1 /* 2 Defines the basic matrix operations for the AIJ (compressed row) 3 matrix storage format. 4 */ 5 6 7 #include <../src/mat/impls/aij/seq/aij.h> /*I "petscmat.h" I*/ 8 #include <petscblaslapack.h> 9 #include <petscbt.h> 10 #include <petsc/private/kernels/blocktranspose.h> 11 12 PetscErrorCode MatSeqAIJSetTypeFromOptions(Mat A) 13 { 14 PetscErrorCode ierr; 15 PetscBool flg; 16 char type[256]; 17 18 PetscFunctionBegin; 19 ierr = PetscObjectOptionsBegin((PetscObject)A); 20 ierr = PetscOptionsFList("-mat_seqaij_type","Matrix SeqAIJ type","MatSeqAIJSetType",MatSeqAIJList,"seqaij",type,256,&flg);CHKERRQ(ierr); 21 if (flg) { 22 ierr = MatSeqAIJSetType(A,type);CHKERRQ(ierr); 23 } 24 ierr = PetscOptionsEnd();CHKERRQ(ierr); 25 PetscFunctionReturn(0); 26 } 27 28 PetscErrorCode MatGetColumnNorms_SeqAIJ(Mat A,NormType type,PetscReal *norms) 29 { 30 PetscErrorCode ierr; 31 PetscInt i,m,n; 32 Mat_SeqAIJ *aij = (Mat_SeqAIJ*)A->data; 33 34 PetscFunctionBegin; 35 ierr = MatGetSize(A,&m,&n);CHKERRQ(ierr); 36 ierr = PetscArrayzero(norms,n);CHKERRQ(ierr); 37 if (type == NORM_2) { 38 for (i=0; i<aij->i[m]; i++) { 39 norms[aij->j[i]] += PetscAbsScalar(aij->a[i]*aij->a[i]); 40 } 41 } else if (type == NORM_1) { 42 for (i=0; i<aij->i[m]; i++) { 43 norms[aij->j[i]] += PetscAbsScalar(aij->a[i]); 44 } 45 } else if (type == NORM_INFINITY) { 46 for (i=0; i<aij->i[m]; i++) { 47 norms[aij->j[i]] = PetscMax(PetscAbsScalar(aij->a[i]),norms[aij->j[i]]); 48 } 49 } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Unknown NormType"); 50 51 if (type == NORM_2) { 52 for (i=0; i<n; i++) norms[i] = PetscSqrtReal(norms[i]); 53 } 54 PetscFunctionReturn(0); 55 } 56 57 PetscErrorCode MatFindOffBlockDiagonalEntries_SeqAIJ(Mat A,IS *is) 58 { 59 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 60 PetscInt i,m=A->rmap->n,cnt = 0, bs = A->rmap->bs; 61 const PetscInt *jj = a->j,*ii = a->i; 62 PetscInt *rows; 63 PetscErrorCode ierr; 64 65 PetscFunctionBegin; 66 for (i=0; i<m; i++) { 67 if ((ii[i] != ii[i+1]) && ((jj[ii[i]] < bs*(i/bs)) || (jj[ii[i+1]-1] > bs*((i+bs)/bs)-1))) { 68 cnt++; 69 } 70 } 71 ierr = PetscMalloc1(cnt,&rows);CHKERRQ(ierr); 72 cnt = 0; 73 for (i=0; i<m; i++) { 74 if ((ii[i] != ii[i+1]) && ((jj[ii[i]] < bs*(i/bs)) || (jj[ii[i+1]-1] > bs*((i+bs)/bs)-1))) { 75 rows[cnt] = i; 76 cnt++; 77 } 78 } 79 ierr = ISCreateGeneral(PETSC_COMM_SELF,cnt,rows,PETSC_OWN_POINTER,is);CHKERRQ(ierr); 80 PetscFunctionReturn(0); 81 } 82 83 PetscErrorCode MatFindZeroDiagonals_SeqAIJ_Private(Mat A,PetscInt *nrows,PetscInt **zrows) 84 { 85 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 86 const MatScalar *aa = a->a; 87 PetscInt i,m=A->rmap->n,cnt = 0; 88 const PetscInt *ii = a->i,*jj = a->j,*diag; 89 PetscInt *rows; 90 PetscErrorCode ierr; 91 92 PetscFunctionBegin; 93 ierr = MatMarkDiagonal_SeqAIJ(A);CHKERRQ(ierr); 94 diag = a->diag; 95 for (i=0; i<m; i++) { 96 if ((diag[i] >= ii[i+1]) || (jj[diag[i]] != i) || (aa[diag[i]] == 0.0)) { 97 cnt++; 98 } 99 } 100 ierr = PetscMalloc1(cnt,&rows);CHKERRQ(ierr); 101 cnt = 0; 102 for (i=0; i<m; i++) { 103 if ((diag[i] >= ii[i+1]) || (jj[diag[i]] != i) || (aa[diag[i]] == 0.0)) { 104 rows[cnt++] = i; 105 } 106 } 107 *nrows = cnt; 108 *zrows = rows; 109 PetscFunctionReturn(0); 110 } 111 112 PetscErrorCode MatFindZeroDiagonals_SeqAIJ(Mat A,IS *zrows) 113 { 114 PetscInt nrows,*rows; 115 PetscErrorCode ierr; 116 117 PetscFunctionBegin; 118 *zrows = NULL; 119 ierr = MatFindZeroDiagonals_SeqAIJ_Private(A,&nrows,&rows);CHKERRQ(ierr); 120 ierr = ISCreateGeneral(PetscObjectComm((PetscObject)A),nrows,rows,PETSC_OWN_POINTER,zrows);CHKERRQ(ierr); 121 PetscFunctionReturn(0); 122 } 123 124 PetscErrorCode MatFindNonzeroRows_SeqAIJ(Mat A,IS *keptrows) 125 { 126 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 127 const MatScalar *aa; 128 PetscInt m=A->rmap->n,cnt = 0; 129 const PetscInt *ii; 130 PetscInt n,i,j,*rows; 131 PetscErrorCode ierr; 132 133 PetscFunctionBegin; 134 *keptrows = NULL; 135 ii = a->i; 136 for (i=0; i<m; i++) { 137 n = ii[i+1] - ii[i]; 138 if (!n) { 139 cnt++; 140 goto ok1; 141 } 142 aa = a->a + ii[i]; 143 for (j=0; j<n; j++) { 144 if (aa[j] != 0.0) goto ok1; 145 } 146 cnt++; 147 ok1:; 148 } 149 if (!cnt) PetscFunctionReturn(0); 150 ierr = PetscMalloc1(A->rmap->n-cnt,&rows);CHKERRQ(ierr); 151 cnt = 0; 152 for (i=0; i<m; i++) { 153 n = ii[i+1] - ii[i]; 154 if (!n) continue; 155 aa = a->a + ii[i]; 156 for (j=0; j<n; j++) { 157 if (aa[j] != 0.0) { 158 rows[cnt++] = i; 159 break; 160 } 161 } 162 } 163 ierr = ISCreateGeneral(PETSC_COMM_SELF,cnt,rows,PETSC_OWN_POINTER,keptrows);CHKERRQ(ierr); 164 PetscFunctionReturn(0); 165 } 166 167 PetscErrorCode MatDiagonalSet_SeqAIJ(Mat Y,Vec D,InsertMode is) 168 { 169 PetscErrorCode ierr; 170 Mat_SeqAIJ *aij = (Mat_SeqAIJ*) Y->data; 171 PetscInt i,m = Y->rmap->n; 172 const PetscInt *diag; 173 MatScalar *aa = aij->a; 174 const PetscScalar *v; 175 PetscBool missing; 176 #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA) 177 PetscBool inserted = PETSC_FALSE; 178 #endif 179 180 PetscFunctionBegin; 181 if (Y->assembled) { 182 ierr = MatMissingDiagonal_SeqAIJ(Y,&missing,NULL);CHKERRQ(ierr); 183 if (!missing) { 184 diag = aij->diag; 185 ierr = VecGetArrayRead(D,&v);CHKERRQ(ierr); 186 if (is == INSERT_VALUES) { 187 #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA) 188 inserted = PETSC_TRUE; 189 #endif 190 for (i=0; i<m; i++) { 191 aa[diag[i]] = v[i]; 192 } 193 } else { 194 for (i=0; i<m; i++) { 195 #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA) 196 if (v[i] != 0.0) inserted = PETSC_TRUE; 197 #endif 198 aa[diag[i]] += v[i]; 199 } 200 } 201 #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA) 202 if (inserted) Y->offloadmask = PETSC_OFFLOAD_CPU; 203 #endif 204 ierr = VecRestoreArrayRead(D,&v);CHKERRQ(ierr); 205 PetscFunctionReturn(0); 206 } 207 ierr = MatSeqAIJInvalidateDiagonal(Y);CHKERRQ(ierr); 208 } 209 ierr = MatDiagonalSet_Default(Y,D,is);CHKERRQ(ierr); 210 PetscFunctionReturn(0); 211 } 212 213 PetscErrorCode MatGetRowIJ_SeqAIJ(Mat A,PetscInt oshift,PetscBool symmetric,PetscBool inodecompressed,PetscInt *m,const PetscInt *ia[],const PetscInt *ja[],PetscBool *done) 214 { 215 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 216 PetscErrorCode ierr; 217 PetscInt i,ishift; 218 219 PetscFunctionBegin; 220 *m = A->rmap->n; 221 if (!ia) PetscFunctionReturn(0); 222 ishift = 0; 223 if (symmetric && !A->structurally_symmetric) { 224 ierr = MatToSymmetricIJ_SeqAIJ(A->rmap->n,a->i,a->j,PETSC_TRUE,ishift,oshift,(PetscInt**)ia,(PetscInt**)ja);CHKERRQ(ierr); 225 } else if (oshift == 1) { 226 PetscInt *tia; 227 PetscInt nz = a->i[A->rmap->n]; 228 /* malloc space and add 1 to i and j indices */ 229 ierr = PetscMalloc1(A->rmap->n+1,&tia);CHKERRQ(ierr); 230 for (i=0; i<A->rmap->n+1; i++) tia[i] = a->i[i] + 1; 231 *ia = tia; 232 if (ja) { 233 PetscInt *tja; 234 ierr = PetscMalloc1(nz+1,&tja);CHKERRQ(ierr); 235 for (i=0; i<nz; i++) tja[i] = a->j[i] + 1; 236 *ja = tja; 237 } 238 } else { 239 *ia = a->i; 240 if (ja) *ja = a->j; 241 } 242 PetscFunctionReturn(0); 243 } 244 245 PetscErrorCode MatRestoreRowIJ_SeqAIJ(Mat A,PetscInt oshift,PetscBool symmetric,PetscBool inodecompressed,PetscInt *n,const PetscInt *ia[],const PetscInt *ja[],PetscBool *done) 246 { 247 PetscErrorCode ierr; 248 249 PetscFunctionBegin; 250 if (!ia) PetscFunctionReturn(0); 251 if ((symmetric && !A->structurally_symmetric) || oshift == 1) { 252 ierr = PetscFree(*ia);CHKERRQ(ierr); 253 if (ja) {ierr = PetscFree(*ja);CHKERRQ(ierr);} 254 } 255 PetscFunctionReturn(0); 256 } 257 258 PetscErrorCode MatGetColumnIJ_SeqAIJ(Mat A,PetscInt oshift,PetscBool symmetric,PetscBool inodecompressed,PetscInt *nn,const PetscInt *ia[],const PetscInt *ja[],PetscBool *done) 259 { 260 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 261 PetscErrorCode ierr; 262 PetscInt i,*collengths,*cia,*cja,n = A->cmap->n,m = A->rmap->n; 263 PetscInt nz = a->i[m],row,*jj,mr,col; 264 265 PetscFunctionBegin; 266 *nn = n; 267 if (!ia) PetscFunctionReturn(0); 268 if (symmetric) { 269 ierr = MatToSymmetricIJ_SeqAIJ(A->rmap->n,a->i,a->j,PETSC_TRUE,0,oshift,(PetscInt**)ia,(PetscInt**)ja);CHKERRQ(ierr); 270 } else { 271 ierr = PetscCalloc1(n,&collengths);CHKERRQ(ierr); 272 ierr = PetscMalloc1(n+1,&cia);CHKERRQ(ierr); 273 ierr = PetscMalloc1(nz,&cja);CHKERRQ(ierr); 274 jj = a->j; 275 for (i=0; i<nz; i++) { 276 collengths[jj[i]]++; 277 } 278 cia[0] = oshift; 279 for (i=0; i<n; i++) { 280 cia[i+1] = cia[i] + collengths[i]; 281 } 282 ierr = PetscArrayzero(collengths,n);CHKERRQ(ierr); 283 jj = a->j; 284 for (row=0; row<m; row++) { 285 mr = a->i[row+1] - a->i[row]; 286 for (i=0; i<mr; i++) { 287 col = *jj++; 288 289 cja[cia[col] + collengths[col]++ - oshift] = row + oshift; 290 } 291 } 292 ierr = PetscFree(collengths);CHKERRQ(ierr); 293 *ia = cia; *ja = cja; 294 } 295 PetscFunctionReturn(0); 296 } 297 298 PetscErrorCode MatRestoreColumnIJ_SeqAIJ(Mat A,PetscInt oshift,PetscBool symmetric,PetscBool inodecompressed,PetscInt *n,const PetscInt *ia[],const PetscInt *ja[],PetscBool *done) 299 { 300 PetscErrorCode ierr; 301 302 PetscFunctionBegin; 303 if (!ia) PetscFunctionReturn(0); 304 305 ierr = PetscFree(*ia);CHKERRQ(ierr); 306 ierr = PetscFree(*ja);CHKERRQ(ierr); 307 PetscFunctionReturn(0); 308 } 309 310 /* 311 MatGetColumnIJ_SeqAIJ_Color() and MatRestoreColumnIJ_SeqAIJ_Color() are customized from 312 MatGetColumnIJ_SeqAIJ() and MatRestoreColumnIJ_SeqAIJ() by adding an output 313 spidx[], index of a->a, to be used in MatTransposeColoringCreate_SeqAIJ() and MatFDColoringCreate_SeqXAIJ() 314 */ 315 PetscErrorCode MatGetColumnIJ_SeqAIJ_Color(Mat A,PetscInt oshift,PetscBool symmetric,PetscBool inodecompressed,PetscInt *nn,const PetscInt *ia[],const PetscInt *ja[],PetscInt *spidx[],PetscBool *done) 316 { 317 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 318 PetscErrorCode ierr; 319 PetscInt i,*collengths,*cia,*cja,n = A->cmap->n,m = A->rmap->n; 320 PetscInt nz = a->i[m],row,mr,col,tmp; 321 PetscInt *cspidx; 322 const PetscInt *jj; 323 324 PetscFunctionBegin; 325 *nn = n; 326 if (!ia) PetscFunctionReturn(0); 327 328 ierr = PetscCalloc1(n,&collengths);CHKERRQ(ierr); 329 ierr = PetscMalloc1(n+1,&cia);CHKERRQ(ierr); 330 ierr = PetscMalloc1(nz,&cja);CHKERRQ(ierr); 331 ierr = PetscMalloc1(nz,&cspidx);CHKERRQ(ierr); 332 jj = a->j; 333 for (i=0; i<nz; i++) { 334 collengths[jj[i]]++; 335 } 336 cia[0] = oshift; 337 for (i=0; i<n; i++) { 338 cia[i+1] = cia[i] + collengths[i]; 339 } 340 ierr = PetscArrayzero(collengths,n);CHKERRQ(ierr); 341 jj = a->j; 342 for (row=0; row<m; row++) { 343 mr = a->i[row+1] - a->i[row]; 344 for (i=0; i<mr; i++) { 345 col = *jj++; 346 tmp = cia[col] + collengths[col]++ - oshift; 347 cspidx[tmp] = a->i[row] + i; /* index of a->j */ 348 cja[tmp] = row + oshift; 349 } 350 } 351 ierr = PetscFree(collengths);CHKERRQ(ierr); 352 *ia = cia; 353 *ja = cja; 354 *spidx = cspidx; 355 PetscFunctionReturn(0); 356 } 357 358 PetscErrorCode MatRestoreColumnIJ_SeqAIJ_Color(Mat A,PetscInt oshift,PetscBool symmetric,PetscBool inodecompressed,PetscInt *n,const PetscInt *ia[],const PetscInt *ja[],PetscInt *spidx[],PetscBool *done) 359 { 360 PetscErrorCode ierr; 361 362 PetscFunctionBegin; 363 ierr = MatRestoreColumnIJ_SeqAIJ(A,oshift,symmetric,inodecompressed,n,ia,ja,done);CHKERRQ(ierr); 364 ierr = PetscFree(*spidx);CHKERRQ(ierr); 365 PetscFunctionReturn(0); 366 } 367 368 PetscErrorCode MatSetValuesRow_SeqAIJ(Mat A,PetscInt row,const PetscScalar v[]) 369 { 370 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 371 PetscInt *ai = a->i; 372 PetscErrorCode ierr; 373 374 PetscFunctionBegin; 375 ierr = PetscArraycpy(a->a+ai[row],v,ai[row+1]-ai[row]);CHKERRQ(ierr); 376 #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA) 377 if (A->offloadmask != PETSC_OFFLOAD_UNALLOCATED && ai[row+1]-ai[row]) A->offloadmask = PETSC_OFFLOAD_CPU; 378 #endif 379 PetscFunctionReturn(0); 380 } 381 382 /* 383 MatSeqAIJSetValuesLocalFast - An optimized version of MatSetValuesLocal() for SeqAIJ matrices with several assumptions 384 385 - a single row of values is set with each call 386 - no row or column indices are negative or (in error) larger than the number of rows or columns 387 - the values are always added to the matrix, not set 388 - no new locations are introduced in the nonzero structure of the matrix 389 390 This does NOT assume the global column indices are sorted 391 392 */ 393 394 #include <petsc/private/isimpl.h> 395 PetscErrorCode MatSeqAIJSetValuesLocalFast(Mat A,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],const PetscScalar v[],InsertMode is) 396 { 397 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 398 PetscInt low,high,t,row,nrow,i,col,l; 399 const PetscInt *rp,*ai = a->i,*ailen = a->ilen,*aj = a->j; 400 PetscInt lastcol = -1; 401 MatScalar *ap,value,*aa = a->a; 402 const PetscInt *ridx = A->rmap->mapping->indices,*cidx = A->cmap->mapping->indices; 403 404 row = ridx[im[0]]; 405 rp = aj + ai[row]; 406 ap = aa + ai[row]; 407 nrow = ailen[row]; 408 low = 0; 409 high = nrow; 410 for (l=0; l<n; l++) { /* loop over added columns */ 411 col = cidx[in[l]]; 412 value = v[l]; 413 414 if (col <= lastcol) low = 0; 415 else high = nrow; 416 lastcol = col; 417 while (high-low > 5) { 418 t = (low+high)/2; 419 if (rp[t] > col) high = t; 420 else low = t; 421 } 422 for (i=low; i<high; i++) { 423 if (rp[i] == col) { 424 ap[i] += value; 425 low = i + 1; 426 break; 427 } 428 } 429 } 430 #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA) 431 if (A->offloadmask != PETSC_OFFLOAD_UNALLOCATED && m && n) A->offloadmask = PETSC_OFFLOAD_CPU; 432 #endif 433 return 0; 434 } 435 436 PetscErrorCode MatSetValues_SeqAIJ(Mat A,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],const PetscScalar v[],InsertMode is) 437 { 438 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 439 PetscInt *rp,k,low,high,t,ii,row,nrow,i,col,l,rmax,N; 440 PetscInt *imax = a->imax,*ai = a->i,*ailen = a->ilen; 441 PetscErrorCode ierr; 442 PetscInt *aj = a->j,nonew = a->nonew,lastcol = -1; 443 MatScalar *ap=NULL,value=0.0,*aa = a->a; 444 PetscBool ignorezeroentries = a->ignorezeroentries; 445 PetscBool roworiented = a->roworiented; 446 #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA) 447 PetscBool inserted = PETSC_FALSE; 448 #endif 449 450 PetscFunctionBegin; 451 for (k=0; k<m; k++) { /* loop over added rows */ 452 row = im[k]; 453 if (row < 0) continue; 454 if (PetscUnlikelyDebug(row >= A->rmap->n)) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Row too large: row %D max %D",row,A->rmap->n-1); 455 rp = aj + ai[row]; 456 if (!A->structure_only) ap = aa + ai[row]; 457 rmax = imax[row]; nrow = ailen[row]; 458 low = 0; 459 high = nrow; 460 for (l=0; l<n; l++) { /* loop over added columns */ 461 if (in[l] < 0) continue; 462 if (PetscUnlikelyDebug(in[l] >= A->cmap->n)) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Column too large: col %D max %D",in[l],A->cmap->n-1); 463 col = in[l]; 464 if (v && !A->structure_only) value = roworiented ? v[l + k*n] : v[k + l*m]; 465 if (!A->structure_only && value == 0.0 && ignorezeroentries && is == ADD_VALUES && row != col) continue; 466 467 if (col <= lastcol) low = 0; 468 else high = nrow; 469 lastcol = col; 470 while (high-low > 5) { 471 t = (low+high)/2; 472 if (rp[t] > col) high = t; 473 else low = t; 474 } 475 for (i=low; i<high; i++) { 476 if (rp[i] > col) break; 477 if (rp[i] == col) { 478 if (!A->structure_only) { 479 if (is == ADD_VALUES) { 480 ap[i] += value; 481 (void)PetscLogFlops(1.0); 482 } 483 else ap[i] = value; 484 #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA) 485 inserted = PETSC_TRUE; 486 #endif 487 } 488 low = i + 1; 489 goto noinsert; 490 } 491 } 492 if (value == 0.0 && ignorezeroentries && row != col) goto noinsert; 493 if (nonew == 1) goto noinsert; 494 if (nonew == -1) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero at (%D,%D) in the matrix",row,col); 495 if (A->structure_only) { 496 MatSeqXAIJReallocateAIJ_structure_only(A,A->rmap->n,1,nrow,row,col,rmax,ai,aj,rp,imax,nonew,MatScalar); 497 } else { 498 MatSeqXAIJReallocateAIJ(A,A->rmap->n,1,nrow,row,col,rmax,aa,ai,aj,rp,ap,imax,nonew,MatScalar); 499 } 500 N = nrow++ - 1; a->nz++; high++; 501 /* shift up all the later entries in this row */ 502 ierr = PetscArraymove(rp+i+1,rp+i,N-i+1);CHKERRQ(ierr); 503 rp[i] = col; 504 if (!A->structure_only){ 505 ierr = PetscArraymove(ap+i+1,ap+i,N-i+1);CHKERRQ(ierr); 506 ap[i] = value; 507 } 508 low = i + 1; 509 A->nonzerostate++; 510 #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA) 511 inserted = PETSC_TRUE; 512 #endif 513 noinsert:; 514 } 515 ailen[row] = nrow; 516 } 517 #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA) 518 if (A->offloadmask != PETSC_OFFLOAD_UNALLOCATED && inserted) A->offloadmask = PETSC_OFFLOAD_CPU; 519 #endif 520 PetscFunctionReturn(0); 521 } 522 523 524 PetscErrorCode MatSetValues_SeqAIJ_SortedFullNoPreallocation(Mat A,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],const PetscScalar v[],InsertMode is) 525 { 526 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 527 PetscInt *rp,k,row; 528 PetscInt *ai = a->i; 529 PetscErrorCode ierr; 530 PetscInt *aj = a->j; 531 MatScalar *aa = a->a,*ap; 532 533 PetscFunctionBegin; 534 if (A->was_assembled) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Cannot call on assembled matrix."); 535 if (m*n+a->nz > a->maxnz) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Number of entries in matrix will be larger than maximum nonzeros allocated for %D in MatSeqAIJSetTotalPreallocation()",a->maxnz); 536 for (k=0; k<m; k++) { /* loop over added rows */ 537 row = im[k]; 538 rp = aj + ai[row]; 539 ap = aa + ai[row]; 540 541 ierr = PetscMemcpy(rp,in,n*sizeof(PetscInt));CHKERRQ(ierr); 542 if (!A->structure_only) { 543 if (v) { 544 ierr = PetscMemcpy(ap,v,n*sizeof(PetscScalar));CHKERRQ(ierr); 545 v += n; 546 } else { 547 ierr = PetscMemzero(ap,n*sizeof(PetscScalar));CHKERRQ(ierr); 548 } 549 } 550 a->ilen[row] = n; 551 a->imax[row] = n; 552 a->i[row+1] = a->i[row]+n; 553 a->nz += n; 554 } 555 #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA) 556 if (A->offloadmask != PETSC_OFFLOAD_UNALLOCATED && m && n) A->offloadmask = PETSC_OFFLOAD_CPU; 557 #endif 558 PetscFunctionReturn(0); 559 } 560 561 /*@ 562 MatSeqAIJSetTotalPreallocation - Sets an upper bound on the total number of expected nonzeros in the matrix. 563 564 Input Parameters: 565 + A - the SeqAIJ matrix 566 - nztotal - bound on the number of nonzeros 567 568 Level: advanced 569 570 Notes: 571 This can be called if you will be provided the matrix row by row (from row zero) with sorted column indices for each row. 572 Simply call MatSetValues() after this call to provide the matrix entries in the usual manner. This matrix may be used 573 as always with multiple matrix assemblies. 574 575 .seealso: MatSetOption(), MAT_SORTED_FULL, MatSetValues(), MatSeqAIJSetPreallocation() 576 @*/ 577 578 PetscErrorCode MatSeqAIJSetTotalPreallocation(Mat A,PetscInt nztotal) 579 { 580 PetscErrorCode ierr; 581 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 582 583 PetscFunctionBegin; 584 ierr = PetscLayoutSetUp(A->rmap);CHKERRQ(ierr); 585 ierr = PetscLayoutSetUp(A->cmap);CHKERRQ(ierr); 586 a->maxnz = nztotal; 587 if (!a->imax) { 588 ierr = PetscMalloc1(A->rmap->n,&a->imax);CHKERRQ(ierr); 589 ierr = PetscLogObjectMemory((PetscObject)A,A->rmap->n*sizeof(PetscInt));CHKERRQ(ierr); 590 } 591 if (!a->ilen) { 592 ierr = PetscMalloc1(A->rmap->n,&a->ilen);CHKERRQ(ierr); 593 ierr = PetscLogObjectMemory((PetscObject)A,A->rmap->n*sizeof(PetscInt));CHKERRQ(ierr); 594 } else { 595 ierr = PetscMemzero(a->ilen,A->rmap->n*sizeof(PetscInt));CHKERRQ(ierr); 596 } 597 598 /* allocate the matrix space */ 599 if (A->structure_only) { 600 ierr = PetscMalloc1(nztotal,&a->j);CHKERRQ(ierr); 601 ierr = PetscMalloc1(A->rmap->n+1,&a->i);CHKERRQ(ierr); 602 ierr = PetscLogObjectMemory((PetscObject)A,(A->rmap->n+1)*sizeof(PetscInt)+nztotal*sizeof(PetscInt));CHKERRQ(ierr); 603 } else { 604 ierr = PetscMalloc3(nztotal,&a->a,nztotal,&a->j,A->rmap->n+1,&a->i);CHKERRQ(ierr); 605 ierr = PetscLogObjectMemory((PetscObject)A,(A->rmap->n+1)*sizeof(PetscInt)+nztotal*(sizeof(PetscScalar)+sizeof(PetscInt)));CHKERRQ(ierr); 606 } 607 a->i[0] = 0; 608 if (A->structure_only) { 609 a->singlemalloc = PETSC_FALSE; 610 a->free_a = PETSC_FALSE; 611 } else { 612 a->singlemalloc = PETSC_TRUE; 613 a->free_a = PETSC_TRUE; 614 } 615 a->free_ij = PETSC_TRUE; 616 A->ops->setvalues = MatSetValues_SeqAIJ_SortedFullNoPreallocation; 617 A->preallocated = PETSC_TRUE; 618 PetscFunctionReturn(0); 619 } 620 621 PetscErrorCode MatSetValues_SeqAIJ_SortedFull(Mat A,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],const PetscScalar v[],InsertMode is) 622 { 623 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 624 PetscInt *rp,k,row; 625 PetscInt *ai = a->i,*ailen = a->ilen; 626 PetscErrorCode ierr; 627 PetscInt *aj = a->j; 628 MatScalar *aa = a->a,*ap; 629 630 PetscFunctionBegin; 631 for (k=0; k<m; k++) { /* loop over added rows */ 632 row = im[k]; 633 if (PetscUnlikelyDebug(n > a->imax[row])) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Preallocation for row %D does not match number of columns provided",n); 634 rp = aj + ai[row]; 635 ap = aa + ai[row]; 636 if (!A->was_assembled) { 637 ierr = PetscMemcpy(rp,in,n*sizeof(PetscInt));CHKERRQ(ierr); 638 } 639 if (!A->structure_only) { 640 if (v) { 641 ierr = PetscMemcpy(ap,v,n*sizeof(PetscScalar));CHKERRQ(ierr); 642 v += n; 643 } else { 644 ierr = PetscMemzero(ap,n*sizeof(PetscScalar));CHKERRQ(ierr); 645 } 646 } 647 ailen[row] = n; 648 a->nz += n; 649 } 650 #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA) 651 if (A->offloadmask != PETSC_OFFLOAD_UNALLOCATED && m && n) A->offloadmask = PETSC_OFFLOAD_CPU; 652 #endif 653 PetscFunctionReturn(0); 654 } 655 656 657 PetscErrorCode MatGetValues_SeqAIJ(Mat A,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],PetscScalar v[]) 658 { 659 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 660 PetscInt *rp,k,low,high,t,row,nrow,i,col,l,*aj = a->j; 661 PetscInt *ai = a->i,*ailen = a->ilen; 662 MatScalar *ap,*aa = a->a; 663 664 PetscFunctionBegin; 665 for (k=0; k<m; k++) { /* loop over rows */ 666 row = im[k]; 667 if (row < 0) {v += n; continue;} /* SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative row: %D",row); */ 668 if (row >= A->rmap->n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Row too large: row %D max %D",row,A->rmap->n-1); 669 rp = aj + ai[row]; ap = aa + ai[row]; 670 nrow = ailen[row]; 671 for (l=0; l<n; l++) { /* loop over columns */ 672 if (in[l] < 0) {v++; continue;} /* SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative column: %D",in[l]); */ 673 if (in[l] >= A->cmap->n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Column too large: col %D max %D",in[l],A->cmap->n-1); 674 col = in[l]; 675 high = nrow; low = 0; /* assume unsorted */ 676 while (high-low > 5) { 677 t = (low+high)/2; 678 if (rp[t] > col) high = t; 679 else low = t; 680 } 681 for (i=low; i<high; i++) { 682 if (rp[i] > col) break; 683 if (rp[i] == col) { 684 *v++ = ap[i]; 685 goto finished; 686 } 687 } 688 *v++ = 0.0; 689 finished:; 690 } 691 } 692 PetscFunctionReturn(0); 693 } 694 695 PetscErrorCode MatView_SeqAIJ_Binary(Mat mat,PetscViewer viewer) 696 { 697 Mat_SeqAIJ *A = (Mat_SeqAIJ*)mat->data; 698 PetscInt header[4],M,N,m,nz,i; 699 PetscInt *rowlens; 700 PetscErrorCode ierr; 701 702 PetscFunctionBegin; 703 ierr = PetscViewerSetUp(viewer);CHKERRQ(ierr); 704 705 M = mat->rmap->N; 706 N = mat->cmap->N; 707 m = mat->rmap->n; 708 nz = A->nz; 709 710 /* write matrix header */ 711 header[0] = MAT_FILE_CLASSID; 712 header[1] = M; header[2] = N; header[3] = nz; 713 ierr = PetscViewerBinaryWrite(viewer,header,4,PETSC_INT);CHKERRQ(ierr); 714 715 /* fill in and store row lengths */ 716 ierr = PetscMalloc1(m,&rowlens);CHKERRQ(ierr); 717 for (i=0; i<m; i++) rowlens[i] = A->i[i+1] - A->i[i]; 718 ierr = PetscViewerBinaryWrite(viewer,rowlens,m,PETSC_INT);CHKERRQ(ierr); 719 ierr = PetscFree(rowlens);CHKERRQ(ierr); 720 /* store column indices */ 721 ierr = PetscViewerBinaryWrite(viewer,A->j,nz,PETSC_INT);CHKERRQ(ierr); 722 /* store nonzero values */ 723 ierr = PetscViewerBinaryWrite(viewer,A->a,nz,PETSC_SCALAR);CHKERRQ(ierr); 724 725 /* write block size option to the viewer's .info file */ 726 ierr = MatView_Binary_BlockSizes(mat,viewer);CHKERRQ(ierr); 727 PetscFunctionReturn(0); 728 } 729 730 static PetscErrorCode MatView_SeqAIJ_ASCII_structonly(Mat A,PetscViewer viewer) 731 { 732 PetscErrorCode ierr; 733 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 734 PetscInt i,k,m=A->rmap->N; 735 736 PetscFunctionBegin; 737 ierr = PetscViewerASCIIUseTabs(viewer,PETSC_FALSE);CHKERRQ(ierr); 738 for (i=0; i<m; i++) { 739 ierr = PetscViewerASCIIPrintf(viewer,"row %D:",i);CHKERRQ(ierr); 740 for (k=a->i[i]; k<a->i[i+1]; k++) { 741 ierr = PetscViewerASCIIPrintf(viewer," (%D) ",a->j[k]);CHKERRQ(ierr); 742 } 743 ierr = PetscViewerASCIIPrintf(viewer,"\n");CHKERRQ(ierr); 744 } 745 ierr = PetscViewerASCIIUseTabs(viewer,PETSC_TRUE);CHKERRQ(ierr); 746 PetscFunctionReturn(0); 747 } 748 749 extern PetscErrorCode MatSeqAIJFactorInfo_Matlab(Mat,PetscViewer); 750 751 PetscErrorCode MatView_SeqAIJ_ASCII(Mat A,PetscViewer viewer) 752 { 753 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 754 PetscErrorCode ierr; 755 PetscInt i,j,m = A->rmap->n; 756 const char *name; 757 PetscViewerFormat format; 758 759 PetscFunctionBegin; 760 if (A->structure_only) { 761 ierr = MatView_SeqAIJ_ASCII_structonly(A,viewer);CHKERRQ(ierr); 762 PetscFunctionReturn(0); 763 } 764 765 ierr = PetscViewerGetFormat(viewer,&format);CHKERRQ(ierr); 766 if (format == PETSC_VIEWER_ASCII_MATLAB) { 767 PetscInt nofinalvalue = 0; 768 if (m && ((a->i[m] == a->i[m-1]) || (a->j[a->nz-1] != A->cmap->n-1))) { 769 /* Need a dummy value to ensure the dimension of the matrix. */ 770 nofinalvalue = 1; 771 } 772 ierr = PetscViewerASCIIUseTabs(viewer,PETSC_FALSE);CHKERRQ(ierr); 773 ierr = PetscViewerASCIIPrintf(viewer,"%% Size = %D %D \n",m,A->cmap->n);CHKERRQ(ierr); 774 ierr = PetscViewerASCIIPrintf(viewer,"%% Nonzeros = %D \n",a->nz);CHKERRQ(ierr); 775 #if defined(PETSC_USE_COMPLEX) 776 ierr = PetscViewerASCIIPrintf(viewer,"zzz = zeros(%D,4);\n",a->nz+nofinalvalue);CHKERRQ(ierr); 777 #else 778 ierr = PetscViewerASCIIPrintf(viewer,"zzz = zeros(%D,3);\n",a->nz+nofinalvalue);CHKERRQ(ierr); 779 #endif 780 ierr = PetscViewerASCIIPrintf(viewer,"zzz = [\n");CHKERRQ(ierr); 781 782 for (i=0; i<m; i++) { 783 for (j=a->i[i]; j<a->i[i+1]; j++) { 784 #if defined(PETSC_USE_COMPLEX) 785 ierr = PetscViewerASCIIPrintf(viewer,"%D %D %18.16e %18.16e\n",i+1,a->j[j]+1,(double)PetscRealPart(a->a[j]),(double)PetscImaginaryPart(a->a[j]));CHKERRQ(ierr); 786 #else 787 ierr = PetscViewerASCIIPrintf(viewer,"%D %D %18.16e\n",i+1,a->j[j]+1,(double)a->a[j]);CHKERRQ(ierr); 788 #endif 789 } 790 } 791 if (nofinalvalue) { 792 #if defined(PETSC_USE_COMPLEX) 793 ierr = PetscViewerASCIIPrintf(viewer,"%D %D %18.16e %18.16e\n",m,A->cmap->n,0.,0.);CHKERRQ(ierr); 794 #else 795 ierr = PetscViewerASCIIPrintf(viewer,"%D %D %18.16e\n",m,A->cmap->n,0.0);CHKERRQ(ierr); 796 #endif 797 } 798 ierr = PetscObjectGetName((PetscObject)A,&name);CHKERRQ(ierr); 799 ierr = PetscViewerASCIIPrintf(viewer,"];\n %s = spconvert(zzz);\n",name);CHKERRQ(ierr); 800 ierr = PetscViewerASCIIUseTabs(viewer,PETSC_TRUE);CHKERRQ(ierr); 801 } else if (format == PETSC_VIEWER_ASCII_FACTOR_INFO || format == PETSC_VIEWER_ASCII_INFO || format == PETSC_VIEWER_ASCII_INFO_DETAIL) { 802 PetscFunctionReturn(0); 803 } else if (format == PETSC_VIEWER_ASCII_COMMON) { 804 ierr = PetscViewerASCIIUseTabs(viewer,PETSC_FALSE);CHKERRQ(ierr); 805 for (i=0; i<m; i++) { 806 ierr = PetscViewerASCIIPrintf(viewer,"row %D:",i);CHKERRQ(ierr); 807 for (j=a->i[i]; j<a->i[i+1]; j++) { 808 #if defined(PETSC_USE_COMPLEX) 809 if (PetscImaginaryPart(a->a[j]) > 0.0 && PetscRealPart(a->a[j]) != 0.0) { 810 ierr = PetscViewerASCIIPrintf(viewer," (%D, %g + %g i)",a->j[j],(double)PetscRealPart(a->a[j]),(double)PetscImaginaryPart(a->a[j]));CHKERRQ(ierr); 811 } else if (PetscImaginaryPart(a->a[j]) < 0.0 && PetscRealPart(a->a[j]) != 0.0) { 812 ierr = PetscViewerASCIIPrintf(viewer," (%D, %g - %g i)",a->j[j],(double)PetscRealPart(a->a[j]),(double)-PetscImaginaryPart(a->a[j]));CHKERRQ(ierr); 813 } else if (PetscRealPart(a->a[j]) != 0.0) { 814 ierr = PetscViewerASCIIPrintf(viewer," (%D, %g) ",a->j[j],(double)PetscRealPart(a->a[j]));CHKERRQ(ierr); 815 } 816 #else 817 if (a->a[j] != 0.0) {ierr = PetscViewerASCIIPrintf(viewer," (%D, %g) ",a->j[j],(double)a->a[j]);CHKERRQ(ierr);} 818 #endif 819 } 820 ierr = PetscViewerASCIIPrintf(viewer,"\n");CHKERRQ(ierr); 821 } 822 ierr = PetscViewerASCIIUseTabs(viewer,PETSC_TRUE);CHKERRQ(ierr); 823 } else if (format == PETSC_VIEWER_ASCII_SYMMODU) { 824 PetscInt nzd=0,fshift=1,*sptr; 825 ierr = PetscViewerASCIIUseTabs(viewer,PETSC_FALSE);CHKERRQ(ierr); 826 ierr = PetscMalloc1(m+1,&sptr);CHKERRQ(ierr); 827 for (i=0; i<m; i++) { 828 sptr[i] = nzd+1; 829 for (j=a->i[i]; j<a->i[i+1]; j++) { 830 if (a->j[j] >= i) { 831 #if defined(PETSC_USE_COMPLEX) 832 if (PetscImaginaryPart(a->a[j]) != 0.0 || PetscRealPart(a->a[j]) != 0.0) nzd++; 833 #else 834 if (a->a[j] != 0.0) nzd++; 835 #endif 836 } 837 } 838 } 839 sptr[m] = nzd+1; 840 ierr = PetscViewerASCIIPrintf(viewer," %D %D\n\n",m,nzd);CHKERRQ(ierr); 841 for (i=0; i<m+1; i+=6) { 842 if (i+4<m) { 843 ierr = PetscViewerASCIIPrintf(viewer," %D %D %D %D %D %D\n",sptr[i],sptr[i+1],sptr[i+2],sptr[i+3],sptr[i+4],sptr[i+5]);CHKERRQ(ierr); 844 } else if (i+3<m) { 845 ierr = PetscViewerASCIIPrintf(viewer," %D %D %D %D %D\n",sptr[i],sptr[i+1],sptr[i+2],sptr[i+3],sptr[i+4]);CHKERRQ(ierr); 846 } else if (i+2<m) { 847 ierr = PetscViewerASCIIPrintf(viewer," %D %D %D %D\n",sptr[i],sptr[i+1],sptr[i+2],sptr[i+3]);CHKERRQ(ierr); 848 } else if (i+1<m) { 849 ierr = PetscViewerASCIIPrintf(viewer," %D %D %D\n",sptr[i],sptr[i+1],sptr[i+2]);CHKERRQ(ierr); 850 } else if (i<m) { 851 ierr = PetscViewerASCIIPrintf(viewer," %D %D\n",sptr[i],sptr[i+1]);CHKERRQ(ierr); 852 } else { 853 ierr = PetscViewerASCIIPrintf(viewer," %D\n",sptr[i]);CHKERRQ(ierr); 854 } 855 } 856 ierr = PetscViewerASCIIPrintf(viewer,"\n");CHKERRQ(ierr); 857 ierr = PetscFree(sptr);CHKERRQ(ierr); 858 for (i=0; i<m; i++) { 859 for (j=a->i[i]; j<a->i[i+1]; j++) { 860 if (a->j[j] >= i) {ierr = PetscViewerASCIIPrintf(viewer," %D ",a->j[j]+fshift);CHKERRQ(ierr);} 861 } 862 ierr = PetscViewerASCIIPrintf(viewer,"\n");CHKERRQ(ierr); 863 } 864 ierr = PetscViewerASCIIPrintf(viewer,"\n");CHKERRQ(ierr); 865 for (i=0; i<m; i++) { 866 for (j=a->i[i]; j<a->i[i+1]; j++) { 867 if (a->j[j] >= i) { 868 #if defined(PETSC_USE_COMPLEX) 869 if (PetscImaginaryPart(a->a[j]) != 0.0 || PetscRealPart(a->a[j]) != 0.0) { 870 ierr = PetscViewerASCIIPrintf(viewer," %18.16e %18.16e ",(double)PetscRealPart(a->a[j]),(double)PetscImaginaryPart(a->a[j]));CHKERRQ(ierr); 871 } 872 #else 873 if (a->a[j] != 0.0) {ierr = PetscViewerASCIIPrintf(viewer," %18.16e ",(double)a->a[j]);CHKERRQ(ierr);} 874 #endif 875 } 876 } 877 ierr = PetscViewerASCIIPrintf(viewer,"\n");CHKERRQ(ierr); 878 } 879 ierr = PetscViewerASCIIUseTabs(viewer,PETSC_TRUE);CHKERRQ(ierr); 880 } else if (format == PETSC_VIEWER_ASCII_DENSE) { 881 PetscInt cnt = 0,jcnt; 882 PetscScalar value; 883 #if defined(PETSC_USE_COMPLEX) 884 PetscBool realonly = PETSC_TRUE; 885 886 for (i=0; i<a->i[m]; i++) { 887 if (PetscImaginaryPart(a->a[i]) != 0.0) { 888 realonly = PETSC_FALSE; 889 break; 890 } 891 } 892 #endif 893 894 ierr = PetscViewerASCIIUseTabs(viewer,PETSC_FALSE);CHKERRQ(ierr); 895 for (i=0; i<m; i++) { 896 jcnt = 0; 897 for (j=0; j<A->cmap->n; j++) { 898 if (jcnt < a->i[i+1]-a->i[i] && j == a->j[cnt]) { 899 value = a->a[cnt++]; 900 jcnt++; 901 } else { 902 value = 0.0; 903 } 904 #if defined(PETSC_USE_COMPLEX) 905 if (realonly) { 906 ierr = PetscViewerASCIIPrintf(viewer," %7.5e ",(double)PetscRealPart(value));CHKERRQ(ierr); 907 } else { 908 ierr = PetscViewerASCIIPrintf(viewer," %7.5e+%7.5e i ",(double)PetscRealPart(value),(double)PetscImaginaryPart(value));CHKERRQ(ierr); 909 } 910 #else 911 ierr = PetscViewerASCIIPrintf(viewer," %7.5e ",(double)value);CHKERRQ(ierr); 912 #endif 913 } 914 ierr = PetscViewerASCIIPrintf(viewer,"\n");CHKERRQ(ierr); 915 } 916 ierr = PetscViewerASCIIUseTabs(viewer,PETSC_TRUE);CHKERRQ(ierr); 917 } else if (format == PETSC_VIEWER_ASCII_MATRIXMARKET) { 918 PetscInt fshift=1; 919 ierr = PetscViewerASCIIUseTabs(viewer,PETSC_FALSE);CHKERRQ(ierr); 920 #if defined(PETSC_USE_COMPLEX) 921 ierr = PetscViewerASCIIPrintf(viewer,"%%%%MatrixMarket matrix coordinate complex general\n");CHKERRQ(ierr); 922 #else 923 ierr = PetscViewerASCIIPrintf(viewer,"%%%%MatrixMarket matrix coordinate real general\n");CHKERRQ(ierr); 924 #endif 925 ierr = PetscViewerASCIIPrintf(viewer,"%D %D %D\n", m, A->cmap->n, a->nz);CHKERRQ(ierr); 926 for (i=0; i<m; i++) { 927 for (j=a->i[i]; j<a->i[i+1]; j++) { 928 #if defined(PETSC_USE_COMPLEX) 929 ierr = PetscViewerASCIIPrintf(viewer,"%D %D %g %g\n", i+fshift,a->j[j]+fshift,(double)PetscRealPart(a->a[j]),(double)PetscImaginaryPart(a->a[j]));CHKERRQ(ierr); 930 #else 931 ierr = PetscViewerASCIIPrintf(viewer,"%D %D %g\n", i+fshift, a->j[j]+fshift, (double)a->a[j]);CHKERRQ(ierr); 932 #endif 933 } 934 } 935 ierr = PetscViewerASCIIUseTabs(viewer,PETSC_TRUE);CHKERRQ(ierr); 936 } else { 937 ierr = PetscViewerASCIIUseTabs(viewer,PETSC_FALSE);CHKERRQ(ierr); 938 if (A->factortype) { 939 for (i=0; i<m; i++) { 940 ierr = PetscViewerASCIIPrintf(viewer,"row %D:",i);CHKERRQ(ierr); 941 /* L part */ 942 for (j=a->i[i]; j<a->i[i+1]; j++) { 943 #if defined(PETSC_USE_COMPLEX) 944 if (PetscImaginaryPart(a->a[j]) > 0.0) { 945 ierr = PetscViewerASCIIPrintf(viewer," (%D, %g + %g i)",a->j[j],(double)PetscRealPart(a->a[j]),(double)PetscImaginaryPart(a->a[j]));CHKERRQ(ierr); 946 } else if (PetscImaginaryPart(a->a[j]) < 0.0) { 947 ierr = PetscViewerASCIIPrintf(viewer," (%D, %g - %g i)",a->j[j],(double)PetscRealPart(a->a[j]),(double)(-PetscImaginaryPart(a->a[j])));CHKERRQ(ierr); 948 } else { 949 ierr = PetscViewerASCIIPrintf(viewer," (%D, %g) ",a->j[j],(double)PetscRealPart(a->a[j]));CHKERRQ(ierr); 950 } 951 #else 952 ierr = PetscViewerASCIIPrintf(viewer," (%D, %g) ",a->j[j],(double)a->a[j]);CHKERRQ(ierr); 953 #endif 954 } 955 /* diagonal */ 956 j = a->diag[i]; 957 #if defined(PETSC_USE_COMPLEX) 958 if (PetscImaginaryPart(a->a[j]) > 0.0) { 959 ierr = PetscViewerASCIIPrintf(viewer," (%D, %g + %g i)",a->j[j],(double)PetscRealPart(1.0/a->a[j]),(double)PetscImaginaryPart(1.0/a->a[j]));CHKERRQ(ierr); 960 } else if (PetscImaginaryPart(a->a[j]) < 0.0) { 961 ierr = PetscViewerASCIIPrintf(viewer," (%D, %g - %g i)",a->j[j],(double)PetscRealPart(1.0/a->a[j]),(double)(-PetscImaginaryPart(1.0/a->a[j])));CHKERRQ(ierr); 962 } else { 963 ierr = PetscViewerASCIIPrintf(viewer," (%D, %g) ",a->j[j],(double)PetscRealPart(1.0/a->a[j]));CHKERRQ(ierr); 964 } 965 #else 966 ierr = PetscViewerASCIIPrintf(viewer," (%D, %g) ",a->j[j],(double)(1.0/a->a[j]));CHKERRQ(ierr); 967 #endif 968 969 /* U part */ 970 for (j=a->diag[i+1]+1; j<a->diag[i]; j++) { 971 #if defined(PETSC_USE_COMPLEX) 972 if (PetscImaginaryPart(a->a[j]) > 0.0) { 973 ierr = PetscViewerASCIIPrintf(viewer," (%D, %g + %g i)",a->j[j],(double)PetscRealPart(a->a[j]),(double)PetscImaginaryPart(a->a[j]));CHKERRQ(ierr); 974 } else if (PetscImaginaryPart(a->a[j]) < 0.0) { 975 ierr = PetscViewerASCIIPrintf(viewer," (%D, %g - %g i)",a->j[j],(double)PetscRealPart(a->a[j]),(double)(-PetscImaginaryPart(a->a[j])));CHKERRQ(ierr); 976 } else { 977 ierr = PetscViewerASCIIPrintf(viewer," (%D, %g) ",a->j[j],(double)PetscRealPart(a->a[j]));CHKERRQ(ierr); 978 } 979 #else 980 ierr = PetscViewerASCIIPrintf(viewer," (%D, %g) ",a->j[j],(double)a->a[j]);CHKERRQ(ierr); 981 #endif 982 } 983 ierr = PetscViewerASCIIPrintf(viewer,"\n");CHKERRQ(ierr); 984 } 985 } else { 986 for (i=0; i<m; i++) { 987 ierr = PetscViewerASCIIPrintf(viewer,"row %D:",i);CHKERRQ(ierr); 988 for (j=a->i[i]; j<a->i[i+1]; j++) { 989 #if defined(PETSC_USE_COMPLEX) 990 if (PetscImaginaryPart(a->a[j]) > 0.0) { 991 ierr = PetscViewerASCIIPrintf(viewer," (%D, %g + %g i)",a->j[j],(double)PetscRealPart(a->a[j]),(double)PetscImaginaryPart(a->a[j]));CHKERRQ(ierr); 992 } else if (PetscImaginaryPart(a->a[j]) < 0.0) { 993 ierr = PetscViewerASCIIPrintf(viewer," (%D, %g - %g i)",a->j[j],(double)PetscRealPart(a->a[j]),(double)-PetscImaginaryPart(a->a[j]));CHKERRQ(ierr); 994 } else { 995 ierr = PetscViewerASCIIPrintf(viewer," (%D, %g) ",a->j[j],(double)PetscRealPart(a->a[j]));CHKERRQ(ierr); 996 } 997 #else 998 ierr = PetscViewerASCIIPrintf(viewer," (%D, %g) ",a->j[j],(double)a->a[j]);CHKERRQ(ierr); 999 #endif 1000 } 1001 ierr = PetscViewerASCIIPrintf(viewer,"\n");CHKERRQ(ierr); 1002 } 1003 } 1004 ierr = PetscViewerASCIIUseTabs(viewer,PETSC_TRUE);CHKERRQ(ierr); 1005 } 1006 ierr = PetscViewerFlush(viewer);CHKERRQ(ierr); 1007 PetscFunctionReturn(0); 1008 } 1009 1010 #include <petscdraw.h> 1011 PetscErrorCode MatView_SeqAIJ_Draw_Zoom(PetscDraw draw,void *Aa) 1012 { 1013 Mat A = (Mat) Aa; 1014 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 1015 PetscErrorCode ierr; 1016 PetscInt i,j,m = A->rmap->n; 1017 int color; 1018 PetscReal xl,yl,xr,yr,x_l,x_r,y_l,y_r; 1019 PetscViewer viewer; 1020 PetscViewerFormat format; 1021 1022 PetscFunctionBegin; 1023 ierr = PetscObjectQuery((PetscObject)A,"Zoomviewer",(PetscObject*)&viewer);CHKERRQ(ierr); 1024 ierr = PetscViewerGetFormat(viewer,&format);CHKERRQ(ierr); 1025 ierr = PetscDrawGetCoordinates(draw,&xl,&yl,&xr,&yr);CHKERRQ(ierr); 1026 1027 /* loop over matrix elements drawing boxes */ 1028 1029 if (format != PETSC_VIEWER_DRAW_CONTOUR) { 1030 ierr = PetscDrawCollectiveBegin(draw);CHKERRQ(ierr); 1031 /* Blue for negative, Cyan for zero and Red for positive */ 1032 color = PETSC_DRAW_BLUE; 1033 for (i=0; i<m; i++) { 1034 y_l = m - i - 1.0; y_r = y_l + 1.0; 1035 for (j=a->i[i]; j<a->i[i+1]; j++) { 1036 x_l = a->j[j]; x_r = x_l + 1.0; 1037 if (PetscRealPart(a->a[j]) >= 0.) continue; 1038 ierr = PetscDrawRectangle(draw,x_l,y_l,x_r,y_r,color,color,color,color);CHKERRQ(ierr); 1039 } 1040 } 1041 color = PETSC_DRAW_CYAN; 1042 for (i=0; i<m; i++) { 1043 y_l = m - i - 1.0; y_r = y_l + 1.0; 1044 for (j=a->i[i]; j<a->i[i+1]; j++) { 1045 x_l = a->j[j]; x_r = x_l + 1.0; 1046 if (a->a[j] != 0.) continue; 1047 ierr = PetscDrawRectangle(draw,x_l,y_l,x_r,y_r,color,color,color,color);CHKERRQ(ierr); 1048 } 1049 } 1050 color = PETSC_DRAW_RED; 1051 for (i=0; i<m; i++) { 1052 y_l = m - i - 1.0; y_r = y_l + 1.0; 1053 for (j=a->i[i]; j<a->i[i+1]; j++) { 1054 x_l = a->j[j]; x_r = x_l + 1.0; 1055 if (PetscRealPart(a->a[j]) <= 0.) continue; 1056 ierr = PetscDrawRectangle(draw,x_l,y_l,x_r,y_r,color,color,color,color);CHKERRQ(ierr); 1057 } 1058 } 1059 ierr = PetscDrawCollectiveEnd(draw);CHKERRQ(ierr); 1060 } else { 1061 /* use contour shading to indicate magnitude of values */ 1062 /* first determine max of all nonzero values */ 1063 PetscReal minv = 0.0, maxv = 0.0; 1064 PetscInt nz = a->nz, count = 0; 1065 PetscDraw popup; 1066 1067 for (i=0; i<nz; i++) { 1068 if (PetscAbsScalar(a->a[i]) > maxv) maxv = PetscAbsScalar(a->a[i]); 1069 } 1070 if (minv >= maxv) maxv = minv + PETSC_SMALL; 1071 ierr = PetscDrawGetPopup(draw,&popup);CHKERRQ(ierr); 1072 ierr = PetscDrawScalePopup(popup,minv,maxv);CHKERRQ(ierr); 1073 1074 ierr = PetscDrawCollectiveBegin(draw);CHKERRQ(ierr); 1075 for (i=0; i<m; i++) { 1076 y_l = m - i - 1.0; 1077 y_r = y_l + 1.0; 1078 for (j=a->i[i]; j<a->i[i+1]; j++) { 1079 x_l = a->j[j]; 1080 x_r = x_l + 1.0; 1081 color = PetscDrawRealToColor(PetscAbsScalar(a->a[count]),minv,maxv); 1082 ierr = PetscDrawRectangle(draw,x_l,y_l,x_r,y_r,color,color,color,color);CHKERRQ(ierr); 1083 count++; 1084 } 1085 } 1086 ierr = PetscDrawCollectiveEnd(draw);CHKERRQ(ierr); 1087 } 1088 PetscFunctionReturn(0); 1089 } 1090 1091 #include <petscdraw.h> 1092 PetscErrorCode MatView_SeqAIJ_Draw(Mat A,PetscViewer viewer) 1093 { 1094 PetscErrorCode ierr; 1095 PetscDraw draw; 1096 PetscReal xr,yr,xl,yl,h,w; 1097 PetscBool isnull; 1098 1099 PetscFunctionBegin; 1100 ierr = PetscViewerDrawGetDraw(viewer,0,&draw);CHKERRQ(ierr); 1101 ierr = PetscDrawIsNull(draw,&isnull);CHKERRQ(ierr); 1102 if (isnull) PetscFunctionReturn(0); 1103 1104 xr = A->cmap->n; yr = A->rmap->n; h = yr/10.0; w = xr/10.0; 1105 xr += w; yr += h; xl = -w; yl = -h; 1106 ierr = PetscDrawSetCoordinates(draw,xl,yl,xr,yr);CHKERRQ(ierr); 1107 ierr = PetscObjectCompose((PetscObject)A,"Zoomviewer",(PetscObject)viewer);CHKERRQ(ierr); 1108 ierr = PetscDrawZoom(draw,MatView_SeqAIJ_Draw_Zoom,A);CHKERRQ(ierr); 1109 ierr = PetscObjectCompose((PetscObject)A,"Zoomviewer",NULL);CHKERRQ(ierr); 1110 ierr = PetscDrawSave(draw);CHKERRQ(ierr); 1111 PetscFunctionReturn(0); 1112 } 1113 1114 PetscErrorCode MatView_SeqAIJ(Mat A,PetscViewer viewer) 1115 { 1116 PetscErrorCode ierr; 1117 PetscBool iascii,isbinary,isdraw; 1118 1119 PetscFunctionBegin; 1120 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii);CHKERRQ(ierr); 1121 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERBINARY,&isbinary);CHKERRQ(ierr); 1122 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERDRAW,&isdraw);CHKERRQ(ierr); 1123 if (iascii) { 1124 ierr = MatView_SeqAIJ_ASCII(A,viewer);CHKERRQ(ierr); 1125 } else if (isbinary) { 1126 ierr = MatView_SeqAIJ_Binary(A,viewer);CHKERRQ(ierr); 1127 } else if (isdraw) { 1128 ierr = MatView_SeqAIJ_Draw(A,viewer);CHKERRQ(ierr); 1129 } 1130 ierr = MatView_SeqAIJ_Inode(A,viewer);CHKERRQ(ierr); 1131 PetscFunctionReturn(0); 1132 } 1133 1134 PetscErrorCode MatAssemblyEnd_SeqAIJ(Mat A,MatAssemblyType mode) 1135 { 1136 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 1137 PetscErrorCode ierr; 1138 PetscInt fshift = 0,i,*ai = a->i,*aj = a->j,*imax = a->imax; 1139 PetscInt m = A->rmap->n,*ip,N,*ailen = a->ilen,rmax = 0; 1140 MatScalar *aa = a->a,*ap; 1141 PetscReal ratio = 0.6; 1142 1143 PetscFunctionBegin; 1144 if (mode == MAT_FLUSH_ASSEMBLY) PetscFunctionReturn(0); 1145 ierr = MatSeqAIJInvalidateDiagonal(A);CHKERRQ(ierr); 1146 if (A->was_assembled && A->ass_nonzerostate == A->nonzerostate) PetscFunctionReturn(0); 1147 1148 if (m) rmax = ailen[0]; /* determine row with most nonzeros */ 1149 for (i=1; i<m; i++) { 1150 /* move each row back by the amount of empty slots (fshift) before it*/ 1151 fshift += imax[i-1] - ailen[i-1]; 1152 rmax = PetscMax(rmax,ailen[i]); 1153 if (fshift) { 1154 ip = aj + ai[i]; 1155 ap = aa + ai[i]; 1156 N = ailen[i]; 1157 ierr = PetscArraymove(ip-fshift,ip,N);CHKERRQ(ierr); 1158 if (!A->structure_only) { 1159 ierr = PetscArraymove(ap-fshift,ap,N);CHKERRQ(ierr); 1160 } 1161 } 1162 ai[i] = ai[i-1] + ailen[i-1]; 1163 } 1164 if (m) { 1165 fshift += imax[m-1] - ailen[m-1]; 1166 ai[m] = ai[m-1] + ailen[m-1]; 1167 } 1168 1169 /* reset ilen and imax for each row */ 1170 a->nonzerorowcnt = 0; 1171 if (A->structure_only) { 1172 ierr = PetscFree(a->imax);CHKERRQ(ierr); 1173 ierr = PetscFree(a->ilen);CHKERRQ(ierr); 1174 } else { /* !A->structure_only */ 1175 for (i=0; i<m; i++) { 1176 ailen[i] = imax[i] = ai[i+1] - ai[i]; 1177 a->nonzerorowcnt += ((ai[i+1] - ai[i]) > 0); 1178 } 1179 } 1180 a->nz = ai[m]; 1181 if (fshift && a->nounused == -1) SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_PLIB, "Unused space detected in matrix: %D X %D, %D unneeded", m, A->cmap->n, fshift); 1182 1183 ierr = MatMarkDiagonal_SeqAIJ(A);CHKERRQ(ierr); 1184 ierr = PetscInfo4(A,"Matrix size: %D X %D; storage space: %D unneeded,%D used\n",m,A->cmap->n,fshift,a->nz);CHKERRQ(ierr); 1185 ierr = PetscInfo1(A,"Number of mallocs during MatSetValues() is %D\n",a->reallocs);CHKERRQ(ierr); 1186 ierr = PetscInfo1(A,"Maximum nonzeros in any row is %D\n",rmax);CHKERRQ(ierr); 1187 1188 A->info.mallocs += a->reallocs; 1189 a->reallocs = 0; 1190 A->info.nz_unneeded = (PetscReal)fshift; 1191 a->rmax = rmax; 1192 1193 if (!A->structure_only) { 1194 ierr = MatCheckCompressedRow(A,a->nonzerorowcnt,&a->compressedrow,a->i,m,ratio);CHKERRQ(ierr); 1195 } 1196 ierr = MatAssemblyEnd_SeqAIJ_Inode(A,mode);CHKERRQ(ierr); 1197 PetscFunctionReturn(0); 1198 } 1199 1200 PetscErrorCode MatRealPart_SeqAIJ(Mat A) 1201 { 1202 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 1203 PetscInt i,nz = a->nz; 1204 MatScalar *aa = a->a; 1205 PetscErrorCode ierr; 1206 1207 PetscFunctionBegin; 1208 for (i=0; i<nz; i++) aa[i] = PetscRealPart(aa[i]); 1209 ierr = MatSeqAIJInvalidateDiagonal(A);CHKERRQ(ierr); 1210 #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA) 1211 if (A->offloadmask != PETSC_OFFLOAD_UNALLOCATED) A->offloadmask = PETSC_OFFLOAD_CPU; 1212 #endif 1213 PetscFunctionReturn(0); 1214 } 1215 1216 PetscErrorCode MatImaginaryPart_SeqAIJ(Mat A) 1217 { 1218 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 1219 PetscInt i,nz = a->nz; 1220 MatScalar *aa = a->a; 1221 PetscErrorCode ierr; 1222 1223 PetscFunctionBegin; 1224 for (i=0; i<nz; i++) aa[i] = PetscImaginaryPart(aa[i]); 1225 ierr = MatSeqAIJInvalidateDiagonal(A);CHKERRQ(ierr); 1226 #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA) 1227 if (A->offloadmask != PETSC_OFFLOAD_UNALLOCATED) A->offloadmask = PETSC_OFFLOAD_CPU; 1228 #endif 1229 PetscFunctionReturn(0); 1230 } 1231 1232 PetscErrorCode MatZeroEntries_SeqAIJ(Mat A) 1233 { 1234 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 1235 PetscErrorCode ierr; 1236 1237 PetscFunctionBegin; 1238 ierr = PetscArrayzero(a->a,a->i[A->rmap->n]);CHKERRQ(ierr); 1239 ierr = MatSeqAIJInvalidateDiagonal(A);CHKERRQ(ierr); 1240 #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA) 1241 if (A->offloadmask != PETSC_OFFLOAD_UNALLOCATED) A->offloadmask = PETSC_OFFLOAD_CPU; 1242 #endif 1243 PetscFunctionReturn(0); 1244 } 1245 1246 PetscErrorCode MatDestroy_SeqAIJ(Mat A) 1247 { 1248 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 1249 PetscErrorCode ierr; 1250 1251 PetscFunctionBegin; 1252 #if defined(PETSC_USE_LOG) 1253 PetscLogObjectState((PetscObject)A,"Rows=%D, Cols=%D, NZ=%D",A->rmap->n,A->cmap->n,a->nz); 1254 #endif 1255 ierr = MatSeqXAIJFreeAIJ(A,&a->a,&a->j,&a->i);CHKERRQ(ierr); 1256 ierr = ISDestroy(&a->row);CHKERRQ(ierr); 1257 ierr = ISDestroy(&a->col);CHKERRQ(ierr); 1258 ierr = PetscFree(a->diag);CHKERRQ(ierr); 1259 ierr = PetscFree(a->ibdiag);CHKERRQ(ierr); 1260 ierr = PetscFree(a->imax);CHKERRQ(ierr); 1261 ierr = PetscFree(a->ilen);CHKERRQ(ierr); 1262 ierr = PetscFree(a->ipre);CHKERRQ(ierr); 1263 ierr = PetscFree3(a->idiag,a->mdiag,a->ssor_work);CHKERRQ(ierr); 1264 ierr = PetscFree(a->solve_work);CHKERRQ(ierr); 1265 ierr = ISDestroy(&a->icol);CHKERRQ(ierr); 1266 ierr = PetscFree(a->saved_values);CHKERRQ(ierr); 1267 ierr = PetscFree2(a->compressedrow.i,a->compressedrow.rindex);CHKERRQ(ierr); 1268 1269 ierr = MatDestroy_SeqAIJ_Inode(A);CHKERRQ(ierr); 1270 ierr = PetscFree(A->data);CHKERRQ(ierr); 1271 1272 /* MatMatMultNumeric_SeqAIJ_SeqAIJ_Sorted may allocate this. 1273 That function is so heavily used (sometimes in an hidden way through multnumeric function pointers) 1274 that is hard to properly add this data to the MatProduct data. We free it here to avoid 1275 users reusing the matrix object with different data to incur in obscure segmentation faults 1276 due to different matrix sizes */ 1277 ierr = PetscObjectCompose((PetscObject)A,"__PETSc__ab_dense",NULL);CHKERRQ(ierr); 1278 1279 ierr = PetscObjectChangeTypeName((PetscObject)A,NULL);CHKERRQ(ierr); 1280 ierr = PetscObjectComposeFunction((PetscObject)A,"MatSeqAIJSetColumnIndices_C",NULL);CHKERRQ(ierr); 1281 ierr = PetscObjectComposeFunction((PetscObject)A,"MatStoreValues_C",NULL);CHKERRQ(ierr); 1282 ierr = PetscObjectComposeFunction((PetscObject)A,"MatRetrieveValues_C",NULL);CHKERRQ(ierr); 1283 ierr = PetscObjectComposeFunction((PetscObject)A,"MatConvert_seqaij_seqsbaij_C",NULL);CHKERRQ(ierr); 1284 ierr = PetscObjectComposeFunction((PetscObject)A,"MatConvert_seqaij_seqbaij_C",NULL);CHKERRQ(ierr); 1285 ierr = PetscObjectComposeFunction((PetscObject)A,"MatConvert_seqaij_seqaijperm_C",NULL);CHKERRQ(ierr); 1286 1287 #if defined(PETSC_HAVE_CUDA) 1288 ierr = PetscObjectComposeFunction((PetscObject)A,"MatConvert_seqaij_seqaijcusparse_C",NULL);CHKERRQ(ierr); 1289 ierr = PetscObjectComposeFunction((PetscObject)A,"MatProductSetFromOptions_seqaijcusparse_seqaij_C",NULL);CHKERRQ(ierr); 1290 #endif 1291 ierr = PetscObjectComposeFunction((PetscObject)A,"MatConvert_seqaij_seqaijcrl_C",NULL);CHKERRQ(ierr); 1292 #if defined(PETSC_HAVE_ELEMENTAL) 1293 ierr = PetscObjectComposeFunction((PetscObject)A,"MatConvert_seqaij_elemental_C",NULL);CHKERRQ(ierr); 1294 #endif 1295 #if defined(PETSC_HAVE_SCALAPACK) 1296 ierr = PetscObjectComposeFunction((PetscObject)A,"MatConvert_seqaij_scalapack_C",NULL);CHKERRQ(ierr); 1297 #endif 1298 #if defined(PETSC_HAVE_HYPRE) 1299 ierr = PetscObjectComposeFunction((PetscObject)A,"MatConvert_seqaij_hypre_C",NULL);CHKERRQ(ierr); 1300 ierr = PetscObjectComposeFunction((PetscObject)A,"MatProductSetFromOptions_transpose_seqaij_seqaij_C",NULL);CHKERRQ(ierr); 1301 #endif 1302 ierr = PetscObjectComposeFunction((PetscObject)A,"MatConvert_seqaij_seqdense_C",NULL);CHKERRQ(ierr); 1303 ierr = PetscObjectComposeFunction((PetscObject)A,"MatConvert_seqaij_seqsell_C",NULL);CHKERRQ(ierr); 1304 ierr = PetscObjectComposeFunction((PetscObject)A,"MatConvert_seqaij_is_C",NULL);CHKERRQ(ierr); 1305 ierr = PetscObjectComposeFunction((PetscObject)A,"MatIsTranspose_C",NULL);CHKERRQ(ierr); 1306 ierr = PetscObjectComposeFunction((PetscObject)A,"MatSeqAIJSetPreallocation_C",NULL);CHKERRQ(ierr); 1307 ierr = PetscObjectComposeFunction((PetscObject)A,"MatResetPreallocation_C",NULL);CHKERRQ(ierr); 1308 ierr = PetscObjectComposeFunction((PetscObject)A,"MatSeqAIJSetPreallocationCSR_C",NULL);CHKERRQ(ierr); 1309 ierr = PetscObjectComposeFunction((PetscObject)A,"MatReorderForNonzeroDiagonal_C",NULL);CHKERRQ(ierr); 1310 ierr = PetscObjectComposeFunction((PetscObject)A,"MatProductSetFromOptions_is_seqaij_C",NULL);CHKERRQ(ierr); 1311 ierr = PetscObjectComposeFunction((PetscObject)A,"MatProductSetFromOptions_seqdense_seqaij_C",NULL);CHKERRQ(ierr); 1312 ierr = PetscObjectComposeFunction((PetscObject)A,"MatProductSetFromOptions_seqaij_seqaij_C",NULL);CHKERRQ(ierr); 1313 PetscFunctionReturn(0); 1314 } 1315 1316 PetscErrorCode MatSetOption_SeqAIJ(Mat A,MatOption op,PetscBool flg) 1317 { 1318 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 1319 PetscErrorCode ierr; 1320 1321 PetscFunctionBegin; 1322 switch (op) { 1323 case MAT_ROW_ORIENTED: 1324 a->roworiented = flg; 1325 break; 1326 case MAT_KEEP_NONZERO_PATTERN: 1327 a->keepnonzeropattern = flg; 1328 break; 1329 case MAT_NEW_NONZERO_LOCATIONS: 1330 a->nonew = (flg ? 0 : 1); 1331 break; 1332 case MAT_NEW_NONZERO_LOCATION_ERR: 1333 a->nonew = (flg ? -1 : 0); 1334 break; 1335 case MAT_NEW_NONZERO_ALLOCATION_ERR: 1336 a->nonew = (flg ? -2 : 0); 1337 break; 1338 case MAT_UNUSED_NONZERO_LOCATION_ERR: 1339 a->nounused = (flg ? -1 : 0); 1340 break; 1341 case MAT_IGNORE_ZERO_ENTRIES: 1342 a->ignorezeroentries = flg; 1343 break; 1344 case MAT_SPD: 1345 case MAT_SYMMETRIC: 1346 case MAT_STRUCTURALLY_SYMMETRIC: 1347 case MAT_HERMITIAN: 1348 case MAT_SYMMETRY_ETERNAL: 1349 case MAT_STRUCTURE_ONLY: 1350 /* These options are handled directly by MatSetOption() */ 1351 break; 1352 case MAT_NEW_DIAGONALS: 1353 case MAT_IGNORE_OFF_PROC_ENTRIES: 1354 case MAT_USE_HASH_TABLE: 1355 ierr = PetscInfo1(A,"Option %s ignored\n",MatOptions[op]);CHKERRQ(ierr); 1356 break; 1357 case MAT_USE_INODES: 1358 /* Not an error because MatSetOption_SeqAIJ_Inode handles this one */ 1359 break; 1360 case MAT_SUBMAT_SINGLEIS: 1361 A->submat_singleis = flg; 1362 break; 1363 case MAT_SORTED_FULL: 1364 if (flg) A->ops->setvalues = MatSetValues_SeqAIJ_SortedFull; 1365 else A->ops->setvalues = MatSetValues_SeqAIJ; 1366 break; 1367 default: 1368 SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"unknown option %d",op); 1369 } 1370 ierr = MatSetOption_SeqAIJ_Inode(A,op,flg);CHKERRQ(ierr); 1371 PetscFunctionReturn(0); 1372 } 1373 1374 PetscErrorCode MatGetDiagonal_SeqAIJ(Mat A,Vec v) 1375 { 1376 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 1377 PetscErrorCode ierr; 1378 PetscInt i,j,n,*ai=a->i,*aj=a->j; 1379 PetscScalar *aa=a->a,*x; 1380 1381 PetscFunctionBegin; 1382 ierr = VecGetLocalSize(v,&n);CHKERRQ(ierr); 1383 if (n != A->rmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Nonconforming matrix and vector"); 1384 1385 if (A->factortype == MAT_FACTOR_ILU || A->factortype == MAT_FACTOR_LU) { 1386 PetscInt *diag=a->diag; 1387 ierr = VecGetArrayWrite(v,&x);CHKERRQ(ierr); 1388 for (i=0; i<n; i++) x[i] = 1.0/aa[diag[i]]; 1389 ierr = VecRestoreArrayWrite(v,&x);CHKERRQ(ierr); 1390 PetscFunctionReturn(0); 1391 } 1392 1393 ierr = VecGetArrayWrite(v,&x);CHKERRQ(ierr); 1394 for (i=0; i<n; i++) { 1395 x[i] = 0.0; 1396 for (j=ai[i]; j<ai[i+1]; j++) { 1397 if (aj[j] == i) { 1398 x[i] = aa[j]; 1399 break; 1400 } 1401 } 1402 } 1403 ierr = VecRestoreArrayWrite(v,&x);CHKERRQ(ierr); 1404 PetscFunctionReturn(0); 1405 } 1406 1407 #include <../src/mat/impls/aij/seq/ftn-kernels/fmult.h> 1408 PetscErrorCode MatMultTransposeAdd_SeqAIJ(Mat A,Vec xx,Vec zz,Vec yy) 1409 { 1410 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 1411 PetscScalar *y; 1412 const PetscScalar *x; 1413 PetscErrorCode ierr; 1414 PetscInt m = A->rmap->n; 1415 #if !defined(PETSC_USE_FORTRAN_KERNEL_MULTTRANSPOSEAIJ) 1416 const MatScalar *v; 1417 PetscScalar alpha; 1418 PetscInt n,i,j; 1419 const PetscInt *idx,*ii,*ridx=NULL; 1420 Mat_CompressedRow cprow = a->compressedrow; 1421 PetscBool usecprow = cprow.use; 1422 #endif 1423 1424 PetscFunctionBegin; 1425 if (zz != yy) {ierr = VecCopy(zz,yy);CHKERRQ(ierr);} 1426 ierr = VecGetArrayRead(xx,&x);CHKERRQ(ierr); 1427 ierr = VecGetArray(yy,&y);CHKERRQ(ierr); 1428 1429 #if defined(PETSC_USE_FORTRAN_KERNEL_MULTTRANSPOSEAIJ) 1430 fortranmulttransposeaddaij_(&m,x,a->i,a->j,a->a,y); 1431 #else 1432 if (usecprow) { 1433 m = cprow.nrows; 1434 ii = cprow.i; 1435 ridx = cprow.rindex; 1436 } else { 1437 ii = a->i; 1438 } 1439 for (i=0; i<m; i++) { 1440 idx = a->j + ii[i]; 1441 v = a->a + ii[i]; 1442 n = ii[i+1] - ii[i]; 1443 if (usecprow) { 1444 alpha = x[ridx[i]]; 1445 } else { 1446 alpha = x[i]; 1447 } 1448 for (j=0; j<n; j++) y[idx[j]] += alpha*v[j]; 1449 } 1450 #endif 1451 ierr = PetscLogFlops(2.0*a->nz);CHKERRQ(ierr); 1452 ierr = VecRestoreArrayRead(xx,&x);CHKERRQ(ierr); 1453 ierr = VecRestoreArray(yy,&y);CHKERRQ(ierr); 1454 PetscFunctionReturn(0); 1455 } 1456 1457 PetscErrorCode MatMultTranspose_SeqAIJ(Mat A,Vec xx,Vec yy) 1458 { 1459 PetscErrorCode ierr; 1460 1461 PetscFunctionBegin; 1462 ierr = VecSet(yy,0.0);CHKERRQ(ierr); 1463 ierr = MatMultTransposeAdd_SeqAIJ(A,xx,yy,yy);CHKERRQ(ierr); 1464 PetscFunctionReturn(0); 1465 } 1466 1467 #include <../src/mat/impls/aij/seq/ftn-kernels/fmult.h> 1468 1469 PetscErrorCode MatMult_SeqAIJ(Mat A,Vec xx,Vec yy) 1470 { 1471 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 1472 PetscScalar *y; 1473 const PetscScalar *x; 1474 const MatScalar *aa; 1475 PetscErrorCode ierr; 1476 PetscInt m=A->rmap->n; 1477 const PetscInt *aj,*ii,*ridx=NULL; 1478 PetscInt n,i; 1479 PetscScalar sum; 1480 PetscBool usecprow=a->compressedrow.use; 1481 1482 #if defined(PETSC_HAVE_PRAGMA_DISJOINT) 1483 #pragma disjoint(*x,*y,*aa) 1484 #endif 1485 1486 PetscFunctionBegin; 1487 ierr = VecGetArrayRead(xx,&x);CHKERRQ(ierr); 1488 ierr = VecGetArray(yy,&y);CHKERRQ(ierr); 1489 ii = a->i; 1490 if (usecprow) { /* use compressed row format */ 1491 ierr = PetscArrayzero(y,m);CHKERRQ(ierr); 1492 m = a->compressedrow.nrows; 1493 ii = a->compressedrow.i; 1494 ridx = a->compressedrow.rindex; 1495 for (i=0; i<m; i++) { 1496 n = ii[i+1] - ii[i]; 1497 aj = a->j + ii[i]; 1498 aa = a->a + ii[i]; 1499 sum = 0.0; 1500 PetscSparseDensePlusDot(sum,x,aa,aj,n); 1501 /* for (j=0; j<n; j++) sum += (*aa++)*x[*aj++]; */ 1502 y[*ridx++] = sum; 1503 } 1504 } else { /* do not use compressed row format */ 1505 #if defined(PETSC_USE_FORTRAN_KERNEL_MULTAIJ) 1506 aj = a->j; 1507 aa = a->a; 1508 fortranmultaij_(&m,x,ii,aj,aa,y); 1509 #else 1510 for (i=0; i<m; i++) { 1511 n = ii[i+1] - ii[i]; 1512 aj = a->j + ii[i]; 1513 aa = a->a + ii[i]; 1514 sum = 0.0; 1515 PetscSparseDensePlusDot(sum,x,aa,aj,n); 1516 y[i] = sum; 1517 } 1518 #endif 1519 } 1520 ierr = PetscLogFlops(2.0*a->nz - a->nonzerorowcnt);CHKERRQ(ierr); 1521 ierr = VecRestoreArrayRead(xx,&x);CHKERRQ(ierr); 1522 ierr = VecRestoreArray(yy,&y);CHKERRQ(ierr); 1523 PetscFunctionReturn(0); 1524 } 1525 1526 PetscErrorCode MatMultMax_SeqAIJ(Mat A,Vec xx,Vec yy) 1527 { 1528 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 1529 PetscScalar *y; 1530 const PetscScalar *x; 1531 const MatScalar *aa; 1532 PetscErrorCode ierr; 1533 PetscInt m=A->rmap->n; 1534 const PetscInt *aj,*ii,*ridx=NULL; 1535 PetscInt n,i,nonzerorow=0; 1536 PetscScalar sum; 1537 PetscBool usecprow=a->compressedrow.use; 1538 1539 #if defined(PETSC_HAVE_PRAGMA_DISJOINT) 1540 #pragma disjoint(*x,*y,*aa) 1541 #endif 1542 1543 PetscFunctionBegin; 1544 ierr = VecGetArrayRead(xx,&x);CHKERRQ(ierr); 1545 ierr = VecGetArray(yy,&y);CHKERRQ(ierr); 1546 if (usecprow) { /* use compressed row format */ 1547 m = a->compressedrow.nrows; 1548 ii = a->compressedrow.i; 1549 ridx = a->compressedrow.rindex; 1550 for (i=0; i<m; i++) { 1551 n = ii[i+1] - ii[i]; 1552 aj = a->j + ii[i]; 1553 aa = a->a + ii[i]; 1554 sum = 0.0; 1555 nonzerorow += (n>0); 1556 PetscSparseDenseMaxDot(sum,x,aa,aj,n); 1557 /* for (j=0; j<n; j++) sum += (*aa++)*x[*aj++]; */ 1558 y[*ridx++] = sum; 1559 } 1560 } else { /* do not use compressed row format */ 1561 ii = a->i; 1562 for (i=0; i<m; i++) { 1563 n = ii[i+1] - ii[i]; 1564 aj = a->j + ii[i]; 1565 aa = a->a + ii[i]; 1566 sum = 0.0; 1567 nonzerorow += (n>0); 1568 PetscSparseDenseMaxDot(sum,x,aa,aj,n); 1569 y[i] = sum; 1570 } 1571 } 1572 ierr = PetscLogFlops(2.0*a->nz - nonzerorow);CHKERRQ(ierr); 1573 ierr = VecRestoreArrayRead(xx,&x);CHKERRQ(ierr); 1574 ierr = VecRestoreArray(yy,&y);CHKERRQ(ierr); 1575 PetscFunctionReturn(0); 1576 } 1577 1578 PetscErrorCode MatMultAddMax_SeqAIJ(Mat A,Vec xx,Vec yy,Vec zz) 1579 { 1580 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 1581 PetscScalar *y,*z; 1582 const PetscScalar *x; 1583 const MatScalar *aa; 1584 PetscErrorCode ierr; 1585 PetscInt m = A->rmap->n,*aj,*ii; 1586 PetscInt n,i,*ridx=NULL; 1587 PetscScalar sum; 1588 PetscBool usecprow=a->compressedrow.use; 1589 1590 PetscFunctionBegin; 1591 ierr = VecGetArrayRead(xx,&x);CHKERRQ(ierr); 1592 ierr = VecGetArrayPair(yy,zz,&y,&z);CHKERRQ(ierr); 1593 if (usecprow) { /* use compressed row format */ 1594 if (zz != yy) { 1595 ierr = PetscArraycpy(z,y,m);CHKERRQ(ierr); 1596 } 1597 m = a->compressedrow.nrows; 1598 ii = a->compressedrow.i; 1599 ridx = a->compressedrow.rindex; 1600 for (i=0; i<m; i++) { 1601 n = ii[i+1] - ii[i]; 1602 aj = a->j + ii[i]; 1603 aa = a->a + ii[i]; 1604 sum = y[*ridx]; 1605 PetscSparseDenseMaxDot(sum,x,aa,aj,n); 1606 z[*ridx++] = sum; 1607 } 1608 } else { /* do not use compressed row format */ 1609 ii = a->i; 1610 for (i=0; i<m; i++) { 1611 n = ii[i+1] - ii[i]; 1612 aj = a->j + ii[i]; 1613 aa = a->a + ii[i]; 1614 sum = y[i]; 1615 PetscSparseDenseMaxDot(sum,x,aa,aj,n); 1616 z[i] = sum; 1617 } 1618 } 1619 ierr = PetscLogFlops(2.0*a->nz);CHKERRQ(ierr); 1620 ierr = VecRestoreArrayRead(xx,&x);CHKERRQ(ierr); 1621 ierr = VecRestoreArrayPair(yy,zz,&y,&z);CHKERRQ(ierr); 1622 PetscFunctionReturn(0); 1623 } 1624 1625 #include <../src/mat/impls/aij/seq/ftn-kernels/fmultadd.h> 1626 PetscErrorCode MatMultAdd_SeqAIJ(Mat A,Vec xx,Vec yy,Vec zz) 1627 { 1628 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 1629 PetscScalar *y,*z; 1630 const PetscScalar *x; 1631 const MatScalar *aa; 1632 PetscErrorCode ierr; 1633 const PetscInt *aj,*ii,*ridx=NULL; 1634 PetscInt m = A->rmap->n,n,i; 1635 PetscScalar sum; 1636 PetscBool usecprow=a->compressedrow.use; 1637 1638 PetscFunctionBegin; 1639 ierr = VecGetArrayRead(xx,&x);CHKERRQ(ierr); 1640 ierr = VecGetArrayPair(yy,zz,&y,&z);CHKERRQ(ierr); 1641 if (usecprow) { /* use compressed row format */ 1642 if (zz != yy) { 1643 ierr = PetscArraycpy(z,y,m);CHKERRQ(ierr); 1644 } 1645 m = a->compressedrow.nrows; 1646 ii = a->compressedrow.i; 1647 ridx = a->compressedrow.rindex; 1648 for (i=0; i<m; i++) { 1649 n = ii[i+1] - ii[i]; 1650 aj = a->j + ii[i]; 1651 aa = a->a + ii[i]; 1652 sum = y[*ridx]; 1653 PetscSparseDensePlusDot(sum,x,aa,aj,n); 1654 z[*ridx++] = sum; 1655 } 1656 } else { /* do not use compressed row format */ 1657 ii = a->i; 1658 #if defined(PETSC_USE_FORTRAN_KERNEL_MULTADDAIJ) 1659 aj = a->j; 1660 aa = a->a; 1661 fortranmultaddaij_(&m,x,ii,aj,aa,y,z); 1662 #else 1663 for (i=0; i<m; i++) { 1664 n = ii[i+1] - ii[i]; 1665 aj = a->j + ii[i]; 1666 aa = a->a + ii[i]; 1667 sum = y[i]; 1668 PetscSparseDensePlusDot(sum,x,aa,aj,n); 1669 z[i] = sum; 1670 } 1671 #endif 1672 } 1673 ierr = PetscLogFlops(2.0*a->nz);CHKERRQ(ierr); 1674 ierr = VecRestoreArrayRead(xx,&x);CHKERRQ(ierr); 1675 ierr = VecRestoreArrayPair(yy,zz,&y,&z);CHKERRQ(ierr); 1676 PetscFunctionReturn(0); 1677 } 1678 1679 /* 1680 Adds diagonal pointers to sparse matrix structure. 1681 */ 1682 PetscErrorCode MatMarkDiagonal_SeqAIJ(Mat A) 1683 { 1684 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 1685 PetscErrorCode ierr; 1686 PetscInt i,j,m = A->rmap->n; 1687 1688 PetscFunctionBegin; 1689 if (!a->diag) { 1690 ierr = PetscMalloc1(m,&a->diag);CHKERRQ(ierr); 1691 ierr = PetscLogObjectMemory((PetscObject)A, m*sizeof(PetscInt));CHKERRQ(ierr); 1692 } 1693 for (i=0; i<A->rmap->n; i++) { 1694 a->diag[i] = a->i[i+1]; 1695 for (j=a->i[i]; j<a->i[i+1]; j++) { 1696 if (a->j[j] == i) { 1697 a->diag[i] = j; 1698 break; 1699 } 1700 } 1701 } 1702 PetscFunctionReturn(0); 1703 } 1704 1705 PetscErrorCode MatShift_SeqAIJ(Mat A,PetscScalar v) 1706 { 1707 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 1708 const PetscInt *diag = (const PetscInt*)a->diag; 1709 const PetscInt *ii = (const PetscInt*) a->i; 1710 PetscInt i,*mdiag = NULL; 1711 PetscErrorCode ierr; 1712 PetscInt cnt = 0; /* how many diagonals are missing */ 1713 1714 PetscFunctionBegin; 1715 if (!A->preallocated || !a->nz) { 1716 ierr = MatSeqAIJSetPreallocation(A,1,NULL);CHKERRQ(ierr); 1717 ierr = MatShift_Basic(A,v);CHKERRQ(ierr); 1718 PetscFunctionReturn(0); 1719 } 1720 1721 if (a->diagonaldense) { 1722 cnt = 0; 1723 } else { 1724 ierr = PetscCalloc1(A->rmap->n,&mdiag);CHKERRQ(ierr); 1725 for (i=0; i<A->rmap->n; i++) { 1726 if (diag[i] >= ii[i+1]) { 1727 cnt++; 1728 mdiag[i] = 1; 1729 } 1730 } 1731 } 1732 if (!cnt) { 1733 ierr = MatShift_Basic(A,v);CHKERRQ(ierr); 1734 } else { 1735 PetscScalar *olda = a->a; /* preserve pointers to current matrix nonzeros structure and values */ 1736 PetscInt *oldj = a->j, *oldi = a->i; 1737 PetscBool singlemalloc = a->singlemalloc,free_a = a->free_a,free_ij = a->free_ij; 1738 1739 a->a = NULL; 1740 a->j = NULL; 1741 a->i = NULL; 1742 /* increase the values in imax for each row where a diagonal is being inserted then reallocate the matrix data structures */ 1743 for (i=0; i<A->rmap->n; i++) { 1744 a->imax[i] += mdiag[i]; 1745 a->imax[i] = PetscMin(a->imax[i],A->cmap->n); 1746 } 1747 ierr = MatSeqAIJSetPreallocation_SeqAIJ(A,0,a->imax);CHKERRQ(ierr); 1748 1749 /* copy old values into new matrix data structure */ 1750 for (i=0; i<A->rmap->n; i++) { 1751 ierr = MatSetValues(A,1,&i,a->imax[i] - mdiag[i],&oldj[oldi[i]],&olda[oldi[i]],ADD_VALUES);CHKERRQ(ierr); 1752 if (i < A->cmap->n) { 1753 ierr = MatSetValue(A,i,i,v,ADD_VALUES);CHKERRQ(ierr); 1754 } 1755 } 1756 ierr = MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1757 ierr = MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1758 if (singlemalloc) { 1759 ierr = PetscFree3(olda,oldj,oldi);CHKERRQ(ierr); 1760 } else { 1761 if (free_a) {ierr = PetscFree(olda);CHKERRQ(ierr);} 1762 if (free_ij) {ierr = PetscFree(oldj);CHKERRQ(ierr);} 1763 if (free_ij) {ierr = PetscFree(oldi);CHKERRQ(ierr);} 1764 } 1765 } 1766 ierr = PetscFree(mdiag);CHKERRQ(ierr); 1767 a->diagonaldense = PETSC_TRUE; 1768 PetscFunctionReturn(0); 1769 } 1770 1771 /* 1772 Checks for missing diagonals 1773 */ 1774 PetscErrorCode MatMissingDiagonal_SeqAIJ(Mat A,PetscBool *missing,PetscInt *d) 1775 { 1776 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 1777 PetscInt *diag,*ii = a->i,i; 1778 PetscErrorCode ierr; 1779 1780 PetscFunctionBegin; 1781 *missing = PETSC_FALSE; 1782 if (A->rmap->n > 0 && !ii) { 1783 *missing = PETSC_TRUE; 1784 if (d) *d = 0; 1785 ierr = PetscInfo(A,"Matrix has no entries therefore is missing diagonal\n");CHKERRQ(ierr); 1786 } else { 1787 PetscInt n; 1788 n = PetscMin(A->rmap->n, A->cmap->n); 1789 diag = a->diag; 1790 for (i=0; i<n; i++) { 1791 if (diag[i] >= ii[i+1]) { 1792 *missing = PETSC_TRUE; 1793 if (d) *d = i; 1794 ierr = PetscInfo1(A,"Matrix is missing diagonal number %D\n",i);CHKERRQ(ierr); 1795 break; 1796 } 1797 } 1798 } 1799 PetscFunctionReturn(0); 1800 } 1801 1802 #include <petscblaslapack.h> 1803 #include <petsc/private/kernels/blockinvert.h> 1804 1805 /* 1806 Note that values is allocated externally by the PC and then passed into this routine 1807 */ 1808 PetscErrorCode MatInvertVariableBlockDiagonal_SeqAIJ(Mat A,PetscInt nblocks,const PetscInt *bsizes,PetscScalar *diag) 1809 { 1810 PetscErrorCode ierr; 1811 PetscInt n = A->rmap->n, i, ncnt = 0, *indx,j,bsizemax = 0,*v_pivots; 1812 PetscBool allowzeropivot,zeropivotdetected=PETSC_FALSE; 1813 const PetscReal shift = 0.0; 1814 PetscInt ipvt[5]; 1815 PetscScalar work[25],*v_work; 1816 1817 PetscFunctionBegin; 1818 allowzeropivot = PetscNot(A->erroriffailure); 1819 for (i=0; i<nblocks; i++) ncnt += bsizes[i]; 1820 if (ncnt != n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Total blocksizes %D doesn't match number matrix rows %D",ncnt,n); 1821 for (i=0; i<nblocks; i++) { 1822 bsizemax = PetscMax(bsizemax,bsizes[i]); 1823 } 1824 ierr = PetscMalloc1(bsizemax,&indx);CHKERRQ(ierr); 1825 if (bsizemax > 7) { 1826 ierr = PetscMalloc2(bsizemax,&v_work,bsizemax,&v_pivots);CHKERRQ(ierr); 1827 } 1828 ncnt = 0; 1829 for (i=0; i<nblocks; i++) { 1830 for (j=0; j<bsizes[i]; j++) indx[j] = ncnt+j; 1831 ierr = MatGetValues(A,bsizes[i],indx,bsizes[i],indx,diag);CHKERRQ(ierr); 1832 switch (bsizes[i]) { 1833 case 1: 1834 *diag = 1.0/(*diag); 1835 break; 1836 case 2: 1837 ierr = PetscKernel_A_gets_inverse_A_2(diag,shift,allowzeropivot,&zeropivotdetected);CHKERRQ(ierr); 1838 if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT; 1839 ierr = PetscKernel_A_gets_transpose_A_2(diag);CHKERRQ(ierr); 1840 break; 1841 case 3: 1842 ierr = PetscKernel_A_gets_inverse_A_3(diag,shift,allowzeropivot,&zeropivotdetected);CHKERRQ(ierr); 1843 if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT; 1844 ierr = PetscKernel_A_gets_transpose_A_3(diag);CHKERRQ(ierr); 1845 break; 1846 case 4: 1847 ierr = PetscKernel_A_gets_inverse_A_4(diag,shift,allowzeropivot,&zeropivotdetected);CHKERRQ(ierr); 1848 if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT; 1849 ierr = PetscKernel_A_gets_transpose_A_4(diag);CHKERRQ(ierr); 1850 break; 1851 case 5: 1852 ierr = PetscKernel_A_gets_inverse_A_5(diag,ipvt,work,shift,allowzeropivot,&zeropivotdetected);CHKERRQ(ierr); 1853 if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT; 1854 ierr = PetscKernel_A_gets_transpose_A_5(diag);CHKERRQ(ierr); 1855 break; 1856 case 6: 1857 ierr = PetscKernel_A_gets_inverse_A_6(diag,shift,allowzeropivot,&zeropivotdetected);CHKERRQ(ierr); 1858 if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT; 1859 ierr = PetscKernel_A_gets_transpose_A_6(diag);CHKERRQ(ierr); 1860 break; 1861 case 7: 1862 ierr = PetscKernel_A_gets_inverse_A_7(diag,shift,allowzeropivot,&zeropivotdetected);CHKERRQ(ierr); 1863 if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT; 1864 ierr = PetscKernel_A_gets_transpose_A_7(diag);CHKERRQ(ierr); 1865 break; 1866 default: 1867 ierr = PetscKernel_A_gets_inverse_A(bsizes[i],diag,v_pivots,v_work,allowzeropivot,&zeropivotdetected);CHKERRQ(ierr); 1868 if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT; 1869 ierr = PetscKernel_A_gets_transpose_A_N(diag,bsizes[i]);CHKERRQ(ierr); 1870 } 1871 ncnt += bsizes[i]; 1872 diag += bsizes[i]*bsizes[i]; 1873 } 1874 if (bsizemax > 7) { 1875 ierr = PetscFree2(v_work,v_pivots);CHKERRQ(ierr); 1876 } 1877 ierr = PetscFree(indx);CHKERRQ(ierr); 1878 PetscFunctionReturn(0); 1879 } 1880 1881 /* 1882 Negative shift indicates do not generate an error if there is a zero diagonal, just invert it anyways 1883 */ 1884 PetscErrorCode MatInvertDiagonal_SeqAIJ(Mat A,PetscScalar omega,PetscScalar fshift) 1885 { 1886 Mat_SeqAIJ *a = (Mat_SeqAIJ*) A->data; 1887 PetscErrorCode ierr; 1888 PetscInt i,*diag,m = A->rmap->n; 1889 MatScalar *v = a->a; 1890 PetscScalar *idiag,*mdiag; 1891 1892 PetscFunctionBegin; 1893 if (a->idiagvalid) PetscFunctionReturn(0); 1894 ierr = MatMarkDiagonal_SeqAIJ(A);CHKERRQ(ierr); 1895 diag = a->diag; 1896 if (!a->idiag) { 1897 ierr = PetscMalloc3(m,&a->idiag,m,&a->mdiag,m,&a->ssor_work);CHKERRQ(ierr); 1898 ierr = PetscLogObjectMemory((PetscObject)A, 3*m*sizeof(PetscScalar));CHKERRQ(ierr); 1899 v = a->a; 1900 } 1901 mdiag = a->mdiag; 1902 idiag = a->idiag; 1903 1904 if (omega == 1.0 && PetscRealPart(fshift) <= 0.0) { 1905 for (i=0; i<m; i++) { 1906 mdiag[i] = v[diag[i]]; 1907 if (!PetscAbsScalar(mdiag[i])) { /* zero diagonal */ 1908 if (PetscRealPart(fshift)) { 1909 ierr = PetscInfo1(A,"Zero diagonal on row %D\n",i);CHKERRQ(ierr); 1910 A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT; 1911 A->factorerror_zeropivot_value = 0.0; 1912 A->factorerror_zeropivot_row = i; 1913 } else SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_INCOMP,"Zero diagonal on row %D",i); 1914 } 1915 idiag[i] = 1.0/v[diag[i]]; 1916 } 1917 ierr = PetscLogFlops(m);CHKERRQ(ierr); 1918 } else { 1919 for (i=0; i<m; i++) { 1920 mdiag[i] = v[diag[i]]; 1921 idiag[i] = omega/(fshift + v[diag[i]]); 1922 } 1923 ierr = PetscLogFlops(2.0*m);CHKERRQ(ierr); 1924 } 1925 a->idiagvalid = PETSC_TRUE; 1926 PetscFunctionReturn(0); 1927 } 1928 1929 #include <../src/mat/impls/aij/seq/ftn-kernels/frelax.h> 1930 PetscErrorCode MatSOR_SeqAIJ(Mat A,Vec bb,PetscReal omega,MatSORType flag,PetscReal fshift,PetscInt its,PetscInt lits,Vec xx) 1931 { 1932 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 1933 PetscScalar *x,d,sum,*t,scale; 1934 const MatScalar *v,*idiag=NULL,*mdiag; 1935 const PetscScalar *b, *bs,*xb, *ts; 1936 PetscErrorCode ierr; 1937 PetscInt n,m = A->rmap->n,i; 1938 const PetscInt *idx,*diag; 1939 1940 PetscFunctionBegin; 1941 its = its*lits; 1942 1943 if (fshift != a->fshift || omega != a->omega) a->idiagvalid = PETSC_FALSE; /* must recompute idiag[] */ 1944 if (!a->idiagvalid) {ierr = MatInvertDiagonal_SeqAIJ(A,omega,fshift);CHKERRQ(ierr);} 1945 a->fshift = fshift; 1946 a->omega = omega; 1947 1948 diag = a->diag; 1949 t = a->ssor_work; 1950 idiag = a->idiag; 1951 mdiag = a->mdiag; 1952 1953 ierr = VecGetArray(xx,&x);CHKERRQ(ierr); 1954 ierr = VecGetArrayRead(bb,&b);CHKERRQ(ierr); 1955 /* We count flops by assuming the upper triangular and lower triangular parts have the same number of nonzeros */ 1956 if (flag == SOR_APPLY_UPPER) { 1957 /* apply (U + D/omega) to the vector */ 1958 bs = b; 1959 for (i=0; i<m; i++) { 1960 d = fshift + mdiag[i]; 1961 n = a->i[i+1] - diag[i] - 1; 1962 idx = a->j + diag[i] + 1; 1963 v = a->a + diag[i] + 1; 1964 sum = b[i]*d/omega; 1965 PetscSparseDensePlusDot(sum,bs,v,idx,n); 1966 x[i] = sum; 1967 } 1968 ierr = VecRestoreArray(xx,&x);CHKERRQ(ierr); 1969 ierr = VecRestoreArrayRead(bb,&b);CHKERRQ(ierr); 1970 ierr = PetscLogFlops(a->nz);CHKERRQ(ierr); 1971 PetscFunctionReturn(0); 1972 } 1973 1974 if (flag == SOR_APPLY_LOWER) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"SOR_APPLY_LOWER is not implemented"); 1975 else if (flag & SOR_EISENSTAT) { 1976 /* Let A = L + U + D; where L is lower triangular, 1977 U is upper triangular, E = D/omega; This routine applies 1978 1979 (L + E)^{-1} A (U + E)^{-1} 1980 1981 to a vector efficiently using Eisenstat's trick. 1982 */ 1983 scale = (2.0/omega) - 1.0; 1984 1985 /* x = (E + U)^{-1} b */ 1986 for (i=m-1; i>=0; i--) { 1987 n = a->i[i+1] - diag[i] - 1; 1988 idx = a->j + diag[i] + 1; 1989 v = a->a + diag[i] + 1; 1990 sum = b[i]; 1991 PetscSparseDenseMinusDot(sum,x,v,idx,n); 1992 x[i] = sum*idiag[i]; 1993 } 1994 1995 /* t = b - (2*E - D)x */ 1996 v = a->a; 1997 for (i=0; i<m; i++) t[i] = b[i] - scale*(v[*diag++])*x[i]; 1998 1999 /* t = (E + L)^{-1}t */ 2000 ts = t; 2001 diag = a->diag; 2002 for (i=0; i<m; i++) { 2003 n = diag[i] - a->i[i]; 2004 idx = a->j + a->i[i]; 2005 v = a->a + a->i[i]; 2006 sum = t[i]; 2007 PetscSparseDenseMinusDot(sum,ts,v,idx,n); 2008 t[i] = sum*idiag[i]; 2009 /* x = x + t */ 2010 x[i] += t[i]; 2011 } 2012 2013 ierr = PetscLogFlops(6.0*m-1 + 2.0*a->nz);CHKERRQ(ierr); 2014 ierr = VecRestoreArray(xx,&x);CHKERRQ(ierr); 2015 ierr = VecRestoreArrayRead(bb,&b);CHKERRQ(ierr); 2016 PetscFunctionReturn(0); 2017 } 2018 if (flag & SOR_ZERO_INITIAL_GUESS) { 2019 if (flag & SOR_FORWARD_SWEEP || flag & SOR_LOCAL_FORWARD_SWEEP) { 2020 for (i=0; i<m; i++) { 2021 n = diag[i] - a->i[i]; 2022 idx = a->j + a->i[i]; 2023 v = a->a + a->i[i]; 2024 sum = b[i]; 2025 PetscSparseDenseMinusDot(sum,x,v,idx,n); 2026 t[i] = sum; 2027 x[i] = sum*idiag[i]; 2028 } 2029 xb = t; 2030 ierr = PetscLogFlops(a->nz);CHKERRQ(ierr); 2031 } else xb = b; 2032 if (flag & SOR_BACKWARD_SWEEP || flag & SOR_LOCAL_BACKWARD_SWEEP) { 2033 for (i=m-1; i>=0; i--) { 2034 n = a->i[i+1] - diag[i] - 1; 2035 idx = a->j + diag[i] + 1; 2036 v = a->a + diag[i] + 1; 2037 sum = xb[i]; 2038 PetscSparseDenseMinusDot(sum,x,v,idx,n); 2039 if (xb == b) { 2040 x[i] = sum*idiag[i]; 2041 } else { 2042 x[i] = (1-omega)*x[i] + sum*idiag[i]; /* omega in idiag */ 2043 } 2044 } 2045 ierr = PetscLogFlops(a->nz);CHKERRQ(ierr); /* assumes 1/2 in upper */ 2046 } 2047 its--; 2048 } 2049 while (its--) { 2050 if (flag & SOR_FORWARD_SWEEP || flag & SOR_LOCAL_FORWARD_SWEEP) { 2051 for (i=0; i<m; i++) { 2052 /* lower */ 2053 n = diag[i] - a->i[i]; 2054 idx = a->j + a->i[i]; 2055 v = a->a + a->i[i]; 2056 sum = b[i]; 2057 PetscSparseDenseMinusDot(sum,x,v,idx,n); 2058 t[i] = sum; /* save application of the lower-triangular part */ 2059 /* upper */ 2060 n = a->i[i+1] - diag[i] - 1; 2061 idx = a->j + diag[i] + 1; 2062 v = a->a + diag[i] + 1; 2063 PetscSparseDenseMinusDot(sum,x,v,idx,n); 2064 x[i] = (1. - omega)*x[i] + sum*idiag[i]; /* omega in idiag */ 2065 } 2066 xb = t; 2067 ierr = PetscLogFlops(2.0*a->nz);CHKERRQ(ierr); 2068 } else xb = b; 2069 if (flag & SOR_BACKWARD_SWEEP || flag & SOR_LOCAL_BACKWARD_SWEEP) { 2070 for (i=m-1; i>=0; i--) { 2071 sum = xb[i]; 2072 if (xb == b) { 2073 /* whole matrix (no checkpointing available) */ 2074 n = a->i[i+1] - a->i[i]; 2075 idx = a->j + a->i[i]; 2076 v = a->a + a->i[i]; 2077 PetscSparseDenseMinusDot(sum,x,v,idx,n); 2078 x[i] = (1. - omega)*x[i] + (sum + mdiag[i]*x[i])*idiag[i]; 2079 } else { /* lower-triangular part has been saved, so only apply upper-triangular */ 2080 n = a->i[i+1] - diag[i] - 1; 2081 idx = a->j + diag[i] + 1; 2082 v = a->a + diag[i] + 1; 2083 PetscSparseDenseMinusDot(sum,x,v,idx,n); 2084 x[i] = (1. - omega)*x[i] + sum*idiag[i]; /* omega in idiag */ 2085 } 2086 } 2087 if (xb == b) { 2088 ierr = PetscLogFlops(2.0*a->nz);CHKERRQ(ierr); 2089 } else { 2090 ierr = PetscLogFlops(a->nz);CHKERRQ(ierr); /* assumes 1/2 in upper */ 2091 } 2092 } 2093 } 2094 ierr = VecRestoreArray(xx,&x);CHKERRQ(ierr); 2095 ierr = VecRestoreArrayRead(bb,&b);CHKERRQ(ierr); 2096 PetscFunctionReturn(0); 2097 } 2098 2099 2100 PetscErrorCode MatGetInfo_SeqAIJ(Mat A,MatInfoType flag,MatInfo *info) 2101 { 2102 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 2103 2104 PetscFunctionBegin; 2105 info->block_size = 1.0; 2106 info->nz_allocated = a->maxnz; 2107 info->nz_used = a->nz; 2108 info->nz_unneeded = (a->maxnz - a->nz); 2109 info->assemblies = A->num_ass; 2110 info->mallocs = A->info.mallocs; 2111 info->memory = ((PetscObject)A)->mem; 2112 if (A->factortype) { 2113 info->fill_ratio_given = A->info.fill_ratio_given; 2114 info->fill_ratio_needed = A->info.fill_ratio_needed; 2115 info->factor_mallocs = A->info.factor_mallocs; 2116 } else { 2117 info->fill_ratio_given = 0; 2118 info->fill_ratio_needed = 0; 2119 info->factor_mallocs = 0; 2120 } 2121 PetscFunctionReturn(0); 2122 } 2123 2124 PetscErrorCode MatZeroRows_SeqAIJ(Mat A,PetscInt N,const PetscInt rows[],PetscScalar diag,Vec x,Vec b) 2125 { 2126 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 2127 PetscInt i,m = A->rmap->n - 1; 2128 PetscErrorCode ierr; 2129 const PetscScalar *xx; 2130 PetscScalar *bb; 2131 PetscInt d = 0; 2132 2133 PetscFunctionBegin; 2134 if (x && b) { 2135 ierr = VecGetArrayRead(x,&xx);CHKERRQ(ierr); 2136 ierr = VecGetArray(b,&bb);CHKERRQ(ierr); 2137 for (i=0; i<N; i++) { 2138 if (rows[i] < 0 || rows[i] > m) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"row %D out of range", rows[i]); 2139 if (rows[i] >= A->cmap->n) continue; 2140 bb[rows[i]] = diag*xx[rows[i]]; 2141 } 2142 ierr = VecRestoreArrayRead(x,&xx);CHKERRQ(ierr); 2143 ierr = VecRestoreArray(b,&bb);CHKERRQ(ierr); 2144 } 2145 2146 if (a->keepnonzeropattern) { 2147 for (i=0; i<N; i++) { 2148 if (rows[i] < 0 || rows[i] > m) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"row %D out of range", rows[i]); 2149 ierr = PetscArrayzero(&a->a[a->i[rows[i]]],a->ilen[rows[i]]);CHKERRQ(ierr); 2150 } 2151 if (diag != 0.0) { 2152 for (i=0; i<N; i++) { 2153 d = rows[i]; 2154 if (rows[i] >= A->cmap->n) continue; 2155 if (a->diag[d] >= a->i[d+1]) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Matrix is missing diagonal entry in the zeroed row %D",d); 2156 } 2157 for (i=0; i<N; i++) { 2158 if (rows[i] >= A->cmap->n) continue; 2159 a->a[a->diag[rows[i]]] = diag; 2160 } 2161 } 2162 } else { 2163 if (diag != 0.0) { 2164 for (i=0; i<N; i++) { 2165 if (rows[i] < 0 || rows[i] > m) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"row %D out of range", rows[i]); 2166 if (a->ilen[rows[i]] > 0) { 2167 if (rows[i] >= A->cmap->n) { 2168 a->ilen[rows[i]] = 0; 2169 } else { 2170 a->ilen[rows[i]] = 1; 2171 a->a[a->i[rows[i]]] = diag; 2172 a->j[a->i[rows[i]]] = rows[i]; 2173 } 2174 } else if (rows[i] < A->cmap->n) { /* in case row was completely empty */ 2175 ierr = MatSetValues_SeqAIJ(A,1,&rows[i],1,&rows[i],&diag,INSERT_VALUES);CHKERRQ(ierr); 2176 } 2177 } 2178 } else { 2179 for (i=0; i<N; i++) { 2180 if (rows[i] < 0 || rows[i] > m) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"row %D out of range", rows[i]); 2181 a->ilen[rows[i]] = 0; 2182 } 2183 } 2184 A->nonzerostate++; 2185 } 2186 #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA) 2187 if (A->offloadmask != PETSC_OFFLOAD_UNALLOCATED) A->offloadmask = PETSC_OFFLOAD_CPU; 2188 #endif 2189 ierr = (*A->ops->assemblyend)(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2190 PetscFunctionReturn(0); 2191 } 2192 2193 PetscErrorCode MatZeroRowsColumns_SeqAIJ(Mat A,PetscInt N,const PetscInt rows[],PetscScalar diag,Vec x,Vec b) 2194 { 2195 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 2196 PetscInt i,j,m = A->rmap->n - 1,d = 0; 2197 PetscErrorCode ierr; 2198 PetscBool missing,*zeroed,vecs = PETSC_FALSE; 2199 const PetscScalar *xx; 2200 PetscScalar *bb; 2201 2202 PetscFunctionBegin; 2203 if (x && b) { 2204 ierr = VecGetArrayRead(x,&xx);CHKERRQ(ierr); 2205 ierr = VecGetArray(b,&bb);CHKERRQ(ierr); 2206 vecs = PETSC_TRUE; 2207 } 2208 ierr = PetscCalloc1(A->rmap->n,&zeroed);CHKERRQ(ierr); 2209 for (i=0; i<N; i++) { 2210 if (rows[i] < 0 || rows[i] > m) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"row %D out of range", rows[i]); 2211 ierr = PetscArrayzero(&a->a[a->i[rows[i]]],a->ilen[rows[i]]);CHKERRQ(ierr); 2212 2213 zeroed[rows[i]] = PETSC_TRUE; 2214 } 2215 for (i=0; i<A->rmap->n; i++) { 2216 if (!zeroed[i]) { 2217 for (j=a->i[i]; j<a->i[i+1]; j++) { 2218 if (a->j[j] < A->rmap->n && zeroed[a->j[j]]) { 2219 if (vecs) bb[i] -= a->a[j]*xx[a->j[j]]; 2220 a->a[j] = 0.0; 2221 } 2222 } 2223 } else if (vecs && i < A->cmap->N) bb[i] = diag*xx[i]; 2224 } 2225 if (x && b) { 2226 ierr = VecRestoreArrayRead(x,&xx);CHKERRQ(ierr); 2227 ierr = VecRestoreArray(b,&bb);CHKERRQ(ierr); 2228 } 2229 ierr = PetscFree(zeroed);CHKERRQ(ierr); 2230 if (diag != 0.0) { 2231 ierr = MatMissingDiagonal_SeqAIJ(A,&missing,&d);CHKERRQ(ierr); 2232 if (missing) { 2233 for (i=0; i<N; i++) { 2234 if (rows[i] >= A->cmap->N) continue; 2235 if (a->nonew && rows[i] >= d) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Matrix is missing diagonal entry in row %D (%D)",d,rows[i]); 2236 ierr = MatSetValues_SeqAIJ(A,1,&rows[i],1,&rows[i],&diag,INSERT_VALUES);CHKERRQ(ierr); 2237 } 2238 } else { 2239 for (i=0; i<N; i++) { 2240 a->a[a->diag[rows[i]]] = diag; 2241 } 2242 } 2243 } 2244 #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA) 2245 if (A->offloadmask != PETSC_OFFLOAD_UNALLOCATED) A->offloadmask = PETSC_OFFLOAD_CPU; 2246 #endif 2247 ierr = (*A->ops->assemblyend)(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2248 PetscFunctionReturn(0); 2249 } 2250 2251 PetscErrorCode MatGetRow_SeqAIJ(Mat A,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v) 2252 { 2253 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 2254 PetscInt *itmp; 2255 2256 PetscFunctionBegin; 2257 if (row < 0 || row >= A->rmap->n) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Row %D out of range",row); 2258 2259 *nz = a->i[row+1] - a->i[row]; 2260 if (v) *v = a->a + a->i[row]; 2261 if (idx) { 2262 itmp = a->j + a->i[row]; 2263 if (*nz) *idx = itmp; 2264 else *idx = NULL; 2265 } 2266 PetscFunctionReturn(0); 2267 } 2268 2269 /* remove this function? */ 2270 PetscErrorCode MatRestoreRow_SeqAIJ(Mat A,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v) 2271 { 2272 PetscFunctionBegin; 2273 PetscFunctionReturn(0); 2274 } 2275 2276 PetscErrorCode MatNorm_SeqAIJ(Mat A,NormType type,PetscReal *nrm) 2277 { 2278 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 2279 MatScalar *v = a->a; 2280 PetscReal sum = 0.0; 2281 PetscErrorCode ierr; 2282 PetscInt i,j; 2283 2284 PetscFunctionBegin; 2285 if (type == NORM_FROBENIUS) { 2286 #if defined(PETSC_USE_REAL___FP16) 2287 PetscBLASInt one = 1,nz = a->nz; 2288 *nrm = BLASnrm2_(&nz,v,&one); 2289 #else 2290 for (i=0; i<a->nz; i++) { 2291 sum += PetscRealPart(PetscConj(*v)*(*v)); v++; 2292 } 2293 *nrm = PetscSqrtReal(sum); 2294 #endif 2295 ierr = PetscLogFlops(2.0*a->nz);CHKERRQ(ierr); 2296 } else if (type == NORM_1) { 2297 PetscReal *tmp; 2298 PetscInt *jj = a->j; 2299 ierr = PetscCalloc1(A->cmap->n+1,&tmp);CHKERRQ(ierr); 2300 *nrm = 0.0; 2301 for (j=0; j<a->nz; j++) { 2302 tmp[*jj++] += PetscAbsScalar(*v); v++; 2303 } 2304 for (j=0; j<A->cmap->n; j++) { 2305 if (tmp[j] > *nrm) *nrm = tmp[j]; 2306 } 2307 ierr = PetscFree(tmp);CHKERRQ(ierr); 2308 ierr = PetscLogFlops(PetscMax(a->nz-1,0));CHKERRQ(ierr); 2309 } else if (type == NORM_INFINITY) { 2310 *nrm = 0.0; 2311 for (j=0; j<A->rmap->n; j++) { 2312 v = a->a + a->i[j]; 2313 sum = 0.0; 2314 for (i=0; i<a->i[j+1]-a->i[j]; i++) { 2315 sum += PetscAbsScalar(*v); v++; 2316 } 2317 if (sum > *nrm) *nrm = sum; 2318 } 2319 ierr = PetscLogFlops(PetscMax(a->nz-1,0));CHKERRQ(ierr); 2320 } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"No support for two norm"); 2321 PetscFunctionReturn(0); 2322 } 2323 2324 /* Merged from MatGetSymbolicTranspose_SeqAIJ() - replace MatGetSymbolicTranspose_SeqAIJ()? */ 2325 PetscErrorCode MatTransposeSymbolic_SeqAIJ(Mat A,Mat *B) 2326 { 2327 PetscErrorCode ierr; 2328 PetscInt i,j,anzj; 2329 Mat_SeqAIJ *a=(Mat_SeqAIJ*)A->data,*b; 2330 PetscInt an=A->cmap->N,am=A->rmap->N; 2331 PetscInt *ati,*atj,*atfill,*ai=a->i,*aj=a->j; 2332 2333 PetscFunctionBegin; 2334 /* Allocate space for symbolic transpose info and work array */ 2335 ierr = PetscCalloc1(an+1,&ati);CHKERRQ(ierr); 2336 ierr = PetscMalloc1(ai[am],&atj);CHKERRQ(ierr); 2337 ierr = PetscMalloc1(an,&atfill);CHKERRQ(ierr); 2338 2339 /* Walk through aj and count ## of non-zeros in each row of A^T. */ 2340 /* Note: offset by 1 for fast conversion into csr format. */ 2341 for (i=0;i<ai[am];i++) ati[aj[i]+1] += 1; 2342 /* Form ati for csr format of A^T. */ 2343 for (i=0;i<an;i++) ati[i+1] += ati[i]; 2344 2345 /* Copy ati into atfill so we have locations of the next free space in atj */ 2346 ierr = PetscArraycpy(atfill,ati,an);CHKERRQ(ierr); 2347 2348 /* Walk through A row-wise and mark nonzero entries of A^T. */ 2349 for (i=0;i<am;i++) { 2350 anzj = ai[i+1] - ai[i]; 2351 for (j=0;j<anzj;j++) { 2352 atj[atfill[*aj]] = i; 2353 atfill[*aj++] += 1; 2354 } 2355 } 2356 2357 /* Clean up temporary space and complete requests. */ 2358 ierr = PetscFree(atfill);CHKERRQ(ierr); 2359 ierr = MatCreateSeqAIJWithArrays(PetscObjectComm((PetscObject)A),an,am,ati,atj,NULL,B);CHKERRQ(ierr); 2360 ierr = MatSetBlockSizes(*B,PetscAbs(A->cmap->bs),PetscAbs(A->rmap->bs));CHKERRQ(ierr); 2361 ierr = MatSetType(*B,((PetscObject)A)->type_name);CHKERRQ(ierr); 2362 2363 b = (Mat_SeqAIJ*)((*B)->data); 2364 b->free_a = PETSC_FALSE; 2365 b->free_ij = PETSC_TRUE; 2366 b->nonew = 0; 2367 PetscFunctionReturn(0); 2368 } 2369 2370 PetscErrorCode MatIsTranspose_SeqAIJ(Mat A,Mat B,PetscReal tol,PetscBool *f) 2371 { 2372 Mat_SeqAIJ *aij = (Mat_SeqAIJ*) A->data,*bij = (Mat_SeqAIJ*) B->data; 2373 PetscInt *adx,*bdx,*aii,*bii,*aptr,*bptr; 2374 MatScalar *va,*vb; 2375 PetscErrorCode ierr; 2376 PetscInt ma,na,mb,nb, i; 2377 2378 PetscFunctionBegin; 2379 ierr = MatGetSize(A,&ma,&na);CHKERRQ(ierr); 2380 ierr = MatGetSize(B,&mb,&nb);CHKERRQ(ierr); 2381 if (ma!=nb || na!=mb) { 2382 *f = PETSC_FALSE; 2383 PetscFunctionReturn(0); 2384 } 2385 aii = aij->i; bii = bij->i; 2386 adx = aij->j; bdx = bij->j; 2387 va = aij->a; vb = bij->a; 2388 ierr = PetscMalloc1(ma,&aptr);CHKERRQ(ierr); 2389 ierr = PetscMalloc1(mb,&bptr);CHKERRQ(ierr); 2390 for (i=0; i<ma; i++) aptr[i] = aii[i]; 2391 for (i=0; i<mb; i++) bptr[i] = bii[i]; 2392 2393 *f = PETSC_TRUE; 2394 for (i=0; i<ma; i++) { 2395 while (aptr[i]<aii[i+1]) { 2396 PetscInt idc,idr; 2397 PetscScalar vc,vr; 2398 /* column/row index/value */ 2399 idc = adx[aptr[i]]; 2400 idr = bdx[bptr[idc]]; 2401 vc = va[aptr[i]]; 2402 vr = vb[bptr[idc]]; 2403 if (i!=idr || PetscAbsScalar(vc-vr) > tol) { 2404 *f = PETSC_FALSE; 2405 goto done; 2406 } else { 2407 aptr[i]++; 2408 if (B || i!=idc) bptr[idc]++; 2409 } 2410 } 2411 } 2412 done: 2413 ierr = PetscFree(aptr);CHKERRQ(ierr); 2414 ierr = PetscFree(bptr);CHKERRQ(ierr); 2415 PetscFunctionReturn(0); 2416 } 2417 2418 PetscErrorCode MatIsHermitianTranspose_SeqAIJ(Mat A,Mat B,PetscReal tol,PetscBool *f) 2419 { 2420 Mat_SeqAIJ *aij = (Mat_SeqAIJ*) A->data,*bij = (Mat_SeqAIJ*) B->data; 2421 PetscInt *adx,*bdx,*aii,*bii,*aptr,*bptr; 2422 MatScalar *va,*vb; 2423 PetscErrorCode ierr; 2424 PetscInt ma,na,mb,nb, i; 2425 2426 PetscFunctionBegin; 2427 ierr = MatGetSize(A,&ma,&na);CHKERRQ(ierr); 2428 ierr = MatGetSize(B,&mb,&nb);CHKERRQ(ierr); 2429 if (ma!=nb || na!=mb) { 2430 *f = PETSC_FALSE; 2431 PetscFunctionReturn(0); 2432 } 2433 aii = aij->i; bii = bij->i; 2434 adx = aij->j; bdx = bij->j; 2435 va = aij->a; vb = bij->a; 2436 ierr = PetscMalloc1(ma,&aptr);CHKERRQ(ierr); 2437 ierr = PetscMalloc1(mb,&bptr);CHKERRQ(ierr); 2438 for (i=0; i<ma; i++) aptr[i] = aii[i]; 2439 for (i=0; i<mb; i++) bptr[i] = bii[i]; 2440 2441 *f = PETSC_TRUE; 2442 for (i=0; i<ma; i++) { 2443 while (aptr[i]<aii[i+1]) { 2444 PetscInt idc,idr; 2445 PetscScalar vc,vr; 2446 /* column/row index/value */ 2447 idc = adx[aptr[i]]; 2448 idr = bdx[bptr[idc]]; 2449 vc = va[aptr[i]]; 2450 vr = vb[bptr[idc]]; 2451 if (i!=idr || PetscAbsScalar(vc-PetscConj(vr)) > tol) { 2452 *f = PETSC_FALSE; 2453 goto done; 2454 } else { 2455 aptr[i]++; 2456 if (B || i!=idc) bptr[idc]++; 2457 } 2458 } 2459 } 2460 done: 2461 ierr = PetscFree(aptr);CHKERRQ(ierr); 2462 ierr = PetscFree(bptr);CHKERRQ(ierr); 2463 PetscFunctionReturn(0); 2464 } 2465 2466 PetscErrorCode MatIsSymmetric_SeqAIJ(Mat A,PetscReal tol,PetscBool *f) 2467 { 2468 PetscErrorCode ierr; 2469 2470 PetscFunctionBegin; 2471 ierr = MatIsTranspose_SeqAIJ(A,A,tol,f);CHKERRQ(ierr); 2472 PetscFunctionReturn(0); 2473 } 2474 2475 PetscErrorCode MatIsHermitian_SeqAIJ(Mat A,PetscReal tol,PetscBool *f) 2476 { 2477 PetscErrorCode ierr; 2478 2479 PetscFunctionBegin; 2480 ierr = MatIsHermitianTranspose_SeqAIJ(A,A,tol,f);CHKERRQ(ierr); 2481 PetscFunctionReturn(0); 2482 } 2483 2484 PetscErrorCode MatDiagonalScale_SeqAIJ(Mat A,Vec ll,Vec rr) 2485 { 2486 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 2487 const PetscScalar *l,*r; 2488 PetscScalar x; 2489 MatScalar *v; 2490 PetscErrorCode ierr; 2491 PetscInt i,j,m = A->rmap->n,n = A->cmap->n,M,nz = a->nz; 2492 const PetscInt *jj; 2493 2494 PetscFunctionBegin; 2495 if (ll) { 2496 /* The local size is used so that VecMPI can be passed to this routine 2497 by MatDiagonalScale_MPIAIJ */ 2498 ierr = VecGetLocalSize(ll,&m);CHKERRQ(ierr); 2499 if (m != A->rmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Left scaling vector wrong length"); 2500 ierr = VecGetArrayRead(ll,&l);CHKERRQ(ierr); 2501 v = a->a; 2502 for (i=0; i<m; i++) { 2503 x = l[i]; 2504 M = a->i[i+1] - a->i[i]; 2505 for (j=0; j<M; j++) (*v++) *= x; 2506 } 2507 ierr = VecRestoreArrayRead(ll,&l);CHKERRQ(ierr); 2508 ierr = PetscLogFlops(nz);CHKERRQ(ierr); 2509 } 2510 if (rr) { 2511 ierr = VecGetLocalSize(rr,&n);CHKERRQ(ierr); 2512 if (n != A->cmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Right scaling vector wrong length"); 2513 ierr = VecGetArrayRead(rr,&r);CHKERRQ(ierr); 2514 v = a->a; jj = a->j; 2515 for (i=0; i<nz; i++) (*v++) *= r[*jj++]; 2516 ierr = VecRestoreArrayRead(rr,&r);CHKERRQ(ierr); 2517 ierr = PetscLogFlops(nz);CHKERRQ(ierr); 2518 } 2519 ierr = MatSeqAIJInvalidateDiagonal(A);CHKERRQ(ierr); 2520 #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA) 2521 if (A->offloadmask != PETSC_OFFLOAD_UNALLOCATED) A->offloadmask = PETSC_OFFLOAD_CPU; 2522 #endif 2523 PetscFunctionReturn(0); 2524 } 2525 2526 PetscErrorCode MatCreateSubMatrix_SeqAIJ(Mat A,IS isrow,IS iscol,PetscInt csize,MatReuse scall,Mat *B) 2527 { 2528 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data,*c; 2529 PetscErrorCode ierr; 2530 PetscInt *smap,i,k,kstart,kend,oldcols = A->cmap->n,*lens; 2531 PetscInt row,mat_i,*mat_j,tcol,first,step,*mat_ilen,sum,lensi; 2532 const PetscInt *irow,*icol; 2533 PetscInt nrows,ncols; 2534 PetscInt *starts,*j_new,*i_new,*aj = a->j,*ai = a->i,ii,*ailen = a->ilen; 2535 MatScalar *a_new,*mat_a; 2536 Mat C; 2537 PetscBool stride; 2538 2539 PetscFunctionBegin; 2540 2541 ierr = ISGetIndices(isrow,&irow);CHKERRQ(ierr); 2542 ierr = ISGetLocalSize(isrow,&nrows);CHKERRQ(ierr); 2543 ierr = ISGetLocalSize(iscol,&ncols);CHKERRQ(ierr); 2544 2545 ierr = PetscObjectTypeCompare((PetscObject)iscol,ISSTRIDE,&stride);CHKERRQ(ierr); 2546 if (stride) { 2547 ierr = ISStrideGetInfo(iscol,&first,&step);CHKERRQ(ierr); 2548 } else { 2549 first = 0; 2550 step = 0; 2551 } 2552 if (stride && step == 1) { 2553 /* special case of contiguous rows */ 2554 ierr = PetscMalloc2(nrows,&lens,nrows,&starts);CHKERRQ(ierr); 2555 /* loop over new rows determining lens and starting points */ 2556 for (i=0; i<nrows; i++) { 2557 kstart = ai[irow[i]]; 2558 kend = kstart + ailen[irow[i]]; 2559 starts[i] = kstart; 2560 for (k=kstart; k<kend; k++) { 2561 if (aj[k] >= first) { 2562 starts[i] = k; 2563 break; 2564 } 2565 } 2566 sum = 0; 2567 while (k < kend) { 2568 if (aj[k++] >= first+ncols) break; 2569 sum++; 2570 } 2571 lens[i] = sum; 2572 } 2573 /* create submatrix */ 2574 if (scall == MAT_REUSE_MATRIX) { 2575 PetscInt n_cols,n_rows; 2576 ierr = MatGetSize(*B,&n_rows,&n_cols);CHKERRQ(ierr); 2577 if (n_rows != nrows || n_cols != ncols) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Reused submatrix wrong size"); 2578 ierr = MatZeroEntries(*B);CHKERRQ(ierr); 2579 C = *B; 2580 } else { 2581 PetscInt rbs,cbs; 2582 ierr = MatCreate(PetscObjectComm((PetscObject)A),&C);CHKERRQ(ierr); 2583 ierr = MatSetSizes(C,nrows,ncols,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 2584 ierr = ISGetBlockSize(isrow,&rbs);CHKERRQ(ierr); 2585 ierr = ISGetBlockSize(iscol,&cbs);CHKERRQ(ierr); 2586 ierr = MatSetBlockSizes(C,rbs,cbs);CHKERRQ(ierr); 2587 ierr = MatSetType(C,((PetscObject)A)->type_name);CHKERRQ(ierr); 2588 ierr = MatSeqAIJSetPreallocation_SeqAIJ(C,0,lens);CHKERRQ(ierr); 2589 } 2590 c = (Mat_SeqAIJ*)C->data; 2591 2592 /* loop over rows inserting into submatrix */ 2593 a_new = c->a; 2594 j_new = c->j; 2595 i_new = c->i; 2596 2597 for (i=0; i<nrows; i++) { 2598 ii = starts[i]; 2599 lensi = lens[i]; 2600 for (k=0; k<lensi; k++) { 2601 *j_new++ = aj[ii+k] - first; 2602 } 2603 ierr = PetscArraycpy(a_new,a->a + starts[i],lensi);CHKERRQ(ierr); 2604 a_new += lensi; 2605 i_new[i+1] = i_new[i] + lensi; 2606 c->ilen[i] = lensi; 2607 } 2608 ierr = PetscFree2(lens,starts);CHKERRQ(ierr); 2609 } else { 2610 ierr = ISGetIndices(iscol,&icol);CHKERRQ(ierr); 2611 ierr = PetscCalloc1(oldcols,&smap);CHKERRQ(ierr); 2612 ierr = PetscMalloc1(1+nrows,&lens);CHKERRQ(ierr); 2613 for (i=0; i<ncols; i++) { 2614 if (PetscUnlikelyDebug(icol[i] >= oldcols)) SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Requesting column beyond largest column icol[%D] %D >= A->cmap->n %D",i,icol[i],oldcols); 2615 smap[icol[i]] = i+1; 2616 } 2617 2618 /* determine lens of each row */ 2619 for (i=0; i<nrows; i++) { 2620 kstart = ai[irow[i]]; 2621 kend = kstart + a->ilen[irow[i]]; 2622 lens[i] = 0; 2623 for (k=kstart; k<kend; k++) { 2624 if (smap[aj[k]]) { 2625 lens[i]++; 2626 } 2627 } 2628 } 2629 /* Create and fill new matrix */ 2630 if (scall == MAT_REUSE_MATRIX) { 2631 PetscBool equal; 2632 2633 c = (Mat_SeqAIJ*)((*B)->data); 2634 if ((*B)->rmap->n != nrows || (*B)->cmap->n != ncols) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Cannot reuse matrix. wrong size"); 2635 ierr = PetscArraycmp(c->ilen,lens,(*B)->rmap->n,&equal);CHKERRQ(ierr); 2636 if (!equal) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Cannot reuse matrix. wrong no of nonzeros"); 2637 ierr = PetscArrayzero(c->ilen,(*B)->rmap->n);CHKERRQ(ierr); 2638 C = *B; 2639 } else { 2640 PetscInt rbs,cbs; 2641 ierr = MatCreate(PetscObjectComm((PetscObject)A),&C);CHKERRQ(ierr); 2642 ierr = MatSetSizes(C,nrows,ncols,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 2643 ierr = ISGetBlockSize(isrow,&rbs);CHKERRQ(ierr); 2644 ierr = ISGetBlockSize(iscol,&cbs);CHKERRQ(ierr); 2645 ierr = MatSetBlockSizes(C,rbs,cbs);CHKERRQ(ierr); 2646 ierr = MatSetType(C,((PetscObject)A)->type_name);CHKERRQ(ierr); 2647 ierr = MatSeqAIJSetPreallocation_SeqAIJ(C,0,lens);CHKERRQ(ierr); 2648 } 2649 c = (Mat_SeqAIJ*)(C->data); 2650 for (i=0; i<nrows; i++) { 2651 row = irow[i]; 2652 kstart = ai[row]; 2653 kend = kstart + a->ilen[row]; 2654 mat_i = c->i[i]; 2655 mat_j = c->j + mat_i; 2656 mat_a = c->a + mat_i; 2657 mat_ilen = c->ilen + i; 2658 for (k=kstart; k<kend; k++) { 2659 if ((tcol=smap[a->j[k]])) { 2660 *mat_j++ = tcol - 1; 2661 *mat_a++ = a->a[k]; 2662 (*mat_ilen)++; 2663 2664 } 2665 } 2666 } 2667 /* Free work space */ 2668 ierr = ISRestoreIndices(iscol,&icol);CHKERRQ(ierr); 2669 ierr = PetscFree(smap);CHKERRQ(ierr); 2670 ierr = PetscFree(lens);CHKERRQ(ierr); 2671 /* sort */ 2672 for (i = 0; i < nrows; i++) { 2673 PetscInt ilen; 2674 2675 mat_i = c->i[i]; 2676 mat_j = c->j + mat_i; 2677 mat_a = c->a + mat_i; 2678 ilen = c->ilen[i]; 2679 ierr = PetscSortIntWithScalarArray(ilen,mat_j,mat_a);CHKERRQ(ierr); 2680 } 2681 } 2682 #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA) 2683 ierr = MatBindToCPU(C,A->boundtocpu);CHKERRQ(ierr); 2684 #endif 2685 ierr = MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2686 ierr = MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2687 2688 ierr = ISRestoreIndices(isrow,&irow);CHKERRQ(ierr); 2689 *B = C; 2690 PetscFunctionReturn(0); 2691 } 2692 2693 PetscErrorCode MatGetMultiProcBlock_SeqAIJ(Mat mat,MPI_Comm subComm,MatReuse scall,Mat *subMat) 2694 { 2695 PetscErrorCode ierr; 2696 Mat B; 2697 2698 PetscFunctionBegin; 2699 if (scall == MAT_INITIAL_MATRIX) { 2700 ierr = MatCreate(subComm,&B);CHKERRQ(ierr); 2701 ierr = MatSetSizes(B,mat->rmap->n,mat->cmap->n,mat->rmap->n,mat->cmap->n);CHKERRQ(ierr); 2702 ierr = MatSetBlockSizesFromMats(B,mat,mat);CHKERRQ(ierr); 2703 ierr = MatSetType(B,MATSEQAIJ);CHKERRQ(ierr); 2704 ierr = MatDuplicateNoCreate_SeqAIJ(B,mat,MAT_COPY_VALUES,PETSC_TRUE);CHKERRQ(ierr); 2705 *subMat = B; 2706 } else { 2707 ierr = MatCopy_SeqAIJ(mat,*subMat,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 2708 } 2709 PetscFunctionReturn(0); 2710 } 2711 2712 PetscErrorCode MatILUFactor_SeqAIJ(Mat inA,IS row,IS col,const MatFactorInfo *info) 2713 { 2714 Mat_SeqAIJ *a = (Mat_SeqAIJ*)inA->data; 2715 PetscErrorCode ierr; 2716 Mat outA; 2717 PetscBool row_identity,col_identity; 2718 2719 PetscFunctionBegin; 2720 if (info->levels != 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Only levels=0 supported for in-place ilu"); 2721 2722 ierr = ISIdentity(row,&row_identity);CHKERRQ(ierr); 2723 ierr = ISIdentity(col,&col_identity);CHKERRQ(ierr); 2724 2725 outA = inA; 2726 outA->factortype = MAT_FACTOR_LU; 2727 ierr = PetscFree(inA->solvertype);CHKERRQ(ierr); 2728 ierr = PetscStrallocpy(MATSOLVERPETSC,&inA->solvertype);CHKERRQ(ierr); 2729 2730 ierr = PetscObjectReference((PetscObject)row);CHKERRQ(ierr); 2731 ierr = ISDestroy(&a->row);CHKERRQ(ierr); 2732 2733 a->row = row; 2734 2735 ierr = PetscObjectReference((PetscObject)col);CHKERRQ(ierr); 2736 ierr = ISDestroy(&a->col);CHKERRQ(ierr); 2737 2738 a->col = col; 2739 2740 /* Create the inverse permutation so that it can be used in MatLUFactorNumeric() */ 2741 ierr = ISDestroy(&a->icol);CHKERRQ(ierr); 2742 ierr = ISInvertPermutation(col,PETSC_DECIDE,&a->icol);CHKERRQ(ierr); 2743 ierr = PetscLogObjectParent((PetscObject)inA,(PetscObject)a->icol);CHKERRQ(ierr); 2744 2745 if (!a->solve_work) { /* this matrix may have been factored before */ 2746 ierr = PetscMalloc1(inA->rmap->n+1,&a->solve_work);CHKERRQ(ierr); 2747 ierr = PetscLogObjectMemory((PetscObject)inA, (inA->rmap->n+1)*sizeof(PetscScalar));CHKERRQ(ierr); 2748 } 2749 2750 ierr = MatMarkDiagonal_SeqAIJ(inA);CHKERRQ(ierr); 2751 if (row_identity && col_identity) { 2752 ierr = MatLUFactorNumeric_SeqAIJ_inplace(outA,inA,info);CHKERRQ(ierr); 2753 } else { 2754 ierr = MatLUFactorNumeric_SeqAIJ_InplaceWithPerm(outA,inA,info);CHKERRQ(ierr); 2755 } 2756 PetscFunctionReturn(0); 2757 } 2758 2759 PetscErrorCode MatScale_SeqAIJ(Mat inA,PetscScalar alpha) 2760 { 2761 Mat_SeqAIJ *a = (Mat_SeqAIJ*)inA->data; 2762 PetscScalar oalpha = alpha; 2763 PetscErrorCode ierr; 2764 PetscBLASInt one = 1,bnz; 2765 2766 PetscFunctionBegin; 2767 ierr = PetscBLASIntCast(a->nz,&bnz);CHKERRQ(ierr); 2768 PetscStackCallBLAS("BLASscal",BLASscal_(&bnz,&oalpha,a->a,&one)); 2769 ierr = PetscLogFlops(a->nz);CHKERRQ(ierr); 2770 ierr = MatSeqAIJInvalidateDiagonal(inA);CHKERRQ(ierr); 2771 #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA) 2772 if (inA->offloadmask != PETSC_OFFLOAD_UNALLOCATED) inA->offloadmask = PETSC_OFFLOAD_CPU; 2773 #endif 2774 PetscFunctionReturn(0); 2775 } 2776 2777 PetscErrorCode MatDestroySubMatrix_Private(Mat_SubSppt *submatj) 2778 { 2779 PetscErrorCode ierr; 2780 PetscInt i; 2781 2782 PetscFunctionBegin; 2783 if (!submatj->id) { /* delete data that are linked only to submats[id=0] */ 2784 ierr = PetscFree4(submatj->sbuf1,submatj->ptr,submatj->tmp,submatj->ctr);CHKERRQ(ierr); 2785 2786 for (i=0; i<submatj->nrqr; ++i) { 2787 ierr = PetscFree(submatj->sbuf2[i]);CHKERRQ(ierr); 2788 } 2789 ierr = PetscFree3(submatj->sbuf2,submatj->req_size,submatj->req_source1);CHKERRQ(ierr); 2790 2791 if (submatj->rbuf1) { 2792 ierr = PetscFree(submatj->rbuf1[0]);CHKERRQ(ierr); 2793 ierr = PetscFree(submatj->rbuf1);CHKERRQ(ierr); 2794 } 2795 2796 for (i=0; i<submatj->nrqs; ++i) { 2797 ierr = PetscFree(submatj->rbuf3[i]);CHKERRQ(ierr); 2798 } 2799 ierr = PetscFree3(submatj->req_source2,submatj->rbuf2,submatj->rbuf3);CHKERRQ(ierr); 2800 ierr = PetscFree(submatj->pa);CHKERRQ(ierr); 2801 } 2802 2803 #if defined(PETSC_USE_CTABLE) 2804 ierr = PetscTableDestroy((PetscTable*)&submatj->rmap);CHKERRQ(ierr); 2805 if (submatj->cmap_loc) {ierr = PetscFree(submatj->cmap_loc);CHKERRQ(ierr);} 2806 ierr = PetscFree(submatj->rmap_loc);CHKERRQ(ierr); 2807 #else 2808 ierr = PetscFree(submatj->rmap);CHKERRQ(ierr); 2809 #endif 2810 2811 if (!submatj->allcolumns) { 2812 #if defined(PETSC_USE_CTABLE) 2813 ierr = PetscTableDestroy((PetscTable*)&submatj->cmap);CHKERRQ(ierr); 2814 #else 2815 ierr = PetscFree(submatj->cmap);CHKERRQ(ierr); 2816 #endif 2817 } 2818 ierr = PetscFree(submatj->row2proc);CHKERRQ(ierr); 2819 2820 ierr = PetscFree(submatj);CHKERRQ(ierr); 2821 PetscFunctionReturn(0); 2822 } 2823 2824 PetscErrorCode MatDestroySubMatrix_SeqAIJ(Mat C) 2825 { 2826 PetscErrorCode ierr; 2827 Mat_SeqAIJ *c = (Mat_SeqAIJ*)C->data; 2828 Mat_SubSppt *submatj = c->submatis1; 2829 2830 PetscFunctionBegin; 2831 ierr = (*submatj->destroy)(C);CHKERRQ(ierr); 2832 ierr = MatDestroySubMatrix_Private(submatj);CHKERRQ(ierr); 2833 PetscFunctionReturn(0); 2834 } 2835 2836 PetscErrorCode MatDestroySubMatrices_SeqAIJ(PetscInt n,Mat *mat[]) 2837 { 2838 PetscErrorCode ierr; 2839 PetscInt i; 2840 Mat C; 2841 Mat_SeqAIJ *c; 2842 Mat_SubSppt *submatj; 2843 2844 PetscFunctionBegin; 2845 for (i=0; i<n; i++) { 2846 C = (*mat)[i]; 2847 c = (Mat_SeqAIJ*)C->data; 2848 submatj = c->submatis1; 2849 if (submatj) { 2850 if (--((PetscObject)C)->refct <= 0) { 2851 ierr = (*submatj->destroy)(C);CHKERRQ(ierr); 2852 ierr = MatDestroySubMatrix_Private(submatj);CHKERRQ(ierr); 2853 ierr = PetscFree(C->defaultvectype);CHKERRQ(ierr); 2854 ierr = PetscLayoutDestroy(&C->rmap);CHKERRQ(ierr); 2855 ierr = PetscLayoutDestroy(&C->cmap);CHKERRQ(ierr); 2856 ierr = PetscHeaderDestroy(&C);CHKERRQ(ierr); 2857 } 2858 } else { 2859 ierr = MatDestroy(&C);CHKERRQ(ierr); 2860 } 2861 } 2862 2863 /* Destroy Dummy submatrices created for reuse */ 2864 ierr = MatDestroySubMatrices_Dummy(n,mat);CHKERRQ(ierr); 2865 2866 ierr = PetscFree(*mat);CHKERRQ(ierr); 2867 PetscFunctionReturn(0); 2868 } 2869 2870 PetscErrorCode MatCreateSubMatrices_SeqAIJ(Mat A,PetscInt n,const IS irow[],const IS icol[],MatReuse scall,Mat *B[]) 2871 { 2872 PetscErrorCode ierr; 2873 PetscInt i; 2874 2875 PetscFunctionBegin; 2876 if (scall == MAT_INITIAL_MATRIX) { 2877 ierr = PetscCalloc1(n+1,B);CHKERRQ(ierr); 2878 } 2879 2880 for (i=0; i<n; i++) { 2881 ierr = MatCreateSubMatrix_SeqAIJ(A,irow[i],icol[i],PETSC_DECIDE,scall,&(*B)[i]);CHKERRQ(ierr); 2882 } 2883 PetscFunctionReturn(0); 2884 } 2885 2886 PetscErrorCode MatIncreaseOverlap_SeqAIJ(Mat A,PetscInt is_max,IS is[],PetscInt ov) 2887 { 2888 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 2889 PetscErrorCode ierr; 2890 PetscInt row,i,j,k,l,m,n,*nidx,isz,val; 2891 const PetscInt *idx; 2892 PetscInt start,end,*ai,*aj; 2893 PetscBT table; 2894 2895 PetscFunctionBegin; 2896 m = A->rmap->n; 2897 ai = a->i; 2898 aj = a->j; 2899 2900 if (ov < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"illegal negative overlap value used"); 2901 2902 ierr = PetscMalloc1(m+1,&nidx);CHKERRQ(ierr); 2903 ierr = PetscBTCreate(m,&table);CHKERRQ(ierr); 2904 2905 for (i=0; i<is_max; i++) { 2906 /* Initialize the two local arrays */ 2907 isz = 0; 2908 ierr = PetscBTMemzero(m,table);CHKERRQ(ierr); 2909 2910 /* Extract the indices, assume there can be duplicate entries */ 2911 ierr = ISGetIndices(is[i],&idx);CHKERRQ(ierr); 2912 ierr = ISGetLocalSize(is[i],&n);CHKERRQ(ierr); 2913 2914 /* Enter these into the temp arrays. I.e., mark table[row], enter row into new index */ 2915 for (j=0; j<n; ++j) { 2916 if (!PetscBTLookupSet(table,idx[j])) nidx[isz++] = idx[j]; 2917 } 2918 ierr = ISRestoreIndices(is[i],&idx);CHKERRQ(ierr); 2919 ierr = ISDestroy(&is[i]);CHKERRQ(ierr); 2920 2921 k = 0; 2922 for (j=0; j<ov; j++) { /* for each overlap */ 2923 n = isz; 2924 for (; k<n; k++) { /* do only those rows in nidx[k], which are not done yet */ 2925 row = nidx[k]; 2926 start = ai[row]; 2927 end = ai[row+1]; 2928 for (l = start; l<end; l++) { 2929 val = aj[l]; 2930 if (!PetscBTLookupSet(table,val)) nidx[isz++] = val; 2931 } 2932 } 2933 } 2934 ierr = ISCreateGeneral(PETSC_COMM_SELF,isz,nidx,PETSC_COPY_VALUES,(is+i));CHKERRQ(ierr); 2935 } 2936 ierr = PetscBTDestroy(&table);CHKERRQ(ierr); 2937 ierr = PetscFree(nidx);CHKERRQ(ierr); 2938 PetscFunctionReturn(0); 2939 } 2940 2941 /* -------------------------------------------------------------- */ 2942 PetscErrorCode MatPermute_SeqAIJ(Mat A,IS rowp,IS colp,Mat *B) 2943 { 2944 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 2945 PetscErrorCode ierr; 2946 PetscInt i,nz = 0,m = A->rmap->n,n = A->cmap->n; 2947 const PetscInt *row,*col; 2948 PetscInt *cnew,j,*lens; 2949 IS icolp,irowp; 2950 PetscInt *cwork = NULL; 2951 PetscScalar *vwork = NULL; 2952 2953 PetscFunctionBegin; 2954 ierr = ISInvertPermutation(rowp,PETSC_DECIDE,&irowp);CHKERRQ(ierr); 2955 ierr = ISGetIndices(irowp,&row);CHKERRQ(ierr); 2956 ierr = ISInvertPermutation(colp,PETSC_DECIDE,&icolp);CHKERRQ(ierr); 2957 ierr = ISGetIndices(icolp,&col);CHKERRQ(ierr); 2958 2959 /* determine lengths of permuted rows */ 2960 ierr = PetscMalloc1(m+1,&lens);CHKERRQ(ierr); 2961 for (i=0; i<m; i++) lens[row[i]] = a->i[i+1] - a->i[i]; 2962 ierr = MatCreate(PetscObjectComm((PetscObject)A),B);CHKERRQ(ierr); 2963 ierr = MatSetSizes(*B,m,n,m,n);CHKERRQ(ierr); 2964 ierr = MatSetBlockSizesFromMats(*B,A,A);CHKERRQ(ierr); 2965 ierr = MatSetType(*B,((PetscObject)A)->type_name);CHKERRQ(ierr); 2966 ierr = MatSeqAIJSetPreallocation_SeqAIJ(*B,0,lens);CHKERRQ(ierr); 2967 ierr = PetscFree(lens);CHKERRQ(ierr); 2968 2969 ierr = PetscMalloc1(n,&cnew);CHKERRQ(ierr); 2970 for (i=0; i<m; i++) { 2971 ierr = MatGetRow_SeqAIJ(A,i,&nz,&cwork,&vwork);CHKERRQ(ierr); 2972 for (j=0; j<nz; j++) cnew[j] = col[cwork[j]]; 2973 ierr = MatSetValues_SeqAIJ(*B,1,&row[i],nz,cnew,vwork,INSERT_VALUES);CHKERRQ(ierr); 2974 ierr = MatRestoreRow_SeqAIJ(A,i,&nz,&cwork,&vwork);CHKERRQ(ierr); 2975 } 2976 ierr = PetscFree(cnew);CHKERRQ(ierr); 2977 2978 (*B)->assembled = PETSC_FALSE; 2979 2980 #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA) 2981 ierr = MatBindToCPU(*B,A->boundtocpu);CHKERRQ(ierr); 2982 #endif 2983 ierr = MatAssemblyBegin(*B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2984 ierr = MatAssemblyEnd(*B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2985 ierr = ISRestoreIndices(irowp,&row);CHKERRQ(ierr); 2986 ierr = ISRestoreIndices(icolp,&col);CHKERRQ(ierr); 2987 ierr = ISDestroy(&irowp);CHKERRQ(ierr); 2988 ierr = ISDestroy(&icolp);CHKERRQ(ierr); 2989 if (rowp == colp) { 2990 if (A->symmetric) { 2991 ierr = MatSetOption(*B,MAT_SYMMETRIC,PETSC_TRUE);CHKERRQ(ierr); 2992 } 2993 if (A->hermitian) { 2994 ierr = MatSetOption(*B,MAT_HERMITIAN,PETSC_TRUE);CHKERRQ(ierr); 2995 } 2996 } 2997 PetscFunctionReturn(0); 2998 } 2999 3000 PetscErrorCode MatCopy_SeqAIJ(Mat A,Mat B,MatStructure str) 3001 { 3002 PetscErrorCode ierr; 3003 3004 PetscFunctionBegin; 3005 /* If the two matrices have the same copy implementation, use fast copy. */ 3006 if (str == SAME_NONZERO_PATTERN && (A->ops->copy == B->ops->copy)) { 3007 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 3008 Mat_SeqAIJ *b = (Mat_SeqAIJ*)B->data; 3009 3010 if (a->i[A->rmap->n] != b->i[B->rmap->n]) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_INCOMP,"Number of nonzeros in two matrices are different %D != %D",a->i[A->rmap->n],b->i[B->rmap->n]); 3011 ierr = PetscArraycpy(b->a,a->a,a->i[A->rmap->n]);CHKERRQ(ierr); 3012 ierr = PetscObjectStateIncrease((PetscObject)B);CHKERRQ(ierr); 3013 } else { 3014 ierr = MatCopy_Basic(A,B,str);CHKERRQ(ierr); 3015 } 3016 PetscFunctionReturn(0); 3017 } 3018 3019 PetscErrorCode MatSetUp_SeqAIJ(Mat A) 3020 { 3021 PetscErrorCode ierr; 3022 3023 PetscFunctionBegin; 3024 ierr = MatSeqAIJSetPreallocation_SeqAIJ(A,PETSC_DEFAULT,NULL);CHKERRQ(ierr); 3025 PetscFunctionReturn(0); 3026 } 3027 3028 PETSC_INTERN PetscErrorCode MatSeqAIJGetArray_SeqAIJ(Mat A,PetscScalar *array[]) 3029 { 3030 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 3031 3032 PetscFunctionBegin; 3033 *array = a->a; 3034 PetscFunctionReturn(0); 3035 } 3036 3037 PETSC_INTERN PetscErrorCode MatSeqAIJRestoreArray_SeqAIJ(Mat A,PetscScalar *array[]) 3038 { 3039 PetscFunctionBegin; 3040 *array = NULL; 3041 PetscFunctionReturn(0); 3042 } 3043 3044 /* 3045 Computes the number of nonzeros per row needed for preallocation when X and Y 3046 have different nonzero structure. 3047 */ 3048 PetscErrorCode MatAXPYGetPreallocation_SeqX_private(PetscInt m,const PetscInt *xi,const PetscInt *xj,const PetscInt *yi,const PetscInt *yj,PetscInt *nnz) 3049 { 3050 PetscInt i,j,k,nzx,nzy; 3051 3052 PetscFunctionBegin; 3053 /* Set the number of nonzeros in the new matrix */ 3054 for (i=0; i<m; i++) { 3055 const PetscInt *xjj = xj+xi[i],*yjj = yj+yi[i]; 3056 nzx = xi[i+1] - xi[i]; 3057 nzy = yi[i+1] - yi[i]; 3058 nnz[i] = 0; 3059 for (j=0,k=0; j<nzx; j++) { /* Point in X */ 3060 for (; k<nzy && yjj[k]<xjj[j]; k++) nnz[i]++; /* Catch up to X */ 3061 if (k<nzy && yjj[k]==xjj[j]) k++; /* Skip duplicate */ 3062 nnz[i]++; 3063 } 3064 for (; k<nzy; k++) nnz[i]++; 3065 } 3066 PetscFunctionReturn(0); 3067 } 3068 3069 PetscErrorCode MatAXPYGetPreallocation_SeqAIJ(Mat Y,Mat X,PetscInt *nnz) 3070 { 3071 PetscInt m = Y->rmap->N; 3072 Mat_SeqAIJ *x = (Mat_SeqAIJ*)X->data; 3073 Mat_SeqAIJ *y = (Mat_SeqAIJ*)Y->data; 3074 PetscErrorCode ierr; 3075 3076 PetscFunctionBegin; 3077 /* Set the number of nonzeros in the new matrix */ 3078 ierr = MatAXPYGetPreallocation_SeqX_private(m,x->i,x->j,y->i,y->j,nnz);CHKERRQ(ierr); 3079 PetscFunctionReturn(0); 3080 } 3081 3082 PetscErrorCode MatAXPY_SeqAIJ(Mat Y,PetscScalar a,Mat X,MatStructure str) 3083 { 3084 PetscErrorCode ierr; 3085 Mat_SeqAIJ *x = (Mat_SeqAIJ*)X->data,*y = (Mat_SeqAIJ*)Y->data; 3086 3087 PetscFunctionBegin; 3088 if (str == DIFFERENT_NONZERO_PATTERN) { 3089 if (x->nz == y->nz) { 3090 PetscBool e; 3091 ierr = PetscArraycmp(x->i,y->i,Y->rmap->n+1,&e);CHKERRQ(ierr); 3092 if (e) { 3093 ierr = PetscArraycmp(x->j,y->j,y->nz,&e);CHKERRQ(ierr); 3094 if (e) { 3095 str = SAME_NONZERO_PATTERN; 3096 } 3097 } 3098 } 3099 } 3100 if (str == SAME_NONZERO_PATTERN) { 3101 PetscScalar alpha = a; 3102 PetscBLASInt one = 1,bnz; 3103 3104 ierr = PetscBLASIntCast(x->nz,&bnz);CHKERRQ(ierr); 3105 PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&bnz,&alpha,x->a,&one,y->a,&one)); 3106 ierr = MatSeqAIJInvalidateDiagonal(Y);CHKERRQ(ierr); 3107 ierr = PetscObjectStateIncrease((PetscObject)Y);CHKERRQ(ierr); 3108 /* the MatAXPY_Basic* subroutines calls MatAssembly, so the matrix on the GPU will be updated */ 3109 #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA) 3110 if (Y->offloadmask != PETSC_OFFLOAD_UNALLOCATED) { 3111 Y->offloadmask = PETSC_OFFLOAD_CPU; 3112 } 3113 #endif 3114 } else if (str == SUBSET_NONZERO_PATTERN) { /* nonzeros of X is a subset of Y's */ 3115 ierr = MatAXPY_Basic(Y,a,X,str);CHKERRQ(ierr); 3116 } else { 3117 Mat B; 3118 PetscInt *nnz; 3119 ierr = PetscMalloc1(Y->rmap->N,&nnz);CHKERRQ(ierr); 3120 ierr = MatCreate(PetscObjectComm((PetscObject)Y),&B);CHKERRQ(ierr); 3121 ierr = PetscObjectSetName((PetscObject)B,((PetscObject)Y)->name);CHKERRQ(ierr); 3122 ierr = MatSetLayouts(B,Y->rmap,Y->cmap);CHKERRQ(ierr); 3123 ierr = MatSetType(B,(MatType) ((PetscObject)Y)->type_name);CHKERRQ(ierr); 3124 ierr = MatAXPYGetPreallocation_SeqAIJ(Y,X,nnz);CHKERRQ(ierr); 3125 ierr = MatSeqAIJSetPreallocation(B,0,nnz);CHKERRQ(ierr); 3126 ierr = MatAXPY_BasicWithPreallocation(B,Y,a,X,str);CHKERRQ(ierr); 3127 ierr = MatHeaderReplace(Y,&B);CHKERRQ(ierr); 3128 ierr = PetscFree(nnz);CHKERRQ(ierr); 3129 } 3130 PetscFunctionReturn(0); 3131 } 3132 3133 PetscErrorCode MatConjugate_SeqAIJ(Mat mat) 3134 { 3135 #if defined(PETSC_USE_COMPLEX) 3136 Mat_SeqAIJ *aij = (Mat_SeqAIJ*)mat->data; 3137 PetscInt i,nz; 3138 PetscScalar *a; 3139 3140 PetscFunctionBegin; 3141 nz = aij->nz; 3142 a = aij->a; 3143 for (i=0; i<nz; i++) a[i] = PetscConj(a[i]); 3144 #if defined(PETSC_HAVE_VIENNACL) || defined(PETSC_HAVE_CUDA) 3145 if (mat->offloadmask != PETSC_OFFLOAD_UNALLOCATED) mat->offloadmask = PETSC_OFFLOAD_CPU; 3146 #endif 3147 #else 3148 PetscFunctionBegin; 3149 #endif 3150 PetscFunctionReturn(0); 3151 } 3152 3153 PetscErrorCode MatGetRowMaxAbs_SeqAIJ(Mat A,Vec v,PetscInt idx[]) 3154 { 3155 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 3156 PetscErrorCode ierr; 3157 PetscInt i,j,m = A->rmap->n,*ai,*aj,ncols,n; 3158 PetscReal atmp; 3159 PetscScalar *x; 3160 MatScalar *aa; 3161 3162 PetscFunctionBegin; 3163 if (A->factortype) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Not for factored matrix"); 3164 aa = a->a; 3165 ai = a->i; 3166 aj = a->j; 3167 3168 ierr = VecSet(v,0.0);CHKERRQ(ierr); 3169 ierr = VecGetArray(v,&x);CHKERRQ(ierr); 3170 ierr = VecGetLocalSize(v,&n);CHKERRQ(ierr); 3171 if (n != A->rmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Nonconforming matrix and vector"); 3172 for (i=0; i<m; i++) { 3173 ncols = ai[1] - ai[0]; ai++; 3174 x[i] = 0.0; 3175 for (j=0; j<ncols; j++) { 3176 atmp = PetscAbsScalar(*aa); 3177 if (PetscAbsScalar(x[i]) < atmp) {x[i] = atmp; if (idx) idx[i] = *aj;} 3178 aa++; aj++; 3179 } 3180 } 3181 ierr = VecRestoreArray(v,&x);CHKERRQ(ierr); 3182 PetscFunctionReturn(0); 3183 } 3184 3185 PetscErrorCode MatGetRowMax_SeqAIJ(Mat A,Vec v,PetscInt idx[]) 3186 { 3187 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 3188 PetscErrorCode ierr; 3189 PetscInt i,j,m = A->rmap->n,*ai,*aj,ncols,n; 3190 PetscScalar *x; 3191 MatScalar *aa; 3192 3193 PetscFunctionBegin; 3194 if (A->factortype) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Not for factored matrix"); 3195 aa = a->a; 3196 ai = a->i; 3197 aj = a->j; 3198 3199 ierr = VecSet(v,0.0);CHKERRQ(ierr); 3200 ierr = VecGetArray(v,&x);CHKERRQ(ierr); 3201 ierr = VecGetLocalSize(v,&n);CHKERRQ(ierr); 3202 if (n != A->rmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Nonconforming matrix and vector"); 3203 for (i=0; i<m; i++) { 3204 ncols = ai[1] - ai[0]; ai++; 3205 if (ncols == A->cmap->n) { /* row is dense */ 3206 x[i] = *aa; if (idx) idx[i] = 0; 3207 } else { /* row is sparse so already KNOW maximum is 0.0 or higher */ 3208 x[i] = 0.0; 3209 if (idx) { 3210 idx[i] = 0; /* in case ncols is zero */ 3211 for (j=0;j<ncols;j++) { /* find first implicit 0.0 in the row */ 3212 if (aj[j] > j) { 3213 idx[i] = j; 3214 break; 3215 } 3216 } 3217 } 3218 } 3219 for (j=0; j<ncols; j++) { 3220 if (PetscRealPart(x[i]) < PetscRealPart(*aa)) {x[i] = *aa; if (idx) idx[i] = *aj;} 3221 aa++; aj++; 3222 } 3223 } 3224 ierr = VecRestoreArray(v,&x);CHKERRQ(ierr); 3225 PetscFunctionReturn(0); 3226 } 3227 3228 PetscErrorCode MatGetRowMinAbs_SeqAIJ(Mat A,Vec v,PetscInt idx[]) 3229 { 3230 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 3231 PetscErrorCode ierr; 3232 PetscInt i,j,m = A->rmap->n,*ai,*aj,ncols,n; 3233 PetscReal atmp; 3234 PetscScalar *x; 3235 MatScalar *aa; 3236 3237 PetscFunctionBegin; 3238 if (A->factortype) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Not for factored matrix"); 3239 aa = a->a; 3240 ai = a->i; 3241 aj = a->j; 3242 3243 ierr = VecSet(v,0.0);CHKERRQ(ierr); 3244 ierr = VecGetArray(v,&x);CHKERRQ(ierr); 3245 ierr = VecGetLocalSize(v,&n);CHKERRQ(ierr); 3246 if (n != A->rmap->n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Nonconforming matrix and vector, %D vs. %D rows", A->rmap->n, n); 3247 for (i=0; i<m; i++) { 3248 ncols = ai[1] - ai[0]; ai++; 3249 if (ncols) { 3250 /* Get first nonzero */ 3251 for (j = 0; j < ncols; j++) { 3252 atmp = PetscAbsScalar(aa[j]); 3253 if (atmp > 1.0e-12) { 3254 x[i] = atmp; 3255 if (idx) idx[i] = aj[j]; 3256 break; 3257 } 3258 } 3259 if (j == ncols) {x[i] = PetscAbsScalar(*aa); if (idx) idx[i] = *aj;} 3260 } else { 3261 x[i] = 0.0; if (idx) idx[i] = 0; 3262 } 3263 for (j = 0; j < ncols; j++) { 3264 atmp = PetscAbsScalar(*aa); 3265 if (atmp > 1.0e-12 && PetscAbsScalar(x[i]) > atmp) {x[i] = atmp; if (idx) idx[i] = *aj;} 3266 aa++; aj++; 3267 } 3268 } 3269 ierr = VecRestoreArray(v,&x);CHKERRQ(ierr); 3270 PetscFunctionReturn(0); 3271 } 3272 3273 PetscErrorCode MatGetRowMin_SeqAIJ(Mat A,Vec v,PetscInt idx[]) 3274 { 3275 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 3276 PetscErrorCode ierr; 3277 PetscInt i,j,m = A->rmap->n,ncols,n; 3278 const PetscInt *ai,*aj; 3279 PetscScalar *x; 3280 const MatScalar *aa; 3281 3282 PetscFunctionBegin; 3283 if (A->factortype) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Not for factored matrix"); 3284 aa = a->a; 3285 ai = a->i; 3286 aj = a->j; 3287 3288 ierr = VecSet(v,0.0);CHKERRQ(ierr); 3289 ierr = VecGetArray(v,&x);CHKERRQ(ierr); 3290 ierr = VecGetLocalSize(v,&n);CHKERRQ(ierr); 3291 if (n != A->rmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Nonconforming matrix and vector"); 3292 for (i=0; i<m; i++) { 3293 ncols = ai[1] - ai[0]; ai++; 3294 if (ncols == A->cmap->n) { /* row is dense */ 3295 x[i] = *aa; if (idx) idx[i] = 0; 3296 } else { /* row is sparse so already KNOW minimum is 0.0 or lower */ 3297 x[i] = 0.0; 3298 if (idx) { /* find first implicit 0.0 in the row */ 3299 idx[i] = 0; /* in case ncols is zero */ 3300 for (j=0; j<ncols; j++) { 3301 if (aj[j] > j) { 3302 idx[i] = j; 3303 break; 3304 } 3305 } 3306 } 3307 } 3308 for (j=0; j<ncols; j++) { 3309 if (PetscRealPart(x[i]) > PetscRealPart(*aa)) {x[i] = *aa; if (idx) idx[i] = *aj;} 3310 aa++; aj++; 3311 } 3312 } 3313 ierr = VecRestoreArray(v,&x);CHKERRQ(ierr); 3314 PetscFunctionReturn(0); 3315 } 3316 3317 PetscErrorCode MatInvertBlockDiagonal_SeqAIJ(Mat A,const PetscScalar **values) 3318 { 3319 Mat_SeqAIJ *a = (Mat_SeqAIJ*) A->data; 3320 PetscErrorCode ierr; 3321 PetscInt i,bs = PetscAbs(A->rmap->bs),mbs = A->rmap->n/bs,ipvt[5],bs2 = bs*bs,*v_pivots,ij[7],*IJ,j; 3322 MatScalar *diag,work[25],*v_work; 3323 const PetscReal shift = 0.0; 3324 PetscBool allowzeropivot,zeropivotdetected=PETSC_FALSE; 3325 3326 PetscFunctionBegin; 3327 allowzeropivot = PetscNot(A->erroriffailure); 3328 if (a->ibdiagvalid) { 3329 if (values) *values = a->ibdiag; 3330 PetscFunctionReturn(0); 3331 } 3332 ierr = MatMarkDiagonal_SeqAIJ(A);CHKERRQ(ierr); 3333 if (!a->ibdiag) { 3334 ierr = PetscMalloc1(bs2*mbs,&a->ibdiag);CHKERRQ(ierr); 3335 ierr = PetscLogObjectMemory((PetscObject)A,bs2*mbs*sizeof(PetscScalar));CHKERRQ(ierr); 3336 } 3337 diag = a->ibdiag; 3338 if (values) *values = a->ibdiag; 3339 /* factor and invert each block */ 3340 switch (bs) { 3341 case 1: 3342 for (i=0; i<mbs; i++) { 3343 ierr = MatGetValues(A,1,&i,1,&i,diag+i);CHKERRQ(ierr); 3344 if (PetscAbsScalar(diag[i] + shift) < PETSC_MACHINE_EPSILON) { 3345 if (allowzeropivot) { 3346 A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT; 3347 A->factorerror_zeropivot_value = PetscAbsScalar(diag[i]); 3348 A->factorerror_zeropivot_row = i; 3349 ierr = PetscInfo3(A,"Zero pivot, row %D pivot %g tolerance %g\n",i,(double)PetscAbsScalar(diag[i]),(double)PETSC_MACHINE_EPSILON);CHKERRQ(ierr); 3350 } else SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_MAT_LU_ZRPVT,"Zero pivot, row %D pivot %g tolerance %g",i,(double)PetscAbsScalar(diag[i]),(double)PETSC_MACHINE_EPSILON); 3351 } 3352 diag[i] = (PetscScalar)1.0 / (diag[i] + shift); 3353 } 3354 break; 3355 case 2: 3356 for (i=0; i<mbs; i++) { 3357 ij[0] = 2*i; ij[1] = 2*i + 1; 3358 ierr = MatGetValues(A,2,ij,2,ij,diag);CHKERRQ(ierr); 3359 ierr = PetscKernel_A_gets_inverse_A_2(diag,shift,allowzeropivot,&zeropivotdetected);CHKERRQ(ierr); 3360 if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT; 3361 ierr = PetscKernel_A_gets_transpose_A_2(diag);CHKERRQ(ierr); 3362 diag += 4; 3363 } 3364 break; 3365 case 3: 3366 for (i=0; i<mbs; i++) { 3367 ij[0] = 3*i; ij[1] = 3*i + 1; ij[2] = 3*i + 2; 3368 ierr = MatGetValues(A,3,ij,3,ij,diag);CHKERRQ(ierr); 3369 ierr = PetscKernel_A_gets_inverse_A_3(diag,shift,allowzeropivot,&zeropivotdetected);CHKERRQ(ierr); 3370 if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT; 3371 ierr = PetscKernel_A_gets_transpose_A_3(diag);CHKERRQ(ierr); 3372 diag += 9; 3373 } 3374 break; 3375 case 4: 3376 for (i=0; i<mbs; i++) { 3377 ij[0] = 4*i; ij[1] = 4*i + 1; ij[2] = 4*i + 2; ij[3] = 4*i + 3; 3378 ierr = MatGetValues(A,4,ij,4,ij,diag);CHKERRQ(ierr); 3379 ierr = PetscKernel_A_gets_inverse_A_4(diag,shift,allowzeropivot,&zeropivotdetected);CHKERRQ(ierr); 3380 if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT; 3381 ierr = PetscKernel_A_gets_transpose_A_4(diag);CHKERRQ(ierr); 3382 diag += 16; 3383 } 3384 break; 3385 case 5: 3386 for (i=0; i<mbs; i++) { 3387 ij[0] = 5*i; ij[1] = 5*i + 1; ij[2] = 5*i + 2; ij[3] = 5*i + 3; ij[4] = 5*i + 4; 3388 ierr = MatGetValues(A,5,ij,5,ij,diag);CHKERRQ(ierr); 3389 ierr = PetscKernel_A_gets_inverse_A_5(diag,ipvt,work,shift,allowzeropivot,&zeropivotdetected);CHKERRQ(ierr); 3390 if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT; 3391 ierr = PetscKernel_A_gets_transpose_A_5(diag);CHKERRQ(ierr); 3392 diag += 25; 3393 } 3394 break; 3395 case 6: 3396 for (i=0; i<mbs; i++) { 3397 ij[0] = 6*i; ij[1] = 6*i + 1; ij[2] = 6*i + 2; ij[3] = 6*i + 3; ij[4] = 6*i + 4; ij[5] = 6*i + 5; 3398 ierr = MatGetValues(A,6,ij,6,ij,diag);CHKERRQ(ierr); 3399 ierr = PetscKernel_A_gets_inverse_A_6(diag,shift,allowzeropivot,&zeropivotdetected);CHKERRQ(ierr); 3400 if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT; 3401 ierr = PetscKernel_A_gets_transpose_A_6(diag);CHKERRQ(ierr); 3402 diag += 36; 3403 } 3404 break; 3405 case 7: 3406 for (i=0; i<mbs; i++) { 3407 ij[0] = 7*i; ij[1] = 7*i + 1; ij[2] = 7*i + 2; ij[3] = 7*i + 3; ij[4] = 7*i + 4; ij[5] = 7*i + 5; ij[5] = 7*i + 6; 3408 ierr = MatGetValues(A,7,ij,7,ij,diag);CHKERRQ(ierr); 3409 ierr = PetscKernel_A_gets_inverse_A_7(diag,shift,allowzeropivot,&zeropivotdetected);CHKERRQ(ierr); 3410 if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT; 3411 ierr = PetscKernel_A_gets_transpose_A_7(diag);CHKERRQ(ierr); 3412 diag += 49; 3413 } 3414 break; 3415 default: 3416 ierr = PetscMalloc3(bs,&v_work,bs,&v_pivots,bs,&IJ);CHKERRQ(ierr); 3417 for (i=0; i<mbs; i++) { 3418 for (j=0; j<bs; j++) { 3419 IJ[j] = bs*i + j; 3420 } 3421 ierr = MatGetValues(A,bs,IJ,bs,IJ,diag);CHKERRQ(ierr); 3422 ierr = PetscKernel_A_gets_inverse_A(bs,diag,v_pivots,v_work,allowzeropivot,&zeropivotdetected);CHKERRQ(ierr); 3423 if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT; 3424 ierr = PetscKernel_A_gets_transpose_A_N(diag,bs);CHKERRQ(ierr); 3425 diag += bs2; 3426 } 3427 ierr = PetscFree3(v_work,v_pivots,IJ);CHKERRQ(ierr); 3428 } 3429 a->ibdiagvalid = PETSC_TRUE; 3430 PetscFunctionReturn(0); 3431 } 3432 3433 static PetscErrorCode MatSetRandom_SeqAIJ(Mat x,PetscRandom rctx) 3434 { 3435 PetscErrorCode ierr; 3436 Mat_SeqAIJ *aij = (Mat_SeqAIJ*)x->data; 3437 PetscScalar a; 3438 PetscInt m,n,i,j,col; 3439 3440 PetscFunctionBegin; 3441 if (!x->assembled) { 3442 ierr = MatGetSize(x,&m,&n);CHKERRQ(ierr); 3443 for (i=0; i<m; i++) { 3444 for (j=0; j<aij->imax[i]; j++) { 3445 ierr = PetscRandomGetValue(rctx,&a);CHKERRQ(ierr); 3446 col = (PetscInt)(n*PetscRealPart(a)); 3447 ierr = MatSetValues(x,1,&i,1,&col,&a,ADD_VALUES);CHKERRQ(ierr); 3448 } 3449 } 3450 } else { 3451 for (i=0; i<aij->nz; i++) {ierr = PetscRandomGetValue(rctx,aij->a+i);CHKERRQ(ierr);} 3452 } 3453 ierr = MatAssemblyBegin(x,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3454 ierr = MatAssemblyEnd(x,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3455 PetscFunctionReturn(0); 3456 } 3457 3458 /* Like MatSetRandom_SeqAIJ, but do not set values on columns in range of [low, high) */ 3459 PetscErrorCode MatSetRandomSkipColumnRange_SeqAIJ_Private(Mat x,PetscInt low,PetscInt high,PetscRandom rctx) 3460 { 3461 PetscErrorCode ierr; 3462 Mat_SeqAIJ *aij = (Mat_SeqAIJ*)x->data; 3463 PetscScalar a; 3464 PetscInt m,n,i,j,col,nskip; 3465 3466 PetscFunctionBegin; 3467 nskip = high - low; 3468 ierr = MatGetSize(x,&m,&n);CHKERRQ(ierr); 3469 n -= nskip; /* shrink number of columns where nonzeros can be set */ 3470 for (i=0; i<m; i++) { 3471 for (j=0; j<aij->imax[i]; j++) { 3472 ierr = PetscRandomGetValue(rctx,&a);CHKERRQ(ierr); 3473 col = (PetscInt)(n*PetscRealPart(a)); 3474 if (col >= low) col += nskip; /* shift col rightward to skip the hole */ 3475 ierr = MatSetValues(x,1,&i,1,&col,&a,ADD_VALUES);CHKERRQ(ierr); 3476 } 3477 } 3478 ierr = MatAssemblyBegin(x,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3479 ierr = MatAssemblyEnd(x,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3480 PetscFunctionReturn(0); 3481 } 3482 3483 3484 /* -------------------------------------------------------------------*/ 3485 static struct _MatOps MatOps_Values = { MatSetValues_SeqAIJ, 3486 MatGetRow_SeqAIJ, 3487 MatRestoreRow_SeqAIJ, 3488 MatMult_SeqAIJ, 3489 /* 4*/ MatMultAdd_SeqAIJ, 3490 MatMultTranspose_SeqAIJ, 3491 MatMultTransposeAdd_SeqAIJ, 3492 NULL, 3493 NULL, 3494 NULL, 3495 /* 10*/ NULL, 3496 MatLUFactor_SeqAIJ, 3497 NULL, 3498 MatSOR_SeqAIJ, 3499 MatTranspose_SeqAIJ, 3500 /*1 5*/ MatGetInfo_SeqAIJ, 3501 MatEqual_SeqAIJ, 3502 MatGetDiagonal_SeqAIJ, 3503 MatDiagonalScale_SeqAIJ, 3504 MatNorm_SeqAIJ, 3505 /* 20*/ NULL, 3506 MatAssemblyEnd_SeqAIJ, 3507 MatSetOption_SeqAIJ, 3508 MatZeroEntries_SeqAIJ, 3509 /* 24*/ MatZeroRows_SeqAIJ, 3510 NULL, 3511 NULL, 3512 NULL, 3513 NULL, 3514 /* 29*/ MatSetUp_SeqAIJ, 3515 NULL, 3516 NULL, 3517 NULL, 3518 NULL, 3519 /* 34*/ MatDuplicate_SeqAIJ, 3520 NULL, 3521 NULL, 3522 MatILUFactor_SeqAIJ, 3523 NULL, 3524 /* 39*/ MatAXPY_SeqAIJ, 3525 MatCreateSubMatrices_SeqAIJ, 3526 MatIncreaseOverlap_SeqAIJ, 3527 MatGetValues_SeqAIJ, 3528 MatCopy_SeqAIJ, 3529 /* 44*/ MatGetRowMax_SeqAIJ, 3530 MatScale_SeqAIJ, 3531 MatShift_SeqAIJ, 3532 MatDiagonalSet_SeqAIJ, 3533 MatZeroRowsColumns_SeqAIJ, 3534 /* 49*/ MatSetRandom_SeqAIJ, 3535 MatGetRowIJ_SeqAIJ, 3536 MatRestoreRowIJ_SeqAIJ, 3537 MatGetColumnIJ_SeqAIJ, 3538 MatRestoreColumnIJ_SeqAIJ, 3539 /* 54*/ MatFDColoringCreate_SeqXAIJ, 3540 NULL, 3541 NULL, 3542 MatPermute_SeqAIJ, 3543 NULL, 3544 /* 59*/ NULL, 3545 MatDestroy_SeqAIJ, 3546 MatView_SeqAIJ, 3547 NULL, 3548 NULL, 3549 /* 64*/ NULL, 3550 MatMatMatMultNumeric_SeqAIJ_SeqAIJ_SeqAIJ, 3551 NULL, 3552 NULL, 3553 NULL, 3554 /* 69*/ MatGetRowMaxAbs_SeqAIJ, 3555 MatGetRowMinAbs_SeqAIJ, 3556 NULL, 3557 NULL, 3558 NULL, 3559 /* 74*/ NULL, 3560 MatFDColoringApply_AIJ, 3561 NULL, 3562 NULL, 3563 NULL, 3564 /* 79*/ MatFindZeroDiagonals_SeqAIJ, 3565 NULL, 3566 NULL, 3567 NULL, 3568 MatLoad_SeqAIJ, 3569 /* 84*/ MatIsSymmetric_SeqAIJ, 3570 MatIsHermitian_SeqAIJ, 3571 NULL, 3572 NULL, 3573 NULL, 3574 /* 89*/ NULL, 3575 NULL, 3576 MatMatMultNumeric_SeqAIJ_SeqAIJ, 3577 NULL, 3578 NULL, 3579 /* 94*/ MatPtAPNumeric_SeqAIJ_SeqAIJ_SparseAxpy, 3580 NULL, 3581 NULL, 3582 MatMatTransposeMultNumeric_SeqAIJ_SeqAIJ, 3583 NULL, 3584 /* 99*/ MatProductSetFromOptions_SeqAIJ, 3585 NULL, 3586 NULL, 3587 MatConjugate_SeqAIJ, 3588 NULL, 3589 /*104*/ MatSetValuesRow_SeqAIJ, 3590 MatRealPart_SeqAIJ, 3591 MatImaginaryPart_SeqAIJ, 3592 NULL, 3593 NULL, 3594 /*109*/ MatMatSolve_SeqAIJ, 3595 NULL, 3596 MatGetRowMin_SeqAIJ, 3597 NULL, 3598 MatMissingDiagonal_SeqAIJ, 3599 /*114*/ NULL, 3600 NULL, 3601 NULL, 3602 NULL, 3603 NULL, 3604 /*119*/ NULL, 3605 NULL, 3606 NULL, 3607 NULL, 3608 MatGetMultiProcBlock_SeqAIJ, 3609 /*124*/ MatFindNonzeroRows_SeqAIJ, 3610 MatGetColumnNorms_SeqAIJ, 3611 MatInvertBlockDiagonal_SeqAIJ, 3612 MatInvertVariableBlockDiagonal_SeqAIJ, 3613 NULL, 3614 /*129*/ NULL, 3615 NULL, 3616 NULL, 3617 MatTransposeMatMultNumeric_SeqAIJ_SeqAIJ, 3618 MatTransposeColoringCreate_SeqAIJ, 3619 /*134*/ MatTransColoringApplySpToDen_SeqAIJ, 3620 MatTransColoringApplyDenToSp_SeqAIJ, 3621 NULL, 3622 NULL, 3623 MatRARtNumeric_SeqAIJ_SeqAIJ, 3624 /*139*/NULL, 3625 NULL, 3626 NULL, 3627 MatFDColoringSetUp_SeqXAIJ, 3628 MatFindOffBlockDiagonalEntries_SeqAIJ, 3629 MatCreateMPIMatConcatenateSeqMat_SeqAIJ, 3630 /*145*/MatDestroySubMatrices_SeqAIJ, 3631 NULL, 3632 NULL 3633 }; 3634 3635 PetscErrorCode MatSeqAIJSetColumnIndices_SeqAIJ(Mat mat,PetscInt *indices) 3636 { 3637 Mat_SeqAIJ *aij = (Mat_SeqAIJ*)mat->data; 3638 PetscInt i,nz,n; 3639 3640 PetscFunctionBegin; 3641 nz = aij->maxnz; 3642 n = mat->rmap->n; 3643 for (i=0; i<nz; i++) { 3644 aij->j[i] = indices[i]; 3645 } 3646 aij->nz = nz; 3647 for (i=0; i<n; i++) { 3648 aij->ilen[i] = aij->imax[i]; 3649 } 3650 PetscFunctionReturn(0); 3651 } 3652 3653 /* 3654 * When a sparse matrix has many zero columns, we should compact them out to save the space 3655 * This happens in MatPtAPSymbolic_MPIAIJ_MPIAIJ_scalable() 3656 * */ 3657 PetscErrorCode MatSeqAIJCompactOutExtraColumns_SeqAIJ(Mat mat, ISLocalToGlobalMapping *mapping) 3658 { 3659 Mat_SeqAIJ *aij = (Mat_SeqAIJ*)mat->data; 3660 PetscTable gid1_lid1; 3661 PetscTablePosition tpos; 3662 PetscInt gid,lid,i,j,ncols,ec; 3663 PetscInt *garray; 3664 PetscErrorCode ierr; 3665 3666 PetscFunctionBegin; 3667 PetscValidHeaderSpecific(mat,MAT_CLASSID,1); 3668 PetscValidPointer(mapping,2); 3669 /* use a table */ 3670 ierr = PetscTableCreate(mat->rmap->n,mat->cmap->N+1,&gid1_lid1);CHKERRQ(ierr); 3671 ec = 0; 3672 for (i=0; i<mat->rmap->n; i++) { 3673 ncols = aij->i[i+1] - aij->i[i]; 3674 for (j=0; j<ncols; j++) { 3675 PetscInt data,gid1 = aij->j[aij->i[i] + j] + 1; 3676 ierr = PetscTableFind(gid1_lid1,gid1,&data);CHKERRQ(ierr); 3677 if (!data) { 3678 /* one based table */ 3679 ierr = PetscTableAdd(gid1_lid1,gid1,++ec,INSERT_VALUES);CHKERRQ(ierr); 3680 } 3681 } 3682 } 3683 /* form array of columns we need */ 3684 ierr = PetscMalloc1(ec+1,&garray);CHKERRQ(ierr); 3685 ierr = PetscTableGetHeadPosition(gid1_lid1,&tpos);CHKERRQ(ierr); 3686 while (tpos) { 3687 ierr = PetscTableGetNext(gid1_lid1,&tpos,&gid,&lid);CHKERRQ(ierr); 3688 gid--; 3689 lid--; 3690 garray[lid] = gid; 3691 } 3692 ierr = PetscSortInt(ec,garray);CHKERRQ(ierr); /* sort, and rebuild */ 3693 ierr = PetscTableRemoveAll(gid1_lid1);CHKERRQ(ierr); 3694 for (i=0; i<ec; i++) { 3695 ierr = PetscTableAdd(gid1_lid1,garray[i]+1,i+1,INSERT_VALUES);CHKERRQ(ierr); 3696 } 3697 /* compact out the extra columns in B */ 3698 for (i=0; i<mat->rmap->n; i++) { 3699 ncols = aij->i[i+1] - aij->i[i]; 3700 for (j=0; j<ncols; j++) { 3701 PetscInt gid1 = aij->j[aij->i[i] + j] + 1; 3702 ierr = PetscTableFind(gid1_lid1,gid1,&lid);CHKERRQ(ierr); 3703 lid--; 3704 aij->j[aij->i[i] + j] = lid; 3705 } 3706 } 3707 ierr = PetscLayoutDestroy(&mat->cmap);CHKERRQ(ierr); 3708 ierr = PetscLayoutCreateFromSizes(PetscObjectComm((PetscObject)mat),ec,ec,1,&mat->cmap);CHKERRQ(ierr); 3709 ierr = PetscTableDestroy(&gid1_lid1);CHKERRQ(ierr); 3710 ierr = ISLocalToGlobalMappingCreate(PETSC_COMM_SELF,mat->cmap->bs,mat->cmap->n,garray,PETSC_OWN_POINTER,mapping);CHKERRQ(ierr); 3711 ierr = ISLocalToGlobalMappingSetType(*mapping,ISLOCALTOGLOBALMAPPINGHASH);CHKERRQ(ierr); 3712 PetscFunctionReturn(0); 3713 } 3714 3715 /*@ 3716 MatSeqAIJSetColumnIndices - Set the column indices for all the rows 3717 in the matrix. 3718 3719 Input Parameters: 3720 + mat - the SeqAIJ matrix 3721 - indices - the column indices 3722 3723 Level: advanced 3724 3725 Notes: 3726 This can be called if you have precomputed the nonzero structure of the 3727 matrix and want to provide it to the matrix object to improve the performance 3728 of the MatSetValues() operation. 3729 3730 You MUST have set the correct numbers of nonzeros per row in the call to 3731 MatCreateSeqAIJ(), and the columns indices MUST be sorted. 3732 3733 MUST be called before any calls to MatSetValues(); 3734 3735 The indices should start with zero, not one. 3736 3737 @*/ 3738 PetscErrorCode MatSeqAIJSetColumnIndices(Mat mat,PetscInt *indices) 3739 { 3740 PetscErrorCode ierr; 3741 3742 PetscFunctionBegin; 3743 PetscValidHeaderSpecific(mat,MAT_CLASSID,1); 3744 PetscValidPointer(indices,2); 3745 ierr = PetscUseMethod(mat,"MatSeqAIJSetColumnIndices_C",(Mat,PetscInt*),(mat,indices));CHKERRQ(ierr); 3746 PetscFunctionReturn(0); 3747 } 3748 3749 /* ----------------------------------------------------------------------------------------*/ 3750 3751 PetscErrorCode MatStoreValues_SeqAIJ(Mat mat) 3752 { 3753 Mat_SeqAIJ *aij = (Mat_SeqAIJ*)mat->data; 3754 PetscErrorCode ierr; 3755 size_t nz = aij->i[mat->rmap->n]; 3756 3757 PetscFunctionBegin; 3758 if (!aij->nonew) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ORDER,"Must call MatSetOption(A,MAT_NEW_NONZERO_LOCATIONS,PETSC_FALSE);first"); 3759 3760 /* allocate space for values if not already there */ 3761 if (!aij->saved_values) { 3762 ierr = PetscMalloc1(nz+1,&aij->saved_values);CHKERRQ(ierr); 3763 ierr = PetscLogObjectMemory((PetscObject)mat,(nz+1)*sizeof(PetscScalar));CHKERRQ(ierr); 3764 } 3765 3766 /* copy values over */ 3767 ierr = PetscArraycpy(aij->saved_values,aij->a,nz);CHKERRQ(ierr); 3768 PetscFunctionReturn(0); 3769 } 3770 3771 /*@ 3772 MatStoreValues - Stashes a copy of the matrix values; this allows, for 3773 example, reuse of the linear part of a Jacobian, while recomputing the 3774 nonlinear portion. 3775 3776 Collect on Mat 3777 3778 Input Parameters: 3779 . mat - the matrix (currently only AIJ matrices support this option) 3780 3781 Level: advanced 3782 3783 Common Usage, with SNESSolve(): 3784 $ Create Jacobian matrix 3785 $ Set linear terms into matrix 3786 $ Apply boundary conditions to matrix, at this time matrix must have 3787 $ final nonzero structure (i.e. setting the nonlinear terms and applying 3788 $ boundary conditions again will not change the nonzero structure 3789 $ ierr = MatSetOption(mat,MAT_NEW_NONZERO_LOCATIONS,PETSC_FALSE); 3790 $ ierr = MatStoreValues(mat); 3791 $ Call SNESSetJacobian() with matrix 3792 $ In your Jacobian routine 3793 $ ierr = MatRetrieveValues(mat); 3794 $ Set nonlinear terms in matrix 3795 3796 Common Usage without SNESSolve(), i.e. when you handle nonlinear solve yourself: 3797 $ // build linear portion of Jacobian 3798 $ ierr = MatSetOption(mat,MAT_NEW_NONZERO_LOCATIONS,PETSC_FALSE); 3799 $ ierr = MatStoreValues(mat); 3800 $ loop over nonlinear iterations 3801 $ ierr = MatRetrieveValues(mat); 3802 $ // call MatSetValues(mat,...) to set nonliner portion of Jacobian 3803 $ // call MatAssemblyBegin/End() on matrix 3804 $ Solve linear system with Jacobian 3805 $ endloop 3806 3807 Notes: 3808 Matrix must already be assemblied before calling this routine 3809 Must set the matrix option MatSetOption(mat,MAT_NEW_NONZERO_LOCATIONS,PETSC_FALSE); before 3810 calling this routine. 3811 3812 When this is called multiple times it overwrites the previous set of stored values 3813 and does not allocated additional space. 3814 3815 .seealso: MatRetrieveValues() 3816 3817 @*/ 3818 PetscErrorCode MatStoreValues(Mat mat) 3819 { 3820 PetscErrorCode ierr; 3821 3822 PetscFunctionBegin; 3823 PetscValidHeaderSpecific(mat,MAT_CLASSID,1); 3824 if (!mat->assembled) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Not for unassembled matrix"); 3825 if (mat->factortype) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Not for factored matrix"); 3826 ierr = PetscUseMethod(mat,"MatStoreValues_C",(Mat),(mat));CHKERRQ(ierr); 3827 PetscFunctionReturn(0); 3828 } 3829 3830 PetscErrorCode MatRetrieveValues_SeqAIJ(Mat mat) 3831 { 3832 Mat_SeqAIJ *aij = (Mat_SeqAIJ*)mat->data; 3833 PetscErrorCode ierr; 3834 PetscInt nz = aij->i[mat->rmap->n]; 3835 3836 PetscFunctionBegin; 3837 if (!aij->nonew) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ORDER,"Must call MatSetOption(A,MAT_NEW_NONZERO_LOCATIONS,PETSC_FALSE);first"); 3838 if (!aij->saved_values) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ORDER,"Must call MatStoreValues(A);first"); 3839 /* copy values over */ 3840 ierr = PetscArraycpy(aij->a,aij->saved_values,nz);CHKERRQ(ierr); 3841 PetscFunctionReturn(0); 3842 } 3843 3844 /*@ 3845 MatRetrieveValues - Retrieves the copy of the matrix values; this allows, for 3846 example, reuse of the linear part of a Jacobian, while recomputing the 3847 nonlinear portion. 3848 3849 Collect on Mat 3850 3851 Input Parameters: 3852 . mat - the matrix (currently only AIJ matrices support this option) 3853 3854 Level: advanced 3855 3856 .seealso: MatStoreValues() 3857 3858 @*/ 3859 PetscErrorCode MatRetrieveValues(Mat mat) 3860 { 3861 PetscErrorCode ierr; 3862 3863 PetscFunctionBegin; 3864 PetscValidHeaderSpecific(mat,MAT_CLASSID,1); 3865 if (!mat->assembled) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Not for unassembled matrix"); 3866 if (mat->factortype) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Not for factored matrix"); 3867 ierr = PetscUseMethod(mat,"MatRetrieveValues_C",(Mat),(mat));CHKERRQ(ierr); 3868 PetscFunctionReturn(0); 3869 } 3870 3871 3872 /* --------------------------------------------------------------------------------*/ 3873 /*@C 3874 MatCreateSeqAIJ - Creates a sparse matrix in AIJ (compressed row) format 3875 (the default parallel PETSc format). For good matrix assembly performance 3876 the user should preallocate the matrix storage by setting the parameter nz 3877 (or the array nnz). By setting these parameters accurately, performance 3878 during matrix assembly can be increased by more than a factor of 50. 3879 3880 Collective 3881 3882 Input Parameters: 3883 + comm - MPI communicator, set to PETSC_COMM_SELF 3884 . m - number of rows 3885 . n - number of columns 3886 . nz - number of nonzeros per row (same for all rows) 3887 - nnz - array containing the number of nonzeros in the various rows 3888 (possibly different for each row) or NULL 3889 3890 Output Parameter: 3891 . A - the matrix 3892 3893 It is recommended that one use the MatCreate(), MatSetType() and/or MatSetFromOptions(), 3894 MatXXXXSetPreallocation() paradigm instead of this routine directly. 3895 [MatXXXXSetPreallocation() is, for example, MatSeqAIJSetPreallocation] 3896 3897 Notes: 3898 If nnz is given then nz is ignored 3899 3900 The AIJ format (also called the Yale sparse matrix format or 3901 compressed row storage), is fully compatible with standard Fortran 77 3902 storage. That is, the stored row and column indices can begin at 3903 either one (as in Fortran) or zero. See the users' manual for details. 3904 3905 Specify the preallocated storage with either nz or nnz (not both). 3906 Set nz=PETSC_DEFAULT and nnz=NULL for PETSc to control dynamic memory 3907 allocation. For large problems you MUST preallocate memory or you 3908 will get TERRIBLE performance, see the users' manual chapter on matrices. 3909 3910 By default, this format uses inodes (identical nodes) when possible, to 3911 improve numerical efficiency of matrix-vector products and solves. We 3912 search for consecutive rows with the same nonzero structure, thereby 3913 reusing matrix information to achieve increased efficiency. 3914 3915 Options Database Keys: 3916 + -mat_no_inode - Do not use inodes 3917 - -mat_inode_limit <limit> - Sets inode limit (max limit=5) 3918 3919 Level: intermediate 3920 3921 .seealso: MatCreate(), MatCreateAIJ(), MatSetValues(), MatSeqAIJSetColumnIndices(), MatCreateSeqAIJWithArrays() 3922 3923 @*/ 3924 PetscErrorCode MatCreateSeqAIJ(MPI_Comm comm,PetscInt m,PetscInt n,PetscInt nz,const PetscInt nnz[],Mat *A) 3925 { 3926 PetscErrorCode ierr; 3927 3928 PetscFunctionBegin; 3929 ierr = MatCreate(comm,A);CHKERRQ(ierr); 3930 ierr = MatSetSizes(*A,m,n,m,n);CHKERRQ(ierr); 3931 ierr = MatSetType(*A,MATSEQAIJ);CHKERRQ(ierr); 3932 ierr = MatSeqAIJSetPreallocation_SeqAIJ(*A,nz,nnz);CHKERRQ(ierr); 3933 PetscFunctionReturn(0); 3934 } 3935 3936 /*@C 3937 MatSeqAIJSetPreallocation - For good matrix assembly performance 3938 the user should preallocate the matrix storage by setting the parameter nz 3939 (or the array nnz). By setting these parameters accurately, performance 3940 during matrix assembly can be increased by more than a factor of 50. 3941 3942 Collective 3943 3944 Input Parameters: 3945 + B - The matrix 3946 . nz - number of nonzeros per row (same for all rows) 3947 - nnz - array containing the number of nonzeros in the various rows 3948 (possibly different for each row) or NULL 3949 3950 Notes: 3951 If nnz is given then nz is ignored 3952 3953 The AIJ format (also called the Yale sparse matrix format or 3954 compressed row storage), is fully compatible with standard Fortran 77 3955 storage. That is, the stored row and column indices can begin at 3956 either one (as in Fortran) or zero. See the users' manual for details. 3957 3958 Specify the preallocated storage with either nz or nnz (not both). 3959 Set nz=PETSC_DEFAULT and nnz=NULL for PETSc to control dynamic memory 3960 allocation. For large problems you MUST preallocate memory or you 3961 will get TERRIBLE performance, see the users' manual chapter on matrices. 3962 3963 You can call MatGetInfo() to get information on how effective the preallocation was; 3964 for example the fields mallocs,nz_allocated,nz_used,nz_unneeded; 3965 You can also run with the option -info and look for messages with the string 3966 malloc in them to see if additional memory allocation was needed. 3967 3968 Developers: Use nz of MAT_SKIP_ALLOCATION to not allocate any space for the matrix 3969 entries or columns indices 3970 3971 By default, this format uses inodes (identical nodes) when possible, to 3972 improve numerical efficiency of matrix-vector products and solves. We 3973 search for consecutive rows with the same nonzero structure, thereby 3974 reusing matrix information to achieve increased efficiency. 3975 3976 Options Database Keys: 3977 + -mat_no_inode - Do not use inodes 3978 - -mat_inode_limit <limit> - Sets inode limit (max limit=5) 3979 3980 Level: intermediate 3981 3982 .seealso: MatCreate(), MatCreateAIJ(), MatSetValues(), MatSeqAIJSetColumnIndices(), MatCreateSeqAIJWithArrays(), MatGetInfo(), 3983 MatSeqAIJSetTotalPreallocation() 3984 3985 @*/ 3986 PetscErrorCode MatSeqAIJSetPreallocation(Mat B,PetscInt nz,const PetscInt nnz[]) 3987 { 3988 PetscErrorCode ierr; 3989 3990 PetscFunctionBegin; 3991 PetscValidHeaderSpecific(B,MAT_CLASSID,1); 3992 PetscValidType(B,1); 3993 ierr = PetscTryMethod(B,"MatSeqAIJSetPreallocation_C",(Mat,PetscInt,const PetscInt[]),(B,nz,nnz));CHKERRQ(ierr); 3994 PetscFunctionReturn(0); 3995 } 3996 3997 PetscErrorCode MatSeqAIJSetPreallocation_SeqAIJ(Mat B,PetscInt nz,const PetscInt *nnz) 3998 { 3999 Mat_SeqAIJ *b; 4000 PetscBool skipallocation = PETSC_FALSE,realalloc = PETSC_FALSE; 4001 PetscErrorCode ierr; 4002 PetscInt i; 4003 4004 PetscFunctionBegin; 4005 if (nz >= 0 || nnz) realalloc = PETSC_TRUE; 4006 if (nz == MAT_SKIP_ALLOCATION) { 4007 skipallocation = PETSC_TRUE; 4008 nz = 0; 4009 } 4010 ierr = PetscLayoutSetUp(B->rmap);CHKERRQ(ierr); 4011 ierr = PetscLayoutSetUp(B->cmap);CHKERRQ(ierr); 4012 4013 if (nz == PETSC_DEFAULT || nz == PETSC_DECIDE) nz = 5; 4014 if (nz < 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"nz cannot be less than 0: value %D",nz); 4015 if (PetscUnlikelyDebug(nnz)) { 4016 for (i=0; i<B->rmap->n; i++) { 4017 if (nnz[i] < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"nnz cannot be less than 0: local row %D value %D",i,nnz[i]); 4018 if (nnz[i] > B->cmap->n) SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"nnz cannot be greater than row length: local row %D value %d rowlength %D",i,nnz[i],B->cmap->n); 4019 } 4020 } 4021 4022 B->preallocated = PETSC_TRUE; 4023 4024 b = (Mat_SeqAIJ*)B->data; 4025 4026 if (!skipallocation) { 4027 if (!b->imax) { 4028 ierr = PetscMalloc1(B->rmap->n,&b->imax);CHKERRQ(ierr); 4029 ierr = PetscLogObjectMemory((PetscObject)B,B->rmap->n*sizeof(PetscInt));CHKERRQ(ierr); 4030 } 4031 if (!b->ilen) { 4032 /* b->ilen will count nonzeros in each row so far. */ 4033 ierr = PetscCalloc1(B->rmap->n,&b->ilen);CHKERRQ(ierr); 4034 ierr = PetscLogObjectMemory((PetscObject)B,B->rmap->n*sizeof(PetscInt));CHKERRQ(ierr); 4035 } else { 4036 ierr = PetscMemzero(b->ilen,B->rmap->n*sizeof(PetscInt));CHKERRQ(ierr); 4037 } 4038 if (!b->ipre) { 4039 ierr = PetscMalloc1(B->rmap->n,&b->ipre);CHKERRQ(ierr); 4040 ierr = PetscLogObjectMemory((PetscObject)B,B->rmap->n*sizeof(PetscInt));CHKERRQ(ierr); 4041 } 4042 if (!nnz) { 4043 if (nz == PETSC_DEFAULT || nz == PETSC_DECIDE) nz = 10; 4044 else if (nz < 0) nz = 1; 4045 nz = PetscMin(nz,B->cmap->n); 4046 for (i=0; i<B->rmap->n; i++) b->imax[i] = nz; 4047 nz = nz*B->rmap->n; 4048 } else { 4049 PetscInt64 nz64 = 0; 4050 for (i=0; i<B->rmap->n; i++) {b->imax[i] = nnz[i]; nz64 += nnz[i];} 4051 ierr = PetscIntCast(nz64,&nz);CHKERRQ(ierr); 4052 } 4053 4054 /* allocate the matrix space */ 4055 /* FIXME: should B's old memory be unlogged? */ 4056 ierr = MatSeqXAIJFreeAIJ(B,&b->a,&b->j,&b->i);CHKERRQ(ierr); 4057 if (B->structure_only) { 4058 ierr = PetscMalloc1(nz,&b->j);CHKERRQ(ierr); 4059 ierr = PetscMalloc1(B->rmap->n+1,&b->i);CHKERRQ(ierr); 4060 ierr = PetscLogObjectMemory((PetscObject)B,(B->rmap->n+1)*sizeof(PetscInt)+nz*sizeof(PetscInt));CHKERRQ(ierr); 4061 } else { 4062 ierr = PetscMalloc3(nz,&b->a,nz,&b->j,B->rmap->n+1,&b->i);CHKERRQ(ierr); 4063 ierr = PetscLogObjectMemory((PetscObject)B,(B->rmap->n+1)*sizeof(PetscInt)+nz*(sizeof(PetscScalar)+sizeof(PetscInt)));CHKERRQ(ierr); 4064 } 4065 b->i[0] = 0; 4066 for (i=1; i<B->rmap->n+1; i++) { 4067 b->i[i] = b->i[i-1] + b->imax[i-1]; 4068 } 4069 if (B->structure_only) { 4070 b->singlemalloc = PETSC_FALSE; 4071 b->free_a = PETSC_FALSE; 4072 } else { 4073 b->singlemalloc = PETSC_TRUE; 4074 b->free_a = PETSC_TRUE; 4075 } 4076 b->free_ij = PETSC_TRUE; 4077 } else { 4078 b->free_a = PETSC_FALSE; 4079 b->free_ij = PETSC_FALSE; 4080 } 4081 4082 if (b->ipre && nnz != b->ipre && b->imax) { 4083 /* reserve user-requested sparsity */ 4084 ierr = PetscArraycpy(b->ipre,b->imax,B->rmap->n);CHKERRQ(ierr); 4085 } 4086 4087 4088 b->nz = 0; 4089 b->maxnz = nz; 4090 B->info.nz_unneeded = (double)b->maxnz; 4091 if (realalloc) { 4092 ierr = MatSetOption(B,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_TRUE);CHKERRQ(ierr); 4093 } 4094 B->was_assembled = PETSC_FALSE; 4095 B->assembled = PETSC_FALSE; 4096 PetscFunctionReturn(0); 4097 } 4098 4099 4100 PetscErrorCode MatResetPreallocation_SeqAIJ(Mat A) 4101 { 4102 Mat_SeqAIJ *a; 4103 PetscInt i; 4104 PetscErrorCode ierr; 4105 4106 PetscFunctionBegin; 4107 PetscValidHeaderSpecific(A,MAT_CLASSID,1); 4108 4109 /* Check local size. If zero, then return */ 4110 if (!A->rmap->n) PetscFunctionReturn(0); 4111 4112 a = (Mat_SeqAIJ*)A->data; 4113 /* if no saved info, we error out */ 4114 if (!a->ipre) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_NULL,"No saved preallocation info \n"); 4115 4116 if (!a->i || !a->j || !a->a || !a->imax || !a->ilen) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_NULL,"Memory info is incomplete, and can not reset preallocation \n"); 4117 4118 ierr = PetscArraycpy(a->imax,a->ipre,A->rmap->n);CHKERRQ(ierr); 4119 ierr = PetscArrayzero(a->ilen,A->rmap->n);CHKERRQ(ierr); 4120 a->i[0] = 0; 4121 for (i=1; i<A->rmap->n+1; i++) { 4122 a->i[i] = a->i[i-1] + a->imax[i-1]; 4123 } 4124 A->preallocated = PETSC_TRUE; 4125 a->nz = 0; 4126 a->maxnz = a->i[A->rmap->n]; 4127 A->info.nz_unneeded = (double)a->maxnz; 4128 A->was_assembled = PETSC_FALSE; 4129 A->assembled = PETSC_FALSE; 4130 PetscFunctionReturn(0); 4131 } 4132 4133 /*@ 4134 MatSeqAIJSetPreallocationCSR - Allocates memory for a sparse sequential matrix in AIJ format. 4135 4136 Input Parameters: 4137 + B - the matrix 4138 . i - the indices into j for the start of each row (starts with zero) 4139 . j - the column indices for each row (starts with zero) these must be sorted for each row 4140 - v - optional values in the matrix 4141 4142 Level: developer 4143 4144 Notes: 4145 The i,j,v values are COPIED with this routine; to avoid the copy use MatCreateSeqAIJWithArrays() 4146 4147 This routine may be called multiple times with different nonzero patterns (or the same nonzero pattern). The nonzero 4148 structure will be the union of all the previous nonzero structures. 4149 4150 Developer Notes: 4151 An optimization could be added to the implementation where it checks if the i, and j are identical to the current i and j and 4152 then just copies the v values directly with PetscMemcpy(). 4153 4154 This routine could also take a PetscCopyMode argument to allow sharing the values instead of always copying them. 4155 4156 .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatSeqAIJSetPreallocation(), MatCreateSeqAIJ(), MATSEQAIJ, MatResetPreallocation() 4157 @*/ 4158 PetscErrorCode MatSeqAIJSetPreallocationCSR(Mat B,const PetscInt i[],const PetscInt j[],const PetscScalar v[]) 4159 { 4160 PetscErrorCode ierr; 4161 4162 PetscFunctionBegin; 4163 PetscValidHeaderSpecific(B,MAT_CLASSID,1); 4164 PetscValidType(B,1); 4165 ierr = PetscTryMethod(B,"MatSeqAIJSetPreallocationCSR_C",(Mat,const PetscInt[],const PetscInt[],const PetscScalar[]),(B,i,j,v));CHKERRQ(ierr); 4166 PetscFunctionReturn(0); 4167 } 4168 4169 PetscErrorCode MatSeqAIJSetPreallocationCSR_SeqAIJ(Mat B,const PetscInt Ii[],const PetscInt J[],const PetscScalar v[]) 4170 { 4171 PetscInt i; 4172 PetscInt m,n; 4173 PetscInt nz; 4174 PetscInt *nnz; 4175 PetscErrorCode ierr; 4176 4177 PetscFunctionBegin; 4178 if (Ii[0]) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE, "Ii[0] must be 0 it is %D", Ii[0]); 4179 4180 ierr = PetscLayoutSetUp(B->rmap);CHKERRQ(ierr); 4181 ierr = PetscLayoutSetUp(B->cmap);CHKERRQ(ierr); 4182 4183 ierr = MatGetSize(B, &m, &n);CHKERRQ(ierr); 4184 ierr = PetscMalloc1(m+1, &nnz);CHKERRQ(ierr); 4185 for (i = 0; i < m; i++) { 4186 nz = Ii[i+1]- Ii[i]; 4187 if (nz < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE, "Local row %D has a negative number of columns %D", i, nnz); 4188 nnz[i] = nz; 4189 } 4190 ierr = MatSeqAIJSetPreallocation(B, 0, nnz);CHKERRQ(ierr); 4191 ierr = PetscFree(nnz);CHKERRQ(ierr); 4192 4193 for (i = 0; i < m; i++) { 4194 ierr = MatSetValues_SeqAIJ(B, 1, &i, Ii[i+1] - Ii[i], J+Ii[i], v ? v + Ii[i] : NULL, INSERT_VALUES);CHKERRQ(ierr); 4195 } 4196 4197 ierr = MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4198 ierr = MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4199 4200 ierr = MatSetOption(B,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE);CHKERRQ(ierr); 4201 PetscFunctionReturn(0); 4202 } 4203 4204 #include <../src/mat/impls/dense/seq/dense.h> 4205 #include <petsc/private/kernels/petscaxpy.h> 4206 4207 /* 4208 Computes (B'*A')' since computing B*A directly is untenable 4209 4210 n p p 4211 [ ] [ ] [ ] 4212 m [ A ] * n [ B ] = m [ C ] 4213 [ ] [ ] [ ] 4214 4215 */ 4216 PetscErrorCode MatMatMultNumeric_SeqDense_SeqAIJ(Mat A,Mat B,Mat C) 4217 { 4218 PetscErrorCode ierr; 4219 Mat_SeqDense *sub_a = (Mat_SeqDense*)A->data; 4220 Mat_SeqAIJ *sub_b = (Mat_SeqAIJ*)B->data; 4221 Mat_SeqDense *sub_c = (Mat_SeqDense*)C->data; 4222 PetscInt i,j,n,m,q,p; 4223 const PetscInt *ii,*idx; 4224 const PetscScalar *b,*a,*a_q; 4225 PetscScalar *c,*c_q; 4226 PetscInt clda = sub_c->lda; 4227 PetscInt alda = sub_a->lda; 4228 4229 PetscFunctionBegin; 4230 m = A->rmap->n; 4231 n = A->cmap->n; 4232 p = B->cmap->n; 4233 a = sub_a->v; 4234 b = sub_b->a; 4235 c = sub_c->v; 4236 if (clda == m) { 4237 ierr = PetscArrayzero(c,m*p);CHKERRQ(ierr); 4238 } else { 4239 for (j=0;j<p;j++) 4240 for (i=0;i<m;i++) 4241 c[j*clda + i] = 0.0; 4242 } 4243 ii = sub_b->i; 4244 idx = sub_b->j; 4245 for (i=0; i<n; i++) { 4246 q = ii[i+1] - ii[i]; 4247 while (q-->0) { 4248 c_q = c + clda*(*idx); 4249 a_q = a + alda*i; 4250 PetscKernelAXPY(c_q,*b,a_q,m); 4251 idx++; 4252 b++; 4253 } 4254 } 4255 PetscFunctionReturn(0); 4256 } 4257 4258 PetscErrorCode MatMatMultSymbolic_SeqDense_SeqAIJ(Mat A,Mat B,PetscReal fill,Mat C) 4259 { 4260 PetscErrorCode ierr; 4261 PetscInt m=A->rmap->n,n=B->cmap->n; 4262 PetscBool cisdense; 4263 4264 PetscFunctionBegin; 4265 if (A->cmap->n != B->rmap->n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"A->cmap->n %D != B->rmap->n %D\n",A->cmap->n,B->rmap->n); 4266 ierr = MatSetSizes(C,m,n,m,n);CHKERRQ(ierr); 4267 ierr = MatSetBlockSizesFromMats(C,A,B);CHKERRQ(ierr); 4268 ierr = PetscObjectTypeCompareAny((PetscObject)C,&cisdense,MATSEQDENSE,MATSEQDENSECUDA,"");CHKERRQ(ierr); 4269 if (!cisdense) { 4270 ierr = MatSetType(C,MATDENSE);CHKERRQ(ierr); 4271 } 4272 ierr = MatSetUp(C);CHKERRQ(ierr); 4273 4274 C->ops->matmultnumeric = MatMatMultNumeric_SeqDense_SeqAIJ; 4275 PetscFunctionReturn(0); 4276 } 4277 4278 /* ----------------------------------------------------------------*/ 4279 /*MC 4280 MATSEQAIJ - MATSEQAIJ = "seqaij" - A matrix type to be used for sequential sparse matrices, 4281 based on compressed sparse row format. 4282 4283 Options Database Keys: 4284 . -mat_type seqaij - sets the matrix type to "seqaij" during a call to MatSetFromOptions() 4285 4286 Level: beginner 4287 4288 Notes: 4289 MatSetValues() may be called for this matrix type with a NULL argument for the numerical values, 4290 in this case the values associated with the rows and columns one passes in are set to zero 4291 in the matrix 4292 4293 MatSetOptions(,MAT_STRUCTURE_ONLY,PETSC_TRUE) may be called for this matrix type. In this no 4294 space is allocated for the nonzero entries and any entries passed with MatSetValues() are ignored 4295 4296 Developer Notes: 4297 It would be nice if all matrix formats supported passing NULL in for the numerical values 4298 4299 .seealso: MatCreateSeqAIJ(), MatSetFromOptions(), MatSetType(), MatCreate(), MatType 4300 M*/ 4301 4302 /*MC 4303 MATAIJ - MATAIJ = "aij" - A matrix type to be used for sparse matrices. 4304 4305 This matrix type is identical to MATSEQAIJ when constructed with a single process communicator, 4306 and MATMPIAIJ otherwise. As a result, for single process communicators, 4307 MatSeqAIJSetPreallocation is supported, and similarly MatMPIAIJSetPreallocation() is supported 4308 for communicators controlling multiple processes. It is recommended that you call both of 4309 the above preallocation routines for simplicity. 4310 4311 Options Database Keys: 4312 . -mat_type aij - sets the matrix type to "aij" during a call to MatSetFromOptions() 4313 4314 Developer Notes: 4315 Subclasses include MATAIJCUSPARSE, MATAIJPERM, MATAIJSELL, MATAIJMKL, MATAIJCRL, and also automatically switches over to use inodes when 4316 enough exist. 4317 4318 Level: beginner 4319 4320 .seealso: MatCreateAIJ(), MatCreateSeqAIJ(), MATSEQAIJ,MATMPIAIJ 4321 M*/ 4322 4323 /*MC 4324 MATAIJCRL - MATAIJCRL = "aijcrl" - A matrix type to be used for sparse matrices. 4325 4326 This matrix type is identical to MATSEQAIJCRL when constructed with a single process communicator, 4327 and MATMPIAIJCRL otherwise. As a result, for single process communicators, 4328 MatSeqAIJSetPreallocation() is supported, and similarly MatMPIAIJSetPreallocation() is supported 4329 for communicators controlling multiple processes. It is recommended that you call both of 4330 the above preallocation routines for simplicity. 4331 4332 Options Database Keys: 4333 . -mat_type aijcrl - sets the matrix type to "aijcrl" during a call to MatSetFromOptions() 4334 4335 Level: beginner 4336 4337 .seealso: MatCreateMPIAIJCRL,MATSEQAIJCRL,MATMPIAIJCRL, MATSEQAIJCRL, MATMPIAIJCRL 4338 M*/ 4339 4340 PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqAIJCRL(Mat,MatType,MatReuse,Mat*); 4341 #if defined(PETSC_HAVE_ELEMENTAL) 4342 PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_Elemental(Mat,MatType,MatReuse,Mat*); 4343 #endif 4344 #if defined(PETSC_HAVE_SCALAPACK) 4345 PETSC_INTERN PetscErrorCode MatConvert_AIJ_ScaLAPACK(Mat,MatType,MatReuse,Mat*); 4346 #endif 4347 #if defined(PETSC_HAVE_HYPRE) 4348 PETSC_INTERN PetscErrorCode MatConvert_AIJ_HYPRE(Mat A,MatType,MatReuse,Mat*); 4349 #endif 4350 PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqDense(Mat,MatType,MatReuse,Mat*); 4351 4352 PETSC_EXTERN PetscErrorCode MatConvert_SeqAIJ_SeqSELL(Mat,MatType,MatReuse,Mat*); 4353 PETSC_INTERN PetscErrorCode MatConvert_XAIJ_IS(Mat,MatType,MatReuse,Mat*); 4354 PETSC_INTERN PetscErrorCode MatProductSetFromOptions_IS_XAIJ(Mat); 4355 4356 /*@C 4357 MatSeqAIJGetArray - gives read/write access to the array where the data for a MATSEQAIJ matrix is stored 4358 4359 Not Collective 4360 4361 Input Parameter: 4362 . mat - a MATSEQAIJ matrix 4363 4364 Output Parameter: 4365 . array - pointer to the data 4366 4367 Level: intermediate 4368 4369 .seealso: MatSeqAIJRestoreArray(), MatSeqAIJGetArrayF90() 4370 @*/ 4371 PetscErrorCode MatSeqAIJGetArray(Mat A,PetscScalar **array) 4372 { 4373 PetscErrorCode ierr; 4374 4375 PetscFunctionBegin; 4376 ierr = PetscUseMethod(A,"MatSeqAIJGetArray_C",(Mat,PetscScalar**),(A,array));CHKERRQ(ierr); 4377 PetscFunctionReturn(0); 4378 } 4379 4380 /*@C 4381 MatSeqAIJGetArrayRead - gives read-only access to the array where the data for a MATSEQAIJ matrix is stored 4382 4383 Not Collective 4384 4385 Input Parameter: 4386 . mat - a MATSEQAIJ matrix 4387 4388 Output Parameter: 4389 . array - pointer to the data 4390 4391 Level: intermediate 4392 4393 .seealso: MatSeqAIJGetArray(), MatSeqAIJRestoreArrayRead() 4394 @*/ 4395 PetscErrorCode MatSeqAIJGetArrayRead(Mat A,const PetscScalar **array) 4396 { 4397 #if defined(PETSC_HAVE_CUDA) || defined(PETSC_HAVE_VIENNACL) 4398 PetscOffloadMask oval; 4399 #endif 4400 PetscErrorCode ierr; 4401 4402 PetscFunctionBegin; 4403 #if defined(PETSC_HAVE_CUDA) || defined(PETSC_HAVE_VIENNACL) 4404 oval = A->offloadmask; 4405 #endif 4406 ierr = MatSeqAIJGetArray(A,(PetscScalar**)array);CHKERRQ(ierr); 4407 #if defined(PETSC_HAVE_CUDA) || defined(PETSC_HAVE_VIENNACL) 4408 if (oval == PETSC_OFFLOAD_GPU || oval == PETSC_OFFLOAD_BOTH) A->offloadmask = PETSC_OFFLOAD_BOTH; 4409 #endif 4410 PetscFunctionReturn(0); 4411 } 4412 4413 /*@C 4414 MatSeqAIJRestoreArrayRead - restore the read-only access array obtained from MatSeqAIJGetArrayRead 4415 4416 Not Collective 4417 4418 Input Parameter: 4419 . mat - a MATSEQAIJ matrix 4420 4421 Output Parameter: 4422 . array - pointer to the data 4423 4424 Level: intermediate 4425 4426 .seealso: MatSeqAIJGetArray(), MatSeqAIJGetArrayRead() 4427 @*/ 4428 PetscErrorCode MatSeqAIJRestoreArrayRead(Mat A,const PetscScalar **array) 4429 { 4430 #if defined(PETSC_HAVE_CUDA) || defined(PETSC_HAVE_VIENNACL) 4431 PetscOffloadMask oval; 4432 #endif 4433 PetscErrorCode ierr; 4434 4435 PetscFunctionBegin; 4436 #if defined(PETSC_HAVE_CUDA) || defined(PETSC_HAVE_VIENNACL) 4437 oval = A->offloadmask; 4438 #endif 4439 ierr = MatSeqAIJRestoreArray(A,(PetscScalar**)array);CHKERRQ(ierr); 4440 #if defined(PETSC_HAVE_CUDA) || defined(PETSC_HAVE_VIENNACL) 4441 A->offloadmask = oval; 4442 #endif 4443 PetscFunctionReturn(0); 4444 } 4445 4446 /*@C 4447 MatSeqAIJGetMaxRowNonzeros - returns the maximum number of nonzeros in any row 4448 4449 Not Collective 4450 4451 Input Parameter: 4452 . mat - a MATSEQAIJ matrix 4453 4454 Output Parameter: 4455 . nz - the maximum number of nonzeros in any row 4456 4457 Level: intermediate 4458 4459 .seealso: MatSeqAIJRestoreArray(), MatSeqAIJGetArrayF90() 4460 @*/ 4461 PetscErrorCode MatSeqAIJGetMaxRowNonzeros(Mat A,PetscInt *nz) 4462 { 4463 Mat_SeqAIJ *aij = (Mat_SeqAIJ*)A->data; 4464 4465 PetscFunctionBegin; 4466 *nz = aij->rmax; 4467 PetscFunctionReturn(0); 4468 } 4469 4470 /*@C 4471 MatSeqAIJRestoreArray - returns access to the array where the data for a MATSEQAIJ matrix is stored obtained by MatSeqAIJGetArray() 4472 4473 Not Collective 4474 4475 Input Parameters: 4476 + mat - a MATSEQAIJ matrix 4477 - array - pointer to the data 4478 4479 Level: intermediate 4480 4481 .seealso: MatSeqAIJGetArray(), MatSeqAIJRestoreArrayF90() 4482 @*/ 4483 PetscErrorCode MatSeqAIJRestoreArray(Mat A,PetscScalar **array) 4484 { 4485 PetscErrorCode ierr; 4486 4487 PetscFunctionBegin; 4488 ierr = PetscUseMethod(A,"MatSeqAIJRestoreArray_C",(Mat,PetscScalar**),(A,array));CHKERRQ(ierr); 4489 PetscFunctionReturn(0); 4490 } 4491 4492 #if defined(PETSC_HAVE_CUDA) 4493 PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqAIJCUSPARSE(Mat); 4494 #endif 4495 4496 PETSC_EXTERN PetscErrorCode MatCreate_SeqAIJ(Mat B) 4497 { 4498 Mat_SeqAIJ *b; 4499 PetscErrorCode ierr; 4500 PetscMPIInt size; 4501 4502 PetscFunctionBegin; 4503 ierr = MPI_Comm_size(PetscObjectComm((PetscObject)B),&size);CHKERRQ(ierr); 4504 if (size > 1) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Comm must be of size 1"); 4505 4506 ierr = PetscNewLog(B,&b);CHKERRQ(ierr); 4507 4508 B->data = (void*)b; 4509 4510 ierr = PetscMemcpy(B->ops,&MatOps_Values,sizeof(struct _MatOps));CHKERRQ(ierr); 4511 if (B->sortedfull) B->ops->setvalues = MatSetValues_SeqAIJ_SortedFull; 4512 4513 b->row = NULL; 4514 b->col = NULL; 4515 b->icol = NULL; 4516 b->reallocs = 0; 4517 b->ignorezeroentries = PETSC_FALSE; 4518 b->roworiented = PETSC_TRUE; 4519 b->nonew = 0; 4520 b->diag = NULL; 4521 b->solve_work = NULL; 4522 B->spptr = NULL; 4523 b->saved_values = NULL; 4524 b->idiag = NULL; 4525 b->mdiag = NULL; 4526 b->ssor_work = NULL; 4527 b->omega = 1.0; 4528 b->fshift = 0.0; 4529 b->idiagvalid = PETSC_FALSE; 4530 b->ibdiagvalid = PETSC_FALSE; 4531 b->keepnonzeropattern = PETSC_FALSE; 4532 4533 ierr = PetscObjectChangeTypeName((PetscObject)B,MATSEQAIJ);CHKERRQ(ierr); 4534 ierr = PetscObjectComposeFunction((PetscObject)B,"MatSeqAIJGetArray_C",MatSeqAIJGetArray_SeqAIJ);CHKERRQ(ierr); 4535 ierr = PetscObjectComposeFunction((PetscObject)B,"MatSeqAIJRestoreArray_C",MatSeqAIJRestoreArray_SeqAIJ);CHKERRQ(ierr); 4536 4537 #if defined(PETSC_HAVE_MATLAB_ENGINE) 4538 ierr = PetscObjectComposeFunction((PetscObject)B,"PetscMatlabEnginePut_C",MatlabEnginePut_SeqAIJ);CHKERRQ(ierr); 4539 ierr = PetscObjectComposeFunction((PetscObject)B,"PetscMatlabEngineGet_C",MatlabEngineGet_SeqAIJ);CHKERRQ(ierr); 4540 #endif 4541 4542 ierr = PetscObjectComposeFunction((PetscObject)B,"MatSeqAIJSetColumnIndices_C",MatSeqAIJSetColumnIndices_SeqAIJ);CHKERRQ(ierr); 4543 ierr = PetscObjectComposeFunction((PetscObject)B,"MatStoreValues_C",MatStoreValues_SeqAIJ);CHKERRQ(ierr); 4544 ierr = PetscObjectComposeFunction((PetscObject)B,"MatRetrieveValues_C",MatRetrieveValues_SeqAIJ);CHKERRQ(ierr); 4545 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqaij_seqsbaij_C",MatConvert_SeqAIJ_SeqSBAIJ);CHKERRQ(ierr); 4546 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqaij_seqbaij_C",MatConvert_SeqAIJ_SeqBAIJ);CHKERRQ(ierr); 4547 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqaij_seqaijperm_C",MatConvert_SeqAIJ_SeqAIJPERM);CHKERRQ(ierr); 4548 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqaij_seqaijsell_C",MatConvert_SeqAIJ_SeqAIJSELL);CHKERRQ(ierr); 4549 #if defined(PETSC_HAVE_MKL_SPARSE) 4550 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqaij_seqaijmkl_C",MatConvert_SeqAIJ_SeqAIJMKL);CHKERRQ(ierr); 4551 #endif 4552 #if defined(PETSC_HAVE_CUDA) 4553 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqaij_seqaijcusparse_C",MatConvert_SeqAIJ_SeqAIJCUSPARSE);CHKERRQ(ierr); 4554 ierr = PetscObjectComposeFunction((PetscObject)B,"MatProductSetFromOptions_seqaijcusparse_seqaij_C",MatProductSetFromOptions_SeqAIJ);CHKERRQ(ierr); 4555 #endif 4556 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqaij_seqaijcrl_C",MatConvert_SeqAIJ_SeqAIJCRL);CHKERRQ(ierr); 4557 #if defined(PETSC_HAVE_ELEMENTAL) 4558 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqaij_elemental_C",MatConvert_SeqAIJ_Elemental);CHKERRQ(ierr); 4559 #endif 4560 #if defined(PETSC_HAVE_SCALAPACK) 4561 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqaij_scalapack_C",MatConvert_AIJ_ScaLAPACK);CHKERRQ(ierr); 4562 #endif 4563 #if defined(PETSC_HAVE_HYPRE) 4564 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqaij_hypre_C",MatConvert_AIJ_HYPRE);CHKERRQ(ierr); 4565 ierr = PetscObjectComposeFunction((PetscObject)B,"MatProductSetFromOptions_transpose_seqaij_seqaij_C",MatProductSetFromOptions_Transpose_AIJ_AIJ);CHKERRQ(ierr); 4566 #endif 4567 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqaij_seqdense_C",MatConvert_SeqAIJ_SeqDense);CHKERRQ(ierr); 4568 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqaij_seqsell_C",MatConvert_SeqAIJ_SeqSELL);CHKERRQ(ierr); 4569 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_seqaij_is_C",MatConvert_XAIJ_IS);CHKERRQ(ierr); 4570 ierr = PetscObjectComposeFunction((PetscObject)B,"MatIsTranspose_C",MatIsTranspose_SeqAIJ);CHKERRQ(ierr); 4571 ierr = PetscObjectComposeFunction((PetscObject)B,"MatIsHermitianTranspose_C",MatIsTranspose_SeqAIJ);CHKERRQ(ierr); 4572 ierr = PetscObjectComposeFunction((PetscObject)B,"MatSeqAIJSetPreallocation_C",MatSeqAIJSetPreallocation_SeqAIJ);CHKERRQ(ierr); 4573 ierr = PetscObjectComposeFunction((PetscObject)B,"MatResetPreallocation_C",MatResetPreallocation_SeqAIJ);CHKERRQ(ierr); 4574 ierr = PetscObjectComposeFunction((PetscObject)B,"MatSeqAIJSetPreallocationCSR_C",MatSeqAIJSetPreallocationCSR_SeqAIJ);CHKERRQ(ierr); 4575 ierr = PetscObjectComposeFunction((PetscObject)B,"MatReorderForNonzeroDiagonal_C",MatReorderForNonzeroDiagonal_SeqAIJ);CHKERRQ(ierr); 4576 ierr = PetscObjectComposeFunction((PetscObject)B,"MatProductSetFromOptions_is_seqaij_C",MatProductSetFromOptions_IS_XAIJ);CHKERRQ(ierr); 4577 ierr = PetscObjectComposeFunction((PetscObject)B,"MatProductSetFromOptions_seqdense_seqaij_C",MatProductSetFromOptions_SeqDense_SeqAIJ);CHKERRQ(ierr); 4578 ierr = PetscObjectComposeFunction((PetscObject)B,"MatProductSetFromOptions_seqaij_seqaij_C",MatProductSetFromOptions_SeqAIJ);CHKERRQ(ierr); 4579 ierr = MatCreate_SeqAIJ_Inode(B);CHKERRQ(ierr); 4580 ierr = PetscObjectChangeTypeName((PetscObject)B,MATSEQAIJ);CHKERRQ(ierr); 4581 ierr = MatSeqAIJSetTypeFromOptions(B);CHKERRQ(ierr); /* this allows changing the matrix subtype to say MATSEQAIJPERM */ 4582 PetscFunctionReturn(0); 4583 } 4584 4585 /* 4586 Given a matrix generated with MatGetFactor() duplicates all the information in A into B 4587 */ 4588 PetscErrorCode MatDuplicateNoCreate_SeqAIJ(Mat C,Mat A,MatDuplicateOption cpvalues,PetscBool mallocmatspace) 4589 { 4590 Mat_SeqAIJ *c = (Mat_SeqAIJ*)C->data,*a = (Mat_SeqAIJ*)A->data; 4591 PetscErrorCode ierr; 4592 PetscInt m = A->rmap->n,i; 4593 4594 PetscFunctionBegin; 4595 if (!A->assembled && cpvalues!=MAT_DO_NOT_COPY_VALUES) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Cannot duplicate unassembled matrix"); 4596 4597 C->factortype = A->factortype; 4598 c->row = NULL; 4599 c->col = NULL; 4600 c->icol = NULL; 4601 c->reallocs = 0; 4602 4603 C->assembled = PETSC_TRUE; 4604 4605 ierr = PetscLayoutReference(A->rmap,&C->rmap);CHKERRQ(ierr); 4606 ierr = PetscLayoutReference(A->cmap,&C->cmap);CHKERRQ(ierr); 4607 4608 ierr = PetscMalloc1(m,&c->imax);CHKERRQ(ierr); 4609 ierr = PetscMemcpy(c->imax,a->imax,m*sizeof(PetscInt));CHKERRQ(ierr); 4610 ierr = PetscMalloc1(m,&c->ilen);CHKERRQ(ierr); 4611 ierr = PetscMemcpy(c->ilen,a->ilen,m*sizeof(PetscInt));CHKERRQ(ierr); 4612 ierr = PetscLogObjectMemory((PetscObject)C, 2*m*sizeof(PetscInt));CHKERRQ(ierr); 4613 4614 /* allocate the matrix space */ 4615 if (mallocmatspace) { 4616 ierr = PetscMalloc3(a->i[m],&c->a,a->i[m],&c->j,m+1,&c->i);CHKERRQ(ierr); 4617 ierr = PetscLogObjectMemory((PetscObject)C, a->i[m]*(sizeof(PetscScalar)+sizeof(PetscInt))+(m+1)*sizeof(PetscInt));CHKERRQ(ierr); 4618 4619 c->singlemalloc = PETSC_TRUE; 4620 4621 ierr = PetscArraycpy(c->i,a->i,m+1);CHKERRQ(ierr); 4622 if (m > 0) { 4623 ierr = PetscArraycpy(c->j,a->j,a->i[m]);CHKERRQ(ierr); 4624 if (cpvalues == MAT_COPY_VALUES) { 4625 ierr = PetscArraycpy(c->a,a->a,a->i[m]);CHKERRQ(ierr); 4626 } else { 4627 ierr = PetscArrayzero(c->a,a->i[m]);CHKERRQ(ierr); 4628 } 4629 } 4630 } 4631 4632 c->ignorezeroentries = a->ignorezeroentries; 4633 c->roworiented = a->roworiented; 4634 c->nonew = a->nonew; 4635 if (a->diag) { 4636 ierr = PetscMalloc1(m+1,&c->diag);CHKERRQ(ierr); 4637 ierr = PetscMemcpy(c->diag,a->diag,m*sizeof(PetscInt));CHKERRQ(ierr); 4638 ierr = PetscLogObjectMemory((PetscObject)C,(m+1)*sizeof(PetscInt));CHKERRQ(ierr); 4639 } else c->diag = NULL; 4640 4641 c->solve_work = NULL; 4642 c->saved_values = NULL; 4643 c->idiag = NULL; 4644 c->ssor_work = NULL; 4645 c->keepnonzeropattern = a->keepnonzeropattern; 4646 c->free_a = PETSC_TRUE; 4647 c->free_ij = PETSC_TRUE; 4648 4649 c->rmax = a->rmax; 4650 c->nz = a->nz; 4651 c->maxnz = a->nz; /* Since we allocate exactly the right amount */ 4652 C->preallocated = PETSC_TRUE; 4653 4654 c->compressedrow.use = a->compressedrow.use; 4655 c->compressedrow.nrows = a->compressedrow.nrows; 4656 if (a->compressedrow.use) { 4657 i = a->compressedrow.nrows; 4658 ierr = PetscMalloc2(i+1,&c->compressedrow.i,i,&c->compressedrow.rindex);CHKERRQ(ierr); 4659 ierr = PetscArraycpy(c->compressedrow.i,a->compressedrow.i,i+1);CHKERRQ(ierr); 4660 ierr = PetscArraycpy(c->compressedrow.rindex,a->compressedrow.rindex,i);CHKERRQ(ierr); 4661 } else { 4662 c->compressedrow.use = PETSC_FALSE; 4663 c->compressedrow.i = NULL; 4664 c->compressedrow.rindex = NULL; 4665 } 4666 c->nonzerorowcnt = a->nonzerorowcnt; 4667 C->nonzerostate = A->nonzerostate; 4668 4669 ierr = MatDuplicate_SeqAIJ_Inode(A,cpvalues,&C);CHKERRQ(ierr); 4670 ierr = PetscFunctionListDuplicate(((PetscObject)A)->qlist,&((PetscObject)C)->qlist);CHKERRQ(ierr); 4671 PetscFunctionReturn(0); 4672 } 4673 4674 PetscErrorCode MatDuplicate_SeqAIJ(Mat A,MatDuplicateOption cpvalues,Mat *B) 4675 { 4676 PetscErrorCode ierr; 4677 4678 PetscFunctionBegin; 4679 ierr = MatCreate(PetscObjectComm((PetscObject)A),B);CHKERRQ(ierr); 4680 ierr = MatSetSizes(*B,A->rmap->n,A->cmap->n,A->rmap->n,A->cmap->n);CHKERRQ(ierr); 4681 if (!(A->rmap->n % A->rmap->bs) && !(A->cmap->n % A->cmap->bs)) { 4682 ierr = MatSetBlockSizesFromMats(*B,A,A);CHKERRQ(ierr); 4683 } 4684 ierr = MatSetType(*B,((PetscObject)A)->type_name);CHKERRQ(ierr); 4685 ierr = MatDuplicateNoCreate_SeqAIJ(*B,A,cpvalues,PETSC_TRUE);CHKERRQ(ierr); 4686 PetscFunctionReturn(0); 4687 } 4688 4689 PetscErrorCode MatLoad_SeqAIJ(Mat newMat, PetscViewer viewer) 4690 { 4691 PetscBool isbinary, ishdf5; 4692 PetscErrorCode ierr; 4693 4694 PetscFunctionBegin; 4695 PetscValidHeaderSpecific(newMat,MAT_CLASSID,1); 4696 PetscValidHeaderSpecific(viewer,PETSC_VIEWER_CLASSID,2); 4697 /* force binary viewer to load .info file if it has not yet done so */ 4698 ierr = PetscViewerSetUp(viewer);CHKERRQ(ierr); 4699 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERBINARY,&isbinary);CHKERRQ(ierr); 4700 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERHDF5, &ishdf5);CHKERRQ(ierr); 4701 if (isbinary) { 4702 ierr = MatLoad_SeqAIJ_Binary(newMat,viewer);CHKERRQ(ierr); 4703 } else if (ishdf5) { 4704 #if defined(PETSC_HAVE_HDF5) 4705 ierr = MatLoad_AIJ_HDF5(newMat,viewer);CHKERRQ(ierr); 4706 #else 4707 SETERRQ(PetscObjectComm((PetscObject)newMat),PETSC_ERR_SUP,"HDF5 not supported in this build.\nPlease reconfigure using --download-hdf5"); 4708 #endif 4709 } else { 4710 SETERRQ2(PetscObjectComm((PetscObject)newMat),PETSC_ERR_SUP,"Viewer type %s not yet supported for reading %s matrices",((PetscObject)viewer)->type_name,((PetscObject)newMat)->type_name); 4711 } 4712 PetscFunctionReturn(0); 4713 } 4714 4715 PetscErrorCode MatLoad_SeqAIJ_Binary(Mat mat, PetscViewer viewer) 4716 { 4717 Mat_SeqAIJ *a = (Mat_SeqAIJ*)mat->data; 4718 PetscErrorCode ierr; 4719 PetscInt header[4],*rowlens,M,N,nz,sum,rows,cols,i; 4720 4721 PetscFunctionBegin; 4722 ierr = PetscViewerSetUp(viewer);CHKERRQ(ierr); 4723 4724 /* read in matrix header */ 4725 ierr = PetscViewerBinaryRead(viewer,header,4,NULL,PETSC_INT);CHKERRQ(ierr); 4726 if (header[0] != MAT_FILE_CLASSID) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED,"Not a matrix object in file"); 4727 M = header[1]; N = header[2]; nz = header[3]; 4728 if (M < 0) SETERRQ1(PetscObjectComm((PetscObject)viewer),PETSC_ERR_FILE_UNEXPECTED,"Matrix row size (%D) in file is negative",M); 4729 if (N < 0) SETERRQ1(PetscObjectComm((PetscObject)viewer),PETSC_ERR_FILE_UNEXPECTED,"Matrix column size (%D) in file is negative",N); 4730 if (nz < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED,"Matrix stored in special format on disk, cannot load as SeqAIJ"); 4731 4732 /* set block sizes from the viewer's .info file */ 4733 ierr = MatLoad_Binary_BlockSizes(mat,viewer);CHKERRQ(ierr); 4734 /* set local and global sizes if not set already */ 4735 if (mat->rmap->n < 0) mat->rmap->n = M; 4736 if (mat->cmap->n < 0) mat->cmap->n = N; 4737 if (mat->rmap->N < 0) mat->rmap->N = M; 4738 if (mat->cmap->N < 0) mat->cmap->N = N; 4739 ierr = PetscLayoutSetUp(mat->rmap);CHKERRQ(ierr); 4740 ierr = PetscLayoutSetUp(mat->cmap);CHKERRQ(ierr); 4741 4742 /* check if the matrix sizes are correct */ 4743 ierr = MatGetSize(mat,&rows,&cols);CHKERRQ(ierr); 4744 if (M != rows || N != cols) SETERRQ4(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED, "Matrix in file of different sizes (%D, %D) than the input matrix (%D, %D)",M,N,rows,cols); 4745 4746 /* read in row lengths */ 4747 ierr = PetscMalloc1(M,&rowlens);CHKERRQ(ierr); 4748 ierr = PetscViewerBinaryRead(viewer,rowlens,M,NULL,PETSC_INT);CHKERRQ(ierr); 4749 /* check if sum(rowlens) is same as nz */ 4750 sum = 0; for (i=0; i<M; i++) sum += rowlens[i]; 4751 if (sum != nz) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED,"Inconsistent matrix data in file: nonzeros = %D, sum-row-lengths = %D\n",nz,sum); 4752 /* preallocate and check sizes */ 4753 ierr = MatSeqAIJSetPreallocation_SeqAIJ(mat,0,rowlens);CHKERRQ(ierr); 4754 ierr = MatGetSize(mat,&rows,&cols);CHKERRQ(ierr); 4755 if (M != rows || N != cols) SETERRQ4(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED, "Matrix in file of different length (%D, %D) than the input matrix (%D, %D)",M,N,rows,cols); 4756 /* store row lengths */ 4757 ierr = PetscArraycpy(a->ilen,rowlens,M);CHKERRQ(ierr); 4758 ierr = PetscFree(rowlens);CHKERRQ(ierr); 4759 4760 /* fill in "i" row pointers */ 4761 a->i[0] = 0; for (i=0; i<M; i++) a->i[i+1] = a->i[i] + a->ilen[i]; 4762 /* read in "j" column indices */ 4763 ierr = PetscViewerBinaryRead(viewer,a->j,nz,NULL,PETSC_INT);CHKERRQ(ierr); 4764 /* read in "a" nonzero values */ 4765 ierr = PetscViewerBinaryRead(viewer,a->a,nz,NULL,PETSC_SCALAR);CHKERRQ(ierr); 4766 4767 ierr = MatAssemblyBegin(mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4768 ierr = MatAssemblyEnd(mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4769 PetscFunctionReturn(0); 4770 } 4771 4772 PetscErrorCode MatEqual_SeqAIJ(Mat A,Mat B,PetscBool * flg) 4773 { 4774 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data,*b = (Mat_SeqAIJ*)B->data; 4775 PetscErrorCode ierr; 4776 #if defined(PETSC_USE_COMPLEX) 4777 PetscInt k; 4778 #endif 4779 4780 PetscFunctionBegin; 4781 /* If the matrix dimensions are not equal,or no of nonzeros */ 4782 if ((A->rmap->n != B->rmap->n) || (A->cmap->n != B->cmap->n) ||(a->nz != b->nz)) { 4783 *flg = PETSC_FALSE; 4784 PetscFunctionReturn(0); 4785 } 4786 4787 /* if the a->i are the same */ 4788 ierr = PetscArraycmp(a->i,b->i,A->rmap->n+1,flg);CHKERRQ(ierr); 4789 if (!*flg) PetscFunctionReturn(0); 4790 4791 /* if a->j are the same */ 4792 ierr = PetscArraycmp(a->j,b->j,a->nz,flg);CHKERRQ(ierr); 4793 if (!*flg) PetscFunctionReturn(0); 4794 4795 /* if a->a are the same */ 4796 #if defined(PETSC_USE_COMPLEX) 4797 for (k=0; k<a->nz; k++) { 4798 if (PetscRealPart(a->a[k]) != PetscRealPart(b->a[k]) || PetscImaginaryPart(a->a[k]) != PetscImaginaryPart(b->a[k])) { 4799 *flg = PETSC_FALSE; 4800 PetscFunctionReturn(0); 4801 } 4802 } 4803 #else 4804 ierr = PetscArraycmp(a->a,b->a,a->nz,flg);CHKERRQ(ierr); 4805 #endif 4806 PetscFunctionReturn(0); 4807 } 4808 4809 /*@ 4810 MatCreateSeqAIJWithArrays - Creates an sequential AIJ matrix using matrix elements (in CSR format) 4811 provided by the user. 4812 4813 Collective 4814 4815 Input Parameters: 4816 + comm - must be an MPI communicator of size 1 4817 . m - number of rows 4818 . n - number of columns 4819 . i - row indices; that is i[0] = 0, i[row] = i[row-1] + number of elements in that row of the matrix 4820 . j - column indices 4821 - a - matrix values 4822 4823 Output Parameter: 4824 . mat - the matrix 4825 4826 Level: intermediate 4827 4828 Notes: 4829 The i, j, and a arrays are not copied by this routine, the user must free these arrays 4830 once the matrix is destroyed and not before 4831 4832 You cannot set new nonzero locations into this matrix, that will generate an error. 4833 4834 The i and j indices are 0 based 4835 4836 The format which is used for the sparse matrix input, is equivalent to a 4837 row-major ordering.. i.e for the following matrix, the input data expected is 4838 as shown 4839 4840 $ 1 0 0 4841 $ 2 0 3 4842 $ 4 5 6 4843 $ 4844 $ i = {0,1,3,6} [size = nrow+1 = 3+1] 4845 $ j = {0,0,2,0,1,2} [size = 6]; values must be sorted for each row 4846 $ v = {1,2,3,4,5,6} [size = 6] 4847 4848 4849 .seealso: MatCreate(), MatCreateAIJ(), MatCreateSeqAIJ(), MatCreateMPIAIJWithArrays(), MatMPIAIJSetPreallocationCSR() 4850 4851 @*/ 4852 PetscErrorCode MatCreateSeqAIJWithArrays(MPI_Comm comm,PetscInt m,PetscInt n,PetscInt i[],PetscInt j[],PetscScalar a[],Mat *mat) 4853 { 4854 PetscErrorCode ierr; 4855 PetscInt ii; 4856 Mat_SeqAIJ *aij; 4857 PetscInt jj; 4858 4859 PetscFunctionBegin; 4860 if (m > 0 && i[0]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"i (row indices) must start with 0"); 4861 ierr = MatCreate(comm,mat);CHKERRQ(ierr); 4862 ierr = MatSetSizes(*mat,m,n,m,n);CHKERRQ(ierr); 4863 /* ierr = MatSetBlockSizes(*mat,,);CHKERRQ(ierr); */ 4864 ierr = MatSetType(*mat,MATSEQAIJ);CHKERRQ(ierr); 4865 ierr = MatSeqAIJSetPreallocation_SeqAIJ(*mat,MAT_SKIP_ALLOCATION,NULL);CHKERRQ(ierr); 4866 aij = (Mat_SeqAIJ*)(*mat)->data; 4867 ierr = PetscMalloc1(m,&aij->imax);CHKERRQ(ierr); 4868 ierr = PetscMalloc1(m,&aij->ilen);CHKERRQ(ierr); 4869 4870 aij->i = i; 4871 aij->j = j; 4872 aij->a = a; 4873 aij->singlemalloc = PETSC_FALSE; 4874 aij->nonew = -1; /*this indicates that inserting a new value in the matrix that generates a new nonzero is an error*/ 4875 aij->free_a = PETSC_FALSE; 4876 aij->free_ij = PETSC_FALSE; 4877 4878 for (ii=0; ii<m; ii++) { 4879 aij->ilen[ii] = aij->imax[ii] = i[ii+1] - i[ii]; 4880 if (PetscDefined(USE_DEBUG)) { 4881 if (i[ii+1] - i[ii] < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative row length in i (row indices) row = %D length = %D",ii,i[ii+1] - i[ii]); 4882 for (jj=i[ii]+1; jj<i[ii+1]; jj++) { 4883 if (j[jj] < j[jj-1]) SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Column entry number %D (actual column %D) in row %D is not sorted",jj-i[ii],j[jj],ii); 4884 if (j[jj] == j[jj-1]) SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Column entry number %D (actual column %D) in row %D is identical to previous entry",jj-i[ii],j[jj],ii); 4885 } 4886 } 4887 } 4888 if (PetscDefined(USE_DEBUG)) { 4889 for (ii=0; ii<aij->i[m]; ii++) { 4890 if (j[ii] < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative column index at location = %D index = %D",ii,j[ii]); 4891 if (j[ii] > n - 1) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Column index to large at location = %D index = %D",ii,j[ii]); 4892 } 4893 } 4894 4895 ierr = MatAssemblyBegin(*mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4896 ierr = MatAssemblyEnd(*mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4897 PetscFunctionReturn(0); 4898 } 4899 /*@C 4900 MatCreateSeqAIJFromTriple - Creates an sequential AIJ matrix using matrix elements (in COO format) 4901 provided by the user. 4902 4903 Collective 4904 4905 Input Parameters: 4906 + comm - must be an MPI communicator of size 1 4907 . m - number of rows 4908 . n - number of columns 4909 . i - row indices 4910 . j - column indices 4911 . a - matrix values 4912 . nz - number of nonzeros 4913 - idx - 0 or 1 based 4914 4915 Output Parameter: 4916 . mat - the matrix 4917 4918 Level: intermediate 4919 4920 Notes: 4921 The i and j indices are 0 based 4922 4923 The format which is used for the sparse matrix input, is equivalent to a 4924 row-major ordering.. i.e for the following matrix, the input data expected is 4925 as shown: 4926 4927 1 0 0 4928 2 0 3 4929 4 5 6 4930 4931 i = {0,1,1,2,2,2} 4932 j = {0,0,2,0,1,2} 4933 v = {1,2,3,4,5,6} 4934 4935 4936 .seealso: MatCreate(), MatCreateAIJ(), MatCreateSeqAIJ(), MatCreateSeqAIJWithArrays(), MatMPIAIJSetPreallocationCSR() 4937 4938 @*/ 4939 PetscErrorCode MatCreateSeqAIJFromTriple(MPI_Comm comm,PetscInt m,PetscInt n,PetscInt i[],PetscInt j[],PetscScalar a[],Mat *mat,PetscInt nz,PetscBool idx) 4940 { 4941 PetscErrorCode ierr; 4942 PetscInt ii, *nnz, one = 1,row,col; 4943 4944 4945 PetscFunctionBegin; 4946 ierr = PetscCalloc1(m,&nnz);CHKERRQ(ierr); 4947 for (ii = 0; ii < nz; ii++) { 4948 nnz[i[ii] - !!idx] += 1; 4949 } 4950 ierr = MatCreate(comm,mat);CHKERRQ(ierr); 4951 ierr = MatSetSizes(*mat,m,n,m,n);CHKERRQ(ierr); 4952 ierr = MatSetType(*mat,MATSEQAIJ);CHKERRQ(ierr); 4953 ierr = MatSeqAIJSetPreallocation_SeqAIJ(*mat,0,nnz);CHKERRQ(ierr); 4954 for (ii = 0; ii < nz; ii++) { 4955 if (idx) { 4956 row = i[ii] - 1; 4957 col = j[ii] - 1; 4958 } else { 4959 row = i[ii]; 4960 col = j[ii]; 4961 } 4962 ierr = MatSetValues(*mat,one,&row,one,&col,&a[ii],ADD_VALUES);CHKERRQ(ierr); 4963 } 4964 ierr = MatAssemblyBegin(*mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4965 ierr = MatAssemblyEnd(*mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4966 ierr = PetscFree(nnz);CHKERRQ(ierr); 4967 PetscFunctionReturn(0); 4968 } 4969 4970 PetscErrorCode MatSeqAIJInvalidateDiagonal(Mat A) 4971 { 4972 Mat_SeqAIJ *a=(Mat_SeqAIJ*)A->data; 4973 PetscErrorCode ierr; 4974 4975 PetscFunctionBegin; 4976 a->idiagvalid = PETSC_FALSE; 4977 a->ibdiagvalid = PETSC_FALSE; 4978 4979 ierr = MatSeqAIJInvalidateDiagonal_Inode(A);CHKERRQ(ierr); 4980 PetscFunctionReturn(0); 4981 } 4982 4983 PetscErrorCode MatCreateMPIMatConcatenateSeqMat_SeqAIJ(MPI_Comm comm,Mat inmat,PetscInt n,MatReuse scall,Mat *outmat) 4984 { 4985 PetscErrorCode ierr; 4986 PetscMPIInt size; 4987 4988 PetscFunctionBegin; 4989 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 4990 if (size == 1) { 4991 if (scall == MAT_INITIAL_MATRIX) { 4992 ierr = MatDuplicate(inmat,MAT_COPY_VALUES,outmat);CHKERRQ(ierr); 4993 } else { 4994 ierr = MatCopy(inmat,*outmat,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 4995 } 4996 } else { 4997 ierr = MatCreateMPIMatConcatenateSeqMat_MPIAIJ(comm,inmat,n,scall,outmat);CHKERRQ(ierr); 4998 } 4999 PetscFunctionReturn(0); 5000 } 5001 5002 /* 5003 Permute A into C's *local* index space using rowemb,colemb. 5004 The embedding are supposed to be injections and the above implies that the range of rowemb is a subset 5005 of [0,m), colemb is in [0,n). 5006 If pattern == DIFFERENT_NONZERO_PATTERN, C is preallocated according to A. 5007 */ 5008 PetscErrorCode MatSetSeqMat_SeqAIJ(Mat C,IS rowemb,IS colemb,MatStructure pattern,Mat B) 5009 { 5010 /* If making this function public, change the error returned in this function away from _PLIB. */ 5011 PetscErrorCode ierr; 5012 Mat_SeqAIJ *Baij; 5013 PetscBool seqaij; 5014 PetscInt m,n,*nz,i,j,count; 5015 PetscScalar v; 5016 const PetscInt *rowindices,*colindices; 5017 5018 PetscFunctionBegin; 5019 if (!B) PetscFunctionReturn(0); 5020 /* Check to make sure the target matrix (and embeddings) are compatible with C and each other. */ 5021 ierr = PetscObjectBaseTypeCompare((PetscObject)B,MATSEQAIJ,&seqaij);CHKERRQ(ierr); 5022 if (!seqaij) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Input matrix is of wrong type"); 5023 if (rowemb) { 5024 ierr = ISGetLocalSize(rowemb,&m);CHKERRQ(ierr); 5025 if (m != B->rmap->n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Row IS of size %D is incompatible with matrix row size %D",m,B->rmap->n); 5026 } else { 5027 if (C->rmap->n != B->rmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Input matrix is row-incompatible with the target matrix"); 5028 } 5029 if (colemb) { 5030 ierr = ISGetLocalSize(colemb,&n);CHKERRQ(ierr); 5031 if (n != B->cmap->n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Diag col IS of size %D is incompatible with input matrix col size %D",n,B->cmap->n); 5032 } else { 5033 if (C->cmap->n != B->cmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Input matrix is col-incompatible with the target matrix"); 5034 } 5035 5036 Baij = (Mat_SeqAIJ*)(B->data); 5037 if (pattern == DIFFERENT_NONZERO_PATTERN) { 5038 ierr = PetscMalloc1(B->rmap->n,&nz);CHKERRQ(ierr); 5039 for (i=0; i<B->rmap->n; i++) { 5040 nz[i] = Baij->i[i+1] - Baij->i[i]; 5041 } 5042 ierr = MatSeqAIJSetPreallocation(C,0,nz);CHKERRQ(ierr); 5043 ierr = PetscFree(nz);CHKERRQ(ierr); 5044 } 5045 if (pattern == SUBSET_NONZERO_PATTERN) { 5046 ierr = MatZeroEntries(C);CHKERRQ(ierr); 5047 } 5048 count = 0; 5049 rowindices = NULL; 5050 colindices = NULL; 5051 if (rowemb) { 5052 ierr = ISGetIndices(rowemb,&rowindices);CHKERRQ(ierr); 5053 } 5054 if (colemb) { 5055 ierr = ISGetIndices(colemb,&colindices);CHKERRQ(ierr); 5056 } 5057 for (i=0; i<B->rmap->n; i++) { 5058 PetscInt row; 5059 row = i; 5060 if (rowindices) row = rowindices[i]; 5061 for (j=Baij->i[i]; j<Baij->i[i+1]; j++) { 5062 PetscInt col; 5063 col = Baij->j[count]; 5064 if (colindices) col = colindices[col]; 5065 v = Baij->a[count]; 5066 ierr = MatSetValues(C,1,&row,1,&col,&v,INSERT_VALUES);CHKERRQ(ierr); 5067 ++count; 5068 } 5069 } 5070 /* FIXME: set C's nonzerostate correctly. */ 5071 /* Assembly for C is necessary. */ 5072 C->preallocated = PETSC_TRUE; 5073 C->assembled = PETSC_TRUE; 5074 C->was_assembled = PETSC_FALSE; 5075 PetscFunctionReturn(0); 5076 } 5077 5078 PetscFunctionList MatSeqAIJList = NULL; 5079 5080 /*@C 5081 MatSeqAIJSetType - Converts a MATSEQAIJ matrix to a subtype 5082 5083 Collective on Mat 5084 5085 Input Parameters: 5086 + mat - the matrix object 5087 - matype - matrix type 5088 5089 Options Database Key: 5090 . -mat_seqai_type <method> - for example seqaijcrl 5091 5092 5093 Level: intermediate 5094 5095 .seealso: PCSetType(), VecSetType(), MatCreate(), MatType, Mat 5096 @*/ 5097 PetscErrorCode MatSeqAIJSetType(Mat mat, MatType matype) 5098 { 5099 PetscErrorCode ierr,(*r)(Mat,MatType,MatReuse,Mat*); 5100 PetscBool sametype; 5101 5102 PetscFunctionBegin; 5103 PetscValidHeaderSpecific(mat,MAT_CLASSID,1); 5104 ierr = PetscObjectTypeCompare((PetscObject)mat,matype,&sametype);CHKERRQ(ierr); 5105 if (sametype) PetscFunctionReturn(0); 5106 5107 ierr = PetscFunctionListFind(MatSeqAIJList,matype,&r);CHKERRQ(ierr); 5108 if (!r) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_UNKNOWN_TYPE,"Unknown Mat type given: %s",matype); 5109 ierr = (*r)(mat,matype,MAT_INPLACE_MATRIX,&mat);CHKERRQ(ierr); 5110 PetscFunctionReturn(0); 5111 } 5112 5113 5114 /*@C 5115 MatSeqAIJRegister - - Adds a new sub-matrix type for sequential AIJ matrices 5116 5117 Not Collective 5118 5119 Input Parameters: 5120 + name - name of a new user-defined matrix type, for example MATSEQAIJCRL 5121 - function - routine to convert to subtype 5122 5123 Notes: 5124 MatSeqAIJRegister() may be called multiple times to add several user-defined solvers. 5125 5126 5127 Then, your matrix can be chosen with the procedural interface at runtime via the option 5128 $ -mat_seqaij_type my_mat 5129 5130 Level: advanced 5131 5132 .seealso: MatSeqAIJRegisterAll() 5133 5134 5135 Level: advanced 5136 @*/ 5137 PetscErrorCode MatSeqAIJRegister(const char sname[],PetscErrorCode (*function)(Mat,MatType,MatReuse,Mat *)) 5138 { 5139 PetscErrorCode ierr; 5140 5141 PetscFunctionBegin; 5142 ierr = MatInitializePackage();CHKERRQ(ierr); 5143 ierr = PetscFunctionListAdd(&MatSeqAIJList,sname,function);CHKERRQ(ierr); 5144 PetscFunctionReturn(0); 5145 } 5146 5147 PetscBool MatSeqAIJRegisterAllCalled = PETSC_FALSE; 5148 5149 /*@C 5150 MatSeqAIJRegisterAll - Registers all of the matrix subtypes of SeqAIJ 5151 5152 Not Collective 5153 5154 Level: advanced 5155 5156 Developers Note: CUSP and CUSPARSE do not yet support the MatConvert_SeqAIJ..() paradigm and thus cannot be registered here 5157 5158 .seealso: MatRegisterAll(), MatSeqAIJRegister() 5159 @*/ 5160 PetscErrorCode MatSeqAIJRegisterAll(void) 5161 { 5162 PetscErrorCode ierr; 5163 5164 PetscFunctionBegin; 5165 if (MatSeqAIJRegisterAllCalled) PetscFunctionReturn(0); 5166 MatSeqAIJRegisterAllCalled = PETSC_TRUE; 5167 5168 ierr = MatSeqAIJRegister(MATSEQAIJCRL, MatConvert_SeqAIJ_SeqAIJCRL);CHKERRQ(ierr); 5169 ierr = MatSeqAIJRegister(MATSEQAIJPERM, MatConvert_SeqAIJ_SeqAIJPERM);CHKERRQ(ierr); 5170 ierr = MatSeqAIJRegister(MATSEQAIJSELL, MatConvert_SeqAIJ_SeqAIJSELL);CHKERRQ(ierr); 5171 #if defined(PETSC_HAVE_MKL_SPARSE) 5172 ierr = MatSeqAIJRegister(MATSEQAIJMKL, MatConvert_SeqAIJ_SeqAIJMKL);CHKERRQ(ierr); 5173 #endif 5174 #if defined(PETSC_HAVE_VIENNACL) && defined(PETSC_HAVE_VIENNACL_NO_CUDA) 5175 ierr = MatSeqAIJRegister(MATMPIAIJVIENNACL, MatConvert_SeqAIJ_SeqAIJViennaCL);CHKERRQ(ierr); 5176 #endif 5177 PetscFunctionReturn(0); 5178 } 5179 5180 /* 5181 Special version for direct calls from Fortran 5182 */ 5183 #include <petsc/private/fortranimpl.h> 5184 #if defined(PETSC_HAVE_FORTRAN_CAPS) 5185 #define matsetvaluesseqaij_ MATSETVALUESSEQAIJ 5186 #elif !defined(PETSC_HAVE_FORTRAN_UNDERSCORE) 5187 #define matsetvaluesseqaij_ matsetvaluesseqaij 5188 #endif 5189 5190 /* Change these macros so can be used in void function */ 5191 #undef CHKERRQ 5192 #define CHKERRQ(ierr) CHKERRABORT(PetscObjectComm((PetscObject)A),ierr) 5193 #undef SETERRQ2 5194 #define SETERRQ2(comm,ierr,b,c,d) CHKERRABORT(comm,ierr) 5195 #undef SETERRQ3 5196 #define SETERRQ3(comm,ierr,b,c,d,e) CHKERRABORT(comm,ierr) 5197 5198 PETSC_EXTERN void matsetvaluesseqaij_(Mat *AA,PetscInt *mm,const PetscInt im[],PetscInt *nn,const PetscInt in[],const PetscScalar v[],InsertMode *isis, PetscErrorCode *_ierr) 5199 { 5200 Mat A = *AA; 5201 PetscInt m = *mm, n = *nn; 5202 InsertMode is = *isis; 5203 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 5204 PetscInt *rp,k,low,high,t,ii,row,nrow,i,col,l,rmax,N; 5205 PetscInt *imax,*ai,*ailen; 5206 PetscErrorCode ierr; 5207 PetscInt *aj,nonew = a->nonew,lastcol = -1; 5208 MatScalar *ap,value,*aa; 5209 PetscBool ignorezeroentries = a->ignorezeroentries; 5210 PetscBool roworiented = a->roworiented; 5211 5212 PetscFunctionBegin; 5213 MatCheckPreallocated(A,1); 5214 imax = a->imax; 5215 ai = a->i; 5216 ailen = a->ilen; 5217 aj = a->j; 5218 aa = a->a; 5219 5220 for (k=0; k<m; k++) { /* loop over added rows */ 5221 row = im[k]; 5222 if (row < 0) continue; 5223 if (PetscUnlikelyDebug(row >= A->rmap->n)) SETERRABORT(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_OUTOFRANGE,"Row too large"); 5224 rp = aj + ai[row]; ap = aa + ai[row]; 5225 rmax = imax[row]; nrow = ailen[row]; 5226 low = 0; 5227 high = nrow; 5228 for (l=0; l<n; l++) { /* loop over added columns */ 5229 if (in[l] < 0) continue; 5230 if (PetscUnlikelyDebug(in[l] >= A->cmap->n)) SETERRABORT(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_OUTOFRANGE,"Column too large"); 5231 col = in[l]; 5232 if (roworiented) value = v[l + k*n]; 5233 else value = v[k + l*m]; 5234 5235 if (value == 0.0 && ignorezeroentries && (is == ADD_VALUES)) continue; 5236 5237 if (col <= lastcol) low = 0; 5238 else high = nrow; 5239 lastcol = col; 5240 while (high-low > 5) { 5241 t = (low+high)/2; 5242 if (rp[t] > col) high = t; 5243 else low = t; 5244 } 5245 for (i=low; i<high; i++) { 5246 if (rp[i] > col) break; 5247 if (rp[i] == col) { 5248 if (is == ADD_VALUES) ap[i] += value; 5249 else ap[i] = value; 5250 goto noinsert; 5251 } 5252 } 5253 if (value == 0.0 && ignorezeroentries) goto noinsert; 5254 if (nonew == 1) goto noinsert; 5255 if (nonew == -1) SETERRABORT(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero in the matrix"); 5256 MatSeqXAIJReallocateAIJ(A,A->rmap->n,1,nrow,row,col,rmax,aa,ai,aj,rp,ap,imax,nonew,MatScalar); 5257 N = nrow++ - 1; a->nz++; high++; 5258 /* shift up all the later entries in this row */ 5259 for (ii=N; ii>=i; ii--) { 5260 rp[ii+1] = rp[ii]; 5261 ap[ii+1] = ap[ii]; 5262 } 5263 rp[i] = col; 5264 ap[i] = value; 5265 A->nonzerostate++; 5266 noinsert:; 5267 low = i + 1; 5268 } 5269 ailen[row] = nrow; 5270 } 5271 PetscFunctionReturnVoid(); 5272 } 5273