1 /* 2 Defines the basic matrix operations for the AIJ (compressed row) 3 matrix storage format. 4 */ 5 6 #include <../src/mat/impls/aij/seq/aij.h> /*I "petscmat.h" I*/ 7 #include <petscblaslapack.h> 8 #include <petscbt.h> 9 #include <petsc/private/kernels/blocktranspose.h> 10 11 PetscErrorCode MatSeqAIJSetTypeFromOptions(Mat A) 12 { 13 PetscBool flg; 14 char type[256]; 15 16 PetscFunctionBegin; 17 PetscObjectOptionsBegin((PetscObject)A); 18 PetscCall(PetscOptionsFList("-mat_seqaij_type", "Matrix SeqAIJ type", "MatSeqAIJSetType", MatSeqAIJList, "seqaij", type, 256, &flg)); 19 if (flg) PetscCall(MatSeqAIJSetType(A, type)); 20 PetscOptionsEnd(); 21 PetscFunctionReturn(0); 22 } 23 24 PetscErrorCode MatGetColumnReductions_SeqAIJ(Mat A, PetscInt type, PetscReal *reductions) 25 { 26 PetscInt i, m, n; 27 Mat_SeqAIJ *aij = (Mat_SeqAIJ *)A->data; 28 29 PetscFunctionBegin; 30 PetscCall(MatGetSize(A, &m, &n)); 31 PetscCall(PetscArrayzero(reductions, n)); 32 if (type == NORM_2) { 33 for (i = 0; i < aij->i[m]; i++) reductions[aij->j[i]] += PetscAbsScalar(aij->a[i] * aij->a[i]); 34 } else if (type == NORM_1) { 35 for (i = 0; i < aij->i[m]; i++) reductions[aij->j[i]] += PetscAbsScalar(aij->a[i]); 36 } else if (type == NORM_INFINITY) { 37 for (i = 0; i < aij->i[m]; i++) reductions[aij->j[i]] = PetscMax(PetscAbsScalar(aij->a[i]), reductions[aij->j[i]]); 38 } else if (type == REDUCTION_SUM_REALPART || type == REDUCTION_MEAN_REALPART) { 39 for (i = 0; i < aij->i[m]; i++) reductions[aij->j[i]] += PetscRealPart(aij->a[i]); 40 } else if (type == REDUCTION_SUM_IMAGINARYPART || type == REDUCTION_MEAN_IMAGINARYPART) { 41 for (i = 0; i < aij->i[m]; i++) reductions[aij->j[i]] += PetscImaginaryPart(aij->a[i]); 42 } else SETERRQ(PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Unknown reduction type"); 43 44 if (type == NORM_2) { 45 for (i = 0; i < n; i++) reductions[i] = PetscSqrtReal(reductions[i]); 46 } else if (type == REDUCTION_MEAN_REALPART || type == REDUCTION_MEAN_IMAGINARYPART) { 47 for (i = 0; i < n; i++) reductions[i] /= m; 48 } 49 PetscFunctionReturn(0); 50 } 51 52 PetscErrorCode MatFindOffBlockDiagonalEntries_SeqAIJ(Mat A, IS *is) 53 { 54 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 55 PetscInt i, m = A->rmap->n, cnt = 0, bs = A->rmap->bs; 56 const PetscInt *jj = a->j, *ii = a->i; 57 PetscInt *rows; 58 59 PetscFunctionBegin; 60 for (i = 0; i < m; i++) { 61 if ((ii[i] != ii[i + 1]) && ((jj[ii[i]] < bs * (i / bs)) || (jj[ii[i + 1] - 1] > bs * ((i + bs) / bs) - 1))) cnt++; 62 } 63 PetscCall(PetscMalloc1(cnt, &rows)); 64 cnt = 0; 65 for (i = 0; i < m; i++) { 66 if ((ii[i] != ii[i + 1]) && ((jj[ii[i]] < bs * (i / bs)) || (jj[ii[i + 1] - 1] > bs * ((i + bs) / bs) - 1))) { 67 rows[cnt] = i; 68 cnt++; 69 } 70 } 71 PetscCall(ISCreateGeneral(PETSC_COMM_SELF, cnt, rows, PETSC_OWN_POINTER, is)); 72 PetscFunctionReturn(0); 73 } 74 75 PetscErrorCode MatFindZeroDiagonals_SeqAIJ_Private(Mat A, PetscInt *nrows, PetscInt **zrows) 76 { 77 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 78 const MatScalar *aa; 79 PetscInt i, m = A->rmap->n, cnt = 0; 80 const PetscInt *ii = a->i, *jj = a->j, *diag; 81 PetscInt *rows; 82 83 PetscFunctionBegin; 84 PetscCall(MatSeqAIJGetArrayRead(A, &aa)); 85 PetscCall(MatMarkDiagonal_SeqAIJ(A)); 86 diag = a->diag; 87 for (i = 0; i < m; i++) { 88 if ((diag[i] >= ii[i + 1]) || (jj[diag[i]] != i) || (aa[diag[i]] == 0.0)) cnt++; 89 } 90 PetscCall(PetscMalloc1(cnt, &rows)); 91 cnt = 0; 92 for (i = 0; i < m; i++) { 93 if ((diag[i] >= ii[i + 1]) || (jj[diag[i]] != i) || (aa[diag[i]] == 0.0)) rows[cnt++] = i; 94 } 95 *nrows = cnt; 96 *zrows = rows; 97 PetscCall(MatSeqAIJRestoreArrayRead(A, &aa)); 98 PetscFunctionReturn(0); 99 } 100 101 PetscErrorCode MatFindZeroDiagonals_SeqAIJ(Mat A, IS *zrows) 102 { 103 PetscInt nrows, *rows; 104 105 PetscFunctionBegin; 106 *zrows = NULL; 107 PetscCall(MatFindZeroDiagonals_SeqAIJ_Private(A, &nrows, &rows)); 108 PetscCall(ISCreateGeneral(PetscObjectComm((PetscObject)A), nrows, rows, PETSC_OWN_POINTER, zrows)); 109 PetscFunctionReturn(0); 110 } 111 112 PetscErrorCode MatFindNonzeroRows_SeqAIJ(Mat A, IS *keptrows) 113 { 114 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 115 const MatScalar *aa; 116 PetscInt m = A->rmap->n, cnt = 0; 117 const PetscInt *ii; 118 PetscInt n, i, j, *rows; 119 120 PetscFunctionBegin; 121 PetscCall(MatSeqAIJGetArrayRead(A, &aa)); 122 *keptrows = NULL; 123 ii = a->i; 124 for (i = 0; i < m; i++) { 125 n = ii[i + 1] - ii[i]; 126 if (!n) { 127 cnt++; 128 goto ok1; 129 } 130 for (j = ii[i]; j < ii[i + 1]; j++) { 131 if (aa[j] != 0.0) goto ok1; 132 } 133 cnt++; 134 ok1:; 135 } 136 if (!cnt) { 137 PetscCall(MatSeqAIJRestoreArrayRead(A, &aa)); 138 PetscFunctionReturn(0); 139 } 140 PetscCall(PetscMalloc1(A->rmap->n - cnt, &rows)); 141 cnt = 0; 142 for (i = 0; i < m; i++) { 143 n = ii[i + 1] - ii[i]; 144 if (!n) continue; 145 for (j = ii[i]; j < ii[i + 1]; j++) { 146 if (aa[j] != 0.0) { 147 rows[cnt++] = i; 148 break; 149 } 150 } 151 } 152 PetscCall(MatSeqAIJRestoreArrayRead(A, &aa)); 153 PetscCall(ISCreateGeneral(PETSC_COMM_SELF, cnt, rows, PETSC_OWN_POINTER, keptrows)); 154 PetscFunctionReturn(0); 155 } 156 157 PetscErrorCode MatDiagonalSet_SeqAIJ(Mat Y, Vec D, InsertMode is) 158 { 159 Mat_SeqAIJ *aij = (Mat_SeqAIJ *)Y->data; 160 PetscInt i, m = Y->rmap->n; 161 const PetscInt *diag; 162 MatScalar *aa; 163 const PetscScalar *v; 164 PetscBool missing; 165 166 PetscFunctionBegin; 167 if (Y->assembled) { 168 PetscCall(MatMissingDiagonal_SeqAIJ(Y, &missing, NULL)); 169 if (!missing) { 170 diag = aij->diag; 171 PetscCall(VecGetArrayRead(D, &v)); 172 PetscCall(MatSeqAIJGetArray(Y, &aa)); 173 if (is == INSERT_VALUES) { 174 for (i = 0; i < m; i++) aa[diag[i]] = v[i]; 175 } else { 176 for (i = 0; i < m; i++) aa[diag[i]] += v[i]; 177 } 178 PetscCall(MatSeqAIJRestoreArray(Y, &aa)); 179 PetscCall(VecRestoreArrayRead(D, &v)); 180 PetscFunctionReturn(0); 181 } 182 PetscCall(MatSeqAIJInvalidateDiagonal(Y)); 183 } 184 PetscCall(MatDiagonalSet_Default(Y, D, is)); 185 PetscFunctionReturn(0); 186 } 187 188 PetscErrorCode MatGetRowIJ_SeqAIJ(Mat A, PetscInt oshift, PetscBool symmetric, PetscBool inodecompressed, PetscInt *m, const PetscInt *ia[], const PetscInt *ja[], PetscBool *done) 189 { 190 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 191 PetscInt i, ishift; 192 193 PetscFunctionBegin; 194 if (m) *m = A->rmap->n; 195 if (!ia) PetscFunctionReturn(0); 196 ishift = 0; 197 if (symmetric && A->structurally_symmetric != PETSC_BOOL3_TRUE) { 198 PetscCall(MatToSymmetricIJ_SeqAIJ(A->rmap->n, a->i, a->j, PETSC_TRUE, ishift, oshift, (PetscInt **)ia, (PetscInt **)ja)); 199 } else if (oshift == 1) { 200 PetscInt *tia; 201 PetscInt nz = a->i[A->rmap->n]; 202 /* malloc space and add 1 to i and j indices */ 203 PetscCall(PetscMalloc1(A->rmap->n + 1, &tia)); 204 for (i = 0; i < A->rmap->n + 1; i++) tia[i] = a->i[i] + 1; 205 *ia = tia; 206 if (ja) { 207 PetscInt *tja; 208 PetscCall(PetscMalloc1(nz + 1, &tja)); 209 for (i = 0; i < nz; i++) tja[i] = a->j[i] + 1; 210 *ja = tja; 211 } 212 } else { 213 *ia = a->i; 214 if (ja) *ja = a->j; 215 } 216 PetscFunctionReturn(0); 217 } 218 219 PetscErrorCode MatRestoreRowIJ_SeqAIJ(Mat A, PetscInt oshift, PetscBool symmetric, PetscBool inodecompressed, PetscInt *n, const PetscInt *ia[], const PetscInt *ja[], PetscBool *done) 220 { 221 PetscFunctionBegin; 222 if (!ia) PetscFunctionReturn(0); 223 if ((symmetric && A->structurally_symmetric != PETSC_BOOL3_TRUE) || oshift == 1) { 224 PetscCall(PetscFree(*ia)); 225 if (ja) PetscCall(PetscFree(*ja)); 226 } 227 PetscFunctionReturn(0); 228 } 229 230 PetscErrorCode MatGetColumnIJ_SeqAIJ(Mat A, PetscInt oshift, PetscBool symmetric, PetscBool inodecompressed, PetscInt *nn, const PetscInt *ia[], const PetscInt *ja[], PetscBool *done) 231 { 232 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 233 PetscInt i, *collengths, *cia, *cja, n = A->cmap->n, m = A->rmap->n; 234 PetscInt nz = a->i[m], row, *jj, mr, col; 235 236 PetscFunctionBegin; 237 *nn = n; 238 if (!ia) PetscFunctionReturn(0); 239 if (symmetric) { 240 PetscCall(MatToSymmetricIJ_SeqAIJ(A->rmap->n, a->i, a->j, PETSC_TRUE, 0, oshift, (PetscInt **)ia, (PetscInt **)ja)); 241 } else { 242 PetscCall(PetscCalloc1(n, &collengths)); 243 PetscCall(PetscMalloc1(n + 1, &cia)); 244 PetscCall(PetscMalloc1(nz, &cja)); 245 jj = a->j; 246 for (i = 0; i < nz; i++) collengths[jj[i]]++; 247 cia[0] = oshift; 248 for (i = 0; i < n; i++) cia[i + 1] = cia[i] + collengths[i]; 249 PetscCall(PetscArrayzero(collengths, n)); 250 jj = a->j; 251 for (row = 0; row < m; row++) { 252 mr = a->i[row + 1] - a->i[row]; 253 for (i = 0; i < mr; i++) { 254 col = *jj++; 255 256 cja[cia[col] + collengths[col]++ - oshift] = row + oshift; 257 } 258 } 259 PetscCall(PetscFree(collengths)); 260 *ia = cia; 261 *ja = cja; 262 } 263 PetscFunctionReturn(0); 264 } 265 266 PetscErrorCode MatRestoreColumnIJ_SeqAIJ(Mat A, PetscInt oshift, PetscBool symmetric, PetscBool inodecompressed, PetscInt *n, const PetscInt *ia[], const PetscInt *ja[], PetscBool *done) 267 { 268 PetscFunctionBegin; 269 if (!ia) PetscFunctionReturn(0); 270 271 PetscCall(PetscFree(*ia)); 272 PetscCall(PetscFree(*ja)); 273 PetscFunctionReturn(0); 274 } 275 276 /* 277 MatGetColumnIJ_SeqAIJ_Color() and MatRestoreColumnIJ_SeqAIJ_Color() are customized from 278 MatGetColumnIJ_SeqAIJ() and MatRestoreColumnIJ_SeqAIJ() by adding an output 279 spidx[], index of a->a, to be used in MatTransposeColoringCreate_SeqAIJ() and MatFDColoringCreate_SeqXAIJ() 280 */ 281 PetscErrorCode MatGetColumnIJ_SeqAIJ_Color(Mat A, PetscInt oshift, PetscBool symmetric, PetscBool inodecompressed, PetscInt *nn, const PetscInt *ia[], const PetscInt *ja[], PetscInt *spidx[], PetscBool *done) 282 { 283 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 284 PetscInt i, *collengths, *cia, *cja, n = A->cmap->n, m = A->rmap->n; 285 PetscInt nz = a->i[m], row, mr, col, tmp; 286 PetscInt *cspidx; 287 const PetscInt *jj; 288 289 PetscFunctionBegin; 290 *nn = n; 291 if (!ia) PetscFunctionReturn(0); 292 293 PetscCall(PetscCalloc1(n, &collengths)); 294 PetscCall(PetscMalloc1(n + 1, &cia)); 295 PetscCall(PetscMalloc1(nz, &cja)); 296 PetscCall(PetscMalloc1(nz, &cspidx)); 297 jj = a->j; 298 for (i = 0; i < nz; i++) collengths[jj[i]]++; 299 cia[0] = oshift; 300 for (i = 0; i < n; i++) cia[i + 1] = cia[i] + collengths[i]; 301 PetscCall(PetscArrayzero(collengths, n)); 302 jj = a->j; 303 for (row = 0; row < m; row++) { 304 mr = a->i[row + 1] - a->i[row]; 305 for (i = 0; i < mr; i++) { 306 col = *jj++; 307 tmp = cia[col] + collengths[col]++ - oshift; 308 cspidx[tmp] = a->i[row] + i; /* index of a->j */ 309 cja[tmp] = row + oshift; 310 } 311 } 312 PetscCall(PetscFree(collengths)); 313 *ia = cia; 314 *ja = cja; 315 *spidx = cspidx; 316 PetscFunctionReturn(0); 317 } 318 319 PetscErrorCode MatRestoreColumnIJ_SeqAIJ_Color(Mat A, PetscInt oshift, PetscBool symmetric, PetscBool inodecompressed, PetscInt *n, const PetscInt *ia[], const PetscInt *ja[], PetscInt *spidx[], PetscBool *done) 320 { 321 PetscFunctionBegin; 322 PetscCall(MatRestoreColumnIJ_SeqAIJ(A, oshift, symmetric, inodecompressed, n, ia, ja, done)); 323 PetscCall(PetscFree(*spidx)); 324 PetscFunctionReturn(0); 325 } 326 327 PetscErrorCode MatSetValuesRow_SeqAIJ(Mat A, PetscInt row, const PetscScalar v[]) 328 { 329 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 330 PetscInt *ai = a->i; 331 PetscScalar *aa; 332 333 PetscFunctionBegin; 334 PetscCall(MatSeqAIJGetArray(A, &aa)); 335 PetscCall(PetscArraycpy(aa + ai[row], v, ai[row + 1] - ai[row])); 336 PetscCall(MatSeqAIJRestoreArray(A, &aa)); 337 PetscFunctionReturn(0); 338 } 339 340 /* 341 MatSeqAIJSetValuesLocalFast - An optimized version of MatSetValuesLocal() for SeqAIJ matrices with several assumptions 342 343 - a single row of values is set with each call 344 - no row or column indices are negative or (in error) larger than the number of rows or columns 345 - the values are always added to the matrix, not set 346 - no new locations are introduced in the nonzero structure of the matrix 347 348 This does NOT assume the global column indices are sorted 349 350 */ 351 352 #include <petsc/private/isimpl.h> 353 PetscErrorCode MatSeqAIJSetValuesLocalFast(Mat A, PetscInt m, const PetscInt im[], PetscInt n, const PetscInt in[], const PetscScalar v[], InsertMode is) 354 { 355 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 356 PetscInt low, high, t, row, nrow, i, col, l; 357 const PetscInt *rp, *ai = a->i, *ailen = a->ilen, *aj = a->j; 358 PetscInt lastcol = -1; 359 MatScalar *ap, value, *aa; 360 const PetscInt *ridx = A->rmap->mapping->indices, *cidx = A->cmap->mapping->indices; 361 362 PetscFunctionBegin; 363 PetscCall(MatSeqAIJGetArray(A, &aa)); 364 row = ridx[im[0]]; 365 rp = aj + ai[row]; 366 ap = aa + ai[row]; 367 nrow = ailen[row]; 368 low = 0; 369 high = nrow; 370 for (l = 0; l < n; l++) { /* loop over added columns */ 371 col = cidx[in[l]]; 372 value = v[l]; 373 374 if (col <= lastcol) low = 0; 375 else high = nrow; 376 lastcol = col; 377 while (high - low > 5) { 378 t = (low + high) / 2; 379 if (rp[t] > col) high = t; 380 else low = t; 381 } 382 for (i = low; i < high; i++) { 383 if (rp[i] == col) { 384 ap[i] += value; 385 low = i + 1; 386 break; 387 } 388 } 389 } 390 PetscCall(MatSeqAIJRestoreArray(A, &aa)); 391 return 0; 392 } 393 394 PetscErrorCode MatSetValues_SeqAIJ(Mat A, PetscInt m, const PetscInt im[], PetscInt n, const PetscInt in[], const PetscScalar v[], InsertMode is) 395 { 396 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 397 PetscInt *rp, k, low, high, t, ii, row, nrow, i, col, l, rmax, N; 398 PetscInt *imax = a->imax, *ai = a->i, *ailen = a->ilen; 399 PetscInt *aj = a->j, nonew = a->nonew, lastcol = -1; 400 MatScalar *ap = NULL, value = 0.0, *aa; 401 PetscBool ignorezeroentries = a->ignorezeroentries; 402 PetscBool roworiented = a->roworiented; 403 404 PetscFunctionBegin; 405 PetscCall(MatSeqAIJGetArray(A, &aa)); 406 for (k = 0; k < m; k++) { /* loop over added rows */ 407 row = im[k]; 408 if (row < 0) continue; 409 PetscCheck(row < A->rmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Row too large: row %" PetscInt_FMT " max %" PetscInt_FMT, row, A->rmap->n - 1); 410 rp = aj + ai[row]; 411 if (!A->structure_only) ap = aa + ai[row]; 412 rmax = imax[row]; 413 nrow = ailen[row]; 414 low = 0; 415 high = nrow; 416 for (l = 0; l < n; l++) { /* loop over added columns */ 417 if (in[l] < 0) continue; 418 PetscCheck(in[l] < A->cmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Column too large: col %" PetscInt_FMT " max %" PetscInt_FMT, in[l], A->cmap->n - 1); 419 col = in[l]; 420 if (v && !A->structure_only) value = roworiented ? v[l + k * n] : v[k + l * m]; 421 if (!A->structure_only && value == 0.0 && ignorezeroentries && is == ADD_VALUES && row != col) continue; 422 423 if (col <= lastcol) low = 0; 424 else high = nrow; 425 lastcol = col; 426 while (high - low > 5) { 427 t = (low + high) / 2; 428 if (rp[t] > col) high = t; 429 else low = t; 430 } 431 for (i = low; i < high; i++) { 432 if (rp[i] > col) break; 433 if (rp[i] == col) { 434 if (!A->structure_only) { 435 if (is == ADD_VALUES) { 436 ap[i] += value; 437 (void)PetscLogFlops(1.0); 438 } else ap[i] = value; 439 } 440 low = i + 1; 441 goto noinsert; 442 } 443 } 444 if (value == 0.0 && ignorezeroentries && row != col) goto noinsert; 445 if (nonew == 1) goto noinsert; 446 PetscCheck(nonew != -1, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Inserting a new nonzero at (%" PetscInt_FMT ",%" PetscInt_FMT ") in the matrix", row, col); 447 if (A->structure_only) { 448 MatSeqXAIJReallocateAIJ_structure_only(A, A->rmap->n, 1, nrow, row, col, rmax, ai, aj, rp, imax, nonew, MatScalar); 449 } else { 450 MatSeqXAIJReallocateAIJ(A, A->rmap->n, 1, nrow, row, col, rmax, aa, ai, aj, rp, ap, imax, nonew, MatScalar); 451 } 452 N = nrow++ - 1; 453 a->nz++; 454 high++; 455 /* shift up all the later entries in this row */ 456 PetscCall(PetscArraymove(rp + i + 1, rp + i, N - i + 1)); 457 rp[i] = col; 458 if (!A->structure_only) { 459 PetscCall(PetscArraymove(ap + i + 1, ap + i, N - i + 1)); 460 ap[i] = value; 461 } 462 low = i + 1; 463 A->nonzerostate++; 464 noinsert:; 465 } 466 ailen[row] = nrow; 467 } 468 PetscCall(MatSeqAIJRestoreArray(A, &aa)); 469 PetscFunctionReturn(0); 470 } 471 472 PetscErrorCode MatSetValues_SeqAIJ_SortedFullNoPreallocation(Mat A, PetscInt m, const PetscInt im[], PetscInt n, const PetscInt in[], const PetscScalar v[], InsertMode is) 473 { 474 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 475 PetscInt *rp, k, row; 476 PetscInt *ai = a->i; 477 PetscInt *aj = a->j; 478 MatScalar *aa, *ap; 479 480 PetscFunctionBegin; 481 PetscCheck(!A->was_assembled, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Cannot call on assembled matrix."); 482 PetscCheck(m * n + a->nz <= a->maxnz, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Number of entries in matrix will be larger than maximum nonzeros allocated for %" PetscInt_FMT " in MatSeqAIJSetTotalPreallocation()", a->maxnz); 483 484 PetscCall(MatSeqAIJGetArray(A, &aa)); 485 for (k = 0; k < m; k++) { /* loop over added rows */ 486 row = im[k]; 487 rp = aj + ai[row]; 488 ap = aa + ai[row]; 489 490 PetscCall(PetscMemcpy(rp, in, n * sizeof(PetscInt))); 491 if (!A->structure_only) { 492 if (v) { 493 PetscCall(PetscMemcpy(ap, v, n * sizeof(PetscScalar))); 494 v += n; 495 } else { 496 PetscCall(PetscMemzero(ap, n * sizeof(PetscScalar))); 497 } 498 } 499 a->ilen[row] = n; 500 a->imax[row] = n; 501 a->i[row + 1] = a->i[row] + n; 502 a->nz += n; 503 } 504 PetscCall(MatSeqAIJRestoreArray(A, &aa)); 505 PetscFunctionReturn(0); 506 } 507 508 /*@ 509 MatSeqAIJSetTotalPreallocation - Sets an upper bound on the total number of expected nonzeros in the matrix. 510 511 Input Parameters: 512 + A - the `MATSEQAIJ` matrix 513 - nztotal - bound on the number of nonzeros 514 515 Level: advanced 516 517 Notes: 518 This can be called if you will be provided the matrix row by row (from row zero) with sorted column indices for each row. 519 Simply call `MatSetValues()` after this call to provide the matrix entries in the usual manner. This matrix may be used 520 as always with multiple matrix assemblies. 521 522 .seealso: `MatSetOption()`, `MAT_SORTED_FULL`, `MatSetValues()`, `MatSeqAIJSetPreallocation()` 523 @*/ 524 525 PetscErrorCode MatSeqAIJSetTotalPreallocation(Mat A, PetscInt nztotal) 526 { 527 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 528 529 PetscFunctionBegin; 530 PetscCall(PetscLayoutSetUp(A->rmap)); 531 PetscCall(PetscLayoutSetUp(A->cmap)); 532 a->maxnz = nztotal; 533 if (!a->imax) { PetscCall(PetscMalloc1(A->rmap->n, &a->imax)); } 534 if (!a->ilen) { 535 PetscCall(PetscMalloc1(A->rmap->n, &a->ilen)); 536 } else { 537 PetscCall(PetscMemzero(a->ilen, A->rmap->n * sizeof(PetscInt))); 538 } 539 540 /* allocate the matrix space */ 541 if (A->structure_only) { 542 PetscCall(PetscMalloc1(nztotal, &a->j)); 543 PetscCall(PetscMalloc1(A->rmap->n + 1, &a->i)); 544 } else { 545 PetscCall(PetscMalloc3(nztotal, &a->a, nztotal, &a->j, A->rmap->n + 1, &a->i)); 546 } 547 a->i[0] = 0; 548 if (A->structure_only) { 549 a->singlemalloc = PETSC_FALSE; 550 a->free_a = PETSC_FALSE; 551 } else { 552 a->singlemalloc = PETSC_TRUE; 553 a->free_a = PETSC_TRUE; 554 } 555 a->free_ij = PETSC_TRUE; 556 A->ops->setvalues = MatSetValues_SeqAIJ_SortedFullNoPreallocation; 557 A->preallocated = PETSC_TRUE; 558 PetscFunctionReturn(0); 559 } 560 561 PetscErrorCode MatSetValues_SeqAIJ_SortedFull(Mat A, PetscInt m, const PetscInt im[], PetscInt n, const PetscInt in[], const PetscScalar v[], InsertMode is) 562 { 563 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 564 PetscInt *rp, k, row; 565 PetscInt *ai = a->i, *ailen = a->ilen; 566 PetscInt *aj = a->j; 567 MatScalar *aa, *ap; 568 569 PetscFunctionBegin; 570 PetscCall(MatSeqAIJGetArray(A, &aa)); 571 for (k = 0; k < m; k++) { /* loop over added rows */ 572 row = im[k]; 573 PetscCheck(n <= a->imax[row], PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Preallocation for row %" PetscInt_FMT " does not match number of columns provided", n); 574 rp = aj + ai[row]; 575 ap = aa + ai[row]; 576 if (!A->was_assembled) PetscCall(PetscMemcpy(rp, in, n * sizeof(PetscInt))); 577 if (!A->structure_only) { 578 if (v) { 579 PetscCall(PetscMemcpy(ap, v, n * sizeof(PetscScalar))); 580 v += n; 581 } else { 582 PetscCall(PetscMemzero(ap, n * sizeof(PetscScalar))); 583 } 584 } 585 ailen[row] = n; 586 a->nz += n; 587 } 588 PetscCall(MatSeqAIJRestoreArray(A, &aa)); 589 PetscFunctionReturn(0); 590 } 591 592 PetscErrorCode MatGetValues_SeqAIJ(Mat A, PetscInt m, const PetscInt im[], PetscInt n, const PetscInt in[], PetscScalar v[]) 593 { 594 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 595 PetscInt *rp, k, low, high, t, row, nrow, i, col, l, *aj = a->j; 596 PetscInt *ai = a->i, *ailen = a->ilen; 597 const MatScalar *ap, *aa; 598 599 PetscFunctionBegin; 600 PetscCall(MatSeqAIJGetArrayRead(A, &aa)); 601 for (k = 0; k < m; k++) { /* loop over rows */ 602 row = im[k]; 603 if (row < 0) { 604 v += n; 605 continue; 606 } /* negative row */ 607 PetscCheck(row < A->rmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Row too large: row %" PetscInt_FMT " max %" PetscInt_FMT, row, A->rmap->n - 1); 608 rp = aj + ai[row]; 609 ap = aa + ai[row]; 610 nrow = ailen[row]; 611 for (l = 0; l < n; l++) { /* loop over columns */ 612 if (in[l] < 0) { 613 v++; 614 continue; 615 } /* negative column */ 616 PetscCheck(in[l] < A->cmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Column too large: col %" PetscInt_FMT " max %" PetscInt_FMT, in[l], A->cmap->n - 1); 617 col = in[l]; 618 high = nrow; 619 low = 0; /* assume unsorted */ 620 while (high - low > 5) { 621 t = (low + high) / 2; 622 if (rp[t] > col) high = t; 623 else low = t; 624 } 625 for (i = low; i < high; i++) { 626 if (rp[i] > col) break; 627 if (rp[i] == col) { 628 *v++ = ap[i]; 629 goto finished; 630 } 631 } 632 *v++ = 0.0; 633 finished:; 634 } 635 } 636 PetscCall(MatSeqAIJRestoreArrayRead(A, &aa)); 637 PetscFunctionReturn(0); 638 } 639 640 PetscErrorCode MatView_SeqAIJ_Binary(Mat mat, PetscViewer viewer) 641 { 642 Mat_SeqAIJ *A = (Mat_SeqAIJ *)mat->data; 643 const PetscScalar *av; 644 PetscInt header[4], M, N, m, nz, i; 645 PetscInt *rowlens; 646 647 PetscFunctionBegin; 648 PetscCall(PetscViewerSetUp(viewer)); 649 650 M = mat->rmap->N; 651 N = mat->cmap->N; 652 m = mat->rmap->n; 653 nz = A->nz; 654 655 /* write matrix header */ 656 header[0] = MAT_FILE_CLASSID; 657 header[1] = M; 658 header[2] = N; 659 header[3] = nz; 660 PetscCall(PetscViewerBinaryWrite(viewer, header, 4, PETSC_INT)); 661 662 /* fill in and store row lengths */ 663 PetscCall(PetscMalloc1(m, &rowlens)); 664 for (i = 0; i < m; i++) rowlens[i] = A->i[i + 1] - A->i[i]; 665 PetscCall(PetscViewerBinaryWrite(viewer, rowlens, m, PETSC_INT)); 666 PetscCall(PetscFree(rowlens)); 667 /* store column indices */ 668 PetscCall(PetscViewerBinaryWrite(viewer, A->j, nz, PETSC_INT)); 669 /* store nonzero values */ 670 PetscCall(MatSeqAIJGetArrayRead(mat, &av)); 671 PetscCall(PetscViewerBinaryWrite(viewer, av, nz, PETSC_SCALAR)); 672 PetscCall(MatSeqAIJRestoreArrayRead(mat, &av)); 673 674 /* write block size option to the viewer's .info file */ 675 PetscCall(MatView_Binary_BlockSizes(mat, viewer)); 676 PetscFunctionReturn(0); 677 } 678 679 static PetscErrorCode MatView_SeqAIJ_ASCII_structonly(Mat A, PetscViewer viewer) 680 { 681 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 682 PetscInt i, k, m = A->rmap->N; 683 684 PetscFunctionBegin; 685 PetscCall(PetscViewerASCIIUseTabs(viewer, PETSC_FALSE)); 686 for (i = 0; i < m; i++) { 687 PetscCall(PetscViewerASCIIPrintf(viewer, "row %" PetscInt_FMT ":", i)); 688 for (k = a->i[i]; k < a->i[i + 1]; k++) PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ") ", a->j[k])); 689 PetscCall(PetscViewerASCIIPrintf(viewer, "\n")); 690 } 691 PetscCall(PetscViewerASCIIUseTabs(viewer, PETSC_TRUE)); 692 PetscFunctionReturn(0); 693 } 694 695 extern PetscErrorCode MatSeqAIJFactorInfo_Matlab(Mat, PetscViewer); 696 697 PetscErrorCode MatView_SeqAIJ_ASCII(Mat A, PetscViewer viewer) 698 { 699 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 700 const PetscScalar *av; 701 PetscInt i, j, m = A->rmap->n; 702 const char *name; 703 PetscViewerFormat format; 704 705 PetscFunctionBegin; 706 if (A->structure_only) { 707 PetscCall(MatView_SeqAIJ_ASCII_structonly(A, viewer)); 708 PetscFunctionReturn(0); 709 } 710 711 PetscCall(PetscViewerGetFormat(viewer, &format)); 712 if (format == PETSC_VIEWER_ASCII_FACTOR_INFO || format == PETSC_VIEWER_ASCII_INFO || format == PETSC_VIEWER_ASCII_INFO_DETAIL) PetscFunctionReturn(0); 713 714 /* trigger copy to CPU if needed */ 715 PetscCall(MatSeqAIJGetArrayRead(A, &av)); 716 PetscCall(MatSeqAIJRestoreArrayRead(A, &av)); 717 if (format == PETSC_VIEWER_ASCII_MATLAB) { 718 PetscInt nofinalvalue = 0; 719 if (m && ((a->i[m] == a->i[m - 1]) || (a->j[a->nz - 1] != A->cmap->n - 1))) { 720 /* Need a dummy value to ensure the dimension of the matrix. */ 721 nofinalvalue = 1; 722 } 723 PetscCall(PetscViewerASCIIUseTabs(viewer, PETSC_FALSE)); 724 PetscCall(PetscViewerASCIIPrintf(viewer, "%% Size = %" PetscInt_FMT " %" PetscInt_FMT " \n", m, A->cmap->n)); 725 PetscCall(PetscViewerASCIIPrintf(viewer, "%% Nonzeros = %" PetscInt_FMT " \n", a->nz)); 726 #if defined(PETSC_USE_COMPLEX) 727 PetscCall(PetscViewerASCIIPrintf(viewer, "zzz = zeros(%" PetscInt_FMT ",4);\n", a->nz + nofinalvalue)); 728 #else 729 PetscCall(PetscViewerASCIIPrintf(viewer, "zzz = zeros(%" PetscInt_FMT ",3);\n", a->nz + nofinalvalue)); 730 #endif 731 PetscCall(PetscViewerASCIIPrintf(viewer, "zzz = [\n")); 732 733 for (i = 0; i < m; i++) { 734 for (j = a->i[i]; j < a->i[i + 1]; j++) { 735 #if defined(PETSC_USE_COMPLEX) 736 PetscCall(PetscViewerASCIIPrintf(viewer, "%" PetscInt_FMT " %" PetscInt_FMT " %18.16e %18.16e\n", i + 1, a->j[j] + 1, (double)PetscRealPart(a->a[j]), (double)PetscImaginaryPart(a->a[j]))); 737 #else 738 PetscCall(PetscViewerASCIIPrintf(viewer, "%" PetscInt_FMT " %" PetscInt_FMT " %18.16e\n", i + 1, a->j[j] + 1, (double)a->a[j])); 739 #endif 740 } 741 } 742 if (nofinalvalue) { 743 #if defined(PETSC_USE_COMPLEX) 744 PetscCall(PetscViewerASCIIPrintf(viewer, "%" PetscInt_FMT " %" PetscInt_FMT " %18.16e %18.16e\n", m, A->cmap->n, 0., 0.)); 745 #else 746 PetscCall(PetscViewerASCIIPrintf(viewer, "%" PetscInt_FMT " %" PetscInt_FMT " %18.16e\n", m, A->cmap->n, 0.0)); 747 #endif 748 } 749 PetscCall(PetscObjectGetName((PetscObject)A, &name)); 750 PetscCall(PetscViewerASCIIPrintf(viewer, "];\n %s = spconvert(zzz);\n", name)); 751 PetscCall(PetscViewerASCIIUseTabs(viewer, PETSC_TRUE)); 752 } else if (format == PETSC_VIEWER_ASCII_COMMON) { 753 PetscCall(PetscViewerASCIIUseTabs(viewer, PETSC_FALSE)); 754 for (i = 0; i < m; i++) { 755 PetscCall(PetscViewerASCIIPrintf(viewer, "row %" PetscInt_FMT ":", i)); 756 for (j = a->i[i]; j < a->i[i + 1]; j++) { 757 #if defined(PETSC_USE_COMPLEX) 758 if (PetscImaginaryPart(a->a[j]) > 0.0 && PetscRealPart(a->a[j]) != 0.0) { 759 PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g + %g i)", a->j[j], (double)PetscRealPart(a->a[j]), (double)PetscImaginaryPart(a->a[j]))); 760 } else if (PetscImaginaryPart(a->a[j]) < 0.0 && PetscRealPart(a->a[j]) != 0.0) { 761 PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g - %g i)", a->j[j], (double)PetscRealPart(a->a[j]), (double)-PetscImaginaryPart(a->a[j]))); 762 } else if (PetscRealPart(a->a[j]) != 0.0) { 763 PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g) ", a->j[j], (double)PetscRealPart(a->a[j]))); 764 } 765 #else 766 if (a->a[j] != 0.0) PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g) ", a->j[j], (double)a->a[j])); 767 #endif 768 } 769 PetscCall(PetscViewerASCIIPrintf(viewer, "\n")); 770 } 771 PetscCall(PetscViewerASCIIUseTabs(viewer, PETSC_TRUE)); 772 } else if (format == PETSC_VIEWER_ASCII_SYMMODU) { 773 PetscInt nzd = 0, fshift = 1, *sptr; 774 PetscCall(PetscViewerASCIIUseTabs(viewer, PETSC_FALSE)); 775 PetscCall(PetscMalloc1(m + 1, &sptr)); 776 for (i = 0; i < m; i++) { 777 sptr[i] = nzd + 1; 778 for (j = a->i[i]; j < a->i[i + 1]; j++) { 779 if (a->j[j] >= i) { 780 #if defined(PETSC_USE_COMPLEX) 781 if (PetscImaginaryPart(a->a[j]) != 0.0 || PetscRealPart(a->a[j]) != 0.0) nzd++; 782 #else 783 if (a->a[j] != 0.0) nzd++; 784 #endif 785 } 786 } 787 } 788 sptr[m] = nzd + 1; 789 PetscCall(PetscViewerASCIIPrintf(viewer, " %" PetscInt_FMT " %" PetscInt_FMT "\n\n", m, nzd)); 790 for (i = 0; i < m + 1; i += 6) { 791 if (i + 4 < m) { 792 PetscCall(PetscViewerASCIIPrintf(viewer, " %" PetscInt_FMT " %" PetscInt_FMT " %" PetscInt_FMT " %" PetscInt_FMT " %" PetscInt_FMT " %" PetscInt_FMT "\n", sptr[i], sptr[i + 1], sptr[i + 2], sptr[i + 3], sptr[i + 4], sptr[i + 5])); 793 } else if (i + 3 < m) { 794 PetscCall(PetscViewerASCIIPrintf(viewer, " %" PetscInt_FMT " %" PetscInt_FMT " %" PetscInt_FMT " %" PetscInt_FMT " %" PetscInt_FMT "\n", sptr[i], sptr[i + 1], sptr[i + 2], sptr[i + 3], sptr[i + 4])); 795 } else if (i + 2 < m) { 796 PetscCall(PetscViewerASCIIPrintf(viewer, " %" PetscInt_FMT " %" PetscInt_FMT " %" PetscInt_FMT " %" PetscInt_FMT "\n", sptr[i], sptr[i + 1], sptr[i + 2], sptr[i + 3])); 797 } else if (i + 1 < m) { 798 PetscCall(PetscViewerASCIIPrintf(viewer, " %" PetscInt_FMT " %" PetscInt_FMT " %" PetscInt_FMT "\n", sptr[i], sptr[i + 1], sptr[i + 2])); 799 } else if (i < m) { 800 PetscCall(PetscViewerASCIIPrintf(viewer, " %" PetscInt_FMT " %" PetscInt_FMT "\n", sptr[i], sptr[i + 1])); 801 } else { 802 PetscCall(PetscViewerASCIIPrintf(viewer, " %" PetscInt_FMT "\n", sptr[i])); 803 } 804 } 805 PetscCall(PetscViewerASCIIPrintf(viewer, "\n")); 806 PetscCall(PetscFree(sptr)); 807 for (i = 0; i < m; i++) { 808 for (j = a->i[i]; j < a->i[i + 1]; j++) { 809 if (a->j[j] >= i) PetscCall(PetscViewerASCIIPrintf(viewer, " %" PetscInt_FMT " ", a->j[j] + fshift)); 810 } 811 PetscCall(PetscViewerASCIIPrintf(viewer, "\n")); 812 } 813 PetscCall(PetscViewerASCIIPrintf(viewer, "\n")); 814 for (i = 0; i < m; i++) { 815 for (j = a->i[i]; j < a->i[i + 1]; j++) { 816 if (a->j[j] >= i) { 817 #if defined(PETSC_USE_COMPLEX) 818 if (PetscImaginaryPart(a->a[j]) != 0.0 || PetscRealPart(a->a[j]) != 0.0) PetscCall(PetscViewerASCIIPrintf(viewer, " %18.16e %18.16e ", (double)PetscRealPart(a->a[j]), (double)PetscImaginaryPart(a->a[j]))); 819 #else 820 if (a->a[j] != 0.0) PetscCall(PetscViewerASCIIPrintf(viewer, " %18.16e ", (double)a->a[j])); 821 #endif 822 } 823 } 824 PetscCall(PetscViewerASCIIPrintf(viewer, "\n")); 825 } 826 PetscCall(PetscViewerASCIIUseTabs(viewer, PETSC_TRUE)); 827 } else if (format == PETSC_VIEWER_ASCII_DENSE) { 828 PetscInt cnt = 0, jcnt; 829 PetscScalar value; 830 #if defined(PETSC_USE_COMPLEX) 831 PetscBool realonly = PETSC_TRUE; 832 833 for (i = 0; i < a->i[m]; i++) { 834 if (PetscImaginaryPart(a->a[i]) != 0.0) { 835 realonly = PETSC_FALSE; 836 break; 837 } 838 } 839 #endif 840 841 PetscCall(PetscViewerASCIIUseTabs(viewer, PETSC_FALSE)); 842 for (i = 0; i < m; i++) { 843 jcnt = 0; 844 for (j = 0; j < A->cmap->n; j++) { 845 if (jcnt < a->i[i + 1] - a->i[i] && j == a->j[cnt]) { 846 value = a->a[cnt++]; 847 jcnt++; 848 } else { 849 value = 0.0; 850 } 851 #if defined(PETSC_USE_COMPLEX) 852 if (realonly) { 853 PetscCall(PetscViewerASCIIPrintf(viewer, " %7.5e ", (double)PetscRealPart(value))); 854 } else { 855 PetscCall(PetscViewerASCIIPrintf(viewer, " %7.5e+%7.5e i ", (double)PetscRealPart(value), (double)PetscImaginaryPart(value))); 856 } 857 #else 858 PetscCall(PetscViewerASCIIPrintf(viewer, " %7.5e ", (double)value)); 859 #endif 860 } 861 PetscCall(PetscViewerASCIIPrintf(viewer, "\n")); 862 } 863 PetscCall(PetscViewerASCIIUseTabs(viewer, PETSC_TRUE)); 864 } else if (format == PETSC_VIEWER_ASCII_MATRIXMARKET) { 865 PetscInt fshift = 1; 866 PetscCall(PetscViewerASCIIUseTabs(viewer, PETSC_FALSE)); 867 #if defined(PETSC_USE_COMPLEX) 868 PetscCall(PetscViewerASCIIPrintf(viewer, "%%%%MatrixMarket matrix coordinate complex general\n")); 869 #else 870 PetscCall(PetscViewerASCIIPrintf(viewer, "%%%%MatrixMarket matrix coordinate real general\n")); 871 #endif 872 PetscCall(PetscViewerASCIIPrintf(viewer, "%" PetscInt_FMT " %" PetscInt_FMT " %" PetscInt_FMT "\n", m, A->cmap->n, a->nz)); 873 for (i = 0; i < m; i++) { 874 for (j = a->i[i]; j < a->i[i + 1]; j++) { 875 #if defined(PETSC_USE_COMPLEX) 876 PetscCall(PetscViewerASCIIPrintf(viewer, "%" PetscInt_FMT " %" PetscInt_FMT " %g %g\n", i + fshift, a->j[j] + fshift, (double)PetscRealPart(a->a[j]), (double)PetscImaginaryPart(a->a[j]))); 877 #else 878 PetscCall(PetscViewerASCIIPrintf(viewer, "%" PetscInt_FMT " %" PetscInt_FMT " %g\n", i + fshift, a->j[j] + fshift, (double)a->a[j])); 879 #endif 880 } 881 } 882 PetscCall(PetscViewerASCIIUseTabs(viewer, PETSC_TRUE)); 883 } else { 884 PetscCall(PetscViewerASCIIUseTabs(viewer, PETSC_FALSE)); 885 if (A->factortype) { 886 for (i = 0; i < m; i++) { 887 PetscCall(PetscViewerASCIIPrintf(viewer, "row %" PetscInt_FMT ":", i)); 888 /* L part */ 889 for (j = a->i[i]; j < a->i[i + 1]; j++) { 890 #if defined(PETSC_USE_COMPLEX) 891 if (PetscImaginaryPart(a->a[j]) > 0.0) { 892 PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g + %g i)", a->j[j], (double)PetscRealPart(a->a[j]), (double)PetscImaginaryPart(a->a[j]))); 893 } else if (PetscImaginaryPart(a->a[j]) < 0.0) { 894 PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g - %g i)", a->j[j], (double)PetscRealPart(a->a[j]), (double)(-PetscImaginaryPart(a->a[j])))); 895 } else { 896 PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g) ", a->j[j], (double)PetscRealPart(a->a[j]))); 897 } 898 #else 899 PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g) ", a->j[j], (double)a->a[j])); 900 #endif 901 } 902 /* diagonal */ 903 j = a->diag[i]; 904 #if defined(PETSC_USE_COMPLEX) 905 if (PetscImaginaryPart(a->a[j]) > 0.0) { 906 PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g + %g i)", a->j[j], (double)PetscRealPart(1.0 / a->a[j]), (double)PetscImaginaryPart(1.0 / a->a[j]))); 907 } else if (PetscImaginaryPart(a->a[j]) < 0.0) { 908 PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g - %g i)", a->j[j], (double)PetscRealPart(1.0 / a->a[j]), (double)(-PetscImaginaryPart(1.0 / a->a[j])))); 909 } else { 910 PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g) ", a->j[j], (double)PetscRealPart(1.0 / a->a[j]))); 911 } 912 #else 913 PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g) ", a->j[j], (double)(1.0 / a->a[j]))); 914 #endif 915 916 /* U part */ 917 for (j = a->diag[i + 1] + 1; j < a->diag[i]; j++) { 918 #if defined(PETSC_USE_COMPLEX) 919 if (PetscImaginaryPart(a->a[j]) > 0.0) { 920 PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g + %g i)", a->j[j], (double)PetscRealPart(a->a[j]), (double)PetscImaginaryPart(a->a[j]))); 921 } else if (PetscImaginaryPart(a->a[j]) < 0.0) { 922 PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g - %g i)", a->j[j], (double)PetscRealPart(a->a[j]), (double)(-PetscImaginaryPart(a->a[j])))); 923 } else { 924 PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g) ", a->j[j], (double)PetscRealPart(a->a[j]))); 925 } 926 #else 927 PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g) ", a->j[j], (double)a->a[j])); 928 #endif 929 } 930 PetscCall(PetscViewerASCIIPrintf(viewer, "\n")); 931 } 932 } else { 933 for (i = 0; i < m; i++) { 934 PetscCall(PetscViewerASCIIPrintf(viewer, "row %" PetscInt_FMT ":", i)); 935 for (j = a->i[i]; j < a->i[i + 1]; j++) { 936 #if defined(PETSC_USE_COMPLEX) 937 if (PetscImaginaryPart(a->a[j]) > 0.0) { 938 PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g + %g i)", a->j[j], (double)PetscRealPart(a->a[j]), (double)PetscImaginaryPart(a->a[j]))); 939 } else if (PetscImaginaryPart(a->a[j]) < 0.0) { 940 PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g - %g i)", a->j[j], (double)PetscRealPart(a->a[j]), (double)-PetscImaginaryPart(a->a[j]))); 941 } else { 942 PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g) ", a->j[j], (double)PetscRealPart(a->a[j]))); 943 } 944 #else 945 PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g) ", a->j[j], (double)a->a[j])); 946 #endif 947 } 948 PetscCall(PetscViewerASCIIPrintf(viewer, "\n")); 949 } 950 } 951 PetscCall(PetscViewerASCIIUseTabs(viewer, PETSC_TRUE)); 952 } 953 PetscCall(PetscViewerFlush(viewer)); 954 PetscFunctionReturn(0); 955 } 956 957 #include <petscdraw.h> 958 PetscErrorCode MatView_SeqAIJ_Draw_Zoom(PetscDraw draw, void *Aa) 959 { 960 Mat A = (Mat)Aa; 961 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 962 PetscInt i, j, m = A->rmap->n; 963 int color; 964 PetscReal xl, yl, xr, yr, x_l, x_r, y_l, y_r; 965 PetscViewer viewer; 966 PetscViewerFormat format; 967 const PetscScalar *aa; 968 969 PetscFunctionBegin; 970 PetscCall(PetscObjectQuery((PetscObject)A, "Zoomviewer", (PetscObject *)&viewer)); 971 PetscCall(PetscViewerGetFormat(viewer, &format)); 972 PetscCall(PetscDrawGetCoordinates(draw, &xl, &yl, &xr, &yr)); 973 974 /* loop over matrix elements drawing boxes */ 975 PetscCall(MatSeqAIJGetArrayRead(A, &aa)); 976 if (format != PETSC_VIEWER_DRAW_CONTOUR) { 977 PetscDrawCollectiveBegin(draw); 978 /* Blue for negative, Cyan for zero and Red for positive */ 979 color = PETSC_DRAW_BLUE; 980 for (i = 0; i < m; i++) { 981 y_l = m - i - 1.0; 982 y_r = y_l + 1.0; 983 for (j = a->i[i]; j < a->i[i + 1]; j++) { 984 x_l = a->j[j]; 985 x_r = x_l + 1.0; 986 if (PetscRealPart(aa[j]) >= 0.) continue; 987 PetscCall(PetscDrawRectangle(draw, x_l, y_l, x_r, y_r, color, color, color, color)); 988 } 989 } 990 color = PETSC_DRAW_CYAN; 991 for (i = 0; i < m; i++) { 992 y_l = m - i - 1.0; 993 y_r = y_l + 1.0; 994 for (j = a->i[i]; j < a->i[i + 1]; j++) { 995 x_l = a->j[j]; 996 x_r = x_l + 1.0; 997 if (aa[j] != 0.) continue; 998 PetscCall(PetscDrawRectangle(draw, x_l, y_l, x_r, y_r, color, color, color, color)); 999 } 1000 } 1001 color = PETSC_DRAW_RED; 1002 for (i = 0; i < m; i++) { 1003 y_l = m - i - 1.0; 1004 y_r = y_l + 1.0; 1005 for (j = a->i[i]; j < a->i[i + 1]; j++) { 1006 x_l = a->j[j]; 1007 x_r = x_l + 1.0; 1008 if (PetscRealPart(aa[j]) <= 0.) continue; 1009 PetscCall(PetscDrawRectangle(draw, x_l, y_l, x_r, y_r, color, color, color, color)); 1010 } 1011 } 1012 PetscDrawCollectiveEnd(draw); 1013 } else { 1014 /* use contour shading to indicate magnitude of values */ 1015 /* first determine max of all nonzero values */ 1016 PetscReal minv = 0.0, maxv = 0.0; 1017 PetscInt nz = a->nz, count = 0; 1018 PetscDraw popup; 1019 1020 for (i = 0; i < nz; i++) { 1021 if (PetscAbsScalar(aa[i]) > maxv) maxv = PetscAbsScalar(aa[i]); 1022 } 1023 if (minv >= maxv) maxv = minv + PETSC_SMALL; 1024 PetscCall(PetscDrawGetPopup(draw, &popup)); 1025 PetscCall(PetscDrawScalePopup(popup, minv, maxv)); 1026 1027 PetscDrawCollectiveBegin(draw); 1028 for (i = 0; i < m; i++) { 1029 y_l = m - i - 1.0; 1030 y_r = y_l + 1.0; 1031 for (j = a->i[i]; j < a->i[i + 1]; j++) { 1032 x_l = a->j[j]; 1033 x_r = x_l + 1.0; 1034 color = PetscDrawRealToColor(PetscAbsScalar(aa[count]), minv, maxv); 1035 PetscCall(PetscDrawRectangle(draw, x_l, y_l, x_r, y_r, color, color, color, color)); 1036 count++; 1037 } 1038 } 1039 PetscDrawCollectiveEnd(draw); 1040 } 1041 PetscCall(MatSeqAIJRestoreArrayRead(A, &aa)); 1042 PetscFunctionReturn(0); 1043 } 1044 1045 #include <petscdraw.h> 1046 PetscErrorCode MatView_SeqAIJ_Draw(Mat A, PetscViewer viewer) 1047 { 1048 PetscDraw draw; 1049 PetscReal xr, yr, xl, yl, h, w; 1050 PetscBool isnull; 1051 1052 PetscFunctionBegin; 1053 PetscCall(PetscViewerDrawGetDraw(viewer, 0, &draw)); 1054 PetscCall(PetscDrawIsNull(draw, &isnull)); 1055 if (isnull) PetscFunctionReturn(0); 1056 1057 xr = A->cmap->n; 1058 yr = A->rmap->n; 1059 h = yr / 10.0; 1060 w = xr / 10.0; 1061 xr += w; 1062 yr += h; 1063 xl = -w; 1064 yl = -h; 1065 PetscCall(PetscDrawSetCoordinates(draw, xl, yl, xr, yr)); 1066 PetscCall(PetscObjectCompose((PetscObject)A, "Zoomviewer", (PetscObject)viewer)); 1067 PetscCall(PetscDrawZoom(draw, MatView_SeqAIJ_Draw_Zoom, A)); 1068 PetscCall(PetscObjectCompose((PetscObject)A, "Zoomviewer", NULL)); 1069 PetscCall(PetscDrawSave(draw)); 1070 PetscFunctionReturn(0); 1071 } 1072 1073 PetscErrorCode MatView_SeqAIJ(Mat A, PetscViewer viewer) 1074 { 1075 PetscBool iascii, isbinary, isdraw; 1076 1077 PetscFunctionBegin; 1078 PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERASCII, &iascii)); 1079 PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERBINARY, &isbinary)); 1080 PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERDRAW, &isdraw)); 1081 if (iascii) PetscCall(MatView_SeqAIJ_ASCII(A, viewer)); 1082 else if (isbinary) PetscCall(MatView_SeqAIJ_Binary(A, viewer)); 1083 else if (isdraw) PetscCall(MatView_SeqAIJ_Draw(A, viewer)); 1084 PetscCall(MatView_SeqAIJ_Inode(A, viewer)); 1085 PetscFunctionReturn(0); 1086 } 1087 1088 PetscErrorCode MatAssemblyEnd_SeqAIJ(Mat A, MatAssemblyType mode) 1089 { 1090 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 1091 PetscInt fshift = 0, i, *ai = a->i, *aj = a->j, *imax = a->imax; 1092 PetscInt m = A->rmap->n, *ip, N, *ailen = a->ilen, rmax = 0; 1093 MatScalar *aa = a->a, *ap; 1094 PetscReal ratio = 0.6; 1095 1096 PetscFunctionBegin; 1097 if (mode == MAT_FLUSH_ASSEMBLY) PetscFunctionReturn(0); 1098 PetscCall(MatSeqAIJInvalidateDiagonal(A)); 1099 if (A->was_assembled && A->ass_nonzerostate == A->nonzerostate) { 1100 /* we need to respect users asking to use or not the inodes routine in between matrix assemblies */ 1101 PetscCall(MatAssemblyEnd_SeqAIJ_Inode(A, mode)); 1102 PetscFunctionReturn(0); 1103 } 1104 1105 if (m) rmax = ailen[0]; /* determine row with most nonzeros */ 1106 for (i = 1; i < m; i++) { 1107 /* move each row back by the amount of empty slots (fshift) before it*/ 1108 fshift += imax[i - 1] - ailen[i - 1]; 1109 rmax = PetscMax(rmax, ailen[i]); 1110 if (fshift) { 1111 ip = aj + ai[i]; 1112 ap = aa + ai[i]; 1113 N = ailen[i]; 1114 PetscCall(PetscArraymove(ip - fshift, ip, N)); 1115 if (!A->structure_only) PetscCall(PetscArraymove(ap - fshift, ap, N)); 1116 } 1117 ai[i] = ai[i - 1] + ailen[i - 1]; 1118 } 1119 if (m) { 1120 fshift += imax[m - 1] - ailen[m - 1]; 1121 ai[m] = ai[m - 1] + ailen[m - 1]; 1122 } 1123 /* reset ilen and imax for each row */ 1124 a->nonzerorowcnt = 0; 1125 if (A->structure_only) { 1126 PetscCall(PetscFree(a->imax)); 1127 PetscCall(PetscFree(a->ilen)); 1128 } else { /* !A->structure_only */ 1129 for (i = 0; i < m; i++) { 1130 ailen[i] = imax[i] = ai[i + 1] - ai[i]; 1131 a->nonzerorowcnt += ((ai[i + 1] - ai[i]) > 0); 1132 } 1133 } 1134 a->nz = ai[m]; 1135 PetscCheck(!fshift || a->nounused != -1, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Unused space detected in matrix: %" PetscInt_FMT " X %" PetscInt_FMT ", %" PetscInt_FMT " unneeded", m, A->cmap->n, fshift); 1136 1137 PetscCall(MatMarkDiagonal_SeqAIJ(A)); 1138 PetscCall(PetscInfo(A, "Matrix size: %" PetscInt_FMT " X %" PetscInt_FMT "; storage space: %" PetscInt_FMT " unneeded,%" PetscInt_FMT " used\n", m, A->cmap->n, fshift, a->nz)); 1139 PetscCall(PetscInfo(A, "Number of mallocs during MatSetValues() is %" PetscInt_FMT "\n", a->reallocs)); 1140 PetscCall(PetscInfo(A, "Maximum nonzeros in any row is %" PetscInt_FMT "\n", rmax)); 1141 1142 A->info.mallocs += a->reallocs; 1143 a->reallocs = 0; 1144 A->info.nz_unneeded = (PetscReal)fshift; 1145 a->rmax = rmax; 1146 1147 if (!A->structure_only) PetscCall(MatCheckCompressedRow(A, a->nonzerorowcnt, &a->compressedrow, a->i, m, ratio)); 1148 PetscCall(MatAssemblyEnd_SeqAIJ_Inode(A, mode)); 1149 PetscFunctionReturn(0); 1150 } 1151 1152 PetscErrorCode MatRealPart_SeqAIJ(Mat A) 1153 { 1154 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 1155 PetscInt i, nz = a->nz; 1156 MatScalar *aa; 1157 1158 PetscFunctionBegin; 1159 PetscCall(MatSeqAIJGetArray(A, &aa)); 1160 for (i = 0; i < nz; i++) aa[i] = PetscRealPart(aa[i]); 1161 PetscCall(MatSeqAIJRestoreArray(A, &aa)); 1162 PetscCall(MatSeqAIJInvalidateDiagonal(A)); 1163 PetscFunctionReturn(0); 1164 } 1165 1166 PetscErrorCode MatImaginaryPart_SeqAIJ(Mat A) 1167 { 1168 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 1169 PetscInt i, nz = a->nz; 1170 MatScalar *aa; 1171 1172 PetscFunctionBegin; 1173 PetscCall(MatSeqAIJGetArray(A, &aa)); 1174 for (i = 0; i < nz; i++) aa[i] = PetscImaginaryPart(aa[i]); 1175 PetscCall(MatSeqAIJRestoreArray(A, &aa)); 1176 PetscCall(MatSeqAIJInvalidateDiagonal(A)); 1177 PetscFunctionReturn(0); 1178 } 1179 1180 PetscErrorCode MatZeroEntries_SeqAIJ(Mat A) 1181 { 1182 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 1183 MatScalar *aa; 1184 1185 PetscFunctionBegin; 1186 PetscCall(MatSeqAIJGetArrayWrite(A, &aa)); 1187 PetscCall(PetscArrayzero(aa, a->i[A->rmap->n])); 1188 PetscCall(MatSeqAIJRestoreArrayWrite(A, &aa)); 1189 PetscCall(MatSeqAIJInvalidateDiagonal(A)); 1190 PetscFunctionReturn(0); 1191 } 1192 1193 PETSC_INTERN PetscErrorCode MatResetPreallocationCOO_SeqAIJ(Mat A) 1194 { 1195 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 1196 1197 PetscFunctionBegin; 1198 PetscCall(PetscFree(a->perm)); 1199 PetscCall(PetscFree(a->jmap)); 1200 PetscFunctionReturn(0); 1201 } 1202 1203 PetscErrorCode MatDestroy_SeqAIJ(Mat A) 1204 { 1205 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 1206 1207 PetscFunctionBegin; 1208 #if defined(PETSC_USE_LOG) 1209 PetscLogObjectState((PetscObject)A, "Rows=%" PetscInt_FMT ", Cols=%" PetscInt_FMT ", NZ=%" PetscInt_FMT, A->rmap->n, A->cmap->n, a->nz); 1210 #endif 1211 PetscCall(MatSeqXAIJFreeAIJ(A, &a->a, &a->j, &a->i)); 1212 PetscCall(MatResetPreallocationCOO_SeqAIJ(A)); 1213 PetscCall(ISDestroy(&a->row)); 1214 PetscCall(ISDestroy(&a->col)); 1215 PetscCall(PetscFree(a->diag)); 1216 PetscCall(PetscFree(a->ibdiag)); 1217 PetscCall(PetscFree(a->imax)); 1218 PetscCall(PetscFree(a->ilen)); 1219 PetscCall(PetscFree(a->ipre)); 1220 PetscCall(PetscFree3(a->idiag, a->mdiag, a->ssor_work)); 1221 PetscCall(PetscFree(a->solve_work)); 1222 PetscCall(ISDestroy(&a->icol)); 1223 PetscCall(PetscFree(a->saved_values)); 1224 PetscCall(PetscFree2(a->compressedrow.i, a->compressedrow.rindex)); 1225 PetscCall(MatDestroy_SeqAIJ_Inode(A)); 1226 PetscCall(PetscFree(A->data)); 1227 1228 /* MatMatMultNumeric_SeqAIJ_SeqAIJ_Sorted may allocate this. 1229 That function is so heavily used (sometimes in an hidden way through multnumeric function pointers) 1230 that is hard to properly add this data to the MatProduct data. We free it here to avoid 1231 users reusing the matrix object with different data to incur in obscure segmentation faults 1232 due to different matrix sizes */ 1233 PetscCall(PetscObjectCompose((PetscObject)A, "__PETSc__ab_dense", NULL)); 1234 1235 PetscCall(PetscObjectChangeTypeName((PetscObject)A, NULL)); 1236 PetscCall(PetscObjectComposeFunction((PetscObject)A, "PetscMatlabEnginePut_C", NULL)); 1237 PetscCall(PetscObjectComposeFunction((PetscObject)A, "PetscMatlabEngineGet_C", NULL)); 1238 PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatSeqAIJSetColumnIndices_C", NULL)); 1239 PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatStoreValues_C", NULL)); 1240 PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatRetrieveValues_C", NULL)); 1241 PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_seqsbaij_C", NULL)); 1242 PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_seqbaij_C", NULL)); 1243 PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_seqaijperm_C", NULL)); 1244 PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_seqaijsell_C", NULL)); 1245 #if defined(PETSC_HAVE_MKL_SPARSE) 1246 PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_seqaijmkl_C", NULL)); 1247 #endif 1248 #if defined(PETSC_HAVE_CUDA) 1249 PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_seqaijcusparse_C", NULL)); 1250 PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_seqaijcusparse_seqaij_C", NULL)); 1251 PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_seqaij_seqaijcusparse_C", NULL)); 1252 #endif 1253 #if defined(PETSC_HAVE_HIP) 1254 PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_seqaijhipsparse_C", NULL)); 1255 PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_seqaijhipsparse_seqaij_C", NULL)); 1256 PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_seqaij_seqaijhipsparse_C", NULL)); 1257 #endif 1258 #if defined(PETSC_HAVE_KOKKOS_KERNELS) 1259 PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_seqaijkokkos_C", NULL)); 1260 #endif 1261 PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_seqaijcrl_C", NULL)); 1262 #if defined(PETSC_HAVE_ELEMENTAL) 1263 PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_elemental_C", NULL)); 1264 #endif 1265 #if defined(PETSC_HAVE_SCALAPACK) 1266 PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_scalapack_C", NULL)); 1267 #endif 1268 #if defined(PETSC_HAVE_HYPRE) 1269 PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_hypre_C", NULL)); 1270 PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_transpose_seqaij_seqaij_C", NULL)); 1271 #endif 1272 PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_seqdense_C", NULL)); 1273 PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_seqsell_C", NULL)); 1274 PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_is_C", NULL)); 1275 PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatIsTranspose_C", NULL)); 1276 PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatIsHermitianTranspose_C", NULL)); 1277 PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatSeqAIJSetPreallocation_C", NULL)); 1278 PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatResetPreallocation_C", NULL)); 1279 PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatSeqAIJSetPreallocationCSR_C", NULL)); 1280 PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatReorderForNonzeroDiagonal_C", NULL)); 1281 PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_is_seqaij_C", NULL)); 1282 PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_seqdense_seqaij_C", NULL)); 1283 PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_seqaij_seqaij_C", NULL)); 1284 PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatSeqAIJKron_C", NULL)); 1285 PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatSetPreallocationCOO_C", NULL)); 1286 PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatSetValuesCOO_C", NULL)); 1287 PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatFactorGetSolverType_C", NULL)); 1288 /* these calls do not belong here: the subclasses Duplicate/Destroy are wrong */ 1289 PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaijsell_seqaij_C", NULL)); 1290 PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaijperm_seqaij_C", NULL)); 1291 PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_seqaijviennacl_C", NULL)); 1292 PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_seqaijviennacl_seqdense_C", NULL)); 1293 PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_seqaijviennacl_seqaij_C", NULL)); 1294 PetscFunctionReturn(0); 1295 } 1296 1297 PetscErrorCode MatSetOption_SeqAIJ(Mat A, MatOption op, PetscBool flg) 1298 { 1299 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 1300 1301 PetscFunctionBegin; 1302 switch (op) { 1303 case MAT_ROW_ORIENTED: 1304 a->roworiented = flg; 1305 break; 1306 case MAT_KEEP_NONZERO_PATTERN: 1307 a->keepnonzeropattern = flg; 1308 break; 1309 case MAT_NEW_NONZERO_LOCATIONS: 1310 a->nonew = (flg ? 0 : 1); 1311 break; 1312 case MAT_NEW_NONZERO_LOCATION_ERR: 1313 a->nonew = (flg ? -1 : 0); 1314 break; 1315 case MAT_NEW_NONZERO_ALLOCATION_ERR: 1316 a->nonew = (flg ? -2 : 0); 1317 break; 1318 case MAT_UNUSED_NONZERO_LOCATION_ERR: 1319 a->nounused = (flg ? -1 : 0); 1320 break; 1321 case MAT_IGNORE_ZERO_ENTRIES: 1322 a->ignorezeroentries = flg; 1323 break; 1324 case MAT_SPD: 1325 case MAT_SYMMETRIC: 1326 case MAT_STRUCTURALLY_SYMMETRIC: 1327 case MAT_HERMITIAN: 1328 case MAT_SYMMETRY_ETERNAL: 1329 case MAT_STRUCTURE_ONLY: 1330 case MAT_STRUCTURAL_SYMMETRY_ETERNAL: 1331 case MAT_SPD_ETERNAL: 1332 /* if the diagonal matrix is square it inherits some of the properties above */ 1333 break; 1334 case MAT_FORCE_DIAGONAL_ENTRIES: 1335 case MAT_IGNORE_OFF_PROC_ENTRIES: 1336 case MAT_USE_HASH_TABLE: 1337 PetscCall(PetscInfo(A, "Option %s ignored\n", MatOptions[op])); 1338 break; 1339 case MAT_USE_INODES: 1340 PetscCall(MatSetOption_SeqAIJ_Inode(A, MAT_USE_INODES, flg)); 1341 break; 1342 case MAT_SUBMAT_SINGLEIS: 1343 A->submat_singleis = flg; 1344 break; 1345 case MAT_SORTED_FULL: 1346 if (flg) A->ops->setvalues = MatSetValues_SeqAIJ_SortedFull; 1347 else A->ops->setvalues = MatSetValues_SeqAIJ; 1348 break; 1349 case MAT_FORM_EXPLICIT_TRANSPOSE: 1350 A->form_explicit_transpose = flg; 1351 break; 1352 default: 1353 SETERRQ(PETSC_COMM_SELF, PETSC_ERR_SUP, "unknown option %d", op); 1354 } 1355 PetscFunctionReturn(0); 1356 } 1357 1358 PetscErrorCode MatGetDiagonal_SeqAIJ(Mat A, Vec v) 1359 { 1360 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 1361 PetscInt i, j, n, *ai = a->i, *aj = a->j; 1362 PetscScalar *x; 1363 const PetscScalar *aa; 1364 1365 PetscFunctionBegin; 1366 PetscCall(VecGetLocalSize(v, &n)); 1367 PetscCheck(n == A->rmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Nonconforming matrix and vector"); 1368 PetscCall(MatSeqAIJGetArrayRead(A, &aa)); 1369 if (A->factortype == MAT_FACTOR_ILU || A->factortype == MAT_FACTOR_LU) { 1370 PetscInt *diag = a->diag; 1371 PetscCall(VecGetArrayWrite(v, &x)); 1372 for (i = 0; i < n; i++) x[i] = 1.0 / aa[diag[i]]; 1373 PetscCall(VecRestoreArrayWrite(v, &x)); 1374 PetscCall(MatSeqAIJRestoreArrayRead(A, &aa)); 1375 PetscFunctionReturn(0); 1376 } 1377 1378 PetscCall(VecGetArrayWrite(v, &x)); 1379 for (i = 0; i < n; i++) { 1380 x[i] = 0.0; 1381 for (j = ai[i]; j < ai[i + 1]; j++) { 1382 if (aj[j] == i) { 1383 x[i] = aa[j]; 1384 break; 1385 } 1386 } 1387 } 1388 PetscCall(VecRestoreArrayWrite(v, &x)); 1389 PetscCall(MatSeqAIJRestoreArrayRead(A, &aa)); 1390 PetscFunctionReturn(0); 1391 } 1392 1393 #include <../src/mat/impls/aij/seq/ftn-kernels/fmult.h> 1394 PetscErrorCode MatMultTransposeAdd_SeqAIJ(Mat A, Vec xx, Vec zz, Vec yy) 1395 { 1396 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 1397 const MatScalar *aa; 1398 PetscScalar *y; 1399 const PetscScalar *x; 1400 PetscInt m = A->rmap->n; 1401 #if !defined(PETSC_USE_FORTRAN_KERNEL_MULTTRANSPOSEAIJ) 1402 const MatScalar *v; 1403 PetscScalar alpha; 1404 PetscInt n, i, j; 1405 const PetscInt *idx, *ii, *ridx = NULL; 1406 Mat_CompressedRow cprow = a->compressedrow; 1407 PetscBool usecprow = cprow.use; 1408 #endif 1409 1410 PetscFunctionBegin; 1411 if (zz != yy) PetscCall(VecCopy(zz, yy)); 1412 PetscCall(VecGetArrayRead(xx, &x)); 1413 PetscCall(VecGetArray(yy, &y)); 1414 PetscCall(MatSeqAIJGetArrayRead(A, &aa)); 1415 1416 #if defined(PETSC_USE_FORTRAN_KERNEL_MULTTRANSPOSEAIJ) 1417 fortranmulttransposeaddaij_(&m, x, a->i, a->j, aa, y); 1418 #else 1419 if (usecprow) { 1420 m = cprow.nrows; 1421 ii = cprow.i; 1422 ridx = cprow.rindex; 1423 } else { 1424 ii = a->i; 1425 } 1426 for (i = 0; i < m; i++) { 1427 idx = a->j + ii[i]; 1428 v = aa + ii[i]; 1429 n = ii[i + 1] - ii[i]; 1430 if (usecprow) { 1431 alpha = x[ridx[i]]; 1432 } else { 1433 alpha = x[i]; 1434 } 1435 for (j = 0; j < n; j++) y[idx[j]] += alpha * v[j]; 1436 } 1437 #endif 1438 PetscCall(PetscLogFlops(2.0 * a->nz)); 1439 PetscCall(VecRestoreArrayRead(xx, &x)); 1440 PetscCall(VecRestoreArray(yy, &y)); 1441 PetscCall(MatSeqAIJRestoreArrayRead(A, &aa)); 1442 PetscFunctionReturn(0); 1443 } 1444 1445 PetscErrorCode MatMultTranspose_SeqAIJ(Mat A, Vec xx, Vec yy) 1446 { 1447 PetscFunctionBegin; 1448 PetscCall(VecSet(yy, 0.0)); 1449 PetscCall(MatMultTransposeAdd_SeqAIJ(A, xx, yy, yy)); 1450 PetscFunctionReturn(0); 1451 } 1452 1453 #include <../src/mat/impls/aij/seq/ftn-kernels/fmult.h> 1454 1455 PetscErrorCode MatMult_SeqAIJ(Mat A, Vec xx, Vec yy) 1456 { 1457 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 1458 PetscScalar *y; 1459 const PetscScalar *x; 1460 const MatScalar *aa, *a_a; 1461 PetscInt m = A->rmap->n; 1462 const PetscInt *aj, *ii, *ridx = NULL; 1463 PetscInt n, i; 1464 PetscScalar sum; 1465 PetscBool usecprow = a->compressedrow.use; 1466 1467 #if defined(PETSC_HAVE_PRAGMA_DISJOINT) 1468 #pragma disjoint(*x, *y, *aa) 1469 #endif 1470 1471 PetscFunctionBegin; 1472 if (a->inode.use && a->inode.checked) { 1473 PetscCall(MatMult_SeqAIJ_Inode(A, xx, yy)); 1474 PetscFunctionReturn(0); 1475 } 1476 PetscCall(MatSeqAIJGetArrayRead(A, &a_a)); 1477 PetscCall(VecGetArrayRead(xx, &x)); 1478 PetscCall(VecGetArray(yy, &y)); 1479 ii = a->i; 1480 if (usecprow) { /* use compressed row format */ 1481 PetscCall(PetscArrayzero(y, m)); 1482 m = a->compressedrow.nrows; 1483 ii = a->compressedrow.i; 1484 ridx = a->compressedrow.rindex; 1485 for (i = 0; i < m; i++) { 1486 n = ii[i + 1] - ii[i]; 1487 aj = a->j + ii[i]; 1488 aa = a_a + ii[i]; 1489 sum = 0.0; 1490 PetscSparseDensePlusDot(sum, x, aa, aj, n); 1491 /* for (j=0; j<n; j++) sum += (*aa++)*x[*aj++]; */ 1492 y[*ridx++] = sum; 1493 } 1494 } else { /* do not use compressed row format */ 1495 #if defined(PETSC_USE_FORTRAN_KERNEL_MULTAIJ) 1496 aj = a->j; 1497 aa = a_a; 1498 fortranmultaij_(&m, x, ii, aj, aa, y); 1499 #else 1500 for (i = 0; i < m; i++) { 1501 n = ii[i + 1] - ii[i]; 1502 aj = a->j + ii[i]; 1503 aa = a_a + ii[i]; 1504 sum = 0.0; 1505 PetscSparseDensePlusDot(sum, x, aa, aj, n); 1506 y[i] = sum; 1507 } 1508 #endif 1509 } 1510 PetscCall(PetscLogFlops(2.0 * a->nz - a->nonzerorowcnt)); 1511 PetscCall(VecRestoreArrayRead(xx, &x)); 1512 PetscCall(VecRestoreArray(yy, &y)); 1513 PetscCall(MatSeqAIJRestoreArrayRead(A, &a_a)); 1514 PetscFunctionReturn(0); 1515 } 1516 1517 PetscErrorCode MatMultMax_SeqAIJ(Mat A, Vec xx, Vec yy) 1518 { 1519 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 1520 PetscScalar *y; 1521 const PetscScalar *x; 1522 const MatScalar *aa, *a_a; 1523 PetscInt m = A->rmap->n; 1524 const PetscInt *aj, *ii, *ridx = NULL; 1525 PetscInt n, i, nonzerorow = 0; 1526 PetscScalar sum; 1527 PetscBool usecprow = a->compressedrow.use; 1528 1529 #if defined(PETSC_HAVE_PRAGMA_DISJOINT) 1530 #pragma disjoint(*x, *y, *aa) 1531 #endif 1532 1533 PetscFunctionBegin; 1534 PetscCall(MatSeqAIJGetArrayRead(A, &a_a)); 1535 PetscCall(VecGetArrayRead(xx, &x)); 1536 PetscCall(VecGetArray(yy, &y)); 1537 if (usecprow) { /* use compressed row format */ 1538 m = a->compressedrow.nrows; 1539 ii = a->compressedrow.i; 1540 ridx = a->compressedrow.rindex; 1541 for (i = 0; i < m; i++) { 1542 n = ii[i + 1] - ii[i]; 1543 aj = a->j + ii[i]; 1544 aa = a_a + ii[i]; 1545 sum = 0.0; 1546 nonzerorow += (n > 0); 1547 PetscSparseDenseMaxDot(sum, x, aa, aj, n); 1548 /* for (j=0; j<n; j++) sum += (*aa++)*x[*aj++]; */ 1549 y[*ridx++] = sum; 1550 } 1551 } else { /* do not use compressed row format */ 1552 ii = a->i; 1553 for (i = 0; i < m; i++) { 1554 n = ii[i + 1] - ii[i]; 1555 aj = a->j + ii[i]; 1556 aa = a_a + ii[i]; 1557 sum = 0.0; 1558 nonzerorow += (n > 0); 1559 PetscSparseDenseMaxDot(sum, x, aa, aj, n); 1560 y[i] = sum; 1561 } 1562 } 1563 PetscCall(PetscLogFlops(2.0 * a->nz - nonzerorow)); 1564 PetscCall(VecRestoreArrayRead(xx, &x)); 1565 PetscCall(VecRestoreArray(yy, &y)); 1566 PetscCall(MatSeqAIJRestoreArrayRead(A, &a_a)); 1567 PetscFunctionReturn(0); 1568 } 1569 1570 PetscErrorCode MatMultAddMax_SeqAIJ(Mat A, Vec xx, Vec yy, Vec zz) 1571 { 1572 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 1573 PetscScalar *y, *z; 1574 const PetscScalar *x; 1575 const MatScalar *aa, *a_a; 1576 PetscInt m = A->rmap->n, *aj, *ii; 1577 PetscInt n, i, *ridx = NULL; 1578 PetscScalar sum; 1579 PetscBool usecprow = a->compressedrow.use; 1580 1581 PetscFunctionBegin; 1582 PetscCall(MatSeqAIJGetArrayRead(A, &a_a)); 1583 PetscCall(VecGetArrayRead(xx, &x)); 1584 PetscCall(VecGetArrayPair(yy, zz, &y, &z)); 1585 if (usecprow) { /* use compressed row format */ 1586 if (zz != yy) PetscCall(PetscArraycpy(z, y, m)); 1587 m = a->compressedrow.nrows; 1588 ii = a->compressedrow.i; 1589 ridx = a->compressedrow.rindex; 1590 for (i = 0; i < m; i++) { 1591 n = ii[i + 1] - ii[i]; 1592 aj = a->j + ii[i]; 1593 aa = a_a + ii[i]; 1594 sum = y[*ridx]; 1595 PetscSparseDenseMaxDot(sum, x, aa, aj, n); 1596 z[*ridx++] = sum; 1597 } 1598 } else { /* do not use compressed row format */ 1599 ii = a->i; 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[i]; 1605 PetscSparseDenseMaxDot(sum, x, aa, aj, n); 1606 z[i] = sum; 1607 } 1608 } 1609 PetscCall(PetscLogFlops(2.0 * a->nz)); 1610 PetscCall(VecRestoreArrayRead(xx, &x)); 1611 PetscCall(VecRestoreArrayPair(yy, zz, &y, &z)); 1612 PetscCall(MatSeqAIJRestoreArrayRead(A, &a_a)); 1613 PetscFunctionReturn(0); 1614 } 1615 1616 #include <../src/mat/impls/aij/seq/ftn-kernels/fmultadd.h> 1617 PetscErrorCode MatMultAdd_SeqAIJ(Mat A, Vec xx, Vec yy, Vec zz) 1618 { 1619 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 1620 PetscScalar *y, *z; 1621 const PetscScalar *x; 1622 const MatScalar *aa, *a_a; 1623 const PetscInt *aj, *ii, *ridx = NULL; 1624 PetscInt m = A->rmap->n, n, i; 1625 PetscScalar sum; 1626 PetscBool usecprow = a->compressedrow.use; 1627 1628 PetscFunctionBegin; 1629 if (a->inode.use && a->inode.checked) { 1630 PetscCall(MatMultAdd_SeqAIJ_Inode(A, xx, yy, zz)); 1631 PetscFunctionReturn(0); 1632 } 1633 PetscCall(MatSeqAIJGetArrayRead(A, &a_a)); 1634 PetscCall(VecGetArrayRead(xx, &x)); 1635 PetscCall(VecGetArrayPair(yy, zz, &y, &z)); 1636 if (usecprow) { /* use compressed row format */ 1637 if (zz != yy) PetscCall(PetscArraycpy(z, y, m)); 1638 m = a->compressedrow.nrows; 1639 ii = a->compressedrow.i; 1640 ridx = a->compressedrow.rindex; 1641 for (i = 0; i < m; i++) { 1642 n = ii[i + 1] - ii[i]; 1643 aj = a->j + ii[i]; 1644 aa = a_a + ii[i]; 1645 sum = y[*ridx]; 1646 PetscSparseDensePlusDot(sum, x, aa, aj, n); 1647 z[*ridx++] = sum; 1648 } 1649 } else { /* do not use compressed row format */ 1650 ii = a->i; 1651 #if defined(PETSC_USE_FORTRAN_KERNEL_MULTADDAIJ) 1652 aj = a->j; 1653 aa = a_a; 1654 fortranmultaddaij_(&m, x, ii, aj, aa, y, z); 1655 #else 1656 for (i = 0; i < m; i++) { 1657 n = ii[i + 1] - ii[i]; 1658 aj = a->j + ii[i]; 1659 aa = a_a + ii[i]; 1660 sum = y[i]; 1661 PetscSparseDensePlusDot(sum, x, aa, aj, n); 1662 z[i] = sum; 1663 } 1664 #endif 1665 } 1666 PetscCall(PetscLogFlops(2.0 * a->nz)); 1667 PetscCall(VecRestoreArrayRead(xx, &x)); 1668 PetscCall(VecRestoreArrayPair(yy, zz, &y, &z)); 1669 PetscCall(MatSeqAIJRestoreArrayRead(A, &a_a)); 1670 PetscFunctionReturn(0); 1671 } 1672 1673 /* 1674 Adds diagonal pointers to sparse matrix structure. 1675 */ 1676 PetscErrorCode MatMarkDiagonal_SeqAIJ(Mat A) 1677 { 1678 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 1679 PetscInt i, j, m = A->rmap->n; 1680 PetscBool alreadySet = PETSC_TRUE; 1681 1682 PetscFunctionBegin; 1683 if (!a->diag) { 1684 PetscCall(PetscMalloc1(m, &a->diag)); 1685 alreadySet = PETSC_FALSE; 1686 } 1687 for (i = 0; i < A->rmap->n; i++) { 1688 /* If A's diagonal is already correctly set, this fast track enables cheap and repeated MatMarkDiagonal_SeqAIJ() calls */ 1689 if (alreadySet) { 1690 PetscInt pos = a->diag[i]; 1691 if (pos >= a->i[i] && pos < a->i[i + 1] && a->j[pos] == i) continue; 1692 } 1693 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 PetscInt cnt = 0; /* how many diagonals are missing */ 1712 1713 PetscFunctionBegin; 1714 if (!A->preallocated || !a->nz) { 1715 PetscCall(MatSeqAIJSetPreallocation(A, 1, NULL)); 1716 PetscCall(MatShift_Basic(A, v)); 1717 PetscFunctionReturn(0); 1718 } 1719 1720 if (a->diagonaldense) { 1721 cnt = 0; 1722 } else { 1723 PetscCall(PetscCalloc1(A->rmap->n, &mdiag)); 1724 for (i = 0; i < A->rmap->n; i++) { 1725 if (i < A->cmap->n && diag[i] >= ii[i + 1]) { /* 'out of range' rows never have diagonals */ 1726 cnt++; 1727 mdiag[i] = 1; 1728 } 1729 } 1730 } 1731 if (!cnt) { 1732 PetscCall(MatShift_Basic(A, v)); 1733 } else { 1734 PetscScalar *olda = a->a; /* preserve pointers to current matrix nonzeros structure and values */ 1735 PetscInt *oldj = a->j, *oldi = a->i; 1736 PetscBool singlemalloc = a->singlemalloc, free_a = a->free_a, free_ij = a->free_ij; 1737 1738 a->a = NULL; 1739 a->j = NULL; 1740 a->i = NULL; 1741 /* increase the values in imax for each row where a diagonal is being inserted then reallocate the matrix data structures */ 1742 for (i = 0; i < PetscMin(A->rmap->n, A->cmap->n); i++) a->imax[i] += mdiag[i]; 1743 PetscCall(MatSeqAIJSetPreallocation_SeqAIJ(A, 0, a->imax)); 1744 1745 /* copy old values into new matrix data structure */ 1746 for (i = 0; i < A->rmap->n; i++) { 1747 PetscCall(MatSetValues(A, 1, &i, a->imax[i] - mdiag[i], &oldj[oldi[i]], &olda[oldi[i]], ADD_VALUES)); 1748 if (i < A->cmap->n) PetscCall(MatSetValue(A, i, i, v, ADD_VALUES)); 1749 } 1750 PetscCall(MatAssemblyBegin(A, MAT_FINAL_ASSEMBLY)); 1751 PetscCall(MatAssemblyEnd(A, MAT_FINAL_ASSEMBLY)); 1752 if (singlemalloc) { 1753 PetscCall(PetscFree3(olda, oldj, oldi)); 1754 } else { 1755 if (free_a) PetscCall(PetscFree(olda)); 1756 if (free_ij) PetscCall(PetscFree(oldj)); 1757 if (free_ij) PetscCall(PetscFree(oldi)); 1758 } 1759 } 1760 PetscCall(PetscFree(mdiag)); 1761 a->diagonaldense = PETSC_TRUE; 1762 PetscFunctionReturn(0); 1763 } 1764 1765 /* 1766 Checks for missing diagonals 1767 */ 1768 PetscErrorCode MatMissingDiagonal_SeqAIJ(Mat A, PetscBool *missing, PetscInt *d) 1769 { 1770 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 1771 PetscInt *diag, *ii = a->i, i; 1772 1773 PetscFunctionBegin; 1774 *missing = PETSC_FALSE; 1775 if (A->rmap->n > 0 && !ii) { 1776 *missing = PETSC_TRUE; 1777 if (d) *d = 0; 1778 PetscCall(PetscInfo(A, "Matrix has no entries therefore is missing diagonal\n")); 1779 } else { 1780 PetscInt n; 1781 n = PetscMin(A->rmap->n, A->cmap->n); 1782 diag = a->diag; 1783 for (i = 0; i < n; i++) { 1784 if (diag[i] >= ii[i + 1]) { 1785 *missing = PETSC_TRUE; 1786 if (d) *d = i; 1787 PetscCall(PetscInfo(A, "Matrix is missing diagonal number %" PetscInt_FMT "\n", i)); 1788 break; 1789 } 1790 } 1791 } 1792 PetscFunctionReturn(0); 1793 } 1794 1795 #include <petscblaslapack.h> 1796 #include <petsc/private/kernels/blockinvert.h> 1797 1798 /* 1799 Note that values is allocated externally by the PC and then passed into this routine 1800 */ 1801 PetscErrorCode MatInvertVariableBlockDiagonal_SeqAIJ(Mat A, PetscInt nblocks, const PetscInt *bsizes, PetscScalar *diag) 1802 { 1803 PetscInt n = A->rmap->n, i, ncnt = 0, *indx, j, bsizemax = 0, *v_pivots; 1804 PetscBool allowzeropivot, zeropivotdetected = PETSC_FALSE; 1805 const PetscReal shift = 0.0; 1806 PetscInt ipvt[5]; 1807 PetscCount flops = 0; 1808 PetscScalar work[25], *v_work; 1809 1810 PetscFunctionBegin; 1811 allowzeropivot = PetscNot(A->erroriffailure); 1812 for (i = 0; i < nblocks; i++) ncnt += bsizes[i]; 1813 PetscCheck(ncnt == n, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Total blocksizes %" PetscInt_FMT " doesn't match number matrix rows %" PetscInt_FMT, ncnt, n); 1814 for (i = 0; i < nblocks; i++) bsizemax = PetscMax(bsizemax, bsizes[i]); 1815 PetscCall(PetscMalloc1(bsizemax, &indx)); 1816 if (bsizemax > 7) PetscCall(PetscMalloc2(bsizemax, &v_work, bsizemax, &v_pivots)); 1817 ncnt = 0; 1818 for (i = 0; i < nblocks; i++) { 1819 for (j = 0; j < bsizes[i]; j++) indx[j] = ncnt + j; 1820 PetscCall(MatGetValues(A, bsizes[i], indx, bsizes[i], indx, diag)); 1821 switch (bsizes[i]) { 1822 case 1: 1823 *diag = 1.0 / (*diag); 1824 break; 1825 case 2: 1826 PetscCall(PetscKernel_A_gets_inverse_A_2(diag, shift, allowzeropivot, &zeropivotdetected)); 1827 if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT; 1828 PetscCall(PetscKernel_A_gets_transpose_A_2(diag)); 1829 break; 1830 case 3: 1831 PetscCall(PetscKernel_A_gets_inverse_A_3(diag, shift, allowzeropivot, &zeropivotdetected)); 1832 if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT; 1833 PetscCall(PetscKernel_A_gets_transpose_A_3(diag)); 1834 break; 1835 case 4: 1836 PetscCall(PetscKernel_A_gets_inverse_A_4(diag, shift, allowzeropivot, &zeropivotdetected)); 1837 if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT; 1838 PetscCall(PetscKernel_A_gets_transpose_A_4(diag)); 1839 break; 1840 case 5: 1841 PetscCall(PetscKernel_A_gets_inverse_A_5(diag, ipvt, work, shift, allowzeropivot, &zeropivotdetected)); 1842 if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT; 1843 PetscCall(PetscKernel_A_gets_transpose_A_5(diag)); 1844 break; 1845 case 6: 1846 PetscCall(PetscKernel_A_gets_inverse_A_6(diag, shift, allowzeropivot, &zeropivotdetected)); 1847 if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT; 1848 PetscCall(PetscKernel_A_gets_transpose_A_6(diag)); 1849 break; 1850 case 7: 1851 PetscCall(PetscKernel_A_gets_inverse_A_7(diag, shift, allowzeropivot, &zeropivotdetected)); 1852 if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT; 1853 PetscCall(PetscKernel_A_gets_transpose_A_7(diag)); 1854 break; 1855 default: 1856 PetscCall(PetscKernel_A_gets_inverse_A(bsizes[i], diag, v_pivots, v_work, allowzeropivot, &zeropivotdetected)); 1857 if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT; 1858 PetscCall(PetscKernel_A_gets_transpose_A_N(diag, bsizes[i])); 1859 } 1860 ncnt += bsizes[i]; 1861 diag += bsizes[i] * bsizes[i]; 1862 flops += 2 * PetscPowInt(bsizes[i], 3) / 3; 1863 } 1864 PetscLogFlops(flops); 1865 if (bsizemax > 7) PetscCall(PetscFree2(v_work, v_pivots)); 1866 PetscCall(PetscFree(indx)); 1867 PetscFunctionReturn(0); 1868 } 1869 1870 /* 1871 Negative shift indicates do not generate an error if there is a zero diagonal, just invert it anyways 1872 */ 1873 PetscErrorCode MatInvertDiagonal_SeqAIJ(Mat A, PetscScalar omega, PetscScalar fshift) 1874 { 1875 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 1876 PetscInt i, *diag, m = A->rmap->n; 1877 const MatScalar *v; 1878 PetscScalar *idiag, *mdiag; 1879 1880 PetscFunctionBegin; 1881 if (a->idiagvalid) PetscFunctionReturn(0); 1882 PetscCall(MatMarkDiagonal_SeqAIJ(A)); 1883 diag = a->diag; 1884 if (!a->idiag) { PetscCall(PetscMalloc3(m, &a->idiag, m, &a->mdiag, m, &a->ssor_work)); } 1885 1886 mdiag = a->mdiag; 1887 idiag = a->idiag; 1888 PetscCall(MatSeqAIJGetArrayRead(A, &v)); 1889 if (omega == 1.0 && PetscRealPart(fshift) <= 0.0) { 1890 for (i = 0; i < m; i++) { 1891 mdiag[i] = v[diag[i]]; 1892 if (!PetscAbsScalar(mdiag[i])) { /* zero diagonal */ 1893 if (PetscRealPart(fshift)) { 1894 PetscCall(PetscInfo(A, "Zero diagonal on row %" PetscInt_FMT "\n", i)); 1895 A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT; 1896 A->factorerror_zeropivot_value = 0.0; 1897 A->factorerror_zeropivot_row = i; 1898 } else SETERRQ(PETSC_COMM_SELF, PETSC_ERR_ARG_INCOMP, "Zero diagonal on row %" PetscInt_FMT, i); 1899 } 1900 idiag[i] = 1.0 / v[diag[i]]; 1901 } 1902 PetscCall(PetscLogFlops(m)); 1903 } else { 1904 for (i = 0; i < m; i++) { 1905 mdiag[i] = v[diag[i]]; 1906 idiag[i] = omega / (fshift + v[diag[i]]); 1907 } 1908 PetscCall(PetscLogFlops(2.0 * m)); 1909 } 1910 a->idiagvalid = PETSC_TRUE; 1911 PetscCall(MatSeqAIJRestoreArrayRead(A, &v)); 1912 PetscFunctionReturn(0); 1913 } 1914 1915 #include <../src/mat/impls/aij/seq/ftn-kernels/frelax.h> 1916 PetscErrorCode MatSOR_SeqAIJ(Mat A, Vec bb, PetscReal omega, MatSORType flag, PetscReal fshift, PetscInt its, PetscInt lits, Vec xx) 1917 { 1918 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 1919 PetscScalar *x, d, sum, *t, scale; 1920 const MatScalar *v, *idiag = NULL, *mdiag, *aa; 1921 const PetscScalar *b, *bs, *xb, *ts; 1922 PetscInt n, m = A->rmap->n, i; 1923 const PetscInt *idx, *diag; 1924 1925 PetscFunctionBegin; 1926 if (a->inode.use && a->inode.checked && omega == 1.0 && fshift == 0.0) { 1927 PetscCall(MatSOR_SeqAIJ_Inode(A, bb, omega, flag, fshift, its, lits, xx)); 1928 PetscFunctionReturn(0); 1929 } 1930 its = its * lits; 1931 1932 if (fshift != a->fshift || omega != a->omega) a->idiagvalid = PETSC_FALSE; /* must recompute idiag[] */ 1933 if (!a->idiagvalid) PetscCall(MatInvertDiagonal_SeqAIJ(A, omega, fshift)); 1934 a->fshift = fshift; 1935 a->omega = omega; 1936 1937 diag = a->diag; 1938 t = a->ssor_work; 1939 idiag = a->idiag; 1940 mdiag = a->mdiag; 1941 1942 PetscCall(MatSeqAIJGetArrayRead(A, &aa)); 1943 PetscCall(VecGetArray(xx, &x)); 1944 PetscCall(VecGetArrayRead(bb, &b)); 1945 /* We count flops by assuming the upper triangular and lower triangular parts have the same number of nonzeros */ 1946 if (flag == SOR_APPLY_UPPER) { 1947 /* apply (U + D/omega) to the vector */ 1948 bs = b; 1949 for (i = 0; i < m; i++) { 1950 d = fshift + mdiag[i]; 1951 n = a->i[i + 1] - diag[i] - 1; 1952 idx = a->j + diag[i] + 1; 1953 v = aa + diag[i] + 1; 1954 sum = b[i] * d / omega; 1955 PetscSparseDensePlusDot(sum, bs, v, idx, n); 1956 x[i] = sum; 1957 } 1958 PetscCall(VecRestoreArray(xx, &x)); 1959 PetscCall(VecRestoreArrayRead(bb, &b)); 1960 PetscCall(MatSeqAIJRestoreArrayRead(A, &aa)); 1961 PetscCall(PetscLogFlops(a->nz)); 1962 PetscFunctionReturn(0); 1963 } 1964 1965 PetscCheck(flag != SOR_APPLY_LOWER, PETSC_COMM_SELF, PETSC_ERR_SUP, "SOR_APPLY_LOWER is not implemented"); 1966 if (flag & SOR_EISENSTAT) { 1967 /* Let A = L + U + D; where L is lower triangular, 1968 U is upper triangular, E = D/omega; This routine applies 1969 1970 (L + E)^{-1} A (U + E)^{-1} 1971 1972 to a vector efficiently using Eisenstat's trick. 1973 */ 1974 scale = (2.0 / omega) - 1.0; 1975 1976 /* x = (E + U)^{-1} b */ 1977 for (i = m - 1; i >= 0; i--) { 1978 n = a->i[i + 1] - diag[i] - 1; 1979 idx = a->j + diag[i] + 1; 1980 v = aa + diag[i] + 1; 1981 sum = b[i]; 1982 PetscSparseDenseMinusDot(sum, x, v, idx, n); 1983 x[i] = sum * idiag[i]; 1984 } 1985 1986 /* t = b - (2*E - D)x */ 1987 v = aa; 1988 for (i = 0; i < m; i++) t[i] = b[i] - scale * (v[*diag++]) * x[i]; 1989 1990 /* t = (E + L)^{-1}t */ 1991 ts = t; 1992 diag = a->diag; 1993 for (i = 0; i < m; i++) { 1994 n = diag[i] - a->i[i]; 1995 idx = a->j + a->i[i]; 1996 v = aa + a->i[i]; 1997 sum = t[i]; 1998 PetscSparseDenseMinusDot(sum, ts, v, idx, n); 1999 t[i] = sum * idiag[i]; 2000 /* x = x + t */ 2001 x[i] += t[i]; 2002 } 2003 2004 PetscCall(PetscLogFlops(6.0 * m - 1 + 2.0 * a->nz)); 2005 PetscCall(VecRestoreArray(xx, &x)); 2006 PetscCall(VecRestoreArrayRead(bb, &b)); 2007 PetscFunctionReturn(0); 2008 } 2009 if (flag & SOR_ZERO_INITIAL_GUESS) { 2010 if (flag & SOR_FORWARD_SWEEP || flag & SOR_LOCAL_FORWARD_SWEEP) { 2011 for (i = 0; i < m; i++) { 2012 n = diag[i] - a->i[i]; 2013 idx = a->j + a->i[i]; 2014 v = aa + a->i[i]; 2015 sum = b[i]; 2016 PetscSparseDenseMinusDot(sum, x, v, idx, n); 2017 t[i] = sum; 2018 x[i] = sum * idiag[i]; 2019 } 2020 xb = t; 2021 PetscCall(PetscLogFlops(a->nz)); 2022 } else xb = b; 2023 if (flag & SOR_BACKWARD_SWEEP || flag & SOR_LOCAL_BACKWARD_SWEEP) { 2024 for (i = m - 1; i >= 0; i--) { 2025 n = a->i[i + 1] - diag[i] - 1; 2026 idx = a->j + diag[i] + 1; 2027 v = aa + diag[i] + 1; 2028 sum = xb[i]; 2029 PetscSparseDenseMinusDot(sum, x, v, idx, n); 2030 if (xb == b) { 2031 x[i] = sum * idiag[i]; 2032 } else { 2033 x[i] = (1 - omega) * x[i] + sum * idiag[i]; /* omega in idiag */ 2034 } 2035 } 2036 PetscCall(PetscLogFlops(a->nz)); /* assumes 1/2 in upper */ 2037 } 2038 its--; 2039 } 2040 while (its--) { 2041 if (flag & SOR_FORWARD_SWEEP || flag & SOR_LOCAL_FORWARD_SWEEP) { 2042 for (i = 0; i < m; i++) { 2043 /* lower */ 2044 n = diag[i] - a->i[i]; 2045 idx = a->j + a->i[i]; 2046 v = aa + a->i[i]; 2047 sum = b[i]; 2048 PetscSparseDenseMinusDot(sum, x, v, idx, n); 2049 t[i] = sum; /* save application of the lower-triangular part */ 2050 /* upper */ 2051 n = a->i[i + 1] - diag[i] - 1; 2052 idx = a->j + diag[i] + 1; 2053 v = aa + diag[i] + 1; 2054 PetscSparseDenseMinusDot(sum, x, v, idx, n); 2055 x[i] = (1. - omega) * x[i] + sum * idiag[i]; /* omega in idiag */ 2056 } 2057 xb = t; 2058 PetscCall(PetscLogFlops(2.0 * a->nz)); 2059 } else xb = b; 2060 if (flag & SOR_BACKWARD_SWEEP || flag & SOR_LOCAL_BACKWARD_SWEEP) { 2061 for (i = m - 1; i >= 0; i--) { 2062 sum = xb[i]; 2063 if (xb == b) { 2064 /* whole matrix (no checkpointing available) */ 2065 n = a->i[i + 1] - a->i[i]; 2066 idx = a->j + a->i[i]; 2067 v = aa + a->i[i]; 2068 PetscSparseDenseMinusDot(sum, x, v, idx, n); 2069 x[i] = (1. - omega) * x[i] + (sum + mdiag[i] * x[i]) * idiag[i]; 2070 } else { /* lower-triangular part has been saved, so only apply upper-triangular */ 2071 n = a->i[i + 1] - diag[i] - 1; 2072 idx = a->j + diag[i] + 1; 2073 v = aa + diag[i] + 1; 2074 PetscSparseDenseMinusDot(sum, x, v, idx, n); 2075 x[i] = (1. - omega) * x[i] + sum * idiag[i]; /* omega in idiag */ 2076 } 2077 } 2078 if (xb == b) { 2079 PetscCall(PetscLogFlops(2.0 * a->nz)); 2080 } else { 2081 PetscCall(PetscLogFlops(a->nz)); /* assumes 1/2 in upper */ 2082 } 2083 } 2084 } 2085 PetscCall(MatSeqAIJRestoreArrayRead(A, &aa)); 2086 PetscCall(VecRestoreArray(xx, &x)); 2087 PetscCall(VecRestoreArrayRead(bb, &b)); 2088 PetscFunctionReturn(0); 2089 } 2090 2091 PetscErrorCode MatGetInfo_SeqAIJ(Mat A, MatInfoType flag, MatInfo *info) 2092 { 2093 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 2094 2095 PetscFunctionBegin; 2096 info->block_size = 1.0; 2097 info->nz_allocated = a->maxnz; 2098 info->nz_used = a->nz; 2099 info->nz_unneeded = (a->maxnz - a->nz); 2100 info->assemblies = A->num_ass; 2101 info->mallocs = A->info.mallocs; 2102 info->memory = 0; /* REVIEW ME */ 2103 if (A->factortype) { 2104 info->fill_ratio_given = A->info.fill_ratio_given; 2105 info->fill_ratio_needed = A->info.fill_ratio_needed; 2106 info->factor_mallocs = A->info.factor_mallocs; 2107 } else { 2108 info->fill_ratio_given = 0; 2109 info->fill_ratio_needed = 0; 2110 info->factor_mallocs = 0; 2111 } 2112 PetscFunctionReturn(0); 2113 } 2114 2115 PetscErrorCode MatZeroRows_SeqAIJ(Mat A, PetscInt N, const PetscInt rows[], PetscScalar diag, Vec x, Vec b) 2116 { 2117 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 2118 PetscInt i, m = A->rmap->n - 1; 2119 const PetscScalar *xx; 2120 PetscScalar *bb, *aa; 2121 PetscInt d = 0; 2122 2123 PetscFunctionBegin; 2124 if (x && b) { 2125 PetscCall(VecGetArrayRead(x, &xx)); 2126 PetscCall(VecGetArray(b, &bb)); 2127 for (i = 0; i < N; i++) { 2128 PetscCheck(rows[i] >= 0 && rows[i] <= m, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "row %" PetscInt_FMT " out of range", rows[i]); 2129 if (rows[i] >= A->cmap->n) continue; 2130 bb[rows[i]] = diag * xx[rows[i]]; 2131 } 2132 PetscCall(VecRestoreArrayRead(x, &xx)); 2133 PetscCall(VecRestoreArray(b, &bb)); 2134 } 2135 2136 PetscCall(MatSeqAIJGetArray(A, &aa)); 2137 if (a->keepnonzeropattern) { 2138 for (i = 0; i < N; i++) { 2139 PetscCheck(rows[i] >= 0 && rows[i] <= m, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "row %" PetscInt_FMT " out of range", rows[i]); 2140 PetscCall(PetscArrayzero(&aa[a->i[rows[i]]], a->ilen[rows[i]])); 2141 } 2142 if (diag != 0.0) { 2143 for (i = 0; i < N; i++) { 2144 d = rows[i]; 2145 if (rows[i] >= A->cmap->n) continue; 2146 PetscCheck(a->diag[d] < a->i[d + 1], PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Matrix is missing diagonal entry in the zeroed row %" PetscInt_FMT, d); 2147 } 2148 for (i = 0; i < N; i++) { 2149 if (rows[i] >= A->cmap->n) continue; 2150 aa[a->diag[rows[i]]] = diag; 2151 } 2152 } 2153 } else { 2154 if (diag != 0.0) { 2155 for (i = 0; i < N; i++) { 2156 PetscCheck(rows[i] >= 0 && rows[i] <= m, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "row %" PetscInt_FMT " out of range", rows[i]); 2157 if (a->ilen[rows[i]] > 0) { 2158 if (rows[i] >= A->cmap->n) { 2159 a->ilen[rows[i]] = 0; 2160 } else { 2161 a->ilen[rows[i]] = 1; 2162 aa[a->i[rows[i]]] = diag; 2163 a->j[a->i[rows[i]]] = rows[i]; 2164 } 2165 } else if (rows[i] < A->cmap->n) { /* in case row was completely empty */ 2166 PetscCall(MatSetValues_SeqAIJ(A, 1, &rows[i], 1, &rows[i], &diag, INSERT_VALUES)); 2167 } 2168 } 2169 } else { 2170 for (i = 0; i < N; i++) { 2171 PetscCheck(rows[i] >= 0 && rows[i] <= m, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "row %" PetscInt_FMT " out of range", rows[i]); 2172 a->ilen[rows[i]] = 0; 2173 } 2174 } 2175 A->nonzerostate++; 2176 } 2177 PetscCall(MatSeqAIJRestoreArray(A, &aa)); 2178 PetscUseTypeMethod(A, assemblyend, MAT_FINAL_ASSEMBLY); 2179 PetscFunctionReturn(0); 2180 } 2181 2182 PetscErrorCode MatZeroRowsColumns_SeqAIJ(Mat A, PetscInt N, const PetscInt rows[], PetscScalar diag, Vec x, Vec b) 2183 { 2184 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 2185 PetscInt i, j, m = A->rmap->n - 1, d = 0; 2186 PetscBool missing, *zeroed, vecs = PETSC_FALSE; 2187 const PetscScalar *xx; 2188 PetscScalar *bb, *aa; 2189 2190 PetscFunctionBegin; 2191 if (!N) PetscFunctionReturn(0); 2192 PetscCall(MatSeqAIJGetArray(A, &aa)); 2193 if (x && b) { 2194 PetscCall(VecGetArrayRead(x, &xx)); 2195 PetscCall(VecGetArray(b, &bb)); 2196 vecs = PETSC_TRUE; 2197 } 2198 PetscCall(PetscCalloc1(A->rmap->n, &zeroed)); 2199 for (i = 0; i < N; i++) { 2200 PetscCheck(rows[i] >= 0 && rows[i] <= m, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "row %" PetscInt_FMT " out of range", rows[i]); 2201 PetscCall(PetscArrayzero(&aa[a->i[rows[i]]], a->ilen[rows[i]])); 2202 2203 zeroed[rows[i]] = PETSC_TRUE; 2204 } 2205 for (i = 0; i < A->rmap->n; i++) { 2206 if (!zeroed[i]) { 2207 for (j = a->i[i]; j < a->i[i + 1]; j++) { 2208 if (a->j[j] < A->rmap->n && zeroed[a->j[j]]) { 2209 if (vecs) bb[i] -= aa[j] * xx[a->j[j]]; 2210 aa[j] = 0.0; 2211 } 2212 } 2213 } else if (vecs && i < A->cmap->N) bb[i] = diag * xx[i]; 2214 } 2215 if (x && b) { 2216 PetscCall(VecRestoreArrayRead(x, &xx)); 2217 PetscCall(VecRestoreArray(b, &bb)); 2218 } 2219 PetscCall(PetscFree(zeroed)); 2220 if (diag != 0.0) { 2221 PetscCall(MatMissingDiagonal_SeqAIJ(A, &missing, &d)); 2222 if (missing) { 2223 for (i = 0; i < N; i++) { 2224 if (rows[i] >= A->cmap->N) continue; 2225 PetscCheck(!a->nonew || rows[i] < d, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Matrix is missing diagonal entry in row %" PetscInt_FMT " (%" PetscInt_FMT ")", d, rows[i]); 2226 PetscCall(MatSetValues_SeqAIJ(A, 1, &rows[i], 1, &rows[i], &diag, INSERT_VALUES)); 2227 } 2228 } else { 2229 for (i = 0; i < N; i++) aa[a->diag[rows[i]]] = diag; 2230 } 2231 } 2232 PetscCall(MatSeqAIJRestoreArray(A, &aa)); 2233 PetscUseTypeMethod(A, assemblyend, MAT_FINAL_ASSEMBLY); 2234 PetscFunctionReturn(0); 2235 } 2236 2237 PetscErrorCode MatGetRow_SeqAIJ(Mat A, PetscInt row, PetscInt *nz, PetscInt **idx, PetscScalar **v) 2238 { 2239 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 2240 const PetscScalar *aa; 2241 PetscInt *itmp; 2242 2243 PetscFunctionBegin; 2244 PetscCall(MatSeqAIJGetArrayRead(A, &aa)); 2245 *nz = a->i[row + 1] - a->i[row]; 2246 if (v) *v = (PetscScalar *)(aa + a->i[row]); 2247 if (idx) { 2248 itmp = a->j + a->i[row]; 2249 if (*nz) *idx = itmp; 2250 else *idx = NULL; 2251 } 2252 PetscCall(MatSeqAIJRestoreArrayRead(A, &aa)); 2253 PetscFunctionReturn(0); 2254 } 2255 2256 PetscErrorCode MatRestoreRow_SeqAIJ(Mat A, PetscInt row, PetscInt *nz, PetscInt **idx, PetscScalar **v) 2257 { 2258 PetscFunctionBegin; 2259 if (nz) *nz = 0; 2260 if (idx) *idx = NULL; 2261 if (v) *v = NULL; 2262 PetscFunctionReturn(0); 2263 } 2264 2265 PetscErrorCode MatNorm_SeqAIJ(Mat A, NormType type, PetscReal *nrm) 2266 { 2267 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 2268 const MatScalar *v; 2269 PetscReal sum = 0.0; 2270 PetscInt i, j; 2271 2272 PetscFunctionBegin; 2273 PetscCall(MatSeqAIJGetArrayRead(A, &v)); 2274 if (type == NORM_FROBENIUS) { 2275 #if defined(PETSC_USE_REAL___FP16) 2276 PetscBLASInt one = 1, nz = a->nz; 2277 PetscCallBLAS("BLASnrm2", *nrm = BLASnrm2_(&nz, v, &one)); 2278 #else 2279 for (i = 0; i < a->nz; i++) { 2280 sum += PetscRealPart(PetscConj(*v) * (*v)); 2281 v++; 2282 } 2283 *nrm = PetscSqrtReal(sum); 2284 #endif 2285 PetscCall(PetscLogFlops(2.0 * a->nz)); 2286 } else if (type == NORM_1) { 2287 PetscReal *tmp; 2288 PetscInt *jj = a->j; 2289 PetscCall(PetscCalloc1(A->cmap->n + 1, &tmp)); 2290 *nrm = 0.0; 2291 for (j = 0; j < a->nz; j++) { 2292 tmp[*jj++] += PetscAbsScalar(*v); 2293 v++; 2294 } 2295 for (j = 0; j < A->cmap->n; j++) { 2296 if (tmp[j] > *nrm) *nrm = tmp[j]; 2297 } 2298 PetscCall(PetscFree(tmp)); 2299 PetscCall(PetscLogFlops(PetscMax(a->nz - 1, 0))); 2300 } else if (type == NORM_INFINITY) { 2301 *nrm = 0.0; 2302 for (j = 0; j < A->rmap->n; j++) { 2303 const PetscScalar *v2 = v + a->i[j]; 2304 sum = 0.0; 2305 for (i = 0; i < a->i[j + 1] - a->i[j]; i++) { 2306 sum += PetscAbsScalar(*v2); 2307 v2++; 2308 } 2309 if (sum > *nrm) *nrm = sum; 2310 } 2311 PetscCall(PetscLogFlops(PetscMax(a->nz - 1, 0))); 2312 } else SETERRQ(PETSC_COMM_SELF, PETSC_ERR_SUP, "No support for two norm"); 2313 PetscCall(MatSeqAIJRestoreArrayRead(A, &v)); 2314 PetscFunctionReturn(0); 2315 } 2316 2317 PetscErrorCode MatIsTranspose_SeqAIJ(Mat A, Mat B, PetscReal tol, PetscBool *f) 2318 { 2319 Mat_SeqAIJ *aij = (Mat_SeqAIJ *)A->data, *bij = (Mat_SeqAIJ *)B->data; 2320 PetscInt *adx, *bdx, *aii, *bii, *aptr, *bptr; 2321 const MatScalar *va, *vb; 2322 PetscInt ma, na, mb, nb, i; 2323 2324 PetscFunctionBegin; 2325 PetscCall(MatGetSize(A, &ma, &na)); 2326 PetscCall(MatGetSize(B, &mb, &nb)); 2327 if (ma != nb || na != mb) { 2328 *f = PETSC_FALSE; 2329 PetscFunctionReturn(0); 2330 } 2331 PetscCall(MatSeqAIJGetArrayRead(A, &va)); 2332 PetscCall(MatSeqAIJGetArrayRead(B, &vb)); 2333 aii = aij->i; 2334 bii = bij->i; 2335 adx = aij->j; 2336 bdx = bij->j; 2337 PetscCall(PetscMalloc1(ma, &aptr)); 2338 PetscCall(PetscMalloc1(mb, &bptr)); 2339 for (i = 0; i < ma; i++) aptr[i] = aii[i]; 2340 for (i = 0; i < mb; i++) bptr[i] = bii[i]; 2341 2342 *f = PETSC_TRUE; 2343 for (i = 0; i < ma; i++) { 2344 while (aptr[i] < aii[i + 1]) { 2345 PetscInt idc, idr; 2346 PetscScalar vc, vr; 2347 /* column/row index/value */ 2348 idc = adx[aptr[i]]; 2349 idr = bdx[bptr[idc]]; 2350 vc = va[aptr[i]]; 2351 vr = vb[bptr[idc]]; 2352 if (i != idr || PetscAbsScalar(vc - vr) > tol) { 2353 *f = PETSC_FALSE; 2354 goto done; 2355 } else { 2356 aptr[i]++; 2357 if (B || i != idc) bptr[idc]++; 2358 } 2359 } 2360 } 2361 done: 2362 PetscCall(PetscFree(aptr)); 2363 PetscCall(PetscFree(bptr)); 2364 PetscCall(MatSeqAIJRestoreArrayRead(A, &va)); 2365 PetscCall(MatSeqAIJRestoreArrayRead(B, &vb)); 2366 PetscFunctionReturn(0); 2367 } 2368 2369 PetscErrorCode MatIsHermitianTranspose_SeqAIJ(Mat A, Mat B, PetscReal tol, PetscBool *f) 2370 { 2371 Mat_SeqAIJ *aij = (Mat_SeqAIJ *)A->data, *bij = (Mat_SeqAIJ *)B->data; 2372 PetscInt *adx, *bdx, *aii, *bii, *aptr, *bptr; 2373 MatScalar *va, *vb; 2374 PetscInt ma, na, mb, nb, i; 2375 2376 PetscFunctionBegin; 2377 PetscCall(MatGetSize(A, &ma, &na)); 2378 PetscCall(MatGetSize(B, &mb, &nb)); 2379 if (ma != nb || na != mb) { 2380 *f = PETSC_FALSE; 2381 PetscFunctionReturn(0); 2382 } 2383 aii = aij->i; 2384 bii = bij->i; 2385 adx = aij->j; 2386 bdx = bij->j; 2387 va = aij->a; 2388 vb = bij->a; 2389 PetscCall(PetscMalloc1(ma, &aptr)); 2390 PetscCall(PetscMalloc1(mb, &bptr)); 2391 for (i = 0; i < ma; i++) aptr[i] = aii[i]; 2392 for (i = 0; i < mb; i++) bptr[i] = bii[i]; 2393 2394 *f = PETSC_TRUE; 2395 for (i = 0; i < ma; i++) { 2396 while (aptr[i] < aii[i + 1]) { 2397 PetscInt idc, idr; 2398 PetscScalar vc, vr; 2399 /* column/row index/value */ 2400 idc = adx[aptr[i]]; 2401 idr = bdx[bptr[idc]]; 2402 vc = va[aptr[i]]; 2403 vr = vb[bptr[idc]]; 2404 if (i != idr || PetscAbsScalar(vc - PetscConj(vr)) > tol) { 2405 *f = PETSC_FALSE; 2406 goto done; 2407 } else { 2408 aptr[i]++; 2409 if (B || i != idc) bptr[idc]++; 2410 } 2411 } 2412 } 2413 done: 2414 PetscCall(PetscFree(aptr)); 2415 PetscCall(PetscFree(bptr)); 2416 PetscFunctionReturn(0); 2417 } 2418 2419 PetscErrorCode MatIsSymmetric_SeqAIJ(Mat A, PetscReal tol, PetscBool *f) 2420 { 2421 PetscFunctionBegin; 2422 PetscCall(MatIsTranspose_SeqAIJ(A, A, tol, f)); 2423 PetscFunctionReturn(0); 2424 } 2425 2426 PetscErrorCode MatIsHermitian_SeqAIJ(Mat A, PetscReal tol, PetscBool *f) 2427 { 2428 PetscFunctionBegin; 2429 PetscCall(MatIsHermitianTranspose_SeqAIJ(A, A, tol, f)); 2430 PetscFunctionReturn(0); 2431 } 2432 2433 PetscErrorCode MatDiagonalScale_SeqAIJ(Mat A, Vec ll, Vec rr) 2434 { 2435 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 2436 const PetscScalar *l, *r; 2437 PetscScalar x; 2438 MatScalar *v; 2439 PetscInt i, j, m = A->rmap->n, n = A->cmap->n, M, nz = a->nz; 2440 const PetscInt *jj; 2441 2442 PetscFunctionBegin; 2443 if (ll) { 2444 /* The local size is used so that VecMPI can be passed to this routine 2445 by MatDiagonalScale_MPIAIJ */ 2446 PetscCall(VecGetLocalSize(ll, &m)); 2447 PetscCheck(m == A->rmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Left scaling vector wrong length"); 2448 PetscCall(VecGetArrayRead(ll, &l)); 2449 PetscCall(MatSeqAIJGetArray(A, &v)); 2450 for (i = 0; i < m; i++) { 2451 x = l[i]; 2452 M = a->i[i + 1] - a->i[i]; 2453 for (j = 0; j < M; j++) (*v++) *= x; 2454 } 2455 PetscCall(VecRestoreArrayRead(ll, &l)); 2456 PetscCall(PetscLogFlops(nz)); 2457 PetscCall(MatSeqAIJRestoreArray(A, &v)); 2458 } 2459 if (rr) { 2460 PetscCall(VecGetLocalSize(rr, &n)); 2461 PetscCheck(n == A->cmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Right scaling vector wrong length"); 2462 PetscCall(VecGetArrayRead(rr, &r)); 2463 PetscCall(MatSeqAIJGetArray(A, &v)); 2464 jj = a->j; 2465 for (i = 0; i < nz; i++) (*v++) *= r[*jj++]; 2466 PetscCall(MatSeqAIJRestoreArray(A, &v)); 2467 PetscCall(VecRestoreArrayRead(rr, &r)); 2468 PetscCall(PetscLogFlops(nz)); 2469 } 2470 PetscCall(MatSeqAIJInvalidateDiagonal(A)); 2471 PetscFunctionReturn(0); 2472 } 2473 2474 PetscErrorCode MatCreateSubMatrix_SeqAIJ(Mat A, IS isrow, IS iscol, PetscInt csize, MatReuse scall, Mat *B) 2475 { 2476 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data, *c; 2477 PetscInt *smap, i, k, kstart, kend, oldcols = A->cmap->n, *lens; 2478 PetscInt row, mat_i, *mat_j, tcol, first, step, *mat_ilen, sum, lensi; 2479 const PetscInt *irow, *icol; 2480 const PetscScalar *aa; 2481 PetscInt nrows, ncols; 2482 PetscInt *starts, *j_new, *i_new, *aj = a->j, *ai = a->i, ii, *ailen = a->ilen; 2483 MatScalar *a_new, *mat_a; 2484 Mat C; 2485 PetscBool stride; 2486 2487 PetscFunctionBegin; 2488 PetscCall(ISGetIndices(isrow, &irow)); 2489 PetscCall(ISGetLocalSize(isrow, &nrows)); 2490 PetscCall(ISGetLocalSize(iscol, &ncols)); 2491 2492 PetscCall(PetscObjectTypeCompare((PetscObject)iscol, ISSTRIDE, &stride)); 2493 if (stride) { 2494 PetscCall(ISStrideGetInfo(iscol, &first, &step)); 2495 } else { 2496 first = 0; 2497 step = 0; 2498 } 2499 if (stride && step == 1) { 2500 /* special case of contiguous rows */ 2501 PetscCall(PetscMalloc2(nrows, &lens, nrows, &starts)); 2502 /* loop over new rows determining lens and starting points */ 2503 for (i = 0; i < nrows; i++) { 2504 kstart = ai[irow[i]]; 2505 kend = kstart + ailen[irow[i]]; 2506 starts[i] = kstart; 2507 for (k = kstart; k < kend; k++) { 2508 if (aj[k] >= first) { 2509 starts[i] = k; 2510 break; 2511 } 2512 } 2513 sum = 0; 2514 while (k < kend) { 2515 if (aj[k++] >= first + ncols) break; 2516 sum++; 2517 } 2518 lens[i] = sum; 2519 } 2520 /* create submatrix */ 2521 if (scall == MAT_REUSE_MATRIX) { 2522 PetscInt n_cols, n_rows; 2523 PetscCall(MatGetSize(*B, &n_rows, &n_cols)); 2524 PetscCheck(n_rows == nrows && n_cols == ncols, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Reused submatrix wrong size"); 2525 PetscCall(MatZeroEntries(*B)); 2526 C = *B; 2527 } else { 2528 PetscInt rbs, cbs; 2529 PetscCall(MatCreate(PetscObjectComm((PetscObject)A), &C)); 2530 PetscCall(MatSetSizes(C, nrows, ncols, PETSC_DETERMINE, PETSC_DETERMINE)); 2531 PetscCall(ISGetBlockSize(isrow, &rbs)); 2532 PetscCall(ISGetBlockSize(iscol, &cbs)); 2533 PetscCall(MatSetBlockSizes(C, rbs, cbs)); 2534 PetscCall(MatSetType(C, ((PetscObject)A)->type_name)); 2535 PetscCall(MatSeqAIJSetPreallocation_SeqAIJ(C, 0, lens)); 2536 } 2537 c = (Mat_SeqAIJ *)C->data; 2538 2539 /* loop over rows inserting into submatrix */ 2540 a_new = c->a; 2541 j_new = c->j; 2542 i_new = c->i; 2543 PetscCall(MatSeqAIJGetArrayRead(A, &aa)); 2544 for (i = 0; i < nrows; i++) { 2545 ii = starts[i]; 2546 lensi = lens[i]; 2547 for (k = 0; k < lensi; k++) *j_new++ = aj[ii + k] - first; 2548 PetscCall(PetscArraycpy(a_new, aa + starts[i], lensi)); 2549 a_new += lensi; 2550 i_new[i + 1] = i_new[i] + lensi; 2551 c->ilen[i] = lensi; 2552 } 2553 PetscCall(MatSeqAIJRestoreArrayRead(A, &aa)); 2554 PetscCall(PetscFree2(lens, starts)); 2555 } else { 2556 PetscCall(ISGetIndices(iscol, &icol)); 2557 PetscCall(PetscCalloc1(oldcols, &smap)); 2558 PetscCall(PetscMalloc1(1 + nrows, &lens)); 2559 for (i = 0; i < ncols; i++) { 2560 PetscCheck(icol[i] < oldcols, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Requesting column beyond largest column icol[%" PetscInt_FMT "] %" PetscInt_FMT " >= A->cmap->n %" PetscInt_FMT, i, icol[i], oldcols); 2561 smap[icol[i]] = i + 1; 2562 } 2563 2564 /* determine lens of each row */ 2565 for (i = 0; i < nrows; i++) { 2566 kstart = ai[irow[i]]; 2567 kend = kstart + a->ilen[irow[i]]; 2568 lens[i] = 0; 2569 for (k = kstart; k < kend; k++) { 2570 if (smap[aj[k]]) lens[i]++; 2571 } 2572 } 2573 /* Create and fill new matrix */ 2574 if (scall == MAT_REUSE_MATRIX) { 2575 PetscBool equal; 2576 2577 c = (Mat_SeqAIJ *)((*B)->data); 2578 PetscCheck((*B)->rmap->n == nrows && (*B)->cmap->n == ncols, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Cannot reuse matrix. wrong size"); 2579 PetscCall(PetscArraycmp(c->ilen, lens, (*B)->rmap->n, &equal)); 2580 PetscCheck(equal, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Cannot reuse matrix. wrong no of nonzeros"); 2581 PetscCall(PetscArrayzero(c->ilen, (*B)->rmap->n)); 2582 C = *B; 2583 } else { 2584 PetscInt rbs, cbs; 2585 PetscCall(MatCreate(PetscObjectComm((PetscObject)A), &C)); 2586 PetscCall(MatSetSizes(C, nrows, ncols, PETSC_DETERMINE, PETSC_DETERMINE)); 2587 PetscCall(ISGetBlockSize(isrow, &rbs)); 2588 PetscCall(ISGetBlockSize(iscol, &cbs)); 2589 PetscCall(MatSetBlockSizes(C, rbs, cbs)); 2590 PetscCall(MatSetType(C, ((PetscObject)A)->type_name)); 2591 PetscCall(MatSeqAIJSetPreallocation_SeqAIJ(C, 0, lens)); 2592 } 2593 PetscCall(MatSeqAIJGetArrayRead(A, &aa)); 2594 c = (Mat_SeqAIJ *)(C->data); 2595 for (i = 0; i < nrows; i++) { 2596 row = irow[i]; 2597 kstart = ai[row]; 2598 kend = kstart + a->ilen[row]; 2599 mat_i = c->i[i]; 2600 mat_j = c->j + mat_i; 2601 mat_a = c->a + mat_i; 2602 mat_ilen = c->ilen + i; 2603 for (k = kstart; k < kend; k++) { 2604 if ((tcol = smap[a->j[k]])) { 2605 *mat_j++ = tcol - 1; 2606 *mat_a++ = aa[k]; 2607 (*mat_ilen)++; 2608 } 2609 } 2610 } 2611 PetscCall(MatSeqAIJRestoreArrayRead(A, &aa)); 2612 /* Free work space */ 2613 PetscCall(ISRestoreIndices(iscol, &icol)); 2614 PetscCall(PetscFree(smap)); 2615 PetscCall(PetscFree(lens)); 2616 /* sort */ 2617 for (i = 0; i < nrows; i++) { 2618 PetscInt ilen; 2619 2620 mat_i = c->i[i]; 2621 mat_j = c->j + mat_i; 2622 mat_a = c->a + mat_i; 2623 ilen = c->ilen[i]; 2624 PetscCall(PetscSortIntWithScalarArray(ilen, mat_j, mat_a)); 2625 } 2626 } 2627 #if defined(PETSC_HAVE_DEVICE) 2628 PetscCall(MatBindToCPU(C, A->boundtocpu)); 2629 #endif 2630 PetscCall(MatAssemblyBegin(C, MAT_FINAL_ASSEMBLY)); 2631 PetscCall(MatAssemblyEnd(C, MAT_FINAL_ASSEMBLY)); 2632 2633 PetscCall(ISRestoreIndices(isrow, &irow)); 2634 *B = C; 2635 PetscFunctionReturn(0); 2636 } 2637 2638 PetscErrorCode MatGetMultiProcBlock_SeqAIJ(Mat mat, MPI_Comm subComm, MatReuse scall, Mat *subMat) 2639 { 2640 Mat B; 2641 2642 PetscFunctionBegin; 2643 if (scall == MAT_INITIAL_MATRIX) { 2644 PetscCall(MatCreate(subComm, &B)); 2645 PetscCall(MatSetSizes(B, mat->rmap->n, mat->cmap->n, mat->rmap->n, mat->cmap->n)); 2646 PetscCall(MatSetBlockSizesFromMats(B, mat, mat)); 2647 PetscCall(MatSetType(B, MATSEQAIJ)); 2648 PetscCall(MatDuplicateNoCreate_SeqAIJ(B, mat, MAT_COPY_VALUES, PETSC_TRUE)); 2649 *subMat = B; 2650 } else { 2651 PetscCall(MatCopy_SeqAIJ(mat, *subMat, SAME_NONZERO_PATTERN)); 2652 } 2653 PetscFunctionReturn(0); 2654 } 2655 2656 PetscErrorCode MatILUFactor_SeqAIJ(Mat inA, IS row, IS col, const MatFactorInfo *info) 2657 { 2658 Mat_SeqAIJ *a = (Mat_SeqAIJ *)inA->data; 2659 Mat outA; 2660 PetscBool row_identity, col_identity; 2661 2662 PetscFunctionBegin; 2663 PetscCheck(info->levels == 0, PETSC_COMM_SELF, PETSC_ERR_SUP, "Only levels=0 supported for in-place ilu"); 2664 2665 PetscCall(ISIdentity(row, &row_identity)); 2666 PetscCall(ISIdentity(col, &col_identity)); 2667 2668 outA = inA; 2669 outA->factortype = MAT_FACTOR_LU; 2670 PetscCall(PetscFree(inA->solvertype)); 2671 PetscCall(PetscStrallocpy(MATSOLVERPETSC, &inA->solvertype)); 2672 2673 PetscCall(PetscObjectReference((PetscObject)row)); 2674 PetscCall(ISDestroy(&a->row)); 2675 2676 a->row = row; 2677 2678 PetscCall(PetscObjectReference((PetscObject)col)); 2679 PetscCall(ISDestroy(&a->col)); 2680 2681 a->col = col; 2682 2683 /* Create the inverse permutation so that it can be used in MatLUFactorNumeric() */ 2684 PetscCall(ISDestroy(&a->icol)); 2685 PetscCall(ISInvertPermutation(col, PETSC_DECIDE, &a->icol)); 2686 2687 if (!a->solve_work) { /* this matrix may have been factored before */ 2688 PetscCall(PetscMalloc1(inA->rmap->n + 1, &a->solve_work)); 2689 } 2690 2691 PetscCall(MatMarkDiagonal_SeqAIJ(inA)); 2692 if (row_identity && col_identity) { 2693 PetscCall(MatLUFactorNumeric_SeqAIJ_inplace(outA, inA, info)); 2694 } else { 2695 PetscCall(MatLUFactorNumeric_SeqAIJ_InplaceWithPerm(outA, inA, info)); 2696 } 2697 PetscFunctionReturn(0); 2698 } 2699 2700 PetscErrorCode MatScale_SeqAIJ(Mat inA, PetscScalar alpha) 2701 { 2702 Mat_SeqAIJ *a = (Mat_SeqAIJ *)inA->data; 2703 PetscScalar *v; 2704 PetscBLASInt one = 1, bnz; 2705 2706 PetscFunctionBegin; 2707 PetscCall(MatSeqAIJGetArray(inA, &v)); 2708 PetscCall(PetscBLASIntCast(a->nz, &bnz)); 2709 PetscCallBLAS("BLASscal", BLASscal_(&bnz, &alpha, v, &one)); 2710 PetscCall(PetscLogFlops(a->nz)); 2711 PetscCall(MatSeqAIJRestoreArray(inA, &v)); 2712 PetscCall(MatSeqAIJInvalidateDiagonal(inA)); 2713 PetscFunctionReturn(0); 2714 } 2715 2716 PetscErrorCode MatDestroySubMatrix_Private(Mat_SubSppt *submatj) 2717 { 2718 PetscInt i; 2719 2720 PetscFunctionBegin; 2721 if (!submatj->id) { /* delete data that are linked only to submats[id=0] */ 2722 PetscCall(PetscFree4(submatj->sbuf1, submatj->ptr, submatj->tmp, submatj->ctr)); 2723 2724 for (i = 0; i < submatj->nrqr; ++i) PetscCall(PetscFree(submatj->sbuf2[i])); 2725 PetscCall(PetscFree3(submatj->sbuf2, submatj->req_size, submatj->req_source1)); 2726 2727 if (submatj->rbuf1) { 2728 PetscCall(PetscFree(submatj->rbuf1[0])); 2729 PetscCall(PetscFree(submatj->rbuf1)); 2730 } 2731 2732 for (i = 0; i < submatj->nrqs; ++i) PetscCall(PetscFree(submatj->rbuf3[i])); 2733 PetscCall(PetscFree3(submatj->req_source2, submatj->rbuf2, submatj->rbuf3)); 2734 PetscCall(PetscFree(submatj->pa)); 2735 } 2736 2737 #if defined(PETSC_USE_CTABLE) 2738 PetscCall(PetscHMapIDestroy(&submatj->rmap)); 2739 if (submatj->cmap_loc) PetscCall(PetscFree(submatj->cmap_loc)); 2740 PetscCall(PetscFree(submatj->rmap_loc)); 2741 #else 2742 PetscCall(PetscFree(submatj->rmap)); 2743 #endif 2744 2745 if (!submatj->allcolumns) { 2746 #if defined(PETSC_USE_CTABLE) 2747 PetscCall(PetscHMapIDestroy((PetscHMapI *)&submatj->cmap)); 2748 #else 2749 PetscCall(PetscFree(submatj->cmap)); 2750 #endif 2751 } 2752 PetscCall(PetscFree(submatj->row2proc)); 2753 2754 PetscCall(PetscFree(submatj)); 2755 PetscFunctionReturn(0); 2756 } 2757 2758 PetscErrorCode MatDestroySubMatrix_SeqAIJ(Mat C) 2759 { 2760 Mat_SeqAIJ *c = (Mat_SeqAIJ *)C->data; 2761 Mat_SubSppt *submatj = c->submatis1; 2762 2763 PetscFunctionBegin; 2764 PetscCall((*submatj->destroy)(C)); 2765 PetscCall(MatDestroySubMatrix_Private(submatj)); 2766 PetscFunctionReturn(0); 2767 } 2768 2769 /* Note this has code duplication with MatDestroySubMatrices_SeqBAIJ() */ 2770 PetscErrorCode MatDestroySubMatrices_SeqAIJ(PetscInt n, Mat *mat[]) 2771 { 2772 PetscInt i; 2773 Mat C; 2774 Mat_SeqAIJ *c; 2775 Mat_SubSppt *submatj; 2776 2777 PetscFunctionBegin; 2778 for (i = 0; i < n; i++) { 2779 C = (*mat)[i]; 2780 c = (Mat_SeqAIJ *)C->data; 2781 submatj = c->submatis1; 2782 if (submatj) { 2783 if (--((PetscObject)C)->refct <= 0) { 2784 PetscCall(PetscFree(C->factorprefix)); 2785 PetscCall((*submatj->destroy)(C)); 2786 PetscCall(MatDestroySubMatrix_Private(submatj)); 2787 PetscCall(PetscFree(C->defaultvectype)); 2788 PetscCall(PetscFree(C->defaultrandtype)); 2789 PetscCall(PetscLayoutDestroy(&C->rmap)); 2790 PetscCall(PetscLayoutDestroy(&C->cmap)); 2791 PetscCall(PetscHeaderDestroy(&C)); 2792 } 2793 } else { 2794 PetscCall(MatDestroy(&C)); 2795 } 2796 } 2797 2798 /* Destroy Dummy submatrices created for reuse */ 2799 PetscCall(MatDestroySubMatrices_Dummy(n, mat)); 2800 2801 PetscCall(PetscFree(*mat)); 2802 PetscFunctionReturn(0); 2803 } 2804 2805 PetscErrorCode MatCreateSubMatrices_SeqAIJ(Mat A, PetscInt n, const IS irow[], const IS icol[], MatReuse scall, Mat *B[]) 2806 { 2807 PetscInt i; 2808 2809 PetscFunctionBegin; 2810 if (scall == MAT_INITIAL_MATRIX) PetscCall(PetscCalloc1(n + 1, B)); 2811 2812 for (i = 0; i < n; i++) PetscCall(MatCreateSubMatrix_SeqAIJ(A, irow[i], icol[i], PETSC_DECIDE, scall, &(*B)[i])); 2813 PetscFunctionReturn(0); 2814 } 2815 2816 PetscErrorCode MatIncreaseOverlap_SeqAIJ(Mat A, PetscInt is_max, IS is[], PetscInt ov) 2817 { 2818 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 2819 PetscInt row, i, j, k, l, ll, m, n, *nidx, isz, val; 2820 const PetscInt *idx; 2821 PetscInt start, end, *ai, *aj, bs = (A->rmap->bs > 0 && A->rmap->bs == A->cmap->bs) ? A->rmap->bs : 1; 2822 PetscBT table; 2823 2824 PetscFunctionBegin; 2825 m = A->rmap->n / bs; 2826 ai = a->i; 2827 aj = a->j; 2828 2829 PetscCheck(ov >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "illegal negative overlap value used"); 2830 2831 PetscCall(PetscMalloc1(m + 1, &nidx)); 2832 PetscCall(PetscBTCreate(m, &table)); 2833 2834 for (i = 0; i < is_max; i++) { 2835 /* Initialize the two local arrays */ 2836 isz = 0; 2837 PetscCall(PetscBTMemzero(m, table)); 2838 2839 /* Extract the indices, assume there can be duplicate entries */ 2840 PetscCall(ISGetIndices(is[i], &idx)); 2841 PetscCall(ISGetLocalSize(is[i], &n)); 2842 2843 if (bs > 1) { 2844 /* Enter these into the temp arrays. I.e., mark table[row], enter row into new index */ 2845 for (j = 0; j < n; ++j) { 2846 if (!PetscBTLookupSet(table, idx[j] / bs)) nidx[isz++] = idx[j] / bs; 2847 } 2848 PetscCall(ISRestoreIndices(is[i], &idx)); 2849 PetscCall(ISDestroy(&is[i])); 2850 2851 k = 0; 2852 for (j = 0; j < ov; j++) { /* for each overlap */ 2853 n = isz; 2854 for (; k < n; k++) { /* do only those rows in nidx[k], which are not done yet */ 2855 for (ll = 0; ll < bs; ll++) { 2856 row = bs * nidx[k] + ll; 2857 start = ai[row]; 2858 end = ai[row + 1]; 2859 for (l = start; l < end; l++) { 2860 val = aj[l] / bs; 2861 if (!PetscBTLookupSet(table, val)) nidx[isz++] = val; 2862 } 2863 } 2864 } 2865 } 2866 PetscCall(ISCreateBlock(PETSC_COMM_SELF, bs, isz, nidx, PETSC_COPY_VALUES, (is + i))); 2867 } else { 2868 /* Enter these into the temp arrays. I.e., mark table[row], enter row into new index */ 2869 for (j = 0; j < n; ++j) { 2870 if (!PetscBTLookupSet(table, idx[j])) nidx[isz++] = idx[j]; 2871 } 2872 PetscCall(ISRestoreIndices(is[i], &idx)); 2873 PetscCall(ISDestroy(&is[i])); 2874 2875 k = 0; 2876 for (j = 0; j < ov; j++) { /* for each overlap */ 2877 n = isz; 2878 for (; k < n; k++) { /* do only those rows in nidx[k], which are not done yet */ 2879 row = nidx[k]; 2880 start = ai[row]; 2881 end = ai[row + 1]; 2882 for (l = start; l < end; l++) { 2883 val = aj[l]; 2884 if (!PetscBTLookupSet(table, val)) nidx[isz++] = val; 2885 } 2886 } 2887 } 2888 PetscCall(ISCreateGeneral(PETSC_COMM_SELF, isz, nidx, PETSC_COPY_VALUES, (is + i))); 2889 } 2890 } 2891 PetscCall(PetscBTDestroy(&table)); 2892 PetscCall(PetscFree(nidx)); 2893 PetscFunctionReturn(0); 2894 } 2895 2896 /* -------------------------------------------------------------- */ 2897 PetscErrorCode MatPermute_SeqAIJ(Mat A, IS rowp, IS colp, Mat *B) 2898 { 2899 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 2900 PetscInt i, nz = 0, m = A->rmap->n, n = A->cmap->n; 2901 const PetscInt *row, *col; 2902 PetscInt *cnew, j, *lens; 2903 IS icolp, irowp; 2904 PetscInt *cwork = NULL; 2905 PetscScalar *vwork = NULL; 2906 2907 PetscFunctionBegin; 2908 PetscCall(ISInvertPermutation(rowp, PETSC_DECIDE, &irowp)); 2909 PetscCall(ISGetIndices(irowp, &row)); 2910 PetscCall(ISInvertPermutation(colp, PETSC_DECIDE, &icolp)); 2911 PetscCall(ISGetIndices(icolp, &col)); 2912 2913 /* determine lengths of permuted rows */ 2914 PetscCall(PetscMalloc1(m + 1, &lens)); 2915 for (i = 0; i < m; i++) lens[row[i]] = a->i[i + 1] - a->i[i]; 2916 PetscCall(MatCreate(PetscObjectComm((PetscObject)A), B)); 2917 PetscCall(MatSetSizes(*B, m, n, m, n)); 2918 PetscCall(MatSetBlockSizesFromMats(*B, A, A)); 2919 PetscCall(MatSetType(*B, ((PetscObject)A)->type_name)); 2920 PetscCall(MatSeqAIJSetPreallocation_SeqAIJ(*B, 0, lens)); 2921 PetscCall(PetscFree(lens)); 2922 2923 PetscCall(PetscMalloc1(n, &cnew)); 2924 for (i = 0; i < m; i++) { 2925 PetscCall(MatGetRow_SeqAIJ(A, i, &nz, &cwork, &vwork)); 2926 for (j = 0; j < nz; j++) cnew[j] = col[cwork[j]]; 2927 PetscCall(MatSetValues_SeqAIJ(*B, 1, &row[i], nz, cnew, vwork, INSERT_VALUES)); 2928 PetscCall(MatRestoreRow_SeqAIJ(A, i, &nz, &cwork, &vwork)); 2929 } 2930 PetscCall(PetscFree(cnew)); 2931 2932 (*B)->assembled = PETSC_FALSE; 2933 2934 #if defined(PETSC_HAVE_DEVICE) 2935 PetscCall(MatBindToCPU(*B, A->boundtocpu)); 2936 #endif 2937 PetscCall(MatAssemblyBegin(*B, MAT_FINAL_ASSEMBLY)); 2938 PetscCall(MatAssemblyEnd(*B, MAT_FINAL_ASSEMBLY)); 2939 PetscCall(ISRestoreIndices(irowp, &row)); 2940 PetscCall(ISRestoreIndices(icolp, &col)); 2941 PetscCall(ISDestroy(&irowp)); 2942 PetscCall(ISDestroy(&icolp)); 2943 if (rowp == colp) PetscCall(MatPropagateSymmetryOptions(A, *B)); 2944 PetscFunctionReturn(0); 2945 } 2946 2947 PetscErrorCode MatCopy_SeqAIJ(Mat A, Mat B, MatStructure str) 2948 { 2949 PetscFunctionBegin; 2950 /* If the two matrices have the same copy implementation, use fast copy. */ 2951 if (str == SAME_NONZERO_PATTERN && (A->ops->copy == B->ops->copy)) { 2952 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 2953 Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data; 2954 const PetscScalar *aa; 2955 2956 PetscCall(MatSeqAIJGetArrayRead(A, &aa)); 2957 PetscCheck(a->i[A->rmap->n] == b->i[B->rmap->n], PETSC_COMM_SELF, PETSC_ERR_ARG_INCOMP, "Number of nonzeros in two matrices are different %" PetscInt_FMT " != %" PetscInt_FMT, a->i[A->rmap->n], b->i[B->rmap->n]); 2958 PetscCall(PetscArraycpy(b->a, aa, a->i[A->rmap->n])); 2959 PetscCall(PetscObjectStateIncrease((PetscObject)B)); 2960 PetscCall(MatSeqAIJRestoreArrayRead(A, &aa)); 2961 } else { 2962 PetscCall(MatCopy_Basic(A, B, str)); 2963 } 2964 PetscFunctionReturn(0); 2965 } 2966 2967 PetscErrorCode MatSetUp_SeqAIJ(Mat A) 2968 { 2969 PetscFunctionBegin; 2970 PetscCall(MatSeqAIJSetPreallocation_SeqAIJ(A, PETSC_DEFAULT, NULL)); 2971 PetscFunctionReturn(0); 2972 } 2973 2974 PETSC_INTERN PetscErrorCode MatSeqAIJGetArray_SeqAIJ(Mat A, PetscScalar *array[]) 2975 { 2976 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 2977 2978 PetscFunctionBegin; 2979 *array = a->a; 2980 PetscFunctionReturn(0); 2981 } 2982 2983 PETSC_INTERN PetscErrorCode MatSeqAIJRestoreArray_SeqAIJ(Mat A, PetscScalar *array[]) 2984 { 2985 PetscFunctionBegin; 2986 *array = NULL; 2987 PetscFunctionReturn(0); 2988 } 2989 2990 /* 2991 Computes the number of nonzeros per row needed for preallocation when X and Y 2992 have different nonzero structure. 2993 */ 2994 PetscErrorCode MatAXPYGetPreallocation_SeqX_private(PetscInt m, const PetscInt *xi, const PetscInt *xj, const PetscInt *yi, const PetscInt *yj, PetscInt *nnz) 2995 { 2996 PetscInt i, j, k, nzx, nzy; 2997 2998 PetscFunctionBegin; 2999 /* Set the number of nonzeros in the new matrix */ 3000 for (i = 0; i < m; i++) { 3001 const PetscInt *xjj = xj + xi[i], *yjj = yj + yi[i]; 3002 nzx = xi[i + 1] - xi[i]; 3003 nzy = yi[i + 1] - yi[i]; 3004 nnz[i] = 0; 3005 for (j = 0, k = 0; j < nzx; j++) { /* Point in X */ 3006 for (; k < nzy && yjj[k] < xjj[j]; k++) nnz[i]++; /* Catch up to X */ 3007 if (k < nzy && yjj[k] == xjj[j]) k++; /* Skip duplicate */ 3008 nnz[i]++; 3009 } 3010 for (; k < nzy; k++) nnz[i]++; 3011 } 3012 PetscFunctionReturn(0); 3013 } 3014 3015 PetscErrorCode MatAXPYGetPreallocation_SeqAIJ(Mat Y, Mat X, PetscInt *nnz) 3016 { 3017 PetscInt m = Y->rmap->N; 3018 Mat_SeqAIJ *x = (Mat_SeqAIJ *)X->data; 3019 Mat_SeqAIJ *y = (Mat_SeqAIJ *)Y->data; 3020 3021 PetscFunctionBegin; 3022 /* Set the number of nonzeros in the new matrix */ 3023 PetscCall(MatAXPYGetPreallocation_SeqX_private(m, x->i, x->j, y->i, y->j, nnz)); 3024 PetscFunctionReturn(0); 3025 } 3026 3027 PetscErrorCode MatAXPY_SeqAIJ(Mat Y, PetscScalar a, Mat X, MatStructure str) 3028 { 3029 Mat_SeqAIJ *x = (Mat_SeqAIJ *)X->data, *y = (Mat_SeqAIJ *)Y->data; 3030 3031 PetscFunctionBegin; 3032 if (str == UNKNOWN_NONZERO_PATTERN || (PetscDefined(USE_DEBUG) && str == SAME_NONZERO_PATTERN)) { 3033 PetscBool e = x->nz == y->nz ? PETSC_TRUE : PETSC_FALSE; 3034 if (e) { 3035 PetscCall(PetscArraycmp(x->i, y->i, Y->rmap->n + 1, &e)); 3036 if (e) { 3037 PetscCall(PetscArraycmp(x->j, y->j, y->nz, &e)); 3038 if (e) str = SAME_NONZERO_PATTERN; 3039 } 3040 } 3041 if (!e) PetscCheck(str != SAME_NONZERO_PATTERN, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "MatStructure is not SAME_NONZERO_PATTERN"); 3042 } 3043 if (str == SAME_NONZERO_PATTERN) { 3044 const PetscScalar *xa; 3045 PetscScalar *ya, alpha = a; 3046 PetscBLASInt one = 1, bnz; 3047 3048 PetscCall(PetscBLASIntCast(x->nz, &bnz)); 3049 PetscCall(MatSeqAIJGetArray(Y, &ya)); 3050 PetscCall(MatSeqAIJGetArrayRead(X, &xa)); 3051 PetscCallBLAS("BLASaxpy", BLASaxpy_(&bnz, &alpha, xa, &one, ya, &one)); 3052 PetscCall(MatSeqAIJRestoreArrayRead(X, &xa)); 3053 PetscCall(MatSeqAIJRestoreArray(Y, &ya)); 3054 PetscCall(PetscLogFlops(2.0 * bnz)); 3055 PetscCall(MatSeqAIJInvalidateDiagonal(Y)); 3056 PetscCall(PetscObjectStateIncrease((PetscObject)Y)); 3057 } else if (str == SUBSET_NONZERO_PATTERN) { /* nonzeros of X is a subset of Y's */ 3058 PetscCall(MatAXPY_Basic(Y, a, X, str)); 3059 } else { 3060 Mat B; 3061 PetscInt *nnz; 3062 PetscCall(PetscMalloc1(Y->rmap->N, &nnz)); 3063 PetscCall(MatCreate(PetscObjectComm((PetscObject)Y), &B)); 3064 PetscCall(PetscObjectSetName((PetscObject)B, ((PetscObject)Y)->name)); 3065 PetscCall(MatSetLayouts(B, Y->rmap, Y->cmap)); 3066 PetscCall(MatSetType(B, ((PetscObject)Y)->type_name)); 3067 PetscCall(MatAXPYGetPreallocation_SeqAIJ(Y, X, nnz)); 3068 PetscCall(MatSeqAIJSetPreallocation(B, 0, nnz)); 3069 PetscCall(MatAXPY_BasicWithPreallocation(B, Y, a, X, str)); 3070 PetscCall(MatHeaderMerge(Y, &B)); 3071 PetscCall(MatSeqAIJCheckInode(Y)); 3072 PetscCall(PetscFree(nnz)); 3073 } 3074 PetscFunctionReturn(0); 3075 } 3076 3077 PETSC_INTERN PetscErrorCode MatConjugate_SeqAIJ(Mat mat) 3078 { 3079 #if defined(PETSC_USE_COMPLEX) 3080 Mat_SeqAIJ *aij = (Mat_SeqAIJ *)mat->data; 3081 PetscInt i, nz; 3082 PetscScalar *a; 3083 3084 PetscFunctionBegin; 3085 nz = aij->nz; 3086 PetscCall(MatSeqAIJGetArray(mat, &a)); 3087 for (i = 0; i < nz; i++) a[i] = PetscConj(a[i]); 3088 PetscCall(MatSeqAIJRestoreArray(mat, &a)); 3089 #else 3090 PetscFunctionBegin; 3091 #endif 3092 PetscFunctionReturn(0); 3093 } 3094 3095 PetscErrorCode MatGetRowMaxAbs_SeqAIJ(Mat A, Vec v, PetscInt idx[]) 3096 { 3097 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 3098 PetscInt i, j, m = A->rmap->n, *ai, *aj, ncols, n; 3099 PetscReal atmp; 3100 PetscScalar *x; 3101 const MatScalar *aa, *av; 3102 3103 PetscFunctionBegin; 3104 PetscCheck(!A->factortype, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Not for factored matrix"); 3105 PetscCall(MatSeqAIJGetArrayRead(A, &av)); 3106 aa = av; 3107 ai = a->i; 3108 aj = a->j; 3109 3110 PetscCall(VecSet(v, 0.0)); 3111 PetscCall(VecGetArrayWrite(v, &x)); 3112 PetscCall(VecGetLocalSize(v, &n)); 3113 PetscCheck(n == A->rmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Nonconforming matrix and vector"); 3114 for (i = 0; i < m; i++) { 3115 ncols = ai[1] - ai[0]; 3116 ai++; 3117 for (j = 0; j < ncols; j++) { 3118 atmp = PetscAbsScalar(*aa); 3119 if (PetscAbsScalar(x[i]) < atmp) { 3120 x[i] = atmp; 3121 if (idx) idx[i] = *aj; 3122 } 3123 aa++; 3124 aj++; 3125 } 3126 } 3127 PetscCall(VecRestoreArrayWrite(v, &x)); 3128 PetscCall(MatSeqAIJRestoreArrayRead(A, &av)); 3129 PetscFunctionReturn(0); 3130 } 3131 3132 PetscErrorCode MatGetRowMax_SeqAIJ(Mat A, Vec v, PetscInt idx[]) 3133 { 3134 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 3135 PetscInt i, j, m = A->rmap->n, *ai, *aj, ncols, n; 3136 PetscScalar *x; 3137 const MatScalar *aa, *av; 3138 3139 PetscFunctionBegin; 3140 PetscCheck(!A->factortype, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Not for factored matrix"); 3141 PetscCall(MatSeqAIJGetArrayRead(A, &av)); 3142 aa = av; 3143 ai = a->i; 3144 aj = a->j; 3145 3146 PetscCall(VecSet(v, 0.0)); 3147 PetscCall(VecGetArrayWrite(v, &x)); 3148 PetscCall(VecGetLocalSize(v, &n)); 3149 PetscCheck(n == A->rmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Nonconforming matrix and vector"); 3150 for (i = 0; i < m; i++) { 3151 ncols = ai[1] - ai[0]; 3152 ai++; 3153 if (ncols == A->cmap->n) { /* row is dense */ 3154 x[i] = *aa; 3155 if (idx) idx[i] = 0; 3156 } else { /* row is sparse so already KNOW maximum is 0.0 or higher */ 3157 x[i] = 0.0; 3158 if (idx) { 3159 for (j = 0; j < ncols; j++) { /* find first implicit 0.0 in the row */ 3160 if (aj[j] > j) { 3161 idx[i] = j; 3162 break; 3163 } 3164 } 3165 /* in case first implicit 0.0 in the row occurs at ncols-th column */ 3166 if (j == ncols && j < A->cmap->n) idx[i] = j; 3167 } 3168 } 3169 for (j = 0; j < ncols; j++) { 3170 if (PetscRealPart(x[i]) < PetscRealPart(*aa)) { 3171 x[i] = *aa; 3172 if (idx) idx[i] = *aj; 3173 } 3174 aa++; 3175 aj++; 3176 } 3177 } 3178 PetscCall(VecRestoreArrayWrite(v, &x)); 3179 PetscCall(MatSeqAIJRestoreArrayRead(A, &av)); 3180 PetscFunctionReturn(0); 3181 } 3182 3183 PetscErrorCode MatGetRowMinAbs_SeqAIJ(Mat A, Vec v, PetscInt idx[]) 3184 { 3185 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 3186 PetscInt i, j, m = A->rmap->n, *ai, *aj, ncols, n; 3187 PetscScalar *x; 3188 const MatScalar *aa, *av; 3189 3190 PetscFunctionBegin; 3191 PetscCall(MatSeqAIJGetArrayRead(A, &av)); 3192 aa = av; 3193 ai = a->i; 3194 aj = a->j; 3195 3196 PetscCall(VecSet(v, 0.0)); 3197 PetscCall(VecGetArrayWrite(v, &x)); 3198 PetscCall(VecGetLocalSize(v, &n)); 3199 PetscCheck(n == m, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Nonconforming matrix and vector, %" PetscInt_FMT " vs. %" PetscInt_FMT " rows", m, n); 3200 for (i = 0; i < m; i++) { 3201 ncols = ai[1] - ai[0]; 3202 ai++; 3203 if (ncols == A->cmap->n) { /* row is dense */ 3204 x[i] = *aa; 3205 if (idx) idx[i] = 0; 3206 } else { /* row is sparse so already KNOW minimum is 0.0 or higher */ 3207 x[i] = 0.0; 3208 if (idx) { /* find first implicit 0.0 in the row */ 3209 for (j = 0; j < ncols; j++) { 3210 if (aj[j] > j) { 3211 idx[i] = j; 3212 break; 3213 } 3214 } 3215 /* in case first implicit 0.0 in the row occurs at ncols-th column */ 3216 if (j == ncols && j < A->cmap->n) idx[i] = j; 3217 } 3218 } 3219 for (j = 0; j < ncols; j++) { 3220 if (PetscAbsScalar(x[i]) > PetscAbsScalar(*aa)) { 3221 x[i] = *aa; 3222 if (idx) idx[i] = *aj; 3223 } 3224 aa++; 3225 aj++; 3226 } 3227 } 3228 PetscCall(VecRestoreArrayWrite(v, &x)); 3229 PetscCall(MatSeqAIJRestoreArrayRead(A, &av)); 3230 PetscFunctionReturn(0); 3231 } 3232 3233 PetscErrorCode MatGetRowMin_SeqAIJ(Mat A, Vec v, PetscInt idx[]) 3234 { 3235 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 3236 PetscInt i, j, m = A->rmap->n, ncols, n; 3237 const PetscInt *ai, *aj; 3238 PetscScalar *x; 3239 const MatScalar *aa, *av; 3240 3241 PetscFunctionBegin; 3242 PetscCheck(!A->factortype, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Not for factored matrix"); 3243 PetscCall(MatSeqAIJGetArrayRead(A, &av)); 3244 aa = av; 3245 ai = a->i; 3246 aj = a->j; 3247 3248 PetscCall(VecSet(v, 0.0)); 3249 PetscCall(VecGetArrayWrite(v, &x)); 3250 PetscCall(VecGetLocalSize(v, &n)); 3251 PetscCheck(n == m, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Nonconforming matrix and vector"); 3252 for (i = 0; i < m; i++) { 3253 ncols = ai[1] - ai[0]; 3254 ai++; 3255 if (ncols == A->cmap->n) { /* row is dense */ 3256 x[i] = *aa; 3257 if (idx) idx[i] = 0; 3258 } else { /* row is sparse so already KNOW minimum is 0.0 or lower */ 3259 x[i] = 0.0; 3260 if (idx) { /* find first implicit 0.0 in the row */ 3261 for (j = 0; j < ncols; j++) { 3262 if (aj[j] > j) { 3263 idx[i] = j; 3264 break; 3265 } 3266 } 3267 /* in case first implicit 0.0 in the row occurs at ncols-th column */ 3268 if (j == ncols && j < A->cmap->n) idx[i] = j; 3269 } 3270 } 3271 for (j = 0; j < ncols; j++) { 3272 if (PetscRealPart(x[i]) > PetscRealPart(*aa)) { 3273 x[i] = *aa; 3274 if (idx) idx[i] = *aj; 3275 } 3276 aa++; 3277 aj++; 3278 } 3279 } 3280 PetscCall(VecRestoreArrayWrite(v, &x)); 3281 PetscCall(MatSeqAIJRestoreArrayRead(A, &av)); 3282 PetscFunctionReturn(0); 3283 } 3284 3285 PetscErrorCode MatInvertBlockDiagonal_SeqAIJ(Mat A, const PetscScalar **values) 3286 { 3287 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 3288 PetscInt i, bs = PetscAbs(A->rmap->bs), mbs = A->rmap->n / bs, ipvt[5], bs2 = bs * bs, *v_pivots, ij[7], *IJ, j; 3289 MatScalar *diag, work[25], *v_work; 3290 const PetscReal shift = 0.0; 3291 PetscBool allowzeropivot, zeropivotdetected = PETSC_FALSE; 3292 3293 PetscFunctionBegin; 3294 allowzeropivot = PetscNot(A->erroriffailure); 3295 if (a->ibdiagvalid) { 3296 if (values) *values = a->ibdiag; 3297 PetscFunctionReturn(0); 3298 } 3299 PetscCall(MatMarkDiagonal_SeqAIJ(A)); 3300 if (!a->ibdiag) { PetscCall(PetscMalloc1(bs2 * mbs, &a->ibdiag)); } 3301 diag = a->ibdiag; 3302 if (values) *values = a->ibdiag; 3303 /* factor and invert each block */ 3304 switch (bs) { 3305 case 1: 3306 for (i = 0; i < mbs; i++) { 3307 PetscCall(MatGetValues(A, 1, &i, 1, &i, diag + i)); 3308 if (PetscAbsScalar(diag[i] + shift) < PETSC_MACHINE_EPSILON) { 3309 if (allowzeropivot) { 3310 A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT; 3311 A->factorerror_zeropivot_value = PetscAbsScalar(diag[i]); 3312 A->factorerror_zeropivot_row = i; 3313 PetscCall(PetscInfo(A, "Zero pivot, row %" PetscInt_FMT " pivot %g tolerance %g\n", i, (double)PetscAbsScalar(diag[i]), (double)PETSC_MACHINE_EPSILON)); 3314 } else SETERRQ(PETSC_COMM_SELF, PETSC_ERR_MAT_LU_ZRPVT, "Zero pivot, row %" PetscInt_FMT " pivot %g tolerance %g", i, (double)PetscAbsScalar(diag[i]), (double)PETSC_MACHINE_EPSILON); 3315 } 3316 diag[i] = (PetscScalar)1.0 / (diag[i] + shift); 3317 } 3318 break; 3319 case 2: 3320 for (i = 0; i < mbs; i++) { 3321 ij[0] = 2 * i; 3322 ij[1] = 2 * i + 1; 3323 PetscCall(MatGetValues(A, 2, ij, 2, ij, diag)); 3324 PetscCall(PetscKernel_A_gets_inverse_A_2(diag, shift, allowzeropivot, &zeropivotdetected)); 3325 if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT; 3326 PetscCall(PetscKernel_A_gets_transpose_A_2(diag)); 3327 diag += 4; 3328 } 3329 break; 3330 case 3: 3331 for (i = 0; i < mbs; i++) { 3332 ij[0] = 3 * i; 3333 ij[1] = 3 * i + 1; 3334 ij[2] = 3 * i + 2; 3335 PetscCall(MatGetValues(A, 3, ij, 3, ij, diag)); 3336 PetscCall(PetscKernel_A_gets_inverse_A_3(diag, shift, allowzeropivot, &zeropivotdetected)); 3337 if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT; 3338 PetscCall(PetscKernel_A_gets_transpose_A_3(diag)); 3339 diag += 9; 3340 } 3341 break; 3342 case 4: 3343 for (i = 0; i < mbs; i++) { 3344 ij[0] = 4 * i; 3345 ij[1] = 4 * i + 1; 3346 ij[2] = 4 * i + 2; 3347 ij[3] = 4 * i + 3; 3348 PetscCall(MatGetValues(A, 4, ij, 4, ij, diag)); 3349 PetscCall(PetscKernel_A_gets_inverse_A_4(diag, shift, allowzeropivot, &zeropivotdetected)); 3350 if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT; 3351 PetscCall(PetscKernel_A_gets_transpose_A_4(diag)); 3352 diag += 16; 3353 } 3354 break; 3355 case 5: 3356 for (i = 0; i < mbs; i++) { 3357 ij[0] = 5 * i; 3358 ij[1] = 5 * i + 1; 3359 ij[2] = 5 * i + 2; 3360 ij[3] = 5 * i + 3; 3361 ij[4] = 5 * i + 4; 3362 PetscCall(MatGetValues(A, 5, ij, 5, ij, diag)); 3363 PetscCall(PetscKernel_A_gets_inverse_A_5(diag, ipvt, work, shift, allowzeropivot, &zeropivotdetected)); 3364 if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT; 3365 PetscCall(PetscKernel_A_gets_transpose_A_5(diag)); 3366 diag += 25; 3367 } 3368 break; 3369 case 6: 3370 for (i = 0; i < mbs; i++) { 3371 ij[0] = 6 * i; 3372 ij[1] = 6 * i + 1; 3373 ij[2] = 6 * i + 2; 3374 ij[3] = 6 * i + 3; 3375 ij[4] = 6 * i + 4; 3376 ij[5] = 6 * i + 5; 3377 PetscCall(MatGetValues(A, 6, ij, 6, ij, diag)); 3378 PetscCall(PetscKernel_A_gets_inverse_A_6(diag, shift, allowzeropivot, &zeropivotdetected)); 3379 if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT; 3380 PetscCall(PetscKernel_A_gets_transpose_A_6(diag)); 3381 diag += 36; 3382 } 3383 break; 3384 case 7: 3385 for (i = 0; i < mbs; i++) { 3386 ij[0] = 7 * i; 3387 ij[1] = 7 * i + 1; 3388 ij[2] = 7 * i + 2; 3389 ij[3] = 7 * i + 3; 3390 ij[4] = 7 * i + 4; 3391 ij[5] = 7 * i + 5; 3392 ij[6] = 7 * i + 6; 3393 PetscCall(MatGetValues(A, 7, ij, 7, ij, diag)); 3394 PetscCall(PetscKernel_A_gets_inverse_A_7(diag, shift, allowzeropivot, &zeropivotdetected)); 3395 if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT; 3396 PetscCall(PetscKernel_A_gets_transpose_A_7(diag)); 3397 diag += 49; 3398 } 3399 break; 3400 default: 3401 PetscCall(PetscMalloc3(bs, &v_work, bs, &v_pivots, bs, &IJ)); 3402 for (i = 0; i < mbs; i++) { 3403 for (j = 0; j < bs; j++) IJ[j] = bs * i + j; 3404 PetscCall(MatGetValues(A, bs, IJ, bs, IJ, diag)); 3405 PetscCall(PetscKernel_A_gets_inverse_A(bs, diag, v_pivots, v_work, allowzeropivot, &zeropivotdetected)); 3406 if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT; 3407 PetscCall(PetscKernel_A_gets_transpose_A_N(diag, bs)); 3408 diag += bs2; 3409 } 3410 PetscCall(PetscFree3(v_work, v_pivots, IJ)); 3411 } 3412 a->ibdiagvalid = PETSC_TRUE; 3413 PetscFunctionReturn(0); 3414 } 3415 3416 static PetscErrorCode MatSetRandom_SeqAIJ(Mat x, PetscRandom rctx) 3417 { 3418 Mat_SeqAIJ *aij = (Mat_SeqAIJ *)x->data; 3419 PetscScalar a, *aa; 3420 PetscInt m, n, i, j, col; 3421 3422 PetscFunctionBegin; 3423 if (!x->assembled) { 3424 PetscCall(MatGetSize(x, &m, &n)); 3425 for (i = 0; i < m; i++) { 3426 for (j = 0; j < aij->imax[i]; j++) { 3427 PetscCall(PetscRandomGetValue(rctx, &a)); 3428 col = (PetscInt)(n * PetscRealPart(a)); 3429 PetscCall(MatSetValues(x, 1, &i, 1, &col, &a, ADD_VALUES)); 3430 } 3431 } 3432 } else { 3433 PetscCall(MatSeqAIJGetArrayWrite(x, &aa)); 3434 for (i = 0; i < aij->nz; i++) PetscCall(PetscRandomGetValue(rctx, aa + i)); 3435 PetscCall(MatSeqAIJRestoreArrayWrite(x, &aa)); 3436 } 3437 PetscCall(MatAssemblyBegin(x, MAT_FINAL_ASSEMBLY)); 3438 PetscCall(MatAssemblyEnd(x, MAT_FINAL_ASSEMBLY)); 3439 PetscFunctionReturn(0); 3440 } 3441 3442 /* Like MatSetRandom_SeqAIJ, but do not set values on columns in range of [low, high) */ 3443 PetscErrorCode MatSetRandomSkipColumnRange_SeqAIJ_Private(Mat x, PetscInt low, PetscInt high, PetscRandom rctx) 3444 { 3445 Mat_SeqAIJ *aij = (Mat_SeqAIJ *)x->data; 3446 PetscScalar a; 3447 PetscInt m, n, i, j, col, nskip; 3448 3449 PetscFunctionBegin; 3450 nskip = high - low; 3451 PetscCall(MatGetSize(x, &m, &n)); 3452 n -= nskip; /* shrink number of columns where nonzeros can be set */ 3453 for (i = 0; i < m; i++) { 3454 for (j = 0; j < aij->imax[i]; j++) { 3455 PetscCall(PetscRandomGetValue(rctx, &a)); 3456 col = (PetscInt)(n * PetscRealPart(a)); 3457 if (col >= low) col += nskip; /* shift col rightward to skip the hole */ 3458 PetscCall(MatSetValues(x, 1, &i, 1, &col, &a, ADD_VALUES)); 3459 } 3460 } 3461 PetscCall(MatAssemblyBegin(x, MAT_FINAL_ASSEMBLY)); 3462 PetscCall(MatAssemblyEnd(x, MAT_FINAL_ASSEMBLY)); 3463 PetscFunctionReturn(0); 3464 } 3465 3466 /* -------------------------------------------------------------------*/ 3467 static struct _MatOps MatOps_Values = {MatSetValues_SeqAIJ, 3468 MatGetRow_SeqAIJ, 3469 MatRestoreRow_SeqAIJ, 3470 MatMult_SeqAIJ, 3471 /* 4*/ MatMultAdd_SeqAIJ, 3472 MatMultTranspose_SeqAIJ, 3473 MatMultTransposeAdd_SeqAIJ, 3474 NULL, 3475 NULL, 3476 NULL, 3477 /* 10*/ NULL, 3478 MatLUFactor_SeqAIJ, 3479 NULL, 3480 MatSOR_SeqAIJ, 3481 MatTranspose_SeqAIJ, 3482 /*1 5*/ MatGetInfo_SeqAIJ, 3483 MatEqual_SeqAIJ, 3484 MatGetDiagonal_SeqAIJ, 3485 MatDiagonalScale_SeqAIJ, 3486 MatNorm_SeqAIJ, 3487 /* 20*/ NULL, 3488 MatAssemblyEnd_SeqAIJ, 3489 MatSetOption_SeqAIJ, 3490 MatZeroEntries_SeqAIJ, 3491 /* 24*/ MatZeroRows_SeqAIJ, 3492 NULL, 3493 NULL, 3494 NULL, 3495 NULL, 3496 /* 29*/ MatSetUp_SeqAIJ, 3497 NULL, 3498 NULL, 3499 NULL, 3500 NULL, 3501 /* 34*/ MatDuplicate_SeqAIJ, 3502 NULL, 3503 NULL, 3504 MatILUFactor_SeqAIJ, 3505 NULL, 3506 /* 39*/ MatAXPY_SeqAIJ, 3507 MatCreateSubMatrices_SeqAIJ, 3508 MatIncreaseOverlap_SeqAIJ, 3509 MatGetValues_SeqAIJ, 3510 MatCopy_SeqAIJ, 3511 /* 44*/ MatGetRowMax_SeqAIJ, 3512 MatScale_SeqAIJ, 3513 MatShift_SeqAIJ, 3514 MatDiagonalSet_SeqAIJ, 3515 MatZeroRowsColumns_SeqAIJ, 3516 /* 49*/ MatSetRandom_SeqAIJ, 3517 MatGetRowIJ_SeqAIJ, 3518 MatRestoreRowIJ_SeqAIJ, 3519 MatGetColumnIJ_SeqAIJ, 3520 MatRestoreColumnIJ_SeqAIJ, 3521 /* 54*/ MatFDColoringCreate_SeqXAIJ, 3522 NULL, 3523 NULL, 3524 MatPermute_SeqAIJ, 3525 NULL, 3526 /* 59*/ NULL, 3527 MatDestroy_SeqAIJ, 3528 MatView_SeqAIJ, 3529 NULL, 3530 NULL, 3531 /* 64*/ NULL, 3532 MatMatMatMultNumeric_SeqAIJ_SeqAIJ_SeqAIJ, 3533 NULL, 3534 NULL, 3535 NULL, 3536 /* 69*/ MatGetRowMaxAbs_SeqAIJ, 3537 MatGetRowMinAbs_SeqAIJ, 3538 NULL, 3539 NULL, 3540 NULL, 3541 /* 74*/ NULL, 3542 MatFDColoringApply_AIJ, 3543 NULL, 3544 NULL, 3545 NULL, 3546 /* 79*/ MatFindZeroDiagonals_SeqAIJ, 3547 NULL, 3548 NULL, 3549 NULL, 3550 MatLoad_SeqAIJ, 3551 /* 84*/ MatIsSymmetric_SeqAIJ, 3552 MatIsHermitian_SeqAIJ, 3553 NULL, 3554 NULL, 3555 NULL, 3556 /* 89*/ NULL, 3557 NULL, 3558 MatMatMultNumeric_SeqAIJ_SeqAIJ, 3559 NULL, 3560 NULL, 3561 /* 94*/ MatPtAPNumeric_SeqAIJ_SeqAIJ_SparseAxpy, 3562 NULL, 3563 NULL, 3564 MatMatTransposeMultNumeric_SeqAIJ_SeqAIJ, 3565 NULL, 3566 /* 99*/ MatProductSetFromOptions_SeqAIJ, 3567 NULL, 3568 NULL, 3569 MatConjugate_SeqAIJ, 3570 NULL, 3571 /*104*/ MatSetValuesRow_SeqAIJ, 3572 MatRealPart_SeqAIJ, 3573 MatImaginaryPart_SeqAIJ, 3574 NULL, 3575 NULL, 3576 /*109*/ MatMatSolve_SeqAIJ, 3577 NULL, 3578 MatGetRowMin_SeqAIJ, 3579 NULL, 3580 MatMissingDiagonal_SeqAIJ, 3581 /*114*/ NULL, 3582 NULL, 3583 NULL, 3584 NULL, 3585 NULL, 3586 /*119*/ NULL, 3587 NULL, 3588 NULL, 3589 NULL, 3590 MatGetMultiProcBlock_SeqAIJ, 3591 /*124*/ MatFindNonzeroRows_SeqAIJ, 3592 MatGetColumnReductions_SeqAIJ, 3593 MatInvertBlockDiagonal_SeqAIJ, 3594 MatInvertVariableBlockDiagonal_SeqAIJ, 3595 NULL, 3596 /*129*/ NULL, 3597 NULL, 3598 NULL, 3599 MatTransposeMatMultNumeric_SeqAIJ_SeqAIJ, 3600 MatTransposeColoringCreate_SeqAIJ, 3601 /*134*/ MatTransColoringApplySpToDen_SeqAIJ, 3602 MatTransColoringApplyDenToSp_SeqAIJ, 3603 NULL, 3604 NULL, 3605 MatRARtNumeric_SeqAIJ_SeqAIJ, 3606 /*139*/ NULL, 3607 NULL, 3608 NULL, 3609 MatFDColoringSetUp_SeqXAIJ, 3610 MatFindOffBlockDiagonalEntries_SeqAIJ, 3611 MatCreateMPIMatConcatenateSeqMat_SeqAIJ, 3612 /*145*/ MatDestroySubMatrices_SeqAIJ, 3613 NULL, 3614 NULL, 3615 MatCreateGraph_Simple_AIJ, 3616 NULL, 3617 /*150*/ MatTransposeSymbolic_SeqAIJ, 3618 MatEliminateZeros_SeqAIJ}; 3619 3620 PetscErrorCode MatSeqAIJSetColumnIndices_SeqAIJ(Mat mat, PetscInt *indices) 3621 { 3622 Mat_SeqAIJ *aij = (Mat_SeqAIJ *)mat->data; 3623 PetscInt i, nz, n; 3624 3625 PetscFunctionBegin; 3626 nz = aij->maxnz; 3627 n = mat->rmap->n; 3628 for (i = 0; i < nz; i++) aij->j[i] = indices[i]; 3629 aij->nz = nz; 3630 for (i = 0; i < n; i++) aij->ilen[i] = aij->imax[i]; 3631 PetscFunctionReturn(0); 3632 } 3633 3634 /* 3635 * Given a sparse matrix with global column indices, compact it by using a local column space. 3636 * The result matrix helps saving memory in other algorithms, such as MatPtAPSymbolic_MPIAIJ_MPIAIJ_scalable() 3637 */ 3638 PetscErrorCode MatSeqAIJCompactOutExtraColumns_SeqAIJ(Mat mat, ISLocalToGlobalMapping *mapping) 3639 { 3640 Mat_SeqAIJ *aij = (Mat_SeqAIJ *)mat->data; 3641 PetscHMapI gid1_lid1; 3642 PetscHashIter tpos; 3643 PetscInt gid, lid, i, ec, nz = aij->nz; 3644 PetscInt *garray, *jj = aij->j; 3645 3646 PetscFunctionBegin; 3647 PetscValidHeaderSpecific(mat, MAT_CLASSID, 1); 3648 PetscValidPointer(mapping, 2); 3649 /* use a table */ 3650 PetscCall(PetscHMapICreateWithSize(mat->rmap->n, &gid1_lid1)); 3651 ec = 0; 3652 for (i = 0; i < nz; i++) { 3653 PetscInt data, gid1 = jj[i] + 1; 3654 PetscCall(PetscHMapIGetWithDefault(gid1_lid1, gid1, 0, &data)); 3655 if (!data) { 3656 /* one based table */ 3657 PetscCall(PetscHMapISet(gid1_lid1, gid1, ++ec)); 3658 } 3659 } 3660 /* form array of columns we need */ 3661 PetscCall(PetscMalloc1(ec, &garray)); 3662 PetscHashIterBegin(gid1_lid1, tpos); 3663 while (!PetscHashIterAtEnd(gid1_lid1, tpos)) { 3664 PetscHashIterGetKey(gid1_lid1, tpos, gid); 3665 PetscHashIterGetVal(gid1_lid1, tpos, lid); 3666 PetscHashIterNext(gid1_lid1, tpos); 3667 gid--; 3668 lid--; 3669 garray[lid] = gid; 3670 } 3671 PetscCall(PetscSortInt(ec, garray)); /* sort, and rebuild */ 3672 PetscCall(PetscHMapIClear(gid1_lid1)); 3673 for (i = 0; i < ec; i++) PetscCall(PetscHMapISet(gid1_lid1, garray[i] + 1, i + 1)); 3674 /* compact out the extra columns in B */ 3675 for (i = 0; i < nz; i++) { 3676 PetscInt gid1 = jj[i] + 1; 3677 PetscCall(PetscHMapIGetWithDefault(gid1_lid1, gid1, 0, &lid)); 3678 lid--; 3679 jj[i] = lid; 3680 } 3681 PetscCall(PetscLayoutDestroy(&mat->cmap)); 3682 PetscCall(PetscHMapIDestroy(&gid1_lid1)); 3683 PetscCall(PetscLayoutCreateFromSizes(PetscObjectComm((PetscObject)mat), ec, ec, 1, &mat->cmap)); 3684 PetscCall(ISLocalToGlobalMappingCreate(PETSC_COMM_SELF, mat->cmap->bs, mat->cmap->n, garray, PETSC_OWN_POINTER, mapping)); 3685 PetscCall(ISLocalToGlobalMappingSetType(*mapping, ISLOCALTOGLOBALMAPPINGHASH)); 3686 PetscFunctionReturn(0); 3687 } 3688 3689 /*@ 3690 MatSeqAIJSetColumnIndices - Set the column indices for all the rows 3691 in the matrix. 3692 3693 Input Parameters: 3694 + mat - the `MATSEQAIJ` matrix 3695 - indices - the column indices 3696 3697 Level: advanced 3698 3699 Notes: 3700 This can be called if you have precomputed the nonzero structure of the 3701 matrix and want to provide it to the matrix object to improve the performance 3702 of the `MatSetValues()` operation. 3703 3704 You MUST have set the correct numbers of nonzeros per row in the call to 3705 `MatCreateSeqAIJ()`, and the columns indices MUST be sorted. 3706 3707 MUST be called before any calls to `MatSetValues()` 3708 3709 The indices should start with zero, not one. 3710 3711 @*/ 3712 PetscErrorCode MatSeqAIJSetColumnIndices(Mat mat, PetscInt *indices) 3713 { 3714 PetscFunctionBegin; 3715 PetscValidHeaderSpecific(mat, MAT_CLASSID, 1); 3716 PetscValidIntPointer(indices, 2); 3717 PetscUseMethod(mat, "MatSeqAIJSetColumnIndices_C", (Mat, PetscInt *), (mat, indices)); 3718 PetscFunctionReturn(0); 3719 } 3720 3721 /* ----------------------------------------------------------------------------------------*/ 3722 3723 PetscErrorCode MatStoreValues_SeqAIJ(Mat mat) 3724 { 3725 Mat_SeqAIJ *aij = (Mat_SeqAIJ *)mat->data; 3726 size_t nz = aij->i[mat->rmap->n]; 3727 3728 PetscFunctionBegin; 3729 PetscCheck(aij->nonew, PETSC_COMM_SELF, PETSC_ERR_ORDER, "Must call MatSetOption(A,MAT_NEW_NONZERO_LOCATIONS,PETSC_FALSE);first"); 3730 3731 /* allocate space for values if not already there */ 3732 if (!aij->saved_values) { PetscCall(PetscMalloc1(nz + 1, &aij->saved_values)); } 3733 3734 /* copy values over */ 3735 PetscCall(PetscArraycpy(aij->saved_values, aij->a, nz)); 3736 PetscFunctionReturn(0); 3737 } 3738 3739 /*@ 3740 MatStoreValues - Stashes a copy of the matrix values; this allows, for 3741 example, reuse of the linear part of a Jacobian, while recomputing the 3742 nonlinear portion. 3743 3744 Logically Collect 3745 3746 Input Parameters: 3747 . mat - the matrix (currently only `MATAIJ` matrices support this option) 3748 3749 Level: advanced 3750 3751 Common Usage, with `SNESSolve()`: 3752 $ Create Jacobian matrix 3753 $ Set linear terms into matrix 3754 $ Apply boundary conditions to matrix, at this time matrix must have 3755 $ final nonzero structure (i.e. setting the nonlinear terms and applying 3756 $ boundary conditions again will not change the nonzero structure 3757 $ ierr = MatSetOption(mat,MAT_NEW_NONZERO_LOCATIONS,PETSC_FALSE); 3758 $ ierr = MatStoreValues(mat); 3759 $ Call SNESSetJacobian() with matrix 3760 $ In your Jacobian routine 3761 $ ierr = MatRetrieveValues(mat); 3762 $ Set nonlinear terms in matrix 3763 3764 Common Usage without SNESSolve(), i.e. when you handle nonlinear solve yourself: 3765 $ // build linear portion of Jacobian 3766 $ ierr = MatSetOption(mat,MAT_NEW_NONZERO_LOCATIONS,PETSC_FALSE); 3767 $ ierr = MatStoreValues(mat); 3768 $ loop over nonlinear iterations 3769 $ ierr = MatRetrieveValues(mat); 3770 $ // call MatSetValues(mat,...) to set nonliner portion of Jacobian 3771 $ // call MatAssemblyBegin/End() on matrix 3772 $ Solve linear system with Jacobian 3773 $ endloop 3774 3775 Notes: 3776 Matrix must already be assemblied before calling this routine 3777 Must set the matrix option `MatSetOption`(mat,`MAT_NEW_NONZERO_LOCATIONS`,`PETSC_FALSE`); before 3778 calling this routine. 3779 3780 When this is called multiple times it overwrites the previous set of stored values 3781 and does not allocated additional space. 3782 3783 .seealso: `MatRetrieveValues()` 3784 @*/ 3785 PetscErrorCode MatStoreValues(Mat mat) 3786 { 3787 PetscFunctionBegin; 3788 PetscValidHeaderSpecific(mat, MAT_CLASSID, 1); 3789 PetscCheck(mat->assembled, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Not for unassembled matrix"); 3790 PetscCheck(!mat->factortype, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Not for factored matrix"); 3791 PetscUseMethod(mat, "MatStoreValues_C", (Mat), (mat)); 3792 PetscFunctionReturn(0); 3793 } 3794 3795 PetscErrorCode MatRetrieveValues_SeqAIJ(Mat mat) 3796 { 3797 Mat_SeqAIJ *aij = (Mat_SeqAIJ *)mat->data; 3798 PetscInt nz = aij->i[mat->rmap->n]; 3799 3800 PetscFunctionBegin; 3801 PetscCheck(aij->nonew, PETSC_COMM_SELF, PETSC_ERR_ORDER, "Must call MatSetOption(A,MAT_NEW_NONZERO_LOCATIONS,PETSC_FALSE);first"); 3802 PetscCheck(aij->saved_values, PETSC_COMM_SELF, PETSC_ERR_ORDER, "Must call MatStoreValues(A);first"); 3803 /* copy values over */ 3804 PetscCall(PetscArraycpy(aij->a, aij->saved_values, nz)); 3805 PetscFunctionReturn(0); 3806 } 3807 3808 /*@ 3809 MatRetrieveValues - Retrieves the copy of the matrix values; this allows, for 3810 example, reuse of the linear part of a Jacobian, while recomputing the 3811 nonlinear portion. 3812 3813 Logically Collect 3814 3815 Input Parameters: 3816 . mat - the matrix (currently only `MATAIJ` matrices support this option) 3817 3818 Level: advanced 3819 3820 .seealso: `MatStoreValues()` 3821 @*/ 3822 PetscErrorCode MatRetrieveValues(Mat mat) 3823 { 3824 PetscFunctionBegin; 3825 PetscValidHeaderSpecific(mat, MAT_CLASSID, 1); 3826 PetscCheck(mat->assembled, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Not for unassembled matrix"); 3827 PetscCheck(!mat->factortype, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Not for factored matrix"); 3828 PetscUseMethod(mat, "MatRetrieveValues_C", (Mat), (mat)); 3829 PetscFunctionReturn(0); 3830 } 3831 3832 /* --------------------------------------------------------------------------------*/ 3833 /*@C 3834 MatCreateSeqAIJ - Creates a sparse matrix in `MATSEQAIJ` (compressed row) format 3835 (the default parallel PETSc format). For good matrix assembly performance 3836 the user should preallocate the matrix storage by setting the parameter nz 3837 (or the array nnz). By setting these parameters accurately, performance 3838 during matrix assembly can be increased by more than a factor of 50. 3839 3840 Collective 3841 3842 Input Parameters: 3843 + comm - MPI communicator, set to `PETSC_COMM_SELF` 3844 . m - number of rows 3845 . n - number of columns 3846 . nz - number of nonzeros per row (same for all rows) 3847 - nnz - array containing the number of nonzeros in the various rows 3848 (possibly different for each row) or NULL 3849 3850 Output Parameter: 3851 . A - the matrix 3852 3853 It is recommended that one use the `MatCreate()`, `MatSetType()` and/or `MatSetFromOptions()`, 3854 MatXXXXSetPreallocation() paradigm instead of this routine directly. 3855 [MatXXXXSetPreallocation() is, for example, `MatSeqAIJSetPreallocation()`] 3856 3857 Notes: 3858 If nnz is given then nz is ignored 3859 3860 The AIJ format, also called 3861 compressed row storage, is fully compatible with standard Fortran 77 3862 storage. That is, the stored row and column indices can begin at 3863 either one (as in Fortran) or zero. See the users' manual for details. 3864 3865 Specify the preallocated storage with either nz or nnz (not both). 3866 Set nz = `PETSC_DEFAULT` and nnz = NULL for PETSc to control dynamic memory 3867 allocation. For large problems you MUST preallocate memory or you 3868 will get TERRIBLE performance, see the users' manual chapter on matrices. 3869 3870 By default, this format uses inodes (identical nodes) when possible, to 3871 improve numerical efficiency of matrix-vector products and solves. We 3872 search for consecutive rows with the same nonzero structure, thereby 3873 reusing matrix information to achieve increased efficiency. 3874 3875 Options Database Keys: 3876 + -mat_no_inode - Do not use inodes 3877 - -mat_inode_limit <limit> - Sets inode limit (max limit=5) 3878 3879 Level: intermediate 3880 3881 .seealso: [Sparse Matrix Creation](sec_matsparse), `MatCreate()`, `MatCreateAIJ()`, `MatSetValues()`, `MatSeqAIJSetColumnIndices()`, `MatCreateSeqAIJWithArrays()` 3882 @*/ 3883 PetscErrorCode MatCreateSeqAIJ(MPI_Comm comm, PetscInt m, PetscInt n, PetscInt nz, const PetscInt nnz[], Mat *A) 3884 { 3885 PetscFunctionBegin; 3886 PetscCall(MatCreate(comm, A)); 3887 PetscCall(MatSetSizes(*A, m, n, m, n)); 3888 PetscCall(MatSetType(*A, MATSEQAIJ)); 3889 PetscCall(MatSeqAIJSetPreallocation_SeqAIJ(*A, nz, nnz)); 3890 PetscFunctionReturn(0); 3891 } 3892 3893 /*@C 3894 MatSeqAIJSetPreallocation - For good matrix assembly performance 3895 the user should preallocate the matrix storage by setting the parameter nz 3896 (or the array nnz). By setting these parameters accurately, performance 3897 during matrix assembly can be increased by more than a factor of 50. 3898 3899 Collective 3900 3901 Input Parameters: 3902 + B - The matrix 3903 . nz - number of nonzeros per row (same for all rows) 3904 - nnz - array containing the number of nonzeros in the various rows 3905 (possibly different for each row) or NULL 3906 3907 Notes: 3908 If nnz is given then nz is ignored 3909 3910 The `MATSEQAIJ` format also called 3911 compressed row storage, is fully compatible with standard Fortran 77 3912 storage. That is, the stored row and column indices can begin at 3913 either one (as in Fortran) or zero. See the users' manual for details. 3914 3915 Specify the preallocated storage with either nz or nnz (not both). 3916 Set nz = `PETSC_DEFAULT` and nnz = NULL for PETSc to control dynamic memory 3917 allocation. For large problems you MUST preallocate memory or you 3918 will get TERRIBLE performance, see the users' manual chapter on matrices. 3919 3920 You can call `MatGetInfo()` to get information on how effective the preallocation was; 3921 for example the fields mallocs,nz_allocated,nz_used,nz_unneeded; 3922 You can also run with the option -info and look for messages with the string 3923 malloc in them to see if additional memory allocation was needed. 3924 3925 Developer Notes: 3926 Use nz of `MAT_SKIP_ALLOCATION` to not allocate any space for the matrix 3927 entries or columns indices 3928 3929 By default, this format uses inodes (identical nodes) when possible, to 3930 improve numerical efficiency of matrix-vector products and solves. We 3931 search for consecutive rows with the same nonzero structure, thereby 3932 reusing matrix information to achieve increased efficiency. 3933 3934 Options Database Keys: 3935 + -mat_no_inode - Do not use inodes 3936 - -mat_inode_limit <limit> - Sets inode limit (max limit=5) 3937 3938 Level: intermediate 3939 3940 .seealso: `MatCreate()`, `MatCreateAIJ()`, `MatSetValues()`, `MatSeqAIJSetColumnIndices()`, `MatCreateSeqAIJWithArrays()`, `MatGetInfo()`, 3941 `MatSeqAIJSetTotalPreallocation()` 3942 @*/ 3943 PetscErrorCode MatSeqAIJSetPreallocation(Mat B, PetscInt nz, const PetscInt nnz[]) 3944 { 3945 PetscFunctionBegin; 3946 PetscValidHeaderSpecific(B, MAT_CLASSID, 1); 3947 PetscValidType(B, 1); 3948 PetscTryMethod(B, "MatSeqAIJSetPreallocation_C", (Mat, PetscInt, const PetscInt[]), (B, nz, nnz)); 3949 PetscFunctionReturn(0); 3950 } 3951 3952 PetscErrorCode MatSeqAIJSetPreallocation_SeqAIJ(Mat B, PetscInt nz, const PetscInt *nnz) 3953 { 3954 Mat_SeqAIJ *b; 3955 PetscBool skipallocation = PETSC_FALSE, realalloc = PETSC_FALSE; 3956 PetscInt i; 3957 3958 PetscFunctionBegin; 3959 if (nz >= 0 || nnz) realalloc = PETSC_TRUE; 3960 if (nz == MAT_SKIP_ALLOCATION) { 3961 skipallocation = PETSC_TRUE; 3962 nz = 0; 3963 } 3964 PetscCall(PetscLayoutSetUp(B->rmap)); 3965 PetscCall(PetscLayoutSetUp(B->cmap)); 3966 3967 if (nz == PETSC_DEFAULT || nz == PETSC_DECIDE) nz = 5; 3968 PetscCheck(nz >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "nz cannot be less than 0: value %" PetscInt_FMT, nz); 3969 if (PetscUnlikelyDebug(nnz)) { 3970 for (i = 0; i < B->rmap->n; i++) { 3971 PetscCheck(nnz[i] >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "nnz cannot be less than 0: local row %" PetscInt_FMT " value %" PetscInt_FMT, i, nnz[i]); 3972 PetscCheck(nnz[i] <= B->cmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "nnz cannot be greater than row length: local row %" PetscInt_FMT " value %" PetscInt_FMT " rowlength %" PetscInt_FMT, i, nnz[i], B->cmap->n); 3973 } 3974 } 3975 3976 B->preallocated = PETSC_TRUE; 3977 3978 b = (Mat_SeqAIJ *)B->data; 3979 3980 if (!skipallocation) { 3981 if (!b->imax) { PetscCall(PetscMalloc1(B->rmap->n, &b->imax)); } 3982 if (!b->ilen) { 3983 /* b->ilen will count nonzeros in each row so far. */ 3984 PetscCall(PetscCalloc1(B->rmap->n, &b->ilen)); 3985 } else { 3986 PetscCall(PetscMemzero(b->ilen, B->rmap->n * sizeof(PetscInt))); 3987 } 3988 if (!b->ipre) { PetscCall(PetscMalloc1(B->rmap->n, &b->ipre)); } 3989 if (!nnz) { 3990 if (nz == PETSC_DEFAULT || nz == PETSC_DECIDE) nz = 10; 3991 else if (nz < 0) nz = 1; 3992 nz = PetscMin(nz, B->cmap->n); 3993 for (i = 0; i < B->rmap->n; i++) b->imax[i] = nz; 3994 nz = nz * B->rmap->n; 3995 } else { 3996 PetscInt64 nz64 = 0; 3997 for (i = 0; i < B->rmap->n; i++) { 3998 b->imax[i] = nnz[i]; 3999 nz64 += nnz[i]; 4000 } 4001 PetscCall(PetscIntCast(nz64, &nz)); 4002 } 4003 4004 /* allocate the matrix space */ 4005 /* FIXME: should B's old memory be unlogged? */ 4006 PetscCall(MatSeqXAIJFreeAIJ(B, &b->a, &b->j, &b->i)); 4007 if (B->structure_only) { 4008 PetscCall(PetscMalloc1(nz, &b->j)); 4009 PetscCall(PetscMalloc1(B->rmap->n + 1, &b->i)); 4010 } else { 4011 PetscCall(PetscMalloc3(nz, &b->a, nz, &b->j, B->rmap->n + 1, &b->i)); 4012 } 4013 b->i[0] = 0; 4014 for (i = 1; i < B->rmap->n + 1; i++) b->i[i] = b->i[i - 1] + b->imax[i - 1]; 4015 if (B->structure_only) { 4016 b->singlemalloc = PETSC_FALSE; 4017 b->free_a = PETSC_FALSE; 4018 } else { 4019 b->singlemalloc = PETSC_TRUE; 4020 b->free_a = PETSC_TRUE; 4021 } 4022 b->free_ij = PETSC_TRUE; 4023 } else { 4024 b->free_a = PETSC_FALSE; 4025 b->free_ij = PETSC_FALSE; 4026 } 4027 4028 if (b->ipre && nnz != b->ipre && b->imax) { 4029 /* reserve user-requested sparsity */ 4030 PetscCall(PetscArraycpy(b->ipre, b->imax, B->rmap->n)); 4031 } 4032 4033 b->nz = 0; 4034 b->maxnz = nz; 4035 B->info.nz_unneeded = (double)b->maxnz; 4036 if (realalloc) PetscCall(MatSetOption(B, MAT_NEW_NONZERO_ALLOCATION_ERR, PETSC_TRUE)); 4037 B->was_assembled = PETSC_FALSE; 4038 B->assembled = PETSC_FALSE; 4039 /* We simply deem preallocation has changed nonzero state. Updating the state 4040 will give clients (like AIJKokkos) a chance to know something has happened. 4041 */ 4042 B->nonzerostate++; 4043 PetscFunctionReturn(0); 4044 } 4045 4046 PetscErrorCode MatResetPreallocation_SeqAIJ(Mat A) 4047 { 4048 Mat_SeqAIJ *a; 4049 PetscInt i; 4050 4051 PetscFunctionBegin; 4052 PetscValidHeaderSpecific(A, MAT_CLASSID, 1); 4053 4054 /* Check local size. If zero, then return */ 4055 if (!A->rmap->n) PetscFunctionReturn(0); 4056 4057 a = (Mat_SeqAIJ *)A->data; 4058 /* if no saved info, we error out */ 4059 PetscCheck(a->ipre, PETSC_COMM_SELF, PETSC_ERR_ARG_NULL, "No saved preallocation info "); 4060 4061 PetscCheck(a->i && a->j && a->a && a->imax && a->ilen, PETSC_COMM_SELF, PETSC_ERR_ARG_NULL, "Memory info is incomplete, and can not reset preallocation "); 4062 4063 PetscCall(PetscArraycpy(a->imax, a->ipre, A->rmap->n)); 4064 PetscCall(PetscArrayzero(a->ilen, A->rmap->n)); 4065 a->i[0] = 0; 4066 for (i = 1; i < A->rmap->n + 1; i++) a->i[i] = a->i[i - 1] + a->imax[i - 1]; 4067 A->preallocated = PETSC_TRUE; 4068 a->nz = 0; 4069 a->maxnz = a->i[A->rmap->n]; 4070 A->info.nz_unneeded = (double)a->maxnz; 4071 A->was_assembled = PETSC_FALSE; 4072 A->assembled = PETSC_FALSE; 4073 PetscFunctionReturn(0); 4074 } 4075 4076 /*@ 4077 MatSeqAIJSetPreallocationCSR - Allocates memory for a sparse sequential matrix in `MATSEQAIJ` format. 4078 4079 Input Parameters: 4080 + B - the matrix 4081 . i - the indices into j for the start of each row (starts with zero) 4082 . j - the column indices for each row (starts with zero) these must be sorted for each row 4083 - v - optional values in the matrix 4084 4085 Level: developer 4086 4087 Notes: 4088 The i,j,v values are COPIED with this routine; to avoid the copy use `MatCreateSeqAIJWithArrays()` 4089 4090 This routine may be called multiple times with different nonzero patterns (or the same nonzero pattern). The nonzero 4091 structure will be the union of all the previous nonzero structures. 4092 4093 Developer Notes: 4094 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 4095 then just copies the v values directly with `PetscMemcpy()`. 4096 4097 This routine could also take a `PetscCopyMode` argument to allow sharing the values instead of always copying them. 4098 4099 .seealso: `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatSeqAIJSetPreallocation()`, `MatCreateSeqAIJ()`, `MATSEQAIJ`, `MatResetPreallocation()` 4100 @*/ 4101 PetscErrorCode MatSeqAIJSetPreallocationCSR(Mat B, const PetscInt i[], const PetscInt j[], const PetscScalar v[]) 4102 { 4103 PetscFunctionBegin; 4104 PetscValidHeaderSpecific(B, MAT_CLASSID, 1); 4105 PetscValidType(B, 1); 4106 PetscTryMethod(B, "MatSeqAIJSetPreallocationCSR_C", (Mat, const PetscInt[], const PetscInt[], const PetscScalar[]), (B, i, j, v)); 4107 PetscFunctionReturn(0); 4108 } 4109 4110 PetscErrorCode MatSeqAIJSetPreallocationCSR_SeqAIJ(Mat B, const PetscInt Ii[], const PetscInt J[], const PetscScalar v[]) 4111 { 4112 PetscInt i; 4113 PetscInt m, n; 4114 PetscInt nz; 4115 PetscInt *nnz; 4116 4117 PetscFunctionBegin; 4118 PetscCheck(Ii[0] == 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Ii[0] must be 0 it is %" PetscInt_FMT, Ii[0]); 4119 4120 PetscCall(PetscLayoutSetUp(B->rmap)); 4121 PetscCall(PetscLayoutSetUp(B->cmap)); 4122 4123 PetscCall(MatGetSize(B, &m, &n)); 4124 PetscCall(PetscMalloc1(m + 1, &nnz)); 4125 for (i = 0; i < m; i++) { 4126 nz = Ii[i + 1] - Ii[i]; 4127 PetscCheck(nz >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Local row %" PetscInt_FMT " has a negative number of columns %" PetscInt_FMT, i, nz); 4128 nnz[i] = nz; 4129 } 4130 PetscCall(MatSeqAIJSetPreallocation(B, 0, nnz)); 4131 PetscCall(PetscFree(nnz)); 4132 4133 for (i = 0; i < m; i++) PetscCall(MatSetValues_SeqAIJ(B, 1, &i, Ii[i + 1] - Ii[i], J + Ii[i], v ? v + Ii[i] : NULL, INSERT_VALUES)); 4134 4135 PetscCall(MatAssemblyBegin(B, MAT_FINAL_ASSEMBLY)); 4136 PetscCall(MatAssemblyEnd(B, MAT_FINAL_ASSEMBLY)); 4137 4138 PetscCall(MatSetOption(B, MAT_NEW_NONZERO_LOCATION_ERR, PETSC_TRUE)); 4139 PetscFunctionReturn(0); 4140 } 4141 4142 /*@ 4143 MatSeqAIJKron - Computes C, the Kronecker product of A and B. 4144 4145 Input Parameters: 4146 + A - left-hand side matrix 4147 . B - right-hand side matrix 4148 - reuse - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX` 4149 4150 Output Parameter: 4151 . C - Kronecker product of A and B 4152 4153 Level: intermediate 4154 4155 Note: 4156 `MAT_REUSE_MATRIX` can only be used when the nonzero structure of the product matrix has not changed from that last call to `MatSeqAIJKron()`. 4157 4158 .seealso: `MatCreateSeqAIJ()`, `MATSEQAIJ`, `MATKAIJ`, `MatReuse` 4159 @*/ 4160 PetscErrorCode MatSeqAIJKron(Mat A, Mat B, MatReuse reuse, Mat *C) 4161 { 4162 PetscFunctionBegin; 4163 PetscValidHeaderSpecific(A, MAT_CLASSID, 1); 4164 PetscValidType(A, 1); 4165 PetscValidHeaderSpecific(B, MAT_CLASSID, 2); 4166 PetscValidType(B, 2); 4167 PetscValidPointer(C, 4); 4168 if (reuse == MAT_REUSE_MATRIX) { 4169 PetscValidHeaderSpecific(*C, MAT_CLASSID, 4); 4170 PetscValidType(*C, 4); 4171 } 4172 PetscTryMethod(A, "MatSeqAIJKron_C", (Mat, Mat, MatReuse, Mat *), (A, B, reuse, C)); 4173 PetscFunctionReturn(0); 4174 } 4175 4176 PetscErrorCode MatSeqAIJKron_SeqAIJ(Mat A, Mat B, MatReuse reuse, Mat *C) 4177 { 4178 Mat newmat; 4179 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 4180 Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data; 4181 PetscScalar *v; 4182 const PetscScalar *aa, *ba; 4183 PetscInt *i, *j, m, n, p, q, nnz = 0, am = A->rmap->n, bm = B->rmap->n, an = A->cmap->n, bn = B->cmap->n; 4184 PetscBool flg; 4185 4186 PetscFunctionBegin; 4187 PetscCheck(!A->factortype, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Not for factored matrix"); 4188 PetscCheck(A->assembled, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Not for unassembled matrix"); 4189 PetscCheck(!B->factortype, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Not for factored matrix"); 4190 PetscCheck(B->assembled, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Not for unassembled matrix"); 4191 PetscCall(PetscObjectTypeCompare((PetscObject)B, MATSEQAIJ, &flg)); 4192 PetscCheck(flg, PETSC_COMM_SELF, PETSC_ERR_SUP, "MatType %s", ((PetscObject)B)->type_name); 4193 PetscCheck(reuse == MAT_INITIAL_MATRIX || reuse == MAT_REUSE_MATRIX, PETSC_COMM_SELF, PETSC_ERR_SUP, "MatReuse %d", (int)reuse); 4194 if (reuse == MAT_INITIAL_MATRIX) { 4195 PetscCall(PetscMalloc2(am * bm + 1, &i, a->i[am] * b->i[bm], &j)); 4196 PetscCall(MatCreate(PETSC_COMM_SELF, &newmat)); 4197 PetscCall(MatSetSizes(newmat, am * bm, an * bn, am * bm, an * bn)); 4198 PetscCall(MatSetType(newmat, MATAIJ)); 4199 i[0] = 0; 4200 for (m = 0; m < am; ++m) { 4201 for (p = 0; p < bm; ++p) { 4202 i[m * bm + p + 1] = i[m * bm + p] + (a->i[m + 1] - a->i[m]) * (b->i[p + 1] - b->i[p]); 4203 for (n = a->i[m]; n < a->i[m + 1]; ++n) { 4204 for (q = b->i[p]; q < b->i[p + 1]; ++q) j[nnz++] = a->j[n] * bn + b->j[q]; 4205 } 4206 } 4207 } 4208 PetscCall(MatSeqAIJSetPreallocationCSR(newmat, i, j, NULL)); 4209 *C = newmat; 4210 PetscCall(PetscFree2(i, j)); 4211 nnz = 0; 4212 } 4213 PetscCall(MatSeqAIJGetArray(*C, &v)); 4214 PetscCall(MatSeqAIJGetArrayRead(A, &aa)); 4215 PetscCall(MatSeqAIJGetArrayRead(B, &ba)); 4216 for (m = 0; m < am; ++m) { 4217 for (p = 0; p < bm; ++p) { 4218 for (n = a->i[m]; n < a->i[m + 1]; ++n) { 4219 for (q = b->i[p]; q < b->i[p + 1]; ++q) v[nnz++] = aa[n] * ba[q]; 4220 } 4221 } 4222 } 4223 PetscCall(MatSeqAIJRestoreArray(*C, &v)); 4224 PetscCall(MatSeqAIJRestoreArrayRead(A, &aa)); 4225 PetscCall(MatSeqAIJRestoreArrayRead(B, &ba)); 4226 PetscFunctionReturn(0); 4227 } 4228 4229 #include <../src/mat/impls/dense/seq/dense.h> 4230 #include <petsc/private/kernels/petscaxpy.h> 4231 4232 /* 4233 Computes (B'*A')' since computing B*A directly is untenable 4234 4235 n p p 4236 [ ] [ ] [ ] 4237 m [ A ] * n [ B ] = m [ C ] 4238 [ ] [ ] [ ] 4239 4240 */ 4241 PetscErrorCode MatMatMultNumeric_SeqDense_SeqAIJ(Mat A, Mat B, Mat C) 4242 { 4243 Mat_SeqDense *sub_a = (Mat_SeqDense *)A->data; 4244 Mat_SeqAIJ *sub_b = (Mat_SeqAIJ *)B->data; 4245 Mat_SeqDense *sub_c = (Mat_SeqDense *)C->data; 4246 PetscInt i, j, n, m, q, p; 4247 const PetscInt *ii, *idx; 4248 const PetscScalar *b, *a, *a_q; 4249 PetscScalar *c, *c_q; 4250 PetscInt clda = sub_c->lda; 4251 PetscInt alda = sub_a->lda; 4252 4253 PetscFunctionBegin; 4254 m = A->rmap->n; 4255 n = A->cmap->n; 4256 p = B->cmap->n; 4257 a = sub_a->v; 4258 b = sub_b->a; 4259 c = sub_c->v; 4260 if (clda == m) { 4261 PetscCall(PetscArrayzero(c, m * p)); 4262 } else { 4263 for (j = 0; j < p; j++) 4264 for (i = 0; i < m; i++) c[j * clda + i] = 0.0; 4265 } 4266 ii = sub_b->i; 4267 idx = sub_b->j; 4268 for (i = 0; i < n; i++) { 4269 q = ii[i + 1] - ii[i]; 4270 while (q-- > 0) { 4271 c_q = c + clda * (*idx); 4272 a_q = a + alda * i; 4273 PetscKernelAXPY(c_q, *b, a_q, m); 4274 idx++; 4275 b++; 4276 } 4277 } 4278 PetscFunctionReturn(0); 4279 } 4280 4281 PetscErrorCode MatMatMultSymbolic_SeqDense_SeqAIJ(Mat A, Mat B, PetscReal fill, Mat C) 4282 { 4283 PetscInt m = A->rmap->n, n = B->cmap->n; 4284 PetscBool cisdense; 4285 4286 PetscFunctionBegin; 4287 PetscCheck(A->cmap->n == B->rmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "A->cmap->n %" PetscInt_FMT " != B->rmap->n %" PetscInt_FMT, A->cmap->n, B->rmap->n); 4288 PetscCall(MatSetSizes(C, m, n, m, n)); 4289 PetscCall(MatSetBlockSizesFromMats(C, A, B)); 4290 PetscCall(PetscObjectTypeCompareAny((PetscObject)C, &cisdense, MATSEQDENSE, MATSEQDENSECUDA, MATSEQDENSEHIP, "")); 4291 if (!cisdense) PetscCall(MatSetType(C, MATDENSE)); 4292 PetscCall(MatSetUp(C)); 4293 4294 C->ops->matmultnumeric = MatMatMultNumeric_SeqDense_SeqAIJ; 4295 PetscFunctionReturn(0); 4296 } 4297 4298 /* ----------------------------------------------------------------*/ 4299 /*MC 4300 MATSEQAIJ - MATSEQAIJ = "seqaij" - A matrix type to be used for sequential sparse matrices, 4301 based on compressed sparse row format. 4302 4303 Options Database Keys: 4304 . -mat_type seqaij - sets the matrix type to "seqaij" during a call to MatSetFromOptions() 4305 4306 Level: beginner 4307 4308 Notes: 4309 `MatSetValues()` may be called for this matrix type with a NULL argument for the numerical values, 4310 in this case the values associated with the rows and columns one passes in are set to zero 4311 in the matrix 4312 4313 `MatSetOptions`(,`MAT_STRUCTURE_ONLY`,`PETSC_TRUE`) may be called for this matrix type. In this no 4314 space is allocated for the nonzero entries and any entries passed with `MatSetValues()` are ignored 4315 4316 Developer Note: 4317 It would be nice if all matrix formats supported passing NULL in for the numerical values 4318 4319 .seealso: `MatCreateSeqAIJ()`, `MatSetFromOptions()`, `MatSetType()`, `MatCreate()`, `MatType`, `MATSELL`, `MATSEQSELL`, `MATMPISELL` 4320 M*/ 4321 4322 /*MC 4323 MATAIJ - MATAIJ = "aij" - A matrix type to be used for sparse matrices. 4324 4325 This matrix type is identical to `MATSEQAIJ` when constructed with a single process communicator, 4326 and `MATMPIAIJ` otherwise. As a result, for single process communicators, 4327 `MatSeqAIJSetPreallocation()` is supported, and similarly `MatMPIAIJSetPreallocation()` is supported 4328 for communicators controlling multiple processes. It is recommended that you call both of 4329 the above preallocation routines for simplicity. 4330 4331 Options Database Keys: 4332 . -mat_type aij - sets the matrix type to "aij" during a call to `MatSetFromOptions()` 4333 4334 Note: 4335 Subclasses include `MATAIJCUSPARSE`, `MATAIJPERM`, `MATAIJSELL`, `MATAIJMKL`, `MATAIJCRL`, and also automatically switches over to use inodes when 4336 enough exist. 4337 4338 Level: beginner 4339 4340 .seealso: `MatCreateAIJ()`, `MatCreateSeqAIJ()`, `MATSEQAIJ`, `MATMPIAIJ`, `MATSELL`, `MATSEQSELL`, `MATMPISELL` 4341 M*/ 4342 4343 /*MC 4344 MATAIJCRL - MATAIJCRL = "aijcrl" - A matrix type to be used for sparse matrices. 4345 4346 This matrix type is identical to `MATSEQAIJCRL` when constructed with a single process communicator, 4347 and `MATMPIAIJCRL` otherwise. As a result, for single process communicators, 4348 `MatSeqAIJSetPreallocation()` is supported, and similarly `MatMPIAIJSetPreallocation()` is supported 4349 for communicators controlling multiple processes. It is recommended that you call both of 4350 the above preallocation routines for simplicity. 4351 4352 Options Database Keys: 4353 . -mat_type aijcrl - sets the matrix type to "aijcrl" during a call to `MatSetFromOptions()` 4354 4355 Level: beginner 4356 4357 .seealso: `MatCreateMPIAIJCRL`, `MATSEQAIJCRL`, `MATMPIAIJCRL`, `MATSEQAIJCRL`, `MATMPIAIJCRL` 4358 M*/ 4359 4360 PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqAIJCRL(Mat, MatType, MatReuse, Mat *); 4361 #if defined(PETSC_HAVE_ELEMENTAL) 4362 PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_Elemental(Mat, MatType, MatReuse, Mat *); 4363 #endif 4364 #if defined(PETSC_HAVE_SCALAPACK) 4365 PETSC_INTERN PetscErrorCode MatConvert_AIJ_ScaLAPACK(Mat, MatType, MatReuse, Mat *); 4366 #endif 4367 #if defined(PETSC_HAVE_HYPRE) 4368 PETSC_INTERN PetscErrorCode MatConvert_AIJ_HYPRE(Mat A, MatType, MatReuse, Mat *); 4369 #endif 4370 4371 PETSC_EXTERN PetscErrorCode MatConvert_SeqAIJ_SeqSELL(Mat, MatType, MatReuse, Mat *); 4372 PETSC_INTERN PetscErrorCode MatConvert_XAIJ_IS(Mat, MatType, MatReuse, Mat *); 4373 PETSC_INTERN PetscErrorCode MatProductSetFromOptions_IS_XAIJ(Mat); 4374 4375 /*@C 4376 MatSeqAIJGetArray - gives read/write access to the array where the data for a `MATSEQAIJ` matrix is stored 4377 4378 Not Collective 4379 4380 Input Parameter: 4381 . mat - a `MATSEQAIJ` matrix 4382 4383 Output Parameter: 4384 . array - pointer to the data 4385 4386 Level: intermediate 4387 4388 .seealso: `MatSeqAIJRestoreArray()`, `MatSeqAIJGetArrayF90()` 4389 @*/ 4390 PetscErrorCode MatSeqAIJGetArray(Mat A, PetscScalar **array) 4391 { 4392 Mat_SeqAIJ *aij = (Mat_SeqAIJ *)A->data; 4393 4394 PetscFunctionBegin; 4395 if (aij->ops->getarray) { 4396 PetscCall((*aij->ops->getarray)(A, array)); 4397 } else { 4398 *array = aij->a; 4399 } 4400 PetscFunctionReturn(0); 4401 } 4402 4403 /*@C 4404 MatSeqAIJRestoreArray - returns access to the array where the data for a `MATSEQAIJ` matrix is stored obtained by `MatSeqAIJGetArray()` 4405 4406 Not Collective 4407 4408 Input Parameters: 4409 + mat - a `MATSEQAIJ` matrix 4410 - array - pointer to the data 4411 4412 Level: intermediate 4413 4414 .seealso: `MatSeqAIJGetArray()`, `MatSeqAIJRestoreArrayF90()` 4415 @*/ 4416 PetscErrorCode MatSeqAIJRestoreArray(Mat A, PetscScalar **array) 4417 { 4418 Mat_SeqAIJ *aij = (Mat_SeqAIJ *)A->data; 4419 4420 PetscFunctionBegin; 4421 if (aij->ops->restorearray) { 4422 PetscCall((*aij->ops->restorearray)(A, array)); 4423 } else { 4424 *array = NULL; 4425 } 4426 PetscCall(MatSeqAIJInvalidateDiagonal(A)); 4427 PetscCall(PetscObjectStateIncrease((PetscObject)A)); 4428 PetscFunctionReturn(0); 4429 } 4430 4431 /*@C 4432 MatSeqAIJGetArrayRead - gives read-only access to the array where the data for a `MATSEQAIJ` matrix is stored 4433 4434 Not Collective 4435 4436 Input Parameter: 4437 . mat - a `MATSEQAIJ` matrix 4438 4439 Output Parameter: 4440 . array - pointer to the data 4441 4442 Level: intermediate 4443 4444 .seealso: `MatSeqAIJGetArray()`, `MatSeqAIJRestoreArrayRead()` 4445 @*/ 4446 PetscErrorCode MatSeqAIJGetArrayRead(Mat A, const PetscScalar **array) 4447 { 4448 Mat_SeqAIJ *aij = (Mat_SeqAIJ *)A->data; 4449 4450 PetscFunctionBegin; 4451 if (aij->ops->getarrayread) { 4452 PetscCall((*aij->ops->getarrayread)(A, array)); 4453 } else { 4454 *array = aij->a; 4455 } 4456 PetscFunctionReturn(0); 4457 } 4458 4459 /*@C 4460 MatSeqAIJRestoreArrayRead - restore the read-only access array obtained from `MatSeqAIJGetArrayRead()` 4461 4462 Not Collective 4463 4464 Input Parameter: 4465 . mat - a `MATSEQAIJ` matrix 4466 4467 Output Parameter: 4468 . array - pointer to the data 4469 4470 Level: intermediate 4471 4472 .seealso: `MatSeqAIJGetArray()`, `MatSeqAIJGetArrayRead()` 4473 @*/ 4474 PetscErrorCode MatSeqAIJRestoreArrayRead(Mat A, const PetscScalar **array) 4475 { 4476 Mat_SeqAIJ *aij = (Mat_SeqAIJ *)A->data; 4477 4478 PetscFunctionBegin; 4479 if (aij->ops->restorearrayread) { 4480 PetscCall((*aij->ops->restorearrayread)(A, array)); 4481 } else { 4482 *array = NULL; 4483 } 4484 PetscFunctionReturn(0); 4485 } 4486 4487 /*@C 4488 MatSeqAIJGetArrayWrite - gives write-only access to the array where the data for a `MATSEQAIJ` matrix is stored 4489 4490 Not Collective 4491 4492 Input Parameter: 4493 . mat - a `MATSEQAIJ` matrix 4494 4495 Output Parameter: 4496 . array - pointer to the data 4497 4498 Level: intermediate 4499 4500 .seealso: `MatSeqAIJGetArray()`, `MatSeqAIJRestoreArrayRead()` 4501 @*/ 4502 PetscErrorCode MatSeqAIJGetArrayWrite(Mat A, PetscScalar **array) 4503 { 4504 Mat_SeqAIJ *aij = (Mat_SeqAIJ *)A->data; 4505 4506 PetscFunctionBegin; 4507 if (aij->ops->getarraywrite) { 4508 PetscCall((*aij->ops->getarraywrite)(A, array)); 4509 } else { 4510 *array = aij->a; 4511 } 4512 PetscCall(MatSeqAIJInvalidateDiagonal(A)); 4513 PetscCall(PetscObjectStateIncrease((PetscObject)A)); 4514 PetscFunctionReturn(0); 4515 } 4516 4517 /*@C 4518 MatSeqAIJRestoreArrayWrite - restore the read-only access array obtained from MatSeqAIJGetArrayRead 4519 4520 Not Collective 4521 4522 Input Parameter: 4523 . mat - a MATSEQAIJ matrix 4524 4525 Output Parameter: 4526 . array - pointer to the data 4527 4528 Level: intermediate 4529 4530 .seealso: `MatSeqAIJGetArray()`, `MatSeqAIJGetArrayRead()` 4531 @*/ 4532 PetscErrorCode MatSeqAIJRestoreArrayWrite(Mat A, PetscScalar **array) 4533 { 4534 Mat_SeqAIJ *aij = (Mat_SeqAIJ *)A->data; 4535 4536 PetscFunctionBegin; 4537 if (aij->ops->restorearraywrite) { 4538 PetscCall((*aij->ops->restorearraywrite)(A, array)); 4539 } else { 4540 *array = NULL; 4541 } 4542 PetscFunctionReturn(0); 4543 } 4544 4545 /*@C 4546 MatSeqAIJGetCSRAndMemType - Get the CSR arrays and the memory type of the `MATSEQAIJ` matrix 4547 4548 Not Collective 4549 4550 Input Parameter: 4551 . mat - a matrix of type `MATSEQAIJ` or its subclasses 4552 4553 Output Parameters: 4554 + i - row map array of the matrix 4555 . j - column index array of the matrix 4556 . a - data array of the matrix 4557 - memtype - memory type of the arrays 4558 4559 Notes: 4560 Any of the output parameters can be NULL, in which case the corresponding value is not returned. 4561 If mat is a device matrix, the arrays are on the device. Otherwise, they are on the host. 4562 4563 One can call this routine on a preallocated but not assembled matrix to just get the memory of the CSR underneath the matrix. 4564 If the matrix is assembled, the data array 'a' is guaranteed to have the latest values of the matrix. 4565 4566 Level: Developer 4567 4568 .seealso: `MatSeqAIJGetArray()`, `MatSeqAIJGetArrayRead()` 4569 @*/ 4570 PetscErrorCode MatSeqAIJGetCSRAndMemType(Mat mat, const PetscInt **i, const PetscInt **j, PetscScalar **a, PetscMemType *mtype) 4571 { 4572 Mat_SeqAIJ *aij = (Mat_SeqAIJ *)mat->data; 4573 4574 PetscFunctionBegin; 4575 PetscCheck(mat->preallocated, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "matrix is not preallocated"); 4576 if (aij->ops->getcsrandmemtype) { 4577 PetscCall((*aij->ops->getcsrandmemtype)(mat, i, j, a, mtype)); 4578 } else { 4579 if (i) *i = aij->i; 4580 if (j) *j = aij->j; 4581 if (a) *a = aij->a; 4582 if (mtype) *mtype = PETSC_MEMTYPE_HOST; 4583 } 4584 PetscFunctionReturn(0); 4585 } 4586 4587 /*@C 4588 MatSeqAIJGetMaxRowNonzeros - returns the maximum number of nonzeros in any row 4589 4590 Not Collective 4591 4592 Input Parameter: 4593 . mat - a `MATSEQAIJ` matrix 4594 4595 Output Parameter: 4596 . nz - the maximum number of nonzeros in any row 4597 4598 Level: intermediate 4599 4600 .seealso: `MatSeqAIJRestoreArray()`, `MatSeqAIJGetArrayF90()` 4601 @*/ 4602 PetscErrorCode MatSeqAIJGetMaxRowNonzeros(Mat A, PetscInt *nz) 4603 { 4604 Mat_SeqAIJ *aij = (Mat_SeqAIJ *)A->data; 4605 4606 PetscFunctionBegin; 4607 *nz = aij->rmax; 4608 PetscFunctionReturn(0); 4609 } 4610 4611 PetscErrorCode MatSetPreallocationCOO_SeqAIJ(Mat mat, PetscCount coo_n, PetscInt coo_i[], PetscInt coo_j[]) 4612 { 4613 MPI_Comm comm; 4614 PetscInt *i, *j; 4615 PetscInt M, N, row; 4616 PetscCount k, p, q, nneg, nnz, start, end; /* Index the coo array, so use PetscCount as their type */ 4617 PetscInt *Ai; /* Change to PetscCount once we use it for row pointers */ 4618 PetscInt *Aj; 4619 PetscScalar *Aa; 4620 Mat_SeqAIJ *seqaij = (Mat_SeqAIJ *)(mat->data); 4621 MatType rtype; 4622 PetscCount *perm, *jmap; 4623 4624 PetscFunctionBegin; 4625 PetscCall(MatResetPreallocationCOO_SeqAIJ(mat)); 4626 PetscCall(PetscObjectGetComm((PetscObject)mat, &comm)); 4627 PetscCall(MatGetSize(mat, &M, &N)); 4628 i = coo_i; 4629 j = coo_j; 4630 PetscCall(PetscMalloc1(coo_n, &perm)); 4631 for (k = 0; k < coo_n; k++) { /* Ignore entries with negative row or col indices */ 4632 if (j[k] < 0) i[k] = -1; 4633 perm[k] = k; 4634 } 4635 4636 /* Sort by row */ 4637 PetscCall(PetscSortIntWithIntCountArrayPair(coo_n, i, j, perm)); 4638 for (k = 0; k < coo_n; k++) { 4639 if (i[k] >= 0) break; 4640 } /* Advance k to the first row with a non-negative index */ 4641 nneg = k; 4642 PetscCall(PetscMalloc1(coo_n - nneg + 1, &jmap)); /* +1 to make a CSR-like data structure. jmap[i] originally is the number of repeats for i-th nonzero */ 4643 nnz = 0; /* Total number of unique nonzeros to be counted */ 4644 jmap++; /* Inc jmap by 1 for convenience */ 4645 4646 PetscCall(PetscCalloc1(M + 1, &Ai)); /* CSR of A */ 4647 PetscCall(PetscMalloc1(coo_n - nneg, &Aj)); /* We have at most coo_n-nneg unique nonzeros */ 4648 4649 /* In each row, sort by column, then unique column indices to get row length */ 4650 Ai++; /* Inc by 1 for convenience */ 4651 q = 0; /* q-th unique nonzero, with q starting from 0 */ 4652 while (k < coo_n) { 4653 row = i[k]; 4654 start = k; /* [start,end) indices for this row */ 4655 while (k < coo_n && i[k] == row) k++; 4656 end = k; 4657 PetscCall(PetscSortIntWithCountArray(end - start, j + start, perm + start)); 4658 /* Find number of unique col entries in this row */ 4659 Aj[q] = j[start]; /* Log the first nonzero in this row */ 4660 jmap[q] = 1; /* Number of repeats of this nozero entry */ 4661 Ai[row] = 1; 4662 nnz++; 4663 4664 for (p = start + 1; p < end; p++) { /* Scan remaining nonzero in this row */ 4665 if (j[p] != j[p - 1]) { /* Meet a new nonzero */ 4666 q++; 4667 jmap[q] = 1; 4668 Aj[q] = j[p]; 4669 Ai[row]++; 4670 nnz++; 4671 } else { 4672 jmap[q]++; 4673 } 4674 } 4675 q++; /* Move to next row and thus next unique nonzero */ 4676 } 4677 4678 Ai--; /* Back to the beginning of Ai[] */ 4679 for (k = 0; k < M; k++) Ai[k + 1] += Ai[k]; 4680 jmap--; /* Back to the beginning of jmap[] */ 4681 jmap[0] = 0; 4682 for (k = 0; k < nnz; k++) jmap[k + 1] += jmap[k]; 4683 if (nnz < coo_n - nneg) { /* Realloc with actual number of unique nonzeros */ 4684 PetscCount *jmap_new; 4685 PetscInt *Aj_new; 4686 4687 PetscCall(PetscMalloc1(nnz + 1, &jmap_new)); 4688 PetscCall(PetscArraycpy(jmap_new, jmap, nnz + 1)); 4689 PetscCall(PetscFree(jmap)); 4690 jmap = jmap_new; 4691 4692 PetscCall(PetscMalloc1(nnz, &Aj_new)); 4693 PetscCall(PetscArraycpy(Aj_new, Aj, nnz)); 4694 PetscCall(PetscFree(Aj)); 4695 Aj = Aj_new; 4696 } 4697 4698 if (nneg) { /* Discard heading entries with negative indices in perm[], as we'll access it from index 0 in MatSetValuesCOO */ 4699 PetscCount *perm_new; 4700 4701 PetscCall(PetscMalloc1(coo_n - nneg, &perm_new)); 4702 PetscCall(PetscArraycpy(perm_new, perm + nneg, coo_n - nneg)); 4703 PetscCall(PetscFree(perm)); 4704 perm = perm_new; 4705 } 4706 4707 PetscCall(MatGetRootType_Private(mat, &rtype)); 4708 PetscCall(PetscCalloc1(nnz, &Aa)); /* Zero the matrix */ 4709 PetscCall(MatSetSeqAIJWithArrays_private(PETSC_COMM_SELF, M, N, Ai, Aj, Aa, rtype, mat)); 4710 4711 seqaij->singlemalloc = PETSC_FALSE; /* Ai, Aj and Aa are not allocated in one big malloc */ 4712 seqaij->free_a = seqaij->free_ij = PETSC_TRUE; /* Let newmat own Ai, Aj and Aa */ 4713 /* Record COO fields */ 4714 seqaij->coo_n = coo_n; 4715 seqaij->Atot = coo_n - nneg; /* Annz is seqaij->nz, so no need to record that again */ 4716 seqaij->jmap = jmap; /* of length nnz+1 */ 4717 seqaij->perm = perm; 4718 PetscFunctionReturn(0); 4719 } 4720 4721 static PetscErrorCode MatSetValuesCOO_SeqAIJ(Mat A, const PetscScalar v[], InsertMode imode) 4722 { 4723 Mat_SeqAIJ *aseq = (Mat_SeqAIJ *)A->data; 4724 PetscCount i, j, Annz = aseq->nz; 4725 PetscCount *perm = aseq->perm, *jmap = aseq->jmap; 4726 PetscScalar *Aa; 4727 4728 PetscFunctionBegin; 4729 PetscCall(MatSeqAIJGetArray(A, &Aa)); 4730 for (i = 0; i < Annz; i++) { 4731 PetscScalar sum = 0.0; 4732 for (j = jmap[i]; j < jmap[i + 1]; j++) sum += v[perm[j]]; 4733 Aa[i] = (imode == INSERT_VALUES ? 0.0 : Aa[i]) + sum; 4734 } 4735 PetscCall(MatSeqAIJRestoreArray(A, &Aa)); 4736 PetscFunctionReturn(0); 4737 } 4738 4739 #if defined(PETSC_HAVE_CUDA) 4740 PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqAIJCUSPARSE(Mat, MatType, MatReuse, Mat *); 4741 #endif 4742 #if defined(PETSC_HAVE_HIP) 4743 PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqAIJHIPSPARSE(Mat, MatType, MatReuse, Mat *); 4744 #endif 4745 #if defined(PETSC_HAVE_KOKKOS_KERNELS) 4746 PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqAIJKokkos(Mat, MatType, MatReuse, Mat *); 4747 #endif 4748 4749 PETSC_EXTERN PetscErrorCode MatCreate_SeqAIJ(Mat B) 4750 { 4751 Mat_SeqAIJ *b; 4752 PetscMPIInt size; 4753 4754 PetscFunctionBegin; 4755 PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)B), &size)); 4756 PetscCheck(size <= 1, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Comm must be of size 1"); 4757 4758 PetscCall(PetscNew(&b)); 4759 4760 B->data = (void *)b; 4761 4762 PetscCall(PetscMemcpy(B->ops, &MatOps_Values, sizeof(struct _MatOps))); 4763 if (B->sortedfull) B->ops->setvalues = MatSetValues_SeqAIJ_SortedFull; 4764 4765 b->row = NULL; 4766 b->col = NULL; 4767 b->icol = NULL; 4768 b->reallocs = 0; 4769 b->ignorezeroentries = PETSC_FALSE; 4770 b->roworiented = PETSC_TRUE; 4771 b->nonew = 0; 4772 b->diag = NULL; 4773 b->solve_work = NULL; 4774 B->spptr = NULL; 4775 b->saved_values = NULL; 4776 b->idiag = NULL; 4777 b->mdiag = NULL; 4778 b->ssor_work = NULL; 4779 b->omega = 1.0; 4780 b->fshift = 0.0; 4781 b->idiagvalid = PETSC_FALSE; 4782 b->ibdiagvalid = PETSC_FALSE; 4783 b->keepnonzeropattern = PETSC_FALSE; 4784 4785 PetscCall(PetscObjectChangeTypeName((PetscObject)B, MATSEQAIJ)); 4786 #if defined(PETSC_HAVE_MATLAB) 4787 PetscCall(PetscObjectComposeFunction((PetscObject)B, "PetscMatlabEnginePut_C", MatlabEnginePut_SeqAIJ)); 4788 PetscCall(PetscObjectComposeFunction((PetscObject)B, "PetscMatlabEngineGet_C", MatlabEngineGet_SeqAIJ)); 4789 #endif 4790 PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatSeqAIJSetColumnIndices_C", MatSeqAIJSetColumnIndices_SeqAIJ)); 4791 PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatStoreValues_C", MatStoreValues_SeqAIJ)); 4792 PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatRetrieveValues_C", MatRetrieveValues_SeqAIJ)); 4793 PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_seqsbaij_C", MatConvert_SeqAIJ_SeqSBAIJ)); 4794 PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_seqbaij_C", MatConvert_SeqAIJ_SeqBAIJ)); 4795 PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_seqaijperm_C", MatConvert_SeqAIJ_SeqAIJPERM)); 4796 PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_seqaijsell_C", MatConvert_SeqAIJ_SeqAIJSELL)); 4797 #if defined(PETSC_HAVE_MKL_SPARSE) 4798 PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_seqaijmkl_C", MatConvert_SeqAIJ_SeqAIJMKL)); 4799 #endif 4800 #if defined(PETSC_HAVE_CUDA) 4801 PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_seqaijcusparse_C", MatConvert_SeqAIJ_SeqAIJCUSPARSE)); 4802 PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_seqaijcusparse_seqaij_C", MatProductSetFromOptions_SeqAIJ)); 4803 PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_seqaij_seqaijcusparse_C", MatProductSetFromOptions_SeqAIJ)); 4804 #endif 4805 #if defined(PETSC_HAVE_HIP) 4806 PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_seqaijhipsparse_C", MatConvert_SeqAIJ_SeqAIJHIPSPARSE)); 4807 PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_seqaijhipsparse_seqaij_C", MatProductSetFromOptions_SeqAIJ)); 4808 PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_seqaij_seqaijhipsparse_C", MatProductSetFromOptions_SeqAIJ)); 4809 #endif 4810 #if defined(PETSC_HAVE_KOKKOS_KERNELS) 4811 PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_seqaijkokkos_C", MatConvert_SeqAIJ_SeqAIJKokkos)); 4812 #endif 4813 PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_seqaijcrl_C", MatConvert_SeqAIJ_SeqAIJCRL)); 4814 #if defined(PETSC_HAVE_ELEMENTAL) 4815 PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_elemental_C", MatConvert_SeqAIJ_Elemental)); 4816 #endif 4817 #if defined(PETSC_HAVE_SCALAPACK) 4818 PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_scalapack_C", MatConvert_AIJ_ScaLAPACK)); 4819 #endif 4820 #if defined(PETSC_HAVE_HYPRE) 4821 PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_hypre_C", MatConvert_AIJ_HYPRE)); 4822 PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_transpose_seqaij_seqaij_C", MatProductSetFromOptions_Transpose_AIJ_AIJ)); 4823 #endif 4824 PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_seqdense_C", MatConvert_SeqAIJ_SeqDense)); 4825 PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_seqsell_C", MatConvert_SeqAIJ_SeqSELL)); 4826 PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_is_C", MatConvert_XAIJ_IS)); 4827 PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatIsTranspose_C", MatIsTranspose_SeqAIJ)); 4828 PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatIsHermitianTranspose_C", MatIsTranspose_SeqAIJ)); 4829 PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatSeqAIJSetPreallocation_C", MatSeqAIJSetPreallocation_SeqAIJ)); 4830 PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatResetPreallocation_C", MatResetPreallocation_SeqAIJ)); 4831 PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatSeqAIJSetPreallocationCSR_C", MatSeqAIJSetPreallocationCSR_SeqAIJ)); 4832 PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatReorderForNonzeroDiagonal_C", MatReorderForNonzeroDiagonal_SeqAIJ)); 4833 PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_is_seqaij_C", MatProductSetFromOptions_IS_XAIJ)); 4834 PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_seqdense_seqaij_C", MatProductSetFromOptions_SeqDense_SeqAIJ)); 4835 PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_seqaij_seqaij_C", MatProductSetFromOptions_SeqAIJ)); 4836 PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatSeqAIJKron_C", MatSeqAIJKron_SeqAIJ)); 4837 PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatSetPreallocationCOO_C", MatSetPreallocationCOO_SeqAIJ)); 4838 PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatSetValuesCOO_C", MatSetValuesCOO_SeqAIJ)); 4839 PetscCall(MatCreate_SeqAIJ_Inode(B)); 4840 PetscCall(PetscObjectChangeTypeName((PetscObject)B, MATSEQAIJ)); 4841 PetscCall(MatSeqAIJSetTypeFromOptions(B)); /* this allows changing the matrix subtype to say MATSEQAIJPERM */ 4842 PetscFunctionReturn(0); 4843 } 4844 4845 /* 4846 Given a matrix generated with MatGetFactor() duplicates all the information in A into C 4847 */ 4848 PetscErrorCode MatDuplicateNoCreate_SeqAIJ(Mat C, Mat A, MatDuplicateOption cpvalues, PetscBool mallocmatspace) 4849 { 4850 Mat_SeqAIJ *c = (Mat_SeqAIJ *)C->data, *a = (Mat_SeqAIJ *)A->data; 4851 PetscInt m = A->rmap->n, i; 4852 4853 PetscFunctionBegin; 4854 PetscCheck(A->assembled || cpvalues == MAT_DO_NOT_COPY_VALUES, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Cannot duplicate unassembled matrix"); 4855 4856 C->factortype = A->factortype; 4857 c->row = NULL; 4858 c->col = NULL; 4859 c->icol = NULL; 4860 c->reallocs = 0; 4861 4862 C->assembled = A->assembled; 4863 C->preallocated = A->preallocated; 4864 4865 if (A->preallocated) { 4866 PetscCall(PetscLayoutReference(A->rmap, &C->rmap)); 4867 PetscCall(PetscLayoutReference(A->cmap, &C->cmap)); 4868 4869 PetscCall(PetscMalloc1(m, &c->imax)); 4870 PetscCall(PetscMemcpy(c->imax, a->imax, m * sizeof(PetscInt))); 4871 PetscCall(PetscMalloc1(m, &c->ilen)); 4872 PetscCall(PetscMemcpy(c->ilen, a->ilen, m * sizeof(PetscInt))); 4873 4874 /* allocate the matrix space */ 4875 if (mallocmatspace) { 4876 PetscCall(PetscMalloc3(a->i[m], &c->a, a->i[m], &c->j, m + 1, &c->i)); 4877 4878 c->singlemalloc = PETSC_TRUE; 4879 4880 PetscCall(PetscArraycpy(c->i, a->i, m + 1)); 4881 if (m > 0) { 4882 PetscCall(PetscArraycpy(c->j, a->j, a->i[m])); 4883 if (cpvalues == MAT_COPY_VALUES) { 4884 const PetscScalar *aa; 4885 4886 PetscCall(MatSeqAIJGetArrayRead(A, &aa)); 4887 PetscCall(PetscArraycpy(c->a, aa, a->i[m])); 4888 PetscCall(MatSeqAIJGetArrayRead(A, &aa)); 4889 } else { 4890 PetscCall(PetscArrayzero(c->a, a->i[m])); 4891 } 4892 } 4893 } 4894 4895 c->ignorezeroentries = a->ignorezeroentries; 4896 c->roworiented = a->roworiented; 4897 c->nonew = a->nonew; 4898 if (a->diag) { 4899 PetscCall(PetscMalloc1(m + 1, &c->diag)); 4900 PetscCall(PetscMemcpy(c->diag, a->diag, m * sizeof(PetscInt))); 4901 } else c->diag = NULL; 4902 4903 c->solve_work = NULL; 4904 c->saved_values = NULL; 4905 c->idiag = NULL; 4906 c->ssor_work = NULL; 4907 c->keepnonzeropattern = a->keepnonzeropattern; 4908 c->free_a = PETSC_TRUE; 4909 c->free_ij = PETSC_TRUE; 4910 4911 c->rmax = a->rmax; 4912 c->nz = a->nz; 4913 c->maxnz = a->nz; /* Since we allocate exactly the right amount */ 4914 4915 c->compressedrow.use = a->compressedrow.use; 4916 c->compressedrow.nrows = a->compressedrow.nrows; 4917 if (a->compressedrow.use) { 4918 i = a->compressedrow.nrows; 4919 PetscCall(PetscMalloc2(i + 1, &c->compressedrow.i, i, &c->compressedrow.rindex)); 4920 PetscCall(PetscArraycpy(c->compressedrow.i, a->compressedrow.i, i + 1)); 4921 PetscCall(PetscArraycpy(c->compressedrow.rindex, a->compressedrow.rindex, i)); 4922 } else { 4923 c->compressedrow.use = PETSC_FALSE; 4924 c->compressedrow.i = NULL; 4925 c->compressedrow.rindex = NULL; 4926 } 4927 c->nonzerorowcnt = a->nonzerorowcnt; 4928 C->nonzerostate = A->nonzerostate; 4929 4930 PetscCall(MatDuplicate_SeqAIJ_Inode(A, cpvalues, &C)); 4931 } 4932 PetscCall(PetscFunctionListDuplicate(((PetscObject)A)->qlist, &((PetscObject)C)->qlist)); 4933 PetscFunctionReturn(0); 4934 } 4935 4936 PetscErrorCode MatDuplicate_SeqAIJ(Mat A, MatDuplicateOption cpvalues, Mat *B) 4937 { 4938 PetscFunctionBegin; 4939 PetscCall(MatCreate(PetscObjectComm((PetscObject)A), B)); 4940 PetscCall(MatSetSizes(*B, A->rmap->n, A->cmap->n, A->rmap->n, A->cmap->n)); 4941 if (!(A->rmap->n % A->rmap->bs) && !(A->cmap->n % A->cmap->bs)) PetscCall(MatSetBlockSizesFromMats(*B, A, A)); 4942 PetscCall(MatSetType(*B, ((PetscObject)A)->type_name)); 4943 PetscCall(MatDuplicateNoCreate_SeqAIJ(*B, A, cpvalues, PETSC_TRUE)); 4944 PetscFunctionReturn(0); 4945 } 4946 4947 PetscErrorCode MatLoad_SeqAIJ(Mat newMat, PetscViewer viewer) 4948 { 4949 PetscBool isbinary, ishdf5; 4950 4951 PetscFunctionBegin; 4952 PetscValidHeaderSpecific(newMat, MAT_CLASSID, 1); 4953 PetscValidHeaderSpecific(viewer, PETSC_VIEWER_CLASSID, 2); 4954 /* force binary viewer to load .info file if it has not yet done so */ 4955 PetscCall(PetscViewerSetUp(viewer)); 4956 PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERBINARY, &isbinary)); 4957 PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERHDF5, &ishdf5)); 4958 if (isbinary) { 4959 PetscCall(MatLoad_SeqAIJ_Binary(newMat, viewer)); 4960 } else if (ishdf5) { 4961 #if defined(PETSC_HAVE_HDF5) 4962 PetscCall(MatLoad_AIJ_HDF5(newMat, viewer)); 4963 #else 4964 SETERRQ(PetscObjectComm((PetscObject)newMat), PETSC_ERR_SUP, "HDF5 not supported in this build.\nPlease reconfigure using --download-hdf5"); 4965 #endif 4966 } else { 4967 SETERRQ(PetscObjectComm((PetscObject)newMat), PETSC_ERR_SUP, "Viewer type %s not yet supported for reading %s matrices", ((PetscObject)viewer)->type_name, ((PetscObject)newMat)->type_name); 4968 } 4969 PetscFunctionReturn(0); 4970 } 4971 4972 PetscErrorCode MatLoad_SeqAIJ_Binary(Mat mat, PetscViewer viewer) 4973 { 4974 Mat_SeqAIJ *a = (Mat_SeqAIJ *)mat->data; 4975 PetscInt header[4], *rowlens, M, N, nz, sum, rows, cols, i; 4976 4977 PetscFunctionBegin; 4978 PetscCall(PetscViewerSetUp(viewer)); 4979 4980 /* read in matrix header */ 4981 PetscCall(PetscViewerBinaryRead(viewer, header, 4, NULL, PETSC_INT)); 4982 PetscCheck(header[0] == MAT_FILE_CLASSID, PETSC_COMM_SELF, PETSC_ERR_FILE_UNEXPECTED, "Not a matrix object in file"); 4983 M = header[1]; 4984 N = header[2]; 4985 nz = header[3]; 4986 PetscCheck(M >= 0, PetscObjectComm((PetscObject)viewer), PETSC_ERR_FILE_UNEXPECTED, "Matrix row size (%" PetscInt_FMT ") in file is negative", M); 4987 PetscCheck(N >= 0, PetscObjectComm((PetscObject)viewer), PETSC_ERR_FILE_UNEXPECTED, "Matrix column size (%" PetscInt_FMT ") in file is negative", N); 4988 PetscCheck(nz >= 0, PETSC_COMM_SELF, PETSC_ERR_FILE_UNEXPECTED, "Matrix stored in special format on disk, cannot load as SeqAIJ"); 4989 4990 /* set block sizes from the viewer's .info file */ 4991 PetscCall(MatLoad_Binary_BlockSizes(mat, viewer)); 4992 /* set local and global sizes if not set already */ 4993 if (mat->rmap->n < 0) mat->rmap->n = M; 4994 if (mat->cmap->n < 0) mat->cmap->n = N; 4995 if (mat->rmap->N < 0) mat->rmap->N = M; 4996 if (mat->cmap->N < 0) mat->cmap->N = N; 4997 PetscCall(PetscLayoutSetUp(mat->rmap)); 4998 PetscCall(PetscLayoutSetUp(mat->cmap)); 4999 5000 /* check if the matrix sizes are correct */ 5001 PetscCall(MatGetSize(mat, &rows, &cols)); 5002 PetscCheck(M == rows && N == cols, PETSC_COMM_SELF, PETSC_ERR_FILE_UNEXPECTED, "Matrix in file of different sizes (%" PetscInt_FMT ", %" PetscInt_FMT ") than the input matrix (%" PetscInt_FMT ", %" PetscInt_FMT ")", M, N, rows, cols); 5003 5004 /* read in row lengths */ 5005 PetscCall(PetscMalloc1(M, &rowlens)); 5006 PetscCall(PetscViewerBinaryRead(viewer, rowlens, M, NULL, PETSC_INT)); 5007 /* check if sum(rowlens) is same as nz */ 5008 sum = 0; 5009 for (i = 0; i < M; i++) sum += rowlens[i]; 5010 PetscCheck(sum == nz, PETSC_COMM_SELF, PETSC_ERR_FILE_UNEXPECTED, "Inconsistent matrix data in file: nonzeros = %" PetscInt_FMT ", sum-row-lengths = %" PetscInt_FMT, nz, sum); 5011 /* preallocate and check sizes */ 5012 PetscCall(MatSeqAIJSetPreallocation_SeqAIJ(mat, 0, rowlens)); 5013 PetscCall(MatGetSize(mat, &rows, &cols)); 5014 PetscCheck(M == rows && N == cols, PETSC_COMM_SELF, PETSC_ERR_FILE_UNEXPECTED, "Matrix in file of different length (%" PetscInt_FMT ", %" PetscInt_FMT ") than the input matrix (%" PetscInt_FMT ", %" PetscInt_FMT ")", M, N, rows, cols); 5015 /* store row lengths */ 5016 PetscCall(PetscArraycpy(a->ilen, rowlens, M)); 5017 PetscCall(PetscFree(rowlens)); 5018 5019 /* fill in "i" row pointers */ 5020 a->i[0] = 0; 5021 for (i = 0; i < M; i++) a->i[i + 1] = a->i[i] + a->ilen[i]; 5022 /* read in "j" column indices */ 5023 PetscCall(PetscViewerBinaryRead(viewer, a->j, nz, NULL, PETSC_INT)); 5024 /* read in "a" nonzero values */ 5025 PetscCall(PetscViewerBinaryRead(viewer, a->a, nz, NULL, PETSC_SCALAR)); 5026 5027 PetscCall(MatAssemblyBegin(mat, MAT_FINAL_ASSEMBLY)); 5028 PetscCall(MatAssemblyEnd(mat, MAT_FINAL_ASSEMBLY)); 5029 PetscFunctionReturn(0); 5030 } 5031 5032 PetscErrorCode MatEqual_SeqAIJ(Mat A, Mat B, PetscBool *flg) 5033 { 5034 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data, *b = (Mat_SeqAIJ *)B->data; 5035 const PetscScalar *aa, *ba; 5036 #if defined(PETSC_USE_COMPLEX) 5037 PetscInt k; 5038 #endif 5039 5040 PetscFunctionBegin; 5041 /* If the matrix dimensions are not equal,or no of nonzeros */ 5042 if ((A->rmap->n != B->rmap->n) || (A->cmap->n != B->cmap->n) || (a->nz != b->nz)) { 5043 *flg = PETSC_FALSE; 5044 PetscFunctionReturn(0); 5045 } 5046 5047 /* if the a->i are the same */ 5048 PetscCall(PetscArraycmp(a->i, b->i, A->rmap->n + 1, flg)); 5049 if (!*flg) PetscFunctionReturn(0); 5050 5051 /* if a->j are the same */ 5052 PetscCall(PetscArraycmp(a->j, b->j, a->nz, flg)); 5053 if (!*flg) PetscFunctionReturn(0); 5054 5055 PetscCall(MatSeqAIJGetArrayRead(A, &aa)); 5056 PetscCall(MatSeqAIJGetArrayRead(B, &ba)); 5057 /* if a->a are the same */ 5058 #if defined(PETSC_USE_COMPLEX) 5059 for (k = 0; k < a->nz; k++) { 5060 if (PetscRealPart(aa[k]) != PetscRealPart(ba[k]) || PetscImaginaryPart(aa[k]) != PetscImaginaryPart(ba[k])) { 5061 *flg = PETSC_FALSE; 5062 PetscFunctionReturn(0); 5063 } 5064 } 5065 #else 5066 PetscCall(PetscArraycmp(aa, ba, a->nz, flg)); 5067 #endif 5068 PetscCall(MatSeqAIJRestoreArrayRead(A, &aa)); 5069 PetscCall(MatSeqAIJRestoreArrayRead(B, &ba)); 5070 PetscFunctionReturn(0); 5071 } 5072 5073 /*@ 5074 MatCreateSeqAIJWithArrays - Creates an sequential `MATSEQAIJ` matrix using matrix elements (in CSR format) 5075 provided by the user. 5076 5077 Collective 5078 5079 Input Parameters: 5080 + comm - must be an MPI communicator of size 1 5081 . m - number of rows 5082 . n - number of columns 5083 . i - row indices; that is i[0] = 0, i[row] = i[row-1] + number of elements in that row of the matrix 5084 . j - column indices 5085 - a - matrix values 5086 5087 Output Parameter: 5088 . mat - the matrix 5089 5090 Level: intermediate 5091 5092 Notes: 5093 The i, j, and a arrays are not copied by this routine, the user must free these arrays 5094 once the matrix is destroyed and not before 5095 5096 You cannot set new nonzero locations into this matrix, that will generate an error. 5097 5098 The i and j indices are 0 based 5099 5100 The format which is used for the sparse matrix input, is equivalent to a 5101 row-major ordering.. i.e for the following matrix, the input data expected is 5102 as shown 5103 5104 $ 1 0 0 5105 $ 2 0 3 5106 $ 4 5 6 5107 $ 5108 $ i = {0,1,3,6} [size = nrow+1 = 3+1] 5109 $ j = {0,0,2,0,1,2} [size = 6]; values must be sorted for each row 5110 $ v = {1,2,3,4,5,6} [size = 6] 5111 5112 .seealso: `MatCreate()`, `MatCreateAIJ()`, `MatCreateSeqAIJ()`, `MatCreateMPIAIJWithArrays()`, `MatMPIAIJSetPreallocationCSR()` 5113 @*/ 5114 PetscErrorCode MatCreateSeqAIJWithArrays(MPI_Comm comm, PetscInt m, PetscInt n, PetscInt i[], PetscInt j[], PetscScalar a[], Mat *mat) 5115 { 5116 PetscInt ii; 5117 Mat_SeqAIJ *aij; 5118 PetscInt jj; 5119 5120 PetscFunctionBegin; 5121 PetscCheck(m <= 0 || i[0] == 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "i (row indices) must start with 0"); 5122 PetscCall(MatCreate(comm, mat)); 5123 PetscCall(MatSetSizes(*mat, m, n, m, n)); 5124 /* PetscCall(MatSetBlockSizes(*mat,,)); */ 5125 PetscCall(MatSetType(*mat, MATSEQAIJ)); 5126 PetscCall(MatSeqAIJSetPreallocation_SeqAIJ(*mat, MAT_SKIP_ALLOCATION, NULL)); 5127 aij = (Mat_SeqAIJ *)(*mat)->data; 5128 PetscCall(PetscMalloc1(m, &aij->imax)); 5129 PetscCall(PetscMalloc1(m, &aij->ilen)); 5130 5131 aij->i = i; 5132 aij->j = j; 5133 aij->a = a; 5134 aij->singlemalloc = PETSC_FALSE; 5135 aij->nonew = -1; /*this indicates that inserting a new value in the matrix that generates a new nonzero is an error*/ 5136 aij->free_a = PETSC_FALSE; 5137 aij->free_ij = PETSC_FALSE; 5138 5139 for (ii = 0, aij->nonzerorowcnt = 0, aij->rmax = 0; ii < m; ii++) { 5140 aij->ilen[ii] = aij->imax[ii] = i[ii + 1] - i[ii]; 5141 if (PetscDefined(USE_DEBUG)) { 5142 PetscCheck(i[ii + 1] - i[ii] >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Negative row length in i (row indices) row = %" PetscInt_FMT " length = %" PetscInt_FMT, ii, i[ii + 1] - i[ii]); 5143 for (jj = i[ii] + 1; jj < i[ii + 1]; jj++) { 5144 PetscCheck(j[jj] >= j[jj - 1], PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Column entry number %" PetscInt_FMT " (actual column %" PetscInt_FMT ") in row %" PetscInt_FMT " is not sorted", jj - i[ii], j[jj], ii); 5145 PetscCheck(j[jj] != j[jj - 1], PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Column entry number %" PetscInt_FMT " (actual column %" PetscInt_FMT ") in row %" PetscInt_FMT " is identical to previous entry", jj - i[ii], j[jj], ii); 5146 } 5147 } 5148 } 5149 if (PetscDefined(USE_DEBUG)) { 5150 for (ii = 0; ii < aij->i[m]; ii++) { 5151 PetscCheck(j[ii] >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Negative column index at location = %" PetscInt_FMT " index = %" PetscInt_FMT, ii, j[ii]); 5152 PetscCheck(j[ii] <= n - 1, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Column index to large at location = %" PetscInt_FMT " index = %" PetscInt_FMT, ii, j[ii]); 5153 } 5154 } 5155 5156 PetscCall(MatAssemblyBegin(*mat, MAT_FINAL_ASSEMBLY)); 5157 PetscCall(MatAssemblyEnd(*mat, MAT_FINAL_ASSEMBLY)); 5158 PetscFunctionReturn(0); 5159 } 5160 5161 /*@ 5162 MatCreateSeqAIJFromTriple - Creates an sequential `MATSEQAIJ` matrix using matrix elements (in COO format) 5163 provided by the user. 5164 5165 Collective 5166 5167 Input Parameters: 5168 + comm - must be an MPI communicator of size 1 5169 . m - number of rows 5170 . n - number of columns 5171 . i - row indices 5172 . j - column indices 5173 . a - matrix values 5174 . nz - number of nonzeros 5175 - idx - if the i and j indices start with 1 use `PETSC_TRUE` otherwise use `PETSC_FALSE` 5176 5177 Output Parameter: 5178 . mat - the matrix 5179 5180 Level: intermediate 5181 5182 Example: 5183 For the following matrix, the input data expected is as shown (using 0 based indexing) 5184 .vb 5185 1 0 0 5186 2 0 3 5187 4 5 6 5188 5189 i = {0,1,1,2,2,2} 5190 j = {0,0,2,0,1,2} 5191 v = {1,2,3,4,5,6} 5192 .ve 5193 Notes: 5194 Instead of using this function, users should also consider `MatSetPreallocationCOO()` and `MatSetValuesCOO()`, which allow repeated or remote entries, 5195 and are particularly useful in iterative applications. 5196 5197 .seealso: `MatCreate()`, `MatCreateAIJ()`, `MatCreateSeqAIJ()`, `MatCreateSeqAIJWithArrays()`, `MatMPIAIJSetPreallocationCSR()`, `MatSetValuesCOO()`, `MatSetPreallocationCOO()` 5198 @*/ 5199 PetscErrorCode MatCreateSeqAIJFromTriple(MPI_Comm comm, PetscInt m, PetscInt n, PetscInt i[], PetscInt j[], PetscScalar a[], Mat *mat, PetscInt nz, PetscBool idx) 5200 { 5201 PetscInt ii, *nnz, one = 1, row, col; 5202 5203 PetscFunctionBegin; 5204 PetscCall(PetscCalloc1(m, &nnz)); 5205 for (ii = 0; ii < nz; ii++) nnz[i[ii] - !!idx] += 1; 5206 PetscCall(MatCreate(comm, mat)); 5207 PetscCall(MatSetSizes(*mat, m, n, m, n)); 5208 PetscCall(MatSetType(*mat, MATSEQAIJ)); 5209 PetscCall(MatSeqAIJSetPreallocation_SeqAIJ(*mat, 0, nnz)); 5210 for (ii = 0; ii < nz; ii++) { 5211 if (idx) { 5212 row = i[ii] - 1; 5213 col = j[ii] - 1; 5214 } else { 5215 row = i[ii]; 5216 col = j[ii]; 5217 } 5218 PetscCall(MatSetValues(*mat, one, &row, one, &col, &a[ii], ADD_VALUES)); 5219 } 5220 PetscCall(MatAssemblyBegin(*mat, MAT_FINAL_ASSEMBLY)); 5221 PetscCall(MatAssemblyEnd(*mat, MAT_FINAL_ASSEMBLY)); 5222 PetscCall(PetscFree(nnz)); 5223 PetscFunctionReturn(0); 5224 } 5225 5226 PetscErrorCode MatSeqAIJInvalidateDiagonal(Mat A) 5227 { 5228 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 5229 5230 PetscFunctionBegin; 5231 a->idiagvalid = PETSC_FALSE; 5232 a->ibdiagvalid = PETSC_FALSE; 5233 5234 PetscCall(MatSeqAIJInvalidateDiagonal_Inode(A)); 5235 PetscFunctionReturn(0); 5236 } 5237 5238 PetscErrorCode MatCreateMPIMatConcatenateSeqMat_SeqAIJ(MPI_Comm comm, Mat inmat, PetscInt n, MatReuse scall, Mat *outmat) 5239 { 5240 PetscFunctionBegin; 5241 PetscCall(MatCreateMPIMatConcatenateSeqMat_MPIAIJ(comm, inmat, n, scall, outmat)); 5242 PetscFunctionReturn(0); 5243 } 5244 5245 /* 5246 Permute A into C's *local* index space using rowemb,colemb. 5247 The embedding are supposed to be injections and the above implies that the range of rowemb is a subset 5248 of [0,m), colemb is in [0,n). 5249 If pattern == DIFFERENT_NONZERO_PATTERN, C is preallocated according to A. 5250 */ 5251 PetscErrorCode MatSetSeqMat_SeqAIJ(Mat C, IS rowemb, IS colemb, MatStructure pattern, Mat B) 5252 { 5253 /* If making this function public, change the error returned in this function away from _PLIB. */ 5254 Mat_SeqAIJ *Baij; 5255 PetscBool seqaij; 5256 PetscInt m, n, *nz, i, j, count; 5257 PetscScalar v; 5258 const PetscInt *rowindices, *colindices; 5259 5260 PetscFunctionBegin; 5261 if (!B) PetscFunctionReturn(0); 5262 /* Check to make sure the target matrix (and embeddings) are compatible with C and each other. */ 5263 PetscCall(PetscObjectBaseTypeCompare((PetscObject)B, MATSEQAIJ, &seqaij)); 5264 PetscCheck(seqaij, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Input matrix is of wrong type"); 5265 if (rowemb) { 5266 PetscCall(ISGetLocalSize(rowemb, &m)); 5267 PetscCheck(m == B->rmap->n, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Row IS of size %" PetscInt_FMT " is incompatible with matrix row size %" PetscInt_FMT, m, B->rmap->n); 5268 } else { 5269 PetscCheck(C->rmap->n == B->rmap->n, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Input matrix is row-incompatible with the target matrix"); 5270 } 5271 if (colemb) { 5272 PetscCall(ISGetLocalSize(colemb, &n)); 5273 PetscCheck(n == B->cmap->n, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Diag col IS of size %" PetscInt_FMT " is incompatible with input matrix col size %" PetscInt_FMT, n, B->cmap->n); 5274 } else { 5275 PetscCheck(C->cmap->n == B->cmap->n, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Input matrix is col-incompatible with the target matrix"); 5276 } 5277 5278 Baij = (Mat_SeqAIJ *)(B->data); 5279 if (pattern == DIFFERENT_NONZERO_PATTERN) { 5280 PetscCall(PetscMalloc1(B->rmap->n, &nz)); 5281 for (i = 0; i < B->rmap->n; i++) nz[i] = Baij->i[i + 1] - Baij->i[i]; 5282 PetscCall(MatSeqAIJSetPreallocation(C, 0, nz)); 5283 PetscCall(PetscFree(nz)); 5284 } 5285 if (pattern == SUBSET_NONZERO_PATTERN) PetscCall(MatZeroEntries(C)); 5286 count = 0; 5287 rowindices = NULL; 5288 colindices = NULL; 5289 if (rowemb) PetscCall(ISGetIndices(rowemb, &rowindices)); 5290 if (colemb) PetscCall(ISGetIndices(colemb, &colindices)); 5291 for (i = 0; i < B->rmap->n; i++) { 5292 PetscInt row; 5293 row = i; 5294 if (rowindices) row = rowindices[i]; 5295 for (j = Baij->i[i]; j < Baij->i[i + 1]; j++) { 5296 PetscInt col; 5297 col = Baij->j[count]; 5298 if (colindices) col = colindices[col]; 5299 v = Baij->a[count]; 5300 PetscCall(MatSetValues(C, 1, &row, 1, &col, &v, INSERT_VALUES)); 5301 ++count; 5302 } 5303 } 5304 /* FIXME: set C's nonzerostate correctly. */ 5305 /* Assembly for C is necessary. */ 5306 C->preallocated = PETSC_TRUE; 5307 C->assembled = PETSC_TRUE; 5308 C->was_assembled = PETSC_FALSE; 5309 PetscFunctionReturn(0); 5310 } 5311 5312 PetscErrorCode MatEliminateZeros_SeqAIJ(Mat A) 5313 { 5314 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 5315 MatScalar *aa = a->a; 5316 PetscInt m = A->rmap->n, fshift = 0, fshift_prev = 0, i, k; 5317 PetscInt *ailen = a->ilen, *imax = a->imax, *ai = a->i, *aj = a->j, rmax = 0; 5318 5319 PetscFunctionBegin; 5320 PetscCheck(A->assembled, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Cannot eliminate zeros for unassembled matrix"); 5321 if (m) rmax = ailen[0]; /* determine row with most nonzeros */ 5322 for (i = 1; i <= m; i++) { 5323 /* move each nonzero entry back by the amount of zero slots (fshift) before it*/ 5324 for (k = ai[i - 1]; k < ai[i]; k++) { 5325 if (aa[k] == 0 && aj[k] != i - 1) fshift++; 5326 else { 5327 if (aa[k] == 0 && aj[k] == i - 1) PetscCall(PetscInfo(A, "Keep the diagonal zero at row %" PetscInt_FMT "\n", i - 1)); 5328 aa[k - fshift] = aa[k]; 5329 aj[k - fshift] = aj[k]; 5330 } 5331 } 5332 ai[i - 1] -= fshift_prev; // safe to update ai[i-1] now since it will not be used in the next iteration 5333 fshift_prev = fshift; 5334 /* reset ilen and imax for each row */ 5335 ailen[i - 1] = imax[i - 1] = ai[i] - fshift - ai[i - 1]; 5336 a->nonzerorowcnt += ((ai[i] - fshift - ai[i - 1]) > 0); 5337 rmax = PetscMax(rmax, ailen[i - 1]); 5338 } 5339 if (m) { 5340 ai[m] -= fshift; 5341 a->nz = ai[m]; 5342 } 5343 PetscCall(PetscInfo(A, "Matrix size: %" PetscInt_FMT " X %" PetscInt_FMT "; zeros eliminated: %" PetscInt_FMT "; nonzeros left: %" PetscInt_FMT "\n", m, A->cmap->n, fshift, a->nz)); 5344 A->nonzerostate -= fshift; 5345 A->info.nz_unneeded += (PetscReal)fshift; 5346 a->rmax = rmax; 5347 if (a->inode.use && a->inode.checked) PetscCall(MatSeqAIJCheckInode(A)); 5348 PetscCall(MatAssemblyBegin(A, MAT_FINAL_ASSEMBLY)); 5349 PetscCall(MatAssemblyEnd(A, MAT_FINAL_ASSEMBLY)); 5350 PetscFunctionReturn(0); 5351 } 5352 5353 PetscFunctionList MatSeqAIJList = NULL; 5354 5355 /*@C 5356 MatSeqAIJSetType - Converts a `MATSEQAIJ` matrix to a subtype 5357 5358 Collective 5359 5360 Input Parameters: 5361 + mat - the matrix object 5362 - matype - matrix type 5363 5364 Options Database Key: 5365 . -mat_seqai_type <method> - for example seqaijcrl 5366 5367 Level: intermediate 5368 5369 .seealso: `PCSetType()`, `VecSetType()`, `MatCreate()`, `MatType`, `Mat` 5370 @*/ 5371 PetscErrorCode MatSeqAIJSetType(Mat mat, MatType matype) 5372 { 5373 PetscBool sametype; 5374 PetscErrorCode (*r)(Mat, MatType, MatReuse, Mat *); 5375 5376 PetscFunctionBegin; 5377 PetscValidHeaderSpecific(mat, MAT_CLASSID, 1); 5378 PetscCall(PetscObjectTypeCompare((PetscObject)mat, matype, &sametype)); 5379 if (sametype) PetscFunctionReturn(0); 5380 5381 PetscCall(PetscFunctionListFind(MatSeqAIJList, matype, &r)); 5382 PetscCheck(r, PETSC_COMM_SELF, PETSC_ERR_ARG_UNKNOWN_TYPE, "Unknown Mat type given: %s", matype); 5383 PetscCall((*r)(mat, matype, MAT_INPLACE_MATRIX, &mat)); 5384 PetscFunctionReturn(0); 5385 } 5386 5387 /*@C 5388 MatSeqAIJRegister - - Adds a new sub-matrix type for sequential `MATSEQAIJ` matrices 5389 5390 Not Collective 5391 5392 Input Parameters: 5393 + name - name of a new user-defined matrix type, for example `MATSEQAIJCRL` 5394 - function - routine to convert to subtype 5395 5396 Notes: 5397 `MatSeqAIJRegister()` may be called multiple times to add several user-defined solvers. 5398 5399 Then, your matrix can be chosen with the procedural interface at runtime via the option 5400 $ -mat_seqaij_type my_mat 5401 5402 Level: advanced 5403 5404 .seealso: `MatSeqAIJRegisterAll()` 5405 @*/ 5406 PetscErrorCode MatSeqAIJRegister(const char sname[], PetscErrorCode (*function)(Mat, MatType, MatReuse, Mat *)) 5407 { 5408 PetscFunctionBegin; 5409 PetscCall(MatInitializePackage()); 5410 PetscCall(PetscFunctionListAdd(&MatSeqAIJList, sname, function)); 5411 PetscFunctionReturn(0); 5412 } 5413 5414 PetscBool MatSeqAIJRegisterAllCalled = PETSC_FALSE; 5415 5416 /*@C 5417 MatSeqAIJRegisterAll - Registers all of the matrix subtypes of `MATSSEQAIJ` 5418 5419 Not Collective 5420 5421 Level: advanced 5422 5423 .seealso: `MatRegisterAll()`, `MatSeqAIJRegister()` 5424 @*/ 5425 PetscErrorCode MatSeqAIJRegisterAll(void) 5426 { 5427 PetscFunctionBegin; 5428 if (MatSeqAIJRegisterAllCalled) PetscFunctionReturn(0); 5429 MatSeqAIJRegisterAllCalled = PETSC_TRUE; 5430 5431 PetscCall(MatSeqAIJRegister(MATSEQAIJCRL, MatConvert_SeqAIJ_SeqAIJCRL)); 5432 PetscCall(MatSeqAIJRegister(MATSEQAIJPERM, MatConvert_SeqAIJ_SeqAIJPERM)); 5433 PetscCall(MatSeqAIJRegister(MATSEQAIJSELL, MatConvert_SeqAIJ_SeqAIJSELL)); 5434 #if defined(PETSC_HAVE_MKL_SPARSE) 5435 PetscCall(MatSeqAIJRegister(MATSEQAIJMKL, MatConvert_SeqAIJ_SeqAIJMKL)); 5436 #endif 5437 #if defined(PETSC_HAVE_CUDA) 5438 PetscCall(MatSeqAIJRegister(MATSEQAIJCUSPARSE, MatConvert_SeqAIJ_SeqAIJCUSPARSE)); 5439 #endif 5440 #if defined(PETSC_HAVE_HIP) 5441 PetscCall(MatSeqAIJRegister(MATSEQAIJHIPSPARSE, MatConvert_SeqAIJ_SeqAIJHIPSPARSE)); 5442 #endif 5443 #if defined(PETSC_HAVE_KOKKOS_KERNELS) 5444 PetscCall(MatSeqAIJRegister(MATSEQAIJKOKKOS, MatConvert_SeqAIJ_SeqAIJKokkos)); 5445 #endif 5446 #if defined(PETSC_HAVE_VIENNACL) && defined(PETSC_HAVE_VIENNACL_NO_CUDA) 5447 PetscCall(MatSeqAIJRegister(MATMPIAIJVIENNACL, MatConvert_SeqAIJ_SeqAIJViennaCL)); 5448 #endif 5449 PetscFunctionReturn(0); 5450 } 5451 5452 /* 5453 Special version for direct calls from Fortran 5454 */ 5455 #include <petsc/private/fortranimpl.h> 5456 #if defined(PETSC_HAVE_FORTRAN_CAPS) 5457 #define matsetvaluesseqaij_ MATSETVALUESSEQAIJ 5458 #elif !defined(PETSC_HAVE_FORTRAN_UNDERSCORE) 5459 #define matsetvaluesseqaij_ matsetvaluesseqaij 5460 #endif 5461 5462 /* Change these macros so can be used in void function */ 5463 5464 /* Change these macros so can be used in void function */ 5465 /* Identical to PetscCallVoid, except it assigns to *_ierr */ 5466 #undef PetscCall 5467 #define PetscCall(...) \ 5468 do { \ 5469 PetscErrorCode ierr_msv_mpiaij = __VA_ARGS__; \ 5470 if (PetscUnlikely(ierr_msv_mpiaij)) { \ 5471 *_ierr = PetscError(PETSC_COMM_SELF, __LINE__, PETSC_FUNCTION_NAME, __FILE__, ierr_msv_mpiaij, PETSC_ERROR_REPEAT, " "); \ 5472 return; \ 5473 } \ 5474 } while (0) 5475 5476 #undef SETERRQ 5477 #define SETERRQ(comm, ierr, ...) \ 5478 do { \ 5479 *_ierr = PetscError(comm, __LINE__, PETSC_FUNCTION_NAME, __FILE__, ierr, PETSC_ERROR_INITIAL, __VA_ARGS__); \ 5480 return; \ 5481 } while (0) 5482 5483 PETSC_EXTERN void matsetvaluesseqaij_(Mat *AA, PetscInt *mm, const PetscInt im[], PetscInt *nn, const PetscInt in[], const PetscScalar v[], InsertMode *isis, PetscErrorCode *_ierr) 5484 { 5485 Mat A = *AA; 5486 PetscInt m = *mm, n = *nn; 5487 InsertMode is = *isis; 5488 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 5489 PetscInt *rp, k, low, high, t, ii, row, nrow, i, col, l, rmax, N; 5490 PetscInt *imax, *ai, *ailen; 5491 PetscInt *aj, nonew = a->nonew, lastcol = -1; 5492 MatScalar *ap, value, *aa; 5493 PetscBool ignorezeroentries = a->ignorezeroentries; 5494 PetscBool roworiented = a->roworiented; 5495 5496 PetscFunctionBegin; 5497 MatCheckPreallocated(A, 1); 5498 imax = a->imax; 5499 ai = a->i; 5500 ailen = a->ilen; 5501 aj = a->j; 5502 aa = a->a; 5503 5504 for (k = 0; k < m; k++) { /* loop over added rows */ 5505 row = im[k]; 5506 if (row < 0) continue; 5507 PetscCheck(row < A->rmap->n, PetscObjectComm((PetscObject)A), PETSC_ERR_ARG_OUTOFRANGE, "Row too large"); 5508 rp = aj + ai[row]; 5509 ap = aa + ai[row]; 5510 rmax = imax[row]; 5511 nrow = ailen[row]; 5512 low = 0; 5513 high = nrow; 5514 for (l = 0; l < n; l++) { /* loop over added columns */ 5515 if (in[l] < 0) continue; 5516 PetscCheck(in[l] < A->cmap->n, PetscObjectComm((PetscObject)A), PETSC_ERR_ARG_OUTOFRANGE, "Column too large"); 5517 col = in[l]; 5518 if (roworiented) value = v[l + k * n]; 5519 else value = v[k + l * m]; 5520 5521 if (value == 0.0 && ignorezeroentries && (is == ADD_VALUES)) continue; 5522 5523 if (col <= lastcol) low = 0; 5524 else high = nrow; 5525 lastcol = col; 5526 while (high - low > 5) { 5527 t = (low + high) / 2; 5528 if (rp[t] > col) high = t; 5529 else low = t; 5530 } 5531 for (i = low; i < high; i++) { 5532 if (rp[i] > col) break; 5533 if (rp[i] == col) { 5534 if (is == ADD_VALUES) ap[i] += value; 5535 else ap[i] = value; 5536 goto noinsert; 5537 } 5538 } 5539 if (value == 0.0 && ignorezeroentries) goto noinsert; 5540 if (nonew == 1) goto noinsert; 5541 PetscCheck(nonew != -1, PetscObjectComm((PetscObject)A), PETSC_ERR_ARG_OUTOFRANGE, "Inserting a new nonzero in the matrix"); 5542 MatSeqXAIJReallocateAIJ(A, A->rmap->n, 1, nrow, row, col, rmax, aa, ai, aj, rp, ap, imax, nonew, MatScalar); 5543 N = nrow++ - 1; 5544 a->nz++; 5545 high++; 5546 /* shift up all the later entries in this row */ 5547 for (ii = N; ii >= i; ii--) { 5548 rp[ii + 1] = rp[ii]; 5549 ap[ii + 1] = ap[ii]; 5550 } 5551 rp[i] = col; 5552 ap[i] = value; 5553 A->nonzerostate++; 5554 noinsert:; 5555 low = i + 1; 5556 } 5557 ailen[row] = nrow; 5558 } 5559 PetscFunctionReturnVoid(); 5560 } 5561 /* Undefining these here since they were redefined from their original definition above! No 5562 * other PETSc functions should be defined past this point, as it is impossible to recover the 5563 * original definitions */ 5564 #undef PetscCall 5565 #undef SETERRQ 5566