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