1 2 #include <../src/mat/impls/baij/mpi/mpibaij.h> /*I "petscmat.h" I*/ 3 #include <../src/mat/impls/sbaij/mpi/mpisbaij.h> 4 #include <../src/mat/impls/sbaij/seq/sbaij.h> 5 #include <petscblaslapack.h> 6 7 #if defined(PETSC_HAVE_ELEMENTAL) 8 PETSC_EXTERN PetscErrorCode MatConvert_MPISBAIJ_Elemental(Mat,MatType,MatReuse,Mat*); 9 #endif 10 #undef __FUNCT__ 11 #define __FUNCT__ "MatStoreValues_MPISBAIJ" 12 PetscErrorCode MatStoreValues_MPISBAIJ(Mat mat) 13 { 14 Mat_MPISBAIJ *aij = (Mat_MPISBAIJ*)mat->data; 15 PetscErrorCode ierr; 16 17 PetscFunctionBegin; 18 ierr = MatStoreValues(aij->A);CHKERRQ(ierr); 19 ierr = MatStoreValues(aij->B);CHKERRQ(ierr); 20 PetscFunctionReturn(0); 21 } 22 23 #undef __FUNCT__ 24 #define __FUNCT__ "MatRetrieveValues_MPISBAIJ" 25 PetscErrorCode MatRetrieveValues_MPISBAIJ(Mat mat) 26 { 27 Mat_MPISBAIJ *aij = (Mat_MPISBAIJ*)mat->data; 28 PetscErrorCode ierr; 29 30 PetscFunctionBegin; 31 ierr = MatRetrieveValues(aij->A);CHKERRQ(ierr); 32 ierr = MatRetrieveValues(aij->B);CHKERRQ(ierr); 33 PetscFunctionReturn(0); 34 } 35 36 #define MatSetValues_SeqSBAIJ_A_Private(row,col,value,addv,orow,ocol) \ 37 { \ 38 \ 39 brow = row/bs; \ 40 rp = aj + ai[brow]; ap = aa + bs2*ai[brow]; \ 41 rmax = aimax[brow]; nrow = ailen[brow]; \ 42 bcol = col/bs; \ 43 ridx = row % bs; cidx = col % bs; \ 44 low = 0; high = nrow; \ 45 while (high-low > 3) { \ 46 t = (low+high)/2; \ 47 if (rp[t] > bcol) high = t; \ 48 else low = t; \ 49 } \ 50 for (_i=low; _i<high; _i++) { \ 51 if (rp[_i] > bcol) break; \ 52 if (rp[_i] == bcol) { \ 53 bap = ap + bs2*_i + bs*cidx + ridx; \ 54 if (addv == ADD_VALUES) *bap += value; \ 55 else *bap = value; \ 56 goto a_noinsert; \ 57 } \ 58 } \ 59 if (a->nonew == 1) goto a_noinsert; \ 60 if (a->nonew == -1) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero at global row/column (%D, %D) into matrix", orow, ocol); \ 61 MatSeqXAIJReallocateAIJ(A,a->mbs,bs2,nrow,brow,bcol,rmax,aa,ai,aj,rp,ap,aimax,a->nonew,MatScalar); \ 62 N = nrow++ - 1; \ 63 /* shift up all the later entries in this row */ \ 64 for (ii=N; ii>=_i; ii--) { \ 65 rp[ii+1] = rp[ii]; \ 66 ierr = PetscMemcpy(ap+bs2*(ii+1),ap+bs2*(ii),bs2*sizeof(MatScalar));CHKERRQ(ierr); \ 67 } \ 68 if (N>=_i) { ierr = PetscMemzero(ap+bs2*_i,bs2*sizeof(MatScalar));CHKERRQ(ierr); } \ 69 rp[_i] = bcol; \ 70 ap[bs2*_i + bs*cidx + ridx] = value; \ 71 A->nonzerostate++;\ 72 a_noinsert:; \ 73 ailen[brow] = nrow; \ 74 } 75 76 #define MatSetValues_SeqSBAIJ_B_Private(row,col,value,addv,orow,ocol) \ 77 { \ 78 brow = row/bs; \ 79 rp = bj + bi[brow]; ap = ba + bs2*bi[brow]; \ 80 rmax = bimax[brow]; nrow = bilen[brow]; \ 81 bcol = col/bs; \ 82 ridx = row % bs; cidx = col % bs; \ 83 low = 0; high = nrow; \ 84 while (high-low > 3) { \ 85 t = (low+high)/2; \ 86 if (rp[t] > bcol) high = t; \ 87 else low = t; \ 88 } \ 89 for (_i=low; _i<high; _i++) { \ 90 if (rp[_i] > bcol) break; \ 91 if (rp[_i] == bcol) { \ 92 bap = ap + bs2*_i + bs*cidx + ridx; \ 93 if (addv == ADD_VALUES) *bap += value; \ 94 else *bap = value; \ 95 goto b_noinsert; \ 96 } \ 97 } \ 98 if (b->nonew == 1) goto b_noinsert; \ 99 if (b->nonew == -1) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero at global row/column (%D, %D) into matrix", orow, ocol); \ 100 MatSeqXAIJReallocateAIJ(B,b->mbs,bs2,nrow,brow,bcol,rmax,ba,bi,bj,rp,ap,bimax,b->nonew,MatScalar); \ 101 N = nrow++ - 1; \ 102 /* shift up all the later entries in this row */ \ 103 for (ii=N; ii>=_i; ii--) { \ 104 rp[ii+1] = rp[ii]; \ 105 ierr = PetscMemcpy(ap+bs2*(ii+1),ap+bs2*(ii),bs2*sizeof(MatScalar));CHKERRQ(ierr); \ 106 } \ 107 if (N>=_i) { ierr = PetscMemzero(ap+bs2*_i,bs2*sizeof(MatScalar));CHKERRQ(ierr);} \ 108 rp[_i] = bcol; \ 109 ap[bs2*_i + bs*cidx + ridx] = value; \ 110 B->nonzerostate++;\ 111 b_noinsert:; \ 112 bilen[brow] = nrow; \ 113 } 114 115 /* Only add/insert a(i,j) with i<=j (blocks). 116 Any a(i,j) with i>j input by user is ingored. 117 */ 118 #undef __FUNCT__ 119 #define __FUNCT__ "MatSetValues_MPISBAIJ" 120 PetscErrorCode MatSetValues_MPISBAIJ(Mat mat,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],const PetscScalar v[],InsertMode addv) 121 { 122 Mat_MPISBAIJ *baij = (Mat_MPISBAIJ*)mat->data; 123 MatScalar value; 124 PetscBool roworiented = baij->roworiented; 125 PetscErrorCode ierr; 126 PetscInt i,j,row,col; 127 PetscInt rstart_orig=mat->rmap->rstart; 128 PetscInt rend_orig =mat->rmap->rend,cstart_orig=mat->cmap->rstart; 129 PetscInt cend_orig =mat->cmap->rend,bs=mat->rmap->bs; 130 131 /* Some Variables required in the macro */ 132 Mat A = baij->A; 133 Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ*)(A)->data; 134 PetscInt *aimax=a->imax,*ai=a->i,*ailen=a->ilen,*aj=a->j; 135 MatScalar *aa =a->a; 136 137 Mat B = baij->B; 138 Mat_SeqBAIJ *b = (Mat_SeqBAIJ*)(B)->data; 139 PetscInt *bimax=b->imax,*bi=b->i,*bilen=b->ilen,*bj=b->j; 140 MatScalar *ba =b->a; 141 142 PetscInt *rp,ii,nrow,_i,rmax,N,brow,bcol; 143 PetscInt low,high,t,ridx,cidx,bs2=a->bs2; 144 MatScalar *ap,*bap; 145 146 /* for stash */ 147 PetscInt n_loc, *in_loc = NULL; 148 MatScalar *v_loc = NULL; 149 150 PetscFunctionBegin; 151 if (!baij->donotstash) { 152 if (n > baij->n_loc) { 153 ierr = PetscFree(baij->in_loc);CHKERRQ(ierr); 154 ierr = PetscFree(baij->v_loc);CHKERRQ(ierr); 155 ierr = PetscMalloc1(n,&baij->in_loc);CHKERRQ(ierr); 156 ierr = PetscMalloc1(n,&baij->v_loc);CHKERRQ(ierr); 157 158 baij->n_loc = n; 159 } 160 in_loc = baij->in_loc; 161 v_loc = baij->v_loc; 162 } 163 164 for (i=0; i<m; i++) { 165 if (im[i] < 0) continue; 166 #if defined(PETSC_USE_DEBUG) 167 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); 168 #endif 169 if (im[i] >= rstart_orig && im[i] < rend_orig) { /* this processor entry */ 170 row = im[i] - rstart_orig; /* local row index */ 171 for (j=0; j<n; j++) { 172 if (im[i]/bs > in[j]/bs) { 173 if (a->ignore_ltriangular) { 174 continue; /* ignore lower triangular blocks */ 175 } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_USER,"Lower triangular value cannot be set for sbaij format. Ignoring these values, run with -mat_ignore_lower_triangular or call MatSetOption(mat,MAT_IGNORE_LOWER_TRIANGULAR,PETSC_TRUE)"); 176 } 177 if (in[j] >= cstart_orig && in[j] < cend_orig) { /* diag entry (A) */ 178 col = in[j] - cstart_orig; /* local col index */ 179 brow = row/bs; bcol = col/bs; 180 if (brow > bcol) continue; /* ignore lower triangular blocks of A */ 181 if (roworiented) value = v[i*n+j]; 182 else value = v[i+j*m]; 183 MatSetValues_SeqSBAIJ_A_Private(row,col,value,addv,im[i],in[j]); 184 /* ierr = MatSetValues_SeqBAIJ(baij->A,1,&row,1,&col,&value,addv);CHKERRQ(ierr); */ 185 } else if (in[j] < 0) continue; 186 #if defined(PETSC_USE_DEBUG) 187 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); 188 #endif 189 else { /* off-diag entry (B) */ 190 if (mat->was_assembled) { 191 if (!baij->colmap) { 192 ierr = MatCreateColmap_MPIBAIJ_Private(mat);CHKERRQ(ierr); 193 } 194 #if defined(PETSC_USE_CTABLE) 195 ierr = PetscTableFind(baij->colmap,in[j]/bs + 1,&col);CHKERRQ(ierr); 196 col = col - 1; 197 #else 198 col = baij->colmap[in[j]/bs] - 1; 199 #endif 200 if (col < 0 && !((Mat_SeqSBAIJ*)(baij->A->data))->nonew) { 201 ierr = MatDisAssemble_MPISBAIJ(mat);CHKERRQ(ierr); 202 col = in[j]; 203 /* Reinitialize the variables required by MatSetValues_SeqBAIJ_B_Private() */ 204 B = baij->B; 205 b = (Mat_SeqBAIJ*)(B)->data; 206 bimax= b->imax;bi=b->i;bilen=b->ilen;bj=b->j; 207 ba = b->a; 208 } else col += in[j]%bs; 209 } else col = in[j]; 210 if (roworiented) value = v[i*n+j]; 211 else value = v[i+j*m]; 212 MatSetValues_SeqSBAIJ_B_Private(row,col,value,addv,im[i],in[j]); 213 /* ierr = MatSetValues_SeqBAIJ(baij->B,1,&row,1,&col,&value,addv);CHKERRQ(ierr); */ 214 } 215 } 216 } else { /* off processor entry */ 217 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]); 218 if (!baij->donotstash) { 219 mat->assembled = PETSC_FALSE; 220 n_loc = 0; 221 for (j=0; j<n; j++) { 222 if (im[i]/bs > in[j]/bs) continue; /* ignore lower triangular blocks */ 223 in_loc[n_loc] = in[j]; 224 if (roworiented) { 225 v_loc[n_loc] = v[i*n+j]; 226 } else { 227 v_loc[n_loc] = v[j*m+i]; 228 } 229 n_loc++; 230 } 231 ierr = MatStashValuesRow_Private(&mat->stash,im[i],n_loc,in_loc,v_loc,PETSC_FALSE);CHKERRQ(ierr); 232 } 233 } 234 } 235 PetscFunctionReturn(0); 236 } 237 238 #undef __FUNCT__ 239 #define __FUNCT__ "MatSetValuesBlocked_SeqSBAIJ_Inlined" 240 PETSC_STATIC_INLINE PetscErrorCode MatSetValuesBlocked_SeqSBAIJ_Inlined(Mat A,PetscInt row,PetscInt col,const PetscScalar v[],InsertMode is,PetscInt orow,PetscInt ocol) 241 { 242 Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ*)A->data; 243 PetscErrorCode ierr; 244 PetscInt *rp,low,high,t,ii,jj,nrow,i,rmax,N; 245 PetscInt *imax =a->imax,*ai=a->i,*ailen=a->ilen; 246 PetscInt *aj =a->j,nonew=a->nonew,bs2=a->bs2,bs=A->rmap->bs; 247 PetscBool roworiented=a->roworiented; 248 const PetscScalar *value = v; 249 MatScalar *ap,*aa = a->a,*bap; 250 251 PetscFunctionBegin; 252 if (col < row) { 253 if (a->ignore_ltriangular) PetscFunctionReturn(0); /* ignore lower triangular block */ 254 else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_USER,"Lower triangular value cannot be set for sbaij format. Ignoring these values, run with -mat_ignore_lower_triangular or call MatSetOption(mat,MAT_IGNORE_LOWER_TRIANGULAR,PETSC_TRUE)"); 255 } 256 rp = aj + ai[row]; 257 ap = aa + bs2*ai[row]; 258 rmax = imax[row]; 259 nrow = ailen[row]; 260 value = v; 261 low = 0; 262 high = nrow; 263 264 while (high-low > 7) { 265 t = (low+high)/2; 266 if (rp[t] > col) high = t; 267 else low = t; 268 } 269 for (i=low; i<high; i++) { 270 if (rp[i] > col) break; 271 if (rp[i] == col) { 272 bap = ap + bs2*i; 273 if (roworiented) { 274 if (is == ADD_VALUES) { 275 for (ii=0; ii<bs; ii++) { 276 for (jj=ii; jj<bs2; jj+=bs) { 277 bap[jj] += *value++; 278 } 279 } 280 } else { 281 for (ii=0; ii<bs; ii++) { 282 for (jj=ii; jj<bs2; jj+=bs) { 283 bap[jj] = *value++; 284 } 285 } 286 } 287 } else { 288 if (is == ADD_VALUES) { 289 for (ii=0; ii<bs; ii++) { 290 for (jj=0; jj<bs; jj++) { 291 *bap++ += *value++; 292 } 293 } 294 } else { 295 for (ii=0; ii<bs; ii++) { 296 for (jj=0; jj<bs; jj++) { 297 *bap++ = *value++; 298 } 299 } 300 } 301 } 302 goto noinsert2; 303 } 304 } 305 if (nonew == 1) goto noinsert2; 306 if (nonew == -1) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new block index nonzero block (%D, %D) in the matrix", orow, ocol); 307 MatSeqXAIJReallocateAIJ(A,a->mbs,bs2,nrow,row,col,rmax,aa,ai,aj,rp,ap,imax,nonew,MatScalar); 308 N = nrow++ - 1; high++; 309 /* shift up all the later entries in this row */ 310 for (ii=N; ii>=i; ii--) { 311 rp[ii+1] = rp[ii]; 312 ierr = PetscMemcpy(ap+bs2*(ii+1),ap+bs2*(ii),bs2*sizeof(MatScalar));CHKERRQ(ierr); 313 } 314 if (N >= i) { 315 ierr = PetscMemzero(ap+bs2*i,bs2*sizeof(MatScalar));CHKERRQ(ierr); 316 } 317 rp[i] = col; 318 bap = ap + bs2*i; 319 if (roworiented) { 320 for (ii=0; ii<bs; ii++) { 321 for (jj=ii; jj<bs2; jj+=bs) { 322 bap[jj] = *value++; 323 } 324 } 325 } else { 326 for (ii=0; ii<bs; ii++) { 327 for (jj=0; jj<bs; jj++) { 328 *bap++ = *value++; 329 } 330 } 331 } 332 noinsert2:; 333 ailen[row] = nrow; 334 PetscFunctionReturn(0); 335 } 336 337 #undef __FUNCT__ 338 #define __FUNCT__ "MatSetValuesBlocked_SeqBAIJ_Inlined" 339 /* 340 This routine is exactly duplicated in mpibaij.c 341 */ 342 PETSC_STATIC_INLINE PetscErrorCode MatSetValuesBlocked_SeqBAIJ_Inlined(Mat A,PetscInt row,PetscInt col,const PetscScalar v[],InsertMode is,PetscInt orow,PetscInt ocol) 343 { 344 Mat_SeqBAIJ *a = (Mat_SeqBAIJ*)A->data; 345 PetscInt *rp,low,high,t,ii,jj,nrow,i,rmax,N; 346 PetscInt *imax=a->imax,*ai=a->i,*ailen=a->ilen; 347 PetscErrorCode ierr; 348 PetscInt *aj =a->j,nonew=a->nonew,bs2=a->bs2,bs=A->rmap->bs; 349 PetscBool roworiented=a->roworiented; 350 const PetscScalar *value = v; 351 MatScalar *ap,*aa = a->a,*bap; 352 353 PetscFunctionBegin; 354 rp = aj + ai[row]; 355 ap = aa + bs2*ai[row]; 356 rmax = imax[row]; 357 nrow = ailen[row]; 358 low = 0; 359 high = nrow; 360 value = v; 361 while (high-low > 7) { 362 t = (low+high)/2; 363 if (rp[t] > col) high = t; 364 else low = t; 365 } 366 for (i=low; i<high; i++) { 367 if (rp[i] > col) break; 368 if (rp[i] == col) { 369 bap = ap + bs2*i; 370 if (roworiented) { 371 if (is == ADD_VALUES) { 372 for (ii=0; ii<bs; ii++) { 373 for (jj=ii; jj<bs2; jj+=bs) { 374 bap[jj] += *value++; 375 } 376 } 377 } else { 378 for (ii=0; ii<bs; ii++) { 379 for (jj=ii; jj<bs2; jj+=bs) { 380 bap[jj] = *value++; 381 } 382 } 383 } 384 } else { 385 if (is == ADD_VALUES) { 386 for (ii=0; ii<bs; ii++,value+=bs) { 387 for (jj=0; jj<bs; jj++) { 388 bap[jj] += value[jj]; 389 } 390 bap += bs; 391 } 392 } else { 393 for (ii=0; ii<bs; ii++,value+=bs) { 394 for (jj=0; jj<bs; jj++) { 395 bap[jj] = value[jj]; 396 } 397 bap += bs; 398 } 399 } 400 } 401 goto noinsert2; 402 } 403 } 404 if (nonew == 1) goto noinsert2; 405 if (nonew == -1) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new global block indexed nonzero block (%D, %D) in the matrix", orow, ocol); 406 MatSeqXAIJReallocateAIJ(A,a->mbs,bs2,nrow,row,col,rmax,aa,ai,aj,rp,ap,imax,nonew,MatScalar); 407 N = nrow++ - 1; high++; 408 /* shift up all the later entries in this row */ 409 for (ii=N; ii>=i; ii--) { 410 rp[ii+1] = rp[ii]; 411 ierr = PetscMemcpy(ap+bs2*(ii+1),ap+bs2*(ii),bs2*sizeof(MatScalar));CHKERRQ(ierr); 412 } 413 if (N >= i) { 414 ierr = PetscMemzero(ap+bs2*i,bs2*sizeof(MatScalar));CHKERRQ(ierr); 415 } 416 rp[i] = col; 417 bap = ap + bs2*i; 418 if (roworiented) { 419 for (ii=0; ii<bs; ii++) { 420 for (jj=ii; jj<bs2; jj+=bs) { 421 bap[jj] = *value++; 422 } 423 } 424 } else { 425 for (ii=0; ii<bs; ii++) { 426 for (jj=0; jj<bs; jj++) { 427 *bap++ = *value++; 428 } 429 } 430 } 431 noinsert2:; 432 ailen[row] = nrow; 433 PetscFunctionReturn(0); 434 } 435 436 #undef __FUNCT__ 437 #define __FUNCT__ "MatSetValuesBlocked_MPISBAIJ" 438 /* 439 This routine could be optimized by removing the need for the block copy below and passing stride information 440 to the above inline routines; similarly in MatSetValuesBlocked_MPIBAIJ() 441 */ 442 PetscErrorCode MatSetValuesBlocked_MPISBAIJ(Mat mat,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],const MatScalar v[],InsertMode addv) 443 { 444 Mat_MPISBAIJ *baij = (Mat_MPISBAIJ*)mat->data; 445 const MatScalar *value; 446 MatScalar *barray =baij->barray; 447 PetscBool roworiented = baij->roworiented,ignore_ltriangular = ((Mat_SeqSBAIJ*)baij->A->data)->ignore_ltriangular; 448 PetscErrorCode ierr; 449 PetscInt i,j,ii,jj,row,col,rstart=baij->rstartbs; 450 PetscInt rend=baij->rendbs,cstart=baij->rstartbs,stepval; 451 PetscInt cend=baij->rendbs,bs=mat->rmap->bs,bs2=baij->bs2; 452 453 PetscFunctionBegin; 454 if (!barray) { 455 ierr = PetscMalloc1(bs2,&barray);CHKERRQ(ierr); 456 baij->barray = barray; 457 } 458 459 if (roworiented) { 460 stepval = (n-1)*bs; 461 } else { 462 stepval = (m-1)*bs; 463 } 464 for (i=0; i<m; i++) { 465 if (im[i] < 0) continue; 466 #if defined(PETSC_USE_DEBUG) 467 if (im[i] >= baij->Mbs) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Block indexed row too large %D max %D",im[i],baij->Mbs-1); 468 #endif 469 if (im[i] >= rstart && im[i] < rend) { 470 row = im[i] - rstart; 471 for (j=0; j<n; j++) { 472 if (im[i] > in[j]) { 473 if (ignore_ltriangular) continue; /* ignore lower triangular blocks */ 474 else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_USER,"Lower triangular value cannot be set for sbaij format. Ignoring these values, run with -mat_ignore_lower_triangular or call MatSetOption(mat,MAT_IGNORE_LOWER_TRIANGULAR,PETSC_TRUE)"); 475 } 476 /* If NumCol = 1 then a copy is not required */ 477 if ((roworiented) && (n == 1)) { 478 barray = (MatScalar*) v + i*bs2; 479 } else if ((!roworiented) && (m == 1)) { 480 barray = (MatScalar*) v + j*bs2; 481 } else { /* Here a copy is required */ 482 if (roworiented) { 483 value = v + i*(stepval+bs)*bs + j*bs; 484 } else { 485 value = v + j*(stepval+bs)*bs + i*bs; 486 } 487 for (ii=0; ii<bs; ii++,value+=stepval) { 488 for (jj=0; jj<bs; jj++) { 489 *barray++ = *value++; 490 } 491 } 492 barray -=bs2; 493 } 494 495 if (in[j] >= cstart && in[j] < cend) { 496 col = in[j] - cstart; 497 ierr = MatSetValuesBlocked_SeqSBAIJ_Inlined(baij->A,row,col,barray,addv,im[i],in[j]);CHKERRQ(ierr); 498 } else if (in[j] < 0) continue; 499 #if defined(PETSC_USE_DEBUG) 500 else if (in[j] >= baij->Nbs) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Block indexed column too large %D max %D",in[j],baij->Nbs-1); 501 #endif 502 else { 503 if (mat->was_assembled) { 504 if (!baij->colmap) { 505 ierr = MatCreateColmap_MPIBAIJ_Private(mat);CHKERRQ(ierr); 506 } 507 508 #if defined(PETSC_USE_DEBUG) 509 #if defined(PETSC_USE_CTABLE) 510 { PetscInt data; 511 ierr = PetscTableFind(baij->colmap,in[j]+1,&data);CHKERRQ(ierr); 512 if ((data - 1) % bs) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Incorrect colmap"); 513 } 514 #else 515 if ((baij->colmap[in[j]] - 1) % bs) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Incorrect colmap"); 516 #endif 517 #endif 518 #if defined(PETSC_USE_CTABLE) 519 ierr = PetscTableFind(baij->colmap,in[j]+1,&col);CHKERRQ(ierr); 520 col = (col - 1)/bs; 521 #else 522 col = (baij->colmap[in[j]] - 1)/bs; 523 #endif 524 if (col < 0 && !((Mat_SeqBAIJ*)(baij->A->data))->nonew) { 525 ierr = MatDisAssemble_MPISBAIJ(mat);CHKERRQ(ierr); 526 col = in[j]; 527 } 528 } else col = in[j]; 529 ierr = MatSetValuesBlocked_SeqBAIJ_Inlined(baij->B,row,col,barray,addv,im[i],in[j]);CHKERRQ(ierr); 530 } 531 } 532 } else { 533 if (mat->nooffprocentries) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Setting off process block indexed row %D even though MatSetOption(,MAT_NO_OFF_PROC_ENTRIES,PETSC_TRUE) was set",im[i]); 534 if (!baij->donotstash) { 535 if (roworiented) { 536 ierr = MatStashValuesRowBlocked_Private(&mat->bstash,im[i],n,in,v,m,n,i);CHKERRQ(ierr); 537 } else { 538 ierr = MatStashValuesColBlocked_Private(&mat->bstash,im[i],n,in,v,m,n,i);CHKERRQ(ierr); 539 } 540 } 541 } 542 } 543 PetscFunctionReturn(0); 544 } 545 546 #undef __FUNCT__ 547 #define __FUNCT__ "MatGetValues_MPISBAIJ" 548 PetscErrorCode MatGetValues_MPISBAIJ(Mat mat,PetscInt m,const PetscInt idxm[],PetscInt n,const PetscInt idxn[],PetscScalar v[]) 549 { 550 Mat_MPISBAIJ *baij = (Mat_MPISBAIJ*)mat->data; 551 PetscErrorCode ierr; 552 PetscInt bs = mat->rmap->bs,i,j,bsrstart = mat->rmap->rstart,bsrend = mat->rmap->rend; 553 PetscInt bscstart = mat->cmap->rstart,bscend = mat->cmap->rend,row,col,data; 554 555 PetscFunctionBegin; 556 for (i=0; i<m; i++) { 557 if (idxm[i] < 0) continue; /* SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative row: %D",idxm[i]); */ 558 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); 559 if (idxm[i] >= bsrstart && idxm[i] < bsrend) { 560 row = idxm[i] - bsrstart; 561 for (j=0; j<n; j++) { 562 if (idxn[j] < 0) continue; /* SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative column %D",idxn[j]); */ 563 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); 564 if (idxn[j] >= bscstart && idxn[j] < bscend) { 565 col = idxn[j] - bscstart; 566 ierr = MatGetValues_SeqSBAIJ(baij->A,1,&row,1,&col,v+i*n+j);CHKERRQ(ierr); 567 } else { 568 if (!baij->colmap) { 569 ierr = MatCreateColmap_MPIBAIJ_Private(mat);CHKERRQ(ierr); 570 } 571 #if defined(PETSC_USE_CTABLE) 572 ierr = PetscTableFind(baij->colmap,idxn[j]/bs+1,&data);CHKERRQ(ierr); 573 data--; 574 #else 575 data = baij->colmap[idxn[j]/bs]-1; 576 #endif 577 if ((data < 0) || (baij->garray[data/bs] != idxn[j]/bs)) *(v+i*n+j) = 0.0; 578 else { 579 col = data + idxn[j]%bs; 580 ierr = MatGetValues_SeqBAIJ(baij->B,1,&row,1,&col,v+i*n+j);CHKERRQ(ierr); 581 } 582 } 583 } 584 } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Only local values currently supported"); 585 } 586 PetscFunctionReturn(0); 587 } 588 589 #undef __FUNCT__ 590 #define __FUNCT__ "MatNorm_MPISBAIJ" 591 PetscErrorCode MatNorm_MPISBAIJ(Mat mat,NormType type,PetscReal *norm) 592 { 593 Mat_MPISBAIJ *baij = (Mat_MPISBAIJ*)mat->data; 594 PetscErrorCode ierr; 595 PetscReal sum[2],*lnorm2; 596 597 PetscFunctionBegin; 598 if (baij->size == 1) { 599 ierr = MatNorm(baij->A,type,norm);CHKERRQ(ierr); 600 } else { 601 if (type == NORM_FROBENIUS) { 602 ierr = PetscMalloc1(2,&lnorm2);CHKERRQ(ierr); 603 ierr = MatNorm(baij->A,type,lnorm2);CHKERRQ(ierr); 604 *lnorm2 = (*lnorm2)*(*lnorm2); lnorm2++; /* squar power of norm(A) */ 605 ierr = MatNorm(baij->B,type,lnorm2);CHKERRQ(ierr); 606 *lnorm2 = (*lnorm2)*(*lnorm2); lnorm2--; /* squar power of norm(B) */ 607 ierr = MPI_Allreduce(lnorm2,sum,2,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 608 *norm = PetscSqrtReal(sum[0] + 2*sum[1]); 609 ierr = PetscFree(lnorm2);CHKERRQ(ierr); 610 } else if (type == NORM_INFINITY || type == NORM_1) { /* max row/column sum */ 611 Mat_SeqSBAIJ *amat=(Mat_SeqSBAIJ*)baij->A->data; 612 Mat_SeqBAIJ *bmat=(Mat_SeqBAIJ*)baij->B->data; 613 PetscReal *rsum,*rsum2,vabs; 614 PetscInt *jj,*garray=baij->garray,rstart=baij->rstartbs,nz; 615 PetscInt brow,bcol,col,bs=baij->A->rmap->bs,row,grow,gcol,mbs=amat->mbs; 616 MatScalar *v; 617 618 ierr = PetscMalloc2(mat->cmap->N,&rsum,mat->cmap->N,&rsum2);CHKERRQ(ierr); 619 ierr = PetscMemzero(rsum,mat->cmap->N*sizeof(PetscReal));CHKERRQ(ierr); 620 /* Amat */ 621 v = amat->a; jj = amat->j; 622 for (brow=0; brow<mbs; brow++) { 623 grow = bs*(rstart + brow); 624 nz = amat->i[brow+1] - amat->i[brow]; 625 for (bcol=0; bcol<nz; bcol++) { 626 gcol = bs*(rstart + *jj); jj++; 627 for (col=0; col<bs; col++) { 628 for (row=0; row<bs; row++) { 629 vabs = PetscAbsScalar(*v); v++; 630 rsum[gcol+col] += vabs; 631 /* non-diagonal block */ 632 if (bcol > 0 && vabs > 0.0) rsum[grow+row] += vabs; 633 } 634 } 635 } 636 } 637 /* Bmat */ 638 v = bmat->a; jj = bmat->j; 639 for (brow=0; brow<mbs; brow++) { 640 grow = bs*(rstart + brow); 641 nz = bmat->i[brow+1] - bmat->i[brow]; 642 for (bcol=0; bcol<nz; bcol++) { 643 gcol = bs*garray[*jj]; jj++; 644 for (col=0; col<bs; col++) { 645 for (row=0; row<bs; row++) { 646 vabs = PetscAbsScalar(*v); v++; 647 rsum[gcol+col] += vabs; 648 rsum[grow+row] += vabs; 649 } 650 } 651 } 652 } 653 ierr = MPI_Allreduce(rsum,rsum2,mat->cmap->N,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 654 *norm = 0.0; 655 for (col=0; col<mat->cmap->N; col++) { 656 if (rsum2[col] > *norm) *norm = rsum2[col]; 657 } 658 ierr = PetscFree2(rsum,rsum2);CHKERRQ(ierr); 659 } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"No support for this norm yet"); 660 } 661 PetscFunctionReturn(0); 662 } 663 664 #undef __FUNCT__ 665 #define __FUNCT__ "MatAssemblyBegin_MPISBAIJ" 666 PetscErrorCode MatAssemblyBegin_MPISBAIJ(Mat mat,MatAssemblyType mode) 667 { 668 Mat_MPISBAIJ *baij = (Mat_MPISBAIJ*)mat->data; 669 PetscErrorCode ierr; 670 PetscInt nstash,reallocs; 671 InsertMode addv; 672 673 PetscFunctionBegin; 674 if (baij->donotstash || mat->nooffprocentries) PetscFunctionReturn(0); 675 676 /* make sure all processors are either in INSERTMODE or ADDMODE */ 677 ierr = MPI_Allreduce((PetscEnum*)&mat->insertmode,(PetscEnum*)&addv,1,MPIU_ENUM,MPI_BOR,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 678 if (addv == (ADD_VALUES|INSERT_VALUES)) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_ARG_WRONGSTATE,"Some processors inserted others added"); 679 mat->insertmode = addv; /* in case this processor had no cache */ 680 681 ierr = MatStashScatterBegin_Private(mat,&mat->stash,mat->rmap->range);CHKERRQ(ierr); 682 ierr = MatStashScatterBegin_Private(mat,&mat->bstash,baij->rangebs);CHKERRQ(ierr); 683 ierr = MatStashGetInfo_Private(&mat->stash,&nstash,&reallocs);CHKERRQ(ierr); 684 ierr = PetscInfo2(mat,"Stash has %D entries,uses %D mallocs.\n",nstash,reallocs);CHKERRQ(ierr); 685 ierr = MatStashGetInfo_Private(&mat->stash,&nstash,&reallocs);CHKERRQ(ierr); 686 ierr = PetscInfo2(mat,"Block-Stash has %D entries, uses %D mallocs.\n",nstash,reallocs);CHKERRQ(ierr); 687 PetscFunctionReturn(0); 688 } 689 690 #undef __FUNCT__ 691 #define __FUNCT__ "MatAssemblyEnd_MPISBAIJ" 692 PetscErrorCode MatAssemblyEnd_MPISBAIJ(Mat mat,MatAssemblyType mode) 693 { 694 Mat_MPISBAIJ *baij=(Mat_MPISBAIJ*)mat->data; 695 Mat_SeqSBAIJ *a =(Mat_SeqSBAIJ*)baij->A->data; 696 PetscErrorCode ierr; 697 PetscInt i,j,rstart,ncols,flg,bs2=baij->bs2; 698 PetscInt *row,*col; 699 PetscBool other_disassembled; 700 PetscMPIInt n; 701 PetscBool r1,r2,r3; 702 MatScalar *val; 703 InsertMode addv = mat->insertmode; 704 705 /* do not use 'b=(Mat_SeqBAIJ*)baij->B->data' as B can be reset in disassembly */ 706 PetscFunctionBegin; 707 if (!baij->donotstash && !mat->nooffprocentries) { 708 while (1) { 709 ierr = MatStashScatterGetMesg_Private(&mat->stash,&n,&row,&col,&val,&flg);CHKERRQ(ierr); 710 if (!flg) break; 711 712 for (i=0; i<n;) { 713 /* Now identify the consecutive vals belonging to the same row */ 714 for (j=i,rstart=row[j]; j<n; j++) { 715 if (row[j] != rstart) break; 716 } 717 if (j < n) ncols = j-i; 718 else ncols = n-i; 719 /* Now assemble all these values with a single function call */ 720 ierr = MatSetValues_MPISBAIJ(mat,1,row+i,ncols,col+i,val+i,addv);CHKERRQ(ierr); 721 i = j; 722 } 723 } 724 ierr = MatStashScatterEnd_Private(&mat->stash);CHKERRQ(ierr); 725 /* Now process the block-stash. Since the values are stashed column-oriented, 726 set the roworiented flag to column oriented, and after MatSetValues() 727 restore the original flags */ 728 r1 = baij->roworiented; 729 r2 = a->roworiented; 730 r3 = ((Mat_SeqBAIJ*)baij->B->data)->roworiented; 731 732 baij->roworiented = PETSC_FALSE; 733 a->roworiented = PETSC_FALSE; 734 735 ((Mat_SeqBAIJ*)baij->B->data)->roworiented = PETSC_FALSE; /* b->roworinted */ 736 while (1) { 737 ierr = MatStashScatterGetMesg_Private(&mat->bstash,&n,&row,&col,&val,&flg);CHKERRQ(ierr); 738 if (!flg) break; 739 740 for (i=0; i<n;) { 741 /* Now identify the consecutive vals belonging to the same row */ 742 for (j=i,rstart=row[j]; j<n; j++) { 743 if (row[j] != rstart) break; 744 } 745 if (j < n) ncols = j-i; 746 else ncols = n-i; 747 ierr = MatSetValuesBlocked_MPISBAIJ(mat,1,row+i,ncols,col+i,val+i*bs2,addv);CHKERRQ(ierr); 748 i = j; 749 } 750 } 751 ierr = MatStashScatterEnd_Private(&mat->bstash);CHKERRQ(ierr); 752 753 baij->roworiented = r1; 754 a->roworiented = r2; 755 756 ((Mat_SeqBAIJ*)baij->B->data)->roworiented = r3; /* b->roworinted */ 757 } 758 759 ierr = MatAssemblyBegin(baij->A,mode);CHKERRQ(ierr); 760 ierr = MatAssemblyEnd(baij->A,mode);CHKERRQ(ierr); 761 762 /* determine if any processor has disassembled, if so we must 763 also disassemble ourselfs, in order that we may reassemble. */ 764 /* 765 if nonzero structure of submatrix B cannot change then we know that 766 no processor disassembled thus we can skip this stuff 767 */ 768 if (!((Mat_SeqBAIJ*)baij->B->data)->nonew) { 769 ierr = MPI_Allreduce(&mat->was_assembled,&other_disassembled,1,MPIU_BOOL,MPI_PROD,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 770 if (mat->was_assembled && !other_disassembled) { 771 ierr = MatDisAssemble_MPISBAIJ(mat);CHKERRQ(ierr); 772 } 773 } 774 775 if (!mat->was_assembled && mode == MAT_FINAL_ASSEMBLY) { 776 ierr = MatSetUpMultiply_MPISBAIJ(mat);CHKERRQ(ierr); /* setup Mvctx and sMvctx */ 777 } 778 ierr = MatAssemblyBegin(baij->B,mode);CHKERRQ(ierr); 779 ierr = MatAssemblyEnd(baij->B,mode);CHKERRQ(ierr); 780 781 ierr = PetscFree2(baij->rowvalues,baij->rowindices);CHKERRQ(ierr); 782 783 baij->rowvalues = 0; 784 785 /* if no new nonzero locations are allowed in matrix then only set the matrix state the first time through */ 786 if ((!mat->was_assembled && mode == MAT_FINAL_ASSEMBLY) || !((Mat_SeqBAIJ*)(baij->A->data))->nonew) { 787 PetscObjectState state = baij->A->nonzerostate + baij->B->nonzerostate; 788 ierr = MPI_Allreduce(&state,&mat->nonzerostate,1,MPIU_INT64,MPI_SUM,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 789 } 790 PetscFunctionReturn(0); 791 } 792 793 extern PetscErrorCode MatView_SeqSBAIJ_ASCII(Mat,PetscViewer); 794 extern PetscErrorCode MatSetValues_MPIBAIJ(Mat,PetscInt,const PetscInt[],PetscInt,const PetscInt[],const PetscScalar[],InsertMode); 795 #include <petscdraw.h> 796 #undef __FUNCT__ 797 #define __FUNCT__ "MatView_MPISBAIJ_ASCIIorDraworSocket" 798 static PetscErrorCode MatView_MPISBAIJ_ASCIIorDraworSocket(Mat mat,PetscViewer viewer) 799 { 800 Mat_MPISBAIJ *baij = (Mat_MPISBAIJ*)mat->data; 801 PetscErrorCode ierr; 802 PetscInt bs = mat->rmap->bs; 803 PetscMPIInt rank = baij->rank; 804 PetscBool iascii,isdraw; 805 PetscViewer sviewer; 806 PetscViewerFormat format; 807 808 PetscFunctionBegin; 809 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii);CHKERRQ(ierr); 810 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERDRAW,&isdraw);CHKERRQ(ierr); 811 if (iascii) { 812 ierr = PetscViewerGetFormat(viewer,&format);CHKERRQ(ierr); 813 if (format == PETSC_VIEWER_ASCII_INFO_DETAIL) { 814 MatInfo info; 815 ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr); 816 ierr = MatGetInfo(mat,MAT_LOCAL,&info);CHKERRQ(ierr); 817 ierr = PetscViewerASCIISynchronizedAllow(viewer,PETSC_TRUE);CHKERRQ(ierr); 818 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"[%d] Local rows %D nz %D nz alloced %D bs %D mem %D\n",rank,mat->rmap->n,(PetscInt)info.nz_used,(PetscInt)info.nz_allocated,mat->rmap->bs,(PetscInt)info.memory);CHKERRQ(ierr); 819 ierr = MatGetInfo(baij->A,MAT_LOCAL,&info);CHKERRQ(ierr); 820 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"[%d] on-diagonal part: nz %D \n",rank,(PetscInt)info.nz_used);CHKERRQ(ierr); 821 ierr = MatGetInfo(baij->B,MAT_LOCAL,&info);CHKERRQ(ierr); 822 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"[%d] off-diagonal part: nz %D \n",rank,(PetscInt)info.nz_used);CHKERRQ(ierr); 823 ierr = PetscViewerFlush(viewer);CHKERRQ(ierr); 824 ierr = PetscViewerASCIISynchronizedAllow(viewer,PETSC_FALSE);CHKERRQ(ierr); 825 ierr = PetscViewerASCIIPrintf(viewer,"Information on VecScatter used in matrix-vector product: \n");CHKERRQ(ierr); 826 ierr = VecScatterView(baij->Mvctx,viewer);CHKERRQ(ierr); 827 PetscFunctionReturn(0); 828 } else if (format == PETSC_VIEWER_ASCII_INFO) { 829 ierr = PetscViewerASCIIPrintf(viewer," block size is %D\n",bs);CHKERRQ(ierr); 830 PetscFunctionReturn(0); 831 } else if (format == PETSC_VIEWER_ASCII_FACTOR_INFO) { 832 PetscFunctionReturn(0); 833 } 834 } 835 836 if (isdraw) { 837 PetscDraw draw; 838 PetscBool isnull; 839 ierr = PetscViewerDrawGetDraw(viewer,0,&draw);CHKERRQ(ierr); 840 ierr = PetscDrawIsNull(draw,&isnull);CHKERRQ(ierr); if (isnull) PetscFunctionReturn(0); 841 } 842 843 { 844 /* assemble the entire matrix onto first processor. */ 845 Mat A; 846 Mat_SeqSBAIJ *Aloc; 847 Mat_SeqBAIJ *Bloc; 848 PetscInt M = mat->rmap->N,N = mat->cmap->N,*ai,*aj,col,i,j,k,*rvals,mbs = baij->mbs; 849 MatScalar *a; 850 const char *matname; 851 852 /* Should this be the same type as mat? */ 853 ierr = MatCreate(PetscObjectComm((PetscObject)mat),&A);CHKERRQ(ierr); 854 if (!rank) { 855 ierr = MatSetSizes(A,M,N,M,N);CHKERRQ(ierr); 856 } else { 857 ierr = MatSetSizes(A,0,0,M,N);CHKERRQ(ierr); 858 } 859 ierr = MatSetType(A,MATMPISBAIJ);CHKERRQ(ierr); 860 ierr = MatMPISBAIJSetPreallocation(A,mat->rmap->bs,0,NULL,0,NULL);CHKERRQ(ierr); 861 ierr = MatSetOption(A,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 862 ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)A);CHKERRQ(ierr); 863 864 /* copy over the A part */ 865 Aloc = (Mat_SeqSBAIJ*)baij->A->data; 866 ai = Aloc->i; aj = Aloc->j; a = Aloc->a; 867 ierr = PetscMalloc1(bs,&rvals);CHKERRQ(ierr); 868 869 for (i=0; i<mbs; i++) { 870 rvals[0] = bs*(baij->rstartbs + i); 871 for (j=1; j<bs; j++) rvals[j] = rvals[j-1] + 1; 872 for (j=ai[i]; j<ai[i+1]; j++) { 873 col = (baij->cstartbs+aj[j])*bs; 874 for (k=0; k<bs; k++) { 875 ierr = MatSetValues_MPISBAIJ(A,bs,rvals,1,&col,a,INSERT_VALUES);CHKERRQ(ierr); 876 col++; 877 a += bs; 878 } 879 } 880 } 881 /* copy over the B part */ 882 Bloc = (Mat_SeqBAIJ*)baij->B->data; 883 ai = Bloc->i; aj = Bloc->j; a = Bloc->a; 884 for (i=0; i<mbs; i++) { 885 886 rvals[0] = bs*(baij->rstartbs + i); 887 for (j=1; j<bs; j++) rvals[j] = rvals[j-1] + 1; 888 for (j=ai[i]; j<ai[i+1]; j++) { 889 col = baij->garray[aj[j]]*bs; 890 for (k=0; k<bs; k++) { 891 ierr = MatSetValues_MPIBAIJ(A,bs,rvals,1,&col,a,INSERT_VALUES);CHKERRQ(ierr); 892 col++; 893 a += bs; 894 } 895 } 896 } 897 ierr = PetscFree(rvals);CHKERRQ(ierr); 898 ierr = MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 899 ierr = MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 900 /* 901 Everyone has to call to draw the matrix since the graphics waits are 902 synchronized across all processors that share the PetscDraw object 903 */ 904 ierr = PetscViewerGetSingleton(viewer,&sviewer);CHKERRQ(ierr); 905 ierr = PetscObjectGetName((PetscObject)mat,&matname);CHKERRQ(ierr); 906 if (!rank) { 907 ierr = PetscObjectSetName((PetscObject)((Mat_MPISBAIJ*)(A->data))->A,matname);CHKERRQ(ierr); 908 ierr = MatView_SeqSBAIJ_ASCII(((Mat_MPISBAIJ*)(A->data))->A,sviewer);CHKERRQ(ierr); 909 } 910 ierr = PetscViewerRestoreSingleton(viewer,&sviewer);CHKERRQ(ierr); 911 ierr = MatDestroy(&A);CHKERRQ(ierr); 912 } 913 PetscFunctionReturn(0); 914 } 915 916 #undef __FUNCT__ 917 #define __FUNCT__ "MatView_MPISBAIJ" 918 PetscErrorCode MatView_MPISBAIJ(Mat mat,PetscViewer viewer) 919 { 920 PetscErrorCode ierr; 921 PetscBool iascii,isdraw,issocket,isbinary; 922 923 PetscFunctionBegin; 924 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii);CHKERRQ(ierr); 925 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERDRAW,&isdraw);CHKERRQ(ierr); 926 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERSOCKET,&issocket);CHKERRQ(ierr); 927 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERBINARY,&isbinary);CHKERRQ(ierr); 928 if (iascii || isdraw || issocket || isbinary) { 929 ierr = MatView_MPISBAIJ_ASCIIorDraworSocket(mat,viewer);CHKERRQ(ierr); 930 } 931 PetscFunctionReturn(0); 932 } 933 934 #undef __FUNCT__ 935 #define __FUNCT__ "MatDestroy_MPISBAIJ" 936 PetscErrorCode MatDestroy_MPISBAIJ(Mat mat) 937 { 938 Mat_MPISBAIJ *baij = (Mat_MPISBAIJ*)mat->data; 939 PetscErrorCode ierr; 940 941 PetscFunctionBegin; 942 #if defined(PETSC_USE_LOG) 943 PetscLogObjectState((PetscObject)mat,"Rows=%D,Cols=%D",mat->rmap->N,mat->cmap->N); 944 #endif 945 ierr = MatStashDestroy_Private(&mat->stash);CHKERRQ(ierr); 946 ierr = MatStashDestroy_Private(&mat->bstash);CHKERRQ(ierr); 947 ierr = MatDestroy(&baij->A);CHKERRQ(ierr); 948 ierr = MatDestroy(&baij->B);CHKERRQ(ierr); 949 #if defined(PETSC_USE_CTABLE) 950 ierr = PetscTableDestroy(&baij->colmap);CHKERRQ(ierr); 951 #else 952 ierr = PetscFree(baij->colmap);CHKERRQ(ierr); 953 #endif 954 ierr = PetscFree(baij->garray);CHKERRQ(ierr); 955 ierr = VecDestroy(&baij->lvec);CHKERRQ(ierr); 956 ierr = VecScatterDestroy(&baij->Mvctx);CHKERRQ(ierr); 957 ierr = VecDestroy(&baij->slvec0);CHKERRQ(ierr); 958 ierr = VecDestroy(&baij->slvec0b);CHKERRQ(ierr); 959 ierr = VecDestroy(&baij->slvec1);CHKERRQ(ierr); 960 ierr = VecDestroy(&baij->slvec1a);CHKERRQ(ierr); 961 ierr = VecDestroy(&baij->slvec1b);CHKERRQ(ierr); 962 ierr = VecScatterDestroy(&baij->sMvctx);CHKERRQ(ierr); 963 ierr = PetscFree2(baij->rowvalues,baij->rowindices);CHKERRQ(ierr); 964 ierr = PetscFree(baij->barray);CHKERRQ(ierr); 965 ierr = PetscFree(baij->hd);CHKERRQ(ierr); 966 ierr = VecDestroy(&baij->diag);CHKERRQ(ierr); 967 ierr = VecDestroy(&baij->bb1);CHKERRQ(ierr); 968 ierr = VecDestroy(&baij->xx1);CHKERRQ(ierr); 969 #if defined(PETSC_USE_REAL_MAT_SINGLE) 970 ierr = PetscFree(baij->setvaluescopy);CHKERRQ(ierr); 971 #endif 972 ierr = PetscFree(baij->in_loc);CHKERRQ(ierr); 973 ierr = PetscFree(baij->v_loc);CHKERRQ(ierr); 974 ierr = PetscFree(baij->rangebs);CHKERRQ(ierr); 975 ierr = PetscFree(mat->data);CHKERRQ(ierr); 976 977 ierr = PetscObjectChangeTypeName((PetscObject)mat,0);CHKERRQ(ierr); 978 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatStoreValues_C",NULL);CHKERRQ(ierr); 979 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatRetrieveValues_C",NULL);CHKERRQ(ierr); 980 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatGetDiagonalBlock_C",NULL);CHKERRQ(ierr); 981 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatMPISBAIJSetPreallocation_C",NULL);CHKERRQ(ierr); 982 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatConvert_mpisbaij_mpisbstrm_C",NULL);CHKERRQ(ierr); 983 #if defined(PETSC_HAVE_ELEMENTAL) 984 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatConvert_mpisbaij_elemental_C",NULL);CHKERRQ(ierr); 985 #endif 986 PetscFunctionReturn(0); 987 } 988 989 #undef __FUNCT__ 990 #define __FUNCT__ "MatMult_MPISBAIJ_Hermitian" 991 PetscErrorCode MatMult_MPISBAIJ_Hermitian(Mat A,Vec xx,Vec yy) 992 { 993 Mat_MPISBAIJ *a = (Mat_MPISBAIJ*)A->data; 994 PetscErrorCode ierr; 995 PetscInt nt,mbs=a->mbs,bs=A->rmap->bs; 996 PetscScalar *x,*from; 997 998 PetscFunctionBegin; 999 ierr = VecGetLocalSize(xx,&nt);CHKERRQ(ierr); 1000 if (nt != A->cmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Incompatible partition of A and xx"); 1001 1002 /* diagonal part */ 1003 ierr = (*a->A->ops->mult)(a->A,xx,a->slvec1a);CHKERRQ(ierr); 1004 ierr = VecSet(a->slvec1b,0.0);CHKERRQ(ierr); 1005 1006 /* subdiagonal part */ 1007 ierr = (*a->B->ops->multhermitiantranspose)(a->B,xx,a->slvec0b);CHKERRQ(ierr); 1008 1009 /* copy x into the vec slvec0 */ 1010 ierr = VecGetArray(a->slvec0,&from);CHKERRQ(ierr); 1011 ierr = VecGetArray(xx,&x);CHKERRQ(ierr); 1012 1013 ierr = PetscMemcpy(from,x,bs*mbs*sizeof(MatScalar));CHKERRQ(ierr); 1014 ierr = VecRestoreArray(a->slvec0,&from);CHKERRQ(ierr); 1015 ierr = VecRestoreArray(xx,&x);CHKERRQ(ierr); 1016 1017 ierr = VecScatterBegin(a->sMvctx,a->slvec0,a->slvec1,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1018 ierr = VecScatterEnd(a->sMvctx,a->slvec0,a->slvec1,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1019 /* supperdiagonal part */ 1020 ierr = (*a->B->ops->multadd)(a->B,a->slvec1b,a->slvec1a,yy);CHKERRQ(ierr); 1021 PetscFunctionReturn(0); 1022 } 1023 1024 #undef __FUNCT__ 1025 #define __FUNCT__ "MatMult_MPISBAIJ" 1026 PetscErrorCode MatMult_MPISBAIJ(Mat A,Vec xx,Vec yy) 1027 { 1028 Mat_MPISBAIJ *a = (Mat_MPISBAIJ*)A->data; 1029 PetscErrorCode ierr; 1030 PetscInt nt,mbs=a->mbs,bs=A->rmap->bs; 1031 PetscScalar *from; 1032 const PetscScalar *x; 1033 1034 PetscFunctionBegin; 1035 ierr = VecGetLocalSize(xx,&nt);CHKERRQ(ierr); 1036 if (nt != A->cmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Incompatible partition of A and xx"); 1037 1038 /* diagonal part */ 1039 ierr = (*a->A->ops->mult)(a->A,xx,a->slvec1a);CHKERRQ(ierr); 1040 ierr = VecSet(a->slvec1b,0.0);CHKERRQ(ierr); 1041 1042 /* subdiagonal part */ 1043 ierr = (*a->B->ops->multtranspose)(a->B,xx,a->slvec0b);CHKERRQ(ierr); 1044 1045 /* copy x into the vec slvec0 */ 1046 ierr = VecGetArray(a->slvec0,&from);CHKERRQ(ierr); 1047 ierr = VecGetArrayRead(xx,&x);CHKERRQ(ierr); 1048 1049 ierr = PetscMemcpy(from,x,bs*mbs*sizeof(MatScalar));CHKERRQ(ierr); 1050 ierr = VecRestoreArray(a->slvec0,&from);CHKERRQ(ierr); 1051 ierr = VecRestoreArrayRead(xx,&x);CHKERRQ(ierr); 1052 1053 ierr = VecScatterBegin(a->sMvctx,a->slvec0,a->slvec1,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1054 ierr = VecScatterEnd(a->sMvctx,a->slvec0,a->slvec1,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1055 /* supperdiagonal part */ 1056 ierr = (*a->B->ops->multadd)(a->B,a->slvec1b,a->slvec1a,yy);CHKERRQ(ierr); 1057 PetscFunctionReturn(0); 1058 } 1059 1060 #undef __FUNCT__ 1061 #define __FUNCT__ "MatMult_MPISBAIJ_2comm" 1062 PetscErrorCode MatMult_MPISBAIJ_2comm(Mat A,Vec xx,Vec yy) 1063 { 1064 Mat_MPISBAIJ *a = (Mat_MPISBAIJ*)A->data; 1065 PetscErrorCode ierr; 1066 PetscInt nt; 1067 1068 PetscFunctionBegin; 1069 ierr = VecGetLocalSize(xx,&nt);CHKERRQ(ierr); 1070 if (nt != A->cmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Incompatible partition of A and xx"); 1071 1072 ierr = VecGetLocalSize(yy,&nt);CHKERRQ(ierr); 1073 if (nt != A->rmap->N) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Incompatible parition of A and yy"); 1074 1075 ierr = VecScatterBegin(a->Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1076 /* do diagonal part */ 1077 ierr = (*a->A->ops->mult)(a->A,xx,yy);CHKERRQ(ierr); 1078 /* do supperdiagonal part */ 1079 ierr = VecScatterEnd(a->Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1080 ierr = (*a->B->ops->multadd)(a->B,a->lvec,yy,yy);CHKERRQ(ierr); 1081 /* do subdiagonal part */ 1082 ierr = (*a->B->ops->multtranspose)(a->B,xx,a->lvec);CHKERRQ(ierr); 1083 ierr = VecScatterBegin(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1084 ierr = VecScatterEnd(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1085 PetscFunctionReturn(0); 1086 } 1087 1088 #undef __FUNCT__ 1089 #define __FUNCT__ "MatMultAdd_MPISBAIJ" 1090 PetscErrorCode MatMultAdd_MPISBAIJ(Mat A,Vec xx,Vec yy,Vec zz) 1091 { 1092 Mat_MPISBAIJ *a = (Mat_MPISBAIJ*)A->data; 1093 PetscErrorCode ierr; 1094 PetscInt mbs=a->mbs,bs=A->rmap->bs; 1095 PetscScalar *from,zero=0.0; 1096 const PetscScalar *x; 1097 1098 PetscFunctionBegin; 1099 /* 1100 PetscSynchronizedPrintf(PetscObjectComm((PetscObject)A)," MatMultAdd is called ...\n"); 1101 PetscSynchronizedFlush(PetscObjectComm((PetscObject)A),PETSC_STDOUT); 1102 */ 1103 /* diagonal part */ 1104 ierr = (*a->A->ops->multadd)(a->A,xx,yy,a->slvec1a);CHKERRQ(ierr); 1105 ierr = VecSet(a->slvec1b,zero);CHKERRQ(ierr); 1106 1107 /* subdiagonal part */ 1108 ierr = (*a->B->ops->multtranspose)(a->B,xx,a->slvec0b);CHKERRQ(ierr); 1109 1110 /* copy x into the vec slvec0 */ 1111 ierr = VecGetArray(a->slvec0,&from);CHKERRQ(ierr); 1112 ierr = VecGetArrayRead(xx,&x);CHKERRQ(ierr); 1113 ierr = PetscMemcpy(from,x,bs*mbs*sizeof(MatScalar));CHKERRQ(ierr); 1114 ierr = VecRestoreArray(a->slvec0,&from);CHKERRQ(ierr); 1115 1116 ierr = VecScatterBegin(a->sMvctx,a->slvec0,a->slvec1,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1117 ierr = VecRestoreArrayRead(xx,&x);CHKERRQ(ierr); 1118 ierr = VecScatterEnd(a->sMvctx,a->slvec0,a->slvec1,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1119 1120 /* supperdiagonal part */ 1121 ierr = (*a->B->ops->multadd)(a->B,a->slvec1b,a->slvec1a,zz);CHKERRQ(ierr); 1122 PetscFunctionReturn(0); 1123 } 1124 1125 #undef __FUNCT__ 1126 #define __FUNCT__ "MatMultAdd_MPISBAIJ_2comm" 1127 PetscErrorCode MatMultAdd_MPISBAIJ_2comm(Mat A,Vec xx,Vec yy,Vec zz) 1128 { 1129 Mat_MPISBAIJ *a = (Mat_MPISBAIJ*)A->data; 1130 PetscErrorCode ierr; 1131 1132 PetscFunctionBegin; 1133 ierr = VecScatterBegin(a->Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1134 /* do diagonal part */ 1135 ierr = (*a->A->ops->multadd)(a->A,xx,yy,zz);CHKERRQ(ierr); 1136 /* do supperdiagonal part */ 1137 ierr = VecScatterEnd(a->Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1138 ierr = (*a->B->ops->multadd)(a->B,a->lvec,zz,zz);CHKERRQ(ierr); 1139 1140 /* do subdiagonal part */ 1141 ierr = (*a->B->ops->multtranspose)(a->B,xx,a->lvec);CHKERRQ(ierr); 1142 ierr = VecScatterBegin(a->Mvctx,a->lvec,zz,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1143 ierr = VecScatterEnd(a->Mvctx,a->lvec,zz,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1144 PetscFunctionReturn(0); 1145 } 1146 1147 /* 1148 This only works correctly for square matrices where the subblock A->A is the 1149 diagonal block 1150 */ 1151 #undef __FUNCT__ 1152 #define __FUNCT__ "MatGetDiagonal_MPISBAIJ" 1153 PetscErrorCode MatGetDiagonal_MPISBAIJ(Mat A,Vec v) 1154 { 1155 Mat_MPISBAIJ *a = (Mat_MPISBAIJ*)A->data; 1156 PetscErrorCode ierr; 1157 1158 PetscFunctionBegin; 1159 /* if (a->rmap->N != a->cmap->N) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Supports only square matrix where A->A is diag block"); */ 1160 ierr = MatGetDiagonal(a->A,v);CHKERRQ(ierr); 1161 PetscFunctionReturn(0); 1162 } 1163 1164 #undef __FUNCT__ 1165 #define __FUNCT__ "MatScale_MPISBAIJ" 1166 PetscErrorCode MatScale_MPISBAIJ(Mat A,PetscScalar aa) 1167 { 1168 Mat_MPISBAIJ *a = (Mat_MPISBAIJ*)A->data; 1169 PetscErrorCode ierr; 1170 1171 PetscFunctionBegin; 1172 ierr = MatScale(a->A,aa);CHKERRQ(ierr); 1173 ierr = MatScale(a->B,aa);CHKERRQ(ierr); 1174 PetscFunctionReturn(0); 1175 } 1176 1177 #undef __FUNCT__ 1178 #define __FUNCT__ "MatGetRow_MPISBAIJ" 1179 PetscErrorCode MatGetRow_MPISBAIJ(Mat matin,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v) 1180 { 1181 Mat_MPISBAIJ *mat = (Mat_MPISBAIJ*)matin->data; 1182 PetscScalar *vworkA,*vworkB,**pvA,**pvB,*v_p; 1183 PetscErrorCode ierr; 1184 PetscInt bs = matin->rmap->bs,bs2 = mat->bs2,i,*cworkA,*cworkB,**pcA,**pcB; 1185 PetscInt nztot,nzA,nzB,lrow,brstart = matin->rmap->rstart,brend = matin->rmap->rend; 1186 PetscInt *cmap,*idx_p,cstart = mat->rstartbs; 1187 1188 PetscFunctionBegin; 1189 if (mat->getrowactive) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Already active"); 1190 mat->getrowactive = PETSC_TRUE; 1191 1192 if (!mat->rowvalues && (idx || v)) { 1193 /* 1194 allocate enough space to hold information from the longest row. 1195 */ 1196 Mat_SeqSBAIJ *Aa = (Mat_SeqSBAIJ*)mat->A->data; 1197 Mat_SeqBAIJ *Ba = (Mat_SeqBAIJ*)mat->B->data; 1198 PetscInt max = 1,mbs = mat->mbs,tmp; 1199 for (i=0; i<mbs; i++) { 1200 tmp = Aa->i[i+1] - Aa->i[i] + Ba->i[i+1] - Ba->i[i]; /* row length */ 1201 if (max < tmp) max = tmp; 1202 } 1203 ierr = PetscMalloc2(max*bs2,&mat->rowvalues,max*bs2,&mat->rowindices);CHKERRQ(ierr); 1204 } 1205 1206 if (row < brstart || row >= brend) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Only local rows"); 1207 lrow = row - brstart; /* local row index */ 1208 1209 pvA = &vworkA; pcA = &cworkA; pvB = &vworkB; pcB = &cworkB; 1210 if (!v) {pvA = 0; pvB = 0;} 1211 if (!idx) {pcA = 0; if (!v) pcB = 0;} 1212 ierr = (*mat->A->ops->getrow)(mat->A,lrow,&nzA,pcA,pvA);CHKERRQ(ierr); 1213 ierr = (*mat->B->ops->getrow)(mat->B,lrow,&nzB,pcB,pvB);CHKERRQ(ierr); 1214 nztot = nzA + nzB; 1215 1216 cmap = mat->garray; 1217 if (v || idx) { 1218 if (nztot) { 1219 /* Sort by increasing column numbers, assuming A and B already sorted */ 1220 PetscInt imark = -1; 1221 if (v) { 1222 *v = v_p = mat->rowvalues; 1223 for (i=0; i<nzB; i++) { 1224 if (cmap[cworkB[i]/bs] < cstart) v_p[i] = vworkB[i]; 1225 else break; 1226 } 1227 imark = i; 1228 for (i=0; i<nzA; i++) v_p[imark+i] = vworkA[i]; 1229 for (i=imark; i<nzB; i++) v_p[nzA+i] = vworkB[i]; 1230 } 1231 if (idx) { 1232 *idx = idx_p = mat->rowindices; 1233 if (imark > -1) { 1234 for (i=0; i<imark; i++) { 1235 idx_p[i] = cmap[cworkB[i]/bs]*bs + cworkB[i]%bs; 1236 } 1237 } else { 1238 for (i=0; i<nzB; i++) { 1239 if (cmap[cworkB[i]/bs] < cstart) idx_p[i] = cmap[cworkB[i]/bs]*bs + cworkB[i]%bs; 1240 else break; 1241 } 1242 imark = i; 1243 } 1244 for (i=0; i<nzA; i++) idx_p[imark+i] = cstart*bs + cworkA[i]; 1245 for (i=imark; i<nzB; i++) idx_p[nzA+i] = cmap[cworkB[i]/bs]*bs + cworkB[i]%bs ; 1246 } 1247 } else { 1248 if (idx) *idx = 0; 1249 if (v) *v = 0; 1250 } 1251 } 1252 *nz = nztot; 1253 ierr = (*mat->A->ops->restorerow)(mat->A,lrow,&nzA,pcA,pvA);CHKERRQ(ierr); 1254 ierr = (*mat->B->ops->restorerow)(mat->B,lrow,&nzB,pcB,pvB);CHKERRQ(ierr); 1255 PetscFunctionReturn(0); 1256 } 1257 1258 #undef __FUNCT__ 1259 #define __FUNCT__ "MatRestoreRow_MPISBAIJ" 1260 PetscErrorCode MatRestoreRow_MPISBAIJ(Mat mat,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v) 1261 { 1262 Mat_MPISBAIJ *baij = (Mat_MPISBAIJ*)mat->data; 1263 1264 PetscFunctionBegin; 1265 if (!baij->getrowactive) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"MatGetRow() must be called first"); 1266 baij->getrowactive = PETSC_FALSE; 1267 PetscFunctionReturn(0); 1268 } 1269 1270 #undef __FUNCT__ 1271 #define __FUNCT__ "MatGetRowUpperTriangular_MPISBAIJ" 1272 PetscErrorCode MatGetRowUpperTriangular_MPISBAIJ(Mat A) 1273 { 1274 Mat_MPISBAIJ *a = (Mat_MPISBAIJ*)A->data; 1275 Mat_SeqSBAIJ *aA = (Mat_SeqSBAIJ*)a->A->data; 1276 1277 PetscFunctionBegin; 1278 aA->getrow_utriangular = PETSC_TRUE; 1279 PetscFunctionReturn(0); 1280 } 1281 #undef __FUNCT__ 1282 #define __FUNCT__ "MatRestoreRowUpperTriangular_MPISBAIJ" 1283 PetscErrorCode MatRestoreRowUpperTriangular_MPISBAIJ(Mat A) 1284 { 1285 Mat_MPISBAIJ *a = (Mat_MPISBAIJ*)A->data; 1286 Mat_SeqSBAIJ *aA = (Mat_SeqSBAIJ*)a->A->data; 1287 1288 PetscFunctionBegin; 1289 aA->getrow_utriangular = PETSC_FALSE; 1290 PetscFunctionReturn(0); 1291 } 1292 1293 #undef __FUNCT__ 1294 #define __FUNCT__ "MatRealPart_MPISBAIJ" 1295 PetscErrorCode MatRealPart_MPISBAIJ(Mat A) 1296 { 1297 Mat_MPISBAIJ *a = (Mat_MPISBAIJ*)A->data; 1298 PetscErrorCode ierr; 1299 1300 PetscFunctionBegin; 1301 ierr = MatRealPart(a->A);CHKERRQ(ierr); 1302 ierr = MatRealPart(a->B);CHKERRQ(ierr); 1303 PetscFunctionReturn(0); 1304 } 1305 1306 #undef __FUNCT__ 1307 #define __FUNCT__ "MatImaginaryPart_MPISBAIJ" 1308 PetscErrorCode MatImaginaryPart_MPISBAIJ(Mat A) 1309 { 1310 Mat_MPISBAIJ *a = (Mat_MPISBAIJ*)A->data; 1311 PetscErrorCode ierr; 1312 1313 PetscFunctionBegin; 1314 ierr = MatImaginaryPart(a->A);CHKERRQ(ierr); 1315 ierr = MatImaginaryPart(a->B);CHKERRQ(ierr); 1316 PetscFunctionReturn(0); 1317 } 1318 1319 /* Check if isrow is a subset of iscol_local, called by MatGetSubMatrix_MPISBAIJ() 1320 Input: isrow - distributed(parallel), 1321 iscol_local - locally owned (seq) 1322 */ 1323 #undef __FUNCT__ 1324 #define __FUNCT__ "ISEqual_private" 1325 PetscErrorCode ISEqual_private(IS isrow,IS iscol_local,PetscBool *flg) 1326 { 1327 PetscErrorCode ierr; 1328 PetscInt sz1,sz2,*a1,*a2,i,j,k,nmatch; 1329 const PetscInt *ptr1,*ptr2; 1330 1331 PetscFunctionBegin; 1332 ierr = ISGetLocalSize(isrow,&sz1);CHKERRQ(ierr); 1333 ierr = ISGetLocalSize(iscol_local,&sz2);CHKERRQ(ierr); 1334 if (sz1 > sz2) { 1335 *flg = PETSC_FALSE; 1336 PetscFunctionReturn(0); 1337 } 1338 1339 ierr = ISGetIndices(isrow,&ptr1);CHKERRQ(ierr); 1340 ierr = ISGetIndices(iscol_local,&ptr2);CHKERRQ(ierr); 1341 1342 ierr = PetscMalloc1(sz1,&a1);CHKERRQ(ierr); 1343 ierr = PetscMalloc1(sz2,&a2);CHKERRQ(ierr); 1344 ierr = PetscMemcpy(a1,ptr1,sz1*sizeof(PetscInt));CHKERRQ(ierr); 1345 ierr = PetscMemcpy(a2,ptr2,sz2*sizeof(PetscInt));CHKERRQ(ierr); 1346 ierr = PetscSortInt(sz1,a1);CHKERRQ(ierr); 1347 ierr = PetscSortInt(sz2,a2);CHKERRQ(ierr); 1348 1349 nmatch=0; 1350 k = 0; 1351 for (i=0; i<sz1; i++){ 1352 for (j=k; j<sz2; j++){ 1353 if (a1[i] == a2[j]) { 1354 k = j; nmatch++; 1355 break; 1356 } 1357 } 1358 } 1359 ierr = ISRestoreIndices(isrow,&ptr1);CHKERRQ(ierr); 1360 ierr = ISRestoreIndices(iscol_local,&ptr2);CHKERRQ(ierr); 1361 ierr = PetscFree(a1);CHKERRQ(ierr); 1362 ierr = PetscFree(a2);CHKERRQ(ierr); 1363 if (nmatch < sz1) { 1364 *flg = PETSC_FALSE; 1365 } else { 1366 *flg = PETSC_TRUE; 1367 } 1368 PetscFunctionReturn(0); 1369 } 1370 1371 #undef __FUNCT__ 1372 #define __FUNCT__ "MatGetSubMatrix_MPISBAIJ" 1373 PetscErrorCode MatGetSubMatrix_MPISBAIJ(Mat mat,IS isrow,IS iscol,MatReuse call,Mat *newmat) 1374 { 1375 PetscErrorCode ierr; 1376 IS iscol_local; 1377 PetscInt csize; 1378 PetscBool isequal; 1379 1380 PetscFunctionBegin; 1381 ierr = ISGetLocalSize(iscol,&csize);CHKERRQ(ierr); 1382 if (call == MAT_REUSE_MATRIX) { 1383 ierr = PetscObjectQuery((PetscObject)*newmat,"ISAllGather",(PetscObject*)&iscol_local);CHKERRQ(ierr); 1384 if (!iscol_local) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Submatrix passed in was not used before, cannot reuse"); 1385 } else { 1386 ierr = ISAllGather(iscol,&iscol_local);CHKERRQ(ierr); 1387 ierr = ISEqual_private(isrow,iscol_local,&isequal);CHKERRQ(ierr); 1388 if (!isequal) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_INCOMP,"For symmetric format, iscol must equal isrow"); 1389 } 1390 1391 /* now call MatGetSubMatrix_MPIBAIJ() */ 1392 ierr = MatGetSubMatrix_MPIBAIJ_Private(mat,isrow,iscol_local,csize,call,newmat);CHKERRQ(ierr); 1393 if (call == MAT_INITIAL_MATRIX) { 1394 ierr = PetscObjectCompose((PetscObject)*newmat,"ISAllGather",(PetscObject)iscol_local);CHKERRQ(ierr); 1395 ierr = ISDestroy(&iscol_local);CHKERRQ(ierr); 1396 } 1397 PetscFunctionReturn(0); 1398 } 1399 1400 #undef __FUNCT__ 1401 #define __FUNCT__ "MatZeroEntries_MPISBAIJ" 1402 PetscErrorCode MatZeroEntries_MPISBAIJ(Mat A) 1403 { 1404 Mat_MPISBAIJ *l = (Mat_MPISBAIJ*)A->data; 1405 PetscErrorCode ierr; 1406 1407 PetscFunctionBegin; 1408 ierr = MatZeroEntries(l->A);CHKERRQ(ierr); 1409 ierr = MatZeroEntries(l->B);CHKERRQ(ierr); 1410 PetscFunctionReturn(0); 1411 } 1412 1413 #undef __FUNCT__ 1414 #define __FUNCT__ "MatGetInfo_MPISBAIJ" 1415 PetscErrorCode MatGetInfo_MPISBAIJ(Mat matin,MatInfoType flag,MatInfo *info) 1416 { 1417 Mat_MPISBAIJ *a = (Mat_MPISBAIJ*)matin->data; 1418 Mat A = a->A,B = a->B; 1419 PetscErrorCode ierr; 1420 PetscReal isend[5],irecv[5]; 1421 1422 PetscFunctionBegin; 1423 info->block_size = (PetscReal)matin->rmap->bs; 1424 1425 ierr = MatGetInfo(A,MAT_LOCAL,info);CHKERRQ(ierr); 1426 1427 isend[0] = info->nz_used; isend[1] = info->nz_allocated; isend[2] = info->nz_unneeded; 1428 isend[3] = info->memory; isend[4] = info->mallocs; 1429 1430 ierr = MatGetInfo(B,MAT_LOCAL,info);CHKERRQ(ierr); 1431 1432 isend[0] += info->nz_used; isend[1] += info->nz_allocated; isend[2] += info->nz_unneeded; 1433 isend[3] += info->memory; isend[4] += info->mallocs; 1434 if (flag == MAT_LOCAL) { 1435 info->nz_used = isend[0]; 1436 info->nz_allocated = isend[1]; 1437 info->nz_unneeded = isend[2]; 1438 info->memory = isend[3]; 1439 info->mallocs = isend[4]; 1440 } else if (flag == MAT_GLOBAL_MAX) { 1441 ierr = MPI_Allreduce(isend,irecv,5,MPIU_REAL,MPIU_MAX,PetscObjectComm((PetscObject)matin));CHKERRQ(ierr); 1442 1443 info->nz_used = irecv[0]; 1444 info->nz_allocated = irecv[1]; 1445 info->nz_unneeded = irecv[2]; 1446 info->memory = irecv[3]; 1447 info->mallocs = irecv[4]; 1448 } else if (flag == MAT_GLOBAL_SUM) { 1449 ierr = MPI_Allreduce(isend,irecv,5,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)matin));CHKERRQ(ierr); 1450 1451 info->nz_used = irecv[0]; 1452 info->nz_allocated = irecv[1]; 1453 info->nz_unneeded = irecv[2]; 1454 info->memory = irecv[3]; 1455 info->mallocs = irecv[4]; 1456 } else SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Unknown MatInfoType argument %d",(int)flag); 1457 info->fill_ratio_given = 0; /* no parallel LU/ILU/Cholesky */ 1458 info->fill_ratio_needed = 0; 1459 info->factor_mallocs = 0; 1460 PetscFunctionReturn(0); 1461 } 1462 1463 #undef __FUNCT__ 1464 #define __FUNCT__ "MatSetOption_MPISBAIJ" 1465 PetscErrorCode MatSetOption_MPISBAIJ(Mat A,MatOption op,PetscBool flg) 1466 { 1467 Mat_MPISBAIJ *a = (Mat_MPISBAIJ*)A->data; 1468 Mat_SeqSBAIJ *aA = (Mat_SeqSBAIJ*)a->A->data; 1469 PetscErrorCode ierr; 1470 1471 PetscFunctionBegin; 1472 switch (op) { 1473 case MAT_NEW_NONZERO_LOCATIONS: 1474 case MAT_NEW_NONZERO_ALLOCATION_ERR: 1475 case MAT_UNUSED_NONZERO_LOCATION_ERR: 1476 case MAT_KEEP_NONZERO_PATTERN: 1477 case MAT_NEW_NONZERO_LOCATION_ERR: 1478 ierr = MatSetOption(a->A,op,flg);CHKERRQ(ierr); 1479 ierr = MatSetOption(a->B,op,flg);CHKERRQ(ierr); 1480 break; 1481 case MAT_ROW_ORIENTED: 1482 a->roworiented = flg; 1483 1484 ierr = MatSetOption(a->A,op,flg);CHKERRQ(ierr); 1485 ierr = MatSetOption(a->B,op,flg);CHKERRQ(ierr); 1486 break; 1487 case MAT_NEW_DIAGONALS: 1488 ierr = PetscInfo1(A,"Option %s ignored\n",MatOptions[op]);CHKERRQ(ierr); 1489 break; 1490 case MAT_IGNORE_OFF_PROC_ENTRIES: 1491 a->donotstash = flg; 1492 break; 1493 case MAT_USE_HASH_TABLE: 1494 a->ht_flag = flg; 1495 break; 1496 case MAT_HERMITIAN: 1497 if (!A->assembled) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Must call MatAssemblyEnd() first"); 1498 ierr = MatSetOption(a->A,op,flg);CHKERRQ(ierr); 1499 1500 A->ops->mult = MatMult_MPISBAIJ_Hermitian; 1501 break; 1502 case MAT_SPD: 1503 A->spd_set = PETSC_TRUE; 1504 A->spd = flg; 1505 if (flg) { 1506 A->symmetric = PETSC_TRUE; 1507 A->structurally_symmetric = PETSC_TRUE; 1508 A->symmetric_set = PETSC_TRUE; 1509 A->structurally_symmetric_set = PETSC_TRUE; 1510 } 1511 break; 1512 case MAT_SYMMETRIC: 1513 ierr = MatSetOption(a->A,op,flg);CHKERRQ(ierr); 1514 break; 1515 case MAT_STRUCTURALLY_SYMMETRIC: 1516 ierr = MatSetOption(a->A,op,flg);CHKERRQ(ierr); 1517 break; 1518 case MAT_SYMMETRY_ETERNAL: 1519 if (!flg) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Matrix must be symmetric"); 1520 ierr = PetscInfo1(A,"Option %s ignored\n",MatOptions[op]);CHKERRQ(ierr); 1521 break; 1522 case MAT_IGNORE_LOWER_TRIANGULAR: 1523 aA->ignore_ltriangular = flg; 1524 break; 1525 case MAT_ERROR_LOWER_TRIANGULAR: 1526 aA->ignore_ltriangular = flg; 1527 break; 1528 case MAT_GETROW_UPPERTRIANGULAR: 1529 aA->getrow_utriangular = flg; 1530 break; 1531 default: 1532 SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"unknown option %d",op); 1533 } 1534 PetscFunctionReturn(0); 1535 } 1536 1537 #undef __FUNCT__ 1538 #define __FUNCT__ "MatTranspose_MPISBAIJ" 1539 PetscErrorCode MatTranspose_MPISBAIJ(Mat A,MatReuse reuse,Mat *B) 1540 { 1541 PetscErrorCode ierr; 1542 1543 PetscFunctionBegin; 1544 if (MAT_INITIAL_MATRIX || *B != A) { 1545 ierr = MatDuplicate(A,MAT_COPY_VALUES,B);CHKERRQ(ierr); 1546 } 1547 PetscFunctionReturn(0); 1548 } 1549 1550 #undef __FUNCT__ 1551 #define __FUNCT__ "MatDiagonalScale_MPISBAIJ" 1552 PetscErrorCode MatDiagonalScale_MPISBAIJ(Mat mat,Vec ll,Vec rr) 1553 { 1554 Mat_MPISBAIJ *baij = (Mat_MPISBAIJ*)mat->data; 1555 Mat a = baij->A, b=baij->B; 1556 PetscErrorCode ierr; 1557 PetscInt nv,m,n; 1558 PetscBool flg; 1559 1560 PetscFunctionBegin; 1561 if (ll != rr) { 1562 ierr = VecEqual(ll,rr,&flg);CHKERRQ(ierr); 1563 if (!flg) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"For symmetric format, left and right scaling vectors must be same\n"); 1564 } 1565 if (!ll) PetscFunctionReturn(0); 1566 1567 ierr = MatGetLocalSize(mat,&m,&n);CHKERRQ(ierr); 1568 if (m != n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"For symmetric format, local size %d %d must be same",m,n); 1569 1570 ierr = VecGetLocalSize(rr,&nv);CHKERRQ(ierr); 1571 if (nv!=n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Left and right vector non-conforming local size"); 1572 1573 ierr = VecScatterBegin(baij->Mvctx,rr,baij->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1574 1575 /* left diagonalscale the off-diagonal part */ 1576 ierr = (*b->ops->diagonalscale)(b,ll,NULL);CHKERRQ(ierr); 1577 1578 /* scale the diagonal part */ 1579 ierr = (*a->ops->diagonalscale)(a,ll,rr);CHKERRQ(ierr); 1580 1581 /* right diagonalscale the off-diagonal part */ 1582 ierr = VecScatterEnd(baij->Mvctx,rr,baij->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1583 ierr = (*b->ops->diagonalscale)(b,NULL,baij->lvec);CHKERRQ(ierr); 1584 PetscFunctionReturn(0); 1585 } 1586 1587 #undef __FUNCT__ 1588 #define __FUNCT__ "MatSetUnfactored_MPISBAIJ" 1589 PetscErrorCode MatSetUnfactored_MPISBAIJ(Mat A) 1590 { 1591 Mat_MPISBAIJ *a = (Mat_MPISBAIJ*)A->data; 1592 PetscErrorCode ierr; 1593 1594 PetscFunctionBegin; 1595 ierr = MatSetUnfactored(a->A);CHKERRQ(ierr); 1596 PetscFunctionReturn(0); 1597 } 1598 1599 static PetscErrorCode MatDuplicate_MPISBAIJ(Mat,MatDuplicateOption,Mat*); 1600 1601 #undef __FUNCT__ 1602 #define __FUNCT__ "MatEqual_MPISBAIJ" 1603 PetscErrorCode MatEqual_MPISBAIJ(Mat A,Mat B,PetscBool *flag) 1604 { 1605 Mat_MPISBAIJ *matB = (Mat_MPISBAIJ*)B->data,*matA = (Mat_MPISBAIJ*)A->data; 1606 Mat a,b,c,d; 1607 PetscBool flg; 1608 PetscErrorCode ierr; 1609 1610 PetscFunctionBegin; 1611 a = matA->A; b = matA->B; 1612 c = matB->A; d = matB->B; 1613 1614 ierr = MatEqual(a,c,&flg);CHKERRQ(ierr); 1615 if (flg) { 1616 ierr = MatEqual(b,d,&flg);CHKERRQ(ierr); 1617 } 1618 ierr = MPI_Allreduce(&flg,flag,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)A));CHKERRQ(ierr); 1619 PetscFunctionReturn(0); 1620 } 1621 1622 #undef __FUNCT__ 1623 #define __FUNCT__ "MatCopy_MPISBAIJ" 1624 PetscErrorCode MatCopy_MPISBAIJ(Mat A,Mat B,MatStructure str) 1625 { 1626 PetscErrorCode ierr; 1627 Mat_MPISBAIJ *a = (Mat_MPISBAIJ*)A->data; 1628 Mat_MPISBAIJ *b = (Mat_MPISBAIJ*)B->data; 1629 1630 PetscFunctionBegin; 1631 /* If the two matrices don't have the same copy implementation, they aren't compatible for fast copy. */ 1632 if ((str != SAME_NONZERO_PATTERN) || (A->ops->copy != B->ops->copy)) { 1633 ierr = MatGetRowUpperTriangular(A);CHKERRQ(ierr); 1634 ierr = MatCopy_Basic(A,B,str);CHKERRQ(ierr); 1635 ierr = MatRestoreRowUpperTriangular(A);CHKERRQ(ierr); 1636 } else { 1637 ierr = MatCopy(a->A,b->A,str);CHKERRQ(ierr); 1638 ierr = MatCopy(a->B,b->B,str);CHKERRQ(ierr); 1639 } 1640 PetscFunctionReturn(0); 1641 } 1642 1643 #undef __FUNCT__ 1644 #define __FUNCT__ "MatSetUp_MPISBAIJ" 1645 PetscErrorCode MatSetUp_MPISBAIJ(Mat A) 1646 { 1647 PetscErrorCode ierr; 1648 1649 PetscFunctionBegin; 1650 ierr = MatMPISBAIJSetPreallocation(A,A->rmap->bs,PETSC_DEFAULT,0,PETSC_DEFAULT,0);CHKERRQ(ierr); 1651 PetscFunctionReturn(0); 1652 } 1653 1654 #undef __FUNCT__ 1655 #define __FUNCT__ "MatAXPY_MPISBAIJ" 1656 PetscErrorCode MatAXPY_MPISBAIJ(Mat Y,PetscScalar a,Mat X,MatStructure str) 1657 { 1658 PetscErrorCode ierr; 1659 Mat_MPISBAIJ *xx=(Mat_MPISBAIJ*)X->data,*yy=(Mat_MPISBAIJ*)Y->data; 1660 PetscBLASInt bnz,one=1; 1661 Mat_SeqSBAIJ *xa,*ya; 1662 Mat_SeqBAIJ *xb,*yb; 1663 1664 PetscFunctionBegin; 1665 if (str == SAME_NONZERO_PATTERN) { 1666 PetscScalar alpha = a; 1667 xa = (Mat_SeqSBAIJ*)xx->A->data; 1668 ya = (Mat_SeqSBAIJ*)yy->A->data; 1669 ierr = PetscBLASIntCast(xa->nz,&bnz);CHKERRQ(ierr); 1670 PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&bnz,&alpha,xa->a,&one,ya->a,&one)); 1671 xb = (Mat_SeqBAIJ*)xx->B->data; 1672 yb = (Mat_SeqBAIJ*)yy->B->data; 1673 ierr = PetscBLASIntCast(xb->nz,&bnz);CHKERRQ(ierr); 1674 PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&bnz,&alpha,xb->a,&one,yb->a,&one)); 1675 ierr = PetscObjectStateIncrease((PetscObject)Y);CHKERRQ(ierr); 1676 } else if (str == SUBSET_NONZERO_PATTERN) { /* nonzeros of X is a subset of Y's */ 1677 ierr = MatSetOption(X,MAT_GETROW_UPPERTRIANGULAR,PETSC_TRUE);CHKERRQ(ierr); 1678 ierr = MatAXPY_Basic(Y,a,X,str);CHKERRQ(ierr); 1679 ierr = MatSetOption(X,MAT_GETROW_UPPERTRIANGULAR,PETSC_FALSE);CHKERRQ(ierr); 1680 } else { 1681 Mat B; 1682 PetscInt *nnz_d,*nnz_o,bs=Y->rmap->bs; 1683 if (bs != X->rmap->bs) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Matrices must have same block size"); 1684 ierr = MatGetRowUpperTriangular(X);CHKERRQ(ierr); 1685 ierr = MatGetRowUpperTriangular(Y);CHKERRQ(ierr); 1686 ierr = PetscMalloc1(yy->A->rmap->N,&nnz_d);CHKERRQ(ierr); 1687 ierr = PetscMalloc1(yy->B->rmap->N,&nnz_o);CHKERRQ(ierr); 1688 ierr = MatCreate(PetscObjectComm((PetscObject)Y),&B);CHKERRQ(ierr); 1689 ierr = PetscObjectSetName((PetscObject)B,((PetscObject)Y)->name);CHKERRQ(ierr); 1690 ierr = MatSetSizes(B,Y->rmap->n,Y->cmap->n,Y->rmap->N,Y->cmap->N);CHKERRQ(ierr); 1691 ierr = MatSetBlockSizesFromMats(B,Y,Y);CHKERRQ(ierr); 1692 ierr = MatSetType(B,MATMPISBAIJ);CHKERRQ(ierr); 1693 ierr = MatAXPYGetPreallocation_SeqSBAIJ(yy->A,xx->A,nnz_d);CHKERRQ(ierr); 1694 ierr = MatAXPYGetPreallocation_MPIBAIJ(yy->B,yy->garray,xx->B,xx->garray,nnz_o);CHKERRQ(ierr); 1695 ierr = MatMPISBAIJSetPreallocation(B,bs,0,nnz_d,0,nnz_o);CHKERRQ(ierr); 1696 ierr = MatAXPY_BasicWithPreallocation(B,Y,a,X,str);CHKERRQ(ierr); 1697 ierr = MatHeaderReplace(Y,B);CHKERRQ(ierr); 1698 ierr = PetscFree(nnz_d);CHKERRQ(ierr); 1699 ierr = PetscFree(nnz_o);CHKERRQ(ierr); 1700 ierr = MatRestoreRowUpperTriangular(X);CHKERRQ(ierr); 1701 ierr = MatRestoreRowUpperTriangular(Y);CHKERRQ(ierr); 1702 } 1703 PetscFunctionReturn(0); 1704 } 1705 1706 #undef __FUNCT__ 1707 #define __FUNCT__ "MatGetSubMatrices_MPISBAIJ" 1708 PetscErrorCode MatGetSubMatrices_MPISBAIJ(Mat A,PetscInt n,const IS irow[],const IS icol[],MatReuse scall,Mat *B[]) 1709 { 1710 PetscErrorCode ierr; 1711 PetscInt i; 1712 PetscBool flg; 1713 1714 PetscFunctionBegin; 1715 ierr = MatGetSubMatrices_MPIBAIJ(A,n,irow,icol,scall,B);CHKERRQ(ierr); 1716 for (i=0; i<n; i++) { 1717 ierr = ISEqual(irow[i],icol[i],&flg);CHKERRQ(ierr); 1718 if (!flg) { /* *B[i] is non-symmetric, set flag */ 1719 ierr = MatSetOption(*B[i],MAT_SYMMETRIC,PETSC_FALSE);CHKERRQ(ierr); 1720 } 1721 } 1722 PetscFunctionReturn(0); 1723 } 1724 1725 #undef __FUNCT__ 1726 #define __FUNCT__ "MatShift_MPISBAIJ" 1727 PetscErrorCode MatShift_MPISBAIJ(Mat Y,PetscScalar a) 1728 { 1729 PetscErrorCode ierr; 1730 Mat_MPISBAIJ *maij = (Mat_MPISBAIJ*)Y->data; 1731 Mat_SeqSBAIJ *aij = (Mat_SeqSBAIJ*)maij->A->data,*bij = (Mat_SeqSBAIJ*)maij->B->data; 1732 1733 PetscFunctionBegin; 1734 if (!aij->nz && !bij->nz) { 1735 ierr = MatMPISBAIJSetPreallocation(Y,Y->rmap->bs,1,NULL,0,NULL);CHKERRQ(ierr); 1736 } 1737 ierr = MatShift_Basic(Y,a);CHKERRQ(ierr); 1738 PetscFunctionReturn(0); 1739 } 1740 1741 /* -------------------------------------------------------------------*/ 1742 static struct _MatOps MatOps_Values = {MatSetValues_MPISBAIJ, 1743 MatGetRow_MPISBAIJ, 1744 MatRestoreRow_MPISBAIJ, 1745 MatMult_MPISBAIJ, 1746 /* 4*/ MatMultAdd_MPISBAIJ, 1747 MatMult_MPISBAIJ, /* transpose versions are same as non-transpose */ 1748 MatMultAdd_MPISBAIJ, 1749 0, 1750 0, 1751 0, 1752 /* 10*/ 0, 1753 0, 1754 0, 1755 MatSOR_MPISBAIJ, 1756 MatTranspose_MPISBAIJ, 1757 /* 15*/ MatGetInfo_MPISBAIJ, 1758 MatEqual_MPISBAIJ, 1759 MatGetDiagonal_MPISBAIJ, 1760 MatDiagonalScale_MPISBAIJ, 1761 MatNorm_MPISBAIJ, 1762 /* 20*/ MatAssemblyBegin_MPISBAIJ, 1763 MatAssemblyEnd_MPISBAIJ, 1764 MatSetOption_MPISBAIJ, 1765 MatZeroEntries_MPISBAIJ, 1766 /* 24*/ 0, 1767 0, 1768 0, 1769 0, 1770 0, 1771 /* 29*/ MatSetUp_MPISBAIJ, 1772 0, 1773 0, 1774 0, 1775 0, 1776 /* 34*/ MatDuplicate_MPISBAIJ, 1777 0, 1778 0, 1779 0, 1780 0, 1781 /* 39*/ MatAXPY_MPISBAIJ, 1782 MatGetSubMatrices_MPISBAIJ, 1783 MatIncreaseOverlap_MPISBAIJ, 1784 MatGetValues_MPISBAIJ, 1785 MatCopy_MPISBAIJ, 1786 /* 44*/ 0, 1787 MatScale_MPISBAIJ, 1788 MatShift_MPISBAIJ, 1789 0, 1790 0, 1791 /* 49*/ 0, 1792 0, 1793 0, 1794 0, 1795 0, 1796 /* 54*/ 0, 1797 0, 1798 MatSetUnfactored_MPISBAIJ, 1799 0, 1800 MatSetValuesBlocked_MPISBAIJ, 1801 /* 59*/ MatGetSubMatrix_MPISBAIJ, 1802 0, 1803 0, 1804 0, 1805 0, 1806 /* 64*/ 0, 1807 0, 1808 0, 1809 0, 1810 0, 1811 /* 69*/ MatGetRowMaxAbs_MPISBAIJ, 1812 0, 1813 0, 1814 0, 1815 0, 1816 /* 74*/ 0, 1817 0, 1818 0, 1819 0, 1820 0, 1821 /* 79*/ 0, 1822 0, 1823 0, 1824 0, 1825 MatLoad_MPISBAIJ, 1826 /* 84*/ 0, 1827 0, 1828 0, 1829 0, 1830 0, 1831 /* 89*/ 0, 1832 0, 1833 0, 1834 0, 1835 0, 1836 /* 94*/ 0, 1837 0, 1838 0, 1839 0, 1840 0, 1841 /* 99*/ 0, 1842 0, 1843 0, 1844 0, 1845 0, 1846 /*104*/ 0, 1847 MatRealPart_MPISBAIJ, 1848 MatImaginaryPart_MPISBAIJ, 1849 MatGetRowUpperTriangular_MPISBAIJ, 1850 MatRestoreRowUpperTriangular_MPISBAIJ, 1851 /*109*/ 0, 1852 0, 1853 0, 1854 0, 1855 0, 1856 /*114*/ 0, 1857 0, 1858 0, 1859 0, 1860 0, 1861 /*119*/ 0, 1862 0, 1863 0, 1864 0, 1865 0, 1866 /*124*/ 0, 1867 0, 1868 0, 1869 0, 1870 0, 1871 /*129*/ 0, 1872 0, 1873 0, 1874 0, 1875 0, 1876 /*134*/ 0, 1877 0, 1878 0, 1879 0, 1880 0, 1881 /*139*/ 0, 1882 0, 1883 0, 1884 0, 1885 0, 1886 /*144*/MatCreateMPIMatConcatenateSeqMat_MPISBAIJ 1887 }; 1888 1889 #undef __FUNCT__ 1890 #define __FUNCT__ "MatGetDiagonalBlock_MPISBAIJ" 1891 PetscErrorCode MatGetDiagonalBlock_MPISBAIJ(Mat A,Mat *a) 1892 { 1893 PetscFunctionBegin; 1894 *a = ((Mat_MPISBAIJ*)A->data)->A; 1895 PetscFunctionReturn(0); 1896 } 1897 1898 #undef __FUNCT__ 1899 #define __FUNCT__ "MatMPISBAIJSetPreallocation_MPISBAIJ" 1900 PetscErrorCode MatMPISBAIJSetPreallocation_MPISBAIJ(Mat B,PetscInt bs,PetscInt d_nz,const PetscInt *d_nnz,PetscInt o_nz,const PetscInt *o_nnz) 1901 { 1902 Mat_MPISBAIJ *b; 1903 PetscErrorCode ierr; 1904 PetscInt i,mbs,Mbs; 1905 1906 PetscFunctionBegin; 1907 ierr = MatSetBlockSize(B,PetscAbs(bs));CHKERRQ(ierr); 1908 ierr = PetscLayoutSetUp(B->rmap);CHKERRQ(ierr); 1909 ierr = PetscLayoutSetUp(B->cmap);CHKERRQ(ierr); 1910 ierr = PetscLayoutGetBlockSize(B->rmap,&bs);CHKERRQ(ierr); 1911 1912 b = (Mat_MPISBAIJ*)B->data; 1913 mbs = B->rmap->n/bs; 1914 Mbs = B->rmap->N/bs; 1915 if (mbs*bs != B->rmap->n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"No of local rows %D must be divisible by blocksize %D",B->rmap->N,bs); 1916 1917 B->rmap->bs = bs; 1918 b->bs2 = bs*bs; 1919 b->mbs = mbs; 1920 b->Mbs = Mbs; 1921 b->nbs = B->cmap->n/bs; 1922 b->Nbs = B->cmap->N/bs; 1923 1924 for (i=0; i<=b->size; i++) { 1925 b->rangebs[i] = B->rmap->range[i]/bs; 1926 } 1927 b->rstartbs = B->rmap->rstart/bs; 1928 b->rendbs = B->rmap->rend/bs; 1929 1930 b->cstartbs = B->cmap->rstart/bs; 1931 b->cendbs = B->cmap->rend/bs; 1932 1933 if (!B->preallocated) { 1934 ierr = MatCreate(PETSC_COMM_SELF,&b->A);CHKERRQ(ierr); 1935 ierr = MatSetSizes(b->A,B->rmap->n,B->cmap->n,B->rmap->n,B->cmap->n);CHKERRQ(ierr); 1936 ierr = MatSetType(b->A,MATSEQSBAIJ);CHKERRQ(ierr); 1937 ierr = PetscLogObjectParent((PetscObject)B,(PetscObject)b->A);CHKERRQ(ierr); 1938 ierr = MatCreate(PETSC_COMM_SELF,&b->B);CHKERRQ(ierr); 1939 ierr = MatSetSizes(b->B,B->rmap->n,B->cmap->N,B->rmap->n,B->cmap->N);CHKERRQ(ierr); 1940 ierr = MatSetType(b->B,MATSEQBAIJ);CHKERRQ(ierr); 1941 ierr = PetscLogObjectParent((PetscObject)B,(PetscObject)b->B);CHKERRQ(ierr); 1942 ierr = MatStashCreate_Private(PetscObjectComm((PetscObject)B),bs,&B->bstash);CHKERRQ(ierr); 1943 } 1944 1945 ierr = MatSeqSBAIJSetPreallocation(b->A,bs,d_nz,d_nnz);CHKERRQ(ierr); 1946 ierr = MatSeqBAIJSetPreallocation(b->B,bs,o_nz,o_nnz);CHKERRQ(ierr); 1947 1948 B->preallocated = PETSC_TRUE; 1949 PetscFunctionReturn(0); 1950 } 1951 1952 #undef __FUNCT__ 1953 #define __FUNCT__ "MatMPISBAIJSetPreallocationCSR_MPISBAIJ" 1954 PetscErrorCode MatMPISBAIJSetPreallocationCSR_MPISBAIJ(Mat B,PetscInt bs,const PetscInt ii[],const PetscInt jj[],const PetscScalar V[]) 1955 { 1956 PetscInt m,rstart,cstart,cend; 1957 PetscInt i,j,d,nz,nz_max=0,*d_nnz=0,*o_nnz=0; 1958 const PetscInt *JJ =0; 1959 PetscScalar *values=0; 1960 PetscErrorCode ierr; 1961 1962 PetscFunctionBegin; 1963 if (bs < 1) SETERRQ1(PetscObjectComm((PetscObject)B),PETSC_ERR_ARG_OUTOFRANGE,"Invalid block size specified, must be positive but it is %D",bs); 1964 ierr = PetscLayoutSetBlockSize(B->rmap,bs);CHKERRQ(ierr); 1965 ierr = PetscLayoutSetBlockSize(B->cmap,bs);CHKERRQ(ierr); 1966 ierr = PetscLayoutSetUp(B->rmap);CHKERRQ(ierr); 1967 ierr = PetscLayoutSetUp(B->cmap);CHKERRQ(ierr); 1968 ierr = PetscLayoutGetBlockSize(B->rmap,&bs);CHKERRQ(ierr); 1969 m = B->rmap->n/bs; 1970 rstart = B->rmap->rstart/bs; 1971 cstart = B->cmap->rstart/bs; 1972 cend = B->cmap->rend/bs; 1973 1974 if (ii[0]) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"ii[0] must be 0 but it is %D",ii[0]); 1975 ierr = PetscMalloc2(m,&d_nnz,m,&o_nnz);CHKERRQ(ierr); 1976 for (i=0; i<m; i++) { 1977 nz = ii[i+1] - ii[i]; 1978 if (nz < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Local row %D has a negative number of columns %D",i,nz); 1979 nz_max = PetscMax(nz_max,nz); 1980 JJ = jj + ii[i]; 1981 for (j=0; j<nz; j++) { 1982 if (*JJ >= cstart) break; 1983 JJ++; 1984 } 1985 d = 0; 1986 for (; j<nz; j++) { 1987 if (*JJ++ >= cend) break; 1988 d++; 1989 } 1990 d_nnz[i] = d; 1991 o_nnz[i] = nz - d; 1992 } 1993 ierr = MatMPISBAIJSetPreallocation(B,bs,0,d_nnz,0,o_nnz);CHKERRQ(ierr); 1994 ierr = PetscFree2(d_nnz,o_nnz);CHKERRQ(ierr); 1995 1996 values = (PetscScalar*)V; 1997 if (!values) { 1998 ierr = PetscMalloc1(bs*bs*nz_max,&values);CHKERRQ(ierr); 1999 ierr = PetscMemzero(values,bs*bs*nz_max*sizeof(PetscScalar));CHKERRQ(ierr); 2000 } 2001 for (i=0; i<m; i++) { 2002 PetscInt row = i + rstart; 2003 PetscInt ncols = ii[i+1] - ii[i]; 2004 const PetscInt *icols = jj + ii[i]; 2005 const PetscScalar *svals = values + (V ? (bs*bs*ii[i]) : 0); 2006 ierr = MatSetValuesBlocked_MPISBAIJ(B,1,&row,ncols,icols,svals,INSERT_VALUES);CHKERRQ(ierr); 2007 } 2008 2009 if (!V) { ierr = PetscFree(values);CHKERRQ(ierr); } 2010 ierr = MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2011 ierr = MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2012 ierr = MatSetOption(B,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE);CHKERRQ(ierr); 2013 PetscFunctionReturn(0); 2014 } 2015 2016 /*MC 2017 MATMPISBAIJ - MATMPISBAIJ = "mpisbaij" - A matrix type to be used for distributed symmetric sparse block matrices, 2018 based on block compressed sparse row format. Only the upper triangular portion of the "diagonal" portion of 2019 the matrix is stored. 2020 2021 For complex numbers by default this matrix is symmetric, NOT Hermitian symmetric. To make it Hermitian symmetric you 2022 can call MatSetOption(Mat, MAT_HERMITIAN); 2023 2024 Options Database Keys: 2025 . -mat_type mpisbaij - sets the matrix type to "mpisbaij" during a call to MatSetFromOptions() 2026 2027 Level: beginner 2028 2029 .seealso: MatCreateMPISBAIJ 2030 M*/ 2031 2032 PETSC_EXTERN PetscErrorCode MatConvert_MPISBAIJ_MPISBSTRM(Mat,MatType,MatReuse,Mat*); 2033 2034 #undef __FUNCT__ 2035 #define __FUNCT__ "MatCreate_MPISBAIJ" 2036 PETSC_EXTERN PetscErrorCode MatCreate_MPISBAIJ(Mat B) 2037 { 2038 Mat_MPISBAIJ *b; 2039 PetscErrorCode ierr; 2040 PetscBool flg = PETSC_FALSE; 2041 2042 PetscFunctionBegin; 2043 ierr = PetscNewLog(B,&b);CHKERRQ(ierr); 2044 B->data = (void*)b; 2045 ierr = PetscMemcpy(B->ops,&MatOps_Values,sizeof(struct _MatOps));CHKERRQ(ierr); 2046 2047 B->ops->destroy = MatDestroy_MPISBAIJ; 2048 B->ops->view = MatView_MPISBAIJ; 2049 B->assembled = PETSC_FALSE; 2050 B->insertmode = NOT_SET_VALUES; 2051 2052 ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)B),&b->rank);CHKERRQ(ierr); 2053 ierr = MPI_Comm_size(PetscObjectComm((PetscObject)B),&b->size);CHKERRQ(ierr); 2054 2055 /* build local table of row and column ownerships */ 2056 ierr = PetscMalloc1(b->size+2,&b->rangebs);CHKERRQ(ierr); 2057 2058 /* build cache for off array entries formed */ 2059 ierr = MatStashCreate_Private(PetscObjectComm((PetscObject)B),1,&B->stash);CHKERRQ(ierr); 2060 2061 b->donotstash = PETSC_FALSE; 2062 b->colmap = NULL; 2063 b->garray = NULL; 2064 b->roworiented = PETSC_TRUE; 2065 2066 /* stuff used in block assembly */ 2067 b->barray = 0; 2068 2069 /* stuff used for matrix vector multiply */ 2070 b->lvec = 0; 2071 b->Mvctx = 0; 2072 b->slvec0 = 0; 2073 b->slvec0b = 0; 2074 b->slvec1 = 0; 2075 b->slvec1a = 0; 2076 b->slvec1b = 0; 2077 b->sMvctx = 0; 2078 2079 /* stuff for MatGetRow() */ 2080 b->rowindices = 0; 2081 b->rowvalues = 0; 2082 b->getrowactive = PETSC_FALSE; 2083 2084 /* hash table stuff */ 2085 b->ht = 0; 2086 b->hd = 0; 2087 b->ht_size = 0; 2088 b->ht_flag = PETSC_FALSE; 2089 b->ht_fact = 0; 2090 b->ht_total_ct = 0; 2091 b->ht_insert_ct = 0; 2092 2093 /* stuff for MatGetSubMatrices_MPIBAIJ_local() */ 2094 b->ijonly = PETSC_FALSE; 2095 2096 b->in_loc = 0; 2097 b->v_loc = 0; 2098 b->n_loc = 0; 2099 2100 ierr = PetscObjectComposeFunction((PetscObject)B,"MatStoreValues_C",MatStoreValues_MPISBAIJ);CHKERRQ(ierr); 2101 ierr = PetscObjectComposeFunction((PetscObject)B,"MatRetrieveValues_C",MatRetrieveValues_MPISBAIJ);CHKERRQ(ierr); 2102 ierr = PetscObjectComposeFunction((PetscObject)B,"MatGetDiagonalBlock_C",MatGetDiagonalBlock_MPISBAIJ);CHKERRQ(ierr); 2103 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMPISBAIJSetPreallocation_C",MatMPISBAIJSetPreallocation_MPISBAIJ);CHKERRQ(ierr); 2104 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMPISBAIJSetPreallocationCSR_C",MatMPISBAIJSetPreallocationCSR_MPISBAIJ);CHKERRQ(ierr); 2105 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpisbaij_mpisbstrm_C",MatConvert_MPISBAIJ_MPISBSTRM);CHKERRQ(ierr); 2106 #if defined(PETSC_HAVE_ELEMENTAL) 2107 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpisbaij_elemental_C",MatConvert_MPISBAIJ_Elemental);CHKERRQ(ierr); 2108 #endif 2109 2110 B->symmetric = PETSC_TRUE; 2111 B->structurally_symmetric = PETSC_TRUE; 2112 B->symmetric_set = PETSC_TRUE; 2113 B->structurally_symmetric_set = PETSC_TRUE; 2114 2115 ierr = PetscObjectChangeTypeName((PetscObject)B,MATMPISBAIJ);CHKERRQ(ierr); 2116 ierr = PetscOptionsBegin(PetscObjectComm((PetscObject)B),NULL,"Options for loading MPISBAIJ matrix 1","Mat");CHKERRQ(ierr); 2117 ierr = PetscOptionsBool("-mat_use_hash_table","Use hash table to save memory in constructing matrix","MatSetOption",flg,&flg,NULL);CHKERRQ(ierr); 2118 if (flg) { 2119 PetscReal fact = 1.39; 2120 ierr = MatSetOption(B,MAT_USE_HASH_TABLE,PETSC_TRUE);CHKERRQ(ierr); 2121 ierr = PetscOptionsReal("-mat_use_hash_table","Use hash table factor","MatMPIBAIJSetHashTableFactor",fact,&fact,NULL);CHKERRQ(ierr); 2122 if (fact <= 1.0) fact = 1.39; 2123 ierr = MatMPIBAIJSetHashTableFactor(B,fact);CHKERRQ(ierr); 2124 ierr = PetscInfo1(B,"Hash table Factor used %5.2f\n",fact);CHKERRQ(ierr); 2125 } 2126 ierr = PetscOptionsEnd();CHKERRQ(ierr); 2127 PetscFunctionReturn(0); 2128 } 2129 2130 /*MC 2131 MATSBAIJ - MATSBAIJ = "sbaij" - A matrix type to be used for symmetric block sparse matrices. 2132 2133 This matrix type is identical to MATSEQSBAIJ when constructed with a single process communicator, 2134 and MATMPISBAIJ otherwise. 2135 2136 Options Database Keys: 2137 . -mat_type sbaij - sets the matrix type to "sbaij" during a call to MatSetFromOptions() 2138 2139 Level: beginner 2140 2141 .seealso: MatCreateMPISBAIJ,MATSEQSBAIJ,MATMPISBAIJ 2142 M*/ 2143 2144 #undef __FUNCT__ 2145 #define __FUNCT__ "MatMPISBAIJSetPreallocation" 2146 /*@C 2147 MatMPISBAIJSetPreallocation - For good matrix assembly performance 2148 the user should preallocate the matrix storage by setting the parameters 2149 d_nz (or d_nnz) and o_nz (or o_nnz). By setting these parameters accurately, 2150 performance can be increased by more than a factor of 50. 2151 2152 Collective on Mat 2153 2154 Input Parameters: 2155 + B - the matrix 2156 . bs - size of block, the blocks are ALWAYS square. One can use MatSetBlockSizes() to set a different row and column blocksize but the row 2157 blocksize always defines the size of the blocks. The column blocksize sets the blocksize of the vectors obtained with MatCreateVecs() 2158 . d_nz - number of block nonzeros per block row in diagonal portion of local 2159 submatrix (same for all local rows) 2160 . d_nnz - array containing the number of block nonzeros in the various block rows 2161 in the upper triangular and diagonal part of the in diagonal portion of the local 2162 (possibly different for each block row) or NULL. If you plan to factor the matrix you must leave room 2163 for the diagonal entry and set a value even if it is zero. 2164 . o_nz - number of block nonzeros per block row in the off-diagonal portion of local 2165 submatrix (same for all local rows). 2166 - o_nnz - array containing the number of nonzeros in the various block rows of the 2167 off-diagonal portion of the local submatrix that is right of the diagonal 2168 (possibly different for each block row) or NULL. 2169 2170 2171 Options Database Keys: 2172 . -mat_no_unroll - uses code that does not unroll the loops in the 2173 block calculations (much slower) 2174 . -mat_block_size - size of the blocks to use 2175 2176 Notes: 2177 2178 If PETSC_DECIDE or PETSC_DETERMINE is used for a particular argument on one processor 2179 than it must be used on all processors that share the object for that argument. 2180 2181 If the *_nnz parameter is given then the *_nz parameter is ignored 2182 2183 Storage Information: 2184 For a square global matrix we define each processor's diagonal portion 2185 to be its local rows and the corresponding columns (a square submatrix); 2186 each processor's off-diagonal portion encompasses the remainder of the 2187 local matrix (a rectangular submatrix). 2188 2189 The user can specify preallocated storage for the diagonal part of 2190 the local submatrix with either d_nz or d_nnz (not both). Set 2191 d_nz=PETSC_DEFAULT and d_nnz=NULL for PETSc to control dynamic 2192 memory allocation. Likewise, specify preallocated storage for the 2193 off-diagonal part of the local submatrix with o_nz or o_nnz (not both). 2194 2195 You can call MatGetInfo() to get information on how effective the preallocation was; 2196 for example the fields mallocs,nz_allocated,nz_used,nz_unneeded; 2197 You can also run with the option -info and look for messages with the string 2198 malloc in them to see if additional memory allocation was needed. 2199 2200 Consider a processor that owns rows 3, 4 and 5 of a parallel matrix. In 2201 the figure below we depict these three local rows and all columns (0-11). 2202 2203 .vb 2204 0 1 2 3 4 5 6 7 8 9 10 11 2205 -------------------------- 2206 row 3 |. . . d d d o o o o o o 2207 row 4 |. . . d d d o o o o o o 2208 row 5 |. . . d d d o o o o o o 2209 -------------------------- 2210 .ve 2211 2212 Thus, any entries in the d locations are stored in the d (diagonal) 2213 submatrix, and any entries in the o locations are stored in the 2214 o (off-diagonal) submatrix. Note that the d matrix is stored in 2215 MatSeqSBAIJ format and the o submatrix in MATSEQBAIJ format. 2216 2217 Now d_nz should indicate the number of block nonzeros per row in the upper triangular 2218 plus the diagonal part of the d matrix, 2219 and o_nz should indicate the number of block nonzeros per row in the o matrix 2220 2221 In general, for PDE problems in which most nonzeros are near the diagonal, 2222 one expects d_nz >> o_nz. For large problems you MUST preallocate memory 2223 or you will get TERRIBLE performance; see the users' manual chapter on 2224 matrices. 2225 2226 Level: intermediate 2227 2228 .keywords: matrix, block, aij, compressed row, sparse, parallel 2229 2230 .seealso: MatCreate(), MatCreateSeqSBAIJ(), MatSetValues(), MatCreateBAIJ(), PetscSplitOwnership() 2231 @*/ 2232 PetscErrorCode MatMPISBAIJSetPreallocation(Mat B,PetscInt bs,PetscInt d_nz,const PetscInt d_nnz[],PetscInt o_nz,const PetscInt o_nnz[]) 2233 { 2234 PetscErrorCode ierr; 2235 2236 PetscFunctionBegin; 2237 PetscValidHeaderSpecific(B,MAT_CLASSID,1); 2238 PetscValidType(B,1); 2239 PetscValidLogicalCollectiveInt(B,bs,2); 2240 ierr = PetscTryMethod(B,"MatMPISBAIJSetPreallocation_C",(Mat,PetscInt,PetscInt,const PetscInt[],PetscInt,const PetscInt[]),(B,bs,d_nz,d_nnz,o_nz,o_nnz));CHKERRQ(ierr); 2241 PetscFunctionReturn(0); 2242 } 2243 2244 #undef __FUNCT__ 2245 #define __FUNCT__ "MatCreateSBAIJ" 2246 /*@C 2247 MatCreateSBAIJ - Creates a sparse parallel matrix in symmetric block AIJ format 2248 (block compressed row). For good matrix assembly performance 2249 the user should preallocate the matrix storage by setting the parameters 2250 d_nz (or d_nnz) and o_nz (or o_nnz). By setting these parameters accurately, 2251 performance can be increased by more than a factor of 50. 2252 2253 Collective on MPI_Comm 2254 2255 Input Parameters: 2256 + comm - MPI communicator 2257 . bs - size of block, the blocks are ALWAYS square. One can use MatSetBlockSizes() to set a different row and column blocksize but the row 2258 blocksize always defines the size of the blocks. The column blocksize sets the blocksize of the vectors obtained with MatCreateVecs() 2259 . m - number of local rows (or PETSC_DECIDE to have calculated if M is given) 2260 This value should be the same as the local size used in creating the 2261 y vector for the matrix-vector product y = Ax. 2262 . n - number of local columns (or PETSC_DECIDE to have calculated if N is given) 2263 This value should be the same as the local size used in creating the 2264 x vector for the matrix-vector product y = Ax. 2265 . M - number of global rows (or PETSC_DETERMINE to have calculated if m is given) 2266 . N - number of global columns (or PETSC_DETERMINE to have calculated if n is given) 2267 . d_nz - number of block nonzeros per block row in diagonal portion of local 2268 submatrix (same for all local rows) 2269 . d_nnz - array containing the number of block nonzeros in the various block rows 2270 in the upper triangular portion of the in diagonal portion of the local 2271 (possibly different for each block block row) or NULL. 2272 If you plan to factor the matrix you must leave room for the diagonal entry and 2273 set its value even if it is zero. 2274 . o_nz - number of block nonzeros per block row in the off-diagonal portion of local 2275 submatrix (same for all local rows). 2276 - o_nnz - array containing the number of nonzeros in the various block rows of the 2277 off-diagonal portion of the local submatrix (possibly different for 2278 each block row) or NULL. 2279 2280 Output Parameter: 2281 . A - the matrix 2282 2283 Options Database Keys: 2284 . -mat_no_unroll - uses code that does not unroll the loops in the 2285 block calculations (much slower) 2286 . -mat_block_size - size of the blocks to use 2287 . -mat_mpi - use the parallel matrix data structures even on one processor 2288 (defaults to using SeqBAIJ format on one processor) 2289 2290 It is recommended that one use the MatCreate(), MatSetType() and/or MatSetFromOptions(), 2291 MatXXXXSetPreallocation() paradgm instead of this routine directly. 2292 [MatXXXXSetPreallocation() is, for example, MatSeqAIJSetPreallocation] 2293 2294 Notes: 2295 The number of rows and columns must be divisible by blocksize. 2296 This matrix type does not support complex Hermitian operation. 2297 2298 The user MUST specify either the local or global matrix dimensions 2299 (possibly both). 2300 2301 If PETSC_DECIDE or PETSC_DETERMINE is used for a particular argument on one processor 2302 than it must be used on all processors that share the object for that argument. 2303 2304 If the *_nnz parameter is given then the *_nz parameter is ignored 2305 2306 Storage Information: 2307 For a square global matrix we define each processor's diagonal portion 2308 to be its local rows and the corresponding columns (a square submatrix); 2309 each processor's off-diagonal portion encompasses the remainder of the 2310 local matrix (a rectangular submatrix). 2311 2312 The user can specify preallocated storage for the diagonal part of 2313 the local submatrix with either d_nz or d_nnz (not both). Set 2314 d_nz=PETSC_DEFAULT and d_nnz=NULL for PETSc to control dynamic 2315 memory allocation. Likewise, specify preallocated storage for the 2316 off-diagonal part of the local submatrix with o_nz or o_nnz (not both). 2317 2318 Consider a processor that owns rows 3, 4 and 5 of a parallel matrix. In 2319 the figure below we depict these three local rows and all columns (0-11). 2320 2321 .vb 2322 0 1 2 3 4 5 6 7 8 9 10 11 2323 -------------------------- 2324 row 3 |. . . d d d o o o o o o 2325 row 4 |. . . d d d o o o o o o 2326 row 5 |. . . d d d o o o o o o 2327 -------------------------- 2328 .ve 2329 2330 Thus, any entries in the d locations are stored in the d (diagonal) 2331 submatrix, and any entries in the o locations are stored in the 2332 o (off-diagonal) submatrix. Note that the d matrix is stored in 2333 MatSeqSBAIJ format and the o submatrix in MATSEQBAIJ format. 2334 2335 Now d_nz should indicate the number of block nonzeros per row in the upper triangular 2336 plus the diagonal part of the d matrix, 2337 and o_nz should indicate the number of block nonzeros per row in the o matrix. 2338 In general, for PDE problems in which most nonzeros are near the diagonal, 2339 one expects d_nz >> o_nz. For large problems you MUST preallocate memory 2340 or you will get TERRIBLE performance; see the users' manual chapter on 2341 matrices. 2342 2343 Level: intermediate 2344 2345 .keywords: matrix, block, aij, compressed row, sparse, parallel 2346 2347 .seealso: MatCreate(), MatCreateSeqSBAIJ(), MatSetValues(), MatCreateBAIJ() 2348 @*/ 2349 2350 PetscErrorCode MatCreateSBAIJ(MPI_Comm comm,PetscInt bs,PetscInt m,PetscInt n,PetscInt M,PetscInt N,PetscInt d_nz,const PetscInt d_nnz[],PetscInt o_nz,const PetscInt o_nnz[],Mat *A) 2351 { 2352 PetscErrorCode ierr; 2353 PetscMPIInt size; 2354 2355 PetscFunctionBegin; 2356 ierr = MatCreate(comm,A);CHKERRQ(ierr); 2357 ierr = MatSetSizes(*A,m,n,M,N);CHKERRQ(ierr); 2358 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 2359 if (size > 1) { 2360 ierr = MatSetType(*A,MATMPISBAIJ);CHKERRQ(ierr); 2361 ierr = MatMPISBAIJSetPreallocation(*A,bs,d_nz,d_nnz,o_nz,o_nnz);CHKERRQ(ierr); 2362 } else { 2363 ierr = MatSetType(*A,MATSEQSBAIJ);CHKERRQ(ierr); 2364 ierr = MatSeqSBAIJSetPreallocation(*A,bs,d_nz,d_nnz);CHKERRQ(ierr); 2365 } 2366 PetscFunctionReturn(0); 2367 } 2368 2369 2370 #undef __FUNCT__ 2371 #define __FUNCT__ "MatDuplicate_MPISBAIJ" 2372 static PetscErrorCode MatDuplicate_MPISBAIJ(Mat matin,MatDuplicateOption cpvalues,Mat *newmat) 2373 { 2374 Mat mat; 2375 Mat_MPISBAIJ *a,*oldmat = (Mat_MPISBAIJ*)matin->data; 2376 PetscErrorCode ierr; 2377 PetscInt len=0,nt,bs=matin->rmap->bs,mbs=oldmat->mbs; 2378 PetscScalar *array; 2379 2380 PetscFunctionBegin; 2381 *newmat = 0; 2382 2383 ierr = MatCreate(PetscObjectComm((PetscObject)matin),&mat);CHKERRQ(ierr); 2384 ierr = MatSetSizes(mat,matin->rmap->n,matin->cmap->n,matin->rmap->N,matin->cmap->N);CHKERRQ(ierr); 2385 ierr = MatSetType(mat,((PetscObject)matin)->type_name);CHKERRQ(ierr); 2386 ierr = PetscMemcpy(mat->ops,matin->ops,sizeof(struct _MatOps));CHKERRQ(ierr); 2387 ierr = PetscLayoutReference(matin->rmap,&mat->rmap);CHKERRQ(ierr); 2388 ierr = PetscLayoutReference(matin->cmap,&mat->cmap);CHKERRQ(ierr); 2389 2390 mat->factortype = matin->factortype; 2391 mat->preallocated = PETSC_TRUE; 2392 mat->assembled = PETSC_TRUE; 2393 mat->insertmode = NOT_SET_VALUES; 2394 2395 a = (Mat_MPISBAIJ*)mat->data; 2396 a->bs2 = oldmat->bs2; 2397 a->mbs = oldmat->mbs; 2398 a->nbs = oldmat->nbs; 2399 a->Mbs = oldmat->Mbs; 2400 a->Nbs = oldmat->Nbs; 2401 2402 2403 a->size = oldmat->size; 2404 a->rank = oldmat->rank; 2405 a->donotstash = oldmat->donotstash; 2406 a->roworiented = oldmat->roworiented; 2407 a->rowindices = 0; 2408 a->rowvalues = 0; 2409 a->getrowactive = PETSC_FALSE; 2410 a->barray = 0; 2411 a->rstartbs = oldmat->rstartbs; 2412 a->rendbs = oldmat->rendbs; 2413 a->cstartbs = oldmat->cstartbs; 2414 a->cendbs = oldmat->cendbs; 2415 2416 /* hash table stuff */ 2417 a->ht = 0; 2418 a->hd = 0; 2419 a->ht_size = 0; 2420 a->ht_flag = oldmat->ht_flag; 2421 a->ht_fact = oldmat->ht_fact; 2422 a->ht_total_ct = 0; 2423 a->ht_insert_ct = 0; 2424 2425 ierr = PetscMemcpy(a->rangebs,oldmat->rangebs,(a->size+2)*sizeof(PetscInt));CHKERRQ(ierr); 2426 if (oldmat->colmap) { 2427 #if defined(PETSC_USE_CTABLE) 2428 ierr = PetscTableCreateCopy(oldmat->colmap,&a->colmap);CHKERRQ(ierr); 2429 #else 2430 ierr = PetscMalloc1(a->Nbs,&a->colmap);CHKERRQ(ierr); 2431 ierr = PetscLogObjectMemory((PetscObject)mat,(a->Nbs)*sizeof(PetscInt));CHKERRQ(ierr); 2432 ierr = PetscMemcpy(a->colmap,oldmat->colmap,(a->Nbs)*sizeof(PetscInt));CHKERRQ(ierr); 2433 #endif 2434 } else a->colmap = 0; 2435 2436 if (oldmat->garray && (len = ((Mat_SeqBAIJ*)(oldmat->B->data))->nbs)) { 2437 ierr = PetscMalloc1(len,&a->garray);CHKERRQ(ierr); 2438 ierr = PetscLogObjectMemory((PetscObject)mat,len*sizeof(PetscInt));CHKERRQ(ierr); 2439 ierr = PetscMemcpy(a->garray,oldmat->garray,len*sizeof(PetscInt));CHKERRQ(ierr); 2440 } else a->garray = 0; 2441 2442 ierr = MatStashCreate_Private(PetscObjectComm((PetscObject)matin),matin->rmap->bs,&mat->bstash);CHKERRQ(ierr); 2443 ierr = VecDuplicate(oldmat->lvec,&a->lvec);CHKERRQ(ierr); 2444 ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)a->lvec);CHKERRQ(ierr); 2445 ierr = VecScatterCopy(oldmat->Mvctx,&a->Mvctx);CHKERRQ(ierr); 2446 ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)a->Mvctx);CHKERRQ(ierr); 2447 2448 ierr = VecDuplicate(oldmat->slvec0,&a->slvec0);CHKERRQ(ierr); 2449 ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)a->slvec0);CHKERRQ(ierr); 2450 ierr = VecDuplicate(oldmat->slvec1,&a->slvec1);CHKERRQ(ierr); 2451 ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)a->slvec1);CHKERRQ(ierr); 2452 2453 ierr = VecGetLocalSize(a->slvec1,&nt);CHKERRQ(ierr); 2454 ierr = VecGetArray(a->slvec1,&array);CHKERRQ(ierr); 2455 ierr = VecCreateSeqWithArray(PETSC_COMM_SELF,1,bs*mbs,array,&a->slvec1a);CHKERRQ(ierr); 2456 ierr = VecCreateSeqWithArray(PETSC_COMM_SELF,1,nt-bs*mbs,array+bs*mbs,&a->slvec1b);CHKERRQ(ierr); 2457 ierr = VecRestoreArray(a->slvec1,&array);CHKERRQ(ierr); 2458 ierr = VecGetArray(a->slvec0,&array);CHKERRQ(ierr); 2459 ierr = VecCreateSeqWithArray(PETSC_COMM_SELF,1,nt-bs*mbs,array+bs*mbs,&a->slvec0b);CHKERRQ(ierr); 2460 ierr = VecRestoreArray(a->slvec0,&array);CHKERRQ(ierr); 2461 ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)a->slvec0);CHKERRQ(ierr); 2462 ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)a->slvec1);CHKERRQ(ierr); 2463 ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)a->slvec0b);CHKERRQ(ierr); 2464 ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)a->slvec1a);CHKERRQ(ierr); 2465 ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)a->slvec1b);CHKERRQ(ierr); 2466 2467 /* ierr = VecScatterCopy(oldmat->sMvctx,&a->sMvctx); - not written yet, replaced by the lazy trick: */ 2468 ierr = PetscObjectReference((PetscObject)oldmat->sMvctx);CHKERRQ(ierr); 2469 a->sMvctx = oldmat->sMvctx; 2470 ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)a->sMvctx);CHKERRQ(ierr); 2471 2472 ierr = MatDuplicate(oldmat->A,cpvalues,&a->A);CHKERRQ(ierr); 2473 ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)a->A);CHKERRQ(ierr); 2474 ierr = MatDuplicate(oldmat->B,cpvalues,&a->B);CHKERRQ(ierr); 2475 ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)a->B);CHKERRQ(ierr); 2476 ierr = PetscFunctionListDuplicate(((PetscObject)matin)->qlist,&((PetscObject)mat)->qlist);CHKERRQ(ierr); 2477 *newmat = mat; 2478 PetscFunctionReturn(0); 2479 } 2480 2481 #undef __FUNCT__ 2482 #define __FUNCT__ "MatLoad_MPISBAIJ" 2483 PetscErrorCode MatLoad_MPISBAIJ(Mat newmat,PetscViewer viewer) 2484 { 2485 PetscErrorCode ierr; 2486 PetscInt i,nz,j,rstart,rend; 2487 PetscScalar *vals,*buf; 2488 MPI_Comm comm; 2489 MPI_Status status; 2490 PetscMPIInt rank,size,tag = ((PetscObject)viewer)->tag,*sndcounts = 0,*browners,maxnz,*rowners,mmbs; 2491 PetscInt header[4],*rowlengths = 0,M,N,m,*cols,*locrowlens; 2492 PetscInt *procsnz = 0,jj,*mycols,*ibuf; 2493 PetscInt bs = newmat->rmap->bs,Mbs,mbs,extra_rows; 2494 PetscInt *dlens,*odlens,*mask,*masked1,*masked2,rowcount,odcount; 2495 PetscInt dcount,kmax,k,nzcount,tmp,sizesset=1,grows,gcols; 2496 int fd; 2497 2498 PetscFunctionBegin; 2499 /* force binary viewer to load .info file if it has not yet done so */ 2500 ierr = PetscViewerSetUp(viewer);CHKERRQ(ierr); 2501 ierr = PetscObjectGetComm((PetscObject)viewer,&comm);CHKERRQ(ierr); 2502 ierr = PetscOptionsBegin(comm,NULL,"Options for loading MPISBAIJ matrix 2","Mat");CHKERRQ(ierr); 2503 ierr = PetscOptionsInt("-matload_block_size","Set the blocksize used to store the matrix","MatLoad",bs,&bs,NULL);CHKERRQ(ierr); 2504 ierr = PetscOptionsEnd();CHKERRQ(ierr); 2505 if (bs < 0) bs = 1; 2506 2507 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 2508 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 2509 ierr = PetscViewerBinaryGetDescriptor(viewer,&fd);CHKERRQ(ierr); 2510 if (!rank) { 2511 ierr = PetscBinaryRead(fd,(char*)header,4,PETSC_INT);CHKERRQ(ierr); 2512 if (header[0] != MAT_FILE_CLASSID) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED,"not matrix object"); 2513 if (header[3] < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED,"Matrix stored in special format, cannot load as MPISBAIJ"); 2514 } 2515 2516 if (newmat->rmap->n < 0 && newmat->rmap->N < 0 && newmat->cmap->n < 0 && newmat->cmap->N < 0) sizesset = 0; 2517 2518 ierr = MPI_Bcast(header+1,3,MPIU_INT,0,comm);CHKERRQ(ierr); 2519 M = header[1]; 2520 N = header[2]; 2521 2522 /* If global rows/cols are set to PETSC_DECIDE, set it to the sizes given in the file */ 2523 if (sizesset && newmat->rmap->N < 0) newmat->rmap->N = M; 2524 if (sizesset && newmat->cmap->N < 0) newmat->cmap->N = N; 2525 2526 /* If global sizes are set, check if they are consistent with that given in the file */ 2527 if (sizesset) { 2528 ierr = MatGetSize(newmat,&grows,&gcols);CHKERRQ(ierr); 2529 } 2530 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); 2531 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); 2532 2533 if (M != N) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Can only do square matrices"); 2534 2535 /* 2536 This code adds extra rows to make sure the number of rows is 2537 divisible by the blocksize 2538 */ 2539 Mbs = M/bs; 2540 extra_rows = bs - M + bs*(Mbs); 2541 if (extra_rows == bs) extra_rows = 0; 2542 else Mbs++; 2543 if (extra_rows &&!rank) { 2544 ierr = PetscInfo(viewer,"Padding loaded matrix to match blocksize\n");CHKERRQ(ierr); 2545 } 2546 2547 /* determine ownership of all rows */ 2548 if (newmat->rmap->n < 0) { /* PETSC_DECIDE */ 2549 mbs = Mbs/size + ((Mbs % size) > rank); 2550 m = mbs*bs; 2551 } else { /* User Set */ 2552 m = newmat->rmap->n; 2553 mbs = m/bs; 2554 } 2555 ierr = PetscMalloc2(size+1,&rowners,size+1,&browners);CHKERRQ(ierr); 2556 ierr = PetscMPIIntCast(mbs,&mmbs);CHKERRQ(ierr); 2557 ierr = MPI_Allgather(&mmbs,1,MPI_INT,rowners+1,1,MPI_INT,comm);CHKERRQ(ierr); 2558 rowners[0] = 0; 2559 for (i=2; i<=size; i++) rowners[i] += rowners[i-1]; 2560 for (i=0; i<=size; i++) browners[i] = rowners[i]*bs; 2561 rstart = rowners[rank]; 2562 rend = rowners[rank+1]; 2563 2564 /* distribute row lengths to all processors */ 2565 ierr = PetscMalloc1((rend-rstart)*bs,&locrowlens);CHKERRQ(ierr); 2566 if (!rank) { 2567 ierr = PetscMalloc1(M+extra_rows,&rowlengths);CHKERRQ(ierr); 2568 ierr = PetscBinaryRead(fd,rowlengths,M,PETSC_INT);CHKERRQ(ierr); 2569 for (i=0; i<extra_rows; i++) rowlengths[M+i] = 1; 2570 ierr = PetscMalloc1(size,&sndcounts);CHKERRQ(ierr); 2571 for (i=0; i<size; i++) sndcounts[i] = browners[i+1] - browners[i]; 2572 ierr = MPI_Scatterv(rowlengths,sndcounts,browners,MPIU_INT,locrowlens,(rend-rstart)*bs,MPIU_INT,0,comm);CHKERRQ(ierr); 2573 ierr = PetscFree(sndcounts);CHKERRQ(ierr); 2574 } else { 2575 ierr = MPI_Scatterv(0,0,0,MPIU_INT,locrowlens,(rend-rstart)*bs,MPIU_INT,0,comm);CHKERRQ(ierr); 2576 } 2577 2578 if (!rank) { /* procs[0] */ 2579 /* calculate the number of nonzeros on each processor */ 2580 ierr = PetscMalloc1(size,&procsnz);CHKERRQ(ierr); 2581 ierr = PetscMemzero(procsnz,size*sizeof(PetscInt));CHKERRQ(ierr); 2582 for (i=0; i<size; i++) { 2583 for (j=rowners[i]*bs; j< rowners[i+1]*bs; j++) { 2584 procsnz[i] += rowlengths[j]; 2585 } 2586 } 2587 ierr = PetscFree(rowlengths);CHKERRQ(ierr); 2588 2589 /* determine max buffer needed and allocate it */ 2590 maxnz = 0; 2591 for (i=0; i<size; i++) { 2592 maxnz = PetscMax(maxnz,procsnz[i]); 2593 } 2594 ierr = PetscMalloc1(maxnz,&cols);CHKERRQ(ierr); 2595 2596 /* read in my part of the matrix column indices */ 2597 nz = procsnz[0]; 2598 ierr = PetscMalloc1(nz,&ibuf);CHKERRQ(ierr); 2599 mycols = ibuf; 2600 if (size == 1) nz -= extra_rows; 2601 ierr = PetscBinaryRead(fd,mycols,nz,PETSC_INT);CHKERRQ(ierr); 2602 if (size == 1) { 2603 for (i=0; i< extra_rows; i++) mycols[nz+i] = M+i; 2604 } 2605 2606 /* read in every ones (except the last) and ship off */ 2607 for (i=1; i<size-1; i++) { 2608 nz = procsnz[i]; 2609 ierr = PetscBinaryRead(fd,cols,nz,PETSC_INT);CHKERRQ(ierr); 2610 ierr = MPI_Send(cols,nz,MPIU_INT,i,tag,comm);CHKERRQ(ierr); 2611 } 2612 /* read in the stuff for the last proc */ 2613 if (size != 1) { 2614 nz = procsnz[size-1] - extra_rows; /* the extra rows are not on the disk */ 2615 ierr = PetscBinaryRead(fd,cols,nz,PETSC_INT);CHKERRQ(ierr); 2616 for (i=0; i<extra_rows; i++) cols[nz+i] = M+i; 2617 ierr = MPI_Send(cols,nz+extra_rows,MPIU_INT,size-1,tag,comm);CHKERRQ(ierr); 2618 } 2619 ierr = PetscFree(cols);CHKERRQ(ierr); 2620 } else { /* procs[i], i>0 */ 2621 /* determine buffer space needed for message */ 2622 nz = 0; 2623 for (i=0; i<m; i++) nz += locrowlens[i]; 2624 ierr = PetscMalloc1(nz,&ibuf);CHKERRQ(ierr); 2625 mycols = ibuf; 2626 /* receive message of column indices*/ 2627 ierr = MPI_Recv(mycols,nz,MPIU_INT,0,tag,comm,&status);CHKERRQ(ierr); 2628 ierr = MPI_Get_count(&status,MPIU_INT,&maxnz);CHKERRQ(ierr); 2629 if (maxnz != nz) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED,"something is wrong with file"); 2630 } 2631 2632 /* loop over local rows, determining number of off diagonal entries */ 2633 ierr = PetscMalloc2(rend-rstart,&dlens,rend-rstart,&odlens);CHKERRQ(ierr); 2634 ierr = PetscMalloc3(Mbs,&mask,Mbs,&masked1,Mbs,&masked2);CHKERRQ(ierr); 2635 ierr = PetscMemzero(mask,Mbs*sizeof(PetscInt));CHKERRQ(ierr); 2636 ierr = PetscMemzero(masked1,Mbs*sizeof(PetscInt));CHKERRQ(ierr); 2637 ierr = PetscMemzero(masked2,Mbs*sizeof(PetscInt));CHKERRQ(ierr); 2638 rowcount = 0; 2639 nzcount = 0; 2640 for (i=0; i<mbs; i++) { 2641 dcount = 0; 2642 odcount = 0; 2643 for (j=0; j<bs; j++) { 2644 kmax = locrowlens[rowcount]; 2645 for (k=0; k<kmax; k++) { 2646 tmp = mycols[nzcount++]/bs; /* block col. index */ 2647 if (!mask[tmp]) { 2648 mask[tmp] = 1; 2649 if (tmp < rstart || tmp >= rend) masked2[odcount++] = tmp; /* entry in off-diag portion */ 2650 else masked1[dcount++] = tmp; /* entry in diag portion */ 2651 } 2652 } 2653 rowcount++; 2654 } 2655 2656 dlens[i] = dcount; /* d_nzz[i] */ 2657 odlens[i] = odcount; /* o_nzz[i] */ 2658 2659 /* zero out the mask elements we set */ 2660 for (j=0; j<dcount; j++) mask[masked1[j]] = 0; 2661 for (j=0; j<odcount; j++) mask[masked2[j]] = 0; 2662 } 2663 if (!sizesset) { 2664 ierr = MatSetSizes(newmat,m,m,M+extra_rows,N+extra_rows);CHKERRQ(ierr); 2665 } 2666 ierr = MatMPISBAIJSetPreallocation(newmat,bs,0,dlens,0,odlens);CHKERRQ(ierr); 2667 ierr = MatSetOption(newmat,MAT_IGNORE_LOWER_TRIANGULAR,PETSC_TRUE);CHKERRQ(ierr); 2668 2669 if (!rank) { 2670 ierr = PetscMalloc1(maxnz,&buf);CHKERRQ(ierr); 2671 /* read in my part of the matrix numerical values */ 2672 nz = procsnz[0]; 2673 vals = buf; 2674 mycols = ibuf; 2675 if (size == 1) nz -= extra_rows; 2676 ierr = PetscBinaryRead(fd,vals,nz,PETSC_SCALAR);CHKERRQ(ierr); 2677 if (size == 1) { 2678 for (i=0; i< extra_rows; i++) vals[nz+i] = 1.0; 2679 } 2680 2681 /* insert into matrix */ 2682 jj = rstart*bs; 2683 for (i=0; i<m; i++) { 2684 ierr = MatSetValues(newmat,1,&jj,locrowlens[i],mycols,vals,INSERT_VALUES);CHKERRQ(ierr); 2685 mycols += locrowlens[i]; 2686 vals += locrowlens[i]; 2687 jj++; 2688 } 2689 2690 /* read in other processors (except the last one) and ship out */ 2691 for (i=1; i<size-1; i++) { 2692 nz = procsnz[i]; 2693 vals = buf; 2694 ierr = PetscBinaryRead(fd,vals,nz,PETSC_SCALAR);CHKERRQ(ierr); 2695 ierr = MPI_Send(vals,nz,MPIU_SCALAR,i,((PetscObject)newmat)->tag,comm);CHKERRQ(ierr); 2696 } 2697 /* the last proc */ 2698 if (size != 1) { 2699 nz = procsnz[i] - extra_rows; 2700 vals = buf; 2701 ierr = PetscBinaryRead(fd,vals,nz,PETSC_SCALAR);CHKERRQ(ierr); 2702 for (i=0; i<extra_rows; i++) vals[nz+i] = 1.0; 2703 ierr = MPI_Send(vals,nz+extra_rows,MPIU_SCALAR,size-1,((PetscObject)newmat)->tag,comm);CHKERRQ(ierr); 2704 } 2705 ierr = PetscFree(procsnz);CHKERRQ(ierr); 2706 2707 } else { 2708 /* receive numeric values */ 2709 ierr = PetscMalloc1(nz,&buf);CHKERRQ(ierr); 2710 2711 /* receive message of values*/ 2712 vals = buf; 2713 mycols = ibuf; 2714 ierr = MPI_Recv(vals,nz,MPIU_SCALAR,0,((PetscObject)newmat)->tag,comm,&status);CHKERRQ(ierr); 2715 ierr = MPI_Get_count(&status,MPIU_SCALAR,&maxnz);CHKERRQ(ierr); 2716 if (maxnz != nz) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED,"something is wrong with file"); 2717 2718 /* insert into matrix */ 2719 jj = rstart*bs; 2720 for (i=0; i<m; i++) { 2721 ierr = MatSetValues_MPISBAIJ(newmat,1,&jj,locrowlens[i],mycols,vals,INSERT_VALUES);CHKERRQ(ierr); 2722 mycols += locrowlens[i]; 2723 vals += locrowlens[i]; 2724 jj++; 2725 } 2726 } 2727 2728 ierr = PetscFree(locrowlens);CHKERRQ(ierr); 2729 ierr = PetscFree(buf);CHKERRQ(ierr); 2730 ierr = PetscFree(ibuf);CHKERRQ(ierr); 2731 ierr = PetscFree2(rowners,browners);CHKERRQ(ierr); 2732 ierr = PetscFree2(dlens,odlens);CHKERRQ(ierr); 2733 ierr = PetscFree3(mask,masked1,masked2);CHKERRQ(ierr); 2734 ierr = MatAssemblyBegin(newmat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2735 ierr = MatAssemblyEnd(newmat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2736 PetscFunctionReturn(0); 2737 } 2738 2739 #undef __FUNCT__ 2740 #define __FUNCT__ "MatMPISBAIJSetHashTableFactor" 2741 /*XXXXX@ 2742 MatMPISBAIJSetHashTableFactor - Sets the factor required to compute the size of the HashTable. 2743 2744 Input Parameters: 2745 . mat - the matrix 2746 . fact - factor 2747 2748 Not Collective on Mat, each process can have a different hash factor 2749 2750 Level: advanced 2751 2752 Notes: 2753 This can also be set by the command line option: -mat_use_hash_table fact 2754 2755 .keywords: matrix, hashtable, factor, HT 2756 2757 .seealso: MatSetOption() 2758 @XXXXX*/ 2759 2760 2761 #undef __FUNCT__ 2762 #define __FUNCT__ "MatGetRowMaxAbs_MPISBAIJ" 2763 PetscErrorCode MatGetRowMaxAbs_MPISBAIJ(Mat A,Vec v,PetscInt idx[]) 2764 { 2765 Mat_MPISBAIJ *a = (Mat_MPISBAIJ*)A->data; 2766 Mat_SeqBAIJ *b = (Mat_SeqBAIJ*)(a->B)->data; 2767 PetscReal atmp; 2768 PetscReal *work,*svalues,*rvalues; 2769 PetscErrorCode ierr; 2770 PetscInt i,bs,mbs,*bi,*bj,brow,j,ncols,krow,kcol,col,row,Mbs,bcol; 2771 PetscMPIInt rank,size; 2772 PetscInt *rowners_bs,dest,count,source; 2773 PetscScalar *va; 2774 MatScalar *ba; 2775 MPI_Status stat; 2776 2777 PetscFunctionBegin; 2778 if (idx) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Send email to petsc-maint@mcs.anl.gov"); 2779 ierr = MatGetRowMaxAbs(a->A,v,NULL);CHKERRQ(ierr); 2780 ierr = VecGetArray(v,&va);CHKERRQ(ierr); 2781 2782 ierr = MPI_Comm_size(PetscObjectComm((PetscObject)A),&size);CHKERRQ(ierr); 2783 ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)A),&rank);CHKERRQ(ierr); 2784 2785 bs = A->rmap->bs; 2786 mbs = a->mbs; 2787 Mbs = a->Mbs; 2788 ba = b->a; 2789 bi = b->i; 2790 bj = b->j; 2791 2792 /* find ownerships */ 2793 rowners_bs = A->rmap->range; 2794 2795 /* each proc creates an array to be distributed */ 2796 ierr = PetscMalloc1(bs*Mbs,&work);CHKERRQ(ierr); 2797 ierr = PetscMemzero(work,bs*Mbs*sizeof(PetscReal));CHKERRQ(ierr); 2798 2799 /* row_max for B */ 2800 if (rank != size-1) { 2801 for (i=0; i<mbs; i++) { 2802 ncols = bi[1] - bi[0]; bi++; 2803 brow = bs*i; 2804 for (j=0; j<ncols; j++) { 2805 bcol = bs*(*bj); 2806 for (kcol=0; kcol<bs; kcol++) { 2807 col = bcol + kcol; /* local col index */ 2808 col += rowners_bs[rank+1]; /* global col index */ 2809 for (krow=0; krow<bs; krow++) { 2810 atmp = PetscAbsScalar(*ba); ba++; 2811 row = brow + krow; /* local row index */ 2812 if (PetscRealPart(va[row]) < atmp) va[row] = atmp; 2813 if (work[col] < atmp) work[col] = atmp; 2814 } 2815 } 2816 bj++; 2817 } 2818 } 2819 2820 /* send values to its owners */ 2821 for (dest=rank+1; dest<size; dest++) { 2822 svalues = work + rowners_bs[dest]; 2823 count = rowners_bs[dest+1]-rowners_bs[dest]; 2824 ierr = MPI_Send(svalues,count,MPIU_REAL,dest,rank,PetscObjectComm((PetscObject)A));CHKERRQ(ierr); 2825 } 2826 } 2827 2828 /* receive values */ 2829 if (rank) { 2830 rvalues = work; 2831 count = rowners_bs[rank+1]-rowners_bs[rank]; 2832 for (source=0; source<rank; source++) { 2833 ierr = MPI_Recv(rvalues,count,MPIU_REAL,MPI_ANY_SOURCE,MPI_ANY_TAG,PetscObjectComm((PetscObject)A),&stat);CHKERRQ(ierr); 2834 /* process values */ 2835 for (i=0; i<count; i++) { 2836 if (PetscRealPart(va[i]) < rvalues[i]) va[i] = rvalues[i]; 2837 } 2838 } 2839 } 2840 2841 ierr = VecRestoreArray(v,&va);CHKERRQ(ierr); 2842 ierr = PetscFree(work);CHKERRQ(ierr); 2843 PetscFunctionReturn(0); 2844 } 2845 2846 #undef __FUNCT__ 2847 #define __FUNCT__ "MatSOR_MPISBAIJ" 2848 PetscErrorCode MatSOR_MPISBAIJ(Mat matin,Vec bb,PetscReal omega,MatSORType flag,PetscReal fshift,PetscInt its,PetscInt lits,Vec xx) 2849 { 2850 Mat_MPISBAIJ *mat = (Mat_MPISBAIJ*)matin->data; 2851 PetscErrorCode ierr; 2852 PetscInt mbs=mat->mbs,bs=matin->rmap->bs; 2853 PetscScalar *x,*ptr,*from; 2854 Vec bb1; 2855 const PetscScalar *b; 2856 2857 PetscFunctionBegin; 2858 if (its <= 0 || lits <= 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Relaxation requires global its %D and local its %D both positive",its,lits); 2859 if (bs > 1) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"SSOR for block size > 1 is not yet implemented"); 2860 2861 if (flag == SOR_APPLY_UPPER) { 2862 ierr = (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);CHKERRQ(ierr); 2863 PetscFunctionReturn(0); 2864 } 2865 2866 if ((flag & SOR_LOCAL_SYMMETRIC_SWEEP) == SOR_LOCAL_SYMMETRIC_SWEEP) { 2867 if (flag & SOR_ZERO_INITIAL_GUESS) { 2868 ierr = (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,lits,xx);CHKERRQ(ierr); 2869 its--; 2870 } 2871 2872 ierr = VecDuplicate(bb,&bb1);CHKERRQ(ierr); 2873 while (its--) { 2874 2875 /* lower triangular part: slvec0b = - B^T*xx */ 2876 ierr = (*mat->B->ops->multtranspose)(mat->B,xx,mat->slvec0b);CHKERRQ(ierr); 2877 2878 /* copy xx into slvec0a */ 2879 ierr = VecGetArray(mat->slvec0,&ptr);CHKERRQ(ierr); 2880 ierr = VecGetArray(xx,&x);CHKERRQ(ierr); 2881 ierr = PetscMemcpy(ptr,x,bs*mbs*sizeof(MatScalar));CHKERRQ(ierr); 2882 ierr = VecRestoreArray(mat->slvec0,&ptr);CHKERRQ(ierr); 2883 2884 ierr = VecScale(mat->slvec0,-1.0);CHKERRQ(ierr); 2885 2886 /* copy bb into slvec1a */ 2887 ierr = VecGetArray(mat->slvec1,&ptr);CHKERRQ(ierr); 2888 ierr = VecGetArrayRead(bb,&b);CHKERRQ(ierr); 2889 ierr = PetscMemcpy(ptr,b,bs*mbs*sizeof(MatScalar));CHKERRQ(ierr); 2890 ierr = VecRestoreArray(mat->slvec1,&ptr);CHKERRQ(ierr); 2891 2892 /* set slvec1b = 0 */ 2893 ierr = VecSet(mat->slvec1b,0.0);CHKERRQ(ierr); 2894 2895 ierr = VecScatterBegin(mat->sMvctx,mat->slvec0,mat->slvec1,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2896 ierr = VecRestoreArray(xx,&x);CHKERRQ(ierr); 2897 ierr = VecRestoreArrayRead(bb,&b);CHKERRQ(ierr); 2898 ierr = VecScatterEnd(mat->sMvctx,mat->slvec0,mat->slvec1,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2899 2900 /* upper triangular part: bb1 = bb1 - B*x */ 2901 ierr = (*mat->B->ops->multadd)(mat->B,mat->slvec1b,mat->slvec1a,bb1);CHKERRQ(ierr); 2902 2903 /* local diagonal sweep */ 2904 ierr = (*mat->A->ops->sor)(mat->A,bb1,omega,SOR_SYMMETRIC_SWEEP,fshift,lits,lits,xx);CHKERRQ(ierr); 2905 } 2906 ierr = VecDestroy(&bb1);CHKERRQ(ierr); 2907 } else if ((flag & SOR_LOCAL_FORWARD_SWEEP) && (its == 1) && (flag & SOR_ZERO_INITIAL_GUESS)) { 2908 ierr = (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);CHKERRQ(ierr); 2909 } else if ((flag & SOR_LOCAL_BACKWARD_SWEEP) && (its == 1) && (flag & SOR_ZERO_INITIAL_GUESS)) { 2910 ierr = (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);CHKERRQ(ierr); 2911 } else if (flag & SOR_EISENSTAT) { 2912 Vec xx1; 2913 PetscBool hasop; 2914 const PetscScalar *diag; 2915 PetscScalar *sl,scale = (omega - 2.0)/omega; 2916 PetscInt i,n; 2917 2918 if (!mat->xx1) { 2919 ierr = VecDuplicate(bb,&mat->xx1);CHKERRQ(ierr); 2920 ierr = VecDuplicate(bb,&mat->bb1);CHKERRQ(ierr); 2921 } 2922 xx1 = mat->xx1; 2923 bb1 = mat->bb1; 2924 2925 ierr = (*mat->A->ops->sor)(mat->A,bb,omega,(MatSORType)(SOR_ZERO_INITIAL_GUESS | SOR_LOCAL_BACKWARD_SWEEP),fshift,lits,1,xx);CHKERRQ(ierr); 2926 2927 if (!mat->diag) { 2928 /* this is wrong for same matrix with new nonzero values */ 2929 ierr = MatCreateVecs(matin,&mat->diag,NULL);CHKERRQ(ierr); 2930 ierr = MatGetDiagonal(matin,mat->diag);CHKERRQ(ierr); 2931 } 2932 ierr = MatHasOperation(matin,MATOP_MULT_DIAGONAL_BLOCK,&hasop);CHKERRQ(ierr); 2933 2934 if (hasop) { 2935 ierr = MatMultDiagonalBlock(matin,xx,bb1);CHKERRQ(ierr); 2936 ierr = VecAYPX(mat->slvec1a,scale,bb);CHKERRQ(ierr); 2937 } else { 2938 /* 2939 These two lines are replaced by code that may be a bit faster for a good compiler 2940 ierr = VecPointwiseMult(mat->slvec1a,mat->diag,xx);CHKERRQ(ierr); 2941 ierr = VecAYPX(mat->slvec1a,scale,bb);CHKERRQ(ierr); 2942 */ 2943 ierr = VecGetArray(mat->slvec1a,&sl);CHKERRQ(ierr); 2944 ierr = VecGetArrayRead(mat->diag,&diag);CHKERRQ(ierr); 2945 ierr = VecGetArrayRead(bb,&b);CHKERRQ(ierr); 2946 ierr = VecGetArray(xx,&x);CHKERRQ(ierr); 2947 ierr = VecGetLocalSize(xx,&n);CHKERRQ(ierr); 2948 if (omega == 1.0) { 2949 for (i=0; i<n; i++) sl[i] = b[i] - diag[i]*x[i]; 2950 ierr = PetscLogFlops(2.0*n);CHKERRQ(ierr); 2951 } else { 2952 for (i=0; i<n; i++) sl[i] = b[i] + scale*diag[i]*x[i]; 2953 ierr = PetscLogFlops(3.0*n);CHKERRQ(ierr); 2954 } 2955 ierr = VecRestoreArray(mat->slvec1a,&sl);CHKERRQ(ierr); 2956 ierr = VecRestoreArrayRead(mat->diag,&diag);CHKERRQ(ierr); 2957 ierr = VecRestoreArrayRead(bb,&b);CHKERRQ(ierr); 2958 ierr = VecRestoreArray(xx,&x);CHKERRQ(ierr); 2959 } 2960 2961 /* multiply off-diagonal portion of matrix */ 2962 ierr = VecSet(mat->slvec1b,0.0);CHKERRQ(ierr); 2963 ierr = (*mat->B->ops->multtranspose)(mat->B,xx,mat->slvec0b);CHKERRQ(ierr); 2964 ierr = VecGetArray(mat->slvec0,&from);CHKERRQ(ierr); 2965 ierr = VecGetArray(xx,&x);CHKERRQ(ierr); 2966 ierr = PetscMemcpy(from,x,bs*mbs*sizeof(MatScalar));CHKERRQ(ierr); 2967 ierr = VecRestoreArray(mat->slvec0,&from);CHKERRQ(ierr); 2968 ierr = VecRestoreArray(xx,&x);CHKERRQ(ierr); 2969 ierr = VecScatterBegin(mat->sMvctx,mat->slvec0,mat->slvec1,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2970 ierr = VecScatterEnd(mat->sMvctx,mat->slvec0,mat->slvec1,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2971 ierr = (*mat->B->ops->multadd)(mat->B,mat->slvec1b,mat->slvec1a,mat->slvec1a);CHKERRQ(ierr); 2972 2973 /* local sweep */ 2974 ierr = (*mat->A->ops->sor)(mat->A,mat->slvec1a,omega,(MatSORType)(SOR_ZERO_INITIAL_GUESS | SOR_LOCAL_FORWARD_SWEEP),fshift,lits,1,xx1);CHKERRQ(ierr); 2975 ierr = VecAXPY(xx,1.0,xx1);CHKERRQ(ierr); 2976 } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"MatSORType is not supported for SBAIJ matrix format"); 2977 PetscFunctionReturn(0); 2978 } 2979 2980 #undef __FUNCT__ 2981 #define __FUNCT__ "MatSOR_MPISBAIJ_2comm" 2982 PetscErrorCode MatSOR_MPISBAIJ_2comm(Mat matin,Vec bb,PetscReal omega,MatSORType flag,PetscReal fshift,PetscInt its,PetscInt lits,Vec xx) 2983 { 2984 Mat_MPISBAIJ *mat = (Mat_MPISBAIJ*)matin->data; 2985 PetscErrorCode ierr; 2986 Vec lvec1,bb1; 2987 2988 PetscFunctionBegin; 2989 if (its <= 0 || lits <= 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Relaxation requires global its %D and local its %D both positive",its,lits); 2990 if (matin->rmap->bs > 1) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"SSOR for block size > 1 is not yet implemented"); 2991 2992 if ((flag & SOR_LOCAL_SYMMETRIC_SWEEP) == SOR_LOCAL_SYMMETRIC_SWEEP) { 2993 if (flag & SOR_ZERO_INITIAL_GUESS) { 2994 ierr = (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,lits,xx);CHKERRQ(ierr); 2995 its--; 2996 } 2997 2998 ierr = VecDuplicate(mat->lvec,&lvec1);CHKERRQ(ierr); 2999 ierr = VecDuplicate(bb,&bb1);CHKERRQ(ierr); 3000 while (its--) { 3001 ierr = VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3002 3003 /* lower diagonal part: bb1 = bb - B^T*xx */ 3004 ierr = (*mat->B->ops->multtranspose)(mat->B,xx,lvec1);CHKERRQ(ierr); 3005 ierr = VecScale(lvec1,-1.0);CHKERRQ(ierr); 3006 3007 ierr = VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 3008 ierr = VecCopy(bb,bb1);CHKERRQ(ierr); 3009 ierr = VecScatterBegin(mat->Mvctx,lvec1,bb1,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3010 3011 /* upper diagonal part: bb1 = bb1 - B*x */ 3012 ierr = VecScale(mat->lvec,-1.0);CHKERRQ(ierr); 3013 ierr = (*mat->B->ops->multadd)(mat->B,mat->lvec,bb1,bb1);CHKERRQ(ierr); 3014 3015 ierr = VecScatterEnd(mat->Mvctx,lvec1,bb1,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 3016 3017 /* diagonal sweep */ 3018 ierr = (*mat->A->ops->sor)(mat->A,bb1,omega,SOR_SYMMETRIC_SWEEP,fshift,lits,lits,xx);CHKERRQ(ierr); 3019 } 3020 ierr = VecDestroy(&lvec1);CHKERRQ(ierr); 3021 ierr = VecDestroy(&bb1);CHKERRQ(ierr); 3022 } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"MatSORType is not supported for SBAIJ matrix format"); 3023 PetscFunctionReturn(0); 3024 } 3025 3026 #undef __FUNCT__ 3027 #define __FUNCT__ "MatCreateMPISBAIJWithArrays" 3028 /*@ 3029 MatCreateMPISBAIJWithArrays - creates a MPI SBAIJ matrix using arrays that contain in standard 3030 CSR format the local rows. 3031 3032 Collective on MPI_Comm 3033 3034 Input Parameters: 3035 + comm - MPI communicator 3036 . bs - the block size, only a block size of 1 is supported 3037 . m - number of local rows (Cannot be PETSC_DECIDE) 3038 . n - This value should be the same as the local size used in creating the 3039 x vector for the matrix-vector product y = Ax. (or PETSC_DECIDE to have 3040 calculated if N is given) For square matrices n is almost always m. 3041 . M - number of global rows (or PETSC_DETERMINE to have calculated if m is given) 3042 . N - number of global columns (or PETSC_DETERMINE to have calculated if n is given) 3043 . i - row indices 3044 . j - column indices 3045 - a - matrix values 3046 3047 Output Parameter: 3048 . mat - the matrix 3049 3050 Level: intermediate 3051 3052 Notes: 3053 The i, j, and a arrays ARE copied by this routine into the internal format used by PETSc; 3054 thus you CANNOT change the matrix entries by changing the values of a[] after you have 3055 called this routine. Use MatCreateMPIAIJWithSplitArrays() to avoid needing to copy the arrays. 3056 3057 The i and j indices are 0 based, and i indices are indices corresponding to the local j array. 3058 3059 .keywords: matrix, aij, compressed row, sparse, parallel 3060 3061 .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatMPIAIJSetPreallocationCSR(), 3062 MPIAIJ, MatCreateAIJ(), MatCreateMPIAIJWithSplitArrays() 3063 @*/ 3064 PetscErrorCode MatCreateMPISBAIJWithArrays(MPI_Comm comm,PetscInt bs,PetscInt m,PetscInt n,PetscInt M,PetscInt N,const PetscInt i[],const PetscInt j[],const PetscScalar a[],Mat *mat) 3065 { 3066 PetscErrorCode ierr; 3067 3068 3069 PetscFunctionBegin; 3070 if (i[0]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"i (row indices) must start with 0"); 3071 if (m < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"local number of rows (m) cannot be PETSC_DECIDE, or negative"); 3072 ierr = MatCreate(comm,mat);CHKERRQ(ierr); 3073 ierr = MatSetSizes(*mat,m,n,M,N);CHKERRQ(ierr); 3074 ierr = MatSetType(*mat,MATMPISBAIJ);CHKERRQ(ierr); 3075 ierr = MatMPISBAIJSetPreallocationCSR(*mat,bs,i,j,a);CHKERRQ(ierr); 3076 PetscFunctionReturn(0); 3077 } 3078 3079 3080 #undef __FUNCT__ 3081 #define __FUNCT__ "MatMPISBAIJSetPreallocationCSR" 3082 /*@C 3083 MatMPISBAIJSetPreallocationCSR - Allocates memory for a sparse parallel matrix in BAIJ format 3084 (the default parallel PETSc format). 3085 3086 Collective on MPI_Comm 3087 3088 Input Parameters: 3089 + B - the matrix 3090 . bs - the block size 3091 . i - the indices into j for the start of each local row (starts with zero) 3092 . j - the column indices for each local row (starts with zero) these must be sorted for each row 3093 - v - optional values in the matrix 3094 3095 Level: developer 3096 3097 .keywords: matrix, aij, compressed row, sparse, parallel 3098 3099 .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIBAIJSetPreallocation(), MatCreateAIJ(), MPIAIJ 3100 @*/ 3101 PetscErrorCode MatMPISBAIJSetPreallocationCSR(Mat B,PetscInt bs,const PetscInt i[],const PetscInt j[], const PetscScalar v[]) 3102 { 3103 PetscErrorCode ierr; 3104 3105 PetscFunctionBegin; 3106 ierr = PetscTryMethod(B,"MatMPISBAIJSetPreallocationCSR_C",(Mat,PetscInt,const PetscInt[],const PetscInt[],const PetscScalar[]),(B,bs,i,j,v));CHKERRQ(ierr); 3107 PetscFunctionReturn(0); 3108 } 3109 3110 #undef __FUNCT__ 3111 #define __FUNCT__ "MatCreateMPIMatConcatenateSeqMat_MPISBAIJ" 3112 PetscErrorCode MatCreateMPIMatConcatenateSeqMat_MPISBAIJ(MPI_Comm comm,Mat inmat,PetscInt n,MatReuse scall,Mat *outmat) 3113 { 3114 PetscErrorCode ierr; 3115 PetscInt m,N,i,rstart,nnz,Ii,bs,cbs; 3116 PetscInt *indx; 3117 PetscScalar *values; 3118 3119 PetscFunctionBegin; 3120 ierr = MatGetSize(inmat,&m,&N);CHKERRQ(ierr); 3121 if (scall == MAT_INITIAL_MATRIX) { /* symbolic phase */ 3122 Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ*)inmat->data; 3123 PetscInt *dnz,*onz,sum,bs,cbs,mbs,Nbs; 3124 PetscInt *bindx,rmax=a->rmax,j; 3125 3126 ierr = MatGetBlockSizes(inmat,&bs,&cbs);CHKERRQ(ierr); 3127 mbs = m/bs; Nbs = N/cbs; 3128 if (n == PETSC_DECIDE) { 3129 ierr = PetscSplitOwnership(comm,&n,&Nbs);CHKERRQ(ierr); 3130 } 3131 /* Check sum(n) = Nbs */ 3132 ierr = MPI_Allreduce(&n,&sum,1,MPIU_INT,MPI_SUM,comm);CHKERRQ(ierr); 3133 if (sum != Nbs) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_INCOMP,"Sum of local columns != global columns %d",Nbs); 3134 3135 ierr = MPI_Scan(&mbs, &rstart,1,MPIU_INT,MPI_SUM,comm);CHKERRQ(ierr); 3136 rstart -= mbs; 3137 3138 ierr = PetscMalloc1(rmax,&bindx);CHKERRQ(ierr); 3139 ierr = MatPreallocateInitialize(comm,mbs,n,dnz,onz);CHKERRQ(ierr); 3140 ierr = MatSetOption(inmat,MAT_GETROW_UPPERTRIANGULAR,PETSC_TRUE);CHKERRQ(ierr); 3141 for (i=0; i<mbs; i++) { 3142 ierr = MatGetRow_SeqSBAIJ(inmat,i*bs,&nnz,&indx,NULL);CHKERRQ(ierr); /* non-blocked nnz and indx */ 3143 nnz = nnz/bs; 3144 for (j=0; j<nnz; j++) bindx[j] = indx[j*bs]/bs; 3145 ierr = MatPreallocateSet(i+rstart,nnz,bindx,dnz,onz);CHKERRQ(ierr); 3146 ierr = MatRestoreRow_SeqSBAIJ(inmat,i*bs,&nnz,&indx,NULL);CHKERRQ(ierr); 3147 } 3148 ierr = MatSetOption(inmat,MAT_GETROW_UPPERTRIANGULAR,PETSC_FALSE);CHKERRQ(ierr); 3149 ierr = PetscFree(bindx);CHKERRQ(ierr); 3150 3151 ierr = MatCreate(comm,outmat);CHKERRQ(ierr); 3152 ierr = MatSetSizes(*outmat,m,n*bs,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 3153 ierr = MatSetBlockSizes(*outmat,bs,cbs);CHKERRQ(ierr); 3154 ierr = MatSetType(*outmat,MATMPISBAIJ);CHKERRQ(ierr); 3155 ierr = MatMPISBAIJSetPreallocation(*outmat,bs,0,dnz,0,onz);CHKERRQ(ierr); 3156 ierr = MatPreallocateFinalize(dnz,onz);CHKERRQ(ierr); 3157 } 3158 3159 /* numeric phase */ 3160 ierr = MatGetBlockSizes(inmat,&bs,&cbs);CHKERRQ(ierr); 3161 ierr = MatGetOwnershipRange(*outmat,&rstart,NULL);CHKERRQ(ierr); 3162 3163 ierr = MatSetOption(inmat,MAT_GETROW_UPPERTRIANGULAR,PETSC_TRUE);CHKERRQ(ierr); 3164 for (i=0; i<m; i++) { 3165 ierr = MatGetRow_SeqSBAIJ(inmat,i,&nnz,&indx,&values);CHKERRQ(ierr); 3166 Ii = i + rstart; 3167 ierr = MatSetValues(*outmat,1,&Ii,nnz,indx,values,INSERT_VALUES);CHKERRQ(ierr); 3168 ierr = MatRestoreRow_SeqSBAIJ(inmat,i,&nnz,&indx,&values);CHKERRQ(ierr); 3169 } 3170 ierr = MatSetOption(inmat,MAT_GETROW_UPPERTRIANGULAR,PETSC_FALSE);CHKERRQ(ierr); 3171 ierr = MatAssemblyBegin(*outmat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3172 ierr = MatAssemblyEnd(*outmat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3173 PetscFunctionReturn(0); 3174 } 3175