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: MatGetDiagonalMarkers(SeqAIJ, 1)
21: static PetscErrorCode MatSeqAIJSetTypeFromOptions(Mat A)
22: {
23: PetscBool flg;
24: char type[256];
26: PetscFunctionBegin;
27: PetscObjectOptionsBegin((PetscObject)A);
28: PetscCall(PetscOptionsFList("-mat_seqaij_type", "Matrix SeqAIJ type", "MatSeqAIJSetType", MatSeqAIJList, "seqaij", type, sizeof(type), &flg));
29: if (flg) PetscCall(MatSeqAIJSetType(A, type));
30: PetscOptionsEnd();
31: PetscFunctionReturn(PETSC_SUCCESS);
32: }
34: static PetscErrorCode MatGetColumnReductions_SeqAIJ(Mat A, PetscInt type, PetscReal *reductions)
35: {
36: PetscInt i, m, n;
37: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)A->data;
39: PetscFunctionBegin;
40: PetscCall(MatGetSize(A, &m, &n));
41: PetscCall(PetscArrayzero(reductions, n));
42: if (type == NORM_2) {
43: for (i = 0; i < aij->i[m]; i++) reductions[aij->j[i]] += PetscAbsScalar(aij->a[i] * aij->a[i]);
44: } else if (type == NORM_1) {
45: for (i = 0; i < aij->i[m]; i++) reductions[aij->j[i]] += PetscAbsScalar(aij->a[i]);
46: } else if (type == NORM_INFINITY) {
47: for (i = 0; i < aij->i[m]; i++) reductions[aij->j[i]] = PetscMax(PetscAbsScalar(aij->a[i]), reductions[aij->j[i]]);
48: } else if (type == REDUCTION_SUM_REALPART || type == REDUCTION_MEAN_REALPART) {
49: for (i = 0; i < aij->i[m]; i++) reductions[aij->j[i]] += PetscRealPart(aij->a[i]);
50: } else if (type == REDUCTION_SUM_IMAGINARYPART || type == REDUCTION_MEAN_IMAGINARYPART) {
51: for (i = 0; i < aij->i[m]; i++) reductions[aij->j[i]] += PetscImaginaryPart(aij->a[i]);
52: } else SETERRQ(PetscObjectComm((PetscObject)A), PETSC_ERR_ARG_WRONG, "Unknown reduction type");
54: if (type == NORM_2) {
55: for (i = 0; i < n; i++) reductions[i] = PetscSqrtReal(reductions[i]);
56: } else if (type == REDUCTION_MEAN_REALPART || type == REDUCTION_MEAN_IMAGINARYPART) {
57: for (i = 0; i < n; i++) reductions[i] /= m;
58: }
59: PetscFunctionReturn(PETSC_SUCCESS);
60: }
62: static PetscErrorCode MatFindOffBlockDiagonalEntries_SeqAIJ(Mat A, IS *is)
63: {
64: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
65: PetscInt i, m = A->rmap->n, cnt = 0, bs = A->rmap->bs;
66: const PetscInt *jj = a->j, *ii = a->i;
67: PetscInt *rows;
69: PetscFunctionBegin;
70: for (i = 0; i < m; i++) {
71: if ((ii[i] != ii[i + 1]) && ((jj[ii[i]] < bs * (i / bs)) || (jj[ii[i + 1] - 1] > bs * ((i + bs) / bs) - 1))) cnt++;
72: }
73: PetscCall(PetscMalloc1(cnt, &rows));
74: cnt = 0;
75: for (i = 0; i < m; i++) {
76: if ((ii[i] != ii[i + 1]) && ((jj[ii[i]] < bs * (i / bs)) || (jj[ii[i + 1] - 1] > bs * ((i + bs) / bs) - 1))) {
77: rows[cnt] = i;
78: cnt++;
79: }
80: }
81: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, cnt, rows, PETSC_OWN_POINTER, is));
82: PetscFunctionReturn(PETSC_SUCCESS);
83: }
85: PetscErrorCode MatFindZeroDiagonals_SeqAIJ_Private(Mat A, PetscInt *nrows, PetscInt **zrows)
86: {
87: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
88: const MatScalar *aa;
89: PetscInt i, m = A->rmap->n, cnt = 0;
90: const PetscInt *ii = a->i, *jj = a->j, *diag;
91: PetscInt *rows;
93: PetscFunctionBegin;
94: PetscCall(MatSeqAIJGetArrayRead(A, &aa));
95: PetscCall(MatGetDiagonalMarkers_SeqAIJ(A, &diag, NULL));
96: for (i = 0; i < m; i++) {
97: if ((diag[i] >= ii[i + 1]) || (jj[diag[i]] != i) || (aa[diag[i]] == 0.0)) cnt++;
98: }
99: PetscCall(PetscMalloc1(cnt, &rows));
100: cnt = 0;
101: for (i = 0; i < m; i++) {
102: if ((diag[i] >= ii[i + 1]) || (jj[diag[i]] != i) || (aa[diag[i]] == 0.0)) rows[cnt++] = i;
103: }
104: *nrows = cnt;
105: *zrows = rows;
106: PetscCall(MatSeqAIJRestoreArrayRead(A, &aa));
107: PetscFunctionReturn(PETSC_SUCCESS);
108: }
110: static PetscErrorCode MatFindZeroDiagonals_SeqAIJ(Mat A, IS *zrows)
111: {
112: PetscInt nrows, *rows;
114: PetscFunctionBegin;
115: *zrows = NULL;
116: PetscCall(MatFindZeroDiagonals_SeqAIJ_Private(A, &nrows, &rows));
117: PetscCall(ISCreateGeneral(PetscObjectComm((PetscObject)A), nrows, rows, PETSC_OWN_POINTER, zrows));
118: PetscFunctionReturn(PETSC_SUCCESS);
119: }
121: static PetscErrorCode MatFindNonzeroRows_SeqAIJ(Mat A, IS *keptrows)
122: {
123: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
124: const MatScalar *aa;
125: PetscInt m = A->rmap->n, cnt = 0;
126: const PetscInt *ii;
127: PetscInt n, i, j, *rows;
129: PetscFunctionBegin;
130: PetscCall(MatSeqAIJGetArrayRead(A, &aa));
131: *keptrows = NULL;
132: ii = a->i;
133: for (i = 0; i < m; i++) {
134: n = ii[i + 1] - ii[i];
135: if (!n) {
136: cnt++;
137: goto ok1;
138: }
139: for (j = ii[i]; j < ii[i + 1]; j++) {
140: if (aa[j] != 0.0) goto ok1;
141: }
142: cnt++;
143: ok1:;
144: }
145: if (!cnt) {
146: PetscCall(MatSeqAIJRestoreArrayRead(A, &aa));
147: PetscFunctionReturn(PETSC_SUCCESS);
148: }
149: PetscCall(PetscMalloc1(A->rmap->n - cnt, &rows));
150: cnt = 0;
151: for (i = 0; i < m; i++) {
152: n = ii[i + 1] - ii[i];
153: if (!n) continue;
154: for (j = ii[i]; j < ii[i + 1]; j++) {
155: if (aa[j] != 0.0) {
156: rows[cnt++] = i;
157: break;
158: }
159: }
160: }
161: PetscCall(MatSeqAIJRestoreArrayRead(A, &aa));
162: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, cnt, rows, PETSC_OWN_POINTER, keptrows));
163: PetscFunctionReturn(PETSC_SUCCESS);
164: }
166: PetscErrorCode MatDiagonalSet_SeqAIJ(Mat Y, Vec D, InsertMode is)
167: {
168: PetscInt i, m = Y->rmap->n;
169: const PetscInt *diag;
170: MatScalar *aa;
171: const PetscScalar *v;
172: PetscBool diagDense;
174: PetscFunctionBegin;
175: if (Y->assembled) {
176: PetscCall(MatGetDiagonalMarkers_SeqAIJ(Y, &diag, &diagDense));
177: if (diagDense) {
178: PetscCall(VecGetArrayRead(D, &v));
179: PetscCall(MatSeqAIJGetArray(Y, &aa));
180: if (is == INSERT_VALUES) {
181: for (i = 0; i < m; i++) aa[diag[i]] = v[i];
182: } else {
183: for (i = 0; i < m; i++) aa[diag[i]] += v[i];
184: }
185: PetscCall(MatSeqAIJRestoreArray(Y, &aa));
186: PetscCall(VecRestoreArrayRead(D, &v));
187: PetscFunctionReturn(PETSC_SUCCESS);
188: }
189: }
190: PetscCall(MatDiagonalSet_Default(Y, D, is));
191: PetscFunctionReturn(PETSC_SUCCESS);
192: }
194: PetscErrorCode MatGetRowIJ_SeqAIJ(Mat A, PetscInt oshift, PetscBool symmetric, PetscBool inodecompressed, PetscInt *m, const PetscInt *ia[], const PetscInt *ja[], PetscBool *done)
195: {
196: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
197: PetscInt i, ishift;
199: PetscFunctionBegin;
200: if (m) *m = A->rmap->n;
201: if (!ia) PetscFunctionReturn(PETSC_SUCCESS);
202: ishift = 0;
203: if (symmetric && A->structurally_symmetric != PETSC_BOOL3_TRUE) {
204: PetscCall(MatToSymmetricIJ_SeqAIJ(A->rmap->n, a->i, a->j, PETSC_TRUE, ishift, oshift, (PetscInt **)ia, (PetscInt **)ja));
205: } else if (oshift == 1) {
206: PetscInt *tia;
207: PetscInt nz = a->i[A->rmap->n];
209: /* malloc space and add 1 to i and j indices */
210: PetscCall(PetscMalloc1(A->rmap->n + 1, &tia));
211: for (i = 0; i < A->rmap->n + 1; i++) tia[i] = a->i[i] + 1;
212: *ia = tia;
213: if (ja) {
214: 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: #include <petsc/private/isimpl.h>
350: /*@
351: MatSeqAIJSetValuesLocalFast - An optimized version of `MatSetValuesLocal()` for `MATSEQAIJ` matrices, valid under
352: several restrictive assumptions.
354: Not Collective
356: Input Parameters:
357: + A - the `MATSEQAIJ` matrix
358: . m - the number of rows being set (must be 1)
359: . im - array of length `m` giving the local row index
360: . n - the number of columns being set
361: . in - array of length `n` giving the local column indices
362: . v - array of length `n` of values to add
363: - is - the insert mode (must be `ADD_VALUES`)
365: Level: developer
367: Notes:
368: This routine requires that a single row of values is set with each call, that no row or column
369: index is negative or larger than the number of rows or columns, that values are always added
370: (not inserted), and that no new nonzero locations are introduced.
372: The global column indices are not assumed to be sorted.
374: .seealso: `Mat`, `MATSEQAIJ`, `MatSetValuesLocal()`, `MatSetValues()`
375: @*/
376: PetscErrorCode MatSeqAIJSetValuesLocalFast(Mat A, PetscInt m, const PetscInt im[], PetscInt n, const PetscInt in[], const PetscScalar v[], InsertMode is)
377: {
378: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
379: PetscInt low, high, t, row, nrow, i, col, l;
380: const PetscInt *rp, *ai = a->i, *ailen = a->ilen, *aj = a->j;
381: PetscInt lastcol = -1;
382: MatScalar *ap, value, *aa;
383: const PetscInt *ridx = A->rmap->mapping->indices, *cidx = A->cmap->mapping->indices;
385: PetscFunctionBegin;
386: PetscCall(MatSeqAIJGetArray(A, &aa));
387: row = ridx[im[0]];
388: rp = aj + ai[row];
389: ap = aa + ai[row];
390: nrow = ailen[row];
391: low = 0;
392: high = nrow;
393: for (l = 0; l < n; l++) { /* loop over added columns */
394: col = cidx[in[l]];
395: value = v[l];
397: if (col <= lastcol) low = 0;
398: else high = nrow;
399: lastcol = col;
400: while (high - low > 5) {
401: t = (low + high) / 2;
402: if (rp[t] > col) high = t;
403: else low = t;
404: }
405: for (i = low; i < high; i++) {
406: if (rp[i] == col) {
407: ap[i] += value;
408: low = i + 1;
409: break;
410: }
411: }
412: }
413: PetscCall(MatSeqAIJRestoreArray(A, &aa));
414: return PETSC_SUCCESS;
415: }
417: PetscErrorCode MatSetValues_SeqAIJ(Mat A, PetscInt m, const PetscInt im[], PetscInt n, const PetscInt in[], const PetscScalar v[], InsertMode is)
418: {
419: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
420: PetscInt *rp, k, low, high, t, ii, row, nrow, i, col, l, rmax, N;
421: PetscInt *imax = a->imax, *ai = a->i, *ailen = a->ilen;
422: PetscInt *aj = a->j, nonew = a->nonew, lastcol = -1;
423: MatScalar *ap = NULL, value = 0.0, *aa;
424: PetscBool ignorezeroentries = a->ignorezeroentries;
425: PetscBool roworiented = a->roworiented;
427: PetscFunctionBegin;
428: PetscCall(MatSeqAIJGetArray(A, &aa));
429: for (k = 0; k < m; k++) { /* loop over added rows */
430: row = im[k];
431: if (row < 0) continue;
432: 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);
433: rp = PetscSafePointerPlusOffset(aj, ai[row]);
434: if (!A->structure_only) ap = PetscSafePointerPlusOffset(aa, ai[row]);
435: rmax = imax[row];
436: nrow = ailen[row];
437: low = 0;
438: high = nrow;
439: for (l = 0; l < n; l++) { /* loop over added columns */
440: if (in[l] < 0) continue;
441: 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);
442: col = in[l];
443: if (v && !A->structure_only) value = roworiented ? v[l + k * n] : v[k + l * m];
444: if (!A->structure_only && value == 0.0 && ignorezeroentries && is == ADD_VALUES && row != col) continue;
446: if (col <= lastcol) low = 0;
447: else high = nrow;
448: lastcol = col;
449: while (high - low > 5) {
450: t = (low + high) / 2;
451: if (rp[t] > col) high = t;
452: else low = t;
453: }
454: for (i = low; i < high; i++) {
455: if (rp[i] > col) break;
456: if (rp[i] == col) {
457: if (!A->structure_only) {
458: if (is == ADD_VALUES) {
459: ap[i] += value;
460: (void)PetscLogFlops(1.0);
461: } else ap[i] = value;
462: }
463: low = i + 1;
464: goto noinsert;
465: }
466: }
467: if (value == 0.0 && ignorezeroentries && row != col) goto noinsert;
468: if (nonew == 1) goto noinsert;
469: PetscCheck(nonew != -1, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Inserting a new nonzero at (%" PetscInt_FMT ",%" PetscInt_FMT ") in the matrix", row, col);
470: if (A->structure_only) {
471: MatSeqXAIJReallocateAIJ_structure_only(A, A->rmap->n, 1, nrow, row, col, rmax, ai, aj, rp, imax, nonew, MatScalar);
472: } else {
473: MatSeqXAIJReallocateAIJ(A, A->rmap->n, 1, nrow, row, col, rmax, aa, ai, aj, rp, ap, imax, nonew, MatScalar);
474: }
475: N = nrow++ - 1;
476: a->nz++;
477: high++;
478: /* shift up all the later entries in this row */
479: PetscCall(PetscArraymove(rp + i + 1, rp + i, N - i + 1));
480: rp[i] = col;
481: if (!A->structure_only) {
482: PetscCall(PetscArraymove(ap + i + 1, ap + i, N - i + 1));
483: ap[i] = value;
484: }
485: low = i + 1;
486: noinsert:;
487: }
488: ailen[row] = nrow;
489: }
490: PetscCall(MatSeqAIJRestoreArray(A, &aa));
491: PetscFunctionReturn(PETSC_SUCCESS);
492: }
494: static PetscErrorCode MatSetValues_SeqAIJ_SortedFullNoPreallocation(Mat A, PetscInt m, const PetscInt im[], PetscInt n, const PetscInt in[], const PetscScalar v[], InsertMode is)
495: {
496: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
497: PetscInt *rp, k, row;
498: PetscInt *ai = a->i;
499: PetscInt *aj = a->j;
500: MatScalar *aa, *ap;
502: PetscFunctionBegin;
503: PetscCheck(!A->was_assembled, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Cannot call on assembled matrix.");
504: 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);
506: PetscCall(MatSeqAIJGetArray(A, &aa));
507: for (k = 0; k < m; k++) { /* loop over added rows */
508: row = im[k];
509: rp = aj + ai[row];
510: ap = PetscSafePointerPlusOffset(aa, ai[row]);
512: PetscCall(PetscArraycpy(rp, in, n));
513: if (!A->structure_only) {
514: if (v) {
515: PetscCall(PetscArraycpy(ap, v, n));
516: v += n;
517: } else {
518: PetscCall(PetscMemzero(ap, n * sizeof(PetscScalar)));
519: }
520: }
521: a->ilen[row] = n;
522: a->imax[row] = n;
523: a->i[row + 1] = a->i[row] + n;
524: a->nz += n;
525: }
526: PetscCall(MatSeqAIJRestoreArray(A, &aa));
527: PetscFunctionReturn(PETSC_SUCCESS);
528: }
530: /*@
531: MatSeqAIJSetTotalPreallocation - Sets an upper bound on the total number of expected nonzeros in the matrix.
533: Input Parameters:
534: + A - the `MATSEQAIJ` matrix
535: - nztotal - bound on the number of nonzeros
537: Level: advanced
539: Notes:
540: This can be called if you will be provided the matrix row by row (from row zero) with sorted column indices for each row.
541: Simply call `MatSetValues()` after this call to provide the matrix entries in the usual manner. This matrix may be used
542: as always with multiple matrix assemblies.
544: .seealso: [](ch_matrices), `Mat`, `MatSetOption()`, `MAT_SORTED_FULL`, `MatSetValues()`, `MatSeqAIJSetPreallocation()`
545: @*/
546: PetscErrorCode MatSeqAIJSetTotalPreallocation(Mat A, PetscInt nztotal)
547: {
548: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
550: PetscFunctionBegin;
551: PetscCall(PetscLayoutSetUp(A->rmap));
552: PetscCall(PetscLayoutSetUp(A->cmap));
553: a->maxnz = nztotal;
554: if (!a->imax) PetscCall(PetscMalloc1(A->rmap->n, &a->imax));
555: if (!a->ilen) {
556: PetscCall(PetscMalloc1(A->rmap->n, &a->ilen));
557: } else {
558: PetscCall(PetscMemzero(a->ilen, A->rmap->n * sizeof(PetscInt)));
559: }
561: /* allocate the matrix space */
562: PetscCall(PetscShmgetAllocateArray(A->rmap->n + 1, sizeof(PetscInt), (void **)&a->i));
563: PetscCall(PetscShmgetAllocateArray(nztotal, sizeof(PetscInt), (void **)&a->j));
564: a->free_ij = PETSC_TRUE;
565: if (A->structure_only) {
566: a->free_a = PETSC_FALSE;
567: } else {
568: PetscCall(PetscShmgetAllocateArray(nztotal, sizeof(PetscScalar), (void **)&a->a));
569: a->free_a = PETSC_TRUE;
570: }
571: a->i[0] = 0;
572: A->ops->setvalues = MatSetValues_SeqAIJ_SortedFullNoPreallocation;
573: A->preallocated = PETSC_TRUE;
574: PetscFunctionReturn(PETSC_SUCCESS);
575: }
577: static PetscErrorCode MatSetValues_SeqAIJ_SortedFull(Mat A, PetscInt m, const PetscInt im[], PetscInt n, const PetscInt in[], const PetscScalar v[], InsertMode is)
578: {
579: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
580: PetscInt *rp, k, row;
581: PetscInt *ai = a->i, *ailen = a->ilen;
582: PetscInt *aj = a->j;
583: MatScalar *aa, *ap;
585: PetscFunctionBegin;
586: PetscCall(MatSeqAIJGetArray(A, &aa));
587: for (k = 0; k < m; k++) { /* loop over added rows */
588: row = im[k];
589: 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);
590: rp = aj + ai[row];
591: ap = aa + ai[row];
592: if (!A->was_assembled) PetscCall(PetscArraycpy(rp, in, n));
593: if (!A->structure_only) {
594: if (v) {
595: PetscCall(PetscArraycpy(ap, v, n));
596: v += n;
597: } else {
598: PetscCall(PetscMemzero(ap, n * sizeof(PetscScalar)));
599: }
600: }
601: ailen[row] = n;
602: a->nz += n;
603: }
604: PetscCall(MatSeqAIJRestoreArray(A, &aa));
605: PetscFunctionReturn(PETSC_SUCCESS);
606: }
608: static PetscErrorCode MatGetValues_SeqAIJ(Mat A, PetscInt m, const PetscInt im[], PetscInt n, const PetscInt in[], PetscScalar v[])
609: {
610: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
611: PetscInt *rp, k, low, high, t, row, nrow, i, col, l, *aj = a->j;
612: PetscInt *ai = a->i, *ailen = a->ilen;
613: const MatScalar *ap, *aa;
614: PetscBool hyprecoo;
615: PetscBool roworiented = a->roworiented;
616: PetscScalar *value;
618: PetscFunctionBegin;
619: PetscCall(PetscStrcmp("_internal_COO_mat_for_hypre", ((PetscObject)A)->name, &hyprecoo));
621: PetscCall(MatSeqAIJGetArrayRead(A, &aa));
622: for (k = 0; k < m; k++) { /* loop over rows */
623: row = im[k];
624: if (row < 0) continue; /* negative row */
625: 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);
626: rp = PetscSafePointerPlusOffset(aj, ai[row]);
627: ap = PetscSafePointerPlusOffset(aa, ai[row]);
628: nrow = ailen[row];
629: for (l = 0; l < n; l++) { /* loop over columns */
630: if (in[l] < 0) continue; /* negative column */
631: value = roworiented ? &v[l + k * n] : &v[k + l * m];
632: 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);
633: col = in[l];
634: /* hypre coo mat stores its diagonal at the front, out of sort */
635: if (hyprecoo) {
636: if (col == rp[0]) {
637: *value = ap[0];
638: goto finished;
639: }
640: low = 1;
641: } else low = 0;
642: high = nrow;
643: /* assume sorted */
644: while (high - low > 5) {
645: t = (low + high) / 2;
646: if (rp[t] > col) high = t;
647: else low = t;
648: }
649: for (i = low; i < high; i++) {
650: if (rp[i] > col) break;
651: if (rp[i] == col) {
652: *value = ap[i];
653: goto finished;
654: }
655: }
656: *value = 0.0;
657: finished:;
658: }
659: }
660: PetscCall(MatSeqAIJRestoreArrayRead(A, &aa));
661: PetscFunctionReturn(PETSC_SUCCESS);
662: }
664: static PetscErrorCode MatView_SeqAIJ_Binary(Mat mat, PetscViewer viewer)
665: {
666: Mat_SeqAIJ *A = (Mat_SeqAIJ *)mat->data;
667: const PetscScalar *av;
668: PetscInt header[4], M, N, m, nz, i;
669: PetscInt *rowlens;
671: PetscFunctionBegin;
672: PetscCall(PetscViewerSetUp(viewer));
674: M = mat->rmap->N;
675: N = mat->cmap->N;
676: m = mat->rmap->n;
677: nz = A->nz;
679: /* write matrix header */
680: header[0] = MAT_FILE_CLASSID;
681: header[1] = M;
682: header[2] = N;
683: header[3] = nz;
684: PetscCall(PetscViewerBinaryWrite(viewer, header, 4, PETSC_INT));
686: /* fill in and store row lengths */
687: PetscCall(PetscMalloc1(m, &rowlens));
688: for (i = 0; i < m; i++) rowlens[i] = A->i[i + 1] - A->i[i];
689: if (PetscDefined(USE_DEBUG)) {
690: PetscInt mnz = 0;
692: for (i = 0; i < m; i++) mnz += rowlens[i];
693: PetscCheck(nz == mnz, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Row lens %" PetscInt_FMT " do not sum to nz %" PetscInt_FMT, mnz, nz);
694: }
695: PetscCall(PetscViewerBinaryWrite(viewer, rowlens, m, PETSC_INT));
696: PetscCall(PetscFree(rowlens));
697: /* store column indices */
698: PetscCall(PetscViewerBinaryWrite(viewer, A->j, nz, PETSC_INT));
699: /* store nonzero values */
700: PetscCall(MatSeqAIJGetArrayRead(mat, &av));
701: PetscCall(PetscViewerBinaryWrite(viewer, av, nz, PETSC_SCALAR));
702: PetscCall(MatSeqAIJRestoreArrayRead(mat, &av));
704: /* write block size option to the viewer's .info file */
705: PetscCall(MatView_Binary_BlockSizes(mat, viewer));
706: PetscFunctionReturn(PETSC_SUCCESS);
707: }
709: static PetscErrorCode MatView_SeqAIJ_ASCII_structonly(Mat A, PetscViewer viewer)
710: {
711: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
712: PetscInt i, k, m = A->rmap->N;
714: PetscFunctionBegin;
715: PetscCall(PetscViewerASCIIUseTabs(viewer, PETSC_FALSE));
716: for (i = 0; i < m; i++) {
717: PetscCall(PetscViewerASCIIPrintf(viewer, "row %" PetscInt_FMT ":", i));
718: for (k = a->i[i]; k < a->i[i + 1]; k++) PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ") ", a->j[k]));
719: PetscCall(PetscViewerASCIIPrintf(viewer, "\n"));
720: }
721: PetscCall(PetscViewerASCIIUseTabs(viewer, PETSC_TRUE));
722: PetscFunctionReturn(PETSC_SUCCESS);
723: }
725: static PetscErrorCode MatView_SeqAIJ_ASCII(Mat A, PetscViewer viewer)
726: {
727: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
728: const PetscScalar *av;
729: PetscInt i, j, m = A->rmap->n;
730: const char *name;
731: PetscViewerFormat format;
733: PetscFunctionBegin;
734: if (A->structure_only) {
735: PetscCall(MatView_SeqAIJ_ASCII_structonly(A, viewer));
736: PetscFunctionReturn(PETSC_SUCCESS);
737: }
739: PetscCall(PetscViewerGetFormat(viewer, &format));
740: // By petsc's rule, even PETSC_VIEWER_ASCII_INFO_DETAIL doesn't print matrix entries
741: if (format == PETSC_VIEWER_ASCII_FACTOR_INFO || format == PETSC_VIEWER_ASCII_INFO || format == PETSC_VIEWER_ASCII_INFO_DETAIL) PetscFunctionReturn(PETSC_SUCCESS);
743: /* trigger copy to CPU if needed */
744: PetscCall(MatSeqAIJGetArrayRead(A, &av));
745: PetscCall(MatSeqAIJRestoreArrayRead(A, &av));
746: if (format == PETSC_VIEWER_ASCII_MATLAB) {
747: PetscInt nofinalvalue = 0;
748: if (m && ((a->i[m] == a->i[m - 1]) || (a->j[a->nz - 1] != A->cmap->n - 1))) {
749: /* Need a dummy value to ensure the dimension of the matrix. */
750: nofinalvalue = 1;
751: }
752: PetscCall(PetscViewerASCIIUseTabs(viewer, PETSC_FALSE));
753: PetscCall(PetscViewerASCIIPrintf(viewer, "%% Size = %" PetscInt_FMT " %" PetscInt_FMT " \n", m, A->cmap->n));
754: PetscCall(PetscViewerASCIIPrintf(viewer, "%% Nonzeros = %" PetscInt_FMT " \n", a->nz));
755: PetscCall(PetscViewerASCIIPrintf(viewer, "zzz = zeros(%" PetscInt_FMT ",%d);\n", a->nz + nofinalvalue, PetscDefined(USE_COMPLEX) ? 4 : 3));
756: PetscCall(PetscViewerASCIIPrintf(viewer, "zzz = [\n"));
758: for (i = 0; i < m; i++) {
759: for (j = a->i[i]; j < a->i[i + 1]; j++) {
760: #if PetscDefined(USE_COMPLEX)
761: 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])));
762: #else
763: PetscCall(PetscViewerASCIIPrintf(viewer, "%" PetscInt_FMT " %" PetscInt_FMT " %18.16e\n", i + 1, a->j[j] + 1, (double)a->a[j]));
764: #endif
765: }
766: }
767: if (nofinalvalue) {
768: if (PetscDefined(USE_COMPLEX)) PetscCall(PetscViewerASCIIPrintf(viewer, "%" PetscInt_FMT " %" PetscInt_FMT " %18.16e %18.16e\n", m, A->cmap->n, 0., 0.));
769: else PetscCall(PetscViewerASCIIPrintf(viewer, "%" PetscInt_FMT " %" PetscInt_FMT " %18.16e\n", m, A->cmap->n, 0.0));
770: }
771: PetscCall(PetscObjectGetName((PetscObject)A, &name));
772: PetscCall(PetscViewerASCIIPrintf(viewer, "];\n %s = spconvert(zzz);\n", name));
773: PetscCall(PetscViewerASCIIUseTabs(viewer, PETSC_TRUE));
774: } else if (format == PETSC_VIEWER_ASCII_COMMON) {
775: PetscCall(PetscViewerASCIIUseTabs(viewer, PETSC_FALSE));
776: for (i = 0; i < m; i++) {
777: PetscCall(PetscViewerASCIIPrintf(viewer, "row %" PetscInt_FMT ":", i));
778: for (j = a->i[i]; j < a->i[i + 1]; j++) {
779: #if PetscDefined(USE_COMPLEX)
780: if (PetscImaginaryPart(a->a[j]) > 0.0 && PetscRealPart(a->a[j]) != 0.0) {
781: PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g + %g i)", a->j[j], (double)PetscRealPart(a->a[j]), (double)PetscImaginaryPart(a->a[j])));
782: } else if (PetscImaginaryPart(a->a[j]) < 0.0 && PetscRealPart(a->a[j]) != 0.0) {
783: PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g - %g i)", a->j[j], (double)PetscRealPart(a->a[j]), (double)-PetscImaginaryPart(a->a[j])));
784: } else if (PetscRealPart(a->a[j]) != 0.0) {
785: PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g) ", a->j[j], (double)PetscRealPart(a->a[j])));
786: }
787: #else
788: if (a->a[j] != 0.0) PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g) ", a->j[j], (double)a->a[j]));
789: #endif
790: }
791: PetscCall(PetscViewerASCIIPrintf(viewer, "\n"));
792: }
793: PetscCall(PetscViewerASCIIUseTabs(viewer, PETSC_TRUE));
794: } else if (format == PETSC_VIEWER_ASCII_SYMMODU) {
795: PetscInt nzd = 0, fshift = 1, *sptr;
796: PetscCall(PetscViewerASCIIUseTabs(viewer, PETSC_FALSE));
797: PetscCall(PetscMalloc1(m + 1, &sptr));
798: for (i = 0; i < m; i++) {
799: sptr[i] = nzd + 1;
800: for (j = a->i[i]; j < a->i[i + 1]; j++) {
801: if (a->j[j] >= i) {
802: if (PetscDefined(USE_COMPLEX) ? (PetscImaginaryPart(a->a[j]) != 0.0 || PetscRealPart(a->a[j]) != 0.0) : a->a[j] != 0.0) nzd++;
803: }
804: }
805: }
806: sptr[m] = nzd + 1;
807: PetscCall(PetscViewerASCIIPrintf(viewer, " %" PetscInt_FMT " %" PetscInt_FMT "\n\n", m, nzd));
808: for (i = 0; i < m + 1; i += 6) {
809: if (i + 4 < m) {
810: 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]));
811: } else if (i + 3 < m) {
812: 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]));
813: } else if (i + 2 < m) {
814: PetscCall(PetscViewerASCIIPrintf(viewer, " %" PetscInt_FMT " %" PetscInt_FMT " %" PetscInt_FMT " %" PetscInt_FMT "\n", sptr[i], sptr[i + 1], sptr[i + 2], sptr[i + 3]));
815: } else if (i + 1 < m) {
816: PetscCall(PetscViewerASCIIPrintf(viewer, " %" PetscInt_FMT " %" PetscInt_FMT " %" PetscInt_FMT "\n", sptr[i], sptr[i + 1], sptr[i + 2]));
817: } else if (i < m) {
818: PetscCall(PetscViewerASCIIPrintf(viewer, " %" PetscInt_FMT " %" PetscInt_FMT "\n", sptr[i], sptr[i + 1]));
819: } else {
820: PetscCall(PetscViewerASCIIPrintf(viewer, " %" PetscInt_FMT "\n", sptr[i]));
821: }
822: }
823: PetscCall(PetscViewerASCIIPrintf(viewer, "\n"));
824: PetscCall(PetscFree(sptr));
825: for (i = 0; i < m; i++) {
826: for (j = a->i[i]; j < a->i[i + 1]; j++) {
827: if (a->j[j] >= i) PetscCall(PetscViewerASCIIPrintf(viewer, " %" PetscInt_FMT " ", a->j[j] + fshift));
828: }
829: PetscCall(PetscViewerASCIIPrintf(viewer, "\n"));
830: }
831: PetscCall(PetscViewerASCIIPrintf(viewer, "\n"));
832: for (i = 0; i < m; i++) {
833: for (j = a->i[i]; j < a->i[i + 1]; j++) {
834: if (a->j[j] >= i) {
835: #if PetscDefined(USE_COMPLEX)
836: 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])));
837: #else
838: if (a->a[j] != 0.0) PetscCall(PetscViewerASCIIPrintf(viewer, " %18.16e ", (double)a->a[j]));
839: #endif
840: }
841: }
842: PetscCall(PetscViewerASCIIPrintf(viewer, "\n"));
843: }
844: PetscCall(PetscViewerASCIIUseTabs(viewer, PETSC_TRUE));
845: } else if (format == PETSC_VIEWER_ASCII_DENSE) {
846: PetscInt cnt = 0, jcnt;
847: PetscScalar value;
848: PetscBool realonly = PETSC_TRUE;
850: if (PetscDefined(USE_COMPLEX)) {
851: for (i = 0; i < a->i[m]; i++) {
852: if (PetscImaginaryPart(a->a[i]) != 0.0) {
853: realonly = PETSC_FALSE;
854: break;
855: }
856: }
857: }
859: PetscCall(PetscViewerASCIIUseTabs(viewer, PETSC_FALSE));
860: for (i = 0; i < m; i++) {
861: jcnt = 0;
862: for (j = 0; j < A->cmap->n; j++) {
863: if (jcnt < a->i[i + 1] - a->i[i] && j == a->j[cnt]) {
864: value = a->a[cnt++];
865: jcnt++;
866: } else {
867: value = 0.0;
868: }
869: if (!PetscDefined(USE_COMPLEX) || realonly) {
870: PetscCall(PetscViewerASCIIPrintf(viewer, " %7.5e ", (double)PetscRealPart(value)));
871: } else {
872: PetscCall(PetscViewerASCIIPrintf(viewer, " %7.5e+%7.5e i ", (double)PetscRealPart(value), (double)PetscImaginaryPart(value)));
873: }
874: }
875: PetscCall(PetscViewerASCIIPrintf(viewer, "\n"));
876: }
877: PetscCall(PetscViewerASCIIUseTabs(viewer, PETSC_TRUE));
878: } else if (format == PETSC_VIEWER_ASCII_MATRIXMARKET) {
879: PetscInt fshift = 1;
880: PetscCall(PetscViewerASCIIUseTabs(viewer, PETSC_FALSE));
881: PetscCall(PetscViewerASCIIPrintf(viewer, "%%%%MatrixMarket matrix coordinate %s general\n", PetscDefined(USE_COMPLEX) ? "complex" : "real"));
882: PetscCall(PetscViewerASCIIPrintf(viewer, "%" PetscInt_FMT " %" PetscInt_FMT " %" PetscInt_FMT "\n", m, A->cmap->n, a->nz));
883: for (i = 0; i < m; i++) {
884: for (j = a->i[i]; j < a->i[i + 1]; j++) {
885: #if PetscDefined(USE_COMPLEX)
886: 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])));
887: #else
888: PetscCall(PetscViewerASCIIPrintf(viewer, "%" PetscInt_FMT " %" PetscInt_FMT " %g\n", i + fshift, a->j[j] + fshift, (double)a->a[j]));
889: #endif
890: }
891: }
892: PetscCall(PetscViewerASCIIUseTabs(viewer, PETSC_TRUE));
893: } else {
894: PetscCall(PetscViewerASCIIUseTabs(viewer, PETSC_FALSE));
895: if (A->factortype) {
896: const PetscInt *adiag;
898: PetscCall(MatGetDiagonalMarkers_SeqAIJ(A, &adiag, NULL));
899: for (i = 0; i < m; i++) {
900: PetscCall(PetscViewerASCIIPrintf(viewer, "row %" PetscInt_FMT ":", i));
901: /* L part */
902: for (j = a->i[i]; j < a->i[i + 1]; j++) {
903: if (PetscDefined(USE_COMPLEX) && PetscImaginaryPart(a->a[j]) > 0.0) {
904: PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g + %g i)", a->j[j], (double)PetscRealPart(a->a[j]), (double)PetscImaginaryPart(a->a[j])));
905: } else if (PetscDefined(USE_COMPLEX) && PetscImaginaryPart(a->a[j]) < 0.0) {
906: PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g - %g i)", a->j[j], (double)PetscRealPart(a->a[j]), (double)(-PetscImaginaryPart(a->a[j]))));
907: } else {
908: PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g) ", a->j[j], (double)PetscRealPart(a->a[j])));
909: }
910: }
911: /* diagonal */
912: j = adiag[i];
913: if (PetscDefined(USE_COMPLEX) && 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 (PetscDefined(USE_COMPLEX) && 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: }
921: /* U part */
922: for (j = adiag[i + 1] + 1; j < adiag[i]; j++) {
923: if (PetscDefined(USE_COMPLEX) && PetscImaginaryPart(a->a[j]) > 0.0) {
924: PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g + %g i)", a->j[j], (double)PetscRealPart(a->a[j]), (double)PetscImaginaryPart(a->a[j])));
925: } else if (PetscDefined(USE_COMPLEX) && PetscImaginaryPart(a->a[j]) < 0.0) {
926: PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g - %g i)", a->j[j], (double)PetscRealPart(a->a[j]), (double)(-PetscImaginaryPart(a->a[j]))));
927: } else {
928: PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g) ", a->j[j], (double)PetscRealPart(a->a[j])));
929: }
930: }
931: PetscCall(PetscViewerASCIIPrintf(viewer, "\n"));
932: }
933: } else {
934: for (i = 0; i < m; i++) {
935: PetscCall(PetscViewerASCIIPrintf(viewer, "row %" PetscInt_FMT ":", i));
936: for (j = a->i[i]; j < a->i[i + 1]; j++) {
937: if (PetscDefined(USE_COMPLEX) && PetscImaginaryPart(a->a[j]) > 0.0) {
938: PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g + %g i)", a->j[j], (double)PetscRealPart(a->a[j]), (double)PetscImaginaryPart(a->a[j])));
939: } else if (PetscDefined(USE_COMPLEX) && PetscImaginaryPart(a->a[j]) < 0.0) {
940: PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g - %g i)", a->j[j], (double)PetscRealPart(a->a[j]), (double)-PetscImaginaryPart(a->a[j])));
941: } else {
942: PetscCall(PetscViewerASCIIPrintf(viewer, " (%" PetscInt_FMT ", %g) ", a->j[j], (double)PetscRealPart(a->a[j])));
943: }
944: }
945: PetscCall(PetscViewerASCIIPrintf(viewer, "\n"));
946: }
947: }
948: PetscCall(PetscViewerASCIIUseTabs(viewer, PETSC_TRUE));
949: }
950: PetscCall(PetscViewerFlush(viewer));
951: PetscFunctionReturn(PETSC_SUCCESS);
952: }
954: #include <petscdraw.h>
955: static PetscErrorCode MatView_SeqAIJ_Draw_Zoom(PetscDraw draw, void *Aa)
956: {
957: Mat A = (Mat)Aa;
958: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
959: PetscInt i, j, m = A->rmap->n;
960: int color;
961: PetscReal xl, yl, xr, yr, x_l, x_r, y_l, y_r;
962: PetscViewer viewer;
963: PetscViewerFormat format;
964: const PetscScalar *aa;
966: PetscFunctionBegin;
967: PetscCall(PetscObjectQuery((PetscObject)A, "Zoomviewer", (PetscObject *)&viewer));
968: PetscCall(PetscViewerGetFormat(viewer, &format));
969: PetscCall(PetscDrawGetCoordinates(draw, &xl, &yl, &xr, &yr));
971: /* loop over matrix elements drawing boxes */
972: PetscCall(MatSeqAIJGetArrayRead(A, &aa));
973: if (format != PETSC_VIEWER_DRAW_CONTOUR) {
974: PetscDrawCollectiveBegin(draw);
975: /* Blue for negative, Cyan for zero and Red for positive */
976: color = PETSC_DRAW_BLUE;
977: for (i = 0; i < m; i++) {
978: y_l = m - i - 1.0;
979: y_r = y_l + 1.0;
980: for (j = a->i[i]; j < a->i[i + 1]; j++) {
981: x_l = a->j[j];
982: x_r = x_l + 1.0;
983: if (PetscRealPart(aa[j]) >= 0.) continue;
984: PetscCall(PetscDrawRectangle(draw, x_l, y_l, x_r, y_r, color, color, color, color));
985: }
986: }
987: color = PETSC_DRAW_CYAN;
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 (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_RED;
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 (PetscRealPart(aa[j]) <= 0.) continue;
1006: PetscCall(PetscDrawRectangle(draw, x_l, y_l, x_r, y_r, color, color, color, color));
1007: }
1008: }
1009: PetscDrawCollectiveEnd(draw);
1010: } else {
1011: /* use contour shading to indicate magnitude of values */
1012: /* first determine max of all nonzero values */
1013: PetscReal minv = 0.0, maxv = 0.0;
1014: PetscInt nz = a->nz, count = 0;
1015: PetscDraw popup;
1017: for (i = 0; i < nz; i++) {
1018: if (PetscAbsScalar(aa[i]) > maxv) maxv = PetscAbsScalar(aa[i]);
1019: }
1020: if (minv >= maxv) maxv = minv + PETSC_SMALL;
1021: PetscCall(PetscDrawGetPopup(draw, &popup));
1022: PetscCall(PetscDrawScalePopup(popup, minv, maxv));
1024: PetscDrawCollectiveBegin(draw);
1025: for (i = 0; i < m; i++) {
1026: y_l = m - i - 1.0;
1027: y_r = y_l + 1.0;
1028: for (j = a->i[i]; j < a->i[i + 1]; j++) {
1029: x_l = a->j[j];
1030: x_r = x_l + 1.0;
1031: color = PetscDrawRealToColor(PetscAbsScalar(aa[count]), minv, maxv);
1032: PetscCall(PetscDrawRectangle(draw, x_l, y_l, x_r, y_r, color, color, color, color));
1033: count++;
1034: }
1035: }
1036: PetscDrawCollectiveEnd(draw);
1037: }
1038: PetscCall(MatSeqAIJRestoreArrayRead(A, &aa));
1039: PetscFunctionReturn(PETSC_SUCCESS);
1040: }
1042: #include <petscdraw.h>
1043: static PetscErrorCode MatView_SeqAIJ_Draw(Mat A, PetscViewer viewer)
1044: {
1045: PetscDraw draw;
1046: PetscReal xr, yr, xl, yl, h, w;
1047: PetscBool isnull;
1049: PetscFunctionBegin;
1050: PetscCall(PetscViewerDrawGetDraw(viewer, 0, &draw));
1051: PetscCall(PetscDrawIsNull(draw, &isnull));
1052: if (isnull) PetscFunctionReturn(PETSC_SUCCESS);
1054: xr = A->cmap->n;
1055: yr = A->rmap->n;
1056: h = yr / 10.0;
1057: w = xr / 10.0;
1058: xr += w;
1059: yr += h;
1060: xl = -w;
1061: yl = -h;
1062: PetscCall(PetscDrawSetCoordinates(draw, xl, yl, xr, yr));
1063: PetscCall(PetscObjectCompose((PetscObject)A, "Zoomviewer", (PetscObject)viewer));
1064: PetscCall(PetscDrawZoom(draw, MatView_SeqAIJ_Draw_Zoom, A));
1065: PetscCall(PetscObjectCompose((PetscObject)A, "Zoomviewer", NULL));
1066: PetscCall(PetscDrawSave(draw));
1067: PetscFunctionReturn(PETSC_SUCCESS);
1068: }
1070: PetscErrorCode MatView_SeqAIJ(Mat A, PetscViewer viewer)
1071: {
1072: PetscBool isascii, isbinary, isdraw;
1074: PetscFunctionBegin;
1075: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERASCII, &isascii));
1076: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERBINARY, &isbinary));
1077: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERDRAW, &isdraw));
1078: if (isascii) PetscCall(MatView_SeqAIJ_ASCII(A, viewer));
1079: else if (isbinary) PetscCall(MatView_SeqAIJ_Binary(A, viewer));
1080: else if (isdraw) PetscCall(MatView_SeqAIJ_Draw(A, viewer));
1081: PetscCall(MatView_SeqAIJ_Inode(A, viewer));
1082: PetscFunctionReturn(PETSC_SUCCESS);
1083: }
1085: PetscErrorCode MatAssemblyEnd_SeqAIJ(Mat A, MatAssemblyType mode)
1086: {
1087: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
1088: PetscInt fshift = 0, i, *ai = a->i, *aj = a->j, *imax = a->imax;
1089: PetscInt m = A->rmap->n, *ip, N, *ailen = a->ilen, rmax = 0;
1090: MatScalar *aa = a->a, *ap;
1091: PetscReal ratio = 0.6;
1093: PetscFunctionBegin;
1094: if (mode == MAT_FLUSH_ASSEMBLY) PetscFunctionReturn(PETSC_SUCCESS);
1095: if (A->was_assembled && A->ass_nonzerostate == A->nonzerostate) {
1096: /* we need to respect users asking to use or not the inodes routine in between matrix assemblies, e.g., via MatSetOption(A, MAT_USE_INODES, val) */
1097: PetscCall(MatAssemblyEnd_SeqAIJ_Inode(A, mode)); /* read the sparsity pattern */
1098: PetscFunctionReturn(PETSC_SUCCESS);
1099: }
1101: if (m) rmax = ailen[0]; /* determine row with most nonzeros */
1102: for (i = 1; i < m; i++) {
1103: /* move each row back by the amount of empty slots (fshift) before it*/
1104: fshift += imax[i - 1] - ailen[i - 1];
1105: rmax = PetscMax(rmax, ailen[i]);
1106: if (fshift) {
1107: ip = aj + ai[i];
1108: ap = aa + ai[i];
1109: N = ailen[i];
1110: PetscCall(PetscArraymove(ip - fshift, ip, N));
1111: if (!A->structure_only) PetscCall(PetscArraymove(ap - fshift, ap, N));
1112: }
1113: ai[i] = ai[i - 1] + ailen[i - 1];
1114: }
1115: if (m) {
1116: fshift += imax[m - 1] - ailen[m - 1];
1117: ai[m] = ai[m - 1] + ailen[m - 1];
1118: }
1119: /* reset ilen and imax for each row */
1120: a->nonzerorowcnt = 0;
1121: if (A->structure_only) {
1122: PetscCall(PetscFree(a->imax));
1123: PetscCall(PetscFree(a->ilen));
1124: } else { /* !A->structure_only */
1125: for (i = 0; i < m; i++) {
1126: ailen[i] = imax[i] = ai[i + 1] - ai[i];
1127: a->nonzerorowcnt += ((ai[i + 1] - ai[i]) > 0);
1128: }
1129: }
1130: a->nz = ai[m];
1131: 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);
1132: 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));
1133: PetscCall(PetscInfo(A, "Number of mallocs during MatSetValues() is %" PetscInt_FMT "\n", a->reallocs));
1134: PetscCall(PetscInfo(A, "Maximum nonzeros in any row is %" PetscInt_FMT "\n", rmax));
1136: A->info.mallocs += a->reallocs;
1137: a->reallocs = 0;
1138: A->info.nz_unneeded = (PetscReal)fshift;
1139: a->rmax = rmax;
1141: if (!A->structure_only) PetscCall(MatCheckCompressedRow(A, a->nonzerorowcnt, &a->compressedrow, a->i, m, ratio));
1142: PetscCall(MatAssemblyEnd_SeqAIJ_Inode(A, mode));
1143: PetscFunctionReturn(PETSC_SUCCESS);
1144: }
1146: static PetscErrorCode MatRealPart_SeqAIJ(Mat A)
1147: {
1148: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
1149: PetscInt i, nz = a->nz;
1150: MatScalar *aa;
1152: PetscFunctionBegin;
1153: PetscCall(MatSeqAIJGetArray(A, &aa));
1154: for (i = 0; i < nz; i++) aa[i] = PetscRealPart(aa[i]);
1155: PetscCall(MatSeqAIJRestoreArray(A, &aa));
1156: PetscFunctionReturn(PETSC_SUCCESS);
1157: }
1159: static PetscErrorCode MatImaginaryPart_SeqAIJ(Mat A)
1160: {
1161: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
1162: PetscInt i, nz = a->nz;
1163: MatScalar *aa;
1165: PetscFunctionBegin;
1166: PetscCall(MatSeqAIJGetArray(A, &aa));
1167: for (i = 0; i < nz; i++) aa[i] = PetscImaginaryPart(aa[i]);
1168: PetscCall(MatSeqAIJRestoreArray(A, &aa));
1169: PetscFunctionReturn(PETSC_SUCCESS);
1170: }
1172: PetscErrorCode MatZeroEntries_SeqAIJ(Mat A)
1173: {
1174: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
1175: MatScalar *aa;
1177: PetscFunctionBegin;
1178: PetscCall(MatSeqAIJGetArrayWrite(A, &aa));
1179: PetscCall(PetscArrayzero(aa, a->i[A->rmap->n]));
1180: PetscCall(MatSeqAIJRestoreArrayWrite(A, &aa));
1181: PetscFunctionReturn(PETSC_SUCCESS);
1182: }
1184: static PetscErrorCode MatReset_SeqAIJ(Mat A)
1185: {
1186: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
1188: PetscFunctionBegin;
1189: if (A->hash_active) {
1190: A->ops[0] = a->cops;
1191: PetscCall(PetscHMapIJVDestroy(&a->ht));
1192: PetscCall(PetscFree(a->dnz));
1193: A->hash_active = PETSC_FALSE;
1194: }
1196: PetscCall(PetscLogObjectState((PetscObject)A, "Rows=%" PetscInt_FMT ", Cols=%" PetscInt_FMT ", NZ=%" PetscInt_FMT, A->rmap->n, A->cmap->n, a->nz));
1197: PetscCall(MatSeqXAIJFreeAIJ(A, &a->a, &a->j, &a->i));
1198: PetscCall(ISDestroy(&a->row));
1199: PetscCall(ISDestroy(&a->col));
1200: PetscCall(PetscFree(a->diag));
1201: PetscCall(PetscFree(a->ibdiag));
1202: a->ibdiagsize = 0;
1203: PetscCall(PetscFree(a->imax));
1204: PetscCall(PetscFree(a->ilen));
1205: PetscCall(PetscFree(a->ipre));
1206: PetscCall(PetscFree3(a->idiag, a->mdiag, a->ssor_work));
1207: PetscCall(PetscFree(a->solve_work));
1208: PetscCall(ISDestroy(&a->icol));
1209: PetscCall(PetscFree(a->saved_values));
1210: a->compressedrow.use = PETSC_FALSE;
1211: PetscCall(PetscFree2(a->compressedrow.i, a->compressedrow.rindex));
1212: PetscCall(MatDestroy_SeqAIJ_Inode(A));
1213: PetscFunctionReturn(PETSC_SUCCESS);
1214: }
1216: static PetscErrorCode MatResetHash_SeqAIJ(Mat A)
1217: {
1218: PetscFunctionBegin;
1219: PetscCall(MatReset_SeqAIJ(A));
1220: PetscCall(MatCreate_SeqAIJ_Inode(A));
1221: PetscCall(MatSetUp_Seq_Hash(A));
1222: A->nonzerostate++;
1223: PetscFunctionReturn(PETSC_SUCCESS);
1224: }
1226: PetscErrorCode MatDestroy_SeqAIJ(Mat A)
1227: {
1228: PetscFunctionBegin;
1229: PetscCall(MatReset_SeqAIJ(A));
1230: PetscCall(PetscFree(A->data));
1232: /* MatMatMultNumeric_SeqAIJ_SeqAIJ_Sorted may allocate this.
1233: That function is so heavily used (sometimes in an hidden way through multnumeric function pointers)
1234: that is hard to properly add this data to the MatProduct data. We free it here to avoid
1235: users reusing the matrix object with different data to incur in obscure segmentation faults
1236: due to different matrix sizes */
1237: PetscCall(PetscObjectCompose((PetscObject)A, "__PETSc__ab_dense", NULL));
1239: PetscCall(PetscObjectChangeTypeName((PetscObject)A, NULL));
1240: PetscCall(PetscObjectComposeFunction((PetscObject)A, "PetscMatlabEnginePut_C", NULL));
1241: PetscCall(PetscObjectComposeFunction((PetscObject)A, "PetscMatlabEngineGet_C", NULL));
1242: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatSeqAIJSetColumnIndices_C", NULL));
1243: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatStoreValues_C", NULL));
1244: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatRetrieveValues_C", NULL));
1245: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_seqsbaij_C", NULL));
1246: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_seqbaij_C", NULL));
1247: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_seqaijperm_C", NULL));
1248: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_seqaijsell_C", NULL));
1249: #if PetscDefined(HAVE_MKL_SPARSE)
1250: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_seqaijmkl_C", NULL));
1251: #endif
1252: #if PetscDefined(HAVE_CUDA)
1253: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_seqaijcusparse_C", NULL));
1254: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_seqaijcusparse_seqaij_C", NULL));
1255: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_seqaij_seqaijcusparse_C", NULL));
1256: #endif
1257: #if PetscDefined(HAVE_HIP)
1258: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_seqaijhipsparse_C", NULL));
1259: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_seqaijhipsparse_seqaij_C", NULL));
1260: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_seqaij_seqaijhipsparse_C", NULL));
1261: #endif
1262: #if PetscDefined(HAVE_KOKKOS_KERNELS)
1263: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_seqaijkokkos_C", NULL));
1264: #endif
1265: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_seqaijcrl_C", NULL));
1266: #if PetscDefined(HAVE_ELEMENTAL)
1267: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_elemental_C", NULL));
1268: #endif
1269: #if PetscDefined(HAVE_SCALAPACK) && (PetscDefined(USE_REAL_SINGLE) || PetscDefined(USE_REAL_DOUBLE))
1270: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_scalapack_C", NULL));
1271: #endif
1272: #if PetscDefined(HAVE_HYPRE)
1273: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_hypre_C", NULL));
1274: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_transpose_seqaij_seqaij_C", NULL));
1275: #endif
1276: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_seqdense_C", NULL));
1277: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_seqsell_C", NULL));
1278: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_is_C", NULL));
1279: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatIsTranspose_C", NULL));
1280: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatIsHermitianTranspose_C", NULL));
1281: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatSeqAIJSetPreallocation_C", NULL));
1282: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatResetPreallocation_C", NULL));
1283: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatResetHash_C", NULL));
1284: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatSeqAIJSetPreallocationCSR_C", NULL));
1285: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatReorderForNonzeroDiagonal_C", NULL));
1286: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_is_seqaij_C", NULL));
1287: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_seqdense_seqaij_C", NULL));
1288: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_seqaij_seqaij_C", NULL));
1289: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatSeqAIJKron_C", NULL));
1290: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatSetPreallocationCOO_C", NULL));
1291: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatSetValuesCOO_C", NULL));
1292: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatFactorGetSolverType_C", NULL));
1293: /* these calls do not belong here: the subclasses Duplicate/Destroy are wrong */
1294: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaijsell_seqaij_C", NULL));
1295: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaijperm_seqaij_C", NULL));
1296: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_seqaijviennacl_C", NULL));
1297: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_seqaijviennacl_seqdense_C", NULL));
1298: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_seqaijviennacl_seqaij_C", NULL));
1299: PetscFunctionReturn(PETSC_SUCCESS);
1300: }
1302: PetscErrorCode MatSetOption_SeqAIJ(Mat A, MatOption op, PetscBool flg)
1303: {
1304: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
1306: PetscFunctionBegin;
1307: switch (op) {
1308: case MAT_ROW_ORIENTED:
1309: a->roworiented = flg;
1310: break;
1311: case MAT_KEEP_NONZERO_PATTERN:
1312: a->keepnonzeropattern = flg;
1313: break;
1314: case MAT_NEW_NONZERO_LOCATIONS:
1315: a->nonew = (flg ? 0 : 1);
1316: break;
1317: case MAT_NEW_NONZERO_LOCATION_ERR:
1318: a->nonew = (flg ? -1 : 0);
1319: break;
1320: case MAT_NEW_NONZERO_ALLOCATION_ERR:
1321: a->nonew = (flg ? -2 : 0);
1322: break;
1323: case MAT_UNUSED_NONZERO_LOCATION_ERR:
1324: a->nounused = (flg ? -1 : 0);
1325: break;
1326: case MAT_IGNORE_ZERO_ENTRIES:
1327: a->ignorezeroentries = flg;
1328: break;
1329: case MAT_USE_INODES:
1330: PetscCall(MatSetOption_SeqAIJ_Inode(A, MAT_USE_INODES, flg));
1331: break;
1332: case MAT_SUBMAT_SINGLEIS:
1333: A->submat_singleis = flg;
1334: break;
1335: case MAT_SORTED_FULL:
1336: if (flg) A->ops->setvalues = MatSetValues_SeqAIJ_SortedFull;
1337: else A->ops->setvalues = MatSetValues_SeqAIJ;
1338: break;
1339: case MAT_FORM_EXPLICIT_TRANSPOSE:
1340: A->form_explicit_transpose = flg;
1341: break;
1342: default:
1343: break;
1344: }
1345: PetscFunctionReturn(PETSC_SUCCESS);
1346: }
1348: PETSC_INTERN PetscErrorCode MatGetDiagonal_SeqAIJ(Mat A, Vec v)
1349: {
1350: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
1351: PetscInt n, *ai = a->i;
1352: PetscScalar *x;
1353: const PetscScalar *aa;
1354: const PetscInt *diag;
1355: PetscBool diagDense;
1357: PetscFunctionBegin;
1358: PetscCall(VecGetLocalSize(v, &n));
1359: PetscCheck(n == A->rmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Nonconforming matrix and vector");
1360: PetscCall(MatSeqAIJGetArrayRead(A, &aa));
1361: if (A->factortype == MAT_FACTOR_ILU || A->factortype == MAT_FACTOR_LU) {
1362: PetscCall(MatGetDiagonalMarkers_SeqAIJ(A, &diag, NULL));
1363: PetscCall(VecGetArrayWrite(v, &x));
1364: for (PetscInt i = 0; i < n; i++) x[i] = 1.0 / aa[diag[i]];
1365: PetscCall(VecRestoreArrayWrite(v, &x));
1366: PetscCall(MatSeqAIJRestoreArrayRead(A, &aa));
1367: PetscFunctionReturn(PETSC_SUCCESS);
1368: }
1370: PetscCheck(A->factortype == MAT_FACTOR_NONE, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Not for factor matrices that are not ILU or LU");
1371: PetscCall(MatGetDiagonalMarkers_SeqAIJ(A, &diag, &diagDense));
1372: PetscCall(VecGetArrayWrite(v, &x));
1373: if (diagDense) {
1374: for (PetscInt i = 0; i < n; i++) x[i] = aa[diag[i]];
1375: } else {
1376: for (PetscInt i = 0; i < n; i++) x[i] = (diag[i] == ai[i + 1]) ? 0.0 : aa[diag[i]];
1377: }
1378: PetscCall(VecRestoreArrayWrite(v, &x));
1379: PetscCall(MatSeqAIJRestoreArrayRead(A, &aa));
1380: PetscFunctionReturn(PETSC_SUCCESS);
1381: }
1383: #include <../src/mat/impls/aij/seq/ftn-kernels/fmult.h>
1384: PetscErrorCode MatMultTransposeAdd_SeqAIJ(Mat A, Vec xx, Vec zz, Vec yy)
1385: {
1386: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
1387: const MatScalar *aa;
1388: PetscScalar *y;
1389: const PetscScalar *x;
1390: PetscInt m = A->rmap->n;
1391: #if !PetscDefined(USE_FORTRAN_KERNEL_MULTTRANSPOSEAIJ)
1392: const MatScalar *v;
1393: PetscScalar alpha;
1394: PetscInt n, i, j;
1395: const PetscInt *idx, *ii, *ridx = NULL;
1396: Mat_CompressedRow cprow = a->compressedrow;
1397: PetscBool usecprow = cprow.use;
1398: #endif
1400: PetscFunctionBegin;
1401: if (zz != yy) PetscCall(VecCopy(zz, yy));
1402: PetscCall(VecGetArrayRead(xx, &x));
1403: PetscCall(VecGetArray(yy, &y));
1404: PetscCall(MatSeqAIJGetArrayRead(A, &aa));
1406: #if PetscDefined(USE_FORTRAN_KERNEL_MULTTRANSPOSEAIJ)
1407: fortranmulttransposeaddaij_(&m, x, a->i, a->j, aa, y);
1408: #else
1409: if (usecprow) {
1410: m = cprow.nrows;
1411: ii = cprow.i;
1412: ridx = cprow.rindex;
1413: } else {
1414: ii = a->i;
1415: }
1416: for (i = 0; i < m; i++) {
1417: idx = a->j + ii[i];
1418: v = aa + ii[i];
1419: n = ii[i + 1] - ii[i];
1420: if (usecprow) {
1421: alpha = x[ridx[i]];
1422: } else {
1423: alpha = x[i];
1424: }
1425: for (j = 0; j < n; j++) y[idx[j]] += alpha * v[j];
1426: }
1427: #endif
1428: PetscCall(PetscLogFlops(2.0 * a->nz));
1429: PetscCall(VecRestoreArrayRead(xx, &x));
1430: PetscCall(VecRestoreArray(yy, &y));
1431: PetscCall(MatSeqAIJRestoreArrayRead(A, &aa));
1432: PetscFunctionReturn(PETSC_SUCCESS);
1433: }
1435: PetscErrorCode MatMultTranspose_SeqAIJ(Mat A, Vec xx, Vec yy)
1436: {
1437: PetscFunctionBegin;
1438: PetscCall(VecSet(yy, 0.0));
1439: PetscCall(MatMultTransposeAdd_SeqAIJ(A, xx, yy, yy));
1440: PetscFunctionReturn(PETSC_SUCCESS);
1441: }
1443: #include <../src/mat/impls/aij/seq/ftn-kernels/fmult.h>
1445: PetscErrorCode MatMult_SeqAIJ(Mat A, Vec xx, Vec yy)
1446: {
1447: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
1448: PetscScalar *y;
1449: const PetscScalar *x;
1450: const MatScalar *a_a;
1451: PetscInt m = A->rmap->n;
1452: const PetscInt *ii, *ridx = NULL;
1453: PetscBool usecprow = a->compressedrow.use;
1455: #if PetscDefined(HAVE_PRAGMA_DISJOINT)
1456: #pragma disjoint(*x, *y, *aa)
1457: #endif
1459: PetscFunctionBegin;
1460: if (a->inode.use && a->inode.checked) {
1461: PetscCall(MatMult_SeqAIJ_Inode(A, xx, yy));
1462: PetscFunctionReturn(PETSC_SUCCESS);
1463: }
1464: PetscCall(MatSeqAIJGetArrayRead(A, &a_a));
1465: PetscCall(VecGetArrayRead(xx, &x));
1466: PetscCall(VecGetArray(yy, &y));
1467: ii = a->i;
1468: if (usecprow) { /* use compressed row format */
1469: PetscCall(PetscArrayzero(y, m));
1470: m = a->compressedrow.nrows;
1471: ii = a->compressedrow.i;
1472: ridx = a->compressedrow.rindex;
1473: PetscPragmaUseOMPKernels(parallel for)
1474: for (PetscInt i = 0; i < m; i++) {
1475: PetscInt n = ii[i + 1] - ii[i];
1476: const PetscInt *aj = a->j + ii[i];
1477: const PetscScalar *aa = a_a + ii[i];
1478: PetscScalar sum = 0.0;
1479: PetscSparseDensePlusDot(sum, x, aa, aj, n);
1480: /* for (j=0; j<n; j++) sum += (*aa++)*x[*aj++]; */
1481: y[ridx[i]] = sum;
1482: }
1483: } else { /* do not use compressed row format */
1484: #if PetscDefined(USE_FORTRAN_KERNEL_MULTAIJ)
1485: fortranmultaij_(&m, x, ii, a->j, a_a, y);
1486: #else
1487: PetscPragmaUseOMPKernels(parallel for)
1488: for (PetscInt i = 0; i < m; i++) {
1489: PetscInt n = ii[i + 1] - ii[i];
1490: const PetscInt *aj = a->j + ii[i];
1491: const PetscScalar *aa = a_a + ii[i];
1492: PetscScalar sum = 0.0;
1493: PetscSparseDensePlusDot(sum, x, aa, aj, n);
1494: y[i] = sum;
1495: }
1496: #endif
1497: }
1498: PetscCall(PetscLogFlops(2.0 * a->nz - a->nonzerorowcnt));
1499: PetscCall(VecRestoreArrayRead(xx, &x));
1500: PetscCall(VecRestoreArray(yy, &y));
1501: PetscCall(MatSeqAIJRestoreArrayRead(A, &a_a));
1502: PetscFunctionReturn(PETSC_SUCCESS);
1503: }
1505: // HACK!!!!! Used by src/mat/tests/ex170.c
1506: PETSC_EXTERN PetscErrorCode MatMultMax_SeqAIJ(Mat A, Vec xx, Vec yy)
1507: {
1508: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
1509: PetscScalar *y;
1510: const PetscScalar *x;
1511: const MatScalar *aa, *a_a;
1512: PetscInt m = A->rmap->n;
1513: const PetscInt *aj, *ii, *ridx = NULL;
1514: PetscInt n, i, nonzerorow = 0;
1515: PetscScalar sum;
1516: PetscBool usecprow = a->compressedrow.use;
1518: #if PetscDefined(HAVE_PRAGMA_DISJOINT)
1519: #pragma disjoint(*x, *y, *aa)
1520: #endif
1522: PetscFunctionBegin;
1523: PetscCall(MatSeqAIJGetArrayRead(A, &a_a));
1524: PetscCall(VecGetArrayRead(xx, &x));
1525: PetscCall(VecGetArray(yy, &y));
1526: if (usecprow) { /* use compressed row format */
1527: m = a->compressedrow.nrows;
1528: ii = a->compressedrow.i;
1529: ridx = a->compressedrow.rindex;
1530: for (i = 0; i < m; i++) {
1531: n = ii[i + 1] - ii[i];
1532: aj = a->j + ii[i];
1533: aa = a_a + ii[i];
1534: sum = 0.0;
1535: nonzerorow += (n > 0);
1536: PetscSparseDenseMaxDot(sum, x, aa, aj, n);
1537: /* for (j=0; j<n; j++) sum += (*aa++)*x[*aj++]; */
1538: y[*ridx++] = sum;
1539: }
1540: } else { /* do not use compressed row format */
1541: ii = a->i;
1542: for (i = 0; i < m; i++) {
1543: n = ii[i + 1] - ii[i];
1544: aj = a->j + ii[i];
1545: aa = a_a + ii[i];
1546: sum = 0.0;
1547: nonzerorow += (n > 0);
1548: PetscSparseDenseMaxDot(sum, x, aa, aj, n);
1549: y[i] = sum;
1550: }
1551: }
1552: PetscCall(PetscLogFlops(2.0 * a->nz - nonzerorow));
1553: PetscCall(VecRestoreArrayRead(xx, &x));
1554: PetscCall(VecRestoreArray(yy, &y));
1555: PetscCall(MatSeqAIJRestoreArrayRead(A, &a_a));
1556: PetscFunctionReturn(PETSC_SUCCESS);
1557: }
1559: // HACK!!!!! Used by src/mat/tests/ex170.c
1560: PETSC_EXTERN PetscErrorCode MatMultAddMax_SeqAIJ(Mat A, Vec xx, Vec yy, Vec zz)
1561: {
1562: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
1563: PetscScalar *y, *z;
1564: const PetscScalar *x;
1565: const MatScalar *aa, *a_a;
1566: PetscInt m = A->rmap->n, *aj, *ii;
1567: PetscInt n, i, *ridx = NULL;
1568: PetscScalar sum;
1569: PetscBool usecprow = a->compressedrow.use;
1571: PetscFunctionBegin;
1572: PetscCall(MatSeqAIJGetArrayRead(A, &a_a));
1573: PetscCall(VecGetArrayRead(xx, &x));
1574: PetscCall(VecGetArrayPair(yy, zz, &y, &z));
1575: if (usecprow) { /* use compressed row format */
1576: if (zz != yy) PetscCall(PetscArraycpy(z, y, m));
1577: m = a->compressedrow.nrows;
1578: ii = a->compressedrow.i;
1579: ridx = a->compressedrow.rindex;
1580: for (i = 0; i < m; i++) {
1581: n = ii[i + 1] - ii[i];
1582: aj = a->j + ii[i];
1583: aa = a_a + ii[i];
1584: sum = y[*ridx];
1585: PetscSparseDenseMaxDot(sum, x, aa, aj, n);
1586: z[*ridx++] = sum;
1587: }
1588: } else { /* do not use compressed row format */
1589: ii = a->i;
1590: for (i = 0; i < m; i++) {
1591: n = ii[i + 1] - ii[i];
1592: aj = a->j + ii[i];
1593: aa = a_a + ii[i];
1594: sum = y[i];
1595: PetscSparseDenseMaxDot(sum, x, aa, aj, n);
1596: z[i] = sum;
1597: }
1598: }
1599: PetscCall(PetscLogFlops(2.0 * a->nz));
1600: PetscCall(VecRestoreArrayRead(xx, &x));
1601: PetscCall(VecRestoreArrayPair(yy, zz, &y, &z));
1602: PetscCall(MatSeqAIJRestoreArrayRead(A, &a_a));
1603: PetscFunctionReturn(PETSC_SUCCESS);
1604: }
1606: #include <../src/mat/impls/aij/seq/ftn-kernels/fmultadd.h>
1607: PetscErrorCode MatMultAdd_SeqAIJ(Mat A, Vec xx, Vec yy, Vec zz)
1608: {
1609: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
1610: PetscScalar *y, *z;
1611: const PetscScalar *x;
1612: const MatScalar *a_a;
1613: const PetscInt *ii, *ridx = NULL;
1614: PetscInt m = A->rmap->n;
1615: PetscBool usecprow = a->compressedrow.use;
1617: PetscFunctionBegin;
1618: if (a->inode.use && a->inode.checked) {
1619: PetscCall(MatMultAdd_SeqAIJ_Inode(A, xx, yy, zz));
1620: PetscFunctionReturn(PETSC_SUCCESS);
1621: }
1622: PetscCall(MatSeqAIJGetArrayRead(A, &a_a));
1623: PetscCall(VecGetArrayRead(xx, &x));
1624: PetscCall(VecGetArrayPair(yy, zz, &y, &z));
1625: if (usecprow) { /* use compressed row format */
1626: if (zz != yy) PetscCall(PetscArraycpy(z, y, m));
1627: m = a->compressedrow.nrows;
1628: ii = a->compressedrow.i;
1629: ridx = a->compressedrow.rindex;
1630: for (PetscInt i = 0; i < m; i++) {
1631: PetscInt n = ii[i + 1] - ii[i];
1632: const PetscInt *aj = a->j + ii[i];
1633: const PetscScalar *aa = a_a + ii[i];
1634: PetscScalar sum = y[*ridx];
1635: PetscSparseDensePlusDot(sum, x, aa, aj, n);
1636: z[*ridx++] = sum;
1637: }
1638: } else { /* do not use compressed row format */
1639: ii = a->i;
1640: #if PetscDefined(USE_FORTRAN_KERNEL_MULTADDAIJ)
1641: fortranmultaddaij_(&m, x, ii, a->j, a_a, y, z);
1642: #else
1643: PetscPragmaUseOMPKernels(parallel for)
1644: for (PetscInt i = 0; i < m; i++) {
1645: PetscInt n = ii[i + 1] - ii[i];
1646: const PetscInt *aj = a->j + ii[i];
1647: const PetscScalar *aa = a_a + ii[i];
1648: PetscScalar sum = y[i];
1649: PetscSparseDensePlusDot(sum, x, aa, aj, n);
1650: z[i] = sum;
1651: }
1652: #endif
1653: }
1654: PetscCall(PetscLogFlops(2.0 * a->nz));
1655: PetscCall(VecRestoreArrayRead(xx, &x));
1656: PetscCall(VecRestoreArrayPair(yy, zz, &y, &z));
1657: PetscCall(MatSeqAIJRestoreArrayRead(A, &a_a));
1658: PetscFunctionReturn(PETSC_SUCCESS);
1659: }
1661: static PetscErrorCode MatShift_SeqAIJ(Mat A, PetscScalar v)
1662: {
1663: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
1664: const PetscInt *diag;
1665: const PetscInt *ii = (const PetscInt *)a->i;
1666: PetscBool diagDense;
1668: PetscFunctionBegin;
1669: if (!A->preallocated || !a->nz) {
1670: PetscCall(MatSeqAIJSetPreallocation(A, 1, NULL));
1671: PetscCall(MatShift_Basic(A, v));
1672: PetscFunctionReturn(PETSC_SUCCESS);
1673: }
1675: PetscCall(MatGetDiagonalMarkers_SeqAIJ(A, &diag, &diagDense));
1676: if (diagDense) {
1677: PetscScalar *Aa;
1679: PetscCall(MatSeqAIJGetArray(A, &Aa));
1680: for (PetscInt i = 0; i < A->rmap->n; i++) Aa[diag[i]] += v;
1681: PetscCall(MatSeqAIJRestoreArray(A, &Aa));
1682: } else {
1683: PetscScalar *olda = a->a; /* preserve pointers to current matrix nonzeros structure and values */
1684: PetscInt *oldj = a->j, *oldi = a->i;
1685: PetscBool free_a = a->free_a, free_ij = a->free_ij;
1686: const PetscScalar *Aa;
1687: PetscInt *mdiag = NULL;
1689: PetscCall(PetscCalloc1(A->rmap->n, &mdiag));
1690: for (PetscInt i = 0; i < A->rmap->n; i++) {
1691: if (i < A->cmap->n && diag[i] >= ii[i + 1]) { /* 'out of range' rows never have diagonals */
1692: mdiag[i] = 1;
1693: }
1694: }
1695: PetscCall(MatSeqAIJGetArrayRead(A, &Aa)); // sync the host
1696: PetscCall(MatSeqAIJRestoreArrayRead(A, &Aa));
1698: a->a = NULL;
1699: a->j = NULL;
1700: a->i = NULL;
1701: /* increase the values in imax for each row where a diagonal is being inserted then reallocate the matrix data structures */
1702: for (PetscInt i = 0; i < PetscMin(A->rmap->n, A->cmap->n); i++) a->imax[i] += mdiag[i];
1703: PetscCall(MatSeqAIJSetPreallocation_SeqAIJ(A, 0, a->imax));
1705: /* copy old values into new matrix data structure */
1706: for (PetscInt i = 0; i < A->rmap->n; i++) {
1707: PetscCall(MatSetValues(A, 1, &i, a->imax[i] - mdiag[i], &oldj[oldi[i]], &olda[oldi[i]], ADD_VALUES));
1708: if (i < A->cmap->n) PetscCall(MatSetValue(A, i, i, v, ADD_VALUES));
1709: }
1710: PetscCall(MatAssemblyBegin(A, MAT_FINAL_ASSEMBLY));
1711: PetscCall(MatAssemblyEnd(A, MAT_FINAL_ASSEMBLY));
1712: if (free_a) PetscCall(PetscShmgetDeallocateArray((void **)&olda));
1713: if (free_ij) PetscCall(PetscShmgetDeallocateArray((void **)&oldj));
1714: if (free_ij) PetscCall(PetscShmgetDeallocateArray((void **)&oldi));
1715: PetscCall(PetscFree(mdiag));
1716: }
1717: PetscFunctionReturn(PETSC_SUCCESS);
1718: }
1720: #include <petscblaslapack.h>
1721: #include <petsc/private/kernels/blockinvert.h>
1723: /*
1724: Note that values is allocated externally by the PC and then passed into this routine
1725: */
1726: static PetscErrorCode MatInvertVariableBlockDiagonal_SeqAIJ(Mat A, PetscInt nblocks, const PetscInt *bsizes, PetscScalar *diag)
1727: {
1728: PetscInt n = A->rmap->n, i, ncnt = 0, *indx, j, bsizemax = 0, *v_pivots;
1729: PetscBool allowzeropivot, zeropivotdetected = PETSC_FALSE;
1730: const PetscReal shift = 0.0;
1731: PetscInt ipvt[5];
1732: PetscCount flops = 0;
1733: PetscScalar work[25], *v_work;
1735: PetscFunctionBegin;
1736: allowzeropivot = PetscNot(A->erroriffailure);
1737: for (i = 0; i < nblocks; i++) ncnt += bsizes[i];
1738: PetscCheck(ncnt == n, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Total blocksizes %" PetscInt_FMT " doesn't match number matrix rows %" PetscInt_FMT, ncnt, n);
1739: for (i = 0; i < nblocks; i++) bsizemax = PetscMax(bsizemax, bsizes[i]);
1740: PetscCall(PetscMalloc1(bsizemax, &indx));
1741: if (bsizemax > 7) PetscCall(PetscMalloc2(bsizemax, &v_work, bsizemax, &v_pivots));
1742: ncnt = 0;
1743: for (i = 0; i < nblocks; i++) {
1744: for (j = 0; j < bsizes[i]; j++) indx[j] = ncnt + j;
1745: PetscCall(MatGetValues(A, bsizes[i], indx, bsizes[i], indx, diag));
1746: switch (bsizes[i]) {
1747: case 1:
1748: *diag = 1.0 / (*diag);
1749: break;
1750: case 2:
1751: PetscCall(PetscKernel_A_gets_inverse_A_2(diag, shift, allowzeropivot, &zeropivotdetected));
1752: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
1753: PetscCall(PetscKernel_A_gets_transpose_A_2(diag));
1754: break;
1755: case 3:
1756: PetscCall(PetscKernel_A_gets_inverse_A_3(diag, shift, allowzeropivot, &zeropivotdetected));
1757: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
1758: PetscCall(PetscKernel_A_gets_transpose_A_3(diag));
1759: break;
1760: case 4:
1761: PetscCall(PetscKernel_A_gets_inverse_A_4(diag, shift, allowzeropivot, &zeropivotdetected));
1762: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
1763: PetscCall(PetscKernel_A_gets_transpose_A_4(diag));
1764: break;
1765: case 5:
1766: PetscCall(PetscKernel_A_gets_inverse_A_5(diag, ipvt, work, shift, allowzeropivot, &zeropivotdetected));
1767: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
1768: PetscCall(PetscKernel_A_gets_transpose_A_5(diag));
1769: break;
1770: case 6:
1771: PetscCall(PetscKernel_A_gets_inverse_A_6(diag, shift, allowzeropivot, &zeropivotdetected));
1772: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
1773: PetscCall(PetscKernel_A_gets_transpose_A_6(diag));
1774: break;
1775: case 7:
1776: PetscCall(PetscKernel_A_gets_inverse_A_7(diag, shift, allowzeropivot, &zeropivotdetected));
1777: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
1778: PetscCall(PetscKernel_A_gets_transpose_A_7(diag));
1779: break;
1780: default:
1781: PetscCall(PetscKernel_A_gets_inverse_A(bsizes[i], diag, v_pivots, v_work, allowzeropivot, &zeropivotdetected));
1782: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
1783: PetscCall(PetscKernel_A_gets_transpose_A_N(diag, bsizes[i]));
1784: }
1785: ncnt += bsizes[i];
1786: diag += bsizes[i] * bsizes[i];
1787: flops += 2 * PetscPowInt64(bsizes[i], 3) / 3;
1788: }
1789: PetscCall(PetscLogFlops(flops));
1790: if (bsizemax > 7) PetscCall(PetscFree2(v_work, v_pivots));
1791: PetscCall(PetscFree(indx));
1792: PetscFunctionReturn(PETSC_SUCCESS);
1793: }
1795: /*
1796: Negative shift indicates do not generate an error if there is a zero diagonal, just invert it anyways
1797: */
1798: static PetscErrorCode MatInvertDiagonalForSOR_SeqAIJ(Mat A, PetscScalar omega, PetscScalar fshift)
1799: {
1800: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
1801: PetscInt i, m = A->rmap->n;
1802: const MatScalar *v;
1803: PetscScalar *idiag, *mdiag;
1804: PetscBool diagDense;
1805: const PetscInt *diag;
1807: PetscFunctionBegin;
1808: if (a->idiagState == ((PetscObject)A)->state && a->omega == omega && a->fshift == fshift) PetscFunctionReturn(PETSC_SUCCESS);
1809: PetscCall(MatGetDiagonalMarkers_SeqAIJ(A, &diag, &diagDense));
1810: PetscCheck(diagDense, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Matrix must have all diagonal locations to invert them");
1811: if (!a->idiag) PetscCall(PetscMalloc3(m, &a->idiag, m, &a->mdiag, m, &a->ssor_work));
1813: mdiag = a->mdiag;
1814: idiag = a->idiag;
1815: PetscCall(MatSeqAIJGetArrayRead(A, &v));
1816: if (omega == 1.0 && PetscRealPart(fshift) <= 0.0) {
1817: for (i = 0; i < m; i++) {
1818: mdiag[i] = v[diag[i]];
1819: if (!PetscAbsScalar(mdiag[i])) { /* zero diagonal */
1820: PetscCheck(PetscRealPart(fshift), PETSC_COMM_SELF, PETSC_ERR_ARG_INCOMP, "Zero diagonal on row %" PetscInt_FMT, i);
1821: PetscCall(PetscInfo(A, "Zero diagonal on row %" PetscInt_FMT "\n", i));
1822: A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
1823: A->factorerror_zeropivot_value = 0.0;
1824: A->factorerror_zeropivot_row = i;
1825: }
1826: idiag[i] = 1.0 / v[diag[i]];
1827: }
1828: PetscCall(PetscLogFlops(m));
1829: } else {
1830: for (i = 0; i < m; i++) {
1831: mdiag[i] = v[diag[i]];
1832: idiag[i] = omega / (fshift + v[diag[i]]);
1833: }
1834: PetscCall(PetscLogFlops(2.0 * m));
1835: }
1836: PetscCall(MatSeqAIJRestoreArrayRead(A, &v));
1837: a->idiagState = ((PetscObject)A)->state;
1838: a->omega = omega;
1839: a->fshift = fshift;
1840: PetscFunctionReturn(PETSC_SUCCESS);
1841: }
1843: PetscErrorCode MatSOR_SeqAIJ(Mat A, Vec bb, PetscReal omega, MatSORType flag, PetscReal fshift, PetscInt its, PetscInt lits, Vec xx)
1844: {
1845: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
1846: PetscScalar *x, d, sum, *t, scale;
1847: const MatScalar *v, *idiag = NULL, *mdiag, *aa;
1848: const PetscScalar *b, *bs, *xb, *ts;
1849: PetscInt n, m = A->rmap->n, i;
1850: const PetscInt *idx, *diag;
1852: PetscFunctionBegin;
1853: if (a->inode.use && a->inode.checked && omega == 1.0 && fshift == 0.0) {
1854: PetscCall(MatSOR_SeqAIJ_Inode(A, bb, omega, flag, fshift, its, lits, xx));
1855: PetscFunctionReturn(PETSC_SUCCESS);
1856: }
1857: its = its * lits;
1858: PetscCall(MatInvertDiagonalForSOR_SeqAIJ(A, omega, fshift));
1859: PetscCall(MatGetDiagonalMarkers_SeqAIJ(A, &diag, NULL));
1860: t = a->ssor_work;
1861: idiag = a->idiag;
1862: mdiag = a->mdiag;
1864: PetscCall(MatSeqAIJGetArrayRead(A, &aa));
1865: PetscCall(VecGetArray(xx, &x));
1866: PetscCall(VecGetArrayRead(bb, &b));
1867: /* We count flops by assuming the upper triangular and lower triangular parts have the same number of nonzeros */
1868: if (flag == SOR_APPLY_UPPER) {
1869: /* apply (U + D/omega) to the vector */
1870: bs = b;
1871: for (i = 0; i < m; i++) {
1872: d = fshift + mdiag[i];
1873: n = a->i[i + 1] - diag[i] - 1;
1874: idx = a->j + diag[i] + 1;
1875: v = aa + diag[i] + 1;
1876: sum = b[i] * d / omega;
1877: PetscSparseDensePlusDot(sum, bs, v, idx, n);
1878: x[i] = sum;
1879: }
1880: PetscCall(VecRestoreArray(xx, &x));
1881: PetscCall(VecRestoreArrayRead(bb, &b));
1882: PetscCall(MatSeqAIJRestoreArrayRead(A, &aa));
1883: PetscCall(PetscLogFlops(a->nz));
1884: PetscFunctionReturn(PETSC_SUCCESS);
1885: }
1887: PetscCheck(flag != SOR_APPLY_LOWER, PETSC_COMM_SELF, PETSC_ERR_SUP, "SOR_APPLY_LOWER is not implemented");
1888: if (flag & SOR_EISENSTAT) {
1889: /* Let A = L + U + D; where L is lower triangular,
1890: U is upper triangular, E = D/omega; This routine applies
1892: (L + E)^{-1} A (U + E)^{-1}
1894: to a vector efficiently using Eisenstat's trick.
1895: */
1896: scale = (2.0 / omega) - 1.0;
1898: /* x = (E + U)^{-1} b */
1899: for (i = m - 1; i >= 0; i--) {
1900: n = a->i[i + 1] - diag[i] - 1;
1901: idx = a->j + diag[i] + 1;
1902: v = aa + diag[i] + 1;
1903: sum = b[i];
1904: PetscSparseDenseMinusDot(sum, x, v, idx, n);
1905: x[i] = sum * idiag[i];
1906: }
1908: /* t = b - (2*E - D)x */
1909: v = aa;
1910: for (i = 0; i < m; i++) t[i] = b[i] - scale * (v[*diag++]) * x[i];
1912: /* t = (E + L)^{-1}t */
1913: ts = t;
1914: diag = a->diag;
1915: for (i = 0; i < m; i++) {
1916: n = diag[i] - a->i[i];
1917: idx = a->j + a->i[i];
1918: v = aa + a->i[i];
1919: sum = t[i];
1920: PetscSparseDenseMinusDot(sum, ts, v, idx, n);
1921: t[i] = sum * idiag[i];
1922: /* x = x + t */
1923: x[i] += t[i];
1924: }
1926: PetscCall(PetscLogFlops(6.0 * m - 1 + 2.0 * a->nz));
1927: PetscCall(VecRestoreArray(xx, &x));
1928: PetscCall(VecRestoreArrayRead(bb, &b));
1929: PetscFunctionReturn(PETSC_SUCCESS);
1930: }
1931: if (flag & SOR_ZERO_INITIAL_GUESS) {
1932: if (flag & SOR_FORWARD_SWEEP || flag & SOR_LOCAL_FORWARD_SWEEP) {
1933: for (i = 0; i < m; i++) {
1934: n = diag[i] - a->i[i];
1935: idx = a->j + a->i[i];
1936: v = aa + a->i[i];
1937: sum = b[i];
1938: PetscSparseDenseMinusDot(sum, x, v, idx, n);
1939: t[i] = sum;
1940: x[i] = sum * idiag[i];
1941: }
1942: xb = t;
1943: PetscCall(PetscLogFlops(a->nz));
1944: } else xb = b;
1945: if (flag & SOR_BACKWARD_SWEEP || flag & SOR_LOCAL_BACKWARD_SWEEP) {
1946: for (i = m - 1; i >= 0; i--) {
1947: n = a->i[i + 1] - diag[i] - 1;
1948: idx = a->j + diag[i] + 1;
1949: v = aa + diag[i] + 1;
1950: sum = xb[i];
1951: PetscSparseDenseMinusDot(sum, x, v, idx, n);
1952: if (xb == b) {
1953: x[i] = sum * idiag[i];
1954: } else {
1955: x[i] = (1 - omega) * x[i] + sum * idiag[i]; /* omega in idiag */
1956: }
1957: }
1958: PetscCall(PetscLogFlops(a->nz)); /* assumes 1/2 in upper */
1959: }
1960: its--;
1961: }
1962: while (its--) {
1963: if (flag & SOR_FORWARD_SWEEP || flag & SOR_LOCAL_FORWARD_SWEEP) {
1964: for (i = 0; i < m; i++) {
1965: /* lower */
1966: n = diag[i] - a->i[i];
1967: idx = a->j + a->i[i];
1968: v = aa + a->i[i];
1969: sum = b[i];
1970: PetscSparseDenseMinusDot(sum, x, v, idx, n);
1971: t[i] = sum; /* save application of the lower-triangular part */
1972: /* upper */
1973: n = a->i[i + 1] - diag[i] - 1;
1974: idx = a->j + diag[i] + 1;
1975: v = aa + diag[i] + 1;
1976: PetscSparseDenseMinusDot(sum, x, v, idx, n);
1977: x[i] = (1. - omega) * x[i] + sum * idiag[i]; /* omega in idiag */
1978: }
1979: xb = t;
1980: PetscCall(PetscLogFlops(2.0 * a->nz));
1981: } else xb = b;
1982: if (flag & SOR_BACKWARD_SWEEP || flag & SOR_LOCAL_BACKWARD_SWEEP) {
1983: for (i = m - 1; i >= 0; i--) {
1984: sum = xb[i];
1985: if (xb == b) {
1986: /* whole matrix (no checkpointing available) */
1987: n = a->i[i + 1] - a->i[i];
1988: idx = a->j + a->i[i];
1989: v = aa + a->i[i];
1990: PetscSparseDenseMinusDot(sum, x, v, idx, n);
1991: x[i] = (1. - omega) * x[i] + (sum + mdiag[i] * x[i]) * idiag[i];
1992: } else { /* lower-triangular part has been saved, so only apply upper-triangular */
1993: n = a->i[i + 1] - diag[i] - 1;
1994: idx = a->j + diag[i] + 1;
1995: v = aa + diag[i] + 1;
1996: PetscSparseDenseMinusDot(sum, x, v, idx, n);
1997: x[i] = (1. - omega) * x[i] + sum * idiag[i]; /* omega in idiag */
1998: }
1999: }
2000: if (xb == b) PetscCall(PetscLogFlops(2.0 * a->nz));
2001: else PetscCall(PetscLogFlops(a->nz)); /* assumes 1/2 in upper */
2002: }
2003: }
2004: PetscCall(MatSeqAIJRestoreArrayRead(A, &aa));
2005: PetscCall(VecRestoreArray(xx, &x));
2006: PetscCall(VecRestoreArrayRead(bb, &b));
2007: PetscFunctionReturn(PETSC_SUCCESS);
2008: }
2010: static PetscErrorCode MatGetInfo_SeqAIJ(Mat A, MatInfoType flag, MatInfo *info)
2011: {
2012: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
2014: PetscFunctionBegin;
2015: info->block_size = 1.0;
2016: info->nz_allocated = a->maxnz;
2017: info->nz_used = a->nz;
2018: info->nz_unneeded = (a->maxnz - a->nz);
2019: info->assemblies = A->num_ass;
2020: info->mallocs = A->info.mallocs;
2021: info->memory = 0; /* REVIEW ME */
2022: if (A->factortype) {
2023: info->fill_ratio_given = A->info.fill_ratio_given;
2024: info->fill_ratio_needed = A->info.fill_ratio_needed;
2025: info->factor_mallocs = A->info.factor_mallocs;
2026: } else {
2027: info->fill_ratio_given = 0;
2028: info->fill_ratio_needed = 0;
2029: info->factor_mallocs = 0;
2030: }
2031: PetscFunctionReturn(PETSC_SUCCESS);
2032: }
2034: static PetscErrorCode MatZeroRows_SeqAIJ(Mat A, PetscInt N, const PetscInt rows[], PetscScalar diagv, Vec x, Vec b)
2035: {
2036: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
2037: PetscInt i, m = A->rmap->n - 1;
2038: const PetscScalar *xx;
2039: PetscScalar *bb, *aa;
2040: PetscInt d = 0;
2041: const PetscInt *diag;
2043: PetscFunctionBegin;
2044: if (x && b) {
2045: PetscCall(VecGetArrayRead(x, &xx));
2046: PetscCall(VecGetArray(b, &bb));
2047: for (i = 0; i < N; i++) {
2048: PetscCheck(rows[i] >= 0 && rows[i] <= m, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "row %" PetscInt_FMT " out of range", rows[i]);
2049: if (rows[i] >= A->cmap->n) continue;
2050: bb[rows[i]] = diagv * xx[rows[i]];
2051: }
2052: PetscCall(VecRestoreArrayRead(x, &xx));
2053: PetscCall(VecRestoreArray(b, &bb));
2054: }
2056: PetscCall(MatGetDiagonalMarkers_SeqAIJ(A, &diag, NULL));
2057: PetscCall(MatSeqAIJGetArray(A, &aa));
2058: if (a->keepnonzeropattern) {
2059: for (i = 0; i < N; i++) {
2060: PetscCheck(rows[i] >= 0 && rows[i] <= m, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "row %" PetscInt_FMT " out of range", rows[i]);
2061: PetscCall(PetscArrayzero(&aa[a->i[rows[i]]], a->ilen[rows[i]]));
2062: }
2063: if (diagv != 0.0) {
2064: for (i = 0; i < N; i++) {
2065: d = rows[i];
2066: if (d >= A->cmap->n) continue;
2067: PetscCheck(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);
2068: aa[diag[d]] = diagv;
2069: }
2070: }
2071: } else {
2072: if (diagv != 0.0) {
2073: for (i = 0; i < N; i++) {
2074: PetscCheck(rows[i] >= 0 && rows[i] <= m, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "row %" PetscInt_FMT " out of range", rows[i]);
2075: if (a->ilen[rows[i]] > 0) {
2076: if (rows[i] >= A->cmap->n) {
2077: a->ilen[rows[i]] = 0;
2078: } else {
2079: a->ilen[rows[i]] = 1;
2080: aa[a->i[rows[i]]] = diagv;
2081: a->j[a->i[rows[i]]] = rows[i];
2082: }
2083: } else if (rows[i] < A->cmap->n) { /* in case row was completely empty */
2084: PetscCall(MatSetValues_SeqAIJ(A, 1, &rows[i], 1, &rows[i], &diagv, INSERT_VALUES));
2085: }
2086: }
2087: } else {
2088: for (i = 0; i < N; i++) {
2089: PetscCheck(rows[i] >= 0 && rows[i] <= m, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "row %" PetscInt_FMT " out of range", rows[i]);
2090: a->ilen[rows[i]] = 0;
2091: }
2092: }
2093: A->nonzerostate++;
2094: }
2095: PetscCall(MatSeqAIJRestoreArray(A, &aa));
2096: PetscUseTypeMethod(A, assemblyend, MAT_FINAL_ASSEMBLY);
2097: PetscFunctionReturn(PETSC_SUCCESS);
2098: }
2100: static PetscErrorCode MatZeroRowsColumns_SeqAIJ(Mat A, PetscInt N, const PetscInt rows[], PetscScalar diagv, Vec x, Vec b)
2101: {
2102: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
2103: PetscInt i, j, m = A->rmap->n - 1, d = 0;
2104: PetscBool *zeroed, vecs = PETSC_FALSE;
2105: const PetscScalar *xx;
2106: PetscScalar *bb, *aa;
2107: const PetscInt *diag;
2108: PetscBool diagDense;
2110: PetscFunctionBegin;
2111: if (!N) PetscFunctionReturn(PETSC_SUCCESS);
2112: PetscCall(MatGetDiagonalMarkers_SeqAIJ(A, &diag, &diagDense));
2113: PetscCall(MatSeqAIJGetArray(A, &aa));
2114: if (x && b) {
2115: PetscCall(VecGetArrayRead(x, &xx));
2116: PetscCall(VecGetArray(b, &bb));
2117: vecs = PETSC_TRUE;
2118: }
2119: PetscCall(PetscCalloc1(A->rmap->n, &zeroed));
2120: for (i = 0; i < N; i++) {
2121: PetscCheck(rows[i] >= 0 && rows[i] <= m, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "row %" PetscInt_FMT " out of range", rows[i]);
2122: PetscCall(PetscArrayzero(PetscSafePointerPlusOffset(aa, a->i[rows[i]]), a->ilen[rows[i]]));
2124: zeroed[rows[i]] = PETSC_TRUE;
2125: }
2126: for (i = 0; i < A->rmap->n; i++) {
2127: if (!zeroed[i]) {
2128: for (j = a->i[i]; j < a->i[i + 1]; j++) {
2129: if (a->j[j] < A->rmap->n && zeroed[a->j[j]]) {
2130: if (vecs) bb[i] -= aa[j] * xx[a->j[j]];
2131: aa[j] = 0.0;
2132: }
2133: }
2134: } else if (vecs && i < A->cmap->N) bb[i] = diagv * xx[i];
2135: }
2136: if (x && b) {
2137: PetscCall(VecRestoreArrayRead(x, &xx));
2138: PetscCall(VecRestoreArray(b, &bb));
2139: }
2140: PetscCall(PetscFree(zeroed));
2141: if (diagv != 0.0) {
2142: if (!diagDense) {
2143: for (i = 0; i < N; i++) {
2144: if (rows[i] >= A->cmap->N) continue;
2145: 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]);
2146: PetscCall(MatSetValues_SeqAIJ(A, 1, &rows[i], 1, &rows[i], &diagv, INSERT_VALUES));
2147: }
2148: } else {
2149: for (i = 0; i < N; i++) aa[diag[rows[i]]] = diagv;
2150: }
2151: }
2152: PetscCall(MatSeqAIJRestoreArray(A, &aa));
2153: if (!diagDense) PetscUseTypeMethod(A, assemblyend, MAT_FINAL_ASSEMBLY);
2154: PetscFunctionReturn(PETSC_SUCCESS);
2155: }
2157: PetscErrorCode MatGetRow_SeqAIJ(Mat A, PetscInt row, PetscInt *nz, PetscInt **idx, PetscScalar **v)
2158: {
2159: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
2160: const PetscScalar *aa;
2162: PetscFunctionBegin;
2163: PetscCall(MatSeqAIJGetArrayRead(A, &aa));
2164: *nz = a->i[row + 1] - a->i[row];
2165: if (v) *v = PetscSafePointerPlusOffset((PetscScalar *)aa, a->i[row]);
2166: if (idx) {
2167: if (*nz && a->j) *idx = a->j + a->i[row];
2168: else *idx = NULL;
2169: }
2170: PetscCall(MatSeqAIJRestoreArrayRead(A, &aa));
2171: PetscFunctionReturn(PETSC_SUCCESS);
2172: }
2174: PetscErrorCode MatRestoreRow_SeqAIJ(Mat A, PetscInt row, PetscInt *nz, PetscInt **idx, PetscScalar **v)
2175: {
2176: PetscFunctionBegin;
2177: PetscFunctionReturn(PETSC_SUCCESS);
2178: }
2180: static PetscErrorCode MatNorm_SeqAIJ(Mat A, NormType type, PetscReal *nrm)
2181: {
2182: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
2183: const MatScalar *v;
2184: PetscReal sum = 0.0;
2185: PetscInt i, j;
2187: PetscFunctionBegin;
2188: PetscCall(MatSeqAIJGetArrayRead(A, &v));
2189: if (type == NORM_FROBENIUS) {
2190: #if PetscDefined(USE_REAL___FP16)
2191: PetscBLASInt one = 1, nz = a->nz;
2192: PetscCallBLAS("BLASnrm2", *nrm = BLASnrm2_(&nz, v, &one));
2193: #else
2194: for (i = 0; i < a->nz; i++) {
2195: sum += PetscRealPart(PetscConj(*v) * (*v));
2196: v++;
2197: }
2198: *nrm = PetscSqrtReal(sum);
2199: #endif
2200: PetscCall(PetscLogFlops(2.0 * a->nz));
2201: } else if (type == NORM_1) {
2202: PetscReal *tmp;
2203: PetscInt *jj = a->j;
2204: PetscCall(PetscCalloc1(A->cmap->n, &tmp));
2205: *nrm = 0.0;
2206: for (j = 0; j < a->nz; j++) {
2207: tmp[*jj++] += PetscAbsScalar(*v);
2208: v++;
2209: }
2210: for (j = 0; j < A->cmap->n; j++) {
2211: if (tmp[j] > *nrm) *nrm = tmp[j];
2212: }
2213: PetscCall(PetscFree(tmp));
2214: PetscCall(PetscLogFlops(a->nz));
2215: } else if (type == NORM_INFINITY) {
2216: *nrm = 0.0;
2217: for (j = 0; j < A->rmap->n; j++) {
2218: const PetscScalar *v2 = PetscSafePointerPlusOffset(v, a->i[j]);
2219: sum = 0.0;
2220: for (i = 0; i < a->i[j + 1] - a->i[j]; i++) {
2221: sum += PetscAbsScalar(*v2);
2222: v2++;
2223: }
2224: if (sum > *nrm) *nrm = sum;
2225: }
2226: PetscCall(PetscLogFlops(a->nz));
2227: } else SETERRQ(PETSC_COMM_SELF, PETSC_ERR_SUP, "No support for two norm");
2228: PetscCall(MatSeqAIJRestoreArrayRead(A, &v));
2229: PetscFunctionReturn(PETSC_SUCCESS);
2230: }
2232: static PetscErrorCode MatIsTranspose_SeqAIJ(Mat A, Mat B, PetscReal tol, PetscBool *f)
2233: {
2234: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)A->data, *bij = (Mat_SeqAIJ *)B->data;
2235: PetscInt *adx, *bdx, *aii, *bii, *aptr, *bptr;
2236: const MatScalar *va, *vb;
2237: PetscInt ma, na, mb, nb, i;
2239: PetscFunctionBegin;
2240: PetscCall(MatGetSize(A, &ma, &na));
2241: PetscCall(MatGetSize(B, &mb, &nb));
2242: if (ma != nb || na != mb) {
2243: *f = PETSC_FALSE;
2244: PetscFunctionReturn(PETSC_SUCCESS);
2245: }
2246: PetscCall(MatSeqAIJGetArrayRead(A, &va));
2247: PetscCall(MatSeqAIJGetArrayRead(B, &vb));
2248: aii = aij->i;
2249: bii = bij->i;
2250: adx = aij->j;
2251: bdx = bij->j;
2252: PetscCall(PetscMalloc1(ma, &aptr));
2253: PetscCall(PetscMalloc1(mb, &bptr));
2254: for (i = 0; i < ma; i++) aptr[i] = aii[i];
2255: for (i = 0; i < mb; i++) bptr[i] = bii[i];
2257: *f = PETSC_TRUE;
2258: for (i = 0; i < ma; i++) {
2259: while (aptr[i] < aii[i + 1]) {
2260: PetscInt idc, idr;
2261: PetscScalar vc, vr;
2262: /* column/row index/value */
2263: idc = adx[aptr[i]];
2264: idr = bdx[bptr[idc]];
2265: vc = va[aptr[i]];
2266: vr = vb[bptr[idc]];
2267: if (i != idr || PetscAbsScalar(vc - vr) > tol) {
2268: *f = PETSC_FALSE;
2269: goto done;
2270: } else {
2271: aptr[i]++;
2272: if (B || i != idc) bptr[idc]++;
2273: }
2274: }
2275: }
2276: done:
2277: PetscCall(PetscFree(aptr));
2278: PetscCall(PetscFree(bptr));
2279: PetscCall(MatSeqAIJRestoreArrayRead(A, &va));
2280: PetscCall(MatSeqAIJRestoreArrayRead(B, &vb));
2281: PetscFunctionReturn(PETSC_SUCCESS);
2282: }
2284: static PetscErrorCode MatIsHermitianTranspose_SeqAIJ(Mat A, Mat B, PetscReal tol, PetscBool *f)
2285: {
2286: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)A->data, *bij = (Mat_SeqAIJ *)B->data;
2287: PetscInt *adx, *bdx, *aii, *bii, *aptr, *bptr;
2288: MatScalar *va, *vb;
2289: PetscInt ma, na, mb, nb, i;
2291: PetscFunctionBegin;
2292: PetscCall(MatGetSize(A, &ma, &na));
2293: PetscCall(MatGetSize(B, &mb, &nb));
2294: if (ma != nb || na != mb) {
2295: *f = PETSC_FALSE;
2296: PetscFunctionReturn(PETSC_SUCCESS);
2297: }
2298: aii = aij->i;
2299: bii = bij->i;
2300: adx = aij->j;
2301: bdx = bij->j;
2302: va = aij->a;
2303: vb = bij->a;
2304: PetscCall(PetscMalloc1(ma, &aptr));
2305: PetscCall(PetscMalloc1(mb, &bptr));
2306: for (i = 0; i < ma; i++) aptr[i] = aii[i];
2307: for (i = 0; i < mb; i++) bptr[i] = bii[i];
2309: *f = PETSC_TRUE;
2310: for (i = 0; i < ma; i++) {
2311: while (aptr[i] < aii[i + 1]) {
2312: PetscInt idc, idr;
2313: PetscScalar vc, vr;
2314: /* column/row index/value */
2315: idc = adx[aptr[i]];
2316: idr = bdx[bptr[idc]];
2317: vc = va[aptr[i]];
2318: vr = vb[bptr[idc]];
2319: if (i != idr || PetscAbsScalar(vc - PetscConj(vr)) > tol) {
2320: *f = PETSC_FALSE;
2321: goto done;
2322: } else {
2323: aptr[i]++;
2324: if (B || i != idc) bptr[idc]++;
2325: }
2326: }
2327: }
2328: done:
2329: PetscCall(PetscFree(aptr));
2330: PetscCall(PetscFree(bptr));
2331: PetscFunctionReturn(PETSC_SUCCESS);
2332: }
2334: PetscErrorCode MatDiagonalScale_SeqAIJ(Mat A, Vec ll, Vec rr)
2335: {
2336: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
2337: const PetscScalar *l, *r;
2338: PetscScalar x;
2339: MatScalar *v;
2340: PetscInt i, j, m = A->rmap->n, n = A->cmap->n, M, nz = a->nz;
2341: const PetscInt *jj;
2343: PetscFunctionBegin;
2344: if (ll) {
2345: /* The local size is used so that VecMPI can be passed to this routine
2346: by MatDiagonalScale_MPIAIJ */
2347: PetscCall(VecGetLocalSize(ll, &m));
2348: PetscCheck(m == A->rmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Left scaling vector wrong length");
2349: PetscCall(VecGetArrayRead(ll, &l));
2350: PetscCall(MatSeqAIJGetArray(A, &v));
2351: for (i = 0; i < m; i++) {
2352: x = l[i];
2353: M = a->i[i + 1] - a->i[i];
2354: for (j = 0; j < M; j++) (*v++) *= x;
2355: }
2356: PetscCall(VecRestoreArrayRead(ll, &l));
2357: PetscCall(PetscLogFlops(nz));
2358: PetscCall(MatSeqAIJRestoreArray(A, &v));
2359: }
2360: if (rr) {
2361: PetscCall(VecGetLocalSize(rr, &n));
2362: PetscCheck(n == A->cmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Right scaling vector wrong length");
2363: PetscCall(VecGetArrayRead(rr, &r));
2364: PetscCall(MatSeqAIJGetArray(A, &v));
2365: jj = a->j;
2366: for (i = 0; i < nz; i++) (*v++) *= r[*jj++];
2367: PetscCall(MatSeqAIJRestoreArray(A, &v));
2368: PetscCall(VecRestoreArrayRead(rr, &r));
2369: PetscCall(PetscLogFlops(nz));
2370: }
2371: PetscFunctionReturn(PETSC_SUCCESS);
2372: }
2374: PetscErrorCode MatCreateSubMatrix_SeqAIJ(Mat A, IS isrow, IS iscol, PetscInt csize, MatReuse scall, Mat *B)
2375: {
2376: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data, *c;
2377: PetscInt *smap, i, k, kstart, kend, oldcols = A->cmap->n, *lens;
2378: PetscInt row, mat_i, *mat_j, tcol, first, step, *mat_ilen, sum, lensi;
2379: const PetscInt *irow, *icol;
2380: const PetscScalar *aa;
2381: PetscInt nrows, ncols;
2382: PetscInt *starts, *j_new, *i_new, *aj = a->j, *ai = a->i, ii, *ailen = a->ilen;
2383: MatScalar *a_new, *mat_a, *c_a;
2384: Mat C;
2385: PetscBool stride;
2387: PetscFunctionBegin;
2388: PetscCall(ISGetIndices(isrow, &irow));
2389: PetscCall(ISGetLocalSize(isrow, &nrows));
2390: PetscCall(ISGetLocalSize(iscol, &ncols));
2392: PetscCall(PetscObjectTypeCompare((PetscObject)iscol, ISSTRIDE, &stride));
2393: if (stride) {
2394: PetscCall(ISStrideGetInfo(iscol, &first, &step));
2395: } else {
2396: first = 0;
2397: step = 0;
2398: }
2399: if (stride && step == 1) {
2400: /* special case of contiguous rows */
2401: PetscCall(PetscMalloc2(nrows, &lens, nrows, &starts));
2402: /* loop over new rows determining lens and starting points */
2403: for (i = 0; i < nrows; i++) {
2404: kstart = ai[irow[i]];
2405: kend = kstart + ailen[irow[i]];
2406: starts[i] = kstart;
2407: for (k = kstart; k < kend; k++) {
2408: if (aj[k] >= first) {
2409: starts[i] = k;
2410: break;
2411: }
2412: }
2413: sum = 0;
2414: while (k < kend) {
2415: if (aj[k++] >= first + ncols) break;
2416: sum++;
2417: }
2418: lens[i] = sum;
2419: }
2420: /* create submatrix */
2421: if (scall == MAT_REUSE_MATRIX) {
2422: PetscInt n_cols, n_rows;
2423: PetscCall(MatGetSize(*B, &n_rows, &n_cols));
2424: PetscCheck(n_rows == nrows && n_cols == ncols, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Reused submatrix wrong size");
2425: PetscCall(MatZeroEntries(*B));
2426: C = *B;
2427: } else {
2428: PetscInt rbs, cbs;
2429: PetscCall(MatCreate(PetscObjectComm((PetscObject)A), &C));
2430: PetscCall(MatSetSizes(C, nrows, ncols, PETSC_DETERMINE, PETSC_DETERMINE));
2431: PetscCall(ISGetBlockSize(isrow, &rbs));
2432: PetscCall(ISGetBlockSize(iscol, &cbs));
2433: PetscCall(MatSetBlockSizes(C, rbs, cbs));
2434: PetscCall(MatSetType(C, ((PetscObject)A)->type_name));
2435: PetscCall(MatSeqAIJSetPreallocation_SeqAIJ(C, 0, lens));
2436: }
2437: c = (Mat_SeqAIJ *)C->data;
2439: /* loop over rows inserting into submatrix */
2440: PetscCall(MatSeqAIJGetArrayWrite(C, &a_new)); // Not 'a_new = c->a-new', since that raw usage ignores offload state of C
2441: j_new = c->j;
2442: i_new = c->i;
2443: PetscCall(MatSeqAIJGetArrayRead(A, &aa));
2444: for (i = 0; i < nrows; i++) {
2445: ii = starts[i];
2446: lensi = lens[i];
2447: if (lensi) {
2448: for (k = 0; k < lensi; k++) *j_new++ = aj[ii + k] - first;
2449: PetscCall(PetscArraycpy(a_new, aa + starts[i], lensi));
2450: a_new += lensi;
2451: }
2452: i_new[i + 1] = i_new[i] + lensi;
2453: c->ilen[i] = lensi;
2454: }
2455: PetscCall(MatSeqAIJRestoreArrayWrite(C, &a_new)); // Set C's offload state properly
2456: PetscCall(MatSeqAIJRestoreArrayRead(A, &aa));
2457: PetscCall(PetscFree2(lens, starts));
2458: } else {
2459: PetscCall(ISGetIndices(iscol, &icol));
2460: PetscCall(PetscCalloc1(oldcols, &smap));
2461: PetscCall(PetscMalloc1(1 + nrows, &lens));
2462: for (i = 0; i < ncols; i++) {
2463: 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);
2464: smap[icol[i]] = i + 1;
2465: }
2467: /* determine lens of each row */
2468: for (i = 0; i < nrows; i++) {
2469: kstart = ai[irow[i]];
2470: kend = kstart + a->ilen[irow[i]];
2471: lens[i] = 0;
2472: for (k = kstart; k < kend; k++) {
2473: if (smap[aj[k]]) lens[i]++;
2474: }
2475: }
2476: /* Create and fill new matrix */
2477: if (scall == MAT_REUSE_MATRIX) {
2478: PetscBool equal;
2480: c = (Mat_SeqAIJ *)((*B)->data);
2481: PetscCheck((*B)->rmap->n == nrows && (*B)->cmap->n == ncols, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Cannot reuse matrix. wrong size");
2482: PetscCall(PetscArraycmp(c->ilen, lens, (*B)->rmap->n, &equal));
2483: PetscCheck(equal, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Cannot reuse matrix. wrong number of nonzeros");
2484: PetscCall(PetscArrayzero(c->ilen, (*B)->rmap->n));
2485: C = *B;
2486: } else {
2487: PetscInt rbs, cbs;
2488: PetscCall(MatCreate(PetscObjectComm((PetscObject)A), &C));
2489: PetscCall(MatSetSizes(C, nrows, ncols, PETSC_DETERMINE, PETSC_DETERMINE));
2490: PetscCall(ISGetBlockSize(isrow, &rbs));
2491: PetscCall(ISGetBlockSize(iscol, &cbs));
2492: if (rbs > 1 || cbs > 1) PetscCall(MatSetBlockSizes(C, rbs, cbs));
2493: PetscCall(MatSetType(C, ((PetscObject)A)->type_name));
2494: PetscCall(MatSeqAIJSetPreallocation_SeqAIJ(C, 0, lens));
2495: }
2496: PetscCall(MatSeqAIJGetArrayRead(A, &aa));
2498: c = (Mat_SeqAIJ *)C->data;
2499: PetscCall(MatSeqAIJGetArrayWrite(C, &c_a)); // Not 'c->a', since that raw usage ignores offload state of C
2500: for (i = 0; i < nrows; i++) {
2501: row = irow[i];
2502: kstart = ai[row];
2503: kend = kstart + a->ilen[row];
2504: mat_i = c->i[i];
2505: mat_j = PetscSafePointerPlusOffset(c->j, mat_i);
2506: mat_a = PetscSafePointerPlusOffset(c_a, mat_i);
2507: mat_ilen = c->ilen + i;
2508: for (k = kstart; k < kend; k++) {
2509: if ((tcol = smap[a->j[k]])) {
2510: *mat_j++ = tcol - 1;
2511: *mat_a++ = aa[k];
2512: (*mat_ilen)++;
2513: }
2514: }
2515: }
2516: PetscCall(MatSeqAIJRestoreArrayRead(A, &aa));
2517: /* Free work space */
2518: PetscCall(ISRestoreIndices(iscol, &icol));
2519: PetscCall(PetscFree(smap));
2520: PetscCall(PetscFree(lens));
2521: /* sort */
2522: for (i = 0; i < nrows; i++) {
2523: PetscInt ilen;
2525: mat_i = c->i[i];
2526: mat_j = PetscSafePointerPlusOffset(c->j, mat_i);
2527: mat_a = PetscSafePointerPlusOffset(c_a, mat_i);
2528: ilen = c->ilen[i];
2529: PetscCall(PetscSortIntWithScalarArray(ilen, mat_j, mat_a));
2530: }
2531: PetscCall(MatSeqAIJRestoreArrayWrite(C, &c_a));
2532: }
2533: #if PetscDefined(HAVE_DEVICE)
2534: PetscCall(MatBindToCPU(C, A->boundtocpu));
2535: #endif
2536: PetscCall(MatAssemblyBegin(C, MAT_FINAL_ASSEMBLY));
2537: PetscCall(MatAssemblyEnd(C, MAT_FINAL_ASSEMBLY));
2539: PetscCall(ISRestoreIndices(isrow, &irow));
2540: *B = C;
2541: PetscFunctionReturn(PETSC_SUCCESS);
2542: }
2544: static PetscErrorCode MatGetMultiProcBlock_SeqAIJ(Mat mat, MPI_Comm subComm, MatReuse scall, Mat *subMat)
2545: {
2546: Mat B;
2548: PetscFunctionBegin;
2549: if (scall == MAT_INITIAL_MATRIX) {
2550: PetscCall(MatCreate(subComm, &B));
2551: PetscCall(MatSetSizes(B, mat->rmap->n, mat->cmap->n, mat->rmap->n, mat->cmap->n));
2552: PetscCall(MatSetBlockSizesFromMats(B, mat, mat));
2553: PetscCall(MatSetType(B, MATSEQAIJ));
2554: PetscCall(MatDuplicateNoCreate_SeqAIJ(B, mat, MAT_COPY_VALUES, PETSC_TRUE));
2555: *subMat = B;
2556: } else {
2557: PetscCall(MatCopy_SeqAIJ(mat, *subMat, SAME_NONZERO_PATTERN));
2558: }
2559: PetscFunctionReturn(PETSC_SUCCESS);
2560: }
2562: static PetscErrorCode MatILUFactor_SeqAIJ(Mat inA, IS row, IS col, const MatFactorInfo *info)
2563: {
2564: Mat_SeqAIJ *a = (Mat_SeqAIJ *)inA->data;
2565: Mat outA;
2566: PetscBool row_identity, col_identity;
2568: PetscFunctionBegin;
2569: PetscCheck(info->levels == 0, PETSC_COMM_SELF, PETSC_ERR_SUP, "Only levels=0 supported for in-place ilu");
2571: PetscCall(ISIdentity(row, &row_identity));
2572: PetscCall(ISIdentity(col, &col_identity));
2574: outA = inA;
2575: PetscCall(PetscFree(inA->solvertype));
2576: PetscCall(PetscStrallocpy(MATSOLVERPETSC, &inA->solvertype));
2578: PetscCall(PetscObjectReference((PetscObject)row));
2579: PetscCall(ISDestroy(&a->row));
2581: a->row = row;
2583: PetscCall(PetscObjectReference((PetscObject)col));
2584: PetscCall(ISDestroy(&a->col));
2586: a->col = col;
2588: /* Create the inverse permutation so that it can be used in MatLUFactorNumeric() */
2589: PetscCall(ISDestroy(&a->icol));
2590: PetscCall(ISInvertPermutation(col, PETSC_DECIDE, &a->icol));
2592: if (!a->solve_work) { /* this matrix may have been factored before */
2593: PetscCall(PetscMalloc1(inA->rmap->n, &a->solve_work));
2594: }
2596: if (row_identity && col_identity) {
2597: PetscCall(MatLUFactorNumeric_SeqAIJ_inplace(outA, inA, info));
2598: } else {
2599: PetscCall(MatLUFactorNumeric_SeqAIJ_InplaceWithPerm(outA, inA, info));
2600: }
2601: outA->factortype = MAT_FACTOR_LU;
2602: PetscFunctionReturn(PETSC_SUCCESS);
2603: }
2605: PetscErrorCode MatScale_SeqAIJ(Mat inA, PetscScalar alpha)
2606: {
2607: Mat_SeqAIJ *a = (Mat_SeqAIJ *)inA->data;
2608: PetscScalar *v;
2609: PetscBLASInt one = 1, bnz;
2611: PetscFunctionBegin;
2612: PetscCall(MatSeqAIJGetArray(inA, &v));
2613: PetscCall(PetscBLASIntCast(a->nz, &bnz));
2614: PetscCallBLAS("BLASscal", BLASscal_(&bnz, &alpha, v, &one));
2615: PetscCall(PetscLogFlops(a->nz));
2616: PetscCall(MatSeqAIJRestoreArray(inA, &v));
2617: PetscFunctionReturn(PETSC_SUCCESS);
2618: }
2620: PetscErrorCode MatDestroySubMatrix_Private(Mat_SubSppt *submatj)
2621: {
2622: PetscInt i;
2624: PetscFunctionBegin;
2625: if (!submatj->id) { /* delete data that are linked only to submats[id=0] */
2626: PetscCall(PetscFree4(submatj->sbuf1, submatj->ptr, submatj->tmp, submatj->ctr));
2628: for (i = 0; i < submatj->nrqr; ++i) PetscCall(PetscFree(submatj->sbuf2[i]));
2629: PetscCall(PetscFree3(submatj->sbuf2, submatj->req_size, submatj->req_source1));
2631: if (submatj->rbuf1) {
2632: PetscCall(PetscFree(submatj->rbuf1[0]));
2633: PetscCall(PetscFree(submatj->rbuf1));
2634: }
2636: for (i = 0; i < submatj->nrqs; ++i) PetscCall(PetscFree(submatj->rbuf3[i]));
2637: PetscCall(PetscFree3(submatj->req_source2, submatj->rbuf2, submatj->rbuf3));
2638: PetscCall(PetscFree(submatj->pa));
2639: }
2641: #if PetscDefined(USE_CTABLE)
2642: PetscCall(PetscHMapIDestroy(&submatj->rmap));
2643: PetscCall(PetscFree(submatj->cmap_loc));
2644: PetscCall(PetscFree(submatj->rmap_loc));
2645: #else
2646: PetscCall(PetscFree(submatj->rmap));
2647: #endif
2649: if (!submatj->allcolumns) {
2650: #if PetscDefined(USE_CTABLE)
2651: PetscCall(PetscHMapIDestroy(&submatj->cmap));
2652: #else
2653: PetscCall(PetscFree(submatj->cmap));
2654: #endif
2655: }
2656: PetscCall(PetscFree(submatj->row2proc));
2658: PetscCall(PetscFree(submatj));
2659: PetscFunctionReturn(PETSC_SUCCESS);
2660: }
2662: PetscErrorCode MatDestroySubMatrix_SeqAIJ(Mat C)
2663: {
2664: Mat_SeqAIJ *c = (Mat_SeqAIJ *)C->data;
2665: Mat_SubSppt *submatj = c->submatis1;
2667: PetscFunctionBegin;
2668: PetscCall((*submatj->destroy)(C));
2669: PetscCall(MatDestroySubMatrix_Private(submatj));
2670: PetscFunctionReturn(PETSC_SUCCESS);
2671: }
2673: /* Note this has code duplication with MatDestroySubMatrices_SeqBAIJ() */
2674: static PetscErrorCode MatDestroySubMatrices_SeqAIJ(PetscInt n, Mat *mat[])
2675: {
2676: PetscInt i;
2677: Mat C;
2678: Mat_SeqAIJ *c;
2679: Mat_SubSppt *submatj;
2681: PetscFunctionBegin;
2682: for (i = 0; i < n; i++) {
2683: C = (*mat)[i];
2684: c = (Mat_SeqAIJ *)C->data;
2685: submatj = c->submatis1;
2686: if (submatj) {
2687: if (--((PetscObject)C)->refct <= 0) {
2688: PetscCall(PetscFree(C->factorprefix));
2689: PetscCall((*submatj->destroy)(C));
2690: PetscCall(MatDestroySubMatrix_Private(submatj));
2691: PetscCall(PetscFree(C->defaultvectype));
2692: PetscCall(PetscFree(C->defaultrandtype));
2693: PetscCall(PetscLayoutDestroy(&C->rmap));
2694: PetscCall(PetscLayoutDestroy(&C->cmap));
2695: PetscCall(PetscHeaderDestroy(&C));
2696: }
2697: } else {
2698: PetscCall(MatDestroy(&C));
2699: }
2700: }
2702: /* Destroy Dummy submatrices created for reuse */
2703: PetscCall(MatDestroySubMatrices_Dummy(n, mat));
2705: PetscCall(PetscFree(*mat));
2706: PetscFunctionReturn(PETSC_SUCCESS);
2707: }
2709: static PetscErrorCode MatCreateSubMatrices_SeqAIJ(Mat A, PetscInt n, const IS irow[], const IS icol[], MatReuse scall, Mat *B[])
2710: {
2711: PetscInt i;
2713: PetscFunctionBegin;
2714: if (scall == MAT_INITIAL_MATRIX) PetscCall(PetscCalloc1(n + 1, B));
2716: for (i = 0; i < n; i++) PetscCall(MatCreateSubMatrix_SeqAIJ(A, irow[i], icol[i], PETSC_DECIDE, scall, &(*B)[i]));
2717: PetscFunctionReturn(PETSC_SUCCESS);
2718: }
2720: static PetscErrorCode MatIncreaseOverlap_SeqAIJ(Mat A, PetscInt is_max, IS is[], PetscInt ov)
2721: {
2722: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
2723: PetscInt row, i, j, k, l, ll, m, n, *nidx, isz, val;
2724: const PetscInt *idx;
2725: PetscInt start, end, *ai, *aj, bs = A->rmap->bs == A->cmap->bs ? A->rmap->bs : 1;
2726: PetscBT table;
2728: PetscFunctionBegin;
2729: m = A->rmap->n / bs;
2730: ai = a->i;
2731: aj = a->j;
2733: PetscCheck(ov >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "illegal negative overlap value used");
2735: PetscCall(PetscMalloc1(m + 1, &nidx));
2736: PetscCall(PetscBTCreate(m, &table));
2738: for (i = 0; i < is_max; i++) {
2739: /* Initialize the two local arrays */
2740: isz = 0;
2741: PetscCall(PetscBTMemzero(m, table));
2743: /* Extract the indices, assume there can be duplicate entries */
2744: PetscCall(ISGetIndices(is[i], &idx));
2745: PetscCall(ISGetLocalSize(is[i], &n));
2747: if (bs > 1) {
2748: /* Enter these into the temp arrays. I.e., mark table[row], enter row into new index */
2749: for (j = 0; j < n; ++j) {
2750: if (!PetscBTLookupSet(table, idx[j] / bs)) nidx[isz++] = idx[j] / bs;
2751: }
2752: PetscCall(ISRestoreIndices(is[i], &idx));
2753: PetscCall(ISDestroy(&is[i]));
2755: k = 0;
2756: for (j = 0; j < ov; j++) { /* for each overlap */
2757: n = isz;
2758: for (; k < n; k++) { /* do only those rows in nidx[k], which are not done yet */
2759: for (ll = 0; ll < bs; ll++) {
2760: row = bs * nidx[k] + ll;
2761: start = ai[row];
2762: end = ai[row + 1];
2763: for (l = start; l < end; l++) {
2764: val = aj[l] / bs;
2765: if (!PetscBTLookupSet(table, val)) nidx[isz++] = val;
2766: }
2767: }
2768: }
2769: }
2770: PetscCall(ISCreateBlock(PETSC_COMM_SELF, bs, isz, nidx, PETSC_COPY_VALUES, is + i));
2771: } else {
2772: /* Enter these into the temp arrays. I.e., mark table[row], enter row into new index */
2773: for (j = 0; j < n; ++j) {
2774: if (!PetscBTLookupSet(table, idx[j])) nidx[isz++] = idx[j];
2775: }
2776: PetscCall(ISRestoreIndices(is[i], &idx));
2777: PetscCall(ISDestroy(&is[i]));
2779: k = 0;
2780: for (j = 0; j < ov; j++) { /* for each overlap */
2781: n = isz;
2782: for (; k < n; k++) { /* do only those rows in nidx[k], which are not done yet */
2783: row = nidx[k];
2784: start = ai[row];
2785: end = ai[row + 1];
2786: for (l = start; l < end; l++) {
2787: val = aj[l];
2788: if (!PetscBTLookupSet(table, val)) nidx[isz++] = val;
2789: }
2790: }
2791: }
2792: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, isz, nidx, PETSC_COPY_VALUES, is + i));
2793: }
2794: }
2795: PetscCall(PetscBTDestroy(&table));
2796: PetscCall(PetscFree(nidx));
2797: PetscFunctionReturn(PETSC_SUCCESS);
2798: }
2800: static PetscErrorCode MatPermute_SeqAIJ(Mat A, IS rowp, IS colp, Mat *B)
2801: {
2802: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
2803: PetscInt i, nz = 0, m = A->rmap->n, n = A->cmap->n;
2804: const PetscInt *row, *col;
2805: PetscInt *cnew, j, *lens;
2806: IS icolp, irowp;
2807: PetscInt *cwork = NULL;
2808: PetscScalar *vwork = NULL;
2810: PetscFunctionBegin;
2811: PetscCall(ISInvertPermutation(rowp, PETSC_DECIDE, &irowp));
2812: PetscCall(ISGetIndices(irowp, &row));
2813: PetscCall(ISInvertPermutation(colp, PETSC_DECIDE, &icolp));
2814: PetscCall(ISGetIndices(icolp, &col));
2816: /* determine lengths of permuted rows */
2817: PetscCall(PetscMalloc1(m + 1, &lens));
2818: for (i = 0; i < m; i++) lens[row[i]] = a->i[i + 1] - a->i[i];
2819: PetscCall(MatCreate(PetscObjectComm((PetscObject)A), B));
2820: PetscCall(MatSetSizes(*B, m, n, m, n));
2821: PetscCall(MatSetBlockSizesFromMats(*B, A, A));
2822: PetscCall(MatSetType(*B, ((PetscObject)A)->type_name));
2823: PetscCall(MatSeqAIJSetPreallocation_SeqAIJ(*B, 0, lens));
2824: PetscCall(PetscFree(lens));
2826: PetscCall(PetscMalloc1(n, &cnew));
2827: for (i = 0; i < m; i++) {
2828: PetscCall(MatGetRow_SeqAIJ(A, i, &nz, &cwork, &vwork));
2829: for (j = 0; j < nz; j++) cnew[j] = col[cwork[j]];
2830: PetscCall(MatSetValues_SeqAIJ(*B, 1, &row[i], nz, cnew, vwork, INSERT_VALUES));
2831: PetscCall(MatRestoreRow_SeqAIJ(A, i, &nz, &cwork, &vwork));
2832: }
2833: PetscCall(PetscFree(cnew));
2835: (*B)->assembled = PETSC_FALSE;
2837: #if PetscDefined(HAVE_DEVICE)
2838: PetscCall(MatBindToCPU(*B, A->boundtocpu));
2839: #endif
2840: PetscCall(MatAssemblyBegin(*B, MAT_FINAL_ASSEMBLY));
2841: PetscCall(MatAssemblyEnd(*B, MAT_FINAL_ASSEMBLY));
2842: PetscCall(ISRestoreIndices(irowp, &row));
2843: PetscCall(ISRestoreIndices(icolp, &col));
2844: PetscCall(ISDestroy(&irowp));
2845: PetscCall(ISDestroy(&icolp));
2846: if (rowp == colp) PetscCall(MatPropagateSymmetryOptions(A, *B));
2847: PetscFunctionReturn(PETSC_SUCCESS);
2848: }
2850: PetscErrorCode MatCopy_SeqAIJ(Mat A, Mat B, MatStructure str)
2851: {
2852: PetscFunctionBegin;
2853: /* If the two matrices have the same copy implementation, use fast copy. */
2854: if (str == SAME_NONZERO_PATTERN && (A->ops->copy == B->ops->copy)) {
2855: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
2856: Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data;
2857: const PetscScalar *aa;
2858: PetscScalar *bb;
2860: PetscCall(MatSeqAIJGetArrayRead(A, &aa));
2861: PetscCall(MatSeqAIJGetArrayWrite(B, &bb));
2863: 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]);
2864: PetscCall(PetscArraycpy(bb, aa, a->i[A->rmap->n]));
2865: PetscCall(PetscObjectStateIncrease((PetscObject)B));
2866: PetscCall(MatSeqAIJRestoreArrayRead(A, &aa));
2867: PetscCall(MatSeqAIJRestoreArrayWrite(B, &bb));
2868: } else {
2869: PetscCall(MatCopy_Basic(A, B, str));
2870: }
2871: PetscFunctionReturn(PETSC_SUCCESS);
2872: }
2874: PETSC_INTERN PetscErrorCode MatSeqAIJGetArray_SeqAIJ(Mat A, PetscScalar *array[])
2875: {
2876: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
2878: PetscFunctionBegin;
2879: *array = a->a;
2880: PetscFunctionReturn(PETSC_SUCCESS);
2881: }
2883: PETSC_INTERN PetscErrorCode MatSeqAIJRestoreArray_SeqAIJ(Mat A, PetscScalar *array[])
2884: {
2885: PetscFunctionBegin;
2886: *array = NULL;
2887: PetscFunctionReturn(PETSC_SUCCESS);
2888: }
2890: /*
2891: Computes the number of nonzeros per row needed for preallocation when X and Y
2892: have different nonzero structure.
2893: */
2894: PetscErrorCode MatAXPYGetPreallocation_SeqX_private(PetscInt m, const PetscInt *xi, const PetscInt *xj, const PetscInt *yi, const PetscInt *yj, PetscInt *nnz)
2895: {
2896: PetscInt i, j, k, nzx, nzy;
2898: PetscFunctionBegin;
2899: /* Set the number of nonzeros in the new matrix */
2900: for (i = 0; i < m; i++) {
2901: const PetscInt *xjj = PetscSafePointerPlusOffset(xj, xi[i]), *yjj = PetscSafePointerPlusOffset(yj, yi[i]);
2902: nzx = xi[i + 1] - xi[i];
2903: nzy = yi[i + 1] - yi[i];
2904: nnz[i] = 0;
2905: for (j = 0, k = 0; j < nzx; j++) { /* Point in X */
2906: for (; k < nzy && yjj[k] < xjj[j]; k++) nnz[i]++; /* Catch up to X */
2907: if (k < nzy && yjj[k] == xjj[j]) k++; /* Skip duplicate */
2908: nnz[i]++;
2909: }
2910: for (; k < nzy; k++) nnz[i]++;
2911: }
2912: PetscFunctionReturn(PETSC_SUCCESS);
2913: }
2915: PetscErrorCode MatAXPYGetPreallocation_SeqAIJ(Mat Y, Mat X, PetscInt *nnz)
2916: {
2917: PetscInt m = Y->rmap->N;
2918: Mat_SeqAIJ *x = (Mat_SeqAIJ *)X->data;
2919: Mat_SeqAIJ *y = (Mat_SeqAIJ *)Y->data;
2921: PetscFunctionBegin;
2922: /* Set the number of nonzeros in the new matrix */
2923: PetscCall(MatAXPYGetPreallocation_SeqX_private(m, x->i, x->j, y->i, y->j, nnz));
2924: PetscFunctionReturn(PETSC_SUCCESS);
2925: }
2927: PetscErrorCode MatAXPY_SeqAIJ(Mat Y, PetscScalar a, Mat X, MatStructure str)
2928: {
2929: Mat_SeqAIJ *x = (Mat_SeqAIJ *)X->data, *y = (Mat_SeqAIJ *)Y->data;
2931: PetscFunctionBegin;
2932: if (str == UNKNOWN_NONZERO_PATTERN || (PetscDefined(USE_DEBUG) && str == SAME_NONZERO_PATTERN)) {
2933: PetscBool e = x->nz == y->nz ? PETSC_TRUE : PETSC_FALSE;
2934: if (e) {
2935: PetscCall(PetscArraycmp(x->i, y->i, Y->rmap->n + 1, &e));
2936: if (e) {
2937: PetscCall(PetscArraycmp(x->j, y->j, y->nz, &e));
2938: if (e) str = SAME_NONZERO_PATTERN;
2939: }
2940: }
2941: if (!e) PetscCheck(str != SAME_NONZERO_PATTERN, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "MatStructure is not SAME_NONZERO_PATTERN");
2942: }
2943: if (str == SAME_NONZERO_PATTERN) {
2944: const PetscScalar *xa;
2945: PetscScalar *ya, alpha = a;
2946: PetscBLASInt one = 1, bnz;
2948: PetscCall(PetscBLASIntCast(x->nz, &bnz));
2949: PetscCall(MatSeqAIJGetArray(Y, &ya));
2950: PetscCall(MatSeqAIJGetArrayRead(X, &xa));
2951: PetscCallBLAS("BLASaxpy", BLASaxpy_(&bnz, &alpha, xa, &one, ya, &one));
2952: PetscCall(MatSeqAIJRestoreArrayRead(X, &xa));
2953: PetscCall(MatSeqAIJRestoreArray(Y, &ya));
2954: PetscCall(PetscLogFlops(2.0 * bnz));
2955: PetscCall(PetscObjectStateIncrease((PetscObject)Y));
2956: } else if (str == SUBSET_NONZERO_PATTERN) { /* nonzeros of X is a subset of Y's */
2957: PetscCall(MatAXPY_Basic(Y, a, X, str));
2958: } else {
2959: Mat B;
2960: PetscInt *nnz;
2961: PetscCall(PetscMalloc1(Y->rmap->N, &nnz));
2962: PetscCall(MatCreate(PetscObjectComm((PetscObject)Y), &B));
2963: PetscCall(PetscObjectSetName((PetscObject)B, ((PetscObject)Y)->name));
2964: PetscCall(MatSetLayouts(B, Y->rmap, Y->cmap));
2965: PetscCall(MatSetType(B, ((PetscObject)Y)->type_name));
2966: PetscCall(MatAXPYGetPreallocation_SeqAIJ(Y, X, nnz));
2967: PetscCall(MatSeqAIJSetPreallocation(B, 0, nnz));
2968: PetscCall(MatAXPY_BasicWithPreallocation(B, Y, a, X, str));
2969: PetscCall(MatHeaderMerge(Y, &B));
2970: PetscCall(MatSeqAIJCheckInode(Y));
2971: PetscCall(PetscFree(nnz));
2972: }
2973: PetscFunctionReturn(PETSC_SUCCESS);
2974: }
2976: PETSC_INTERN PetscErrorCode MatConjugate_SeqAIJ(Mat mat)
2977: {
2978: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)mat->data;
2979: PetscInt i, nz = aij->nz;
2980: PetscScalar *a;
2982: PetscFunctionBegin;
2983: PetscCall(MatSeqAIJGetArray(mat, &a));
2984: for (i = 0; i < nz; i++) a[i] = PetscConj(a[i]);
2985: PetscCall(MatSeqAIJRestoreArray(mat, &a));
2986: PetscFunctionReturn(PETSC_SUCCESS);
2987: }
2989: static PetscErrorCode MatGetRowMaxAbs_SeqAIJ(Mat A, Vec v, PetscInt idx[])
2990: {
2991: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
2992: PetscInt i, j, m = A->rmap->n, *ai, *aj, ncols, n;
2993: PetscReal atmp;
2994: PetscScalar *x;
2995: const MatScalar *aa, *av;
2997: PetscFunctionBegin;
2998: PetscCheck(!A->factortype, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Not for factored matrix");
2999: PetscCall(MatSeqAIJGetArrayRead(A, &av));
3000: aa = av;
3001: ai = a->i;
3002: aj = a->j;
3004: PetscCall(VecGetArrayWrite(v, &x));
3005: PetscCall(VecGetLocalSize(v, &n));
3006: PetscCheck(n == A->rmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Nonconforming matrix and vector");
3007: for (i = 0; i < m; i++) {
3008: ncols = ai[1] - ai[0];
3009: ai++;
3010: x[i] = 0;
3011: for (j = 0; j < ncols; j++) {
3012: atmp = PetscAbsScalar(*aa);
3013: if (PetscAbsScalar(x[i]) < atmp) {
3014: x[i] = atmp;
3015: if (idx) idx[i] = *aj;
3016: }
3017: aa++;
3018: aj++;
3019: }
3020: }
3021: PetscCall(VecRestoreArrayWrite(v, &x));
3022: PetscCall(MatSeqAIJRestoreArrayRead(A, &av));
3023: PetscFunctionReturn(PETSC_SUCCESS);
3024: }
3026: static PetscErrorCode MatGetRowSumAbs_SeqAIJ(Mat A, Vec v)
3027: {
3028: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
3029: PetscInt i, j, m = A->rmap->n, *ai, ncols, n;
3030: PetscScalar *x;
3031: const MatScalar *aa, *av;
3033: PetscFunctionBegin;
3034: PetscCheck(!A->factortype, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Not for factored matrix");
3035: PetscCall(MatSeqAIJGetArrayRead(A, &av));
3036: aa = av;
3037: ai = a->i;
3039: PetscCall(VecGetArrayWrite(v, &x));
3040: PetscCall(VecGetLocalSize(v, &n));
3041: PetscCheck(n == A->rmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Nonconforming matrix and vector");
3042: for (i = 0; i < m; i++) {
3043: ncols = ai[1] - ai[0];
3044: ai++;
3045: x[i] = 0;
3046: for (j = 0; j < ncols; j++) {
3047: x[i] += PetscAbsScalar(*aa);
3048: aa++;
3049: }
3050: }
3051: PetscCall(VecRestoreArrayWrite(v, &x));
3052: PetscCall(MatSeqAIJRestoreArrayRead(A, &av));
3053: PetscFunctionReturn(PETSC_SUCCESS);
3054: }
3056: static PetscErrorCode MatGetRowMax_SeqAIJ(Mat A, Vec v, PetscInt idx[])
3057: {
3058: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
3059: PetscInt i, j, m = A->rmap->n, *ai, *aj, ncols, n;
3060: PetscScalar *x;
3061: const MatScalar *aa, *av;
3063: PetscFunctionBegin;
3064: PetscCheck(!A->factortype, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Not for factored matrix");
3065: PetscCall(MatSeqAIJGetArrayRead(A, &av));
3066: aa = av;
3067: ai = a->i;
3068: aj = a->j;
3070: PetscCall(VecGetArrayWrite(v, &x));
3071: PetscCall(VecGetLocalSize(v, &n));
3072: PetscCheck(n == A->rmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Nonconforming matrix and vector");
3073: for (i = 0; i < m; i++) {
3074: ncols = ai[1] - ai[0];
3075: ai++;
3076: if (ncols == A->cmap->n) { /* row is dense */
3077: x[i] = *aa;
3078: if (idx) idx[i] = 0;
3079: } else { /* row is sparse so already KNOW maximum is 0.0 or higher */
3080: x[i] = 0.0;
3081: if (idx) {
3082: for (j = 0; j < ncols; j++) { /* find first implicit 0.0 in the row */
3083: if (aj[j] > j) {
3084: idx[i] = j;
3085: break;
3086: }
3087: }
3088: /* in case first implicit 0.0 in the row occurs at ncols-th column */
3089: if (j == ncols && j < A->cmap->n) idx[i] = j;
3090: }
3091: }
3092: for (j = 0; j < ncols; j++) {
3093: if (PetscRealPart(x[i]) < PetscRealPart(*aa)) {
3094: x[i] = *aa;
3095: if (idx) idx[i] = *aj;
3096: }
3097: aa++;
3098: aj++;
3099: }
3100: }
3101: PetscCall(VecRestoreArrayWrite(v, &x));
3102: PetscCall(MatSeqAIJRestoreArrayRead(A, &av));
3103: PetscFunctionReturn(PETSC_SUCCESS);
3104: }
3106: static PetscErrorCode MatGetRowMinAbs_SeqAIJ(Mat A, Vec v, PetscInt idx[])
3107: {
3108: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
3109: PetscInt i, j, m = A->rmap->n, *ai, *aj, ncols, n;
3110: PetscScalar *x;
3111: const MatScalar *aa, *av;
3113: PetscFunctionBegin;
3114: PetscCall(MatSeqAIJGetArrayRead(A, &av));
3115: aa = av;
3116: ai = a->i;
3117: aj = a->j;
3119: PetscCall(VecGetArrayWrite(v, &x));
3120: PetscCall(VecGetLocalSize(v, &n));
3121: PetscCheck(n == m, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Nonconforming matrix and vector, %" PetscInt_FMT " vs. %" PetscInt_FMT " rows", m, n);
3122: for (i = 0; i < m; i++) {
3123: ncols = ai[1] - ai[0];
3124: ai++;
3125: if (ncols == A->cmap->n) { /* row is dense */
3126: x[i] = *aa;
3127: if (idx) idx[i] = 0;
3128: } else { /* row is sparse so already KNOW minimum is 0.0 or higher */
3129: x[i] = 0.0;
3130: if (idx) { /* find first implicit 0.0 in the row */
3131: for (j = 0; j < ncols; j++) {
3132: if (aj[j] > j) {
3133: idx[i] = j;
3134: break;
3135: }
3136: }
3137: /* in case first implicit 0.0 in the row occurs at ncols-th column */
3138: if (j == ncols && j < A->cmap->n) idx[i] = j;
3139: }
3140: }
3141: for (j = 0; j < ncols; j++) {
3142: if (PetscAbsScalar(x[i]) > PetscAbsScalar(*aa)) {
3143: x[i] = *aa;
3144: if (idx) idx[i] = *aj;
3145: }
3146: aa++;
3147: aj++;
3148: }
3149: }
3150: PetscCall(VecRestoreArrayWrite(v, &x));
3151: PetscCall(MatSeqAIJRestoreArrayRead(A, &av));
3152: PetscFunctionReturn(PETSC_SUCCESS);
3153: }
3155: static PetscErrorCode MatGetRowMin_SeqAIJ(Mat A, Vec v, PetscInt idx[])
3156: {
3157: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
3158: PetscInt i, j, m = A->rmap->n, ncols, n;
3159: const PetscInt *ai, *aj;
3160: PetscScalar *x;
3161: const MatScalar *aa, *av;
3163: PetscFunctionBegin;
3164: PetscCheck(!A->factortype, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Not for factored matrix");
3165: PetscCall(MatSeqAIJGetArrayRead(A, &av));
3166: aa = av;
3167: ai = a->i;
3168: aj = a->j;
3170: PetscCall(VecGetArrayWrite(v, &x));
3171: PetscCall(VecGetLocalSize(v, &n));
3172: PetscCheck(n == m, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Nonconforming matrix and vector");
3173: for (i = 0; i < m; i++) {
3174: ncols = ai[1] - ai[0];
3175: ai++;
3176: if (ncols == A->cmap->n) { /* row is dense */
3177: x[i] = *aa;
3178: if (idx) idx[i] = 0;
3179: } else { /* row is sparse so already KNOW minimum is 0.0 or lower */
3180: x[i] = 0.0;
3181: if (idx) { /* find first implicit 0.0 in the row */
3182: for (j = 0; j < ncols; j++) {
3183: if (aj[j] > j) {
3184: idx[i] = j;
3185: break;
3186: }
3187: }
3188: /* in case first implicit 0.0 in the row occurs at ncols-th column */
3189: if (j == ncols && j < A->cmap->n) idx[i] = j;
3190: }
3191: }
3192: for (j = 0; j < ncols; j++) {
3193: if (PetscRealPart(x[i]) > PetscRealPart(*aa)) {
3194: x[i] = *aa;
3195: if (idx) idx[i] = *aj;
3196: }
3197: aa++;
3198: aj++;
3199: }
3200: }
3201: PetscCall(VecRestoreArrayWrite(v, &x));
3202: PetscCall(MatSeqAIJRestoreArrayRead(A, &av));
3203: PetscFunctionReturn(PETSC_SUCCESS);
3204: }
3206: static PetscErrorCode MatInvertBlockDiagonal_SeqAIJ(Mat A, const PetscScalar **values)
3207: {
3208: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
3209: PetscInt i, bs = A->rmap->bs, mbs = A->rmap->n / bs, ipvt[5], bs2 = bs * bs, *v_pivots, ij[7], *IJ, j;
3210: MatScalar *diag, work[25], *v_work;
3211: const PetscReal shift = 0.0;
3212: PetscBool allowzeropivot, zeropivotdetected = PETSC_FALSE;
3214: PetscFunctionBegin;
3215: allowzeropivot = PetscNot(A->erroriffailure);
3216: if (a->ibdiag && a->ibdiagsize == bs2 * mbs && a->ibdiagState == ((PetscObject)A)->state) {
3217: if (values) *values = a->ibdiag;
3218: PetscFunctionReturn(PETSC_SUCCESS);
3219: }
3220: /* reallocate only when the length changes, so that the pointer stays valid for callers that
3221: hold on to it, such as PCSetUp_PBJacobi_Host() */
3222: if (!a->ibdiag || a->ibdiagsize != bs2 * mbs) {
3223: PetscCall(PetscFree(a->ibdiag));
3224: PetscCall(PetscMalloc1(bs2 * mbs, &a->ibdiag));
3225: a->ibdiagsize = bs2 * mbs;
3226: }
3227: diag = a->ibdiag;
3228: if (values) *values = a->ibdiag;
3229: /* factor and invert each block */
3230: switch (bs) {
3231: case 1:
3232: for (i = 0; i < mbs; i++) {
3233: PetscCall(MatGetValues(A, 1, &i, 1, &i, diag + i));
3234: if (PetscAbsScalar(diag[i] + shift) < PETSC_MACHINE_EPSILON) {
3235: PetscCheck(allowzeropivot, 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);
3236: A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
3237: A->factorerror_zeropivot_value = PetscAbsScalar(diag[i]);
3238: A->factorerror_zeropivot_row = i;
3239: PetscCall(PetscInfo(A, "Zero pivot, row %" PetscInt_FMT " pivot %g tolerance %g\n", i, (double)PetscAbsScalar(diag[i]), (double)PETSC_MACHINE_EPSILON));
3240: }
3241: diag[i] = (PetscScalar)1.0 / (diag[i] + shift);
3242: }
3243: break;
3244: case 2:
3245: for (i = 0; i < mbs; i++) {
3246: ij[0] = 2 * i;
3247: ij[1] = 2 * i + 1;
3248: PetscCall(MatGetValues(A, 2, ij, 2, ij, diag));
3249: PetscCall(PetscKernel_A_gets_inverse_A_2(diag, shift, allowzeropivot, &zeropivotdetected));
3250: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
3251: PetscCall(PetscKernel_A_gets_transpose_A_2(diag));
3252: diag += 4;
3253: }
3254: break;
3255: case 3:
3256: for (i = 0; i < mbs; i++) {
3257: ij[0] = 3 * i;
3258: ij[1] = 3 * i + 1;
3259: ij[2] = 3 * i + 2;
3260: PetscCall(MatGetValues(A, 3, ij, 3, ij, diag));
3261: PetscCall(PetscKernel_A_gets_inverse_A_3(diag, shift, allowzeropivot, &zeropivotdetected));
3262: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
3263: PetscCall(PetscKernel_A_gets_transpose_A_3(diag));
3264: diag += 9;
3265: }
3266: break;
3267: case 4:
3268: for (i = 0; i < mbs; i++) {
3269: ij[0] = 4 * i;
3270: ij[1] = 4 * i + 1;
3271: ij[2] = 4 * i + 2;
3272: ij[3] = 4 * i + 3;
3273: PetscCall(MatGetValues(A, 4, ij, 4, ij, diag));
3274: PetscCall(PetscKernel_A_gets_inverse_A_4(diag, shift, allowzeropivot, &zeropivotdetected));
3275: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
3276: PetscCall(PetscKernel_A_gets_transpose_A_4(diag));
3277: diag += 16;
3278: }
3279: break;
3280: case 5:
3281: for (i = 0; i < mbs; i++) {
3282: ij[0] = 5 * i;
3283: ij[1] = 5 * i + 1;
3284: ij[2] = 5 * i + 2;
3285: ij[3] = 5 * i + 3;
3286: ij[4] = 5 * i + 4;
3287: PetscCall(MatGetValues(A, 5, ij, 5, ij, diag));
3288: PetscCall(PetscKernel_A_gets_inverse_A_5(diag, ipvt, work, shift, allowzeropivot, &zeropivotdetected));
3289: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
3290: PetscCall(PetscKernel_A_gets_transpose_A_5(diag));
3291: diag += 25;
3292: }
3293: break;
3294: case 6:
3295: for (i = 0; i < mbs; i++) {
3296: ij[0] = 6 * i;
3297: ij[1] = 6 * i + 1;
3298: ij[2] = 6 * i + 2;
3299: ij[3] = 6 * i + 3;
3300: ij[4] = 6 * i + 4;
3301: ij[5] = 6 * i + 5;
3302: PetscCall(MatGetValues(A, 6, ij, 6, ij, diag));
3303: PetscCall(PetscKernel_A_gets_inverse_A_6(diag, shift, allowzeropivot, &zeropivotdetected));
3304: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
3305: PetscCall(PetscKernel_A_gets_transpose_A_6(diag));
3306: diag += 36;
3307: }
3308: break;
3309: case 7:
3310: for (i = 0; i < mbs; i++) {
3311: ij[0] = 7 * i;
3312: ij[1] = 7 * i + 1;
3313: ij[2] = 7 * i + 2;
3314: ij[3] = 7 * i + 3;
3315: ij[4] = 7 * i + 4;
3316: ij[5] = 7 * i + 5;
3317: ij[6] = 7 * i + 6;
3318: PetscCall(MatGetValues(A, 7, ij, 7, ij, diag));
3319: PetscCall(PetscKernel_A_gets_inverse_A_7(diag, shift, allowzeropivot, &zeropivotdetected));
3320: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
3321: PetscCall(PetscKernel_A_gets_transpose_A_7(diag));
3322: diag += 49;
3323: }
3324: break;
3325: default:
3326: PetscCall(PetscMalloc3(bs, &v_work, bs, &v_pivots, bs, &IJ));
3327: for (i = 0; i < mbs; i++) {
3328: for (j = 0; j < bs; j++) IJ[j] = bs * i + j;
3329: PetscCall(MatGetValues(A, bs, IJ, bs, IJ, diag));
3330: PetscCall(PetscKernel_A_gets_inverse_A(bs, diag, v_pivots, v_work, allowzeropivot, &zeropivotdetected));
3331: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
3332: PetscCall(PetscKernel_A_gets_transpose_A_N(diag, bs));
3333: diag += bs2;
3334: }
3335: PetscCall(PetscFree3(v_work, v_pivots, IJ));
3336: }
3337: a->ibdiagState = ((PetscObject)A)->state;
3338: PetscFunctionReturn(PETSC_SUCCESS);
3339: }
3341: static PetscErrorCode MatSetRandom_SeqAIJ(Mat x, PetscRandom rctx)
3342: {
3343: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)x->data;
3344: PetscScalar a, *aa;
3345: PetscInt m, n, i, j, col;
3347: PetscFunctionBegin;
3348: if (!x->assembled) {
3349: PetscCall(MatGetSize(x, &m, &n));
3350: for (i = 0; i < m; i++) {
3351: for (j = 0; j < aij->imax[i]; j++) {
3352: PetscCall(PetscRandomGetValue(rctx, &a));
3353: col = (PetscInt)(n * PetscRealPart(a));
3354: PetscCall(MatSetValues(x, 1, &i, 1, &col, &a, ADD_VALUES));
3355: }
3356: }
3357: } else {
3358: PetscCall(MatSeqAIJGetArrayWrite(x, &aa));
3359: for (i = 0; i < aij->nz; i++) PetscCall(PetscRandomGetValue(rctx, aa + i));
3360: PetscCall(MatSeqAIJRestoreArrayWrite(x, &aa));
3361: }
3362: PetscCall(MatAssemblyBegin(x, MAT_FINAL_ASSEMBLY));
3363: PetscCall(MatAssemblyEnd(x, MAT_FINAL_ASSEMBLY));
3364: PetscFunctionReturn(PETSC_SUCCESS);
3365: }
3367: /* Like MatSetRandom_SeqAIJ, but do not set values on columns in range of [low, high) */
3368: PetscErrorCode MatSetRandomSkipColumnRange_SeqAIJ_Private(Mat x, PetscInt low, PetscInt high, PetscRandom rctx)
3369: {
3370: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)x->data;
3371: PetscScalar a;
3372: PetscInt m, n, i, j, col, nskip;
3374: PetscFunctionBegin;
3375: nskip = high - low;
3376: PetscCall(MatGetSize(x, &m, &n));
3377: n -= nskip; /* shrink number of columns where nonzeros can be set */
3378: for (i = 0; i < m; i++) {
3379: for (j = 0; j < aij->imax[i]; j++) {
3380: PetscCall(PetscRandomGetValue(rctx, &a));
3381: col = (PetscInt)(n * PetscRealPart(a));
3382: if (col >= low) col += nskip; /* shift col rightward to skip the hole */
3383: PetscCall(MatSetValues(x, 1, &i, 1, &col, &a, ADD_VALUES));
3384: }
3385: }
3386: PetscCall(MatAssemblyBegin(x, MAT_FINAL_ASSEMBLY));
3387: PetscCall(MatAssemblyEnd(x, MAT_FINAL_ASSEMBLY));
3388: PetscFunctionReturn(PETSC_SUCCESS);
3389: }
3391: static struct _MatOps MatOps_Values = {MatSetValues_SeqAIJ,
3392: MatGetRow_SeqAIJ,
3393: MatRestoreRow_SeqAIJ,
3394: MatMult_SeqAIJ,
3395: /* 4*/ MatMultAdd_SeqAIJ,
3396: MatMultTranspose_SeqAIJ,
3397: MatMultTransposeAdd_SeqAIJ,
3398: NULL,
3399: NULL,
3400: NULL,
3401: /* 10*/ NULL,
3402: MatLUFactor_SeqAIJ,
3403: NULL,
3404: MatSOR_SeqAIJ,
3405: MatTranspose_SeqAIJ,
3406: /* 15*/ MatGetInfo_SeqAIJ,
3407: MatEqual_SeqAIJ,
3408: MatGetDiagonal_SeqAIJ,
3409: MatDiagonalScale_SeqAIJ,
3410: MatNorm_SeqAIJ,
3411: /* 20*/ NULL,
3412: MatAssemblyEnd_SeqAIJ,
3413: MatSetOption_SeqAIJ,
3414: MatZeroEntries_SeqAIJ,
3415: /* 24*/ MatZeroRows_SeqAIJ,
3416: NULL,
3417: NULL,
3418: NULL,
3419: NULL,
3420: /* 29*/ MatSetUp_Seq_Hash,
3421: NULL,
3422: NULL,
3423: NULL,
3424: NULL,
3425: /* 34*/ MatDuplicate_SeqAIJ,
3426: NULL,
3427: NULL,
3428: MatILUFactor_SeqAIJ,
3429: NULL,
3430: /* 39*/ MatAXPY_SeqAIJ,
3431: MatCreateSubMatrices_SeqAIJ,
3432: MatIncreaseOverlap_SeqAIJ,
3433: MatGetValues_SeqAIJ,
3434: MatCopy_SeqAIJ,
3435: /* 44*/ MatGetRowMax_SeqAIJ,
3436: MatScale_SeqAIJ,
3437: MatShift_SeqAIJ,
3438: MatDiagonalSet_SeqAIJ,
3439: MatZeroRowsColumns_SeqAIJ,
3440: /* 49*/ MatSetRandom_SeqAIJ,
3441: MatGetRowIJ_SeqAIJ,
3442: MatRestoreRowIJ_SeqAIJ,
3443: MatGetColumnIJ_SeqAIJ,
3444: MatRestoreColumnIJ_SeqAIJ,
3445: /* 54*/ MatFDColoringCreate_SeqXAIJ,
3446: NULL,
3447: NULL,
3448: MatPermute_SeqAIJ,
3449: NULL,
3450: /* 59*/ NULL,
3451: MatDestroy_SeqAIJ,
3452: MatView_SeqAIJ,
3453: NULL,
3454: NULL,
3455: /* 64*/ MatMatMatMultNumeric_SeqAIJ_SeqAIJ_SeqAIJ,
3456: NULL,
3457: NULL,
3458: NULL,
3459: MatGetRowMaxAbs_SeqAIJ,
3460: /* 69*/ MatGetRowMinAbs_SeqAIJ,
3461: NULL,
3462: NULL,
3463: MatFDColoringApply_AIJ,
3464: NULL,
3465: /* 74*/ MatFindZeroDiagonals_SeqAIJ,
3466: NULL,
3467: NULL,
3468: NULL,
3469: MatLoad_SeqAIJ,
3470: /* 79*/ NULL,
3471: NULL,
3472: NULL,
3473: NULL,
3474: NULL,
3475: /* 84*/ NULL,
3476: MatMatMultNumeric_SeqAIJ_SeqAIJ,
3477: MatPtAPNumeric_SeqAIJ_SeqAIJ_SparseAxpy,
3478: NULL,
3479: MatMatTransposeMultNumeric_SeqAIJ_SeqAIJ,
3480: /* 90*/ NULL,
3481: MatProductSetFromOptions_SeqAIJ,
3482: NULL,
3483: NULL,
3484: MatConjugate_SeqAIJ,
3485: /* 94*/ NULL,
3486: MatSetValuesRow_SeqAIJ,
3487: MatRealPart_SeqAIJ,
3488: MatImaginaryPart_SeqAIJ,
3489: NULL,
3490: /* 99*/ NULL,
3491: MatMatSolve_SeqAIJ,
3492: NULL,
3493: MatGetRowMin_SeqAIJ,
3494: NULL,
3495: /*104*/ NULL,
3496: NULL,
3497: NULL,
3498: NULL,
3499: NULL,
3500: /*109*/ NULL,
3501: NULL,
3502: NULL,
3503: NULL,
3504: MatGetMultiProcBlock_SeqAIJ,
3505: /*114*/ MatFindNonzeroRows_SeqAIJ,
3506: MatGetColumnReductions_SeqAIJ,
3507: MatInvertBlockDiagonal_SeqAIJ,
3508: MatInvertVariableBlockDiagonal_SeqAIJ,
3509: NULL,
3510: /*119*/ NULL,
3511: MatTransposeMatMultNumeric_SeqAIJ_SeqAIJ,
3512: MatTransposeColoringCreate_SeqAIJ,
3513: MatTransColoringApplySpToDen_SeqAIJ,
3514: MatTransColoringApplyDenToSp_SeqAIJ,
3515: /*124*/ MatRARtNumeric_SeqAIJ_SeqAIJ,
3516: NULL,
3517: NULL,
3518: MatFDColoringSetUp_SeqXAIJ,
3519: MatFindOffBlockDiagonalEntries_SeqAIJ,
3520: /*129*/ MatCreateMPIMatConcatenateSeqMat_SeqAIJ,
3521: MatDestroySubMatrices_SeqAIJ,
3522: NULL,
3523: NULL,
3524: MatCreateGraph_Simple_AIJ,
3525: /*134*/ MatTransposeSymbolic_SeqAIJ,
3526: MatEliminateZeros_SeqAIJ,
3527: MatGetRowSumAbs_SeqAIJ,
3528: NULL,
3529: NULL,
3530: /*139*/ NULL,
3531: MatCopyHashToXAIJ_Seq_Hash,
3532: NULL,
3533: NULL,
3534: NULL,
3535: /*144*/ NULL,
3536: NULL,
3537: NULL,
3538: NULL};
3540: static PetscErrorCode MatSeqAIJSetColumnIndices_SeqAIJ(Mat mat, PetscInt *indices)
3541: {
3542: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)mat->data;
3543: PetscInt i, nz, n;
3545: PetscFunctionBegin;
3546: nz = aij->maxnz;
3547: n = mat->rmap->n;
3548: for (i = 0; i < nz; i++) aij->j[i] = indices[i];
3549: aij->nz = nz;
3550: for (i = 0; i < n; i++) aij->ilen[i] = aij->imax[i];
3551: PetscFunctionReturn(PETSC_SUCCESS);
3552: }
3554: /*
3555: * Given a sparse matrix with global column indices, compact it by using a local column space.
3556: * The result matrix helps saving memory in other algorithms, such as MatPtAPSymbolic_MPIAIJ_MPIAIJ_scalable()
3557: */
3558: PetscErrorCode MatSeqAIJCompactOutExtraColumns_SeqAIJ(Mat mat, ISLocalToGlobalMapping *mapping)
3559: {
3560: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)mat->data;
3561: PetscHMapI gid1_lid1;
3562: PetscHashIter tpos;
3563: PetscInt gid, lid, i, ec, nz = aij->nz;
3564: PetscInt *garray, *jj = aij->j;
3566: PetscFunctionBegin;
3568: PetscAssertPointer(mapping, 2);
3569: /* use a table */
3570: PetscCall(PetscHMapICreateWithSize(mat->rmap->n, &gid1_lid1));
3571: ec = 0;
3572: for (i = 0; i < nz; i++) {
3573: PetscInt data, gid1 = jj[i] + 1;
3574: PetscCall(PetscHMapIGetWithDefault(gid1_lid1, gid1, 0, &data));
3575: if (!data) {
3576: /* one based table */
3577: PetscCall(PetscHMapISet(gid1_lid1, gid1, ++ec));
3578: }
3579: }
3580: /* form array of columns we need */
3581: PetscCall(PetscMalloc1(ec, &garray));
3582: PetscHashIterBegin(gid1_lid1, tpos);
3583: while (!PetscHashIterAtEnd(gid1_lid1, tpos)) {
3584: PetscHashIterGetKey(gid1_lid1, tpos, gid);
3585: PetscHashIterGetVal(gid1_lid1, tpos, lid);
3586: PetscHashIterNext(gid1_lid1, tpos);
3587: gid--;
3588: lid--;
3589: garray[lid] = gid;
3590: }
3591: PetscCall(PetscSortInt(ec, garray)); /* sort, and rebuild */
3592: PetscCall(PetscHMapIClear(gid1_lid1));
3593: for (i = 0; i < ec; i++) PetscCall(PetscHMapISet(gid1_lid1, garray[i] + 1, i + 1));
3594: /* compact out the extra columns in B */
3595: for (i = 0; i < nz; i++) {
3596: PetscInt gid1 = jj[i] + 1;
3597: PetscCall(PetscHMapIGetWithDefault(gid1_lid1, gid1, 0, &lid));
3598: lid--;
3599: jj[i] = lid;
3600: }
3601: PetscCall(PetscLayoutDestroy(&mat->cmap));
3602: PetscCall(PetscHMapIDestroy(&gid1_lid1));
3603: PetscCall(PetscLayoutCreateFromSizes(PetscObjectComm((PetscObject)mat), ec, ec, 1, &mat->cmap));
3604: PetscCall(ISLocalToGlobalMappingCreate(PETSC_COMM_SELF, mat->cmap->bs, mat->cmap->n, garray, PETSC_OWN_POINTER, mapping));
3605: PetscCall(ISLocalToGlobalMappingSetType(*mapping, ISLOCALTOGLOBALMAPPINGHASH));
3606: PetscFunctionReturn(PETSC_SUCCESS);
3607: }
3609: /*@
3610: MatSeqAIJSetColumnIndices - Set the column indices for all the rows
3611: in the matrix.
3613: Input Parameters:
3614: + mat - the `MATSEQAIJ` matrix
3615: - indices - the column indices
3617: Level: advanced
3619: Notes:
3620: This can be called if you have precomputed the nonzero structure of the
3621: matrix and want to provide it to the matrix object to improve the performance
3622: of the `MatSetValues()` operation.
3624: You MUST have set the correct numbers of nonzeros per row in the call to
3625: `MatCreateSeqAIJ()`, and the columns indices MUST be sorted.
3627: MUST be called before any calls to `MatSetValues()`
3629: The indices should start with zero, not one.
3631: .seealso: [](ch_matrices), `Mat`, `MATSEQAIJ`
3632: @*/
3633: PetscErrorCode MatSeqAIJSetColumnIndices(Mat mat, PetscInt *indices)
3634: {
3635: PetscFunctionBegin;
3637: PetscAssertPointer(indices, 2);
3638: PetscUseMethod(mat, "MatSeqAIJSetColumnIndices_C", (Mat, PetscInt *), (mat, indices));
3639: PetscFunctionReturn(PETSC_SUCCESS);
3640: }
3642: static PetscErrorCode MatStoreValues_SeqAIJ(Mat mat)
3643: {
3644: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)mat->data;
3645: size_t nz = aij->i[mat->rmap->n];
3647: PetscFunctionBegin;
3648: PetscCheck(aij->nonew, PETSC_COMM_SELF, PETSC_ERR_ORDER, "Must call MatSetOption(A,MAT_NEW_NONZERO_LOCATIONS,PETSC_FALSE);first");
3650: /* allocate space for values if not already there */
3651: if (!aij->saved_values) PetscCall(PetscMalloc1(nz + 1, &aij->saved_values));
3653: /* copy values over */
3654: PetscCall(PetscArraycpy(aij->saved_values, aij->a, nz));
3655: PetscFunctionReturn(PETSC_SUCCESS);
3656: }
3658: /*@
3659: MatStoreValues - Stashes a copy of the matrix values; this allows reusing of the linear part of a Jacobian, while recomputing only the
3660: nonlinear portion.
3662: Logically Collect
3664: Input Parameter:
3665: . mat - the matrix (currently only `MATAIJ` matrices support this option)
3667: Level: advanced
3669: Example Usage:
3670: .vb
3671: Using SNES
3672: Create Jacobian matrix
3673: Set linear terms into matrix
3674: Apply boundary conditions to matrix, at this time matrix must have
3675: final nonzero structure (i.e. setting the nonlinear terms and applying
3676: boundary conditions again will not change the nonzero structure
3677: MatSetOption(mat, MAT_NEW_NONZERO_LOCATIONS, PETSC_FALSE);
3678: MatStoreValues(mat);
3679: Call SNESSetJacobian() with matrix
3680: In your Jacobian routine
3681: MatRetrieveValues(mat);
3682: Set nonlinear terms in matrix
3684: Without `SNESSolve()`, i.e. when you handle nonlinear solve yourself:
3685: // build linear portion of Jacobian
3686: MatSetOption(mat, MAT_NEW_NONZERO_LOCATIONS, PETSC_FALSE);
3687: MatStoreValues(mat);
3688: loop over nonlinear iterations
3689: MatRetrieveValues(mat);
3690: // call MatSetValues(mat,...) to set nonliner portion of Jacobian
3691: // call MatAssemblyBegin/End() on matrix
3692: Solve linear system with Jacobian
3693: endloop
3694: .ve
3696: Notes:
3697: Matrix must already be assembled before calling this routine
3698: Must set the matrix option `MatSetOption`(mat,`MAT_NEW_NONZERO_LOCATIONS`,`PETSC_FALSE`); before
3699: calling this routine.
3701: When this is called multiple times it overwrites the previous set of stored values
3702: and does not allocated additional space.
3704: .seealso: [](ch_matrices), `Mat`, `MatRetrieveValues()`
3705: @*/
3706: PetscErrorCode MatStoreValues(Mat mat)
3707: {
3708: PetscFunctionBegin;
3710: PetscCheck(mat->assembled, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Not for unassembled matrix");
3711: PetscCheck(!mat->factortype, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Not for factored matrix");
3712: PetscUseMethod(mat, "MatStoreValues_C", (Mat), (mat));
3713: PetscFunctionReturn(PETSC_SUCCESS);
3714: }
3716: static PetscErrorCode MatRetrieveValues_SeqAIJ(Mat mat)
3717: {
3718: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)mat->data;
3719: PetscInt nz = aij->i[mat->rmap->n];
3721: PetscFunctionBegin;
3722: PetscCheck(aij->nonew, PETSC_COMM_SELF, PETSC_ERR_ORDER, "Must call MatSetOption(A,MAT_NEW_NONZERO_LOCATIONS,PETSC_FALSE);first");
3723: PetscCheck(aij->saved_values, PETSC_COMM_SELF, PETSC_ERR_ORDER, "Must call MatStoreValues(A);first");
3724: /* copy values over */
3725: PetscCall(PetscArraycpy(aij->a, aij->saved_values, nz));
3726: PetscFunctionReturn(PETSC_SUCCESS);
3727: }
3729: /*@
3730: MatRetrieveValues - Retrieves the copy of the matrix values that was stored with `MatStoreValues()`
3732: Logically Collect
3734: Input Parameter:
3735: . mat - the matrix (currently only `MATAIJ` matrices support this option)
3737: Level: advanced
3739: .seealso: [](ch_matrices), `Mat`, `MatStoreValues()`
3740: @*/
3741: PetscErrorCode MatRetrieveValues(Mat mat)
3742: {
3743: PetscFunctionBegin;
3745: PetscCheck(mat->assembled, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Not for unassembled matrix");
3746: PetscCheck(!mat->factortype, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Not for factored matrix");
3747: PetscUseMethod(mat, "MatRetrieveValues_C", (Mat), (mat));
3748: PetscCall(PetscObjectStateIncrease((PetscObject)mat));
3749: PetscFunctionReturn(PETSC_SUCCESS);
3750: }
3752: /*@
3753: MatCreateSeqAIJ - Creates a sparse matrix in `MATSEQAIJ` (compressed row) format
3754: (the default parallel PETSc format). For good matrix assembly performance
3755: the user should preallocate the matrix storage by setting the parameter `nz`
3756: (or the array `nnz`).
3758: Collective
3760: Input Parameters:
3761: + comm - MPI communicator, set to `PETSC_COMM_SELF`
3762: . m - number of rows
3763: . n - number of columns
3764: . nz - number of nonzeros per row (same for all rows)
3765: - nnz - array containing the number of nonzeros in the various rows
3766: (possibly different for each row) or NULL
3768: Output Parameter:
3769: . A - the matrix
3771: Options Database Keys:
3772: + -mat_no_inode - Do not use inodes
3773: - -mat_inode_limit limit - Sets inode limit (max limit=5)
3775: Level: intermediate
3777: Notes:
3778: It is recommend to use `MatCreateFromOptions()` instead of this routine
3780: If `nnz` is given then `nz` is ignored
3782: The `MATSEQAIJ` format, also called
3783: compressed row storage, is fully compatible with standard Fortran
3784: storage. That is, the stored row and column indices can begin at
3785: either one (as in Fortran) or zero.
3787: Specify the preallocated storage with either `nz` or `nnz` (not both).
3788: Set `nz` = `PETSC_DEFAULT` and `nnz` = `NULL` for PETSc to control dynamic memory
3789: allocation.
3791: By default, this format uses inodes (identical nodes) when possible, to
3792: improve numerical efficiency of matrix-vector products and solves. We
3793: search for consecutive rows with the same nonzero structure, thereby
3794: reusing matrix information to achieve increased efficiency.
3796: .seealso: [](ch_matrices), `Mat`, [Sparse Matrix Creation](sec_matsparse), `MatCreate()`, `MatCreateAIJ()`, `MatSetValues()`, `MatSeqAIJSetColumnIndices()`, `MatCreateSeqAIJWithArrays()`
3797: @*/
3798: PetscErrorCode MatCreateSeqAIJ(MPI_Comm comm, PetscInt m, PetscInt n, PetscInt nz, const PetscInt nnz[], Mat *A)
3799: {
3800: PetscFunctionBegin;
3801: PetscCall(MatCreate(comm, A));
3802: PetscCall(MatSetSizes(*A, m, n, m, n));
3803: PetscCall(MatSetType(*A, MATSEQAIJ));
3804: PetscCall(MatSeqAIJSetPreallocation_SeqAIJ(*A, nz, nnz));
3805: PetscFunctionReturn(PETSC_SUCCESS);
3806: }
3808: /*@
3809: MatSeqAIJSetPreallocation - For good matrix assembly performance
3810: the user should preallocate the matrix storage by setting the parameter nz
3811: (or the array nnz). By setting these parameters accurately, performance
3812: during matrix assembly can be increased by more than a factor of 50.
3814: Collective
3816: Input Parameters:
3817: + B - The matrix
3818: . nz - number of nonzeros per row (same for all rows)
3819: - nnz - array containing the number of nonzeros in the various rows
3820: (possibly different for each row) or NULL
3822: Options Database Keys:
3823: + -mat_no_inode - Do not use inodes
3824: - -mat_inode_limit limit - Sets inode limit (max limit=5)
3826: Level: intermediate
3828: Notes:
3829: If `nnz` is given then `nz` is ignored
3831: The `MATSEQAIJ` format also called
3832: compressed row storage, is fully compatible with standard Fortran
3833: storage. That is, the stored row and column indices can begin at
3834: either one (as in Fortran) or zero. See the users' manual for details.
3836: Specify the preallocated storage with either `nz` or `nnz` (not both).
3837: Set nz = `PETSC_DEFAULT` and `nnz` = `NULL` for PETSc to control dynamic memory
3838: allocation.
3840: You can call `MatGetInfo()` to get information on how effective the preallocation was;
3841: for example the fields mallocs,nz_allocated,nz_used,nz_unneeded;
3842: You can also run with the option -info and look for messages with the string
3843: malloc in them to see if additional memory allocation was needed.
3845: Developer Notes:
3846: Use nz of `MAT_SKIP_ALLOCATION` to not allocate any space for the matrix
3847: entries or columns indices
3849: By default, this format uses inodes (identical nodes) when possible, to
3850: improve numerical efficiency of matrix-vector products and solves. We
3851: search for consecutive rows with the same nonzero structure, thereby
3852: reusing matrix information to achieve increased efficiency.
3854: .seealso: [](ch_matrices), `Mat`, `MatCreate()`, `MatCreateAIJ()`, `MatSetValues()`, `MatSeqAIJSetColumnIndices()`, `MatCreateSeqAIJWithArrays()`, `MatGetInfo()`,
3855: `MatSeqAIJSetTotalPreallocation()`
3856: @*/
3857: PetscErrorCode MatSeqAIJSetPreallocation(Mat B, PetscInt nz, const PetscInt nnz[])
3858: {
3859: PetscFunctionBegin;
3862: PetscTryMethod(B, "MatSeqAIJSetPreallocation_C", (Mat, PetscInt, const PetscInt[]), (B, nz, nnz));
3863: PetscFunctionReturn(PETSC_SUCCESS);
3864: }
3866: PetscErrorCode MatSeqAIJSetPreallocation_SeqAIJ(Mat B, PetscInt nz, const PetscInt *nnz)
3867: {
3868: Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data;
3869: PetscBool skipallocation = PETSC_FALSE, realalloc = PETSC_FALSE;
3870: PetscInt i;
3872: PetscFunctionBegin;
3873: if (B->hash_active) {
3874: B->ops[0] = b->cops;
3875: PetscCall(PetscHMapIJVDestroy(&b->ht));
3876: PetscCall(PetscFree(b->dnz));
3877: B->hash_active = PETSC_FALSE;
3878: }
3879: if (nz >= 0 || nnz) realalloc = PETSC_TRUE;
3880: if (nz == MAT_SKIP_ALLOCATION) {
3881: skipallocation = PETSC_TRUE;
3882: nz = 0;
3883: }
3884: PetscCall(PetscLayoutSetUp(B->rmap));
3885: PetscCall(PetscLayoutSetUp(B->cmap));
3887: if (nz == PETSC_DEFAULT || nz == PETSC_DECIDE) nz = 5;
3888: PetscCheck(nz >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "nz cannot be less than 0: value %" PetscInt_FMT, nz);
3889: if (nnz) {
3890: for (i = 0; i < B->rmap->n; i++) {
3891: 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]);
3892: 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);
3893: }
3894: }
3896: B->preallocated = PETSC_TRUE;
3897: if (!skipallocation) {
3898: if (!b->imax) PetscCall(PetscMalloc1(B->rmap->n, &b->imax));
3899: if (!b->ilen) {
3900: /* b->ilen will count nonzeros in each row so far. */
3901: PetscCall(PetscCalloc1(B->rmap->n, &b->ilen));
3902: } else {
3903: PetscCall(PetscMemzero(b->ilen, B->rmap->n * sizeof(PetscInt)));
3904: }
3905: if (!b->ipre) PetscCall(PetscMalloc1(B->rmap->n, &b->ipre));
3906: if (!nnz) {
3907: if (nz == PETSC_DEFAULT || nz == PETSC_DECIDE) nz = 10;
3908: else if (nz < 0) nz = 1;
3909: nz = PetscMin(nz, B->cmap->n);
3910: for (i = 0; i < B->rmap->n; i++) b->imax[i] = nz;
3911: PetscCall(PetscIntMultError(nz, B->rmap->n, &nz));
3912: } else {
3913: PetscInt64 nz64 = 0;
3914: for (i = 0; i < B->rmap->n; i++) {
3915: b->imax[i] = nnz[i];
3916: nz64 += nnz[i];
3917: }
3918: PetscCall(PetscIntCast(nz64, &nz));
3919: }
3921: /* allocate the matrix space */
3922: PetscCall(MatSeqXAIJFreeAIJ(B, &b->a, &b->j, &b->i));
3923: PetscCall(PetscShmgetAllocateArray(nz, sizeof(PetscInt), (void **)&b->j));
3924: PetscCall(PetscShmgetAllocateArray(B->rmap->n + 1, sizeof(PetscInt), (void **)&b->i));
3925: b->free_ij = PETSC_TRUE;
3926: if (B->structure_only) {
3927: b->free_a = PETSC_FALSE;
3928: } else {
3929: PetscCall(PetscShmgetAllocateArray(nz, sizeof(PetscScalar), (void **)&b->a));
3930: b->free_a = PETSC_TRUE;
3931: }
3932: b->i[0] = 0;
3933: for (i = 1; i < B->rmap->n + 1; i++) b->i[i] = b->i[i - 1] + b->imax[i - 1];
3934: } else {
3935: b->free_a = PETSC_FALSE;
3936: b->free_ij = PETSC_FALSE;
3937: }
3939: if (b->ipre && nnz != b->ipre && b->imax) {
3940: /* reserve user-requested sparsity */
3941: PetscCall(PetscArraycpy(b->ipre, b->imax, B->rmap->n));
3942: }
3944: b->nz = 0;
3945: b->maxnz = nz;
3946: B->info.nz_unneeded = (double)b->maxnz;
3947: if (realalloc) PetscCall(MatSetOption(B, MAT_NEW_NONZERO_ALLOCATION_ERR, PETSC_TRUE));
3948: B->was_assembled = PETSC_FALSE;
3949: B->assembled = PETSC_FALSE;
3950: /* We simply deem preallocation has changed nonzero state. Updating the state
3951: will give clients (like AIJKokkos) a chance to know something has happened.
3952: */
3953: B->nonzerostate++;
3954: PetscFunctionReturn(PETSC_SUCCESS);
3955: }
3957: PetscErrorCode MatResetPreallocation_SeqAIJ_Private(Mat A, PetscBool *memoryreset)
3958: {
3959: Mat_SeqAIJ *a;
3960: PetscInt i;
3961: PetscBool skipreset;
3963: PetscFunctionBegin;
3966: PetscCheck(A->insertmode == NOT_SET_VALUES, PETSC_COMM_SELF, PETSC_ERR_SUP, "Cannot reset preallocation after setting some values but not yet calling MatAssemblyBegin()/MatAssemblyEnd()");
3967: if (A->num_ass == 0) PetscFunctionReturn(PETSC_SUCCESS);
3969: /* Check local size. If zero, then return */
3970: if (!A->rmap->n) PetscFunctionReturn(PETSC_SUCCESS);
3972: a = (Mat_SeqAIJ *)A->data;
3973: /* if no saved info, we error out */
3974: PetscCheck(a->ipre, PETSC_COMM_SELF, PETSC_ERR_ARG_NULL, "No saved preallocation info ");
3976: PetscCheck(a->i && a->imax && a->ilen, PETSC_COMM_SELF, PETSC_ERR_ARG_NULL, "Memory info is incomplete, and cannot reset preallocation ");
3978: PetscCall(PetscArraycmp(a->ipre, a->ilen, A->rmap->n, &skipreset));
3979: if (skipreset) PetscCall(MatZeroEntries(A));
3980: else {
3981: PetscCall(PetscArraycpy(a->imax, a->ipre, A->rmap->n));
3982: PetscCall(PetscArrayzero(a->ilen, A->rmap->n));
3983: a->i[0] = 0;
3984: for (i = 1; i < A->rmap->n + 1; i++) a->i[i] = a->i[i - 1] + a->imax[i - 1];
3985: A->preallocated = PETSC_TRUE;
3986: a->nz = 0;
3987: a->maxnz = a->i[A->rmap->n];
3988: A->info.nz_unneeded = (double)a->maxnz;
3989: A->was_assembled = PETSC_FALSE;
3990: A->assembled = PETSC_FALSE;
3991: A->nonzerostate++;
3992: /* Log that the state of this object has changed; this will help guarantee that preconditioners get re-setup */
3993: PetscCall(PetscObjectStateIncrease((PetscObject)A));
3994: }
3995: if (memoryreset) *memoryreset = (PetscBool)!skipreset;
3996: PetscFunctionReturn(PETSC_SUCCESS);
3997: }
3999: static PetscErrorCode MatResetPreallocation_SeqAIJ(Mat A)
4000: {
4001: PetscFunctionBegin;
4002: PetscCall(MatResetPreallocation_SeqAIJ_Private(A, NULL));
4003: PetscFunctionReturn(PETSC_SUCCESS);
4004: }
4006: /*@
4007: MatSeqAIJSetPreallocationCSR - Allocates memory for a sparse sequential matrix in `MATSEQAIJ` format.
4009: Input Parameters:
4010: + B - the matrix
4011: . i - the indices into `j` for the start of each row (indices start with zero)
4012: . j - the column indices for each row (indices start with zero) these must be sorted for each row
4013: - v - optional values in the matrix, use `NULL` if not provided
4015: Level: developer
4017: Notes:
4018: The `i`,`j`,`v` values are COPIED with this routine; to avoid the copy use `MatCreateSeqAIJWithArrays()`
4020: This routine may be called multiple times with different nonzero patterns (or the same nonzero pattern). The nonzero
4021: structure will be the union of all the previous nonzero structures.
4023: Developer Notes:
4024: 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
4025: then just copies the `v` values directly with `PetscMemcpy()`.
4027: This routine could also take a `PetscCopyMode` argument to allow sharing the values instead of always copying them.
4029: .seealso: [](ch_matrices), `Mat`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatSeqAIJSetPreallocation()`, `MATSEQAIJ`, `MatResetPreallocation()`
4030: @*/
4031: PetscErrorCode MatSeqAIJSetPreallocationCSR(Mat B, const PetscInt i[], const PetscInt j[], const PetscScalar v[])
4032: {
4033: PetscFunctionBegin;
4036: PetscTryMethod(B, "MatSeqAIJSetPreallocationCSR_C", (Mat, const PetscInt[], const PetscInt[], const PetscScalar[]), (B, i, j, v));
4037: PetscFunctionReturn(PETSC_SUCCESS);
4038: }
4040: static PetscErrorCode MatSeqAIJSetPreallocationCSR_SeqAIJ(Mat B, const PetscInt Ii[], const PetscInt J[], const PetscScalar v[])
4041: {
4042: PetscInt i;
4043: PetscInt m, n;
4044: PetscInt nz;
4045: PetscInt *nnz;
4047: PetscFunctionBegin;
4048: PetscCheck(Ii[0] == 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Ii[0] must be 0 it is %" PetscInt_FMT, Ii[0]);
4050: PetscCall(PetscLayoutSetUp(B->rmap));
4051: PetscCall(PetscLayoutSetUp(B->cmap));
4053: PetscCall(MatGetSize(B, &m, &n));
4054: PetscCall(PetscMalloc1(m + 1, &nnz));
4055: for (i = 0; i < m; i++) {
4056: nz = Ii[i + 1] - Ii[i];
4057: PetscCheck(nz >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Local row %" PetscInt_FMT " has a negative number of columns %" PetscInt_FMT, i, nz);
4058: nnz[i] = nz;
4059: }
4060: PetscCall(MatSeqAIJSetPreallocation(B, 0, nnz));
4061: PetscCall(PetscFree(nnz));
4063: 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));
4065: PetscCall(MatAssemblyBegin(B, MAT_FINAL_ASSEMBLY));
4066: PetscCall(MatAssemblyEnd(B, MAT_FINAL_ASSEMBLY));
4068: PetscCall(MatSetOption(B, MAT_NEW_NONZERO_LOCATION_ERR, PETSC_TRUE));
4069: PetscFunctionReturn(PETSC_SUCCESS);
4070: }
4072: /*@
4073: MatSeqAIJKron - Computes `C`, the Kronecker product of `A` and `B`.
4075: Input Parameters:
4076: + A - left-hand side matrix
4077: . B - right-hand side matrix
4078: - reuse - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX`
4080: Output Parameter:
4081: . C - Kronecker product of `A` and `B`
4083: Level: intermediate
4085: Note:
4086: `MAT_REUSE_MATRIX` can only be used when the nonzero structure of the product matrix has not changed from that last call to `MatSeqAIJKron()`.
4088: .seealso: [](ch_matrices), `Mat`, `MatCreateSeqAIJ()`, `MATSEQAIJ`, `MATKAIJ`, `MatReuse`
4089: @*/
4090: PetscErrorCode MatSeqAIJKron(Mat A, Mat B, MatReuse reuse, Mat *C)
4091: {
4092: PetscFunctionBegin;
4097: PetscAssertPointer(C, 4);
4098: if (reuse == MAT_REUSE_MATRIX) {
4101: }
4102: PetscTryMethod(A, "MatSeqAIJKron_C", (Mat, Mat, MatReuse, Mat *), (A, B, reuse, C));
4103: PetscFunctionReturn(PETSC_SUCCESS);
4104: }
4106: static PetscErrorCode MatSeqAIJKron_SeqAIJ(Mat A, Mat B, MatReuse reuse, Mat *C)
4107: {
4108: Mat newmat;
4109: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
4110: Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data;
4111: PetscScalar *v;
4112: const PetscScalar *aa, *ba;
4113: 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;
4114: PetscBool flg;
4116: PetscFunctionBegin;
4117: PetscCheck(!A->factortype, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Not for factored matrix");
4118: PetscCheck(A->assembled, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Not for unassembled matrix");
4119: PetscCheck(!B->factortype, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Not for factored matrix");
4120: PetscCheck(B->assembled, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Not for unassembled matrix");
4121: PetscCall(PetscObjectTypeCompare((PetscObject)B, MATSEQAIJ, &flg));
4122: PetscCheck(flg, PETSC_COMM_SELF, PETSC_ERR_SUP, "MatType %s", ((PetscObject)B)->type_name);
4123: PetscCheck(reuse == MAT_INITIAL_MATRIX || reuse == MAT_REUSE_MATRIX, PETSC_COMM_SELF, PETSC_ERR_SUP, "MatReuse %d", (int)reuse);
4124: if (reuse == MAT_INITIAL_MATRIX) {
4125: PetscCall(PetscMalloc2(am * bm + 1, &i, a->i[am] * b->i[bm], &j));
4126: PetscCall(MatCreate(PETSC_COMM_SELF, &newmat));
4127: PetscCall(MatSetSizes(newmat, am * bm, an * bn, am * bm, an * bn));
4128: PetscCall(MatSetType(newmat, MATAIJ));
4129: i[0] = 0;
4130: for (m = 0; m < am; ++m) {
4131: for (p = 0; p < bm; ++p) {
4132: i[m * bm + p + 1] = i[m * bm + p] + (a->i[m + 1] - a->i[m]) * (b->i[p + 1] - b->i[p]);
4133: for (n = a->i[m]; n < a->i[m + 1]; ++n) {
4134: for (q = b->i[p]; q < b->i[p + 1]; ++q) j[nnz++] = a->j[n] * bn + b->j[q];
4135: }
4136: }
4137: }
4138: PetscCall(MatSeqAIJSetPreallocationCSR(newmat, i, j, NULL));
4139: *C = newmat;
4140: PetscCall(PetscFree2(i, j));
4141: nnz = 0;
4142: }
4143: PetscCall(MatSeqAIJGetArray(*C, &v));
4144: PetscCall(MatSeqAIJGetArrayRead(A, &aa));
4145: PetscCall(MatSeqAIJGetArrayRead(B, &ba));
4146: for (m = 0; m < am; ++m) {
4147: for (p = 0; p < bm; ++p) {
4148: for (n = a->i[m]; n < a->i[m + 1]; ++n) {
4149: for (q = b->i[p]; q < b->i[p + 1]; ++q) v[nnz++] = aa[n] * ba[q];
4150: }
4151: }
4152: }
4153: PetscCall(MatSeqAIJRestoreArray(*C, &v));
4154: PetscCall(MatSeqAIJRestoreArrayRead(A, &aa));
4155: PetscCall(MatSeqAIJRestoreArrayRead(B, &ba));
4156: PetscFunctionReturn(PETSC_SUCCESS);
4157: }
4159: #include <../src/mat/impls/dense/seq/dense.h>
4160: #include <petsc/private/kernels/petscaxpy.h>
4162: /*
4163: Computes (B'*A')' since computing B*A directly is untenable
4165: n p p
4166: [ ] [ ] [ ]
4167: m [ A ] * n [ B ] = m [ C ]
4168: [ ] [ ] [ ]
4170: */
4171: PetscErrorCode MatMatMultNumeric_SeqDense_SeqAIJ(Mat A, Mat B, Mat C)
4172: {
4173: Mat_SeqDense *sub_a = (Mat_SeqDense *)A->data;
4174: Mat_SeqAIJ *sub_b = (Mat_SeqAIJ *)B->data;
4175: Mat_SeqDense *sub_c = (Mat_SeqDense *)C->data;
4176: PetscInt i, j, n, m, q, p;
4177: const PetscInt *ii, *idx;
4178: const PetscScalar *b, *a, *a_q;
4179: PetscScalar *c, *c_q;
4180: PetscInt clda = sub_c->lda;
4181: PetscInt alda = sub_a->lda;
4183: PetscFunctionBegin;
4184: m = A->rmap->n;
4185: n = A->cmap->n;
4186: p = B->cmap->n;
4187: a = sub_a->v;
4188: b = sub_b->a;
4189: c = sub_c->v;
4190: if (clda == m) {
4191: PetscCall(PetscArrayzero(c, m * p));
4192: } else {
4193: for (j = 0; j < p; j++)
4194: for (i = 0; i < m; i++) c[j * clda + i] = 0.0;
4195: }
4196: ii = sub_b->i;
4197: idx = sub_b->j;
4198: for (i = 0; i < n; i++) {
4199: q = ii[i + 1] - ii[i];
4200: while (q-- > 0) {
4201: c_q = c + clda * (*idx);
4202: a_q = a + alda * i;
4203: PetscKernelAXPY(c_q, *b, a_q, m);
4204: idx++;
4205: b++;
4206: }
4207: }
4208: PetscFunctionReturn(PETSC_SUCCESS);
4209: }
4211: PetscErrorCode MatMatMultSymbolic_SeqDense_SeqAIJ(Mat A, Mat B, PetscReal fill, Mat C)
4212: {
4213: PetscInt m = A->rmap->n, n = B->cmap->n;
4214: PetscBool cisdense;
4216: PetscFunctionBegin;
4217: 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);
4218: PetscCall(MatSetSizes(C, m, n, m, n));
4219: PetscCall(MatSetBlockSizesFromMats(C, A, B));
4220: PetscCall(PetscObjectTypeCompareAny((PetscObject)C, &cisdense, MATSEQDENSE, MATSEQDENSECUDA, MATSEQDENSEHIP, ""));
4221: if (!cisdense) PetscCall(MatSetType(C, MATDENSE));
4222: PetscCall(MatSetUp(C));
4224: C->ops->matmultnumeric = MatMatMultNumeric_SeqDense_SeqAIJ;
4225: PetscFunctionReturn(PETSC_SUCCESS);
4226: }
4228: /*MC
4229: MATSEQAIJ - MATSEQAIJ = "seqaij" - A matrix type to be used for sequential sparse matrices,
4230: based on compressed sparse row format.
4232: Options Database Key:
4233: . -mat_type seqaij - sets the matrix type to "seqaij" during a call to MatSetFromOptions()
4235: Level: beginner
4237: Notes:
4238: `MatSetValues()` may be called for this matrix type with a `NULL` argument for the numerical values,
4239: in this case the values associated with the rows and columns one passes in are set to zero
4240: in the matrix
4242: `MatSetOptions`(,`MAT_STRUCTURE_ONLY`,`PETSC_TRUE`) may be called for this matrix type. In this no
4243: space is allocated for the nonzero entries and any entries passed with `MatSetValues()` are ignored
4245: Developer Note:
4246: It would be nice if all matrix formats supported passing `NULL` in for the numerical values
4248: .seealso: [](ch_matrices), `Mat`, `MatCreateSeqAIJ()`, `MatSetFromOptions()`, `MatSetType()`, `MatCreate()`, `MatType`, `MATSELL`, `MATSEQSELL`, `MATMPISELL`
4249: M*/
4251: /*MC
4252: MATAIJ - MATAIJ = "aij" - A matrix type to be used for sparse matrices.
4254: This matrix type is identical to `MATSEQAIJ` when constructed with a single process communicator,
4255: and `MATMPIAIJ` otherwise. As a result, for single process communicators,
4256: `MatSeqAIJSetPreallocation()` is supported, and similarly `MatMPIAIJSetPreallocation()` is supported
4257: for communicators controlling multiple processes. It is recommended that you call both of
4258: the above preallocation routines for simplicity.
4260: Options Database Key:
4261: . -mat_type aij - sets the matrix type to "aij" during a call to `MatSetFromOptions()`
4263: Level: beginner
4265: Note:
4266: Subclasses include `MATAIJCUSPARSE`, `MATAIJPERM`, `MATAIJSELL`, `MATAIJMKL`, `MATAIJCRL`, and also automatically switches over to use inodes when
4267: enough exist.
4269: .seealso: [](ch_matrices), `Mat`, `MatCreateAIJ()`, `MatCreateSeqAIJ()`, `MATSEQAIJ`, `MATMPIAIJ`, `MATSELL`, `MATSEQSELL`, `MATMPISELL`
4270: M*/
4272: PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqAIJCRL(Mat, MatType, MatReuse, Mat *);
4273: #if PetscDefined(HAVE_ELEMENTAL)
4274: PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_Elemental(Mat, MatType, MatReuse, Mat *);
4275: #endif
4276: #if PetscDefined(HAVE_SCALAPACK) && (PetscDefined(USE_REAL_SINGLE) || PetscDefined(USE_REAL_DOUBLE))
4277: PETSC_INTERN PetscErrorCode MatConvert_AIJ_ScaLAPACK(Mat, MatType, MatReuse, Mat *);
4278: #endif
4279: #if PetscDefined(HAVE_HYPRE)
4280: PETSC_INTERN PetscErrorCode MatConvert_AIJ_HYPRE(Mat A, MatType, MatReuse, Mat *);
4281: #endif
4283: PETSC_EXTERN PetscErrorCode MatConvert_SeqAIJ_SeqSELL(Mat, MatType, MatReuse, Mat *);
4284: PETSC_INTERN PetscErrorCode MatConvert_XAIJ_IS(Mat, MatType, MatReuse, Mat *);
4285: PETSC_INTERN PetscErrorCode MatProductSetFromOptions_IS_XAIJ(Mat);
4287: /*@
4288: MatSeqAIJGetArray - gives read/write access to the array where the data for a `MATSEQAIJ` matrix is stored
4290: Not Collective
4292: Input Parameter:
4293: . A - a `MATSEQAIJ` matrix
4295: Output Parameter:
4296: . array - pointer to the data
4298: Level: intermediate
4300: .seealso: [](ch_matrices), `Mat`, `MatSeqAIJRestoreArray()`
4301: @*/
4302: PetscErrorCode MatSeqAIJGetArray(Mat A, PetscScalar *array[])
4303: {
4304: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)A->data;
4306: PetscFunctionBegin;
4307: if (aij->ops->getarray) {
4308: PetscCall((*aij->ops->getarray)(A, array));
4309: } else {
4310: *array = aij->a;
4311: }
4312: PetscFunctionReturn(PETSC_SUCCESS);
4313: }
4315: /*@
4316: MatSeqAIJRestoreArray - returns access to the array where the data for a `MATSEQAIJ` matrix is stored obtained by `MatSeqAIJGetArray()`
4318: Not Collective
4320: Input Parameters:
4321: + A - a `MATSEQAIJ` matrix
4322: - array - pointer to the data
4324: Level: intermediate
4326: .seealso: [](ch_matrices), `Mat`, `MatSeqAIJGetArray()`
4327: @*/
4328: PetscErrorCode MatSeqAIJRestoreArray(Mat A, PetscScalar *array[])
4329: {
4330: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)A->data;
4332: PetscFunctionBegin;
4333: if (aij->ops->restorearray) {
4334: PetscCall((*aij->ops->restorearray)(A, array));
4335: } else {
4336: *array = NULL;
4337: }
4338: PetscCall(PetscObjectStateIncrease((PetscObject)A));
4339: PetscFunctionReturn(PETSC_SUCCESS);
4340: }
4342: /*@
4343: MatSeqAIJGetArrayRead - gives read-only access to the array where the data for a `MATSEQAIJ` matrix is stored
4345: Not Collective
4347: Input Parameter:
4348: . A - a `MATSEQAIJ` matrix
4350: Output Parameter:
4351: . array - pointer to the data
4353: Level: intermediate
4355: .seealso: [](ch_matrices), `Mat`, `MatSeqAIJGetArray()`, `MatSeqAIJRestoreArrayRead()`
4356: @*/
4357: PetscErrorCode MatSeqAIJGetArrayRead(Mat A, const PetscScalar *array[])
4358: {
4359: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)A->data;
4361: PetscFunctionBegin;
4362: if (aij->ops->getarrayread) {
4363: PetscCall((*aij->ops->getarrayread)(A, array));
4364: } else {
4365: *array = aij->a;
4366: }
4367: PetscFunctionReturn(PETSC_SUCCESS);
4368: }
4370: /*@
4371: MatSeqAIJRestoreArrayRead - restore the read-only access array obtained from `MatSeqAIJGetArrayRead()`
4373: Not Collective
4375: Input Parameter:
4376: . A - a `MATSEQAIJ` matrix
4378: Output Parameter:
4379: . array - pointer to the data
4381: Level: intermediate
4383: .seealso: [](ch_matrices), `Mat`, `MatSeqAIJGetArray()`, `MatSeqAIJGetArrayRead()`
4384: @*/
4385: PetscErrorCode MatSeqAIJRestoreArrayRead(Mat A, const PetscScalar *array[])
4386: {
4387: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)A->data;
4389: PetscFunctionBegin;
4390: if (aij->ops->restorearrayread) {
4391: PetscCall((*aij->ops->restorearrayread)(A, array));
4392: } else {
4393: *array = NULL;
4394: }
4395: PetscFunctionReturn(PETSC_SUCCESS);
4396: }
4398: /*@
4399: MatSeqAIJGetArrayWrite - gives write-only access to the array where the data for a `MATSEQAIJ` matrix is stored
4401: Not Collective
4403: Input Parameter:
4404: . A - a `MATSEQAIJ` matrix
4406: Output Parameter:
4407: . array - pointer to the data
4409: Level: intermediate
4411: .seealso: [](ch_matrices), `Mat`, `MatSeqAIJGetArray()`, `MatSeqAIJRestoreArrayWrite()`
4412: @*/
4413: PetscErrorCode MatSeqAIJGetArrayWrite(Mat A, PetscScalar *array[])
4414: {
4415: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)A->data;
4417: PetscFunctionBegin;
4418: if (aij->ops->getarraywrite) {
4419: PetscCall((*aij->ops->getarraywrite)(A, array));
4420: } else {
4421: *array = aij->a;
4422: }
4423: PetscCall(PetscObjectStateIncrease((PetscObject)A));
4424: PetscFunctionReturn(PETSC_SUCCESS);
4425: }
4427: /*@
4428: MatSeqAIJRestoreArrayWrite - restore the write-only access array obtained from `MatSeqAIJGetArrayWrite()`
4430: Not Collective
4432: Input Parameter:
4433: . A - a `MATSEQAIJ` matrix
4435: Output Parameter:
4436: . array - pointer to the data
4438: Level: intermediate
4440: .seealso: [](ch_matrices), `Mat`, `MatSeqAIJGetArray()`, `MatSeqAIJGetArrayWrite()`
4441: @*/
4442: PetscErrorCode MatSeqAIJRestoreArrayWrite(Mat A, PetscScalar *array[])
4443: {
4444: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)A->data;
4446: PetscFunctionBegin;
4447: if (aij->ops->restorearraywrite) {
4448: PetscCall((*aij->ops->restorearraywrite)(A, array));
4449: } else {
4450: *array = NULL;
4451: }
4452: PetscFunctionReturn(PETSC_SUCCESS);
4453: }
4455: /*@
4456: MatSeqAIJGetCSRAndMemType - Get the CSR arrays and the memory type of the `MATSEQAIJ` matrix
4458: Not Collective; No Fortran Support
4460: Input Parameter:
4461: . mat - a matrix of type `MATSEQAIJ` or its subclasses
4463: Output Parameters:
4464: + i - row map array of the matrix
4465: . j - column index array of the matrix
4466: . a - data array of the matrix
4467: - mtype - memory type of the arrays
4469: Level: developer
4471: Notes:
4472: Any of the output parameters can be `NULL`, in which case the corresponding value is not returned.
4473: If mat is a device matrix, the arrays are on the device. Otherwise, they are on the host.
4475: One can call this routine on a preallocated but not assembled matrix to just get the memory of the CSR underneath the matrix.
4476: If the matrix is assembled, the data array `a` is guaranteed to have the latest values of the matrix.
4478: .seealso: [](ch_matrices), `Mat`, `MatSeqAIJGetArray()`, `MatSeqAIJGetArrayRead()`
4479: @*/
4480: PetscErrorCode MatSeqAIJGetCSRAndMemType(Mat mat, const PetscInt *i[], const PetscInt *j[], PetscScalar *a[], PetscMemType *mtype)
4481: {
4482: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)mat->data;
4484: PetscFunctionBegin;
4485: PetscCheck(mat->preallocated, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "matrix is not preallocated");
4486: if (aij->ops->getcsrandmemtype) {
4487: PetscCall((*aij->ops->getcsrandmemtype)(mat, i, j, a, mtype));
4488: } else {
4489: if (i) *i = aij->i;
4490: if (j) *j = aij->j;
4491: if (a) *a = aij->a;
4492: if (mtype) *mtype = PETSC_MEMTYPE_HOST;
4493: }
4494: PetscFunctionReturn(PETSC_SUCCESS);
4495: }
4497: /*@
4498: MatSeqAIJGetMaxRowNonzeros - returns the maximum number of nonzeros in any row
4500: Not Collective
4502: Input Parameter:
4503: . A - a `MATSEQAIJ` matrix
4505: Output Parameter:
4506: . nz - the maximum number of nonzeros in any row
4508: Level: intermediate
4510: .seealso: [](ch_matrices), `Mat`, `MatSeqAIJRestoreArray()`
4511: @*/
4512: PetscErrorCode MatSeqAIJGetMaxRowNonzeros(Mat A, PetscInt *nz)
4513: {
4514: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)A->data;
4516: PetscFunctionBegin;
4517: *nz = aij->rmax;
4518: PetscFunctionReturn(PETSC_SUCCESS);
4519: }
4521: static PetscErrorCode MatCOOStructDestroy_SeqAIJ(PetscCtxRt data)
4522: {
4523: MatCOOStruct_SeqAIJ *coo = *(MatCOOStruct_SeqAIJ **)data;
4525: PetscFunctionBegin;
4526: PetscCall(PetscFree(coo->perm));
4527: PetscCall(PetscFree(coo->jmap));
4528: PetscCall(PetscFree(coo));
4529: PetscFunctionReturn(PETSC_SUCCESS);
4530: }
4532: PetscErrorCode MatSetPreallocationCOO_SeqAIJ(Mat mat, PetscCount coo_n, PetscInt coo_i[], PetscInt coo_j[])
4533: {
4534: MPI_Comm comm;
4535: PetscInt *i, *j;
4536: PetscInt M, N, row, iprev;
4537: PetscCount k, p, q, nneg, nnz, start, end; /* Index the coo array, so use PetscCount as their type */
4538: PetscInt *Ai; /* Change to PetscCount once we use it for row pointers */
4539: PetscInt *Aj;
4540: PetscScalar *Aa;
4541: Mat_SeqAIJ *seqaij = (Mat_SeqAIJ *)mat->data;
4542: MatType rtype;
4543: PetscCount *perm, *jmap;
4544: MatCOOStruct_SeqAIJ *coo;
4545: PetscBool isorted;
4546: PetscBool hypre;
4548: PetscFunctionBegin;
4549: PetscCall(PetscObjectGetComm((PetscObject)mat, &comm));
4550: PetscCall(MatGetSize(mat, &M, &N));
4551: i = coo_i;
4552: j = coo_j;
4553: PetscCall(PetscMalloc1(coo_n, &perm));
4555: /* 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) */
4556: isorted = PETSC_TRUE;
4557: iprev = PETSC_INT_MIN;
4558: for (k = 0; k < coo_n; k++) {
4559: if (j[k] < 0) i[k] = -1;
4560: if (isorted) {
4561: if (i[k] < iprev) isorted = PETSC_FALSE;
4562: else iprev = i[k];
4563: }
4564: perm[k] = k;
4565: }
4567: /* Sort by row if not already */
4568: if (!isorted) PetscCall(PetscSortIntWithIntCountArrayPair(coo_n, i, j, perm));
4569: PetscCheck(coo_n == 0 || i[coo_n - 1] < M, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "COO row index %" PetscInt_FMT " is >= the matrix row size %" PetscInt_FMT, i[coo_n - 1], M);
4571: /* Advance k to the first row with a non-negative index */
4572: for (k = 0; k < coo_n; k++)
4573: if (i[k] >= 0) break;
4574: nneg = k;
4575: 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 */
4576: nnz = 0; /* Total number of unique nonzeros to be counted */
4577: jmap++; /* Inc jmap by 1 for convenience */
4579: PetscCall(PetscShmgetAllocateArray(M + 1, sizeof(PetscInt), (void **)&Ai)); /* CSR of A */
4580: PetscCall(PetscArrayzero(Ai, M + 1));
4581: PetscCall(PetscShmgetAllocateArray(coo_n - nneg, sizeof(PetscInt), (void **)&Aj)); /* We have at most coo_n-nneg unique nonzeros */
4583: PetscCall(PetscStrcmp("_internal_COO_mat_for_hypre", ((PetscObject)mat)->name, &hypre));
4585: /* In each row, sort by column, then unique column indices to get row length */
4586: Ai++; /* Inc by 1 for convenience */
4587: q = 0; /* q-th unique nonzero, with q starting from 0 */
4588: while (k < coo_n) {
4589: PetscBool strictly_sorted; // this row is strictly sorted?
4590: PetscInt jprev;
4592: /* get [start,end) indices for this row; also check if cols in this row are strictly sorted */
4593: row = i[k];
4594: start = k;
4595: jprev = PETSC_INT_MIN;
4596: strictly_sorted = PETSC_TRUE;
4597: while (k < coo_n && i[k] == row) {
4598: if (strictly_sorted) {
4599: if (j[k] <= jprev) strictly_sorted = PETSC_FALSE;
4600: else jprev = j[k];
4601: }
4602: k++;
4603: }
4604: end = k;
4606: /* hack for HYPRE: swap min column to diag so that diagonal values will go first */
4607: if (hypre) {
4608: PetscInt minj = PETSC_INT_MAX;
4609: PetscBool hasdiag = PETSC_FALSE;
4611: if (strictly_sorted) { // fast path to swap the first and the diag
4612: PetscCount tmp;
4613: for (p = start; p < end; p++) {
4614: if (j[p] == row && p != start) {
4615: j[p] = j[start]; // swap j[], so that the diagonal value will go first (manipulated by perm[])
4616: j[start] = row;
4617: tmp = perm[start];
4618: perm[start] = perm[p]; // also swap perm[] so we can save the call to PetscSortIntWithCountArray() below
4619: perm[p] = tmp;
4620: break;
4621: }
4622: }
4623: } else {
4624: for (p = start; p < end; p++) {
4625: hasdiag = (PetscBool)(hasdiag || (j[p] == row));
4626: minj = PetscMin(minj, j[p]);
4627: }
4629: if (hasdiag) {
4630: for (p = start; p < end; p++) {
4631: if (j[p] == minj) j[p] = row;
4632: else if (j[p] == row) j[p] = minj;
4633: }
4634: }
4635: }
4636: }
4637: // sort by columns in a row. perm[] indicates their original order
4638: if (!strictly_sorted) PetscCall(PetscSortIntWithCountArray(end - start, j + start, perm + start));
4639: PetscCheck(end == start || j[end - 1] < N, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "COO column index %" PetscInt_FMT " is >= the matrix column size %" PetscInt_FMT, j[end - 1], N);
4641: if (strictly_sorted) { // fast path to set Aj[], jmap[], Ai[], nnz, q
4642: for (p = start; p < end; p++, q++) {
4643: Aj[q] = j[p];
4644: jmap[q] = 1;
4645: }
4646: PetscCall(PetscIntCast(end - start, Ai + row));
4647: nnz += Ai[row]; // q is already advanced
4648: } else {
4649: /* Find number of unique col entries in this row */
4650: Aj[q] = j[start]; /* Log the first nonzero in this row */
4651: jmap[q] = 1; /* Number of repeats of this nonzero entry */
4652: Ai[row] = 1;
4653: nnz++;
4655: for (p = start + 1; p < end; p++) { /* Scan remaining nonzero in this row */
4656: if (j[p] != j[p - 1]) { /* Meet a new nonzero */
4657: q++;
4658: jmap[q] = 1;
4659: Aj[q] = j[p];
4660: Ai[row]++;
4661: nnz++;
4662: } else {
4663: jmap[q]++;
4664: }
4665: }
4666: q++; /* Move to next row and thus next unique nonzero */
4667: }
4668: }
4670: Ai--; /* Back to the beginning of Ai[] */
4671: for (k = 0; k < M; k++) Ai[k + 1] += Ai[k];
4672: jmap--; // Back to the beginning of jmap[]
4673: jmap[0] = 0;
4674: for (k = 0; k < nnz; k++) jmap[k + 1] += jmap[k];
4676: if (nnz < coo_n - nneg) { /* Reallocate with actual number of unique nonzeros */
4677: PetscCount *jmap_new;
4678: PetscInt *Aj_new;
4680: PetscCall(PetscMalloc1(nnz + 1, &jmap_new));
4681: PetscCall(PetscArraycpy(jmap_new, jmap, nnz + 1));
4682: PetscCall(PetscFree(jmap));
4683: jmap = jmap_new;
4685: PetscCall(PetscShmgetAllocateArray(nnz, sizeof(PetscInt), (void **)&Aj_new));
4686: PetscCall(PetscArraycpy(Aj_new, Aj, nnz));
4687: PetscCall(PetscShmgetDeallocateArray((void **)&Aj));
4688: Aj = Aj_new;
4689: }
4691: if (nneg) { /* Discard heading entries with negative indices in perm[], as we'll access it from index 0 in MatSetValuesCOO */
4692: PetscCount *perm_new;
4694: PetscCall(PetscMalloc1(coo_n - nneg, &perm_new));
4695: PetscCall(PetscArraycpy(perm_new, perm + nneg, coo_n - nneg));
4696: PetscCall(PetscFree(perm));
4697: perm = perm_new;
4698: }
4700: PetscCall(MatGetRootType_Private(mat, &rtype));
4701: PetscCall(PetscShmgetAllocateArray(nnz, sizeof(PetscScalar), (void **)&Aa));
4702: PetscCall(PetscArrayzero(Aa, nnz));
4703: PetscCall(MatSetSeqAIJWithArrays_private(PETSC_COMM_SELF, M, N, Ai, Aj, Aa, rtype, mat));
4705: seqaij->free_a = seqaij->free_ij = PETSC_TRUE; /* Let newmat own Ai, Aj and Aa */
4707: // Put the COO struct in a container and then attach that to the matrix
4708: PetscCall(PetscMalloc1(1, &coo));
4709: PetscCall(PetscIntCast(nnz, &coo->nz));
4710: coo->n = coo_n;
4711: coo->Atot = coo_n - nneg; // Annz is seqaij->nz, so no need to record that again
4712: coo->jmap = jmap; // of length nnz+1
4713: coo->perm = perm;
4714: PetscCall(PetscObjectContainerCompose((PetscObject)mat, "__PETSc_MatCOOStruct_Host", coo, MatCOOStructDestroy_SeqAIJ));
4715: PetscFunctionReturn(PETSC_SUCCESS);
4716: }
4718: static PetscErrorCode MatSetValuesCOO_SeqAIJ(Mat A, const PetscScalar v[], InsertMode imode)
4719: {
4720: Mat_SeqAIJ *aseq = (Mat_SeqAIJ *)A->data;
4721: PetscCount i, j, Annz = aseq->nz;
4722: PetscCount *perm, *jmap;
4723: PetscScalar *Aa;
4724: PetscContainer container;
4725: MatCOOStruct_SeqAIJ *coo;
4727: PetscFunctionBegin;
4728: PetscCall(PetscObjectQuery((PetscObject)A, "__PETSc_MatCOOStruct_Host", (PetscObject *)&container));
4729: PetscCheck(container, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Not found MatCOOStruct on this matrix");
4730: PetscCall(PetscContainerGetPointer(container, &coo));
4731: perm = coo->perm;
4732: jmap = coo->jmap;
4733: PetscCall(MatSeqAIJGetArray(A, &Aa));
4734: for (i = 0; i < Annz; i++) {
4735: PetscScalar sum = 0.0;
4736: for (j = jmap[i]; j < jmap[i + 1]; j++) sum += v[perm[j]];
4737: Aa[i] = (imode == INSERT_VALUES ? 0.0 : Aa[i]) + sum;
4738: }
4739: PetscCall(MatSeqAIJRestoreArray(A, &Aa));
4740: PetscFunctionReturn(PETSC_SUCCESS);
4741: }
4743: #if PetscDefined(HAVE_CUDA)
4744: PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqAIJCUSPARSE(Mat, MatType, MatReuse, Mat *);
4745: #endif
4746: #if PetscDefined(HAVE_HIP)
4747: PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqAIJHIPSPARSE(Mat, MatType, MatReuse, Mat *);
4748: #endif
4749: #if PetscDefined(HAVE_KOKKOS_KERNELS)
4750: PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqAIJKokkos(Mat, MatType, MatReuse, Mat *);
4751: #endif
4753: PETSC_EXTERN PetscErrorCode MatCreate_SeqAIJ(Mat B)
4754: {
4755: Mat_SeqAIJ *b;
4756: PetscMPIInt size;
4758: PetscFunctionBegin;
4759: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)B), &size));
4760: PetscCheck(size <= 1, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Comm must be of size 1");
4762: PetscCall(PetscNew(&b));
4764: B->data = (void *)b;
4765: B->ops[0] = MatOps_Values;
4766: if (B->sortedfull) B->ops->setvalues = MatSetValues_SeqAIJ_SortedFull;
4768: b->row = NULL;
4769: b->col = NULL;
4770: b->icol = NULL;
4771: b->reallocs = 0;
4772: b->ignorezeroentries = PETSC_FALSE;
4773: b->roworiented = PETSC_TRUE;
4774: b->nonew = 0;
4775: b->diag = NULL;
4776: b->solve_work = NULL;
4777: B->spptr = NULL;
4778: b->saved_values = NULL;
4779: b->idiag = NULL;
4780: b->mdiag = NULL;
4781: b->ssor_work = NULL;
4782: b->omega = 1.0;
4783: b->fshift = 0.0;
4784: b->ibdiag = NULL;
4785: b->keepnonzeropattern = PETSC_FALSE;
4787: PetscCall(PetscObjectChangeTypeName((PetscObject)B, MATSEQAIJ));
4788: #if PetscDefined(HAVE_MATLAB)
4789: PetscCall(PetscObjectComposeFunction((PetscObject)B, "PetscMatlabEnginePut_C", MatlabEnginePut_SeqAIJ));
4790: PetscCall(PetscObjectComposeFunction((PetscObject)B, "PetscMatlabEngineGet_C", MatlabEngineGet_SeqAIJ));
4791: #endif
4792: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatSeqAIJSetColumnIndices_C", MatSeqAIJSetColumnIndices_SeqAIJ));
4793: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatStoreValues_C", MatStoreValues_SeqAIJ));
4794: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatRetrieveValues_C", MatRetrieveValues_SeqAIJ));
4795: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_seqsbaij_C", MatConvert_SeqAIJ_SeqSBAIJ));
4796: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_seqbaij_C", MatConvert_SeqAIJ_SeqBAIJ));
4797: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_seqaijperm_C", MatConvert_SeqAIJ_SeqAIJPERM));
4798: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_seqaijsell_C", MatConvert_SeqAIJ_SeqAIJSELL));
4799: #if PetscDefined(HAVE_MKL_SPARSE)
4800: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_seqaijmkl_C", MatConvert_SeqAIJ_SeqAIJMKL));
4801: #endif
4802: #if PetscDefined(HAVE_CUDA)
4803: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_seqaijcusparse_C", MatConvert_SeqAIJ_SeqAIJCUSPARSE));
4804: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_seqaijcusparse_seqaij_C", MatProductSetFromOptions_SeqAIJ));
4805: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_seqaij_seqaijcusparse_C", MatProductSetFromOptions_SeqAIJ));
4806: #endif
4807: #if PetscDefined(HAVE_HIP)
4808: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_seqaijhipsparse_C", MatConvert_SeqAIJ_SeqAIJHIPSPARSE));
4809: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_seqaijhipsparse_seqaij_C", MatProductSetFromOptions_SeqAIJ));
4810: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_seqaij_seqaijhipsparse_C", MatProductSetFromOptions_SeqAIJ));
4811: #endif
4812: #if PetscDefined(HAVE_KOKKOS_KERNELS)
4813: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_seqaijkokkos_C", MatConvert_SeqAIJ_SeqAIJKokkos));
4814: #endif
4815: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_seqaijcrl_C", MatConvert_SeqAIJ_SeqAIJCRL));
4816: #if PetscDefined(HAVE_ELEMENTAL)
4817: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_elemental_C", MatConvert_SeqAIJ_Elemental));
4818: #endif
4819: #if PetscDefined(HAVE_SCALAPACK) && (PetscDefined(USE_REAL_SINGLE) || PetscDefined(USE_REAL_DOUBLE))
4820: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_scalapack_C", MatConvert_AIJ_ScaLAPACK));
4821: #endif
4822: #if PetscDefined(HAVE_HYPRE)
4823: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_hypre_C", MatConvert_AIJ_HYPRE));
4824: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_transpose_seqaij_seqaij_C", MatProductSetFromOptions_Transpose_AIJ_AIJ));
4825: #endif
4826: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_seqdense_C", MatConvert_SeqAIJ_SeqDense));
4827: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_seqsell_C", MatConvert_SeqAIJ_SeqSELL));
4828: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_is_C", MatConvert_XAIJ_IS));
4829: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatIsTranspose_C", MatIsTranspose_SeqAIJ));
4830: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatIsHermitianTranspose_C", MatIsHermitianTranspose_SeqAIJ));
4831: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatSeqAIJSetPreallocation_C", MatSeqAIJSetPreallocation_SeqAIJ));
4832: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatResetPreallocation_C", MatResetPreallocation_SeqAIJ));
4833: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatResetHash_C", MatResetHash_SeqAIJ));
4834: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatSeqAIJSetPreallocationCSR_C", MatSeqAIJSetPreallocationCSR_SeqAIJ));
4835: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatReorderForNonzeroDiagonal_C", MatReorderForNonzeroDiagonal_SeqAIJ));
4836: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_is_seqaij_C", MatProductSetFromOptions_IS_XAIJ));
4837: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_seqdense_seqaij_C", MatProductSetFromOptions_SeqDense_SeqAIJ));
4838: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_seqaij_seqaij_C", MatProductSetFromOptions_SeqAIJ));
4839: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatSeqAIJKron_C", MatSeqAIJKron_SeqAIJ));
4840: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatSetPreallocationCOO_C", MatSetPreallocationCOO_SeqAIJ));
4841: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatSetValuesCOO_C", MatSetValuesCOO_SeqAIJ));
4842: PetscCall(MatCreate_SeqAIJ_Inode(B));
4843: PetscCall(PetscObjectChangeTypeName((PetscObject)B, MATSEQAIJ));
4844: PetscCall(MatSeqAIJSetTypeFromOptions(B)); /* this allows changing the matrix subtype to say MATSEQAIJPERM */
4845: PetscFunctionReturn(PETSC_SUCCESS);
4846: }
4848: /*
4849: Given a matrix generated with MatGetFactor() duplicates all the information in A into C
4850: */
4851: PetscErrorCode MatDuplicateNoCreate_SeqAIJ(Mat C, Mat A, MatDuplicateOption cpvalues, PetscBool mallocmatspace)
4852: {
4853: Mat_SeqAIJ *c = (Mat_SeqAIJ *)C->data, *a = (Mat_SeqAIJ *)A->data;
4854: PetscInt m = A->rmap->n, i;
4856: PetscFunctionBegin;
4857: PetscCheck(A->assembled || cpvalues == MAT_DO_NOT_COPY_VALUES, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Cannot duplicate unassembled matrix");
4859: C->factortype = A->factortype;
4860: c->row = NULL;
4861: c->col = NULL;
4862: c->icol = NULL;
4863: c->reallocs = 0;
4864: C->assembled = A->assembled;
4866: if (A->preallocated) {
4867: PetscCall(PetscLayoutReference(A->rmap, &C->rmap));
4868: PetscCall(PetscLayoutReference(A->cmap, &C->cmap));
4870: if (!A->hash_active) {
4871: PetscCall(PetscMalloc1(m, &c->imax));
4872: PetscCall(PetscArraycpy(c->imax, a->imax, m));
4873: PetscCall(PetscMalloc1(m, &c->ilen));
4874: PetscCall(PetscArraycpy(c->ilen, a->ilen, m));
4876: /* allocate the matrix space */
4877: if (mallocmatspace) {
4878: PetscCall(PetscShmgetAllocateArray(a->i[m], sizeof(PetscScalar), (void **)&c->a));
4879: PetscCall(PetscShmgetAllocateArray(a->i[m], sizeof(PetscInt), (void **)&c->j));
4880: PetscCall(PetscShmgetAllocateArray(m + 1, sizeof(PetscInt), (void **)&c->i));
4881: PetscCall(PetscArraycpy(c->i, a->i, m + 1));
4882: c->free_a = PETSC_TRUE;
4883: c->free_ij = PETSC_TRUE;
4884: if (m > 0) {
4885: PetscCall(PetscArraycpy(c->j, a->j, a->i[m]));
4886: if (cpvalues == MAT_COPY_VALUES) {
4887: const PetscScalar *aa;
4889: PetscCall(MatSeqAIJGetArrayRead(A, &aa));
4890: PetscCall(PetscArraycpy(c->a, aa, a->i[m]));
4891: PetscCall(MatSeqAIJGetArrayRead(A, &aa));
4892: } else {
4893: PetscCall(PetscArrayzero(c->a, a->i[m]));
4894: }
4895: }
4896: }
4897: C->preallocated = PETSC_TRUE;
4898: } else {
4899: PetscCheck(mallocmatspace, PetscObjectComm((PetscObject)A), PETSC_ERR_ARG_WRONGSTATE, "Cannot malloc matrix memory from a non-preallocated matrix");
4900: PetscCall(MatSetUp(C));
4901: }
4903: c->ignorezeroentries = a->ignorezeroentries;
4904: c->roworiented = a->roworiented;
4905: c->nonew = a->nonew;
4906: c->solve_work = NULL;
4907: c->saved_values = NULL;
4908: c->idiag = NULL;
4909: c->ssor_work = NULL;
4910: c->keepnonzeropattern = a->keepnonzeropattern;
4912: c->rmax = a->rmax;
4913: c->nz = a->nz;
4914: c->maxnz = a->nz; /* Since we allocate exactly the right amount */
4916: c->compressedrow.use = a->compressedrow.use;
4917: c->compressedrow.nrows = a->compressedrow.nrows;
4918: if (a->compressedrow.use) {
4919: i = a->compressedrow.nrows;
4920: PetscCall(PetscMalloc2(i + 1, &c->compressedrow.i, i, &c->compressedrow.rindex));
4921: PetscCall(PetscArraycpy(c->compressedrow.i, a->compressedrow.i, i + 1));
4922: PetscCall(PetscArraycpy(c->compressedrow.rindex, a->compressedrow.rindex, i));
4923: } else {
4924: c->compressedrow.use = PETSC_FALSE;
4925: c->compressedrow.i = NULL;
4926: c->compressedrow.rindex = NULL;
4927: }
4928: c->nonzerorowcnt = a->nonzerorowcnt;
4929: C->nonzerostate = A->nonzerostate;
4931: PetscCall(MatDuplicate_SeqAIJ_Inode(A, cpvalues, &C));
4932: }
4933: PetscCall(PetscFunctionListDuplicate(((PetscObject)A)->qlist, &((PetscObject)C)->qlist));
4934: PetscFunctionReturn(PETSC_SUCCESS);
4935: }
4937: PetscErrorCode MatDuplicate_SeqAIJ(Mat A, MatDuplicateOption cpvalues, Mat *B)
4938: {
4939: PetscFunctionBegin;
4940: PetscCall(MatCreate(PetscObjectComm((PetscObject)A), B));
4941: PetscCall(MatSetSizes(*B, A->rmap->n, A->cmap->n, A->rmap->n, A->cmap->n));
4942: if (!(A->rmap->n % A->rmap->bs) && !(A->cmap->n % A->cmap->bs)) PetscCall(MatSetBlockSizesFromMats(*B, A, A));
4943: PetscCall(MatSetType(*B, ((PetscObject)A)->type_name));
4944: PetscCall(MatDuplicateNoCreate_SeqAIJ(*B, A, cpvalues, PETSC_TRUE));
4945: PetscFunctionReturn(PETSC_SUCCESS);
4946: }
4948: PetscErrorCode MatLoad_SeqAIJ(Mat newMat, PetscViewer viewer)
4949: {
4950: PetscBool isbinary, ishdf5;
4952: PetscFunctionBegin;
4955: /* force binary viewer to load .info file if it has not yet done so */
4956: PetscCall(PetscViewerSetUp(viewer));
4957: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERBINARY, &isbinary));
4958: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERHDF5, &ishdf5));
4959: if (isbinary) {
4960: PetscCall(MatLoad_SeqAIJ_Binary(newMat, viewer));
4961: } else if (ishdf5) {
4962: #if PetscDefined(HAVE_HDF5)
4963: PetscCall(MatLoad_AIJ_HDF5(newMat, viewer));
4964: #else
4965: SETERRQ(PetscObjectComm((PetscObject)newMat), PETSC_ERR_SUP, "HDF5 not supported in this build.\nPlease reconfigure using --download-hdf5");
4966: #endif
4967: } else {
4968: 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);
4969: }
4970: PetscFunctionReturn(PETSC_SUCCESS);
4971: }
4973: PetscErrorCode MatLoad_SeqAIJ_Binary(Mat mat, PetscViewer viewer)
4974: {
4975: Mat_SeqAIJ *a = (Mat_SeqAIJ *)mat->data;
4976: PetscInt header[4], *rowlens, M, N, nz, sum, rows, cols, i;
4978: PetscFunctionBegin;
4979: PetscCall(PetscViewerSetUp(viewer));
4981: /* read in matrix header */
4982: PetscCall(PetscViewerBinaryRead(viewer, header, 4, NULL, PETSC_INT));
4983: PetscCheck(header[0] == MAT_FILE_CLASSID, PETSC_COMM_SELF, PETSC_ERR_FILE_UNEXPECTED, "Not a matrix object in file");
4984: M = header[1];
4985: N = header[2];
4986: nz = header[3];
4987: PetscCheck(M >= 0, PetscObjectComm((PetscObject)viewer), PETSC_ERR_FILE_UNEXPECTED, "Matrix row size (%" PetscInt_FMT ") in file is negative", M);
4988: PetscCheck(N >= 0, PetscObjectComm((PetscObject)viewer), PETSC_ERR_FILE_UNEXPECTED, "Matrix column size (%" PetscInt_FMT ") in file is negative", N);
4989: PetscCheck(nz >= 0, PETSC_COMM_SELF, PETSC_ERR_FILE_UNEXPECTED, "Matrix stored in special format on disk, cannot load as SeqAIJ");
4991: /* set block sizes from the viewer's .info file */
4992: PetscCall(MatLoad_Binary_BlockSizes(mat, viewer));
4993: /* set local and global sizes if not set already */
4994: if (mat->rmap->n < 0) mat->rmap->n = M;
4995: if (mat->cmap->n < 0) mat->cmap->n = N;
4996: if (mat->rmap->N < 0) mat->rmap->N = M;
4997: if (mat->cmap->N < 0) mat->cmap->N = N;
4998: PetscCall(PetscLayoutSetUp(mat->rmap));
4999: PetscCall(PetscLayoutSetUp(mat->cmap));
5001: /* check if the matrix sizes are correct */
5002: PetscCall(MatGetSize(mat, &rows, &cols));
5003: 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);
5005: /* read in row lengths */
5006: PetscCall(PetscMalloc1(M, &rowlens));
5007: PetscCall(PetscViewerBinaryRead(viewer, rowlens, M, NULL, PETSC_INT));
5008: /* check if sum(rowlens) is same as nz */
5009: sum = 0;
5010: for (i = 0; i < M; i++) sum += rowlens[i];
5011: 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);
5012: /* preallocate and check sizes */
5013: PetscCall(MatSeqAIJSetPreallocation_SeqAIJ(mat, 0, rowlens));
5014: PetscCall(MatGetSize(mat, &rows, &cols));
5015: 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);
5016: /* store row lengths */
5017: PetscCall(PetscArraycpy(a->ilen, rowlens, M));
5018: PetscCall(PetscFree(rowlens));
5020: /* fill in "i" row pointers */
5021: a->i[0] = 0;
5022: for (i = 0; i < M; i++) a->i[i + 1] = a->i[i] + a->ilen[i];
5023: /* read in "j" column indices */
5024: PetscCall(PetscViewerBinaryRead(viewer, a->j, nz, NULL, PETSC_INT));
5025: /* read in "a" nonzero values */
5026: PetscCall(PetscViewerBinaryRead(viewer, a->a, nz, NULL, PETSC_SCALAR));
5028: PetscCall(MatAssemblyBegin(mat, MAT_FINAL_ASSEMBLY));
5029: PetscCall(MatAssemblyEnd(mat, MAT_FINAL_ASSEMBLY));
5030: PetscFunctionReturn(PETSC_SUCCESS);
5031: }
5033: PetscErrorCode MatEqual_SeqAIJ(Mat A, Mat B, PetscBool *flg)
5034: {
5035: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data, *b = (Mat_SeqAIJ *)B->data;
5036: const PetscScalar *aa, *ba;
5038: PetscFunctionBegin;
5039: /* If the matrix dimensions are not equal,or no of nonzeros */
5040: if ((A->rmap->n != B->rmap->n) || (A->cmap->n != B->cmap->n) || (a->nz != b->nz)) {
5041: *flg = PETSC_FALSE;
5042: PetscFunctionReturn(PETSC_SUCCESS);
5043: }
5045: /* if the a->i are the same */
5046: PetscCall(PetscArraycmp(a->i, b->i, A->rmap->n + 1, flg));
5047: if (!*flg) PetscFunctionReturn(PETSC_SUCCESS);
5049: /* if a->j are the same */
5050: PetscCall(PetscArraycmp(a->j, b->j, a->nz, flg));
5051: if (!*flg) PetscFunctionReturn(PETSC_SUCCESS);
5053: PetscCall(MatSeqAIJGetArrayRead(A, &aa));
5054: PetscCall(MatSeqAIJGetArrayRead(B, &ba));
5055: /* if a->a are the same */
5056: PetscCall(PetscArraycmp(aa, ba, a->nz, flg));
5057: PetscCall(MatSeqAIJRestoreArrayRead(A, &aa));
5058: PetscCall(MatSeqAIJRestoreArrayRead(B, &ba));
5059: PetscFunctionReturn(PETSC_SUCCESS);
5060: }
5062: /*@
5063: MatCreateSeqAIJWithArrays - Creates an sequential `MATSEQAIJ` matrix using matrix elements (in CSR format)
5064: provided by the user.
5066: Collective
5068: Input Parameters:
5069: + comm - must be an MPI communicator of size 1
5070: . m - number of rows
5071: . n - number of columns
5072: . i - row indices; that is i[0] = 0, i[row] = i[row-1] + number of elements in that row of the matrix
5073: . j - column indices
5074: - a - matrix values
5076: Output Parameter:
5077: . mat - the matrix
5079: Level: intermediate
5081: Notes:
5082: The `i`, `j`, and `a` arrays are not copied by this routine, the user must free these arrays
5083: once the matrix is destroyed and not before
5085: You cannot set new nonzero locations into this matrix, that will generate an error.
5087: The `i` and `j` indices are 0 based
5089: The format which is used for the sparse matrix input, is equivalent to a
5090: row-major ordering.. i.e for the following matrix, the input data expected is
5091: as shown
5092: .vb
5093: 1 0 0
5094: 2 0 3
5095: 4 5 6
5097: i = {0,1,3,6} [size = nrow+1 = 3+1]
5098: j = {0,0,2,0,1,2} [size = 6]; values must be sorted for each row
5099: v = {1,2,3,4,5,6} [size = 6]
5100: .ve
5102: .seealso: [](ch_matrices), `Mat`, `MatCreate()`, `MatCreateAIJ()`, `MatCreateSeqAIJ()`, `MatCreateMPIAIJWithArrays()`, `MatMPIAIJSetPreallocationCSR()`
5103: @*/
5104: PetscErrorCode MatCreateSeqAIJWithArrays(MPI_Comm comm, PetscInt m, PetscInt n, PetscInt i[], PetscInt j[], PetscScalar a[], Mat *mat)
5105: {
5106: PetscInt ii;
5107: Mat_SeqAIJ *aij;
5109: PetscFunctionBegin;
5110: PetscCheck(m <= 0 || i[0] == 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "i (row indices) must start with 0");
5111: PetscCall(MatCreate(comm, mat));
5112: PetscCall(MatSetSizes(*mat, m, n, m, n));
5113: /* PetscCall(MatSetBlockSizes(*mat,,)); */
5114: PetscCall(MatSetType(*mat, MATSEQAIJ));
5115: PetscCall(MatSeqAIJSetPreallocation_SeqAIJ(*mat, MAT_SKIP_ALLOCATION, NULL));
5116: aij = (Mat_SeqAIJ *)(*mat)->data;
5117: PetscCall(PetscMalloc1(m, &aij->imax));
5118: PetscCall(PetscMalloc1(m, &aij->ilen));
5120: aij->i = i;
5121: aij->j = j;
5122: aij->a = a;
5123: aij->nonew = -1; /*this indicates that inserting a new value in the matrix that generates a new nonzero is an error*/
5124: aij->free_a = PETSC_FALSE;
5125: aij->free_ij = PETSC_FALSE;
5127: for (ii = 0, aij->nonzerorowcnt = 0, aij->rmax = 0; ii < m; ii++) {
5128: aij->ilen[ii] = aij->imax[ii] = i[ii + 1] - i[ii];
5129: if (PetscDefined(USE_DEBUG)) {
5130: 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]);
5131: for (PetscInt jj = i[ii] + 1; jj < i[ii + 1]; jj++) {
5132: 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);
5133: 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);
5134: }
5135: }
5136: }
5137: if (PetscDefined(USE_DEBUG)) {
5138: for (ii = 0; ii < aij->i[m]; ii++) {
5139: PetscCheck(j[ii] >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Negative column index at location = %" PetscInt_FMT " index = %" PetscInt_FMT, ii, j[ii]);
5140: 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);
5141: }
5142: }
5144: PetscCall(MatAssemblyBegin(*mat, MAT_FINAL_ASSEMBLY));
5145: PetscCall(MatAssemblyEnd(*mat, MAT_FINAL_ASSEMBLY));
5146: PetscFunctionReturn(PETSC_SUCCESS);
5147: }
5149: /*@
5150: MatCreateSeqAIJFromTriple - Creates an sequential `MATSEQAIJ` matrix using matrix elements (in COO format)
5151: provided by the user.
5153: Collective
5155: Input Parameters:
5156: + comm - must be an MPI communicator of size 1
5157: . m - number of rows
5158: . n - number of columns
5159: . i - row indices
5160: . j - column indices
5161: . a - matrix values
5162: . nz - number of nonzeros
5163: - idx - if the `i` and `j` indices start with 1 use `PETSC_TRUE` otherwise use `PETSC_FALSE`
5165: Output Parameter:
5166: . mat - the matrix
5168: Level: intermediate
5170: Example:
5171: For the following matrix, the input data expected is as shown (using 0 based indexing)
5172: .vb
5173: 1 0 0
5174: 2 0 3
5175: 4 5 6
5177: i = {0,1,1,2,2,2}
5178: j = {0,0,2,0,1,2}
5179: v = {1,2,3,4,5,6}
5180: .ve
5182: Note:
5183: Instead of using this function, users should also consider `MatSetPreallocationCOO()` and `MatSetValuesCOO()`, which allow repeated or remote entries,
5184: and are particularly useful in iterative applications.
5186: .seealso: [](ch_matrices), `Mat`, `MatCreate()`, `MatCreateAIJ()`, `MatCreateSeqAIJ()`, `MatCreateSeqAIJWithArrays()`, `MatMPIAIJSetPreallocationCSR()`, `MatSetValuesCOO()`, `MatSetPreallocationCOO()`
5187: @*/
5188: PetscErrorCode MatCreateSeqAIJFromTriple(MPI_Comm comm, PetscInt m, PetscInt n, PetscInt i[], PetscInt j[], PetscScalar a[], Mat *mat, PetscCount nz, PetscBool idx)
5189: {
5190: PetscInt ii, *nnz, one = 1, row, col;
5192: PetscFunctionBegin;
5193: PetscCall(PetscCalloc1(m, &nnz));
5194: for (ii = 0; ii < nz; ii++) nnz[i[ii] - !!idx] += 1;
5195: PetscCall(MatCreate(comm, mat));
5196: PetscCall(MatSetSizes(*mat, m, n, m, n));
5197: PetscCall(MatSetType(*mat, MATSEQAIJ));
5198: PetscCall(MatSeqAIJSetPreallocation_SeqAIJ(*mat, 0, nnz));
5199: for (ii = 0; ii < nz; ii++) {
5200: if (idx) {
5201: row = i[ii] - 1;
5202: col = j[ii] - 1;
5203: } else {
5204: row = i[ii];
5205: col = j[ii];
5206: }
5207: PetscCall(MatSetValues(*mat, one, &row, one, &col, &a[ii], ADD_VALUES));
5208: }
5209: PetscCall(MatAssemblyBegin(*mat, MAT_FINAL_ASSEMBLY));
5210: PetscCall(MatAssemblyEnd(*mat, MAT_FINAL_ASSEMBLY));
5211: PetscCall(PetscFree(nnz));
5212: PetscFunctionReturn(PETSC_SUCCESS);
5213: }
5215: PetscErrorCode MatCreateMPIMatConcatenateSeqMat_SeqAIJ(MPI_Comm comm, Mat inmat, PetscInt n, MatReuse scall, Mat *outmat)
5216: {
5217: PetscFunctionBegin;
5218: PetscCall(MatCreateMPIMatConcatenateSeqMat_MPIAIJ(comm, inmat, n, scall, outmat));
5219: PetscFunctionReturn(PETSC_SUCCESS);
5220: }
5222: /*
5223: Permute A into C's *local* index space using rowemb,colemb.
5224: The embedding are supposed to be injections and the above implies that the range of rowemb is a subset
5225: of [0,m), colemb is in [0,n).
5226: If pattern == DIFFERENT_NONZERO_PATTERN, C is preallocated according to A.
5227: */
5228: PetscErrorCode MatSetSeqMat_SeqAIJ(Mat C, IS rowemb, IS colemb, MatStructure pattern, Mat B)
5229: {
5230: /* If making this function public, change the error returned in this function away from _PLIB. */
5231: Mat_SeqAIJ *Baij;
5232: PetscBool seqaij;
5233: PetscInt m, n, *nz, i, j, count;
5234: PetscScalar v;
5235: const PetscInt *rowindices, *colindices;
5237: PetscFunctionBegin;
5238: if (!B) PetscFunctionReturn(PETSC_SUCCESS);
5239: /* Check to make sure the target matrix (and embeddings) are compatible with C and each other. */
5240: PetscCall(PetscObjectBaseTypeCompare((PetscObject)B, MATSEQAIJ, &seqaij));
5241: PetscCheck(seqaij, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Input matrix is of wrong type");
5242: if (rowemb) {
5243: PetscCall(ISGetLocalSize(rowemb, &m));
5244: 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);
5245: } else PetscCheck(C->rmap->n == B->rmap->n, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Input matrix is row-incompatible with the target matrix");
5246: if (colemb) {
5247: PetscCall(ISGetLocalSize(colemb, &n));
5248: 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);
5249: } else PetscCheck(C->cmap->n == B->cmap->n, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Input matrix is col-incompatible with the target matrix");
5251: Baij = (Mat_SeqAIJ *)B->data;
5252: rowindices = NULL;
5253: if (rowemb) PetscCall(ISGetIndices(rowemb, &rowindices));
5254: if (pattern == DIFFERENT_NONZERO_PATTERN) {
5255: PetscCall(PetscMalloc1(C->rmap->n, &nz));
5256: if (rowemb) {
5257: PetscCall(PetscArrayzero(nz, C->rmap->n));
5258: for (i = 0; i < B->rmap->n; i++) nz[rowindices[i]] = Baij->i[i + 1] - Baij->i[i];
5259: } else {
5260: for (i = 0; i < B->rmap->n; i++) nz[i] = Baij->i[i + 1] - Baij->i[i];
5261: }
5262: PetscCall(MatSeqAIJSetPreallocation(C, 0, nz));
5263: PetscCall(PetscFree(nz));
5264: }
5265: if (pattern == SUBSET_NONZERO_PATTERN) PetscCall(MatZeroEntries(C));
5266: count = 0;
5267: colindices = NULL;
5268: if (colemb) PetscCall(ISGetIndices(colemb, &colindices));
5269: for (i = 0; i < B->rmap->n; i++) {
5270: PetscInt row;
5271: row = i;
5272: if (rowindices) row = rowindices[i];
5273: for (j = Baij->i[i]; j < Baij->i[i + 1]; j++) {
5274: PetscInt col;
5275: col = Baij->j[count];
5276: if (colindices) col = colindices[col];
5277: v = Baij->a[count];
5278: PetscCall(MatSetValues(C, 1, &row, 1, &col, &v, INSERT_VALUES));
5279: ++count;
5280: }
5281: }
5282: if (colemb) PetscCall(ISRestoreIndices(colemb, &colindices));
5283: if (rowemb) PetscCall(ISRestoreIndices(rowemb, &rowindices));
5284: /* FIXME: set C's nonzerostate correctly. */
5285: /* Assembly for C is necessary. */
5286: C->preallocated = PETSC_TRUE;
5287: C->assembled = PETSC_TRUE;
5288: C->was_assembled = PETSC_FALSE;
5289: PetscFunctionReturn(PETSC_SUCCESS);
5290: }
5292: PetscErrorCode MatEliminateZeros_SeqAIJ(Mat A, PetscBool keep)
5293: {
5294: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
5295: MatScalar *aa = a->a;
5296: PetscInt m = A->rmap->n, fshift = 0, fshift_prev = 0, i, k;
5297: PetscInt *ailen = a->ilen, *imax = a->imax, *ai = a->i, *aj = a->j, rmax = 0;
5299: PetscFunctionBegin;
5300: PetscCheck(A->assembled, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Cannot eliminate zeros for unassembled matrix");
5301: if (m) rmax = ailen[0]; /* determine row with most nonzeros */
5302: for (i = 1, a->nonzerorowcnt = 0; i <= m; i++) {
5303: /* move each nonzero entry back by the amount of zero slots (fshift) before it*/
5304: for (k = ai[i - 1]; k < ai[i]; k++) {
5305: if (aa[k] == 0 && (aj[k] != i - 1 || !keep)) fshift++;
5306: else {
5307: if (aa[k] == 0 && aj[k] == i - 1) PetscCall(PetscInfo(A, "Keep the diagonal zero at row %" PetscInt_FMT "\n", i - 1));
5308: aa[k - fshift] = aa[k];
5309: aj[k - fshift] = aj[k];
5310: }
5311: }
5312: ai[i - 1] -= fshift_prev; // safe to update ai[i-1] now since it will not be used in the next iteration
5313: fshift_prev = fshift;
5314: /* reset ilen and imax for each row */
5315: ailen[i - 1] = imax[i - 1] = ai[i] - fshift - ai[i - 1];
5316: a->nonzerorowcnt += ((ai[i] - fshift - ai[i - 1]) > 0);
5317: rmax = PetscMax(rmax, ailen[i - 1]);
5318: }
5319: if (fshift) {
5320: if (m) {
5321: ai[m] -= fshift;
5322: a->nz = ai[m];
5323: }
5324: 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));
5325: A->nonzerostate++;
5326: A->info.nz_unneeded += (PetscReal)fshift;
5327: a->rmax = rmax;
5328: if (a->inode.use && a->inode.checked) PetscCall(MatSeqAIJCheckInode(A));
5329: PetscCall(MatAssemblyBegin(A, MAT_FINAL_ASSEMBLY));
5330: PetscCall(MatAssemblyEnd(A, MAT_FINAL_ASSEMBLY));
5331: }
5332: PetscFunctionReturn(PETSC_SUCCESS);
5333: }
5335: PetscFunctionList MatSeqAIJList = NULL;
5337: /*@
5338: MatSeqAIJSetType - Converts a `MATSEQAIJ` matrix to a subtype
5340: Collective
5342: Input Parameters:
5343: + mat - the matrix object
5344: - matype - matrix type
5346: Options Database Key:
5347: . -mat_seqaij_type method - for example seqaijcrl
5349: Level: intermediate
5351: .seealso: [](ch_matrices), `Mat`, `PCSetType()`, `VecSetType()`, `MatCreate()`, `MatType`
5352: @*/
5353: PetscErrorCode MatSeqAIJSetType(Mat mat, MatType matype)
5354: {
5355: PetscBool sametype;
5356: PetscErrorCode (*r)(Mat, MatType, MatReuse, Mat *);
5358: PetscFunctionBegin;
5360: PetscCall(PetscObjectTypeCompare((PetscObject)mat, matype, &sametype));
5361: if (sametype) PetscFunctionReturn(PETSC_SUCCESS);
5363: PetscCall(PetscFunctionListFind(MatSeqAIJList, matype, &r));
5364: PetscCheck(r, PetscObjectComm((PetscObject)mat), PETSC_ERR_ARG_UNKNOWN_TYPE, "Unknown Mat type given: %s", matype);
5365: PetscCall((*r)(mat, matype, MAT_INPLACE_MATRIX, &mat));
5366: PetscFunctionReturn(PETSC_SUCCESS);
5367: }
5369: /*@
5370: MatSeqAIJRegister - - Adds a new sub-matrix type for sequential `MATSEQAIJ` matrices
5372: Not Collective, No Fortran Support
5374: Input Parameters:
5375: + sname - name of a new user-defined matrix type, for example `MATSEQAIJCRL`
5376: - function - routine to convert to subtype
5378: Level: advanced
5380: Notes:
5381: `MatSeqAIJRegister()` may be called multiple times to add several user-defined solvers.
5383: Then, your matrix can be chosen with the procedural interface at runtime via the option
5384: .vb
5385: -mat_seqaij_type my_mat
5386: .ve
5388: .seealso: [](ch_matrices), `Mat`, `MatSeqAIJRegisterAll()`
5389: @*/
5390: PetscErrorCode MatSeqAIJRegister(const char sname[], PetscErrorCode (*function)(Mat, MatType, MatReuse, Mat *))
5391: {
5392: PetscFunctionBegin;
5393: PetscCall(MatInitializePackage());
5394: PetscCall(PetscFunctionListAdd(&MatSeqAIJList, sname, function));
5395: PetscFunctionReturn(PETSC_SUCCESS);
5396: }
5398: PetscBool MatSeqAIJRegisterAllCalled = PETSC_FALSE;
5400: /*@
5401: MatSeqAIJRegisterAll - Registers all of the matrix subtypes of `MATSSEQAIJ`
5403: Not Collective
5405: Level: advanced
5407: Note:
5408: This registers the versions of `MATSEQAIJ` for GPUs
5410: .seealso: [](ch_matrices), `Mat`, `MatRegisterAll()`, `MatSeqAIJRegister()`
5411: @*/
5412: PetscErrorCode MatSeqAIJRegisterAll(void)
5413: {
5414: PetscFunctionBegin;
5415: if (MatSeqAIJRegisterAllCalled) PetscFunctionReturn(PETSC_SUCCESS);
5416: MatSeqAIJRegisterAllCalled = PETSC_TRUE;
5418: PetscCall(MatSeqAIJRegister(MATSEQAIJCRL, MatConvert_SeqAIJ_SeqAIJCRL));
5419: PetscCall(MatSeqAIJRegister(MATSEQAIJPERM, MatConvert_SeqAIJ_SeqAIJPERM));
5420: PetscCall(MatSeqAIJRegister(MATSEQAIJSELL, MatConvert_SeqAIJ_SeqAIJSELL));
5421: #if PetscDefined(HAVE_MKL_SPARSE)
5422: PetscCall(MatSeqAIJRegister(MATSEQAIJMKL, MatConvert_SeqAIJ_SeqAIJMKL));
5423: #endif
5424: #if PetscDefined(HAVE_CUDA)
5425: PetscCall(MatSeqAIJRegister(MATSEQAIJCUSPARSE, MatConvert_SeqAIJ_SeqAIJCUSPARSE));
5426: #endif
5427: #if PetscDefined(HAVE_HIP)
5428: PetscCall(MatSeqAIJRegister(MATSEQAIJHIPSPARSE, MatConvert_SeqAIJ_SeqAIJHIPSPARSE));
5429: #endif
5430: #if PetscDefined(HAVE_KOKKOS_KERNELS)
5431: PetscCall(MatSeqAIJRegister(MATSEQAIJKOKKOS, MatConvert_SeqAIJ_SeqAIJKokkos));
5432: #endif
5433: #if PetscDefined(HAVE_VIENNACL) && PetscDefined(HAVE_VIENNACL_NO_CUDA)
5434: PetscCall(MatSeqAIJRegister(MATMPIAIJVIENNACL, MatConvert_SeqAIJ_SeqAIJViennaCL));
5435: #endif
5436: PetscFunctionReturn(PETSC_SUCCESS);
5437: }
5439: /*
5440: Special version for direct calls from Fortran
5441: */
5442: #if PetscDefined(HAVE_FORTRAN_CAPS)
5443: #define matsetvaluesseqaij_ MATSETVALUESSEQAIJ
5444: #elif !PetscDefined(HAVE_FORTRAN_UNDERSCORE)
5445: #define matsetvaluesseqaij_ matsetvaluesseqaij
5446: #endif
5448: /* Change these macros so can be used in void function */
5450: /* Change these macros so can be used in void function */
5451: /* Identical to PetscCallVoid, except it assigns to *_ierr */
5452: #undef PetscCall
5453: #define PetscCall(...) \
5454: do { \
5455: PetscErrorCode ierr_msv_mpiaij = __VA_ARGS__; \
5456: if (PetscUnlikely(ierr_msv_mpiaij)) { \
5457: *_ierr = PetscError(PETSC_COMM_SELF, __LINE__, PETSC_FUNCTION_NAME, __FILE__, ierr_msv_mpiaij, PETSC_ERROR_REPEAT, " "); \
5458: return; \
5459: } \
5460: } while (0)
5462: #undef SETERRQ
5463: #define SETERRQ(comm, ierr, ...) \
5464: do { \
5465: *_ierr = PetscError(comm, __LINE__, PETSC_FUNCTION_NAME, __FILE__, ierr, PETSC_ERROR_INITIAL, __VA_ARGS__); \
5466: return; \
5467: } while (0)
5469: PETSC_EXTERN void matsetvaluesseqaij_(Mat *AA, PetscInt *mm, const PetscInt im[], PetscInt *nn, const PetscInt in[], const PetscScalar v[], InsertMode *isis, PetscErrorCode *_ierr)
5470: {
5471: Mat A = *AA;
5472: PetscInt m = *mm, n = *nn;
5473: InsertMode is = *isis;
5474: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
5475: PetscInt *rp, k, low, high, t, ii, row, nrow, i, col, l, rmax, N;
5476: PetscInt *imax, *ai, *ailen;
5477: PetscInt *aj, nonew = a->nonew, lastcol = -1;
5478: MatScalar *ap, value, *aa;
5479: PetscBool ignorezeroentries = a->ignorezeroentries;
5480: PetscBool roworiented = a->roworiented;
5482: PetscFunctionBegin;
5483: MatCheckPreallocated(A, 1);
5484: imax = a->imax;
5485: ai = a->i;
5486: ailen = a->ilen;
5487: aj = a->j;
5488: aa = a->a;
5490: for (k = 0; k < m; k++) { /* loop over added rows */
5491: row = im[k];
5492: if (row < 0) continue;
5493: PetscCheck(row < A->rmap->n, PetscObjectComm((PetscObject)A), PETSC_ERR_ARG_OUTOFRANGE, "Row too large");
5494: rp = aj + ai[row];
5495: ap = aa + ai[row];
5496: rmax = imax[row];
5497: nrow = ailen[row];
5498: low = 0;
5499: high = nrow;
5500: for (l = 0; l < n; l++) { /* loop over added columns */
5501: if (in[l] < 0) continue;
5502: PetscCheck(in[l] < A->cmap->n, PetscObjectComm((PetscObject)A), PETSC_ERR_ARG_OUTOFRANGE, "Column too large");
5503: col = in[l];
5504: if (roworiented) value = v[l + k * n];
5505: else value = v[k + l * m];
5507: if (value == 0.0 && ignorezeroentries && (is == ADD_VALUES)) continue;
5509: if (col <= lastcol) low = 0;
5510: else high = nrow;
5511: lastcol = col;
5512: while (high - low > 5) {
5513: t = (low + high) / 2;
5514: if (rp[t] > col) high = t;
5515: else low = t;
5516: }
5517: for (i = low; i < high; i++) {
5518: if (rp[i] > col) break;
5519: if (rp[i] == col) {
5520: if (is == ADD_VALUES) ap[i] += value;
5521: else ap[i] = value;
5522: goto noinsert;
5523: }
5524: }
5525: if (value == 0.0 && ignorezeroentries) goto noinsert;
5526: if (nonew == 1) goto noinsert;
5527: PetscCheck(nonew != -1, PetscObjectComm((PetscObject)A), PETSC_ERR_ARG_OUTOFRANGE, "Inserting a new nonzero in the matrix");
5528: MatSeqXAIJReallocateAIJ(A, A->rmap->n, 1, nrow, row, col, rmax, aa, ai, aj, rp, ap, imax, nonew, MatScalar);
5529: N = nrow++ - 1;
5530: a->nz++;
5531: high++;
5532: /* shift up all the later entries in this row */
5533: for (ii = N; ii >= i; ii--) {
5534: rp[ii + 1] = rp[ii];
5535: ap[ii + 1] = ap[ii];
5536: }
5537: rp[i] = col;
5538: ap[i] = value;
5539: noinsert:;
5540: low = i + 1;
5541: }
5542: ailen[row] = nrow;
5543: }
5544: PetscFunctionReturnVoid();
5545: }
5546: /* Undefining these here since they were redefined from their original definition above! No
5547: * other PETSc functions should be defined past this point, as it is impossible to recover the
5548: * original definitions */
5549: #undef PetscCall
5550: #undef SETERRQ