1 2 #include <../src/mat/impls/aij/mpi/mpiaij.h> /*I "petscmat.h" I*/ 3 #include <petsc-private/vecimpl.h> 4 #include <petscblaslapack.h> 5 #include <petscsf.h> 6 7 /*MC 8 MATAIJ - MATAIJ = "aij" - A matrix type to be used for sparse matrices. 9 10 This matrix type is identical to MATSEQAIJ when constructed with a single process communicator, 11 and MATMPIAIJ otherwise. As a result, for single process communicators, 12 MatSeqAIJSetPreallocation is supported, and similarly MatMPIAIJSetPreallocation is supported 13 for communicators controlling multiple processes. It is recommended that you call both of 14 the above preallocation routines for simplicity. 15 16 Options Database Keys: 17 . -mat_type aij - sets the matrix type to "aij" during a call to MatSetFromOptions() 18 19 Developer Notes: Subclasses include MATAIJCUSP, MATAIJCUSPARSE, MATAIJPERM, MATAIJCRL, and also automatically switches over to use inodes when 20 enough exist. 21 22 Level: beginner 23 24 .seealso: MatCreateAIJ(), MatCreateSeqAIJ(), MATSEQAIJ,MATMPIAIJ 25 M*/ 26 27 /*MC 28 MATAIJCRL - MATAIJCRL = "aijcrl" - A matrix type to be used for sparse matrices. 29 30 This matrix type is identical to MATSEQAIJCRL when constructed with a single process communicator, 31 and MATMPIAIJCRL otherwise. As a result, for single process communicators, 32 MatSeqAIJSetPreallocation() is supported, and similarly MatMPIAIJSetPreallocation() is supported 33 for communicators controlling multiple processes. It is recommended that you call both of 34 the above preallocation routines for simplicity. 35 36 Options Database Keys: 37 . -mat_type aijcrl - sets the matrix type to "aijcrl" during a call to MatSetFromOptions() 38 39 Level: beginner 40 41 .seealso: MatCreateMPIAIJCRL,MATSEQAIJCRL,MATMPIAIJCRL, MATSEQAIJCRL, MATMPIAIJCRL 42 M*/ 43 44 #undef __FUNCT__ 45 #define __FUNCT__ "MatFindNonzeroRows_MPIAIJ" 46 PetscErrorCode MatFindNonzeroRows_MPIAIJ(Mat M,IS *keptrows) 47 { 48 PetscErrorCode ierr; 49 Mat_MPIAIJ *mat = (Mat_MPIAIJ*)M->data; 50 Mat_SeqAIJ *a = (Mat_SeqAIJ*)mat->A->data; 51 Mat_SeqAIJ *b = (Mat_SeqAIJ*)mat->B->data; 52 const PetscInt *ia,*ib; 53 const MatScalar *aa,*bb; 54 PetscInt na,nb,i,j,*rows,cnt=0,n0rows; 55 PetscInt m = M->rmap->n,rstart = M->rmap->rstart; 56 57 PetscFunctionBegin; 58 *keptrows = 0; 59 ia = a->i; 60 ib = b->i; 61 for (i=0; i<m; i++) { 62 na = ia[i+1] - ia[i]; 63 nb = ib[i+1] - ib[i]; 64 if (!na && !nb) { 65 cnt++; 66 goto ok1; 67 } 68 aa = a->a + ia[i]; 69 for (j=0; j<na; j++) { 70 if (aa[j] != 0.0) goto ok1; 71 } 72 bb = b->a + ib[i]; 73 for (j=0; j <nb; j++) { 74 if (bb[j] != 0.0) goto ok1; 75 } 76 cnt++; 77 ok1:; 78 } 79 ierr = MPI_Allreduce(&cnt,&n0rows,1,MPIU_INT,MPIU_SUM,PetscObjectComm((PetscObject)M));CHKERRQ(ierr); 80 if (!n0rows) PetscFunctionReturn(0); 81 ierr = PetscMalloc1((M->rmap->n-cnt),&rows);CHKERRQ(ierr); 82 cnt = 0; 83 for (i=0; i<m; i++) { 84 na = ia[i+1] - ia[i]; 85 nb = ib[i+1] - ib[i]; 86 if (!na && !nb) continue; 87 aa = a->a + ia[i]; 88 for (j=0; j<na;j++) { 89 if (aa[j] != 0.0) { 90 rows[cnt++] = rstart + i; 91 goto ok2; 92 } 93 } 94 bb = b->a + ib[i]; 95 for (j=0; j<nb; j++) { 96 if (bb[j] != 0.0) { 97 rows[cnt++] = rstart + i; 98 goto ok2; 99 } 100 } 101 ok2:; 102 } 103 ierr = ISCreateGeneral(PetscObjectComm((PetscObject)M),cnt,rows,PETSC_OWN_POINTER,keptrows);CHKERRQ(ierr); 104 PetscFunctionReturn(0); 105 } 106 107 #undef __FUNCT__ 108 #define __FUNCT__ "MatDiagonalSet_MPIAIJ" 109 PetscErrorCode MatDiagonalSet_MPIAIJ(Mat Y,Vec D,InsertMode is) 110 { 111 PetscErrorCode ierr; 112 Mat_MPIAIJ *aij = (Mat_MPIAIJ*) Y->data; 113 114 PetscFunctionBegin; 115 if (Y->assembled && Y->rmap->rstart == Y->cmap->rstart && Y->rmap->rend == Y->cmap->rend) { 116 ierr = MatDiagonalSet(aij->A,D,is);CHKERRQ(ierr); 117 } else { 118 ierr = MatDiagonalSet_Default(Y,D,is);CHKERRQ(ierr); 119 } 120 PetscFunctionReturn(0); 121 } 122 123 124 #undef __FUNCT__ 125 #define __FUNCT__ "MatFindZeroDiagonals_MPIAIJ" 126 PetscErrorCode MatFindZeroDiagonals_MPIAIJ(Mat M,IS *zrows) 127 { 128 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)M->data; 129 PetscErrorCode ierr; 130 PetscInt i,rstart,nrows,*rows; 131 132 PetscFunctionBegin; 133 *zrows = NULL; 134 ierr = MatFindZeroDiagonals_SeqAIJ_Private(aij->A,&nrows,&rows);CHKERRQ(ierr); 135 ierr = MatGetOwnershipRange(M,&rstart,NULL);CHKERRQ(ierr); 136 for (i=0; i<nrows; i++) rows[i] += rstart; 137 ierr = ISCreateGeneral(PetscObjectComm((PetscObject)M),nrows,rows,PETSC_OWN_POINTER,zrows);CHKERRQ(ierr); 138 PetscFunctionReturn(0); 139 } 140 141 #undef __FUNCT__ 142 #define __FUNCT__ "MatGetColumnNorms_MPIAIJ" 143 PetscErrorCode MatGetColumnNorms_MPIAIJ(Mat A,NormType type,PetscReal *norms) 144 { 145 PetscErrorCode ierr; 146 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)A->data; 147 PetscInt i,n,*garray = aij->garray; 148 Mat_SeqAIJ *a_aij = (Mat_SeqAIJ*) aij->A->data; 149 Mat_SeqAIJ *b_aij = (Mat_SeqAIJ*) aij->B->data; 150 PetscReal *work; 151 152 PetscFunctionBegin; 153 ierr = MatGetSize(A,NULL,&n);CHKERRQ(ierr); 154 ierr = PetscCalloc1(n,&work);CHKERRQ(ierr); 155 if (type == NORM_2) { 156 for (i=0; i<a_aij->i[aij->A->rmap->n]; i++) { 157 work[A->cmap->rstart + a_aij->j[i]] += PetscAbsScalar(a_aij->a[i]*a_aij->a[i]); 158 } 159 for (i=0; i<b_aij->i[aij->B->rmap->n]; i++) { 160 work[garray[b_aij->j[i]]] += PetscAbsScalar(b_aij->a[i]*b_aij->a[i]); 161 } 162 } else if (type == NORM_1) { 163 for (i=0; i<a_aij->i[aij->A->rmap->n]; i++) { 164 work[A->cmap->rstart + a_aij->j[i]] += PetscAbsScalar(a_aij->a[i]); 165 } 166 for (i=0; i<b_aij->i[aij->B->rmap->n]; i++) { 167 work[garray[b_aij->j[i]]] += PetscAbsScalar(b_aij->a[i]); 168 } 169 } else if (type == NORM_INFINITY) { 170 for (i=0; i<a_aij->i[aij->A->rmap->n]; i++) { 171 work[A->cmap->rstart + a_aij->j[i]] = PetscMax(PetscAbsScalar(a_aij->a[i]), work[A->cmap->rstart + a_aij->j[i]]); 172 } 173 for (i=0; i<b_aij->i[aij->B->rmap->n]; i++) { 174 work[garray[b_aij->j[i]]] = PetscMax(PetscAbsScalar(b_aij->a[i]),work[garray[b_aij->j[i]]]); 175 } 176 177 } else SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_WRONG,"Unknown NormType"); 178 if (type == NORM_INFINITY) { 179 ierr = MPI_Allreduce(work,norms,n,MPIU_REAL,MPIU_MAX,PetscObjectComm((PetscObject)A));CHKERRQ(ierr); 180 } else { 181 ierr = MPI_Allreduce(work,norms,n,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)A));CHKERRQ(ierr); 182 } 183 ierr = PetscFree(work);CHKERRQ(ierr); 184 if (type == NORM_2) { 185 for (i=0; i<n; i++) norms[i] = PetscSqrtReal(norms[i]); 186 } 187 PetscFunctionReturn(0); 188 } 189 190 #undef __FUNCT__ 191 #define __FUNCT__ "MatDistribute_MPIAIJ" 192 /* 193 Distributes a SeqAIJ matrix across a set of processes. Code stolen from 194 MatLoad_MPIAIJ(). Horrible lack of reuse. Should be a routine for each matrix type. 195 196 Only for square matrices 197 198 Used by a preconditioner, hence PETSC_EXTERN 199 */ 200 PETSC_EXTERN PetscErrorCode MatDistribute_MPIAIJ(MPI_Comm comm,Mat gmat,PetscInt m,MatReuse reuse,Mat *inmat) 201 { 202 PetscMPIInt rank,size; 203 PetscInt *rowners,*dlens,*olens,i,rstart,rend,j,jj,nz = 0,*gmataj,cnt,row,*ld,bses[2]; 204 PetscErrorCode ierr; 205 Mat mat; 206 Mat_SeqAIJ *gmata; 207 PetscMPIInt tag; 208 MPI_Status status; 209 PetscBool aij; 210 MatScalar *gmataa,*ao,*ad,*gmataarestore=0; 211 212 PetscFunctionBegin; 213 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 214 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 215 if (!rank) { 216 ierr = PetscObjectTypeCompare((PetscObject)gmat,MATSEQAIJ,&aij);CHKERRQ(ierr); 217 if (!aij) SETERRQ1(PetscObjectComm((PetscObject)gmat),PETSC_ERR_SUP,"Currently no support for input matrix of type %s\n",((PetscObject)gmat)->type_name); 218 } 219 if (reuse == MAT_INITIAL_MATRIX) { 220 ierr = MatCreate(comm,&mat);CHKERRQ(ierr); 221 ierr = MatSetSizes(mat,m,m,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 222 ierr = MatGetBlockSizes(gmat,&bses[0],&bses[1]);CHKERRQ(ierr); 223 ierr = MPI_Bcast(bses,2,MPIU_INT,0,comm);CHKERRQ(ierr); 224 ierr = MatSetBlockSizes(mat,bses[0],bses[1]);CHKERRQ(ierr); 225 ierr = MatSetType(mat,MATAIJ);CHKERRQ(ierr); 226 ierr = PetscMalloc1((size+1),&rowners);CHKERRQ(ierr); 227 ierr = PetscMalloc2(m,&dlens,m,&olens);CHKERRQ(ierr); 228 ierr = MPI_Allgather(&m,1,MPIU_INT,rowners+1,1,MPIU_INT,comm);CHKERRQ(ierr); 229 230 rowners[0] = 0; 231 for (i=2; i<=size; i++) rowners[i] += rowners[i-1]; 232 rstart = rowners[rank]; 233 rend = rowners[rank+1]; 234 ierr = PetscObjectGetNewTag((PetscObject)mat,&tag);CHKERRQ(ierr); 235 if (!rank) { 236 gmata = (Mat_SeqAIJ*) gmat->data; 237 /* send row lengths to all processors */ 238 for (i=0; i<m; i++) dlens[i] = gmata->ilen[i]; 239 for (i=1; i<size; i++) { 240 ierr = MPI_Send(gmata->ilen + rowners[i],rowners[i+1]-rowners[i],MPIU_INT,i,tag,comm);CHKERRQ(ierr); 241 } 242 /* determine number diagonal and off-diagonal counts */ 243 ierr = PetscMemzero(olens,m*sizeof(PetscInt));CHKERRQ(ierr); 244 ierr = PetscCalloc1(m,&ld);CHKERRQ(ierr); 245 jj = 0; 246 for (i=0; i<m; i++) { 247 for (j=0; j<dlens[i]; j++) { 248 if (gmata->j[jj] < rstart) ld[i]++; 249 if (gmata->j[jj] < rstart || gmata->j[jj] >= rend) olens[i]++; 250 jj++; 251 } 252 } 253 /* send column indices to other processes */ 254 for (i=1; i<size; i++) { 255 nz = gmata->i[rowners[i+1]]-gmata->i[rowners[i]]; 256 ierr = MPI_Send(&nz,1,MPIU_INT,i,tag,comm);CHKERRQ(ierr); 257 ierr = MPI_Send(gmata->j + gmata->i[rowners[i]],nz,MPIU_INT,i,tag,comm);CHKERRQ(ierr); 258 } 259 260 /* send numerical values to other processes */ 261 for (i=1; i<size; i++) { 262 nz = gmata->i[rowners[i+1]]-gmata->i[rowners[i]]; 263 ierr = MPI_Send(gmata->a + gmata->i[rowners[i]],nz,MPIU_SCALAR,i,tag,comm);CHKERRQ(ierr); 264 } 265 gmataa = gmata->a; 266 gmataj = gmata->j; 267 268 } else { 269 /* receive row lengths */ 270 ierr = MPI_Recv(dlens,m,MPIU_INT,0,tag,comm,&status);CHKERRQ(ierr); 271 /* receive column indices */ 272 ierr = MPI_Recv(&nz,1,MPIU_INT,0,tag,comm,&status);CHKERRQ(ierr); 273 ierr = PetscMalloc2(nz,&gmataa,nz,&gmataj);CHKERRQ(ierr); 274 ierr = MPI_Recv(gmataj,nz,MPIU_INT,0,tag,comm,&status);CHKERRQ(ierr); 275 /* determine number diagonal and off-diagonal counts */ 276 ierr = PetscMemzero(olens,m*sizeof(PetscInt));CHKERRQ(ierr); 277 ierr = PetscCalloc1(m,&ld);CHKERRQ(ierr); 278 jj = 0; 279 for (i=0; i<m; i++) { 280 for (j=0; j<dlens[i]; j++) { 281 if (gmataj[jj] < rstart) ld[i]++; 282 if (gmataj[jj] < rstart || gmataj[jj] >= rend) olens[i]++; 283 jj++; 284 } 285 } 286 /* receive numerical values */ 287 ierr = PetscMemzero(gmataa,nz*sizeof(PetscScalar));CHKERRQ(ierr); 288 ierr = MPI_Recv(gmataa,nz,MPIU_SCALAR,0,tag,comm,&status);CHKERRQ(ierr); 289 } 290 /* set preallocation */ 291 for (i=0; i<m; i++) { 292 dlens[i] -= olens[i]; 293 } 294 ierr = MatSeqAIJSetPreallocation(mat,0,dlens);CHKERRQ(ierr); 295 ierr = MatMPIAIJSetPreallocation(mat,0,dlens,0,olens);CHKERRQ(ierr); 296 297 for (i=0; i<m; i++) { 298 dlens[i] += olens[i]; 299 } 300 cnt = 0; 301 for (i=0; i<m; i++) { 302 row = rstart + i; 303 ierr = MatSetValues(mat,1,&row,dlens[i],gmataj+cnt,gmataa+cnt,INSERT_VALUES);CHKERRQ(ierr); 304 cnt += dlens[i]; 305 } 306 if (rank) { 307 ierr = PetscFree2(gmataa,gmataj);CHKERRQ(ierr); 308 } 309 ierr = PetscFree2(dlens,olens);CHKERRQ(ierr); 310 ierr = PetscFree(rowners);CHKERRQ(ierr); 311 312 ((Mat_MPIAIJ*)(mat->data))->ld = ld; 313 314 *inmat = mat; 315 } else { /* column indices are already set; only need to move over numerical values from process 0 */ 316 Mat_SeqAIJ *Ad = (Mat_SeqAIJ*)((Mat_MPIAIJ*)((*inmat)->data))->A->data; 317 Mat_SeqAIJ *Ao = (Mat_SeqAIJ*)((Mat_MPIAIJ*)((*inmat)->data))->B->data; 318 mat = *inmat; 319 ierr = PetscObjectGetNewTag((PetscObject)mat,&tag);CHKERRQ(ierr); 320 if (!rank) { 321 /* send numerical values to other processes */ 322 gmata = (Mat_SeqAIJ*) gmat->data; 323 ierr = MatGetOwnershipRanges(mat,(const PetscInt**)&rowners);CHKERRQ(ierr); 324 gmataa = gmata->a; 325 for (i=1; i<size; i++) { 326 nz = gmata->i[rowners[i+1]]-gmata->i[rowners[i]]; 327 ierr = MPI_Send(gmataa + gmata->i[rowners[i]],nz,MPIU_SCALAR,i,tag,comm);CHKERRQ(ierr); 328 } 329 nz = gmata->i[rowners[1]]-gmata->i[rowners[0]]; 330 } else { 331 /* receive numerical values from process 0*/ 332 nz = Ad->nz + Ao->nz; 333 ierr = PetscMalloc1(nz,&gmataa);CHKERRQ(ierr); gmataarestore = gmataa; 334 ierr = MPI_Recv(gmataa,nz,MPIU_SCALAR,0,tag,comm,&status);CHKERRQ(ierr); 335 } 336 /* transfer numerical values into the diagonal A and off diagonal B parts of mat */ 337 ld = ((Mat_MPIAIJ*)(mat->data))->ld; 338 ad = Ad->a; 339 ao = Ao->a; 340 if (mat->rmap->n) { 341 i = 0; 342 nz = ld[i]; ierr = PetscMemcpy(ao,gmataa,nz*sizeof(PetscScalar));CHKERRQ(ierr); ao += nz; gmataa += nz; 343 nz = Ad->i[i+1] - Ad->i[i]; ierr = PetscMemcpy(ad,gmataa,nz*sizeof(PetscScalar));CHKERRQ(ierr); ad += nz; gmataa += nz; 344 } 345 for (i=1; i<mat->rmap->n; i++) { 346 nz = Ao->i[i] - Ao->i[i-1] - ld[i-1] + ld[i]; ierr = PetscMemcpy(ao,gmataa,nz*sizeof(PetscScalar));CHKERRQ(ierr); ao += nz; gmataa += nz; 347 nz = Ad->i[i+1] - Ad->i[i]; ierr = PetscMemcpy(ad,gmataa,nz*sizeof(PetscScalar));CHKERRQ(ierr); ad += nz; gmataa += nz; 348 } 349 i--; 350 if (mat->rmap->n) { 351 nz = Ao->i[i+1] - Ao->i[i] - ld[i]; ierr = PetscMemcpy(ao,gmataa,nz*sizeof(PetscScalar));CHKERRQ(ierr); 352 } 353 if (rank) { 354 ierr = PetscFree(gmataarestore);CHKERRQ(ierr); 355 } 356 } 357 ierr = MatAssemblyBegin(mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 358 ierr = MatAssemblyEnd(mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 359 PetscFunctionReturn(0); 360 } 361 362 /* 363 Local utility routine that creates a mapping from the global column 364 number to the local number in the off-diagonal part of the local 365 storage of the matrix. When PETSC_USE_CTABLE is used this is scalable at 366 a slightly higher hash table cost; without it it is not scalable (each processor 367 has an order N integer array but is fast to acess. 368 */ 369 #undef __FUNCT__ 370 #define __FUNCT__ "MatCreateColmap_MPIAIJ_Private" 371 PetscErrorCode MatCreateColmap_MPIAIJ_Private(Mat mat) 372 { 373 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 374 PetscErrorCode ierr; 375 PetscInt n = aij->B->cmap->n,i; 376 377 PetscFunctionBegin; 378 if (!aij->garray) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"MPIAIJ Matrix was assembled but is missing garray"); 379 #if defined(PETSC_USE_CTABLE) 380 ierr = PetscTableCreate(n,mat->cmap->N+1,&aij->colmap);CHKERRQ(ierr); 381 for (i=0; i<n; i++) { 382 ierr = PetscTableAdd(aij->colmap,aij->garray[i]+1,i+1,INSERT_VALUES);CHKERRQ(ierr); 383 } 384 #else 385 ierr = PetscCalloc1((mat->cmap->N+1),&aij->colmap);CHKERRQ(ierr); 386 ierr = PetscLogObjectMemory((PetscObject)mat,(mat->cmap->N+1)*sizeof(PetscInt));CHKERRQ(ierr); 387 for (i=0; i<n; i++) aij->colmap[aij->garray[i]] = i+1; 388 #endif 389 PetscFunctionReturn(0); 390 } 391 392 #define MatSetValues_SeqAIJ_A_Private(row,col,value,addv) \ 393 { \ 394 if (col <= lastcol1) low1 = 0; \ 395 else high1 = nrow1; \ 396 lastcol1 = col;\ 397 while (high1-low1 > 5) { \ 398 t = (low1+high1)/2; \ 399 if (rp1[t] > col) high1 = t; \ 400 else low1 = t; \ 401 } \ 402 for (_i=low1; _i<high1; _i++) { \ 403 if (rp1[_i] > col) break; \ 404 if (rp1[_i] == col) { \ 405 if (addv == ADD_VALUES) ap1[_i] += value; \ 406 else ap1[_i] = value; \ 407 goto a_noinsert; \ 408 } \ 409 } \ 410 if (value == 0.0 && ignorezeroentries) {low1 = 0; high1 = nrow1;goto a_noinsert;} \ 411 if (nonew == 1) {low1 = 0; high1 = nrow1; goto a_noinsert;} \ 412 if (nonew == -1) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero (%D, %D) into matrix", row, col); \ 413 MatSeqXAIJReallocateAIJ(A,am,1,nrow1,row,col,rmax1,aa,ai,aj,rp1,ap1,aimax,nonew,MatScalar); \ 414 N = nrow1++ - 1; a->nz++; high1++; \ 415 /* shift up all the later entries in this row */ \ 416 for (ii=N; ii>=_i; ii--) { \ 417 rp1[ii+1] = rp1[ii]; \ 418 ap1[ii+1] = ap1[ii]; \ 419 } \ 420 rp1[_i] = col; \ 421 ap1[_i] = value; \ 422 A->nonzerostate++;\ 423 a_noinsert: ; \ 424 ailen[row] = nrow1; \ 425 } 426 427 428 #define MatSetValues_SeqAIJ_B_Private(row,col,value,addv) \ 429 { \ 430 if (col <= lastcol2) low2 = 0; \ 431 else high2 = nrow2; \ 432 lastcol2 = col; \ 433 while (high2-low2 > 5) { \ 434 t = (low2+high2)/2; \ 435 if (rp2[t] > col) high2 = t; \ 436 else low2 = t; \ 437 } \ 438 for (_i=low2; _i<high2; _i++) { \ 439 if (rp2[_i] > col) break; \ 440 if (rp2[_i] == col) { \ 441 if (addv == ADD_VALUES) ap2[_i] += value; \ 442 else ap2[_i] = value; \ 443 goto b_noinsert; \ 444 } \ 445 } \ 446 if (value == 0.0 && ignorezeroentries) {low2 = 0; high2 = nrow2; goto b_noinsert;} \ 447 if (nonew == 1) {low2 = 0; high2 = nrow2; goto b_noinsert;} \ 448 if (nonew == -1) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero (%D, %D) into matrix", row, col); \ 449 MatSeqXAIJReallocateAIJ(B,bm,1,nrow2,row,col,rmax2,ba,bi,bj,rp2,ap2,bimax,nonew,MatScalar); \ 450 N = nrow2++ - 1; b->nz++; high2++; \ 451 /* shift up all the later entries in this row */ \ 452 for (ii=N; ii>=_i; ii--) { \ 453 rp2[ii+1] = rp2[ii]; \ 454 ap2[ii+1] = ap2[ii]; \ 455 } \ 456 rp2[_i] = col; \ 457 ap2[_i] = value; \ 458 B->nonzerostate++; \ 459 b_noinsert: ; \ 460 bilen[row] = nrow2; \ 461 } 462 463 #undef __FUNCT__ 464 #define __FUNCT__ "MatSetValuesRow_MPIAIJ" 465 PetscErrorCode MatSetValuesRow_MPIAIJ(Mat A,PetscInt row,const PetscScalar v[]) 466 { 467 Mat_MPIAIJ *mat = (Mat_MPIAIJ*)A->data; 468 Mat_SeqAIJ *a = (Mat_SeqAIJ*)mat->A->data,*b = (Mat_SeqAIJ*)mat->B->data; 469 PetscErrorCode ierr; 470 PetscInt l,*garray = mat->garray,diag; 471 472 PetscFunctionBegin; 473 /* code only works for square matrices A */ 474 475 /* find size of row to the left of the diagonal part */ 476 ierr = MatGetOwnershipRange(A,&diag,0);CHKERRQ(ierr); 477 row = row - diag; 478 for (l=0; l<b->i[row+1]-b->i[row]; l++) { 479 if (garray[b->j[b->i[row]+l]] > diag) break; 480 } 481 ierr = PetscMemcpy(b->a+b->i[row],v,l*sizeof(PetscScalar));CHKERRQ(ierr); 482 483 /* diagonal part */ 484 ierr = PetscMemcpy(a->a+a->i[row],v+l,(a->i[row+1]-a->i[row])*sizeof(PetscScalar));CHKERRQ(ierr); 485 486 /* right of diagonal part */ 487 ierr = PetscMemcpy(b->a+b->i[row]+l,v+l+a->i[row+1]-a->i[row],(b->i[row+1]-b->i[row]-l)*sizeof(PetscScalar));CHKERRQ(ierr); 488 PetscFunctionReturn(0); 489 } 490 491 #undef __FUNCT__ 492 #define __FUNCT__ "MatSetValues_MPIAIJ" 493 PetscErrorCode MatSetValues_MPIAIJ(Mat mat,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],const PetscScalar v[],InsertMode addv) 494 { 495 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 496 PetscScalar value; 497 PetscErrorCode ierr; 498 PetscInt i,j,rstart = mat->rmap->rstart,rend = mat->rmap->rend; 499 PetscInt cstart = mat->cmap->rstart,cend = mat->cmap->rend,row,col; 500 PetscBool roworiented = aij->roworiented; 501 502 /* Some Variables required in the macro */ 503 Mat A = aij->A; 504 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 505 PetscInt *aimax = a->imax,*ai = a->i,*ailen = a->ilen,*aj = a->j; 506 MatScalar *aa = a->a; 507 PetscBool ignorezeroentries = a->ignorezeroentries; 508 Mat B = aij->B; 509 Mat_SeqAIJ *b = (Mat_SeqAIJ*)B->data; 510 PetscInt *bimax = b->imax,*bi = b->i,*bilen = b->ilen,*bj = b->j,bm = aij->B->rmap->n,am = aij->A->rmap->n; 511 MatScalar *ba = b->a; 512 513 PetscInt *rp1,*rp2,ii,nrow1,nrow2,_i,rmax1,rmax2,N,low1,high1,low2,high2,t,lastcol1,lastcol2; 514 PetscInt nonew; 515 MatScalar *ap1,*ap2; 516 517 PetscFunctionBegin; 518 for (i=0; i<m; i++) { 519 if (im[i] < 0) continue; 520 #if defined(PETSC_USE_DEBUG) 521 if (im[i] >= mat->rmap->N) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Row too large: row %D max %D",im[i],mat->rmap->N-1); 522 #endif 523 if (im[i] >= rstart && im[i] < rend) { 524 row = im[i] - rstart; 525 lastcol1 = -1; 526 rp1 = aj + ai[row]; 527 ap1 = aa + ai[row]; 528 rmax1 = aimax[row]; 529 nrow1 = ailen[row]; 530 low1 = 0; 531 high1 = nrow1; 532 lastcol2 = -1; 533 rp2 = bj + bi[row]; 534 ap2 = ba + bi[row]; 535 rmax2 = bimax[row]; 536 nrow2 = bilen[row]; 537 low2 = 0; 538 high2 = nrow2; 539 540 for (j=0; j<n; j++) { 541 if (roworiented) value = v[i*n+j]; 542 else value = v[i+j*m]; 543 if (ignorezeroentries && value == 0.0 && (addv == ADD_VALUES)) continue; 544 if (in[j] >= cstart && in[j] < cend) { 545 col = in[j] - cstart; 546 nonew = a->nonew; 547 MatSetValues_SeqAIJ_A_Private(row,col,value,addv); 548 } else if (in[j] < 0) continue; 549 #if defined(PETSC_USE_DEBUG) 550 else if (in[j] >= mat->cmap->N) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Column too large: col %D max %D",in[j],mat->cmap->N-1); 551 #endif 552 else { 553 if (mat->was_assembled) { 554 if (!aij->colmap) { 555 ierr = MatCreateColmap_MPIAIJ_Private(mat);CHKERRQ(ierr); 556 } 557 #if defined(PETSC_USE_CTABLE) 558 ierr = PetscTableFind(aij->colmap,in[j]+1,&col);CHKERRQ(ierr); 559 col--; 560 #else 561 col = aij->colmap[in[j]] - 1; 562 #endif 563 if (col < 0 && !((Mat_SeqAIJ*)(aij->B->data))->nonew) { 564 ierr = MatDisAssemble_MPIAIJ(mat);CHKERRQ(ierr); 565 col = in[j]; 566 /* Reinitialize the variables required by MatSetValues_SeqAIJ_B_Private() */ 567 B = aij->B; 568 b = (Mat_SeqAIJ*)B->data; 569 bimax = b->imax; bi = b->i; bilen = b->ilen; bj = b->j; ba = b->a; 570 rp2 = bj + bi[row]; 571 ap2 = ba + bi[row]; 572 rmax2 = bimax[row]; 573 nrow2 = bilen[row]; 574 low2 = 0; 575 high2 = nrow2; 576 bm = aij->B->rmap->n; 577 ba = b->a; 578 } else if (col < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero (%D, %D) into matrix", im[i], in[j]); 579 } else col = in[j]; 580 nonew = b->nonew; 581 MatSetValues_SeqAIJ_B_Private(row,col,value,addv); 582 } 583 } 584 } else { 585 if (mat->nooffprocentries) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Setting off process row %D even though MatSetOption(,MAT_NO_OFF_PROC_ENTRIES,PETSC_TRUE) was set",im[i]); 586 if (!aij->donotstash) { 587 mat->assembled = PETSC_FALSE; 588 if (roworiented) { 589 ierr = MatStashValuesRow_Private(&mat->stash,im[i],n,in,v+i*n,(PetscBool)(ignorezeroentries && (addv == ADD_VALUES)));CHKERRQ(ierr); 590 } else { 591 ierr = MatStashValuesCol_Private(&mat->stash,im[i],n,in,v+i,m,(PetscBool)(ignorezeroentries && (addv == ADD_VALUES)));CHKERRQ(ierr); 592 } 593 } 594 } 595 } 596 PetscFunctionReturn(0); 597 } 598 599 #undef __FUNCT__ 600 #define __FUNCT__ "MatGetValues_MPIAIJ" 601 PetscErrorCode MatGetValues_MPIAIJ(Mat mat,PetscInt m,const PetscInt idxm[],PetscInt n,const PetscInt idxn[],PetscScalar v[]) 602 { 603 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 604 PetscErrorCode ierr; 605 PetscInt i,j,rstart = mat->rmap->rstart,rend = mat->rmap->rend; 606 PetscInt cstart = mat->cmap->rstart,cend = mat->cmap->rend,row,col; 607 608 PetscFunctionBegin; 609 for (i=0; i<m; i++) { 610 if (idxm[i] < 0) continue; /* SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative row: %D",idxm[i]);*/ 611 if (idxm[i] >= mat->rmap->N) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Row too large: row %D max %D",idxm[i],mat->rmap->N-1); 612 if (idxm[i] >= rstart && idxm[i] < rend) { 613 row = idxm[i] - rstart; 614 for (j=0; j<n; j++) { 615 if (idxn[j] < 0) continue; /* SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative column: %D",idxn[j]); */ 616 if (idxn[j] >= mat->cmap->N) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Column too large: col %D max %D",idxn[j],mat->cmap->N-1); 617 if (idxn[j] >= cstart && idxn[j] < cend) { 618 col = idxn[j] - cstart; 619 ierr = MatGetValues(aij->A,1,&row,1,&col,v+i*n+j);CHKERRQ(ierr); 620 } else { 621 if (!aij->colmap) { 622 ierr = MatCreateColmap_MPIAIJ_Private(mat);CHKERRQ(ierr); 623 } 624 #if defined(PETSC_USE_CTABLE) 625 ierr = PetscTableFind(aij->colmap,idxn[j]+1,&col);CHKERRQ(ierr); 626 col--; 627 #else 628 col = aij->colmap[idxn[j]] - 1; 629 #endif 630 if ((col < 0) || (aij->garray[col] != idxn[j])) *(v+i*n+j) = 0.0; 631 else { 632 ierr = MatGetValues(aij->B,1,&row,1,&col,v+i*n+j);CHKERRQ(ierr); 633 } 634 } 635 } 636 } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Only local values currently supported"); 637 } 638 PetscFunctionReturn(0); 639 } 640 641 extern PetscErrorCode MatMultDiagonalBlock_MPIAIJ(Mat,Vec,Vec); 642 643 #undef __FUNCT__ 644 #define __FUNCT__ "MatAssemblyBegin_MPIAIJ" 645 PetscErrorCode MatAssemblyBegin_MPIAIJ(Mat mat,MatAssemblyType mode) 646 { 647 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 648 PetscErrorCode ierr; 649 PetscInt nstash,reallocs; 650 InsertMode addv; 651 652 PetscFunctionBegin; 653 if (aij->donotstash || mat->nooffprocentries) PetscFunctionReturn(0); 654 655 /* make sure all processors are either in INSERTMODE or ADDMODE */ 656 ierr = MPI_Allreduce((PetscEnum*)&mat->insertmode,(PetscEnum*)&addv,1,MPIU_ENUM,MPI_BOR,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 657 if (addv == (ADD_VALUES|INSERT_VALUES)) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_ARG_WRONGSTATE,"Some processors inserted others added"); 658 mat->insertmode = addv; /* in case this processor had no cache */ 659 660 ierr = MatStashScatterBegin_Private(mat,&mat->stash,mat->rmap->range);CHKERRQ(ierr); 661 ierr = MatStashGetInfo_Private(&mat->stash,&nstash,&reallocs);CHKERRQ(ierr); 662 ierr = PetscInfo2(aij->A,"Stash has %D entries, uses %D mallocs.\n",nstash,reallocs);CHKERRQ(ierr); 663 PetscFunctionReturn(0); 664 } 665 666 #undef __FUNCT__ 667 #define __FUNCT__ "MatAssemblyEnd_MPIAIJ" 668 PetscErrorCode MatAssemblyEnd_MPIAIJ(Mat mat,MatAssemblyType mode) 669 { 670 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 671 Mat_SeqAIJ *a = (Mat_SeqAIJ*)aij->A->data; 672 PetscErrorCode ierr; 673 PetscMPIInt n; 674 PetscInt i,j,rstart,ncols,flg; 675 PetscInt *row,*col; 676 PetscBool other_disassembled; 677 PetscScalar *val; 678 InsertMode addv = mat->insertmode; 679 680 /* do not use 'b = (Mat_SeqAIJ*)aij->B->data' as B can be reset in disassembly */ 681 682 PetscFunctionBegin; 683 if (!aij->donotstash && !mat->nooffprocentries) { 684 while (1) { 685 ierr = MatStashScatterGetMesg_Private(&mat->stash,&n,&row,&col,&val,&flg);CHKERRQ(ierr); 686 if (!flg) break; 687 688 for (i=0; i<n; ) { 689 /* Now identify the consecutive vals belonging to the same row */ 690 for (j=i,rstart=row[j]; j<n; j++) { 691 if (row[j] != rstart) break; 692 } 693 if (j < n) ncols = j-i; 694 else ncols = n-i; 695 /* Now assemble all these values with a single function call */ 696 ierr = MatSetValues_MPIAIJ(mat,1,row+i,ncols,col+i,val+i,addv);CHKERRQ(ierr); 697 698 i = j; 699 } 700 } 701 ierr = MatStashScatterEnd_Private(&mat->stash);CHKERRQ(ierr); 702 } 703 ierr = MatAssemblyBegin(aij->A,mode);CHKERRQ(ierr); 704 ierr = MatAssemblyEnd(aij->A,mode);CHKERRQ(ierr); 705 706 /* determine if any processor has disassembled, if so we must 707 also disassemble ourselfs, in order that we may reassemble. */ 708 /* 709 if nonzero structure of submatrix B cannot change then we know that 710 no processor disassembled thus we can skip this stuff 711 */ 712 if (!((Mat_SeqAIJ*)aij->B->data)->nonew) { 713 ierr = MPI_Allreduce(&mat->was_assembled,&other_disassembled,1,MPIU_BOOL,MPI_PROD,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 714 if (mat->was_assembled && !other_disassembled) { 715 ierr = MatDisAssemble_MPIAIJ(mat);CHKERRQ(ierr); 716 } 717 } 718 if (!mat->was_assembled && mode == MAT_FINAL_ASSEMBLY) { 719 ierr = MatSetUpMultiply_MPIAIJ(mat);CHKERRQ(ierr); 720 } 721 ierr = MatSetOption(aij->B,MAT_USE_INODES,PETSC_FALSE);CHKERRQ(ierr); 722 ierr = MatAssemblyBegin(aij->B,mode);CHKERRQ(ierr); 723 ierr = MatAssemblyEnd(aij->B,mode);CHKERRQ(ierr); 724 725 ierr = PetscFree2(aij->rowvalues,aij->rowindices);CHKERRQ(ierr); 726 727 aij->rowvalues = 0; 728 729 /* used by MatAXPY() */ 730 a->xtoy = 0; ((Mat_SeqAIJ*)aij->B->data)->xtoy = 0; /* b->xtoy = 0 */ 731 a->XtoY = 0; ((Mat_SeqAIJ*)aij->B->data)->XtoY = 0; /* b->XtoY = 0 */ 732 733 ierr = VecDestroy(&aij->diag);CHKERRQ(ierr); 734 if (a->inode.size) mat->ops->multdiagonalblock = MatMultDiagonalBlock_MPIAIJ; 735 736 /* if no new nonzero locations are allowed in matrix then only set the matrix state the first time through */ 737 if ((!mat->was_assembled && mode == MAT_FINAL_ASSEMBLY) || !((Mat_SeqAIJ*)(aij->A->data))->nonew) { 738 PetscObjectState state = aij->A->nonzerostate + aij->B->nonzerostate; 739 ierr = MPI_Allreduce(&state,&mat->nonzerostate,1,MPIU_INT64,MPI_SUM,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 740 } 741 PetscFunctionReturn(0); 742 } 743 744 #undef __FUNCT__ 745 #define __FUNCT__ "MatZeroEntries_MPIAIJ" 746 PetscErrorCode MatZeroEntries_MPIAIJ(Mat A) 747 { 748 Mat_MPIAIJ *l = (Mat_MPIAIJ*)A->data; 749 PetscErrorCode ierr; 750 751 PetscFunctionBegin; 752 ierr = MatZeroEntries(l->A);CHKERRQ(ierr); 753 ierr = MatZeroEntries(l->B);CHKERRQ(ierr); 754 PetscFunctionReturn(0); 755 } 756 757 #undef __FUNCT__ 758 #define __FUNCT__ "MatZeroRows_MPIAIJ" 759 PetscErrorCode MatZeroRows_MPIAIJ(Mat A,PetscInt N,const PetscInt rows[],PetscScalar diag,Vec x,Vec b) 760 { 761 Mat_MPIAIJ *mat = (Mat_MPIAIJ *) A->data; 762 PetscInt *owners = A->rmap->range; 763 PetscInt n = A->rmap->n; 764 PetscSF sf; 765 PetscInt *lrows; 766 PetscSFNode *rrows; 767 PetscInt r, p = 0, len = 0; 768 PetscErrorCode ierr; 769 770 PetscFunctionBegin; 771 /* Create SF where leaves are input rows and roots are owned rows */ 772 ierr = PetscMalloc1(n, &lrows);CHKERRQ(ierr); 773 for (r = 0; r < n; ++r) lrows[r] = -1; 774 if (!A->nooffproczerorows) {ierr = PetscMalloc1(N, &rrows);CHKERRQ(ierr);} 775 for (r = 0; r < N; ++r) { 776 const PetscInt idx = rows[r]; 777 if (idx < 0 || A->rmap->N <= idx) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Row %D out of range [0,%D)",idx,A->rmap->N); 778 if (idx < owners[p] || owners[p+1] <= idx) { /* short-circuit the search if the last p owns this row too */ 779 ierr = PetscLayoutFindOwner(A->rmap,idx,&p);CHKERRQ(ierr); 780 } 781 if (A->nooffproczerorows) { 782 if (p != mat->rank) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"MAT_NO_OFF_PROC_ZERO_ROWS set, but row %D is not owned by rank %d",idx,mat->rank); 783 lrows[len++] = idx - owners[p]; 784 } else { 785 rrows[r].rank = p; 786 rrows[r].index = rows[r] - owners[p]; 787 } 788 } 789 if (!A->nooffproczerorows) { 790 ierr = PetscSFCreate(PetscObjectComm((PetscObject) A), &sf);CHKERRQ(ierr); 791 ierr = PetscSFSetGraph(sf, n, N, NULL, PETSC_OWN_POINTER, rrows, PETSC_OWN_POINTER);CHKERRQ(ierr); 792 /* Collect flags for rows to be zeroed */ 793 ierr = PetscSFReduceBegin(sf, MPIU_INT, (PetscInt*)rows, lrows, MPI_LOR);CHKERRQ(ierr); 794 ierr = PetscSFReduceEnd(sf, MPIU_INT, (PetscInt*)rows, lrows, MPI_LOR);CHKERRQ(ierr); 795 ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 796 /* Compress and put in row numbers */ 797 for (r = 0; r < n; ++r) if (lrows[r] >= 0) lrows[len++] = r; 798 } 799 /* fix right hand side if needed */ 800 if (x && b) { 801 const PetscScalar *xx; 802 PetscScalar *bb; 803 804 ierr = VecGetArrayRead(x, &xx);CHKERRQ(ierr); 805 ierr = VecGetArray(b, &bb);CHKERRQ(ierr); 806 for (r = 0; r < len; ++r) bb[lrows[r]] = diag*xx[lrows[r]]; 807 ierr = VecRestoreArrayRead(x, &xx);CHKERRQ(ierr); 808 ierr = VecRestoreArray(b, &bb);CHKERRQ(ierr); 809 } 810 /* Must zero l->B before l->A because the (diag) case below may put values into l->B*/ 811 ierr = MatZeroRows(mat->B, len, lrows, 0.0, NULL, NULL);CHKERRQ(ierr); 812 if ((diag != 0.0) && (mat->A->rmap->N == mat->A->cmap->N)) { 813 ierr = MatZeroRows(mat->A, len, lrows, diag, NULL, NULL);CHKERRQ(ierr); 814 } else if (diag != 0.0) { 815 ierr = MatZeroRows(mat->A, len, lrows, 0.0, NULL, NULL);CHKERRQ(ierr); 816 if (((Mat_SeqAIJ *) mat->A->data)->nonew) SETERRQ(PETSC_COMM_SELF, PETSC_ERR_SUP, "MatZeroRows() on rectangular matrices cannot be used with the Mat options\nMAT_NEW_NONZERO_LOCATIONS,MAT_NEW_NONZERO_LOCATION_ERR,MAT_NEW_NONZERO_ALLOCATION_ERR"); 817 for (r = 0; r < len; ++r) { 818 const PetscInt row = lrows[r] + A->rmap->rstart; 819 ierr = MatSetValues(A, 1, &row, 1, &row, &diag, INSERT_VALUES);CHKERRQ(ierr); 820 } 821 ierr = MatAssemblyBegin(A, MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 822 ierr = MatAssemblyEnd(A, MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 823 } else { 824 ierr = MatZeroRows(mat->A, len, lrows, 0.0, NULL, NULL);CHKERRQ(ierr); 825 } 826 ierr = PetscFree(lrows);CHKERRQ(ierr); 827 828 /* only change matrix nonzero state if pattern was allowed to be changed */ 829 if (!((Mat_SeqAIJ*)(mat->A->data))->keepnonzeropattern) { 830 PetscObjectState state = mat->A->nonzerostate + mat->B->nonzerostate; 831 ierr = MPI_Allreduce(&state,&A->nonzerostate,1,MPIU_INT64,MPI_SUM,PetscObjectComm((PetscObject)A));CHKERRQ(ierr); 832 } 833 PetscFunctionReturn(0); 834 } 835 836 #undef __FUNCT__ 837 #define __FUNCT__ "MatZeroRowsColumns_MPIAIJ" 838 PetscErrorCode MatZeroRowsColumns_MPIAIJ(Mat A,PetscInt N,const PetscInt rows[],PetscScalar diag,Vec x,Vec b) 839 { 840 Mat_MPIAIJ *l = (Mat_MPIAIJ*)A->data; 841 PetscErrorCode ierr; 842 PetscMPIInt n = A->rmap->n; 843 PetscInt i,j,r,m,p = 0,len = 0; 844 PetscInt *lrows,*owners = A->rmap->range; 845 PetscSFNode *rrows; 846 PetscSF sf; 847 const PetscScalar *xx; 848 PetscScalar *bb,*mask; 849 Vec xmask,lmask; 850 Mat_SeqAIJ *aij = (Mat_SeqAIJ*)l->B->data; 851 const PetscInt *aj, *ii,*ridx; 852 PetscScalar *aa; 853 854 PetscFunctionBegin; 855 /* Create SF where leaves are input rows and roots are owned rows */ 856 ierr = PetscMalloc1(n, &lrows);CHKERRQ(ierr); 857 for (r = 0; r < n; ++r) lrows[r] = -1; 858 ierr = PetscMalloc1(N, &rrows);CHKERRQ(ierr); 859 for (r = 0; r < N; ++r) { 860 const PetscInt idx = rows[r]; 861 if (idx < 0 || A->rmap->N <= idx) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Row %D out of range [0,%D)",idx,A->rmap->N); 862 if (idx < owners[p] || owners[p+1] <= idx) { /* short-circuit the search if the last p owns this row too */ 863 ierr = PetscLayoutFindOwner(A->rmap,idx,&p);CHKERRQ(ierr); 864 } 865 rrows[r].rank = p; 866 rrows[r].index = rows[r] - owners[p]; 867 } 868 ierr = PetscSFCreate(PetscObjectComm((PetscObject) A), &sf);CHKERRQ(ierr); 869 ierr = PetscSFSetGraph(sf, n, N, NULL, PETSC_OWN_POINTER, rrows, PETSC_OWN_POINTER);CHKERRQ(ierr); 870 /* Collect flags for rows to be zeroed */ 871 ierr = PetscSFReduceBegin(sf, MPIU_INT, (PetscInt *) rows, lrows, MPI_LOR);CHKERRQ(ierr); 872 ierr = PetscSFReduceEnd(sf, MPIU_INT, (PetscInt *) rows, lrows, MPI_LOR);CHKERRQ(ierr); 873 ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 874 /* Compress and put in row numbers */ 875 for (r = 0; r < n; ++r) if (lrows[r] >= 0) lrows[len++] = r; 876 /* zero diagonal part of matrix */ 877 ierr = MatZeroRowsColumns(l->A,len,lrows,diag,x,b);CHKERRQ(ierr); 878 /* handle off diagonal part of matrix */ 879 ierr = MatCreateVecs(A,&xmask,NULL);CHKERRQ(ierr); 880 ierr = VecDuplicate(l->lvec,&lmask);CHKERRQ(ierr); 881 ierr = VecGetArray(xmask,&bb);CHKERRQ(ierr); 882 for (i=0; i<len; i++) bb[lrows[i]] = 1; 883 ierr = VecRestoreArray(xmask,&bb);CHKERRQ(ierr); 884 ierr = VecScatterBegin(l->Mvctx,xmask,lmask,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 885 ierr = VecScatterEnd(l->Mvctx,xmask,lmask,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 886 ierr = VecDestroy(&xmask);CHKERRQ(ierr); 887 if (x) { 888 ierr = VecScatterBegin(l->Mvctx,x,l->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 889 ierr = VecScatterEnd(l->Mvctx,x,l->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 890 ierr = VecGetArrayRead(l->lvec,&xx);CHKERRQ(ierr); 891 ierr = VecGetArray(b,&bb);CHKERRQ(ierr); 892 } 893 ierr = VecGetArray(lmask,&mask);CHKERRQ(ierr); 894 /* remove zeroed rows of off diagonal matrix */ 895 ii = aij->i; 896 for (i=0; i<len; i++) { 897 ierr = PetscMemzero(aij->a + ii[lrows[i]],(ii[lrows[i]+1] - ii[lrows[i]])*sizeof(PetscScalar));CHKERRQ(ierr); 898 } 899 /* loop over all elements of off process part of matrix zeroing removed columns*/ 900 if (aij->compressedrow.use) { 901 m = aij->compressedrow.nrows; 902 ii = aij->compressedrow.i; 903 ridx = aij->compressedrow.rindex; 904 for (i=0; i<m; i++) { 905 n = ii[i+1] - ii[i]; 906 aj = aij->j + ii[i]; 907 aa = aij->a + ii[i]; 908 909 for (j=0; j<n; j++) { 910 if (PetscAbsScalar(mask[*aj])) { 911 if (b) bb[*ridx] -= *aa*xx[*aj]; 912 *aa = 0.0; 913 } 914 aa++; 915 aj++; 916 } 917 ridx++; 918 } 919 } else { /* do not use compressed row format */ 920 m = l->B->rmap->n; 921 for (i=0; i<m; i++) { 922 n = ii[i+1] - ii[i]; 923 aj = aij->j + ii[i]; 924 aa = aij->a + ii[i]; 925 for (j=0; j<n; j++) { 926 if (PetscAbsScalar(mask[*aj])) { 927 if (b) bb[i] -= *aa*xx[*aj]; 928 *aa = 0.0; 929 } 930 aa++; 931 aj++; 932 } 933 } 934 } 935 if (x) { 936 ierr = VecRestoreArray(b,&bb);CHKERRQ(ierr); 937 ierr = VecRestoreArrayRead(l->lvec,&xx);CHKERRQ(ierr); 938 } 939 ierr = VecRestoreArray(lmask,&mask);CHKERRQ(ierr); 940 ierr = VecDestroy(&lmask);CHKERRQ(ierr); 941 ierr = PetscFree(lrows);CHKERRQ(ierr); 942 943 /* only change matrix nonzero state if pattern was allowed to be changed */ 944 if (!((Mat_SeqAIJ*)(l->A->data))->keepnonzeropattern) { 945 PetscObjectState state = l->A->nonzerostate + l->B->nonzerostate; 946 ierr = MPI_Allreduce(&state,&A->nonzerostate,1,MPIU_INT64,MPI_SUM,PetscObjectComm((PetscObject)A));CHKERRQ(ierr); 947 } 948 PetscFunctionReturn(0); 949 } 950 951 #undef __FUNCT__ 952 #define __FUNCT__ "MatMult_MPIAIJ" 953 PetscErrorCode MatMult_MPIAIJ(Mat A,Vec xx,Vec yy) 954 { 955 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 956 PetscErrorCode ierr; 957 PetscInt nt; 958 959 PetscFunctionBegin; 960 ierr = VecGetLocalSize(xx,&nt);CHKERRQ(ierr); 961 if (nt != A->cmap->n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Incompatible partition of A (%D) and xx (%D)",A->cmap->n,nt); 962 ierr = VecScatterBegin(a->Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 963 ierr = (*a->A->ops->mult)(a->A,xx,yy);CHKERRQ(ierr); 964 ierr = VecScatterEnd(a->Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 965 ierr = (*a->B->ops->multadd)(a->B,a->lvec,yy,yy);CHKERRQ(ierr); 966 PetscFunctionReturn(0); 967 } 968 969 #undef __FUNCT__ 970 #define __FUNCT__ "MatMultDiagonalBlock_MPIAIJ" 971 PetscErrorCode MatMultDiagonalBlock_MPIAIJ(Mat A,Vec bb,Vec xx) 972 { 973 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 974 PetscErrorCode ierr; 975 976 PetscFunctionBegin; 977 ierr = MatMultDiagonalBlock(a->A,bb,xx);CHKERRQ(ierr); 978 PetscFunctionReturn(0); 979 } 980 981 #undef __FUNCT__ 982 #define __FUNCT__ "MatMultAdd_MPIAIJ" 983 PetscErrorCode MatMultAdd_MPIAIJ(Mat A,Vec xx,Vec yy,Vec zz) 984 { 985 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 986 PetscErrorCode ierr; 987 988 PetscFunctionBegin; 989 ierr = VecScatterBegin(a->Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 990 ierr = (*a->A->ops->multadd)(a->A,xx,yy,zz);CHKERRQ(ierr); 991 ierr = VecScatterEnd(a->Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 992 ierr = (*a->B->ops->multadd)(a->B,a->lvec,zz,zz);CHKERRQ(ierr); 993 PetscFunctionReturn(0); 994 } 995 996 #undef __FUNCT__ 997 #define __FUNCT__ "MatMultTranspose_MPIAIJ" 998 PetscErrorCode MatMultTranspose_MPIAIJ(Mat A,Vec xx,Vec yy) 999 { 1000 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 1001 PetscErrorCode ierr; 1002 PetscBool merged; 1003 1004 PetscFunctionBegin; 1005 ierr = VecScatterGetMerged(a->Mvctx,&merged);CHKERRQ(ierr); 1006 /* do nondiagonal part */ 1007 ierr = (*a->B->ops->multtranspose)(a->B,xx,a->lvec);CHKERRQ(ierr); 1008 if (!merged) { 1009 /* send it on its way */ 1010 ierr = VecScatterBegin(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1011 /* do local part */ 1012 ierr = (*a->A->ops->multtranspose)(a->A,xx,yy);CHKERRQ(ierr); 1013 /* receive remote parts: note this assumes the values are not actually */ 1014 /* added in yy until the next line, */ 1015 ierr = VecScatterEnd(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1016 } else { 1017 /* do local part */ 1018 ierr = (*a->A->ops->multtranspose)(a->A,xx,yy);CHKERRQ(ierr); 1019 /* send it on its way */ 1020 ierr = VecScatterBegin(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1021 /* values actually were received in the Begin() but we need to call this nop */ 1022 ierr = VecScatterEnd(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1023 } 1024 PetscFunctionReturn(0); 1025 } 1026 1027 #undef __FUNCT__ 1028 #define __FUNCT__ "MatIsTranspose_MPIAIJ" 1029 PetscErrorCode MatIsTranspose_MPIAIJ(Mat Amat,Mat Bmat,PetscReal tol,PetscBool *f) 1030 { 1031 MPI_Comm comm; 1032 Mat_MPIAIJ *Aij = (Mat_MPIAIJ*) Amat->data, *Bij; 1033 Mat Adia = Aij->A, Bdia, Aoff,Boff,*Aoffs,*Boffs; 1034 IS Me,Notme; 1035 PetscErrorCode ierr; 1036 PetscInt M,N,first,last,*notme,i; 1037 PetscMPIInt size; 1038 1039 PetscFunctionBegin; 1040 /* Easy test: symmetric diagonal block */ 1041 Bij = (Mat_MPIAIJ*) Bmat->data; Bdia = Bij->A; 1042 ierr = MatIsTranspose(Adia,Bdia,tol,f);CHKERRQ(ierr); 1043 if (!*f) PetscFunctionReturn(0); 1044 ierr = PetscObjectGetComm((PetscObject)Amat,&comm);CHKERRQ(ierr); 1045 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 1046 if (size == 1) PetscFunctionReturn(0); 1047 1048 /* Hard test: off-diagonal block. This takes a MatGetSubMatrix. */ 1049 ierr = MatGetSize(Amat,&M,&N);CHKERRQ(ierr); 1050 ierr = MatGetOwnershipRange(Amat,&first,&last);CHKERRQ(ierr); 1051 ierr = PetscMalloc1((N-last+first),¬me);CHKERRQ(ierr); 1052 for (i=0; i<first; i++) notme[i] = i; 1053 for (i=last; i<M; i++) notme[i-last+first] = i; 1054 ierr = ISCreateGeneral(MPI_COMM_SELF,N-last+first,notme,PETSC_COPY_VALUES,&Notme);CHKERRQ(ierr); 1055 ierr = ISCreateStride(MPI_COMM_SELF,last-first,first,1,&Me);CHKERRQ(ierr); 1056 ierr = MatGetSubMatrices(Amat,1,&Me,&Notme,MAT_INITIAL_MATRIX,&Aoffs);CHKERRQ(ierr); 1057 Aoff = Aoffs[0]; 1058 ierr = MatGetSubMatrices(Bmat,1,&Notme,&Me,MAT_INITIAL_MATRIX,&Boffs);CHKERRQ(ierr); 1059 Boff = Boffs[0]; 1060 ierr = MatIsTranspose(Aoff,Boff,tol,f);CHKERRQ(ierr); 1061 ierr = MatDestroyMatrices(1,&Aoffs);CHKERRQ(ierr); 1062 ierr = MatDestroyMatrices(1,&Boffs);CHKERRQ(ierr); 1063 ierr = ISDestroy(&Me);CHKERRQ(ierr); 1064 ierr = ISDestroy(&Notme);CHKERRQ(ierr); 1065 ierr = PetscFree(notme);CHKERRQ(ierr); 1066 PetscFunctionReturn(0); 1067 } 1068 1069 #undef __FUNCT__ 1070 #define __FUNCT__ "MatMultTransposeAdd_MPIAIJ" 1071 PetscErrorCode MatMultTransposeAdd_MPIAIJ(Mat A,Vec xx,Vec yy,Vec zz) 1072 { 1073 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 1074 PetscErrorCode ierr; 1075 1076 PetscFunctionBegin; 1077 /* do nondiagonal part */ 1078 ierr = (*a->B->ops->multtranspose)(a->B,xx,a->lvec);CHKERRQ(ierr); 1079 /* send it on its way */ 1080 ierr = VecScatterBegin(a->Mvctx,a->lvec,zz,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1081 /* do local part */ 1082 ierr = (*a->A->ops->multtransposeadd)(a->A,xx,yy,zz);CHKERRQ(ierr); 1083 /* receive remote parts */ 1084 ierr = VecScatterEnd(a->Mvctx,a->lvec,zz,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1085 PetscFunctionReturn(0); 1086 } 1087 1088 /* 1089 This only works correctly for square matrices where the subblock A->A is the 1090 diagonal block 1091 */ 1092 #undef __FUNCT__ 1093 #define __FUNCT__ "MatGetDiagonal_MPIAIJ" 1094 PetscErrorCode MatGetDiagonal_MPIAIJ(Mat A,Vec v) 1095 { 1096 PetscErrorCode ierr; 1097 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 1098 1099 PetscFunctionBegin; 1100 if (A->rmap->N != A->cmap->N) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"Supports only square matrix where A->A is diag block"); 1101 if (A->rmap->rstart != A->cmap->rstart || A->rmap->rend != A->cmap->rend) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"row partition must equal col partition"); 1102 ierr = MatGetDiagonal(a->A,v);CHKERRQ(ierr); 1103 PetscFunctionReturn(0); 1104 } 1105 1106 #undef __FUNCT__ 1107 #define __FUNCT__ "MatScale_MPIAIJ" 1108 PetscErrorCode MatScale_MPIAIJ(Mat A,PetscScalar aa) 1109 { 1110 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 1111 PetscErrorCode ierr; 1112 1113 PetscFunctionBegin; 1114 ierr = MatScale(a->A,aa);CHKERRQ(ierr); 1115 ierr = MatScale(a->B,aa);CHKERRQ(ierr); 1116 PetscFunctionReturn(0); 1117 } 1118 1119 #undef __FUNCT__ 1120 #define __FUNCT__ "MatDestroy_Redundant" 1121 PetscErrorCode MatDestroy_Redundant(Mat_Redundant **redundant) 1122 { 1123 PetscErrorCode ierr; 1124 Mat_Redundant *redund = *redundant; 1125 PetscInt i; 1126 1127 PetscFunctionBegin; 1128 *redundant = NULL; 1129 if (redund){ 1130 if (redund->matseq) { /* via MatGetSubMatrices() */ 1131 ierr = ISDestroy(&redund->isrow);CHKERRQ(ierr); 1132 ierr = ISDestroy(&redund->iscol);CHKERRQ(ierr); 1133 ierr = MatDestroy(&redund->matseq[0]);CHKERRQ(ierr); 1134 ierr = PetscFree(redund->matseq);CHKERRQ(ierr); 1135 } else { 1136 ierr = PetscFree2(redund->send_rank,redund->recv_rank);CHKERRQ(ierr); 1137 ierr = PetscFree(redund->sbuf_j);CHKERRQ(ierr); 1138 ierr = PetscFree(redund->sbuf_a);CHKERRQ(ierr); 1139 for (i=0; i<redund->nrecvs; i++) { 1140 ierr = PetscFree(redund->rbuf_j[i]);CHKERRQ(ierr); 1141 ierr = PetscFree(redund->rbuf_a[i]);CHKERRQ(ierr); 1142 } 1143 ierr = PetscFree4(redund->sbuf_nz,redund->rbuf_nz,redund->rbuf_j,redund->rbuf_a);CHKERRQ(ierr); 1144 } 1145 1146 if (redund->subcomm) { 1147 ierr = PetscCommDestroy(&redund->subcomm);CHKERRQ(ierr); 1148 } 1149 ierr = PetscFree(redund);CHKERRQ(ierr); 1150 } 1151 PetscFunctionReturn(0); 1152 } 1153 1154 #undef __FUNCT__ 1155 #define __FUNCT__ "MatDestroy_MPIAIJ" 1156 PetscErrorCode MatDestroy_MPIAIJ(Mat mat) 1157 { 1158 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 1159 PetscErrorCode ierr; 1160 1161 PetscFunctionBegin; 1162 #if defined(PETSC_USE_LOG) 1163 PetscLogObjectState((PetscObject)mat,"Rows=%D, Cols=%D",mat->rmap->N,mat->cmap->N); 1164 #endif 1165 ierr = MatDestroy_Redundant(&mat->redundant);CHKERRQ(ierr); 1166 ierr = MatStashDestroy_Private(&mat->stash);CHKERRQ(ierr); 1167 ierr = VecDestroy(&aij->diag);CHKERRQ(ierr); 1168 ierr = MatDestroy(&aij->A);CHKERRQ(ierr); 1169 ierr = MatDestroy(&aij->B);CHKERRQ(ierr); 1170 #if defined(PETSC_USE_CTABLE) 1171 ierr = PetscTableDestroy(&aij->colmap);CHKERRQ(ierr); 1172 #else 1173 ierr = PetscFree(aij->colmap);CHKERRQ(ierr); 1174 #endif 1175 ierr = PetscFree(aij->garray);CHKERRQ(ierr); 1176 ierr = VecDestroy(&aij->lvec);CHKERRQ(ierr); 1177 ierr = VecScatterDestroy(&aij->Mvctx);CHKERRQ(ierr); 1178 ierr = PetscFree2(aij->rowvalues,aij->rowindices);CHKERRQ(ierr); 1179 ierr = PetscFree(aij->ld);CHKERRQ(ierr); 1180 ierr = PetscFree(mat->data);CHKERRQ(ierr); 1181 1182 ierr = PetscObjectChangeTypeName((PetscObject)mat,0);CHKERRQ(ierr); 1183 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatStoreValues_C",NULL);CHKERRQ(ierr); 1184 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatRetrieveValues_C",NULL);CHKERRQ(ierr); 1185 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatGetDiagonalBlock_C",NULL);CHKERRQ(ierr); 1186 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatIsTranspose_C",NULL);CHKERRQ(ierr); 1187 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatMPIAIJSetPreallocation_C",NULL);CHKERRQ(ierr); 1188 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatMPIAIJSetPreallocationCSR_C",NULL);CHKERRQ(ierr); 1189 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatDiagonalScaleLocal_C",NULL);CHKERRQ(ierr); 1190 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatConvert_mpiaij_mpisbaij_C",NULL);CHKERRQ(ierr); 1191 #if defined(PETSC_HAVE_ELEMENTAL) 1192 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatConvert_mpiaij_elemental_C",NULL);CHKERRQ(ierr); 1193 #endif 1194 PetscFunctionReturn(0); 1195 } 1196 1197 #undef __FUNCT__ 1198 #define __FUNCT__ "MatView_MPIAIJ_Binary" 1199 PetscErrorCode MatView_MPIAIJ_Binary(Mat mat,PetscViewer viewer) 1200 { 1201 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 1202 Mat_SeqAIJ *A = (Mat_SeqAIJ*)aij->A->data; 1203 Mat_SeqAIJ *B = (Mat_SeqAIJ*)aij->B->data; 1204 PetscErrorCode ierr; 1205 PetscMPIInt rank,size,tag = ((PetscObject)viewer)->tag; 1206 int fd; 1207 PetscInt nz,header[4],*row_lengths,*range=0,rlen,i; 1208 PetscInt nzmax,*column_indices,j,k,col,*garray = aij->garray,cnt,cstart = mat->cmap->rstart,rnz = 0; 1209 PetscScalar *column_values; 1210 PetscInt message_count,flowcontrolcount; 1211 FILE *file; 1212 1213 PetscFunctionBegin; 1214 ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr); 1215 ierr = MPI_Comm_size(PetscObjectComm((PetscObject)mat),&size);CHKERRQ(ierr); 1216 nz = A->nz + B->nz; 1217 if (!rank) { 1218 header[0] = MAT_FILE_CLASSID; 1219 header[1] = mat->rmap->N; 1220 header[2] = mat->cmap->N; 1221 1222 ierr = MPI_Reduce(&nz,&header[3],1,MPIU_INT,MPI_SUM,0,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1223 ierr = PetscViewerBinaryGetDescriptor(viewer,&fd);CHKERRQ(ierr); 1224 ierr = PetscBinaryWrite(fd,header,4,PETSC_INT,PETSC_TRUE);CHKERRQ(ierr); 1225 /* get largest number of rows any processor has */ 1226 rlen = mat->rmap->n; 1227 range = mat->rmap->range; 1228 for (i=1; i<size; i++) rlen = PetscMax(rlen,range[i+1] - range[i]); 1229 } else { 1230 ierr = MPI_Reduce(&nz,0,1,MPIU_INT,MPI_SUM,0,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1231 rlen = mat->rmap->n; 1232 } 1233 1234 /* load up the local row counts */ 1235 ierr = PetscMalloc1((rlen+1),&row_lengths);CHKERRQ(ierr); 1236 for (i=0; i<mat->rmap->n; i++) row_lengths[i] = A->i[i+1] - A->i[i] + B->i[i+1] - B->i[i]; 1237 1238 /* store the row lengths to the file */ 1239 ierr = PetscViewerFlowControlStart(viewer,&message_count,&flowcontrolcount);CHKERRQ(ierr); 1240 if (!rank) { 1241 ierr = PetscBinaryWrite(fd,row_lengths,mat->rmap->n,PETSC_INT,PETSC_TRUE);CHKERRQ(ierr); 1242 for (i=1; i<size; i++) { 1243 ierr = PetscViewerFlowControlStepMaster(viewer,i,&message_count,flowcontrolcount);CHKERRQ(ierr); 1244 rlen = range[i+1] - range[i]; 1245 ierr = MPIULong_Recv(row_lengths,rlen,MPIU_INT,i,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1246 ierr = PetscBinaryWrite(fd,row_lengths,rlen,PETSC_INT,PETSC_TRUE);CHKERRQ(ierr); 1247 } 1248 ierr = PetscViewerFlowControlEndMaster(viewer,&message_count);CHKERRQ(ierr); 1249 } else { 1250 ierr = PetscViewerFlowControlStepWorker(viewer,rank,&message_count);CHKERRQ(ierr); 1251 ierr = MPIULong_Send(row_lengths,mat->rmap->n,MPIU_INT,0,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1252 ierr = PetscViewerFlowControlEndWorker(viewer,&message_count);CHKERRQ(ierr); 1253 } 1254 ierr = PetscFree(row_lengths);CHKERRQ(ierr); 1255 1256 /* load up the local column indices */ 1257 nzmax = nz; /* th processor needs space a largest processor needs */ 1258 ierr = MPI_Reduce(&nz,&nzmax,1,MPIU_INT,MPI_MAX,0,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1259 ierr = PetscMalloc1((nzmax+1),&column_indices);CHKERRQ(ierr); 1260 cnt = 0; 1261 for (i=0; i<mat->rmap->n; i++) { 1262 for (j=B->i[i]; j<B->i[i+1]; j++) { 1263 if ((col = garray[B->j[j]]) > cstart) break; 1264 column_indices[cnt++] = col; 1265 } 1266 for (k=A->i[i]; k<A->i[i+1]; k++) column_indices[cnt++] = A->j[k] + cstart; 1267 for (; j<B->i[i+1]; j++) column_indices[cnt++] = garray[B->j[j]]; 1268 } 1269 if (cnt != A->nz + B->nz) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_LIB,"Internal PETSc error: cnt = %D nz = %D",cnt,A->nz+B->nz); 1270 1271 /* store the column indices to the file */ 1272 ierr = PetscViewerFlowControlStart(viewer,&message_count,&flowcontrolcount);CHKERRQ(ierr); 1273 if (!rank) { 1274 MPI_Status status; 1275 ierr = PetscBinaryWrite(fd,column_indices,nz,PETSC_INT,PETSC_TRUE);CHKERRQ(ierr); 1276 for (i=1; i<size; i++) { 1277 ierr = PetscViewerFlowControlStepMaster(viewer,i,&message_count,flowcontrolcount);CHKERRQ(ierr); 1278 ierr = MPI_Recv(&rnz,1,MPIU_INT,i,tag,PetscObjectComm((PetscObject)mat),&status);CHKERRQ(ierr); 1279 if (rnz > nzmax) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_LIB,"Internal PETSc error: nz = %D nzmax = %D",nz,nzmax); 1280 ierr = MPIULong_Recv(column_indices,rnz,MPIU_INT,i,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1281 ierr = PetscBinaryWrite(fd,column_indices,rnz,PETSC_INT,PETSC_TRUE);CHKERRQ(ierr); 1282 } 1283 ierr = PetscViewerFlowControlEndMaster(viewer,&message_count);CHKERRQ(ierr); 1284 } else { 1285 ierr = PetscViewerFlowControlStepWorker(viewer,rank,&message_count);CHKERRQ(ierr); 1286 ierr = MPI_Send(&nz,1,MPIU_INT,0,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1287 ierr = MPIULong_Send(column_indices,nz,MPIU_INT,0,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1288 ierr = PetscViewerFlowControlEndWorker(viewer,&message_count);CHKERRQ(ierr); 1289 } 1290 ierr = PetscFree(column_indices);CHKERRQ(ierr); 1291 1292 /* load up the local column values */ 1293 ierr = PetscMalloc1((nzmax+1),&column_values);CHKERRQ(ierr); 1294 cnt = 0; 1295 for (i=0; i<mat->rmap->n; i++) { 1296 for (j=B->i[i]; j<B->i[i+1]; j++) { 1297 if (garray[B->j[j]] > cstart) break; 1298 column_values[cnt++] = B->a[j]; 1299 } 1300 for (k=A->i[i]; k<A->i[i+1]; k++) column_values[cnt++] = A->a[k]; 1301 for (; j<B->i[i+1]; j++) column_values[cnt++] = B->a[j]; 1302 } 1303 if (cnt != A->nz + B->nz) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Internal PETSc error: cnt = %D nz = %D",cnt,A->nz+B->nz); 1304 1305 /* store the column values to the file */ 1306 ierr = PetscViewerFlowControlStart(viewer,&message_count,&flowcontrolcount);CHKERRQ(ierr); 1307 if (!rank) { 1308 MPI_Status status; 1309 ierr = PetscBinaryWrite(fd,column_values,nz,PETSC_SCALAR,PETSC_TRUE);CHKERRQ(ierr); 1310 for (i=1; i<size; i++) { 1311 ierr = PetscViewerFlowControlStepMaster(viewer,i,&message_count,flowcontrolcount);CHKERRQ(ierr); 1312 ierr = MPI_Recv(&rnz,1,MPIU_INT,i,tag,PetscObjectComm((PetscObject)mat),&status);CHKERRQ(ierr); 1313 if (rnz > nzmax) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_LIB,"Internal PETSc error: nz = %D nzmax = %D",nz,nzmax); 1314 ierr = MPIULong_Recv(column_values,rnz,MPIU_SCALAR,i,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1315 ierr = PetscBinaryWrite(fd,column_values,rnz,PETSC_SCALAR,PETSC_TRUE);CHKERRQ(ierr); 1316 } 1317 ierr = PetscViewerFlowControlEndMaster(viewer,&message_count);CHKERRQ(ierr); 1318 } else { 1319 ierr = PetscViewerFlowControlStepWorker(viewer,rank,&message_count);CHKERRQ(ierr); 1320 ierr = MPI_Send(&nz,1,MPIU_INT,0,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1321 ierr = MPIULong_Send(column_values,nz,MPIU_SCALAR,0,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1322 ierr = PetscViewerFlowControlEndWorker(viewer,&message_count);CHKERRQ(ierr); 1323 } 1324 ierr = PetscFree(column_values);CHKERRQ(ierr); 1325 1326 ierr = PetscViewerBinaryGetInfoPointer(viewer,&file);CHKERRQ(ierr); 1327 if (file) fprintf(file,"-matload_block_size %d\n",(int)PetscAbs(mat->rmap->bs)); 1328 PetscFunctionReturn(0); 1329 } 1330 1331 #include <petscdraw.h> 1332 #undef __FUNCT__ 1333 #define __FUNCT__ "MatView_MPIAIJ_ASCIIorDraworSocket" 1334 PetscErrorCode MatView_MPIAIJ_ASCIIorDraworSocket(Mat mat,PetscViewer viewer) 1335 { 1336 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 1337 PetscErrorCode ierr; 1338 PetscMPIInt rank = aij->rank,size = aij->size; 1339 PetscBool isdraw,iascii,isbinary; 1340 PetscViewer sviewer; 1341 PetscViewerFormat format; 1342 1343 PetscFunctionBegin; 1344 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERDRAW,&isdraw);CHKERRQ(ierr); 1345 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii);CHKERRQ(ierr); 1346 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERBINARY,&isbinary);CHKERRQ(ierr); 1347 if (iascii) { 1348 ierr = PetscViewerGetFormat(viewer,&format);CHKERRQ(ierr); 1349 if (format == PETSC_VIEWER_ASCII_INFO_DETAIL) { 1350 MatInfo info; 1351 PetscBool inodes; 1352 1353 ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr); 1354 ierr = MatGetInfo(mat,MAT_LOCAL,&info);CHKERRQ(ierr); 1355 ierr = MatInodeGetInodeSizes(aij->A,NULL,(PetscInt**)&inodes,NULL);CHKERRQ(ierr); 1356 ierr = PetscViewerASCIISynchronizedAllow(viewer,PETSC_TRUE);CHKERRQ(ierr); 1357 if (!inodes) { 1358 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"[%d] Local rows %D nz %D nz alloced %D mem %D, not using I-node routines\n", 1359 rank,mat->rmap->n,(PetscInt)info.nz_used,(PetscInt)info.nz_allocated,(PetscInt)info.memory);CHKERRQ(ierr); 1360 } else { 1361 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"[%d] Local rows %D nz %D nz alloced %D mem %D, using I-node routines\n", 1362 rank,mat->rmap->n,(PetscInt)info.nz_used,(PetscInt)info.nz_allocated,(PetscInt)info.memory);CHKERRQ(ierr); 1363 } 1364 ierr = MatGetInfo(aij->A,MAT_LOCAL,&info);CHKERRQ(ierr); 1365 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"[%d] on-diagonal part: nz %D \n",rank,(PetscInt)info.nz_used);CHKERRQ(ierr); 1366 ierr = MatGetInfo(aij->B,MAT_LOCAL,&info);CHKERRQ(ierr); 1367 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"[%d] off-diagonal part: nz %D \n",rank,(PetscInt)info.nz_used);CHKERRQ(ierr); 1368 ierr = PetscViewerFlush(viewer);CHKERRQ(ierr); 1369 ierr = PetscViewerASCIISynchronizedAllow(viewer,PETSC_FALSE);CHKERRQ(ierr); 1370 ierr = PetscViewerASCIIPrintf(viewer,"Information on VecScatter used in matrix-vector product: \n");CHKERRQ(ierr); 1371 ierr = VecScatterView(aij->Mvctx,viewer);CHKERRQ(ierr); 1372 PetscFunctionReturn(0); 1373 } else if (format == PETSC_VIEWER_ASCII_INFO) { 1374 PetscInt inodecount,inodelimit,*inodes; 1375 ierr = MatInodeGetInodeSizes(aij->A,&inodecount,&inodes,&inodelimit);CHKERRQ(ierr); 1376 if (inodes) { 1377 ierr = PetscViewerASCIIPrintf(viewer,"using I-node (on process 0) routines: found %D nodes, limit used is %D\n",inodecount,inodelimit);CHKERRQ(ierr); 1378 } else { 1379 ierr = PetscViewerASCIIPrintf(viewer,"not using I-node (on process 0) routines\n");CHKERRQ(ierr); 1380 } 1381 PetscFunctionReturn(0); 1382 } else if (format == PETSC_VIEWER_ASCII_FACTOR_INFO) { 1383 PetscFunctionReturn(0); 1384 } 1385 } else if (isbinary) { 1386 if (size == 1) { 1387 ierr = PetscObjectSetName((PetscObject)aij->A,((PetscObject)mat)->name);CHKERRQ(ierr); 1388 ierr = MatView(aij->A,viewer);CHKERRQ(ierr); 1389 } else { 1390 ierr = MatView_MPIAIJ_Binary(mat,viewer);CHKERRQ(ierr); 1391 } 1392 PetscFunctionReturn(0); 1393 } else if (isdraw) { 1394 PetscDraw draw; 1395 PetscBool isnull; 1396 ierr = PetscViewerDrawGetDraw(viewer,0,&draw);CHKERRQ(ierr); 1397 ierr = PetscDrawIsNull(draw,&isnull);CHKERRQ(ierr); if (isnull) PetscFunctionReturn(0); 1398 } 1399 1400 { 1401 /* assemble the entire matrix onto first processor. */ 1402 Mat A; 1403 Mat_SeqAIJ *Aloc; 1404 PetscInt M = mat->rmap->N,N = mat->cmap->N,m,*ai,*aj,row,*cols,i,*ct; 1405 MatScalar *a; 1406 1407 ierr = MatCreate(PetscObjectComm((PetscObject)mat),&A);CHKERRQ(ierr); 1408 if (!rank) { 1409 ierr = MatSetSizes(A,M,N,M,N);CHKERRQ(ierr); 1410 } else { 1411 ierr = MatSetSizes(A,0,0,M,N);CHKERRQ(ierr); 1412 } 1413 /* This is just a temporary matrix, so explicitly using MATMPIAIJ is probably best */ 1414 ierr = MatSetType(A,MATMPIAIJ);CHKERRQ(ierr); 1415 ierr = MatMPIAIJSetPreallocation(A,0,NULL,0,NULL);CHKERRQ(ierr); 1416 ierr = MatSetOption(A,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 1417 ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)A);CHKERRQ(ierr); 1418 1419 /* copy over the A part */ 1420 Aloc = (Mat_SeqAIJ*)aij->A->data; 1421 m = aij->A->rmap->n; ai = Aloc->i; aj = Aloc->j; a = Aloc->a; 1422 row = mat->rmap->rstart; 1423 for (i=0; i<ai[m]; i++) aj[i] += mat->cmap->rstart; 1424 for (i=0; i<m; i++) { 1425 ierr = MatSetValues(A,1,&row,ai[i+1]-ai[i],aj,a,INSERT_VALUES);CHKERRQ(ierr); 1426 row++; 1427 a += ai[i+1]-ai[i]; aj += ai[i+1]-ai[i]; 1428 } 1429 aj = Aloc->j; 1430 for (i=0; i<ai[m]; i++) aj[i] -= mat->cmap->rstart; 1431 1432 /* copy over the B part */ 1433 Aloc = (Mat_SeqAIJ*)aij->B->data; 1434 m = aij->B->rmap->n; ai = Aloc->i; aj = Aloc->j; a = Aloc->a; 1435 row = mat->rmap->rstart; 1436 ierr = PetscMalloc1((ai[m]+1),&cols);CHKERRQ(ierr); 1437 ct = cols; 1438 for (i=0; i<ai[m]; i++) cols[i] = aij->garray[aj[i]]; 1439 for (i=0; i<m; i++) { 1440 ierr = MatSetValues(A,1,&row,ai[i+1]-ai[i],cols,a,INSERT_VALUES);CHKERRQ(ierr); 1441 row++; 1442 a += ai[i+1]-ai[i]; cols += ai[i+1]-ai[i]; 1443 } 1444 ierr = PetscFree(ct);CHKERRQ(ierr); 1445 ierr = MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1446 ierr = MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1447 /* 1448 Everyone has to call to draw the matrix since the graphics waits are 1449 synchronized across all processors that share the PetscDraw object 1450 */ 1451 ierr = PetscViewerGetSingleton(viewer,&sviewer);CHKERRQ(ierr); 1452 if (!rank) { 1453 ierr = MatView_SeqAIJ(((Mat_MPIAIJ*)(A->data))->A,sviewer);CHKERRQ(ierr); 1454 } 1455 ierr = PetscViewerRestoreSingleton(viewer,&sviewer);CHKERRQ(ierr); 1456 ierr = MatDestroy(&A);CHKERRQ(ierr); 1457 } 1458 PetscFunctionReturn(0); 1459 } 1460 1461 #undef __FUNCT__ 1462 #define __FUNCT__ "MatView_MPIAIJ" 1463 PetscErrorCode MatView_MPIAIJ(Mat mat,PetscViewer viewer) 1464 { 1465 PetscErrorCode ierr; 1466 PetscBool iascii,isdraw,issocket,isbinary; 1467 1468 PetscFunctionBegin; 1469 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii);CHKERRQ(ierr); 1470 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERDRAW,&isdraw);CHKERRQ(ierr); 1471 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERBINARY,&isbinary);CHKERRQ(ierr); 1472 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERSOCKET,&issocket);CHKERRQ(ierr); 1473 if (iascii || isdraw || isbinary || issocket) { 1474 ierr = MatView_MPIAIJ_ASCIIorDraworSocket(mat,viewer);CHKERRQ(ierr); 1475 } 1476 PetscFunctionReturn(0); 1477 } 1478 1479 #undef __FUNCT__ 1480 #define __FUNCT__ "MatSOR_MPIAIJ" 1481 PetscErrorCode MatSOR_MPIAIJ(Mat matin,Vec bb,PetscReal omega,MatSORType flag,PetscReal fshift,PetscInt its,PetscInt lits,Vec xx) 1482 { 1483 Mat_MPIAIJ *mat = (Mat_MPIAIJ*)matin->data; 1484 PetscErrorCode ierr; 1485 Vec bb1 = 0; 1486 PetscBool hasop; 1487 1488 PetscFunctionBegin; 1489 if (flag == SOR_APPLY_UPPER) { 1490 ierr = (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);CHKERRQ(ierr); 1491 PetscFunctionReturn(0); 1492 } 1493 1494 if (its > 1 || ~flag & SOR_ZERO_INITIAL_GUESS || flag & SOR_EISENSTAT) { 1495 ierr = VecDuplicate(bb,&bb1);CHKERRQ(ierr); 1496 } 1497 1498 if ((flag & SOR_LOCAL_SYMMETRIC_SWEEP) == SOR_LOCAL_SYMMETRIC_SWEEP) { 1499 if (flag & SOR_ZERO_INITIAL_GUESS) { 1500 ierr = (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);CHKERRQ(ierr); 1501 its--; 1502 } 1503 1504 while (its--) { 1505 ierr = VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1506 ierr = VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1507 1508 /* update rhs: bb1 = bb - B*x */ 1509 ierr = VecScale(mat->lvec,-1.0);CHKERRQ(ierr); 1510 ierr = (*mat->B->ops->multadd)(mat->B,mat->lvec,bb,bb1);CHKERRQ(ierr); 1511 1512 /* local sweep */ 1513 ierr = (*mat->A->ops->sor)(mat->A,bb1,omega,SOR_SYMMETRIC_SWEEP,fshift,lits,1,xx);CHKERRQ(ierr); 1514 } 1515 } else if (flag & SOR_LOCAL_FORWARD_SWEEP) { 1516 if (flag & SOR_ZERO_INITIAL_GUESS) { 1517 ierr = (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);CHKERRQ(ierr); 1518 its--; 1519 } 1520 while (its--) { 1521 ierr = VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1522 ierr = VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1523 1524 /* update rhs: bb1 = bb - B*x */ 1525 ierr = VecScale(mat->lvec,-1.0);CHKERRQ(ierr); 1526 ierr = (*mat->B->ops->multadd)(mat->B,mat->lvec,bb,bb1);CHKERRQ(ierr); 1527 1528 /* local sweep */ 1529 ierr = (*mat->A->ops->sor)(mat->A,bb1,omega,SOR_FORWARD_SWEEP,fshift,lits,1,xx);CHKERRQ(ierr); 1530 } 1531 } else if (flag & SOR_LOCAL_BACKWARD_SWEEP) { 1532 if (flag & SOR_ZERO_INITIAL_GUESS) { 1533 ierr = (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);CHKERRQ(ierr); 1534 its--; 1535 } 1536 while (its--) { 1537 ierr = VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1538 ierr = VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1539 1540 /* update rhs: bb1 = bb - B*x */ 1541 ierr = VecScale(mat->lvec,-1.0);CHKERRQ(ierr); 1542 ierr = (*mat->B->ops->multadd)(mat->B,mat->lvec,bb,bb1);CHKERRQ(ierr); 1543 1544 /* local sweep */ 1545 ierr = (*mat->A->ops->sor)(mat->A,bb1,omega,SOR_BACKWARD_SWEEP,fshift,lits,1,xx);CHKERRQ(ierr); 1546 } 1547 } else if (flag & SOR_EISENSTAT) { 1548 Vec xx1; 1549 1550 ierr = VecDuplicate(bb,&xx1);CHKERRQ(ierr); 1551 ierr = (*mat->A->ops->sor)(mat->A,bb,omega,(MatSORType)(SOR_ZERO_INITIAL_GUESS | SOR_LOCAL_BACKWARD_SWEEP),fshift,lits,1,xx);CHKERRQ(ierr); 1552 1553 ierr = VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1554 ierr = VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1555 if (!mat->diag) { 1556 ierr = MatCreateVecs(matin,&mat->diag,NULL);CHKERRQ(ierr); 1557 ierr = MatGetDiagonal(matin,mat->diag);CHKERRQ(ierr); 1558 } 1559 ierr = MatHasOperation(matin,MATOP_MULT_DIAGONAL_BLOCK,&hasop);CHKERRQ(ierr); 1560 if (hasop) { 1561 ierr = MatMultDiagonalBlock(matin,xx,bb1);CHKERRQ(ierr); 1562 } else { 1563 ierr = VecPointwiseMult(bb1,mat->diag,xx);CHKERRQ(ierr); 1564 } 1565 ierr = VecAYPX(bb1,(omega-2.0)/omega,bb);CHKERRQ(ierr); 1566 1567 ierr = MatMultAdd(mat->B,mat->lvec,bb1,bb1);CHKERRQ(ierr); 1568 1569 /* local sweep */ 1570 ierr = (*mat->A->ops->sor)(mat->A,bb1,omega,(MatSORType)(SOR_ZERO_INITIAL_GUESS | SOR_LOCAL_FORWARD_SWEEP),fshift,lits,1,xx1);CHKERRQ(ierr); 1571 ierr = VecAXPY(xx,1.0,xx1);CHKERRQ(ierr); 1572 ierr = VecDestroy(&xx1);CHKERRQ(ierr); 1573 } else SETERRQ(PetscObjectComm((PetscObject)matin),PETSC_ERR_SUP,"Parallel SOR not supported"); 1574 1575 ierr = VecDestroy(&bb1);CHKERRQ(ierr); 1576 PetscFunctionReturn(0); 1577 } 1578 1579 #undef __FUNCT__ 1580 #define __FUNCT__ "MatPermute_MPIAIJ" 1581 PetscErrorCode MatPermute_MPIAIJ(Mat A,IS rowp,IS colp,Mat *B) 1582 { 1583 Mat aA,aB,Aperm; 1584 const PetscInt *rwant,*cwant,*gcols,*ai,*bi,*aj,*bj; 1585 PetscScalar *aa,*ba; 1586 PetscInt i,j,m,n,ng,anz,bnz,*dnnz,*onnz,*tdnnz,*tonnz,*rdest,*cdest,*work,*gcdest; 1587 PetscSF rowsf,sf; 1588 IS parcolp = NULL; 1589 PetscBool done; 1590 PetscErrorCode ierr; 1591 1592 PetscFunctionBegin; 1593 ierr = MatGetLocalSize(A,&m,&n);CHKERRQ(ierr); 1594 ierr = ISGetIndices(rowp,&rwant);CHKERRQ(ierr); 1595 ierr = ISGetIndices(colp,&cwant);CHKERRQ(ierr); 1596 ierr = PetscMalloc3(PetscMax(m,n),&work,m,&rdest,n,&cdest);CHKERRQ(ierr); 1597 1598 /* Invert row permutation to find out where my rows should go */ 1599 ierr = PetscSFCreate(PetscObjectComm((PetscObject)A),&rowsf);CHKERRQ(ierr); 1600 ierr = PetscSFSetGraphLayout(rowsf,A->rmap,A->rmap->n,NULL,PETSC_OWN_POINTER,rwant);CHKERRQ(ierr); 1601 ierr = PetscSFSetFromOptions(rowsf);CHKERRQ(ierr); 1602 for (i=0; i<m; i++) work[i] = A->rmap->rstart + i; 1603 ierr = PetscSFReduceBegin(rowsf,MPIU_INT,work,rdest,MPIU_REPLACE);CHKERRQ(ierr); 1604 ierr = PetscSFReduceEnd(rowsf,MPIU_INT,work,rdest,MPIU_REPLACE);CHKERRQ(ierr); 1605 1606 /* Invert column permutation to find out where my columns should go */ 1607 ierr = PetscSFCreate(PetscObjectComm((PetscObject)A),&sf);CHKERRQ(ierr); 1608 ierr = PetscSFSetGraphLayout(sf,A->cmap,A->cmap->n,NULL,PETSC_OWN_POINTER,cwant);CHKERRQ(ierr); 1609 ierr = PetscSFSetFromOptions(sf);CHKERRQ(ierr); 1610 for (i=0; i<n; i++) work[i] = A->cmap->rstart + i; 1611 ierr = PetscSFReduceBegin(sf,MPIU_INT,work,cdest,MPIU_REPLACE);CHKERRQ(ierr); 1612 ierr = PetscSFReduceEnd(sf,MPIU_INT,work,cdest,MPIU_REPLACE);CHKERRQ(ierr); 1613 ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 1614 1615 ierr = ISRestoreIndices(rowp,&rwant);CHKERRQ(ierr); 1616 ierr = ISRestoreIndices(colp,&cwant);CHKERRQ(ierr); 1617 ierr = MatMPIAIJGetSeqAIJ(A,&aA,&aB,&gcols);CHKERRQ(ierr); 1618 1619 /* Find out where my gcols should go */ 1620 ierr = MatGetSize(aB,NULL,&ng);CHKERRQ(ierr); 1621 ierr = PetscMalloc1(ng,&gcdest);CHKERRQ(ierr); 1622 ierr = PetscSFCreate(PetscObjectComm((PetscObject)A),&sf);CHKERRQ(ierr); 1623 ierr = PetscSFSetGraphLayout(sf,A->cmap,ng,NULL,PETSC_OWN_POINTER,gcols);CHKERRQ(ierr); 1624 ierr = PetscSFSetFromOptions(sf);CHKERRQ(ierr); 1625 ierr = PetscSFBcastBegin(sf,MPIU_INT,cdest,gcdest);CHKERRQ(ierr); 1626 ierr = PetscSFBcastEnd(sf,MPIU_INT,cdest,gcdest);CHKERRQ(ierr); 1627 ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 1628 1629 ierr = PetscCalloc4(m,&dnnz,m,&onnz,m,&tdnnz,m,&tonnz);CHKERRQ(ierr); 1630 ierr = MatGetRowIJ(aA,0,PETSC_FALSE,PETSC_FALSE,&anz,&ai,&aj,&done);CHKERRQ(ierr); 1631 ierr = MatGetRowIJ(aB,0,PETSC_FALSE,PETSC_FALSE,&bnz,&bi,&bj,&done);CHKERRQ(ierr); 1632 for (i=0; i<m; i++) { 1633 PetscInt row = rdest[i],rowner; 1634 ierr = PetscLayoutFindOwner(A->rmap,row,&rowner);CHKERRQ(ierr); 1635 for (j=ai[i]; j<ai[i+1]; j++) { 1636 PetscInt cowner,col = cdest[aj[j]]; 1637 ierr = PetscLayoutFindOwner(A->cmap,col,&cowner);CHKERRQ(ierr); /* Could build an index for the columns to eliminate this search */ 1638 if (rowner == cowner) dnnz[i]++; 1639 else onnz[i]++; 1640 } 1641 for (j=bi[i]; j<bi[i+1]; j++) { 1642 PetscInt cowner,col = gcdest[bj[j]]; 1643 ierr = PetscLayoutFindOwner(A->cmap,col,&cowner);CHKERRQ(ierr); 1644 if (rowner == cowner) dnnz[i]++; 1645 else onnz[i]++; 1646 } 1647 } 1648 ierr = PetscSFBcastBegin(rowsf,MPIU_INT,dnnz,tdnnz);CHKERRQ(ierr); 1649 ierr = PetscSFBcastEnd(rowsf,MPIU_INT,dnnz,tdnnz);CHKERRQ(ierr); 1650 ierr = PetscSFBcastBegin(rowsf,MPIU_INT,onnz,tonnz);CHKERRQ(ierr); 1651 ierr = PetscSFBcastEnd(rowsf,MPIU_INT,onnz,tonnz);CHKERRQ(ierr); 1652 ierr = PetscSFDestroy(&rowsf);CHKERRQ(ierr); 1653 1654 ierr = MatCreateAIJ(PetscObjectComm((PetscObject)A),A->rmap->n,A->cmap->n,A->rmap->N,A->cmap->N,0,tdnnz,0,tonnz,&Aperm);CHKERRQ(ierr); 1655 ierr = MatSeqAIJGetArray(aA,&aa);CHKERRQ(ierr); 1656 ierr = MatSeqAIJGetArray(aB,&ba);CHKERRQ(ierr); 1657 for (i=0; i<m; i++) { 1658 PetscInt *acols = dnnz,*bcols = onnz; /* Repurpose now-unneeded arrays */ 1659 PetscInt j0,rowlen; 1660 rowlen = ai[i+1] - ai[i]; 1661 for (j0=j=0; j<rowlen; j0=j) { /* rowlen could be larger than number of rows m, so sum in batches */ 1662 for ( ; j<PetscMin(rowlen,j0+m); j++) acols[j-j0] = cdest[aj[ai[i]+j]]; 1663 ierr = MatSetValues(Aperm,1,&rdest[i],j-j0,acols,aa+ai[i]+j0,INSERT_VALUES);CHKERRQ(ierr); 1664 } 1665 rowlen = bi[i+1] - bi[i]; 1666 for (j0=j=0; j<rowlen; j0=j) { 1667 for ( ; j<PetscMin(rowlen,j0+m); j++) bcols[j-j0] = gcdest[bj[bi[i]+j]]; 1668 ierr = MatSetValues(Aperm,1,&rdest[i],j-j0,bcols,ba+bi[i]+j0,INSERT_VALUES);CHKERRQ(ierr); 1669 } 1670 } 1671 ierr = MatAssemblyBegin(Aperm,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1672 ierr = MatAssemblyEnd(Aperm,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1673 ierr = MatRestoreRowIJ(aA,0,PETSC_FALSE,PETSC_FALSE,&anz,&ai,&aj,&done);CHKERRQ(ierr); 1674 ierr = MatRestoreRowIJ(aB,0,PETSC_FALSE,PETSC_FALSE,&bnz,&bi,&bj,&done);CHKERRQ(ierr); 1675 ierr = MatSeqAIJRestoreArray(aA,&aa);CHKERRQ(ierr); 1676 ierr = MatSeqAIJRestoreArray(aB,&ba);CHKERRQ(ierr); 1677 ierr = PetscFree4(dnnz,onnz,tdnnz,tonnz);CHKERRQ(ierr); 1678 ierr = PetscFree3(work,rdest,cdest);CHKERRQ(ierr); 1679 ierr = PetscFree(gcdest);CHKERRQ(ierr); 1680 if (parcolp) {ierr = ISDestroy(&colp);CHKERRQ(ierr);} 1681 *B = Aperm; 1682 PetscFunctionReturn(0); 1683 } 1684 1685 #undef __FUNCT__ 1686 #define __FUNCT__ "MatGetInfo_MPIAIJ" 1687 PetscErrorCode MatGetInfo_MPIAIJ(Mat matin,MatInfoType flag,MatInfo *info) 1688 { 1689 Mat_MPIAIJ *mat = (Mat_MPIAIJ*)matin->data; 1690 Mat A = mat->A,B = mat->B; 1691 PetscErrorCode ierr; 1692 PetscReal isend[5],irecv[5]; 1693 1694 PetscFunctionBegin; 1695 info->block_size = 1.0; 1696 ierr = MatGetInfo(A,MAT_LOCAL,info);CHKERRQ(ierr); 1697 1698 isend[0] = info->nz_used; isend[1] = info->nz_allocated; isend[2] = info->nz_unneeded; 1699 isend[3] = info->memory; isend[4] = info->mallocs; 1700 1701 ierr = MatGetInfo(B,MAT_LOCAL,info);CHKERRQ(ierr); 1702 1703 isend[0] += info->nz_used; isend[1] += info->nz_allocated; isend[2] += info->nz_unneeded; 1704 isend[3] += info->memory; isend[4] += info->mallocs; 1705 if (flag == MAT_LOCAL) { 1706 info->nz_used = isend[0]; 1707 info->nz_allocated = isend[1]; 1708 info->nz_unneeded = isend[2]; 1709 info->memory = isend[3]; 1710 info->mallocs = isend[4]; 1711 } else if (flag == MAT_GLOBAL_MAX) { 1712 ierr = MPI_Allreduce(isend,irecv,5,MPIU_REAL,MPIU_MAX,PetscObjectComm((PetscObject)matin));CHKERRQ(ierr); 1713 1714 info->nz_used = irecv[0]; 1715 info->nz_allocated = irecv[1]; 1716 info->nz_unneeded = irecv[2]; 1717 info->memory = irecv[3]; 1718 info->mallocs = irecv[4]; 1719 } else if (flag == MAT_GLOBAL_SUM) { 1720 ierr = MPI_Allreduce(isend,irecv,5,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)matin));CHKERRQ(ierr); 1721 1722 info->nz_used = irecv[0]; 1723 info->nz_allocated = irecv[1]; 1724 info->nz_unneeded = irecv[2]; 1725 info->memory = irecv[3]; 1726 info->mallocs = irecv[4]; 1727 } 1728 info->fill_ratio_given = 0; /* no parallel LU/ILU/Cholesky */ 1729 info->fill_ratio_needed = 0; 1730 info->factor_mallocs = 0; 1731 PetscFunctionReturn(0); 1732 } 1733 1734 #undef __FUNCT__ 1735 #define __FUNCT__ "MatSetOption_MPIAIJ" 1736 PetscErrorCode MatSetOption_MPIAIJ(Mat A,MatOption op,PetscBool flg) 1737 { 1738 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 1739 PetscErrorCode ierr; 1740 1741 PetscFunctionBegin; 1742 switch (op) { 1743 case MAT_NEW_NONZERO_LOCATIONS: 1744 case MAT_NEW_NONZERO_ALLOCATION_ERR: 1745 case MAT_UNUSED_NONZERO_LOCATION_ERR: 1746 case MAT_KEEP_NONZERO_PATTERN: 1747 case MAT_NEW_NONZERO_LOCATION_ERR: 1748 case MAT_USE_INODES: 1749 case MAT_IGNORE_ZERO_ENTRIES: 1750 MatCheckPreallocated(A,1); 1751 ierr = MatSetOption(a->A,op,flg);CHKERRQ(ierr); 1752 ierr = MatSetOption(a->B,op,flg);CHKERRQ(ierr); 1753 break; 1754 case MAT_ROW_ORIENTED: 1755 a->roworiented = flg; 1756 1757 ierr = MatSetOption(a->A,op,flg);CHKERRQ(ierr); 1758 ierr = MatSetOption(a->B,op,flg);CHKERRQ(ierr); 1759 break; 1760 case MAT_NEW_DIAGONALS: 1761 ierr = PetscInfo1(A,"Option %s ignored\n",MatOptions[op]);CHKERRQ(ierr); 1762 break; 1763 case MAT_IGNORE_OFF_PROC_ENTRIES: 1764 a->donotstash = flg; 1765 break; 1766 case MAT_SPD: 1767 A->spd_set = PETSC_TRUE; 1768 A->spd = flg; 1769 if (flg) { 1770 A->symmetric = PETSC_TRUE; 1771 A->structurally_symmetric = PETSC_TRUE; 1772 A->symmetric_set = PETSC_TRUE; 1773 A->structurally_symmetric_set = PETSC_TRUE; 1774 } 1775 break; 1776 case MAT_SYMMETRIC: 1777 ierr = MatSetOption(a->A,op,flg);CHKERRQ(ierr); 1778 break; 1779 case MAT_STRUCTURALLY_SYMMETRIC: 1780 ierr = MatSetOption(a->A,op,flg);CHKERRQ(ierr); 1781 break; 1782 case MAT_HERMITIAN: 1783 ierr = MatSetOption(a->A,op,flg);CHKERRQ(ierr); 1784 break; 1785 case MAT_SYMMETRY_ETERNAL: 1786 ierr = MatSetOption(a->A,op,flg);CHKERRQ(ierr); 1787 break; 1788 default: 1789 SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"unknown option %d",op); 1790 } 1791 PetscFunctionReturn(0); 1792 } 1793 1794 #undef __FUNCT__ 1795 #define __FUNCT__ "MatGetRow_MPIAIJ" 1796 PetscErrorCode MatGetRow_MPIAIJ(Mat matin,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v) 1797 { 1798 Mat_MPIAIJ *mat = (Mat_MPIAIJ*)matin->data; 1799 PetscScalar *vworkA,*vworkB,**pvA,**pvB,*v_p; 1800 PetscErrorCode ierr; 1801 PetscInt i,*cworkA,*cworkB,**pcA,**pcB,cstart = matin->cmap->rstart; 1802 PetscInt nztot,nzA,nzB,lrow,rstart = matin->rmap->rstart,rend = matin->rmap->rend; 1803 PetscInt *cmap,*idx_p; 1804 1805 PetscFunctionBegin; 1806 if (mat->getrowactive) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Already active"); 1807 mat->getrowactive = PETSC_TRUE; 1808 1809 if (!mat->rowvalues && (idx || v)) { 1810 /* 1811 allocate enough space to hold information from the longest row. 1812 */ 1813 Mat_SeqAIJ *Aa = (Mat_SeqAIJ*)mat->A->data,*Ba = (Mat_SeqAIJ*)mat->B->data; 1814 PetscInt max = 1,tmp; 1815 for (i=0; i<matin->rmap->n; i++) { 1816 tmp = Aa->i[i+1] - Aa->i[i] + Ba->i[i+1] - Ba->i[i]; 1817 if (max < tmp) max = tmp; 1818 } 1819 ierr = PetscMalloc2(max,&mat->rowvalues,max,&mat->rowindices);CHKERRQ(ierr); 1820 } 1821 1822 if (row < rstart || row >= rend) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Only local rows"); 1823 lrow = row - rstart; 1824 1825 pvA = &vworkA; pcA = &cworkA; pvB = &vworkB; pcB = &cworkB; 1826 if (!v) {pvA = 0; pvB = 0;} 1827 if (!idx) {pcA = 0; if (!v) pcB = 0;} 1828 ierr = (*mat->A->ops->getrow)(mat->A,lrow,&nzA,pcA,pvA);CHKERRQ(ierr); 1829 ierr = (*mat->B->ops->getrow)(mat->B,lrow,&nzB,pcB,pvB);CHKERRQ(ierr); 1830 nztot = nzA + nzB; 1831 1832 cmap = mat->garray; 1833 if (v || idx) { 1834 if (nztot) { 1835 /* Sort by increasing column numbers, assuming A and B already sorted */ 1836 PetscInt imark = -1; 1837 if (v) { 1838 *v = v_p = mat->rowvalues; 1839 for (i=0; i<nzB; i++) { 1840 if (cmap[cworkB[i]] < cstart) v_p[i] = vworkB[i]; 1841 else break; 1842 } 1843 imark = i; 1844 for (i=0; i<nzA; i++) v_p[imark+i] = vworkA[i]; 1845 for (i=imark; i<nzB; i++) v_p[nzA+i] = vworkB[i]; 1846 } 1847 if (idx) { 1848 *idx = idx_p = mat->rowindices; 1849 if (imark > -1) { 1850 for (i=0; i<imark; i++) { 1851 idx_p[i] = cmap[cworkB[i]]; 1852 } 1853 } else { 1854 for (i=0; i<nzB; i++) { 1855 if (cmap[cworkB[i]] < cstart) idx_p[i] = cmap[cworkB[i]]; 1856 else break; 1857 } 1858 imark = i; 1859 } 1860 for (i=0; i<nzA; i++) idx_p[imark+i] = cstart + cworkA[i]; 1861 for (i=imark; i<nzB; i++) idx_p[nzA+i] = cmap[cworkB[i]]; 1862 } 1863 } else { 1864 if (idx) *idx = 0; 1865 if (v) *v = 0; 1866 } 1867 } 1868 *nz = nztot; 1869 ierr = (*mat->A->ops->restorerow)(mat->A,lrow,&nzA,pcA,pvA);CHKERRQ(ierr); 1870 ierr = (*mat->B->ops->restorerow)(mat->B,lrow,&nzB,pcB,pvB);CHKERRQ(ierr); 1871 PetscFunctionReturn(0); 1872 } 1873 1874 #undef __FUNCT__ 1875 #define __FUNCT__ "MatRestoreRow_MPIAIJ" 1876 PetscErrorCode MatRestoreRow_MPIAIJ(Mat mat,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v) 1877 { 1878 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 1879 1880 PetscFunctionBegin; 1881 if (!aij->getrowactive) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"MatGetRow() must be called first"); 1882 aij->getrowactive = PETSC_FALSE; 1883 PetscFunctionReturn(0); 1884 } 1885 1886 #undef __FUNCT__ 1887 #define __FUNCT__ "MatNorm_MPIAIJ" 1888 PetscErrorCode MatNorm_MPIAIJ(Mat mat,NormType type,PetscReal *norm) 1889 { 1890 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 1891 Mat_SeqAIJ *amat = (Mat_SeqAIJ*)aij->A->data,*bmat = (Mat_SeqAIJ*)aij->B->data; 1892 PetscErrorCode ierr; 1893 PetscInt i,j,cstart = mat->cmap->rstart; 1894 PetscReal sum = 0.0; 1895 MatScalar *v; 1896 1897 PetscFunctionBegin; 1898 if (aij->size == 1) { 1899 ierr = MatNorm(aij->A,type,norm);CHKERRQ(ierr); 1900 } else { 1901 if (type == NORM_FROBENIUS) { 1902 v = amat->a; 1903 for (i=0; i<amat->nz; i++) { 1904 sum += PetscRealPart(PetscConj(*v)*(*v)); v++; 1905 } 1906 v = bmat->a; 1907 for (i=0; i<bmat->nz; i++) { 1908 sum += PetscRealPart(PetscConj(*v)*(*v)); v++; 1909 } 1910 ierr = MPI_Allreduce(&sum,norm,1,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1911 *norm = PetscSqrtReal(*norm); 1912 } else if (type == NORM_1) { /* max column norm */ 1913 PetscReal *tmp,*tmp2; 1914 PetscInt *jj,*garray = aij->garray; 1915 ierr = PetscCalloc1((mat->cmap->N+1),&tmp);CHKERRQ(ierr); 1916 ierr = PetscMalloc1((mat->cmap->N+1),&tmp2);CHKERRQ(ierr); 1917 *norm = 0.0; 1918 v = amat->a; jj = amat->j; 1919 for (j=0; j<amat->nz; j++) { 1920 tmp[cstart + *jj++] += PetscAbsScalar(*v); v++; 1921 } 1922 v = bmat->a; jj = bmat->j; 1923 for (j=0; j<bmat->nz; j++) { 1924 tmp[garray[*jj++]] += PetscAbsScalar(*v); v++; 1925 } 1926 ierr = MPI_Allreduce(tmp,tmp2,mat->cmap->N,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1927 for (j=0; j<mat->cmap->N; j++) { 1928 if (tmp2[j] > *norm) *norm = tmp2[j]; 1929 } 1930 ierr = PetscFree(tmp);CHKERRQ(ierr); 1931 ierr = PetscFree(tmp2);CHKERRQ(ierr); 1932 } else if (type == NORM_INFINITY) { /* max row norm */ 1933 PetscReal ntemp = 0.0; 1934 for (j=0; j<aij->A->rmap->n; j++) { 1935 v = amat->a + amat->i[j]; 1936 sum = 0.0; 1937 for (i=0; i<amat->i[j+1]-amat->i[j]; i++) { 1938 sum += PetscAbsScalar(*v); v++; 1939 } 1940 v = bmat->a + bmat->i[j]; 1941 for (i=0; i<bmat->i[j+1]-bmat->i[j]; i++) { 1942 sum += PetscAbsScalar(*v); v++; 1943 } 1944 if (sum > ntemp) ntemp = sum; 1945 } 1946 ierr = MPI_Allreduce(&ntemp,norm,1,MPIU_REAL,MPIU_MAX,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1947 } else SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"No support for two norm"); 1948 } 1949 PetscFunctionReturn(0); 1950 } 1951 1952 #undef __FUNCT__ 1953 #define __FUNCT__ "MatTranspose_MPIAIJ" 1954 PetscErrorCode MatTranspose_MPIAIJ(Mat A,MatReuse reuse,Mat *matout) 1955 { 1956 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 1957 Mat_SeqAIJ *Aloc=(Mat_SeqAIJ*)a->A->data,*Bloc=(Mat_SeqAIJ*)a->B->data; 1958 PetscErrorCode ierr; 1959 PetscInt M = A->rmap->N,N = A->cmap->N,ma,na,mb,nb,*ai,*aj,*bi,*bj,row,*cols,*cols_tmp,i; 1960 PetscInt cstart = A->cmap->rstart,ncol; 1961 Mat B; 1962 MatScalar *array; 1963 1964 PetscFunctionBegin; 1965 if (reuse == MAT_REUSE_MATRIX && A == *matout && M != N) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_SIZ,"Square matrix only for in-place"); 1966 1967 ma = A->rmap->n; na = A->cmap->n; mb = a->B->rmap->n; nb = a->B->cmap->n; 1968 ai = Aloc->i; aj = Aloc->j; 1969 bi = Bloc->i; bj = Bloc->j; 1970 if (reuse == MAT_INITIAL_MATRIX || *matout == A) { 1971 PetscInt *d_nnz,*g_nnz,*o_nnz; 1972 PetscSFNode *oloc; 1973 PETSC_UNUSED PetscSF sf; 1974 1975 ierr = PetscMalloc4(na,&d_nnz,na,&o_nnz,nb,&g_nnz,nb,&oloc);CHKERRQ(ierr); 1976 /* compute d_nnz for preallocation */ 1977 ierr = PetscMemzero(d_nnz,na*sizeof(PetscInt));CHKERRQ(ierr); 1978 for (i=0; i<ai[ma]; i++) { 1979 d_nnz[aj[i]]++; 1980 aj[i] += cstart; /* global col index to be used by MatSetValues() */ 1981 } 1982 /* compute local off-diagonal contributions */ 1983 ierr = PetscMemzero(g_nnz,nb*sizeof(PetscInt));CHKERRQ(ierr); 1984 for (i=0; i<bi[ma]; i++) g_nnz[bj[i]]++; 1985 /* map those to global */ 1986 ierr = PetscSFCreate(PetscObjectComm((PetscObject)A),&sf);CHKERRQ(ierr); 1987 ierr = PetscSFSetGraphLayout(sf,A->cmap,nb,NULL,PETSC_USE_POINTER,a->garray);CHKERRQ(ierr); 1988 ierr = PetscSFSetFromOptions(sf);CHKERRQ(ierr); 1989 ierr = PetscMemzero(o_nnz,na*sizeof(PetscInt));CHKERRQ(ierr); 1990 ierr = PetscSFReduceBegin(sf,MPIU_INT,g_nnz,o_nnz,MPIU_SUM);CHKERRQ(ierr); 1991 ierr = PetscSFReduceEnd(sf,MPIU_INT,g_nnz,o_nnz,MPIU_SUM);CHKERRQ(ierr); 1992 ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 1993 1994 ierr = MatCreate(PetscObjectComm((PetscObject)A),&B);CHKERRQ(ierr); 1995 ierr = MatSetSizes(B,A->cmap->n,A->rmap->n,N,M);CHKERRQ(ierr); 1996 ierr = MatSetBlockSizes(B,PetscAbs(A->cmap->bs),PetscAbs(A->rmap->bs));CHKERRQ(ierr); 1997 ierr = MatSetType(B,((PetscObject)A)->type_name);CHKERRQ(ierr); 1998 ierr = MatMPIAIJSetPreallocation(B,0,d_nnz,0,o_nnz);CHKERRQ(ierr); 1999 ierr = PetscFree4(d_nnz,o_nnz,g_nnz,oloc);CHKERRQ(ierr); 2000 } else { 2001 B = *matout; 2002 ierr = MatSetOption(B,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_TRUE);CHKERRQ(ierr); 2003 for (i=0; i<ai[ma]; i++) aj[i] += cstart; /* global col index to be used by MatSetValues() */ 2004 } 2005 2006 /* copy over the A part */ 2007 array = Aloc->a; 2008 row = A->rmap->rstart; 2009 for (i=0; i<ma; i++) { 2010 ncol = ai[i+1]-ai[i]; 2011 ierr = MatSetValues(B,ncol,aj,1,&row,array,INSERT_VALUES);CHKERRQ(ierr); 2012 row++; 2013 array += ncol; aj += ncol; 2014 } 2015 aj = Aloc->j; 2016 for (i=0; i<ai[ma]; i++) aj[i] -= cstart; /* resume local col index */ 2017 2018 /* copy over the B part */ 2019 ierr = PetscCalloc1(bi[mb],&cols);CHKERRQ(ierr); 2020 array = Bloc->a; 2021 row = A->rmap->rstart; 2022 for (i=0; i<bi[mb]; i++) cols[i] = a->garray[bj[i]]; 2023 cols_tmp = cols; 2024 for (i=0; i<mb; i++) { 2025 ncol = bi[i+1]-bi[i]; 2026 ierr = MatSetValues(B,ncol,cols_tmp,1,&row,array,INSERT_VALUES);CHKERRQ(ierr); 2027 row++; 2028 array += ncol; cols_tmp += ncol; 2029 } 2030 ierr = PetscFree(cols);CHKERRQ(ierr); 2031 2032 ierr = MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2033 ierr = MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2034 if (reuse == MAT_INITIAL_MATRIX || *matout != A) { 2035 *matout = B; 2036 } else { 2037 ierr = MatHeaderMerge(A,B);CHKERRQ(ierr); 2038 } 2039 PetscFunctionReturn(0); 2040 } 2041 2042 #undef __FUNCT__ 2043 #define __FUNCT__ "MatDiagonalScale_MPIAIJ" 2044 PetscErrorCode MatDiagonalScale_MPIAIJ(Mat mat,Vec ll,Vec rr) 2045 { 2046 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 2047 Mat a = aij->A,b = aij->B; 2048 PetscErrorCode ierr; 2049 PetscInt s1,s2,s3; 2050 2051 PetscFunctionBegin; 2052 ierr = MatGetLocalSize(mat,&s2,&s3);CHKERRQ(ierr); 2053 if (rr) { 2054 ierr = VecGetLocalSize(rr,&s1);CHKERRQ(ierr); 2055 if (s1!=s3) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"right vector non-conforming local size"); 2056 /* Overlap communication with computation. */ 2057 ierr = VecScatterBegin(aij->Mvctx,rr,aij->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2058 } 2059 if (ll) { 2060 ierr = VecGetLocalSize(ll,&s1);CHKERRQ(ierr); 2061 if (s1!=s2) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"left vector non-conforming local size"); 2062 ierr = (*b->ops->diagonalscale)(b,ll,0);CHKERRQ(ierr); 2063 } 2064 /* scale the diagonal block */ 2065 ierr = (*a->ops->diagonalscale)(a,ll,rr);CHKERRQ(ierr); 2066 2067 if (rr) { 2068 /* Do a scatter end and then right scale the off-diagonal block */ 2069 ierr = VecScatterEnd(aij->Mvctx,rr,aij->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2070 ierr = (*b->ops->diagonalscale)(b,0,aij->lvec);CHKERRQ(ierr); 2071 } 2072 PetscFunctionReturn(0); 2073 } 2074 2075 #undef __FUNCT__ 2076 #define __FUNCT__ "MatSetUnfactored_MPIAIJ" 2077 PetscErrorCode MatSetUnfactored_MPIAIJ(Mat A) 2078 { 2079 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 2080 PetscErrorCode ierr; 2081 2082 PetscFunctionBegin; 2083 ierr = MatSetUnfactored(a->A);CHKERRQ(ierr); 2084 PetscFunctionReturn(0); 2085 } 2086 2087 #undef __FUNCT__ 2088 #define __FUNCT__ "MatEqual_MPIAIJ" 2089 PetscErrorCode MatEqual_MPIAIJ(Mat A,Mat B,PetscBool *flag) 2090 { 2091 Mat_MPIAIJ *matB = (Mat_MPIAIJ*)B->data,*matA = (Mat_MPIAIJ*)A->data; 2092 Mat a,b,c,d; 2093 PetscBool flg; 2094 PetscErrorCode ierr; 2095 2096 PetscFunctionBegin; 2097 a = matA->A; b = matA->B; 2098 c = matB->A; d = matB->B; 2099 2100 ierr = MatEqual(a,c,&flg);CHKERRQ(ierr); 2101 if (flg) { 2102 ierr = MatEqual(b,d,&flg);CHKERRQ(ierr); 2103 } 2104 ierr = MPI_Allreduce(&flg,flag,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)A));CHKERRQ(ierr); 2105 PetscFunctionReturn(0); 2106 } 2107 2108 #undef __FUNCT__ 2109 #define __FUNCT__ "MatCopy_MPIAIJ" 2110 PetscErrorCode MatCopy_MPIAIJ(Mat A,Mat B,MatStructure str) 2111 { 2112 PetscErrorCode ierr; 2113 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 2114 Mat_MPIAIJ *b = (Mat_MPIAIJ*)B->data; 2115 2116 PetscFunctionBegin; 2117 /* If the two matrices don't have the same copy implementation, they aren't compatible for fast copy. */ 2118 if ((str != SAME_NONZERO_PATTERN) || (A->ops->copy != B->ops->copy)) { 2119 /* because of the column compression in the off-processor part of the matrix a->B, 2120 the number of columns in a->B and b->B may be different, hence we cannot call 2121 the MatCopy() directly on the two parts. If need be, we can provide a more 2122 efficient copy than the MatCopy_Basic() by first uncompressing the a->B matrices 2123 then copying the submatrices */ 2124 ierr = MatCopy_Basic(A,B,str);CHKERRQ(ierr); 2125 } else { 2126 ierr = MatCopy(a->A,b->A,str);CHKERRQ(ierr); 2127 ierr = MatCopy(a->B,b->B,str);CHKERRQ(ierr); 2128 } 2129 PetscFunctionReturn(0); 2130 } 2131 2132 #undef __FUNCT__ 2133 #define __FUNCT__ "MatSetUp_MPIAIJ" 2134 PetscErrorCode MatSetUp_MPIAIJ(Mat A) 2135 { 2136 PetscErrorCode ierr; 2137 2138 PetscFunctionBegin; 2139 ierr = MatMPIAIJSetPreallocation(A,PETSC_DEFAULT,0,PETSC_DEFAULT,0);CHKERRQ(ierr); 2140 PetscFunctionReturn(0); 2141 } 2142 2143 /* 2144 Computes the number of nonzeros per row needed for preallocation when X and Y 2145 have different nonzero structure. 2146 */ 2147 #undef __FUNCT__ 2148 #define __FUNCT__ "MatAXPYGetPreallocation_MPIX_private" 2149 PetscErrorCode MatAXPYGetPreallocation_MPIX_private(PetscInt m,const PetscInt *xi,const PetscInt *xj,const PetscInt *xltog,const PetscInt *yi,const PetscInt *yj,const PetscInt *yltog,PetscInt *nnz) 2150 { 2151 PetscInt i,j,k,nzx,nzy; 2152 2153 PetscFunctionBegin; 2154 /* Set the number of nonzeros in the new matrix */ 2155 for (i=0; i<m; i++) { 2156 const PetscInt *xjj = xj+xi[i],*yjj = yj+yi[i]; 2157 nzx = xi[i+1] - xi[i]; 2158 nzy = yi[i+1] - yi[i]; 2159 nnz[i] = 0; 2160 for (j=0,k=0; j<nzx; j++) { /* Point in X */ 2161 for (; k<nzy && yltog[yjj[k]]<xltog[xjj[j]]; k++) nnz[i]++; /* Catch up to X */ 2162 if (k<nzy && yltog[yjj[k]]==xltog[xjj[j]]) k++; /* Skip duplicate */ 2163 nnz[i]++; 2164 } 2165 for (; k<nzy; k++) nnz[i]++; 2166 } 2167 PetscFunctionReturn(0); 2168 } 2169 2170 /* This is the same as MatAXPYGetPreallocation_SeqAIJ, except that the local-to-global map is provided */ 2171 #undef __FUNCT__ 2172 #define __FUNCT__ "MatAXPYGetPreallocation_MPIAIJ" 2173 static PetscErrorCode MatAXPYGetPreallocation_MPIAIJ(Mat Y,const PetscInt *yltog,Mat X,const PetscInt *xltog,PetscInt *nnz) 2174 { 2175 PetscErrorCode ierr; 2176 PetscInt m = Y->rmap->N; 2177 Mat_SeqAIJ *x = (Mat_SeqAIJ*)X->data; 2178 Mat_SeqAIJ *y = (Mat_SeqAIJ*)Y->data; 2179 2180 PetscFunctionBegin; 2181 ierr = MatAXPYGetPreallocation_MPIX_private(m,x->i,x->j,xltog,y->i,y->j,yltog,nnz);CHKERRQ(ierr); 2182 PetscFunctionReturn(0); 2183 } 2184 2185 #undef __FUNCT__ 2186 #define __FUNCT__ "MatAXPY_MPIAIJ" 2187 PetscErrorCode MatAXPY_MPIAIJ(Mat Y,PetscScalar a,Mat X,MatStructure str) 2188 { 2189 PetscErrorCode ierr; 2190 PetscInt i; 2191 Mat_MPIAIJ *xx = (Mat_MPIAIJ*)X->data,*yy = (Mat_MPIAIJ*)Y->data; 2192 PetscBLASInt bnz,one=1; 2193 Mat_SeqAIJ *x,*y; 2194 2195 PetscFunctionBegin; 2196 if (str == SAME_NONZERO_PATTERN) { 2197 PetscScalar alpha = a; 2198 x = (Mat_SeqAIJ*)xx->A->data; 2199 ierr = PetscBLASIntCast(x->nz,&bnz);CHKERRQ(ierr); 2200 y = (Mat_SeqAIJ*)yy->A->data; 2201 PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&bnz,&alpha,x->a,&one,y->a,&one)); 2202 x = (Mat_SeqAIJ*)xx->B->data; 2203 y = (Mat_SeqAIJ*)yy->B->data; 2204 ierr = PetscBLASIntCast(x->nz,&bnz);CHKERRQ(ierr); 2205 PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&bnz,&alpha,x->a,&one,y->a,&one)); 2206 ierr = PetscObjectStateIncrease((PetscObject)Y);CHKERRQ(ierr); 2207 } else if (str == SUBSET_NONZERO_PATTERN) { 2208 ierr = MatAXPY_SeqAIJ(yy->A,a,xx->A,str);CHKERRQ(ierr); 2209 2210 x = (Mat_SeqAIJ*)xx->B->data; 2211 y = (Mat_SeqAIJ*)yy->B->data; 2212 if (y->xtoy && y->XtoY != xx->B) { 2213 ierr = PetscFree(y->xtoy);CHKERRQ(ierr); 2214 ierr = MatDestroy(&y->XtoY);CHKERRQ(ierr); 2215 } 2216 if (!y->xtoy) { /* get xtoy */ 2217 ierr = MatAXPYGetxtoy_Private(xx->B->rmap->n,x->i,x->j,xx->garray,y->i,y->j,yy->garray,&y->xtoy);CHKERRQ(ierr); 2218 y->XtoY = xx->B; 2219 ierr = PetscObjectReference((PetscObject)xx->B);CHKERRQ(ierr); 2220 } 2221 for (i=0; i<x->nz; i++) y->a[y->xtoy[i]] += a*(x->a[i]); 2222 ierr = PetscObjectStateIncrease((PetscObject)Y);CHKERRQ(ierr); 2223 } else { 2224 Mat B; 2225 PetscInt *nnz_d,*nnz_o; 2226 ierr = PetscMalloc1(yy->A->rmap->N,&nnz_d);CHKERRQ(ierr); 2227 ierr = PetscMalloc1(yy->B->rmap->N,&nnz_o);CHKERRQ(ierr); 2228 ierr = MatCreate(PetscObjectComm((PetscObject)Y),&B);CHKERRQ(ierr); 2229 ierr = PetscObjectSetName((PetscObject)B,((PetscObject)Y)->name);CHKERRQ(ierr); 2230 ierr = MatSetSizes(B,Y->rmap->n,Y->cmap->n,Y->rmap->N,Y->cmap->N);CHKERRQ(ierr); 2231 ierr = MatSetBlockSizesFromMats(B,Y,Y);CHKERRQ(ierr); 2232 ierr = MatSetType(B,MATMPIAIJ);CHKERRQ(ierr); 2233 ierr = MatAXPYGetPreallocation_SeqAIJ(yy->A,xx->A,nnz_d);CHKERRQ(ierr); 2234 ierr = MatAXPYGetPreallocation_MPIAIJ(yy->B,yy->garray,xx->B,xx->garray,nnz_o);CHKERRQ(ierr); 2235 ierr = MatMPIAIJSetPreallocation(B,0,nnz_d,0,nnz_o);CHKERRQ(ierr); 2236 ierr = MatAXPY_BasicWithPreallocation(B,Y,a,X,str);CHKERRQ(ierr); 2237 ierr = MatHeaderReplace(Y,B);CHKERRQ(ierr); 2238 ierr = PetscFree(nnz_d);CHKERRQ(ierr); 2239 ierr = PetscFree(nnz_o);CHKERRQ(ierr); 2240 } 2241 PetscFunctionReturn(0); 2242 } 2243 2244 extern PetscErrorCode MatConjugate_SeqAIJ(Mat); 2245 2246 #undef __FUNCT__ 2247 #define __FUNCT__ "MatConjugate_MPIAIJ" 2248 PetscErrorCode MatConjugate_MPIAIJ(Mat mat) 2249 { 2250 #if defined(PETSC_USE_COMPLEX) 2251 PetscErrorCode ierr; 2252 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 2253 2254 PetscFunctionBegin; 2255 ierr = MatConjugate_SeqAIJ(aij->A);CHKERRQ(ierr); 2256 ierr = MatConjugate_SeqAIJ(aij->B);CHKERRQ(ierr); 2257 #else 2258 PetscFunctionBegin; 2259 #endif 2260 PetscFunctionReturn(0); 2261 } 2262 2263 #undef __FUNCT__ 2264 #define __FUNCT__ "MatRealPart_MPIAIJ" 2265 PetscErrorCode MatRealPart_MPIAIJ(Mat A) 2266 { 2267 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 2268 PetscErrorCode ierr; 2269 2270 PetscFunctionBegin; 2271 ierr = MatRealPart(a->A);CHKERRQ(ierr); 2272 ierr = MatRealPart(a->B);CHKERRQ(ierr); 2273 PetscFunctionReturn(0); 2274 } 2275 2276 #undef __FUNCT__ 2277 #define __FUNCT__ "MatImaginaryPart_MPIAIJ" 2278 PetscErrorCode MatImaginaryPart_MPIAIJ(Mat A) 2279 { 2280 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 2281 PetscErrorCode ierr; 2282 2283 PetscFunctionBegin; 2284 ierr = MatImaginaryPart(a->A);CHKERRQ(ierr); 2285 ierr = MatImaginaryPart(a->B);CHKERRQ(ierr); 2286 PetscFunctionReturn(0); 2287 } 2288 2289 #if defined(PETSC_HAVE_PBGL) 2290 2291 #include <boost/parallel/mpi/bsp_process_group.hpp> 2292 #include <boost/graph/distributed/ilu_default_graph.hpp> 2293 #include <boost/graph/distributed/ilu_0_block.hpp> 2294 #include <boost/graph/distributed/ilu_preconditioner.hpp> 2295 #include <boost/graph/distributed/petsc/interface.hpp> 2296 #include <boost/multi_array.hpp> 2297 #include <boost/parallel/distributed_property_map->hpp> 2298 2299 #undef __FUNCT__ 2300 #define __FUNCT__ "MatILUFactorSymbolic_MPIAIJ" 2301 /* 2302 This uses the parallel ILU factorization of Peter Gottschling <pgottsch@osl.iu.edu> 2303 */ 2304 PetscErrorCode MatILUFactorSymbolic_MPIAIJ(Mat fact,Mat A, IS isrow, IS iscol, const MatFactorInfo *info) 2305 { 2306 namespace petsc = boost::distributed::petsc; 2307 2308 namespace graph_dist = boost::graph::distributed; 2309 using boost::graph::distributed::ilu_default::process_group_type; 2310 using boost::graph::ilu_permuted; 2311 2312 PetscBool row_identity, col_identity; 2313 PetscContainer c; 2314 PetscInt m, n, M, N; 2315 PetscErrorCode ierr; 2316 2317 PetscFunctionBegin; 2318 if (info->levels != 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Only levels = 0 supported for parallel ilu"); 2319 ierr = ISIdentity(isrow, &row_identity);CHKERRQ(ierr); 2320 ierr = ISIdentity(iscol, &col_identity);CHKERRQ(ierr); 2321 if (!row_identity || !col_identity) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Row and column permutations must be identity for parallel ILU"); 2322 2323 process_group_type pg; 2324 typedef graph_dist::ilu_default::ilu_level_graph_type lgraph_type; 2325 lgraph_type *lgraph_p = new lgraph_type(petsc::num_global_vertices(A), pg, petsc::matrix_distribution(A, pg)); 2326 lgraph_type& level_graph = *lgraph_p; 2327 graph_dist::ilu_default::graph_type& graph(level_graph.graph); 2328 2329 petsc::read_matrix(A, graph, get(boost::edge_weight, graph)); 2330 ilu_permuted(level_graph); 2331 2332 /* put together the new matrix */ 2333 ierr = MatCreate(PetscObjectComm((PetscObject)A), fact);CHKERRQ(ierr); 2334 ierr = MatGetLocalSize(A, &m, &n);CHKERRQ(ierr); 2335 ierr = MatGetSize(A, &M, &N);CHKERRQ(ierr); 2336 ierr = MatSetSizes(fact, m, n, M, N);CHKERRQ(ierr); 2337 ierr = MatSetBlockSizesFromMats(fact,A,A);CHKERRQ(ierr); 2338 ierr = MatSetType(fact, ((PetscObject)A)->type_name);CHKERRQ(ierr); 2339 ierr = MatAssemblyBegin(fact, MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2340 ierr = MatAssemblyEnd(fact, MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2341 2342 ierr = PetscContainerCreate(PetscObjectComm((PetscObject)A), &c); 2343 ierr = PetscContainerSetPointer(c, lgraph_p); 2344 ierr = PetscObjectCompose((PetscObject) (fact), "graph", (PetscObject) c); 2345 ierr = PetscContainerDestroy(&c); 2346 PetscFunctionReturn(0); 2347 } 2348 2349 #undef __FUNCT__ 2350 #define __FUNCT__ "MatLUFactorNumeric_MPIAIJ" 2351 PetscErrorCode MatLUFactorNumeric_MPIAIJ(Mat B,Mat A, const MatFactorInfo *info) 2352 { 2353 PetscFunctionBegin; 2354 PetscFunctionReturn(0); 2355 } 2356 2357 #undef __FUNCT__ 2358 #define __FUNCT__ "MatSolve_MPIAIJ" 2359 /* 2360 This uses the parallel ILU factorization of Peter Gottschling <pgottsch@osl.iu.edu> 2361 */ 2362 PetscErrorCode MatSolve_MPIAIJ(Mat A, Vec b, Vec x) 2363 { 2364 namespace graph_dist = boost::graph::distributed; 2365 2366 typedef graph_dist::ilu_default::ilu_level_graph_type lgraph_type; 2367 lgraph_type *lgraph_p; 2368 PetscContainer c; 2369 PetscErrorCode ierr; 2370 2371 PetscFunctionBegin; 2372 ierr = PetscObjectQuery((PetscObject) A, "graph", (PetscObject*) &c);CHKERRQ(ierr); 2373 ierr = PetscContainerGetPointer(c, (void**) &lgraph_p);CHKERRQ(ierr); 2374 ierr = VecCopy(b, x);CHKERRQ(ierr); 2375 2376 PetscScalar *array_x; 2377 ierr = VecGetArray(x, &array_x);CHKERRQ(ierr); 2378 PetscInt sx; 2379 ierr = VecGetSize(x, &sx);CHKERRQ(ierr); 2380 2381 PetscScalar *array_b; 2382 ierr = VecGetArray(b, &array_b);CHKERRQ(ierr); 2383 PetscInt sb; 2384 ierr = VecGetSize(b, &sb);CHKERRQ(ierr); 2385 2386 lgraph_type& level_graph = *lgraph_p; 2387 graph_dist::ilu_default::graph_type& graph(level_graph.graph); 2388 2389 typedef boost::multi_array_ref<PetscScalar, 1> array_ref_type; 2390 array_ref_type ref_b(array_b, boost::extents[num_vertices(graph)]); 2391 array_ref_type ref_x(array_x, boost::extents[num_vertices(graph)]); 2392 2393 typedef boost::iterator_property_map<array_ref_type::iterator, 2394 boost::property_map<graph_dist::ilu_default::graph_type, boost::vertex_index_t>::type> gvector_type; 2395 gvector_type vector_b(ref_b.begin(), get(boost::vertex_index, graph)); 2396 gvector_type vector_x(ref_x.begin(), get(boost::vertex_index, graph)); 2397 2398 ilu_set_solve(*lgraph_p, vector_b, vector_x); 2399 PetscFunctionReturn(0); 2400 } 2401 #endif 2402 2403 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->subcomm = subcomm; 2745 redund->type = PETSC_SUBCOMM_INTERLACED; 2746 } 2747 PetscFunctionReturn(0); 2748 } 2749 2750 #undef __FUNCT__ 2751 #define __FUNCT__ "MatGetRedundantMatrix_MPIAIJ" 2752 PetscErrorCode MatGetRedundantMatrix_MPIAIJ(Mat mat,PetscInt nsubcomm,MPI_Comm subcomm,MatReuse reuse,Mat *matredundant) 2753 { 2754 PetscErrorCode ierr; 2755 MPI_Comm comm; 2756 PetscMPIInt size; 2757 PetscInt mloc_sub,rstart,rend,M=mat->rmap->N,N=mat->cmap->N; 2758 Mat_Redundant *redund=NULL; 2759 PetscSubcomm psubcomm=NULL; 2760 MPI_Comm subcomm_in=subcomm; 2761 Mat *matseq; 2762 IS isrow,iscol; 2763 2764 PetscFunctionBegin; 2765 if (subcomm_in == MPI_COMM_NULL && reuse == MAT_INITIAL_MATRIX) { /* get subcomm if user does not provide subcomm */ 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 ierr = PetscCommDuplicate(psubcomm->comm,&subcomm,NULL);CHKERRQ(ierr); 2776 if (psubcomm->type == PETSC_SUBCOMM_INTERLACED) { 2777 ierr = MatGetRedundantMatrix_MPIAIJ_interlaced(mat,nsubcomm,subcomm,reuse,matredundant);CHKERRQ(ierr); 2778 ierr = PetscSubcommDestroy(&psubcomm);CHKERRQ(ierr); 2779 PetscFunctionReturn(0); 2780 } 2781 ierr = PetscSubcommDestroy(&psubcomm);CHKERRQ(ierr); 2782 } 2783 2784 /* get isrow and iscol for redund->type == PETSC_SUBCOMM_CONTIGUOUS */ 2785 if (reuse == MAT_INITIAL_MATRIX) { 2786 /* create a local sequential matrix matseq[0] */ 2787 mloc_sub = PETSC_DECIDE; 2788 ierr = PetscSplitOwnership(subcomm,&mloc_sub,&M);CHKERRQ(ierr); 2789 ierr = MPI_Scan(&mloc_sub,&rend,1,MPIU_INT,MPI_SUM,subcomm);CHKERRQ(ierr); 2790 rstart = rend - mloc_sub; 2791 ierr = ISCreateStride(PETSC_COMM_SELF,mloc_sub,rstart,1,&isrow);CHKERRQ(ierr); 2792 ierr = ISCreateStride(PETSC_COMM_SELF,N,0,1,&iscol);CHKERRQ(ierr); 2793 } else { /* reuse == MAT_REUSE_MATRIX */ 2794 /* retrieve subcomm */ 2795 ierr = PetscObjectGetComm((PetscObject)(*matredundant),&subcomm);CHKERRQ(ierr); 2796 redund = (*matredundant)->redundant; 2797 if (redund->type == PETSC_SUBCOMM_INTERLACED) { 2798 ierr = MatGetRedundantMatrix_MPIAIJ_interlaced(mat,nsubcomm,subcomm,reuse,matredundant);CHKERRQ(ierr); 2799 PetscFunctionReturn(0); 2800 } 2801 2802 isrow = redund->isrow; 2803 iscol = redund->iscol; 2804 matseq = redund->matseq; 2805 } 2806 ierr = MatGetSubMatrices(mat,1,&isrow,&iscol,reuse,&matseq);CHKERRQ(ierr); 2807 ierr = MatCreateMPIMatConcatenateSeqMat(subcomm,matseq[0],PETSC_DECIDE,reuse,matredundant);CHKERRQ(ierr); 2808 2809 if (reuse == MAT_INITIAL_MATRIX) { 2810 /* create a supporting struct and attach it to C for reuse */ 2811 ierr = PetscNewLog(*matredundant,&redund);CHKERRQ(ierr); 2812 (*matredundant)->redundant = redund; 2813 redund->isrow = isrow; 2814 redund->iscol = iscol; 2815 redund->matseq = matseq; 2816 redund->subcomm = subcomm; 2817 redund->type = PETSC_SUBCOMM_CONTIGUOUS; 2818 } 2819 PetscFunctionReturn(0); 2820 } 2821 2822 #undef __FUNCT__ 2823 #define __FUNCT__ "MatGetRowMaxAbs_MPIAIJ" 2824 PetscErrorCode MatGetRowMaxAbs_MPIAIJ(Mat A, Vec v, PetscInt idx[]) 2825 { 2826 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 2827 PetscErrorCode ierr; 2828 PetscInt i,*idxb = 0; 2829 PetscScalar *va,*vb; 2830 Vec vtmp; 2831 2832 PetscFunctionBegin; 2833 ierr = MatGetRowMaxAbs(a->A,v,idx);CHKERRQ(ierr); 2834 ierr = VecGetArray(v,&va);CHKERRQ(ierr); 2835 if (idx) { 2836 for (i=0; i<A->rmap->n; i++) { 2837 if (PetscAbsScalar(va[i])) idx[i] += A->cmap->rstart; 2838 } 2839 } 2840 2841 ierr = VecCreateSeq(PETSC_COMM_SELF,A->rmap->n,&vtmp);CHKERRQ(ierr); 2842 if (idx) { 2843 ierr = PetscMalloc1(A->rmap->n,&idxb);CHKERRQ(ierr); 2844 } 2845 ierr = MatGetRowMaxAbs(a->B,vtmp,idxb);CHKERRQ(ierr); 2846 ierr = VecGetArray(vtmp,&vb);CHKERRQ(ierr); 2847 2848 for (i=0; i<A->rmap->n; i++) { 2849 if (PetscAbsScalar(va[i]) < PetscAbsScalar(vb[i])) { 2850 va[i] = vb[i]; 2851 if (idx) idx[i] = a->garray[idxb[i]]; 2852 } 2853 } 2854 2855 ierr = VecRestoreArray(v,&va);CHKERRQ(ierr); 2856 ierr = VecRestoreArray(vtmp,&vb);CHKERRQ(ierr); 2857 ierr = PetscFree(idxb);CHKERRQ(ierr); 2858 ierr = VecDestroy(&vtmp);CHKERRQ(ierr); 2859 PetscFunctionReturn(0); 2860 } 2861 2862 #undef __FUNCT__ 2863 #define __FUNCT__ "MatGetRowMinAbs_MPIAIJ" 2864 PetscErrorCode MatGetRowMinAbs_MPIAIJ(Mat A, Vec v, PetscInt idx[]) 2865 { 2866 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 2867 PetscErrorCode ierr; 2868 PetscInt i,*idxb = 0; 2869 PetscScalar *va,*vb; 2870 Vec vtmp; 2871 2872 PetscFunctionBegin; 2873 ierr = MatGetRowMinAbs(a->A,v,idx);CHKERRQ(ierr); 2874 ierr = VecGetArray(v,&va);CHKERRQ(ierr); 2875 if (idx) { 2876 for (i=0; i<A->cmap->n; i++) { 2877 if (PetscAbsScalar(va[i])) idx[i] += A->cmap->rstart; 2878 } 2879 } 2880 2881 ierr = VecCreateSeq(PETSC_COMM_SELF,A->rmap->n,&vtmp);CHKERRQ(ierr); 2882 if (idx) { 2883 ierr = PetscMalloc1(A->rmap->n,&idxb);CHKERRQ(ierr); 2884 } 2885 ierr = MatGetRowMinAbs(a->B,vtmp,idxb);CHKERRQ(ierr); 2886 ierr = VecGetArray(vtmp,&vb);CHKERRQ(ierr); 2887 2888 for (i=0; i<A->rmap->n; i++) { 2889 if (PetscAbsScalar(va[i]) > PetscAbsScalar(vb[i])) { 2890 va[i] = vb[i]; 2891 if (idx) idx[i] = a->garray[idxb[i]]; 2892 } 2893 } 2894 2895 ierr = VecRestoreArray(v,&va);CHKERRQ(ierr); 2896 ierr = VecRestoreArray(vtmp,&vb);CHKERRQ(ierr); 2897 ierr = PetscFree(idxb);CHKERRQ(ierr); 2898 ierr = VecDestroy(&vtmp);CHKERRQ(ierr); 2899 PetscFunctionReturn(0); 2900 } 2901 2902 #undef __FUNCT__ 2903 #define __FUNCT__ "MatGetRowMin_MPIAIJ" 2904 PetscErrorCode MatGetRowMin_MPIAIJ(Mat A, Vec v, PetscInt idx[]) 2905 { 2906 Mat_MPIAIJ *mat = (Mat_MPIAIJ*) A->data; 2907 PetscInt n = A->rmap->n; 2908 PetscInt cstart = A->cmap->rstart; 2909 PetscInt *cmap = mat->garray; 2910 PetscInt *diagIdx, *offdiagIdx; 2911 Vec diagV, offdiagV; 2912 PetscScalar *a, *diagA, *offdiagA; 2913 PetscInt r; 2914 PetscErrorCode ierr; 2915 2916 PetscFunctionBegin; 2917 ierr = PetscMalloc2(n,&diagIdx,n,&offdiagIdx);CHKERRQ(ierr); 2918 ierr = VecCreateSeq(PetscObjectComm((PetscObject)A), n, &diagV);CHKERRQ(ierr); 2919 ierr = VecCreateSeq(PetscObjectComm((PetscObject)A), n, &offdiagV);CHKERRQ(ierr); 2920 ierr = MatGetRowMin(mat->A, diagV, diagIdx);CHKERRQ(ierr); 2921 ierr = MatGetRowMin(mat->B, offdiagV, offdiagIdx);CHKERRQ(ierr); 2922 ierr = VecGetArray(v, &a);CHKERRQ(ierr); 2923 ierr = VecGetArray(diagV, &diagA);CHKERRQ(ierr); 2924 ierr = VecGetArray(offdiagV, &offdiagA);CHKERRQ(ierr); 2925 for (r = 0; r < n; ++r) { 2926 if (PetscAbsScalar(diagA[r]) <= PetscAbsScalar(offdiagA[r])) { 2927 a[r] = diagA[r]; 2928 idx[r] = cstart + diagIdx[r]; 2929 } else { 2930 a[r] = offdiagA[r]; 2931 idx[r] = cmap[offdiagIdx[r]]; 2932 } 2933 } 2934 ierr = VecRestoreArray(v, &a);CHKERRQ(ierr); 2935 ierr = VecRestoreArray(diagV, &diagA);CHKERRQ(ierr); 2936 ierr = VecRestoreArray(offdiagV, &offdiagA);CHKERRQ(ierr); 2937 ierr = VecDestroy(&diagV);CHKERRQ(ierr); 2938 ierr = VecDestroy(&offdiagV);CHKERRQ(ierr); 2939 ierr = PetscFree2(diagIdx, offdiagIdx);CHKERRQ(ierr); 2940 PetscFunctionReturn(0); 2941 } 2942 2943 #undef __FUNCT__ 2944 #define __FUNCT__ "MatGetRowMax_MPIAIJ" 2945 PetscErrorCode MatGetRowMax_MPIAIJ(Mat A, Vec v, PetscInt idx[]) 2946 { 2947 Mat_MPIAIJ *mat = (Mat_MPIAIJ*) A->data; 2948 PetscInt n = A->rmap->n; 2949 PetscInt cstart = A->cmap->rstart; 2950 PetscInt *cmap = mat->garray; 2951 PetscInt *diagIdx, *offdiagIdx; 2952 Vec diagV, offdiagV; 2953 PetscScalar *a, *diagA, *offdiagA; 2954 PetscInt r; 2955 PetscErrorCode ierr; 2956 2957 PetscFunctionBegin; 2958 ierr = PetscMalloc2(n,&diagIdx,n,&offdiagIdx);CHKERRQ(ierr); 2959 ierr = VecCreateSeq(PETSC_COMM_SELF, n, &diagV);CHKERRQ(ierr); 2960 ierr = VecCreateSeq(PETSC_COMM_SELF, n, &offdiagV);CHKERRQ(ierr); 2961 ierr = MatGetRowMax(mat->A, diagV, diagIdx);CHKERRQ(ierr); 2962 ierr = MatGetRowMax(mat->B, offdiagV, offdiagIdx);CHKERRQ(ierr); 2963 ierr = VecGetArray(v, &a);CHKERRQ(ierr); 2964 ierr = VecGetArray(diagV, &diagA);CHKERRQ(ierr); 2965 ierr = VecGetArray(offdiagV, &offdiagA);CHKERRQ(ierr); 2966 for (r = 0; r < n; ++r) { 2967 if (PetscAbsScalar(diagA[r]) >= PetscAbsScalar(offdiagA[r])) { 2968 a[r] = diagA[r]; 2969 idx[r] = cstart + diagIdx[r]; 2970 } else { 2971 a[r] = offdiagA[r]; 2972 idx[r] = cmap[offdiagIdx[r]]; 2973 } 2974 } 2975 ierr = VecRestoreArray(v, &a);CHKERRQ(ierr); 2976 ierr = VecRestoreArray(diagV, &diagA);CHKERRQ(ierr); 2977 ierr = VecRestoreArray(offdiagV, &offdiagA);CHKERRQ(ierr); 2978 ierr = VecDestroy(&diagV);CHKERRQ(ierr); 2979 ierr = VecDestroy(&offdiagV);CHKERRQ(ierr); 2980 ierr = PetscFree2(diagIdx, offdiagIdx);CHKERRQ(ierr); 2981 PetscFunctionReturn(0); 2982 } 2983 2984 #undef __FUNCT__ 2985 #define __FUNCT__ "MatGetSeqNonzeroStructure_MPIAIJ" 2986 PetscErrorCode MatGetSeqNonzeroStructure_MPIAIJ(Mat mat,Mat *newmat) 2987 { 2988 PetscErrorCode ierr; 2989 Mat *dummy; 2990 2991 PetscFunctionBegin; 2992 ierr = MatGetSubMatrix_MPIAIJ_All(mat,MAT_DO_NOT_GET_VALUES,MAT_INITIAL_MATRIX,&dummy);CHKERRQ(ierr); 2993 *newmat = *dummy; 2994 ierr = PetscFree(dummy);CHKERRQ(ierr); 2995 PetscFunctionReturn(0); 2996 } 2997 2998 #undef __FUNCT__ 2999 #define __FUNCT__ "MatInvertBlockDiagonal_MPIAIJ" 3000 PetscErrorCode MatInvertBlockDiagonal_MPIAIJ(Mat A,const PetscScalar **values) 3001 { 3002 Mat_MPIAIJ *a = (Mat_MPIAIJ*) A->data; 3003 PetscErrorCode ierr; 3004 3005 PetscFunctionBegin; 3006 ierr = MatInvertBlockDiagonal(a->A,values);CHKERRQ(ierr); 3007 PetscFunctionReturn(0); 3008 } 3009 3010 #undef __FUNCT__ 3011 #define __FUNCT__ "MatSetRandom_MPIAIJ" 3012 static PetscErrorCode MatSetRandom_MPIAIJ(Mat x,PetscRandom rctx) 3013 { 3014 PetscErrorCode ierr; 3015 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)x->data; 3016 3017 PetscFunctionBegin; 3018 ierr = MatSetRandom(aij->A,rctx);CHKERRQ(ierr); 3019 ierr = MatSetRandom(aij->B,rctx);CHKERRQ(ierr); 3020 ierr = MatAssemblyBegin(x,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3021 ierr = MatAssemblyEnd(x,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3022 PetscFunctionReturn(0); 3023 } 3024 3025 /* -------------------------------------------------------------------*/ 3026 static struct _MatOps MatOps_Values = {MatSetValues_MPIAIJ, 3027 MatGetRow_MPIAIJ, 3028 MatRestoreRow_MPIAIJ, 3029 MatMult_MPIAIJ, 3030 /* 4*/ MatMultAdd_MPIAIJ, 3031 MatMultTranspose_MPIAIJ, 3032 MatMultTransposeAdd_MPIAIJ, 3033 #if defined(PETSC_HAVE_PBGL) 3034 MatSolve_MPIAIJ, 3035 #else 3036 0, 3037 #endif 3038 0, 3039 0, 3040 /*10*/ 0, 3041 0, 3042 0, 3043 MatSOR_MPIAIJ, 3044 MatTranspose_MPIAIJ, 3045 /*15*/ MatGetInfo_MPIAIJ, 3046 MatEqual_MPIAIJ, 3047 MatGetDiagonal_MPIAIJ, 3048 MatDiagonalScale_MPIAIJ, 3049 MatNorm_MPIAIJ, 3050 /*20*/ MatAssemblyBegin_MPIAIJ, 3051 MatAssemblyEnd_MPIAIJ, 3052 MatSetOption_MPIAIJ, 3053 MatZeroEntries_MPIAIJ, 3054 /*24*/ MatZeroRows_MPIAIJ, 3055 0, 3056 #if defined(PETSC_HAVE_PBGL) 3057 0, 3058 #else 3059 0, 3060 #endif 3061 0, 3062 0, 3063 /*29*/ MatSetUp_MPIAIJ, 3064 #if defined(PETSC_HAVE_PBGL) 3065 0, 3066 #else 3067 0, 3068 #endif 3069 0, 3070 0, 3071 0, 3072 /*34*/ MatDuplicate_MPIAIJ, 3073 0, 3074 0, 3075 0, 3076 0, 3077 /*39*/ MatAXPY_MPIAIJ, 3078 MatGetSubMatrices_MPIAIJ, 3079 MatIncreaseOverlap_MPIAIJ, 3080 MatGetValues_MPIAIJ, 3081 MatCopy_MPIAIJ, 3082 /*44*/ MatGetRowMax_MPIAIJ, 3083 MatScale_MPIAIJ, 3084 0, 3085 MatDiagonalSet_MPIAIJ, 3086 MatZeroRowsColumns_MPIAIJ, 3087 /*49*/ MatSetRandom_MPIAIJ, 3088 0, 3089 0, 3090 0, 3091 0, 3092 /*54*/ MatFDColoringCreate_MPIXAIJ, 3093 0, 3094 MatSetUnfactored_MPIAIJ, 3095 MatPermute_MPIAIJ, 3096 0, 3097 /*59*/ MatGetSubMatrix_MPIAIJ, 3098 MatDestroy_MPIAIJ, 3099 MatView_MPIAIJ, 3100 0, 3101 MatMatMatMult_MPIAIJ_MPIAIJ_MPIAIJ, 3102 /*64*/ MatMatMatMultSymbolic_MPIAIJ_MPIAIJ_MPIAIJ, 3103 MatMatMatMultNumeric_MPIAIJ_MPIAIJ_MPIAIJ, 3104 0, 3105 0, 3106 0, 3107 /*69*/ MatGetRowMaxAbs_MPIAIJ, 3108 MatGetRowMinAbs_MPIAIJ, 3109 0, 3110 MatSetColoring_MPIAIJ, 3111 0, 3112 MatSetValuesAdifor_MPIAIJ, 3113 /*75*/ MatFDColoringApply_AIJ, 3114 0, 3115 0, 3116 0, 3117 MatFindZeroDiagonals_MPIAIJ, 3118 /*80*/ 0, 3119 0, 3120 0, 3121 /*83*/ MatLoad_MPIAIJ, 3122 0, 3123 0, 3124 0, 3125 0, 3126 0, 3127 /*89*/ MatMatMult_MPIAIJ_MPIAIJ, 3128 MatMatMultSymbolic_MPIAIJ_MPIAIJ, 3129 MatMatMultNumeric_MPIAIJ_MPIAIJ, 3130 MatPtAP_MPIAIJ_MPIAIJ, 3131 MatPtAPSymbolic_MPIAIJ_MPIAIJ, 3132 /*94*/ MatPtAPNumeric_MPIAIJ_MPIAIJ, 3133 0, 3134 0, 3135 0, 3136 0, 3137 /*99*/ 0, 3138 0, 3139 0, 3140 MatConjugate_MPIAIJ, 3141 0, 3142 /*104*/MatSetValuesRow_MPIAIJ, 3143 MatRealPart_MPIAIJ, 3144 MatImaginaryPart_MPIAIJ, 3145 0, 3146 0, 3147 /*109*/0, 3148 MatGetRedundantMatrix_MPIAIJ, 3149 MatGetRowMin_MPIAIJ, 3150 0, 3151 0, 3152 /*114*/MatGetSeqNonzeroStructure_MPIAIJ, 3153 0, 3154 0, 3155 0, 3156 0, 3157 /*119*/0, 3158 0, 3159 0, 3160 0, 3161 MatGetMultiProcBlock_MPIAIJ, 3162 /*124*/MatFindNonzeroRows_MPIAIJ, 3163 MatGetColumnNorms_MPIAIJ, 3164 MatInvertBlockDiagonal_MPIAIJ, 3165 0, 3166 MatGetSubMatricesParallel_MPIAIJ, 3167 /*129*/0, 3168 MatTransposeMatMult_MPIAIJ_MPIAIJ, 3169 MatTransposeMatMultSymbolic_MPIAIJ_MPIAIJ, 3170 MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ, 3171 0, 3172 /*134*/0, 3173 0, 3174 0, 3175 0, 3176 0, 3177 /*139*/0, 3178 0, 3179 0, 3180 MatFDColoringSetUp_MPIXAIJ, 3181 0, 3182 /*144*/MatCreateMPIMatConcatenateSeqMat_MPIAIJ 3183 }; 3184 3185 /* ----------------------------------------------------------------------------------------*/ 3186 3187 #undef __FUNCT__ 3188 #define __FUNCT__ "MatStoreValues_MPIAIJ" 3189 PetscErrorCode MatStoreValues_MPIAIJ(Mat mat) 3190 { 3191 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 3192 PetscErrorCode ierr; 3193 3194 PetscFunctionBegin; 3195 ierr = MatStoreValues(aij->A);CHKERRQ(ierr); 3196 ierr = MatStoreValues(aij->B);CHKERRQ(ierr); 3197 PetscFunctionReturn(0); 3198 } 3199 3200 #undef __FUNCT__ 3201 #define __FUNCT__ "MatRetrieveValues_MPIAIJ" 3202 PetscErrorCode MatRetrieveValues_MPIAIJ(Mat mat) 3203 { 3204 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 3205 PetscErrorCode ierr; 3206 3207 PetscFunctionBegin; 3208 ierr = MatRetrieveValues(aij->A);CHKERRQ(ierr); 3209 ierr = MatRetrieveValues(aij->B);CHKERRQ(ierr); 3210 PetscFunctionReturn(0); 3211 } 3212 3213 #undef __FUNCT__ 3214 #define __FUNCT__ "MatMPIAIJSetPreallocation_MPIAIJ" 3215 PetscErrorCode MatMPIAIJSetPreallocation_MPIAIJ(Mat B,PetscInt d_nz,const PetscInt d_nnz[],PetscInt o_nz,const PetscInt o_nnz[]) 3216 { 3217 Mat_MPIAIJ *b; 3218 PetscErrorCode ierr; 3219 3220 PetscFunctionBegin; 3221 ierr = PetscLayoutSetUp(B->rmap);CHKERRQ(ierr); 3222 ierr = PetscLayoutSetUp(B->cmap);CHKERRQ(ierr); 3223 b = (Mat_MPIAIJ*)B->data; 3224 3225 if (!B->preallocated) { 3226 /* Explicitly create 2 MATSEQAIJ matrices. */ 3227 ierr = MatCreate(PETSC_COMM_SELF,&b->A);CHKERRQ(ierr); 3228 ierr = MatSetSizes(b->A,B->rmap->n,B->cmap->n,B->rmap->n,B->cmap->n);CHKERRQ(ierr); 3229 ierr = MatSetBlockSizesFromMats(b->A,B,B);CHKERRQ(ierr); 3230 ierr = MatSetType(b->A,MATSEQAIJ);CHKERRQ(ierr); 3231 ierr = PetscLogObjectParent((PetscObject)B,(PetscObject)b->A);CHKERRQ(ierr); 3232 ierr = MatCreate(PETSC_COMM_SELF,&b->B);CHKERRQ(ierr); 3233 ierr = MatSetSizes(b->B,B->rmap->n,B->cmap->N,B->rmap->n,B->cmap->N);CHKERRQ(ierr); 3234 ierr = MatSetBlockSizesFromMats(b->B,B,B);CHKERRQ(ierr); 3235 ierr = MatSetType(b->B,MATSEQAIJ);CHKERRQ(ierr); 3236 ierr = PetscLogObjectParent((PetscObject)B,(PetscObject)b->B);CHKERRQ(ierr); 3237 } 3238 3239 ierr = MatSeqAIJSetPreallocation(b->A,d_nz,d_nnz);CHKERRQ(ierr); 3240 ierr = MatSeqAIJSetPreallocation(b->B,o_nz,o_nnz);CHKERRQ(ierr); 3241 B->preallocated = PETSC_TRUE; 3242 PetscFunctionReturn(0); 3243 } 3244 3245 #undef __FUNCT__ 3246 #define __FUNCT__ "MatDuplicate_MPIAIJ" 3247 PetscErrorCode MatDuplicate_MPIAIJ(Mat matin,MatDuplicateOption cpvalues,Mat *newmat) 3248 { 3249 Mat mat; 3250 Mat_MPIAIJ *a,*oldmat = (Mat_MPIAIJ*)matin->data; 3251 PetscErrorCode ierr; 3252 3253 PetscFunctionBegin; 3254 *newmat = 0; 3255 ierr = MatCreate(PetscObjectComm((PetscObject)matin),&mat);CHKERRQ(ierr); 3256 ierr = MatSetSizes(mat,matin->rmap->n,matin->cmap->n,matin->rmap->N,matin->cmap->N);CHKERRQ(ierr); 3257 ierr = MatSetBlockSizesFromMats(mat,matin,matin);CHKERRQ(ierr); 3258 ierr = MatSetType(mat,((PetscObject)matin)->type_name);CHKERRQ(ierr); 3259 ierr = PetscMemcpy(mat->ops,matin->ops,sizeof(struct _MatOps));CHKERRQ(ierr); 3260 a = (Mat_MPIAIJ*)mat->data; 3261 3262 mat->factortype = matin->factortype; 3263 mat->assembled = PETSC_TRUE; 3264 mat->insertmode = NOT_SET_VALUES; 3265 mat->preallocated = PETSC_TRUE; 3266 3267 a->size = oldmat->size; 3268 a->rank = oldmat->rank; 3269 a->donotstash = oldmat->donotstash; 3270 a->roworiented = oldmat->roworiented; 3271 a->rowindices = 0; 3272 a->rowvalues = 0; 3273 a->getrowactive = PETSC_FALSE; 3274 3275 ierr = PetscLayoutReference(matin->rmap,&mat->rmap);CHKERRQ(ierr); 3276 ierr = PetscLayoutReference(matin->cmap,&mat->cmap);CHKERRQ(ierr); 3277 3278 if (oldmat->colmap) { 3279 #if defined(PETSC_USE_CTABLE) 3280 ierr = PetscTableCreateCopy(oldmat->colmap,&a->colmap);CHKERRQ(ierr); 3281 #else 3282 ierr = PetscMalloc1((mat->cmap->N),&a->colmap);CHKERRQ(ierr); 3283 ierr = PetscLogObjectMemory((PetscObject)mat,(mat->cmap->N)*sizeof(PetscInt));CHKERRQ(ierr); 3284 ierr = PetscMemcpy(a->colmap,oldmat->colmap,(mat->cmap->N)*sizeof(PetscInt));CHKERRQ(ierr); 3285 #endif 3286 } else a->colmap = 0; 3287 if (oldmat->garray) { 3288 PetscInt len; 3289 len = oldmat->B->cmap->n; 3290 ierr = PetscMalloc1((len+1),&a->garray);CHKERRQ(ierr); 3291 ierr = PetscLogObjectMemory((PetscObject)mat,len*sizeof(PetscInt));CHKERRQ(ierr); 3292 if (len) { ierr = PetscMemcpy(a->garray,oldmat->garray,len*sizeof(PetscInt));CHKERRQ(ierr); } 3293 } else a->garray = 0; 3294 3295 ierr = VecDuplicate(oldmat->lvec,&a->lvec);CHKERRQ(ierr); 3296 ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)a->lvec);CHKERRQ(ierr); 3297 ierr = VecScatterCopy(oldmat->Mvctx,&a->Mvctx);CHKERRQ(ierr); 3298 ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)a->Mvctx);CHKERRQ(ierr); 3299 ierr = MatDuplicate(oldmat->A,cpvalues,&a->A);CHKERRQ(ierr); 3300 ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)a->A);CHKERRQ(ierr); 3301 ierr = MatDuplicate(oldmat->B,cpvalues,&a->B);CHKERRQ(ierr); 3302 ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)a->B);CHKERRQ(ierr); 3303 ierr = PetscFunctionListDuplicate(((PetscObject)matin)->qlist,&((PetscObject)mat)->qlist);CHKERRQ(ierr); 3304 *newmat = mat; 3305 PetscFunctionReturn(0); 3306 } 3307 3308 3309 3310 #undef __FUNCT__ 3311 #define __FUNCT__ "MatLoad_MPIAIJ" 3312 PetscErrorCode MatLoad_MPIAIJ(Mat newMat, PetscViewer viewer) 3313 { 3314 PetscScalar *vals,*svals; 3315 MPI_Comm comm; 3316 PetscErrorCode ierr; 3317 PetscMPIInt rank,size,tag = ((PetscObject)viewer)->tag; 3318 PetscInt i,nz,j,rstart,rend,mmax,maxnz = 0,grows,gcols; 3319 PetscInt header[4],*rowlengths = 0,M,N,m,*cols; 3320 PetscInt *ourlens = NULL,*procsnz = NULL,*offlens = NULL,jj,*mycols,*smycols; 3321 PetscInt cend,cstart,n,*rowners,sizesset=1; 3322 int fd; 3323 PetscInt bs = 1; 3324 3325 PetscFunctionBegin; 3326 ierr = PetscObjectGetComm((PetscObject)viewer,&comm);CHKERRQ(ierr); 3327 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 3328 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 3329 if (!rank) { 3330 ierr = PetscViewerBinaryGetDescriptor(viewer,&fd);CHKERRQ(ierr); 3331 ierr = PetscBinaryRead(fd,(char*)header,4,PETSC_INT);CHKERRQ(ierr); 3332 if (header[0] != MAT_FILE_CLASSID) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED,"not matrix object"); 3333 } 3334 3335 ierr = PetscOptionsBegin(comm,NULL,"Options for loading SEQAIJ matrix","Mat");CHKERRQ(ierr); 3336 ierr = PetscOptionsInt("-matload_block_size","Set the blocksize used to store the matrix","MatLoad",bs,&bs,NULL);CHKERRQ(ierr); 3337 ierr = PetscOptionsEnd();CHKERRQ(ierr); 3338 3339 if (newMat->rmap->n < 0 && newMat->rmap->N < 0 && newMat->cmap->n < 0 && newMat->cmap->N < 0) sizesset = 0; 3340 3341 ierr = MPI_Bcast(header+1,3,MPIU_INT,0,comm);CHKERRQ(ierr); 3342 M = header[1]; N = header[2]; 3343 /* If global rows/cols are set to PETSC_DECIDE, set it to the sizes given in the file */ 3344 if (sizesset && newMat->rmap->N < 0) newMat->rmap->N = M; 3345 if (sizesset && newMat->cmap->N < 0) newMat->cmap->N = N; 3346 3347 /* If global sizes are set, check if they are consistent with that given in the file */ 3348 if (sizesset) { 3349 ierr = MatGetSize(newMat,&grows,&gcols);CHKERRQ(ierr); 3350 } 3351 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); 3352 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); 3353 3354 /* determine ownership of all (block) rows */ 3355 if (M%bs) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED, "Inconsistent # of rows (%d) and block size (%d)",M,bs); 3356 if (newMat->rmap->n < 0) m = bs*((M/bs)/size + (((M/bs) % size) > rank)); /* PETSC_DECIDE */ 3357 else m = newMat->rmap->n; /* Set by user */ 3358 3359 ierr = PetscMalloc1((size+1),&rowners);CHKERRQ(ierr); 3360 ierr = MPI_Allgather(&m,1,MPIU_INT,rowners+1,1,MPIU_INT,comm);CHKERRQ(ierr); 3361 3362 /* First process needs enough room for process with most rows */ 3363 if (!rank) { 3364 mmax = rowners[1]; 3365 for (i=2; i<=size; i++) { 3366 mmax = PetscMax(mmax, rowners[i]); 3367 } 3368 } else mmax = -1; /* unused, but compilers complain */ 3369 3370 rowners[0] = 0; 3371 for (i=2; i<=size; i++) { 3372 rowners[i] += rowners[i-1]; 3373 } 3374 rstart = rowners[rank]; 3375 rend = rowners[rank+1]; 3376 3377 /* distribute row lengths to all processors */ 3378 ierr = PetscMalloc2(m,&ourlens,m,&offlens);CHKERRQ(ierr); 3379 if (!rank) { 3380 ierr = PetscBinaryRead(fd,ourlens,m,PETSC_INT);CHKERRQ(ierr); 3381 ierr = PetscMalloc1(mmax,&rowlengths);CHKERRQ(ierr); 3382 ierr = PetscCalloc1(size,&procsnz);CHKERRQ(ierr); 3383 for (j=0; j<m; j++) { 3384 procsnz[0] += ourlens[j]; 3385 } 3386 for (i=1; i<size; i++) { 3387 ierr = PetscBinaryRead(fd,rowlengths,rowners[i+1]-rowners[i],PETSC_INT);CHKERRQ(ierr); 3388 /* calculate the number of nonzeros on each processor */ 3389 for (j=0; j<rowners[i+1]-rowners[i]; j++) { 3390 procsnz[i] += rowlengths[j]; 3391 } 3392 ierr = MPIULong_Send(rowlengths,rowners[i+1]-rowners[i],MPIU_INT,i,tag,comm);CHKERRQ(ierr); 3393 } 3394 ierr = PetscFree(rowlengths);CHKERRQ(ierr); 3395 } else { 3396 ierr = MPIULong_Recv(ourlens,m,MPIU_INT,0,tag,comm);CHKERRQ(ierr); 3397 } 3398 3399 if (!rank) { 3400 /* determine max buffer needed and allocate it */ 3401 maxnz = 0; 3402 for (i=0; i<size; i++) { 3403 maxnz = PetscMax(maxnz,procsnz[i]); 3404 } 3405 ierr = PetscMalloc1(maxnz,&cols);CHKERRQ(ierr); 3406 3407 /* read in my part of the matrix column indices */ 3408 nz = procsnz[0]; 3409 ierr = PetscMalloc1(nz,&mycols);CHKERRQ(ierr); 3410 ierr = PetscBinaryRead(fd,mycols,nz,PETSC_INT);CHKERRQ(ierr); 3411 3412 /* read in every one elses and ship off */ 3413 for (i=1; i<size; i++) { 3414 nz = procsnz[i]; 3415 ierr = PetscBinaryRead(fd,cols,nz,PETSC_INT);CHKERRQ(ierr); 3416 ierr = MPIULong_Send(cols,nz,MPIU_INT,i,tag,comm);CHKERRQ(ierr); 3417 } 3418 ierr = PetscFree(cols);CHKERRQ(ierr); 3419 } else { 3420 /* determine buffer space needed for message */ 3421 nz = 0; 3422 for (i=0; i<m; i++) { 3423 nz += ourlens[i]; 3424 } 3425 ierr = PetscMalloc1(nz,&mycols);CHKERRQ(ierr); 3426 3427 /* receive message of column indices*/ 3428 ierr = MPIULong_Recv(mycols,nz,MPIU_INT,0,tag,comm);CHKERRQ(ierr); 3429 } 3430 3431 /* determine column ownership if matrix is not square */ 3432 if (N != M) { 3433 if (newMat->cmap->n < 0) n = N/size + ((N % size) > rank); 3434 else n = newMat->cmap->n; 3435 ierr = MPI_Scan(&n,&cend,1,MPIU_INT,MPI_SUM,comm);CHKERRQ(ierr); 3436 cstart = cend - n; 3437 } else { 3438 cstart = rstart; 3439 cend = rend; 3440 n = cend - cstart; 3441 } 3442 3443 /* loop over local rows, determining number of off diagonal entries */ 3444 ierr = PetscMemzero(offlens,m*sizeof(PetscInt));CHKERRQ(ierr); 3445 jj = 0; 3446 for (i=0; i<m; i++) { 3447 for (j=0; j<ourlens[i]; j++) { 3448 if (mycols[jj] < cstart || mycols[jj] >= cend) offlens[i]++; 3449 jj++; 3450 } 3451 } 3452 3453 for (i=0; i<m; i++) { 3454 ourlens[i] -= offlens[i]; 3455 } 3456 if (!sizesset) { 3457 ierr = MatSetSizes(newMat,m,n,M,N);CHKERRQ(ierr); 3458 } 3459 3460 if (bs > 1) {ierr = MatSetBlockSize(newMat,bs);CHKERRQ(ierr);} 3461 3462 ierr = MatMPIAIJSetPreallocation(newMat,0,ourlens,0,offlens);CHKERRQ(ierr); 3463 3464 for (i=0; i<m; i++) { 3465 ourlens[i] += offlens[i]; 3466 } 3467 3468 if (!rank) { 3469 ierr = PetscMalloc1((maxnz+1),&vals);CHKERRQ(ierr); 3470 3471 /* read in my part of the matrix numerical values */ 3472 nz = procsnz[0]; 3473 ierr = PetscBinaryRead(fd,vals,nz,PETSC_SCALAR);CHKERRQ(ierr); 3474 3475 /* insert into matrix */ 3476 jj = rstart; 3477 smycols = mycols; 3478 svals = vals; 3479 for (i=0; i<m; i++) { 3480 ierr = MatSetValues_MPIAIJ(newMat,1,&jj,ourlens[i],smycols,svals,INSERT_VALUES);CHKERRQ(ierr); 3481 smycols += ourlens[i]; 3482 svals += ourlens[i]; 3483 jj++; 3484 } 3485 3486 /* read in other processors and ship out */ 3487 for (i=1; i<size; i++) { 3488 nz = procsnz[i]; 3489 ierr = PetscBinaryRead(fd,vals,nz,PETSC_SCALAR);CHKERRQ(ierr); 3490 ierr = MPIULong_Send(vals,nz,MPIU_SCALAR,i,((PetscObject)newMat)->tag,comm);CHKERRQ(ierr); 3491 } 3492 ierr = PetscFree(procsnz);CHKERRQ(ierr); 3493 } else { 3494 /* receive numeric values */ 3495 ierr = PetscMalloc1((nz+1),&vals);CHKERRQ(ierr); 3496 3497 /* receive message of values*/ 3498 ierr = MPIULong_Recv(vals,nz,MPIU_SCALAR,0,((PetscObject)newMat)->tag,comm);CHKERRQ(ierr); 3499 3500 /* insert into matrix */ 3501 jj = rstart; 3502 smycols = mycols; 3503 svals = vals; 3504 for (i=0; i<m; i++) { 3505 ierr = MatSetValues_MPIAIJ(newMat,1,&jj,ourlens[i],smycols,svals,INSERT_VALUES);CHKERRQ(ierr); 3506 smycols += ourlens[i]; 3507 svals += ourlens[i]; 3508 jj++; 3509 } 3510 } 3511 ierr = PetscFree2(ourlens,offlens);CHKERRQ(ierr); 3512 ierr = PetscFree(vals);CHKERRQ(ierr); 3513 ierr = PetscFree(mycols);CHKERRQ(ierr); 3514 ierr = PetscFree(rowners);CHKERRQ(ierr); 3515 ierr = MatAssemblyBegin(newMat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3516 ierr = MatAssemblyEnd(newMat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3517 PetscFunctionReturn(0); 3518 } 3519 3520 #undef __FUNCT__ 3521 #define __FUNCT__ "MatGetSubMatrix_MPIAIJ" 3522 PetscErrorCode MatGetSubMatrix_MPIAIJ(Mat mat,IS isrow,IS iscol,MatReuse call,Mat *newmat) 3523 { 3524 PetscErrorCode ierr; 3525 IS iscol_local; 3526 PetscInt csize; 3527 3528 PetscFunctionBegin; 3529 ierr = ISGetLocalSize(iscol,&csize);CHKERRQ(ierr); 3530 if (call == MAT_REUSE_MATRIX) { 3531 ierr = PetscObjectQuery((PetscObject)*newmat,"ISAllGather",(PetscObject*)&iscol_local);CHKERRQ(ierr); 3532 if (!iscol_local) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Submatrix passed in was not used before, cannot reuse"); 3533 } else { 3534 PetscInt cbs; 3535 ierr = ISGetBlockSize(iscol,&cbs);CHKERRQ(ierr); 3536 ierr = ISAllGather(iscol,&iscol_local);CHKERRQ(ierr); 3537 ierr = ISSetBlockSize(iscol_local,cbs);CHKERRQ(ierr); 3538 } 3539 ierr = MatGetSubMatrix_MPIAIJ_Private(mat,isrow,iscol_local,csize,call,newmat);CHKERRQ(ierr); 3540 if (call == MAT_INITIAL_MATRIX) { 3541 ierr = PetscObjectCompose((PetscObject)*newmat,"ISAllGather",(PetscObject)iscol_local);CHKERRQ(ierr); 3542 ierr = ISDestroy(&iscol_local);CHKERRQ(ierr); 3543 } 3544 PetscFunctionReturn(0); 3545 } 3546 3547 extern PetscErrorCode MatGetSubMatrices_MPIAIJ_Local(Mat,PetscInt,const IS[],const IS[],MatReuse,PetscBool*,Mat*); 3548 #undef __FUNCT__ 3549 #define __FUNCT__ "MatGetSubMatrix_MPIAIJ_Private" 3550 /* 3551 Not great since it makes two copies of the submatrix, first an SeqAIJ 3552 in local and then by concatenating the local matrices the end result. 3553 Writing it directly would be much like MatGetSubMatrices_MPIAIJ() 3554 3555 Note: This requires a sequential iscol with all indices. 3556 */ 3557 PetscErrorCode MatGetSubMatrix_MPIAIJ_Private(Mat mat,IS isrow,IS iscol,PetscInt csize,MatReuse call,Mat *newmat) 3558 { 3559 PetscErrorCode ierr; 3560 PetscMPIInt rank,size; 3561 PetscInt i,m,n,rstart,row,rend,nz,*cwork,j,bs,cbs; 3562 PetscInt *ii,*jj,nlocal,*dlens,*olens,dlen,olen,jend,mglobal,ncol; 3563 PetscBool allcolumns, colflag; 3564 Mat M,Mreuse; 3565 MatScalar *vwork,*aa; 3566 MPI_Comm comm; 3567 Mat_SeqAIJ *aij; 3568 3569 PetscFunctionBegin; 3570 ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 3571 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 3572 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 3573 3574 ierr = ISIdentity(iscol,&colflag);CHKERRQ(ierr); 3575 ierr = ISGetLocalSize(iscol,&ncol);CHKERRQ(ierr); 3576 if (colflag && ncol == mat->cmap->N) { 3577 allcolumns = PETSC_TRUE; 3578 } else { 3579 allcolumns = PETSC_FALSE; 3580 } 3581 if (call == MAT_REUSE_MATRIX) { 3582 ierr = PetscObjectQuery((PetscObject)*newmat,"SubMatrix",(PetscObject*)&Mreuse);CHKERRQ(ierr); 3583 if (!Mreuse) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Submatrix passed in was not used before, cannot reuse"); 3584 ierr = MatGetSubMatrices_MPIAIJ_Local(mat,1,&isrow,&iscol,MAT_REUSE_MATRIX,&allcolumns,&Mreuse);CHKERRQ(ierr); 3585 } else { 3586 ierr = MatGetSubMatrices_MPIAIJ_Local(mat,1,&isrow,&iscol,MAT_INITIAL_MATRIX,&allcolumns,&Mreuse);CHKERRQ(ierr); 3587 } 3588 3589 /* 3590 m - number of local rows 3591 n - number of columns (same on all processors) 3592 rstart - first row in new global matrix generated 3593 */ 3594 ierr = MatGetSize(Mreuse,&m,&n);CHKERRQ(ierr); 3595 ierr = MatGetBlockSizes(Mreuse,&bs,&cbs);CHKERRQ(ierr); 3596 if (call == MAT_INITIAL_MATRIX) { 3597 aij = (Mat_SeqAIJ*)(Mreuse)->data; 3598 ii = aij->i; 3599 jj = aij->j; 3600 3601 /* 3602 Determine the number of non-zeros in the diagonal and off-diagonal 3603 portions of the matrix in order to do correct preallocation 3604 */ 3605 3606 /* first get start and end of "diagonal" columns */ 3607 if (csize == PETSC_DECIDE) { 3608 ierr = ISGetSize(isrow,&mglobal);CHKERRQ(ierr); 3609 if (mglobal == n) { /* square matrix */ 3610 nlocal = m; 3611 } else { 3612 nlocal = n/size + ((n % size) > rank); 3613 } 3614 } else { 3615 nlocal = csize; 3616 } 3617 ierr = MPI_Scan(&nlocal,&rend,1,MPIU_INT,MPI_SUM,comm);CHKERRQ(ierr); 3618 rstart = rend - nlocal; 3619 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); 3620 3621 /* next, compute all the lengths */ 3622 ierr = PetscMalloc1((2*m+1),&dlens);CHKERRQ(ierr); 3623 olens = dlens + m; 3624 for (i=0; i<m; i++) { 3625 jend = ii[i+1] - ii[i]; 3626 olen = 0; 3627 dlen = 0; 3628 for (j=0; j<jend; j++) { 3629 if (*jj < rstart || *jj >= rend) olen++; 3630 else dlen++; 3631 jj++; 3632 } 3633 olens[i] = olen; 3634 dlens[i] = dlen; 3635 } 3636 ierr = MatCreate(comm,&M);CHKERRQ(ierr); 3637 ierr = MatSetSizes(M,m,nlocal,PETSC_DECIDE,n);CHKERRQ(ierr); 3638 ierr = MatSetBlockSizes(M,bs,cbs);CHKERRQ(ierr); 3639 ierr = MatSetType(M,((PetscObject)mat)->type_name);CHKERRQ(ierr); 3640 ierr = MatMPIAIJSetPreallocation(M,0,dlens,0,olens);CHKERRQ(ierr); 3641 ierr = PetscFree(dlens);CHKERRQ(ierr); 3642 } else { 3643 PetscInt ml,nl; 3644 3645 M = *newmat; 3646 ierr = MatGetLocalSize(M,&ml,&nl);CHKERRQ(ierr); 3647 if (ml != m) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Previous matrix must be same size/layout as request"); 3648 ierr = MatZeroEntries(M);CHKERRQ(ierr); 3649 /* 3650 The next two lines are needed so we may call MatSetValues_MPIAIJ() below directly, 3651 rather than the slower MatSetValues(). 3652 */ 3653 M->was_assembled = PETSC_TRUE; 3654 M->assembled = PETSC_FALSE; 3655 } 3656 ierr = MatGetOwnershipRange(M,&rstart,&rend);CHKERRQ(ierr); 3657 aij = (Mat_SeqAIJ*)(Mreuse)->data; 3658 ii = aij->i; 3659 jj = aij->j; 3660 aa = aij->a; 3661 for (i=0; i<m; i++) { 3662 row = rstart + i; 3663 nz = ii[i+1] - ii[i]; 3664 cwork = jj; jj += nz; 3665 vwork = aa; aa += nz; 3666 ierr = MatSetValues_MPIAIJ(M,1,&row,nz,cwork,vwork,INSERT_VALUES);CHKERRQ(ierr); 3667 } 3668 3669 ierr = MatAssemblyBegin(M,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3670 ierr = MatAssemblyEnd(M,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3671 *newmat = M; 3672 3673 /* save submatrix used in processor for next request */ 3674 if (call == MAT_INITIAL_MATRIX) { 3675 ierr = PetscObjectCompose((PetscObject)M,"SubMatrix",(PetscObject)Mreuse);CHKERRQ(ierr); 3676 ierr = MatDestroy(&Mreuse);CHKERRQ(ierr); 3677 } 3678 PetscFunctionReturn(0); 3679 } 3680 3681 #undef __FUNCT__ 3682 #define __FUNCT__ "MatMPIAIJSetPreallocationCSR_MPIAIJ" 3683 PetscErrorCode MatMPIAIJSetPreallocationCSR_MPIAIJ(Mat B,const PetscInt Ii[],const PetscInt J[],const PetscScalar v[]) 3684 { 3685 PetscInt m,cstart, cend,j,nnz,i,d; 3686 PetscInt *d_nnz,*o_nnz,nnz_max = 0,rstart,ii; 3687 const PetscInt *JJ; 3688 PetscScalar *values; 3689 PetscErrorCode ierr; 3690 3691 PetscFunctionBegin; 3692 if (Ii[0]) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Ii[0] must be 0 it is %D",Ii[0]); 3693 3694 ierr = PetscLayoutSetUp(B->rmap);CHKERRQ(ierr); 3695 ierr = PetscLayoutSetUp(B->cmap);CHKERRQ(ierr); 3696 m = B->rmap->n; 3697 cstart = B->cmap->rstart; 3698 cend = B->cmap->rend; 3699 rstart = B->rmap->rstart; 3700 3701 ierr = PetscMalloc2(m,&d_nnz,m,&o_nnz);CHKERRQ(ierr); 3702 3703 #if defined(PETSC_USE_DEBUGGING) 3704 for (i=0; i<m; i++) { 3705 nnz = Ii[i+1]- Ii[i]; 3706 JJ = J + Ii[i]; 3707 if (nnz < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Local row %D has a negative %D number of columns",i,nnz); 3708 if (nnz && (JJ[0] < 0)) SETERRRQ1(PETSC_ERR_ARG_WRONGSTATE,"Row %D starts with negative column index",i,j); 3709 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); 3710 } 3711 #endif 3712 3713 for (i=0; i<m; i++) { 3714 nnz = Ii[i+1]- Ii[i]; 3715 JJ = J + Ii[i]; 3716 nnz_max = PetscMax(nnz_max,nnz); 3717 d = 0; 3718 for (j=0; j<nnz; j++) { 3719 if (cstart <= JJ[j] && JJ[j] < cend) d++; 3720 } 3721 d_nnz[i] = d; 3722 o_nnz[i] = nnz - d; 3723 } 3724 ierr = MatMPIAIJSetPreallocation(B,0,d_nnz,0,o_nnz);CHKERRQ(ierr); 3725 ierr = PetscFree2(d_nnz,o_nnz);CHKERRQ(ierr); 3726 3727 if (v) values = (PetscScalar*)v; 3728 else { 3729 ierr = PetscCalloc1((nnz_max+1),&values);CHKERRQ(ierr); 3730 } 3731 3732 for (i=0; i<m; i++) { 3733 ii = i + rstart; 3734 nnz = Ii[i+1]- Ii[i]; 3735 ierr = MatSetValues_MPIAIJ(B,1,&ii,nnz,J+Ii[i],values+(v ? Ii[i] : 0),INSERT_VALUES);CHKERRQ(ierr); 3736 } 3737 ierr = MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3738 ierr = MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3739 3740 if (!v) { 3741 ierr = PetscFree(values);CHKERRQ(ierr); 3742 } 3743 ierr = MatSetOption(B,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE);CHKERRQ(ierr); 3744 PetscFunctionReturn(0); 3745 } 3746 3747 #undef __FUNCT__ 3748 #define __FUNCT__ "MatMPIAIJSetPreallocationCSR" 3749 /*@ 3750 MatMPIAIJSetPreallocationCSR - Allocates memory for a sparse parallel matrix in AIJ format 3751 (the default parallel PETSc format). 3752 3753 Collective on MPI_Comm 3754 3755 Input Parameters: 3756 + B - the matrix 3757 . i - the indices into j for the start of each local row (starts with zero) 3758 . j - the column indices for each local row (starts with zero) 3759 - v - optional values in the matrix 3760 3761 Level: developer 3762 3763 Notes: 3764 The i, j, and a arrays ARE copied by this routine into the internal format used by PETSc; 3765 thus you CANNOT change the matrix entries by changing the values of a[] after you have 3766 called this routine. Use MatCreateMPIAIJWithSplitArrays() to avoid needing to copy the arrays. 3767 3768 The i and j indices are 0 based, and i indices are indices corresponding to the local j array. 3769 3770 The format which is used for the sparse matrix input, is equivalent to a 3771 row-major ordering.. i.e for the following matrix, the input data expected is 3772 as shown: 3773 3774 1 0 0 3775 2 0 3 P0 3776 ------- 3777 4 5 6 P1 3778 3779 Process0 [P0]: rows_owned=[0,1] 3780 i = {0,1,3} [size = nrow+1 = 2+1] 3781 j = {0,0,2} [size = nz = 6] 3782 v = {1,2,3} [size = nz = 6] 3783 3784 Process1 [P1]: rows_owned=[2] 3785 i = {0,3} [size = nrow+1 = 1+1] 3786 j = {0,1,2} [size = nz = 6] 3787 v = {4,5,6} [size = nz = 6] 3788 3789 .keywords: matrix, aij, compressed row, sparse, parallel 3790 3791 .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatCreateAIJ(), MPIAIJ, 3792 MatCreateSeqAIJWithArrays(), MatCreateMPIAIJWithSplitArrays() 3793 @*/ 3794 PetscErrorCode MatMPIAIJSetPreallocationCSR(Mat B,const PetscInt i[],const PetscInt j[], const PetscScalar v[]) 3795 { 3796 PetscErrorCode ierr; 3797 3798 PetscFunctionBegin; 3799 ierr = PetscTryMethod(B,"MatMPIAIJSetPreallocationCSR_C",(Mat,const PetscInt[],const PetscInt[],const PetscScalar[]),(B,i,j,v));CHKERRQ(ierr); 3800 PetscFunctionReturn(0); 3801 } 3802 3803 #undef __FUNCT__ 3804 #define __FUNCT__ "MatMPIAIJSetPreallocation" 3805 /*@C 3806 MatMPIAIJSetPreallocation - Preallocates memory for a sparse parallel matrix in AIJ format 3807 (the default parallel PETSc format). For good matrix assembly performance 3808 the user should preallocate the matrix storage by setting the parameters 3809 d_nz (or d_nnz) and o_nz (or o_nnz). By setting these parameters accurately, 3810 performance can be increased by more than a factor of 50. 3811 3812 Collective on MPI_Comm 3813 3814 Input Parameters: 3815 + B - the matrix 3816 . d_nz - number of nonzeros per row in DIAGONAL portion of local submatrix 3817 (same value is used for all local rows) 3818 . d_nnz - array containing the number of nonzeros in the various rows of the 3819 DIAGONAL portion of the local submatrix (possibly different for each row) 3820 or NULL, if d_nz is used to specify the nonzero structure. 3821 The size of this array is equal to the number of local rows, i.e 'm'. 3822 For matrices that will be factored, you must leave room for (and set) 3823 the diagonal entry even if it is zero. 3824 . o_nz - number of nonzeros per row in the OFF-DIAGONAL portion of local 3825 submatrix (same value is used for all local rows). 3826 - o_nnz - array containing the number of nonzeros in the various rows of the 3827 OFF-DIAGONAL portion of the local submatrix (possibly different for 3828 each row) or NULL, if o_nz is used to specify the nonzero 3829 structure. The size of this array is equal to the number 3830 of local rows, i.e 'm'. 3831 3832 If the *_nnz parameter is given then the *_nz parameter is ignored 3833 3834 The AIJ format (also called the Yale sparse matrix format or 3835 compressed row storage (CSR)), is fully compatible with standard Fortran 77 3836 storage. The stored row and column indices begin with zero. 3837 See Users-Manual: ch_mat for details. 3838 3839 The parallel matrix is partitioned such that the first m0 rows belong to 3840 process 0, the next m1 rows belong to process 1, the next m2 rows belong 3841 to process 2 etc.. where m0,m1,m2... are the input parameter 'm'. 3842 3843 The DIAGONAL portion of the local submatrix of a processor can be defined 3844 as the submatrix which is obtained by extraction the part corresponding to 3845 the rows r1-r2 and columns c1-c2 of the global matrix, where r1 is the 3846 first row that belongs to the processor, r2 is the last row belonging to 3847 the this processor, and c1-c2 is range of indices of the local part of a 3848 vector suitable for applying the matrix to. This is an mxn matrix. In the 3849 common case of a square matrix, the row and column ranges are the same and 3850 the DIAGONAL part is also square. The remaining portion of the local 3851 submatrix (mxN) constitute the OFF-DIAGONAL portion. 3852 3853 If o_nnz, d_nnz are specified, then o_nz, and d_nz are ignored. 3854 3855 You can call MatGetInfo() to get information on how effective the preallocation was; 3856 for example the fields mallocs,nz_allocated,nz_used,nz_unneeded; 3857 You can also run with the option -info and look for messages with the string 3858 malloc in them to see if additional memory allocation was needed. 3859 3860 Example usage: 3861 3862 Consider the following 8x8 matrix with 34 non-zero values, that is 3863 assembled across 3 processors. Lets assume that proc0 owns 3 rows, 3864 proc1 owns 3 rows, proc2 owns 2 rows. This division can be shown 3865 as follows: 3866 3867 .vb 3868 1 2 0 | 0 3 0 | 0 4 3869 Proc0 0 5 6 | 7 0 0 | 8 0 3870 9 0 10 | 11 0 0 | 12 0 3871 ------------------------------------- 3872 13 0 14 | 15 16 17 | 0 0 3873 Proc1 0 18 0 | 19 20 21 | 0 0 3874 0 0 0 | 22 23 0 | 24 0 3875 ------------------------------------- 3876 Proc2 25 26 27 | 0 0 28 | 29 0 3877 30 0 0 | 31 32 33 | 0 34 3878 .ve 3879 3880 This can be represented as a collection of submatrices as: 3881 3882 .vb 3883 A B C 3884 D E F 3885 G H I 3886 .ve 3887 3888 Where the submatrices A,B,C are owned by proc0, D,E,F are 3889 owned by proc1, G,H,I are owned by proc2. 3890 3891 The 'm' parameters for proc0,proc1,proc2 are 3,3,2 respectively. 3892 The 'n' parameters for proc0,proc1,proc2 are 3,3,2 respectively. 3893 The 'M','N' parameters are 8,8, and have the same values on all procs. 3894 3895 The DIAGONAL submatrices corresponding to proc0,proc1,proc2 are 3896 submatrices [A], [E], [I] respectively. The OFF-DIAGONAL submatrices 3897 corresponding to proc0,proc1,proc2 are [BC], [DF], [GH] respectively. 3898 Internally, each processor stores the DIAGONAL part, and the OFF-DIAGONAL 3899 part as SeqAIJ matrices. for eg: proc1 will store [E] as a SeqAIJ 3900 matrix, ans [DF] as another SeqAIJ matrix. 3901 3902 When d_nz, o_nz parameters are specified, d_nz storage elements are 3903 allocated for every row of the local diagonal submatrix, and o_nz 3904 storage locations are allocated for every row of the OFF-DIAGONAL submat. 3905 One way to choose d_nz and o_nz is to use the max nonzerors per local 3906 rows for each of the local DIAGONAL, and the OFF-DIAGONAL submatrices. 3907 In this case, the values of d_nz,o_nz are: 3908 .vb 3909 proc0 : dnz = 2, o_nz = 2 3910 proc1 : dnz = 3, o_nz = 2 3911 proc2 : dnz = 1, o_nz = 4 3912 .ve 3913 We are allocating m*(d_nz+o_nz) storage locations for every proc. This 3914 translates to 3*(2+2)=12 for proc0, 3*(3+2)=15 for proc1, 2*(1+4)=10 3915 for proc3. i.e we are using 12+15+10=37 storage locations to store 3916 34 values. 3917 3918 When d_nnz, o_nnz parameters are specified, the storage is specified 3919 for every row, coresponding to both DIAGONAL and OFF-DIAGONAL submatrices. 3920 In the above case the values for d_nnz,o_nnz are: 3921 .vb 3922 proc0: d_nnz = [2,2,2] and o_nnz = [2,2,2] 3923 proc1: d_nnz = [3,3,2] and o_nnz = [2,1,1] 3924 proc2: d_nnz = [1,1] and o_nnz = [4,4] 3925 .ve 3926 Here the space allocated is sum of all the above values i.e 34, and 3927 hence pre-allocation is perfect. 3928 3929 Level: intermediate 3930 3931 .keywords: matrix, aij, compressed row, sparse, parallel 3932 3933 .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatCreateAIJ(), MatMPIAIJSetPreallocationCSR(), 3934 MPIAIJ, MatGetInfo(), PetscSplitOwnership() 3935 @*/ 3936 PetscErrorCode MatMPIAIJSetPreallocation(Mat B,PetscInt d_nz,const PetscInt d_nnz[],PetscInt o_nz,const PetscInt o_nnz[]) 3937 { 3938 PetscErrorCode ierr; 3939 3940 PetscFunctionBegin; 3941 PetscValidHeaderSpecific(B,MAT_CLASSID,1); 3942 PetscValidType(B,1); 3943 ierr = PetscTryMethod(B,"MatMPIAIJSetPreallocation_C",(Mat,PetscInt,const PetscInt[],PetscInt,const PetscInt[]),(B,d_nz,d_nnz,o_nz,o_nnz));CHKERRQ(ierr); 3944 PetscFunctionReturn(0); 3945 } 3946 3947 #undef __FUNCT__ 3948 #define __FUNCT__ "MatCreateMPIAIJWithArrays" 3949 /*@ 3950 MatCreateMPIAIJWithArrays - creates a MPI AIJ matrix using arrays that contain in standard 3951 CSR format the local rows. 3952 3953 Collective on MPI_Comm 3954 3955 Input Parameters: 3956 + comm - MPI communicator 3957 . m - number of local rows (Cannot be PETSC_DECIDE) 3958 . n - This value should be the same as the local size used in creating the 3959 x vector for the matrix-vector product y = Ax. (or PETSC_DECIDE to have 3960 calculated if N is given) For square matrices n is almost always m. 3961 . M - number of global rows (or PETSC_DETERMINE to have calculated if m is given) 3962 . N - number of global columns (or PETSC_DETERMINE to have calculated if n is given) 3963 . i - row indices 3964 . j - column indices 3965 - a - matrix values 3966 3967 Output Parameter: 3968 . mat - the matrix 3969 3970 Level: intermediate 3971 3972 Notes: 3973 The i, j, and a arrays ARE copied by this routine into the internal format used by PETSc; 3974 thus you CANNOT change the matrix entries by changing the values of a[] after you have 3975 called this routine. Use MatCreateMPIAIJWithSplitArrays() to avoid needing to copy the arrays. 3976 3977 The i and j indices are 0 based, and i indices are indices corresponding to the local j array. 3978 3979 The format which is used for the sparse matrix input, is equivalent to a 3980 row-major ordering.. i.e for the following matrix, the input data expected is 3981 as shown: 3982 3983 1 0 0 3984 2 0 3 P0 3985 ------- 3986 4 5 6 P1 3987 3988 Process0 [P0]: rows_owned=[0,1] 3989 i = {0,1,3} [size = nrow+1 = 2+1] 3990 j = {0,0,2} [size = nz = 6] 3991 v = {1,2,3} [size = nz = 6] 3992 3993 Process1 [P1]: rows_owned=[2] 3994 i = {0,3} [size = nrow+1 = 1+1] 3995 j = {0,1,2} [size = nz = 6] 3996 v = {4,5,6} [size = nz = 6] 3997 3998 .keywords: matrix, aij, compressed row, sparse, parallel 3999 4000 .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatMPIAIJSetPreallocationCSR(), 4001 MPIAIJ, MatCreateAIJ(), MatCreateMPIAIJWithSplitArrays() 4002 @*/ 4003 PetscErrorCode MatCreateMPIAIJWithArrays(MPI_Comm comm,PetscInt m,PetscInt n,PetscInt M,PetscInt N,const PetscInt i[],const PetscInt j[],const PetscScalar a[],Mat *mat) 4004 { 4005 PetscErrorCode ierr; 4006 4007 PetscFunctionBegin; 4008 if (i[0]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"i (row indices) must start with 0"); 4009 if (m < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"local number of rows (m) cannot be PETSC_DECIDE, or negative"); 4010 ierr = MatCreate(comm,mat);CHKERRQ(ierr); 4011 ierr = MatSetSizes(*mat,m,n,M,N);CHKERRQ(ierr); 4012 /* ierr = MatSetBlockSizes(M,bs,cbs);CHKERRQ(ierr); */ 4013 ierr = MatSetType(*mat,MATMPIAIJ);CHKERRQ(ierr); 4014 ierr = MatMPIAIJSetPreallocationCSR(*mat,i,j,a);CHKERRQ(ierr); 4015 PetscFunctionReturn(0); 4016 } 4017 4018 #undef __FUNCT__ 4019 #define __FUNCT__ "MatCreateAIJ" 4020 /*@C 4021 MatCreateAIJ - Creates a sparse parallel matrix in AIJ format 4022 (the default parallel PETSc format). For good matrix assembly performance 4023 the user should preallocate the matrix storage by setting the parameters 4024 d_nz (or d_nnz) and o_nz (or o_nnz). By setting these parameters accurately, 4025 performance can be increased by more than a factor of 50. 4026 4027 Collective on MPI_Comm 4028 4029 Input Parameters: 4030 + comm - MPI communicator 4031 . m - number of local rows (or PETSC_DECIDE to have calculated if M is given) 4032 This value should be the same as the local size used in creating the 4033 y vector for the matrix-vector product y = Ax. 4034 . n - This value should be the same as the local size used in creating the 4035 x vector for the matrix-vector product y = Ax. (or PETSC_DECIDE to have 4036 calculated if N is given) For square matrices n is almost always m. 4037 . M - number of global rows (or PETSC_DETERMINE to have calculated if m is given) 4038 . N - number of global columns (or PETSC_DETERMINE to have calculated if n is given) 4039 . d_nz - number of nonzeros per row in DIAGONAL portion of local submatrix 4040 (same value is used for all local rows) 4041 . d_nnz - array containing the number of nonzeros in the various rows of the 4042 DIAGONAL portion of the local submatrix (possibly different for each row) 4043 or NULL, if d_nz is used to specify the nonzero structure. 4044 The size of this array is equal to the number of local rows, i.e 'm'. 4045 . o_nz - number of nonzeros per row in the OFF-DIAGONAL portion of local 4046 submatrix (same value is used for all local rows). 4047 - o_nnz - array containing the number of nonzeros in the various rows of the 4048 OFF-DIAGONAL portion of the local submatrix (possibly different for 4049 each row) or NULL, if o_nz is used to specify the nonzero 4050 structure. The size of this array is equal to the number 4051 of local rows, i.e 'm'. 4052 4053 Output Parameter: 4054 . A - the matrix 4055 4056 It is recommended that one use the MatCreate(), MatSetType() and/or MatSetFromOptions(), 4057 MatXXXXSetPreallocation() paradgm instead of this routine directly. 4058 [MatXXXXSetPreallocation() is, for example, MatSeqAIJSetPreallocation] 4059 4060 Notes: 4061 If the *_nnz parameter is given then the *_nz parameter is ignored 4062 4063 m,n,M,N parameters specify the size of the matrix, and its partitioning across 4064 processors, while d_nz,d_nnz,o_nz,o_nnz parameters specify the approximate 4065 storage requirements for this matrix. 4066 4067 If PETSC_DECIDE or PETSC_DETERMINE is used for a particular argument on one 4068 processor than it must be used on all processors that share the object for 4069 that argument. 4070 4071 The user MUST specify either the local or global matrix dimensions 4072 (possibly both). 4073 4074 The parallel matrix is partitioned across processors such that the 4075 first m0 rows belong to process 0, the next m1 rows belong to 4076 process 1, the next m2 rows belong to process 2 etc.. where 4077 m0,m1,m2,.. are the input parameter 'm'. i.e each processor stores 4078 values corresponding to [m x N] submatrix. 4079 4080 The columns are logically partitioned with the n0 columns belonging 4081 to 0th partition, the next n1 columns belonging to the next 4082 partition etc.. where n0,n1,n2... are the input parameter 'n'. 4083 4084 The DIAGONAL portion of the local submatrix on any given processor 4085 is the submatrix corresponding to the rows and columns m,n 4086 corresponding to the given processor. i.e diagonal matrix on 4087 process 0 is [m0 x n0], diagonal matrix on process 1 is [m1 x n1] 4088 etc. The remaining portion of the local submatrix [m x (N-n)] 4089 constitute the OFF-DIAGONAL portion. The example below better 4090 illustrates this concept. 4091 4092 For a square global matrix we define each processor's diagonal portion 4093 to be its local rows and the corresponding columns (a square submatrix); 4094 each processor's off-diagonal portion encompasses the remainder of the 4095 local matrix (a rectangular submatrix). 4096 4097 If o_nnz, d_nnz are specified, then o_nz, and d_nz are ignored. 4098 4099 When calling this routine with a single process communicator, a matrix of 4100 type SEQAIJ is returned. If a matrix of type MPIAIJ is desired for this 4101 type of communicator, use the construction mechanism: 4102 MatCreate(...,&A); MatSetType(A,MATMPIAIJ); MatSetSizes(A, m,n,M,N); MatMPIAIJSetPreallocation(A,...); 4103 4104 By default, this format uses inodes (identical nodes) when possible. 4105 We search for consecutive rows with the same nonzero structure, thereby 4106 reusing matrix information to achieve increased efficiency. 4107 4108 Options Database Keys: 4109 + -mat_no_inode - Do not use inodes 4110 . -mat_inode_limit <limit> - Sets inode limit (max limit=5) 4111 - -mat_aij_oneindex - Internally use indexing starting at 1 4112 rather than 0. Note that when calling MatSetValues(), 4113 the user still MUST index entries starting at 0! 4114 4115 4116 Example usage: 4117 4118 Consider the following 8x8 matrix with 34 non-zero values, that is 4119 assembled across 3 processors. Lets assume that proc0 owns 3 rows, 4120 proc1 owns 3 rows, proc2 owns 2 rows. This division can be shown 4121 as follows: 4122 4123 .vb 4124 1 2 0 | 0 3 0 | 0 4 4125 Proc0 0 5 6 | 7 0 0 | 8 0 4126 9 0 10 | 11 0 0 | 12 0 4127 ------------------------------------- 4128 13 0 14 | 15 16 17 | 0 0 4129 Proc1 0 18 0 | 19 20 21 | 0 0 4130 0 0 0 | 22 23 0 | 24 0 4131 ------------------------------------- 4132 Proc2 25 26 27 | 0 0 28 | 29 0 4133 30 0 0 | 31 32 33 | 0 34 4134 .ve 4135 4136 This can be represented as a collection of submatrices as: 4137 4138 .vb 4139 A B C 4140 D E F 4141 G H I 4142 .ve 4143 4144 Where the submatrices A,B,C are owned by proc0, D,E,F are 4145 owned by proc1, G,H,I are owned by proc2. 4146 4147 The 'm' parameters for proc0,proc1,proc2 are 3,3,2 respectively. 4148 The 'n' parameters for proc0,proc1,proc2 are 3,3,2 respectively. 4149 The 'M','N' parameters are 8,8, and have the same values on all procs. 4150 4151 The DIAGONAL submatrices corresponding to proc0,proc1,proc2 are 4152 submatrices [A], [E], [I] respectively. The OFF-DIAGONAL submatrices 4153 corresponding to proc0,proc1,proc2 are [BC], [DF], [GH] respectively. 4154 Internally, each processor stores the DIAGONAL part, and the OFF-DIAGONAL 4155 part as SeqAIJ matrices. for eg: proc1 will store [E] as a SeqAIJ 4156 matrix, ans [DF] as another SeqAIJ matrix. 4157 4158 When d_nz, o_nz parameters are specified, d_nz storage elements are 4159 allocated for every row of the local diagonal submatrix, and o_nz 4160 storage locations are allocated for every row of the OFF-DIAGONAL submat. 4161 One way to choose d_nz and o_nz is to use the max nonzerors per local 4162 rows for each of the local DIAGONAL, and the OFF-DIAGONAL submatrices. 4163 In this case, the values of d_nz,o_nz are: 4164 .vb 4165 proc0 : dnz = 2, o_nz = 2 4166 proc1 : dnz = 3, o_nz = 2 4167 proc2 : dnz = 1, o_nz = 4 4168 .ve 4169 We are allocating m*(d_nz+o_nz) storage locations for every proc. This 4170 translates to 3*(2+2)=12 for proc0, 3*(3+2)=15 for proc1, 2*(1+4)=10 4171 for proc3. i.e we are using 12+15+10=37 storage locations to store 4172 34 values. 4173 4174 When d_nnz, o_nnz parameters are specified, the storage is specified 4175 for every row, coresponding to both DIAGONAL and OFF-DIAGONAL submatrices. 4176 In the above case the values for d_nnz,o_nnz are: 4177 .vb 4178 proc0: d_nnz = [2,2,2] and o_nnz = [2,2,2] 4179 proc1: d_nnz = [3,3,2] and o_nnz = [2,1,1] 4180 proc2: d_nnz = [1,1] and o_nnz = [4,4] 4181 .ve 4182 Here the space allocated is sum of all the above values i.e 34, and 4183 hence pre-allocation is perfect. 4184 4185 Level: intermediate 4186 4187 .keywords: matrix, aij, compressed row, sparse, parallel 4188 4189 .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatMPIAIJSetPreallocationCSR(), 4190 MPIAIJ, MatCreateMPIAIJWithArrays() 4191 @*/ 4192 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) 4193 { 4194 PetscErrorCode ierr; 4195 PetscMPIInt size; 4196 4197 PetscFunctionBegin; 4198 ierr = MatCreate(comm,A);CHKERRQ(ierr); 4199 ierr = MatSetSizes(*A,m,n,M,N);CHKERRQ(ierr); 4200 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 4201 if (size > 1) { 4202 ierr = MatSetType(*A,MATMPIAIJ);CHKERRQ(ierr); 4203 ierr = MatMPIAIJSetPreallocation(*A,d_nz,d_nnz,o_nz,o_nnz);CHKERRQ(ierr); 4204 } else { 4205 ierr = MatSetType(*A,MATSEQAIJ);CHKERRQ(ierr); 4206 ierr = MatSeqAIJSetPreallocation(*A,d_nz,d_nnz);CHKERRQ(ierr); 4207 } 4208 PetscFunctionReturn(0); 4209 } 4210 4211 #undef __FUNCT__ 4212 #define __FUNCT__ "MatMPIAIJGetSeqAIJ" 4213 PetscErrorCode MatMPIAIJGetSeqAIJ(Mat A,Mat *Ad,Mat *Ao,const PetscInt *colmap[]) 4214 { 4215 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 4216 4217 PetscFunctionBegin; 4218 if (Ad) *Ad = a->A; 4219 if (Ao) *Ao = a->B; 4220 if (colmap) *colmap = a->garray; 4221 PetscFunctionReturn(0); 4222 } 4223 4224 #undef __FUNCT__ 4225 #define __FUNCT__ "MatSetColoring_MPIAIJ" 4226 PetscErrorCode MatSetColoring_MPIAIJ(Mat A,ISColoring coloring) 4227 { 4228 PetscErrorCode ierr; 4229 PetscInt i; 4230 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 4231 4232 PetscFunctionBegin; 4233 if (coloring->ctype == IS_COLORING_GLOBAL) { 4234 ISColoringValue *allcolors,*colors; 4235 ISColoring ocoloring; 4236 4237 /* set coloring for diagonal portion */ 4238 ierr = MatSetColoring_SeqAIJ(a->A,coloring);CHKERRQ(ierr); 4239 4240 /* set coloring for off-diagonal portion */ 4241 ierr = ISAllGatherColors(PetscObjectComm((PetscObject)A),coloring->n,coloring->colors,NULL,&allcolors);CHKERRQ(ierr); 4242 ierr = PetscMalloc1((a->B->cmap->n+1),&colors);CHKERRQ(ierr); 4243 for (i=0; i<a->B->cmap->n; i++) { 4244 colors[i] = allcolors[a->garray[i]]; 4245 } 4246 ierr = PetscFree(allcolors);CHKERRQ(ierr); 4247 ierr = ISColoringCreate(MPI_COMM_SELF,coloring->n,a->B->cmap->n,colors,&ocoloring);CHKERRQ(ierr); 4248 ierr = MatSetColoring_SeqAIJ(a->B,ocoloring);CHKERRQ(ierr); 4249 ierr = ISColoringDestroy(&ocoloring);CHKERRQ(ierr); 4250 } else if (coloring->ctype == IS_COLORING_GHOSTED) { 4251 ISColoringValue *colors; 4252 PetscInt *larray; 4253 ISColoring ocoloring; 4254 4255 /* set coloring for diagonal portion */ 4256 ierr = PetscMalloc1((a->A->cmap->n+1),&larray);CHKERRQ(ierr); 4257 for (i=0; i<a->A->cmap->n; i++) { 4258 larray[i] = i + A->cmap->rstart; 4259 } 4260 ierr = ISGlobalToLocalMappingApply(A->cmap->mapping,IS_GTOLM_MASK,a->A->cmap->n,larray,NULL,larray);CHKERRQ(ierr); 4261 ierr = PetscMalloc1((a->A->cmap->n+1),&colors);CHKERRQ(ierr); 4262 for (i=0; i<a->A->cmap->n; i++) { 4263 colors[i] = coloring->colors[larray[i]]; 4264 } 4265 ierr = PetscFree(larray);CHKERRQ(ierr); 4266 ierr = ISColoringCreate(PETSC_COMM_SELF,coloring->n,a->A->cmap->n,colors,&ocoloring);CHKERRQ(ierr); 4267 ierr = MatSetColoring_SeqAIJ(a->A,ocoloring);CHKERRQ(ierr); 4268 ierr = ISColoringDestroy(&ocoloring);CHKERRQ(ierr); 4269 4270 /* set coloring for off-diagonal portion */ 4271 ierr = PetscMalloc1((a->B->cmap->n+1),&larray);CHKERRQ(ierr); 4272 ierr = ISGlobalToLocalMappingApply(A->cmap->mapping,IS_GTOLM_MASK,a->B->cmap->n,a->garray,NULL,larray);CHKERRQ(ierr); 4273 ierr = PetscMalloc1((a->B->cmap->n+1),&colors);CHKERRQ(ierr); 4274 for (i=0; i<a->B->cmap->n; i++) { 4275 colors[i] = coloring->colors[larray[i]]; 4276 } 4277 ierr = PetscFree(larray);CHKERRQ(ierr); 4278 ierr = ISColoringCreate(MPI_COMM_SELF,coloring->n,a->B->cmap->n,colors,&ocoloring);CHKERRQ(ierr); 4279 ierr = MatSetColoring_SeqAIJ(a->B,ocoloring);CHKERRQ(ierr); 4280 ierr = ISColoringDestroy(&ocoloring);CHKERRQ(ierr); 4281 } else SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"No support ISColoringType %d",(int)coloring->ctype); 4282 PetscFunctionReturn(0); 4283 } 4284 4285 #undef __FUNCT__ 4286 #define __FUNCT__ "MatSetValuesAdifor_MPIAIJ" 4287 PetscErrorCode MatSetValuesAdifor_MPIAIJ(Mat A,PetscInt nl,void *advalues) 4288 { 4289 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 4290 PetscErrorCode ierr; 4291 4292 PetscFunctionBegin; 4293 ierr = MatSetValuesAdifor_SeqAIJ(a->A,nl,advalues);CHKERRQ(ierr); 4294 ierr = MatSetValuesAdifor_SeqAIJ(a->B,nl,advalues);CHKERRQ(ierr); 4295 PetscFunctionReturn(0); 4296 } 4297 4298 #undef __FUNCT__ 4299 #define __FUNCT__ "MatCreateMPIAIJConcatenateSeqAIJSymbolic" 4300 PetscErrorCode MatCreateMPIAIJConcatenateSeqAIJSymbolic(MPI_Comm comm,Mat inmat,PetscInt n,Mat *outmat) 4301 { 4302 PetscErrorCode ierr; 4303 PetscInt m,N,i,rstart,nnz,*dnz,*onz,sum,bs,cbs; 4304 PetscInt *indx; 4305 4306 PetscFunctionBegin; 4307 /* This routine will ONLY return MPIAIJ type matrix */ 4308 ierr = MatGetSize(inmat,&m,&N);CHKERRQ(ierr); 4309 ierr = MatGetBlockSizes(inmat,&bs,&cbs);CHKERRQ(ierr); 4310 if (n == PETSC_DECIDE) { 4311 ierr = PetscSplitOwnership(comm,&n,&N);CHKERRQ(ierr); 4312 } 4313 /* Check sum(n) = N */ 4314 ierr = MPI_Allreduce(&n,&sum,1,MPIU_INT,MPI_SUM,comm);CHKERRQ(ierr); 4315 if (sum != N) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_INCOMP,"Sum of local columns != global columns %d",N); 4316 4317 ierr = MPI_Scan(&m, &rstart,1,MPIU_INT,MPI_SUM,comm);CHKERRQ(ierr); 4318 rstart -= m; 4319 4320 ierr = MatPreallocateInitialize(comm,m,n,dnz,onz);CHKERRQ(ierr); 4321 for (i=0; i<m; i++) { 4322 ierr = MatGetRow_SeqAIJ(inmat,i,&nnz,&indx,NULL);CHKERRQ(ierr); 4323 ierr = MatPreallocateSet(i+rstart,nnz,indx,dnz,onz);CHKERRQ(ierr); 4324 ierr = MatRestoreRow_SeqAIJ(inmat,i,&nnz,&indx,NULL);CHKERRQ(ierr); 4325 } 4326 4327 ierr = MatCreate(comm,outmat);CHKERRQ(ierr); 4328 ierr = MatSetSizes(*outmat,m,n,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 4329 ierr = MatSetBlockSizes(*outmat,bs,cbs);CHKERRQ(ierr); 4330 ierr = MatSetType(*outmat,MATMPIAIJ);CHKERRQ(ierr); 4331 ierr = MatMPIAIJSetPreallocation(*outmat,0,dnz,0,onz);CHKERRQ(ierr); 4332 ierr = MatPreallocateFinalize(dnz,onz);CHKERRQ(ierr); 4333 PetscFunctionReturn(0); 4334 } 4335 4336 #undef __FUNCT__ 4337 #define __FUNCT__ "MatCreateMPIAIJConcatenateSeqAIJNumeric" 4338 PetscErrorCode MatCreateMPIAIJConcatenateSeqAIJNumeric(MPI_Comm comm,Mat inmat,PetscInt n,Mat outmat) 4339 { 4340 PetscErrorCode ierr; 4341 PetscInt m,N,i,rstart,nnz,Ii; 4342 PetscInt *indx; 4343 PetscScalar *values; 4344 4345 PetscFunctionBegin; 4346 ierr = MatGetSize(inmat,&m,&N);CHKERRQ(ierr); 4347 ierr = MatGetOwnershipRange(outmat,&rstart,NULL);CHKERRQ(ierr); 4348 for (i=0; i<m; i++) { 4349 ierr = MatGetRow_SeqAIJ(inmat,i,&nnz,&indx,&values);CHKERRQ(ierr); 4350 Ii = i + rstart; 4351 ierr = MatSetValues(outmat,1,&Ii,nnz,indx,values,INSERT_VALUES);CHKERRQ(ierr); 4352 ierr = MatRestoreRow_SeqAIJ(inmat,i,&nnz,&indx,&values);CHKERRQ(ierr); 4353 } 4354 ierr = MatAssemblyBegin(outmat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4355 ierr = MatAssemblyEnd(outmat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4356 PetscFunctionReturn(0); 4357 } 4358 4359 #undef __FUNCT__ 4360 #define __FUNCT__ "MatCreateMPIMatConcatenateSeqMat_MPIAIJ" 4361 PetscErrorCode MatCreateMPIMatConcatenateSeqMat_MPIAIJ(MPI_Comm comm,Mat inmat,PetscInt n,MatReuse scall,Mat *outmat) 4362 { 4363 PetscErrorCode ierr; 4364 PetscMPIInt size; 4365 4366 PetscFunctionBegin; 4367 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 4368 ierr = PetscLogEventBegin(MAT_Merge,inmat,0,0,0);CHKERRQ(ierr); 4369 if (size == 1) { 4370 if (scall == MAT_INITIAL_MATRIX) { 4371 ierr = MatDuplicate(inmat,MAT_COPY_VALUES,outmat);CHKERRQ(ierr); 4372 } else { 4373 ierr = MatCopy(inmat,*outmat,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 4374 } 4375 } else { 4376 if (scall == MAT_INITIAL_MATRIX) { 4377 ierr = MatCreateMPIAIJConcatenateSeqAIJSymbolic(comm,inmat,n,outmat);CHKERRQ(ierr); 4378 } 4379 ierr = MatCreateMPIAIJConcatenateSeqAIJNumeric(comm,inmat,n,*outmat);CHKERRQ(ierr); 4380 } 4381 ierr = PetscLogEventEnd(MAT_Merge,inmat,0,0,0);CHKERRQ(ierr); 4382 PetscFunctionReturn(0); 4383 } 4384 4385 #undef __FUNCT__ 4386 #define __FUNCT__ "MatFileSplit" 4387 PetscErrorCode MatFileSplit(Mat A,char *outfile) 4388 { 4389 PetscErrorCode ierr; 4390 PetscMPIInt rank; 4391 PetscInt m,N,i,rstart,nnz; 4392 size_t len; 4393 const PetscInt *indx; 4394 PetscViewer out; 4395 char *name; 4396 Mat B; 4397 const PetscScalar *values; 4398 4399 PetscFunctionBegin; 4400 ierr = MatGetLocalSize(A,&m,0);CHKERRQ(ierr); 4401 ierr = MatGetSize(A,0,&N);CHKERRQ(ierr); 4402 /* Should this be the type of the diagonal block of A? */ 4403 ierr = MatCreate(PETSC_COMM_SELF,&B);CHKERRQ(ierr); 4404 ierr = MatSetSizes(B,m,N,m,N);CHKERRQ(ierr); 4405 ierr = MatSetBlockSizesFromMats(B,A,A);CHKERRQ(ierr); 4406 ierr = MatSetType(B,MATSEQAIJ);CHKERRQ(ierr); 4407 ierr = MatSeqAIJSetPreallocation(B,0,NULL);CHKERRQ(ierr); 4408 ierr = MatGetOwnershipRange(A,&rstart,0);CHKERRQ(ierr); 4409 for (i=0; i<m; i++) { 4410 ierr = MatGetRow(A,i+rstart,&nnz,&indx,&values);CHKERRQ(ierr); 4411 ierr = MatSetValues(B,1,&i,nnz,indx,values,INSERT_VALUES);CHKERRQ(ierr); 4412 ierr = MatRestoreRow(A,i+rstart,&nnz,&indx,&values);CHKERRQ(ierr); 4413 } 4414 ierr = MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4415 ierr = MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4416 4417 ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)A),&rank);CHKERRQ(ierr); 4418 ierr = PetscStrlen(outfile,&len);CHKERRQ(ierr); 4419 ierr = PetscMalloc1((len+5),&name);CHKERRQ(ierr); 4420 sprintf(name,"%s.%d",outfile,rank); 4421 ierr = PetscViewerBinaryOpen(PETSC_COMM_SELF,name,FILE_MODE_APPEND,&out);CHKERRQ(ierr); 4422 ierr = PetscFree(name);CHKERRQ(ierr); 4423 ierr = MatView(B,out);CHKERRQ(ierr); 4424 ierr = PetscViewerDestroy(&out);CHKERRQ(ierr); 4425 ierr = MatDestroy(&B);CHKERRQ(ierr); 4426 PetscFunctionReturn(0); 4427 } 4428 4429 extern PetscErrorCode MatDestroy_MPIAIJ(Mat); 4430 #undef __FUNCT__ 4431 #define __FUNCT__ "MatDestroy_MPIAIJ_SeqsToMPI" 4432 PetscErrorCode MatDestroy_MPIAIJ_SeqsToMPI(Mat A) 4433 { 4434 PetscErrorCode ierr; 4435 Mat_Merge_SeqsToMPI *merge; 4436 PetscContainer container; 4437 4438 PetscFunctionBegin; 4439 ierr = PetscObjectQuery((PetscObject)A,"MatMergeSeqsToMPI",(PetscObject*)&container);CHKERRQ(ierr); 4440 if (container) { 4441 ierr = PetscContainerGetPointer(container,(void**)&merge);CHKERRQ(ierr); 4442 ierr = PetscFree(merge->id_r);CHKERRQ(ierr); 4443 ierr = PetscFree(merge->len_s);CHKERRQ(ierr); 4444 ierr = PetscFree(merge->len_r);CHKERRQ(ierr); 4445 ierr = PetscFree(merge->bi);CHKERRQ(ierr); 4446 ierr = PetscFree(merge->bj);CHKERRQ(ierr); 4447 ierr = PetscFree(merge->buf_ri[0]);CHKERRQ(ierr); 4448 ierr = PetscFree(merge->buf_ri);CHKERRQ(ierr); 4449 ierr = PetscFree(merge->buf_rj[0]);CHKERRQ(ierr); 4450 ierr = PetscFree(merge->buf_rj);CHKERRQ(ierr); 4451 ierr = PetscFree(merge->coi);CHKERRQ(ierr); 4452 ierr = PetscFree(merge->coj);CHKERRQ(ierr); 4453 ierr = PetscFree(merge->owners_co);CHKERRQ(ierr); 4454 ierr = PetscLayoutDestroy(&merge->rowmap);CHKERRQ(ierr); 4455 ierr = PetscFree(merge);CHKERRQ(ierr); 4456 ierr = PetscObjectCompose((PetscObject)A,"MatMergeSeqsToMPI",0);CHKERRQ(ierr); 4457 } 4458 ierr = MatDestroy_MPIAIJ(A);CHKERRQ(ierr); 4459 PetscFunctionReturn(0); 4460 } 4461 4462 #include <../src/mat/utils/freespace.h> 4463 #include <petscbt.h> 4464 4465 #undef __FUNCT__ 4466 #define __FUNCT__ "MatCreateMPIAIJSumSeqAIJNumeric" 4467 PetscErrorCode MatCreateMPIAIJSumSeqAIJNumeric(Mat seqmat,Mat mpimat) 4468 { 4469 PetscErrorCode ierr; 4470 MPI_Comm comm; 4471 Mat_SeqAIJ *a =(Mat_SeqAIJ*)seqmat->data; 4472 PetscMPIInt size,rank,taga,*len_s; 4473 PetscInt N=mpimat->cmap->N,i,j,*owners,*ai=a->i,*aj; 4474 PetscInt proc,m; 4475 PetscInt **buf_ri,**buf_rj; 4476 PetscInt k,anzi,*bj_i,*bi,*bj,arow,bnzi,nextaj; 4477 PetscInt nrows,**buf_ri_k,**nextrow,**nextai; 4478 MPI_Request *s_waits,*r_waits; 4479 MPI_Status *status; 4480 MatScalar *aa=a->a; 4481 MatScalar **abuf_r,*ba_i; 4482 Mat_Merge_SeqsToMPI *merge; 4483 PetscContainer container; 4484 4485 PetscFunctionBegin; 4486 ierr = PetscObjectGetComm((PetscObject)mpimat,&comm);CHKERRQ(ierr); 4487 ierr = PetscLogEventBegin(MAT_Seqstompinum,seqmat,0,0,0);CHKERRQ(ierr); 4488 4489 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 4490 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 4491 4492 ierr = PetscObjectQuery((PetscObject)mpimat,"MatMergeSeqsToMPI",(PetscObject*)&container);CHKERRQ(ierr); 4493 ierr = PetscContainerGetPointer(container,(void**)&merge);CHKERRQ(ierr); 4494 4495 bi = merge->bi; 4496 bj = merge->bj; 4497 buf_ri = merge->buf_ri; 4498 buf_rj = merge->buf_rj; 4499 4500 ierr = PetscMalloc1(size,&status);CHKERRQ(ierr); 4501 owners = merge->rowmap->range; 4502 len_s = merge->len_s; 4503 4504 /* send and recv matrix values */ 4505 /*-----------------------------*/ 4506 ierr = PetscObjectGetNewTag((PetscObject)mpimat,&taga);CHKERRQ(ierr); 4507 ierr = PetscPostIrecvScalar(comm,taga,merge->nrecv,merge->id_r,merge->len_r,&abuf_r,&r_waits);CHKERRQ(ierr); 4508 4509 ierr = PetscMalloc1((merge->nsend+1),&s_waits);CHKERRQ(ierr); 4510 for (proc=0,k=0; proc<size; proc++) { 4511 if (!len_s[proc]) continue; 4512 i = owners[proc]; 4513 ierr = MPI_Isend(aa+ai[i],len_s[proc],MPIU_MATSCALAR,proc,taga,comm,s_waits+k);CHKERRQ(ierr); 4514 k++; 4515 } 4516 4517 if (merge->nrecv) {ierr = MPI_Waitall(merge->nrecv,r_waits,status);CHKERRQ(ierr);} 4518 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,s_waits,status);CHKERRQ(ierr);} 4519 ierr = PetscFree(status);CHKERRQ(ierr); 4520 4521 ierr = PetscFree(s_waits);CHKERRQ(ierr); 4522 ierr = PetscFree(r_waits);CHKERRQ(ierr); 4523 4524 /* insert mat values of mpimat */ 4525 /*----------------------------*/ 4526 ierr = PetscMalloc1(N,&ba_i);CHKERRQ(ierr); 4527 ierr = PetscMalloc3(merge->nrecv,&buf_ri_k,merge->nrecv,&nextrow,merge->nrecv,&nextai);CHKERRQ(ierr); 4528 4529 for (k=0; k<merge->nrecv; k++) { 4530 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 4531 nrows = *(buf_ri_k[k]); 4532 nextrow[k] = buf_ri_k[k]+1; /* next row number of k-th recved i-structure */ 4533 nextai[k] = buf_ri_k[k] + (nrows + 1); /* poins to the next i-structure of k-th recved i-structure */ 4534 } 4535 4536 /* set values of ba */ 4537 m = merge->rowmap->n; 4538 for (i=0; i<m; i++) { 4539 arow = owners[rank] + i; 4540 bj_i = bj+bi[i]; /* col indices of the i-th row of mpimat */ 4541 bnzi = bi[i+1] - bi[i]; 4542 ierr = PetscMemzero(ba_i,bnzi*sizeof(PetscScalar));CHKERRQ(ierr); 4543 4544 /* add local non-zero vals of this proc's seqmat into ba */ 4545 anzi = ai[arow+1] - ai[arow]; 4546 aj = a->j + ai[arow]; 4547 aa = a->a + ai[arow]; 4548 nextaj = 0; 4549 for (j=0; nextaj<anzi; j++) { 4550 if (*(bj_i + j) == aj[nextaj]) { /* bcol == acol */ 4551 ba_i[j] += aa[nextaj++]; 4552 } 4553 } 4554 4555 /* add received vals into ba */ 4556 for (k=0; k<merge->nrecv; k++) { /* k-th received message */ 4557 /* i-th row */ 4558 if (i == *nextrow[k]) { 4559 anzi = *(nextai[k]+1) - *nextai[k]; 4560 aj = buf_rj[k] + *(nextai[k]); 4561 aa = abuf_r[k] + *(nextai[k]); 4562 nextaj = 0; 4563 for (j=0; nextaj<anzi; j++) { 4564 if (*(bj_i + j) == aj[nextaj]) { /* bcol == acol */ 4565 ba_i[j] += aa[nextaj++]; 4566 } 4567 } 4568 nextrow[k]++; nextai[k]++; 4569 } 4570 } 4571 ierr = MatSetValues(mpimat,1,&arow,bnzi,bj_i,ba_i,INSERT_VALUES);CHKERRQ(ierr); 4572 } 4573 ierr = MatAssemblyBegin(mpimat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4574 ierr = MatAssemblyEnd(mpimat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4575 4576 ierr = PetscFree(abuf_r[0]);CHKERRQ(ierr); 4577 ierr = PetscFree(abuf_r);CHKERRQ(ierr); 4578 ierr = PetscFree(ba_i);CHKERRQ(ierr); 4579 ierr = PetscFree3(buf_ri_k,nextrow,nextai);CHKERRQ(ierr); 4580 ierr = PetscLogEventEnd(MAT_Seqstompinum,seqmat,0,0,0);CHKERRQ(ierr); 4581 PetscFunctionReturn(0); 4582 } 4583 4584 extern PetscErrorCode MatDestroy_MPIAIJ_SeqsToMPI(Mat); 4585 4586 #undef __FUNCT__ 4587 #define __FUNCT__ "MatCreateMPIAIJSumSeqAIJSymbolic" 4588 PetscErrorCode MatCreateMPIAIJSumSeqAIJSymbolic(MPI_Comm comm,Mat seqmat,PetscInt m,PetscInt n,Mat *mpimat) 4589 { 4590 PetscErrorCode ierr; 4591 Mat B_mpi; 4592 Mat_SeqAIJ *a=(Mat_SeqAIJ*)seqmat->data; 4593 PetscMPIInt size,rank,tagi,tagj,*len_s,*len_si,*len_ri; 4594 PetscInt **buf_rj,**buf_ri,**buf_ri_k; 4595 PetscInt M=seqmat->rmap->n,N=seqmat->cmap->n,i,*owners,*ai=a->i,*aj=a->j; 4596 PetscInt len,proc,*dnz,*onz,bs,cbs; 4597 PetscInt k,anzi,*bi,*bj,*lnk,nlnk,arow,bnzi,nspacedouble=0; 4598 PetscInt nrows,*buf_s,*buf_si,*buf_si_i,**nextrow,**nextai; 4599 MPI_Request *si_waits,*sj_waits,*ri_waits,*rj_waits; 4600 MPI_Status *status; 4601 PetscFreeSpaceList free_space=NULL,current_space=NULL; 4602 PetscBT lnkbt; 4603 Mat_Merge_SeqsToMPI *merge; 4604 PetscContainer container; 4605 4606 PetscFunctionBegin; 4607 ierr = PetscLogEventBegin(MAT_Seqstompisym,seqmat,0,0,0);CHKERRQ(ierr); 4608 4609 /* make sure it is a PETSc comm */ 4610 ierr = PetscCommDuplicate(comm,&comm,NULL);CHKERRQ(ierr); 4611 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 4612 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 4613 4614 ierr = PetscNew(&merge);CHKERRQ(ierr); 4615 ierr = PetscMalloc1(size,&status);CHKERRQ(ierr); 4616 4617 /* determine row ownership */ 4618 /*---------------------------------------------------------*/ 4619 ierr = PetscLayoutCreate(comm,&merge->rowmap);CHKERRQ(ierr); 4620 ierr = PetscLayoutSetLocalSize(merge->rowmap,m);CHKERRQ(ierr); 4621 ierr = PetscLayoutSetSize(merge->rowmap,M);CHKERRQ(ierr); 4622 ierr = PetscLayoutSetBlockSize(merge->rowmap,1);CHKERRQ(ierr); 4623 ierr = PetscLayoutSetUp(merge->rowmap);CHKERRQ(ierr); 4624 ierr = PetscMalloc1(size,&len_si);CHKERRQ(ierr); 4625 ierr = PetscMalloc1(size,&merge->len_s);CHKERRQ(ierr); 4626 4627 m = merge->rowmap->n; 4628 owners = merge->rowmap->range; 4629 4630 /* determine the number of messages to send, their lengths */ 4631 /*---------------------------------------------------------*/ 4632 len_s = merge->len_s; 4633 4634 len = 0; /* length of buf_si[] */ 4635 merge->nsend = 0; 4636 for (proc=0; proc<size; proc++) { 4637 len_si[proc] = 0; 4638 if (proc == rank) { 4639 len_s[proc] = 0; 4640 } else { 4641 len_si[proc] = owners[proc+1] - owners[proc] + 1; 4642 len_s[proc] = ai[owners[proc+1]] - ai[owners[proc]]; /* num of rows to be sent to [proc] */ 4643 } 4644 if (len_s[proc]) { 4645 merge->nsend++; 4646 nrows = 0; 4647 for (i=owners[proc]; i<owners[proc+1]; i++) { 4648 if (ai[i+1] > ai[i]) nrows++; 4649 } 4650 len_si[proc] = 2*(nrows+1); 4651 len += len_si[proc]; 4652 } 4653 } 4654 4655 /* determine the number and length of messages to receive for ij-structure */ 4656 /*-------------------------------------------------------------------------*/ 4657 ierr = PetscGatherNumberOfMessages(comm,NULL,len_s,&merge->nrecv);CHKERRQ(ierr); 4658 ierr = PetscGatherMessageLengths2(comm,merge->nsend,merge->nrecv,len_s,len_si,&merge->id_r,&merge->len_r,&len_ri);CHKERRQ(ierr); 4659 4660 /* post the Irecv of j-structure */ 4661 /*-------------------------------*/ 4662 ierr = PetscCommGetNewTag(comm,&tagj);CHKERRQ(ierr); 4663 ierr = PetscPostIrecvInt(comm,tagj,merge->nrecv,merge->id_r,merge->len_r,&buf_rj,&rj_waits);CHKERRQ(ierr); 4664 4665 /* post the Isend of j-structure */ 4666 /*--------------------------------*/ 4667 ierr = PetscMalloc2(merge->nsend,&si_waits,merge->nsend,&sj_waits);CHKERRQ(ierr); 4668 4669 for (proc=0, k=0; proc<size; proc++) { 4670 if (!len_s[proc]) continue; 4671 i = owners[proc]; 4672 ierr = MPI_Isend(aj+ai[i],len_s[proc],MPIU_INT,proc,tagj,comm,sj_waits+k);CHKERRQ(ierr); 4673 k++; 4674 } 4675 4676 /* receives and sends of j-structure are complete */ 4677 /*------------------------------------------------*/ 4678 if (merge->nrecv) {ierr = MPI_Waitall(merge->nrecv,rj_waits,status);CHKERRQ(ierr);} 4679 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,sj_waits,status);CHKERRQ(ierr);} 4680 4681 /* send and recv i-structure */ 4682 /*---------------------------*/ 4683 ierr = PetscCommGetNewTag(comm,&tagi);CHKERRQ(ierr); 4684 ierr = PetscPostIrecvInt(comm,tagi,merge->nrecv,merge->id_r,len_ri,&buf_ri,&ri_waits);CHKERRQ(ierr); 4685 4686 ierr = PetscMalloc1((len+1),&buf_s);CHKERRQ(ierr); 4687 buf_si = buf_s; /* points to the beginning of k-th msg to be sent */ 4688 for (proc=0,k=0; proc<size; proc++) { 4689 if (!len_s[proc]) continue; 4690 /* form outgoing message for i-structure: 4691 buf_si[0]: nrows to be sent 4692 [1:nrows]: row index (global) 4693 [nrows+1:2*nrows+1]: i-structure index 4694 */ 4695 /*-------------------------------------------*/ 4696 nrows = len_si[proc]/2 - 1; 4697 buf_si_i = buf_si + nrows+1; 4698 buf_si[0] = nrows; 4699 buf_si_i[0] = 0; 4700 nrows = 0; 4701 for (i=owners[proc]; i<owners[proc+1]; i++) { 4702 anzi = ai[i+1] - ai[i]; 4703 if (anzi) { 4704 buf_si_i[nrows+1] = buf_si_i[nrows] + anzi; /* i-structure */ 4705 buf_si[nrows+1] = i-owners[proc]; /* local row index */ 4706 nrows++; 4707 } 4708 } 4709 ierr = MPI_Isend(buf_si,len_si[proc],MPIU_INT,proc,tagi,comm,si_waits+k);CHKERRQ(ierr); 4710 k++; 4711 buf_si += len_si[proc]; 4712 } 4713 4714 if (merge->nrecv) {ierr = MPI_Waitall(merge->nrecv,ri_waits,status);CHKERRQ(ierr);} 4715 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,si_waits,status);CHKERRQ(ierr);} 4716 4717 ierr = PetscInfo2(seqmat,"nsend: %D, nrecv: %D\n",merge->nsend,merge->nrecv);CHKERRQ(ierr); 4718 for (i=0; i<merge->nrecv; i++) { 4719 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); 4720 } 4721 4722 ierr = PetscFree(len_si);CHKERRQ(ierr); 4723 ierr = PetscFree(len_ri);CHKERRQ(ierr); 4724 ierr = PetscFree(rj_waits);CHKERRQ(ierr); 4725 ierr = PetscFree2(si_waits,sj_waits);CHKERRQ(ierr); 4726 ierr = PetscFree(ri_waits);CHKERRQ(ierr); 4727 ierr = PetscFree(buf_s);CHKERRQ(ierr); 4728 ierr = PetscFree(status);CHKERRQ(ierr); 4729 4730 /* compute a local seq matrix in each processor */ 4731 /*----------------------------------------------*/ 4732 /* allocate bi array and free space for accumulating nonzero column info */ 4733 ierr = PetscMalloc1((m+1),&bi);CHKERRQ(ierr); 4734 bi[0] = 0; 4735 4736 /* create and initialize a linked list */ 4737 nlnk = N+1; 4738 ierr = PetscLLCreate(N,N,nlnk,lnk,lnkbt);CHKERRQ(ierr); 4739 4740 /* initial FreeSpace size is 2*(num of local nnz(seqmat)) */ 4741 len = ai[owners[rank+1]] - ai[owners[rank]]; 4742 ierr = PetscFreeSpaceGet((PetscInt)(2*len+1),&free_space);CHKERRQ(ierr); 4743 4744 current_space = free_space; 4745 4746 /* determine symbolic info for each local row */ 4747 ierr = PetscMalloc3(merge->nrecv,&buf_ri_k,merge->nrecv,&nextrow,merge->nrecv,&nextai);CHKERRQ(ierr); 4748 4749 for (k=0; k<merge->nrecv; k++) { 4750 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 4751 nrows = *buf_ri_k[k]; 4752 nextrow[k] = buf_ri_k[k] + 1; /* next row number of k-th recved i-structure */ 4753 nextai[k] = buf_ri_k[k] + (nrows + 1); /* poins to the next i-structure of k-th recved i-structure */ 4754 } 4755 4756 ierr = MatPreallocateInitialize(comm,m,n,dnz,onz);CHKERRQ(ierr); 4757 len = 0; 4758 for (i=0; i<m; i++) { 4759 bnzi = 0; 4760 /* add local non-zero cols of this proc's seqmat into lnk */ 4761 arow = owners[rank] + i; 4762 anzi = ai[arow+1] - ai[arow]; 4763 aj = a->j + ai[arow]; 4764 ierr = PetscLLAddSorted(anzi,aj,N,nlnk,lnk,lnkbt);CHKERRQ(ierr); 4765 bnzi += nlnk; 4766 /* add received col data into lnk */ 4767 for (k=0; k<merge->nrecv; k++) { /* k-th received message */ 4768 if (i == *nextrow[k]) { /* i-th row */ 4769 anzi = *(nextai[k]+1) - *nextai[k]; 4770 aj = buf_rj[k] + *nextai[k]; 4771 ierr = PetscLLAddSorted(anzi,aj,N,nlnk,lnk,lnkbt);CHKERRQ(ierr); 4772 bnzi += nlnk; 4773 nextrow[k]++; nextai[k]++; 4774 } 4775 } 4776 if (len < bnzi) len = bnzi; /* =max(bnzi) */ 4777 4778 /* if free space is not available, make more free space */ 4779 if (current_space->local_remaining<bnzi) { 4780 ierr = PetscFreeSpaceGet(bnzi+current_space->total_array_size,¤t_space);CHKERRQ(ierr); 4781 nspacedouble++; 4782 } 4783 /* copy data into free space, then initialize lnk */ 4784 ierr = PetscLLClean(N,N,bnzi,lnk,current_space->array,lnkbt);CHKERRQ(ierr); 4785 ierr = MatPreallocateSet(i+owners[rank],bnzi,current_space->array,dnz,onz);CHKERRQ(ierr); 4786 4787 current_space->array += bnzi; 4788 current_space->local_used += bnzi; 4789 current_space->local_remaining -= bnzi; 4790 4791 bi[i+1] = bi[i] + bnzi; 4792 } 4793 4794 ierr = PetscFree3(buf_ri_k,nextrow,nextai);CHKERRQ(ierr); 4795 4796 ierr = PetscMalloc1((bi[m]+1),&bj);CHKERRQ(ierr); 4797 ierr = PetscFreeSpaceContiguous(&free_space,bj);CHKERRQ(ierr); 4798 ierr = PetscLLDestroy(lnk,lnkbt);CHKERRQ(ierr); 4799 4800 /* create symbolic parallel matrix B_mpi */ 4801 /*---------------------------------------*/ 4802 ierr = MatGetBlockSizes(seqmat,&bs,&cbs);CHKERRQ(ierr); 4803 ierr = MatCreate(comm,&B_mpi);CHKERRQ(ierr); 4804 if (n==PETSC_DECIDE) { 4805 ierr = MatSetSizes(B_mpi,m,n,PETSC_DETERMINE,N);CHKERRQ(ierr); 4806 } else { 4807 ierr = MatSetSizes(B_mpi,m,n,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 4808 } 4809 ierr = MatSetBlockSizes(B_mpi,bs,cbs);CHKERRQ(ierr); 4810 ierr = MatSetType(B_mpi,MATMPIAIJ);CHKERRQ(ierr); 4811 ierr = MatMPIAIJSetPreallocation(B_mpi,0,dnz,0,onz);CHKERRQ(ierr); 4812 ierr = MatPreallocateFinalize(dnz,onz);CHKERRQ(ierr); 4813 ierr = MatSetOption(B_mpi,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 4814 4815 /* B_mpi is not ready for use - assembly will be done by MatCreateMPIAIJSumSeqAIJNumeric() */ 4816 B_mpi->assembled = PETSC_FALSE; 4817 B_mpi->ops->destroy = MatDestroy_MPIAIJ_SeqsToMPI; 4818 merge->bi = bi; 4819 merge->bj = bj; 4820 merge->buf_ri = buf_ri; 4821 merge->buf_rj = buf_rj; 4822 merge->coi = NULL; 4823 merge->coj = NULL; 4824 merge->owners_co = NULL; 4825 4826 ierr = PetscCommDestroy(&comm);CHKERRQ(ierr); 4827 4828 /* attach the supporting struct to B_mpi for reuse */ 4829 ierr = PetscContainerCreate(PETSC_COMM_SELF,&container);CHKERRQ(ierr); 4830 ierr = PetscContainerSetPointer(container,merge);CHKERRQ(ierr); 4831 ierr = PetscObjectCompose((PetscObject)B_mpi,"MatMergeSeqsToMPI",(PetscObject)container);CHKERRQ(ierr); 4832 ierr = PetscContainerDestroy(&container);CHKERRQ(ierr); 4833 *mpimat = B_mpi; 4834 4835 ierr = PetscLogEventEnd(MAT_Seqstompisym,seqmat,0,0,0);CHKERRQ(ierr); 4836 PetscFunctionReturn(0); 4837 } 4838 4839 #undef __FUNCT__ 4840 #define __FUNCT__ "MatCreateMPIAIJSumSeqAIJ" 4841 /*@C 4842 MatCreateMPIAIJSumSeqAIJ - Creates a MPIAIJ matrix by adding sequential 4843 matrices from each processor 4844 4845 Collective on MPI_Comm 4846 4847 Input Parameters: 4848 + comm - the communicators the parallel matrix will live on 4849 . seqmat - the input sequential matrices 4850 . m - number of local rows (or PETSC_DECIDE) 4851 . n - number of local columns (or PETSC_DECIDE) 4852 - scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX 4853 4854 Output Parameter: 4855 . mpimat - the parallel matrix generated 4856 4857 Level: advanced 4858 4859 Notes: 4860 The dimensions of the sequential matrix in each processor MUST be the same. 4861 The input seqmat is included into the container "Mat_Merge_SeqsToMPI", and will be 4862 destroyed when mpimat is destroyed. Call PetscObjectQuery() to access seqmat. 4863 @*/ 4864 PetscErrorCode MatCreateMPIAIJSumSeqAIJ(MPI_Comm comm,Mat seqmat,PetscInt m,PetscInt n,MatReuse scall,Mat *mpimat) 4865 { 4866 PetscErrorCode ierr; 4867 PetscMPIInt size; 4868 4869 PetscFunctionBegin; 4870 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 4871 if (size == 1) { 4872 ierr = PetscLogEventBegin(MAT_Seqstompi,seqmat,0,0,0);CHKERRQ(ierr); 4873 if (scall == MAT_INITIAL_MATRIX) { 4874 ierr = MatDuplicate(seqmat,MAT_COPY_VALUES,mpimat);CHKERRQ(ierr); 4875 } else { 4876 ierr = MatCopy(seqmat,*mpimat,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 4877 } 4878 ierr = PetscLogEventEnd(MAT_Seqstompi,seqmat,0,0,0);CHKERRQ(ierr); 4879 PetscFunctionReturn(0); 4880 } 4881 ierr = PetscLogEventBegin(MAT_Seqstompi,seqmat,0,0,0);CHKERRQ(ierr); 4882 if (scall == MAT_INITIAL_MATRIX) { 4883 ierr = MatCreateMPIAIJSumSeqAIJSymbolic(comm,seqmat,m,n,mpimat);CHKERRQ(ierr); 4884 } 4885 ierr = MatCreateMPIAIJSumSeqAIJNumeric(seqmat,*mpimat);CHKERRQ(ierr); 4886 ierr = PetscLogEventEnd(MAT_Seqstompi,seqmat,0,0,0);CHKERRQ(ierr); 4887 PetscFunctionReturn(0); 4888 } 4889 4890 #undef __FUNCT__ 4891 #define __FUNCT__ "MatMPIAIJGetLocalMat" 4892 /*@ 4893 MatMPIAIJGetLocalMat - Creates a SeqAIJ from a MPIAIJ matrix by taking all its local rows and putting them into a sequential vector with 4894 mlocal rows and n columns. Where mlocal is the row count obtained with MatGetLocalSize() and n is the global column count obtained 4895 with MatGetSize() 4896 4897 Not Collective 4898 4899 Input Parameters: 4900 + A - the matrix 4901 . scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX 4902 4903 Output Parameter: 4904 . A_loc - the local sequential matrix generated 4905 4906 Level: developer 4907 4908 .seealso: MatGetOwnerShipRange(), MatMPIAIJGetLocalMatCondensed() 4909 4910 @*/ 4911 PetscErrorCode MatMPIAIJGetLocalMat(Mat A,MatReuse scall,Mat *A_loc) 4912 { 4913 PetscErrorCode ierr; 4914 Mat_MPIAIJ *mpimat=(Mat_MPIAIJ*)A->data; 4915 Mat_SeqAIJ *mat,*a,*b; 4916 PetscInt *ai,*aj,*bi,*bj,*cmap=mpimat->garray; 4917 MatScalar *aa,*ba,*cam; 4918 PetscScalar *ca; 4919 PetscInt am=A->rmap->n,i,j,k,cstart=A->cmap->rstart; 4920 PetscInt *ci,*cj,col,ncols_d,ncols_o,jo; 4921 PetscBool match; 4922 MPI_Comm comm; 4923 PetscMPIInt size; 4924 4925 PetscFunctionBegin; 4926 ierr = PetscObjectTypeCompare((PetscObject)A,MATMPIAIJ,&match);CHKERRQ(ierr); 4927 if (!match) SETERRQ(PetscObjectComm((PetscObject)A), PETSC_ERR_SUP,"Requires MPIAIJ matrix as input"); 4928 ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 4929 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 4930 if (size == 1 && scall == MAT_REUSE_MATRIX) PetscFunctionReturn(0); 4931 4932 ierr = PetscLogEventBegin(MAT_Getlocalmat,A,0,0,0);CHKERRQ(ierr); 4933 a = (Mat_SeqAIJ*)(mpimat->A)->data; 4934 b = (Mat_SeqAIJ*)(mpimat->B)->data; 4935 ai = a->i; aj = a->j; bi = b->i; bj = b->j; 4936 aa = a->a; ba = b->a; 4937 if (scall == MAT_INITIAL_MATRIX) { 4938 if (size == 1) { 4939 ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,am,A->cmap->N,ai,aj,aa,A_loc);CHKERRQ(ierr); 4940 PetscFunctionReturn(0); 4941 } 4942 4943 ierr = PetscMalloc1((1+am),&ci);CHKERRQ(ierr); 4944 ci[0] = 0; 4945 for (i=0; i<am; i++) { 4946 ci[i+1] = ci[i] + (ai[i+1] - ai[i]) + (bi[i+1] - bi[i]); 4947 } 4948 ierr = PetscMalloc1((1+ci[am]),&cj);CHKERRQ(ierr); 4949 ierr = PetscMalloc1((1+ci[am]),&ca);CHKERRQ(ierr); 4950 k = 0; 4951 for (i=0; i<am; i++) { 4952 ncols_o = bi[i+1] - bi[i]; 4953 ncols_d = ai[i+1] - ai[i]; 4954 /* off-diagonal portion of A */ 4955 for (jo=0; jo<ncols_o; jo++) { 4956 col = cmap[*bj]; 4957 if (col >= cstart) break; 4958 cj[k] = col; bj++; 4959 ca[k++] = *ba++; 4960 } 4961 /* diagonal portion of A */ 4962 for (j=0; j<ncols_d; j++) { 4963 cj[k] = cstart + *aj++; 4964 ca[k++] = *aa++; 4965 } 4966 /* off-diagonal portion of A */ 4967 for (j=jo; j<ncols_o; j++) { 4968 cj[k] = cmap[*bj++]; 4969 ca[k++] = *ba++; 4970 } 4971 } 4972 /* put together the new matrix */ 4973 ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,am,A->cmap->N,ci,cj,ca,A_loc);CHKERRQ(ierr); 4974 /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */ 4975 /* Since these are PETSc arrays, change flags to free them as necessary. */ 4976 mat = (Mat_SeqAIJ*)(*A_loc)->data; 4977 mat->free_a = PETSC_TRUE; 4978 mat->free_ij = PETSC_TRUE; 4979 mat->nonew = 0; 4980 } else if (scall == MAT_REUSE_MATRIX) { 4981 mat=(Mat_SeqAIJ*)(*A_loc)->data; 4982 ci = mat->i; cj = mat->j; cam = mat->a; 4983 for (i=0; i<am; i++) { 4984 /* off-diagonal portion of A */ 4985 ncols_o = bi[i+1] - bi[i]; 4986 for (jo=0; jo<ncols_o; jo++) { 4987 col = cmap[*bj]; 4988 if (col >= cstart) break; 4989 *cam++ = *ba++; bj++; 4990 } 4991 /* diagonal portion of A */ 4992 ncols_d = ai[i+1] - ai[i]; 4993 for (j=0; j<ncols_d; j++) *cam++ = *aa++; 4994 /* off-diagonal portion of A */ 4995 for (j=jo; j<ncols_o; j++) { 4996 *cam++ = *ba++; bj++; 4997 } 4998 } 4999 } else SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Invalid MatReuse %d",(int)scall); 5000 ierr = PetscLogEventEnd(MAT_Getlocalmat,A,0,0,0);CHKERRQ(ierr); 5001 PetscFunctionReturn(0); 5002 } 5003 5004 #undef __FUNCT__ 5005 #define __FUNCT__ "MatMPIAIJGetLocalMatCondensed" 5006 /*@C 5007 MatMPIAIJGetLocalMatCondensed - Creates a SeqAIJ matrix from an MPIAIJ matrix by taking all its local rows and NON-ZERO columns 5008 5009 Not Collective 5010 5011 Input Parameters: 5012 + A - the matrix 5013 . scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX 5014 - row, col - index sets of rows and columns to extract (or NULL) 5015 5016 Output Parameter: 5017 . A_loc - the local sequential matrix generated 5018 5019 Level: developer 5020 5021 .seealso: MatGetOwnershipRange(), MatMPIAIJGetLocalMat() 5022 5023 @*/ 5024 PetscErrorCode MatMPIAIJGetLocalMatCondensed(Mat A,MatReuse scall,IS *row,IS *col,Mat *A_loc) 5025 { 5026 Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data; 5027 PetscErrorCode ierr; 5028 PetscInt i,start,end,ncols,nzA,nzB,*cmap,imark,*idx; 5029 IS isrowa,iscola; 5030 Mat *aloc; 5031 PetscBool match; 5032 5033 PetscFunctionBegin; 5034 ierr = PetscObjectTypeCompare((PetscObject)A,MATMPIAIJ,&match);CHKERRQ(ierr); 5035 if (!match) SETERRQ(PetscObjectComm((PetscObject)A), PETSC_ERR_SUP,"Requires MPIAIJ matrix as input"); 5036 ierr = PetscLogEventBegin(MAT_Getlocalmatcondensed,A,0,0,0);CHKERRQ(ierr); 5037 if (!row) { 5038 start = A->rmap->rstart; end = A->rmap->rend; 5039 ierr = ISCreateStride(PETSC_COMM_SELF,end-start,start,1,&isrowa);CHKERRQ(ierr); 5040 } else { 5041 isrowa = *row; 5042 } 5043 if (!col) { 5044 start = A->cmap->rstart; 5045 cmap = a->garray; 5046 nzA = a->A->cmap->n; 5047 nzB = a->B->cmap->n; 5048 ierr = PetscMalloc1((nzA+nzB), &idx);CHKERRQ(ierr); 5049 ncols = 0; 5050 for (i=0; i<nzB; i++) { 5051 if (cmap[i] < start) idx[ncols++] = cmap[i]; 5052 else break; 5053 } 5054 imark = i; 5055 for (i=0; i<nzA; i++) idx[ncols++] = start + i; 5056 for (i=imark; i<nzB; i++) idx[ncols++] = cmap[i]; 5057 ierr = ISCreateGeneral(PETSC_COMM_SELF,ncols,idx,PETSC_OWN_POINTER,&iscola);CHKERRQ(ierr); 5058 } else { 5059 iscola = *col; 5060 } 5061 if (scall != MAT_INITIAL_MATRIX) { 5062 ierr = PetscMalloc(sizeof(Mat),&aloc);CHKERRQ(ierr); 5063 aloc[0] = *A_loc; 5064 } 5065 ierr = MatGetSubMatrices(A,1,&isrowa,&iscola,scall,&aloc);CHKERRQ(ierr); 5066 *A_loc = aloc[0]; 5067 ierr = PetscFree(aloc);CHKERRQ(ierr); 5068 if (!row) { 5069 ierr = ISDestroy(&isrowa);CHKERRQ(ierr); 5070 } 5071 if (!col) { 5072 ierr = ISDestroy(&iscola);CHKERRQ(ierr); 5073 } 5074 ierr = PetscLogEventEnd(MAT_Getlocalmatcondensed,A,0,0,0);CHKERRQ(ierr); 5075 PetscFunctionReturn(0); 5076 } 5077 5078 #undef __FUNCT__ 5079 #define __FUNCT__ "MatGetBrowsOfAcols" 5080 /*@C 5081 MatGetBrowsOfAcols - Creates a SeqAIJ matrix by taking rows of B that equal to nonzero columns of local A 5082 5083 Collective on Mat 5084 5085 Input Parameters: 5086 + A,B - the matrices in mpiaij format 5087 . scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX 5088 - rowb, colb - index sets of rows and columns of B to extract (or NULL) 5089 5090 Output Parameter: 5091 + rowb, colb - index sets of rows and columns of B to extract 5092 - B_seq - the sequential matrix generated 5093 5094 Level: developer 5095 5096 @*/ 5097 PetscErrorCode MatGetBrowsOfAcols(Mat A,Mat B,MatReuse scall,IS *rowb,IS *colb,Mat *B_seq) 5098 { 5099 Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data; 5100 PetscErrorCode ierr; 5101 PetscInt *idx,i,start,ncols,nzA,nzB,*cmap,imark; 5102 IS isrowb,iscolb; 5103 Mat *bseq=NULL; 5104 5105 PetscFunctionBegin; 5106 if (A->cmap->rstart != B->rmap->rstart || A->cmap->rend != B->rmap->rend) { 5107 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); 5108 } 5109 ierr = PetscLogEventBegin(MAT_GetBrowsOfAcols,A,B,0,0);CHKERRQ(ierr); 5110 5111 if (scall == MAT_INITIAL_MATRIX) { 5112 start = A->cmap->rstart; 5113 cmap = a->garray; 5114 nzA = a->A->cmap->n; 5115 nzB = a->B->cmap->n; 5116 ierr = PetscMalloc1((nzA+nzB), &idx);CHKERRQ(ierr); 5117 ncols = 0; 5118 for (i=0; i<nzB; i++) { /* row < local row index */ 5119 if (cmap[i] < start) idx[ncols++] = cmap[i]; 5120 else break; 5121 } 5122 imark = i; 5123 for (i=0; i<nzA; i++) idx[ncols++] = start + i; /* local rows */ 5124 for (i=imark; i<nzB; i++) idx[ncols++] = cmap[i]; /* row > local row index */ 5125 ierr = ISCreateGeneral(PETSC_COMM_SELF,ncols,idx,PETSC_OWN_POINTER,&isrowb);CHKERRQ(ierr); 5126 ierr = ISCreateStride(PETSC_COMM_SELF,B->cmap->N,0,1,&iscolb);CHKERRQ(ierr); 5127 } else { 5128 if (!rowb || !colb) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"IS rowb and colb must be provided for MAT_REUSE_MATRIX"); 5129 isrowb = *rowb; iscolb = *colb; 5130 ierr = PetscMalloc(sizeof(Mat),&bseq);CHKERRQ(ierr); 5131 bseq[0] = *B_seq; 5132 } 5133 ierr = MatGetSubMatrices(B,1,&isrowb,&iscolb,scall,&bseq);CHKERRQ(ierr); 5134 *B_seq = bseq[0]; 5135 ierr = PetscFree(bseq);CHKERRQ(ierr); 5136 if (!rowb) { 5137 ierr = ISDestroy(&isrowb);CHKERRQ(ierr); 5138 } else { 5139 *rowb = isrowb; 5140 } 5141 if (!colb) { 5142 ierr = ISDestroy(&iscolb);CHKERRQ(ierr); 5143 } else { 5144 *colb = iscolb; 5145 } 5146 ierr = PetscLogEventEnd(MAT_GetBrowsOfAcols,A,B,0,0);CHKERRQ(ierr); 5147 PetscFunctionReturn(0); 5148 } 5149 5150 #undef __FUNCT__ 5151 #define __FUNCT__ "MatGetBrowsOfAoCols_MPIAIJ" 5152 /* 5153 MatGetBrowsOfAoCols_MPIAIJ - Creates a SeqAIJ matrix by taking rows of B that equal to nonzero columns 5154 of the OFF-DIAGONAL portion of local A 5155 5156 Collective on Mat 5157 5158 Input Parameters: 5159 + A,B - the matrices in mpiaij format 5160 - scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX 5161 5162 Output Parameter: 5163 + startsj_s - starting point in B's sending j-arrays, saved for MAT_REUSE (or NULL) 5164 . startsj_r - starting point in B's receiving j-arrays, saved for MAT_REUSE (or NULL) 5165 . bufa_ptr - array for sending matrix values, saved for MAT_REUSE (or NULL) 5166 - B_oth - the sequential matrix generated with size aBn=a->B->cmap->n by B->cmap->N 5167 5168 Level: developer 5169 5170 */ 5171 PetscErrorCode MatGetBrowsOfAoCols_MPIAIJ(Mat A,Mat B,MatReuse scall,PetscInt **startsj_s,PetscInt **startsj_r,MatScalar **bufa_ptr,Mat *B_oth) 5172 { 5173 VecScatter_MPI_General *gen_to,*gen_from; 5174 PetscErrorCode ierr; 5175 Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data; 5176 Mat_SeqAIJ *b_oth; 5177 VecScatter ctx =a->Mvctx; 5178 MPI_Comm comm; 5179 PetscMPIInt *rprocs,*sprocs,tag=((PetscObject)ctx)->tag,rank; 5180 PetscInt *rowlen,*bufj,*bufJ,ncols,aBn=a->B->cmap->n,row,*b_othi,*b_othj; 5181 PetscScalar *rvalues,*svalues; 5182 MatScalar *b_otha,*bufa,*bufA; 5183 PetscInt i,j,k,l,ll,nrecvs,nsends,nrows,*srow,*rstarts,*rstartsj = 0,*sstarts,*sstartsj,len; 5184 MPI_Request *rwaits = NULL,*swaits = NULL; 5185 MPI_Status *sstatus,rstatus; 5186 PetscMPIInt jj,size; 5187 PetscInt *cols,sbs,rbs; 5188 PetscScalar *vals; 5189 5190 PetscFunctionBegin; 5191 ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 5192 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 5193 if (size == 1) PetscFunctionReturn(0); 5194 5195 if (A->cmap->rstart != B->rmap->rstart || A->cmap->rend != B->rmap->rend) { 5196 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); 5197 } 5198 ierr = PetscLogEventBegin(MAT_GetBrowsOfAocols,A,B,0,0);CHKERRQ(ierr); 5199 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 5200 5201 gen_to = (VecScatter_MPI_General*)ctx->todata; 5202 gen_from = (VecScatter_MPI_General*)ctx->fromdata; 5203 rvalues = gen_from->values; /* holds the length of receiving row */ 5204 svalues = gen_to->values; /* holds the length of sending row */ 5205 nrecvs = gen_from->n; 5206 nsends = gen_to->n; 5207 5208 ierr = PetscMalloc2(nrecvs,&rwaits,nsends,&swaits);CHKERRQ(ierr); 5209 srow = gen_to->indices; /* local row index to be sent */ 5210 sstarts = gen_to->starts; 5211 sprocs = gen_to->procs; 5212 sstatus = gen_to->sstatus; 5213 sbs = gen_to->bs; 5214 rstarts = gen_from->starts; 5215 rprocs = gen_from->procs; 5216 rbs = gen_from->bs; 5217 5218 if (!startsj_s || !bufa_ptr) scall = MAT_INITIAL_MATRIX; 5219 if (scall == MAT_INITIAL_MATRIX) { 5220 /* i-array */ 5221 /*---------*/ 5222 /* post receives */ 5223 for (i=0; i<nrecvs; i++) { 5224 rowlen = (PetscInt*)rvalues + rstarts[i]*rbs; 5225 nrows = (rstarts[i+1]-rstarts[i])*rbs; /* num of indices to be received */ 5226 ierr = MPI_Irecv(rowlen,nrows,MPIU_INT,rprocs[i],tag,comm,rwaits+i);CHKERRQ(ierr); 5227 } 5228 5229 /* pack the outgoing message */ 5230 ierr = PetscMalloc2(nsends+1,&sstartsj,nrecvs+1,&rstartsj);CHKERRQ(ierr); 5231 5232 sstartsj[0] = 0; 5233 rstartsj[0] = 0; 5234 len = 0; /* total length of j or a array to be sent */ 5235 k = 0; 5236 for (i=0; i<nsends; i++) { 5237 rowlen = (PetscInt*)svalues + sstarts[i]*sbs; 5238 nrows = sstarts[i+1]-sstarts[i]; /* num of block rows */ 5239 for (j=0; j<nrows; j++) { 5240 row = srow[k] + B->rmap->range[rank]; /* global row idx */ 5241 for (l=0; l<sbs; l++) { 5242 ierr = MatGetRow_MPIAIJ(B,row+l,&ncols,NULL,NULL);CHKERRQ(ierr); /* rowlength */ 5243 5244 rowlen[j*sbs+l] = ncols; 5245 5246 len += ncols; 5247 ierr = MatRestoreRow_MPIAIJ(B,row+l,&ncols,NULL,NULL);CHKERRQ(ierr); 5248 } 5249 k++; 5250 } 5251 ierr = MPI_Isend(rowlen,nrows*sbs,MPIU_INT,sprocs[i],tag,comm,swaits+i);CHKERRQ(ierr); 5252 5253 sstartsj[i+1] = len; /* starting point of (i+1)-th outgoing msg in bufj and bufa */ 5254 } 5255 /* recvs and sends of i-array are completed */ 5256 i = nrecvs; 5257 while (i--) { 5258 ierr = MPI_Waitany(nrecvs,rwaits,&jj,&rstatus);CHKERRQ(ierr); 5259 } 5260 if (nsends) {ierr = MPI_Waitall(nsends,swaits,sstatus);CHKERRQ(ierr);} 5261 5262 /* allocate buffers for sending j and a arrays */ 5263 ierr = PetscMalloc1((len+1),&bufj);CHKERRQ(ierr); 5264 ierr = PetscMalloc1((len+1),&bufa);CHKERRQ(ierr); 5265 5266 /* create i-array of B_oth */ 5267 ierr = PetscMalloc1((aBn+2),&b_othi);CHKERRQ(ierr); 5268 5269 b_othi[0] = 0; 5270 len = 0; /* total length of j or a array to be received */ 5271 k = 0; 5272 for (i=0; i<nrecvs; i++) { 5273 rowlen = (PetscInt*)rvalues + rstarts[i]*rbs; 5274 nrows = rbs*(rstarts[i+1]-rstarts[i]); /* num of rows to be recieved */ 5275 for (j=0; j<nrows; j++) { 5276 b_othi[k+1] = b_othi[k] + rowlen[j]; 5277 len += rowlen[j]; k++; 5278 } 5279 rstartsj[i+1] = len; /* starting point of (i+1)-th incoming msg in bufj and bufa */ 5280 } 5281 5282 /* allocate space for j and a arrrays of B_oth */ 5283 ierr = PetscMalloc1((b_othi[aBn]+1),&b_othj);CHKERRQ(ierr); 5284 ierr = PetscMalloc1((b_othi[aBn]+1),&b_otha);CHKERRQ(ierr); 5285 5286 /* j-array */ 5287 /*---------*/ 5288 /* post receives of j-array */ 5289 for (i=0; i<nrecvs; i++) { 5290 nrows = rstartsj[i+1]-rstartsj[i]; /* length of the msg received */ 5291 ierr = MPI_Irecv(b_othj+rstartsj[i],nrows,MPIU_INT,rprocs[i],tag,comm,rwaits+i);CHKERRQ(ierr); 5292 } 5293 5294 /* pack the outgoing message j-array */ 5295 k = 0; 5296 for (i=0; i<nsends; i++) { 5297 nrows = sstarts[i+1]-sstarts[i]; /* num of block rows */ 5298 bufJ = bufj+sstartsj[i]; 5299 for (j=0; j<nrows; j++) { 5300 row = srow[k++] + B->rmap->range[rank]; /* global row idx */ 5301 for (ll=0; ll<sbs; ll++) { 5302 ierr = MatGetRow_MPIAIJ(B,row+ll,&ncols,&cols,NULL);CHKERRQ(ierr); 5303 for (l=0; l<ncols; l++) { 5304 *bufJ++ = cols[l]; 5305 } 5306 ierr = MatRestoreRow_MPIAIJ(B,row+ll,&ncols,&cols,NULL);CHKERRQ(ierr); 5307 } 5308 } 5309 ierr = MPI_Isend(bufj+sstartsj[i],sstartsj[i+1]-sstartsj[i],MPIU_INT,sprocs[i],tag,comm,swaits+i);CHKERRQ(ierr); 5310 } 5311 5312 /* recvs and sends of j-array are completed */ 5313 i = nrecvs; 5314 while (i--) { 5315 ierr = MPI_Waitany(nrecvs,rwaits,&jj,&rstatus);CHKERRQ(ierr); 5316 } 5317 if (nsends) {ierr = MPI_Waitall(nsends,swaits,sstatus);CHKERRQ(ierr);} 5318 } else if (scall == MAT_REUSE_MATRIX) { 5319 sstartsj = *startsj_s; 5320 rstartsj = *startsj_r; 5321 bufa = *bufa_ptr; 5322 b_oth = (Mat_SeqAIJ*)(*B_oth)->data; 5323 b_otha = b_oth->a; 5324 } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE, "Matrix P does not posses an object container"); 5325 5326 /* a-array */ 5327 /*---------*/ 5328 /* post receives of a-array */ 5329 for (i=0; i<nrecvs; i++) { 5330 nrows = rstartsj[i+1]-rstartsj[i]; /* length of the msg received */ 5331 ierr = MPI_Irecv(b_otha+rstartsj[i],nrows,MPIU_SCALAR,rprocs[i],tag,comm,rwaits+i);CHKERRQ(ierr); 5332 } 5333 5334 /* pack the outgoing message a-array */ 5335 k = 0; 5336 for (i=0; i<nsends; i++) { 5337 nrows = sstarts[i+1]-sstarts[i]; /* num of block rows */ 5338 bufA = bufa+sstartsj[i]; 5339 for (j=0; j<nrows; j++) { 5340 row = srow[k++] + B->rmap->range[rank]; /* global row idx */ 5341 for (ll=0; ll<sbs; ll++) { 5342 ierr = MatGetRow_MPIAIJ(B,row+ll,&ncols,NULL,&vals);CHKERRQ(ierr); 5343 for (l=0; l<ncols; l++) { 5344 *bufA++ = vals[l]; 5345 } 5346 ierr = MatRestoreRow_MPIAIJ(B,row+ll,&ncols,NULL,&vals);CHKERRQ(ierr); 5347 } 5348 } 5349 ierr = MPI_Isend(bufa+sstartsj[i],sstartsj[i+1]-sstartsj[i],MPIU_SCALAR,sprocs[i],tag,comm,swaits+i);CHKERRQ(ierr); 5350 } 5351 /* recvs and sends of a-array are completed */ 5352 i = nrecvs; 5353 while (i--) { 5354 ierr = MPI_Waitany(nrecvs,rwaits,&jj,&rstatus);CHKERRQ(ierr); 5355 } 5356 if (nsends) {ierr = MPI_Waitall(nsends,swaits,sstatus);CHKERRQ(ierr);} 5357 ierr = PetscFree2(rwaits,swaits);CHKERRQ(ierr); 5358 5359 if (scall == MAT_INITIAL_MATRIX) { 5360 /* put together the new matrix */ 5361 ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,aBn,B->cmap->N,b_othi,b_othj,b_otha,B_oth);CHKERRQ(ierr); 5362 5363 /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */ 5364 /* Since these are PETSc arrays, change flags to free them as necessary. */ 5365 b_oth = (Mat_SeqAIJ*)(*B_oth)->data; 5366 b_oth->free_a = PETSC_TRUE; 5367 b_oth->free_ij = PETSC_TRUE; 5368 b_oth->nonew = 0; 5369 5370 ierr = PetscFree(bufj);CHKERRQ(ierr); 5371 if (!startsj_s || !bufa_ptr) { 5372 ierr = PetscFree2(sstartsj,rstartsj);CHKERRQ(ierr); 5373 ierr = PetscFree(bufa_ptr);CHKERRQ(ierr); 5374 } else { 5375 *startsj_s = sstartsj; 5376 *startsj_r = rstartsj; 5377 *bufa_ptr = bufa; 5378 } 5379 } 5380 ierr = PetscLogEventEnd(MAT_GetBrowsOfAocols,A,B,0,0);CHKERRQ(ierr); 5381 PetscFunctionReturn(0); 5382 } 5383 5384 #undef __FUNCT__ 5385 #define __FUNCT__ "MatGetCommunicationStructs" 5386 /*@C 5387 MatGetCommunicationStructs - Provides access to the communication structures used in matrix-vector multiplication. 5388 5389 Not Collective 5390 5391 Input Parameters: 5392 . A - The matrix in mpiaij format 5393 5394 Output Parameter: 5395 + lvec - The local vector holding off-process values from the argument to a matrix-vector product 5396 . colmap - A map from global column index to local index into lvec 5397 - multScatter - A scatter from the argument of a matrix-vector product to lvec 5398 5399 Level: developer 5400 5401 @*/ 5402 #if defined(PETSC_USE_CTABLE) 5403 PetscErrorCode MatGetCommunicationStructs(Mat A, Vec *lvec, PetscTable *colmap, VecScatter *multScatter) 5404 #else 5405 PetscErrorCode MatGetCommunicationStructs(Mat A, Vec *lvec, PetscInt *colmap[], VecScatter *multScatter) 5406 #endif 5407 { 5408 Mat_MPIAIJ *a; 5409 5410 PetscFunctionBegin; 5411 PetscValidHeaderSpecific(A, MAT_CLASSID, 1); 5412 PetscValidPointer(lvec, 2); 5413 PetscValidPointer(colmap, 3); 5414 PetscValidPointer(multScatter, 4); 5415 a = (Mat_MPIAIJ*) A->data; 5416 if (lvec) *lvec = a->lvec; 5417 if (colmap) *colmap = a->colmap; 5418 if (multScatter) *multScatter = a->Mvctx; 5419 PetscFunctionReturn(0); 5420 } 5421 5422 PETSC_EXTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJCRL(Mat,MatType,MatReuse,Mat*); 5423 PETSC_EXTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJPERM(Mat,MatType,MatReuse,Mat*); 5424 PETSC_EXTERN PetscErrorCode MatConvert_MPIAIJ_MPISBAIJ(Mat,MatType,MatReuse,Mat*); 5425 #if defined(PETSC_HAVE_ELEMENTAL) 5426 PETSC_EXTERN PetscErrorCode MatConvert_MPIAIJ_Elemental(Mat,MatType,MatReuse,Mat*); 5427 #endif 5428 5429 #undef __FUNCT__ 5430 #define __FUNCT__ "MatMatMultNumeric_MPIDense_MPIAIJ" 5431 /* 5432 Computes (B'*A')' since computing B*A directly is untenable 5433 5434 n p p 5435 ( ) ( ) ( ) 5436 m ( A ) * n ( B ) = m ( C ) 5437 ( ) ( ) ( ) 5438 5439 */ 5440 PetscErrorCode MatMatMultNumeric_MPIDense_MPIAIJ(Mat A,Mat B,Mat C) 5441 { 5442 PetscErrorCode ierr; 5443 Mat At,Bt,Ct; 5444 5445 PetscFunctionBegin; 5446 ierr = MatTranspose(A,MAT_INITIAL_MATRIX,&At);CHKERRQ(ierr); 5447 ierr = MatTranspose(B,MAT_INITIAL_MATRIX,&Bt);CHKERRQ(ierr); 5448 ierr = MatMatMult(Bt,At,MAT_INITIAL_MATRIX,1.0,&Ct);CHKERRQ(ierr); 5449 ierr = MatDestroy(&At);CHKERRQ(ierr); 5450 ierr = MatDestroy(&Bt);CHKERRQ(ierr); 5451 ierr = MatTranspose(Ct,MAT_REUSE_MATRIX,&C);CHKERRQ(ierr); 5452 ierr = MatDestroy(&Ct);CHKERRQ(ierr); 5453 PetscFunctionReturn(0); 5454 } 5455 5456 #undef __FUNCT__ 5457 #define __FUNCT__ "MatMatMultSymbolic_MPIDense_MPIAIJ" 5458 PetscErrorCode MatMatMultSymbolic_MPIDense_MPIAIJ(Mat A,Mat B,PetscReal fill,Mat *C) 5459 { 5460 PetscErrorCode ierr; 5461 PetscInt m=A->rmap->n,n=B->cmap->n; 5462 Mat Cmat; 5463 5464 PetscFunctionBegin; 5465 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); 5466 ierr = MatCreate(PetscObjectComm((PetscObject)A),&Cmat);CHKERRQ(ierr); 5467 ierr = MatSetSizes(Cmat,m,n,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 5468 ierr = MatSetBlockSizesFromMats(Cmat,A,B);CHKERRQ(ierr); 5469 ierr = MatSetType(Cmat,MATMPIDENSE);CHKERRQ(ierr); 5470 ierr = MatMPIDenseSetPreallocation(Cmat,NULL);CHKERRQ(ierr); 5471 ierr = MatAssemblyBegin(Cmat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5472 ierr = MatAssemblyEnd(Cmat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5473 5474 Cmat->ops->matmultnumeric = MatMatMultNumeric_MPIDense_MPIAIJ; 5475 5476 *C = Cmat; 5477 PetscFunctionReturn(0); 5478 } 5479 5480 /* ----------------------------------------------------------------*/ 5481 #undef __FUNCT__ 5482 #define __FUNCT__ "MatMatMult_MPIDense_MPIAIJ" 5483 PetscErrorCode MatMatMult_MPIDense_MPIAIJ(Mat A,Mat B,MatReuse scall,PetscReal fill,Mat *C) 5484 { 5485 PetscErrorCode ierr; 5486 5487 PetscFunctionBegin; 5488 if (scall == MAT_INITIAL_MATRIX) { 5489 ierr = PetscLogEventBegin(MAT_MatMultSymbolic,A,B,0,0);CHKERRQ(ierr); 5490 ierr = MatMatMultSymbolic_MPIDense_MPIAIJ(A,B,fill,C);CHKERRQ(ierr); 5491 ierr = PetscLogEventEnd(MAT_MatMultSymbolic,A,B,0,0);CHKERRQ(ierr); 5492 } 5493 ierr = PetscLogEventBegin(MAT_MatMultNumeric,A,B,0,0);CHKERRQ(ierr); 5494 ierr = MatMatMultNumeric_MPIDense_MPIAIJ(A,B,*C);CHKERRQ(ierr); 5495 ierr = PetscLogEventEnd(MAT_MatMultNumeric,A,B,0,0);CHKERRQ(ierr); 5496 PetscFunctionReturn(0); 5497 } 5498 5499 #if defined(PETSC_HAVE_MUMPS) 5500 PETSC_EXTERN PetscErrorCode MatGetFactor_aij_mumps(Mat,MatFactorType,Mat*); 5501 #endif 5502 #if defined(PETSC_HAVE_PASTIX) 5503 PETSC_EXTERN PetscErrorCode MatGetFactor_mpiaij_pastix(Mat,MatFactorType,Mat*); 5504 #endif 5505 #if defined(PETSC_HAVE_SUPERLU_DIST) 5506 PETSC_EXTERN PetscErrorCode MatGetFactor_mpiaij_superlu_dist(Mat,MatFactorType,Mat*); 5507 #endif 5508 #if defined(PETSC_HAVE_CLIQUE) 5509 PETSC_EXTERN PetscErrorCode MatGetFactor_aij_clique(Mat,MatFactorType,Mat*); 5510 #endif 5511 5512 /*MC 5513 MATMPIAIJ - MATMPIAIJ = "mpiaij" - A matrix type to be used for parallel sparse matrices. 5514 5515 Options Database Keys: 5516 . -mat_type mpiaij - sets the matrix type to "mpiaij" during a call to MatSetFromOptions() 5517 5518 Level: beginner 5519 5520 .seealso: MatCreateAIJ() 5521 M*/ 5522 5523 #undef __FUNCT__ 5524 #define __FUNCT__ "MatCreate_MPIAIJ" 5525 PETSC_EXTERN PetscErrorCode MatCreate_MPIAIJ(Mat B) 5526 { 5527 Mat_MPIAIJ *b; 5528 PetscErrorCode ierr; 5529 PetscMPIInt size; 5530 5531 PetscFunctionBegin; 5532 ierr = MPI_Comm_size(PetscObjectComm((PetscObject)B),&size);CHKERRQ(ierr); 5533 5534 ierr = PetscNewLog(B,&b);CHKERRQ(ierr); 5535 B->data = (void*)b; 5536 ierr = PetscMemcpy(B->ops,&MatOps_Values,sizeof(struct _MatOps));CHKERRQ(ierr); 5537 B->assembled = PETSC_FALSE; 5538 B->insertmode = NOT_SET_VALUES; 5539 b->size = size; 5540 5541 ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)B),&b->rank);CHKERRQ(ierr); 5542 5543 /* build cache for off array entries formed */ 5544 ierr = MatStashCreate_Private(PetscObjectComm((PetscObject)B),1,&B->stash);CHKERRQ(ierr); 5545 5546 b->donotstash = PETSC_FALSE; 5547 b->colmap = 0; 5548 b->garray = 0; 5549 b->roworiented = PETSC_TRUE; 5550 5551 /* stuff used for matrix vector multiply */ 5552 b->lvec = NULL; 5553 b->Mvctx = NULL; 5554 5555 /* stuff for MatGetRow() */ 5556 b->rowindices = 0; 5557 b->rowvalues = 0; 5558 b->getrowactive = PETSC_FALSE; 5559 5560 /* flexible pointer used in CUSP/CUSPARSE classes */ 5561 b->spptr = NULL; 5562 5563 #if defined(PETSC_HAVE_MUMPS) 5564 ierr = PetscObjectComposeFunction((PetscObject)B,"MatGetFactor_mumps_C",MatGetFactor_aij_mumps);CHKERRQ(ierr); 5565 #endif 5566 #if defined(PETSC_HAVE_PASTIX) 5567 ierr = PetscObjectComposeFunction((PetscObject)B,"MatGetFactor_pastix_C",MatGetFactor_mpiaij_pastix);CHKERRQ(ierr); 5568 #endif 5569 #if defined(PETSC_HAVE_SUPERLU_DIST) 5570 ierr = PetscObjectComposeFunction((PetscObject)B,"MatGetFactor_superlu_dist_C",MatGetFactor_mpiaij_superlu_dist);CHKERRQ(ierr); 5571 #endif 5572 #if defined(PETSC_HAVE_CLIQUE) 5573 ierr = PetscObjectComposeFunction((PetscObject)B,"MatGetFactor_clique_C",MatGetFactor_aij_clique);CHKERRQ(ierr); 5574 #endif 5575 ierr = PetscObjectComposeFunction((PetscObject)B,"MatStoreValues_C",MatStoreValues_MPIAIJ);CHKERRQ(ierr); 5576 ierr = PetscObjectComposeFunction((PetscObject)B,"MatRetrieveValues_C",MatRetrieveValues_MPIAIJ);CHKERRQ(ierr); 5577 ierr = PetscObjectComposeFunction((PetscObject)B,"MatGetDiagonalBlock_C",MatGetDiagonalBlock_MPIAIJ);CHKERRQ(ierr); 5578 ierr = PetscObjectComposeFunction((PetscObject)B,"MatIsTranspose_C",MatIsTranspose_MPIAIJ);CHKERRQ(ierr); 5579 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMPIAIJSetPreallocation_C",MatMPIAIJSetPreallocation_MPIAIJ);CHKERRQ(ierr); 5580 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMPIAIJSetPreallocationCSR_C",MatMPIAIJSetPreallocationCSR_MPIAIJ);CHKERRQ(ierr); 5581 ierr = PetscObjectComposeFunction((PetscObject)B,"MatDiagonalScaleLocal_C",MatDiagonalScaleLocal_MPIAIJ);CHKERRQ(ierr); 5582 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_mpiaijperm_C",MatConvert_MPIAIJ_MPIAIJPERM);CHKERRQ(ierr); 5583 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_mpiaijcrl_C",MatConvert_MPIAIJ_MPIAIJCRL);CHKERRQ(ierr); 5584 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_mpisbaij_C",MatConvert_MPIAIJ_MPISBAIJ);CHKERRQ(ierr); 5585 #if defined(PETSC_HAVE_ELEMENTAL) 5586 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_elemental_C",MatConvert_MPIAIJ_Elemental);CHKERRQ(ierr); 5587 #endif 5588 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMatMult_mpidense_mpiaij_C",MatMatMult_MPIDense_MPIAIJ);CHKERRQ(ierr); 5589 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMatMultSymbolic_mpidense_mpiaij_C",MatMatMultSymbolic_MPIDense_MPIAIJ);CHKERRQ(ierr); 5590 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMatMultNumeric_mpidense_mpiaij_C",MatMatMultNumeric_MPIDense_MPIAIJ);CHKERRQ(ierr); 5591 ierr = PetscObjectChangeTypeName((PetscObject)B,MATMPIAIJ);CHKERRQ(ierr); 5592 PetscFunctionReturn(0); 5593 } 5594 5595 #undef __FUNCT__ 5596 #define __FUNCT__ "MatCreateMPIAIJWithSplitArrays" 5597 /*@ 5598 MatCreateMPIAIJWithSplitArrays - creates a MPI AIJ matrix using arrays that contain the "diagonal" 5599 and "off-diagonal" part of the matrix in CSR format. 5600 5601 Collective on MPI_Comm 5602 5603 Input Parameters: 5604 + comm - MPI communicator 5605 . m - number of local rows (Cannot be PETSC_DECIDE) 5606 . n - This value should be the same as the local size used in creating the 5607 x vector for the matrix-vector product y = Ax. (or PETSC_DECIDE to have 5608 calculated if N is given) For square matrices n is almost always m. 5609 . M - number of global rows (or PETSC_DETERMINE to have calculated if m is given) 5610 . N - number of global columns (or PETSC_DETERMINE to have calculated if n is given) 5611 . i - row indices for "diagonal" portion of matrix 5612 . j - column indices 5613 . a - matrix values 5614 . oi - row indices for "off-diagonal" portion of matrix 5615 . oj - column indices 5616 - oa - matrix values 5617 5618 Output Parameter: 5619 . mat - the matrix 5620 5621 Level: advanced 5622 5623 Notes: 5624 The i, j, and a arrays ARE NOT copied by this routine into the internal format used by PETSc. The user 5625 must free the arrays once the matrix has been destroyed and not before. 5626 5627 The i and j indices are 0 based 5628 5629 See MatCreateAIJ() for the definition of "diagonal" and "off-diagonal" portion of the matrix 5630 5631 This sets local rows and cannot be used to set off-processor values. 5632 5633 Use of this routine is discouraged because it is inflexible and cumbersome to use. It is extremely rare that a 5634 legacy application natively assembles into exactly this split format. The code to do so is nontrivial and does 5635 not easily support in-place reassembly. It is recommended to use MatSetValues() (or a variant thereof) because 5636 the resulting assembly is easier to implement, will work with any matrix format, and the user does not have to 5637 keep track of the underlying array. Use MatSetOption(A,MAT_IGNORE_OFF_PROC_ENTRIES,PETSC_TRUE) to disable all 5638 communication if it is known that only local entries will be set. 5639 5640 .keywords: matrix, aij, compressed row, sparse, parallel 5641 5642 .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatMPIAIJSetPreallocationCSR(), 5643 MPIAIJ, MatCreateAIJ(), MatCreateMPIAIJWithArrays() 5644 @*/ 5645 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) 5646 { 5647 PetscErrorCode ierr; 5648 Mat_MPIAIJ *maij; 5649 5650 PetscFunctionBegin; 5651 if (m < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"local number of rows (m) cannot be PETSC_DECIDE, or negative"); 5652 if (i[0]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"i (row indices) must start with 0"); 5653 if (oi[0]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"oi (row indices) must start with 0"); 5654 ierr = MatCreate(comm,mat);CHKERRQ(ierr); 5655 ierr = MatSetSizes(*mat,m,n,M,N);CHKERRQ(ierr); 5656 ierr = MatSetType(*mat,MATMPIAIJ);CHKERRQ(ierr); 5657 maij = (Mat_MPIAIJ*) (*mat)->data; 5658 5659 (*mat)->preallocated = PETSC_TRUE; 5660 5661 ierr = PetscLayoutSetUp((*mat)->rmap);CHKERRQ(ierr); 5662 ierr = PetscLayoutSetUp((*mat)->cmap);CHKERRQ(ierr); 5663 5664 ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,m,n,i,j,a,&maij->A);CHKERRQ(ierr); 5665 ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,m,(*mat)->cmap->N,oi,oj,oa,&maij->B);CHKERRQ(ierr); 5666 5667 ierr = MatAssemblyBegin(maij->A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5668 ierr = MatAssemblyEnd(maij->A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5669 ierr = MatAssemblyBegin(maij->B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5670 ierr = MatAssemblyEnd(maij->B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5671 5672 ierr = MatAssemblyBegin(*mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5673 ierr = MatAssemblyEnd(*mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5674 ierr = MatSetOption(*mat,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE);CHKERRQ(ierr); 5675 PetscFunctionReturn(0); 5676 } 5677 5678 /* 5679 Special version for direct calls from Fortran 5680 */ 5681 #include <petsc-private/fortranimpl.h> 5682 5683 #if defined(PETSC_HAVE_FORTRAN_CAPS) 5684 #define matsetvaluesmpiaij_ MATSETVALUESMPIAIJ 5685 #elif !defined(PETSC_HAVE_FORTRAN_UNDERSCORE) 5686 #define matsetvaluesmpiaij_ matsetvaluesmpiaij 5687 #endif 5688 5689 /* Change these macros so can be used in void function */ 5690 #undef CHKERRQ 5691 #define CHKERRQ(ierr) CHKERRABORT(PETSC_COMM_WORLD,ierr) 5692 #undef SETERRQ2 5693 #define SETERRQ2(comm,ierr,b,c,d) CHKERRABORT(comm,ierr) 5694 #undef SETERRQ3 5695 #define SETERRQ3(comm,ierr,b,c,d,e) CHKERRABORT(comm,ierr) 5696 #undef SETERRQ 5697 #define SETERRQ(c,ierr,b) CHKERRABORT(c,ierr) 5698 5699 #undef __FUNCT__ 5700 #define __FUNCT__ "matsetvaluesmpiaij_" 5701 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) 5702 { 5703 Mat mat = *mmat; 5704 PetscInt m = *mm, n = *mn; 5705 InsertMode addv = *maddv; 5706 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 5707 PetscScalar value; 5708 PetscErrorCode ierr; 5709 5710 MatCheckPreallocated(mat,1); 5711 if (mat->insertmode == NOT_SET_VALUES) mat->insertmode = addv; 5712 5713 #if defined(PETSC_USE_DEBUG) 5714 else if (mat->insertmode != addv) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Cannot mix add values and insert values"); 5715 #endif 5716 { 5717 PetscInt i,j,rstart = mat->rmap->rstart,rend = mat->rmap->rend; 5718 PetscInt cstart = mat->cmap->rstart,cend = mat->cmap->rend,row,col; 5719 PetscBool roworiented = aij->roworiented; 5720 5721 /* Some Variables required in the macro */ 5722 Mat A = aij->A; 5723 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 5724 PetscInt *aimax = a->imax,*ai = a->i,*ailen = a->ilen,*aj = a->j; 5725 MatScalar *aa = a->a; 5726 PetscBool ignorezeroentries = (((a->ignorezeroentries)&&(addv==ADD_VALUES)) ? PETSC_TRUE : PETSC_FALSE); 5727 Mat B = aij->B; 5728 Mat_SeqAIJ *b = (Mat_SeqAIJ*)B->data; 5729 PetscInt *bimax = b->imax,*bi = b->i,*bilen = b->ilen,*bj = b->j,bm = aij->B->rmap->n,am = aij->A->rmap->n; 5730 MatScalar *ba = b->a; 5731 5732 PetscInt *rp1,*rp2,ii,nrow1,nrow2,_i,rmax1,rmax2,N,low1,high1,low2,high2,t,lastcol1,lastcol2; 5733 PetscInt nonew = a->nonew; 5734 MatScalar *ap1,*ap2; 5735 5736 PetscFunctionBegin; 5737 for (i=0; i<m; i++) { 5738 if (im[i] < 0) continue; 5739 #if defined(PETSC_USE_DEBUG) 5740 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); 5741 #endif 5742 if (im[i] >= rstart && im[i] < rend) { 5743 row = im[i] - rstart; 5744 lastcol1 = -1; 5745 rp1 = aj + ai[row]; 5746 ap1 = aa + ai[row]; 5747 rmax1 = aimax[row]; 5748 nrow1 = ailen[row]; 5749 low1 = 0; 5750 high1 = nrow1; 5751 lastcol2 = -1; 5752 rp2 = bj + bi[row]; 5753 ap2 = ba + bi[row]; 5754 rmax2 = bimax[row]; 5755 nrow2 = bilen[row]; 5756 low2 = 0; 5757 high2 = nrow2; 5758 5759 for (j=0; j<n; j++) { 5760 if (roworiented) value = v[i*n+j]; 5761 else value = v[i+j*m]; 5762 if (ignorezeroentries && value == 0.0 && (addv == ADD_VALUES)) continue; 5763 if (in[j] >= cstart && in[j] < cend) { 5764 col = in[j] - cstart; 5765 MatSetValues_SeqAIJ_A_Private(row,col,value,addv); 5766 } else if (in[j] < 0) continue; 5767 #if defined(PETSC_USE_DEBUG) 5768 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); 5769 #endif 5770 else { 5771 if (mat->was_assembled) { 5772 if (!aij->colmap) { 5773 ierr = MatCreateColmap_MPIAIJ_Private(mat);CHKERRQ(ierr); 5774 } 5775 #if defined(PETSC_USE_CTABLE) 5776 ierr = PetscTableFind(aij->colmap,in[j]+1,&col);CHKERRQ(ierr); 5777 col--; 5778 #else 5779 col = aij->colmap[in[j]] - 1; 5780 #endif 5781 if (col < 0 && !((Mat_SeqAIJ*)(aij->A->data))->nonew) { 5782 ierr = MatDisAssemble_MPIAIJ(mat);CHKERRQ(ierr); 5783 col = in[j]; 5784 /* Reinitialize the variables required by MatSetValues_SeqAIJ_B_Private() */ 5785 B = aij->B; 5786 b = (Mat_SeqAIJ*)B->data; 5787 bimax = b->imax; bi = b->i; bilen = b->ilen; bj = b->j; 5788 rp2 = bj + bi[row]; 5789 ap2 = ba + bi[row]; 5790 rmax2 = bimax[row]; 5791 nrow2 = bilen[row]; 5792 low2 = 0; 5793 high2 = nrow2; 5794 bm = aij->B->rmap->n; 5795 ba = b->a; 5796 } 5797 } else col = in[j]; 5798 MatSetValues_SeqAIJ_B_Private(row,col,value,addv); 5799 } 5800 } 5801 } else if (!aij->donotstash) { 5802 if (roworiented) { 5803 ierr = MatStashValuesRow_Private(&mat->stash,im[i],n,in,v+i*n,(PetscBool)(ignorezeroentries && (addv == ADD_VALUES)));CHKERRQ(ierr); 5804 } else { 5805 ierr = MatStashValuesCol_Private(&mat->stash,im[i],n,in,v+i,m,(PetscBool)(ignorezeroentries && (addv == ADD_VALUES)));CHKERRQ(ierr); 5806 } 5807 } 5808 } 5809 } 5810 PetscFunctionReturnVoid(); 5811 } 5812 5813