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->psubcomm) { 1147 ierr = PetscSubcommDestroy(&redund->psubcomm);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 2404 #undef __FUNCT__ 2405 #define __FUNCT__ "MatGetRedundantMatrix_MPIAIJ_interlaced" 2406 PetscErrorCode MatGetRedundantMatrix_MPIAIJ_interlaced(Mat mat,PetscInt nsubcomm,MPI_Comm subcomm,MatReuse reuse,Mat *matredundant) 2407 { 2408 PetscMPIInt rank,size; 2409 MPI_Comm comm; 2410 PetscErrorCode ierr; 2411 PetscInt nsends=0,nrecvs=0,i,rownz_max=0,M=mat->rmap->N,N=mat->cmap->N; 2412 PetscMPIInt *send_rank= NULL,*recv_rank=NULL,subrank,subsize; 2413 PetscInt *rowrange = mat->rmap->range; 2414 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 2415 Mat A = aij->A,B=aij->B,C=*matredundant; 2416 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data,*b=(Mat_SeqAIJ*)B->data; 2417 PetscScalar *sbuf_a; 2418 PetscInt nzlocal=a->nz+b->nz; 2419 PetscInt j,cstart=mat->cmap->rstart,cend=mat->cmap->rend,row,nzA,nzB,ncols,*cworkA,*cworkB; 2420 PetscInt rstart=mat->rmap->rstart,rend=mat->rmap->rend,*bmap=aij->garray; 2421 PetscInt *cols,ctmp,lwrite,*rptr,l,*sbuf_j; 2422 MatScalar *aworkA,*aworkB; 2423 PetscScalar *vals; 2424 PetscMPIInt tag1,tag2,tag3,imdex; 2425 MPI_Request *s_waits1=NULL,*s_waits2=NULL,*s_waits3=NULL; 2426 MPI_Request *r_waits1=NULL,*r_waits2=NULL,*r_waits3=NULL; 2427 MPI_Status recv_status,*send_status; 2428 PetscInt *sbuf_nz=NULL,*rbuf_nz=NULL,count; 2429 PetscInt **rbuf_j=NULL; 2430 PetscScalar **rbuf_a=NULL; 2431 Mat_Redundant *redund =NULL; 2432 2433 PetscFunctionBegin; 2434 ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 2435 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 2436 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 2437 ierr = MPI_Comm_rank(subcomm,&subrank);CHKERRQ(ierr); 2438 ierr = MPI_Comm_size(subcomm,&subsize);CHKERRQ(ierr); 2439 2440 if (reuse == MAT_REUSE_MATRIX) { 2441 if (M != mat->rmap->N || N != mat->cmap->N) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Cannot reuse matrix. Wrong global size"); 2442 2443 redund = C->redundant; 2444 if (nzlocal != redund->nzlocal) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Cannot reuse matrix. Wrong nzlocal"); 2445 2446 nsends = redund->nsends; 2447 nrecvs = redund->nrecvs; 2448 send_rank = redund->send_rank; 2449 recv_rank = redund->recv_rank; 2450 sbuf_nz = redund->sbuf_nz; 2451 rbuf_nz = redund->rbuf_nz; 2452 sbuf_j = redund->sbuf_j; 2453 sbuf_a = redund->sbuf_a; 2454 rbuf_j = redund->rbuf_j; 2455 rbuf_a = redund->rbuf_a; 2456 } 2457 2458 if (reuse == MAT_INITIAL_MATRIX) { 2459 PetscInt nleftover,np_subcomm; 2460 2461 /* get the destination processors' id send_rank, nsends and nrecvs */ 2462 ierr = PetscMalloc2(size,&send_rank,size,&recv_rank);CHKERRQ(ierr); 2463 2464 np_subcomm = size/nsubcomm; 2465 nleftover = size - nsubcomm*np_subcomm; 2466 2467 /* block of codes below is specific for INTERLACED */ 2468 /* ------------------------------------------------*/ 2469 nsends = 0; nrecvs = 0; 2470 for (i=0; i<size; i++) { 2471 if (subrank == i/nsubcomm && i != rank) { /* my_subrank == other's subrank */ 2472 send_rank[nsends++] = i; 2473 recv_rank[nrecvs++] = i; 2474 } 2475 } 2476 if (rank >= size - nleftover) { /* this proc is a leftover processor */ 2477 i = size-nleftover-1; 2478 j = 0; 2479 while (j < nsubcomm - nleftover) { 2480 send_rank[nsends++] = i; 2481 i--; j++; 2482 } 2483 } 2484 2485 if (nleftover && subsize == size/nsubcomm && subrank==subsize-1) { /* this proc recvs from leftover processors */ 2486 for (i=0; i<nleftover; i++) { 2487 recv_rank[nrecvs++] = size-nleftover+i; 2488 } 2489 } 2490 /*----------------------------------------------*/ 2491 2492 /* allocate sbuf_j, sbuf_a */ 2493 i = nzlocal + rowrange[rank+1] - rowrange[rank] + 2; 2494 ierr = PetscMalloc1(i,&sbuf_j);CHKERRQ(ierr); 2495 ierr = PetscMalloc1((nzlocal+1),&sbuf_a);CHKERRQ(ierr); 2496 /* 2497 ierr = PetscSynchronizedPrintf(comm,"[%d] nsends %d, nrecvs %d\n",rank,nsends,nrecvs);CHKERRQ(ierr); 2498 ierr = PetscSynchronizedFlush(comm,PETSC_STDOUT);CHKERRQ(ierr); 2499 */ 2500 } /* endof if (reuse == MAT_INITIAL_MATRIX) */ 2501 2502 /* copy mat's local entries into the buffers */ 2503 if (reuse == MAT_INITIAL_MATRIX) { 2504 rownz_max = 0; 2505 rptr = sbuf_j; 2506 cols = sbuf_j + rend-rstart + 1; 2507 vals = sbuf_a; 2508 rptr[0] = 0; 2509 for (i=0; i<rend-rstart; i++) { 2510 row = i + rstart; 2511 nzA = a->i[i+1] - a->i[i]; nzB = b->i[i+1] - b->i[i]; 2512 ncols = nzA + nzB; 2513 cworkA = a->j + a->i[i]; cworkB = b->j + b->i[i]; 2514 aworkA = a->a + a->i[i]; aworkB = b->a + b->i[i]; 2515 /* load the column indices for this row into cols */ 2516 lwrite = 0; 2517 for (l=0; l<nzB; l++) { 2518 if ((ctmp = bmap[cworkB[l]]) < cstart) { 2519 vals[lwrite] = aworkB[l]; 2520 cols[lwrite++] = ctmp; 2521 } 2522 } 2523 for (l=0; l<nzA; l++) { 2524 vals[lwrite] = aworkA[l]; 2525 cols[lwrite++] = cstart + cworkA[l]; 2526 } 2527 for (l=0; l<nzB; l++) { 2528 if ((ctmp = bmap[cworkB[l]]) >= cend) { 2529 vals[lwrite] = aworkB[l]; 2530 cols[lwrite++] = ctmp; 2531 } 2532 } 2533 vals += ncols; 2534 cols += ncols; 2535 rptr[i+1] = rptr[i] + ncols; 2536 if (rownz_max < ncols) rownz_max = ncols; 2537 } 2538 if (rptr[rend-rstart] != a->nz + b->nz) SETERRQ4(PETSC_COMM_SELF,PETSC_ERR_PLIB, "rptr[%d] %d != %d + %d",rend-rstart,rptr[rend-rstart+1],a->nz,b->nz); 2539 } else { /* only copy matrix values into sbuf_a */ 2540 rptr = sbuf_j; 2541 vals = sbuf_a; 2542 rptr[0] = 0; 2543 for (i=0; i<rend-rstart; i++) { 2544 row = i + rstart; 2545 nzA = a->i[i+1] - a->i[i]; nzB = b->i[i+1] - b->i[i]; 2546 ncols = nzA + nzB; 2547 cworkB = b->j + b->i[i]; 2548 aworkA = a->a + a->i[i]; 2549 aworkB = b->a + b->i[i]; 2550 lwrite = 0; 2551 for (l=0; l<nzB; l++) { 2552 if ((ctmp = bmap[cworkB[l]]) < cstart) vals[lwrite++] = aworkB[l]; 2553 } 2554 for (l=0; l<nzA; l++) vals[lwrite++] = aworkA[l]; 2555 for (l=0; l<nzB; l++) { 2556 if ((ctmp = bmap[cworkB[l]]) >= cend) vals[lwrite++] = aworkB[l]; 2557 } 2558 vals += ncols; 2559 rptr[i+1] = rptr[i] + ncols; 2560 } 2561 } /* endof if (reuse == MAT_INITIAL_MATRIX) */ 2562 2563 /* send nzlocal to others, and recv other's nzlocal */ 2564 /*--------------------------------------------------*/ 2565 if (reuse == MAT_INITIAL_MATRIX) { 2566 ierr = PetscMalloc2(3*(nsends + nrecvs)+1,&s_waits3,nsends+1,&send_status);CHKERRQ(ierr); 2567 2568 s_waits2 = s_waits3 + nsends; 2569 s_waits1 = s_waits2 + nsends; 2570 r_waits1 = s_waits1 + nsends; 2571 r_waits2 = r_waits1 + nrecvs; 2572 r_waits3 = r_waits2 + nrecvs; 2573 } else { 2574 ierr = PetscMalloc2(nsends + nrecvs +1,&s_waits3,nsends+1,&send_status);CHKERRQ(ierr); 2575 2576 r_waits3 = s_waits3 + nsends; 2577 } 2578 2579 ierr = PetscObjectGetNewTag((PetscObject)mat,&tag3);CHKERRQ(ierr); 2580 if (reuse == MAT_INITIAL_MATRIX) { 2581 /* get new tags to keep the communication clean */ 2582 ierr = PetscObjectGetNewTag((PetscObject)mat,&tag1);CHKERRQ(ierr); 2583 ierr = PetscObjectGetNewTag((PetscObject)mat,&tag2);CHKERRQ(ierr); 2584 ierr = PetscMalloc4(nsends,&sbuf_nz,nrecvs,&rbuf_nz,nrecvs,&rbuf_j,nrecvs,&rbuf_a);CHKERRQ(ierr); 2585 2586 /* post receives of other's nzlocal */ 2587 for (i=0; i<nrecvs; i++) { 2588 ierr = MPI_Irecv(rbuf_nz+i,1,MPIU_INT,MPI_ANY_SOURCE,tag1,comm,r_waits1+i);CHKERRQ(ierr); 2589 } 2590 /* send nzlocal to others */ 2591 for (i=0; i<nsends; i++) { 2592 sbuf_nz[i] = nzlocal; 2593 ierr = MPI_Isend(sbuf_nz+i,1,MPIU_INT,send_rank[i],tag1,comm,s_waits1+i);CHKERRQ(ierr); 2594 } 2595 /* wait on receives of nzlocal; allocate space for rbuf_j, rbuf_a */ 2596 count = nrecvs; 2597 while (count) { 2598 ierr = MPI_Waitany(nrecvs,r_waits1,&imdex,&recv_status);CHKERRQ(ierr); 2599 2600 recv_rank[imdex] = recv_status.MPI_SOURCE; 2601 /* allocate rbuf_a and rbuf_j; then post receives of rbuf_j */ 2602 ierr = PetscMalloc1((rbuf_nz[imdex]+1),&rbuf_a[imdex]);CHKERRQ(ierr); 2603 2604 i = rowrange[recv_status.MPI_SOURCE+1] - rowrange[recv_status.MPI_SOURCE]; /* number of expected mat->i */ 2605 2606 rbuf_nz[imdex] += i + 2; 2607 2608 ierr = PetscMalloc1(rbuf_nz[imdex],&rbuf_j[imdex]);CHKERRQ(ierr); 2609 ierr = MPI_Irecv(rbuf_j[imdex],rbuf_nz[imdex],MPIU_INT,recv_status.MPI_SOURCE,tag2,comm,r_waits2+imdex);CHKERRQ(ierr); 2610 count--; 2611 } 2612 /* wait on sends of nzlocal */ 2613 if (nsends) {ierr = MPI_Waitall(nsends,s_waits1,send_status);CHKERRQ(ierr);} 2614 /* send mat->i,j to others, and recv from other's */ 2615 /*------------------------------------------------*/ 2616 for (i=0; i<nsends; i++) { 2617 j = nzlocal + rowrange[rank+1] - rowrange[rank] + 1; 2618 ierr = MPI_Isend(sbuf_j,j,MPIU_INT,send_rank[i],tag2,comm,s_waits2+i);CHKERRQ(ierr); 2619 } 2620 /* wait on receives of mat->i,j */ 2621 /*------------------------------*/ 2622 count = nrecvs; 2623 while (count) { 2624 ierr = MPI_Waitany(nrecvs,r_waits2,&imdex,&recv_status);CHKERRQ(ierr); 2625 if (recv_rank[imdex] != recv_status.MPI_SOURCE) SETERRQ2(PETSC_COMM_SELF,1, "recv_rank %d != MPI_SOURCE %d",recv_rank[imdex],recv_status.MPI_SOURCE); 2626 count--; 2627 } 2628 /* wait on sends of mat->i,j */ 2629 /*---------------------------*/ 2630 if (nsends) { 2631 ierr = MPI_Waitall(nsends,s_waits2,send_status);CHKERRQ(ierr); 2632 } 2633 } /* endof if (reuse == MAT_INITIAL_MATRIX) */ 2634 2635 /* post receives, send and receive mat->a */ 2636 /*----------------------------------------*/ 2637 for (imdex=0; imdex<nrecvs; imdex++) { 2638 ierr = MPI_Irecv(rbuf_a[imdex],rbuf_nz[imdex],MPIU_SCALAR,recv_rank[imdex],tag3,comm,r_waits3+imdex);CHKERRQ(ierr); 2639 } 2640 for (i=0; i<nsends; i++) { 2641 ierr = MPI_Isend(sbuf_a,nzlocal,MPIU_SCALAR,send_rank[i],tag3,comm,s_waits3+i);CHKERRQ(ierr); 2642 } 2643 count = nrecvs; 2644 while (count) { 2645 ierr = MPI_Waitany(nrecvs,r_waits3,&imdex,&recv_status);CHKERRQ(ierr); 2646 if (recv_rank[imdex] != recv_status.MPI_SOURCE) SETERRQ2(PETSC_COMM_SELF,1, "recv_rank %d != MPI_SOURCE %d",recv_rank[imdex],recv_status.MPI_SOURCE); 2647 count--; 2648 } 2649 if (nsends) { 2650 ierr = MPI_Waitall(nsends,s_waits3,send_status);CHKERRQ(ierr); 2651 } 2652 2653 ierr = PetscFree2(s_waits3,send_status);CHKERRQ(ierr); 2654 2655 /* create redundant matrix */ 2656 /*-------------------------*/ 2657 if (reuse == MAT_INITIAL_MATRIX) { 2658 const PetscInt *range; 2659 PetscInt rstart_sub,rend_sub,mloc_sub; 2660 2661 /* compute rownz_max for preallocation */ 2662 for (imdex=0; imdex<nrecvs; imdex++) { 2663 j = rowrange[recv_rank[imdex]+1] - rowrange[recv_rank[imdex]]; 2664 rptr = rbuf_j[imdex]; 2665 for (i=0; i<j; i++) { 2666 ncols = rptr[i+1] - rptr[i]; 2667 if (rownz_max < ncols) rownz_max = ncols; 2668 } 2669 } 2670 2671 ierr = MatCreate(subcomm,&C);CHKERRQ(ierr); 2672 2673 /* get local size of redundant matrix 2674 - mloc_sub is chosen for PETSC_SUBCOMM_INTERLACED, works for other types, but may not efficient! */ 2675 ierr = MatGetOwnershipRanges(mat,&range);CHKERRQ(ierr); 2676 rstart_sub = range[nsubcomm*subrank]; 2677 if (subrank+1 < subsize) { /* not the last proc in subcomm */ 2678 rend_sub = range[nsubcomm*(subrank+1)]; 2679 } else { 2680 rend_sub = mat->rmap->N; 2681 } 2682 mloc_sub = rend_sub - rstart_sub; 2683 2684 if (M == N) { 2685 ierr = MatSetSizes(C,mloc_sub,mloc_sub,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 2686 } else { /* non-square matrix */ 2687 ierr = MatSetSizes(C,mloc_sub,PETSC_DECIDE,PETSC_DECIDE,mat->cmap->N);CHKERRQ(ierr); 2688 } 2689 ierr = MatSetBlockSizesFromMats(C,mat,mat);CHKERRQ(ierr); 2690 ierr = MatSetFromOptions(C);CHKERRQ(ierr); 2691 ierr = MatSeqAIJSetPreallocation(C,rownz_max,NULL);CHKERRQ(ierr); 2692 ierr = MatMPIAIJSetPreallocation(C,rownz_max,NULL,rownz_max,NULL);CHKERRQ(ierr); 2693 } else { 2694 C = *matredundant; 2695 } 2696 2697 /* insert local matrix entries */ 2698 rptr = sbuf_j; 2699 cols = sbuf_j + rend-rstart + 1; 2700 vals = sbuf_a; 2701 for (i=0; i<rend-rstart; i++) { 2702 row = i + rstart; 2703 ncols = rptr[i+1] - rptr[i]; 2704 ierr = MatSetValues(C,1,&row,ncols,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 2705 vals += ncols; 2706 cols += ncols; 2707 } 2708 /* insert received matrix entries */ 2709 for (imdex=0; imdex<nrecvs; imdex++) { 2710 rstart = rowrange[recv_rank[imdex]]; 2711 rend = rowrange[recv_rank[imdex]+1]; 2712 /* printf("[%d] insert rows %d - %d\n",rank,rstart,rend-1); */ 2713 rptr = rbuf_j[imdex]; 2714 cols = rbuf_j[imdex] + rend-rstart + 1; 2715 vals = rbuf_a[imdex]; 2716 for (i=0; i<rend-rstart; i++) { 2717 row = i + rstart; 2718 ncols = rptr[i+1] - rptr[i]; 2719 ierr = MatSetValues(C,1,&row,ncols,cols,vals,INSERT_VALUES);CHKERRQ(ierr); 2720 vals += ncols; 2721 cols += ncols; 2722 } 2723 } 2724 ierr = MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2725 ierr = MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2726 2727 if (reuse == MAT_INITIAL_MATRIX) { 2728 *matredundant = C; 2729 2730 /* create a supporting struct and attach it to C for reuse */ 2731 ierr = PetscNewLog(C,&redund);CHKERRQ(ierr); 2732 C->redundant = redund; 2733 redund->nzlocal = nzlocal; 2734 redund->nsends = nsends; 2735 redund->nrecvs = nrecvs; 2736 redund->send_rank = send_rank; 2737 redund->recv_rank = recv_rank; 2738 redund->sbuf_nz = sbuf_nz; 2739 redund->rbuf_nz = rbuf_nz; 2740 redund->sbuf_j = sbuf_j; 2741 redund->sbuf_a = sbuf_a; 2742 redund->rbuf_j = rbuf_j; 2743 redund->rbuf_a = rbuf_a; 2744 redund->psubcomm = NULL; 2745 } 2746 PetscFunctionReturn(0); 2747 } 2748 2749 #undef __FUNCT__ 2750 #define __FUNCT__ "MatGetRedundantMatrix_MPIAIJ" 2751 PetscErrorCode MatGetRedundantMatrix_MPIAIJ(Mat mat,PetscInt nsubcomm,MPI_Comm subcomm,MatReuse reuse,Mat *matredundant) 2752 { 2753 PetscErrorCode ierr; 2754 MPI_Comm comm; 2755 PetscMPIInt size,subsize; 2756 PetscInt mloc_sub,rstart,rend,M=mat->rmap->N,N=mat->cmap->N; 2757 Mat_Redundant *redund=NULL; 2758 PetscSubcomm psubcomm=NULL; 2759 MPI_Comm subcomm_in=subcomm; 2760 Mat *matseq; 2761 IS isrow,iscol; 2762 2763 PetscFunctionBegin; 2764 if (subcomm_in == MPI_COMM_NULL) { /* user does not provide subcomm */ 2765 if (reuse == MAT_INITIAL_MATRIX) { 2766 /* create psubcomm, then get subcomm */ 2767 ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 2768 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 2769 if (nsubcomm < 1 || nsubcomm > size) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"nsubcomm must between 1 and %D",size); 2770 2771 ierr = PetscSubcommCreate(comm,&psubcomm);CHKERRQ(ierr); 2772 ierr = PetscSubcommSetNumber(psubcomm,nsubcomm);CHKERRQ(ierr); 2773 ierr = PetscSubcommSetType(psubcomm,PETSC_SUBCOMM_CONTIGUOUS);CHKERRQ(ierr); 2774 ierr = PetscSubcommSetFromOptions(psubcomm);CHKERRQ(ierr); 2775 subcomm = psubcomm->comm; 2776 } else { /* retrieve psubcomm and subcomm */ 2777 ierr = PetscObjectGetComm((PetscObject)(*matredundant),&subcomm);CHKERRQ(ierr); 2778 ierr = MPI_Comm_size(subcomm,&subsize);CHKERRQ(ierr); 2779 redund = (*matredundant)->redundant; 2780 psubcomm = redund->psubcomm; 2781 } 2782 if (psubcomm->type == PETSC_SUBCOMM_INTERLACED) { 2783 ierr = MatGetRedundantMatrix_MPIAIJ_interlaced(mat,nsubcomm,subcomm,reuse,matredundant);CHKERRQ(ierr); 2784 if (reuse == MAT_INITIAL_MATRIX) { /* psubcomm is created in this routine, free it in MatDestroy_Redundant() */ 2785 ierr = MPI_Comm_size(psubcomm->comm,&subsize);CHKERRQ(ierr); 2786 (*matredundant)->redundant->psubcomm = psubcomm; 2787 } 2788 PetscFunctionReturn(0); 2789 } 2790 } 2791 2792 /* use MPI subcomm via MatGetSubMatrices(); use subcomm_in or psubcomm->comm (psubcomm->type != INTERLACED) */ 2793 ierr = MPI_Comm_size(subcomm,&subsize);CHKERRQ(ierr); 2794 if (reuse == MAT_INITIAL_MATRIX) { 2795 /* create a local sequential matrix matseq[0] */ 2796 mloc_sub = PETSC_DECIDE; 2797 ierr = PetscSplitOwnership(subcomm,&mloc_sub,&M);CHKERRQ(ierr); 2798 ierr = MPI_Scan(&mloc_sub,&rend,1,MPIU_INT,MPI_SUM,subcomm);CHKERRQ(ierr); 2799 rstart = rend - mloc_sub; 2800 ierr = ISCreateStride(PETSC_COMM_SELF,mloc_sub,rstart,1,&isrow);CHKERRQ(ierr); 2801 ierr = ISCreateStride(PETSC_COMM_SELF,N,0,1,&iscol);CHKERRQ(ierr); 2802 } else { /* reuse == MAT_REUSE_MATRIX */ 2803 redund = (*matredundant)->redundant; 2804 isrow = redund->isrow; 2805 iscol = redund->iscol; 2806 matseq = redund->matseq; 2807 } 2808 ierr = MatGetSubMatrices(mat,1,&isrow,&iscol,reuse,&matseq);CHKERRQ(ierr); 2809 ierr = MatCreateMPIMatConcatenateSeqMat(subcomm,matseq[0],PETSC_DECIDE,reuse,matredundant);CHKERRQ(ierr); 2810 2811 if (reuse == MAT_INITIAL_MATRIX) { 2812 /* create a supporting struct and attach it to C for reuse */ 2813 ierr = PetscNewLog(*matredundant,&redund);CHKERRQ(ierr); 2814 (*matredundant)->redundant = redund; 2815 redund->isrow = isrow; 2816 redund->iscol = iscol; 2817 redund->matseq = matseq; 2818 redund->psubcomm = psubcomm; 2819 } 2820 PetscFunctionReturn(0); 2821 } 2822 2823 #undef __FUNCT__ 2824 #define __FUNCT__ "MatGetRowMaxAbs_MPIAIJ" 2825 PetscErrorCode MatGetRowMaxAbs_MPIAIJ(Mat A, Vec v, PetscInt idx[]) 2826 { 2827 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 2828 PetscErrorCode ierr; 2829 PetscInt i,*idxb = 0; 2830 PetscScalar *va,*vb; 2831 Vec vtmp; 2832 2833 PetscFunctionBegin; 2834 ierr = MatGetRowMaxAbs(a->A,v,idx);CHKERRQ(ierr); 2835 ierr = VecGetArray(v,&va);CHKERRQ(ierr); 2836 if (idx) { 2837 for (i=0; i<A->rmap->n; i++) { 2838 if (PetscAbsScalar(va[i])) idx[i] += A->cmap->rstart; 2839 } 2840 } 2841 2842 ierr = VecCreateSeq(PETSC_COMM_SELF,A->rmap->n,&vtmp);CHKERRQ(ierr); 2843 if (idx) { 2844 ierr = PetscMalloc1(A->rmap->n,&idxb);CHKERRQ(ierr); 2845 } 2846 ierr = MatGetRowMaxAbs(a->B,vtmp,idxb);CHKERRQ(ierr); 2847 ierr = VecGetArray(vtmp,&vb);CHKERRQ(ierr); 2848 2849 for (i=0; i<A->rmap->n; i++) { 2850 if (PetscAbsScalar(va[i]) < PetscAbsScalar(vb[i])) { 2851 va[i] = vb[i]; 2852 if (idx) idx[i] = a->garray[idxb[i]]; 2853 } 2854 } 2855 2856 ierr = VecRestoreArray(v,&va);CHKERRQ(ierr); 2857 ierr = VecRestoreArray(vtmp,&vb);CHKERRQ(ierr); 2858 ierr = PetscFree(idxb);CHKERRQ(ierr); 2859 ierr = VecDestroy(&vtmp);CHKERRQ(ierr); 2860 PetscFunctionReturn(0); 2861 } 2862 2863 #undef __FUNCT__ 2864 #define __FUNCT__ "MatGetRowMinAbs_MPIAIJ" 2865 PetscErrorCode MatGetRowMinAbs_MPIAIJ(Mat A, Vec v, PetscInt idx[]) 2866 { 2867 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 2868 PetscErrorCode ierr; 2869 PetscInt i,*idxb = 0; 2870 PetscScalar *va,*vb; 2871 Vec vtmp; 2872 2873 PetscFunctionBegin; 2874 ierr = MatGetRowMinAbs(a->A,v,idx);CHKERRQ(ierr); 2875 ierr = VecGetArray(v,&va);CHKERRQ(ierr); 2876 if (idx) { 2877 for (i=0; i<A->cmap->n; i++) { 2878 if (PetscAbsScalar(va[i])) idx[i] += A->cmap->rstart; 2879 } 2880 } 2881 2882 ierr = VecCreateSeq(PETSC_COMM_SELF,A->rmap->n,&vtmp);CHKERRQ(ierr); 2883 if (idx) { 2884 ierr = PetscMalloc1(A->rmap->n,&idxb);CHKERRQ(ierr); 2885 } 2886 ierr = MatGetRowMinAbs(a->B,vtmp,idxb);CHKERRQ(ierr); 2887 ierr = VecGetArray(vtmp,&vb);CHKERRQ(ierr); 2888 2889 for (i=0; i<A->rmap->n; i++) { 2890 if (PetscAbsScalar(va[i]) > PetscAbsScalar(vb[i])) { 2891 va[i] = vb[i]; 2892 if (idx) idx[i] = a->garray[idxb[i]]; 2893 } 2894 } 2895 2896 ierr = VecRestoreArray(v,&va);CHKERRQ(ierr); 2897 ierr = VecRestoreArray(vtmp,&vb);CHKERRQ(ierr); 2898 ierr = PetscFree(idxb);CHKERRQ(ierr); 2899 ierr = VecDestroy(&vtmp);CHKERRQ(ierr); 2900 PetscFunctionReturn(0); 2901 } 2902 2903 #undef __FUNCT__ 2904 #define __FUNCT__ "MatGetRowMin_MPIAIJ" 2905 PetscErrorCode MatGetRowMin_MPIAIJ(Mat A, Vec v, PetscInt idx[]) 2906 { 2907 Mat_MPIAIJ *mat = (Mat_MPIAIJ*) A->data; 2908 PetscInt n = A->rmap->n; 2909 PetscInt cstart = A->cmap->rstart; 2910 PetscInt *cmap = mat->garray; 2911 PetscInt *diagIdx, *offdiagIdx; 2912 Vec diagV, offdiagV; 2913 PetscScalar *a, *diagA, *offdiagA; 2914 PetscInt r; 2915 PetscErrorCode ierr; 2916 2917 PetscFunctionBegin; 2918 ierr = PetscMalloc2(n,&diagIdx,n,&offdiagIdx);CHKERRQ(ierr); 2919 ierr = VecCreateSeq(PetscObjectComm((PetscObject)A), n, &diagV);CHKERRQ(ierr); 2920 ierr = VecCreateSeq(PetscObjectComm((PetscObject)A), n, &offdiagV);CHKERRQ(ierr); 2921 ierr = MatGetRowMin(mat->A, diagV, diagIdx);CHKERRQ(ierr); 2922 ierr = MatGetRowMin(mat->B, offdiagV, offdiagIdx);CHKERRQ(ierr); 2923 ierr = VecGetArray(v, &a);CHKERRQ(ierr); 2924 ierr = VecGetArray(diagV, &diagA);CHKERRQ(ierr); 2925 ierr = VecGetArray(offdiagV, &offdiagA);CHKERRQ(ierr); 2926 for (r = 0; r < n; ++r) { 2927 if (PetscAbsScalar(diagA[r]) <= PetscAbsScalar(offdiagA[r])) { 2928 a[r] = diagA[r]; 2929 idx[r] = cstart + diagIdx[r]; 2930 } else { 2931 a[r] = offdiagA[r]; 2932 idx[r] = cmap[offdiagIdx[r]]; 2933 } 2934 } 2935 ierr = VecRestoreArray(v, &a);CHKERRQ(ierr); 2936 ierr = VecRestoreArray(diagV, &diagA);CHKERRQ(ierr); 2937 ierr = VecRestoreArray(offdiagV, &offdiagA);CHKERRQ(ierr); 2938 ierr = VecDestroy(&diagV);CHKERRQ(ierr); 2939 ierr = VecDestroy(&offdiagV);CHKERRQ(ierr); 2940 ierr = PetscFree2(diagIdx, offdiagIdx);CHKERRQ(ierr); 2941 PetscFunctionReturn(0); 2942 } 2943 2944 #undef __FUNCT__ 2945 #define __FUNCT__ "MatGetRowMax_MPIAIJ" 2946 PetscErrorCode MatGetRowMax_MPIAIJ(Mat A, Vec v, PetscInt idx[]) 2947 { 2948 Mat_MPIAIJ *mat = (Mat_MPIAIJ*) A->data; 2949 PetscInt n = A->rmap->n; 2950 PetscInt cstart = A->cmap->rstart; 2951 PetscInt *cmap = mat->garray; 2952 PetscInt *diagIdx, *offdiagIdx; 2953 Vec diagV, offdiagV; 2954 PetscScalar *a, *diagA, *offdiagA; 2955 PetscInt r; 2956 PetscErrorCode ierr; 2957 2958 PetscFunctionBegin; 2959 ierr = PetscMalloc2(n,&diagIdx,n,&offdiagIdx);CHKERRQ(ierr); 2960 ierr = VecCreateSeq(PETSC_COMM_SELF, n, &diagV);CHKERRQ(ierr); 2961 ierr = VecCreateSeq(PETSC_COMM_SELF, n, &offdiagV);CHKERRQ(ierr); 2962 ierr = MatGetRowMax(mat->A, diagV, diagIdx);CHKERRQ(ierr); 2963 ierr = MatGetRowMax(mat->B, offdiagV, offdiagIdx);CHKERRQ(ierr); 2964 ierr = VecGetArray(v, &a);CHKERRQ(ierr); 2965 ierr = VecGetArray(diagV, &diagA);CHKERRQ(ierr); 2966 ierr = VecGetArray(offdiagV, &offdiagA);CHKERRQ(ierr); 2967 for (r = 0; r < n; ++r) { 2968 if (PetscAbsScalar(diagA[r]) >= PetscAbsScalar(offdiagA[r])) { 2969 a[r] = diagA[r]; 2970 idx[r] = cstart + diagIdx[r]; 2971 } else { 2972 a[r] = offdiagA[r]; 2973 idx[r] = cmap[offdiagIdx[r]]; 2974 } 2975 } 2976 ierr = VecRestoreArray(v, &a);CHKERRQ(ierr); 2977 ierr = VecRestoreArray(diagV, &diagA);CHKERRQ(ierr); 2978 ierr = VecRestoreArray(offdiagV, &offdiagA);CHKERRQ(ierr); 2979 ierr = VecDestroy(&diagV);CHKERRQ(ierr); 2980 ierr = VecDestroy(&offdiagV);CHKERRQ(ierr); 2981 ierr = PetscFree2(diagIdx, offdiagIdx);CHKERRQ(ierr); 2982 PetscFunctionReturn(0); 2983 } 2984 2985 #undef __FUNCT__ 2986 #define __FUNCT__ "MatGetSeqNonzeroStructure_MPIAIJ" 2987 PetscErrorCode MatGetSeqNonzeroStructure_MPIAIJ(Mat mat,Mat *newmat) 2988 { 2989 PetscErrorCode ierr; 2990 Mat *dummy; 2991 2992 PetscFunctionBegin; 2993 ierr = MatGetSubMatrix_MPIAIJ_All(mat,MAT_DO_NOT_GET_VALUES,MAT_INITIAL_MATRIX,&dummy);CHKERRQ(ierr); 2994 *newmat = *dummy; 2995 ierr = PetscFree(dummy);CHKERRQ(ierr); 2996 PetscFunctionReturn(0); 2997 } 2998 2999 #undef __FUNCT__ 3000 #define __FUNCT__ "MatInvertBlockDiagonal_MPIAIJ" 3001 PetscErrorCode MatInvertBlockDiagonal_MPIAIJ(Mat A,const PetscScalar **values) 3002 { 3003 Mat_MPIAIJ *a = (Mat_MPIAIJ*) A->data; 3004 PetscErrorCode ierr; 3005 3006 PetscFunctionBegin; 3007 ierr = MatInvertBlockDiagonal(a->A,values);CHKERRQ(ierr); 3008 PetscFunctionReturn(0); 3009 } 3010 3011 #undef __FUNCT__ 3012 #define __FUNCT__ "MatSetRandom_MPIAIJ" 3013 static PetscErrorCode MatSetRandom_MPIAIJ(Mat x,PetscRandom rctx) 3014 { 3015 PetscErrorCode ierr; 3016 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)x->data; 3017 3018 PetscFunctionBegin; 3019 ierr = MatSetRandom(aij->A,rctx);CHKERRQ(ierr); 3020 ierr = MatSetRandom(aij->B,rctx);CHKERRQ(ierr); 3021 ierr = MatAssemblyBegin(x,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3022 ierr = MatAssemblyEnd(x,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3023 PetscFunctionReturn(0); 3024 } 3025 3026 /* -------------------------------------------------------------------*/ 3027 static struct _MatOps MatOps_Values = {MatSetValues_MPIAIJ, 3028 MatGetRow_MPIAIJ, 3029 MatRestoreRow_MPIAIJ, 3030 MatMult_MPIAIJ, 3031 /* 4*/ MatMultAdd_MPIAIJ, 3032 MatMultTranspose_MPIAIJ, 3033 MatMultTransposeAdd_MPIAIJ, 3034 #if defined(PETSC_HAVE_PBGL) 3035 MatSolve_MPIAIJ, 3036 #else 3037 0, 3038 #endif 3039 0, 3040 0, 3041 /*10*/ 0, 3042 0, 3043 0, 3044 MatSOR_MPIAIJ, 3045 MatTranspose_MPIAIJ, 3046 /*15*/ MatGetInfo_MPIAIJ, 3047 MatEqual_MPIAIJ, 3048 MatGetDiagonal_MPIAIJ, 3049 MatDiagonalScale_MPIAIJ, 3050 MatNorm_MPIAIJ, 3051 /*20*/ MatAssemblyBegin_MPIAIJ, 3052 MatAssemblyEnd_MPIAIJ, 3053 MatSetOption_MPIAIJ, 3054 MatZeroEntries_MPIAIJ, 3055 /*24*/ MatZeroRows_MPIAIJ, 3056 0, 3057 #if defined(PETSC_HAVE_PBGL) 3058 0, 3059 #else 3060 0, 3061 #endif 3062 0, 3063 0, 3064 /*29*/ MatSetUp_MPIAIJ, 3065 #if defined(PETSC_HAVE_PBGL) 3066 0, 3067 #else 3068 0, 3069 #endif 3070 0, 3071 0, 3072 0, 3073 /*34*/ MatDuplicate_MPIAIJ, 3074 0, 3075 0, 3076 0, 3077 0, 3078 /*39*/ MatAXPY_MPIAIJ, 3079 MatGetSubMatrices_MPIAIJ, 3080 MatIncreaseOverlap_MPIAIJ, 3081 MatGetValues_MPIAIJ, 3082 MatCopy_MPIAIJ, 3083 /*44*/ MatGetRowMax_MPIAIJ, 3084 MatScale_MPIAIJ, 3085 0, 3086 MatDiagonalSet_MPIAIJ, 3087 MatZeroRowsColumns_MPIAIJ, 3088 /*49*/ MatSetRandom_MPIAIJ, 3089 0, 3090 0, 3091 0, 3092 0, 3093 /*54*/ MatFDColoringCreate_MPIXAIJ, 3094 0, 3095 MatSetUnfactored_MPIAIJ, 3096 MatPermute_MPIAIJ, 3097 0, 3098 /*59*/ MatGetSubMatrix_MPIAIJ, 3099 MatDestroy_MPIAIJ, 3100 MatView_MPIAIJ, 3101 0, 3102 MatMatMatMult_MPIAIJ_MPIAIJ_MPIAIJ, 3103 /*64*/ MatMatMatMultSymbolic_MPIAIJ_MPIAIJ_MPIAIJ, 3104 MatMatMatMultNumeric_MPIAIJ_MPIAIJ_MPIAIJ, 3105 0, 3106 0, 3107 0, 3108 /*69*/ MatGetRowMaxAbs_MPIAIJ, 3109 MatGetRowMinAbs_MPIAIJ, 3110 0, 3111 MatSetColoring_MPIAIJ, 3112 0, 3113 MatSetValuesAdifor_MPIAIJ, 3114 /*75*/ MatFDColoringApply_AIJ, 3115 0, 3116 0, 3117 0, 3118 MatFindZeroDiagonals_MPIAIJ, 3119 /*80*/ 0, 3120 0, 3121 0, 3122 /*83*/ MatLoad_MPIAIJ, 3123 0, 3124 0, 3125 0, 3126 0, 3127 0, 3128 /*89*/ MatMatMult_MPIAIJ_MPIAIJ, 3129 MatMatMultSymbolic_MPIAIJ_MPIAIJ, 3130 MatMatMultNumeric_MPIAIJ_MPIAIJ, 3131 MatPtAP_MPIAIJ_MPIAIJ, 3132 MatPtAPSymbolic_MPIAIJ_MPIAIJ, 3133 /*94*/ MatPtAPNumeric_MPIAIJ_MPIAIJ, 3134 0, 3135 0, 3136 0, 3137 0, 3138 /*99*/ 0, 3139 0, 3140 0, 3141 MatConjugate_MPIAIJ, 3142 0, 3143 /*104*/MatSetValuesRow_MPIAIJ, 3144 MatRealPart_MPIAIJ, 3145 MatImaginaryPart_MPIAIJ, 3146 0, 3147 0, 3148 /*109*/0, 3149 MatGetRedundantMatrix_MPIAIJ, 3150 MatGetRowMin_MPIAIJ, 3151 0, 3152 0, 3153 /*114*/MatGetSeqNonzeroStructure_MPIAIJ, 3154 0, 3155 0, 3156 0, 3157 0, 3158 /*119*/0, 3159 0, 3160 0, 3161 0, 3162 MatGetMultiProcBlock_MPIAIJ, 3163 /*124*/MatFindNonzeroRows_MPIAIJ, 3164 MatGetColumnNorms_MPIAIJ, 3165 MatInvertBlockDiagonal_MPIAIJ, 3166 0, 3167 MatGetSubMatricesParallel_MPIAIJ, 3168 /*129*/0, 3169 MatTransposeMatMult_MPIAIJ_MPIAIJ, 3170 MatTransposeMatMultSymbolic_MPIAIJ_MPIAIJ, 3171 MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ, 3172 0, 3173 /*134*/0, 3174 0, 3175 0, 3176 0, 3177 0, 3178 /*139*/0, 3179 0, 3180 0, 3181 MatFDColoringSetUp_MPIXAIJ, 3182 0, 3183 /*144*/MatCreateMPIMatConcatenateSeqMat_MPIAIJ 3184 }; 3185 3186 /* ----------------------------------------------------------------------------------------*/ 3187 3188 #undef __FUNCT__ 3189 #define __FUNCT__ "MatStoreValues_MPIAIJ" 3190 PetscErrorCode MatStoreValues_MPIAIJ(Mat mat) 3191 { 3192 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 3193 PetscErrorCode ierr; 3194 3195 PetscFunctionBegin; 3196 ierr = MatStoreValues(aij->A);CHKERRQ(ierr); 3197 ierr = MatStoreValues(aij->B);CHKERRQ(ierr); 3198 PetscFunctionReturn(0); 3199 } 3200 3201 #undef __FUNCT__ 3202 #define __FUNCT__ "MatRetrieveValues_MPIAIJ" 3203 PetscErrorCode MatRetrieveValues_MPIAIJ(Mat mat) 3204 { 3205 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 3206 PetscErrorCode ierr; 3207 3208 PetscFunctionBegin; 3209 ierr = MatRetrieveValues(aij->A);CHKERRQ(ierr); 3210 ierr = MatRetrieveValues(aij->B);CHKERRQ(ierr); 3211 PetscFunctionReturn(0); 3212 } 3213 3214 #undef __FUNCT__ 3215 #define __FUNCT__ "MatMPIAIJSetPreallocation_MPIAIJ" 3216 PetscErrorCode MatMPIAIJSetPreallocation_MPIAIJ(Mat B,PetscInt d_nz,const PetscInt d_nnz[],PetscInt o_nz,const PetscInt o_nnz[]) 3217 { 3218 Mat_MPIAIJ *b; 3219 PetscErrorCode ierr; 3220 3221 PetscFunctionBegin; 3222 ierr = PetscLayoutSetUp(B->rmap);CHKERRQ(ierr); 3223 ierr = PetscLayoutSetUp(B->cmap);CHKERRQ(ierr); 3224 b = (Mat_MPIAIJ*)B->data; 3225 3226 if (!B->preallocated) { 3227 /* Explicitly create 2 MATSEQAIJ matrices. */ 3228 ierr = MatCreate(PETSC_COMM_SELF,&b->A);CHKERRQ(ierr); 3229 ierr = MatSetSizes(b->A,B->rmap->n,B->cmap->n,B->rmap->n,B->cmap->n);CHKERRQ(ierr); 3230 ierr = MatSetBlockSizesFromMats(b->A,B,B);CHKERRQ(ierr); 3231 ierr = MatSetType(b->A,MATSEQAIJ);CHKERRQ(ierr); 3232 ierr = PetscLogObjectParent((PetscObject)B,(PetscObject)b->A);CHKERRQ(ierr); 3233 ierr = MatCreate(PETSC_COMM_SELF,&b->B);CHKERRQ(ierr); 3234 ierr = MatSetSizes(b->B,B->rmap->n,B->cmap->N,B->rmap->n,B->cmap->N);CHKERRQ(ierr); 3235 ierr = MatSetBlockSizesFromMats(b->B,B,B);CHKERRQ(ierr); 3236 ierr = MatSetType(b->B,MATSEQAIJ);CHKERRQ(ierr); 3237 ierr = PetscLogObjectParent((PetscObject)B,(PetscObject)b->B);CHKERRQ(ierr); 3238 } 3239 3240 ierr = MatSeqAIJSetPreallocation(b->A,d_nz,d_nnz);CHKERRQ(ierr); 3241 ierr = MatSeqAIJSetPreallocation(b->B,o_nz,o_nnz);CHKERRQ(ierr); 3242 B->preallocated = PETSC_TRUE; 3243 PetscFunctionReturn(0); 3244 } 3245 3246 #undef __FUNCT__ 3247 #define __FUNCT__ "MatDuplicate_MPIAIJ" 3248 PetscErrorCode MatDuplicate_MPIAIJ(Mat matin,MatDuplicateOption cpvalues,Mat *newmat) 3249 { 3250 Mat mat; 3251 Mat_MPIAIJ *a,*oldmat = (Mat_MPIAIJ*)matin->data; 3252 PetscErrorCode ierr; 3253 3254 PetscFunctionBegin; 3255 *newmat = 0; 3256 ierr = MatCreate(PetscObjectComm((PetscObject)matin),&mat);CHKERRQ(ierr); 3257 ierr = MatSetSizes(mat,matin->rmap->n,matin->cmap->n,matin->rmap->N,matin->cmap->N);CHKERRQ(ierr); 3258 ierr = MatSetBlockSizesFromMats(mat,matin,matin);CHKERRQ(ierr); 3259 ierr = MatSetType(mat,((PetscObject)matin)->type_name);CHKERRQ(ierr); 3260 ierr = PetscMemcpy(mat->ops,matin->ops,sizeof(struct _MatOps));CHKERRQ(ierr); 3261 a = (Mat_MPIAIJ*)mat->data; 3262 3263 mat->factortype = matin->factortype; 3264 mat->assembled = PETSC_TRUE; 3265 mat->insertmode = NOT_SET_VALUES; 3266 mat->preallocated = PETSC_TRUE; 3267 3268 a->size = oldmat->size; 3269 a->rank = oldmat->rank; 3270 a->donotstash = oldmat->donotstash; 3271 a->roworiented = oldmat->roworiented; 3272 a->rowindices = 0; 3273 a->rowvalues = 0; 3274 a->getrowactive = PETSC_FALSE; 3275 3276 ierr = PetscLayoutReference(matin->rmap,&mat->rmap);CHKERRQ(ierr); 3277 ierr = PetscLayoutReference(matin->cmap,&mat->cmap);CHKERRQ(ierr); 3278 3279 if (oldmat->colmap) { 3280 #if defined(PETSC_USE_CTABLE) 3281 ierr = PetscTableCreateCopy(oldmat->colmap,&a->colmap);CHKERRQ(ierr); 3282 #else 3283 ierr = PetscMalloc1((mat->cmap->N),&a->colmap);CHKERRQ(ierr); 3284 ierr = PetscLogObjectMemory((PetscObject)mat,(mat->cmap->N)*sizeof(PetscInt));CHKERRQ(ierr); 3285 ierr = PetscMemcpy(a->colmap,oldmat->colmap,(mat->cmap->N)*sizeof(PetscInt));CHKERRQ(ierr); 3286 #endif 3287 } else a->colmap = 0; 3288 if (oldmat->garray) { 3289 PetscInt len; 3290 len = oldmat->B->cmap->n; 3291 ierr = PetscMalloc1((len+1),&a->garray);CHKERRQ(ierr); 3292 ierr = PetscLogObjectMemory((PetscObject)mat,len*sizeof(PetscInt));CHKERRQ(ierr); 3293 if (len) { ierr = PetscMemcpy(a->garray,oldmat->garray,len*sizeof(PetscInt));CHKERRQ(ierr); } 3294 } else a->garray = 0; 3295 3296 ierr = VecDuplicate(oldmat->lvec,&a->lvec);CHKERRQ(ierr); 3297 ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)a->lvec);CHKERRQ(ierr); 3298 ierr = VecScatterCopy(oldmat->Mvctx,&a->Mvctx);CHKERRQ(ierr); 3299 ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)a->Mvctx);CHKERRQ(ierr); 3300 ierr = MatDuplicate(oldmat->A,cpvalues,&a->A);CHKERRQ(ierr); 3301 ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)a->A);CHKERRQ(ierr); 3302 ierr = MatDuplicate(oldmat->B,cpvalues,&a->B);CHKERRQ(ierr); 3303 ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)a->B);CHKERRQ(ierr); 3304 ierr = PetscFunctionListDuplicate(((PetscObject)matin)->qlist,&((PetscObject)mat)->qlist);CHKERRQ(ierr); 3305 *newmat = mat; 3306 PetscFunctionReturn(0); 3307 } 3308 3309 3310 3311 #undef __FUNCT__ 3312 #define __FUNCT__ "MatLoad_MPIAIJ" 3313 PetscErrorCode MatLoad_MPIAIJ(Mat newMat, PetscViewer viewer) 3314 { 3315 PetscScalar *vals,*svals; 3316 MPI_Comm comm; 3317 PetscErrorCode ierr; 3318 PetscMPIInt rank,size,tag = ((PetscObject)viewer)->tag; 3319 PetscInt i,nz,j,rstart,rend,mmax,maxnz = 0,grows,gcols; 3320 PetscInt header[4],*rowlengths = 0,M,N,m,*cols; 3321 PetscInt *ourlens = NULL,*procsnz = NULL,*offlens = NULL,jj,*mycols,*smycols; 3322 PetscInt cend,cstart,n,*rowners,sizesset=1; 3323 int fd; 3324 PetscInt bs = 1; 3325 3326 PetscFunctionBegin; 3327 ierr = PetscObjectGetComm((PetscObject)viewer,&comm);CHKERRQ(ierr); 3328 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 3329 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 3330 if (!rank) { 3331 ierr = PetscViewerBinaryGetDescriptor(viewer,&fd);CHKERRQ(ierr); 3332 ierr = PetscBinaryRead(fd,(char*)header,4,PETSC_INT);CHKERRQ(ierr); 3333 if (header[0] != MAT_FILE_CLASSID) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED,"not matrix object"); 3334 } 3335 3336 ierr = PetscOptionsBegin(comm,NULL,"Options for loading SEQAIJ matrix","Mat");CHKERRQ(ierr); 3337 ierr = PetscOptionsInt("-matload_block_size","Set the blocksize used to store the matrix","MatLoad",bs,&bs,NULL);CHKERRQ(ierr); 3338 ierr = PetscOptionsEnd();CHKERRQ(ierr); 3339 3340 if (newMat->rmap->n < 0 && newMat->rmap->N < 0 && newMat->cmap->n < 0 && newMat->cmap->N < 0) sizesset = 0; 3341 3342 ierr = MPI_Bcast(header+1,3,MPIU_INT,0,comm);CHKERRQ(ierr); 3343 M = header[1]; N = header[2]; 3344 /* If global rows/cols are set to PETSC_DECIDE, set it to the sizes given in the file */ 3345 if (sizesset && newMat->rmap->N < 0) newMat->rmap->N = M; 3346 if (sizesset && newMat->cmap->N < 0) newMat->cmap->N = N; 3347 3348 /* If global sizes are set, check if they are consistent with that given in the file */ 3349 if (sizesset) { 3350 ierr = MatGetSize(newMat,&grows,&gcols);CHKERRQ(ierr); 3351 } 3352 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); 3353 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); 3354 3355 /* determine ownership of all (block) rows */ 3356 if (M%bs) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED, "Inconsistent # of rows (%d) and block size (%d)",M,bs); 3357 if (newMat->rmap->n < 0) m = bs*((M/bs)/size + (((M/bs) % size) > rank)); /* PETSC_DECIDE */ 3358 else m = newMat->rmap->n; /* Set by user */ 3359 3360 ierr = PetscMalloc1((size+1),&rowners);CHKERRQ(ierr); 3361 ierr = MPI_Allgather(&m,1,MPIU_INT,rowners+1,1,MPIU_INT,comm);CHKERRQ(ierr); 3362 3363 /* First process needs enough room for process with most rows */ 3364 if (!rank) { 3365 mmax = rowners[1]; 3366 for (i=2; i<=size; i++) { 3367 mmax = PetscMax(mmax, rowners[i]); 3368 } 3369 } else mmax = -1; /* unused, but compilers complain */ 3370 3371 rowners[0] = 0; 3372 for (i=2; i<=size; i++) { 3373 rowners[i] += rowners[i-1]; 3374 } 3375 rstart = rowners[rank]; 3376 rend = rowners[rank+1]; 3377 3378 /* distribute row lengths to all processors */ 3379 ierr = PetscMalloc2(m,&ourlens,m,&offlens);CHKERRQ(ierr); 3380 if (!rank) { 3381 ierr = PetscBinaryRead(fd,ourlens,m,PETSC_INT);CHKERRQ(ierr); 3382 ierr = PetscMalloc1(mmax,&rowlengths);CHKERRQ(ierr); 3383 ierr = PetscCalloc1(size,&procsnz);CHKERRQ(ierr); 3384 for (j=0; j<m; j++) { 3385 procsnz[0] += ourlens[j]; 3386 } 3387 for (i=1; i<size; i++) { 3388 ierr = PetscBinaryRead(fd,rowlengths,rowners[i+1]-rowners[i],PETSC_INT);CHKERRQ(ierr); 3389 /* calculate the number of nonzeros on each processor */ 3390 for (j=0; j<rowners[i+1]-rowners[i]; j++) { 3391 procsnz[i] += rowlengths[j]; 3392 } 3393 ierr = MPIULong_Send(rowlengths,rowners[i+1]-rowners[i],MPIU_INT,i,tag,comm);CHKERRQ(ierr); 3394 } 3395 ierr = PetscFree(rowlengths);CHKERRQ(ierr); 3396 } else { 3397 ierr = MPIULong_Recv(ourlens,m,MPIU_INT,0,tag,comm);CHKERRQ(ierr); 3398 } 3399 3400 if (!rank) { 3401 /* determine max buffer needed and allocate it */ 3402 maxnz = 0; 3403 for (i=0; i<size; i++) { 3404 maxnz = PetscMax(maxnz,procsnz[i]); 3405 } 3406 ierr = PetscMalloc1(maxnz,&cols);CHKERRQ(ierr); 3407 3408 /* read in my part of the matrix column indices */ 3409 nz = procsnz[0]; 3410 ierr = PetscMalloc1(nz,&mycols);CHKERRQ(ierr); 3411 ierr = PetscBinaryRead(fd,mycols,nz,PETSC_INT);CHKERRQ(ierr); 3412 3413 /* read in every one elses and ship off */ 3414 for (i=1; i<size; i++) { 3415 nz = procsnz[i]; 3416 ierr = PetscBinaryRead(fd,cols,nz,PETSC_INT);CHKERRQ(ierr); 3417 ierr = MPIULong_Send(cols,nz,MPIU_INT,i,tag,comm);CHKERRQ(ierr); 3418 } 3419 ierr = PetscFree(cols);CHKERRQ(ierr); 3420 } else { 3421 /* determine buffer space needed for message */ 3422 nz = 0; 3423 for (i=0; i<m; i++) { 3424 nz += ourlens[i]; 3425 } 3426 ierr = PetscMalloc1(nz,&mycols);CHKERRQ(ierr); 3427 3428 /* receive message of column indices*/ 3429 ierr = MPIULong_Recv(mycols,nz,MPIU_INT,0,tag,comm);CHKERRQ(ierr); 3430 } 3431 3432 /* determine column ownership if matrix is not square */ 3433 if (N != M) { 3434 if (newMat->cmap->n < 0) n = N/size + ((N % size) > rank); 3435 else n = newMat->cmap->n; 3436 ierr = MPI_Scan(&n,&cend,1,MPIU_INT,MPI_SUM,comm);CHKERRQ(ierr); 3437 cstart = cend - n; 3438 } else { 3439 cstart = rstart; 3440 cend = rend; 3441 n = cend - cstart; 3442 } 3443 3444 /* loop over local rows, determining number of off diagonal entries */ 3445 ierr = PetscMemzero(offlens,m*sizeof(PetscInt));CHKERRQ(ierr); 3446 jj = 0; 3447 for (i=0; i<m; i++) { 3448 for (j=0; j<ourlens[i]; j++) { 3449 if (mycols[jj] < cstart || mycols[jj] >= cend) offlens[i]++; 3450 jj++; 3451 } 3452 } 3453 3454 for (i=0; i<m; i++) { 3455 ourlens[i] -= offlens[i]; 3456 } 3457 if (!sizesset) { 3458 ierr = MatSetSizes(newMat,m,n,M,N);CHKERRQ(ierr); 3459 } 3460 3461 if (bs > 1) {ierr = MatSetBlockSize(newMat,bs);CHKERRQ(ierr);} 3462 3463 ierr = MatMPIAIJSetPreallocation(newMat,0,ourlens,0,offlens);CHKERRQ(ierr); 3464 3465 for (i=0; i<m; i++) { 3466 ourlens[i] += offlens[i]; 3467 } 3468 3469 if (!rank) { 3470 ierr = PetscMalloc1((maxnz+1),&vals);CHKERRQ(ierr); 3471 3472 /* read in my part of the matrix numerical values */ 3473 nz = procsnz[0]; 3474 ierr = PetscBinaryRead(fd,vals,nz,PETSC_SCALAR);CHKERRQ(ierr); 3475 3476 /* insert into matrix */ 3477 jj = rstart; 3478 smycols = mycols; 3479 svals = vals; 3480 for (i=0; i<m; i++) { 3481 ierr = MatSetValues_MPIAIJ(newMat,1,&jj,ourlens[i],smycols,svals,INSERT_VALUES);CHKERRQ(ierr); 3482 smycols += ourlens[i]; 3483 svals += ourlens[i]; 3484 jj++; 3485 } 3486 3487 /* read in other processors and ship out */ 3488 for (i=1; i<size; i++) { 3489 nz = procsnz[i]; 3490 ierr = PetscBinaryRead(fd,vals,nz,PETSC_SCALAR);CHKERRQ(ierr); 3491 ierr = MPIULong_Send(vals,nz,MPIU_SCALAR,i,((PetscObject)newMat)->tag,comm);CHKERRQ(ierr); 3492 } 3493 ierr = PetscFree(procsnz);CHKERRQ(ierr); 3494 } else { 3495 /* receive numeric values */ 3496 ierr = PetscMalloc1((nz+1),&vals);CHKERRQ(ierr); 3497 3498 /* receive message of values*/ 3499 ierr = MPIULong_Recv(vals,nz,MPIU_SCALAR,0,((PetscObject)newMat)->tag,comm);CHKERRQ(ierr); 3500 3501 /* insert into matrix */ 3502 jj = rstart; 3503 smycols = mycols; 3504 svals = vals; 3505 for (i=0; i<m; i++) { 3506 ierr = MatSetValues_MPIAIJ(newMat,1,&jj,ourlens[i],smycols,svals,INSERT_VALUES);CHKERRQ(ierr); 3507 smycols += ourlens[i]; 3508 svals += ourlens[i]; 3509 jj++; 3510 } 3511 } 3512 ierr = PetscFree2(ourlens,offlens);CHKERRQ(ierr); 3513 ierr = PetscFree(vals);CHKERRQ(ierr); 3514 ierr = PetscFree(mycols);CHKERRQ(ierr); 3515 ierr = PetscFree(rowners);CHKERRQ(ierr); 3516 ierr = MatAssemblyBegin(newMat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3517 ierr = MatAssemblyEnd(newMat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3518 PetscFunctionReturn(0); 3519 } 3520 3521 #undef __FUNCT__ 3522 #define __FUNCT__ "MatGetSubMatrix_MPIAIJ" 3523 PetscErrorCode MatGetSubMatrix_MPIAIJ(Mat mat,IS isrow,IS iscol,MatReuse call,Mat *newmat) 3524 { 3525 PetscErrorCode ierr; 3526 IS iscol_local; 3527 PetscInt csize; 3528 3529 PetscFunctionBegin; 3530 ierr = ISGetLocalSize(iscol,&csize);CHKERRQ(ierr); 3531 if (call == MAT_REUSE_MATRIX) { 3532 ierr = PetscObjectQuery((PetscObject)*newmat,"ISAllGather",(PetscObject*)&iscol_local);CHKERRQ(ierr); 3533 if (!iscol_local) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Submatrix passed in was not used before, cannot reuse"); 3534 } else { 3535 PetscInt cbs; 3536 ierr = ISGetBlockSize(iscol,&cbs);CHKERRQ(ierr); 3537 ierr = ISAllGather(iscol,&iscol_local);CHKERRQ(ierr); 3538 ierr = ISSetBlockSize(iscol_local,cbs);CHKERRQ(ierr); 3539 } 3540 ierr = MatGetSubMatrix_MPIAIJ_Private(mat,isrow,iscol_local,csize,call,newmat);CHKERRQ(ierr); 3541 if (call == MAT_INITIAL_MATRIX) { 3542 ierr = PetscObjectCompose((PetscObject)*newmat,"ISAllGather",(PetscObject)iscol_local);CHKERRQ(ierr); 3543 ierr = ISDestroy(&iscol_local);CHKERRQ(ierr); 3544 } 3545 PetscFunctionReturn(0); 3546 } 3547 3548 extern PetscErrorCode MatGetSubMatrices_MPIAIJ_Local(Mat,PetscInt,const IS[],const IS[],MatReuse,PetscBool*,Mat*); 3549 #undef __FUNCT__ 3550 #define __FUNCT__ "MatGetSubMatrix_MPIAIJ_Private" 3551 /* 3552 Not great since it makes two copies of the submatrix, first an SeqAIJ 3553 in local and then by concatenating the local matrices the end result. 3554 Writing it directly would be much like MatGetSubMatrices_MPIAIJ() 3555 3556 Note: This requires a sequential iscol with all indices. 3557 */ 3558 PetscErrorCode MatGetSubMatrix_MPIAIJ_Private(Mat mat,IS isrow,IS iscol,PetscInt csize,MatReuse call,Mat *newmat) 3559 { 3560 PetscErrorCode ierr; 3561 PetscMPIInt rank,size; 3562 PetscInt i,m,n,rstart,row,rend,nz,*cwork,j,bs,cbs; 3563 PetscInt *ii,*jj,nlocal,*dlens,*olens,dlen,olen,jend,mglobal,ncol; 3564 PetscBool allcolumns, colflag; 3565 Mat M,Mreuse; 3566 MatScalar *vwork,*aa; 3567 MPI_Comm comm; 3568 Mat_SeqAIJ *aij; 3569 3570 PetscFunctionBegin; 3571 ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 3572 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 3573 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 3574 3575 ierr = ISIdentity(iscol,&colflag);CHKERRQ(ierr); 3576 ierr = ISGetLocalSize(iscol,&ncol);CHKERRQ(ierr); 3577 if (colflag && ncol == mat->cmap->N) { 3578 allcolumns = PETSC_TRUE; 3579 } else { 3580 allcolumns = PETSC_FALSE; 3581 } 3582 if (call == MAT_REUSE_MATRIX) { 3583 ierr = PetscObjectQuery((PetscObject)*newmat,"SubMatrix",(PetscObject*)&Mreuse);CHKERRQ(ierr); 3584 if (!Mreuse) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Submatrix passed in was not used before, cannot reuse"); 3585 ierr = MatGetSubMatrices_MPIAIJ_Local(mat,1,&isrow,&iscol,MAT_REUSE_MATRIX,&allcolumns,&Mreuse);CHKERRQ(ierr); 3586 } else { 3587 ierr = MatGetSubMatrices_MPIAIJ_Local(mat,1,&isrow,&iscol,MAT_INITIAL_MATRIX,&allcolumns,&Mreuse);CHKERRQ(ierr); 3588 } 3589 3590 /* 3591 m - number of local rows 3592 n - number of columns (same on all processors) 3593 rstart - first row in new global matrix generated 3594 */ 3595 ierr = MatGetSize(Mreuse,&m,&n);CHKERRQ(ierr); 3596 ierr = MatGetBlockSizes(Mreuse,&bs,&cbs);CHKERRQ(ierr); 3597 if (call == MAT_INITIAL_MATRIX) { 3598 aij = (Mat_SeqAIJ*)(Mreuse)->data; 3599 ii = aij->i; 3600 jj = aij->j; 3601 3602 /* 3603 Determine the number of non-zeros in the diagonal and off-diagonal 3604 portions of the matrix in order to do correct preallocation 3605 */ 3606 3607 /* first get start and end of "diagonal" columns */ 3608 if (csize == PETSC_DECIDE) { 3609 ierr = ISGetSize(isrow,&mglobal);CHKERRQ(ierr); 3610 if (mglobal == n) { /* square matrix */ 3611 nlocal = m; 3612 } else { 3613 nlocal = n/size + ((n % size) > rank); 3614 } 3615 } else { 3616 nlocal = csize; 3617 } 3618 ierr = MPI_Scan(&nlocal,&rend,1,MPIU_INT,MPI_SUM,comm);CHKERRQ(ierr); 3619 rstart = rend - nlocal; 3620 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); 3621 3622 /* next, compute all the lengths */ 3623 ierr = PetscMalloc1((2*m+1),&dlens);CHKERRQ(ierr); 3624 olens = dlens + m; 3625 for (i=0; i<m; i++) { 3626 jend = ii[i+1] - ii[i]; 3627 olen = 0; 3628 dlen = 0; 3629 for (j=0; j<jend; j++) { 3630 if (*jj < rstart || *jj >= rend) olen++; 3631 else dlen++; 3632 jj++; 3633 } 3634 olens[i] = olen; 3635 dlens[i] = dlen; 3636 } 3637 ierr = MatCreate(comm,&M);CHKERRQ(ierr); 3638 ierr = MatSetSizes(M,m,nlocal,PETSC_DECIDE,n);CHKERRQ(ierr); 3639 ierr = MatSetBlockSizes(M,bs,cbs);CHKERRQ(ierr); 3640 ierr = MatSetType(M,((PetscObject)mat)->type_name);CHKERRQ(ierr); 3641 ierr = MatMPIAIJSetPreallocation(M,0,dlens,0,olens);CHKERRQ(ierr); 3642 ierr = PetscFree(dlens);CHKERRQ(ierr); 3643 } else { 3644 PetscInt ml,nl; 3645 3646 M = *newmat; 3647 ierr = MatGetLocalSize(M,&ml,&nl);CHKERRQ(ierr); 3648 if (ml != m) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Previous matrix must be same size/layout as request"); 3649 ierr = MatZeroEntries(M);CHKERRQ(ierr); 3650 /* 3651 The next two lines are needed so we may call MatSetValues_MPIAIJ() below directly, 3652 rather than the slower MatSetValues(). 3653 */ 3654 M->was_assembled = PETSC_TRUE; 3655 M->assembled = PETSC_FALSE; 3656 } 3657 ierr = MatGetOwnershipRange(M,&rstart,&rend);CHKERRQ(ierr); 3658 aij = (Mat_SeqAIJ*)(Mreuse)->data; 3659 ii = aij->i; 3660 jj = aij->j; 3661 aa = aij->a; 3662 for (i=0; i<m; i++) { 3663 row = rstart + i; 3664 nz = ii[i+1] - ii[i]; 3665 cwork = jj; jj += nz; 3666 vwork = aa; aa += nz; 3667 ierr = MatSetValues_MPIAIJ(M,1,&row,nz,cwork,vwork,INSERT_VALUES);CHKERRQ(ierr); 3668 } 3669 3670 ierr = MatAssemblyBegin(M,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3671 ierr = MatAssemblyEnd(M,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3672 *newmat = M; 3673 3674 /* save submatrix used in processor for next request */ 3675 if (call == MAT_INITIAL_MATRIX) { 3676 ierr = PetscObjectCompose((PetscObject)M,"SubMatrix",(PetscObject)Mreuse);CHKERRQ(ierr); 3677 ierr = MatDestroy(&Mreuse);CHKERRQ(ierr); 3678 } 3679 PetscFunctionReturn(0); 3680 } 3681 3682 #undef __FUNCT__ 3683 #define __FUNCT__ "MatMPIAIJSetPreallocationCSR_MPIAIJ" 3684 PetscErrorCode MatMPIAIJSetPreallocationCSR_MPIAIJ(Mat B,const PetscInt Ii[],const PetscInt J[],const PetscScalar v[]) 3685 { 3686 PetscInt m,cstart, cend,j,nnz,i,d; 3687 PetscInt *d_nnz,*o_nnz,nnz_max = 0,rstart,ii; 3688 const PetscInt *JJ; 3689 PetscScalar *values; 3690 PetscErrorCode ierr; 3691 3692 PetscFunctionBegin; 3693 if (Ii[0]) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Ii[0] must be 0 it is %D",Ii[0]); 3694 3695 ierr = PetscLayoutSetUp(B->rmap);CHKERRQ(ierr); 3696 ierr = PetscLayoutSetUp(B->cmap);CHKERRQ(ierr); 3697 m = B->rmap->n; 3698 cstart = B->cmap->rstart; 3699 cend = B->cmap->rend; 3700 rstart = B->rmap->rstart; 3701 3702 ierr = PetscMalloc2(m,&d_nnz,m,&o_nnz);CHKERRQ(ierr); 3703 3704 #if defined(PETSC_USE_DEBUGGING) 3705 for (i=0; i<m; i++) { 3706 nnz = Ii[i+1]- Ii[i]; 3707 JJ = J + Ii[i]; 3708 if (nnz < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Local row %D has a negative %D number of columns",i,nnz); 3709 if (nnz && (JJ[0] < 0)) SETERRRQ1(PETSC_ERR_ARG_WRONGSTATE,"Row %D starts with negative column index",i,j); 3710 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); 3711 } 3712 #endif 3713 3714 for (i=0; i<m; i++) { 3715 nnz = Ii[i+1]- Ii[i]; 3716 JJ = J + Ii[i]; 3717 nnz_max = PetscMax(nnz_max,nnz); 3718 d = 0; 3719 for (j=0; j<nnz; j++) { 3720 if (cstart <= JJ[j] && JJ[j] < cend) d++; 3721 } 3722 d_nnz[i] = d; 3723 o_nnz[i] = nnz - d; 3724 } 3725 ierr = MatMPIAIJSetPreallocation(B,0,d_nnz,0,o_nnz);CHKERRQ(ierr); 3726 ierr = PetscFree2(d_nnz,o_nnz);CHKERRQ(ierr); 3727 3728 if (v) values = (PetscScalar*)v; 3729 else { 3730 ierr = PetscCalloc1((nnz_max+1),&values);CHKERRQ(ierr); 3731 } 3732 3733 for (i=0; i<m; i++) { 3734 ii = i + rstart; 3735 nnz = Ii[i+1]- Ii[i]; 3736 ierr = MatSetValues_MPIAIJ(B,1,&ii,nnz,J+Ii[i],values+(v ? Ii[i] : 0),INSERT_VALUES);CHKERRQ(ierr); 3737 } 3738 ierr = MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3739 ierr = MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3740 3741 if (!v) { 3742 ierr = PetscFree(values);CHKERRQ(ierr); 3743 } 3744 ierr = MatSetOption(B,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE);CHKERRQ(ierr); 3745 PetscFunctionReturn(0); 3746 } 3747 3748 #undef __FUNCT__ 3749 #define __FUNCT__ "MatMPIAIJSetPreallocationCSR" 3750 /*@ 3751 MatMPIAIJSetPreallocationCSR - Allocates memory for a sparse parallel matrix in AIJ format 3752 (the default parallel PETSc format). 3753 3754 Collective on MPI_Comm 3755 3756 Input Parameters: 3757 + B - the matrix 3758 . i - the indices into j for the start of each local row (starts with zero) 3759 . j - the column indices for each local row (starts with zero) 3760 - v - optional values in the matrix 3761 3762 Level: developer 3763 3764 Notes: 3765 The i, j, and a arrays ARE copied by this routine into the internal format used by PETSc; 3766 thus you CANNOT change the matrix entries by changing the values of a[] after you have 3767 called this routine. Use MatCreateMPIAIJWithSplitArrays() to avoid needing to copy the arrays. 3768 3769 The i and j indices are 0 based, and i indices are indices corresponding to the local j array. 3770 3771 The format which is used for the sparse matrix input, is equivalent to a 3772 row-major ordering.. i.e for the following matrix, the input data expected is 3773 as shown: 3774 3775 1 0 0 3776 2 0 3 P0 3777 ------- 3778 4 5 6 P1 3779 3780 Process0 [P0]: rows_owned=[0,1] 3781 i = {0,1,3} [size = nrow+1 = 2+1] 3782 j = {0,0,2} [size = nz = 6] 3783 v = {1,2,3} [size = nz = 6] 3784 3785 Process1 [P1]: rows_owned=[2] 3786 i = {0,3} [size = nrow+1 = 1+1] 3787 j = {0,1,2} [size = nz = 6] 3788 v = {4,5,6} [size = nz = 6] 3789 3790 .keywords: matrix, aij, compressed row, sparse, parallel 3791 3792 .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatCreateAIJ(), MPIAIJ, 3793 MatCreateSeqAIJWithArrays(), MatCreateMPIAIJWithSplitArrays() 3794 @*/ 3795 PetscErrorCode MatMPIAIJSetPreallocationCSR(Mat B,const PetscInt i[],const PetscInt j[], const PetscScalar v[]) 3796 { 3797 PetscErrorCode ierr; 3798 3799 PetscFunctionBegin; 3800 ierr = PetscTryMethod(B,"MatMPIAIJSetPreallocationCSR_C",(Mat,const PetscInt[],const PetscInt[],const PetscScalar[]),(B,i,j,v));CHKERRQ(ierr); 3801 PetscFunctionReturn(0); 3802 } 3803 3804 #undef __FUNCT__ 3805 #define __FUNCT__ "MatMPIAIJSetPreallocation" 3806 /*@C 3807 MatMPIAIJSetPreallocation - Preallocates memory for a sparse parallel matrix in AIJ format 3808 (the default parallel PETSc format). For good matrix assembly performance 3809 the user should preallocate the matrix storage by setting the parameters 3810 d_nz (or d_nnz) and o_nz (or o_nnz). By setting these parameters accurately, 3811 performance can be increased by more than a factor of 50. 3812 3813 Collective on MPI_Comm 3814 3815 Input Parameters: 3816 + B - the matrix 3817 . d_nz - number of nonzeros per row in DIAGONAL portion of local submatrix 3818 (same value is used for all local rows) 3819 . d_nnz - array containing the number of nonzeros in the various rows of the 3820 DIAGONAL portion of the local submatrix (possibly different for each row) 3821 or NULL, if d_nz is used to specify the nonzero structure. 3822 The size of this array is equal to the number of local rows, i.e 'm'. 3823 For matrices that will be factored, you must leave room for (and set) 3824 the diagonal entry even if it is zero. 3825 . o_nz - number of nonzeros per row in the OFF-DIAGONAL portion of local 3826 submatrix (same value is used for all local rows). 3827 - o_nnz - array containing the number of nonzeros in the various rows of the 3828 OFF-DIAGONAL portion of the local submatrix (possibly different for 3829 each row) or NULL, if o_nz is used to specify the nonzero 3830 structure. The size of this array is equal to the number 3831 of local rows, i.e 'm'. 3832 3833 If the *_nnz parameter is given then the *_nz parameter is ignored 3834 3835 The AIJ format (also called the Yale sparse matrix format or 3836 compressed row storage (CSR)), is fully compatible with standard Fortran 77 3837 storage. The stored row and column indices begin with zero. 3838 See Users-Manual: ch_mat for details. 3839 3840 The parallel matrix is partitioned such that the first m0 rows belong to 3841 process 0, the next m1 rows belong to process 1, the next m2 rows belong 3842 to process 2 etc.. where m0,m1,m2... are the input parameter 'm'. 3843 3844 The DIAGONAL portion of the local submatrix of a processor can be defined 3845 as the submatrix which is obtained by extraction the part corresponding to 3846 the rows r1-r2 and columns c1-c2 of the global matrix, where r1 is the 3847 first row that belongs to the processor, r2 is the last row belonging to 3848 the this processor, and c1-c2 is range of indices of the local part of a 3849 vector suitable for applying the matrix to. This is an mxn matrix. In the 3850 common case of a square matrix, the row and column ranges are the same and 3851 the DIAGONAL part is also square. The remaining portion of the local 3852 submatrix (mxN) constitute the OFF-DIAGONAL portion. 3853 3854 If o_nnz, d_nnz are specified, then o_nz, and d_nz are ignored. 3855 3856 You can call MatGetInfo() to get information on how effective the preallocation was; 3857 for example the fields mallocs,nz_allocated,nz_used,nz_unneeded; 3858 You can also run with the option -info and look for messages with the string 3859 malloc in them to see if additional memory allocation was needed. 3860 3861 Example usage: 3862 3863 Consider the following 8x8 matrix with 34 non-zero values, that is 3864 assembled across 3 processors. Lets assume that proc0 owns 3 rows, 3865 proc1 owns 3 rows, proc2 owns 2 rows. This division can be shown 3866 as follows: 3867 3868 .vb 3869 1 2 0 | 0 3 0 | 0 4 3870 Proc0 0 5 6 | 7 0 0 | 8 0 3871 9 0 10 | 11 0 0 | 12 0 3872 ------------------------------------- 3873 13 0 14 | 15 16 17 | 0 0 3874 Proc1 0 18 0 | 19 20 21 | 0 0 3875 0 0 0 | 22 23 0 | 24 0 3876 ------------------------------------- 3877 Proc2 25 26 27 | 0 0 28 | 29 0 3878 30 0 0 | 31 32 33 | 0 34 3879 .ve 3880 3881 This can be represented as a collection of submatrices as: 3882 3883 .vb 3884 A B C 3885 D E F 3886 G H I 3887 .ve 3888 3889 Where the submatrices A,B,C are owned by proc0, D,E,F are 3890 owned by proc1, G,H,I are owned by proc2. 3891 3892 The 'm' parameters for proc0,proc1,proc2 are 3,3,2 respectively. 3893 The 'n' parameters for proc0,proc1,proc2 are 3,3,2 respectively. 3894 The 'M','N' parameters are 8,8, and have the same values on all procs. 3895 3896 The DIAGONAL submatrices corresponding to proc0,proc1,proc2 are 3897 submatrices [A], [E], [I] respectively. The OFF-DIAGONAL submatrices 3898 corresponding to proc0,proc1,proc2 are [BC], [DF], [GH] respectively. 3899 Internally, each processor stores the DIAGONAL part, and the OFF-DIAGONAL 3900 part as SeqAIJ matrices. for eg: proc1 will store [E] as a SeqAIJ 3901 matrix, ans [DF] as another SeqAIJ matrix. 3902 3903 When d_nz, o_nz parameters are specified, d_nz storage elements are 3904 allocated for every row of the local diagonal submatrix, and o_nz 3905 storage locations are allocated for every row of the OFF-DIAGONAL submat. 3906 One way to choose d_nz and o_nz is to use the max nonzerors per local 3907 rows for each of the local DIAGONAL, and the OFF-DIAGONAL submatrices. 3908 In this case, the values of d_nz,o_nz are: 3909 .vb 3910 proc0 : dnz = 2, o_nz = 2 3911 proc1 : dnz = 3, o_nz = 2 3912 proc2 : dnz = 1, o_nz = 4 3913 .ve 3914 We are allocating m*(d_nz+o_nz) storage locations for every proc. This 3915 translates to 3*(2+2)=12 for proc0, 3*(3+2)=15 for proc1, 2*(1+4)=10 3916 for proc3. i.e we are using 12+15+10=37 storage locations to store 3917 34 values. 3918 3919 When d_nnz, o_nnz parameters are specified, the storage is specified 3920 for every row, coresponding to both DIAGONAL and OFF-DIAGONAL submatrices. 3921 In the above case the values for d_nnz,o_nnz are: 3922 .vb 3923 proc0: d_nnz = [2,2,2] and o_nnz = [2,2,2] 3924 proc1: d_nnz = [3,3,2] and o_nnz = [2,1,1] 3925 proc2: d_nnz = [1,1] and o_nnz = [4,4] 3926 .ve 3927 Here the space allocated is sum of all the above values i.e 34, and 3928 hence pre-allocation is perfect. 3929 3930 Level: intermediate 3931 3932 .keywords: matrix, aij, compressed row, sparse, parallel 3933 3934 .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatCreateAIJ(), MatMPIAIJSetPreallocationCSR(), 3935 MPIAIJ, MatGetInfo(), PetscSplitOwnership() 3936 @*/ 3937 PetscErrorCode MatMPIAIJSetPreallocation(Mat B,PetscInt d_nz,const PetscInt d_nnz[],PetscInt o_nz,const PetscInt o_nnz[]) 3938 { 3939 PetscErrorCode ierr; 3940 3941 PetscFunctionBegin; 3942 PetscValidHeaderSpecific(B,MAT_CLASSID,1); 3943 PetscValidType(B,1); 3944 ierr = PetscTryMethod(B,"MatMPIAIJSetPreallocation_C",(Mat,PetscInt,const PetscInt[],PetscInt,const PetscInt[]),(B,d_nz,d_nnz,o_nz,o_nnz));CHKERRQ(ierr); 3945 PetscFunctionReturn(0); 3946 } 3947 3948 #undef __FUNCT__ 3949 #define __FUNCT__ "MatCreateMPIAIJWithArrays" 3950 /*@ 3951 MatCreateMPIAIJWithArrays - creates a MPI AIJ matrix using arrays that contain in standard 3952 CSR format the local rows. 3953 3954 Collective on MPI_Comm 3955 3956 Input Parameters: 3957 + comm - MPI communicator 3958 . m - number of local rows (Cannot be PETSC_DECIDE) 3959 . n - This value should be the same as the local size used in creating the 3960 x vector for the matrix-vector product y = Ax. (or PETSC_DECIDE to have 3961 calculated if N is given) For square matrices n is almost always m. 3962 . M - number of global rows (or PETSC_DETERMINE to have calculated if m is given) 3963 . N - number of global columns (or PETSC_DETERMINE to have calculated if n is given) 3964 . i - row indices 3965 . j - column indices 3966 - a - matrix values 3967 3968 Output Parameter: 3969 . mat - the matrix 3970 3971 Level: intermediate 3972 3973 Notes: 3974 The i, j, and a arrays ARE copied by this routine into the internal format used by PETSc; 3975 thus you CANNOT change the matrix entries by changing the values of a[] after you have 3976 called this routine. Use MatCreateMPIAIJWithSplitArrays() to avoid needing to copy the arrays. 3977 3978 The i and j indices are 0 based, and i indices are indices corresponding to the local j array. 3979 3980 The format which is used for the sparse matrix input, is equivalent to a 3981 row-major ordering.. i.e for the following matrix, the input data expected is 3982 as shown: 3983 3984 1 0 0 3985 2 0 3 P0 3986 ------- 3987 4 5 6 P1 3988 3989 Process0 [P0]: rows_owned=[0,1] 3990 i = {0,1,3} [size = nrow+1 = 2+1] 3991 j = {0,0,2} [size = nz = 6] 3992 v = {1,2,3} [size = nz = 6] 3993 3994 Process1 [P1]: rows_owned=[2] 3995 i = {0,3} [size = nrow+1 = 1+1] 3996 j = {0,1,2} [size = nz = 6] 3997 v = {4,5,6} [size = nz = 6] 3998 3999 .keywords: matrix, aij, compressed row, sparse, parallel 4000 4001 .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatMPIAIJSetPreallocationCSR(), 4002 MPIAIJ, MatCreateAIJ(), MatCreateMPIAIJWithSplitArrays() 4003 @*/ 4004 PetscErrorCode MatCreateMPIAIJWithArrays(MPI_Comm comm,PetscInt m,PetscInt n,PetscInt M,PetscInt N,const PetscInt i[],const PetscInt j[],const PetscScalar a[],Mat *mat) 4005 { 4006 PetscErrorCode ierr; 4007 4008 PetscFunctionBegin; 4009 if (i[0]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"i (row indices) must start with 0"); 4010 if (m < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"local number of rows (m) cannot be PETSC_DECIDE, or negative"); 4011 ierr = MatCreate(comm,mat);CHKERRQ(ierr); 4012 ierr = MatSetSizes(*mat,m,n,M,N);CHKERRQ(ierr); 4013 /* ierr = MatSetBlockSizes(M,bs,cbs);CHKERRQ(ierr); */ 4014 ierr = MatSetType(*mat,MATMPIAIJ);CHKERRQ(ierr); 4015 ierr = MatMPIAIJSetPreallocationCSR(*mat,i,j,a);CHKERRQ(ierr); 4016 PetscFunctionReturn(0); 4017 } 4018 4019 #undef __FUNCT__ 4020 #define __FUNCT__ "MatCreateAIJ" 4021 /*@C 4022 MatCreateAIJ - Creates a sparse parallel matrix in AIJ format 4023 (the default parallel PETSc format). For good matrix assembly performance 4024 the user should preallocate the matrix storage by setting the parameters 4025 d_nz (or d_nnz) and o_nz (or o_nnz). By setting these parameters accurately, 4026 performance can be increased by more than a factor of 50. 4027 4028 Collective on MPI_Comm 4029 4030 Input Parameters: 4031 + comm - MPI communicator 4032 . m - number of local rows (or PETSC_DECIDE to have calculated if M is given) 4033 This value should be the same as the local size used in creating the 4034 y vector for the matrix-vector product y = Ax. 4035 . n - This value should be the same as the local size used in creating the 4036 x vector for the matrix-vector product y = Ax. (or PETSC_DECIDE to have 4037 calculated if N is given) For square matrices n is almost always m. 4038 . M - number of global rows (or PETSC_DETERMINE to have calculated if m is given) 4039 . N - number of global columns (or PETSC_DETERMINE to have calculated if n is given) 4040 . d_nz - number of nonzeros per row in DIAGONAL portion of local submatrix 4041 (same value is used for all local rows) 4042 . d_nnz - array containing the number of nonzeros in the various rows of the 4043 DIAGONAL portion of the local submatrix (possibly different for each row) 4044 or NULL, if d_nz is used to specify the nonzero structure. 4045 The size of this array is equal to the number of local rows, i.e 'm'. 4046 . o_nz - number of nonzeros per row in the OFF-DIAGONAL portion of local 4047 submatrix (same value is used for all local rows). 4048 - o_nnz - array containing the number of nonzeros in the various rows of the 4049 OFF-DIAGONAL portion of the local submatrix (possibly different for 4050 each row) or NULL, if o_nz is used to specify the nonzero 4051 structure. The size of this array is equal to the number 4052 of local rows, i.e 'm'. 4053 4054 Output Parameter: 4055 . A - the matrix 4056 4057 It is recommended that one use the MatCreate(), MatSetType() and/or MatSetFromOptions(), 4058 MatXXXXSetPreallocation() paradgm instead of this routine directly. 4059 [MatXXXXSetPreallocation() is, for example, MatSeqAIJSetPreallocation] 4060 4061 Notes: 4062 If the *_nnz parameter is given then the *_nz parameter is ignored 4063 4064 m,n,M,N parameters specify the size of the matrix, and its partitioning across 4065 processors, while d_nz,d_nnz,o_nz,o_nnz parameters specify the approximate 4066 storage requirements for this matrix. 4067 4068 If PETSC_DECIDE or PETSC_DETERMINE is used for a particular argument on one 4069 processor than it must be used on all processors that share the object for 4070 that argument. 4071 4072 The user MUST specify either the local or global matrix dimensions 4073 (possibly both). 4074 4075 The parallel matrix is partitioned across processors such that the 4076 first m0 rows belong to process 0, the next m1 rows belong to 4077 process 1, the next m2 rows belong to process 2 etc.. where 4078 m0,m1,m2,.. are the input parameter 'm'. i.e each processor stores 4079 values corresponding to [m x N] submatrix. 4080 4081 The columns are logically partitioned with the n0 columns belonging 4082 to 0th partition, the next n1 columns belonging to the next 4083 partition etc.. where n0,n1,n2... are the input parameter 'n'. 4084 4085 The DIAGONAL portion of the local submatrix on any given processor 4086 is the submatrix corresponding to the rows and columns m,n 4087 corresponding to the given processor. i.e diagonal matrix on 4088 process 0 is [m0 x n0], diagonal matrix on process 1 is [m1 x n1] 4089 etc. The remaining portion of the local submatrix [m x (N-n)] 4090 constitute the OFF-DIAGONAL portion. The example below better 4091 illustrates this concept. 4092 4093 For a square global matrix we define each processor's diagonal portion 4094 to be its local rows and the corresponding columns (a square submatrix); 4095 each processor's off-diagonal portion encompasses the remainder of the 4096 local matrix (a rectangular submatrix). 4097 4098 If o_nnz, d_nnz are specified, then o_nz, and d_nz are ignored. 4099 4100 When calling this routine with a single process communicator, a matrix of 4101 type SEQAIJ is returned. If a matrix of type MPIAIJ is desired for this 4102 type of communicator, use the construction mechanism: 4103 MatCreate(...,&A); MatSetType(A,MATMPIAIJ); MatSetSizes(A, m,n,M,N); MatMPIAIJSetPreallocation(A,...); 4104 4105 By default, this format uses inodes (identical nodes) when possible. 4106 We search for consecutive rows with the same nonzero structure, thereby 4107 reusing matrix information to achieve increased efficiency. 4108 4109 Options Database Keys: 4110 + -mat_no_inode - Do not use inodes 4111 . -mat_inode_limit <limit> - Sets inode limit (max limit=5) 4112 - -mat_aij_oneindex - Internally use indexing starting at 1 4113 rather than 0. Note that when calling MatSetValues(), 4114 the user still MUST index entries starting at 0! 4115 4116 4117 Example usage: 4118 4119 Consider the following 8x8 matrix with 34 non-zero values, that is 4120 assembled across 3 processors. Lets assume that proc0 owns 3 rows, 4121 proc1 owns 3 rows, proc2 owns 2 rows. This division can be shown 4122 as follows: 4123 4124 .vb 4125 1 2 0 | 0 3 0 | 0 4 4126 Proc0 0 5 6 | 7 0 0 | 8 0 4127 9 0 10 | 11 0 0 | 12 0 4128 ------------------------------------- 4129 13 0 14 | 15 16 17 | 0 0 4130 Proc1 0 18 0 | 19 20 21 | 0 0 4131 0 0 0 | 22 23 0 | 24 0 4132 ------------------------------------- 4133 Proc2 25 26 27 | 0 0 28 | 29 0 4134 30 0 0 | 31 32 33 | 0 34 4135 .ve 4136 4137 This can be represented as a collection of submatrices as: 4138 4139 .vb 4140 A B C 4141 D E F 4142 G H I 4143 .ve 4144 4145 Where the submatrices A,B,C are owned by proc0, D,E,F are 4146 owned by proc1, G,H,I are owned by proc2. 4147 4148 The 'm' parameters for proc0,proc1,proc2 are 3,3,2 respectively. 4149 The 'n' parameters for proc0,proc1,proc2 are 3,3,2 respectively. 4150 The 'M','N' parameters are 8,8, and have the same values on all procs. 4151 4152 The DIAGONAL submatrices corresponding to proc0,proc1,proc2 are 4153 submatrices [A], [E], [I] respectively. The OFF-DIAGONAL submatrices 4154 corresponding to proc0,proc1,proc2 are [BC], [DF], [GH] respectively. 4155 Internally, each processor stores the DIAGONAL part, and the OFF-DIAGONAL 4156 part as SeqAIJ matrices. for eg: proc1 will store [E] as a SeqAIJ 4157 matrix, ans [DF] as another SeqAIJ matrix. 4158 4159 When d_nz, o_nz parameters are specified, d_nz storage elements are 4160 allocated for every row of the local diagonal submatrix, and o_nz 4161 storage locations are allocated for every row of the OFF-DIAGONAL submat. 4162 One way to choose d_nz and o_nz is to use the max nonzerors per local 4163 rows for each of the local DIAGONAL, and the OFF-DIAGONAL submatrices. 4164 In this case, the values of d_nz,o_nz are: 4165 .vb 4166 proc0 : dnz = 2, o_nz = 2 4167 proc1 : dnz = 3, o_nz = 2 4168 proc2 : dnz = 1, o_nz = 4 4169 .ve 4170 We are allocating m*(d_nz+o_nz) storage locations for every proc. This 4171 translates to 3*(2+2)=12 for proc0, 3*(3+2)=15 for proc1, 2*(1+4)=10 4172 for proc3. i.e we are using 12+15+10=37 storage locations to store 4173 34 values. 4174 4175 When d_nnz, o_nnz parameters are specified, the storage is specified 4176 for every row, coresponding to both DIAGONAL and OFF-DIAGONAL submatrices. 4177 In the above case the values for d_nnz,o_nnz are: 4178 .vb 4179 proc0: d_nnz = [2,2,2] and o_nnz = [2,2,2] 4180 proc1: d_nnz = [3,3,2] and o_nnz = [2,1,1] 4181 proc2: d_nnz = [1,1] and o_nnz = [4,4] 4182 .ve 4183 Here the space allocated is sum of all the above values i.e 34, and 4184 hence pre-allocation is perfect. 4185 4186 Level: intermediate 4187 4188 .keywords: matrix, aij, compressed row, sparse, parallel 4189 4190 .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatMPIAIJSetPreallocationCSR(), 4191 MPIAIJ, MatCreateMPIAIJWithArrays() 4192 @*/ 4193 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) 4194 { 4195 PetscErrorCode ierr; 4196 PetscMPIInt size; 4197 4198 PetscFunctionBegin; 4199 ierr = MatCreate(comm,A);CHKERRQ(ierr); 4200 ierr = MatSetSizes(*A,m,n,M,N);CHKERRQ(ierr); 4201 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 4202 if (size > 1) { 4203 ierr = MatSetType(*A,MATMPIAIJ);CHKERRQ(ierr); 4204 ierr = MatMPIAIJSetPreallocation(*A,d_nz,d_nnz,o_nz,o_nnz);CHKERRQ(ierr); 4205 } else { 4206 ierr = MatSetType(*A,MATSEQAIJ);CHKERRQ(ierr); 4207 ierr = MatSeqAIJSetPreallocation(*A,d_nz,d_nnz);CHKERRQ(ierr); 4208 } 4209 PetscFunctionReturn(0); 4210 } 4211 4212 #undef __FUNCT__ 4213 #define __FUNCT__ "MatMPIAIJGetSeqAIJ" 4214 PetscErrorCode MatMPIAIJGetSeqAIJ(Mat A,Mat *Ad,Mat *Ao,const PetscInt *colmap[]) 4215 { 4216 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 4217 4218 PetscFunctionBegin; 4219 if (Ad) *Ad = a->A; 4220 if (Ao) *Ao = a->B; 4221 if (colmap) *colmap = a->garray; 4222 PetscFunctionReturn(0); 4223 } 4224 4225 #undef __FUNCT__ 4226 #define __FUNCT__ "MatSetColoring_MPIAIJ" 4227 PetscErrorCode MatSetColoring_MPIAIJ(Mat A,ISColoring coloring) 4228 { 4229 PetscErrorCode ierr; 4230 PetscInt i; 4231 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 4232 4233 PetscFunctionBegin; 4234 if (coloring->ctype == IS_COLORING_GLOBAL) { 4235 ISColoringValue *allcolors,*colors; 4236 ISColoring ocoloring; 4237 4238 /* set coloring for diagonal portion */ 4239 ierr = MatSetColoring_SeqAIJ(a->A,coloring);CHKERRQ(ierr); 4240 4241 /* set coloring for off-diagonal portion */ 4242 ierr = ISAllGatherColors(PetscObjectComm((PetscObject)A),coloring->n,coloring->colors,NULL,&allcolors);CHKERRQ(ierr); 4243 ierr = PetscMalloc1((a->B->cmap->n+1),&colors);CHKERRQ(ierr); 4244 for (i=0; i<a->B->cmap->n; i++) { 4245 colors[i] = allcolors[a->garray[i]]; 4246 } 4247 ierr = PetscFree(allcolors);CHKERRQ(ierr); 4248 ierr = ISColoringCreate(MPI_COMM_SELF,coloring->n,a->B->cmap->n,colors,&ocoloring);CHKERRQ(ierr); 4249 ierr = MatSetColoring_SeqAIJ(a->B,ocoloring);CHKERRQ(ierr); 4250 ierr = ISColoringDestroy(&ocoloring);CHKERRQ(ierr); 4251 } else if (coloring->ctype == IS_COLORING_GHOSTED) { 4252 ISColoringValue *colors; 4253 PetscInt *larray; 4254 ISColoring ocoloring; 4255 4256 /* set coloring for diagonal portion */ 4257 ierr = PetscMalloc1((a->A->cmap->n+1),&larray);CHKERRQ(ierr); 4258 for (i=0; i<a->A->cmap->n; i++) { 4259 larray[i] = i + A->cmap->rstart; 4260 } 4261 ierr = ISGlobalToLocalMappingApply(A->cmap->mapping,IS_GTOLM_MASK,a->A->cmap->n,larray,NULL,larray);CHKERRQ(ierr); 4262 ierr = PetscMalloc1((a->A->cmap->n+1),&colors);CHKERRQ(ierr); 4263 for (i=0; i<a->A->cmap->n; i++) { 4264 colors[i] = coloring->colors[larray[i]]; 4265 } 4266 ierr = PetscFree(larray);CHKERRQ(ierr); 4267 ierr = ISColoringCreate(PETSC_COMM_SELF,coloring->n,a->A->cmap->n,colors,&ocoloring);CHKERRQ(ierr); 4268 ierr = MatSetColoring_SeqAIJ(a->A,ocoloring);CHKERRQ(ierr); 4269 ierr = ISColoringDestroy(&ocoloring);CHKERRQ(ierr); 4270 4271 /* set coloring for off-diagonal portion */ 4272 ierr = PetscMalloc1((a->B->cmap->n+1),&larray);CHKERRQ(ierr); 4273 ierr = ISGlobalToLocalMappingApply(A->cmap->mapping,IS_GTOLM_MASK,a->B->cmap->n,a->garray,NULL,larray);CHKERRQ(ierr); 4274 ierr = PetscMalloc1((a->B->cmap->n+1),&colors);CHKERRQ(ierr); 4275 for (i=0; i<a->B->cmap->n; i++) { 4276 colors[i] = coloring->colors[larray[i]]; 4277 } 4278 ierr = PetscFree(larray);CHKERRQ(ierr); 4279 ierr = ISColoringCreate(MPI_COMM_SELF,coloring->n,a->B->cmap->n,colors,&ocoloring);CHKERRQ(ierr); 4280 ierr = MatSetColoring_SeqAIJ(a->B,ocoloring);CHKERRQ(ierr); 4281 ierr = ISColoringDestroy(&ocoloring);CHKERRQ(ierr); 4282 } else SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"No support ISColoringType %d",(int)coloring->ctype); 4283 PetscFunctionReturn(0); 4284 } 4285 4286 #undef __FUNCT__ 4287 #define __FUNCT__ "MatSetValuesAdifor_MPIAIJ" 4288 PetscErrorCode MatSetValuesAdifor_MPIAIJ(Mat A,PetscInt nl,void *advalues) 4289 { 4290 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 4291 PetscErrorCode ierr; 4292 4293 PetscFunctionBegin; 4294 ierr = MatSetValuesAdifor_SeqAIJ(a->A,nl,advalues);CHKERRQ(ierr); 4295 ierr = MatSetValuesAdifor_SeqAIJ(a->B,nl,advalues);CHKERRQ(ierr); 4296 PetscFunctionReturn(0); 4297 } 4298 4299 #undef __FUNCT__ 4300 #define __FUNCT__ "MatCreateMPIAIJConcatenateSeqAIJSymbolic" 4301 PetscErrorCode MatCreateMPIAIJConcatenateSeqAIJSymbolic(MPI_Comm comm,Mat inmat,PetscInt n,Mat *outmat) 4302 { 4303 PetscErrorCode ierr; 4304 PetscInt m,N,i,rstart,nnz,*dnz,*onz,sum,bs,cbs; 4305 PetscInt *indx; 4306 4307 PetscFunctionBegin; 4308 /* This routine will ONLY return MPIAIJ type matrix */ 4309 ierr = MatGetSize(inmat,&m,&N);CHKERRQ(ierr); 4310 ierr = MatGetBlockSizes(inmat,&bs,&cbs);CHKERRQ(ierr); 4311 if (n == PETSC_DECIDE) { 4312 ierr = PetscSplitOwnership(comm,&n,&N);CHKERRQ(ierr); 4313 } 4314 /* Check sum(n) = N */ 4315 ierr = MPI_Allreduce(&n,&sum,1,MPIU_INT,MPI_SUM,comm);CHKERRQ(ierr); 4316 if (sum != N) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_INCOMP,"Sum of local columns != global columns %d",N); 4317 4318 ierr = MPI_Scan(&m, &rstart,1,MPIU_INT,MPI_SUM,comm);CHKERRQ(ierr); 4319 rstart -= m; 4320 4321 ierr = MatPreallocateInitialize(comm,m,n,dnz,onz);CHKERRQ(ierr); 4322 for (i=0; i<m; i++) { 4323 ierr = MatGetRow_SeqAIJ(inmat,i,&nnz,&indx,NULL);CHKERRQ(ierr); 4324 ierr = MatPreallocateSet(i+rstart,nnz,indx,dnz,onz);CHKERRQ(ierr); 4325 ierr = MatRestoreRow_SeqAIJ(inmat,i,&nnz,&indx,NULL);CHKERRQ(ierr); 4326 } 4327 4328 ierr = MatCreate(comm,outmat);CHKERRQ(ierr); 4329 ierr = MatSetSizes(*outmat,m,n,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 4330 ierr = MatSetBlockSizes(*outmat,bs,cbs);CHKERRQ(ierr); 4331 ierr = MatSetType(*outmat,MATMPIAIJ);CHKERRQ(ierr); 4332 ierr = MatMPIAIJSetPreallocation(*outmat,0,dnz,0,onz);CHKERRQ(ierr); 4333 ierr = MatPreallocateFinalize(dnz,onz);CHKERRQ(ierr); 4334 PetscFunctionReturn(0); 4335 } 4336 4337 #undef __FUNCT__ 4338 #define __FUNCT__ "MatCreateMPIAIJConcatenateSeqAIJNumeric" 4339 PetscErrorCode MatCreateMPIAIJConcatenateSeqAIJNumeric(MPI_Comm comm,Mat inmat,PetscInt n,Mat outmat) 4340 { 4341 PetscErrorCode ierr; 4342 PetscInt m,N,i,rstart,nnz,Ii; 4343 PetscInt *indx; 4344 PetscScalar *values; 4345 4346 PetscFunctionBegin; 4347 ierr = MatGetSize(inmat,&m,&N);CHKERRQ(ierr); 4348 ierr = MatGetOwnershipRange(outmat,&rstart,NULL);CHKERRQ(ierr); 4349 for (i=0; i<m; i++) { 4350 ierr = MatGetRow_SeqAIJ(inmat,i,&nnz,&indx,&values);CHKERRQ(ierr); 4351 Ii = i + rstart; 4352 ierr = MatSetValues(outmat,1,&Ii,nnz,indx,values,INSERT_VALUES);CHKERRQ(ierr); 4353 ierr = MatRestoreRow_SeqAIJ(inmat,i,&nnz,&indx,&values);CHKERRQ(ierr); 4354 } 4355 ierr = MatAssemblyBegin(outmat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4356 ierr = MatAssemblyEnd(outmat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4357 PetscFunctionReturn(0); 4358 } 4359 4360 #undef __FUNCT__ 4361 #define __FUNCT__ "MatCreateMPIMatConcatenateSeqMat_MPIAIJ" 4362 PetscErrorCode MatCreateMPIMatConcatenateSeqMat_MPIAIJ(MPI_Comm comm,Mat inmat,PetscInt n,MatReuse scall,Mat *outmat) 4363 { 4364 PetscErrorCode ierr; 4365 PetscMPIInt size; 4366 4367 PetscFunctionBegin; 4368 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 4369 ierr = PetscLogEventBegin(MAT_Merge,inmat,0,0,0);CHKERRQ(ierr); 4370 if (size == 1) { 4371 if (scall == MAT_INITIAL_MATRIX) { 4372 ierr = MatDuplicate(inmat,MAT_COPY_VALUES,outmat);CHKERRQ(ierr); 4373 } else { 4374 ierr = MatCopy(inmat,*outmat,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 4375 } 4376 } else { 4377 if (scall == MAT_INITIAL_MATRIX) { 4378 ierr = MatCreateMPIAIJConcatenateSeqAIJSymbolic(comm,inmat,n,outmat);CHKERRQ(ierr); 4379 } 4380 ierr = MatCreateMPIAIJConcatenateSeqAIJNumeric(comm,inmat,n,*outmat);CHKERRQ(ierr); 4381 } 4382 ierr = PetscLogEventEnd(MAT_Merge,inmat,0,0,0);CHKERRQ(ierr); 4383 PetscFunctionReturn(0); 4384 } 4385 4386 #undef __FUNCT__ 4387 #define __FUNCT__ "MatFileSplit" 4388 PetscErrorCode MatFileSplit(Mat A,char *outfile) 4389 { 4390 PetscErrorCode ierr; 4391 PetscMPIInt rank; 4392 PetscInt m,N,i,rstart,nnz; 4393 size_t len; 4394 const PetscInt *indx; 4395 PetscViewer out; 4396 char *name; 4397 Mat B; 4398 const PetscScalar *values; 4399 4400 PetscFunctionBegin; 4401 ierr = MatGetLocalSize(A,&m,0);CHKERRQ(ierr); 4402 ierr = MatGetSize(A,0,&N);CHKERRQ(ierr); 4403 /* Should this be the type of the diagonal block of A? */ 4404 ierr = MatCreate(PETSC_COMM_SELF,&B);CHKERRQ(ierr); 4405 ierr = MatSetSizes(B,m,N,m,N);CHKERRQ(ierr); 4406 ierr = MatSetBlockSizesFromMats(B,A,A);CHKERRQ(ierr); 4407 ierr = MatSetType(B,MATSEQAIJ);CHKERRQ(ierr); 4408 ierr = MatSeqAIJSetPreallocation(B,0,NULL);CHKERRQ(ierr); 4409 ierr = MatGetOwnershipRange(A,&rstart,0);CHKERRQ(ierr); 4410 for (i=0; i<m; i++) { 4411 ierr = MatGetRow(A,i+rstart,&nnz,&indx,&values);CHKERRQ(ierr); 4412 ierr = MatSetValues(B,1,&i,nnz,indx,values,INSERT_VALUES);CHKERRQ(ierr); 4413 ierr = MatRestoreRow(A,i+rstart,&nnz,&indx,&values);CHKERRQ(ierr); 4414 } 4415 ierr = MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4416 ierr = MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4417 4418 ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)A),&rank);CHKERRQ(ierr); 4419 ierr = PetscStrlen(outfile,&len);CHKERRQ(ierr); 4420 ierr = PetscMalloc1((len+5),&name);CHKERRQ(ierr); 4421 sprintf(name,"%s.%d",outfile,rank); 4422 ierr = PetscViewerBinaryOpen(PETSC_COMM_SELF,name,FILE_MODE_APPEND,&out);CHKERRQ(ierr); 4423 ierr = PetscFree(name);CHKERRQ(ierr); 4424 ierr = MatView(B,out);CHKERRQ(ierr); 4425 ierr = PetscViewerDestroy(&out);CHKERRQ(ierr); 4426 ierr = MatDestroy(&B);CHKERRQ(ierr); 4427 PetscFunctionReturn(0); 4428 } 4429 4430 extern PetscErrorCode MatDestroy_MPIAIJ(Mat); 4431 #undef __FUNCT__ 4432 #define __FUNCT__ "MatDestroy_MPIAIJ_SeqsToMPI" 4433 PetscErrorCode MatDestroy_MPIAIJ_SeqsToMPI(Mat A) 4434 { 4435 PetscErrorCode ierr; 4436 Mat_Merge_SeqsToMPI *merge; 4437 PetscContainer container; 4438 4439 PetscFunctionBegin; 4440 ierr = PetscObjectQuery((PetscObject)A,"MatMergeSeqsToMPI",(PetscObject*)&container);CHKERRQ(ierr); 4441 if (container) { 4442 ierr = PetscContainerGetPointer(container,(void**)&merge);CHKERRQ(ierr); 4443 ierr = PetscFree(merge->id_r);CHKERRQ(ierr); 4444 ierr = PetscFree(merge->len_s);CHKERRQ(ierr); 4445 ierr = PetscFree(merge->len_r);CHKERRQ(ierr); 4446 ierr = PetscFree(merge->bi);CHKERRQ(ierr); 4447 ierr = PetscFree(merge->bj);CHKERRQ(ierr); 4448 ierr = PetscFree(merge->buf_ri[0]);CHKERRQ(ierr); 4449 ierr = PetscFree(merge->buf_ri);CHKERRQ(ierr); 4450 ierr = PetscFree(merge->buf_rj[0]);CHKERRQ(ierr); 4451 ierr = PetscFree(merge->buf_rj);CHKERRQ(ierr); 4452 ierr = PetscFree(merge->coi);CHKERRQ(ierr); 4453 ierr = PetscFree(merge->coj);CHKERRQ(ierr); 4454 ierr = PetscFree(merge->owners_co);CHKERRQ(ierr); 4455 ierr = PetscLayoutDestroy(&merge->rowmap);CHKERRQ(ierr); 4456 ierr = PetscFree(merge);CHKERRQ(ierr); 4457 ierr = PetscObjectCompose((PetscObject)A,"MatMergeSeqsToMPI",0);CHKERRQ(ierr); 4458 } 4459 ierr = MatDestroy_MPIAIJ(A);CHKERRQ(ierr); 4460 PetscFunctionReturn(0); 4461 } 4462 4463 #include <../src/mat/utils/freespace.h> 4464 #include <petscbt.h> 4465 4466 #undef __FUNCT__ 4467 #define __FUNCT__ "MatCreateMPIAIJSumSeqAIJNumeric" 4468 PetscErrorCode MatCreateMPIAIJSumSeqAIJNumeric(Mat seqmat,Mat mpimat) 4469 { 4470 PetscErrorCode ierr; 4471 MPI_Comm comm; 4472 Mat_SeqAIJ *a =(Mat_SeqAIJ*)seqmat->data; 4473 PetscMPIInt size,rank,taga,*len_s; 4474 PetscInt N=mpimat->cmap->N,i,j,*owners,*ai=a->i,*aj; 4475 PetscInt proc,m; 4476 PetscInt **buf_ri,**buf_rj; 4477 PetscInt k,anzi,*bj_i,*bi,*bj,arow,bnzi,nextaj; 4478 PetscInt nrows,**buf_ri_k,**nextrow,**nextai; 4479 MPI_Request *s_waits,*r_waits; 4480 MPI_Status *status; 4481 MatScalar *aa=a->a; 4482 MatScalar **abuf_r,*ba_i; 4483 Mat_Merge_SeqsToMPI *merge; 4484 PetscContainer container; 4485 4486 PetscFunctionBegin; 4487 ierr = PetscObjectGetComm((PetscObject)mpimat,&comm);CHKERRQ(ierr); 4488 ierr = PetscLogEventBegin(MAT_Seqstompinum,seqmat,0,0,0);CHKERRQ(ierr); 4489 4490 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 4491 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 4492 4493 ierr = PetscObjectQuery((PetscObject)mpimat,"MatMergeSeqsToMPI",(PetscObject*)&container);CHKERRQ(ierr); 4494 ierr = PetscContainerGetPointer(container,(void**)&merge);CHKERRQ(ierr); 4495 4496 bi = merge->bi; 4497 bj = merge->bj; 4498 buf_ri = merge->buf_ri; 4499 buf_rj = merge->buf_rj; 4500 4501 ierr = PetscMalloc1(size,&status);CHKERRQ(ierr); 4502 owners = merge->rowmap->range; 4503 len_s = merge->len_s; 4504 4505 /* send and recv matrix values */ 4506 /*-----------------------------*/ 4507 ierr = PetscObjectGetNewTag((PetscObject)mpimat,&taga);CHKERRQ(ierr); 4508 ierr = PetscPostIrecvScalar(comm,taga,merge->nrecv,merge->id_r,merge->len_r,&abuf_r,&r_waits);CHKERRQ(ierr); 4509 4510 ierr = PetscMalloc1((merge->nsend+1),&s_waits);CHKERRQ(ierr); 4511 for (proc=0,k=0; proc<size; proc++) { 4512 if (!len_s[proc]) continue; 4513 i = owners[proc]; 4514 ierr = MPI_Isend(aa+ai[i],len_s[proc],MPIU_MATSCALAR,proc,taga,comm,s_waits+k);CHKERRQ(ierr); 4515 k++; 4516 } 4517 4518 if (merge->nrecv) {ierr = MPI_Waitall(merge->nrecv,r_waits,status);CHKERRQ(ierr);} 4519 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,s_waits,status);CHKERRQ(ierr);} 4520 ierr = PetscFree(status);CHKERRQ(ierr); 4521 4522 ierr = PetscFree(s_waits);CHKERRQ(ierr); 4523 ierr = PetscFree(r_waits);CHKERRQ(ierr); 4524 4525 /* insert mat values of mpimat */ 4526 /*----------------------------*/ 4527 ierr = PetscMalloc1(N,&ba_i);CHKERRQ(ierr); 4528 ierr = PetscMalloc3(merge->nrecv,&buf_ri_k,merge->nrecv,&nextrow,merge->nrecv,&nextai);CHKERRQ(ierr); 4529 4530 for (k=0; k<merge->nrecv; k++) { 4531 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 4532 nrows = *(buf_ri_k[k]); 4533 nextrow[k] = buf_ri_k[k]+1; /* next row number of k-th recved i-structure */ 4534 nextai[k] = buf_ri_k[k] + (nrows + 1); /* poins to the next i-structure of k-th recved i-structure */ 4535 } 4536 4537 /* set values of ba */ 4538 m = merge->rowmap->n; 4539 for (i=0; i<m; i++) { 4540 arow = owners[rank] + i; 4541 bj_i = bj+bi[i]; /* col indices of the i-th row of mpimat */ 4542 bnzi = bi[i+1] - bi[i]; 4543 ierr = PetscMemzero(ba_i,bnzi*sizeof(PetscScalar));CHKERRQ(ierr); 4544 4545 /* add local non-zero vals of this proc's seqmat into ba */ 4546 anzi = ai[arow+1] - ai[arow]; 4547 aj = a->j + ai[arow]; 4548 aa = a->a + ai[arow]; 4549 nextaj = 0; 4550 for (j=0; nextaj<anzi; j++) { 4551 if (*(bj_i + j) == aj[nextaj]) { /* bcol == acol */ 4552 ba_i[j] += aa[nextaj++]; 4553 } 4554 } 4555 4556 /* add received vals into ba */ 4557 for (k=0; k<merge->nrecv; k++) { /* k-th received message */ 4558 /* i-th row */ 4559 if (i == *nextrow[k]) { 4560 anzi = *(nextai[k]+1) - *nextai[k]; 4561 aj = buf_rj[k] + *(nextai[k]); 4562 aa = abuf_r[k] + *(nextai[k]); 4563 nextaj = 0; 4564 for (j=0; nextaj<anzi; j++) { 4565 if (*(bj_i + j) == aj[nextaj]) { /* bcol == acol */ 4566 ba_i[j] += aa[nextaj++]; 4567 } 4568 } 4569 nextrow[k]++; nextai[k]++; 4570 } 4571 } 4572 ierr = MatSetValues(mpimat,1,&arow,bnzi,bj_i,ba_i,INSERT_VALUES);CHKERRQ(ierr); 4573 } 4574 ierr = MatAssemblyBegin(mpimat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4575 ierr = MatAssemblyEnd(mpimat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4576 4577 ierr = PetscFree(abuf_r[0]);CHKERRQ(ierr); 4578 ierr = PetscFree(abuf_r);CHKERRQ(ierr); 4579 ierr = PetscFree(ba_i);CHKERRQ(ierr); 4580 ierr = PetscFree3(buf_ri_k,nextrow,nextai);CHKERRQ(ierr); 4581 ierr = PetscLogEventEnd(MAT_Seqstompinum,seqmat,0,0,0);CHKERRQ(ierr); 4582 PetscFunctionReturn(0); 4583 } 4584 4585 extern PetscErrorCode MatDestroy_MPIAIJ_SeqsToMPI(Mat); 4586 4587 #undef __FUNCT__ 4588 #define __FUNCT__ "MatCreateMPIAIJSumSeqAIJSymbolic" 4589 PetscErrorCode MatCreateMPIAIJSumSeqAIJSymbolic(MPI_Comm comm,Mat seqmat,PetscInt m,PetscInt n,Mat *mpimat) 4590 { 4591 PetscErrorCode ierr; 4592 Mat B_mpi; 4593 Mat_SeqAIJ *a=(Mat_SeqAIJ*)seqmat->data; 4594 PetscMPIInt size,rank,tagi,tagj,*len_s,*len_si,*len_ri; 4595 PetscInt **buf_rj,**buf_ri,**buf_ri_k; 4596 PetscInt M=seqmat->rmap->n,N=seqmat->cmap->n,i,*owners,*ai=a->i,*aj=a->j; 4597 PetscInt len,proc,*dnz,*onz,bs,cbs; 4598 PetscInt k,anzi,*bi,*bj,*lnk,nlnk,arow,bnzi,nspacedouble=0; 4599 PetscInt nrows,*buf_s,*buf_si,*buf_si_i,**nextrow,**nextai; 4600 MPI_Request *si_waits,*sj_waits,*ri_waits,*rj_waits; 4601 MPI_Status *status; 4602 PetscFreeSpaceList free_space=NULL,current_space=NULL; 4603 PetscBT lnkbt; 4604 Mat_Merge_SeqsToMPI *merge; 4605 PetscContainer container; 4606 4607 PetscFunctionBegin; 4608 ierr = PetscLogEventBegin(MAT_Seqstompisym,seqmat,0,0,0);CHKERRQ(ierr); 4609 4610 /* make sure it is a PETSc comm */ 4611 ierr = PetscCommDuplicate(comm,&comm,NULL);CHKERRQ(ierr); 4612 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 4613 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 4614 4615 ierr = PetscNew(&merge);CHKERRQ(ierr); 4616 ierr = PetscMalloc1(size,&status);CHKERRQ(ierr); 4617 4618 /* determine row ownership */ 4619 /*---------------------------------------------------------*/ 4620 ierr = PetscLayoutCreate(comm,&merge->rowmap);CHKERRQ(ierr); 4621 ierr = PetscLayoutSetLocalSize(merge->rowmap,m);CHKERRQ(ierr); 4622 ierr = PetscLayoutSetSize(merge->rowmap,M);CHKERRQ(ierr); 4623 ierr = PetscLayoutSetBlockSize(merge->rowmap,1);CHKERRQ(ierr); 4624 ierr = PetscLayoutSetUp(merge->rowmap);CHKERRQ(ierr); 4625 ierr = PetscMalloc1(size,&len_si);CHKERRQ(ierr); 4626 ierr = PetscMalloc1(size,&merge->len_s);CHKERRQ(ierr); 4627 4628 m = merge->rowmap->n; 4629 owners = merge->rowmap->range; 4630 4631 /* determine the number of messages to send, their lengths */ 4632 /*---------------------------------------------------------*/ 4633 len_s = merge->len_s; 4634 4635 len = 0; /* length of buf_si[] */ 4636 merge->nsend = 0; 4637 for (proc=0; proc<size; proc++) { 4638 len_si[proc] = 0; 4639 if (proc == rank) { 4640 len_s[proc] = 0; 4641 } else { 4642 len_si[proc] = owners[proc+1] - owners[proc] + 1; 4643 len_s[proc] = ai[owners[proc+1]] - ai[owners[proc]]; /* num of rows to be sent to [proc] */ 4644 } 4645 if (len_s[proc]) { 4646 merge->nsend++; 4647 nrows = 0; 4648 for (i=owners[proc]; i<owners[proc+1]; i++) { 4649 if (ai[i+1] > ai[i]) nrows++; 4650 } 4651 len_si[proc] = 2*(nrows+1); 4652 len += len_si[proc]; 4653 } 4654 } 4655 4656 /* determine the number and length of messages to receive for ij-structure */ 4657 /*-------------------------------------------------------------------------*/ 4658 ierr = PetscGatherNumberOfMessages(comm,NULL,len_s,&merge->nrecv);CHKERRQ(ierr); 4659 ierr = PetscGatherMessageLengths2(comm,merge->nsend,merge->nrecv,len_s,len_si,&merge->id_r,&merge->len_r,&len_ri);CHKERRQ(ierr); 4660 4661 /* post the Irecv of j-structure */ 4662 /*-------------------------------*/ 4663 ierr = PetscCommGetNewTag(comm,&tagj);CHKERRQ(ierr); 4664 ierr = PetscPostIrecvInt(comm,tagj,merge->nrecv,merge->id_r,merge->len_r,&buf_rj,&rj_waits);CHKERRQ(ierr); 4665 4666 /* post the Isend of j-structure */ 4667 /*--------------------------------*/ 4668 ierr = PetscMalloc2(merge->nsend,&si_waits,merge->nsend,&sj_waits);CHKERRQ(ierr); 4669 4670 for (proc=0, k=0; proc<size; proc++) { 4671 if (!len_s[proc]) continue; 4672 i = owners[proc]; 4673 ierr = MPI_Isend(aj+ai[i],len_s[proc],MPIU_INT,proc,tagj,comm,sj_waits+k);CHKERRQ(ierr); 4674 k++; 4675 } 4676 4677 /* receives and sends of j-structure are complete */ 4678 /*------------------------------------------------*/ 4679 if (merge->nrecv) {ierr = MPI_Waitall(merge->nrecv,rj_waits,status);CHKERRQ(ierr);} 4680 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,sj_waits,status);CHKERRQ(ierr);} 4681 4682 /* send and recv i-structure */ 4683 /*---------------------------*/ 4684 ierr = PetscCommGetNewTag(comm,&tagi);CHKERRQ(ierr); 4685 ierr = PetscPostIrecvInt(comm,tagi,merge->nrecv,merge->id_r,len_ri,&buf_ri,&ri_waits);CHKERRQ(ierr); 4686 4687 ierr = PetscMalloc1((len+1),&buf_s);CHKERRQ(ierr); 4688 buf_si = buf_s; /* points to the beginning of k-th msg to be sent */ 4689 for (proc=0,k=0; proc<size; proc++) { 4690 if (!len_s[proc]) continue; 4691 /* form outgoing message for i-structure: 4692 buf_si[0]: nrows to be sent 4693 [1:nrows]: row index (global) 4694 [nrows+1:2*nrows+1]: i-structure index 4695 */ 4696 /*-------------------------------------------*/ 4697 nrows = len_si[proc]/2 - 1; 4698 buf_si_i = buf_si + nrows+1; 4699 buf_si[0] = nrows; 4700 buf_si_i[0] = 0; 4701 nrows = 0; 4702 for (i=owners[proc]; i<owners[proc+1]; i++) { 4703 anzi = ai[i+1] - ai[i]; 4704 if (anzi) { 4705 buf_si_i[nrows+1] = buf_si_i[nrows] + anzi; /* i-structure */ 4706 buf_si[nrows+1] = i-owners[proc]; /* local row index */ 4707 nrows++; 4708 } 4709 } 4710 ierr = MPI_Isend(buf_si,len_si[proc],MPIU_INT,proc,tagi,comm,si_waits+k);CHKERRQ(ierr); 4711 k++; 4712 buf_si += len_si[proc]; 4713 } 4714 4715 if (merge->nrecv) {ierr = MPI_Waitall(merge->nrecv,ri_waits,status);CHKERRQ(ierr);} 4716 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,si_waits,status);CHKERRQ(ierr);} 4717 4718 ierr = PetscInfo2(seqmat,"nsend: %D, nrecv: %D\n",merge->nsend,merge->nrecv);CHKERRQ(ierr); 4719 for (i=0; i<merge->nrecv; i++) { 4720 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); 4721 } 4722 4723 ierr = PetscFree(len_si);CHKERRQ(ierr); 4724 ierr = PetscFree(len_ri);CHKERRQ(ierr); 4725 ierr = PetscFree(rj_waits);CHKERRQ(ierr); 4726 ierr = PetscFree2(si_waits,sj_waits);CHKERRQ(ierr); 4727 ierr = PetscFree(ri_waits);CHKERRQ(ierr); 4728 ierr = PetscFree(buf_s);CHKERRQ(ierr); 4729 ierr = PetscFree(status);CHKERRQ(ierr); 4730 4731 /* compute a local seq matrix in each processor */ 4732 /*----------------------------------------------*/ 4733 /* allocate bi array and free space for accumulating nonzero column info */ 4734 ierr = PetscMalloc1((m+1),&bi);CHKERRQ(ierr); 4735 bi[0] = 0; 4736 4737 /* create and initialize a linked list */ 4738 nlnk = N+1; 4739 ierr = PetscLLCreate(N,N,nlnk,lnk,lnkbt);CHKERRQ(ierr); 4740 4741 /* initial FreeSpace size is 2*(num of local nnz(seqmat)) */ 4742 len = ai[owners[rank+1]] - ai[owners[rank]]; 4743 ierr = PetscFreeSpaceGet((PetscInt)(2*len+1),&free_space);CHKERRQ(ierr); 4744 4745 current_space = free_space; 4746 4747 /* determine symbolic info for each local row */ 4748 ierr = PetscMalloc3(merge->nrecv,&buf_ri_k,merge->nrecv,&nextrow,merge->nrecv,&nextai);CHKERRQ(ierr); 4749 4750 for (k=0; k<merge->nrecv; k++) { 4751 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 4752 nrows = *buf_ri_k[k]; 4753 nextrow[k] = buf_ri_k[k] + 1; /* next row number of k-th recved i-structure */ 4754 nextai[k] = buf_ri_k[k] + (nrows + 1); /* poins to the next i-structure of k-th recved i-structure */ 4755 } 4756 4757 ierr = MatPreallocateInitialize(comm,m,n,dnz,onz);CHKERRQ(ierr); 4758 len = 0; 4759 for (i=0; i<m; i++) { 4760 bnzi = 0; 4761 /* add local non-zero cols of this proc's seqmat into lnk */ 4762 arow = owners[rank] + i; 4763 anzi = ai[arow+1] - ai[arow]; 4764 aj = a->j + ai[arow]; 4765 ierr = PetscLLAddSorted(anzi,aj,N,nlnk,lnk,lnkbt);CHKERRQ(ierr); 4766 bnzi += nlnk; 4767 /* add received col data into lnk */ 4768 for (k=0; k<merge->nrecv; k++) { /* k-th received message */ 4769 if (i == *nextrow[k]) { /* i-th row */ 4770 anzi = *(nextai[k]+1) - *nextai[k]; 4771 aj = buf_rj[k] + *nextai[k]; 4772 ierr = PetscLLAddSorted(anzi,aj,N,nlnk,lnk,lnkbt);CHKERRQ(ierr); 4773 bnzi += nlnk; 4774 nextrow[k]++; nextai[k]++; 4775 } 4776 } 4777 if (len < bnzi) len = bnzi; /* =max(bnzi) */ 4778 4779 /* if free space is not available, make more free space */ 4780 if (current_space->local_remaining<bnzi) { 4781 ierr = PetscFreeSpaceGet(bnzi+current_space->total_array_size,¤t_space);CHKERRQ(ierr); 4782 nspacedouble++; 4783 } 4784 /* copy data into free space, then initialize lnk */ 4785 ierr = PetscLLClean(N,N,bnzi,lnk,current_space->array,lnkbt);CHKERRQ(ierr); 4786 ierr = MatPreallocateSet(i+owners[rank],bnzi,current_space->array,dnz,onz);CHKERRQ(ierr); 4787 4788 current_space->array += bnzi; 4789 current_space->local_used += bnzi; 4790 current_space->local_remaining -= bnzi; 4791 4792 bi[i+1] = bi[i] + bnzi; 4793 } 4794 4795 ierr = PetscFree3(buf_ri_k,nextrow,nextai);CHKERRQ(ierr); 4796 4797 ierr = PetscMalloc1((bi[m]+1),&bj);CHKERRQ(ierr); 4798 ierr = PetscFreeSpaceContiguous(&free_space,bj);CHKERRQ(ierr); 4799 ierr = PetscLLDestroy(lnk,lnkbt);CHKERRQ(ierr); 4800 4801 /* create symbolic parallel matrix B_mpi */ 4802 /*---------------------------------------*/ 4803 ierr = MatGetBlockSizes(seqmat,&bs,&cbs);CHKERRQ(ierr); 4804 ierr = MatCreate(comm,&B_mpi);CHKERRQ(ierr); 4805 if (n==PETSC_DECIDE) { 4806 ierr = MatSetSizes(B_mpi,m,n,PETSC_DETERMINE,N);CHKERRQ(ierr); 4807 } else { 4808 ierr = MatSetSizes(B_mpi,m,n,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 4809 } 4810 ierr = MatSetBlockSizes(B_mpi,bs,cbs);CHKERRQ(ierr); 4811 ierr = MatSetType(B_mpi,MATMPIAIJ);CHKERRQ(ierr); 4812 ierr = MatMPIAIJSetPreallocation(B_mpi,0,dnz,0,onz);CHKERRQ(ierr); 4813 ierr = MatPreallocateFinalize(dnz,onz);CHKERRQ(ierr); 4814 ierr = MatSetOption(B_mpi,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 4815 4816 /* B_mpi is not ready for use - assembly will be done by MatCreateMPIAIJSumSeqAIJNumeric() */ 4817 B_mpi->assembled = PETSC_FALSE; 4818 B_mpi->ops->destroy = MatDestroy_MPIAIJ_SeqsToMPI; 4819 merge->bi = bi; 4820 merge->bj = bj; 4821 merge->buf_ri = buf_ri; 4822 merge->buf_rj = buf_rj; 4823 merge->coi = NULL; 4824 merge->coj = NULL; 4825 merge->owners_co = NULL; 4826 4827 ierr = PetscCommDestroy(&comm);CHKERRQ(ierr); 4828 4829 /* attach the supporting struct to B_mpi for reuse */ 4830 ierr = PetscContainerCreate(PETSC_COMM_SELF,&container);CHKERRQ(ierr); 4831 ierr = PetscContainerSetPointer(container,merge);CHKERRQ(ierr); 4832 ierr = PetscObjectCompose((PetscObject)B_mpi,"MatMergeSeqsToMPI",(PetscObject)container);CHKERRQ(ierr); 4833 ierr = PetscContainerDestroy(&container);CHKERRQ(ierr); 4834 *mpimat = B_mpi; 4835 4836 ierr = PetscLogEventEnd(MAT_Seqstompisym,seqmat,0,0,0);CHKERRQ(ierr); 4837 PetscFunctionReturn(0); 4838 } 4839 4840 #undef __FUNCT__ 4841 #define __FUNCT__ "MatCreateMPIAIJSumSeqAIJ" 4842 /*@C 4843 MatCreateMPIAIJSumSeqAIJ - Creates a MPIAIJ matrix by adding sequential 4844 matrices from each processor 4845 4846 Collective on MPI_Comm 4847 4848 Input Parameters: 4849 + comm - the communicators the parallel matrix will live on 4850 . seqmat - the input sequential matrices 4851 . m - number of local rows (or PETSC_DECIDE) 4852 . n - number of local columns (or PETSC_DECIDE) 4853 - scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX 4854 4855 Output Parameter: 4856 . mpimat - the parallel matrix generated 4857 4858 Level: advanced 4859 4860 Notes: 4861 The dimensions of the sequential matrix in each processor MUST be the same. 4862 The input seqmat is included into the container "Mat_Merge_SeqsToMPI", and will be 4863 destroyed when mpimat is destroyed. Call PetscObjectQuery() to access seqmat. 4864 @*/ 4865 PetscErrorCode MatCreateMPIAIJSumSeqAIJ(MPI_Comm comm,Mat seqmat,PetscInt m,PetscInt n,MatReuse scall,Mat *mpimat) 4866 { 4867 PetscErrorCode ierr; 4868 PetscMPIInt size; 4869 4870 PetscFunctionBegin; 4871 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 4872 if (size == 1) { 4873 ierr = PetscLogEventBegin(MAT_Seqstompi,seqmat,0,0,0);CHKERRQ(ierr); 4874 if (scall == MAT_INITIAL_MATRIX) { 4875 ierr = MatDuplicate(seqmat,MAT_COPY_VALUES,mpimat);CHKERRQ(ierr); 4876 } else { 4877 ierr = MatCopy(seqmat,*mpimat,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 4878 } 4879 ierr = PetscLogEventEnd(MAT_Seqstompi,seqmat,0,0,0);CHKERRQ(ierr); 4880 PetscFunctionReturn(0); 4881 } 4882 ierr = PetscLogEventBegin(MAT_Seqstompi,seqmat,0,0,0);CHKERRQ(ierr); 4883 if (scall == MAT_INITIAL_MATRIX) { 4884 ierr = MatCreateMPIAIJSumSeqAIJSymbolic(comm,seqmat,m,n,mpimat);CHKERRQ(ierr); 4885 } 4886 ierr = MatCreateMPIAIJSumSeqAIJNumeric(seqmat,*mpimat);CHKERRQ(ierr); 4887 ierr = PetscLogEventEnd(MAT_Seqstompi,seqmat,0,0,0);CHKERRQ(ierr); 4888 PetscFunctionReturn(0); 4889 } 4890 4891 #undef __FUNCT__ 4892 #define __FUNCT__ "MatMPIAIJGetLocalMat" 4893 /*@ 4894 MatMPIAIJGetLocalMat - Creates a SeqAIJ from a MPIAIJ matrix by taking all its local rows and putting them into a sequential vector with 4895 mlocal rows and n columns. Where mlocal is the row count obtained with MatGetLocalSize() and n is the global column count obtained 4896 with MatGetSize() 4897 4898 Not Collective 4899 4900 Input Parameters: 4901 + A - the matrix 4902 . scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX 4903 4904 Output Parameter: 4905 . A_loc - the local sequential matrix generated 4906 4907 Level: developer 4908 4909 .seealso: MatGetOwnerShipRange(), MatMPIAIJGetLocalMatCondensed() 4910 4911 @*/ 4912 PetscErrorCode MatMPIAIJGetLocalMat(Mat A,MatReuse scall,Mat *A_loc) 4913 { 4914 PetscErrorCode ierr; 4915 Mat_MPIAIJ *mpimat=(Mat_MPIAIJ*)A->data; 4916 Mat_SeqAIJ *mat,*a,*b; 4917 PetscInt *ai,*aj,*bi,*bj,*cmap=mpimat->garray; 4918 MatScalar *aa,*ba,*cam; 4919 PetscScalar *ca; 4920 PetscInt am=A->rmap->n,i,j,k,cstart=A->cmap->rstart; 4921 PetscInt *ci,*cj,col,ncols_d,ncols_o,jo; 4922 PetscBool match; 4923 MPI_Comm comm; 4924 PetscMPIInt size; 4925 4926 PetscFunctionBegin; 4927 ierr = PetscObjectTypeCompare((PetscObject)A,MATMPIAIJ,&match);CHKERRQ(ierr); 4928 if (!match) SETERRQ(PetscObjectComm((PetscObject)A), PETSC_ERR_SUP,"Requires MPIAIJ matrix as input"); 4929 ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 4930 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 4931 if (size == 1 && scall == MAT_REUSE_MATRIX) PetscFunctionReturn(0); 4932 4933 ierr = PetscLogEventBegin(MAT_Getlocalmat,A,0,0,0);CHKERRQ(ierr); 4934 a = (Mat_SeqAIJ*)(mpimat->A)->data; 4935 b = (Mat_SeqAIJ*)(mpimat->B)->data; 4936 ai = a->i; aj = a->j; bi = b->i; bj = b->j; 4937 aa = a->a; ba = b->a; 4938 if (scall == MAT_INITIAL_MATRIX) { 4939 if (size == 1) { 4940 ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,am,A->cmap->N,ai,aj,aa,A_loc);CHKERRQ(ierr); 4941 PetscFunctionReturn(0); 4942 } 4943 4944 ierr = PetscMalloc1((1+am),&ci);CHKERRQ(ierr); 4945 ci[0] = 0; 4946 for (i=0; i<am; i++) { 4947 ci[i+1] = ci[i] + (ai[i+1] - ai[i]) + (bi[i+1] - bi[i]); 4948 } 4949 ierr = PetscMalloc1((1+ci[am]),&cj);CHKERRQ(ierr); 4950 ierr = PetscMalloc1((1+ci[am]),&ca);CHKERRQ(ierr); 4951 k = 0; 4952 for (i=0; i<am; i++) { 4953 ncols_o = bi[i+1] - bi[i]; 4954 ncols_d = ai[i+1] - ai[i]; 4955 /* off-diagonal portion of A */ 4956 for (jo=0; jo<ncols_o; jo++) { 4957 col = cmap[*bj]; 4958 if (col >= cstart) break; 4959 cj[k] = col; bj++; 4960 ca[k++] = *ba++; 4961 } 4962 /* diagonal portion of A */ 4963 for (j=0; j<ncols_d; j++) { 4964 cj[k] = cstart + *aj++; 4965 ca[k++] = *aa++; 4966 } 4967 /* off-diagonal portion of A */ 4968 for (j=jo; j<ncols_o; j++) { 4969 cj[k] = cmap[*bj++]; 4970 ca[k++] = *ba++; 4971 } 4972 } 4973 /* put together the new matrix */ 4974 ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,am,A->cmap->N,ci,cj,ca,A_loc);CHKERRQ(ierr); 4975 /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */ 4976 /* Since these are PETSc arrays, change flags to free them as necessary. */ 4977 mat = (Mat_SeqAIJ*)(*A_loc)->data; 4978 mat->free_a = PETSC_TRUE; 4979 mat->free_ij = PETSC_TRUE; 4980 mat->nonew = 0; 4981 } else if (scall == MAT_REUSE_MATRIX) { 4982 mat=(Mat_SeqAIJ*)(*A_loc)->data; 4983 ci = mat->i; cj = mat->j; cam = mat->a; 4984 for (i=0; i<am; i++) { 4985 /* off-diagonal portion of A */ 4986 ncols_o = bi[i+1] - bi[i]; 4987 for (jo=0; jo<ncols_o; jo++) { 4988 col = cmap[*bj]; 4989 if (col >= cstart) break; 4990 *cam++ = *ba++; bj++; 4991 } 4992 /* diagonal portion of A */ 4993 ncols_d = ai[i+1] - ai[i]; 4994 for (j=0; j<ncols_d; j++) *cam++ = *aa++; 4995 /* off-diagonal portion of A */ 4996 for (j=jo; j<ncols_o; j++) { 4997 *cam++ = *ba++; bj++; 4998 } 4999 } 5000 } else SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Invalid MatReuse %d",(int)scall); 5001 ierr = PetscLogEventEnd(MAT_Getlocalmat,A,0,0,0);CHKERRQ(ierr); 5002 PetscFunctionReturn(0); 5003 } 5004 5005 #undef __FUNCT__ 5006 #define __FUNCT__ "MatMPIAIJGetLocalMatCondensed" 5007 /*@C 5008 MatMPIAIJGetLocalMatCondensed - Creates a SeqAIJ matrix from an MPIAIJ matrix by taking all its local rows and NON-ZERO columns 5009 5010 Not Collective 5011 5012 Input Parameters: 5013 + A - the matrix 5014 . scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX 5015 - row, col - index sets of rows and columns to extract (or NULL) 5016 5017 Output Parameter: 5018 . A_loc - the local sequential matrix generated 5019 5020 Level: developer 5021 5022 .seealso: MatGetOwnershipRange(), MatMPIAIJGetLocalMat() 5023 5024 @*/ 5025 PetscErrorCode MatMPIAIJGetLocalMatCondensed(Mat A,MatReuse scall,IS *row,IS *col,Mat *A_loc) 5026 { 5027 Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data; 5028 PetscErrorCode ierr; 5029 PetscInt i,start,end,ncols,nzA,nzB,*cmap,imark,*idx; 5030 IS isrowa,iscola; 5031 Mat *aloc; 5032 PetscBool match; 5033 5034 PetscFunctionBegin; 5035 ierr = PetscObjectTypeCompare((PetscObject)A,MATMPIAIJ,&match);CHKERRQ(ierr); 5036 if (!match) SETERRQ(PetscObjectComm((PetscObject)A), PETSC_ERR_SUP,"Requires MPIAIJ matrix as input"); 5037 ierr = PetscLogEventBegin(MAT_Getlocalmatcondensed,A,0,0,0);CHKERRQ(ierr); 5038 if (!row) { 5039 start = A->rmap->rstart; end = A->rmap->rend; 5040 ierr = ISCreateStride(PETSC_COMM_SELF,end-start,start,1,&isrowa);CHKERRQ(ierr); 5041 } else { 5042 isrowa = *row; 5043 } 5044 if (!col) { 5045 start = A->cmap->rstart; 5046 cmap = a->garray; 5047 nzA = a->A->cmap->n; 5048 nzB = a->B->cmap->n; 5049 ierr = PetscMalloc1((nzA+nzB), &idx);CHKERRQ(ierr); 5050 ncols = 0; 5051 for (i=0; i<nzB; i++) { 5052 if (cmap[i] < start) idx[ncols++] = cmap[i]; 5053 else break; 5054 } 5055 imark = i; 5056 for (i=0; i<nzA; i++) idx[ncols++] = start + i; 5057 for (i=imark; i<nzB; i++) idx[ncols++] = cmap[i]; 5058 ierr = ISCreateGeneral(PETSC_COMM_SELF,ncols,idx,PETSC_OWN_POINTER,&iscola);CHKERRQ(ierr); 5059 } else { 5060 iscola = *col; 5061 } 5062 if (scall != MAT_INITIAL_MATRIX) { 5063 ierr = PetscMalloc(sizeof(Mat),&aloc);CHKERRQ(ierr); 5064 aloc[0] = *A_loc; 5065 } 5066 ierr = MatGetSubMatrices(A,1,&isrowa,&iscola,scall,&aloc);CHKERRQ(ierr); 5067 *A_loc = aloc[0]; 5068 ierr = PetscFree(aloc);CHKERRQ(ierr); 5069 if (!row) { 5070 ierr = ISDestroy(&isrowa);CHKERRQ(ierr); 5071 } 5072 if (!col) { 5073 ierr = ISDestroy(&iscola);CHKERRQ(ierr); 5074 } 5075 ierr = PetscLogEventEnd(MAT_Getlocalmatcondensed,A,0,0,0);CHKERRQ(ierr); 5076 PetscFunctionReturn(0); 5077 } 5078 5079 #undef __FUNCT__ 5080 #define __FUNCT__ "MatGetBrowsOfAcols" 5081 /*@C 5082 MatGetBrowsOfAcols - Creates a SeqAIJ matrix by taking rows of B that equal to nonzero columns of local A 5083 5084 Collective on Mat 5085 5086 Input Parameters: 5087 + A,B - the matrices in mpiaij format 5088 . scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX 5089 - rowb, colb - index sets of rows and columns of B to extract (or NULL) 5090 5091 Output Parameter: 5092 + rowb, colb - index sets of rows and columns of B to extract 5093 - B_seq - the sequential matrix generated 5094 5095 Level: developer 5096 5097 @*/ 5098 PetscErrorCode MatGetBrowsOfAcols(Mat A,Mat B,MatReuse scall,IS *rowb,IS *colb,Mat *B_seq) 5099 { 5100 Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data; 5101 PetscErrorCode ierr; 5102 PetscInt *idx,i,start,ncols,nzA,nzB,*cmap,imark; 5103 IS isrowb,iscolb; 5104 Mat *bseq=NULL; 5105 5106 PetscFunctionBegin; 5107 if (A->cmap->rstart != B->rmap->rstart || A->cmap->rend != B->rmap->rend) { 5108 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); 5109 } 5110 ierr = PetscLogEventBegin(MAT_GetBrowsOfAcols,A,B,0,0);CHKERRQ(ierr); 5111 5112 if (scall == MAT_INITIAL_MATRIX) { 5113 start = A->cmap->rstart; 5114 cmap = a->garray; 5115 nzA = a->A->cmap->n; 5116 nzB = a->B->cmap->n; 5117 ierr = PetscMalloc1((nzA+nzB), &idx);CHKERRQ(ierr); 5118 ncols = 0; 5119 for (i=0; i<nzB; i++) { /* row < local row index */ 5120 if (cmap[i] < start) idx[ncols++] = cmap[i]; 5121 else break; 5122 } 5123 imark = i; 5124 for (i=0; i<nzA; i++) idx[ncols++] = start + i; /* local rows */ 5125 for (i=imark; i<nzB; i++) idx[ncols++] = cmap[i]; /* row > local row index */ 5126 ierr = ISCreateGeneral(PETSC_COMM_SELF,ncols,idx,PETSC_OWN_POINTER,&isrowb);CHKERRQ(ierr); 5127 ierr = ISCreateStride(PETSC_COMM_SELF,B->cmap->N,0,1,&iscolb);CHKERRQ(ierr); 5128 } else { 5129 if (!rowb || !colb) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"IS rowb and colb must be provided for MAT_REUSE_MATRIX"); 5130 isrowb = *rowb; iscolb = *colb; 5131 ierr = PetscMalloc(sizeof(Mat),&bseq);CHKERRQ(ierr); 5132 bseq[0] = *B_seq; 5133 } 5134 ierr = MatGetSubMatrices(B,1,&isrowb,&iscolb,scall,&bseq);CHKERRQ(ierr); 5135 *B_seq = bseq[0]; 5136 ierr = PetscFree(bseq);CHKERRQ(ierr); 5137 if (!rowb) { 5138 ierr = ISDestroy(&isrowb);CHKERRQ(ierr); 5139 } else { 5140 *rowb = isrowb; 5141 } 5142 if (!colb) { 5143 ierr = ISDestroy(&iscolb);CHKERRQ(ierr); 5144 } else { 5145 *colb = iscolb; 5146 } 5147 ierr = PetscLogEventEnd(MAT_GetBrowsOfAcols,A,B,0,0);CHKERRQ(ierr); 5148 PetscFunctionReturn(0); 5149 } 5150 5151 #undef __FUNCT__ 5152 #define __FUNCT__ "MatGetBrowsOfAoCols_MPIAIJ" 5153 /* 5154 MatGetBrowsOfAoCols_MPIAIJ - Creates a SeqAIJ matrix by taking rows of B that equal to nonzero columns 5155 of the OFF-DIAGONAL portion of local A 5156 5157 Collective on Mat 5158 5159 Input Parameters: 5160 + A,B - the matrices in mpiaij format 5161 - scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX 5162 5163 Output Parameter: 5164 + startsj_s - starting point in B's sending j-arrays, saved for MAT_REUSE (or NULL) 5165 . startsj_r - starting point in B's receiving j-arrays, saved for MAT_REUSE (or NULL) 5166 . bufa_ptr - array for sending matrix values, saved for MAT_REUSE (or NULL) 5167 - B_oth - the sequential matrix generated with size aBn=a->B->cmap->n by B->cmap->N 5168 5169 Level: developer 5170 5171 */ 5172 PetscErrorCode MatGetBrowsOfAoCols_MPIAIJ(Mat A,Mat B,MatReuse scall,PetscInt **startsj_s,PetscInt **startsj_r,MatScalar **bufa_ptr,Mat *B_oth) 5173 { 5174 VecScatter_MPI_General *gen_to,*gen_from; 5175 PetscErrorCode ierr; 5176 Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data; 5177 Mat_SeqAIJ *b_oth; 5178 VecScatter ctx =a->Mvctx; 5179 MPI_Comm comm; 5180 PetscMPIInt *rprocs,*sprocs,tag=((PetscObject)ctx)->tag,rank; 5181 PetscInt *rowlen,*bufj,*bufJ,ncols,aBn=a->B->cmap->n,row,*b_othi,*b_othj; 5182 PetscScalar *rvalues,*svalues; 5183 MatScalar *b_otha,*bufa,*bufA; 5184 PetscInt i,j,k,l,ll,nrecvs,nsends,nrows,*srow,*rstarts,*rstartsj = 0,*sstarts,*sstartsj,len; 5185 MPI_Request *rwaits = NULL,*swaits = NULL; 5186 MPI_Status *sstatus,rstatus; 5187 PetscMPIInt jj,size; 5188 PetscInt *cols,sbs,rbs; 5189 PetscScalar *vals; 5190 5191 PetscFunctionBegin; 5192 ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 5193 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 5194 if (size == 1) PetscFunctionReturn(0); 5195 5196 if (A->cmap->rstart != B->rmap->rstart || A->cmap->rend != B->rmap->rend) { 5197 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); 5198 } 5199 ierr = PetscLogEventBegin(MAT_GetBrowsOfAocols,A,B,0,0);CHKERRQ(ierr); 5200 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 5201 5202 gen_to = (VecScatter_MPI_General*)ctx->todata; 5203 gen_from = (VecScatter_MPI_General*)ctx->fromdata; 5204 rvalues = gen_from->values; /* holds the length of receiving row */ 5205 svalues = gen_to->values; /* holds the length of sending row */ 5206 nrecvs = gen_from->n; 5207 nsends = gen_to->n; 5208 5209 ierr = PetscMalloc2(nrecvs,&rwaits,nsends,&swaits);CHKERRQ(ierr); 5210 srow = gen_to->indices; /* local row index to be sent */ 5211 sstarts = gen_to->starts; 5212 sprocs = gen_to->procs; 5213 sstatus = gen_to->sstatus; 5214 sbs = gen_to->bs; 5215 rstarts = gen_from->starts; 5216 rprocs = gen_from->procs; 5217 rbs = gen_from->bs; 5218 5219 if (!startsj_s || !bufa_ptr) scall = MAT_INITIAL_MATRIX; 5220 if (scall == MAT_INITIAL_MATRIX) { 5221 /* i-array */ 5222 /*---------*/ 5223 /* post receives */ 5224 for (i=0; i<nrecvs; i++) { 5225 rowlen = (PetscInt*)rvalues + rstarts[i]*rbs; 5226 nrows = (rstarts[i+1]-rstarts[i])*rbs; /* num of indices to be received */ 5227 ierr = MPI_Irecv(rowlen,nrows,MPIU_INT,rprocs[i],tag,comm,rwaits+i);CHKERRQ(ierr); 5228 } 5229 5230 /* pack the outgoing message */ 5231 ierr = PetscMalloc2(nsends+1,&sstartsj,nrecvs+1,&rstartsj);CHKERRQ(ierr); 5232 5233 sstartsj[0] = 0; 5234 rstartsj[0] = 0; 5235 len = 0; /* total length of j or a array to be sent */ 5236 k = 0; 5237 for (i=0; i<nsends; i++) { 5238 rowlen = (PetscInt*)svalues + sstarts[i]*sbs; 5239 nrows = sstarts[i+1]-sstarts[i]; /* num of block rows */ 5240 for (j=0; j<nrows; j++) { 5241 row = srow[k] + B->rmap->range[rank]; /* global row idx */ 5242 for (l=0; l<sbs; l++) { 5243 ierr = MatGetRow_MPIAIJ(B,row+l,&ncols,NULL,NULL);CHKERRQ(ierr); /* rowlength */ 5244 5245 rowlen[j*sbs+l] = ncols; 5246 5247 len += ncols; 5248 ierr = MatRestoreRow_MPIAIJ(B,row+l,&ncols,NULL,NULL);CHKERRQ(ierr); 5249 } 5250 k++; 5251 } 5252 ierr = MPI_Isend(rowlen,nrows*sbs,MPIU_INT,sprocs[i],tag,comm,swaits+i);CHKERRQ(ierr); 5253 5254 sstartsj[i+1] = len; /* starting point of (i+1)-th outgoing msg in bufj and bufa */ 5255 } 5256 /* recvs and sends of i-array are completed */ 5257 i = nrecvs; 5258 while (i--) { 5259 ierr = MPI_Waitany(nrecvs,rwaits,&jj,&rstatus);CHKERRQ(ierr); 5260 } 5261 if (nsends) {ierr = MPI_Waitall(nsends,swaits,sstatus);CHKERRQ(ierr);} 5262 5263 /* allocate buffers for sending j and a arrays */ 5264 ierr = PetscMalloc1((len+1),&bufj);CHKERRQ(ierr); 5265 ierr = PetscMalloc1((len+1),&bufa);CHKERRQ(ierr); 5266 5267 /* create i-array of B_oth */ 5268 ierr = PetscMalloc1((aBn+2),&b_othi);CHKERRQ(ierr); 5269 5270 b_othi[0] = 0; 5271 len = 0; /* total length of j or a array to be received */ 5272 k = 0; 5273 for (i=0; i<nrecvs; i++) { 5274 rowlen = (PetscInt*)rvalues + rstarts[i]*rbs; 5275 nrows = rbs*(rstarts[i+1]-rstarts[i]); /* num of rows to be recieved */ 5276 for (j=0; j<nrows; j++) { 5277 b_othi[k+1] = b_othi[k] + rowlen[j]; 5278 len += rowlen[j]; k++; 5279 } 5280 rstartsj[i+1] = len; /* starting point of (i+1)-th incoming msg in bufj and bufa */ 5281 } 5282 5283 /* allocate space for j and a arrrays of B_oth */ 5284 ierr = PetscMalloc1((b_othi[aBn]+1),&b_othj);CHKERRQ(ierr); 5285 ierr = PetscMalloc1((b_othi[aBn]+1),&b_otha);CHKERRQ(ierr); 5286 5287 /* j-array */ 5288 /*---------*/ 5289 /* post receives of j-array */ 5290 for (i=0; i<nrecvs; i++) { 5291 nrows = rstartsj[i+1]-rstartsj[i]; /* length of the msg received */ 5292 ierr = MPI_Irecv(b_othj+rstartsj[i],nrows,MPIU_INT,rprocs[i],tag,comm,rwaits+i);CHKERRQ(ierr); 5293 } 5294 5295 /* pack the outgoing message j-array */ 5296 k = 0; 5297 for (i=0; i<nsends; i++) { 5298 nrows = sstarts[i+1]-sstarts[i]; /* num of block rows */ 5299 bufJ = bufj+sstartsj[i]; 5300 for (j=0; j<nrows; j++) { 5301 row = srow[k++] + B->rmap->range[rank]; /* global row idx */ 5302 for (ll=0; ll<sbs; ll++) { 5303 ierr = MatGetRow_MPIAIJ(B,row+ll,&ncols,&cols,NULL);CHKERRQ(ierr); 5304 for (l=0; l<ncols; l++) { 5305 *bufJ++ = cols[l]; 5306 } 5307 ierr = MatRestoreRow_MPIAIJ(B,row+ll,&ncols,&cols,NULL);CHKERRQ(ierr); 5308 } 5309 } 5310 ierr = MPI_Isend(bufj+sstartsj[i],sstartsj[i+1]-sstartsj[i],MPIU_INT,sprocs[i],tag,comm,swaits+i);CHKERRQ(ierr); 5311 } 5312 5313 /* recvs and sends of j-array are completed */ 5314 i = nrecvs; 5315 while (i--) { 5316 ierr = MPI_Waitany(nrecvs,rwaits,&jj,&rstatus);CHKERRQ(ierr); 5317 } 5318 if (nsends) {ierr = MPI_Waitall(nsends,swaits,sstatus);CHKERRQ(ierr);} 5319 } else if (scall == MAT_REUSE_MATRIX) { 5320 sstartsj = *startsj_s; 5321 rstartsj = *startsj_r; 5322 bufa = *bufa_ptr; 5323 b_oth = (Mat_SeqAIJ*)(*B_oth)->data; 5324 b_otha = b_oth->a; 5325 } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE, "Matrix P does not posses an object container"); 5326 5327 /* a-array */ 5328 /*---------*/ 5329 /* post receives of a-array */ 5330 for (i=0; i<nrecvs; i++) { 5331 nrows = rstartsj[i+1]-rstartsj[i]; /* length of the msg received */ 5332 ierr = MPI_Irecv(b_otha+rstartsj[i],nrows,MPIU_SCALAR,rprocs[i],tag,comm,rwaits+i);CHKERRQ(ierr); 5333 } 5334 5335 /* pack the outgoing message a-array */ 5336 k = 0; 5337 for (i=0; i<nsends; i++) { 5338 nrows = sstarts[i+1]-sstarts[i]; /* num of block rows */ 5339 bufA = bufa+sstartsj[i]; 5340 for (j=0; j<nrows; j++) { 5341 row = srow[k++] + B->rmap->range[rank]; /* global row idx */ 5342 for (ll=0; ll<sbs; ll++) { 5343 ierr = MatGetRow_MPIAIJ(B,row+ll,&ncols,NULL,&vals);CHKERRQ(ierr); 5344 for (l=0; l<ncols; l++) { 5345 *bufA++ = vals[l]; 5346 } 5347 ierr = MatRestoreRow_MPIAIJ(B,row+ll,&ncols,NULL,&vals);CHKERRQ(ierr); 5348 } 5349 } 5350 ierr = MPI_Isend(bufa+sstartsj[i],sstartsj[i+1]-sstartsj[i],MPIU_SCALAR,sprocs[i],tag,comm,swaits+i);CHKERRQ(ierr); 5351 } 5352 /* recvs and sends of a-array are completed */ 5353 i = nrecvs; 5354 while (i--) { 5355 ierr = MPI_Waitany(nrecvs,rwaits,&jj,&rstatus);CHKERRQ(ierr); 5356 } 5357 if (nsends) {ierr = MPI_Waitall(nsends,swaits,sstatus);CHKERRQ(ierr);} 5358 ierr = PetscFree2(rwaits,swaits);CHKERRQ(ierr); 5359 5360 if (scall == MAT_INITIAL_MATRIX) { 5361 /* put together the new matrix */ 5362 ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,aBn,B->cmap->N,b_othi,b_othj,b_otha,B_oth);CHKERRQ(ierr); 5363 5364 /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */ 5365 /* Since these are PETSc arrays, change flags to free them as necessary. */ 5366 b_oth = (Mat_SeqAIJ*)(*B_oth)->data; 5367 b_oth->free_a = PETSC_TRUE; 5368 b_oth->free_ij = PETSC_TRUE; 5369 b_oth->nonew = 0; 5370 5371 ierr = PetscFree(bufj);CHKERRQ(ierr); 5372 if (!startsj_s || !bufa_ptr) { 5373 ierr = PetscFree2(sstartsj,rstartsj);CHKERRQ(ierr); 5374 ierr = PetscFree(bufa_ptr);CHKERRQ(ierr); 5375 } else { 5376 *startsj_s = sstartsj; 5377 *startsj_r = rstartsj; 5378 *bufa_ptr = bufa; 5379 } 5380 } 5381 ierr = PetscLogEventEnd(MAT_GetBrowsOfAocols,A,B,0,0);CHKERRQ(ierr); 5382 PetscFunctionReturn(0); 5383 } 5384 5385 #undef __FUNCT__ 5386 #define __FUNCT__ "MatGetCommunicationStructs" 5387 /*@C 5388 MatGetCommunicationStructs - Provides access to the communication structures used in matrix-vector multiplication. 5389 5390 Not Collective 5391 5392 Input Parameters: 5393 . A - The matrix in mpiaij format 5394 5395 Output Parameter: 5396 + lvec - The local vector holding off-process values from the argument to a matrix-vector product 5397 . colmap - A map from global column index to local index into lvec 5398 - multScatter - A scatter from the argument of a matrix-vector product to lvec 5399 5400 Level: developer 5401 5402 @*/ 5403 #if defined(PETSC_USE_CTABLE) 5404 PetscErrorCode MatGetCommunicationStructs(Mat A, Vec *lvec, PetscTable *colmap, VecScatter *multScatter) 5405 #else 5406 PetscErrorCode MatGetCommunicationStructs(Mat A, Vec *lvec, PetscInt *colmap[], VecScatter *multScatter) 5407 #endif 5408 { 5409 Mat_MPIAIJ *a; 5410 5411 PetscFunctionBegin; 5412 PetscValidHeaderSpecific(A, MAT_CLASSID, 1); 5413 PetscValidPointer(lvec, 2); 5414 PetscValidPointer(colmap, 3); 5415 PetscValidPointer(multScatter, 4); 5416 a = (Mat_MPIAIJ*) A->data; 5417 if (lvec) *lvec = a->lvec; 5418 if (colmap) *colmap = a->colmap; 5419 if (multScatter) *multScatter = a->Mvctx; 5420 PetscFunctionReturn(0); 5421 } 5422 5423 PETSC_EXTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJCRL(Mat,MatType,MatReuse,Mat*); 5424 PETSC_EXTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJPERM(Mat,MatType,MatReuse,Mat*); 5425 PETSC_EXTERN PetscErrorCode MatConvert_MPIAIJ_MPISBAIJ(Mat,MatType,MatReuse,Mat*); 5426 #if defined(PETSC_HAVE_ELEMENTAL) 5427 PETSC_EXTERN PetscErrorCode MatConvert_MPIAIJ_Elemental(Mat,MatType,MatReuse,Mat*); 5428 #endif 5429 5430 #undef __FUNCT__ 5431 #define __FUNCT__ "MatMatMultNumeric_MPIDense_MPIAIJ" 5432 /* 5433 Computes (B'*A')' since computing B*A directly is untenable 5434 5435 n p p 5436 ( ) ( ) ( ) 5437 m ( A ) * n ( B ) = m ( C ) 5438 ( ) ( ) ( ) 5439 5440 */ 5441 PetscErrorCode MatMatMultNumeric_MPIDense_MPIAIJ(Mat A,Mat B,Mat C) 5442 { 5443 PetscErrorCode ierr; 5444 Mat At,Bt,Ct; 5445 5446 PetscFunctionBegin; 5447 ierr = MatTranspose(A,MAT_INITIAL_MATRIX,&At);CHKERRQ(ierr); 5448 ierr = MatTranspose(B,MAT_INITIAL_MATRIX,&Bt);CHKERRQ(ierr); 5449 ierr = MatMatMult(Bt,At,MAT_INITIAL_MATRIX,1.0,&Ct);CHKERRQ(ierr); 5450 ierr = MatDestroy(&At);CHKERRQ(ierr); 5451 ierr = MatDestroy(&Bt);CHKERRQ(ierr); 5452 ierr = MatTranspose(Ct,MAT_REUSE_MATRIX,&C);CHKERRQ(ierr); 5453 ierr = MatDestroy(&Ct);CHKERRQ(ierr); 5454 PetscFunctionReturn(0); 5455 } 5456 5457 #undef __FUNCT__ 5458 #define __FUNCT__ "MatMatMultSymbolic_MPIDense_MPIAIJ" 5459 PetscErrorCode MatMatMultSymbolic_MPIDense_MPIAIJ(Mat A,Mat B,PetscReal fill,Mat *C) 5460 { 5461 PetscErrorCode ierr; 5462 PetscInt m=A->rmap->n,n=B->cmap->n; 5463 Mat Cmat; 5464 5465 PetscFunctionBegin; 5466 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); 5467 ierr = MatCreate(PetscObjectComm((PetscObject)A),&Cmat);CHKERRQ(ierr); 5468 ierr = MatSetSizes(Cmat,m,n,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 5469 ierr = MatSetBlockSizesFromMats(Cmat,A,B);CHKERRQ(ierr); 5470 ierr = MatSetType(Cmat,MATMPIDENSE);CHKERRQ(ierr); 5471 ierr = MatMPIDenseSetPreallocation(Cmat,NULL);CHKERRQ(ierr); 5472 ierr = MatAssemblyBegin(Cmat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5473 ierr = MatAssemblyEnd(Cmat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5474 5475 Cmat->ops->matmultnumeric = MatMatMultNumeric_MPIDense_MPIAIJ; 5476 5477 *C = Cmat; 5478 PetscFunctionReturn(0); 5479 } 5480 5481 /* ----------------------------------------------------------------*/ 5482 #undef __FUNCT__ 5483 #define __FUNCT__ "MatMatMult_MPIDense_MPIAIJ" 5484 PetscErrorCode MatMatMult_MPIDense_MPIAIJ(Mat A,Mat B,MatReuse scall,PetscReal fill,Mat *C) 5485 { 5486 PetscErrorCode ierr; 5487 5488 PetscFunctionBegin; 5489 if (scall == MAT_INITIAL_MATRIX) { 5490 ierr = PetscLogEventBegin(MAT_MatMultSymbolic,A,B,0,0);CHKERRQ(ierr); 5491 ierr = MatMatMultSymbolic_MPIDense_MPIAIJ(A,B,fill,C);CHKERRQ(ierr); 5492 ierr = PetscLogEventEnd(MAT_MatMultSymbolic,A,B,0,0);CHKERRQ(ierr); 5493 } 5494 ierr = PetscLogEventBegin(MAT_MatMultNumeric,A,B,0,0);CHKERRQ(ierr); 5495 ierr = MatMatMultNumeric_MPIDense_MPIAIJ(A,B,*C);CHKERRQ(ierr); 5496 ierr = PetscLogEventEnd(MAT_MatMultNumeric,A,B,0,0);CHKERRQ(ierr); 5497 PetscFunctionReturn(0); 5498 } 5499 5500 #if defined(PETSC_HAVE_MUMPS) 5501 PETSC_EXTERN PetscErrorCode MatGetFactor_aij_mumps(Mat,MatFactorType,Mat*); 5502 #endif 5503 #if defined(PETSC_HAVE_PASTIX) 5504 PETSC_EXTERN PetscErrorCode MatGetFactor_mpiaij_pastix(Mat,MatFactorType,Mat*); 5505 #endif 5506 #if defined(PETSC_HAVE_SUPERLU_DIST) 5507 PETSC_EXTERN PetscErrorCode MatGetFactor_mpiaij_superlu_dist(Mat,MatFactorType,Mat*); 5508 #endif 5509 #if defined(PETSC_HAVE_CLIQUE) 5510 PETSC_EXTERN PetscErrorCode MatGetFactor_aij_clique(Mat,MatFactorType,Mat*); 5511 #endif 5512 5513 /*MC 5514 MATMPIAIJ - MATMPIAIJ = "mpiaij" - A matrix type to be used for parallel sparse matrices. 5515 5516 Options Database Keys: 5517 . -mat_type mpiaij - sets the matrix type to "mpiaij" during a call to MatSetFromOptions() 5518 5519 Level: beginner 5520 5521 .seealso: MatCreateAIJ() 5522 M*/ 5523 5524 #undef __FUNCT__ 5525 #define __FUNCT__ "MatCreate_MPIAIJ" 5526 PETSC_EXTERN PetscErrorCode MatCreate_MPIAIJ(Mat B) 5527 { 5528 Mat_MPIAIJ *b; 5529 PetscErrorCode ierr; 5530 PetscMPIInt size; 5531 5532 PetscFunctionBegin; 5533 ierr = MPI_Comm_size(PetscObjectComm((PetscObject)B),&size);CHKERRQ(ierr); 5534 5535 ierr = PetscNewLog(B,&b);CHKERRQ(ierr); 5536 B->data = (void*)b; 5537 ierr = PetscMemcpy(B->ops,&MatOps_Values,sizeof(struct _MatOps));CHKERRQ(ierr); 5538 B->assembled = PETSC_FALSE; 5539 B->insertmode = NOT_SET_VALUES; 5540 b->size = size; 5541 5542 ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)B),&b->rank);CHKERRQ(ierr); 5543 5544 /* build cache for off array entries formed */ 5545 ierr = MatStashCreate_Private(PetscObjectComm((PetscObject)B),1,&B->stash);CHKERRQ(ierr); 5546 5547 b->donotstash = PETSC_FALSE; 5548 b->colmap = 0; 5549 b->garray = 0; 5550 b->roworiented = PETSC_TRUE; 5551 5552 /* stuff used for matrix vector multiply */ 5553 b->lvec = NULL; 5554 b->Mvctx = NULL; 5555 5556 /* stuff for MatGetRow() */ 5557 b->rowindices = 0; 5558 b->rowvalues = 0; 5559 b->getrowactive = PETSC_FALSE; 5560 5561 /* flexible pointer used in CUSP/CUSPARSE classes */ 5562 b->spptr = NULL; 5563 5564 #if defined(PETSC_HAVE_MUMPS) 5565 ierr = PetscObjectComposeFunction((PetscObject)B,"MatGetFactor_mumps_C",MatGetFactor_aij_mumps);CHKERRQ(ierr); 5566 #endif 5567 #if defined(PETSC_HAVE_PASTIX) 5568 ierr = PetscObjectComposeFunction((PetscObject)B,"MatGetFactor_pastix_C",MatGetFactor_mpiaij_pastix);CHKERRQ(ierr); 5569 #endif 5570 #if defined(PETSC_HAVE_SUPERLU_DIST) 5571 ierr = PetscObjectComposeFunction((PetscObject)B,"MatGetFactor_superlu_dist_C",MatGetFactor_mpiaij_superlu_dist);CHKERRQ(ierr); 5572 #endif 5573 #if defined(PETSC_HAVE_CLIQUE) 5574 ierr = PetscObjectComposeFunction((PetscObject)B,"MatGetFactor_clique_C",MatGetFactor_aij_clique);CHKERRQ(ierr); 5575 #endif 5576 ierr = PetscObjectComposeFunction((PetscObject)B,"MatStoreValues_C",MatStoreValues_MPIAIJ);CHKERRQ(ierr); 5577 ierr = PetscObjectComposeFunction((PetscObject)B,"MatRetrieveValues_C",MatRetrieveValues_MPIAIJ);CHKERRQ(ierr); 5578 ierr = PetscObjectComposeFunction((PetscObject)B,"MatGetDiagonalBlock_C",MatGetDiagonalBlock_MPIAIJ);CHKERRQ(ierr); 5579 ierr = PetscObjectComposeFunction((PetscObject)B,"MatIsTranspose_C",MatIsTranspose_MPIAIJ);CHKERRQ(ierr); 5580 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMPIAIJSetPreallocation_C",MatMPIAIJSetPreallocation_MPIAIJ);CHKERRQ(ierr); 5581 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMPIAIJSetPreallocationCSR_C",MatMPIAIJSetPreallocationCSR_MPIAIJ);CHKERRQ(ierr); 5582 ierr = PetscObjectComposeFunction((PetscObject)B,"MatDiagonalScaleLocal_C",MatDiagonalScaleLocal_MPIAIJ);CHKERRQ(ierr); 5583 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_mpiaijperm_C",MatConvert_MPIAIJ_MPIAIJPERM);CHKERRQ(ierr); 5584 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_mpiaijcrl_C",MatConvert_MPIAIJ_MPIAIJCRL);CHKERRQ(ierr); 5585 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_mpisbaij_C",MatConvert_MPIAIJ_MPISBAIJ);CHKERRQ(ierr); 5586 #if defined(PETSC_HAVE_ELEMENTAL) 5587 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_elemental_C",MatConvert_MPIAIJ_Elemental);CHKERRQ(ierr); 5588 #endif 5589 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMatMult_mpidense_mpiaij_C",MatMatMult_MPIDense_MPIAIJ);CHKERRQ(ierr); 5590 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMatMultSymbolic_mpidense_mpiaij_C",MatMatMultSymbolic_MPIDense_MPIAIJ);CHKERRQ(ierr); 5591 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMatMultNumeric_mpidense_mpiaij_C",MatMatMultNumeric_MPIDense_MPIAIJ);CHKERRQ(ierr); 5592 ierr = PetscObjectChangeTypeName((PetscObject)B,MATMPIAIJ);CHKERRQ(ierr); 5593 PetscFunctionReturn(0); 5594 } 5595 5596 #undef __FUNCT__ 5597 #define __FUNCT__ "MatCreateMPIAIJWithSplitArrays" 5598 /*@ 5599 MatCreateMPIAIJWithSplitArrays - creates a MPI AIJ matrix using arrays that contain the "diagonal" 5600 and "off-diagonal" part of the matrix in CSR format. 5601 5602 Collective on MPI_Comm 5603 5604 Input Parameters: 5605 + comm - MPI communicator 5606 . m - number of local rows (Cannot be PETSC_DECIDE) 5607 . n - This value should be the same as the local size used in creating the 5608 x vector for the matrix-vector product y = Ax. (or PETSC_DECIDE to have 5609 calculated if N is given) For square matrices n is almost always m. 5610 . M - number of global rows (or PETSC_DETERMINE to have calculated if m is given) 5611 . N - number of global columns (or PETSC_DETERMINE to have calculated if n is given) 5612 . i - row indices for "diagonal" portion of matrix 5613 . j - column indices 5614 . a - matrix values 5615 . oi - row indices for "off-diagonal" portion of matrix 5616 . oj - column indices 5617 - oa - matrix values 5618 5619 Output Parameter: 5620 . mat - the matrix 5621 5622 Level: advanced 5623 5624 Notes: 5625 The i, j, and a arrays ARE NOT copied by this routine into the internal format used by PETSc. The user 5626 must free the arrays once the matrix has been destroyed and not before. 5627 5628 The i and j indices are 0 based 5629 5630 See MatCreateAIJ() for the definition of "diagonal" and "off-diagonal" portion of the matrix 5631 5632 This sets local rows and cannot be used to set off-processor values. 5633 5634 Use of this routine is discouraged because it is inflexible and cumbersome to use. It is extremely rare that a 5635 legacy application natively assembles into exactly this split format. The code to do so is nontrivial and does 5636 not easily support in-place reassembly. It is recommended to use MatSetValues() (or a variant thereof) because 5637 the resulting assembly is easier to implement, will work with any matrix format, and the user does not have to 5638 keep track of the underlying array. Use MatSetOption(A,MAT_IGNORE_OFF_PROC_ENTRIES,PETSC_TRUE) to disable all 5639 communication if it is known that only local entries will be set. 5640 5641 .keywords: matrix, aij, compressed row, sparse, parallel 5642 5643 .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatMPIAIJSetPreallocationCSR(), 5644 MPIAIJ, MatCreateAIJ(), MatCreateMPIAIJWithArrays() 5645 @*/ 5646 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) 5647 { 5648 PetscErrorCode ierr; 5649 Mat_MPIAIJ *maij; 5650 5651 PetscFunctionBegin; 5652 if (m < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"local number of rows (m) cannot be PETSC_DECIDE, or negative"); 5653 if (i[0]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"i (row indices) must start with 0"); 5654 if (oi[0]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"oi (row indices) must start with 0"); 5655 ierr = MatCreate(comm,mat);CHKERRQ(ierr); 5656 ierr = MatSetSizes(*mat,m,n,M,N);CHKERRQ(ierr); 5657 ierr = MatSetType(*mat,MATMPIAIJ);CHKERRQ(ierr); 5658 maij = (Mat_MPIAIJ*) (*mat)->data; 5659 5660 (*mat)->preallocated = PETSC_TRUE; 5661 5662 ierr = PetscLayoutSetUp((*mat)->rmap);CHKERRQ(ierr); 5663 ierr = PetscLayoutSetUp((*mat)->cmap);CHKERRQ(ierr); 5664 5665 ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,m,n,i,j,a,&maij->A);CHKERRQ(ierr); 5666 ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,m,(*mat)->cmap->N,oi,oj,oa,&maij->B);CHKERRQ(ierr); 5667 5668 ierr = MatAssemblyBegin(maij->A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5669 ierr = MatAssemblyEnd(maij->A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5670 ierr = MatAssemblyBegin(maij->B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5671 ierr = MatAssemblyEnd(maij->B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5672 5673 ierr = MatAssemblyBegin(*mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5674 ierr = MatAssemblyEnd(*mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5675 ierr = MatSetOption(*mat,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE);CHKERRQ(ierr); 5676 PetscFunctionReturn(0); 5677 } 5678 5679 /* 5680 Special version for direct calls from Fortran 5681 */ 5682 #include <petsc-private/fortranimpl.h> 5683 5684 #if defined(PETSC_HAVE_FORTRAN_CAPS) 5685 #define matsetvaluesmpiaij_ MATSETVALUESMPIAIJ 5686 #elif !defined(PETSC_HAVE_FORTRAN_UNDERSCORE) 5687 #define matsetvaluesmpiaij_ matsetvaluesmpiaij 5688 #endif 5689 5690 /* Change these macros so can be used in void function */ 5691 #undef CHKERRQ 5692 #define CHKERRQ(ierr) CHKERRABORT(PETSC_COMM_WORLD,ierr) 5693 #undef SETERRQ2 5694 #define SETERRQ2(comm,ierr,b,c,d) CHKERRABORT(comm,ierr) 5695 #undef SETERRQ3 5696 #define SETERRQ3(comm,ierr,b,c,d,e) CHKERRABORT(comm,ierr) 5697 #undef SETERRQ 5698 #define SETERRQ(c,ierr,b) CHKERRABORT(c,ierr) 5699 5700 #undef __FUNCT__ 5701 #define __FUNCT__ "matsetvaluesmpiaij_" 5702 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) 5703 { 5704 Mat mat = *mmat; 5705 PetscInt m = *mm, n = *mn; 5706 InsertMode addv = *maddv; 5707 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 5708 PetscScalar value; 5709 PetscErrorCode ierr; 5710 5711 MatCheckPreallocated(mat,1); 5712 if (mat->insertmode == NOT_SET_VALUES) mat->insertmode = addv; 5713 5714 #if defined(PETSC_USE_DEBUG) 5715 else if (mat->insertmode != addv) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Cannot mix add values and insert values"); 5716 #endif 5717 { 5718 PetscInt i,j,rstart = mat->rmap->rstart,rend = mat->rmap->rend; 5719 PetscInt cstart = mat->cmap->rstart,cend = mat->cmap->rend,row,col; 5720 PetscBool roworiented = aij->roworiented; 5721 5722 /* Some Variables required in the macro */ 5723 Mat A = aij->A; 5724 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 5725 PetscInt *aimax = a->imax,*ai = a->i,*ailen = a->ilen,*aj = a->j; 5726 MatScalar *aa = a->a; 5727 PetscBool ignorezeroentries = (((a->ignorezeroentries)&&(addv==ADD_VALUES)) ? PETSC_TRUE : PETSC_FALSE); 5728 Mat B = aij->B; 5729 Mat_SeqAIJ *b = (Mat_SeqAIJ*)B->data; 5730 PetscInt *bimax = b->imax,*bi = b->i,*bilen = b->ilen,*bj = b->j,bm = aij->B->rmap->n,am = aij->A->rmap->n; 5731 MatScalar *ba = b->a; 5732 5733 PetscInt *rp1,*rp2,ii,nrow1,nrow2,_i,rmax1,rmax2,N,low1,high1,low2,high2,t,lastcol1,lastcol2; 5734 PetscInt nonew = a->nonew; 5735 MatScalar *ap1,*ap2; 5736 5737 PetscFunctionBegin; 5738 for (i=0; i<m; i++) { 5739 if (im[i] < 0) continue; 5740 #if defined(PETSC_USE_DEBUG) 5741 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); 5742 #endif 5743 if (im[i] >= rstart && im[i] < rend) { 5744 row = im[i] - rstart; 5745 lastcol1 = -1; 5746 rp1 = aj + ai[row]; 5747 ap1 = aa + ai[row]; 5748 rmax1 = aimax[row]; 5749 nrow1 = ailen[row]; 5750 low1 = 0; 5751 high1 = nrow1; 5752 lastcol2 = -1; 5753 rp2 = bj + bi[row]; 5754 ap2 = ba + bi[row]; 5755 rmax2 = bimax[row]; 5756 nrow2 = bilen[row]; 5757 low2 = 0; 5758 high2 = nrow2; 5759 5760 for (j=0; j<n; j++) { 5761 if (roworiented) value = v[i*n+j]; 5762 else value = v[i+j*m]; 5763 if (ignorezeroentries && value == 0.0 && (addv == ADD_VALUES)) continue; 5764 if (in[j] >= cstart && in[j] < cend) { 5765 col = in[j] - cstart; 5766 MatSetValues_SeqAIJ_A_Private(row,col,value,addv); 5767 } else if (in[j] < 0) continue; 5768 #if defined(PETSC_USE_DEBUG) 5769 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); 5770 #endif 5771 else { 5772 if (mat->was_assembled) { 5773 if (!aij->colmap) { 5774 ierr = MatCreateColmap_MPIAIJ_Private(mat);CHKERRQ(ierr); 5775 } 5776 #if defined(PETSC_USE_CTABLE) 5777 ierr = PetscTableFind(aij->colmap,in[j]+1,&col);CHKERRQ(ierr); 5778 col--; 5779 #else 5780 col = aij->colmap[in[j]] - 1; 5781 #endif 5782 if (col < 0 && !((Mat_SeqAIJ*)(aij->A->data))->nonew) { 5783 ierr = MatDisAssemble_MPIAIJ(mat);CHKERRQ(ierr); 5784 col = in[j]; 5785 /* Reinitialize the variables required by MatSetValues_SeqAIJ_B_Private() */ 5786 B = aij->B; 5787 b = (Mat_SeqAIJ*)B->data; 5788 bimax = b->imax; bi = b->i; bilen = b->ilen; bj = b->j; 5789 rp2 = bj + bi[row]; 5790 ap2 = ba + bi[row]; 5791 rmax2 = bimax[row]; 5792 nrow2 = bilen[row]; 5793 low2 = 0; 5794 high2 = nrow2; 5795 bm = aij->B->rmap->n; 5796 ba = b->a; 5797 } 5798 } else col = in[j]; 5799 MatSetValues_SeqAIJ_B_Private(row,col,value,addv); 5800 } 5801 } 5802 } else if (!aij->donotstash) { 5803 if (roworiented) { 5804 ierr = MatStashValuesRow_Private(&mat->stash,im[i],n,in,v+i*n,(PetscBool)(ignorezeroentries && (addv == ADD_VALUES)));CHKERRQ(ierr); 5805 } else { 5806 ierr = MatStashValuesCol_Private(&mat->stash,im[i],n,in,v+i,m,(PetscBool)(ignorezeroentries && (addv == ADD_VALUES)));CHKERRQ(ierr); 5807 } 5808 } 5809 } 5810 } 5811 PetscFunctionReturnVoid(); 5812 } 5813 5814