Actual source code: aij.c
1: /*
2: Defines the basic matrix operations for the AIJ (compressed row)
3: matrix storage format.
4: */
6: #include <../src/mat/impls/aij/seq/aij.h>
7: #include <petscblaslapack.h>
8: #include <petscbt.h>
9: #include <petsc/private/kernels/blocktranspose.h>
11: PetscErrorCode MatSeqAIJSetTypeFromOptions(Mat A)
12: {
13: PetscBool flg;
14: char type[256];
16: PetscObjectOptionsBegin((PetscObject)A);
17: PetscOptionsFList("-mat_seqaij_type", "Matrix SeqAIJ type", "MatSeqAIJSetType", MatSeqAIJList, "seqaij", type, 256, &flg);
18: if (flg) MatSeqAIJSetType(A, type);
19: PetscOptionsEnd();
20: return 0;
21: }
23: PetscErrorCode MatGetColumnReductions_SeqAIJ(Mat A, PetscInt type, PetscReal *reductions)
24: {
25: PetscInt i, m, n;
26: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)A->data;
28: MatGetSize(A, &m, &n);
29: PetscArrayzero(reductions, n);
30: if (type == NORM_2) {
31: for (i = 0; i < aij->i[m]; i++) reductions[aij->j[i]] += PetscAbsScalar(aij->a[i] * aij->a[i]);
32: } else if (type == NORM_1) {
33: for (i = 0; i < aij->i[m]; i++) reductions[aij->j[i]] += PetscAbsScalar(aij->a[i]);
34: } else if (type == NORM_INFINITY) {
35: for (i = 0; i < aij->i[m]; i++) reductions[aij->j[i]] = PetscMax(PetscAbsScalar(aij->a[i]), reductions[aij->j[i]]);
36: } else if (type == REDUCTION_SUM_REALPART || type == REDUCTION_MEAN_REALPART) {
37: for (i = 0; i < aij->i[m]; i++) reductions[aij->j[i]] += PetscRealPart(aij->a[i]);
38: } else if (type == REDUCTION_SUM_IMAGINARYPART || type == REDUCTION_MEAN_IMAGINARYPART) {
39: for (i = 0; i < aij->i[m]; i++) reductions[aij->j[i]] += PetscImaginaryPart(aij->a[i]);
40: } else SETERRQ(PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Unknown reduction type");
42: if (type == NORM_2) {
43: for (i = 0; i < n; i++) reductions[i] = PetscSqrtReal(reductions[i]);
44: } else if (type == REDUCTION_MEAN_REALPART || type == REDUCTION_MEAN_IMAGINARYPART) {
45: for (i = 0; i < n; i++) reductions[i] /= m;
46: }
47: return 0;
48: }
50: PetscErrorCode MatFindOffBlockDiagonalEntries_SeqAIJ(Mat A, IS *is)
51: {
52: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
53: PetscInt i, m = A->rmap->n, cnt = 0, bs = A->rmap->bs;
54: const PetscInt *jj = a->j, *ii = a->i;
55: PetscInt *rows;
57: for (i = 0; i < m; i++) {
58: if ((ii[i] != ii[i + 1]) && ((jj[ii[i]] < bs * (i / bs)) || (jj[ii[i + 1] - 1] > bs * ((i + bs) / bs) - 1))) cnt++;
59: }
60: PetscMalloc1(cnt, &rows);
61: cnt = 0;
62: for (i = 0; i < m; i++) {
63: if ((ii[i] != ii[i + 1]) && ((jj[ii[i]] < bs * (i / bs)) || (jj[ii[i + 1] - 1] > bs * ((i + bs) / bs) - 1))) {
64: rows[cnt] = i;
65: cnt++;
66: }
67: }
68: ISCreateGeneral(PETSC_COMM_SELF, cnt, rows, PETSC_OWN_POINTER, is);
69: return 0;
70: }
72: PetscErrorCode MatFindZeroDiagonals_SeqAIJ_Private(Mat A, PetscInt *nrows, PetscInt **zrows)
73: {
74: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
75: const MatScalar *aa;
76: PetscInt i, m = A->rmap->n, cnt = 0;
77: const PetscInt *ii = a->i, *jj = a->j, *diag;
78: PetscInt *rows;
80: MatSeqAIJGetArrayRead(A, &aa);
81: MatMarkDiagonal_SeqAIJ(A);
82: diag = a->diag;
83: for (i = 0; i < m; i++) {
84: if ((diag[i] >= ii[i + 1]) || (jj[diag[i]] != i) || (aa[diag[i]] == 0.0)) cnt++;
85: }
86: PetscMalloc1(cnt, &rows);
87: cnt = 0;
88: for (i = 0; i < m; i++) {
89: if ((diag[i] >= ii[i + 1]) || (jj[diag[i]] != i) || (aa[diag[i]] == 0.0)) rows[cnt++] = i;
90: }
91: *nrows = cnt;
92: *zrows = rows;
93: MatSeqAIJRestoreArrayRead(A, &aa);
94: return 0;
95: }
97: PetscErrorCode MatFindZeroDiagonals_SeqAIJ(Mat A, IS *zrows)
98: {
99: PetscInt nrows, *rows;
101: *zrows = NULL;
102: MatFindZeroDiagonals_SeqAIJ_Private(A, &nrows, &rows);
103: ISCreateGeneral(PetscObjectComm((PetscObject)A), nrows, rows, PETSC_OWN_POINTER, zrows);
104: return 0;
105: }
107: PetscErrorCode MatFindNonzeroRows_SeqAIJ(Mat A, IS *keptrows)
108: {
109: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
110: const MatScalar *aa;
111: PetscInt m = A->rmap->n, cnt = 0;
112: const PetscInt *ii;
113: PetscInt n, i, j, *rows;
115: MatSeqAIJGetArrayRead(A, &aa);
116: *keptrows = NULL;
117: ii = a->i;
118: for (i = 0; i < m; i++) {
119: n = ii[i + 1] - ii[i];
120: if (!n) {
121: cnt++;
122: goto ok1;
123: }
124: for (j = ii[i]; j < ii[i + 1]; j++) {
125: if (aa[j] != 0.0) goto ok1;
126: }
127: cnt++;
128: ok1:;
129: }
130: if (!cnt) {
131: MatSeqAIJRestoreArrayRead(A, &aa);
132: return 0;
133: }
134: PetscMalloc1(A->rmap->n - cnt, &rows);
135: cnt = 0;
136: for (i = 0; i < m; i++) {
137: n = ii[i + 1] - ii[i];
138: if (!n) continue;
139: for (j = ii[i]; j < ii[i + 1]; j++) {
140: if (aa[j] != 0.0) {
141: rows[cnt++] = i;
142: break;
143: }
144: }
145: }
146: MatSeqAIJRestoreArrayRead(A, &aa);
147: ISCreateGeneral(PETSC_COMM_SELF, cnt, rows, PETSC_OWN_POINTER, keptrows);
148: return 0;
149: }
151: PetscErrorCode MatDiagonalSet_SeqAIJ(Mat Y, Vec D, InsertMode is)
152: {
153: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)Y->data;
154: PetscInt i, m = Y->rmap->n;
155: const PetscInt *diag;
156: MatScalar *aa;
157: const PetscScalar *v;
158: PetscBool missing;
160: if (Y->assembled) {
161: MatMissingDiagonal_SeqAIJ(Y, &missing, NULL);
162: if (!missing) {
163: diag = aij->diag;
164: VecGetArrayRead(D, &v);
165: MatSeqAIJGetArray(Y, &aa);
166: if (is == INSERT_VALUES) {
167: for (i = 0; i < m; i++) aa[diag[i]] = v[i];
168: } else {
169: for (i = 0; i < m; i++) aa[diag[i]] += v[i];
170: }
171: MatSeqAIJRestoreArray(Y, &aa);
172: VecRestoreArrayRead(D, &v);
173: return 0;
174: }
175: MatSeqAIJInvalidateDiagonal(Y);
176: }
177: MatDiagonalSet_Default(Y, D, is);
178: return 0;
179: }
181: PetscErrorCode MatGetRowIJ_SeqAIJ(Mat A, PetscInt oshift, PetscBool symmetric, PetscBool inodecompressed, PetscInt *m, const PetscInt *ia[], const PetscInt *ja[], PetscBool *done)
182: {
183: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
184: PetscInt i, ishift;
186: if (m) *m = A->rmap->n;
187: if (!ia) return 0;
188: ishift = 0;
189: if (symmetric && A->structurally_symmetric != PETSC_BOOL3_TRUE) {
190: MatToSymmetricIJ_SeqAIJ(A->rmap->n, a->i, a->j, PETSC_TRUE, ishift, oshift, (PetscInt **)ia, (PetscInt **)ja);
191: } else if (oshift == 1) {
192: PetscInt *tia;
193: PetscInt nz = a->i[A->rmap->n];
194: /* malloc space and add 1 to i and j indices */
195: PetscMalloc1(A->rmap->n + 1, &tia);
196: for (i = 0; i < A->rmap->n + 1; i++) tia[i] = a->i[i] + 1;
197: *ia = tia;
198: if (ja) {
199: PetscInt *tja;
200: PetscMalloc1(nz + 1, &tja);
201: for (i = 0; i < nz; i++) tja[i] = a->j[i] + 1;
202: *ja = tja;
203: }
204: } else {
205: *ia = a->i;
206: if (ja) *ja = a->j;
207: }
208: return 0;
209: }
211: PetscErrorCode MatRestoreRowIJ_SeqAIJ(Mat A, PetscInt oshift, PetscBool symmetric, PetscBool inodecompressed, PetscInt *n, const PetscInt *ia[], const PetscInt *ja[], PetscBool *done)
212: {
213: if (!ia) return 0;
214: if ((symmetric && A->structurally_symmetric != PETSC_BOOL3_TRUE) || oshift == 1) {
215: PetscFree(*ia);
216: if (ja) PetscFree(*ja);
217: }
218: return 0;
219: }
221: PetscErrorCode MatGetColumnIJ_SeqAIJ(Mat A, PetscInt oshift, PetscBool symmetric, PetscBool inodecompressed, PetscInt *nn, const PetscInt *ia[], const PetscInt *ja[], PetscBool *done)
222: {
223: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
224: PetscInt i, *collengths, *cia, *cja, n = A->cmap->n, m = A->rmap->n;
225: PetscInt nz = a->i[m], row, *jj, mr, col;
227: *nn = n;
228: if (!ia) return 0;
229: if (symmetric) {
230: MatToSymmetricIJ_SeqAIJ(A->rmap->n, a->i, a->j, PETSC_TRUE, 0, oshift, (PetscInt **)ia, (PetscInt **)ja);
231: } else {
232: PetscCalloc1(n, &collengths);
233: PetscMalloc1(n + 1, &cia);
234: PetscMalloc1(nz, &cja);
235: jj = a->j;
236: for (i = 0; i < nz; i++) collengths[jj[i]]++;
237: cia[0] = oshift;
238: for (i = 0; i < n; i++) cia[i + 1] = cia[i] + collengths[i];
239: PetscArrayzero(collengths, n);
240: jj = a->j;
241: for (row = 0; row < m; row++) {
242: mr = a->i[row + 1] - a->i[row];
243: for (i = 0; i < mr; i++) {
244: col = *jj++;
246: cja[cia[col] + collengths[col]++ - oshift] = row + oshift;
247: }
248: }
249: PetscFree(collengths);
250: *ia = cia;
251: *ja = cja;
252: }
253: return 0;
254: }
256: PetscErrorCode MatRestoreColumnIJ_SeqAIJ(Mat A, PetscInt oshift, PetscBool symmetric, PetscBool inodecompressed, PetscInt *n, const PetscInt *ia[], const PetscInt *ja[], PetscBool *done)
257: {
258: if (!ia) return 0;
260: PetscFree(*ia);
261: PetscFree(*ja);
262: return 0;
263: }
265: /*
266: MatGetColumnIJ_SeqAIJ_Color() and MatRestoreColumnIJ_SeqAIJ_Color() are customized from
267: MatGetColumnIJ_SeqAIJ() and MatRestoreColumnIJ_SeqAIJ() by adding an output
268: spidx[], index of a->a, to be used in MatTransposeColoringCreate_SeqAIJ() and MatFDColoringCreate_SeqXAIJ()
269: */
270: PetscErrorCode MatGetColumnIJ_SeqAIJ_Color(Mat A, PetscInt oshift, PetscBool symmetric, PetscBool inodecompressed, PetscInt *nn, const PetscInt *ia[], const PetscInt *ja[], PetscInt *spidx[], PetscBool *done)
271: {
272: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
273: PetscInt i, *collengths, *cia, *cja, n = A->cmap->n, m = A->rmap->n;
274: PetscInt nz = a->i[m], row, mr, col, tmp;
275: PetscInt *cspidx;
276: const PetscInt *jj;
278: *nn = n;
279: if (!ia) return 0;
281: PetscCalloc1(n, &collengths);
282: PetscMalloc1(n + 1, &cia);
283: PetscMalloc1(nz, &cja);
284: PetscMalloc1(nz, &cspidx);
285: jj = a->j;
286: for (i = 0; i < nz; i++) collengths[jj[i]]++;
287: cia[0] = oshift;
288: for (i = 0; i < n; i++) cia[i + 1] = cia[i] + collengths[i];
289: PetscArrayzero(collengths, n);
290: jj = a->j;
291: for (row = 0; row < m; row++) {
292: mr = a->i[row + 1] - a->i[row];
293: for (i = 0; i < mr; i++) {
294: col = *jj++;
295: tmp = cia[col] + collengths[col]++ - oshift;
296: cspidx[tmp] = a->i[row] + i; /* index of a->j */
297: cja[tmp] = row + oshift;
298: }
299: }
300: PetscFree(collengths);
301: *ia = cia;
302: *ja = cja;
303: *spidx = cspidx;
304: return 0;
305: }
307: PetscErrorCode MatRestoreColumnIJ_SeqAIJ_Color(Mat A, PetscInt oshift, PetscBool symmetric, PetscBool inodecompressed, PetscInt *n, const PetscInt *ia[], const PetscInt *ja[], PetscInt *spidx[], PetscBool *done)
308: {
309: MatRestoreColumnIJ_SeqAIJ(A, oshift, symmetric, inodecompressed, n, ia, ja, done);
310: PetscFree(*spidx);
311: return 0;
312: }
314: PetscErrorCode MatSetValuesRow_SeqAIJ(Mat A, PetscInt row, const PetscScalar v[])
315: {
316: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
317: PetscInt *ai = a->i;
318: PetscScalar *aa;
320: MatSeqAIJGetArray(A, &aa);
321: PetscArraycpy(aa + ai[row], v, ai[row + 1] - ai[row]);
322: MatSeqAIJRestoreArray(A, &aa);
323: return 0;
324: }
326: /*
327: MatSeqAIJSetValuesLocalFast - An optimized version of MatSetValuesLocal() for SeqAIJ matrices with several assumptions
329: - a single row of values is set with each call
330: - no row or column indices are negative or (in error) larger than the number of rows or columns
331: - the values are always added to the matrix, not set
332: - no new locations are introduced in the nonzero structure of the matrix
334: This does NOT assume the global column indices are sorted
336: */
338: #include <petsc/private/isimpl.h>
339: PetscErrorCode MatSeqAIJSetValuesLocalFast(Mat A, PetscInt m, const PetscInt im[], PetscInt n, const PetscInt in[], const PetscScalar v[], InsertMode is)
340: {
341: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
342: PetscInt low, high, t, row, nrow, i, col, l;
343: const PetscInt *rp, *ai = a->i, *ailen = a->ilen, *aj = a->j;
344: PetscInt lastcol = -1;
345: MatScalar *ap, value, *aa;
346: const PetscInt *ridx = A->rmap->mapping->indices, *cidx = A->cmap->mapping->indices;
348: MatSeqAIJGetArray(A, &aa);
349: row = ridx[im[0]];
350: rp = aj + ai[row];
351: ap = aa + ai[row];
352: nrow = ailen[row];
353: low = 0;
354: high = nrow;
355: for (l = 0; l < n; l++) { /* loop over added columns */
356: col = cidx[in[l]];
357: value = v[l];
359: if (col <= lastcol) low = 0;
360: else high = nrow;
361: lastcol = col;
362: while (high - low > 5) {
363: t = (low + high) / 2;
364: if (rp[t] > col) high = t;
365: else low = t;
366: }
367: for (i = low; i < high; i++) {
368: if (rp[i] == col) {
369: ap[i] += value;
370: low = i + 1;
371: break;
372: }
373: }
374: }
375: MatSeqAIJRestoreArray(A, &aa);
376: return 0;
377: }
379: PetscErrorCode MatSetValues_SeqAIJ(Mat A, PetscInt m, const PetscInt im[], PetscInt n, const PetscInt in[], const PetscScalar v[], InsertMode is)
380: {
381: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
382: PetscInt *rp, k, low, high, t, ii, row, nrow, i, col, l, rmax, N;
383: PetscInt *imax = a->imax, *ai = a->i, *ailen = a->ilen;
384: PetscInt *aj = a->j, nonew = a->nonew, lastcol = -1;
385: MatScalar *ap = NULL, value = 0.0, *aa;
386: PetscBool ignorezeroentries = a->ignorezeroentries;
387: PetscBool roworiented = a->roworiented;
389: MatSeqAIJGetArray(A, &aa);
390: for (k = 0; k < m; k++) { /* loop over added rows */
391: row = im[k];
392: if (row < 0) continue;
394: rp = aj + ai[row];
395: if (!A->structure_only) ap = aa + ai[row];
396: rmax = imax[row];
397: nrow = ailen[row];
398: low = 0;
399: high = nrow;
400: for (l = 0; l < n; l++) { /* loop over added columns */
401: if (in[l] < 0) continue;
403: col = in[l];
404: if (v && !A->structure_only) value = roworiented ? v[l + k * n] : v[k + l * m];
405: if (!A->structure_only && value == 0.0 && ignorezeroentries && is == ADD_VALUES && row != col) continue;
407: if (col <= lastcol) low = 0;
408: else high = nrow;
409: lastcol = col;
410: while (high - low > 5) {
411: t = (low + high) / 2;
412: if (rp[t] > col) high = t;
413: else low = t;
414: }
415: for (i = low; i < high; i++) {
416: if (rp[i] > col) break;
417: if (rp[i] == col) {
418: if (!A->structure_only) {
419: if (is == ADD_VALUES) {
420: ap[i] += value;
421: (void)PetscLogFlops(1.0);
422: } else ap[i] = value;
423: }
424: low = i + 1;
425: goto noinsert;
426: }
427: }
428: if (value == 0.0 && ignorezeroentries && row != col) goto noinsert;
429: if (nonew == 1) goto noinsert;
431: if (A->structure_only) {
432: MatSeqXAIJReallocateAIJ_structure_only(A, A->rmap->n, 1, nrow, row, col, rmax, ai, aj, rp, imax, nonew, MatScalar);
433: } else {
434: MatSeqXAIJReallocateAIJ(A, A->rmap->n, 1, nrow, row, col, rmax, aa, ai, aj, rp, ap, imax, nonew, MatScalar);
435: }
436: N = nrow++ - 1;
437: a->nz++;
438: high++;
439: /* shift up all the later entries in this row */
440: PetscArraymove(rp + i + 1, rp + i, N - i + 1);
441: rp[i] = col;
442: if (!A->structure_only) {
443: PetscArraymove(ap + i + 1, ap + i, N - i + 1);
444: ap[i] = value;
445: }
446: low = i + 1;
447: A->nonzerostate++;
448: noinsert:;
449: }
450: ailen[row] = nrow;
451: }
452: MatSeqAIJRestoreArray(A, &aa);
453: return 0;
454: }
456: PetscErrorCode MatSetValues_SeqAIJ_SortedFullNoPreallocation(Mat A, PetscInt m, const PetscInt im[], PetscInt n, const PetscInt in[], const PetscScalar v[], InsertMode is)
457: {
458: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
459: PetscInt *rp, k, row;
460: PetscInt *ai = a->i;
461: PetscInt *aj = a->j;
462: MatScalar *aa, *ap;
467: MatSeqAIJGetArray(A, &aa);
468: for (k = 0; k < m; k++) { /* loop over added rows */
469: row = im[k];
470: rp = aj + ai[row];
471: ap = aa + ai[row];
473: PetscMemcpy(rp, in, n * sizeof(PetscInt));
474: if (!A->structure_only) {
475: if (v) {
476: PetscMemcpy(ap, v, n * sizeof(PetscScalar));
477: v += n;
478: } else {
479: PetscMemzero(ap, n * sizeof(PetscScalar));
480: }
481: }
482: a->ilen[row] = n;
483: a->imax[row] = n;
484: a->i[row + 1] = a->i[row] + n;
485: a->nz += n;
486: }
487: MatSeqAIJRestoreArray(A, &aa);
488: return 0;
489: }
491: /*@
492: MatSeqAIJSetTotalPreallocation - Sets an upper bound on the total number of expected nonzeros in the matrix.
494: Input Parameters:
495: + A - the `MATSEQAIJ` matrix
496: - nztotal - bound on the number of nonzeros
498: Level: advanced
500: Notes:
501: This can be called if you will be provided the matrix row by row (from row zero) with sorted column indices for each row.
502: Simply call `MatSetValues()` after this call to provide the matrix entries in the usual manner. This matrix may be used
503: as always with multiple matrix assemblies.
505: .seealso: `MatSetOption()`, `MAT_SORTED_FULL`, `MatSetValues()`, `MatSeqAIJSetPreallocation()`
506: @*/
508: PetscErrorCode MatSeqAIJSetTotalPreallocation(Mat A, PetscInt nztotal)
509: {
510: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
512: PetscLayoutSetUp(A->rmap);
513: PetscLayoutSetUp(A->cmap);
514: a->maxnz = nztotal;
515: if (!a->imax) { PetscMalloc1(A->rmap->n, &a->imax); }
516: if (!a->ilen) {
517: PetscMalloc1(A->rmap->n, &a->ilen);
518: } else {
519: PetscMemzero(a->ilen, A->rmap->n * sizeof(PetscInt));
520: }
522: /* allocate the matrix space */
523: if (A->structure_only) {
524: PetscMalloc1(nztotal, &a->j);
525: PetscMalloc1(A->rmap->n + 1, &a->i);
526: } else {
527: PetscMalloc3(nztotal, &a->a, nztotal, &a->j, A->rmap->n + 1, &a->i);
528: }
529: a->i[0] = 0;
530: if (A->structure_only) {
531: a->singlemalloc = PETSC_FALSE;
532: a->free_a = PETSC_FALSE;
533: } else {
534: a->singlemalloc = PETSC_TRUE;
535: a->free_a = PETSC_TRUE;
536: }
537: a->free_ij = PETSC_TRUE;
538: A->ops->setvalues = MatSetValues_SeqAIJ_SortedFullNoPreallocation;
539: A->preallocated = PETSC_TRUE;
540: return 0;
541: }
543: PetscErrorCode MatSetValues_SeqAIJ_SortedFull(Mat A, PetscInt m, const PetscInt im[], PetscInt n, const PetscInt in[], const PetscScalar v[], InsertMode is)
544: {
545: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
546: PetscInt *rp, k, row;
547: PetscInt *ai = a->i, *ailen = a->ilen;
548: PetscInt *aj = a->j;
549: MatScalar *aa, *ap;
551: MatSeqAIJGetArray(A, &aa);
552: for (k = 0; k < m; k++) { /* loop over added rows */
553: row = im[k];
555: rp = aj + ai[row];
556: ap = aa + ai[row];
557: if (!A->was_assembled) PetscMemcpy(rp, in, n * sizeof(PetscInt));
558: if (!A->structure_only) {
559: if (v) {
560: PetscMemcpy(ap, v, n * sizeof(PetscScalar));
561: v += n;
562: } else {
563: PetscMemzero(ap, n * sizeof(PetscScalar));
564: }
565: }
566: ailen[row] = n;
567: a->nz += n;
568: }
569: MatSeqAIJRestoreArray(A, &aa);
570: return 0;
571: }
573: PetscErrorCode MatGetValues_SeqAIJ(Mat A, PetscInt m, const PetscInt im[], PetscInt n, const PetscInt in[], PetscScalar v[])
574: {
575: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
576: PetscInt *rp, k, low, high, t, row, nrow, i, col, l, *aj = a->j;
577: PetscInt *ai = a->i, *ailen = a->ilen;
578: const MatScalar *ap, *aa;
580: MatSeqAIJGetArrayRead(A, &aa);
581: for (k = 0; k < m; k++) { /* loop over rows */
582: row = im[k];
583: if (row < 0) {
584: v += n;
585: continue;
586: } /* negative row */
588: rp = aj + ai[row];
589: ap = aa + ai[row];
590: nrow = ailen[row];
591: for (l = 0; l < n; l++) { /* loop over columns */
592: if (in[l] < 0) {
593: v++;
594: continue;
595: } /* negative column */
597: col = in[l];
598: high = nrow;
599: low = 0; /* assume unsorted */
600: while (high - low > 5) {
601: t = (low + high) / 2;
602: if (rp[t] > col) high = t;
603: else low = t;
604: }
605: for (i = low; i < high; i++) {
606: if (rp[i] > col) break;
607: if (rp[i] == col) {
608: *v++ = ap[i];
609: goto finished;
610: }
611: }
612: *v++ = 0.0;
613: finished:;
614: }
615: }
616: MatSeqAIJRestoreArrayRead(A, &aa);
617: return 0;
618: }
620: PetscErrorCode MatView_SeqAIJ_Binary(Mat mat, PetscViewer viewer)
621: {
622: Mat_SeqAIJ *A = (Mat_SeqAIJ *)mat->data;
623: const PetscScalar *av;
624: PetscInt header[4], M, N, m, nz, i;
625: PetscInt *rowlens;
627: PetscViewerSetUp(viewer);
629: M = mat->rmap->N;
630: N = mat->cmap->N;
631: m = mat->rmap->n;
632: nz = A->nz;
634: /* write matrix header */
635: header[0] = MAT_FILE_CLASSID;
636: header[1] = M;
637: header[2] = N;
638: header[3] = nz;
639: PetscViewerBinaryWrite(viewer, header, 4, PETSC_INT);
641: /* fill in and store row lengths */
642: PetscMalloc1(m, &rowlens);
643: for (i = 0; i < m; i++) rowlens[i] = A->i[i + 1] - A->i[i];
644: PetscViewerBinaryWrite(viewer, rowlens, m, PETSC_INT);
645: PetscFree(rowlens);
646: /* store column indices */
647: PetscViewerBinaryWrite(viewer, A->j, nz, PETSC_INT);
648: /* store nonzero values */
649: MatSeqAIJGetArrayRead(mat, &av);
650: PetscViewerBinaryWrite(viewer, av, nz, PETSC_SCALAR);
651: MatSeqAIJRestoreArrayRead(mat, &av);
653: /* write block size option to the viewer's .info file */
654: MatView_Binary_BlockSizes(mat, viewer);
655: return 0;
656: }
658: static PetscErrorCode MatView_SeqAIJ_ASCII_structonly(Mat A, PetscViewer viewer)
659: {
660: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
661: PetscInt i, k, m = A->rmap->N;
663: PetscViewerASCIIUseTabs(viewer, PETSC_FALSE);
664: for (i = 0; i < m; i++) {
665: PetscViewerASCIIPrintf(viewer, "row %" PetscInt_FMT ":", i);
666: for (k = a->i[i]; k < a->i[i + 1]; k++) PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ") ", a->j[k]);
667: PetscViewerASCIIPrintf(viewer, "\n");
668: }
669: PetscViewerASCIIUseTabs(viewer, PETSC_TRUE);
670: return 0;
671: }
673: extern PetscErrorCode MatSeqAIJFactorInfo_Matlab(Mat, PetscViewer);
675: PetscErrorCode MatView_SeqAIJ_ASCII(Mat A, PetscViewer viewer)
676: {
677: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
678: const PetscScalar *av;
679: PetscInt i, j, m = A->rmap->n;
680: const char *name;
681: PetscViewerFormat format;
683: if (A->structure_only) {
684: MatView_SeqAIJ_ASCII_structonly(A, viewer);
685: return 0;
686: }
688: PetscViewerGetFormat(viewer, &format);
689: if (format == PETSC_VIEWER_ASCII_FACTOR_INFO || format == PETSC_VIEWER_ASCII_INFO || format == PETSC_VIEWER_ASCII_INFO_DETAIL) return 0;
691: /* trigger copy to CPU if needed */
692: MatSeqAIJGetArrayRead(A, &av);
693: MatSeqAIJRestoreArrayRead(A, &av);
694: if (format == PETSC_VIEWER_ASCII_MATLAB) {
695: PetscInt nofinalvalue = 0;
696: if (m && ((a->i[m] == a->i[m - 1]) || (a->j[a->nz - 1] != A->cmap->n - 1))) {
697: /* Need a dummy value to ensure the dimension of the matrix. */
698: nofinalvalue = 1;
699: }
700: PetscViewerASCIIUseTabs(viewer, PETSC_FALSE);
701: PetscViewerASCIIPrintf(viewer, "%% Size = %" PetscInt_FMT " %" PetscInt_FMT " \n", m, A->cmap->n);
702: PetscViewerASCIIPrintf(viewer, "%% Nonzeros = %" PetscInt_FMT " \n", a->nz);
703: #if defined(PETSC_USE_COMPLEX)
704: PetscViewerASCIIPrintf(viewer, "zzz = zeros(%" PetscInt_FMT ",4);\n", a->nz + nofinalvalue);
705: #else
706: PetscViewerASCIIPrintf(viewer, "zzz = zeros(%" PetscInt_FMT ",3);\n", a->nz + nofinalvalue);
707: #endif
708: PetscViewerASCIIPrintf(viewer, "zzz = [\n");
710: for (i = 0; i < m; i++) {
711: for (j = a->i[i]; j < a->i[i + 1]; j++) {
712: #if defined(PETSC_USE_COMPLEX)
713: 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]));
714: #else
715: PetscViewerASCIIPrintf(viewer, "%" PetscInt_FMT " %" PetscInt_FMT " %18.16e\n", i + 1, a->j[j] + 1, (double)a->a[j]);
716: #endif
717: }
718: }
719: if (nofinalvalue) {
720: #if defined(PETSC_USE_COMPLEX)
721: PetscViewerASCIIPrintf(viewer, "%" PetscInt_FMT " %" PetscInt_FMT " %18.16e %18.16e\n", m, A->cmap->n, 0., 0.);
722: #else
723: PetscViewerASCIIPrintf(viewer, "%" PetscInt_FMT " %" PetscInt_FMT " %18.16e\n", m, A->cmap->n, 0.0);
724: #endif
725: }
726: PetscObjectGetName((PetscObject)A, &name);
727: PetscViewerASCIIPrintf(viewer, "];\n %s = spconvert(zzz);\n", name);
728: PetscViewerASCIIUseTabs(viewer, PETSC_TRUE);
729: } else if (format == PETSC_VIEWER_ASCII_COMMON) {
730: PetscViewerASCIIUseTabs(viewer, PETSC_FALSE);
731: for (i = 0; i < m; i++) {
732: PetscViewerASCIIPrintf(viewer, "row %" PetscInt_FMT ":", i);
733: for (j = a->i[i]; j < a->i[i + 1]; j++) {
734: #if defined(PETSC_USE_COMPLEX)
735: if (PetscImaginaryPart(a->a[j]) > 0.0 && PetscRealPart(a->a[j]) != 0.0) {
736: PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g + %g i)", a->j[j], (double)PetscRealPart(a->a[j]), (double)PetscImaginaryPart(a->a[j]));
737: } else if (PetscImaginaryPart(a->a[j]) < 0.0 && PetscRealPart(a->a[j]) != 0.0) {
738: PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g - %g i)", a->j[j], (double)PetscRealPart(a->a[j]), (double)-PetscImaginaryPart(a->a[j]));
739: } else if (PetscRealPart(a->a[j]) != 0.0) {
740: PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g) ", a->j[j], (double)PetscRealPart(a->a[j]));
741: }
742: #else
743: if (a->a[j] != 0.0) PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g) ", a->j[j], (double)a->a[j]);
744: #endif
745: }
746: PetscViewerASCIIPrintf(viewer, "\n");
747: }
748: PetscViewerASCIIUseTabs(viewer, PETSC_TRUE);
749: } else if (format == PETSC_VIEWER_ASCII_SYMMODU) {
750: PetscInt nzd = 0, fshift = 1, *sptr;
751: PetscViewerASCIIUseTabs(viewer, PETSC_FALSE);
752: PetscMalloc1(m + 1, &sptr);
753: for (i = 0; i < m; i++) {
754: sptr[i] = nzd + 1;
755: for (j = a->i[i]; j < a->i[i + 1]; j++) {
756: if (a->j[j] >= i) {
757: #if defined(PETSC_USE_COMPLEX)
758: if (PetscImaginaryPart(a->a[j]) != 0.0 || PetscRealPart(a->a[j]) != 0.0) nzd++;
759: #else
760: if (a->a[j] != 0.0) nzd++;
761: #endif
762: }
763: }
764: }
765: sptr[m] = nzd + 1;
766: PetscViewerASCIIPrintf(viewer, " %" PetscInt_FMT " %" PetscInt_FMT "\n\n", m, nzd);
767: for (i = 0; i < m + 1; i += 6) {
768: if (i + 4 < m) {
769: 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]);
770: } else if (i + 3 < m) {
771: 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]);
772: } else if (i + 2 < m) {
773: PetscViewerASCIIPrintf(viewer, " %" PetscInt_FMT " %" PetscInt_FMT " %" PetscInt_FMT " %" PetscInt_FMT "\n", sptr[i], sptr[i + 1], sptr[i + 2], sptr[i + 3]);
774: } else if (i + 1 < m) {
775: PetscViewerASCIIPrintf(viewer, " %" PetscInt_FMT " %" PetscInt_FMT " %" PetscInt_FMT "\n", sptr[i], sptr[i + 1], sptr[i + 2]);
776: } else if (i < m) {
777: PetscViewerASCIIPrintf(viewer, " %" PetscInt_FMT " %" PetscInt_FMT "\n", sptr[i], sptr[i + 1]);
778: } else {
779: PetscViewerASCIIPrintf(viewer, " %" PetscInt_FMT "\n", sptr[i]);
780: }
781: }
782: PetscViewerASCIIPrintf(viewer, "\n");
783: PetscFree(sptr);
784: for (i = 0; i < m; i++) {
785: for (j = a->i[i]; j < a->i[i + 1]; j++) {
786: if (a->j[j] >= i) PetscViewerASCIIPrintf(viewer, " %" PetscInt_FMT " ", a->j[j] + fshift);
787: }
788: PetscViewerASCIIPrintf(viewer, "\n");
789: }
790: PetscViewerASCIIPrintf(viewer, "\n");
791: for (i = 0; i < m; i++) {
792: for (j = a->i[i]; j < a->i[i + 1]; j++) {
793: if (a->j[j] >= i) {
794: #if defined(PETSC_USE_COMPLEX)
795: if (PetscImaginaryPart(a->a[j]) != 0.0 || PetscRealPart(a->a[j]) != 0.0) PetscViewerASCIIPrintf(viewer, " %18.16e %18.16e ", (double)PetscRealPart(a->a[j]), (double)PetscImaginaryPart(a->a[j]));
796: #else
797: if (a->a[j] != 0.0) PetscViewerASCIIPrintf(viewer, " %18.16e ", (double)a->a[j]);
798: #endif
799: }
800: }
801: PetscViewerASCIIPrintf(viewer, "\n");
802: }
803: PetscViewerASCIIUseTabs(viewer, PETSC_TRUE);
804: } else if (format == PETSC_VIEWER_ASCII_DENSE) {
805: PetscInt cnt = 0, jcnt;
806: PetscScalar value;
807: #if defined(PETSC_USE_COMPLEX)
808: PetscBool realonly = PETSC_TRUE;
810: for (i = 0; i < a->i[m]; i++) {
811: if (PetscImaginaryPart(a->a[i]) != 0.0) {
812: realonly = PETSC_FALSE;
813: break;
814: }
815: }
816: #endif
818: PetscViewerASCIIUseTabs(viewer, PETSC_FALSE);
819: for (i = 0; i < m; i++) {
820: jcnt = 0;
821: for (j = 0; j < A->cmap->n; j++) {
822: if (jcnt < a->i[i + 1] - a->i[i] && j == a->j[cnt]) {
823: value = a->a[cnt++];
824: jcnt++;
825: } else {
826: value = 0.0;
827: }
828: #if defined(PETSC_USE_COMPLEX)
829: if (realonly) {
830: PetscViewerASCIIPrintf(viewer, " %7.5e ", (double)PetscRealPart(value));
831: } else {
832: PetscViewerASCIIPrintf(viewer, " %7.5e+%7.5e i ", (double)PetscRealPart(value), (double)PetscImaginaryPart(value));
833: }
834: #else
835: PetscViewerASCIIPrintf(viewer, " %7.5e ", (double)value);
836: #endif
837: }
838: PetscViewerASCIIPrintf(viewer, "\n");
839: }
840: PetscViewerASCIIUseTabs(viewer, PETSC_TRUE);
841: } else if (format == PETSC_VIEWER_ASCII_MATRIXMARKET) {
842: PetscInt fshift = 1;
843: PetscViewerASCIIUseTabs(viewer, PETSC_FALSE);
844: #if defined(PETSC_USE_COMPLEX)
845: PetscViewerASCIIPrintf(viewer, "%%%%MatrixMarket matrix coordinate complex general\n");
846: #else
847: PetscViewerASCIIPrintf(viewer, "%%%%MatrixMarket matrix coordinate real general\n");
848: #endif
849: PetscViewerASCIIPrintf(viewer, "%" PetscInt_FMT " %" PetscInt_FMT " %" PetscInt_FMT "\n", m, A->cmap->n, a->nz);
850: for (i = 0; i < m; i++) {
851: for (j = a->i[i]; j < a->i[i + 1]; j++) {
852: #if defined(PETSC_USE_COMPLEX)
853: 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]));
854: #else
855: PetscViewerASCIIPrintf(viewer, "%" PetscInt_FMT " %" PetscInt_FMT " %g\n", i + fshift, a->j[j] + fshift, (double)a->a[j]);
856: #endif
857: }
858: }
859: PetscViewerASCIIUseTabs(viewer, PETSC_TRUE);
860: } else {
861: PetscViewerASCIIUseTabs(viewer, PETSC_FALSE);
862: if (A->factortype) {
863: for (i = 0; i < m; i++) {
864: PetscViewerASCIIPrintf(viewer, "row %" PetscInt_FMT ":", i);
865: /* L part */
866: for (j = a->i[i]; j < a->i[i + 1]; j++) {
867: #if defined(PETSC_USE_COMPLEX)
868: if (PetscImaginaryPart(a->a[j]) > 0.0) {
869: PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g + %g i)", a->j[j], (double)PetscRealPart(a->a[j]), (double)PetscImaginaryPart(a->a[j]));
870: } else if (PetscImaginaryPart(a->a[j]) < 0.0) {
871: PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g - %g i)", a->j[j], (double)PetscRealPart(a->a[j]), (double)(-PetscImaginaryPart(a->a[j])));
872: } else {
873: PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g) ", a->j[j], (double)PetscRealPart(a->a[j]));
874: }
875: #else
876: PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g) ", a->j[j], (double)a->a[j]);
877: #endif
878: }
879: /* diagonal */
880: j = a->diag[i];
881: #if defined(PETSC_USE_COMPLEX)
882: if (PetscImaginaryPart(a->a[j]) > 0.0) {
883: PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g + %g i)", a->j[j], (double)PetscRealPart(1.0 / a->a[j]), (double)PetscImaginaryPart(1.0 / a->a[j]));
884: } else if (PetscImaginaryPart(a->a[j]) < 0.0) {
885: PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g - %g i)", a->j[j], (double)PetscRealPart(1.0 / a->a[j]), (double)(-PetscImaginaryPart(1.0 / a->a[j])));
886: } else {
887: PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g) ", a->j[j], (double)PetscRealPart(1.0 / a->a[j]));
888: }
889: #else
890: PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g) ", a->j[j], (double)(1.0 / a->a[j]));
891: #endif
893: /* U part */
894: for (j = a->diag[i + 1] + 1; j < a->diag[i]; j++) {
895: #if defined(PETSC_USE_COMPLEX)
896: if (PetscImaginaryPart(a->a[j]) > 0.0) {
897: PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g + %g i)", a->j[j], (double)PetscRealPart(a->a[j]), (double)PetscImaginaryPart(a->a[j]));
898: } else if (PetscImaginaryPart(a->a[j]) < 0.0) {
899: PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g - %g i)", a->j[j], (double)PetscRealPart(a->a[j]), (double)(-PetscImaginaryPart(a->a[j])));
900: } else {
901: PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g) ", a->j[j], (double)PetscRealPart(a->a[j]));
902: }
903: #else
904: PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g) ", a->j[j], (double)a->a[j]);
905: #endif
906: }
907: PetscViewerASCIIPrintf(viewer, "\n");
908: }
909: } else {
910: for (i = 0; i < m; i++) {
911: PetscViewerASCIIPrintf(viewer, "row %" PetscInt_FMT ":", i);
912: for (j = a->i[i]; j < a->i[i + 1]; j++) {
913: #if defined(PETSC_USE_COMPLEX)
914: if (PetscImaginaryPart(a->a[j]) > 0.0) {
915: PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g + %g i)", a->j[j], (double)PetscRealPart(a->a[j]), (double)PetscImaginaryPart(a->a[j]));
916: } else if (PetscImaginaryPart(a->a[j]) < 0.0) {
917: PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g - %g i)", a->j[j], (double)PetscRealPart(a->a[j]), (double)-PetscImaginaryPart(a->a[j]));
918: } else {
919: PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g) ", a->j[j], (double)PetscRealPart(a->a[j]));
920: }
921: #else
922: PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g) ", a->j[j], (double)a->a[j]);
923: #endif
924: }
925: PetscViewerASCIIPrintf(viewer, "\n");
926: }
927: }
928: PetscViewerASCIIUseTabs(viewer, PETSC_TRUE);
929: }
930: PetscViewerFlush(viewer);
931: return 0;
932: }
934: #include <petscdraw.h>
935: PetscErrorCode MatView_SeqAIJ_Draw_Zoom(PetscDraw draw, void *Aa)
936: {
937: Mat A = (Mat)Aa;
938: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
939: PetscInt i, j, m = A->rmap->n;
940: int color;
941: PetscReal xl, yl, xr, yr, x_l, x_r, y_l, y_r;
942: PetscViewer viewer;
943: PetscViewerFormat format;
944: const PetscScalar *aa;
946: PetscObjectQuery((PetscObject)A, "Zoomviewer", (PetscObject *)&viewer);
947: PetscViewerGetFormat(viewer, &format);
948: PetscDrawGetCoordinates(draw, &xl, &yl, &xr, &yr);
950: /* loop over matrix elements drawing boxes */
951: MatSeqAIJGetArrayRead(A, &aa);
952: if (format != PETSC_VIEWER_DRAW_CONTOUR) {
953: PetscDrawCollectiveBegin(draw);
954: /* Blue for negative, Cyan for zero and Red for positive */
955: color = PETSC_DRAW_BLUE;
956: for (i = 0; i < m; i++) {
957: y_l = m - i - 1.0;
958: y_r = y_l + 1.0;
959: for (j = a->i[i]; j < a->i[i + 1]; j++) {
960: x_l = a->j[j];
961: x_r = x_l + 1.0;
962: if (PetscRealPart(aa[j]) >= 0.) continue;
963: PetscDrawRectangle(draw, x_l, y_l, x_r, y_r, color, color, color, color);
964: }
965: }
966: color = PETSC_DRAW_CYAN;
967: for (i = 0; i < m; i++) {
968: y_l = m - i - 1.0;
969: y_r = y_l + 1.0;
970: for (j = a->i[i]; j < a->i[i + 1]; j++) {
971: x_l = a->j[j];
972: x_r = x_l + 1.0;
973: if (aa[j] != 0.) continue;
974: PetscDrawRectangle(draw, x_l, y_l, x_r, y_r, color, color, color, color);
975: }
976: }
977: color = PETSC_DRAW_RED;
978: for (i = 0; i < m; i++) {
979: y_l = m - i - 1.0;
980: y_r = y_l + 1.0;
981: for (j = a->i[i]; j < a->i[i + 1]; j++) {
982: x_l = a->j[j];
983: x_r = x_l + 1.0;
984: if (PetscRealPart(aa[j]) <= 0.) continue;
985: PetscDrawRectangle(draw, x_l, y_l, x_r, y_r, color, color, color, color);
986: }
987: }
988: PetscDrawCollectiveEnd(draw);
989: } else {
990: /* use contour shading to indicate magnitude of values */
991: /* first determine max of all nonzero values */
992: PetscReal minv = 0.0, maxv = 0.0;
993: PetscInt nz = a->nz, count = 0;
994: PetscDraw popup;
996: for (i = 0; i < nz; i++) {
997: if (PetscAbsScalar(aa[i]) > maxv) maxv = PetscAbsScalar(aa[i]);
998: }
999: if (minv >= maxv) maxv = minv + PETSC_SMALL;
1000: PetscDrawGetPopup(draw, &popup);
1001: PetscDrawScalePopup(popup, minv, maxv);
1003: PetscDrawCollectiveBegin(draw);
1004: for (i = 0; i < m; i++) {
1005: y_l = m - i - 1.0;
1006: y_r = y_l + 1.0;
1007: for (j = a->i[i]; j < a->i[i + 1]; j++) {
1008: x_l = a->j[j];
1009: x_r = x_l + 1.0;
1010: color = PetscDrawRealToColor(PetscAbsScalar(aa[count]), minv, maxv);
1011: PetscDrawRectangle(draw, x_l, y_l, x_r, y_r, color, color, color, color);
1012: count++;
1013: }
1014: }
1015: PetscDrawCollectiveEnd(draw);
1016: }
1017: MatSeqAIJRestoreArrayRead(A, &aa);
1018: return 0;
1019: }
1021: #include <petscdraw.h>
1022: PetscErrorCode MatView_SeqAIJ_Draw(Mat A, PetscViewer viewer)
1023: {
1024: PetscDraw draw;
1025: PetscReal xr, yr, xl, yl, h, w;
1026: PetscBool isnull;
1028: PetscViewerDrawGetDraw(viewer, 0, &draw);
1029: PetscDrawIsNull(draw, &isnull);
1030: if (isnull) return 0;
1032: xr = A->cmap->n;
1033: yr = A->rmap->n;
1034: h = yr / 10.0;
1035: w = xr / 10.0;
1036: xr += w;
1037: yr += h;
1038: xl = -w;
1039: yl = -h;
1040: PetscDrawSetCoordinates(draw, xl, yl, xr, yr);
1041: PetscObjectCompose((PetscObject)A, "Zoomviewer", (PetscObject)viewer);
1042: PetscDrawZoom(draw, MatView_SeqAIJ_Draw_Zoom, A);
1043: PetscObjectCompose((PetscObject)A, "Zoomviewer", NULL);
1044: PetscDrawSave(draw);
1045: return 0;
1046: }
1048: PetscErrorCode MatView_SeqAIJ(Mat A, PetscViewer viewer)
1049: {
1050: PetscBool iascii, isbinary, isdraw;
1052: PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERASCII, &iascii);
1053: PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERBINARY, &isbinary);
1054: PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERDRAW, &isdraw);
1055: if (iascii) MatView_SeqAIJ_ASCII(A, viewer);
1056: else if (isbinary) MatView_SeqAIJ_Binary(A, viewer);
1057: else if (isdraw) MatView_SeqAIJ_Draw(A, viewer);
1058: MatView_SeqAIJ_Inode(A, viewer);
1059: return 0;
1060: }
1062: PetscErrorCode MatAssemblyEnd_SeqAIJ(Mat A, MatAssemblyType mode)
1063: {
1064: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
1065: PetscInt fshift = 0, i, *ai = a->i, *aj = a->j, *imax = a->imax;
1066: PetscInt m = A->rmap->n, *ip, N, *ailen = a->ilen, rmax = 0;
1067: MatScalar *aa = a->a, *ap;
1068: PetscReal ratio = 0.6;
1070: if (mode == MAT_FLUSH_ASSEMBLY) return 0;
1071: MatSeqAIJInvalidateDiagonal(A);
1072: if (A->was_assembled && A->ass_nonzerostate == A->nonzerostate) {
1073: /* we need to respect users asking to use or not the inodes routine in between matrix assemblies */
1074: MatAssemblyEnd_SeqAIJ_Inode(A, mode);
1075: return 0;
1076: }
1078: if (m) rmax = ailen[0]; /* determine row with most nonzeros */
1079: for (i = 1; i < m; i++) {
1080: /* move each row back by the amount of empty slots (fshift) before it*/
1081: fshift += imax[i - 1] - ailen[i - 1];
1082: rmax = PetscMax(rmax, ailen[i]);
1083: if (fshift) {
1084: ip = aj + ai[i];
1085: ap = aa + ai[i];
1086: N = ailen[i];
1087: PetscArraymove(ip - fshift, ip, N);
1088: if (!A->structure_only) PetscArraymove(ap - fshift, ap, N);
1089: }
1090: ai[i] = ai[i - 1] + ailen[i - 1];
1091: }
1092: if (m) {
1093: fshift += imax[m - 1] - ailen[m - 1];
1094: ai[m] = ai[m - 1] + ailen[m - 1];
1095: }
1097: /* reset ilen and imax for each row */
1098: a->nonzerorowcnt = 0;
1099: if (A->structure_only) {
1100: PetscFree(a->imax);
1101: PetscFree(a->ilen);
1102: } else { /* !A->structure_only */
1103: for (i = 0; i < m; i++) {
1104: ailen[i] = imax[i] = ai[i + 1] - ai[i];
1105: a->nonzerorowcnt += ((ai[i + 1] - ai[i]) > 0);
1106: }
1107: }
1108: a->nz = ai[m];
1111: MatMarkDiagonal_SeqAIJ(A);
1112: 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);
1113: PetscInfo(A, "Number of mallocs during MatSetValues() is %" PetscInt_FMT "\n", a->reallocs);
1114: PetscInfo(A, "Maximum nonzeros in any row is %" PetscInt_FMT "\n", rmax);
1116: A->info.mallocs += a->reallocs;
1117: a->reallocs = 0;
1118: A->info.nz_unneeded = (PetscReal)fshift;
1119: a->rmax = rmax;
1121: if (!A->structure_only) MatCheckCompressedRow(A, a->nonzerorowcnt, &a->compressedrow, a->i, m, ratio);
1122: MatAssemblyEnd_SeqAIJ_Inode(A, mode);
1123: return 0;
1124: }
1126: PetscErrorCode MatRealPart_SeqAIJ(Mat A)
1127: {
1128: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
1129: PetscInt i, nz = a->nz;
1130: MatScalar *aa;
1132: MatSeqAIJGetArray(A, &aa);
1133: for (i = 0; i < nz; i++) aa[i] = PetscRealPart(aa[i]);
1134: MatSeqAIJRestoreArray(A, &aa);
1135: MatSeqAIJInvalidateDiagonal(A);
1136: return 0;
1137: }
1139: PetscErrorCode MatImaginaryPart_SeqAIJ(Mat A)
1140: {
1141: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
1142: PetscInt i, nz = a->nz;
1143: MatScalar *aa;
1145: MatSeqAIJGetArray(A, &aa);
1146: for (i = 0; i < nz; i++) aa[i] = PetscImaginaryPart(aa[i]);
1147: MatSeqAIJRestoreArray(A, &aa);
1148: MatSeqAIJInvalidateDiagonal(A);
1149: return 0;
1150: }
1152: PetscErrorCode MatZeroEntries_SeqAIJ(Mat A)
1153: {
1154: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
1155: MatScalar *aa;
1157: MatSeqAIJGetArrayWrite(A, &aa);
1158: PetscArrayzero(aa, a->i[A->rmap->n]);
1159: MatSeqAIJRestoreArrayWrite(A, &aa);
1160: MatSeqAIJInvalidateDiagonal(A);
1161: return 0;
1162: }
1164: PETSC_INTERN PetscErrorCode MatResetPreallocationCOO_SeqAIJ(Mat A)
1165: {
1166: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
1168: PetscFree(a->perm);
1169: PetscFree(a->jmap);
1170: return 0;
1171: }
1173: PetscErrorCode MatDestroy_SeqAIJ(Mat A)
1174: {
1175: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
1177: #if defined(PETSC_USE_LOG)
1178: PetscLogObjectState((PetscObject)A, "Rows=%" PetscInt_FMT ", Cols=%" PetscInt_FMT ", NZ=%" PetscInt_FMT, A->rmap->n, A->cmap->n, a->nz);
1179: #endif
1180: MatSeqXAIJFreeAIJ(A, &a->a, &a->j, &a->i);
1181: MatResetPreallocationCOO_SeqAIJ(A);
1182: ISDestroy(&a->row);
1183: ISDestroy(&a->col);
1184: PetscFree(a->diag);
1185: PetscFree(a->ibdiag);
1186: PetscFree(a->imax);
1187: PetscFree(a->ilen);
1188: PetscFree(a->ipre);
1189: PetscFree3(a->idiag, a->mdiag, a->ssor_work);
1190: PetscFree(a->solve_work);
1191: ISDestroy(&a->icol);
1192: PetscFree(a->saved_values);
1193: PetscFree2(a->compressedrow.i, a->compressedrow.rindex);
1194: MatDestroy_SeqAIJ_Inode(A);
1195: PetscFree(A->data);
1197: /* MatMatMultNumeric_SeqAIJ_SeqAIJ_Sorted may allocate this.
1198: That function is so heavily used (sometimes in an hidden way through multnumeric function pointers)
1199: that is hard to properly add this data to the MatProduct data. We free it here to avoid
1200: users reusing the matrix object with different data to incur in obscure segmentation faults
1201: due to different matrix sizes */
1202: PetscObjectCompose((PetscObject)A, "__PETSc__ab_dense", NULL);
1204: PetscObjectChangeTypeName((PetscObject)A, NULL);
1205: PetscObjectComposeFunction((PetscObject)A, "PetscMatlabEnginePut_C", NULL);
1206: PetscObjectComposeFunction((PetscObject)A, "PetscMatlabEngineGet_C", NULL);
1207: PetscObjectComposeFunction((PetscObject)A, "MatSeqAIJSetColumnIndices_C", NULL);
1208: PetscObjectComposeFunction((PetscObject)A, "MatStoreValues_C", NULL);
1209: PetscObjectComposeFunction((PetscObject)A, "MatRetrieveValues_C", NULL);
1210: PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_seqsbaij_C", NULL);
1211: PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_seqbaij_C", NULL);
1212: PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_seqaijperm_C", NULL);
1213: PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_seqaijsell_C", NULL);
1214: #if defined(PETSC_HAVE_MKL_SPARSE)
1215: PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_seqaijmkl_C", NULL);
1216: #endif
1217: #if defined(PETSC_HAVE_CUDA)
1218: PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_seqaijcusparse_C", NULL);
1219: PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_seqaijcusparse_seqaij_C", NULL);
1220: PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_seqaij_seqaijcusparse_C", NULL);
1221: #endif
1222: #if defined(PETSC_HAVE_KOKKOS_KERNELS)
1223: PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_seqaijkokkos_C", NULL);
1224: #endif
1225: PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_seqaijcrl_C", NULL);
1226: #if defined(PETSC_HAVE_ELEMENTAL)
1227: PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_elemental_C", NULL);
1228: #endif
1229: #if defined(PETSC_HAVE_SCALAPACK)
1230: PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_scalapack_C", NULL);
1231: #endif
1232: #if defined(PETSC_HAVE_HYPRE)
1233: PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_hypre_C", NULL);
1234: PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_transpose_seqaij_seqaij_C", NULL);
1235: #endif
1236: PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_seqdense_C", NULL);
1237: PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_seqsell_C", NULL);
1238: PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_is_C", NULL);
1239: PetscObjectComposeFunction((PetscObject)A, "MatIsTranspose_C", NULL);
1240: PetscObjectComposeFunction((PetscObject)A, "MatIsHermitianTranspose_C", NULL);
1241: PetscObjectComposeFunction((PetscObject)A, "MatSeqAIJSetPreallocation_C", NULL);
1242: PetscObjectComposeFunction((PetscObject)A, "MatResetPreallocation_C", NULL);
1243: PetscObjectComposeFunction((PetscObject)A, "MatSeqAIJSetPreallocationCSR_C", NULL);
1244: PetscObjectComposeFunction((PetscObject)A, "MatReorderForNonzeroDiagonal_C", NULL);
1245: PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_is_seqaij_C", NULL);
1246: PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_seqdense_seqaij_C", NULL);
1247: PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_seqaij_seqaij_C", NULL);
1248: PetscObjectComposeFunction((PetscObject)A, "MatSeqAIJKron_C", NULL);
1249: PetscObjectComposeFunction((PetscObject)A, "MatSetPreallocationCOO_C", NULL);
1250: PetscObjectComposeFunction((PetscObject)A, "MatSetValuesCOO_C", NULL);
1251: PetscObjectComposeFunction((PetscObject)A, "MatFactorGetSolverType_C", NULL);
1252: /* these calls do not belong here: the subclasses Duplicate/Destroy are wrong */
1253: PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaijsell_seqaij_C", NULL);
1254: PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaijperm_seqaij_C", NULL);
1255: PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_seqaijviennacl_C", NULL);
1256: PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_seqaijviennacl_seqdense_C", NULL);
1257: PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_seqaijviennacl_seqaij_C", NULL);
1258: return 0;
1259: }
1261: PetscErrorCode MatSetOption_SeqAIJ(Mat A, MatOption op, PetscBool flg)
1262: {
1263: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
1265: switch (op) {
1266: case MAT_ROW_ORIENTED:
1267: a->roworiented = flg;
1268: break;
1269: case MAT_KEEP_NONZERO_PATTERN:
1270: a->keepnonzeropattern = flg;
1271: break;
1272: case MAT_NEW_NONZERO_LOCATIONS:
1273: a->nonew = (flg ? 0 : 1);
1274: break;
1275: case MAT_NEW_NONZERO_LOCATION_ERR:
1276: a->nonew = (flg ? -1 : 0);
1277: break;
1278: case MAT_NEW_NONZERO_ALLOCATION_ERR:
1279: a->nonew = (flg ? -2 : 0);
1280: break;
1281: case MAT_UNUSED_NONZERO_LOCATION_ERR:
1282: a->nounused = (flg ? -1 : 0);
1283: break;
1284: case MAT_IGNORE_ZERO_ENTRIES:
1285: a->ignorezeroentries = flg;
1286: break;
1287: case MAT_SPD:
1288: case MAT_SYMMETRIC:
1289: case MAT_STRUCTURALLY_SYMMETRIC:
1290: case MAT_HERMITIAN:
1291: case MAT_SYMMETRY_ETERNAL:
1292: case MAT_STRUCTURE_ONLY:
1293: case MAT_STRUCTURAL_SYMMETRY_ETERNAL:
1294: case MAT_SPD_ETERNAL:
1295: /* if the diagonal matrix is square it inherits some of the properties above */
1296: break;
1297: case MAT_FORCE_DIAGONAL_ENTRIES:
1298: case MAT_IGNORE_OFF_PROC_ENTRIES:
1299: case MAT_USE_HASH_TABLE:
1300: PetscInfo(A, "Option %s ignored\n", MatOptions[op]);
1301: break;
1302: case MAT_USE_INODES:
1303: MatSetOption_SeqAIJ_Inode(A, MAT_USE_INODES, flg);
1304: break;
1305: case MAT_SUBMAT_SINGLEIS:
1306: A->submat_singleis = flg;
1307: break;
1308: case MAT_SORTED_FULL:
1309: if (flg) A->ops->setvalues = MatSetValues_SeqAIJ_SortedFull;
1310: else A->ops->setvalues = MatSetValues_SeqAIJ;
1311: break;
1312: case MAT_FORM_EXPLICIT_TRANSPOSE:
1313: A->form_explicit_transpose = flg;
1314: break;
1315: default:
1316: SETERRQ(PETSC_COMM_SELF, PETSC_ERR_SUP, "unknown option %d", op);
1317: }
1318: return 0;
1319: }
1321: PetscErrorCode MatGetDiagonal_SeqAIJ(Mat A, Vec v)
1322: {
1323: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
1324: PetscInt i, j, n, *ai = a->i, *aj = a->j;
1325: PetscScalar *x;
1326: const PetscScalar *aa;
1328: VecGetLocalSize(v, &n);
1330: MatSeqAIJGetArrayRead(A, &aa);
1331: if (A->factortype == MAT_FACTOR_ILU || A->factortype == MAT_FACTOR_LU) {
1332: PetscInt *diag = a->diag;
1333: VecGetArrayWrite(v, &x);
1334: for (i = 0; i < n; i++) x[i] = 1.0 / aa[diag[i]];
1335: VecRestoreArrayWrite(v, &x);
1336: MatSeqAIJRestoreArrayRead(A, &aa);
1337: return 0;
1338: }
1340: VecGetArrayWrite(v, &x);
1341: for (i = 0; i < n; i++) {
1342: x[i] = 0.0;
1343: for (j = ai[i]; j < ai[i + 1]; j++) {
1344: if (aj[j] == i) {
1345: x[i] = aa[j];
1346: break;
1347: }
1348: }
1349: }
1350: VecRestoreArrayWrite(v, &x);
1351: MatSeqAIJRestoreArrayRead(A, &aa);
1352: return 0;
1353: }
1355: #include <../src/mat/impls/aij/seq/ftn-kernels/fmult.h>
1356: PetscErrorCode MatMultTransposeAdd_SeqAIJ(Mat A, Vec xx, Vec zz, Vec yy)
1357: {
1358: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
1359: const MatScalar *aa;
1360: PetscScalar *y;
1361: const PetscScalar *x;
1362: PetscInt m = A->rmap->n;
1363: #if !defined(PETSC_USE_FORTRAN_KERNEL_MULTTRANSPOSEAIJ)
1364: const MatScalar *v;
1365: PetscScalar alpha;
1366: PetscInt n, i, j;
1367: const PetscInt *idx, *ii, *ridx = NULL;
1368: Mat_CompressedRow cprow = a->compressedrow;
1369: PetscBool usecprow = cprow.use;
1370: #endif
1372: if (zz != yy) VecCopy(zz, yy);
1373: VecGetArrayRead(xx, &x);
1374: VecGetArray(yy, &y);
1375: MatSeqAIJGetArrayRead(A, &aa);
1377: #if defined(PETSC_USE_FORTRAN_KERNEL_MULTTRANSPOSEAIJ)
1378: fortranmulttransposeaddaij_(&m, x, a->i, a->j, aa, y);
1379: #else
1380: if (usecprow) {
1381: m = cprow.nrows;
1382: ii = cprow.i;
1383: ridx = cprow.rindex;
1384: } else {
1385: ii = a->i;
1386: }
1387: for (i = 0; i < m; i++) {
1388: idx = a->j + ii[i];
1389: v = aa + ii[i];
1390: n = ii[i + 1] - ii[i];
1391: if (usecprow) {
1392: alpha = x[ridx[i]];
1393: } else {
1394: alpha = x[i];
1395: }
1396: for (j = 0; j < n; j++) y[idx[j]] += alpha * v[j];
1397: }
1398: #endif
1399: PetscLogFlops(2.0 * a->nz);
1400: VecRestoreArrayRead(xx, &x);
1401: VecRestoreArray(yy, &y);
1402: MatSeqAIJRestoreArrayRead(A, &aa);
1403: return 0;
1404: }
1406: PetscErrorCode MatMultTranspose_SeqAIJ(Mat A, Vec xx, Vec yy)
1407: {
1408: VecSet(yy, 0.0);
1409: MatMultTransposeAdd_SeqAIJ(A, xx, yy, yy);
1410: return 0;
1411: }
1413: #include <../src/mat/impls/aij/seq/ftn-kernels/fmult.h>
1415: PetscErrorCode MatMult_SeqAIJ(Mat A, Vec xx, Vec yy)
1416: {
1417: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
1418: PetscScalar *y;
1419: const PetscScalar *x;
1420: const MatScalar *aa, *a_a;
1421: PetscInt m = A->rmap->n;
1422: const PetscInt *aj, *ii, *ridx = NULL;
1423: PetscInt n, i;
1424: PetscScalar sum;
1425: PetscBool usecprow = a->compressedrow.use;
1427: #if defined(PETSC_HAVE_PRAGMA_DISJOINT)
1428: #pragma disjoint(*x, *y, *aa)
1429: #endif
1431: if (a->inode.use && a->inode.checked) {
1432: MatMult_SeqAIJ_Inode(A, xx, yy);
1433: return 0;
1434: }
1435: MatSeqAIJGetArrayRead(A, &a_a);
1436: VecGetArrayRead(xx, &x);
1437: VecGetArray(yy, &y);
1438: ii = a->i;
1439: if (usecprow) { /* use compressed row format */
1440: PetscArrayzero(y, m);
1441: m = a->compressedrow.nrows;
1442: ii = a->compressedrow.i;
1443: ridx = a->compressedrow.rindex;
1444: for (i = 0; i < m; i++) {
1445: n = ii[i + 1] - ii[i];
1446: aj = a->j + ii[i];
1447: aa = a_a + ii[i];
1448: sum = 0.0;
1449: PetscSparseDensePlusDot(sum, x, aa, aj, n);
1450: /* for (j=0; j<n; j++) sum += (*aa++)*x[*aj++]; */
1451: y[*ridx++] = sum;
1452: }
1453: } else { /* do not use compressed row format */
1454: #if defined(PETSC_USE_FORTRAN_KERNEL_MULTAIJ)
1455: aj = a->j;
1456: aa = a_a;
1457: fortranmultaij_(&m, x, ii, aj, aa, y);
1458: #else
1459: for (i = 0; i < m; i++) {
1460: n = ii[i + 1] - ii[i];
1461: aj = a->j + ii[i];
1462: aa = a_a + ii[i];
1463: sum = 0.0;
1464: PetscSparseDensePlusDot(sum, x, aa, aj, n);
1465: y[i] = sum;
1466: }
1467: #endif
1468: }
1469: PetscLogFlops(2.0 * a->nz - a->nonzerorowcnt);
1470: VecRestoreArrayRead(xx, &x);
1471: VecRestoreArray(yy, &y);
1472: MatSeqAIJRestoreArrayRead(A, &a_a);
1473: return 0;
1474: }
1476: PetscErrorCode MatMultMax_SeqAIJ(Mat A, Vec xx, Vec yy)
1477: {
1478: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
1479: PetscScalar *y;
1480: const PetscScalar *x;
1481: const MatScalar *aa, *a_a;
1482: PetscInt m = A->rmap->n;
1483: const PetscInt *aj, *ii, *ridx = NULL;
1484: PetscInt n, i, nonzerorow = 0;
1485: PetscScalar sum;
1486: PetscBool usecprow = a->compressedrow.use;
1488: #if defined(PETSC_HAVE_PRAGMA_DISJOINT)
1489: #pragma disjoint(*x, *y, *aa)
1490: #endif
1492: MatSeqAIJGetArrayRead(A, &a_a);
1493: VecGetArrayRead(xx, &x);
1494: VecGetArray(yy, &y);
1495: if (usecprow) { /* use compressed row format */
1496: m = a->compressedrow.nrows;
1497: ii = a->compressedrow.i;
1498: ridx = a->compressedrow.rindex;
1499: for (i = 0; i < m; i++) {
1500: n = ii[i + 1] - ii[i];
1501: aj = a->j + ii[i];
1502: aa = a_a + ii[i];
1503: sum = 0.0;
1504: nonzerorow += (n > 0);
1505: PetscSparseDenseMaxDot(sum, x, aa, aj, n);
1506: /* for (j=0; j<n; j++) sum += (*aa++)*x[*aj++]; */
1507: y[*ridx++] = sum;
1508: }
1509: } else { /* do not use compressed row format */
1510: ii = a->i;
1511: for (i = 0; i < m; i++) {
1512: n = ii[i + 1] - ii[i];
1513: aj = a->j + ii[i];
1514: aa = a_a + ii[i];
1515: sum = 0.0;
1516: nonzerorow += (n > 0);
1517: PetscSparseDenseMaxDot(sum, x, aa, aj, n);
1518: y[i] = sum;
1519: }
1520: }
1521: PetscLogFlops(2.0 * a->nz - nonzerorow);
1522: VecRestoreArrayRead(xx, &x);
1523: VecRestoreArray(yy, &y);
1524: MatSeqAIJRestoreArrayRead(A, &a_a);
1525: return 0;
1526: }
1528: PetscErrorCode MatMultAddMax_SeqAIJ(Mat A, Vec xx, Vec yy, Vec zz)
1529: {
1530: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
1531: PetscScalar *y, *z;
1532: const PetscScalar *x;
1533: const MatScalar *aa, *a_a;
1534: PetscInt m = A->rmap->n, *aj, *ii;
1535: PetscInt n, i, *ridx = NULL;
1536: PetscScalar sum;
1537: PetscBool usecprow = a->compressedrow.use;
1539: MatSeqAIJGetArrayRead(A, &a_a);
1540: VecGetArrayRead(xx, &x);
1541: VecGetArrayPair(yy, zz, &y, &z);
1542: if (usecprow) { /* use compressed row format */
1543: if (zz != yy) PetscArraycpy(z, y, m);
1544: m = a->compressedrow.nrows;
1545: ii = a->compressedrow.i;
1546: ridx = a->compressedrow.rindex;
1547: for (i = 0; i < m; i++) {
1548: n = ii[i + 1] - ii[i];
1549: aj = a->j + ii[i];
1550: aa = a_a + ii[i];
1551: sum = y[*ridx];
1552: PetscSparseDenseMaxDot(sum, x, aa, aj, n);
1553: z[*ridx++] = sum;
1554: }
1555: } else { /* do not use compressed row format */
1556: ii = a->i;
1557: for (i = 0; i < m; i++) {
1558: n = ii[i + 1] - ii[i];
1559: aj = a->j + ii[i];
1560: aa = a_a + ii[i];
1561: sum = y[i];
1562: PetscSparseDenseMaxDot(sum, x, aa, aj, n);
1563: z[i] = sum;
1564: }
1565: }
1566: PetscLogFlops(2.0 * a->nz);
1567: VecRestoreArrayRead(xx, &x);
1568: VecRestoreArrayPair(yy, zz, &y, &z);
1569: MatSeqAIJRestoreArrayRead(A, &a_a);
1570: return 0;
1571: }
1573: #include <../src/mat/impls/aij/seq/ftn-kernels/fmultadd.h>
1574: PetscErrorCode MatMultAdd_SeqAIJ(Mat A, Vec xx, Vec yy, Vec zz)
1575: {
1576: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
1577: PetscScalar *y, *z;
1578: const PetscScalar *x;
1579: const MatScalar *aa, *a_a;
1580: const PetscInt *aj, *ii, *ridx = NULL;
1581: PetscInt m = A->rmap->n, n, i;
1582: PetscScalar sum;
1583: PetscBool usecprow = a->compressedrow.use;
1585: if (a->inode.use && a->inode.checked) {
1586: MatMultAdd_SeqAIJ_Inode(A, xx, yy, zz);
1587: return 0;
1588: }
1589: MatSeqAIJGetArrayRead(A, &a_a);
1590: VecGetArrayRead(xx, &x);
1591: VecGetArrayPair(yy, zz, &y, &z);
1592: if (usecprow) { /* use compressed row format */
1593: if (zz != yy) PetscArraycpy(z, y, m);
1594: m = a->compressedrow.nrows;
1595: ii = a->compressedrow.i;
1596: ridx = a->compressedrow.rindex;
1597: for (i = 0; i < m; i++) {
1598: n = ii[i + 1] - ii[i];
1599: aj = a->j + ii[i];
1600: aa = a_a + ii[i];
1601: sum = y[*ridx];
1602: PetscSparseDensePlusDot(sum, x, aa, aj, n);
1603: z[*ridx++] = sum;
1604: }
1605: } else { /* do not use compressed row format */
1606: ii = a->i;
1607: #if defined(PETSC_USE_FORTRAN_KERNEL_MULTADDAIJ)
1608: aj = a->j;
1609: aa = a_a;
1610: fortranmultaddaij_(&m, x, ii, aj, aa, y, z);
1611: #else
1612: for (i = 0; i < m; i++) {
1613: n = ii[i + 1] - ii[i];
1614: aj = a->j + ii[i];
1615: aa = a_a + ii[i];
1616: sum = y[i];
1617: PetscSparseDensePlusDot(sum, x, aa, aj, n);
1618: z[i] = sum;
1619: }
1620: #endif
1621: }
1622: PetscLogFlops(2.0 * a->nz);
1623: VecRestoreArrayRead(xx, &x);
1624: VecRestoreArrayPair(yy, zz, &y, &z);
1625: MatSeqAIJRestoreArrayRead(A, &a_a);
1626: return 0;
1627: }
1629: /*
1630: Adds diagonal pointers to sparse matrix structure.
1631: */
1632: PetscErrorCode MatMarkDiagonal_SeqAIJ(Mat A)
1633: {
1634: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
1635: PetscInt i, j, m = A->rmap->n;
1636: PetscBool alreadySet = PETSC_TRUE;
1638: if (!a->diag) {
1639: PetscMalloc1(m, &a->diag);
1640: alreadySet = PETSC_FALSE;
1641: }
1642: for (i = 0; i < A->rmap->n; i++) {
1643: /* If A's diagonal is already correctly set, this fast track enables cheap and repeated MatMarkDiagonal_SeqAIJ() calls */
1644: if (alreadySet) {
1645: PetscInt pos = a->diag[i];
1646: if (pos >= a->i[i] && pos < a->i[i + 1] && a->j[pos] == i) continue;
1647: }
1649: a->diag[i] = a->i[i + 1];
1650: for (j = a->i[i]; j < a->i[i + 1]; j++) {
1651: if (a->j[j] == i) {
1652: a->diag[i] = j;
1653: break;
1654: }
1655: }
1656: }
1657: return 0;
1658: }
1660: PetscErrorCode MatShift_SeqAIJ(Mat A, PetscScalar v)
1661: {
1662: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
1663: const PetscInt *diag = (const PetscInt *)a->diag;
1664: const PetscInt *ii = (const PetscInt *)a->i;
1665: PetscInt i, *mdiag = NULL;
1666: PetscInt cnt = 0; /* how many diagonals are missing */
1668: if (!A->preallocated || !a->nz) {
1669: MatSeqAIJSetPreallocation(A, 1, NULL);
1670: MatShift_Basic(A, v);
1671: return 0;
1672: }
1674: if (a->diagonaldense) {
1675: cnt = 0;
1676: } else {
1677: PetscCalloc1(A->rmap->n, &mdiag);
1678: for (i = 0; i < A->rmap->n; i++) {
1679: if (i < A->cmap->n && diag[i] >= ii[i + 1]) { /* 'out of range' rows never have diagonals */
1680: cnt++;
1681: mdiag[i] = 1;
1682: }
1683: }
1684: }
1685: if (!cnt) {
1686: MatShift_Basic(A, v);
1687: } else {
1688: PetscScalar *olda = a->a; /* preserve pointers to current matrix nonzeros structure and values */
1689: PetscInt *oldj = a->j, *oldi = a->i;
1690: PetscBool singlemalloc = a->singlemalloc, free_a = a->free_a, free_ij = a->free_ij;
1692: a->a = NULL;
1693: a->j = NULL;
1694: a->i = NULL;
1695: /* increase the values in imax for each row where a diagonal is being inserted then reallocate the matrix data structures */
1696: for (i = 0; i < PetscMin(A->rmap->n, A->cmap->n); i++) a->imax[i] += mdiag[i];
1697: MatSeqAIJSetPreallocation_SeqAIJ(A, 0, a->imax);
1699: /* copy old values into new matrix data structure */
1700: for (i = 0; i < A->rmap->n; i++) {
1701: MatSetValues(A, 1, &i, a->imax[i] - mdiag[i], &oldj[oldi[i]], &olda[oldi[i]], ADD_VALUES);
1702: if (i < A->cmap->n) MatSetValue(A, i, i, v, ADD_VALUES);
1703: }
1704: MatAssemblyBegin(A, MAT_FINAL_ASSEMBLY);
1705: MatAssemblyEnd(A, MAT_FINAL_ASSEMBLY);
1706: if (singlemalloc) {
1707: PetscFree3(olda, oldj, oldi);
1708: } else {
1709: if (free_a) PetscFree(olda);
1710: if (free_ij) PetscFree(oldj);
1711: if (free_ij) PetscFree(oldi);
1712: }
1713: }
1714: PetscFree(mdiag);
1715: a->diagonaldense = PETSC_TRUE;
1716: return 0;
1717: }
1719: /*
1720: Checks for missing diagonals
1721: */
1722: PetscErrorCode MatMissingDiagonal_SeqAIJ(Mat A, PetscBool *missing, PetscInt *d)
1723: {
1724: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
1725: PetscInt *diag, *ii = a->i, i;
1727: *missing = PETSC_FALSE;
1728: if (A->rmap->n > 0 && !ii) {
1729: *missing = PETSC_TRUE;
1730: if (d) *d = 0;
1731: PetscInfo(A, "Matrix has no entries therefore is missing diagonal\n");
1732: } else {
1733: PetscInt n;
1734: n = PetscMin(A->rmap->n, A->cmap->n);
1735: diag = a->diag;
1736: for (i = 0; i < n; i++) {
1737: if (diag[i] >= ii[i + 1]) {
1738: *missing = PETSC_TRUE;
1739: if (d) *d = i;
1740: PetscInfo(A, "Matrix is missing diagonal number %" PetscInt_FMT "\n", i);
1741: break;
1742: }
1743: }
1744: }
1745: return 0;
1746: }
1748: #include <petscblaslapack.h>
1749: #include <petsc/private/kernels/blockinvert.h>
1751: /*
1752: Note that values is allocated externally by the PC and then passed into this routine
1753: */
1754: PetscErrorCode MatInvertVariableBlockDiagonal_SeqAIJ(Mat A, PetscInt nblocks, const PetscInt *bsizes, PetscScalar *diag)
1755: {
1756: PetscInt n = A->rmap->n, i, ncnt = 0, *indx, j, bsizemax = 0, *v_pivots;
1757: PetscBool allowzeropivot, zeropivotdetected = PETSC_FALSE;
1758: const PetscReal shift = 0.0;
1759: PetscInt ipvt[5];
1760: PetscCount flops = 0;
1761: PetscScalar work[25], *v_work;
1763: allowzeropivot = PetscNot(A->erroriffailure);
1764: for (i = 0; i < nblocks; i++) ncnt += bsizes[i];
1766: for (i = 0; i < nblocks; i++) bsizemax = PetscMax(bsizemax, bsizes[i]);
1767: PetscMalloc1(bsizemax, &indx);
1768: if (bsizemax > 7) PetscMalloc2(bsizemax, &v_work, bsizemax, &v_pivots);
1769: ncnt = 0;
1770: for (i = 0; i < nblocks; i++) {
1771: for (j = 0; j < bsizes[i]; j++) indx[j] = ncnt + j;
1772: MatGetValues(A, bsizes[i], indx, bsizes[i], indx, diag);
1773: switch (bsizes[i]) {
1774: case 1:
1775: *diag = 1.0 / (*diag);
1776: break;
1777: case 2:
1778: PetscKernel_A_gets_inverse_A_2(diag, shift, allowzeropivot, &zeropivotdetected);
1779: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
1780: PetscKernel_A_gets_transpose_A_2(diag);
1781: break;
1782: case 3:
1783: PetscKernel_A_gets_inverse_A_3(diag, shift, allowzeropivot, &zeropivotdetected);
1784: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
1785: PetscKernel_A_gets_transpose_A_3(diag);
1786: break;
1787: case 4:
1788: PetscKernel_A_gets_inverse_A_4(diag, shift, allowzeropivot, &zeropivotdetected);
1789: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
1790: PetscKernel_A_gets_transpose_A_4(diag);
1791: break;
1792: case 5:
1793: PetscKernel_A_gets_inverse_A_5(diag, ipvt, work, shift, allowzeropivot, &zeropivotdetected);
1794: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
1795: PetscKernel_A_gets_transpose_A_5(diag);
1796: break;
1797: case 6:
1798: PetscKernel_A_gets_inverse_A_6(diag, shift, allowzeropivot, &zeropivotdetected);
1799: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
1800: PetscKernel_A_gets_transpose_A_6(diag);
1801: break;
1802: case 7:
1803: PetscKernel_A_gets_inverse_A_7(diag, shift, allowzeropivot, &zeropivotdetected);
1804: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
1805: PetscKernel_A_gets_transpose_A_7(diag);
1806: break;
1807: default:
1808: PetscKernel_A_gets_inverse_A(bsizes[i], diag, v_pivots, v_work, allowzeropivot, &zeropivotdetected);
1809: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
1810: PetscKernel_A_gets_transpose_A_N(diag, bsizes[i]);
1811: }
1812: ncnt += bsizes[i];
1813: diag += bsizes[i] * bsizes[i];
1814: flops += 2 * PetscPowInt(bsizes[i], 3) / 3;
1815: }
1816: PetscLogFlops(flops);
1817: if (bsizemax > 7) PetscFree2(v_work, v_pivots);
1818: PetscFree(indx);
1819: return 0;
1820: }
1822: /*
1823: Negative shift indicates do not generate an error if there is a zero diagonal, just invert it anyways
1824: */
1825: PetscErrorCode MatInvertDiagonal_SeqAIJ(Mat A, PetscScalar omega, PetscScalar fshift)
1826: {
1827: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
1828: PetscInt i, *diag, m = A->rmap->n;
1829: const MatScalar *v;
1830: PetscScalar *idiag, *mdiag;
1832: if (a->idiagvalid) return 0;
1833: MatMarkDiagonal_SeqAIJ(A);
1834: diag = a->diag;
1835: if (!a->idiag) { PetscMalloc3(m, &a->idiag, m, &a->mdiag, m, &a->ssor_work); }
1837: mdiag = a->mdiag;
1838: idiag = a->idiag;
1839: MatSeqAIJGetArrayRead(A, &v);
1840: if (omega == 1.0 && PetscRealPart(fshift) <= 0.0) {
1841: for (i = 0; i < m; i++) {
1842: mdiag[i] = v[diag[i]];
1843: if (!PetscAbsScalar(mdiag[i])) { /* zero diagonal */
1844: if (PetscRealPart(fshift)) {
1845: PetscInfo(A, "Zero diagonal on row %" PetscInt_FMT "\n", i);
1846: A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
1847: A->factorerror_zeropivot_value = 0.0;
1848: A->factorerror_zeropivot_row = i;
1849: } else SETERRQ(PETSC_COMM_SELF, PETSC_ERR_ARG_INCOMP, "Zero diagonal on row %" PetscInt_FMT, i);
1850: }
1851: idiag[i] = 1.0 / v[diag[i]];
1852: }
1853: PetscLogFlops(m);
1854: } else {
1855: for (i = 0; i < m; i++) {
1856: mdiag[i] = v[diag[i]];
1857: idiag[i] = omega / (fshift + v[diag[i]]);
1858: }
1859: PetscLogFlops(2.0 * m);
1860: }
1861: a->idiagvalid = PETSC_TRUE;
1862: MatSeqAIJRestoreArrayRead(A, &v);
1863: return 0;
1864: }
1866: #include <../src/mat/impls/aij/seq/ftn-kernels/frelax.h>
1867: PetscErrorCode MatSOR_SeqAIJ(Mat A, Vec bb, PetscReal omega, MatSORType flag, PetscReal fshift, PetscInt its, PetscInt lits, Vec xx)
1868: {
1869: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
1870: PetscScalar *x, d, sum, *t, scale;
1871: const MatScalar *v, *idiag = NULL, *mdiag, *aa;
1872: const PetscScalar *b, *bs, *xb, *ts;
1873: PetscInt n, m = A->rmap->n, i;
1874: const PetscInt *idx, *diag;
1876: if (a->inode.use && a->inode.checked && omega == 1.0 && fshift == 0.0) {
1877: MatSOR_SeqAIJ_Inode(A, bb, omega, flag, fshift, its, lits, xx);
1878: return 0;
1879: }
1880: its = its * lits;
1882: if (fshift != a->fshift || omega != a->omega) a->idiagvalid = PETSC_FALSE; /* must recompute idiag[] */
1883: if (!a->idiagvalid) MatInvertDiagonal_SeqAIJ(A, omega, fshift);
1884: a->fshift = fshift;
1885: a->omega = omega;
1887: diag = a->diag;
1888: t = a->ssor_work;
1889: idiag = a->idiag;
1890: mdiag = a->mdiag;
1892: MatSeqAIJGetArrayRead(A, &aa);
1893: VecGetArray(xx, &x);
1894: VecGetArrayRead(bb, &b);
1895: /* We count flops by assuming the upper triangular and lower triangular parts have the same number of nonzeros */
1896: if (flag == SOR_APPLY_UPPER) {
1897: /* apply (U + D/omega) to the vector */
1898: bs = b;
1899: for (i = 0; i < m; i++) {
1900: d = fshift + mdiag[i];
1901: n = a->i[i + 1] - diag[i] - 1;
1902: idx = a->j + diag[i] + 1;
1903: v = aa + diag[i] + 1;
1904: sum = b[i] * d / omega;
1905: PetscSparseDensePlusDot(sum, bs, v, idx, n);
1906: x[i] = sum;
1907: }
1908: VecRestoreArray(xx, &x);
1909: VecRestoreArrayRead(bb, &b);
1910: MatSeqAIJRestoreArrayRead(A, &aa);
1911: PetscLogFlops(a->nz);
1912: return 0;
1913: }
1916: if (flag & SOR_EISENSTAT) {
1917: /* Let A = L + U + D; where L is lower triangular,
1918: U is upper triangular, E = D/omega; This routine applies
1920: (L + E)^{-1} A (U + E)^{-1}
1922: to a vector efficiently using Eisenstat's trick.
1923: */
1924: scale = (2.0 / omega) - 1.0;
1926: /* x = (E + U)^{-1} b */
1927: for (i = m - 1; i >= 0; i--) {
1928: n = a->i[i + 1] - diag[i] - 1;
1929: idx = a->j + diag[i] + 1;
1930: v = aa + diag[i] + 1;
1931: sum = b[i];
1932: PetscSparseDenseMinusDot(sum, x, v, idx, n);
1933: x[i] = sum * idiag[i];
1934: }
1936: /* t = b - (2*E - D)x */
1937: v = aa;
1938: for (i = 0; i < m; i++) t[i] = b[i] - scale * (v[*diag++]) * x[i];
1940: /* t = (E + L)^{-1}t */
1941: ts = t;
1942: diag = a->diag;
1943: for (i = 0; i < m; i++) {
1944: n = diag[i] - a->i[i];
1945: idx = a->j + a->i[i];
1946: v = aa + a->i[i];
1947: sum = t[i];
1948: PetscSparseDenseMinusDot(sum, ts, v, idx, n);
1949: t[i] = sum * idiag[i];
1950: /* x = x + t */
1951: x[i] += t[i];
1952: }
1954: PetscLogFlops(6.0 * m - 1 + 2.0 * a->nz);
1955: VecRestoreArray(xx, &x);
1956: VecRestoreArrayRead(bb, &b);
1957: return 0;
1958: }
1959: if (flag & SOR_ZERO_INITIAL_GUESS) {
1960: if (flag & SOR_FORWARD_SWEEP || flag & SOR_LOCAL_FORWARD_SWEEP) {
1961: for (i = 0; i < m; i++) {
1962: n = diag[i] - a->i[i];
1963: idx = a->j + a->i[i];
1964: v = aa + a->i[i];
1965: sum = b[i];
1966: PetscSparseDenseMinusDot(sum, x, v, idx, n);
1967: t[i] = sum;
1968: x[i] = sum * idiag[i];
1969: }
1970: xb = t;
1971: PetscLogFlops(a->nz);
1972: } else xb = b;
1973: if (flag & SOR_BACKWARD_SWEEP || flag & SOR_LOCAL_BACKWARD_SWEEP) {
1974: for (i = m - 1; i >= 0; i--) {
1975: n = a->i[i + 1] - diag[i] - 1;
1976: idx = a->j + diag[i] + 1;
1977: v = aa + diag[i] + 1;
1978: sum = xb[i];
1979: PetscSparseDenseMinusDot(sum, x, v, idx, n);
1980: if (xb == b) {
1981: x[i] = sum * idiag[i];
1982: } else {
1983: x[i] = (1 - omega) * x[i] + sum * idiag[i]; /* omega in idiag */
1984: }
1985: }
1986: PetscLogFlops(a->nz); /* assumes 1/2 in upper */
1987: }
1988: its--;
1989: }
1990: while (its--) {
1991: if (flag & SOR_FORWARD_SWEEP || flag & SOR_LOCAL_FORWARD_SWEEP) {
1992: for (i = 0; i < m; i++) {
1993: /* lower */
1994: n = diag[i] - a->i[i];
1995: idx = a->j + a->i[i];
1996: v = aa + a->i[i];
1997: sum = b[i];
1998: PetscSparseDenseMinusDot(sum, x, v, idx, n);
1999: t[i] = sum; /* save application of the lower-triangular part */
2000: /* upper */
2001: n = a->i[i + 1] - diag[i] - 1;
2002: idx = a->j + diag[i] + 1;
2003: v = aa + diag[i] + 1;
2004: PetscSparseDenseMinusDot(sum, x, v, idx, n);
2005: x[i] = (1. - omega) * x[i] + sum * idiag[i]; /* omega in idiag */
2006: }
2007: xb = t;
2008: PetscLogFlops(2.0 * a->nz);
2009: } else xb = b;
2010: if (flag & SOR_BACKWARD_SWEEP || flag & SOR_LOCAL_BACKWARD_SWEEP) {
2011: for (i = m - 1; i >= 0; i--) {
2012: sum = xb[i];
2013: if (xb == b) {
2014: /* whole matrix (no checkpointing available) */
2015: n = a->i[i + 1] - a->i[i];
2016: idx = a->j + a->i[i];
2017: v = aa + a->i[i];
2018: PetscSparseDenseMinusDot(sum, x, v, idx, n);
2019: x[i] = (1. - omega) * x[i] + (sum + mdiag[i] * x[i]) * idiag[i];
2020: } else { /* lower-triangular part has been saved, so only apply upper-triangular */
2021: n = a->i[i + 1] - diag[i] - 1;
2022: idx = a->j + diag[i] + 1;
2023: v = aa + diag[i] + 1;
2024: PetscSparseDenseMinusDot(sum, x, v, idx, n);
2025: x[i] = (1. - omega) * x[i] + sum * idiag[i]; /* omega in idiag */
2026: }
2027: }
2028: if (xb == b) {
2029: PetscLogFlops(2.0 * a->nz);
2030: } else {
2031: PetscLogFlops(a->nz); /* assumes 1/2 in upper */
2032: }
2033: }
2034: }
2035: MatSeqAIJRestoreArrayRead(A, &aa);
2036: VecRestoreArray(xx, &x);
2037: VecRestoreArrayRead(bb, &b);
2038: return 0;
2039: }
2041: PetscErrorCode MatGetInfo_SeqAIJ(Mat A, MatInfoType flag, MatInfo *info)
2042: {
2043: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
2045: info->block_size = 1.0;
2046: info->nz_allocated = a->maxnz;
2047: info->nz_used = a->nz;
2048: info->nz_unneeded = (a->maxnz - a->nz);
2049: info->assemblies = A->num_ass;
2050: info->mallocs = A->info.mallocs;
2051: info->memory = 0; /* REVIEW ME */
2052: if (A->factortype) {
2053: info->fill_ratio_given = A->info.fill_ratio_given;
2054: info->fill_ratio_needed = A->info.fill_ratio_needed;
2055: info->factor_mallocs = A->info.factor_mallocs;
2056: } else {
2057: info->fill_ratio_given = 0;
2058: info->fill_ratio_needed = 0;
2059: info->factor_mallocs = 0;
2060: }
2061: return 0;
2062: }
2064: PetscErrorCode MatZeroRows_SeqAIJ(Mat A, PetscInt N, const PetscInt rows[], PetscScalar diag, Vec x, Vec b)
2065: {
2066: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
2067: PetscInt i, m = A->rmap->n - 1;
2068: const PetscScalar *xx;
2069: PetscScalar *bb, *aa;
2070: PetscInt d = 0;
2072: if (x && b) {
2073: VecGetArrayRead(x, &xx);
2074: VecGetArray(b, &bb);
2075: for (i = 0; i < N; i++) {
2077: if (rows[i] >= A->cmap->n) continue;
2078: bb[rows[i]] = diag * xx[rows[i]];
2079: }
2080: VecRestoreArrayRead(x, &xx);
2081: VecRestoreArray(b, &bb);
2082: }
2084: MatSeqAIJGetArray(A, &aa);
2085: if (a->keepnonzeropattern) {
2086: for (i = 0; i < N; i++) {
2088: PetscArrayzero(&aa[a->i[rows[i]]], a->ilen[rows[i]]);
2089: }
2090: if (diag != 0.0) {
2091: for (i = 0; i < N; i++) {
2092: d = rows[i];
2093: if (rows[i] >= A->cmap->n) continue;
2095: }
2096: for (i = 0; i < N; i++) {
2097: if (rows[i] >= A->cmap->n) continue;
2098: aa[a->diag[rows[i]]] = diag;
2099: }
2100: }
2101: } else {
2102: if (diag != 0.0) {
2103: for (i = 0; i < N; i++) {
2105: if (a->ilen[rows[i]] > 0) {
2106: if (rows[i] >= A->cmap->n) {
2107: a->ilen[rows[i]] = 0;
2108: } else {
2109: a->ilen[rows[i]] = 1;
2110: aa[a->i[rows[i]]] = diag;
2111: a->j[a->i[rows[i]]] = rows[i];
2112: }
2113: } else if (rows[i] < A->cmap->n) { /* in case row was completely empty */
2114: MatSetValues_SeqAIJ(A, 1, &rows[i], 1, &rows[i], &diag, INSERT_VALUES);
2115: }
2116: }
2117: } else {
2118: for (i = 0; i < N; i++) {
2120: a->ilen[rows[i]] = 0;
2121: }
2122: }
2123: A->nonzerostate++;
2124: }
2125: MatSeqAIJRestoreArray(A, &aa);
2126: PetscUseTypeMethod(A, assemblyend, MAT_FINAL_ASSEMBLY);
2127: return 0;
2128: }
2130: PetscErrorCode MatZeroRowsColumns_SeqAIJ(Mat A, PetscInt N, const PetscInt rows[], PetscScalar diag, Vec x, Vec b)
2131: {
2132: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
2133: PetscInt i, j, m = A->rmap->n - 1, d = 0;
2134: PetscBool missing, *zeroed, vecs = PETSC_FALSE;
2135: const PetscScalar *xx;
2136: PetscScalar *bb, *aa;
2138: if (!N) return 0;
2139: MatSeqAIJGetArray(A, &aa);
2140: if (x && b) {
2141: VecGetArrayRead(x, &xx);
2142: VecGetArray(b, &bb);
2143: vecs = PETSC_TRUE;
2144: }
2145: PetscCalloc1(A->rmap->n, &zeroed);
2146: for (i = 0; i < N; i++) {
2148: PetscArrayzero(&aa[a->i[rows[i]]], a->ilen[rows[i]]);
2150: zeroed[rows[i]] = PETSC_TRUE;
2151: }
2152: for (i = 0; i < A->rmap->n; i++) {
2153: if (!zeroed[i]) {
2154: for (j = a->i[i]; j < a->i[i + 1]; j++) {
2155: if (a->j[j] < A->rmap->n && zeroed[a->j[j]]) {
2156: if (vecs) bb[i] -= aa[j] * xx[a->j[j]];
2157: aa[j] = 0.0;
2158: }
2159: }
2160: } else if (vecs && i < A->cmap->N) bb[i] = diag * xx[i];
2161: }
2162: if (x && b) {
2163: VecRestoreArrayRead(x, &xx);
2164: VecRestoreArray(b, &bb);
2165: }
2166: PetscFree(zeroed);
2167: if (diag != 0.0) {
2168: MatMissingDiagonal_SeqAIJ(A, &missing, &d);
2169: if (missing) {
2170: for (i = 0; i < N; i++) {
2171: if (rows[i] >= A->cmap->N) continue;
2173: MatSetValues_SeqAIJ(A, 1, &rows[i], 1, &rows[i], &diag, INSERT_VALUES);
2174: }
2175: } else {
2176: for (i = 0; i < N; i++) aa[a->diag[rows[i]]] = diag;
2177: }
2178: }
2179: MatSeqAIJRestoreArray(A, &aa);
2180: PetscUseTypeMethod(A, assemblyend, MAT_FINAL_ASSEMBLY);
2181: return 0;
2182: }
2184: PetscErrorCode MatGetRow_SeqAIJ(Mat A, PetscInt row, PetscInt *nz, PetscInt **idx, PetscScalar **v)
2185: {
2186: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
2187: const PetscScalar *aa;
2188: PetscInt *itmp;
2190: MatSeqAIJGetArrayRead(A, &aa);
2191: *nz = a->i[row + 1] - a->i[row];
2192: if (v) *v = (PetscScalar *)(aa + a->i[row]);
2193: if (idx) {
2194: itmp = a->j + a->i[row];
2195: if (*nz) *idx = itmp;
2196: else *idx = NULL;
2197: }
2198: MatSeqAIJRestoreArrayRead(A, &aa);
2199: return 0;
2200: }
2202: PetscErrorCode MatRestoreRow_SeqAIJ(Mat A, PetscInt row, PetscInt *nz, PetscInt **idx, PetscScalar **v)
2203: {
2204: if (nz) *nz = 0;
2205: if (idx) *idx = NULL;
2206: if (v) *v = NULL;
2207: return 0;
2208: }
2210: PetscErrorCode MatNorm_SeqAIJ(Mat A, NormType type, PetscReal *nrm)
2211: {
2212: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
2213: const MatScalar *v;
2214: PetscReal sum = 0.0;
2215: PetscInt i, j;
2217: MatSeqAIJGetArrayRead(A, &v);
2218: if (type == NORM_FROBENIUS) {
2219: #if defined(PETSC_USE_REAL___FP16)
2220: PetscBLASInt one = 1, nz = a->nz;
2221: PetscCallBLAS("BLASnrm2", *nrm = BLASnrm2_(&nz, v, &one));
2222: #else
2223: for (i = 0; i < a->nz; i++) {
2224: sum += PetscRealPart(PetscConj(*v) * (*v));
2225: v++;
2226: }
2227: *nrm = PetscSqrtReal(sum);
2228: #endif
2229: PetscLogFlops(2.0 * a->nz);
2230: } else if (type == NORM_1) {
2231: PetscReal *tmp;
2232: PetscInt *jj = a->j;
2233: PetscCalloc1(A->cmap->n + 1, &tmp);
2234: *nrm = 0.0;
2235: for (j = 0; j < a->nz; j++) {
2236: tmp[*jj++] += PetscAbsScalar(*v);
2237: v++;
2238: }
2239: for (j = 0; j < A->cmap->n; j++) {
2240: if (tmp[j] > *nrm) *nrm = tmp[j];
2241: }
2242: PetscFree(tmp);
2243: PetscLogFlops(PetscMax(a->nz - 1, 0));
2244: } else if (type == NORM_INFINITY) {
2245: *nrm = 0.0;
2246: for (j = 0; j < A->rmap->n; j++) {
2247: const PetscScalar *v2 = v + a->i[j];
2248: sum = 0.0;
2249: for (i = 0; i < a->i[j + 1] - a->i[j]; i++) {
2250: sum += PetscAbsScalar(*v2);
2251: v2++;
2252: }
2253: if (sum > *nrm) *nrm = sum;
2254: }
2255: PetscLogFlops(PetscMax(a->nz - 1, 0));
2256: } else SETERRQ(PETSC_COMM_SELF, PETSC_ERR_SUP, "No support for two norm");
2257: MatSeqAIJRestoreArrayRead(A, &v);
2258: return 0;
2259: }
2261: PetscErrorCode MatIsTranspose_SeqAIJ(Mat A, Mat B, PetscReal tol, PetscBool *f)
2262: {
2263: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)A->data, *bij = (Mat_SeqAIJ *)B->data;
2264: PetscInt *adx, *bdx, *aii, *bii, *aptr, *bptr;
2265: const MatScalar *va, *vb;
2266: PetscInt ma, na, mb, nb, i;
2268: MatGetSize(A, &ma, &na);
2269: MatGetSize(B, &mb, &nb);
2270: if (ma != nb || na != mb) {
2271: *f = PETSC_FALSE;
2272: return 0;
2273: }
2274: MatSeqAIJGetArrayRead(A, &va);
2275: MatSeqAIJGetArrayRead(B, &vb);
2276: aii = aij->i;
2277: bii = bij->i;
2278: adx = aij->j;
2279: bdx = bij->j;
2280: PetscMalloc1(ma, &aptr);
2281: PetscMalloc1(mb, &bptr);
2282: for (i = 0; i < ma; i++) aptr[i] = aii[i];
2283: for (i = 0; i < mb; i++) bptr[i] = bii[i];
2285: *f = PETSC_TRUE;
2286: for (i = 0; i < ma; i++) {
2287: while (aptr[i] < aii[i + 1]) {
2288: PetscInt idc, idr;
2289: PetscScalar vc, vr;
2290: /* column/row index/value */
2291: idc = adx[aptr[i]];
2292: idr = bdx[bptr[idc]];
2293: vc = va[aptr[i]];
2294: vr = vb[bptr[idc]];
2295: if (i != idr || PetscAbsScalar(vc - vr) > tol) {
2296: *f = PETSC_FALSE;
2297: goto done;
2298: } else {
2299: aptr[i]++;
2300: if (B || i != idc) bptr[idc]++;
2301: }
2302: }
2303: }
2304: done:
2305: PetscFree(aptr);
2306: PetscFree(bptr);
2307: MatSeqAIJRestoreArrayRead(A, &va);
2308: MatSeqAIJRestoreArrayRead(B, &vb);
2309: return 0;
2310: }
2312: PetscErrorCode MatIsHermitianTranspose_SeqAIJ(Mat A, Mat B, PetscReal tol, PetscBool *f)
2313: {
2314: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)A->data, *bij = (Mat_SeqAIJ *)B->data;
2315: PetscInt *adx, *bdx, *aii, *bii, *aptr, *bptr;
2316: MatScalar *va, *vb;
2317: PetscInt ma, na, mb, nb, i;
2319: MatGetSize(A, &ma, &na);
2320: MatGetSize(B, &mb, &nb);
2321: if (ma != nb || na != mb) {
2322: *f = PETSC_FALSE;
2323: return 0;
2324: }
2325: aii = aij->i;
2326: bii = bij->i;
2327: adx = aij->j;
2328: bdx = bij->j;
2329: va = aij->a;
2330: vb = bij->a;
2331: PetscMalloc1(ma, &aptr);
2332: PetscMalloc1(mb, &bptr);
2333: for (i = 0; i < ma; i++) aptr[i] = aii[i];
2334: for (i = 0; i < mb; i++) bptr[i] = bii[i];
2336: *f = PETSC_TRUE;
2337: for (i = 0; i < ma; i++) {
2338: while (aptr[i] < aii[i + 1]) {
2339: PetscInt idc, idr;
2340: PetscScalar vc, vr;
2341: /* column/row index/value */
2342: idc = adx[aptr[i]];
2343: idr = bdx[bptr[idc]];
2344: vc = va[aptr[i]];
2345: vr = vb[bptr[idc]];
2346: if (i != idr || PetscAbsScalar(vc - PetscConj(vr)) > tol) {
2347: *f = PETSC_FALSE;
2348: goto done;
2349: } else {
2350: aptr[i]++;
2351: if (B || i != idc) bptr[idc]++;
2352: }
2353: }
2354: }
2355: done:
2356: PetscFree(aptr);
2357: PetscFree(bptr);
2358: return 0;
2359: }
2361: PetscErrorCode MatIsSymmetric_SeqAIJ(Mat A, PetscReal tol, PetscBool *f)
2362: {
2363: MatIsTranspose_SeqAIJ(A, A, tol, f);
2364: return 0;
2365: }
2367: PetscErrorCode MatIsHermitian_SeqAIJ(Mat A, PetscReal tol, PetscBool *f)
2368: {
2369: MatIsHermitianTranspose_SeqAIJ(A, A, tol, f);
2370: return 0;
2371: }
2373: PetscErrorCode MatDiagonalScale_SeqAIJ(Mat A, Vec ll, Vec rr)
2374: {
2375: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
2376: const PetscScalar *l, *r;
2377: PetscScalar x;
2378: MatScalar *v;
2379: PetscInt i, j, m = A->rmap->n, n = A->cmap->n, M, nz = a->nz;
2380: const PetscInt *jj;
2382: if (ll) {
2383: /* The local size is used so that VecMPI can be passed to this routine
2384: by MatDiagonalScale_MPIAIJ */
2385: VecGetLocalSize(ll, &m);
2387: VecGetArrayRead(ll, &l);
2388: MatSeqAIJGetArray(A, &v);
2389: for (i = 0; i < m; i++) {
2390: x = l[i];
2391: M = a->i[i + 1] - a->i[i];
2392: for (j = 0; j < M; j++) (*v++) *= x;
2393: }
2394: VecRestoreArrayRead(ll, &l);
2395: PetscLogFlops(nz);
2396: MatSeqAIJRestoreArray(A, &v);
2397: }
2398: if (rr) {
2399: VecGetLocalSize(rr, &n);
2401: VecGetArrayRead(rr, &r);
2402: MatSeqAIJGetArray(A, &v);
2403: jj = a->j;
2404: for (i = 0; i < nz; i++) (*v++) *= r[*jj++];
2405: MatSeqAIJRestoreArray(A, &v);
2406: VecRestoreArrayRead(rr, &r);
2407: PetscLogFlops(nz);
2408: }
2409: MatSeqAIJInvalidateDiagonal(A);
2410: return 0;
2411: }
2413: PetscErrorCode MatCreateSubMatrix_SeqAIJ(Mat A, IS isrow, IS iscol, PetscInt csize, MatReuse scall, Mat *B)
2414: {
2415: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data, *c;
2416: PetscInt *smap, i, k, kstart, kend, oldcols = A->cmap->n, *lens;
2417: PetscInt row, mat_i, *mat_j, tcol, first, step, *mat_ilen, sum, lensi;
2418: const PetscInt *irow, *icol;
2419: const PetscScalar *aa;
2420: PetscInt nrows, ncols;
2421: PetscInt *starts, *j_new, *i_new, *aj = a->j, *ai = a->i, ii, *ailen = a->ilen;
2422: MatScalar *a_new, *mat_a, *c_a;
2423: Mat C;
2424: PetscBool stride;
2426: ISGetIndices(isrow, &irow);
2427: ISGetLocalSize(isrow, &nrows);
2428: ISGetLocalSize(iscol, &ncols);
2430: PetscObjectTypeCompare((PetscObject)iscol, ISSTRIDE, &stride);
2431: if (stride) {
2432: ISStrideGetInfo(iscol, &first, &step);
2433: } else {
2434: first = 0;
2435: step = 0;
2436: }
2437: if (stride && step == 1) {
2438: /* special case of contiguous rows */
2439: PetscMalloc2(nrows, &lens, nrows, &starts);
2440: /* loop over new rows determining lens and starting points */
2441: for (i = 0; i < nrows; i++) {
2442: kstart = ai[irow[i]];
2443: kend = kstart + ailen[irow[i]];
2444: starts[i] = kstart;
2445: for (k = kstart; k < kend; k++) {
2446: if (aj[k] >= first) {
2447: starts[i] = k;
2448: break;
2449: }
2450: }
2451: sum = 0;
2452: while (k < kend) {
2453: if (aj[k++] >= first + ncols) break;
2454: sum++;
2455: }
2456: lens[i] = sum;
2457: }
2458: /* create submatrix */
2459: if (scall == MAT_REUSE_MATRIX) {
2460: PetscInt n_cols, n_rows;
2461: MatGetSize(*B, &n_rows, &n_cols);
2463: MatZeroEntries(*B);
2464: C = *B;
2465: } else {
2466: PetscInt rbs, cbs;
2467: MatCreate(PetscObjectComm((PetscObject)A), &C);
2468: MatSetSizes(C, nrows, ncols, PETSC_DETERMINE, PETSC_DETERMINE);
2469: ISGetBlockSize(isrow, &rbs);
2470: ISGetBlockSize(iscol, &cbs);
2471: MatSetBlockSizes(C, rbs, cbs);
2472: MatSetType(C, ((PetscObject)A)->type_name);
2473: MatSeqAIJSetPreallocation_SeqAIJ(C, 0, lens);
2474: }
2475: c = (Mat_SeqAIJ *)C->data;
2477: /* loop over rows inserting into submatrix */
2478: MatSeqAIJGetArrayWrite(C, &a_new); // Not 'a_new = c->a-new', since that raw usage ignores offload state of C
2479: j_new = c->j;
2480: i_new = c->i;
2481: MatSeqAIJGetArrayRead(A, &aa);
2482: for (i = 0; i < nrows; i++) {
2483: ii = starts[i];
2484: lensi = lens[i];
2485: for (k = 0; k < lensi; k++) *j_new++ = aj[ii + k] - first;
2486: PetscArraycpy(a_new, aa + starts[i], lensi);
2487: a_new += lensi;
2488: i_new[i + 1] = i_new[i] + lensi;
2489: c->ilen[i] = lensi;
2490: }
2491: MatSeqAIJRestoreArrayWrite(C, &a_new); // Set C's offload state properly
2492: MatSeqAIJRestoreArrayRead(A, &aa);
2493: PetscFree2(lens, starts);
2494: } else {
2495: ISGetIndices(iscol, &icol);
2496: PetscCalloc1(oldcols, &smap);
2497: PetscMalloc1(1 + nrows, &lens);
2498: for (i = 0; i < ncols; i++) {
2500: smap[icol[i]] = i + 1;
2501: }
2503: /* determine lens of each row */
2504: for (i = 0; i < nrows; i++) {
2505: kstart = ai[irow[i]];
2506: kend = kstart + a->ilen[irow[i]];
2507: lens[i] = 0;
2508: for (k = kstart; k < kend; k++) {
2509: if (smap[aj[k]]) lens[i]++;
2510: }
2511: }
2512: /* Create and fill new matrix */
2513: if (scall == MAT_REUSE_MATRIX) {
2514: PetscBool equal;
2516: c = (Mat_SeqAIJ *)((*B)->data);
2518: PetscArraycmp(c->ilen, lens, (*B)->rmap->n, &equal);
2520: PetscArrayzero(c->ilen, (*B)->rmap->n);
2521: C = *B;
2522: } else {
2523: PetscInt rbs, cbs;
2524: MatCreate(PetscObjectComm((PetscObject)A), &C);
2525: MatSetSizes(C, nrows, ncols, PETSC_DETERMINE, PETSC_DETERMINE);
2526: ISGetBlockSize(isrow, &rbs);
2527: ISGetBlockSize(iscol, &cbs);
2528: MatSetBlockSizes(C, rbs, cbs);
2529: MatSetType(C, ((PetscObject)A)->type_name);
2530: MatSeqAIJSetPreallocation_SeqAIJ(C, 0, lens);
2531: }
2532: MatSeqAIJGetArrayRead(A, &aa);
2534: c = (Mat_SeqAIJ *)(C->data);
2535: MatSeqAIJGetArrayWrite(C, &c_a); // Not 'c->a', since that raw usage ignores offload state of C
2536: for (i = 0; i < nrows; i++) {
2537: row = irow[i];
2538: kstart = ai[row];
2539: kend = kstart + a->ilen[row];
2540: mat_i = c->i[i];
2541: mat_j = c->j + mat_i;
2542: mat_a = c_a + mat_i;
2543: mat_ilen = c->ilen + i;
2544: for (k = kstart; k < kend; k++) {
2545: if ((tcol = smap[a->j[k]])) {
2546: *mat_j++ = tcol - 1;
2547: *mat_a++ = aa[k];
2548: (*mat_ilen)++;
2549: }
2550: }
2551: }
2552: MatSeqAIJRestoreArrayRead(A, &aa);
2553: /* Free work space */
2554: ISRestoreIndices(iscol, &icol);
2555: PetscFree(smap);
2556: PetscFree(lens);
2557: /* sort */
2558: for (i = 0; i < nrows; i++) {
2559: PetscInt ilen;
2561: mat_i = c->i[i];
2562: mat_j = c->j + mat_i;
2563: mat_a = c_a + mat_i;
2564: ilen = c->ilen[i];
2565: PetscSortIntWithScalarArray(ilen, mat_j, mat_a);
2566: }
2567: MatSeqAIJRestoreArrayWrite(C, &c_a);
2568: }
2569: #if defined(PETSC_HAVE_DEVICE)
2570: MatBindToCPU(C, A->boundtocpu);
2571: #endif
2572: MatAssemblyBegin(C, MAT_FINAL_ASSEMBLY);
2573: MatAssemblyEnd(C, MAT_FINAL_ASSEMBLY);
2575: ISRestoreIndices(isrow, &irow);
2576: *B = C;
2577: return 0;
2578: }
2580: PetscErrorCode MatGetMultiProcBlock_SeqAIJ(Mat mat, MPI_Comm subComm, MatReuse scall, Mat *subMat)
2581: {
2582: Mat B;
2584: if (scall == MAT_INITIAL_MATRIX) {
2585: MatCreate(subComm, &B);
2586: MatSetSizes(B, mat->rmap->n, mat->cmap->n, mat->rmap->n, mat->cmap->n);
2587: MatSetBlockSizesFromMats(B, mat, mat);
2588: MatSetType(B, MATSEQAIJ);
2589: MatDuplicateNoCreate_SeqAIJ(B, mat, MAT_COPY_VALUES, PETSC_TRUE);
2590: *subMat = B;
2591: } else {
2592: MatCopy_SeqAIJ(mat, *subMat, SAME_NONZERO_PATTERN);
2593: }
2594: return 0;
2595: }
2597: PetscErrorCode MatILUFactor_SeqAIJ(Mat inA, IS row, IS col, const MatFactorInfo *info)
2598: {
2599: Mat_SeqAIJ *a = (Mat_SeqAIJ *)inA->data;
2600: Mat outA;
2601: PetscBool row_identity, col_identity;
2605: ISIdentity(row, &row_identity);
2606: ISIdentity(col, &col_identity);
2608: outA = inA;
2609: outA->factortype = MAT_FACTOR_LU;
2610: PetscFree(inA->solvertype);
2611: PetscStrallocpy(MATSOLVERPETSC, &inA->solvertype);
2613: PetscObjectReference((PetscObject)row);
2614: ISDestroy(&a->row);
2616: a->row = row;
2618: PetscObjectReference((PetscObject)col);
2619: ISDestroy(&a->col);
2621: a->col = col;
2623: /* Create the inverse permutation so that it can be used in MatLUFactorNumeric() */
2624: ISDestroy(&a->icol);
2625: ISInvertPermutation(col, PETSC_DECIDE, &a->icol);
2627: if (!a->solve_work) { /* this matrix may have been factored before */
2628: PetscMalloc1(inA->rmap->n + 1, &a->solve_work);
2629: }
2631: MatMarkDiagonal_SeqAIJ(inA);
2632: if (row_identity && col_identity) {
2633: MatLUFactorNumeric_SeqAIJ_inplace(outA, inA, info);
2634: } else {
2635: MatLUFactorNumeric_SeqAIJ_InplaceWithPerm(outA, inA, info);
2636: }
2637: return 0;
2638: }
2640: PetscErrorCode MatScale_SeqAIJ(Mat inA, PetscScalar alpha)
2641: {
2642: Mat_SeqAIJ *a = (Mat_SeqAIJ *)inA->data;
2643: PetscScalar *v;
2644: PetscBLASInt one = 1, bnz;
2646: MatSeqAIJGetArray(inA, &v);
2647: PetscBLASIntCast(a->nz, &bnz);
2648: PetscCallBLAS("BLASscal", BLASscal_(&bnz, &alpha, v, &one));
2649: PetscLogFlops(a->nz);
2650: MatSeqAIJRestoreArray(inA, &v);
2651: MatSeqAIJInvalidateDiagonal(inA);
2652: return 0;
2653: }
2655: PetscErrorCode MatDestroySubMatrix_Private(Mat_SubSppt *submatj)
2656: {
2657: PetscInt i;
2659: if (!submatj->id) { /* delete data that are linked only to submats[id=0] */
2660: PetscFree4(submatj->sbuf1, submatj->ptr, submatj->tmp, submatj->ctr);
2662: for (i = 0; i < submatj->nrqr; ++i) PetscFree(submatj->sbuf2[i]);
2663: PetscFree3(submatj->sbuf2, submatj->req_size, submatj->req_source1);
2665: if (submatj->rbuf1) {
2666: PetscFree(submatj->rbuf1[0]);
2667: PetscFree(submatj->rbuf1);
2668: }
2670: for (i = 0; i < submatj->nrqs; ++i) PetscFree(submatj->rbuf3[i]);
2671: PetscFree3(submatj->req_source2, submatj->rbuf2, submatj->rbuf3);
2672: PetscFree(submatj->pa);
2673: }
2675: #if defined(PETSC_USE_CTABLE)
2676: PetscTableDestroy((PetscTable *)&submatj->rmap);
2677: if (submatj->cmap_loc) PetscFree(submatj->cmap_loc);
2678: PetscFree(submatj->rmap_loc);
2679: #else
2680: PetscFree(submatj->rmap);
2681: #endif
2683: if (!submatj->allcolumns) {
2684: #if defined(PETSC_USE_CTABLE)
2685: PetscTableDestroy((PetscTable *)&submatj->cmap);
2686: #else
2687: PetscFree(submatj->cmap);
2688: #endif
2689: }
2690: PetscFree(submatj->row2proc);
2692: PetscFree(submatj);
2693: return 0;
2694: }
2696: PetscErrorCode MatDestroySubMatrix_SeqAIJ(Mat C)
2697: {
2698: Mat_SeqAIJ *c = (Mat_SeqAIJ *)C->data;
2699: Mat_SubSppt *submatj = c->submatis1;
2701: (*submatj->destroy)(C);
2702: MatDestroySubMatrix_Private(submatj);
2703: return 0;
2704: }
2706: /* Note this has code duplication with MatDestroySubMatrices_SeqBAIJ() */
2707: PetscErrorCode MatDestroySubMatrices_SeqAIJ(PetscInt n, Mat *mat[])
2708: {
2709: PetscInt i;
2710: Mat C;
2711: Mat_SeqAIJ *c;
2712: Mat_SubSppt *submatj;
2714: for (i = 0; i < n; i++) {
2715: C = (*mat)[i];
2716: c = (Mat_SeqAIJ *)C->data;
2717: submatj = c->submatis1;
2718: if (submatj) {
2719: if (--((PetscObject)C)->refct <= 0) {
2720: PetscFree(C->factorprefix);
2721: (*submatj->destroy)(C);
2722: MatDestroySubMatrix_Private(submatj);
2723: PetscFree(C->defaultvectype);
2724: PetscFree(C->defaultrandtype);
2725: PetscLayoutDestroy(&C->rmap);
2726: PetscLayoutDestroy(&C->cmap);
2727: PetscHeaderDestroy(&C);
2728: }
2729: } else {
2730: MatDestroy(&C);
2731: }
2732: }
2734: /* Destroy Dummy submatrices created for reuse */
2735: MatDestroySubMatrices_Dummy(n, mat);
2737: PetscFree(*mat);
2738: return 0;
2739: }
2741: PetscErrorCode MatCreateSubMatrices_SeqAIJ(Mat A, PetscInt n, const IS irow[], const IS icol[], MatReuse scall, Mat *B[])
2742: {
2743: PetscInt i;
2745: if (scall == MAT_INITIAL_MATRIX) PetscCalloc1(n + 1, B);
2747: for (i = 0; i < n; i++) MatCreateSubMatrix_SeqAIJ(A, irow[i], icol[i], PETSC_DECIDE, scall, &(*B)[i]);
2748: return 0;
2749: }
2751: PetscErrorCode MatIncreaseOverlap_SeqAIJ(Mat A, PetscInt is_max, IS is[], PetscInt ov)
2752: {
2753: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
2754: PetscInt row, i, j, k, l, ll, m, n, *nidx, isz, val;
2755: const PetscInt *idx;
2756: PetscInt start, end, *ai, *aj, bs = (A->rmap->bs > 0 && A->rmap->bs == A->cmap->bs) ? A->rmap->bs : 1;
2757: PetscBT table;
2759: m = A->rmap->n / bs;
2760: ai = a->i;
2761: aj = a->j;
2765: PetscMalloc1(m + 1, &nidx);
2766: PetscBTCreate(m, &table);
2768: for (i = 0; i < is_max; i++) {
2769: /* Initialize the two local arrays */
2770: isz = 0;
2771: PetscBTMemzero(m, table);
2773: /* Extract the indices, assume there can be duplicate entries */
2774: ISGetIndices(is[i], &idx);
2775: ISGetLocalSize(is[i], &n);
2777: if (bs > 1) {
2778: /* Enter these into the temp arrays. I.e., mark table[row], enter row into new index */
2779: for (j = 0; j < n; ++j) {
2780: if (!PetscBTLookupSet(table, idx[j] / bs)) nidx[isz++] = idx[j] / bs;
2781: }
2782: ISRestoreIndices(is[i], &idx);
2783: ISDestroy(&is[i]);
2785: k = 0;
2786: for (j = 0; j < ov; j++) { /* for each overlap */
2787: n = isz;
2788: for (; k < n; k++) { /* do only those rows in nidx[k], which are not done yet */
2789: for (ll = 0; ll < bs; ll++) {
2790: row = bs * nidx[k] + ll;
2791: start = ai[row];
2792: end = ai[row + 1];
2793: for (l = start; l < end; l++) {
2794: val = aj[l] / bs;
2795: if (!PetscBTLookupSet(table, val)) nidx[isz++] = val;
2796: }
2797: }
2798: }
2799: }
2800: ISCreateBlock(PETSC_COMM_SELF, bs, isz, nidx, PETSC_COPY_VALUES, (is + i));
2801: } else {
2802: /* Enter these into the temp arrays. I.e., mark table[row], enter row into new index */
2803: for (j = 0; j < n; ++j) {
2804: if (!PetscBTLookupSet(table, idx[j])) nidx[isz++] = idx[j];
2805: }
2806: ISRestoreIndices(is[i], &idx);
2807: ISDestroy(&is[i]);
2809: k = 0;
2810: for (j = 0; j < ov; j++) { /* for each overlap */
2811: n = isz;
2812: for (; k < n; k++) { /* do only those rows in nidx[k], which are not done yet */
2813: row = nidx[k];
2814: start = ai[row];
2815: end = ai[row + 1];
2816: for (l = start; l < end; l++) {
2817: val = aj[l];
2818: if (!PetscBTLookupSet(table, val)) nidx[isz++] = val;
2819: }
2820: }
2821: }
2822: ISCreateGeneral(PETSC_COMM_SELF, isz, nidx, PETSC_COPY_VALUES, (is + i));
2823: }
2824: }
2825: PetscBTDestroy(&table);
2826: PetscFree(nidx);
2827: return 0;
2828: }
2830: /* -------------------------------------------------------------- */
2831: PetscErrorCode MatPermute_SeqAIJ(Mat A, IS rowp, IS colp, Mat *B)
2832: {
2833: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
2834: PetscInt i, nz = 0, m = A->rmap->n, n = A->cmap->n;
2835: const PetscInt *row, *col;
2836: PetscInt *cnew, j, *lens;
2837: IS icolp, irowp;
2838: PetscInt *cwork = NULL;
2839: PetscScalar *vwork = NULL;
2841: ISInvertPermutation(rowp, PETSC_DECIDE, &irowp);
2842: ISGetIndices(irowp, &row);
2843: ISInvertPermutation(colp, PETSC_DECIDE, &icolp);
2844: ISGetIndices(icolp, &col);
2846: /* determine lengths of permuted rows */
2847: PetscMalloc1(m + 1, &lens);
2848: for (i = 0; i < m; i++) lens[row[i]] = a->i[i + 1] - a->i[i];
2849: MatCreate(PetscObjectComm((PetscObject)A), B);
2850: MatSetSizes(*B, m, n, m, n);
2851: MatSetBlockSizesFromMats(*B, A, A);
2852: MatSetType(*B, ((PetscObject)A)->type_name);
2853: MatSeqAIJSetPreallocation_SeqAIJ(*B, 0, lens);
2854: PetscFree(lens);
2856: PetscMalloc1(n, &cnew);
2857: for (i = 0; i < m; i++) {
2858: MatGetRow_SeqAIJ(A, i, &nz, &cwork, &vwork);
2859: for (j = 0; j < nz; j++) cnew[j] = col[cwork[j]];
2860: MatSetValues_SeqAIJ(*B, 1, &row[i], nz, cnew, vwork, INSERT_VALUES);
2861: MatRestoreRow_SeqAIJ(A, i, &nz, &cwork, &vwork);
2862: }
2863: PetscFree(cnew);
2865: (*B)->assembled = PETSC_FALSE;
2867: #if defined(PETSC_HAVE_DEVICE)
2868: MatBindToCPU(*B, A->boundtocpu);
2869: #endif
2870: MatAssemblyBegin(*B, MAT_FINAL_ASSEMBLY);
2871: MatAssemblyEnd(*B, MAT_FINAL_ASSEMBLY);
2872: ISRestoreIndices(irowp, &row);
2873: ISRestoreIndices(icolp, &col);
2874: ISDestroy(&irowp);
2875: ISDestroy(&icolp);
2876: if (rowp == colp) MatPropagateSymmetryOptions(A, *B);
2877: return 0;
2878: }
2880: PetscErrorCode MatCopy_SeqAIJ(Mat A, Mat B, MatStructure str)
2881: {
2882: /* If the two matrices have the same copy implementation, use fast copy. */
2883: if (str == SAME_NONZERO_PATTERN && (A->ops->copy == B->ops->copy)) {
2884: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
2885: Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data;
2886: const PetscScalar *aa;
2888: MatSeqAIJGetArrayRead(A, &aa);
2890: PetscArraycpy(b->a, aa, a->i[A->rmap->n]);
2891: PetscObjectStateIncrease((PetscObject)B);
2892: MatSeqAIJRestoreArrayRead(A, &aa);
2893: } else {
2894: MatCopy_Basic(A, B, str);
2895: }
2896: return 0;
2897: }
2899: PetscErrorCode MatSetUp_SeqAIJ(Mat A)
2900: {
2901: MatSeqAIJSetPreallocation_SeqAIJ(A, PETSC_DEFAULT, NULL);
2902: return 0;
2903: }
2905: PETSC_INTERN PetscErrorCode MatSeqAIJGetArray_SeqAIJ(Mat A, PetscScalar *array[])
2906: {
2907: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
2909: *array = a->a;
2910: return 0;
2911: }
2913: PETSC_INTERN PetscErrorCode MatSeqAIJRestoreArray_SeqAIJ(Mat A, PetscScalar *array[])
2914: {
2915: *array = NULL;
2916: return 0;
2917: }
2919: /*
2920: Computes the number of nonzeros per row needed for preallocation when X and Y
2921: have different nonzero structure.
2922: */
2923: PetscErrorCode MatAXPYGetPreallocation_SeqX_private(PetscInt m, const PetscInt *xi, const PetscInt *xj, const PetscInt *yi, const PetscInt *yj, PetscInt *nnz)
2924: {
2925: PetscInt i, j, k, nzx, nzy;
2927: /* Set the number of nonzeros in the new matrix */
2928: for (i = 0; i < m; i++) {
2929: const PetscInt *xjj = xj + xi[i], *yjj = yj + yi[i];
2930: nzx = xi[i + 1] - xi[i];
2931: nzy = yi[i + 1] - yi[i];
2932: nnz[i] = 0;
2933: for (j = 0, k = 0; j < nzx; j++) { /* Point in X */
2934: for (; k < nzy && yjj[k] < xjj[j]; k++) nnz[i]++; /* Catch up to X */
2935: if (k < nzy && yjj[k] == xjj[j]) k++; /* Skip duplicate */
2936: nnz[i]++;
2937: }
2938: for (; k < nzy; k++) nnz[i]++;
2939: }
2940: return 0;
2941: }
2943: PetscErrorCode MatAXPYGetPreallocation_SeqAIJ(Mat Y, Mat X, PetscInt *nnz)
2944: {
2945: PetscInt m = Y->rmap->N;
2946: Mat_SeqAIJ *x = (Mat_SeqAIJ *)X->data;
2947: Mat_SeqAIJ *y = (Mat_SeqAIJ *)Y->data;
2949: /* Set the number of nonzeros in the new matrix */
2950: MatAXPYGetPreallocation_SeqX_private(m, x->i, x->j, y->i, y->j, nnz);
2951: return 0;
2952: }
2954: PetscErrorCode MatAXPY_SeqAIJ(Mat Y, PetscScalar a, Mat X, MatStructure str)
2955: {
2956: Mat_SeqAIJ *x = (Mat_SeqAIJ *)X->data, *y = (Mat_SeqAIJ *)Y->data;
2958: if (str == UNKNOWN_NONZERO_PATTERN || (PetscDefined(USE_DEBUG) && str == SAME_NONZERO_PATTERN)) {
2959: PetscBool e = x->nz == y->nz ? PETSC_TRUE : PETSC_FALSE;
2960: if (e) {
2961: PetscArraycmp(x->i, y->i, Y->rmap->n + 1, &e);
2962: if (e) {
2963: PetscArraycmp(x->j, y->j, y->nz, &e);
2964: if (e) str = SAME_NONZERO_PATTERN;
2965: }
2966: }
2968: }
2969: if (str == SAME_NONZERO_PATTERN) {
2970: const PetscScalar *xa;
2971: PetscScalar *ya, alpha = a;
2972: PetscBLASInt one = 1, bnz;
2974: PetscBLASIntCast(x->nz, &bnz);
2975: MatSeqAIJGetArray(Y, &ya);
2976: MatSeqAIJGetArrayRead(X, &xa);
2977: PetscCallBLAS("BLASaxpy", BLASaxpy_(&bnz, &alpha, xa, &one, ya, &one));
2978: MatSeqAIJRestoreArrayRead(X, &xa);
2979: MatSeqAIJRestoreArray(Y, &ya);
2980: PetscLogFlops(2.0 * bnz);
2981: MatSeqAIJInvalidateDiagonal(Y);
2982: PetscObjectStateIncrease((PetscObject)Y);
2983: } else if (str == SUBSET_NONZERO_PATTERN) { /* nonzeros of X is a subset of Y's */
2984: MatAXPY_Basic(Y, a, X, str);
2985: } else {
2986: Mat B;
2987: PetscInt *nnz;
2988: PetscMalloc1(Y->rmap->N, &nnz);
2989: MatCreate(PetscObjectComm((PetscObject)Y), &B);
2990: PetscObjectSetName((PetscObject)B, ((PetscObject)Y)->name);
2991: MatSetLayouts(B, Y->rmap, Y->cmap);
2992: MatSetType(B, ((PetscObject)Y)->type_name);
2993: MatAXPYGetPreallocation_SeqAIJ(Y, X, nnz);
2994: MatSeqAIJSetPreallocation(B, 0, nnz);
2995: MatAXPY_BasicWithPreallocation(B, Y, a, X, str);
2996: MatHeaderMerge(Y, &B);
2997: MatSeqAIJCheckInode(Y);
2998: PetscFree(nnz);
2999: }
3000: return 0;
3001: }
3003: PETSC_INTERN PetscErrorCode MatConjugate_SeqAIJ(Mat mat)
3004: {
3005: #if defined(PETSC_USE_COMPLEX)
3006: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)mat->data;
3007: PetscInt i, nz;
3008: PetscScalar *a;
3010: nz = aij->nz;
3011: MatSeqAIJGetArray(mat, &a);
3012: for (i = 0; i < nz; i++) a[i] = PetscConj(a[i]);
3013: MatSeqAIJRestoreArray(mat, &a);
3014: #else
3015: #endif
3016: return 0;
3017: }
3019: PetscErrorCode MatGetRowMaxAbs_SeqAIJ(Mat A, Vec v, PetscInt idx[])
3020: {
3021: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
3022: PetscInt i, j, m = A->rmap->n, *ai, *aj, ncols, n;
3023: PetscReal atmp;
3024: PetscScalar *x;
3025: const MatScalar *aa, *av;
3028: MatSeqAIJGetArrayRead(A, &av);
3029: aa = av;
3030: ai = a->i;
3031: aj = a->j;
3033: VecSet(v, 0.0);
3034: VecGetArrayWrite(v, &x);
3035: VecGetLocalSize(v, &n);
3037: for (i = 0; i < m; i++) {
3038: ncols = ai[1] - ai[0];
3039: ai++;
3040: for (j = 0; j < ncols; j++) {
3041: atmp = PetscAbsScalar(*aa);
3042: if (PetscAbsScalar(x[i]) < atmp) {
3043: x[i] = atmp;
3044: if (idx) idx[i] = *aj;
3045: }
3046: aa++;
3047: aj++;
3048: }
3049: }
3050: VecRestoreArrayWrite(v, &x);
3051: MatSeqAIJRestoreArrayRead(A, &av);
3052: return 0;
3053: }
3055: PetscErrorCode MatGetRowMax_SeqAIJ(Mat A, Vec v, PetscInt idx[])
3056: {
3057: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
3058: PetscInt i, j, m = A->rmap->n, *ai, *aj, ncols, n;
3059: PetscScalar *x;
3060: const MatScalar *aa, *av;
3063: MatSeqAIJGetArrayRead(A, &av);
3064: aa = av;
3065: ai = a->i;
3066: aj = a->j;
3068: VecSet(v, 0.0);
3069: VecGetArrayWrite(v, &x);
3070: VecGetLocalSize(v, &n);
3072: for (i = 0; i < m; i++) {
3073: ncols = ai[1] - ai[0];
3074: ai++;
3075: if (ncols == A->cmap->n) { /* row is dense */
3076: x[i] = *aa;
3077: if (idx) idx[i] = 0;
3078: } else { /* row is sparse so already KNOW maximum is 0.0 or higher */
3079: x[i] = 0.0;
3080: if (idx) {
3081: for (j = 0; j < ncols; j++) { /* find first implicit 0.0 in the row */
3082: if (aj[j] > j) {
3083: idx[i] = j;
3084: break;
3085: }
3086: }
3087: /* in case first implicit 0.0 in the row occurs at ncols-th column */
3088: if (j == ncols && j < A->cmap->n) idx[i] = j;
3089: }
3090: }
3091: for (j = 0; j < ncols; j++) {
3092: if (PetscRealPart(x[i]) < PetscRealPart(*aa)) {
3093: x[i] = *aa;
3094: if (idx) idx[i] = *aj;
3095: }
3096: aa++;
3097: aj++;
3098: }
3099: }
3100: VecRestoreArrayWrite(v, &x);
3101: MatSeqAIJRestoreArrayRead(A, &av);
3102: return 0;
3103: }
3105: PetscErrorCode MatGetRowMinAbs_SeqAIJ(Mat A, Vec v, PetscInt idx[])
3106: {
3107: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
3108: PetscInt i, j, m = A->rmap->n, *ai, *aj, ncols, n;
3109: PetscScalar *x;
3110: const MatScalar *aa, *av;
3112: MatSeqAIJGetArrayRead(A, &av);
3113: aa = av;
3114: ai = a->i;
3115: aj = a->j;
3117: VecSet(v, 0.0);
3118: VecGetArrayWrite(v, &x);
3119: VecGetLocalSize(v, &n);
3121: for (i = 0; i < m; i++) {
3122: ncols = ai[1] - ai[0];
3123: ai++;
3124: if (ncols == A->cmap->n) { /* row is dense */
3125: x[i] = *aa;
3126: if (idx) idx[i] = 0;
3127: } else { /* row is sparse so already KNOW minimum is 0.0 or higher */
3128: x[i] = 0.0;
3129: if (idx) { /* find first implicit 0.0 in the row */
3130: for (j = 0; j < ncols; j++) {
3131: if (aj[j] > j) {
3132: idx[i] = j;
3133: break;
3134: }
3135: }
3136: /* in case first implicit 0.0 in the row occurs at ncols-th column */
3137: if (j == ncols && j < A->cmap->n) idx[i] = j;
3138: }
3139: }
3140: for (j = 0; j < ncols; j++) {
3141: if (PetscAbsScalar(x[i]) > PetscAbsScalar(*aa)) {
3142: x[i] = *aa;
3143: if (idx) idx[i] = *aj;
3144: }
3145: aa++;
3146: aj++;
3147: }
3148: }
3149: VecRestoreArrayWrite(v, &x);
3150: MatSeqAIJRestoreArrayRead(A, &av);
3151: return 0;
3152: }
3154: PetscErrorCode MatGetRowMin_SeqAIJ(Mat A, Vec v, PetscInt idx[])
3155: {
3156: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
3157: PetscInt i, j, m = A->rmap->n, ncols, n;
3158: const PetscInt *ai, *aj;
3159: PetscScalar *x;
3160: const MatScalar *aa, *av;
3163: MatSeqAIJGetArrayRead(A, &av);
3164: aa = av;
3165: ai = a->i;
3166: aj = a->j;
3168: VecSet(v, 0.0);
3169: VecGetArrayWrite(v, &x);
3170: VecGetLocalSize(v, &n);
3172: for (i = 0; i < m; i++) {
3173: ncols = ai[1] - ai[0];
3174: ai++;
3175: if (ncols == A->cmap->n) { /* row is dense */
3176: x[i] = *aa;
3177: if (idx) idx[i] = 0;
3178: } else { /* row is sparse so already KNOW minimum is 0.0 or lower */
3179: x[i] = 0.0;
3180: if (idx) { /* find first implicit 0.0 in the row */
3181: for (j = 0; j < ncols; j++) {
3182: if (aj[j] > j) {
3183: idx[i] = j;
3184: break;
3185: }
3186: }
3187: /* in case first implicit 0.0 in the row occurs at ncols-th column */
3188: if (j == ncols && j < A->cmap->n) idx[i] = j;
3189: }
3190: }
3191: for (j = 0; j < ncols; j++) {
3192: if (PetscRealPart(x[i]) > PetscRealPart(*aa)) {
3193: x[i] = *aa;
3194: if (idx) idx[i] = *aj;
3195: }
3196: aa++;
3197: aj++;
3198: }
3199: }
3200: VecRestoreArrayWrite(v, &x);
3201: MatSeqAIJRestoreArrayRead(A, &av);
3202: return 0;
3203: }
3205: PetscErrorCode MatInvertBlockDiagonal_SeqAIJ(Mat A, const PetscScalar **values)
3206: {
3207: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
3208: PetscInt i, bs = PetscAbs(A->rmap->bs), mbs = A->rmap->n / bs, ipvt[5], bs2 = bs * bs, *v_pivots, ij[7], *IJ, j;
3209: MatScalar *diag, work[25], *v_work;
3210: const PetscReal shift = 0.0;
3211: PetscBool allowzeropivot, zeropivotdetected = PETSC_FALSE;
3213: allowzeropivot = PetscNot(A->erroriffailure);
3214: if (a->ibdiagvalid) {
3215: if (values) *values = a->ibdiag;
3216: return 0;
3217: }
3218: MatMarkDiagonal_SeqAIJ(A);
3219: if (!a->ibdiag) { PetscMalloc1(bs2 * mbs, &a->ibdiag); }
3220: diag = a->ibdiag;
3221: if (values) *values = a->ibdiag;
3222: /* factor and invert each block */
3223: switch (bs) {
3224: case 1:
3225: for (i = 0; i < mbs; i++) {
3226: MatGetValues(A, 1, &i, 1, &i, diag + i);
3227: if (PetscAbsScalar(diag[i] + shift) < PETSC_MACHINE_EPSILON) {
3228: if (allowzeropivot) {
3229: A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
3230: A->factorerror_zeropivot_value = PetscAbsScalar(diag[i]);
3231: A->factorerror_zeropivot_row = i;
3232: PetscInfo(A, "Zero pivot, row %" PetscInt_FMT " pivot %g tolerance %g\n", i, (double)PetscAbsScalar(diag[i]), (double)PETSC_MACHINE_EPSILON);
3233: } 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);
3234: }
3235: diag[i] = (PetscScalar)1.0 / (diag[i] + shift);
3236: }
3237: break;
3238: case 2:
3239: for (i = 0; i < mbs; i++) {
3240: ij[0] = 2 * i;
3241: ij[1] = 2 * i + 1;
3242: MatGetValues(A, 2, ij, 2, ij, diag);
3243: PetscKernel_A_gets_inverse_A_2(diag, shift, allowzeropivot, &zeropivotdetected);
3244: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
3245: PetscKernel_A_gets_transpose_A_2(diag);
3246: diag += 4;
3247: }
3248: break;
3249: case 3:
3250: for (i = 0; i < mbs; i++) {
3251: ij[0] = 3 * i;
3252: ij[1] = 3 * i + 1;
3253: ij[2] = 3 * i + 2;
3254: MatGetValues(A, 3, ij, 3, ij, diag);
3255: PetscKernel_A_gets_inverse_A_3(diag, shift, allowzeropivot, &zeropivotdetected);
3256: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
3257: PetscKernel_A_gets_transpose_A_3(diag);
3258: diag += 9;
3259: }
3260: break;
3261: case 4:
3262: for (i = 0; i < mbs; i++) {
3263: ij[0] = 4 * i;
3264: ij[1] = 4 * i + 1;
3265: ij[2] = 4 * i + 2;
3266: ij[3] = 4 * i + 3;
3267: MatGetValues(A, 4, ij, 4, ij, diag);
3268: PetscKernel_A_gets_inverse_A_4(diag, shift, allowzeropivot, &zeropivotdetected);
3269: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
3270: PetscKernel_A_gets_transpose_A_4(diag);
3271: diag += 16;
3272: }
3273: break;
3274: case 5:
3275: for (i = 0; i < mbs; i++) {
3276: ij[0] = 5 * i;
3277: ij[1] = 5 * i + 1;
3278: ij[2] = 5 * i + 2;
3279: ij[3] = 5 * i + 3;
3280: ij[4] = 5 * i + 4;
3281: MatGetValues(A, 5, ij, 5, ij, diag);
3282: PetscKernel_A_gets_inverse_A_5(diag, ipvt, work, shift, allowzeropivot, &zeropivotdetected);
3283: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
3284: PetscKernel_A_gets_transpose_A_5(diag);
3285: diag += 25;
3286: }
3287: break;
3288: case 6:
3289: for (i = 0; i < mbs; i++) {
3290: ij[0] = 6 * i;
3291: ij[1] = 6 * i + 1;
3292: ij[2] = 6 * i + 2;
3293: ij[3] = 6 * i + 3;
3294: ij[4] = 6 * i + 4;
3295: ij[5] = 6 * i + 5;
3296: MatGetValues(A, 6, ij, 6, ij, diag);
3297: PetscKernel_A_gets_inverse_A_6(diag, shift, allowzeropivot, &zeropivotdetected);
3298: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
3299: PetscKernel_A_gets_transpose_A_6(diag);
3300: diag += 36;
3301: }
3302: break;
3303: case 7:
3304: for (i = 0; i < mbs; i++) {
3305: ij[0] = 7 * i;
3306: ij[1] = 7 * i + 1;
3307: ij[2] = 7 * i + 2;
3308: ij[3] = 7 * i + 3;
3309: ij[4] = 7 * i + 4;
3310: ij[5] = 7 * i + 5;
3311: ij[6] = 7 * i + 6;
3312: MatGetValues(A, 7, ij, 7, ij, diag);
3313: PetscKernel_A_gets_inverse_A_7(diag, shift, allowzeropivot, &zeropivotdetected);
3314: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
3315: PetscKernel_A_gets_transpose_A_7(diag);
3316: diag += 49;
3317: }
3318: break;
3319: default:
3320: PetscMalloc3(bs, &v_work, bs, &v_pivots, bs, &IJ);
3321: for (i = 0; i < mbs; i++) {
3322: for (j = 0; j < bs; j++) IJ[j] = bs * i + j;
3323: MatGetValues(A, bs, IJ, bs, IJ, diag);
3324: PetscKernel_A_gets_inverse_A(bs, diag, v_pivots, v_work, allowzeropivot, &zeropivotdetected);
3325: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
3326: PetscKernel_A_gets_transpose_A_N(diag, bs);
3327: diag += bs2;
3328: }
3329: PetscFree3(v_work, v_pivots, IJ);
3330: }
3331: a->ibdiagvalid = PETSC_TRUE;
3332: return 0;
3333: }
3335: static PetscErrorCode MatSetRandom_SeqAIJ(Mat x, PetscRandom rctx)
3336: {
3337: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)x->data;
3338: PetscScalar a, *aa;
3339: PetscInt m, n, i, j, col;
3341: if (!x->assembled) {
3342: MatGetSize(x, &m, &n);
3343: for (i = 0; i < m; i++) {
3344: for (j = 0; j < aij->imax[i]; j++) {
3345: PetscRandomGetValue(rctx, &a);
3346: col = (PetscInt)(n * PetscRealPart(a));
3347: MatSetValues(x, 1, &i, 1, &col, &a, ADD_VALUES);
3348: }
3349: }
3350: } else {
3351: MatSeqAIJGetArrayWrite(x, &aa);
3352: for (i = 0; i < aij->nz; i++) PetscRandomGetValue(rctx, aa + i);
3353: MatSeqAIJRestoreArrayWrite(x, &aa);
3354: }
3355: MatAssemblyBegin(x, MAT_FINAL_ASSEMBLY);
3356: MatAssemblyEnd(x, MAT_FINAL_ASSEMBLY);
3357: return 0;
3358: }
3360: /* Like MatSetRandom_SeqAIJ, but do not set values on columns in range of [low, high) */
3361: PetscErrorCode MatSetRandomSkipColumnRange_SeqAIJ_Private(Mat x, PetscInt low, PetscInt high, PetscRandom rctx)
3362: {
3363: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)x->data;
3364: PetscScalar a;
3365: PetscInt m, n, i, j, col, nskip;
3367: nskip = high - low;
3368: MatGetSize(x, &m, &n);
3369: n -= nskip; /* shrink number of columns where nonzeros can be set */
3370: for (i = 0; i < m; i++) {
3371: for (j = 0; j < aij->imax[i]; j++) {
3372: PetscRandomGetValue(rctx, &a);
3373: col = (PetscInt)(n * PetscRealPart(a));
3374: if (col >= low) col += nskip; /* shift col rightward to skip the hole */
3375: MatSetValues(x, 1, &i, 1, &col, &a, ADD_VALUES);
3376: }
3377: }
3378: MatAssemblyBegin(x, MAT_FINAL_ASSEMBLY);
3379: MatAssemblyEnd(x, MAT_FINAL_ASSEMBLY);
3380: return 0;
3381: }
3383: /* -------------------------------------------------------------------*/
3384: static struct _MatOps MatOps_Values = {MatSetValues_SeqAIJ,
3385: MatGetRow_SeqAIJ,
3386: MatRestoreRow_SeqAIJ,
3387: MatMult_SeqAIJ,
3388: /* 4*/ MatMultAdd_SeqAIJ,
3389: MatMultTranspose_SeqAIJ,
3390: MatMultTransposeAdd_SeqAIJ,
3391: NULL,
3392: NULL,
3393: NULL,
3394: /* 10*/ NULL,
3395: MatLUFactor_SeqAIJ,
3396: NULL,
3397: MatSOR_SeqAIJ,
3398: MatTranspose_SeqAIJ,
3399: /*1 5*/ MatGetInfo_SeqAIJ,
3400: MatEqual_SeqAIJ,
3401: MatGetDiagonal_SeqAIJ,
3402: MatDiagonalScale_SeqAIJ,
3403: MatNorm_SeqAIJ,
3404: /* 20*/ NULL,
3405: MatAssemblyEnd_SeqAIJ,
3406: MatSetOption_SeqAIJ,
3407: MatZeroEntries_SeqAIJ,
3408: /* 24*/ MatZeroRows_SeqAIJ,
3409: NULL,
3410: NULL,
3411: NULL,
3412: NULL,
3413: /* 29*/ MatSetUp_SeqAIJ,
3414: NULL,
3415: NULL,
3416: NULL,
3417: NULL,
3418: /* 34*/ MatDuplicate_SeqAIJ,
3419: NULL,
3420: NULL,
3421: MatILUFactor_SeqAIJ,
3422: NULL,
3423: /* 39*/ MatAXPY_SeqAIJ,
3424: MatCreateSubMatrices_SeqAIJ,
3425: MatIncreaseOverlap_SeqAIJ,
3426: MatGetValues_SeqAIJ,
3427: MatCopy_SeqAIJ,
3428: /* 44*/ MatGetRowMax_SeqAIJ,
3429: MatScale_SeqAIJ,
3430: MatShift_SeqAIJ,
3431: MatDiagonalSet_SeqAIJ,
3432: MatZeroRowsColumns_SeqAIJ,
3433: /* 49*/ MatSetRandom_SeqAIJ,
3434: MatGetRowIJ_SeqAIJ,
3435: MatRestoreRowIJ_SeqAIJ,
3436: MatGetColumnIJ_SeqAIJ,
3437: MatRestoreColumnIJ_SeqAIJ,
3438: /* 54*/ MatFDColoringCreate_SeqXAIJ,
3439: NULL,
3440: NULL,
3441: MatPermute_SeqAIJ,
3442: NULL,
3443: /* 59*/ NULL,
3444: MatDestroy_SeqAIJ,
3445: MatView_SeqAIJ,
3446: NULL,
3447: NULL,
3448: /* 64*/ NULL,
3449: MatMatMatMultNumeric_SeqAIJ_SeqAIJ_SeqAIJ,
3450: NULL,
3451: NULL,
3452: NULL,
3453: /* 69*/ MatGetRowMaxAbs_SeqAIJ,
3454: MatGetRowMinAbs_SeqAIJ,
3455: NULL,
3456: NULL,
3457: NULL,
3458: /* 74*/ NULL,
3459: MatFDColoringApply_AIJ,
3460: NULL,
3461: NULL,
3462: NULL,
3463: /* 79*/ MatFindZeroDiagonals_SeqAIJ,
3464: NULL,
3465: NULL,
3466: NULL,
3467: MatLoad_SeqAIJ,
3468: /* 84*/ MatIsSymmetric_SeqAIJ,
3469: MatIsHermitian_SeqAIJ,
3470: NULL,
3471: NULL,
3472: NULL,
3473: /* 89*/ NULL,
3474: NULL,
3475: MatMatMultNumeric_SeqAIJ_SeqAIJ,
3476: NULL,
3477: NULL,
3478: /* 94*/ MatPtAPNumeric_SeqAIJ_SeqAIJ_SparseAxpy,
3479: NULL,
3480: NULL,
3481: MatMatTransposeMultNumeric_SeqAIJ_SeqAIJ,
3482: NULL,
3483: /* 99*/ MatProductSetFromOptions_SeqAIJ,
3484: NULL,
3485: NULL,
3486: MatConjugate_SeqAIJ,
3487: NULL,
3488: /*104*/ MatSetValuesRow_SeqAIJ,
3489: MatRealPart_SeqAIJ,
3490: MatImaginaryPart_SeqAIJ,
3491: NULL,
3492: NULL,
3493: /*109*/ MatMatSolve_SeqAIJ,
3494: NULL,
3495: MatGetRowMin_SeqAIJ,
3496: NULL,
3497: MatMissingDiagonal_SeqAIJ,
3498: /*114*/ NULL,
3499: NULL,
3500: NULL,
3501: NULL,
3502: NULL,
3503: /*119*/ NULL,
3504: NULL,
3505: NULL,
3506: NULL,
3507: MatGetMultiProcBlock_SeqAIJ,
3508: /*124*/ MatFindNonzeroRows_SeqAIJ,
3509: MatGetColumnReductions_SeqAIJ,
3510: MatInvertBlockDiagonal_SeqAIJ,
3511: MatInvertVariableBlockDiagonal_SeqAIJ,
3512: NULL,
3513: /*129*/ NULL,
3514: NULL,
3515: NULL,
3516: MatTransposeMatMultNumeric_SeqAIJ_SeqAIJ,
3517: MatTransposeColoringCreate_SeqAIJ,
3518: /*134*/ MatTransColoringApplySpToDen_SeqAIJ,
3519: MatTransColoringApplyDenToSp_SeqAIJ,
3520: NULL,
3521: NULL,
3522: MatRARtNumeric_SeqAIJ_SeqAIJ,
3523: /*139*/ NULL,
3524: NULL,
3525: NULL,
3526: MatFDColoringSetUp_SeqXAIJ,
3527: MatFindOffBlockDiagonalEntries_SeqAIJ,
3528: MatCreateMPIMatConcatenateSeqMat_SeqAIJ,
3529: /*145*/ MatDestroySubMatrices_SeqAIJ,
3530: NULL,
3531: NULL,
3532: MatCreateGraph_Simple_AIJ,
3533: NULL,
3534: /*150*/ MatTransposeSymbolic_SeqAIJ};
3536: PetscErrorCode MatSeqAIJSetColumnIndices_SeqAIJ(Mat mat, PetscInt *indices)
3537: {
3538: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)mat->data;
3539: PetscInt i, nz, n;
3541: nz = aij->maxnz;
3542: n = mat->rmap->n;
3543: for (i = 0; i < nz; i++) aij->j[i] = indices[i];
3544: aij->nz = nz;
3545: for (i = 0; i < n; i++) aij->ilen[i] = aij->imax[i];
3546: return 0;
3547: }
3549: /*
3550: * Given a sparse matrix with global column indices, compact it by using a local column space.
3551: * The result matrix helps saving memory in other algorithms, such as MatPtAPSymbolic_MPIAIJ_MPIAIJ_scalable()
3552: */
3553: PetscErrorCode MatSeqAIJCompactOutExtraColumns_SeqAIJ(Mat mat, ISLocalToGlobalMapping *mapping)
3554: {
3555: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)mat->data;
3556: PetscTable gid1_lid1;
3557: PetscTablePosition tpos;
3558: PetscInt gid, lid, i, ec, nz = aij->nz;
3559: PetscInt *garray, *jj = aij->j;
3563: /* use a table */
3564: PetscTableCreate(mat->rmap->n, mat->cmap->N + 1, &gid1_lid1);
3565: ec = 0;
3566: for (i = 0; i < nz; i++) {
3567: PetscInt data, gid1 = jj[i] + 1;
3568: PetscTableFind(gid1_lid1, gid1, &data);
3569: if (!data) {
3570: /* one based table */
3571: PetscTableAdd(gid1_lid1, gid1, ++ec, INSERT_VALUES);
3572: }
3573: }
3574: /* form array of columns we need */
3575: PetscMalloc1(ec, &garray);
3576: PetscTableGetHeadPosition(gid1_lid1, &tpos);
3577: while (tpos) {
3578: PetscTableGetNext(gid1_lid1, &tpos, &gid, &lid);
3579: gid--;
3580: lid--;
3581: garray[lid] = gid;
3582: }
3583: PetscSortInt(ec, garray); /* sort, and rebuild */
3584: PetscTableRemoveAll(gid1_lid1);
3585: for (i = 0; i < ec; i++) PetscTableAdd(gid1_lid1, garray[i] + 1, i + 1, INSERT_VALUES);
3586: /* compact out the extra columns in B */
3587: for (i = 0; i < nz; i++) {
3588: PetscInt gid1 = jj[i] + 1;
3589: PetscTableFind(gid1_lid1, gid1, &lid);
3590: lid--;
3591: jj[i] = lid;
3592: }
3593: PetscLayoutDestroy(&mat->cmap);
3594: PetscTableDestroy(&gid1_lid1);
3595: PetscLayoutCreateFromSizes(PetscObjectComm((PetscObject)mat), ec, ec, 1, &mat->cmap);
3596: ISLocalToGlobalMappingCreate(PETSC_COMM_SELF, mat->cmap->bs, mat->cmap->n, garray, PETSC_OWN_POINTER, mapping);
3597: ISLocalToGlobalMappingSetType(*mapping, ISLOCALTOGLOBALMAPPINGHASH);
3598: return 0;
3599: }
3601: /*@
3602: MatSeqAIJSetColumnIndices - Set the column indices for all the rows
3603: in the matrix.
3605: Input Parameters:
3606: + mat - the `MATSEQAIJ` matrix
3607: - indices - the column indices
3609: Level: advanced
3611: Notes:
3612: This can be called if you have precomputed the nonzero structure of the
3613: matrix and want to provide it to the matrix object to improve the performance
3614: of the `MatSetValues()` operation.
3616: You MUST have set the correct numbers of nonzeros per row in the call to
3617: `MatCreateSeqAIJ()`, and the columns indices MUST be sorted.
3619: MUST be called before any calls to `MatSetValues()`
3621: The indices should start with zero, not one.
3623: @*/
3624: PetscErrorCode MatSeqAIJSetColumnIndices(Mat mat, PetscInt *indices)
3625: {
3628: PetscUseMethod(mat, "MatSeqAIJSetColumnIndices_C", (Mat, PetscInt *), (mat, indices));
3629: return 0;
3630: }
3632: /* ----------------------------------------------------------------------------------------*/
3634: PetscErrorCode MatStoreValues_SeqAIJ(Mat mat)
3635: {
3636: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)mat->data;
3637: size_t nz = aij->i[mat->rmap->n];
3641: /* allocate space for values if not already there */
3642: if (!aij->saved_values) { PetscMalloc1(nz + 1, &aij->saved_values); }
3644: /* copy values over */
3645: PetscArraycpy(aij->saved_values, aij->a, nz);
3646: return 0;
3647: }
3649: /*@
3650: MatStoreValues - Stashes a copy of the matrix values; this allows, for
3651: example, reuse of the linear part of a Jacobian, while recomputing the
3652: nonlinear portion.
3654: Logically Collect
3656: Input Parameters:
3657: . mat - the matrix (currently only `MATAIJ` matrices support this option)
3659: Level: advanced
3661: Common Usage, with `SNESSolve()`:
3662: $ Create Jacobian matrix
3663: $ Set linear terms into matrix
3664: $ Apply boundary conditions to matrix, at this time matrix must have
3665: $ final nonzero structure (i.e. setting the nonlinear terms and applying
3666: $ boundary conditions again will not change the nonzero structure
3667: $ MatSetOption(mat,MAT_NEW_NONZERO_LOCATIONS,PETSC_FALSE);
3668: $ MatStoreValues(mat);
3669: $ Call SNESSetJacobian() with matrix
3670: $ In your Jacobian routine
3671: $ MatRetrieveValues(mat);
3672: $ Set nonlinear terms in matrix
3674: Common Usage without SNESSolve(), i.e. when you handle nonlinear solve yourself:
3675: $ // build linear portion of Jacobian
3676: $ MatSetOption(mat,MAT_NEW_NONZERO_LOCATIONS,PETSC_FALSE);
3677: $ MatStoreValues(mat);
3678: $ loop over nonlinear iterations
3679: $ MatRetrieveValues(mat);
3680: $ // call MatSetValues(mat,...) to set nonliner portion of Jacobian
3681: $ // call MatAssemblyBegin/End() on matrix
3682: $ Solve linear system with Jacobian
3683: $ endloop
3685: Notes:
3686: Matrix must already be assembled before calling this routine
3687: Must set the matrix option `MatSetOption`(mat,`MAT_NEW_NONZERO_LOCATIONS`,`PETSC_FALSE`); before
3688: calling this routine.
3690: When this is called multiple times it overwrites the previous set of stored values
3691: and does not allocated additional space.
3693: .seealso: `MatRetrieveValues()`
3694: @*/
3695: PetscErrorCode MatStoreValues(Mat mat)
3696: {
3700: PetscUseMethod(mat, "MatStoreValues_C", (Mat), (mat));
3701: return 0;
3702: }
3704: PetscErrorCode MatRetrieveValues_SeqAIJ(Mat mat)
3705: {
3706: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)mat->data;
3707: PetscInt nz = aij->i[mat->rmap->n];
3711: /* copy values over */
3712: PetscArraycpy(aij->a, aij->saved_values, nz);
3713: return 0;
3714: }
3716: /*@
3717: MatRetrieveValues - Retrieves the copy of the matrix values; this allows, for
3718: example, reuse of the linear part of a Jacobian, while recomputing the
3719: nonlinear portion.
3721: Logically Collect
3723: Input Parameters:
3724: . mat - the matrix (currently only `MATAIJ` matrices support this option)
3726: Level: advanced
3728: .seealso: `MatStoreValues()`
3729: @*/
3730: PetscErrorCode MatRetrieveValues(Mat mat)
3731: {
3735: PetscUseMethod(mat, "MatRetrieveValues_C", (Mat), (mat));
3736: return 0;
3737: }
3739: /* --------------------------------------------------------------------------------*/
3740: /*@C
3741: MatCreateSeqAIJ - Creates a sparse matrix in `MATSEQAIJ` (compressed row) format
3742: (the default parallel PETSc format). For good matrix assembly performance
3743: the user should preallocate the matrix storage by setting the parameter nz
3744: (or the array nnz). By setting these parameters accurately, performance
3745: during matrix assembly can be increased by more than a factor of 50.
3747: Collective
3749: Input Parameters:
3750: + comm - MPI communicator, set to `PETSC_COMM_SELF`
3751: . m - number of rows
3752: . n - number of columns
3753: . nz - number of nonzeros per row (same for all rows)
3754: - nnz - array containing the number of nonzeros in the various rows
3755: (possibly different for each row) or NULL
3757: Output Parameter:
3758: . A - the matrix
3760: It is recommended that one use the `MatCreate()`, `MatSetType()` and/or `MatSetFromOptions()`,
3761: MatXXXXSetPreallocation() paradigm instead of this routine directly.
3762: [MatXXXXSetPreallocation() is, for example, `MatSeqAIJSetPreallocation()`]
3764: Notes:
3765: If nnz is given then nz is ignored
3767: The AIJ format, also called
3768: compressed row storage, is fully compatible with standard Fortran 77
3769: storage. That is, the stored row and column indices can begin at
3770: either one (as in Fortran) or zero. See the users' manual for details.
3772: Specify the preallocated storage with either nz or nnz (not both).
3773: Set nz = `PETSC_DEFAULT` and nnz = NULL for PETSc to control dynamic memory
3774: allocation. For large problems you MUST preallocate memory or you
3775: will get TERRIBLE performance, see the users' manual chapter on matrices.
3777: By default, this format uses inodes (identical nodes) when possible, to
3778: improve numerical efficiency of matrix-vector products and solves. We
3779: search for consecutive rows with the same nonzero structure, thereby
3780: reusing matrix information to achieve increased efficiency.
3782: Options Database Keys:
3783: + -mat_no_inode - Do not use inodes
3784: - -mat_inode_limit <limit> - Sets inode limit (max limit=5)
3786: Level: intermediate
3788: .seealso: [Sparse Matrix Creation](sec_matsparse), `MatCreate()`, `MatCreateAIJ()`, `MatSetValues()`, `MatSeqAIJSetColumnIndices()`, `MatCreateSeqAIJWithArrays()`
3789: @*/
3790: PetscErrorCode MatCreateSeqAIJ(MPI_Comm comm, PetscInt m, PetscInt n, PetscInt nz, const PetscInt nnz[], Mat *A)
3791: {
3792: MatCreate(comm, A);
3793: MatSetSizes(*A, m, n, m, n);
3794: MatSetType(*A, MATSEQAIJ);
3795: MatSeqAIJSetPreallocation_SeqAIJ(*A, nz, nnz);
3796: return 0;
3797: }
3799: /*@C
3800: MatSeqAIJSetPreallocation - For good matrix assembly performance
3801: the user should preallocate the matrix storage by setting the parameter nz
3802: (or the array nnz). By setting these parameters accurately, performance
3803: during matrix assembly can be increased by more than a factor of 50.
3805: Collective
3807: Input Parameters:
3808: + B - The matrix
3809: . nz - number of nonzeros per row (same for all rows)
3810: - nnz - array containing the number of nonzeros in the various rows
3811: (possibly different for each row) or NULL
3813: Notes:
3814: If nnz is given then nz is ignored
3816: The `MATSEQAIJ` format also called
3817: compressed row storage, is fully compatible with standard Fortran 77
3818: storage. That is, the stored row and column indices can begin at
3819: either one (as in Fortran) or zero. See the users' manual for details.
3821: Specify the preallocated storage with either nz or nnz (not both).
3822: Set nz = `PETSC_DEFAULT` and nnz = NULL for PETSc to control dynamic memory
3823: allocation. For large problems you MUST preallocate memory or you
3824: will get TERRIBLE performance, see the users' manual chapter on matrices.
3826: You can call `MatGetInfo()` to get information on how effective the preallocation was;
3827: for example the fields mallocs,nz_allocated,nz_used,nz_unneeded;
3828: You can also run with the option -info and look for messages with the string
3829: malloc in them to see if additional memory allocation was needed.
3831: Developer Notes:
3832: Use nz of `MAT_SKIP_ALLOCATION` to not allocate any space for the matrix
3833: entries or columns indices
3835: By default, this format uses inodes (identical nodes) when possible, to
3836: improve numerical efficiency of matrix-vector products and solves. We
3837: search for consecutive rows with the same nonzero structure, thereby
3838: reusing matrix information to achieve increased efficiency.
3840: Options Database Keys:
3841: + -mat_no_inode - Do not use inodes
3842: - -mat_inode_limit <limit> - Sets inode limit (max limit=5)
3844: Level: intermediate
3846: .seealso: `MatCreate()`, `MatCreateAIJ()`, `MatSetValues()`, `MatSeqAIJSetColumnIndices()`, `MatCreateSeqAIJWithArrays()`, `MatGetInfo()`,
3847: `MatSeqAIJSetTotalPreallocation()`
3848: @*/
3849: PetscErrorCode MatSeqAIJSetPreallocation(Mat B, PetscInt nz, const PetscInt nnz[])
3850: {
3853: PetscTryMethod(B, "MatSeqAIJSetPreallocation_C", (Mat, PetscInt, const PetscInt[]), (B, nz, nnz));
3854: return 0;
3855: }
3857: PetscErrorCode MatSeqAIJSetPreallocation_SeqAIJ(Mat B, PetscInt nz, const PetscInt *nnz)
3858: {
3859: Mat_SeqAIJ *b;
3860: PetscBool skipallocation = PETSC_FALSE, realalloc = PETSC_FALSE;
3861: PetscInt i;
3863: if (nz >= 0 || nnz) realalloc = PETSC_TRUE;
3864: if (nz == MAT_SKIP_ALLOCATION) {
3865: skipallocation = PETSC_TRUE;
3866: nz = 0;
3867: }
3868: PetscLayoutSetUp(B->rmap);
3869: PetscLayoutSetUp(B->cmap);
3871: if (nz == PETSC_DEFAULT || nz == PETSC_DECIDE) nz = 5;
3873: if (PetscUnlikelyDebug(nnz)) {
3874: for (i = 0; i < B->rmap->n; i++) {
3877: }
3878: }
3880: B->preallocated = PETSC_TRUE;
3882: b = (Mat_SeqAIJ *)B->data;
3884: if (!skipallocation) {
3885: if (!b->imax) { PetscMalloc1(B->rmap->n, &b->imax); }
3886: if (!b->ilen) {
3887: /* b->ilen will count nonzeros in each row so far. */
3888: PetscCalloc1(B->rmap->n, &b->ilen);
3889: } else {
3890: PetscMemzero(b->ilen, B->rmap->n * sizeof(PetscInt));
3891: }
3892: if (!b->ipre) { PetscMalloc1(B->rmap->n, &b->ipre); }
3893: if (!nnz) {
3894: if (nz == PETSC_DEFAULT || nz == PETSC_DECIDE) nz = 10;
3895: else if (nz < 0) nz = 1;
3896: nz = PetscMin(nz, B->cmap->n);
3897: for (i = 0; i < B->rmap->n; i++) b->imax[i] = nz;
3898: nz = nz * B->rmap->n;
3899: } else {
3900: PetscInt64 nz64 = 0;
3901: for (i = 0; i < B->rmap->n; i++) {
3902: b->imax[i] = nnz[i];
3903: nz64 += nnz[i];
3904: }
3905: PetscIntCast(nz64, &nz);
3906: }
3908: /* allocate the matrix space */
3909: /* FIXME: should B's old memory be unlogged? */
3910: MatSeqXAIJFreeAIJ(B, &b->a, &b->j, &b->i);
3911: if (B->structure_only) {
3912: PetscMalloc1(nz, &b->j);
3913: PetscMalloc1(B->rmap->n + 1, &b->i);
3914: } else {
3915: PetscMalloc3(nz, &b->a, nz, &b->j, B->rmap->n + 1, &b->i);
3916: }
3917: b->i[0] = 0;
3918: for (i = 1; i < B->rmap->n + 1; i++) b->i[i] = b->i[i - 1] + b->imax[i - 1];
3919: if (B->structure_only) {
3920: b->singlemalloc = PETSC_FALSE;
3921: b->free_a = PETSC_FALSE;
3922: } else {
3923: b->singlemalloc = PETSC_TRUE;
3924: b->free_a = PETSC_TRUE;
3925: }
3926: b->free_ij = PETSC_TRUE;
3927: } else {
3928: b->free_a = PETSC_FALSE;
3929: b->free_ij = PETSC_FALSE;
3930: }
3932: if (b->ipre && nnz != b->ipre && b->imax) {
3933: /* reserve user-requested sparsity */
3934: PetscArraycpy(b->ipre, b->imax, B->rmap->n);
3935: }
3937: b->nz = 0;
3938: b->maxnz = nz;
3939: B->info.nz_unneeded = (double)b->maxnz;
3940: if (realalloc) MatSetOption(B, MAT_NEW_NONZERO_ALLOCATION_ERR, PETSC_TRUE);
3941: B->was_assembled = PETSC_FALSE;
3942: B->assembled = PETSC_FALSE;
3943: /* We simply deem preallocation has changed nonzero state. Updating the state
3944: will give clients (like AIJKokkos) a chance to know something has happened.
3945: */
3946: B->nonzerostate++;
3947: return 0;
3948: }
3950: PetscErrorCode MatResetPreallocation_SeqAIJ(Mat A)
3951: {
3952: Mat_SeqAIJ *a;
3953: PetscInt i;
3957: /* Check local size. If zero, then return */
3958: if (!A->rmap->n) return 0;
3960: a = (Mat_SeqAIJ *)A->data;
3961: /* if no saved info, we error out */
3966: PetscArraycpy(a->imax, a->ipre, A->rmap->n);
3967: PetscArrayzero(a->ilen, A->rmap->n);
3968: a->i[0] = 0;
3969: for (i = 1; i < A->rmap->n + 1; i++) a->i[i] = a->i[i - 1] + a->imax[i - 1];
3970: A->preallocated = PETSC_TRUE;
3971: a->nz = 0;
3972: a->maxnz = a->i[A->rmap->n];
3973: A->info.nz_unneeded = (double)a->maxnz;
3974: A->was_assembled = PETSC_FALSE;
3975: A->assembled = PETSC_FALSE;
3976: return 0;
3977: }
3979: /*@
3980: MatSeqAIJSetPreallocationCSR - Allocates memory for a sparse sequential matrix in `MATSEQAIJ` format.
3982: Input Parameters:
3983: + B - the matrix
3984: . i - the indices into j for the start of each row (starts with zero)
3985: . j - the column indices for each row (starts with zero) these must be sorted for each row
3986: - v - optional values in the matrix
3988: Level: developer
3990: Notes:
3991: The i,j,v values are COPIED with this routine; to avoid the copy use `MatCreateSeqAIJWithArrays()`
3993: This routine may be called multiple times with different nonzero patterns (or the same nonzero pattern). The nonzero
3994: structure will be the union of all the previous nonzero structures.
3996: Developer Notes:
3997: 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
3998: then just copies the v values directly with `PetscMemcpy()`.
4000: This routine could also take a `PetscCopyMode` argument to allow sharing the values instead of always copying them.
4002: .seealso: `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatSeqAIJSetPreallocation()`, `MatCreateSeqAIJ()`, `MATSEQAIJ`, `MatResetPreallocation()`
4003: @*/
4004: PetscErrorCode MatSeqAIJSetPreallocationCSR(Mat B, const PetscInt i[], const PetscInt j[], const PetscScalar v[])
4005: {
4008: PetscTryMethod(B, "MatSeqAIJSetPreallocationCSR_C", (Mat, const PetscInt[], const PetscInt[], const PetscScalar[]), (B, i, j, v));
4009: return 0;
4010: }
4012: PetscErrorCode MatSeqAIJSetPreallocationCSR_SeqAIJ(Mat B, const PetscInt Ii[], const PetscInt J[], const PetscScalar v[])
4013: {
4014: PetscInt i;
4015: PetscInt m, n;
4016: PetscInt nz;
4017: PetscInt *nnz;
4021: PetscLayoutSetUp(B->rmap);
4022: PetscLayoutSetUp(B->cmap);
4024: MatGetSize(B, &m, &n);
4025: PetscMalloc1(m + 1, &nnz);
4026: for (i = 0; i < m; i++) {
4027: nz = Ii[i + 1] - Ii[i];
4029: nnz[i] = nz;
4030: }
4031: MatSeqAIJSetPreallocation(B, 0, nnz);
4032: PetscFree(nnz);
4034: for (i = 0; i < m; i++) MatSetValues_SeqAIJ(B, 1, &i, Ii[i + 1] - Ii[i], J + Ii[i], v ? v + Ii[i] : NULL, INSERT_VALUES);
4036: MatAssemblyBegin(B, MAT_FINAL_ASSEMBLY);
4037: MatAssemblyEnd(B, MAT_FINAL_ASSEMBLY);
4039: MatSetOption(B, MAT_NEW_NONZERO_LOCATION_ERR, PETSC_TRUE);
4040: return 0;
4041: }
4043: /*@
4044: MatSeqAIJKron - Computes C, the Kronecker product of A and B.
4046: Input Parameters:
4047: + A - left-hand side matrix
4048: . B - right-hand side matrix
4049: - reuse - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX`
4051: Output Parameter:
4052: . C - Kronecker product of A and B
4054: Level: intermediate
4056: Note:
4057: `MAT_REUSE_MATRIX` can only be used when the nonzero structure of the product matrix has not changed from that last call to `MatSeqAIJKron()`.
4059: .seealso: `MatCreateSeqAIJ()`, `MATSEQAIJ`, `MATKAIJ`, `MatReuse`
4060: @*/
4061: PetscErrorCode MatSeqAIJKron(Mat A, Mat B, MatReuse reuse, Mat *C)
4062: {
4068: if (reuse == MAT_REUSE_MATRIX) {
4071: }
4072: PetscTryMethod(A, "MatSeqAIJKron_C", (Mat, Mat, MatReuse, Mat *), (A, B, reuse, C));
4073: return 0;
4074: }
4076: PetscErrorCode MatSeqAIJKron_SeqAIJ(Mat A, Mat B, MatReuse reuse, Mat *C)
4077: {
4078: Mat newmat;
4079: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
4080: Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data;
4081: PetscScalar *v;
4082: const PetscScalar *aa, *ba;
4083: 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;
4084: PetscBool flg;
4090: PetscObjectTypeCompare((PetscObject)B, MATSEQAIJ, &flg);
4093: if (reuse == MAT_INITIAL_MATRIX) {
4094: PetscMalloc2(am * bm + 1, &i, a->i[am] * b->i[bm], &j);
4095: MatCreate(PETSC_COMM_SELF, &newmat);
4096: MatSetSizes(newmat, am * bm, an * bn, am * bm, an * bn);
4097: MatSetType(newmat, MATAIJ);
4098: i[0] = 0;
4099: for (m = 0; m < am; ++m) {
4100: for (p = 0; p < bm; ++p) {
4101: i[m * bm + p + 1] = i[m * bm + p] + (a->i[m + 1] - a->i[m]) * (b->i[p + 1] - b->i[p]);
4102: for (n = a->i[m]; n < a->i[m + 1]; ++n) {
4103: for (q = b->i[p]; q < b->i[p + 1]; ++q) j[nnz++] = a->j[n] * bn + b->j[q];
4104: }
4105: }
4106: }
4107: MatSeqAIJSetPreallocationCSR(newmat, i, j, NULL);
4108: *C = newmat;
4109: PetscFree2(i, j);
4110: nnz = 0;
4111: }
4112: MatSeqAIJGetArray(*C, &v);
4113: MatSeqAIJGetArrayRead(A, &aa);
4114: MatSeqAIJGetArrayRead(B, &ba);
4115: for (m = 0; m < am; ++m) {
4116: for (p = 0; p < bm; ++p) {
4117: for (n = a->i[m]; n < a->i[m + 1]; ++n) {
4118: for (q = b->i[p]; q < b->i[p + 1]; ++q) v[nnz++] = aa[n] * ba[q];
4119: }
4120: }
4121: }
4122: MatSeqAIJRestoreArray(*C, &v);
4123: MatSeqAIJRestoreArrayRead(A, &aa);
4124: MatSeqAIJRestoreArrayRead(B, &ba);
4125: return 0;
4126: }
4128: #include <../src/mat/impls/dense/seq/dense.h>
4129: #include <petsc/private/kernels/petscaxpy.h>
4131: /*
4132: Computes (B'*A')' since computing B*A directly is untenable
4134: n p p
4135: [ ] [ ] [ ]
4136: m [ A ] * n [ B ] = m [ C ]
4137: [ ] [ ] [ ]
4139: */
4140: PetscErrorCode MatMatMultNumeric_SeqDense_SeqAIJ(Mat A, Mat B, Mat C)
4141: {
4142: Mat_SeqDense *sub_a = (Mat_SeqDense *)A->data;
4143: Mat_SeqAIJ *sub_b = (Mat_SeqAIJ *)B->data;
4144: Mat_SeqDense *sub_c = (Mat_SeqDense *)C->data;
4145: PetscInt i, j, n, m, q, p;
4146: const PetscInt *ii, *idx;
4147: const PetscScalar *b, *a, *a_q;
4148: PetscScalar *c, *c_q;
4149: PetscInt clda = sub_c->lda;
4150: PetscInt alda = sub_a->lda;
4152: m = A->rmap->n;
4153: n = A->cmap->n;
4154: p = B->cmap->n;
4155: a = sub_a->v;
4156: b = sub_b->a;
4157: c = sub_c->v;
4158: if (clda == m) {
4159: PetscArrayzero(c, m * p);
4160: } else {
4161: for (j = 0; j < p; j++)
4162: for (i = 0; i < m; i++) c[j * clda + i] = 0.0;
4163: }
4164: ii = sub_b->i;
4165: idx = sub_b->j;
4166: for (i = 0; i < n; i++) {
4167: q = ii[i + 1] - ii[i];
4168: while (q-- > 0) {
4169: c_q = c + clda * (*idx);
4170: a_q = a + alda * i;
4171: PetscKernelAXPY(c_q, *b, a_q, m);
4172: idx++;
4173: b++;
4174: }
4175: }
4176: return 0;
4177: }
4179: PetscErrorCode MatMatMultSymbolic_SeqDense_SeqAIJ(Mat A, Mat B, PetscReal fill, Mat C)
4180: {
4181: PetscInt m = A->rmap->n, n = B->cmap->n;
4182: PetscBool cisdense;
4185: MatSetSizes(C, m, n, m, n);
4186: MatSetBlockSizesFromMats(C, A, B);
4187: PetscObjectTypeCompareAny((PetscObject)C, &cisdense, MATSEQDENSE, MATSEQDENSECUDA, "");
4188: if (!cisdense) MatSetType(C, MATDENSE);
4189: MatSetUp(C);
4191: C->ops->matmultnumeric = MatMatMultNumeric_SeqDense_SeqAIJ;
4192: return 0;
4193: }
4195: /* ----------------------------------------------------------------*/
4196: /*MC
4197: MATSEQAIJ - MATSEQAIJ = "seqaij" - A matrix type to be used for sequential sparse matrices,
4198: based on compressed sparse row format.
4200: Options Database Keys:
4201: . -mat_type seqaij - sets the matrix type to "seqaij" during a call to MatSetFromOptions()
4203: Level: beginner
4205: Notes:
4206: `MatSetValues()` may be called for this matrix type with a NULL argument for the numerical values,
4207: in this case the values associated with the rows and columns one passes in are set to zero
4208: in the matrix
4210: `MatSetOptions`(,`MAT_STRUCTURE_ONLY`,`PETSC_TRUE`) may be called for this matrix type. In this no
4211: space is allocated for the nonzero entries and any entries passed with `MatSetValues()` are ignored
4213: Developer Note:
4214: It would be nice if all matrix formats supported passing NULL in for the numerical values
4216: .seealso: `MatCreateSeqAIJ()`, `MatSetFromOptions()`, `MatSetType()`, `MatCreate()`, `MatType`, `MATSELL`, `MATSEQSELL`, `MATMPISELL`
4217: M*/
4219: /*MC
4220: MATAIJ - MATAIJ = "aij" - A matrix type to be used for sparse matrices.
4222: This matrix type is identical to `MATSEQAIJ` when constructed with a single process communicator,
4223: and `MATMPIAIJ` otherwise. As a result, for single process communicators,
4224: `MatSeqAIJSetPreallocation()` is supported, and similarly `MatMPIAIJSetPreallocation()` is supported
4225: for communicators controlling multiple processes. It is recommended that you call both of
4226: the above preallocation routines for simplicity.
4228: Options Database Keys:
4229: . -mat_type aij - sets the matrix type to "aij" during a call to `MatSetFromOptions()`
4231: Note:
4232: Subclasses include `MATAIJCUSPARSE`, `MATAIJPERM`, `MATAIJSELL`, `MATAIJMKL`, `MATAIJCRL`, and also automatically switches over to use inodes when
4233: enough exist.
4235: Level: beginner
4237: .seealso: `MatCreateAIJ()`, `MatCreateSeqAIJ()`, `MATSEQAIJ`, `MATMPIAIJ`, `MATSELL`, `MATSEQSELL`, `MATMPISELL`
4238: M*/
4240: /*MC
4241: MATAIJCRL - MATAIJCRL = "aijcrl" - A matrix type to be used for sparse matrices.
4243: This matrix type is identical to `MATSEQAIJCRL` when constructed with a single process communicator,
4244: and `MATMPIAIJCRL` otherwise. As a result, for single process communicators,
4245: `MatSeqAIJSetPreallocation()` is supported, and similarly `MatMPIAIJSetPreallocation()` is supported
4246: for communicators controlling multiple processes. It is recommended that you call both of
4247: the above preallocation routines for simplicity.
4249: Options Database Keys:
4250: . -mat_type aijcrl - sets the matrix type to "aijcrl" during a call to `MatSetFromOptions()`
4252: Level: beginner
4254: .seealso: `MatCreateMPIAIJCRL`, `MATSEQAIJCRL`, `MATMPIAIJCRL`, `MATSEQAIJCRL`, `MATMPIAIJCRL`
4255: M*/
4257: PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqAIJCRL(Mat, MatType, MatReuse, Mat *);
4258: #if defined(PETSC_HAVE_ELEMENTAL)
4259: PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_Elemental(Mat, MatType, MatReuse, Mat *);
4260: #endif
4261: #if defined(PETSC_HAVE_SCALAPACK)
4262: PETSC_INTERN PetscErrorCode MatConvert_AIJ_ScaLAPACK(Mat, MatType, MatReuse, Mat *);
4263: #endif
4264: #if defined(PETSC_HAVE_HYPRE)
4265: PETSC_INTERN PetscErrorCode MatConvert_AIJ_HYPRE(Mat A, MatType, MatReuse, Mat *);
4266: #endif
4268: PETSC_EXTERN PetscErrorCode MatConvert_SeqAIJ_SeqSELL(Mat, MatType, MatReuse, Mat *);
4269: PETSC_INTERN PetscErrorCode MatConvert_XAIJ_IS(Mat, MatType, MatReuse, Mat *);
4270: PETSC_INTERN PetscErrorCode MatProductSetFromOptions_IS_XAIJ(Mat);
4272: /*@C
4273: MatSeqAIJGetArray - gives read/write access to the array where the data for a `MATSEQAIJ` matrix is stored
4275: Not Collective
4277: Input Parameter:
4278: . mat - a `MATSEQAIJ` matrix
4280: Output Parameter:
4281: . array - pointer to the data
4283: Level: intermediate
4285: .seealso: `MatSeqAIJRestoreArray()`, `MatSeqAIJGetArrayF90()`
4286: @*/
4287: PetscErrorCode MatSeqAIJGetArray(Mat A, PetscScalar **array)
4288: {
4289: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)A->data;
4291: if (aij->ops->getarray) {
4292: (*aij->ops->getarray)(A, array);
4293: } else {
4294: *array = aij->a;
4295: }
4296: return 0;
4297: }
4299: /*@C
4300: MatSeqAIJRestoreArray - returns access to the array where the data for a `MATSEQAIJ` matrix is stored obtained by `MatSeqAIJGetArray()`
4302: Not Collective
4304: Input Parameters:
4305: + mat - a `MATSEQAIJ` matrix
4306: - array - pointer to the data
4308: Level: intermediate
4310: .seealso: `MatSeqAIJGetArray()`, `MatSeqAIJRestoreArrayF90()`
4311: @*/
4312: PetscErrorCode MatSeqAIJRestoreArray(Mat A, PetscScalar **array)
4313: {
4314: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)A->data;
4316: if (aij->ops->restorearray) {
4317: (*aij->ops->restorearray)(A, array);
4318: } else {
4319: *array = NULL;
4320: }
4321: MatSeqAIJInvalidateDiagonal(A);
4322: PetscObjectStateIncrease((PetscObject)A);
4323: return 0;
4324: }
4326: /*@C
4327: MatSeqAIJGetArrayRead - gives read-only access to the array where the data for a `MATSEQAIJ` matrix is stored
4329: Not Collective
4331: Input Parameter:
4332: . mat - a `MATSEQAIJ` matrix
4334: Output Parameter:
4335: . array - pointer to the data
4337: Level: intermediate
4339: .seealso: `MatSeqAIJGetArray()`, `MatSeqAIJRestoreArrayRead()`
4340: @*/
4341: PetscErrorCode MatSeqAIJGetArrayRead(Mat A, const PetscScalar **array)
4342: {
4343: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)A->data;
4345: if (aij->ops->getarrayread) {
4346: (*aij->ops->getarrayread)(A, array);
4347: } else {
4348: *array = aij->a;
4349: }
4350: return 0;
4351: }
4353: /*@C
4354: MatSeqAIJRestoreArrayRead - restore the read-only access array obtained from `MatSeqAIJGetArrayRead()`
4356: Not Collective
4358: Input Parameter:
4359: . mat - a `MATSEQAIJ` matrix
4361: Output Parameter:
4362: . array - pointer to the data
4364: Level: intermediate
4366: .seealso: `MatSeqAIJGetArray()`, `MatSeqAIJGetArrayRead()`
4367: @*/
4368: PetscErrorCode MatSeqAIJRestoreArrayRead(Mat A, const PetscScalar **array)
4369: {
4370: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)A->data;
4372: if (aij->ops->restorearrayread) {
4373: (*aij->ops->restorearrayread)(A, array);
4374: } else {
4375: *array = NULL;
4376: }
4377: return 0;
4378: }
4380: /*@C
4381: MatSeqAIJGetArrayWrite - gives write-only access to the array where the data for a `MATSEQAIJ` matrix is stored
4383: Not Collective
4385: Input Parameter:
4386: . mat - a `MATSEQAIJ` matrix
4388: Output Parameter:
4389: . array - pointer to the data
4391: Level: intermediate
4393: .seealso: `MatSeqAIJGetArray()`, `MatSeqAIJRestoreArrayRead()`
4394: @*/
4395: PetscErrorCode MatSeqAIJGetArrayWrite(Mat A, PetscScalar **array)
4396: {
4397: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)A->data;
4399: if (aij->ops->getarraywrite) {
4400: (*aij->ops->getarraywrite)(A, array);
4401: } else {
4402: *array = aij->a;
4403: }
4404: MatSeqAIJInvalidateDiagonal(A);
4405: PetscObjectStateIncrease((PetscObject)A);
4406: return 0;
4407: }
4409: /*@C
4410: MatSeqAIJRestoreArrayWrite - restore the read-only access array obtained from MatSeqAIJGetArrayRead
4412: Not Collective
4414: Input Parameter:
4415: . mat - a MATSEQAIJ matrix
4417: Output Parameter:
4418: . array - pointer to the data
4420: Level: intermediate
4422: .seealso: `MatSeqAIJGetArray()`, `MatSeqAIJGetArrayRead()`
4423: @*/
4424: PetscErrorCode MatSeqAIJRestoreArrayWrite(Mat A, PetscScalar **array)
4425: {
4426: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)A->data;
4428: if (aij->ops->restorearraywrite) {
4429: (*aij->ops->restorearraywrite)(A, array);
4430: } else {
4431: *array = NULL;
4432: }
4433: return 0;
4434: }
4436: /*@C
4437: MatSeqAIJGetCSRAndMemType - Get the CSR arrays and the memory type of the `MATSEQAIJ` matrix
4439: Not Collective
4441: Input Parameter:
4442: . mat - a matrix of type `MATSEQAIJ` or its subclasses
4444: Output Parameters:
4445: + i - row map array of the matrix
4446: . j - column index array of the matrix
4447: . a - data array of the matrix
4448: - memtype - memory type of the arrays
4450: Notes:
4451: Any of the output parameters can be NULL, in which case the corresponding value is not returned.
4452: If mat is a device matrix, the arrays are on the device. Otherwise, they are on the host.
4454: One can call this routine on a preallocated but not assembled matrix to just get the memory of the CSR underneath the matrix.
4455: If the matrix is assembled, the data array 'a' is guaranteed to have the latest values of the matrix.
4457: Level: Developer
4459: .seealso: `MatSeqAIJGetArray()`, `MatSeqAIJGetArrayRead()`
4460: @*/
4461: PetscErrorCode MatSeqAIJGetCSRAndMemType(Mat mat, const PetscInt **i, const PetscInt **j, PetscScalar **a, PetscMemType *mtype)
4462: {
4463: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)mat->data;
4466: if (aij->ops->getcsrandmemtype) {
4467: (*aij->ops->getcsrandmemtype)(mat, i, j, a, mtype);
4468: } else {
4469: if (i) *i = aij->i;
4470: if (j) *j = aij->j;
4471: if (a) *a = aij->a;
4472: if (mtype) *mtype = PETSC_MEMTYPE_HOST;
4473: }
4474: return 0;
4475: }
4477: /*@C
4478: MatSeqAIJGetMaxRowNonzeros - returns the maximum number of nonzeros in any row
4480: Not Collective
4482: Input Parameter:
4483: . mat - a `MATSEQAIJ` matrix
4485: Output Parameter:
4486: . nz - the maximum number of nonzeros in any row
4488: Level: intermediate
4490: .seealso: `MatSeqAIJRestoreArray()`, `MatSeqAIJGetArrayF90()`
4491: @*/
4492: PetscErrorCode MatSeqAIJGetMaxRowNonzeros(Mat A, PetscInt *nz)
4493: {
4494: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)A->data;
4496: *nz = aij->rmax;
4497: return 0;
4498: }
4500: PetscErrorCode MatSetPreallocationCOO_SeqAIJ(Mat mat, PetscCount coo_n, PetscInt coo_i[], PetscInt coo_j[])
4501: {
4502: MPI_Comm comm;
4503: PetscInt *i, *j;
4504: PetscInt M, N, row;
4505: PetscCount k, p, q, nneg, nnz, start, end; /* Index the coo array, so use PetscCount as their type */
4506: PetscInt *Ai; /* Change to PetscCount once we use it for row pointers */
4507: PetscInt *Aj;
4508: PetscScalar *Aa;
4509: Mat_SeqAIJ *seqaij = (Mat_SeqAIJ *)(mat->data);
4510: MatType rtype;
4511: PetscCount *perm, *jmap;
4513: MatResetPreallocationCOO_SeqAIJ(mat);
4514: PetscObjectGetComm((PetscObject)mat, &comm);
4515: MatGetSize(mat, &M, &N);
4516: i = coo_i;
4517: j = coo_j;
4518: PetscMalloc1(coo_n, &perm);
4519: for (k = 0; k < coo_n; k++) { /* Ignore entries with negative row or col indices */
4520: if (j[k] < 0) i[k] = -1;
4521: perm[k] = k;
4522: }
4524: /* Sort by row */
4525: PetscSortIntWithIntCountArrayPair(coo_n, i, j, perm);
4526: for (k = 0; k < coo_n; k++) {
4527: if (i[k] >= 0) break;
4528: } /* Advance k to the first row with a non-negative index */
4529: nneg = k;
4530: 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 */
4531: nnz = 0; /* Total number of unique nonzeros to be counted */
4532: jmap++; /* Inc jmap by 1 for convenience */
4534: PetscCalloc1(M + 1, &Ai); /* CSR of A */
4535: PetscMalloc1(coo_n - nneg, &Aj); /* We have at most coo_n-nneg unique nonzeros */
4537: /* In each row, sort by column, then unique column indices to get row length */
4538: Ai++; /* Inc by 1 for convenience */
4539: q = 0; /* q-th unique nonzero, with q starting from 0 */
4540: while (k < coo_n) {
4541: row = i[k];
4542: start = k; /* [start,end) indices for this row */
4543: while (k < coo_n && i[k] == row) k++;
4544: end = k;
4545: PetscSortIntWithCountArray(end - start, j + start, perm + start);
4546: /* Find number of unique col entries in this row */
4547: Aj[q] = j[start]; /* Log the first nonzero in this row */
4548: jmap[q] = 1; /* Number of repeats of this nozero entry */
4549: Ai[row] = 1;
4550: nnz++;
4552: for (p = start + 1; p < end; p++) { /* Scan remaining nonzero in this row */
4553: if (j[p] != j[p - 1]) { /* Meet a new nonzero */
4554: q++;
4555: jmap[q] = 1;
4556: Aj[q] = j[p];
4557: Ai[row]++;
4558: nnz++;
4559: } else {
4560: jmap[q]++;
4561: }
4562: }
4563: q++; /* Move to next row and thus next unique nonzero */
4564: }
4566: Ai--; /* Back to the beginning of Ai[] */
4567: for (k = 0; k < M; k++) Ai[k + 1] += Ai[k];
4568: jmap--; /* Back to the beginning of jmap[] */
4569: jmap[0] = 0;
4570: for (k = 0; k < nnz; k++) jmap[k + 1] += jmap[k];
4571: if (nnz < coo_n - nneg) { /* Realloc with actual number of unique nonzeros */
4572: PetscCount *jmap_new;
4573: PetscInt *Aj_new;
4575: PetscMalloc1(nnz + 1, &jmap_new);
4576: PetscArraycpy(jmap_new, jmap, nnz + 1);
4577: PetscFree(jmap);
4578: jmap = jmap_new;
4580: PetscMalloc1(nnz, &Aj_new);
4581: PetscArraycpy(Aj_new, Aj, nnz);
4582: PetscFree(Aj);
4583: Aj = Aj_new;
4584: }
4586: if (nneg) { /* Discard heading entries with negative indices in perm[], as we'll access it from index 0 in MatSetValuesCOO */
4587: PetscCount *perm_new;
4589: PetscMalloc1(coo_n - nneg, &perm_new);
4590: PetscArraycpy(perm_new, perm + nneg, coo_n - nneg);
4591: PetscFree(perm);
4592: perm = perm_new;
4593: }
4595: MatGetRootType_Private(mat, &rtype);
4596: PetscCalloc1(nnz, &Aa); /* Zero the matrix */
4597: MatSetSeqAIJWithArrays_private(PETSC_COMM_SELF, M, N, Ai, Aj, Aa, rtype, mat);
4599: seqaij->singlemalloc = PETSC_FALSE; /* Ai, Aj and Aa are not allocated in one big malloc */
4600: seqaij->free_a = seqaij->free_ij = PETSC_TRUE; /* Let newmat own Ai, Aj and Aa */
4601: /* Record COO fields */
4602: seqaij->coo_n = coo_n;
4603: seqaij->Atot = coo_n - nneg; /* Annz is seqaij->nz, so no need to record that again */
4604: seqaij->jmap = jmap; /* of length nnz+1 */
4605: seqaij->perm = perm;
4606: return 0;
4607: }
4609: static PetscErrorCode MatSetValuesCOO_SeqAIJ(Mat A, const PetscScalar v[], InsertMode imode)
4610: {
4611: Mat_SeqAIJ *aseq = (Mat_SeqAIJ *)A->data;
4612: PetscCount i, j, Annz = aseq->nz;
4613: PetscCount *perm = aseq->perm, *jmap = aseq->jmap;
4614: PetscScalar *Aa;
4616: MatSeqAIJGetArray(A, &Aa);
4617: for (i = 0; i < Annz; i++) {
4618: PetscScalar sum = 0.0;
4619: for (j = jmap[i]; j < jmap[i + 1]; j++) sum += v[perm[j]];
4620: Aa[i] = (imode == INSERT_VALUES ? 0.0 : Aa[i]) + sum;
4621: }
4622: MatSeqAIJRestoreArray(A, &Aa);
4623: return 0;
4624: }
4626: #if defined(PETSC_HAVE_CUDA)
4627: PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqAIJCUSPARSE(Mat, MatType, MatReuse, Mat *);
4628: #endif
4629: #if defined(PETSC_HAVE_KOKKOS_KERNELS)
4630: PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqAIJKokkos(Mat, MatType, MatReuse, Mat *);
4631: #endif
4633: PETSC_EXTERN PetscErrorCode MatCreate_SeqAIJ(Mat B)
4634: {
4635: Mat_SeqAIJ *b;
4636: PetscMPIInt size;
4638: MPI_Comm_size(PetscObjectComm((PetscObject)B), &size);
4641: PetscNew(&b);
4643: B->data = (void *)b;
4645: PetscMemcpy(B->ops, &MatOps_Values, sizeof(struct _MatOps));
4646: if (B->sortedfull) B->ops->setvalues = MatSetValues_SeqAIJ_SortedFull;
4648: b->row = NULL;
4649: b->col = NULL;
4650: b->icol = NULL;
4651: b->reallocs = 0;
4652: b->ignorezeroentries = PETSC_FALSE;
4653: b->roworiented = PETSC_TRUE;
4654: b->nonew = 0;
4655: b->diag = NULL;
4656: b->solve_work = NULL;
4657: B->spptr = NULL;
4658: b->saved_values = NULL;
4659: b->idiag = NULL;
4660: b->mdiag = NULL;
4661: b->ssor_work = NULL;
4662: b->omega = 1.0;
4663: b->fshift = 0.0;
4664: b->idiagvalid = PETSC_FALSE;
4665: b->ibdiagvalid = PETSC_FALSE;
4666: b->keepnonzeropattern = PETSC_FALSE;
4668: PetscObjectChangeTypeName((PetscObject)B, MATSEQAIJ);
4669: #if defined(PETSC_HAVE_MATLAB)
4670: PetscObjectComposeFunction((PetscObject)B, "PetscMatlabEnginePut_C", MatlabEnginePut_SeqAIJ);
4671: PetscObjectComposeFunction((PetscObject)B, "PetscMatlabEngineGet_C", MatlabEngineGet_SeqAIJ);
4672: #endif
4673: PetscObjectComposeFunction((PetscObject)B, "MatSeqAIJSetColumnIndices_C", MatSeqAIJSetColumnIndices_SeqAIJ);
4674: PetscObjectComposeFunction((PetscObject)B, "MatStoreValues_C", MatStoreValues_SeqAIJ);
4675: PetscObjectComposeFunction((PetscObject)B, "MatRetrieveValues_C", MatRetrieveValues_SeqAIJ);
4676: PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_seqsbaij_C", MatConvert_SeqAIJ_SeqSBAIJ);
4677: PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_seqbaij_C", MatConvert_SeqAIJ_SeqBAIJ);
4678: PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_seqaijperm_C", MatConvert_SeqAIJ_SeqAIJPERM);
4679: PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_seqaijsell_C", MatConvert_SeqAIJ_SeqAIJSELL);
4680: #if defined(PETSC_HAVE_MKL_SPARSE)
4681: PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_seqaijmkl_C", MatConvert_SeqAIJ_SeqAIJMKL);
4682: #endif
4683: #if defined(PETSC_HAVE_CUDA)
4684: PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_seqaijcusparse_C", MatConvert_SeqAIJ_SeqAIJCUSPARSE);
4685: PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_seqaijcusparse_seqaij_C", MatProductSetFromOptions_SeqAIJ);
4686: PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_seqaij_seqaijcusparse_C", MatProductSetFromOptions_SeqAIJ);
4687: #endif
4688: #if defined(PETSC_HAVE_KOKKOS_KERNELS)
4689: PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_seqaijkokkos_C", MatConvert_SeqAIJ_SeqAIJKokkos);
4690: #endif
4691: PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_seqaijcrl_C", MatConvert_SeqAIJ_SeqAIJCRL);
4692: #if defined(PETSC_HAVE_ELEMENTAL)
4693: PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_elemental_C", MatConvert_SeqAIJ_Elemental);
4694: #endif
4695: #if defined(PETSC_HAVE_SCALAPACK)
4696: PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_scalapack_C", MatConvert_AIJ_ScaLAPACK);
4697: #endif
4698: #if defined(PETSC_HAVE_HYPRE)
4699: PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_hypre_C", MatConvert_AIJ_HYPRE);
4700: PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_transpose_seqaij_seqaij_C", MatProductSetFromOptions_Transpose_AIJ_AIJ);
4701: #endif
4702: PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_seqdense_C", MatConvert_SeqAIJ_SeqDense);
4703: PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_seqsell_C", MatConvert_SeqAIJ_SeqSELL);
4704: PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_is_C", MatConvert_XAIJ_IS);
4705: PetscObjectComposeFunction((PetscObject)B, "MatIsTranspose_C", MatIsTranspose_SeqAIJ);
4706: PetscObjectComposeFunction((PetscObject)B, "MatIsHermitianTranspose_C", MatIsTranspose_SeqAIJ);
4707: PetscObjectComposeFunction((PetscObject)B, "MatSeqAIJSetPreallocation_C", MatSeqAIJSetPreallocation_SeqAIJ);
4708: PetscObjectComposeFunction((PetscObject)B, "MatResetPreallocation_C", MatResetPreallocation_SeqAIJ);
4709: PetscObjectComposeFunction((PetscObject)B, "MatSeqAIJSetPreallocationCSR_C", MatSeqAIJSetPreallocationCSR_SeqAIJ);
4710: PetscObjectComposeFunction((PetscObject)B, "MatReorderForNonzeroDiagonal_C", MatReorderForNonzeroDiagonal_SeqAIJ);
4711: PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_is_seqaij_C", MatProductSetFromOptions_IS_XAIJ);
4712: PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_seqdense_seqaij_C", MatProductSetFromOptions_SeqDense_SeqAIJ);
4713: PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_seqaij_seqaij_C", MatProductSetFromOptions_SeqAIJ);
4714: PetscObjectComposeFunction((PetscObject)B, "MatSeqAIJKron_C", MatSeqAIJKron_SeqAIJ);
4715: PetscObjectComposeFunction((PetscObject)B, "MatSetPreallocationCOO_C", MatSetPreallocationCOO_SeqAIJ);
4716: PetscObjectComposeFunction((PetscObject)B, "MatSetValuesCOO_C", MatSetValuesCOO_SeqAIJ);
4717: MatCreate_SeqAIJ_Inode(B);
4718: PetscObjectChangeTypeName((PetscObject)B, MATSEQAIJ);
4719: MatSeqAIJSetTypeFromOptions(B); /* this allows changing the matrix subtype to say MATSEQAIJPERM */
4720: return 0;
4721: }
4723: /*
4724: Given a matrix generated with MatGetFactor() duplicates all the information in A into C
4725: */
4726: PetscErrorCode MatDuplicateNoCreate_SeqAIJ(Mat C, Mat A, MatDuplicateOption cpvalues, PetscBool mallocmatspace)
4727: {
4728: Mat_SeqAIJ *c = (Mat_SeqAIJ *)C->data, *a = (Mat_SeqAIJ *)A->data;
4729: PetscInt m = A->rmap->n, i;
4733: C->factortype = A->factortype;
4734: c->row = NULL;
4735: c->col = NULL;
4736: c->icol = NULL;
4737: c->reallocs = 0;
4739: C->assembled = A->assembled;
4740: C->preallocated = A->preallocated;
4742: if (A->preallocated) {
4743: PetscLayoutReference(A->rmap, &C->rmap);
4744: PetscLayoutReference(A->cmap, &C->cmap);
4746: PetscMalloc1(m, &c->imax);
4747: PetscMemcpy(c->imax, a->imax, m * sizeof(PetscInt));
4748: PetscMalloc1(m, &c->ilen);
4749: PetscMemcpy(c->ilen, a->ilen, m * sizeof(PetscInt));
4751: /* allocate the matrix space */
4752: if (mallocmatspace) {
4753: PetscMalloc3(a->i[m], &c->a, a->i[m], &c->j, m + 1, &c->i);
4755: c->singlemalloc = PETSC_TRUE;
4757: PetscArraycpy(c->i, a->i, m + 1);
4758: if (m > 0) {
4759: PetscArraycpy(c->j, a->j, a->i[m]);
4760: if (cpvalues == MAT_COPY_VALUES) {
4761: const PetscScalar *aa;
4763: MatSeqAIJGetArrayRead(A, &aa);
4764: PetscArraycpy(c->a, aa, a->i[m]);
4765: MatSeqAIJGetArrayRead(A, &aa);
4766: } else {
4767: PetscArrayzero(c->a, a->i[m]);
4768: }
4769: }
4770: }
4772: c->ignorezeroentries = a->ignorezeroentries;
4773: c->roworiented = a->roworiented;
4774: c->nonew = a->nonew;
4775: if (a->diag) {
4776: PetscMalloc1(m + 1, &c->diag);
4777: PetscMemcpy(c->diag, a->diag, m * sizeof(PetscInt));
4778: } else c->diag = NULL;
4780: c->solve_work = NULL;
4781: c->saved_values = NULL;
4782: c->idiag = NULL;
4783: c->ssor_work = NULL;
4784: c->keepnonzeropattern = a->keepnonzeropattern;
4785: c->free_a = PETSC_TRUE;
4786: c->free_ij = PETSC_TRUE;
4788: c->rmax = a->rmax;
4789: c->nz = a->nz;
4790: c->maxnz = a->nz; /* Since we allocate exactly the right amount */
4792: c->compressedrow.use = a->compressedrow.use;
4793: c->compressedrow.nrows = a->compressedrow.nrows;
4794: if (a->compressedrow.use) {
4795: i = a->compressedrow.nrows;
4796: PetscMalloc2(i + 1, &c->compressedrow.i, i, &c->compressedrow.rindex);
4797: PetscArraycpy(c->compressedrow.i, a->compressedrow.i, i + 1);
4798: PetscArraycpy(c->compressedrow.rindex, a->compressedrow.rindex, i);
4799: } else {
4800: c->compressedrow.use = PETSC_FALSE;
4801: c->compressedrow.i = NULL;
4802: c->compressedrow.rindex = NULL;
4803: }
4804: c->nonzerorowcnt = a->nonzerorowcnt;
4805: C->nonzerostate = A->nonzerostate;
4807: MatDuplicate_SeqAIJ_Inode(A, cpvalues, &C);
4808: }
4809: PetscFunctionListDuplicate(((PetscObject)A)->qlist, &((PetscObject)C)->qlist);
4810: return 0;
4811: }
4813: PetscErrorCode MatDuplicate_SeqAIJ(Mat A, MatDuplicateOption cpvalues, Mat *B)
4814: {
4815: MatCreate(PetscObjectComm((PetscObject)A), B);
4816: MatSetSizes(*B, A->rmap->n, A->cmap->n, A->rmap->n, A->cmap->n);
4817: if (!(A->rmap->n % A->rmap->bs) && !(A->cmap->n % A->cmap->bs)) MatSetBlockSizesFromMats(*B, A, A);
4818: MatSetType(*B, ((PetscObject)A)->type_name);
4819: MatDuplicateNoCreate_SeqAIJ(*B, A, cpvalues, PETSC_TRUE);
4820: return 0;
4821: }
4823: PetscErrorCode MatLoad_SeqAIJ(Mat newMat, PetscViewer viewer)
4824: {
4825: PetscBool isbinary, ishdf5;
4829: /* force binary viewer to load .info file if it has not yet done so */
4830: PetscViewerSetUp(viewer);
4831: PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERBINARY, &isbinary);
4832: PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERHDF5, &ishdf5);
4833: if (isbinary) {
4834: MatLoad_SeqAIJ_Binary(newMat, viewer);
4835: } else if (ishdf5) {
4836: #if defined(PETSC_HAVE_HDF5)
4837: MatLoad_AIJ_HDF5(newMat, viewer);
4838: #else
4839: SETERRQ(PetscObjectComm((PetscObject)newMat), PETSC_ERR_SUP, "HDF5 not supported in this build.\nPlease reconfigure using --download-hdf5");
4840: #endif
4841: } else {
4842: 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);
4843: }
4844: return 0;
4845: }
4847: PetscErrorCode MatLoad_SeqAIJ_Binary(Mat mat, PetscViewer viewer)
4848: {
4849: Mat_SeqAIJ *a = (Mat_SeqAIJ *)mat->data;
4850: PetscInt header[4], *rowlens, M, N, nz, sum, rows, cols, i;
4852: PetscViewerSetUp(viewer);
4854: /* read in matrix header */
4855: PetscViewerBinaryRead(viewer, header, 4, NULL, PETSC_INT);
4857: M = header[1];
4858: N = header[2];
4859: nz = header[3];
4864: /* set block sizes from the viewer's .info file */
4865: MatLoad_Binary_BlockSizes(mat, viewer);
4866: /* set local and global sizes if not set already */
4867: if (mat->rmap->n < 0) mat->rmap->n = M;
4868: if (mat->cmap->n < 0) mat->cmap->n = N;
4869: if (mat->rmap->N < 0) mat->rmap->N = M;
4870: if (mat->cmap->N < 0) mat->cmap->N = N;
4871: PetscLayoutSetUp(mat->rmap);
4872: PetscLayoutSetUp(mat->cmap);
4874: /* check if the matrix sizes are correct */
4875: MatGetSize(mat, &rows, &cols);
4878: /* read in row lengths */
4879: PetscMalloc1(M, &rowlens);
4880: PetscViewerBinaryRead(viewer, rowlens, M, NULL, PETSC_INT);
4881: /* check if sum(rowlens) is same as nz */
4882: sum = 0;
4883: for (i = 0; i < M; i++) sum += rowlens[i];
4885: /* preallocate and check sizes */
4886: MatSeqAIJSetPreallocation_SeqAIJ(mat, 0, rowlens);
4887: MatGetSize(mat, &rows, &cols);
4889: /* store row lengths */
4890: PetscArraycpy(a->ilen, rowlens, M);
4891: PetscFree(rowlens);
4893: /* fill in "i" row pointers */
4894: a->i[0] = 0;
4895: for (i = 0; i < M; i++) a->i[i + 1] = a->i[i] + a->ilen[i];
4896: /* read in "j" column indices */
4897: PetscViewerBinaryRead(viewer, a->j, nz, NULL, PETSC_INT);
4898: /* read in "a" nonzero values */
4899: PetscViewerBinaryRead(viewer, a->a, nz, NULL, PETSC_SCALAR);
4901: MatAssemblyBegin(mat, MAT_FINAL_ASSEMBLY);
4902: MatAssemblyEnd(mat, MAT_FINAL_ASSEMBLY);
4903: return 0;
4904: }
4906: PetscErrorCode MatEqual_SeqAIJ(Mat A, Mat B, PetscBool *flg)
4907: {
4908: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data, *b = (Mat_SeqAIJ *)B->data;
4909: const PetscScalar *aa, *ba;
4910: #if defined(PETSC_USE_COMPLEX)
4911: PetscInt k;
4912: #endif
4914: /* If the matrix dimensions are not equal,or no of nonzeros */
4915: if ((A->rmap->n != B->rmap->n) || (A->cmap->n != B->cmap->n) || (a->nz != b->nz)) {
4916: *flg = PETSC_FALSE;
4917: return 0;
4918: }
4920: /* if the a->i are the same */
4921: PetscArraycmp(a->i, b->i, A->rmap->n + 1, flg);
4922: if (!*flg) return 0;
4924: /* if a->j are the same */
4925: PetscArraycmp(a->j, b->j, a->nz, flg);
4926: if (!*flg) return 0;
4928: MatSeqAIJGetArrayRead(A, &aa);
4929: MatSeqAIJGetArrayRead(B, &ba);
4930: /* if a->a are the same */
4931: #if defined(PETSC_USE_COMPLEX)
4932: for (k = 0; k < a->nz; k++) {
4933: if (PetscRealPart(aa[k]) != PetscRealPart(ba[k]) || PetscImaginaryPart(aa[k]) != PetscImaginaryPart(ba[k])) {
4934: *flg = PETSC_FALSE;
4935: return 0;
4936: }
4937: }
4938: #else
4939: PetscArraycmp(aa, ba, a->nz, flg);
4940: #endif
4941: MatSeqAIJRestoreArrayRead(A, &aa);
4942: MatSeqAIJRestoreArrayRead(B, &ba);
4943: return 0;
4944: }
4946: /*@
4947: MatCreateSeqAIJWithArrays - Creates an sequential `MATSEQAIJ` matrix using matrix elements (in CSR format)
4948: provided by the user.
4950: Collective
4952: Input Parameters:
4953: + comm - must be an MPI communicator of size 1
4954: . m - number of rows
4955: . n - number of columns
4956: . i - row indices; that is i[0] = 0, i[row] = i[row-1] + number of elements in that row of the matrix
4957: . j - column indices
4958: - a - matrix values
4960: Output Parameter:
4961: . mat - the matrix
4963: Level: intermediate
4965: Notes:
4966: The i, j, and a arrays are not copied by this routine, the user must free these arrays
4967: once the matrix is destroyed and not before
4969: You cannot set new nonzero locations into this matrix, that will generate an error.
4971: The i and j indices are 0 based
4973: The format which is used for the sparse matrix input, is equivalent to a
4974: row-major ordering.. i.e for the following matrix, the input data expected is
4975: as shown
4977: $ 1 0 0
4978: $ 2 0 3
4979: $ 4 5 6
4980: $
4981: $ i = {0,1,3,6} [size = nrow+1 = 3+1]
4982: $ j = {0,0,2,0,1,2} [size = 6]; values must be sorted for each row
4983: $ v = {1,2,3,4,5,6} [size = 6]
4985: .seealso: `MatCreate()`, `MatCreateAIJ()`, `MatCreateSeqAIJ()`, `MatCreateMPIAIJWithArrays()`, `MatMPIAIJSetPreallocationCSR()`
4986: @*/
4987: PetscErrorCode MatCreateSeqAIJWithArrays(MPI_Comm comm, PetscInt m, PetscInt n, PetscInt i[], PetscInt j[], PetscScalar a[], Mat *mat)
4988: {
4989: PetscInt ii;
4990: Mat_SeqAIJ *aij;
4991: PetscInt jj;
4994: MatCreate(comm, mat);
4995: MatSetSizes(*mat, m, n, m, n);
4996: /* MatSetBlockSizes(*mat,,); */
4997: MatSetType(*mat, MATSEQAIJ);
4998: MatSeqAIJSetPreallocation_SeqAIJ(*mat, MAT_SKIP_ALLOCATION, NULL);
4999: aij = (Mat_SeqAIJ *)(*mat)->data;
5000: PetscMalloc1(m, &aij->imax);
5001: PetscMalloc1(m, &aij->ilen);
5003: aij->i = i;
5004: aij->j = j;
5005: aij->a = a;
5006: aij->singlemalloc = PETSC_FALSE;
5007: aij->nonew = -1; /*this indicates that inserting a new value in the matrix that generates a new nonzero is an error*/
5008: aij->free_a = PETSC_FALSE;
5009: aij->free_ij = PETSC_FALSE;
5011: for (ii = 0, aij->nonzerorowcnt = 0, aij->rmax = 0; ii < m; ii++) {
5012: aij->ilen[ii] = aij->imax[ii] = i[ii + 1] - i[ii];
5013: if (PetscDefined(USE_DEBUG)) {
5015: for (jj = i[ii] + 1; jj < i[ii + 1]; jj++) {
5018: }
5019: }
5020: }
5021: if (PetscDefined(USE_DEBUG)) {
5022: for (ii = 0; ii < aij->i[m]; ii++) {
5025: }
5026: }
5028: MatAssemblyBegin(*mat, MAT_FINAL_ASSEMBLY);
5029: MatAssemblyEnd(*mat, MAT_FINAL_ASSEMBLY);
5030: return 0;
5031: }
5033: /*@
5034: MatCreateSeqAIJFromTriple - Creates an sequential `MATSEQAIJ` matrix using matrix elements (in COO format)
5035: provided by the user.
5037: Collective
5039: Input Parameters:
5040: + comm - must be an MPI communicator of size 1
5041: . m - number of rows
5042: . n - number of columns
5043: . i - row indices
5044: . j - column indices
5045: . a - matrix values
5046: . nz - number of nonzeros
5047: - idx - if the i and j indices start with 1 use `PETSC_TRUE` otherwise use `PETSC_FALSE`
5049: Output Parameter:
5050: . mat - the matrix
5052: Level: intermediate
5054: Example:
5055: For the following matrix, the input data expected is as shown (using 0 based indexing)
5056: .vb
5057: 1 0 0
5058: 2 0 3
5059: 4 5 6
5061: i = {0,1,1,2,2,2}
5062: j = {0,0,2,0,1,2}
5063: v = {1,2,3,4,5,6}
5064: .ve
5065: Notes:
5066: Instead of using this function, users should also consider `MatSetPreallocationCOO()` and `MatSetValuesCOO()`, which allow repeated or remote entries,
5067: and are particularly useful in iterative applications.
5069: .seealso: `MatCreate()`, `MatCreateAIJ()`, `MatCreateSeqAIJ()`, `MatCreateSeqAIJWithArrays()`, `MatMPIAIJSetPreallocationCSR()`, `MatSetValuesCOO()`, `MatSetPreallocationCOO()`
5070: @*/
5071: PetscErrorCode MatCreateSeqAIJFromTriple(MPI_Comm comm, PetscInt m, PetscInt n, PetscInt i[], PetscInt j[], PetscScalar a[], Mat *mat, PetscInt nz, PetscBool idx)
5072: {
5073: PetscInt ii, *nnz, one = 1, row, col;
5075: PetscCalloc1(m, &nnz);
5076: for (ii = 0; ii < nz; ii++) nnz[i[ii] - !!idx] += 1;
5077: MatCreate(comm, mat);
5078: MatSetSizes(*mat, m, n, m, n);
5079: MatSetType(*mat, MATSEQAIJ);
5080: MatSeqAIJSetPreallocation_SeqAIJ(*mat, 0, nnz);
5081: for (ii = 0; ii < nz; ii++) {
5082: if (idx) {
5083: row = i[ii] - 1;
5084: col = j[ii] - 1;
5085: } else {
5086: row = i[ii];
5087: col = j[ii];
5088: }
5089: MatSetValues(*mat, one, &row, one, &col, &a[ii], ADD_VALUES);
5090: }
5091: MatAssemblyBegin(*mat, MAT_FINAL_ASSEMBLY);
5092: MatAssemblyEnd(*mat, MAT_FINAL_ASSEMBLY);
5093: PetscFree(nnz);
5094: return 0;
5095: }
5097: PetscErrorCode MatSeqAIJInvalidateDiagonal(Mat A)
5098: {
5099: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
5101: a->idiagvalid = PETSC_FALSE;
5102: a->ibdiagvalid = PETSC_FALSE;
5104: MatSeqAIJInvalidateDiagonal_Inode(A);
5105: return 0;
5106: }
5108: PetscErrorCode MatCreateMPIMatConcatenateSeqMat_SeqAIJ(MPI_Comm comm, Mat inmat, PetscInt n, MatReuse scall, Mat *outmat)
5109: {
5110: MatCreateMPIMatConcatenateSeqMat_MPIAIJ(comm, inmat, n, scall, outmat);
5111: return 0;
5112: }
5114: /*
5115: Permute A into C's *local* index space using rowemb,colemb.
5116: The embedding are supposed to be injections and the above implies that the range of rowemb is a subset
5117: of [0,m), colemb is in [0,n).
5118: If pattern == DIFFERENT_NONZERO_PATTERN, C is preallocated according to A.
5119: */
5120: PetscErrorCode MatSetSeqMat_SeqAIJ(Mat C, IS rowemb, IS colemb, MatStructure pattern, Mat B)
5121: {
5122: /* If making this function public, change the error returned in this function away from _PLIB. */
5123: Mat_SeqAIJ *Baij;
5124: PetscBool seqaij;
5125: PetscInt m, n, *nz, i, j, count;
5126: PetscScalar v;
5127: const PetscInt *rowindices, *colindices;
5129: if (!B) return 0;
5130: /* Check to make sure the target matrix (and embeddings) are compatible with C and each other. */
5131: PetscObjectBaseTypeCompare((PetscObject)B, MATSEQAIJ, &seqaij);
5133: if (rowemb) {
5134: ISGetLocalSize(rowemb, &m);
5136: } else {
5138: }
5139: if (colemb) {
5140: ISGetLocalSize(colemb, &n);
5142: } else {
5144: }
5146: Baij = (Mat_SeqAIJ *)(B->data);
5147: if (pattern == DIFFERENT_NONZERO_PATTERN) {
5148: PetscMalloc1(B->rmap->n, &nz);
5149: for (i = 0; i < B->rmap->n; i++) nz[i] = Baij->i[i + 1] - Baij->i[i];
5150: MatSeqAIJSetPreallocation(C, 0, nz);
5151: PetscFree(nz);
5152: }
5153: if (pattern == SUBSET_NONZERO_PATTERN) MatZeroEntries(C);
5154: count = 0;
5155: rowindices = NULL;
5156: colindices = NULL;
5157: if (rowemb) ISGetIndices(rowemb, &rowindices);
5158: if (colemb) ISGetIndices(colemb, &colindices);
5159: for (i = 0; i < B->rmap->n; i++) {
5160: PetscInt row;
5161: row = i;
5162: if (rowindices) row = rowindices[i];
5163: for (j = Baij->i[i]; j < Baij->i[i + 1]; j++) {
5164: PetscInt col;
5165: col = Baij->j[count];
5166: if (colindices) col = colindices[col];
5167: v = Baij->a[count];
5168: MatSetValues(C, 1, &row, 1, &col, &v, INSERT_VALUES);
5169: ++count;
5170: }
5171: }
5172: /* FIXME: set C's nonzerostate correctly. */
5173: /* Assembly for C is necessary. */
5174: C->preallocated = PETSC_TRUE;
5175: C->assembled = PETSC_TRUE;
5176: C->was_assembled = PETSC_FALSE;
5177: return 0;
5178: }
5180: PetscFunctionList MatSeqAIJList = NULL;
5182: /*@C
5183: MatSeqAIJSetType - Converts a `MATSEQAIJ` matrix to a subtype
5185: Collective
5187: Input Parameters:
5188: + mat - the matrix object
5189: - matype - matrix type
5191: Options Database Key:
5192: . -mat_seqaij_type <method> - for example seqaijcrl
5194: Level: intermediate
5196: .seealso: `PCSetType()`, `VecSetType()`, `MatCreate()`, `MatType`, `Mat`
5197: @*/
5198: PetscErrorCode MatSeqAIJSetType(Mat mat, MatType matype)
5199: {
5200: PetscBool sametype;
5201: PetscErrorCode (*r)(Mat, MatType, MatReuse, Mat *);
5204: PetscObjectTypeCompare((PetscObject)mat, matype, &sametype);
5205: if (sametype) return 0;
5207: PetscFunctionListFind(MatSeqAIJList, matype, &r);
5209: (*r)(mat, matype, MAT_INPLACE_MATRIX, &mat);
5210: return 0;
5211: }
5213: /*@C
5214: MatSeqAIJRegister - - Adds a new sub-matrix type for sequential `MATSEQAIJ` matrices
5216: Not Collective
5218: Input Parameters:
5219: + name - name of a new user-defined matrix type, for example `MATSEQAIJCRL`
5220: - function - routine to convert to subtype
5222: Notes:
5223: `MatSeqAIJRegister()` may be called multiple times to add several user-defined solvers.
5225: Then, your matrix can be chosen with the procedural interface at runtime via the option
5226: $ -mat_seqaij_type my_mat
5228: Level: advanced
5230: .seealso: `MatSeqAIJRegisterAll()`
5231: @*/
5232: PetscErrorCode MatSeqAIJRegister(const char sname[], PetscErrorCode (*function)(Mat, MatType, MatReuse, Mat *))
5233: {
5234: MatInitializePackage();
5235: PetscFunctionListAdd(&MatSeqAIJList, sname, function);
5236: return 0;
5237: }
5239: PetscBool MatSeqAIJRegisterAllCalled = PETSC_FALSE;
5241: /*@C
5242: MatSeqAIJRegisterAll - Registers all of the matrix subtypes of `MATSSEQAIJ`
5244: Not Collective
5246: Level: advanced
5248: .seealso: `MatRegisterAll()`, `MatSeqAIJRegister()`
5249: @*/
5250: PetscErrorCode MatSeqAIJRegisterAll(void)
5251: {
5252: if (MatSeqAIJRegisterAllCalled) return 0;
5253: MatSeqAIJRegisterAllCalled = PETSC_TRUE;
5255: MatSeqAIJRegister(MATSEQAIJCRL, MatConvert_SeqAIJ_SeqAIJCRL);
5256: MatSeqAIJRegister(MATSEQAIJPERM, MatConvert_SeqAIJ_SeqAIJPERM);
5257: MatSeqAIJRegister(MATSEQAIJSELL, MatConvert_SeqAIJ_SeqAIJSELL);
5258: #if defined(PETSC_HAVE_MKL_SPARSE)
5259: MatSeqAIJRegister(MATSEQAIJMKL, MatConvert_SeqAIJ_SeqAIJMKL);
5260: #endif
5261: #if defined(PETSC_HAVE_CUDA)
5262: MatSeqAIJRegister(MATSEQAIJCUSPARSE, MatConvert_SeqAIJ_SeqAIJCUSPARSE);
5263: #endif
5264: #if defined(PETSC_HAVE_KOKKOS_KERNELS)
5265: MatSeqAIJRegister(MATSEQAIJKOKKOS, MatConvert_SeqAIJ_SeqAIJKokkos);
5266: #endif
5267: #if defined(PETSC_HAVE_VIENNACL) && defined(PETSC_HAVE_VIENNACL_NO_CUDA)
5268: MatSeqAIJRegister(MATMPIAIJVIENNACL, MatConvert_SeqAIJ_SeqAIJViennaCL);
5269: #endif
5270: return 0;
5271: }
5273: /*
5274: Special version for direct calls from Fortran
5275: */
5276: #include <petsc/private/fortranimpl.h>
5277: #if defined(PETSC_HAVE_FORTRAN_CAPS)
5278: #define matsetvaluesseqaij_ MATSETVALUESSEQAIJ
5279: #elif !defined(PETSC_HAVE_FORTRAN_UNDERSCORE)
5280: #define matsetvaluesseqaij_ matsetvaluesseqaij
5281: #endif
5283: /* Change these macros so can be used in void function */
5285: /* Change these macros so can be used in void function */
5286: /* Identical to PetscCallVoid, except it assigns to *_ierr */
5287: #undef PetscCall
5288: #define PetscCall(...) \
5289: do { \
5290: PetscErrorCode ierr_msv_mpiaij = __VA_ARGS__; \
5291: if (PetscUnlikely(ierr_msv_mpiaij)) { \
5292: *_PetscError(PETSC_COMM_SELF, __LINE__, PETSC_FUNCTION_NAME, __FILE__, ierr_msv_mpiaij, PETSC_ERROR_REPEAT, " "); \
5293: return; \
5294: } \
5295: } while (0)
5297: #undef SETERRQ
5298: #define SETERRQ(comm, ierr, ...) \
5299: do { \
5300: *_PetscError(comm, __LINE__, PETSC_FUNCTION_NAME, __FILE__, ierr, PETSC_ERROR_INITIAL, __VA_ARGS__); \
5301: return; \
5302: } while (0)
5304: PETSC_EXTERN void matsetvaluesseqaij_(Mat *AA, PetscInt *mm, const PetscInt im[], PetscInt *nn, const PetscInt in[], const PetscScalar v[], InsertMode *isis, PetscErrorCode *_ierr)
5305: {
5306: Mat A = *AA;
5307: PetscInt m = *mm, n = *nn;
5308: InsertMode is = *isis;
5309: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
5310: PetscInt *rp, k, low, high, t, ii, row, nrow, i, col, l, rmax, N;
5311: PetscInt *imax, *ai, *ailen;
5312: PetscInt *aj, nonew = a->nonew, lastcol = -1;
5313: MatScalar *ap, value, *aa;
5314: PetscBool ignorezeroentries = a->ignorezeroentries;
5315: PetscBool roworiented = a->roworiented;
5317: MatCheckPreallocated(A, 1);
5318: imax = a->imax;
5319: ai = a->i;
5320: ailen = a->ilen;
5321: aj = a->j;
5322: aa = a->a;
5324: for (k = 0; k < m; k++) { /* loop over added rows */
5325: row = im[k];
5326: if (row < 0) continue;
5328: rp = aj + ai[row];
5329: ap = aa + ai[row];
5330: rmax = imax[row];
5331: nrow = ailen[row];
5332: low = 0;
5333: high = nrow;
5334: for (l = 0; l < n; l++) { /* loop over added columns */
5335: if (in[l] < 0) continue;
5337: col = in[l];
5338: if (roworiented) value = v[l + k * n];
5339: else value = v[k + l * m];
5341: if (value == 0.0 && ignorezeroentries && (is == ADD_VALUES)) continue;
5343: if (col <= lastcol) low = 0;
5344: else high = nrow;
5345: lastcol = col;
5346: while (high - low > 5) {
5347: t = (low + high) / 2;
5348: if (rp[t] > col) high = t;
5349: else low = t;
5350: }
5351: for (i = low; i < high; i++) {
5352: if (rp[i] > col) break;
5353: if (rp[i] == col) {
5354: if (is == ADD_VALUES) ap[i] += value;
5355: else ap[i] = value;
5356: goto noinsert;
5357: }
5358: }
5359: if (value == 0.0 && ignorezeroentries) goto noinsert;
5360: if (nonew == 1) goto noinsert;
5362: MatSeqXAIJReallocateAIJ(A, A->rmap->n, 1, nrow, row, col, rmax, aa, ai, aj, rp, ap, imax, nonew, MatScalar);
5363: N = nrow++ - 1;
5364: a->nz++;
5365: high++;
5366: /* shift up all the later entries in this row */
5367: for (ii = N; ii >= i; ii--) {
5368: rp[ii + 1] = rp[ii];
5369: ap[ii + 1] = ap[ii];
5370: }
5371: rp[i] = col;
5372: ap[i] = value;
5373: A->nonzerostate++;
5374: noinsert:;
5375: low = i + 1;
5376: }
5377: ailen[row] = nrow;
5378: }
5379: return;
5380: }
5381: /* Undefining these here since they were redefined from their original definition above! No
5382: * other PETSc functions should be defined past this point, as it is impossible to recover the
5383: * original definitions */
5384: #undef PetscCall
5385: #undef SETERRQ