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: /* defines MatSetValues_Seq_Hash(), MatAssemblyEnd_Seq_Hash(), MatSetUp_Seq_Hash() */
12: #define TYPE AIJ
13: #define TYPE_BS
14: #include "../src/mat/impls/aij/seq/seqhashmatsetvalues.h"
15: #include "../src/mat/impls/aij/seq/seqhashmat.h"
16: #undef TYPE
17: #undef TYPE_BS
19: static PetscErrorCode MatSeqAIJSetTypeFromOptions(Mat A)
20: {
21: PetscBool flg;
22: char type[256];
24: PetscFunctionBegin;
25: PetscObjectOptionsBegin((PetscObject)A);
26: PetscCall(PetscOptionsFList("-mat_seqaij_type", "Matrix SeqAIJ type", "MatSeqAIJSetType", MatSeqAIJList, "seqaij", type, 256, &flg));
27: if (flg) PetscCall(MatSeqAIJSetType(A, type));
28: PetscOptionsEnd();
29: PetscFunctionReturn(PETSC_SUCCESS);
30: }
32: static PetscErrorCode MatGetColumnReductions_SeqAIJ(Mat A, PetscInt type, PetscReal *reductions)
33: {
34: PetscInt i, m, n;
35: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)A->data;
37: PetscFunctionBegin;
38: PetscCall(MatGetSize(A, &m, &n));
39: PetscCall(PetscArrayzero(reductions, n));
40: if (type == NORM_2) {
41: for (i = 0; i < aij->i[m]; i++) reductions[aij->j[i]] += PetscAbsScalar(aij->a[i] * aij->a[i]);
42: } else if (type == NORM_1) {
43: for (i = 0; i < aij->i[m]; i++) reductions[aij->j[i]] += PetscAbsScalar(aij->a[i]);
44: } else if (type == NORM_INFINITY) {
45: for (i = 0; i < aij->i[m]; i++) reductions[aij->j[i]] = PetscMax(PetscAbsScalar(aij->a[i]), reductions[aij->j[i]]);
46: } else if (type == REDUCTION_SUM_REALPART || type == REDUCTION_MEAN_REALPART) {
47: for (i = 0; i < aij->i[m]; i++) reductions[aij->j[i]] += PetscRealPart(aij->a[i]);
48: } else if (type == REDUCTION_SUM_IMAGINARYPART || type == REDUCTION_MEAN_IMAGINARYPART) {
49: for (i = 0; i < aij->i[m]; i++) reductions[aij->j[i]] += PetscImaginaryPart(aij->a[i]);
50: } else SETERRQ(PetscObjectComm((PetscObject)A), PETSC_ERR_ARG_WRONG, "Unknown reduction type");
52: if (type == NORM_2) {
53: for (i = 0; i < n; i++) reductions[i] = PetscSqrtReal(reductions[i]);
54: } else if (type == REDUCTION_MEAN_REALPART || type == REDUCTION_MEAN_IMAGINARYPART) {
55: for (i = 0; i < n; i++) reductions[i] /= m;
56: }
57: PetscFunctionReturn(PETSC_SUCCESS);
58: }
60: static PetscErrorCode MatFindOffBlockDiagonalEntries_SeqAIJ(Mat A, IS *is)
61: {
62: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
63: PetscInt i, m = A->rmap->n, cnt = 0, bs = A->rmap->bs;
64: const PetscInt *jj = a->j, *ii = a->i;
65: PetscInt *rows;
67: PetscFunctionBegin;
68: for (i = 0; i < m; i++) {
69: if ((ii[i] != ii[i + 1]) && ((jj[ii[i]] < bs * (i / bs)) || (jj[ii[i + 1] - 1] > bs * ((i + bs) / bs) - 1))) cnt++;
70: }
71: PetscCall(PetscMalloc1(cnt, &rows));
72: cnt = 0;
73: for (i = 0; i < m; i++) {
74: if ((ii[i] != ii[i + 1]) && ((jj[ii[i]] < bs * (i / bs)) || (jj[ii[i + 1] - 1] > bs * ((i + bs) / bs) - 1))) {
75: rows[cnt] = i;
76: cnt++;
77: }
78: }
79: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, cnt, rows, PETSC_OWN_POINTER, is));
80: PetscFunctionReturn(PETSC_SUCCESS);
81: }
83: PetscErrorCode MatFindZeroDiagonals_SeqAIJ_Private(Mat A, PetscInt *nrows, PetscInt **zrows)
84: {
85: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
86: const MatScalar *aa;
87: PetscInt i, m = A->rmap->n, cnt = 0;
88: const PetscInt *ii = a->i, *jj = a->j, *diag;
89: PetscInt *rows;
91: PetscFunctionBegin;
92: PetscCall(MatSeqAIJGetArrayRead(A, &aa));
93: PetscCall(MatMarkDiagonal_SeqAIJ(A));
94: diag = a->diag;
95: for (i = 0; i < m; i++) {
96: if ((diag[i] >= ii[i + 1]) || (jj[diag[i]] != i) || (aa[diag[i]] == 0.0)) cnt++;
97: }
98: PetscCall(PetscMalloc1(cnt, &rows));
99: cnt = 0;
100: for (i = 0; i < m; i++) {
101: if ((diag[i] >= ii[i + 1]) || (jj[diag[i]] != i) || (aa[diag[i]] == 0.0)) rows[cnt++] = i;
102: }
103: *nrows = cnt;
104: *zrows = rows;
105: PetscCall(MatSeqAIJRestoreArrayRead(A, &aa));
106: PetscFunctionReturn(PETSC_SUCCESS);
107: }
109: static PetscErrorCode MatFindZeroDiagonals_SeqAIJ(Mat A, IS *zrows)
110: {
111: PetscInt nrows, *rows;
113: PetscFunctionBegin;
114: *zrows = NULL;
115: PetscCall(MatFindZeroDiagonals_SeqAIJ_Private(A, &nrows, &rows));
116: PetscCall(ISCreateGeneral(PetscObjectComm((PetscObject)A), nrows, rows, PETSC_OWN_POINTER, zrows));
117: PetscFunctionReturn(PETSC_SUCCESS);
118: }
120: static PetscErrorCode MatFindNonzeroRows_SeqAIJ(Mat A, IS *keptrows)
121: {
122: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
123: const MatScalar *aa;
124: PetscInt m = A->rmap->n, cnt = 0;
125: const PetscInt *ii;
126: PetscInt n, i, j, *rows;
128: PetscFunctionBegin;
129: PetscCall(MatSeqAIJGetArrayRead(A, &aa));
130: *keptrows = NULL;
131: ii = a->i;
132: for (i = 0; i < m; i++) {
133: n = ii[i + 1] - ii[i];
134: if (!n) {
135: cnt++;
136: goto ok1;
137: }
138: for (j = ii[i]; j < ii[i + 1]; j++) {
139: if (aa[j] != 0.0) goto ok1;
140: }
141: cnt++;
142: ok1:;
143: }
144: if (!cnt) {
145: PetscCall(MatSeqAIJRestoreArrayRead(A, &aa));
146: PetscFunctionReturn(PETSC_SUCCESS);
147: }
148: PetscCall(PetscMalloc1(A->rmap->n - cnt, &rows));
149: cnt = 0;
150: for (i = 0; i < m; i++) {
151: n = ii[i + 1] - ii[i];
152: if (!n) continue;
153: for (j = ii[i]; j < ii[i + 1]; j++) {
154: if (aa[j] != 0.0) {
155: rows[cnt++] = i;
156: break;
157: }
158: }
159: }
160: PetscCall(MatSeqAIJRestoreArrayRead(A, &aa));
161: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, cnt, rows, PETSC_OWN_POINTER, keptrows));
162: PetscFunctionReturn(PETSC_SUCCESS);
163: }
165: PetscErrorCode MatDiagonalSet_SeqAIJ(Mat Y, Vec D, InsertMode is)
166: {
167: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)Y->data;
168: PetscInt i, m = Y->rmap->n;
169: const PetscInt *diag;
170: MatScalar *aa;
171: const PetscScalar *v;
172: PetscBool missing;
174: PetscFunctionBegin;
175: if (Y->assembled) {
176: PetscCall(MatMissingDiagonal_SeqAIJ(Y, &missing, NULL));
177: if (!missing) {
178: diag = aij->diag;
179: PetscCall(VecGetArrayRead(D, &v));
180: PetscCall(MatSeqAIJGetArray(Y, &aa));
181: if (is == INSERT_VALUES) {
182: for (i = 0; i < m; i++) aa[diag[i]] = v[i];
183: } else {
184: for (i = 0; i < m; i++) aa[diag[i]] += v[i];
185: }
186: PetscCall(MatSeqAIJRestoreArray(Y, &aa));
187: PetscCall(VecRestoreArrayRead(D, &v));
188: PetscFunctionReturn(PETSC_SUCCESS);
189: }
190: PetscCall(MatSeqAIJInvalidateDiagonal(Y));
191: }
192: PetscCall(MatDiagonalSet_Default(Y, D, is));
193: PetscFunctionReturn(PETSC_SUCCESS);
194: }
196: PetscErrorCode MatGetRowIJ_SeqAIJ(Mat A, PetscInt oshift, PetscBool symmetric, PetscBool inodecompressed, PetscInt *m, const PetscInt *ia[], const PetscInt *ja[], PetscBool *done)
197: {
198: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
199: PetscInt i, ishift;
201: PetscFunctionBegin;
202: if (m) *m = A->rmap->n;
203: if (!ia) PetscFunctionReturn(PETSC_SUCCESS);
204: ishift = 0;
205: if (symmetric && A->structurally_symmetric != PETSC_BOOL3_TRUE) {
206: PetscCall(MatToSymmetricIJ_SeqAIJ(A->rmap->n, a->i, a->j, PETSC_TRUE, ishift, oshift, (PetscInt **)ia, (PetscInt **)ja));
207: } else if (oshift == 1) {
208: PetscInt *tia;
209: PetscInt nz = a->i[A->rmap->n];
210: /* malloc space and add 1 to i and j indices */
211: PetscCall(PetscMalloc1(A->rmap->n + 1, &tia));
212: for (i = 0; i < A->rmap->n + 1; i++) tia[i] = a->i[i] + 1;
213: *ia = tia;
214: if (ja) {
215: PetscInt *tja;
216: PetscCall(PetscMalloc1(nz + 1, &tja));
217: for (i = 0; i < nz; i++) tja[i] = a->j[i] + 1;
218: *ja = tja;
219: }
220: } else {
221: *ia = a->i;
222: if (ja) *ja = a->j;
223: }
224: PetscFunctionReturn(PETSC_SUCCESS);
225: }
227: PetscErrorCode MatRestoreRowIJ_SeqAIJ(Mat A, PetscInt oshift, PetscBool symmetric, PetscBool inodecompressed, PetscInt *n, const PetscInt *ia[], const PetscInt *ja[], PetscBool *done)
228: {
229: PetscFunctionBegin;
230: if (!ia) PetscFunctionReturn(PETSC_SUCCESS);
231: if ((symmetric && A->structurally_symmetric != PETSC_BOOL3_TRUE) || oshift == 1) {
232: PetscCall(PetscFree(*ia));
233: if (ja) PetscCall(PetscFree(*ja));
234: }
235: PetscFunctionReturn(PETSC_SUCCESS);
236: }
238: PetscErrorCode MatGetColumnIJ_SeqAIJ(Mat A, PetscInt oshift, PetscBool symmetric, PetscBool inodecompressed, PetscInt *nn, const PetscInt *ia[], const PetscInt *ja[], PetscBool *done)
239: {
240: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
241: PetscInt i, *collengths, *cia, *cja, n = A->cmap->n, m = A->rmap->n;
242: PetscInt nz = a->i[m], row, *jj, mr, col;
244: PetscFunctionBegin;
245: *nn = n;
246: if (!ia) PetscFunctionReturn(PETSC_SUCCESS);
247: if (symmetric) {
248: PetscCall(MatToSymmetricIJ_SeqAIJ(A->rmap->n, a->i, a->j, PETSC_TRUE, 0, oshift, (PetscInt **)ia, (PetscInt **)ja));
249: } else {
250: PetscCall(PetscCalloc1(n, &collengths));
251: PetscCall(PetscMalloc1(n + 1, &cia));
252: PetscCall(PetscMalloc1(nz, &cja));
253: jj = a->j;
254: for (i = 0; i < nz; i++) collengths[jj[i]]++;
255: cia[0] = oshift;
256: for (i = 0; i < n; i++) cia[i + 1] = cia[i] + collengths[i];
257: PetscCall(PetscArrayzero(collengths, n));
258: jj = a->j;
259: for (row = 0; row < m; row++) {
260: mr = a->i[row + 1] - a->i[row];
261: for (i = 0; i < mr; i++) {
262: col = *jj++;
264: cja[cia[col] + collengths[col]++ - oshift] = row + oshift;
265: }
266: }
267: PetscCall(PetscFree(collengths));
268: *ia = cia;
269: *ja = cja;
270: }
271: PetscFunctionReturn(PETSC_SUCCESS);
272: }
274: PetscErrorCode MatRestoreColumnIJ_SeqAIJ(Mat A, PetscInt oshift, PetscBool symmetric, PetscBool inodecompressed, PetscInt *n, const PetscInt *ia[], const PetscInt *ja[], PetscBool *done)
275: {
276: PetscFunctionBegin;
277: if (!ia) PetscFunctionReturn(PETSC_SUCCESS);
279: PetscCall(PetscFree(*ia));
280: PetscCall(PetscFree(*ja));
281: PetscFunctionReturn(PETSC_SUCCESS);
282: }
284: /*
285: MatGetColumnIJ_SeqAIJ_Color() and MatRestoreColumnIJ_SeqAIJ_Color() are customized from
286: MatGetColumnIJ_SeqAIJ() and MatRestoreColumnIJ_SeqAIJ() by adding an output
287: spidx[], index of a->a, to be used in MatTransposeColoringCreate_SeqAIJ() and MatFDColoringCreate_SeqXAIJ()
288: */
289: PetscErrorCode MatGetColumnIJ_SeqAIJ_Color(Mat A, PetscInt oshift, PetscBool symmetric, PetscBool inodecompressed, PetscInt *nn, const PetscInt *ia[], const PetscInt *ja[], PetscInt *spidx[], PetscBool *done)
290: {
291: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
292: PetscInt i, *collengths, *cia, *cja, n = A->cmap->n, m = A->rmap->n;
293: PetscInt nz = a->i[m], row, mr, col, tmp;
294: PetscInt *cspidx;
295: const PetscInt *jj;
297: PetscFunctionBegin;
298: *nn = n;
299: if (!ia) PetscFunctionReturn(PETSC_SUCCESS);
301: PetscCall(PetscCalloc1(n, &collengths));
302: PetscCall(PetscMalloc1(n + 1, &cia));
303: PetscCall(PetscMalloc1(nz, &cja));
304: PetscCall(PetscMalloc1(nz, &cspidx));
305: jj = a->j;
306: for (i = 0; i < nz; i++) collengths[jj[i]]++;
307: cia[0] = oshift;
308: for (i = 0; i < n; i++) cia[i + 1] = cia[i] + collengths[i];
309: PetscCall(PetscArrayzero(collengths, n));
310: jj = a->j;
311: for (row = 0; row < m; row++) {
312: mr = a->i[row + 1] - a->i[row];
313: for (i = 0; i < mr; i++) {
314: col = *jj++;
315: tmp = cia[col] + collengths[col]++ - oshift;
316: cspidx[tmp] = a->i[row] + i; /* index of a->j */
317: cja[tmp] = row + oshift;
318: }
319: }
320: PetscCall(PetscFree(collengths));
321: *ia = cia;
322: *ja = cja;
323: *spidx = cspidx;
324: PetscFunctionReturn(PETSC_SUCCESS);
325: }
327: PetscErrorCode MatRestoreColumnIJ_SeqAIJ_Color(Mat A, PetscInt oshift, PetscBool symmetric, PetscBool inodecompressed, PetscInt *n, const PetscInt *ia[], const PetscInt *ja[], PetscInt *spidx[], PetscBool *done)
328: {
329: PetscFunctionBegin;
330: PetscCall(MatRestoreColumnIJ_SeqAIJ(A, oshift, symmetric, inodecompressed, n, ia, ja, done));
331: PetscCall(PetscFree(*spidx));
332: PetscFunctionReturn(PETSC_SUCCESS);
333: }
335: static PetscErrorCode MatSetValuesRow_SeqAIJ(Mat A, PetscInt row, const PetscScalar v[])
336: {
337: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
338: PetscInt *ai = a->i;
339: PetscScalar *aa;
341: PetscFunctionBegin;
342: PetscCall(MatSeqAIJGetArray(A, &aa));
343: PetscCall(PetscArraycpy(aa + ai[row], v, ai[row + 1] - ai[row]));
344: PetscCall(MatSeqAIJRestoreArray(A, &aa));
345: PetscFunctionReturn(PETSC_SUCCESS);
346: }
348: /*
349: MatSeqAIJSetValuesLocalFast - An optimized version of MatSetValuesLocal() for SeqAIJ matrices with several assumptions
351: - a single row of values is set with each call
352: - no row or column indices are negative or (in error) larger than the number of rows or columns
353: - the values are always added to the matrix, not set
354: - no new locations are introduced in the nonzero structure of the matrix
356: This does NOT assume the global column indices are sorted
358: */
360: #include <petsc/private/isimpl.h>
361: PetscErrorCode MatSeqAIJSetValuesLocalFast(Mat A, PetscInt m, const PetscInt im[], PetscInt n, const PetscInt in[], const PetscScalar v[], InsertMode is)
362: {
363: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
364: PetscInt low, high, t, row, nrow, i, col, l;
365: const PetscInt *rp, *ai = a->i, *ailen = a->ilen, *aj = a->j;
366: PetscInt lastcol = -1;
367: MatScalar *ap, value, *aa;
368: const PetscInt *ridx = A->rmap->mapping->indices, *cidx = A->cmap->mapping->indices;
370: PetscFunctionBegin;
371: PetscCall(MatSeqAIJGetArray(A, &aa));
372: row = ridx[im[0]];
373: rp = aj + ai[row];
374: ap = aa + ai[row];
375: nrow = ailen[row];
376: low = 0;
377: high = nrow;
378: for (l = 0; l < n; l++) { /* loop over added columns */
379: col = cidx[in[l]];
380: value = v[l];
382: if (col <= lastcol) low = 0;
383: else high = nrow;
384: lastcol = col;
385: while (high - low > 5) {
386: t = (low + high) / 2;
387: if (rp[t] > col) high = t;
388: else low = t;
389: }
390: for (i = low; i < high; i++) {
391: if (rp[i] == col) {
392: ap[i] += value;
393: low = i + 1;
394: break;
395: }
396: }
397: }
398: PetscCall(MatSeqAIJRestoreArray(A, &aa));
399: return PETSC_SUCCESS;
400: }
402: PetscErrorCode MatSetValues_SeqAIJ(Mat A, PetscInt m, const PetscInt im[], PetscInt n, const PetscInt in[], const PetscScalar v[], InsertMode is)
403: {
404: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
405: PetscInt *rp, k, low, high, t, ii, row, nrow, i, col, l, rmax, N;
406: PetscInt *imax = a->imax, *ai = a->i, *ailen = a->ilen;
407: PetscInt *aj = a->j, nonew = a->nonew, lastcol = -1;
408: MatScalar *ap = NULL, value = 0.0, *aa;
409: PetscBool ignorezeroentries = a->ignorezeroentries;
410: PetscBool roworiented = a->roworiented;
412: PetscFunctionBegin;
413: PetscCall(MatSeqAIJGetArray(A, &aa));
414: for (k = 0; k < m; k++) { /* loop over added rows */
415: row = im[k];
416: if (row < 0) continue;
417: PetscCheck(row < A->rmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Row too large: row %" PetscInt_FMT " max %" PetscInt_FMT, row, A->rmap->n - 1);
418: rp = PetscSafePointerPlusOffset(aj, ai[row]);
419: if (!A->structure_only) ap = PetscSafePointerPlusOffset(aa, ai[row]);
420: rmax = imax[row];
421: nrow = ailen[row];
422: low = 0;
423: high = nrow;
424: for (l = 0; l < n; l++) { /* loop over added columns */
425: if (in[l] < 0) continue;
426: PetscCheck(in[l] < A->cmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Column too large: col %" PetscInt_FMT " max %" PetscInt_FMT, in[l], A->cmap->n - 1);
427: col = in[l];
428: if (v && !A->structure_only) value = roworiented ? v[l + k * n] : v[k + l * m];
429: if (!A->structure_only && value == 0.0 && ignorezeroentries && is == ADD_VALUES && row != col) continue;
431: if (col <= lastcol) low = 0;
432: else high = nrow;
433: lastcol = col;
434: while (high - low > 5) {
435: t = (low + high) / 2;
436: if (rp[t] > col) high = t;
437: else low = t;
438: }
439: for (i = low; i < high; i++) {
440: if (rp[i] > col) break;
441: if (rp[i] == col) {
442: if (!A->structure_only) {
443: if (is == ADD_VALUES) {
444: ap[i] += value;
445: (void)PetscLogFlops(1.0);
446: } else ap[i] = value;
447: }
448: low = i + 1;
449: goto noinsert;
450: }
451: }
452: if (value == 0.0 && ignorezeroentries && row != col) goto noinsert;
453: if (nonew == 1) goto noinsert;
454: PetscCheck(nonew != -1, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Inserting a new nonzero at (%" PetscInt_FMT ",%" PetscInt_FMT ") in the matrix", row, col);
455: if (A->structure_only) {
456: MatSeqXAIJReallocateAIJ_structure_only(A, A->rmap->n, 1, nrow, row, col, rmax, ai, aj, rp, imax, nonew, MatScalar);
457: } else {
458: MatSeqXAIJReallocateAIJ(A, A->rmap->n, 1, nrow, row, col, rmax, aa, ai, aj, rp, ap, imax, nonew, MatScalar);
459: }
460: N = nrow++ - 1;
461: a->nz++;
462: high++;
463: /* shift up all the later entries in this row */
464: PetscCall(PetscArraymove(rp + i + 1, rp + i, N - i + 1));
465: rp[i] = col;
466: if (!A->structure_only) {
467: PetscCall(PetscArraymove(ap + i + 1, ap + i, N - i + 1));
468: ap[i] = value;
469: }
470: low = i + 1;
471: noinsert:;
472: }
473: ailen[row] = nrow;
474: }
475: PetscCall(MatSeqAIJRestoreArray(A, &aa));
476: PetscFunctionReturn(PETSC_SUCCESS);
477: }
479: static PetscErrorCode MatSetValues_SeqAIJ_SortedFullNoPreallocation(Mat A, PetscInt m, const PetscInt im[], PetscInt n, const PetscInt in[], const PetscScalar v[], InsertMode is)
480: {
481: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
482: PetscInt *rp, k, row;
483: PetscInt *ai = a->i;
484: PetscInt *aj = a->j;
485: MatScalar *aa, *ap;
487: PetscFunctionBegin;
488: PetscCheck(!A->was_assembled, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Cannot call on assembled matrix.");
489: PetscCheck(m * n + a->nz <= a->maxnz, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Number of entries in matrix will be larger than maximum nonzeros allocated for %" PetscInt_FMT " in MatSeqAIJSetTotalPreallocation()", a->maxnz);
491: PetscCall(MatSeqAIJGetArray(A, &aa));
492: for (k = 0; k < m; k++) { /* loop over added rows */
493: row = im[k];
494: rp = aj + ai[row];
495: ap = PetscSafePointerPlusOffset(aa, ai[row]);
497: PetscCall(PetscMemcpy(rp, in, n * sizeof(PetscInt)));
498: if (!A->structure_only) {
499: if (v) {
500: PetscCall(PetscMemcpy(ap, v, n * sizeof(PetscScalar)));
501: v += n;
502: } else {
503: PetscCall(PetscMemzero(ap, n * sizeof(PetscScalar)));
504: }
505: }
506: a->ilen[row] = n;
507: a->imax[row] = n;
508: a->i[row + 1] = a->i[row] + n;
509: a->nz += n;
510: }
511: PetscCall(MatSeqAIJRestoreArray(A, &aa));
512: PetscFunctionReturn(PETSC_SUCCESS);
513: }
515: /*@
516: MatSeqAIJSetTotalPreallocation - Sets an upper bound on the total number of expected nonzeros in the matrix.
518: Input Parameters:
519: + A - the `MATSEQAIJ` matrix
520: - nztotal - bound on the number of nonzeros
522: Level: advanced
524: Notes:
525: This can be called if you will be provided the matrix row by row (from row zero) with sorted column indices for each row.
526: Simply call `MatSetValues()` after this call to provide the matrix entries in the usual manner. This matrix may be used
527: as always with multiple matrix assemblies.
529: .seealso: [](ch_matrices), `Mat`, `MatSetOption()`, `MAT_SORTED_FULL`, `MatSetValues()`, `MatSeqAIJSetPreallocation()`
530: @*/
531: PetscErrorCode MatSeqAIJSetTotalPreallocation(Mat A, PetscInt nztotal)
532: {
533: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
535: PetscFunctionBegin;
536: PetscCall(PetscLayoutSetUp(A->rmap));
537: PetscCall(PetscLayoutSetUp(A->cmap));
538: a->maxnz = nztotal;
539: if (!a->imax) { PetscCall(PetscMalloc1(A->rmap->n, &a->imax)); }
540: if (!a->ilen) {
541: PetscCall(PetscMalloc1(A->rmap->n, &a->ilen));
542: } else {
543: PetscCall(PetscMemzero(a->ilen, A->rmap->n * sizeof(PetscInt)));
544: }
546: /* allocate the matrix space */
547: PetscCall(PetscShmgetAllocateArray(A->rmap->n + 1, sizeof(PetscInt), (void **)&a->i));
548: PetscCall(PetscShmgetAllocateArray(nztotal, sizeof(PetscInt), (void **)&a->j));
549: a->free_ij = PETSC_TRUE;
550: if (A->structure_only) {
551: a->free_a = PETSC_FALSE;
552: } else {
553: PetscCall(PetscShmgetAllocateArray(nztotal, sizeof(PetscScalar), (void **)&a->a));
554: a->free_a = PETSC_TRUE;
555: }
556: a->i[0] = 0;
557: A->ops->setvalues = MatSetValues_SeqAIJ_SortedFullNoPreallocation;
558: A->preallocated = PETSC_TRUE;
559: PetscFunctionReturn(PETSC_SUCCESS);
560: }
562: static PetscErrorCode MatSetValues_SeqAIJ_SortedFull(Mat A, PetscInt m, const PetscInt im[], PetscInt n, const PetscInt in[], const PetscScalar v[], InsertMode is)
563: {
564: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
565: PetscInt *rp, k, row;
566: PetscInt *ai = a->i, *ailen = a->ilen;
567: PetscInt *aj = a->j;
568: MatScalar *aa, *ap;
570: PetscFunctionBegin;
571: PetscCall(MatSeqAIJGetArray(A, &aa));
572: for (k = 0; k < m; k++) { /* loop over added rows */
573: row = im[k];
574: PetscCheck(n <= a->imax[row], PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Preallocation for row %" PetscInt_FMT " does not match number of columns provided", n);
575: rp = aj + ai[row];
576: ap = aa + ai[row];
577: if (!A->was_assembled) PetscCall(PetscMemcpy(rp, in, n * sizeof(PetscInt)));
578: if (!A->structure_only) {
579: if (v) {
580: PetscCall(PetscMemcpy(ap, v, n * sizeof(PetscScalar)));
581: v += n;
582: } else {
583: PetscCall(PetscMemzero(ap, n * sizeof(PetscScalar)));
584: }
585: }
586: ailen[row] = n;
587: a->nz += n;
588: }
589: PetscCall(MatSeqAIJRestoreArray(A, &aa));
590: PetscFunctionReturn(PETSC_SUCCESS);
591: }
593: static PetscErrorCode MatGetValues_SeqAIJ(Mat A, PetscInt m, const PetscInt im[], PetscInt n, const PetscInt in[], PetscScalar v[])
594: {
595: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
596: PetscInt *rp, k, low, high, t, row, nrow, i, col, l, *aj = a->j;
597: PetscInt *ai = a->i, *ailen = a->ilen;
598: const MatScalar *ap, *aa;
600: PetscFunctionBegin;
601: PetscCall(MatSeqAIJGetArrayRead(A, &aa));
602: for (k = 0; k < m; k++) { /* loop over rows */
603: row = im[k];
604: if (row < 0) {
605: v += n;
606: continue;
607: } /* negative row */
608: PetscCheck(row < A->rmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Row too large: row %" PetscInt_FMT " max %" PetscInt_FMT, row, A->rmap->n - 1);
609: rp = PetscSafePointerPlusOffset(aj, ai[row]);
610: ap = PetscSafePointerPlusOffset(aa, ai[row]);
611: nrow = ailen[row];
612: for (l = 0; l < n; l++) { /* loop over columns */
613: if (in[l] < 0) {
614: v++;
615: continue;
616: } /* negative column */
617: PetscCheck(in[l] < A->cmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Column too large: col %" PetscInt_FMT " max %" PetscInt_FMT, in[l], A->cmap->n - 1);
618: col = in[l];
619: high = nrow;
620: low = 0; /* assume unsorted */
621: while (high - low > 5) {
622: t = (low + high) / 2;
623: if (rp[t] > col) high = t;
624: else low = t;
625: }
626: for (i = low; i < high; i++) {
627: if (rp[i] > col) break;
628: if (rp[i] == col) {
629: *v++ = ap[i];
630: goto finished;
631: }
632: }
633: *v++ = 0.0;
634: finished:;
635: }
636: }
637: PetscCall(MatSeqAIJRestoreArrayRead(A, &aa));
638: PetscFunctionReturn(PETSC_SUCCESS);
639: }
641: static PetscErrorCode MatView_SeqAIJ_Binary(Mat mat, PetscViewer viewer)
642: {
643: Mat_SeqAIJ *A = (Mat_SeqAIJ *)mat->data;
644: const PetscScalar *av;
645: PetscInt header[4], M, N, m, nz, i;
646: PetscInt *rowlens;
648: PetscFunctionBegin;
649: PetscCall(PetscViewerSetUp(viewer));
651: M = mat->rmap->N;
652: N = mat->cmap->N;
653: m = mat->rmap->n;
654: nz = A->nz;
656: /* write matrix header */
657: header[0] = MAT_FILE_CLASSID;
658: header[1] = M;
659: header[2] = N;
660: header[3] = nz;
661: PetscCall(PetscViewerBinaryWrite(viewer, header, 4, PETSC_INT));
663: /* fill in and store row lengths */
664: PetscCall(PetscMalloc1(m, &rowlens));
665: for (i = 0; i < m; i++) rowlens[i] = A->i[i + 1] - A->i[i];
666: if (PetscDefined(USE_DEBUG)) {
667: PetscInt mnz = 0;
669: for (i = 0; i < m; i++) mnz += rowlens[i];
670: PetscCheck(nz == mnz, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Row lens %" PetscInt_FMT " do not sum to nz %" PetscInt_FMT, mnz, nz);
671: }
672: PetscCall(PetscViewerBinaryWrite(viewer, rowlens, m, PETSC_INT));
673: PetscCall(PetscFree(rowlens));
674: /* store column indices */
675: PetscCall(PetscViewerBinaryWrite(viewer, A->j, nz, PETSC_INT));
676: /* store nonzero values */
677: PetscCall(MatSeqAIJGetArrayRead(mat, &av));
678: PetscCall(PetscViewerBinaryWrite(viewer, av, nz, PETSC_SCALAR));
679: PetscCall(MatSeqAIJRestoreArrayRead(mat, &av));
681: /* write block size option to the viewer's .info file */
682: PetscCall(MatView_Binary_BlockSizes(mat, viewer));
683: PetscFunctionReturn(PETSC_SUCCESS);
684: }
686: static PetscErrorCode MatView_SeqAIJ_ASCII_structonly(Mat A, PetscViewer viewer)
687: {
688: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
689: PetscInt i, k, m = A->rmap->N;
691: PetscFunctionBegin;
692: PetscCall(PetscViewerASCIIUseTabs(viewer, PETSC_FALSE));
693: for (i = 0; i < m; i++) {
694: PetscCall(PetscViewerASCIIPrintf(viewer, "row %" PetscInt_FMT ":", i));
695: for (k = a->i[i]; k < a->i[i + 1]; k++) PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ") ", a->j[k]));
696: PetscCall(PetscViewerASCIIPrintf(viewer, "\n"));
697: }
698: PetscCall(PetscViewerASCIIUseTabs(viewer, PETSC_TRUE));
699: PetscFunctionReturn(PETSC_SUCCESS);
700: }
702: extern PetscErrorCode MatSeqAIJFactorInfo_Matlab(Mat, PetscViewer);
704: static PetscErrorCode MatView_SeqAIJ_ASCII(Mat A, PetscViewer viewer)
705: {
706: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
707: const PetscScalar *av;
708: PetscInt i, j, m = A->rmap->n;
709: const char *name;
710: PetscViewerFormat format;
712: PetscFunctionBegin;
713: if (A->structure_only) {
714: PetscCall(MatView_SeqAIJ_ASCII_structonly(A, viewer));
715: PetscFunctionReturn(PETSC_SUCCESS);
716: }
718: PetscCall(PetscViewerGetFormat(viewer, &format));
719: // By petsc's rule, even PETSC_VIEWER_ASCII_INFO_DETAIL doesn't print matrix entries
720: if (format == PETSC_VIEWER_ASCII_FACTOR_INFO || format == PETSC_VIEWER_ASCII_INFO || format == PETSC_VIEWER_ASCII_INFO_DETAIL) PetscFunctionReturn(PETSC_SUCCESS);
722: /* trigger copy to CPU if needed */
723: PetscCall(MatSeqAIJGetArrayRead(A, &av));
724: PetscCall(MatSeqAIJRestoreArrayRead(A, &av));
725: if (format == PETSC_VIEWER_ASCII_MATLAB) {
726: PetscInt nofinalvalue = 0;
727: if (m && ((a->i[m] == a->i[m - 1]) || (a->j[a->nz - 1] != A->cmap->n - 1))) {
728: /* Need a dummy value to ensure the dimension of the matrix. */
729: nofinalvalue = 1;
730: }
731: PetscCall(PetscViewerASCIIUseTabs(viewer, PETSC_FALSE));
732: PetscCall(PetscViewerASCIIPrintf(viewer, "%% Size = %" PetscInt_FMT " %" PetscInt_FMT " \n", m, A->cmap->n));
733: PetscCall(PetscViewerASCIIPrintf(viewer, "%% Nonzeros = %" PetscInt_FMT " \n", a->nz));
734: #if defined(PETSC_USE_COMPLEX)
735: PetscCall(PetscViewerASCIIPrintf(viewer, "zzz = zeros(%" PetscInt_FMT ",4);\n", a->nz + nofinalvalue));
736: #else
737: PetscCall(PetscViewerASCIIPrintf(viewer, "zzz = zeros(%" PetscInt_FMT ",3);\n", a->nz + nofinalvalue));
738: #endif
739: PetscCall(PetscViewerASCIIPrintf(viewer, "zzz = [\n"));
741: for (i = 0; i < m; i++) {
742: for (j = a->i[i]; j < a->i[i + 1]; j++) {
743: #if defined(PETSC_USE_COMPLEX)
744: PetscCall(PetscViewerASCIIPrintf(viewer, "%" PetscInt_FMT " %" PetscInt_FMT " %18.16e %18.16e\n", i + 1, a->j[j] + 1, (double)PetscRealPart(a->a[j]), (double)PetscImaginaryPart(a->a[j])));
745: #else
746: PetscCall(PetscViewerASCIIPrintf(viewer, "%" PetscInt_FMT " %" PetscInt_FMT " %18.16e\n", i + 1, a->j[j] + 1, (double)a->a[j]));
747: #endif
748: }
749: }
750: if (nofinalvalue) {
751: #if defined(PETSC_USE_COMPLEX)
752: PetscCall(PetscViewerASCIIPrintf(viewer, "%" PetscInt_FMT " %" PetscInt_FMT " %18.16e %18.16e\n", m, A->cmap->n, 0., 0.));
753: #else
754: PetscCall(PetscViewerASCIIPrintf(viewer, "%" PetscInt_FMT " %" PetscInt_FMT " %18.16e\n", m, A->cmap->n, 0.0));
755: #endif
756: }
757: PetscCall(PetscObjectGetName((PetscObject)A, &name));
758: PetscCall(PetscViewerASCIIPrintf(viewer, "];\n %s = spconvert(zzz);\n", name));
759: PetscCall(PetscViewerASCIIUseTabs(viewer, PETSC_TRUE));
760: } else if (format == PETSC_VIEWER_ASCII_COMMON) {
761: PetscCall(PetscViewerASCIIUseTabs(viewer, PETSC_FALSE));
762: for (i = 0; i < m; i++) {
763: PetscCall(PetscViewerASCIIPrintf(viewer, "row %" PetscInt_FMT ":", i));
764: for (j = a->i[i]; j < a->i[i + 1]; j++) {
765: #if defined(PETSC_USE_COMPLEX)
766: if (PetscImaginaryPart(a->a[j]) > 0.0 && PetscRealPart(a->a[j]) != 0.0) {
767: PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g + %g i)", a->j[j], (double)PetscRealPart(a->a[j]), (double)PetscImaginaryPart(a->a[j])));
768: } else if (PetscImaginaryPart(a->a[j]) < 0.0 && PetscRealPart(a->a[j]) != 0.0) {
769: PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g - %g i)", a->j[j], (double)PetscRealPart(a->a[j]), (double)-PetscImaginaryPart(a->a[j])));
770: } else if (PetscRealPart(a->a[j]) != 0.0) {
771: PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g) ", a->j[j], (double)PetscRealPart(a->a[j])));
772: }
773: #else
774: if (a->a[j] != 0.0) PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g) ", a->j[j], (double)a->a[j]));
775: #endif
776: }
777: PetscCall(PetscViewerASCIIPrintf(viewer, "\n"));
778: }
779: PetscCall(PetscViewerASCIIUseTabs(viewer, PETSC_TRUE));
780: } else if (format == PETSC_VIEWER_ASCII_SYMMODU) {
781: PetscInt nzd = 0, fshift = 1, *sptr;
782: PetscCall(PetscViewerASCIIUseTabs(viewer, PETSC_FALSE));
783: PetscCall(PetscMalloc1(m + 1, &sptr));
784: for (i = 0; i < m; i++) {
785: sptr[i] = nzd + 1;
786: for (j = a->i[i]; j < a->i[i + 1]; j++) {
787: if (a->j[j] >= i) {
788: #if defined(PETSC_USE_COMPLEX)
789: if (PetscImaginaryPart(a->a[j]) != 0.0 || PetscRealPart(a->a[j]) != 0.0) nzd++;
790: #else
791: if (a->a[j] != 0.0) nzd++;
792: #endif
793: }
794: }
795: }
796: sptr[m] = nzd + 1;
797: PetscCall(PetscViewerASCIIPrintf(viewer, " %" PetscInt_FMT " %" PetscInt_FMT "\n\n", m, nzd));
798: for (i = 0; i < m + 1; i += 6) {
799: if (i + 4 < m) {
800: PetscCall(PetscViewerASCIIPrintf(viewer, " %" PetscInt_FMT " %" PetscInt_FMT " %" PetscInt_FMT " %" PetscInt_FMT " %" PetscInt_FMT " %" PetscInt_FMT "\n", sptr[i], sptr[i + 1], sptr[i + 2], sptr[i + 3], sptr[i + 4], sptr[i + 5]));
801: } else if (i + 3 < m) {
802: PetscCall(PetscViewerASCIIPrintf(viewer, " %" PetscInt_FMT " %" PetscInt_FMT " %" PetscInt_FMT " %" PetscInt_FMT " %" PetscInt_FMT "\n", sptr[i], sptr[i + 1], sptr[i + 2], sptr[i + 3], sptr[i + 4]));
803: } else if (i + 2 < m) {
804: PetscCall(PetscViewerASCIIPrintf(viewer, " %" PetscInt_FMT " %" PetscInt_FMT " %" PetscInt_FMT " %" PetscInt_FMT "\n", sptr[i], sptr[i + 1], sptr[i + 2], sptr[i + 3]));
805: } else if (i + 1 < m) {
806: PetscCall(PetscViewerASCIIPrintf(viewer, " %" PetscInt_FMT " %" PetscInt_FMT " %" PetscInt_FMT "\n", sptr[i], sptr[i + 1], sptr[i + 2]));
807: } else if (i < m) {
808: PetscCall(PetscViewerASCIIPrintf(viewer, " %" PetscInt_FMT " %" PetscInt_FMT "\n", sptr[i], sptr[i + 1]));
809: } else {
810: PetscCall(PetscViewerASCIIPrintf(viewer, " %" PetscInt_FMT "\n", sptr[i]));
811: }
812: }
813: PetscCall(PetscViewerASCIIPrintf(viewer, "\n"));
814: PetscCall(PetscFree(sptr));
815: for (i = 0; i < m; i++) {
816: for (j = a->i[i]; j < a->i[i + 1]; j++) {
817: if (a->j[j] >= i) PetscCall(PetscViewerASCIIPrintf(viewer, " %" PetscInt_FMT " ", a->j[j] + fshift));
818: }
819: PetscCall(PetscViewerASCIIPrintf(viewer, "\n"));
820: }
821: PetscCall(PetscViewerASCIIPrintf(viewer, "\n"));
822: for (i = 0; i < m; i++) {
823: for (j = a->i[i]; j < a->i[i + 1]; j++) {
824: if (a->j[j] >= i) {
825: #if defined(PETSC_USE_COMPLEX)
826: if (PetscImaginaryPart(a->a[j]) != 0.0 || PetscRealPart(a->a[j]) != 0.0) PetscCall(PetscViewerASCIIPrintf(viewer, " %18.16e %18.16e ", (double)PetscRealPart(a->a[j]), (double)PetscImaginaryPart(a->a[j])));
827: #else
828: if (a->a[j] != 0.0) PetscCall(PetscViewerASCIIPrintf(viewer, " %18.16e ", (double)a->a[j]));
829: #endif
830: }
831: }
832: PetscCall(PetscViewerASCIIPrintf(viewer, "\n"));
833: }
834: PetscCall(PetscViewerASCIIUseTabs(viewer, PETSC_TRUE));
835: } else if (format == PETSC_VIEWER_ASCII_DENSE) {
836: PetscInt cnt = 0, jcnt;
837: PetscScalar value;
838: #if defined(PETSC_USE_COMPLEX)
839: PetscBool realonly = PETSC_TRUE;
841: for (i = 0; i < a->i[m]; i++) {
842: if (PetscImaginaryPart(a->a[i]) != 0.0) {
843: realonly = PETSC_FALSE;
844: break;
845: }
846: }
847: #endif
849: PetscCall(PetscViewerASCIIUseTabs(viewer, PETSC_FALSE));
850: for (i = 0; i < m; i++) {
851: jcnt = 0;
852: for (j = 0; j < A->cmap->n; j++) {
853: if (jcnt < a->i[i + 1] - a->i[i] && j == a->j[cnt]) {
854: value = a->a[cnt++];
855: jcnt++;
856: } else {
857: value = 0.0;
858: }
859: #if defined(PETSC_USE_COMPLEX)
860: if (realonly) {
861: PetscCall(PetscViewerASCIIPrintf(viewer, " %7.5e ", (double)PetscRealPart(value)));
862: } else {
863: PetscCall(PetscViewerASCIIPrintf(viewer, " %7.5e+%7.5e i ", (double)PetscRealPart(value), (double)PetscImaginaryPart(value)));
864: }
865: #else
866: PetscCall(PetscViewerASCIIPrintf(viewer, " %7.5e ", (double)value));
867: #endif
868: }
869: PetscCall(PetscViewerASCIIPrintf(viewer, "\n"));
870: }
871: PetscCall(PetscViewerASCIIUseTabs(viewer, PETSC_TRUE));
872: } else if (format == PETSC_VIEWER_ASCII_MATRIXMARKET) {
873: PetscInt fshift = 1;
874: PetscCall(PetscViewerASCIIUseTabs(viewer, PETSC_FALSE));
875: #if defined(PETSC_USE_COMPLEX)
876: PetscCall(PetscViewerASCIIPrintf(viewer, "%%%%MatrixMarket matrix coordinate complex general\n"));
877: #else
878: PetscCall(PetscViewerASCIIPrintf(viewer, "%%%%MatrixMarket matrix coordinate real general\n"));
879: #endif
880: PetscCall(PetscViewerASCIIPrintf(viewer, "%" PetscInt_FMT " %" PetscInt_FMT " %" PetscInt_FMT "\n", m, A->cmap->n, a->nz));
881: for (i = 0; i < m; i++) {
882: for (j = a->i[i]; j < a->i[i + 1]; j++) {
883: #if defined(PETSC_USE_COMPLEX)
884: PetscCall(PetscViewerASCIIPrintf(viewer, "%" PetscInt_FMT " %" PetscInt_FMT " %g %g\n", i + fshift, a->j[j] + fshift, (double)PetscRealPart(a->a[j]), (double)PetscImaginaryPart(a->a[j])));
885: #else
886: PetscCall(PetscViewerASCIIPrintf(viewer, "%" PetscInt_FMT " %" PetscInt_FMT " %g\n", i + fshift, a->j[j] + fshift, (double)a->a[j]));
887: #endif
888: }
889: }
890: PetscCall(PetscViewerASCIIUseTabs(viewer, PETSC_TRUE));
891: } else {
892: PetscCall(PetscViewerASCIIUseTabs(viewer, PETSC_FALSE));
893: if (A->factortype) {
894: for (i = 0; i < m; i++) {
895: PetscCall(PetscViewerASCIIPrintf(viewer, "row %" PetscInt_FMT ":", i));
896: /* L part */
897: for (j = a->i[i]; j < a->i[i + 1]; j++) {
898: #if defined(PETSC_USE_COMPLEX)
899: if (PetscImaginaryPart(a->a[j]) > 0.0) {
900: PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g + %g i)", a->j[j], (double)PetscRealPart(a->a[j]), (double)PetscImaginaryPart(a->a[j])));
901: } else if (PetscImaginaryPart(a->a[j]) < 0.0) {
902: PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g - %g i)", a->j[j], (double)PetscRealPart(a->a[j]), (double)(-PetscImaginaryPart(a->a[j]))));
903: } else {
904: PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g) ", a->j[j], (double)PetscRealPart(a->a[j])));
905: }
906: #else
907: PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g) ", a->j[j], (double)a->a[j]));
908: #endif
909: }
910: /* diagonal */
911: j = a->diag[i];
912: #if defined(PETSC_USE_COMPLEX)
913: if (PetscImaginaryPart(a->a[j]) > 0.0) {
914: PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g + %g i)", a->j[j], (double)PetscRealPart(1 / a->a[j]), (double)PetscImaginaryPart(1 / a->a[j])));
915: } else if (PetscImaginaryPart(a->a[j]) < 0.0) {
916: PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g - %g i)", a->j[j], (double)PetscRealPart(1 / a->a[j]), (double)(-PetscImaginaryPart(1 / a->a[j]))));
917: } else {
918: PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g) ", a->j[j], (double)PetscRealPart(1 / a->a[j])));
919: }
920: #else
921: PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g) ", a->j[j], (double)(1 / a->a[j])));
922: #endif
924: /* U part */
925: for (j = a->diag[i + 1] + 1; j < a->diag[i]; j++) {
926: #if defined(PETSC_USE_COMPLEX)
927: if (PetscImaginaryPart(a->a[j]) > 0.0) {
928: PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g + %g i)", a->j[j], (double)PetscRealPart(a->a[j]), (double)PetscImaginaryPart(a->a[j])));
929: } else if (PetscImaginaryPart(a->a[j]) < 0.0) {
930: PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g - %g i)", a->j[j], (double)PetscRealPart(a->a[j]), (double)(-PetscImaginaryPart(a->a[j]))));
931: } else {
932: PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g) ", a->j[j], (double)PetscRealPart(a->a[j])));
933: }
934: #else
935: PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g) ", a->j[j], (double)a->a[j]));
936: #endif
937: }
938: PetscCall(PetscViewerASCIIPrintf(viewer, "\n"));
939: }
940: } else {
941: for (i = 0; i < m; i++) {
942: PetscCall(PetscViewerASCIIPrintf(viewer, "row %" PetscInt_FMT ":", i));
943: for (j = a->i[i]; j < a->i[i + 1]; j++) {
944: #if defined(PETSC_USE_COMPLEX)
945: if (PetscImaginaryPart(a->a[j]) > 0.0) {
946: PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g + %g i)", a->j[j], (double)PetscRealPart(a->a[j]), (double)PetscImaginaryPart(a->a[j])));
947: } else if (PetscImaginaryPart(a->a[j]) < 0.0) {
948: PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g - %g i)", a->j[j], (double)PetscRealPart(a->a[j]), (double)-PetscImaginaryPart(a->a[j])));
949: } else {
950: PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g) ", a->j[j], (double)PetscRealPart(a->a[j])));
951: }
952: #else
953: PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g) ", a->j[j], (double)a->a[j]));
954: #endif
955: }
956: PetscCall(PetscViewerASCIIPrintf(viewer, "\n"));
957: }
958: }
959: PetscCall(PetscViewerASCIIUseTabs(viewer, PETSC_TRUE));
960: }
961: PetscCall(PetscViewerFlush(viewer));
962: PetscFunctionReturn(PETSC_SUCCESS);
963: }
965: #include <petscdraw.h>
966: static PetscErrorCode MatView_SeqAIJ_Draw_Zoom(PetscDraw draw, void *Aa)
967: {
968: Mat A = (Mat)Aa;
969: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
970: PetscInt i, j, m = A->rmap->n;
971: int color;
972: PetscReal xl, yl, xr, yr, x_l, x_r, y_l, y_r;
973: PetscViewer viewer;
974: PetscViewerFormat format;
975: const PetscScalar *aa;
977: PetscFunctionBegin;
978: PetscCall(PetscObjectQuery((PetscObject)A, "Zoomviewer", (PetscObject *)&viewer));
979: PetscCall(PetscViewerGetFormat(viewer, &format));
980: PetscCall(PetscDrawGetCoordinates(draw, &xl, &yl, &xr, &yr));
982: /* loop over matrix elements drawing boxes */
983: PetscCall(MatSeqAIJGetArrayRead(A, &aa));
984: if (format != PETSC_VIEWER_DRAW_CONTOUR) {
985: PetscDrawCollectiveBegin(draw);
986: /* Blue for negative, Cyan for zero and Red for positive */
987: color = PETSC_DRAW_BLUE;
988: for (i = 0; i < m; i++) {
989: y_l = m - i - 1.0;
990: y_r = y_l + 1.0;
991: for (j = a->i[i]; j < a->i[i + 1]; j++) {
992: x_l = a->j[j];
993: x_r = x_l + 1.0;
994: if (PetscRealPart(aa[j]) >= 0.) continue;
995: PetscCall(PetscDrawRectangle(draw, x_l, y_l, x_r, y_r, color, color, color, color));
996: }
997: }
998: color = PETSC_DRAW_CYAN;
999: for (i = 0; i < m; i++) {
1000: y_l = m - i - 1.0;
1001: y_r = y_l + 1.0;
1002: for (j = a->i[i]; j < a->i[i + 1]; j++) {
1003: x_l = a->j[j];
1004: x_r = x_l + 1.0;
1005: if (aa[j] != 0.) continue;
1006: PetscCall(PetscDrawRectangle(draw, x_l, y_l, x_r, y_r, color, color, color, color));
1007: }
1008: }
1009: color = PETSC_DRAW_RED;
1010: for (i = 0; i < m; i++) {
1011: y_l = m - i - 1.0;
1012: y_r = y_l + 1.0;
1013: for (j = a->i[i]; j < a->i[i + 1]; j++) {
1014: x_l = a->j[j];
1015: x_r = x_l + 1.0;
1016: if (PetscRealPart(aa[j]) <= 0.) continue;
1017: PetscCall(PetscDrawRectangle(draw, x_l, y_l, x_r, y_r, color, color, color, color));
1018: }
1019: }
1020: PetscDrawCollectiveEnd(draw);
1021: } else {
1022: /* use contour shading to indicate magnitude of values */
1023: /* first determine max of all nonzero values */
1024: PetscReal minv = 0.0, maxv = 0.0;
1025: PetscInt nz = a->nz, count = 0;
1026: PetscDraw popup;
1028: for (i = 0; i < nz; i++) {
1029: if (PetscAbsScalar(aa[i]) > maxv) maxv = PetscAbsScalar(aa[i]);
1030: }
1031: if (minv >= maxv) maxv = minv + PETSC_SMALL;
1032: PetscCall(PetscDrawGetPopup(draw, &popup));
1033: PetscCall(PetscDrawScalePopup(popup, minv, maxv));
1035: PetscDrawCollectiveBegin(draw);
1036: for (i = 0; i < m; i++) {
1037: y_l = m - i - 1.0;
1038: y_r = y_l + 1.0;
1039: for (j = a->i[i]; j < a->i[i + 1]; j++) {
1040: x_l = a->j[j];
1041: x_r = x_l + 1.0;
1042: color = PetscDrawRealToColor(PetscAbsScalar(aa[count]), minv, maxv);
1043: PetscCall(PetscDrawRectangle(draw, x_l, y_l, x_r, y_r, color, color, color, color));
1044: count++;
1045: }
1046: }
1047: PetscDrawCollectiveEnd(draw);
1048: }
1049: PetscCall(MatSeqAIJRestoreArrayRead(A, &aa));
1050: PetscFunctionReturn(PETSC_SUCCESS);
1051: }
1053: #include <petscdraw.h>
1054: static PetscErrorCode MatView_SeqAIJ_Draw(Mat A, PetscViewer viewer)
1055: {
1056: PetscDraw draw;
1057: PetscReal xr, yr, xl, yl, h, w;
1058: PetscBool isnull;
1060: PetscFunctionBegin;
1061: PetscCall(PetscViewerDrawGetDraw(viewer, 0, &draw));
1062: PetscCall(PetscDrawIsNull(draw, &isnull));
1063: if (isnull) PetscFunctionReturn(PETSC_SUCCESS);
1065: xr = A->cmap->n;
1066: yr = A->rmap->n;
1067: h = yr / 10.0;
1068: w = xr / 10.0;
1069: xr += w;
1070: yr += h;
1071: xl = -w;
1072: yl = -h;
1073: PetscCall(PetscDrawSetCoordinates(draw, xl, yl, xr, yr));
1074: PetscCall(PetscObjectCompose((PetscObject)A, "Zoomviewer", (PetscObject)viewer));
1075: PetscCall(PetscDrawZoom(draw, MatView_SeqAIJ_Draw_Zoom, A));
1076: PetscCall(PetscObjectCompose((PetscObject)A, "Zoomviewer", NULL));
1077: PetscCall(PetscDrawSave(draw));
1078: PetscFunctionReturn(PETSC_SUCCESS);
1079: }
1081: PetscErrorCode MatView_SeqAIJ(Mat A, PetscViewer viewer)
1082: {
1083: PetscBool iascii, isbinary, isdraw;
1085: PetscFunctionBegin;
1086: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERASCII, &iascii));
1087: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERBINARY, &isbinary));
1088: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERDRAW, &isdraw));
1089: if (iascii) PetscCall(MatView_SeqAIJ_ASCII(A, viewer));
1090: else if (isbinary) PetscCall(MatView_SeqAIJ_Binary(A, viewer));
1091: else if (isdraw) PetscCall(MatView_SeqAIJ_Draw(A, viewer));
1092: PetscCall(MatView_SeqAIJ_Inode(A, viewer));
1093: PetscFunctionReturn(PETSC_SUCCESS);
1094: }
1096: PetscErrorCode MatAssemblyEnd_SeqAIJ(Mat A, MatAssemblyType mode)
1097: {
1098: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
1099: PetscInt fshift = 0, i, *ai = a->i, *aj = a->j, *imax = a->imax;
1100: PetscInt m = A->rmap->n, *ip, N, *ailen = a->ilen, rmax = 0, n;
1101: MatScalar *aa = a->a, *ap;
1102: PetscReal ratio = 0.6;
1104: PetscFunctionBegin;
1105: if (mode == MAT_FLUSH_ASSEMBLY) PetscFunctionReturn(PETSC_SUCCESS);
1106: PetscCall(MatSeqAIJInvalidateDiagonal(A));
1107: if (A->was_assembled && A->ass_nonzerostate == A->nonzerostate) {
1108: /* we need to respect users asking to use or not the inodes routine in between matrix assemblies */
1109: PetscCall(MatAssemblyEnd_SeqAIJ_Inode(A, mode));
1110: PetscFunctionReturn(PETSC_SUCCESS);
1111: }
1113: if (m) rmax = ailen[0]; /* determine row with most nonzeros */
1114: for (i = 1; i < m; i++) {
1115: /* move each row back by the amount of empty slots (fshift) before it*/
1116: fshift += imax[i - 1] - ailen[i - 1];
1117: rmax = PetscMax(rmax, ailen[i]);
1118: if (fshift) {
1119: ip = aj + ai[i];
1120: ap = aa + ai[i];
1121: N = ailen[i];
1122: PetscCall(PetscArraymove(ip - fshift, ip, N));
1123: if (!A->structure_only) PetscCall(PetscArraymove(ap - fshift, ap, N));
1124: }
1125: ai[i] = ai[i - 1] + ailen[i - 1];
1126: }
1127: if (m) {
1128: fshift += imax[m - 1] - ailen[m - 1];
1129: ai[m] = ai[m - 1] + ailen[m - 1];
1130: }
1131: /* reset ilen and imax for each row */
1132: a->nonzerorowcnt = 0;
1133: if (A->structure_only) {
1134: PetscCall(PetscFree(a->imax));
1135: PetscCall(PetscFree(a->ilen));
1136: } else { /* !A->structure_only */
1137: for (i = 0; i < m; i++) {
1138: ailen[i] = imax[i] = ai[i + 1] - ai[i];
1139: a->nonzerorowcnt += ((ai[i + 1] - ai[i]) > 0);
1140: }
1141: }
1142: a->nz = ai[m];
1143: PetscCheck(!fshift || a->nounused != -1, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Unused space detected in matrix: %" PetscInt_FMT " X %" PetscInt_FMT ", %" PetscInt_FMT " unneeded", m, A->cmap->n, fshift);
1144: PetscCall(MatMarkDiagonal_SeqAIJ(A)); // since diagonal info is used a lot, it is helpful to set them up at the end of assembly
1145: a->diagonaldense = PETSC_TRUE;
1146: n = PetscMin(A->rmap->n, A->cmap->n);
1147: for (i = 0; i < n; i++) {
1148: if (a->diag[i] >= ai[i + 1]) {
1149: a->diagonaldense = PETSC_FALSE;
1150: break;
1151: }
1152: }
1153: PetscCall(PetscInfo(A, "Matrix size: %" PetscInt_FMT " X %" PetscInt_FMT "; storage space: %" PetscInt_FMT " unneeded,%" PetscInt_FMT " used\n", m, A->cmap->n, fshift, a->nz));
1154: PetscCall(PetscInfo(A, "Number of mallocs during MatSetValues() is %" PetscInt_FMT "\n", a->reallocs));
1155: PetscCall(PetscInfo(A, "Maximum nonzeros in any row is %" PetscInt_FMT "\n", rmax));
1157: A->info.mallocs += a->reallocs;
1158: a->reallocs = 0;
1159: A->info.nz_unneeded = (PetscReal)fshift;
1160: a->rmax = rmax;
1162: if (!A->structure_only) PetscCall(MatCheckCompressedRow(A, a->nonzerorowcnt, &a->compressedrow, a->i, m, ratio));
1163: PetscCall(MatAssemblyEnd_SeqAIJ_Inode(A, mode));
1164: PetscFunctionReturn(PETSC_SUCCESS);
1165: }
1167: static PetscErrorCode MatRealPart_SeqAIJ(Mat A)
1168: {
1169: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
1170: PetscInt i, nz = a->nz;
1171: MatScalar *aa;
1173: PetscFunctionBegin;
1174: PetscCall(MatSeqAIJGetArray(A, &aa));
1175: for (i = 0; i < nz; i++) aa[i] = PetscRealPart(aa[i]);
1176: PetscCall(MatSeqAIJRestoreArray(A, &aa));
1177: PetscCall(MatSeqAIJInvalidateDiagonal(A));
1178: PetscFunctionReturn(PETSC_SUCCESS);
1179: }
1181: static PetscErrorCode MatImaginaryPart_SeqAIJ(Mat A)
1182: {
1183: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
1184: PetscInt i, nz = a->nz;
1185: MatScalar *aa;
1187: PetscFunctionBegin;
1188: PetscCall(MatSeqAIJGetArray(A, &aa));
1189: for (i = 0; i < nz; i++) aa[i] = PetscImaginaryPart(aa[i]);
1190: PetscCall(MatSeqAIJRestoreArray(A, &aa));
1191: PetscCall(MatSeqAIJInvalidateDiagonal(A));
1192: PetscFunctionReturn(PETSC_SUCCESS);
1193: }
1195: PetscErrorCode MatZeroEntries_SeqAIJ(Mat A)
1196: {
1197: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
1198: MatScalar *aa;
1200: PetscFunctionBegin;
1201: PetscCall(MatSeqAIJGetArrayWrite(A, &aa));
1202: PetscCall(PetscArrayzero(aa, a->i[A->rmap->n]));
1203: PetscCall(MatSeqAIJRestoreArrayWrite(A, &aa));
1204: PetscCall(MatSeqAIJInvalidateDiagonal(A));
1205: PetscFunctionReturn(PETSC_SUCCESS);
1206: }
1208: static PetscErrorCode MatReset_SeqAIJ(Mat A)
1209: {
1210: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
1212: PetscFunctionBegin;
1213: if (A->hash_active) {
1214: A->ops[0] = a->cops;
1215: PetscCall(PetscHMapIJVDestroy(&a->ht));
1216: PetscCall(PetscFree(a->dnz));
1217: A->hash_active = PETSC_FALSE;
1218: }
1220: PetscCall(PetscLogObjectState((PetscObject)A, "Rows=%" PetscInt_FMT ", Cols=%" PetscInt_FMT ", NZ=%" PetscInt_FMT, A->rmap->n, A->cmap->n, a->nz));
1221: PetscCall(MatSeqXAIJFreeAIJ(A, &a->a, &a->j, &a->i));
1222: PetscCall(ISDestroy(&a->row));
1223: PetscCall(ISDestroy(&a->col));
1224: PetscCall(PetscFree(a->diag));
1225: PetscCall(PetscFree(a->ibdiag));
1226: PetscCall(PetscFree(a->imax));
1227: PetscCall(PetscFree(a->ilen));
1228: PetscCall(PetscFree(a->ipre));
1229: PetscCall(PetscFree3(a->idiag, a->mdiag, a->ssor_work));
1230: PetscCall(PetscFree(a->solve_work));
1231: PetscCall(ISDestroy(&a->icol));
1232: PetscCall(PetscFree(a->saved_values));
1233: PetscCall(PetscFree2(a->compressedrow.i, a->compressedrow.rindex));
1234: PetscCall(MatDestroy_SeqAIJ_Inode(A));
1235: PetscFunctionReturn(PETSC_SUCCESS);
1236: }
1238: static PetscErrorCode MatResetHash_SeqAIJ(Mat A)
1239: {
1240: PetscFunctionBegin;
1241: PetscCall(MatReset_SeqAIJ(A));
1242: PetscCall(MatCreate_SeqAIJ_Inode(A));
1243: PetscCall(MatSetUp_Seq_Hash(A));
1244: A->nonzerostate++;
1245: PetscFunctionReturn(PETSC_SUCCESS);
1246: }
1248: PetscErrorCode MatDestroy_SeqAIJ(Mat A)
1249: {
1250: PetscFunctionBegin;
1251: PetscCall(MatReset_SeqAIJ(A));
1252: PetscCall(PetscFree(A->data));
1254: /* MatMatMultNumeric_SeqAIJ_SeqAIJ_Sorted may allocate this.
1255: That function is so heavily used (sometimes in an hidden way through multnumeric function pointers)
1256: that is hard to properly add this data to the MatProduct data. We free it here to avoid
1257: users reusing the matrix object with different data to incur in obscure segmentation faults
1258: due to different matrix sizes */
1259: PetscCall(PetscObjectCompose((PetscObject)A, "__PETSc__ab_dense", NULL));
1261: PetscCall(PetscObjectChangeTypeName((PetscObject)A, NULL));
1262: PetscCall(PetscObjectComposeFunction((PetscObject)A, "PetscMatlabEnginePut_C", NULL));
1263: PetscCall(PetscObjectComposeFunction((PetscObject)A, "PetscMatlabEngineGet_C", NULL));
1264: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatSeqAIJSetColumnIndices_C", NULL));
1265: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatStoreValues_C", NULL));
1266: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatRetrieveValues_C", NULL));
1267: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_seqsbaij_C", NULL));
1268: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_seqbaij_C", NULL));
1269: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_seqaijperm_C", NULL));
1270: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_seqaijsell_C", NULL));
1271: #if defined(PETSC_HAVE_MKL_SPARSE)
1272: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_seqaijmkl_C", NULL));
1273: #endif
1274: #if defined(PETSC_HAVE_CUDA)
1275: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_seqaijcusparse_C", NULL));
1276: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_seqaijcusparse_seqaij_C", NULL));
1277: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_seqaij_seqaijcusparse_C", NULL));
1278: #endif
1279: #if defined(PETSC_HAVE_HIP)
1280: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_seqaijhipsparse_C", NULL));
1281: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_seqaijhipsparse_seqaij_C", NULL));
1282: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_seqaij_seqaijhipsparse_C", NULL));
1283: #endif
1284: #if defined(PETSC_HAVE_KOKKOS_KERNELS)
1285: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_seqaijkokkos_C", NULL));
1286: #endif
1287: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_seqaijcrl_C", NULL));
1288: #if defined(PETSC_HAVE_ELEMENTAL)
1289: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_elemental_C", NULL));
1290: #endif
1291: #if defined(PETSC_HAVE_SCALAPACK)
1292: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_scalapack_C", NULL));
1293: #endif
1294: #if defined(PETSC_HAVE_HYPRE)
1295: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_hypre_C", NULL));
1296: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_transpose_seqaij_seqaij_C", NULL));
1297: #endif
1298: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_seqdense_C", NULL));
1299: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_seqsell_C", NULL));
1300: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_is_C", NULL));
1301: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatIsTranspose_C", NULL));
1302: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatIsHermitianTranspose_C", NULL));
1303: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatSeqAIJSetPreallocation_C", NULL));
1304: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatResetPreallocation_C", NULL));
1305: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatResetHash_C", NULL));
1306: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatSeqAIJSetPreallocationCSR_C", NULL));
1307: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatReorderForNonzeroDiagonal_C", NULL));
1308: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_is_seqaij_C", NULL));
1309: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_seqdense_seqaij_C", NULL));
1310: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_seqaij_seqaij_C", NULL));
1311: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatSeqAIJKron_C", NULL));
1312: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatSetPreallocationCOO_C", NULL));
1313: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatSetValuesCOO_C", NULL));
1314: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatFactorGetSolverType_C", NULL));
1315: /* these calls do not belong here: the subclasses Duplicate/Destroy are wrong */
1316: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaijsell_seqaij_C", NULL));
1317: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaijperm_seqaij_C", NULL));
1318: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_seqaijviennacl_C", NULL));
1319: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_seqaijviennacl_seqdense_C", NULL));
1320: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_seqaijviennacl_seqaij_C", NULL));
1321: PetscFunctionReturn(PETSC_SUCCESS);
1322: }
1324: PetscErrorCode MatSetOption_SeqAIJ(Mat A, MatOption op, PetscBool flg)
1325: {
1326: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
1328: PetscFunctionBegin;
1329: switch (op) {
1330: case MAT_ROW_ORIENTED:
1331: a->roworiented = flg;
1332: break;
1333: case MAT_KEEP_NONZERO_PATTERN:
1334: a->keepnonzeropattern = flg;
1335: break;
1336: case MAT_NEW_NONZERO_LOCATIONS:
1337: a->nonew = (flg ? 0 : 1);
1338: break;
1339: case MAT_NEW_NONZERO_LOCATION_ERR:
1340: a->nonew = (flg ? -1 : 0);
1341: break;
1342: case MAT_NEW_NONZERO_ALLOCATION_ERR:
1343: a->nonew = (flg ? -2 : 0);
1344: break;
1345: case MAT_UNUSED_NONZERO_LOCATION_ERR:
1346: a->nounused = (flg ? -1 : 0);
1347: break;
1348: case MAT_IGNORE_ZERO_ENTRIES:
1349: a->ignorezeroentries = flg;
1350: break;
1351: case MAT_SPD:
1352: case MAT_SYMMETRIC:
1353: case MAT_STRUCTURALLY_SYMMETRIC:
1354: case MAT_HERMITIAN:
1355: case MAT_SYMMETRY_ETERNAL:
1356: case MAT_STRUCTURE_ONLY:
1357: case MAT_STRUCTURAL_SYMMETRY_ETERNAL:
1358: case MAT_SPD_ETERNAL:
1359: /* if the diagonal matrix is square it inherits some of the properties above */
1360: break;
1361: case MAT_FORCE_DIAGONAL_ENTRIES:
1362: case MAT_IGNORE_OFF_PROC_ENTRIES:
1363: case MAT_USE_HASH_TABLE:
1364: PetscCall(PetscInfo(A, "Option %s ignored\n", MatOptions[op]));
1365: break;
1366: case MAT_USE_INODES:
1367: PetscCall(MatSetOption_SeqAIJ_Inode(A, MAT_USE_INODES, flg));
1368: break;
1369: case MAT_SUBMAT_SINGLEIS:
1370: A->submat_singleis = flg;
1371: break;
1372: case MAT_SORTED_FULL:
1373: if (flg) A->ops->setvalues = MatSetValues_SeqAIJ_SortedFull;
1374: else A->ops->setvalues = MatSetValues_SeqAIJ;
1375: break;
1376: case MAT_FORM_EXPLICIT_TRANSPOSE:
1377: A->form_explicit_transpose = flg;
1378: break;
1379: default:
1380: SETERRQ(PETSC_COMM_SELF, PETSC_ERR_SUP, "unknown option %d", op);
1381: }
1382: PetscFunctionReturn(PETSC_SUCCESS);
1383: }
1385: static PetscErrorCode MatGetDiagonal_SeqAIJ(Mat A, Vec v)
1386: {
1387: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
1388: PetscInt i, j, n, *ai = a->i, *aj = a->j;
1389: PetscScalar *x;
1390: const PetscScalar *aa;
1392: PetscFunctionBegin;
1393: PetscCall(VecGetLocalSize(v, &n));
1394: PetscCheck(n == A->rmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Nonconforming matrix and vector");
1395: PetscCall(MatSeqAIJGetArrayRead(A, &aa));
1396: if (A->factortype == MAT_FACTOR_ILU || A->factortype == MAT_FACTOR_LU) {
1397: PetscInt *diag = a->diag;
1398: PetscCall(VecGetArrayWrite(v, &x));
1399: for (i = 0; i < n; i++) x[i] = 1.0 / aa[diag[i]];
1400: PetscCall(VecRestoreArrayWrite(v, &x));
1401: PetscCall(MatSeqAIJRestoreArrayRead(A, &aa));
1402: PetscFunctionReturn(PETSC_SUCCESS);
1403: }
1405: PetscCall(VecGetArrayWrite(v, &x));
1406: for (i = 0; i < n; i++) {
1407: x[i] = 0.0;
1408: for (j = ai[i]; j < ai[i + 1]; j++) {
1409: if (aj[j] == i) {
1410: x[i] = aa[j];
1411: break;
1412: }
1413: }
1414: }
1415: PetscCall(VecRestoreArrayWrite(v, &x));
1416: PetscCall(MatSeqAIJRestoreArrayRead(A, &aa));
1417: PetscFunctionReturn(PETSC_SUCCESS);
1418: }
1420: #include <../src/mat/impls/aij/seq/ftn-kernels/fmult.h>
1421: PetscErrorCode MatMultTransposeAdd_SeqAIJ(Mat A, Vec xx, Vec zz, Vec yy)
1422: {
1423: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
1424: const MatScalar *aa;
1425: PetscScalar *y;
1426: const PetscScalar *x;
1427: PetscInt m = A->rmap->n;
1428: #if !defined(PETSC_USE_FORTRAN_KERNEL_MULTTRANSPOSEAIJ)
1429: const MatScalar *v;
1430: PetscScalar alpha;
1431: PetscInt n, i, j;
1432: const PetscInt *idx, *ii, *ridx = NULL;
1433: Mat_CompressedRow cprow = a->compressedrow;
1434: PetscBool usecprow = cprow.use;
1435: #endif
1437: PetscFunctionBegin;
1438: if (zz != yy) PetscCall(VecCopy(zz, yy));
1439: PetscCall(VecGetArrayRead(xx, &x));
1440: PetscCall(VecGetArray(yy, &y));
1441: PetscCall(MatSeqAIJGetArrayRead(A, &aa));
1443: #if defined(PETSC_USE_FORTRAN_KERNEL_MULTTRANSPOSEAIJ)
1444: fortranmulttransposeaddaij_(&m, x, a->i, a->j, aa, y);
1445: #else
1446: if (usecprow) {
1447: m = cprow.nrows;
1448: ii = cprow.i;
1449: ridx = cprow.rindex;
1450: } else {
1451: ii = a->i;
1452: }
1453: for (i = 0; i < m; i++) {
1454: idx = a->j + ii[i];
1455: v = aa + ii[i];
1456: n = ii[i + 1] - ii[i];
1457: if (usecprow) {
1458: alpha = x[ridx[i]];
1459: } else {
1460: alpha = x[i];
1461: }
1462: for (j = 0; j < n; j++) y[idx[j]] += alpha * v[j];
1463: }
1464: #endif
1465: PetscCall(PetscLogFlops(2.0 * a->nz));
1466: PetscCall(VecRestoreArrayRead(xx, &x));
1467: PetscCall(VecRestoreArray(yy, &y));
1468: PetscCall(MatSeqAIJRestoreArrayRead(A, &aa));
1469: PetscFunctionReturn(PETSC_SUCCESS);
1470: }
1472: PetscErrorCode MatMultTranspose_SeqAIJ(Mat A, Vec xx, Vec yy)
1473: {
1474: PetscFunctionBegin;
1475: PetscCall(VecSet(yy, 0.0));
1476: PetscCall(MatMultTransposeAdd_SeqAIJ(A, xx, yy, yy));
1477: PetscFunctionReturn(PETSC_SUCCESS);
1478: }
1480: #include <../src/mat/impls/aij/seq/ftn-kernels/fmult.h>
1482: PetscErrorCode MatMult_SeqAIJ(Mat A, Vec xx, Vec yy)
1483: {
1484: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
1485: PetscScalar *y;
1486: const PetscScalar *x;
1487: const MatScalar *a_a;
1488: PetscInt m = A->rmap->n;
1489: const PetscInt *ii, *ridx = NULL;
1490: PetscBool usecprow = a->compressedrow.use;
1492: #if defined(PETSC_HAVE_PRAGMA_DISJOINT)
1493: #pragma disjoint(*x, *y, *aa)
1494: #endif
1496: PetscFunctionBegin;
1497: if (a->inode.use && a->inode.checked) {
1498: PetscCall(MatMult_SeqAIJ_Inode(A, xx, yy));
1499: PetscFunctionReturn(PETSC_SUCCESS);
1500: }
1501: PetscCall(MatSeqAIJGetArrayRead(A, &a_a));
1502: PetscCall(VecGetArrayRead(xx, &x));
1503: PetscCall(VecGetArray(yy, &y));
1504: ii = a->i;
1505: if (usecprow) { /* use compressed row format */
1506: PetscCall(PetscArrayzero(y, m));
1507: m = a->compressedrow.nrows;
1508: ii = a->compressedrow.i;
1509: ridx = a->compressedrow.rindex;
1510: PetscPragmaUseOMPKernels(parallel for)
1511: for (PetscInt i = 0; i < m; i++) {
1512: PetscInt n = ii[i + 1] - ii[i];
1513: const PetscInt *aj = a->j + ii[i];
1514: const PetscScalar *aa = a_a + ii[i];
1515: PetscScalar sum = 0.0;
1516: PetscSparseDensePlusDot(sum, x, aa, aj, n);
1517: /* for (j=0; j<n; j++) sum += (*aa++)*x[*aj++]; */
1518: y[*ridx++] = sum;
1519: }
1520: } else { /* do not use compressed row format */
1521: #if defined(PETSC_USE_FORTRAN_KERNEL_MULTAIJ)
1522: fortranmultaij_(&m, x, ii, a->j, a_a, y);
1523: #else
1524: PetscPragmaUseOMPKernels(parallel for)
1525: for (PetscInt i = 0; i < m; i++) {
1526: PetscInt n = ii[i + 1] - ii[i];
1527: const PetscInt *aj = a->j + ii[i];
1528: const PetscScalar *aa = a_a + ii[i];
1529: PetscScalar sum = 0.0;
1530: PetscSparseDensePlusDot(sum, x, aa, aj, n);
1531: y[i] = sum;
1532: }
1533: #endif
1534: }
1535: PetscCall(PetscLogFlops(2.0 * a->nz - a->nonzerorowcnt));
1536: PetscCall(VecRestoreArrayRead(xx, &x));
1537: PetscCall(VecRestoreArray(yy, &y));
1538: PetscCall(MatSeqAIJRestoreArrayRead(A, &a_a));
1539: PetscFunctionReturn(PETSC_SUCCESS);
1540: }
1542: // HACK!!!!! Used by src/mat/tests/ex170.c
1543: PETSC_EXTERN PetscErrorCode MatMultMax_SeqAIJ(Mat A, Vec xx, Vec yy)
1544: {
1545: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
1546: PetscScalar *y;
1547: const PetscScalar *x;
1548: const MatScalar *aa, *a_a;
1549: PetscInt m = A->rmap->n;
1550: const PetscInt *aj, *ii, *ridx = NULL;
1551: PetscInt n, i, nonzerorow = 0;
1552: PetscScalar sum;
1553: PetscBool usecprow = a->compressedrow.use;
1555: #if defined(PETSC_HAVE_PRAGMA_DISJOINT)
1556: #pragma disjoint(*x, *y, *aa)
1557: #endif
1559: PetscFunctionBegin;
1560: PetscCall(MatSeqAIJGetArrayRead(A, &a_a));
1561: PetscCall(VecGetArrayRead(xx, &x));
1562: PetscCall(VecGetArray(yy, &y));
1563: if (usecprow) { /* use compressed row format */
1564: m = a->compressedrow.nrows;
1565: ii = a->compressedrow.i;
1566: ridx = a->compressedrow.rindex;
1567: for (i = 0; i < m; i++) {
1568: n = ii[i + 1] - ii[i];
1569: aj = a->j + ii[i];
1570: aa = a_a + ii[i];
1571: sum = 0.0;
1572: nonzerorow += (n > 0);
1573: PetscSparseDenseMaxDot(sum, x, aa, aj, n);
1574: /* for (j=0; j<n; j++) sum += (*aa++)*x[*aj++]; */
1575: y[*ridx++] = sum;
1576: }
1577: } else { /* do not use compressed row format */
1578: ii = a->i;
1579: for (i = 0; i < m; i++) {
1580: n = ii[i + 1] - ii[i];
1581: aj = a->j + ii[i];
1582: aa = a_a + ii[i];
1583: sum = 0.0;
1584: nonzerorow += (n > 0);
1585: PetscSparseDenseMaxDot(sum, x, aa, aj, n);
1586: y[i] = sum;
1587: }
1588: }
1589: PetscCall(PetscLogFlops(2.0 * a->nz - nonzerorow));
1590: PetscCall(VecRestoreArrayRead(xx, &x));
1591: PetscCall(VecRestoreArray(yy, &y));
1592: PetscCall(MatSeqAIJRestoreArrayRead(A, &a_a));
1593: PetscFunctionReturn(PETSC_SUCCESS);
1594: }
1596: // HACK!!!!! Used by src/mat/tests/ex170.c
1597: PETSC_EXTERN PetscErrorCode MatMultAddMax_SeqAIJ(Mat A, Vec xx, Vec yy, Vec zz)
1598: {
1599: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
1600: PetscScalar *y, *z;
1601: const PetscScalar *x;
1602: const MatScalar *aa, *a_a;
1603: PetscInt m = A->rmap->n, *aj, *ii;
1604: PetscInt n, i, *ridx = NULL;
1605: PetscScalar sum;
1606: PetscBool usecprow = a->compressedrow.use;
1608: PetscFunctionBegin;
1609: PetscCall(MatSeqAIJGetArrayRead(A, &a_a));
1610: PetscCall(VecGetArrayRead(xx, &x));
1611: PetscCall(VecGetArrayPair(yy, zz, &y, &z));
1612: if (usecprow) { /* use compressed row format */
1613: if (zz != yy) PetscCall(PetscArraycpy(z, y, m));
1614: m = a->compressedrow.nrows;
1615: ii = a->compressedrow.i;
1616: ridx = a->compressedrow.rindex;
1617: for (i = 0; i < m; i++) {
1618: n = ii[i + 1] - ii[i];
1619: aj = a->j + ii[i];
1620: aa = a_a + ii[i];
1621: sum = y[*ridx];
1622: PetscSparseDenseMaxDot(sum, x, aa, aj, n);
1623: z[*ridx++] = sum;
1624: }
1625: } else { /* do not use compressed row format */
1626: ii = a->i;
1627: for (i = 0; i < m; i++) {
1628: n = ii[i + 1] - ii[i];
1629: aj = a->j + ii[i];
1630: aa = a_a + ii[i];
1631: sum = y[i];
1632: PetscSparseDenseMaxDot(sum, x, aa, aj, n);
1633: z[i] = sum;
1634: }
1635: }
1636: PetscCall(PetscLogFlops(2.0 * a->nz));
1637: PetscCall(VecRestoreArrayRead(xx, &x));
1638: PetscCall(VecRestoreArrayPair(yy, zz, &y, &z));
1639: PetscCall(MatSeqAIJRestoreArrayRead(A, &a_a));
1640: PetscFunctionReturn(PETSC_SUCCESS);
1641: }
1643: #include <../src/mat/impls/aij/seq/ftn-kernels/fmultadd.h>
1644: PetscErrorCode MatMultAdd_SeqAIJ(Mat A, Vec xx, Vec yy, Vec zz)
1645: {
1646: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
1647: PetscScalar *y, *z;
1648: const PetscScalar *x;
1649: const MatScalar *a_a;
1650: const PetscInt *ii, *ridx = NULL;
1651: PetscInt m = A->rmap->n;
1652: PetscBool usecprow = a->compressedrow.use;
1654: PetscFunctionBegin;
1655: if (a->inode.use && a->inode.checked) {
1656: PetscCall(MatMultAdd_SeqAIJ_Inode(A, xx, yy, zz));
1657: PetscFunctionReturn(PETSC_SUCCESS);
1658: }
1659: PetscCall(MatSeqAIJGetArrayRead(A, &a_a));
1660: PetscCall(VecGetArrayRead(xx, &x));
1661: PetscCall(VecGetArrayPair(yy, zz, &y, &z));
1662: if (usecprow) { /* use compressed row format */
1663: if (zz != yy) PetscCall(PetscArraycpy(z, y, m));
1664: m = a->compressedrow.nrows;
1665: ii = a->compressedrow.i;
1666: ridx = a->compressedrow.rindex;
1667: for (PetscInt i = 0; i < m; i++) {
1668: PetscInt n = ii[i + 1] - ii[i];
1669: const PetscInt *aj = a->j + ii[i];
1670: const PetscScalar *aa = a_a + ii[i];
1671: PetscScalar sum = y[*ridx];
1672: PetscSparseDensePlusDot(sum, x, aa, aj, n);
1673: z[*ridx++] = sum;
1674: }
1675: } else { /* do not use compressed row format */
1676: ii = a->i;
1677: #if defined(PETSC_USE_FORTRAN_KERNEL_MULTADDAIJ)
1678: fortranmultaddaij_(&m, x, ii, a->j, a_a, y, z);
1679: #else
1680: PetscPragmaUseOMPKernels(parallel for)
1681: for (PetscInt i = 0; i < m; i++) {
1682: PetscInt n = ii[i + 1] - ii[i];
1683: const PetscInt *aj = a->j + ii[i];
1684: const PetscScalar *aa = a_a + ii[i];
1685: PetscScalar sum = y[i];
1686: PetscSparseDensePlusDot(sum, x, aa, aj, n);
1687: z[i] = sum;
1688: }
1689: #endif
1690: }
1691: PetscCall(PetscLogFlops(2.0 * a->nz));
1692: PetscCall(VecRestoreArrayRead(xx, &x));
1693: PetscCall(VecRestoreArrayPair(yy, zz, &y, &z));
1694: PetscCall(MatSeqAIJRestoreArrayRead(A, &a_a));
1695: PetscFunctionReturn(PETSC_SUCCESS);
1696: }
1698: /*
1699: Adds diagonal pointers to sparse matrix nonzero structure.
1700: */
1701: PetscErrorCode MatMarkDiagonal_SeqAIJ(Mat A)
1702: {
1703: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
1704: PetscInt i, j, m = A->rmap->n;
1705: PetscBool alreadySet = PETSC_TRUE;
1707: PetscFunctionBegin;
1708: if (!a->diag) {
1709: PetscCall(PetscMalloc1(m, &a->diag));
1710: alreadySet = PETSC_FALSE;
1711: }
1712: for (i = 0; i < A->rmap->n; i++) {
1713: /* If A's diagonal is already correctly set, this fast track enables cheap and repeated MatMarkDiagonal_SeqAIJ() calls */
1714: if (alreadySet) {
1715: PetscInt pos = a->diag[i];
1716: if (pos >= a->i[i] && pos < a->i[i + 1] && a->j[pos] == i) continue;
1717: }
1719: a->diag[i] = a->i[i + 1];
1720: for (j = a->i[i]; j < a->i[i + 1]; j++) {
1721: if (a->j[j] == i) {
1722: a->diag[i] = j;
1723: break;
1724: }
1725: }
1726: }
1727: PetscFunctionReturn(PETSC_SUCCESS);
1728: }
1730: static PetscErrorCode MatShift_SeqAIJ(Mat A, PetscScalar v)
1731: {
1732: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
1733: const PetscInt *diag = (const PetscInt *)a->diag;
1734: const PetscInt *ii = (const PetscInt *)a->i;
1735: PetscInt i, *mdiag = NULL;
1736: PetscInt cnt = 0; /* how many diagonals are missing */
1738: PetscFunctionBegin;
1739: if (!A->preallocated || !a->nz) {
1740: PetscCall(MatSeqAIJSetPreallocation(A, 1, NULL));
1741: PetscCall(MatShift_Basic(A, v));
1742: PetscFunctionReturn(PETSC_SUCCESS);
1743: }
1745: if (a->diagonaldense) {
1746: cnt = 0;
1747: } else {
1748: PetscCall(PetscCalloc1(A->rmap->n, &mdiag));
1749: for (i = 0; i < A->rmap->n; i++) {
1750: if (i < A->cmap->n && diag[i] >= ii[i + 1]) { /* 'out of range' rows never have diagonals */
1751: cnt++;
1752: mdiag[i] = 1;
1753: }
1754: }
1755: }
1756: if (!cnt) {
1757: PetscCall(MatShift_Basic(A, v));
1758: } else {
1759: PetscScalar *olda = a->a; /* preserve pointers to current matrix nonzeros structure and values */
1760: PetscInt *oldj = a->j, *oldi = a->i;
1761: PetscBool free_a = a->free_a, free_ij = a->free_ij;
1762: const PetscScalar *Aa;
1764: PetscCall(MatSeqAIJGetArrayRead(A, &Aa)); // sync the host
1765: PetscCall(MatSeqAIJRestoreArrayRead(A, &Aa));
1767: a->a = NULL;
1768: a->j = NULL;
1769: a->i = NULL;
1770: /* increase the values in imax for each row where a diagonal is being inserted then reallocate the matrix data structures */
1771: for (i = 0; i < PetscMin(A->rmap->n, A->cmap->n); i++) a->imax[i] += mdiag[i];
1772: PetscCall(MatSeqAIJSetPreallocation_SeqAIJ(A, 0, a->imax));
1774: /* copy old values into new matrix data structure */
1775: for (i = 0; i < A->rmap->n; i++) {
1776: PetscCall(MatSetValues(A, 1, &i, a->imax[i] - mdiag[i], &oldj[oldi[i]], &olda[oldi[i]], ADD_VALUES));
1777: if (i < A->cmap->n) PetscCall(MatSetValue(A, i, i, v, ADD_VALUES));
1778: }
1779: PetscCall(MatAssemblyBegin(A, MAT_FINAL_ASSEMBLY));
1780: PetscCall(MatAssemblyEnd(A, MAT_FINAL_ASSEMBLY));
1781: if (free_a) PetscCall(PetscShmgetDeallocateArray((void **)&olda));
1782: if (free_ij) PetscCall(PetscShmgetDeallocateArray((void **)&oldj));
1783: if (free_ij) PetscCall(PetscShmgetDeallocateArray((void **)&oldi));
1784: }
1785: PetscCall(PetscFree(mdiag));
1786: a->diagonaldense = PETSC_TRUE;
1787: PetscFunctionReturn(PETSC_SUCCESS);
1788: }
1790: /*
1791: Checks for missing diagonals
1792: */
1793: PetscErrorCode MatMissingDiagonal_SeqAIJ(Mat A, PetscBool *missing, PetscInt *d)
1794: {
1795: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
1796: PetscInt *diag, *ii = a->i, i;
1798: PetscFunctionBegin;
1799: *missing = PETSC_FALSE;
1800: if (A->rmap->n > 0 && !ii) {
1801: *missing = PETSC_TRUE;
1802: if (d) *d = 0;
1803: PetscCall(PetscInfo(A, "Matrix has no entries therefore is missing diagonal\n"));
1804: } else {
1805: PetscInt n;
1806: n = PetscMin(A->rmap->n, A->cmap->n);
1807: diag = a->diag;
1808: for (i = 0; i < n; i++) {
1809: if (diag[i] >= ii[i + 1]) {
1810: *missing = PETSC_TRUE;
1811: if (d) *d = i;
1812: PetscCall(PetscInfo(A, "Matrix is missing diagonal number %" PetscInt_FMT "\n", i));
1813: break;
1814: }
1815: }
1816: }
1817: PetscFunctionReturn(PETSC_SUCCESS);
1818: }
1820: #include <petscblaslapack.h>
1821: #include <petsc/private/kernels/blockinvert.h>
1823: /*
1824: Note that values is allocated externally by the PC and then passed into this routine
1825: */
1826: static PetscErrorCode MatInvertVariableBlockDiagonal_SeqAIJ(Mat A, PetscInt nblocks, const PetscInt *bsizes, PetscScalar *diag)
1827: {
1828: PetscInt n = A->rmap->n, i, ncnt = 0, *indx, j, bsizemax = 0, *v_pivots;
1829: PetscBool allowzeropivot, zeropivotdetected = PETSC_FALSE;
1830: const PetscReal shift = 0.0;
1831: PetscInt ipvt[5];
1832: PetscCount flops = 0;
1833: PetscScalar work[25], *v_work;
1835: PetscFunctionBegin;
1836: allowzeropivot = PetscNot(A->erroriffailure);
1837: for (i = 0; i < nblocks; i++) ncnt += bsizes[i];
1838: PetscCheck(ncnt == n, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Total blocksizes %" PetscInt_FMT " doesn't match number matrix rows %" PetscInt_FMT, ncnt, n);
1839: for (i = 0; i < nblocks; i++) bsizemax = PetscMax(bsizemax, bsizes[i]);
1840: PetscCall(PetscMalloc1(bsizemax, &indx));
1841: if (bsizemax > 7) PetscCall(PetscMalloc2(bsizemax, &v_work, bsizemax, &v_pivots));
1842: ncnt = 0;
1843: for (i = 0; i < nblocks; i++) {
1844: for (j = 0; j < bsizes[i]; j++) indx[j] = ncnt + j;
1845: PetscCall(MatGetValues(A, bsizes[i], indx, bsizes[i], indx, diag));
1846: switch (bsizes[i]) {
1847: case 1:
1848: *diag = 1.0 / (*diag);
1849: break;
1850: case 2:
1851: PetscCall(PetscKernel_A_gets_inverse_A_2(diag, shift, allowzeropivot, &zeropivotdetected));
1852: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
1853: PetscCall(PetscKernel_A_gets_transpose_A_2(diag));
1854: break;
1855: case 3:
1856: PetscCall(PetscKernel_A_gets_inverse_A_3(diag, shift, allowzeropivot, &zeropivotdetected));
1857: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
1858: PetscCall(PetscKernel_A_gets_transpose_A_3(diag));
1859: break;
1860: case 4:
1861: PetscCall(PetscKernel_A_gets_inverse_A_4(diag, shift, allowzeropivot, &zeropivotdetected));
1862: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
1863: PetscCall(PetscKernel_A_gets_transpose_A_4(diag));
1864: break;
1865: case 5:
1866: PetscCall(PetscKernel_A_gets_inverse_A_5(diag, ipvt, work, shift, allowzeropivot, &zeropivotdetected));
1867: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
1868: PetscCall(PetscKernel_A_gets_transpose_A_5(diag));
1869: break;
1870: case 6:
1871: PetscCall(PetscKernel_A_gets_inverse_A_6(diag, shift, allowzeropivot, &zeropivotdetected));
1872: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
1873: PetscCall(PetscKernel_A_gets_transpose_A_6(diag));
1874: break;
1875: case 7:
1876: PetscCall(PetscKernel_A_gets_inverse_A_7(diag, shift, allowzeropivot, &zeropivotdetected));
1877: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
1878: PetscCall(PetscKernel_A_gets_transpose_A_7(diag));
1879: break;
1880: default:
1881: PetscCall(PetscKernel_A_gets_inverse_A(bsizes[i], diag, v_pivots, v_work, allowzeropivot, &zeropivotdetected));
1882: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
1883: PetscCall(PetscKernel_A_gets_transpose_A_N(diag, bsizes[i]));
1884: }
1885: ncnt += bsizes[i];
1886: diag += bsizes[i] * bsizes[i];
1887: flops += 2 * PetscPowInt64(bsizes[i], 3) / 3;
1888: }
1889: PetscCall(PetscLogFlops(flops));
1890: if (bsizemax > 7) PetscCall(PetscFree2(v_work, v_pivots));
1891: PetscCall(PetscFree(indx));
1892: PetscFunctionReturn(PETSC_SUCCESS);
1893: }
1895: /*
1896: Negative shift indicates do not generate an error if there is a zero diagonal, just invert it anyways
1897: */
1898: static PetscErrorCode MatInvertDiagonal_SeqAIJ(Mat A, PetscScalar omega, PetscScalar fshift)
1899: {
1900: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
1901: PetscInt i, *diag, m = A->rmap->n;
1902: const MatScalar *v;
1903: PetscScalar *idiag, *mdiag;
1905: PetscFunctionBegin;
1906: if (a->idiagvalid) PetscFunctionReturn(PETSC_SUCCESS);
1907: PetscCall(MatMarkDiagonal_SeqAIJ(A));
1908: diag = a->diag;
1909: if (!a->idiag) { PetscCall(PetscMalloc3(m, &a->idiag, m, &a->mdiag, m, &a->ssor_work)); }
1911: mdiag = a->mdiag;
1912: idiag = a->idiag;
1913: PetscCall(MatSeqAIJGetArrayRead(A, &v));
1914: if (omega == 1.0 && PetscRealPart(fshift) <= 0.0) {
1915: for (i = 0; i < m; i++) {
1916: mdiag[i] = v[diag[i]];
1917: if (!PetscAbsScalar(mdiag[i])) { /* zero diagonal */
1918: if (PetscRealPart(fshift)) {
1919: PetscCall(PetscInfo(A, "Zero diagonal on row %" PetscInt_FMT "\n", i));
1920: A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
1921: A->factorerror_zeropivot_value = 0.0;
1922: A->factorerror_zeropivot_row = i;
1923: } else SETERRQ(PETSC_COMM_SELF, PETSC_ERR_ARG_INCOMP, "Zero diagonal on row %" PetscInt_FMT, i);
1924: }
1925: idiag[i] = 1.0 / v[diag[i]];
1926: }
1927: PetscCall(PetscLogFlops(m));
1928: } else {
1929: for (i = 0; i < m; i++) {
1930: mdiag[i] = v[diag[i]];
1931: idiag[i] = omega / (fshift + v[diag[i]]);
1932: }
1933: PetscCall(PetscLogFlops(2.0 * m));
1934: }
1935: a->idiagvalid = PETSC_TRUE;
1936: PetscCall(MatSeqAIJRestoreArrayRead(A, &v));
1937: PetscFunctionReturn(PETSC_SUCCESS);
1938: }
1940: PetscErrorCode MatSOR_SeqAIJ(Mat A, Vec bb, PetscReal omega, MatSORType flag, PetscReal fshift, PetscInt its, PetscInt lits, Vec xx)
1941: {
1942: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
1943: PetscScalar *x, d, sum, *t, scale;
1944: const MatScalar *v, *idiag = NULL, *mdiag, *aa;
1945: const PetscScalar *b, *bs, *xb, *ts;
1946: PetscInt n, m = A->rmap->n, i;
1947: const PetscInt *idx, *diag;
1949: PetscFunctionBegin;
1950: if (a->inode.use && a->inode.checked && omega == 1.0 && fshift == 0.0) {
1951: PetscCall(MatSOR_SeqAIJ_Inode(A, bb, omega, flag, fshift, its, lits, xx));
1952: PetscFunctionReturn(PETSC_SUCCESS);
1953: }
1954: its = its * lits;
1956: if (fshift != a->fshift || omega != a->omega) a->idiagvalid = PETSC_FALSE; /* must recompute idiag[] */
1957: if (!a->idiagvalid) PetscCall(MatInvertDiagonal_SeqAIJ(A, omega, fshift));
1958: a->fshift = fshift;
1959: a->omega = omega;
1961: diag = a->diag;
1962: t = a->ssor_work;
1963: idiag = a->idiag;
1964: mdiag = a->mdiag;
1966: PetscCall(MatSeqAIJGetArrayRead(A, &aa));
1967: PetscCall(VecGetArray(xx, &x));
1968: PetscCall(VecGetArrayRead(bb, &b));
1969: /* We count flops by assuming the upper triangular and lower triangular parts have the same number of nonzeros */
1970: if (flag == SOR_APPLY_UPPER) {
1971: /* apply (U + D/omega) to the vector */
1972: bs = b;
1973: for (i = 0; i < m; i++) {
1974: d = fshift + mdiag[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 = b[i] * d / omega;
1979: PetscSparseDensePlusDot(sum, bs, v, idx, n);
1980: x[i] = sum;
1981: }
1982: PetscCall(VecRestoreArray(xx, &x));
1983: PetscCall(VecRestoreArrayRead(bb, &b));
1984: PetscCall(MatSeqAIJRestoreArrayRead(A, &aa));
1985: PetscCall(PetscLogFlops(a->nz));
1986: PetscFunctionReturn(PETSC_SUCCESS);
1987: }
1989: PetscCheck(flag != SOR_APPLY_LOWER, PETSC_COMM_SELF, PETSC_ERR_SUP, "SOR_APPLY_LOWER is not implemented");
1990: if (flag & SOR_EISENSTAT) {
1991: /* Let A = L + U + D; where L is lower triangular,
1992: U is upper triangular, E = D/omega; This routine applies
1994: (L + E)^{-1} A (U + E)^{-1}
1996: to a vector efficiently using Eisenstat's trick.
1997: */
1998: scale = (2.0 / omega) - 1.0;
2000: /* x = (E + U)^{-1} b */
2001: for (i = m - 1; i >= 0; i--) {
2002: n = a->i[i + 1] - diag[i] - 1;
2003: idx = a->j + diag[i] + 1;
2004: v = aa + diag[i] + 1;
2005: sum = b[i];
2006: PetscSparseDenseMinusDot(sum, x, v, idx, n);
2007: x[i] = sum * idiag[i];
2008: }
2010: /* t = b - (2*E - D)x */
2011: v = aa;
2012: for (i = 0; i < m; i++) t[i] = b[i] - scale * (v[*diag++]) * x[i];
2014: /* t = (E + L)^{-1}t */
2015: ts = t;
2016: diag = a->diag;
2017: for (i = 0; i < m; i++) {
2018: n = diag[i] - a->i[i];
2019: idx = a->j + a->i[i];
2020: v = aa + a->i[i];
2021: sum = t[i];
2022: PetscSparseDenseMinusDot(sum, ts, v, idx, n);
2023: t[i] = sum * idiag[i];
2024: /* x = x + t */
2025: x[i] += t[i];
2026: }
2028: PetscCall(PetscLogFlops(6.0 * m - 1 + 2.0 * a->nz));
2029: PetscCall(VecRestoreArray(xx, &x));
2030: PetscCall(VecRestoreArrayRead(bb, &b));
2031: PetscFunctionReturn(PETSC_SUCCESS);
2032: }
2033: if (flag & SOR_ZERO_INITIAL_GUESS) {
2034: if (flag & SOR_FORWARD_SWEEP || flag & SOR_LOCAL_FORWARD_SWEEP) {
2035: for (i = 0; i < m; i++) {
2036: n = diag[i] - a->i[i];
2037: idx = a->j + a->i[i];
2038: v = aa + a->i[i];
2039: sum = b[i];
2040: PetscSparseDenseMinusDot(sum, x, v, idx, n);
2041: t[i] = sum;
2042: x[i] = sum * idiag[i];
2043: }
2044: xb = t;
2045: PetscCall(PetscLogFlops(a->nz));
2046: } else xb = b;
2047: if (flag & SOR_BACKWARD_SWEEP || flag & SOR_LOCAL_BACKWARD_SWEEP) {
2048: for (i = m - 1; i >= 0; i--) {
2049: n = a->i[i + 1] - diag[i] - 1;
2050: idx = a->j + diag[i] + 1;
2051: v = aa + diag[i] + 1;
2052: sum = xb[i];
2053: PetscSparseDenseMinusDot(sum, x, v, idx, n);
2054: if (xb == b) {
2055: x[i] = sum * idiag[i];
2056: } else {
2057: x[i] = (1 - omega) * x[i] + sum * idiag[i]; /* omega in idiag */
2058: }
2059: }
2060: PetscCall(PetscLogFlops(a->nz)); /* assumes 1/2 in upper */
2061: }
2062: its--;
2063: }
2064: while (its--) {
2065: if (flag & SOR_FORWARD_SWEEP || flag & SOR_LOCAL_FORWARD_SWEEP) {
2066: for (i = 0; i < m; i++) {
2067: /* lower */
2068: n = diag[i] - a->i[i];
2069: idx = a->j + a->i[i];
2070: v = aa + a->i[i];
2071: sum = b[i];
2072: PetscSparseDenseMinusDot(sum, x, v, idx, n);
2073: t[i] = sum; /* save application of the lower-triangular part */
2074: /* upper */
2075: n = a->i[i + 1] - diag[i] - 1;
2076: idx = a->j + diag[i] + 1;
2077: v = aa + diag[i] + 1;
2078: PetscSparseDenseMinusDot(sum, x, v, idx, n);
2079: x[i] = (1. - omega) * x[i] + sum * idiag[i]; /* omega in idiag */
2080: }
2081: xb = t;
2082: PetscCall(PetscLogFlops(2.0 * a->nz));
2083: } else xb = b;
2084: if (flag & SOR_BACKWARD_SWEEP || flag & SOR_LOCAL_BACKWARD_SWEEP) {
2085: for (i = m - 1; i >= 0; i--) {
2086: sum = xb[i];
2087: if (xb == b) {
2088: /* whole matrix (no checkpointing available) */
2089: n = a->i[i + 1] - a->i[i];
2090: idx = a->j + a->i[i];
2091: v = aa + a->i[i];
2092: PetscSparseDenseMinusDot(sum, x, v, idx, n);
2093: x[i] = (1. - omega) * x[i] + (sum + mdiag[i] * x[i]) * idiag[i];
2094: } else { /* lower-triangular part has been saved, so only apply upper-triangular */
2095: n = a->i[i + 1] - diag[i] - 1;
2096: idx = a->j + diag[i] + 1;
2097: v = aa + diag[i] + 1;
2098: PetscSparseDenseMinusDot(sum, x, v, idx, n);
2099: x[i] = (1. - omega) * x[i] + sum * idiag[i]; /* omega in idiag */
2100: }
2101: }
2102: if (xb == b) {
2103: PetscCall(PetscLogFlops(2.0 * a->nz));
2104: } else {
2105: PetscCall(PetscLogFlops(a->nz)); /* assumes 1/2 in upper */
2106: }
2107: }
2108: }
2109: PetscCall(MatSeqAIJRestoreArrayRead(A, &aa));
2110: PetscCall(VecRestoreArray(xx, &x));
2111: PetscCall(VecRestoreArrayRead(bb, &b));
2112: PetscFunctionReturn(PETSC_SUCCESS);
2113: }
2115: static PetscErrorCode MatGetInfo_SeqAIJ(Mat A, MatInfoType flag, MatInfo *info)
2116: {
2117: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
2119: PetscFunctionBegin;
2120: info->block_size = 1.0;
2121: info->nz_allocated = a->maxnz;
2122: info->nz_used = a->nz;
2123: info->nz_unneeded = (a->maxnz - a->nz);
2124: info->assemblies = A->num_ass;
2125: info->mallocs = A->info.mallocs;
2126: info->memory = 0; /* REVIEW ME */
2127: if (A->factortype) {
2128: info->fill_ratio_given = A->info.fill_ratio_given;
2129: info->fill_ratio_needed = A->info.fill_ratio_needed;
2130: info->factor_mallocs = A->info.factor_mallocs;
2131: } else {
2132: info->fill_ratio_given = 0;
2133: info->fill_ratio_needed = 0;
2134: info->factor_mallocs = 0;
2135: }
2136: PetscFunctionReturn(PETSC_SUCCESS);
2137: }
2139: static PetscErrorCode MatZeroRows_SeqAIJ(Mat A, PetscInt N, const PetscInt rows[], PetscScalar diag, Vec x, Vec b)
2140: {
2141: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
2142: PetscInt i, m = A->rmap->n - 1;
2143: const PetscScalar *xx;
2144: PetscScalar *bb, *aa;
2145: PetscInt d = 0;
2147: PetscFunctionBegin;
2148: if (x && b) {
2149: PetscCall(VecGetArrayRead(x, &xx));
2150: PetscCall(VecGetArray(b, &bb));
2151: for (i = 0; i < N; i++) {
2152: PetscCheck(rows[i] >= 0 && rows[i] <= m, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "row %" PetscInt_FMT " out of range", rows[i]);
2153: if (rows[i] >= A->cmap->n) continue;
2154: bb[rows[i]] = diag * xx[rows[i]];
2155: }
2156: PetscCall(VecRestoreArrayRead(x, &xx));
2157: PetscCall(VecRestoreArray(b, &bb));
2158: }
2160: PetscCall(MatSeqAIJGetArray(A, &aa));
2161: if (a->keepnonzeropattern) {
2162: for (i = 0; i < N; i++) {
2163: PetscCheck(rows[i] >= 0 && rows[i] <= m, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "row %" PetscInt_FMT " out of range", rows[i]);
2164: PetscCall(PetscArrayzero(&aa[a->i[rows[i]]], a->ilen[rows[i]]));
2165: }
2166: if (diag != 0.0) {
2167: for (i = 0; i < N; i++) {
2168: d = rows[i];
2169: if (rows[i] >= A->cmap->n) continue;
2170: PetscCheck(a->diag[d] < a->i[d + 1], PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Matrix is missing diagonal entry in the zeroed row %" PetscInt_FMT, d);
2171: }
2172: for (i = 0; i < N; i++) {
2173: if (rows[i] >= A->cmap->n) continue;
2174: aa[a->diag[rows[i]]] = diag;
2175: }
2176: }
2177: } else {
2178: if (diag != 0.0) {
2179: for (i = 0; i < N; i++) {
2180: PetscCheck(rows[i] >= 0 && rows[i] <= m, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "row %" PetscInt_FMT " out of range", rows[i]);
2181: if (a->ilen[rows[i]] > 0) {
2182: if (rows[i] >= A->cmap->n) {
2183: a->ilen[rows[i]] = 0;
2184: } else {
2185: a->ilen[rows[i]] = 1;
2186: aa[a->i[rows[i]]] = diag;
2187: a->j[a->i[rows[i]]] = rows[i];
2188: }
2189: } else if (rows[i] < A->cmap->n) { /* in case row was completely empty */
2190: PetscCall(MatSetValues_SeqAIJ(A, 1, &rows[i], 1, &rows[i], &diag, INSERT_VALUES));
2191: }
2192: }
2193: } else {
2194: for (i = 0; i < N; i++) {
2195: PetscCheck(rows[i] >= 0 && rows[i] <= m, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "row %" PetscInt_FMT " out of range", rows[i]);
2196: a->ilen[rows[i]] = 0;
2197: }
2198: }
2199: A->nonzerostate++;
2200: }
2201: PetscCall(MatSeqAIJRestoreArray(A, &aa));
2202: PetscUseTypeMethod(A, assemblyend, MAT_FINAL_ASSEMBLY);
2203: PetscFunctionReturn(PETSC_SUCCESS);
2204: }
2206: static PetscErrorCode MatZeroRowsColumns_SeqAIJ(Mat A, PetscInt N, const PetscInt rows[], PetscScalar diag, Vec x, Vec b)
2207: {
2208: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
2209: PetscInt i, j, m = A->rmap->n - 1, d = 0;
2210: PetscBool missing, *zeroed, vecs = PETSC_FALSE;
2211: const PetscScalar *xx;
2212: PetscScalar *bb, *aa;
2214: PetscFunctionBegin;
2215: if (!N) PetscFunctionReturn(PETSC_SUCCESS);
2216: PetscCall(MatSeqAIJGetArray(A, &aa));
2217: if (x && b) {
2218: PetscCall(VecGetArrayRead(x, &xx));
2219: PetscCall(VecGetArray(b, &bb));
2220: vecs = PETSC_TRUE;
2221: }
2222: PetscCall(PetscCalloc1(A->rmap->n, &zeroed));
2223: for (i = 0; i < N; i++) {
2224: PetscCheck(rows[i] >= 0 && rows[i] <= m, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "row %" PetscInt_FMT " out of range", rows[i]);
2225: PetscCall(PetscArrayzero(PetscSafePointerPlusOffset(aa, a->i[rows[i]]), a->ilen[rows[i]]));
2227: zeroed[rows[i]] = PETSC_TRUE;
2228: }
2229: for (i = 0; i < A->rmap->n; i++) {
2230: if (!zeroed[i]) {
2231: for (j = a->i[i]; j < a->i[i + 1]; j++) {
2232: if (a->j[j] < A->rmap->n && zeroed[a->j[j]]) {
2233: if (vecs) bb[i] -= aa[j] * xx[a->j[j]];
2234: aa[j] = 0.0;
2235: }
2236: }
2237: } else if (vecs && i < A->cmap->N) bb[i] = diag * xx[i];
2238: }
2239: if (x && b) {
2240: PetscCall(VecRestoreArrayRead(x, &xx));
2241: PetscCall(VecRestoreArray(b, &bb));
2242: }
2243: PetscCall(PetscFree(zeroed));
2244: if (diag != 0.0) {
2245: PetscCall(MatMissingDiagonal_SeqAIJ(A, &missing, &d));
2246: if (missing) {
2247: for (i = 0; i < N; i++) {
2248: if (rows[i] >= A->cmap->N) continue;
2249: PetscCheck(!a->nonew || rows[i] < d, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Matrix is missing diagonal entry in row %" PetscInt_FMT " (%" PetscInt_FMT ")", d, rows[i]);
2250: PetscCall(MatSetValues_SeqAIJ(A, 1, &rows[i], 1, &rows[i], &diag, INSERT_VALUES));
2251: }
2252: } else {
2253: for (i = 0; i < N; i++) aa[a->diag[rows[i]]] = diag;
2254: }
2255: }
2256: PetscCall(MatSeqAIJRestoreArray(A, &aa));
2257: PetscUseTypeMethod(A, assemblyend, MAT_FINAL_ASSEMBLY);
2258: PetscFunctionReturn(PETSC_SUCCESS);
2259: }
2261: PetscErrorCode MatGetRow_SeqAIJ(Mat A, PetscInt row, PetscInt *nz, PetscInt **idx, PetscScalar **v)
2262: {
2263: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
2264: const PetscScalar *aa;
2266: PetscFunctionBegin;
2267: PetscCall(MatSeqAIJGetArrayRead(A, &aa));
2268: *nz = a->i[row + 1] - a->i[row];
2269: if (v) *v = PetscSafePointerPlusOffset((PetscScalar *)aa, a->i[row]);
2270: if (idx) {
2271: if (*nz && a->j) *idx = a->j + a->i[row];
2272: else *idx = NULL;
2273: }
2274: PetscCall(MatSeqAIJRestoreArrayRead(A, &aa));
2275: PetscFunctionReturn(PETSC_SUCCESS);
2276: }
2278: PetscErrorCode MatRestoreRow_SeqAIJ(Mat A, PetscInt row, PetscInt *nz, PetscInt **idx, PetscScalar **v)
2279: {
2280: PetscFunctionBegin;
2281: PetscFunctionReturn(PETSC_SUCCESS);
2282: }
2284: static PetscErrorCode MatNorm_SeqAIJ(Mat A, NormType type, PetscReal *nrm)
2285: {
2286: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
2287: const MatScalar *v;
2288: PetscReal sum = 0.0;
2289: PetscInt i, j;
2291: PetscFunctionBegin;
2292: PetscCall(MatSeqAIJGetArrayRead(A, &v));
2293: if (type == NORM_FROBENIUS) {
2294: #if defined(PETSC_USE_REAL___FP16)
2295: PetscBLASInt one = 1, nz = a->nz;
2296: PetscCallBLAS("BLASnrm2", *nrm = BLASnrm2_(&nz, v, &one));
2297: #else
2298: for (i = 0; i < a->nz; i++) {
2299: sum += PetscRealPart(PetscConj(*v) * (*v));
2300: v++;
2301: }
2302: *nrm = PetscSqrtReal(sum);
2303: #endif
2304: PetscCall(PetscLogFlops(2.0 * a->nz));
2305: } else if (type == NORM_1) {
2306: PetscReal *tmp;
2307: PetscInt *jj = a->j;
2308: PetscCall(PetscCalloc1(A->cmap->n + 1, &tmp));
2309: *nrm = 0.0;
2310: for (j = 0; j < a->nz; j++) {
2311: tmp[*jj++] += PetscAbsScalar(*v);
2312: v++;
2313: }
2314: for (j = 0; j < A->cmap->n; j++) {
2315: if (tmp[j] > *nrm) *nrm = tmp[j];
2316: }
2317: PetscCall(PetscFree(tmp));
2318: PetscCall(PetscLogFlops(PetscMax(a->nz - 1, 0)));
2319: } else if (type == NORM_INFINITY) {
2320: *nrm = 0.0;
2321: for (j = 0; j < A->rmap->n; j++) {
2322: const PetscScalar *v2 = PetscSafePointerPlusOffset(v, a->i[j]);
2323: sum = 0.0;
2324: for (i = 0; i < a->i[j + 1] - a->i[j]; i++) {
2325: sum += PetscAbsScalar(*v2);
2326: v2++;
2327: }
2328: if (sum > *nrm) *nrm = sum;
2329: }
2330: PetscCall(PetscLogFlops(PetscMax(a->nz - 1, 0)));
2331: } else SETERRQ(PETSC_COMM_SELF, PETSC_ERR_SUP, "No support for two norm");
2332: PetscCall(MatSeqAIJRestoreArrayRead(A, &v));
2333: PetscFunctionReturn(PETSC_SUCCESS);
2334: }
2336: static PetscErrorCode MatIsTranspose_SeqAIJ(Mat A, Mat B, PetscReal tol, PetscBool *f)
2337: {
2338: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)A->data, *bij = (Mat_SeqAIJ *)B->data;
2339: PetscInt *adx, *bdx, *aii, *bii, *aptr, *bptr;
2340: const MatScalar *va, *vb;
2341: PetscInt ma, na, mb, nb, i;
2343: PetscFunctionBegin;
2344: PetscCall(MatGetSize(A, &ma, &na));
2345: PetscCall(MatGetSize(B, &mb, &nb));
2346: if (ma != nb || na != mb) {
2347: *f = PETSC_FALSE;
2348: PetscFunctionReturn(PETSC_SUCCESS);
2349: }
2350: PetscCall(MatSeqAIJGetArrayRead(A, &va));
2351: PetscCall(MatSeqAIJGetArrayRead(B, &vb));
2352: aii = aij->i;
2353: bii = bij->i;
2354: adx = aij->j;
2355: bdx = bij->j;
2356: PetscCall(PetscMalloc1(ma, &aptr));
2357: PetscCall(PetscMalloc1(mb, &bptr));
2358: for (i = 0; i < ma; i++) aptr[i] = aii[i];
2359: for (i = 0; i < mb; i++) bptr[i] = bii[i];
2361: *f = PETSC_TRUE;
2362: for (i = 0; i < ma; i++) {
2363: while (aptr[i] < aii[i + 1]) {
2364: PetscInt idc, idr;
2365: PetscScalar vc, vr;
2366: /* column/row index/value */
2367: idc = adx[aptr[i]];
2368: idr = bdx[bptr[idc]];
2369: vc = va[aptr[i]];
2370: vr = vb[bptr[idc]];
2371: if (i != idr || PetscAbsScalar(vc - vr) > tol) {
2372: *f = PETSC_FALSE;
2373: goto done;
2374: } else {
2375: aptr[i]++;
2376: if (B || i != idc) bptr[idc]++;
2377: }
2378: }
2379: }
2380: done:
2381: PetscCall(PetscFree(aptr));
2382: PetscCall(PetscFree(bptr));
2383: PetscCall(MatSeqAIJRestoreArrayRead(A, &va));
2384: PetscCall(MatSeqAIJRestoreArrayRead(B, &vb));
2385: PetscFunctionReturn(PETSC_SUCCESS);
2386: }
2388: static PetscErrorCode MatIsHermitianTranspose_SeqAIJ(Mat A, Mat B, PetscReal tol, PetscBool *f)
2389: {
2390: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)A->data, *bij = (Mat_SeqAIJ *)B->data;
2391: PetscInt *adx, *bdx, *aii, *bii, *aptr, *bptr;
2392: MatScalar *va, *vb;
2393: PetscInt ma, na, mb, nb, i;
2395: PetscFunctionBegin;
2396: PetscCall(MatGetSize(A, &ma, &na));
2397: PetscCall(MatGetSize(B, &mb, &nb));
2398: if (ma != nb || na != mb) {
2399: *f = PETSC_FALSE;
2400: PetscFunctionReturn(PETSC_SUCCESS);
2401: }
2402: aii = aij->i;
2403: bii = bij->i;
2404: adx = aij->j;
2405: bdx = bij->j;
2406: va = aij->a;
2407: vb = bij->a;
2408: PetscCall(PetscMalloc1(ma, &aptr));
2409: PetscCall(PetscMalloc1(mb, &bptr));
2410: for (i = 0; i < ma; i++) aptr[i] = aii[i];
2411: for (i = 0; i < mb; i++) bptr[i] = bii[i];
2413: *f = PETSC_TRUE;
2414: for (i = 0; i < ma; i++) {
2415: while (aptr[i] < aii[i + 1]) {
2416: PetscInt idc, idr;
2417: PetscScalar vc, vr;
2418: /* column/row index/value */
2419: idc = adx[aptr[i]];
2420: idr = bdx[bptr[idc]];
2421: vc = va[aptr[i]];
2422: vr = vb[bptr[idc]];
2423: if (i != idr || PetscAbsScalar(vc - PetscConj(vr)) > tol) {
2424: *f = PETSC_FALSE;
2425: goto done;
2426: } else {
2427: aptr[i]++;
2428: if (B || i != idc) bptr[idc]++;
2429: }
2430: }
2431: }
2432: done:
2433: PetscCall(PetscFree(aptr));
2434: PetscCall(PetscFree(bptr));
2435: PetscFunctionReturn(PETSC_SUCCESS);
2436: }
2438: PetscErrorCode MatDiagonalScale_SeqAIJ(Mat A, Vec ll, Vec rr)
2439: {
2440: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
2441: const PetscScalar *l, *r;
2442: PetscScalar x;
2443: MatScalar *v;
2444: PetscInt i, j, m = A->rmap->n, n = A->cmap->n, M, nz = a->nz;
2445: const PetscInt *jj;
2447: PetscFunctionBegin;
2448: if (ll) {
2449: /* The local size is used so that VecMPI can be passed to this routine
2450: by MatDiagonalScale_MPIAIJ */
2451: PetscCall(VecGetLocalSize(ll, &m));
2452: PetscCheck(m == A->rmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Left scaling vector wrong length");
2453: PetscCall(VecGetArrayRead(ll, &l));
2454: PetscCall(MatSeqAIJGetArray(A, &v));
2455: for (i = 0; i < m; i++) {
2456: x = l[i];
2457: M = a->i[i + 1] - a->i[i];
2458: for (j = 0; j < M; j++) (*v++) *= x;
2459: }
2460: PetscCall(VecRestoreArrayRead(ll, &l));
2461: PetscCall(PetscLogFlops(nz));
2462: PetscCall(MatSeqAIJRestoreArray(A, &v));
2463: }
2464: if (rr) {
2465: PetscCall(VecGetLocalSize(rr, &n));
2466: PetscCheck(n == A->cmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Right scaling vector wrong length");
2467: PetscCall(VecGetArrayRead(rr, &r));
2468: PetscCall(MatSeqAIJGetArray(A, &v));
2469: jj = a->j;
2470: for (i = 0; i < nz; i++) (*v++) *= r[*jj++];
2471: PetscCall(MatSeqAIJRestoreArray(A, &v));
2472: PetscCall(VecRestoreArrayRead(rr, &r));
2473: PetscCall(PetscLogFlops(nz));
2474: }
2475: PetscCall(MatSeqAIJInvalidateDiagonal(A));
2476: PetscFunctionReturn(PETSC_SUCCESS);
2477: }
2479: PetscErrorCode MatCreateSubMatrix_SeqAIJ(Mat A, IS isrow, IS iscol, PetscInt csize, MatReuse scall, Mat *B)
2480: {
2481: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data, *c;
2482: PetscInt *smap, i, k, kstart, kend, oldcols = A->cmap->n, *lens;
2483: PetscInt row, mat_i, *mat_j, tcol, first, step, *mat_ilen, sum, lensi;
2484: const PetscInt *irow, *icol;
2485: const PetscScalar *aa;
2486: PetscInt nrows, ncols;
2487: PetscInt *starts, *j_new, *i_new, *aj = a->j, *ai = a->i, ii, *ailen = a->ilen;
2488: MatScalar *a_new, *mat_a, *c_a;
2489: Mat C;
2490: PetscBool stride;
2492: PetscFunctionBegin;
2493: PetscCall(ISGetIndices(isrow, &irow));
2494: PetscCall(ISGetLocalSize(isrow, &nrows));
2495: PetscCall(ISGetLocalSize(iscol, &ncols));
2497: PetscCall(PetscObjectTypeCompare((PetscObject)iscol, ISSTRIDE, &stride));
2498: if (stride) {
2499: PetscCall(ISStrideGetInfo(iscol, &first, &step));
2500: } else {
2501: first = 0;
2502: step = 0;
2503: }
2504: if (stride && step == 1) {
2505: /* special case of contiguous rows */
2506: PetscCall(PetscMalloc2(nrows, &lens, nrows, &starts));
2507: /* loop over new rows determining lens and starting points */
2508: for (i = 0; i < nrows; i++) {
2509: kstart = ai[irow[i]];
2510: kend = kstart + ailen[irow[i]];
2511: starts[i] = kstart;
2512: for (k = kstart; k < kend; k++) {
2513: if (aj[k] >= first) {
2514: starts[i] = k;
2515: break;
2516: }
2517: }
2518: sum = 0;
2519: while (k < kend) {
2520: if (aj[k++] >= first + ncols) break;
2521: sum++;
2522: }
2523: lens[i] = sum;
2524: }
2525: /* create submatrix */
2526: if (scall == MAT_REUSE_MATRIX) {
2527: PetscInt n_cols, n_rows;
2528: PetscCall(MatGetSize(*B, &n_rows, &n_cols));
2529: PetscCheck(n_rows == nrows && n_cols == ncols, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Reused submatrix wrong size");
2530: PetscCall(MatZeroEntries(*B));
2531: C = *B;
2532: } else {
2533: PetscInt rbs, cbs;
2534: PetscCall(MatCreate(PetscObjectComm((PetscObject)A), &C));
2535: PetscCall(MatSetSizes(C, nrows, ncols, PETSC_DETERMINE, PETSC_DETERMINE));
2536: PetscCall(ISGetBlockSize(isrow, &rbs));
2537: PetscCall(ISGetBlockSize(iscol, &cbs));
2538: PetscCall(MatSetBlockSizes(C, rbs, cbs));
2539: PetscCall(MatSetType(C, ((PetscObject)A)->type_name));
2540: PetscCall(MatSeqAIJSetPreallocation_SeqAIJ(C, 0, lens));
2541: }
2542: c = (Mat_SeqAIJ *)C->data;
2544: /* loop over rows inserting into submatrix */
2545: PetscCall(MatSeqAIJGetArrayWrite(C, &a_new)); // Not 'a_new = c->a-new', since that raw usage ignores offload state of C
2546: j_new = c->j;
2547: i_new = c->i;
2548: PetscCall(MatSeqAIJGetArrayRead(A, &aa));
2549: for (i = 0; i < nrows; i++) {
2550: ii = starts[i];
2551: lensi = lens[i];
2552: if (lensi) {
2553: for (k = 0; k < lensi; k++) *j_new++ = aj[ii + k] - first;
2554: PetscCall(PetscArraycpy(a_new, aa + starts[i], lensi));
2555: a_new += lensi;
2556: }
2557: i_new[i + 1] = i_new[i] + lensi;
2558: c->ilen[i] = lensi;
2559: }
2560: PetscCall(MatSeqAIJRestoreArrayWrite(C, &a_new)); // Set C's offload state properly
2561: PetscCall(MatSeqAIJRestoreArrayRead(A, &aa));
2562: PetscCall(PetscFree2(lens, starts));
2563: } else {
2564: PetscCall(ISGetIndices(iscol, &icol));
2565: PetscCall(PetscCalloc1(oldcols, &smap));
2566: PetscCall(PetscMalloc1(1 + nrows, &lens));
2567: for (i = 0; i < ncols; i++) {
2568: PetscCheck(icol[i] < oldcols, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Requesting column beyond largest column icol[%" PetscInt_FMT "] %" PetscInt_FMT " >= A->cmap->n %" PetscInt_FMT, i, icol[i], oldcols);
2569: smap[icol[i]] = i + 1;
2570: }
2572: /* determine lens of each row */
2573: for (i = 0; i < nrows; i++) {
2574: kstart = ai[irow[i]];
2575: kend = kstart + a->ilen[irow[i]];
2576: lens[i] = 0;
2577: for (k = kstart; k < kend; k++) {
2578: if (smap[aj[k]]) lens[i]++;
2579: }
2580: }
2581: /* Create and fill new matrix */
2582: if (scall == MAT_REUSE_MATRIX) {
2583: PetscBool equal;
2585: c = (Mat_SeqAIJ *)((*B)->data);
2586: PetscCheck((*B)->rmap->n == nrows && (*B)->cmap->n == ncols, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Cannot reuse matrix. wrong size");
2587: PetscCall(PetscArraycmp(c->ilen, lens, (*B)->rmap->n, &equal));
2588: PetscCheck(equal, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Cannot reuse matrix. wrong number of nonzeros");
2589: PetscCall(PetscArrayzero(c->ilen, (*B)->rmap->n));
2590: C = *B;
2591: } else {
2592: PetscInt rbs, cbs;
2593: PetscCall(MatCreate(PetscObjectComm((PetscObject)A), &C));
2594: PetscCall(MatSetSizes(C, nrows, ncols, PETSC_DETERMINE, PETSC_DETERMINE));
2595: PetscCall(ISGetBlockSize(isrow, &rbs));
2596: PetscCall(ISGetBlockSize(iscol, &cbs));
2597: if (rbs > 1 || cbs > 1) PetscCall(MatSetBlockSizes(C, rbs, cbs));
2598: PetscCall(MatSetType(C, ((PetscObject)A)->type_name));
2599: PetscCall(MatSeqAIJSetPreallocation_SeqAIJ(C, 0, lens));
2600: }
2601: PetscCall(MatSeqAIJGetArrayRead(A, &aa));
2603: c = (Mat_SeqAIJ *)C->data;
2604: PetscCall(MatSeqAIJGetArrayWrite(C, &c_a)); // Not 'c->a', since that raw usage ignores offload state of C
2605: for (i = 0; i < nrows; i++) {
2606: row = irow[i];
2607: kstart = ai[row];
2608: kend = kstart + a->ilen[row];
2609: mat_i = c->i[i];
2610: mat_j = PetscSafePointerPlusOffset(c->j, mat_i);
2611: mat_a = PetscSafePointerPlusOffset(c_a, mat_i);
2612: mat_ilen = c->ilen + i;
2613: for (k = kstart; k < kend; k++) {
2614: if ((tcol = smap[a->j[k]])) {
2615: *mat_j++ = tcol - 1;
2616: *mat_a++ = aa[k];
2617: (*mat_ilen)++;
2618: }
2619: }
2620: }
2621: PetscCall(MatSeqAIJRestoreArrayRead(A, &aa));
2622: /* Free work space */
2623: PetscCall(ISRestoreIndices(iscol, &icol));
2624: PetscCall(PetscFree(smap));
2625: PetscCall(PetscFree(lens));
2626: /* sort */
2627: for (i = 0; i < nrows; i++) {
2628: PetscInt ilen;
2630: mat_i = c->i[i];
2631: mat_j = PetscSafePointerPlusOffset(c->j, mat_i);
2632: mat_a = PetscSafePointerPlusOffset(c_a, mat_i);
2633: ilen = c->ilen[i];
2634: PetscCall(PetscSortIntWithScalarArray(ilen, mat_j, mat_a));
2635: }
2636: PetscCall(MatSeqAIJRestoreArrayWrite(C, &c_a));
2637: }
2638: #if defined(PETSC_HAVE_DEVICE)
2639: PetscCall(MatBindToCPU(C, A->boundtocpu));
2640: #endif
2641: PetscCall(MatAssemblyBegin(C, MAT_FINAL_ASSEMBLY));
2642: PetscCall(MatAssemblyEnd(C, MAT_FINAL_ASSEMBLY));
2644: PetscCall(ISRestoreIndices(isrow, &irow));
2645: *B = C;
2646: PetscFunctionReturn(PETSC_SUCCESS);
2647: }
2649: static PetscErrorCode MatGetMultiProcBlock_SeqAIJ(Mat mat, MPI_Comm subComm, MatReuse scall, Mat *subMat)
2650: {
2651: Mat B;
2653: PetscFunctionBegin;
2654: if (scall == MAT_INITIAL_MATRIX) {
2655: PetscCall(MatCreate(subComm, &B));
2656: PetscCall(MatSetSizes(B, mat->rmap->n, mat->cmap->n, mat->rmap->n, mat->cmap->n));
2657: PetscCall(MatSetBlockSizesFromMats(B, mat, mat));
2658: PetscCall(MatSetType(B, MATSEQAIJ));
2659: PetscCall(MatDuplicateNoCreate_SeqAIJ(B, mat, MAT_COPY_VALUES, PETSC_TRUE));
2660: *subMat = B;
2661: } else {
2662: PetscCall(MatCopy_SeqAIJ(mat, *subMat, SAME_NONZERO_PATTERN));
2663: }
2664: PetscFunctionReturn(PETSC_SUCCESS);
2665: }
2667: static PetscErrorCode MatILUFactor_SeqAIJ(Mat inA, IS row, IS col, const MatFactorInfo *info)
2668: {
2669: Mat_SeqAIJ *a = (Mat_SeqAIJ *)inA->data;
2670: Mat outA;
2671: PetscBool row_identity, col_identity;
2673: PetscFunctionBegin;
2674: PetscCheck(info->levels == 0, PETSC_COMM_SELF, PETSC_ERR_SUP, "Only levels=0 supported for in-place ilu");
2676: PetscCall(ISIdentity(row, &row_identity));
2677: PetscCall(ISIdentity(col, &col_identity));
2679: outA = inA;
2680: outA->factortype = MAT_FACTOR_LU;
2681: PetscCall(PetscFree(inA->solvertype));
2682: PetscCall(PetscStrallocpy(MATSOLVERPETSC, &inA->solvertype));
2684: PetscCall(PetscObjectReference((PetscObject)row));
2685: PetscCall(ISDestroy(&a->row));
2687: a->row = row;
2689: PetscCall(PetscObjectReference((PetscObject)col));
2690: PetscCall(ISDestroy(&a->col));
2692: a->col = col;
2694: /* Create the inverse permutation so that it can be used in MatLUFactorNumeric() */
2695: PetscCall(ISDestroy(&a->icol));
2696: PetscCall(ISInvertPermutation(col, PETSC_DECIDE, &a->icol));
2698: if (!a->solve_work) { /* this matrix may have been factored before */
2699: PetscCall(PetscMalloc1(inA->rmap->n + 1, &a->solve_work));
2700: }
2702: PetscCall(MatMarkDiagonal_SeqAIJ(inA));
2703: if (row_identity && col_identity) {
2704: PetscCall(MatLUFactorNumeric_SeqAIJ_inplace(outA, inA, info));
2705: } else {
2706: PetscCall(MatLUFactorNumeric_SeqAIJ_InplaceWithPerm(outA, inA, info));
2707: }
2708: PetscFunctionReturn(PETSC_SUCCESS);
2709: }
2711: PetscErrorCode MatScale_SeqAIJ(Mat inA, PetscScalar alpha)
2712: {
2713: Mat_SeqAIJ *a = (Mat_SeqAIJ *)inA->data;
2714: PetscScalar *v;
2715: PetscBLASInt one = 1, bnz;
2717: PetscFunctionBegin;
2718: PetscCall(MatSeqAIJGetArray(inA, &v));
2719: PetscCall(PetscBLASIntCast(a->nz, &bnz));
2720: PetscCallBLAS("BLASscal", BLASscal_(&bnz, &alpha, v, &one));
2721: PetscCall(PetscLogFlops(a->nz));
2722: PetscCall(MatSeqAIJRestoreArray(inA, &v));
2723: PetscCall(MatSeqAIJInvalidateDiagonal(inA));
2724: PetscFunctionReturn(PETSC_SUCCESS);
2725: }
2727: PetscErrorCode MatDestroySubMatrix_Private(Mat_SubSppt *submatj)
2728: {
2729: PetscInt i;
2731: PetscFunctionBegin;
2732: if (!submatj->id) { /* delete data that are linked only to submats[id=0] */
2733: PetscCall(PetscFree4(submatj->sbuf1, submatj->ptr, submatj->tmp, submatj->ctr));
2735: for (i = 0; i < submatj->nrqr; ++i) PetscCall(PetscFree(submatj->sbuf2[i]));
2736: PetscCall(PetscFree3(submatj->sbuf2, submatj->req_size, submatj->req_source1));
2738: if (submatj->rbuf1) {
2739: PetscCall(PetscFree(submatj->rbuf1[0]));
2740: PetscCall(PetscFree(submatj->rbuf1));
2741: }
2743: for (i = 0; i < submatj->nrqs; ++i) PetscCall(PetscFree(submatj->rbuf3[i]));
2744: PetscCall(PetscFree3(submatj->req_source2, submatj->rbuf2, submatj->rbuf3));
2745: PetscCall(PetscFree(submatj->pa));
2746: }
2748: #if defined(PETSC_USE_CTABLE)
2749: PetscCall(PetscHMapIDestroy(&submatj->rmap));
2750: if (submatj->cmap_loc) PetscCall(PetscFree(submatj->cmap_loc));
2751: PetscCall(PetscFree(submatj->rmap_loc));
2752: #else
2753: PetscCall(PetscFree(submatj->rmap));
2754: #endif
2756: if (!submatj->allcolumns) {
2757: #if defined(PETSC_USE_CTABLE)
2758: PetscCall(PetscHMapIDestroy(&submatj->cmap));
2759: #else
2760: PetscCall(PetscFree(submatj->cmap));
2761: #endif
2762: }
2763: PetscCall(PetscFree(submatj->row2proc));
2765: PetscCall(PetscFree(submatj));
2766: PetscFunctionReturn(PETSC_SUCCESS);
2767: }
2769: PetscErrorCode MatDestroySubMatrix_SeqAIJ(Mat C)
2770: {
2771: Mat_SeqAIJ *c = (Mat_SeqAIJ *)C->data;
2772: Mat_SubSppt *submatj = c->submatis1;
2774: PetscFunctionBegin;
2775: PetscCall((*submatj->destroy)(C));
2776: PetscCall(MatDestroySubMatrix_Private(submatj));
2777: PetscFunctionReturn(PETSC_SUCCESS);
2778: }
2780: /* Note this has code duplication with MatDestroySubMatrices_SeqBAIJ() */
2781: static PetscErrorCode MatDestroySubMatrices_SeqAIJ(PetscInt n, Mat *mat[])
2782: {
2783: PetscInt i;
2784: Mat C;
2785: Mat_SeqAIJ *c;
2786: Mat_SubSppt *submatj;
2788: PetscFunctionBegin;
2789: for (i = 0; i < n; i++) {
2790: C = (*mat)[i];
2791: c = (Mat_SeqAIJ *)C->data;
2792: submatj = c->submatis1;
2793: if (submatj) {
2794: if (--((PetscObject)C)->refct <= 0) {
2795: PetscCall(PetscFree(C->factorprefix));
2796: PetscCall((*submatj->destroy)(C));
2797: PetscCall(MatDestroySubMatrix_Private(submatj));
2798: PetscCall(PetscFree(C->defaultvectype));
2799: PetscCall(PetscFree(C->defaultrandtype));
2800: PetscCall(PetscLayoutDestroy(&C->rmap));
2801: PetscCall(PetscLayoutDestroy(&C->cmap));
2802: PetscCall(PetscHeaderDestroy(&C));
2803: }
2804: } else {
2805: PetscCall(MatDestroy(&C));
2806: }
2807: }
2809: /* Destroy Dummy submatrices created for reuse */
2810: PetscCall(MatDestroySubMatrices_Dummy(n, mat));
2812: PetscCall(PetscFree(*mat));
2813: PetscFunctionReturn(PETSC_SUCCESS);
2814: }
2816: static PetscErrorCode MatCreateSubMatrices_SeqAIJ(Mat A, PetscInt n, const IS irow[], const IS icol[], MatReuse scall, Mat *B[])
2817: {
2818: PetscInt i;
2820: PetscFunctionBegin;
2821: if (scall == MAT_INITIAL_MATRIX) PetscCall(PetscCalloc1(n + 1, B));
2823: for (i = 0; i < n; i++) PetscCall(MatCreateSubMatrix_SeqAIJ(A, irow[i], icol[i], PETSC_DECIDE, scall, &(*B)[i]));
2824: PetscFunctionReturn(PETSC_SUCCESS);
2825: }
2827: static PetscErrorCode MatIncreaseOverlap_SeqAIJ(Mat A, PetscInt is_max, IS is[], PetscInt ov)
2828: {
2829: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
2830: PetscInt row, i, j, k, l, ll, m, n, *nidx, isz, val;
2831: const PetscInt *idx;
2832: PetscInt start, end, *ai, *aj, bs = (A->rmap->bs > 0 && A->rmap->bs == A->cmap->bs) ? A->rmap->bs : 1;
2833: PetscBT table;
2835: PetscFunctionBegin;
2836: m = A->rmap->n / bs;
2837: ai = a->i;
2838: aj = a->j;
2840: PetscCheck(ov >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "illegal negative overlap value used");
2842: PetscCall(PetscMalloc1(m + 1, &nidx));
2843: PetscCall(PetscBTCreate(m, &table));
2845: for (i = 0; i < is_max; i++) {
2846: /* Initialize the two local arrays */
2847: isz = 0;
2848: PetscCall(PetscBTMemzero(m, table));
2850: /* Extract the indices, assume there can be duplicate entries */
2851: PetscCall(ISGetIndices(is[i], &idx));
2852: PetscCall(ISGetLocalSize(is[i], &n));
2854: if (bs > 1) {
2855: /* Enter these into the temp arrays. I.e., mark table[row], enter row into new index */
2856: for (j = 0; j < n; ++j) {
2857: if (!PetscBTLookupSet(table, idx[j] / bs)) nidx[isz++] = idx[j] / bs;
2858: }
2859: PetscCall(ISRestoreIndices(is[i], &idx));
2860: PetscCall(ISDestroy(&is[i]));
2862: k = 0;
2863: for (j = 0; j < ov; j++) { /* for each overlap */
2864: n = isz;
2865: for (; k < n; k++) { /* do only those rows in nidx[k], which are not done yet */
2866: for (ll = 0; ll < bs; ll++) {
2867: row = bs * nidx[k] + ll;
2868: start = ai[row];
2869: end = ai[row + 1];
2870: for (l = start; l < end; l++) {
2871: val = aj[l] / bs;
2872: if (!PetscBTLookupSet(table, val)) nidx[isz++] = val;
2873: }
2874: }
2875: }
2876: }
2877: PetscCall(ISCreateBlock(PETSC_COMM_SELF, bs, isz, nidx, PETSC_COPY_VALUES, is + i));
2878: } else {
2879: /* Enter these into the temp arrays. I.e., mark table[row], enter row into new index */
2880: for (j = 0; j < n; ++j) {
2881: if (!PetscBTLookupSet(table, idx[j])) nidx[isz++] = idx[j];
2882: }
2883: PetscCall(ISRestoreIndices(is[i], &idx));
2884: PetscCall(ISDestroy(&is[i]));
2886: k = 0;
2887: for (j = 0; j < ov; j++) { /* for each overlap */
2888: n = isz;
2889: for (; k < n; k++) { /* do only those rows in nidx[k], which are not done yet */
2890: row = nidx[k];
2891: start = ai[row];
2892: end = ai[row + 1];
2893: for (l = start; l < end; l++) {
2894: val = aj[l];
2895: if (!PetscBTLookupSet(table, val)) nidx[isz++] = val;
2896: }
2897: }
2898: }
2899: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, isz, nidx, PETSC_COPY_VALUES, is + i));
2900: }
2901: }
2902: PetscCall(PetscBTDestroy(&table));
2903: PetscCall(PetscFree(nidx));
2904: PetscFunctionReturn(PETSC_SUCCESS);
2905: }
2907: static PetscErrorCode MatPermute_SeqAIJ(Mat A, IS rowp, IS colp, Mat *B)
2908: {
2909: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
2910: PetscInt i, nz = 0, m = A->rmap->n, n = A->cmap->n;
2911: const PetscInt *row, *col;
2912: PetscInt *cnew, j, *lens;
2913: IS icolp, irowp;
2914: PetscInt *cwork = NULL;
2915: PetscScalar *vwork = NULL;
2917: PetscFunctionBegin;
2918: PetscCall(ISInvertPermutation(rowp, PETSC_DECIDE, &irowp));
2919: PetscCall(ISGetIndices(irowp, &row));
2920: PetscCall(ISInvertPermutation(colp, PETSC_DECIDE, &icolp));
2921: PetscCall(ISGetIndices(icolp, &col));
2923: /* determine lengths of permuted rows */
2924: PetscCall(PetscMalloc1(m + 1, &lens));
2925: for (i = 0; i < m; i++) lens[row[i]] = a->i[i + 1] - a->i[i];
2926: PetscCall(MatCreate(PetscObjectComm((PetscObject)A), B));
2927: PetscCall(MatSetSizes(*B, m, n, m, n));
2928: PetscCall(MatSetBlockSizesFromMats(*B, A, A));
2929: PetscCall(MatSetType(*B, ((PetscObject)A)->type_name));
2930: PetscCall(MatSeqAIJSetPreallocation_SeqAIJ(*B, 0, lens));
2931: PetscCall(PetscFree(lens));
2933: PetscCall(PetscMalloc1(n, &cnew));
2934: for (i = 0; i < m; i++) {
2935: PetscCall(MatGetRow_SeqAIJ(A, i, &nz, &cwork, &vwork));
2936: for (j = 0; j < nz; j++) cnew[j] = col[cwork[j]];
2937: PetscCall(MatSetValues_SeqAIJ(*B, 1, &row[i], nz, cnew, vwork, INSERT_VALUES));
2938: PetscCall(MatRestoreRow_SeqAIJ(A, i, &nz, &cwork, &vwork));
2939: }
2940: PetscCall(PetscFree(cnew));
2942: (*B)->assembled = PETSC_FALSE;
2944: #if defined(PETSC_HAVE_DEVICE)
2945: PetscCall(MatBindToCPU(*B, A->boundtocpu));
2946: #endif
2947: PetscCall(MatAssemblyBegin(*B, MAT_FINAL_ASSEMBLY));
2948: PetscCall(MatAssemblyEnd(*B, MAT_FINAL_ASSEMBLY));
2949: PetscCall(ISRestoreIndices(irowp, &row));
2950: PetscCall(ISRestoreIndices(icolp, &col));
2951: PetscCall(ISDestroy(&irowp));
2952: PetscCall(ISDestroy(&icolp));
2953: if (rowp == colp) PetscCall(MatPropagateSymmetryOptions(A, *B));
2954: PetscFunctionReturn(PETSC_SUCCESS);
2955: }
2957: PetscErrorCode MatCopy_SeqAIJ(Mat A, Mat B, MatStructure str)
2958: {
2959: PetscFunctionBegin;
2960: /* If the two matrices have the same copy implementation, use fast copy. */
2961: if (str == SAME_NONZERO_PATTERN && (A->ops->copy == B->ops->copy)) {
2962: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
2963: Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data;
2964: const PetscScalar *aa;
2966: PetscCall(MatSeqAIJGetArrayRead(A, &aa));
2967: PetscCheck(a->i[A->rmap->n] == b->i[B->rmap->n], PETSC_COMM_SELF, PETSC_ERR_ARG_INCOMP, "Number of nonzeros in two matrices are different %" PetscInt_FMT " != %" PetscInt_FMT, a->i[A->rmap->n], b->i[B->rmap->n]);
2968: PetscCall(PetscArraycpy(b->a, aa, a->i[A->rmap->n]));
2969: PetscCall(PetscObjectStateIncrease((PetscObject)B));
2970: PetscCall(MatSeqAIJRestoreArrayRead(A, &aa));
2971: } else {
2972: PetscCall(MatCopy_Basic(A, B, str));
2973: }
2974: PetscFunctionReturn(PETSC_SUCCESS);
2975: }
2977: PETSC_INTERN PetscErrorCode MatSeqAIJGetArray_SeqAIJ(Mat A, PetscScalar *array[])
2978: {
2979: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
2981: PetscFunctionBegin;
2982: *array = a->a;
2983: PetscFunctionReturn(PETSC_SUCCESS);
2984: }
2986: PETSC_INTERN PetscErrorCode MatSeqAIJRestoreArray_SeqAIJ(Mat A, PetscScalar *array[])
2987: {
2988: PetscFunctionBegin;
2989: *array = NULL;
2990: PetscFunctionReturn(PETSC_SUCCESS);
2991: }
2993: /*
2994: Computes the number of nonzeros per row needed for preallocation when X and Y
2995: have different nonzero structure.
2996: */
2997: PetscErrorCode MatAXPYGetPreallocation_SeqX_private(PetscInt m, const PetscInt *xi, const PetscInt *xj, const PetscInt *yi, const PetscInt *yj, PetscInt *nnz)
2998: {
2999: PetscInt i, j, k, nzx, nzy;
3001: PetscFunctionBegin;
3002: /* Set the number of nonzeros in the new matrix */
3003: for (i = 0; i < m; i++) {
3004: const PetscInt *xjj = PetscSafePointerPlusOffset(xj, xi[i]), *yjj = PetscSafePointerPlusOffset(yj, yi[i]);
3005: nzx = xi[i + 1] - xi[i];
3006: nzy = yi[i + 1] - yi[i];
3007: nnz[i] = 0;
3008: for (j = 0, k = 0; j < nzx; j++) { /* Point in X */
3009: for (; k < nzy && yjj[k] < xjj[j]; k++) nnz[i]++; /* Catch up to X */
3010: if (k < nzy && yjj[k] == xjj[j]) k++; /* Skip duplicate */
3011: nnz[i]++;
3012: }
3013: for (; k < nzy; k++) nnz[i]++;
3014: }
3015: PetscFunctionReturn(PETSC_SUCCESS);
3016: }
3018: PetscErrorCode MatAXPYGetPreallocation_SeqAIJ(Mat Y, Mat X, PetscInt *nnz)
3019: {
3020: PetscInt m = Y->rmap->N;
3021: Mat_SeqAIJ *x = (Mat_SeqAIJ *)X->data;
3022: Mat_SeqAIJ *y = (Mat_SeqAIJ *)Y->data;
3024: PetscFunctionBegin;
3025: /* Set the number of nonzeros in the new matrix */
3026: PetscCall(MatAXPYGetPreallocation_SeqX_private(m, x->i, x->j, y->i, y->j, nnz));
3027: PetscFunctionReturn(PETSC_SUCCESS);
3028: }
3030: PetscErrorCode MatAXPY_SeqAIJ(Mat Y, PetscScalar a, Mat X, MatStructure str)
3031: {
3032: Mat_SeqAIJ *x = (Mat_SeqAIJ *)X->data, *y = (Mat_SeqAIJ *)Y->data;
3034: PetscFunctionBegin;
3035: if (str == UNKNOWN_NONZERO_PATTERN || (PetscDefined(USE_DEBUG) && str == SAME_NONZERO_PATTERN)) {
3036: PetscBool e = x->nz == y->nz ? PETSC_TRUE : PETSC_FALSE;
3037: if (e) {
3038: PetscCall(PetscArraycmp(x->i, y->i, Y->rmap->n + 1, &e));
3039: if (e) {
3040: PetscCall(PetscArraycmp(x->j, y->j, y->nz, &e));
3041: if (e) str = SAME_NONZERO_PATTERN;
3042: }
3043: }
3044: if (!e) PetscCheck(str != SAME_NONZERO_PATTERN, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "MatStructure is not SAME_NONZERO_PATTERN");
3045: }
3046: if (str == SAME_NONZERO_PATTERN) {
3047: const PetscScalar *xa;
3048: PetscScalar *ya, alpha = a;
3049: PetscBLASInt one = 1, bnz;
3051: PetscCall(PetscBLASIntCast(x->nz, &bnz));
3052: PetscCall(MatSeqAIJGetArray(Y, &ya));
3053: PetscCall(MatSeqAIJGetArrayRead(X, &xa));
3054: PetscCallBLAS("BLASaxpy", BLASaxpy_(&bnz, &alpha, xa, &one, ya, &one));
3055: PetscCall(MatSeqAIJRestoreArrayRead(X, &xa));
3056: PetscCall(MatSeqAIJRestoreArray(Y, &ya));
3057: PetscCall(PetscLogFlops(2.0 * bnz));
3058: PetscCall(MatSeqAIJInvalidateDiagonal(Y));
3059: PetscCall(PetscObjectStateIncrease((PetscObject)Y));
3060: } else if (str == SUBSET_NONZERO_PATTERN) { /* nonzeros of X is a subset of Y's */
3061: PetscCall(MatAXPY_Basic(Y, a, X, str));
3062: } else {
3063: Mat B;
3064: PetscInt *nnz;
3065: PetscCall(PetscMalloc1(Y->rmap->N, &nnz));
3066: PetscCall(MatCreate(PetscObjectComm((PetscObject)Y), &B));
3067: PetscCall(PetscObjectSetName((PetscObject)B, ((PetscObject)Y)->name));
3068: PetscCall(MatSetLayouts(B, Y->rmap, Y->cmap));
3069: PetscCall(MatSetType(B, ((PetscObject)Y)->type_name));
3070: PetscCall(MatAXPYGetPreallocation_SeqAIJ(Y, X, nnz));
3071: PetscCall(MatSeqAIJSetPreallocation(B, 0, nnz));
3072: PetscCall(MatAXPY_BasicWithPreallocation(B, Y, a, X, str));
3073: PetscCall(MatHeaderMerge(Y, &B));
3074: PetscCall(MatSeqAIJCheckInode(Y));
3075: PetscCall(PetscFree(nnz));
3076: }
3077: PetscFunctionReturn(PETSC_SUCCESS);
3078: }
3080: PETSC_INTERN PetscErrorCode MatConjugate_SeqAIJ(Mat mat)
3081: {
3082: #if defined(PETSC_USE_COMPLEX)
3083: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)mat->data;
3084: PetscInt i, nz;
3085: PetscScalar *a;
3087: PetscFunctionBegin;
3088: nz = aij->nz;
3089: PetscCall(MatSeqAIJGetArray(mat, &a));
3090: for (i = 0; i < nz; i++) a[i] = PetscConj(a[i]);
3091: PetscCall(MatSeqAIJRestoreArray(mat, &a));
3092: #else
3093: PetscFunctionBegin;
3094: #endif
3095: PetscFunctionReturn(PETSC_SUCCESS);
3096: }
3098: static PetscErrorCode MatGetRowMaxAbs_SeqAIJ(Mat A, Vec v, PetscInt idx[])
3099: {
3100: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
3101: PetscInt i, j, m = A->rmap->n, *ai, *aj, ncols, n;
3102: PetscReal atmp;
3103: PetscScalar *x;
3104: const MatScalar *aa, *av;
3106: PetscFunctionBegin;
3107: PetscCheck(!A->factortype, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Not for factored matrix");
3108: PetscCall(MatSeqAIJGetArrayRead(A, &av));
3109: aa = av;
3110: ai = a->i;
3111: aj = a->j;
3113: PetscCall(VecSet(v, 0.0));
3114: PetscCall(VecGetArrayWrite(v, &x));
3115: PetscCall(VecGetLocalSize(v, &n));
3116: PetscCheck(n == A->rmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Nonconforming matrix and vector");
3117: for (i = 0; i < m; i++) {
3118: ncols = ai[1] - ai[0];
3119: ai++;
3120: for (j = 0; j < ncols; j++) {
3121: atmp = PetscAbsScalar(*aa);
3122: if (PetscAbsScalar(x[i]) < atmp) {
3123: x[i] = atmp;
3124: if (idx) idx[i] = *aj;
3125: }
3126: aa++;
3127: aj++;
3128: }
3129: }
3130: PetscCall(VecRestoreArrayWrite(v, &x));
3131: PetscCall(MatSeqAIJRestoreArrayRead(A, &av));
3132: PetscFunctionReturn(PETSC_SUCCESS);
3133: }
3135: static PetscErrorCode MatGetRowSumAbs_SeqAIJ(Mat A, Vec v)
3136: {
3137: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
3138: PetscInt i, j, m = A->rmap->n, *ai, ncols, n;
3139: PetscScalar *x;
3140: const MatScalar *aa, *av;
3142: PetscFunctionBegin;
3143: PetscCheck(!A->factortype, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Not for factored matrix");
3144: PetscCall(MatSeqAIJGetArrayRead(A, &av));
3145: aa = av;
3146: ai = a->i;
3148: PetscCall(VecSet(v, 0.0));
3149: PetscCall(VecGetArrayWrite(v, &x));
3150: PetscCall(VecGetLocalSize(v, &n));
3151: PetscCheck(n == A->rmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Nonconforming matrix and vector");
3152: for (i = 0; i < m; i++) {
3153: ncols = ai[1] - ai[0];
3154: ai++;
3155: for (j = 0; j < ncols; j++) {
3156: x[i] += PetscAbsScalar(*aa);
3157: aa++;
3158: }
3159: }
3160: PetscCall(VecRestoreArrayWrite(v, &x));
3161: PetscCall(MatSeqAIJRestoreArrayRead(A, &av));
3162: PetscFunctionReturn(PETSC_SUCCESS);
3163: }
3165: static PetscErrorCode MatGetRowMax_SeqAIJ(Mat A, Vec v, PetscInt idx[])
3166: {
3167: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
3168: PetscInt i, j, m = A->rmap->n, *ai, *aj, ncols, n;
3169: PetscScalar *x;
3170: const MatScalar *aa, *av;
3172: PetscFunctionBegin;
3173: PetscCheck(!A->factortype, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Not for factored matrix");
3174: PetscCall(MatSeqAIJGetArrayRead(A, &av));
3175: aa = av;
3176: ai = a->i;
3177: aj = a->j;
3179: PetscCall(VecSet(v, 0.0));
3180: PetscCall(VecGetArrayWrite(v, &x));
3181: PetscCall(VecGetLocalSize(v, &n));
3182: PetscCheck(n == A->rmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Nonconforming matrix and vector");
3183: for (i = 0; i < m; i++) {
3184: ncols = ai[1] - ai[0];
3185: ai++;
3186: if (ncols == A->cmap->n) { /* row is dense */
3187: x[i] = *aa;
3188: if (idx) idx[i] = 0;
3189: } else { /* row is sparse so already KNOW maximum is 0.0 or higher */
3190: x[i] = 0.0;
3191: if (idx) {
3192: for (j = 0; j < ncols; j++) { /* find first implicit 0.0 in the row */
3193: if (aj[j] > j) {
3194: idx[i] = j;
3195: break;
3196: }
3197: }
3198: /* in case first implicit 0.0 in the row occurs at ncols-th column */
3199: if (j == ncols && j < A->cmap->n) idx[i] = j;
3200: }
3201: }
3202: for (j = 0; j < ncols; j++) {
3203: if (PetscRealPart(x[i]) < PetscRealPart(*aa)) {
3204: x[i] = *aa;
3205: if (idx) idx[i] = *aj;
3206: }
3207: aa++;
3208: aj++;
3209: }
3210: }
3211: PetscCall(VecRestoreArrayWrite(v, &x));
3212: PetscCall(MatSeqAIJRestoreArrayRead(A, &av));
3213: PetscFunctionReturn(PETSC_SUCCESS);
3214: }
3216: static PetscErrorCode MatGetRowMinAbs_SeqAIJ(Mat A, Vec v, PetscInt idx[])
3217: {
3218: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
3219: PetscInt i, j, m = A->rmap->n, *ai, *aj, ncols, n;
3220: PetscScalar *x;
3221: const MatScalar *aa, *av;
3223: PetscFunctionBegin;
3224: PetscCall(MatSeqAIJGetArrayRead(A, &av));
3225: aa = av;
3226: ai = a->i;
3227: aj = a->j;
3229: PetscCall(VecSet(v, 0.0));
3230: PetscCall(VecGetArrayWrite(v, &x));
3231: PetscCall(VecGetLocalSize(v, &n));
3232: PetscCheck(n == m, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Nonconforming matrix and vector, %" PetscInt_FMT " vs. %" PetscInt_FMT " rows", m, n);
3233: for (i = 0; i < m; i++) {
3234: ncols = ai[1] - ai[0];
3235: ai++;
3236: if (ncols == A->cmap->n) { /* row is dense */
3237: x[i] = *aa;
3238: if (idx) idx[i] = 0;
3239: } else { /* row is sparse so already KNOW minimum is 0.0 or higher */
3240: x[i] = 0.0;
3241: if (idx) { /* find first implicit 0.0 in the row */
3242: for (j = 0; j < ncols; j++) {
3243: if (aj[j] > j) {
3244: idx[i] = j;
3245: break;
3246: }
3247: }
3248: /* in case first implicit 0.0 in the row occurs at ncols-th column */
3249: if (j == ncols && j < A->cmap->n) idx[i] = j;
3250: }
3251: }
3252: for (j = 0; j < ncols; j++) {
3253: if (PetscAbsScalar(x[i]) > PetscAbsScalar(*aa)) {
3254: x[i] = *aa;
3255: if (idx) idx[i] = *aj;
3256: }
3257: aa++;
3258: aj++;
3259: }
3260: }
3261: PetscCall(VecRestoreArrayWrite(v, &x));
3262: PetscCall(MatSeqAIJRestoreArrayRead(A, &av));
3263: PetscFunctionReturn(PETSC_SUCCESS);
3264: }
3266: static PetscErrorCode MatGetRowMin_SeqAIJ(Mat A, Vec v, PetscInt idx[])
3267: {
3268: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
3269: PetscInt i, j, m = A->rmap->n, ncols, n;
3270: const PetscInt *ai, *aj;
3271: PetscScalar *x;
3272: const MatScalar *aa, *av;
3274: PetscFunctionBegin;
3275: PetscCheck(!A->factortype, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Not for factored matrix");
3276: PetscCall(MatSeqAIJGetArrayRead(A, &av));
3277: aa = av;
3278: ai = a->i;
3279: aj = a->j;
3281: PetscCall(VecSet(v, 0.0));
3282: PetscCall(VecGetArrayWrite(v, &x));
3283: PetscCall(VecGetLocalSize(v, &n));
3284: PetscCheck(n == m, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Nonconforming matrix and vector");
3285: for (i = 0; i < m; i++) {
3286: ncols = ai[1] - ai[0];
3287: ai++;
3288: if (ncols == A->cmap->n) { /* row is dense */
3289: x[i] = *aa;
3290: if (idx) idx[i] = 0;
3291: } else { /* row is sparse so already KNOW minimum is 0.0 or lower */
3292: x[i] = 0.0;
3293: if (idx) { /* find first implicit 0.0 in the row */
3294: for (j = 0; j < ncols; j++) {
3295: if (aj[j] > j) {
3296: idx[i] = j;
3297: break;
3298: }
3299: }
3300: /* in case first implicit 0.0 in the row occurs at ncols-th column */
3301: if (j == ncols && j < A->cmap->n) idx[i] = j;
3302: }
3303: }
3304: for (j = 0; j < ncols; j++) {
3305: if (PetscRealPart(x[i]) > PetscRealPart(*aa)) {
3306: x[i] = *aa;
3307: if (idx) idx[i] = *aj;
3308: }
3309: aa++;
3310: aj++;
3311: }
3312: }
3313: PetscCall(VecRestoreArrayWrite(v, &x));
3314: PetscCall(MatSeqAIJRestoreArrayRead(A, &av));
3315: PetscFunctionReturn(PETSC_SUCCESS);
3316: }
3318: static PetscErrorCode MatInvertBlockDiagonal_SeqAIJ(Mat A, const PetscScalar **values)
3319: {
3320: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
3321: PetscInt i, bs = PetscAbs(A->rmap->bs), mbs = A->rmap->n / bs, ipvt[5], bs2 = bs * bs, *v_pivots, ij[7], *IJ, j;
3322: MatScalar *diag, work[25], *v_work;
3323: const PetscReal shift = 0.0;
3324: PetscBool allowzeropivot, zeropivotdetected = PETSC_FALSE;
3326: PetscFunctionBegin;
3327: allowzeropivot = PetscNot(A->erroriffailure);
3328: if (a->ibdiagvalid) {
3329: if (values) *values = a->ibdiag;
3330: PetscFunctionReturn(PETSC_SUCCESS);
3331: }
3332: PetscCall(MatMarkDiagonal_SeqAIJ(A));
3333: if (!a->ibdiag) { PetscCall(PetscMalloc1(bs2 * mbs, &a->ibdiag)); }
3334: diag = a->ibdiag;
3335: if (values) *values = a->ibdiag;
3336: /* factor and invert each block */
3337: switch (bs) {
3338: case 1:
3339: for (i = 0; i < mbs; i++) {
3340: PetscCall(MatGetValues(A, 1, &i, 1, &i, diag + i));
3341: if (PetscAbsScalar(diag[i] + shift) < PETSC_MACHINE_EPSILON) {
3342: if (allowzeropivot) {
3343: A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
3344: A->factorerror_zeropivot_value = PetscAbsScalar(diag[i]);
3345: A->factorerror_zeropivot_row = i;
3346: PetscCall(PetscInfo(A, "Zero pivot, row %" PetscInt_FMT " pivot %g tolerance %g\n", i, (double)PetscAbsScalar(diag[i]), (double)PETSC_MACHINE_EPSILON));
3347: } 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);
3348: }
3349: diag[i] = (PetscScalar)1.0 / (diag[i] + shift);
3350: }
3351: break;
3352: case 2:
3353: for (i = 0; i < mbs; i++) {
3354: ij[0] = 2 * i;
3355: ij[1] = 2 * i + 1;
3356: PetscCall(MatGetValues(A, 2, ij, 2, ij, diag));
3357: PetscCall(PetscKernel_A_gets_inverse_A_2(diag, shift, allowzeropivot, &zeropivotdetected));
3358: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
3359: PetscCall(PetscKernel_A_gets_transpose_A_2(diag));
3360: diag += 4;
3361: }
3362: break;
3363: case 3:
3364: for (i = 0; i < mbs; i++) {
3365: ij[0] = 3 * i;
3366: ij[1] = 3 * i + 1;
3367: ij[2] = 3 * i + 2;
3368: PetscCall(MatGetValues(A, 3, ij, 3, ij, diag));
3369: PetscCall(PetscKernel_A_gets_inverse_A_3(diag, shift, allowzeropivot, &zeropivotdetected));
3370: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
3371: PetscCall(PetscKernel_A_gets_transpose_A_3(diag));
3372: diag += 9;
3373: }
3374: break;
3375: case 4:
3376: for (i = 0; i < mbs; i++) {
3377: ij[0] = 4 * i;
3378: ij[1] = 4 * i + 1;
3379: ij[2] = 4 * i + 2;
3380: ij[3] = 4 * i + 3;
3381: PetscCall(MatGetValues(A, 4, ij, 4, ij, diag));
3382: PetscCall(PetscKernel_A_gets_inverse_A_4(diag, shift, allowzeropivot, &zeropivotdetected));
3383: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
3384: PetscCall(PetscKernel_A_gets_transpose_A_4(diag));
3385: diag += 16;
3386: }
3387: break;
3388: case 5:
3389: for (i = 0; i < mbs; i++) {
3390: ij[0] = 5 * i;
3391: ij[1] = 5 * i + 1;
3392: ij[2] = 5 * i + 2;
3393: ij[3] = 5 * i + 3;
3394: ij[4] = 5 * i + 4;
3395: PetscCall(MatGetValues(A, 5, ij, 5, ij, diag));
3396: PetscCall(PetscKernel_A_gets_inverse_A_5(diag, ipvt, work, shift, allowzeropivot, &zeropivotdetected));
3397: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
3398: PetscCall(PetscKernel_A_gets_transpose_A_5(diag));
3399: diag += 25;
3400: }
3401: break;
3402: case 6:
3403: for (i = 0; i < mbs; i++) {
3404: ij[0] = 6 * i;
3405: ij[1] = 6 * i + 1;
3406: ij[2] = 6 * i + 2;
3407: ij[3] = 6 * i + 3;
3408: ij[4] = 6 * i + 4;
3409: ij[5] = 6 * i + 5;
3410: PetscCall(MatGetValues(A, 6, ij, 6, ij, diag));
3411: PetscCall(PetscKernel_A_gets_inverse_A_6(diag, shift, allowzeropivot, &zeropivotdetected));
3412: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
3413: PetscCall(PetscKernel_A_gets_transpose_A_6(diag));
3414: diag += 36;
3415: }
3416: break;
3417: case 7:
3418: for (i = 0; i < mbs; i++) {
3419: ij[0] = 7 * i;
3420: ij[1] = 7 * i + 1;
3421: ij[2] = 7 * i + 2;
3422: ij[3] = 7 * i + 3;
3423: ij[4] = 7 * i + 4;
3424: ij[5] = 7 * i + 5;
3425: ij[6] = 7 * i + 6;
3426: PetscCall(MatGetValues(A, 7, ij, 7, ij, diag));
3427: PetscCall(PetscKernel_A_gets_inverse_A_7(diag, shift, allowzeropivot, &zeropivotdetected));
3428: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
3429: PetscCall(PetscKernel_A_gets_transpose_A_7(diag));
3430: diag += 49;
3431: }
3432: break;
3433: default:
3434: PetscCall(PetscMalloc3(bs, &v_work, bs, &v_pivots, bs, &IJ));
3435: for (i = 0; i < mbs; i++) {
3436: for (j = 0; j < bs; j++) IJ[j] = bs * i + j;
3437: PetscCall(MatGetValues(A, bs, IJ, bs, IJ, diag));
3438: PetscCall(PetscKernel_A_gets_inverse_A(bs, diag, v_pivots, v_work, allowzeropivot, &zeropivotdetected));
3439: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
3440: PetscCall(PetscKernel_A_gets_transpose_A_N(diag, bs));
3441: diag += bs2;
3442: }
3443: PetscCall(PetscFree3(v_work, v_pivots, IJ));
3444: }
3445: a->ibdiagvalid = PETSC_TRUE;
3446: PetscFunctionReturn(PETSC_SUCCESS);
3447: }
3449: static PetscErrorCode MatSetRandom_SeqAIJ(Mat x, PetscRandom rctx)
3450: {
3451: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)x->data;
3452: PetscScalar a, *aa;
3453: PetscInt m, n, i, j, col;
3455: PetscFunctionBegin;
3456: if (!x->assembled) {
3457: PetscCall(MatGetSize(x, &m, &n));
3458: for (i = 0; i < m; i++) {
3459: for (j = 0; j < aij->imax[i]; j++) {
3460: PetscCall(PetscRandomGetValue(rctx, &a));
3461: col = (PetscInt)(n * PetscRealPart(a));
3462: PetscCall(MatSetValues(x, 1, &i, 1, &col, &a, ADD_VALUES));
3463: }
3464: }
3465: } else {
3466: PetscCall(MatSeqAIJGetArrayWrite(x, &aa));
3467: for (i = 0; i < aij->nz; i++) PetscCall(PetscRandomGetValue(rctx, aa + i));
3468: PetscCall(MatSeqAIJRestoreArrayWrite(x, &aa));
3469: }
3470: PetscCall(MatAssemblyBegin(x, MAT_FINAL_ASSEMBLY));
3471: PetscCall(MatAssemblyEnd(x, MAT_FINAL_ASSEMBLY));
3472: PetscFunctionReturn(PETSC_SUCCESS);
3473: }
3475: /* Like MatSetRandom_SeqAIJ, but do not set values on columns in range of [low, high) */
3476: PetscErrorCode MatSetRandomSkipColumnRange_SeqAIJ_Private(Mat x, PetscInt low, PetscInt high, PetscRandom rctx)
3477: {
3478: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)x->data;
3479: PetscScalar a;
3480: PetscInt m, n, i, j, col, nskip;
3482: PetscFunctionBegin;
3483: nskip = high - low;
3484: PetscCall(MatGetSize(x, &m, &n));
3485: n -= nskip; /* shrink number of columns where nonzeros can be set */
3486: for (i = 0; i < m; i++) {
3487: for (j = 0; j < aij->imax[i]; j++) {
3488: PetscCall(PetscRandomGetValue(rctx, &a));
3489: col = (PetscInt)(n * PetscRealPart(a));
3490: if (col >= low) col += nskip; /* shift col rightward to skip the hole */
3491: PetscCall(MatSetValues(x, 1, &i, 1, &col, &a, ADD_VALUES));
3492: }
3493: }
3494: PetscCall(MatAssemblyBegin(x, MAT_FINAL_ASSEMBLY));
3495: PetscCall(MatAssemblyEnd(x, MAT_FINAL_ASSEMBLY));
3496: PetscFunctionReturn(PETSC_SUCCESS);
3497: }
3499: static struct _MatOps MatOps_Values = {MatSetValues_SeqAIJ,
3500: MatGetRow_SeqAIJ,
3501: MatRestoreRow_SeqAIJ,
3502: MatMult_SeqAIJ,
3503: /* 4*/ MatMultAdd_SeqAIJ,
3504: MatMultTranspose_SeqAIJ,
3505: MatMultTransposeAdd_SeqAIJ,
3506: NULL,
3507: NULL,
3508: NULL,
3509: /* 10*/ NULL,
3510: MatLUFactor_SeqAIJ,
3511: NULL,
3512: MatSOR_SeqAIJ,
3513: MatTranspose_SeqAIJ,
3514: /*1 5*/ MatGetInfo_SeqAIJ,
3515: MatEqual_SeqAIJ,
3516: MatGetDiagonal_SeqAIJ,
3517: MatDiagonalScale_SeqAIJ,
3518: MatNorm_SeqAIJ,
3519: /* 20*/ NULL,
3520: MatAssemblyEnd_SeqAIJ,
3521: MatSetOption_SeqAIJ,
3522: MatZeroEntries_SeqAIJ,
3523: /* 24*/ MatZeroRows_SeqAIJ,
3524: NULL,
3525: NULL,
3526: NULL,
3527: NULL,
3528: /* 29*/ MatSetUp_Seq_Hash,
3529: NULL,
3530: NULL,
3531: NULL,
3532: NULL,
3533: /* 34*/ MatDuplicate_SeqAIJ,
3534: NULL,
3535: NULL,
3536: MatILUFactor_SeqAIJ,
3537: NULL,
3538: /* 39*/ MatAXPY_SeqAIJ,
3539: MatCreateSubMatrices_SeqAIJ,
3540: MatIncreaseOverlap_SeqAIJ,
3541: MatGetValues_SeqAIJ,
3542: MatCopy_SeqAIJ,
3543: /* 44*/ MatGetRowMax_SeqAIJ,
3544: MatScale_SeqAIJ,
3545: MatShift_SeqAIJ,
3546: MatDiagonalSet_SeqAIJ,
3547: MatZeroRowsColumns_SeqAIJ,
3548: /* 49*/ MatSetRandom_SeqAIJ,
3549: MatGetRowIJ_SeqAIJ,
3550: MatRestoreRowIJ_SeqAIJ,
3551: MatGetColumnIJ_SeqAIJ,
3552: MatRestoreColumnIJ_SeqAIJ,
3553: /* 54*/ MatFDColoringCreate_SeqXAIJ,
3554: NULL,
3555: NULL,
3556: MatPermute_SeqAIJ,
3557: NULL,
3558: /* 59*/ NULL,
3559: MatDestroy_SeqAIJ,
3560: MatView_SeqAIJ,
3561: NULL,
3562: NULL,
3563: /* 64*/ NULL,
3564: MatMatMatMultNumeric_SeqAIJ_SeqAIJ_SeqAIJ,
3565: NULL,
3566: NULL,
3567: NULL,
3568: /* 69*/ MatGetRowMaxAbs_SeqAIJ,
3569: MatGetRowMinAbs_SeqAIJ,
3570: NULL,
3571: NULL,
3572: NULL,
3573: /* 74*/ NULL,
3574: MatFDColoringApply_AIJ,
3575: NULL,
3576: NULL,
3577: NULL,
3578: /* 79*/ MatFindZeroDiagonals_SeqAIJ,
3579: NULL,
3580: NULL,
3581: NULL,
3582: MatLoad_SeqAIJ,
3583: /* 84*/ NULL,
3584: NULL,
3585: NULL,
3586: NULL,
3587: NULL,
3588: /* 89*/ NULL,
3589: NULL,
3590: MatMatMultNumeric_SeqAIJ_SeqAIJ,
3591: NULL,
3592: NULL,
3593: /* 94*/ MatPtAPNumeric_SeqAIJ_SeqAIJ_SparseAxpy,
3594: NULL,
3595: NULL,
3596: MatMatTransposeMultNumeric_SeqAIJ_SeqAIJ,
3597: NULL,
3598: /* 99*/ MatProductSetFromOptions_SeqAIJ,
3599: NULL,
3600: NULL,
3601: MatConjugate_SeqAIJ,
3602: NULL,
3603: /*104*/ MatSetValuesRow_SeqAIJ,
3604: MatRealPart_SeqAIJ,
3605: MatImaginaryPart_SeqAIJ,
3606: NULL,
3607: NULL,
3608: /*109*/ MatMatSolve_SeqAIJ,
3609: NULL,
3610: MatGetRowMin_SeqAIJ,
3611: NULL,
3612: MatMissingDiagonal_SeqAIJ,
3613: /*114*/ NULL,
3614: NULL,
3615: NULL,
3616: NULL,
3617: NULL,
3618: /*119*/ NULL,
3619: NULL,
3620: NULL,
3621: NULL,
3622: MatGetMultiProcBlock_SeqAIJ,
3623: /*124*/ MatFindNonzeroRows_SeqAIJ,
3624: MatGetColumnReductions_SeqAIJ,
3625: MatInvertBlockDiagonal_SeqAIJ,
3626: MatInvertVariableBlockDiagonal_SeqAIJ,
3627: NULL,
3628: /*129*/ NULL,
3629: NULL,
3630: NULL,
3631: MatTransposeMatMultNumeric_SeqAIJ_SeqAIJ,
3632: MatTransposeColoringCreate_SeqAIJ,
3633: /*134*/ MatTransColoringApplySpToDen_SeqAIJ,
3634: MatTransColoringApplyDenToSp_SeqAIJ,
3635: NULL,
3636: NULL,
3637: MatRARtNumeric_SeqAIJ_SeqAIJ,
3638: /*139*/ NULL,
3639: NULL,
3640: NULL,
3641: MatFDColoringSetUp_SeqXAIJ,
3642: MatFindOffBlockDiagonalEntries_SeqAIJ,
3643: MatCreateMPIMatConcatenateSeqMat_SeqAIJ,
3644: /*145*/ MatDestroySubMatrices_SeqAIJ,
3645: NULL,
3646: NULL,
3647: MatCreateGraph_Simple_AIJ,
3648: NULL,
3649: /*150*/ MatTransposeSymbolic_SeqAIJ,
3650: MatEliminateZeros_SeqAIJ,
3651: MatGetRowSumAbs_SeqAIJ,
3652: NULL,
3653: NULL,
3654: /*155*/ NULL,
3655: MatCopyHashToXAIJ_Seq_Hash};
3657: static PetscErrorCode MatSeqAIJSetColumnIndices_SeqAIJ(Mat mat, PetscInt *indices)
3658: {
3659: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)mat->data;
3660: PetscInt i, nz, n;
3662: PetscFunctionBegin;
3663: nz = aij->maxnz;
3664: n = mat->rmap->n;
3665: for (i = 0; i < nz; i++) aij->j[i] = indices[i];
3666: aij->nz = nz;
3667: for (i = 0; i < n; i++) aij->ilen[i] = aij->imax[i];
3668: PetscFunctionReturn(PETSC_SUCCESS);
3669: }
3671: /*
3672: * Given a sparse matrix with global column indices, compact it by using a local column space.
3673: * The result matrix helps saving memory in other algorithms, such as MatPtAPSymbolic_MPIAIJ_MPIAIJ_scalable()
3674: */
3675: PetscErrorCode MatSeqAIJCompactOutExtraColumns_SeqAIJ(Mat mat, ISLocalToGlobalMapping *mapping)
3676: {
3677: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)mat->data;
3678: PetscHMapI gid1_lid1;
3679: PetscHashIter tpos;
3680: PetscInt gid, lid, i, ec, nz = aij->nz;
3681: PetscInt *garray, *jj = aij->j;
3683: PetscFunctionBegin;
3685: PetscAssertPointer(mapping, 2);
3686: /* use a table */
3687: PetscCall(PetscHMapICreateWithSize(mat->rmap->n, &gid1_lid1));
3688: ec = 0;
3689: for (i = 0; i < nz; i++) {
3690: PetscInt data, gid1 = jj[i] + 1;
3691: PetscCall(PetscHMapIGetWithDefault(gid1_lid1, gid1, 0, &data));
3692: if (!data) {
3693: /* one based table */
3694: PetscCall(PetscHMapISet(gid1_lid1, gid1, ++ec));
3695: }
3696: }
3697: /* form array of columns we need */
3698: PetscCall(PetscMalloc1(ec, &garray));
3699: PetscHashIterBegin(gid1_lid1, tpos);
3700: while (!PetscHashIterAtEnd(gid1_lid1, tpos)) {
3701: PetscHashIterGetKey(gid1_lid1, tpos, gid);
3702: PetscHashIterGetVal(gid1_lid1, tpos, lid);
3703: PetscHashIterNext(gid1_lid1, tpos);
3704: gid--;
3705: lid--;
3706: garray[lid] = gid;
3707: }
3708: PetscCall(PetscSortInt(ec, garray)); /* sort, and rebuild */
3709: PetscCall(PetscHMapIClear(gid1_lid1));
3710: for (i = 0; i < ec; i++) PetscCall(PetscHMapISet(gid1_lid1, garray[i] + 1, i + 1));
3711: /* compact out the extra columns in B */
3712: for (i = 0; i < nz; i++) {
3713: PetscInt gid1 = jj[i] + 1;
3714: PetscCall(PetscHMapIGetWithDefault(gid1_lid1, gid1, 0, &lid));
3715: lid--;
3716: jj[i] = lid;
3717: }
3718: PetscCall(PetscLayoutDestroy(&mat->cmap));
3719: PetscCall(PetscHMapIDestroy(&gid1_lid1));
3720: PetscCall(PetscLayoutCreateFromSizes(PetscObjectComm((PetscObject)mat), ec, ec, 1, &mat->cmap));
3721: PetscCall(ISLocalToGlobalMappingCreate(PETSC_COMM_SELF, mat->cmap->bs, mat->cmap->n, garray, PETSC_OWN_POINTER, mapping));
3722: PetscCall(ISLocalToGlobalMappingSetType(*mapping, ISLOCALTOGLOBALMAPPINGHASH));
3723: PetscFunctionReturn(PETSC_SUCCESS);
3724: }
3726: /*@
3727: MatSeqAIJSetColumnIndices - Set the column indices for all the rows
3728: in the matrix.
3730: Input Parameters:
3731: + mat - the `MATSEQAIJ` matrix
3732: - indices - the column indices
3734: Level: advanced
3736: Notes:
3737: This can be called if you have precomputed the nonzero structure of the
3738: matrix and want to provide it to the matrix object to improve the performance
3739: of the `MatSetValues()` operation.
3741: You MUST have set the correct numbers of nonzeros per row in the call to
3742: `MatCreateSeqAIJ()`, and the columns indices MUST be sorted.
3744: MUST be called before any calls to `MatSetValues()`
3746: The indices should start with zero, not one.
3748: .seealso: [](ch_matrices), `Mat`, `MATSEQAIJ`
3749: @*/
3750: PetscErrorCode MatSeqAIJSetColumnIndices(Mat mat, PetscInt *indices)
3751: {
3752: PetscFunctionBegin;
3754: PetscAssertPointer(indices, 2);
3755: PetscUseMethod(mat, "MatSeqAIJSetColumnIndices_C", (Mat, PetscInt *), (mat, indices));
3756: PetscFunctionReturn(PETSC_SUCCESS);
3757: }
3759: static PetscErrorCode MatStoreValues_SeqAIJ(Mat mat)
3760: {
3761: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)mat->data;
3762: size_t nz = aij->i[mat->rmap->n];
3764: PetscFunctionBegin;
3765: PetscCheck(aij->nonew, PETSC_COMM_SELF, PETSC_ERR_ORDER, "Must call MatSetOption(A,MAT_NEW_NONZERO_LOCATIONS,PETSC_FALSE);first");
3767: /* allocate space for values if not already there */
3768: if (!aij->saved_values) { PetscCall(PetscMalloc1(nz + 1, &aij->saved_values)); }
3770: /* copy values over */
3771: PetscCall(PetscArraycpy(aij->saved_values, aij->a, nz));
3772: PetscFunctionReturn(PETSC_SUCCESS);
3773: }
3775: /*@
3776: MatStoreValues - Stashes a copy of the matrix values; this allows reusing of the linear part of a Jacobian, while recomputing only the
3777: nonlinear portion.
3779: Logically Collect
3781: Input Parameter:
3782: . mat - the matrix (currently only `MATAIJ` matrices support this option)
3784: Level: advanced
3786: Example Usage:
3787: .vb
3788: Using SNES
3789: Create Jacobian matrix
3790: Set linear terms into matrix
3791: Apply boundary conditions to matrix, at this time matrix must have
3792: final nonzero structure (i.e. setting the nonlinear terms and applying
3793: boundary conditions again will not change the nonzero structure
3794: MatSetOption(mat, MAT_NEW_NONZERO_LOCATIONS, PETSC_FALSE);
3795: MatStoreValues(mat);
3796: Call SNESSetJacobian() with matrix
3797: In your Jacobian routine
3798: MatRetrieveValues(mat);
3799: Set nonlinear terms in matrix
3801: Without `SNESSolve()`, i.e. when you handle nonlinear solve yourself:
3802: // build linear portion of Jacobian
3803: MatSetOption(mat, MAT_NEW_NONZERO_LOCATIONS, PETSC_FALSE);
3804: MatStoreValues(mat);
3805: loop over nonlinear iterations
3806: MatRetrieveValues(mat);
3807: // call MatSetValues(mat,...) to set nonliner portion of Jacobian
3808: // call MatAssemblyBegin/End() on matrix
3809: Solve linear system with Jacobian
3810: endloop
3811: .ve
3813: Notes:
3814: Matrix must already be assembled before calling this routine
3815: Must set the matrix option `MatSetOption`(mat,`MAT_NEW_NONZERO_LOCATIONS`,`PETSC_FALSE`); before
3816: calling this routine.
3818: When this is called multiple times it overwrites the previous set of stored values
3819: and does not allocated additional space.
3821: .seealso: [](ch_matrices), `Mat`, `MatRetrieveValues()`
3822: @*/
3823: PetscErrorCode MatStoreValues(Mat mat)
3824: {
3825: PetscFunctionBegin;
3827: PetscCheck(mat->assembled, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Not for unassembled matrix");
3828: PetscCheck(!mat->factortype, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Not for factored matrix");
3829: PetscUseMethod(mat, "MatStoreValues_C", (Mat), (mat));
3830: PetscFunctionReturn(PETSC_SUCCESS);
3831: }
3833: static PetscErrorCode MatRetrieveValues_SeqAIJ(Mat mat)
3834: {
3835: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)mat->data;
3836: PetscInt nz = aij->i[mat->rmap->n];
3838: PetscFunctionBegin;
3839: PetscCheck(aij->nonew, PETSC_COMM_SELF, PETSC_ERR_ORDER, "Must call MatSetOption(A,MAT_NEW_NONZERO_LOCATIONS,PETSC_FALSE);first");
3840: PetscCheck(aij->saved_values, PETSC_COMM_SELF, PETSC_ERR_ORDER, "Must call MatStoreValues(A);first");
3841: /* copy values over */
3842: PetscCall(PetscArraycpy(aij->a, aij->saved_values, nz));
3843: PetscFunctionReturn(PETSC_SUCCESS);
3844: }
3846: /*@
3847: MatRetrieveValues - Retrieves the copy of the matrix values that was stored with `MatStoreValues()`
3849: Logically Collect
3851: Input Parameter:
3852: . mat - the matrix (currently only `MATAIJ` matrices support this option)
3854: Level: advanced
3856: .seealso: [](ch_matrices), `Mat`, `MatStoreValues()`
3857: @*/
3858: PetscErrorCode MatRetrieveValues(Mat mat)
3859: {
3860: PetscFunctionBegin;
3862: PetscCheck(mat->assembled, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Not for unassembled matrix");
3863: PetscCheck(!mat->factortype, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Not for factored matrix");
3864: PetscUseMethod(mat, "MatRetrieveValues_C", (Mat), (mat));
3865: PetscFunctionReturn(PETSC_SUCCESS);
3866: }
3868: /*@
3869: MatCreateSeqAIJ - Creates a sparse matrix in `MATSEQAIJ` (compressed row) format
3870: (the default parallel PETSc format). For good matrix assembly performance
3871: the user should preallocate the matrix storage by setting the parameter `nz`
3872: (or the array `nnz`).
3874: Collective
3876: Input Parameters:
3877: + comm - MPI communicator, set to `PETSC_COMM_SELF`
3878: . m - number of rows
3879: . n - number of columns
3880: . nz - number of nonzeros per row (same for all rows)
3881: - nnz - array containing the number of nonzeros in the various rows
3882: (possibly different for each row) or NULL
3884: Output Parameter:
3885: . A - the matrix
3887: Options Database Keys:
3888: + -mat_no_inode - Do not use inodes
3889: - -mat_inode_limit <limit> - Sets inode limit (max limit=5)
3891: Level: intermediate
3893: Notes:
3894: It is recommend to use `MatCreateFromOptions()` instead of this routine
3896: If `nnz` is given then `nz` is ignored
3898: The `MATSEQAIJ` format, also called
3899: compressed row storage, is fully compatible with standard Fortran
3900: storage. That is, the stored row and column indices can begin at
3901: either one (as in Fortran) or zero.
3903: Specify the preallocated storage with either `nz` or `nnz` (not both).
3904: Set `nz` = `PETSC_DEFAULT` and `nnz` = `NULL` for PETSc to control dynamic memory
3905: allocation.
3907: By default, this format uses inodes (identical nodes) when possible, to
3908: improve numerical efficiency of matrix-vector products and solves. We
3909: search for consecutive rows with the same nonzero structure, thereby
3910: reusing matrix information to achieve increased efficiency.
3912: .seealso: [](ch_matrices), `Mat`, [Sparse Matrix Creation](sec_matsparse), `MatCreate()`, `MatCreateAIJ()`, `MatSetValues()`, `MatSeqAIJSetColumnIndices()`, `MatCreateSeqAIJWithArrays()`
3913: @*/
3914: PetscErrorCode MatCreateSeqAIJ(MPI_Comm comm, PetscInt m, PetscInt n, PetscInt nz, const PetscInt nnz[], Mat *A)
3915: {
3916: PetscFunctionBegin;
3917: PetscCall(MatCreate(comm, A));
3918: PetscCall(MatSetSizes(*A, m, n, m, n));
3919: PetscCall(MatSetType(*A, MATSEQAIJ));
3920: PetscCall(MatSeqAIJSetPreallocation_SeqAIJ(*A, nz, nnz));
3921: PetscFunctionReturn(PETSC_SUCCESS);
3922: }
3924: /*@
3925: MatSeqAIJSetPreallocation - For good matrix assembly performance
3926: the user should preallocate the matrix storage by setting the parameter nz
3927: (or the array nnz). By setting these parameters accurately, performance
3928: during matrix assembly can be increased by more than a factor of 50.
3930: Collective
3932: Input Parameters:
3933: + B - The matrix
3934: . nz - number of nonzeros per row (same for all rows)
3935: - nnz - array containing the number of nonzeros in the various rows
3936: (possibly different for each row) or NULL
3938: Options Database Keys:
3939: + -mat_no_inode - Do not use inodes
3940: - -mat_inode_limit <limit> - Sets inode limit (max limit=5)
3942: Level: intermediate
3944: Notes:
3945: If `nnz` is given then `nz` is ignored
3947: The `MATSEQAIJ` format also called
3948: compressed row storage, is fully compatible with standard Fortran
3949: storage. That is, the stored row and column indices can begin at
3950: either one (as in Fortran) or zero. See the users' manual for details.
3952: Specify the preallocated storage with either `nz` or `nnz` (not both).
3953: Set nz = `PETSC_DEFAULT` and `nnz` = `NULL` for PETSc to control dynamic memory
3954: allocation.
3956: You can call `MatGetInfo()` to get information on how effective the preallocation was;
3957: for example the fields mallocs,nz_allocated,nz_used,nz_unneeded;
3958: You can also run with the option -info and look for messages with the string
3959: malloc in them to see if additional memory allocation was needed.
3961: Developer Notes:
3962: Use nz of `MAT_SKIP_ALLOCATION` to not allocate any space for the matrix
3963: entries or columns indices
3965: By default, this format uses inodes (identical nodes) when possible, to
3966: improve numerical efficiency of matrix-vector products and solves. We
3967: search for consecutive rows with the same nonzero structure, thereby
3968: reusing matrix information to achieve increased efficiency.
3970: .seealso: [](ch_matrices), `Mat`, `MatCreate()`, `MatCreateAIJ()`, `MatSetValues()`, `MatSeqAIJSetColumnIndices()`, `MatCreateSeqAIJWithArrays()`, `MatGetInfo()`,
3971: `MatSeqAIJSetTotalPreallocation()`
3972: @*/
3973: PetscErrorCode MatSeqAIJSetPreallocation(Mat B, PetscInt nz, const PetscInt nnz[])
3974: {
3975: PetscFunctionBegin;
3978: PetscTryMethod(B, "MatSeqAIJSetPreallocation_C", (Mat, PetscInt, const PetscInt[]), (B, nz, nnz));
3979: PetscFunctionReturn(PETSC_SUCCESS);
3980: }
3982: PetscErrorCode MatSeqAIJSetPreallocation_SeqAIJ(Mat B, PetscInt nz, const PetscInt *nnz)
3983: {
3984: Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data;
3985: PetscBool skipallocation = PETSC_FALSE, realalloc = PETSC_FALSE;
3986: PetscInt i;
3988: PetscFunctionBegin;
3989: if (B->hash_active) {
3990: B->ops[0] = b->cops;
3991: PetscCall(PetscHMapIJVDestroy(&b->ht));
3992: PetscCall(PetscFree(b->dnz));
3993: B->hash_active = PETSC_FALSE;
3994: }
3995: if (nz >= 0 || nnz) realalloc = PETSC_TRUE;
3996: if (nz == MAT_SKIP_ALLOCATION) {
3997: skipallocation = PETSC_TRUE;
3998: nz = 0;
3999: }
4000: PetscCall(PetscLayoutSetUp(B->rmap));
4001: PetscCall(PetscLayoutSetUp(B->cmap));
4003: if (nz == PETSC_DEFAULT || nz == PETSC_DECIDE) nz = 5;
4004: PetscCheck(nz >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "nz cannot be less than 0: value %" PetscInt_FMT, nz);
4005: if (nnz) {
4006: for (i = 0; i < B->rmap->n; i++) {
4007: PetscCheck(nnz[i] >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "nnz cannot be less than 0: local row %" PetscInt_FMT " value %" PetscInt_FMT, i, nnz[i]);
4008: PetscCheck(nnz[i] <= B->cmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "nnz cannot be greater than row length: local row %" PetscInt_FMT " value %" PetscInt_FMT " rowlength %" PetscInt_FMT, i, nnz[i], B->cmap->n);
4009: }
4010: }
4012: B->preallocated = PETSC_TRUE;
4013: if (!skipallocation) {
4014: if (!b->imax) { PetscCall(PetscMalloc1(B->rmap->n, &b->imax)); }
4015: if (!b->ilen) {
4016: /* b->ilen will count nonzeros in each row so far. */
4017: PetscCall(PetscCalloc1(B->rmap->n, &b->ilen));
4018: } else {
4019: PetscCall(PetscMemzero(b->ilen, B->rmap->n * sizeof(PetscInt)));
4020: }
4021: if (!b->ipre) PetscCall(PetscMalloc1(B->rmap->n, &b->ipre));
4022: if (!nnz) {
4023: if (nz == PETSC_DEFAULT || nz == PETSC_DECIDE) nz = 10;
4024: else if (nz < 0) nz = 1;
4025: nz = PetscMin(nz, B->cmap->n);
4026: for (i = 0; i < B->rmap->n; i++) b->imax[i] = nz;
4027: PetscCall(PetscIntMultError(nz, B->rmap->n, &nz));
4028: } else {
4029: PetscInt64 nz64 = 0;
4030: for (i = 0; i < B->rmap->n; i++) {
4031: b->imax[i] = nnz[i];
4032: nz64 += nnz[i];
4033: }
4034: PetscCall(PetscIntCast(nz64, &nz));
4035: }
4037: /* allocate the matrix space */
4038: PetscCall(MatSeqXAIJFreeAIJ(B, &b->a, &b->j, &b->i));
4039: PetscCall(PetscShmgetAllocateArray(nz, sizeof(PetscInt), (void **)&b->j));
4040: PetscCall(PetscShmgetAllocateArray(B->rmap->n + 1, sizeof(PetscInt), (void **)&b->i));
4041: b->free_ij = PETSC_TRUE;
4042: if (B->structure_only) {
4043: b->free_a = PETSC_FALSE;
4044: } else {
4045: PetscCall(PetscShmgetAllocateArray(nz, sizeof(PetscScalar), (void **)&b->a));
4046: b->free_a = PETSC_TRUE;
4047: }
4048: b->i[0] = 0;
4049: for (i = 1; i < B->rmap->n + 1; i++) b->i[i] = b->i[i - 1] + b->imax[i - 1];
4050: } else {
4051: b->free_a = PETSC_FALSE;
4052: b->free_ij = PETSC_FALSE;
4053: }
4055: if (b->ipre && nnz != b->ipre && b->imax) {
4056: /* reserve user-requested sparsity */
4057: PetscCall(PetscArraycpy(b->ipre, b->imax, B->rmap->n));
4058: }
4060: b->nz = 0;
4061: b->maxnz = nz;
4062: B->info.nz_unneeded = (double)b->maxnz;
4063: if (realalloc) PetscCall(MatSetOption(B, MAT_NEW_NONZERO_ALLOCATION_ERR, PETSC_TRUE));
4064: B->was_assembled = PETSC_FALSE;
4065: B->assembled = PETSC_FALSE;
4066: /* We simply deem preallocation has changed nonzero state. Updating the state
4067: will give clients (like AIJKokkos) a chance to know something has happened.
4068: */
4069: B->nonzerostate++;
4070: PetscFunctionReturn(PETSC_SUCCESS);
4071: }
4073: static PetscErrorCode MatResetPreallocation_SeqAIJ(Mat A)
4074: {
4075: Mat_SeqAIJ *a;
4076: PetscInt i;
4077: PetscBool skipreset;
4079: PetscFunctionBegin;
4082: /* Check local size. If zero, then return */
4083: if (!A->rmap->n) PetscFunctionReturn(PETSC_SUCCESS);
4085: a = (Mat_SeqAIJ *)A->data;
4086: /* if no saved info, we error out */
4087: PetscCheck(a->ipre, PETSC_COMM_SELF, PETSC_ERR_ARG_NULL, "No saved preallocation info ");
4089: PetscCheck(a->i && a->imax && a->ilen, PETSC_COMM_SELF, PETSC_ERR_ARG_NULL, "Memory info is incomplete, and can not reset preallocation ");
4091: PetscCall(PetscArraycmp(a->ipre, a->ilen, A->rmap->n, &skipreset));
4092: if (!skipreset) {
4093: PetscCall(PetscArraycpy(a->imax, a->ipre, A->rmap->n));
4094: PetscCall(PetscArrayzero(a->ilen, A->rmap->n));
4095: a->i[0] = 0;
4096: for (i = 1; i < A->rmap->n + 1; i++) a->i[i] = a->i[i - 1] + a->imax[i - 1];
4097: A->preallocated = PETSC_TRUE;
4098: a->nz = 0;
4099: a->maxnz = a->i[A->rmap->n];
4100: A->info.nz_unneeded = (double)a->maxnz;
4101: A->was_assembled = PETSC_FALSE;
4102: A->assembled = PETSC_FALSE;
4103: }
4104: PetscFunctionReturn(PETSC_SUCCESS);
4105: }
4107: /*@
4108: MatSeqAIJSetPreallocationCSR - Allocates memory for a sparse sequential matrix in `MATSEQAIJ` format.
4110: Input Parameters:
4111: + B - the matrix
4112: . i - the indices into `j` for the start of each row (indices start with zero)
4113: . j - the column indices for each row (indices start with zero) these must be sorted for each row
4114: - v - optional values in the matrix, use `NULL` if not provided
4116: Level: developer
4118: Notes:
4119: The `i`,`j`,`v` values are COPIED with this routine; to avoid the copy use `MatCreateSeqAIJWithArrays()`
4121: This routine may be called multiple times with different nonzero patterns (or the same nonzero pattern). The nonzero
4122: structure will be the union of all the previous nonzero structures.
4124: Developer Notes:
4125: 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
4126: then just copies the `v` values directly with `PetscMemcpy()`.
4128: This routine could also take a `PetscCopyMode` argument to allow sharing the values instead of always copying them.
4130: .seealso: [](ch_matrices), `Mat`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatSeqAIJSetPreallocation()`, `MATSEQAIJ`, `MatResetPreallocation()`
4131: @*/
4132: PetscErrorCode MatSeqAIJSetPreallocationCSR(Mat B, const PetscInt i[], const PetscInt j[], const PetscScalar v[])
4133: {
4134: PetscFunctionBegin;
4137: PetscTryMethod(B, "MatSeqAIJSetPreallocationCSR_C", (Mat, const PetscInt[], const PetscInt[], const PetscScalar[]), (B, i, j, v));
4138: PetscFunctionReturn(PETSC_SUCCESS);
4139: }
4141: static PetscErrorCode MatSeqAIJSetPreallocationCSR_SeqAIJ(Mat B, const PetscInt Ii[], const PetscInt J[], const PetscScalar v[])
4142: {
4143: PetscInt i;
4144: PetscInt m, n;
4145: PetscInt nz;
4146: PetscInt *nnz;
4148: PetscFunctionBegin;
4149: PetscCheck(Ii[0] == 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Ii[0] must be 0 it is %" PetscInt_FMT, Ii[0]);
4151: PetscCall(PetscLayoutSetUp(B->rmap));
4152: PetscCall(PetscLayoutSetUp(B->cmap));
4154: PetscCall(MatGetSize(B, &m, &n));
4155: PetscCall(PetscMalloc1(m + 1, &nnz));
4156: for (i = 0; i < m; i++) {
4157: nz = Ii[i + 1] - Ii[i];
4158: PetscCheck(nz >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Local row %" PetscInt_FMT " has a negative number of columns %" PetscInt_FMT, i, nz);
4159: nnz[i] = nz;
4160: }
4161: PetscCall(MatSeqAIJSetPreallocation(B, 0, nnz));
4162: PetscCall(PetscFree(nnz));
4164: for (i = 0; i < m; i++) PetscCall(MatSetValues_SeqAIJ(B, 1, &i, Ii[i + 1] - Ii[i], J + Ii[i], PetscSafePointerPlusOffset(v, Ii[i]), INSERT_VALUES));
4166: PetscCall(MatAssemblyBegin(B, MAT_FINAL_ASSEMBLY));
4167: PetscCall(MatAssemblyEnd(B, MAT_FINAL_ASSEMBLY));
4169: PetscCall(MatSetOption(B, MAT_NEW_NONZERO_LOCATION_ERR, PETSC_TRUE));
4170: PetscFunctionReturn(PETSC_SUCCESS);
4171: }
4173: /*@
4174: MatSeqAIJKron - Computes `C`, the Kronecker product of `A` and `B`.
4176: Input Parameters:
4177: + A - left-hand side matrix
4178: . B - right-hand side matrix
4179: - reuse - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX`
4181: Output Parameter:
4182: . C - Kronecker product of `A` and `B`
4184: Level: intermediate
4186: Note:
4187: `MAT_REUSE_MATRIX` can only be used when the nonzero structure of the product matrix has not changed from that last call to `MatSeqAIJKron()`.
4189: .seealso: [](ch_matrices), `Mat`, `MatCreateSeqAIJ()`, `MATSEQAIJ`, `MATKAIJ`, `MatReuse`
4190: @*/
4191: PetscErrorCode MatSeqAIJKron(Mat A, Mat B, MatReuse reuse, Mat *C)
4192: {
4193: PetscFunctionBegin;
4198: PetscAssertPointer(C, 4);
4199: if (reuse == MAT_REUSE_MATRIX) {
4202: }
4203: PetscTryMethod(A, "MatSeqAIJKron_C", (Mat, Mat, MatReuse, Mat *), (A, B, reuse, C));
4204: PetscFunctionReturn(PETSC_SUCCESS);
4205: }
4207: static PetscErrorCode MatSeqAIJKron_SeqAIJ(Mat A, Mat B, MatReuse reuse, Mat *C)
4208: {
4209: Mat newmat;
4210: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
4211: Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data;
4212: PetscScalar *v;
4213: const PetscScalar *aa, *ba;
4214: 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;
4215: PetscBool flg;
4217: PetscFunctionBegin;
4218: PetscCheck(!A->factortype, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Not for factored matrix");
4219: PetscCheck(A->assembled, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Not for unassembled matrix");
4220: PetscCheck(!B->factortype, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Not for factored matrix");
4221: PetscCheck(B->assembled, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Not for unassembled matrix");
4222: PetscCall(PetscObjectTypeCompare((PetscObject)B, MATSEQAIJ, &flg));
4223: PetscCheck(flg, PETSC_COMM_SELF, PETSC_ERR_SUP, "MatType %s", ((PetscObject)B)->type_name);
4224: PetscCheck(reuse == MAT_INITIAL_MATRIX || reuse == MAT_REUSE_MATRIX, PETSC_COMM_SELF, PETSC_ERR_SUP, "MatReuse %d", (int)reuse);
4225: if (reuse == MAT_INITIAL_MATRIX) {
4226: PetscCall(PetscMalloc2(am * bm + 1, &i, a->i[am] * b->i[bm], &j));
4227: PetscCall(MatCreate(PETSC_COMM_SELF, &newmat));
4228: PetscCall(MatSetSizes(newmat, am * bm, an * bn, am * bm, an * bn));
4229: PetscCall(MatSetType(newmat, MATAIJ));
4230: i[0] = 0;
4231: for (m = 0; m < am; ++m) {
4232: for (p = 0; p < bm; ++p) {
4233: i[m * bm + p + 1] = i[m * bm + p] + (a->i[m + 1] - a->i[m]) * (b->i[p + 1] - b->i[p]);
4234: for (n = a->i[m]; n < a->i[m + 1]; ++n) {
4235: for (q = b->i[p]; q < b->i[p + 1]; ++q) j[nnz++] = a->j[n] * bn + b->j[q];
4236: }
4237: }
4238: }
4239: PetscCall(MatSeqAIJSetPreallocationCSR(newmat, i, j, NULL));
4240: *C = newmat;
4241: PetscCall(PetscFree2(i, j));
4242: nnz = 0;
4243: }
4244: PetscCall(MatSeqAIJGetArray(*C, &v));
4245: PetscCall(MatSeqAIJGetArrayRead(A, &aa));
4246: PetscCall(MatSeqAIJGetArrayRead(B, &ba));
4247: for (m = 0; m < am; ++m) {
4248: for (p = 0; p < bm; ++p) {
4249: for (n = a->i[m]; n < a->i[m + 1]; ++n) {
4250: for (q = b->i[p]; q < b->i[p + 1]; ++q) v[nnz++] = aa[n] * ba[q];
4251: }
4252: }
4253: }
4254: PetscCall(MatSeqAIJRestoreArray(*C, &v));
4255: PetscCall(MatSeqAIJRestoreArrayRead(A, &aa));
4256: PetscCall(MatSeqAIJRestoreArrayRead(B, &ba));
4257: PetscFunctionReturn(PETSC_SUCCESS);
4258: }
4260: #include <../src/mat/impls/dense/seq/dense.h>
4261: #include <petsc/private/kernels/petscaxpy.h>
4263: /*
4264: Computes (B'*A')' since computing B*A directly is untenable
4266: n p p
4267: [ ] [ ] [ ]
4268: m [ A ] * n [ B ] = m [ C ]
4269: [ ] [ ] [ ]
4271: */
4272: PetscErrorCode MatMatMultNumeric_SeqDense_SeqAIJ(Mat A, Mat B, Mat C)
4273: {
4274: Mat_SeqDense *sub_a = (Mat_SeqDense *)A->data;
4275: Mat_SeqAIJ *sub_b = (Mat_SeqAIJ *)B->data;
4276: Mat_SeqDense *sub_c = (Mat_SeqDense *)C->data;
4277: PetscInt i, j, n, m, q, p;
4278: const PetscInt *ii, *idx;
4279: const PetscScalar *b, *a, *a_q;
4280: PetscScalar *c, *c_q;
4281: PetscInt clda = sub_c->lda;
4282: PetscInt alda = sub_a->lda;
4284: PetscFunctionBegin;
4285: m = A->rmap->n;
4286: n = A->cmap->n;
4287: p = B->cmap->n;
4288: a = sub_a->v;
4289: b = sub_b->a;
4290: c = sub_c->v;
4291: if (clda == m) {
4292: PetscCall(PetscArrayzero(c, m * p));
4293: } else {
4294: for (j = 0; j < p; j++)
4295: for (i = 0; i < m; i++) c[j * clda + i] = 0.0;
4296: }
4297: ii = sub_b->i;
4298: idx = sub_b->j;
4299: for (i = 0; i < n; i++) {
4300: q = ii[i + 1] - ii[i];
4301: while (q-- > 0) {
4302: c_q = c + clda * (*idx);
4303: a_q = a + alda * i;
4304: PetscKernelAXPY(c_q, *b, a_q, m);
4305: idx++;
4306: b++;
4307: }
4308: }
4309: PetscFunctionReturn(PETSC_SUCCESS);
4310: }
4312: PetscErrorCode MatMatMultSymbolic_SeqDense_SeqAIJ(Mat A, Mat B, PetscReal fill, Mat C)
4313: {
4314: PetscInt m = A->rmap->n, n = B->cmap->n;
4315: PetscBool cisdense;
4317: PetscFunctionBegin;
4318: PetscCheck(A->cmap->n == B->rmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "A->cmap->n %" PetscInt_FMT " != B->rmap->n %" PetscInt_FMT, A->cmap->n, B->rmap->n);
4319: PetscCall(MatSetSizes(C, m, n, m, n));
4320: PetscCall(MatSetBlockSizesFromMats(C, A, B));
4321: PetscCall(PetscObjectTypeCompareAny((PetscObject)C, &cisdense, MATSEQDENSE, MATSEQDENSECUDA, MATSEQDENSEHIP, ""));
4322: if (!cisdense) PetscCall(MatSetType(C, MATDENSE));
4323: PetscCall(MatSetUp(C));
4325: C->ops->matmultnumeric = MatMatMultNumeric_SeqDense_SeqAIJ;
4326: PetscFunctionReturn(PETSC_SUCCESS);
4327: }
4329: /*MC
4330: MATSEQAIJ - MATSEQAIJ = "seqaij" - A matrix type to be used for sequential sparse matrices,
4331: based on compressed sparse row format.
4333: Options Database Key:
4334: . -mat_type seqaij - sets the matrix type to "seqaij" during a call to MatSetFromOptions()
4336: Level: beginner
4338: Notes:
4339: `MatSetValues()` may be called for this matrix type with a `NULL` argument for the numerical values,
4340: in this case the values associated with the rows and columns one passes in are set to zero
4341: in the matrix
4343: `MatSetOptions`(,`MAT_STRUCTURE_ONLY`,`PETSC_TRUE`) may be called for this matrix type. In this no
4344: space is allocated for the nonzero entries and any entries passed with `MatSetValues()` are ignored
4346: Developer Note:
4347: It would be nice if all matrix formats supported passing `NULL` in for the numerical values
4349: .seealso: [](ch_matrices), `Mat`, `MatCreateSeqAIJ()`, `MatSetFromOptions()`, `MatSetType()`, `MatCreate()`, `MatType`, `MATSELL`, `MATSEQSELL`, `MATMPISELL`
4350: M*/
4352: /*MC
4353: MATAIJ - MATAIJ = "aij" - A matrix type to be used for sparse matrices.
4355: This matrix type is identical to `MATSEQAIJ` when constructed with a single process communicator,
4356: and `MATMPIAIJ` otherwise. As a result, for single process communicators,
4357: `MatSeqAIJSetPreallocation()` is supported, and similarly `MatMPIAIJSetPreallocation()` is supported
4358: for communicators controlling multiple processes. It is recommended that you call both of
4359: the above preallocation routines for simplicity.
4361: Options Database Key:
4362: . -mat_type aij - sets the matrix type to "aij" during a call to `MatSetFromOptions()`
4364: Level: beginner
4366: Note:
4367: Subclasses include `MATAIJCUSPARSE`, `MATAIJPERM`, `MATAIJSELL`, `MATAIJMKL`, `MATAIJCRL`, and also automatically switches over to use inodes when
4368: enough exist.
4370: .seealso: [](ch_matrices), `Mat`, `MatCreateAIJ()`, `MatCreateSeqAIJ()`, `MATSEQAIJ`, `MATMPIAIJ`, `MATSELL`, `MATSEQSELL`, `MATMPISELL`
4371: M*/
4373: /*MC
4374: MATAIJCRL - MATAIJCRL = "aijcrl" - A matrix type to be used for sparse matrices.
4376: Options Database Key:
4377: . -mat_type aijcrl - sets the matrix type to "aijcrl" during a call to `MatSetFromOptions()`
4379: Level: beginner
4381: Note:
4382: This matrix type is identical to `MATSEQAIJCRL` when constructed with a single process communicator,
4383: and `MATMPIAIJCRL` otherwise. As a result, for single process communicators,
4384: `MatSeqAIJSetPreallocation()` is supported, and similarly `MatMPIAIJSetPreallocation()` is supported
4385: for communicators controlling multiple processes. It is recommended that you call both of
4386: the above preallocation routines for simplicity.
4388: .seealso: [](ch_matrices), `Mat`, `MatCreateMPIAIJCRL`, `MATSEQAIJCRL`, `MATMPIAIJCRL`, `MATSEQAIJCRL`, `MATMPIAIJCRL`
4389: M*/
4391: PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqAIJCRL(Mat, MatType, MatReuse, Mat *);
4392: #if defined(PETSC_HAVE_ELEMENTAL)
4393: PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_Elemental(Mat, MatType, MatReuse, Mat *);
4394: #endif
4395: #if defined(PETSC_HAVE_SCALAPACK)
4396: PETSC_INTERN PetscErrorCode MatConvert_AIJ_ScaLAPACK(Mat, MatType, MatReuse, Mat *);
4397: #endif
4398: #if defined(PETSC_HAVE_HYPRE)
4399: PETSC_INTERN PetscErrorCode MatConvert_AIJ_HYPRE(Mat A, MatType, MatReuse, Mat *);
4400: #endif
4402: PETSC_EXTERN PetscErrorCode MatConvert_SeqAIJ_SeqSELL(Mat, MatType, MatReuse, Mat *);
4403: PETSC_INTERN PetscErrorCode MatConvert_XAIJ_IS(Mat, MatType, MatReuse, Mat *);
4404: PETSC_INTERN PetscErrorCode MatProductSetFromOptions_IS_XAIJ(Mat);
4406: /*@C
4407: MatSeqAIJGetArray - gives read/write access to the array where the data for a `MATSEQAIJ` matrix is stored
4409: Not Collective
4411: Input Parameter:
4412: . A - a `MATSEQAIJ` matrix
4414: Output Parameter:
4415: . array - pointer to the data
4417: Level: intermediate
4419: Fortran Notes:
4420: `MatSeqAIJGetArray()` Fortran binding is deprecated (since PETSc 3.19), use `MatSeqAIJGetArrayF90()`
4422: .seealso: [](ch_matrices), `Mat`, `MatSeqAIJRestoreArray()`, `MatSeqAIJGetArrayF90()`
4423: @*/
4424: PetscErrorCode MatSeqAIJGetArray(Mat A, PetscScalar *array[])
4425: {
4426: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)A->data;
4428: PetscFunctionBegin;
4429: if (aij->ops->getarray) {
4430: PetscCall((*aij->ops->getarray)(A, array));
4431: } else {
4432: *array = aij->a;
4433: }
4434: PetscFunctionReturn(PETSC_SUCCESS);
4435: }
4437: /*@C
4438: MatSeqAIJRestoreArray - returns access to the array where the data for a `MATSEQAIJ` matrix is stored obtained by `MatSeqAIJGetArray()`
4440: Not Collective
4442: Input Parameters:
4443: + A - a `MATSEQAIJ` matrix
4444: - array - pointer to the data
4446: Level: intermediate
4448: Fortran Notes:
4449: `MatSeqAIJRestoreArray()` Fortran binding is deprecated (since PETSc 3.19), use `MatSeqAIJRestoreArrayF90()`
4451: .seealso: [](ch_matrices), `Mat`, `MatSeqAIJGetArray()`, `MatSeqAIJRestoreArrayF90()`
4452: @*/
4453: PetscErrorCode MatSeqAIJRestoreArray(Mat A, PetscScalar *array[])
4454: {
4455: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)A->data;
4457: PetscFunctionBegin;
4458: if (aij->ops->restorearray) {
4459: PetscCall((*aij->ops->restorearray)(A, array));
4460: } else {
4461: *array = NULL;
4462: }
4463: PetscCall(MatSeqAIJInvalidateDiagonal(A));
4464: PetscCall(PetscObjectStateIncrease((PetscObject)A));
4465: PetscFunctionReturn(PETSC_SUCCESS);
4466: }
4468: /*@C
4469: MatSeqAIJGetArrayRead - gives read-only access to the array where the data for a `MATSEQAIJ` matrix is stored
4471: Not Collective; No Fortran Support
4473: Input Parameter:
4474: . A - a `MATSEQAIJ` matrix
4476: Output Parameter:
4477: . array - pointer to the data
4479: Level: intermediate
4481: .seealso: [](ch_matrices), `Mat`, `MatSeqAIJGetArray()`, `MatSeqAIJRestoreArrayRead()`
4482: @*/
4483: PetscErrorCode MatSeqAIJGetArrayRead(Mat A, const PetscScalar *array[])
4484: {
4485: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)A->data;
4487: PetscFunctionBegin;
4488: if (aij->ops->getarrayread) {
4489: PetscCall((*aij->ops->getarrayread)(A, array));
4490: } else {
4491: *array = aij->a;
4492: }
4493: PetscFunctionReturn(PETSC_SUCCESS);
4494: }
4496: /*@C
4497: MatSeqAIJRestoreArrayRead - restore the read-only access array obtained from `MatSeqAIJGetArrayRead()`
4499: Not Collective; No Fortran Support
4501: Input Parameter:
4502: . A - a `MATSEQAIJ` matrix
4504: Output Parameter:
4505: . array - pointer to the data
4507: Level: intermediate
4509: .seealso: [](ch_matrices), `Mat`, `MatSeqAIJGetArray()`, `MatSeqAIJGetArrayRead()`
4510: @*/
4511: PetscErrorCode MatSeqAIJRestoreArrayRead(Mat A, const PetscScalar *array[])
4512: {
4513: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)A->data;
4515: PetscFunctionBegin;
4516: if (aij->ops->restorearrayread) {
4517: PetscCall((*aij->ops->restorearrayread)(A, array));
4518: } else {
4519: *array = NULL;
4520: }
4521: PetscFunctionReturn(PETSC_SUCCESS);
4522: }
4524: /*@C
4525: MatSeqAIJGetArrayWrite - gives write-only access to the array where the data for a `MATSEQAIJ` matrix is stored
4527: Not Collective; No Fortran Support
4529: Input Parameter:
4530: . A - a `MATSEQAIJ` matrix
4532: Output Parameter:
4533: . array - pointer to the data
4535: Level: intermediate
4537: .seealso: [](ch_matrices), `Mat`, `MatSeqAIJGetArray()`, `MatSeqAIJRestoreArrayRead()`
4538: @*/
4539: PetscErrorCode MatSeqAIJGetArrayWrite(Mat A, PetscScalar *array[])
4540: {
4541: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)A->data;
4543: PetscFunctionBegin;
4544: if (aij->ops->getarraywrite) {
4545: PetscCall((*aij->ops->getarraywrite)(A, array));
4546: } else {
4547: *array = aij->a;
4548: }
4549: PetscCall(MatSeqAIJInvalidateDiagonal(A));
4550: PetscCall(PetscObjectStateIncrease((PetscObject)A));
4551: PetscFunctionReturn(PETSC_SUCCESS);
4552: }
4554: /*@C
4555: MatSeqAIJRestoreArrayWrite - restore the read-only access array obtained from MatSeqAIJGetArrayRead
4557: Not Collective; No Fortran Support
4559: Input Parameter:
4560: . A - a MATSEQAIJ matrix
4562: Output Parameter:
4563: . array - pointer to the data
4565: Level: intermediate
4567: .seealso: [](ch_matrices), `Mat`, `MatSeqAIJGetArray()`, `MatSeqAIJGetArrayRead()`
4568: @*/
4569: PetscErrorCode MatSeqAIJRestoreArrayWrite(Mat A, PetscScalar *array[])
4570: {
4571: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)A->data;
4573: PetscFunctionBegin;
4574: if (aij->ops->restorearraywrite) {
4575: PetscCall((*aij->ops->restorearraywrite)(A, array));
4576: } else {
4577: *array = NULL;
4578: }
4579: PetscFunctionReturn(PETSC_SUCCESS);
4580: }
4582: /*@C
4583: MatSeqAIJGetCSRAndMemType - Get the CSR arrays and the memory type of the `MATSEQAIJ` matrix
4585: Not Collective; No Fortran Support
4587: Input Parameter:
4588: . mat - a matrix of type `MATSEQAIJ` or its subclasses
4590: Output Parameters:
4591: + i - row map array of the matrix
4592: . j - column index array of the matrix
4593: . a - data array of the matrix
4594: - mtype - memory type of the arrays
4596: Level: developer
4598: Notes:
4599: Any of the output parameters can be `NULL`, in which case the corresponding value is not returned.
4600: If mat is a device matrix, the arrays are on the device. Otherwise, they are on the host.
4602: One can call this routine on a preallocated but not assembled matrix to just get the memory of the CSR underneath the matrix.
4603: If the matrix is assembled, the data array `a` is guaranteed to have the latest values of the matrix.
4605: .seealso: [](ch_matrices), `Mat`, `MatSeqAIJGetArray()`, `MatSeqAIJGetArrayRead()`
4606: @*/
4607: PetscErrorCode MatSeqAIJGetCSRAndMemType(Mat mat, const PetscInt *i[], const PetscInt *j[], PetscScalar *a[], PetscMemType *mtype)
4608: {
4609: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)mat->data;
4611: PetscFunctionBegin;
4612: PetscCheck(mat->preallocated, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "matrix is not preallocated");
4613: if (aij->ops->getcsrandmemtype) {
4614: PetscCall((*aij->ops->getcsrandmemtype)(mat, i, j, a, mtype));
4615: } else {
4616: if (i) *i = aij->i;
4617: if (j) *j = aij->j;
4618: if (a) *a = aij->a;
4619: if (mtype) *mtype = PETSC_MEMTYPE_HOST;
4620: }
4621: PetscFunctionReturn(PETSC_SUCCESS);
4622: }
4624: /*@
4625: MatSeqAIJGetMaxRowNonzeros - returns the maximum number of nonzeros in any row
4627: Not Collective
4629: Input Parameter:
4630: . A - a `MATSEQAIJ` matrix
4632: Output Parameter:
4633: . nz - the maximum number of nonzeros in any row
4635: Level: intermediate
4637: .seealso: [](ch_matrices), `Mat`, `MatSeqAIJRestoreArray()`, `MatSeqAIJGetArrayF90()`
4638: @*/
4639: PetscErrorCode MatSeqAIJGetMaxRowNonzeros(Mat A, PetscInt *nz)
4640: {
4641: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)A->data;
4643: PetscFunctionBegin;
4644: *nz = aij->rmax;
4645: PetscFunctionReturn(PETSC_SUCCESS);
4646: }
4648: static PetscErrorCode MatCOOStructDestroy_SeqAIJ(void **data)
4649: {
4650: MatCOOStruct_SeqAIJ *coo = (MatCOOStruct_SeqAIJ *)*data;
4652: PetscFunctionBegin;
4653: PetscCall(PetscFree(coo->perm));
4654: PetscCall(PetscFree(coo->jmap));
4655: PetscCall(PetscFree(coo));
4656: PetscFunctionReturn(PETSC_SUCCESS);
4657: }
4659: PetscErrorCode MatSetPreallocationCOO_SeqAIJ(Mat mat, PetscCount coo_n, PetscInt coo_i[], PetscInt coo_j[])
4660: {
4661: MPI_Comm comm;
4662: PetscInt *i, *j;
4663: PetscInt M, N, row, iprev;
4664: PetscCount k, p, q, nneg, nnz, start, end; /* Index the coo array, so use PetscCount as their type */
4665: PetscInt *Ai; /* Change to PetscCount once we use it for row pointers */
4666: PetscInt *Aj;
4667: PetscScalar *Aa;
4668: Mat_SeqAIJ *seqaij = (Mat_SeqAIJ *)mat->data;
4669: MatType rtype;
4670: PetscCount *perm, *jmap;
4671: MatCOOStruct_SeqAIJ *coo;
4672: PetscBool isorted;
4673: PetscBool hypre;
4674: const char *name;
4676: PetscFunctionBegin;
4677: PetscCall(PetscObjectGetComm((PetscObject)mat, &comm));
4678: PetscCall(MatGetSize(mat, &M, &N));
4679: i = coo_i;
4680: j = coo_j;
4681: PetscCall(PetscMalloc1(coo_n, &perm));
4683: /* Ignore entries with negative row or col indices; at the same time, check if i[] is already sorted (e.g., MatConvert_AlJ_HYPRE results in this case) */
4684: isorted = PETSC_TRUE;
4685: iprev = PETSC_INT_MIN;
4686: for (k = 0; k < coo_n; k++) {
4687: if (j[k] < 0) i[k] = -1;
4688: if (isorted) {
4689: if (i[k] < iprev) isorted = PETSC_FALSE;
4690: else iprev = i[k];
4691: }
4692: perm[k] = k;
4693: }
4695: /* Sort by row if not already */
4696: if (!isorted) PetscCall(PetscSortIntWithIntCountArrayPair(coo_n, i, j, perm));
4698: /* Advance k to the first row with a non-negative index */
4699: for (k = 0; k < coo_n; k++)
4700: if (i[k] >= 0) break;
4701: nneg = k;
4702: PetscCall(PetscMalloc1(coo_n - nneg + 1, &jmap)); /* +1 to make a CSR-like data structure. jmap[i] originally is the number of repeats for i-th nonzero */
4703: nnz = 0; /* Total number of unique nonzeros to be counted */
4704: jmap++; /* Inc jmap by 1 for convenience */
4706: PetscCall(PetscShmgetAllocateArray(M + 1, sizeof(PetscInt), (void **)&Ai)); /* CSR of A */
4707: PetscCall(PetscArrayzero(Ai, M + 1));
4708: PetscCall(PetscShmgetAllocateArray(coo_n - nneg, sizeof(PetscInt), (void **)&Aj)); /* We have at most coo_n-nneg unique nonzeros */
4710: PetscCall(PetscObjectGetName((PetscObject)mat, &name));
4711: PetscCall(PetscStrcmp("_internal_COO_mat_for_hypre", name, &hypre));
4713: /* In each row, sort by column, then unique column indices to get row length */
4714: Ai++; /* Inc by 1 for convenience */
4715: q = 0; /* q-th unique nonzero, with q starting from 0 */
4716: while (k < coo_n) {
4717: PetscBool strictly_sorted; // this row is strictly sorted?
4718: PetscInt jprev;
4720: /* get [start,end) indices for this row; also check if cols in this row are strictly sorted */
4721: row = i[k];
4722: start = k;
4723: jprev = PETSC_INT_MIN;
4724: strictly_sorted = PETSC_TRUE;
4725: while (k < coo_n && i[k] == row) {
4726: if (strictly_sorted) {
4727: if (j[k] <= jprev) strictly_sorted = PETSC_FALSE;
4728: else jprev = j[k];
4729: }
4730: k++;
4731: }
4732: end = k;
4734: /* hack for HYPRE: swap min column to diag so that diagonal values will go first */
4735: if (hypre) {
4736: PetscInt minj = PETSC_INT_MAX;
4737: PetscBool hasdiag = PETSC_FALSE;
4739: if (strictly_sorted) { // fast path to swap the first and the diag
4740: PetscCount tmp;
4741: for (p = start; p < end; p++) {
4742: if (j[p] == row && p != start) {
4743: j[p] = j[start]; // swap j[], so that the diagonal value will go first (manipulated by perm[])
4744: j[start] = row;
4745: tmp = perm[start];
4746: perm[start] = perm[p]; // also swap perm[] so we can save the call to PetscSortIntWithCountArray() below
4747: perm[p] = tmp;
4748: break;
4749: }
4750: }
4751: } else {
4752: for (p = start; p < end; p++) {
4753: hasdiag = (PetscBool)(hasdiag || (j[p] == row));
4754: minj = PetscMin(minj, j[p]);
4755: }
4757: if (hasdiag) {
4758: for (p = start; p < end; p++) {
4759: if (j[p] == minj) j[p] = row;
4760: else if (j[p] == row) j[p] = minj;
4761: }
4762: }
4763: }
4764: }
4765: // sort by columns in a row. perm[] indicates their original order
4766: if (!strictly_sorted) PetscCall(PetscSortIntWithCountArray(end - start, j + start, perm + start));
4768: if (strictly_sorted) { // fast path to set Aj[], jmap[], Ai[], nnz, q
4769: for (p = start; p < end; p++, q++) {
4770: Aj[q] = j[p];
4771: jmap[q] = 1;
4772: }
4773: PetscCall(PetscIntCast(end - start, Ai + row));
4774: nnz += Ai[row]; // q is already advanced
4775: } else {
4776: /* Find number of unique col entries in this row */
4777: Aj[q] = j[start]; /* Log the first nonzero in this row */
4778: jmap[q] = 1; /* Number of repeats of this nonzero entry */
4779: Ai[row] = 1;
4780: nnz++;
4782: for (p = start + 1; p < end; p++) { /* Scan remaining nonzero in this row */
4783: if (j[p] != j[p - 1]) { /* Meet a new nonzero */
4784: q++;
4785: jmap[q] = 1;
4786: Aj[q] = j[p];
4787: Ai[row]++;
4788: nnz++;
4789: } else {
4790: jmap[q]++;
4791: }
4792: }
4793: q++; /* Move to next row and thus next unique nonzero */
4794: }
4795: }
4797: Ai--; /* Back to the beginning of Ai[] */
4798: for (k = 0; k < M; k++) Ai[k + 1] += Ai[k];
4799: jmap--; // Back to the beginning of jmap[]
4800: jmap[0] = 0;
4801: for (k = 0; k < nnz; k++) jmap[k + 1] += jmap[k];
4803: if (nnz < coo_n - nneg) { /* Reallocate with actual number of unique nonzeros */
4804: PetscCount *jmap_new;
4805: PetscInt *Aj_new;
4807: PetscCall(PetscMalloc1(nnz + 1, &jmap_new));
4808: PetscCall(PetscArraycpy(jmap_new, jmap, nnz + 1));
4809: PetscCall(PetscFree(jmap));
4810: jmap = jmap_new;
4812: PetscCall(PetscShmgetAllocateArray(nnz, sizeof(PetscInt), (void **)&Aj_new));
4813: PetscCall(PetscArraycpy(Aj_new, Aj, nnz));
4814: PetscCall(PetscShmgetDeallocateArray((void **)&Aj));
4815: Aj = Aj_new;
4816: }
4818: if (nneg) { /* Discard heading entries with negative indices in perm[], as we'll access it from index 0 in MatSetValuesCOO */
4819: PetscCount *perm_new;
4821: PetscCall(PetscMalloc1(coo_n - nneg, &perm_new));
4822: PetscCall(PetscArraycpy(perm_new, perm + nneg, coo_n - nneg));
4823: PetscCall(PetscFree(perm));
4824: perm = perm_new;
4825: }
4827: PetscCall(MatGetRootType_Private(mat, &rtype));
4828: PetscCall(PetscShmgetAllocateArray(nnz, sizeof(PetscScalar), (void **)&Aa));
4829: PetscCall(PetscArrayzero(Aa, nnz));
4830: PetscCall(MatSetSeqAIJWithArrays_private(PETSC_COMM_SELF, M, N, Ai, Aj, Aa, rtype, mat));
4832: seqaij->free_a = seqaij->free_ij = PETSC_TRUE; /* Let newmat own Ai, Aj and Aa */
4834: // Put the COO struct in a container and then attach that to the matrix
4835: PetscCall(PetscMalloc1(1, &coo));
4836: PetscCall(PetscIntCast(nnz, &coo->nz));
4837: coo->n = coo_n;
4838: coo->Atot = coo_n - nneg; // Annz is seqaij->nz, so no need to record that again
4839: coo->jmap = jmap; // of length nnz+1
4840: coo->perm = perm;
4841: PetscCall(PetscObjectContainerCompose((PetscObject)mat, "__PETSc_MatCOOStruct_Host", coo, MatCOOStructDestroy_SeqAIJ));
4842: PetscFunctionReturn(PETSC_SUCCESS);
4843: }
4845: static PetscErrorCode MatSetValuesCOO_SeqAIJ(Mat A, const PetscScalar v[], InsertMode imode)
4846: {
4847: Mat_SeqAIJ *aseq = (Mat_SeqAIJ *)A->data;
4848: PetscCount i, j, Annz = aseq->nz;
4849: PetscCount *perm, *jmap;
4850: PetscScalar *Aa;
4851: PetscContainer container;
4852: MatCOOStruct_SeqAIJ *coo;
4854: PetscFunctionBegin;
4855: PetscCall(PetscObjectQuery((PetscObject)A, "__PETSc_MatCOOStruct_Host", (PetscObject *)&container));
4856: PetscCheck(container, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Not found MatCOOStruct on this matrix");
4857: PetscCall(PetscContainerGetPointer(container, (void **)&coo));
4858: perm = coo->perm;
4859: jmap = coo->jmap;
4860: PetscCall(MatSeqAIJGetArray(A, &Aa));
4861: for (i = 0; i < Annz; i++) {
4862: PetscScalar sum = 0.0;
4863: for (j = jmap[i]; j < jmap[i + 1]; j++) sum += v[perm[j]];
4864: Aa[i] = (imode == INSERT_VALUES ? 0.0 : Aa[i]) + sum;
4865: }
4866: PetscCall(MatSeqAIJRestoreArray(A, &Aa));
4867: PetscFunctionReturn(PETSC_SUCCESS);
4868: }
4870: #if defined(PETSC_HAVE_CUDA)
4871: PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqAIJCUSPARSE(Mat, MatType, MatReuse, Mat *);
4872: #endif
4873: #if defined(PETSC_HAVE_HIP)
4874: PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqAIJHIPSPARSE(Mat, MatType, MatReuse, Mat *);
4875: #endif
4876: #if defined(PETSC_HAVE_KOKKOS_KERNELS)
4877: PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqAIJKokkos(Mat, MatType, MatReuse, Mat *);
4878: #endif
4880: PETSC_EXTERN PetscErrorCode MatCreate_SeqAIJ(Mat B)
4881: {
4882: Mat_SeqAIJ *b;
4883: PetscMPIInt size;
4885: PetscFunctionBegin;
4886: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)B), &size));
4887: PetscCheck(size <= 1, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Comm must be of size 1");
4889: PetscCall(PetscNew(&b));
4891: B->data = (void *)b;
4892: B->ops[0] = MatOps_Values;
4893: if (B->sortedfull) B->ops->setvalues = MatSetValues_SeqAIJ_SortedFull;
4895: b->row = NULL;
4896: b->col = NULL;
4897: b->icol = NULL;
4898: b->reallocs = 0;
4899: b->ignorezeroentries = PETSC_FALSE;
4900: b->roworiented = PETSC_TRUE;
4901: b->nonew = 0;
4902: b->diag = NULL;
4903: b->solve_work = NULL;
4904: B->spptr = NULL;
4905: b->saved_values = NULL;
4906: b->idiag = NULL;
4907: b->mdiag = NULL;
4908: b->ssor_work = NULL;
4909: b->omega = 1.0;
4910: b->fshift = 0.0;
4911: b->idiagvalid = PETSC_FALSE;
4912: b->ibdiagvalid = PETSC_FALSE;
4913: b->keepnonzeropattern = PETSC_FALSE;
4915: PetscCall(PetscObjectChangeTypeName((PetscObject)B, MATSEQAIJ));
4916: #if defined(PETSC_HAVE_MATLAB)
4917: PetscCall(PetscObjectComposeFunction((PetscObject)B, "PetscMatlabEnginePut_C", MatlabEnginePut_SeqAIJ));
4918: PetscCall(PetscObjectComposeFunction((PetscObject)B, "PetscMatlabEngineGet_C", MatlabEngineGet_SeqAIJ));
4919: #endif
4920: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatSeqAIJSetColumnIndices_C", MatSeqAIJSetColumnIndices_SeqAIJ));
4921: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatStoreValues_C", MatStoreValues_SeqAIJ));
4922: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatRetrieveValues_C", MatRetrieveValues_SeqAIJ));
4923: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_seqsbaij_C", MatConvert_SeqAIJ_SeqSBAIJ));
4924: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_seqbaij_C", MatConvert_SeqAIJ_SeqBAIJ));
4925: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_seqaijperm_C", MatConvert_SeqAIJ_SeqAIJPERM));
4926: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_seqaijsell_C", MatConvert_SeqAIJ_SeqAIJSELL));
4927: #if defined(PETSC_HAVE_MKL_SPARSE)
4928: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_seqaijmkl_C", MatConvert_SeqAIJ_SeqAIJMKL));
4929: #endif
4930: #if defined(PETSC_HAVE_CUDA)
4931: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_seqaijcusparse_C", MatConvert_SeqAIJ_SeqAIJCUSPARSE));
4932: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_seqaijcusparse_seqaij_C", MatProductSetFromOptions_SeqAIJ));
4933: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_seqaij_seqaijcusparse_C", MatProductSetFromOptions_SeqAIJ));
4934: #endif
4935: #if defined(PETSC_HAVE_HIP)
4936: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_seqaijhipsparse_C", MatConvert_SeqAIJ_SeqAIJHIPSPARSE));
4937: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_seqaijhipsparse_seqaij_C", MatProductSetFromOptions_SeqAIJ));
4938: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_seqaij_seqaijhipsparse_C", MatProductSetFromOptions_SeqAIJ));
4939: #endif
4940: #if defined(PETSC_HAVE_KOKKOS_KERNELS)
4941: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_seqaijkokkos_C", MatConvert_SeqAIJ_SeqAIJKokkos));
4942: #endif
4943: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_seqaijcrl_C", MatConvert_SeqAIJ_SeqAIJCRL));
4944: #if defined(PETSC_HAVE_ELEMENTAL)
4945: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_elemental_C", MatConvert_SeqAIJ_Elemental));
4946: #endif
4947: #if defined(PETSC_HAVE_SCALAPACK)
4948: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_scalapack_C", MatConvert_AIJ_ScaLAPACK));
4949: #endif
4950: #if defined(PETSC_HAVE_HYPRE)
4951: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_hypre_C", MatConvert_AIJ_HYPRE));
4952: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_transpose_seqaij_seqaij_C", MatProductSetFromOptions_Transpose_AIJ_AIJ));
4953: #endif
4954: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_seqdense_C", MatConvert_SeqAIJ_SeqDense));
4955: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_seqsell_C", MatConvert_SeqAIJ_SeqSELL));
4956: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_is_C", MatConvert_XAIJ_IS));
4957: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatIsTranspose_C", MatIsTranspose_SeqAIJ));
4958: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatIsHermitianTranspose_C", MatIsHermitianTranspose_SeqAIJ));
4959: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatSeqAIJSetPreallocation_C", MatSeqAIJSetPreallocation_SeqAIJ));
4960: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatResetPreallocation_C", MatResetPreallocation_SeqAIJ));
4961: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatResetHash_C", MatResetHash_SeqAIJ));
4962: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatSeqAIJSetPreallocationCSR_C", MatSeqAIJSetPreallocationCSR_SeqAIJ));
4963: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatReorderForNonzeroDiagonal_C", MatReorderForNonzeroDiagonal_SeqAIJ));
4964: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_is_seqaij_C", MatProductSetFromOptions_IS_XAIJ));
4965: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_seqdense_seqaij_C", MatProductSetFromOptions_SeqDense_SeqAIJ));
4966: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_seqaij_seqaij_C", MatProductSetFromOptions_SeqAIJ));
4967: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatSeqAIJKron_C", MatSeqAIJKron_SeqAIJ));
4968: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatSetPreallocationCOO_C", MatSetPreallocationCOO_SeqAIJ));
4969: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatSetValuesCOO_C", MatSetValuesCOO_SeqAIJ));
4970: PetscCall(MatCreate_SeqAIJ_Inode(B));
4971: PetscCall(PetscObjectChangeTypeName((PetscObject)B, MATSEQAIJ));
4972: PetscCall(MatSeqAIJSetTypeFromOptions(B)); /* this allows changing the matrix subtype to say MATSEQAIJPERM */
4973: PetscFunctionReturn(PETSC_SUCCESS);
4974: }
4976: /*
4977: Given a matrix generated with MatGetFactor() duplicates all the information in A into C
4978: */
4979: PetscErrorCode MatDuplicateNoCreate_SeqAIJ(Mat C, Mat A, MatDuplicateOption cpvalues, PetscBool mallocmatspace)
4980: {
4981: Mat_SeqAIJ *c = (Mat_SeqAIJ *)C->data, *a = (Mat_SeqAIJ *)A->data;
4982: PetscInt m = A->rmap->n, i;
4984: PetscFunctionBegin;
4985: PetscCheck(A->assembled || cpvalues == MAT_DO_NOT_COPY_VALUES, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Cannot duplicate unassembled matrix");
4987: C->factortype = A->factortype;
4988: c->row = NULL;
4989: c->col = NULL;
4990: c->icol = NULL;
4991: c->reallocs = 0;
4992: c->diagonaldense = a->diagonaldense;
4994: C->assembled = A->assembled;
4996: if (A->preallocated) {
4997: PetscCall(PetscLayoutReference(A->rmap, &C->rmap));
4998: PetscCall(PetscLayoutReference(A->cmap, &C->cmap));
5000: if (!A->hash_active) {
5001: PetscCall(PetscMalloc1(m, &c->imax));
5002: PetscCall(PetscMemcpy(c->imax, a->imax, m * sizeof(PetscInt)));
5003: PetscCall(PetscMalloc1(m, &c->ilen));
5004: PetscCall(PetscMemcpy(c->ilen, a->ilen, m * sizeof(PetscInt)));
5006: /* allocate the matrix space */
5007: if (mallocmatspace) {
5008: PetscCall(PetscShmgetAllocateArray(a->i[m], sizeof(PetscScalar), (void **)&c->a));
5009: PetscCall(PetscShmgetAllocateArray(a->i[m], sizeof(PetscInt), (void **)&c->j));
5010: PetscCall(PetscShmgetAllocateArray(m + 1, sizeof(PetscInt), (void **)&c->i));
5011: PetscCall(PetscArraycpy(c->i, a->i, m + 1));
5012: c->free_a = PETSC_TRUE;
5013: c->free_ij = PETSC_TRUE;
5014: if (m > 0) {
5015: PetscCall(PetscArraycpy(c->j, a->j, a->i[m]));
5016: if (cpvalues == MAT_COPY_VALUES) {
5017: const PetscScalar *aa;
5019: PetscCall(MatSeqAIJGetArrayRead(A, &aa));
5020: PetscCall(PetscArraycpy(c->a, aa, a->i[m]));
5021: PetscCall(MatSeqAIJGetArrayRead(A, &aa));
5022: } else {
5023: PetscCall(PetscArrayzero(c->a, a->i[m]));
5024: }
5025: }
5026: }
5027: C->preallocated = PETSC_TRUE;
5028: } else {
5029: PetscCheck(mallocmatspace, PetscObjectComm((PetscObject)A), PETSC_ERR_ARG_WRONGSTATE, "Cannot malloc matrix memory from a non-preallocated matrix");
5030: PetscCall(MatSetUp(C));
5031: }
5033: c->ignorezeroentries = a->ignorezeroentries;
5034: c->roworiented = a->roworiented;
5035: c->nonew = a->nonew;
5036: if (a->diag) {
5037: PetscCall(PetscMalloc1(m + 1, &c->diag));
5038: PetscCall(PetscMemcpy(c->diag, a->diag, m * sizeof(PetscInt)));
5039: } else c->diag = NULL;
5041: c->solve_work = NULL;
5042: c->saved_values = NULL;
5043: c->idiag = NULL;
5044: c->ssor_work = NULL;
5045: c->keepnonzeropattern = a->keepnonzeropattern;
5047: c->rmax = a->rmax;
5048: c->nz = a->nz;
5049: c->maxnz = a->nz; /* Since we allocate exactly the right amount */
5051: c->compressedrow.use = a->compressedrow.use;
5052: c->compressedrow.nrows = a->compressedrow.nrows;
5053: if (a->compressedrow.use) {
5054: i = a->compressedrow.nrows;
5055: PetscCall(PetscMalloc2(i + 1, &c->compressedrow.i, i, &c->compressedrow.rindex));
5056: PetscCall(PetscArraycpy(c->compressedrow.i, a->compressedrow.i, i + 1));
5057: PetscCall(PetscArraycpy(c->compressedrow.rindex, a->compressedrow.rindex, i));
5058: } else {
5059: c->compressedrow.use = PETSC_FALSE;
5060: c->compressedrow.i = NULL;
5061: c->compressedrow.rindex = NULL;
5062: }
5063: c->nonzerorowcnt = a->nonzerorowcnt;
5064: C->nonzerostate = A->nonzerostate;
5066: PetscCall(MatDuplicate_SeqAIJ_Inode(A, cpvalues, &C));
5067: }
5068: PetscCall(PetscFunctionListDuplicate(((PetscObject)A)->qlist, &((PetscObject)C)->qlist));
5069: PetscFunctionReturn(PETSC_SUCCESS);
5070: }
5072: PetscErrorCode MatDuplicate_SeqAIJ(Mat A, MatDuplicateOption cpvalues, Mat *B)
5073: {
5074: PetscFunctionBegin;
5075: PetscCall(MatCreate(PetscObjectComm((PetscObject)A), B));
5076: PetscCall(MatSetSizes(*B, A->rmap->n, A->cmap->n, A->rmap->n, A->cmap->n));
5077: if (!(A->rmap->n % A->rmap->bs) && !(A->cmap->n % A->cmap->bs)) PetscCall(MatSetBlockSizesFromMats(*B, A, A));
5078: PetscCall(MatSetType(*B, ((PetscObject)A)->type_name));
5079: PetscCall(MatDuplicateNoCreate_SeqAIJ(*B, A, cpvalues, PETSC_TRUE));
5080: PetscFunctionReturn(PETSC_SUCCESS);
5081: }
5083: PetscErrorCode MatLoad_SeqAIJ(Mat newMat, PetscViewer viewer)
5084: {
5085: PetscBool isbinary, ishdf5;
5087: PetscFunctionBegin;
5090: /* force binary viewer to load .info file if it has not yet done so */
5091: PetscCall(PetscViewerSetUp(viewer));
5092: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERBINARY, &isbinary));
5093: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERHDF5, &ishdf5));
5094: if (isbinary) {
5095: PetscCall(MatLoad_SeqAIJ_Binary(newMat, viewer));
5096: } else if (ishdf5) {
5097: #if defined(PETSC_HAVE_HDF5)
5098: PetscCall(MatLoad_AIJ_HDF5(newMat, viewer));
5099: #else
5100: SETERRQ(PetscObjectComm((PetscObject)newMat), PETSC_ERR_SUP, "HDF5 not supported in this build.\nPlease reconfigure using --download-hdf5");
5101: #endif
5102: } else {
5103: 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);
5104: }
5105: PetscFunctionReturn(PETSC_SUCCESS);
5106: }
5108: PetscErrorCode MatLoad_SeqAIJ_Binary(Mat mat, PetscViewer viewer)
5109: {
5110: Mat_SeqAIJ *a = (Mat_SeqAIJ *)mat->data;
5111: PetscInt header[4], *rowlens, M, N, nz, sum, rows, cols, i;
5113: PetscFunctionBegin;
5114: PetscCall(PetscViewerSetUp(viewer));
5116: /* read in matrix header */
5117: PetscCall(PetscViewerBinaryRead(viewer, header, 4, NULL, PETSC_INT));
5118: PetscCheck(header[0] == MAT_FILE_CLASSID, PETSC_COMM_SELF, PETSC_ERR_FILE_UNEXPECTED, "Not a matrix object in file");
5119: M = header[1];
5120: N = header[2];
5121: nz = header[3];
5122: PetscCheck(M >= 0, PetscObjectComm((PetscObject)viewer), PETSC_ERR_FILE_UNEXPECTED, "Matrix row size (%" PetscInt_FMT ") in file is negative", M);
5123: PetscCheck(N >= 0, PetscObjectComm((PetscObject)viewer), PETSC_ERR_FILE_UNEXPECTED, "Matrix column size (%" PetscInt_FMT ") in file is negative", N);
5124: PetscCheck(nz >= 0, PETSC_COMM_SELF, PETSC_ERR_FILE_UNEXPECTED, "Matrix stored in special format on disk, cannot load as SeqAIJ");
5126: /* set block sizes from the viewer's .info file */
5127: PetscCall(MatLoad_Binary_BlockSizes(mat, viewer));
5128: /* set local and global sizes if not set already */
5129: if (mat->rmap->n < 0) mat->rmap->n = M;
5130: if (mat->cmap->n < 0) mat->cmap->n = N;
5131: if (mat->rmap->N < 0) mat->rmap->N = M;
5132: if (mat->cmap->N < 0) mat->cmap->N = N;
5133: PetscCall(PetscLayoutSetUp(mat->rmap));
5134: PetscCall(PetscLayoutSetUp(mat->cmap));
5136: /* check if the matrix sizes are correct */
5137: PetscCall(MatGetSize(mat, &rows, &cols));
5138: PetscCheck(M == rows && N == cols, PETSC_COMM_SELF, PETSC_ERR_FILE_UNEXPECTED, "Matrix in file of different sizes (%" PetscInt_FMT ", %" PetscInt_FMT ") than the input matrix (%" PetscInt_FMT ", %" PetscInt_FMT ")", M, N, rows, cols);
5140: /* read in row lengths */
5141: PetscCall(PetscMalloc1(M, &rowlens));
5142: PetscCall(PetscViewerBinaryRead(viewer, rowlens, M, NULL, PETSC_INT));
5143: /* check if sum(rowlens) is same as nz */
5144: sum = 0;
5145: for (i = 0; i < M; i++) sum += rowlens[i];
5146: PetscCheck(sum == nz, PETSC_COMM_SELF, PETSC_ERR_FILE_UNEXPECTED, "Inconsistent matrix data in file: nonzeros = %" PetscInt_FMT ", sum-row-lengths = %" PetscInt_FMT, nz, sum);
5147: /* preallocate and check sizes */
5148: PetscCall(MatSeqAIJSetPreallocation_SeqAIJ(mat, 0, rowlens));
5149: PetscCall(MatGetSize(mat, &rows, &cols));
5150: PetscCheck(M == rows && N == cols, PETSC_COMM_SELF, PETSC_ERR_FILE_UNEXPECTED, "Matrix in file of different length (%" PetscInt_FMT ", %" PetscInt_FMT ") than the input matrix (%" PetscInt_FMT ", %" PetscInt_FMT ")", M, N, rows, cols);
5151: /* store row lengths */
5152: PetscCall(PetscArraycpy(a->ilen, rowlens, M));
5153: PetscCall(PetscFree(rowlens));
5155: /* fill in "i" row pointers */
5156: a->i[0] = 0;
5157: for (i = 0; i < M; i++) a->i[i + 1] = a->i[i] + a->ilen[i];
5158: /* read in "j" column indices */
5159: PetscCall(PetscViewerBinaryRead(viewer, a->j, nz, NULL, PETSC_INT));
5160: /* read in "a" nonzero values */
5161: PetscCall(PetscViewerBinaryRead(viewer, a->a, nz, NULL, PETSC_SCALAR));
5163: PetscCall(MatAssemblyBegin(mat, MAT_FINAL_ASSEMBLY));
5164: PetscCall(MatAssemblyEnd(mat, MAT_FINAL_ASSEMBLY));
5165: PetscFunctionReturn(PETSC_SUCCESS);
5166: }
5168: PetscErrorCode MatEqual_SeqAIJ(Mat A, Mat B, PetscBool *flg)
5169: {
5170: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data, *b = (Mat_SeqAIJ *)B->data;
5171: const PetscScalar *aa, *ba;
5172: #if defined(PETSC_USE_COMPLEX)
5173: PetscInt k;
5174: #endif
5176: PetscFunctionBegin;
5177: /* If the matrix dimensions are not equal,or no of nonzeros */
5178: if ((A->rmap->n != B->rmap->n) || (A->cmap->n != B->cmap->n) || (a->nz != b->nz)) {
5179: *flg = PETSC_FALSE;
5180: PetscFunctionReturn(PETSC_SUCCESS);
5181: }
5183: /* if the a->i are the same */
5184: PetscCall(PetscArraycmp(a->i, b->i, A->rmap->n + 1, flg));
5185: if (!*flg) PetscFunctionReturn(PETSC_SUCCESS);
5187: /* if a->j are the same */
5188: PetscCall(PetscArraycmp(a->j, b->j, a->nz, flg));
5189: if (!*flg) PetscFunctionReturn(PETSC_SUCCESS);
5191: PetscCall(MatSeqAIJGetArrayRead(A, &aa));
5192: PetscCall(MatSeqAIJGetArrayRead(B, &ba));
5193: /* if a->a are the same */
5194: #if defined(PETSC_USE_COMPLEX)
5195: for (k = 0; k < a->nz; k++) {
5196: if (PetscRealPart(aa[k]) != PetscRealPart(ba[k]) || PetscImaginaryPart(aa[k]) != PetscImaginaryPart(ba[k])) {
5197: *flg = PETSC_FALSE;
5198: PetscFunctionReturn(PETSC_SUCCESS);
5199: }
5200: }
5201: #else
5202: PetscCall(PetscArraycmp(aa, ba, a->nz, flg));
5203: #endif
5204: PetscCall(MatSeqAIJRestoreArrayRead(A, &aa));
5205: PetscCall(MatSeqAIJRestoreArrayRead(B, &ba));
5206: PetscFunctionReturn(PETSC_SUCCESS);
5207: }
5209: /*@
5210: MatCreateSeqAIJWithArrays - Creates an sequential `MATSEQAIJ` matrix using matrix elements (in CSR format)
5211: provided by the user.
5213: Collective
5215: Input Parameters:
5216: + comm - must be an MPI communicator of size 1
5217: . m - number of rows
5218: . n - number of columns
5219: . i - row indices; that is i[0] = 0, i[row] = i[row-1] + number of elements in that row of the matrix
5220: . j - column indices
5221: - a - matrix values
5223: Output Parameter:
5224: . mat - the matrix
5226: Level: intermediate
5228: Notes:
5229: The `i`, `j`, and `a` arrays are not copied by this routine, the user must free these arrays
5230: once the matrix is destroyed and not before
5232: You cannot set new nonzero locations into this matrix, that will generate an error.
5234: The `i` and `j` indices are 0 based
5236: The format which is used for the sparse matrix input, is equivalent to a
5237: row-major ordering.. i.e for the following matrix, the input data expected is
5238: as shown
5239: .vb
5240: 1 0 0
5241: 2 0 3
5242: 4 5 6
5244: i = {0,1,3,6} [size = nrow+1 = 3+1]
5245: j = {0,0,2,0,1,2} [size = 6]; values must be sorted for each row
5246: v = {1,2,3,4,5,6} [size = 6]
5247: .ve
5249: .seealso: [](ch_matrices), `Mat`, `MatCreate()`, `MatCreateAIJ()`, `MatCreateSeqAIJ()`, `MatCreateMPIAIJWithArrays()`, `MatMPIAIJSetPreallocationCSR()`
5250: @*/
5251: PetscErrorCode MatCreateSeqAIJWithArrays(MPI_Comm comm, PetscInt m, PetscInt n, PetscInt i[], PetscInt j[], PetscScalar a[], Mat *mat)
5252: {
5253: PetscInt ii;
5254: Mat_SeqAIJ *aij;
5255: PetscInt jj;
5257: PetscFunctionBegin;
5258: PetscCheck(m <= 0 || i[0] == 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "i (row indices) must start with 0");
5259: PetscCall(MatCreate(comm, mat));
5260: PetscCall(MatSetSizes(*mat, m, n, m, n));
5261: /* PetscCall(MatSetBlockSizes(*mat,,)); */
5262: PetscCall(MatSetType(*mat, MATSEQAIJ));
5263: PetscCall(MatSeqAIJSetPreallocation_SeqAIJ(*mat, MAT_SKIP_ALLOCATION, NULL));
5264: aij = (Mat_SeqAIJ *)(*mat)->data;
5265: PetscCall(PetscMalloc1(m, &aij->imax));
5266: PetscCall(PetscMalloc1(m, &aij->ilen));
5268: aij->i = i;
5269: aij->j = j;
5270: aij->a = a;
5271: aij->nonew = -1; /*this indicates that inserting a new value in the matrix that generates a new nonzero is an error*/
5272: aij->free_a = PETSC_FALSE;
5273: aij->free_ij = PETSC_FALSE;
5275: for (ii = 0, aij->nonzerorowcnt = 0, aij->rmax = 0; ii < m; ii++) {
5276: aij->ilen[ii] = aij->imax[ii] = i[ii + 1] - i[ii];
5277: if (PetscDefined(USE_DEBUG)) {
5278: PetscCheck(i[ii + 1] - i[ii] >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Negative row length in i (row indices) row = %" PetscInt_FMT " length = %" PetscInt_FMT, ii, i[ii + 1] - i[ii]);
5279: for (jj = i[ii] + 1; jj < i[ii + 1]; jj++) {
5280: PetscCheck(j[jj] >= j[jj - 1], PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Column entry number %" PetscInt_FMT " (actual column %" PetscInt_FMT ") in row %" PetscInt_FMT " is not sorted", jj - i[ii], j[jj], ii);
5281: PetscCheck(j[jj] != j[jj - 1], PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Column entry number %" PetscInt_FMT " (actual column %" PetscInt_FMT ") in row %" PetscInt_FMT " is identical to previous entry", jj - i[ii], j[jj], ii);
5282: }
5283: }
5284: }
5285: if (PetscDefined(USE_DEBUG)) {
5286: for (ii = 0; ii < aij->i[m]; ii++) {
5287: PetscCheck(j[ii] >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Negative column index at location = %" PetscInt_FMT " index = %" PetscInt_FMT, ii, j[ii]);
5288: PetscCheck(j[ii] <= n - 1, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Column index to large at location = %" PetscInt_FMT " index = %" PetscInt_FMT " last column = %" PetscInt_FMT, ii, j[ii], n - 1);
5289: }
5290: }
5292: PetscCall(MatAssemblyBegin(*mat, MAT_FINAL_ASSEMBLY));
5293: PetscCall(MatAssemblyEnd(*mat, MAT_FINAL_ASSEMBLY));
5294: PetscFunctionReturn(PETSC_SUCCESS);
5295: }
5297: /*@
5298: MatCreateSeqAIJFromTriple - Creates an sequential `MATSEQAIJ` matrix using matrix elements (in COO format)
5299: provided by the user.
5301: Collective
5303: Input Parameters:
5304: + comm - must be an MPI communicator of size 1
5305: . m - number of rows
5306: . n - number of columns
5307: . i - row indices
5308: . j - column indices
5309: . a - matrix values
5310: . nz - number of nonzeros
5311: - idx - if the `i` and `j` indices start with 1 use `PETSC_TRUE` otherwise use `PETSC_FALSE`
5313: Output Parameter:
5314: . mat - the matrix
5316: Level: intermediate
5318: Example:
5319: For the following matrix, the input data expected is as shown (using 0 based indexing)
5320: .vb
5321: 1 0 0
5322: 2 0 3
5323: 4 5 6
5325: i = {0,1,1,2,2,2}
5326: j = {0,0,2,0,1,2}
5327: v = {1,2,3,4,5,6}
5328: .ve
5330: Note:
5331: Instead of using this function, users should also consider `MatSetPreallocationCOO()` and `MatSetValuesCOO()`, which allow repeated or remote entries,
5332: and are particularly useful in iterative applications.
5334: .seealso: [](ch_matrices), `Mat`, `MatCreate()`, `MatCreateAIJ()`, `MatCreateSeqAIJ()`, `MatCreateSeqAIJWithArrays()`, `MatMPIAIJSetPreallocationCSR()`, `MatSetValuesCOO()`, `MatSetPreallocationCOO()`
5335: @*/
5336: PetscErrorCode MatCreateSeqAIJFromTriple(MPI_Comm comm, PetscInt m, PetscInt n, PetscInt i[], PetscInt j[], PetscScalar a[], Mat *mat, PetscInt nz, PetscBool idx)
5337: {
5338: PetscInt ii, *nnz, one = 1, row, col;
5340: PetscFunctionBegin;
5341: PetscCall(PetscCalloc1(m, &nnz));
5342: for (ii = 0; ii < nz; ii++) nnz[i[ii] - !!idx] += 1;
5343: PetscCall(MatCreate(comm, mat));
5344: PetscCall(MatSetSizes(*mat, m, n, m, n));
5345: PetscCall(MatSetType(*mat, MATSEQAIJ));
5346: PetscCall(MatSeqAIJSetPreallocation_SeqAIJ(*mat, 0, nnz));
5347: for (ii = 0; ii < nz; ii++) {
5348: if (idx) {
5349: row = i[ii] - 1;
5350: col = j[ii] - 1;
5351: } else {
5352: row = i[ii];
5353: col = j[ii];
5354: }
5355: PetscCall(MatSetValues(*mat, one, &row, one, &col, &a[ii], ADD_VALUES));
5356: }
5357: PetscCall(MatAssemblyBegin(*mat, MAT_FINAL_ASSEMBLY));
5358: PetscCall(MatAssemblyEnd(*mat, MAT_FINAL_ASSEMBLY));
5359: PetscCall(PetscFree(nnz));
5360: PetscFunctionReturn(PETSC_SUCCESS);
5361: }
5363: PetscErrorCode MatSeqAIJInvalidateDiagonal(Mat A)
5364: {
5365: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
5367: PetscFunctionBegin;
5368: a->idiagvalid = PETSC_FALSE;
5369: a->ibdiagvalid = PETSC_FALSE;
5371: PetscCall(MatSeqAIJInvalidateDiagonal_Inode(A));
5372: PetscFunctionReturn(PETSC_SUCCESS);
5373: }
5375: PetscErrorCode MatCreateMPIMatConcatenateSeqMat_SeqAIJ(MPI_Comm comm, Mat inmat, PetscInt n, MatReuse scall, Mat *outmat)
5376: {
5377: PetscFunctionBegin;
5378: PetscCall(MatCreateMPIMatConcatenateSeqMat_MPIAIJ(comm, inmat, n, scall, outmat));
5379: PetscFunctionReturn(PETSC_SUCCESS);
5380: }
5382: /*
5383: Permute A into C's *local* index space using rowemb,colemb.
5384: The embedding are supposed to be injections and the above implies that the range of rowemb is a subset
5385: of [0,m), colemb is in [0,n).
5386: If pattern == DIFFERENT_NONZERO_PATTERN, C is preallocated according to A.
5387: */
5388: PetscErrorCode MatSetSeqMat_SeqAIJ(Mat C, IS rowemb, IS colemb, MatStructure pattern, Mat B)
5389: {
5390: /* If making this function public, change the error returned in this function away from _PLIB. */
5391: Mat_SeqAIJ *Baij;
5392: PetscBool seqaij;
5393: PetscInt m, n, *nz, i, j, count;
5394: PetscScalar v;
5395: const PetscInt *rowindices, *colindices;
5397: PetscFunctionBegin;
5398: if (!B) PetscFunctionReturn(PETSC_SUCCESS);
5399: /* Check to make sure the target matrix (and embeddings) are compatible with C and each other. */
5400: PetscCall(PetscObjectBaseTypeCompare((PetscObject)B, MATSEQAIJ, &seqaij));
5401: PetscCheck(seqaij, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Input matrix is of wrong type");
5402: if (rowemb) {
5403: PetscCall(ISGetLocalSize(rowemb, &m));
5404: PetscCheck(m == B->rmap->n, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Row IS of size %" PetscInt_FMT " is incompatible with matrix row size %" PetscInt_FMT, m, B->rmap->n);
5405: } else {
5406: PetscCheck(C->rmap->n == B->rmap->n, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Input matrix is row-incompatible with the target matrix");
5407: }
5408: if (colemb) {
5409: PetscCall(ISGetLocalSize(colemb, &n));
5410: PetscCheck(n == B->cmap->n, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Diag col IS of size %" PetscInt_FMT " is incompatible with input matrix col size %" PetscInt_FMT, n, B->cmap->n);
5411: } else {
5412: PetscCheck(C->cmap->n == B->cmap->n, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Input matrix is col-incompatible with the target matrix");
5413: }
5415: Baij = (Mat_SeqAIJ *)B->data;
5416: if (pattern == DIFFERENT_NONZERO_PATTERN) {
5417: PetscCall(PetscMalloc1(B->rmap->n, &nz));
5418: for (i = 0; i < B->rmap->n; i++) nz[i] = Baij->i[i + 1] - Baij->i[i];
5419: PetscCall(MatSeqAIJSetPreallocation(C, 0, nz));
5420: PetscCall(PetscFree(nz));
5421: }
5422: if (pattern == SUBSET_NONZERO_PATTERN) PetscCall(MatZeroEntries(C));
5423: count = 0;
5424: rowindices = NULL;
5425: colindices = NULL;
5426: if (rowemb) PetscCall(ISGetIndices(rowemb, &rowindices));
5427: if (colemb) PetscCall(ISGetIndices(colemb, &colindices));
5428: for (i = 0; i < B->rmap->n; i++) {
5429: PetscInt row;
5430: row = i;
5431: if (rowindices) row = rowindices[i];
5432: for (j = Baij->i[i]; j < Baij->i[i + 1]; j++) {
5433: PetscInt col;
5434: col = Baij->j[count];
5435: if (colindices) col = colindices[col];
5436: v = Baij->a[count];
5437: PetscCall(MatSetValues(C, 1, &row, 1, &col, &v, INSERT_VALUES));
5438: ++count;
5439: }
5440: }
5441: /* FIXME: set C's nonzerostate correctly. */
5442: /* Assembly for C is necessary. */
5443: C->preallocated = PETSC_TRUE;
5444: C->assembled = PETSC_TRUE;
5445: C->was_assembled = PETSC_FALSE;
5446: PetscFunctionReturn(PETSC_SUCCESS);
5447: }
5449: PetscErrorCode MatEliminateZeros_SeqAIJ(Mat A, PetscBool keep)
5450: {
5451: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
5452: MatScalar *aa = a->a;
5453: PetscInt m = A->rmap->n, fshift = 0, fshift_prev = 0, i, k;
5454: PetscInt *ailen = a->ilen, *imax = a->imax, *ai = a->i, *aj = a->j, rmax = 0;
5456: PetscFunctionBegin;
5457: PetscCheck(A->assembled, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Cannot eliminate zeros for unassembled matrix");
5458: if (m) rmax = ailen[0]; /* determine row with most nonzeros */
5459: for (i = 1; i <= m; i++) {
5460: /* move each nonzero entry back by the amount of zero slots (fshift) before it*/
5461: for (k = ai[i - 1]; k < ai[i]; k++) {
5462: if (aa[k] == 0 && (aj[k] != i - 1 || !keep)) fshift++;
5463: else {
5464: if (aa[k] == 0 && aj[k] == i - 1) PetscCall(PetscInfo(A, "Keep the diagonal zero at row %" PetscInt_FMT "\n", i - 1));
5465: aa[k - fshift] = aa[k];
5466: aj[k - fshift] = aj[k];
5467: }
5468: }
5469: ai[i - 1] -= fshift_prev; // safe to update ai[i-1] now since it will not be used in the next iteration
5470: fshift_prev = fshift;
5471: /* reset ilen and imax for each row */
5472: ailen[i - 1] = imax[i - 1] = ai[i] - fshift - ai[i - 1];
5473: a->nonzerorowcnt += ((ai[i] - fshift - ai[i - 1]) > 0);
5474: rmax = PetscMax(rmax, ailen[i - 1]);
5475: }
5476: if (fshift) {
5477: if (m) {
5478: ai[m] -= fshift;
5479: a->nz = ai[m];
5480: }
5481: PetscCall(PetscInfo(A, "Matrix size: %" PetscInt_FMT " X %" PetscInt_FMT "; zeros eliminated: %" PetscInt_FMT "; nonzeros left: %" PetscInt_FMT "\n", m, A->cmap->n, fshift, a->nz));
5482: A->nonzerostate++;
5483: A->info.nz_unneeded += (PetscReal)fshift;
5484: a->rmax = rmax;
5485: if (a->inode.use && a->inode.checked) PetscCall(MatSeqAIJCheckInode(A));
5486: PetscCall(MatAssemblyBegin(A, MAT_FINAL_ASSEMBLY));
5487: PetscCall(MatAssemblyEnd(A, MAT_FINAL_ASSEMBLY));
5488: }
5489: PetscFunctionReturn(PETSC_SUCCESS);
5490: }
5492: PetscFunctionList MatSeqAIJList = NULL;
5494: /*@
5495: MatSeqAIJSetType - Converts a `MATSEQAIJ` matrix to a subtype
5497: Collective
5499: Input Parameters:
5500: + mat - the matrix object
5501: - matype - matrix type
5503: Options Database Key:
5504: . -mat_seqaij_type <method> - for example seqaijcrl
5506: Level: intermediate
5508: .seealso: [](ch_matrices), `Mat`, `PCSetType()`, `VecSetType()`, `MatCreate()`, `MatType`
5509: @*/
5510: PetscErrorCode MatSeqAIJSetType(Mat mat, MatType matype)
5511: {
5512: PetscBool sametype;
5513: PetscErrorCode (*r)(Mat, MatType, MatReuse, Mat *);
5515: PetscFunctionBegin;
5517: PetscCall(PetscObjectTypeCompare((PetscObject)mat, matype, &sametype));
5518: if (sametype) PetscFunctionReturn(PETSC_SUCCESS);
5520: PetscCall(PetscFunctionListFind(MatSeqAIJList, matype, &r));
5521: PetscCheck(r, PetscObjectComm((PetscObject)mat), PETSC_ERR_ARG_UNKNOWN_TYPE, "Unknown Mat type given: %s", matype);
5522: PetscCall((*r)(mat, matype, MAT_INPLACE_MATRIX, &mat));
5523: PetscFunctionReturn(PETSC_SUCCESS);
5524: }
5526: /*@C
5527: MatSeqAIJRegister - - Adds a new sub-matrix type for sequential `MATSEQAIJ` matrices
5529: Not Collective, No Fortran Support
5531: Input Parameters:
5532: + sname - name of a new user-defined matrix type, for example `MATSEQAIJCRL`
5533: - function - routine to convert to subtype
5535: Level: advanced
5537: Notes:
5538: `MatSeqAIJRegister()` may be called multiple times to add several user-defined solvers.
5540: Then, your matrix can be chosen with the procedural interface at runtime via the option
5541: $ -mat_seqaij_type my_mat
5543: .seealso: [](ch_matrices), `Mat`, `MatSeqAIJRegisterAll()`
5544: @*/
5545: PetscErrorCode MatSeqAIJRegister(const char sname[], PetscErrorCode (*function)(Mat, MatType, MatReuse, Mat *))
5546: {
5547: PetscFunctionBegin;
5548: PetscCall(MatInitializePackage());
5549: PetscCall(PetscFunctionListAdd(&MatSeqAIJList, sname, function));
5550: PetscFunctionReturn(PETSC_SUCCESS);
5551: }
5553: PetscBool MatSeqAIJRegisterAllCalled = PETSC_FALSE;
5555: /*@C
5556: MatSeqAIJRegisterAll - Registers all of the matrix subtypes of `MATSSEQAIJ`
5558: Not Collective
5560: Level: advanced
5562: Note:
5563: This registers the versions of `MATSEQAIJ` for GPUs
5565: .seealso: [](ch_matrices), `Mat`, `MatRegisterAll()`, `MatSeqAIJRegister()`
5566: @*/
5567: PetscErrorCode MatSeqAIJRegisterAll(void)
5568: {
5569: PetscFunctionBegin;
5570: if (MatSeqAIJRegisterAllCalled) PetscFunctionReturn(PETSC_SUCCESS);
5571: MatSeqAIJRegisterAllCalled = PETSC_TRUE;
5573: PetscCall(MatSeqAIJRegister(MATSEQAIJCRL, MatConvert_SeqAIJ_SeqAIJCRL));
5574: PetscCall(MatSeqAIJRegister(MATSEQAIJPERM, MatConvert_SeqAIJ_SeqAIJPERM));
5575: PetscCall(MatSeqAIJRegister(MATSEQAIJSELL, MatConvert_SeqAIJ_SeqAIJSELL));
5576: #if defined(PETSC_HAVE_MKL_SPARSE)
5577: PetscCall(MatSeqAIJRegister(MATSEQAIJMKL, MatConvert_SeqAIJ_SeqAIJMKL));
5578: #endif
5579: #if defined(PETSC_HAVE_CUDA)
5580: PetscCall(MatSeqAIJRegister(MATSEQAIJCUSPARSE, MatConvert_SeqAIJ_SeqAIJCUSPARSE));
5581: #endif
5582: #if defined(PETSC_HAVE_HIP)
5583: PetscCall(MatSeqAIJRegister(MATSEQAIJHIPSPARSE, MatConvert_SeqAIJ_SeqAIJHIPSPARSE));
5584: #endif
5585: #if defined(PETSC_HAVE_KOKKOS_KERNELS)
5586: PetscCall(MatSeqAIJRegister(MATSEQAIJKOKKOS, MatConvert_SeqAIJ_SeqAIJKokkos));
5587: #endif
5588: #if defined(PETSC_HAVE_VIENNACL) && defined(PETSC_HAVE_VIENNACL_NO_CUDA)
5589: PetscCall(MatSeqAIJRegister(MATMPIAIJVIENNACL, MatConvert_SeqAIJ_SeqAIJViennaCL));
5590: #endif
5591: PetscFunctionReturn(PETSC_SUCCESS);
5592: }
5594: /*
5595: Special version for direct calls from Fortran
5596: */
5597: #if defined(PETSC_HAVE_FORTRAN_CAPS)
5598: #define matsetvaluesseqaij_ MATSETVALUESSEQAIJ
5599: #elif !defined(PETSC_HAVE_FORTRAN_UNDERSCORE)
5600: #define matsetvaluesseqaij_ matsetvaluesseqaij
5601: #endif
5603: /* Change these macros so can be used in void function */
5605: /* Change these macros so can be used in void function */
5606: /* Identical to PetscCallVoid, except it assigns to *_ierr */
5607: #undef PetscCall
5608: #define PetscCall(...) \
5609: do { \
5610: PetscErrorCode ierr_msv_mpiaij = __VA_ARGS__; \
5611: if (PetscUnlikely(ierr_msv_mpiaij)) { \
5612: *_ierr = PetscError(PETSC_COMM_SELF, __LINE__, PETSC_FUNCTION_NAME, __FILE__, ierr_msv_mpiaij, PETSC_ERROR_REPEAT, " "); \
5613: return; \
5614: } \
5615: } while (0)
5617: #undef SETERRQ
5618: #define SETERRQ(comm, ierr, ...) \
5619: do { \
5620: *_ierr = PetscError(comm, __LINE__, PETSC_FUNCTION_NAME, __FILE__, ierr, PETSC_ERROR_INITIAL, __VA_ARGS__); \
5621: return; \
5622: } while (0)
5624: PETSC_EXTERN void matsetvaluesseqaij_(Mat *AA, PetscInt *mm, const PetscInt im[], PetscInt *nn, const PetscInt in[], const PetscScalar v[], InsertMode *isis, PetscErrorCode *_ierr)
5625: {
5626: Mat A = *AA;
5627: PetscInt m = *mm, n = *nn;
5628: InsertMode is = *isis;
5629: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
5630: PetscInt *rp, k, low, high, t, ii, row, nrow, i, col, l, rmax, N;
5631: PetscInt *imax, *ai, *ailen;
5632: PetscInt *aj, nonew = a->nonew, lastcol = -1;
5633: MatScalar *ap, value, *aa;
5634: PetscBool ignorezeroentries = a->ignorezeroentries;
5635: PetscBool roworiented = a->roworiented;
5637: PetscFunctionBegin;
5638: MatCheckPreallocated(A, 1);
5639: imax = a->imax;
5640: ai = a->i;
5641: ailen = a->ilen;
5642: aj = a->j;
5643: aa = a->a;
5645: for (k = 0; k < m; k++) { /* loop over added rows */
5646: row = im[k];
5647: if (row < 0) continue;
5648: PetscCheck(row < A->rmap->n, PetscObjectComm((PetscObject)A), PETSC_ERR_ARG_OUTOFRANGE, "Row too large");
5649: rp = aj + ai[row];
5650: ap = aa + ai[row];
5651: rmax = imax[row];
5652: nrow = ailen[row];
5653: low = 0;
5654: high = nrow;
5655: for (l = 0; l < n; l++) { /* loop over added columns */
5656: if (in[l] < 0) continue;
5657: PetscCheck(in[l] < A->cmap->n, PetscObjectComm((PetscObject)A), PETSC_ERR_ARG_OUTOFRANGE, "Column too large");
5658: col = in[l];
5659: if (roworiented) value = v[l + k * n];
5660: else value = v[k + l * m];
5662: if (value == 0.0 && ignorezeroentries && (is == ADD_VALUES)) continue;
5664: if (col <= lastcol) low = 0;
5665: else high = nrow;
5666: lastcol = col;
5667: while (high - low > 5) {
5668: t = (low + high) / 2;
5669: if (rp[t] > col) high = t;
5670: else low = t;
5671: }
5672: for (i = low; i < high; i++) {
5673: if (rp[i] > col) break;
5674: if (rp[i] == col) {
5675: if (is == ADD_VALUES) ap[i] += value;
5676: else ap[i] = value;
5677: goto noinsert;
5678: }
5679: }
5680: if (value == 0.0 && ignorezeroentries) goto noinsert;
5681: if (nonew == 1) goto noinsert;
5682: PetscCheck(nonew != -1, PetscObjectComm((PetscObject)A), PETSC_ERR_ARG_OUTOFRANGE, "Inserting a new nonzero in the matrix");
5683: MatSeqXAIJReallocateAIJ(A, A->rmap->n, 1, nrow, row, col, rmax, aa, ai, aj, rp, ap, imax, nonew, MatScalar);
5684: N = nrow++ - 1;
5685: a->nz++;
5686: high++;
5687: /* shift up all the later entries in this row */
5688: for (ii = N; ii >= i; ii--) {
5689: rp[ii + 1] = rp[ii];
5690: ap[ii + 1] = ap[ii];
5691: }
5692: rp[i] = col;
5693: ap[i] = value;
5694: noinsert:;
5695: low = i + 1;
5696: }
5697: ailen[row] = nrow;
5698: }
5699: PetscFunctionReturnVoid();
5700: }
5701: /* Undefining these here since they were redefined from their original definition above! No
5702: * other PETSc functions should be defined past this point, as it is impossible to recover the
5703: * original definitions */
5704: #undef PetscCall
5705: #undef SETERRQ