1 2 3 #include <../src/mat/impls/aij/mpi/mpiaij.h> /*I "petscmat.h" I*/ 4 #include <petsc/private/vecimpl.h> 5 #include <petsc/private/vecscatterimpl.h> 6 #include <petsc/private/isimpl.h> 7 #include <petscblaslapack.h> 8 #include <petscsf.h> 9 10 /*MC 11 MATAIJ - MATAIJ = "aij" - A matrix type to be used for sparse matrices. 12 13 This matrix type is identical to MATSEQAIJ when constructed with a single process communicator, 14 and MATMPIAIJ otherwise. As a result, for single process communicators, 15 MatSeqAIJSetPreallocation is supported, and similarly MatMPIAIJSetPreallocation() is supported 16 for communicators controlling multiple processes. It is recommended that you call both of 17 the above preallocation routines for simplicity. 18 19 Options Database Keys: 20 . -mat_type aij - sets the matrix type to "aij" during a call to MatSetFromOptions() 21 22 Developer Notes: 23 Subclasses include MATAIJCUSP, MATAIJCUSPARSE, MATAIJPERM, MATAIJSELL, MATAIJMKL, MATAIJCRL, and also automatically switches over to use inodes when 24 enough exist. 25 26 Level: beginner 27 28 .seealso: MatCreateAIJ(), MatCreateSeqAIJ(), MATSEQAIJ, MATMPIAIJ 29 M*/ 30 31 /*MC 32 MATAIJCRL - MATAIJCRL = "aijcrl" - A matrix type to be used for sparse matrices. 33 34 This matrix type is identical to MATSEQAIJCRL when constructed with a single process communicator, 35 and MATMPIAIJCRL otherwise. As a result, for single process communicators, 36 MatSeqAIJSetPreallocation() is supported, and similarly MatMPIAIJSetPreallocation() is supported 37 for communicators controlling multiple processes. It is recommended that you call both of 38 the above preallocation routines for simplicity. 39 40 Options Database Keys: 41 . -mat_type aijcrl - sets the matrix type to "aijcrl" during a call to MatSetFromOptions() 42 43 Level: beginner 44 45 .seealso: MatCreateMPIAIJCRL,MATSEQAIJCRL,MATMPIAIJCRL, MATSEQAIJCRL, MATMPIAIJCRL 46 M*/ 47 48 PetscErrorCode MatSetBlockSizes_MPIAIJ(Mat M, PetscInt rbs, PetscInt cbs) 49 { 50 PetscErrorCode ierr; 51 Mat_MPIAIJ *mat = (Mat_MPIAIJ*)M->data; 52 53 PetscFunctionBegin; 54 if (mat->A) { 55 ierr = MatSetBlockSizes(mat->A,rbs,cbs);CHKERRQ(ierr); 56 ierr = MatSetBlockSizes(mat->B,rbs,1);CHKERRQ(ierr); 57 } 58 PetscFunctionReturn(0); 59 } 60 61 PetscErrorCode MatFindNonzeroRows_MPIAIJ(Mat M,IS *keptrows) 62 { 63 PetscErrorCode ierr; 64 Mat_MPIAIJ *mat = (Mat_MPIAIJ*)M->data; 65 Mat_SeqAIJ *a = (Mat_SeqAIJ*)mat->A->data; 66 Mat_SeqAIJ *b = (Mat_SeqAIJ*)mat->B->data; 67 const PetscInt *ia,*ib; 68 const MatScalar *aa,*bb; 69 PetscInt na,nb,i,j,*rows,cnt=0,n0rows; 70 PetscInt m = M->rmap->n,rstart = M->rmap->rstart; 71 72 PetscFunctionBegin; 73 *keptrows = 0; 74 ia = a->i; 75 ib = b->i; 76 for (i=0; i<m; i++) { 77 na = ia[i+1] - ia[i]; 78 nb = ib[i+1] - ib[i]; 79 if (!na && !nb) { 80 cnt++; 81 goto ok1; 82 } 83 aa = a->a + ia[i]; 84 for (j=0; j<na; j++) { 85 if (aa[j] != 0.0) goto ok1; 86 } 87 bb = b->a + ib[i]; 88 for (j=0; j <nb; j++) { 89 if (bb[j] != 0.0) goto ok1; 90 } 91 cnt++; 92 ok1:; 93 } 94 ierr = MPIU_Allreduce(&cnt,&n0rows,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)M));CHKERRQ(ierr); 95 if (!n0rows) PetscFunctionReturn(0); 96 ierr = PetscMalloc1(M->rmap->n-cnt,&rows);CHKERRQ(ierr); 97 cnt = 0; 98 for (i=0; i<m; i++) { 99 na = ia[i+1] - ia[i]; 100 nb = ib[i+1] - ib[i]; 101 if (!na && !nb) continue; 102 aa = a->a + ia[i]; 103 for (j=0; j<na;j++) { 104 if (aa[j] != 0.0) { 105 rows[cnt++] = rstart + i; 106 goto ok2; 107 } 108 } 109 bb = b->a + ib[i]; 110 for (j=0; j<nb; j++) { 111 if (bb[j] != 0.0) { 112 rows[cnt++] = rstart + i; 113 goto ok2; 114 } 115 } 116 ok2:; 117 } 118 ierr = ISCreateGeneral(PetscObjectComm((PetscObject)M),cnt,rows,PETSC_OWN_POINTER,keptrows);CHKERRQ(ierr); 119 PetscFunctionReturn(0); 120 } 121 122 PetscErrorCode MatDiagonalSet_MPIAIJ(Mat Y,Vec D,InsertMode is) 123 { 124 PetscErrorCode ierr; 125 Mat_MPIAIJ *aij = (Mat_MPIAIJ*) Y->data; 126 PetscBool cong; 127 128 PetscFunctionBegin; 129 ierr = MatHasCongruentLayouts(Y,&cong);CHKERRQ(ierr); 130 if (Y->assembled && cong) { 131 ierr = MatDiagonalSet(aij->A,D,is);CHKERRQ(ierr); 132 } else { 133 ierr = MatDiagonalSet_Default(Y,D,is);CHKERRQ(ierr); 134 } 135 PetscFunctionReturn(0); 136 } 137 138 PetscErrorCode MatFindZeroDiagonals_MPIAIJ(Mat M,IS *zrows) 139 { 140 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)M->data; 141 PetscErrorCode ierr; 142 PetscInt i,rstart,nrows,*rows; 143 144 PetscFunctionBegin; 145 *zrows = NULL; 146 ierr = MatFindZeroDiagonals_SeqAIJ_Private(aij->A,&nrows,&rows);CHKERRQ(ierr); 147 ierr = MatGetOwnershipRange(M,&rstart,NULL);CHKERRQ(ierr); 148 for (i=0; i<nrows; i++) rows[i] += rstart; 149 ierr = ISCreateGeneral(PetscObjectComm((PetscObject)M),nrows,rows,PETSC_OWN_POINTER,zrows);CHKERRQ(ierr); 150 PetscFunctionReturn(0); 151 } 152 153 PetscErrorCode MatGetColumnNorms_MPIAIJ(Mat A,NormType type,PetscReal *norms) 154 { 155 PetscErrorCode ierr; 156 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)A->data; 157 PetscInt i,n,*garray = aij->garray; 158 Mat_SeqAIJ *a_aij = (Mat_SeqAIJ*) aij->A->data; 159 Mat_SeqAIJ *b_aij = (Mat_SeqAIJ*) aij->B->data; 160 PetscReal *work; 161 162 PetscFunctionBegin; 163 ierr = MatGetSize(A,NULL,&n);CHKERRQ(ierr); 164 ierr = PetscCalloc1(n,&work);CHKERRQ(ierr); 165 if (type == NORM_2) { 166 for (i=0; i<a_aij->i[aij->A->rmap->n]; i++) { 167 work[A->cmap->rstart + a_aij->j[i]] += PetscAbsScalar(a_aij->a[i]*a_aij->a[i]); 168 } 169 for (i=0; i<b_aij->i[aij->B->rmap->n]; i++) { 170 work[garray[b_aij->j[i]]] += PetscAbsScalar(b_aij->a[i]*b_aij->a[i]); 171 } 172 } else if (type == NORM_1) { 173 for (i=0; i<a_aij->i[aij->A->rmap->n]; i++) { 174 work[A->cmap->rstart + a_aij->j[i]] += PetscAbsScalar(a_aij->a[i]); 175 } 176 for (i=0; i<b_aij->i[aij->B->rmap->n]; i++) { 177 work[garray[b_aij->j[i]]] += PetscAbsScalar(b_aij->a[i]); 178 } 179 } else if (type == NORM_INFINITY) { 180 for (i=0; i<a_aij->i[aij->A->rmap->n]; i++) { 181 work[A->cmap->rstart + a_aij->j[i]] = PetscMax(PetscAbsScalar(a_aij->a[i]), work[A->cmap->rstart + a_aij->j[i]]); 182 } 183 for (i=0; i<b_aij->i[aij->B->rmap->n]; i++) { 184 work[garray[b_aij->j[i]]] = PetscMax(PetscAbsScalar(b_aij->a[i]),work[garray[b_aij->j[i]]]); 185 } 186 187 } else SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_WRONG,"Unknown NormType"); 188 if (type == NORM_INFINITY) { 189 ierr = MPIU_Allreduce(work,norms,n,MPIU_REAL,MPIU_MAX,PetscObjectComm((PetscObject)A));CHKERRQ(ierr); 190 } else { 191 ierr = MPIU_Allreduce(work,norms,n,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)A));CHKERRQ(ierr); 192 } 193 ierr = PetscFree(work);CHKERRQ(ierr); 194 if (type == NORM_2) { 195 for (i=0; i<n; i++) norms[i] = PetscSqrtReal(norms[i]); 196 } 197 PetscFunctionReturn(0); 198 } 199 200 PetscErrorCode MatFindOffBlockDiagonalEntries_MPIAIJ(Mat A,IS *is) 201 { 202 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 203 IS sis,gis; 204 PetscErrorCode ierr; 205 const PetscInt *isis,*igis; 206 PetscInt n,*iis,nsis,ngis,rstart,i; 207 208 PetscFunctionBegin; 209 ierr = MatFindOffBlockDiagonalEntries(a->A,&sis);CHKERRQ(ierr); 210 ierr = MatFindNonzeroRows(a->B,&gis);CHKERRQ(ierr); 211 ierr = ISGetSize(gis,&ngis);CHKERRQ(ierr); 212 ierr = ISGetSize(sis,&nsis);CHKERRQ(ierr); 213 ierr = ISGetIndices(sis,&isis);CHKERRQ(ierr); 214 ierr = ISGetIndices(gis,&igis);CHKERRQ(ierr); 215 216 ierr = PetscMalloc1(ngis+nsis,&iis);CHKERRQ(ierr); 217 ierr = PetscArraycpy(iis,igis,ngis);CHKERRQ(ierr); 218 ierr = PetscArraycpy(iis+ngis,isis,nsis);CHKERRQ(ierr); 219 n = ngis + nsis; 220 ierr = PetscSortRemoveDupsInt(&n,iis);CHKERRQ(ierr); 221 ierr = MatGetOwnershipRange(A,&rstart,NULL);CHKERRQ(ierr); 222 for (i=0; i<n; i++) iis[i] += rstart; 223 ierr = ISCreateGeneral(PetscObjectComm((PetscObject)A),n,iis,PETSC_OWN_POINTER,is);CHKERRQ(ierr); 224 225 ierr = ISRestoreIndices(sis,&isis);CHKERRQ(ierr); 226 ierr = ISRestoreIndices(gis,&igis);CHKERRQ(ierr); 227 ierr = ISDestroy(&sis);CHKERRQ(ierr); 228 ierr = ISDestroy(&gis);CHKERRQ(ierr); 229 PetscFunctionReturn(0); 230 } 231 232 /* 233 Distributes a SeqAIJ matrix across a set of processes. Code stolen from 234 MatLoad_MPIAIJ(). Horrible lack of reuse. Should be a routine for each matrix type. 235 236 Only for square matrices 237 238 Used by a preconditioner, hence PETSC_EXTERN 239 */ 240 PETSC_EXTERN PetscErrorCode MatDistribute_MPIAIJ(MPI_Comm comm,Mat gmat,PetscInt m,MatReuse reuse,Mat *inmat) 241 { 242 PetscMPIInt rank,size; 243 PetscInt *rowners,*dlens,*olens,i,rstart,rend,j,jj,nz = 0,*gmataj,cnt,row,*ld,bses[2]; 244 PetscErrorCode ierr; 245 Mat mat; 246 Mat_SeqAIJ *gmata; 247 PetscMPIInt tag; 248 MPI_Status status; 249 PetscBool aij; 250 MatScalar *gmataa,*ao,*ad,*gmataarestore=0; 251 252 PetscFunctionBegin; 253 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 254 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 255 if (!rank) { 256 ierr = PetscObjectTypeCompare((PetscObject)gmat,MATSEQAIJ,&aij);CHKERRQ(ierr); 257 if (!aij) SETERRQ1(PetscObjectComm((PetscObject)gmat),PETSC_ERR_SUP,"Currently no support for input matrix of type %s\n",((PetscObject)gmat)->type_name); 258 } 259 if (reuse == MAT_INITIAL_MATRIX) { 260 ierr = MatCreate(comm,&mat);CHKERRQ(ierr); 261 ierr = MatSetSizes(mat,m,m,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 262 ierr = MatGetBlockSizes(gmat,&bses[0],&bses[1]);CHKERRQ(ierr); 263 ierr = MPI_Bcast(bses,2,MPIU_INT,0,comm);CHKERRQ(ierr); 264 ierr = MatSetBlockSizes(mat,bses[0],bses[1]);CHKERRQ(ierr); 265 ierr = MatSetType(mat,MATAIJ);CHKERRQ(ierr); 266 ierr = PetscMalloc1(size+1,&rowners);CHKERRQ(ierr); 267 ierr = PetscMalloc2(m,&dlens,m,&olens);CHKERRQ(ierr); 268 ierr = MPI_Allgather(&m,1,MPIU_INT,rowners+1,1,MPIU_INT,comm);CHKERRQ(ierr); 269 270 rowners[0] = 0; 271 for (i=2; i<=size; i++) rowners[i] += rowners[i-1]; 272 rstart = rowners[rank]; 273 rend = rowners[rank+1]; 274 ierr = PetscObjectGetNewTag((PetscObject)mat,&tag);CHKERRQ(ierr); 275 if (!rank) { 276 gmata = (Mat_SeqAIJ*) gmat->data; 277 /* send row lengths to all processors */ 278 for (i=0; i<m; i++) dlens[i] = gmata->ilen[i]; 279 for (i=1; i<size; i++) { 280 ierr = MPI_Send(gmata->ilen + rowners[i],rowners[i+1]-rowners[i],MPIU_INT,i,tag,comm);CHKERRQ(ierr); 281 } 282 /* determine number diagonal and off-diagonal counts */ 283 ierr = PetscArrayzero(olens,m);CHKERRQ(ierr); 284 ierr = PetscCalloc1(m,&ld);CHKERRQ(ierr); 285 jj = 0; 286 for (i=0; i<m; i++) { 287 for (j=0; j<dlens[i]; j++) { 288 if (gmata->j[jj] < rstart) ld[i]++; 289 if (gmata->j[jj] < rstart || gmata->j[jj] >= rend) olens[i]++; 290 jj++; 291 } 292 } 293 /* send column indices to other processes */ 294 for (i=1; i<size; i++) { 295 nz = gmata->i[rowners[i+1]]-gmata->i[rowners[i]]; 296 ierr = MPI_Send(&nz,1,MPIU_INT,i,tag,comm);CHKERRQ(ierr); 297 ierr = MPI_Send(gmata->j + gmata->i[rowners[i]],nz,MPIU_INT,i,tag,comm);CHKERRQ(ierr); 298 } 299 300 /* send numerical values to other processes */ 301 for (i=1; i<size; i++) { 302 nz = gmata->i[rowners[i+1]]-gmata->i[rowners[i]]; 303 ierr = MPI_Send(gmata->a + gmata->i[rowners[i]],nz,MPIU_SCALAR,i,tag,comm);CHKERRQ(ierr); 304 } 305 gmataa = gmata->a; 306 gmataj = gmata->j; 307 308 } else { 309 /* receive row lengths */ 310 ierr = MPI_Recv(dlens,m,MPIU_INT,0,tag,comm,&status);CHKERRQ(ierr); 311 /* receive column indices */ 312 ierr = MPI_Recv(&nz,1,MPIU_INT,0,tag,comm,&status);CHKERRQ(ierr); 313 ierr = PetscMalloc2(nz,&gmataa,nz,&gmataj);CHKERRQ(ierr); 314 ierr = MPI_Recv(gmataj,nz,MPIU_INT,0,tag,comm,&status);CHKERRQ(ierr); 315 /* determine number diagonal and off-diagonal counts */ 316 ierr = PetscArrayzero(olens,m);CHKERRQ(ierr); 317 ierr = PetscCalloc1(m,&ld);CHKERRQ(ierr); 318 jj = 0; 319 for (i=0; i<m; i++) { 320 for (j=0; j<dlens[i]; j++) { 321 if (gmataj[jj] < rstart) ld[i]++; 322 if (gmataj[jj] < rstart || gmataj[jj] >= rend) olens[i]++; 323 jj++; 324 } 325 } 326 /* receive numerical values */ 327 ierr = PetscArrayzero(gmataa,nz);CHKERRQ(ierr); 328 ierr = MPI_Recv(gmataa,nz,MPIU_SCALAR,0,tag,comm,&status);CHKERRQ(ierr); 329 } 330 /* set preallocation */ 331 for (i=0; i<m; i++) { 332 dlens[i] -= olens[i]; 333 } 334 ierr = MatSeqAIJSetPreallocation(mat,0,dlens);CHKERRQ(ierr); 335 ierr = MatMPIAIJSetPreallocation(mat,0,dlens,0,olens);CHKERRQ(ierr); 336 337 for (i=0; i<m; i++) { 338 dlens[i] += olens[i]; 339 } 340 cnt = 0; 341 for (i=0; i<m; i++) { 342 row = rstart + i; 343 ierr = MatSetValues(mat,1,&row,dlens[i],gmataj+cnt,gmataa+cnt,INSERT_VALUES);CHKERRQ(ierr); 344 cnt += dlens[i]; 345 } 346 if (rank) { 347 ierr = PetscFree2(gmataa,gmataj);CHKERRQ(ierr); 348 } 349 ierr = PetscFree2(dlens,olens);CHKERRQ(ierr); 350 ierr = PetscFree(rowners);CHKERRQ(ierr); 351 352 ((Mat_MPIAIJ*)(mat->data))->ld = ld; 353 354 *inmat = mat; 355 } else { /* column indices are already set; only need to move over numerical values from process 0 */ 356 Mat_SeqAIJ *Ad = (Mat_SeqAIJ*)((Mat_MPIAIJ*)((*inmat)->data))->A->data; 357 Mat_SeqAIJ *Ao = (Mat_SeqAIJ*)((Mat_MPIAIJ*)((*inmat)->data))->B->data; 358 mat = *inmat; 359 ierr = PetscObjectGetNewTag((PetscObject)mat,&tag);CHKERRQ(ierr); 360 if (!rank) { 361 /* send numerical values to other processes */ 362 gmata = (Mat_SeqAIJ*) gmat->data; 363 ierr = MatGetOwnershipRanges(mat,(const PetscInt**)&rowners);CHKERRQ(ierr); 364 gmataa = gmata->a; 365 for (i=1; i<size; i++) { 366 nz = gmata->i[rowners[i+1]]-gmata->i[rowners[i]]; 367 ierr = MPI_Send(gmataa + gmata->i[rowners[i]],nz,MPIU_SCALAR,i,tag,comm);CHKERRQ(ierr); 368 } 369 nz = gmata->i[rowners[1]]-gmata->i[rowners[0]]; 370 } else { 371 /* receive numerical values from process 0*/ 372 nz = Ad->nz + Ao->nz; 373 ierr = PetscMalloc1(nz,&gmataa);CHKERRQ(ierr); gmataarestore = gmataa; 374 ierr = MPI_Recv(gmataa,nz,MPIU_SCALAR,0,tag,comm,&status);CHKERRQ(ierr); 375 } 376 /* transfer numerical values into the diagonal A and off diagonal B parts of mat */ 377 ld = ((Mat_MPIAIJ*)(mat->data))->ld; 378 ad = Ad->a; 379 ao = Ao->a; 380 if (mat->rmap->n) { 381 i = 0; 382 nz = ld[i]; ierr = PetscArraycpy(ao,gmataa,nz);CHKERRQ(ierr); ao += nz; gmataa += nz; 383 nz = Ad->i[i+1] - Ad->i[i]; ierr = PetscArraycpy(ad,gmataa,nz);CHKERRQ(ierr); ad += nz; gmataa += nz; 384 } 385 for (i=1; i<mat->rmap->n; i++) { 386 nz = Ao->i[i] - Ao->i[i-1] - ld[i-1] + ld[i]; ierr = PetscArraycpy(ao,gmataa,nz);CHKERRQ(ierr); ao += nz; gmataa += nz; 387 nz = Ad->i[i+1] - Ad->i[i]; ierr = PetscArraycpy(ad,gmataa,nz);CHKERRQ(ierr); ad += nz; gmataa += nz; 388 } 389 i--; 390 if (mat->rmap->n) { 391 nz = Ao->i[i+1] - Ao->i[i] - ld[i]; ierr = PetscArraycpy(ao,gmataa,nz);CHKERRQ(ierr); 392 } 393 if (rank) { 394 ierr = PetscFree(gmataarestore);CHKERRQ(ierr); 395 } 396 } 397 ierr = MatAssemblyBegin(mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 398 ierr = MatAssemblyEnd(mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 399 PetscFunctionReturn(0); 400 } 401 402 /* 403 Local utility routine that creates a mapping from the global column 404 number to the local number in the off-diagonal part of the local 405 storage of the matrix. When PETSC_USE_CTABLE is used this is scalable at 406 a slightly higher hash table cost; without it it is not scalable (each processor 407 has an order N integer array but is fast to acess. 408 */ 409 PetscErrorCode MatCreateColmap_MPIAIJ_Private(Mat mat) 410 { 411 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 412 PetscErrorCode ierr; 413 PetscInt n = aij->B->cmap->n,i; 414 415 PetscFunctionBegin; 416 if (!aij->garray) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"MPIAIJ Matrix was assembled but is missing garray"); 417 #if defined(PETSC_USE_CTABLE) 418 ierr = PetscTableCreate(n,mat->cmap->N+1,&aij->colmap);CHKERRQ(ierr); 419 for (i=0; i<n; i++) { 420 ierr = PetscTableAdd(aij->colmap,aij->garray[i]+1,i+1,INSERT_VALUES);CHKERRQ(ierr); 421 } 422 #else 423 ierr = PetscCalloc1(mat->cmap->N+1,&aij->colmap);CHKERRQ(ierr); 424 ierr = PetscLogObjectMemory((PetscObject)mat,(mat->cmap->N+1)*sizeof(PetscInt));CHKERRQ(ierr); 425 for (i=0; i<n; i++) aij->colmap[aij->garray[i]] = i+1; 426 #endif 427 PetscFunctionReturn(0); 428 } 429 430 #define MatSetValues_SeqAIJ_A_Private(row,col,value,addv,orow,ocol) \ 431 { \ 432 if (col <= lastcol1) low1 = 0; \ 433 else high1 = nrow1; \ 434 lastcol1 = col;\ 435 while (high1-low1 > 5) { \ 436 t = (low1+high1)/2; \ 437 if (rp1[t] > col) high1 = t; \ 438 else low1 = t; \ 439 } \ 440 for (_i=low1; _i<high1; _i++) { \ 441 if (rp1[_i] > col) break; \ 442 if (rp1[_i] == col) { \ 443 if (addv == ADD_VALUES) { \ 444 ap1[_i] += value; \ 445 /* Not sure LogFlops will slow dow the code or not */ \ 446 (void)PetscLogFlops(1.0); \ 447 } \ 448 else ap1[_i] = value; \ 449 goto a_noinsert; \ 450 } \ 451 } \ 452 if (value == 0.0 && ignorezeroentries && row != col) {low1 = 0; high1 = nrow1;goto a_noinsert;} \ 453 if (nonew == 1) {low1 = 0; high1 = nrow1; goto a_noinsert;} \ 454 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); \ 455 MatSeqXAIJReallocateAIJ(A,am,1,nrow1,row,col,rmax1,aa,ai,aj,rp1,ap1,aimax,nonew,MatScalar); \ 456 N = nrow1++ - 1; a->nz++; high1++; \ 457 /* shift up all the later entries in this row */ \ 458 ierr = PetscArraymove(rp1+_i+1,rp1+_i,N-_i+1);CHKERRQ(ierr);\ 459 ierr = PetscArraymove(ap1+_i+1,ap1+_i,N-_i+1);CHKERRQ(ierr);\ 460 rp1[_i] = col; \ 461 ap1[_i] = value; \ 462 A->nonzerostate++;\ 463 a_noinsert: ; \ 464 ailen[row] = nrow1; \ 465 } 466 467 #define MatSetValues_SeqAIJ_B_Private(row,col,value,addv,orow,ocol) \ 468 { \ 469 if (col <= lastcol2) low2 = 0; \ 470 else high2 = nrow2; \ 471 lastcol2 = col; \ 472 while (high2-low2 > 5) { \ 473 t = (low2+high2)/2; \ 474 if (rp2[t] > col) high2 = t; \ 475 else low2 = t; \ 476 } \ 477 for (_i=low2; _i<high2; _i++) { \ 478 if (rp2[_i] > col) break; \ 479 if (rp2[_i] == col) { \ 480 if (addv == ADD_VALUES) { \ 481 ap2[_i] += value; \ 482 (void)PetscLogFlops(1.0); \ 483 } \ 484 else ap2[_i] = value; \ 485 goto b_noinsert; \ 486 } \ 487 } \ 488 if (value == 0.0 && ignorezeroentries) {low2 = 0; high2 = nrow2; goto b_noinsert;} \ 489 if (nonew == 1) {low2 = 0; high2 = nrow2; goto b_noinsert;} \ 490 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); \ 491 MatSeqXAIJReallocateAIJ(B,bm,1,nrow2,row,col,rmax2,ba,bi,bj,rp2,ap2,bimax,nonew,MatScalar); \ 492 N = nrow2++ - 1; b->nz++; high2++; \ 493 /* shift up all the later entries in this row */ \ 494 ierr = PetscArraymove(rp2+_i+1,rp2+_i,N-_i+1);CHKERRQ(ierr);\ 495 ierr = PetscArraymove(ap2+_i+1,ap2+_i,N-_i+1);CHKERRQ(ierr);\ 496 rp2[_i] = col; \ 497 ap2[_i] = value; \ 498 B->nonzerostate++; \ 499 b_noinsert: ; \ 500 bilen[row] = nrow2; \ 501 } 502 503 PetscErrorCode MatSetValuesRow_MPIAIJ(Mat A,PetscInt row,const PetscScalar v[]) 504 { 505 Mat_MPIAIJ *mat = (Mat_MPIAIJ*)A->data; 506 Mat_SeqAIJ *a = (Mat_SeqAIJ*)mat->A->data,*b = (Mat_SeqAIJ*)mat->B->data; 507 PetscErrorCode ierr; 508 PetscInt l,*garray = mat->garray,diag; 509 510 PetscFunctionBegin; 511 /* code only works for square matrices A */ 512 513 /* find size of row to the left of the diagonal part */ 514 ierr = MatGetOwnershipRange(A,&diag,0);CHKERRQ(ierr); 515 row = row - diag; 516 for (l=0; l<b->i[row+1]-b->i[row]; l++) { 517 if (garray[b->j[b->i[row]+l]] > diag) break; 518 } 519 ierr = PetscArraycpy(b->a+b->i[row],v,l);CHKERRQ(ierr); 520 521 /* diagonal part */ 522 ierr = PetscArraycpy(a->a+a->i[row],v+l,(a->i[row+1]-a->i[row]));CHKERRQ(ierr); 523 524 /* right of diagonal part */ 525 ierr = PetscArraycpy(b->a+b->i[row]+l,v+l+a->i[row+1]-a->i[row],b->i[row+1]-b->i[row]-l);CHKERRQ(ierr); 526 PetscFunctionReturn(0); 527 } 528 529 PetscErrorCode MatSetValues_MPIAIJ(Mat mat,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],const PetscScalar v[],InsertMode addv) 530 { 531 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 532 PetscScalar value = 0.0; 533 PetscErrorCode ierr; 534 PetscInt i,j,rstart = mat->rmap->rstart,rend = mat->rmap->rend; 535 PetscInt cstart = mat->cmap->rstart,cend = mat->cmap->rend,row,col; 536 PetscBool roworiented = aij->roworiented; 537 538 /* Some Variables required in the macro */ 539 Mat A = aij->A; 540 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 541 PetscInt *aimax = a->imax,*ai = a->i,*ailen = a->ilen,*aj = a->j; 542 MatScalar *aa = a->a; 543 PetscBool ignorezeroentries = a->ignorezeroentries; 544 Mat B = aij->B; 545 Mat_SeqAIJ *b = (Mat_SeqAIJ*)B->data; 546 PetscInt *bimax = b->imax,*bi = b->i,*bilen = b->ilen,*bj = b->j,bm = aij->B->rmap->n,am = aij->A->rmap->n; 547 MatScalar *ba = b->a; 548 549 PetscInt *rp1,*rp2,ii,nrow1,nrow2,_i,rmax1,rmax2,N,low1,high1,low2,high2,t,lastcol1,lastcol2; 550 PetscInt nonew; 551 MatScalar *ap1,*ap2; 552 553 PetscFunctionBegin; 554 for (i=0; i<m; i++) { 555 if (im[i] < 0) continue; 556 #if defined(PETSC_USE_DEBUG) 557 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); 558 #endif 559 if (im[i] >= rstart && im[i] < rend) { 560 row = im[i] - rstart; 561 lastcol1 = -1; 562 rp1 = aj + ai[row]; 563 ap1 = aa + ai[row]; 564 rmax1 = aimax[row]; 565 nrow1 = ailen[row]; 566 low1 = 0; 567 high1 = nrow1; 568 lastcol2 = -1; 569 rp2 = bj + bi[row]; 570 ap2 = ba + bi[row]; 571 rmax2 = bimax[row]; 572 nrow2 = bilen[row]; 573 low2 = 0; 574 high2 = nrow2; 575 576 for (j=0; j<n; j++) { 577 if (v) value = roworiented ? v[i*n+j] : v[i+j*m]; 578 if (in[j] >= cstart && in[j] < cend) { 579 col = in[j] - cstart; 580 nonew = a->nonew; 581 if (ignorezeroentries && value == 0.0 && (addv == ADD_VALUES) && row != col) continue; 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) { 614 if (1 == ((Mat_SeqAIJ*)(aij->B->data))->nonew) { 615 ierr = PetscInfo3(mat,"Skipping of insertion of new nonzero location in off-diagonal portion of matrix %g(%D,%D)\n",(double)PetscRealPart(value),im[i],in[j]);CHKERRQ(ierr); 616 } else SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero at global row/column (%D, %D) into matrix", im[i], in[j]); 617 } 618 } else col = in[j]; 619 nonew = b->nonew; 620 MatSetValues_SeqAIJ_B_Private(row,col,value,addv,im[i],in[j]); 621 } 622 } 623 } else { 624 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]); 625 if (!aij->donotstash) { 626 mat->assembled = PETSC_FALSE; 627 if (roworiented) { 628 ierr = MatStashValuesRow_Private(&mat->stash,im[i],n,in,v+i*n,(PetscBool)(ignorezeroentries && (addv == ADD_VALUES)));CHKERRQ(ierr); 629 } else { 630 ierr = MatStashValuesCol_Private(&mat->stash,im[i],n,in,v+i,m,(PetscBool)(ignorezeroentries && (addv == ADD_VALUES)));CHKERRQ(ierr); 631 } 632 } 633 } 634 } 635 PetscFunctionReturn(0); 636 } 637 638 /* 639 This function sets the j and ilen arrays (of the diagonal and off-diagonal part) of an MPIAIJ-matrix. 640 The values in mat_i have to be sorted and the values in mat_j have to be sorted for each row (CSR-like). 641 No off-processor parts off the matrix are allowed here and mat->was_assembled has to be PETSC_FALSE. 642 */ 643 PetscErrorCode MatSetValues_MPIAIJ_CopyFromCSRFormat_Symbolic(Mat mat,const PetscInt mat_j[],const PetscInt mat_i[]) 644 { 645 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 646 Mat A = aij->A; /* diagonal part of the matrix */ 647 Mat B = aij->B; /* offdiagonal part of the matrix */ 648 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 649 Mat_SeqAIJ *b = (Mat_SeqAIJ*)B->data; 650 PetscInt cstart = mat->cmap->rstart,cend = mat->cmap->rend,col; 651 PetscInt *ailen = a->ilen,*aj = a->j; 652 PetscInt *bilen = b->ilen,*bj = b->j; 653 PetscInt am = aij->A->rmap->n,j; 654 PetscInt diag_so_far = 0,dnz; 655 PetscInt offd_so_far = 0,onz; 656 657 PetscFunctionBegin; 658 /* Iterate over all rows of the matrix */ 659 for (j=0; j<am; j++) { 660 dnz = onz = 0; 661 /* Iterate over all non-zero columns of the current row */ 662 for (col=mat_i[j]; col<mat_i[j+1]; col++) { 663 /* If column is in the diagonal */ 664 if (mat_j[col] >= cstart && mat_j[col] < cend) { 665 aj[diag_so_far++] = mat_j[col] - cstart; 666 dnz++; 667 } else { /* off-diagonal entries */ 668 bj[offd_so_far++] = mat_j[col]; 669 onz++; 670 } 671 } 672 ailen[j] = dnz; 673 bilen[j] = onz; 674 } 675 PetscFunctionReturn(0); 676 } 677 678 /* 679 This function sets the local j, a and ilen arrays (of the diagonal and off-diagonal part) of an MPIAIJ-matrix. 680 The values in mat_i have to be sorted and the values in mat_j have to be sorted for each row (CSR-like). 681 No off-processor parts off the matrix are allowed here, they are set at a later point by MatSetValues_MPIAIJ. 682 Also, mat->was_assembled has to be false, otherwise the statement aj[rowstart_diag+dnz_row] = mat_j[col] - cstart; 683 would not be true and the more complex MatSetValues_MPIAIJ has to be used. 684 */ 685 PetscErrorCode MatSetValues_MPIAIJ_CopyFromCSRFormat(Mat mat,const PetscInt mat_j[],const PetscInt mat_i[],const PetscScalar mat_a[]) 686 { 687 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 688 Mat A = aij->A; /* diagonal part of the matrix */ 689 Mat B = aij->B; /* offdiagonal part of the matrix */ 690 Mat_SeqAIJ *aijd =(Mat_SeqAIJ*)(aij->A)->data,*aijo=(Mat_SeqAIJ*)(aij->B)->data; 691 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 692 Mat_SeqAIJ *b = (Mat_SeqAIJ*)B->data; 693 PetscInt cstart = mat->cmap->rstart,cend = mat->cmap->rend; 694 PetscInt *ailen = a->ilen,*aj = a->j; 695 PetscInt *bilen = b->ilen,*bj = b->j; 696 PetscInt am = aij->A->rmap->n,j; 697 PetscInt *full_diag_i=aijd->i,*full_offd_i=aijo->i; /* These variables can also include non-local elements, which are set at a later point. */ 698 PetscInt col,dnz_row,onz_row,rowstart_diag,rowstart_offd; 699 PetscScalar *aa = a->a,*ba = b->a; 700 701 PetscFunctionBegin; 702 /* Iterate over all rows of the matrix */ 703 for (j=0; j<am; j++) { 704 dnz_row = onz_row = 0; 705 rowstart_offd = full_offd_i[j]; 706 rowstart_diag = full_diag_i[j]; 707 /* Iterate over all non-zero columns of the current row */ 708 for (col=mat_i[j]; col<mat_i[j+1]; col++) { 709 /* If column is in the diagonal */ 710 if (mat_j[col] >= cstart && mat_j[col] < cend) { 711 aj[rowstart_diag+dnz_row] = mat_j[col] - cstart; 712 aa[rowstart_diag+dnz_row] = mat_a[col]; 713 dnz_row++; 714 } else { /* off-diagonal entries */ 715 bj[rowstart_offd+onz_row] = mat_j[col]; 716 ba[rowstart_offd+onz_row] = mat_a[col]; 717 onz_row++; 718 } 719 } 720 ailen[j] = dnz_row; 721 bilen[j] = onz_row; 722 } 723 PetscFunctionReturn(0); 724 } 725 726 PetscErrorCode MatGetValues_MPIAIJ(Mat mat,PetscInt m,const PetscInt idxm[],PetscInt n,const PetscInt idxn[],PetscScalar v[]) 727 { 728 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 729 PetscErrorCode ierr; 730 PetscInt i,j,rstart = mat->rmap->rstart,rend = mat->rmap->rend; 731 PetscInt cstart = mat->cmap->rstart,cend = mat->cmap->rend,row,col; 732 733 PetscFunctionBegin; 734 for (i=0; i<m; i++) { 735 if (idxm[i] < 0) continue; /* SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative row: %D",idxm[i]);*/ 736 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); 737 if (idxm[i] >= rstart && idxm[i] < rend) { 738 row = idxm[i] - rstart; 739 for (j=0; j<n; j++) { 740 if (idxn[j] < 0) continue; /* SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative column: %D",idxn[j]); */ 741 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); 742 if (idxn[j] >= cstart && idxn[j] < cend) { 743 col = idxn[j] - cstart; 744 ierr = MatGetValues(aij->A,1,&row,1,&col,v+i*n+j);CHKERRQ(ierr); 745 } else { 746 if (!aij->colmap) { 747 ierr = MatCreateColmap_MPIAIJ_Private(mat);CHKERRQ(ierr); 748 } 749 #if defined(PETSC_USE_CTABLE) 750 ierr = PetscTableFind(aij->colmap,idxn[j]+1,&col);CHKERRQ(ierr); 751 col--; 752 #else 753 col = aij->colmap[idxn[j]] - 1; 754 #endif 755 if ((col < 0) || (aij->garray[col] != idxn[j])) *(v+i*n+j) = 0.0; 756 else { 757 ierr = MatGetValues(aij->B,1,&row,1,&col,v+i*n+j);CHKERRQ(ierr); 758 } 759 } 760 } 761 } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Only local values currently supported"); 762 } 763 PetscFunctionReturn(0); 764 } 765 766 extern PetscErrorCode MatMultDiagonalBlock_MPIAIJ(Mat,Vec,Vec); 767 768 PetscErrorCode MatAssemblyBegin_MPIAIJ(Mat mat,MatAssemblyType mode) 769 { 770 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 771 PetscErrorCode ierr; 772 PetscInt nstash,reallocs; 773 774 PetscFunctionBegin; 775 if (aij->donotstash || mat->nooffprocentries) PetscFunctionReturn(0); 776 777 ierr = MatStashScatterBegin_Private(mat,&mat->stash,mat->rmap->range);CHKERRQ(ierr); 778 ierr = MatStashGetInfo_Private(&mat->stash,&nstash,&reallocs);CHKERRQ(ierr); 779 ierr = PetscInfo2(aij->A,"Stash has %D entries, uses %D mallocs.\n",nstash,reallocs);CHKERRQ(ierr); 780 PetscFunctionReturn(0); 781 } 782 783 PetscErrorCode MatAssemblyEnd_MPIAIJ(Mat mat,MatAssemblyType mode) 784 { 785 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 786 Mat_SeqAIJ *a = (Mat_SeqAIJ*)aij->A->data; 787 PetscErrorCode ierr; 788 PetscMPIInt n; 789 PetscInt i,j,rstart,ncols,flg; 790 PetscInt *row,*col; 791 PetscBool other_disassembled; 792 PetscScalar *val; 793 794 /* do not use 'b = (Mat_SeqAIJ*)aij->B->data' as B can be reset in disassembly */ 795 796 PetscFunctionBegin; 797 if (!aij->donotstash && !mat->nooffprocentries) { 798 while (1) { 799 ierr = MatStashScatterGetMesg_Private(&mat->stash,&n,&row,&col,&val,&flg);CHKERRQ(ierr); 800 if (!flg) break; 801 802 for (i=0; i<n; ) { 803 /* Now identify the consecutive vals belonging to the same row */ 804 for (j=i,rstart=row[j]; j<n; j++) { 805 if (row[j] != rstart) break; 806 } 807 if (j < n) ncols = j-i; 808 else ncols = n-i; 809 /* Now assemble all these values with a single function call */ 810 ierr = MatSetValues_MPIAIJ(mat,1,row+i,ncols,col+i,val+i,mat->insertmode);CHKERRQ(ierr); 811 812 i = j; 813 } 814 } 815 ierr = MatStashScatterEnd_Private(&mat->stash);CHKERRQ(ierr); 816 } 817 ierr = MatAssemblyBegin(aij->A,mode);CHKERRQ(ierr); 818 ierr = MatAssemblyEnd(aij->A,mode);CHKERRQ(ierr); 819 820 /* determine if any processor has disassembled, if so we must 821 also disassemble ourself, in order that we may reassemble. */ 822 /* 823 if nonzero structure of submatrix B cannot change then we know that 824 no processor disassembled thus we can skip this stuff 825 */ 826 if (!((Mat_SeqAIJ*)aij->B->data)->nonew) { 827 ierr = MPIU_Allreduce(&mat->was_assembled,&other_disassembled,1,MPIU_BOOL,MPI_PROD,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 828 if (mat->was_assembled && !other_disassembled) { 829 ierr = MatDisAssemble_MPIAIJ(mat);CHKERRQ(ierr); 830 } 831 } 832 if (!mat->was_assembled && mode == MAT_FINAL_ASSEMBLY) { 833 ierr = MatSetUpMultiply_MPIAIJ(mat);CHKERRQ(ierr); 834 } 835 ierr = MatSetOption(aij->B,MAT_USE_INODES,PETSC_FALSE);CHKERRQ(ierr); 836 ierr = MatAssemblyBegin(aij->B,mode);CHKERRQ(ierr); 837 ierr = MatAssemblyEnd(aij->B,mode);CHKERRQ(ierr); 838 839 ierr = PetscFree2(aij->rowvalues,aij->rowindices);CHKERRQ(ierr); 840 841 aij->rowvalues = 0; 842 843 ierr = VecDestroy(&aij->diag);CHKERRQ(ierr); 844 if (a->inode.size) mat->ops->multdiagonalblock = MatMultDiagonalBlock_MPIAIJ; 845 846 /* if no new nonzero locations are allowed in matrix then only set the matrix state the first time through */ 847 if ((!mat->was_assembled && mode == MAT_FINAL_ASSEMBLY) || !((Mat_SeqAIJ*)(aij->A->data))->nonew) { 848 PetscObjectState state = aij->A->nonzerostate + aij->B->nonzerostate; 849 ierr = MPIU_Allreduce(&state,&mat->nonzerostate,1,MPIU_INT64,MPI_SUM,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 850 } 851 PetscFunctionReturn(0); 852 } 853 854 PetscErrorCode MatZeroEntries_MPIAIJ(Mat A) 855 { 856 Mat_MPIAIJ *l = (Mat_MPIAIJ*)A->data; 857 PetscErrorCode ierr; 858 859 PetscFunctionBegin; 860 ierr = MatZeroEntries(l->A);CHKERRQ(ierr); 861 ierr = MatZeroEntries(l->B);CHKERRQ(ierr); 862 PetscFunctionReturn(0); 863 } 864 865 PetscErrorCode MatZeroRows_MPIAIJ(Mat A,PetscInt N,const PetscInt rows[],PetscScalar diag,Vec x,Vec b) 866 { 867 Mat_MPIAIJ *mat = (Mat_MPIAIJ *) A->data; 868 PetscObjectState sA, sB; 869 PetscInt *lrows; 870 PetscInt r, len; 871 PetscBool cong, lch, gch; 872 PetscErrorCode ierr; 873 874 PetscFunctionBegin; 875 /* get locally owned rows */ 876 ierr = MatZeroRowsMapLocal_Private(A,N,rows,&len,&lrows);CHKERRQ(ierr); 877 ierr = MatHasCongruentLayouts(A,&cong);CHKERRQ(ierr); 878 /* fix right hand side if needed */ 879 if (x && b) { 880 const PetscScalar *xx; 881 PetscScalar *bb; 882 883 if (!cong) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"Need matching row/col layout"); 884 ierr = VecGetArrayRead(x, &xx);CHKERRQ(ierr); 885 ierr = VecGetArray(b, &bb);CHKERRQ(ierr); 886 for (r = 0; r < len; ++r) bb[lrows[r]] = diag*xx[lrows[r]]; 887 ierr = VecRestoreArrayRead(x, &xx);CHKERRQ(ierr); 888 ierr = VecRestoreArray(b, &bb);CHKERRQ(ierr); 889 } 890 891 sA = mat->A->nonzerostate; 892 sB = mat->B->nonzerostate; 893 894 if (diag != 0.0 && cong) { 895 ierr = MatZeroRows(mat->A, len, lrows, diag, NULL, NULL);CHKERRQ(ierr); 896 ierr = MatZeroRows(mat->B, len, lrows, 0.0, NULL, NULL);CHKERRQ(ierr); 897 } else if (diag != 0.0) { /* non-square or non congruent layouts -> if keepnonzeropattern is false, we allow for new insertion */ 898 Mat_SeqAIJ *aijA = (Mat_SeqAIJ*)mat->A->data; 899 Mat_SeqAIJ *aijB = (Mat_SeqAIJ*)mat->B->data; 900 PetscInt nnwA, nnwB; 901 PetscBool nnzA, nnzB; 902 903 nnwA = aijA->nonew; 904 nnwB = aijB->nonew; 905 nnzA = aijA->keepnonzeropattern; 906 nnzB = aijB->keepnonzeropattern; 907 if (!nnzA) { 908 ierr = PetscInfo(mat->A,"Requested to not keep the pattern and add a nonzero diagonal; may encounter reallocations on diagonal block.\n");CHKERRQ(ierr); 909 aijA->nonew = 0; 910 } 911 if (!nnzB) { 912 ierr = PetscInfo(mat->B,"Requested to not keep the pattern and add a nonzero diagonal; may encounter reallocations on off-diagonal block.\n");CHKERRQ(ierr); 913 aijB->nonew = 0; 914 } 915 /* Must zero here before the next loop */ 916 ierr = MatZeroRows(mat->A, len, lrows, 0.0, NULL, NULL);CHKERRQ(ierr); 917 ierr = MatZeroRows(mat->B, len, lrows, 0.0, NULL, NULL);CHKERRQ(ierr); 918 for (r = 0; r < len; ++r) { 919 const PetscInt row = lrows[r] + A->rmap->rstart; 920 if (row >= A->cmap->N) continue; 921 ierr = MatSetValues(A, 1, &row, 1, &row, &diag, INSERT_VALUES);CHKERRQ(ierr); 922 } 923 aijA->nonew = nnwA; 924 aijB->nonew = nnwB; 925 } else { 926 ierr = MatZeroRows(mat->A, len, lrows, 0.0, NULL, NULL);CHKERRQ(ierr); 927 ierr = MatZeroRows(mat->B, len, lrows, 0.0, NULL, NULL);CHKERRQ(ierr); 928 } 929 ierr = PetscFree(lrows);CHKERRQ(ierr); 930 ierr = MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 931 ierr = MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 932 933 /* reduce nonzerostate */ 934 lch = (PetscBool)(sA != mat->A->nonzerostate || sB != mat->B->nonzerostate); 935 ierr = MPIU_Allreduce(&lch,&gch,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)A));CHKERRQ(ierr); 936 if (gch) A->nonzerostate++; 937 PetscFunctionReturn(0); 938 } 939 940 PetscErrorCode MatZeroRowsColumns_MPIAIJ(Mat A,PetscInt N,const PetscInt rows[],PetscScalar diag,Vec x,Vec b) 941 { 942 Mat_MPIAIJ *l = (Mat_MPIAIJ*)A->data; 943 PetscErrorCode ierr; 944 PetscMPIInt n = A->rmap->n; 945 PetscInt i,j,r,m,p = 0,len = 0; 946 PetscInt *lrows,*owners = A->rmap->range; 947 PetscSFNode *rrows; 948 PetscSF sf; 949 const PetscScalar *xx; 950 PetscScalar *bb,*mask; 951 Vec xmask,lmask; 952 Mat_SeqAIJ *aij = (Mat_SeqAIJ*)l->B->data; 953 const PetscInt *aj, *ii,*ridx; 954 PetscScalar *aa; 955 956 PetscFunctionBegin; 957 /* Create SF where leaves are input rows and roots are owned rows */ 958 ierr = PetscMalloc1(n, &lrows);CHKERRQ(ierr); 959 for (r = 0; r < n; ++r) lrows[r] = -1; 960 ierr = PetscMalloc1(N, &rrows);CHKERRQ(ierr); 961 for (r = 0; r < N; ++r) { 962 const PetscInt idx = rows[r]; 963 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); 964 if (idx < owners[p] || owners[p+1] <= idx) { /* short-circuit the search if the last p owns this row too */ 965 ierr = PetscLayoutFindOwner(A->rmap,idx,&p);CHKERRQ(ierr); 966 } 967 rrows[r].rank = p; 968 rrows[r].index = rows[r] - owners[p]; 969 } 970 ierr = PetscSFCreate(PetscObjectComm((PetscObject) A), &sf);CHKERRQ(ierr); 971 ierr = PetscSFSetGraph(sf, n, N, NULL, PETSC_OWN_POINTER, rrows, PETSC_OWN_POINTER);CHKERRQ(ierr); 972 /* Collect flags for rows to be zeroed */ 973 ierr = PetscSFReduceBegin(sf, MPIU_INT, (PetscInt *) rows, lrows, MPI_LOR);CHKERRQ(ierr); 974 ierr = PetscSFReduceEnd(sf, MPIU_INT, (PetscInt *) rows, lrows, MPI_LOR);CHKERRQ(ierr); 975 ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 976 /* Compress and put in row numbers */ 977 for (r = 0; r < n; ++r) if (lrows[r] >= 0) lrows[len++] = r; 978 /* zero diagonal part of matrix */ 979 ierr = MatZeroRowsColumns(l->A,len,lrows,diag,x,b);CHKERRQ(ierr); 980 /* handle off diagonal part of matrix */ 981 ierr = MatCreateVecs(A,&xmask,NULL);CHKERRQ(ierr); 982 ierr = VecDuplicate(l->lvec,&lmask);CHKERRQ(ierr); 983 ierr = VecGetArray(xmask,&bb);CHKERRQ(ierr); 984 for (i=0; i<len; i++) bb[lrows[i]] = 1; 985 ierr = VecRestoreArray(xmask,&bb);CHKERRQ(ierr); 986 ierr = VecScatterBegin(l->Mvctx,xmask,lmask,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 987 ierr = VecScatterEnd(l->Mvctx,xmask,lmask,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 988 ierr = VecDestroy(&xmask);CHKERRQ(ierr); 989 if (x && b) { /* this code is buggy when the row and column layout don't match */ 990 PetscBool cong; 991 992 ierr = MatHasCongruentLayouts(A,&cong);CHKERRQ(ierr); 993 if (!cong) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"Need matching row/col layout"); 994 ierr = VecScatterBegin(l->Mvctx,x,l->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 995 ierr = VecScatterEnd(l->Mvctx,x,l->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 996 ierr = VecGetArrayRead(l->lvec,&xx);CHKERRQ(ierr); 997 ierr = VecGetArray(b,&bb);CHKERRQ(ierr); 998 } 999 ierr = VecGetArray(lmask,&mask);CHKERRQ(ierr); 1000 /* remove zeroed rows of off diagonal matrix */ 1001 ii = aij->i; 1002 for (i=0; i<len; i++) { 1003 ierr = PetscArrayzero(aij->a + ii[lrows[i]],ii[lrows[i]+1] - ii[lrows[i]]);CHKERRQ(ierr); 1004 } 1005 /* loop over all elements of off process part of matrix zeroing removed columns*/ 1006 if (aij->compressedrow.use) { 1007 m = aij->compressedrow.nrows; 1008 ii = aij->compressedrow.i; 1009 ridx = aij->compressedrow.rindex; 1010 for (i=0; i<m; i++) { 1011 n = ii[i+1] - ii[i]; 1012 aj = aij->j + ii[i]; 1013 aa = aij->a + ii[i]; 1014 1015 for (j=0; j<n; j++) { 1016 if (PetscAbsScalar(mask[*aj])) { 1017 if (b) bb[*ridx] -= *aa*xx[*aj]; 1018 *aa = 0.0; 1019 } 1020 aa++; 1021 aj++; 1022 } 1023 ridx++; 1024 } 1025 } else { /* do not use compressed row format */ 1026 m = l->B->rmap->n; 1027 for (i=0; i<m; i++) { 1028 n = ii[i+1] - ii[i]; 1029 aj = aij->j + ii[i]; 1030 aa = aij->a + ii[i]; 1031 for (j=0; j<n; j++) { 1032 if (PetscAbsScalar(mask[*aj])) { 1033 if (b) bb[i] -= *aa*xx[*aj]; 1034 *aa = 0.0; 1035 } 1036 aa++; 1037 aj++; 1038 } 1039 } 1040 } 1041 if (x && b) { 1042 ierr = VecRestoreArray(b,&bb);CHKERRQ(ierr); 1043 ierr = VecRestoreArrayRead(l->lvec,&xx);CHKERRQ(ierr); 1044 } 1045 ierr = VecRestoreArray(lmask,&mask);CHKERRQ(ierr); 1046 ierr = VecDestroy(&lmask);CHKERRQ(ierr); 1047 ierr = PetscFree(lrows);CHKERRQ(ierr); 1048 1049 /* only change matrix nonzero state if pattern was allowed to be changed */ 1050 if (!((Mat_SeqAIJ*)(l->A->data))->keepnonzeropattern) { 1051 PetscObjectState state = l->A->nonzerostate + l->B->nonzerostate; 1052 ierr = MPIU_Allreduce(&state,&A->nonzerostate,1,MPIU_INT64,MPI_SUM,PetscObjectComm((PetscObject)A));CHKERRQ(ierr); 1053 } 1054 PetscFunctionReturn(0); 1055 } 1056 1057 PetscErrorCode MatMult_MPIAIJ(Mat A,Vec xx,Vec yy) 1058 { 1059 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 1060 PetscErrorCode ierr; 1061 PetscInt nt; 1062 VecScatter Mvctx = a->Mvctx; 1063 1064 PetscFunctionBegin; 1065 ierr = VecGetLocalSize(xx,&nt);CHKERRQ(ierr); 1066 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); 1067 1068 ierr = VecScatterBegin(Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1069 ierr = (*a->A->ops->mult)(a->A,xx,yy);CHKERRQ(ierr); 1070 ierr = VecScatterEnd(Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1071 ierr = (*a->B->ops->multadd)(a->B,a->lvec,yy,yy);CHKERRQ(ierr); 1072 PetscFunctionReturn(0); 1073 } 1074 1075 PetscErrorCode MatMultDiagonalBlock_MPIAIJ(Mat A,Vec bb,Vec xx) 1076 { 1077 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 1078 PetscErrorCode ierr; 1079 1080 PetscFunctionBegin; 1081 ierr = MatMultDiagonalBlock(a->A,bb,xx);CHKERRQ(ierr); 1082 PetscFunctionReturn(0); 1083 } 1084 1085 PetscErrorCode MatMultAdd_MPIAIJ(Mat A,Vec xx,Vec yy,Vec zz) 1086 { 1087 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 1088 PetscErrorCode ierr; 1089 VecScatter Mvctx = a->Mvctx; 1090 1091 PetscFunctionBegin; 1092 if (a->Mvctx_mpi1_flg) Mvctx = a->Mvctx_mpi1; 1093 ierr = VecScatterBegin(Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1094 ierr = (*a->A->ops->multadd)(a->A,xx,yy,zz);CHKERRQ(ierr); 1095 ierr = VecScatterEnd(Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1096 ierr = (*a->B->ops->multadd)(a->B,a->lvec,zz,zz);CHKERRQ(ierr); 1097 PetscFunctionReturn(0); 1098 } 1099 1100 PetscErrorCode MatMultTranspose_MPIAIJ(Mat A,Vec xx,Vec yy) 1101 { 1102 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 1103 PetscErrorCode ierr; 1104 1105 PetscFunctionBegin; 1106 /* do nondiagonal part */ 1107 ierr = (*a->B->ops->multtranspose)(a->B,xx,a->lvec);CHKERRQ(ierr); 1108 /* do local part */ 1109 ierr = (*a->A->ops->multtranspose)(a->A,xx,yy);CHKERRQ(ierr); 1110 /* add partial results together */ 1111 ierr = VecScatterBegin(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1112 ierr = VecScatterEnd(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1113 PetscFunctionReturn(0); 1114 } 1115 1116 PetscErrorCode MatIsTranspose_MPIAIJ(Mat Amat,Mat Bmat,PetscReal tol,PetscBool *f) 1117 { 1118 MPI_Comm comm; 1119 Mat_MPIAIJ *Aij = (Mat_MPIAIJ*) Amat->data, *Bij; 1120 Mat Adia = Aij->A, Bdia, Aoff,Boff,*Aoffs,*Boffs; 1121 IS Me,Notme; 1122 PetscErrorCode ierr; 1123 PetscInt M,N,first,last,*notme,i; 1124 PetscBool lf; 1125 PetscMPIInt size; 1126 1127 PetscFunctionBegin; 1128 /* Easy test: symmetric diagonal block */ 1129 Bij = (Mat_MPIAIJ*) Bmat->data; Bdia = Bij->A; 1130 ierr = MatIsTranspose(Adia,Bdia,tol,&lf);CHKERRQ(ierr); 1131 ierr = MPIU_Allreduce(&lf,f,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)Amat));CHKERRQ(ierr); 1132 if (!*f) PetscFunctionReturn(0); 1133 ierr = PetscObjectGetComm((PetscObject)Amat,&comm);CHKERRQ(ierr); 1134 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 1135 if (size == 1) PetscFunctionReturn(0); 1136 1137 /* Hard test: off-diagonal block. This takes a MatCreateSubMatrix. */ 1138 ierr = MatGetSize(Amat,&M,&N);CHKERRQ(ierr); 1139 ierr = MatGetOwnershipRange(Amat,&first,&last);CHKERRQ(ierr); 1140 ierr = PetscMalloc1(N-last+first,¬me);CHKERRQ(ierr); 1141 for (i=0; i<first; i++) notme[i] = i; 1142 for (i=last; i<M; i++) notme[i-last+first] = i; 1143 ierr = ISCreateGeneral(MPI_COMM_SELF,N-last+first,notme,PETSC_COPY_VALUES,&Notme);CHKERRQ(ierr); 1144 ierr = ISCreateStride(MPI_COMM_SELF,last-first,first,1,&Me);CHKERRQ(ierr); 1145 ierr = MatCreateSubMatrices(Amat,1,&Me,&Notme,MAT_INITIAL_MATRIX,&Aoffs);CHKERRQ(ierr); 1146 Aoff = Aoffs[0]; 1147 ierr = MatCreateSubMatrices(Bmat,1,&Notme,&Me,MAT_INITIAL_MATRIX,&Boffs);CHKERRQ(ierr); 1148 Boff = Boffs[0]; 1149 ierr = MatIsTranspose(Aoff,Boff,tol,f);CHKERRQ(ierr); 1150 ierr = MatDestroyMatrices(1,&Aoffs);CHKERRQ(ierr); 1151 ierr = MatDestroyMatrices(1,&Boffs);CHKERRQ(ierr); 1152 ierr = ISDestroy(&Me);CHKERRQ(ierr); 1153 ierr = ISDestroy(&Notme);CHKERRQ(ierr); 1154 ierr = PetscFree(notme);CHKERRQ(ierr); 1155 PetscFunctionReturn(0); 1156 } 1157 1158 PetscErrorCode MatIsSymmetric_MPIAIJ(Mat A,PetscReal tol,PetscBool *f) 1159 { 1160 PetscErrorCode ierr; 1161 1162 PetscFunctionBegin; 1163 ierr = MatIsTranspose_MPIAIJ(A,A,tol,f);CHKERRQ(ierr); 1164 PetscFunctionReturn(0); 1165 } 1166 1167 PetscErrorCode MatMultTransposeAdd_MPIAIJ(Mat A,Vec xx,Vec yy,Vec zz) 1168 { 1169 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 1170 PetscErrorCode ierr; 1171 1172 PetscFunctionBegin; 1173 /* do nondiagonal part */ 1174 ierr = (*a->B->ops->multtranspose)(a->B,xx,a->lvec);CHKERRQ(ierr); 1175 /* do local part */ 1176 ierr = (*a->A->ops->multtransposeadd)(a->A,xx,yy,zz);CHKERRQ(ierr); 1177 /* add partial results together */ 1178 ierr = VecScatterBegin(a->Mvctx,a->lvec,zz,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1179 ierr = VecScatterEnd(a->Mvctx,a->lvec,zz,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1180 PetscFunctionReturn(0); 1181 } 1182 1183 /* 1184 This only works correctly for square matrices where the subblock A->A is the 1185 diagonal block 1186 */ 1187 PetscErrorCode MatGetDiagonal_MPIAIJ(Mat A,Vec v) 1188 { 1189 PetscErrorCode ierr; 1190 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 1191 1192 PetscFunctionBegin; 1193 if (A->rmap->N != A->cmap->N) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"Supports only square matrix where A->A is diag block"); 1194 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"); 1195 ierr = MatGetDiagonal(a->A,v);CHKERRQ(ierr); 1196 PetscFunctionReturn(0); 1197 } 1198 1199 PetscErrorCode MatScale_MPIAIJ(Mat A,PetscScalar aa) 1200 { 1201 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 1202 PetscErrorCode ierr; 1203 1204 PetscFunctionBegin; 1205 ierr = MatScale(a->A,aa);CHKERRQ(ierr); 1206 ierr = MatScale(a->B,aa);CHKERRQ(ierr); 1207 PetscFunctionReturn(0); 1208 } 1209 1210 PetscErrorCode MatDestroy_MPIAIJ(Mat mat) 1211 { 1212 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 1213 PetscErrorCode ierr; 1214 1215 PetscFunctionBegin; 1216 #if defined(PETSC_USE_LOG) 1217 PetscLogObjectState((PetscObject)mat,"Rows=%D, Cols=%D",mat->rmap->N,mat->cmap->N); 1218 #endif 1219 ierr = MatStashDestroy_Private(&mat->stash);CHKERRQ(ierr); 1220 ierr = VecDestroy(&aij->diag);CHKERRQ(ierr); 1221 ierr = MatDestroy(&aij->A);CHKERRQ(ierr); 1222 ierr = MatDestroy(&aij->B);CHKERRQ(ierr); 1223 #if defined(PETSC_USE_CTABLE) 1224 ierr = PetscTableDestroy(&aij->colmap);CHKERRQ(ierr); 1225 #else 1226 ierr = PetscFree(aij->colmap);CHKERRQ(ierr); 1227 #endif 1228 ierr = PetscFree(aij->garray);CHKERRQ(ierr); 1229 ierr = VecDestroy(&aij->lvec);CHKERRQ(ierr); 1230 ierr = VecScatterDestroy(&aij->Mvctx);CHKERRQ(ierr); 1231 if (aij->Mvctx_mpi1) {ierr = VecScatterDestroy(&aij->Mvctx_mpi1);CHKERRQ(ierr);} 1232 ierr = PetscFree2(aij->rowvalues,aij->rowindices);CHKERRQ(ierr); 1233 ierr = PetscFree(aij->ld);CHKERRQ(ierr); 1234 ierr = PetscFree(mat->data);CHKERRQ(ierr); 1235 1236 ierr = PetscObjectChangeTypeName((PetscObject)mat,0);CHKERRQ(ierr); 1237 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatStoreValues_C",NULL);CHKERRQ(ierr); 1238 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatRetrieveValues_C",NULL);CHKERRQ(ierr); 1239 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatIsTranspose_C",NULL);CHKERRQ(ierr); 1240 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatMPIAIJSetPreallocation_C",NULL);CHKERRQ(ierr); 1241 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatResetPreallocation_C",NULL);CHKERRQ(ierr); 1242 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatMPIAIJSetPreallocationCSR_C",NULL);CHKERRQ(ierr); 1243 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatDiagonalScaleLocal_C",NULL);CHKERRQ(ierr); 1244 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatConvert_mpiaij_mpisbaij_C",NULL);CHKERRQ(ierr); 1245 #if defined(PETSC_HAVE_ELEMENTAL) 1246 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatConvert_mpiaij_elemental_C",NULL);CHKERRQ(ierr); 1247 #endif 1248 #if defined(PETSC_HAVE_HYPRE) 1249 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatConvert_mpiaij_hypre_C",NULL);CHKERRQ(ierr); 1250 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatMatMatMult_transpose_mpiaij_mpiaij_C",NULL);CHKERRQ(ierr); 1251 #endif 1252 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatConvert_mpiaij_is_C",NULL);CHKERRQ(ierr); 1253 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatPtAP_is_mpiaij_C",NULL);CHKERRQ(ierr); 1254 PetscFunctionReturn(0); 1255 } 1256 1257 PetscErrorCode MatView_MPIAIJ_Binary(Mat mat,PetscViewer viewer) 1258 { 1259 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 1260 Mat_SeqAIJ *A = (Mat_SeqAIJ*)aij->A->data; 1261 Mat_SeqAIJ *B = (Mat_SeqAIJ*)aij->B->data; 1262 PetscErrorCode ierr; 1263 PetscMPIInt rank,size,tag = ((PetscObject)viewer)->tag; 1264 int fd; 1265 PetscInt nz,header[4],*row_lengths,*range=0,rlen,i; 1266 PetscInt nzmax,*column_indices,j,k,col,*garray = aij->garray,cnt,cstart = mat->cmap->rstart,rnz = 0; 1267 PetscScalar *column_values; 1268 PetscInt message_count,flowcontrolcount; 1269 FILE *file; 1270 1271 PetscFunctionBegin; 1272 ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr); 1273 ierr = MPI_Comm_size(PetscObjectComm((PetscObject)mat),&size);CHKERRQ(ierr); 1274 nz = A->nz + B->nz; 1275 ierr = PetscViewerBinaryGetDescriptor(viewer,&fd);CHKERRQ(ierr); 1276 if (!rank) { 1277 header[0] = MAT_FILE_CLASSID; 1278 header[1] = mat->rmap->N; 1279 header[2] = mat->cmap->N; 1280 1281 ierr = MPI_Reduce(&nz,&header[3],1,MPIU_INT,MPI_SUM,0,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1282 ierr = PetscBinaryWrite(fd,header,4,PETSC_INT,PETSC_TRUE);CHKERRQ(ierr); 1283 /* get largest number of rows any processor has */ 1284 rlen = mat->rmap->n; 1285 range = mat->rmap->range; 1286 for (i=1; i<size; i++) rlen = PetscMax(rlen,range[i+1] - range[i]); 1287 } else { 1288 ierr = MPI_Reduce(&nz,0,1,MPIU_INT,MPI_SUM,0,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1289 rlen = mat->rmap->n; 1290 } 1291 1292 /* load up the local row counts */ 1293 ierr = PetscMalloc1(rlen+1,&row_lengths);CHKERRQ(ierr); 1294 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]; 1295 1296 /* store the row lengths to the file */ 1297 ierr = PetscViewerFlowControlStart(viewer,&message_count,&flowcontrolcount);CHKERRQ(ierr); 1298 if (!rank) { 1299 ierr = PetscBinaryWrite(fd,row_lengths,mat->rmap->n,PETSC_INT,PETSC_TRUE);CHKERRQ(ierr); 1300 for (i=1; i<size; i++) { 1301 ierr = PetscViewerFlowControlStepMaster(viewer,i,&message_count,flowcontrolcount);CHKERRQ(ierr); 1302 rlen = range[i+1] - range[i]; 1303 ierr = MPIULong_Recv(row_lengths,rlen,MPIU_INT,i,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1304 ierr = PetscBinaryWrite(fd,row_lengths,rlen,PETSC_INT,PETSC_TRUE);CHKERRQ(ierr); 1305 } 1306 ierr = PetscViewerFlowControlEndMaster(viewer,&message_count);CHKERRQ(ierr); 1307 } else { 1308 ierr = PetscViewerFlowControlStepWorker(viewer,rank,&message_count);CHKERRQ(ierr); 1309 ierr = MPIULong_Send(row_lengths,mat->rmap->n,MPIU_INT,0,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1310 ierr = PetscViewerFlowControlEndWorker(viewer,&message_count);CHKERRQ(ierr); 1311 } 1312 ierr = PetscFree(row_lengths);CHKERRQ(ierr); 1313 1314 /* load up the local column indices */ 1315 nzmax = nz; /* th processor needs space a largest processor needs */ 1316 ierr = MPI_Reduce(&nz,&nzmax,1,MPIU_INT,MPI_MAX,0,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1317 ierr = PetscMalloc1(nzmax+1,&column_indices);CHKERRQ(ierr); 1318 cnt = 0; 1319 for (i=0; i<mat->rmap->n; i++) { 1320 for (j=B->i[i]; j<B->i[i+1]; j++) { 1321 if ((col = garray[B->j[j]]) > cstart) break; 1322 column_indices[cnt++] = col; 1323 } 1324 for (k=A->i[i]; k<A->i[i+1]; k++) column_indices[cnt++] = A->j[k] + cstart; 1325 for (; j<B->i[i+1]; j++) column_indices[cnt++] = garray[B->j[j]]; 1326 } 1327 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); 1328 1329 /* store the column indices to the file */ 1330 ierr = PetscViewerFlowControlStart(viewer,&message_count,&flowcontrolcount);CHKERRQ(ierr); 1331 if (!rank) { 1332 MPI_Status status; 1333 ierr = PetscBinaryWrite(fd,column_indices,nz,PETSC_INT,PETSC_TRUE);CHKERRQ(ierr); 1334 for (i=1; i<size; i++) { 1335 ierr = PetscViewerFlowControlStepMaster(viewer,i,&message_count,flowcontrolcount);CHKERRQ(ierr); 1336 ierr = MPI_Recv(&rnz,1,MPIU_INT,i,tag,PetscObjectComm((PetscObject)mat),&status);CHKERRQ(ierr); 1337 if (rnz > nzmax) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_LIB,"Internal PETSc error: nz = %D nzmax = %D",nz,nzmax); 1338 ierr = MPIULong_Recv(column_indices,rnz,MPIU_INT,i,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1339 ierr = PetscBinaryWrite(fd,column_indices,rnz,PETSC_INT,PETSC_TRUE);CHKERRQ(ierr); 1340 } 1341 ierr = PetscViewerFlowControlEndMaster(viewer,&message_count);CHKERRQ(ierr); 1342 } else { 1343 ierr = PetscViewerFlowControlStepWorker(viewer,rank,&message_count);CHKERRQ(ierr); 1344 ierr = MPI_Send(&nz,1,MPIU_INT,0,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1345 ierr = MPIULong_Send(column_indices,nz,MPIU_INT,0,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1346 ierr = PetscViewerFlowControlEndWorker(viewer,&message_count);CHKERRQ(ierr); 1347 } 1348 ierr = PetscFree(column_indices);CHKERRQ(ierr); 1349 1350 /* load up the local column values */ 1351 ierr = PetscMalloc1(nzmax+1,&column_values);CHKERRQ(ierr); 1352 cnt = 0; 1353 for (i=0; i<mat->rmap->n; i++) { 1354 for (j=B->i[i]; j<B->i[i+1]; j++) { 1355 if (garray[B->j[j]] > cstart) break; 1356 column_values[cnt++] = B->a[j]; 1357 } 1358 for (k=A->i[i]; k<A->i[i+1]; k++) column_values[cnt++] = A->a[k]; 1359 for (; j<B->i[i+1]; j++) column_values[cnt++] = B->a[j]; 1360 } 1361 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); 1362 1363 /* store the column values to the file */ 1364 ierr = PetscViewerFlowControlStart(viewer,&message_count,&flowcontrolcount);CHKERRQ(ierr); 1365 if (!rank) { 1366 MPI_Status status; 1367 ierr = PetscBinaryWrite(fd,column_values,nz,PETSC_SCALAR,PETSC_TRUE);CHKERRQ(ierr); 1368 for (i=1; i<size; i++) { 1369 ierr = PetscViewerFlowControlStepMaster(viewer,i,&message_count,flowcontrolcount);CHKERRQ(ierr); 1370 ierr = MPI_Recv(&rnz,1,MPIU_INT,i,tag,PetscObjectComm((PetscObject)mat),&status);CHKERRQ(ierr); 1371 if (rnz > nzmax) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_LIB,"Internal PETSc error: nz = %D nzmax = %D",nz,nzmax); 1372 ierr = MPIULong_Recv(column_values,rnz,MPIU_SCALAR,i,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1373 ierr = PetscBinaryWrite(fd,column_values,rnz,PETSC_SCALAR,PETSC_TRUE);CHKERRQ(ierr); 1374 } 1375 ierr = PetscViewerFlowControlEndMaster(viewer,&message_count);CHKERRQ(ierr); 1376 } else { 1377 ierr = PetscViewerFlowControlStepWorker(viewer,rank,&message_count);CHKERRQ(ierr); 1378 ierr = MPI_Send(&nz,1,MPIU_INT,0,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1379 ierr = MPIULong_Send(column_values,nz,MPIU_SCALAR,0,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1380 ierr = PetscViewerFlowControlEndWorker(viewer,&message_count);CHKERRQ(ierr); 1381 } 1382 ierr = PetscFree(column_values);CHKERRQ(ierr); 1383 1384 ierr = PetscViewerBinaryGetInfoPointer(viewer,&file);CHKERRQ(ierr); 1385 if (file) fprintf(file,"-matload_block_size %d\n",(int)PetscAbs(mat->rmap->bs)); 1386 PetscFunctionReturn(0); 1387 } 1388 1389 #include <petscdraw.h> 1390 PetscErrorCode MatView_MPIAIJ_ASCIIorDraworSocket(Mat mat,PetscViewer viewer) 1391 { 1392 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 1393 PetscErrorCode ierr; 1394 PetscMPIInt rank = aij->rank,size = aij->size; 1395 PetscBool isdraw,iascii,isbinary; 1396 PetscViewer sviewer; 1397 PetscViewerFormat format; 1398 1399 PetscFunctionBegin; 1400 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERDRAW,&isdraw);CHKERRQ(ierr); 1401 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii);CHKERRQ(ierr); 1402 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERBINARY,&isbinary);CHKERRQ(ierr); 1403 if (iascii) { 1404 ierr = PetscViewerGetFormat(viewer,&format);CHKERRQ(ierr); 1405 if (format == PETSC_VIEWER_LOAD_BALANCE) { 1406 PetscInt i,nmax = 0,nmin = PETSC_MAX_INT,navg = 0,*nz,nzlocal = ((Mat_SeqAIJ*) (aij->A->data))->nz + ((Mat_SeqAIJ*) (aij->B->data))->nz; 1407 ierr = PetscMalloc1(size,&nz);CHKERRQ(ierr); 1408 ierr = MPI_Allgather(&nzlocal,1,MPIU_INT,nz,1,MPIU_INT,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1409 for (i=0; i<(PetscInt)size; i++) { 1410 nmax = PetscMax(nmax,nz[i]); 1411 nmin = PetscMin(nmin,nz[i]); 1412 navg += nz[i]; 1413 } 1414 ierr = PetscFree(nz);CHKERRQ(ierr); 1415 navg = navg/size; 1416 ierr = PetscViewerASCIIPrintf(viewer,"Load Balance - Nonzeros: Min %D avg %D max %D\n",nmin,navg,nmax);CHKERRQ(ierr); 1417 PetscFunctionReturn(0); 1418 } 1419 ierr = PetscViewerGetFormat(viewer,&format);CHKERRQ(ierr); 1420 if (format == PETSC_VIEWER_ASCII_INFO_DETAIL) { 1421 MatInfo info; 1422 PetscBool inodes; 1423 1424 ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr); 1425 ierr = MatGetInfo(mat,MAT_LOCAL,&info);CHKERRQ(ierr); 1426 ierr = MatInodeGetInodeSizes(aij->A,NULL,(PetscInt**)&inodes,NULL);CHKERRQ(ierr); 1427 ierr = PetscViewerASCIIPushSynchronized(viewer);CHKERRQ(ierr); 1428 if (!inodes) { 1429 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"[%d] Local rows %D nz %D nz alloced %D mem %g, not using I-node routines\n", 1430 rank,mat->rmap->n,(PetscInt)info.nz_used,(PetscInt)info.nz_allocated,(double)info.memory);CHKERRQ(ierr); 1431 } else { 1432 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"[%d] Local rows %D nz %D nz alloced %D mem %g, using I-node routines\n", 1433 rank,mat->rmap->n,(PetscInt)info.nz_used,(PetscInt)info.nz_allocated,(double)info.memory);CHKERRQ(ierr); 1434 } 1435 ierr = MatGetInfo(aij->A,MAT_LOCAL,&info);CHKERRQ(ierr); 1436 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"[%d] on-diagonal part: nz %D \n",rank,(PetscInt)info.nz_used);CHKERRQ(ierr); 1437 ierr = MatGetInfo(aij->B,MAT_LOCAL,&info);CHKERRQ(ierr); 1438 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"[%d] off-diagonal part: nz %D \n",rank,(PetscInt)info.nz_used);CHKERRQ(ierr); 1439 ierr = PetscViewerFlush(viewer);CHKERRQ(ierr); 1440 ierr = PetscViewerASCIIPopSynchronized(viewer);CHKERRQ(ierr); 1441 ierr = PetscViewerASCIIPrintf(viewer,"Information on VecScatter used in matrix-vector product: \n");CHKERRQ(ierr); 1442 ierr = VecScatterView(aij->Mvctx,viewer);CHKERRQ(ierr); 1443 PetscFunctionReturn(0); 1444 } else if (format == PETSC_VIEWER_ASCII_INFO) { 1445 PetscInt inodecount,inodelimit,*inodes; 1446 ierr = MatInodeGetInodeSizes(aij->A,&inodecount,&inodes,&inodelimit);CHKERRQ(ierr); 1447 if (inodes) { 1448 ierr = PetscViewerASCIIPrintf(viewer,"using I-node (on process 0) routines: found %D nodes, limit used is %D\n",inodecount,inodelimit);CHKERRQ(ierr); 1449 } else { 1450 ierr = PetscViewerASCIIPrintf(viewer,"not using I-node (on process 0) routines\n");CHKERRQ(ierr); 1451 } 1452 PetscFunctionReturn(0); 1453 } else if (format == PETSC_VIEWER_ASCII_FACTOR_INFO) { 1454 PetscFunctionReturn(0); 1455 } 1456 } else if (isbinary) { 1457 if (size == 1) { 1458 ierr = PetscObjectSetName((PetscObject)aij->A,((PetscObject)mat)->name);CHKERRQ(ierr); 1459 ierr = MatView(aij->A,viewer);CHKERRQ(ierr); 1460 } else { 1461 ierr = MatView_MPIAIJ_Binary(mat,viewer);CHKERRQ(ierr); 1462 } 1463 PetscFunctionReturn(0); 1464 } else if (iascii && size == 1) { 1465 ierr = PetscObjectSetName((PetscObject)aij->A,((PetscObject)mat)->name);CHKERRQ(ierr); 1466 ierr = MatView(aij->A,viewer);CHKERRQ(ierr); 1467 PetscFunctionReturn(0); 1468 } else if (isdraw) { 1469 PetscDraw draw; 1470 PetscBool isnull; 1471 ierr = PetscViewerDrawGetDraw(viewer,0,&draw);CHKERRQ(ierr); 1472 ierr = PetscDrawIsNull(draw,&isnull);CHKERRQ(ierr); 1473 if (isnull) PetscFunctionReturn(0); 1474 } 1475 1476 { /* assemble the entire matrix onto first processor */ 1477 Mat A = NULL, Av; 1478 IS isrow,iscol; 1479 1480 ierr = ISCreateStride(PetscObjectComm((PetscObject)mat),!rank ? mat->rmap->N : 0,0,1,&isrow);CHKERRQ(ierr); 1481 ierr = ISCreateStride(PetscObjectComm((PetscObject)mat),!rank ? mat->cmap->N : 0,0,1,&iscol);CHKERRQ(ierr); 1482 ierr = MatCreateSubMatrix(mat,isrow,iscol,MAT_INITIAL_MATRIX,&A);CHKERRQ(ierr); 1483 ierr = MatMPIAIJGetSeqAIJ(A,&Av,NULL,NULL);CHKERRQ(ierr); 1484 /* The commented code uses MatCreateSubMatrices instead */ 1485 /* 1486 Mat *AA, A = NULL, Av; 1487 IS isrow,iscol; 1488 1489 ierr = ISCreateStride(PetscObjectComm((PetscObject)mat),!rank ? mat->rmap->N : 0,0,1,&isrow);CHKERRQ(ierr); 1490 ierr = ISCreateStride(PetscObjectComm((PetscObject)mat),!rank ? mat->cmap->N : 0,0,1,&iscol);CHKERRQ(ierr); 1491 ierr = MatCreateSubMatrices(mat,1,&isrow,&iscol,MAT_INITIAL_MATRIX,&AA);CHKERRQ(ierr); 1492 if (!rank) { 1493 ierr = PetscObjectReference((PetscObject)AA[0]);CHKERRQ(ierr); 1494 A = AA[0]; 1495 Av = AA[0]; 1496 } 1497 ierr = MatDestroySubMatrices(1,&AA);CHKERRQ(ierr); 1498 */ 1499 ierr = ISDestroy(&iscol);CHKERRQ(ierr); 1500 ierr = ISDestroy(&isrow);CHKERRQ(ierr); 1501 /* 1502 Everyone has to call to draw the matrix since the graphics waits are 1503 synchronized across all processors that share the PetscDraw object 1504 */ 1505 ierr = PetscViewerGetSubViewer(viewer,PETSC_COMM_SELF,&sviewer);CHKERRQ(ierr); 1506 if (!rank) { 1507 if (((PetscObject)mat)->name) { 1508 ierr = PetscObjectSetName((PetscObject)Av,((PetscObject)mat)->name);CHKERRQ(ierr); 1509 } 1510 ierr = MatView_SeqAIJ(Av,sviewer);CHKERRQ(ierr); 1511 } 1512 ierr = PetscViewerRestoreSubViewer(viewer,PETSC_COMM_SELF,&sviewer);CHKERRQ(ierr); 1513 ierr = PetscViewerFlush(viewer);CHKERRQ(ierr); 1514 ierr = MatDestroy(&A);CHKERRQ(ierr); 1515 } 1516 PetscFunctionReturn(0); 1517 } 1518 1519 PetscErrorCode MatView_MPIAIJ(Mat mat,PetscViewer viewer) 1520 { 1521 PetscErrorCode ierr; 1522 PetscBool iascii,isdraw,issocket,isbinary; 1523 1524 PetscFunctionBegin; 1525 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii);CHKERRQ(ierr); 1526 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERDRAW,&isdraw);CHKERRQ(ierr); 1527 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERBINARY,&isbinary);CHKERRQ(ierr); 1528 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERSOCKET,&issocket);CHKERRQ(ierr); 1529 if (iascii || isdraw || isbinary || issocket) { 1530 ierr = MatView_MPIAIJ_ASCIIorDraworSocket(mat,viewer);CHKERRQ(ierr); 1531 } 1532 PetscFunctionReturn(0); 1533 } 1534 1535 PetscErrorCode MatSOR_MPIAIJ(Mat matin,Vec bb,PetscReal omega,MatSORType flag,PetscReal fshift,PetscInt its,PetscInt lits,Vec xx) 1536 { 1537 Mat_MPIAIJ *mat = (Mat_MPIAIJ*)matin->data; 1538 PetscErrorCode ierr; 1539 Vec bb1 = 0; 1540 PetscBool hasop; 1541 1542 PetscFunctionBegin; 1543 if (flag == SOR_APPLY_UPPER) { 1544 ierr = (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);CHKERRQ(ierr); 1545 PetscFunctionReturn(0); 1546 } 1547 1548 if (its > 1 || ~flag & SOR_ZERO_INITIAL_GUESS || flag & SOR_EISENSTAT) { 1549 ierr = VecDuplicate(bb,&bb1);CHKERRQ(ierr); 1550 } 1551 1552 if ((flag & SOR_LOCAL_SYMMETRIC_SWEEP) == SOR_LOCAL_SYMMETRIC_SWEEP) { 1553 if (flag & SOR_ZERO_INITIAL_GUESS) { 1554 ierr = (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);CHKERRQ(ierr); 1555 its--; 1556 } 1557 1558 while (its--) { 1559 ierr = VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1560 ierr = VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1561 1562 /* update rhs: bb1 = bb - B*x */ 1563 ierr = VecScale(mat->lvec,-1.0);CHKERRQ(ierr); 1564 ierr = (*mat->B->ops->multadd)(mat->B,mat->lvec,bb,bb1);CHKERRQ(ierr); 1565 1566 /* local sweep */ 1567 ierr = (*mat->A->ops->sor)(mat->A,bb1,omega,SOR_SYMMETRIC_SWEEP,fshift,lits,1,xx);CHKERRQ(ierr); 1568 } 1569 } else if (flag & SOR_LOCAL_FORWARD_SWEEP) { 1570 if (flag & SOR_ZERO_INITIAL_GUESS) { 1571 ierr = (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);CHKERRQ(ierr); 1572 its--; 1573 } 1574 while (its--) { 1575 ierr = VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1576 ierr = VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1577 1578 /* update rhs: bb1 = bb - B*x */ 1579 ierr = VecScale(mat->lvec,-1.0);CHKERRQ(ierr); 1580 ierr = (*mat->B->ops->multadd)(mat->B,mat->lvec,bb,bb1);CHKERRQ(ierr); 1581 1582 /* local sweep */ 1583 ierr = (*mat->A->ops->sor)(mat->A,bb1,omega,SOR_FORWARD_SWEEP,fshift,lits,1,xx);CHKERRQ(ierr); 1584 } 1585 } else if (flag & SOR_LOCAL_BACKWARD_SWEEP) { 1586 if (flag & SOR_ZERO_INITIAL_GUESS) { 1587 ierr = (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);CHKERRQ(ierr); 1588 its--; 1589 } 1590 while (its--) { 1591 ierr = VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1592 ierr = VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1593 1594 /* update rhs: bb1 = bb - B*x */ 1595 ierr = VecScale(mat->lvec,-1.0);CHKERRQ(ierr); 1596 ierr = (*mat->B->ops->multadd)(mat->B,mat->lvec,bb,bb1);CHKERRQ(ierr); 1597 1598 /* local sweep */ 1599 ierr = (*mat->A->ops->sor)(mat->A,bb1,omega,SOR_BACKWARD_SWEEP,fshift,lits,1,xx);CHKERRQ(ierr); 1600 } 1601 } else if (flag & SOR_EISENSTAT) { 1602 Vec xx1; 1603 1604 ierr = VecDuplicate(bb,&xx1);CHKERRQ(ierr); 1605 ierr = (*mat->A->ops->sor)(mat->A,bb,omega,(MatSORType)(SOR_ZERO_INITIAL_GUESS | SOR_LOCAL_BACKWARD_SWEEP),fshift,lits,1,xx);CHKERRQ(ierr); 1606 1607 ierr = VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1608 ierr = VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1609 if (!mat->diag) { 1610 ierr = MatCreateVecs(matin,&mat->diag,NULL);CHKERRQ(ierr); 1611 ierr = MatGetDiagonal(matin,mat->diag);CHKERRQ(ierr); 1612 } 1613 ierr = MatHasOperation(matin,MATOP_MULT_DIAGONAL_BLOCK,&hasop);CHKERRQ(ierr); 1614 if (hasop) { 1615 ierr = MatMultDiagonalBlock(matin,xx,bb1);CHKERRQ(ierr); 1616 } else { 1617 ierr = VecPointwiseMult(bb1,mat->diag,xx);CHKERRQ(ierr); 1618 } 1619 ierr = VecAYPX(bb1,(omega-2.0)/omega,bb);CHKERRQ(ierr); 1620 1621 ierr = MatMultAdd(mat->B,mat->lvec,bb1,bb1);CHKERRQ(ierr); 1622 1623 /* local sweep */ 1624 ierr = (*mat->A->ops->sor)(mat->A,bb1,omega,(MatSORType)(SOR_ZERO_INITIAL_GUESS | SOR_LOCAL_FORWARD_SWEEP),fshift,lits,1,xx1);CHKERRQ(ierr); 1625 ierr = VecAXPY(xx,1.0,xx1);CHKERRQ(ierr); 1626 ierr = VecDestroy(&xx1);CHKERRQ(ierr); 1627 } else SETERRQ(PetscObjectComm((PetscObject)matin),PETSC_ERR_SUP,"Parallel SOR not supported"); 1628 1629 ierr = VecDestroy(&bb1);CHKERRQ(ierr); 1630 1631 matin->factorerrortype = mat->A->factorerrortype; 1632 PetscFunctionReturn(0); 1633 } 1634 1635 PetscErrorCode MatPermute_MPIAIJ(Mat A,IS rowp,IS colp,Mat *B) 1636 { 1637 Mat aA,aB,Aperm; 1638 const PetscInt *rwant,*cwant,*gcols,*ai,*bi,*aj,*bj; 1639 PetscScalar *aa,*ba; 1640 PetscInt i,j,m,n,ng,anz,bnz,*dnnz,*onnz,*tdnnz,*tonnz,*rdest,*cdest,*work,*gcdest; 1641 PetscSF rowsf,sf; 1642 IS parcolp = NULL; 1643 PetscBool done; 1644 PetscErrorCode ierr; 1645 1646 PetscFunctionBegin; 1647 ierr = MatGetLocalSize(A,&m,&n);CHKERRQ(ierr); 1648 ierr = ISGetIndices(rowp,&rwant);CHKERRQ(ierr); 1649 ierr = ISGetIndices(colp,&cwant);CHKERRQ(ierr); 1650 ierr = PetscMalloc3(PetscMax(m,n),&work,m,&rdest,n,&cdest);CHKERRQ(ierr); 1651 1652 /* Invert row permutation to find out where my rows should go */ 1653 ierr = PetscSFCreate(PetscObjectComm((PetscObject)A),&rowsf);CHKERRQ(ierr); 1654 ierr = PetscSFSetGraphLayout(rowsf,A->rmap,A->rmap->n,NULL,PETSC_OWN_POINTER,rwant);CHKERRQ(ierr); 1655 ierr = PetscSFSetFromOptions(rowsf);CHKERRQ(ierr); 1656 for (i=0; i<m; i++) work[i] = A->rmap->rstart + i; 1657 ierr = PetscSFReduceBegin(rowsf,MPIU_INT,work,rdest,MPIU_REPLACE);CHKERRQ(ierr); 1658 ierr = PetscSFReduceEnd(rowsf,MPIU_INT,work,rdest,MPIU_REPLACE);CHKERRQ(ierr); 1659 1660 /* Invert column permutation to find out where my columns should go */ 1661 ierr = PetscSFCreate(PetscObjectComm((PetscObject)A),&sf);CHKERRQ(ierr); 1662 ierr = PetscSFSetGraphLayout(sf,A->cmap,A->cmap->n,NULL,PETSC_OWN_POINTER,cwant);CHKERRQ(ierr); 1663 ierr = PetscSFSetFromOptions(sf);CHKERRQ(ierr); 1664 for (i=0; i<n; i++) work[i] = A->cmap->rstart + i; 1665 ierr = PetscSFReduceBegin(sf,MPIU_INT,work,cdest,MPIU_REPLACE);CHKERRQ(ierr); 1666 ierr = PetscSFReduceEnd(sf,MPIU_INT,work,cdest,MPIU_REPLACE);CHKERRQ(ierr); 1667 ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 1668 1669 ierr = ISRestoreIndices(rowp,&rwant);CHKERRQ(ierr); 1670 ierr = ISRestoreIndices(colp,&cwant);CHKERRQ(ierr); 1671 ierr = MatMPIAIJGetSeqAIJ(A,&aA,&aB,&gcols);CHKERRQ(ierr); 1672 1673 /* Find out where my gcols should go */ 1674 ierr = MatGetSize(aB,NULL,&ng);CHKERRQ(ierr); 1675 ierr = PetscMalloc1(ng,&gcdest);CHKERRQ(ierr); 1676 ierr = PetscSFCreate(PetscObjectComm((PetscObject)A),&sf);CHKERRQ(ierr); 1677 ierr = PetscSFSetGraphLayout(sf,A->cmap,ng,NULL,PETSC_OWN_POINTER,gcols);CHKERRQ(ierr); 1678 ierr = PetscSFSetFromOptions(sf);CHKERRQ(ierr); 1679 ierr = PetscSFBcastBegin(sf,MPIU_INT,cdest,gcdest);CHKERRQ(ierr); 1680 ierr = PetscSFBcastEnd(sf,MPIU_INT,cdest,gcdest);CHKERRQ(ierr); 1681 ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 1682 1683 ierr = PetscCalloc4(m,&dnnz,m,&onnz,m,&tdnnz,m,&tonnz);CHKERRQ(ierr); 1684 ierr = MatGetRowIJ(aA,0,PETSC_FALSE,PETSC_FALSE,&anz,&ai,&aj,&done);CHKERRQ(ierr); 1685 ierr = MatGetRowIJ(aB,0,PETSC_FALSE,PETSC_FALSE,&bnz,&bi,&bj,&done);CHKERRQ(ierr); 1686 for (i=0; i<m; i++) { 1687 PetscInt row = rdest[i],rowner; 1688 ierr = PetscLayoutFindOwner(A->rmap,row,&rowner);CHKERRQ(ierr); 1689 for (j=ai[i]; j<ai[i+1]; j++) { 1690 PetscInt cowner,col = cdest[aj[j]]; 1691 ierr = PetscLayoutFindOwner(A->cmap,col,&cowner);CHKERRQ(ierr); /* Could build an index for the columns to eliminate this search */ 1692 if (rowner == cowner) dnnz[i]++; 1693 else onnz[i]++; 1694 } 1695 for (j=bi[i]; j<bi[i+1]; j++) { 1696 PetscInt cowner,col = gcdest[bj[j]]; 1697 ierr = PetscLayoutFindOwner(A->cmap,col,&cowner);CHKERRQ(ierr); 1698 if (rowner == cowner) dnnz[i]++; 1699 else onnz[i]++; 1700 } 1701 } 1702 ierr = PetscSFBcastBegin(rowsf,MPIU_INT,dnnz,tdnnz);CHKERRQ(ierr); 1703 ierr = PetscSFBcastEnd(rowsf,MPIU_INT,dnnz,tdnnz);CHKERRQ(ierr); 1704 ierr = PetscSFBcastBegin(rowsf,MPIU_INT,onnz,tonnz);CHKERRQ(ierr); 1705 ierr = PetscSFBcastEnd(rowsf,MPIU_INT,onnz,tonnz);CHKERRQ(ierr); 1706 ierr = PetscSFDestroy(&rowsf);CHKERRQ(ierr); 1707 1708 ierr = MatCreateAIJ(PetscObjectComm((PetscObject)A),A->rmap->n,A->cmap->n,A->rmap->N,A->cmap->N,0,tdnnz,0,tonnz,&Aperm);CHKERRQ(ierr); 1709 ierr = MatSeqAIJGetArray(aA,&aa);CHKERRQ(ierr); 1710 ierr = MatSeqAIJGetArray(aB,&ba);CHKERRQ(ierr); 1711 for (i=0; i<m; i++) { 1712 PetscInt *acols = dnnz,*bcols = onnz; /* Repurpose now-unneeded arrays */ 1713 PetscInt j0,rowlen; 1714 rowlen = ai[i+1] - ai[i]; 1715 for (j0=j=0; j<rowlen; j0=j) { /* rowlen could be larger than number of rows m, so sum in batches */ 1716 for ( ; j<PetscMin(rowlen,j0+m); j++) acols[j-j0] = cdest[aj[ai[i]+j]]; 1717 ierr = MatSetValues(Aperm,1,&rdest[i],j-j0,acols,aa+ai[i]+j0,INSERT_VALUES);CHKERRQ(ierr); 1718 } 1719 rowlen = bi[i+1] - bi[i]; 1720 for (j0=j=0; j<rowlen; j0=j) { 1721 for ( ; j<PetscMin(rowlen,j0+m); j++) bcols[j-j0] = gcdest[bj[bi[i]+j]]; 1722 ierr = MatSetValues(Aperm,1,&rdest[i],j-j0,bcols,ba+bi[i]+j0,INSERT_VALUES);CHKERRQ(ierr); 1723 } 1724 } 1725 ierr = MatAssemblyBegin(Aperm,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1726 ierr = MatAssemblyEnd(Aperm,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1727 ierr = MatRestoreRowIJ(aA,0,PETSC_FALSE,PETSC_FALSE,&anz,&ai,&aj,&done);CHKERRQ(ierr); 1728 ierr = MatRestoreRowIJ(aB,0,PETSC_FALSE,PETSC_FALSE,&bnz,&bi,&bj,&done);CHKERRQ(ierr); 1729 ierr = MatSeqAIJRestoreArray(aA,&aa);CHKERRQ(ierr); 1730 ierr = MatSeqAIJRestoreArray(aB,&ba);CHKERRQ(ierr); 1731 ierr = PetscFree4(dnnz,onnz,tdnnz,tonnz);CHKERRQ(ierr); 1732 ierr = PetscFree3(work,rdest,cdest);CHKERRQ(ierr); 1733 ierr = PetscFree(gcdest);CHKERRQ(ierr); 1734 if (parcolp) {ierr = ISDestroy(&colp);CHKERRQ(ierr);} 1735 *B = Aperm; 1736 PetscFunctionReturn(0); 1737 } 1738 1739 PetscErrorCode MatGetGhosts_MPIAIJ(Mat mat,PetscInt *nghosts,const PetscInt *ghosts[]) 1740 { 1741 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 1742 PetscErrorCode ierr; 1743 1744 PetscFunctionBegin; 1745 ierr = MatGetSize(aij->B,NULL,nghosts);CHKERRQ(ierr); 1746 if (ghosts) *ghosts = aij->garray; 1747 PetscFunctionReturn(0); 1748 } 1749 1750 PetscErrorCode MatGetInfo_MPIAIJ(Mat matin,MatInfoType flag,MatInfo *info) 1751 { 1752 Mat_MPIAIJ *mat = (Mat_MPIAIJ*)matin->data; 1753 Mat A = mat->A,B = mat->B; 1754 PetscErrorCode ierr; 1755 PetscReal isend[5],irecv[5]; 1756 1757 PetscFunctionBegin; 1758 info->block_size = 1.0; 1759 ierr = MatGetInfo(A,MAT_LOCAL,info);CHKERRQ(ierr); 1760 1761 isend[0] = info->nz_used; isend[1] = info->nz_allocated; isend[2] = info->nz_unneeded; 1762 isend[3] = info->memory; isend[4] = info->mallocs; 1763 1764 ierr = MatGetInfo(B,MAT_LOCAL,info);CHKERRQ(ierr); 1765 1766 isend[0] += info->nz_used; isend[1] += info->nz_allocated; isend[2] += info->nz_unneeded; 1767 isend[3] += info->memory; isend[4] += info->mallocs; 1768 if (flag == MAT_LOCAL) { 1769 info->nz_used = isend[0]; 1770 info->nz_allocated = isend[1]; 1771 info->nz_unneeded = isend[2]; 1772 info->memory = isend[3]; 1773 info->mallocs = isend[4]; 1774 } else if (flag == MAT_GLOBAL_MAX) { 1775 ierr = MPIU_Allreduce(isend,irecv,5,MPIU_REAL,MPIU_MAX,PetscObjectComm((PetscObject)matin));CHKERRQ(ierr); 1776 1777 info->nz_used = irecv[0]; 1778 info->nz_allocated = irecv[1]; 1779 info->nz_unneeded = irecv[2]; 1780 info->memory = irecv[3]; 1781 info->mallocs = irecv[4]; 1782 } else if (flag == MAT_GLOBAL_SUM) { 1783 ierr = MPIU_Allreduce(isend,irecv,5,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)matin));CHKERRQ(ierr); 1784 1785 info->nz_used = irecv[0]; 1786 info->nz_allocated = irecv[1]; 1787 info->nz_unneeded = irecv[2]; 1788 info->memory = irecv[3]; 1789 info->mallocs = irecv[4]; 1790 } 1791 info->fill_ratio_given = 0; /* no parallel LU/ILU/Cholesky */ 1792 info->fill_ratio_needed = 0; 1793 info->factor_mallocs = 0; 1794 PetscFunctionReturn(0); 1795 } 1796 1797 PetscErrorCode MatSetOption_MPIAIJ(Mat A,MatOption op,PetscBool flg) 1798 { 1799 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 1800 PetscErrorCode ierr; 1801 1802 PetscFunctionBegin; 1803 switch (op) { 1804 case MAT_NEW_NONZERO_LOCATIONS: 1805 case MAT_NEW_NONZERO_ALLOCATION_ERR: 1806 case MAT_UNUSED_NONZERO_LOCATION_ERR: 1807 case MAT_KEEP_NONZERO_PATTERN: 1808 case MAT_NEW_NONZERO_LOCATION_ERR: 1809 case MAT_USE_INODES: 1810 case MAT_IGNORE_ZERO_ENTRIES: 1811 MatCheckPreallocated(A,1); 1812 ierr = MatSetOption(a->A,op,flg);CHKERRQ(ierr); 1813 ierr = MatSetOption(a->B,op,flg);CHKERRQ(ierr); 1814 break; 1815 case MAT_ROW_ORIENTED: 1816 MatCheckPreallocated(A,1); 1817 a->roworiented = flg; 1818 1819 ierr = MatSetOption(a->A,op,flg);CHKERRQ(ierr); 1820 ierr = MatSetOption(a->B,op,flg);CHKERRQ(ierr); 1821 break; 1822 case MAT_NEW_DIAGONALS: 1823 case MAT_SORTED_FULL: 1824 ierr = PetscInfo1(A,"Option %s ignored\n",MatOptions[op]);CHKERRQ(ierr); 1825 break; 1826 case MAT_IGNORE_OFF_PROC_ENTRIES: 1827 a->donotstash = flg; 1828 break; 1829 /* Symmetry flags are handled directly by MatSetOption() and they don't affect preallocation */ 1830 case MAT_SPD: 1831 case MAT_SYMMETRIC: 1832 case MAT_STRUCTURALLY_SYMMETRIC: 1833 case MAT_HERMITIAN: 1834 case MAT_SYMMETRY_ETERNAL: 1835 break; 1836 case MAT_SUBMAT_SINGLEIS: 1837 A->submat_singleis = flg; 1838 break; 1839 case MAT_STRUCTURE_ONLY: 1840 /* The option is handled directly by MatSetOption() */ 1841 break; 1842 default: 1843 SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"unknown option %d",op); 1844 } 1845 PetscFunctionReturn(0); 1846 } 1847 1848 PetscErrorCode MatGetRow_MPIAIJ(Mat matin,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v) 1849 { 1850 Mat_MPIAIJ *mat = (Mat_MPIAIJ*)matin->data; 1851 PetscScalar *vworkA,*vworkB,**pvA,**pvB,*v_p; 1852 PetscErrorCode ierr; 1853 PetscInt i,*cworkA,*cworkB,**pcA,**pcB,cstart = matin->cmap->rstart; 1854 PetscInt nztot,nzA,nzB,lrow,rstart = matin->rmap->rstart,rend = matin->rmap->rend; 1855 PetscInt *cmap,*idx_p; 1856 1857 PetscFunctionBegin; 1858 if (mat->getrowactive) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Already active"); 1859 mat->getrowactive = PETSC_TRUE; 1860 1861 if (!mat->rowvalues && (idx || v)) { 1862 /* 1863 allocate enough space to hold information from the longest row. 1864 */ 1865 Mat_SeqAIJ *Aa = (Mat_SeqAIJ*)mat->A->data,*Ba = (Mat_SeqAIJ*)mat->B->data; 1866 PetscInt max = 1,tmp; 1867 for (i=0; i<matin->rmap->n; i++) { 1868 tmp = Aa->i[i+1] - Aa->i[i] + Ba->i[i+1] - Ba->i[i]; 1869 if (max < tmp) max = tmp; 1870 } 1871 ierr = PetscMalloc2(max,&mat->rowvalues,max,&mat->rowindices);CHKERRQ(ierr); 1872 } 1873 1874 if (row < rstart || row >= rend) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Only local rows"); 1875 lrow = row - rstart; 1876 1877 pvA = &vworkA; pcA = &cworkA; pvB = &vworkB; pcB = &cworkB; 1878 if (!v) {pvA = 0; pvB = 0;} 1879 if (!idx) {pcA = 0; if (!v) pcB = 0;} 1880 ierr = (*mat->A->ops->getrow)(mat->A,lrow,&nzA,pcA,pvA);CHKERRQ(ierr); 1881 ierr = (*mat->B->ops->getrow)(mat->B,lrow,&nzB,pcB,pvB);CHKERRQ(ierr); 1882 nztot = nzA + nzB; 1883 1884 cmap = mat->garray; 1885 if (v || idx) { 1886 if (nztot) { 1887 /* Sort by increasing column numbers, assuming A and B already sorted */ 1888 PetscInt imark = -1; 1889 if (v) { 1890 *v = v_p = mat->rowvalues; 1891 for (i=0; i<nzB; i++) { 1892 if (cmap[cworkB[i]] < cstart) v_p[i] = vworkB[i]; 1893 else break; 1894 } 1895 imark = i; 1896 for (i=0; i<nzA; i++) v_p[imark+i] = vworkA[i]; 1897 for (i=imark; i<nzB; i++) v_p[nzA+i] = vworkB[i]; 1898 } 1899 if (idx) { 1900 *idx = idx_p = mat->rowindices; 1901 if (imark > -1) { 1902 for (i=0; i<imark; i++) { 1903 idx_p[i] = cmap[cworkB[i]]; 1904 } 1905 } else { 1906 for (i=0; i<nzB; i++) { 1907 if (cmap[cworkB[i]] < cstart) idx_p[i] = cmap[cworkB[i]]; 1908 else break; 1909 } 1910 imark = i; 1911 } 1912 for (i=0; i<nzA; i++) idx_p[imark+i] = cstart + cworkA[i]; 1913 for (i=imark; i<nzB; i++) idx_p[nzA+i] = cmap[cworkB[i]]; 1914 } 1915 } else { 1916 if (idx) *idx = 0; 1917 if (v) *v = 0; 1918 } 1919 } 1920 *nz = nztot; 1921 ierr = (*mat->A->ops->restorerow)(mat->A,lrow,&nzA,pcA,pvA);CHKERRQ(ierr); 1922 ierr = (*mat->B->ops->restorerow)(mat->B,lrow,&nzB,pcB,pvB);CHKERRQ(ierr); 1923 PetscFunctionReturn(0); 1924 } 1925 1926 PetscErrorCode MatRestoreRow_MPIAIJ(Mat mat,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v) 1927 { 1928 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 1929 1930 PetscFunctionBegin; 1931 if (!aij->getrowactive) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"MatGetRow() must be called first"); 1932 aij->getrowactive = PETSC_FALSE; 1933 PetscFunctionReturn(0); 1934 } 1935 1936 PetscErrorCode MatNorm_MPIAIJ(Mat mat,NormType type,PetscReal *norm) 1937 { 1938 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 1939 Mat_SeqAIJ *amat = (Mat_SeqAIJ*)aij->A->data,*bmat = (Mat_SeqAIJ*)aij->B->data; 1940 PetscErrorCode ierr; 1941 PetscInt i,j,cstart = mat->cmap->rstart; 1942 PetscReal sum = 0.0; 1943 MatScalar *v; 1944 1945 PetscFunctionBegin; 1946 if (aij->size == 1) { 1947 ierr = MatNorm(aij->A,type,norm);CHKERRQ(ierr); 1948 } else { 1949 if (type == NORM_FROBENIUS) { 1950 v = amat->a; 1951 for (i=0; i<amat->nz; i++) { 1952 sum += PetscRealPart(PetscConj(*v)*(*v)); v++; 1953 } 1954 v = bmat->a; 1955 for (i=0; i<bmat->nz; i++) { 1956 sum += PetscRealPart(PetscConj(*v)*(*v)); v++; 1957 } 1958 ierr = MPIU_Allreduce(&sum,norm,1,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1959 *norm = PetscSqrtReal(*norm); 1960 ierr = PetscLogFlops(2*amat->nz+2*bmat->nz);CHKERRQ(ierr); 1961 } else if (type == NORM_1) { /* max column norm */ 1962 PetscReal *tmp,*tmp2; 1963 PetscInt *jj,*garray = aij->garray; 1964 ierr = PetscCalloc1(mat->cmap->N+1,&tmp);CHKERRQ(ierr); 1965 ierr = PetscMalloc1(mat->cmap->N+1,&tmp2);CHKERRQ(ierr); 1966 *norm = 0.0; 1967 v = amat->a; jj = amat->j; 1968 for (j=0; j<amat->nz; j++) { 1969 tmp[cstart + *jj++] += PetscAbsScalar(*v); v++; 1970 } 1971 v = bmat->a; jj = bmat->j; 1972 for (j=0; j<bmat->nz; j++) { 1973 tmp[garray[*jj++]] += PetscAbsScalar(*v); v++; 1974 } 1975 ierr = MPIU_Allreduce(tmp,tmp2,mat->cmap->N,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1976 for (j=0; j<mat->cmap->N; j++) { 1977 if (tmp2[j] > *norm) *norm = tmp2[j]; 1978 } 1979 ierr = PetscFree(tmp);CHKERRQ(ierr); 1980 ierr = PetscFree(tmp2);CHKERRQ(ierr); 1981 ierr = PetscLogFlops(PetscMax(amat->nz+bmat->nz-1,0));CHKERRQ(ierr); 1982 } else if (type == NORM_INFINITY) { /* max row norm */ 1983 PetscReal ntemp = 0.0; 1984 for (j=0; j<aij->A->rmap->n; j++) { 1985 v = amat->a + amat->i[j]; 1986 sum = 0.0; 1987 for (i=0; i<amat->i[j+1]-amat->i[j]; i++) { 1988 sum += PetscAbsScalar(*v); v++; 1989 } 1990 v = bmat->a + bmat->i[j]; 1991 for (i=0; i<bmat->i[j+1]-bmat->i[j]; i++) { 1992 sum += PetscAbsScalar(*v); v++; 1993 } 1994 if (sum > ntemp) ntemp = sum; 1995 } 1996 ierr = MPIU_Allreduce(&ntemp,norm,1,MPIU_REAL,MPIU_MAX,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1997 ierr = PetscLogFlops(PetscMax(amat->nz+bmat->nz-1,0));CHKERRQ(ierr); 1998 } else SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"No support for two norm"); 1999 } 2000 PetscFunctionReturn(0); 2001 } 2002 2003 PetscErrorCode MatTranspose_MPIAIJ(Mat A,MatReuse reuse,Mat *matout) 2004 { 2005 Mat_MPIAIJ *a =(Mat_MPIAIJ*)A->data,*b; 2006 Mat_SeqAIJ *Aloc =(Mat_SeqAIJ*)a->A->data,*Bloc=(Mat_SeqAIJ*)a->B->data,*sub_B_diag; 2007 PetscInt M = A->rmap->N,N=A->cmap->N,ma,na,mb,nb,row,*cols,*cols_tmp,*B_diag_ilen,i,ncol,A_diag_ncol; 2008 const PetscInt *ai,*aj,*bi,*bj,*B_diag_i; 2009 PetscErrorCode ierr; 2010 Mat B,A_diag,*B_diag; 2011 const MatScalar *array; 2012 2013 PetscFunctionBegin; 2014 ma = A->rmap->n; na = A->cmap->n; mb = a->B->rmap->n; nb = a->B->cmap->n; 2015 ai = Aloc->i; aj = Aloc->j; 2016 bi = Bloc->i; bj = Bloc->j; 2017 if (reuse == MAT_INITIAL_MATRIX || *matout == A) { 2018 PetscInt *d_nnz,*g_nnz,*o_nnz; 2019 PetscSFNode *oloc; 2020 PETSC_UNUSED PetscSF sf; 2021 2022 ierr = PetscMalloc4(na,&d_nnz,na,&o_nnz,nb,&g_nnz,nb,&oloc);CHKERRQ(ierr); 2023 /* compute d_nnz for preallocation */ 2024 ierr = PetscArrayzero(d_nnz,na);CHKERRQ(ierr); 2025 for (i=0; i<ai[ma]; i++) { 2026 d_nnz[aj[i]]++; 2027 } 2028 /* compute local off-diagonal contributions */ 2029 ierr = PetscArrayzero(g_nnz,nb);CHKERRQ(ierr); 2030 for (i=0; i<bi[ma]; i++) g_nnz[bj[i]]++; 2031 /* map those to global */ 2032 ierr = PetscSFCreate(PetscObjectComm((PetscObject)A),&sf);CHKERRQ(ierr); 2033 ierr = PetscSFSetGraphLayout(sf,A->cmap,nb,NULL,PETSC_USE_POINTER,a->garray);CHKERRQ(ierr); 2034 ierr = PetscSFSetFromOptions(sf);CHKERRQ(ierr); 2035 ierr = PetscArrayzero(o_nnz,na);CHKERRQ(ierr); 2036 ierr = PetscSFReduceBegin(sf,MPIU_INT,g_nnz,o_nnz,MPIU_SUM);CHKERRQ(ierr); 2037 ierr = PetscSFReduceEnd(sf,MPIU_INT,g_nnz,o_nnz,MPIU_SUM);CHKERRQ(ierr); 2038 ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 2039 2040 ierr = MatCreate(PetscObjectComm((PetscObject)A),&B);CHKERRQ(ierr); 2041 ierr = MatSetSizes(B,A->cmap->n,A->rmap->n,N,M);CHKERRQ(ierr); 2042 ierr = MatSetBlockSizes(B,PetscAbs(A->cmap->bs),PetscAbs(A->rmap->bs));CHKERRQ(ierr); 2043 ierr = MatSetType(B,((PetscObject)A)->type_name);CHKERRQ(ierr); 2044 ierr = MatMPIAIJSetPreallocation(B,0,d_nnz,0,o_nnz);CHKERRQ(ierr); 2045 ierr = PetscFree4(d_nnz,o_nnz,g_nnz,oloc);CHKERRQ(ierr); 2046 } else { 2047 B = *matout; 2048 ierr = MatSetOption(B,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_TRUE);CHKERRQ(ierr); 2049 } 2050 2051 b = (Mat_MPIAIJ*)B->data; 2052 A_diag = a->A; 2053 B_diag = &b->A; 2054 sub_B_diag = (Mat_SeqAIJ*)(*B_diag)->data; 2055 A_diag_ncol = A_diag->cmap->N; 2056 B_diag_ilen = sub_B_diag->ilen; 2057 B_diag_i = sub_B_diag->i; 2058 2059 /* Set ilen for diagonal of B */ 2060 for (i=0; i<A_diag_ncol; i++) { 2061 B_diag_ilen[i] = B_diag_i[i+1] - B_diag_i[i]; 2062 } 2063 2064 /* Transpose the diagonal part of the matrix. In contrast to the offdiagonal part, this can be done 2065 very quickly (=without using MatSetValues), because all writes are local. */ 2066 ierr = MatTranspose(A_diag,MAT_REUSE_MATRIX,B_diag);CHKERRQ(ierr); 2067 2068 /* copy over the B part */ 2069 ierr = PetscMalloc1(bi[mb],&cols);CHKERRQ(ierr); 2070 array = Bloc->a; 2071 row = A->rmap->rstart; 2072 for (i=0; i<bi[mb]; i++) cols[i] = a->garray[bj[i]]; 2073 cols_tmp = cols; 2074 for (i=0; i<mb; i++) { 2075 ncol = bi[i+1]-bi[i]; 2076 ierr = MatSetValues(B,ncol,cols_tmp,1,&row,array,INSERT_VALUES);CHKERRQ(ierr); 2077 row++; 2078 array += ncol; cols_tmp += ncol; 2079 } 2080 ierr = PetscFree(cols);CHKERRQ(ierr); 2081 2082 ierr = MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2083 ierr = MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2084 if (reuse == MAT_INITIAL_MATRIX || reuse == MAT_REUSE_MATRIX) { 2085 *matout = B; 2086 } else { 2087 ierr = MatHeaderMerge(A,&B);CHKERRQ(ierr); 2088 } 2089 PetscFunctionReturn(0); 2090 } 2091 2092 PetscErrorCode MatDiagonalScale_MPIAIJ(Mat mat,Vec ll,Vec rr) 2093 { 2094 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 2095 Mat a = aij->A,b = aij->B; 2096 PetscErrorCode ierr; 2097 PetscInt s1,s2,s3; 2098 2099 PetscFunctionBegin; 2100 ierr = MatGetLocalSize(mat,&s2,&s3);CHKERRQ(ierr); 2101 if (rr) { 2102 ierr = VecGetLocalSize(rr,&s1);CHKERRQ(ierr); 2103 if (s1!=s3) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"right vector non-conforming local size"); 2104 /* Overlap communication with computation. */ 2105 ierr = VecScatterBegin(aij->Mvctx,rr,aij->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2106 } 2107 if (ll) { 2108 ierr = VecGetLocalSize(ll,&s1);CHKERRQ(ierr); 2109 if (s1!=s2) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"left vector non-conforming local size"); 2110 ierr = (*b->ops->diagonalscale)(b,ll,0);CHKERRQ(ierr); 2111 } 2112 /* scale the diagonal block */ 2113 ierr = (*a->ops->diagonalscale)(a,ll,rr);CHKERRQ(ierr); 2114 2115 if (rr) { 2116 /* Do a scatter end and then right scale the off-diagonal block */ 2117 ierr = VecScatterEnd(aij->Mvctx,rr,aij->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2118 ierr = (*b->ops->diagonalscale)(b,0,aij->lvec);CHKERRQ(ierr); 2119 } 2120 PetscFunctionReturn(0); 2121 } 2122 2123 PetscErrorCode MatSetUnfactored_MPIAIJ(Mat A) 2124 { 2125 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 2126 PetscErrorCode ierr; 2127 2128 PetscFunctionBegin; 2129 ierr = MatSetUnfactored(a->A);CHKERRQ(ierr); 2130 PetscFunctionReturn(0); 2131 } 2132 2133 PetscErrorCode MatEqual_MPIAIJ(Mat A,Mat B,PetscBool *flag) 2134 { 2135 Mat_MPIAIJ *matB = (Mat_MPIAIJ*)B->data,*matA = (Mat_MPIAIJ*)A->data; 2136 Mat a,b,c,d; 2137 PetscBool flg; 2138 PetscErrorCode ierr; 2139 2140 PetscFunctionBegin; 2141 a = matA->A; b = matA->B; 2142 c = matB->A; d = matB->B; 2143 2144 ierr = MatEqual(a,c,&flg);CHKERRQ(ierr); 2145 if (flg) { 2146 ierr = MatEqual(b,d,&flg);CHKERRQ(ierr); 2147 } 2148 ierr = MPIU_Allreduce(&flg,flag,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)A));CHKERRQ(ierr); 2149 PetscFunctionReturn(0); 2150 } 2151 2152 PetscErrorCode MatCopy_MPIAIJ(Mat A,Mat B,MatStructure str) 2153 { 2154 PetscErrorCode ierr; 2155 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 2156 Mat_MPIAIJ *b = (Mat_MPIAIJ*)B->data; 2157 2158 PetscFunctionBegin; 2159 /* If the two matrices don't have the same copy implementation, they aren't compatible for fast copy. */ 2160 if ((str != SAME_NONZERO_PATTERN) || (A->ops->copy != B->ops->copy)) { 2161 /* because of the column compression in the off-processor part of the matrix a->B, 2162 the number of columns in a->B and b->B may be different, hence we cannot call 2163 the MatCopy() directly on the two parts. If need be, we can provide a more 2164 efficient copy than the MatCopy_Basic() by first uncompressing the a->B matrices 2165 then copying the submatrices */ 2166 ierr = MatCopy_Basic(A,B,str);CHKERRQ(ierr); 2167 } else { 2168 ierr = MatCopy(a->A,b->A,str);CHKERRQ(ierr); 2169 ierr = MatCopy(a->B,b->B,str);CHKERRQ(ierr); 2170 } 2171 ierr = PetscObjectStateIncrease((PetscObject)B);CHKERRQ(ierr); 2172 PetscFunctionReturn(0); 2173 } 2174 2175 PetscErrorCode MatSetUp_MPIAIJ(Mat A) 2176 { 2177 PetscErrorCode ierr; 2178 2179 PetscFunctionBegin; 2180 ierr = MatMPIAIJSetPreallocation(A,PETSC_DEFAULT,0,PETSC_DEFAULT,0);CHKERRQ(ierr); 2181 PetscFunctionReturn(0); 2182 } 2183 2184 /* 2185 Computes the number of nonzeros per row needed for preallocation when X and Y 2186 have different nonzero structure. 2187 */ 2188 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) 2189 { 2190 PetscInt i,j,k,nzx,nzy; 2191 2192 PetscFunctionBegin; 2193 /* Set the number of nonzeros in the new matrix */ 2194 for (i=0; i<m; i++) { 2195 const PetscInt *xjj = xj+xi[i],*yjj = yj+yi[i]; 2196 nzx = xi[i+1] - xi[i]; 2197 nzy = yi[i+1] - yi[i]; 2198 nnz[i] = 0; 2199 for (j=0,k=0; j<nzx; j++) { /* Point in X */ 2200 for (; k<nzy && yltog[yjj[k]]<xltog[xjj[j]]; k++) nnz[i]++; /* Catch up to X */ 2201 if (k<nzy && yltog[yjj[k]]==xltog[xjj[j]]) k++; /* Skip duplicate */ 2202 nnz[i]++; 2203 } 2204 for (; k<nzy; k++) nnz[i]++; 2205 } 2206 PetscFunctionReturn(0); 2207 } 2208 2209 /* This is the same as MatAXPYGetPreallocation_SeqAIJ, except that the local-to-global map is provided */ 2210 static PetscErrorCode MatAXPYGetPreallocation_MPIAIJ(Mat Y,const PetscInt *yltog,Mat X,const PetscInt *xltog,PetscInt *nnz) 2211 { 2212 PetscErrorCode ierr; 2213 PetscInt m = Y->rmap->N; 2214 Mat_SeqAIJ *x = (Mat_SeqAIJ*)X->data; 2215 Mat_SeqAIJ *y = (Mat_SeqAIJ*)Y->data; 2216 2217 PetscFunctionBegin; 2218 ierr = MatAXPYGetPreallocation_MPIX_private(m,x->i,x->j,xltog,y->i,y->j,yltog,nnz);CHKERRQ(ierr); 2219 PetscFunctionReturn(0); 2220 } 2221 2222 PetscErrorCode MatAXPY_MPIAIJ(Mat Y,PetscScalar a,Mat X,MatStructure str) 2223 { 2224 PetscErrorCode ierr; 2225 Mat_MPIAIJ *xx = (Mat_MPIAIJ*)X->data,*yy = (Mat_MPIAIJ*)Y->data; 2226 PetscBLASInt bnz,one=1; 2227 Mat_SeqAIJ *x,*y; 2228 2229 PetscFunctionBegin; 2230 if (str == SAME_NONZERO_PATTERN) { 2231 PetscScalar alpha = a; 2232 x = (Mat_SeqAIJ*)xx->A->data; 2233 ierr = PetscBLASIntCast(x->nz,&bnz);CHKERRQ(ierr); 2234 y = (Mat_SeqAIJ*)yy->A->data; 2235 PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&bnz,&alpha,x->a,&one,y->a,&one)); 2236 x = (Mat_SeqAIJ*)xx->B->data; 2237 y = (Mat_SeqAIJ*)yy->B->data; 2238 ierr = PetscBLASIntCast(x->nz,&bnz);CHKERRQ(ierr); 2239 PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&bnz,&alpha,x->a,&one,y->a,&one)); 2240 ierr = PetscObjectStateIncrease((PetscObject)Y);CHKERRQ(ierr); 2241 } else if (str == SUBSET_NONZERO_PATTERN) { /* nonzeros of X is a subset of Y's */ 2242 ierr = MatAXPY_Basic(Y,a,X,str);CHKERRQ(ierr); 2243 } else { 2244 Mat B; 2245 PetscInt *nnz_d,*nnz_o; 2246 ierr = PetscMalloc1(yy->A->rmap->N,&nnz_d);CHKERRQ(ierr); 2247 ierr = PetscMalloc1(yy->B->rmap->N,&nnz_o);CHKERRQ(ierr); 2248 ierr = MatCreate(PetscObjectComm((PetscObject)Y),&B);CHKERRQ(ierr); 2249 ierr = PetscObjectSetName((PetscObject)B,((PetscObject)Y)->name);CHKERRQ(ierr); 2250 ierr = MatSetSizes(B,Y->rmap->n,Y->cmap->n,Y->rmap->N,Y->cmap->N);CHKERRQ(ierr); 2251 ierr = MatSetBlockSizesFromMats(B,Y,Y);CHKERRQ(ierr); 2252 ierr = MatSetType(B,MATMPIAIJ);CHKERRQ(ierr); 2253 ierr = MatAXPYGetPreallocation_SeqAIJ(yy->A,xx->A,nnz_d);CHKERRQ(ierr); 2254 ierr = MatAXPYGetPreallocation_MPIAIJ(yy->B,yy->garray,xx->B,xx->garray,nnz_o);CHKERRQ(ierr); 2255 ierr = MatMPIAIJSetPreallocation(B,0,nnz_d,0,nnz_o);CHKERRQ(ierr); 2256 ierr = MatAXPY_BasicWithPreallocation(B,Y,a,X,str);CHKERRQ(ierr); 2257 ierr = MatHeaderReplace(Y,&B);CHKERRQ(ierr); 2258 ierr = PetscFree(nnz_d);CHKERRQ(ierr); 2259 ierr = PetscFree(nnz_o);CHKERRQ(ierr); 2260 } 2261 PetscFunctionReturn(0); 2262 } 2263 2264 extern PetscErrorCode MatConjugate_SeqAIJ(Mat); 2265 2266 PetscErrorCode MatConjugate_MPIAIJ(Mat mat) 2267 { 2268 #if defined(PETSC_USE_COMPLEX) 2269 PetscErrorCode ierr; 2270 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 2271 2272 PetscFunctionBegin; 2273 ierr = MatConjugate_SeqAIJ(aij->A);CHKERRQ(ierr); 2274 ierr = MatConjugate_SeqAIJ(aij->B);CHKERRQ(ierr); 2275 #else 2276 PetscFunctionBegin; 2277 #endif 2278 PetscFunctionReturn(0); 2279 } 2280 2281 PetscErrorCode MatRealPart_MPIAIJ(Mat A) 2282 { 2283 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 2284 PetscErrorCode ierr; 2285 2286 PetscFunctionBegin; 2287 ierr = MatRealPart(a->A);CHKERRQ(ierr); 2288 ierr = MatRealPart(a->B);CHKERRQ(ierr); 2289 PetscFunctionReturn(0); 2290 } 2291 2292 PetscErrorCode MatImaginaryPart_MPIAIJ(Mat A) 2293 { 2294 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 2295 PetscErrorCode ierr; 2296 2297 PetscFunctionBegin; 2298 ierr = MatImaginaryPart(a->A);CHKERRQ(ierr); 2299 ierr = MatImaginaryPart(a->B);CHKERRQ(ierr); 2300 PetscFunctionReturn(0); 2301 } 2302 2303 PetscErrorCode MatGetRowMaxAbs_MPIAIJ(Mat A, Vec v, PetscInt idx[]) 2304 { 2305 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 2306 PetscErrorCode ierr; 2307 PetscInt i,*idxb = 0; 2308 PetscScalar *va,*vb; 2309 Vec vtmp; 2310 2311 PetscFunctionBegin; 2312 ierr = MatGetRowMaxAbs(a->A,v,idx);CHKERRQ(ierr); 2313 ierr = VecGetArray(v,&va);CHKERRQ(ierr); 2314 if (idx) { 2315 for (i=0; i<A->rmap->n; i++) { 2316 if (PetscAbsScalar(va[i])) idx[i] += A->cmap->rstart; 2317 } 2318 } 2319 2320 ierr = VecCreateSeq(PETSC_COMM_SELF,A->rmap->n,&vtmp);CHKERRQ(ierr); 2321 if (idx) { 2322 ierr = PetscMalloc1(A->rmap->n,&idxb);CHKERRQ(ierr); 2323 } 2324 ierr = MatGetRowMaxAbs(a->B,vtmp,idxb);CHKERRQ(ierr); 2325 ierr = VecGetArray(vtmp,&vb);CHKERRQ(ierr); 2326 2327 for (i=0; i<A->rmap->n; i++) { 2328 if (PetscAbsScalar(va[i]) < PetscAbsScalar(vb[i])) { 2329 va[i] = vb[i]; 2330 if (idx) idx[i] = a->garray[idxb[i]]; 2331 } 2332 } 2333 2334 ierr = VecRestoreArray(v,&va);CHKERRQ(ierr); 2335 ierr = VecRestoreArray(vtmp,&vb);CHKERRQ(ierr); 2336 ierr = PetscFree(idxb);CHKERRQ(ierr); 2337 ierr = VecDestroy(&vtmp);CHKERRQ(ierr); 2338 PetscFunctionReturn(0); 2339 } 2340 2341 PetscErrorCode MatGetRowMinAbs_MPIAIJ(Mat A, Vec v, PetscInt idx[]) 2342 { 2343 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 2344 PetscErrorCode ierr; 2345 PetscInt i,*idxb = 0; 2346 PetscScalar *va,*vb; 2347 Vec vtmp; 2348 2349 PetscFunctionBegin; 2350 ierr = MatGetRowMinAbs(a->A,v,idx);CHKERRQ(ierr); 2351 ierr = VecGetArray(v,&va);CHKERRQ(ierr); 2352 if (idx) { 2353 for (i=0; i<A->cmap->n; i++) { 2354 if (PetscAbsScalar(va[i])) idx[i] += A->cmap->rstart; 2355 } 2356 } 2357 2358 ierr = VecCreateSeq(PETSC_COMM_SELF,A->rmap->n,&vtmp);CHKERRQ(ierr); 2359 if (idx) { 2360 ierr = PetscMalloc1(A->rmap->n,&idxb);CHKERRQ(ierr); 2361 } 2362 ierr = MatGetRowMinAbs(a->B,vtmp,idxb);CHKERRQ(ierr); 2363 ierr = VecGetArray(vtmp,&vb);CHKERRQ(ierr); 2364 2365 for (i=0; i<A->rmap->n; i++) { 2366 if (PetscAbsScalar(va[i]) > PetscAbsScalar(vb[i])) { 2367 va[i] = vb[i]; 2368 if (idx) idx[i] = a->garray[idxb[i]]; 2369 } 2370 } 2371 2372 ierr = VecRestoreArray(v,&va);CHKERRQ(ierr); 2373 ierr = VecRestoreArray(vtmp,&vb);CHKERRQ(ierr); 2374 ierr = PetscFree(idxb);CHKERRQ(ierr); 2375 ierr = VecDestroy(&vtmp);CHKERRQ(ierr); 2376 PetscFunctionReturn(0); 2377 } 2378 2379 PetscErrorCode MatGetRowMin_MPIAIJ(Mat A, Vec v, PetscInt idx[]) 2380 { 2381 Mat_MPIAIJ *mat = (Mat_MPIAIJ*) A->data; 2382 PetscInt n = A->rmap->n; 2383 PetscInt cstart = A->cmap->rstart; 2384 PetscInt *cmap = mat->garray; 2385 PetscInt *diagIdx, *offdiagIdx; 2386 Vec diagV, offdiagV; 2387 PetscScalar *a, *diagA, *offdiagA; 2388 PetscInt r; 2389 PetscErrorCode ierr; 2390 2391 PetscFunctionBegin; 2392 ierr = PetscMalloc2(n,&diagIdx,n,&offdiagIdx);CHKERRQ(ierr); 2393 ierr = VecCreateSeq(PetscObjectComm((PetscObject)A), n, &diagV);CHKERRQ(ierr); 2394 ierr = VecCreateSeq(PetscObjectComm((PetscObject)A), n, &offdiagV);CHKERRQ(ierr); 2395 ierr = MatGetRowMin(mat->A, diagV, diagIdx);CHKERRQ(ierr); 2396 ierr = MatGetRowMin(mat->B, offdiagV, offdiagIdx);CHKERRQ(ierr); 2397 ierr = VecGetArray(v, &a);CHKERRQ(ierr); 2398 ierr = VecGetArray(diagV, &diagA);CHKERRQ(ierr); 2399 ierr = VecGetArray(offdiagV, &offdiagA);CHKERRQ(ierr); 2400 for (r = 0; r < n; ++r) { 2401 if (PetscAbsScalar(diagA[r]) <= PetscAbsScalar(offdiagA[r])) { 2402 a[r] = diagA[r]; 2403 idx[r] = cstart + diagIdx[r]; 2404 } else { 2405 a[r] = offdiagA[r]; 2406 idx[r] = cmap[offdiagIdx[r]]; 2407 } 2408 } 2409 ierr = VecRestoreArray(v, &a);CHKERRQ(ierr); 2410 ierr = VecRestoreArray(diagV, &diagA);CHKERRQ(ierr); 2411 ierr = VecRestoreArray(offdiagV, &offdiagA);CHKERRQ(ierr); 2412 ierr = VecDestroy(&diagV);CHKERRQ(ierr); 2413 ierr = VecDestroy(&offdiagV);CHKERRQ(ierr); 2414 ierr = PetscFree2(diagIdx, offdiagIdx);CHKERRQ(ierr); 2415 PetscFunctionReturn(0); 2416 } 2417 2418 PetscErrorCode MatGetRowMax_MPIAIJ(Mat A, Vec v, PetscInt idx[]) 2419 { 2420 Mat_MPIAIJ *mat = (Mat_MPIAIJ*) A->data; 2421 PetscInt n = A->rmap->n; 2422 PetscInt cstart = A->cmap->rstart; 2423 PetscInt *cmap = mat->garray; 2424 PetscInt *diagIdx, *offdiagIdx; 2425 Vec diagV, offdiagV; 2426 PetscScalar *a, *diagA, *offdiagA; 2427 PetscInt r; 2428 PetscErrorCode ierr; 2429 2430 PetscFunctionBegin; 2431 ierr = PetscMalloc2(n,&diagIdx,n,&offdiagIdx);CHKERRQ(ierr); 2432 ierr = VecCreateSeq(PETSC_COMM_SELF, n, &diagV);CHKERRQ(ierr); 2433 ierr = VecCreateSeq(PETSC_COMM_SELF, n, &offdiagV);CHKERRQ(ierr); 2434 ierr = MatGetRowMax(mat->A, diagV, diagIdx);CHKERRQ(ierr); 2435 ierr = MatGetRowMax(mat->B, offdiagV, offdiagIdx);CHKERRQ(ierr); 2436 ierr = VecGetArray(v, &a);CHKERRQ(ierr); 2437 ierr = VecGetArray(diagV, &diagA);CHKERRQ(ierr); 2438 ierr = VecGetArray(offdiagV, &offdiagA);CHKERRQ(ierr); 2439 for (r = 0; r < n; ++r) { 2440 if (PetscAbsScalar(diagA[r]) >= PetscAbsScalar(offdiagA[r])) { 2441 a[r] = diagA[r]; 2442 idx[r] = cstart + diagIdx[r]; 2443 } else { 2444 a[r] = offdiagA[r]; 2445 idx[r] = cmap[offdiagIdx[r]]; 2446 } 2447 } 2448 ierr = VecRestoreArray(v, &a);CHKERRQ(ierr); 2449 ierr = VecRestoreArray(diagV, &diagA);CHKERRQ(ierr); 2450 ierr = VecRestoreArray(offdiagV, &offdiagA);CHKERRQ(ierr); 2451 ierr = VecDestroy(&diagV);CHKERRQ(ierr); 2452 ierr = VecDestroy(&offdiagV);CHKERRQ(ierr); 2453 ierr = PetscFree2(diagIdx, offdiagIdx);CHKERRQ(ierr); 2454 PetscFunctionReturn(0); 2455 } 2456 2457 PetscErrorCode MatGetSeqNonzeroStructure_MPIAIJ(Mat mat,Mat *newmat) 2458 { 2459 PetscErrorCode ierr; 2460 Mat *dummy; 2461 2462 PetscFunctionBegin; 2463 ierr = MatCreateSubMatrix_MPIAIJ_All(mat,MAT_DO_NOT_GET_VALUES,MAT_INITIAL_MATRIX,&dummy);CHKERRQ(ierr); 2464 *newmat = *dummy; 2465 ierr = PetscFree(dummy);CHKERRQ(ierr); 2466 PetscFunctionReturn(0); 2467 } 2468 2469 PetscErrorCode MatInvertBlockDiagonal_MPIAIJ(Mat A,const PetscScalar **values) 2470 { 2471 Mat_MPIAIJ *a = (Mat_MPIAIJ*) A->data; 2472 PetscErrorCode ierr; 2473 2474 PetscFunctionBegin; 2475 ierr = MatInvertBlockDiagonal(a->A,values);CHKERRQ(ierr); 2476 A->factorerrortype = a->A->factorerrortype; 2477 PetscFunctionReturn(0); 2478 } 2479 2480 static PetscErrorCode MatSetRandom_MPIAIJ(Mat x,PetscRandom rctx) 2481 { 2482 PetscErrorCode ierr; 2483 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)x->data; 2484 2485 PetscFunctionBegin; 2486 if (!x->assembled && !x->preallocated) SETERRQ(PetscObjectComm((PetscObject)x), PETSC_ERR_ARG_WRONGSTATE, "MatSetRandom on an unassembled and unpreallocated MATMPIAIJ is not allowed"); 2487 ierr = MatSetRandom(aij->A,rctx);CHKERRQ(ierr); 2488 if (x->assembled) { 2489 ierr = MatSetRandom(aij->B,rctx);CHKERRQ(ierr); 2490 } else { 2491 ierr = MatSetRandomSkipColumnRange_SeqAIJ_Private(aij->B,x->cmap->rstart,x->cmap->rend,rctx);CHKERRQ(ierr); 2492 } 2493 ierr = MatAssemblyBegin(x,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2494 ierr = MatAssemblyEnd(x,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2495 PetscFunctionReturn(0); 2496 } 2497 2498 PetscErrorCode MatMPIAIJSetUseScalableIncreaseOverlap_MPIAIJ(Mat A,PetscBool sc) 2499 { 2500 PetscFunctionBegin; 2501 if (sc) A->ops->increaseoverlap = MatIncreaseOverlap_MPIAIJ_Scalable; 2502 else A->ops->increaseoverlap = MatIncreaseOverlap_MPIAIJ; 2503 PetscFunctionReturn(0); 2504 } 2505 2506 /*@ 2507 MatMPIAIJSetUseScalableIncreaseOverlap - Determine if the matrix uses a scalable algorithm to compute the overlap 2508 2509 Collective on Mat 2510 2511 Input Parameters: 2512 + A - the matrix 2513 - sc - PETSC_TRUE indicates use the scalable algorithm (default is not to use the scalable algorithm) 2514 2515 Level: advanced 2516 2517 @*/ 2518 PetscErrorCode MatMPIAIJSetUseScalableIncreaseOverlap(Mat A,PetscBool sc) 2519 { 2520 PetscErrorCode ierr; 2521 2522 PetscFunctionBegin; 2523 ierr = PetscTryMethod(A,"MatMPIAIJSetUseScalableIncreaseOverlap_C",(Mat,PetscBool),(A,sc));CHKERRQ(ierr); 2524 PetscFunctionReturn(0); 2525 } 2526 2527 PetscErrorCode MatSetFromOptions_MPIAIJ(PetscOptionItems *PetscOptionsObject,Mat A) 2528 { 2529 PetscErrorCode ierr; 2530 PetscBool sc = PETSC_FALSE,flg; 2531 2532 PetscFunctionBegin; 2533 ierr = PetscOptionsHead(PetscOptionsObject,"MPIAIJ options");CHKERRQ(ierr); 2534 if (A->ops->increaseoverlap == MatIncreaseOverlap_MPIAIJ_Scalable) sc = PETSC_TRUE; 2535 ierr = PetscOptionsBool("-mat_increase_overlap_scalable","Use a scalable algorithm to compute the overlap","MatIncreaseOverlap",sc,&sc,&flg);CHKERRQ(ierr); 2536 if (flg) { 2537 ierr = MatMPIAIJSetUseScalableIncreaseOverlap(A,sc);CHKERRQ(ierr); 2538 } 2539 ierr = PetscOptionsTail();CHKERRQ(ierr); 2540 PetscFunctionReturn(0); 2541 } 2542 2543 PetscErrorCode MatShift_MPIAIJ(Mat Y,PetscScalar a) 2544 { 2545 PetscErrorCode ierr; 2546 Mat_MPIAIJ *maij = (Mat_MPIAIJ*)Y->data; 2547 Mat_SeqAIJ *aij = (Mat_SeqAIJ*)maij->A->data; 2548 2549 PetscFunctionBegin; 2550 if (!Y->preallocated) { 2551 ierr = MatMPIAIJSetPreallocation(Y,1,NULL,0,NULL);CHKERRQ(ierr); 2552 } else if (!aij->nz) { 2553 PetscInt nonew = aij->nonew; 2554 ierr = MatSeqAIJSetPreallocation(maij->A,1,NULL);CHKERRQ(ierr); 2555 aij->nonew = nonew; 2556 } 2557 ierr = MatShift_Basic(Y,a);CHKERRQ(ierr); 2558 PetscFunctionReturn(0); 2559 } 2560 2561 PetscErrorCode MatMissingDiagonal_MPIAIJ(Mat A,PetscBool *missing,PetscInt *d) 2562 { 2563 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 2564 PetscErrorCode ierr; 2565 2566 PetscFunctionBegin; 2567 if (A->rmap->n != A->cmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Only works for square matrices"); 2568 ierr = MatMissingDiagonal(a->A,missing,d);CHKERRQ(ierr); 2569 if (d) { 2570 PetscInt rstart; 2571 ierr = MatGetOwnershipRange(A,&rstart,NULL);CHKERRQ(ierr); 2572 *d += rstart; 2573 2574 } 2575 PetscFunctionReturn(0); 2576 } 2577 2578 PetscErrorCode MatInvertVariableBlockDiagonal_MPIAIJ(Mat A,PetscInt nblocks,const PetscInt *bsizes,PetscScalar *diag) 2579 { 2580 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 2581 PetscErrorCode ierr; 2582 2583 PetscFunctionBegin; 2584 ierr = MatInvertVariableBlockDiagonal(a->A,nblocks,bsizes,diag);CHKERRQ(ierr); 2585 PetscFunctionReturn(0); 2586 } 2587 2588 /* -------------------------------------------------------------------*/ 2589 static struct _MatOps MatOps_Values = {MatSetValues_MPIAIJ, 2590 MatGetRow_MPIAIJ, 2591 MatRestoreRow_MPIAIJ, 2592 MatMult_MPIAIJ, 2593 /* 4*/ MatMultAdd_MPIAIJ, 2594 MatMultTranspose_MPIAIJ, 2595 MatMultTransposeAdd_MPIAIJ, 2596 0, 2597 0, 2598 0, 2599 /*10*/ 0, 2600 0, 2601 0, 2602 MatSOR_MPIAIJ, 2603 MatTranspose_MPIAIJ, 2604 /*15*/ MatGetInfo_MPIAIJ, 2605 MatEqual_MPIAIJ, 2606 MatGetDiagonal_MPIAIJ, 2607 MatDiagonalScale_MPIAIJ, 2608 MatNorm_MPIAIJ, 2609 /*20*/ MatAssemblyBegin_MPIAIJ, 2610 MatAssemblyEnd_MPIAIJ, 2611 MatSetOption_MPIAIJ, 2612 MatZeroEntries_MPIAIJ, 2613 /*24*/ MatZeroRows_MPIAIJ, 2614 0, 2615 0, 2616 0, 2617 0, 2618 /*29*/ MatSetUp_MPIAIJ, 2619 0, 2620 0, 2621 MatGetDiagonalBlock_MPIAIJ, 2622 0, 2623 /*34*/ MatDuplicate_MPIAIJ, 2624 0, 2625 0, 2626 0, 2627 0, 2628 /*39*/ MatAXPY_MPIAIJ, 2629 MatCreateSubMatrices_MPIAIJ, 2630 MatIncreaseOverlap_MPIAIJ, 2631 MatGetValues_MPIAIJ, 2632 MatCopy_MPIAIJ, 2633 /*44*/ MatGetRowMax_MPIAIJ, 2634 MatScale_MPIAIJ, 2635 MatShift_MPIAIJ, 2636 MatDiagonalSet_MPIAIJ, 2637 MatZeroRowsColumns_MPIAIJ, 2638 /*49*/ MatSetRandom_MPIAIJ, 2639 0, 2640 0, 2641 0, 2642 0, 2643 /*54*/ MatFDColoringCreate_MPIXAIJ, 2644 0, 2645 MatSetUnfactored_MPIAIJ, 2646 MatPermute_MPIAIJ, 2647 0, 2648 /*59*/ MatCreateSubMatrix_MPIAIJ, 2649 MatDestroy_MPIAIJ, 2650 MatView_MPIAIJ, 2651 0, 2652 MatMatMatMult_MPIAIJ_MPIAIJ_MPIAIJ, 2653 /*64*/ MatMatMatMultSymbolic_MPIAIJ_MPIAIJ_MPIAIJ, 2654 MatMatMatMultNumeric_MPIAIJ_MPIAIJ_MPIAIJ, 2655 0, 2656 0, 2657 0, 2658 /*69*/ MatGetRowMaxAbs_MPIAIJ, 2659 MatGetRowMinAbs_MPIAIJ, 2660 0, 2661 0, 2662 0, 2663 0, 2664 /*75*/ MatFDColoringApply_AIJ, 2665 MatSetFromOptions_MPIAIJ, 2666 0, 2667 0, 2668 MatFindZeroDiagonals_MPIAIJ, 2669 /*80*/ 0, 2670 0, 2671 0, 2672 /*83*/ MatLoad_MPIAIJ, 2673 MatIsSymmetric_MPIAIJ, 2674 0, 2675 0, 2676 0, 2677 0, 2678 /*89*/ MatMatMult_MPIAIJ_MPIAIJ, 2679 MatMatMultSymbolic_MPIAIJ_MPIAIJ, 2680 MatMatMultNumeric_MPIAIJ_MPIAIJ, 2681 MatPtAP_MPIAIJ_MPIAIJ, 2682 MatPtAPSymbolic_MPIAIJ_MPIAIJ, 2683 /*94*/ MatPtAPNumeric_MPIAIJ_MPIAIJ, 2684 0, 2685 0, 2686 0, 2687 0, 2688 /*99*/ 0, 2689 0, 2690 0, 2691 MatConjugate_MPIAIJ, 2692 0, 2693 /*104*/MatSetValuesRow_MPIAIJ, 2694 MatRealPart_MPIAIJ, 2695 MatImaginaryPart_MPIAIJ, 2696 0, 2697 0, 2698 /*109*/0, 2699 0, 2700 MatGetRowMin_MPIAIJ, 2701 0, 2702 MatMissingDiagonal_MPIAIJ, 2703 /*114*/MatGetSeqNonzeroStructure_MPIAIJ, 2704 0, 2705 MatGetGhosts_MPIAIJ, 2706 0, 2707 0, 2708 /*119*/0, 2709 0, 2710 0, 2711 0, 2712 MatGetMultiProcBlock_MPIAIJ, 2713 /*124*/MatFindNonzeroRows_MPIAIJ, 2714 MatGetColumnNorms_MPIAIJ, 2715 MatInvertBlockDiagonal_MPIAIJ, 2716 MatInvertVariableBlockDiagonal_MPIAIJ, 2717 MatCreateSubMatricesMPI_MPIAIJ, 2718 /*129*/0, 2719 MatTransposeMatMult_MPIAIJ_MPIAIJ, 2720 MatTransposeMatMultSymbolic_MPIAIJ_MPIAIJ, 2721 MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ, 2722 0, 2723 /*134*/0, 2724 0, 2725 MatRARt_MPIAIJ_MPIAIJ, 2726 0, 2727 0, 2728 /*139*/MatSetBlockSizes_MPIAIJ, 2729 0, 2730 0, 2731 MatFDColoringSetUp_MPIXAIJ, 2732 MatFindOffBlockDiagonalEntries_MPIAIJ, 2733 /*144*/MatCreateMPIMatConcatenateSeqMat_MPIAIJ 2734 }; 2735 2736 /* ----------------------------------------------------------------------------------------*/ 2737 2738 PetscErrorCode MatStoreValues_MPIAIJ(Mat mat) 2739 { 2740 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 2741 PetscErrorCode ierr; 2742 2743 PetscFunctionBegin; 2744 ierr = MatStoreValues(aij->A);CHKERRQ(ierr); 2745 ierr = MatStoreValues(aij->B);CHKERRQ(ierr); 2746 PetscFunctionReturn(0); 2747 } 2748 2749 PetscErrorCode MatRetrieveValues_MPIAIJ(Mat mat) 2750 { 2751 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 2752 PetscErrorCode ierr; 2753 2754 PetscFunctionBegin; 2755 ierr = MatRetrieveValues(aij->A);CHKERRQ(ierr); 2756 ierr = MatRetrieveValues(aij->B);CHKERRQ(ierr); 2757 PetscFunctionReturn(0); 2758 } 2759 2760 PetscErrorCode MatMPIAIJSetPreallocation_MPIAIJ(Mat B,PetscInt d_nz,const PetscInt d_nnz[],PetscInt o_nz,const PetscInt o_nnz[]) 2761 { 2762 Mat_MPIAIJ *b; 2763 PetscErrorCode ierr; 2764 PetscMPIInt size; 2765 2766 PetscFunctionBegin; 2767 ierr = PetscLayoutSetUp(B->rmap);CHKERRQ(ierr); 2768 ierr = PetscLayoutSetUp(B->cmap);CHKERRQ(ierr); 2769 b = (Mat_MPIAIJ*)B->data; 2770 2771 #if defined(PETSC_USE_CTABLE) 2772 ierr = PetscTableDestroy(&b->colmap);CHKERRQ(ierr); 2773 #else 2774 ierr = PetscFree(b->colmap);CHKERRQ(ierr); 2775 #endif 2776 ierr = PetscFree(b->garray);CHKERRQ(ierr); 2777 ierr = VecDestroy(&b->lvec);CHKERRQ(ierr); 2778 ierr = VecScatterDestroy(&b->Mvctx);CHKERRQ(ierr); 2779 2780 /* Because the B will have been resized we simply destroy it and create a new one each time */ 2781 ierr = MPI_Comm_size(PetscObjectComm((PetscObject)B),&size);CHKERRQ(ierr); 2782 ierr = MatDestroy(&b->B);CHKERRQ(ierr); 2783 ierr = MatCreate(PETSC_COMM_SELF,&b->B);CHKERRQ(ierr); 2784 ierr = MatSetSizes(b->B,B->rmap->n,size > 1 ? B->cmap->N : 0,B->rmap->n,size > 1 ? B->cmap->N : 0);CHKERRQ(ierr); 2785 ierr = MatSetBlockSizesFromMats(b->B,B,B);CHKERRQ(ierr); 2786 ierr = MatSetType(b->B,MATSEQAIJ);CHKERRQ(ierr); 2787 ierr = PetscLogObjectParent((PetscObject)B,(PetscObject)b->B);CHKERRQ(ierr); 2788 2789 if (!B->preallocated) { 2790 ierr = MatCreate(PETSC_COMM_SELF,&b->A);CHKERRQ(ierr); 2791 ierr = MatSetSizes(b->A,B->rmap->n,B->cmap->n,B->rmap->n,B->cmap->n);CHKERRQ(ierr); 2792 ierr = MatSetBlockSizesFromMats(b->A,B,B);CHKERRQ(ierr); 2793 ierr = MatSetType(b->A,MATSEQAIJ);CHKERRQ(ierr); 2794 ierr = PetscLogObjectParent((PetscObject)B,(PetscObject)b->A);CHKERRQ(ierr); 2795 } 2796 2797 ierr = MatSeqAIJSetPreallocation(b->A,d_nz,d_nnz);CHKERRQ(ierr); 2798 ierr = MatSeqAIJSetPreallocation(b->B,o_nz,o_nnz);CHKERRQ(ierr); 2799 B->preallocated = PETSC_TRUE; 2800 B->was_assembled = PETSC_FALSE; 2801 B->assembled = PETSC_FALSE; 2802 PetscFunctionReturn(0); 2803 } 2804 2805 PetscErrorCode MatResetPreallocation_MPIAIJ(Mat B) 2806 { 2807 Mat_MPIAIJ *b; 2808 PetscErrorCode ierr; 2809 2810 PetscFunctionBegin; 2811 PetscValidHeaderSpecific(B,MAT_CLASSID,1); 2812 ierr = PetscLayoutSetUp(B->rmap);CHKERRQ(ierr); 2813 ierr = PetscLayoutSetUp(B->cmap);CHKERRQ(ierr); 2814 b = (Mat_MPIAIJ*)B->data; 2815 2816 #if defined(PETSC_USE_CTABLE) 2817 ierr = PetscTableDestroy(&b->colmap);CHKERRQ(ierr); 2818 #else 2819 ierr = PetscFree(b->colmap);CHKERRQ(ierr); 2820 #endif 2821 ierr = PetscFree(b->garray);CHKERRQ(ierr); 2822 ierr = VecDestroy(&b->lvec);CHKERRQ(ierr); 2823 ierr = VecScatterDestroy(&b->Mvctx);CHKERRQ(ierr); 2824 2825 ierr = MatResetPreallocation(b->A);CHKERRQ(ierr); 2826 ierr = MatResetPreallocation(b->B);CHKERRQ(ierr); 2827 B->preallocated = PETSC_TRUE; 2828 B->was_assembled = PETSC_FALSE; 2829 B->assembled = PETSC_FALSE; 2830 PetscFunctionReturn(0); 2831 } 2832 2833 PetscErrorCode MatDuplicate_MPIAIJ(Mat matin,MatDuplicateOption cpvalues,Mat *newmat) 2834 { 2835 Mat mat; 2836 Mat_MPIAIJ *a,*oldmat = (Mat_MPIAIJ*)matin->data; 2837 PetscErrorCode ierr; 2838 2839 PetscFunctionBegin; 2840 *newmat = 0; 2841 ierr = MatCreate(PetscObjectComm((PetscObject)matin),&mat);CHKERRQ(ierr); 2842 ierr = MatSetSizes(mat,matin->rmap->n,matin->cmap->n,matin->rmap->N,matin->cmap->N);CHKERRQ(ierr); 2843 ierr = MatSetBlockSizesFromMats(mat,matin,matin);CHKERRQ(ierr); 2844 ierr = MatSetType(mat,((PetscObject)matin)->type_name);CHKERRQ(ierr); 2845 a = (Mat_MPIAIJ*)mat->data; 2846 2847 mat->factortype = matin->factortype; 2848 mat->assembled = PETSC_TRUE; 2849 mat->insertmode = NOT_SET_VALUES; 2850 mat->preallocated = PETSC_TRUE; 2851 2852 a->size = oldmat->size; 2853 a->rank = oldmat->rank; 2854 a->donotstash = oldmat->donotstash; 2855 a->roworiented = oldmat->roworiented; 2856 a->rowindices = 0; 2857 a->rowvalues = 0; 2858 a->getrowactive = PETSC_FALSE; 2859 2860 ierr = PetscLayoutReference(matin->rmap,&mat->rmap);CHKERRQ(ierr); 2861 ierr = PetscLayoutReference(matin->cmap,&mat->cmap);CHKERRQ(ierr); 2862 2863 if (oldmat->colmap) { 2864 #if defined(PETSC_USE_CTABLE) 2865 ierr = PetscTableCreateCopy(oldmat->colmap,&a->colmap);CHKERRQ(ierr); 2866 #else 2867 ierr = PetscMalloc1(mat->cmap->N,&a->colmap);CHKERRQ(ierr); 2868 ierr = PetscLogObjectMemory((PetscObject)mat,(mat->cmap->N)*sizeof(PetscInt));CHKERRQ(ierr); 2869 ierr = PetscArraycpy(a->colmap,oldmat->colmap,mat->cmap->N);CHKERRQ(ierr); 2870 #endif 2871 } else a->colmap = 0; 2872 if (oldmat->garray) { 2873 PetscInt len; 2874 len = oldmat->B->cmap->n; 2875 ierr = PetscMalloc1(len+1,&a->garray);CHKERRQ(ierr); 2876 ierr = PetscLogObjectMemory((PetscObject)mat,len*sizeof(PetscInt));CHKERRQ(ierr); 2877 if (len) { ierr = PetscArraycpy(a->garray,oldmat->garray,len);CHKERRQ(ierr); } 2878 } else a->garray = 0; 2879 2880 ierr = VecDuplicate(oldmat->lvec,&a->lvec);CHKERRQ(ierr); 2881 ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)a->lvec);CHKERRQ(ierr); 2882 ierr = VecScatterCopy(oldmat->Mvctx,&a->Mvctx);CHKERRQ(ierr); 2883 ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)a->Mvctx);CHKERRQ(ierr); 2884 2885 if (oldmat->Mvctx_mpi1) { 2886 ierr = VecScatterCopy(oldmat->Mvctx_mpi1,&a->Mvctx_mpi1);CHKERRQ(ierr); 2887 ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)a->Mvctx_mpi1);CHKERRQ(ierr); 2888 } 2889 2890 ierr = MatDuplicate(oldmat->A,cpvalues,&a->A);CHKERRQ(ierr); 2891 ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)a->A);CHKERRQ(ierr); 2892 ierr = MatDuplicate(oldmat->B,cpvalues,&a->B);CHKERRQ(ierr); 2893 ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)a->B);CHKERRQ(ierr); 2894 ierr = PetscFunctionListDuplicate(((PetscObject)matin)->qlist,&((PetscObject)mat)->qlist);CHKERRQ(ierr); 2895 *newmat = mat; 2896 PetscFunctionReturn(0); 2897 } 2898 2899 PetscErrorCode MatLoad_MPIAIJ(Mat newMat, PetscViewer viewer) 2900 { 2901 PetscBool isbinary, ishdf5; 2902 PetscErrorCode ierr; 2903 2904 PetscFunctionBegin; 2905 PetscValidHeaderSpecific(newMat,MAT_CLASSID,1); 2906 PetscValidHeaderSpecific(viewer,PETSC_VIEWER_CLASSID,2); 2907 /* force binary viewer to load .info file if it has not yet done so */ 2908 ierr = PetscViewerSetUp(viewer);CHKERRQ(ierr); 2909 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERBINARY,&isbinary);CHKERRQ(ierr); 2910 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERHDF5, &ishdf5);CHKERRQ(ierr); 2911 if (isbinary) { 2912 ierr = MatLoad_MPIAIJ_Binary(newMat,viewer);CHKERRQ(ierr); 2913 } else if (ishdf5) { 2914 #if defined(PETSC_HAVE_HDF5) 2915 ierr = MatLoad_AIJ_HDF5(newMat,viewer);CHKERRQ(ierr); 2916 #else 2917 SETERRQ(PetscObjectComm((PetscObject)newMat),PETSC_ERR_SUP,"HDF5 not supported in this build.\nPlease reconfigure using --download-hdf5"); 2918 #endif 2919 } else { 2920 SETERRQ2(PetscObjectComm((PetscObject)newMat),PETSC_ERR_SUP,"Viewer type %s not yet supported for reading %s matrices",((PetscObject)viewer)->type_name,((PetscObject)newMat)->type_name); 2921 } 2922 PetscFunctionReturn(0); 2923 } 2924 2925 PetscErrorCode MatLoad_MPIAIJ_Binary(Mat newMat, PetscViewer viewer) 2926 { 2927 PetscScalar *vals,*svals; 2928 MPI_Comm comm; 2929 PetscErrorCode ierr; 2930 PetscMPIInt rank,size,tag = ((PetscObject)viewer)->tag; 2931 PetscInt i,nz,j,rstart,rend,mmax,maxnz = 0; 2932 PetscInt header[4],*rowlengths = 0,M,N,m,*cols; 2933 PetscInt *ourlens = NULL,*procsnz = NULL,*offlens = NULL,jj,*mycols,*smycols; 2934 PetscInt cend,cstart,n,*rowners; 2935 int fd; 2936 PetscInt bs = newMat->rmap->bs; 2937 2938 PetscFunctionBegin; 2939 ierr = PetscObjectGetComm((PetscObject)viewer,&comm);CHKERRQ(ierr); 2940 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 2941 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 2942 ierr = PetscViewerBinaryGetDescriptor(viewer,&fd);CHKERRQ(ierr); 2943 if (!rank) { 2944 ierr = PetscBinaryRead(fd,(char*)header,4,NULL,PETSC_INT);CHKERRQ(ierr); 2945 if (header[0] != MAT_FILE_CLASSID) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED,"not matrix object"); 2946 if (header[3] < 0) SETERRQ(PetscObjectComm((PetscObject)newMat),PETSC_ERR_FILE_UNEXPECTED,"Matrix stored in special format on disk,cannot load as MATMPIAIJ"); 2947 } 2948 2949 ierr = PetscOptionsBegin(comm,NULL,"Options for loading MATMPIAIJ matrix","Mat");CHKERRQ(ierr); 2950 ierr = PetscOptionsInt("-matload_block_size","Set the blocksize used to store the matrix","MatLoad",bs,&bs,NULL);CHKERRQ(ierr); 2951 ierr = PetscOptionsEnd();CHKERRQ(ierr); 2952 if (bs < 0) bs = 1; 2953 2954 ierr = MPI_Bcast(header+1,3,MPIU_INT,0,comm);CHKERRQ(ierr); 2955 M = header[1]; N = header[2]; 2956 2957 /* If global sizes are set, check if they are consistent with that given in the file */ 2958 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); 2959 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); 2960 2961 /* determine ownership of all (block) rows */ 2962 if (M%bs) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED, "Inconsistent # of rows (%d) and block size (%d)",M,bs); 2963 if (newMat->rmap->n < 0) m = bs*((M/bs)/size + (((M/bs) % size) > rank)); /* PETSC_DECIDE */ 2964 else m = newMat->rmap->n; /* Set by user */ 2965 2966 ierr = PetscMalloc1(size+1,&rowners);CHKERRQ(ierr); 2967 ierr = MPI_Allgather(&m,1,MPIU_INT,rowners+1,1,MPIU_INT,comm);CHKERRQ(ierr); 2968 2969 /* First process needs enough room for process with most rows */ 2970 if (!rank) { 2971 mmax = rowners[1]; 2972 for (i=2; i<=size; i++) { 2973 mmax = PetscMax(mmax, rowners[i]); 2974 } 2975 } else mmax = -1; /* unused, but compilers complain */ 2976 2977 rowners[0] = 0; 2978 for (i=2; i<=size; i++) { 2979 rowners[i] += rowners[i-1]; 2980 } 2981 rstart = rowners[rank]; 2982 rend = rowners[rank+1]; 2983 2984 /* distribute row lengths to all processors */ 2985 ierr = PetscMalloc2(m,&ourlens,m,&offlens);CHKERRQ(ierr); 2986 if (!rank) { 2987 ierr = PetscBinaryRead(fd,ourlens,m,NULL,PETSC_INT);CHKERRQ(ierr); 2988 ierr = PetscMalloc1(mmax,&rowlengths);CHKERRQ(ierr); 2989 ierr = PetscCalloc1(size,&procsnz);CHKERRQ(ierr); 2990 for (j=0; j<m; j++) { 2991 procsnz[0] += ourlens[j]; 2992 } 2993 for (i=1; i<size; i++) { 2994 ierr = PetscBinaryRead(fd,rowlengths,rowners[i+1]-rowners[i],NULL,PETSC_INT);CHKERRQ(ierr); 2995 /* calculate the number of nonzeros on each processor */ 2996 for (j=0; j<rowners[i+1]-rowners[i]; j++) { 2997 procsnz[i] += rowlengths[j]; 2998 } 2999 ierr = MPIULong_Send(rowlengths,rowners[i+1]-rowners[i],MPIU_INT,i,tag,comm);CHKERRQ(ierr); 3000 } 3001 ierr = PetscFree(rowlengths);CHKERRQ(ierr); 3002 } else { 3003 ierr = MPIULong_Recv(ourlens,m,MPIU_INT,0,tag,comm);CHKERRQ(ierr); 3004 } 3005 3006 if (!rank) { 3007 /* determine max buffer needed and allocate it */ 3008 maxnz = 0; 3009 for (i=0; i<size; i++) { 3010 maxnz = PetscMax(maxnz,procsnz[i]); 3011 } 3012 ierr = PetscMalloc1(maxnz,&cols);CHKERRQ(ierr); 3013 3014 /* read in my part of the matrix column indices */ 3015 nz = procsnz[0]; 3016 ierr = PetscMalloc1(nz,&mycols);CHKERRQ(ierr); 3017 ierr = PetscBinaryRead(fd,mycols,nz,NULL,PETSC_INT);CHKERRQ(ierr); 3018 3019 /* read in every one elses and ship off */ 3020 for (i=1; i<size; i++) { 3021 nz = procsnz[i]; 3022 ierr = PetscBinaryRead(fd,cols,nz,NULL,PETSC_INT);CHKERRQ(ierr); 3023 ierr = MPIULong_Send(cols,nz,MPIU_INT,i,tag,comm);CHKERRQ(ierr); 3024 } 3025 ierr = PetscFree(cols);CHKERRQ(ierr); 3026 } else { 3027 /* determine buffer space needed for message */ 3028 nz = 0; 3029 for (i=0; i<m; i++) { 3030 nz += ourlens[i]; 3031 } 3032 ierr = PetscMalloc1(nz,&mycols);CHKERRQ(ierr); 3033 3034 /* receive message of column indices*/ 3035 ierr = MPIULong_Recv(mycols,nz,MPIU_INT,0,tag,comm);CHKERRQ(ierr); 3036 } 3037 3038 /* determine column ownership if matrix is not square */ 3039 if (N != M) { 3040 if (newMat->cmap->n < 0) n = N/size + ((N % size) > rank); 3041 else n = newMat->cmap->n; 3042 ierr = MPI_Scan(&n,&cend,1,MPIU_INT,MPI_SUM,comm);CHKERRQ(ierr); 3043 cstart = cend - n; 3044 } else { 3045 cstart = rstart; 3046 cend = rend; 3047 n = cend - cstart; 3048 } 3049 3050 /* loop over local rows, determining number of off diagonal entries */ 3051 ierr = PetscArrayzero(offlens,m);CHKERRQ(ierr); 3052 jj = 0; 3053 for (i=0; i<m; i++) { 3054 for (j=0; j<ourlens[i]; j++) { 3055 if (mycols[jj] < cstart || mycols[jj] >= cend) offlens[i]++; 3056 jj++; 3057 } 3058 } 3059 3060 for (i=0; i<m; i++) { 3061 ourlens[i] -= offlens[i]; 3062 } 3063 ierr = MatSetSizes(newMat,m,n,M,N);CHKERRQ(ierr); 3064 3065 if (bs > 1) {ierr = MatSetBlockSize(newMat,bs);CHKERRQ(ierr);} 3066 3067 ierr = MatMPIAIJSetPreallocation(newMat,0,ourlens,0,offlens);CHKERRQ(ierr); 3068 3069 for (i=0; i<m; i++) { 3070 ourlens[i] += offlens[i]; 3071 } 3072 3073 if (!rank) { 3074 ierr = PetscMalloc1(maxnz+1,&vals);CHKERRQ(ierr); 3075 3076 /* read in my part of the matrix numerical values */ 3077 nz = procsnz[0]; 3078 ierr = PetscBinaryRead(fd,vals,nz,NULL,PETSC_SCALAR);CHKERRQ(ierr); 3079 3080 /* insert into matrix */ 3081 jj = rstart; 3082 smycols = mycols; 3083 svals = vals; 3084 for (i=0; i<m; i++) { 3085 ierr = MatSetValues_MPIAIJ(newMat,1,&jj,ourlens[i],smycols,svals,INSERT_VALUES);CHKERRQ(ierr); 3086 smycols += ourlens[i]; 3087 svals += ourlens[i]; 3088 jj++; 3089 } 3090 3091 /* read in other processors and ship out */ 3092 for (i=1; i<size; i++) { 3093 nz = procsnz[i]; 3094 ierr = PetscBinaryRead(fd,vals,nz,NULL,PETSC_SCALAR);CHKERRQ(ierr); 3095 ierr = MPIULong_Send(vals,nz,MPIU_SCALAR,i,((PetscObject)newMat)->tag,comm);CHKERRQ(ierr); 3096 } 3097 ierr = PetscFree(procsnz);CHKERRQ(ierr); 3098 } else { 3099 /* receive numeric values */ 3100 ierr = PetscMalloc1(nz+1,&vals);CHKERRQ(ierr); 3101 3102 /* receive message of values*/ 3103 ierr = MPIULong_Recv(vals,nz,MPIU_SCALAR,0,((PetscObject)newMat)->tag,comm);CHKERRQ(ierr); 3104 3105 /* insert into matrix */ 3106 jj = rstart; 3107 smycols = mycols; 3108 svals = vals; 3109 for (i=0; i<m; i++) { 3110 ierr = MatSetValues_MPIAIJ(newMat,1,&jj,ourlens[i],smycols,svals,INSERT_VALUES);CHKERRQ(ierr); 3111 smycols += ourlens[i]; 3112 svals += ourlens[i]; 3113 jj++; 3114 } 3115 } 3116 ierr = PetscFree2(ourlens,offlens);CHKERRQ(ierr); 3117 ierr = PetscFree(vals);CHKERRQ(ierr); 3118 ierr = PetscFree(mycols);CHKERRQ(ierr); 3119 ierr = PetscFree(rowners);CHKERRQ(ierr); 3120 ierr = MatAssemblyBegin(newMat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3121 ierr = MatAssemblyEnd(newMat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3122 PetscFunctionReturn(0); 3123 } 3124 3125 /* Not scalable because of ISAllGather() unless getting all columns. */ 3126 PetscErrorCode ISGetSeqIS_Private(Mat mat,IS iscol,IS *isseq) 3127 { 3128 PetscErrorCode ierr; 3129 IS iscol_local; 3130 PetscBool isstride; 3131 PetscMPIInt lisstride=0,gisstride; 3132 3133 PetscFunctionBegin; 3134 /* check if we are grabbing all columns*/ 3135 ierr = PetscObjectTypeCompare((PetscObject)iscol,ISSTRIDE,&isstride);CHKERRQ(ierr); 3136 3137 if (isstride) { 3138 PetscInt start,len,mstart,mlen; 3139 ierr = ISStrideGetInfo(iscol,&start,NULL);CHKERRQ(ierr); 3140 ierr = ISGetLocalSize(iscol,&len);CHKERRQ(ierr); 3141 ierr = MatGetOwnershipRangeColumn(mat,&mstart,&mlen);CHKERRQ(ierr); 3142 if (mstart == start && mlen-mstart == len) lisstride = 1; 3143 } 3144 3145 ierr = MPIU_Allreduce(&lisstride,&gisstride,1,MPI_INT,MPI_MIN,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 3146 if (gisstride) { 3147 PetscInt N; 3148 ierr = MatGetSize(mat,NULL,&N);CHKERRQ(ierr); 3149 ierr = ISCreateStride(PetscObjectComm((PetscObject)mat),N,0,1,&iscol_local);CHKERRQ(ierr); 3150 ierr = ISSetIdentity(iscol_local);CHKERRQ(ierr); 3151 ierr = PetscInfo(mat,"Optimizing for obtaining all columns of the matrix; skipping ISAllGather()\n");CHKERRQ(ierr); 3152 } else { 3153 PetscInt cbs; 3154 ierr = ISGetBlockSize(iscol,&cbs);CHKERRQ(ierr); 3155 ierr = ISAllGather(iscol,&iscol_local);CHKERRQ(ierr); 3156 ierr = ISSetBlockSize(iscol_local,cbs);CHKERRQ(ierr); 3157 } 3158 3159 *isseq = iscol_local; 3160 PetscFunctionReturn(0); 3161 } 3162 3163 /* 3164 Used by MatCreateSubMatrix_MPIAIJ_SameRowColDist() to avoid ISAllGather() and global size of iscol_local 3165 (see MatCreateSubMatrix_MPIAIJ_nonscalable) 3166 3167 Input Parameters: 3168 mat - matrix 3169 isrow - parallel row index set; its local indices are a subset of local columns of mat, 3170 i.e., mat->rstart <= isrow[i] < mat->rend 3171 iscol - parallel column index set; its local indices are a subset of local columns of mat, 3172 i.e., mat->cstart <= iscol[i] < mat->cend 3173 Output Parameter: 3174 isrow_d,iscol_d - sequential row and column index sets for retrieving mat->A 3175 iscol_o - sequential column index set for retrieving mat->B 3176 garray - column map; garray[i] indicates global location of iscol_o[i] in iscol 3177 */ 3178 PetscErrorCode ISGetSeqIS_SameColDist_Private(Mat mat,IS isrow,IS iscol,IS *isrow_d,IS *iscol_d,IS *iscol_o,const PetscInt *garray[]) 3179 { 3180 PetscErrorCode ierr; 3181 Vec x,cmap; 3182 const PetscInt *is_idx; 3183 PetscScalar *xarray,*cmaparray; 3184 PetscInt ncols,isstart,*idx,m,rstart,*cmap1,count; 3185 Mat_MPIAIJ *a=(Mat_MPIAIJ*)mat->data; 3186 Mat B=a->B; 3187 Vec lvec=a->lvec,lcmap; 3188 PetscInt i,cstart,cend,Bn=B->cmap->N; 3189 MPI_Comm comm; 3190 VecScatter Mvctx=a->Mvctx; 3191 3192 PetscFunctionBegin; 3193 ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 3194 ierr = ISGetLocalSize(iscol,&ncols);CHKERRQ(ierr); 3195 3196 /* (1) iscol is a sub-column vector of mat, pad it with '-1.' to form a full vector x */ 3197 ierr = MatCreateVecs(mat,&x,NULL);CHKERRQ(ierr); 3198 ierr = VecSet(x,-1.0);CHKERRQ(ierr); 3199 ierr = VecDuplicate(x,&cmap);CHKERRQ(ierr); 3200 ierr = VecSet(cmap,-1.0);CHKERRQ(ierr); 3201 3202 /* Get start indices */ 3203 ierr = MPI_Scan(&ncols,&isstart,1,MPIU_INT,MPI_SUM,comm);CHKERRQ(ierr); 3204 isstart -= ncols; 3205 ierr = MatGetOwnershipRangeColumn(mat,&cstart,&cend);CHKERRQ(ierr); 3206 3207 ierr = ISGetIndices(iscol,&is_idx);CHKERRQ(ierr); 3208 ierr = VecGetArray(x,&xarray);CHKERRQ(ierr); 3209 ierr = VecGetArray(cmap,&cmaparray);CHKERRQ(ierr); 3210 ierr = PetscMalloc1(ncols,&idx);CHKERRQ(ierr); 3211 for (i=0; i<ncols; i++) { 3212 xarray[is_idx[i]-cstart] = (PetscScalar)is_idx[i]; 3213 cmaparray[is_idx[i]-cstart] = i + isstart; /* global index of iscol[i] */ 3214 idx[i] = is_idx[i]-cstart; /* local index of iscol[i] */ 3215 } 3216 ierr = VecRestoreArray(x,&xarray);CHKERRQ(ierr); 3217 ierr = VecRestoreArray(cmap,&cmaparray);CHKERRQ(ierr); 3218 ierr = ISRestoreIndices(iscol,&is_idx);CHKERRQ(ierr); 3219 3220 /* Get iscol_d */ 3221 ierr = ISCreateGeneral(PETSC_COMM_SELF,ncols,idx,PETSC_OWN_POINTER,iscol_d);CHKERRQ(ierr); 3222 ierr = ISGetBlockSize(iscol,&i);CHKERRQ(ierr); 3223 ierr = ISSetBlockSize(*iscol_d,i);CHKERRQ(ierr); 3224 3225 /* Get isrow_d */ 3226 ierr = ISGetLocalSize(isrow,&m);CHKERRQ(ierr); 3227 rstart = mat->rmap->rstart; 3228 ierr = PetscMalloc1(m,&idx);CHKERRQ(ierr); 3229 ierr = ISGetIndices(isrow,&is_idx);CHKERRQ(ierr); 3230 for (i=0; i<m; i++) idx[i] = is_idx[i]-rstart; 3231 ierr = ISRestoreIndices(isrow,&is_idx);CHKERRQ(ierr); 3232 3233 ierr = ISCreateGeneral(PETSC_COMM_SELF,m,idx,PETSC_OWN_POINTER,isrow_d);CHKERRQ(ierr); 3234 ierr = ISGetBlockSize(isrow,&i);CHKERRQ(ierr); 3235 ierr = ISSetBlockSize(*isrow_d,i);CHKERRQ(ierr); 3236 3237 /* (2) Scatter x and cmap using aij->Mvctx to get their off-process portions (see MatMult_MPIAIJ) */ 3238 ierr = VecScatterBegin(Mvctx,x,lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3239 ierr = VecScatterEnd(Mvctx,x,lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3240 3241 ierr = VecDuplicate(lvec,&lcmap);CHKERRQ(ierr); 3242 3243 ierr = VecScatterBegin(Mvctx,cmap,lcmap,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3244 ierr = VecScatterEnd(Mvctx,cmap,lcmap,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3245 3246 /* (3) create sequential iscol_o (a subset of iscol) and isgarray */ 3247 /* off-process column indices */ 3248 count = 0; 3249 ierr = PetscMalloc1(Bn,&idx);CHKERRQ(ierr); 3250 ierr = PetscMalloc1(Bn,&cmap1);CHKERRQ(ierr); 3251 3252 ierr = VecGetArray(lvec,&xarray);CHKERRQ(ierr); 3253 ierr = VecGetArray(lcmap,&cmaparray);CHKERRQ(ierr); 3254 for (i=0; i<Bn; i++) { 3255 if (PetscRealPart(xarray[i]) > -1.0) { 3256 idx[count] = i; /* local column index in off-diagonal part B */ 3257 cmap1[count] = (PetscInt)PetscRealPart(cmaparray[i]); /* column index in submat */ 3258 count++; 3259 } 3260 } 3261 ierr = VecRestoreArray(lvec,&xarray);CHKERRQ(ierr); 3262 ierr = VecRestoreArray(lcmap,&cmaparray);CHKERRQ(ierr); 3263 3264 ierr = ISCreateGeneral(PETSC_COMM_SELF,count,idx,PETSC_COPY_VALUES,iscol_o);CHKERRQ(ierr); 3265 /* cannot ensure iscol_o has same blocksize as iscol! */ 3266 3267 ierr = PetscFree(idx);CHKERRQ(ierr); 3268 *garray = cmap1; 3269 3270 ierr = VecDestroy(&x);CHKERRQ(ierr); 3271 ierr = VecDestroy(&cmap);CHKERRQ(ierr); 3272 ierr = VecDestroy(&lcmap);CHKERRQ(ierr); 3273 PetscFunctionReturn(0); 3274 } 3275 3276 /* isrow and iscol have same processor distribution as mat, output *submat is a submatrix of local mat */ 3277 PetscErrorCode MatCreateSubMatrix_MPIAIJ_SameRowColDist(Mat mat,IS isrow,IS iscol,MatReuse call,Mat *submat) 3278 { 3279 PetscErrorCode ierr; 3280 Mat_MPIAIJ *a = (Mat_MPIAIJ*)mat->data,*asub; 3281 Mat M = NULL; 3282 MPI_Comm comm; 3283 IS iscol_d,isrow_d,iscol_o; 3284 Mat Asub = NULL,Bsub = NULL; 3285 PetscInt n; 3286 3287 PetscFunctionBegin; 3288 ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 3289 3290 if (call == MAT_REUSE_MATRIX) { 3291 /* Retrieve isrow_d, iscol_d and iscol_o from submat */ 3292 ierr = PetscObjectQuery((PetscObject)*submat,"isrow_d",(PetscObject*)&isrow_d);CHKERRQ(ierr); 3293 if (!isrow_d) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"isrow_d passed in was not used before, cannot reuse"); 3294 3295 ierr = PetscObjectQuery((PetscObject)*submat,"iscol_d",(PetscObject*)&iscol_d);CHKERRQ(ierr); 3296 if (!iscol_d) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"iscol_d passed in was not used before, cannot reuse"); 3297 3298 ierr = PetscObjectQuery((PetscObject)*submat,"iscol_o",(PetscObject*)&iscol_o);CHKERRQ(ierr); 3299 if (!iscol_o) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"iscol_o passed in was not used before, cannot reuse"); 3300 3301 /* Update diagonal and off-diagonal portions of submat */ 3302 asub = (Mat_MPIAIJ*)(*submat)->data; 3303 ierr = MatCreateSubMatrix_SeqAIJ(a->A,isrow_d,iscol_d,PETSC_DECIDE,MAT_REUSE_MATRIX,&asub->A);CHKERRQ(ierr); 3304 ierr = ISGetLocalSize(iscol_o,&n);CHKERRQ(ierr); 3305 if (n) { 3306 ierr = MatCreateSubMatrix_SeqAIJ(a->B,isrow_d,iscol_o,PETSC_DECIDE,MAT_REUSE_MATRIX,&asub->B);CHKERRQ(ierr); 3307 } 3308 ierr = MatAssemblyBegin(*submat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3309 ierr = MatAssemblyEnd(*submat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3310 3311 } else { /* call == MAT_INITIAL_MATRIX) */ 3312 const PetscInt *garray; 3313 PetscInt BsubN; 3314 3315 /* Create isrow_d, iscol_d, iscol_o and isgarray (replace isgarray with array?) */ 3316 ierr = ISGetSeqIS_SameColDist_Private(mat,isrow,iscol,&isrow_d,&iscol_d,&iscol_o,&garray);CHKERRQ(ierr); 3317 3318 /* Create local submatrices Asub and Bsub */ 3319 ierr = MatCreateSubMatrix_SeqAIJ(a->A,isrow_d,iscol_d,PETSC_DECIDE,MAT_INITIAL_MATRIX,&Asub);CHKERRQ(ierr); 3320 ierr = MatCreateSubMatrix_SeqAIJ(a->B,isrow_d,iscol_o,PETSC_DECIDE,MAT_INITIAL_MATRIX,&Bsub);CHKERRQ(ierr); 3321 3322 /* Create submatrix M */ 3323 ierr = MatCreateMPIAIJWithSeqAIJ(comm,Asub,Bsub,garray,&M);CHKERRQ(ierr); 3324 3325 /* If Bsub has empty columns, compress iscol_o such that it will retrieve condensed Bsub from a->B during reuse */ 3326 asub = (Mat_MPIAIJ*)M->data; 3327 3328 ierr = ISGetLocalSize(iscol_o,&BsubN);CHKERRQ(ierr); 3329 n = asub->B->cmap->N; 3330 if (BsubN > n) { 3331 /* This case can be tested using ~petsc/src/tao/bound/examples/tutorials/runplate2_3 */ 3332 const PetscInt *idx; 3333 PetscInt i,j,*idx_new,*subgarray = asub->garray; 3334 ierr = PetscInfo2(M,"submatrix Bn %D != BsubN %D, update iscol_o\n",n,BsubN);CHKERRQ(ierr); 3335 3336 ierr = PetscMalloc1(n,&idx_new);CHKERRQ(ierr); 3337 j = 0; 3338 ierr = ISGetIndices(iscol_o,&idx);CHKERRQ(ierr); 3339 for (i=0; i<n; i++) { 3340 if (j >= BsubN) break; 3341 while (subgarray[i] > garray[j]) j++; 3342 3343 if (subgarray[i] == garray[j]) { 3344 idx_new[i] = idx[j++]; 3345 } else SETERRQ4(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"subgarray[%D]=%D cannot < garray[%D]=%D",i,subgarray[i],j,garray[j]); 3346 } 3347 ierr = ISRestoreIndices(iscol_o,&idx);CHKERRQ(ierr); 3348 3349 ierr = ISDestroy(&iscol_o);CHKERRQ(ierr); 3350 ierr = ISCreateGeneral(PETSC_COMM_SELF,n,idx_new,PETSC_OWN_POINTER,&iscol_o);CHKERRQ(ierr); 3351 3352 } else if (BsubN < n) { 3353 SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Columns of Bsub cannot be smaller than B's",BsubN,asub->B->cmap->N); 3354 } 3355 3356 ierr = PetscFree(garray);CHKERRQ(ierr); 3357 *submat = M; 3358 3359 /* Save isrow_d, iscol_d and iscol_o used in processor for next request */ 3360 ierr = PetscObjectCompose((PetscObject)M,"isrow_d",(PetscObject)isrow_d);CHKERRQ(ierr); 3361 ierr = ISDestroy(&isrow_d);CHKERRQ(ierr); 3362 3363 ierr = PetscObjectCompose((PetscObject)M,"iscol_d",(PetscObject)iscol_d);CHKERRQ(ierr); 3364 ierr = ISDestroy(&iscol_d);CHKERRQ(ierr); 3365 3366 ierr = PetscObjectCompose((PetscObject)M,"iscol_o",(PetscObject)iscol_o);CHKERRQ(ierr); 3367 ierr = ISDestroy(&iscol_o);CHKERRQ(ierr); 3368 } 3369 PetscFunctionReturn(0); 3370 } 3371 3372 PetscErrorCode MatCreateSubMatrix_MPIAIJ(Mat mat,IS isrow,IS iscol,MatReuse call,Mat *newmat) 3373 { 3374 PetscErrorCode ierr; 3375 IS iscol_local=NULL,isrow_d; 3376 PetscInt csize; 3377 PetscInt n,i,j,start,end; 3378 PetscBool sameRowDist=PETSC_FALSE,sameDist[2],tsameDist[2]; 3379 MPI_Comm comm; 3380 3381 PetscFunctionBegin; 3382 /* If isrow has same processor distribution as mat, 3383 call MatCreateSubMatrix_MPIAIJ_SameRowDist() to avoid using a hash table with global size of iscol */ 3384 if (call == MAT_REUSE_MATRIX) { 3385 ierr = PetscObjectQuery((PetscObject)*newmat,"isrow_d",(PetscObject*)&isrow_d);CHKERRQ(ierr); 3386 if (isrow_d) { 3387 sameRowDist = PETSC_TRUE; 3388 tsameDist[1] = PETSC_TRUE; /* sameColDist */ 3389 } else { 3390 ierr = PetscObjectQuery((PetscObject)*newmat,"SubIScol",(PetscObject*)&iscol_local);CHKERRQ(ierr); 3391 if (iscol_local) { 3392 sameRowDist = PETSC_TRUE; 3393 tsameDist[1] = PETSC_FALSE; /* !sameColDist */ 3394 } 3395 } 3396 } else { 3397 /* Check if isrow has same processor distribution as mat */ 3398 sameDist[0] = PETSC_FALSE; 3399 ierr = ISGetLocalSize(isrow,&n);CHKERRQ(ierr); 3400 if (!n) { 3401 sameDist[0] = PETSC_TRUE; 3402 } else { 3403 ierr = ISGetMinMax(isrow,&i,&j);CHKERRQ(ierr); 3404 ierr = MatGetOwnershipRange(mat,&start,&end);CHKERRQ(ierr); 3405 if (i >= start && j < end) { 3406 sameDist[0] = PETSC_TRUE; 3407 } 3408 } 3409 3410 /* Check if iscol has same processor distribution as mat */ 3411 sameDist[1] = PETSC_FALSE; 3412 ierr = ISGetLocalSize(iscol,&n);CHKERRQ(ierr); 3413 if (!n) { 3414 sameDist[1] = PETSC_TRUE; 3415 } else { 3416 ierr = ISGetMinMax(iscol,&i,&j);CHKERRQ(ierr); 3417 ierr = MatGetOwnershipRangeColumn(mat,&start,&end);CHKERRQ(ierr); 3418 if (i >= start && j < end) sameDist[1] = PETSC_TRUE; 3419 } 3420 3421 ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 3422 ierr = MPIU_Allreduce(&sameDist,&tsameDist,2,MPIU_BOOL,MPI_LAND,comm);CHKERRQ(ierr); 3423 sameRowDist = tsameDist[0]; 3424 } 3425 3426 if (sameRowDist) { 3427 if (tsameDist[1]) { /* sameRowDist & sameColDist */ 3428 /* isrow and iscol have same processor distribution as mat */ 3429 ierr = MatCreateSubMatrix_MPIAIJ_SameRowColDist(mat,isrow,iscol,call,newmat);CHKERRQ(ierr); 3430 PetscFunctionReturn(0); 3431 } else { /* sameRowDist */ 3432 /* isrow has same processor distribution as mat */ 3433 if (call == MAT_INITIAL_MATRIX) { 3434 PetscBool sorted; 3435 ierr = ISGetSeqIS_Private(mat,iscol,&iscol_local);CHKERRQ(ierr); 3436 ierr = ISGetLocalSize(iscol_local,&n);CHKERRQ(ierr); /* local size of iscol_local = global columns of newmat */ 3437 ierr = ISGetSize(iscol,&i);CHKERRQ(ierr); 3438 if (n != i) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"n %d != size of iscol %d",n,i); 3439 3440 ierr = ISSorted(iscol_local,&sorted);CHKERRQ(ierr); 3441 if (sorted) { 3442 /* MatCreateSubMatrix_MPIAIJ_SameRowDist() requires iscol_local be sorted; it can have duplicate indices */ 3443 ierr = MatCreateSubMatrix_MPIAIJ_SameRowDist(mat,isrow,iscol,iscol_local,MAT_INITIAL_MATRIX,newmat);CHKERRQ(ierr); 3444 PetscFunctionReturn(0); 3445 } 3446 } else { /* call == MAT_REUSE_MATRIX */ 3447 IS iscol_sub; 3448 ierr = PetscObjectQuery((PetscObject)*newmat,"SubIScol",(PetscObject*)&iscol_sub);CHKERRQ(ierr); 3449 if (iscol_sub) { 3450 ierr = MatCreateSubMatrix_MPIAIJ_SameRowDist(mat,isrow,iscol,NULL,call,newmat);CHKERRQ(ierr); 3451 PetscFunctionReturn(0); 3452 } 3453 } 3454 } 3455 } 3456 3457 /* General case: iscol -> iscol_local which has global size of iscol */ 3458 if (call == MAT_REUSE_MATRIX) { 3459 ierr = PetscObjectQuery((PetscObject)*newmat,"ISAllGather",(PetscObject*)&iscol_local);CHKERRQ(ierr); 3460 if (!iscol_local) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Submatrix passed in was not used before, cannot reuse"); 3461 } else { 3462 if (!iscol_local) { 3463 ierr = ISGetSeqIS_Private(mat,iscol,&iscol_local);CHKERRQ(ierr); 3464 } 3465 } 3466 3467 ierr = ISGetLocalSize(iscol,&csize);CHKERRQ(ierr); 3468 ierr = MatCreateSubMatrix_MPIAIJ_nonscalable(mat,isrow,iscol_local,csize,call,newmat);CHKERRQ(ierr); 3469 3470 if (call == MAT_INITIAL_MATRIX) { 3471 ierr = PetscObjectCompose((PetscObject)*newmat,"ISAllGather",(PetscObject)iscol_local);CHKERRQ(ierr); 3472 ierr = ISDestroy(&iscol_local);CHKERRQ(ierr); 3473 } 3474 PetscFunctionReturn(0); 3475 } 3476 3477 /*@C 3478 MatCreateMPIAIJWithSeqAIJ - creates a MPIAIJ matrix using SeqAIJ matrices that contain the "diagonal" 3479 and "off-diagonal" part of the matrix in CSR format. 3480 3481 Collective 3482 3483 Input Parameters: 3484 + comm - MPI communicator 3485 . A - "diagonal" portion of matrix 3486 . B - "off-diagonal" portion of matrix, may have empty columns, will be destroyed by this routine 3487 - garray - global index of B columns 3488 3489 Output Parameter: 3490 . mat - the matrix, with input A as its local diagonal matrix 3491 Level: advanced 3492 3493 Notes: 3494 See MatCreateAIJ() for the definition of "diagonal" and "off-diagonal" portion of the matrix. 3495 A becomes part of output mat, B is destroyed by this routine. The user cannot use A and B anymore. 3496 3497 .seealso: MatCreateMPIAIJWithSplitArrays() 3498 @*/ 3499 PetscErrorCode MatCreateMPIAIJWithSeqAIJ(MPI_Comm comm,Mat A,Mat B,const PetscInt garray[],Mat *mat) 3500 { 3501 PetscErrorCode ierr; 3502 Mat_MPIAIJ *maij; 3503 Mat_SeqAIJ *b=(Mat_SeqAIJ*)B->data,*bnew; 3504 PetscInt *oi=b->i,*oj=b->j,i,nz,col; 3505 PetscScalar *oa=b->a; 3506 Mat Bnew; 3507 PetscInt m,n,N; 3508 3509 PetscFunctionBegin; 3510 ierr = MatCreate(comm,mat);CHKERRQ(ierr); 3511 ierr = MatGetSize(A,&m,&n);CHKERRQ(ierr); 3512 if (m != B->rmap->N) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Am %D != Bm %D",m,B->rmap->N); 3513 if (A->rmap->bs != B->rmap->bs) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"A row bs %D != B row bs %D",A->rmap->bs,B->rmap->bs); 3514 /* remove check below; When B is created using iscol_o from ISGetSeqIS_SameColDist_Private(), its bs may not be same as A */ 3515 /* if (A->cmap->bs != B->cmap->bs) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"A column bs %D != B column bs %D",A->cmap->bs,B->cmap->bs); */ 3516 3517 /* Get global columns of mat */ 3518 ierr = MPIU_Allreduce(&n,&N,1,MPIU_INT,MPI_SUM,comm);CHKERRQ(ierr); 3519 3520 ierr = MatSetSizes(*mat,m,n,PETSC_DECIDE,N);CHKERRQ(ierr); 3521 ierr = MatSetType(*mat,MATMPIAIJ);CHKERRQ(ierr); 3522 ierr = MatSetBlockSizes(*mat,A->rmap->bs,A->cmap->bs);CHKERRQ(ierr); 3523 maij = (Mat_MPIAIJ*)(*mat)->data; 3524 3525 (*mat)->preallocated = PETSC_TRUE; 3526 3527 ierr = PetscLayoutSetUp((*mat)->rmap);CHKERRQ(ierr); 3528 ierr = PetscLayoutSetUp((*mat)->cmap);CHKERRQ(ierr); 3529 3530 /* Set A as diagonal portion of *mat */ 3531 maij->A = A; 3532 3533 nz = oi[m]; 3534 for (i=0; i<nz; i++) { 3535 col = oj[i]; 3536 oj[i] = garray[col]; 3537 } 3538 3539 /* Set Bnew as off-diagonal portion of *mat */ 3540 ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,m,N,oi,oj,oa,&Bnew);CHKERRQ(ierr); 3541 bnew = (Mat_SeqAIJ*)Bnew->data; 3542 bnew->maxnz = b->maxnz; /* allocated nonzeros of B */ 3543 maij->B = Bnew; 3544 3545 if (B->rmap->N != Bnew->rmap->N) SETERRQ2(PETSC_COMM_SELF,0,"BN %d != BnewN %d",B->rmap->N,Bnew->rmap->N); 3546 3547 b->singlemalloc = PETSC_FALSE; /* B arrays are shared by Bnew */ 3548 b->free_a = PETSC_FALSE; 3549 b->free_ij = PETSC_FALSE; 3550 ierr = MatDestroy(&B);CHKERRQ(ierr); 3551 3552 bnew->singlemalloc = PETSC_TRUE; /* arrays will be freed by MatDestroy(&Bnew) */ 3553 bnew->free_a = PETSC_TRUE; 3554 bnew->free_ij = PETSC_TRUE; 3555 3556 /* condense columns of maij->B */ 3557 ierr = MatSetOption(*mat,MAT_NO_OFF_PROC_ENTRIES,PETSC_TRUE);CHKERRQ(ierr); 3558 ierr = MatAssemblyBegin(*mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3559 ierr = MatAssemblyEnd(*mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3560 ierr = MatSetOption(*mat,MAT_NO_OFF_PROC_ENTRIES,PETSC_FALSE);CHKERRQ(ierr); 3561 ierr = MatSetOption(*mat,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE);CHKERRQ(ierr); 3562 PetscFunctionReturn(0); 3563 } 3564 3565 extern PetscErrorCode MatCreateSubMatrices_MPIAIJ_SingleIS_Local(Mat,PetscInt,const IS[],const IS[],MatReuse,PetscBool,Mat*); 3566 3567 PetscErrorCode MatCreateSubMatrix_MPIAIJ_SameRowDist(Mat mat,IS isrow,IS iscol,IS iscol_local,MatReuse call,Mat *newmat) 3568 { 3569 PetscErrorCode ierr; 3570 PetscInt i,m,n,rstart,row,rend,nz,j,bs,cbs; 3571 PetscInt *ii,*jj,nlocal,*dlens,*olens,dlen,olen,jend,mglobal; 3572 Mat_MPIAIJ *a=(Mat_MPIAIJ*)mat->data; 3573 Mat M,Msub,B=a->B; 3574 MatScalar *aa; 3575 Mat_SeqAIJ *aij; 3576 PetscInt *garray = a->garray,*colsub,Ncols; 3577 PetscInt count,Bn=B->cmap->N,cstart=mat->cmap->rstart,cend=mat->cmap->rend; 3578 IS iscol_sub,iscmap; 3579 const PetscInt *is_idx,*cmap; 3580 PetscBool allcolumns=PETSC_FALSE; 3581 MPI_Comm comm; 3582 3583 PetscFunctionBegin; 3584 ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 3585 3586 if (call == MAT_REUSE_MATRIX) { 3587 ierr = PetscObjectQuery((PetscObject)*newmat,"SubIScol",(PetscObject*)&iscol_sub);CHKERRQ(ierr); 3588 if (!iscol_sub) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"SubIScol passed in was not used before, cannot reuse"); 3589 ierr = ISGetLocalSize(iscol_sub,&count);CHKERRQ(ierr); 3590 3591 ierr = PetscObjectQuery((PetscObject)*newmat,"Subcmap",(PetscObject*)&iscmap);CHKERRQ(ierr); 3592 if (!iscmap) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Subcmap passed in was not used before, cannot reuse"); 3593 3594 ierr = PetscObjectQuery((PetscObject)*newmat,"SubMatrix",(PetscObject*)&Msub);CHKERRQ(ierr); 3595 if (!Msub) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Submatrix passed in was not used before, cannot reuse"); 3596 3597 ierr = MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat,1,&isrow,&iscol_sub,MAT_REUSE_MATRIX,PETSC_FALSE,&Msub);CHKERRQ(ierr); 3598 3599 } else { /* call == MAT_INITIAL_MATRIX) */ 3600 PetscBool flg; 3601 3602 ierr = ISGetLocalSize(iscol,&n);CHKERRQ(ierr); 3603 ierr = ISGetSize(iscol,&Ncols);CHKERRQ(ierr); 3604 3605 /* (1) iscol -> nonscalable iscol_local */ 3606 /* Check for special case: each processor gets entire matrix columns */ 3607 ierr = ISIdentity(iscol_local,&flg);CHKERRQ(ierr); 3608 if (flg && n == mat->cmap->N) allcolumns = PETSC_TRUE; 3609 if (allcolumns) { 3610 iscol_sub = iscol_local; 3611 ierr = PetscObjectReference((PetscObject)iscol_local);CHKERRQ(ierr); 3612 ierr = ISCreateStride(PETSC_COMM_SELF,n,0,1,&iscmap);CHKERRQ(ierr); 3613 3614 } else { 3615 /* (2) iscol_local -> iscol_sub and iscmap. Implementation below requires iscol_local be sorted, it can have duplicate indices */ 3616 PetscInt *idx,*cmap1,k; 3617 ierr = PetscMalloc1(Ncols,&idx);CHKERRQ(ierr); 3618 ierr = PetscMalloc1(Ncols,&cmap1);CHKERRQ(ierr); 3619 ierr = ISGetIndices(iscol_local,&is_idx);CHKERRQ(ierr); 3620 count = 0; 3621 k = 0; 3622 for (i=0; i<Ncols; i++) { 3623 j = is_idx[i]; 3624 if (j >= cstart && j < cend) { 3625 /* diagonal part of mat */ 3626 idx[count] = j; 3627 cmap1[count++] = i; /* column index in submat */ 3628 } else if (Bn) { 3629 /* off-diagonal part of mat */ 3630 if (j == garray[k]) { 3631 idx[count] = j; 3632 cmap1[count++] = i; /* column index in submat */ 3633 } else if (j > garray[k]) { 3634 while (j > garray[k] && k < Bn-1) k++; 3635 if (j == garray[k]) { 3636 idx[count] = j; 3637 cmap1[count++] = i; /* column index in submat */ 3638 } 3639 } 3640 } 3641 } 3642 ierr = ISRestoreIndices(iscol_local,&is_idx);CHKERRQ(ierr); 3643 3644 ierr = ISCreateGeneral(PETSC_COMM_SELF,count,idx,PETSC_OWN_POINTER,&iscol_sub);CHKERRQ(ierr); 3645 ierr = ISGetBlockSize(iscol,&cbs);CHKERRQ(ierr); 3646 ierr = ISSetBlockSize(iscol_sub,cbs);CHKERRQ(ierr); 3647 3648 ierr = ISCreateGeneral(PetscObjectComm((PetscObject)iscol_local),count,cmap1,PETSC_OWN_POINTER,&iscmap);CHKERRQ(ierr); 3649 } 3650 3651 /* (3) Create sequential Msub */ 3652 ierr = MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat,1,&isrow,&iscol_sub,MAT_INITIAL_MATRIX,allcolumns,&Msub);CHKERRQ(ierr); 3653 } 3654 3655 ierr = ISGetLocalSize(iscol_sub,&count);CHKERRQ(ierr); 3656 aij = (Mat_SeqAIJ*)(Msub)->data; 3657 ii = aij->i; 3658 ierr = ISGetIndices(iscmap,&cmap);CHKERRQ(ierr); 3659 3660 /* 3661 m - number of local rows 3662 Ncols - number of columns (same on all processors) 3663 rstart - first row in new global matrix generated 3664 */ 3665 ierr = MatGetSize(Msub,&m,NULL);CHKERRQ(ierr); 3666 3667 if (call == MAT_INITIAL_MATRIX) { 3668 /* (4) Create parallel newmat */ 3669 PetscMPIInt rank,size; 3670 PetscInt csize; 3671 3672 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 3673 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 3674 3675 /* 3676 Determine the number of non-zeros in the diagonal and off-diagonal 3677 portions of the matrix in order to do correct preallocation 3678 */ 3679 3680 /* first get start and end of "diagonal" columns */ 3681 ierr = ISGetLocalSize(iscol,&csize);CHKERRQ(ierr); 3682 if (csize == PETSC_DECIDE) { 3683 ierr = ISGetSize(isrow,&mglobal);CHKERRQ(ierr); 3684 if (mglobal == Ncols) { /* square matrix */ 3685 nlocal = m; 3686 } else { 3687 nlocal = Ncols/size + ((Ncols % size) > rank); 3688 } 3689 } else { 3690 nlocal = csize; 3691 } 3692 ierr = MPI_Scan(&nlocal,&rend,1,MPIU_INT,MPI_SUM,comm);CHKERRQ(ierr); 3693 rstart = rend - nlocal; 3694 if (rank == size - 1 && rend != Ncols) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Local column sizes %D do not add up to total number of columns %D",rend,Ncols); 3695 3696 /* next, compute all the lengths */ 3697 jj = aij->j; 3698 ierr = PetscMalloc1(2*m+1,&dlens);CHKERRQ(ierr); 3699 olens = dlens + m; 3700 for (i=0; i<m; i++) { 3701 jend = ii[i+1] - ii[i]; 3702 olen = 0; 3703 dlen = 0; 3704 for (j=0; j<jend; j++) { 3705 if (cmap[*jj] < rstart || cmap[*jj] >= rend) olen++; 3706 else dlen++; 3707 jj++; 3708 } 3709 olens[i] = olen; 3710 dlens[i] = dlen; 3711 } 3712 3713 ierr = ISGetBlockSize(isrow,&bs);CHKERRQ(ierr); 3714 ierr = ISGetBlockSize(iscol,&cbs);CHKERRQ(ierr); 3715 3716 ierr = MatCreate(comm,&M);CHKERRQ(ierr); 3717 ierr = MatSetSizes(M,m,nlocal,PETSC_DECIDE,Ncols);CHKERRQ(ierr); 3718 ierr = MatSetBlockSizes(M,bs,cbs);CHKERRQ(ierr); 3719 ierr = MatSetType(M,((PetscObject)mat)->type_name);CHKERRQ(ierr); 3720 ierr = MatMPIAIJSetPreallocation(M,0,dlens,0,olens);CHKERRQ(ierr); 3721 ierr = PetscFree(dlens);CHKERRQ(ierr); 3722 3723 } else { /* call == MAT_REUSE_MATRIX */ 3724 M = *newmat; 3725 ierr = MatGetLocalSize(M,&i,NULL);CHKERRQ(ierr); 3726 if (i != m) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Previous matrix must be same size/layout as request"); 3727 ierr = MatZeroEntries(M);CHKERRQ(ierr); 3728 /* 3729 The next two lines are needed so we may call MatSetValues_MPIAIJ() below directly, 3730 rather than the slower MatSetValues(). 3731 */ 3732 M->was_assembled = PETSC_TRUE; 3733 M->assembled = PETSC_FALSE; 3734 } 3735 3736 /* (5) Set values of Msub to *newmat */ 3737 ierr = PetscMalloc1(count,&colsub);CHKERRQ(ierr); 3738 ierr = MatGetOwnershipRange(M,&rstart,NULL);CHKERRQ(ierr); 3739 3740 jj = aij->j; 3741 aa = aij->a; 3742 for (i=0; i<m; i++) { 3743 row = rstart + i; 3744 nz = ii[i+1] - ii[i]; 3745 for (j=0; j<nz; j++) colsub[j] = cmap[jj[j]]; 3746 ierr = MatSetValues_MPIAIJ(M,1,&row,nz,colsub,aa,INSERT_VALUES);CHKERRQ(ierr); 3747 jj += nz; aa += nz; 3748 } 3749 ierr = ISRestoreIndices(iscmap,&cmap);CHKERRQ(ierr); 3750 3751 ierr = MatAssemblyBegin(M,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3752 ierr = MatAssemblyEnd(M,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3753 3754 ierr = PetscFree(colsub);CHKERRQ(ierr); 3755 3756 /* save Msub, iscol_sub and iscmap used in processor for next request */ 3757 if (call == MAT_INITIAL_MATRIX) { 3758 *newmat = M; 3759 ierr = PetscObjectCompose((PetscObject)(*newmat),"SubMatrix",(PetscObject)Msub);CHKERRQ(ierr); 3760 ierr = MatDestroy(&Msub);CHKERRQ(ierr); 3761 3762 ierr = PetscObjectCompose((PetscObject)(*newmat),"SubIScol",(PetscObject)iscol_sub);CHKERRQ(ierr); 3763 ierr = ISDestroy(&iscol_sub);CHKERRQ(ierr); 3764 3765 ierr = PetscObjectCompose((PetscObject)(*newmat),"Subcmap",(PetscObject)iscmap);CHKERRQ(ierr); 3766 ierr = ISDestroy(&iscmap);CHKERRQ(ierr); 3767 3768 if (iscol_local) { 3769 ierr = PetscObjectCompose((PetscObject)(*newmat),"ISAllGather",(PetscObject)iscol_local);CHKERRQ(ierr); 3770 ierr = ISDestroy(&iscol_local);CHKERRQ(ierr); 3771 } 3772 } 3773 PetscFunctionReturn(0); 3774 } 3775 3776 /* 3777 Not great since it makes two copies of the submatrix, first an SeqAIJ 3778 in local and then by concatenating the local matrices the end result. 3779 Writing it directly would be much like MatCreateSubMatrices_MPIAIJ() 3780 3781 Note: This requires a sequential iscol with all indices. 3782 */ 3783 PetscErrorCode MatCreateSubMatrix_MPIAIJ_nonscalable(Mat mat,IS isrow,IS iscol,PetscInt csize,MatReuse call,Mat *newmat) 3784 { 3785 PetscErrorCode ierr; 3786 PetscMPIInt rank,size; 3787 PetscInt i,m,n,rstart,row,rend,nz,*cwork,j,bs,cbs; 3788 PetscInt *ii,*jj,nlocal,*dlens,*olens,dlen,olen,jend,mglobal; 3789 Mat M,Mreuse; 3790 MatScalar *aa,*vwork; 3791 MPI_Comm comm; 3792 Mat_SeqAIJ *aij; 3793 PetscBool colflag,allcolumns=PETSC_FALSE; 3794 3795 PetscFunctionBegin; 3796 ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 3797 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 3798 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 3799 3800 /* Check for special case: each processor gets entire matrix columns */ 3801 ierr = ISIdentity(iscol,&colflag);CHKERRQ(ierr); 3802 ierr = ISGetLocalSize(iscol,&n);CHKERRQ(ierr); 3803 if (colflag && n == mat->cmap->N) allcolumns = PETSC_TRUE; 3804 3805 if (call == MAT_REUSE_MATRIX) { 3806 ierr = PetscObjectQuery((PetscObject)*newmat,"SubMatrix",(PetscObject*)&Mreuse);CHKERRQ(ierr); 3807 if (!Mreuse) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Submatrix passed in was not used before, cannot reuse"); 3808 ierr = MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat,1,&isrow,&iscol,MAT_REUSE_MATRIX,allcolumns,&Mreuse);CHKERRQ(ierr); 3809 } else { 3810 ierr = MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat,1,&isrow,&iscol,MAT_INITIAL_MATRIX,allcolumns,&Mreuse);CHKERRQ(ierr); 3811 } 3812 3813 /* 3814 m - number of local rows 3815 n - number of columns (same on all processors) 3816 rstart - first row in new global matrix generated 3817 */ 3818 ierr = MatGetSize(Mreuse,&m,&n);CHKERRQ(ierr); 3819 ierr = MatGetBlockSizes(Mreuse,&bs,&cbs);CHKERRQ(ierr); 3820 if (call == MAT_INITIAL_MATRIX) { 3821 aij = (Mat_SeqAIJ*)(Mreuse)->data; 3822 ii = aij->i; 3823 jj = aij->j; 3824 3825 /* 3826 Determine the number of non-zeros in the diagonal and off-diagonal 3827 portions of the matrix in order to do correct preallocation 3828 */ 3829 3830 /* first get start and end of "diagonal" columns */ 3831 if (csize == PETSC_DECIDE) { 3832 ierr = ISGetSize(isrow,&mglobal);CHKERRQ(ierr); 3833 if (mglobal == n) { /* square matrix */ 3834 nlocal = m; 3835 } else { 3836 nlocal = n/size + ((n % size) > rank); 3837 } 3838 } else { 3839 nlocal = csize; 3840 } 3841 ierr = MPI_Scan(&nlocal,&rend,1,MPIU_INT,MPI_SUM,comm);CHKERRQ(ierr); 3842 rstart = rend - nlocal; 3843 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); 3844 3845 /* next, compute all the lengths */ 3846 ierr = PetscMalloc1(2*m+1,&dlens);CHKERRQ(ierr); 3847 olens = dlens + m; 3848 for (i=0; i<m; i++) { 3849 jend = ii[i+1] - ii[i]; 3850 olen = 0; 3851 dlen = 0; 3852 for (j=0; j<jend; j++) { 3853 if (*jj < rstart || *jj >= rend) olen++; 3854 else dlen++; 3855 jj++; 3856 } 3857 olens[i] = olen; 3858 dlens[i] = dlen; 3859 } 3860 ierr = MatCreate(comm,&M);CHKERRQ(ierr); 3861 ierr = MatSetSizes(M,m,nlocal,PETSC_DECIDE,n);CHKERRQ(ierr); 3862 ierr = MatSetBlockSizes(M,bs,cbs);CHKERRQ(ierr); 3863 ierr = MatSetType(M,((PetscObject)mat)->type_name);CHKERRQ(ierr); 3864 ierr = MatMPIAIJSetPreallocation(M,0,dlens,0,olens);CHKERRQ(ierr); 3865 ierr = PetscFree(dlens);CHKERRQ(ierr); 3866 } else { 3867 PetscInt ml,nl; 3868 3869 M = *newmat; 3870 ierr = MatGetLocalSize(M,&ml,&nl);CHKERRQ(ierr); 3871 if (ml != m) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Previous matrix must be same size/layout as request"); 3872 ierr = MatZeroEntries(M);CHKERRQ(ierr); 3873 /* 3874 The next two lines are needed so we may call MatSetValues_MPIAIJ() below directly, 3875 rather than the slower MatSetValues(). 3876 */ 3877 M->was_assembled = PETSC_TRUE; 3878 M->assembled = PETSC_FALSE; 3879 } 3880 ierr = MatGetOwnershipRange(M,&rstart,&rend);CHKERRQ(ierr); 3881 aij = (Mat_SeqAIJ*)(Mreuse)->data; 3882 ii = aij->i; 3883 jj = aij->j; 3884 aa = aij->a; 3885 for (i=0; i<m; i++) { 3886 row = rstart + i; 3887 nz = ii[i+1] - ii[i]; 3888 cwork = jj; jj += nz; 3889 vwork = aa; aa += nz; 3890 ierr = MatSetValues_MPIAIJ(M,1,&row,nz,cwork,vwork,INSERT_VALUES);CHKERRQ(ierr); 3891 } 3892 3893 ierr = MatAssemblyBegin(M,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3894 ierr = MatAssemblyEnd(M,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3895 *newmat = M; 3896 3897 /* save submatrix used in processor for next request */ 3898 if (call == MAT_INITIAL_MATRIX) { 3899 ierr = PetscObjectCompose((PetscObject)M,"SubMatrix",(PetscObject)Mreuse);CHKERRQ(ierr); 3900 ierr = MatDestroy(&Mreuse);CHKERRQ(ierr); 3901 } 3902 PetscFunctionReturn(0); 3903 } 3904 3905 PetscErrorCode MatMPIAIJSetPreallocationCSR_MPIAIJ(Mat B,const PetscInt Ii[],const PetscInt J[],const PetscScalar v[]) 3906 { 3907 PetscInt m,cstart, cend,j,nnz,i,d; 3908 PetscInt *d_nnz,*o_nnz,nnz_max = 0,rstart,ii; 3909 const PetscInt *JJ; 3910 PetscErrorCode ierr; 3911 PetscBool nooffprocentries; 3912 3913 PetscFunctionBegin; 3914 if (Ii[0]) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Ii[0] must be 0 it is %D",Ii[0]); 3915 3916 ierr = PetscLayoutSetUp(B->rmap);CHKERRQ(ierr); 3917 ierr = PetscLayoutSetUp(B->cmap);CHKERRQ(ierr); 3918 m = B->rmap->n; 3919 cstart = B->cmap->rstart; 3920 cend = B->cmap->rend; 3921 rstart = B->rmap->rstart; 3922 3923 ierr = PetscCalloc2(m,&d_nnz,m,&o_nnz);CHKERRQ(ierr); 3924 3925 #if defined(PETSC_USE_DEBUG) 3926 for (i=0; i<m; i++) { 3927 nnz = Ii[i+1]- Ii[i]; 3928 JJ = J + Ii[i]; 3929 if (nnz < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Local row %D has a negative %D number of columns",i,nnz); 3930 if (nnz && (JJ[0] < 0)) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Row %D starts with negative column index",i,JJ[0]); 3931 if (nnz && (JJ[nnz-1] >= B->cmap->N)) SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Row %D ends with too large a column index %D (max allowed %D)",i,JJ[nnz-1],B->cmap->N); 3932 } 3933 #endif 3934 3935 for (i=0; i<m; i++) { 3936 nnz = Ii[i+1]- Ii[i]; 3937 JJ = J + Ii[i]; 3938 nnz_max = PetscMax(nnz_max,nnz); 3939 d = 0; 3940 for (j=0; j<nnz; j++) { 3941 if (cstart <= JJ[j] && JJ[j] < cend) d++; 3942 } 3943 d_nnz[i] = d; 3944 o_nnz[i] = nnz - d; 3945 } 3946 ierr = MatMPIAIJSetPreallocation(B,0,d_nnz,0,o_nnz);CHKERRQ(ierr); 3947 ierr = PetscFree2(d_nnz,o_nnz);CHKERRQ(ierr); 3948 3949 for (i=0; i<m; i++) { 3950 ii = i + rstart; 3951 ierr = MatSetValues_MPIAIJ(B,1,&ii,Ii[i+1] - Ii[i],J+Ii[i], v ? v + Ii[i] : NULL,INSERT_VALUES);CHKERRQ(ierr); 3952 } 3953 nooffprocentries = B->nooffprocentries; 3954 B->nooffprocentries = PETSC_TRUE; 3955 ierr = MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3956 ierr = MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3957 B->nooffprocentries = nooffprocentries; 3958 3959 ierr = MatSetOption(B,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE);CHKERRQ(ierr); 3960 PetscFunctionReturn(0); 3961 } 3962 3963 /*@ 3964 MatMPIAIJSetPreallocationCSR - Allocates memory for a sparse parallel matrix in AIJ format 3965 (the default parallel PETSc format). 3966 3967 Collective 3968 3969 Input Parameters: 3970 + B - the matrix 3971 . i - the indices into j for the start of each local row (starts with zero) 3972 . j - the column indices for each local row (starts with zero) 3973 - v - optional values in the matrix 3974 3975 Level: developer 3976 3977 Notes: 3978 The i, j, and v arrays ARE copied by this routine into the internal format used by PETSc; 3979 thus you CANNOT change the matrix entries by changing the values of v[] after you have 3980 called this routine. Use MatCreateMPIAIJWithSplitArrays() to avoid needing to copy the arrays. 3981 3982 The i and j indices are 0 based, and i indices are indices corresponding to the local j array. 3983 3984 The format which is used for the sparse matrix input, is equivalent to a 3985 row-major ordering.. i.e for the following matrix, the input data expected is 3986 as shown 3987 3988 $ 1 0 0 3989 $ 2 0 3 P0 3990 $ ------- 3991 $ 4 5 6 P1 3992 $ 3993 $ Process0 [P0]: rows_owned=[0,1] 3994 $ i = {0,1,3} [size = nrow+1 = 2+1] 3995 $ j = {0,0,2} [size = 3] 3996 $ v = {1,2,3} [size = 3] 3997 $ 3998 $ Process1 [P1]: rows_owned=[2] 3999 $ i = {0,3} [size = nrow+1 = 1+1] 4000 $ j = {0,1,2} [size = 3] 4001 $ v = {4,5,6} [size = 3] 4002 4003 .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatCreateAIJ(), MATMPIAIJ, 4004 MatCreateSeqAIJWithArrays(), MatCreateMPIAIJWithSplitArrays() 4005 @*/ 4006 PetscErrorCode MatMPIAIJSetPreallocationCSR(Mat B,const PetscInt i[],const PetscInt j[], const PetscScalar v[]) 4007 { 4008 PetscErrorCode ierr; 4009 4010 PetscFunctionBegin; 4011 ierr = PetscTryMethod(B,"MatMPIAIJSetPreallocationCSR_C",(Mat,const PetscInt[],const PetscInt[],const PetscScalar[]),(B,i,j,v));CHKERRQ(ierr); 4012 PetscFunctionReturn(0); 4013 } 4014 4015 /*@C 4016 MatMPIAIJSetPreallocation - Preallocates memory for a sparse parallel matrix in AIJ format 4017 (the default parallel PETSc format). For good matrix assembly performance 4018 the user should preallocate the matrix storage by setting the parameters 4019 d_nz (or d_nnz) and o_nz (or o_nnz). By setting these parameters accurately, 4020 performance can be increased by more than a factor of 50. 4021 4022 Collective 4023 4024 Input Parameters: 4025 + B - the matrix 4026 . d_nz - number of nonzeros per row in DIAGONAL portion of local submatrix 4027 (same value is used for all local rows) 4028 . d_nnz - array containing the number of nonzeros in the various rows of the 4029 DIAGONAL portion of the local submatrix (possibly different for each row) 4030 or NULL (PETSC_NULL_INTEGER in Fortran), if d_nz is used to specify the nonzero structure. 4031 The size of this array is equal to the number of local rows, i.e 'm'. 4032 For matrices that will be factored, you must leave room for (and set) 4033 the diagonal entry even if it is zero. 4034 . o_nz - number of nonzeros per row in the OFF-DIAGONAL portion of local 4035 submatrix (same value is used for all local rows). 4036 - o_nnz - array containing the number of nonzeros in the various rows of the 4037 OFF-DIAGONAL portion of the local submatrix (possibly different for 4038 each row) or NULL (PETSC_NULL_INTEGER in Fortran), if o_nz is used to specify the nonzero 4039 structure. The size of this array is equal to the number 4040 of local rows, i.e 'm'. 4041 4042 If the *_nnz parameter is given then the *_nz parameter is ignored 4043 4044 The AIJ format (also called the Yale sparse matrix format or 4045 compressed row storage (CSR)), is fully compatible with standard Fortran 77 4046 storage. The stored row and column indices begin with zero. 4047 See Users-Manual: ch_mat for details. 4048 4049 The parallel matrix is partitioned such that the first m0 rows belong to 4050 process 0, the next m1 rows belong to process 1, the next m2 rows belong 4051 to process 2 etc.. where m0,m1,m2... are the input parameter 'm'. 4052 4053 The DIAGONAL portion of the local submatrix of a processor can be defined 4054 as the submatrix which is obtained by extraction the part corresponding to 4055 the rows r1-r2 and columns c1-c2 of the global matrix, where r1 is the 4056 first row that belongs to the processor, r2 is the last row belonging to 4057 the this processor, and c1-c2 is range of indices of the local part of a 4058 vector suitable for applying the matrix to. This is an mxn matrix. In the 4059 common case of a square matrix, the row and column ranges are the same and 4060 the DIAGONAL part is also square. The remaining portion of the local 4061 submatrix (mxN) constitute the OFF-DIAGONAL portion. 4062 4063 If o_nnz, d_nnz are specified, then o_nz, and d_nz are ignored. 4064 4065 You can call MatGetInfo() to get information on how effective the preallocation was; 4066 for example the fields mallocs,nz_allocated,nz_used,nz_unneeded; 4067 You can also run with the option -info and look for messages with the string 4068 malloc in them to see if additional memory allocation was needed. 4069 4070 Example usage: 4071 4072 Consider the following 8x8 matrix with 34 non-zero values, that is 4073 assembled across 3 processors. Lets assume that proc0 owns 3 rows, 4074 proc1 owns 3 rows, proc2 owns 2 rows. This division can be shown 4075 as follows: 4076 4077 .vb 4078 1 2 0 | 0 3 0 | 0 4 4079 Proc0 0 5 6 | 7 0 0 | 8 0 4080 9 0 10 | 11 0 0 | 12 0 4081 ------------------------------------- 4082 13 0 14 | 15 16 17 | 0 0 4083 Proc1 0 18 0 | 19 20 21 | 0 0 4084 0 0 0 | 22 23 0 | 24 0 4085 ------------------------------------- 4086 Proc2 25 26 27 | 0 0 28 | 29 0 4087 30 0 0 | 31 32 33 | 0 34 4088 .ve 4089 4090 This can be represented as a collection of submatrices as: 4091 4092 .vb 4093 A B C 4094 D E F 4095 G H I 4096 .ve 4097 4098 Where the submatrices A,B,C are owned by proc0, D,E,F are 4099 owned by proc1, G,H,I are owned by proc2. 4100 4101 The 'm' parameters for proc0,proc1,proc2 are 3,3,2 respectively. 4102 The 'n' parameters for proc0,proc1,proc2 are 3,3,2 respectively. 4103 The 'M','N' parameters are 8,8, and have the same values on all procs. 4104 4105 The DIAGONAL submatrices corresponding to proc0,proc1,proc2 are 4106 submatrices [A], [E], [I] respectively. The OFF-DIAGONAL submatrices 4107 corresponding to proc0,proc1,proc2 are [BC], [DF], [GH] respectively. 4108 Internally, each processor stores the DIAGONAL part, and the OFF-DIAGONAL 4109 part as SeqAIJ matrices. for eg: proc1 will store [E] as a SeqAIJ 4110 matrix, ans [DF] as another SeqAIJ matrix. 4111 4112 When d_nz, o_nz parameters are specified, d_nz storage elements are 4113 allocated for every row of the local diagonal submatrix, and o_nz 4114 storage locations are allocated for every row of the OFF-DIAGONAL submat. 4115 One way to choose d_nz and o_nz is to use the max nonzerors per local 4116 rows for each of the local DIAGONAL, and the OFF-DIAGONAL submatrices. 4117 In this case, the values of d_nz,o_nz are: 4118 .vb 4119 proc0 : dnz = 2, o_nz = 2 4120 proc1 : dnz = 3, o_nz = 2 4121 proc2 : dnz = 1, o_nz = 4 4122 .ve 4123 We are allocating m*(d_nz+o_nz) storage locations for every proc. This 4124 translates to 3*(2+2)=12 for proc0, 3*(3+2)=15 for proc1, 2*(1+4)=10 4125 for proc3. i.e we are using 12+15+10=37 storage locations to store 4126 34 values. 4127 4128 When d_nnz, o_nnz parameters are specified, the storage is specified 4129 for every row, coresponding to both DIAGONAL and OFF-DIAGONAL submatrices. 4130 In the above case the values for d_nnz,o_nnz are: 4131 .vb 4132 proc0: d_nnz = [2,2,2] and o_nnz = [2,2,2] 4133 proc1: d_nnz = [3,3,2] and o_nnz = [2,1,1] 4134 proc2: d_nnz = [1,1] and o_nnz = [4,4] 4135 .ve 4136 Here the space allocated is sum of all the above values i.e 34, and 4137 hence pre-allocation is perfect. 4138 4139 Level: intermediate 4140 4141 .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatCreateAIJ(), MatMPIAIJSetPreallocationCSR(), 4142 MATMPIAIJ, MatGetInfo(), PetscSplitOwnership() 4143 @*/ 4144 PetscErrorCode MatMPIAIJSetPreallocation(Mat B,PetscInt d_nz,const PetscInt d_nnz[],PetscInt o_nz,const PetscInt o_nnz[]) 4145 { 4146 PetscErrorCode ierr; 4147 4148 PetscFunctionBegin; 4149 PetscValidHeaderSpecific(B,MAT_CLASSID,1); 4150 PetscValidType(B,1); 4151 ierr = PetscTryMethod(B,"MatMPIAIJSetPreallocation_C",(Mat,PetscInt,const PetscInt[],PetscInt,const PetscInt[]),(B,d_nz,d_nnz,o_nz,o_nnz));CHKERRQ(ierr); 4152 PetscFunctionReturn(0); 4153 } 4154 4155 /*@ 4156 MatCreateMPIAIJWithArrays - creates a MPI AIJ matrix using arrays that contain in standard 4157 CSR format for the local rows. 4158 4159 Collective 4160 4161 Input Parameters: 4162 + comm - MPI communicator 4163 . m - number of local rows (Cannot be PETSC_DECIDE) 4164 . n - This value should be the same as the local size used in creating the 4165 x vector for the matrix-vector product y = Ax. (or PETSC_DECIDE to have 4166 calculated if N is given) For square matrices n is almost always m. 4167 . M - number of global rows (or PETSC_DETERMINE to have calculated if m is given) 4168 . N - number of global columns (or PETSC_DETERMINE to have calculated if n is given) 4169 . i - row indices; that is i[0] = 0, i[row] = i[row-1] + number of elements in that row of the matrix 4170 . j - column indices 4171 - a - matrix values 4172 4173 Output Parameter: 4174 . mat - the matrix 4175 4176 Level: intermediate 4177 4178 Notes: 4179 The i, j, and a arrays ARE copied by this routine into the internal format used by PETSc; 4180 thus you CANNOT change the matrix entries by changing the values of a[] after you have 4181 called this routine. Use MatCreateMPIAIJWithSplitArrays() to avoid needing to copy the arrays. 4182 4183 The i and j indices are 0 based, and i indices are indices corresponding to the local j array. 4184 4185 The format which is used for the sparse matrix input, is equivalent to a 4186 row-major ordering.. i.e for the following matrix, the input data expected is 4187 as shown 4188 4189 Once you have created the matrix you can update it with new numerical values using MatUpdateMPIAIJWithArrays 4190 4191 $ 1 0 0 4192 $ 2 0 3 P0 4193 $ ------- 4194 $ 4 5 6 P1 4195 $ 4196 $ Process0 [P0]: rows_owned=[0,1] 4197 $ i = {0,1,3} [size = nrow+1 = 2+1] 4198 $ j = {0,0,2} [size = 3] 4199 $ v = {1,2,3} [size = 3] 4200 $ 4201 $ Process1 [P1]: rows_owned=[2] 4202 $ i = {0,3} [size = nrow+1 = 1+1] 4203 $ j = {0,1,2} [size = 3] 4204 $ v = {4,5,6} [size = 3] 4205 4206 .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatMPIAIJSetPreallocationCSR(), 4207 MATMPIAIJ, MatCreateAIJ(), MatCreateMPIAIJWithSplitArrays(), MatUpdateMPIAIJWithArrays() 4208 @*/ 4209 PetscErrorCode MatCreateMPIAIJWithArrays(MPI_Comm comm,PetscInt m,PetscInt n,PetscInt M,PetscInt N,const PetscInt i[],const PetscInt j[],const PetscScalar a[],Mat *mat) 4210 { 4211 PetscErrorCode ierr; 4212 4213 PetscFunctionBegin; 4214 if (i && i[0]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"i (row indices) must start with 0"); 4215 if (m < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"local number of rows (m) cannot be PETSC_DECIDE, or negative"); 4216 ierr = MatCreate(comm,mat);CHKERRQ(ierr); 4217 ierr = MatSetSizes(*mat,m,n,M,N);CHKERRQ(ierr); 4218 /* ierr = MatSetBlockSizes(M,bs,cbs);CHKERRQ(ierr); */ 4219 ierr = MatSetType(*mat,MATMPIAIJ);CHKERRQ(ierr); 4220 ierr = MatMPIAIJSetPreallocationCSR(*mat,i,j,a);CHKERRQ(ierr); 4221 PetscFunctionReturn(0); 4222 } 4223 4224 /*@ 4225 MatUpdateMPIAIJWithArrays - updates a MPI AIJ matrix using arrays that contain in standard 4226 CSR format for the local rows. Only the numerical values are updated the other arrays must be identical 4227 4228 Collective 4229 4230 Input Parameters: 4231 + mat - the matrix 4232 . m - number of local rows (Cannot be PETSC_DECIDE) 4233 . n - This value should be the same as the local size used in creating the 4234 x vector for the matrix-vector product y = Ax. (or PETSC_DECIDE to have 4235 calculated if N is given) For square matrices n is almost always m. 4236 . M - number of global rows (or PETSC_DETERMINE to have calculated if m is given) 4237 . N - number of global columns (or PETSC_DETERMINE to have calculated if n is given) 4238 . Ii - row indices; that is Ii[0] = 0, Ii[row] = Ii[row-1] + number of elements in that row of the matrix 4239 . J - column indices 4240 - v - matrix values 4241 4242 Level: intermediate 4243 4244 .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatMPIAIJSetPreallocationCSR(), 4245 MATMPIAIJ, MatCreateAIJ(), MatCreateMPIAIJWithSplitArrays(), MatUpdateMPIAIJWithArrays() 4246 @*/ 4247 PetscErrorCode MatUpdateMPIAIJWithArrays(Mat mat,PetscInt m,PetscInt n,PetscInt M,PetscInt N,const PetscInt Ii[],const PetscInt J[],const PetscScalar v[]) 4248 { 4249 PetscErrorCode ierr; 4250 PetscInt cstart,nnz,i,j; 4251 PetscInt *ld; 4252 PetscBool nooffprocentries; 4253 Mat_MPIAIJ *Aij = (Mat_MPIAIJ*)mat->data; 4254 Mat_SeqAIJ *Ad = (Mat_SeqAIJ*)Aij->A->data, *Ao = (Mat_SeqAIJ*)Aij->B->data; 4255 PetscScalar *ad = Ad->a, *ao = Ao->a; 4256 const PetscInt *Adi = Ad->i; 4257 PetscInt ldi,Iii,md; 4258 4259 PetscFunctionBegin; 4260 if (Ii[0]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"i (row indices) must start with 0"); 4261 if (m < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"local number of rows (m) cannot be PETSC_DECIDE, or negative"); 4262 if (m != mat->rmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Local number of rows cannot change from call to MatUpdateMPIAIJWithArrays()"); 4263 if (n != mat->cmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Local number of columns cannot change from call to MatUpdateMPIAIJWithArrays()"); 4264 4265 cstart = mat->cmap->rstart; 4266 if (!Aij->ld) { 4267 /* count number of entries below block diagonal */ 4268 ierr = PetscCalloc1(m,&ld);CHKERRQ(ierr); 4269 Aij->ld = ld; 4270 for (i=0; i<m; i++) { 4271 nnz = Ii[i+1]- Ii[i]; 4272 j = 0; 4273 while (J[j] < cstart && j < nnz) {j++;} 4274 J += nnz; 4275 ld[i] = j; 4276 } 4277 } else { 4278 ld = Aij->ld; 4279 } 4280 4281 for (i=0; i<m; i++) { 4282 nnz = Ii[i+1]- Ii[i]; 4283 Iii = Ii[i]; 4284 ldi = ld[i]; 4285 md = Adi[i+1]-Adi[i]; 4286 ierr = PetscArraycpy(ao,v + Iii,ldi);CHKERRQ(ierr); 4287 ierr = PetscArraycpy(ad,v + Iii + ldi,md);CHKERRQ(ierr); 4288 ierr = PetscArraycpy(ao + ldi,v + Iii + ldi + md,nnz - ldi - md);CHKERRQ(ierr); 4289 ad += md; 4290 ao += nnz - md; 4291 } 4292 nooffprocentries = mat->nooffprocentries; 4293 mat->nooffprocentries = PETSC_TRUE; 4294 ierr = PetscObjectStateIncrease((PetscObject)Aij->A);CHKERRQ(ierr); 4295 ierr = PetscObjectStateIncrease((PetscObject)Aij->B);CHKERRQ(ierr); 4296 ierr = PetscObjectStateIncrease((PetscObject)mat);CHKERRQ(ierr); 4297 ierr = MatAssemblyBegin(mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4298 ierr = MatAssemblyEnd(mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4299 mat->nooffprocentries = nooffprocentries; 4300 PetscFunctionReturn(0); 4301 } 4302 4303 /*@C 4304 MatCreateAIJ - Creates a sparse parallel matrix in AIJ format 4305 (the default parallel PETSc format). For good matrix assembly performance 4306 the user should preallocate the matrix storage by setting the parameters 4307 d_nz (or d_nnz) and o_nz (or o_nnz). By setting these parameters accurately, 4308 performance can be increased by more than a factor of 50. 4309 4310 Collective 4311 4312 Input Parameters: 4313 + comm - MPI communicator 4314 . m - number of local rows (or PETSC_DECIDE to have calculated if M is given) 4315 This value should be the same as the local size used in creating the 4316 y vector for the matrix-vector product y = Ax. 4317 . n - This value should be the same as the local size used in creating the 4318 x vector for the matrix-vector product y = Ax. (or PETSC_DECIDE to have 4319 calculated if N is given) For square matrices n is almost always m. 4320 . M - number of global rows (or PETSC_DETERMINE to have calculated if m is given) 4321 . N - number of global columns (or PETSC_DETERMINE to have calculated if n is given) 4322 . d_nz - number of nonzeros per row in DIAGONAL portion of local submatrix 4323 (same value is used for all local rows) 4324 . d_nnz - array containing the number of nonzeros in the various rows of the 4325 DIAGONAL portion of the local submatrix (possibly different for each row) 4326 or NULL, if d_nz is used to specify the nonzero structure. 4327 The size of this array is equal to the number of local rows, i.e 'm'. 4328 . o_nz - number of nonzeros per row in the OFF-DIAGONAL portion of local 4329 submatrix (same value is used for all local rows). 4330 - o_nnz - array containing the number of nonzeros in the various rows of the 4331 OFF-DIAGONAL portion of the local submatrix (possibly different for 4332 each row) or NULL, if o_nz is used to specify the nonzero 4333 structure. The size of this array is equal to the number 4334 of local rows, i.e 'm'. 4335 4336 Output Parameter: 4337 . A - the matrix 4338 4339 It is recommended that one use the MatCreate(), MatSetType() and/or MatSetFromOptions(), 4340 MatXXXXSetPreallocation() paradigm instead of this routine directly. 4341 [MatXXXXSetPreallocation() is, for example, MatSeqAIJSetPreallocation] 4342 4343 Notes: 4344 If the *_nnz parameter is given then the *_nz parameter is ignored 4345 4346 m,n,M,N parameters specify the size of the matrix, and its partitioning across 4347 processors, while d_nz,d_nnz,o_nz,o_nnz parameters specify the approximate 4348 storage requirements for this matrix. 4349 4350 If PETSC_DECIDE or PETSC_DETERMINE is used for a particular argument on one 4351 processor than it must be used on all processors that share the object for 4352 that argument. 4353 4354 The user MUST specify either the local or global matrix dimensions 4355 (possibly both). 4356 4357 The parallel matrix is partitioned across processors such that the 4358 first m0 rows belong to process 0, the next m1 rows belong to 4359 process 1, the next m2 rows belong to process 2 etc.. where 4360 m0,m1,m2,.. are the input parameter 'm'. i.e each processor stores 4361 values corresponding to [m x N] submatrix. 4362 4363 The columns are logically partitioned with the n0 columns belonging 4364 to 0th partition, the next n1 columns belonging to the next 4365 partition etc.. where n0,n1,n2... are the input parameter 'n'. 4366 4367 The DIAGONAL portion of the local submatrix on any given processor 4368 is the submatrix corresponding to the rows and columns m,n 4369 corresponding to the given processor. i.e diagonal matrix on 4370 process 0 is [m0 x n0], diagonal matrix on process 1 is [m1 x n1] 4371 etc. The remaining portion of the local submatrix [m x (N-n)] 4372 constitute the OFF-DIAGONAL portion. The example below better 4373 illustrates this concept. 4374 4375 For a square global matrix we define each processor's diagonal portion 4376 to be its local rows and the corresponding columns (a square submatrix); 4377 each processor's off-diagonal portion encompasses the remainder of the 4378 local matrix (a rectangular submatrix). 4379 4380 If o_nnz, d_nnz are specified, then o_nz, and d_nz are ignored. 4381 4382 When calling this routine with a single process communicator, a matrix of 4383 type SEQAIJ is returned. If a matrix of type MPIAIJ is desired for this 4384 type of communicator, use the construction mechanism 4385 .vb 4386 MatCreate(...,&A); MatSetType(A,MATMPIAIJ); MatSetSizes(A, m,n,M,N); MatMPIAIJSetPreallocation(A,...); 4387 .ve 4388 4389 $ MatCreate(...,&A); 4390 $ MatSetType(A,MATMPIAIJ); 4391 $ MatSetSizes(A, m,n,M,N); 4392 $ MatMPIAIJSetPreallocation(A,...); 4393 4394 By default, this format uses inodes (identical nodes) when possible. 4395 We search for consecutive rows with the same nonzero structure, thereby 4396 reusing matrix information to achieve increased efficiency. 4397 4398 Options Database Keys: 4399 + -mat_no_inode - Do not use inodes 4400 - -mat_inode_limit <limit> - Sets inode limit (max limit=5) 4401 4402 4403 4404 Example usage: 4405 4406 Consider the following 8x8 matrix with 34 non-zero values, that is 4407 assembled across 3 processors. Lets assume that proc0 owns 3 rows, 4408 proc1 owns 3 rows, proc2 owns 2 rows. This division can be shown 4409 as follows 4410 4411 .vb 4412 1 2 0 | 0 3 0 | 0 4 4413 Proc0 0 5 6 | 7 0 0 | 8 0 4414 9 0 10 | 11 0 0 | 12 0 4415 ------------------------------------- 4416 13 0 14 | 15 16 17 | 0 0 4417 Proc1 0 18 0 | 19 20 21 | 0 0 4418 0 0 0 | 22 23 0 | 24 0 4419 ------------------------------------- 4420 Proc2 25 26 27 | 0 0 28 | 29 0 4421 30 0 0 | 31 32 33 | 0 34 4422 .ve 4423 4424 This can be represented as a collection of submatrices as 4425 4426 .vb 4427 A B C 4428 D E F 4429 G H I 4430 .ve 4431 4432 Where the submatrices A,B,C are owned by proc0, D,E,F are 4433 owned by proc1, G,H,I are owned by proc2. 4434 4435 The 'm' parameters for proc0,proc1,proc2 are 3,3,2 respectively. 4436 The 'n' parameters for proc0,proc1,proc2 are 3,3,2 respectively. 4437 The 'M','N' parameters are 8,8, and have the same values on all procs. 4438 4439 The DIAGONAL submatrices corresponding to proc0,proc1,proc2 are 4440 submatrices [A], [E], [I] respectively. The OFF-DIAGONAL submatrices 4441 corresponding to proc0,proc1,proc2 are [BC], [DF], [GH] respectively. 4442 Internally, each processor stores the DIAGONAL part, and the OFF-DIAGONAL 4443 part as SeqAIJ matrices. for eg: proc1 will store [E] as a SeqAIJ 4444 matrix, ans [DF] as another SeqAIJ matrix. 4445 4446 When d_nz, o_nz parameters are specified, d_nz storage elements are 4447 allocated for every row of the local diagonal submatrix, and o_nz 4448 storage locations are allocated for every row of the OFF-DIAGONAL submat. 4449 One way to choose d_nz and o_nz is to use the max nonzerors per local 4450 rows for each of the local DIAGONAL, and the OFF-DIAGONAL submatrices. 4451 In this case, the values of d_nz,o_nz are 4452 .vb 4453 proc0 : dnz = 2, o_nz = 2 4454 proc1 : dnz = 3, o_nz = 2 4455 proc2 : dnz = 1, o_nz = 4 4456 .ve 4457 We are allocating m*(d_nz+o_nz) storage locations for every proc. This 4458 translates to 3*(2+2)=12 for proc0, 3*(3+2)=15 for proc1, 2*(1+4)=10 4459 for proc3. i.e we are using 12+15+10=37 storage locations to store 4460 34 values. 4461 4462 When d_nnz, o_nnz parameters are specified, the storage is specified 4463 for every row, coresponding to both DIAGONAL and OFF-DIAGONAL submatrices. 4464 In the above case the values for d_nnz,o_nnz are 4465 .vb 4466 proc0: d_nnz = [2,2,2] and o_nnz = [2,2,2] 4467 proc1: d_nnz = [3,3,2] and o_nnz = [2,1,1] 4468 proc2: d_nnz = [1,1] and o_nnz = [4,4] 4469 .ve 4470 Here the space allocated is sum of all the above values i.e 34, and 4471 hence pre-allocation is perfect. 4472 4473 Level: intermediate 4474 4475 .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatMPIAIJSetPreallocationCSR(), 4476 MATMPIAIJ, MatCreateMPIAIJWithArrays() 4477 @*/ 4478 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) 4479 { 4480 PetscErrorCode ierr; 4481 PetscMPIInt size; 4482 4483 PetscFunctionBegin; 4484 ierr = MatCreate(comm,A);CHKERRQ(ierr); 4485 ierr = MatSetSizes(*A,m,n,M,N);CHKERRQ(ierr); 4486 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 4487 if (size > 1) { 4488 ierr = MatSetType(*A,MATMPIAIJ);CHKERRQ(ierr); 4489 ierr = MatMPIAIJSetPreallocation(*A,d_nz,d_nnz,o_nz,o_nnz);CHKERRQ(ierr); 4490 } else { 4491 ierr = MatSetType(*A,MATSEQAIJ);CHKERRQ(ierr); 4492 ierr = MatSeqAIJSetPreallocation(*A,d_nz,d_nnz);CHKERRQ(ierr); 4493 } 4494 PetscFunctionReturn(0); 4495 } 4496 4497 PetscErrorCode MatMPIAIJGetSeqAIJ(Mat A,Mat *Ad,Mat *Ao,const PetscInt *colmap[]) 4498 { 4499 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 4500 PetscBool flg; 4501 PetscErrorCode ierr; 4502 4503 PetscFunctionBegin; 4504 ierr = PetscStrbeginswith(((PetscObject)A)->type_name,MATMPIAIJ,&flg);CHKERRQ(ierr); 4505 if (!flg) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"This function requires a MATMPIAIJ matrix as input"); 4506 if (Ad) *Ad = a->A; 4507 if (Ao) *Ao = a->B; 4508 if (colmap) *colmap = a->garray; 4509 PetscFunctionReturn(0); 4510 } 4511 4512 PetscErrorCode MatCreateMPIMatConcatenateSeqMat_MPIAIJ(MPI_Comm comm,Mat inmat,PetscInt n,MatReuse scall,Mat *outmat) 4513 { 4514 PetscErrorCode ierr; 4515 PetscInt m,N,i,rstart,nnz,Ii; 4516 PetscInt *indx; 4517 PetscScalar *values; 4518 4519 PetscFunctionBegin; 4520 ierr = MatGetSize(inmat,&m,&N);CHKERRQ(ierr); 4521 if (scall == MAT_INITIAL_MATRIX) { /* symbolic phase */ 4522 PetscInt *dnz,*onz,sum,bs,cbs; 4523 4524 if (n == PETSC_DECIDE) { 4525 ierr = PetscSplitOwnership(comm,&n,&N);CHKERRQ(ierr); 4526 } 4527 /* Check sum(n) = N */ 4528 ierr = MPIU_Allreduce(&n,&sum,1,MPIU_INT,MPI_SUM,comm);CHKERRQ(ierr); 4529 if (sum != N) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_INCOMP,"Sum of local columns %D != global columns %D",sum,N); 4530 4531 ierr = MPI_Scan(&m, &rstart,1,MPIU_INT,MPI_SUM,comm);CHKERRQ(ierr); 4532 rstart -= m; 4533 4534 ierr = MatPreallocateInitialize(comm,m,n,dnz,onz);CHKERRQ(ierr); 4535 for (i=0; i<m; i++) { 4536 ierr = MatGetRow_SeqAIJ(inmat,i,&nnz,&indx,NULL);CHKERRQ(ierr); 4537 ierr = MatPreallocateSet(i+rstart,nnz,indx,dnz,onz);CHKERRQ(ierr); 4538 ierr = MatRestoreRow_SeqAIJ(inmat,i,&nnz,&indx,NULL);CHKERRQ(ierr); 4539 } 4540 4541 ierr = MatCreate(comm,outmat);CHKERRQ(ierr); 4542 ierr = MatSetSizes(*outmat,m,n,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 4543 ierr = MatGetBlockSizes(inmat,&bs,&cbs);CHKERRQ(ierr); 4544 ierr = MatSetBlockSizes(*outmat,bs,cbs);CHKERRQ(ierr); 4545 ierr = MatSetType(*outmat,MATAIJ);CHKERRQ(ierr); 4546 ierr = MatSeqAIJSetPreallocation(*outmat,0,dnz);CHKERRQ(ierr); 4547 ierr = MatMPIAIJSetPreallocation(*outmat,0,dnz,0,onz);CHKERRQ(ierr); 4548 ierr = MatPreallocateFinalize(dnz,onz);CHKERRQ(ierr); 4549 } 4550 4551 /* numeric phase */ 4552 ierr = MatGetOwnershipRange(*outmat,&rstart,NULL);CHKERRQ(ierr); 4553 for (i=0; i<m; i++) { 4554 ierr = MatGetRow_SeqAIJ(inmat,i,&nnz,&indx,&values);CHKERRQ(ierr); 4555 Ii = i + rstart; 4556 ierr = MatSetValues(*outmat,1,&Ii,nnz,indx,values,INSERT_VALUES);CHKERRQ(ierr); 4557 ierr = MatRestoreRow_SeqAIJ(inmat,i,&nnz,&indx,&values);CHKERRQ(ierr); 4558 } 4559 ierr = MatAssemblyBegin(*outmat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4560 ierr = MatAssemblyEnd(*outmat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4561 PetscFunctionReturn(0); 4562 } 4563 4564 PetscErrorCode MatFileSplit(Mat A,char *outfile) 4565 { 4566 PetscErrorCode ierr; 4567 PetscMPIInt rank; 4568 PetscInt m,N,i,rstart,nnz; 4569 size_t len; 4570 const PetscInt *indx; 4571 PetscViewer out; 4572 char *name; 4573 Mat B; 4574 const PetscScalar *values; 4575 4576 PetscFunctionBegin; 4577 ierr = MatGetLocalSize(A,&m,0);CHKERRQ(ierr); 4578 ierr = MatGetSize(A,0,&N);CHKERRQ(ierr); 4579 /* Should this be the type of the diagonal block of A? */ 4580 ierr = MatCreate(PETSC_COMM_SELF,&B);CHKERRQ(ierr); 4581 ierr = MatSetSizes(B,m,N,m,N);CHKERRQ(ierr); 4582 ierr = MatSetBlockSizesFromMats(B,A,A);CHKERRQ(ierr); 4583 ierr = MatSetType(B,MATSEQAIJ);CHKERRQ(ierr); 4584 ierr = MatSeqAIJSetPreallocation(B,0,NULL);CHKERRQ(ierr); 4585 ierr = MatGetOwnershipRange(A,&rstart,0);CHKERRQ(ierr); 4586 for (i=0; i<m; i++) { 4587 ierr = MatGetRow(A,i+rstart,&nnz,&indx,&values);CHKERRQ(ierr); 4588 ierr = MatSetValues(B,1,&i,nnz,indx,values,INSERT_VALUES);CHKERRQ(ierr); 4589 ierr = MatRestoreRow(A,i+rstart,&nnz,&indx,&values);CHKERRQ(ierr); 4590 } 4591 ierr = MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4592 ierr = MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4593 4594 ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)A),&rank);CHKERRQ(ierr); 4595 ierr = PetscStrlen(outfile,&len);CHKERRQ(ierr); 4596 ierr = PetscMalloc1(len+5,&name);CHKERRQ(ierr); 4597 sprintf(name,"%s.%d",outfile,rank); 4598 ierr = PetscViewerBinaryOpen(PETSC_COMM_SELF,name,FILE_MODE_APPEND,&out);CHKERRQ(ierr); 4599 ierr = PetscFree(name);CHKERRQ(ierr); 4600 ierr = MatView(B,out);CHKERRQ(ierr); 4601 ierr = PetscViewerDestroy(&out);CHKERRQ(ierr); 4602 ierr = MatDestroy(&B);CHKERRQ(ierr); 4603 PetscFunctionReturn(0); 4604 } 4605 4606 PetscErrorCode MatDestroy_MPIAIJ_SeqsToMPI(Mat A) 4607 { 4608 PetscErrorCode ierr; 4609 Mat_Merge_SeqsToMPI *merge; 4610 PetscContainer container; 4611 4612 PetscFunctionBegin; 4613 ierr = PetscObjectQuery((PetscObject)A,"MatMergeSeqsToMPI",(PetscObject*)&container);CHKERRQ(ierr); 4614 if (container) { 4615 ierr = PetscContainerGetPointer(container,(void**)&merge);CHKERRQ(ierr); 4616 ierr = PetscFree(merge->id_r);CHKERRQ(ierr); 4617 ierr = PetscFree(merge->len_s);CHKERRQ(ierr); 4618 ierr = PetscFree(merge->len_r);CHKERRQ(ierr); 4619 ierr = PetscFree(merge->bi);CHKERRQ(ierr); 4620 ierr = PetscFree(merge->bj);CHKERRQ(ierr); 4621 ierr = PetscFree(merge->buf_ri[0]);CHKERRQ(ierr); 4622 ierr = PetscFree(merge->buf_ri);CHKERRQ(ierr); 4623 ierr = PetscFree(merge->buf_rj[0]);CHKERRQ(ierr); 4624 ierr = PetscFree(merge->buf_rj);CHKERRQ(ierr); 4625 ierr = PetscFree(merge->coi);CHKERRQ(ierr); 4626 ierr = PetscFree(merge->coj);CHKERRQ(ierr); 4627 ierr = PetscFree(merge->owners_co);CHKERRQ(ierr); 4628 ierr = PetscLayoutDestroy(&merge->rowmap);CHKERRQ(ierr); 4629 ierr = PetscFree(merge);CHKERRQ(ierr); 4630 ierr = PetscObjectCompose((PetscObject)A,"MatMergeSeqsToMPI",0);CHKERRQ(ierr); 4631 } 4632 ierr = MatDestroy_MPIAIJ(A);CHKERRQ(ierr); 4633 PetscFunctionReturn(0); 4634 } 4635 4636 #include <../src/mat/utils/freespace.h> 4637 #include <petscbt.h> 4638 4639 PetscErrorCode MatCreateMPIAIJSumSeqAIJNumeric(Mat seqmat,Mat mpimat) 4640 { 4641 PetscErrorCode ierr; 4642 MPI_Comm comm; 4643 Mat_SeqAIJ *a =(Mat_SeqAIJ*)seqmat->data; 4644 PetscMPIInt size,rank,taga,*len_s; 4645 PetscInt N=mpimat->cmap->N,i,j,*owners,*ai=a->i,*aj; 4646 PetscInt proc,m; 4647 PetscInt **buf_ri,**buf_rj; 4648 PetscInt k,anzi,*bj_i,*bi,*bj,arow,bnzi,nextaj; 4649 PetscInt nrows,**buf_ri_k,**nextrow,**nextai; 4650 MPI_Request *s_waits,*r_waits; 4651 MPI_Status *status; 4652 MatScalar *aa=a->a; 4653 MatScalar **abuf_r,*ba_i; 4654 Mat_Merge_SeqsToMPI *merge; 4655 PetscContainer container; 4656 4657 PetscFunctionBegin; 4658 ierr = PetscObjectGetComm((PetscObject)mpimat,&comm);CHKERRQ(ierr); 4659 ierr = PetscLogEventBegin(MAT_Seqstompinum,seqmat,0,0,0);CHKERRQ(ierr); 4660 4661 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 4662 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 4663 4664 ierr = PetscObjectQuery((PetscObject)mpimat,"MatMergeSeqsToMPI",(PetscObject*)&container);CHKERRQ(ierr); 4665 ierr = PetscContainerGetPointer(container,(void**)&merge);CHKERRQ(ierr); 4666 4667 bi = merge->bi; 4668 bj = merge->bj; 4669 buf_ri = merge->buf_ri; 4670 buf_rj = merge->buf_rj; 4671 4672 ierr = PetscMalloc1(size,&status);CHKERRQ(ierr); 4673 owners = merge->rowmap->range; 4674 len_s = merge->len_s; 4675 4676 /* send and recv matrix values */ 4677 /*-----------------------------*/ 4678 ierr = PetscObjectGetNewTag((PetscObject)mpimat,&taga);CHKERRQ(ierr); 4679 ierr = PetscPostIrecvScalar(comm,taga,merge->nrecv,merge->id_r,merge->len_r,&abuf_r,&r_waits);CHKERRQ(ierr); 4680 4681 ierr = PetscMalloc1(merge->nsend+1,&s_waits);CHKERRQ(ierr); 4682 for (proc=0,k=0; proc<size; proc++) { 4683 if (!len_s[proc]) continue; 4684 i = owners[proc]; 4685 ierr = MPI_Isend(aa+ai[i],len_s[proc],MPIU_MATSCALAR,proc,taga,comm,s_waits+k);CHKERRQ(ierr); 4686 k++; 4687 } 4688 4689 if (merge->nrecv) {ierr = MPI_Waitall(merge->nrecv,r_waits,status);CHKERRQ(ierr);} 4690 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,s_waits,status);CHKERRQ(ierr);} 4691 ierr = PetscFree(status);CHKERRQ(ierr); 4692 4693 ierr = PetscFree(s_waits);CHKERRQ(ierr); 4694 ierr = PetscFree(r_waits);CHKERRQ(ierr); 4695 4696 /* insert mat values of mpimat */ 4697 /*----------------------------*/ 4698 ierr = PetscMalloc1(N,&ba_i);CHKERRQ(ierr); 4699 ierr = PetscMalloc3(merge->nrecv,&buf_ri_k,merge->nrecv,&nextrow,merge->nrecv,&nextai);CHKERRQ(ierr); 4700 4701 for (k=0; k<merge->nrecv; k++) { 4702 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 4703 nrows = *(buf_ri_k[k]); 4704 nextrow[k] = buf_ri_k[k]+1; /* next row number of k-th recved i-structure */ 4705 nextai[k] = buf_ri_k[k] + (nrows + 1); /* poins to the next i-structure of k-th recved i-structure */ 4706 } 4707 4708 /* set values of ba */ 4709 m = merge->rowmap->n; 4710 for (i=0; i<m; i++) { 4711 arow = owners[rank] + i; 4712 bj_i = bj+bi[i]; /* col indices of the i-th row of mpimat */ 4713 bnzi = bi[i+1] - bi[i]; 4714 ierr = PetscArrayzero(ba_i,bnzi);CHKERRQ(ierr); 4715 4716 /* add local non-zero vals of this proc's seqmat into ba */ 4717 anzi = ai[arow+1] - ai[arow]; 4718 aj = a->j + ai[arow]; 4719 aa = a->a + ai[arow]; 4720 nextaj = 0; 4721 for (j=0; nextaj<anzi; j++) { 4722 if (*(bj_i + j) == aj[nextaj]) { /* bcol == acol */ 4723 ba_i[j] += aa[nextaj++]; 4724 } 4725 } 4726 4727 /* add received vals into ba */ 4728 for (k=0; k<merge->nrecv; k++) { /* k-th received message */ 4729 /* i-th row */ 4730 if (i == *nextrow[k]) { 4731 anzi = *(nextai[k]+1) - *nextai[k]; 4732 aj = buf_rj[k] + *(nextai[k]); 4733 aa = abuf_r[k] + *(nextai[k]); 4734 nextaj = 0; 4735 for (j=0; nextaj<anzi; j++) { 4736 if (*(bj_i + j) == aj[nextaj]) { /* bcol == acol */ 4737 ba_i[j] += aa[nextaj++]; 4738 } 4739 } 4740 nextrow[k]++; nextai[k]++; 4741 } 4742 } 4743 ierr = MatSetValues(mpimat,1,&arow,bnzi,bj_i,ba_i,INSERT_VALUES);CHKERRQ(ierr); 4744 } 4745 ierr = MatAssemblyBegin(mpimat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4746 ierr = MatAssemblyEnd(mpimat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4747 4748 ierr = PetscFree(abuf_r[0]);CHKERRQ(ierr); 4749 ierr = PetscFree(abuf_r);CHKERRQ(ierr); 4750 ierr = PetscFree(ba_i);CHKERRQ(ierr); 4751 ierr = PetscFree3(buf_ri_k,nextrow,nextai);CHKERRQ(ierr); 4752 ierr = PetscLogEventEnd(MAT_Seqstompinum,seqmat,0,0,0);CHKERRQ(ierr); 4753 PetscFunctionReturn(0); 4754 } 4755 4756 PetscErrorCode MatCreateMPIAIJSumSeqAIJSymbolic(MPI_Comm comm,Mat seqmat,PetscInt m,PetscInt n,Mat *mpimat) 4757 { 4758 PetscErrorCode ierr; 4759 Mat B_mpi; 4760 Mat_SeqAIJ *a=(Mat_SeqAIJ*)seqmat->data; 4761 PetscMPIInt size,rank,tagi,tagj,*len_s,*len_si,*len_ri; 4762 PetscInt **buf_rj,**buf_ri,**buf_ri_k; 4763 PetscInt M=seqmat->rmap->n,N=seqmat->cmap->n,i,*owners,*ai=a->i,*aj=a->j; 4764 PetscInt len,proc,*dnz,*onz,bs,cbs; 4765 PetscInt k,anzi,*bi,*bj,*lnk,nlnk,arow,bnzi,nspacedouble=0; 4766 PetscInt nrows,*buf_s,*buf_si,*buf_si_i,**nextrow,**nextai; 4767 MPI_Request *si_waits,*sj_waits,*ri_waits,*rj_waits; 4768 MPI_Status *status; 4769 PetscFreeSpaceList free_space=NULL,current_space=NULL; 4770 PetscBT lnkbt; 4771 Mat_Merge_SeqsToMPI *merge; 4772 PetscContainer container; 4773 4774 PetscFunctionBegin; 4775 ierr = PetscLogEventBegin(MAT_Seqstompisym,seqmat,0,0,0);CHKERRQ(ierr); 4776 4777 /* make sure it is a PETSc comm */ 4778 ierr = PetscCommDuplicate(comm,&comm,NULL);CHKERRQ(ierr); 4779 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 4780 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 4781 4782 ierr = PetscNew(&merge);CHKERRQ(ierr); 4783 ierr = PetscMalloc1(size,&status);CHKERRQ(ierr); 4784 4785 /* determine row ownership */ 4786 /*---------------------------------------------------------*/ 4787 ierr = PetscLayoutCreate(comm,&merge->rowmap);CHKERRQ(ierr); 4788 ierr = PetscLayoutSetLocalSize(merge->rowmap,m);CHKERRQ(ierr); 4789 ierr = PetscLayoutSetSize(merge->rowmap,M);CHKERRQ(ierr); 4790 ierr = PetscLayoutSetBlockSize(merge->rowmap,1);CHKERRQ(ierr); 4791 ierr = PetscLayoutSetUp(merge->rowmap);CHKERRQ(ierr); 4792 ierr = PetscMalloc1(size,&len_si);CHKERRQ(ierr); 4793 ierr = PetscMalloc1(size,&merge->len_s);CHKERRQ(ierr); 4794 4795 m = merge->rowmap->n; 4796 owners = merge->rowmap->range; 4797 4798 /* determine the number of messages to send, their lengths */ 4799 /*---------------------------------------------------------*/ 4800 len_s = merge->len_s; 4801 4802 len = 0; /* length of buf_si[] */ 4803 merge->nsend = 0; 4804 for (proc=0; proc<size; proc++) { 4805 len_si[proc] = 0; 4806 if (proc == rank) { 4807 len_s[proc] = 0; 4808 } else { 4809 len_si[proc] = owners[proc+1] - owners[proc] + 1; 4810 len_s[proc] = ai[owners[proc+1]] - ai[owners[proc]]; /* num of rows to be sent to [proc] */ 4811 } 4812 if (len_s[proc]) { 4813 merge->nsend++; 4814 nrows = 0; 4815 for (i=owners[proc]; i<owners[proc+1]; i++) { 4816 if (ai[i+1] > ai[i]) nrows++; 4817 } 4818 len_si[proc] = 2*(nrows+1); 4819 len += len_si[proc]; 4820 } 4821 } 4822 4823 /* determine the number and length of messages to receive for ij-structure */ 4824 /*-------------------------------------------------------------------------*/ 4825 ierr = PetscGatherNumberOfMessages(comm,NULL,len_s,&merge->nrecv);CHKERRQ(ierr); 4826 ierr = PetscGatherMessageLengths2(comm,merge->nsend,merge->nrecv,len_s,len_si,&merge->id_r,&merge->len_r,&len_ri);CHKERRQ(ierr); 4827 4828 /* post the Irecv of j-structure */ 4829 /*-------------------------------*/ 4830 ierr = PetscCommGetNewTag(comm,&tagj);CHKERRQ(ierr); 4831 ierr = PetscPostIrecvInt(comm,tagj,merge->nrecv,merge->id_r,merge->len_r,&buf_rj,&rj_waits);CHKERRQ(ierr); 4832 4833 /* post the Isend of j-structure */ 4834 /*--------------------------------*/ 4835 ierr = PetscMalloc2(merge->nsend,&si_waits,merge->nsend,&sj_waits);CHKERRQ(ierr); 4836 4837 for (proc=0, k=0; proc<size; proc++) { 4838 if (!len_s[proc]) continue; 4839 i = owners[proc]; 4840 ierr = MPI_Isend(aj+ai[i],len_s[proc],MPIU_INT,proc,tagj,comm,sj_waits+k);CHKERRQ(ierr); 4841 k++; 4842 } 4843 4844 /* receives and sends of j-structure are complete */ 4845 /*------------------------------------------------*/ 4846 if (merge->nrecv) {ierr = MPI_Waitall(merge->nrecv,rj_waits,status);CHKERRQ(ierr);} 4847 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,sj_waits,status);CHKERRQ(ierr);} 4848 4849 /* send and recv i-structure */ 4850 /*---------------------------*/ 4851 ierr = PetscCommGetNewTag(comm,&tagi);CHKERRQ(ierr); 4852 ierr = PetscPostIrecvInt(comm,tagi,merge->nrecv,merge->id_r,len_ri,&buf_ri,&ri_waits);CHKERRQ(ierr); 4853 4854 ierr = PetscMalloc1(len+1,&buf_s);CHKERRQ(ierr); 4855 buf_si = buf_s; /* points to the beginning of k-th msg to be sent */ 4856 for (proc=0,k=0; proc<size; proc++) { 4857 if (!len_s[proc]) continue; 4858 /* form outgoing message for i-structure: 4859 buf_si[0]: nrows to be sent 4860 [1:nrows]: row index (global) 4861 [nrows+1:2*nrows+1]: i-structure index 4862 */ 4863 /*-------------------------------------------*/ 4864 nrows = len_si[proc]/2 - 1; 4865 buf_si_i = buf_si + nrows+1; 4866 buf_si[0] = nrows; 4867 buf_si_i[0] = 0; 4868 nrows = 0; 4869 for (i=owners[proc]; i<owners[proc+1]; i++) { 4870 anzi = ai[i+1] - ai[i]; 4871 if (anzi) { 4872 buf_si_i[nrows+1] = buf_si_i[nrows] + anzi; /* i-structure */ 4873 buf_si[nrows+1] = i-owners[proc]; /* local row index */ 4874 nrows++; 4875 } 4876 } 4877 ierr = MPI_Isend(buf_si,len_si[proc],MPIU_INT,proc,tagi,comm,si_waits+k);CHKERRQ(ierr); 4878 k++; 4879 buf_si += len_si[proc]; 4880 } 4881 4882 if (merge->nrecv) {ierr = MPI_Waitall(merge->nrecv,ri_waits,status);CHKERRQ(ierr);} 4883 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,si_waits,status);CHKERRQ(ierr);} 4884 4885 ierr = PetscInfo2(seqmat,"nsend: %D, nrecv: %D\n",merge->nsend,merge->nrecv);CHKERRQ(ierr); 4886 for (i=0; i<merge->nrecv; i++) { 4887 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); 4888 } 4889 4890 ierr = PetscFree(len_si);CHKERRQ(ierr); 4891 ierr = PetscFree(len_ri);CHKERRQ(ierr); 4892 ierr = PetscFree(rj_waits);CHKERRQ(ierr); 4893 ierr = PetscFree2(si_waits,sj_waits);CHKERRQ(ierr); 4894 ierr = PetscFree(ri_waits);CHKERRQ(ierr); 4895 ierr = PetscFree(buf_s);CHKERRQ(ierr); 4896 ierr = PetscFree(status);CHKERRQ(ierr); 4897 4898 /* compute a local seq matrix in each processor */ 4899 /*----------------------------------------------*/ 4900 /* allocate bi array and free space for accumulating nonzero column info */ 4901 ierr = PetscMalloc1(m+1,&bi);CHKERRQ(ierr); 4902 bi[0] = 0; 4903 4904 /* create and initialize a linked list */ 4905 nlnk = N+1; 4906 ierr = PetscLLCreate(N,N,nlnk,lnk,lnkbt);CHKERRQ(ierr); 4907 4908 /* initial FreeSpace size is 2*(num of local nnz(seqmat)) */ 4909 len = ai[owners[rank+1]] - ai[owners[rank]]; 4910 ierr = PetscFreeSpaceGet(PetscIntMultTruncate(2,len)+1,&free_space);CHKERRQ(ierr); 4911 4912 current_space = free_space; 4913 4914 /* determine symbolic info for each local row */ 4915 ierr = PetscMalloc3(merge->nrecv,&buf_ri_k,merge->nrecv,&nextrow,merge->nrecv,&nextai);CHKERRQ(ierr); 4916 4917 for (k=0; k<merge->nrecv; k++) { 4918 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 4919 nrows = *buf_ri_k[k]; 4920 nextrow[k] = buf_ri_k[k] + 1; /* next row number of k-th recved i-structure */ 4921 nextai[k] = buf_ri_k[k] + (nrows + 1); /* poins to the next i-structure of k-th recved i-structure */ 4922 } 4923 4924 ierr = MatPreallocateInitialize(comm,m,n,dnz,onz);CHKERRQ(ierr); 4925 len = 0; 4926 for (i=0; i<m; i++) { 4927 bnzi = 0; 4928 /* add local non-zero cols of this proc's seqmat into lnk */ 4929 arow = owners[rank] + i; 4930 anzi = ai[arow+1] - ai[arow]; 4931 aj = a->j + ai[arow]; 4932 ierr = PetscLLAddSorted(anzi,aj,N,nlnk,lnk,lnkbt);CHKERRQ(ierr); 4933 bnzi += nlnk; 4934 /* add received col data into lnk */ 4935 for (k=0; k<merge->nrecv; k++) { /* k-th received message */ 4936 if (i == *nextrow[k]) { /* i-th row */ 4937 anzi = *(nextai[k]+1) - *nextai[k]; 4938 aj = buf_rj[k] + *nextai[k]; 4939 ierr = PetscLLAddSorted(anzi,aj,N,nlnk,lnk,lnkbt);CHKERRQ(ierr); 4940 bnzi += nlnk; 4941 nextrow[k]++; nextai[k]++; 4942 } 4943 } 4944 if (len < bnzi) len = bnzi; /* =max(bnzi) */ 4945 4946 /* if free space is not available, make more free space */ 4947 if (current_space->local_remaining<bnzi) { 4948 ierr = PetscFreeSpaceGet(PetscIntSumTruncate(bnzi,current_space->total_array_size),¤t_space);CHKERRQ(ierr); 4949 nspacedouble++; 4950 } 4951 /* copy data into free space, then initialize lnk */ 4952 ierr = PetscLLClean(N,N,bnzi,lnk,current_space->array,lnkbt);CHKERRQ(ierr); 4953 ierr = MatPreallocateSet(i+owners[rank],bnzi,current_space->array,dnz,onz);CHKERRQ(ierr); 4954 4955 current_space->array += bnzi; 4956 current_space->local_used += bnzi; 4957 current_space->local_remaining -= bnzi; 4958 4959 bi[i+1] = bi[i] + bnzi; 4960 } 4961 4962 ierr = PetscFree3(buf_ri_k,nextrow,nextai);CHKERRQ(ierr); 4963 4964 ierr = PetscMalloc1(bi[m]+1,&bj);CHKERRQ(ierr); 4965 ierr = PetscFreeSpaceContiguous(&free_space,bj);CHKERRQ(ierr); 4966 ierr = PetscLLDestroy(lnk,lnkbt);CHKERRQ(ierr); 4967 4968 /* create symbolic parallel matrix B_mpi */ 4969 /*---------------------------------------*/ 4970 ierr = MatGetBlockSizes(seqmat,&bs,&cbs);CHKERRQ(ierr); 4971 ierr = MatCreate(comm,&B_mpi);CHKERRQ(ierr); 4972 if (n==PETSC_DECIDE) { 4973 ierr = MatSetSizes(B_mpi,m,n,PETSC_DETERMINE,N);CHKERRQ(ierr); 4974 } else { 4975 ierr = MatSetSizes(B_mpi,m,n,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 4976 } 4977 ierr = MatSetBlockSizes(B_mpi,bs,cbs);CHKERRQ(ierr); 4978 ierr = MatSetType(B_mpi,MATMPIAIJ);CHKERRQ(ierr); 4979 ierr = MatMPIAIJSetPreallocation(B_mpi,0,dnz,0,onz);CHKERRQ(ierr); 4980 ierr = MatPreallocateFinalize(dnz,onz);CHKERRQ(ierr); 4981 ierr = MatSetOption(B_mpi,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 4982 4983 /* B_mpi is not ready for use - assembly will be done by MatCreateMPIAIJSumSeqAIJNumeric() */ 4984 B_mpi->assembled = PETSC_FALSE; 4985 B_mpi->ops->destroy = MatDestroy_MPIAIJ_SeqsToMPI; 4986 merge->bi = bi; 4987 merge->bj = bj; 4988 merge->buf_ri = buf_ri; 4989 merge->buf_rj = buf_rj; 4990 merge->coi = NULL; 4991 merge->coj = NULL; 4992 merge->owners_co = NULL; 4993 4994 ierr = PetscCommDestroy(&comm);CHKERRQ(ierr); 4995 4996 /* attach the supporting struct to B_mpi for reuse */ 4997 ierr = PetscContainerCreate(PETSC_COMM_SELF,&container);CHKERRQ(ierr); 4998 ierr = PetscContainerSetPointer(container,merge);CHKERRQ(ierr); 4999 ierr = PetscObjectCompose((PetscObject)B_mpi,"MatMergeSeqsToMPI",(PetscObject)container);CHKERRQ(ierr); 5000 ierr = PetscContainerDestroy(&container);CHKERRQ(ierr); 5001 *mpimat = B_mpi; 5002 5003 ierr = PetscLogEventEnd(MAT_Seqstompisym,seqmat,0,0,0);CHKERRQ(ierr); 5004 PetscFunctionReturn(0); 5005 } 5006 5007 /*@C 5008 MatCreateMPIAIJSumSeqAIJ - Creates a MATMPIAIJ matrix by adding sequential 5009 matrices from each processor 5010 5011 Collective 5012 5013 Input Parameters: 5014 + comm - the communicators the parallel matrix will live on 5015 . seqmat - the input sequential matrices 5016 . m - number of local rows (or PETSC_DECIDE) 5017 . n - number of local columns (or PETSC_DECIDE) 5018 - scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX 5019 5020 Output Parameter: 5021 . mpimat - the parallel matrix generated 5022 5023 Level: advanced 5024 5025 Notes: 5026 The dimensions of the sequential matrix in each processor MUST be the same. 5027 The input seqmat is included into the container "Mat_Merge_SeqsToMPI", and will be 5028 destroyed when mpimat is destroyed. Call PetscObjectQuery() to access seqmat. 5029 @*/ 5030 PetscErrorCode MatCreateMPIAIJSumSeqAIJ(MPI_Comm comm,Mat seqmat,PetscInt m,PetscInt n,MatReuse scall,Mat *mpimat) 5031 { 5032 PetscErrorCode ierr; 5033 PetscMPIInt size; 5034 5035 PetscFunctionBegin; 5036 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 5037 if (size == 1) { 5038 ierr = PetscLogEventBegin(MAT_Seqstompi,seqmat,0,0,0);CHKERRQ(ierr); 5039 if (scall == MAT_INITIAL_MATRIX) { 5040 ierr = MatDuplicate(seqmat,MAT_COPY_VALUES,mpimat);CHKERRQ(ierr); 5041 } else { 5042 ierr = MatCopy(seqmat,*mpimat,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 5043 } 5044 ierr = PetscLogEventEnd(MAT_Seqstompi,seqmat,0,0,0);CHKERRQ(ierr); 5045 PetscFunctionReturn(0); 5046 } 5047 ierr = PetscLogEventBegin(MAT_Seqstompi,seqmat,0,0,0);CHKERRQ(ierr); 5048 if (scall == MAT_INITIAL_MATRIX) { 5049 ierr = MatCreateMPIAIJSumSeqAIJSymbolic(comm,seqmat,m,n,mpimat);CHKERRQ(ierr); 5050 } 5051 ierr = MatCreateMPIAIJSumSeqAIJNumeric(seqmat,*mpimat);CHKERRQ(ierr); 5052 ierr = PetscLogEventEnd(MAT_Seqstompi,seqmat,0,0,0);CHKERRQ(ierr); 5053 PetscFunctionReturn(0); 5054 } 5055 5056 /*@ 5057 MatMPIAIJGetLocalMat - Creates a SeqAIJ from a MATMPIAIJ matrix by taking all its local rows and putting them into a sequential matrix with 5058 mlocal rows and n columns. Where mlocal is the row count obtained with MatGetLocalSize() and n is the global column count obtained 5059 with MatGetSize() 5060 5061 Not Collective 5062 5063 Input Parameters: 5064 + A - the matrix 5065 . scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX 5066 5067 Output Parameter: 5068 . A_loc - the local sequential matrix generated 5069 5070 Level: developer 5071 5072 .seealso: MatGetOwnershipRange(), MatMPIAIJGetLocalMatCondensed() 5073 5074 @*/ 5075 PetscErrorCode MatMPIAIJGetLocalMat(Mat A,MatReuse scall,Mat *A_loc) 5076 { 5077 PetscErrorCode ierr; 5078 Mat_MPIAIJ *mpimat=(Mat_MPIAIJ*)A->data; 5079 Mat_SeqAIJ *mat,*a,*b; 5080 PetscInt *ai,*aj,*bi,*bj,*cmap=mpimat->garray; 5081 MatScalar *aa,*ba,*cam; 5082 PetscScalar *ca; 5083 PetscInt am=A->rmap->n,i,j,k,cstart=A->cmap->rstart; 5084 PetscInt *ci,*cj,col,ncols_d,ncols_o,jo; 5085 PetscBool match; 5086 MPI_Comm comm; 5087 PetscMPIInt size; 5088 5089 PetscFunctionBegin; 5090 ierr = PetscStrbeginswith(((PetscObject)A)->type_name,MATMPIAIJ,&match);CHKERRQ(ierr); 5091 if (!match) SETERRQ(PetscObjectComm((PetscObject)A), PETSC_ERR_SUP,"Requires MATMPIAIJ matrix as input"); 5092 ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 5093 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 5094 if (size == 1 && scall == MAT_REUSE_MATRIX) PetscFunctionReturn(0); 5095 5096 ierr = PetscLogEventBegin(MAT_Getlocalmat,A,0,0,0);CHKERRQ(ierr); 5097 a = (Mat_SeqAIJ*)(mpimat->A)->data; 5098 b = (Mat_SeqAIJ*)(mpimat->B)->data; 5099 ai = a->i; aj = a->j; bi = b->i; bj = b->j; 5100 aa = a->a; ba = b->a; 5101 if (scall == MAT_INITIAL_MATRIX) { 5102 if (size == 1) { 5103 ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,am,A->cmap->N,ai,aj,aa,A_loc);CHKERRQ(ierr); 5104 PetscFunctionReturn(0); 5105 } 5106 5107 ierr = PetscMalloc1(1+am,&ci);CHKERRQ(ierr); 5108 ci[0] = 0; 5109 for (i=0; i<am; i++) { 5110 ci[i+1] = ci[i] + (ai[i+1] - ai[i]) + (bi[i+1] - bi[i]); 5111 } 5112 ierr = PetscMalloc1(1+ci[am],&cj);CHKERRQ(ierr); 5113 ierr = PetscMalloc1(1+ci[am],&ca);CHKERRQ(ierr); 5114 k = 0; 5115 for (i=0; i<am; i++) { 5116 ncols_o = bi[i+1] - bi[i]; 5117 ncols_d = ai[i+1] - ai[i]; 5118 /* off-diagonal portion of A */ 5119 for (jo=0; jo<ncols_o; jo++) { 5120 col = cmap[*bj]; 5121 if (col >= cstart) break; 5122 cj[k] = col; bj++; 5123 ca[k++] = *ba++; 5124 } 5125 /* diagonal portion of A */ 5126 for (j=0; j<ncols_d; j++) { 5127 cj[k] = cstart + *aj++; 5128 ca[k++] = *aa++; 5129 } 5130 /* off-diagonal portion of A */ 5131 for (j=jo; j<ncols_o; j++) { 5132 cj[k] = cmap[*bj++]; 5133 ca[k++] = *ba++; 5134 } 5135 } 5136 /* put together the new matrix */ 5137 ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,am,A->cmap->N,ci,cj,ca,A_loc);CHKERRQ(ierr); 5138 /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */ 5139 /* Since these are PETSc arrays, change flags to free them as necessary. */ 5140 mat = (Mat_SeqAIJ*)(*A_loc)->data; 5141 mat->free_a = PETSC_TRUE; 5142 mat->free_ij = PETSC_TRUE; 5143 mat->nonew = 0; 5144 } else if (scall == MAT_REUSE_MATRIX) { 5145 mat=(Mat_SeqAIJ*)(*A_loc)->data; 5146 ci = mat->i; cj = mat->j; cam = mat->a; 5147 for (i=0; i<am; i++) { 5148 /* off-diagonal portion of A */ 5149 ncols_o = bi[i+1] - bi[i]; 5150 for (jo=0; jo<ncols_o; jo++) { 5151 col = cmap[*bj]; 5152 if (col >= cstart) break; 5153 *cam++ = *ba++; bj++; 5154 } 5155 /* diagonal portion of A */ 5156 ncols_d = ai[i+1] - ai[i]; 5157 for (j=0; j<ncols_d; j++) *cam++ = *aa++; 5158 /* off-diagonal portion of A */ 5159 for (j=jo; j<ncols_o; j++) { 5160 *cam++ = *ba++; bj++; 5161 } 5162 } 5163 } else SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Invalid MatReuse %d",(int)scall); 5164 ierr = PetscLogEventEnd(MAT_Getlocalmat,A,0,0,0);CHKERRQ(ierr); 5165 PetscFunctionReturn(0); 5166 } 5167 5168 /*@C 5169 MatMPIAIJGetLocalMatCondensed - Creates a SeqAIJ matrix from an MATMPIAIJ matrix by taking all its local rows and NON-ZERO columns 5170 5171 Not Collective 5172 5173 Input Parameters: 5174 + A - the matrix 5175 . scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX 5176 - row, col - index sets of rows and columns to extract (or NULL) 5177 5178 Output Parameter: 5179 . A_loc - the local sequential matrix generated 5180 5181 Level: developer 5182 5183 .seealso: MatGetOwnershipRange(), MatMPIAIJGetLocalMat() 5184 5185 @*/ 5186 PetscErrorCode MatMPIAIJGetLocalMatCondensed(Mat A,MatReuse scall,IS *row,IS *col,Mat *A_loc) 5187 { 5188 Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data; 5189 PetscErrorCode ierr; 5190 PetscInt i,start,end,ncols,nzA,nzB,*cmap,imark,*idx; 5191 IS isrowa,iscola; 5192 Mat *aloc; 5193 PetscBool match; 5194 5195 PetscFunctionBegin; 5196 ierr = PetscObjectTypeCompare((PetscObject)A,MATMPIAIJ,&match);CHKERRQ(ierr); 5197 if (!match) SETERRQ(PetscObjectComm((PetscObject)A), PETSC_ERR_SUP,"Requires MATMPIAIJ matrix as input"); 5198 ierr = PetscLogEventBegin(MAT_Getlocalmatcondensed,A,0,0,0);CHKERRQ(ierr); 5199 if (!row) { 5200 start = A->rmap->rstart; end = A->rmap->rend; 5201 ierr = ISCreateStride(PETSC_COMM_SELF,end-start,start,1,&isrowa);CHKERRQ(ierr); 5202 } else { 5203 isrowa = *row; 5204 } 5205 if (!col) { 5206 start = A->cmap->rstart; 5207 cmap = a->garray; 5208 nzA = a->A->cmap->n; 5209 nzB = a->B->cmap->n; 5210 ierr = PetscMalloc1(nzA+nzB, &idx);CHKERRQ(ierr); 5211 ncols = 0; 5212 for (i=0; i<nzB; i++) { 5213 if (cmap[i] < start) idx[ncols++] = cmap[i]; 5214 else break; 5215 } 5216 imark = i; 5217 for (i=0; i<nzA; i++) idx[ncols++] = start + i; 5218 for (i=imark; i<nzB; i++) idx[ncols++] = cmap[i]; 5219 ierr = ISCreateGeneral(PETSC_COMM_SELF,ncols,idx,PETSC_OWN_POINTER,&iscola);CHKERRQ(ierr); 5220 } else { 5221 iscola = *col; 5222 } 5223 if (scall != MAT_INITIAL_MATRIX) { 5224 ierr = PetscMalloc1(1,&aloc);CHKERRQ(ierr); 5225 aloc[0] = *A_loc; 5226 } 5227 ierr = MatCreateSubMatrices(A,1,&isrowa,&iscola,scall,&aloc);CHKERRQ(ierr); 5228 if (!col) { /* attach global id of condensed columns */ 5229 ierr = PetscObjectCompose((PetscObject)aloc[0],"_petsc_GetLocalMatCondensed_iscol",(PetscObject)iscola);CHKERRQ(ierr); 5230 } 5231 *A_loc = aloc[0]; 5232 ierr = PetscFree(aloc);CHKERRQ(ierr); 5233 if (!row) { 5234 ierr = ISDestroy(&isrowa);CHKERRQ(ierr); 5235 } 5236 if (!col) { 5237 ierr = ISDestroy(&iscola);CHKERRQ(ierr); 5238 } 5239 ierr = PetscLogEventEnd(MAT_Getlocalmatcondensed,A,0,0,0);CHKERRQ(ierr); 5240 PetscFunctionReturn(0); 5241 } 5242 5243 /*@C 5244 MatGetBrowsOfAcols - Creates a SeqAIJ matrix by taking rows of B that equal to nonzero columns of local A 5245 5246 Collective on Mat 5247 5248 Input Parameters: 5249 + A,B - the matrices in mpiaij format 5250 . scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX 5251 - rowb, colb - index sets of rows and columns of B to extract (or NULL) 5252 5253 Output Parameter: 5254 + rowb, colb - index sets of rows and columns of B to extract 5255 - B_seq - the sequential matrix generated 5256 5257 Level: developer 5258 5259 @*/ 5260 PetscErrorCode MatGetBrowsOfAcols(Mat A,Mat B,MatReuse scall,IS *rowb,IS *colb,Mat *B_seq) 5261 { 5262 Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data; 5263 PetscErrorCode ierr; 5264 PetscInt *idx,i,start,ncols,nzA,nzB,*cmap,imark; 5265 IS isrowb,iscolb; 5266 Mat *bseq=NULL; 5267 5268 PetscFunctionBegin; 5269 if (A->cmap->rstart != B->rmap->rstart || A->cmap->rend != B->rmap->rend) { 5270 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); 5271 } 5272 ierr = PetscLogEventBegin(MAT_GetBrowsOfAcols,A,B,0,0);CHKERRQ(ierr); 5273 5274 if (scall == MAT_INITIAL_MATRIX) { 5275 start = A->cmap->rstart; 5276 cmap = a->garray; 5277 nzA = a->A->cmap->n; 5278 nzB = a->B->cmap->n; 5279 ierr = PetscMalloc1(nzA+nzB, &idx);CHKERRQ(ierr); 5280 ncols = 0; 5281 for (i=0; i<nzB; i++) { /* row < local row index */ 5282 if (cmap[i] < start) idx[ncols++] = cmap[i]; 5283 else break; 5284 } 5285 imark = i; 5286 for (i=0; i<nzA; i++) idx[ncols++] = start + i; /* local rows */ 5287 for (i=imark; i<nzB; i++) idx[ncols++] = cmap[i]; /* row > local row index */ 5288 ierr = ISCreateGeneral(PETSC_COMM_SELF,ncols,idx,PETSC_OWN_POINTER,&isrowb);CHKERRQ(ierr); 5289 ierr = ISCreateStride(PETSC_COMM_SELF,B->cmap->N,0,1,&iscolb);CHKERRQ(ierr); 5290 } else { 5291 if (!rowb || !colb) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"IS rowb and colb must be provided for MAT_REUSE_MATRIX"); 5292 isrowb = *rowb; iscolb = *colb; 5293 ierr = PetscMalloc1(1,&bseq);CHKERRQ(ierr); 5294 bseq[0] = *B_seq; 5295 } 5296 ierr = MatCreateSubMatrices(B,1,&isrowb,&iscolb,scall,&bseq);CHKERRQ(ierr); 5297 *B_seq = bseq[0]; 5298 ierr = PetscFree(bseq);CHKERRQ(ierr); 5299 if (!rowb) { 5300 ierr = ISDestroy(&isrowb);CHKERRQ(ierr); 5301 } else { 5302 *rowb = isrowb; 5303 } 5304 if (!colb) { 5305 ierr = ISDestroy(&iscolb);CHKERRQ(ierr); 5306 } else { 5307 *colb = iscolb; 5308 } 5309 ierr = PetscLogEventEnd(MAT_GetBrowsOfAcols,A,B,0,0);CHKERRQ(ierr); 5310 PetscFunctionReturn(0); 5311 } 5312 5313 /* 5314 MatGetBrowsOfAoCols_MPIAIJ - Creates a SeqAIJ matrix by taking rows of B that equal to nonzero columns 5315 of the OFF-DIAGONAL portion of local A 5316 5317 Collective on Mat 5318 5319 Input Parameters: 5320 + A,B - the matrices in mpiaij format 5321 - scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX 5322 5323 Output Parameter: 5324 + startsj_s - starting point in B's sending j-arrays, saved for MAT_REUSE (or NULL) 5325 . startsj_r - starting point in B's receiving j-arrays, saved for MAT_REUSE (or NULL) 5326 . bufa_ptr - array for sending matrix values, saved for MAT_REUSE (or NULL) 5327 - B_oth - the sequential matrix generated with size aBn=a->B->cmap->n by B->cmap->N 5328 5329 Developer Notes: This directly accesses information inside the VecScatter associated with the matrix-vector product 5330 for this matrix. This is not desirable.. 5331 5332 Level: developer 5333 5334 */ 5335 PetscErrorCode MatGetBrowsOfAoCols_MPIAIJ(Mat A,Mat B,MatReuse scall,PetscInt **startsj_s,PetscInt **startsj_r,MatScalar **bufa_ptr,Mat *B_oth) 5336 { 5337 PetscErrorCode ierr; 5338 Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data; 5339 Mat_SeqAIJ *b_oth; 5340 VecScatter ctx; 5341 MPI_Comm comm; 5342 const PetscMPIInt *rprocs,*sprocs; 5343 const PetscInt *srow,*rstarts,*sstarts; 5344 PetscInt *rowlen,*bufj,*bufJ,ncols,aBn=a->B->cmap->n,row,*b_othi,*b_othj,*rvalues=NULL,*svalues=NULL,*cols,sbs,rbs; 5345 PetscInt i,j,k=0,l,ll,nrecvs,nsends,nrows,*rstartsj = 0,*sstartsj,len; 5346 PetscScalar *b_otha,*bufa,*bufA,*vals; 5347 MPI_Request *rwaits = NULL,*swaits = NULL; 5348 MPI_Status rstatus; 5349 PetscMPIInt jj,size,tag,rank,nsends_mpi,nrecvs_mpi; 5350 5351 PetscFunctionBegin; 5352 ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 5353 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 5354 5355 if (A->cmap->rstart != B->rmap->rstart || A->cmap->rend != B->rmap->rend) { 5356 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); 5357 } 5358 ierr = PetscLogEventBegin(MAT_GetBrowsOfAocols,A,B,0,0);CHKERRQ(ierr); 5359 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 5360 5361 if (size == 1) { 5362 startsj_s = NULL; 5363 bufa_ptr = NULL; 5364 *B_oth = NULL; 5365 PetscFunctionReturn(0); 5366 } 5367 5368 ctx = a->Mvctx; 5369 tag = ((PetscObject)ctx)->tag; 5370 5371 if (ctx->inuse) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE," Scatter ctx already in use"); 5372 ierr = VecScatterGetRemote_Private(ctx,PETSC_TRUE/*send*/,&nsends,&sstarts,&srow,&sprocs,&sbs);CHKERRQ(ierr); 5373 /* rprocs[] must be ordered so that indices received from them are ordered in rvalues[], which is key to algorithms used in this subroutine */ 5374 ierr = VecScatterGetRemoteOrdered_Private(ctx,PETSC_FALSE/*recv*/,&nrecvs,&rstarts,NULL/*indices not needed*/,&rprocs,&rbs);CHKERRQ(ierr); 5375 ierr = PetscMPIIntCast(nsends,&nsends_mpi);CHKERRQ(ierr); 5376 ierr = PetscMPIIntCast(nrecvs,&nrecvs_mpi);CHKERRQ(ierr); 5377 ierr = PetscMalloc2(nrecvs,&rwaits,nsends,&swaits);CHKERRQ(ierr); 5378 5379 if (!startsj_s || !bufa_ptr) scall = MAT_INITIAL_MATRIX; 5380 if (scall == MAT_INITIAL_MATRIX) { 5381 /* i-array */ 5382 /*---------*/ 5383 /* post receives */ 5384 if (nrecvs) {ierr = PetscMalloc1(rbs*(rstarts[nrecvs] - rstarts[0]),&rvalues);CHKERRQ(ierr);} /* rstarts can be NULL when nrecvs=0 */ 5385 for (i=0; i<nrecvs; i++) { 5386 rowlen = rvalues + rstarts[i]*rbs; 5387 nrows = (rstarts[i+1]-rstarts[i])*rbs; /* num of indices to be received */ 5388 ierr = MPI_Irecv(rowlen,nrows,MPIU_INT,rprocs[i],tag,comm,rwaits+i);CHKERRQ(ierr); 5389 } 5390 5391 /* pack the outgoing message */ 5392 ierr = PetscMalloc2(nsends+1,&sstartsj,nrecvs+1,&rstartsj);CHKERRQ(ierr); 5393 5394 sstartsj[0] = 0; 5395 rstartsj[0] = 0; 5396 len = 0; /* total length of j or a array to be sent */ 5397 if (nsends) { 5398 k = sstarts[0]; /* ATTENTION: sstarts[0] and rstarts[0] are not necessarily zero */ 5399 ierr = PetscMalloc1(sbs*(sstarts[nsends]-sstarts[0]),&svalues);CHKERRQ(ierr); 5400 } 5401 for (i=0; i<nsends; i++) { 5402 rowlen = svalues + (sstarts[i]-sstarts[0])*sbs; 5403 nrows = sstarts[i+1]-sstarts[i]; /* num of block rows */ 5404 for (j=0; j<nrows; j++) { 5405 row = srow[k] + B->rmap->range[rank]; /* global row idx */ 5406 for (l=0; l<sbs; l++) { 5407 ierr = MatGetRow_MPIAIJ(B,row+l,&ncols,NULL,NULL);CHKERRQ(ierr); /* rowlength */ 5408 5409 rowlen[j*sbs+l] = ncols; 5410 5411 len += ncols; 5412 ierr = MatRestoreRow_MPIAIJ(B,row+l,&ncols,NULL,NULL);CHKERRQ(ierr); 5413 } 5414 k++; 5415 } 5416 ierr = MPI_Isend(rowlen,nrows*sbs,MPIU_INT,sprocs[i],tag,comm,swaits+i);CHKERRQ(ierr); 5417 5418 sstartsj[i+1] = len; /* starting point of (i+1)-th outgoing msg in bufj and bufa */ 5419 } 5420 /* recvs and sends of i-array are completed */ 5421 i = nrecvs; 5422 while (i--) { 5423 ierr = MPI_Waitany(nrecvs_mpi,rwaits,&jj,&rstatus);CHKERRQ(ierr); 5424 } 5425 if (nsends) {ierr = MPI_Waitall(nsends_mpi,swaits,MPI_STATUSES_IGNORE);CHKERRQ(ierr);} 5426 ierr = PetscFree(svalues);CHKERRQ(ierr); 5427 5428 /* allocate buffers for sending j and a arrays */ 5429 ierr = PetscMalloc1(len+1,&bufj);CHKERRQ(ierr); 5430 ierr = PetscMalloc1(len+1,&bufa);CHKERRQ(ierr); 5431 5432 /* create i-array of B_oth */ 5433 ierr = PetscMalloc1(aBn+2,&b_othi);CHKERRQ(ierr); 5434 5435 b_othi[0] = 0; 5436 len = 0; /* total length of j or a array to be received */ 5437 k = 0; 5438 for (i=0; i<nrecvs; i++) { 5439 rowlen = rvalues + (rstarts[i]-rstarts[0])*rbs; 5440 nrows = (rstarts[i+1]-rstarts[i])*rbs; /* num of rows to be received */ 5441 for (j=0; j<nrows; j++) { 5442 b_othi[k+1] = b_othi[k] + rowlen[j]; 5443 ierr = PetscIntSumError(rowlen[j],len,&len);CHKERRQ(ierr); 5444 k++; 5445 } 5446 rstartsj[i+1] = len; /* starting point of (i+1)-th incoming msg in bufj and bufa */ 5447 } 5448 ierr = PetscFree(rvalues);CHKERRQ(ierr); 5449 5450 /* allocate space for j and a arrrays of B_oth */ 5451 ierr = PetscMalloc1(b_othi[aBn]+1,&b_othj);CHKERRQ(ierr); 5452 ierr = PetscMalloc1(b_othi[aBn]+1,&b_otha);CHKERRQ(ierr); 5453 5454 /* j-array */ 5455 /*---------*/ 5456 /* post receives of j-array */ 5457 for (i=0; i<nrecvs; i++) { 5458 nrows = rstartsj[i+1]-rstartsj[i]; /* length of the msg received */ 5459 ierr = MPI_Irecv(b_othj+rstartsj[i],nrows,MPIU_INT,rprocs[i],tag,comm,rwaits+i);CHKERRQ(ierr); 5460 } 5461 5462 /* pack the outgoing message j-array */ 5463 if (nsends) k = sstarts[0]; 5464 for (i=0; i<nsends; i++) { 5465 nrows = sstarts[i+1]-sstarts[i]; /* num of block rows */ 5466 bufJ = bufj+sstartsj[i]; 5467 for (j=0; j<nrows; j++) { 5468 row = srow[k++] + B->rmap->range[rank]; /* global row idx */ 5469 for (ll=0; ll<sbs; ll++) { 5470 ierr = MatGetRow_MPIAIJ(B,row+ll,&ncols,&cols,NULL);CHKERRQ(ierr); 5471 for (l=0; l<ncols; l++) { 5472 *bufJ++ = cols[l]; 5473 } 5474 ierr = MatRestoreRow_MPIAIJ(B,row+ll,&ncols,&cols,NULL);CHKERRQ(ierr); 5475 } 5476 } 5477 ierr = MPI_Isend(bufj+sstartsj[i],sstartsj[i+1]-sstartsj[i],MPIU_INT,sprocs[i],tag,comm,swaits+i);CHKERRQ(ierr); 5478 } 5479 5480 /* recvs and sends of j-array are completed */ 5481 i = nrecvs; 5482 while (i--) { 5483 ierr = MPI_Waitany(nrecvs_mpi,rwaits,&jj,&rstatus);CHKERRQ(ierr); 5484 } 5485 if (nsends) {ierr = MPI_Waitall(nsends_mpi,swaits,MPI_STATUSES_IGNORE);CHKERRQ(ierr);} 5486 } else if (scall == MAT_REUSE_MATRIX) { 5487 sstartsj = *startsj_s; 5488 rstartsj = *startsj_r; 5489 bufa = *bufa_ptr; 5490 b_oth = (Mat_SeqAIJ*)(*B_oth)->data; 5491 b_otha = b_oth->a; 5492 } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE, "Matrix P does not posses an object container"); 5493 5494 /* a-array */ 5495 /*---------*/ 5496 /* post receives of a-array */ 5497 for (i=0; i<nrecvs; i++) { 5498 nrows = rstartsj[i+1]-rstartsj[i]; /* length of the msg received */ 5499 ierr = MPI_Irecv(b_otha+rstartsj[i],nrows,MPIU_SCALAR,rprocs[i],tag,comm,rwaits+i);CHKERRQ(ierr); 5500 } 5501 5502 /* pack the outgoing message a-array */ 5503 if (nsends) k = sstarts[0]; 5504 for (i=0; i<nsends; i++) { 5505 nrows = sstarts[i+1]-sstarts[i]; /* num of block rows */ 5506 bufA = bufa+sstartsj[i]; 5507 for (j=0; j<nrows; j++) { 5508 row = srow[k++] + B->rmap->range[rank]; /* global row idx */ 5509 for (ll=0; ll<sbs; ll++) { 5510 ierr = MatGetRow_MPIAIJ(B,row+ll,&ncols,NULL,&vals);CHKERRQ(ierr); 5511 for (l=0; l<ncols; l++) { 5512 *bufA++ = vals[l]; 5513 } 5514 ierr = MatRestoreRow_MPIAIJ(B,row+ll,&ncols,NULL,&vals);CHKERRQ(ierr); 5515 } 5516 } 5517 ierr = MPI_Isend(bufa+sstartsj[i],sstartsj[i+1]-sstartsj[i],MPIU_SCALAR,sprocs[i],tag,comm,swaits+i);CHKERRQ(ierr); 5518 } 5519 /* recvs and sends of a-array are completed */ 5520 i = nrecvs; 5521 while (i--) { 5522 ierr = MPI_Waitany(nrecvs_mpi,rwaits,&jj,&rstatus);CHKERRQ(ierr); 5523 } 5524 if (nsends) {ierr = MPI_Waitall(nsends_mpi,swaits,MPI_STATUSES_IGNORE);CHKERRQ(ierr);} 5525 ierr = PetscFree2(rwaits,swaits);CHKERRQ(ierr); 5526 5527 if (scall == MAT_INITIAL_MATRIX) { 5528 /* put together the new matrix */ 5529 ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,aBn,B->cmap->N,b_othi,b_othj,b_otha,B_oth);CHKERRQ(ierr); 5530 5531 /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */ 5532 /* Since these are PETSc arrays, change flags to free them as necessary. */ 5533 b_oth = (Mat_SeqAIJ*)(*B_oth)->data; 5534 b_oth->free_a = PETSC_TRUE; 5535 b_oth->free_ij = PETSC_TRUE; 5536 b_oth->nonew = 0; 5537 5538 ierr = PetscFree(bufj);CHKERRQ(ierr); 5539 if (!startsj_s || !bufa_ptr) { 5540 ierr = PetscFree2(sstartsj,rstartsj);CHKERRQ(ierr); 5541 ierr = PetscFree(bufa_ptr);CHKERRQ(ierr); 5542 } else { 5543 *startsj_s = sstartsj; 5544 *startsj_r = rstartsj; 5545 *bufa_ptr = bufa; 5546 } 5547 } 5548 5549 ierr = VecScatterRestoreRemote_Private(ctx,PETSC_TRUE,&nsends,&sstarts,&srow,&sprocs,&sbs);CHKERRQ(ierr); 5550 ierr = VecScatterRestoreRemoteOrdered_Private(ctx,PETSC_FALSE,&nrecvs,&rstarts,NULL,&rprocs,&rbs);CHKERRQ(ierr); 5551 ierr = PetscLogEventEnd(MAT_GetBrowsOfAocols,A,B,0,0);CHKERRQ(ierr); 5552 PetscFunctionReturn(0); 5553 } 5554 5555 /*@C 5556 MatGetCommunicationStructs - Provides access to the communication structures used in matrix-vector multiplication. 5557 5558 Not Collective 5559 5560 Input Parameters: 5561 . A - The matrix in mpiaij format 5562 5563 Output Parameter: 5564 + lvec - The local vector holding off-process values from the argument to a matrix-vector product 5565 . colmap - A map from global column index to local index into lvec 5566 - multScatter - A scatter from the argument of a matrix-vector product to lvec 5567 5568 Level: developer 5569 5570 @*/ 5571 #if defined(PETSC_USE_CTABLE) 5572 PetscErrorCode MatGetCommunicationStructs(Mat A, Vec *lvec, PetscTable *colmap, VecScatter *multScatter) 5573 #else 5574 PetscErrorCode MatGetCommunicationStructs(Mat A, Vec *lvec, PetscInt *colmap[], VecScatter *multScatter) 5575 #endif 5576 { 5577 Mat_MPIAIJ *a; 5578 5579 PetscFunctionBegin; 5580 PetscValidHeaderSpecific(A, MAT_CLASSID, 1); 5581 PetscValidPointer(lvec, 2); 5582 PetscValidPointer(colmap, 3); 5583 PetscValidPointer(multScatter, 4); 5584 a = (Mat_MPIAIJ*) A->data; 5585 if (lvec) *lvec = a->lvec; 5586 if (colmap) *colmap = a->colmap; 5587 if (multScatter) *multScatter = a->Mvctx; 5588 PetscFunctionReturn(0); 5589 } 5590 5591 PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJCRL(Mat,MatType,MatReuse,Mat*); 5592 PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJPERM(Mat,MatType,MatReuse,Mat*); 5593 PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJSELL(Mat,MatType,MatReuse,Mat*); 5594 #if defined(PETSC_HAVE_MKL_SPARSE) 5595 PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJMKL(Mat,MatType,MatReuse,Mat*); 5596 #endif 5597 PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPISBAIJ(Mat,MatType,MatReuse,Mat*); 5598 #if defined(PETSC_HAVE_ELEMENTAL) 5599 PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_Elemental(Mat,MatType,MatReuse,Mat*); 5600 #endif 5601 #if defined(PETSC_HAVE_HYPRE) 5602 PETSC_INTERN PetscErrorCode MatConvert_AIJ_HYPRE(Mat,MatType,MatReuse,Mat*); 5603 PETSC_INTERN PetscErrorCode MatMatMatMult_Transpose_AIJ_AIJ(Mat,Mat,Mat,MatReuse,PetscReal,Mat*); 5604 #endif 5605 PETSC_INTERN PetscErrorCode MatConvert_XAIJ_IS(Mat,MatType,MatReuse,Mat*); 5606 PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPISELL(Mat,MatType,MatReuse,Mat*); 5607 PETSC_INTERN PetscErrorCode MatPtAP_IS_XAIJ(Mat,Mat,MatReuse,PetscReal,Mat*); 5608 5609 /* 5610 Computes (B'*A')' since computing B*A directly is untenable 5611 5612 n p p 5613 ( ) ( ) ( ) 5614 m ( A ) * n ( B ) = m ( C ) 5615 ( ) ( ) ( ) 5616 5617 */ 5618 PetscErrorCode MatMatMultNumeric_MPIDense_MPIAIJ(Mat A,Mat B,Mat C) 5619 { 5620 PetscErrorCode ierr; 5621 Mat At,Bt,Ct; 5622 5623 PetscFunctionBegin; 5624 ierr = MatTranspose(A,MAT_INITIAL_MATRIX,&At);CHKERRQ(ierr); 5625 ierr = MatTranspose(B,MAT_INITIAL_MATRIX,&Bt);CHKERRQ(ierr); 5626 ierr = MatMatMult(Bt,At,MAT_INITIAL_MATRIX,1.0,&Ct);CHKERRQ(ierr); 5627 ierr = MatDestroy(&At);CHKERRQ(ierr); 5628 ierr = MatDestroy(&Bt);CHKERRQ(ierr); 5629 ierr = MatTranspose(Ct,MAT_REUSE_MATRIX,&C);CHKERRQ(ierr); 5630 ierr = MatDestroy(&Ct);CHKERRQ(ierr); 5631 PetscFunctionReturn(0); 5632 } 5633 5634 PetscErrorCode MatMatMultSymbolic_MPIDense_MPIAIJ(Mat A,Mat B,PetscReal fill,Mat *C) 5635 { 5636 PetscErrorCode ierr; 5637 PetscInt m=A->rmap->n,n=B->cmap->n; 5638 Mat Cmat; 5639 5640 PetscFunctionBegin; 5641 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); 5642 ierr = MatCreate(PetscObjectComm((PetscObject)A),&Cmat);CHKERRQ(ierr); 5643 ierr = MatSetSizes(Cmat,m,n,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 5644 ierr = MatSetBlockSizesFromMats(Cmat,A,B);CHKERRQ(ierr); 5645 ierr = MatSetType(Cmat,MATMPIDENSE);CHKERRQ(ierr); 5646 ierr = MatMPIDenseSetPreallocation(Cmat,NULL);CHKERRQ(ierr); 5647 ierr = MatAssemblyBegin(Cmat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5648 ierr = MatAssemblyEnd(Cmat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5649 5650 Cmat->ops->matmultnumeric = MatMatMultNumeric_MPIDense_MPIAIJ; 5651 5652 *C = Cmat; 5653 PetscFunctionReturn(0); 5654 } 5655 5656 /* ----------------------------------------------------------------*/ 5657 PETSC_INTERN PetscErrorCode MatMatMult_MPIDense_MPIAIJ(Mat A,Mat B,MatReuse scall,PetscReal fill,Mat *C) 5658 { 5659 PetscErrorCode ierr; 5660 5661 PetscFunctionBegin; 5662 if (scall == MAT_INITIAL_MATRIX) { 5663 ierr = PetscLogEventBegin(MAT_MatMultSymbolic,A,B,0,0);CHKERRQ(ierr); 5664 ierr = MatMatMultSymbolic_MPIDense_MPIAIJ(A,B,fill,C);CHKERRQ(ierr); 5665 ierr = PetscLogEventEnd(MAT_MatMultSymbolic,A,B,0,0);CHKERRQ(ierr); 5666 } 5667 ierr = PetscLogEventBegin(MAT_MatMultNumeric,A,B,0,0);CHKERRQ(ierr); 5668 ierr = MatMatMultNumeric_MPIDense_MPIAIJ(A,B,*C);CHKERRQ(ierr); 5669 ierr = PetscLogEventEnd(MAT_MatMultNumeric,A,B,0,0);CHKERRQ(ierr); 5670 PetscFunctionReturn(0); 5671 } 5672 5673 /*MC 5674 MATMPIAIJ - MATMPIAIJ = "mpiaij" - A matrix type to be used for parallel sparse matrices. 5675 5676 Options Database Keys: 5677 . -mat_type mpiaij - sets the matrix type to "mpiaij" during a call to MatSetFromOptions() 5678 5679 Level: beginner 5680 5681 .seealso: MatCreateAIJ() 5682 M*/ 5683 5684 PETSC_EXTERN PetscErrorCode MatCreate_MPIAIJ(Mat B) 5685 { 5686 Mat_MPIAIJ *b; 5687 PetscErrorCode ierr; 5688 PetscMPIInt size; 5689 5690 PetscFunctionBegin; 5691 ierr = MPI_Comm_size(PetscObjectComm((PetscObject)B),&size);CHKERRQ(ierr); 5692 5693 ierr = PetscNewLog(B,&b);CHKERRQ(ierr); 5694 B->data = (void*)b; 5695 ierr = PetscMemcpy(B->ops,&MatOps_Values,sizeof(struct _MatOps));CHKERRQ(ierr); 5696 B->assembled = PETSC_FALSE; 5697 B->insertmode = NOT_SET_VALUES; 5698 b->size = size; 5699 5700 ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)B),&b->rank);CHKERRQ(ierr); 5701 5702 /* build cache for off array entries formed */ 5703 ierr = MatStashCreate_Private(PetscObjectComm((PetscObject)B),1,&B->stash);CHKERRQ(ierr); 5704 5705 b->donotstash = PETSC_FALSE; 5706 b->colmap = 0; 5707 b->garray = 0; 5708 b->roworiented = PETSC_TRUE; 5709 5710 /* stuff used for matrix vector multiply */ 5711 b->lvec = NULL; 5712 b->Mvctx = NULL; 5713 5714 /* stuff for MatGetRow() */ 5715 b->rowindices = 0; 5716 b->rowvalues = 0; 5717 b->getrowactive = PETSC_FALSE; 5718 5719 /* flexible pointer used in CUSP/CUSPARSE classes */ 5720 b->spptr = NULL; 5721 5722 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMPIAIJSetUseScalableIncreaseOverlap_C",MatMPIAIJSetUseScalableIncreaseOverlap_MPIAIJ);CHKERRQ(ierr); 5723 ierr = PetscObjectComposeFunction((PetscObject)B,"MatStoreValues_C",MatStoreValues_MPIAIJ);CHKERRQ(ierr); 5724 ierr = PetscObjectComposeFunction((PetscObject)B,"MatRetrieveValues_C",MatRetrieveValues_MPIAIJ);CHKERRQ(ierr); 5725 ierr = PetscObjectComposeFunction((PetscObject)B,"MatIsTranspose_C",MatIsTranspose_MPIAIJ);CHKERRQ(ierr); 5726 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMPIAIJSetPreallocation_C",MatMPIAIJSetPreallocation_MPIAIJ);CHKERRQ(ierr); 5727 ierr = PetscObjectComposeFunction((PetscObject)B,"MatResetPreallocation_C",MatResetPreallocation_MPIAIJ);CHKERRQ(ierr); 5728 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMPIAIJSetPreallocationCSR_C",MatMPIAIJSetPreallocationCSR_MPIAIJ);CHKERRQ(ierr); 5729 ierr = PetscObjectComposeFunction((PetscObject)B,"MatDiagonalScaleLocal_C",MatDiagonalScaleLocal_MPIAIJ);CHKERRQ(ierr); 5730 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_mpiaijperm_C",MatConvert_MPIAIJ_MPIAIJPERM);CHKERRQ(ierr); 5731 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_mpiaijsell_C",MatConvert_MPIAIJ_MPIAIJSELL);CHKERRQ(ierr); 5732 #if defined(PETSC_HAVE_MKL_SPARSE) 5733 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_mpiaijmkl_C",MatConvert_MPIAIJ_MPIAIJMKL);CHKERRQ(ierr); 5734 #endif 5735 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_mpiaijcrl_C",MatConvert_MPIAIJ_MPIAIJCRL);CHKERRQ(ierr); 5736 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_mpisbaij_C",MatConvert_MPIAIJ_MPISBAIJ);CHKERRQ(ierr); 5737 #if defined(PETSC_HAVE_ELEMENTAL) 5738 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_elemental_C",MatConvert_MPIAIJ_Elemental);CHKERRQ(ierr); 5739 #endif 5740 #if defined(PETSC_HAVE_HYPRE) 5741 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_hypre_C",MatConvert_AIJ_HYPRE);CHKERRQ(ierr); 5742 #endif 5743 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_is_C",MatConvert_XAIJ_IS);CHKERRQ(ierr); 5744 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_mpisell_C",MatConvert_MPIAIJ_MPISELL);CHKERRQ(ierr); 5745 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMatMult_mpidense_mpiaij_C",MatMatMult_MPIDense_MPIAIJ);CHKERRQ(ierr); 5746 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMatMultSymbolic_mpidense_mpiaij_C",MatMatMultSymbolic_MPIDense_MPIAIJ);CHKERRQ(ierr); 5747 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMatMultNumeric_mpidense_mpiaij_C",MatMatMultNumeric_MPIDense_MPIAIJ);CHKERRQ(ierr); 5748 #if defined(PETSC_HAVE_HYPRE) 5749 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMatMatMult_transpose_mpiaij_mpiaij_C",MatMatMatMult_Transpose_AIJ_AIJ);CHKERRQ(ierr); 5750 #endif 5751 ierr = PetscObjectComposeFunction((PetscObject)B,"MatPtAP_is_mpiaij_C",MatPtAP_IS_XAIJ);CHKERRQ(ierr); 5752 ierr = PetscObjectChangeTypeName((PetscObject)B,MATMPIAIJ);CHKERRQ(ierr); 5753 PetscFunctionReturn(0); 5754 } 5755 5756 /*@C 5757 MatCreateMPIAIJWithSplitArrays - creates a MPI AIJ matrix using arrays that contain the "diagonal" 5758 and "off-diagonal" part of the matrix in CSR format. 5759 5760 Collective 5761 5762 Input Parameters: 5763 + comm - MPI communicator 5764 . m - number of local rows (Cannot be PETSC_DECIDE) 5765 . n - This value should be the same as the local size used in creating the 5766 x vector for the matrix-vector product y = Ax. (or PETSC_DECIDE to have 5767 calculated if N is given) For square matrices n is almost always m. 5768 . M - number of global rows (or PETSC_DETERMINE to have calculated if m is given) 5769 . N - number of global columns (or PETSC_DETERMINE to have calculated if n is given) 5770 . i - row indices for "diagonal" portion of matrix; that is i[0] = 0, i[row] = i[row-1] + number of elements in that row of the matrix 5771 . j - column indices 5772 . a - matrix values 5773 . oi - row indices for "off-diagonal" portion of matrix; that is oi[0] = 0, oi[row] = oi[row-1] + number of elements in that row of the matrix 5774 . oj - column indices 5775 - oa - matrix values 5776 5777 Output Parameter: 5778 . mat - the matrix 5779 5780 Level: advanced 5781 5782 Notes: 5783 The i, j, and a arrays ARE NOT copied by this routine into the internal format used by PETSc. The user 5784 must free the arrays once the matrix has been destroyed and not before. 5785 5786 The i and j indices are 0 based 5787 5788 See MatCreateAIJ() for the definition of "diagonal" and "off-diagonal" portion of the matrix 5789 5790 This sets local rows and cannot be used to set off-processor values. 5791 5792 Use of this routine is discouraged because it is inflexible and cumbersome to use. It is extremely rare that a 5793 legacy application natively assembles into exactly this split format. The code to do so is nontrivial and does 5794 not easily support in-place reassembly. It is recommended to use MatSetValues() (or a variant thereof) because 5795 the resulting assembly is easier to implement, will work with any matrix format, and the user does not have to 5796 keep track of the underlying array. Use MatSetOption(A,MAT_NO_OFF_PROC_ENTRIES,PETSC_TRUE) to disable all 5797 communication if it is known that only local entries will be set. 5798 5799 .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatMPIAIJSetPreallocationCSR(), 5800 MATMPIAIJ, MatCreateAIJ(), MatCreateMPIAIJWithArrays() 5801 @*/ 5802 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) 5803 { 5804 PetscErrorCode ierr; 5805 Mat_MPIAIJ *maij; 5806 5807 PetscFunctionBegin; 5808 if (m < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"local number of rows (m) cannot be PETSC_DECIDE, or negative"); 5809 if (i[0]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"i (row indices) must start with 0"); 5810 if (oi[0]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"oi (row indices) must start with 0"); 5811 ierr = MatCreate(comm,mat);CHKERRQ(ierr); 5812 ierr = MatSetSizes(*mat,m,n,M,N);CHKERRQ(ierr); 5813 ierr = MatSetType(*mat,MATMPIAIJ);CHKERRQ(ierr); 5814 maij = (Mat_MPIAIJ*) (*mat)->data; 5815 5816 (*mat)->preallocated = PETSC_TRUE; 5817 5818 ierr = PetscLayoutSetUp((*mat)->rmap);CHKERRQ(ierr); 5819 ierr = PetscLayoutSetUp((*mat)->cmap);CHKERRQ(ierr); 5820 5821 ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,m,n,i,j,a,&maij->A);CHKERRQ(ierr); 5822 ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,m,(*mat)->cmap->N,oi,oj,oa,&maij->B);CHKERRQ(ierr); 5823 5824 ierr = MatAssemblyBegin(maij->A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5825 ierr = MatAssemblyEnd(maij->A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5826 ierr = MatAssemblyBegin(maij->B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5827 ierr = MatAssemblyEnd(maij->B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5828 5829 ierr = MatSetOption(*mat,MAT_NO_OFF_PROC_ENTRIES,PETSC_TRUE);CHKERRQ(ierr); 5830 ierr = MatAssemblyBegin(*mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5831 ierr = MatAssemblyEnd(*mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5832 ierr = MatSetOption(*mat,MAT_NO_OFF_PROC_ENTRIES,PETSC_FALSE);CHKERRQ(ierr); 5833 ierr = MatSetOption(*mat,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE);CHKERRQ(ierr); 5834 PetscFunctionReturn(0); 5835 } 5836 5837 /* 5838 Special version for direct calls from Fortran 5839 */ 5840 #include <petsc/private/fortranimpl.h> 5841 5842 /* Change these macros so can be used in void function */ 5843 #undef CHKERRQ 5844 #define CHKERRQ(ierr) CHKERRABORT(PETSC_COMM_WORLD,ierr) 5845 #undef SETERRQ2 5846 #define SETERRQ2(comm,ierr,b,c,d) CHKERRABORT(comm,ierr) 5847 #undef SETERRQ3 5848 #define SETERRQ3(comm,ierr,b,c,d,e) CHKERRABORT(comm,ierr) 5849 #undef SETERRQ 5850 #define SETERRQ(c,ierr,b) CHKERRABORT(c,ierr) 5851 5852 #if defined(PETSC_HAVE_FORTRAN_CAPS) 5853 #define matsetvaluesmpiaij_ MATSETVALUESMPIAIJ 5854 #elif !defined(PETSC_HAVE_FORTRAN_UNDERSCORE) 5855 #define matsetvaluesmpiaij_ matsetvaluesmpiaij 5856 #else 5857 #endif 5858 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) 5859 { 5860 Mat mat = *mmat; 5861 PetscInt m = *mm, n = *mn; 5862 InsertMode addv = *maddv; 5863 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 5864 PetscScalar value; 5865 PetscErrorCode ierr; 5866 5867 MatCheckPreallocated(mat,1); 5868 if (mat->insertmode == NOT_SET_VALUES) mat->insertmode = addv; 5869 5870 #if defined(PETSC_USE_DEBUG) 5871 else if (mat->insertmode != addv) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Cannot mix add values and insert values"); 5872 #endif 5873 { 5874 PetscInt i,j,rstart = mat->rmap->rstart,rend = mat->rmap->rend; 5875 PetscInt cstart = mat->cmap->rstart,cend = mat->cmap->rend,row,col; 5876 PetscBool roworiented = aij->roworiented; 5877 5878 /* Some Variables required in the macro */ 5879 Mat A = aij->A; 5880 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 5881 PetscInt *aimax = a->imax,*ai = a->i,*ailen = a->ilen,*aj = a->j; 5882 MatScalar *aa = a->a; 5883 PetscBool ignorezeroentries = (((a->ignorezeroentries)&&(addv==ADD_VALUES)) ? PETSC_TRUE : PETSC_FALSE); 5884 Mat B = aij->B; 5885 Mat_SeqAIJ *b = (Mat_SeqAIJ*)B->data; 5886 PetscInt *bimax = b->imax,*bi = b->i,*bilen = b->ilen,*bj = b->j,bm = aij->B->rmap->n,am = aij->A->rmap->n; 5887 MatScalar *ba = b->a; 5888 5889 PetscInt *rp1,*rp2,ii,nrow1,nrow2,_i,rmax1,rmax2,N,low1,high1,low2,high2,t,lastcol1,lastcol2; 5890 PetscInt nonew = a->nonew; 5891 MatScalar *ap1,*ap2; 5892 5893 PetscFunctionBegin; 5894 for (i=0; i<m; i++) { 5895 if (im[i] < 0) continue; 5896 #if defined(PETSC_USE_DEBUG) 5897 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); 5898 #endif 5899 if (im[i] >= rstart && im[i] < rend) { 5900 row = im[i] - rstart; 5901 lastcol1 = -1; 5902 rp1 = aj + ai[row]; 5903 ap1 = aa + ai[row]; 5904 rmax1 = aimax[row]; 5905 nrow1 = ailen[row]; 5906 low1 = 0; 5907 high1 = nrow1; 5908 lastcol2 = -1; 5909 rp2 = bj + bi[row]; 5910 ap2 = ba + bi[row]; 5911 rmax2 = bimax[row]; 5912 nrow2 = bilen[row]; 5913 low2 = 0; 5914 high2 = nrow2; 5915 5916 for (j=0; j<n; j++) { 5917 if (roworiented) value = v[i*n+j]; 5918 else value = v[i+j*m]; 5919 if (in[j] >= cstart && in[j] < cend) { 5920 col = in[j] - cstart; 5921 if (ignorezeroentries && value == 0.0 && (addv == ADD_VALUES) && row != col) continue; 5922 MatSetValues_SeqAIJ_A_Private(row,col,value,addv,im[i],in[j]); 5923 } else if (in[j] < 0) continue; 5924 #if defined(PETSC_USE_DEBUG) 5925 /* extra brace on SETERRQ2() is required for --with-errorchecking=0 - due to the next 'else' clause */ 5926 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);} 5927 #endif 5928 else { 5929 if (mat->was_assembled) { 5930 if (!aij->colmap) { 5931 ierr = MatCreateColmap_MPIAIJ_Private(mat);CHKERRQ(ierr); 5932 } 5933 #if defined(PETSC_USE_CTABLE) 5934 ierr = PetscTableFind(aij->colmap,in[j]+1,&col);CHKERRQ(ierr); 5935 col--; 5936 #else 5937 col = aij->colmap[in[j]] - 1; 5938 #endif 5939 if (ignorezeroentries && value == 0.0 && (addv == ADD_VALUES) && row != col) continue; 5940 if (col < 0 && !((Mat_SeqAIJ*)(aij->A->data))->nonew) { 5941 ierr = MatDisAssemble_MPIAIJ(mat);CHKERRQ(ierr); 5942 col = in[j]; 5943 /* Reinitialize the variables required by MatSetValues_SeqAIJ_B_Private() */ 5944 B = aij->B; 5945 b = (Mat_SeqAIJ*)B->data; 5946 bimax = b->imax; bi = b->i; bilen = b->ilen; bj = b->j; 5947 rp2 = bj + bi[row]; 5948 ap2 = ba + bi[row]; 5949 rmax2 = bimax[row]; 5950 nrow2 = bilen[row]; 5951 low2 = 0; 5952 high2 = nrow2; 5953 bm = aij->B->rmap->n; 5954 ba = b->a; 5955 } 5956 } else col = in[j]; 5957 MatSetValues_SeqAIJ_B_Private(row,col,value,addv,im[i],in[j]); 5958 } 5959 } 5960 } else if (!aij->donotstash) { 5961 if (roworiented) { 5962 ierr = MatStashValuesRow_Private(&mat->stash,im[i],n,in,v+i*n,(PetscBool)(ignorezeroentries && (addv == ADD_VALUES)));CHKERRQ(ierr); 5963 } else { 5964 ierr = MatStashValuesCol_Private(&mat->stash,im[i],n,in,v+i,m,(PetscBool)(ignorezeroentries && (addv == ADD_VALUES)));CHKERRQ(ierr); 5965 } 5966 } 5967 } 5968 } 5969 PetscFunctionReturnVoid(); 5970 } 5971