #include <boost/parallel/distributed_property_map->hpp> 2293 2294 #undef __FUNCT__ 2295 #define __FUNCT__ "MatILUFactorSymbolic_MPIAIJ" 2296 /* 2297 This uses the parallel ILU factorization of Peter Gottschling <pgottsch@osl.iu.edu> 2298 */ 2299 PetscErrorCode MatILUFactorSymbolic_MPIAIJ(Mat fact,Mat A, IS isrow, IS iscol, const MatFactorInfo *info) 2300 { 2301 namespace petsc = boost::distributed::petsc; 2302 2303 namespace graph_dist = boost::graph::distributed; 2304 using boost::graph::distributed::ilu_default::process_group_type; 2305 using boost::graph::ilu_permuted; 2306 2307 PetscBool row_identity, col_identity; 2308 PetscContainer c; 2309 PetscInt m, n, M, N; 2310 PetscErrorCode ierr; 2311 2312 PetscFunctionBegin; 2313 if (info->levels != 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Only levels = 0 supported for parallel ilu"); 2314 ierr = ISIdentity(isrow, &row_identity);CHKERRQ(ierr); 2315 ierr = ISIdentity(iscol, &col_identity);CHKERRQ(ierr); 2316 if (!row_identity || !col_identity) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Row and column permutations must be identity for parallel ILU"); 2317 2318 process_group_type pg; 2319 typedef graph_dist::ilu_default::ilu_level_graph_type lgraph_type; 2320 lgraph_type *lgraph_p = new lgraph_type(petsc::num_global_vertices(A), pg, petsc::matrix_distribution(A, pg)); 2321 lgraph_type& level_graph = *lgraph_p; 2322 graph_dist::ilu_default::graph_type& graph(level_graph.graph); 2323 2324 petsc::read_matrix(A, graph, get(boost::edge_weight, graph)); 2325 ilu_permuted(level_graph); 2326 2327 /* put together the new matrix */ 2328 ierr = MatCreate(PetscObjectComm((PetscObject)A), fact);CHKERRQ(ierr); 2329 ierr = MatGetLocalSize(A, &m, &n);CHKERRQ(ierr); 2330 ierr = MatGetSize(A, &M, &N);CHKERRQ(ierr); 2331 ierr = MatSetSizes(fact, m, n, M, N);CHKERRQ(ierr); 2332 ierr = MatSetBlockSizesFromMats(fact,A,A);CHKERRQ(ierr); 2333 ierr = MatSetType(fact, ((PetscObject)A)->type_name);CHKERRQ(ierr); 2334 ierr = MatAssemblyBegin(fact, MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2335 ierr = MatAssemblyEnd(fact, MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2336 2337 ierr = PetscContainerCreate(PetscObjectComm((PetscObject)A), &c); 2338 ierr = PetscContainerSetPointer(c, lgraph_p); 2339 ierr = PetscObjectCompose((PetscObject) (fact), "graph", (PetscObject) c); 2340 ierr = PetscContainerDestroy(&c); 2341 PetscFunctionReturn(0); 2342 } 2343 2344 #undef __FUNCT__ 2345 #define __FUNCT__ "MatLUFactorNumeric_MPIAIJ" 2346 PetscErrorCode MatLUFactorNumeric_MPIAIJ(Mat B,Mat A, const MatFactorInfo *info) 2347 { 2348 PetscFunctionBegin; 2349 PetscFunctionReturn(0); 2350 } 2351 2352 #undef __FUNCT__ 2353 #define __FUNCT__ "MatSolve_MPIAIJ" 2354 /* 2355 This uses the parallel ILU factorization of Peter Gottschling <pgottsch@osl.iu.edu> 2356 */ 2357 PetscErrorCode MatSolve_MPIAIJ(Mat A, Vec b, Vec x) 2358 { 2359 namespace graph_dist = boost::graph::distributed; 2360 2361 typedef graph_dist::ilu_default::ilu_level_graph_type lgraph_type; 2362 lgraph_type *lgraph_p; 2363 PetscContainer c; 2364 PetscErrorCode ierr; 2365 2366 PetscFunctionBegin; 2367 ierr = PetscObjectQuery((PetscObject) A, "graph", (PetscObject*) &c);CHKERRQ(ierr); 2368 ierr = PetscContainerGetPointer(c, (void**) &lgraph_p);CHKERRQ(ierr); 2369 ierr = VecCopy(b, x);CHKERRQ(ierr); 2370 2371 PetscScalar *array_x; 2372 ierr = VecGetArray(x, &array_x);CHKERRQ(ierr); 2373 PetscInt sx; 2374 ierr = VecGetSize(x, &sx);CHKERRQ(ierr); 2375 2376 PetscScalar *array_b; 2377 ierr = VecGetArray(b, &array_b);CHKERRQ(ierr); 2378 PetscInt sb; 2379 ierr = VecGetSize(b, &sb);CHKERRQ(ierr); 2380 2381 lgraph_type& level_graph = *lgraph_p; 2382 graph_dist::ilu_default::graph_type& graph(level_graph.graph); 2383 2384 typedef boost::multi_array_ref<PetscScalar, 1> array_ref_type; 2385 array_ref_type ref_b(array_b, boost::extents[num_vertices(graph)]); 2386 array_ref_type ref_x(array_x, boost::extents[num_vertices(graph)]); 2387 2388 typedef boost::iterator_property_map<array_ref_type::iterator, 2389 boost::property_map<graph_dist::ilu_default::graph_type, boost::vertex_index_t>::type> gvector_type; 2390 gvector_type vector_b(ref_b.begin(), get(boost::vertex_index, graph)); 2391 gvector_type vector_x(ref_x.begin(), get(boost::vertex_index, graph)); 2392 2393 ilu_set_solve(*lgraph_p, vector_b, vector_x); 2394 PetscFunctionReturn(0); 2395 } 2396 #endif 2397 2398 #undef __FUNCT__ 2399 #define __FUNCT__ "MatGetRowMaxAbs_MPIAIJ" 2400 PetscErrorCode MatGetRowMaxAbs_MPIAIJ(Mat A, Vec v, PetscInt idx[]) 2401 { 2402 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 2403 PetscErrorCode ierr; 2404 PetscInt i,*idxb = 0; 2405 PetscScalar *va,*vb; 2406 Vec vtmp; 2407 2408 PetscFunctionBegin; 2409 ierr = MatGetRowMaxAbs(a->A,v,idx);CHKERRQ(ierr); 2410 ierr = VecGetArray(v,&va);CHKERRQ(ierr); 2411 if (idx) { 2412 for (i=0; i<A->rmap->n; i++) { 2413 if (PetscAbsScalar(va[i])) idx[i] += A->cmap->rstart; 2414 } 2415 } 2416 2417 ierr = VecCreateSeq(PETSC_COMM_SELF,A->rmap->n,&vtmp);CHKERRQ(ierr); 2418 if (idx) { 2419 ierr = PetscMalloc1(A->rmap->n,&idxb);CHKERRQ(ierr); 2420 } 2421 ierr = MatGetRowMaxAbs(a->B,vtmp,idxb);CHKERRQ(ierr); 2422 ierr = VecGetArray(vtmp,&vb);CHKERRQ(ierr); 2423 2424 for (i=0; i<A->rmap->n; i++) { 2425 if (PetscAbsScalar(va[i]) < PetscAbsScalar(vb[i])) { 2426 va[i] = vb[i]; 2427 if (idx) idx[i] = a->garray[idxb[i]]; 2428 } 2429 } 2430 2431 ierr = VecRestoreArray(v,&va);CHKERRQ(ierr); 2432 ierr = VecRestoreArray(vtmp,&vb);CHKERRQ(ierr); 2433 ierr = PetscFree(idxb);CHKERRQ(ierr); 2434 ierr = VecDestroy(&vtmp);CHKERRQ(ierr); 2435 PetscFunctionReturn(0); 2436 } 2437 2438 #undef __FUNCT__ 2439 #define __FUNCT__ "MatGetRowMinAbs_MPIAIJ" 2440 PetscErrorCode MatGetRowMinAbs_MPIAIJ(Mat A, Vec v, PetscInt idx[]) 2441 { 2442 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 2443 PetscErrorCode ierr; 2444 PetscInt i,*idxb = 0; 2445 PetscScalar *va,*vb; 2446 Vec vtmp; 2447 2448 PetscFunctionBegin; 2449 ierr = MatGetRowMinAbs(a->A,v,idx);CHKERRQ(ierr); 2450 ierr = VecGetArray(v,&va);CHKERRQ(ierr); 2451 if (idx) { 2452 for (i=0; i<A->cmap->n; i++) { 2453 if (PetscAbsScalar(va[i])) idx[i] += A->cmap->rstart; 2454 } 2455 } 2456 2457 ierr = VecCreateSeq(PETSC_COMM_SELF,A->rmap->n,&vtmp);CHKERRQ(ierr); 2458 if (idx) { 2459 ierr = PetscMalloc1(A->rmap->n,&idxb);CHKERRQ(ierr); 2460 } 2461 ierr = MatGetRowMinAbs(a->B,vtmp,idxb);CHKERRQ(ierr); 2462 ierr = VecGetArray(vtmp,&vb);CHKERRQ(ierr); 2463 2464 for (i=0; i<A->rmap->n; i++) { 2465 if (PetscAbsScalar(va[i]) > PetscAbsScalar(vb[i])) { 2466 va[i] = vb[i]; 2467 if (idx) idx[i] = a->garray[idxb[i]]; 2468 } 2469 } 2470 2471 ierr = VecRestoreArray(v,&va);CHKERRQ(ierr); 2472 ierr = VecRestoreArray(vtmp,&vb);CHKERRQ(ierr); 2473 ierr = PetscFree(idxb);CHKERRQ(ierr); 2474 ierr = VecDestroy(&vtmp);CHKERRQ(ierr); 2475 PetscFunctionReturn(0); 2476 } 2477 2478 #undef __FUNCT__ 2479 #define __FUNCT__ "MatGetRowMin_MPIAIJ" 2480 PetscErrorCode MatGetRowMin_MPIAIJ(Mat A, Vec v, PetscInt idx[]) 2481 { 2482 Mat_MPIAIJ *mat = (Mat_MPIAIJ*) A->data; 2483 PetscInt n = A->rmap->n; 2484 PetscInt cstart = A->cmap->rstart; 2485 PetscInt *cmap = mat->garray; 2486 PetscInt *diagIdx, *offdiagIdx; 2487 Vec diagV, offdiagV; 2488 PetscScalar *a, *diagA, *offdiagA; 2489 PetscInt r; 2490 PetscErrorCode ierr; 2491 2492 PetscFunctionBegin; 2493 ierr = PetscMalloc2(n,&diagIdx,n,&offdiagIdx);CHKERRQ(ierr); 2494 ierr = VecCreateSeq(PetscObjectComm((PetscObject)A), n, &diagV);CHKERRQ(ierr); 2495 ierr = VecCreateSeq(PetscObjectComm((PetscObject)A), n, &offdiagV);CHKERRQ(ierr); 2496 ierr = MatGetRowMin(mat->A, diagV, diagIdx);CHKERRQ(ierr); 2497 ierr = MatGetRowMin(mat->B, offdiagV, offdiagIdx);CHKERRQ(ierr); 2498 ierr = VecGetArray(v, &a);CHKERRQ(ierr); 2499 ierr = VecGetArray(diagV, &diagA);CHKERRQ(ierr); 2500 ierr = VecGetArray(offdiagV, &offdiagA);CHKERRQ(ierr); 2501 for (r = 0; r < n; ++r) { 2502 if (PetscAbsScalar(diagA[r]) <= PetscAbsScalar(offdiagA[r])) { 2503 a[r] = diagA[r]; 2504 idx[r] = cstart + diagIdx[r]; 2505 } else { 2506 a[r] = offdiagA[r]; 2507 idx[r] = cmap[offdiagIdx[r]]; 2508 } 2509 } 2510 ierr = VecRestoreArray(v, &a);CHKERRQ(ierr); 2511 ierr = VecRestoreArray(diagV, &diagA);CHKERRQ(ierr); 2512 ierr = VecRestoreArray(offdiagV, &offdiagA);CHKERRQ(ierr); 2513 ierr = VecDestroy(&diagV);CHKERRQ(ierr); 2514 ierr = VecDestroy(&offdiagV);CHKERRQ(ierr); 2515 ierr = PetscFree2(diagIdx, offdiagIdx);CHKERRQ(ierr); 2516 PetscFunctionReturn(0); 2517 } 2518 2519 #undef __FUNCT__ 2520 #define __FUNCT__ "MatGetRowMax_MPIAIJ" 2521 PetscErrorCode MatGetRowMax_MPIAIJ(Mat A, Vec v, PetscInt idx[]) 2522 { 2523 Mat_MPIAIJ *mat = (Mat_MPIAIJ*) A->data; 2524 PetscInt n = A->rmap->n; 2525 PetscInt cstart = A->cmap->rstart; 2526 PetscInt *cmap = mat->garray; 2527 PetscInt *diagIdx, *offdiagIdx; 2528 Vec diagV, offdiagV; 2529 PetscScalar *a, *diagA, *offdiagA; 2530 PetscInt r; 2531 PetscErrorCode ierr; 2532 2533 PetscFunctionBegin; 2534 ierr = PetscMalloc2(n,&diagIdx,n,&offdiagIdx);CHKERRQ(ierr); 2535 ierr = VecCreateSeq(PETSC_COMM_SELF, n, &diagV);CHKERRQ(ierr); 2536 ierr = VecCreateSeq(PETSC_COMM_SELF, n, &offdiagV);CHKERRQ(ierr); 2537 ierr = MatGetRowMax(mat->A, diagV, diagIdx);CHKERRQ(ierr); 2538 ierr = MatGetRowMax(mat->B, offdiagV, offdiagIdx);CHKERRQ(ierr); 2539 ierr = VecGetArray(v, &a);CHKERRQ(ierr); 2540 ierr = VecGetArray(diagV, &diagA);CHKERRQ(ierr); 2541 ierr = VecGetArray(offdiagV, &offdiagA);CHKERRQ(ierr); 2542 for (r = 0; r < n; ++r) { 2543 if (PetscAbsScalar(diagA[r]) >= PetscAbsScalar(offdiagA[r])) { 2544 a[r] = diagA[r]; 2545 idx[r] = cstart + diagIdx[r]; 2546 } else { 2547 a[r] = offdiagA[r]; 2548 idx[r] = cmap[offdiagIdx[r]]; 2549 } 2550 } 2551 ierr = VecRestoreArray(v, &a);CHKERRQ(ierr); 2552 ierr = VecRestoreArray(diagV, &diagA);CHKERRQ(ierr); 2553 ierr = VecRestoreArray(offdiagV, &offdiagA);CHKERRQ(ierr); 2554 ierr = VecDestroy(&diagV);CHKERRQ(ierr); 2555 ierr = VecDestroy(&offdiagV);CHKERRQ(ierr); 2556 ierr = PetscFree2(diagIdx, offdiagIdx);CHKERRQ(ierr); 2557 PetscFunctionReturn(0); 2558 } 2559 2560 #undef __FUNCT__ 2561 #define __FUNCT__ "MatGetSeqNonzeroStructure_MPIAIJ" 2562 PetscErrorCode MatGetSeqNonzeroStructure_MPIAIJ(Mat mat,Mat *newmat) 2563 { 2564 PetscErrorCode ierr; 2565 Mat *dummy; 2566 2567 PetscFunctionBegin; 2568 ierr = MatGetSubMatrix_MPIAIJ_All(mat,MAT_DO_NOT_GET_VALUES,MAT_INITIAL_MATRIX,&dummy);CHKERRQ(ierr); 2569 *newmat = *dummy; 2570 ierr = PetscFree(dummy);CHKERRQ(ierr); 2571 PetscFunctionReturn(0); 2572 } 2573 2574 #undef __FUNCT__ 2575 #define __FUNCT__ "MatInvertBlockDiagonal_MPIAIJ" 2576 PetscErrorCode MatInvertBlockDiagonal_MPIAIJ(Mat A,const PetscScalar **values) 2577 { 2578 Mat_MPIAIJ *a = (Mat_MPIAIJ*) A->data; 2579 PetscErrorCode ierr; 2580 2581 PetscFunctionBegin; 2582 ierr = MatInvertBlockDiagonal(a->A,values);CHKERRQ(ierr); 2583 PetscFunctionReturn(0); 2584 } 2585 2586 #undef __FUNCT__ 2587 #define __FUNCT__ "MatSetRandom_MPIAIJ" 2588 static PetscErrorCode MatSetRandom_MPIAIJ(Mat x,PetscRandom rctx) 2589 { 2590 PetscErrorCode ierr; 2591 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)x->data; 2592 2593 PetscFunctionBegin; 2594 ierr = MatSetRandom(aij->A,rctx);CHKERRQ(ierr); 2595 ierr = MatSetRandom(aij->B,rctx);CHKERRQ(ierr); 2596 ierr = MatAssemblyBegin(x,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2597 ierr = MatAssemblyEnd(x,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2598 PetscFunctionReturn(0); 2599 } 2600 2601 #undef __FUNCT__ 2602 #define __FUNCT__ "MatShift_MPIAIJ" 2603 PetscErrorCode MatShift_MPIAIJ(Mat Y,PetscScalar a) 2604 { 2605 PetscErrorCode ierr; 2606 Mat_MPIAIJ *maij = (Mat_MPIAIJ*)Y->data; 2607 Mat_SeqAIJ *aij = (Mat_SeqAIJ*)maij->A->data,*bij = (Mat_SeqAIJ*)maij->B->data; 2608 2609 PetscFunctionBegin; 2610 if (!aij->nz && !bij->nz) { 2611 ierr = MatMPIAIJSetPreallocation(Y,1,NULL,0,NULL);CHKERRQ(ierr); 2612 } 2613 ierr = MatShift_Basic(Y,a);CHKERRQ(ierr); 2614 PetscFunctionReturn(0); 2615 } 2616 2617 /* -------------------------------------------------------------------*/ 2618 static struct _MatOps MatOps_Values = {MatSetValues_MPIAIJ, 2619 MatGetRow_MPIAIJ, 2620 MatRestoreRow_MPIAIJ, 2621 MatMult_MPIAIJ, 2622 /* 4*/ MatMultAdd_MPIAIJ, 2623 MatMultTranspose_MPIAIJ, 2624 MatMultTransposeAdd_MPIAIJ, 2625 #if defined(PETSC_HAVE_PBGL) 2626 MatSolve_MPIAIJ, 2627 #else 2628 0, 2629 #endif 2630 0, 2631 0, 2632 /*10*/ 0, 2633 0, 2634 0, 2635 MatSOR_MPIAIJ, 2636 MatTranspose_MPIAIJ, 2637 /*15*/ MatGetInfo_MPIAIJ, 2638 MatEqual_MPIAIJ, 2639 MatGetDiagonal_MPIAIJ, 2640 MatDiagonalScale_MPIAIJ, 2641 MatNorm_MPIAIJ, 2642 /*20*/ MatAssemblyBegin_MPIAIJ, 2643 MatAssemblyEnd_MPIAIJ, 2644 MatSetOption_MPIAIJ, 2645 MatZeroEntries_MPIAIJ, 2646 /*24*/ MatZeroRows_MPIAIJ, 2647 0, 2648 #if defined(PETSC_HAVE_PBGL) 2649 0, 2650 #else 2651 0, 2652 #endif 2653 0, 2654 0, 2655 /*29*/ MatSetUp_MPIAIJ, 2656 #if defined(PETSC_HAVE_PBGL) 2657 0, 2658 #else 2659 0, 2660 #endif 2661 0, 2662 0, 2663 0, 2664 /*34*/ MatDuplicate_MPIAIJ, 2665 0, 2666 0, 2667 0, 2668 0, 2669 /*39*/ MatAXPY_MPIAIJ, 2670 MatGetSubMatrices_MPIAIJ, 2671 MatIncreaseOverlap_MPIAIJ, 2672 MatGetValues_MPIAIJ, 2673 MatCopy_MPIAIJ, 2674 /*44*/ MatGetRowMax_MPIAIJ, 2675 MatScale_MPIAIJ, 2676 MatShift_MPIAIJ, 2677 MatDiagonalSet_MPIAIJ, 2678 MatZeroRowsColumns_MPIAIJ, 2679 /*49*/ MatSetRandom_MPIAIJ, 2680 0, 2681 0, 2682 0, 2683 0, 2684 /*54*/ MatFDColoringCreate_MPIXAIJ, 2685 0, 2686 MatSetUnfactored_MPIAIJ, 2687 MatPermute_MPIAIJ, 2688 0, 2689 /*59*/ MatGetSubMatrix_MPIAIJ, 2690 MatDestroy_MPIAIJ, 2691 MatView_MPIAIJ, 2692 0, 2693 MatMatMatMult_MPIAIJ_MPIAIJ_MPIAIJ, 2694 /*64*/ MatMatMatMultSymbolic_MPIAIJ_MPIAIJ_MPIAIJ, 2695 MatMatMatMultNumeric_MPIAIJ_MPIAIJ_MPIAIJ, 2696 0, 2697 0, 2698 0, 2699 /*69*/ MatGetRowMaxAbs_MPIAIJ, 2700 MatGetRowMinAbs_MPIAIJ, 2701 0, 2702 MatSetColoring_MPIAIJ, 2703 0, 2704 MatSetValuesAdifor_MPIAIJ, 2705 /*75*/ MatFDColoringApply_AIJ, 2706 0, 2707 0, 2708 0, 2709 MatFindZeroDiagonals_MPIAIJ, 2710 /*80*/ 0, 2711 0, 2712 0, 2713 /*83*/ MatLoad_MPIAIJ, 2714 0, 2715 0, 2716 0, 2717 0, 2718 0, 2719 /*89*/ MatMatMult_MPIAIJ_MPIAIJ, 2720 MatMatMultSymbolic_MPIAIJ_MPIAIJ, 2721 MatMatMultNumeric_MPIAIJ_MPIAIJ, 2722 MatPtAP_MPIAIJ_MPIAIJ, 2723 MatPtAPSymbolic_MPIAIJ_MPIAIJ, 2724 /*94*/ MatPtAPNumeric_MPIAIJ_MPIAIJ, 2725 0, 2726 0, 2727 0, 2728 0, 2729 /*99*/ 0, 2730 0, 2731 0, 2732 MatConjugate_MPIAIJ, 2733 0, 2734 /*104*/MatSetValuesRow_MPIAIJ, 2735 MatRealPart_MPIAIJ, 2736 MatImaginaryPart_MPIAIJ, 2737 0, 2738 0, 2739 /*109*/0, 2740 0, 2741 MatGetRowMin_MPIAIJ, 2742 0, 2743 0, 2744 /*114*/MatGetSeqNonzeroStructure_MPIAIJ, 2745 0, 2746 MatGetGhosts_MPIAIJ, 2747 0, 2748 0, 2749 /*119*/0, 2750 0, 2751 0, 2752 0, 2753 MatGetMultiProcBlock_MPIAIJ, 2754 /*124*/MatFindNonzeroRows_MPIAIJ, 2755 MatGetColumnNorms_MPIAIJ, 2756 MatInvertBlockDiagonal_MPIAIJ, 2757 0, 2758 MatGetSubMatricesMPI_MPIAIJ, 2759 /*129*/0, 2760 MatTransposeMatMult_MPIAIJ_MPIAIJ, 2761 MatTransposeMatMultSymbolic_MPIAIJ_MPIAIJ, 2762 MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ, 2763 0, 2764 /*134*/0, 2765 0, 2766 0, 2767 0, 2768 0, 2769 /*139*/0, 2770 0, 2771 0, 2772 MatFDColoringSetUp_MPIXAIJ, 2773 MatFindOffBlockDiagonalEntries_MPIAIJ, 2774 /*144*/MatCreateMPIMatConcatenateSeqMat_MPIAIJ 2775 }; 2776 2777 /* ----------------------------------------------------------------------------------------*/ 2778 2779 #undef __FUNCT__ 2780 #define __FUNCT__ "MatStoreValues_MPIAIJ" 2781 PetscErrorCode MatStoreValues_MPIAIJ(Mat mat) 2782 { 2783 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 2784 PetscErrorCode ierr; 2785 2786 PetscFunctionBegin; 2787 ierr = MatStoreValues(aij->A);CHKERRQ(ierr); 2788 ierr = MatStoreValues(aij->B);CHKERRQ(ierr); 2789 PetscFunctionReturn(0); 2790 } 2791 2792 #undef __FUNCT__ 2793 #define __FUNCT__ "MatRetrieveValues_MPIAIJ" 2794 PetscErrorCode MatRetrieveValues_MPIAIJ(Mat mat) 2795 { 2796 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 2797 PetscErrorCode ierr; 2798 2799 PetscFunctionBegin; 2800 ierr = MatRetrieveValues(aij->A);CHKERRQ(ierr); 2801 ierr = MatRetrieveValues(aij->B);CHKERRQ(ierr); 2802 PetscFunctionReturn(0); 2803 } 2804 2805 #undef __FUNCT__ 2806 #define __FUNCT__ "MatMPIAIJSetPreallocation_MPIAIJ" 2807 PetscErrorCode MatMPIAIJSetPreallocation_MPIAIJ(Mat B,PetscInt d_nz,const PetscInt d_nnz[],PetscInt o_nz,const PetscInt o_nnz[]) 2808 { 2809 Mat_MPIAIJ *b; 2810 PetscErrorCode ierr; 2811 2812 PetscFunctionBegin; 2813 ierr = PetscLayoutSetUp(B->rmap);CHKERRQ(ierr); 2814 ierr = PetscLayoutSetUp(B->cmap);CHKERRQ(ierr); 2815 b = (Mat_MPIAIJ*)B->data; 2816 2817 if (!B->preallocated) { 2818 /* Explicitly create 2 MATSEQAIJ matrices. */ 2819 ierr = MatCreate(PETSC_COMM_SELF,&b->A);CHKERRQ(ierr); 2820 ierr = MatSetSizes(b->A,B->rmap->n,B->cmap->n,B->rmap->n,B->cmap->n);CHKERRQ(ierr); 2821 ierr = MatSetBlockSizesFromMats(b->A,B,B);CHKERRQ(ierr); 2822 ierr = MatSetType(b->A,MATSEQAIJ);CHKERRQ(ierr); 2823 ierr = PetscLogObjectParent((PetscObject)B,(PetscObject)b->A);CHKERRQ(ierr); 2824 ierr = MatCreate(PETSC_COMM_SELF,&b->B);CHKERRQ(ierr); 2825 ierr = MatSetSizes(b->B,B->rmap->n,B->cmap->N,B->rmap->n,B->cmap->N);CHKERRQ(ierr); 2826 ierr = MatSetBlockSizesFromMats(b->B,B,B);CHKERRQ(ierr); 2827 ierr = MatSetType(b->B,MATSEQAIJ);CHKERRQ(ierr); 2828 ierr = PetscLogObjectParent((PetscObject)B,(PetscObject)b->B);CHKERRQ(ierr); 2829 } 2830 2831 ierr = MatSeqAIJSetPreallocation(b->A,d_nz,d_nnz);CHKERRQ(ierr); 2832 ierr = MatSeqAIJSetPreallocation(b->B,o_nz,o_nnz);CHKERRQ(ierr); 2833 B->preallocated = PETSC_TRUE; 2834 PetscFunctionReturn(0); 2835 } 2836 2837 #undef __FUNCT__ 2838 #define __FUNCT__ "MatDuplicate_MPIAIJ" 2839 PetscErrorCode MatDuplicate_MPIAIJ(Mat matin,MatDuplicateOption cpvalues,Mat *newmat) 2840 { 2841 Mat mat; 2842 Mat_MPIAIJ *a,*oldmat = (Mat_MPIAIJ*)matin->data; 2843 PetscErrorCode ierr; 2844 2845 PetscFunctionBegin; 2846 *newmat = 0; 2847 ierr = MatCreate(PetscObjectComm((PetscObject)matin),&mat);CHKERRQ(ierr); 2848 ierr = MatSetSizes(mat,matin->rmap->n,matin->cmap->n,matin->rmap->N,matin->cmap->N);CHKERRQ(ierr); 2849 ierr = MatSetBlockSizesFromMats(mat,matin,matin);CHKERRQ(ierr); 2850 ierr = MatSetType(mat,((PetscObject)matin)->type_name);CHKERRQ(ierr); 2851 ierr = PetscMemcpy(mat->ops,matin->ops,sizeof(struct _MatOps));CHKERRQ(ierr); 2852 a = (Mat_MPIAIJ*)mat->data; 2853 2854 mat->factortype = matin->factortype; 2855 mat->assembled = PETSC_TRUE; 2856 mat->insertmode = NOT_SET_VALUES; 2857 mat->preallocated = PETSC_TRUE; 2858 2859 a->size = oldmat->size; 2860 a->rank = oldmat->rank; 2861 a->donotstash = oldmat->donotstash; 2862 a->roworiented = oldmat->roworiented; 2863 a->rowindices = 0; 2864 a->rowvalues = 0; 2865 a->getrowactive = PETSC_FALSE; 2866 2867 ierr = PetscLayoutReference(matin->rmap,&mat->rmap);CHKERRQ(ierr); 2868 ierr = PetscLayoutReference(matin->cmap,&mat->cmap);CHKERRQ(ierr); 2869 2870 if (oldmat->colmap) { 2871 #if defined(PETSC_USE_CTABLE) 2872 ierr = PetscTableCreateCopy(oldmat->colmap,&a->colmap);CHKERRQ(ierr); 2873 #else 2874 ierr = PetscMalloc1(mat->cmap->N,&a->colmap);CHKERRQ(ierr); 2875 ierr = PetscLogObjectMemory((PetscObject)mat,(mat->cmap->N)*sizeof(PetscInt));CHKERRQ(ierr); 2876 ierr = PetscMemcpy(a->colmap,oldmat->colmap,(mat->cmap->N)*sizeof(PetscInt));CHKERRQ(ierr); 2877 #endif 2878 } else a->colmap = 0; 2879 if (oldmat->garray) { 2880 PetscInt len; 2881 len = oldmat->B->cmap->n; 2882 ierr = PetscMalloc1(len+1,&a->garray);CHKERRQ(ierr); 2883 ierr = PetscLogObjectMemory((PetscObject)mat,len*sizeof(PetscInt));CHKERRQ(ierr); 2884 if (len) { ierr = PetscMemcpy(a->garray,oldmat->garray,len*sizeof(PetscInt));CHKERRQ(ierr); } 2885 } else a->garray = 0; 2886 2887 ierr = VecDuplicate(oldmat->lvec,&a->lvec);CHKERRQ(ierr); 2888 ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)a->lvec);CHKERRQ(ierr); 2889 ierr = VecScatterCopy(oldmat->Mvctx,&a->Mvctx);CHKERRQ(ierr); 2890 ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)a->Mvctx);CHKERRQ(ierr); 2891 ierr = MatDuplicate(oldmat->A,cpvalues,&a->A);CHKERRQ(ierr); 2892 ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)a->A);CHKERRQ(ierr); 2893 ierr = MatDuplicate(oldmat->B,cpvalues,&a->B);CHKERRQ(ierr); 2894 ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)a->B);CHKERRQ(ierr); 2895 ierr = PetscFunctionListDuplicate(((PetscObject)matin)->qlist,&((PetscObject)mat)->qlist);CHKERRQ(ierr); 2896 *newmat = mat; 2897 PetscFunctionReturn(0); 2898 } 2899 2900 2901 2902 #undef __FUNCT__ 2903 #define __FUNCT__ "MatLoad_MPIAIJ" 2904 PetscErrorCode MatLoad_MPIAIJ(Mat newMat, PetscViewer viewer) 2905 { 2906 PetscScalar *vals,*svals; 2907 MPI_Comm comm; 2908 PetscErrorCode ierr; 2909 PetscMPIInt rank,size,tag = ((PetscObject)viewer)->tag; 2910 PetscInt i,nz,j,rstart,rend,mmax,maxnz = 0; 2911 PetscInt header[4],*rowlengths = 0,M,N,m,*cols; 2912 PetscInt *ourlens = NULL,*procsnz = NULL,*offlens = NULL,jj,*mycols,*smycols; 2913 PetscInt cend,cstart,n,*rowners; 2914 int fd; 2915 PetscInt bs = newMat->rmap->bs; 2916 2917 PetscFunctionBegin; 2918 /* force binary viewer to load .info file if it has not yet done so */ 2919 ierr = PetscViewerSetUp(viewer);CHKERRQ(ierr); 2920 ierr = PetscObjectGetComm((PetscObject)viewer,&comm);CHKERRQ(ierr); 2921 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 2922 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 2923 ierr = PetscViewerBinaryGetDescriptor(viewer,&fd);CHKERRQ(ierr); 2924 if (!rank) { 2925 ierr = PetscBinaryRead(fd,(char*)header,4,PETSC_INT);CHKERRQ(ierr); 2926 if (header[0] != MAT_FILE_CLASSID) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED,"not matrix object"); 2927 } 2928 2929 ierr = PetscOptionsBegin(comm,NULL,"Options for loading MPIAIJ matrix","Mat");CHKERRQ(ierr); 2930 ierr = PetscOptionsInt("-matload_block_size","Set the blocksize used to store the matrix","MatLoad",bs,&bs,NULL);CHKERRQ(ierr); 2931 ierr = PetscOptionsEnd();CHKERRQ(ierr); 2932 if (bs < 0) bs = 1; 2933 2934 ierr = MPI_Bcast(header+1,3,MPIU_INT,0,comm);CHKERRQ(ierr); 2935 M = header[1]; N = header[2]; 2936 2937 /* If global sizes are set, check if they are consistent with that given in the file */ 2938 if (newMat->rmap->N >= 0 && newMat->rmap->N != M) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED,"Inconsistent # of rows:Matrix in file has (%D) and input matrix has (%D)",newMat->rmap->N,M); 2939 if (newMat->cmap->N >=0 && newMat->cmap->N != N) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED,"Inconsistent # of cols:Matrix in file has (%D) and input matrix has (%D)",newMat->cmap->N,N); 2940 2941 /* determine ownership of all (block) rows */ 2942 if (M%bs) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED, "Inconsistent # of rows (%d) and block size (%d)",M,bs); 2943 if (newMat->rmap->n < 0) m = bs*((M/bs)/size + (((M/bs) % size) > rank)); /* PETSC_DECIDE */ 2944 else m = newMat->rmap->n; /* Set by user */ 2945 2946 ierr = PetscMalloc1(size+1,&rowners);CHKERRQ(ierr); 2947 ierr = MPI_Allgather(&m,1,MPIU_INT,rowners+1,1,MPIU_INT,comm);CHKERRQ(ierr); 2948 2949 /* First process needs enough room for process with most rows */ 2950 if (!rank) { 2951 mmax = rowners[1]; 2952 for (i=2; i<=size; i++) { 2953 mmax = PetscMax(mmax, rowners[i]); 2954 } 2955 } else mmax = -1; /* unused, but compilers complain */ 2956 2957 rowners[0] = 0; 2958 for (i=2; i<=size; i++) { 2959 rowners[i] += rowners[i-1]; 2960 } 2961 rstart = rowners[rank]; 2962 rend = rowners[rank+1]; 2963 2964 /* distribute row lengths to all processors */ 2965 ierr = PetscMalloc2(m,&ourlens,m,&offlens);CHKERRQ(ierr); 2966 if (!rank) { 2967 ierr = PetscBinaryRead(fd,ourlens,m,PETSC_INT);CHKERRQ(ierr); 2968 ierr = PetscMalloc1(mmax,&rowlengths);CHKERRQ(ierr); 2969 ierr = PetscCalloc1(size,&procsnz);CHKERRQ(ierr); 2970 for (j=0; j<m; j++) { 2971 procsnz[0] += ourlens[j]; 2972 } 2973 for (i=1; i<size; i++) { 2974 ierr = PetscBinaryRead(fd,rowlengths,rowners[i+1]-rowners[i],PETSC_INT);CHKERRQ(ierr); 2975 /* calculate the number of nonzeros on each processor */ 2976 for (j=0; j<rowners[i+1]-rowners[i]; j++) { 2977 procsnz[i] += rowlengths[j]; 2978 } 2979 ierr = MPIULong_Send(rowlengths,rowners[i+1]-rowners[i],MPIU_INT,i,tag,comm);CHKERRQ(ierr); 2980 } 2981 ierr = PetscFree(rowlengths);CHKERRQ(ierr); 2982 } else { 2983 ierr = MPIULong_Recv(ourlens,m,MPIU_INT,0,tag,comm);CHKERRQ(ierr); 2984 } 2985 2986 if (!rank) { 2987 /* determine max buffer needed and allocate it */ 2988 maxnz = 0; 2989 for (i=0; i<size; i++) { 2990 maxnz = PetscMax(maxnz,procsnz[i]); 2991 } 2992 ierr = PetscMalloc1(maxnz,&cols);CHKERRQ(ierr); 2993 2994 /* read in my part of the matrix column indices */ 2995 nz = procsnz[0]; 2996 ierr = PetscMalloc1(nz,&mycols);CHKERRQ(ierr); 2997 ierr = PetscBinaryRead(fd,mycols,nz,PETSC_INT);CHKERRQ(ierr); 2998 2999 /* read in every one elses and ship off */ 3000 for (i=1; i<size; i++) { 3001 nz = procsnz[i]; 3002 ierr = PetscBinaryRead(fd,cols,nz,PETSC_INT);CHKERRQ(ierr); 3003 ierr = MPIULong_Send(cols,nz,MPIU_INT,i,tag,comm);CHKERRQ(ierr); 3004 } 3005 ierr = PetscFree(cols);CHKERRQ(ierr); 3006 } else { 3007 /* determine buffer space needed for message */ 3008 nz = 0; 3009 for (i=0; i<m; i++) { 3010 nz += ourlens[i]; 3011 } 3012 ierr = PetscMalloc1(nz,&mycols);CHKERRQ(ierr); 3013 3014 /* receive message of column indices*/ 3015 ierr = MPIULong_Recv(mycols,nz,MPIU_INT,0,tag,comm);CHKERRQ(ierr); 3016 } 3017 3018 /* determine column ownership if matrix is not square */ 3019 if (N != M) { 3020 if (newMat->cmap->n < 0) n = N/size + ((N % size) > rank); 3021 else n = newMat->cmap->n; 3022 ierr = MPI_Scan(&n,&cend,1,MPIU_INT,MPI_SUM,comm);CHKERRQ(ierr); 3023 cstart = cend - n; 3024 } else { 3025 cstart = rstart; 3026 cend = rend; 3027 n = cend - cstart; 3028 } 3029 3030 /* loop over local rows, determining number of off diagonal entries */ 3031 ierr = PetscMemzero(offlens,m*sizeof(PetscInt));CHKERRQ(ierr); 3032 jj = 0; 3033 for (i=0; i<m; i++) { 3034 for (j=0; j<ourlens[i]; j++) { 3035 if (mycols[jj] < cstart || mycols[jj] >= cend) offlens[i]++; 3036 jj++; 3037 } 3038 } 3039 3040 for (i=0; i<m; i++) { 3041 ourlens[i] -= offlens[i]; 3042 } 3043 ierr = MatSetSizes(newMat,m,n,M,N);CHKERRQ(ierr); 3044 3045 if (bs > 1) {ierr = MatSetBlockSize(newMat,bs);CHKERRQ(ierr);} 3046 3047 ierr = MatMPIAIJSetPreallocation(newMat,0,ourlens,0,offlens);CHKERRQ(ierr); 3048 3049 for (i=0; i<m; i++) { 3050 ourlens[i] += offlens[i]; 3051 } 3052 3053 if (!rank) { 3054 ierr = PetscMalloc1(maxnz+1,&vals);CHKERRQ(ierr); 3055 3056 /* read in my part of the matrix numerical values */ 3057 nz = procsnz[0]; 3058 ierr = PetscBinaryRead(fd,vals,nz,PETSC_SCALAR);CHKERRQ(ierr); 3059 3060 /* insert into matrix */ 3061 jj = rstart; 3062 smycols = mycols; 3063 svals = vals; 3064 for (i=0; i<m; i++) { 3065 ierr = MatSetValues_MPIAIJ(newMat,1,&jj,ourlens[i],smycols,svals,INSERT_VALUES);CHKERRQ(ierr); 3066 smycols += ourlens[i]; 3067 svals += ourlens[i]; 3068 jj++; 3069 } 3070 3071 /* read in other processors and ship out */ 3072 for (i=1; i<size; i++) { 3073 nz = procsnz[i]; 3074 ierr = PetscBinaryRead(fd,vals,nz,PETSC_SCALAR);CHKERRQ(ierr); 3075 ierr = MPIULong_Send(vals,nz,MPIU_SCALAR,i,((PetscObject)newMat)->tag,comm);CHKERRQ(ierr); 3076 } 3077 ierr = PetscFree(procsnz);CHKERRQ(ierr); 3078 } else { 3079 /* receive numeric values */ 3080 ierr = PetscMalloc1(nz+1,&vals);CHKERRQ(ierr); 3081 3082 /* receive message of values*/ 3083 ierr = MPIULong_Recv(vals,nz,MPIU_SCALAR,0,((PetscObject)newMat)->tag,comm);CHKERRQ(ierr); 3084 3085 /* insert into matrix */ 3086 jj = rstart; 3087 smycols = mycols; 3088 svals = vals; 3089 for (i=0; i<m; i++) { 3090 ierr = MatSetValues_MPIAIJ(newMat,1,&jj,ourlens[i],smycols,svals,INSERT_VALUES);CHKERRQ(ierr); 3091 smycols += ourlens[i]; 3092 svals += ourlens[i]; 3093 jj++; 3094 } 3095 } 3096 ierr = PetscFree2(ourlens,offlens);CHKERRQ(ierr); 3097 ierr = PetscFree(vals);CHKERRQ(ierr); 3098 ierr = PetscFree(mycols);CHKERRQ(ierr); 3099 ierr = PetscFree(rowners);CHKERRQ(ierr); 3100 ierr = MatAssemblyBegin(newMat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3101 ierr = MatAssemblyEnd(newMat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3102 PetscFunctionReturn(0); 3103 } 3104 3105 #undef __FUNCT__ 3106 #define __FUNCT__ "MatGetSubMatrix_MPIAIJ" 3107 /* TODO: Not scalable because of ISAllGather(). */ 3108 PetscErrorCode MatGetSubMatrix_MPIAIJ(Mat mat,IS isrow,IS iscol,MatReuse call,Mat *newmat) 3109 { 3110 PetscErrorCode ierr; 3111 IS iscol_local; 3112 PetscInt csize; 3113 3114 PetscFunctionBegin; 3115 ierr = ISGetLocalSize(iscol,&csize);CHKERRQ(ierr); 3116 if (call == MAT_REUSE_MATRIX) { 3117 ierr = PetscObjectQuery((PetscObject)*newmat,"ISAllGather",(PetscObject*)&iscol_local);CHKERRQ(ierr); 3118 if (!iscol_local) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Submatrix passed in was not used before, cannot reuse"); 3119 } else { 3120 PetscInt cbs; 3121 ierr = ISGetBlockSize(iscol,&cbs);CHKERRQ(ierr); 3122 ierr = ISAllGather(iscol,&iscol_local);CHKERRQ(ierr); 3123 ierr = ISSetBlockSize(iscol_local,cbs);CHKERRQ(ierr); 3124 } 3125 ierr = MatGetSubMatrix_MPIAIJ_Private(mat,isrow,iscol_local,csize,call,newmat);CHKERRQ(ierr); 3126 if (call == MAT_INITIAL_MATRIX) { 3127 ierr = PetscObjectCompose((PetscObject)*newmat,"ISAllGather",(PetscObject)iscol_local);CHKERRQ(ierr); 3128 ierr = ISDestroy(&iscol_local);CHKERRQ(ierr); 3129 } 3130 PetscFunctionReturn(0); 3131 } 3132 3133 extern PetscErrorCode MatGetSubMatrices_MPIAIJ_Local(Mat,PetscInt,const IS[],const IS[],MatReuse,PetscBool*,Mat*); 3134 #undef __FUNCT__ 3135 #define __FUNCT__ "MatGetSubMatrix_MPIAIJ_Private" 3136 /* 3137 Not great since it makes two copies of the submatrix, first an SeqAIJ 3138 in local and then by concatenating the local matrices the end result. 3139 Writing it directly would be much like MatGetSubMatrices_MPIAIJ() 3140 3141 Note: This requires a sequential iscol with all indices. 3142 */ 3143 PetscErrorCode MatGetSubMatrix_MPIAIJ_Private(Mat mat,IS isrow,IS iscol,PetscInt csize,MatReuse call,Mat *newmat) 3144 { 3145 PetscErrorCode ierr; 3146 PetscMPIInt rank,size; 3147 PetscInt i,m,n,rstart,row,rend,nz,*cwork,j,bs,cbs; 3148 PetscInt *ii,*jj,nlocal,*dlens,*olens,dlen,olen,jend,mglobal,ncol; 3149 PetscBool allcolumns, colflag; 3150 Mat M,Mreuse; 3151 MatScalar *vwork,*aa; 3152 MPI_Comm comm; 3153 Mat_SeqAIJ *aij; 3154 3155 PetscFunctionBegin; 3156 ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 3157 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 3158 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 3159 3160 ierr = ISIdentity(iscol,&colflag);CHKERRQ(ierr); 3161 ierr = ISGetLocalSize(iscol,&ncol);CHKERRQ(ierr); 3162 if (colflag && ncol == mat->cmap->N) { 3163 allcolumns = PETSC_TRUE; 3164 } else { 3165 allcolumns = PETSC_FALSE; 3166 } 3167 if (call == MAT_REUSE_MATRIX) { 3168 ierr = PetscObjectQuery((PetscObject)*newmat,"SubMatrix",(PetscObject*)&Mreuse);CHKERRQ(ierr); 3169 if (!Mreuse) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Submatrix passed in was not used before, cannot reuse"); 3170 ierr = MatGetSubMatrices_MPIAIJ_Local(mat,1,&isrow,&iscol,MAT_REUSE_MATRIX,&allcolumns,&Mreuse);CHKERRQ(ierr); 3171 } else { 3172 ierr = MatGetSubMatrices_MPIAIJ_Local(mat,1,&isrow,&iscol,MAT_INITIAL_MATRIX,&allcolumns,&Mreuse);CHKERRQ(ierr); 3173 } 3174 3175 /* 3176 m - number of local rows 3177 n - number of columns (same on all processors) 3178 rstart - first row in new global matrix generated 3179 */ 3180 ierr = MatGetSize(Mreuse,&m,&n);CHKERRQ(ierr); 3181 ierr = MatGetBlockSizes(Mreuse,&bs,&cbs);CHKERRQ(ierr); 3182 if (call == MAT_INITIAL_MATRIX) { 3183 aij = (Mat_SeqAIJ*)(Mreuse)->data; 3184 ii = aij->i; 3185 jj = aij->j; 3186 3187 /* 3188 Determine the number of non-zeros in the diagonal and off-diagonal 3189 portions of the matrix in order to do correct preallocation 3190 */ 3191 3192 /* first get start and end of "diagonal" columns */ 3193 if (csize == PETSC_DECIDE) { 3194 ierr = ISGetSize(isrow,&mglobal);CHKERRQ(ierr); 3195 if (mglobal == n) { /* square matrix */ 3196 nlocal = m; 3197 } else { 3198 nlocal = n/size + ((n % size) > rank); 3199 } 3200 } else { 3201 nlocal = csize; 3202 } 3203 ierr = MPI_Scan(&nlocal,&rend,1,MPIU_INT,MPI_SUM,comm);CHKERRQ(ierr); 3204 rstart = rend - nlocal; 3205 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); 3206 3207 /* next, compute all the lengths */ 3208 ierr = PetscMalloc1(2*m+1,&dlens);CHKERRQ(ierr); 3209 olens = dlens + m; 3210 for (i=0; i<m; i++) { 3211 jend = ii[i+1] - ii[i]; 3212 olen = 0; 3213 dlen = 0; 3214 for (j=0; j<jend; j++) { 3215 if (*jj < rstart || *jj >= rend) olen++; 3216 else dlen++; 3217 jj++; 3218 } 3219 olens[i] = olen; 3220 dlens[i] = dlen; 3221 } 3222 ierr = MatCreate(comm,&M);CHKERRQ(ierr); 3223 ierr = MatSetSizes(M,m,nlocal,PETSC_DECIDE,n);CHKERRQ(ierr); 3224 ierr = MatSetBlockSizes(M,bs,cbs);CHKERRQ(ierr); 3225 ierr = MatSetType(M,((PetscObject)mat)->type_name);CHKERRQ(ierr); 3226 ierr = MatMPIAIJSetPreallocation(M,0,dlens,0,olens);CHKERRQ(ierr); 3227 ierr = PetscFree(dlens);CHKERRQ(ierr); 3228 } else { 3229 PetscInt ml,nl; 3230 3231 M = *newmat; 3232 ierr = MatGetLocalSize(M,&ml,&nl);CHKERRQ(ierr); 3233 if (ml != m) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Previous matrix must be same size/layout as request"); 3234 ierr = MatZeroEntries(M);CHKERRQ(ierr); 3235 /* 3236 The next two lines are needed so we may call MatSetValues_MPIAIJ() below directly, 3237 rather than the slower MatSetValues(). 3238 */ 3239 M->was_assembled = PETSC_TRUE; 3240 M->assembled = PETSC_FALSE; 3241 } 3242 ierr = MatGetOwnershipRange(M,&rstart,&rend);CHKERRQ(ierr); 3243 aij = (Mat_SeqAIJ*)(Mreuse)->data; 3244 ii = aij->i; 3245 jj = aij->j; 3246 aa = aij->a; 3247 for (i=0; i<m; i++) { 3248 row = rstart + i; 3249 nz = ii[i+1] - ii[i]; 3250 cwork = jj; jj += nz; 3251 vwork = aa; aa += nz; 3252 ierr = MatSetValues_MPIAIJ(M,1,&row,nz,cwork,vwork,INSERT_VALUES);CHKERRQ(ierr); 3253 } 3254 3255 ierr = MatAssemblyBegin(M,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3256 ierr = MatAssemblyEnd(M,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3257 *newmat = M; 3258 3259 /* save submatrix used in processor for next request */ 3260 if (call == MAT_INITIAL_MATRIX) { 3261 ierr = PetscObjectCompose((PetscObject)M,"SubMatrix",(PetscObject)Mreuse);CHKERRQ(ierr); 3262 ierr = MatDestroy(&Mreuse);CHKERRQ(ierr); 3263 } 3264 PetscFunctionReturn(0); 3265 } 3266 3267 #undef __FUNCT__ 3268 #define __FUNCT__ "MatMPIAIJSetPreallocationCSR_MPIAIJ" 3269 PetscErrorCode MatMPIAIJSetPreallocationCSR_MPIAIJ(Mat B,const PetscInt Ii[],const PetscInt J[],const PetscScalar v[]) 3270 { 3271 PetscInt m,cstart, cend,j,nnz,i,d; 3272 PetscInt *d_nnz,*o_nnz,nnz_max = 0,rstart,ii; 3273 const PetscInt *JJ; 3274 PetscScalar *values; 3275 PetscErrorCode ierr; 3276 3277 PetscFunctionBegin; 3278 if (Ii[0]) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Ii[0] must be 0 it is %D",Ii[0]); 3279 3280 ierr = PetscLayoutSetUp(B->rmap);CHKERRQ(ierr); 3281 ierr = PetscLayoutSetUp(B->cmap);CHKERRQ(ierr); 3282 m = B->rmap->n; 3283 cstart = B->cmap->rstart; 3284 cend = B->cmap->rend; 3285 rstart = B->rmap->rstart; 3286 3287 ierr = PetscMalloc2(m,&d_nnz,m,&o_nnz);CHKERRQ(ierr); 3288 3289 #if defined(PETSC_USE_DEBUGGING) 3290 for (i=0; i<m; i++) { 3291 nnz = Ii[i+1]- Ii[i]; 3292 JJ = J + Ii[i]; 3293 if (nnz < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Local row %D has a negative %D number of columns",i,nnz); 3294 if (nnz && (JJ[0] < 0)) SETERRRQ1(PETSC_ERR_ARG_WRONGSTATE,"Row %D starts with negative column index",i,j); 3295 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); 3296 } 3297 #endif 3298 3299 for (i=0; i<m; i++) { 3300 nnz = Ii[i+1]- Ii[i]; 3301 JJ = J + Ii[i]; 3302 nnz_max = PetscMax(nnz_max,nnz); 3303 d = 0; 3304 for (j=0; j<nnz; j++) { 3305 if (cstart <= JJ[j] && JJ[j] < cend) d++; 3306 } 3307 d_nnz[i] = d; 3308 o_nnz[i] = nnz - d; 3309 } 3310 ierr = MatMPIAIJSetPreallocation(B,0,d_nnz,0,o_nnz);CHKERRQ(ierr); 3311 ierr = PetscFree2(d_nnz,o_nnz);CHKERRQ(ierr); 3312 3313 if (v) values = (PetscScalar*)v; 3314 else { 3315 ierr = PetscCalloc1(nnz_max+1,&values);CHKERRQ(ierr); 3316 } 3317 3318 for (i=0; i<m; i++) { 3319 ii = i + rstart; 3320 nnz = Ii[i+1]- Ii[i]; 3321 ierr = MatSetValues_MPIAIJ(B,1,&ii,nnz,J+Ii[i],values+(v ? Ii[i] : 0),INSERT_VALUES);CHKERRQ(ierr); 3322 } 3323 ierr = MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3324 ierr = MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3325 3326 if (!v) { 3327 ierr = PetscFree(values);CHKERRQ(ierr); 3328 } 3329 ierr = MatSetOption(B,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE);CHKERRQ(ierr); 3330 PetscFunctionReturn(0); 3331 } 3332 3333 #undef __FUNCT__ 3334 #define __FUNCT__ "MatMPIAIJSetPreallocationCSR" 3335 /*@ 3336 MatMPIAIJSetPreallocationCSR - Allocates memory for a sparse parallel matrix in AIJ format 3337 (the default parallel PETSc format). 3338 3339 Collective on MPI_Comm 3340 3341 Input Parameters: 3342 + B - the matrix 3343 . i - the indices into j for the start of each local row (starts with zero) 3344 . j - the column indices for each local row (starts with zero) 3345 - v - optional values in the matrix 3346 3347 Level: developer 3348 3349 Notes: 3350 The i, j, and a arrays ARE copied by this routine into the internal format used by PETSc; 3351 thus you CANNOT change the matrix entries by changing the values of a[] after you have 3352 called this routine. Use MatCreateMPIAIJWithSplitArrays() to avoid needing to copy the arrays. 3353 3354 The i and j indices are 0 based, and i indices are indices corresponding to the local j array. 3355 3356 The format which is used for the sparse matrix input, is equivalent to a 3357 row-major ordering.. i.e for the following matrix, the input data expected is 3358 as shown 3359 3360 $ 1 0 0 3361 $ 2 0 3 P0 3362 $ ------- 3363 $ 4 5 6 P1 3364 $ 3365 $ Process0 [P0]: rows_owned=[0,1] 3366 $ i = {0,1,3} [size = nrow+1 = 2+1] 3367 $ j = {0,0,2} [size = 3] 3368 $ v = {1,2,3} [size = 3] 3369 $ 3370 $ Process1 [P1]: rows_owned=[2] 3371 $ i = {0,3} [size = nrow+1 = 1+1] 3372 $ j = {0,1,2} [size = 3] 3373 $ v = {4,5,6} [size = 3] 3374 3375 .keywords: matrix, aij, compressed row, sparse, parallel 3376 3377 .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatCreateAIJ(), MPIAIJ, 3378 MatCreateSeqAIJWithArrays(), MatCreateMPIAIJWithSplitArrays() 3379 @*/ 3380 PetscErrorCode MatMPIAIJSetPreallocationCSR(Mat B,const PetscInt i[],const PetscInt j[], const PetscScalar v[]) 3381 { 3382 PetscErrorCode ierr; 3383 3384 PetscFunctionBegin; 3385 ierr = PetscTryMethod(B,"MatMPIAIJSetPreallocationCSR_C",(Mat,const PetscInt[],const PetscInt[],const PetscScalar[]),(B,i,j,v));CHKERRQ(ierr); 3386 PetscFunctionReturn(0); 3387 } 3388 3389 #undef __FUNCT__ 3390 #define __FUNCT__ "MatMPIAIJSetPreallocation" 3391 /*@C 3392 MatMPIAIJSetPreallocation - Preallocates memory for a sparse parallel matrix in AIJ format 3393 (the default parallel PETSc format). For good matrix assembly performance 3394 the user should preallocate the matrix storage by setting the parameters 3395 d_nz (or d_nnz) and o_nz (or o_nnz). By setting these parameters accurately, 3396 performance can be increased by more than a factor of 50. 3397 3398 Collective on MPI_Comm 3399 3400 Input Parameters: 3401 + B - the matrix 3402 . d_nz - number of nonzeros per row in DIAGONAL portion of local submatrix 3403 (same value is used for all local rows) 3404 . d_nnz - array containing the number of nonzeros in the various rows of the 3405 DIAGONAL portion of the local submatrix (possibly different for each row) 3406 or NULL (PETSC_NULL_INTEGER in Fortran), if d_nz is used to specify the nonzero structure. 3407 The size of this array is equal to the number of local rows, i.e 'm'. 3408 For matrices that will be factored, you must leave room for (and set) 3409 the diagonal entry even if it is zero. 3410 . o_nz - number of nonzeros per row in the OFF-DIAGONAL portion of local 3411 submatrix (same value is used for all local rows). 3412 - o_nnz - array containing the number of nonzeros in the various rows of the 3413 OFF-DIAGONAL portion of the local submatrix (possibly different for 3414 each row) or NULL (PETSC_NULL_INTEGER in Fortran), if o_nz is used to specify the nonzero 3415 structure. The size of this array is equal to the number 3416 of local rows, i.e 'm'. 3417 3418 If the *_nnz parameter is given then the *_nz parameter is ignored 3419 3420 The AIJ format (also called the Yale sparse matrix format or 3421 compressed row storage (CSR)), is fully compatible with standard Fortran 77 3422 storage. The stored row and column indices begin with zero. 3423 See Users-Manual: ch_mat for details. 3424 3425 The parallel matrix is partitioned such that the first m0 rows belong to 3426 process 0, the next m1 rows belong to process 1, the next m2 rows belong 3427 to process 2 etc.. where m0,m1,m2... are the input parameter 'm'. 3428 3429 The DIAGONAL portion of the local submatrix of a processor can be defined 3430 as the submatrix which is obtained by extraction the part corresponding to 3431 the rows r1-r2 and columns c1-c2 of the global matrix, where r1 is the 3432 first row that belongs to the processor, r2 is the last row belonging to 3433 the this processor, and c1-c2 is range of indices of the local part of a 3434 vector suitable for applying the matrix to. This is an mxn matrix. In the 3435 common case of a square matrix, the row and column ranges are the same and 3436 the DIAGONAL part is also square. The remaining portion of the local 3437 submatrix (mxN) constitute the OFF-DIAGONAL portion. 3438 3439 If o_nnz, d_nnz are specified, then o_nz, and d_nz are ignored. 3440 3441 You can call MatGetInfo() to get information on how effective the preallocation was; 3442 for example the fields mallocs,nz_allocated,nz_used,nz_unneeded; 3443 You can also run with the option -info and look for messages with the string 3444 malloc in them to see if additional memory allocation was needed. 3445 3446 Example usage: 3447 3448 Consider the following 8x8 matrix with 34 non-zero values, that is 3449 assembled across 3 processors. Lets assume that proc0 owns 3 rows, 3450 proc1 owns 3 rows, proc2 owns 2 rows. This division can be shown 3451 as follows: 3452 3453 .vb 3454 1 2 0 | 0 3 0 | 0 4 3455 Proc0 0 5 6 | 7 0 0 | 8 0 3456 9 0 10 | 11 0 0 | 12 0 3457 ------------------------------------- 3458 13 0 14 | 15 16 17 | 0 0 3459 Proc1 0 18 0 | 19 20 21 | 0 0 3460 0 0 0 | 22 23 0 | 24 0 3461 ------------------------------------- 3462 Proc2 25 26 27 | 0 0 28 | 29 0 3463 30 0 0 | 31 32 33 | 0 34 3464 .ve 3465 3466 This can be represented as a collection of submatrices as: 3467 3468 .vb 3469 A B C 3470 D E F 3471 G H I 3472 .ve 3473 3474 Where the submatrices A,B,C are owned by proc0, D,E,F are 3475 owned by proc1, G,H,I are owned by proc2. 3476 3477 The 'm' parameters for proc0,proc1,proc2 are 3,3,2 respectively. 3478 The 'n' parameters for proc0,proc1,proc2 are 3,3,2 respectively. 3479 The 'M','N' parameters are 8,8, and have the same values on all procs. 3480 3481 The DIAGONAL submatrices corresponding to proc0,proc1,proc2 are 3482 submatrices [A], [E], [I] respectively. The OFF-DIAGONAL submatrices 3483 corresponding to proc0,proc1,proc2 are [BC], [DF], [GH] respectively. 3484 Internally, each processor stores the DIAGONAL part, and the OFF-DIAGONAL 3485 part as SeqAIJ matrices. for eg: proc1 will store [E] as a SeqAIJ 3486 matrix, ans [DF] as another SeqAIJ matrix. 3487 3488 When d_nz, o_nz parameters are specified, d_nz storage elements are 3489 allocated for every row of the local diagonal submatrix, and o_nz 3490 storage locations are allocated for every row of the OFF-DIAGONAL submat. 3491 One way to choose d_nz and o_nz is to use the max nonzerors per local 3492 rows for each of the local DIAGONAL, and the OFF-DIAGONAL submatrices. 3493 In this case, the values of d_nz,o_nz are: 3494 .vb 3495 proc0 : dnz = 2, o_nz = 2 3496 proc1 : dnz = 3, o_nz = 2 3497 proc2 : dnz = 1, o_nz = 4 3498 .ve 3499 We are allocating m*(d_nz+o_nz) storage locations for every proc. This 3500 translates to 3*(2+2)=12 for proc0, 3*(3+2)=15 for proc1, 2*(1+4)=10 3501 for proc3. i.e we are using 12+15+10=37 storage locations to store 3502 34 values. 3503 3504 When d_nnz, o_nnz parameters are specified, the storage is specified 3505 for every row, coresponding to both DIAGONAL and OFF-DIAGONAL submatrices. 3506 In the above case the values for d_nnz,o_nnz are: 3507 .vb 3508 proc0: d_nnz = [2,2,2] and o_nnz = [2,2,2] 3509 proc1: d_nnz = [3,3,2] and o_nnz = [2,1,1] 3510 proc2: d_nnz = [1,1] and o_nnz = [4,4] 3511 .ve 3512 Here the space allocated is sum of all the above values i.e 34, and 3513 hence pre-allocation is perfect. 3514 3515 Level: intermediate 3516 3517 .keywords: matrix, aij, compressed row, sparse, parallel 3518 3519 .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatCreateAIJ(), MatMPIAIJSetPreallocationCSR(), 3520 MPIAIJ, MatGetInfo(), PetscSplitOwnership() 3521 @*/ 3522 PetscErrorCode MatMPIAIJSetPreallocation(Mat B,PetscInt d_nz,const PetscInt d_nnz[],PetscInt o_nz,const PetscInt o_nnz[]) 3523 { 3524 PetscErrorCode ierr; 3525 3526 PetscFunctionBegin; 3527 PetscValidHeaderSpecific(B,MAT_CLASSID,1); 3528 PetscValidType(B,1); 3529 ierr = PetscTryMethod(B,"MatMPIAIJSetPreallocation_C",(Mat,PetscInt,const PetscInt[],PetscInt,const PetscInt[]),(B,d_nz,d_nnz,o_nz,o_nnz));CHKERRQ(ierr); 3530 PetscFunctionReturn(0); 3531 } 3532 3533 #undef __FUNCT__ 3534 #define __FUNCT__ "MatCreateMPIAIJWithArrays" 3535 /*@ 3536 MatCreateMPIAIJWithArrays - creates a MPI AIJ matrix using arrays that contain in standard 3537 CSR format the local rows. 3538 3539 Collective on MPI_Comm 3540 3541 Input Parameters: 3542 + comm - MPI communicator 3543 . m - number of local rows (Cannot be PETSC_DECIDE) 3544 . n - This value should be the same as the local size used in creating the 3545 x vector for the matrix-vector product y = Ax. (or PETSC_DECIDE to have 3546 calculated if N is given) For square matrices n is almost always m. 3547 . M - number of global rows (or PETSC_DETERMINE to have calculated if m is given) 3548 . N - number of global columns (or PETSC_DETERMINE to have calculated if n is given) 3549 . i - row indices 3550 . j - column indices 3551 - a - matrix values 3552 3553 Output Parameter: 3554 . mat - the matrix 3555 3556 Level: intermediate 3557 3558 Notes: 3559 The i, j, and a arrays ARE copied by this routine into the internal format used by PETSc; 3560 thus you CANNOT change the matrix entries by changing the values of a[] after you have 3561 called this routine. Use MatCreateMPIAIJWithSplitArrays() to avoid needing to copy the arrays. 3562 3563 The i and j indices are 0 based, and i indices are indices corresponding to the local j array. 3564 3565 The format which is used for the sparse matrix input, is equivalent to a 3566 row-major ordering.. i.e for the following matrix, the input data expected is 3567 as shown 3568 3569 $ 1 0 0 3570 $ 2 0 3 P0 3571 $ ------- 3572 $ 4 5 6 P1 3573 $ 3574 $ Process0 [P0]: rows_owned=[0,1] 3575 $ i = {0,1,3} [size = nrow+1 = 2+1] 3576 $ j = {0,0,2} [size = 3] 3577 $ v = {1,2,3} [size = 3] 3578 $ 3579 $ Process1 [P1]: rows_owned=[2] 3580 $ i = {0,3} [size = nrow+1 = 1+1] 3581 $ j = {0,1,2} [size = 3] 3582 $ v = {4,5,6} [size = 3] 3583 3584 .keywords: matrix, aij, compressed row, sparse, parallel 3585 3586 .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatMPIAIJSetPreallocationCSR(), 3587 MPIAIJ, MatCreateAIJ(), MatCreateMPIAIJWithSplitArrays() 3588 @*/ 3589 PetscErrorCode MatCreateMPIAIJWithArrays(MPI_Comm comm,PetscInt m,PetscInt n,PetscInt M,PetscInt N,const PetscInt i[],const PetscInt j[],const PetscScalar a[],Mat *mat) 3590 { 3591 PetscErrorCode ierr; 3592 3593 PetscFunctionBegin; 3594 if (i[0]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"i (row indices) must start with 0"); 3595 if (m < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"local number of rows (m) cannot be PETSC_DECIDE, or negative"); 3596 ierr = MatCreate(comm,mat);CHKERRQ(ierr); 3597 ierr = MatSetSizes(*mat,m,n,M,N);CHKERRQ(ierr); 3598 /* ierr = MatSetBlockSizes(M,bs,cbs);CHKERRQ(ierr); */ 3599 ierr = MatSetType(*mat,MATMPIAIJ);CHKERRQ(ierr); 3600 ierr = MatMPIAIJSetPreallocationCSR(*mat,i,j,a);CHKERRQ(ierr); 3601 PetscFunctionReturn(0); 3602 } 3603 3604 #undef __FUNCT__ 3605 #define __FUNCT__ "MatCreateAIJ" 3606 /*@C 3607 MatCreateAIJ - Creates a sparse parallel matrix in AIJ format 3608 (the default parallel PETSc format). For good matrix assembly performance 3609 the user should preallocate the matrix storage by setting the parameters 3610 d_nz (or d_nnz) and o_nz (or o_nnz). By setting these parameters accurately, 3611 performance can be increased by more than a factor of 50. 3612 3613 Collective on MPI_Comm 3614 3615 Input Parameters: 3616 + comm - MPI communicator 3617 . m - number of local rows (or PETSC_DECIDE to have calculated if M is given) 3618 This value should be the same as the local size used in creating the 3619 y vector for the matrix-vector product y = Ax. 3620 . n - This value should be the same as the local size used in creating the 3621 x vector for the matrix-vector product y = Ax. (or PETSC_DECIDE to have 3622 calculated if N is given) For square matrices n is almost always m. 3623 . M - number of global rows (or PETSC_DETERMINE to have calculated if m is given) 3624 . N - number of global columns (or PETSC_DETERMINE to have calculated if n is given) 3625 . d_nz - number of nonzeros per row in DIAGONAL portion of local submatrix 3626 (same value is used for all local rows) 3627 . d_nnz - array containing the number of nonzeros in the various rows of the 3628 DIAGONAL portion of the local submatrix (possibly different for each row) 3629 or NULL, if d_nz is used to specify the nonzero structure. 3630 The size of this array is equal to the number of local rows, i.e 'm'. 3631 . o_nz - number of nonzeros per row in the OFF-DIAGONAL portion of local 3632 submatrix (same value is used for all local rows). 3633 - o_nnz - array containing the number of nonzeros in the various rows of the 3634 OFF-DIAGONAL portion of the local submatrix (possibly different for 3635 each row) or NULL, if o_nz is used to specify the nonzero 3636 structure. The size of this array is equal to the number 3637 of local rows, i.e 'm'. 3638 3639 Output Parameter: 3640 . A - the matrix 3641 3642 It is recommended that one use the MatCreate(), MatSetType() and/or MatSetFromOptions(), 3643 MatXXXXSetPreallocation() paradgm instead of this routine directly. 3644 [MatXXXXSetPreallocation() is, for example, MatSeqAIJSetPreallocation] 3645 3646 Notes: 3647 If the *_nnz parameter is given then the *_nz parameter is ignored 3648 3649 m,n,M,N parameters specify the size of the matrix, and its partitioning across 3650 processors, while d_nz,d_nnz,o_nz,o_nnz parameters specify the approximate 3651 storage requirements for this matrix. 3652 3653 If PETSC_DECIDE or PETSC_DETERMINE is used for a particular argument on one 3654 processor than it must be used on all processors that share the object for 3655 that argument. 3656 3657 The user MUST specify either the local or global matrix dimensions 3658 (possibly both). 3659 3660 The parallel matrix is partitioned across processors such that the 3661 first m0 rows belong to process 0, the next m1 rows belong to 3662 process 1, the next m2 rows belong to process 2 etc.. where 3663 m0,m1,m2,.. are the input parameter 'm'. i.e each processor stores 3664 values corresponding to [m x N] submatrix. 3665 3666 The columns are logically partitioned with the n0 columns belonging 3667 to 0th partition, the next n1 columns belonging to the next 3668 partition etc.. where n0,n1,n2... are the input parameter 'n'. 3669 3670 The DIAGONAL portion of the local submatrix on any given processor 3671 is the submatrix corresponding to the rows and columns m,n 3672 corresponding to the given processor. i.e diagonal matrix on 3673 process 0 is [m0 x n0], diagonal matrix on process 1 is [m1 x n1] 3674 etc. The remaining portion of the local submatrix [m x (N-n)] 3675 constitute the OFF-DIAGONAL portion. The example below better 3676 illustrates this concept. 3677 3678 For a square global matrix we define each processor's diagonal portion 3679 to be its local rows and the corresponding columns (a square submatrix); 3680 each processor's off-diagonal portion encompasses the remainder of the 3681 local matrix (a rectangular submatrix). 3682 3683 If o_nnz, d_nnz are specified, then o_nz, and d_nz are ignored. 3684 3685 When calling this routine with a single process communicator, a matrix of 3686 type SEQAIJ is returned. If a matrix of type MPIAIJ is desired for this 3687 type of communicator, use the construction mechanism: 3688 MatCreate(...,&A); MatSetType(A,MATMPIAIJ); MatSetSizes(A, m,n,M,N); MatMPIAIJSetPreallocation(A,...); 3689 3690 By default, this format uses inodes (identical nodes) when possible. 3691 We search for consecutive rows with the same nonzero structure, thereby 3692 reusing matrix information to achieve increased efficiency. 3693 3694 Options Database Keys: 3695 + -mat_no_inode - Do not use inodes 3696 . -mat_inode_limit <limit> - Sets inode limit (max limit=5) 3697 - -mat_aij_oneindex - Internally use indexing starting at 1 3698 rather than 0. Note that when calling MatSetValues(), 3699 the user still MUST index entries starting at 0! 3700 3701 3702 Example usage: 3703 3704 Consider the following 8x8 matrix with 34 non-zero values, that is 3705 assembled across 3 processors. Lets assume that proc0 owns 3 rows, 3706 proc1 owns 3 rows, proc2 owns 2 rows. This division can be shown 3707 as follows: 3708 3709 .vb 3710 1 2 0 | 0 3 0 | 0 4 3711 Proc0 0 5 6 | 7 0 0 | 8 0 3712 9 0 10 | 11 0 0 | 12 0 3713 ------------------------------------- 3714 13 0 14 | 15 16 17 | 0 0 3715 Proc1 0 18 0 | 19 20 21 | 0 0 3716 0 0 0 | 22 23 0 | 24 0 3717 ------------------------------------- 3718 Proc2 25 26 27 | 0 0 28 | 29 0 3719 30 0 0 | 31 32 33 | 0 34 3720 .ve 3721 3722 This can be represented as a collection of submatrices as: 3723 3724 .vb 3725 A B C 3726 D E F 3727 G H I 3728 .ve 3729 3730 Where the submatrices A,B,C are owned by proc0, D,E,F are 3731 owned by proc1, G,H,I are owned by proc2. 3732 3733 The 'm' parameters for proc0,proc1,proc2 are 3,3,2 respectively. 3734 The 'n' parameters for proc0,proc1,proc2 are 3,3,2 respectively. 3735 The 'M','N' parameters are 8,8, and have the same values on all procs. 3736 3737 The DIAGONAL submatrices corresponding to proc0,proc1,proc2 are 3738 submatrices [A], [E], [I] respectively. The OFF-DIAGONAL submatrices 3739 corresponding to proc0,proc1,proc2 are [BC], [DF], [GH] respectively. 3740 Internally, each processor stores the DIAGONAL part, and the OFF-DIAGONAL 3741 part as SeqAIJ matrices. for eg: proc1 will store [E] as a SeqAIJ 3742 matrix, ans [DF] as another SeqAIJ matrix. 3743 3744 When d_nz, o_nz parameters are specified, d_nz storage elements are 3745 allocated for every row of the local diagonal submatrix, and o_nz 3746 storage locations are allocated for every row of the OFF-DIAGONAL submat. 3747 One way to choose d_nz and o_nz is to use the max nonzerors per local 3748 rows for each of the local DIAGONAL, and the OFF-DIAGONAL submatrices. 3749 In this case, the values of d_nz,o_nz are: 3750 .vb 3751 proc0 : dnz = 2, o_nz = 2 3752 proc1 : dnz = 3, o_nz = 2 3753 proc2 : dnz = 1, o_nz = 4 3754 .ve 3755 We are allocating m*(d_nz+o_nz) storage locations for every proc. This 3756 translates to 3*(2+2)=12 for proc0, 3*(3+2)=15 for proc1, 2*(1+4)=10 3757 for proc3. i.e we are using 12+15+10=37 storage locations to store 3758 34 values. 3759 3760 When d_nnz, o_nnz parameters are specified, the storage is specified 3761 for every row, coresponding to both DIAGONAL and OFF-DIAGONAL submatrices. 3762 In the above case the values for d_nnz,o_nnz are: 3763 .vb 3764 proc0: d_nnz = [2,2,2] and o_nnz = [2,2,2] 3765 proc1: d_nnz = [3,3,2] and o_nnz = [2,1,1] 3766 proc2: d_nnz = [1,1] and o_nnz = [4,4] 3767 .ve 3768 Here the space allocated is sum of all the above values i.e 34, and 3769 hence pre-allocation is perfect. 3770 3771 Level: intermediate 3772 3773 .keywords: matrix, aij, compressed row, sparse, parallel 3774 3775 .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatMPIAIJSetPreallocationCSR(), 3776 MPIAIJ, MatCreateMPIAIJWithArrays() 3777 @*/ 3778 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) 3779 { 3780 PetscErrorCode ierr; 3781 PetscMPIInt size; 3782 3783 PetscFunctionBegin; 3784 ierr = MatCreate(comm,A);CHKERRQ(ierr); 3785 ierr = MatSetSizes(*A,m,n,M,N);CHKERRQ(ierr); 3786 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 3787 if (size > 1) { 3788 ierr = MatSetType(*A,MATMPIAIJ);CHKERRQ(ierr); 3789 ierr = MatMPIAIJSetPreallocation(*A,d_nz,d_nnz,o_nz,o_nnz);CHKERRQ(ierr); 3790 } else { 3791 ierr = MatSetType(*A,MATSEQAIJ);CHKERRQ(ierr); 3792 ierr = MatSeqAIJSetPreallocation(*A,d_nz,d_nnz);CHKERRQ(ierr); 3793 } 3794 PetscFunctionReturn(0); 3795 } 3796 3797 #undef __FUNCT__ 3798 #define __FUNCT__ "MatMPIAIJGetSeqAIJ" 3799 PetscErrorCode MatMPIAIJGetSeqAIJ(Mat A,Mat *Ad,Mat *Ao,const PetscInt *colmap[]) 3800 { 3801 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 3802 3803 PetscFunctionBegin; 3804 if (Ad) *Ad = a->A; 3805 if (Ao) *Ao = a->B; 3806 if (colmap) *colmap = a->garray; 3807 PetscFunctionReturn(0); 3808 } 3809 3810 #undef __FUNCT__ 3811 #define __FUNCT__ "MatSetColoring_MPIAIJ" 3812 PetscErrorCode MatSetColoring_MPIAIJ(Mat A,ISColoring coloring) 3813 { 3814 PetscErrorCode ierr; 3815 PetscInt i; 3816 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 3817 3818 PetscFunctionBegin; 3819 if (coloring->ctype == IS_COLORING_GLOBAL) { 3820 ISColoringValue *allcolors,*colors; 3821 ISColoring ocoloring; 3822 3823 /* set coloring for diagonal portion */ 3824 ierr = MatSetColoring_SeqAIJ(a->A,coloring);CHKERRQ(ierr); 3825 3826 /* set coloring for off-diagonal portion */ 3827 ierr = ISAllGatherColors(PetscObjectComm((PetscObject)A),coloring->n,coloring->colors,NULL,&allcolors);CHKERRQ(ierr); 3828 ierr = PetscMalloc1(a->B->cmap->n+1,&colors);CHKERRQ(ierr); 3829 for (i=0; i<a->B->cmap->n; i++) { 3830 colors[i] = allcolors[a->garray[i]]; 3831 } 3832 ierr = PetscFree(allcolors);CHKERRQ(ierr); 3833 ierr = ISColoringCreate(MPI_COMM_SELF,coloring->n,a->B->cmap->n,colors,PETSC_OWN_POINTER,&ocoloring);CHKERRQ(ierr); 3834 ierr = MatSetColoring_SeqAIJ(a->B,ocoloring);CHKERRQ(ierr); 3835 ierr = ISColoringDestroy(&ocoloring);CHKERRQ(ierr); 3836 } else if (coloring->ctype == IS_COLORING_GHOSTED) { 3837 ISColoringValue *colors; 3838 PetscInt *larray; 3839 ISColoring ocoloring; 3840 3841 /* set coloring for diagonal portion */ 3842 ierr = PetscMalloc1(a->A->cmap->n+1,&larray);CHKERRQ(ierr); 3843 for (i=0; i<a->A->cmap->n; i++) { 3844 larray[i] = i + A->cmap->rstart; 3845 } 3846 ierr = ISGlobalToLocalMappingApply(A->cmap->mapping,IS_GTOLM_MASK,a->A->cmap->n,larray,NULL,larray);CHKERRQ(ierr); 3847 ierr = PetscMalloc1(a->A->cmap->n+1,&colors);CHKERRQ(ierr); 3848 for (i=0; i<a->A->cmap->n; i++) { 3849 colors[i] = coloring->colors[larray[i]]; 3850 } 3851 ierr = PetscFree(larray);CHKERRQ(ierr); 3852 ierr = ISColoringCreate(PETSC_COMM_SELF,coloring->n,a->A->cmap->n,colors,PETSC_OWN_POINTER,&ocoloring);CHKERRQ(ierr); 3853 ierr = MatSetColoring_SeqAIJ(a->A,ocoloring);CHKERRQ(ierr); 3854 ierr = ISColoringDestroy(&ocoloring);CHKERRQ(ierr); 3855 3856 /* set coloring for off-diagonal portion */ 3857 ierr = PetscMalloc1(a->B->cmap->n+1,&larray);CHKERRQ(ierr); 3858 ierr = ISGlobalToLocalMappingApply(A->cmap->mapping,IS_GTOLM_MASK,a->B->cmap->n,a->garray,NULL,larray);CHKERRQ(ierr); 3859 ierr = PetscMalloc1(a->B->cmap->n+1,&colors);CHKERRQ(ierr); 3860 for (i=0; i<a->B->cmap->n; i++) { 3861 colors[i] = coloring->colors[larray[i]]; 3862 } 3863 ierr = PetscFree(larray);CHKERRQ(ierr); 3864 ierr = ISColoringCreate(MPI_COMM_SELF,coloring->n,a->B->cmap->n,colors,PETSC_OWN_POINTER,&ocoloring);CHKERRQ(ierr); 3865 ierr = MatSetColoring_SeqAIJ(a->B,ocoloring);CHKERRQ(ierr); 3866 ierr = ISColoringDestroy(&ocoloring);CHKERRQ(ierr); 3867 } else SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"No support ISColoringType %d",(int)coloring->ctype); 3868 PetscFunctionReturn(0); 3869 } 3870 3871 #undef __FUNCT__ 3872 #define __FUNCT__ "MatSetValuesAdifor_MPIAIJ" 3873 PetscErrorCode MatSetValuesAdifor_MPIAIJ(Mat A,PetscInt nl,void *advalues) 3874 { 3875 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 3876 PetscErrorCode ierr; 3877 3878 PetscFunctionBegin; 3879 ierr = MatSetValuesAdifor_SeqAIJ(a->A,nl,advalues);CHKERRQ(ierr); 3880 ierr = MatSetValuesAdifor_SeqAIJ(a->B,nl,advalues);CHKERRQ(ierr); 3881 PetscFunctionReturn(0); 3882 } 3883 3884 #undef __FUNCT__ 3885 #define __FUNCT__ "MatCreateMPIMatConcatenateSeqMat_MPIAIJ" 3886 PetscErrorCode MatCreateMPIMatConcatenateSeqMat_MPIAIJ(MPI_Comm comm,Mat inmat,PetscInt n,MatReuse scall,Mat *outmat) 3887 { 3888 PetscErrorCode ierr; 3889 PetscInt m,N,i,rstart,nnz,Ii; 3890 PetscInt *indx; 3891 PetscScalar *values; 3892 3893 PetscFunctionBegin; 3894 ierr = MatGetSize(inmat,&m,&N);CHKERRQ(ierr); 3895 if (scall == MAT_INITIAL_MATRIX) { /* symbolic phase */ 3896 PetscInt *dnz,*onz,sum,bs,cbs; 3897 3898 if (n == PETSC_DECIDE) { 3899 ierr = PetscSplitOwnership(comm,&n,&N);CHKERRQ(ierr); 3900 } 3901 /* Check sum(n) = N */ 3902 ierr = MPI_Allreduce(&n,&sum,1,MPIU_INT,MPI_SUM,comm);CHKERRQ(ierr); 3903 if (sum != N) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_INCOMP,"Sum of local columns != global columns %d",N); 3904 3905 ierr = MPI_Scan(&m, &rstart,1,MPIU_INT,MPI_SUM,comm);CHKERRQ(ierr); 3906 rstart -= m; 3907 3908 ierr = MatPreallocateInitialize(comm,m,n,dnz,onz);CHKERRQ(ierr); 3909 for (i=0; i<m; i++) { 3910 ierr = MatGetRow_SeqAIJ(inmat,i,&nnz,&indx,NULL);CHKERRQ(ierr); 3911 ierr = MatPreallocateSet(i+rstart,nnz,indx,dnz,onz);CHKERRQ(ierr); 3912 ierr = MatRestoreRow_SeqAIJ(inmat,i,&nnz,&indx,NULL);CHKERRQ(ierr); 3913 } 3914 3915 ierr = MatCreate(comm,outmat);CHKERRQ(ierr); 3916 ierr = MatSetSizes(*outmat,m,n,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 3917 ierr = MatGetBlockSizes(inmat,&bs,&cbs);CHKERRQ(ierr); 3918 ierr = MatSetBlockSizes(*outmat,bs,cbs);CHKERRQ(ierr); 3919 ierr = MatSetType(*outmat,MATMPIAIJ);CHKERRQ(ierr); 3920 ierr = MatMPIAIJSetPreallocation(*outmat,0,dnz,0,onz);CHKERRQ(ierr); 3921 ierr = MatPreallocateFinalize(dnz,onz);CHKERRQ(ierr); 3922 } 3923 3924 /* numeric phase */ 3925 ierr = MatGetOwnershipRange(*outmat,&rstart,NULL);CHKERRQ(ierr); 3926 for (i=0; i<m; i++) { 3927 ierr = MatGetRow_SeqAIJ(inmat,i,&nnz,&indx,&values);CHKERRQ(ierr); 3928 Ii = i + rstart; 3929 ierr = MatSetValues(*outmat,1,&Ii,nnz,indx,values,INSERT_VALUES);CHKERRQ(ierr); 3930 ierr = MatRestoreRow_SeqAIJ(inmat,i,&nnz,&indx,&values);CHKERRQ(ierr); 3931 } 3932 ierr = MatAssemblyBegin(*outmat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3933 ierr = MatAssemblyEnd(*outmat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3934 PetscFunctionReturn(0); 3935 } 3936 3937 #undef __FUNCT__ 3938 #define __FUNCT__ "MatFileSplit" 3939 PetscErrorCode MatFileSplit(Mat A,char *outfile) 3940 { 3941 PetscErrorCode ierr; 3942 PetscMPIInt rank; 3943 PetscInt m,N,i,rstart,nnz; 3944 size_t len; 3945 const PetscInt *indx; 3946 PetscViewer out; 3947 char *name; 3948 Mat B; 3949 const PetscScalar *values; 3950 3951 PetscFunctionBegin; 3952 ierr = MatGetLocalSize(A,&m,0);CHKERRQ(ierr); 3953 ierr = MatGetSize(A,0,&N);CHKERRQ(ierr); 3954 /* Should this be the type of the diagonal block of A? */ 3955 ierr = MatCreate(PETSC_COMM_SELF,&B);CHKERRQ(ierr); 3956 ierr = MatSetSizes(B,m,N,m,N);CHKERRQ(ierr); 3957 ierr = MatSetBlockSizesFromMats(B,A,A);CHKERRQ(ierr); 3958 ierr = MatSetType(B,MATSEQAIJ);CHKERRQ(ierr); 3959 ierr = MatSeqAIJSetPreallocation(B,0,NULL);CHKERRQ(ierr); 3960 ierr = MatGetOwnershipRange(A,&rstart,0);CHKERRQ(ierr); 3961 for (i=0; i<m; i++) { 3962 ierr = MatGetRow(A,i+rstart,&nnz,&indx,&values);CHKERRQ(ierr); 3963 ierr = MatSetValues(B,1,&i,nnz,indx,values,INSERT_VALUES);CHKERRQ(ierr); 3964 ierr = MatRestoreRow(A,i+rstart,&nnz,&indx,&values);CHKERRQ(ierr); 3965 } 3966 ierr = MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3967 ierr = MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3968 3969 ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)A),&rank);CHKERRQ(ierr); 3970 ierr = PetscStrlen(outfile,&len);CHKERRQ(ierr); 3971 ierr = PetscMalloc1(len+5,&name);CHKERRQ(ierr); 3972 sprintf(name,"%s.%d",outfile,rank); 3973 ierr = PetscViewerBinaryOpen(PETSC_COMM_SELF,name,FILE_MODE_APPEND,&out);CHKERRQ(ierr); 3974 ierr = PetscFree(name);CHKERRQ(ierr); 3975 ierr = MatView(B,out);CHKERRQ(ierr); 3976 ierr = PetscViewerDestroy(&out);CHKERRQ(ierr); 3977 ierr = MatDestroy(&B);CHKERRQ(ierr); 3978 PetscFunctionReturn(0); 3979 } 3980 3981 extern PetscErrorCode MatDestroy_MPIAIJ(Mat); 3982 #undef __FUNCT__ 3983 #define __FUNCT__ "MatDestroy_MPIAIJ_SeqsToMPI" 3984 PetscErrorCode MatDestroy_MPIAIJ_SeqsToMPI(Mat A) 3985 { 3986 PetscErrorCode ierr; 3987 Mat_Merge_SeqsToMPI *merge; 3988 PetscContainer container; 3989 3990 PetscFunctionBegin; 3991 ierr = PetscObjectQuery((PetscObject)A,"MatMergeSeqsToMPI",(PetscObject*)&container);CHKERRQ(ierr); 3992 if (container) { 3993 ierr = PetscContainerGetPointer(container,(void**)&merge);CHKERRQ(ierr); 3994 ierr = PetscFree(merge->id_r);CHKERRQ(ierr); 3995 ierr = PetscFree(merge->len_s);CHKERRQ(ierr); 3996 ierr = PetscFree(merge->len_r);CHKERRQ(ierr); 3997 ierr = PetscFree(merge->bi);CHKERRQ(ierr); 3998 ierr = PetscFree(merge->bj);CHKERRQ(ierr); 3999 ierr = PetscFree(merge->buf_ri[0]);CHKERRQ(ierr); 4000 ierr = PetscFree(merge->buf_ri);CHKERRQ(ierr); 4001 ierr = PetscFree(merge->buf_rj[0]);CHKERRQ(ierr); 4002 ierr = PetscFree(merge->buf_rj);CHKERRQ(ierr); 4003 ierr = PetscFree(merge->coi);CHKERRQ(ierr); 4004 ierr = PetscFree(merge->coj);CHKERRQ(ierr); 4005 ierr = PetscFree(merge->owners_co);CHKERRQ(ierr); 4006 ierr = PetscLayoutDestroy(&merge->rowmap);CHKERRQ(ierr); 4007 ierr = PetscFree(merge);CHKERRQ(ierr); 4008 ierr = PetscObjectCompose((PetscObject)A,"MatMergeSeqsToMPI",0);CHKERRQ(ierr); 4009 } 4010 ierr = MatDestroy_MPIAIJ(A);CHKERRQ(ierr); 4011 PetscFunctionReturn(0); 4012 } 4013 4014 #include <../src/mat/utils/freespace.h> 4015 #include <petscbt.h> 4016 4017 #undef __FUNCT__ 4018 #define __FUNCT__ "MatCreateMPIAIJSumSeqAIJNumeric" 4019 PetscErrorCode MatCreateMPIAIJSumSeqAIJNumeric(Mat seqmat,Mat mpimat) 4020 { 4021 PetscErrorCode ierr; 4022 MPI_Comm comm; 4023 Mat_SeqAIJ *a =(Mat_SeqAIJ*)seqmat->data; 4024 PetscMPIInt size,rank,taga,*len_s; 4025 PetscInt N=mpimat->cmap->N,i,j,*owners,*ai=a->i,*aj; 4026 PetscInt proc,m; 4027 PetscInt **buf_ri,**buf_rj; 4028 PetscInt k,anzi,*bj_i,*bi,*bj,arow,bnzi,nextaj; 4029 PetscInt nrows,**buf_ri_k,**nextrow,**nextai; 4030 MPI_Request *s_waits,*r_waits; 4031 MPI_Status *status; 4032 MatScalar *aa=a->a; 4033 MatScalar **abuf_r,*ba_i; 4034 Mat_Merge_SeqsToMPI *merge; 4035 PetscContainer container; 4036 4037 PetscFunctionBegin; 4038 ierr = PetscObjectGetComm((PetscObject)mpimat,&comm);CHKERRQ(ierr); 4039 ierr = PetscLogEventBegin(MAT_Seqstompinum,seqmat,0,0,0);CHKERRQ(ierr); 4040 4041 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 4042 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 4043 4044 ierr = PetscObjectQuery((PetscObject)mpimat,"MatMergeSeqsToMPI",(PetscObject*)&container);CHKERRQ(ierr); 4045 ierr = PetscContainerGetPointer(container,(void**)&merge);CHKERRQ(ierr); 4046 4047 bi = merge->bi; 4048 bj = merge->bj; 4049 buf_ri = merge->buf_ri; 4050 buf_rj = merge->buf_rj; 4051 4052 ierr = PetscMalloc1(size,&status);CHKERRQ(ierr); 4053 owners = merge->rowmap->range; 4054 len_s = merge->len_s; 4055 4056 /* send and recv matrix values */ 4057 /*-----------------------------*/ 4058 ierr = PetscObjectGetNewTag((PetscObject)mpimat,&taga);CHKERRQ(ierr); 4059 ierr = PetscPostIrecvScalar(comm,taga,merge->nrecv,merge->id_r,merge->len_r,&abuf_r,&r_waits);CHKERRQ(ierr); 4060 4061 ierr = PetscMalloc1(merge->nsend+1,&s_waits);CHKERRQ(ierr); 4062 for (proc=0,k=0; proc<size; proc++) { 4063 if (!len_s[proc]) continue; 4064 i = owners[proc]; 4065 ierr = MPI_Isend(aa+ai[i],len_s[proc],MPIU_MATSCALAR,proc,taga,comm,s_waits+k);CHKERRQ(ierr); 4066 k++; 4067 } 4068 4069 if (merge->nrecv) {ierr = MPI_Waitall(merge->nrecv,r_waits,status);CHKERRQ(ierr);} 4070 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,s_waits,status);CHKERRQ(ierr);} 4071 ierr = PetscFree(status);CHKERRQ(ierr); 4072 4073 ierr = PetscFree(s_waits);CHKERRQ(ierr); 4074 ierr = PetscFree(r_waits);CHKERRQ(ierr); 4075 4076 /* insert mat values of mpimat */ 4077 /*----------------------------*/ 4078 ierr = PetscMalloc1(N,&ba_i);CHKERRQ(ierr); 4079 ierr = PetscMalloc3(merge->nrecv,&buf_ri_k,merge->nrecv,&nextrow,merge->nrecv,&nextai);CHKERRQ(ierr); 4080 4081 for (k=0; k<merge->nrecv; k++) { 4082 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 4083 nrows = *(buf_ri_k[k]); 4084 nextrow[k] = buf_ri_k[k]+1; /* next row number of k-th recved i-structure */ 4085 nextai[k] = buf_ri_k[k] + (nrows + 1); /* poins to the next i-structure of k-th recved i-structure */ 4086 } 4087 4088 /* set values of ba */ 4089 m = merge->rowmap->n; 4090 for (i=0; i<m; i++) { 4091 arow = owners[rank] + i; 4092 bj_i = bj+bi[i]; /* col indices of the i-th row of mpimat */ 4093 bnzi = bi[i+1] - bi[i]; 4094 ierr = PetscMemzero(ba_i,bnzi*sizeof(PetscScalar));CHKERRQ(ierr); 4095 4096 /* add local non-zero vals of this proc's seqmat into ba */ 4097 anzi = ai[arow+1] - ai[arow]; 4098 aj = a->j + ai[arow]; 4099 aa = a->a + ai[arow]; 4100 nextaj = 0; 4101 for (j=0; nextaj<anzi; j++) { 4102 if (*(bj_i + j) == aj[nextaj]) { /* bcol == acol */ 4103 ba_i[j] += aa[nextaj++]; 4104 } 4105 } 4106 4107 /* add received vals into ba */ 4108 for (k=0; k<merge->nrecv; k++) { /* k-th received message */ 4109 /* i-th row */ 4110 if (i == *nextrow[k]) { 4111 anzi = *(nextai[k]+1) - *nextai[k]; 4112 aj = buf_rj[k] + *(nextai[k]); 4113 aa = abuf_r[k] + *(nextai[k]); 4114 nextaj = 0; 4115 for (j=0; nextaj<anzi; j++) { 4116 if (*(bj_i + j) == aj[nextaj]) { /* bcol == acol */ 4117 ba_i[j] += aa[nextaj++]; 4118 } 4119 } 4120 nextrow[k]++; nextai[k]++; 4121 } 4122 } 4123 ierr = MatSetValues(mpimat,1,&arow,bnzi,bj_i,ba_i,INSERT_VALUES);CHKERRQ(ierr); 4124 } 4125 ierr = MatAssemblyBegin(mpimat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4126 ierr = MatAssemblyEnd(mpimat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4127 4128 ierr = PetscFree(abuf_r[0]);CHKERRQ(ierr); 4129 ierr = PetscFree(abuf_r);CHKERRQ(ierr); 4130 ierr = PetscFree(ba_i);CHKERRQ(ierr); 4131 ierr = PetscFree3(buf_ri_k,nextrow,nextai);CHKERRQ(ierr); 4132 ierr = PetscLogEventEnd(MAT_Seqstompinum,seqmat,0,0,0);CHKERRQ(ierr); 4133 PetscFunctionReturn(0); 4134 } 4135 4136 extern PetscErrorCode MatDestroy_MPIAIJ_SeqsToMPI(Mat); 4137 4138 #undef __FUNCT__ 4139 #define __FUNCT__ "MatCreateMPIAIJSumSeqAIJSymbolic" 4140 PetscErrorCode MatCreateMPIAIJSumSeqAIJSymbolic(MPI_Comm comm,Mat seqmat,PetscInt m,PetscInt n,Mat *mpimat) 4141 { 4142 PetscErrorCode ierr; 4143 Mat B_mpi; 4144 Mat_SeqAIJ *a=(Mat_SeqAIJ*)seqmat->data; 4145 PetscMPIInt size,rank,tagi,tagj,*len_s,*len_si,*len_ri; 4146 PetscInt **buf_rj,**buf_ri,**buf_ri_k; 4147 PetscInt M=seqmat->rmap->n,N=seqmat->cmap->n,i,*owners,*ai=a->i,*aj=a->j; 4148 PetscInt len,proc,*dnz,*onz,bs,cbs; 4149 PetscInt k,anzi,*bi,*bj,*lnk,nlnk,arow,bnzi,nspacedouble=0; 4150 PetscInt nrows,*buf_s,*buf_si,*buf_si_i,**nextrow,**nextai; 4151 MPI_Request *si_waits,*sj_waits,*ri_waits,*rj_waits; 4152 MPI_Status *status; 4153 PetscFreeSpaceList free_space=NULL,current_space=NULL; 4154 PetscBT lnkbt; 4155 Mat_Merge_SeqsToMPI *merge; 4156 PetscContainer container; 4157 4158 PetscFunctionBegin; 4159 ierr = PetscLogEventBegin(MAT_Seqstompisym,seqmat,0,0,0);CHKERRQ(ierr); 4160 4161 /* make sure it is a PETSc comm */ 4162 ierr = PetscCommDuplicate(comm,&comm,NULL);CHKERRQ(ierr); 4163 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 4164 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 4165 4166 ierr = PetscNew(&merge);CHKERRQ(ierr); 4167 ierr = PetscMalloc1(size,&status);CHKERRQ(ierr); 4168 4169 /* determine row ownership */ 4170 /*---------------------------------------------------------*/ 4171 ierr = PetscLayoutCreate(comm,&merge->rowmap);CHKERRQ(ierr); 4172 ierr = PetscLayoutSetLocalSize(merge->rowmap,m);CHKERRQ(ierr); 4173 ierr = PetscLayoutSetSize(merge->rowmap,M);CHKERRQ(ierr); 4174 ierr = PetscLayoutSetBlockSize(merge->rowmap,1);CHKERRQ(ierr); 4175 ierr = PetscLayoutSetUp(merge->rowmap);CHKERRQ(ierr); 4176 ierr = PetscMalloc1(size,&len_si);CHKERRQ(ierr); 4177 ierr = PetscMalloc1(size,&merge->len_s);CHKERRQ(ierr); 4178 4179 m = merge->rowmap->n; 4180 owners = merge->rowmap->range; 4181 4182 /* determine the number of messages to send, their lengths */ 4183 /*---------------------------------------------------------*/ 4184 len_s = merge->len_s; 4185 4186 len = 0; /* length of buf_si[] */ 4187 merge->nsend = 0; 4188 for (proc=0; proc<size; proc++) { 4189 len_si[proc] = 0; 4190 if (proc == rank) { 4191 len_s[proc] = 0; 4192 } else { 4193 len_si[proc] = owners[proc+1] - owners[proc] + 1; 4194 len_s[proc] = ai[owners[proc+1]] - ai[owners[proc]]; /* num of rows to be sent to [proc] */ 4195 } 4196 if (len_s[proc]) { 4197 merge->nsend++; 4198 nrows = 0; 4199 for (i=owners[proc]; i<owners[proc+1]; i++) { 4200 if (ai[i+1] > ai[i]) nrows++; 4201 } 4202 len_si[proc] = 2*(nrows+1); 4203 len += len_si[proc]; 4204 } 4205 } 4206 4207 /* determine the number and length of messages to receive for ij-structure */ 4208 /*-------------------------------------------------------------------------*/ 4209 ierr = PetscGatherNumberOfMessages(comm,NULL,len_s,&merge->nrecv);CHKERRQ(ierr); 4210 ierr = PetscGatherMessageLengths2(comm,merge->nsend,merge->nrecv,len_s,len_si,&merge->id_r,&merge->len_r,&len_ri);CHKERRQ(ierr); 4211 4212 /* post the Irecv of j-structure */ 4213 /*-------------------------------*/ 4214 ierr = PetscCommGetNewTag(comm,&tagj);CHKERRQ(ierr); 4215 ierr = PetscPostIrecvInt(comm,tagj,merge->nrecv,merge->id_r,merge->len_r,&buf_rj,&rj_waits);CHKERRQ(ierr); 4216 4217 /* post the Isend of j-structure */ 4218 /*--------------------------------*/ 4219 ierr = PetscMalloc2(merge->nsend,&si_waits,merge->nsend,&sj_waits);CHKERRQ(ierr); 4220 4221 for (proc=0, k=0; proc<size; proc++) { 4222 if (!len_s[proc]) continue; 4223 i = owners[proc]; 4224 ierr = MPI_Isend(aj+ai[i],len_s[proc],MPIU_INT,proc,tagj,comm,sj_waits+k);CHKERRQ(ierr); 4225 k++; 4226 } 4227 4228 /* receives and sends of j-structure are complete */ 4229 /*------------------------------------------------*/ 4230 if (merge->nrecv) {ierr = MPI_Waitall(merge->nrecv,rj_waits,status);CHKERRQ(ierr);} 4231 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,sj_waits,status);CHKERRQ(ierr);} 4232 4233 /* send and recv i-structure */ 4234 /*---------------------------*/ 4235 ierr = PetscCommGetNewTag(comm,&tagi);CHKERRQ(ierr); 4236 ierr = PetscPostIrecvInt(comm,tagi,merge->nrecv,merge->id_r,len_ri,&buf_ri,&ri_waits);CHKERRQ(ierr); 4237 4238 ierr = PetscMalloc1(len+1,&buf_s);CHKERRQ(ierr); 4239 buf_si = buf_s; /* points to the beginning of k-th msg to be sent */ 4240 for (proc=0,k=0; proc<size; proc++) { 4241 if (!len_s[proc]) continue; 4242 /* form outgoing message for i-structure: 4243 buf_si[0]: nrows to be sent 4244 [1:nrows]: row index (global) 4245 [nrows+1:2*nrows+1]: i-structure index 4246 */ 4247 /*-------------------------------------------*/ 4248 nrows = len_si[proc]/2 - 1; 4249 buf_si_i = buf_si + nrows+1; 4250 buf_si[0] = nrows; 4251 buf_si_i[0] = 0; 4252 nrows = 0; 4253 for (i=owners[proc]; i<owners[proc+1]; i++) { 4254 anzi = ai[i+1] - ai[i]; 4255 if (anzi) { 4256 buf_si_i[nrows+1] = buf_si_i[nrows] + anzi; /* i-structure */ 4257 buf_si[nrows+1] = i-owners[proc]; /* local row index */ 4258 nrows++; 4259 } 4260 } 4261 ierr = MPI_Isend(buf_si,len_si[proc],MPIU_INT,proc,tagi,comm,si_waits+k);CHKERRQ(ierr); 4262 k++; 4263 buf_si += len_si[proc]; 4264 } 4265 4266 if (merge->nrecv) {ierr = MPI_Waitall(merge->nrecv,ri_waits,status);CHKERRQ(ierr);} 4267 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,si_waits,status);CHKERRQ(ierr);} 4268 4269 ierr = PetscInfo2(seqmat,"nsend: %D, nrecv: %D\n",merge->nsend,merge->nrecv);CHKERRQ(ierr); 4270 for (i=0; i<merge->nrecv; i++) { 4271 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); 4272 } 4273 4274 ierr = PetscFree(len_si);CHKERRQ(ierr); 4275 ierr = PetscFree(len_ri);CHKERRQ(ierr); 4276 ierr = PetscFree(rj_waits);CHKERRQ(ierr); 4277 ierr = PetscFree2(si_waits,sj_waits);CHKERRQ(ierr); 4278 ierr = PetscFree(ri_waits);CHKERRQ(ierr); 4279 ierr = PetscFree(buf_s);CHKERRQ(ierr); 4280 ierr = PetscFree(status);CHKERRQ(ierr); 4281 4282 /* compute a local seq matrix in each processor */ 4283 /*----------------------------------------------*/ 4284 /* allocate bi array and free space for accumulating nonzero column info */ 4285 ierr = PetscMalloc1(m+1,&bi);CHKERRQ(ierr); 4286 bi[0] = 0; 4287 4288 /* create and initialize a linked list */ 4289 nlnk = N+1; 4290 ierr = PetscLLCreate(N,N,nlnk,lnk,lnkbt);CHKERRQ(ierr); 4291 4292 /* initial FreeSpace size is 2*(num of local nnz(seqmat)) */ 4293 len = ai[owners[rank+1]] - ai[owners[rank]]; 4294 ierr = PetscFreeSpaceGet((PetscInt)(2*len+1),&free_space);CHKERRQ(ierr); 4295 4296 current_space = free_space; 4297 4298 /* determine symbolic info for each local row */ 4299 ierr = PetscMalloc3(merge->nrecv,&buf_ri_k,merge->nrecv,&nextrow,merge->nrecv,&nextai);CHKERRQ(ierr); 4300 4301 for (k=0; k<merge->nrecv; k++) { 4302 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 4303 nrows = *buf_ri_k[k]; 4304 nextrow[k] = buf_ri_k[k] + 1; /* next row number of k-th recved i-structure */ 4305 nextai[k] = buf_ri_k[k] + (nrows + 1); /* poins to the next i-structure of k-th recved i-structure */ 4306 } 4307 4308 ierr = MatPreallocateInitialize(comm,m,n,dnz,onz);CHKERRQ(ierr); 4309 len = 0; 4310 for (i=0; i<m; i++) { 4311 bnzi = 0; 4312 /* add local non-zero cols of this proc's seqmat into lnk */ 4313 arow = owners[rank] + i; 4314 anzi = ai[arow+1] - ai[arow]; 4315 aj = a->j + ai[arow]; 4316 ierr = PetscLLAddSorted(anzi,aj,N,nlnk,lnk,lnkbt);CHKERRQ(ierr); 4317 bnzi += nlnk; 4318 /* add received col data into lnk */ 4319 for (k=0; k<merge->nrecv; k++) { /* k-th received message */ 4320 if (i == *nextrow[k]) { /* i-th row */ 4321 anzi = *(nextai[k]+1) - *nextai[k]; 4322 aj = buf_rj[k] + *nextai[k]; 4323 ierr = PetscLLAddSorted(anzi,aj,N,nlnk,lnk,lnkbt);CHKERRQ(ierr); 4324 bnzi += nlnk; 4325 nextrow[k]++; nextai[k]++; 4326 } 4327 } 4328 if (len < bnzi) len = bnzi; /* =max(bnzi) */ 4329 4330 /* if free space is not available, make more free space */ 4331 if (current_space->local_remaining<bnzi) { 4332 ierr = PetscFreeSpaceGet(bnzi+current_space->total_array_size,¤t_space);CHKERRQ(ierr); 4333 nspacedouble++; 4334 } 4335 /* copy data into free space, then initialize lnk */ 4336 ierr = PetscLLClean(N,N,bnzi,lnk,current_space->array,lnkbt);CHKERRQ(ierr); 4337 ierr = MatPreallocateSet(i+owners[rank],bnzi,current_space->array,dnz,onz);CHKERRQ(ierr); 4338 4339 current_space->array += bnzi; 4340 current_space->local_used += bnzi; 4341 current_space->local_remaining -= bnzi; 4342 4343 bi[i+1] = bi[i] + bnzi; 4344 } 4345 4346 ierr = PetscFree3(buf_ri_k,nextrow,nextai);CHKERRQ(ierr); 4347 4348 ierr = PetscMalloc1(bi[m]+1,&bj);CHKERRQ(ierr); 4349 ierr = PetscFreeSpaceContiguous(&free_space,bj);CHKERRQ(ierr); 4350 ierr = PetscLLDestroy(lnk,lnkbt);CHKERRQ(ierr); 4351 4352 /* create symbolic parallel matrix B_mpi */ 4353 /*---------------------------------------*/ 4354 ierr = MatGetBlockSizes(seqmat,&bs,&cbs);CHKERRQ(ierr); 4355 ierr = MatCreate(comm,&B_mpi);CHKERRQ(ierr); 4356 if (n==PETSC_DECIDE) { 4357 ierr = MatSetSizes(B_mpi,m,n,PETSC_DETERMINE,N);CHKERRQ(ierr); 4358 } else { 4359 ierr = MatSetSizes(B_mpi,m,n,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 4360 } 4361 ierr = MatSetBlockSizes(B_mpi,bs,cbs);CHKERRQ(ierr); 4362 ierr = MatSetType(B_mpi,MATMPIAIJ);CHKERRQ(ierr); 4363 ierr = MatMPIAIJSetPreallocation(B_mpi,0,dnz,0,onz);CHKERRQ(ierr); 4364 ierr = MatPreallocateFinalize(dnz,onz);CHKERRQ(ierr); 4365 ierr = MatSetOption(B_mpi,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 4366 4367 /* B_mpi is not ready for use - assembly will be done by MatCreateMPIAIJSumSeqAIJNumeric() */ 4368 B_mpi->assembled = PETSC_FALSE; 4369 B_mpi->ops->destroy = MatDestroy_MPIAIJ_SeqsToMPI; 4370 merge->bi = bi; 4371 merge->bj = bj; 4372 merge->buf_ri = buf_ri; 4373 merge->buf_rj = buf_rj; 4374 merge->coi = NULL; 4375 merge->coj = NULL; 4376 merge->owners_co = NULL; 4377 4378 ierr = PetscCommDestroy(&comm);CHKERRQ(ierr); 4379 4380 /* attach the supporting struct to B_mpi for reuse */ 4381 ierr = PetscContainerCreate(PETSC_COMM_SELF,&container);CHKERRQ(ierr); 4382 ierr = PetscContainerSetPointer(container,merge);CHKERRQ(ierr); 4383 ierr = PetscObjectCompose((PetscObject)B_mpi,"MatMergeSeqsToMPI",(PetscObject)container);CHKERRQ(ierr); 4384 ierr = PetscContainerDestroy(&container);CHKERRQ(ierr); 4385 *mpimat = B_mpi; 4386 4387 ierr = PetscLogEventEnd(MAT_Seqstompisym,seqmat,0,0,0);CHKERRQ(ierr); 4388 PetscFunctionReturn(0); 4389 } 4390 4391 #undef __FUNCT__ 4392 #define __FUNCT__ "MatCreateMPIAIJSumSeqAIJ" 4393 /*@C 4394 MatCreateMPIAIJSumSeqAIJ - Creates a MPIAIJ matrix by adding sequential 4395 matrices from each processor 4396 4397 Collective on MPI_Comm 4398 4399 Input Parameters: 4400 + comm - the communicators the parallel matrix will live on 4401 . seqmat - the input sequential matrices 4402 . m - number of local rows (or PETSC_DECIDE) 4403 . n - number of local columns (or PETSC_DECIDE) 4404 - scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX 4405 4406 Output Parameter: 4407 . mpimat - the parallel matrix generated 4408 4409 Level: advanced 4410 4411 Notes: 4412 The dimensions of the sequential matrix in each processor MUST be the same. 4413 The input seqmat is included into the container "Mat_Merge_SeqsToMPI", and will be 4414 destroyed when mpimat is destroyed. Call PetscObjectQuery() to access seqmat. 4415 @*/ 4416 PetscErrorCode MatCreateMPIAIJSumSeqAIJ(MPI_Comm comm,Mat seqmat,PetscInt m,PetscInt n,MatReuse scall,Mat *mpimat) 4417 { 4418 PetscErrorCode ierr; 4419 PetscMPIInt size; 4420 4421 PetscFunctionBegin; 4422 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 4423 if (size == 1) { 4424 ierr = PetscLogEventBegin(MAT_Seqstompi,seqmat,0,0,0);CHKERRQ(ierr); 4425 if (scall == MAT_INITIAL_MATRIX) { 4426 ierr = MatDuplicate(seqmat,MAT_COPY_VALUES,mpimat);CHKERRQ(ierr); 4427 } else { 4428 ierr = MatCopy(seqmat,*mpimat,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 4429 } 4430 ierr = PetscLogEventEnd(MAT_Seqstompi,seqmat,0,0,0);CHKERRQ(ierr); 4431 PetscFunctionReturn(0); 4432 } 4433 ierr = PetscLogEventBegin(MAT_Seqstompi,seqmat,0,0,0);CHKERRQ(ierr); 4434 if (scall == MAT_INITIAL_MATRIX) { 4435 ierr = MatCreateMPIAIJSumSeqAIJSymbolic(comm,seqmat,m,n,mpimat);CHKERRQ(ierr); 4436 } 4437 ierr = MatCreateMPIAIJSumSeqAIJNumeric(seqmat,*mpimat);CHKERRQ(ierr); 4438 ierr = PetscLogEventEnd(MAT_Seqstompi,seqmat,0,0,0);CHKERRQ(ierr); 4439 PetscFunctionReturn(0); 4440 } 4441 4442 #undef __FUNCT__ 4443 #define __FUNCT__ "MatMPIAIJGetLocalMat" 4444 /*@ 4445 MatMPIAIJGetLocalMat - Creates a SeqAIJ from a MPIAIJ matrix by taking all its local rows and putting them into a sequential vector with 4446 mlocal rows and n columns. Where mlocal is the row count obtained with MatGetLocalSize() and n is the global column count obtained 4447 with MatGetSize() 4448 4449 Not Collective 4450 4451 Input Parameters: 4452 + A - the matrix 4453 . scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX 4454 4455 Output Parameter: 4456 . A_loc - the local sequential matrix generated 4457 4458 Level: developer 4459 4460 .seealso: MatGetOwnerShipRange(), MatMPIAIJGetLocalMatCondensed() 4461 4462 @*/ 4463 PetscErrorCode MatMPIAIJGetLocalMat(Mat A,MatReuse scall,Mat *A_loc) 4464 { 4465 PetscErrorCode ierr; 4466 Mat_MPIAIJ *mpimat=(Mat_MPIAIJ*)A->data; 4467 Mat_SeqAIJ *mat,*a,*b; 4468 PetscInt *ai,*aj,*bi,*bj,*cmap=mpimat->garray; 4469 MatScalar *aa,*ba,*cam; 4470 PetscScalar *ca; 4471 PetscInt am=A->rmap->n,i,j,k,cstart=A->cmap->rstart; 4472 PetscInt *ci,*cj,col,ncols_d,ncols_o,jo; 4473 PetscBool match; 4474 MPI_Comm comm; 4475 PetscMPIInt size; 4476 4477 PetscFunctionBegin; 4478 ierr = PetscObjectTypeCompare((PetscObject)A,MATMPIAIJ,&match);CHKERRQ(ierr); 4479 if (!match) SETERRQ(PetscObjectComm((PetscObject)A), PETSC_ERR_SUP,"Requires MPIAIJ matrix as input"); 4480 ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 4481 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 4482 if (size == 1 && scall == MAT_REUSE_MATRIX) PetscFunctionReturn(0); 4483 4484 ierr = PetscLogEventBegin(MAT_Getlocalmat,A,0,0,0);CHKERRQ(ierr); 4485 a = (Mat_SeqAIJ*)(mpimat->A)->data; 4486 b = (Mat_SeqAIJ*)(mpimat->B)->data; 4487 ai = a->i; aj = a->j; bi = b->i; bj = b->j; 4488 aa = a->a; ba = b->a; 4489 if (scall == MAT_INITIAL_MATRIX) { 4490 if (size == 1) { 4491 ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,am,A->cmap->N,ai,aj,aa,A_loc);CHKERRQ(ierr); 4492 PetscFunctionReturn(0); 4493 } 4494 4495 ierr = PetscMalloc1(1+am,&ci);CHKERRQ(ierr); 4496 ci[0] = 0; 4497 for (i=0; i<am; i++) { 4498 ci[i+1] = ci[i] + (ai[i+1] - ai[i]) + (bi[i+1] - bi[i]); 4499 } 4500 ierr = PetscMalloc1(1+ci[am],&cj);CHKERRQ(ierr); 4501 ierr = PetscMalloc1(1+ci[am],&ca);CHKERRQ(ierr); 4502 k = 0; 4503 for (i=0; i<am; i++) { 4504 ncols_o = bi[i+1] - bi[i]; 4505 ncols_d = ai[i+1] - ai[i]; 4506 /* off-diagonal portion of A */ 4507 for (jo=0; jo<ncols_o; jo++) { 4508 col = cmap[*bj]; 4509 if (col >= cstart) break; 4510 cj[k] = col; bj++; 4511 ca[k++] = *ba++; 4512 } 4513 /* diagonal portion of A */ 4514 for (j=0; j<ncols_d; j++) { 4515 cj[k] = cstart + *aj++; 4516 ca[k++] = *aa++; 4517 } 4518 /* off-diagonal portion of A */ 4519 for (j=jo; j<ncols_o; j++) { 4520 cj[k] = cmap[*bj++]; 4521 ca[k++] = *ba++; 4522 } 4523 } 4524 /* put together the new matrix */ 4525 ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,am,A->cmap->N,ci,cj,ca,A_loc);CHKERRQ(ierr); 4526 /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */ 4527 /* Since these are PETSc arrays, change flags to free them as necessary. */ 4528 mat = (Mat_SeqAIJ*)(*A_loc)->data; 4529 mat->free_a = PETSC_TRUE; 4530 mat->free_ij = PETSC_TRUE; 4531 mat->nonew = 0; 4532 } else if (scall == MAT_REUSE_MATRIX) { 4533 mat=(Mat_SeqAIJ*)(*A_loc)->data; 4534 ci = mat->i; cj = mat->j; cam = mat->a; 4535 for (i=0; i<am; i++) { 4536 /* off-diagonal portion of A */ 4537 ncols_o = bi[i+1] - bi[i]; 4538 for (jo=0; jo<ncols_o; jo++) { 4539 col = cmap[*bj]; 4540 if (col >= cstart) break; 4541 *cam++ = *ba++; bj++; 4542 } 4543 /* diagonal portion of A */ 4544 ncols_d = ai[i+1] - ai[i]; 4545 for (j=0; j<ncols_d; j++) *cam++ = *aa++; 4546 /* off-diagonal portion of A */ 4547 for (j=jo; j<ncols_o; j++) { 4548 *cam++ = *ba++; bj++; 4549 } 4550 } 4551 } else SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Invalid MatReuse %d",(int)scall); 4552 ierr = PetscLogEventEnd(MAT_Getlocalmat,A,0,0,0);CHKERRQ(ierr); 4553 PetscFunctionReturn(0); 4554 } 4555 4556 #undef __FUNCT__ 4557 #define __FUNCT__ "MatMPIAIJGetLocalMatCondensed" 4558 /*@C 4559 MatMPIAIJGetLocalMatCondensed - Creates a SeqAIJ matrix from an MPIAIJ matrix by taking all its local rows and NON-ZERO columns 4560 4561 Not Collective 4562 4563 Input Parameters: 4564 + A - the matrix 4565 . scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX 4566 - row, col - index sets of rows and columns to extract (or NULL) 4567 4568 Output Parameter: 4569 . A_loc - the local sequential matrix generated 4570 4571 Level: developer 4572 4573 .seealso: MatGetOwnershipRange(), MatMPIAIJGetLocalMat() 4574 4575 @*/ 4576 PetscErrorCode MatMPIAIJGetLocalMatCondensed(Mat A,MatReuse scall,IS *row,IS *col,Mat *A_loc) 4577 { 4578 Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data; 4579 PetscErrorCode ierr; 4580 PetscInt i,start,end,ncols,nzA,nzB,*cmap,imark,*idx; 4581 IS isrowa,iscola; 4582 Mat *aloc; 4583 PetscBool match; 4584 4585 PetscFunctionBegin; 4586 ierr = PetscObjectTypeCompare((PetscObject)A,MATMPIAIJ,&match);CHKERRQ(ierr); 4587 if (!match) SETERRQ(PetscObjectComm((PetscObject)A), PETSC_ERR_SUP,"Requires MPIAIJ matrix as input"); 4588 ierr = PetscLogEventBegin(MAT_Getlocalmatcondensed,A,0,0,0);CHKERRQ(ierr); 4589 if (!row) { 4590 start = A->rmap->rstart; end = A->rmap->rend; 4591 ierr = ISCreateStride(PETSC_COMM_SELF,end-start,start,1,&isrowa);CHKERRQ(ierr); 4592 } else { 4593 isrowa = *row; 4594 } 4595 if (!col) { 4596 start = A->cmap->rstart; 4597 cmap = a->garray; 4598 nzA = a->A->cmap->n; 4599 nzB = a->B->cmap->n; 4600 ierr = PetscMalloc1(nzA+nzB, &idx);CHKERRQ(ierr); 4601 ncols = 0; 4602 for (i=0; i<nzB; i++) { 4603 if (cmap[i] < start) idx[ncols++] = cmap[i]; 4604 else break; 4605 } 4606 imark = i; 4607 for (i=0; i<nzA; i++) idx[ncols++] = start + i; 4608 for (i=imark; i<nzB; i++) idx[ncols++] = cmap[i]; 4609 ierr = ISCreateGeneral(PETSC_COMM_SELF,ncols,idx,PETSC_OWN_POINTER,&iscola);CHKERRQ(ierr); 4610 } else { 4611 iscola = *col; 4612 } 4613 if (scall != MAT_INITIAL_MATRIX) { 4614 ierr = PetscMalloc1(1,&aloc);CHKERRQ(ierr); 4615 aloc[0] = *A_loc; 4616 } 4617 ierr = MatGetSubMatrices(A,1,&isrowa,&iscola,scall,&aloc);CHKERRQ(ierr); 4618 *A_loc = aloc[0]; 4619 ierr = PetscFree(aloc);CHKERRQ(ierr); 4620 if (!row) { 4621 ierr = ISDestroy(&isrowa);CHKERRQ(ierr); 4622 } 4623 if (!col) { 4624 ierr = ISDestroy(&iscola);CHKERRQ(ierr); 4625 } 4626 ierr = PetscLogEventEnd(MAT_Getlocalmatcondensed,A,0,0,0);CHKERRQ(ierr); 4627 PetscFunctionReturn(0); 4628 } 4629 4630 #undef __FUNCT__ 4631 #define __FUNCT__ "MatGetBrowsOfAcols" 4632 /*@C 4633 MatGetBrowsOfAcols - Creates a SeqAIJ matrix by taking rows of B that equal to nonzero columns of local A 4634 4635 Collective on Mat 4636 4637 Input Parameters: 4638 + A,B - the matrices in mpiaij format 4639 . scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX 4640 - rowb, colb - index sets of rows and columns of B to extract (or NULL) 4641 4642 Output Parameter: 4643 + rowb, colb - index sets of rows and columns of B to extract 4644 - B_seq - the sequential matrix generated 4645 4646 Level: developer 4647 4648 @*/ 4649 PetscErrorCode MatGetBrowsOfAcols(Mat A,Mat B,MatReuse scall,IS *rowb,IS *colb,Mat *B_seq) 4650 { 4651 Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data; 4652 PetscErrorCode ierr; 4653 PetscInt *idx,i,start,ncols,nzA,nzB,*cmap,imark; 4654 IS isrowb,iscolb; 4655 Mat *bseq=NULL; 4656 4657 PetscFunctionBegin; 4658 if (A->cmap->rstart != B->rmap->rstart || A->cmap->rend != B->rmap->rend) { 4659 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); 4660 } 4661 ierr = PetscLogEventBegin(MAT_GetBrowsOfAcols,A,B,0,0);CHKERRQ(ierr); 4662 4663 if (scall == MAT_INITIAL_MATRIX) { 4664 start = A->cmap->rstart; 4665 cmap = a->garray; 4666 nzA = a->A->cmap->n; 4667 nzB = a->B->cmap->n; 4668 ierr = PetscMalloc1(nzA+nzB, &idx);CHKERRQ(ierr); 4669 ncols = 0; 4670 for (i=0; i<nzB; i++) { /* row < local row index */ 4671 if (cmap[i] < start) idx[ncols++] = cmap[i]; 4672 else break; 4673 } 4674 imark = i; 4675 for (i=0; i<nzA; i++) idx[ncols++] = start + i; /* local rows */ 4676 for (i=imark; i<nzB; i++) idx[ncols++] = cmap[i]; /* row > local row index */ 4677 ierr = ISCreateGeneral(PETSC_COMM_SELF,ncols,idx,PETSC_OWN_POINTER,&isrowb);CHKERRQ(ierr); 4678 ierr = ISCreateStride(PETSC_COMM_SELF,B->cmap->N,0,1,&iscolb);CHKERRQ(ierr); 4679 } else { 4680 if (!rowb || !colb) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"IS rowb and colb must be provided for MAT_REUSE_MATRIX"); 4681 isrowb = *rowb; iscolb = *colb; 4682 ierr = PetscMalloc1(1,&bseq);CHKERRQ(ierr); 4683 bseq[0] = *B_seq; 4684 } 4685 ierr = MatGetSubMatrices(B,1,&isrowb,&iscolb,scall,&bseq);CHKERRQ(ierr); 4686 *B_seq = bseq[0]; 4687 ierr = PetscFree(bseq);CHKERRQ(ierr); 4688 if (!rowb) { 4689 ierr = ISDestroy(&isrowb);CHKERRQ(ierr); 4690 } else { 4691 *rowb = isrowb; 4692 } 4693 if (!colb) { 4694 ierr = ISDestroy(&iscolb);CHKERRQ(ierr); 4695 } else { 4696 *colb = iscolb; 4697 } 4698 ierr = PetscLogEventEnd(MAT_GetBrowsOfAcols,A,B,0,0);CHKERRQ(ierr); 4699 PetscFunctionReturn(0); 4700 } 4701 4702 #undef __FUNCT__ 4703 #define __FUNCT__ "MatGetBrowsOfAoCols_MPIAIJ" 4704 /* 4705 MatGetBrowsOfAoCols_MPIAIJ - Creates a SeqAIJ matrix by taking rows of B that equal to nonzero columns 4706 of the OFF-DIAGONAL portion of local A 4707 4708 Collective on Mat 4709 4710 Input Parameters: 4711 + A,B - the matrices in mpiaij format 4712 - scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX 4713 4714 Output Parameter: 4715 + startsj_s - starting point in B's sending j-arrays, saved for MAT_REUSE (or NULL) 4716 . startsj_r - starting point in B's receiving j-arrays, saved for MAT_REUSE (or NULL) 4717 . bufa_ptr - array for sending matrix values, saved for MAT_REUSE (or NULL) 4718 - B_oth - the sequential matrix generated with size aBn=a->B->cmap->n by B->cmap->N 4719 4720 Level: developer 4721 4722 */ 4723 PetscErrorCode MatGetBrowsOfAoCols_MPIAIJ(Mat A,Mat B,MatReuse scall,PetscInt **startsj_s,PetscInt **startsj_r,MatScalar **bufa_ptr,Mat *B_oth) 4724 { 4725 VecScatter_MPI_General *gen_to,*gen_from; 4726 PetscErrorCode ierr; 4727 Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data; 4728 Mat_SeqAIJ *b_oth; 4729 VecScatter ctx =a->Mvctx; 4730 MPI_Comm comm; 4731 PetscMPIInt *rprocs,*sprocs,tag=((PetscObject)ctx)->tag,rank; 4732 PetscInt *rowlen,*bufj,*bufJ,ncols,aBn=a->B->cmap->n,row,*b_othi,*b_othj; 4733 PetscScalar *rvalues,*svalues; 4734 MatScalar *b_otha,*bufa,*bufA; 4735 PetscInt i,j,k,l,ll,nrecvs,nsends,nrows,*srow,*rstarts,*rstartsj = 0,*sstarts,*sstartsj,len; 4736 MPI_Request *rwaits = NULL,*swaits = NULL; 4737 MPI_Status *sstatus,rstatus; 4738 PetscMPIInt jj,size; 4739 PetscInt *cols,sbs,rbs; 4740 PetscScalar *vals; 4741 4742 PetscFunctionBegin; 4743 ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 4744 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 4745 4746 if (A->cmap->rstart != B->rmap->rstart || A->cmap->rend != B->rmap->rend) { 4747 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); 4748 } 4749 ierr = PetscLogEventBegin(MAT_GetBrowsOfAocols,A,B,0,0);CHKERRQ(ierr); 4750 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 4751 4752 gen_to = (VecScatter_MPI_General*)ctx->todata; 4753 gen_from = (VecScatter_MPI_General*)ctx->fromdata; 4754 rvalues = gen_from->values; /* holds the length of receiving row */ 4755 svalues = gen_to->values; /* holds the length of sending row */ 4756 nrecvs = gen_from->n; 4757 nsends = gen_to->n; 4758 4759 ierr = PetscMalloc2(nrecvs,&rwaits,nsends,&swaits);CHKERRQ(ierr); 4760 srow = gen_to->indices; /* local row index to be sent */ 4761 sstarts = gen_to->starts; 4762 sprocs = gen_to->procs; 4763 sstatus = gen_to->sstatus; 4764 sbs = gen_to->bs; 4765 rstarts = gen_from->starts; 4766 rprocs = gen_from->procs; 4767 rbs = gen_from->bs; 4768 4769 if (!startsj_s || !bufa_ptr) scall = MAT_INITIAL_MATRIX; 4770 if (scall == MAT_INITIAL_MATRIX) { 4771 /* i-array */ 4772 /*---------*/ 4773 /* post receives */ 4774 for (i=0; i<nrecvs; i++) { 4775 rowlen = (PetscInt*)rvalues + rstarts[i]*rbs; 4776 nrows = (rstarts[i+1]-rstarts[i])*rbs; /* num of indices to be received */ 4777 ierr = MPI_Irecv(rowlen,nrows,MPIU_INT,rprocs[i],tag,comm,rwaits+i);CHKERRQ(ierr); 4778 } 4779 4780 /* pack the outgoing message */ 4781 ierr = PetscMalloc2(nsends+1,&sstartsj,nrecvs+1,&rstartsj);CHKERRQ(ierr); 4782 4783 sstartsj[0] = 0; 4784 rstartsj[0] = 0; 4785 len = 0; /* total length of j or a array to be sent */ 4786 k = 0; 4787 for (i=0; i<nsends; i++) { 4788 rowlen = (PetscInt*)svalues + sstarts[i]*sbs; 4789 nrows = sstarts[i+1]-sstarts[i]; /* num of block rows */ 4790 for (j=0; j<nrows; j++) { 4791 row = srow[k] + B->rmap->range[rank]; /* global row idx */ 4792 for (l=0; l<sbs; l++) { 4793 ierr = MatGetRow_MPIAIJ(B,row+l,&ncols,NULL,NULL);CHKERRQ(ierr); /* rowlength */ 4794 4795 rowlen[j*sbs+l] = ncols; 4796 4797 len += ncols; 4798 ierr = MatRestoreRow_MPIAIJ(B,row+l,&ncols,NULL,NULL);CHKERRQ(ierr); 4799 } 4800 k++; 4801 } 4802 ierr = MPI_Isend(rowlen,nrows*sbs,MPIU_INT,sprocs[i],tag,comm,swaits+i);CHKERRQ(ierr); 4803 4804 sstartsj[i+1] = len; /* starting point of (i+1)-th outgoing msg in bufj and bufa */ 4805 } 4806 /* recvs and sends of i-array are completed */ 4807 i = nrecvs; 4808 while (i--) { 4809 ierr = MPI_Waitany(nrecvs,rwaits,&jj,&rstatus);CHKERRQ(ierr); 4810 } 4811 if (nsends) {ierr = MPI_Waitall(nsends,swaits,sstatus);CHKERRQ(ierr);} 4812 4813 /* allocate buffers for sending j and a arrays */ 4814 ierr = PetscMalloc1(len+1,&bufj);CHKERRQ(ierr); 4815 ierr = PetscMalloc1(len+1,&bufa);CHKERRQ(ierr); 4816 4817 /* create i-array of B_oth */ 4818 ierr = PetscMalloc1(aBn+2,&b_othi);CHKERRQ(ierr); 4819 4820 b_othi[0] = 0; 4821 len = 0; /* total length of j or a array to be received */ 4822 k = 0; 4823 for (i=0; i<nrecvs; i++) { 4824 rowlen = (PetscInt*)rvalues + rstarts[i]*rbs; 4825 nrows = rbs*(rstarts[i+1]-rstarts[i]); /* num of rows to be recieved */ 4826 for (j=0; j<nrows; j++) { 4827 b_othi[k+1] = b_othi[k] + rowlen[j]; 4828 len += rowlen[j]; k++; 4829 } 4830 rstartsj[i+1] = len; /* starting point of (i+1)-th incoming msg in bufj and bufa */ 4831 } 4832 4833 /* allocate space for j and a arrrays of B_oth */ 4834 ierr = PetscMalloc1(b_othi[aBn]+1,&b_othj);CHKERRQ(ierr); 4835 ierr = PetscMalloc1(b_othi[aBn]+1,&b_otha);CHKERRQ(ierr); 4836 4837 /* j-array */ 4838 /*---------*/ 4839 /* post receives of j-array */ 4840 for (i=0; i<nrecvs; i++) { 4841 nrows = rstartsj[i+1]-rstartsj[i]; /* length of the msg received */ 4842 ierr = MPI_Irecv(b_othj+rstartsj[i],nrows,MPIU_INT,rprocs[i],tag,comm,rwaits+i);CHKERRQ(ierr); 4843 } 4844 4845 /* pack the outgoing message j-array */ 4846 k = 0; 4847 for (i=0; i<nsends; i++) { 4848 nrows = sstarts[i+1]-sstarts[i]; /* num of block rows */ 4849 bufJ = bufj+sstartsj[i]; 4850 for (j=0; j<nrows; j++) { 4851 row = srow[k++] + B->rmap->range[rank]; /* global row idx */ 4852 for (ll=0; ll<sbs; ll++) { 4853 ierr = MatGetRow_MPIAIJ(B,row+ll,&ncols,&cols,NULL);CHKERRQ(ierr); 4854 for (l=0; l<ncols; l++) { 4855 *bufJ++ = cols[l]; 4856 } 4857 ierr = MatRestoreRow_MPIAIJ(B,row+ll,&ncols,&cols,NULL);CHKERRQ(ierr); 4858 } 4859 } 4860 ierr = MPI_Isend(bufj+sstartsj[i],sstartsj[i+1]-sstartsj[i],MPIU_INT,sprocs[i],tag,comm,swaits+i);CHKERRQ(ierr); 4861 } 4862 4863 /* recvs and sends of j-array are completed */ 4864 i = nrecvs; 4865 while (i--) { 4866 ierr = MPI_Waitany(nrecvs,rwaits,&jj,&rstatus);CHKERRQ(ierr); 4867 } 4868 if (nsends) {ierr = MPI_Waitall(nsends,swaits,sstatus);CHKERRQ(ierr);} 4869 } else if (scall == MAT_REUSE_MATRIX) { 4870 sstartsj = *startsj_s; 4871 rstartsj = *startsj_r; 4872 bufa = *bufa_ptr; 4873 b_oth = (Mat_SeqAIJ*)(*B_oth)->data; 4874 b_otha = b_oth->a; 4875 } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE, "Matrix P does not posses an object container"); 4876 4877 /* a-array */ 4878 /*---------*/ 4879 /* post receives of a-array */ 4880 for (i=0; i<nrecvs; i++) { 4881 nrows = rstartsj[i+1]-rstartsj[i]; /* length of the msg received */ 4882 ierr = MPI_Irecv(b_otha+rstartsj[i],nrows,MPIU_SCALAR,rprocs[i],tag,comm,rwaits+i);CHKERRQ(ierr); 4883 } 4884 4885 /* pack the outgoing message a-array */ 4886 k = 0; 4887 for (i=0; i<nsends; i++) { 4888 nrows = sstarts[i+1]-sstarts[i]; /* num of block rows */ 4889 bufA = bufa+sstartsj[i]; 4890 for (j=0; j<nrows; j++) { 4891 row = srow[k++] + B->rmap->range[rank]; /* global row idx */ 4892 for (ll=0; ll<sbs; ll++) { 4893 ierr = MatGetRow_MPIAIJ(B,row+ll,&ncols,NULL,&vals);CHKERRQ(ierr); 4894 for (l=0; l<ncols; l++) { 4895 *bufA++ = vals[l]; 4896 } 4897 ierr = MatRestoreRow_MPIAIJ(B,row+ll,&ncols,NULL,&vals);CHKERRQ(ierr); 4898 } 4899 } 4900 ierr = MPI_Isend(bufa+sstartsj[i],sstartsj[i+1]-sstartsj[i],MPIU_SCALAR,sprocs[i],tag,comm,swaits+i);CHKERRQ(ierr); 4901 } 4902 /* recvs and sends of a-array are completed */ 4903 i = nrecvs; 4904 while (i--) { 4905 ierr = MPI_Waitany(nrecvs,rwaits,&jj,&rstatus);CHKERRQ(ierr); 4906 } 4907 if (nsends) {ierr = MPI_Waitall(nsends,swaits,sstatus);CHKERRQ(ierr);} 4908 ierr = PetscFree2(rwaits,swaits);CHKERRQ(ierr); 4909 4910 if (scall == MAT_INITIAL_MATRIX) { 4911 /* put together the new matrix */ 4912 ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,aBn,B->cmap->N,b_othi,b_othj,b_otha,B_oth);CHKERRQ(ierr); 4913 4914 /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */ 4915 /* Since these are PETSc arrays, change flags to free them as necessary. */ 4916 b_oth = (Mat_SeqAIJ*)(*B_oth)->data; 4917 b_oth->free_a = PETSC_TRUE; 4918 b_oth->free_ij = PETSC_TRUE; 4919 b_oth->nonew = 0; 4920 4921 ierr = PetscFree(bufj);CHKERRQ(ierr); 4922 if (!startsj_s || !bufa_ptr) { 4923 ierr = PetscFree2(sstartsj,rstartsj);CHKERRQ(ierr); 4924 ierr = PetscFree(bufa_ptr);CHKERRQ(ierr); 4925 } else { 4926 *startsj_s = sstartsj; 4927 *startsj_r = rstartsj; 4928 *bufa_ptr = bufa; 4929 } 4930 } 4931 ierr = PetscLogEventEnd(MAT_GetBrowsOfAocols,A,B,0,0);CHKERRQ(ierr); 4932 PetscFunctionReturn(0); 4933 } 4934 4935 #undef __FUNCT__ 4936 #define __FUNCT__ "MatGetCommunicationStructs" 4937 /*@C 4938 MatGetCommunicationStructs - Provides access to the communication structures used in matrix-vector multiplication. 4939 4940 Not Collective 4941 4942 Input Parameters: 4943 . A - The matrix in mpiaij format 4944 4945 Output Parameter: 4946 + lvec - The local vector holding off-process values from the argument to a matrix-vector product 4947 . colmap - A map from global column index to local index into lvec 4948 - multScatter - A scatter from the argument of a matrix-vector product to lvec 4949 4950 Level: developer 4951 4952 @*/ 4953 #if defined(PETSC_USE_CTABLE) 4954 PetscErrorCode MatGetCommunicationStructs(Mat A, Vec *lvec, PetscTable *colmap, VecScatter *multScatter) 4955 #else 4956 PetscErrorCode MatGetCommunicationStructs(Mat A, Vec *lvec, PetscInt *colmap[], VecScatter *multScatter) 4957 #endif 4958 { 4959 Mat_MPIAIJ *a; 4960 4961 PetscFunctionBegin; 4962 PetscValidHeaderSpecific(A, MAT_CLASSID, 1); 4963 PetscValidPointer(lvec, 2); 4964 PetscValidPointer(colmap, 3); 4965 PetscValidPointer(multScatter, 4); 4966 a = (Mat_MPIAIJ*) A->data; 4967 if (lvec) *lvec = a->lvec; 4968 if (colmap) *colmap = a->colmap; 4969 if (multScatter) *multScatter = a->Mvctx; 4970 PetscFunctionReturn(0); 4971 } 4972 4973 PETSC_EXTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJCRL(Mat,MatType,MatReuse,Mat*); 4974 PETSC_EXTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJPERM(Mat,MatType,MatReuse,Mat*); 4975 PETSC_EXTERN PetscErrorCode MatConvert_MPIAIJ_MPISBAIJ(Mat,MatType,MatReuse,Mat*); 4976 #if defined(PETSC_HAVE_ELEMENTAL) 4977 PETSC_EXTERN PetscErrorCode MatConvert_MPIAIJ_Elemental(Mat,MatType,MatReuse,Mat*); 4978 #endif 4979 4980 #undef __FUNCT__ 4981 #define __FUNCT__ "MatMatMultNumeric_MPIDense_MPIAIJ" 4982 /* 4983 Computes (B'*A')' since computing B*A directly is untenable 4984 4985 n p p 4986 ( ) ( ) ( ) 4987 m ( A ) * n ( B ) = m ( C ) 4988 ( ) ( ) ( ) 4989 4990 */ 4991 PetscErrorCode MatMatMultNumeric_MPIDense_MPIAIJ(Mat A,Mat B,Mat C) 4992 { 4993 PetscErrorCode ierr; 4994 Mat At,Bt,Ct; 4995 4996 PetscFunctionBegin; 4997 ierr = MatTranspose(A,MAT_INITIAL_MATRIX,&At);CHKERRQ(ierr); 4998 ierr = MatTranspose(B,MAT_INITIAL_MATRIX,&Bt);CHKERRQ(ierr); 4999 ierr = MatMatMult(Bt,At,MAT_INITIAL_MATRIX,1.0,&Ct);CHKERRQ(ierr); 5000 ierr = MatDestroy(&At);CHKERRQ(ierr); 5001 ierr = MatDestroy(&Bt);CHKERRQ(ierr); 5002 ierr = MatTranspose(Ct,MAT_REUSE_MATRIX,&C);CHKERRQ(ierr); 5003 ierr = MatDestroy(&Ct);CHKERRQ(ierr); 5004 PetscFunctionReturn(0); 5005 } 5006 5007 #undef __FUNCT__ 5008 #define __FUNCT__ "MatMatMultSymbolic_MPIDense_MPIAIJ" 5009 PetscErrorCode MatMatMultSymbolic_MPIDense_MPIAIJ(Mat A,Mat B,PetscReal fill,Mat *C) 5010 { 5011 PetscErrorCode ierr; 5012 PetscInt m=A->rmap->n,n=B->cmap->n; 5013 Mat Cmat; 5014 5015 PetscFunctionBegin; 5016 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); 5017 ierr = MatCreate(PetscObjectComm((PetscObject)A),&Cmat);CHKERRQ(ierr); 5018 ierr = MatSetSizes(Cmat,m,n,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 5019 ierr = MatSetBlockSizesFromMats(Cmat,A,B);CHKERRQ(ierr); 5020 ierr = MatSetType(Cmat,MATMPIDENSE);CHKERRQ(ierr); 5021 ierr = MatMPIDenseSetPreallocation(Cmat,NULL);CHKERRQ(ierr); 5022 ierr = MatAssemblyBegin(Cmat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5023 ierr = MatAssemblyEnd(Cmat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5024 5025 Cmat->ops->matmultnumeric = MatMatMultNumeric_MPIDense_MPIAIJ; 5026 5027 *C = Cmat; 5028 PetscFunctionReturn(0); 5029 } 5030 5031 /* ----------------------------------------------------------------*/ 5032 #undef __FUNCT__ 5033 #define __FUNCT__ "MatMatMult_MPIDense_MPIAIJ" 5034 PetscErrorCode MatMatMult_MPIDense_MPIAIJ(Mat A,Mat B,MatReuse scall,PetscReal fill,Mat *C) 5035 { 5036 PetscErrorCode ierr; 5037 5038 PetscFunctionBegin; 5039 if (scall == MAT_INITIAL_MATRIX) { 5040 ierr = PetscLogEventBegin(MAT_MatMultSymbolic,A,B,0,0);CHKERRQ(ierr); 5041 ierr = MatMatMultSymbolic_MPIDense_MPIAIJ(A,B,fill,C);CHKERRQ(ierr); 5042 ierr = PetscLogEventEnd(MAT_MatMultSymbolic,A,B,0,0);CHKERRQ(ierr); 5043 } 5044 ierr = PetscLogEventBegin(MAT_MatMultNumeric,A,B,0,0);CHKERRQ(ierr); 5045 ierr = MatMatMultNumeric_MPIDense_MPIAIJ(A,B,*C);CHKERRQ(ierr); 5046 ierr = PetscLogEventEnd(MAT_MatMultNumeric,A,B,0,0);CHKERRQ(ierr); 5047 PetscFunctionReturn(0); 5048 } 5049 5050 /*MC 5051 MATMPIAIJ - MATMPIAIJ = "mpiaij" - A matrix type to be used for parallel sparse matrices. 5052 5053 Options Database Keys: 5054 . -mat_type mpiaij - sets the matrix type to "mpiaij" during a call to MatSetFromOptions() 5055 5056 Level: beginner 5057 5058 .seealso: MatCreateAIJ() 5059 M*/ 5060 5061 #undef __FUNCT__ 5062 #define __FUNCT__ "MatCreate_MPIAIJ" 5063 PETSC_EXTERN PetscErrorCode MatCreate_MPIAIJ(Mat B) 5064 { 5065 Mat_MPIAIJ *b; 5066 PetscErrorCode ierr; 5067 PetscMPIInt size; 5068 5069 PetscFunctionBegin; 5070 ierr = MPI_Comm_size(PetscObjectComm((PetscObject)B),&size);CHKERRQ(ierr); 5071 5072 ierr = PetscNewLog(B,&b);CHKERRQ(ierr); 5073 B->data = (void*)b; 5074 ierr = PetscMemcpy(B->ops,&MatOps_Values,sizeof(struct _MatOps));CHKERRQ(ierr); 5075 B->assembled = PETSC_FALSE; 5076 B->insertmode = NOT_SET_VALUES; 5077 b->size = size; 5078 5079 ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)B),&b->rank);CHKERRQ(ierr); 5080 5081 /* build cache for off array entries formed */ 5082 ierr = MatStashCreate_Private(PetscObjectComm((PetscObject)B),1,&B->stash);CHKERRQ(ierr); 5083 5084 b->donotstash = PETSC_FALSE; 5085 b->colmap = 0; 5086 b->garray = 0; 5087 b->roworiented = PETSC_TRUE; 5088 5089 /* stuff used for matrix vector multiply */ 5090 b->lvec = NULL; 5091 b->Mvctx = NULL; 5092 5093 /* stuff for MatGetRow() */ 5094 b->rowindices = 0; 5095 b->rowvalues = 0; 5096 b->getrowactive = PETSC_FALSE; 5097 5098 /* flexible pointer used in CUSP/CUSPARSE classes */ 5099 b->spptr = NULL; 5100 5101 ierr = PetscObjectComposeFunction((PetscObject)B,"MatStoreValues_C",MatStoreValues_MPIAIJ);CHKERRQ(ierr); 5102 ierr = PetscObjectComposeFunction((PetscObject)B,"MatRetrieveValues_C",MatRetrieveValues_MPIAIJ);CHKERRQ(ierr); 5103 ierr = PetscObjectComposeFunction((PetscObject)B,"MatGetDiagonalBlock_C",MatGetDiagonalBlock_MPIAIJ);CHKERRQ(ierr); 5104 ierr = PetscObjectComposeFunction((PetscObject)B,"MatIsTranspose_C",MatIsTranspose_MPIAIJ);CHKERRQ(ierr); 5105 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMPIAIJSetPreallocation_C",MatMPIAIJSetPreallocation_MPIAIJ);CHKERRQ(ierr); 5106 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMPIAIJSetPreallocationCSR_C",MatMPIAIJSetPreallocationCSR_MPIAIJ);CHKERRQ(ierr); 5107 ierr = PetscObjectComposeFunction((PetscObject)B,"MatDiagonalScaleLocal_C",MatDiagonalScaleLocal_MPIAIJ);CHKERRQ(ierr); 5108 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_mpiaijperm_C",MatConvert_MPIAIJ_MPIAIJPERM);CHKERRQ(ierr); 5109 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_mpiaijcrl_C",MatConvert_MPIAIJ_MPIAIJCRL);CHKERRQ(ierr); 5110 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_mpisbaij_C",MatConvert_MPIAIJ_MPISBAIJ);CHKERRQ(ierr); 5111 #if defined(PETSC_HAVE_ELEMENTAL) 5112 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_elemental_C",MatConvert_MPIAIJ_Elemental);CHKERRQ(ierr); 5113 #endif 5114 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMatMult_mpidense_mpiaij_C",MatMatMult_MPIDense_MPIAIJ);CHKERRQ(ierr); 5115 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMatMultSymbolic_mpidense_mpiaij_C",MatMatMultSymbolic_MPIDense_MPIAIJ);CHKERRQ(ierr); 5116 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMatMultNumeric_mpidense_mpiaij_C",MatMatMultNumeric_MPIDense_MPIAIJ);CHKERRQ(ierr); 5117 ierr = PetscObjectChangeTypeName((PetscObject)B,MATMPIAIJ);CHKERRQ(ierr); 5118 PetscFunctionReturn(0); 5119 } 5120 5121 #undef __FUNCT__ 5122 #define __FUNCT__ "MatCreateMPIAIJWithSplitArrays" 5123 /*@C 5124 MatCreateMPIAIJWithSplitArrays - creates a MPI AIJ matrix using arrays that contain the "diagonal" 5125 and "off-diagonal" part of the matrix in CSR format. 5126 5127 Collective on MPI_Comm 5128 5129 Input Parameters: 5130 + comm - MPI communicator 5131 . m - number of local rows (Cannot be PETSC_DECIDE) 5132 . n - This value should be the same as the local size used in creating the 5133 x vector for the matrix-vector product y = Ax. (or PETSC_DECIDE to have 5134 calculated if N is given) For square matrices n is almost always m. 5135 . M - number of global rows (or PETSC_DETERMINE to have calculated if m is given) 5136 . N - number of global columns (or PETSC_DETERMINE to have calculated if n is given) 5137 . i - row indices for "diagonal" portion of matrix 5138 . j - column indices 5139 . a - matrix values 5140 . oi - row indices for "off-diagonal" portion of matrix 5141 . oj - column indices 5142 - oa - matrix values 5143 5144 Output Parameter: 5145 . mat - the matrix 5146 5147 Level: advanced 5148 5149 Notes: 5150 The i, j, and a arrays ARE NOT copied by this routine into the internal format used by PETSc. The user 5151 must free the arrays once the matrix has been destroyed and not before. 5152 5153 The i and j indices are 0 based 5154 5155 See MatCreateAIJ() for the definition of "diagonal" and "off-diagonal" portion of the matrix 5156 5157 This sets local rows and cannot be used to set off-processor values. 5158 5159 Use of this routine is discouraged because it is inflexible and cumbersome to use. It is extremely rare that a 5160 legacy application natively assembles into exactly this split format. The code to do so is nontrivial and does 5161 not easily support in-place reassembly. It is recommended to use MatSetValues() (or a variant thereof) because 5162 the resulting assembly is easier to implement, will work with any matrix format, and the user does not have to 5163 keep track of the underlying array. Use MatSetOption(A,MAT_IGNORE_OFF_PROC_ENTRIES,PETSC_TRUE) to disable all 5164 communication if it is known that only local entries will be set. 5165 5166 .keywords: matrix, aij, compressed row, sparse, parallel 5167 5168 .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatMPIAIJSetPreallocationCSR(), 5169 MPIAIJ, MatCreateAIJ(), MatCreateMPIAIJWithArrays() 5170 @*/ 5171 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) 5172 { 5173 PetscErrorCode ierr; 5174 Mat_MPIAIJ *maij; 5175 5176 PetscFunctionBegin; 5177 if (m < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"local number of rows (m) cannot be PETSC_DECIDE, or negative"); 5178 if (i[0]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"i (row indices) must start with 0"); 5179 if (oi[0]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"oi (row indices) must start with 0"); 5180 ierr = MatCreate(comm,mat);CHKERRQ(ierr); 5181 ierr = MatSetSizes(*mat,m,n,M,N);CHKERRQ(ierr); 5182 ierr = MatSetType(*mat,MATMPIAIJ);CHKERRQ(ierr); 5183 maij = (Mat_MPIAIJ*) (*mat)->data; 5184 5185 (*mat)->preallocated = PETSC_TRUE; 5186 5187 ierr = PetscLayoutSetUp((*mat)->rmap);CHKERRQ(ierr); 5188 ierr = PetscLayoutSetUp((*mat)->cmap);CHKERRQ(ierr); 5189 5190 ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,m,n,i,j,a,&maij->A);CHKERRQ(ierr); 5191 ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,m,(*mat)->cmap->N,oi,oj,oa,&maij->B);CHKERRQ(ierr); 5192 5193 ierr = MatAssemblyBegin(maij->A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5194 ierr = MatAssemblyEnd(maij->A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5195 ierr = MatAssemblyBegin(maij->B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5196 ierr = MatAssemblyEnd(maij->B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5197 5198 ierr = MatAssemblyBegin(*mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5199 ierr = MatAssemblyEnd(*mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5200 ierr = MatSetOption(*mat,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE);CHKERRQ(ierr); 5201 PetscFunctionReturn(0); 5202 } 5203 5204 /* 5205 Special version for direct calls from Fortran 5206 */ 5207 #include <petsc/private/fortranimpl.h> 5208 5209 #if defined(PETSC_HAVE_FORTRAN_CAPS) 5210 #define matsetvaluesmpiaij_ MATSETVALUESMPIAIJ 5211 #elif !defined(PETSC_HAVE_FORTRAN_UNDERSCORE) 5212 #define matsetvaluesmpiaij_ matsetvaluesmpiaij 5213 #endif 5214 5215 /* Change these macros so can be used in void function */ 5216 #undef CHKERRQ 5217 #define CHKERRQ(ierr) CHKERRABORT(PETSC_COMM_WORLD,ierr) 5218 #undef SETERRQ2 5219 #define SETERRQ2(comm,ierr,b,c,d) CHKERRABORT(comm,ierr) 5220 #undef SETERRQ3 5221 #define SETERRQ3(comm,ierr,b,c,d,e) CHKERRABORT(comm,ierr) 5222 #undef SETERRQ 5223 #define SETERRQ(c,ierr,b) CHKERRABORT(c,ierr) 5224 5225 #undef __FUNCT__ 5226 #define __FUNCT__ "matsetvaluesmpiaij_" 5227 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) 5228 { 5229 Mat mat = *mmat; 5230 PetscInt m = *mm, n = *mn; 5231 InsertMode addv = *maddv; 5232 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 5233 PetscScalar value; 5234 PetscErrorCode ierr; 5235 5236 MatCheckPreallocated(mat,1); 5237 if (mat->insertmode == NOT_SET_VALUES) mat->insertmode = addv; 5238 5239 #if defined(PETSC_USE_DEBUG) 5240 else if (mat->insertmode != addv) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Cannot mix add values and insert values"); 5241 #endif 5242 { 5243 PetscInt i,j,rstart = mat->rmap->rstart,rend = mat->rmap->rend; 5244 PetscInt cstart = mat->cmap->rstart,cend = mat->cmap->rend,row,col; 5245 PetscBool roworiented = aij->roworiented; 5246 5247 /* Some Variables required in the macro */ 5248 Mat A = aij->A; 5249 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 5250 PetscInt *aimax = a->imax,*ai = a->i,*ailen = a->ilen,*aj = a->j; 5251 MatScalar *aa = a->a; 5252 PetscBool ignorezeroentries = (((a->ignorezeroentries)&&(addv==ADD_VALUES)) ? PETSC_TRUE : PETSC_FALSE); 5253 Mat B = aij->B; 5254 Mat_SeqAIJ *b = (Mat_SeqAIJ*)B->data; 5255 PetscInt *bimax = b->imax,*bi = b->i,*bilen = b->ilen,*bj = b->j,bm = aij->B->rmap->n,am = aij->A->rmap->n; 5256 MatScalar *ba = b->a; 5257 5258 PetscInt *rp1,*rp2,ii,nrow1,nrow2,_i,rmax1,rmax2,N,low1,high1,low2,high2,t,lastcol1,lastcol2; 5259 PetscInt nonew = a->nonew; 5260 MatScalar *ap1,*ap2; 5261 5262 PetscFunctionBegin; 5263 for (i=0; i<m; i++) { 5264 if (im[i] < 0) continue; 5265 #if defined(PETSC_USE_DEBUG) 5266 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); 5267 #endif 5268 if (im[i] >= rstart && im[i] < rend) { 5269 row = im[i] - rstart; 5270 lastcol1 = -1; 5271 rp1 = aj + ai[row]; 5272 ap1 = aa + ai[row]; 5273 rmax1 = aimax[row]; 5274 nrow1 = ailen[row]; 5275 low1 = 0; 5276 high1 = nrow1; 5277 lastcol2 = -1; 5278 rp2 = bj + bi[row]; 5279 ap2 = ba + bi[row]; 5280 rmax2 = bimax[row]; 5281 nrow2 = bilen[row]; 5282 low2 = 0; 5283 high2 = nrow2; 5284 5285 for (j=0; j<n; j++) { 5286 if (roworiented) value = v[i*n+j]; 5287 else value = v[i+j*m]; 5288 if (ignorezeroentries && value == 0.0 && (addv == ADD_VALUES)) continue; 5289 if (in[j] >= cstart && in[j] < cend) { 5290 col = in[j] - cstart; 5291 MatSetValues_SeqAIJ_A_Private(row,col,value,addv,im[i],in[j]); 5292 } else if (in[j] < 0) continue; 5293 #if defined(PETSC_USE_DEBUG) 5294 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); 5295 #endif 5296 else { 5297 if (mat->was_assembled) { 5298 if (!aij->colmap) { 5299 ierr = MatCreateColmap_MPIAIJ_Private(mat);CHKERRQ(ierr); 5300 } 5301 #if defined(PETSC_USE_CTABLE) 5302 ierr = PetscTableFind(aij->colmap,in[j]+1,&col);CHKERRQ(ierr); 5303 col--; 5304 #else 5305 col = aij->colmap[in[j]] - 1; 5306 #endif 5307 if (col < 0 && !((Mat_SeqAIJ*)(aij->A->data))->nonew) { 5308 ierr = MatDisAssemble_MPIAIJ(mat);CHKERRQ(ierr); 5309 col = in[j]; 5310 /* Reinitialize the variables required by MatSetValues_SeqAIJ_B_Private() */ 5311 B = aij->B; 5312 b = (Mat_SeqAIJ*)B->data; 5313 bimax = b->imax; bi = b->i; bilen = b->ilen; bj = b->j; 5314 rp2 = bj + bi[row]; 5315 ap2 = ba + bi[row]; 5316 rmax2 = bimax[row]; 5317 nrow2 = bilen[row]; 5318 low2 = 0; 5319 high2 = nrow2; 5320 bm = aij->B->rmap->n; 5321 ba = b->a; 5322 } 5323 } else col = in[j]; 5324 MatSetValues_SeqAIJ_B_Private(row,col,value,addv,im[i],in[j]); 5325 } 5326 } 5327 } else if (!aij->donotstash) { 5328 if (roworiented) { 5329 ierr = MatStashValuesRow_Private(&mat->stash,im[i],n,in,v+i*n,(PetscBool)(ignorezeroentries && (addv == ADD_VALUES)));CHKERRQ(ierr); 5330 } else { 5331 ierr = MatStashValuesCol_Private(&mat->stash,im[i],n,in,v+i,m,(PetscBool)(ignorezeroentries && (addv == ADD_VALUES)));CHKERRQ(ierr); 5332 } 5333 } 5334 } 5335 } 5336 PetscFunctionReturnVoid(); 5337 } 5338 5339