1 2 #include <../src/mat/impls/aij/mpi/mpiaij.h> /*I "petscmat.h" I*/ 3 #include <petsc-private/vecimpl.h> 4 #include <petscblaslapack.h> 5 #include <petscsf.h> 6 7 /*MC 8 MATAIJ - MATAIJ = "aij" - A matrix type to be used for sparse matrices. 9 10 This matrix type is identical to MATSEQAIJ when constructed with a single process communicator, 11 and MATMPIAIJ otherwise. As a result, for single process communicators, 12 MatSeqAIJSetPreallocation is supported, and similarly MatMPIAIJSetPreallocation is supported 13 for communicators controlling multiple processes. It is recommended that you call both of 14 the above preallocation routines for simplicity. 15 16 Options Database Keys: 17 . -mat_type aij - sets the matrix type to "aij" during a call to MatSetFromOptions() 18 19 Developer Notes: Subclasses include MATAIJCUSP, MATAIJCUSPARSE, MATAIJPERM, MATAIJCRL, and also automatically switches over to use inodes when 20 enough exist. 21 22 Level: beginner 23 24 .seealso: MatCreateAIJ(), MatCreateSeqAIJ(), MATSEQAIJ,MATMPIAIJ 25 M*/ 26 27 /*MC 28 MATAIJCRL - MATAIJCRL = "aijcrl" - A matrix type to be used for sparse matrices. 29 30 This matrix type is identical to MATSEQAIJCRL when constructed with a single process communicator, 31 and MATMPIAIJCRL otherwise. As a result, for single process communicators, 32 MatSeqAIJSetPreallocation() is supported, and similarly MatMPIAIJSetPreallocation() is supported 33 for communicators controlling multiple processes. It is recommended that you call both of 34 the above preallocation routines for simplicity. 35 36 Options Database Keys: 37 . -mat_type aijcrl - sets the matrix type to "aijcrl" during a call to MatSetFromOptions() 38 39 Level: beginner 40 41 .seealso: MatCreateMPIAIJCRL,MATSEQAIJCRL,MATMPIAIJCRL, MATSEQAIJCRL, MATMPIAIJCRL 42 M*/ 43 44 #undef __FUNCT__ 45 #define __FUNCT__ "MatFindNonzeroRows_MPIAIJ" 46 PetscErrorCode MatFindNonzeroRows_MPIAIJ(Mat M,IS *keptrows) 47 { 48 PetscErrorCode ierr; 49 Mat_MPIAIJ *mat = (Mat_MPIAIJ*)M->data; 50 Mat_SeqAIJ *a = (Mat_SeqAIJ*)mat->A->data; 51 Mat_SeqAIJ *b = (Mat_SeqAIJ*)mat->B->data; 52 const PetscInt *ia,*ib; 53 const MatScalar *aa,*bb; 54 PetscInt na,nb,i,j,*rows,cnt=0,n0rows; 55 PetscInt m = M->rmap->n,rstart = M->rmap->rstart; 56 57 PetscFunctionBegin; 58 *keptrows = 0; 59 ia = a->i; 60 ib = b->i; 61 for (i=0; i<m; i++) { 62 na = ia[i+1] - ia[i]; 63 nb = ib[i+1] - ib[i]; 64 if (!na && !nb) { 65 cnt++; 66 goto ok1; 67 } 68 aa = a->a + ia[i]; 69 for (j=0; j<na; j++) { 70 if (aa[j] != 0.0) goto ok1; 71 } 72 bb = b->a + ib[i]; 73 for (j=0; j <nb; j++) { 74 if (bb[j] != 0.0) goto ok1; 75 } 76 cnt++; 77 ok1:; 78 } 79 ierr = MPI_Allreduce(&cnt,&n0rows,1,MPIU_INT,MPIU_SUM,PetscObjectComm((PetscObject)M));CHKERRQ(ierr); 80 if (!n0rows) PetscFunctionReturn(0); 81 ierr = PetscMalloc1((M->rmap->n-cnt),&rows);CHKERRQ(ierr); 82 cnt = 0; 83 for (i=0; i<m; i++) { 84 na = ia[i+1] - ia[i]; 85 nb = ib[i+1] - ib[i]; 86 if (!na && !nb) continue; 87 aa = a->a + ia[i]; 88 for (j=0; j<na;j++) { 89 if (aa[j] != 0.0) { 90 rows[cnt++] = rstart + i; 91 goto ok2; 92 } 93 } 94 bb = b->a + ib[i]; 95 for (j=0; j<nb; j++) { 96 if (bb[j] != 0.0) { 97 rows[cnt++] = rstart + i; 98 goto ok2; 99 } 100 } 101 ok2:; 102 } 103 ierr = ISCreateGeneral(PetscObjectComm((PetscObject)M),cnt,rows,PETSC_OWN_POINTER,keptrows);CHKERRQ(ierr); 104 PetscFunctionReturn(0); 105 } 106 107 #undef __FUNCT__ 108 #define __FUNCT__ "MatDiagonalSet_MPIAIJ" 109 PetscErrorCode MatDiagonalSet_MPIAIJ(Mat Y,Vec D,InsertMode is) 110 { 111 PetscErrorCode ierr; 112 Mat_MPIAIJ *aij = (Mat_MPIAIJ*) Y->data; 113 114 PetscFunctionBegin; 115 if (Y->assembled && Y->rmap->rstart == Y->cmap->rstart && Y->rmap->rend == Y->cmap->rend) { 116 ierr = MatDiagonalSet(aij->A,D,is);CHKERRQ(ierr); 117 } else { 118 ierr = MatDiagonalSet_Default(Y,D,is);CHKERRQ(ierr); 119 } 120 PetscFunctionReturn(0); 121 } 122 123 124 #undef __FUNCT__ 125 #define __FUNCT__ "MatFindZeroDiagonals_MPIAIJ" 126 PetscErrorCode MatFindZeroDiagonals_MPIAIJ(Mat M,IS *zrows) 127 { 128 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)M->data; 129 PetscErrorCode ierr; 130 PetscInt i,rstart,nrows,*rows; 131 132 PetscFunctionBegin; 133 *zrows = NULL; 134 ierr = MatFindZeroDiagonals_SeqAIJ_Private(aij->A,&nrows,&rows);CHKERRQ(ierr); 135 ierr = MatGetOwnershipRange(M,&rstart,NULL);CHKERRQ(ierr); 136 for (i=0; i<nrows; i++) rows[i] += rstart; 137 ierr = ISCreateGeneral(PetscObjectComm((PetscObject)M),nrows,rows,PETSC_OWN_POINTER,zrows);CHKERRQ(ierr); 138 PetscFunctionReturn(0); 139 } 140 141 #undef __FUNCT__ 142 #define __FUNCT__ "MatGetColumnNorms_MPIAIJ" 143 PetscErrorCode MatGetColumnNorms_MPIAIJ(Mat A,NormType type,PetscReal *norms) 144 { 145 PetscErrorCode ierr; 146 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)A->data; 147 PetscInt i,n,*garray = aij->garray; 148 Mat_SeqAIJ *a_aij = (Mat_SeqAIJ*) aij->A->data; 149 Mat_SeqAIJ *b_aij = (Mat_SeqAIJ*) aij->B->data; 150 PetscReal *work; 151 152 PetscFunctionBegin; 153 ierr = MatGetSize(A,NULL,&n);CHKERRQ(ierr); 154 ierr = PetscCalloc1(n,&work);CHKERRQ(ierr); 155 if (type == NORM_2) { 156 for (i=0; i<a_aij->i[aij->A->rmap->n]; i++) { 157 work[A->cmap->rstart + a_aij->j[i]] += PetscAbsScalar(a_aij->a[i]*a_aij->a[i]); 158 } 159 for (i=0; i<b_aij->i[aij->B->rmap->n]; i++) { 160 work[garray[b_aij->j[i]]] += PetscAbsScalar(b_aij->a[i]*b_aij->a[i]); 161 } 162 } else if (type == NORM_1) { 163 for (i=0; i<a_aij->i[aij->A->rmap->n]; i++) { 164 work[A->cmap->rstart + a_aij->j[i]] += PetscAbsScalar(a_aij->a[i]); 165 } 166 for (i=0; i<b_aij->i[aij->B->rmap->n]; i++) { 167 work[garray[b_aij->j[i]]] += PetscAbsScalar(b_aij->a[i]); 168 } 169 } else if (type == NORM_INFINITY) { 170 for (i=0; i<a_aij->i[aij->A->rmap->n]; i++) { 171 work[A->cmap->rstart + a_aij->j[i]] = PetscMax(PetscAbsScalar(a_aij->a[i]), work[A->cmap->rstart + a_aij->j[i]]); 172 } 173 for (i=0; i<b_aij->i[aij->B->rmap->n]; i++) { 174 work[garray[b_aij->j[i]]] = PetscMax(PetscAbsScalar(b_aij->a[i]),work[garray[b_aij->j[i]]]); 175 } 176 177 } else SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_WRONG,"Unknown NormType"); 178 if (type == NORM_INFINITY) { 179 ierr = MPI_Allreduce(work,norms,n,MPIU_REAL,MPIU_MAX,PetscObjectComm((PetscObject)A));CHKERRQ(ierr); 180 } else { 181 ierr = MPI_Allreduce(work,norms,n,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)A));CHKERRQ(ierr); 182 } 183 ierr = PetscFree(work);CHKERRQ(ierr); 184 if (type == NORM_2) { 185 for (i=0; i<n; i++) norms[i] = PetscSqrtReal(norms[i]); 186 } 187 PetscFunctionReturn(0); 188 } 189 190 #undef __FUNCT__ 191 #define __FUNCT__ "MatDistribute_MPIAIJ" 192 /* 193 Distributes a SeqAIJ matrix across a set of processes. Code stolen from 194 MatLoad_MPIAIJ(). Horrible lack of reuse. Should be a routine for each matrix type. 195 196 Only for square matrices 197 198 Used by a preconditioner, hence PETSC_EXTERN 199 */ 200 PETSC_EXTERN PetscErrorCode MatDistribute_MPIAIJ(MPI_Comm comm,Mat gmat,PetscInt m,MatReuse reuse,Mat *inmat) 201 { 202 PetscMPIInt rank,size; 203 PetscInt *rowners,*dlens,*olens,i,rstart,rend,j,jj,nz = 0,*gmataj,cnt,row,*ld,bses[2]; 204 PetscErrorCode ierr; 205 Mat mat; 206 Mat_SeqAIJ *gmata; 207 PetscMPIInt tag; 208 MPI_Status status; 209 PetscBool aij; 210 MatScalar *gmataa,*ao,*ad,*gmataarestore=0; 211 212 PetscFunctionBegin; 213 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 214 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 215 if (!rank) { 216 ierr = PetscObjectTypeCompare((PetscObject)gmat,MATSEQAIJ,&aij);CHKERRQ(ierr); 217 if (!aij) SETERRQ1(PetscObjectComm((PetscObject)gmat),PETSC_ERR_SUP,"Currently no support for input matrix of type %s\n",((PetscObject)gmat)->type_name); 218 } 219 if (reuse == MAT_INITIAL_MATRIX) { 220 ierr = MatCreate(comm,&mat);CHKERRQ(ierr); 221 ierr = MatSetSizes(mat,m,m,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 222 ierr = MatGetBlockSizes(gmat,&bses[0],&bses[1]);CHKERRQ(ierr); 223 ierr = MPI_Bcast(bses,2,MPIU_INT,0,comm);CHKERRQ(ierr); 224 ierr = MatSetBlockSizes(mat,bses[0],bses[1]);CHKERRQ(ierr); 225 ierr = MatSetType(mat,MATAIJ);CHKERRQ(ierr); 226 ierr = PetscMalloc1((size+1),&rowners);CHKERRQ(ierr); 227 ierr = PetscMalloc2(m,&dlens,m,&olens);CHKERRQ(ierr); 228 ierr = MPI_Allgather(&m,1,MPIU_INT,rowners+1,1,MPIU_INT,comm);CHKERRQ(ierr); 229 230 rowners[0] = 0; 231 for (i=2; i<=size; i++) rowners[i] += rowners[i-1]; 232 rstart = rowners[rank]; 233 rend = rowners[rank+1]; 234 ierr = PetscObjectGetNewTag((PetscObject)mat,&tag);CHKERRQ(ierr); 235 if (!rank) { 236 gmata = (Mat_SeqAIJ*) gmat->data; 237 /* send row lengths to all processors */ 238 for (i=0; i<m; i++) dlens[i] = gmata->ilen[i]; 239 for (i=1; i<size; i++) { 240 ierr = MPI_Send(gmata->ilen + rowners[i],rowners[i+1]-rowners[i],MPIU_INT,i,tag,comm);CHKERRQ(ierr); 241 } 242 /* determine number diagonal and off-diagonal counts */ 243 ierr = PetscMemzero(olens,m*sizeof(PetscInt));CHKERRQ(ierr); 244 ierr = PetscCalloc1(m,&ld);CHKERRQ(ierr); 245 jj = 0; 246 for (i=0; i<m; i++) { 247 for (j=0; j<dlens[i]; j++) { 248 if (gmata->j[jj] < rstart) ld[i]++; 249 if (gmata->j[jj] < rstart || gmata->j[jj] >= rend) olens[i]++; 250 jj++; 251 } 252 } 253 /* send column indices to other processes */ 254 for (i=1; i<size; i++) { 255 nz = gmata->i[rowners[i+1]]-gmata->i[rowners[i]]; 256 ierr = MPI_Send(&nz,1,MPIU_INT,i,tag,comm);CHKERRQ(ierr); 257 ierr = MPI_Send(gmata->j + gmata->i[rowners[i]],nz,MPIU_INT,i,tag,comm);CHKERRQ(ierr); 258 } 259 260 /* send numerical values to other processes */ 261 for (i=1; i<size; i++) { 262 nz = gmata->i[rowners[i+1]]-gmata->i[rowners[i]]; 263 ierr = MPI_Send(gmata->a + gmata->i[rowners[i]],nz,MPIU_SCALAR,i,tag,comm);CHKERRQ(ierr); 264 } 265 gmataa = gmata->a; 266 gmataj = gmata->j; 267 268 } else { 269 /* receive row lengths */ 270 ierr = MPI_Recv(dlens,m,MPIU_INT,0,tag,comm,&status);CHKERRQ(ierr); 271 /* receive column indices */ 272 ierr = MPI_Recv(&nz,1,MPIU_INT,0,tag,comm,&status);CHKERRQ(ierr); 273 ierr = PetscMalloc2(nz,&gmataa,nz,&gmataj);CHKERRQ(ierr); 274 ierr = MPI_Recv(gmataj,nz,MPIU_INT,0,tag,comm,&status);CHKERRQ(ierr); 275 /* determine number diagonal and off-diagonal counts */ 276 ierr = PetscMemzero(olens,m*sizeof(PetscInt));CHKERRQ(ierr); 277 ierr = PetscCalloc1(m,&ld);CHKERRQ(ierr); 278 jj = 0; 279 for (i=0; i<m; i++) { 280 for (j=0; j<dlens[i]; j++) { 281 if (gmataj[jj] < rstart) ld[i]++; 282 if (gmataj[jj] < rstart || gmataj[jj] >= rend) olens[i]++; 283 jj++; 284 } 285 } 286 /* receive numerical values */ 287 ierr = PetscMemzero(gmataa,nz*sizeof(PetscScalar));CHKERRQ(ierr); 288 ierr = MPI_Recv(gmataa,nz,MPIU_SCALAR,0,tag,comm,&status);CHKERRQ(ierr); 289 } 290 /* set preallocation */ 291 for (i=0; i<m; i++) { 292 dlens[i] -= olens[i]; 293 } 294 ierr = MatSeqAIJSetPreallocation(mat,0,dlens);CHKERRQ(ierr); 295 ierr = MatMPIAIJSetPreallocation(mat,0,dlens,0,olens);CHKERRQ(ierr); 296 297 for (i=0; i<m; i++) { 298 dlens[i] += olens[i]; 299 } 300 cnt = 0; 301 for (i=0; i<m; i++) { 302 row = rstart + i; 303 ierr = MatSetValues(mat,1,&row,dlens[i],gmataj+cnt,gmataa+cnt,INSERT_VALUES);CHKERRQ(ierr); 304 cnt += dlens[i]; 305 } 306 if (rank) { 307 ierr = PetscFree2(gmataa,gmataj);CHKERRQ(ierr); 308 } 309 ierr = PetscFree2(dlens,olens);CHKERRQ(ierr); 310 ierr = PetscFree(rowners);CHKERRQ(ierr); 311 312 ((Mat_MPIAIJ*)(mat->data))->ld = ld; 313 314 *inmat = mat; 315 } else { /* column indices are already set; only need to move over numerical values from process 0 */ 316 Mat_SeqAIJ *Ad = (Mat_SeqAIJ*)((Mat_MPIAIJ*)((*inmat)->data))->A->data; 317 Mat_SeqAIJ *Ao = (Mat_SeqAIJ*)((Mat_MPIAIJ*)((*inmat)->data))->B->data; 318 mat = *inmat; 319 ierr = PetscObjectGetNewTag((PetscObject)mat,&tag);CHKERRQ(ierr); 320 if (!rank) { 321 /* send numerical values to other processes */ 322 gmata = (Mat_SeqAIJ*) gmat->data; 323 ierr = MatGetOwnershipRanges(mat,(const PetscInt**)&rowners);CHKERRQ(ierr); 324 gmataa = gmata->a; 325 for (i=1; i<size; i++) { 326 nz = gmata->i[rowners[i+1]]-gmata->i[rowners[i]]; 327 ierr = MPI_Send(gmataa + gmata->i[rowners[i]],nz,MPIU_SCALAR,i,tag,comm);CHKERRQ(ierr); 328 } 329 nz = gmata->i[rowners[1]]-gmata->i[rowners[0]]; 330 } else { 331 /* receive numerical values from process 0*/ 332 nz = Ad->nz + Ao->nz; 333 ierr = PetscMalloc1(nz,&gmataa);CHKERRQ(ierr); gmataarestore = gmataa; 334 ierr = MPI_Recv(gmataa,nz,MPIU_SCALAR,0,tag,comm,&status);CHKERRQ(ierr); 335 } 336 /* transfer numerical values into the diagonal A and off diagonal B parts of mat */ 337 ld = ((Mat_MPIAIJ*)(mat->data))->ld; 338 ad = Ad->a; 339 ao = Ao->a; 340 if (mat->rmap->n) { 341 i = 0; 342 nz = ld[i]; ierr = PetscMemcpy(ao,gmataa,nz*sizeof(PetscScalar));CHKERRQ(ierr); ao += nz; gmataa += nz; 343 nz = Ad->i[i+1] - Ad->i[i]; ierr = PetscMemcpy(ad,gmataa,nz*sizeof(PetscScalar));CHKERRQ(ierr); ad += nz; gmataa += nz; 344 } 345 for (i=1; i<mat->rmap->n; i++) { 346 nz = Ao->i[i] - Ao->i[i-1] - ld[i-1] + ld[i]; ierr = PetscMemcpy(ao,gmataa,nz*sizeof(PetscScalar));CHKERRQ(ierr); ao += nz; gmataa += nz; 347 nz = Ad->i[i+1] - Ad->i[i]; ierr = PetscMemcpy(ad,gmataa,nz*sizeof(PetscScalar));CHKERRQ(ierr); ad += nz; gmataa += nz; 348 } 349 i--; 350 if (mat->rmap->n) { 351 nz = Ao->i[i+1] - Ao->i[i] - ld[i]; ierr = PetscMemcpy(ao,gmataa,nz*sizeof(PetscScalar));CHKERRQ(ierr); 352 } 353 if (rank) { 354 ierr = PetscFree(gmataarestore);CHKERRQ(ierr); 355 } 356 } 357 ierr = MatAssemblyBegin(mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 358 ierr = MatAssemblyEnd(mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 359 PetscFunctionReturn(0); 360 } 361 362 /* 363 Local utility routine that creates a mapping from the global column 364 number to the local number in the off-diagonal part of the local 365 storage of the matrix. When PETSC_USE_CTABLE is used this is scalable at 366 a slightly higher hash table cost; without it it is not scalable (each processor 367 has an order N integer array but is fast to acess. 368 */ 369 #undef __FUNCT__ 370 #define __FUNCT__ "MatCreateColmap_MPIAIJ_Private" 371 PetscErrorCode MatCreateColmap_MPIAIJ_Private(Mat mat) 372 { 373 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 374 PetscErrorCode ierr; 375 PetscInt n = aij->B->cmap->n,i; 376 377 PetscFunctionBegin; 378 if (!aij->garray) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"MPIAIJ Matrix was assembled but is missing garray"); 379 #if defined(PETSC_USE_CTABLE) 380 ierr = PetscTableCreate(n,mat->cmap->N+1,&aij->colmap);CHKERRQ(ierr); 381 for (i=0; i<n; i++) { 382 ierr = PetscTableAdd(aij->colmap,aij->garray[i]+1,i+1,INSERT_VALUES);CHKERRQ(ierr); 383 } 384 #else 385 ierr = PetscCalloc1((mat->cmap->N+1),&aij->colmap);CHKERRQ(ierr); 386 ierr = PetscLogObjectMemory((PetscObject)mat,(mat->cmap->N+1)*sizeof(PetscInt));CHKERRQ(ierr); 387 for (i=0; i<n; i++) aij->colmap[aij->garray[i]] = i+1; 388 #endif 389 PetscFunctionReturn(0); 390 } 391 392 #define MatSetValues_SeqAIJ_A_Private(row,col,value,addv) \ 393 { \ 394 if (col <= lastcol1) low1 = 0; \ 395 else high1 = nrow1; \ 396 lastcol1 = col;\ 397 while (high1-low1 > 5) { \ 398 t = (low1+high1)/2; \ 399 if (rp1[t] > col) high1 = t; \ 400 else low1 = t; \ 401 } \ 402 for (_i=low1; _i<high1; _i++) { \ 403 if (rp1[_i] > col) break; \ 404 if (rp1[_i] == col) { \ 405 if (addv == ADD_VALUES) ap1[_i] += value; \ 406 else ap1[_i] = value; \ 407 goto a_noinsert; \ 408 } \ 409 } \ 410 if (value == 0.0 && ignorezeroentries) {low1 = 0; high1 = nrow1;goto a_noinsert;} \ 411 if (nonew == 1) {low1 = 0; high1 = nrow1; goto a_noinsert;} \ 412 if (nonew == -1) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero (%D, %D) into matrix", row, col); \ 413 MatSeqXAIJReallocateAIJ(A,am,1,nrow1,row,col,rmax1,aa,ai,aj,rp1,ap1,aimax,nonew,MatScalar); \ 414 N = nrow1++ - 1; a->nz++; high1++; \ 415 /* shift up all the later entries in this row */ \ 416 for (ii=N; ii>=_i; ii--) { \ 417 rp1[ii+1] = rp1[ii]; \ 418 ap1[ii+1] = ap1[ii]; \ 419 } \ 420 rp1[_i] = col; \ 421 ap1[_i] = value; \ 422 A->nonzerostate++;\ 423 a_noinsert: ; \ 424 ailen[row] = nrow1; \ 425 } 426 427 428 #define MatSetValues_SeqAIJ_B_Private(row,col,value,addv) \ 429 { \ 430 if (col <= lastcol2) low2 = 0; \ 431 else high2 = nrow2; \ 432 lastcol2 = col; \ 433 while (high2-low2 > 5) { \ 434 t = (low2+high2)/2; \ 435 if (rp2[t] > col) high2 = t; \ 436 else low2 = t; \ 437 } \ 438 for (_i=low2; _i<high2; _i++) { \ 439 if (rp2[_i] > col) break; \ 440 if (rp2[_i] == col) { \ 441 if (addv == ADD_VALUES) ap2[_i] += value; \ 442 else ap2[_i] = value; \ 443 goto b_noinsert; \ 444 } \ 445 } \ 446 if (value == 0.0 && ignorezeroentries) {low2 = 0; high2 = nrow2; goto b_noinsert;} \ 447 if (nonew == 1) {low2 = 0; high2 = nrow2; goto b_noinsert;} \ 448 if (nonew == -1) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero (%D, %D) into matrix", row, col); \ 449 MatSeqXAIJReallocateAIJ(B,bm,1,nrow2,row,col,rmax2,ba,bi,bj,rp2,ap2,bimax,nonew,MatScalar); \ 450 N = nrow2++ - 1; b->nz++; high2++; \ 451 /* shift up all the later entries in this row */ \ 452 for (ii=N; ii>=_i; ii--) { \ 453 rp2[ii+1] = rp2[ii]; \ 454 ap2[ii+1] = ap2[ii]; \ 455 } \ 456 rp2[_i] = col; \ 457 ap2[_i] = value; \ 458 B->nonzerostate++; \ 459 b_noinsert: ; \ 460 bilen[row] = nrow2; \ 461 } 462 463 #undef __FUNCT__ 464 #define __FUNCT__ "MatSetValuesRow_MPIAIJ" 465 PetscErrorCode MatSetValuesRow_MPIAIJ(Mat A,PetscInt row,const PetscScalar v[]) 466 { 467 Mat_MPIAIJ *mat = (Mat_MPIAIJ*)A->data; 468 Mat_SeqAIJ *a = (Mat_SeqAIJ*)mat->A->data,*b = (Mat_SeqAIJ*)mat->B->data; 469 PetscErrorCode ierr; 470 PetscInt l,*garray = mat->garray,diag; 471 472 PetscFunctionBegin; 473 /* code only works for square matrices A */ 474 475 /* find size of row to the left of the diagonal part */ 476 ierr = MatGetOwnershipRange(A,&diag,0);CHKERRQ(ierr); 477 row = row - diag; 478 for (l=0; l<b->i[row+1]-b->i[row]; l++) { 479 if (garray[b->j[b->i[row]+l]] > diag) break; 480 } 481 ierr = PetscMemcpy(b->a+b->i[row],v,l*sizeof(PetscScalar));CHKERRQ(ierr); 482 483 /* diagonal part */ 484 ierr = PetscMemcpy(a->a+a->i[row],v+l,(a->i[row+1]-a->i[row])*sizeof(PetscScalar));CHKERRQ(ierr); 485 486 /* right of diagonal part */ 487 ierr = PetscMemcpy(b->a+b->i[row]+l,v+l+a->i[row+1]-a->i[row],(b->i[row+1]-b->i[row]-l)*sizeof(PetscScalar));CHKERRQ(ierr); 488 PetscFunctionReturn(0); 489 } 490 491 #undef __FUNCT__ 492 #define __FUNCT__ "MatSetValues_MPIAIJ" 493 PetscErrorCode MatSetValues_MPIAIJ(Mat mat,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],const PetscScalar v[],InsertMode addv) 494 { 495 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 496 PetscScalar value; 497 PetscErrorCode ierr; 498 PetscInt i,j,rstart = mat->rmap->rstart,rend = mat->rmap->rend; 499 PetscInt cstart = mat->cmap->rstart,cend = mat->cmap->rend,row,col; 500 PetscBool roworiented = aij->roworiented; 501 502 /* Some Variables required in the macro */ 503 Mat A = aij->A; 504 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 505 PetscInt *aimax = a->imax,*ai = a->i,*ailen = a->ilen,*aj = a->j; 506 MatScalar *aa = a->a; 507 PetscBool ignorezeroentries = a->ignorezeroentries; 508 Mat B = aij->B; 509 Mat_SeqAIJ *b = (Mat_SeqAIJ*)B->data; 510 PetscInt *bimax = b->imax,*bi = b->i,*bilen = b->ilen,*bj = b->j,bm = aij->B->rmap->n,am = aij->A->rmap->n; 511 MatScalar *ba = b->a; 512 513 PetscInt *rp1,*rp2,ii,nrow1,nrow2,_i,rmax1,rmax2,N,low1,high1,low2,high2,t,lastcol1,lastcol2; 514 PetscInt nonew; 515 MatScalar *ap1,*ap2; 516 517 PetscFunctionBegin; 518 for (i=0; i<m; i++) { 519 if (im[i] < 0) continue; 520 #if defined(PETSC_USE_DEBUG) 521 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); 522 #endif 523 if (im[i] >= rstart && im[i] < rend) { 524 row = im[i] - rstart; 525 lastcol1 = -1; 526 rp1 = aj + ai[row]; 527 ap1 = aa + ai[row]; 528 rmax1 = aimax[row]; 529 nrow1 = ailen[row]; 530 low1 = 0; 531 high1 = nrow1; 532 lastcol2 = -1; 533 rp2 = bj + bi[row]; 534 ap2 = ba + bi[row]; 535 rmax2 = bimax[row]; 536 nrow2 = bilen[row]; 537 low2 = 0; 538 high2 = nrow2; 539 540 for (j=0; j<n; j++) { 541 if (roworiented) value = v[i*n+j]; 542 else value = v[i+j*m]; 543 if (ignorezeroentries && value == 0.0 && (addv == ADD_VALUES)) continue; 544 if (in[j] >= cstart && in[j] < cend) { 545 col = in[j] - cstart; 546 nonew = a->nonew; 547 MatSetValues_SeqAIJ_A_Private(row,col,value,addv); 548 } else if (in[j] < 0) continue; 549 #if defined(PETSC_USE_DEBUG) 550 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); 551 #endif 552 else { 553 if (mat->was_assembled) { 554 if (!aij->colmap) { 555 ierr = MatCreateColmap_MPIAIJ_Private(mat);CHKERRQ(ierr); 556 } 557 #if defined(PETSC_USE_CTABLE) 558 ierr = PetscTableFind(aij->colmap,in[j]+1,&col);CHKERRQ(ierr); 559 col--; 560 #else 561 col = aij->colmap[in[j]] - 1; 562 #endif 563 if (col < 0 && !((Mat_SeqAIJ*)(aij->B->data))->nonew) { 564 ierr = MatDisAssemble_MPIAIJ(mat);CHKERRQ(ierr); 565 col = in[j]; 566 /* Reinitialize the variables required by MatSetValues_SeqAIJ_B_Private() */ 567 B = aij->B; 568 b = (Mat_SeqAIJ*)B->data; 569 bimax = b->imax; bi = b->i; bilen = b->ilen; bj = b->j; ba = b->a; 570 rp2 = bj + bi[row]; 571 ap2 = ba + bi[row]; 572 rmax2 = bimax[row]; 573 nrow2 = bilen[row]; 574 low2 = 0; 575 high2 = nrow2; 576 bm = aij->B->rmap->n; 577 ba = b->a; 578 } else if (col < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero (%D, %D) into matrix", im[i], in[j]); 579 } else col = in[j]; 580 nonew = b->nonew; 581 MatSetValues_SeqAIJ_B_Private(row,col,value,addv); 582 } 583 } 584 } else { 585 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]); 586 if (!aij->donotstash) { 587 mat->assembled = PETSC_FALSE; 588 if (roworiented) { 589 ierr = MatStashValuesRow_Private(&mat->stash,im[i],n,in,v+i*n,(PetscBool)(ignorezeroentries && (addv == ADD_VALUES)));CHKERRQ(ierr); 590 } else { 591 ierr = MatStashValuesCol_Private(&mat->stash,im[i],n,in,v+i,m,(PetscBool)(ignorezeroentries && (addv == ADD_VALUES)));CHKERRQ(ierr); 592 } 593 } 594 } 595 } 596 PetscFunctionReturn(0); 597 } 598 599 #undef __FUNCT__ 600 #define __FUNCT__ "MatGetValues_MPIAIJ" 601 PetscErrorCode MatGetValues_MPIAIJ(Mat mat,PetscInt m,const PetscInt idxm[],PetscInt n,const PetscInt idxn[],PetscScalar v[]) 602 { 603 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 604 PetscErrorCode ierr; 605 PetscInt i,j,rstart = mat->rmap->rstart,rend = mat->rmap->rend; 606 PetscInt cstart = mat->cmap->rstart,cend = mat->cmap->rend,row,col; 607 608 PetscFunctionBegin; 609 for (i=0; i<m; i++) { 610 if (idxm[i] < 0) continue; /* SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative row: %D",idxm[i]);*/ 611 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); 612 if (idxm[i] >= rstart && idxm[i] < rend) { 613 row = idxm[i] - rstart; 614 for (j=0; j<n; j++) { 615 if (idxn[j] < 0) continue; /* SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative column: %D",idxn[j]); */ 616 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); 617 if (idxn[j] >= cstart && idxn[j] < cend) { 618 col = idxn[j] - cstart; 619 ierr = MatGetValues(aij->A,1,&row,1,&col,v+i*n+j);CHKERRQ(ierr); 620 } else { 621 if (!aij->colmap) { 622 ierr = MatCreateColmap_MPIAIJ_Private(mat);CHKERRQ(ierr); 623 } 624 #if defined(PETSC_USE_CTABLE) 625 ierr = PetscTableFind(aij->colmap,idxn[j]+1,&col);CHKERRQ(ierr); 626 col--; 627 #else 628 col = aij->colmap[idxn[j]] - 1; 629 #endif 630 if ((col < 0) || (aij->garray[col] != idxn[j])) *(v+i*n+j) = 0.0; 631 else { 632 ierr = MatGetValues(aij->B,1,&row,1,&col,v+i*n+j);CHKERRQ(ierr); 633 } 634 } 635 } 636 } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Only local values currently supported"); 637 } 638 PetscFunctionReturn(0); 639 } 640 641 extern PetscErrorCode MatMultDiagonalBlock_MPIAIJ(Mat,Vec,Vec); 642 643 #undef __FUNCT__ 644 #define __FUNCT__ "MatAssemblyBegin_MPIAIJ" 645 PetscErrorCode MatAssemblyBegin_MPIAIJ(Mat mat,MatAssemblyType mode) 646 { 647 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 648 PetscErrorCode ierr; 649 PetscInt nstash,reallocs; 650 InsertMode addv; 651 652 PetscFunctionBegin; 653 if (aij->donotstash || mat->nooffprocentries) PetscFunctionReturn(0); 654 655 /* make sure all processors are either in INSERTMODE or ADDMODE */ 656 ierr = MPI_Allreduce((PetscEnum*)&mat->insertmode,(PetscEnum*)&addv,1,MPIU_ENUM,MPI_BOR,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 657 if (addv == (ADD_VALUES|INSERT_VALUES)) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_ARG_WRONGSTATE,"Some processors inserted others added"); 658 mat->insertmode = addv; /* in case this processor had no cache */ 659 660 ierr = MatStashScatterBegin_Private(mat,&mat->stash,mat->rmap->range);CHKERRQ(ierr); 661 ierr = MatStashGetInfo_Private(&mat->stash,&nstash,&reallocs);CHKERRQ(ierr); 662 ierr = PetscInfo2(aij->A,"Stash has %D entries, uses %D mallocs.\n",nstash,reallocs);CHKERRQ(ierr); 663 PetscFunctionReturn(0); 664 } 665 666 #undef __FUNCT__ 667 #define __FUNCT__ "MatAssemblyEnd_MPIAIJ" 668 PetscErrorCode MatAssemblyEnd_MPIAIJ(Mat mat,MatAssemblyType mode) 669 { 670 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 671 Mat_SeqAIJ *a = (Mat_SeqAIJ*)aij->A->data; 672 PetscErrorCode ierr; 673 PetscMPIInt n; 674 PetscInt i,j,rstart,ncols,flg; 675 PetscInt *row,*col; 676 PetscBool other_disassembled; 677 PetscScalar *val; 678 InsertMode addv = mat->insertmode; 679 680 /* do not use 'b = (Mat_SeqAIJ*)aij->B->data' as B can be reset in disassembly */ 681 682 PetscFunctionBegin; 683 if (!aij->donotstash && !mat->nooffprocentries) { 684 while (1) { 685 ierr = MatStashScatterGetMesg_Private(&mat->stash,&n,&row,&col,&val,&flg);CHKERRQ(ierr); 686 if (!flg) break; 687 688 for (i=0; i<n; ) { 689 /* Now identify the consecutive vals belonging to the same row */ 690 for (j=i,rstart=row[j]; j<n; j++) { 691 if (row[j] != rstart) break; 692 } 693 if (j < n) ncols = j-i; 694 else ncols = n-i; 695 /* Now assemble all these values with a single function call */ 696 ierr = MatSetValues_MPIAIJ(mat,1,row+i,ncols,col+i,val+i,addv);CHKERRQ(ierr); 697 698 i = j; 699 } 700 } 701 ierr = MatStashScatterEnd_Private(&mat->stash);CHKERRQ(ierr); 702 } 703 ierr = MatAssemblyBegin(aij->A,mode);CHKERRQ(ierr); 704 ierr = MatAssemblyEnd(aij->A,mode);CHKERRQ(ierr); 705 706 /* determine if any processor has disassembled, if so we must 707 also disassemble ourselfs, in order that we may reassemble. */ 708 /* 709 if nonzero structure of submatrix B cannot change then we know that 710 no processor disassembled thus we can skip this stuff 711 */ 712 if (!((Mat_SeqAIJ*)aij->B->data)->nonew) { 713 ierr = MPI_Allreduce(&mat->was_assembled,&other_disassembled,1,MPIU_BOOL,MPI_PROD,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 714 if (mat->was_assembled && !other_disassembled) { 715 ierr = MatDisAssemble_MPIAIJ(mat);CHKERRQ(ierr); 716 } 717 } 718 if (!mat->was_assembled && mode == MAT_FINAL_ASSEMBLY) { 719 ierr = MatSetUpMultiply_MPIAIJ(mat);CHKERRQ(ierr); 720 } 721 ierr = MatSetOption(aij->B,MAT_USE_INODES,PETSC_FALSE);CHKERRQ(ierr); 722 ierr = MatAssemblyBegin(aij->B,mode);CHKERRQ(ierr); 723 ierr = MatAssemblyEnd(aij->B,mode);CHKERRQ(ierr); 724 725 ierr = PetscFree2(aij->rowvalues,aij->rowindices);CHKERRQ(ierr); 726 727 aij->rowvalues = 0; 728 729 /* used by MatAXPY() */ 730 a->xtoy = 0; ((Mat_SeqAIJ*)aij->B->data)->xtoy = 0; /* b->xtoy = 0 */ 731 a->XtoY = 0; ((Mat_SeqAIJ*)aij->B->data)->XtoY = 0; /* b->XtoY = 0 */ 732 733 ierr = VecDestroy(&aij->diag);CHKERRQ(ierr); 734 if (a->inode.size) mat->ops->multdiagonalblock = MatMultDiagonalBlock_MPIAIJ; 735 736 /* if no new nonzero locations are allowed in matrix then only set the matrix state the first time through */ 737 if ((!mat->was_assembled && mode == MAT_FINAL_ASSEMBLY) || !((Mat_SeqAIJ*)(aij->A->data))->nonew) { 738 PetscObjectState state = aij->A->nonzerostate + aij->B->nonzerostate; 739 ierr = MPI_Allreduce(&state,&mat->nonzerostate,1,MPIU_INT64,MPI_SUM,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 740 } 741 PetscFunctionReturn(0); 742 } 743 744 #undef __FUNCT__ 745 #define __FUNCT__ "MatZeroEntries_MPIAIJ" 746 PetscErrorCode MatZeroEntries_MPIAIJ(Mat A) 747 { 748 Mat_MPIAIJ *l = (Mat_MPIAIJ*)A->data; 749 PetscErrorCode ierr; 750 751 PetscFunctionBegin; 752 ierr = MatZeroEntries(l->A);CHKERRQ(ierr); 753 ierr = MatZeroEntries(l->B);CHKERRQ(ierr); 754 PetscFunctionReturn(0); 755 } 756 757 #undef __FUNCT__ 758 #define __FUNCT__ "MatZeroRows_MPIAIJ" 759 PetscErrorCode MatZeroRows_MPIAIJ(Mat A,PetscInt N,const PetscInt rows[],PetscScalar diag,Vec x,Vec b) 760 { 761 Mat_MPIAIJ *mat = (Mat_MPIAIJ *) A->data; 762 PetscInt *owners = A->rmap->range; 763 PetscInt n = A->rmap->n; 764 PetscSF sf; 765 PetscInt *lrows; 766 PetscSFNode *rrows; 767 PetscInt r, p = 0, len = 0; 768 PetscErrorCode ierr; 769 770 PetscFunctionBegin; 771 /* Create SF where leaves are input rows and roots are owned rows */ 772 ierr = PetscMalloc1(n, &lrows);CHKERRQ(ierr); 773 for (r = 0; r < n; ++r) lrows[r] = -1; 774 if (!A->nooffproczerorows) {ierr = PetscMalloc1(N, &rrows);CHKERRQ(ierr);} 775 for (r = 0; r < N; ++r) { 776 const PetscInt idx = rows[r]; 777 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); 778 if (idx < owners[p] || owners[p+1] <= idx) { /* short-circuit the search if the last p owns this row too */ 779 ierr = PetscLayoutFindOwner(A->rmap,idx,&p);CHKERRQ(ierr); 780 } 781 if (A->nooffproczerorows) { 782 if (p != mat->rank) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"MAT_NO_OFF_PROC_ZERO_ROWS set, but row %D is not owned by rank %d",idx,mat->rank); 783 lrows[len++] = idx - owners[p]; 784 } else { 785 rrows[r].rank = p; 786 rrows[r].index = rows[r] - owners[p]; 787 } 788 } 789 if (!A->nooffproczerorows) { 790 ierr = PetscSFCreate(PetscObjectComm((PetscObject) A), &sf);CHKERRQ(ierr); 791 ierr = PetscSFSetGraph(sf, n, N, NULL, PETSC_OWN_POINTER, rrows, PETSC_OWN_POINTER);CHKERRQ(ierr); 792 /* Collect flags for rows to be zeroed */ 793 ierr = PetscSFReduceBegin(sf, MPIU_INT, (PetscInt*)rows, lrows, MPI_LOR);CHKERRQ(ierr); 794 ierr = PetscSFReduceEnd(sf, MPIU_INT, (PetscInt*)rows, lrows, MPI_LOR);CHKERRQ(ierr); 795 ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 796 /* Compress and put in row numbers */ 797 for (r = 0; r < n; ++r) if (lrows[r] >= 0) lrows[len++] = r; 798 } 799 /* fix right hand side if needed */ 800 if (x && b) { 801 const PetscScalar *xx; 802 PetscScalar *bb; 803 804 ierr = VecGetArrayRead(x, &xx);CHKERRQ(ierr); 805 ierr = VecGetArray(b, &bb);CHKERRQ(ierr); 806 for (r = 0; r < len; ++r) bb[lrows[r]] = diag*xx[lrows[r]]; 807 ierr = VecRestoreArrayRead(x, &xx);CHKERRQ(ierr); 808 ierr = VecRestoreArray(b, &bb);CHKERRQ(ierr); 809 } 810 /* Must zero l->B before l->A because the (diag) case below may put values into l->B*/ 811 ierr = MatZeroRows(mat->B, len, lrows, 0.0, NULL, NULL);CHKERRQ(ierr); 812 if ((diag != 0.0) && (mat->A->rmap->N == mat->A->cmap->N)) { 813 ierr = MatZeroRows(mat->A, len, lrows, diag, NULL, NULL);CHKERRQ(ierr); 814 } else if (diag != 0.0) { 815 ierr = MatZeroRows(mat->A, len, lrows, 0.0, NULL, NULL);CHKERRQ(ierr); 816 if (((Mat_SeqAIJ *) mat->A->data)->nonew) SETERRQ(PETSC_COMM_SELF, PETSC_ERR_SUP, "MatZeroRows() on rectangular matrices cannot be used with the Mat options\nMAT_NEW_NONZERO_LOCATIONS,MAT_NEW_NONZERO_LOCATION_ERR,MAT_NEW_NONZERO_ALLOCATION_ERR"); 817 for (r = 0; r < len; ++r) { 818 const PetscInt row = lrows[r] + A->rmap->rstart; 819 ierr = MatSetValues(A, 1, &row, 1, &row, &diag, INSERT_VALUES);CHKERRQ(ierr); 820 } 821 ierr = MatAssemblyBegin(A, MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 822 ierr = MatAssemblyEnd(A, MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 823 } else { 824 ierr = MatZeroRows(mat->A, len, lrows, 0.0, NULL, NULL);CHKERRQ(ierr); 825 } 826 ierr = PetscFree(lrows);CHKERRQ(ierr); 827 828 /* only change matrix nonzero state if pattern was allowed to be changed */ 829 if (!((Mat_SeqAIJ*)(mat->A->data))->keepnonzeropattern) { 830 PetscObjectState state = mat->A->nonzerostate + mat->B->nonzerostate; 831 ierr = MPI_Allreduce(&state,&A->nonzerostate,1,MPIU_INT64,MPI_SUM,PetscObjectComm((PetscObject)A));CHKERRQ(ierr); 832 } 833 PetscFunctionReturn(0); 834 } 835 836 #undef __FUNCT__ 837 #define __FUNCT__ "MatZeroRowsColumns_MPIAIJ" 838 PetscErrorCode MatZeroRowsColumns_MPIAIJ(Mat A,PetscInt N,const PetscInt rows[],PetscScalar diag,Vec x,Vec b) 839 { 840 Mat_MPIAIJ *l = (Mat_MPIAIJ*)A->data; 841 PetscErrorCode ierr; 842 PetscMPIInt n = A->rmap->n; 843 PetscInt i,j,r,m,p = 0,len = 0; 844 PetscInt *lrows,*owners = A->rmap->range; 845 PetscSFNode *rrows; 846 PetscSF sf; 847 const PetscScalar *xx; 848 PetscScalar *bb,*mask; 849 Vec xmask,lmask; 850 Mat_SeqAIJ *aij = (Mat_SeqAIJ*)l->B->data; 851 const PetscInt *aj, *ii,*ridx; 852 PetscScalar *aa; 853 854 PetscFunctionBegin; 855 /* Create SF where leaves are input rows and roots are owned rows */ 856 ierr = PetscMalloc1(n, &lrows);CHKERRQ(ierr); 857 for (r = 0; r < n; ++r) lrows[r] = -1; 858 ierr = PetscMalloc1(N, &rrows);CHKERRQ(ierr); 859 for (r = 0; r < N; ++r) { 860 const PetscInt idx = rows[r]; 861 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); 862 if (idx < owners[p] || owners[p+1] <= idx) { /* short-circuit the search if the last p owns this row too */ 863 ierr = PetscLayoutFindOwner(A->rmap,idx,&p);CHKERRQ(ierr); 864 } 865 rrows[r].rank = p; 866 rrows[r].index = rows[r] - owners[p]; 867 } 868 ierr = PetscSFCreate(PetscObjectComm((PetscObject) A), &sf);CHKERRQ(ierr); 869 ierr = PetscSFSetGraph(sf, n, N, NULL, PETSC_OWN_POINTER, rrows, PETSC_OWN_POINTER);CHKERRQ(ierr); 870 /* Collect flags for rows to be zeroed */ 871 ierr = PetscSFReduceBegin(sf, MPIU_INT, (PetscInt *) rows, lrows, MPI_LOR);CHKERRQ(ierr); 872 ierr = PetscSFReduceEnd(sf, MPIU_INT, (PetscInt *) rows, lrows, MPI_LOR);CHKERRQ(ierr); 873 ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 874 /* Compress and put in row numbers */ 875 for (r = 0; r < n; ++r) if (lrows[r] >= 0) lrows[len++] = r; 876 /* zero diagonal part of matrix */ 877 ierr = MatZeroRowsColumns(l->A,len,lrows,diag,x,b);CHKERRQ(ierr); 878 /* handle off diagonal part of matrix */ 879 ierr = MatCreateVecs(A,&xmask,NULL);CHKERRQ(ierr); 880 ierr = VecDuplicate(l->lvec,&lmask);CHKERRQ(ierr); 881 ierr = VecGetArray(xmask,&bb);CHKERRQ(ierr); 882 for (i=0; i<len; i++) bb[lrows[i]] = 1; 883 ierr = VecRestoreArray(xmask,&bb);CHKERRQ(ierr); 884 ierr = VecScatterBegin(l->Mvctx,xmask,lmask,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 885 ierr = VecScatterEnd(l->Mvctx,xmask,lmask,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 886 ierr = VecDestroy(&xmask);CHKERRQ(ierr); 887 if (x) { 888 ierr = VecScatterBegin(l->Mvctx,x,l->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 889 ierr = VecScatterEnd(l->Mvctx,x,l->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 890 ierr = VecGetArrayRead(l->lvec,&xx);CHKERRQ(ierr); 891 ierr = VecGetArray(b,&bb);CHKERRQ(ierr); 892 } 893 ierr = VecGetArray(lmask,&mask);CHKERRQ(ierr); 894 /* remove zeroed rows of off diagonal matrix */ 895 ii = aij->i; 896 for (i=0; i<len; i++) { 897 ierr = PetscMemzero(aij->a + ii[lrows[i]],(ii[lrows[i]+1] - ii[lrows[i]])*sizeof(PetscScalar));CHKERRQ(ierr); 898 } 899 /* loop over all elements of off process part of matrix zeroing removed columns*/ 900 if (aij->compressedrow.use) { 901 m = aij->compressedrow.nrows; 902 ii = aij->compressedrow.i; 903 ridx = aij->compressedrow.rindex; 904 for (i=0; i<m; i++) { 905 n = ii[i+1] - ii[i]; 906 aj = aij->j + ii[i]; 907 aa = aij->a + ii[i]; 908 909 for (j=0; j<n; j++) { 910 if (PetscAbsScalar(mask[*aj])) { 911 if (b) bb[*ridx] -= *aa*xx[*aj]; 912 *aa = 0.0; 913 } 914 aa++; 915 aj++; 916 } 917 ridx++; 918 } 919 } else { /* do not use compressed row format */ 920 m = l->B->rmap->n; 921 for (i=0; i<m; i++) { 922 n = ii[i+1] - ii[i]; 923 aj = aij->j + ii[i]; 924 aa = aij->a + ii[i]; 925 for (j=0; j<n; j++) { 926 if (PetscAbsScalar(mask[*aj])) { 927 if (b) bb[i] -= *aa*xx[*aj]; 928 *aa = 0.0; 929 } 930 aa++; 931 aj++; 932 } 933 } 934 } 935 if (x) { 936 ierr = VecRestoreArray(b,&bb);CHKERRQ(ierr); 937 ierr = VecRestoreArrayRead(l->lvec,&xx);CHKERRQ(ierr); 938 } 939 ierr = VecRestoreArray(lmask,&mask);CHKERRQ(ierr); 940 ierr = VecDestroy(&lmask);CHKERRQ(ierr); 941 ierr = PetscFree(lrows);CHKERRQ(ierr); 942 943 /* only change matrix nonzero state if pattern was allowed to be changed */ 944 if (!((Mat_SeqAIJ*)(l->A->data))->keepnonzeropattern) { 945 PetscObjectState state = l->A->nonzerostate + l->B->nonzerostate; 946 ierr = MPI_Allreduce(&state,&A->nonzerostate,1,MPIU_INT64,MPI_SUM,PetscObjectComm((PetscObject)A));CHKERRQ(ierr); 947 } 948 PetscFunctionReturn(0); 949 } 950 951 #undef __FUNCT__ 952 #define __FUNCT__ "MatMult_MPIAIJ" 953 PetscErrorCode MatMult_MPIAIJ(Mat A,Vec xx,Vec yy) 954 { 955 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 956 PetscErrorCode ierr; 957 PetscInt nt; 958 959 PetscFunctionBegin; 960 ierr = VecGetLocalSize(xx,&nt);CHKERRQ(ierr); 961 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); 962 ierr = VecScatterBegin(a->Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 963 ierr = (*a->A->ops->mult)(a->A,xx,yy);CHKERRQ(ierr); 964 ierr = VecScatterEnd(a->Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 965 ierr = (*a->B->ops->multadd)(a->B,a->lvec,yy,yy);CHKERRQ(ierr); 966 PetscFunctionReturn(0); 967 } 968 969 #undef __FUNCT__ 970 #define __FUNCT__ "MatMultDiagonalBlock_MPIAIJ" 971 PetscErrorCode MatMultDiagonalBlock_MPIAIJ(Mat A,Vec bb,Vec xx) 972 { 973 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 974 PetscErrorCode ierr; 975 976 PetscFunctionBegin; 977 ierr = MatMultDiagonalBlock(a->A,bb,xx);CHKERRQ(ierr); 978 PetscFunctionReturn(0); 979 } 980 981 #undef __FUNCT__ 982 #define __FUNCT__ "MatMultAdd_MPIAIJ" 983 PetscErrorCode MatMultAdd_MPIAIJ(Mat A,Vec xx,Vec yy,Vec zz) 984 { 985 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 986 PetscErrorCode ierr; 987 988 PetscFunctionBegin; 989 ierr = VecScatterBegin(a->Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 990 ierr = (*a->A->ops->multadd)(a->A,xx,yy,zz);CHKERRQ(ierr); 991 ierr = VecScatterEnd(a->Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 992 ierr = (*a->B->ops->multadd)(a->B,a->lvec,zz,zz);CHKERRQ(ierr); 993 PetscFunctionReturn(0); 994 } 995 996 #undef __FUNCT__ 997 #define __FUNCT__ "MatMultTranspose_MPIAIJ" 998 PetscErrorCode MatMultTranspose_MPIAIJ(Mat A,Vec xx,Vec yy) 999 { 1000 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 1001 PetscErrorCode ierr; 1002 PetscBool merged; 1003 1004 PetscFunctionBegin; 1005 ierr = VecScatterGetMerged(a->Mvctx,&merged);CHKERRQ(ierr); 1006 /* do nondiagonal part */ 1007 ierr = (*a->B->ops->multtranspose)(a->B,xx,a->lvec);CHKERRQ(ierr); 1008 if (!merged) { 1009 /* send it on its way */ 1010 ierr = VecScatterBegin(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1011 /* do local part */ 1012 ierr = (*a->A->ops->multtranspose)(a->A,xx,yy);CHKERRQ(ierr); 1013 /* receive remote parts: note this assumes the values are not actually */ 1014 /* added in yy until the next line, */ 1015 ierr = VecScatterEnd(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1016 } else { 1017 /* do local part */ 1018 ierr = (*a->A->ops->multtranspose)(a->A,xx,yy);CHKERRQ(ierr); 1019 /* send it on its way */ 1020 ierr = VecScatterBegin(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1021 /* values actually were received in the Begin() but we need to call this nop */ 1022 ierr = VecScatterEnd(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1023 } 1024 PetscFunctionReturn(0); 1025 } 1026 1027 #undef __FUNCT__ 1028 #define __FUNCT__ "MatIsTranspose_MPIAIJ" 1029 PetscErrorCode MatIsTranspose_MPIAIJ(Mat Amat,Mat Bmat,PetscReal tol,PetscBool *f) 1030 { 1031 MPI_Comm comm; 1032 Mat_MPIAIJ *Aij = (Mat_MPIAIJ*) Amat->data, *Bij; 1033 Mat Adia = Aij->A, Bdia, Aoff,Boff,*Aoffs,*Boffs; 1034 IS Me,Notme; 1035 PetscErrorCode ierr; 1036 PetscInt M,N,first,last,*notme,i; 1037 PetscMPIInt size; 1038 1039 PetscFunctionBegin; 1040 /* Easy test: symmetric diagonal block */ 1041 Bij = (Mat_MPIAIJ*) Bmat->data; Bdia = Bij->A; 1042 ierr = MatIsTranspose(Adia,Bdia,tol,f);CHKERRQ(ierr); 1043 if (!*f) PetscFunctionReturn(0); 1044 ierr = PetscObjectGetComm((PetscObject)Amat,&comm);CHKERRQ(ierr); 1045 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 1046 if (size == 1) PetscFunctionReturn(0); 1047 1048 /* Hard test: off-diagonal block. This takes a MatGetSubMatrix. */ 1049 ierr = MatGetSize(Amat,&M,&N);CHKERRQ(ierr); 1050 ierr = MatGetOwnershipRange(Amat,&first,&last);CHKERRQ(ierr); 1051 ierr = PetscMalloc1((N-last+first),¬me);CHKERRQ(ierr); 1052 for (i=0; i<first; i++) notme[i] = i; 1053 for (i=last; i<M; i++) notme[i-last+first] = i; 1054 ierr = ISCreateGeneral(MPI_COMM_SELF,N-last+first,notme,PETSC_COPY_VALUES,&Notme);CHKERRQ(ierr); 1055 ierr = ISCreateStride(MPI_COMM_SELF,last-first,first,1,&Me);CHKERRQ(ierr); 1056 ierr = MatGetSubMatrices(Amat,1,&Me,&Notme,MAT_INITIAL_MATRIX,&Aoffs);CHKERRQ(ierr); 1057 Aoff = Aoffs[0]; 1058 ierr = MatGetSubMatrices(Bmat,1,&Notme,&Me,MAT_INITIAL_MATRIX,&Boffs);CHKERRQ(ierr); 1059 Boff = Boffs[0]; 1060 ierr = MatIsTranspose(Aoff,Boff,tol,f);CHKERRQ(ierr); 1061 ierr = MatDestroyMatrices(1,&Aoffs);CHKERRQ(ierr); 1062 ierr = MatDestroyMatrices(1,&Boffs);CHKERRQ(ierr); 1063 ierr = ISDestroy(&Me);CHKERRQ(ierr); 1064 ierr = ISDestroy(&Notme);CHKERRQ(ierr); 1065 ierr = PetscFree(notme);CHKERRQ(ierr); 1066 PetscFunctionReturn(0); 1067 } 1068 1069 #undef __FUNCT__ 1070 #define __FUNCT__ "MatMultTransposeAdd_MPIAIJ" 1071 PetscErrorCode MatMultTransposeAdd_MPIAIJ(Mat A,Vec xx,Vec yy,Vec zz) 1072 { 1073 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 1074 PetscErrorCode ierr; 1075 1076 PetscFunctionBegin; 1077 /* do nondiagonal part */ 1078 ierr = (*a->B->ops->multtranspose)(a->B,xx,a->lvec);CHKERRQ(ierr); 1079 /* send it on its way */ 1080 ierr = VecScatterBegin(a->Mvctx,a->lvec,zz,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1081 /* do local part */ 1082 ierr = (*a->A->ops->multtransposeadd)(a->A,xx,yy,zz);CHKERRQ(ierr); 1083 /* receive remote parts */ 1084 ierr = VecScatterEnd(a->Mvctx,a->lvec,zz,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1085 PetscFunctionReturn(0); 1086 } 1087 1088 /* 1089 This only works correctly for square matrices where the subblock A->A is the 1090 diagonal block 1091 */ 1092 #undef __FUNCT__ 1093 #define __FUNCT__ "MatGetDiagonal_MPIAIJ" 1094 PetscErrorCode MatGetDiagonal_MPIAIJ(Mat A,Vec v) 1095 { 1096 PetscErrorCode ierr; 1097 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 1098 1099 PetscFunctionBegin; 1100 if (A->rmap->N != A->cmap->N) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"Supports only square matrix where A->A is diag block"); 1101 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"); 1102 ierr = MatGetDiagonal(a->A,v);CHKERRQ(ierr); 1103 PetscFunctionReturn(0); 1104 } 1105 1106 #undef __FUNCT__ 1107 #define __FUNCT__ "MatScale_MPIAIJ" 1108 PetscErrorCode MatScale_MPIAIJ(Mat A,PetscScalar aa) 1109 { 1110 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 1111 PetscErrorCode ierr; 1112 1113 PetscFunctionBegin; 1114 ierr = MatScale(a->A,aa);CHKERRQ(ierr); 1115 ierr = MatScale(a->B,aa);CHKERRQ(ierr); 1116 PetscFunctionReturn(0); 1117 } 1118 1119 #undef __FUNCT__ 1120 #define __FUNCT__ "MatDestroy_Redundant" 1121 PetscErrorCode MatDestroy_Redundant(Mat_Redundant **redundant) 1122 { 1123 PetscErrorCode ierr; 1124 Mat_Redundant *redund = *redundant; 1125 PetscInt i; 1126 1127 PetscFunctionBegin; 1128 *redundant = NULL; 1129 if (redund){ 1130 if (redund->matseq) { /* via MatGetSubMatrices() */ 1131 ierr = ISDestroy(&redund->isrow);CHKERRQ(ierr); 1132 ierr = ISDestroy(&redund->iscol);CHKERRQ(ierr); 1133 ierr = MatDestroy(&redund->matseq[0]);CHKERRQ(ierr); 1134 ierr = PetscFree(redund->matseq);CHKERRQ(ierr); 1135 } else { 1136 ierr = PetscFree2(redund->send_rank,redund->recv_rank);CHKERRQ(ierr); 1137 ierr = PetscFree(redund->sbuf_j);CHKERRQ(ierr); 1138 ierr = PetscFree(redund->sbuf_a);CHKERRQ(ierr); 1139 for (i=0; i<redund->nrecvs; i++) { 1140 ierr = PetscFree(redund->rbuf_j[i]);CHKERRQ(ierr); 1141 ierr = PetscFree(redund->rbuf_a[i]);CHKERRQ(ierr); 1142 } 1143 ierr = PetscFree4(redund->sbuf_nz,redund->rbuf_nz,redund->rbuf_j,redund->rbuf_a);CHKERRQ(ierr); 1144 } 1145 1146 if (redund->subcomm) { 1147 ierr = PetscCommDestroy(&redund->subcomm);CHKERRQ(ierr); 1148 } 1149 ierr = PetscFree(redund);CHKERRQ(ierr); 1150 } 1151 PetscFunctionReturn(0); 1152 } 1153 1154 #undef __FUNCT__ 1155 #define __FUNCT__ "MatDestroy_MPIAIJ" 1156 PetscErrorCode MatDestroy_MPIAIJ(Mat mat) 1157 { 1158 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 1159 PetscErrorCode ierr; 1160 1161 PetscFunctionBegin; 1162 #if defined(PETSC_USE_LOG) 1163 PetscLogObjectState((PetscObject)mat,"Rows=%D, Cols=%D",mat->rmap->N,mat->cmap->N); 1164 #endif 1165 ierr = MatDestroy_Redundant(&mat->redundant);CHKERRQ(ierr); 1166 ierr = MatStashDestroy_Private(&mat->stash);CHKERRQ(ierr); 1167 ierr = VecDestroy(&aij->diag);CHKERRQ(ierr); 1168 ierr = MatDestroy(&aij->A);CHKERRQ(ierr); 1169 ierr = MatDestroy(&aij->B);CHKERRQ(ierr); 1170 #if defined(PETSC_USE_CTABLE) 1171 ierr = PetscTableDestroy(&aij->colmap);CHKERRQ(ierr); 1172 #else 1173 ierr = PetscFree(aij->colmap);CHKERRQ(ierr); 1174 #endif 1175 ierr = PetscFree(aij->garray);CHKERRQ(ierr); 1176 ierr = VecDestroy(&aij->lvec);CHKERRQ(ierr); 1177 ierr = VecScatterDestroy(&aij->Mvctx);CHKERRQ(ierr); 1178 ierr = PetscFree2(aij->rowvalues,aij->rowindices);CHKERRQ(ierr); 1179 ierr = PetscFree(aij->ld);CHKERRQ(ierr); 1180 ierr = PetscFree(mat->data);CHKERRQ(ierr); 1181 1182 ierr = PetscObjectChangeTypeName((PetscObject)mat,0);CHKERRQ(ierr); 1183 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatStoreValues_C",NULL);CHKERRQ(ierr); 1184 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatRetrieveValues_C",NULL);CHKERRQ(ierr); 1185 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatGetDiagonalBlock_C",NULL);CHKERRQ(ierr); 1186 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatIsTranspose_C",NULL);CHKERRQ(ierr); 1187 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatMPIAIJSetPreallocation_C",NULL);CHKERRQ(ierr); 1188 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatMPIAIJSetPreallocationCSR_C",NULL);CHKERRQ(ierr); 1189 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatDiagonalScaleLocal_C",NULL);CHKERRQ(ierr); 1190 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatConvert_mpiaij_mpisbaij_C",NULL);CHKERRQ(ierr); 1191 #if defined(PETSC_HAVE_ELEMENTAL) 1192 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatConvert_mpiaij_elemental_C",NULL);CHKERRQ(ierr); 1193 #endif 1194 PetscFunctionReturn(0); 1195 } 1196 1197 #undef __FUNCT__ 1198 #define __FUNCT__ "MatView_MPIAIJ_Binary" 1199 PetscErrorCode MatView_MPIAIJ_Binary(Mat mat,PetscViewer viewer) 1200 { 1201 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 1202 Mat_SeqAIJ *A = (Mat_SeqAIJ*)aij->A->data; 1203 Mat_SeqAIJ *B = (Mat_SeqAIJ*)aij->B->data; 1204 PetscErrorCode ierr; 1205 PetscMPIInt rank,size,tag = ((PetscObject)viewer)->tag; 1206 int fd; 1207 PetscInt nz,header[4],*row_lengths,*range=0,rlen,i; 1208 PetscInt nzmax,*column_indices,j,k,col,*garray = aij->garray,cnt,cstart = mat->cmap->rstart,rnz = 0; 1209 PetscScalar *column_values; 1210 PetscInt message_count,flowcontrolcount; 1211 FILE *file; 1212 1213 PetscFunctionBegin; 1214 ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr); 1215 ierr = MPI_Comm_size(PetscObjectComm((PetscObject)mat),&size);CHKERRQ(ierr); 1216 nz = A->nz + B->nz; 1217 if (!rank) { 1218 header[0] = MAT_FILE_CLASSID; 1219 header[1] = mat->rmap->N; 1220 header[2] = mat->cmap->N; 1221 1222 ierr = MPI_Reduce(&nz,&header[3],1,MPIU_INT,MPI_SUM,0,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1223 ierr = PetscViewerBinaryGetDescriptor(viewer,&fd);CHKERRQ(ierr); 1224 ierr = PetscBinaryWrite(fd,header,4,PETSC_INT,PETSC_TRUE);CHKERRQ(ierr); 1225 /* get largest number of rows any processor has */ 1226 rlen = mat->rmap->n; 1227 range = mat->rmap->range; 1228 for (i=1; i<size; i++) rlen = PetscMax(rlen,range[i+1] - range[i]); 1229 } else { 1230 ierr = MPI_Reduce(&nz,0,1,MPIU_INT,MPI_SUM,0,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1231 rlen = mat->rmap->n; 1232 } 1233 1234 /* load up the local row counts */ 1235 ierr = PetscMalloc1((rlen+1),&row_lengths);CHKERRQ(ierr); 1236 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]; 1237 1238 /* store the row lengths to the file */ 1239 ierr = PetscViewerFlowControlStart(viewer,&message_count,&flowcontrolcount);CHKERRQ(ierr); 1240 if (!rank) { 1241 ierr = PetscBinaryWrite(fd,row_lengths,mat->rmap->n,PETSC_INT,PETSC_TRUE);CHKERRQ(ierr); 1242 for (i=1; i<size; i++) { 1243 ierr = PetscViewerFlowControlStepMaster(viewer,i,&message_count,flowcontrolcount);CHKERRQ(ierr); 1244 rlen = range[i+1] - range[i]; 1245 ierr = MPIULong_Recv(row_lengths,rlen,MPIU_INT,i,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1246 ierr = PetscBinaryWrite(fd,row_lengths,rlen,PETSC_INT,PETSC_TRUE);CHKERRQ(ierr); 1247 } 1248 ierr = PetscViewerFlowControlEndMaster(viewer,&message_count);CHKERRQ(ierr); 1249 } else { 1250 ierr = PetscViewerFlowControlStepWorker(viewer,rank,&message_count);CHKERRQ(ierr); 1251 ierr = MPIULong_Send(row_lengths,mat->rmap->n,MPIU_INT,0,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1252 ierr = PetscViewerFlowControlEndWorker(viewer,&message_count);CHKERRQ(ierr); 1253 } 1254 ierr = PetscFree(row_lengths);CHKERRQ(ierr); 1255 1256 /* load up the local column indices */ 1257 nzmax = nz; /* th processor needs space a largest processor needs */ 1258 ierr = MPI_Reduce(&nz,&nzmax,1,MPIU_INT,MPI_MAX,0,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1259 ierr = PetscMalloc1((nzmax+1),&column_indices);CHKERRQ(ierr); 1260 cnt = 0; 1261 for (i=0; i<mat->rmap->n; i++) { 1262 for (j=B->i[i]; j<B->i[i+1]; j++) { 1263 if ((col = garray[B->j[j]]) > cstart) break; 1264 column_indices[cnt++] = col; 1265 } 1266 for (k=A->i[i]; k<A->i[i+1]; k++) column_indices[cnt++] = A->j[k] + cstart; 1267 for (; j<B->i[i+1]; j++) column_indices[cnt++] = garray[B->j[j]]; 1268 } 1269 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); 1270 1271 /* store the column indices to the file */ 1272 ierr = PetscViewerFlowControlStart(viewer,&message_count,&flowcontrolcount);CHKERRQ(ierr); 1273 if (!rank) { 1274 MPI_Status status; 1275 ierr = PetscBinaryWrite(fd,column_indices,nz,PETSC_INT,PETSC_TRUE);CHKERRQ(ierr); 1276 for (i=1; i<size; i++) { 1277 ierr = PetscViewerFlowControlStepMaster(viewer,i,&message_count,flowcontrolcount);CHKERRQ(ierr); 1278 ierr = MPI_Recv(&rnz,1,MPIU_INT,i,tag,PetscObjectComm((PetscObject)mat),&status);CHKERRQ(ierr); 1279 if (rnz > nzmax) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_LIB,"Internal PETSc error: nz = %D nzmax = %D",nz,nzmax); 1280 ierr = MPIULong_Recv(column_indices,rnz,MPIU_INT,i,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1281 ierr = PetscBinaryWrite(fd,column_indices,rnz,PETSC_INT,PETSC_TRUE);CHKERRQ(ierr); 1282 } 1283 ierr = PetscViewerFlowControlEndMaster(viewer,&message_count);CHKERRQ(ierr); 1284 } else { 1285 ierr = PetscViewerFlowControlStepWorker(viewer,rank,&message_count);CHKERRQ(ierr); 1286 ierr = MPI_Send(&nz,1,MPIU_INT,0,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1287 ierr = MPIULong_Send(column_indices,nz,MPIU_INT,0,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1288 ierr = PetscViewerFlowControlEndWorker(viewer,&message_count);CHKERRQ(ierr); 1289 } 1290 ierr = PetscFree(column_indices);CHKERRQ(ierr); 1291 1292 /* load up the local column values */ 1293 ierr = PetscMalloc1((nzmax+1),&column_values);CHKERRQ(ierr); 1294 cnt = 0; 1295 for (i=0; i<mat->rmap->n; i++) { 1296 for (j=B->i[i]; j<B->i[i+1]; j++) { 1297 if (garray[B->j[j]] > cstart) break; 1298 column_values[cnt++] = B->a[j]; 1299 } 1300 for (k=A->i[i]; k<A->i[i+1]; k++) column_values[cnt++] = A->a[k]; 1301 for (; j<B->i[i+1]; j++) column_values[cnt++] = B->a[j]; 1302 } 1303 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); 1304 1305 /* store the column values to the file */ 1306 ierr = PetscViewerFlowControlStart(viewer,&message_count,&flowcontrolcount);CHKERRQ(ierr); 1307 if (!rank) { 1308 MPI_Status status; 1309 ierr = PetscBinaryWrite(fd,column_values,nz,PETSC_SCALAR,PETSC_TRUE);CHKERRQ(ierr); 1310 for (i=1; i<size; i++) { 1311 ierr = PetscViewerFlowControlStepMaster(viewer,i,&message_count,flowcontrolcount);CHKERRQ(ierr); 1312 ierr = MPI_Recv(&rnz,1,MPIU_INT,i,tag,PetscObjectComm((PetscObject)mat),&status);CHKERRQ(ierr); 1313 if (rnz > nzmax) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_LIB,"Internal PETSc error: nz = %D nzmax = %D",nz,nzmax); 1314 ierr = MPIULong_Recv(column_values,rnz,MPIU_SCALAR,i,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1315 ierr = PetscBinaryWrite(fd,column_values,rnz,PETSC_SCALAR,PETSC_TRUE);CHKERRQ(ierr); 1316 } 1317 ierr = PetscViewerFlowControlEndMaster(viewer,&message_count);CHKERRQ(ierr); 1318 } else { 1319 ierr = PetscViewerFlowControlStepWorker(viewer,rank,&message_count);CHKERRQ(ierr); 1320 ierr = MPI_Send(&nz,1,MPIU_INT,0,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1321 ierr = MPIULong_Send(column_values,nz,MPIU_SCALAR,0,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1322 ierr = PetscViewerFlowControlEndWorker(viewer,&message_count);CHKERRQ(ierr); 1323 } 1324 ierr = PetscFree(column_values);CHKERRQ(ierr); 1325 1326 ierr = PetscViewerBinaryGetInfoPointer(viewer,&file);CHKERRQ(ierr); 1327 if (file) fprintf(file,"-matload_block_size %d\n",(int)PetscAbs(mat->rmap->bs)); 1328 PetscFunctionReturn(0); 1329 } 1330 1331 #include <petscdraw.h> 1332 #undef __FUNCT__ 1333 #define __FUNCT__ "MatView_MPIAIJ_ASCIIorDraworSocket" 1334 PetscErrorCode MatView_MPIAIJ_ASCIIorDraworSocket(Mat mat,PetscViewer viewer) 1335 { 1336 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 1337 PetscErrorCode ierr; 1338 PetscMPIInt rank = aij->rank,size = aij->size; 1339 PetscBool isdraw,iascii,isbinary; 1340 PetscViewer sviewer; 1341 PetscViewerFormat format; 1342 1343 PetscFunctionBegin; 1344 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERDRAW,&isdraw);CHKERRQ(ierr); 1345 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii);CHKERRQ(ierr); 1346 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERBINARY,&isbinary);CHKERRQ(ierr); 1347 if (iascii) { 1348 ierr = PetscViewerGetFormat(viewer,&format);CHKERRQ(ierr); 1349 if (format == PETSC_VIEWER_ASCII_INFO_DETAIL) { 1350 MatInfo info; 1351 PetscBool inodes; 1352 1353 ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr); 1354 ierr = MatGetInfo(mat,MAT_LOCAL,&info);CHKERRQ(ierr); 1355 ierr = MatInodeGetInodeSizes(aij->A,NULL,(PetscInt**)&inodes,NULL);CHKERRQ(ierr); 1356 ierr = PetscViewerASCIISynchronizedAllow(viewer,PETSC_TRUE);CHKERRQ(ierr); 1357 if (!inodes) { 1358 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"[%d] Local rows %D nz %D nz alloced %D mem %D, not using I-node routines\n", 1359 rank,mat->rmap->n,(PetscInt)info.nz_used,(PetscInt)info.nz_allocated,(PetscInt)info.memory);CHKERRQ(ierr); 1360 } else { 1361 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"[%d] Local rows %D nz %D nz alloced %D mem %D, using I-node routines\n", 1362 rank,mat->rmap->n,(PetscInt)info.nz_used,(PetscInt)info.nz_allocated,(PetscInt)info.memory);CHKERRQ(ierr); 1363 } 1364 ierr = MatGetInfo(aij->A,MAT_LOCAL,&info);CHKERRQ(ierr); 1365 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"[%d] on-diagonal part: nz %D \n",rank,(PetscInt)info.nz_used);CHKERRQ(ierr); 1366 ierr = MatGetInfo(aij->B,MAT_LOCAL,&info);CHKERRQ(ierr); 1367 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"[%d] off-diagonal part: nz %D \n",rank,(PetscInt)info.nz_used);CHKERRQ(ierr); 1368 ierr = PetscViewerFlush(viewer);CHKERRQ(ierr); 1369 ierr = PetscViewerASCIISynchronizedAllow(viewer,PETSC_FALSE);CHKERRQ(ierr); 1370 ierr = PetscViewerASCIIPrintf(viewer,"Information on VecScatter used in matrix-vector product: \n");CHKERRQ(ierr); 1371 ierr = VecScatterView(aij->Mvctx,viewer);CHKERRQ(ierr); 1372 PetscFunctionReturn(0); 1373 } else if (format == PETSC_VIEWER_ASCII_INFO) { 1374 PetscInt inodecount,inodelimit,*inodes; 1375 ierr = MatInodeGetInodeSizes(aij->A,&inodecount,&inodes,&inodelimit);CHKERRQ(ierr); 1376 if (inodes) { 1377 ierr = PetscViewerASCIIPrintf(viewer,"using I-node (on process 0) routines: found %D nodes, limit used is %D\n",inodecount,inodelimit);CHKERRQ(ierr); 1378 } else { 1379 ierr = PetscViewerASCIIPrintf(viewer,"not using I-node (on process 0) routines\n");CHKERRQ(ierr); 1380 } 1381 PetscFunctionReturn(0); 1382 } else if (format == PETSC_VIEWER_ASCII_FACTOR_INFO) { 1383 PetscFunctionReturn(0); 1384 } 1385 } else if (isbinary) { 1386 if (size == 1) { 1387 ierr = PetscObjectSetName((PetscObject)aij->A,((PetscObject)mat)->name);CHKERRQ(ierr); 1388 ierr = MatView(aij->A,viewer);CHKERRQ(ierr); 1389 } else { 1390 ierr = MatView_MPIAIJ_Binary(mat,viewer);CHKERRQ(ierr); 1391 } 1392 PetscFunctionReturn(0); 1393 } else if (isdraw) { 1394 PetscDraw draw; 1395 PetscBool isnull; 1396 ierr = PetscViewerDrawGetDraw(viewer,0,&draw);CHKERRQ(ierr); 1397 ierr = PetscDrawIsNull(draw,&isnull);CHKERRQ(ierr); if (isnull) PetscFunctionReturn(0); 1398 } 1399 1400 { 1401 /* assemble the entire matrix onto first processor. */ 1402 Mat A; 1403 Mat_SeqAIJ *Aloc; 1404 PetscInt M = mat->rmap->N,N = mat->cmap->N,m,*ai,*aj,row,*cols,i,*ct; 1405 MatScalar *a; 1406 1407 ierr = MatCreate(PetscObjectComm((PetscObject)mat),&A);CHKERRQ(ierr); 1408 if (!rank) { 1409 ierr = MatSetSizes(A,M,N,M,N);CHKERRQ(ierr); 1410 } else { 1411 ierr = MatSetSizes(A,0,0,M,N);CHKERRQ(ierr); 1412 } 1413 /* This is just a temporary matrix, so explicitly using MATMPIAIJ is probably best */ 1414 ierr = MatSetType(A,MATMPIAIJ);CHKERRQ(ierr); 1415 ierr = MatMPIAIJSetPreallocation(A,0,NULL,0,NULL);CHKERRQ(ierr); 1416 ierr = MatSetOption(A,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 1417 ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)A);CHKERRQ(ierr); 1418 1419 /* copy over the A part */ 1420 Aloc = (Mat_SeqAIJ*)aij->A->data; 1421 m = aij->A->rmap->n; ai = Aloc->i; aj = Aloc->j; a = Aloc->a; 1422 row = mat->rmap->rstart; 1423 for (i=0; i<ai[m]; i++) aj[i] += mat->cmap->rstart; 1424 for (i=0; i<m; i++) { 1425 ierr = MatSetValues(A,1,&row,ai[i+1]-ai[i],aj,a,INSERT_VALUES);CHKERRQ(ierr); 1426 row++; 1427 a += ai[i+1]-ai[i]; aj += ai[i+1]-ai[i]; 1428 } 1429 aj = Aloc->j; 1430 for (i=0; i<ai[m]; i++) aj[i] -= mat->cmap->rstart; 1431 1432 /* copy over the B part */ 1433 Aloc = (Mat_SeqAIJ*)aij->B->data; 1434 m = aij->B->rmap->n; ai = Aloc->i; aj = Aloc->j; a = Aloc->a; 1435 row = mat->rmap->rstart; 1436 ierr = PetscMalloc1((ai[m]+1),&cols);CHKERRQ(ierr); 1437 ct = cols; 1438 for (i=0; i<ai[m]; i++) cols[i] = aij->garray[aj[i]]; 1439 for (i=0; i<m; i++) { 1440 ierr = MatSetValues(A,1,&row,ai[i+1]-ai[i],cols,a,INSERT_VALUES);CHKERRQ(ierr); 1441 row++; 1442 a += ai[i+1]-ai[i]; cols += ai[i+1]-ai[i]; 1443 } 1444 ierr = PetscFree(ct);CHKERRQ(ierr); 1445 ierr = MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1446 ierr = MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1447 /* 1448 Everyone has to call to draw the matrix since the graphics waits are 1449 synchronized across all processors that share the PetscDraw object 1450 */ 1451 ierr = PetscViewerGetSingleton(viewer,&sviewer);CHKERRQ(ierr); 1452 if (!rank) { 1453 ierr = MatView_SeqAIJ(((Mat_MPIAIJ*)(A->data))->A,sviewer);CHKERRQ(ierr); 1454 } 1455 ierr = PetscViewerRestoreSingleton(viewer,&sviewer);CHKERRQ(ierr); 1456 ierr = MatDestroy(&A);CHKERRQ(ierr); 1457 } 1458 PetscFunctionReturn(0); 1459 } 1460 1461 #undef __FUNCT__ 1462 #define __FUNCT__ "MatView_MPIAIJ" 1463 PetscErrorCode MatView_MPIAIJ(Mat mat,PetscViewer viewer) 1464 { 1465 PetscErrorCode ierr; 1466 PetscBool iascii,isdraw,issocket,isbinary; 1467 1468 PetscFunctionBegin; 1469 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii);CHKERRQ(ierr); 1470 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERDRAW,&isdraw);CHKERRQ(ierr); 1471 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERBINARY,&isbinary);CHKERRQ(ierr); 1472 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERSOCKET,&issocket);CHKERRQ(ierr); 1473 if (iascii || isdraw || isbinary || issocket) { 1474 ierr = MatView_MPIAIJ_ASCIIorDraworSocket(mat,viewer);CHKERRQ(ierr); 1475 } 1476 PetscFunctionReturn(0); 1477 } 1478 1479 #undef __FUNCT__ 1480 #define __FUNCT__ "MatSOR_MPIAIJ" 1481 PetscErrorCode MatSOR_MPIAIJ(Mat matin,Vec bb,PetscReal omega,MatSORType flag,PetscReal fshift,PetscInt its,PetscInt lits,Vec xx) 1482 { 1483 Mat_MPIAIJ *mat = (Mat_MPIAIJ*)matin->data; 1484 PetscErrorCode ierr; 1485 Vec bb1 = 0; 1486 PetscBool hasop; 1487 1488 PetscFunctionBegin; 1489 if (flag == SOR_APPLY_UPPER) { 1490 ierr = (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);CHKERRQ(ierr); 1491 PetscFunctionReturn(0); 1492 } 1493 1494 if (its > 1 || ~flag & SOR_ZERO_INITIAL_GUESS || flag & SOR_EISENSTAT) { 1495 ierr = VecDuplicate(bb,&bb1);CHKERRQ(ierr); 1496 } 1497 1498 if ((flag & SOR_LOCAL_SYMMETRIC_SWEEP) == SOR_LOCAL_SYMMETRIC_SWEEP) { 1499 if (flag & SOR_ZERO_INITIAL_GUESS) { 1500 ierr = (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);CHKERRQ(ierr); 1501 its--; 1502 } 1503 1504 while (its--) { 1505 ierr = VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1506 ierr = VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1507 1508 /* update rhs: bb1 = bb - B*x */ 1509 ierr = VecScale(mat->lvec,-1.0);CHKERRQ(ierr); 1510 ierr = (*mat->B->ops->multadd)(mat->B,mat->lvec,bb,bb1);CHKERRQ(ierr); 1511 1512 /* local sweep */ 1513 ierr = (*mat->A->ops->sor)(mat->A,bb1,omega,SOR_SYMMETRIC_SWEEP,fshift,lits,1,xx);CHKERRQ(ierr); 1514 } 1515 } else if (flag & SOR_LOCAL_FORWARD_SWEEP) { 1516 if (flag & SOR_ZERO_INITIAL_GUESS) { 1517 ierr = (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);CHKERRQ(ierr); 1518 its--; 1519 } 1520 while (its--) { 1521 ierr = VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1522 ierr = VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1523 1524 /* update rhs: bb1 = bb - B*x */ 1525 ierr = VecScale(mat->lvec,-1.0);CHKERRQ(ierr); 1526 ierr = (*mat->B->ops->multadd)(mat->B,mat->lvec,bb,bb1);CHKERRQ(ierr); 1527 1528 /* local sweep */ 1529 ierr = (*mat->A->ops->sor)(mat->A,bb1,omega,SOR_FORWARD_SWEEP,fshift,lits,1,xx);CHKERRQ(ierr); 1530 } 1531 } else if (flag & SOR_LOCAL_BACKWARD_SWEEP) { 1532 if (flag & SOR_ZERO_INITIAL_GUESS) { 1533 ierr = (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);CHKERRQ(ierr); 1534 its--; 1535 } 1536 while (its--) { 1537 ierr = VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1538 ierr = VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1539 1540 /* update rhs: bb1 = bb - B*x */ 1541 ierr = VecScale(mat->lvec,-1.0);CHKERRQ(ierr); 1542 ierr = (*mat->B->ops->multadd)(mat->B,mat->lvec,bb,bb1);CHKERRQ(ierr); 1543 1544 /* local sweep */ 1545 ierr = (*mat->A->ops->sor)(mat->A,bb1,omega,SOR_BACKWARD_SWEEP,fshift,lits,1,xx);CHKERRQ(ierr); 1546 } 1547 } else if (flag & SOR_EISENSTAT) { 1548 Vec xx1; 1549 1550 ierr = VecDuplicate(bb,&xx1);CHKERRQ(ierr); 1551 ierr = (*mat->A->ops->sor)(mat->A,bb,omega,(MatSORType)(SOR_ZERO_INITIAL_GUESS | SOR_LOCAL_BACKWARD_SWEEP),fshift,lits,1,xx);CHKERRQ(ierr); 1552 1553 ierr = VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1554 ierr = VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1555 if (!mat->diag) { 1556 ierr = MatCreateVecs(matin,&mat->diag,NULL);CHKERRQ(ierr); 1557 ierr = MatGetDiagonal(matin,mat->diag);CHKERRQ(ierr); 1558 } 1559 ierr = MatHasOperation(matin,MATOP_MULT_DIAGONAL_BLOCK,&hasop);CHKERRQ(ierr); 1560 if (hasop) { 1561 ierr = MatMultDiagonalBlock(matin,xx,bb1);CHKERRQ(ierr); 1562 } else { 1563 ierr = VecPointwiseMult(bb1,mat->diag,xx);CHKERRQ(ierr); 1564 } 1565 ierr = VecAYPX(bb1,(omega-2.0)/omega,bb);CHKERRQ(ierr); 1566 1567 ierr = MatMultAdd(mat->B,mat->lvec,bb1,bb1);CHKERRQ(ierr); 1568 1569 /* local sweep */ 1570 ierr = (*mat->A->ops->sor)(mat->A,bb1,omega,(MatSORType)(SOR_ZERO_INITIAL_GUESS | SOR_LOCAL_FORWARD_SWEEP),fshift,lits,1,xx1);CHKERRQ(ierr); 1571 ierr = VecAXPY(xx,1.0,xx1);CHKERRQ(ierr); 1572 ierr = VecDestroy(&xx1);CHKERRQ(ierr); 1573 } else SETERRQ(PetscObjectComm((PetscObject)matin),PETSC_ERR_SUP,"Parallel SOR not supported"); 1574 1575 ierr = VecDestroy(&bb1);CHKERRQ(ierr); 1576 PetscFunctionReturn(0); 1577 } 1578 1579 #undef __FUNCT__ 1580 #define __FUNCT__ "MatPermute_MPIAIJ" 1581 PetscErrorCode MatPermute_MPIAIJ(Mat A,IS rowp,IS colp,Mat *B) 1582 { 1583 Mat aA,aB,Aperm; 1584 const PetscInt *rwant,*cwant,*gcols,*ai,*bi,*aj,*bj; 1585 PetscScalar *aa,*ba; 1586 PetscInt i,j,m,n,ng,anz,bnz,*dnnz,*onnz,*tdnnz,*tonnz,*rdest,*cdest,*work,*gcdest; 1587 PetscSF rowsf,sf; 1588 IS parcolp = NULL; 1589 PetscBool done; 1590 PetscErrorCode ierr; 1591 1592 PetscFunctionBegin; 1593 ierr = MatGetLocalSize(A,&m,&n);CHKERRQ(ierr); 1594 ierr = ISGetIndices(rowp,&rwant);CHKERRQ(ierr); 1595 ierr = ISGetIndices(colp,&cwant);CHKERRQ(ierr); 1596 ierr = PetscMalloc3(PetscMax(m,n),&work,m,&rdest,n,&cdest);CHKERRQ(ierr); 1597 1598 /* Invert row permutation to find out where my rows should go */ 1599 ierr = PetscSFCreate(PetscObjectComm((PetscObject)A),&rowsf);CHKERRQ(ierr); 1600 ierr = PetscSFSetGraphLayout(rowsf,A->rmap,A->rmap->n,NULL,PETSC_OWN_POINTER,rwant);CHKERRQ(ierr); 1601 ierr = PetscSFSetFromOptions(rowsf);CHKERRQ(ierr); 1602 for (i=0; i<m; i++) work[i] = A->rmap->rstart + i; 1603 ierr = PetscSFReduceBegin(rowsf,MPIU_INT,work,rdest,MPIU_REPLACE);CHKERRQ(ierr); 1604 ierr = PetscSFReduceEnd(rowsf,MPIU_INT,work,rdest,MPIU_REPLACE);CHKERRQ(ierr); 1605 1606 /* Invert column permutation to find out where my columns should go */ 1607 ierr = PetscSFCreate(PetscObjectComm((PetscObject)A),&sf);CHKERRQ(ierr); 1608 ierr = PetscSFSetGraphLayout(sf,A->cmap,A->cmap->n,NULL,PETSC_OWN_POINTER,cwant);CHKERRQ(ierr); 1609 ierr = PetscSFSetFromOptions(sf);CHKERRQ(ierr); 1610 for (i=0; i<n; i++) work[i] = A->cmap->rstart + i; 1611 ierr = PetscSFReduceBegin(sf,MPIU_INT,work,cdest,MPIU_REPLACE);CHKERRQ(ierr); 1612 ierr = PetscSFReduceEnd(sf,MPIU_INT,work,cdest,MPIU_REPLACE);CHKERRQ(ierr); 1613 ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 1614 1615 ierr = ISRestoreIndices(rowp,&rwant);CHKERRQ(ierr); 1616 ierr = ISRestoreIndices(colp,&cwant);CHKERRQ(ierr); 1617 ierr = MatMPIAIJGetSeqAIJ(A,&aA,&aB,&gcols);CHKERRQ(ierr); 1618 1619 /* Find out where my gcols should go */ 1620 ierr = MatGetSize(aB,NULL,&ng);CHKERRQ(ierr); 1621 ierr = PetscMalloc1(ng,&gcdest);CHKERRQ(ierr); 1622 ierr = PetscSFCreate(PetscObjectComm((PetscObject)A),&sf);CHKERRQ(ierr); 1623 ierr = PetscSFSetGraphLayout(sf,A->cmap,ng,NULL,PETSC_OWN_POINTER,gcols);CHKERRQ(ierr); 1624 ierr = PetscSFSetFromOptions(sf);CHKERRQ(ierr); 1625 ierr = PetscSFBcastBegin(sf,MPIU_INT,cdest,gcdest);CHKERRQ(ierr); 1626 ierr = PetscSFBcastEnd(sf,MPIU_INT,cdest,gcdest);CHKERRQ(ierr); 1627 ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 1628 1629 ierr = PetscCalloc4(m,&dnnz,m,&onnz,m,&tdnnz,m,&tonnz);CHKERRQ(ierr); 1630 ierr = MatGetRowIJ(aA,0,PETSC_FALSE,PETSC_FALSE,&anz,&ai,&aj,&done);CHKERRQ(ierr); 1631 ierr = MatGetRowIJ(aB,0,PETSC_FALSE,PETSC_FALSE,&bnz,&bi,&bj,&done);CHKERRQ(ierr); 1632 for (i=0; i<m; i++) { 1633 PetscInt row = rdest[i],rowner; 1634 ierr = PetscLayoutFindOwner(A->rmap,row,&rowner);CHKERRQ(ierr); 1635 for (j=ai[i]; j<ai[i+1]; j++) { 1636 PetscInt cowner,col = cdest[aj[j]]; 1637 ierr = PetscLayoutFindOwner(A->cmap,col,&cowner);CHKERRQ(ierr); /* Could build an index for the columns to eliminate this search */ 1638 if (rowner == cowner) dnnz[i]++; 1639 else onnz[i]++; 1640 } 1641 for (j=bi[i]; j<bi[i+1]; j++) { 1642 PetscInt cowner,col = gcdest[bj[j]]; 1643 ierr = PetscLayoutFindOwner(A->cmap,col,&cowner);CHKERRQ(ierr); 1644 if (rowner == cowner) dnnz[i]++; 1645 else onnz[i]++; 1646 } 1647 } 1648 ierr = PetscSFBcastBegin(rowsf,MPIU_INT,dnnz,tdnnz);CHKERRQ(ierr); 1649 ierr = PetscSFBcastEnd(rowsf,MPIU_INT,dnnz,tdnnz);CHKERRQ(ierr); 1650 ierr = PetscSFBcastBegin(rowsf,MPIU_INT,onnz,tonnz);CHKERRQ(ierr); 1651 ierr = PetscSFBcastEnd(rowsf,MPIU_INT,onnz,tonnz);CHKERRQ(ierr); 1652 ierr = PetscSFDestroy(&rowsf);CHKERRQ(ierr); 1653 1654 ierr = MatCreateAIJ(PetscObjectComm((PetscObject)A),A->rmap->n,A->cmap->n,A->rmap->N,A->cmap->N,0,tdnnz,0,tonnz,&Aperm);CHKERRQ(ierr); 1655 ierr = MatSeqAIJGetArray(aA,&aa);CHKERRQ(ierr); 1656 ierr = MatSeqAIJGetArray(aB,&ba);CHKERRQ(ierr); 1657 for (i=0; i<m; i++) { 1658 PetscInt *acols = dnnz,*bcols = onnz; /* Repurpose now-unneeded arrays */ 1659 PetscInt j0,rowlen; 1660 rowlen = ai[i+1] - ai[i]; 1661 for (j0=j=0; j<rowlen; j0=j) { /* rowlen could be larger than number of rows m, so sum in batches */ 1662 for ( ; j<PetscMin(rowlen,j0+m); j++) acols[j-j0] = cdest[aj[ai[i]+j]]; 1663 ierr = MatSetValues(Aperm,1,&rdest[i],j-j0,acols,aa+ai[i]+j0,INSERT_VALUES);CHKERRQ(ierr); 1664 } 1665 rowlen = bi[i+1] - bi[i]; 1666 for (j0=j=0; j<rowlen; j0=j) { 1667 for ( ; j<PetscMin(rowlen,j0+m); j++) bcols[j-j0] = gcdest[bj[bi[i]+j]]; 1668 ierr = MatSetValues(Aperm,1,&rdest[i],j-j0,bcols,ba+bi[i]+j0,INSERT_VALUES);CHKERRQ(ierr); 1669 } 1670 } 1671 ierr = MatAssemblyBegin(Aperm,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1672 ierr = MatAssemblyEnd(Aperm,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1673 ierr = MatRestoreRowIJ(aA,0,PETSC_FALSE,PETSC_FALSE,&anz,&ai,&aj,&done);CHKERRQ(ierr); 1674 ierr = MatRestoreRowIJ(aB,0,PETSC_FALSE,PETSC_FALSE,&bnz,&bi,&bj,&done);CHKERRQ(ierr); 1675 ierr = MatSeqAIJRestoreArray(aA,&aa);CHKERRQ(ierr); 1676 ierr = MatSeqAIJRestoreArray(aB,&ba);CHKERRQ(ierr); 1677 ierr = PetscFree4(dnnz,onnz,tdnnz,tonnz);CHKERRQ(ierr); 1678 ierr = PetscFree3(work,rdest,cdest);CHKERRQ(ierr); 1679 ierr = PetscFree(gcdest);CHKERRQ(ierr); 1680 if (parcolp) {ierr = ISDestroy(&colp);CHKERRQ(ierr);} 1681 *B = Aperm; 1682 PetscFunctionReturn(0); 1683 } 1684 1685 #undef __FUNCT__ 1686 #define __FUNCT__ "MatGetInfo_MPIAIJ" 1687 PetscErrorCode MatGetInfo_MPIAIJ(Mat matin,MatInfoType flag,MatInfo *info) 1688 { 1689 Mat_MPIAIJ *mat = (Mat_MPIAIJ*)matin->data; 1690 Mat A = mat->A,B = mat->B; 1691 PetscErrorCode ierr; 1692 PetscReal isend[5],irecv[5]; 1693 1694 PetscFunctionBegin; 1695 info->block_size = 1.0; 1696 ierr = MatGetInfo(A,MAT_LOCAL,info);CHKERRQ(ierr); 1697 1698 isend[0] = info->nz_used; isend[1] = info->nz_allocated; isend[2] = info->nz_unneeded; 1699 isend[3] = info->memory; isend[4] = info->mallocs; 1700 1701 ierr = MatGetInfo(B,MAT_LOCAL,info);CHKERRQ(ierr); 1702 1703 isend[0] += info->nz_used; isend[1] += info->nz_allocated; isend[2] += info->nz_unneeded; 1704 isend[3] += info->memory; isend[4] += info->mallocs; 1705 if (flag == MAT_LOCAL) { 1706 info->nz_used = isend[0]; 1707 info->nz_allocated = isend[1]; 1708 info->nz_unneeded = isend[2]; 1709 info->memory = isend[3]; 1710 info->mallocs = isend[4]; 1711 } else if (flag == MAT_GLOBAL_MAX) { 1712 ierr = MPI_Allreduce(isend,irecv,5,MPIU_REAL,MPIU_MAX,PetscObjectComm((PetscObject)matin));CHKERRQ(ierr); 1713 1714 info->nz_used = irecv[0]; 1715 info->nz_allocated = irecv[1]; 1716 info->nz_unneeded = irecv[2]; 1717 info->memory = irecv[3]; 1718 info->mallocs = irecv[4]; 1719 } else if (flag == MAT_GLOBAL_SUM) { 1720 ierr = MPI_Allreduce(isend,irecv,5,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)matin));CHKERRQ(ierr); 1721 1722 info->nz_used = irecv[0]; 1723 info->nz_allocated = irecv[1]; 1724 info->nz_unneeded = irecv[2]; 1725 info->memory = irecv[3]; 1726 info->mallocs = irecv[4]; 1727 } 1728 info->fill_ratio_given = 0; /* no parallel LU/ILU/Cholesky */ 1729 info->fill_ratio_needed = 0; 1730 info->factor_mallocs = 0; 1731 PetscFunctionReturn(0); 1732 } 1733 1734 #undef __FUNCT__ 1735 #define __FUNCT__ "MatSetOption_MPIAIJ" 1736 PetscErrorCode MatSetOption_MPIAIJ(Mat A,MatOption op,PetscBool flg) 1737 { 1738 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 1739 PetscErrorCode ierr; 1740 1741 PetscFunctionBegin; 1742 switch (op) { 1743 case MAT_NEW_NONZERO_LOCATIONS: 1744 case MAT_NEW_NONZERO_ALLOCATION_ERR: 1745 case MAT_UNUSED_NONZERO_LOCATION_ERR: 1746 case MAT_KEEP_NONZERO_PATTERN: 1747 case MAT_NEW_NONZERO_LOCATION_ERR: 1748 case MAT_USE_INODES: 1749 case MAT_IGNORE_ZERO_ENTRIES: 1750 MatCheckPreallocated(A,1); 1751 ierr = MatSetOption(a->A,op,flg);CHKERRQ(ierr); 1752 ierr = MatSetOption(a->B,op,flg);CHKERRQ(ierr); 1753 break; 1754 case MAT_ROW_ORIENTED: 1755 a->roworiented = flg; 1756 1757 ierr = MatSetOption(a->A,op,flg);CHKERRQ(ierr); 1758 ierr = MatSetOption(a->B,op,flg);CHKERRQ(ierr); 1759 break; 1760 case MAT_NEW_DIAGONALS: 1761 ierr = PetscInfo1(A,"Option %s ignored\n",MatOptions[op]);CHKERRQ(ierr); 1762 break; 1763 case MAT_IGNORE_OFF_PROC_ENTRIES: 1764 a->donotstash = flg; 1765 break; 1766 case MAT_SPD: 1767 A->spd_set = PETSC_TRUE; 1768 A->spd = flg; 1769 if (flg) { 1770 A->symmetric = PETSC_TRUE; 1771 A->structurally_symmetric = PETSC_TRUE; 1772 A->symmetric_set = PETSC_TRUE; 1773 A->structurally_symmetric_set = PETSC_TRUE; 1774 } 1775 break; 1776 case MAT_SYMMETRIC: 1777 ierr = MatSetOption(a->A,op,flg);CHKERRQ(ierr); 1778 break; 1779 case MAT_STRUCTURALLY_SYMMETRIC: 1780 ierr = MatSetOption(a->A,op,flg);CHKERRQ(ierr); 1781 break; 1782 case MAT_HERMITIAN: 1783 ierr = MatSetOption(a->A,op,flg);CHKERRQ(ierr); 1784 break; 1785 case MAT_SYMMETRY_ETERNAL: 1786 ierr = MatSetOption(a->A,op,flg);CHKERRQ(ierr); 1787 break; 1788 default: 1789 SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"unknown option %d",op); 1790 } 1791 PetscFunctionReturn(0); 1792 } 1793 1794 #undef __FUNCT__ 1795 #define __FUNCT__ "MatGetRow_MPIAIJ" 1796 PetscErrorCode MatGetRow_MPIAIJ(Mat matin,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v) 1797 { 1798 Mat_MPIAIJ *mat = (Mat_MPIAIJ*)matin->data; 1799 PetscScalar *vworkA,*vworkB,**pvA,**pvB,*v_p; 1800 PetscErrorCode ierr; 1801 PetscInt i,*cworkA,*cworkB,**pcA,**pcB,cstart = matin->cmap->rstart; 1802 PetscInt nztot,nzA,nzB,lrow,rstart = matin->rmap->rstart,rend = matin->rmap->rend; 1803 PetscInt *cmap,*idx_p; 1804 1805 PetscFunctionBegin; 1806 if (mat->getrowactive) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Already active"); 1807 mat->getrowactive = PETSC_TRUE; 1808 1809 if (!mat->rowvalues && (idx || v)) { 1810 /* 1811 allocate enough space to hold information from the longest row. 1812 */ 1813 Mat_SeqAIJ *Aa = (Mat_SeqAIJ*)mat->A->data,*Ba = (Mat_SeqAIJ*)mat->B->data; 1814 PetscInt max = 1,tmp; 1815 for (i=0; i<matin->rmap->n; i++) { 1816 tmp = Aa->i[i+1] - Aa->i[i] + Ba->i[i+1] - Ba->i[i]; 1817 if (max < tmp) max = tmp; 1818 } 1819 ierr = PetscMalloc2(max,&mat->rowvalues,max,&mat->rowindices);CHKERRQ(ierr); 1820 } 1821 1822 if (row < rstart || row >= rend) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Only local rows"); 1823 lrow = row - rstart; 1824 1825 pvA = &vworkA; pcA = &cworkA; pvB = &vworkB; pcB = &cworkB; 1826 if (!v) {pvA = 0; pvB = 0;} 1827 if (!idx) {pcA = 0; if (!v) pcB = 0;} 1828 ierr = (*mat->A->ops->getrow)(mat->A,lrow,&nzA,pcA,pvA);CHKERRQ(ierr); 1829 ierr = (*mat->B->ops->getrow)(mat->B,lrow,&nzB,pcB,pvB);CHKERRQ(ierr); 1830 nztot = nzA + nzB; 1831 1832 cmap = mat->garray; 1833 if (v || idx) { 1834 if (nztot) { 1835 /* Sort by increasing column numbers, assuming A and B already sorted */ 1836 PetscInt imark = -1; 1837 if (v) { 1838 *v = v_p = mat->rowvalues; 1839 for (i=0; i<nzB; i++) { 1840 if (cmap[cworkB[i]] < cstart) v_p[i] = vworkB[i]; 1841 else break; 1842 } 1843 imark = i; 1844 for (i=0; i<nzA; i++) v_p[imark+i] = vworkA[i]; 1845 for (i=imark; i<nzB; i++) v_p[nzA+i] = vworkB[i]; 1846 } 1847 if (idx) { 1848 *idx = idx_p = mat->rowindices; 1849 if (imark > -1) { 1850 for (i=0; i<imark; i++) { 1851 idx_p[i] = cmap[cworkB[i]]; 1852 } 1853 } else { 1854 for (i=0; i<nzB; i++) { 1855 if (cmap[cworkB[i]] < cstart) idx_p[i] = cmap[cworkB[i]]; 1856 else break; 1857 } 1858 imark = i; 1859 } 1860 for (i=0; i<nzA; i++) idx_p[imark+i] = cstart + cworkA[i]; 1861 for (i=imark; i<nzB; i++) idx_p[nzA+i] = cmap[cworkB[i]]; 1862 } 1863 } else { 1864 if (idx) *idx = 0; 1865 if (v) *v = 0; 1866 } 1867 } 1868 *nz = nztot; 1869 ierr = (*mat->A->ops->restorerow)(mat->A,lrow,&nzA,pcA,pvA);CHKERRQ(ierr); 1870 ierr = (*mat->B->ops->restorerow)(mat->B,lrow,&nzB,pcB,pvB);CHKERRQ(ierr); 1871 PetscFunctionReturn(0); 1872 } 1873 1874 #undef __FUNCT__ 1875 #define __FUNCT__ "MatRestoreRow_MPIAIJ" 1876 PetscErrorCode MatRestoreRow_MPIAIJ(Mat mat,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v) 1877 { 1878 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 1879 1880 PetscFunctionBegin; 1881 if (!aij->getrowactive) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"MatGetRow() must be called first"); 1882 aij->getrowactive = PETSC_FALSE; 1883 PetscFunctionReturn(0); 1884 } 1885 1886 #undef __FUNCT__ 1887 #define __FUNCT__ "MatNorm_MPIAIJ" 1888 PetscErrorCode MatNorm_MPIAIJ(Mat mat,NormType type,PetscReal *norm) 1889 { 1890 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 1891 Mat_SeqAIJ *amat = (Mat_SeqAIJ*)aij->A->data,*bmat = (Mat_SeqAIJ*)aij->B->data; 1892 PetscErrorCode ierr; 1893 PetscInt i,j,cstart = mat->cmap->rstart; 1894 PetscReal sum = 0.0; 1895 MatScalar *v; 1896 1897 PetscFunctionBegin; 1898 if (aij->size == 1) { 1899 ierr = MatNorm(aij->A,type,norm);CHKERRQ(ierr); 1900 } else { 1901 if (type == NORM_FROBENIUS) { 1902 v = amat->a; 1903 for (i=0; i<amat->nz; i++) { 1904 sum += PetscRealPart(PetscConj(*v)*(*v)); v++; 1905 } 1906 v = bmat->a; 1907 for (i=0; i<bmat->nz; i++) { 1908 sum += PetscRealPart(PetscConj(*v)*(*v)); v++; 1909 } 1910 ierr = MPI_Allreduce(&sum,norm,1,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1911 *norm = PetscSqrtReal(*norm); 1912 } else if (type == NORM_1) { /* max column norm */ 1913 PetscReal *tmp,*tmp2; 1914 PetscInt *jj,*garray = aij->garray; 1915 ierr = PetscCalloc1((mat->cmap->N+1),&tmp);CHKERRQ(ierr); 1916 ierr = PetscMalloc1((mat->cmap->N+1),&tmp2);CHKERRQ(ierr); 1917 *norm = 0.0; 1918 v = amat->a; jj = amat->j; 1919 for (j=0; j<amat->nz; j++) { 1920 tmp[cstart + *jj++] += PetscAbsScalar(*v); v++; 1921 } 1922 v = bmat->a; jj = bmat->j; 1923 for (j=0; j<bmat->nz; j++) { 1924 tmp[garray[*jj++]] += PetscAbsScalar(*v); v++; 1925 } 1926 ierr = MPI_Allreduce(tmp,tmp2,mat->cmap->N,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1927 for (j=0; j<mat->cmap->N; j++) { 1928 if (tmp2[j] > *norm) *norm = tmp2[j]; 1929 } 1930 ierr = PetscFree(tmp);CHKERRQ(ierr); 1931 ierr = PetscFree(tmp2);CHKERRQ(ierr); 1932 } else if (type == NORM_INFINITY) { /* max row norm */ 1933 PetscReal ntemp = 0.0; 1934 for (j=0; j<aij->A->rmap->n; j++) { 1935 v = amat->a + amat->i[j]; 1936 sum = 0.0; 1937 for (i=0; i<amat->i[j+1]-amat->i[j]; i++) { 1938 sum += PetscAbsScalar(*v); v++; 1939 } 1940 v = bmat->a + bmat->i[j]; 1941 for (i=0; i<bmat->i[j+1]-bmat->i[j]; i++) { 1942 sum += PetscAbsScalar(*v); v++; 1943 } 1944 if (sum > ntemp) ntemp = sum; 1945 } 1946 ierr = MPI_Allreduce(&ntemp,norm,1,MPIU_REAL,MPIU_MAX,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1947 } else SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"No support for two norm"); 1948 } 1949 PetscFunctionReturn(0); 1950 } 1951 1952 #undef __FUNCT__ 1953 #define __FUNCT__ "MatTranspose_MPIAIJ" 1954 PetscErrorCode MatTranspose_MPIAIJ(Mat A,MatReuse reuse,Mat *matout) 1955 { 1956 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 1957 Mat_SeqAIJ *Aloc=(Mat_SeqAIJ*)a->A->data,*Bloc=(Mat_SeqAIJ*)a->B->data; 1958 PetscErrorCode ierr; 1959 PetscInt M = A->rmap->N,N = A->cmap->N,ma,na,mb,nb,*ai,*aj,*bi,*bj,row,*cols,*cols_tmp,i; 1960 PetscInt cstart = A->cmap->rstart,ncol; 1961 Mat B; 1962 MatScalar *array; 1963 1964 PetscFunctionBegin; 1965 if (reuse == MAT_REUSE_MATRIX && A == *matout && M != N) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_SIZ,"Square matrix only for in-place"); 1966 1967 ma = A->rmap->n; na = A->cmap->n; mb = a->B->rmap->n; nb = a->B->cmap->n; 1968 ai = Aloc->i; aj = Aloc->j; 1969 bi = Bloc->i; bj = Bloc->j; 1970 if (reuse == MAT_INITIAL_MATRIX || *matout == A) { 1971 PetscInt *d_nnz,*g_nnz,*o_nnz; 1972 PetscSFNode *oloc; 1973 PETSC_UNUSED PetscSF sf; 1974 1975 ierr = PetscMalloc4(na,&d_nnz,na,&o_nnz,nb,&g_nnz,nb,&oloc);CHKERRQ(ierr); 1976 /* compute d_nnz for preallocation */ 1977 ierr = PetscMemzero(d_nnz,na*sizeof(PetscInt));CHKERRQ(ierr); 1978 for (i=0; i<ai[ma]; i++) { 1979 d_nnz[aj[i]]++; 1980 aj[i] += cstart; /* global col index to be used by MatSetValues() */ 1981 } 1982 /* compute local off-diagonal contributions */ 1983 ierr = PetscMemzero(g_nnz,nb*sizeof(PetscInt));CHKERRQ(ierr); 1984 for (i=0; i<bi[ma]; i++) g_nnz[bj[i]]++; 1985 /* map those to global */ 1986 ierr = PetscSFCreate(PetscObjectComm((PetscObject)A),&sf);CHKERRQ(ierr); 1987 ierr = PetscSFSetGraphLayout(sf,A->cmap,nb,NULL,PETSC_USE_POINTER,a->garray);CHKERRQ(ierr); 1988 ierr = PetscSFSetFromOptions(sf);CHKERRQ(ierr); 1989 ierr = PetscMemzero(o_nnz,na*sizeof(PetscInt));CHKERRQ(ierr); 1990 ierr = PetscSFReduceBegin(sf,MPIU_INT,g_nnz,o_nnz,MPIU_SUM);CHKERRQ(ierr); 1991 ierr = PetscSFReduceEnd(sf,MPIU_INT,g_nnz,o_nnz,MPIU_SUM);CHKERRQ(ierr); 1992 ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 1993 1994 ierr = MatCreate(PetscObjectComm((PetscObject)A),&B);CHKERRQ(ierr); 1995 ierr = MatSetSizes(B,A->cmap->n,A->rmap->n,N,M);CHKERRQ(ierr); 1996 ierr = MatSetBlockSizes(B,PetscAbs(A->cmap->bs),PetscAbs(A->rmap->bs));CHKERRQ(ierr); 1997 ierr = MatSetType(B,((PetscObject)A)->type_name);CHKERRQ(ierr); 1998 ierr = MatMPIAIJSetPreallocation(B,0,d_nnz,0,o_nnz);CHKERRQ(ierr); 1999 ierr = PetscFree4(d_nnz,o_nnz,g_nnz,oloc);CHKERRQ(ierr); 2000 } else { 2001 B = *matout; 2002 ierr = MatSetOption(B,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_TRUE);CHKERRQ(ierr); 2003 for (i=0; i<ai[ma]; i++) aj[i] += cstart; /* global col index to be used by MatSetValues() */ 2004 } 2005 2006 /* copy over the A part */ 2007 array = Aloc->a; 2008 row = A->rmap->rstart; 2009 for (i=0; i<ma; i++) { 2010 ncol = ai[i+1]-ai[i]; 2011 ierr = MatSetValues(B,ncol,aj,1,&row,array,INSERT_VALUES);CHKERRQ(ierr); 2012 row++; 2013 array += ncol; aj += ncol; 2014 } 2015 aj = Aloc->j; 2016 for (i=0; i<ai[ma]; i++) aj[i] -= cstart; /* resume local col index */ 2017 2018 /* copy over the B part */ 2019 ierr = PetscCalloc1(bi[mb],&cols);CHKERRQ(ierr); 2020 array = Bloc->a; 2021 row = A->rmap->rstart; 2022 for (i=0; i<bi[mb]; i++) cols[i] = a->garray[bj[i]]; 2023 cols_tmp = cols; 2024 for (i=0; i<mb; i++) { 2025 ncol = bi[i+1]-bi[i]; 2026 ierr = MatSetValues(B,ncol,cols_tmp,1,&row,array,INSERT_VALUES);CHKERRQ(ierr); 2027 row++; 2028 array += ncol; cols_tmp += ncol; 2029 } 2030 ierr = PetscFree(cols);CHKERRQ(ierr); 2031 2032 ierr = MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2033 ierr = MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2034 if (reuse == MAT_INITIAL_MATRIX || *matout != A) { 2035 *matout = B; 2036 } else { 2037 ierr = MatHeaderMerge(A,B);CHKERRQ(ierr); 2038 } 2039 PetscFunctionReturn(0); 2040 } 2041 2042 #undef __FUNCT__ 2043 #define __FUNCT__ "MatDiagonalScale_MPIAIJ" 2044 PetscErrorCode MatDiagonalScale_MPIAIJ(Mat mat,Vec ll,Vec rr) 2045 { 2046 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 2047 Mat a = aij->A,b = aij->B; 2048 PetscErrorCode ierr; 2049 PetscInt s1,s2,s3; 2050 2051 PetscFunctionBegin; 2052 ierr = MatGetLocalSize(mat,&s2,&s3);CHKERRQ(ierr); 2053 if (rr) { 2054 ierr = VecGetLocalSize(rr,&s1);CHKERRQ(ierr); 2055 if (s1!=s3) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"right vector non-conforming local size"); 2056 /* Overlap communication with computation. */ 2057 ierr = VecScatterBegin(aij->Mvctx,rr,aij->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2058 } 2059 if (ll) { 2060 ierr = VecGetLocalSize(ll,&s1);CHKERRQ(ierr); 2061 if (s1!=s2) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"left vector non-conforming local size"); 2062 ierr = (*b->ops->diagonalscale)(b,ll,0);CHKERRQ(ierr); 2063 } 2064 /* scale the diagonal block */ 2065 ierr = (*a->ops->diagonalscale)(a,ll,rr);CHKERRQ(ierr); 2066 2067 if (rr) { 2068 /* Do a scatter end and then right scale the off-diagonal block */ 2069 ierr = VecScatterEnd(aij->Mvctx,rr,aij->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2070 ierr = (*b->ops->diagonalscale)(b,0,aij->lvec);CHKERRQ(ierr); 2071 } 2072 PetscFunctionReturn(0); 2073 } 2074 2075 #undef __FUNCT__ 2076 #define __FUNCT__ "MatSetUnfactored_MPIAIJ" 2077 PetscErrorCode MatSetUnfactored_MPIAIJ(Mat A) 2078 { 2079 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 2080 PetscErrorCode ierr; 2081 2082 PetscFunctionBegin; 2083 ierr = MatSetUnfactored(a->A);CHKERRQ(ierr); 2084 PetscFunctionReturn(0); 2085 } 2086 2087 #undef __FUNCT__ 2088 #define __FUNCT__ "MatEqual_MPIAIJ" 2089 PetscErrorCode MatEqual_MPIAIJ(Mat A,Mat B,PetscBool *flag) 2090 { 2091 Mat_MPIAIJ *matB = (Mat_MPIAIJ*)B->data,*matA = (Mat_MPIAIJ*)A->data; 2092 Mat a,b,c,d; 2093 PetscBool flg; 2094 PetscErrorCode ierr; 2095 2096 PetscFunctionBegin; 2097 a = matA->A; b = matA->B; 2098 c = matB->A; d = matB->B; 2099 2100 ierr = MatEqual(a,c,&flg);CHKERRQ(ierr); 2101 if (flg) { 2102 ierr = MatEqual(b,d,&flg);CHKERRQ(ierr); 2103 } 2104 ierr = MPI_Allreduce(&flg,flag,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)A));CHKERRQ(ierr); 2105 PetscFunctionReturn(0); 2106 } 2107 2108 #undef __FUNCT__ 2109 #define __FUNCT__ "MatCopy_MPIAIJ" 2110 PetscErrorCode MatCopy_MPIAIJ(Mat A,Mat B,MatStructure str) 2111 { 2112 PetscErrorCode ierr; 2113 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 2114 Mat_MPIAIJ *b = (Mat_MPIAIJ*)B->data; 2115 2116 PetscFunctionBegin; 2117 /* If the two matrices don't have the same copy implementation, they aren't compatible for fast copy. */ 2118 if ((str != SAME_NONZERO_PATTERN) || (A->ops->copy != B->ops->copy)) { 2119 /* because of the column compression in the off-processor part of the matrix a->B, 2120 the number of columns in a->B and b->B may be different, hence we cannot call 2121 the MatCopy() directly on the two parts. If need be, we can provide a more 2122 efficient copy than the MatCopy_Basic() by first uncompressing the a->B matrices 2123 then copying the submatrices */ 2124 ierr = MatCopy_Basic(A,B,str);CHKERRQ(ierr); 2125 } else { 2126 ierr = MatCopy(a->A,b->A,str);CHKERRQ(ierr); 2127 ierr = MatCopy(a->B,b->B,str);CHKERRQ(ierr); 2128 } 2129 PetscFunctionReturn(0); 2130 } 2131 2132 #undef __FUNCT__ 2133 #define __FUNCT__ "MatSetUp_MPIAIJ" 2134 PetscErrorCode MatSetUp_MPIAIJ(Mat A) 2135 { 2136 PetscErrorCode ierr; 2137 2138 PetscFunctionBegin; 2139 ierr = MatMPIAIJSetPreallocation(A,PETSC_DEFAULT,0,PETSC_DEFAULT,0);CHKERRQ(ierr); 2140 PetscFunctionReturn(0); 2141 } 2142 2143 /* 2144 Computes the number of nonzeros per row needed for preallocation when X and Y 2145 have different nonzero structure. 2146 */ 2147 #undef __FUNCT__ 2148 #define __FUNCT__ "MatAXPYGetPreallocation_MPIX_private" 2149 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) 2150 { 2151 PetscInt i,j,k,nzx,nzy; 2152 2153 PetscFunctionBegin; 2154 /* Set the number of nonzeros in the new matrix */ 2155 for (i=0; i<m; i++) { 2156 const PetscInt *xjj = xj+xi[i],*yjj = yj+yi[i]; 2157 nzx = xi[i+1] - xi[i]; 2158 nzy = yi[i+1] - yi[i]; 2159 nnz[i] = 0; 2160 for (j=0,k=0; j<nzx; j++) { /* Point in X */ 2161 for (; k<nzy && yltog[yjj[k]]<xltog[xjj[j]]; k++) nnz[i]++; /* Catch up to X */ 2162 if (k<nzy && yltog[yjj[k]]==xltog[xjj[j]]) k++; /* Skip duplicate */ 2163 nnz[i]++; 2164 } 2165 for (; k<nzy; k++) nnz[i]++; 2166 } 2167 PetscFunctionReturn(0); 2168 } 2169 2170 /* This is the same as MatAXPYGetPreallocation_SeqAIJ, except that the local-to-global map is provided */ 2171 #undef __FUNCT__ 2172 #define __FUNCT__ "MatAXPYGetPreallocation_MPIAIJ" 2173 static PetscErrorCode MatAXPYGetPreallocation_MPIAIJ(Mat Y,const PetscInt *yltog,Mat X,const PetscInt *xltog,PetscInt *nnz) 2174 { 2175 PetscErrorCode ierr; 2176 PetscInt m = Y->rmap->N; 2177 Mat_SeqAIJ *x = (Mat_SeqAIJ*)X->data; 2178 Mat_SeqAIJ *y = (Mat_SeqAIJ*)Y->data; 2179 2180 PetscFunctionBegin; 2181 ierr = MatAXPYGetPreallocation_MPIX_private(m,x->i,x->j,xltog,y->i,y->j,yltog,nnz);CHKERRQ(ierr); 2182 PetscFunctionReturn(0); 2183 } 2184 2185 #undef __FUNCT__ 2186 #define __FUNCT__ "MatAXPY_MPIAIJ" 2187 PetscErrorCode MatAXPY_MPIAIJ(Mat Y,PetscScalar a,Mat X,MatStructure str) 2188 { 2189 PetscErrorCode ierr; 2190 PetscInt i; 2191 Mat_MPIAIJ *xx = (Mat_MPIAIJ*)X->data,*yy = (Mat_MPIAIJ*)Y->data; 2192 PetscBLASInt bnz,one=1; 2193 Mat_SeqAIJ *x,*y; 2194 2195 PetscFunctionBegin; 2196 if (str == SAME_NONZERO_PATTERN) { 2197 PetscScalar alpha = a; 2198 x = (Mat_SeqAIJ*)xx->A->data; 2199 ierr = PetscBLASIntCast(x->nz,&bnz);CHKERRQ(ierr); 2200 y = (Mat_SeqAIJ*)yy->A->data; 2201 PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&bnz,&alpha,x->a,&one,y->a,&one)); 2202 x = (Mat_SeqAIJ*)xx->B->data; 2203 y = (Mat_SeqAIJ*)yy->B->data; 2204 ierr = PetscBLASIntCast(x->nz,&bnz);CHKERRQ(ierr); 2205 PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&bnz,&alpha,x->a,&one,y->a,&one)); 2206 ierr = PetscObjectStateIncrease((PetscObject)Y);CHKERRQ(ierr); 2207 } else if (str == SUBSET_NONZERO_PATTERN) { 2208 ierr = MatAXPY_SeqAIJ(yy->A,a,xx->A,str);CHKERRQ(ierr); 2209 2210 x = (Mat_SeqAIJ*)xx->B->data; 2211 y = (Mat_SeqAIJ*)yy->B->data; 2212 if (y->xtoy && y->XtoY != xx->B) { 2213 ierr = PetscFree(y->xtoy);CHKERRQ(ierr); 2214 ierr = MatDestroy(&y->XtoY);CHKERRQ(ierr); 2215 } 2216 if (!y->xtoy) { /* get xtoy */ 2217 ierr = MatAXPYGetxtoy_Private(xx->B->rmap->n,x->i,x->j,xx->garray,y->i,y->j,yy->garray,&y->xtoy);CHKERRQ(ierr); 2218 y->XtoY = xx->B; 2219 ierr = PetscObjectReference((PetscObject)xx->B);CHKERRQ(ierr); 2220 } 2221 for (i=0; i<x->nz; i++) y->a[y->xtoy[i]] += a*(x->a[i]); 2222 ierr = PetscObjectStateIncrease((PetscObject)Y);CHKERRQ(ierr); 2223 } else { 2224 Mat B; 2225 PetscInt *nnz_d,*nnz_o; 2226 ierr = PetscMalloc1(yy->A->rmap->N,&nnz_d);CHKERRQ(ierr); 2227 ierr = PetscMalloc1(yy->B->rmap->N,&nnz_o);CHKERRQ(ierr); 2228 ierr = MatCreate(PetscObjectComm((PetscObject)Y),&B);CHKERRQ(ierr); 2229 ierr = PetscObjectSetName((PetscObject)B,((PetscObject)Y)->name);CHKERRQ(ierr); 2230 ierr = MatSetSizes(B,Y->rmap->n,Y->cmap->n,Y->rmap->N,Y->cmap->N);CHKERRQ(ierr); 2231 ierr = MatSetBlockSizesFromMats(B,Y,Y);CHKERRQ(ierr); 2232 ierr = MatSetType(B,MATMPIAIJ);CHKERRQ(ierr); 2233 ierr = MatAXPYGetPreallocation_SeqAIJ(yy->A,xx->A,nnz_d);CHKERRQ(ierr); 2234 ierr = MatAXPYGetPreallocation_MPIAIJ(yy->B,yy->garray,xx->B,xx->garray,nnz_o);CHKERRQ(ierr); 2235 ierr = MatMPIAIJSetPreallocation(B,0,nnz_d,0,nnz_o);CHKERRQ(ierr); 2236 ierr = MatAXPY_BasicWithPreallocation(B,Y,a,X,str);CHKERRQ(ierr); 2237 ierr = MatHeaderReplace(Y,B);CHKERRQ(ierr); 2238 ierr = PetscFree(nnz_d);CHKERRQ(ierr); 2239 ierr = PetscFree(nnz_o);CHKERRQ(ierr); 2240 } 2241 PetscFunctionReturn(0); 2242 } 2243 2244 extern PetscErrorCode MatConjugate_SeqAIJ(Mat); 2245 2246 #undef __FUNCT__ 2247 #define __FUNCT__ "MatConjugate_MPIAIJ" 2248 PetscErrorCode MatConjugate_MPIAIJ(Mat mat) 2249 { 2250 #if defined(PETSC_USE_COMPLEX) 2251 PetscErrorCode ierr; 2252 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 2253 2254 PetscFunctionBegin; 2255 ierr = MatConjugate_SeqAIJ(aij->A);CHKERRQ(ierr); 2256 ierr = MatConjugate_SeqAIJ(aij->B);CHKERRQ(ierr); 2257 #else 2258 PetscFunctionBegin; 2259 #endif 2260 PetscFunctionReturn(0); 2261 } 2262 2263 #undef __FUNCT__ 2264 #define __FUNCT__ "MatRealPart_MPIAIJ" 2265 PetscErrorCode MatRealPart_MPIAIJ(Mat A) 2266 { 2267 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 2268 PetscErrorCode ierr; 2269 2270 PetscFunctionBegin; 2271 ierr = MatRealPart(a->A);CHKERRQ(ierr); 2272 ierr = MatRealPart(a->B);CHKERRQ(ierr); 2273 PetscFunctionReturn(0); 2274 } 2275 2276 #undef __FUNCT__ 2277 #define __FUNCT__ "MatImaginaryPart_MPIAIJ" 2278 PetscErrorCode MatImaginaryPart_MPIAIJ(Mat A) 2279 { 2280 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 2281 PetscErrorCode ierr; 2282 2283 PetscFunctionBegin; 2284 ierr = MatImaginaryPart(a->A);CHKERRQ(ierr); 2285 ierr = MatImaginaryPart(a->B);CHKERRQ(ierr); 2286 PetscFunctionReturn(0); 2287 } 2288 2289 #if defined(PETSC_HAVE_PBGL) 2290 2291 #include <boost/parallel/mpi/bsp_process_group.hpp> 2292 #include <boost/graph/distributed/ilu_default_graph.hpp> 2293 #include <boost/graph/distributed/ilu_0_block.hpp> 2294 #include <boost/graph/distributed/ilu_preconditioner.hpp> 2295 #include <boost/graph/distributed/petsc/interface.hpp> 2296 #include <boost/multi_array.hpp> 2297 #include <boost/parallel/distributed_property_map->hpp> 2298 2299 #undef __FUNCT__ 2300 #define __FUNCT__ "MatILUFactorSymbolic_MPIAIJ" 2301 /* 2302 This uses the parallel ILU factorization of Peter Gottschling <pgottsch@osl.iu.edu> 2303 */ 2304 PetscErrorCode MatILUFactorSymbolic_MPIAIJ(Mat fact,Mat A, IS isrow, IS iscol, const MatFactorInfo *info) 2305 { 2306 namespace petsc = boost::distributed::petsc; 2307 2308 namespace graph_dist = boost::graph::distributed; 2309 using boost::graph::distributed::ilu_default::process_group_type; 2310 using boost::graph::ilu_permuted; 2311 2312 PetscBool row_identity, col_identity; 2313 PetscContainer c; 2314 PetscInt m, n, M, N; 2315 PetscErrorCode ierr; 2316 2317 PetscFunctionBegin; 2318 if (info->levels != 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Only levels = 0 supported for parallel ilu"); 2319 ierr = ISIdentity(isrow, &row_identity);CHKERRQ(ierr); 2320 ierr = ISIdentity(iscol, &col_identity);CHKERRQ(ierr); 2321 if (!row_identity || !col_identity) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Row and column permutations must be identity for parallel ILU"); 2322 2323 process_group_type pg; 2324 typedef graph_dist::ilu_default::ilu_level_graph_type lgraph_type; 2325 lgraph_type *lgraph_p = new lgraph_type(petsc::num_global_vertices(A), pg, petsc::matrix_distribution(A, pg)); 2326 lgraph_type& level_graph = *lgraph_p; 2327 graph_dist::ilu_default::graph_type& graph(level_graph.graph); 2328 2329 petsc::read_matrix(A, graph, get(boost::edge_weight, graph)); 2330 ilu_permuted(level_graph); 2331 2332 /* put together the new matrix */ 2333 ierr = MatCreate(PetscObjectComm((PetscObject)A), fact);CHKERRQ(ierr); 2334 ierr = MatGetLocalSize(A, &m, &n);CHKERRQ(ierr); 2335 ierr = MatGetSize(A, &M, &N);CHKERRQ(ierr); 2336 ierr = MatSetSizes(fact, m, n, M, N);CHKERRQ(ierr); 2337 ierr = MatSetBlockSizesFromMats(fact,A,A);CHKERRQ(ierr); 2338 ierr = MatSetType(fact, ((PetscObject)A)->type_name);CHKERRQ(ierr); 2339 ierr = MatAssemblyBegin(fact, MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2340 ierr = MatAssemblyEnd(fact, MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2341 2342 ierr = PetscContainerCreate(PetscObjectComm((PetscObject)A), &c); 2343 ierr = PetscContainerSetPointer(c, lgraph_p); 2344 ierr = PetscObjectCompose((PetscObject) (fact), "graph", (PetscObject) c); 2345 ierr = PetscContainerDestroy(&c); 2346 PetscFunctionReturn(0); 2347 } 2348 2349 #undef __FUNCT__ 2350 #define __FUNCT__ "MatLUFactorNumeric_MPIAIJ" 2351 PetscErrorCode MatLUFactorNumeric_MPIAIJ(Mat B,Mat A, const MatFactorInfo *info) 2352 { 2353 PetscFunctionBegin; 2354 PetscFunctionReturn(0); 2355 } 2356 2357 #undef __FUNCT__ 2358 #define __FUNCT__ "MatSolve_MPIAIJ" 2359 /* 2360 This uses the parallel ILU factorization of Peter Gottschling <pgottsch@osl.iu.edu> 2361 */ 2362 PetscErrorCode MatSolve_MPIAIJ(Mat A, Vec b, Vec x) 2363 { 2364 namespace graph_dist = boost::graph::distributed; 2365 2366 typedef graph_dist::ilu_default::ilu_level_graph_type lgraph_type; 2367 lgraph_type *lgraph_p; 2368 PetscContainer c; 2369 PetscErrorCode ierr; 2370 2371 PetscFunctionBegin; 2372 ierr = PetscObjectQuery((PetscObject) A, "graph", (PetscObject*) &c);CHKERRQ(ierr); 2373 ierr = PetscContainerGetPointer(c, (void**) &lgraph_p);CHKERRQ(ierr); 2374 ierr = VecCopy(b, x);CHKERRQ(ierr); 2375 2376 PetscScalar *array_x; 2377 ierr = VecGetArray(x, &array_x);CHKERRQ(ierr); 2378 PetscInt sx; 2379 ierr = VecGetSize(x, &sx);CHKERRQ(ierr); 2380 2381 PetscScalar *array_b; 2382 ierr = VecGetArray(b, &array_b);CHKERRQ(ierr); 2383 PetscInt sb; 2384 ierr = VecGetSize(b, &sb);CHKERRQ(ierr); 2385 2386 lgraph_type& level_graph = *lgraph_p; 2387 graph_dist::ilu_default::graph_type& graph(level_graph.graph); 2388 2389 typedef boost::multi_array_ref<PetscScalar, 1> array_ref_type; 2390 array_ref_type ref_b(array_b, boost::extents[num_vertices(graph)]); 2391 array_ref_type ref_x(array_x, boost::extents[num_vertices(graph)]); 2392 2393 typedef boost::iterator_property_map<array_ref_type::iterator, 2394 boost::property_map<graph_dist::ilu_default::graph_type, boost::vertex_index_t>::type> gvector_type; 2395 gvector_type vector_b(ref_b.begin(), get(boost::vertex_index, graph)); 2396 gvector_type vector_x(ref_x.begin(), get(boost::vertex_index, graph)); 2397 2398 ilu_set_solve(*lgraph_p, vector_b, vector_x); 2399 PetscFunctionReturn(0); 2400 } 2401 #endif 2402 2403 #undef __FUNCT__ 2404 #define __FUNCT__ "MatGetRowMaxAbs_MPIAIJ" 2405 PetscErrorCode MatGetRowMaxAbs_MPIAIJ(Mat A, Vec v, PetscInt idx[]) 2406 { 2407 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 2408 PetscErrorCode ierr; 2409 PetscInt i,*idxb = 0; 2410 PetscScalar *va,*vb; 2411 Vec vtmp; 2412 2413 PetscFunctionBegin; 2414 ierr = MatGetRowMaxAbs(a->A,v,idx);CHKERRQ(ierr); 2415 ierr = VecGetArray(v,&va);CHKERRQ(ierr); 2416 if (idx) { 2417 for (i=0; i<A->rmap->n; i++) { 2418 if (PetscAbsScalar(va[i])) idx[i] += A->cmap->rstart; 2419 } 2420 } 2421 2422 ierr = VecCreateSeq(PETSC_COMM_SELF,A->rmap->n,&vtmp);CHKERRQ(ierr); 2423 if (idx) { 2424 ierr = PetscMalloc1(A->rmap->n,&idxb);CHKERRQ(ierr); 2425 } 2426 ierr = MatGetRowMaxAbs(a->B,vtmp,idxb);CHKERRQ(ierr); 2427 ierr = VecGetArray(vtmp,&vb);CHKERRQ(ierr); 2428 2429 for (i=0; i<A->rmap->n; i++) { 2430 if (PetscAbsScalar(va[i]) < PetscAbsScalar(vb[i])) { 2431 va[i] = vb[i]; 2432 if (idx) idx[i] = a->garray[idxb[i]]; 2433 } 2434 } 2435 2436 ierr = VecRestoreArray(v,&va);CHKERRQ(ierr); 2437 ierr = VecRestoreArray(vtmp,&vb);CHKERRQ(ierr); 2438 ierr = PetscFree(idxb);CHKERRQ(ierr); 2439 ierr = VecDestroy(&vtmp);CHKERRQ(ierr); 2440 PetscFunctionReturn(0); 2441 } 2442 2443 #undef __FUNCT__ 2444 #define __FUNCT__ "MatGetRowMinAbs_MPIAIJ" 2445 PetscErrorCode MatGetRowMinAbs_MPIAIJ(Mat A, Vec v, PetscInt idx[]) 2446 { 2447 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 2448 PetscErrorCode ierr; 2449 PetscInt i,*idxb = 0; 2450 PetscScalar *va,*vb; 2451 Vec vtmp; 2452 2453 PetscFunctionBegin; 2454 ierr = MatGetRowMinAbs(a->A,v,idx);CHKERRQ(ierr); 2455 ierr = VecGetArray(v,&va);CHKERRQ(ierr); 2456 if (idx) { 2457 for (i=0; i<A->cmap->n; i++) { 2458 if (PetscAbsScalar(va[i])) idx[i] += A->cmap->rstart; 2459 } 2460 } 2461 2462 ierr = VecCreateSeq(PETSC_COMM_SELF,A->rmap->n,&vtmp);CHKERRQ(ierr); 2463 if (idx) { 2464 ierr = PetscMalloc1(A->rmap->n,&idxb);CHKERRQ(ierr); 2465 } 2466 ierr = MatGetRowMinAbs(a->B,vtmp,idxb);CHKERRQ(ierr); 2467 ierr = VecGetArray(vtmp,&vb);CHKERRQ(ierr); 2468 2469 for (i=0; i<A->rmap->n; i++) { 2470 if (PetscAbsScalar(va[i]) > PetscAbsScalar(vb[i])) { 2471 va[i] = vb[i]; 2472 if (idx) idx[i] = a->garray[idxb[i]]; 2473 } 2474 } 2475 2476 ierr = VecRestoreArray(v,&va);CHKERRQ(ierr); 2477 ierr = VecRestoreArray(vtmp,&vb);CHKERRQ(ierr); 2478 ierr = PetscFree(idxb);CHKERRQ(ierr); 2479 ierr = VecDestroy(&vtmp);CHKERRQ(ierr); 2480 PetscFunctionReturn(0); 2481 } 2482 2483 #undef __FUNCT__ 2484 #define __FUNCT__ "MatGetRowMin_MPIAIJ" 2485 PetscErrorCode MatGetRowMin_MPIAIJ(Mat A, Vec v, PetscInt idx[]) 2486 { 2487 Mat_MPIAIJ *mat = (Mat_MPIAIJ*) A->data; 2488 PetscInt n = A->rmap->n; 2489 PetscInt cstart = A->cmap->rstart; 2490 PetscInt *cmap = mat->garray; 2491 PetscInt *diagIdx, *offdiagIdx; 2492 Vec diagV, offdiagV; 2493 PetscScalar *a, *diagA, *offdiagA; 2494 PetscInt r; 2495 PetscErrorCode ierr; 2496 2497 PetscFunctionBegin; 2498 ierr = PetscMalloc2(n,&diagIdx,n,&offdiagIdx);CHKERRQ(ierr); 2499 ierr = VecCreateSeq(PetscObjectComm((PetscObject)A), n, &diagV);CHKERRQ(ierr); 2500 ierr = VecCreateSeq(PetscObjectComm((PetscObject)A), n, &offdiagV);CHKERRQ(ierr); 2501 ierr = MatGetRowMin(mat->A, diagV, diagIdx);CHKERRQ(ierr); 2502 ierr = MatGetRowMin(mat->B, offdiagV, offdiagIdx);CHKERRQ(ierr); 2503 ierr = VecGetArray(v, &a);CHKERRQ(ierr); 2504 ierr = VecGetArray(diagV, &diagA);CHKERRQ(ierr); 2505 ierr = VecGetArray(offdiagV, &offdiagA);CHKERRQ(ierr); 2506 for (r = 0; r < n; ++r) { 2507 if (PetscAbsScalar(diagA[r]) <= PetscAbsScalar(offdiagA[r])) { 2508 a[r] = diagA[r]; 2509 idx[r] = cstart + diagIdx[r]; 2510 } else { 2511 a[r] = offdiagA[r]; 2512 idx[r] = cmap[offdiagIdx[r]]; 2513 } 2514 } 2515 ierr = VecRestoreArray(v, &a);CHKERRQ(ierr); 2516 ierr = VecRestoreArray(diagV, &diagA);CHKERRQ(ierr); 2517 ierr = VecRestoreArray(offdiagV, &offdiagA);CHKERRQ(ierr); 2518 ierr = VecDestroy(&diagV);CHKERRQ(ierr); 2519 ierr = VecDestroy(&offdiagV);CHKERRQ(ierr); 2520 ierr = PetscFree2(diagIdx, offdiagIdx);CHKERRQ(ierr); 2521 PetscFunctionReturn(0); 2522 } 2523 2524 #undef __FUNCT__ 2525 #define __FUNCT__ "MatGetRowMax_MPIAIJ" 2526 PetscErrorCode MatGetRowMax_MPIAIJ(Mat A, Vec v, PetscInt idx[]) 2527 { 2528 Mat_MPIAIJ *mat = (Mat_MPIAIJ*) A->data; 2529 PetscInt n = A->rmap->n; 2530 PetscInt cstart = A->cmap->rstart; 2531 PetscInt *cmap = mat->garray; 2532 PetscInt *diagIdx, *offdiagIdx; 2533 Vec diagV, offdiagV; 2534 PetscScalar *a, *diagA, *offdiagA; 2535 PetscInt r; 2536 PetscErrorCode ierr; 2537 2538 PetscFunctionBegin; 2539 ierr = PetscMalloc2(n,&diagIdx,n,&offdiagIdx);CHKERRQ(ierr); 2540 ierr = VecCreateSeq(PETSC_COMM_SELF, n, &diagV);CHKERRQ(ierr); 2541 ierr = VecCreateSeq(PETSC_COMM_SELF, n, &offdiagV);CHKERRQ(ierr); 2542 ierr = MatGetRowMax(mat->A, diagV, diagIdx);CHKERRQ(ierr); 2543 ierr = MatGetRowMax(mat->B, offdiagV, offdiagIdx);CHKERRQ(ierr); 2544 ierr = VecGetArray(v, &a);CHKERRQ(ierr); 2545 ierr = VecGetArray(diagV, &diagA);CHKERRQ(ierr); 2546 ierr = VecGetArray(offdiagV, &offdiagA);CHKERRQ(ierr); 2547 for (r = 0; r < n; ++r) { 2548 if (PetscAbsScalar(diagA[r]) >= PetscAbsScalar(offdiagA[r])) { 2549 a[r] = diagA[r]; 2550 idx[r] = cstart + diagIdx[r]; 2551 } else { 2552 a[r] = offdiagA[r]; 2553 idx[r] = cmap[offdiagIdx[r]]; 2554 } 2555 } 2556 ierr = VecRestoreArray(v, &a);CHKERRQ(ierr); 2557 ierr = VecRestoreArray(diagV, &diagA);CHKERRQ(ierr); 2558 ierr = VecRestoreArray(offdiagV, &offdiagA);CHKERRQ(ierr); 2559 ierr = VecDestroy(&diagV);CHKERRQ(ierr); 2560 ierr = VecDestroy(&offdiagV);CHKERRQ(ierr); 2561 ierr = PetscFree2(diagIdx, offdiagIdx);CHKERRQ(ierr); 2562 PetscFunctionReturn(0); 2563 } 2564 2565 #undef __FUNCT__ 2566 #define __FUNCT__ "MatGetSeqNonzeroStructure_MPIAIJ" 2567 PetscErrorCode MatGetSeqNonzeroStructure_MPIAIJ(Mat mat,Mat *newmat) 2568 { 2569 PetscErrorCode ierr; 2570 Mat *dummy; 2571 2572 PetscFunctionBegin; 2573 ierr = MatGetSubMatrix_MPIAIJ_All(mat,MAT_DO_NOT_GET_VALUES,MAT_INITIAL_MATRIX,&dummy);CHKERRQ(ierr); 2574 *newmat = *dummy; 2575 ierr = PetscFree(dummy);CHKERRQ(ierr); 2576 PetscFunctionReturn(0); 2577 } 2578 2579 #undef __FUNCT__ 2580 #define __FUNCT__ "MatInvertBlockDiagonal_MPIAIJ" 2581 PetscErrorCode MatInvertBlockDiagonal_MPIAIJ(Mat A,const PetscScalar **values) 2582 { 2583 Mat_MPIAIJ *a = (Mat_MPIAIJ*) A->data; 2584 PetscErrorCode ierr; 2585 2586 PetscFunctionBegin; 2587 ierr = MatInvertBlockDiagonal(a->A,values);CHKERRQ(ierr); 2588 PetscFunctionReturn(0); 2589 } 2590 2591 #undef __FUNCT__ 2592 #define __FUNCT__ "MatSetRandom_MPIAIJ" 2593 static PetscErrorCode MatSetRandom_MPIAIJ(Mat x,PetscRandom rctx) 2594 { 2595 PetscErrorCode ierr; 2596 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)x->data; 2597 2598 PetscFunctionBegin; 2599 ierr = MatSetRandom(aij->A,rctx);CHKERRQ(ierr); 2600 ierr = MatSetRandom(aij->B,rctx);CHKERRQ(ierr); 2601 ierr = MatAssemblyBegin(x,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2602 ierr = MatAssemblyEnd(x,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2603 PetscFunctionReturn(0); 2604 } 2605 2606 /* -------------------------------------------------------------------*/ 2607 static struct _MatOps MatOps_Values = {MatSetValues_MPIAIJ, 2608 MatGetRow_MPIAIJ, 2609 MatRestoreRow_MPIAIJ, 2610 MatMult_MPIAIJ, 2611 /* 4*/ MatMultAdd_MPIAIJ, 2612 MatMultTranspose_MPIAIJ, 2613 MatMultTransposeAdd_MPIAIJ, 2614 #if defined(PETSC_HAVE_PBGL) 2615 MatSolve_MPIAIJ, 2616 #else 2617 0, 2618 #endif 2619 0, 2620 0, 2621 /*10*/ 0, 2622 0, 2623 0, 2624 MatSOR_MPIAIJ, 2625 MatTranspose_MPIAIJ, 2626 /*15*/ MatGetInfo_MPIAIJ, 2627 MatEqual_MPIAIJ, 2628 MatGetDiagonal_MPIAIJ, 2629 MatDiagonalScale_MPIAIJ, 2630 MatNorm_MPIAIJ, 2631 /*20*/ MatAssemblyBegin_MPIAIJ, 2632 MatAssemblyEnd_MPIAIJ, 2633 MatSetOption_MPIAIJ, 2634 MatZeroEntries_MPIAIJ, 2635 /*24*/ MatZeroRows_MPIAIJ, 2636 0, 2637 #if defined(PETSC_HAVE_PBGL) 2638 0, 2639 #else 2640 0, 2641 #endif 2642 0, 2643 0, 2644 /*29*/ MatSetUp_MPIAIJ, 2645 #if defined(PETSC_HAVE_PBGL) 2646 0, 2647 #else 2648 0, 2649 #endif 2650 0, 2651 0, 2652 0, 2653 /*34*/ MatDuplicate_MPIAIJ, 2654 0, 2655 0, 2656 0, 2657 0, 2658 /*39*/ MatAXPY_MPIAIJ, 2659 MatGetSubMatrices_MPIAIJ, 2660 MatIncreaseOverlap_MPIAIJ, 2661 MatGetValues_MPIAIJ, 2662 MatCopy_MPIAIJ, 2663 /*44*/ MatGetRowMax_MPIAIJ, 2664 MatScale_MPIAIJ, 2665 0, 2666 MatDiagonalSet_MPIAIJ, 2667 MatZeroRowsColumns_MPIAIJ, 2668 /*49*/ MatSetRandom_MPIAIJ, 2669 0, 2670 0, 2671 0, 2672 0, 2673 /*54*/ MatFDColoringCreate_MPIXAIJ, 2674 0, 2675 MatSetUnfactored_MPIAIJ, 2676 MatPermute_MPIAIJ, 2677 0, 2678 /*59*/ MatGetSubMatrix_MPIAIJ, 2679 MatDestroy_MPIAIJ, 2680 MatView_MPIAIJ, 2681 0, 2682 MatMatMatMult_MPIAIJ_MPIAIJ_MPIAIJ, 2683 /*64*/ MatMatMatMultSymbolic_MPIAIJ_MPIAIJ_MPIAIJ, 2684 MatMatMatMultNumeric_MPIAIJ_MPIAIJ_MPIAIJ, 2685 0, 2686 0, 2687 0, 2688 /*69*/ MatGetRowMaxAbs_MPIAIJ, 2689 MatGetRowMinAbs_MPIAIJ, 2690 0, 2691 MatSetColoring_MPIAIJ, 2692 0, 2693 MatSetValuesAdifor_MPIAIJ, 2694 /*75*/ MatFDColoringApply_AIJ, 2695 0, 2696 0, 2697 0, 2698 MatFindZeroDiagonals_MPIAIJ, 2699 /*80*/ 0, 2700 0, 2701 0, 2702 /*83*/ MatLoad_MPIAIJ, 2703 0, 2704 0, 2705 0, 2706 0, 2707 0, 2708 /*89*/ MatMatMult_MPIAIJ_MPIAIJ, 2709 MatMatMultSymbolic_MPIAIJ_MPIAIJ, 2710 MatMatMultNumeric_MPIAIJ_MPIAIJ, 2711 MatPtAP_MPIAIJ_MPIAIJ, 2712 MatPtAPSymbolic_MPIAIJ_MPIAIJ, 2713 /*94*/ MatPtAPNumeric_MPIAIJ_MPIAIJ, 2714 0, 2715 0, 2716 0, 2717 0, 2718 /*99*/ 0, 2719 0, 2720 0, 2721 MatConjugate_MPIAIJ, 2722 0, 2723 /*104*/MatSetValuesRow_MPIAIJ, 2724 MatRealPart_MPIAIJ, 2725 MatImaginaryPart_MPIAIJ, 2726 0, 2727 0, 2728 /*109*/0, 2729 0, 2730 MatGetRowMin_MPIAIJ, 2731 0, 2732 0, 2733 /*114*/MatGetSeqNonzeroStructure_MPIAIJ, 2734 0, 2735 0, 2736 0, 2737 0, 2738 /*119*/0, 2739 0, 2740 0, 2741 0, 2742 MatGetMultiProcBlock_MPIAIJ, 2743 /*124*/MatFindNonzeroRows_MPIAIJ, 2744 MatGetColumnNorms_MPIAIJ, 2745 MatInvertBlockDiagonal_MPIAIJ, 2746 0, 2747 MatGetSubMatricesParallel_MPIAIJ, 2748 /*129*/0, 2749 MatTransposeMatMult_MPIAIJ_MPIAIJ, 2750 MatTransposeMatMultSymbolic_MPIAIJ_MPIAIJ, 2751 MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ, 2752 0, 2753 /*134*/0, 2754 0, 2755 0, 2756 0, 2757 0, 2758 /*139*/0, 2759 0, 2760 0, 2761 MatFDColoringSetUp_MPIXAIJ, 2762 0, 2763 /*144*/MatCreateMPIMatConcatenateSeqMat_MPIAIJ 2764 }; 2765 2766 /* ----------------------------------------------------------------------------------------*/ 2767 2768 #undef __FUNCT__ 2769 #define __FUNCT__ "MatStoreValues_MPIAIJ" 2770 PetscErrorCode MatStoreValues_MPIAIJ(Mat mat) 2771 { 2772 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 2773 PetscErrorCode ierr; 2774 2775 PetscFunctionBegin; 2776 ierr = MatStoreValues(aij->A);CHKERRQ(ierr); 2777 ierr = MatStoreValues(aij->B);CHKERRQ(ierr); 2778 PetscFunctionReturn(0); 2779 } 2780 2781 #undef __FUNCT__ 2782 #define __FUNCT__ "MatRetrieveValues_MPIAIJ" 2783 PetscErrorCode MatRetrieveValues_MPIAIJ(Mat mat) 2784 { 2785 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 2786 PetscErrorCode ierr; 2787 2788 PetscFunctionBegin; 2789 ierr = MatRetrieveValues(aij->A);CHKERRQ(ierr); 2790 ierr = MatRetrieveValues(aij->B);CHKERRQ(ierr); 2791 PetscFunctionReturn(0); 2792 } 2793 2794 #undef __FUNCT__ 2795 #define __FUNCT__ "MatMPIAIJSetPreallocation_MPIAIJ" 2796 PetscErrorCode MatMPIAIJSetPreallocation_MPIAIJ(Mat B,PetscInt d_nz,const PetscInt d_nnz[],PetscInt o_nz,const PetscInt o_nnz[]) 2797 { 2798 Mat_MPIAIJ *b; 2799 PetscErrorCode ierr; 2800 2801 PetscFunctionBegin; 2802 ierr = PetscLayoutSetUp(B->rmap);CHKERRQ(ierr); 2803 ierr = PetscLayoutSetUp(B->cmap);CHKERRQ(ierr); 2804 b = (Mat_MPIAIJ*)B->data; 2805 2806 if (!B->preallocated) { 2807 /* Explicitly create 2 MATSEQAIJ matrices. */ 2808 ierr = MatCreate(PETSC_COMM_SELF,&b->A);CHKERRQ(ierr); 2809 ierr = MatSetSizes(b->A,B->rmap->n,B->cmap->n,B->rmap->n,B->cmap->n);CHKERRQ(ierr); 2810 ierr = MatSetBlockSizesFromMats(b->A,B,B);CHKERRQ(ierr); 2811 ierr = MatSetType(b->A,MATSEQAIJ);CHKERRQ(ierr); 2812 ierr = PetscLogObjectParent((PetscObject)B,(PetscObject)b->A);CHKERRQ(ierr); 2813 ierr = MatCreate(PETSC_COMM_SELF,&b->B);CHKERRQ(ierr); 2814 ierr = MatSetSizes(b->B,B->rmap->n,B->cmap->N,B->rmap->n,B->cmap->N);CHKERRQ(ierr); 2815 ierr = MatSetBlockSizesFromMats(b->B,B,B);CHKERRQ(ierr); 2816 ierr = MatSetType(b->B,MATSEQAIJ);CHKERRQ(ierr); 2817 ierr = PetscLogObjectParent((PetscObject)B,(PetscObject)b->B);CHKERRQ(ierr); 2818 } 2819 2820 ierr = MatSeqAIJSetPreallocation(b->A,d_nz,d_nnz);CHKERRQ(ierr); 2821 ierr = MatSeqAIJSetPreallocation(b->B,o_nz,o_nnz);CHKERRQ(ierr); 2822 B->preallocated = PETSC_TRUE; 2823 PetscFunctionReturn(0); 2824 } 2825 2826 #undef __FUNCT__ 2827 #define __FUNCT__ "MatDuplicate_MPIAIJ" 2828 PetscErrorCode MatDuplicate_MPIAIJ(Mat matin,MatDuplicateOption cpvalues,Mat *newmat) 2829 { 2830 Mat mat; 2831 Mat_MPIAIJ *a,*oldmat = (Mat_MPIAIJ*)matin->data; 2832 PetscErrorCode ierr; 2833 2834 PetscFunctionBegin; 2835 *newmat = 0; 2836 ierr = MatCreate(PetscObjectComm((PetscObject)matin),&mat);CHKERRQ(ierr); 2837 ierr = MatSetSizes(mat,matin->rmap->n,matin->cmap->n,matin->rmap->N,matin->cmap->N);CHKERRQ(ierr); 2838 ierr = MatSetBlockSizesFromMats(mat,matin,matin);CHKERRQ(ierr); 2839 ierr = MatSetType(mat,((PetscObject)matin)->type_name);CHKERRQ(ierr); 2840 ierr = PetscMemcpy(mat->ops,matin->ops,sizeof(struct _MatOps));CHKERRQ(ierr); 2841 a = (Mat_MPIAIJ*)mat->data; 2842 2843 mat->factortype = matin->factortype; 2844 mat->assembled = PETSC_TRUE; 2845 mat->insertmode = NOT_SET_VALUES; 2846 mat->preallocated = PETSC_TRUE; 2847 2848 a->size = oldmat->size; 2849 a->rank = oldmat->rank; 2850 a->donotstash = oldmat->donotstash; 2851 a->roworiented = oldmat->roworiented; 2852 a->rowindices = 0; 2853 a->rowvalues = 0; 2854 a->getrowactive = PETSC_FALSE; 2855 2856 ierr = PetscLayoutReference(matin->rmap,&mat->rmap);CHKERRQ(ierr); 2857 ierr = PetscLayoutReference(matin->cmap,&mat->cmap);CHKERRQ(ierr); 2858 2859 if (oldmat->colmap) { 2860 #if defined(PETSC_USE_CTABLE) 2861 ierr = PetscTableCreateCopy(oldmat->colmap,&a->colmap);CHKERRQ(ierr); 2862 #else 2863 ierr = PetscMalloc1((mat->cmap->N),&a->colmap);CHKERRQ(ierr); 2864 ierr = PetscLogObjectMemory((PetscObject)mat,(mat->cmap->N)*sizeof(PetscInt));CHKERRQ(ierr); 2865 ierr = PetscMemcpy(a->colmap,oldmat->colmap,(mat->cmap->N)*sizeof(PetscInt));CHKERRQ(ierr); 2866 #endif 2867 } else a->colmap = 0; 2868 if (oldmat->garray) { 2869 PetscInt len; 2870 len = oldmat->B->cmap->n; 2871 ierr = PetscMalloc1((len+1),&a->garray);CHKERRQ(ierr); 2872 ierr = PetscLogObjectMemory((PetscObject)mat,len*sizeof(PetscInt));CHKERRQ(ierr); 2873 if (len) { ierr = PetscMemcpy(a->garray,oldmat->garray,len*sizeof(PetscInt));CHKERRQ(ierr); } 2874 } else a->garray = 0; 2875 2876 ierr = VecDuplicate(oldmat->lvec,&a->lvec);CHKERRQ(ierr); 2877 ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)a->lvec);CHKERRQ(ierr); 2878 ierr = VecScatterCopy(oldmat->Mvctx,&a->Mvctx);CHKERRQ(ierr); 2879 ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)a->Mvctx);CHKERRQ(ierr); 2880 ierr = MatDuplicate(oldmat->A,cpvalues,&a->A);CHKERRQ(ierr); 2881 ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)a->A);CHKERRQ(ierr); 2882 ierr = MatDuplicate(oldmat->B,cpvalues,&a->B);CHKERRQ(ierr); 2883 ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)a->B);CHKERRQ(ierr); 2884 ierr = PetscFunctionListDuplicate(((PetscObject)matin)->qlist,&((PetscObject)mat)->qlist);CHKERRQ(ierr); 2885 *newmat = mat; 2886 PetscFunctionReturn(0); 2887 } 2888 2889 2890 2891 #undef __FUNCT__ 2892 #define __FUNCT__ "MatLoad_MPIAIJ" 2893 PetscErrorCode MatLoad_MPIAIJ(Mat newMat, PetscViewer viewer) 2894 { 2895 PetscScalar *vals,*svals; 2896 MPI_Comm comm; 2897 PetscErrorCode ierr; 2898 PetscMPIInt rank,size,tag = ((PetscObject)viewer)->tag; 2899 PetscInt i,nz,j,rstart,rend,mmax,maxnz = 0,grows,gcols; 2900 PetscInt header[4],*rowlengths = 0,M,N,m,*cols; 2901 PetscInt *ourlens = NULL,*procsnz = NULL,*offlens = NULL,jj,*mycols,*smycols; 2902 PetscInt cend,cstart,n,*rowners,sizesset=1; 2903 int fd; 2904 PetscInt bs = 1; 2905 2906 PetscFunctionBegin; 2907 ierr = PetscObjectGetComm((PetscObject)viewer,&comm);CHKERRQ(ierr); 2908 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 2909 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 2910 if (!rank) { 2911 ierr = PetscViewerBinaryGetDescriptor(viewer,&fd);CHKERRQ(ierr); 2912 ierr = PetscBinaryRead(fd,(char*)header,4,PETSC_INT);CHKERRQ(ierr); 2913 if (header[0] != MAT_FILE_CLASSID) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED,"not matrix object"); 2914 } 2915 2916 ierr = PetscOptionsBegin(comm,NULL,"Options for loading SEQAIJ matrix","Mat");CHKERRQ(ierr); 2917 ierr = PetscOptionsInt("-matload_block_size","Set the blocksize used to store the matrix","MatLoad",bs,&bs,NULL);CHKERRQ(ierr); 2918 ierr = PetscOptionsEnd();CHKERRQ(ierr); 2919 2920 if (newMat->rmap->n < 0 && newMat->rmap->N < 0 && newMat->cmap->n < 0 && newMat->cmap->N < 0) sizesset = 0; 2921 2922 ierr = MPI_Bcast(header+1,3,MPIU_INT,0,comm);CHKERRQ(ierr); 2923 M = header[1]; N = header[2]; 2924 /* If global rows/cols are set to PETSC_DECIDE, set it to the sizes given in the file */ 2925 if (sizesset && newMat->rmap->N < 0) newMat->rmap->N = M; 2926 if (sizesset && newMat->cmap->N < 0) newMat->cmap->N = N; 2927 2928 /* If global sizes are set, check if they are consistent with that given in the file */ 2929 if (sizesset) { 2930 ierr = MatGetSize(newMat,&grows,&gcols);CHKERRQ(ierr); 2931 } 2932 if (sizesset && newMat->rmap->N != grows) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED, "Inconsistent # of rows:Matrix in file has (%d) and input matrix has (%d)",M,grows); 2933 if (sizesset && newMat->cmap->N != gcols) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED, "Inconsistent # of cols:Matrix in file has (%d) and input matrix has (%d)",N,gcols); 2934 2935 /* determine ownership of all (block) rows */ 2936 if (M%bs) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED, "Inconsistent # of rows (%d) and block size (%d)",M,bs); 2937 if (newMat->rmap->n < 0) m = bs*((M/bs)/size + (((M/bs) % size) > rank)); /* PETSC_DECIDE */ 2938 else m = newMat->rmap->n; /* Set by user */ 2939 2940 ierr = PetscMalloc1((size+1),&rowners);CHKERRQ(ierr); 2941 ierr = MPI_Allgather(&m,1,MPIU_INT,rowners+1,1,MPIU_INT,comm);CHKERRQ(ierr); 2942 2943 /* First process needs enough room for process with most rows */ 2944 if (!rank) { 2945 mmax = rowners[1]; 2946 for (i=2; i<=size; i++) { 2947 mmax = PetscMax(mmax, rowners[i]); 2948 } 2949 } else mmax = -1; /* unused, but compilers complain */ 2950 2951 rowners[0] = 0; 2952 for (i=2; i<=size; i++) { 2953 rowners[i] += rowners[i-1]; 2954 } 2955 rstart = rowners[rank]; 2956 rend = rowners[rank+1]; 2957 2958 /* distribute row lengths to all processors */ 2959 ierr = PetscMalloc2(m,&ourlens,m,&offlens);CHKERRQ(ierr); 2960 if (!rank) { 2961 ierr = PetscBinaryRead(fd,ourlens,m,PETSC_INT);CHKERRQ(ierr); 2962 ierr = PetscMalloc1(mmax,&rowlengths);CHKERRQ(ierr); 2963 ierr = PetscCalloc1(size,&procsnz);CHKERRQ(ierr); 2964 for (j=0; j<m; j++) { 2965 procsnz[0] += ourlens[j]; 2966 } 2967 for (i=1; i<size; i++) { 2968 ierr = PetscBinaryRead(fd,rowlengths,rowners[i+1]-rowners[i],PETSC_INT);CHKERRQ(ierr); 2969 /* calculate the number of nonzeros on each processor */ 2970 for (j=0; j<rowners[i+1]-rowners[i]; j++) { 2971 procsnz[i] += rowlengths[j]; 2972 } 2973 ierr = MPIULong_Send(rowlengths,rowners[i+1]-rowners[i],MPIU_INT,i,tag,comm);CHKERRQ(ierr); 2974 } 2975 ierr = PetscFree(rowlengths);CHKERRQ(ierr); 2976 } else { 2977 ierr = MPIULong_Recv(ourlens,m,MPIU_INT,0,tag,comm);CHKERRQ(ierr); 2978 } 2979 2980 if (!rank) { 2981 /* determine max buffer needed and allocate it */ 2982 maxnz = 0; 2983 for (i=0; i<size; i++) { 2984 maxnz = PetscMax(maxnz,procsnz[i]); 2985 } 2986 ierr = PetscMalloc1(maxnz,&cols);CHKERRQ(ierr); 2987 2988 /* read in my part of the matrix column indices */ 2989 nz = procsnz[0]; 2990 ierr = PetscMalloc1(nz,&mycols);CHKERRQ(ierr); 2991 ierr = PetscBinaryRead(fd,mycols,nz,PETSC_INT);CHKERRQ(ierr); 2992 2993 /* read in every one elses and ship off */ 2994 for (i=1; i<size; i++) { 2995 nz = procsnz[i]; 2996 ierr = PetscBinaryRead(fd,cols,nz,PETSC_INT);CHKERRQ(ierr); 2997 ierr = MPIULong_Send(cols,nz,MPIU_INT,i,tag,comm);CHKERRQ(ierr); 2998 } 2999 ierr = PetscFree(cols);CHKERRQ(ierr); 3000 } else { 3001 /* determine buffer space needed for message */ 3002 nz = 0; 3003 for (i=0; i<m; i++) { 3004 nz += ourlens[i]; 3005 } 3006 ierr = PetscMalloc1(nz,&mycols);CHKERRQ(ierr); 3007 3008 /* receive message of column indices*/ 3009 ierr = MPIULong_Recv(mycols,nz,MPIU_INT,0,tag,comm);CHKERRQ(ierr); 3010 } 3011 3012 /* determine column ownership if matrix is not square */ 3013 if (N != M) { 3014 if (newMat->cmap->n < 0) n = N/size + ((N % size) > rank); 3015 else n = newMat->cmap->n; 3016 ierr = MPI_Scan(&n,&cend,1,MPIU_INT,MPI_SUM,comm);CHKERRQ(ierr); 3017 cstart = cend - n; 3018 } else { 3019 cstart = rstart; 3020 cend = rend; 3021 n = cend - cstart; 3022 } 3023 3024 /* loop over local rows, determining number of off diagonal entries */ 3025 ierr = PetscMemzero(offlens,m*sizeof(PetscInt));CHKERRQ(ierr); 3026 jj = 0; 3027 for (i=0; i<m; i++) { 3028 for (j=0; j<ourlens[i]; j++) { 3029 if (mycols[jj] < cstart || mycols[jj] >= cend) offlens[i]++; 3030 jj++; 3031 } 3032 } 3033 3034 for (i=0; i<m; i++) { 3035 ourlens[i] -= offlens[i]; 3036 } 3037 if (!sizesset) { 3038 ierr = MatSetSizes(newMat,m,n,M,N);CHKERRQ(ierr); 3039 } 3040 3041 if (bs > 1) {ierr = MatSetBlockSize(newMat,bs);CHKERRQ(ierr);} 3042 3043 ierr = MatMPIAIJSetPreallocation(newMat,0,ourlens,0,offlens);CHKERRQ(ierr); 3044 3045 for (i=0; i<m; i++) { 3046 ourlens[i] += offlens[i]; 3047 } 3048 3049 if (!rank) { 3050 ierr = PetscMalloc1((maxnz+1),&vals);CHKERRQ(ierr); 3051 3052 /* read in my part of the matrix numerical values */ 3053 nz = procsnz[0]; 3054 ierr = PetscBinaryRead(fd,vals,nz,PETSC_SCALAR);CHKERRQ(ierr); 3055 3056 /* insert into matrix */ 3057 jj = rstart; 3058 smycols = mycols; 3059 svals = vals; 3060 for (i=0; i<m; i++) { 3061 ierr = MatSetValues_MPIAIJ(newMat,1,&jj,ourlens[i],smycols,svals,INSERT_VALUES);CHKERRQ(ierr); 3062 smycols += ourlens[i]; 3063 svals += ourlens[i]; 3064 jj++; 3065 } 3066 3067 /* read in other processors and ship out */ 3068 for (i=1; i<size; i++) { 3069 nz = procsnz[i]; 3070 ierr = PetscBinaryRead(fd,vals,nz,PETSC_SCALAR);CHKERRQ(ierr); 3071 ierr = MPIULong_Send(vals,nz,MPIU_SCALAR,i,((PetscObject)newMat)->tag,comm);CHKERRQ(ierr); 3072 } 3073 ierr = PetscFree(procsnz);CHKERRQ(ierr); 3074 } else { 3075 /* receive numeric values */ 3076 ierr = PetscMalloc1((nz+1),&vals);CHKERRQ(ierr); 3077 3078 /* receive message of values*/ 3079 ierr = MPIULong_Recv(vals,nz,MPIU_SCALAR,0,((PetscObject)newMat)->tag,comm);CHKERRQ(ierr); 3080 3081 /* insert into matrix */ 3082 jj = rstart; 3083 smycols = mycols; 3084 svals = vals; 3085 for (i=0; i<m; i++) { 3086 ierr = MatSetValues_MPIAIJ(newMat,1,&jj,ourlens[i],smycols,svals,INSERT_VALUES);CHKERRQ(ierr); 3087 smycols += ourlens[i]; 3088 svals += ourlens[i]; 3089 jj++; 3090 } 3091 } 3092 ierr = PetscFree2(ourlens,offlens);CHKERRQ(ierr); 3093 ierr = PetscFree(vals);CHKERRQ(ierr); 3094 ierr = PetscFree(mycols);CHKERRQ(ierr); 3095 ierr = PetscFree(rowners);CHKERRQ(ierr); 3096 ierr = MatAssemblyBegin(newMat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3097 ierr = MatAssemblyEnd(newMat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3098 PetscFunctionReturn(0); 3099 } 3100 3101 #undef __FUNCT__ 3102 #define __FUNCT__ "MatGetSubMatrix_MPIAIJ" 3103 PetscErrorCode MatGetSubMatrix_MPIAIJ(Mat mat,IS isrow,IS iscol,MatReuse call,Mat *newmat) 3104 { 3105 PetscErrorCode ierr; 3106 IS iscol_local; 3107 PetscInt csize; 3108 3109 PetscFunctionBegin; 3110 ierr = ISGetLocalSize(iscol,&csize);CHKERRQ(ierr); 3111 if (call == MAT_REUSE_MATRIX) { 3112 ierr = PetscObjectQuery((PetscObject)*newmat,"ISAllGather",(PetscObject*)&iscol_local);CHKERRQ(ierr); 3113 if (!iscol_local) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Submatrix passed in was not used before, cannot reuse"); 3114 } else { 3115 PetscInt cbs; 3116 ierr = ISGetBlockSize(iscol,&cbs);CHKERRQ(ierr); 3117 ierr = ISAllGather(iscol,&iscol_local);CHKERRQ(ierr); 3118 ierr = ISSetBlockSize(iscol_local,cbs);CHKERRQ(ierr); 3119 } 3120 ierr = MatGetSubMatrix_MPIAIJ_Private(mat,isrow,iscol_local,csize,call,newmat);CHKERRQ(ierr); 3121 if (call == MAT_INITIAL_MATRIX) { 3122 ierr = PetscObjectCompose((PetscObject)*newmat,"ISAllGather",(PetscObject)iscol_local);CHKERRQ(ierr); 3123 ierr = ISDestroy(&iscol_local);CHKERRQ(ierr); 3124 } 3125 PetscFunctionReturn(0); 3126 } 3127 3128 extern PetscErrorCode MatGetSubMatrices_MPIAIJ_Local(Mat,PetscInt,const IS[],const IS[],MatReuse,PetscBool*,Mat*); 3129 #undef __FUNCT__ 3130 #define __FUNCT__ "MatGetSubMatrix_MPIAIJ_Private" 3131 /* 3132 Not great since it makes two copies of the submatrix, first an SeqAIJ 3133 in local and then by concatenating the local matrices the end result. 3134 Writing it directly would be much like MatGetSubMatrices_MPIAIJ() 3135 3136 Note: This requires a sequential iscol with all indices. 3137 */ 3138 PetscErrorCode MatGetSubMatrix_MPIAIJ_Private(Mat mat,IS isrow,IS iscol,PetscInt csize,MatReuse call,Mat *newmat) 3139 { 3140 PetscErrorCode ierr; 3141 PetscMPIInt rank,size; 3142 PetscInt i,m,n,rstart,row,rend,nz,*cwork,j,bs,cbs; 3143 PetscInt *ii,*jj,nlocal,*dlens,*olens,dlen,olen,jend,mglobal,ncol; 3144 PetscBool allcolumns, colflag; 3145 Mat M,Mreuse; 3146 MatScalar *vwork,*aa; 3147 MPI_Comm comm; 3148 Mat_SeqAIJ *aij; 3149 3150 PetscFunctionBegin; 3151 ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 3152 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 3153 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 3154 3155 ierr = ISIdentity(iscol,&colflag);CHKERRQ(ierr); 3156 ierr = ISGetLocalSize(iscol,&ncol);CHKERRQ(ierr); 3157 if (colflag && ncol == mat->cmap->N) { 3158 allcolumns = PETSC_TRUE; 3159 } else { 3160 allcolumns = PETSC_FALSE; 3161 } 3162 if (call == MAT_REUSE_MATRIX) { 3163 ierr = PetscObjectQuery((PetscObject)*newmat,"SubMatrix",(PetscObject*)&Mreuse);CHKERRQ(ierr); 3164 if (!Mreuse) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Submatrix passed in was not used before, cannot reuse"); 3165 ierr = MatGetSubMatrices_MPIAIJ_Local(mat,1,&isrow,&iscol,MAT_REUSE_MATRIX,&allcolumns,&Mreuse);CHKERRQ(ierr); 3166 } else { 3167 ierr = MatGetSubMatrices_MPIAIJ_Local(mat,1,&isrow,&iscol,MAT_INITIAL_MATRIX,&allcolumns,&Mreuse);CHKERRQ(ierr); 3168 } 3169 3170 /* 3171 m - number of local rows 3172 n - number of columns (same on all processors) 3173 rstart - first row in new global matrix generated 3174 */ 3175 ierr = MatGetSize(Mreuse,&m,&n);CHKERRQ(ierr); 3176 ierr = MatGetBlockSizes(Mreuse,&bs,&cbs);CHKERRQ(ierr); 3177 if (call == MAT_INITIAL_MATRIX) { 3178 aij = (Mat_SeqAIJ*)(Mreuse)->data; 3179 ii = aij->i; 3180 jj = aij->j; 3181 3182 /* 3183 Determine the number of non-zeros in the diagonal and off-diagonal 3184 portions of the matrix in order to do correct preallocation 3185 */ 3186 3187 /* first get start and end of "diagonal" columns */ 3188 if (csize == PETSC_DECIDE) { 3189 ierr = ISGetSize(isrow,&mglobal);CHKERRQ(ierr); 3190 if (mglobal == n) { /* square matrix */ 3191 nlocal = m; 3192 } else { 3193 nlocal = n/size + ((n % size) > rank); 3194 } 3195 } else { 3196 nlocal = csize; 3197 } 3198 ierr = MPI_Scan(&nlocal,&rend,1,MPIU_INT,MPI_SUM,comm);CHKERRQ(ierr); 3199 rstart = rend - nlocal; 3200 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); 3201 3202 /* next, compute all the lengths */ 3203 ierr = PetscMalloc1((2*m+1),&dlens);CHKERRQ(ierr); 3204 olens = dlens + m; 3205 for (i=0; i<m; i++) { 3206 jend = ii[i+1] - ii[i]; 3207 olen = 0; 3208 dlen = 0; 3209 for (j=0; j<jend; j++) { 3210 if (*jj < rstart || *jj >= rend) olen++; 3211 else dlen++; 3212 jj++; 3213 } 3214 olens[i] = olen; 3215 dlens[i] = dlen; 3216 } 3217 ierr = MatCreate(comm,&M);CHKERRQ(ierr); 3218 ierr = MatSetSizes(M,m,nlocal,PETSC_DECIDE,n);CHKERRQ(ierr); 3219 ierr = MatSetBlockSizes(M,bs,cbs);CHKERRQ(ierr); 3220 ierr = MatSetType(M,((PetscObject)mat)->type_name);CHKERRQ(ierr); 3221 ierr = MatMPIAIJSetPreallocation(M,0,dlens,0,olens);CHKERRQ(ierr); 3222 ierr = PetscFree(dlens);CHKERRQ(ierr); 3223 } else { 3224 PetscInt ml,nl; 3225 3226 M = *newmat; 3227 ierr = MatGetLocalSize(M,&ml,&nl);CHKERRQ(ierr); 3228 if (ml != m) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Previous matrix must be same size/layout as request"); 3229 ierr = MatZeroEntries(M);CHKERRQ(ierr); 3230 /* 3231 The next two lines are needed so we may call MatSetValues_MPIAIJ() below directly, 3232 rather than the slower MatSetValues(). 3233 */ 3234 M->was_assembled = PETSC_TRUE; 3235 M->assembled = PETSC_FALSE; 3236 } 3237 ierr = MatGetOwnershipRange(M,&rstart,&rend);CHKERRQ(ierr); 3238 aij = (Mat_SeqAIJ*)(Mreuse)->data; 3239 ii = aij->i; 3240 jj = aij->j; 3241 aa = aij->a; 3242 for (i=0; i<m; i++) { 3243 row = rstart + i; 3244 nz = ii[i+1] - ii[i]; 3245 cwork = jj; jj += nz; 3246 vwork = aa; aa += nz; 3247 ierr = MatSetValues_MPIAIJ(M,1,&row,nz,cwork,vwork,INSERT_VALUES);CHKERRQ(ierr); 3248 } 3249 3250 ierr = MatAssemblyBegin(M,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3251 ierr = MatAssemblyEnd(M,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3252 *newmat = M; 3253 3254 /* save submatrix used in processor for next request */ 3255 if (call == MAT_INITIAL_MATRIX) { 3256 ierr = PetscObjectCompose((PetscObject)M,"SubMatrix",(PetscObject)Mreuse);CHKERRQ(ierr); 3257 ierr = MatDestroy(&Mreuse);CHKERRQ(ierr); 3258 } 3259 PetscFunctionReturn(0); 3260 } 3261 3262 #undef __FUNCT__ 3263 #define __FUNCT__ "MatMPIAIJSetPreallocationCSR_MPIAIJ" 3264 PetscErrorCode MatMPIAIJSetPreallocationCSR_MPIAIJ(Mat B,const PetscInt Ii[],const PetscInt J[],const PetscScalar v[]) 3265 { 3266 PetscInt m,cstart, cend,j,nnz,i,d; 3267 PetscInt *d_nnz,*o_nnz,nnz_max = 0,rstart,ii; 3268 const PetscInt *JJ; 3269 PetscScalar *values; 3270 PetscErrorCode ierr; 3271 3272 PetscFunctionBegin; 3273 if (Ii[0]) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Ii[0] must be 0 it is %D",Ii[0]); 3274 3275 ierr = PetscLayoutSetUp(B->rmap);CHKERRQ(ierr); 3276 ierr = PetscLayoutSetUp(B->cmap);CHKERRQ(ierr); 3277 m = B->rmap->n; 3278 cstart = B->cmap->rstart; 3279 cend = B->cmap->rend; 3280 rstart = B->rmap->rstart; 3281 3282 ierr = PetscMalloc2(m,&d_nnz,m,&o_nnz);CHKERRQ(ierr); 3283 3284 #if defined(PETSC_USE_DEBUGGING) 3285 for (i=0; i<m; i++) { 3286 nnz = Ii[i+1]- Ii[i]; 3287 JJ = J + Ii[i]; 3288 if (nnz < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Local row %D has a negative %D number of columns",i,nnz); 3289 if (nnz && (JJ[0] < 0)) SETERRRQ1(PETSC_ERR_ARG_WRONGSTATE,"Row %D starts with negative column index",i,j); 3290 if (nnz && (JJ[nnz-1] >= B->cmap->N) SETERRRQ3(PETSC_ERR_ARG_WRONGSTATE,"Row %D ends with too large a column index %D (max allowed %D)",i,JJ[nnz-1],B->cmap->N); 3291 } 3292 #endif 3293 3294 for (i=0; i<m; i++) { 3295 nnz = Ii[i+1]- Ii[i]; 3296 JJ = J + Ii[i]; 3297 nnz_max = PetscMax(nnz_max,nnz); 3298 d = 0; 3299 for (j=0; j<nnz; j++) { 3300 if (cstart <= JJ[j] && JJ[j] < cend) d++; 3301 } 3302 d_nnz[i] = d; 3303 o_nnz[i] = nnz - d; 3304 } 3305 ierr = MatMPIAIJSetPreallocation(B,0,d_nnz,0,o_nnz);CHKERRQ(ierr); 3306 ierr = PetscFree2(d_nnz,o_nnz);CHKERRQ(ierr); 3307 3308 if (v) values = (PetscScalar*)v; 3309 else { 3310 ierr = PetscCalloc1((nnz_max+1),&values);CHKERRQ(ierr); 3311 } 3312 3313 for (i=0; i<m; i++) { 3314 ii = i + rstart; 3315 nnz = Ii[i+1]- Ii[i]; 3316 ierr = MatSetValues_MPIAIJ(B,1,&ii,nnz,J+Ii[i],values+(v ? Ii[i] : 0),INSERT_VALUES);CHKERRQ(ierr); 3317 } 3318 ierr = MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3319 ierr = MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3320 3321 if (!v) { 3322 ierr = PetscFree(values);CHKERRQ(ierr); 3323 } 3324 ierr = MatSetOption(B,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE);CHKERRQ(ierr); 3325 PetscFunctionReturn(0); 3326 } 3327 3328 #undef __FUNCT__ 3329 #define __FUNCT__ "MatMPIAIJSetPreallocationCSR" 3330 /*@ 3331 MatMPIAIJSetPreallocationCSR - Allocates memory for a sparse parallel matrix in AIJ format 3332 (the default parallel PETSc format). 3333 3334 Collective on MPI_Comm 3335 3336 Input Parameters: 3337 + B - the matrix 3338 . i - the indices into j for the start of each local row (starts with zero) 3339 . j - the column indices for each local row (starts with zero) 3340 - v - optional values in the matrix 3341 3342 Level: developer 3343 3344 Notes: 3345 The i, j, and a arrays ARE copied by this routine into the internal format used by PETSc; 3346 thus you CANNOT change the matrix entries by changing the values of a[] after you have 3347 called this routine. Use MatCreateMPIAIJWithSplitArrays() to avoid needing to copy the arrays. 3348 3349 The i and j indices are 0 based, and i indices are indices corresponding to the local j array. 3350 3351 The format which is used for the sparse matrix input, is equivalent to a 3352 row-major ordering.. i.e for the following matrix, the input data expected is 3353 as shown: 3354 3355 1 0 0 3356 2 0 3 P0 3357 ------- 3358 4 5 6 P1 3359 3360 Process0 [P0]: rows_owned=[0,1] 3361 i = {0,1,3} [size = nrow+1 = 2+1] 3362 j = {0,0,2} [size = nz = 6] 3363 v = {1,2,3} [size = nz = 6] 3364 3365 Process1 [P1]: rows_owned=[2] 3366 i = {0,3} [size = nrow+1 = 1+1] 3367 j = {0,1,2} [size = nz = 6] 3368 v = {4,5,6} [size = nz = 6] 3369 3370 .keywords: matrix, aij, compressed row, sparse, parallel 3371 3372 .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatCreateAIJ(), MPIAIJ, 3373 MatCreateSeqAIJWithArrays(), MatCreateMPIAIJWithSplitArrays() 3374 @*/ 3375 PetscErrorCode MatMPIAIJSetPreallocationCSR(Mat B,const PetscInt i[],const PetscInt j[], const PetscScalar v[]) 3376 { 3377 PetscErrorCode ierr; 3378 3379 PetscFunctionBegin; 3380 ierr = PetscTryMethod(B,"MatMPIAIJSetPreallocationCSR_C",(Mat,const PetscInt[],const PetscInt[],const PetscScalar[]),(B,i,j,v));CHKERRQ(ierr); 3381 PetscFunctionReturn(0); 3382 } 3383 3384 #undef __FUNCT__ 3385 #define __FUNCT__ "MatMPIAIJSetPreallocation" 3386 /*@C 3387 MatMPIAIJSetPreallocation - Preallocates memory for a sparse parallel matrix in AIJ format 3388 (the default parallel PETSc format). For good matrix assembly performance 3389 the user should preallocate the matrix storage by setting the parameters 3390 d_nz (or d_nnz) and o_nz (or o_nnz). By setting these parameters accurately, 3391 performance can be increased by more than a factor of 50. 3392 3393 Collective on MPI_Comm 3394 3395 Input Parameters: 3396 + B - the matrix 3397 . d_nz - number of nonzeros per row in DIAGONAL portion of local submatrix 3398 (same value is used for all local rows) 3399 . d_nnz - array containing the number of nonzeros in the various rows of the 3400 DIAGONAL portion of the local submatrix (possibly different for each row) 3401 or NULL, if d_nz is used to specify the nonzero structure. 3402 The size of this array is equal to the number of local rows, i.e 'm'. 3403 For matrices that will be factored, you must leave room for (and set) 3404 the diagonal entry even if it is zero. 3405 . o_nz - number of nonzeros per row in the OFF-DIAGONAL portion of local 3406 submatrix (same value is used for all local rows). 3407 - o_nnz - array containing the number of nonzeros in the various rows of the 3408 OFF-DIAGONAL portion of the local submatrix (possibly different for 3409 each row) or NULL, if o_nz is used to specify the nonzero 3410 structure. The size of this array is equal to the number 3411 of local rows, i.e 'm'. 3412 3413 If the *_nnz parameter is given then the *_nz parameter is ignored 3414 3415 The AIJ format (also called the Yale sparse matrix format or 3416 compressed row storage (CSR)), is fully compatible with standard Fortran 77 3417 storage. The stored row and column indices begin with zero. 3418 See Users-Manual: ch_mat for details. 3419 3420 The parallel matrix is partitioned such that the first m0 rows belong to 3421 process 0, the next m1 rows belong to process 1, the next m2 rows belong 3422 to process 2 etc.. where m0,m1,m2... are the input parameter 'm'. 3423 3424 The DIAGONAL portion of the local submatrix of a processor can be defined 3425 as the submatrix which is obtained by extraction the part corresponding to 3426 the rows r1-r2 and columns c1-c2 of the global matrix, where r1 is the 3427 first row that belongs to the processor, r2 is the last row belonging to 3428 the this processor, and c1-c2 is range of indices of the local part of a 3429 vector suitable for applying the matrix to. This is an mxn matrix. In the 3430 common case of a square matrix, the row and column ranges are the same and 3431 the DIAGONAL part is also square. The remaining portion of the local 3432 submatrix (mxN) constitute the OFF-DIAGONAL portion. 3433 3434 If o_nnz, d_nnz are specified, then o_nz, and d_nz are ignored. 3435 3436 You can call MatGetInfo() to get information on how effective the preallocation was; 3437 for example the fields mallocs,nz_allocated,nz_used,nz_unneeded; 3438 You can also run with the option -info and look for messages with the string 3439 malloc in them to see if additional memory allocation was needed. 3440 3441 Example usage: 3442 3443 Consider the following 8x8 matrix with 34 non-zero values, that is 3444 assembled across 3 processors. Lets assume that proc0 owns 3 rows, 3445 proc1 owns 3 rows, proc2 owns 2 rows. This division can be shown 3446 as follows: 3447 3448 .vb 3449 1 2 0 | 0 3 0 | 0 4 3450 Proc0 0 5 6 | 7 0 0 | 8 0 3451 9 0 10 | 11 0 0 | 12 0 3452 ------------------------------------- 3453 13 0 14 | 15 16 17 | 0 0 3454 Proc1 0 18 0 | 19 20 21 | 0 0 3455 0 0 0 | 22 23 0 | 24 0 3456 ------------------------------------- 3457 Proc2 25 26 27 | 0 0 28 | 29 0 3458 30 0 0 | 31 32 33 | 0 34 3459 .ve 3460 3461 This can be represented as a collection of submatrices as: 3462 3463 .vb 3464 A B C 3465 D E F 3466 G H I 3467 .ve 3468 3469 Where the submatrices A,B,C are owned by proc0, D,E,F are 3470 owned by proc1, G,H,I are owned by proc2. 3471 3472 The 'm' parameters for proc0,proc1,proc2 are 3,3,2 respectively. 3473 The 'n' parameters for proc0,proc1,proc2 are 3,3,2 respectively. 3474 The 'M','N' parameters are 8,8, and have the same values on all procs. 3475 3476 The DIAGONAL submatrices corresponding to proc0,proc1,proc2 are 3477 submatrices [A], [E], [I] respectively. The OFF-DIAGONAL submatrices 3478 corresponding to proc0,proc1,proc2 are [BC], [DF], [GH] respectively. 3479 Internally, each processor stores the DIAGONAL part, and the OFF-DIAGONAL 3480 part as SeqAIJ matrices. for eg: proc1 will store [E] as a SeqAIJ 3481 matrix, ans [DF] as another SeqAIJ matrix. 3482 3483 When d_nz, o_nz parameters are specified, d_nz storage elements are 3484 allocated for every row of the local diagonal submatrix, and o_nz 3485 storage locations are allocated for every row of the OFF-DIAGONAL submat. 3486 One way to choose d_nz and o_nz is to use the max nonzerors per local 3487 rows for each of the local DIAGONAL, and the OFF-DIAGONAL submatrices. 3488 In this case, the values of d_nz,o_nz are: 3489 .vb 3490 proc0 : dnz = 2, o_nz = 2 3491 proc1 : dnz = 3, o_nz = 2 3492 proc2 : dnz = 1, o_nz = 4 3493 .ve 3494 We are allocating m*(d_nz+o_nz) storage locations for every proc. This 3495 translates to 3*(2+2)=12 for proc0, 3*(3+2)=15 for proc1, 2*(1+4)=10 3496 for proc3. i.e we are using 12+15+10=37 storage locations to store 3497 34 values. 3498 3499 When d_nnz, o_nnz parameters are specified, the storage is specified 3500 for every row, coresponding to both DIAGONAL and OFF-DIAGONAL submatrices. 3501 In the above case the values for d_nnz,o_nnz are: 3502 .vb 3503 proc0: d_nnz = [2,2,2] and o_nnz = [2,2,2] 3504 proc1: d_nnz = [3,3,2] and o_nnz = [2,1,1] 3505 proc2: d_nnz = [1,1] and o_nnz = [4,4] 3506 .ve 3507 Here the space allocated is sum of all the above values i.e 34, and 3508 hence pre-allocation is perfect. 3509 3510 Level: intermediate 3511 3512 .keywords: matrix, aij, compressed row, sparse, parallel 3513 3514 .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatCreateAIJ(), MatMPIAIJSetPreallocationCSR(), 3515 MPIAIJ, MatGetInfo(), PetscSplitOwnership() 3516 @*/ 3517 PetscErrorCode MatMPIAIJSetPreallocation(Mat B,PetscInt d_nz,const PetscInt d_nnz[],PetscInt o_nz,const PetscInt o_nnz[]) 3518 { 3519 PetscErrorCode ierr; 3520 3521 PetscFunctionBegin; 3522 PetscValidHeaderSpecific(B,MAT_CLASSID,1); 3523 PetscValidType(B,1); 3524 ierr = PetscTryMethod(B,"MatMPIAIJSetPreallocation_C",(Mat,PetscInt,const PetscInt[],PetscInt,const PetscInt[]),(B,d_nz,d_nnz,o_nz,o_nnz));CHKERRQ(ierr); 3525 PetscFunctionReturn(0); 3526 } 3527 3528 #undef __FUNCT__ 3529 #define __FUNCT__ "MatCreateMPIAIJWithArrays" 3530 /*@ 3531 MatCreateMPIAIJWithArrays - creates a MPI AIJ matrix using arrays that contain in standard 3532 CSR format the local rows. 3533 3534 Collective on MPI_Comm 3535 3536 Input Parameters: 3537 + comm - MPI communicator 3538 . m - number of local rows (Cannot be PETSC_DECIDE) 3539 . n - This value should be the same as the local size used in creating the 3540 x vector for the matrix-vector product y = Ax. (or PETSC_DECIDE to have 3541 calculated if N is given) For square matrices n is almost always m. 3542 . M - number of global rows (or PETSC_DETERMINE to have calculated if m is given) 3543 . N - number of global columns (or PETSC_DETERMINE to have calculated if n is given) 3544 . i - row indices 3545 . j - column indices 3546 - a - matrix values 3547 3548 Output Parameter: 3549 . mat - the matrix 3550 3551 Level: intermediate 3552 3553 Notes: 3554 The i, j, and a arrays ARE copied by this routine into the internal format used by PETSc; 3555 thus you CANNOT change the matrix entries by changing the values of a[] after you have 3556 called this routine. Use MatCreateMPIAIJWithSplitArrays() to avoid needing to copy the arrays. 3557 3558 The i and j indices are 0 based, and i indices are indices corresponding to the local j array. 3559 3560 The format which is used for the sparse matrix input, is equivalent to a 3561 row-major ordering.. i.e for the following matrix, the input data expected is 3562 as shown: 3563 3564 1 0 0 3565 2 0 3 P0 3566 ------- 3567 4 5 6 P1 3568 3569 Process0 [P0]: rows_owned=[0,1] 3570 i = {0,1,3} [size = nrow+1 = 2+1] 3571 j = {0,0,2} [size = nz = 6] 3572 v = {1,2,3} [size = nz = 6] 3573 3574 Process1 [P1]: rows_owned=[2] 3575 i = {0,3} [size = nrow+1 = 1+1] 3576 j = {0,1,2} [size = nz = 6] 3577 v = {4,5,6} [size = nz = 6] 3578 3579 .keywords: matrix, aij, compressed row, sparse, parallel 3580 3581 .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatMPIAIJSetPreallocationCSR(), 3582 MPIAIJ, MatCreateAIJ(), MatCreateMPIAIJWithSplitArrays() 3583 @*/ 3584 PetscErrorCode MatCreateMPIAIJWithArrays(MPI_Comm comm,PetscInt m,PetscInt n,PetscInt M,PetscInt N,const PetscInt i[],const PetscInt j[],const PetscScalar a[],Mat *mat) 3585 { 3586 PetscErrorCode ierr; 3587 3588 PetscFunctionBegin; 3589 if (i[0]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"i (row indices) must start with 0"); 3590 if (m < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"local number of rows (m) cannot be PETSC_DECIDE, or negative"); 3591 ierr = MatCreate(comm,mat);CHKERRQ(ierr); 3592 ierr = MatSetSizes(*mat,m,n,M,N);CHKERRQ(ierr); 3593 /* ierr = MatSetBlockSizes(M,bs,cbs);CHKERRQ(ierr); */ 3594 ierr = MatSetType(*mat,MATMPIAIJ);CHKERRQ(ierr); 3595 ierr = MatMPIAIJSetPreallocationCSR(*mat,i,j,a);CHKERRQ(ierr); 3596 PetscFunctionReturn(0); 3597 } 3598 3599 #undef __FUNCT__ 3600 #define __FUNCT__ "MatCreateAIJ" 3601 /*@C 3602 MatCreateAIJ - Creates a sparse parallel matrix in AIJ format 3603 (the default parallel PETSc format). For good matrix assembly performance 3604 the user should preallocate the matrix storage by setting the parameters 3605 d_nz (or d_nnz) and o_nz (or o_nnz). By setting these parameters accurately, 3606 performance can be increased by more than a factor of 50. 3607 3608 Collective on MPI_Comm 3609 3610 Input Parameters: 3611 + comm - MPI communicator 3612 . m - number of local rows (or PETSC_DECIDE to have calculated if M is given) 3613 This value should be the same as the local size used in creating the 3614 y vector for the matrix-vector product y = Ax. 3615 . n - This value should be the same as the local size used in creating the 3616 x vector for the matrix-vector product y = Ax. (or PETSC_DECIDE to have 3617 calculated if N is given) For square matrices n is almost always m. 3618 . M - number of global rows (or PETSC_DETERMINE to have calculated if m is given) 3619 . N - number of global columns (or PETSC_DETERMINE to have calculated if n is given) 3620 . d_nz - number of nonzeros per row in DIAGONAL portion of local submatrix 3621 (same value is used for all local rows) 3622 . d_nnz - array containing the number of nonzeros in the various rows of the 3623 DIAGONAL portion of the local submatrix (possibly different for each row) 3624 or NULL, if d_nz is used to specify the nonzero structure. 3625 The size of this array is equal to the number of local rows, i.e 'm'. 3626 . o_nz - number of nonzeros per row in the OFF-DIAGONAL portion of local 3627 submatrix (same value is used for all local rows). 3628 - o_nnz - array containing the number of nonzeros in the various rows of the 3629 OFF-DIAGONAL portion of the local submatrix (possibly different for 3630 each row) or NULL, if o_nz is used to specify the nonzero 3631 structure. The size of this array is equal to the number 3632 of local rows, i.e 'm'. 3633 3634 Output Parameter: 3635 . A - the matrix 3636 3637 It is recommended that one use the MatCreate(), MatSetType() and/or MatSetFromOptions(), 3638 MatXXXXSetPreallocation() paradgm instead of this routine directly. 3639 [MatXXXXSetPreallocation() is, for example, MatSeqAIJSetPreallocation] 3640 3641 Notes: 3642 If the *_nnz parameter is given then the *_nz parameter is ignored 3643 3644 m,n,M,N parameters specify the size of the matrix, and its partitioning across 3645 processors, while d_nz,d_nnz,o_nz,o_nnz parameters specify the approximate 3646 storage requirements for this matrix. 3647 3648 If PETSC_DECIDE or PETSC_DETERMINE is used for a particular argument on one 3649 processor than it must be used on all processors that share the object for 3650 that argument. 3651 3652 The user MUST specify either the local or global matrix dimensions 3653 (possibly both). 3654 3655 The parallel matrix is partitioned across processors such that the 3656 first m0 rows belong to process 0, the next m1 rows belong to 3657 process 1, the next m2 rows belong to process 2 etc.. where 3658 m0,m1,m2,.. are the input parameter 'm'. i.e each processor stores 3659 values corresponding to [m x N] submatrix. 3660 3661 The columns are logically partitioned with the n0 columns belonging 3662 to 0th partition, the next n1 columns belonging to the next 3663 partition etc.. where n0,n1,n2... are the input parameter 'n'. 3664 3665 The DIAGONAL portion of the local submatrix on any given processor 3666 is the submatrix corresponding to the rows and columns m,n 3667 corresponding to the given processor. i.e diagonal matrix on 3668 process 0 is [m0 x n0], diagonal matrix on process 1 is [m1 x n1] 3669 etc. The remaining portion of the local submatrix [m x (N-n)] 3670 constitute the OFF-DIAGONAL portion. The example below better 3671 illustrates this concept. 3672 3673 For a square global matrix we define each processor's diagonal portion 3674 to be its local rows and the corresponding columns (a square submatrix); 3675 each processor's off-diagonal portion encompasses the remainder of the 3676 local matrix (a rectangular submatrix). 3677 3678 If o_nnz, d_nnz are specified, then o_nz, and d_nz are ignored. 3679 3680 When calling this routine with a single process communicator, a matrix of 3681 type SEQAIJ is returned. If a matrix of type MPIAIJ is desired for this 3682 type of communicator, use the construction mechanism: 3683 MatCreate(...,&A); MatSetType(A,MATMPIAIJ); MatSetSizes(A, m,n,M,N); MatMPIAIJSetPreallocation(A,...); 3684 3685 By default, this format uses inodes (identical nodes) when possible. 3686 We search for consecutive rows with the same nonzero structure, thereby 3687 reusing matrix information to achieve increased efficiency. 3688 3689 Options Database Keys: 3690 + -mat_no_inode - Do not use inodes 3691 . -mat_inode_limit <limit> - Sets inode limit (max limit=5) 3692 - -mat_aij_oneindex - Internally use indexing starting at 1 3693 rather than 0. Note that when calling MatSetValues(), 3694 the user still MUST index entries starting at 0! 3695 3696 3697 Example usage: 3698 3699 Consider the following 8x8 matrix with 34 non-zero values, that is 3700 assembled across 3 processors. Lets assume that proc0 owns 3 rows, 3701 proc1 owns 3 rows, proc2 owns 2 rows. This division can be shown 3702 as follows: 3703 3704 .vb 3705 1 2 0 | 0 3 0 | 0 4 3706 Proc0 0 5 6 | 7 0 0 | 8 0 3707 9 0 10 | 11 0 0 | 12 0 3708 ------------------------------------- 3709 13 0 14 | 15 16 17 | 0 0 3710 Proc1 0 18 0 | 19 20 21 | 0 0 3711 0 0 0 | 22 23 0 | 24 0 3712 ------------------------------------- 3713 Proc2 25 26 27 | 0 0 28 | 29 0 3714 30 0 0 | 31 32 33 | 0 34 3715 .ve 3716 3717 This can be represented as a collection of submatrices as: 3718 3719 .vb 3720 A B C 3721 D E F 3722 G H I 3723 .ve 3724 3725 Where the submatrices A,B,C are owned by proc0, D,E,F are 3726 owned by proc1, G,H,I are owned by proc2. 3727 3728 The 'm' parameters for proc0,proc1,proc2 are 3,3,2 respectively. 3729 The 'n' parameters for proc0,proc1,proc2 are 3,3,2 respectively. 3730 The 'M','N' parameters are 8,8, and have the same values on all procs. 3731 3732 The DIAGONAL submatrices corresponding to proc0,proc1,proc2 are 3733 submatrices [A], [E], [I] respectively. The OFF-DIAGONAL submatrices 3734 corresponding to proc0,proc1,proc2 are [BC], [DF], [GH] respectively. 3735 Internally, each processor stores the DIAGONAL part, and the OFF-DIAGONAL 3736 part as SeqAIJ matrices. for eg: proc1 will store [E] as a SeqAIJ 3737 matrix, ans [DF] as another SeqAIJ matrix. 3738 3739 When d_nz, o_nz parameters are specified, d_nz storage elements are 3740 allocated for every row of the local diagonal submatrix, and o_nz 3741 storage locations are allocated for every row of the OFF-DIAGONAL submat. 3742 One way to choose d_nz and o_nz is to use the max nonzerors per local 3743 rows for each of the local DIAGONAL, and the OFF-DIAGONAL submatrices. 3744 In this case, the values of d_nz,o_nz are: 3745 .vb 3746 proc0 : dnz = 2, o_nz = 2 3747 proc1 : dnz = 3, o_nz = 2 3748 proc2 : dnz = 1, o_nz = 4 3749 .ve 3750 We are allocating m*(d_nz+o_nz) storage locations for every proc. This 3751 translates to 3*(2+2)=12 for proc0, 3*(3+2)=15 for proc1, 2*(1+4)=10 3752 for proc3. i.e we are using 12+15+10=37 storage locations to store 3753 34 values. 3754 3755 When d_nnz, o_nnz parameters are specified, the storage is specified 3756 for every row, coresponding to both DIAGONAL and OFF-DIAGONAL submatrices. 3757 In the above case the values for d_nnz,o_nnz are: 3758 .vb 3759 proc0: d_nnz = [2,2,2] and o_nnz = [2,2,2] 3760 proc1: d_nnz = [3,3,2] and o_nnz = [2,1,1] 3761 proc2: d_nnz = [1,1] and o_nnz = [4,4] 3762 .ve 3763 Here the space allocated is sum of all the above values i.e 34, and 3764 hence pre-allocation is perfect. 3765 3766 Level: intermediate 3767 3768 .keywords: matrix, aij, compressed row, sparse, parallel 3769 3770 .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatMPIAIJSetPreallocationCSR(), 3771 MPIAIJ, MatCreateMPIAIJWithArrays() 3772 @*/ 3773 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) 3774 { 3775 PetscErrorCode ierr; 3776 PetscMPIInt size; 3777 3778 PetscFunctionBegin; 3779 ierr = MatCreate(comm,A);CHKERRQ(ierr); 3780 ierr = MatSetSizes(*A,m,n,M,N);CHKERRQ(ierr); 3781 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 3782 if (size > 1) { 3783 ierr = MatSetType(*A,MATMPIAIJ);CHKERRQ(ierr); 3784 ierr = MatMPIAIJSetPreallocation(*A,d_nz,d_nnz,o_nz,o_nnz);CHKERRQ(ierr); 3785 } else { 3786 ierr = MatSetType(*A,MATSEQAIJ);CHKERRQ(ierr); 3787 ierr = MatSeqAIJSetPreallocation(*A,d_nz,d_nnz);CHKERRQ(ierr); 3788 } 3789 PetscFunctionReturn(0); 3790 } 3791 3792 #undef __FUNCT__ 3793 #define __FUNCT__ "MatMPIAIJGetSeqAIJ" 3794 PetscErrorCode MatMPIAIJGetSeqAIJ(Mat A,Mat *Ad,Mat *Ao,const PetscInt *colmap[]) 3795 { 3796 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 3797 3798 PetscFunctionBegin; 3799 if (Ad) *Ad = a->A; 3800 if (Ao) *Ao = a->B; 3801 if (colmap) *colmap = a->garray; 3802 PetscFunctionReturn(0); 3803 } 3804 3805 #undef __FUNCT__ 3806 #define __FUNCT__ "MatSetColoring_MPIAIJ" 3807 PetscErrorCode MatSetColoring_MPIAIJ(Mat A,ISColoring coloring) 3808 { 3809 PetscErrorCode ierr; 3810 PetscInt i; 3811 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 3812 3813 PetscFunctionBegin; 3814 if (coloring->ctype == IS_COLORING_GLOBAL) { 3815 ISColoringValue *allcolors,*colors; 3816 ISColoring ocoloring; 3817 3818 /* set coloring for diagonal portion */ 3819 ierr = MatSetColoring_SeqAIJ(a->A,coloring);CHKERRQ(ierr); 3820 3821 /* set coloring for off-diagonal portion */ 3822 ierr = ISAllGatherColors(PetscObjectComm((PetscObject)A),coloring->n,coloring->colors,NULL,&allcolors);CHKERRQ(ierr); 3823 ierr = PetscMalloc1((a->B->cmap->n+1),&colors);CHKERRQ(ierr); 3824 for (i=0; i<a->B->cmap->n; i++) { 3825 colors[i] = allcolors[a->garray[i]]; 3826 } 3827 ierr = PetscFree(allcolors);CHKERRQ(ierr); 3828 ierr = ISColoringCreate(MPI_COMM_SELF,coloring->n,a->B->cmap->n,colors,&ocoloring);CHKERRQ(ierr); 3829 ierr = MatSetColoring_SeqAIJ(a->B,ocoloring);CHKERRQ(ierr); 3830 ierr = ISColoringDestroy(&ocoloring);CHKERRQ(ierr); 3831 } else if (coloring->ctype == IS_COLORING_GHOSTED) { 3832 ISColoringValue *colors; 3833 PetscInt *larray; 3834 ISColoring ocoloring; 3835 3836 /* set coloring for diagonal portion */ 3837 ierr = PetscMalloc1((a->A->cmap->n+1),&larray);CHKERRQ(ierr); 3838 for (i=0; i<a->A->cmap->n; i++) { 3839 larray[i] = i + A->cmap->rstart; 3840 } 3841 ierr = ISGlobalToLocalMappingApply(A->cmap->mapping,IS_GTOLM_MASK,a->A->cmap->n,larray,NULL,larray);CHKERRQ(ierr); 3842 ierr = PetscMalloc1((a->A->cmap->n+1),&colors);CHKERRQ(ierr); 3843 for (i=0; i<a->A->cmap->n; i++) { 3844 colors[i] = coloring->colors[larray[i]]; 3845 } 3846 ierr = PetscFree(larray);CHKERRQ(ierr); 3847 ierr = ISColoringCreate(PETSC_COMM_SELF,coloring->n,a->A->cmap->n,colors,&ocoloring);CHKERRQ(ierr); 3848 ierr = MatSetColoring_SeqAIJ(a->A,ocoloring);CHKERRQ(ierr); 3849 ierr = ISColoringDestroy(&ocoloring);CHKERRQ(ierr); 3850 3851 /* set coloring for off-diagonal portion */ 3852 ierr = PetscMalloc1((a->B->cmap->n+1),&larray);CHKERRQ(ierr); 3853 ierr = ISGlobalToLocalMappingApply(A->cmap->mapping,IS_GTOLM_MASK,a->B->cmap->n,a->garray,NULL,larray);CHKERRQ(ierr); 3854 ierr = PetscMalloc1((a->B->cmap->n+1),&colors);CHKERRQ(ierr); 3855 for (i=0; i<a->B->cmap->n; i++) { 3856 colors[i] = coloring->colors[larray[i]]; 3857 } 3858 ierr = PetscFree(larray);CHKERRQ(ierr); 3859 ierr = ISColoringCreate(MPI_COMM_SELF,coloring->n,a->B->cmap->n,colors,&ocoloring);CHKERRQ(ierr); 3860 ierr = MatSetColoring_SeqAIJ(a->B,ocoloring);CHKERRQ(ierr); 3861 ierr = ISColoringDestroy(&ocoloring);CHKERRQ(ierr); 3862 } else SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"No support ISColoringType %d",(int)coloring->ctype); 3863 PetscFunctionReturn(0); 3864 } 3865 3866 #undef __FUNCT__ 3867 #define __FUNCT__ "MatSetValuesAdifor_MPIAIJ" 3868 PetscErrorCode MatSetValuesAdifor_MPIAIJ(Mat A,PetscInt nl,void *advalues) 3869 { 3870 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 3871 PetscErrorCode ierr; 3872 3873 PetscFunctionBegin; 3874 ierr = MatSetValuesAdifor_SeqAIJ(a->A,nl,advalues);CHKERRQ(ierr); 3875 ierr = MatSetValuesAdifor_SeqAIJ(a->B,nl,advalues);CHKERRQ(ierr); 3876 PetscFunctionReturn(0); 3877 } 3878 3879 #undef __FUNCT__ 3880 #define __FUNCT__ "MatCreateMPIAIJConcatenateSeqAIJSymbolic" 3881 PetscErrorCode MatCreateMPIAIJConcatenateSeqAIJSymbolic(MPI_Comm comm,Mat inmat,PetscInt n,Mat *outmat) 3882 { 3883 PetscErrorCode ierr; 3884 PetscInt m,N,i,rstart,nnz,*dnz,*onz,sum,bs,cbs; 3885 PetscInt *indx; 3886 3887 PetscFunctionBegin; 3888 /* This routine will ONLY return MPIAIJ type matrix */ 3889 ierr = MatGetSize(inmat,&m,&N);CHKERRQ(ierr); 3890 ierr = MatGetBlockSizes(inmat,&bs,&cbs);CHKERRQ(ierr); 3891 if (n == PETSC_DECIDE) { 3892 ierr = PetscSplitOwnership(comm,&n,&N);CHKERRQ(ierr); 3893 } 3894 /* Check sum(n) = N */ 3895 ierr = MPI_Allreduce(&n,&sum,1,MPIU_INT,MPI_SUM,comm);CHKERRQ(ierr); 3896 if (sum != N) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_INCOMP,"Sum of local columns != global columns %d",N); 3897 3898 ierr = MPI_Scan(&m, &rstart,1,MPIU_INT,MPI_SUM,comm);CHKERRQ(ierr); 3899 rstart -= m; 3900 3901 ierr = MatPreallocateInitialize(comm,m,n,dnz,onz);CHKERRQ(ierr); 3902 for (i=0; i<m; i++) { 3903 ierr = MatGetRow_SeqAIJ(inmat,i,&nnz,&indx,NULL);CHKERRQ(ierr); 3904 ierr = MatPreallocateSet(i+rstart,nnz,indx,dnz,onz);CHKERRQ(ierr); 3905 ierr = MatRestoreRow_SeqAIJ(inmat,i,&nnz,&indx,NULL);CHKERRQ(ierr); 3906 } 3907 3908 ierr = MatCreate(comm,outmat);CHKERRQ(ierr); 3909 ierr = MatSetSizes(*outmat,m,n,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 3910 ierr = MatSetBlockSizes(*outmat,bs,cbs);CHKERRQ(ierr); 3911 ierr = MatSetType(*outmat,MATMPIAIJ);CHKERRQ(ierr); 3912 ierr = MatMPIAIJSetPreallocation(*outmat,0,dnz,0,onz);CHKERRQ(ierr); 3913 ierr = MatPreallocateFinalize(dnz,onz);CHKERRQ(ierr); 3914 PetscFunctionReturn(0); 3915 } 3916 3917 #undef __FUNCT__ 3918 #define __FUNCT__ "MatCreateMPIAIJConcatenateSeqAIJNumeric" 3919 PetscErrorCode MatCreateMPIAIJConcatenateSeqAIJNumeric(MPI_Comm comm,Mat inmat,PetscInt n,Mat outmat) 3920 { 3921 PetscErrorCode ierr; 3922 PetscInt m,N,i,rstart,nnz,Ii; 3923 PetscInt *indx; 3924 PetscScalar *values; 3925 3926 PetscFunctionBegin; 3927 ierr = MatGetSize(inmat,&m,&N);CHKERRQ(ierr); 3928 ierr = MatGetOwnershipRange(outmat,&rstart,NULL);CHKERRQ(ierr); 3929 for (i=0; i<m; i++) { 3930 ierr = MatGetRow_SeqAIJ(inmat,i,&nnz,&indx,&values);CHKERRQ(ierr); 3931 Ii = i + rstart; 3932 ierr = MatSetValues(outmat,1,&Ii,nnz,indx,values,INSERT_VALUES);CHKERRQ(ierr); 3933 ierr = MatRestoreRow_SeqAIJ(inmat,i,&nnz,&indx,&values);CHKERRQ(ierr); 3934 } 3935 ierr = MatAssemblyBegin(outmat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3936 ierr = MatAssemblyEnd(outmat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3937 PetscFunctionReturn(0); 3938 } 3939 3940 #undef __FUNCT__ 3941 #define __FUNCT__ "MatCreateMPIMatConcatenateSeqMat_MPIAIJ" 3942 PetscErrorCode MatCreateMPIMatConcatenateSeqMat_MPIAIJ(MPI_Comm comm,Mat inmat,PetscInt n,MatReuse scall,Mat *outmat) 3943 { 3944 PetscErrorCode ierr; 3945 PetscMPIInt size; 3946 3947 PetscFunctionBegin; 3948 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 3949 ierr = PetscLogEventBegin(MAT_Merge,inmat,0,0,0);CHKERRQ(ierr); 3950 if (size == 1) { 3951 if (scall == MAT_INITIAL_MATRIX) { 3952 ierr = MatDuplicate(inmat,MAT_COPY_VALUES,outmat);CHKERRQ(ierr); 3953 } else { 3954 ierr = MatCopy(inmat,*outmat,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 3955 } 3956 } else { 3957 if (scall == MAT_INITIAL_MATRIX) { 3958 ierr = MatCreateMPIAIJConcatenateSeqAIJSymbolic(comm,inmat,n,outmat);CHKERRQ(ierr); 3959 } 3960 ierr = MatCreateMPIAIJConcatenateSeqAIJNumeric(comm,inmat,n,*outmat);CHKERRQ(ierr); 3961 } 3962 ierr = PetscLogEventEnd(MAT_Merge,inmat,0,0,0);CHKERRQ(ierr); 3963 PetscFunctionReturn(0); 3964 } 3965 3966 #undef __FUNCT__ 3967 #define __FUNCT__ "MatFileSplit" 3968 PetscErrorCode MatFileSplit(Mat A,char *outfile) 3969 { 3970 PetscErrorCode ierr; 3971 PetscMPIInt rank; 3972 PetscInt m,N,i,rstart,nnz; 3973 size_t len; 3974 const PetscInt *indx; 3975 PetscViewer out; 3976 char *name; 3977 Mat B; 3978 const PetscScalar *values; 3979 3980 PetscFunctionBegin; 3981 ierr = MatGetLocalSize(A,&m,0);CHKERRQ(ierr); 3982 ierr = MatGetSize(A,0,&N);CHKERRQ(ierr); 3983 /* Should this be the type of the diagonal block of A? */ 3984 ierr = MatCreate(PETSC_COMM_SELF,&B);CHKERRQ(ierr); 3985 ierr = MatSetSizes(B,m,N,m,N);CHKERRQ(ierr); 3986 ierr = MatSetBlockSizesFromMats(B,A,A);CHKERRQ(ierr); 3987 ierr = MatSetType(B,MATSEQAIJ);CHKERRQ(ierr); 3988 ierr = MatSeqAIJSetPreallocation(B,0,NULL);CHKERRQ(ierr); 3989 ierr = MatGetOwnershipRange(A,&rstart,0);CHKERRQ(ierr); 3990 for (i=0; i<m; i++) { 3991 ierr = MatGetRow(A,i+rstart,&nnz,&indx,&values);CHKERRQ(ierr); 3992 ierr = MatSetValues(B,1,&i,nnz,indx,values,INSERT_VALUES);CHKERRQ(ierr); 3993 ierr = MatRestoreRow(A,i+rstart,&nnz,&indx,&values);CHKERRQ(ierr); 3994 } 3995 ierr = MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3996 ierr = MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3997 3998 ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)A),&rank);CHKERRQ(ierr); 3999 ierr = PetscStrlen(outfile,&len);CHKERRQ(ierr); 4000 ierr = PetscMalloc1((len+5),&name);CHKERRQ(ierr); 4001 sprintf(name,"%s.%d",outfile,rank); 4002 ierr = PetscViewerBinaryOpen(PETSC_COMM_SELF,name,FILE_MODE_APPEND,&out);CHKERRQ(ierr); 4003 ierr = PetscFree(name);CHKERRQ(ierr); 4004 ierr = MatView(B,out);CHKERRQ(ierr); 4005 ierr = PetscViewerDestroy(&out);CHKERRQ(ierr); 4006 ierr = MatDestroy(&B);CHKERRQ(ierr); 4007 PetscFunctionReturn(0); 4008 } 4009 4010 extern PetscErrorCode MatDestroy_MPIAIJ(Mat); 4011 #undef __FUNCT__ 4012 #define __FUNCT__ "MatDestroy_MPIAIJ_SeqsToMPI" 4013 PetscErrorCode MatDestroy_MPIAIJ_SeqsToMPI(Mat A) 4014 { 4015 PetscErrorCode ierr; 4016 Mat_Merge_SeqsToMPI *merge; 4017 PetscContainer container; 4018 4019 PetscFunctionBegin; 4020 ierr = PetscObjectQuery((PetscObject)A,"MatMergeSeqsToMPI",(PetscObject*)&container);CHKERRQ(ierr); 4021 if (container) { 4022 ierr = PetscContainerGetPointer(container,(void**)&merge);CHKERRQ(ierr); 4023 ierr = PetscFree(merge->id_r);CHKERRQ(ierr); 4024 ierr = PetscFree(merge->len_s);CHKERRQ(ierr); 4025 ierr = PetscFree(merge->len_r);CHKERRQ(ierr); 4026 ierr = PetscFree(merge->bi);CHKERRQ(ierr); 4027 ierr = PetscFree(merge->bj);CHKERRQ(ierr); 4028 ierr = PetscFree(merge->buf_ri[0]);CHKERRQ(ierr); 4029 ierr = PetscFree(merge->buf_ri);CHKERRQ(ierr); 4030 ierr = PetscFree(merge->buf_rj[0]);CHKERRQ(ierr); 4031 ierr = PetscFree(merge->buf_rj);CHKERRQ(ierr); 4032 ierr = PetscFree(merge->coi);CHKERRQ(ierr); 4033 ierr = PetscFree(merge->coj);CHKERRQ(ierr); 4034 ierr = PetscFree(merge->owners_co);CHKERRQ(ierr); 4035 ierr = PetscLayoutDestroy(&merge->rowmap);CHKERRQ(ierr); 4036 ierr = PetscFree(merge);CHKERRQ(ierr); 4037 ierr = PetscObjectCompose((PetscObject)A,"MatMergeSeqsToMPI",0);CHKERRQ(ierr); 4038 } 4039 ierr = MatDestroy_MPIAIJ(A);CHKERRQ(ierr); 4040 PetscFunctionReturn(0); 4041 } 4042 4043 #include <../src/mat/utils/freespace.h> 4044 #include <petscbt.h> 4045 4046 #undef __FUNCT__ 4047 #define __FUNCT__ "MatCreateMPIAIJSumSeqAIJNumeric" 4048 PetscErrorCode MatCreateMPIAIJSumSeqAIJNumeric(Mat seqmat,Mat mpimat) 4049 { 4050 PetscErrorCode ierr; 4051 MPI_Comm comm; 4052 Mat_SeqAIJ *a =(Mat_SeqAIJ*)seqmat->data; 4053 PetscMPIInt size,rank,taga,*len_s; 4054 PetscInt N=mpimat->cmap->N,i,j,*owners,*ai=a->i,*aj; 4055 PetscInt proc,m; 4056 PetscInt **buf_ri,**buf_rj; 4057 PetscInt k,anzi,*bj_i,*bi,*bj,arow,bnzi,nextaj; 4058 PetscInt nrows,**buf_ri_k,**nextrow,**nextai; 4059 MPI_Request *s_waits,*r_waits; 4060 MPI_Status *status; 4061 MatScalar *aa=a->a; 4062 MatScalar **abuf_r,*ba_i; 4063 Mat_Merge_SeqsToMPI *merge; 4064 PetscContainer container; 4065 4066 PetscFunctionBegin; 4067 ierr = PetscObjectGetComm((PetscObject)mpimat,&comm);CHKERRQ(ierr); 4068 ierr = PetscLogEventBegin(MAT_Seqstompinum,seqmat,0,0,0);CHKERRQ(ierr); 4069 4070 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 4071 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 4072 4073 ierr = PetscObjectQuery((PetscObject)mpimat,"MatMergeSeqsToMPI",(PetscObject*)&container);CHKERRQ(ierr); 4074 ierr = PetscContainerGetPointer(container,(void**)&merge);CHKERRQ(ierr); 4075 4076 bi = merge->bi; 4077 bj = merge->bj; 4078 buf_ri = merge->buf_ri; 4079 buf_rj = merge->buf_rj; 4080 4081 ierr = PetscMalloc1(size,&status);CHKERRQ(ierr); 4082 owners = merge->rowmap->range; 4083 len_s = merge->len_s; 4084 4085 /* send and recv matrix values */ 4086 /*-----------------------------*/ 4087 ierr = PetscObjectGetNewTag((PetscObject)mpimat,&taga);CHKERRQ(ierr); 4088 ierr = PetscPostIrecvScalar(comm,taga,merge->nrecv,merge->id_r,merge->len_r,&abuf_r,&r_waits);CHKERRQ(ierr); 4089 4090 ierr = PetscMalloc1((merge->nsend+1),&s_waits);CHKERRQ(ierr); 4091 for (proc=0,k=0; proc<size; proc++) { 4092 if (!len_s[proc]) continue; 4093 i = owners[proc]; 4094 ierr = MPI_Isend(aa+ai[i],len_s[proc],MPIU_MATSCALAR,proc,taga,comm,s_waits+k);CHKERRQ(ierr); 4095 k++; 4096 } 4097 4098 if (merge->nrecv) {ierr = MPI_Waitall(merge->nrecv,r_waits,status);CHKERRQ(ierr);} 4099 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,s_waits,status);CHKERRQ(ierr);} 4100 ierr = PetscFree(status);CHKERRQ(ierr); 4101 4102 ierr = PetscFree(s_waits);CHKERRQ(ierr); 4103 ierr = PetscFree(r_waits);CHKERRQ(ierr); 4104 4105 /* insert mat values of mpimat */ 4106 /*----------------------------*/ 4107 ierr = PetscMalloc1(N,&ba_i);CHKERRQ(ierr); 4108 ierr = PetscMalloc3(merge->nrecv,&buf_ri_k,merge->nrecv,&nextrow,merge->nrecv,&nextai);CHKERRQ(ierr); 4109 4110 for (k=0; k<merge->nrecv; k++) { 4111 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 4112 nrows = *(buf_ri_k[k]); 4113 nextrow[k] = buf_ri_k[k]+1; /* next row number of k-th recved i-structure */ 4114 nextai[k] = buf_ri_k[k] + (nrows + 1); /* poins to the next i-structure of k-th recved i-structure */ 4115 } 4116 4117 /* set values of ba */ 4118 m = merge->rowmap->n; 4119 for (i=0; i<m; i++) { 4120 arow = owners[rank] + i; 4121 bj_i = bj+bi[i]; /* col indices of the i-th row of mpimat */ 4122 bnzi = bi[i+1] - bi[i]; 4123 ierr = PetscMemzero(ba_i,bnzi*sizeof(PetscScalar));CHKERRQ(ierr); 4124 4125 /* add local non-zero vals of this proc's seqmat into ba */ 4126 anzi = ai[arow+1] - ai[arow]; 4127 aj = a->j + ai[arow]; 4128 aa = a->a + ai[arow]; 4129 nextaj = 0; 4130 for (j=0; nextaj<anzi; j++) { 4131 if (*(bj_i + j) == aj[nextaj]) { /* bcol == acol */ 4132 ba_i[j] += aa[nextaj++]; 4133 } 4134 } 4135 4136 /* add received vals into ba */ 4137 for (k=0; k<merge->nrecv; k++) { /* k-th received message */ 4138 /* i-th row */ 4139 if (i == *nextrow[k]) { 4140 anzi = *(nextai[k]+1) - *nextai[k]; 4141 aj = buf_rj[k] + *(nextai[k]); 4142 aa = abuf_r[k] + *(nextai[k]); 4143 nextaj = 0; 4144 for (j=0; nextaj<anzi; j++) { 4145 if (*(bj_i + j) == aj[nextaj]) { /* bcol == acol */ 4146 ba_i[j] += aa[nextaj++]; 4147 } 4148 } 4149 nextrow[k]++; nextai[k]++; 4150 } 4151 } 4152 ierr = MatSetValues(mpimat,1,&arow,bnzi,bj_i,ba_i,INSERT_VALUES);CHKERRQ(ierr); 4153 } 4154 ierr = MatAssemblyBegin(mpimat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4155 ierr = MatAssemblyEnd(mpimat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4156 4157 ierr = PetscFree(abuf_r[0]);CHKERRQ(ierr); 4158 ierr = PetscFree(abuf_r);CHKERRQ(ierr); 4159 ierr = PetscFree(ba_i);CHKERRQ(ierr); 4160 ierr = PetscFree3(buf_ri_k,nextrow,nextai);CHKERRQ(ierr); 4161 ierr = PetscLogEventEnd(MAT_Seqstompinum,seqmat,0,0,0);CHKERRQ(ierr); 4162 PetscFunctionReturn(0); 4163 } 4164 4165 extern PetscErrorCode MatDestroy_MPIAIJ_SeqsToMPI(Mat); 4166 4167 #undef __FUNCT__ 4168 #define __FUNCT__ "MatCreateMPIAIJSumSeqAIJSymbolic" 4169 PetscErrorCode MatCreateMPIAIJSumSeqAIJSymbolic(MPI_Comm comm,Mat seqmat,PetscInt m,PetscInt n,Mat *mpimat) 4170 { 4171 PetscErrorCode ierr; 4172 Mat B_mpi; 4173 Mat_SeqAIJ *a=(Mat_SeqAIJ*)seqmat->data; 4174 PetscMPIInt size,rank,tagi,tagj,*len_s,*len_si,*len_ri; 4175 PetscInt **buf_rj,**buf_ri,**buf_ri_k; 4176 PetscInt M=seqmat->rmap->n,N=seqmat->cmap->n,i,*owners,*ai=a->i,*aj=a->j; 4177 PetscInt len,proc,*dnz,*onz,bs,cbs; 4178 PetscInt k,anzi,*bi,*bj,*lnk,nlnk,arow,bnzi,nspacedouble=0; 4179 PetscInt nrows,*buf_s,*buf_si,*buf_si_i,**nextrow,**nextai; 4180 MPI_Request *si_waits,*sj_waits,*ri_waits,*rj_waits; 4181 MPI_Status *status; 4182 PetscFreeSpaceList free_space=NULL,current_space=NULL; 4183 PetscBT lnkbt; 4184 Mat_Merge_SeqsToMPI *merge; 4185 PetscContainer container; 4186 4187 PetscFunctionBegin; 4188 ierr = PetscLogEventBegin(MAT_Seqstompisym,seqmat,0,0,0);CHKERRQ(ierr); 4189 4190 /* make sure it is a PETSc comm */ 4191 ierr = PetscCommDuplicate(comm,&comm,NULL);CHKERRQ(ierr); 4192 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 4193 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 4194 4195 ierr = PetscNew(&merge);CHKERRQ(ierr); 4196 ierr = PetscMalloc1(size,&status);CHKERRQ(ierr); 4197 4198 /* determine row ownership */ 4199 /*---------------------------------------------------------*/ 4200 ierr = PetscLayoutCreate(comm,&merge->rowmap);CHKERRQ(ierr); 4201 ierr = PetscLayoutSetLocalSize(merge->rowmap,m);CHKERRQ(ierr); 4202 ierr = PetscLayoutSetSize(merge->rowmap,M);CHKERRQ(ierr); 4203 ierr = PetscLayoutSetBlockSize(merge->rowmap,1);CHKERRQ(ierr); 4204 ierr = PetscLayoutSetUp(merge->rowmap);CHKERRQ(ierr); 4205 ierr = PetscMalloc1(size,&len_si);CHKERRQ(ierr); 4206 ierr = PetscMalloc1(size,&merge->len_s);CHKERRQ(ierr); 4207 4208 m = merge->rowmap->n; 4209 owners = merge->rowmap->range; 4210 4211 /* determine the number of messages to send, their lengths */ 4212 /*---------------------------------------------------------*/ 4213 len_s = merge->len_s; 4214 4215 len = 0; /* length of buf_si[] */ 4216 merge->nsend = 0; 4217 for (proc=0; proc<size; proc++) { 4218 len_si[proc] = 0; 4219 if (proc == rank) { 4220 len_s[proc] = 0; 4221 } else { 4222 len_si[proc] = owners[proc+1] - owners[proc] + 1; 4223 len_s[proc] = ai[owners[proc+1]] - ai[owners[proc]]; /* num of rows to be sent to [proc] */ 4224 } 4225 if (len_s[proc]) { 4226 merge->nsend++; 4227 nrows = 0; 4228 for (i=owners[proc]; i<owners[proc+1]; i++) { 4229 if (ai[i+1] > ai[i]) nrows++; 4230 } 4231 len_si[proc] = 2*(nrows+1); 4232 len += len_si[proc]; 4233 } 4234 } 4235 4236 /* determine the number and length of messages to receive for ij-structure */ 4237 /*-------------------------------------------------------------------------*/ 4238 ierr = PetscGatherNumberOfMessages(comm,NULL,len_s,&merge->nrecv);CHKERRQ(ierr); 4239 ierr = PetscGatherMessageLengths2(comm,merge->nsend,merge->nrecv,len_s,len_si,&merge->id_r,&merge->len_r,&len_ri);CHKERRQ(ierr); 4240 4241 /* post the Irecv of j-structure */ 4242 /*-------------------------------*/ 4243 ierr = PetscCommGetNewTag(comm,&tagj);CHKERRQ(ierr); 4244 ierr = PetscPostIrecvInt(comm,tagj,merge->nrecv,merge->id_r,merge->len_r,&buf_rj,&rj_waits);CHKERRQ(ierr); 4245 4246 /* post the Isend of j-structure */ 4247 /*--------------------------------*/ 4248 ierr = PetscMalloc2(merge->nsend,&si_waits,merge->nsend,&sj_waits);CHKERRQ(ierr); 4249 4250 for (proc=0, k=0; proc<size; proc++) { 4251 if (!len_s[proc]) continue; 4252 i = owners[proc]; 4253 ierr = MPI_Isend(aj+ai[i],len_s[proc],MPIU_INT,proc,tagj,comm,sj_waits+k);CHKERRQ(ierr); 4254 k++; 4255 } 4256 4257 /* receives and sends of j-structure are complete */ 4258 /*------------------------------------------------*/ 4259 if (merge->nrecv) {ierr = MPI_Waitall(merge->nrecv,rj_waits,status);CHKERRQ(ierr);} 4260 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,sj_waits,status);CHKERRQ(ierr);} 4261 4262 /* send and recv i-structure */ 4263 /*---------------------------*/ 4264 ierr = PetscCommGetNewTag(comm,&tagi);CHKERRQ(ierr); 4265 ierr = PetscPostIrecvInt(comm,tagi,merge->nrecv,merge->id_r,len_ri,&buf_ri,&ri_waits);CHKERRQ(ierr); 4266 4267 ierr = PetscMalloc1((len+1),&buf_s);CHKERRQ(ierr); 4268 buf_si = buf_s; /* points to the beginning of k-th msg to be sent */ 4269 for (proc=0,k=0; proc<size; proc++) { 4270 if (!len_s[proc]) continue; 4271 /* form outgoing message for i-structure: 4272 buf_si[0]: nrows to be sent 4273 [1:nrows]: row index (global) 4274 [nrows+1:2*nrows+1]: i-structure index 4275 */ 4276 /*-------------------------------------------*/ 4277 nrows = len_si[proc]/2 - 1; 4278 buf_si_i = buf_si + nrows+1; 4279 buf_si[0] = nrows; 4280 buf_si_i[0] = 0; 4281 nrows = 0; 4282 for (i=owners[proc]; i<owners[proc+1]; i++) { 4283 anzi = ai[i+1] - ai[i]; 4284 if (anzi) { 4285 buf_si_i[nrows+1] = buf_si_i[nrows] + anzi; /* i-structure */ 4286 buf_si[nrows+1] = i-owners[proc]; /* local row index */ 4287 nrows++; 4288 } 4289 } 4290 ierr = MPI_Isend(buf_si,len_si[proc],MPIU_INT,proc,tagi,comm,si_waits+k);CHKERRQ(ierr); 4291 k++; 4292 buf_si += len_si[proc]; 4293 } 4294 4295 if (merge->nrecv) {ierr = MPI_Waitall(merge->nrecv,ri_waits,status);CHKERRQ(ierr);} 4296 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,si_waits,status);CHKERRQ(ierr);} 4297 4298 ierr = PetscInfo2(seqmat,"nsend: %D, nrecv: %D\n",merge->nsend,merge->nrecv);CHKERRQ(ierr); 4299 for (i=0; i<merge->nrecv; i++) { 4300 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); 4301 } 4302 4303 ierr = PetscFree(len_si);CHKERRQ(ierr); 4304 ierr = PetscFree(len_ri);CHKERRQ(ierr); 4305 ierr = PetscFree(rj_waits);CHKERRQ(ierr); 4306 ierr = PetscFree2(si_waits,sj_waits);CHKERRQ(ierr); 4307 ierr = PetscFree(ri_waits);CHKERRQ(ierr); 4308 ierr = PetscFree(buf_s);CHKERRQ(ierr); 4309 ierr = PetscFree(status);CHKERRQ(ierr); 4310 4311 /* compute a local seq matrix in each processor */ 4312 /*----------------------------------------------*/ 4313 /* allocate bi array and free space for accumulating nonzero column info */ 4314 ierr = PetscMalloc1((m+1),&bi);CHKERRQ(ierr); 4315 bi[0] = 0; 4316 4317 /* create and initialize a linked list */ 4318 nlnk = N+1; 4319 ierr = PetscLLCreate(N,N,nlnk,lnk,lnkbt);CHKERRQ(ierr); 4320 4321 /* initial FreeSpace size is 2*(num of local nnz(seqmat)) */ 4322 len = ai[owners[rank+1]] - ai[owners[rank]]; 4323 ierr = PetscFreeSpaceGet((PetscInt)(2*len+1),&free_space);CHKERRQ(ierr); 4324 4325 current_space = free_space; 4326 4327 /* determine symbolic info for each local row */ 4328 ierr = PetscMalloc3(merge->nrecv,&buf_ri_k,merge->nrecv,&nextrow,merge->nrecv,&nextai);CHKERRQ(ierr); 4329 4330 for (k=0; k<merge->nrecv; k++) { 4331 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 4332 nrows = *buf_ri_k[k]; 4333 nextrow[k] = buf_ri_k[k] + 1; /* next row number of k-th recved i-structure */ 4334 nextai[k] = buf_ri_k[k] + (nrows + 1); /* poins to the next i-structure of k-th recved i-structure */ 4335 } 4336 4337 ierr = MatPreallocateInitialize(comm,m,n,dnz,onz);CHKERRQ(ierr); 4338 len = 0; 4339 for (i=0; i<m; i++) { 4340 bnzi = 0; 4341 /* add local non-zero cols of this proc's seqmat into lnk */ 4342 arow = owners[rank] + i; 4343 anzi = ai[arow+1] - ai[arow]; 4344 aj = a->j + ai[arow]; 4345 ierr = PetscLLAddSorted(anzi,aj,N,nlnk,lnk,lnkbt);CHKERRQ(ierr); 4346 bnzi += nlnk; 4347 /* add received col data into lnk */ 4348 for (k=0; k<merge->nrecv; k++) { /* k-th received message */ 4349 if (i == *nextrow[k]) { /* i-th row */ 4350 anzi = *(nextai[k]+1) - *nextai[k]; 4351 aj = buf_rj[k] + *nextai[k]; 4352 ierr = PetscLLAddSorted(anzi,aj,N,nlnk,lnk,lnkbt);CHKERRQ(ierr); 4353 bnzi += nlnk; 4354 nextrow[k]++; nextai[k]++; 4355 } 4356 } 4357 if (len < bnzi) len = bnzi; /* =max(bnzi) */ 4358 4359 /* if free space is not available, make more free space */ 4360 if (current_space->local_remaining<bnzi) { 4361 ierr = PetscFreeSpaceGet(bnzi+current_space->total_array_size,¤t_space);CHKERRQ(ierr); 4362 nspacedouble++; 4363 } 4364 /* copy data into free space, then initialize lnk */ 4365 ierr = PetscLLClean(N,N,bnzi,lnk,current_space->array,lnkbt);CHKERRQ(ierr); 4366 ierr = MatPreallocateSet(i+owners[rank],bnzi,current_space->array,dnz,onz);CHKERRQ(ierr); 4367 4368 current_space->array += bnzi; 4369 current_space->local_used += bnzi; 4370 current_space->local_remaining -= bnzi; 4371 4372 bi[i+1] = bi[i] + bnzi; 4373 } 4374 4375 ierr = PetscFree3(buf_ri_k,nextrow,nextai);CHKERRQ(ierr); 4376 4377 ierr = PetscMalloc1((bi[m]+1),&bj);CHKERRQ(ierr); 4378 ierr = PetscFreeSpaceContiguous(&free_space,bj);CHKERRQ(ierr); 4379 ierr = PetscLLDestroy(lnk,lnkbt);CHKERRQ(ierr); 4380 4381 /* create symbolic parallel matrix B_mpi */ 4382 /*---------------------------------------*/ 4383 ierr = MatGetBlockSizes(seqmat,&bs,&cbs);CHKERRQ(ierr); 4384 ierr = MatCreate(comm,&B_mpi);CHKERRQ(ierr); 4385 if (n==PETSC_DECIDE) { 4386 ierr = MatSetSizes(B_mpi,m,n,PETSC_DETERMINE,N);CHKERRQ(ierr); 4387 } else { 4388 ierr = MatSetSizes(B_mpi,m,n,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 4389 } 4390 ierr = MatSetBlockSizes(B_mpi,bs,cbs);CHKERRQ(ierr); 4391 ierr = MatSetType(B_mpi,MATMPIAIJ);CHKERRQ(ierr); 4392 ierr = MatMPIAIJSetPreallocation(B_mpi,0,dnz,0,onz);CHKERRQ(ierr); 4393 ierr = MatPreallocateFinalize(dnz,onz);CHKERRQ(ierr); 4394 ierr = MatSetOption(B_mpi,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 4395 4396 /* B_mpi is not ready for use - assembly will be done by MatCreateMPIAIJSumSeqAIJNumeric() */ 4397 B_mpi->assembled = PETSC_FALSE; 4398 B_mpi->ops->destroy = MatDestroy_MPIAIJ_SeqsToMPI; 4399 merge->bi = bi; 4400 merge->bj = bj; 4401 merge->buf_ri = buf_ri; 4402 merge->buf_rj = buf_rj; 4403 merge->coi = NULL; 4404 merge->coj = NULL; 4405 merge->owners_co = NULL; 4406 4407 ierr = PetscCommDestroy(&comm);CHKERRQ(ierr); 4408 4409 /* attach the supporting struct to B_mpi for reuse */ 4410 ierr = PetscContainerCreate(PETSC_COMM_SELF,&container);CHKERRQ(ierr); 4411 ierr = PetscContainerSetPointer(container,merge);CHKERRQ(ierr); 4412 ierr = PetscObjectCompose((PetscObject)B_mpi,"MatMergeSeqsToMPI",(PetscObject)container);CHKERRQ(ierr); 4413 ierr = PetscContainerDestroy(&container);CHKERRQ(ierr); 4414 *mpimat = B_mpi; 4415 4416 ierr = PetscLogEventEnd(MAT_Seqstompisym,seqmat,0,0,0);CHKERRQ(ierr); 4417 PetscFunctionReturn(0); 4418 } 4419 4420 #undef __FUNCT__ 4421 #define __FUNCT__ "MatCreateMPIAIJSumSeqAIJ" 4422 /*@C 4423 MatCreateMPIAIJSumSeqAIJ - Creates a MPIAIJ matrix by adding sequential 4424 matrices from each processor 4425 4426 Collective on MPI_Comm 4427 4428 Input Parameters: 4429 + comm - the communicators the parallel matrix will live on 4430 . seqmat - the input sequential matrices 4431 . m - number of local rows (or PETSC_DECIDE) 4432 . n - number of local columns (or PETSC_DECIDE) 4433 - scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX 4434 4435 Output Parameter: 4436 . mpimat - the parallel matrix generated 4437 4438 Level: advanced 4439 4440 Notes: 4441 The dimensions of the sequential matrix in each processor MUST be the same. 4442 The input seqmat is included into the container "Mat_Merge_SeqsToMPI", and will be 4443 destroyed when mpimat is destroyed. Call PetscObjectQuery() to access seqmat. 4444 @*/ 4445 PetscErrorCode MatCreateMPIAIJSumSeqAIJ(MPI_Comm comm,Mat seqmat,PetscInt m,PetscInt n,MatReuse scall,Mat *mpimat) 4446 { 4447 PetscErrorCode ierr; 4448 PetscMPIInt size; 4449 4450 PetscFunctionBegin; 4451 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 4452 if (size == 1) { 4453 ierr = PetscLogEventBegin(MAT_Seqstompi,seqmat,0,0,0);CHKERRQ(ierr); 4454 if (scall == MAT_INITIAL_MATRIX) { 4455 ierr = MatDuplicate(seqmat,MAT_COPY_VALUES,mpimat);CHKERRQ(ierr); 4456 } else { 4457 ierr = MatCopy(seqmat,*mpimat,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 4458 } 4459 ierr = PetscLogEventEnd(MAT_Seqstompi,seqmat,0,0,0);CHKERRQ(ierr); 4460 PetscFunctionReturn(0); 4461 } 4462 ierr = PetscLogEventBegin(MAT_Seqstompi,seqmat,0,0,0);CHKERRQ(ierr); 4463 if (scall == MAT_INITIAL_MATRIX) { 4464 ierr = MatCreateMPIAIJSumSeqAIJSymbolic(comm,seqmat,m,n,mpimat);CHKERRQ(ierr); 4465 } 4466 ierr = MatCreateMPIAIJSumSeqAIJNumeric(seqmat,*mpimat);CHKERRQ(ierr); 4467 ierr = PetscLogEventEnd(MAT_Seqstompi,seqmat,0,0,0);CHKERRQ(ierr); 4468 PetscFunctionReturn(0); 4469 } 4470 4471 #undef __FUNCT__ 4472 #define __FUNCT__ "MatMPIAIJGetLocalMat" 4473 /*@ 4474 MatMPIAIJGetLocalMat - Creates a SeqAIJ from a MPIAIJ matrix by taking all its local rows and putting them into a sequential vector with 4475 mlocal rows and n columns. Where mlocal is the row count obtained with MatGetLocalSize() and n is the global column count obtained 4476 with MatGetSize() 4477 4478 Not Collective 4479 4480 Input Parameters: 4481 + A - the matrix 4482 . scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX 4483 4484 Output Parameter: 4485 . A_loc - the local sequential matrix generated 4486 4487 Level: developer 4488 4489 .seealso: MatGetOwnerShipRange(), MatMPIAIJGetLocalMatCondensed() 4490 4491 @*/ 4492 PetscErrorCode MatMPIAIJGetLocalMat(Mat A,MatReuse scall,Mat *A_loc) 4493 { 4494 PetscErrorCode ierr; 4495 Mat_MPIAIJ *mpimat=(Mat_MPIAIJ*)A->data; 4496 Mat_SeqAIJ *mat,*a,*b; 4497 PetscInt *ai,*aj,*bi,*bj,*cmap=mpimat->garray; 4498 MatScalar *aa,*ba,*cam; 4499 PetscScalar *ca; 4500 PetscInt am=A->rmap->n,i,j,k,cstart=A->cmap->rstart; 4501 PetscInt *ci,*cj,col,ncols_d,ncols_o,jo; 4502 PetscBool match; 4503 MPI_Comm comm; 4504 PetscMPIInt size; 4505 4506 PetscFunctionBegin; 4507 ierr = PetscObjectTypeCompare((PetscObject)A,MATMPIAIJ,&match);CHKERRQ(ierr); 4508 if (!match) SETERRQ(PetscObjectComm((PetscObject)A), PETSC_ERR_SUP,"Requires MPIAIJ matrix as input"); 4509 ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 4510 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 4511 if (size == 1 && scall == MAT_REUSE_MATRIX) PetscFunctionReturn(0); 4512 4513 ierr = PetscLogEventBegin(MAT_Getlocalmat,A,0,0,0);CHKERRQ(ierr); 4514 a = (Mat_SeqAIJ*)(mpimat->A)->data; 4515 b = (Mat_SeqAIJ*)(mpimat->B)->data; 4516 ai = a->i; aj = a->j; bi = b->i; bj = b->j; 4517 aa = a->a; ba = b->a; 4518 if (scall == MAT_INITIAL_MATRIX) { 4519 if (size == 1) { 4520 ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,am,A->cmap->N,ai,aj,aa,A_loc);CHKERRQ(ierr); 4521 PetscFunctionReturn(0); 4522 } 4523 4524 ierr = PetscMalloc1((1+am),&ci);CHKERRQ(ierr); 4525 ci[0] = 0; 4526 for (i=0; i<am; i++) { 4527 ci[i+1] = ci[i] + (ai[i+1] - ai[i]) + (bi[i+1] - bi[i]); 4528 } 4529 ierr = PetscMalloc1((1+ci[am]),&cj);CHKERRQ(ierr); 4530 ierr = PetscMalloc1((1+ci[am]),&ca);CHKERRQ(ierr); 4531 k = 0; 4532 for (i=0; i<am; i++) { 4533 ncols_o = bi[i+1] - bi[i]; 4534 ncols_d = ai[i+1] - ai[i]; 4535 /* off-diagonal portion of A */ 4536 for (jo=0; jo<ncols_o; jo++) { 4537 col = cmap[*bj]; 4538 if (col >= cstart) break; 4539 cj[k] = col; bj++; 4540 ca[k++] = *ba++; 4541 } 4542 /* diagonal portion of A */ 4543 for (j=0; j<ncols_d; j++) { 4544 cj[k] = cstart + *aj++; 4545 ca[k++] = *aa++; 4546 } 4547 /* off-diagonal portion of A */ 4548 for (j=jo; j<ncols_o; j++) { 4549 cj[k] = cmap[*bj++]; 4550 ca[k++] = *ba++; 4551 } 4552 } 4553 /* put together the new matrix */ 4554 ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,am,A->cmap->N,ci,cj,ca,A_loc);CHKERRQ(ierr); 4555 /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */ 4556 /* Since these are PETSc arrays, change flags to free them as necessary. */ 4557 mat = (Mat_SeqAIJ*)(*A_loc)->data; 4558 mat->free_a = PETSC_TRUE; 4559 mat->free_ij = PETSC_TRUE; 4560 mat->nonew = 0; 4561 } else if (scall == MAT_REUSE_MATRIX) { 4562 mat=(Mat_SeqAIJ*)(*A_loc)->data; 4563 ci = mat->i; cj = mat->j; cam = mat->a; 4564 for (i=0; i<am; i++) { 4565 /* off-diagonal portion of A */ 4566 ncols_o = bi[i+1] - bi[i]; 4567 for (jo=0; jo<ncols_o; jo++) { 4568 col = cmap[*bj]; 4569 if (col >= cstart) break; 4570 *cam++ = *ba++; bj++; 4571 } 4572 /* diagonal portion of A */ 4573 ncols_d = ai[i+1] - ai[i]; 4574 for (j=0; j<ncols_d; j++) *cam++ = *aa++; 4575 /* off-diagonal portion of A */ 4576 for (j=jo; j<ncols_o; j++) { 4577 *cam++ = *ba++; bj++; 4578 } 4579 } 4580 } else SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Invalid MatReuse %d",(int)scall); 4581 ierr = PetscLogEventEnd(MAT_Getlocalmat,A,0,0,0);CHKERRQ(ierr); 4582 PetscFunctionReturn(0); 4583 } 4584 4585 #undef __FUNCT__ 4586 #define __FUNCT__ "MatMPIAIJGetLocalMatCondensed" 4587 /*@C 4588 MatMPIAIJGetLocalMatCondensed - Creates a SeqAIJ matrix from an MPIAIJ matrix by taking all its local rows and NON-ZERO columns 4589 4590 Not Collective 4591 4592 Input Parameters: 4593 + A - the matrix 4594 . scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX 4595 - row, col - index sets of rows and columns to extract (or NULL) 4596 4597 Output Parameter: 4598 . A_loc - the local sequential matrix generated 4599 4600 Level: developer 4601 4602 .seealso: MatGetOwnershipRange(), MatMPIAIJGetLocalMat() 4603 4604 @*/ 4605 PetscErrorCode MatMPIAIJGetLocalMatCondensed(Mat A,MatReuse scall,IS *row,IS *col,Mat *A_loc) 4606 { 4607 Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data; 4608 PetscErrorCode ierr; 4609 PetscInt i,start,end,ncols,nzA,nzB,*cmap,imark,*idx; 4610 IS isrowa,iscola; 4611 Mat *aloc; 4612 PetscBool match; 4613 4614 PetscFunctionBegin; 4615 ierr = PetscObjectTypeCompare((PetscObject)A,MATMPIAIJ,&match);CHKERRQ(ierr); 4616 if (!match) SETERRQ(PetscObjectComm((PetscObject)A), PETSC_ERR_SUP,"Requires MPIAIJ matrix as input"); 4617 ierr = PetscLogEventBegin(MAT_Getlocalmatcondensed,A,0,0,0);CHKERRQ(ierr); 4618 if (!row) { 4619 start = A->rmap->rstart; end = A->rmap->rend; 4620 ierr = ISCreateStride(PETSC_COMM_SELF,end-start,start,1,&isrowa);CHKERRQ(ierr); 4621 } else { 4622 isrowa = *row; 4623 } 4624 if (!col) { 4625 start = A->cmap->rstart; 4626 cmap = a->garray; 4627 nzA = a->A->cmap->n; 4628 nzB = a->B->cmap->n; 4629 ierr = PetscMalloc1((nzA+nzB), &idx);CHKERRQ(ierr); 4630 ncols = 0; 4631 for (i=0; i<nzB; i++) { 4632 if (cmap[i] < start) idx[ncols++] = cmap[i]; 4633 else break; 4634 } 4635 imark = i; 4636 for (i=0; i<nzA; i++) idx[ncols++] = start + i; 4637 for (i=imark; i<nzB; i++) idx[ncols++] = cmap[i]; 4638 ierr = ISCreateGeneral(PETSC_COMM_SELF,ncols,idx,PETSC_OWN_POINTER,&iscola);CHKERRQ(ierr); 4639 } else { 4640 iscola = *col; 4641 } 4642 if (scall != MAT_INITIAL_MATRIX) { 4643 ierr = PetscMalloc(sizeof(Mat),&aloc);CHKERRQ(ierr); 4644 aloc[0] = *A_loc; 4645 } 4646 ierr = MatGetSubMatrices(A,1,&isrowa,&iscola,scall,&aloc);CHKERRQ(ierr); 4647 *A_loc = aloc[0]; 4648 ierr = PetscFree(aloc);CHKERRQ(ierr); 4649 if (!row) { 4650 ierr = ISDestroy(&isrowa);CHKERRQ(ierr); 4651 } 4652 if (!col) { 4653 ierr = ISDestroy(&iscola);CHKERRQ(ierr); 4654 } 4655 ierr = PetscLogEventEnd(MAT_Getlocalmatcondensed,A,0,0,0);CHKERRQ(ierr); 4656 PetscFunctionReturn(0); 4657 } 4658 4659 #undef __FUNCT__ 4660 #define __FUNCT__ "MatGetBrowsOfAcols" 4661 /*@C 4662 MatGetBrowsOfAcols - Creates a SeqAIJ matrix by taking rows of B that equal to nonzero columns of local A 4663 4664 Collective on Mat 4665 4666 Input Parameters: 4667 + A,B - the matrices in mpiaij format 4668 . scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX 4669 - rowb, colb - index sets of rows and columns of B to extract (or NULL) 4670 4671 Output Parameter: 4672 + rowb, colb - index sets of rows and columns of B to extract 4673 - B_seq - the sequential matrix generated 4674 4675 Level: developer 4676 4677 @*/ 4678 PetscErrorCode MatGetBrowsOfAcols(Mat A,Mat B,MatReuse scall,IS *rowb,IS *colb,Mat *B_seq) 4679 { 4680 Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data; 4681 PetscErrorCode ierr; 4682 PetscInt *idx,i,start,ncols,nzA,nzB,*cmap,imark; 4683 IS isrowb,iscolb; 4684 Mat *bseq=NULL; 4685 4686 PetscFunctionBegin; 4687 if (A->cmap->rstart != B->rmap->rstart || A->cmap->rend != B->rmap->rend) { 4688 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); 4689 } 4690 ierr = PetscLogEventBegin(MAT_GetBrowsOfAcols,A,B,0,0);CHKERRQ(ierr); 4691 4692 if (scall == MAT_INITIAL_MATRIX) { 4693 start = A->cmap->rstart; 4694 cmap = a->garray; 4695 nzA = a->A->cmap->n; 4696 nzB = a->B->cmap->n; 4697 ierr = PetscMalloc1((nzA+nzB), &idx);CHKERRQ(ierr); 4698 ncols = 0; 4699 for (i=0; i<nzB; i++) { /* row < local row index */ 4700 if (cmap[i] < start) idx[ncols++] = cmap[i]; 4701 else break; 4702 } 4703 imark = i; 4704 for (i=0; i<nzA; i++) idx[ncols++] = start + i; /* local rows */ 4705 for (i=imark; i<nzB; i++) idx[ncols++] = cmap[i]; /* row > local row index */ 4706 ierr = ISCreateGeneral(PETSC_COMM_SELF,ncols,idx,PETSC_OWN_POINTER,&isrowb);CHKERRQ(ierr); 4707 ierr = ISCreateStride(PETSC_COMM_SELF,B->cmap->N,0,1,&iscolb);CHKERRQ(ierr); 4708 } else { 4709 if (!rowb || !colb) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"IS rowb and colb must be provided for MAT_REUSE_MATRIX"); 4710 isrowb = *rowb; iscolb = *colb; 4711 ierr = PetscMalloc(sizeof(Mat),&bseq);CHKERRQ(ierr); 4712 bseq[0] = *B_seq; 4713 } 4714 ierr = MatGetSubMatrices(B,1,&isrowb,&iscolb,scall,&bseq);CHKERRQ(ierr); 4715 *B_seq = bseq[0]; 4716 ierr = PetscFree(bseq);CHKERRQ(ierr); 4717 if (!rowb) { 4718 ierr = ISDestroy(&isrowb);CHKERRQ(ierr); 4719 } else { 4720 *rowb = isrowb; 4721 } 4722 if (!colb) { 4723 ierr = ISDestroy(&iscolb);CHKERRQ(ierr); 4724 } else { 4725 *colb = iscolb; 4726 } 4727 ierr = PetscLogEventEnd(MAT_GetBrowsOfAcols,A,B,0,0);CHKERRQ(ierr); 4728 PetscFunctionReturn(0); 4729 } 4730 4731 #undef __FUNCT__ 4732 #define __FUNCT__ "MatGetBrowsOfAoCols_MPIAIJ" 4733 /* 4734 MatGetBrowsOfAoCols_MPIAIJ - Creates a SeqAIJ matrix by taking rows of B that equal to nonzero columns 4735 of the OFF-DIAGONAL portion of local A 4736 4737 Collective on Mat 4738 4739 Input Parameters: 4740 + A,B - the matrices in mpiaij format 4741 - scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX 4742 4743 Output Parameter: 4744 + startsj_s - starting point in B's sending j-arrays, saved for MAT_REUSE (or NULL) 4745 . startsj_r - starting point in B's receiving j-arrays, saved for MAT_REUSE (or NULL) 4746 . bufa_ptr - array for sending matrix values, saved for MAT_REUSE (or NULL) 4747 - B_oth - the sequential matrix generated with size aBn=a->B->cmap->n by B->cmap->N 4748 4749 Level: developer 4750 4751 */ 4752 PetscErrorCode MatGetBrowsOfAoCols_MPIAIJ(Mat A,Mat B,MatReuse scall,PetscInt **startsj_s,PetscInt **startsj_r,MatScalar **bufa_ptr,Mat *B_oth) 4753 { 4754 VecScatter_MPI_General *gen_to,*gen_from; 4755 PetscErrorCode ierr; 4756 Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data; 4757 Mat_SeqAIJ *b_oth; 4758 VecScatter ctx =a->Mvctx; 4759 MPI_Comm comm; 4760 PetscMPIInt *rprocs,*sprocs,tag=((PetscObject)ctx)->tag,rank; 4761 PetscInt *rowlen,*bufj,*bufJ,ncols,aBn=a->B->cmap->n,row,*b_othi,*b_othj; 4762 PetscScalar *rvalues,*svalues; 4763 MatScalar *b_otha,*bufa,*bufA; 4764 PetscInt i,j,k,l,ll,nrecvs,nsends,nrows,*srow,*rstarts,*rstartsj = 0,*sstarts,*sstartsj,len; 4765 MPI_Request *rwaits = NULL,*swaits = NULL; 4766 MPI_Status *sstatus,rstatus; 4767 PetscMPIInt jj,size; 4768 PetscInt *cols,sbs,rbs; 4769 PetscScalar *vals; 4770 4771 PetscFunctionBegin; 4772 ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 4773 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 4774 if (size == 1) PetscFunctionReturn(0); 4775 4776 if (A->cmap->rstart != B->rmap->rstart || A->cmap->rend != B->rmap->rend) { 4777 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); 4778 } 4779 ierr = PetscLogEventBegin(MAT_GetBrowsOfAocols,A,B,0,0);CHKERRQ(ierr); 4780 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 4781 4782 gen_to = (VecScatter_MPI_General*)ctx->todata; 4783 gen_from = (VecScatter_MPI_General*)ctx->fromdata; 4784 rvalues = gen_from->values; /* holds the length of receiving row */ 4785 svalues = gen_to->values; /* holds the length of sending row */ 4786 nrecvs = gen_from->n; 4787 nsends = gen_to->n; 4788 4789 ierr = PetscMalloc2(nrecvs,&rwaits,nsends,&swaits);CHKERRQ(ierr); 4790 srow = gen_to->indices; /* local row index to be sent */ 4791 sstarts = gen_to->starts; 4792 sprocs = gen_to->procs; 4793 sstatus = gen_to->sstatus; 4794 sbs = gen_to->bs; 4795 rstarts = gen_from->starts; 4796 rprocs = gen_from->procs; 4797 rbs = gen_from->bs; 4798 4799 if (!startsj_s || !bufa_ptr) scall = MAT_INITIAL_MATRIX; 4800 if (scall == MAT_INITIAL_MATRIX) { 4801 /* i-array */ 4802 /*---------*/ 4803 /* post receives */ 4804 for (i=0; i<nrecvs; i++) { 4805 rowlen = (PetscInt*)rvalues + rstarts[i]*rbs; 4806 nrows = (rstarts[i+1]-rstarts[i])*rbs; /* num of indices to be received */ 4807 ierr = MPI_Irecv(rowlen,nrows,MPIU_INT,rprocs[i],tag,comm,rwaits+i);CHKERRQ(ierr); 4808 } 4809 4810 /* pack the outgoing message */ 4811 ierr = PetscMalloc2(nsends+1,&sstartsj,nrecvs+1,&rstartsj);CHKERRQ(ierr); 4812 4813 sstartsj[0] = 0; 4814 rstartsj[0] = 0; 4815 len = 0; /* total length of j or a array to be sent */ 4816 k = 0; 4817 for (i=0; i<nsends; i++) { 4818 rowlen = (PetscInt*)svalues + sstarts[i]*sbs; 4819 nrows = sstarts[i+1]-sstarts[i]; /* num of block rows */ 4820 for (j=0; j<nrows; j++) { 4821 row = srow[k] + B->rmap->range[rank]; /* global row idx */ 4822 for (l=0; l<sbs; l++) { 4823 ierr = MatGetRow_MPIAIJ(B,row+l,&ncols,NULL,NULL);CHKERRQ(ierr); /* rowlength */ 4824 4825 rowlen[j*sbs+l] = ncols; 4826 4827 len += ncols; 4828 ierr = MatRestoreRow_MPIAIJ(B,row+l,&ncols,NULL,NULL);CHKERRQ(ierr); 4829 } 4830 k++; 4831 } 4832 ierr = MPI_Isend(rowlen,nrows*sbs,MPIU_INT,sprocs[i],tag,comm,swaits+i);CHKERRQ(ierr); 4833 4834 sstartsj[i+1] = len; /* starting point of (i+1)-th outgoing msg in bufj and bufa */ 4835 } 4836 /* recvs and sends of i-array are completed */ 4837 i = nrecvs; 4838 while (i--) { 4839 ierr = MPI_Waitany(nrecvs,rwaits,&jj,&rstatus);CHKERRQ(ierr); 4840 } 4841 if (nsends) {ierr = MPI_Waitall(nsends,swaits,sstatus);CHKERRQ(ierr);} 4842 4843 /* allocate buffers for sending j and a arrays */ 4844 ierr = PetscMalloc1((len+1),&bufj);CHKERRQ(ierr); 4845 ierr = PetscMalloc1((len+1),&bufa);CHKERRQ(ierr); 4846 4847 /* create i-array of B_oth */ 4848 ierr = PetscMalloc1((aBn+2),&b_othi);CHKERRQ(ierr); 4849 4850 b_othi[0] = 0; 4851 len = 0; /* total length of j or a array to be received */ 4852 k = 0; 4853 for (i=0; i<nrecvs; i++) { 4854 rowlen = (PetscInt*)rvalues + rstarts[i]*rbs; 4855 nrows = rbs*(rstarts[i+1]-rstarts[i]); /* num of rows to be recieved */ 4856 for (j=0; j<nrows; j++) { 4857 b_othi[k+1] = b_othi[k] + rowlen[j]; 4858 len += rowlen[j]; k++; 4859 } 4860 rstartsj[i+1] = len; /* starting point of (i+1)-th incoming msg in bufj and bufa */ 4861 } 4862 4863 /* allocate space for j and a arrrays of B_oth */ 4864 ierr = PetscMalloc1((b_othi[aBn]+1),&b_othj);CHKERRQ(ierr); 4865 ierr = PetscMalloc1((b_othi[aBn]+1),&b_otha);CHKERRQ(ierr); 4866 4867 /* j-array */ 4868 /*---------*/ 4869 /* post receives of j-array */ 4870 for (i=0; i<nrecvs; i++) { 4871 nrows = rstartsj[i+1]-rstartsj[i]; /* length of the msg received */ 4872 ierr = MPI_Irecv(b_othj+rstartsj[i],nrows,MPIU_INT,rprocs[i],tag,comm,rwaits+i);CHKERRQ(ierr); 4873 } 4874 4875 /* pack the outgoing message j-array */ 4876 k = 0; 4877 for (i=0; i<nsends; i++) { 4878 nrows = sstarts[i+1]-sstarts[i]; /* num of block rows */ 4879 bufJ = bufj+sstartsj[i]; 4880 for (j=0; j<nrows; j++) { 4881 row = srow[k++] + B->rmap->range[rank]; /* global row idx */ 4882 for (ll=0; ll<sbs; ll++) { 4883 ierr = MatGetRow_MPIAIJ(B,row+ll,&ncols,&cols,NULL);CHKERRQ(ierr); 4884 for (l=0; l<ncols; l++) { 4885 *bufJ++ = cols[l]; 4886 } 4887 ierr = MatRestoreRow_MPIAIJ(B,row+ll,&ncols,&cols,NULL);CHKERRQ(ierr); 4888 } 4889 } 4890 ierr = MPI_Isend(bufj+sstartsj[i],sstartsj[i+1]-sstartsj[i],MPIU_INT,sprocs[i],tag,comm,swaits+i);CHKERRQ(ierr); 4891 } 4892 4893 /* recvs and sends of j-array are completed */ 4894 i = nrecvs; 4895 while (i--) { 4896 ierr = MPI_Waitany(nrecvs,rwaits,&jj,&rstatus);CHKERRQ(ierr); 4897 } 4898 if (nsends) {ierr = MPI_Waitall(nsends,swaits,sstatus);CHKERRQ(ierr);} 4899 } else if (scall == MAT_REUSE_MATRIX) { 4900 sstartsj = *startsj_s; 4901 rstartsj = *startsj_r; 4902 bufa = *bufa_ptr; 4903 b_oth = (Mat_SeqAIJ*)(*B_oth)->data; 4904 b_otha = b_oth->a; 4905 } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE, "Matrix P does not posses an object container"); 4906 4907 /* a-array */ 4908 /*---------*/ 4909 /* post receives of a-array */ 4910 for (i=0; i<nrecvs; i++) { 4911 nrows = rstartsj[i+1]-rstartsj[i]; /* length of the msg received */ 4912 ierr = MPI_Irecv(b_otha+rstartsj[i],nrows,MPIU_SCALAR,rprocs[i],tag,comm,rwaits+i);CHKERRQ(ierr); 4913 } 4914 4915 /* pack the outgoing message a-array */ 4916 k = 0; 4917 for (i=0; i<nsends; i++) { 4918 nrows = sstarts[i+1]-sstarts[i]; /* num of block rows */ 4919 bufA = bufa+sstartsj[i]; 4920 for (j=0; j<nrows; j++) { 4921 row = srow[k++] + B->rmap->range[rank]; /* global row idx */ 4922 for (ll=0; ll<sbs; ll++) { 4923 ierr = MatGetRow_MPIAIJ(B,row+ll,&ncols,NULL,&vals);CHKERRQ(ierr); 4924 for (l=0; l<ncols; l++) { 4925 *bufA++ = vals[l]; 4926 } 4927 ierr = MatRestoreRow_MPIAIJ(B,row+ll,&ncols,NULL,&vals);CHKERRQ(ierr); 4928 } 4929 } 4930 ierr = MPI_Isend(bufa+sstartsj[i],sstartsj[i+1]-sstartsj[i],MPIU_SCALAR,sprocs[i],tag,comm,swaits+i);CHKERRQ(ierr); 4931 } 4932 /* recvs and sends of a-array are completed */ 4933 i = nrecvs; 4934 while (i--) { 4935 ierr = MPI_Waitany(nrecvs,rwaits,&jj,&rstatus);CHKERRQ(ierr); 4936 } 4937 if (nsends) {ierr = MPI_Waitall(nsends,swaits,sstatus);CHKERRQ(ierr);} 4938 ierr = PetscFree2(rwaits,swaits);CHKERRQ(ierr); 4939 4940 if (scall == MAT_INITIAL_MATRIX) { 4941 /* put together the new matrix */ 4942 ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,aBn,B->cmap->N,b_othi,b_othj,b_otha,B_oth);CHKERRQ(ierr); 4943 4944 /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */ 4945 /* Since these are PETSc arrays, change flags to free them as necessary. */ 4946 b_oth = (Mat_SeqAIJ*)(*B_oth)->data; 4947 b_oth->free_a = PETSC_TRUE; 4948 b_oth->free_ij = PETSC_TRUE; 4949 b_oth->nonew = 0; 4950 4951 ierr = PetscFree(bufj);CHKERRQ(ierr); 4952 if (!startsj_s || !bufa_ptr) { 4953 ierr = PetscFree2(sstartsj,rstartsj);CHKERRQ(ierr); 4954 ierr = PetscFree(bufa_ptr);CHKERRQ(ierr); 4955 } else { 4956 *startsj_s = sstartsj; 4957 *startsj_r = rstartsj; 4958 *bufa_ptr = bufa; 4959 } 4960 } 4961 ierr = PetscLogEventEnd(MAT_GetBrowsOfAocols,A,B,0,0);CHKERRQ(ierr); 4962 PetscFunctionReturn(0); 4963 } 4964 4965 #undef __FUNCT__ 4966 #define __FUNCT__ "MatGetCommunicationStructs" 4967 /*@C 4968 MatGetCommunicationStructs - Provides access to the communication structures used in matrix-vector multiplication. 4969 4970 Not Collective 4971 4972 Input Parameters: 4973 . A - The matrix in mpiaij format 4974 4975 Output Parameter: 4976 + lvec - The local vector holding off-process values from the argument to a matrix-vector product 4977 . colmap - A map from global column index to local index into lvec 4978 - multScatter - A scatter from the argument of a matrix-vector product to lvec 4979 4980 Level: developer 4981 4982 @*/ 4983 #if defined(PETSC_USE_CTABLE) 4984 PetscErrorCode MatGetCommunicationStructs(Mat A, Vec *lvec, PetscTable *colmap, VecScatter *multScatter) 4985 #else 4986 PetscErrorCode MatGetCommunicationStructs(Mat A, Vec *lvec, PetscInt *colmap[], VecScatter *multScatter) 4987 #endif 4988 { 4989 Mat_MPIAIJ *a; 4990 4991 PetscFunctionBegin; 4992 PetscValidHeaderSpecific(A, MAT_CLASSID, 1); 4993 PetscValidPointer(lvec, 2); 4994 PetscValidPointer(colmap, 3); 4995 PetscValidPointer(multScatter, 4); 4996 a = (Mat_MPIAIJ*) A->data; 4997 if (lvec) *lvec = a->lvec; 4998 if (colmap) *colmap = a->colmap; 4999 if (multScatter) *multScatter = a->Mvctx; 5000 PetscFunctionReturn(0); 5001 } 5002 5003 PETSC_EXTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJCRL(Mat,MatType,MatReuse,Mat*); 5004 PETSC_EXTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJPERM(Mat,MatType,MatReuse,Mat*); 5005 PETSC_EXTERN PetscErrorCode MatConvert_MPIAIJ_MPISBAIJ(Mat,MatType,MatReuse,Mat*); 5006 #if defined(PETSC_HAVE_ELEMENTAL) 5007 PETSC_EXTERN PetscErrorCode MatConvert_MPIAIJ_Elemental(Mat,MatType,MatReuse,Mat*); 5008 #endif 5009 5010 #undef __FUNCT__ 5011 #define __FUNCT__ "MatMatMultNumeric_MPIDense_MPIAIJ" 5012 /* 5013 Computes (B'*A')' since computing B*A directly is untenable 5014 5015 n p p 5016 ( ) ( ) ( ) 5017 m ( A ) * n ( B ) = m ( C ) 5018 ( ) ( ) ( ) 5019 5020 */ 5021 PetscErrorCode MatMatMultNumeric_MPIDense_MPIAIJ(Mat A,Mat B,Mat C) 5022 { 5023 PetscErrorCode ierr; 5024 Mat At,Bt,Ct; 5025 5026 PetscFunctionBegin; 5027 ierr = MatTranspose(A,MAT_INITIAL_MATRIX,&At);CHKERRQ(ierr); 5028 ierr = MatTranspose(B,MAT_INITIAL_MATRIX,&Bt);CHKERRQ(ierr); 5029 ierr = MatMatMult(Bt,At,MAT_INITIAL_MATRIX,1.0,&Ct);CHKERRQ(ierr); 5030 ierr = MatDestroy(&At);CHKERRQ(ierr); 5031 ierr = MatDestroy(&Bt);CHKERRQ(ierr); 5032 ierr = MatTranspose(Ct,MAT_REUSE_MATRIX,&C);CHKERRQ(ierr); 5033 ierr = MatDestroy(&Ct);CHKERRQ(ierr); 5034 PetscFunctionReturn(0); 5035 } 5036 5037 #undef __FUNCT__ 5038 #define __FUNCT__ "MatMatMultSymbolic_MPIDense_MPIAIJ" 5039 PetscErrorCode MatMatMultSymbolic_MPIDense_MPIAIJ(Mat A,Mat B,PetscReal fill,Mat *C) 5040 { 5041 PetscErrorCode ierr; 5042 PetscInt m=A->rmap->n,n=B->cmap->n; 5043 Mat Cmat; 5044 5045 PetscFunctionBegin; 5046 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); 5047 ierr = MatCreate(PetscObjectComm((PetscObject)A),&Cmat);CHKERRQ(ierr); 5048 ierr = MatSetSizes(Cmat,m,n,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 5049 ierr = MatSetBlockSizesFromMats(Cmat,A,B);CHKERRQ(ierr); 5050 ierr = MatSetType(Cmat,MATMPIDENSE);CHKERRQ(ierr); 5051 ierr = MatMPIDenseSetPreallocation(Cmat,NULL);CHKERRQ(ierr); 5052 ierr = MatAssemblyBegin(Cmat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5053 ierr = MatAssemblyEnd(Cmat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5054 5055 Cmat->ops->matmultnumeric = MatMatMultNumeric_MPIDense_MPIAIJ; 5056 5057 *C = Cmat; 5058 PetscFunctionReturn(0); 5059 } 5060 5061 /* ----------------------------------------------------------------*/ 5062 #undef __FUNCT__ 5063 #define __FUNCT__ "MatMatMult_MPIDense_MPIAIJ" 5064 PetscErrorCode MatMatMult_MPIDense_MPIAIJ(Mat A,Mat B,MatReuse scall,PetscReal fill,Mat *C) 5065 { 5066 PetscErrorCode ierr; 5067 5068 PetscFunctionBegin; 5069 if (scall == MAT_INITIAL_MATRIX) { 5070 ierr = PetscLogEventBegin(MAT_MatMultSymbolic,A,B,0,0);CHKERRQ(ierr); 5071 ierr = MatMatMultSymbolic_MPIDense_MPIAIJ(A,B,fill,C);CHKERRQ(ierr); 5072 ierr = PetscLogEventEnd(MAT_MatMultSymbolic,A,B,0,0);CHKERRQ(ierr); 5073 } 5074 ierr = PetscLogEventBegin(MAT_MatMultNumeric,A,B,0,0);CHKERRQ(ierr); 5075 ierr = MatMatMultNumeric_MPIDense_MPIAIJ(A,B,*C);CHKERRQ(ierr); 5076 ierr = PetscLogEventEnd(MAT_MatMultNumeric,A,B,0,0);CHKERRQ(ierr); 5077 PetscFunctionReturn(0); 5078 } 5079 5080 #if defined(PETSC_HAVE_MUMPS) 5081 PETSC_EXTERN PetscErrorCode MatGetFactor_aij_mumps(Mat,MatFactorType,Mat*); 5082 #endif 5083 #if defined(PETSC_HAVE_PASTIX) 5084 PETSC_EXTERN PetscErrorCode MatGetFactor_mpiaij_pastix(Mat,MatFactorType,Mat*); 5085 #endif 5086 #if defined(PETSC_HAVE_SUPERLU_DIST) 5087 PETSC_EXTERN PetscErrorCode MatGetFactor_mpiaij_superlu_dist(Mat,MatFactorType,Mat*); 5088 #endif 5089 #if defined(PETSC_HAVE_CLIQUE) 5090 PETSC_EXTERN PetscErrorCode MatGetFactor_aij_clique(Mat,MatFactorType,Mat*); 5091 #endif 5092 5093 /*MC 5094 MATMPIAIJ - MATMPIAIJ = "mpiaij" - A matrix type to be used for parallel sparse matrices. 5095 5096 Options Database Keys: 5097 . -mat_type mpiaij - sets the matrix type to "mpiaij" during a call to MatSetFromOptions() 5098 5099 Level: beginner 5100 5101 .seealso: MatCreateAIJ() 5102 M*/ 5103 5104 #undef __FUNCT__ 5105 #define __FUNCT__ "MatCreate_MPIAIJ" 5106 PETSC_EXTERN PetscErrorCode MatCreate_MPIAIJ(Mat B) 5107 { 5108 Mat_MPIAIJ *b; 5109 PetscErrorCode ierr; 5110 PetscMPIInt size; 5111 5112 PetscFunctionBegin; 5113 ierr = MPI_Comm_size(PetscObjectComm((PetscObject)B),&size);CHKERRQ(ierr); 5114 5115 ierr = PetscNewLog(B,&b);CHKERRQ(ierr); 5116 B->data = (void*)b; 5117 ierr = PetscMemcpy(B->ops,&MatOps_Values,sizeof(struct _MatOps));CHKERRQ(ierr); 5118 B->assembled = PETSC_FALSE; 5119 B->insertmode = NOT_SET_VALUES; 5120 b->size = size; 5121 5122 ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)B),&b->rank);CHKERRQ(ierr); 5123 5124 /* build cache for off array entries formed */ 5125 ierr = MatStashCreate_Private(PetscObjectComm((PetscObject)B),1,&B->stash);CHKERRQ(ierr); 5126 5127 b->donotstash = PETSC_FALSE; 5128 b->colmap = 0; 5129 b->garray = 0; 5130 b->roworiented = PETSC_TRUE; 5131 5132 /* stuff used for matrix vector multiply */ 5133 b->lvec = NULL; 5134 b->Mvctx = NULL; 5135 5136 /* stuff for MatGetRow() */ 5137 b->rowindices = 0; 5138 b->rowvalues = 0; 5139 b->getrowactive = PETSC_FALSE; 5140 5141 /* flexible pointer used in CUSP/CUSPARSE classes */ 5142 b->spptr = NULL; 5143 5144 #if defined(PETSC_HAVE_MUMPS) 5145 ierr = PetscObjectComposeFunction((PetscObject)B,"MatGetFactor_mumps_C",MatGetFactor_aij_mumps);CHKERRQ(ierr); 5146 #endif 5147 #if defined(PETSC_HAVE_PASTIX) 5148 ierr = PetscObjectComposeFunction((PetscObject)B,"MatGetFactor_pastix_C",MatGetFactor_mpiaij_pastix);CHKERRQ(ierr); 5149 #endif 5150 #if defined(PETSC_HAVE_SUPERLU_DIST) 5151 ierr = PetscObjectComposeFunction((PetscObject)B,"MatGetFactor_superlu_dist_C",MatGetFactor_mpiaij_superlu_dist);CHKERRQ(ierr); 5152 #endif 5153 #if defined(PETSC_HAVE_CLIQUE) 5154 ierr = PetscObjectComposeFunction((PetscObject)B,"MatGetFactor_clique_C",MatGetFactor_aij_clique);CHKERRQ(ierr); 5155 #endif 5156 ierr = PetscObjectComposeFunction((PetscObject)B,"MatStoreValues_C",MatStoreValues_MPIAIJ);CHKERRQ(ierr); 5157 ierr = PetscObjectComposeFunction((PetscObject)B,"MatRetrieveValues_C",MatRetrieveValues_MPIAIJ);CHKERRQ(ierr); 5158 ierr = PetscObjectComposeFunction((PetscObject)B,"MatGetDiagonalBlock_C",MatGetDiagonalBlock_MPIAIJ);CHKERRQ(ierr); 5159 ierr = PetscObjectComposeFunction((PetscObject)B,"MatIsTranspose_C",MatIsTranspose_MPIAIJ);CHKERRQ(ierr); 5160 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMPIAIJSetPreallocation_C",MatMPIAIJSetPreallocation_MPIAIJ);CHKERRQ(ierr); 5161 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMPIAIJSetPreallocationCSR_C",MatMPIAIJSetPreallocationCSR_MPIAIJ);CHKERRQ(ierr); 5162 ierr = PetscObjectComposeFunction((PetscObject)B,"MatDiagonalScaleLocal_C",MatDiagonalScaleLocal_MPIAIJ);CHKERRQ(ierr); 5163 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_mpiaijperm_C",MatConvert_MPIAIJ_MPIAIJPERM);CHKERRQ(ierr); 5164 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_mpiaijcrl_C",MatConvert_MPIAIJ_MPIAIJCRL);CHKERRQ(ierr); 5165 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_mpisbaij_C",MatConvert_MPIAIJ_MPISBAIJ);CHKERRQ(ierr); 5166 #if defined(PETSC_HAVE_ELEMENTAL) 5167 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_elemental_C",MatConvert_MPIAIJ_Elemental);CHKERRQ(ierr); 5168 #endif 5169 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMatMult_mpidense_mpiaij_C",MatMatMult_MPIDense_MPIAIJ);CHKERRQ(ierr); 5170 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMatMultSymbolic_mpidense_mpiaij_C",MatMatMultSymbolic_MPIDense_MPIAIJ);CHKERRQ(ierr); 5171 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMatMultNumeric_mpidense_mpiaij_C",MatMatMultNumeric_MPIDense_MPIAIJ);CHKERRQ(ierr); 5172 ierr = PetscObjectChangeTypeName((PetscObject)B,MATMPIAIJ);CHKERRQ(ierr); 5173 PetscFunctionReturn(0); 5174 } 5175 5176 #undef __FUNCT__ 5177 #define __FUNCT__ "MatCreateMPIAIJWithSplitArrays" 5178 /*@ 5179 MatCreateMPIAIJWithSplitArrays - creates a MPI AIJ matrix using arrays that contain the "diagonal" 5180 and "off-diagonal" part of the matrix in CSR format. 5181 5182 Collective on MPI_Comm 5183 5184 Input Parameters: 5185 + comm - MPI communicator 5186 . m - number of local rows (Cannot be PETSC_DECIDE) 5187 . n - This value should be the same as the local size used in creating the 5188 x vector for the matrix-vector product y = Ax. (or PETSC_DECIDE to have 5189 calculated if N is given) For square matrices n is almost always m. 5190 . M - number of global rows (or PETSC_DETERMINE to have calculated if m is given) 5191 . N - number of global columns (or PETSC_DETERMINE to have calculated if n is given) 5192 . i - row indices for "diagonal" portion of matrix 5193 . j - column indices 5194 . a - matrix values 5195 . oi - row indices for "off-diagonal" portion of matrix 5196 . oj - column indices 5197 - oa - matrix values 5198 5199 Output Parameter: 5200 . mat - the matrix 5201 5202 Level: advanced 5203 5204 Notes: 5205 The i, j, and a arrays ARE NOT copied by this routine into the internal format used by PETSc. The user 5206 must free the arrays once the matrix has been destroyed and not before. 5207 5208 The i and j indices are 0 based 5209 5210 See MatCreateAIJ() for the definition of "diagonal" and "off-diagonal" portion of the matrix 5211 5212 This sets local rows and cannot be used to set off-processor values. 5213 5214 Use of this routine is discouraged because it is inflexible and cumbersome to use. It is extremely rare that a 5215 legacy application natively assembles into exactly this split format. The code to do so is nontrivial and does 5216 not easily support in-place reassembly. It is recommended to use MatSetValues() (or a variant thereof) because 5217 the resulting assembly is easier to implement, will work with any matrix format, and the user does not have to 5218 keep track of the underlying array. Use MatSetOption(A,MAT_IGNORE_OFF_PROC_ENTRIES,PETSC_TRUE) to disable all 5219 communication if it is known that only local entries will be set. 5220 5221 .keywords: matrix, aij, compressed row, sparse, parallel 5222 5223 .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatMPIAIJSetPreallocationCSR(), 5224 MPIAIJ, MatCreateAIJ(), MatCreateMPIAIJWithArrays() 5225 @*/ 5226 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) 5227 { 5228 PetscErrorCode ierr; 5229 Mat_MPIAIJ *maij; 5230 5231 PetscFunctionBegin; 5232 if (m < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"local number of rows (m) cannot be PETSC_DECIDE, or negative"); 5233 if (i[0]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"i (row indices) must start with 0"); 5234 if (oi[0]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"oi (row indices) must start with 0"); 5235 ierr = MatCreate(comm,mat);CHKERRQ(ierr); 5236 ierr = MatSetSizes(*mat,m,n,M,N);CHKERRQ(ierr); 5237 ierr = MatSetType(*mat,MATMPIAIJ);CHKERRQ(ierr); 5238 maij = (Mat_MPIAIJ*) (*mat)->data; 5239 5240 (*mat)->preallocated = PETSC_TRUE; 5241 5242 ierr = PetscLayoutSetUp((*mat)->rmap);CHKERRQ(ierr); 5243 ierr = PetscLayoutSetUp((*mat)->cmap);CHKERRQ(ierr); 5244 5245 ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,m,n,i,j,a,&maij->A);CHKERRQ(ierr); 5246 ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,m,(*mat)->cmap->N,oi,oj,oa,&maij->B);CHKERRQ(ierr); 5247 5248 ierr = MatAssemblyBegin(maij->A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5249 ierr = MatAssemblyEnd(maij->A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5250 ierr = MatAssemblyBegin(maij->B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5251 ierr = MatAssemblyEnd(maij->B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5252 5253 ierr = MatAssemblyBegin(*mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5254 ierr = MatAssemblyEnd(*mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5255 ierr = MatSetOption(*mat,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE);CHKERRQ(ierr); 5256 PetscFunctionReturn(0); 5257 } 5258 5259 /* 5260 Special version for direct calls from Fortran 5261 */ 5262 #include <petsc-private/fortranimpl.h> 5263 5264 #if defined(PETSC_HAVE_FORTRAN_CAPS) 5265 #define matsetvaluesmpiaij_ MATSETVALUESMPIAIJ 5266 #elif !defined(PETSC_HAVE_FORTRAN_UNDERSCORE) 5267 #define matsetvaluesmpiaij_ matsetvaluesmpiaij 5268 #endif 5269 5270 /* Change these macros so can be used in void function */ 5271 #undef CHKERRQ 5272 #define CHKERRQ(ierr) CHKERRABORT(PETSC_COMM_WORLD,ierr) 5273 #undef SETERRQ2 5274 #define SETERRQ2(comm,ierr,b,c,d) CHKERRABORT(comm,ierr) 5275 #undef SETERRQ3 5276 #define SETERRQ3(comm,ierr,b,c,d,e) CHKERRABORT(comm,ierr) 5277 #undef SETERRQ 5278 #define SETERRQ(c,ierr,b) CHKERRABORT(c,ierr) 5279 5280 #undef __FUNCT__ 5281 #define __FUNCT__ "matsetvaluesmpiaij_" 5282 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) 5283 { 5284 Mat mat = *mmat; 5285 PetscInt m = *mm, n = *mn; 5286 InsertMode addv = *maddv; 5287 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 5288 PetscScalar value; 5289 PetscErrorCode ierr; 5290 5291 MatCheckPreallocated(mat,1); 5292 if (mat->insertmode == NOT_SET_VALUES) mat->insertmode = addv; 5293 5294 #if defined(PETSC_USE_DEBUG) 5295 else if (mat->insertmode != addv) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Cannot mix add values and insert values"); 5296 #endif 5297 { 5298 PetscInt i,j,rstart = mat->rmap->rstart,rend = mat->rmap->rend; 5299 PetscInt cstart = mat->cmap->rstart,cend = mat->cmap->rend,row,col; 5300 PetscBool roworiented = aij->roworiented; 5301 5302 /* Some Variables required in the macro */ 5303 Mat A = aij->A; 5304 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 5305 PetscInt *aimax = a->imax,*ai = a->i,*ailen = a->ilen,*aj = a->j; 5306 MatScalar *aa = a->a; 5307 PetscBool ignorezeroentries = (((a->ignorezeroentries)&&(addv==ADD_VALUES)) ? PETSC_TRUE : PETSC_FALSE); 5308 Mat B = aij->B; 5309 Mat_SeqAIJ *b = (Mat_SeqAIJ*)B->data; 5310 PetscInt *bimax = b->imax,*bi = b->i,*bilen = b->ilen,*bj = b->j,bm = aij->B->rmap->n,am = aij->A->rmap->n; 5311 MatScalar *ba = b->a; 5312 5313 PetscInt *rp1,*rp2,ii,nrow1,nrow2,_i,rmax1,rmax2,N,low1,high1,low2,high2,t,lastcol1,lastcol2; 5314 PetscInt nonew = a->nonew; 5315 MatScalar *ap1,*ap2; 5316 5317 PetscFunctionBegin; 5318 for (i=0; i<m; i++) { 5319 if (im[i] < 0) continue; 5320 #if defined(PETSC_USE_DEBUG) 5321 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); 5322 #endif 5323 if (im[i] >= rstart && im[i] < rend) { 5324 row = im[i] - rstart; 5325 lastcol1 = -1; 5326 rp1 = aj + ai[row]; 5327 ap1 = aa + ai[row]; 5328 rmax1 = aimax[row]; 5329 nrow1 = ailen[row]; 5330 low1 = 0; 5331 high1 = nrow1; 5332 lastcol2 = -1; 5333 rp2 = bj + bi[row]; 5334 ap2 = ba + bi[row]; 5335 rmax2 = bimax[row]; 5336 nrow2 = bilen[row]; 5337 low2 = 0; 5338 high2 = nrow2; 5339 5340 for (j=0; j<n; j++) { 5341 if (roworiented) value = v[i*n+j]; 5342 else value = v[i+j*m]; 5343 if (ignorezeroentries && value == 0.0 && (addv == ADD_VALUES)) continue; 5344 if (in[j] >= cstart && in[j] < cend) { 5345 col = in[j] - cstart; 5346 MatSetValues_SeqAIJ_A_Private(row,col,value,addv); 5347 } else if (in[j] < 0) continue; 5348 #if defined(PETSC_USE_DEBUG) 5349 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); 5350 #endif 5351 else { 5352 if (mat->was_assembled) { 5353 if (!aij->colmap) { 5354 ierr = MatCreateColmap_MPIAIJ_Private(mat);CHKERRQ(ierr); 5355 } 5356 #if defined(PETSC_USE_CTABLE) 5357 ierr = PetscTableFind(aij->colmap,in[j]+1,&col);CHKERRQ(ierr); 5358 col--; 5359 #else 5360 col = aij->colmap[in[j]] - 1; 5361 #endif 5362 if (col < 0 && !((Mat_SeqAIJ*)(aij->A->data))->nonew) { 5363 ierr = MatDisAssemble_MPIAIJ(mat);CHKERRQ(ierr); 5364 col = in[j]; 5365 /* Reinitialize the variables required by MatSetValues_SeqAIJ_B_Private() */ 5366 B = aij->B; 5367 b = (Mat_SeqAIJ*)B->data; 5368 bimax = b->imax; bi = b->i; bilen = b->ilen; bj = b->j; 5369 rp2 = bj + bi[row]; 5370 ap2 = ba + bi[row]; 5371 rmax2 = bimax[row]; 5372 nrow2 = bilen[row]; 5373 low2 = 0; 5374 high2 = nrow2; 5375 bm = aij->B->rmap->n; 5376 ba = b->a; 5377 } 5378 } else col = in[j]; 5379 MatSetValues_SeqAIJ_B_Private(row,col,value,addv); 5380 } 5381 } 5382 } else if (!aij->donotstash) { 5383 if (roworiented) { 5384 ierr = MatStashValuesRow_Private(&mat->stash,im[i],n,in,v+i*n,(PetscBool)(ignorezeroentries && (addv == ADD_VALUES)));CHKERRQ(ierr); 5385 } else { 5386 ierr = MatStashValuesCol_Private(&mat->stash,im[i],n,in,v+i,m,(PetscBool)(ignorezeroentries && (addv == ADD_VALUES)));CHKERRQ(ierr); 5387 } 5388 } 5389 } 5390 } 5391 PetscFunctionReturnVoid(); 5392 } 5393 5394