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 /* defines MatSetValues_Seq_Hash(), MatAssemblyEnd_Seq_Hash(), MatSetUp_Seq_Hash() */ 12 #define TYPE AIJ 13 #define TYPE_BS 14 #include "../src/mat/impls/aij/seq/seqhashmatsetvalues.h" 15 #include "../src/mat/impls/aij/seq/seqhashmat.h" 16 #undef TYPE 17 #undef TYPE_BS 18 19 static PetscErrorCode MatSeqAIJSetTypeFromOptions(Mat A) 20 { 21 PetscBool flg; 22 char type[256]; 23 24 PetscFunctionBegin; 25 PetscObjectOptionsBegin((PetscObject)A); 26 PetscCall(PetscOptionsFList("-mat_seqaij_type", "Matrix SeqAIJ type", "MatSeqAIJSetType", MatSeqAIJList, "seqaij", type, 256, &flg)); 27 if (flg) PetscCall(MatSeqAIJSetType(A, type)); 28 PetscOptionsEnd(); 29 PetscFunctionReturn(PETSC_SUCCESS); 30 } 31 32 static PetscErrorCode MatGetColumnReductions_SeqAIJ(Mat A, PetscInt type, PetscReal *reductions) 33 { 34 PetscInt i, m, n; 35 Mat_SeqAIJ *aij = (Mat_SeqAIJ *)A->data; 36 37 PetscFunctionBegin; 38 PetscCall(MatGetSize(A, &m, &n)); 39 PetscCall(PetscArrayzero(reductions, n)); 40 if (type == NORM_2) { 41 for (i = 0; i < aij->i[m]; i++) reductions[aij->j[i]] += PetscAbsScalar(aij->a[i] * aij->a[i]); 42 } else if (type == NORM_1) { 43 for (i = 0; i < aij->i[m]; i++) reductions[aij->j[i]] += PetscAbsScalar(aij->a[i]); 44 } else if (type == NORM_INFINITY) { 45 for (i = 0; i < aij->i[m]; i++) reductions[aij->j[i]] = PetscMax(PetscAbsScalar(aij->a[i]), reductions[aij->j[i]]); 46 } else if (type == REDUCTION_SUM_REALPART || type == REDUCTION_MEAN_REALPART) { 47 for (i = 0; i < aij->i[m]; i++) reductions[aij->j[i]] += PetscRealPart(aij->a[i]); 48 } else if (type == REDUCTION_SUM_IMAGINARYPART || type == REDUCTION_MEAN_IMAGINARYPART) { 49 for (i = 0; i < aij->i[m]; i++) reductions[aij->j[i]] += PetscImaginaryPart(aij->a[i]); 50 } else SETERRQ(PetscObjectComm((PetscObject)A), PETSC_ERR_ARG_WRONG, "Unknown reduction type"); 51 52 if (type == NORM_2) { 53 for (i = 0; i < n; i++) reductions[i] = PetscSqrtReal(reductions[i]); 54 } else if (type == REDUCTION_MEAN_REALPART || type == REDUCTION_MEAN_IMAGINARYPART) { 55 for (i = 0; i < n; i++) reductions[i] /= m; 56 } 57 PetscFunctionReturn(PETSC_SUCCESS); 58 } 59 60 static PetscErrorCode MatFindOffBlockDiagonalEntries_SeqAIJ(Mat A, IS *is) 61 { 62 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 63 PetscInt i, m = A->rmap->n, cnt = 0, bs = A->rmap->bs; 64 const PetscInt *jj = a->j, *ii = a->i; 65 PetscInt *rows; 66 67 PetscFunctionBegin; 68 for (i = 0; i < m; i++) { 69 if ((ii[i] != ii[i + 1]) && ((jj[ii[i]] < bs * (i / bs)) || (jj[ii[i + 1] - 1] > bs * ((i + bs) / bs) - 1))) cnt++; 70 } 71 PetscCall(PetscMalloc1(cnt, &rows)); 72 cnt = 0; 73 for (i = 0; i < m; i++) { 74 if ((ii[i] != ii[i + 1]) && ((jj[ii[i]] < bs * (i / bs)) || (jj[ii[i + 1] - 1] > bs * ((i + bs) / bs) - 1))) { 75 rows[cnt] = i; 76 cnt++; 77 } 78 } 79 PetscCall(ISCreateGeneral(PETSC_COMM_SELF, cnt, rows, PETSC_OWN_POINTER, is)); 80 PetscFunctionReturn(PETSC_SUCCESS); 81 } 82 83 PetscErrorCode MatFindZeroDiagonals_SeqAIJ_Private(Mat A, PetscInt *nrows, PetscInt **zrows) 84 { 85 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 86 const MatScalar *aa; 87 PetscInt i, m = A->rmap->n, cnt = 0; 88 const PetscInt *ii = a->i, *jj = a->j, *diag; 89 PetscInt *rows; 90 91 PetscFunctionBegin; 92 PetscCall(MatSeqAIJGetArrayRead(A, &aa)); 93 PetscCall(MatMarkDiagonal_SeqAIJ(A)); 94 diag = a->diag; 95 for (i = 0; i < m; i++) { 96 if ((diag[i] >= ii[i + 1]) || (jj[diag[i]] != i) || (aa[diag[i]] == 0.0)) cnt++; 97 } 98 PetscCall(PetscMalloc1(cnt, &rows)); 99 cnt = 0; 100 for (i = 0; i < m; i++) { 101 if ((diag[i] >= ii[i + 1]) || (jj[diag[i]] != i) || (aa[diag[i]] == 0.0)) rows[cnt++] = i; 102 } 103 *nrows = cnt; 104 *zrows = rows; 105 PetscCall(MatSeqAIJRestoreArrayRead(A, &aa)); 106 PetscFunctionReturn(PETSC_SUCCESS); 107 } 108 109 static PetscErrorCode MatFindZeroDiagonals_SeqAIJ(Mat A, IS *zrows) 110 { 111 PetscInt nrows, *rows; 112 113 PetscFunctionBegin; 114 *zrows = NULL; 115 PetscCall(MatFindZeroDiagonals_SeqAIJ_Private(A, &nrows, &rows)); 116 PetscCall(ISCreateGeneral(PetscObjectComm((PetscObject)A), nrows, rows, PETSC_OWN_POINTER, zrows)); 117 PetscFunctionReturn(PETSC_SUCCESS); 118 } 119 120 static PetscErrorCode MatFindNonzeroRows_SeqAIJ(Mat A, IS *keptrows) 121 { 122 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 123 const MatScalar *aa; 124 PetscInt m = A->rmap->n, cnt = 0; 125 const PetscInt *ii; 126 PetscInt n, i, j, *rows; 127 128 PetscFunctionBegin; 129 PetscCall(MatSeqAIJGetArrayRead(A, &aa)); 130 *keptrows = NULL; 131 ii = a->i; 132 for (i = 0; i < m; i++) { 133 n = ii[i + 1] - ii[i]; 134 if (!n) { 135 cnt++; 136 goto ok1; 137 } 138 for (j = ii[i]; j < ii[i + 1]; j++) { 139 if (aa[j] != 0.0) goto ok1; 140 } 141 cnt++; 142 ok1:; 143 } 144 if (!cnt) { 145 PetscCall(MatSeqAIJRestoreArrayRead(A, &aa)); 146 PetscFunctionReturn(PETSC_SUCCESS); 147 } 148 PetscCall(PetscMalloc1(A->rmap->n - cnt, &rows)); 149 cnt = 0; 150 for (i = 0; i < m; i++) { 151 n = ii[i + 1] - ii[i]; 152 if (!n) continue; 153 for (j = ii[i]; j < ii[i + 1]; j++) { 154 if (aa[j] != 0.0) { 155 rows[cnt++] = i; 156 break; 157 } 158 } 159 } 160 PetscCall(MatSeqAIJRestoreArrayRead(A, &aa)); 161 PetscCall(ISCreateGeneral(PETSC_COMM_SELF, cnt, rows, PETSC_OWN_POINTER, keptrows)); 162 PetscFunctionReturn(PETSC_SUCCESS); 163 } 164 165 PetscErrorCode MatDiagonalSet_SeqAIJ(Mat Y, Vec D, InsertMode is) 166 { 167 Mat_SeqAIJ *aij = (Mat_SeqAIJ *)Y->data; 168 PetscInt i, m = Y->rmap->n; 169 const PetscInt *diag; 170 MatScalar *aa; 171 const PetscScalar *v; 172 PetscBool missing; 173 174 PetscFunctionBegin; 175 if (Y->assembled) { 176 PetscCall(MatMissingDiagonal_SeqAIJ(Y, &missing, NULL)); 177 if (!missing) { 178 diag = aij->diag; 179 PetscCall(VecGetArrayRead(D, &v)); 180 PetscCall(MatSeqAIJGetArray(Y, &aa)); 181 if (is == INSERT_VALUES) { 182 for (i = 0; i < m; i++) aa[diag[i]] = v[i]; 183 } else { 184 for (i = 0; i < m; i++) aa[diag[i]] += v[i]; 185 } 186 PetscCall(MatSeqAIJRestoreArray(Y, &aa)); 187 PetscCall(VecRestoreArrayRead(D, &v)); 188 PetscFunctionReturn(PETSC_SUCCESS); 189 } 190 PetscCall(MatSeqAIJInvalidateDiagonal(Y)); 191 } 192 PetscCall(MatDiagonalSet_Default(Y, D, is)); 193 PetscFunctionReturn(PETSC_SUCCESS); 194 } 195 196 PetscErrorCode MatGetRowIJ_SeqAIJ(Mat A, PetscInt oshift, PetscBool symmetric, PetscBool inodecompressed, PetscInt *m, const PetscInt *ia[], const PetscInt *ja[], PetscBool *done) 197 { 198 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 199 PetscInt i, ishift; 200 201 PetscFunctionBegin; 202 if (m) *m = A->rmap->n; 203 if (!ia) PetscFunctionReturn(PETSC_SUCCESS); 204 ishift = 0; 205 if (symmetric && A->structurally_symmetric != PETSC_BOOL3_TRUE) { 206 PetscCall(MatToSymmetricIJ_SeqAIJ(A->rmap->n, a->i, a->j, PETSC_TRUE, ishift, oshift, (PetscInt **)ia, (PetscInt **)ja)); 207 } else if (oshift == 1) { 208 PetscInt *tia; 209 PetscInt nz = a->i[A->rmap->n]; 210 /* malloc space and add 1 to i and j indices */ 211 PetscCall(PetscMalloc1(A->rmap->n + 1, &tia)); 212 for (i = 0; i < A->rmap->n + 1; i++) tia[i] = a->i[i] + 1; 213 *ia = tia; 214 if (ja) { 215 PetscInt *tja; 216 PetscCall(PetscMalloc1(nz + 1, &tja)); 217 for (i = 0; i < nz; i++) tja[i] = a->j[i] + 1; 218 *ja = tja; 219 } 220 } else { 221 *ia = a->i; 222 if (ja) *ja = a->j; 223 } 224 PetscFunctionReturn(PETSC_SUCCESS); 225 } 226 227 PetscErrorCode MatRestoreRowIJ_SeqAIJ(Mat A, PetscInt oshift, PetscBool symmetric, PetscBool inodecompressed, PetscInt *n, const PetscInt *ia[], const PetscInt *ja[], PetscBool *done) 228 { 229 PetscFunctionBegin; 230 if (!ia) PetscFunctionReturn(PETSC_SUCCESS); 231 if ((symmetric && A->structurally_symmetric != PETSC_BOOL3_TRUE) || oshift == 1) { 232 PetscCall(PetscFree(*ia)); 233 if (ja) PetscCall(PetscFree(*ja)); 234 } 235 PetscFunctionReturn(PETSC_SUCCESS); 236 } 237 238 PetscErrorCode MatGetColumnIJ_SeqAIJ(Mat A, PetscInt oshift, PetscBool symmetric, PetscBool inodecompressed, PetscInt *nn, const PetscInt *ia[], const PetscInt *ja[], PetscBool *done) 239 { 240 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 241 PetscInt i, *collengths, *cia, *cja, n = A->cmap->n, m = A->rmap->n; 242 PetscInt nz = a->i[m], row, *jj, mr, col; 243 244 PetscFunctionBegin; 245 *nn = n; 246 if (!ia) PetscFunctionReturn(PETSC_SUCCESS); 247 if (symmetric) { 248 PetscCall(MatToSymmetricIJ_SeqAIJ(A->rmap->n, a->i, a->j, PETSC_TRUE, 0, oshift, (PetscInt **)ia, (PetscInt **)ja)); 249 } else { 250 PetscCall(PetscCalloc1(n, &collengths)); 251 PetscCall(PetscMalloc1(n + 1, &cia)); 252 PetscCall(PetscMalloc1(nz, &cja)); 253 jj = a->j; 254 for (i = 0; i < nz; i++) collengths[jj[i]]++; 255 cia[0] = oshift; 256 for (i = 0; i < n; i++) cia[i + 1] = cia[i] + collengths[i]; 257 PetscCall(PetscArrayzero(collengths, n)); 258 jj = a->j; 259 for (row = 0; row < m; row++) { 260 mr = a->i[row + 1] - a->i[row]; 261 for (i = 0; i < mr; i++) { 262 col = *jj++; 263 264 cja[cia[col] + collengths[col]++ - oshift] = row + oshift; 265 } 266 } 267 PetscCall(PetscFree(collengths)); 268 *ia = cia; 269 *ja = cja; 270 } 271 PetscFunctionReturn(PETSC_SUCCESS); 272 } 273 274 PetscErrorCode MatRestoreColumnIJ_SeqAIJ(Mat A, PetscInt oshift, PetscBool symmetric, PetscBool inodecompressed, PetscInt *n, const PetscInt *ia[], const PetscInt *ja[], PetscBool *done) 275 { 276 PetscFunctionBegin; 277 if (!ia) PetscFunctionReturn(PETSC_SUCCESS); 278 279 PetscCall(PetscFree(*ia)); 280 PetscCall(PetscFree(*ja)); 281 PetscFunctionReturn(PETSC_SUCCESS); 282 } 283 284 /* 285 MatGetColumnIJ_SeqAIJ_Color() and MatRestoreColumnIJ_SeqAIJ_Color() are customized from 286 MatGetColumnIJ_SeqAIJ() and MatRestoreColumnIJ_SeqAIJ() by adding an output 287 spidx[], index of a->a, to be used in MatTransposeColoringCreate_SeqAIJ() and MatFDColoringCreate_SeqXAIJ() 288 */ 289 PetscErrorCode MatGetColumnIJ_SeqAIJ_Color(Mat A, PetscInt oshift, PetscBool symmetric, PetscBool inodecompressed, PetscInt *nn, const PetscInt *ia[], const PetscInt *ja[], PetscInt *spidx[], PetscBool *done) 290 { 291 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 292 PetscInt i, *collengths, *cia, *cja, n = A->cmap->n, m = A->rmap->n; 293 PetscInt nz = a->i[m], row, mr, col, tmp; 294 PetscInt *cspidx; 295 const PetscInt *jj; 296 297 PetscFunctionBegin; 298 *nn = n; 299 if (!ia) PetscFunctionReturn(PETSC_SUCCESS); 300 301 PetscCall(PetscCalloc1(n, &collengths)); 302 PetscCall(PetscMalloc1(n + 1, &cia)); 303 PetscCall(PetscMalloc1(nz, &cja)); 304 PetscCall(PetscMalloc1(nz, &cspidx)); 305 jj = a->j; 306 for (i = 0; i < nz; i++) collengths[jj[i]]++; 307 cia[0] = oshift; 308 for (i = 0; i < n; i++) cia[i + 1] = cia[i] + collengths[i]; 309 PetscCall(PetscArrayzero(collengths, n)); 310 jj = a->j; 311 for (row = 0; row < m; row++) { 312 mr = a->i[row + 1] - a->i[row]; 313 for (i = 0; i < mr; i++) { 314 col = *jj++; 315 tmp = cia[col] + collengths[col]++ - oshift; 316 cspidx[tmp] = a->i[row] + i; /* index of a->j */ 317 cja[tmp] = row + oshift; 318 } 319 } 320 PetscCall(PetscFree(collengths)); 321 *ia = cia; 322 *ja = cja; 323 *spidx = cspidx; 324 PetscFunctionReturn(PETSC_SUCCESS); 325 } 326 327 PetscErrorCode MatRestoreColumnIJ_SeqAIJ_Color(Mat A, PetscInt oshift, PetscBool symmetric, PetscBool inodecompressed, PetscInt *n, const PetscInt *ia[], const PetscInt *ja[], PetscInt *spidx[], PetscBool *done) 328 { 329 PetscFunctionBegin; 330 PetscCall(MatRestoreColumnIJ_SeqAIJ(A, oshift, symmetric, inodecompressed, n, ia, ja, done)); 331 PetscCall(PetscFree(*spidx)); 332 PetscFunctionReturn(PETSC_SUCCESS); 333 } 334 335 static PetscErrorCode MatSetValuesRow_SeqAIJ(Mat A, PetscInt row, const PetscScalar v[]) 336 { 337 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 338 PetscInt *ai = a->i; 339 PetscScalar *aa; 340 341 PetscFunctionBegin; 342 PetscCall(MatSeqAIJGetArray(A, &aa)); 343 PetscCall(PetscArraycpy(aa + ai[row], v, ai[row + 1] - ai[row])); 344 PetscCall(MatSeqAIJRestoreArray(A, &aa)); 345 PetscFunctionReturn(PETSC_SUCCESS); 346 } 347 348 /* 349 MatSeqAIJSetValuesLocalFast - An optimized version of MatSetValuesLocal() for SeqAIJ matrices with several assumptions 350 351 - a single row of values is set with each call 352 - no row or column indices are negative or (in error) larger than the number of rows or columns 353 - the values are always added to the matrix, not set 354 - no new locations are introduced in the nonzero structure of the matrix 355 356 This does NOT assume the global column indices are sorted 357 358 */ 359 360 #include <petsc/private/isimpl.h> 361 PetscErrorCode MatSeqAIJSetValuesLocalFast(Mat A, PetscInt m, const PetscInt im[], PetscInt n, const PetscInt in[], const PetscScalar v[], InsertMode is) 362 { 363 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 364 PetscInt low, high, t, row, nrow, i, col, l; 365 const PetscInt *rp, *ai = a->i, *ailen = a->ilen, *aj = a->j; 366 PetscInt lastcol = -1; 367 MatScalar *ap, value, *aa; 368 const PetscInt *ridx = A->rmap->mapping->indices, *cidx = A->cmap->mapping->indices; 369 370 PetscFunctionBegin; 371 PetscCall(MatSeqAIJGetArray(A, &aa)); 372 row = ridx[im[0]]; 373 rp = aj + ai[row]; 374 ap = aa + ai[row]; 375 nrow = ailen[row]; 376 low = 0; 377 high = nrow; 378 for (l = 0; l < n; l++) { /* loop over added columns */ 379 col = cidx[in[l]]; 380 value = v[l]; 381 382 if (col <= lastcol) low = 0; 383 else high = nrow; 384 lastcol = col; 385 while (high - low > 5) { 386 t = (low + high) / 2; 387 if (rp[t] > col) high = t; 388 else low = t; 389 } 390 for (i = low; i < high; i++) { 391 if (rp[i] == col) { 392 ap[i] += value; 393 low = i + 1; 394 break; 395 } 396 } 397 } 398 PetscCall(MatSeqAIJRestoreArray(A, &aa)); 399 return PETSC_SUCCESS; 400 } 401 402 PetscErrorCode MatSetValues_SeqAIJ(Mat A, PetscInt m, const PetscInt im[], PetscInt n, const PetscInt in[], const PetscScalar v[], InsertMode is) 403 { 404 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 405 PetscInt *rp, k, low, high, t, ii, row, nrow, i, col, l, rmax, N; 406 PetscInt *imax = a->imax, *ai = a->i, *ailen = a->ilen; 407 PetscInt *aj = a->j, nonew = a->nonew, lastcol = -1; 408 MatScalar *ap = NULL, value = 0.0, *aa; 409 PetscBool ignorezeroentries = a->ignorezeroentries; 410 PetscBool roworiented = a->roworiented; 411 412 PetscFunctionBegin; 413 PetscCall(MatSeqAIJGetArray(A, &aa)); 414 for (k = 0; k < m; k++) { /* loop over added rows */ 415 row = im[k]; 416 if (row < 0) continue; 417 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); 418 rp = aj + ai[row]; 419 if (!A->structure_only) ap = aa + ai[row]; 420 rmax = imax[row]; 421 nrow = ailen[row]; 422 low = 0; 423 high = nrow; 424 for (l = 0; l < n; l++) { /* loop over added columns */ 425 if (in[l] < 0) continue; 426 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); 427 col = in[l]; 428 if (v && !A->structure_only) value = roworiented ? v[l + k * n] : v[k + l * m]; 429 if (!A->structure_only && value == 0.0 && ignorezeroentries && is == ADD_VALUES && row != col) continue; 430 431 if (col <= lastcol) low = 0; 432 else high = nrow; 433 lastcol = col; 434 while (high - low > 5) { 435 t = (low + high) / 2; 436 if (rp[t] > col) high = t; 437 else low = t; 438 } 439 for (i = low; i < high; i++) { 440 if (rp[i] > col) break; 441 if (rp[i] == col) { 442 if (!A->structure_only) { 443 if (is == ADD_VALUES) { 444 ap[i] += value; 445 (void)PetscLogFlops(1.0); 446 } else ap[i] = value; 447 } 448 low = i + 1; 449 goto noinsert; 450 } 451 } 452 if (value == 0.0 && ignorezeroentries && row != col) goto noinsert; 453 if (nonew == 1) goto noinsert; 454 PetscCheck(nonew != -1, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Inserting a new nonzero at (%" PetscInt_FMT ",%" PetscInt_FMT ") in the matrix", row, col); 455 if (A->structure_only) { 456 MatSeqXAIJReallocateAIJ_structure_only(A, A->rmap->n, 1, nrow, row, col, rmax, ai, aj, rp, imax, nonew, MatScalar); 457 } else { 458 MatSeqXAIJReallocateAIJ(A, A->rmap->n, 1, nrow, row, col, rmax, aa, ai, aj, rp, ap, imax, nonew, MatScalar); 459 } 460 N = nrow++ - 1; 461 a->nz++; 462 high++; 463 /* shift up all the later entries in this row */ 464 PetscCall(PetscArraymove(rp + i + 1, rp + i, N - i + 1)); 465 rp[i] = col; 466 if (!A->structure_only) { 467 PetscCall(PetscArraymove(ap + i + 1, ap + i, N - i + 1)); 468 ap[i] = value; 469 } 470 low = i + 1; 471 A->nonzerostate++; 472 noinsert:; 473 } 474 ailen[row] = nrow; 475 } 476 PetscCall(MatSeqAIJRestoreArray(A, &aa)); 477 PetscFunctionReturn(PETSC_SUCCESS); 478 } 479 480 static PetscErrorCode MatSetValues_SeqAIJ_SortedFullNoPreallocation(Mat A, PetscInt m, const PetscInt im[], PetscInt n, const PetscInt in[], const PetscScalar v[], InsertMode is) 481 { 482 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 483 PetscInt *rp, k, row; 484 PetscInt *ai = a->i; 485 PetscInt *aj = a->j; 486 MatScalar *aa, *ap; 487 488 PetscFunctionBegin; 489 PetscCheck(!A->was_assembled, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Cannot call on assembled matrix."); 490 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); 491 492 PetscCall(MatSeqAIJGetArray(A, &aa)); 493 for (k = 0; k < m; k++) { /* loop over added rows */ 494 row = im[k]; 495 rp = aj + ai[row]; 496 ap = aa + ai[row]; 497 498 PetscCall(PetscMemcpy(rp, in, n * sizeof(PetscInt))); 499 if (!A->structure_only) { 500 if (v) { 501 PetscCall(PetscMemcpy(ap, v, n * sizeof(PetscScalar))); 502 v += n; 503 } else { 504 PetscCall(PetscMemzero(ap, n * sizeof(PetscScalar))); 505 } 506 } 507 a->ilen[row] = n; 508 a->imax[row] = n; 509 a->i[row + 1] = a->i[row] + n; 510 a->nz += n; 511 } 512 PetscCall(MatSeqAIJRestoreArray(A, &aa)); 513 PetscFunctionReturn(PETSC_SUCCESS); 514 } 515 516 /*@ 517 MatSeqAIJSetTotalPreallocation - Sets an upper bound on the total number of expected nonzeros in the matrix. 518 519 Input Parameters: 520 + A - the `MATSEQAIJ` matrix 521 - nztotal - bound on the number of nonzeros 522 523 Level: advanced 524 525 Notes: 526 This can be called if you will be provided the matrix row by row (from row zero) with sorted column indices for each row. 527 Simply call `MatSetValues()` after this call to provide the matrix entries in the usual manner. This matrix may be used 528 as always with multiple matrix assemblies. 529 530 .seealso: [](ch_matrices), `Mat`, `MatSetOption()`, `MAT_SORTED_FULL`, `MatSetValues()`, `MatSeqAIJSetPreallocation()` 531 @*/ 532 PetscErrorCode MatSeqAIJSetTotalPreallocation(Mat A, PetscInt nztotal) 533 { 534 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 535 536 PetscFunctionBegin; 537 PetscCall(PetscLayoutSetUp(A->rmap)); 538 PetscCall(PetscLayoutSetUp(A->cmap)); 539 a->maxnz = nztotal; 540 if (!a->imax) { PetscCall(PetscMalloc1(A->rmap->n, &a->imax)); } 541 if (!a->ilen) { 542 PetscCall(PetscMalloc1(A->rmap->n, &a->ilen)); 543 } else { 544 PetscCall(PetscMemzero(a->ilen, A->rmap->n * sizeof(PetscInt))); 545 } 546 547 /* allocate the matrix space */ 548 if (A->structure_only) { 549 PetscCall(PetscMalloc1(nztotal, &a->j)); 550 PetscCall(PetscMalloc1(A->rmap->n + 1, &a->i)); 551 } else { 552 PetscCall(PetscMalloc3(nztotal, &a->a, nztotal, &a->j, A->rmap->n + 1, &a->i)); 553 } 554 a->i[0] = 0; 555 if (A->structure_only) { 556 a->singlemalloc = PETSC_FALSE; 557 a->free_a = PETSC_FALSE; 558 } else { 559 a->singlemalloc = PETSC_TRUE; 560 a->free_a = PETSC_TRUE; 561 } 562 a->free_ij = PETSC_TRUE; 563 A->ops->setvalues = MatSetValues_SeqAIJ_SortedFullNoPreallocation; 564 A->preallocated = PETSC_TRUE; 565 PetscFunctionReturn(PETSC_SUCCESS); 566 } 567 568 static PetscErrorCode MatSetValues_SeqAIJ_SortedFull(Mat A, PetscInt m, const PetscInt im[], PetscInt n, const PetscInt in[], const PetscScalar v[], InsertMode is) 569 { 570 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 571 PetscInt *rp, k, row; 572 PetscInt *ai = a->i, *ailen = a->ilen; 573 PetscInt *aj = a->j; 574 MatScalar *aa, *ap; 575 576 PetscFunctionBegin; 577 PetscCall(MatSeqAIJGetArray(A, &aa)); 578 for (k = 0; k < m; k++) { /* loop over added rows */ 579 row = im[k]; 580 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); 581 rp = aj + ai[row]; 582 ap = aa + ai[row]; 583 if (!A->was_assembled) PetscCall(PetscMemcpy(rp, in, n * sizeof(PetscInt))); 584 if (!A->structure_only) { 585 if (v) { 586 PetscCall(PetscMemcpy(ap, v, n * sizeof(PetscScalar))); 587 v += n; 588 } else { 589 PetscCall(PetscMemzero(ap, n * sizeof(PetscScalar))); 590 } 591 } 592 ailen[row] = n; 593 a->nz += n; 594 } 595 PetscCall(MatSeqAIJRestoreArray(A, &aa)); 596 PetscFunctionReturn(PETSC_SUCCESS); 597 } 598 599 static PetscErrorCode MatGetValues_SeqAIJ(Mat A, PetscInt m, const PetscInt im[], PetscInt n, const PetscInt in[], PetscScalar v[]) 600 { 601 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 602 PetscInt *rp, k, low, high, t, row, nrow, i, col, l, *aj = a->j; 603 PetscInt *ai = a->i, *ailen = a->ilen; 604 const MatScalar *ap, *aa; 605 606 PetscFunctionBegin; 607 PetscCall(MatSeqAIJGetArrayRead(A, &aa)); 608 for (k = 0; k < m; k++) { /* loop over rows */ 609 row = im[k]; 610 if (row < 0) { 611 v += n; 612 continue; 613 } /* negative row */ 614 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); 615 rp = aj + ai[row]; 616 ap = aa + ai[row]; 617 nrow = ailen[row]; 618 for (l = 0; l < n; l++) { /* loop over columns */ 619 if (in[l] < 0) { 620 v++; 621 continue; 622 } /* negative column */ 623 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); 624 col = in[l]; 625 high = nrow; 626 low = 0; /* assume unsorted */ 627 while (high - low > 5) { 628 t = (low + high) / 2; 629 if (rp[t] > col) high = t; 630 else low = t; 631 } 632 for (i = low; i < high; i++) { 633 if (rp[i] > col) break; 634 if (rp[i] == col) { 635 *v++ = ap[i]; 636 goto finished; 637 } 638 } 639 *v++ = 0.0; 640 finished:; 641 } 642 } 643 PetscCall(MatSeqAIJRestoreArrayRead(A, &aa)); 644 PetscFunctionReturn(PETSC_SUCCESS); 645 } 646 647 static PetscErrorCode MatView_SeqAIJ_Binary(Mat mat, PetscViewer viewer) 648 { 649 Mat_SeqAIJ *A = (Mat_SeqAIJ *)mat->data; 650 const PetscScalar *av; 651 PetscInt header[4], M, N, m, nz, i; 652 PetscInt *rowlens; 653 654 PetscFunctionBegin; 655 PetscCall(PetscViewerSetUp(viewer)); 656 657 M = mat->rmap->N; 658 N = mat->cmap->N; 659 m = mat->rmap->n; 660 nz = A->nz; 661 662 /* write matrix header */ 663 header[0] = MAT_FILE_CLASSID; 664 header[1] = M; 665 header[2] = N; 666 header[3] = nz; 667 PetscCall(PetscViewerBinaryWrite(viewer, header, 4, PETSC_INT)); 668 669 /* fill in and store row lengths */ 670 PetscCall(PetscMalloc1(m, &rowlens)); 671 for (i = 0; i < m; i++) rowlens[i] = A->i[i + 1] - A->i[i]; 672 PetscCall(PetscViewerBinaryWrite(viewer, rowlens, m, PETSC_INT)); 673 PetscCall(PetscFree(rowlens)); 674 /* store column indices */ 675 PetscCall(PetscViewerBinaryWrite(viewer, A->j, nz, PETSC_INT)); 676 /* store nonzero values */ 677 PetscCall(MatSeqAIJGetArrayRead(mat, &av)); 678 PetscCall(PetscViewerBinaryWrite(viewer, av, nz, PETSC_SCALAR)); 679 PetscCall(MatSeqAIJRestoreArrayRead(mat, &av)); 680 681 /* write block size option to the viewer's .info file */ 682 PetscCall(MatView_Binary_BlockSizes(mat, viewer)); 683 PetscFunctionReturn(PETSC_SUCCESS); 684 } 685 686 static PetscErrorCode MatView_SeqAIJ_ASCII_structonly(Mat A, PetscViewer viewer) 687 { 688 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 689 PetscInt i, k, m = A->rmap->N; 690 691 PetscFunctionBegin; 692 PetscCall(PetscViewerASCIIUseTabs(viewer, PETSC_FALSE)); 693 for (i = 0; i < m; i++) { 694 PetscCall(PetscViewerASCIIPrintf(viewer, "row %" PetscInt_FMT ":", i)); 695 for (k = a->i[i]; k < a->i[i + 1]; k++) PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ") ", a->j[k])); 696 PetscCall(PetscViewerASCIIPrintf(viewer, "\n")); 697 } 698 PetscCall(PetscViewerASCIIUseTabs(viewer, PETSC_TRUE)); 699 PetscFunctionReturn(PETSC_SUCCESS); 700 } 701 702 extern PetscErrorCode MatSeqAIJFactorInfo_Matlab(Mat, PetscViewer); 703 704 static PetscErrorCode MatView_SeqAIJ_ASCII(Mat A, PetscViewer viewer) 705 { 706 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 707 const PetscScalar *av; 708 PetscInt i, j, m = A->rmap->n; 709 const char *name; 710 PetscViewerFormat format; 711 712 PetscFunctionBegin; 713 if (A->structure_only) { 714 PetscCall(MatView_SeqAIJ_ASCII_structonly(A, viewer)); 715 PetscFunctionReturn(PETSC_SUCCESS); 716 } 717 718 PetscCall(PetscViewerGetFormat(viewer, &format)); 719 if (format == PETSC_VIEWER_ASCII_FACTOR_INFO || format == PETSC_VIEWER_ASCII_INFO || format == PETSC_VIEWER_ASCII_INFO_DETAIL) PetscFunctionReturn(PETSC_SUCCESS); 720 721 /* trigger copy to CPU if needed */ 722 PetscCall(MatSeqAIJGetArrayRead(A, &av)); 723 PetscCall(MatSeqAIJRestoreArrayRead(A, &av)); 724 if (format == PETSC_VIEWER_ASCII_MATLAB) { 725 PetscInt nofinalvalue = 0; 726 if (m && ((a->i[m] == a->i[m - 1]) || (a->j[a->nz - 1] != A->cmap->n - 1))) { 727 /* Need a dummy value to ensure the dimension of the matrix. */ 728 nofinalvalue = 1; 729 } 730 PetscCall(PetscViewerASCIIUseTabs(viewer, PETSC_FALSE)); 731 PetscCall(PetscViewerASCIIPrintf(viewer, "%% Size = %" PetscInt_FMT " %" PetscInt_FMT " \n", m, A->cmap->n)); 732 PetscCall(PetscViewerASCIIPrintf(viewer, "%% Nonzeros = %" PetscInt_FMT " \n", a->nz)); 733 #if defined(PETSC_USE_COMPLEX) 734 PetscCall(PetscViewerASCIIPrintf(viewer, "zzz = zeros(%" PetscInt_FMT ",4);\n", a->nz + nofinalvalue)); 735 #else 736 PetscCall(PetscViewerASCIIPrintf(viewer, "zzz = zeros(%" PetscInt_FMT ",3);\n", a->nz + nofinalvalue)); 737 #endif 738 PetscCall(PetscViewerASCIIPrintf(viewer, "zzz = [\n")); 739 740 for (i = 0; i < m; i++) { 741 for (j = a->i[i]; j < a->i[i + 1]; j++) { 742 #if defined(PETSC_USE_COMPLEX) 743 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]))); 744 #else 745 PetscCall(PetscViewerASCIIPrintf(viewer, "%" PetscInt_FMT " %" PetscInt_FMT " %18.16e\n", i + 1, a->j[j] + 1, (double)a->a[j])); 746 #endif 747 } 748 } 749 if (nofinalvalue) { 750 #if defined(PETSC_USE_COMPLEX) 751 PetscCall(PetscViewerASCIIPrintf(viewer, "%" PetscInt_FMT " %" PetscInt_FMT " %18.16e %18.16e\n", m, A->cmap->n, 0., 0.)); 752 #else 753 PetscCall(PetscViewerASCIIPrintf(viewer, "%" PetscInt_FMT " %" PetscInt_FMT " %18.16e\n", m, A->cmap->n, 0.0)); 754 #endif 755 } 756 PetscCall(PetscObjectGetName((PetscObject)A, &name)); 757 PetscCall(PetscViewerASCIIPrintf(viewer, "];\n %s = spconvert(zzz);\n", name)); 758 PetscCall(PetscViewerASCIIUseTabs(viewer, PETSC_TRUE)); 759 } else if (format == PETSC_VIEWER_ASCII_COMMON) { 760 PetscCall(PetscViewerASCIIUseTabs(viewer, PETSC_FALSE)); 761 for (i = 0; i < m; i++) { 762 PetscCall(PetscViewerASCIIPrintf(viewer, "row %" PetscInt_FMT ":", i)); 763 for (j = a->i[i]; j < a->i[i + 1]; j++) { 764 #if defined(PETSC_USE_COMPLEX) 765 if (PetscImaginaryPart(a->a[j]) > 0.0 && PetscRealPart(a->a[j]) != 0.0) { 766 PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g + %g i)", a->j[j], (double)PetscRealPart(a->a[j]), (double)PetscImaginaryPart(a->a[j]))); 767 } else if (PetscImaginaryPart(a->a[j]) < 0.0 && PetscRealPart(a->a[j]) != 0.0) { 768 PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g - %g i)", a->j[j], (double)PetscRealPart(a->a[j]), (double)-PetscImaginaryPart(a->a[j]))); 769 } else if (PetscRealPart(a->a[j]) != 0.0) { 770 PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g) ", a->j[j], (double)PetscRealPart(a->a[j]))); 771 } 772 #else 773 if (a->a[j] != 0.0) PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g) ", a->j[j], (double)a->a[j])); 774 #endif 775 } 776 PetscCall(PetscViewerASCIIPrintf(viewer, "\n")); 777 } 778 PetscCall(PetscViewerASCIIUseTabs(viewer, PETSC_TRUE)); 779 } else if (format == PETSC_VIEWER_ASCII_SYMMODU) { 780 PetscInt nzd = 0, fshift = 1, *sptr; 781 PetscCall(PetscViewerASCIIUseTabs(viewer, PETSC_FALSE)); 782 PetscCall(PetscMalloc1(m + 1, &sptr)); 783 for (i = 0; i < m; i++) { 784 sptr[i] = nzd + 1; 785 for (j = a->i[i]; j < a->i[i + 1]; j++) { 786 if (a->j[j] >= i) { 787 #if defined(PETSC_USE_COMPLEX) 788 if (PetscImaginaryPart(a->a[j]) != 0.0 || PetscRealPart(a->a[j]) != 0.0) nzd++; 789 #else 790 if (a->a[j] != 0.0) nzd++; 791 #endif 792 } 793 } 794 } 795 sptr[m] = nzd + 1; 796 PetscCall(PetscViewerASCIIPrintf(viewer, " %" PetscInt_FMT " %" PetscInt_FMT "\n\n", m, nzd)); 797 for (i = 0; i < m + 1; i += 6) { 798 if (i + 4 < m) { 799 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])); 800 } else if (i + 3 < m) { 801 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])); 802 } else if (i + 2 < m) { 803 PetscCall(PetscViewerASCIIPrintf(viewer, " %" PetscInt_FMT " %" PetscInt_FMT " %" PetscInt_FMT " %" PetscInt_FMT "\n", sptr[i], sptr[i + 1], sptr[i + 2], sptr[i + 3])); 804 } else if (i + 1 < m) { 805 PetscCall(PetscViewerASCIIPrintf(viewer, " %" PetscInt_FMT " %" PetscInt_FMT " %" PetscInt_FMT "\n", sptr[i], sptr[i + 1], sptr[i + 2])); 806 } else if (i < m) { 807 PetscCall(PetscViewerASCIIPrintf(viewer, " %" PetscInt_FMT " %" PetscInt_FMT "\n", sptr[i], sptr[i + 1])); 808 } else { 809 PetscCall(PetscViewerASCIIPrintf(viewer, " %" PetscInt_FMT "\n", sptr[i])); 810 } 811 } 812 PetscCall(PetscViewerASCIIPrintf(viewer, "\n")); 813 PetscCall(PetscFree(sptr)); 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) PetscCall(PetscViewerASCIIPrintf(viewer, " %" PetscInt_FMT " ", a->j[j] + fshift)); 817 } 818 PetscCall(PetscViewerASCIIPrintf(viewer, "\n")); 819 } 820 PetscCall(PetscViewerASCIIPrintf(viewer, "\n")); 821 for (i = 0; i < m; i++) { 822 for (j = a->i[i]; j < a->i[i + 1]; j++) { 823 if (a->j[j] >= i) { 824 #if defined(PETSC_USE_COMPLEX) 825 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]))); 826 #else 827 if (a->a[j] != 0.0) PetscCall(PetscViewerASCIIPrintf(viewer, " %18.16e ", (double)a->a[j])); 828 #endif 829 } 830 } 831 PetscCall(PetscViewerASCIIPrintf(viewer, "\n")); 832 } 833 PetscCall(PetscViewerASCIIUseTabs(viewer, PETSC_TRUE)); 834 } else if (format == PETSC_VIEWER_ASCII_DENSE) { 835 PetscInt cnt = 0, jcnt; 836 PetscScalar value; 837 #if defined(PETSC_USE_COMPLEX) 838 PetscBool realonly = PETSC_TRUE; 839 840 for (i = 0; i < a->i[m]; i++) { 841 if (PetscImaginaryPart(a->a[i]) != 0.0) { 842 realonly = PETSC_FALSE; 843 break; 844 } 845 } 846 #endif 847 848 PetscCall(PetscViewerASCIIUseTabs(viewer, PETSC_FALSE)); 849 for (i = 0; i < m; i++) { 850 jcnt = 0; 851 for (j = 0; j < A->cmap->n; j++) { 852 if (jcnt < a->i[i + 1] - a->i[i] && j == a->j[cnt]) { 853 value = a->a[cnt++]; 854 jcnt++; 855 } else { 856 value = 0.0; 857 } 858 #if defined(PETSC_USE_COMPLEX) 859 if (realonly) { 860 PetscCall(PetscViewerASCIIPrintf(viewer, " %7.5e ", (double)PetscRealPart(value))); 861 } else { 862 PetscCall(PetscViewerASCIIPrintf(viewer, " %7.5e+%7.5e i ", (double)PetscRealPart(value), (double)PetscImaginaryPart(value))); 863 } 864 #else 865 PetscCall(PetscViewerASCIIPrintf(viewer, " %7.5e ", (double)value)); 866 #endif 867 } 868 PetscCall(PetscViewerASCIIPrintf(viewer, "\n")); 869 } 870 PetscCall(PetscViewerASCIIUseTabs(viewer, PETSC_TRUE)); 871 } else if (format == PETSC_VIEWER_ASCII_MATRIXMARKET) { 872 PetscInt fshift = 1; 873 PetscCall(PetscViewerASCIIUseTabs(viewer, PETSC_FALSE)); 874 #if defined(PETSC_USE_COMPLEX) 875 PetscCall(PetscViewerASCIIPrintf(viewer, "%%%%MatrixMarket matrix coordinate complex general\n")); 876 #else 877 PetscCall(PetscViewerASCIIPrintf(viewer, "%%%%MatrixMarket matrix coordinate real general\n")); 878 #endif 879 PetscCall(PetscViewerASCIIPrintf(viewer, "%" PetscInt_FMT " %" PetscInt_FMT " %" PetscInt_FMT "\n", m, A->cmap->n, a->nz)); 880 for (i = 0; i < m; i++) { 881 for (j = a->i[i]; j < a->i[i + 1]; j++) { 882 #if defined(PETSC_USE_COMPLEX) 883 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]))); 884 #else 885 PetscCall(PetscViewerASCIIPrintf(viewer, "%" PetscInt_FMT " %" PetscInt_FMT " %g\n", i + fshift, a->j[j] + fshift, (double)a->a[j])); 886 #endif 887 } 888 } 889 PetscCall(PetscViewerASCIIUseTabs(viewer, PETSC_TRUE)); 890 } else { 891 PetscCall(PetscViewerASCIIUseTabs(viewer, PETSC_FALSE)); 892 if (A->factortype) { 893 for (i = 0; i < m; i++) { 894 PetscCall(PetscViewerASCIIPrintf(viewer, "row %" PetscInt_FMT ":", i)); 895 /* L part */ 896 for (j = a->i[i]; j < a->i[i + 1]; j++) { 897 #if defined(PETSC_USE_COMPLEX) 898 if (PetscImaginaryPart(a->a[j]) > 0.0) { 899 PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g + %g i)", a->j[j], (double)PetscRealPart(a->a[j]), (double)PetscImaginaryPart(a->a[j]))); 900 } else if (PetscImaginaryPart(a->a[j]) < 0.0) { 901 PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g - %g i)", a->j[j], (double)PetscRealPart(a->a[j]), (double)(-PetscImaginaryPart(a->a[j])))); 902 } else { 903 PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g) ", a->j[j], (double)PetscRealPart(a->a[j]))); 904 } 905 #else 906 PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g) ", a->j[j], (double)a->a[j])); 907 #endif 908 } 909 /* diagonal */ 910 j = a->diag[i]; 911 #if defined(PETSC_USE_COMPLEX) 912 if (PetscImaginaryPart(a->a[j]) > 0.0) { 913 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]))); 914 } else if (PetscImaginaryPart(a->a[j]) < 0.0) { 915 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])))); 916 } else { 917 PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g) ", a->j[j], (double)PetscRealPart(1.0 / a->a[j]))); 918 } 919 #else 920 PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g) ", a->j[j], (double)(1.0 / a->a[j]))); 921 #endif 922 923 /* U part */ 924 for (j = a->diag[i + 1] + 1; j < a->diag[i]; j++) { 925 #if defined(PETSC_USE_COMPLEX) 926 if (PetscImaginaryPart(a->a[j]) > 0.0) { 927 PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g + %g i)", a->j[j], (double)PetscRealPart(a->a[j]), (double)PetscImaginaryPart(a->a[j]))); 928 } else if (PetscImaginaryPart(a->a[j]) < 0.0) { 929 PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g - %g i)", a->j[j], (double)PetscRealPart(a->a[j]), (double)(-PetscImaginaryPart(a->a[j])))); 930 } else { 931 PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g) ", a->j[j], (double)PetscRealPart(a->a[j]))); 932 } 933 #else 934 PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g) ", a->j[j], (double)a->a[j])); 935 #endif 936 } 937 PetscCall(PetscViewerASCIIPrintf(viewer, "\n")); 938 } 939 } else { 940 for (i = 0; i < m; i++) { 941 PetscCall(PetscViewerASCIIPrintf(viewer, "row %" PetscInt_FMT ":", i)); 942 for (j = a->i[i]; j < a->i[i + 1]; j++) { 943 #if defined(PETSC_USE_COMPLEX) 944 if (PetscImaginaryPart(a->a[j]) > 0.0) { 945 PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g + %g i)", a->j[j], (double)PetscRealPart(a->a[j]), (double)PetscImaginaryPart(a->a[j]))); 946 } else if (PetscImaginaryPart(a->a[j]) < 0.0) { 947 PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g - %g i)", a->j[j], (double)PetscRealPart(a->a[j]), (double)-PetscImaginaryPart(a->a[j]))); 948 } else { 949 PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g) ", a->j[j], (double)PetscRealPart(a->a[j]))); 950 } 951 #else 952 PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g) ", a->j[j], (double)a->a[j])); 953 #endif 954 } 955 PetscCall(PetscViewerASCIIPrintf(viewer, "\n")); 956 } 957 } 958 PetscCall(PetscViewerASCIIUseTabs(viewer, PETSC_TRUE)); 959 } 960 PetscCall(PetscViewerFlush(viewer)); 961 PetscFunctionReturn(PETSC_SUCCESS); 962 } 963 964 #include <petscdraw.h> 965 static PetscErrorCode MatView_SeqAIJ_Draw_Zoom(PetscDraw draw, void *Aa) 966 { 967 Mat A = (Mat)Aa; 968 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 969 PetscInt i, j, m = A->rmap->n; 970 int color; 971 PetscReal xl, yl, xr, yr, x_l, x_r, y_l, y_r; 972 PetscViewer viewer; 973 PetscViewerFormat format; 974 const PetscScalar *aa; 975 976 PetscFunctionBegin; 977 PetscCall(PetscObjectQuery((PetscObject)A, "Zoomviewer", (PetscObject *)&viewer)); 978 PetscCall(PetscViewerGetFormat(viewer, &format)); 979 PetscCall(PetscDrawGetCoordinates(draw, &xl, &yl, &xr, &yr)); 980 981 /* loop over matrix elements drawing boxes */ 982 PetscCall(MatSeqAIJGetArrayRead(A, &aa)); 983 if (format != PETSC_VIEWER_DRAW_CONTOUR) { 984 PetscDrawCollectiveBegin(draw); 985 /* Blue for negative, Cyan for zero and Red for positive */ 986 color = PETSC_DRAW_BLUE; 987 for (i = 0; i < m; i++) { 988 y_l = m - i - 1.0; 989 y_r = y_l + 1.0; 990 for (j = a->i[i]; j < a->i[i + 1]; j++) { 991 x_l = a->j[j]; 992 x_r = x_l + 1.0; 993 if (PetscRealPart(aa[j]) >= 0.) continue; 994 PetscCall(PetscDrawRectangle(draw, x_l, y_l, x_r, y_r, color, color, color, color)); 995 } 996 } 997 color = PETSC_DRAW_CYAN; 998 for (i = 0; i < m; i++) { 999 y_l = m - i - 1.0; 1000 y_r = y_l + 1.0; 1001 for (j = a->i[i]; j < a->i[i + 1]; j++) { 1002 x_l = a->j[j]; 1003 x_r = x_l + 1.0; 1004 if (aa[j] != 0.) continue; 1005 PetscCall(PetscDrawRectangle(draw, x_l, y_l, x_r, y_r, color, color, color, color)); 1006 } 1007 } 1008 color = PETSC_DRAW_RED; 1009 for (i = 0; i < m; i++) { 1010 y_l = m - i - 1.0; 1011 y_r = y_l + 1.0; 1012 for (j = a->i[i]; j < a->i[i + 1]; j++) { 1013 x_l = a->j[j]; 1014 x_r = x_l + 1.0; 1015 if (PetscRealPart(aa[j]) <= 0.) continue; 1016 PetscCall(PetscDrawRectangle(draw, x_l, y_l, x_r, y_r, color, color, color, color)); 1017 } 1018 } 1019 PetscDrawCollectiveEnd(draw); 1020 } else { 1021 /* use contour shading to indicate magnitude of values */ 1022 /* first determine max of all nonzero values */ 1023 PetscReal minv = 0.0, maxv = 0.0; 1024 PetscInt nz = a->nz, count = 0; 1025 PetscDraw popup; 1026 1027 for (i = 0; i < nz; i++) { 1028 if (PetscAbsScalar(aa[i]) > maxv) maxv = PetscAbsScalar(aa[i]); 1029 } 1030 if (minv >= maxv) maxv = minv + PETSC_SMALL; 1031 PetscCall(PetscDrawGetPopup(draw, &popup)); 1032 PetscCall(PetscDrawScalePopup(popup, minv, maxv)); 1033 1034 PetscDrawCollectiveBegin(draw); 1035 for (i = 0; i < m; i++) { 1036 y_l = m - i - 1.0; 1037 y_r = y_l + 1.0; 1038 for (j = a->i[i]; j < a->i[i + 1]; j++) { 1039 x_l = a->j[j]; 1040 x_r = x_l + 1.0; 1041 color = PetscDrawRealToColor(PetscAbsScalar(aa[count]), minv, maxv); 1042 PetscCall(PetscDrawRectangle(draw, x_l, y_l, x_r, y_r, color, color, color, color)); 1043 count++; 1044 } 1045 } 1046 PetscDrawCollectiveEnd(draw); 1047 } 1048 PetscCall(MatSeqAIJRestoreArrayRead(A, &aa)); 1049 PetscFunctionReturn(PETSC_SUCCESS); 1050 } 1051 1052 #include <petscdraw.h> 1053 static PetscErrorCode MatView_SeqAIJ_Draw(Mat A, PetscViewer viewer) 1054 { 1055 PetscDraw draw; 1056 PetscReal xr, yr, xl, yl, h, w; 1057 PetscBool isnull; 1058 1059 PetscFunctionBegin; 1060 PetscCall(PetscViewerDrawGetDraw(viewer, 0, &draw)); 1061 PetscCall(PetscDrawIsNull(draw, &isnull)); 1062 if (isnull) PetscFunctionReturn(PETSC_SUCCESS); 1063 1064 xr = A->cmap->n; 1065 yr = A->rmap->n; 1066 h = yr / 10.0; 1067 w = xr / 10.0; 1068 xr += w; 1069 yr += h; 1070 xl = -w; 1071 yl = -h; 1072 PetscCall(PetscDrawSetCoordinates(draw, xl, yl, xr, yr)); 1073 PetscCall(PetscObjectCompose((PetscObject)A, "Zoomviewer", (PetscObject)viewer)); 1074 PetscCall(PetscDrawZoom(draw, MatView_SeqAIJ_Draw_Zoom, A)); 1075 PetscCall(PetscObjectCompose((PetscObject)A, "Zoomviewer", NULL)); 1076 PetscCall(PetscDrawSave(draw)); 1077 PetscFunctionReturn(PETSC_SUCCESS); 1078 } 1079 1080 PetscErrorCode MatView_SeqAIJ(Mat A, PetscViewer viewer) 1081 { 1082 PetscBool iascii, isbinary, isdraw; 1083 1084 PetscFunctionBegin; 1085 PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERASCII, &iascii)); 1086 PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERBINARY, &isbinary)); 1087 PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERDRAW, &isdraw)); 1088 if (iascii) PetscCall(MatView_SeqAIJ_ASCII(A, viewer)); 1089 else if (isbinary) PetscCall(MatView_SeqAIJ_Binary(A, viewer)); 1090 else if (isdraw) PetscCall(MatView_SeqAIJ_Draw(A, viewer)); 1091 PetscCall(MatView_SeqAIJ_Inode(A, viewer)); 1092 PetscFunctionReturn(PETSC_SUCCESS); 1093 } 1094 1095 PetscErrorCode MatAssemblyEnd_SeqAIJ(Mat A, MatAssemblyType mode) 1096 { 1097 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 1098 PetscInt fshift = 0, i, *ai = a->i, *aj = a->j, *imax = a->imax; 1099 PetscInt m = A->rmap->n, *ip, N, *ailen = a->ilen, rmax = 0, n; 1100 MatScalar *aa = a->a, *ap; 1101 PetscReal ratio = 0.6; 1102 1103 PetscFunctionBegin; 1104 if (mode == MAT_FLUSH_ASSEMBLY) PetscFunctionReturn(PETSC_SUCCESS); 1105 PetscCall(MatSeqAIJInvalidateDiagonal(A)); 1106 if (A->was_assembled && A->ass_nonzerostate == A->nonzerostate) { 1107 /* we need to respect users asking to use or not the inodes routine in between matrix assemblies */ 1108 PetscCall(MatAssemblyEnd_SeqAIJ_Inode(A, mode)); 1109 PetscFunctionReturn(PETSC_SUCCESS); 1110 } 1111 1112 if (m) rmax = ailen[0]; /* determine row with most nonzeros */ 1113 for (i = 1; i < m; i++) { 1114 /* move each row back by the amount of empty slots (fshift) before it*/ 1115 fshift += imax[i - 1] - ailen[i - 1]; 1116 rmax = PetscMax(rmax, ailen[i]); 1117 if (fshift) { 1118 ip = aj + ai[i]; 1119 ap = aa + ai[i]; 1120 N = ailen[i]; 1121 PetscCall(PetscArraymove(ip - fshift, ip, N)); 1122 if (!A->structure_only) PetscCall(PetscArraymove(ap - fshift, ap, N)); 1123 } 1124 ai[i] = ai[i - 1] + ailen[i - 1]; 1125 } 1126 if (m) { 1127 fshift += imax[m - 1] - ailen[m - 1]; 1128 ai[m] = ai[m - 1] + ailen[m - 1]; 1129 } 1130 /* reset ilen and imax for each row */ 1131 a->nonzerorowcnt = 0; 1132 if (A->structure_only) { 1133 PetscCall(PetscFree(a->imax)); 1134 PetscCall(PetscFree(a->ilen)); 1135 } else { /* !A->structure_only */ 1136 for (i = 0; i < m; i++) { 1137 ailen[i] = imax[i] = ai[i + 1] - ai[i]; 1138 a->nonzerorowcnt += ((ai[i + 1] - ai[i]) > 0); 1139 } 1140 } 1141 a->nz = ai[m]; 1142 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); 1143 PetscCall(MatMarkDiagonal_SeqAIJ(A)); // since diagonal info is used a lot, it is helpful to set them up at the end of assembly 1144 a->diagonaldense = PETSC_TRUE; 1145 n = PetscMin(A->rmap->n, A->cmap->n); 1146 for (i = 0; i < n; i++) { 1147 if (a->diag[i] >= ai[i + 1]) { 1148 a->diagonaldense = PETSC_FALSE; 1149 break; 1150 } 1151 } 1152 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)); 1153 PetscCall(PetscInfo(A, "Number of mallocs during MatSetValues() is %" PetscInt_FMT "\n", a->reallocs)); 1154 PetscCall(PetscInfo(A, "Maximum nonzeros in any row is %" PetscInt_FMT "\n", rmax)); 1155 1156 A->info.mallocs += a->reallocs; 1157 a->reallocs = 0; 1158 A->info.nz_unneeded = (PetscReal)fshift; 1159 a->rmax = rmax; 1160 1161 if (!A->structure_only) PetscCall(MatCheckCompressedRow(A, a->nonzerorowcnt, &a->compressedrow, a->i, m, ratio)); 1162 PetscCall(MatAssemblyEnd_SeqAIJ_Inode(A, mode)); 1163 PetscFunctionReturn(PETSC_SUCCESS); 1164 } 1165 1166 static PetscErrorCode MatRealPart_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] = PetscRealPart(aa[i]); 1175 PetscCall(MatSeqAIJRestoreArray(A, &aa)); 1176 PetscCall(MatSeqAIJInvalidateDiagonal(A)); 1177 PetscFunctionReturn(PETSC_SUCCESS); 1178 } 1179 1180 static PetscErrorCode MatImaginaryPart_SeqAIJ(Mat A) 1181 { 1182 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 1183 PetscInt i, nz = a->nz; 1184 MatScalar *aa; 1185 1186 PetscFunctionBegin; 1187 PetscCall(MatSeqAIJGetArray(A, &aa)); 1188 for (i = 0; i < nz; i++) aa[i] = PetscImaginaryPart(aa[i]); 1189 PetscCall(MatSeqAIJRestoreArray(A, &aa)); 1190 PetscCall(MatSeqAIJInvalidateDiagonal(A)); 1191 PetscFunctionReturn(PETSC_SUCCESS); 1192 } 1193 1194 PetscErrorCode MatZeroEntries_SeqAIJ(Mat A) 1195 { 1196 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 1197 MatScalar *aa; 1198 1199 PetscFunctionBegin; 1200 PetscCall(MatSeqAIJGetArrayWrite(A, &aa)); 1201 PetscCall(PetscArrayzero(aa, a->i[A->rmap->n])); 1202 PetscCall(MatSeqAIJRestoreArrayWrite(A, &aa)); 1203 PetscCall(MatSeqAIJInvalidateDiagonal(A)); 1204 PetscFunctionReturn(PETSC_SUCCESS); 1205 } 1206 1207 PetscErrorCode MatDestroy_SeqAIJ(Mat A) 1208 { 1209 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 1210 1211 PetscFunctionBegin; 1212 if (A->hash_active) { 1213 A->ops[0] = a->cops; 1214 PetscCall(PetscHMapIJVDestroy(&a->ht)); 1215 PetscCall(PetscFree(a->dnz)); 1216 A->hash_active = PETSC_FALSE; 1217 } 1218 1219 PetscCall(PetscLogObjectState((PetscObject)A, "Rows=%" PetscInt_FMT ", Cols=%" PetscInt_FMT ", NZ=%" PetscInt_FMT, A->rmap->n, A->cmap->n, a->nz)); 1220 PetscCall(MatSeqXAIJFreeAIJ(A, &a->a, &a->j, &a->i)); 1221 PetscCall(ISDestroy(&a->row)); 1222 PetscCall(ISDestroy(&a->col)); 1223 PetscCall(PetscFree(a->diag)); 1224 PetscCall(PetscFree(a->ibdiag)); 1225 PetscCall(PetscFree(a->imax)); 1226 PetscCall(PetscFree(a->ilen)); 1227 PetscCall(PetscFree(a->ipre)); 1228 PetscCall(PetscFree3(a->idiag, a->mdiag, a->ssor_work)); 1229 PetscCall(PetscFree(a->solve_work)); 1230 PetscCall(ISDestroy(&a->icol)); 1231 PetscCall(PetscFree(a->saved_values)); 1232 PetscCall(PetscFree2(a->compressedrow.i, a->compressedrow.rindex)); 1233 PetscCall(MatDestroy_SeqAIJ_Inode(A)); 1234 PetscCall(PetscFree(A->data)); 1235 1236 /* MatMatMultNumeric_SeqAIJ_SeqAIJ_Sorted may allocate this. 1237 That function is so heavily used (sometimes in an hidden way through multnumeric function pointers) 1238 that is hard to properly add this data to the MatProduct data. We free it here to avoid 1239 users reusing the matrix object with different data to incur in obscure segmentation faults 1240 due to different matrix sizes */ 1241 PetscCall(PetscObjectCompose((PetscObject)A, "__PETSc__ab_dense", NULL)); 1242 1243 PetscCall(PetscObjectChangeTypeName((PetscObject)A, NULL)); 1244 PetscCall(PetscObjectComposeFunction((PetscObject)A, "PetscMatlabEnginePut_C", NULL)); 1245 PetscCall(PetscObjectComposeFunction((PetscObject)A, "PetscMatlabEngineGet_C", NULL)); 1246 PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatSeqAIJSetColumnIndices_C", NULL)); 1247 PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatStoreValues_C", NULL)); 1248 PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatRetrieveValues_C", NULL)); 1249 PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_seqsbaij_C", NULL)); 1250 PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_seqbaij_C", NULL)); 1251 PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_seqaijperm_C", NULL)); 1252 PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_seqaijsell_C", NULL)); 1253 #if defined(PETSC_HAVE_MKL_SPARSE) 1254 PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_seqaijmkl_C", NULL)); 1255 #endif 1256 #if defined(PETSC_HAVE_CUDA) 1257 PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_seqaijcusparse_C", NULL)); 1258 PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_seqaijcusparse_seqaij_C", NULL)); 1259 PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_seqaij_seqaijcusparse_C", NULL)); 1260 #endif 1261 #if defined(PETSC_HAVE_HIP) 1262 PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_seqaijhipsparse_C", NULL)); 1263 PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_seqaijhipsparse_seqaij_C", NULL)); 1264 PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_seqaij_seqaijhipsparse_C", NULL)); 1265 #endif 1266 #if defined(PETSC_HAVE_KOKKOS_KERNELS) 1267 PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_seqaijkokkos_C", NULL)); 1268 #endif 1269 PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_seqaijcrl_C", NULL)); 1270 #if defined(PETSC_HAVE_ELEMENTAL) 1271 PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_elemental_C", NULL)); 1272 #endif 1273 #if defined(PETSC_HAVE_SCALAPACK) 1274 PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_scalapack_C", NULL)); 1275 #endif 1276 #if defined(PETSC_HAVE_HYPRE) 1277 PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_hypre_C", NULL)); 1278 PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_transpose_seqaij_seqaij_C", NULL)); 1279 #endif 1280 PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_seqdense_C", NULL)); 1281 PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_seqsell_C", NULL)); 1282 PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_is_C", NULL)); 1283 PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatIsTranspose_C", NULL)); 1284 PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatIsHermitianTranspose_C", NULL)); 1285 PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatSeqAIJSetPreallocation_C", NULL)); 1286 PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatResetPreallocation_C", NULL)); 1287 PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatSeqAIJSetPreallocationCSR_C", NULL)); 1288 PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatReorderForNonzeroDiagonal_C", NULL)); 1289 PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_is_seqaij_C", NULL)); 1290 PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_seqdense_seqaij_C", NULL)); 1291 PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_seqaij_seqaij_C", NULL)); 1292 PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatSeqAIJKron_C", NULL)); 1293 PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatSetPreallocationCOO_C", NULL)); 1294 PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatSetValuesCOO_C", NULL)); 1295 PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatFactorGetSolverType_C", NULL)); 1296 /* these calls do not belong here: the subclasses Duplicate/Destroy are wrong */ 1297 PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaijsell_seqaij_C", NULL)); 1298 PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaijperm_seqaij_C", NULL)); 1299 PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_seqaijviennacl_C", NULL)); 1300 PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_seqaijviennacl_seqdense_C", NULL)); 1301 PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_seqaijviennacl_seqaij_C", NULL)); 1302 PetscFunctionReturn(PETSC_SUCCESS); 1303 } 1304 1305 PetscErrorCode MatSetOption_SeqAIJ(Mat A, MatOption op, PetscBool flg) 1306 { 1307 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 1308 1309 PetscFunctionBegin; 1310 switch (op) { 1311 case MAT_ROW_ORIENTED: 1312 a->roworiented = flg; 1313 break; 1314 case MAT_KEEP_NONZERO_PATTERN: 1315 a->keepnonzeropattern = flg; 1316 break; 1317 case MAT_NEW_NONZERO_LOCATIONS: 1318 a->nonew = (flg ? 0 : 1); 1319 break; 1320 case MAT_NEW_NONZERO_LOCATION_ERR: 1321 a->nonew = (flg ? -1 : 0); 1322 break; 1323 case MAT_NEW_NONZERO_ALLOCATION_ERR: 1324 a->nonew = (flg ? -2 : 0); 1325 break; 1326 case MAT_UNUSED_NONZERO_LOCATION_ERR: 1327 a->nounused = (flg ? -1 : 0); 1328 break; 1329 case MAT_IGNORE_ZERO_ENTRIES: 1330 a->ignorezeroentries = flg; 1331 break; 1332 case MAT_SPD: 1333 case MAT_SYMMETRIC: 1334 case MAT_STRUCTURALLY_SYMMETRIC: 1335 case MAT_HERMITIAN: 1336 case MAT_SYMMETRY_ETERNAL: 1337 case MAT_STRUCTURE_ONLY: 1338 case MAT_STRUCTURAL_SYMMETRY_ETERNAL: 1339 case MAT_SPD_ETERNAL: 1340 /* if the diagonal matrix is square it inherits some of the properties above */ 1341 break; 1342 case MAT_FORCE_DIAGONAL_ENTRIES: 1343 case MAT_IGNORE_OFF_PROC_ENTRIES: 1344 case MAT_USE_HASH_TABLE: 1345 PetscCall(PetscInfo(A, "Option %s ignored\n", MatOptions[op])); 1346 break; 1347 case MAT_USE_INODES: 1348 PetscCall(MatSetOption_SeqAIJ_Inode(A, MAT_USE_INODES, flg)); 1349 break; 1350 case MAT_SUBMAT_SINGLEIS: 1351 A->submat_singleis = flg; 1352 break; 1353 case MAT_SORTED_FULL: 1354 if (flg) A->ops->setvalues = MatSetValues_SeqAIJ_SortedFull; 1355 else A->ops->setvalues = MatSetValues_SeqAIJ; 1356 break; 1357 case MAT_FORM_EXPLICIT_TRANSPOSE: 1358 A->form_explicit_transpose = flg; 1359 break; 1360 default: 1361 SETERRQ(PETSC_COMM_SELF, PETSC_ERR_SUP, "unknown option %d", op); 1362 } 1363 PetscFunctionReturn(PETSC_SUCCESS); 1364 } 1365 1366 static PetscErrorCode MatGetDiagonal_SeqAIJ(Mat A, Vec v) 1367 { 1368 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 1369 PetscInt i, j, n, *ai = a->i, *aj = a->j; 1370 PetscScalar *x; 1371 const PetscScalar *aa; 1372 1373 PetscFunctionBegin; 1374 PetscCall(VecGetLocalSize(v, &n)); 1375 PetscCheck(n == A->rmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Nonconforming matrix and vector"); 1376 PetscCall(MatSeqAIJGetArrayRead(A, &aa)); 1377 if (A->factortype == MAT_FACTOR_ILU || A->factortype == MAT_FACTOR_LU) { 1378 PetscInt *diag = a->diag; 1379 PetscCall(VecGetArrayWrite(v, &x)); 1380 for (i = 0; i < n; i++) x[i] = 1.0 / aa[diag[i]]; 1381 PetscCall(VecRestoreArrayWrite(v, &x)); 1382 PetscCall(MatSeqAIJRestoreArrayRead(A, &aa)); 1383 PetscFunctionReturn(PETSC_SUCCESS); 1384 } 1385 1386 PetscCall(VecGetArrayWrite(v, &x)); 1387 for (i = 0; i < n; i++) { 1388 x[i] = 0.0; 1389 for (j = ai[i]; j < ai[i + 1]; j++) { 1390 if (aj[j] == i) { 1391 x[i] = aa[j]; 1392 break; 1393 } 1394 } 1395 } 1396 PetscCall(VecRestoreArrayWrite(v, &x)); 1397 PetscCall(MatSeqAIJRestoreArrayRead(A, &aa)); 1398 PetscFunctionReturn(PETSC_SUCCESS); 1399 } 1400 1401 #include <../src/mat/impls/aij/seq/ftn-kernels/fmult.h> 1402 PetscErrorCode MatMultTransposeAdd_SeqAIJ(Mat A, Vec xx, Vec zz, Vec yy) 1403 { 1404 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 1405 const MatScalar *aa; 1406 PetscScalar *y; 1407 const PetscScalar *x; 1408 PetscInt m = A->rmap->n; 1409 #if !defined(PETSC_USE_FORTRAN_KERNEL_MULTTRANSPOSEAIJ) 1410 const MatScalar *v; 1411 PetscScalar alpha; 1412 PetscInt n, i, j; 1413 const PetscInt *idx, *ii, *ridx = NULL; 1414 Mat_CompressedRow cprow = a->compressedrow; 1415 PetscBool usecprow = cprow.use; 1416 #endif 1417 1418 PetscFunctionBegin; 1419 if (zz != yy) PetscCall(VecCopy(zz, yy)); 1420 PetscCall(VecGetArrayRead(xx, &x)); 1421 PetscCall(VecGetArray(yy, &y)); 1422 PetscCall(MatSeqAIJGetArrayRead(A, &aa)); 1423 1424 #if defined(PETSC_USE_FORTRAN_KERNEL_MULTTRANSPOSEAIJ) 1425 fortranmulttransposeaddaij_(&m, x, a->i, a->j, aa, y); 1426 #else 1427 if (usecprow) { 1428 m = cprow.nrows; 1429 ii = cprow.i; 1430 ridx = cprow.rindex; 1431 } else { 1432 ii = a->i; 1433 } 1434 for (i = 0; i < m; i++) { 1435 idx = a->j + ii[i]; 1436 v = aa + ii[i]; 1437 n = ii[i + 1] - ii[i]; 1438 if (usecprow) { 1439 alpha = x[ridx[i]]; 1440 } else { 1441 alpha = x[i]; 1442 } 1443 for (j = 0; j < n; j++) y[idx[j]] += alpha * v[j]; 1444 } 1445 #endif 1446 PetscCall(PetscLogFlops(2.0 * a->nz)); 1447 PetscCall(VecRestoreArrayRead(xx, &x)); 1448 PetscCall(VecRestoreArray(yy, &y)); 1449 PetscCall(MatSeqAIJRestoreArrayRead(A, &aa)); 1450 PetscFunctionReturn(PETSC_SUCCESS); 1451 } 1452 1453 PetscErrorCode MatMultTranspose_SeqAIJ(Mat A, Vec xx, Vec yy) 1454 { 1455 PetscFunctionBegin; 1456 PetscCall(VecSet(yy, 0.0)); 1457 PetscCall(MatMultTransposeAdd_SeqAIJ(A, xx, yy, yy)); 1458 PetscFunctionReturn(PETSC_SUCCESS); 1459 } 1460 1461 #include <../src/mat/impls/aij/seq/ftn-kernels/fmult.h> 1462 1463 PetscErrorCode MatMult_SeqAIJ(Mat A, Vec xx, Vec yy) 1464 { 1465 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 1466 PetscScalar *y; 1467 const PetscScalar *x; 1468 const MatScalar *aa, *a_a; 1469 PetscInt m = A->rmap->n; 1470 const PetscInt *aj, *ii, *ridx = NULL; 1471 PetscInt n, i; 1472 PetscScalar sum; 1473 PetscBool usecprow = a->compressedrow.use; 1474 1475 #if defined(PETSC_HAVE_PRAGMA_DISJOINT) 1476 #pragma disjoint(*x, *y, *aa) 1477 #endif 1478 1479 PetscFunctionBegin; 1480 if (a->inode.use && a->inode.checked) { 1481 PetscCall(MatMult_SeqAIJ_Inode(A, xx, yy)); 1482 PetscFunctionReturn(PETSC_SUCCESS); 1483 } 1484 PetscCall(MatSeqAIJGetArrayRead(A, &a_a)); 1485 PetscCall(VecGetArrayRead(xx, &x)); 1486 PetscCall(VecGetArray(yy, &y)); 1487 ii = a->i; 1488 if (usecprow) { /* use compressed row format */ 1489 PetscCall(PetscArrayzero(y, m)); 1490 m = a->compressedrow.nrows; 1491 ii = a->compressedrow.i; 1492 ridx = a->compressedrow.rindex; 1493 for (i = 0; i < m; i++) { 1494 n = ii[i + 1] - ii[i]; 1495 aj = a->j + ii[i]; 1496 aa = a_a + ii[i]; 1497 sum = 0.0; 1498 PetscSparseDensePlusDot(sum, x, aa, aj, n); 1499 /* for (j=0; j<n; j++) sum += (*aa++)*x[*aj++]; */ 1500 y[*ridx++] = sum; 1501 } 1502 } else { /* do not use compressed row format */ 1503 #if defined(PETSC_USE_FORTRAN_KERNEL_MULTAIJ) 1504 aj = a->j; 1505 aa = a_a; 1506 fortranmultaij_(&m, x, ii, aj, aa, y); 1507 #else 1508 for (i = 0; i < m; i++) { 1509 n = ii[i + 1] - ii[i]; 1510 aj = a->j + ii[i]; 1511 aa = a_a + ii[i]; 1512 sum = 0.0; 1513 PetscSparseDensePlusDot(sum, x, aa, aj, n); 1514 y[i] = sum; 1515 } 1516 #endif 1517 } 1518 PetscCall(PetscLogFlops(2.0 * a->nz - a->nonzerorowcnt)); 1519 PetscCall(VecRestoreArrayRead(xx, &x)); 1520 PetscCall(VecRestoreArray(yy, &y)); 1521 PetscCall(MatSeqAIJRestoreArrayRead(A, &a_a)); 1522 PetscFunctionReturn(PETSC_SUCCESS); 1523 } 1524 1525 // HACK!!!!! Used by src/mat/tests/ex170.c 1526 PETSC_EXTERN PetscErrorCode MatMultMax_SeqAIJ(Mat A, Vec xx, Vec yy) 1527 { 1528 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 1529 PetscScalar *y; 1530 const PetscScalar *x; 1531 const MatScalar *aa, *a_a; 1532 PetscInt m = A->rmap->n; 1533 const PetscInt *aj, *ii, *ridx = NULL; 1534 PetscInt n, i, nonzerorow = 0; 1535 PetscScalar sum; 1536 PetscBool usecprow = a->compressedrow.use; 1537 1538 #if defined(PETSC_HAVE_PRAGMA_DISJOINT) 1539 #pragma disjoint(*x, *y, *aa) 1540 #endif 1541 1542 PetscFunctionBegin; 1543 PetscCall(MatSeqAIJGetArrayRead(A, &a_a)); 1544 PetscCall(VecGetArrayRead(xx, &x)); 1545 PetscCall(VecGetArray(yy, &y)); 1546 if (usecprow) { /* use compressed row format */ 1547 m = a->compressedrow.nrows; 1548 ii = a->compressedrow.i; 1549 ridx = a->compressedrow.rindex; 1550 for (i = 0; i < m; i++) { 1551 n = ii[i + 1] - ii[i]; 1552 aj = a->j + ii[i]; 1553 aa = a_a + ii[i]; 1554 sum = 0.0; 1555 nonzerorow += (n > 0); 1556 PetscSparseDenseMaxDot(sum, x, aa, aj, n); 1557 /* for (j=0; j<n; j++) sum += (*aa++)*x[*aj++]; */ 1558 y[*ridx++] = sum; 1559 } 1560 } else { /* do not use compressed row format */ 1561 ii = a->i; 1562 for (i = 0; i < m; i++) { 1563 n = ii[i + 1] - ii[i]; 1564 aj = a->j + ii[i]; 1565 aa = a_a + ii[i]; 1566 sum = 0.0; 1567 nonzerorow += (n > 0); 1568 PetscSparseDenseMaxDot(sum, x, aa, aj, n); 1569 y[i] = sum; 1570 } 1571 } 1572 PetscCall(PetscLogFlops(2.0 * a->nz - nonzerorow)); 1573 PetscCall(VecRestoreArrayRead(xx, &x)); 1574 PetscCall(VecRestoreArray(yy, &y)); 1575 PetscCall(MatSeqAIJRestoreArrayRead(A, &a_a)); 1576 PetscFunctionReturn(PETSC_SUCCESS); 1577 } 1578 1579 // HACK!!!!! Used by src/mat/tests/ex170.c 1580 PETSC_EXTERN PetscErrorCode MatMultAddMax_SeqAIJ(Mat A, Vec xx, Vec yy, Vec zz) 1581 { 1582 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 1583 PetscScalar *y, *z; 1584 const PetscScalar *x; 1585 const MatScalar *aa, *a_a; 1586 PetscInt m = A->rmap->n, *aj, *ii; 1587 PetscInt n, i, *ridx = NULL; 1588 PetscScalar sum; 1589 PetscBool usecprow = a->compressedrow.use; 1590 1591 PetscFunctionBegin; 1592 PetscCall(MatSeqAIJGetArrayRead(A, &a_a)); 1593 PetscCall(VecGetArrayRead(xx, &x)); 1594 PetscCall(VecGetArrayPair(yy, zz, &y, &z)); 1595 if (usecprow) { /* use compressed row format */ 1596 if (zz != yy) PetscCall(PetscArraycpy(z, y, m)); 1597 m = a->compressedrow.nrows; 1598 ii = a->compressedrow.i; 1599 ridx = a->compressedrow.rindex; 1600 for (i = 0; i < m; i++) { 1601 n = ii[i + 1] - ii[i]; 1602 aj = a->j + ii[i]; 1603 aa = a_a + ii[i]; 1604 sum = y[*ridx]; 1605 PetscSparseDenseMaxDot(sum, x, aa, aj, n); 1606 z[*ridx++] = sum; 1607 } 1608 } else { /* do not use compressed row format */ 1609 ii = a->i; 1610 for (i = 0; i < m; i++) { 1611 n = ii[i + 1] - ii[i]; 1612 aj = a->j + ii[i]; 1613 aa = a_a + ii[i]; 1614 sum = y[i]; 1615 PetscSparseDenseMaxDot(sum, x, aa, aj, n); 1616 z[i] = sum; 1617 } 1618 } 1619 PetscCall(PetscLogFlops(2.0 * a->nz)); 1620 PetscCall(VecRestoreArrayRead(xx, &x)); 1621 PetscCall(VecRestoreArrayPair(yy, zz, &y, &z)); 1622 PetscCall(MatSeqAIJRestoreArrayRead(A, &a_a)); 1623 PetscFunctionReturn(PETSC_SUCCESS); 1624 } 1625 1626 #include <../src/mat/impls/aij/seq/ftn-kernels/fmultadd.h> 1627 PetscErrorCode MatMultAdd_SeqAIJ(Mat A, Vec xx, Vec yy, Vec zz) 1628 { 1629 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 1630 PetscScalar *y, *z; 1631 const PetscScalar *x; 1632 const MatScalar *aa, *a_a; 1633 const PetscInt *aj, *ii, *ridx = NULL; 1634 PetscInt m = A->rmap->n, n, i; 1635 PetscScalar sum; 1636 PetscBool usecprow = a->compressedrow.use; 1637 1638 PetscFunctionBegin; 1639 if (a->inode.use && a->inode.checked) { 1640 PetscCall(MatMultAdd_SeqAIJ_Inode(A, xx, yy, zz)); 1641 PetscFunctionReturn(PETSC_SUCCESS); 1642 } 1643 PetscCall(MatSeqAIJGetArrayRead(A, &a_a)); 1644 PetscCall(VecGetArrayRead(xx, &x)); 1645 PetscCall(VecGetArrayPair(yy, zz, &y, &z)); 1646 if (usecprow) { /* use compressed row format */ 1647 if (zz != yy) PetscCall(PetscArraycpy(z, y, m)); 1648 m = a->compressedrow.nrows; 1649 ii = a->compressedrow.i; 1650 ridx = a->compressedrow.rindex; 1651 for (i = 0; i < m; i++) { 1652 n = ii[i + 1] - ii[i]; 1653 aj = a->j + ii[i]; 1654 aa = a_a + ii[i]; 1655 sum = y[*ridx]; 1656 PetscSparseDensePlusDot(sum, x, aa, aj, n); 1657 z[*ridx++] = sum; 1658 } 1659 } else { /* do not use compressed row format */ 1660 ii = a->i; 1661 #if defined(PETSC_USE_FORTRAN_KERNEL_MULTADDAIJ) 1662 aj = a->j; 1663 aa = a_a; 1664 fortranmultaddaij_(&m, x, ii, aj, aa, y, z); 1665 #else 1666 for (i = 0; i < m; i++) { 1667 n = ii[i + 1] - ii[i]; 1668 aj = a->j + ii[i]; 1669 aa = a_a + ii[i]; 1670 sum = y[i]; 1671 PetscSparseDensePlusDot(sum, x, aa, aj, n); 1672 z[i] = sum; 1673 } 1674 #endif 1675 } 1676 PetscCall(PetscLogFlops(2.0 * a->nz)); 1677 PetscCall(VecRestoreArrayRead(xx, &x)); 1678 PetscCall(VecRestoreArrayPair(yy, zz, &y, &z)); 1679 PetscCall(MatSeqAIJRestoreArrayRead(A, &a_a)); 1680 PetscFunctionReturn(PETSC_SUCCESS); 1681 } 1682 1683 /* 1684 Adds diagonal pointers to sparse matrix structure. 1685 */ 1686 PetscErrorCode MatMarkDiagonal_SeqAIJ(Mat A) 1687 { 1688 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 1689 PetscInt i, j, m = A->rmap->n; 1690 PetscBool alreadySet = PETSC_TRUE; 1691 1692 PetscFunctionBegin; 1693 if (!a->diag) { 1694 PetscCall(PetscMalloc1(m, &a->diag)); 1695 alreadySet = PETSC_FALSE; 1696 } 1697 for (i = 0; i < A->rmap->n; i++) { 1698 /* If A's diagonal is already correctly set, this fast track enables cheap and repeated MatMarkDiagonal_SeqAIJ() calls */ 1699 if (alreadySet) { 1700 PetscInt pos = a->diag[i]; 1701 if (pos >= a->i[i] && pos < a->i[i + 1] && a->j[pos] == i) continue; 1702 } 1703 1704 a->diag[i] = a->i[i + 1]; 1705 for (j = a->i[i]; j < a->i[i + 1]; j++) { 1706 if (a->j[j] == i) { 1707 a->diag[i] = j; 1708 break; 1709 } 1710 } 1711 } 1712 PetscFunctionReturn(PETSC_SUCCESS); 1713 } 1714 1715 static PetscErrorCode MatShift_SeqAIJ(Mat A, PetscScalar v) 1716 { 1717 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 1718 const PetscInt *diag = (const PetscInt *)a->diag; 1719 const PetscInt *ii = (const PetscInt *)a->i; 1720 PetscInt i, *mdiag = NULL; 1721 PetscInt cnt = 0; /* how many diagonals are missing */ 1722 1723 PetscFunctionBegin; 1724 if (!A->preallocated || !a->nz) { 1725 PetscCall(MatSeqAIJSetPreallocation(A, 1, NULL)); 1726 PetscCall(MatShift_Basic(A, v)); 1727 PetscFunctionReturn(PETSC_SUCCESS); 1728 } 1729 1730 if (a->diagonaldense) { 1731 cnt = 0; 1732 } else { 1733 PetscCall(PetscCalloc1(A->rmap->n, &mdiag)); 1734 for (i = 0; i < A->rmap->n; i++) { 1735 if (i < A->cmap->n && diag[i] >= ii[i + 1]) { /* 'out of range' rows never have diagonals */ 1736 cnt++; 1737 mdiag[i] = 1; 1738 } 1739 } 1740 } 1741 if (!cnt) { 1742 PetscCall(MatShift_Basic(A, v)); 1743 } else { 1744 PetscScalar *olda = a->a; /* preserve pointers to current matrix nonzeros structure and values */ 1745 PetscInt *oldj = a->j, *oldi = a->i; 1746 PetscBool singlemalloc = a->singlemalloc, free_a = a->free_a, free_ij = a->free_ij; 1747 const PetscScalar *Aa; 1748 1749 PetscCall(MatSeqAIJGetArrayRead(A, &Aa)); // sync the host 1750 PetscCall(MatSeqAIJRestoreArrayRead(A, &Aa)); 1751 1752 a->a = NULL; 1753 a->j = NULL; 1754 a->i = NULL; 1755 /* increase the values in imax for each row where a diagonal is being inserted then reallocate the matrix data structures */ 1756 for (i = 0; i < PetscMin(A->rmap->n, A->cmap->n); i++) a->imax[i] += mdiag[i]; 1757 PetscCall(MatSeqAIJSetPreallocation_SeqAIJ(A, 0, a->imax)); 1758 1759 /* copy old values into new matrix data structure */ 1760 for (i = 0; i < A->rmap->n; i++) { 1761 PetscCall(MatSetValues(A, 1, &i, a->imax[i] - mdiag[i], &oldj[oldi[i]], &olda[oldi[i]], ADD_VALUES)); 1762 if (i < A->cmap->n) PetscCall(MatSetValue(A, i, i, v, ADD_VALUES)); 1763 } 1764 PetscCall(MatAssemblyBegin(A, MAT_FINAL_ASSEMBLY)); 1765 PetscCall(MatAssemblyEnd(A, MAT_FINAL_ASSEMBLY)); 1766 if (singlemalloc) { 1767 PetscCall(PetscFree3(olda, oldj, oldi)); 1768 } else { 1769 if (free_a) PetscCall(PetscFree(olda)); 1770 if (free_ij) PetscCall(PetscFree(oldj)); 1771 if (free_ij) PetscCall(PetscFree(oldi)); 1772 } 1773 } 1774 PetscCall(PetscFree(mdiag)); 1775 a->diagonaldense = PETSC_TRUE; 1776 PetscFunctionReturn(PETSC_SUCCESS); 1777 } 1778 1779 /* 1780 Checks for missing diagonals 1781 */ 1782 PetscErrorCode MatMissingDiagonal_SeqAIJ(Mat A, PetscBool *missing, PetscInt *d) 1783 { 1784 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 1785 PetscInt *diag, *ii = a->i, i; 1786 1787 PetscFunctionBegin; 1788 *missing = PETSC_FALSE; 1789 if (A->rmap->n > 0 && !ii) { 1790 *missing = PETSC_TRUE; 1791 if (d) *d = 0; 1792 PetscCall(PetscInfo(A, "Matrix has no entries therefore is missing diagonal\n")); 1793 } else { 1794 PetscInt n; 1795 n = PetscMin(A->rmap->n, A->cmap->n); 1796 diag = a->diag; 1797 for (i = 0; i < n; i++) { 1798 if (diag[i] >= ii[i + 1]) { 1799 *missing = PETSC_TRUE; 1800 if (d) *d = i; 1801 PetscCall(PetscInfo(A, "Matrix is missing diagonal number %" PetscInt_FMT "\n", i)); 1802 break; 1803 } 1804 } 1805 } 1806 PetscFunctionReturn(PETSC_SUCCESS); 1807 } 1808 1809 #include <petscblaslapack.h> 1810 #include <petsc/private/kernels/blockinvert.h> 1811 1812 /* 1813 Note that values is allocated externally by the PC and then passed into this routine 1814 */ 1815 static PetscErrorCode MatInvertVariableBlockDiagonal_SeqAIJ(Mat A, PetscInt nblocks, const PetscInt *bsizes, PetscScalar *diag) 1816 { 1817 PetscInt n = A->rmap->n, i, ncnt = 0, *indx, j, bsizemax = 0, *v_pivots; 1818 PetscBool allowzeropivot, zeropivotdetected = PETSC_FALSE; 1819 const PetscReal shift = 0.0; 1820 PetscInt ipvt[5]; 1821 PetscCount flops = 0; 1822 PetscScalar work[25], *v_work; 1823 1824 PetscFunctionBegin; 1825 allowzeropivot = PetscNot(A->erroriffailure); 1826 for (i = 0; i < nblocks; i++) ncnt += bsizes[i]; 1827 PetscCheck(ncnt == n, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Total blocksizes %" PetscInt_FMT " doesn't match number matrix rows %" PetscInt_FMT, ncnt, n); 1828 for (i = 0; i < nblocks; i++) bsizemax = PetscMax(bsizemax, bsizes[i]); 1829 PetscCall(PetscMalloc1(bsizemax, &indx)); 1830 if (bsizemax > 7) PetscCall(PetscMalloc2(bsizemax, &v_work, bsizemax, &v_pivots)); 1831 ncnt = 0; 1832 for (i = 0; i < nblocks; i++) { 1833 for (j = 0; j < bsizes[i]; j++) indx[j] = ncnt + j; 1834 PetscCall(MatGetValues(A, bsizes[i], indx, bsizes[i], indx, diag)); 1835 switch (bsizes[i]) { 1836 case 1: 1837 *diag = 1.0 / (*diag); 1838 break; 1839 case 2: 1840 PetscCall(PetscKernel_A_gets_inverse_A_2(diag, shift, allowzeropivot, &zeropivotdetected)); 1841 if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT; 1842 PetscCall(PetscKernel_A_gets_transpose_A_2(diag)); 1843 break; 1844 case 3: 1845 PetscCall(PetscKernel_A_gets_inverse_A_3(diag, shift, allowzeropivot, &zeropivotdetected)); 1846 if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT; 1847 PetscCall(PetscKernel_A_gets_transpose_A_3(diag)); 1848 break; 1849 case 4: 1850 PetscCall(PetscKernel_A_gets_inverse_A_4(diag, shift, allowzeropivot, &zeropivotdetected)); 1851 if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT; 1852 PetscCall(PetscKernel_A_gets_transpose_A_4(diag)); 1853 break; 1854 case 5: 1855 PetscCall(PetscKernel_A_gets_inverse_A_5(diag, ipvt, work, shift, allowzeropivot, &zeropivotdetected)); 1856 if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT; 1857 PetscCall(PetscKernel_A_gets_transpose_A_5(diag)); 1858 break; 1859 case 6: 1860 PetscCall(PetscKernel_A_gets_inverse_A_6(diag, shift, allowzeropivot, &zeropivotdetected)); 1861 if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT; 1862 PetscCall(PetscKernel_A_gets_transpose_A_6(diag)); 1863 break; 1864 case 7: 1865 PetscCall(PetscKernel_A_gets_inverse_A_7(diag, shift, allowzeropivot, &zeropivotdetected)); 1866 if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT; 1867 PetscCall(PetscKernel_A_gets_transpose_A_7(diag)); 1868 break; 1869 default: 1870 PetscCall(PetscKernel_A_gets_inverse_A(bsizes[i], diag, v_pivots, v_work, allowzeropivot, &zeropivotdetected)); 1871 if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT; 1872 PetscCall(PetscKernel_A_gets_transpose_A_N(diag, bsizes[i])); 1873 } 1874 ncnt += bsizes[i]; 1875 diag += bsizes[i] * bsizes[i]; 1876 flops += 2 * PetscPowInt(bsizes[i], 3) / 3; 1877 } 1878 PetscCall(PetscLogFlops(flops)); 1879 if (bsizemax > 7) PetscCall(PetscFree2(v_work, v_pivots)); 1880 PetscCall(PetscFree(indx)); 1881 PetscFunctionReturn(PETSC_SUCCESS); 1882 } 1883 1884 /* 1885 Negative shift indicates do not generate an error if there is a zero diagonal, just invert it anyways 1886 */ 1887 static PetscErrorCode MatInvertDiagonal_SeqAIJ(Mat A, PetscScalar omega, PetscScalar fshift) 1888 { 1889 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 1890 PetscInt i, *diag, m = A->rmap->n; 1891 const MatScalar *v; 1892 PetscScalar *idiag, *mdiag; 1893 1894 PetscFunctionBegin; 1895 if (a->idiagvalid) PetscFunctionReturn(PETSC_SUCCESS); 1896 PetscCall(MatMarkDiagonal_SeqAIJ(A)); 1897 diag = a->diag; 1898 if (!a->idiag) { PetscCall(PetscMalloc3(m, &a->idiag, m, &a->mdiag, m, &a->ssor_work)); } 1899 1900 mdiag = a->mdiag; 1901 idiag = a->idiag; 1902 PetscCall(MatSeqAIJGetArrayRead(A, &v)); 1903 if (omega == 1.0 && PetscRealPart(fshift) <= 0.0) { 1904 for (i = 0; i < m; i++) { 1905 mdiag[i] = v[diag[i]]; 1906 if (!PetscAbsScalar(mdiag[i])) { /* zero diagonal */ 1907 if (PetscRealPart(fshift)) { 1908 PetscCall(PetscInfo(A, "Zero diagonal on row %" PetscInt_FMT "\n", i)); 1909 A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT; 1910 A->factorerror_zeropivot_value = 0.0; 1911 A->factorerror_zeropivot_row = i; 1912 } else SETERRQ(PETSC_COMM_SELF, PETSC_ERR_ARG_INCOMP, "Zero diagonal on row %" PetscInt_FMT, i); 1913 } 1914 idiag[i] = 1.0 / v[diag[i]]; 1915 } 1916 PetscCall(PetscLogFlops(m)); 1917 } else { 1918 for (i = 0; i < m; i++) { 1919 mdiag[i] = v[diag[i]]; 1920 idiag[i] = omega / (fshift + v[diag[i]]); 1921 } 1922 PetscCall(PetscLogFlops(2.0 * m)); 1923 } 1924 a->idiagvalid = PETSC_TRUE; 1925 PetscCall(MatSeqAIJRestoreArrayRead(A, &v)); 1926 PetscFunctionReturn(PETSC_SUCCESS); 1927 } 1928 1929 #include <../src/mat/impls/aij/seq/ftn-kernels/frelax.h> 1930 PetscErrorCode MatSOR_SeqAIJ(Mat A, Vec bb, PetscReal omega, MatSORType flag, PetscReal fshift, PetscInt its, PetscInt lits, Vec xx) 1931 { 1932 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 1933 PetscScalar *x, d, sum, *t, scale; 1934 const MatScalar *v, *idiag = NULL, *mdiag, *aa; 1935 const PetscScalar *b, *bs, *xb, *ts; 1936 PetscInt n, m = A->rmap->n, i; 1937 const PetscInt *idx, *diag; 1938 1939 PetscFunctionBegin; 1940 if (a->inode.use && a->inode.checked && omega == 1.0 && fshift == 0.0) { 1941 PetscCall(MatSOR_SeqAIJ_Inode(A, bb, omega, flag, fshift, its, lits, xx)); 1942 PetscFunctionReturn(PETSC_SUCCESS); 1943 } 1944 its = its * lits; 1945 1946 if (fshift != a->fshift || omega != a->omega) a->idiagvalid = PETSC_FALSE; /* must recompute idiag[] */ 1947 if (!a->idiagvalid) PetscCall(MatInvertDiagonal_SeqAIJ(A, omega, fshift)); 1948 a->fshift = fshift; 1949 a->omega = omega; 1950 1951 diag = a->diag; 1952 t = a->ssor_work; 1953 idiag = a->idiag; 1954 mdiag = a->mdiag; 1955 1956 PetscCall(MatSeqAIJGetArrayRead(A, &aa)); 1957 PetscCall(VecGetArray(xx, &x)); 1958 PetscCall(VecGetArrayRead(bb, &b)); 1959 /* We count flops by assuming the upper triangular and lower triangular parts have the same number of nonzeros */ 1960 if (flag == SOR_APPLY_UPPER) { 1961 /* apply (U + D/omega) to the vector */ 1962 bs = b; 1963 for (i = 0; i < m; i++) { 1964 d = fshift + mdiag[i]; 1965 n = a->i[i + 1] - diag[i] - 1; 1966 idx = a->j + diag[i] + 1; 1967 v = aa + diag[i] + 1; 1968 sum = b[i] * d / omega; 1969 PetscSparseDensePlusDot(sum, bs, v, idx, n); 1970 x[i] = sum; 1971 } 1972 PetscCall(VecRestoreArray(xx, &x)); 1973 PetscCall(VecRestoreArrayRead(bb, &b)); 1974 PetscCall(MatSeqAIJRestoreArrayRead(A, &aa)); 1975 PetscCall(PetscLogFlops(a->nz)); 1976 PetscFunctionReturn(PETSC_SUCCESS); 1977 } 1978 1979 PetscCheck(flag != SOR_APPLY_LOWER, PETSC_COMM_SELF, PETSC_ERR_SUP, "SOR_APPLY_LOWER is not implemented"); 1980 if (flag & SOR_EISENSTAT) { 1981 /* Let A = L + U + D; where L is lower triangular, 1982 U is upper triangular, E = D/omega; This routine applies 1983 1984 (L + E)^{-1} A (U + E)^{-1} 1985 1986 to a vector efficiently using Eisenstat's trick. 1987 */ 1988 scale = (2.0 / omega) - 1.0; 1989 1990 /* x = (E + U)^{-1} b */ 1991 for (i = m - 1; i >= 0; i--) { 1992 n = a->i[i + 1] - diag[i] - 1; 1993 idx = a->j + diag[i] + 1; 1994 v = aa + diag[i] + 1; 1995 sum = b[i]; 1996 PetscSparseDenseMinusDot(sum, x, v, idx, n); 1997 x[i] = sum * idiag[i]; 1998 } 1999 2000 /* t = b - (2*E - D)x */ 2001 v = aa; 2002 for (i = 0; i < m; i++) t[i] = b[i] - scale * (v[*diag++]) * x[i]; 2003 2004 /* t = (E + L)^{-1}t */ 2005 ts = t; 2006 diag = a->diag; 2007 for (i = 0; i < m; i++) { 2008 n = diag[i] - a->i[i]; 2009 idx = a->j + a->i[i]; 2010 v = aa + a->i[i]; 2011 sum = t[i]; 2012 PetscSparseDenseMinusDot(sum, ts, v, idx, n); 2013 t[i] = sum * idiag[i]; 2014 /* x = x + t */ 2015 x[i] += t[i]; 2016 } 2017 2018 PetscCall(PetscLogFlops(6.0 * m - 1 + 2.0 * a->nz)); 2019 PetscCall(VecRestoreArray(xx, &x)); 2020 PetscCall(VecRestoreArrayRead(bb, &b)); 2021 PetscFunctionReturn(PETSC_SUCCESS); 2022 } 2023 if (flag & SOR_ZERO_INITIAL_GUESS) { 2024 if (flag & SOR_FORWARD_SWEEP || flag & SOR_LOCAL_FORWARD_SWEEP) { 2025 for (i = 0; i < m; i++) { 2026 n = diag[i] - a->i[i]; 2027 idx = a->j + a->i[i]; 2028 v = aa + a->i[i]; 2029 sum = b[i]; 2030 PetscSparseDenseMinusDot(sum, x, v, idx, n); 2031 t[i] = sum; 2032 x[i] = sum * idiag[i]; 2033 } 2034 xb = t; 2035 PetscCall(PetscLogFlops(a->nz)); 2036 } else xb = b; 2037 if (flag & SOR_BACKWARD_SWEEP || flag & SOR_LOCAL_BACKWARD_SWEEP) { 2038 for (i = m - 1; i >= 0; i--) { 2039 n = a->i[i + 1] - diag[i] - 1; 2040 idx = a->j + diag[i] + 1; 2041 v = aa + diag[i] + 1; 2042 sum = xb[i]; 2043 PetscSparseDenseMinusDot(sum, x, v, idx, n); 2044 if (xb == b) { 2045 x[i] = sum * idiag[i]; 2046 } else { 2047 x[i] = (1 - omega) * x[i] + sum * idiag[i]; /* omega in idiag */ 2048 } 2049 } 2050 PetscCall(PetscLogFlops(a->nz)); /* assumes 1/2 in upper */ 2051 } 2052 its--; 2053 } 2054 while (its--) { 2055 if (flag & SOR_FORWARD_SWEEP || flag & SOR_LOCAL_FORWARD_SWEEP) { 2056 for (i = 0; i < m; i++) { 2057 /* lower */ 2058 n = diag[i] - a->i[i]; 2059 idx = a->j + a->i[i]; 2060 v = aa + a->i[i]; 2061 sum = b[i]; 2062 PetscSparseDenseMinusDot(sum, x, v, idx, n); 2063 t[i] = sum; /* save application of the lower-triangular part */ 2064 /* upper */ 2065 n = a->i[i + 1] - diag[i] - 1; 2066 idx = a->j + diag[i] + 1; 2067 v = aa + diag[i] + 1; 2068 PetscSparseDenseMinusDot(sum, x, v, idx, n); 2069 x[i] = (1. - omega) * x[i] + sum * idiag[i]; /* omega in idiag */ 2070 } 2071 xb = t; 2072 PetscCall(PetscLogFlops(2.0 * a->nz)); 2073 } else xb = b; 2074 if (flag & SOR_BACKWARD_SWEEP || flag & SOR_LOCAL_BACKWARD_SWEEP) { 2075 for (i = m - 1; i >= 0; i--) { 2076 sum = xb[i]; 2077 if (xb == b) { 2078 /* whole matrix (no checkpointing available) */ 2079 n = a->i[i + 1] - a->i[i]; 2080 idx = a->j + a->i[i]; 2081 v = aa + a->i[i]; 2082 PetscSparseDenseMinusDot(sum, x, v, idx, n); 2083 x[i] = (1. - omega) * x[i] + (sum + mdiag[i] * x[i]) * idiag[i]; 2084 } else { /* lower-triangular part has been saved, so only apply upper-triangular */ 2085 n = a->i[i + 1] - diag[i] - 1; 2086 idx = a->j + diag[i] + 1; 2087 v = aa + diag[i] + 1; 2088 PetscSparseDenseMinusDot(sum, x, v, idx, n); 2089 x[i] = (1. - omega) * x[i] + sum * idiag[i]; /* omega in idiag */ 2090 } 2091 } 2092 if (xb == b) { 2093 PetscCall(PetscLogFlops(2.0 * a->nz)); 2094 } else { 2095 PetscCall(PetscLogFlops(a->nz)); /* assumes 1/2 in upper */ 2096 } 2097 } 2098 } 2099 PetscCall(MatSeqAIJRestoreArrayRead(A, &aa)); 2100 PetscCall(VecRestoreArray(xx, &x)); 2101 PetscCall(VecRestoreArrayRead(bb, &b)); 2102 PetscFunctionReturn(PETSC_SUCCESS); 2103 } 2104 2105 static PetscErrorCode MatGetInfo_SeqAIJ(Mat A, MatInfoType flag, MatInfo *info) 2106 { 2107 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 2108 2109 PetscFunctionBegin; 2110 info->block_size = 1.0; 2111 info->nz_allocated = a->maxnz; 2112 info->nz_used = a->nz; 2113 info->nz_unneeded = (a->maxnz - a->nz); 2114 info->assemblies = A->num_ass; 2115 info->mallocs = A->info.mallocs; 2116 info->memory = 0; /* REVIEW ME */ 2117 if (A->factortype) { 2118 info->fill_ratio_given = A->info.fill_ratio_given; 2119 info->fill_ratio_needed = A->info.fill_ratio_needed; 2120 info->factor_mallocs = A->info.factor_mallocs; 2121 } else { 2122 info->fill_ratio_given = 0; 2123 info->fill_ratio_needed = 0; 2124 info->factor_mallocs = 0; 2125 } 2126 PetscFunctionReturn(PETSC_SUCCESS); 2127 } 2128 2129 static PetscErrorCode MatZeroRows_SeqAIJ(Mat A, PetscInt N, const PetscInt rows[], PetscScalar diag, Vec x, Vec b) 2130 { 2131 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 2132 PetscInt i, m = A->rmap->n - 1; 2133 const PetscScalar *xx; 2134 PetscScalar *bb, *aa; 2135 PetscInt d = 0; 2136 2137 PetscFunctionBegin; 2138 if (x && b) { 2139 PetscCall(VecGetArrayRead(x, &xx)); 2140 PetscCall(VecGetArray(b, &bb)); 2141 for (i = 0; i < N; i++) { 2142 PetscCheck(rows[i] >= 0 && rows[i] <= m, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "row %" PetscInt_FMT " out of range", rows[i]); 2143 if (rows[i] >= A->cmap->n) continue; 2144 bb[rows[i]] = diag * xx[rows[i]]; 2145 } 2146 PetscCall(VecRestoreArrayRead(x, &xx)); 2147 PetscCall(VecRestoreArray(b, &bb)); 2148 } 2149 2150 PetscCall(MatSeqAIJGetArray(A, &aa)); 2151 if (a->keepnonzeropattern) { 2152 for (i = 0; i < N; i++) { 2153 PetscCheck(rows[i] >= 0 && rows[i] <= m, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "row %" PetscInt_FMT " out of range", rows[i]); 2154 PetscCall(PetscArrayzero(&aa[a->i[rows[i]]], a->ilen[rows[i]])); 2155 } 2156 if (diag != 0.0) { 2157 for (i = 0; i < N; i++) { 2158 d = rows[i]; 2159 if (rows[i] >= A->cmap->n) continue; 2160 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); 2161 } 2162 for (i = 0; i < N; i++) { 2163 if (rows[i] >= A->cmap->n) continue; 2164 aa[a->diag[rows[i]]] = diag; 2165 } 2166 } 2167 } else { 2168 if (diag != 0.0) { 2169 for (i = 0; i < N; i++) { 2170 PetscCheck(rows[i] >= 0 && rows[i] <= m, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "row %" PetscInt_FMT " out of range", rows[i]); 2171 if (a->ilen[rows[i]] > 0) { 2172 if (rows[i] >= A->cmap->n) { 2173 a->ilen[rows[i]] = 0; 2174 } else { 2175 a->ilen[rows[i]] = 1; 2176 aa[a->i[rows[i]]] = diag; 2177 a->j[a->i[rows[i]]] = rows[i]; 2178 } 2179 } else if (rows[i] < A->cmap->n) { /* in case row was completely empty */ 2180 PetscCall(MatSetValues_SeqAIJ(A, 1, &rows[i], 1, &rows[i], &diag, INSERT_VALUES)); 2181 } 2182 } 2183 } else { 2184 for (i = 0; i < N; i++) { 2185 PetscCheck(rows[i] >= 0 && rows[i] <= m, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "row %" PetscInt_FMT " out of range", rows[i]); 2186 a->ilen[rows[i]] = 0; 2187 } 2188 } 2189 A->nonzerostate++; 2190 } 2191 PetscCall(MatSeqAIJRestoreArray(A, &aa)); 2192 PetscUseTypeMethod(A, assemblyend, MAT_FINAL_ASSEMBLY); 2193 PetscFunctionReturn(PETSC_SUCCESS); 2194 } 2195 2196 static PetscErrorCode MatZeroRowsColumns_SeqAIJ(Mat A, PetscInt N, const PetscInt rows[], PetscScalar diag, Vec x, Vec b) 2197 { 2198 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 2199 PetscInt i, j, m = A->rmap->n - 1, d = 0; 2200 PetscBool missing, *zeroed, vecs = PETSC_FALSE; 2201 const PetscScalar *xx; 2202 PetscScalar *bb, *aa; 2203 2204 PetscFunctionBegin; 2205 if (!N) PetscFunctionReturn(PETSC_SUCCESS); 2206 PetscCall(MatSeqAIJGetArray(A, &aa)); 2207 if (x && b) { 2208 PetscCall(VecGetArrayRead(x, &xx)); 2209 PetscCall(VecGetArray(b, &bb)); 2210 vecs = PETSC_TRUE; 2211 } 2212 PetscCall(PetscCalloc1(A->rmap->n, &zeroed)); 2213 for (i = 0; i < N; i++) { 2214 PetscCheck(rows[i] >= 0 && rows[i] <= m, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "row %" PetscInt_FMT " out of range", rows[i]); 2215 PetscCall(PetscArrayzero(&aa[a->i[rows[i]]], a->ilen[rows[i]])); 2216 2217 zeroed[rows[i]] = PETSC_TRUE; 2218 } 2219 for (i = 0; i < A->rmap->n; i++) { 2220 if (!zeroed[i]) { 2221 for (j = a->i[i]; j < a->i[i + 1]; j++) { 2222 if (a->j[j] < A->rmap->n && zeroed[a->j[j]]) { 2223 if (vecs) bb[i] -= aa[j] * xx[a->j[j]]; 2224 aa[j] = 0.0; 2225 } 2226 } 2227 } else if (vecs && i < A->cmap->N) bb[i] = diag * xx[i]; 2228 } 2229 if (x && b) { 2230 PetscCall(VecRestoreArrayRead(x, &xx)); 2231 PetscCall(VecRestoreArray(b, &bb)); 2232 } 2233 PetscCall(PetscFree(zeroed)); 2234 if (diag != 0.0) { 2235 PetscCall(MatMissingDiagonal_SeqAIJ(A, &missing, &d)); 2236 if (missing) { 2237 for (i = 0; i < N; i++) { 2238 if (rows[i] >= A->cmap->N) continue; 2239 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]); 2240 PetscCall(MatSetValues_SeqAIJ(A, 1, &rows[i], 1, &rows[i], &diag, INSERT_VALUES)); 2241 } 2242 } else { 2243 for (i = 0; i < N; i++) aa[a->diag[rows[i]]] = diag; 2244 } 2245 } 2246 PetscCall(MatSeqAIJRestoreArray(A, &aa)); 2247 PetscUseTypeMethod(A, assemblyend, MAT_FINAL_ASSEMBLY); 2248 PetscFunctionReturn(PETSC_SUCCESS); 2249 } 2250 2251 PetscErrorCode MatGetRow_SeqAIJ(Mat A, PetscInt row, PetscInt *nz, PetscInt **idx, PetscScalar **v) 2252 { 2253 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 2254 const PetscScalar *aa; 2255 2256 PetscFunctionBegin; 2257 PetscCall(MatSeqAIJGetArrayRead(A, &aa)); 2258 *nz = a->i[row + 1] - a->i[row]; 2259 if (v) *v = aa ? (PetscScalar *)(aa + a->i[row]) : NULL; 2260 if (idx) { 2261 if (*nz && a->j) *idx = a->j + a->i[row]; 2262 else *idx = NULL; 2263 } 2264 PetscCall(MatSeqAIJRestoreArrayRead(A, &aa)); 2265 PetscFunctionReturn(PETSC_SUCCESS); 2266 } 2267 2268 PetscErrorCode MatRestoreRow_SeqAIJ(Mat A, PetscInt row, PetscInt *nz, PetscInt **idx, PetscScalar **v) 2269 { 2270 PetscFunctionBegin; 2271 PetscFunctionReturn(PETSC_SUCCESS); 2272 } 2273 2274 static PetscErrorCode MatNorm_SeqAIJ(Mat A, NormType type, PetscReal *nrm) 2275 { 2276 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 2277 const MatScalar *v; 2278 PetscReal sum = 0.0; 2279 PetscInt i, j; 2280 2281 PetscFunctionBegin; 2282 PetscCall(MatSeqAIJGetArrayRead(A, &v)); 2283 if (type == NORM_FROBENIUS) { 2284 #if defined(PETSC_USE_REAL___FP16) 2285 PetscBLASInt one = 1, nz = a->nz; 2286 PetscCallBLAS("BLASnrm2", *nrm = BLASnrm2_(&nz, v, &one)); 2287 #else 2288 for (i = 0; i < a->nz; i++) { 2289 sum += PetscRealPart(PetscConj(*v) * (*v)); 2290 v++; 2291 } 2292 *nrm = PetscSqrtReal(sum); 2293 #endif 2294 PetscCall(PetscLogFlops(2.0 * a->nz)); 2295 } else if (type == NORM_1) { 2296 PetscReal *tmp; 2297 PetscInt *jj = a->j; 2298 PetscCall(PetscCalloc1(A->cmap->n + 1, &tmp)); 2299 *nrm = 0.0; 2300 for (j = 0; j < a->nz; j++) { 2301 tmp[*jj++] += PetscAbsScalar(*v); 2302 v++; 2303 } 2304 for (j = 0; j < A->cmap->n; j++) { 2305 if (tmp[j] > *nrm) *nrm = tmp[j]; 2306 } 2307 PetscCall(PetscFree(tmp)); 2308 PetscCall(PetscLogFlops(PetscMax(a->nz - 1, 0))); 2309 } else if (type == NORM_INFINITY) { 2310 *nrm = 0.0; 2311 for (j = 0; j < A->rmap->n; j++) { 2312 const PetscScalar *v2 = v + a->i[j]; 2313 sum = 0.0; 2314 for (i = 0; i < a->i[j + 1] - a->i[j]; i++) { 2315 sum += PetscAbsScalar(*v2); 2316 v2++; 2317 } 2318 if (sum > *nrm) *nrm = sum; 2319 } 2320 PetscCall(PetscLogFlops(PetscMax(a->nz - 1, 0))); 2321 } else SETERRQ(PETSC_COMM_SELF, PETSC_ERR_SUP, "No support for two norm"); 2322 PetscCall(MatSeqAIJRestoreArrayRead(A, &v)); 2323 PetscFunctionReturn(PETSC_SUCCESS); 2324 } 2325 2326 static PetscErrorCode MatIsTranspose_SeqAIJ(Mat A, Mat B, PetscReal tol, PetscBool *f) 2327 { 2328 Mat_SeqAIJ *aij = (Mat_SeqAIJ *)A->data, *bij = (Mat_SeqAIJ *)B->data; 2329 PetscInt *adx, *bdx, *aii, *bii, *aptr, *bptr; 2330 const MatScalar *va, *vb; 2331 PetscInt ma, na, mb, nb, i; 2332 2333 PetscFunctionBegin; 2334 PetscCall(MatGetSize(A, &ma, &na)); 2335 PetscCall(MatGetSize(B, &mb, &nb)); 2336 if (ma != nb || na != mb) { 2337 *f = PETSC_FALSE; 2338 PetscFunctionReturn(PETSC_SUCCESS); 2339 } 2340 PetscCall(MatSeqAIJGetArrayRead(A, &va)); 2341 PetscCall(MatSeqAIJGetArrayRead(B, &vb)); 2342 aii = aij->i; 2343 bii = bij->i; 2344 adx = aij->j; 2345 bdx = bij->j; 2346 PetscCall(PetscMalloc1(ma, &aptr)); 2347 PetscCall(PetscMalloc1(mb, &bptr)); 2348 for (i = 0; i < ma; i++) aptr[i] = aii[i]; 2349 for (i = 0; i < mb; i++) bptr[i] = bii[i]; 2350 2351 *f = PETSC_TRUE; 2352 for (i = 0; i < ma; i++) { 2353 while (aptr[i] < aii[i + 1]) { 2354 PetscInt idc, idr; 2355 PetscScalar vc, vr; 2356 /* column/row index/value */ 2357 idc = adx[aptr[i]]; 2358 idr = bdx[bptr[idc]]; 2359 vc = va[aptr[i]]; 2360 vr = vb[bptr[idc]]; 2361 if (i != idr || PetscAbsScalar(vc - vr) > tol) { 2362 *f = PETSC_FALSE; 2363 goto done; 2364 } else { 2365 aptr[i]++; 2366 if (B || i != idc) bptr[idc]++; 2367 } 2368 } 2369 } 2370 done: 2371 PetscCall(PetscFree(aptr)); 2372 PetscCall(PetscFree(bptr)); 2373 PetscCall(MatSeqAIJRestoreArrayRead(A, &va)); 2374 PetscCall(MatSeqAIJRestoreArrayRead(B, &vb)); 2375 PetscFunctionReturn(PETSC_SUCCESS); 2376 } 2377 2378 static PetscErrorCode MatIsHermitianTranspose_SeqAIJ(Mat A, Mat B, PetscReal tol, PetscBool *f) 2379 { 2380 Mat_SeqAIJ *aij = (Mat_SeqAIJ *)A->data, *bij = (Mat_SeqAIJ *)B->data; 2381 PetscInt *adx, *bdx, *aii, *bii, *aptr, *bptr; 2382 MatScalar *va, *vb; 2383 PetscInt ma, na, mb, nb, i; 2384 2385 PetscFunctionBegin; 2386 PetscCall(MatGetSize(A, &ma, &na)); 2387 PetscCall(MatGetSize(B, &mb, &nb)); 2388 if (ma != nb || na != mb) { 2389 *f = PETSC_FALSE; 2390 PetscFunctionReturn(PETSC_SUCCESS); 2391 } 2392 aii = aij->i; 2393 bii = bij->i; 2394 adx = aij->j; 2395 bdx = bij->j; 2396 va = aij->a; 2397 vb = bij->a; 2398 PetscCall(PetscMalloc1(ma, &aptr)); 2399 PetscCall(PetscMalloc1(mb, &bptr)); 2400 for (i = 0; i < ma; i++) aptr[i] = aii[i]; 2401 for (i = 0; i < mb; i++) bptr[i] = bii[i]; 2402 2403 *f = PETSC_TRUE; 2404 for (i = 0; i < ma; i++) { 2405 while (aptr[i] < aii[i + 1]) { 2406 PetscInt idc, idr; 2407 PetscScalar vc, vr; 2408 /* column/row index/value */ 2409 idc = adx[aptr[i]]; 2410 idr = bdx[bptr[idc]]; 2411 vc = va[aptr[i]]; 2412 vr = vb[bptr[idc]]; 2413 if (i != idr || PetscAbsScalar(vc - PetscConj(vr)) > tol) { 2414 *f = PETSC_FALSE; 2415 goto done; 2416 } else { 2417 aptr[i]++; 2418 if (B || i != idc) bptr[idc]++; 2419 } 2420 } 2421 } 2422 done: 2423 PetscCall(PetscFree(aptr)); 2424 PetscCall(PetscFree(bptr)); 2425 PetscFunctionReturn(PETSC_SUCCESS); 2426 } 2427 2428 static PetscErrorCode MatIsSymmetric_SeqAIJ(Mat A, PetscReal tol, PetscBool *f) 2429 { 2430 PetscFunctionBegin; 2431 PetscCall(MatIsTranspose_SeqAIJ(A, A, tol, f)); 2432 PetscFunctionReturn(PETSC_SUCCESS); 2433 } 2434 2435 static PetscErrorCode MatIsHermitian_SeqAIJ(Mat A, PetscReal tol, PetscBool *f) 2436 { 2437 PetscFunctionBegin; 2438 PetscCall(MatIsHermitianTranspose_SeqAIJ(A, A, tol, f)); 2439 PetscFunctionReturn(PETSC_SUCCESS); 2440 } 2441 2442 PetscErrorCode MatDiagonalScale_SeqAIJ(Mat A, Vec ll, Vec rr) 2443 { 2444 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 2445 const PetscScalar *l, *r; 2446 PetscScalar x; 2447 MatScalar *v; 2448 PetscInt i, j, m = A->rmap->n, n = A->cmap->n, M, nz = a->nz; 2449 const PetscInt *jj; 2450 2451 PetscFunctionBegin; 2452 if (ll) { 2453 /* The local size is used so that VecMPI can be passed to this routine 2454 by MatDiagonalScale_MPIAIJ */ 2455 PetscCall(VecGetLocalSize(ll, &m)); 2456 PetscCheck(m == A->rmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Left scaling vector wrong length"); 2457 PetscCall(VecGetArrayRead(ll, &l)); 2458 PetscCall(MatSeqAIJGetArray(A, &v)); 2459 for (i = 0; i < m; i++) { 2460 x = l[i]; 2461 M = a->i[i + 1] - a->i[i]; 2462 for (j = 0; j < M; j++) (*v++) *= x; 2463 } 2464 PetscCall(VecRestoreArrayRead(ll, &l)); 2465 PetscCall(PetscLogFlops(nz)); 2466 PetscCall(MatSeqAIJRestoreArray(A, &v)); 2467 } 2468 if (rr) { 2469 PetscCall(VecGetLocalSize(rr, &n)); 2470 PetscCheck(n == A->cmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Right scaling vector wrong length"); 2471 PetscCall(VecGetArrayRead(rr, &r)); 2472 PetscCall(MatSeqAIJGetArray(A, &v)); 2473 jj = a->j; 2474 for (i = 0; i < nz; i++) (*v++) *= r[*jj++]; 2475 PetscCall(MatSeqAIJRestoreArray(A, &v)); 2476 PetscCall(VecRestoreArrayRead(rr, &r)); 2477 PetscCall(PetscLogFlops(nz)); 2478 } 2479 PetscCall(MatSeqAIJInvalidateDiagonal(A)); 2480 PetscFunctionReturn(PETSC_SUCCESS); 2481 } 2482 2483 PetscErrorCode MatCreateSubMatrix_SeqAIJ(Mat A, IS isrow, IS iscol, PetscInt csize, MatReuse scall, Mat *B) 2484 { 2485 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data, *c; 2486 PetscInt *smap, i, k, kstart, kend, oldcols = A->cmap->n, *lens; 2487 PetscInt row, mat_i, *mat_j, tcol, first, step, *mat_ilen, sum, lensi; 2488 const PetscInt *irow, *icol; 2489 const PetscScalar *aa; 2490 PetscInt nrows, ncols; 2491 PetscInt *starts, *j_new, *i_new, *aj = a->j, *ai = a->i, ii, *ailen = a->ilen; 2492 MatScalar *a_new, *mat_a, *c_a; 2493 Mat C; 2494 PetscBool stride; 2495 2496 PetscFunctionBegin; 2497 PetscCall(ISGetIndices(isrow, &irow)); 2498 PetscCall(ISGetLocalSize(isrow, &nrows)); 2499 PetscCall(ISGetLocalSize(iscol, &ncols)); 2500 2501 PetscCall(PetscObjectTypeCompare((PetscObject)iscol, ISSTRIDE, &stride)); 2502 if (stride) { 2503 PetscCall(ISStrideGetInfo(iscol, &first, &step)); 2504 } else { 2505 first = 0; 2506 step = 0; 2507 } 2508 if (stride && step == 1) { 2509 /* special case of contiguous rows */ 2510 PetscCall(PetscMalloc2(nrows, &lens, nrows, &starts)); 2511 /* loop over new rows determining lens and starting points */ 2512 for (i = 0; i < nrows; i++) { 2513 kstart = ai[irow[i]]; 2514 kend = kstart + ailen[irow[i]]; 2515 starts[i] = kstart; 2516 for (k = kstart; k < kend; k++) { 2517 if (aj[k] >= first) { 2518 starts[i] = k; 2519 break; 2520 } 2521 } 2522 sum = 0; 2523 while (k < kend) { 2524 if (aj[k++] >= first + ncols) break; 2525 sum++; 2526 } 2527 lens[i] = sum; 2528 } 2529 /* create submatrix */ 2530 if (scall == MAT_REUSE_MATRIX) { 2531 PetscInt n_cols, n_rows; 2532 PetscCall(MatGetSize(*B, &n_rows, &n_cols)); 2533 PetscCheck(n_rows == nrows && n_cols == ncols, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Reused submatrix wrong size"); 2534 PetscCall(MatZeroEntries(*B)); 2535 C = *B; 2536 } else { 2537 PetscInt rbs, cbs; 2538 PetscCall(MatCreate(PetscObjectComm((PetscObject)A), &C)); 2539 PetscCall(MatSetSizes(C, nrows, ncols, PETSC_DETERMINE, PETSC_DETERMINE)); 2540 PetscCall(ISGetBlockSize(isrow, &rbs)); 2541 PetscCall(ISGetBlockSize(iscol, &cbs)); 2542 PetscCall(MatSetBlockSizes(C, rbs, cbs)); 2543 PetscCall(MatSetType(C, ((PetscObject)A)->type_name)); 2544 PetscCall(MatSeqAIJSetPreallocation_SeqAIJ(C, 0, lens)); 2545 } 2546 c = (Mat_SeqAIJ *)C->data; 2547 2548 /* loop over rows inserting into submatrix */ 2549 PetscCall(MatSeqAIJGetArrayWrite(C, &a_new)); // Not 'a_new = c->a-new', since that raw usage ignores offload state of C 2550 j_new = c->j; 2551 i_new = c->i; 2552 PetscCall(MatSeqAIJGetArrayRead(A, &aa)); 2553 for (i = 0; i < nrows; i++) { 2554 ii = starts[i]; 2555 lensi = lens[i]; 2556 for (k = 0; k < lensi; k++) *j_new++ = aj[ii + k] - first; 2557 PetscCall(PetscArraycpy(a_new, aa + starts[i], lensi)); 2558 a_new += lensi; 2559 i_new[i + 1] = i_new[i] + lensi; 2560 c->ilen[i] = lensi; 2561 } 2562 PetscCall(MatSeqAIJRestoreArrayWrite(C, &a_new)); // Set C's offload state properly 2563 PetscCall(MatSeqAIJRestoreArrayRead(A, &aa)); 2564 PetscCall(PetscFree2(lens, starts)); 2565 } else { 2566 PetscCall(ISGetIndices(iscol, &icol)); 2567 PetscCall(PetscCalloc1(oldcols, &smap)); 2568 PetscCall(PetscMalloc1(1 + nrows, &lens)); 2569 for (i = 0; i < ncols; i++) { 2570 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); 2571 smap[icol[i]] = i + 1; 2572 } 2573 2574 /* determine lens of each row */ 2575 for (i = 0; i < nrows; i++) { 2576 kstart = ai[irow[i]]; 2577 kend = kstart + a->ilen[irow[i]]; 2578 lens[i] = 0; 2579 for (k = kstart; k < kend; k++) { 2580 if (smap[aj[k]]) lens[i]++; 2581 } 2582 } 2583 /* Create and fill new matrix */ 2584 if (scall == MAT_REUSE_MATRIX) { 2585 PetscBool equal; 2586 2587 c = (Mat_SeqAIJ *)((*B)->data); 2588 PetscCheck((*B)->rmap->n == nrows && (*B)->cmap->n == ncols, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Cannot reuse matrix. wrong size"); 2589 PetscCall(PetscArraycmp(c->ilen, lens, (*B)->rmap->n, &equal)); 2590 PetscCheck(equal, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Cannot reuse matrix. wrong number of nonzeros"); 2591 PetscCall(PetscArrayzero(c->ilen, (*B)->rmap->n)); 2592 C = *B; 2593 } else { 2594 PetscInt rbs, cbs; 2595 PetscCall(MatCreate(PetscObjectComm((PetscObject)A), &C)); 2596 PetscCall(MatSetSizes(C, nrows, ncols, PETSC_DETERMINE, PETSC_DETERMINE)); 2597 PetscCall(ISGetBlockSize(isrow, &rbs)); 2598 PetscCall(ISGetBlockSize(iscol, &cbs)); 2599 if (rbs > 1 || cbs > 1) PetscCall(MatSetBlockSizes(C, rbs, cbs)); 2600 PetscCall(MatSetType(C, ((PetscObject)A)->type_name)); 2601 PetscCall(MatSeqAIJSetPreallocation_SeqAIJ(C, 0, lens)); 2602 } 2603 PetscCall(MatSeqAIJGetArrayRead(A, &aa)); 2604 2605 c = (Mat_SeqAIJ *)(C->data); 2606 PetscCall(MatSeqAIJGetArrayWrite(C, &c_a)); // Not 'c->a', since that raw usage ignores offload state of C 2607 for (i = 0; i < nrows; i++) { 2608 row = irow[i]; 2609 kstart = ai[row]; 2610 kend = kstart + a->ilen[row]; 2611 mat_i = c->i[i]; 2612 mat_j = c->j + mat_i; 2613 mat_a = c_a + mat_i; 2614 mat_ilen = c->ilen + i; 2615 for (k = kstart; k < kend; k++) { 2616 if ((tcol = smap[a->j[k]])) { 2617 *mat_j++ = tcol - 1; 2618 *mat_a++ = aa[k]; 2619 (*mat_ilen)++; 2620 } 2621 } 2622 } 2623 PetscCall(MatSeqAIJRestoreArrayRead(A, &aa)); 2624 /* Free work space */ 2625 PetscCall(ISRestoreIndices(iscol, &icol)); 2626 PetscCall(PetscFree(smap)); 2627 PetscCall(PetscFree(lens)); 2628 /* sort */ 2629 for (i = 0; i < nrows; i++) { 2630 PetscInt ilen; 2631 2632 mat_i = c->i[i]; 2633 mat_j = c->j + mat_i; 2634 mat_a = c_a + mat_i; 2635 ilen = c->ilen[i]; 2636 PetscCall(PetscSortIntWithScalarArray(ilen, mat_j, mat_a)); 2637 } 2638 PetscCall(MatSeqAIJRestoreArrayWrite(C, &c_a)); 2639 } 2640 #if defined(PETSC_HAVE_DEVICE) 2641 PetscCall(MatBindToCPU(C, A->boundtocpu)); 2642 #endif 2643 PetscCall(MatAssemblyBegin(C, MAT_FINAL_ASSEMBLY)); 2644 PetscCall(MatAssemblyEnd(C, MAT_FINAL_ASSEMBLY)); 2645 2646 PetscCall(ISRestoreIndices(isrow, &irow)); 2647 *B = C; 2648 PetscFunctionReturn(PETSC_SUCCESS); 2649 } 2650 2651 static PetscErrorCode MatGetMultiProcBlock_SeqAIJ(Mat mat, MPI_Comm subComm, MatReuse scall, Mat *subMat) 2652 { 2653 Mat B; 2654 2655 PetscFunctionBegin; 2656 if (scall == MAT_INITIAL_MATRIX) { 2657 PetscCall(MatCreate(subComm, &B)); 2658 PetscCall(MatSetSizes(B, mat->rmap->n, mat->cmap->n, mat->rmap->n, mat->cmap->n)); 2659 PetscCall(MatSetBlockSizesFromMats(B, mat, mat)); 2660 PetscCall(MatSetType(B, MATSEQAIJ)); 2661 PetscCall(MatDuplicateNoCreate_SeqAIJ(B, mat, MAT_COPY_VALUES, PETSC_TRUE)); 2662 *subMat = B; 2663 } else { 2664 PetscCall(MatCopy_SeqAIJ(mat, *subMat, SAME_NONZERO_PATTERN)); 2665 } 2666 PetscFunctionReturn(PETSC_SUCCESS); 2667 } 2668 2669 static PetscErrorCode MatILUFactor_SeqAIJ(Mat inA, IS row, IS col, const MatFactorInfo *info) 2670 { 2671 Mat_SeqAIJ *a = (Mat_SeqAIJ *)inA->data; 2672 Mat outA; 2673 PetscBool row_identity, col_identity; 2674 2675 PetscFunctionBegin; 2676 PetscCheck(info->levels == 0, PETSC_COMM_SELF, PETSC_ERR_SUP, "Only levels=0 supported for in-place ilu"); 2677 2678 PetscCall(ISIdentity(row, &row_identity)); 2679 PetscCall(ISIdentity(col, &col_identity)); 2680 2681 outA = inA; 2682 outA->factortype = MAT_FACTOR_LU; 2683 PetscCall(PetscFree(inA->solvertype)); 2684 PetscCall(PetscStrallocpy(MATSOLVERPETSC, &inA->solvertype)); 2685 2686 PetscCall(PetscObjectReference((PetscObject)row)); 2687 PetscCall(ISDestroy(&a->row)); 2688 2689 a->row = row; 2690 2691 PetscCall(PetscObjectReference((PetscObject)col)); 2692 PetscCall(ISDestroy(&a->col)); 2693 2694 a->col = col; 2695 2696 /* Create the inverse permutation so that it can be used in MatLUFactorNumeric() */ 2697 PetscCall(ISDestroy(&a->icol)); 2698 PetscCall(ISInvertPermutation(col, PETSC_DECIDE, &a->icol)); 2699 2700 if (!a->solve_work) { /* this matrix may have been factored before */ 2701 PetscCall(PetscMalloc1(inA->rmap->n + 1, &a->solve_work)); 2702 } 2703 2704 PetscCall(MatMarkDiagonal_SeqAIJ(inA)); 2705 if (row_identity && col_identity) { 2706 PetscCall(MatLUFactorNumeric_SeqAIJ_inplace(outA, inA, info)); 2707 } else { 2708 PetscCall(MatLUFactorNumeric_SeqAIJ_InplaceWithPerm(outA, inA, info)); 2709 } 2710 PetscFunctionReturn(PETSC_SUCCESS); 2711 } 2712 2713 PetscErrorCode MatScale_SeqAIJ(Mat inA, PetscScalar alpha) 2714 { 2715 Mat_SeqAIJ *a = (Mat_SeqAIJ *)inA->data; 2716 PetscScalar *v; 2717 PetscBLASInt one = 1, bnz; 2718 2719 PetscFunctionBegin; 2720 PetscCall(MatSeqAIJGetArray(inA, &v)); 2721 PetscCall(PetscBLASIntCast(a->nz, &bnz)); 2722 PetscCallBLAS("BLASscal", BLASscal_(&bnz, &alpha, v, &one)); 2723 PetscCall(PetscLogFlops(a->nz)); 2724 PetscCall(MatSeqAIJRestoreArray(inA, &v)); 2725 PetscCall(MatSeqAIJInvalidateDiagonal(inA)); 2726 PetscFunctionReturn(PETSC_SUCCESS); 2727 } 2728 2729 PetscErrorCode MatDestroySubMatrix_Private(Mat_SubSppt *submatj) 2730 { 2731 PetscInt i; 2732 2733 PetscFunctionBegin; 2734 if (!submatj->id) { /* delete data that are linked only to submats[id=0] */ 2735 PetscCall(PetscFree4(submatj->sbuf1, submatj->ptr, submatj->tmp, submatj->ctr)); 2736 2737 for (i = 0; i < submatj->nrqr; ++i) PetscCall(PetscFree(submatj->sbuf2[i])); 2738 PetscCall(PetscFree3(submatj->sbuf2, submatj->req_size, submatj->req_source1)); 2739 2740 if (submatj->rbuf1) { 2741 PetscCall(PetscFree(submatj->rbuf1[0])); 2742 PetscCall(PetscFree(submatj->rbuf1)); 2743 } 2744 2745 for (i = 0; i < submatj->nrqs; ++i) PetscCall(PetscFree(submatj->rbuf3[i])); 2746 PetscCall(PetscFree3(submatj->req_source2, submatj->rbuf2, submatj->rbuf3)); 2747 PetscCall(PetscFree(submatj->pa)); 2748 } 2749 2750 #if defined(PETSC_USE_CTABLE) 2751 PetscCall(PetscHMapIDestroy(&submatj->rmap)); 2752 if (submatj->cmap_loc) PetscCall(PetscFree(submatj->cmap_loc)); 2753 PetscCall(PetscFree(submatj->rmap_loc)); 2754 #else 2755 PetscCall(PetscFree(submatj->rmap)); 2756 #endif 2757 2758 if (!submatj->allcolumns) { 2759 #if defined(PETSC_USE_CTABLE) 2760 PetscCall(PetscHMapIDestroy((PetscHMapI *)&submatj->cmap)); 2761 #else 2762 PetscCall(PetscFree(submatj->cmap)); 2763 #endif 2764 } 2765 PetscCall(PetscFree(submatj->row2proc)); 2766 2767 PetscCall(PetscFree(submatj)); 2768 PetscFunctionReturn(PETSC_SUCCESS); 2769 } 2770 2771 PetscErrorCode MatDestroySubMatrix_SeqAIJ(Mat C) 2772 { 2773 Mat_SeqAIJ *c = (Mat_SeqAIJ *)C->data; 2774 Mat_SubSppt *submatj = c->submatis1; 2775 2776 PetscFunctionBegin; 2777 PetscCall((*submatj->destroy)(C)); 2778 PetscCall(MatDestroySubMatrix_Private(submatj)); 2779 PetscFunctionReturn(PETSC_SUCCESS); 2780 } 2781 2782 /* Note this has code duplication with MatDestroySubMatrices_SeqBAIJ() */ 2783 static PetscErrorCode MatDestroySubMatrices_SeqAIJ(PetscInt n, Mat *mat[]) 2784 { 2785 PetscInt i; 2786 Mat C; 2787 Mat_SeqAIJ *c; 2788 Mat_SubSppt *submatj; 2789 2790 PetscFunctionBegin; 2791 for (i = 0; i < n; i++) { 2792 C = (*mat)[i]; 2793 c = (Mat_SeqAIJ *)C->data; 2794 submatj = c->submatis1; 2795 if (submatj) { 2796 if (--((PetscObject)C)->refct <= 0) { 2797 PetscCall(PetscFree(C->factorprefix)); 2798 PetscCall((*submatj->destroy)(C)); 2799 PetscCall(MatDestroySubMatrix_Private(submatj)); 2800 PetscCall(PetscFree(C->defaultvectype)); 2801 PetscCall(PetscFree(C->defaultrandtype)); 2802 PetscCall(PetscLayoutDestroy(&C->rmap)); 2803 PetscCall(PetscLayoutDestroy(&C->cmap)); 2804 PetscCall(PetscHeaderDestroy(&C)); 2805 } 2806 } else { 2807 PetscCall(MatDestroy(&C)); 2808 } 2809 } 2810 2811 /* Destroy Dummy submatrices created for reuse */ 2812 PetscCall(MatDestroySubMatrices_Dummy(n, mat)); 2813 2814 PetscCall(PetscFree(*mat)); 2815 PetscFunctionReturn(PETSC_SUCCESS); 2816 } 2817 2818 static PetscErrorCode MatCreateSubMatrices_SeqAIJ(Mat A, PetscInt n, const IS irow[], const IS icol[], MatReuse scall, Mat *B[]) 2819 { 2820 PetscInt i; 2821 2822 PetscFunctionBegin; 2823 if (scall == MAT_INITIAL_MATRIX) PetscCall(PetscCalloc1(n + 1, B)); 2824 2825 for (i = 0; i < n; i++) PetscCall(MatCreateSubMatrix_SeqAIJ(A, irow[i], icol[i], PETSC_DECIDE, scall, &(*B)[i])); 2826 PetscFunctionReturn(PETSC_SUCCESS); 2827 } 2828 2829 static PetscErrorCode MatIncreaseOverlap_SeqAIJ(Mat A, PetscInt is_max, IS is[], PetscInt ov) 2830 { 2831 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 2832 PetscInt row, i, j, k, l, ll, m, n, *nidx, isz, val; 2833 const PetscInt *idx; 2834 PetscInt start, end, *ai, *aj, bs = (A->rmap->bs > 0 && A->rmap->bs == A->cmap->bs) ? A->rmap->bs : 1; 2835 PetscBT table; 2836 2837 PetscFunctionBegin; 2838 m = A->rmap->n / bs; 2839 ai = a->i; 2840 aj = a->j; 2841 2842 PetscCheck(ov >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "illegal negative overlap value used"); 2843 2844 PetscCall(PetscMalloc1(m + 1, &nidx)); 2845 PetscCall(PetscBTCreate(m, &table)); 2846 2847 for (i = 0; i < is_max; i++) { 2848 /* Initialize the two local arrays */ 2849 isz = 0; 2850 PetscCall(PetscBTMemzero(m, table)); 2851 2852 /* Extract the indices, assume there can be duplicate entries */ 2853 PetscCall(ISGetIndices(is[i], &idx)); 2854 PetscCall(ISGetLocalSize(is[i], &n)); 2855 2856 if (bs > 1) { 2857 /* Enter these into the temp arrays. I.e., mark table[row], enter row into new index */ 2858 for (j = 0; j < n; ++j) { 2859 if (!PetscBTLookupSet(table, idx[j] / bs)) nidx[isz++] = idx[j] / bs; 2860 } 2861 PetscCall(ISRestoreIndices(is[i], &idx)); 2862 PetscCall(ISDestroy(&is[i])); 2863 2864 k = 0; 2865 for (j = 0; j < ov; j++) { /* for each overlap */ 2866 n = isz; 2867 for (; k < n; k++) { /* do only those rows in nidx[k], which are not done yet */ 2868 for (ll = 0; ll < bs; ll++) { 2869 row = bs * nidx[k] + ll; 2870 start = ai[row]; 2871 end = ai[row + 1]; 2872 for (l = start; l < end; l++) { 2873 val = aj[l] / bs; 2874 if (!PetscBTLookupSet(table, val)) nidx[isz++] = val; 2875 } 2876 } 2877 } 2878 } 2879 PetscCall(ISCreateBlock(PETSC_COMM_SELF, bs, isz, nidx, PETSC_COPY_VALUES, (is + i))); 2880 } else { 2881 /* Enter these into the temp arrays. I.e., mark table[row], enter row into new index */ 2882 for (j = 0; j < n; ++j) { 2883 if (!PetscBTLookupSet(table, idx[j])) nidx[isz++] = idx[j]; 2884 } 2885 PetscCall(ISRestoreIndices(is[i], &idx)); 2886 PetscCall(ISDestroy(&is[i])); 2887 2888 k = 0; 2889 for (j = 0; j < ov; j++) { /* for each overlap */ 2890 n = isz; 2891 for (; k < n; k++) { /* do only those rows in nidx[k], which are not done yet */ 2892 row = nidx[k]; 2893 start = ai[row]; 2894 end = ai[row + 1]; 2895 for (l = start; l < end; l++) { 2896 val = aj[l]; 2897 if (!PetscBTLookupSet(table, val)) nidx[isz++] = val; 2898 } 2899 } 2900 } 2901 PetscCall(ISCreateGeneral(PETSC_COMM_SELF, isz, nidx, PETSC_COPY_VALUES, (is + i))); 2902 } 2903 } 2904 PetscCall(PetscBTDestroy(&table)); 2905 PetscCall(PetscFree(nidx)); 2906 PetscFunctionReturn(PETSC_SUCCESS); 2907 } 2908 2909 static PetscErrorCode MatPermute_SeqAIJ(Mat A, IS rowp, IS colp, Mat *B) 2910 { 2911 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 2912 PetscInt i, nz = 0, m = A->rmap->n, n = A->cmap->n; 2913 const PetscInt *row, *col; 2914 PetscInt *cnew, j, *lens; 2915 IS icolp, irowp; 2916 PetscInt *cwork = NULL; 2917 PetscScalar *vwork = NULL; 2918 2919 PetscFunctionBegin; 2920 PetscCall(ISInvertPermutation(rowp, PETSC_DECIDE, &irowp)); 2921 PetscCall(ISGetIndices(irowp, &row)); 2922 PetscCall(ISInvertPermutation(colp, PETSC_DECIDE, &icolp)); 2923 PetscCall(ISGetIndices(icolp, &col)); 2924 2925 /* determine lengths of permuted rows */ 2926 PetscCall(PetscMalloc1(m + 1, &lens)); 2927 for (i = 0; i < m; i++) lens[row[i]] = a->i[i + 1] - a->i[i]; 2928 PetscCall(MatCreate(PetscObjectComm((PetscObject)A), B)); 2929 PetscCall(MatSetSizes(*B, m, n, m, n)); 2930 PetscCall(MatSetBlockSizesFromMats(*B, A, A)); 2931 PetscCall(MatSetType(*B, ((PetscObject)A)->type_name)); 2932 PetscCall(MatSeqAIJSetPreallocation_SeqAIJ(*B, 0, lens)); 2933 PetscCall(PetscFree(lens)); 2934 2935 PetscCall(PetscMalloc1(n, &cnew)); 2936 for (i = 0; i < m; i++) { 2937 PetscCall(MatGetRow_SeqAIJ(A, i, &nz, &cwork, &vwork)); 2938 for (j = 0; j < nz; j++) cnew[j] = col[cwork[j]]; 2939 PetscCall(MatSetValues_SeqAIJ(*B, 1, &row[i], nz, cnew, vwork, INSERT_VALUES)); 2940 PetscCall(MatRestoreRow_SeqAIJ(A, i, &nz, &cwork, &vwork)); 2941 } 2942 PetscCall(PetscFree(cnew)); 2943 2944 (*B)->assembled = PETSC_FALSE; 2945 2946 #if defined(PETSC_HAVE_DEVICE) 2947 PetscCall(MatBindToCPU(*B, A->boundtocpu)); 2948 #endif 2949 PetscCall(MatAssemblyBegin(*B, MAT_FINAL_ASSEMBLY)); 2950 PetscCall(MatAssemblyEnd(*B, MAT_FINAL_ASSEMBLY)); 2951 PetscCall(ISRestoreIndices(irowp, &row)); 2952 PetscCall(ISRestoreIndices(icolp, &col)); 2953 PetscCall(ISDestroy(&irowp)); 2954 PetscCall(ISDestroy(&icolp)); 2955 if (rowp == colp) PetscCall(MatPropagateSymmetryOptions(A, *B)); 2956 PetscFunctionReturn(PETSC_SUCCESS); 2957 } 2958 2959 PetscErrorCode MatCopy_SeqAIJ(Mat A, Mat B, MatStructure str) 2960 { 2961 PetscFunctionBegin; 2962 /* If the two matrices have the same copy implementation, use fast copy. */ 2963 if (str == SAME_NONZERO_PATTERN && (A->ops->copy == B->ops->copy)) { 2964 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 2965 Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data; 2966 const PetscScalar *aa; 2967 2968 PetscCall(MatSeqAIJGetArrayRead(A, &aa)); 2969 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]); 2970 PetscCall(PetscArraycpy(b->a, aa, a->i[A->rmap->n])); 2971 PetscCall(PetscObjectStateIncrease((PetscObject)B)); 2972 PetscCall(MatSeqAIJRestoreArrayRead(A, &aa)); 2973 } else { 2974 PetscCall(MatCopy_Basic(A, B, str)); 2975 } 2976 PetscFunctionReturn(PETSC_SUCCESS); 2977 } 2978 2979 PETSC_INTERN PetscErrorCode MatSeqAIJGetArray_SeqAIJ(Mat A, PetscScalar *array[]) 2980 { 2981 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 2982 2983 PetscFunctionBegin; 2984 *array = a->a; 2985 PetscFunctionReturn(PETSC_SUCCESS); 2986 } 2987 2988 PETSC_INTERN PetscErrorCode MatSeqAIJRestoreArray_SeqAIJ(Mat A, PetscScalar *array[]) 2989 { 2990 PetscFunctionBegin; 2991 *array = NULL; 2992 PetscFunctionReturn(PETSC_SUCCESS); 2993 } 2994 2995 /* 2996 Computes the number of nonzeros per row needed for preallocation when X and Y 2997 have different nonzero structure. 2998 */ 2999 PetscErrorCode MatAXPYGetPreallocation_SeqX_private(PetscInt m, const PetscInt *xi, const PetscInt *xj, const PetscInt *yi, const PetscInt *yj, PetscInt *nnz) 3000 { 3001 PetscInt i, j, k, nzx, nzy; 3002 3003 PetscFunctionBegin; 3004 /* Set the number of nonzeros in the new matrix */ 3005 for (i = 0; i < m; i++) { 3006 const PetscInt *xjj = xj + xi[i], *yjj = yj + yi[i]; 3007 nzx = xi[i + 1] - xi[i]; 3008 nzy = yi[i + 1] - yi[i]; 3009 nnz[i] = 0; 3010 for (j = 0, k = 0; j < nzx; j++) { /* Point in X */ 3011 for (; k < nzy && yjj[k] < xjj[j]; k++) nnz[i]++; /* Catch up to X */ 3012 if (k < nzy && yjj[k] == xjj[j]) k++; /* Skip duplicate */ 3013 nnz[i]++; 3014 } 3015 for (; k < nzy; k++) nnz[i]++; 3016 } 3017 PetscFunctionReturn(PETSC_SUCCESS); 3018 } 3019 3020 PetscErrorCode MatAXPYGetPreallocation_SeqAIJ(Mat Y, Mat X, PetscInt *nnz) 3021 { 3022 PetscInt m = Y->rmap->N; 3023 Mat_SeqAIJ *x = (Mat_SeqAIJ *)X->data; 3024 Mat_SeqAIJ *y = (Mat_SeqAIJ *)Y->data; 3025 3026 PetscFunctionBegin; 3027 /* Set the number of nonzeros in the new matrix */ 3028 PetscCall(MatAXPYGetPreallocation_SeqX_private(m, x->i, x->j, y->i, y->j, nnz)); 3029 PetscFunctionReturn(PETSC_SUCCESS); 3030 } 3031 3032 PetscErrorCode MatAXPY_SeqAIJ(Mat Y, PetscScalar a, Mat X, MatStructure str) 3033 { 3034 Mat_SeqAIJ *x = (Mat_SeqAIJ *)X->data, *y = (Mat_SeqAIJ *)Y->data; 3035 3036 PetscFunctionBegin; 3037 if (str == UNKNOWN_NONZERO_PATTERN || (PetscDefined(USE_DEBUG) && str == SAME_NONZERO_PATTERN)) { 3038 PetscBool e = x->nz == y->nz ? PETSC_TRUE : PETSC_FALSE; 3039 if (e) { 3040 PetscCall(PetscArraycmp(x->i, y->i, Y->rmap->n + 1, &e)); 3041 if (e) { 3042 PetscCall(PetscArraycmp(x->j, y->j, y->nz, &e)); 3043 if (e) str = SAME_NONZERO_PATTERN; 3044 } 3045 } 3046 if (!e) PetscCheck(str != SAME_NONZERO_PATTERN, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "MatStructure is not SAME_NONZERO_PATTERN"); 3047 } 3048 if (str == SAME_NONZERO_PATTERN) { 3049 const PetscScalar *xa; 3050 PetscScalar *ya, alpha = a; 3051 PetscBLASInt one = 1, bnz; 3052 3053 PetscCall(PetscBLASIntCast(x->nz, &bnz)); 3054 PetscCall(MatSeqAIJGetArray(Y, &ya)); 3055 PetscCall(MatSeqAIJGetArrayRead(X, &xa)); 3056 PetscCallBLAS("BLASaxpy", BLASaxpy_(&bnz, &alpha, xa, &one, ya, &one)); 3057 PetscCall(MatSeqAIJRestoreArrayRead(X, &xa)); 3058 PetscCall(MatSeqAIJRestoreArray(Y, &ya)); 3059 PetscCall(PetscLogFlops(2.0 * bnz)); 3060 PetscCall(MatSeqAIJInvalidateDiagonal(Y)); 3061 PetscCall(PetscObjectStateIncrease((PetscObject)Y)); 3062 } else if (str == SUBSET_NONZERO_PATTERN) { /* nonzeros of X is a subset of Y's */ 3063 PetscCall(MatAXPY_Basic(Y, a, X, str)); 3064 } else { 3065 Mat B; 3066 PetscInt *nnz; 3067 PetscCall(PetscMalloc1(Y->rmap->N, &nnz)); 3068 PetscCall(MatCreate(PetscObjectComm((PetscObject)Y), &B)); 3069 PetscCall(PetscObjectSetName((PetscObject)B, ((PetscObject)Y)->name)); 3070 PetscCall(MatSetLayouts(B, Y->rmap, Y->cmap)); 3071 PetscCall(MatSetType(B, ((PetscObject)Y)->type_name)); 3072 PetscCall(MatAXPYGetPreallocation_SeqAIJ(Y, X, nnz)); 3073 PetscCall(MatSeqAIJSetPreallocation(B, 0, nnz)); 3074 PetscCall(MatAXPY_BasicWithPreallocation(B, Y, a, X, str)); 3075 PetscCall(MatHeaderMerge(Y, &B)); 3076 PetscCall(MatSeqAIJCheckInode(Y)); 3077 PetscCall(PetscFree(nnz)); 3078 } 3079 PetscFunctionReturn(PETSC_SUCCESS); 3080 } 3081 3082 PETSC_INTERN PetscErrorCode MatConjugate_SeqAIJ(Mat mat) 3083 { 3084 #if defined(PETSC_USE_COMPLEX) 3085 Mat_SeqAIJ *aij = (Mat_SeqAIJ *)mat->data; 3086 PetscInt i, nz; 3087 PetscScalar *a; 3088 3089 PetscFunctionBegin; 3090 nz = aij->nz; 3091 PetscCall(MatSeqAIJGetArray(mat, &a)); 3092 for (i = 0; i < nz; i++) a[i] = PetscConj(a[i]); 3093 PetscCall(MatSeqAIJRestoreArray(mat, &a)); 3094 #else 3095 PetscFunctionBegin; 3096 #endif 3097 PetscFunctionReturn(PETSC_SUCCESS); 3098 } 3099 3100 static PetscErrorCode MatGetRowMaxAbs_SeqAIJ(Mat A, Vec v, PetscInt idx[]) 3101 { 3102 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 3103 PetscInt i, j, m = A->rmap->n, *ai, *aj, ncols, n; 3104 PetscReal atmp; 3105 PetscScalar *x; 3106 const MatScalar *aa, *av; 3107 3108 PetscFunctionBegin; 3109 PetscCheck(!A->factortype, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Not for factored matrix"); 3110 PetscCall(MatSeqAIJGetArrayRead(A, &av)); 3111 aa = av; 3112 ai = a->i; 3113 aj = a->j; 3114 3115 PetscCall(VecSet(v, 0.0)); 3116 PetscCall(VecGetArrayWrite(v, &x)); 3117 PetscCall(VecGetLocalSize(v, &n)); 3118 PetscCheck(n == A->rmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Nonconforming matrix and vector"); 3119 for (i = 0; i < m; i++) { 3120 ncols = ai[1] - ai[0]; 3121 ai++; 3122 for (j = 0; j < ncols; j++) { 3123 atmp = PetscAbsScalar(*aa); 3124 if (PetscAbsScalar(x[i]) < atmp) { 3125 x[i] = atmp; 3126 if (idx) idx[i] = *aj; 3127 } 3128 aa++; 3129 aj++; 3130 } 3131 } 3132 PetscCall(VecRestoreArrayWrite(v, &x)); 3133 PetscCall(MatSeqAIJRestoreArrayRead(A, &av)); 3134 PetscFunctionReturn(PETSC_SUCCESS); 3135 } 3136 3137 static PetscErrorCode MatGetRowMax_SeqAIJ(Mat A, Vec v, PetscInt idx[]) 3138 { 3139 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 3140 PetscInt i, j, m = A->rmap->n, *ai, *aj, ncols, n; 3141 PetscScalar *x; 3142 const MatScalar *aa, *av; 3143 3144 PetscFunctionBegin; 3145 PetscCheck(!A->factortype, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Not for factored matrix"); 3146 PetscCall(MatSeqAIJGetArrayRead(A, &av)); 3147 aa = av; 3148 ai = a->i; 3149 aj = a->j; 3150 3151 PetscCall(VecSet(v, 0.0)); 3152 PetscCall(VecGetArrayWrite(v, &x)); 3153 PetscCall(VecGetLocalSize(v, &n)); 3154 PetscCheck(n == A->rmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Nonconforming matrix and vector"); 3155 for (i = 0; i < m; i++) { 3156 ncols = ai[1] - ai[0]; 3157 ai++; 3158 if (ncols == A->cmap->n) { /* row is dense */ 3159 x[i] = *aa; 3160 if (idx) idx[i] = 0; 3161 } else { /* row is sparse so already KNOW maximum is 0.0 or higher */ 3162 x[i] = 0.0; 3163 if (idx) { 3164 for (j = 0; j < ncols; j++) { /* find first implicit 0.0 in the row */ 3165 if (aj[j] > j) { 3166 idx[i] = j; 3167 break; 3168 } 3169 } 3170 /* in case first implicit 0.0 in the row occurs at ncols-th column */ 3171 if (j == ncols && j < A->cmap->n) idx[i] = j; 3172 } 3173 } 3174 for (j = 0; j < ncols; j++) { 3175 if (PetscRealPart(x[i]) < PetscRealPart(*aa)) { 3176 x[i] = *aa; 3177 if (idx) idx[i] = *aj; 3178 } 3179 aa++; 3180 aj++; 3181 } 3182 } 3183 PetscCall(VecRestoreArrayWrite(v, &x)); 3184 PetscCall(MatSeqAIJRestoreArrayRead(A, &av)); 3185 PetscFunctionReturn(PETSC_SUCCESS); 3186 } 3187 3188 static PetscErrorCode MatGetRowMinAbs_SeqAIJ(Mat A, Vec v, PetscInt idx[]) 3189 { 3190 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 3191 PetscInt i, j, m = A->rmap->n, *ai, *aj, ncols, n; 3192 PetscScalar *x; 3193 const MatScalar *aa, *av; 3194 3195 PetscFunctionBegin; 3196 PetscCall(MatSeqAIJGetArrayRead(A, &av)); 3197 aa = av; 3198 ai = a->i; 3199 aj = a->j; 3200 3201 PetscCall(VecSet(v, 0.0)); 3202 PetscCall(VecGetArrayWrite(v, &x)); 3203 PetscCall(VecGetLocalSize(v, &n)); 3204 PetscCheck(n == m, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Nonconforming matrix and vector, %" PetscInt_FMT " vs. %" PetscInt_FMT " rows", m, n); 3205 for (i = 0; i < m; i++) { 3206 ncols = ai[1] - ai[0]; 3207 ai++; 3208 if (ncols == A->cmap->n) { /* row is dense */ 3209 x[i] = *aa; 3210 if (idx) idx[i] = 0; 3211 } else { /* row is sparse so already KNOW minimum is 0.0 or higher */ 3212 x[i] = 0.0; 3213 if (idx) { /* find first implicit 0.0 in the row */ 3214 for (j = 0; j < ncols; j++) { 3215 if (aj[j] > j) { 3216 idx[i] = j; 3217 break; 3218 } 3219 } 3220 /* in case first implicit 0.0 in the row occurs at ncols-th column */ 3221 if (j == ncols && j < A->cmap->n) idx[i] = j; 3222 } 3223 } 3224 for (j = 0; j < ncols; j++) { 3225 if (PetscAbsScalar(x[i]) > PetscAbsScalar(*aa)) { 3226 x[i] = *aa; 3227 if (idx) idx[i] = *aj; 3228 } 3229 aa++; 3230 aj++; 3231 } 3232 } 3233 PetscCall(VecRestoreArrayWrite(v, &x)); 3234 PetscCall(MatSeqAIJRestoreArrayRead(A, &av)); 3235 PetscFunctionReturn(PETSC_SUCCESS); 3236 } 3237 3238 static PetscErrorCode MatGetRowMin_SeqAIJ(Mat A, Vec v, PetscInt idx[]) 3239 { 3240 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 3241 PetscInt i, j, m = A->rmap->n, ncols, n; 3242 const PetscInt *ai, *aj; 3243 PetscScalar *x; 3244 const MatScalar *aa, *av; 3245 3246 PetscFunctionBegin; 3247 PetscCheck(!A->factortype, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Not for factored matrix"); 3248 PetscCall(MatSeqAIJGetArrayRead(A, &av)); 3249 aa = av; 3250 ai = a->i; 3251 aj = a->j; 3252 3253 PetscCall(VecSet(v, 0.0)); 3254 PetscCall(VecGetArrayWrite(v, &x)); 3255 PetscCall(VecGetLocalSize(v, &n)); 3256 PetscCheck(n == m, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Nonconforming matrix and vector"); 3257 for (i = 0; i < m; i++) { 3258 ncols = ai[1] - ai[0]; 3259 ai++; 3260 if (ncols == A->cmap->n) { /* row is dense */ 3261 x[i] = *aa; 3262 if (idx) idx[i] = 0; 3263 } else { /* row is sparse so already KNOW minimum is 0.0 or lower */ 3264 x[i] = 0.0; 3265 if (idx) { /* find first implicit 0.0 in the row */ 3266 for (j = 0; j < ncols; j++) { 3267 if (aj[j] > j) { 3268 idx[i] = j; 3269 break; 3270 } 3271 } 3272 /* in case first implicit 0.0 in the row occurs at ncols-th column */ 3273 if (j == ncols && j < A->cmap->n) idx[i] = j; 3274 } 3275 } 3276 for (j = 0; j < ncols; j++) { 3277 if (PetscRealPart(x[i]) > PetscRealPart(*aa)) { 3278 x[i] = *aa; 3279 if (idx) idx[i] = *aj; 3280 } 3281 aa++; 3282 aj++; 3283 } 3284 } 3285 PetscCall(VecRestoreArrayWrite(v, &x)); 3286 PetscCall(MatSeqAIJRestoreArrayRead(A, &av)); 3287 PetscFunctionReturn(PETSC_SUCCESS); 3288 } 3289 3290 static PetscErrorCode MatInvertBlockDiagonal_SeqAIJ(Mat A, const PetscScalar **values) 3291 { 3292 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 3293 PetscInt i, bs = PetscAbs(A->rmap->bs), mbs = A->rmap->n / bs, ipvt[5], bs2 = bs * bs, *v_pivots, ij[7], *IJ, j; 3294 MatScalar *diag, work[25], *v_work; 3295 const PetscReal shift = 0.0; 3296 PetscBool allowzeropivot, zeropivotdetected = PETSC_FALSE; 3297 3298 PetscFunctionBegin; 3299 allowzeropivot = PetscNot(A->erroriffailure); 3300 if (a->ibdiagvalid) { 3301 if (values) *values = a->ibdiag; 3302 PetscFunctionReturn(PETSC_SUCCESS); 3303 } 3304 PetscCall(MatMarkDiagonal_SeqAIJ(A)); 3305 if (!a->ibdiag) { PetscCall(PetscMalloc1(bs2 * mbs, &a->ibdiag)); } 3306 diag = a->ibdiag; 3307 if (values) *values = a->ibdiag; 3308 /* factor and invert each block */ 3309 switch (bs) { 3310 case 1: 3311 for (i = 0; i < mbs; i++) { 3312 PetscCall(MatGetValues(A, 1, &i, 1, &i, diag + i)); 3313 if (PetscAbsScalar(diag[i] + shift) < PETSC_MACHINE_EPSILON) { 3314 if (allowzeropivot) { 3315 A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT; 3316 A->factorerror_zeropivot_value = PetscAbsScalar(diag[i]); 3317 A->factorerror_zeropivot_row = i; 3318 PetscCall(PetscInfo(A, "Zero pivot, row %" PetscInt_FMT " pivot %g tolerance %g\n", i, (double)PetscAbsScalar(diag[i]), (double)PETSC_MACHINE_EPSILON)); 3319 } 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); 3320 } 3321 diag[i] = (PetscScalar)1.0 / (diag[i] + shift); 3322 } 3323 break; 3324 case 2: 3325 for (i = 0; i < mbs; i++) { 3326 ij[0] = 2 * i; 3327 ij[1] = 2 * i + 1; 3328 PetscCall(MatGetValues(A, 2, ij, 2, ij, diag)); 3329 PetscCall(PetscKernel_A_gets_inverse_A_2(diag, shift, allowzeropivot, &zeropivotdetected)); 3330 if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT; 3331 PetscCall(PetscKernel_A_gets_transpose_A_2(diag)); 3332 diag += 4; 3333 } 3334 break; 3335 case 3: 3336 for (i = 0; i < mbs; i++) { 3337 ij[0] = 3 * i; 3338 ij[1] = 3 * i + 1; 3339 ij[2] = 3 * i + 2; 3340 PetscCall(MatGetValues(A, 3, ij, 3, ij, diag)); 3341 PetscCall(PetscKernel_A_gets_inverse_A_3(diag, shift, allowzeropivot, &zeropivotdetected)); 3342 if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT; 3343 PetscCall(PetscKernel_A_gets_transpose_A_3(diag)); 3344 diag += 9; 3345 } 3346 break; 3347 case 4: 3348 for (i = 0; i < mbs; i++) { 3349 ij[0] = 4 * i; 3350 ij[1] = 4 * i + 1; 3351 ij[2] = 4 * i + 2; 3352 ij[3] = 4 * i + 3; 3353 PetscCall(MatGetValues(A, 4, ij, 4, ij, diag)); 3354 PetscCall(PetscKernel_A_gets_inverse_A_4(diag, shift, allowzeropivot, &zeropivotdetected)); 3355 if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT; 3356 PetscCall(PetscKernel_A_gets_transpose_A_4(diag)); 3357 diag += 16; 3358 } 3359 break; 3360 case 5: 3361 for (i = 0; i < mbs; i++) { 3362 ij[0] = 5 * i; 3363 ij[1] = 5 * i + 1; 3364 ij[2] = 5 * i + 2; 3365 ij[3] = 5 * i + 3; 3366 ij[4] = 5 * i + 4; 3367 PetscCall(MatGetValues(A, 5, ij, 5, ij, diag)); 3368 PetscCall(PetscKernel_A_gets_inverse_A_5(diag, ipvt, work, shift, allowzeropivot, &zeropivotdetected)); 3369 if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT; 3370 PetscCall(PetscKernel_A_gets_transpose_A_5(diag)); 3371 diag += 25; 3372 } 3373 break; 3374 case 6: 3375 for (i = 0; i < mbs; i++) { 3376 ij[0] = 6 * i; 3377 ij[1] = 6 * i + 1; 3378 ij[2] = 6 * i + 2; 3379 ij[3] = 6 * i + 3; 3380 ij[4] = 6 * i + 4; 3381 ij[5] = 6 * i + 5; 3382 PetscCall(MatGetValues(A, 6, ij, 6, ij, diag)); 3383 PetscCall(PetscKernel_A_gets_inverse_A_6(diag, shift, allowzeropivot, &zeropivotdetected)); 3384 if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT; 3385 PetscCall(PetscKernel_A_gets_transpose_A_6(diag)); 3386 diag += 36; 3387 } 3388 break; 3389 case 7: 3390 for (i = 0; i < mbs; i++) { 3391 ij[0] = 7 * i; 3392 ij[1] = 7 * i + 1; 3393 ij[2] = 7 * i + 2; 3394 ij[3] = 7 * i + 3; 3395 ij[4] = 7 * i + 4; 3396 ij[5] = 7 * i + 5; 3397 ij[6] = 7 * i + 6; 3398 PetscCall(MatGetValues(A, 7, ij, 7, ij, diag)); 3399 PetscCall(PetscKernel_A_gets_inverse_A_7(diag, shift, allowzeropivot, &zeropivotdetected)); 3400 if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT; 3401 PetscCall(PetscKernel_A_gets_transpose_A_7(diag)); 3402 diag += 49; 3403 } 3404 break; 3405 default: 3406 PetscCall(PetscMalloc3(bs, &v_work, bs, &v_pivots, bs, &IJ)); 3407 for (i = 0; i < mbs; i++) { 3408 for (j = 0; j < bs; j++) IJ[j] = bs * i + j; 3409 PetscCall(MatGetValues(A, bs, IJ, bs, IJ, diag)); 3410 PetscCall(PetscKernel_A_gets_inverse_A(bs, diag, v_pivots, v_work, allowzeropivot, &zeropivotdetected)); 3411 if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT; 3412 PetscCall(PetscKernel_A_gets_transpose_A_N(diag, bs)); 3413 diag += bs2; 3414 } 3415 PetscCall(PetscFree3(v_work, v_pivots, IJ)); 3416 } 3417 a->ibdiagvalid = PETSC_TRUE; 3418 PetscFunctionReturn(PETSC_SUCCESS); 3419 } 3420 3421 static PetscErrorCode MatSetRandom_SeqAIJ(Mat x, PetscRandom rctx) 3422 { 3423 Mat_SeqAIJ *aij = (Mat_SeqAIJ *)x->data; 3424 PetscScalar a, *aa; 3425 PetscInt m, n, i, j, col; 3426 3427 PetscFunctionBegin; 3428 if (!x->assembled) { 3429 PetscCall(MatGetSize(x, &m, &n)); 3430 for (i = 0; i < m; i++) { 3431 for (j = 0; j < aij->imax[i]; j++) { 3432 PetscCall(PetscRandomGetValue(rctx, &a)); 3433 col = (PetscInt)(n * PetscRealPart(a)); 3434 PetscCall(MatSetValues(x, 1, &i, 1, &col, &a, ADD_VALUES)); 3435 } 3436 } 3437 } else { 3438 PetscCall(MatSeqAIJGetArrayWrite(x, &aa)); 3439 for (i = 0; i < aij->nz; i++) PetscCall(PetscRandomGetValue(rctx, aa + i)); 3440 PetscCall(MatSeqAIJRestoreArrayWrite(x, &aa)); 3441 } 3442 PetscCall(MatAssemblyBegin(x, MAT_FINAL_ASSEMBLY)); 3443 PetscCall(MatAssemblyEnd(x, MAT_FINAL_ASSEMBLY)); 3444 PetscFunctionReturn(PETSC_SUCCESS); 3445 } 3446 3447 /* Like MatSetRandom_SeqAIJ, but do not set values on columns in range of [low, high) */ 3448 PetscErrorCode MatSetRandomSkipColumnRange_SeqAIJ_Private(Mat x, PetscInt low, PetscInt high, PetscRandom rctx) 3449 { 3450 Mat_SeqAIJ *aij = (Mat_SeqAIJ *)x->data; 3451 PetscScalar a; 3452 PetscInt m, n, i, j, col, nskip; 3453 3454 PetscFunctionBegin; 3455 nskip = high - low; 3456 PetscCall(MatGetSize(x, &m, &n)); 3457 n -= nskip; /* shrink number of columns where nonzeros can be set */ 3458 for (i = 0; i < m; i++) { 3459 for (j = 0; j < aij->imax[i]; j++) { 3460 PetscCall(PetscRandomGetValue(rctx, &a)); 3461 col = (PetscInt)(n * PetscRealPart(a)); 3462 if (col >= low) col += nskip; /* shift col rightward to skip the hole */ 3463 PetscCall(MatSetValues(x, 1, &i, 1, &col, &a, ADD_VALUES)); 3464 } 3465 } 3466 PetscCall(MatAssemblyBegin(x, MAT_FINAL_ASSEMBLY)); 3467 PetscCall(MatAssemblyEnd(x, MAT_FINAL_ASSEMBLY)); 3468 PetscFunctionReturn(PETSC_SUCCESS); 3469 } 3470 3471 static struct _MatOps MatOps_Values = {MatSetValues_SeqAIJ, 3472 MatGetRow_SeqAIJ, 3473 MatRestoreRow_SeqAIJ, 3474 MatMult_SeqAIJ, 3475 /* 4*/ MatMultAdd_SeqAIJ, 3476 MatMultTranspose_SeqAIJ, 3477 MatMultTransposeAdd_SeqAIJ, 3478 NULL, 3479 NULL, 3480 NULL, 3481 /* 10*/ NULL, 3482 MatLUFactor_SeqAIJ, 3483 NULL, 3484 MatSOR_SeqAIJ, 3485 MatTranspose_SeqAIJ, 3486 /*1 5*/ MatGetInfo_SeqAIJ, 3487 MatEqual_SeqAIJ, 3488 MatGetDiagonal_SeqAIJ, 3489 MatDiagonalScale_SeqAIJ, 3490 MatNorm_SeqAIJ, 3491 /* 20*/ NULL, 3492 MatAssemblyEnd_SeqAIJ, 3493 MatSetOption_SeqAIJ, 3494 MatZeroEntries_SeqAIJ, 3495 /* 24*/ MatZeroRows_SeqAIJ, 3496 NULL, 3497 NULL, 3498 NULL, 3499 NULL, 3500 /* 29*/ MatSetUp_Seq_Hash, 3501 NULL, 3502 NULL, 3503 NULL, 3504 NULL, 3505 /* 34*/ MatDuplicate_SeqAIJ, 3506 NULL, 3507 NULL, 3508 MatILUFactor_SeqAIJ, 3509 NULL, 3510 /* 39*/ MatAXPY_SeqAIJ, 3511 MatCreateSubMatrices_SeqAIJ, 3512 MatIncreaseOverlap_SeqAIJ, 3513 MatGetValues_SeqAIJ, 3514 MatCopy_SeqAIJ, 3515 /* 44*/ MatGetRowMax_SeqAIJ, 3516 MatScale_SeqAIJ, 3517 MatShift_SeqAIJ, 3518 MatDiagonalSet_SeqAIJ, 3519 MatZeroRowsColumns_SeqAIJ, 3520 /* 49*/ MatSetRandom_SeqAIJ, 3521 MatGetRowIJ_SeqAIJ, 3522 MatRestoreRowIJ_SeqAIJ, 3523 MatGetColumnIJ_SeqAIJ, 3524 MatRestoreColumnIJ_SeqAIJ, 3525 /* 54*/ MatFDColoringCreate_SeqXAIJ, 3526 NULL, 3527 NULL, 3528 MatPermute_SeqAIJ, 3529 NULL, 3530 /* 59*/ NULL, 3531 MatDestroy_SeqAIJ, 3532 MatView_SeqAIJ, 3533 NULL, 3534 NULL, 3535 /* 64*/ NULL, 3536 MatMatMatMultNumeric_SeqAIJ_SeqAIJ_SeqAIJ, 3537 NULL, 3538 NULL, 3539 NULL, 3540 /* 69*/ MatGetRowMaxAbs_SeqAIJ, 3541 MatGetRowMinAbs_SeqAIJ, 3542 NULL, 3543 NULL, 3544 NULL, 3545 /* 74*/ NULL, 3546 MatFDColoringApply_AIJ, 3547 NULL, 3548 NULL, 3549 NULL, 3550 /* 79*/ MatFindZeroDiagonals_SeqAIJ, 3551 NULL, 3552 NULL, 3553 NULL, 3554 MatLoad_SeqAIJ, 3555 /* 84*/ MatIsSymmetric_SeqAIJ, 3556 MatIsHermitian_SeqAIJ, 3557 NULL, 3558 NULL, 3559 NULL, 3560 /* 89*/ NULL, 3561 NULL, 3562 MatMatMultNumeric_SeqAIJ_SeqAIJ, 3563 NULL, 3564 NULL, 3565 /* 94*/ MatPtAPNumeric_SeqAIJ_SeqAIJ_SparseAxpy, 3566 NULL, 3567 NULL, 3568 MatMatTransposeMultNumeric_SeqAIJ_SeqAIJ, 3569 NULL, 3570 /* 99*/ MatProductSetFromOptions_SeqAIJ, 3571 NULL, 3572 NULL, 3573 MatConjugate_SeqAIJ, 3574 NULL, 3575 /*104*/ MatSetValuesRow_SeqAIJ, 3576 MatRealPart_SeqAIJ, 3577 MatImaginaryPart_SeqAIJ, 3578 NULL, 3579 NULL, 3580 /*109*/ MatMatSolve_SeqAIJ, 3581 NULL, 3582 MatGetRowMin_SeqAIJ, 3583 NULL, 3584 MatMissingDiagonal_SeqAIJ, 3585 /*114*/ NULL, 3586 NULL, 3587 NULL, 3588 NULL, 3589 NULL, 3590 /*119*/ NULL, 3591 NULL, 3592 NULL, 3593 NULL, 3594 MatGetMultiProcBlock_SeqAIJ, 3595 /*124*/ MatFindNonzeroRows_SeqAIJ, 3596 MatGetColumnReductions_SeqAIJ, 3597 MatInvertBlockDiagonal_SeqAIJ, 3598 MatInvertVariableBlockDiagonal_SeqAIJ, 3599 NULL, 3600 /*129*/ NULL, 3601 NULL, 3602 NULL, 3603 MatTransposeMatMultNumeric_SeqAIJ_SeqAIJ, 3604 MatTransposeColoringCreate_SeqAIJ, 3605 /*134*/ MatTransColoringApplySpToDen_SeqAIJ, 3606 MatTransColoringApplyDenToSp_SeqAIJ, 3607 NULL, 3608 NULL, 3609 MatRARtNumeric_SeqAIJ_SeqAIJ, 3610 /*139*/ NULL, 3611 NULL, 3612 NULL, 3613 MatFDColoringSetUp_SeqXAIJ, 3614 MatFindOffBlockDiagonalEntries_SeqAIJ, 3615 MatCreateMPIMatConcatenateSeqMat_SeqAIJ, 3616 /*145*/ MatDestroySubMatrices_SeqAIJ, 3617 NULL, 3618 NULL, 3619 MatCreateGraph_Simple_AIJ, 3620 NULL, 3621 /*150*/ MatTransposeSymbolic_SeqAIJ, 3622 MatEliminateZeros_SeqAIJ}; 3623 3624 static PetscErrorCode MatSeqAIJSetColumnIndices_SeqAIJ(Mat mat, PetscInt *indices) 3625 { 3626 Mat_SeqAIJ *aij = (Mat_SeqAIJ *)mat->data; 3627 PetscInt i, nz, n; 3628 3629 PetscFunctionBegin; 3630 nz = aij->maxnz; 3631 n = mat->rmap->n; 3632 for (i = 0; i < nz; i++) aij->j[i] = indices[i]; 3633 aij->nz = nz; 3634 for (i = 0; i < n; i++) aij->ilen[i] = aij->imax[i]; 3635 PetscFunctionReturn(PETSC_SUCCESS); 3636 } 3637 3638 /* 3639 * Given a sparse matrix with global column indices, compact it by using a local column space. 3640 * The result matrix helps saving memory in other algorithms, such as MatPtAPSymbolic_MPIAIJ_MPIAIJ_scalable() 3641 */ 3642 PetscErrorCode MatSeqAIJCompactOutExtraColumns_SeqAIJ(Mat mat, ISLocalToGlobalMapping *mapping) 3643 { 3644 Mat_SeqAIJ *aij = (Mat_SeqAIJ *)mat->data; 3645 PetscHMapI gid1_lid1; 3646 PetscHashIter tpos; 3647 PetscInt gid, lid, i, ec, nz = aij->nz; 3648 PetscInt *garray, *jj = aij->j; 3649 3650 PetscFunctionBegin; 3651 PetscValidHeaderSpecific(mat, MAT_CLASSID, 1); 3652 PetscAssertPointer(mapping, 2); 3653 /* use a table */ 3654 PetscCall(PetscHMapICreateWithSize(mat->rmap->n, &gid1_lid1)); 3655 ec = 0; 3656 for (i = 0; i < nz; i++) { 3657 PetscInt data, gid1 = jj[i] + 1; 3658 PetscCall(PetscHMapIGetWithDefault(gid1_lid1, gid1, 0, &data)); 3659 if (!data) { 3660 /* one based table */ 3661 PetscCall(PetscHMapISet(gid1_lid1, gid1, ++ec)); 3662 } 3663 } 3664 /* form array of columns we need */ 3665 PetscCall(PetscMalloc1(ec, &garray)); 3666 PetscHashIterBegin(gid1_lid1, tpos); 3667 while (!PetscHashIterAtEnd(gid1_lid1, tpos)) { 3668 PetscHashIterGetKey(gid1_lid1, tpos, gid); 3669 PetscHashIterGetVal(gid1_lid1, tpos, lid); 3670 PetscHashIterNext(gid1_lid1, tpos); 3671 gid--; 3672 lid--; 3673 garray[lid] = gid; 3674 } 3675 PetscCall(PetscSortInt(ec, garray)); /* sort, and rebuild */ 3676 PetscCall(PetscHMapIClear(gid1_lid1)); 3677 for (i = 0; i < ec; i++) PetscCall(PetscHMapISet(gid1_lid1, garray[i] + 1, i + 1)); 3678 /* compact out the extra columns in B */ 3679 for (i = 0; i < nz; i++) { 3680 PetscInt gid1 = jj[i] + 1; 3681 PetscCall(PetscHMapIGetWithDefault(gid1_lid1, gid1, 0, &lid)); 3682 lid--; 3683 jj[i] = lid; 3684 } 3685 PetscCall(PetscLayoutDestroy(&mat->cmap)); 3686 PetscCall(PetscHMapIDestroy(&gid1_lid1)); 3687 PetscCall(PetscLayoutCreateFromSizes(PetscObjectComm((PetscObject)mat), ec, ec, 1, &mat->cmap)); 3688 PetscCall(ISLocalToGlobalMappingCreate(PETSC_COMM_SELF, mat->cmap->bs, mat->cmap->n, garray, PETSC_OWN_POINTER, mapping)); 3689 PetscCall(ISLocalToGlobalMappingSetType(*mapping, ISLOCALTOGLOBALMAPPINGHASH)); 3690 PetscFunctionReturn(PETSC_SUCCESS); 3691 } 3692 3693 /*@ 3694 MatSeqAIJSetColumnIndices - Set the column indices for all the rows 3695 in the matrix. 3696 3697 Input Parameters: 3698 + mat - the `MATSEQAIJ` matrix 3699 - indices - the column indices 3700 3701 Level: advanced 3702 3703 Notes: 3704 This can be called if you have precomputed the nonzero structure of the 3705 matrix and want to provide it to the matrix object to improve the performance 3706 of the `MatSetValues()` operation. 3707 3708 You MUST have set the correct numbers of nonzeros per row in the call to 3709 `MatCreateSeqAIJ()`, and the columns indices MUST be sorted. 3710 3711 MUST be called before any calls to `MatSetValues()` 3712 3713 The indices should start with zero, not one. 3714 3715 .seealso: [](ch_matrices), `Mat`, `MATSEQAIJ` 3716 @*/ 3717 PetscErrorCode MatSeqAIJSetColumnIndices(Mat mat, PetscInt *indices) 3718 { 3719 PetscFunctionBegin; 3720 PetscValidHeaderSpecific(mat, MAT_CLASSID, 1); 3721 PetscAssertPointer(indices, 2); 3722 PetscUseMethod(mat, "MatSeqAIJSetColumnIndices_C", (Mat, PetscInt *), (mat, indices)); 3723 PetscFunctionReturn(PETSC_SUCCESS); 3724 } 3725 3726 static PetscErrorCode MatStoreValues_SeqAIJ(Mat mat) 3727 { 3728 Mat_SeqAIJ *aij = (Mat_SeqAIJ *)mat->data; 3729 size_t nz = aij->i[mat->rmap->n]; 3730 3731 PetscFunctionBegin; 3732 PetscCheck(aij->nonew, PETSC_COMM_SELF, PETSC_ERR_ORDER, "Must call MatSetOption(A,MAT_NEW_NONZERO_LOCATIONS,PETSC_FALSE);first"); 3733 3734 /* allocate space for values if not already there */ 3735 if (!aij->saved_values) { PetscCall(PetscMalloc1(nz + 1, &aij->saved_values)); } 3736 3737 /* copy values over */ 3738 PetscCall(PetscArraycpy(aij->saved_values, aij->a, nz)); 3739 PetscFunctionReturn(PETSC_SUCCESS); 3740 } 3741 3742 /*@ 3743 MatStoreValues - Stashes a copy of the matrix values; this allows reusing of the linear part of a Jacobian, while recomputing only the 3744 nonlinear portion. 3745 3746 Logically Collect 3747 3748 Input Parameter: 3749 . mat - the matrix (currently only `MATAIJ` matrices support this option) 3750 3751 Level: advanced 3752 3753 Example Usage: 3754 .vb 3755 Using SNES 3756 Create Jacobian matrix 3757 Set linear terms into matrix 3758 Apply boundary conditions to matrix, at this time matrix must have 3759 final nonzero structure (i.e. setting the nonlinear terms and applying 3760 boundary conditions again will not change the nonzero structure 3761 MatSetOption(mat, MAT_NEW_NONZERO_LOCATIONS, PETSC_FALSE); 3762 MatStoreValues(mat); 3763 Call SNESSetJacobian() with matrix 3764 In your Jacobian routine 3765 MatRetrieveValues(mat); 3766 Set nonlinear terms in matrix 3767 3768 Without `SNESSolve()`, i.e. when you handle nonlinear solve yourself: 3769 // build linear portion of Jacobian 3770 MatSetOption(mat, MAT_NEW_NONZERO_LOCATIONS, PETSC_FALSE); 3771 MatStoreValues(mat); 3772 loop over nonlinear iterations 3773 MatRetrieveValues(mat); 3774 // call MatSetValues(mat,...) to set nonliner portion of Jacobian 3775 // call MatAssemblyBegin/End() on matrix 3776 Solve linear system with Jacobian 3777 endloop 3778 .ve 3779 3780 Notes: 3781 Matrix must already be assembled before calling this routine 3782 Must set the matrix option `MatSetOption`(mat,`MAT_NEW_NONZERO_LOCATIONS`,`PETSC_FALSE`); before 3783 calling this routine. 3784 3785 When this is called multiple times it overwrites the previous set of stored values 3786 and does not allocated additional space. 3787 3788 .seealso: [](ch_matrices), `Mat`, `MatRetrieveValues()` 3789 @*/ 3790 PetscErrorCode MatStoreValues(Mat mat) 3791 { 3792 PetscFunctionBegin; 3793 PetscValidHeaderSpecific(mat, MAT_CLASSID, 1); 3794 PetscCheck(mat->assembled, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Not for unassembled matrix"); 3795 PetscCheck(!mat->factortype, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Not for factored matrix"); 3796 PetscUseMethod(mat, "MatStoreValues_C", (Mat), (mat)); 3797 PetscFunctionReturn(PETSC_SUCCESS); 3798 } 3799 3800 static PetscErrorCode MatRetrieveValues_SeqAIJ(Mat mat) 3801 { 3802 Mat_SeqAIJ *aij = (Mat_SeqAIJ *)mat->data; 3803 PetscInt nz = aij->i[mat->rmap->n]; 3804 3805 PetscFunctionBegin; 3806 PetscCheck(aij->nonew, PETSC_COMM_SELF, PETSC_ERR_ORDER, "Must call MatSetOption(A,MAT_NEW_NONZERO_LOCATIONS,PETSC_FALSE);first"); 3807 PetscCheck(aij->saved_values, PETSC_COMM_SELF, PETSC_ERR_ORDER, "Must call MatStoreValues(A);first"); 3808 /* copy values over */ 3809 PetscCall(PetscArraycpy(aij->a, aij->saved_values, nz)); 3810 PetscFunctionReturn(PETSC_SUCCESS); 3811 } 3812 3813 /*@ 3814 MatRetrieveValues - Retrieves the copy of the matrix values that was stored with `MatStoreValues()` 3815 3816 Logically Collect 3817 3818 Input Parameter: 3819 . mat - the matrix (currently only `MATAIJ` matrices support this option) 3820 3821 Level: advanced 3822 3823 .seealso: [](ch_matrices), `Mat`, `MatStoreValues()` 3824 @*/ 3825 PetscErrorCode MatRetrieveValues(Mat mat) 3826 { 3827 PetscFunctionBegin; 3828 PetscValidHeaderSpecific(mat, MAT_CLASSID, 1); 3829 PetscCheck(mat->assembled, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Not for unassembled matrix"); 3830 PetscCheck(!mat->factortype, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Not for factored matrix"); 3831 PetscUseMethod(mat, "MatRetrieveValues_C", (Mat), (mat)); 3832 PetscFunctionReturn(PETSC_SUCCESS); 3833 } 3834 3835 /*@C 3836 MatCreateSeqAIJ - Creates a sparse matrix in `MATSEQAIJ` (compressed row) format 3837 (the default parallel PETSc format). For good matrix assembly performance 3838 the user should preallocate the matrix storage by setting the parameter `nz` 3839 (or the array `nnz`). 3840 3841 Collective 3842 3843 Input Parameters: 3844 + comm - MPI communicator, set to `PETSC_COMM_SELF` 3845 . m - number of rows 3846 . n - number of columns 3847 . nz - number of nonzeros per row (same for all rows) 3848 - nnz - array containing the number of nonzeros in the various rows 3849 (possibly different for each row) or NULL 3850 3851 Output Parameter: 3852 . A - the matrix 3853 3854 Options Database Keys: 3855 + -mat_no_inode - Do not use inodes 3856 - -mat_inode_limit <limit> - Sets inode limit (max limit=5) 3857 3858 Level: intermediate 3859 3860 Notes: 3861 It is recommend to use `MatCreateFromOptions()` instead of this routine 3862 3863 If `nnz` is given then `nz` is ignored 3864 3865 The `MATSEQAIJ` format, also called 3866 compressed row storage, is fully compatible with standard Fortran 3867 storage. That is, the stored row and column indices can begin at 3868 either one (as in Fortran) or zero. 3869 3870 Specify the preallocated storage with either `nz` or `nnz` (not both). 3871 Set `nz` = `PETSC_DEFAULT` and `nnz` = `NULL` for PETSc to control dynamic memory 3872 allocation. 3873 3874 By default, this format uses inodes (identical nodes) when possible, to 3875 improve numerical efficiency of matrix-vector products and solves. We 3876 search for consecutive rows with the same nonzero structure, thereby 3877 reusing matrix information to achieve increased efficiency. 3878 3879 .seealso: [](ch_matrices), `Mat`, [Sparse Matrix Creation](sec_matsparse), `MatCreate()`, `MatCreateAIJ()`, `MatSetValues()`, `MatSeqAIJSetColumnIndices()`, `MatCreateSeqAIJWithArrays()` 3880 @*/ 3881 PetscErrorCode MatCreateSeqAIJ(MPI_Comm comm, PetscInt m, PetscInt n, PetscInt nz, const PetscInt nnz[], Mat *A) 3882 { 3883 PetscFunctionBegin; 3884 PetscCall(MatCreate(comm, A)); 3885 PetscCall(MatSetSizes(*A, m, n, m, n)); 3886 PetscCall(MatSetType(*A, MATSEQAIJ)); 3887 PetscCall(MatSeqAIJSetPreallocation_SeqAIJ(*A, nz, nnz)); 3888 PetscFunctionReturn(PETSC_SUCCESS); 3889 } 3890 3891 /*@C 3892 MatSeqAIJSetPreallocation - For good matrix assembly performance 3893 the user should preallocate the matrix storage by setting the parameter nz 3894 (or the array nnz). By setting these parameters accurately, performance 3895 during matrix assembly can be increased by more than a factor of 50. 3896 3897 Collective 3898 3899 Input Parameters: 3900 + B - The matrix 3901 . nz - number of nonzeros per row (same for all rows) 3902 - nnz - array containing the number of nonzeros in the various rows 3903 (possibly different for each row) or NULL 3904 3905 Options Database Keys: 3906 + -mat_no_inode - Do not use inodes 3907 - -mat_inode_limit <limit> - Sets inode limit (max limit=5) 3908 3909 Level: intermediate 3910 3911 Notes: 3912 If `nnz` is given then `nz` is ignored 3913 3914 The `MATSEQAIJ` format also called 3915 compressed row storage, is fully compatible with standard Fortran 3916 storage. That is, the stored row and column indices can begin at 3917 either one (as in Fortran) or zero. See the users' manual for details. 3918 3919 Specify the preallocated storage with either `nz` or `nnz` (not both). 3920 Set nz = `PETSC_DEFAULT` and `nnz` = `NULL` for PETSc to control dynamic memory 3921 allocation. 3922 3923 You can call `MatGetInfo()` to get information on how effective the preallocation was; 3924 for example the fields mallocs,nz_allocated,nz_used,nz_unneeded; 3925 You can also run with the option -info and look for messages with the string 3926 malloc in them to see if additional memory allocation was needed. 3927 3928 Developer Notes: 3929 Use nz of `MAT_SKIP_ALLOCATION` to not allocate any space for the matrix 3930 entries or columns indices 3931 3932 By default, this format uses inodes (identical nodes) when possible, to 3933 improve numerical efficiency of matrix-vector products and solves. We 3934 search for consecutive rows with the same nonzero structure, thereby 3935 reusing matrix information to achieve increased efficiency. 3936 3937 .seealso: [](ch_matrices), `Mat`, `MatCreate()`, `MatCreateAIJ()`, `MatSetValues()`, `MatSeqAIJSetColumnIndices()`, `MatCreateSeqAIJWithArrays()`, `MatGetInfo()`, 3938 `MatSeqAIJSetTotalPreallocation()` 3939 @*/ 3940 PetscErrorCode MatSeqAIJSetPreallocation(Mat B, PetscInt nz, const PetscInt nnz[]) 3941 { 3942 PetscFunctionBegin; 3943 PetscValidHeaderSpecific(B, MAT_CLASSID, 1); 3944 PetscValidType(B, 1); 3945 PetscTryMethod(B, "MatSeqAIJSetPreallocation_C", (Mat, PetscInt, const PetscInt[]), (B, nz, nnz)); 3946 PetscFunctionReturn(PETSC_SUCCESS); 3947 } 3948 3949 PetscErrorCode MatSeqAIJSetPreallocation_SeqAIJ(Mat B, PetscInt nz, const PetscInt *nnz) 3950 { 3951 Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data; 3952 PetscBool skipallocation = PETSC_FALSE, realalloc = PETSC_FALSE; 3953 PetscInt i; 3954 3955 PetscFunctionBegin; 3956 if (B->hash_active) { 3957 B->ops[0] = b->cops; 3958 PetscCall(PetscHMapIJVDestroy(&b->ht)); 3959 PetscCall(PetscFree(b->dnz)); 3960 B->hash_active = PETSC_FALSE; 3961 } 3962 if (nz >= 0 || nnz) realalloc = PETSC_TRUE; 3963 if (nz == MAT_SKIP_ALLOCATION) { 3964 skipallocation = PETSC_TRUE; 3965 nz = 0; 3966 } 3967 PetscCall(PetscLayoutSetUp(B->rmap)); 3968 PetscCall(PetscLayoutSetUp(B->cmap)); 3969 3970 if (nz == PETSC_DEFAULT || nz == PETSC_DECIDE) nz = 5; 3971 PetscCheck(nz >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "nz cannot be less than 0: value %" PetscInt_FMT, nz); 3972 if (PetscUnlikelyDebug(nnz)) { 3973 for (i = 0; i < B->rmap->n; i++) { 3974 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]); 3975 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); 3976 } 3977 } 3978 3979 B->preallocated = PETSC_TRUE; 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 PetscCall(PetscIntMultError(nz, B->rmap->n, &nz)); 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(PETSC_SUCCESS); 4044 } 4045 4046 static PetscErrorCode MatResetPreallocation_SeqAIJ(Mat A) 4047 { 4048 Mat_SeqAIJ *a; 4049 PetscInt i; 4050 PetscBool skipreset; 4051 4052 PetscFunctionBegin; 4053 PetscValidHeaderSpecific(A, MAT_CLASSID, 1); 4054 4055 /* Check local size. If zero, then return */ 4056 if (!A->rmap->n) PetscFunctionReturn(PETSC_SUCCESS); 4057 4058 a = (Mat_SeqAIJ *)A->data; 4059 /* if no saved info, we error out */ 4060 PetscCheck(a->ipre, PETSC_COMM_SELF, PETSC_ERR_ARG_NULL, "No saved preallocation info "); 4061 4062 PetscCheck(a->i && a->imax && a->ilen, PETSC_COMM_SELF, PETSC_ERR_ARG_NULL, "Memory info is incomplete, and can not reset preallocation "); 4063 4064 PetscCall(PetscArraycmp(a->ipre, a->ilen, A->rmap->n, &skipreset)); 4065 if (!skipreset) { 4066 PetscCall(PetscArraycpy(a->imax, a->ipre, A->rmap->n)); 4067 PetscCall(PetscArrayzero(a->ilen, A->rmap->n)); 4068 a->i[0] = 0; 4069 for (i = 1; i < A->rmap->n + 1; i++) a->i[i] = a->i[i - 1] + a->imax[i - 1]; 4070 A->preallocated = PETSC_TRUE; 4071 a->nz = 0; 4072 a->maxnz = a->i[A->rmap->n]; 4073 A->info.nz_unneeded = (double)a->maxnz; 4074 A->was_assembled = PETSC_FALSE; 4075 A->assembled = PETSC_FALSE; 4076 } 4077 PetscFunctionReturn(PETSC_SUCCESS); 4078 } 4079 4080 /*@ 4081 MatSeqAIJSetPreallocationCSR - Allocates memory for a sparse sequential matrix in `MATSEQAIJ` format. 4082 4083 Input Parameters: 4084 + B - the matrix 4085 . i - the indices into j for the start of each row (starts with zero) 4086 . j - the column indices for each row (starts with zero) these must be sorted for each row 4087 - v - optional values in the matrix 4088 4089 Level: developer 4090 4091 Notes: 4092 The `i`,`j`,`v` values are COPIED with this routine; to avoid the copy use `MatCreateSeqAIJWithArrays()` 4093 4094 This routine may be called multiple times with different nonzero patterns (or the same nonzero pattern). The nonzero 4095 structure will be the union of all the previous nonzero structures. 4096 4097 Developer Notes: 4098 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 4099 then just copies the `v` values directly with `PetscMemcpy()`. 4100 4101 This routine could also take a `PetscCopyMode` argument to allow sharing the values instead of always copying them. 4102 4103 .seealso: [](ch_matrices), `Mat`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatSeqAIJSetPreallocation()`, `MATSEQAIJ`, `MatResetPreallocation()` 4104 @*/ 4105 PetscErrorCode MatSeqAIJSetPreallocationCSR(Mat B, const PetscInt i[], const PetscInt j[], const PetscScalar v[]) 4106 { 4107 PetscFunctionBegin; 4108 PetscValidHeaderSpecific(B, MAT_CLASSID, 1); 4109 PetscValidType(B, 1); 4110 PetscTryMethod(B, "MatSeqAIJSetPreallocationCSR_C", (Mat, const PetscInt[], const PetscInt[], const PetscScalar[]), (B, i, j, v)); 4111 PetscFunctionReturn(PETSC_SUCCESS); 4112 } 4113 4114 static PetscErrorCode MatSeqAIJSetPreallocationCSR_SeqAIJ(Mat B, const PetscInt Ii[], const PetscInt J[], const PetscScalar v[]) 4115 { 4116 PetscInt i; 4117 PetscInt m, n; 4118 PetscInt nz; 4119 PetscInt *nnz; 4120 4121 PetscFunctionBegin; 4122 PetscCheck(Ii[0] == 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Ii[0] must be 0 it is %" PetscInt_FMT, Ii[0]); 4123 4124 PetscCall(PetscLayoutSetUp(B->rmap)); 4125 PetscCall(PetscLayoutSetUp(B->cmap)); 4126 4127 PetscCall(MatGetSize(B, &m, &n)); 4128 PetscCall(PetscMalloc1(m + 1, &nnz)); 4129 for (i = 0; i < m; i++) { 4130 nz = Ii[i + 1] - Ii[i]; 4131 PetscCheck(nz >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Local row %" PetscInt_FMT " has a negative number of columns %" PetscInt_FMT, i, nz); 4132 nnz[i] = nz; 4133 } 4134 PetscCall(MatSeqAIJSetPreallocation(B, 0, nnz)); 4135 PetscCall(PetscFree(nnz)); 4136 4137 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)); 4138 4139 PetscCall(MatAssemblyBegin(B, MAT_FINAL_ASSEMBLY)); 4140 PetscCall(MatAssemblyEnd(B, MAT_FINAL_ASSEMBLY)); 4141 4142 PetscCall(MatSetOption(B, MAT_NEW_NONZERO_LOCATION_ERR, PETSC_TRUE)); 4143 PetscFunctionReturn(PETSC_SUCCESS); 4144 } 4145 4146 /*@ 4147 MatSeqAIJKron - Computes `C`, the Kronecker product of `A` and `B`. 4148 4149 Input Parameters: 4150 + A - left-hand side matrix 4151 . B - right-hand side matrix 4152 - reuse - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX` 4153 4154 Output Parameter: 4155 . C - Kronecker product of `A` and `B` 4156 4157 Level: intermediate 4158 4159 Note: 4160 `MAT_REUSE_MATRIX` can only be used when the nonzero structure of the product matrix has not changed from that last call to `MatSeqAIJKron()`. 4161 4162 .seealso: [](ch_matrices), `Mat`, `MatCreateSeqAIJ()`, `MATSEQAIJ`, `MATKAIJ`, `MatReuse` 4163 @*/ 4164 PetscErrorCode MatSeqAIJKron(Mat A, Mat B, MatReuse reuse, Mat *C) 4165 { 4166 PetscFunctionBegin; 4167 PetscValidHeaderSpecific(A, MAT_CLASSID, 1); 4168 PetscValidType(A, 1); 4169 PetscValidHeaderSpecific(B, MAT_CLASSID, 2); 4170 PetscValidType(B, 2); 4171 PetscAssertPointer(C, 4); 4172 if (reuse == MAT_REUSE_MATRIX) { 4173 PetscValidHeaderSpecific(*C, MAT_CLASSID, 4); 4174 PetscValidType(*C, 4); 4175 } 4176 PetscTryMethod(A, "MatSeqAIJKron_C", (Mat, Mat, MatReuse, Mat *), (A, B, reuse, C)); 4177 PetscFunctionReturn(PETSC_SUCCESS); 4178 } 4179 4180 static PetscErrorCode MatSeqAIJKron_SeqAIJ(Mat A, Mat B, MatReuse reuse, Mat *C) 4181 { 4182 Mat newmat; 4183 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 4184 Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data; 4185 PetscScalar *v; 4186 const PetscScalar *aa, *ba; 4187 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; 4188 PetscBool flg; 4189 4190 PetscFunctionBegin; 4191 PetscCheck(!A->factortype, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Not for factored matrix"); 4192 PetscCheck(A->assembled, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Not for unassembled matrix"); 4193 PetscCheck(!B->factortype, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Not for factored matrix"); 4194 PetscCheck(B->assembled, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Not for unassembled matrix"); 4195 PetscCall(PetscObjectTypeCompare((PetscObject)B, MATSEQAIJ, &flg)); 4196 PetscCheck(flg, PETSC_COMM_SELF, PETSC_ERR_SUP, "MatType %s", ((PetscObject)B)->type_name); 4197 PetscCheck(reuse == MAT_INITIAL_MATRIX || reuse == MAT_REUSE_MATRIX, PETSC_COMM_SELF, PETSC_ERR_SUP, "MatReuse %d", (int)reuse); 4198 if (reuse == MAT_INITIAL_MATRIX) { 4199 PetscCall(PetscMalloc2(am * bm + 1, &i, a->i[am] * b->i[bm], &j)); 4200 PetscCall(MatCreate(PETSC_COMM_SELF, &newmat)); 4201 PetscCall(MatSetSizes(newmat, am * bm, an * bn, am * bm, an * bn)); 4202 PetscCall(MatSetType(newmat, MATAIJ)); 4203 i[0] = 0; 4204 for (m = 0; m < am; ++m) { 4205 for (p = 0; p < bm; ++p) { 4206 i[m * bm + p + 1] = i[m * bm + p] + (a->i[m + 1] - a->i[m]) * (b->i[p + 1] - b->i[p]); 4207 for (n = a->i[m]; n < a->i[m + 1]; ++n) { 4208 for (q = b->i[p]; q < b->i[p + 1]; ++q) j[nnz++] = a->j[n] * bn + b->j[q]; 4209 } 4210 } 4211 } 4212 PetscCall(MatSeqAIJSetPreallocationCSR(newmat, i, j, NULL)); 4213 *C = newmat; 4214 PetscCall(PetscFree2(i, j)); 4215 nnz = 0; 4216 } 4217 PetscCall(MatSeqAIJGetArray(*C, &v)); 4218 PetscCall(MatSeqAIJGetArrayRead(A, &aa)); 4219 PetscCall(MatSeqAIJGetArrayRead(B, &ba)); 4220 for (m = 0; m < am; ++m) { 4221 for (p = 0; p < bm; ++p) { 4222 for (n = a->i[m]; n < a->i[m + 1]; ++n) { 4223 for (q = b->i[p]; q < b->i[p + 1]; ++q) v[nnz++] = aa[n] * ba[q]; 4224 } 4225 } 4226 } 4227 PetscCall(MatSeqAIJRestoreArray(*C, &v)); 4228 PetscCall(MatSeqAIJRestoreArrayRead(A, &aa)); 4229 PetscCall(MatSeqAIJRestoreArrayRead(B, &ba)); 4230 PetscFunctionReturn(PETSC_SUCCESS); 4231 } 4232 4233 #include <../src/mat/impls/dense/seq/dense.h> 4234 #include <petsc/private/kernels/petscaxpy.h> 4235 4236 /* 4237 Computes (B'*A')' since computing B*A directly is untenable 4238 4239 n p p 4240 [ ] [ ] [ ] 4241 m [ A ] * n [ B ] = m [ C ] 4242 [ ] [ ] [ ] 4243 4244 */ 4245 PetscErrorCode MatMatMultNumeric_SeqDense_SeqAIJ(Mat A, Mat B, Mat C) 4246 { 4247 Mat_SeqDense *sub_a = (Mat_SeqDense *)A->data; 4248 Mat_SeqAIJ *sub_b = (Mat_SeqAIJ *)B->data; 4249 Mat_SeqDense *sub_c = (Mat_SeqDense *)C->data; 4250 PetscInt i, j, n, m, q, p; 4251 const PetscInt *ii, *idx; 4252 const PetscScalar *b, *a, *a_q; 4253 PetscScalar *c, *c_q; 4254 PetscInt clda = sub_c->lda; 4255 PetscInt alda = sub_a->lda; 4256 4257 PetscFunctionBegin; 4258 m = A->rmap->n; 4259 n = A->cmap->n; 4260 p = B->cmap->n; 4261 a = sub_a->v; 4262 b = sub_b->a; 4263 c = sub_c->v; 4264 if (clda == m) { 4265 PetscCall(PetscArrayzero(c, m * p)); 4266 } else { 4267 for (j = 0; j < p; j++) 4268 for (i = 0; i < m; i++) c[j * clda + i] = 0.0; 4269 } 4270 ii = sub_b->i; 4271 idx = sub_b->j; 4272 for (i = 0; i < n; i++) { 4273 q = ii[i + 1] - ii[i]; 4274 while (q-- > 0) { 4275 c_q = c + clda * (*idx); 4276 a_q = a + alda * i; 4277 PetscKernelAXPY(c_q, *b, a_q, m); 4278 idx++; 4279 b++; 4280 } 4281 } 4282 PetscFunctionReturn(PETSC_SUCCESS); 4283 } 4284 4285 PetscErrorCode MatMatMultSymbolic_SeqDense_SeqAIJ(Mat A, Mat B, PetscReal fill, Mat C) 4286 { 4287 PetscInt m = A->rmap->n, n = B->cmap->n; 4288 PetscBool cisdense; 4289 4290 PetscFunctionBegin; 4291 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); 4292 PetscCall(MatSetSizes(C, m, n, m, n)); 4293 PetscCall(MatSetBlockSizesFromMats(C, A, B)); 4294 PetscCall(PetscObjectTypeCompareAny((PetscObject)C, &cisdense, MATSEQDENSE, MATSEQDENSECUDA, MATSEQDENSEHIP, "")); 4295 if (!cisdense) PetscCall(MatSetType(C, MATDENSE)); 4296 PetscCall(MatSetUp(C)); 4297 4298 C->ops->matmultnumeric = MatMatMultNumeric_SeqDense_SeqAIJ; 4299 PetscFunctionReturn(PETSC_SUCCESS); 4300 } 4301 4302 /*MC 4303 MATSEQAIJ - MATSEQAIJ = "seqaij" - A matrix type to be used for sequential sparse matrices, 4304 based on compressed sparse row format. 4305 4306 Options Database Key: 4307 . -mat_type seqaij - sets the matrix type to "seqaij" during a call to MatSetFromOptions() 4308 4309 Level: beginner 4310 4311 Notes: 4312 `MatSetValues()` may be called for this matrix type with a `NULL` argument for the numerical values, 4313 in this case the values associated with the rows and columns one passes in are set to zero 4314 in the matrix 4315 4316 `MatSetOptions`(,`MAT_STRUCTURE_ONLY`,`PETSC_TRUE`) may be called for this matrix type. In this no 4317 space is allocated for the nonzero entries and any entries passed with `MatSetValues()` are ignored 4318 4319 Developer Note: 4320 It would be nice if all matrix formats supported passing `NULL` in for the numerical values 4321 4322 .seealso: [](ch_matrices), `Mat`, `MatCreateSeqAIJ()`, `MatSetFromOptions()`, `MatSetType()`, `MatCreate()`, `MatType`, `MATSELL`, `MATSEQSELL`, `MATMPISELL` 4323 M*/ 4324 4325 /*MC 4326 MATAIJ - MATAIJ = "aij" - A matrix type to be used for sparse matrices. 4327 4328 This matrix type is identical to `MATSEQAIJ` when constructed with a single process communicator, 4329 and `MATMPIAIJ` otherwise. As a result, for single process communicators, 4330 `MatSeqAIJSetPreallocation()` is supported, and similarly `MatMPIAIJSetPreallocation()` is supported 4331 for communicators controlling multiple processes. It is recommended that you call both of 4332 the above preallocation routines for simplicity. 4333 4334 Options Database Key: 4335 . -mat_type aij - sets the matrix type to "aij" during a call to `MatSetFromOptions()` 4336 4337 Level: beginner 4338 4339 Note: 4340 Subclasses include `MATAIJCUSPARSE`, `MATAIJPERM`, `MATAIJSELL`, `MATAIJMKL`, `MATAIJCRL`, and also automatically switches over to use inodes when 4341 enough exist. 4342 4343 .seealso: [](ch_matrices), `Mat`, `MatCreateAIJ()`, `MatCreateSeqAIJ()`, `MATSEQAIJ`, `MATMPIAIJ`, `MATSELL`, `MATSEQSELL`, `MATMPISELL` 4344 M*/ 4345 4346 /*MC 4347 MATAIJCRL - MATAIJCRL = "aijcrl" - A matrix type to be used for sparse matrices. 4348 4349 Options Database Key: 4350 . -mat_type aijcrl - sets the matrix type to "aijcrl" during a call to `MatSetFromOptions()` 4351 4352 Level: beginner 4353 4354 Note: 4355 This matrix type is identical to `MATSEQAIJCRL` when constructed with a single process communicator, 4356 and `MATMPIAIJCRL` otherwise. As a result, for single process communicators, 4357 `MatSeqAIJSetPreallocation()` is supported, and similarly `MatMPIAIJSetPreallocation()` is supported 4358 for communicators controlling multiple processes. It is recommended that you call both of 4359 the above preallocation routines for simplicity. 4360 4361 .seealso: [](ch_matrices), `Mat`, `MatCreateMPIAIJCRL`, `MATSEQAIJCRL`, `MATMPIAIJCRL`, `MATSEQAIJCRL`, `MATMPIAIJCRL` 4362 M*/ 4363 4364 PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqAIJCRL(Mat, MatType, MatReuse, Mat *); 4365 #if defined(PETSC_HAVE_ELEMENTAL) 4366 PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_Elemental(Mat, MatType, MatReuse, Mat *); 4367 #endif 4368 #if defined(PETSC_HAVE_SCALAPACK) 4369 PETSC_INTERN PetscErrorCode MatConvert_AIJ_ScaLAPACK(Mat, MatType, MatReuse, Mat *); 4370 #endif 4371 #if defined(PETSC_HAVE_HYPRE) 4372 PETSC_INTERN PetscErrorCode MatConvert_AIJ_HYPRE(Mat A, MatType, MatReuse, Mat *); 4373 #endif 4374 4375 PETSC_EXTERN PetscErrorCode MatConvert_SeqAIJ_SeqSELL(Mat, MatType, MatReuse, Mat *); 4376 PETSC_INTERN PetscErrorCode MatConvert_XAIJ_IS(Mat, MatType, MatReuse, Mat *); 4377 PETSC_INTERN PetscErrorCode MatProductSetFromOptions_IS_XAIJ(Mat); 4378 4379 /*@C 4380 MatSeqAIJGetArray - gives read/write access to the array where the data for a `MATSEQAIJ` matrix is stored 4381 4382 Not Collective 4383 4384 Input Parameter: 4385 . A - a `MATSEQAIJ` matrix 4386 4387 Output Parameter: 4388 . array - pointer to the data 4389 4390 Level: intermediate 4391 4392 Fortran Notes: 4393 `MatSeqAIJGetArray()` Fortran binding is deprecated (since PETSc 3.19), use `MatSeqAIJGetArrayF90()` 4394 4395 .seealso: [](ch_matrices), `Mat`, `MatSeqAIJRestoreArray()`, `MatSeqAIJGetArrayF90()` 4396 @*/ 4397 PetscErrorCode MatSeqAIJGetArray(Mat A, PetscScalar **array) 4398 { 4399 Mat_SeqAIJ *aij = (Mat_SeqAIJ *)A->data; 4400 4401 PetscFunctionBegin; 4402 if (aij->ops->getarray) { 4403 PetscCall((*aij->ops->getarray)(A, array)); 4404 } else { 4405 *array = aij->a; 4406 } 4407 PetscFunctionReturn(PETSC_SUCCESS); 4408 } 4409 4410 /*@C 4411 MatSeqAIJRestoreArray - returns access to the array where the data for a `MATSEQAIJ` matrix is stored obtained by `MatSeqAIJGetArray()` 4412 4413 Not Collective 4414 4415 Input Parameters: 4416 + A - a `MATSEQAIJ` matrix 4417 - array - pointer to the data 4418 4419 Level: intermediate 4420 4421 Fortran Notes: 4422 `MatSeqAIJRestoreArray()` Fortran binding is deprecated (since PETSc 3.19), use `MatSeqAIJRestoreArrayF90()` 4423 4424 .seealso: [](ch_matrices), `Mat`, `MatSeqAIJGetArray()`, `MatSeqAIJRestoreArrayF90()` 4425 @*/ 4426 PetscErrorCode MatSeqAIJRestoreArray(Mat A, PetscScalar **array) 4427 { 4428 Mat_SeqAIJ *aij = (Mat_SeqAIJ *)A->data; 4429 4430 PetscFunctionBegin; 4431 if (aij->ops->restorearray) { 4432 PetscCall((*aij->ops->restorearray)(A, array)); 4433 } else { 4434 *array = NULL; 4435 } 4436 PetscCall(MatSeqAIJInvalidateDiagonal(A)); 4437 PetscCall(PetscObjectStateIncrease((PetscObject)A)); 4438 PetscFunctionReturn(PETSC_SUCCESS); 4439 } 4440 4441 /*@C 4442 MatSeqAIJGetArrayRead - gives read-only access to the array where the data for a `MATSEQAIJ` matrix is stored 4443 4444 Not Collective; No Fortran Support 4445 4446 Input Parameter: 4447 . A - a `MATSEQAIJ` matrix 4448 4449 Output Parameter: 4450 . array - pointer to the data 4451 4452 Level: intermediate 4453 4454 .seealso: [](ch_matrices), `Mat`, `MatSeqAIJGetArray()`, `MatSeqAIJRestoreArrayRead()` 4455 @*/ 4456 PetscErrorCode MatSeqAIJGetArrayRead(Mat A, const PetscScalar **array) 4457 { 4458 Mat_SeqAIJ *aij = (Mat_SeqAIJ *)A->data; 4459 4460 PetscFunctionBegin; 4461 if (aij->ops->getarrayread) { 4462 PetscCall((*aij->ops->getarrayread)(A, array)); 4463 } else { 4464 *array = aij->a; 4465 } 4466 PetscFunctionReturn(PETSC_SUCCESS); 4467 } 4468 4469 /*@C 4470 MatSeqAIJRestoreArrayRead - restore the read-only access array obtained from `MatSeqAIJGetArrayRead()` 4471 4472 Not Collective; No Fortran Support 4473 4474 Input Parameter: 4475 . A - a `MATSEQAIJ` matrix 4476 4477 Output Parameter: 4478 . array - pointer to the data 4479 4480 Level: intermediate 4481 4482 .seealso: [](ch_matrices), `Mat`, `MatSeqAIJGetArray()`, `MatSeqAIJGetArrayRead()` 4483 @*/ 4484 PetscErrorCode MatSeqAIJRestoreArrayRead(Mat A, const PetscScalar **array) 4485 { 4486 Mat_SeqAIJ *aij = (Mat_SeqAIJ *)A->data; 4487 4488 PetscFunctionBegin; 4489 if (aij->ops->restorearrayread) { 4490 PetscCall((*aij->ops->restorearrayread)(A, array)); 4491 } else { 4492 *array = NULL; 4493 } 4494 PetscFunctionReturn(PETSC_SUCCESS); 4495 } 4496 4497 /*@C 4498 MatSeqAIJGetArrayWrite - gives write-only access to the array where the data for a `MATSEQAIJ` matrix is stored 4499 4500 Not Collective; No Fortran Support 4501 4502 Input Parameter: 4503 . A - a `MATSEQAIJ` matrix 4504 4505 Output Parameter: 4506 . array - pointer to the data 4507 4508 Level: intermediate 4509 4510 .seealso: [](ch_matrices), `Mat`, `MatSeqAIJGetArray()`, `MatSeqAIJRestoreArrayRead()` 4511 @*/ 4512 PetscErrorCode MatSeqAIJGetArrayWrite(Mat A, PetscScalar **array) 4513 { 4514 Mat_SeqAIJ *aij = (Mat_SeqAIJ *)A->data; 4515 4516 PetscFunctionBegin; 4517 if (aij->ops->getarraywrite) { 4518 PetscCall((*aij->ops->getarraywrite)(A, array)); 4519 } else { 4520 *array = aij->a; 4521 } 4522 PetscCall(MatSeqAIJInvalidateDiagonal(A)); 4523 PetscCall(PetscObjectStateIncrease((PetscObject)A)); 4524 PetscFunctionReturn(PETSC_SUCCESS); 4525 } 4526 4527 /*@C 4528 MatSeqAIJRestoreArrayWrite - restore the read-only access array obtained from MatSeqAIJGetArrayRead 4529 4530 Not Collective; No Fortran Support 4531 4532 Input Parameter: 4533 . A - a MATSEQAIJ matrix 4534 4535 Output Parameter: 4536 . array - pointer to the data 4537 4538 Level: intermediate 4539 4540 .seealso: [](ch_matrices), `Mat`, `MatSeqAIJGetArray()`, `MatSeqAIJGetArrayRead()` 4541 @*/ 4542 PetscErrorCode MatSeqAIJRestoreArrayWrite(Mat A, PetscScalar **array) 4543 { 4544 Mat_SeqAIJ *aij = (Mat_SeqAIJ *)A->data; 4545 4546 PetscFunctionBegin; 4547 if (aij->ops->restorearraywrite) { 4548 PetscCall((*aij->ops->restorearraywrite)(A, array)); 4549 } else { 4550 *array = NULL; 4551 } 4552 PetscFunctionReturn(PETSC_SUCCESS); 4553 } 4554 4555 /*@C 4556 MatSeqAIJGetCSRAndMemType - Get the CSR arrays and the memory type of the `MATSEQAIJ` matrix 4557 4558 Not Collective; No Fortran Support 4559 4560 Input Parameter: 4561 . mat - a matrix of type `MATSEQAIJ` or its subclasses 4562 4563 Output Parameters: 4564 + i - row map array of the matrix 4565 . j - column index array of the matrix 4566 . a - data array of the matrix 4567 - mtype - memory type of the arrays 4568 4569 Level: developer 4570 4571 Notes: 4572 Any of the output parameters can be `NULL`, in which case the corresponding value is not returned. 4573 If mat is a device matrix, the arrays are on the device. Otherwise, they are on the host. 4574 4575 One can call this routine on a preallocated but not assembled matrix to just get the memory of the CSR underneath the matrix. 4576 If the matrix is assembled, the data array `a` is guaranteed to have the latest values of the matrix. 4577 4578 .seealso: [](ch_matrices), `Mat`, `MatSeqAIJGetArray()`, `MatSeqAIJGetArrayRead()` 4579 @*/ 4580 PetscErrorCode MatSeqAIJGetCSRAndMemType(Mat mat, const PetscInt **i, const PetscInt **j, PetscScalar **a, PetscMemType *mtype) 4581 { 4582 Mat_SeqAIJ *aij = (Mat_SeqAIJ *)mat->data; 4583 4584 PetscFunctionBegin; 4585 PetscCheck(mat->preallocated, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "matrix is not preallocated"); 4586 if (aij->ops->getcsrandmemtype) { 4587 PetscCall((*aij->ops->getcsrandmemtype)(mat, i, j, a, mtype)); 4588 } else { 4589 if (i) *i = aij->i; 4590 if (j) *j = aij->j; 4591 if (a) *a = aij->a; 4592 if (mtype) *mtype = PETSC_MEMTYPE_HOST; 4593 } 4594 PetscFunctionReturn(PETSC_SUCCESS); 4595 } 4596 4597 /*@C 4598 MatSeqAIJGetMaxRowNonzeros - returns the maximum number of nonzeros in any row 4599 4600 Not Collective 4601 4602 Input Parameter: 4603 . A - a `MATSEQAIJ` matrix 4604 4605 Output Parameter: 4606 . nz - the maximum number of nonzeros in any row 4607 4608 Level: intermediate 4609 4610 .seealso: [](ch_matrices), `Mat`, `MatSeqAIJRestoreArray()`, `MatSeqAIJGetArrayF90()` 4611 @*/ 4612 PetscErrorCode MatSeqAIJGetMaxRowNonzeros(Mat A, PetscInt *nz) 4613 { 4614 Mat_SeqAIJ *aij = (Mat_SeqAIJ *)A->data; 4615 4616 PetscFunctionBegin; 4617 *nz = aij->rmax; 4618 PetscFunctionReturn(PETSC_SUCCESS); 4619 } 4620 4621 static PetscErrorCode MatCOOStructDestroy_SeqAIJ(void *data) 4622 { 4623 MatCOOStruct_SeqAIJ *coo = (MatCOOStruct_SeqAIJ *)data; 4624 PetscFunctionBegin; 4625 PetscCall(PetscFree(coo->perm)); 4626 PetscCall(PetscFree(coo->jmap)); 4627 PetscCall(PetscFree(coo)); 4628 PetscFunctionReturn(PETSC_SUCCESS); 4629 } 4630 4631 PetscErrorCode MatSetPreallocationCOO_SeqAIJ(Mat mat, PetscCount coo_n, PetscInt coo_i[], PetscInt coo_j[]) 4632 { 4633 MPI_Comm comm; 4634 PetscInt *i, *j; 4635 PetscInt M, N, row; 4636 PetscCount k, p, q, nneg, nnz, start, end; /* Index the coo array, so use PetscCount as their type */ 4637 PetscInt *Ai; /* Change to PetscCount once we use it for row pointers */ 4638 PetscInt *Aj; 4639 PetscScalar *Aa; 4640 Mat_SeqAIJ *seqaij = (Mat_SeqAIJ *)(mat->data); 4641 MatType rtype; 4642 PetscCount *perm, *jmap; 4643 PetscContainer container; 4644 MatCOOStruct_SeqAIJ *coo; 4645 4646 PetscFunctionBegin; 4647 PetscCall(PetscObjectGetComm((PetscObject)mat, &comm)); 4648 PetscCall(MatGetSize(mat, &M, &N)); 4649 i = coo_i; 4650 j = coo_j; 4651 PetscCall(PetscMalloc1(coo_n, &perm)); 4652 for (k = 0; k < coo_n; k++) { /* Ignore entries with negative row or col indices */ 4653 if (j[k] < 0) i[k] = -1; 4654 perm[k] = k; 4655 } 4656 4657 /* Sort by row */ 4658 PetscCall(PetscSortIntWithIntCountArrayPair(coo_n, i, j, perm)); 4659 4660 /* Advance k to the first row with a non-negative index */ 4661 for (k = 0; k < coo_n; k++) 4662 if (i[k] >= 0) break; 4663 nneg = k; 4664 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 */ 4665 nnz = 0; /* Total number of unique nonzeros to be counted */ 4666 jmap++; /* Inc jmap by 1 for convenience */ 4667 4668 PetscCall(PetscCalloc1(M + 1, &Ai)); /* CSR of A */ 4669 PetscCall(PetscMalloc1(coo_n - nneg, &Aj)); /* We have at most coo_n-nneg unique nonzeros */ 4670 4671 /* Support for HYPRE */ 4672 PetscBool hypre; 4673 const char *name; 4674 PetscCall(PetscObjectGetName((PetscObject)mat, &name)); 4675 PetscCall(PetscStrcmp("_internal_COO_mat_for_hypre", name, &hypre)); 4676 4677 /* In each row, sort by column, then unique column indices to get row length */ 4678 Ai++; /* Inc by 1 for convenience */ 4679 q = 0; /* q-th unique nonzero, with q starting from 0 */ 4680 while (k < coo_n) { 4681 row = i[k]; 4682 start = k; /* [start,end) indices for this row */ 4683 while (k < coo_n && i[k] == row) k++; 4684 end = k; 4685 /* hack for HYPRE: swap min column to diag so that diagonal values will go first */ 4686 if (hypre) { 4687 PetscInt minj = PETSC_MAX_INT; 4688 PetscBool hasdiag = PETSC_FALSE; 4689 for (p = start; p < end; p++) { 4690 hasdiag = (PetscBool)(hasdiag || (j[p] == row)); 4691 minj = PetscMin(minj, j[p]); 4692 } 4693 if (hasdiag) { 4694 for (p = start; p < end; p++) { 4695 if (j[p] == minj) j[p] = row; 4696 else if (j[p] == row) j[p] = minj; 4697 } 4698 } 4699 } 4700 PetscCall(PetscSortIntWithCountArray(end - start, j + start, perm + start)); 4701 4702 /* Find number of unique col entries in this row */ 4703 Aj[q] = j[start]; /* Log the first nonzero in this row */ 4704 jmap[q] = 1; /* Number of repeats of this nonzero entry */ 4705 Ai[row] = 1; 4706 nnz++; 4707 4708 for (p = start + 1; p < end; p++) { /* Scan remaining nonzero in this row */ 4709 if (j[p] != j[p - 1]) { /* Meet a new nonzero */ 4710 q++; 4711 jmap[q] = 1; 4712 Aj[q] = j[p]; 4713 Ai[row]++; 4714 nnz++; 4715 } else { 4716 jmap[q]++; 4717 } 4718 } 4719 q++; /* Move to next row and thus next unique nonzero */ 4720 } 4721 Ai--; /* Back to the beginning of Ai[] */ 4722 for (k = 0; k < M; k++) Ai[k + 1] += Ai[k]; 4723 jmap--; /* Back to the beginning of jmap[] */ 4724 jmap[0] = 0; 4725 for (k = 0; k < nnz; k++) jmap[k + 1] += jmap[k]; 4726 if (nnz < coo_n - nneg) { /* Realloc with actual number of unique nonzeros */ 4727 PetscCount *jmap_new; 4728 PetscInt *Aj_new; 4729 4730 PetscCall(PetscMalloc1(nnz + 1, &jmap_new)); 4731 PetscCall(PetscArraycpy(jmap_new, jmap, nnz + 1)); 4732 PetscCall(PetscFree(jmap)); 4733 jmap = jmap_new; 4734 4735 PetscCall(PetscMalloc1(nnz, &Aj_new)); 4736 PetscCall(PetscArraycpy(Aj_new, Aj, nnz)); 4737 PetscCall(PetscFree(Aj)); 4738 Aj = Aj_new; 4739 } 4740 4741 if (nneg) { /* Discard heading entries with negative indices in perm[], as we'll access it from index 0 in MatSetValuesCOO */ 4742 PetscCount *perm_new; 4743 4744 PetscCall(PetscMalloc1(coo_n - nneg, &perm_new)); 4745 PetscCall(PetscArraycpy(perm_new, perm + nneg, coo_n - nneg)); 4746 PetscCall(PetscFree(perm)); 4747 perm = perm_new; 4748 } 4749 4750 PetscCall(MatGetRootType_Private(mat, &rtype)); 4751 PetscCall(PetscCalloc1(nnz, &Aa)); /* Zero the matrix */ 4752 PetscCall(MatSetSeqAIJWithArrays_private(PETSC_COMM_SELF, M, N, Ai, Aj, Aa, rtype, mat)); 4753 4754 seqaij->singlemalloc = PETSC_FALSE; /* Ai, Aj and Aa are not allocated in one big malloc */ 4755 seqaij->free_a = seqaij->free_ij = PETSC_TRUE; /* Let newmat own Ai, Aj and Aa */ 4756 4757 // Put the COO struct in a container and then attach that to the matrix 4758 PetscCall(PetscMalloc1(1, &coo)); 4759 coo->nz = nnz; 4760 coo->n = coo_n; 4761 coo->Atot = coo_n - nneg; // Annz is seqaij->nz, so no need to record that again 4762 coo->jmap = jmap; // of length nnz+1 4763 coo->perm = perm; 4764 PetscCall(PetscContainerCreate(PETSC_COMM_SELF, &container)); 4765 PetscCall(PetscContainerSetPointer(container, coo)); 4766 PetscCall(PetscContainerSetUserDestroy(container, MatCOOStructDestroy_SeqAIJ)); 4767 PetscCall(PetscObjectCompose((PetscObject)mat, "__PETSc_MatCOOStruct_Host", (PetscObject)container)); 4768 PetscCall(PetscContainerDestroy(&container)); 4769 PetscFunctionReturn(PETSC_SUCCESS); 4770 } 4771 4772 static PetscErrorCode MatSetValuesCOO_SeqAIJ(Mat A, const PetscScalar v[], InsertMode imode) 4773 { 4774 Mat_SeqAIJ *aseq = (Mat_SeqAIJ *)A->data; 4775 PetscCount i, j, Annz = aseq->nz; 4776 PetscCount *perm, *jmap; 4777 PetscScalar *Aa; 4778 PetscContainer container; 4779 MatCOOStruct_SeqAIJ *coo; 4780 4781 PetscFunctionBegin; 4782 PetscCall(PetscObjectQuery((PetscObject)A, "__PETSc_MatCOOStruct_Host", (PetscObject *)&container)); 4783 PetscCheck(container, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Not found MatCOOStruct on this matrix"); 4784 PetscCall(PetscContainerGetPointer(container, (void **)&coo)); 4785 perm = coo->perm; 4786 jmap = coo->jmap; 4787 PetscCall(MatSeqAIJGetArray(A, &Aa)); 4788 for (i = 0; i < Annz; i++) { 4789 PetscScalar sum = 0.0; 4790 for (j = jmap[i]; j < jmap[i + 1]; j++) sum += v[perm[j]]; 4791 Aa[i] = (imode == INSERT_VALUES ? 0.0 : Aa[i]) + sum; 4792 } 4793 PetscCall(MatSeqAIJRestoreArray(A, &Aa)); 4794 PetscFunctionReturn(PETSC_SUCCESS); 4795 } 4796 4797 #if defined(PETSC_HAVE_CUDA) 4798 PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqAIJCUSPARSE(Mat, MatType, MatReuse, Mat *); 4799 #endif 4800 #if defined(PETSC_HAVE_HIP) 4801 PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqAIJHIPSPARSE(Mat, MatType, MatReuse, Mat *); 4802 #endif 4803 #if defined(PETSC_HAVE_KOKKOS_KERNELS) 4804 PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqAIJKokkos(Mat, MatType, MatReuse, Mat *); 4805 #endif 4806 4807 PETSC_EXTERN PetscErrorCode MatCreate_SeqAIJ(Mat B) 4808 { 4809 Mat_SeqAIJ *b; 4810 PetscMPIInt size; 4811 4812 PetscFunctionBegin; 4813 PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)B), &size)); 4814 PetscCheck(size <= 1, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Comm must be of size 1"); 4815 4816 PetscCall(PetscNew(&b)); 4817 4818 B->data = (void *)b; 4819 B->ops[0] = MatOps_Values; 4820 if (B->sortedfull) B->ops->setvalues = MatSetValues_SeqAIJ_SortedFull; 4821 4822 b->row = NULL; 4823 b->col = NULL; 4824 b->icol = NULL; 4825 b->reallocs = 0; 4826 b->ignorezeroentries = PETSC_FALSE; 4827 b->roworiented = PETSC_TRUE; 4828 b->nonew = 0; 4829 b->diag = NULL; 4830 b->solve_work = NULL; 4831 B->spptr = NULL; 4832 b->saved_values = NULL; 4833 b->idiag = NULL; 4834 b->mdiag = NULL; 4835 b->ssor_work = NULL; 4836 b->omega = 1.0; 4837 b->fshift = 0.0; 4838 b->idiagvalid = PETSC_FALSE; 4839 b->ibdiagvalid = PETSC_FALSE; 4840 b->keepnonzeropattern = PETSC_FALSE; 4841 4842 PetscCall(PetscObjectChangeTypeName((PetscObject)B, MATSEQAIJ)); 4843 #if defined(PETSC_HAVE_MATLAB) 4844 PetscCall(PetscObjectComposeFunction((PetscObject)B, "PetscMatlabEnginePut_C", MatlabEnginePut_SeqAIJ)); 4845 PetscCall(PetscObjectComposeFunction((PetscObject)B, "PetscMatlabEngineGet_C", MatlabEngineGet_SeqAIJ)); 4846 #endif 4847 PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatSeqAIJSetColumnIndices_C", MatSeqAIJSetColumnIndices_SeqAIJ)); 4848 PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatStoreValues_C", MatStoreValues_SeqAIJ)); 4849 PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatRetrieveValues_C", MatRetrieveValues_SeqAIJ)); 4850 PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_seqsbaij_C", MatConvert_SeqAIJ_SeqSBAIJ)); 4851 PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_seqbaij_C", MatConvert_SeqAIJ_SeqBAIJ)); 4852 PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_seqaijperm_C", MatConvert_SeqAIJ_SeqAIJPERM)); 4853 PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_seqaijsell_C", MatConvert_SeqAIJ_SeqAIJSELL)); 4854 #if defined(PETSC_HAVE_MKL_SPARSE) 4855 PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_seqaijmkl_C", MatConvert_SeqAIJ_SeqAIJMKL)); 4856 #endif 4857 #if defined(PETSC_HAVE_CUDA) 4858 PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_seqaijcusparse_C", MatConvert_SeqAIJ_SeqAIJCUSPARSE)); 4859 PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_seqaijcusparse_seqaij_C", MatProductSetFromOptions_SeqAIJ)); 4860 PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_seqaij_seqaijcusparse_C", MatProductSetFromOptions_SeqAIJ)); 4861 #endif 4862 #if defined(PETSC_HAVE_HIP) 4863 PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_seqaijhipsparse_C", MatConvert_SeqAIJ_SeqAIJHIPSPARSE)); 4864 PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_seqaijhipsparse_seqaij_C", MatProductSetFromOptions_SeqAIJ)); 4865 PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_seqaij_seqaijhipsparse_C", MatProductSetFromOptions_SeqAIJ)); 4866 #endif 4867 #if defined(PETSC_HAVE_KOKKOS_KERNELS) 4868 PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_seqaijkokkos_C", MatConvert_SeqAIJ_SeqAIJKokkos)); 4869 #endif 4870 PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_seqaijcrl_C", MatConvert_SeqAIJ_SeqAIJCRL)); 4871 #if defined(PETSC_HAVE_ELEMENTAL) 4872 PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_elemental_C", MatConvert_SeqAIJ_Elemental)); 4873 #endif 4874 #if defined(PETSC_HAVE_SCALAPACK) 4875 PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_scalapack_C", MatConvert_AIJ_ScaLAPACK)); 4876 #endif 4877 #if defined(PETSC_HAVE_HYPRE) 4878 PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_hypre_C", MatConvert_AIJ_HYPRE)); 4879 PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_transpose_seqaij_seqaij_C", MatProductSetFromOptions_Transpose_AIJ_AIJ)); 4880 #endif 4881 PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_seqdense_C", MatConvert_SeqAIJ_SeqDense)); 4882 PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_seqsell_C", MatConvert_SeqAIJ_SeqSELL)); 4883 PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_is_C", MatConvert_XAIJ_IS)); 4884 PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatIsTranspose_C", MatIsTranspose_SeqAIJ)); 4885 PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatIsHermitianTranspose_C", MatIsTranspose_SeqAIJ)); 4886 PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatSeqAIJSetPreallocation_C", MatSeqAIJSetPreallocation_SeqAIJ)); 4887 PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatResetPreallocation_C", MatResetPreallocation_SeqAIJ)); 4888 PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatSeqAIJSetPreallocationCSR_C", MatSeqAIJSetPreallocationCSR_SeqAIJ)); 4889 PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatReorderForNonzeroDiagonal_C", MatReorderForNonzeroDiagonal_SeqAIJ)); 4890 PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_is_seqaij_C", MatProductSetFromOptions_IS_XAIJ)); 4891 PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_seqdense_seqaij_C", MatProductSetFromOptions_SeqDense_SeqAIJ)); 4892 PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_seqaij_seqaij_C", MatProductSetFromOptions_SeqAIJ)); 4893 PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatSeqAIJKron_C", MatSeqAIJKron_SeqAIJ)); 4894 PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatSetPreallocationCOO_C", MatSetPreallocationCOO_SeqAIJ)); 4895 PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatSetValuesCOO_C", MatSetValuesCOO_SeqAIJ)); 4896 PetscCall(MatCreate_SeqAIJ_Inode(B)); 4897 PetscCall(PetscObjectChangeTypeName((PetscObject)B, MATSEQAIJ)); 4898 PetscCall(MatSeqAIJSetTypeFromOptions(B)); /* this allows changing the matrix subtype to say MATSEQAIJPERM */ 4899 PetscFunctionReturn(PETSC_SUCCESS); 4900 } 4901 4902 /* 4903 Given a matrix generated with MatGetFactor() duplicates all the information in A into C 4904 */ 4905 PetscErrorCode MatDuplicateNoCreate_SeqAIJ(Mat C, Mat A, MatDuplicateOption cpvalues, PetscBool mallocmatspace) 4906 { 4907 Mat_SeqAIJ *c = (Mat_SeqAIJ *)C->data, *a = (Mat_SeqAIJ *)A->data; 4908 PetscInt m = A->rmap->n, i; 4909 4910 PetscFunctionBegin; 4911 PetscCheck(A->assembled || cpvalues == MAT_DO_NOT_COPY_VALUES, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Cannot duplicate unassembled matrix"); 4912 4913 C->factortype = A->factortype; 4914 c->row = NULL; 4915 c->col = NULL; 4916 c->icol = NULL; 4917 c->reallocs = 0; 4918 c->diagonaldense = a->diagonaldense; 4919 4920 C->assembled = A->assembled; 4921 4922 if (A->preallocated) { 4923 PetscCall(PetscLayoutReference(A->rmap, &C->rmap)); 4924 PetscCall(PetscLayoutReference(A->cmap, &C->cmap)); 4925 4926 if (!A->hash_active) { 4927 PetscCall(PetscMalloc1(m, &c->imax)); 4928 PetscCall(PetscMemcpy(c->imax, a->imax, m * sizeof(PetscInt))); 4929 PetscCall(PetscMalloc1(m, &c->ilen)); 4930 PetscCall(PetscMemcpy(c->ilen, a->ilen, m * sizeof(PetscInt))); 4931 4932 /* allocate the matrix space */ 4933 if (mallocmatspace) { 4934 PetscCall(PetscMalloc3(a->i[m], &c->a, a->i[m], &c->j, m + 1, &c->i)); 4935 4936 c->singlemalloc = PETSC_TRUE; 4937 4938 PetscCall(PetscArraycpy(c->i, a->i, m + 1)); 4939 if (m > 0) { 4940 PetscCall(PetscArraycpy(c->j, a->j, a->i[m])); 4941 if (cpvalues == MAT_COPY_VALUES) { 4942 const PetscScalar *aa; 4943 4944 PetscCall(MatSeqAIJGetArrayRead(A, &aa)); 4945 PetscCall(PetscArraycpy(c->a, aa, a->i[m])); 4946 PetscCall(MatSeqAIJGetArrayRead(A, &aa)); 4947 } else { 4948 PetscCall(PetscArrayzero(c->a, a->i[m])); 4949 } 4950 } 4951 } 4952 C->preallocated = PETSC_TRUE; 4953 } else { 4954 PetscCheck(mallocmatspace, PetscObjectComm((PetscObject)A), PETSC_ERR_ARG_WRONGSTATE, "Cannot malloc matrix memory from a non-preallocated matrix"); 4955 PetscCall(MatSetUp(C)); 4956 } 4957 4958 c->ignorezeroentries = a->ignorezeroentries; 4959 c->roworiented = a->roworiented; 4960 c->nonew = a->nonew; 4961 if (a->diag) { 4962 PetscCall(PetscMalloc1(m + 1, &c->diag)); 4963 PetscCall(PetscMemcpy(c->diag, a->diag, m * sizeof(PetscInt))); 4964 } else c->diag = NULL; 4965 4966 c->solve_work = NULL; 4967 c->saved_values = NULL; 4968 c->idiag = NULL; 4969 c->ssor_work = NULL; 4970 c->keepnonzeropattern = a->keepnonzeropattern; 4971 c->free_a = PETSC_TRUE; 4972 c->free_ij = PETSC_TRUE; 4973 4974 c->rmax = a->rmax; 4975 c->nz = a->nz; 4976 c->maxnz = a->nz; /* Since we allocate exactly the right amount */ 4977 4978 c->compressedrow.use = a->compressedrow.use; 4979 c->compressedrow.nrows = a->compressedrow.nrows; 4980 if (a->compressedrow.use) { 4981 i = a->compressedrow.nrows; 4982 PetscCall(PetscMalloc2(i + 1, &c->compressedrow.i, i, &c->compressedrow.rindex)); 4983 PetscCall(PetscArraycpy(c->compressedrow.i, a->compressedrow.i, i + 1)); 4984 PetscCall(PetscArraycpy(c->compressedrow.rindex, a->compressedrow.rindex, i)); 4985 } else { 4986 c->compressedrow.use = PETSC_FALSE; 4987 c->compressedrow.i = NULL; 4988 c->compressedrow.rindex = NULL; 4989 } 4990 c->nonzerorowcnt = a->nonzerorowcnt; 4991 C->nonzerostate = A->nonzerostate; 4992 4993 PetscCall(MatDuplicate_SeqAIJ_Inode(A, cpvalues, &C)); 4994 } 4995 PetscCall(PetscFunctionListDuplicate(((PetscObject)A)->qlist, &((PetscObject)C)->qlist)); 4996 PetscFunctionReturn(PETSC_SUCCESS); 4997 } 4998 4999 PetscErrorCode MatDuplicate_SeqAIJ(Mat A, MatDuplicateOption cpvalues, Mat *B) 5000 { 5001 PetscFunctionBegin; 5002 PetscCall(MatCreate(PetscObjectComm((PetscObject)A), B)); 5003 PetscCall(MatSetSizes(*B, A->rmap->n, A->cmap->n, A->rmap->n, A->cmap->n)); 5004 if (!(A->rmap->n % A->rmap->bs) && !(A->cmap->n % A->cmap->bs)) PetscCall(MatSetBlockSizesFromMats(*B, A, A)); 5005 PetscCall(MatSetType(*B, ((PetscObject)A)->type_name)); 5006 PetscCall(MatDuplicateNoCreate_SeqAIJ(*B, A, cpvalues, PETSC_TRUE)); 5007 PetscFunctionReturn(PETSC_SUCCESS); 5008 } 5009 5010 PetscErrorCode MatLoad_SeqAIJ(Mat newMat, PetscViewer viewer) 5011 { 5012 PetscBool isbinary, ishdf5; 5013 5014 PetscFunctionBegin; 5015 PetscValidHeaderSpecific(newMat, MAT_CLASSID, 1); 5016 PetscValidHeaderSpecific(viewer, PETSC_VIEWER_CLASSID, 2); 5017 /* force binary viewer to load .info file if it has not yet done so */ 5018 PetscCall(PetscViewerSetUp(viewer)); 5019 PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERBINARY, &isbinary)); 5020 PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERHDF5, &ishdf5)); 5021 if (isbinary) { 5022 PetscCall(MatLoad_SeqAIJ_Binary(newMat, viewer)); 5023 } else if (ishdf5) { 5024 #if defined(PETSC_HAVE_HDF5) 5025 PetscCall(MatLoad_AIJ_HDF5(newMat, viewer)); 5026 #else 5027 SETERRQ(PetscObjectComm((PetscObject)newMat), PETSC_ERR_SUP, "HDF5 not supported in this build.\nPlease reconfigure using --download-hdf5"); 5028 #endif 5029 } else { 5030 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); 5031 } 5032 PetscFunctionReturn(PETSC_SUCCESS); 5033 } 5034 5035 PetscErrorCode MatLoad_SeqAIJ_Binary(Mat mat, PetscViewer viewer) 5036 { 5037 Mat_SeqAIJ *a = (Mat_SeqAIJ *)mat->data; 5038 PetscInt header[4], *rowlens, M, N, nz, sum, rows, cols, i; 5039 5040 PetscFunctionBegin; 5041 PetscCall(PetscViewerSetUp(viewer)); 5042 5043 /* read in matrix header */ 5044 PetscCall(PetscViewerBinaryRead(viewer, header, 4, NULL, PETSC_INT)); 5045 PetscCheck(header[0] == MAT_FILE_CLASSID, PETSC_COMM_SELF, PETSC_ERR_FILE_UNEXPECTED, "Not a matrix object in file"); 5046 M = header[1]; 5047 N = header[2]; 5048 nz = header[3]; 5049 PetscCheck(M >= 0, PetscObjectComm((PetscObject)viewer), PETSC_ERR_FILE_UNEXPECTED, "Matrix row size (%" PetscInt_FMT ") in file is negative", M); 5050 PetscCheck(N >= 0, PetscObjectComm((PetscObject)viewer), PETSC_ERR_FILE_UNEXPECTED, "Matrix column size (%" PetscInt_FMT ") in file is negative", N); 5051 PetscCheck(nz >= 0, PETSC_COMM_SELF, PETSC_ERR_FILE_UNEXPECTED, "Matrix stored in special format on disk, cannot load as SeqAIJ"); 5052 5053 /* set block sizes from the viewer's .info file */ 5054 PetscCall(MatLoad_Binary_BlockSizes(mat, viewer)); 5055 /* set local and global sizes if not set already */ 5056 if (mat->rmap->n < 0) mat->rmap->n = M; 5057 if (mat->cmap->n < 0) mat->cmap->n = N; 5058 if (mat->rmap->N < 0) mat->rmap->N = M; 5059 if (mat->cmap->N < 0) mat->cmap->N = N; 5060 PetscCall(PetscLayoutSetUp(mat->rmap)); 5061 PetscCall(PetscLayoutSetUp(mat->cmap)); 5062 5063 /* check if the matrix sizes are correct */ 5064 PetscCall(MatGetSize(mat, &rows, &cols)); 5065 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); 5066 5067 /* read in row lengths */ 5068 PetscCall(PetscMalloc1(M, &rowlens)); 5069 PetscCall(PetscViewerBinaryRead(viewer, rowlens, M, NULL, PETSC_INT)); 5070 /* check if sum(rowlens) is same as nz */ 5071 sum = 0; 5072 for (i = 0; i < M; i++) sum += rowlens[i]; 5073 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); 5074 /* preallocate and check sizes */ 5075 PetscCall(MatSeqAIJSetPreallocation_SeqAIJ(mat, 0, rowlens)); 5076 PetscCall(MatGetSize(mat, &rows, &cols)); 5077 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); 5078 /* store row lengths */ 5079 PetscCall(PetscArraycpy(a->ilen, rowlens, M)); 5080 PetscCall(PetscFree(rowlens)); 5081 5082 /* fill in "i" row pointers */ 5083 a->i[0] = 0; 5084 for (i = 0; i < M; i++) a->i[i + 1] = a->i[i] + a->ilen[i]; 5085 /* read in "j" column indices */ 5086 PetscCall(PetscViewerBinaryRead(viewer, a->j, nz, NULL, PETSC_INT)); 5087 /* read in "a" nonzero values */ 5088 PetscCall(PetscViewerBinaryRead(viewer, a->a, nz, NULL, PETSC_SCALAR)); 5089 5090 PetscCall(MatAssemblyBegin(mat, MAT_FINAL_ASSEMBLY)); 5091 PetscCall(MatAssemblyEnd(mat, MAT_FINAL_ASSEMBLY)); 5092 PetscFunctionReturn(PETSC_SUCCESS); 5093 } 5094 5095 PetscErrorCode MatEqual_SeqAIJ(Mat A, Mat B, PetscBool *flg) 5096 { 5097 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data, *b = (Mat_SeqAIJ *)B->data; 5098 const PetscScalar *aa, *ba; 5099 #if defined(PETSC_USE_COMPLEX) 5100 PetscInt k; 5101 #endif 5102 5103 PetscFunctionBegin; 5104 /* If the matrix dimensions are not equal,or no of nonzeros */ 5105 if ((A->rmap->n != B->rmap->n) || (A->cmap->n != B->cmap->n) || (a->nz != b->nz)) { 5106 *flg = PETSC_FALSE; 5107 PetscFunctionReturn(PETSC_SUCCESS); 5108 } 5109 5110 /* if the a->i are the same */ 5111 PetscCall(PetscArraycmp(a->i, b->i, A->rmap->n + 1, flg)); 5112 if (!*flg) PetscFunctionReturn(PETSC_SUCCESS); 5113 5114 /* if a->j are the same */ 5115 PetscCall(PetscArraycmp(a->j, b->j, a->nz, flg)); 5116 if (!*flg) PetscFunctionReturn(PETSC_SUCCESS); 5117 5118 PetscCall(MatSeqAIJGetArrayRead(A, &aa)); 5119 PetscCall(MatSeqAIJGetArrayRead(B, &ba)); 5120 /* if a->a are the same */ 5121 #if defined(PETSC_USE_COMPLEX) 5122 for (k = 0; k < a->nz; k++) { 5123 if (PetscRealPart(aa[k]) != PetscRealPart(ba[k]) || PetscImaginaryPart(aa[k]) != PetscImaginaryPart(ba[k])) { 5124 *flg = PETSC_FALSE; 5125 PetscFunctionReturn(PETSC_SUCCESS); 5126 } 5127 } 5128 #else 5129 PetscCall(PetscArraycmp(aa, ba, a->nz, flg)); 5130 #endif 5131 PetscCall(MatSeqAIJRestoreArrayRead(A, &aa)); 5132 PetscCall(MatSeqAIJRestoreArrayRead(B, &ba)); 5133 PetscFunctionReturn(PETSC_SUCCESS); 5134 } 5135 5136 /*@ 5137 MatCreateSeqAIJWithArrays - Creates an sequential `MATSEQAIJ` matrix using matrix elements (in CSR format) 5138 provided by the user. 5139 5140 Collective 5141 5142 Input Parameters: 5143 + comm - must be an MPI communicator of size 1 5144 . m - number of rows 5145 . n - number of columns 5146 . i - row indices; that is i[0] = 0, i[row] = i[row-1] + number of elements in that row of the matrix 5147 . j - column indices 5148 - a - matrix values 5149 5150 Output Parameter: 5151 . mat - the matrix 5152 5153 Level: intermediate 5154 5155 Notes: 5156 The `i`, `j`, and `a` arrays are not copied by this routine, the user must free these arrays 5157 once the matrix is destroyed and not before 5158 5159 You cannot set new nonzero locations into this matrix, that will generate an error. 5160 5161 The `i` and `j` indices are 0 based 5162 5163 The format which is used for the sparse matrix input, is equivalent to a 5164 row-major ordering.. i.e for the following matrix, the input data expected is 5165 as shown 5166 .vb 5167 1 0 0 5168 2 0 3 5169 4 5 6 5170 5171 i = {0,1,3,6} [size = nrow+1 = 3+1] 5172 j = {0,0,2,0,1,2} [size = 6]; values must be sorted for each row 5173 v = {1,2,3,4,5,6} [size = 6] 5174 .ve 5175 5176 .seealso: [](ch_matrices), `Mat`, `MatCreate()`, `MatCreateAIJ()`, `MatCreateSeqAIJ()`, `MatCreateMPIAIJWithArrays()`, `MatMPIAIJSetPreallocationCSR()` 5177 @*/ 5178 PetscErrorCode MatCreateSeqAIJWithArrays(MPI_Comm comm, PetscInt m, PetscInt n, PetscInt i[], PetscInt j[], PetscScalar a[], Mat *mat) 5179 { 5180 PetscInt ii; 5181 Mat_SeqAIJ *aij; 5182 PetscInt jj; 5183 5184 PetscFunctionBegin; 5185 PetscCheck(m <= 0 || i[0] == 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "i (row indices) must start with 0"); 5186 PetscCall(MatCreate(comm, mat)); 5187 PetscCall(MatSetSizes(*mat, m, n, m, n)); 5188 /* PetscCall(MatSetBlockSizes(*mat,,)); */ 5189 PetscCall(MatSetType(*mat, MATSEQAIJ)); 5190 PetscCall(MatSeqAIJSetPreallocation_SeqAIJ(*mat, MAT_SKIP_ALLOCATION, NULL)); 5191 aij = (Mat_SeqAIJ *)(*mat)->data; 5192 PetscCall(PetscMalloc1(m, &aij->imax)); 5193 PetscCall(PetscMalloc1(m, &aij->ilen)); 5194 5195 aij->i = i; 5196 aij->j = j; 5197 aij->a = a; 5198 aij->singlemalloc = PETSC_FALSE; 5199 aij->nonew = -1; /*this indicates that inserting a new value in the matrix that generates a new nonzero is an error*/ 5200 aij->free_a = PETSC_FALSE; 5201 aij->free_ij = PETSC_FALSE; 5202 5203 for (ii = 0, aij->nonzerorowcnt = 0, aij->rmax = 0; ii < m; ii++) { 5204 aij->ilen[ii] = aij->imax[ii] = i[ii + 1] - i[ii]; 5205 if (PetscDefined(USE_DEBUG)) { 5206 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]); 5207 for (jj = i[ii] + 1; jj < i[ii + 1]; jj++) { 5208 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); 5209 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); 5210 } 5211 } 5212 } 5213 if (PetscDefined(USE_DEBUG)) { 5214 for (ii = 0; ii < aij->i[m]; ii++) { 5215 PetscCheck(j[ii] >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Negative column index at location = %" PetscInt_FMT " index = %" PetscInt_FMT, ii, j[ii]); 5216 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]); 5217 } 5218 } 5219 5220 PetscCall(MatAssemblyBegin(*mat, MAT_FINAL_ASSEMBLY)); 5221 PetscCall(MatAssemblyEnd(*mat, MAT_FINAL_ASSEMBLY)); 5222 PetscFunctionReturn(PETSC_SUCCESS); 5223 } 5224 5225 /*@ 5226 MatCreateSeqAIJFromTriple - Creates an sequential `MATSEQAIJ` matrix using matrix elements (in COO format) 5227 provided by the user. 5228 5229 Collective 5230 5231 Input Parameters: 5232 + comm - must be an MPI communicator of size 1 5233 . m - number of rows 5234 . n - number of columns 5235 . i - row indices 5236 . j - column indices 5237 . a - matrix values 5238 . nz - number of nonzeros 5239 - idx - if the `i` and `j` indices start with 1 use `PETSC_TRUE` otherwise use `PETSC_FALSE` 5240 5241 Output Parameter: 5242 . mat - the matrix 5243 5244 Level: intermediate 5245 5246 Example: 5247 For the following matrix, the input data expected is as shown (using 0 based indexing) 5248 .vb 5249 1 0 0 5250 2 0 3 5251 4 5 6 5252 5253 i = {0,1,1,2,2,2} 5254 j = {0,0,2,0,1,2} 5255 v = {1,2,3,4,5,6} 5256 .ve 5257 5258 Note: 5259 Instead of using this function, users should also consider `MatSetPreallocationCOO()` and `MatSetValuesCOO()`, which allow repeated or remote entries, 5260 and are particularly useful in iterative applications. 5261 5262 .seealso: [](ch_matrices), `Mat`, `MatCreate()`, `MatCreateAIJ()`, `MatCreateSeqAIJ()`, `MatCreateSeqAIJWithArrays()`, `MatMPIAIJSetPreallocationCSR()`, `MatSetValuesCOO()`, `MatSetPreallocationCOO()` 5263 @*/ 5264 PetscErrorCode MatCreateSeqAIJFromTriple(MPI_Comm comm, PetscInt m, PetscInt n, PetscInt i[], PetscInt j[], PetscScalar a[], Mat *mat, PetscInt nz, PetscBool idx) 5265 { 5266 PetscInt ii, *nnz, one = 1, row, col; 5267 5268 PetscFunctionBegin; 5269 PetscCall(PetscCalloc1(m, &nnz)); 5270 for (ii = 0; ii < nz; ii++) nnz[i[ii] - !!idx] += 1; 5271 PetscCall(MatCreate(comm, mat)); 5272 PetscCall(MatSetSizes(*mat, m, n, m, n)); 5273 PetscCall(MatSetType(*mat, MATSEQAIJ)); 5274 PetscCall(MatSeqAIJSetPreallocation_SeqAIJ(*mat, 0, nnz)); 5275 for (ii = 0; ii < nz; ii++) { 5276 if (idx) { 5277 row = i[ii] - 1; 5278 col = j[ii] - 1; 5279 } else { 5280 row = i[ii]; 5281 col = j[ii]; 5282 } 5283 PetscCall(MatSetValues(*mat, one, &row, one, &col, &a[ii], ADD_VALUES)); 5284 } 5285 PetscCall(MatAssemblyBegin(*mat, MAT_FINAL_ASSEMBLY)); 5286 PetscCall(MatAssemblyEnd(*mat, MAT_FINAL_ASSEMBLY)); 5287 PetscCall(PetscFree(nnz)); 5288 PetscFunctionReturn(PETSC_SUCCESS); 5289 } 5290 5291 PetscErrorCode MatSeqAIJInvalidateDiagonal(Mat A) 5292 { 5293 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 5294 5295 PetscFunctionBegin; 5296 a->idiagvalid = PETSC_FALSE; 5297 a->ibdiagvalid = PETSC_FALSE; 5298 5299 PetscCall(MatSeqAIJInvalidateDiagonal_Inode(A)); 5300 PetscFunctionReturn(PETSC_SUCCESS); 5301 } 5302 5303 PetscErrorCode MatCreateMPIMatConcatenateSeqMat_SeqAIJ(MPI_Comm comm, Mat inmat, PetscInt n, MatReuse scall, Mat *outmat) 5304 { 5305 PetscFunctionBegin; 5306 PetscCall(MatCreateMPIMatConcatenateSeqMat_MPIAIJ(comm, inmat, n, scall, outmat)); 5307 PetscFunctionReturn(PETSC_SUCCESS); 5308 } 5309 5310 /* 5311 Permute A into C's *local* index space using rowemb,colemb. 5312 The embedding are supposed to be injections and the above implies that the range of rowemb is a subset 5313 of [0,m), colemb is in [0,n). 5314 If pattern == DIFFERENT_NONZERO_PATTERN, C is preallocated according to A. 5315 */ 5316 PetscErrorCode MatSetSeqMat_SeqAIJ(Mat C, IS rowemb, IS colemb, MatStructure pattern, Mat B) 5317 { 5318 /* If making this function public, change the error returned in this function away from _PLIB. */ 5319 Mat_SeqAIJ *Baij; 5320 PetscBool seqaij; 5321 PetscInt m, n, *nz, i, j, count; 5322 PetscScalar v; 5323 const PetscInt *rowindices, *colindices; 5324 5325 PetscFunctionBegin; 5326 if (!B) PetscFunctionReturn(PETSC_SUCCESS); 5327 /* Check to make sure the target matrix (and embeddings) are compatible with C and each other. */ 5328 PetscCall(PetscObjectBaseTypeCompare((PetscObject)B, MATSEQAIJ, &seqaij)); 5329 PetscCheck(seqaij, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Input matrix is of wrong type"); 5330 if (rowemb) { 5331 PetscCall(ISGetLocalSize(rowemb, &m)); 5332 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); 5333 } else { 5334 PetscCheck(C->rmap->n == B->rmap->n, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Input matrix is row-incompatible with the target matrix"); 5335 } 5336 if (colemb) { 5337 PetscCall(ISGetLocalSize(colemb, &n)); 5338 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); 5339 } else { 5340 PetscCheck(C->cmap->n == B->cmap->n, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Input matrix is col-incompatible with the target matrix"); 5341 } 5342 5343 Baij = (Mat_SeqAIJ *)(B->data); 5344 if (pattern == DIFFERENT_NONZERO_PATTERN) { 5345 PetscCall(PetscMalloc1(B->rmap->n, &nz)); 5346 for (i = 0; i < B->rmap->n; i++) nz[i] = Baij->i[i + 1] - Baij->i[i]; 5347 PetscCall(MatSeqAIJSetPreallocation(C, 0, nz)); 5348 PetscCall(PetscFree(nz)); 5349 } 5350 if (pattern == SUBSET_NONZERO_PATTERN) PetscCall(MatZeroEntries(C)); 5351 count = 0; 5352 rowindices = NULL; 5353 colindices = NULL; 5354 if (rowemb) PetscCall(ISGetIndices(rowemb, &rowindices)); 5355 if (colemb) PetscCall(ISGetIndices(colemb, &colindices)); 5356 for (i = 0; i < B->rmap->n; i++) { 5357 PetscInt row; 5358 row = i; 5359 if (rowindices) row = rowindices[i]; 5360 for (j = Baij->i[i]; j < Baij->i[i + 1]; j++) { 5361 PetscInt col; 5362 col = Baij->j[count]; 5363 if (colindices) col = colindices[col]; 5364 v = Baij->a[count]; 5365 PetscCall(MatSetValues(C, 1, &row, 1, &col, &v, INSERT_VALUES)); 5366 ++count; 5367 } 5368 } 5369 /* FIXME: set C's nonzerostate correctly. */ 5370 /* Assembly for C is necessary. */ 5371 C->preallocated = PETSC_TRUE; 5372 C->assembled = PETSC_TRUE; 5373 C->was_assembled = PETSC_FALSE; 5374 PetscFunctionReturn(PETSC_SUCCESS); 5375 } 5376 5377 PetscErrorCode MatEliminateZeros_SeqAIJ(Mat A, PetscBool keep) 5378 { 5379 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 5380 MatScalar *aa = a->a; 5381 PetscInt m = A->rmap->n, fshift = 0, fshift_prev = 0, i, k; 5382 PetscInt *ailen = a->ilen, *imax = a->imax, *ai = a->i, *aj = a->j, rmax = 0; 5383 5384 PetscFunctionBegin; 5385 PetscCheck(A->assembled, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Cannot eliminate zeros for unassembled matrix"); 5386 if (m) rmax = ailen[0]; /* determine row with most nonzeros */ 5387 for (i = 1; i <= m; i++) { 5388 /* move each nonzero entry back by the amount of zero slots (fshift) before it*/ 5389 for (k = ai[i - 1]; k < ai[i]; k++) { 5390 if (aa[k] == 0 && (aj[k] != i - 1 || !keep)) fshift++; 5391 else { 5392 if (aa[k] == 0 && aj[k] == i - 1) PetscCall(PetscInfo(A, "Keep the diagonal zero at row %" PetscInt_FMT "\n", i - 1)); 5393 aa[k - fshift] = aa[k]; 5394 aj[k - fshift] = aj[k]; 5395 } 5396 } 5397 ai[i - 1] -= fshift_prev; // safe to update ai[i-1] now since it will not be used in the next iteration 5398 fshift_prev = fshift; 5399 /* reset ilen and imax for each row */ 5400 ailen[i - 1] = imax[i - 1] = ai[i] - fshift - ai[i - 1]; 5401 a->nonzerorowcnt += ((ai[i] - fshift - ai[i - 1]) > 0); 5402 rmax = PetscMax(rmax, ailen[i - 1]); 5403 } 5404 if (fshift) { 5405 if (m) { 5406 ai[m] -= fshift; 5407 a->nz = ai[m]; 5408 } 5409 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)); 5410 A->nonzerostate++; 5411 A->info.nz_unneeded += (PetscReal)fshift; 5412 a->rmax = rmax; 5413 if (a->inode.use && a->inode.checked) PetscCall(MatSeqAIJCheckInode(A)); 5414 PetscCall(MatAssemblyBegin(A, MAT_FINAL_ASSEMBLY)); 5415 PetscCall(MatAssemblyEnd(A, MAT_FINAL_ASSEMBLY)); 5416 } 5417 PetscFunctionReturn(PETSC_SUCCESS); 5418 } 5419 5420 PetscFunctionList MatSeqAIJList = NULL; 5421 5422 /*@C 5423 MatSeqAIJSetType - Converts a `MATSEQAIJ` matrix to a subtype 5424 5425 Collective 5426 5427 Input Parameters: 5428 + mat - the matrix object 5429 - matype - matrix type 5430 5431 Options Database Key: 5432 . -mat_seqaij_type <method> - for example seqaijcrl 5433 5434 Level: intermediate 5435 5436 .seealso: [](ch_matrices), `Mat`, `PCSetType()`, `VecSetType()`, `MatCreate()`, `MatType` 5437 @*/ 5438 PetscErrorCode MatSeqAIJSetType(Mat mat, MatType matype) 5439 { 5440 PetscBool sametype; 5441 PetscErrorCode (*r)(Mat, MatType, MatReuse, Mat *); 5442 5443 PetscFunctionBegin; 5444 PetscValidHeaderSpecific(mat, MAT_CLASSID, 1); 5445 PetscCall(PetscObjectTypeCompare((PetscObject)mat, matype, &sametype)); 5446 if (sametype) PetscFunctionReturn(PETSC_SUCCESS); 5447 5448 PetscCall(PetscFunctionListFind(MatSeqAIJList, matype, &r)); 5449 PetscCheck(r, PetscObjectComm((PetscObject)mat), PETSC_ERR_ARG_UNKNOWN_TYPE, "Unknown Mat type given: %s", matype); 5450 PetscCall((*r)(mat, matype, MAT_INPLACE_MATRIX, &mat)); 5451 PetscFunctionReturn(PETSC_SUCCESS); 5452 } 5453 5454 /*@C 5455 MatSeqAIJRegister - - Adds a new sub-matrix type for sequential `MATSEQAIJ` matrices 5456 5457 Not Collective 5458 5459 Input Parameters: 5460 + sname - name of a new user-defined matrix type, for example `MATSEQAIJCRL` 5461 - function - routine to convert to subtype 5462 5463 Level: advanced 5464 5465 Notes: 5466 `MatSeqAIJRegister()` may be called multiple times to add several user-defined solvers. 5467 5468 Then, your matrix can be chosen with the procedural interface at runtime via the option 5469 $ -mat_seqaij_type my_mat 5470 5471 .seealso: [](ch_matrices), `Mat`, `MatSeqAIJRegisterAll()` 5472 @*/ 5473 PetscErrorCode MatSeqAIJRegister(const char sname[], PetscErrorCode (*function)(Mat, MatType, MatReuse, Mat *)) 5474 { 5475 PetscFunctionBegin; 5476 PetscCall(MatInitializePackage()); 5477 PetscCall(PetscFunctionListAdd(&MatSeqAIJList, sname, function)); 5478 PetscFunctionReturn(PETSC_SUCCESS); 5479 } 5480 5481 PetscBool MatSeqAIJRegisterAllCalled = PETSC_FALSE; 5482 5483 /*@C 5484 MatSeqAIJRegisterAll - Registers all of the matrix subtypes of `MATSSEQAIJ` 5485 5486 Not Collective 5487 5488 Level: advanced 5489 5490 Note: 5491 This registers the versions of `MATSEQAIJ` for GPUs 5492 5493 .seealso: [](ch_matrices), `Mat`, `MatRegisterAll()`, `MatSeqAIJRegister()` 5494 @*/ 5495 PetscErrorCode MatSeqAIJRegisterAll(void) 5496 { 5497 PetscFunctionBegin; 5498 if (MatSeqAIJRegisterAllCalled) PetscFunctionReturn(PETSC_SUCCESS); 5499 MatSeqAIJRegisterAllCalled = PETSC_TRUE; 5500 5501 PetscCall(MatSeqAIJRegister(MATSEQAIJCRL, MatConvert_SeqAIJ_SeqAIJCRL)); 5502 PetscCall(MatSeqAIJRegister(MATSEQAIJPERM, MatConvert_SeqAIJ_SeqAIJPERM)); 5503 PetscCall(MatSeqAIJRegister(MATSEQAIJSELL, MatConvert_SeqAIJ_SeqAIJSELL)); 5504 #if defined(PETSC_HAVE_MKL_SPARSE) 5505 PetscCall(MatSeqAIJRegister(MATSEQAIJMKL, MatConvert_SeqAIJ_SeqAIJMKL)); 5506 #endif 5507 #if defined(PETSC_HAVE_CUDA) 5508 PetscCall(MatSeqAIJRegister(MATSEQAIJCUSPARSE, MatConvert_SeqAIJ_SeqAIJCUSPARSE)); 5509 #endif 5510 #if defined(PETSC_HAVE_HIP) 5511 PetscCall(MatSeqAIJRegister(MATSEQAIJHIPSPARSE, MatConvert_SeqAIJ_SeqAIJHIPSPARSE)); 5512 #endif 5513 #if defined(PETSC_HAVE_KOKKOS_KERNELS) 5514 PetscCall(MatSeqAIJRegister(MATSEQAIJKOKKOS, MatConvert_SeqAIJ_SeqAIJKokkos)); 5515 #endif 5516 #if defined(PETSC_HAVE_VIENNACL) && defined(PETSC_HAVE_VIENNACL_NO_CUDA) 5517 PetscCall(MatSeqAIJRegister(MATMPIAIJVIENNACL, MatConvert_SeqAIJ_SeqAIJViennaCL)); 5518 #endif 5519 PetscFunctionReturn(PETSC_SUCCESS); 5520 } 5521 5522 /* 5523 Special version for direct calls from Fortran 5524 */ 5525 #include <petsc/private/fortranimpl.h> 5526 #if defined(PETSC_HAVE_FORTRAN_CAPS) 5527 #define matsetvaluesseqaij_ MATSETVALUESSEQAIJ 5528 #elif !defined(PETSC_HAVE_FORTRAN_UNDERSCORE) 5529 #define matsetvaluesseqaij_ matsetvaluesseqaij 5530 #endif 5531 5532 /* Change these macros so can be used in void function */ 5533 5534 /* Change these macros so can be used in void function */ 5535 /* Identical to PetscCallVoid, except it assigns to *_ierr */ 5536 #undef PetscCall 5537 #define PetscCall(...) \ 5538 do { \ 5539 PetscErrorCode ierr_msv_mpiaij = __VA_ARGS__; \ 5540 if (PetscUnlikely(ierr_msv_mpiaij)) { \ 5541 *_ierr = PetscError(PETSC_COMM_SELF, __LINE__, PETSC_FUNCTION_NAME, __FILE__, ierr_msv_mpiaij, PETSC_ERROR_REPEAT, " "); \ 5542 return; \ 5543 } \ 5544 } while (0) 5545 5546 #undef SETERRQ 5547 #define SETERRQ(comm, ierr, ...) \ 5548 do { \ 5549 *_ierr = PetscError(comm, __LINE__, PETSC_FUNCTION_NAME, __FILE__, ierr, PETSC_ERROR_INITIAL, __VA_ARGS__); \ 5550 return; \ 5551 } while (0) 5552 5553 PETSC_EXTERN void matsetvaluesseqaij_(Mat *AA, PetscInt *mm, const PetscInt im[], PetscInt *nn, const PetscInt in[], const PetscScalar v[], InsertMode *isis, PetscErrorCode *_ierr) 5554 { 5555 Mat A = *AA; 5556 PetscInt m = *mm, n = *nn; 5557 InsertMode is = *isis; 5558 Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data; 5559 PetscInt *rp, k, low, high, t, ii, row, nrow, i, col, l, rmax, N; 5560 PetscInt *imax, *ai, *ailen; 5561 PetscInt *aj, nonew = a->nonew, lastcol = -1; 5562 MatScalar *ap, value, *aa; 5563 PetscBool ignorezeroentries = a->ignorezeroentries; 5564 PetscBool roworiented = a->roworiented; 5565 5566 PetscFunctionBegin; 5567 MatCheckPreallocated(A, 1); 5568 imax = a->imax; 5569 ai = a->i; 5570 ailen = a->ilen; 5571 aj = a->j; 5572 aa = a->a; 5573 5574 for (k = 0; k < m; k++) { /* loop over added rows */ 5575 row = im[k]; 5576 if (row < 0) continue; 5577 PetscCheck(row < A->rmap->n, PetscObjectComm((PetscObject)A), PETSC_ERR_ARG_OUTOFRANGE, "Row too large"); 5578 rp = aj + ai[row]; 5579 ap = aa + ai[row]; 5580 rmax = imax[row]; 5581 nrow = ailen[row]; 5582 low = 0; 5583 high = nrow; 5584 for (l = 0; l < n; l++) { /* loop over added columns */ 5585 if (in[l] < 0) continue; 5586 PetscCheck(in[l] < A->cmap->n, PetscObjectComm((PetscObject)A), PETSC_ERR_ARG_OUTOFRANGE, "Column too large"); 5587 col = in[l]; 5588 if (roworiented) value = v[l + k * n]; 5589 else value = v[k + l * m]; 5590 5591 if (value == 0.0 && ignorezeroentries && (is == ADD_VALUES)) continue; 5592 5593 if (col <= lastcol) low = 0; 5594 else high = nrow; 5595 lastcol = col; 5596 while (high - low > 5) { 5597 t = (low + high) / 2; 5598 if (rp[t] > col) high = t; 5599 else low = t; 5600 } 5601 for (i = low; i < high; i++) { 5602 if (rp[i] > col) break; 5603 if (rp[i] == col) { 5604 if (is == ADD_VALUES) ap[i] += value; 5605 else ap[i] = value; 5606 goto noinsert; 5607 } 5608 } 5609 if (value == 0.0 && ignorezeroentries) goto noinsert; 5610 if (nonew == 1) goto noinsert; 5611 PetscCheck(nonew != -1, PetscObjectComm((PetscObject)A), PETSC_ERR_ARG_OUTOFRANGE, "Inserting a new nonzero in the matrix"); 5612 MatSeqXAIJReallocateAIJ(A, A->rmap->n, 1, nrow, row, col, rmax, aa, ai, aj, rp, ap, imax, nonew, MatScalar); 5613 N = nrow++ - 1; 5614 a->nz++; 5615 high++; 5616 /* shift up all the later entries in this row */ 5617 for (ii = N; ii >= i; ii--) { 5618 rp[ii + 1] = rp[ii]; 5619 ap[ii + 1] = ap[ii]; 5620 } 5621 rp[i] = col; 5622 ap[i] = value; 5623 A->nonzerostate++; 5624 noinsert:; 5625 low = i + 1; 5626 } 5627 ailen[row] = nrow; 5628 } 5629 PetscFunctionReturnVoid(); 5630 } 5631 /* Undefining these here since they were redefined from their original definition above! No 5632 * other PETSc functions should be defined past this point, as it is impossible to recover the 5633 * original definitions */ 5634 #undef PetscCall 5635 #undef SETERRQ 5636