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