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