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