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, 256, &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: PetscCall(PetscFree(a->imax));
1203: PetscCall(PetscFree(a->ilen));
1204: PetscCall(PetscFree(a->ipre));
1205: PetscCall(PetscFree3(a->idiag, a->mdiag, a->ssor_work));
1206: PetscCall(PetscFree(a->solve_work));
1207: PetscCall(ISDestroy(&a->icol));
1208: PetscCall(PetscFree(a->saved_values));
1209: a->compressedrow.use = PETSC_FALSE;
1210: PetscCall(PetscFree2(a->compressedrow.i, a->compressedrow.rindex));
1211: PetscCall(MatDestroy_SeqAIJ_Inode(A));
1212: PetscFunctionReturn(PETSC_SUCCESS);
1213: }
1215: static PetscErrorCode MatResetHash_SeqAIJ(Mat A)
1216: {
1217: PetscFunctionBegin;
1218: PetscCall(MatReset_SeqAIJ(A));
1219: PetscCall(MatCreate_SeqAIJ_Inode(A));
1220: PetscCall(MatSetUp_Seq_Hash(A));
1221: A->nonzerostate++;
1222: PetscFunctionReturn(PETSC_SUCCESS);
1223: }
1225: PetscErrorCode MatDestroy_SeqAIJ(Mat A)
1226: {
1227: PetscFunctionBegin;
1228: PetscCall(MatReset_SeqAIJ(A));
1229: PetscCall(PetscFree(A->data));
1231: /* MatMatMultNumeric_SeqAIJ_SeqAIJ_Sorted may allocate this.
1232: That function is so heavily used (sometimes in an hidden way through multnumeric function pointers)
1233: that is hard to properly add this data to the MatProduct data. We free it here to avoid
1234: users reusing the matrix object with different data to incur in obscure segmentation faults
1235: due to different matrix sizes */
1236: PetscCall(PetscObjectCompose((PetscObject)A, "__PETSc__ab_dense", NULL));
1238: PetscCall(PetscObjectChangeTypeName((PetscObject)A, NULL));
1239: PetscCall(PetscObjectComposeFunction((PetscObject)A, "PetscMatlabEnginePut_C", NULL));
1240: PetscCall(PetscObjectComposeFunction((PetscObject)A, "PetscMatlabEngineGet_C", NULL));
1241: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatSeqAIJSetColumnIndices_C", NULL));
1242: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatStoreValues_C", NULL));
1243: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatRetrieveValues_C", NULL));
1244: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_seqsbaij_C", NULL));
1245: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_seqbaij_C", NULL));
1246: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_seqaijperm_C", NULL));
1247: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_seqaijsell_C", NULL));
1248: #if PetscDefined(HAVE_MKL_SPARSE)
1249: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_seqaijmkl_C", NULL));
1250: #endif
1251: #if PetscDefined(HAVE_CUDA)
1252: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_seqaijcusparse_C", NULL));
1253: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_seqaijcusparse_seqaij_C", NULL));
1254: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_seqaij_seqaijcusparse_C", NULL));
1255: #endif
1256: #if PetscDefined(HAVE_HIP)
1257: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_seqaijhipsparse_C", NULL));
1258: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_seqaijhipsparse_seqaij_C", NULL));
1259: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_seqaij_seqaijhipsparse_C", NULL));
1260: #endif
1261: #if PetscDefined(HAVE_KOKKOS_KERNELS)
1262: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_seqaijkokkos_C", NULL));
1263: #endif
1264: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_seqaijcrl_C", NULL));
1265: #if PetscDefined(HAVE_ELEMENTAL)
1266: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_elemental_C", NULL));
1267: #endif
1268: #if PetscDefined(HAVE_SCALAPACK) && (PetscDefined(USE_REAL_SINGLE) || PetscDefined(USE_REAL_DOUBLE))
1269: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_scalapack_C", NULL));
1270: #endif
1271: #if PetscDefined(HAVE_HYPRE)
1272: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_hypre_C", NULL));
1273: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_transpose_seqaij_seqaij_C", NULL));
1274: #endif
1275: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_seqdense_C", NULL));
1276: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_seqsell_C", NULL));
1277: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_is_C", NULL));
1278: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatIsTranspose_C", NULL));
1279: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatIsHermitianTranspose_C", NULL));
1280: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatSeqAIJSetPreallocation_C", NULL));
1281: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatResetPreallocation_C", NULL));
1282: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatResetHash_C", NULL));
1283: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatSeqAIJSetPreallocationCSR_C", NULL));
1284: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatReorderForNonzeroDiagonal_C", NULL));
1285: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_is_seqaij_C", NULL));
1286: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_seqdense_seqaij_C", NULL));
1287: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_seqaij_seqaij_C", NULL));
1288: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatSeqAIJKron_C", NULL));
1289: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatSetPreallocationCOO_C", NULL));
1290: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatSetValuesCOO_C", NULL));
1291: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatFactorGetSolverType_C", NULL));
1292: /* these calls do not belong here: the subclasses Duplicate/Destroy are wrong */
1293: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaijsell_seqaij_C", NULL));
1294: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaijperm_seqaij_C", NULL));
1295: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_seqaij_seqaijviennacl_C", NULL));
1296: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_seqaijviennacl_seqdense_C", NULL));
1297: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatProductSetFromOptions_seqaijviennacl_seqaij_C", NULL));
1298: PetscFunctionReturn(PETSC_SUCCESS);
1299: }
1301: PetscErrorCode MatSetOption_SeqAIJ(Mat A, MatOption op, PetscBool flg)
1302: {
1303: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
1305: PetscFunctionBegin;
1306: switch (op) {
1307: case MAT_ROW_ORIENTED:
1308: a->roworiented = flg;
1309: break;
1310: case MAT_KEEP_NONZERO_PATTERN:
1311: a->keepnonzeropattern = flg;
1312: break;
1313: case MAT_NEW_NONZERO_LOCATIONS:
1314: a->nonew = (flg ? 0 : 1);
1315: break;
1316: case MAT_NEW_NONZERO_LOCATION_ERR:
1317: a->nonew = (flg ? -1 : 0);
1318: break;
1319: case MAT_NEW_NONZERO_ALLOCATION_ERR:
1320: a->nonew = (flg ? -2 : 0);
1321: break;
1322: case MAT_UNUSED_NONZERO_LOCATION_ERR:
1323: a->nounused = (flg ? -1 : 0);
1324: break;
1325: case MAT_IGNORE_ZERO_ENTRIES:
1326: a->ignorezeroentries = flg;
1327: break;
1328: case MAT_USE_INODES:
1329: PetscCall(MatSetOption_SeqAIJ_Inode(A, MAT_USE_INODES, flg));
1330: break;
1331: case MAT_SUBMAT_SINGLEIS:
1332: A->submat_singleis = flg;
1333: break;
1334: case MAT_SORTED_FULL:
1335: if (flg) A->ops->setvalues = MatSetValues_SeqAIJ_SortedFull;
1336: else A->ops->setvalues = MatSetValues_SeqAIJ;
1337: break;
1338: case MAT_FORM_EXPLICIT_TRANSPOSE:
1339: A->form_explicit_transpose = flg;
1340: break;
1341: default:
1342: break;
1343: }
1344: PetscFunctionReturn(PETSC_SUCCESS);
1345: }
1347: PETSC_INTERN PetscErrorCode MatGetDiagonal_SeqAIJ(Mat A, Vec v)
1348: {
1349: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
1350: PetscInt n, *ai = a->i;
1351: PetscScalar *x;
1352: const PetscScalar *aa;
1353: const PetscInt *diag;
1354: PetscBool diagDense;
1356: PetscFunctionBegin;
1357: PetscCall(VecGetLocalSize(v, &n));
1358: PetscCheck(n == A->rmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Nonconforming matrix and vector");
1359: PetscCall(MatSeqAIJGetArrayRead(A, &aa));
1360: if (A->factortype == MAT_FACTOR_ILU || A->factortype == MAT_FACTOR_LU) {
1361: PetscCall(MatGetDiagonalMarkers_SeqAIJ(A, &diag, NULL));
1362: PetscCall(VecGetArrayWrite(v, &x));
1363: for (PetscInt i = 0; i < n; i++) x[i] = 1.0 / aa[diag[i]];
1364: PetscCall(VecRestoreArrayWrite(v, &x));
1365: PetscCall(MatSeqAIJRestoreArrayRead(A, &aa));
1366: PetscFunctionReturn(PETSC_SUCCESS);
1367: }
1369: PetscCheck(A->factortype == MAT_FACTOR_NONE, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Not for factor matrices that are not ILU or LU");
1370: PetscCall(MatGetDiagonalMarkers_SeqAIJ(A, &diag, &diagDense));
1371: PetscCall(VecGetArrayWrite(v, &x));
1372: if (diagDense) {
1373: for (PetscInt i = 0; i < n; i++) x[i] = aa[diag[i]];
1374: } else {
1375: for (PetscInt i = 0; i < n; i++) x[i] = (diag[i] == ai[i + 1]) ? 0.0 : aa[diag[i]];
1376: }
1377: PetscCall(VecRestoreArrayWrite(v, &x));
1378: PetscCall(MatSeqAIJRestoreArrayRead(A, &aa));
1379: PetscFunctionReturn(PETSC_SUCCESS);
1380: }
1382: #include <../src/mat/impls/aij/seq/ftn-kernels/fmult.h>
1383: PetscErrorCode MatMultTransposeAdd_SeqAIJ(Mat A, Vec xx, Vec zz, Vec yy)
1384: {
1385: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
1386: const MatScalar *aa;
1387: PetscScalar *y;
1388: const PetscScalar *x;
1389: PetscInt m = A->rmap->n;
1390: #if !PetscDefined(USE_FORTRAN_KERNEL_MULTTRANSPOSEAIJ)
1391: const MatScalar *v;
1392: PetscScalar alpha;
1393: PetscInt n, i, j;
1394: const PetscInt *idx, *ii, *ridx = NULL;
1395: Mat_CompressedRow cprow = a->compressedrow;
1396: PetscBool usecprow = cprow.use;
1397: #endif
1399: PetscFunctionBegin;
1400: if (zz != yy) PetscCall(VecCopy(zz, yy));
1401: PetscCall(VecGetArrayRead(xx, &x));
1402: PetscCall(VecGetArray(yy, &y));
1403: PetscCall(MatSeqAIJGetArrayRead(A, &aa));
1405: #if PetscDefined(USE_FORTRAN_KERNEL_MULTTRANSPOSEAIJ)
1406: fortranmulttransposeaddaij_(&m, x, a->i, a->j, aa, y);
1407: #else
1408: if (usecprow) {
1409: m = cprow.nrows;
1410: ii = cprow.i;
1411: ridx = cprow.rindex;
1412: } else {
1413: ii = a->i;
1414: }
1415: for (i = 0; i < m; i++) {
1416: idx = a->j + ii[i];
1417: v = aa + ii[i];
1418: n = ii[i + 1] - ii[i];
1419: if (usecprow) {
1420: alpha = x[ridx[i]];
1421: } else {
1422: alpha = x[i];
1423: }
1424: for (j = 0; j < n; j++) y[idx[j]] += alpha * v[j];
1425: }
1426: #endif
1427: PetscCall(PetscLogFlops(2.0 * a->nz));
1428: PetscCall(VecRestoreArrayRead(xx, &x));
1429: PetscCall(VecRestoreArray(yy, &y));
1430: PetscCall(MatSeqAIJRestoreArrayRead(A, &aa));
1431: PetscFunctionReturn(PETSC_SUCCESS);
1432: }
1434: PetscErrorCode MatMultTranspose_SeqAIJ(Mat A, Vec xx, Vec yy)
1435: {
1436: PetscFunctionBegin;
1437: PetscCall(VecSet(yy, 0.0));
1438: PetscCall(MatMultTransposeAdd_SeqAIJ(A, xx, yy, yy));
1439: PetscFunctionReturn(PETSC_SUCCESS);
1440: }
1442: #include <../src/mat/impls/aij/seq/ftn-kernels/fmult.h>
1444: PetscErrorCode MatMult_SeqAIJ(Mat A, Vec xx, Vec yy)
1445: {
1446: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
1447: PetscScalar *y;
1448: const PetscScalar *x;
1449: const MatScalar *a_a;
1450: PetscInt m = A->rmap->n;
1451: const PetscInt *ii, *ridx = NULL;
1452: PetscBool usecprow = a->compressedrow.use;
1454: #if PetscDefined(HAVE_PRAGMA_DISJOINT)
1455: #pragma disjoint(*x, *y, *aa)
1456: #endif
1458: PetscFunctionBegin;
1459: if (a->inode.use && a->inode.checked) {
1460: PetscCall(MatMult_SeqAIJ_Inode(A, xx, yy));
1461: PetscFunctionReturn(PETSC_SUCCESS);
1462: }
1463: PetscCall(MatSeqAIJGetArrayRead(A, &a_a));
1464: PetscCall(VecGetArrayRead(xx, &x));
1465: PetscCall(VecGetArray(yy, &y));
1466: ii = a->i;
1467: if (usecprow) { /* use compressed row format */
1468: PetscCall(PetscArrayzero(y, m));
1469: m = a->compressedrow.nrows;
1470: ii = a->compressedrow.i;
1471: ridx = a->compressedrow.rindex;
1472: PetscPragmaUseOMPKernels(parallel for)
1473: for (PetscInt i = 0; i < m; i++) {
1474: PetscInt n = ii[i + 1] - ii[i];
1475: const PetscInt *aj = a->j + ii[i];
1476: const PetscScalar *aa = a_a + ii[i];
1477: PetscScalar sum = 0.0;
1478: PetscSparseDensePlusDot(sum, x, aa, aj, n);
1479: /* for (j=0; j<n; j++) sum += (*aa++)*x[*aj++]; */
1480: y[ridx[i]] = sum;
1481: }
1482: } else { /* do not use compressed row format */
1483: #if PetscDefined(USE_FORTRAN_KERNEL_MULTAIJ)
1484: fortranmultaij_(&m, x, ii, a->j, a_a, y);
1485: #else
1486: PetscPragmaUseOMPKernels(parallel for)
1487: for (PetscInt i = 0; i < m; i++) {
1488: PetscInt n = ii[i + 1] - ii[i];
1489: const PetscInt *aj = a->j + ii[i];
1490: const PetscScalar *aa = a_a + ii[i];
1491: PetscScalar sum = 0.0;
1492: PetscSparseDensePlusDot(sum, x, aa, aj, n);
1493: y[i] = sum;
1494: }
1495: #endif
1496: }
1497: PetscCall(PetscLogFlops(2.0 * a->nz - a->nonzerorowcnt));
1498: PetscCall(VecRestoreArrayRead(xx, &x));
1499: PetscCall(VecRestoreArray(yy, &y));
1500: PetscCall(MatSeqAIJRestoreArrayRead(A, &a_a));
1501: PetscFunctionReturn(PETSC_SUCCESS);
1502: }
1504: // HACK!!!!! Used by src/mat/tests/ex170.c
1505: PETSC_EXTERN PetscErrorCode MatMultMax_SeqAIJ(Mat A, Vec xx, Vec yy)
1506: {
1507: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
1508: PetscScalar *y;
1509: const PetscScalar *x;
1510: const MatScalar *aa, *a_a;
1511: PetscInt m = A->rmap->n;
1512: const PetscInt *aj, *ii, *ridx = NULL;
1513: PetscInt n, i, nonzerorow = 0;
1514: PetscScalar sum;
1515: PetscBool usecprow = a->compressedrow.use;
1517: #if PetscDefined(HAVE_PRAGMA_DISJOINT)
1518: #pragma disjoint(*x, *y, *aa)
1519: #endif
1521: PetscFunctionBegin;
1522: PetscCall(MatSeqAIJGetArrayRead(A, &a_a));
1523: PetscCall(VecGetArrayRead(xx, &x));
1524: PetscCall(VecGetArray(yy, &y));
1525: if (usecprow) { /* use compressed row format */
1526: m = a->compressedrow.nrows;
1527: ii = a->compressedrow.i;
1528: ridx = a->compressedrow.rindex;
1529: for (i = 0; i < m; i++) {
1530: n = ii[i + 1] - ii[i];
1531: aj = a->j + ii[i];
1532: aa = a_a + ii[i];
1533: sum = 0.0;
1534: nonzerorow += (n > 0);
1535: PetscSparseDenseMaxDot(sum, x, aa, aj, n);
1536: /* for (j=0; j<n; j++) sum += (*aa++)*x[*aj++]; */
1537: y[*ridx++] = sum;
1538: }
1539: } else { /* do not use compressed row format */
1540: ii = a->i;
1541: for (i = 0; i < m; i++) {
1542: n = ii[i + 1] - ii[i];
1543: aj = a->j + ii[i];
1544: aa = a_a + ii[i];
1545: sum = 0.0;
1546: nonzerorow += (n > 0);
1547: PetscSparseDenseMaxDot(sum, x, aa, aj, n);
1548: y[i] = sum;
1549: }
1550: }
1551: PetscCall(PetscLogFlops(2.0 * a->nz - nonzerorow));
1552: PetscCall(VecRestoreArrayRead(xx, &x));
1553: PetscCall(VecRestoreArray(yy, &y));
1554: PetscCall(MatSeqAIJRestoreArrayRead(A, &a_a));
1555: PetscFunctionReturn(PETSC_SUCCESS);
1556: }
1558: // HACK!!!!! Used by src/mat/tests/ex170.c
1559: PETSC_EXTERN PetscErrorCode MatMultAddMax_SeqAIJ(Mat A, Vec xx, Vec yy, Vec zz)
1560: {
1561: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
1562: PetscScalar *y, *z;
1563: const PetscScalar *x;
1564: const MatScalar *aa, *a_a;
1565: PetscInt m = A->rmap->n, *aj, *ii;
1566: PetscInt n, i, *ridx = NULL;
1567: PetscScalar sum;
1568: PetscBool usecprow = a->compressedrow.use;
1570: PetscFunctionBegin;
1571: PetscCall(MatSeqAIJGetArrayRead(A, &a_a));
1572: PetscCall(VecGetArrayRead(xx, &x));
1573: PetscCall(VecGetArrayPair(yy, zz, &y, &z));
1574: if (usecprow) { /* use compressed row format */
1575: if (zz != yy) PetscCall(PetscArraycpy(z, y, m));
1576: m = a->compressedrow.nrows;
1577: ii = a->compressedrow.i;
1578: ridx = a->compressedrow.rindex;
1579: for (i = 0; i < m; i++) {
1580: n = ii[i + 1] - ii[i];
1581: aj = a->j + ii[i];
1582: aa = a_a + ii[i];
1583: sum = y[*ridx];
1584: PetscSparseDenseMaxDot(sum, x, aa, aj, n);
1585: z[*ridx++] = sum;
1586: }
1587: } else { /* do not use compressed row format */
1588: ii = a->i;
1589: for (i = 0; i < m; i++) {
1590: n = ii[i + 1] - ii[i];
1591: aj = a->j + ii[i];
1592: aa = a_a + ii[i];
1593: sum = y[i];
1594: PetscSparseDenseMaxDot(sum, x, aa, aj, n);
1595: z[i] = sum;
1596: }
1597: }
1598: PetscCall(PetscLogFlops(2.0 * a->nz));
1599: PetscCall(VecRestoreArrayRead(xx, &x));
1600: PetscCall(VecRestoreArrayPair(yy, zz, &y, &z));
1601: PetscCall(MatSeqAIJRestoreArrayRead(A, &a_a));
1602: PetscFunctionReturn(PETSC_SUCCESS);
1603: }
1605: #include <../src/mat/impls/aij/seq/ftn-kernels/fmultadd.h>
1606: PetscErrorCode MatMultAdd_SeqAIJ(Mat A, Vec xx, Vec yy, Vec zz)
1607: {
1608: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
1609: PetscScalar *y, *z;
1610: const PetscScalar *x;
1611: const MatScalar *a_a;
1612: const PetscInt *ii, *ridx = NULL;
1613: PetscInt m = A->rmap->n;
1614: PetscBool usecprow = a->compressedrow.use;
1616: PetscFunctionBegin;
1617: if (a->inode.use && a->inode.checked) {
1618: PetscCall(MatMultAdd_SeqAIJ_Inode(A, xx, yy, zz));
1619: PetscFunctionReturn(PETSC_SUCCESS);
1620: }
1621: PetscCall(MatSeqAIJGetArrayRead(A, &a_a));
1622: PetscCall(VecGetArrayRead(xx, &x));
1623: PetscCall(VecGetArrayPair(yy, zz, &y, &z));
1624: if (usecprow) { /* use compressed row format */
1625: if (zz != yy) PetscCall(PetscArraycpy(z, y, m));
1626: m = a->compressedrow.nrows;
1627: ii = a->compressedrow.i;
1628: ridx = a->compressedrow.rindex;
1629: for (PetscInt i = 0; i < m; i++) {
1630: PetscInt n = ii[i + 1] - ii[i];
1631: const PetscInt *aj = a->j + ii[i];
1632: const PetscScalar *aa = a_a + ii[i];
1633: PetscScalar sum = y[*ridx];
1634: PetscSparseDensePlusDot(sum, x, aa, aj, n);
1635: z[*ridx++] = sum;
1636: }
1637: } else { /* do not use compressed row format */
1638: ii = a->i;
1639: #if PetscDefined(USE_FORTRAN_KERNEL_MULTADDAIJ)
1640: fortranmultaddaij_(&m, x, ii, a->j, a_a, y, z);
1641: #else
1642: PetscPragmaUseOMPKernels(parallel for)
1643: for (PetscInt i = 0; i < m; i++) {
1644: PetscInt n = ii[i + 1] - ii[i];
1645: const PetscInt *aj = a->j + ii[i];
1646: const PetscScalar *aa = a_a + ii[i];
1647: PetscScalar sum = y[i];
1648: PetscSparseDensePlusDot(sum, x, aa, aj, n);
1649: z[i] = sum;
1650: }
1651: #endif
1652: }
1653: PetscCall(PetscLogFlops(2.0 * a->nz));
1654: PetscCall(VecRestoreArrayRead(xx, &x));
1655: PetscCall(VecRestoreArrayPair(yy, zz, &y, &z));
1656: PetscCall(MatSeqAIJRestoreArrayRead(A, &a_a));
1657: PetscFunctionReturn(PETSC_SUCCESS);
1658: }
1660: static PetscErrorCode MatShift_SeqAIJ(Mat A, PetscScalar v)
1661: {
1662: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
1663: const PetscInt *diag;
1664: const PetscInt *ii = (const PetscInt *)a->i;
1665: PetscBool diagDense;
1667: PetscFunctionBegin;
1668: if (!A->preallocated || !a->nz) {
1669: PetscCall(MatSeqAIJSetPreallocation(A, 1, NULL));
1670: PetscCall(MatShift_Basic(A, v));
1671: PetscFunctionReturn(PETSC_SUCCESS);
1672: }
1674: PetscCall(MatGetDiagonalMarkers_SeqAIJ(A, &diag, &diagDense));
1675: if (diagDense) {
1676: PetscScalar *Aa;
1678: PetscCall(MatSeqAIJGetArray(A, &Aa));
1679: for (PetscInt i = 0; i < A->rmap->n; i++) Aa[diag[i]] += v;
1680: PetscCall(MatSeqAIJRestoreArray(A, &Aa));
1681: } else {
1682: PetscScalar *olda = a->a; /* preserve pointers to current matrix nonzeros structure and values */
1683: PetscInt *oldj = a->j, *oldi = a->i;
1684: PetscBool free_a = a->free_a, free_ij = a->free_ij;
1685: const PetscScalar *Aa;
1686: PetscInt *mdiag = NULL;
1688: PetscCall(PetscCalloc1(A->rmap->n, &mdiag));
1689: for (PetscInt i = 0; i < A->rmap->n; i++) {
1690: if (i < A->cmap->n && diag[i] >= ii[i + 1]) { /* 'out of range' rows never have diagonals */
1691: mdiag[i] = 1;
1692: }
1693: }
1694: PetscCall(MatSeqAIJGetArrayRead(A, &Aa)); // sync the host
1695: PetscCall(MatSeqAIJRestoreArrayRead(A, &Aa));
1697: a->a = NULL;
1698: a->j = NULL;
1699: a->i = NULL;
1700: /* increase the values in imax for each row where a diagonal is being inserted then reallocate the matrix data structures */
1701: for (PetscInt i = 0; i < PetscMin(A->rmap->n, A->cmap->n); i++) a->imax[i] += mdiag[i];
1702: PetscCall(MatSeqAIJSetPreallocation_SeqAIJ(A, 0, a->imax));
1704: /* copy old values into new matrix data structure */
1705: for (PetscInt i = 0; i < A->rmap->n; i++) {
1706: PetscCall(MatSetValues(A, 1, &i, a->imax[i] - mdiag[i], &oldj[oldi[i]], &olda[oldi[i]], ADD_VALUES));
1707: if (i < A->cmap->n) PetscCall(MatSetValue(A, i, i, v, ADD_VALUES));
1708: }
1709: PetscCall(MatAssemblyBegin(A, MAT_FINAL_ASSEMBLY));
1710: PetscCall(MatAssemblyEnd(A, MAT_FINAL_ASSEMBLY));
1711: if (free_a) PetscCall(PetscShmgetDeallocateArray((void **)&olda));
1712: if (free_ij) PetscCall(PetscShmgetDeallocateArray((void **)&oldj));
1713: if (free_ij) PetscCall(PetscShmgetDeallocateArray((void **)&oldi));
1714: PetscCall(PetscFree(mdiag));
1715: }
1716: PetscFunctionReturn(PETSC_SUCCESS);
1717: }
1719: #include <petscblaslapack.h>
1720: #include <petsc/private/kernels/blockinvert.h>
1722: /*
1723: Note that values is allocated externally by the PC and then passed into this routine
1724: */
1725: static PetscErrorCode MatInvertVariableBlockDiagonal_SeqAIJ(Mat A, PetscInt nblocks, const PetscInt *bsizes, PetscScalar *diag)
1726: {
1727: PetscInt n = A->rmap->n, i, ncnt = 0, *indx, j, bsizemax = 0, *v_pivots;
1728: PetscBool allowzeropivot, zeropivotdetected = PETSC_FALSE;
1729: const PetscReal shift = 0.0;
1730: PetscInt ipvt[5];
1731: PetscCount flops = 0;
1732: PetscScalar work[25], *v_work;
1734: PetscFunctionBegin;
1735: allowzeropivot = PetscNot(A->erroriffailure);
1736: for (i = 0; i < nblocks; i++) ncnt += bsizes[i];
1737: PetscCheck(ncnt == n, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Total blocksizes %" PetscInt_FMT " doesn't match number matrix rows %" PetscInt_FMT, ncnt, n);
1738: for (i = 0; i < nblocks; i++) bsizemax = PetscMax(bsizemax, bsizes[i]);
1739: PetscCall(PetscMalloc1(bsizemax, &indx));
1740: if (bsizemax > 7) PetscCall(PetscMalloc2(bsizemax, &v_work, bsizemax, &v_pivots));
1741: ncnt = 0;
1742: for (i = 0; i < nblocks; i++) {
1743: for (j = 0; j < bsizes[i]; j++) indx[j] = ncnt + j;
1744: PetscCall(MatGetValues(A, bsizes[i], indx, bsizes[i], indx, diag));
1745: switch (bsizes[i]) {
1746: case 1:
1747: *diag = 1.0 / (*diag);
1748: break;
1749: case 2:
1750: PetscCall(PetscKernel_A_gets_inverse_A_2(diag, shift, allowzeropivot, &zeropivotdetected));
1751: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
1752: PetscCall(PetscKernel_A_gets_transpose_A_2(diag));
1753: break;
1754: case 3:
1755: PetscCall(PetscKernel_A_gets_inverse_A_3(diag, shift, allowzeropivot, &zeropivotdetected));
1756: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
1757: PetscCall(PetscKernel_A_gets_transpose_A_3(diag));
1758: break;
1759: case 4:
1760: PetscCall(PetscKernel_A_gets_inverse_A_4(diag, shift, allowzeropivot, &zeropivotdetected));
1761: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
1762: PetscCall(PetscKernel_A_gets_transpose_A_4(diag));
1763: break;
1764: case 5:
1765: PetscCall(PetscKernel_A_gets_inverse_A_5(diag, ipvt, work, shift, allowzeropivot, &zeropivotdetected));
1766: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
1767: PetscCall(PetscKernel_A_gets_transpose_A_5(diag));
1768: break;
1769: case 6:
1770: PetscCall(PetscKernel_A_gets_inverse_A_6(diag, shift, allowzeropivot, &zeropivotdetected));
1771: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
1772: PetscCall(PetscKernel_A_gets_transpose_A_6(diag));
1773: break;
1774: case 7:
1775: PetscCall(PetscKernel_A_gets_inverse_A_7(diag, shift, allowzeropivot, &zeropivotdetected));
1776: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
1777: PetscCall(PetscKernel_A_gets_transpose_A_7(diag));
1778: break;
1779: default:
1780: PetscCall(PetscKernel_A_gets_inverse_A(bsizes[i], diag, v_pivots, v_work, allowzeropivot, &zeropivotdetected));
1781: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
1782: PetscCall(PetscKernel_A_gets_transpose_A_N(diag, bsizes[i]));
1783: }
1784: ncnt += bsizes[i];
1785: diag += bsizes[i] * bsizes[i];
1786: flops += 2 * PetscPowInt64(bsizes[i], 3) / 3;
1787: }
1788: PetscCall(PetscLogFlops(flops));
1789: if (bsizemax > 7) PetscCall(PetscFree2(v_work, v_pivots));
1790: PetscCall(PetscFree(indx));
1791: PetscFunctionReturn(PETSC_SUCCESS);
1792: }
1794: /*
1795: Negative shift indicates do not generate an error if there is a zero diagonal, just invert it anyways
1796: */
1797: static PetscErrorCode MatInvertDiagonalForSOR_SeqAIJ(Mat A, PetscScalar omega, PetscScalar fshift)
1798: {
1799: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
1800: PetscInt i, m = A->rmap->n;
1801: const MatScalar *v;
1802: PetscScalar *idiag, *mdiag;
1803: PetscBool diagDense;
1804: const PetscInt *diag;
1806: PetscFunctionBegin;
1807: if (a->idiagState == ((PetscObject)A)->state && a->omega == omega && a->fshift == fshift) PetscFunctionReturn(PETSC_SUCCESS);
1808: PetscCall(MatGetDiagonalMarkers_SeqAIJ(A, &diag, &diagDense));
1809: PetscCheck(diagDense, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Matrix must have all diagonal locations to invert them");
1810: if (!a->idiag) PetscCall(PetscMalloc3(m, &a->idiag, m, &a->mdiag, m, &a->ssor_work));
1812: mdiag = a->mdiag;
1813: idiag = a->idiag;
1814: PetscCall(MatSeqAIJGetArrayRead(A, &v));
1815: if (omega == 1.0 && PetscRealPart(fshift) <= 0.0) {
1816: for (i = 0; i < m; i++) {
1817: mdiag[i] = v[diag[i]];
1818: if (!PetscAbsScalar(mdiag[i])) { /* zero diagonal */
1819: PetscCheck(PetscRealPart(fshift), PETSC_COMM_SELF, PETSC_ERR_ARG_INCOMP, "Zero diagonal on row %" PetscInt_FMT, i);
1820: PetscCall(PetscInfo(A, "Zero diagonal on row %" PetscInt_FMT "\n", i));
1821: A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
1822: A->factorerror_zeropivot_value = 0.0;
1823: A->factorerror_zeropivot_row = i;
1824: }
1825: idiag[i] = 1.0 / v[diag[i]];
1826: }
1827: PetscCall(PetscLogFlops(m));
1828: } else {
1829: for (i = 0; i < m; i++) {
1830: mdiag[i] = v[diag[i]];
1831: idiag[i] = omega / (fshift + v[diag[i]]);
1832: }
1833: PetscCall(PetscLogFlops(2.0 * m));
1834: }
1835: PetscCall(MatSeqAIJRestoreArrayRead(A, &v));
1836: a->idiagState = ((PetscObject)A)->state;
1837: a->omega = omega;
1838: a->fshift = fshift;
1839: PetscFunctionReturn(PETSC_SUCCESS);
1840: }
1842: PetscErrorCode MatSOR_SeqAIJ(Mat A, Vec bb, PetscReal omega, MatSORType flag, PetscReal fshift, PetscInt its, PetscInt lits, Vec xx)
1843: {
1844: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
1845: PetscScalar *x, d, sum, *t, scale;
1846: const MatScalar *v, *idiag = NULL, *mdiag, *aa;
1847: const PetscScalar *b, *bs, *xb, *ts;
1848: PetscInt n, m = A->rmap->n, i;
1849: const PetscInt *idx, *diag;
1851: PetscFunctionBegin;
1852: if (a->inode.use && a->inode.checked && omega == 1.0 && fshift == 0.0) {
1853: PetscCall(MatSOR_SeqAIJ_Inode(A, bb, omega, flag, fshift, its, lits, xx));
1854: PetscFunctionReturn(PETSC_SUCCESS);
1855: }
1856: its = its * lits;
1857: PetscCall(MatInvertDiagonalForSOR_SeqAIJ(A, omega, fshift));
1858: PetscCall(MatGetDiagonalMarkers_SeqAIJ(A, &diag, NULL));
1859: t = a->ssor_work;
1860: idiag = a->idiag;
1861: mdiag = a->mdiag;
1863: PetscCall(MatSeqAIJGetArrayRead(A, &aa));
1864: PetscCall(VecGetArray(xx, &x));
1865: PetscCall(VecGetArrayRead(bb, &b));
1866: /* We count flops by assuming the upper triangular and lower triangular parts have the same number of nonzeros */
1867: if (flag == SOR_APPLY_UPPER) {
1868: /* apply (U + D/omega) to the vector */
1869: bs = b;
1870: for (i = 0; i < m; i++) {
1871: d = fshift + mdiag[i];
1872: n = a->i[i + 1] - diag[i] - 1;
1873: idx = a->j + diag[i] + 1;
1874: v = aa + diag[i] + 1;
1875: sum = b[i] * d / omega;
1876: PetscSparseDensePlusDot(sum, bs, v, idx, n);
1877: x[i] = sum;
1878: }
1879: PetscCall(VecRestoreArray(xx, &x));
1880: PetscCall(VecRestoreArrayRead(bb, &b));
1881: PetscCall(MatSeqAIJRestoreArrayRead(A, &aa));
1882: PetscCall(PetscLogFlops(a->nz));
1883: PetscFunctionReturn(PETSC_SUCCESS);
1884: }
1886: PetscCheck(flag != SOR_APPLY_LOWER, PETSC_COMM_SELF, PETSC_ERR_SUP, "SOR_APPLY_LOWER is not implemented");
1887: if (flag & SOR_EISENSTAT) {
1888: /* Let A = L + U + D; where L is lower triangular,
1889: U is upper triangular, E = D/omega; This routine applies
1891: (L + E)^{-1} A (U + E)^{-1}
1893: to a vector efficiently using Eisenstat's trick.
1894: */
1895: scale = (2.0 / omega) - 1.0;
1897: /* x = (E + U)^{-1} b */
1898: for (i = m - 1; i >= 0; i--) {
1899: n = a->i[i + 1] - diag[i] - 1;
1900: idx = a->j + diag[i] + 1;
1901: v = aa + diag[i] + 1;
1902: sum = b[i];
1903: PetscSparseDenseMinusDot(sum, x, v, idx, n);
1904: x[i] = sum * idiag[i];
1905: }
1907: /* t = b - (2*E - D)x */
1908: v = aa;
1909: for (i = 0; i < m; i++) t[i] = b[i] - scale * (v[*diag++]) * x[i];
1911: /* t = (E + L)^{-1}t */
1912: ts = t;
1913: diag = a->diag;
1914: for (i = 0; i < m; i++) {
1915: n = diag[i] - a->i[i];
1916: idx = a->j + a->i[i];
1917: v = aa + a->i[i];
1918: sum = t[i];
1919: PetscSparseDenseMinusDot(sum, ts, v, idx, n);
1920: t[i] = sum * idiag[i];
1921: /* x = x + t */
1922: x[i] += t[i];
1923: }
1925: PetscCall(PetscLogFlops(6.0 * m - 1 + 2.0 * a->nz));
1926: PetscCall(VecRestoreArray(xx, &x));
1927: PetscCall(VecRestoreArrayRead(bb, &b));
1928: PetscFunctionReturn(PETSC_SUCCESS);
1929: }
1930: if (flag & SOR_ZERO_INITIAL_GUESS) {
1931: if (flag & SOR_FORWARD_SWEEP || flag & SOR_LOCAL_FORWARD_SWEEP) {
1932: for (i = 0; i < m; i++) {
1933: n = diag[i] - a->i[i];
1934: idx = a->j + a->i[i];
1935: v = aa + a->i[i];
1936: sum = b[i];
1937: PetscSparseDenseMinusDot(sum, x, v, idx, n);
1938: t[i] = sum;
1939: x[i] = sum * idiag[i];
1940: }
1941: xb = t;
1942: PetscCall(PetscLogFlops(a->nz));
1943: } else xb = b;
1944: if (flag & SOR_BACKWARD_SWEEP || flag & SOR_LOCAL_BACKWARD_SWEEP) {
1945: for (i = m - 1; i >= 0; i--) {
1946: n = a->i[i + 1] - diag[i] - 1;
1947: idx = a->j + diag[i] + 1;
1948: v = aa + diag[i] + 1;
1949: sum = xb[i];
1950: PetscSparseDenseMinusDot(sum, x, v, idx, n);
1951: if (xb == b) {
1952: x[i] = sum * idiag[i];
1953: } else {
1954: x[i] = (1 - omega) * x[i] + sum * idiag[i]; /* omega in idiag */
1955: }
1956: }
1957: PetscCall(PetscLogFlops(a->nz)); /* assumes 1/2 in upper */
1958: }
1959: its--;
1960: }
1961: while (its--) {
1962: if (flag & SOR_FORWARD_SWEEP || flag & SOR_LOCAL_FORWARD_SWEEP) {
1963: for (i = 0; i < m; i++) {
1964: /* lower */
1965: n = diag[i] - a->i[i];
1966: idx = a->j + a->i[i];
1967: v = aa + a->i[i];
1968: sum = b[i];
1969: PetscSparseDenseMinusDot(sum, x, v, idx, n);
1970: t[i] = sum; /* save application of the lower-triangular part */
1971: /* upper */
1972: n = a->i[i + 1] - diag[i] - 1;
1973: idx = a->j + diag[i] + 1;
1974: v = aa + diag[i] + 1;
1975: PetscSparseDenseMinusDot(sum, x, v, idx, n);
1976: x[i] = (1. - omega) * x[i] + sum * idiag[i]; /* omega in idiag */
1977: }
1978: xb = t;
1979: PetscCall(PetscLogFlops(2.0 * a->nz));
1980: } else xb = b;
1981: if (flag & SOR_BACKWARD_SWEEP || flag & SOR_LOCAL_BACKWARD_SWEEP) {
1982: for (i = m - 1; i >= 0; i--) {
1983: sum = xb[i];
1984: if (xb == b) {
1985: /* whole matrix (no checkpointing available) */
1986: n = a->i[i + 1] - a->i[i];
1987: idx = a->j + a->i[i];
1988: v = aa + a->i[i];
1989: PetscSparseDenseMinusDot(sum, x, v, idx, n);
1990: x[i] = (1. - omega) * x[i] + (sum + mdiag[i] * x[i]) * idiag[i];
1991: } else { /* lower-triangular part has been saved, so only apply upper-triangular */
1992: n = a->i[i + 1] - diag[i] - 1;
1993: idx = a->j + diag[i] + 1;
1994: v = aa + diag[i] + 1;
1995: PetscSparseDenseMinusDot(sum, x, v, idx, n);
1996: x[i] = (1. - omega) * x[i] + sum * idiag[i]; /* omega in idiag */
1997: }
1998: }
1999: if (xb == b) PetscCall(PetscLogFlops(2.0 * a->nz));
2000: else PetscCall(PetscLogFlops(a->nz)); /* assumes 1/2 in upper */
2001: }
2002: }
2003: PetscCall(MatSeqAIJRestoreArrayRead(A, &aa));
2004: PetscCall(VecRestoreArray(xx, &x));
2005: PetscCall(VecRestoreArrayRead(bb, &b));
2006: PetscFunctionReturn(PETSC_SUCCESS);
2007: }
2009: static PetscErrorCode MatGetInfo_SeqAIJ(Mat A, MatInfoType flag, MatInfo *info)
2010: {
2011: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
2013: PetscFunctionBegin;
2014: info->block_size = 1.0;
2015: info->nz_allocated = a->maxnz;
2016: info->nz_used = a->nz;
2017: info->nz_unneeded = (a->maxnz - a->nz);
2018: info->assemblies = A->num_ass;
2019: info->mallocs = A->info.mallocs;
2020: info->memory = 0; /* REVIEW ME */
2021: if (A->factortype) {
2022: info->fill_ratio_given = A->info.fill_ratio_given;
2023: info->fill_ratio_needed = A->info.fill_ratio_needed;
2024: info->factor_mallocs = A->info.factor_mallocs;
2025: } else {
2026: info->fill_ratio_given = 0;
2027: info->fill_ratio_needed = 0;
2028: info->factor_mallocs = 0;
2029: }
2030: PetscFunctionReturn(PETSC_SUCCESS);
2031: }
2033: static PetscErrorCode MatZeroRows_SeqAIJ(Mat A, PetscInt N, const PetscInt rows[], PetscScalar diagv, Vec x, Vec b)
2034: {
2035: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
2036: PetscInt i, m = A->rmap->n - 1;
2037: const PetscScalar *xx;
2038: PetscScalar *bb, *aa;
2039: PetscInt d = 0;
2040: const PetscInt *diag;
2042: PetscFunctionBegin;
2043: if (x && b) {
2044: PetscCall(VecGetArrayRead(x, &xx));
2045: PetscCall(VecGetArray(b, &bb));
2046: for (i = 0; i < N; i++) {
2047: PetscCheck(rows[i] >= 0 && rows[i] <= m, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "row %" PetscInt_FMT " out of range", rows[i]);
2048: if (rows[i] >= A->cmap->n) continue;
2049: bb[rows[i]] = diagv * xx[rows[i]];
2050: }
2051: PetscCall(VecRestoreArrayRead(x, &xx));
2052: PetscCall(VecRestoreArray(b, &bb));
2053: }
2055: PetscCall(MatGetDiagonalMarkers_SeqAIJ(A, &diag, NULL));
2056: PetscCall(MatSeqAIJGetArray(A, &aa));
2057: if (a->keepnonzeropattern) {
2058: for (i = 0; i < N; i++) {
2059: PetscCheck(rows[i] >= 0 && rows[i] <= m, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "row %" PetscInt_FMT " out of range", rows[i]);
2060: PetscCall(PetscArrayzero(&aa[a->i[rows[i]]], a->ilen[rows[i]]));
2061: }
2062: if (diagv != 0.0) {
2063: for (i = 0; i < N; i++) {
2064: d = rows[i];
2065: if (d >= A->cmap->n) continue;
2066: 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);
2067: aa[diag[d]] = diagv;
2068: }
2069: }
2070: } else {
2071: if (diagv != 0.0) {
2072: for (i = 0; i < N; i++) {
2073: PetscCheck(rows[i] >= 0 && rows[i] <= m, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "row %" PetscInt_FMT " out of range", rows[i]);
2074: if (a->ilen[rows[i]] > 0) {
2075: if (rows[i] >= A->cmap->n) {
2076: a->ilen[rows[i]] = 0;
2077: } else {
2078: a->ilen[rows[i]] = 1;
2079: aa[a->i[rows[i]]] = diagv;
2080: a->j[a->i[rows[i]]] = rows[i];
2081: }
2082: } else if (rows[i] < A->cmap->n) { /* in case row was completely empty */
2083: PetscCall(MatSetValues_SeqAIJ(A, 1, &rows[i], 1, &rows[i], &diagv, INSERT_VALUES));
2084: }
2085: }
2086: } else {
2087: for (i = 0; i < N; i++) {
2088: PetscCheck(rows[i] >= 0 && rows[i] <= m, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "row %" PetscInt_FMT " out of range", rows[i]);
2089: a->ilen[rows[i]] = 0;
2090: }
2091: }
2092: A->nonzerostate++;
2093: }
2094: PetscCall(MatSeqAIJRestoreArray(A, &aa));
2095: PetscUseTypeMethod(A, assemblyend, MAT_FINAL_ASSEMBLY);
2096: PetscFunctionReturn(PETSC_SUCCESS);
2097: }
2099: static PetscErrorCode MatZeroRowsColumns_SeqAIJ(Mat A, PetscInt N, const PetscInt rows[], PetscScalar diagv, Vec x, Vec b)
2100: {
2101: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
2102: PetscInt i, j, m = A->rmap->n - 1, d = 0;
2103: PetscBool *zeroed, vecs = PETSC_FALSE;
2104: const PetscScalar *xx;
2105: PetscScalar *bb, *aa;
2106: const PetscInt *diag;
2107: PetscBool diagDense;
2109: PetscFunctionBegin;
2110: if (!N) PetscFunctionReturn(PETSC_SUCCESS);
2111: PetscCall(MatGetDiagonalMarkers_SeqAIJ(A, &diag, &diagDense));
2112: PetscCall(MatSeqAIJGetArray(A, &aa));
2113: if (x && b) {
2114: PetscCall(VecGetArrayRead(x, &xx));
2115: PetscCall(VecGetArray(b, &bb));
2116: vecs = PETSC_TRUE;
2117: }
2118: PetscCall(PetscCalloc1(A->rmap->n, &zeroed));
2119: for (i = 0; i < N; i++) {
2120: PetscCheck(rows[i] >= 0 && rows[i] <= m, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "row %" PetscInt_FMT " out of range", rows[i]);
2121: PetscCall(PetscArrayzero(PetscSafePointerPlusOffset(aa, a->i[rows[i]]), a->ilen[rows[i]]));
2123: zeroed[rows[i]] = PETSC_TRUE;
2124: }
2125: for (i = 0; i < A->rmap->n; i++) {
2126: if (!zeroed[i]) {
2127: for (j = a->i[i]; j < a->i[i + 1]; j++) {
2128: if (a->j[j] < A->rmap->n && zeroed[a->j[j]]) {
2129: if (vecs) bb[i] -= aa[j] * xx[a->j[j]];
2130: aa[j] = 0.0;
2131: }
2132: }
2133: } else if (vecs && i < A->cmap->N) bb[i] = diagv * xx[i];
2134: }
2135: if (x && b) {
2136: PetscCall(VecRestoreArrayRead(x, &xx));
2137: PetscCall(VecRestoreArray(b, &bb));
2138: }
2139: PetscCall(PetscFree(zeroed));
2140: if (diagv != 0.0) {
2141: if (!diagDense) {
2142: for (i = 0; i < N; i++) {
2143: if (rows[i] >= A->cmap->N) continue;
2144: 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]);
2145: PetscCall(MatSetValues_SeqAIJ(A, 1, &rows[i], 1, &rows[i], &diagv, INSERT_VALUES));
2146: }
2147: } else {
2148: for (i = 0; i < N; i++) aa[diag[rows[i]]] = diagv;
2149: }
2150: }
2151: PetscCall(MatSeqAIJRestoreArray(A, &aa));
2152: if (!diagDense) PetscUseTypeMethod(A, assemblyend, MAT_FINAL_ASSEMBLY);
2153: PetscFunctionReturn(PETSC_SUCCESS);
2154: }
2156: PetscErrorCode MatGetRow_SeqAIJ(Mat A, PetscInt row, PetscInt *nz, PetscInt **idx, PetscScalar **v)
2157: {
2158: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
2159: const PetscScalar *aa;
2161: PetscFunctionBegin;
2162: PetscCall(MatSeqAIJGetArrayRead(A, &aa));
2163: *nz = a->i[row + 1] - a->i[row];
2164: if (v) *v = PetscSafePointerPlusOffset((PetscScalar *)aa, a->i[row]);
2165: if (idx) {
2166: if (*nz && a->j) *idx = a->j + a->i[row];
2167: else *idx = NULL;
2168: }
2169: PetscCall(MatSeqAIJRestoreArrayRead(A, &aa));
2170: PetscFunctionReturn(PETSC_SUCCESS);
2171: }
2173: PetscErrorCode MatRestoreRow_SeqAIJ(Mat A, PetscInt row, PetscInt *nz, PetscInt **idx, PetscScalar **v)
2174: {
2175: PetscFunctionBegin;
2176: PetscFunctionReturn(PETSC_SUCCESS);
2177: }
2179: static PetscErrorCode MatNorm_SeqAIJ(Mat A, NormType type, PetscReal *nrm)
2180: {
2181: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
2182: const MatScalar *v;
2183: PetscReal sum = 0.0;
2184: PetscInt i, j;
2186: PetscFunctionBegin;
2187: PetscCall(MatSeqAIJGetArrayRead(A, &v));
2188: if (type == NORM_FROBENIUS) {
2189: #if PetscDefined(USE_REAL___FP16)
2190: PetscBLASInt one = 1, nz = a->nz;
2191: PetscCallBLAS("BLASnrm2", *nrm = BLASnrm2_(&nz, v, &one));
2192: #else
2193: for (i = 0; i < a->nz; i++) {
2194: sum += PetscRealPart(PetscConj(*v) * (*v));
2195: v++;
2196: }
2197: *nrm = PetscSqrtReal(sum);
2198: #endif
2199: PetscCall(PetscLogFlops(2.0 * a->nz));
2200: } else if (type == NORM_1) {
2201: PetscReal *tmp;
2202: PetscInt *jj = a->j;
2203: PetscCall(PetscCalloc1(A->cmap->n, &tmp));
2204: *nrm = 0.0;
2205: for (j = 0; j < a->nz; j++) {
2206: tmp[*jj++] += PetscAbsScalar(*v);
2207: v++;
2208: }
2209: for (j = 0; j < A->cmap->n; j++) {
2210: if (tmp[j] > *nrm) *nrm = tmp[j];
2211: }
2212: PetscCall(PetscFree(tmp));
2213: PetscCall(PetscLogFlops(a->nz));
2214: } else if (type == NORM_INFINITY) {
2215: *nrm = 0.0;
2216: for (j = 0; j < A->rmap->n; j++) {
2217: const PetscScalar *v2 = PetscSafePointerPlusOffset(v, a->i[j]);
2218: sum = 0.0;
2219: for (i = 0; i < a->i[j + 1] - a->i[j]; i++) {
2220: sum += PetscAbsScalar(*v2);
2221: v2++;
2222: }
2223: if (sum > *nrm) *nrm = sum;
2224: }
2225: PetscCall(PetscLogFlops(a->nz));
2226: } else SETERRQ(PETSC_COMM_SELF, PETSC_ERR_SUP, "No support for two norm");
2227: PetscCall(MatSeqAIJRestoreArrayRead(A, &v));
2228: PetscFunctionReturn(PETSC_SUCCESS);
2229: }
2231: static PetscErrorCode MatIsTranspose_SeqAIJ(Mat A, Mat B, PetscReal tol, PetscBool *f)
2232: {
2233: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)A->data, *bij = (Mat_SeqAIJ *)B->data;
2234: PetscInt *adx, *bdx, *aii, *bii, *aptr, *bptr;
2235: const MatScalar *va, *vb;
2236: PetscInt ma, na, mb, nb, i;
2238: PetscFunctionBegin;
2239: PetscCall(MatGetSize(A, &ma, &na));
2240: PetscCall(MatGetSize(B, &mb, &nb));
2241: if (ma != nb || na != mb) {
2242: *f = PETSC_FALSE;
2243: PetscFunctionReturn(PETSC_SUCCESS);
2244: }
2245: PetscCall(MatSeqAIJGetArrayRead(A, &va));
2246: PetscCall(MatSeqAIJGetArrayRead(B, &vb));
2247: aii = aij->i;
2248: bii = bij->i;
2249: adx = aij->j;
2250: bdx = bij->j;
2251: PetscCall(PetscMalloc1(ma, &aptr));
2252: PetscCall(PetscMalloc1(mb, &bptr));
2253: for (i = 0; i < ma; i++) aptr[i] = aii[i];
2254: for (i = 0; i < mb; i++) bptr[i] = bii[i];
2256: *f = PETSC_TRUE;
2257: for (i = 0; i < ma; i++) {
2258: while (aptr[i] < aii[i + 1]) {
2259: PetscInt idc, idr;
2260: PetscScalar vc, vr;
2261: /* column/row index/value */
2262: idc = adx[aptr[i]];
2263: idr = bdx[bptr[idc]];
2264: vc = va[aptr[i]];
2265: vr = vb[bptr[idc]];
2266: if (i != idr || PetscAbsScalar(vc - vr) > tol) {
2267: *f = PETSC_FALSE;
2268: goto done;
2269: } else {
2270: aptr[i]++;
2271: if (B || i != idc) bptr[idc]++;
2272: }
2273: }
2274: }
2275: done:
2276: PetscCall(PetscFree(aptr));
2277: PetscCall(PetscFree(bptr));
2278: PetscCall(MatSeqAIJRestoreArrayRead(A, &va));
2279: PetscCall(MatSeqAIJRestoreArrayRead(B, &vb));
2280: PetscFunctionReturn(PETSC_SUCCESS);
2281: }
2283: static PetscErrorCode MatIsHermitianTranspose_SeqAIJ(Mat A, Mat B, PetscReal tol, PetscBool *f)
2284: {
2285: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)A->data, *bij = (Mat_SeqAIJ *)B->data;
2286: PetscInt *adx, *bdx, *aii, *bii, *aptr, *bptr;
2287: MatScalar *va, *vb;
2288: PetscInt ma, na, mb, nb, i;
2290: PetscFunctionBegin;
2291: PetscCall(MatGetSize(A, &ma, &na));
2292: PetscCall(MatGetSize(B, &mb, &nb));
2293: if (ma != nb || na != mb) {
2294: *f = PETSC_FALSE;
2295: PetscFunctionReturn(PETSC_SUCCESS);
2296: }
2297: aii = aij->i;
2298: bii = bij->i;
2299: adx = aij->j;
2300: bdx = bij->j;
2301: va = aij->a;
2302: vb = bij->a;
2303: PetscCall(PetscMalloc1(ma, &aptr));
2304: PetscCall(PetscMalloc1(mb, &bptr));
2305: for (i = 0; i < ma; i++) aptr[i] = aii[i];
2306: for (i = 0; i < mb; i++) bptr[i] = bii[i];
2308: *f = PETSC_TRUE;
2309: for (i = 0; i < ma; i++) {
2310: while (aptr[i] < aii[i + 1]) {
2311: PetscInt idc, idr;
2312: PetscScalar vc, vr;
2313: /* column/row index/value */
2314: idc = adx[aptr[i]];
2315: idr = bdx[bptr[idc]];
2316: vc = va[aptr[i]];
2317: vr = vb[bptr[idc]];
2318: if (i != idr || PetscAbsScalar(vc - PetscConj(vr)) > tol) {
2319: *f = PETSC_FALSE;
2320: goto done;
2321: } else {
2322: aptr[i]++;
2323: if (B || i != idc) bptr[idc]++;
2324: }
2325: }
2326: }
2327: done:
2328: PetscCall(PetscFree(aptr));
2329: PetscCall(PetscFree(bptr));
2330: PetscFunctionReturn(PETSC_SUCCESS);
2331: }
2333: PetscErrorCode MatDiagonalScale_SeqAIJ(Mat A, Vec ll, Vec rr)
2334: {
2335: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
2336: const PetscScalar *l, *r;
2337: PetscScalar x;
2338: MatScalar *v;
2339: PetscInt i, j, m = A->rmap->n, n = A->cmap->n, M, nz = a->nz;
2340: const PetscInt *jj;
2342: PetscFunctionBegin;
2343: if (ll) {
2344: /* The local size is used so that VecMPI can be passed to this routine
2345: by MatDiagonalScale_MPIAIJ */
2346: PetscCall(VecGetLocalSize(ll, &m));
2347: PetscCheck(m == A->rmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Left scaling vector wrong length");
2348: PetscCall(VecGetArrayRead(ll, &l));
2349: PetscCall(MatSeqAIJGetArray(A, &v));
2350: for (i = 0; i < m; i++) {
2351: x = l[i];
2352: M = a->i[i + 1] - a->i[i];
2353: for (j = 0; j < M; j++) (*v++) *= x;
2354: }
2355: PetscCall(VecRestoreArrayRead(ll, &l));
2356: PetscCall(PetscLogFlops(nz));
2357: PetscCall(MatSeqAIJRestoreArray(A, &v));
2358: }
2359: if (rr) {
2360: PetscCall(VecGetLocalSize(rr, &n));
2361: PetscCheck(n == A->cmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Right scaling vector wrong length");
2362: PetscCall(VecGetArrayRead(rr, &r));
2363: PetscCall(MatSeqAIJGetArray(A, &v));
2364: jj = a->j;
2365: for (i = 0; i < nz; i++) (*v++) *= r[*jj++];
2366: PetscCall(MatSeqAIJRestoreArray(A, &v));
2367: PetscCall(VecRestoreArrayRead(rr, &r));
2368: PetscCall(PetscLogFlops(nz));
2369: }
2370: PetscFunctionReturn(PETSC_SUCCESS);
2371: }
2373: PetscErrorCode MatCreateSubMatrix_SeqAIJ(Mat A, IS isrow, IS iscol, PetscInt csize, MatReuse scall, Mat *B)
2374: {
2375: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data, *c;
2376: PetscInt *smap, i, k, kstart, kend, oldcols = A->cmap->n, *lens;
2377: PetscInt row, mat_i, *mat_j, tcol, first, step, *mat_ilen, sum, lensi;
2378: const PetscInt *irow, *icol;
2379: const PetscScalar *aa;
2380: PetscInt nrows, ncols;
2381: PetscInt *starts, *j_new, *i_new, *aj = a->j, *ai = a->i, ii, *ailen = a->ilen;
2382: MatScalar *a_new, *mat_a, *c_a;
2383: Mat C;
2384: PetscBool stride;
2386: PetscFunctionBegin;
2387: PetscCall(ISGetIndices(isrow, &irow));
2388: PetscCall(ISGetLocalSize(isrow, &nrows));
2389: PetscCall(ISGetLocalSize(iscol, &ncols));
2391: PetscCall(PetscObjectTypeCompare((PetscObject)iscol, ISSTRIDE, &stride));
2392: if (stride) {
2393: PetscCall(ISStrideGetInfo(iscol, &first, &step));
2394: } else {
2395: first = 0;
2396: step = 0;
2397: }
2398: if (stride && step == 1) {
2399: /* special case of contiguous rows */
2400: PetscCall(PetscMalloc2(nrows, &lens, nrows, &starts));
2401: /* loop over new rows determining lens and starting points */
2402: for (i = 0; i < nrows; i++) {
2403: kstart = ai[irow[i]];
2404: kend = kstart + ailen[irow[i]];
2405: starts[i] = kstart;
2406: for (k = kstart; k < kend; k++) {
2407: if (aj[k] >= first) {
2408: starts[i] = k;
2409: break;
2410: }
2411: }
2412: sum = 0;
2413: while (k < kend) {
2414: if (aj[k++] >= first + ncols) break;
2415: sum++;
2416: }
2417: lens[i] = sum;
2418: }
2419: /* create submatrix */
2420: if (scall == MAT_REUSE_MATRIX) {
2421: PetscInt n_cols, n_rows;
2422: PetscCall(MatGetSize(*B, &n_rows, &n_cols));
2423: PetscCheck(n_rows == nrows && n_cols == ncols, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Reused submatrix wrong size");
2424: PetscCall(MatZeroEntries(*B));
2425: C = *B;
2426: } else {
2427: PetscInt rbs, cbs;
2428: PetscCall(MatCreate(PetscObjectComm((PetscObject)A), &C));
2429: PetscCall(MatSetSizes(C, nrows, ncols, PETSC_DETERMINE, PETSC_DETERMINE));
2430: PetscCall(ISGetBlockSize(isrow, &rbs));
2431: PetscCall(ISGetBlockSize(iscol, &cbs));
2432: PetscCall(MatSetBlockSizes(C, rbs, cbs));
2433: PetscCall(MatSetType(C, ((PetscObject)A)->type_name));
2434: PetscCall(MatSeqAIJSetPreallocation_SeqAIJ(C, 0, lens));
2435: }
2436: c = (Mat_SeqAIJ *)C->data;
2438: /* loop over rows inserting into submatrix */
2439: PetscCall(MatSeqAIJGetArrayWrite(C, &a_new)); // Not 'a_new = c->a-new', since that raw usage ignores offload state of C
2440: j_new = c->j;
2441: i_new = c->i;
2442: PetscCall(MatSeqAIJGetArrayRead(A, &aa));
2443: for (i = 0; i < nrows; i++) {
2444: ii = starts[i];
2445: lensi = lens[i];
2446: if (lensi) {
2447: for (k = 0; k < lensi; k++) *j_new++ = aj[ii + k] - first;
2448: PetscCall(PetscArraycpy(a_new, aa + starts[i], lensi));
2449: a_new += lensi;
2450: }
2451: i_new[i + 1] = i_new[i] + lensi;
2452: c->ilen[i] = lensi;
2453: }
2454: PetscCall(MatSeqAIJRestoreArrayWrite(C, &a_new)); // Set C's offload state properly
2455: PetscCall(MatSeqAIJRestoreArrayRead(A, &aa));
2456: PetscCall(PetscFree2(lens, starts));
2457: } else {
2458: PetscCall(ISGetIndices(iscol, &icol));
2459: PetscCall(PetscCalloc1(oldcols, &smap));
2460: PetscCall(PetscMalloc1(1 + nrows, &lens));
2461: for (i = 0; i < ncols; i++) {
2462: 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);
2463: smap[icol[i]] = i + 1;
2464: }
2466: /* determine lens of each row */
2467: for (i = 0; i < nrows; i++) {
2468: kstart = ai[irow[i]];
2469: kend = kstart + a->ilen[irow[i]];
2470: lens[i] = 0;
2471: for (k = kstart; k < kend; k++) {
2472: if (smap[aj[k]]) lens[i]++;
2473: }
2474: }
2475: /* Create and fill new matrix */
2476: if (scall == MAT_REUSE_MATRIX) {
2477: PetscBool equal;
2479: c = (Mat_SeqAIJ *)((*B)->data);
2480: PetscCheck((*B)->rmap->n == nrows && (*B)->cmap->n == ncols, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Cannot reuse matrix. wrong size");
2481: PetscCall(PetscArraycmp(c->ilen, lens, (*B)->rmap->n, &equal));
2482: PetscCheck(equal, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Cannot reuse matrix. wrong number of nonzeros");
2483: PetscCall(PetscArrayzero(c->ilen, (*B)->rmap->n));
2484: C = *B;
2485: } else {
2486: PetscInt rbs, cbs;
2487: PetscCall(MatCreate(PetscObjectComm((PetscObject)A), &C));
2488: PetscCall(MatSetSizes(C, nrows, ncols, PETSC_DETERMINE, PETSC_DETERMINE));
2489: PetscCall(ISGetBlockSize(isrow, &rbs));
2490: PetscCall(ISGetBlockSize(iscol, &cbs));
2491: if (rbs > 1 || cbs > 1) PetscCall(MatSetBlockSizes(C, rbs, cbs));
2492: PetscCall(MatSetType(C, ((PetscObject)A)->type_name));
2493: PetscCall(MatSeqAIJSetPreallocation_SeqAIJ(C, 0, lens));
2494: }
2495: PetscCall(MatSeqAIJGetArrayRead(A, &aa));
2497: c = (Mat_SeqAIJ *)C->data;
2498: PetscCall(MatSeqAIJGetArrayWrite(C, &c_a)); // Not 'c->a', since that raw usage ignores offload state of C
2499: for (i = 0; i < nrows; i++) {
2500: row = irow[i];
2501: kstart = ai[row];
2502: kend = kstart + a->ilen[row];
2503: mat_i = c->i[i];
2504: mat_j = PetscSafePointerPlusOffset(c->j, mat_i);
2505: mat_a = PetscSafePointerPlusOffset(c_a, mat_i);
2506: mat_ilen = c->ilen + i;
2507: for (k = kstart; k < kend; k++) {
2508: if ((tcol = smap[a->j[k]])) {
2509: *mat_j++ = tcol - 1;
2510: *mat_a++ = aa[k];
2511: (*mat_ilen)++;
2512: }
2513: }
2514: }
2515: PetscCall(MatSeqAIJRestoreArrayRead(A, &aa));
2516: /* Free work space */
2517: PetscCall(ISRestoreIndices(iscol, &icol));
2518: PetscCall(PetscFree(smap));
2519: PetscCall(PetscFree(lens));
2520: /* sort */
2521: for (i = 0; i < nrows; i++) {
2522: PetscInt ilen;
2524: mat_i = c->i[i];
2525: mat_j = PetscSafePointerPlusOffset(c->j, mat_i);
2526: mat_a = PetscSafePointerPlusOffset(c_a, mat_i);
2527: ilen = c->ilen[i];
2528: PetscCall(PetscSortIntWithScalarArray(ilen, mat_j, mat_a));
2529: }
2530: PetscCall(MatSeqAIJRestoreArrayWrite(C, &c_a));
2531: }
2532: #if PetscDefined(HAVE_DEVICE)
2533: PetscCall(MatBindToCPU(C, A->boundtocpu));
2534: #endif
2535: PetscCall(MatAssemblyBegin(C, MAT_FINAL_ASSEMBLY));
2536: PetscCall(MatAssemblyEnd(C, MAT_FINAL_ASSEMBLY));
2538: PetscCall(ISRestoreIndices(isrow, &irow));
2539: *B = C;
2540: PetscFunctionReturn(PETSC_SUCCESS);
2541: }
2543: static PetscErrorCode MatGetMultiProcBlock_SeqAIJ(Mat mat, MPI_Comm subComm, MatReuse scall, Mat *subMat)
2544: {
2545: Mat B;
2547: PetscFunctionBegin;
2548: if (scall == MAT_INITIAL_MATRIX) {
2549: PetscCall(MatCreate(subComm, &B));
2550: PetscCall(MatSetSizes(B, mat->rmap->n, mat->cmap->n, mat->rmap->n, mat->cmap->n));
2551: PetscCall(MatSetBlockSizesFromMats(B, mat, mat));
2552: PetscCall(MatSetType(B, MATSEQAIJ));
2553: PetscCall(MatDuplicateNoCreate_SeqAIJ(B, mat, MAT_COPY_VALUES, PETSC_TRUE));
2554: *subMat = B;
2555: } else {
2556: PetscCall(MatCopy_SeqAIJ(mat, *subMat, SAME_NONZERO_PATTERN));
2557: }
2558: PetscFunctionReturn(PETSC_SUCCESS);
2559: }
2561: static PetscErrorCode MatILUFactor_SeqAIJ(Mat inA, IS row, IS col, const MatFactorInfo *info)
2562: {
2563: Mat_SeqAIJ *a = (Mat_SeqAIJ *)inA->data;
2564: Mat outA;
2565: PetscBool row_identity, col_identity;
2567: PetscFunctionBegin;
2568: PetscCheck(info->levels == 0, PETSC_COMM_SELF, PETSC_ERR_SUP, "Only levels=0 supported for in-place ilu");
2570: PetscCall(ISIdentity(row, &row_identity));
2571: PetscCall(ISIdentity(col, &col_identity));
2573: outA = inA;
2574: PetscCall(PetscFree(inA->solvertype));
2575: PetscCall(PetscStrallocpy(MATSOLVERPETSC, &inA->solvertype));
2577: PetscCall(PetscObjectReference((PetscObject)row));
2578: PetscCall(ISDestroy(&a->row));
2580: a->row = row;
2582: PetscCall(PetscObjectReference((PetscObject)col));
2583: PetscCall(ISDestroy(&a->col));
2585: a->col = col;
2587: /* Create the inverse permutation so that it can be used in MatLUFactorNumeric() */
2588: PetscCall(ISDestroy(&a->icol));
2589: PetscCall(ISInvertPermutation(col, PETSC_DECIDE, &a->icol));
2591: if (!a->solve_work) { /* this matrix may have been factored before */
2592: PetscCall(PetscMalloc1(inA->rmap->n, &a->solve_work));
2593: }
2595: if (row_identity && col_identity) {
2596: PetscCall(MatLUFactorNumeric_SeqAIJ_inplace(outA, inA, info));
2597: } else {
2598: PetscCall(MatLUFactorNumeric_SeqAIJ_InplaceWithPerm(outA, inA, info));
2599: }
2600: outA->factortype = MAT_FACTOR_LU;
2601: PetscFunctionReturn(PETSC_SUCCESS);
2602: }
2604: PetscErrorCode MatScale_SeqAIJ(Mat inA, PetscScalar alpha)
2605: {
2606: Mat_SeqAIJ *a = (Mat_SeqAIJ *)inA->data;
2607: PetscScalar *v;
2608: PetscBLASInt one = 1, bnz;
2610: PetscFunctionBegin;
2611: PetscCall(MatSeqAIJGetArray(inA, &v));
2612: PetscCall(PetscBLASIntCast(a->nz, &bnz));
2613: PetscCallBLAS("BLASscal", BLASscal_(&bnz, &alpha, v, &one));
2614: PetscCall(PetscLogFlops(a->nz));
2615: PetscCall(MatSeqAIJRestoreArray(inA, &v));
2616: PetscFunctionReturn(PETSC_SUCCESS);
2617: }
2619: PetscErrorCode MatDestroySubMatrix_Private(Mat_SubSppt *submatj)
2620: {
2621: PetscInt i;
2623: PetscFunctionBegin;
2624: if (!submatj->id) { /* delete data that are linked only to submats[id=0] */
2625: PetscCall(PetscFree4(submatj->sbuf1, submatj->ptr, submatj->tmp, submatj->ctr));
2627: for (i = 0; i < submatj->nrqr; ++i) PetscCall(PetscFree(submatj->sbuf2[i]));
2628: PetscCall(PetscFree3(submatj->sbuf2, submatj->req_size, submatj->req_source1));
2630: if (submatj->rbuf1) {
2631: PetscCall(PetscFree(submatj->rbuf1[0]));
2632: PetscCall(PetscFree(submatj->rbuf1));
2633: }
2635: for (i = 0; i < submatj->nrqs; ++i) PetscCall(PetscFree(submatj->rbuf3[i]));
2636: PetscCall(PetscFree3(submatj->req_source2, submatj->rbuf2, submatj->rbuf3));
2637: PetscCall(PetscFree(submatj->pa));
2638: }
2640: #if PetscDefined(USE_CTABLE)
2641: PetscCall(PetscHMapIDestroy(&submatj->rmap));
2642: PetscCall(PetscFree(submatj->cmap_loc));
2643: PetscCall(PetscFree(submatj->rmap_loc));
2644: #else
2645: PetscCall(PetscFree(submatj->rmap));
2646: #endif
2648: if (!submatj->allcolumns) {
2649: #if PetscDefined(USE_CTABLE)
2650: PetscCall(PetscHMapIDestroy(&submatj->cmap));
2651: #else
2652: PetscCall(PetscFree(submatj->cmap));
2653: #endif
2654: }
2655: PetscCall(PetscFree(submatj->row2proc));
2657: PetscCall(PetscFree(submatj));
2658: PetscFunctionReturn(PETSC_SUCCESS);
2659: }
2661: PetscErrorCode MatDestroySubMatrix_SeqAIJ(Mat C)
2662: {
2663: Mat_SeqAIJ *c = (Mat_SeqAIJ *)C->data;
2664: Mat_SubSppt *submatj = c->submatis1;
2666: PetscFunctionBegin;
2667: PetscCall((*submatj->destroy)(C));
2668: PetscCall(MatDestroySubMatrix_Private(submatj));
2669: PetscFunctionReturn(PETSC_SUCCESS);
2670: }
2672: /* Note this has code duplication with MatDestroySubMatrices_SeqBAIJ() */
2673: static PetscErrorCode MatDestroySubMatrices_SeqAIJ(PetscInt n, Mat *mat[])
2674: {
2675: PetscInt i;
2676: Mat C;
2677: Mat_SeqAIJ *c;
2678: Mat_SubSppt *submatj;
2680: PetscFunctionBegin;
2681: for (i = 0; i < n; i++) {
2682: C = (*mat)[i];
2683: c = (Mat_SeqAIJ *)C->data;
2684: submatj = c->submatis1;
2685: if (submatj) {
2686: if (--((PetscObject)C)->refct <= 0) {
2687: PetscCall(PetscFree(C->factorprefix));
2688: PetscCall((*submatj->destroy)(C));
2689: PetscCall(MatDestroySubMatrix_Private(submatj));
2690: PetscCall(PetscFree(C->defaultvectype));
2691: PetscCall(PetscFree(C->defaultrandtype));
2692: PetscCall(PetscLayoutDestroy(&C->rmap));
2693: PetscCall(PetscLayoutDestroy(&C->cmap));
2694: PetscCall(PetscHeaderDestroy(&C));
2695: }
2696: } else {
2697: PetscCall(MatDestroy(&C));
2698: }
2699: }
2701: /* Destroy Dummy submatrices created for reuse */
2702: PetscCall(MatDestroySubMatrices_Dummy(n, mat));
2704: PetscCall(PetscFree(*mat));
2705: PetscFunctionReturn(PETSC_SUCCESS);
2706: }
2708: static PetscErrorCode MatCreateSubMatrices_SeqAIJ(Mat A, PetscInt n, const IS irow[], const IS icol[], MatReuse scall, Mat *B[])
2709: {
2710: PetscInt i;
2712: PetscFunctionBegin;
2713: if (scall == MAT_INITIAL_MATRIX) PetscCall(PetscCalloc1(n + 1, B));
2715: for (i = 0; i < n; i++) PetscCall(MatCreateSubMatrix_SeqAIJ(A, irow[i], icol[i], PETSC_DECIDE, scall, &(*B)[i]));
2716: PetscFunctionReturn(PETSC_SUCCESS);
2717: }
2719: static PetscErrorCode MatIncreaseOverlap_SeqAIJ(Mat A, PetscInt is_max, IS is[], PetscInt ov)
2720: {
2721: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
2722: PetscInt row, i, j, k, l, ll, m, n, *nidx, isz, val;
2723: const PetscInt *idx;
2724: PetscInt start, end, *ai, *aj, bs = A->rmap->bs == A->cmap->bs ? A->rmap->bs : 1;
2725: PetscBT table;
2727: PetscFunctionBegin;
2728: m = A->rmap->n / bs;
2729: ai = a->i;
2730: aj = a->j;
2732: PetscCheck(ov >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "illegal negative overlap value used");
2734: PetscCall(PetscMalloc1(m + 1, &nidx));
2735: PetscCall(PetscBTCreate(m, &table));
2737: for (i = 0; i < is_max; i++) {
2738: /* Initialize the two local arrays */
2739: isz = 0;
2740: PetscCall(PetscBTMemzero(m, table));
2742: /* Extract the indices, assume there can be duplicate entries */
2743: PetscCall(ISGetIndices(is[i], &idx));
2744: PetscCall(ISGetLocalSize(is[i], &n));
2746: if (bs > 1) {
2747: /* Enter these into the temp arrays. I.e., mark table[row], enter row into new index */
2748: for (j = 0; j < n; ++j) {
2749: if (!PetscBTLookupSet(table, idx[j] / bs)) nidx[isz++] = idx[j] / bs;
2750: }
2751: PetscCall(ISRestoreIndices(is[i], &idx));
2752: PetscCall(ISDestroy(&is[i]));
2754: k = 0;
2755: for (j = 0; j < ov; j++) { /* for each overlap */
2756: n = isz;
2757: for (; k < n; k++) { /* do only those rows in nidx[k], which are not done yet */
2758: for (ll = 0; ll < bs; ll++) {
2759: row = bs * nidx[k] + ll;
2760: start = ai[row];
2761: end = ai[row + 1];
2762: for (l = start; l < end; l++) {
2763: val = aj[l] / bs;
2764: if (!PetscBTLookupSet(table, val)) nidx[isz++] = val;
2765: }
2766: }
2767: }
2768: }
2769: PetscCall(ISCreateBlock(PETSC_COMM_SELF, bs, isz, nidx, PETSC_COPY_VALUES, is + i));
2770: } else {
2771: /* Enter these into the temp arrays. I.e., mark table[row], enter row into new index */
2772: for (j = 0; j < n; ++j) {
2773: if (!PetscBTLookupSet(table, idx[j])) nidx[isz++] = idx[j];
2774: }
2775: PetscCall(ISRestoreIndices(is[i], &idx));
2776: PetscCall(ISDestroy(&is[i]));
2778: k = 0;
2779: for (j = 0; j < ov; j++) { /* for each overlap */
2780: n = isz;
2781: for (; k < n; k++) { /* do only those rows in nidx[k], which are not done yet */
2782: row = nidx[k];
2783: start = ai[row];
2784: end = ai[row + 1];
2785: for (l = start; l < end; l++) {
2786: val = aj[l];
2787: if (!PetscBTLookupSet(table, val)) nidx[isz++] = val;
2788: }
2789: }
2790: }
2791: PetscCall(ISCreateGeneral(PETSC_COMM_SELF, isz, nidx, PETSC_COPY_VALUES, is + i));
2792: }
2793: }
2794: PetscCall(PetscBTDestroy(&table));
2795: PetscCall(PetscFree(nidx));
2796: PetscFunctionReturn(PETSC_SUCCESS);
2797: }
2799: static PetscErrorCode MatPermute_SeqAIJ(Mat A, IS rowp, IS colp, Mat *B)
2800: {
2801: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
2802: PetscInt i, nz = 0, m = A->rmap->n, n = A->cmap->n;
2803: const PetscInt *row, *col;
2804: PetscInt *cnew, j, *lens;
2805: IS icolp, irowp;
2806: PetscInt *cwork = NULL;
2807: PetscScalar *vwork = NULL;
2809: PetscFunctionBegin;
2810: PetscCall(ISInvertPermutation(rowp, PETSC_DECIDE, &irowp));
2811: PetscCall(ISGetIndices(irowp, &row));
2812: PetscCall(ISInvertPermutation(colp, PETSC_DECIDE, &icolp));
2813: PetscCall(ISGetIndices(icolp, &col));
2815: /* determine lengths of permuted rows */
2816: PetscCall(PetscMalloc1(m + 1, &lens));
2817: for (i = 0; i < m; i++) lens[row[i]] = a->i[i + 1] - a->i[i];
2818: PetscCall(MatCreate(PetscObjectComm((PetscObject)A), B));
2819: PetscCall(MatSetSizes(*B, m, n, m, n));
2820: PetscCall(MatSetBlockSizesFromMats(*B, A, A));
2821: PetscCall(MatSetType(*B, ((PetscObject)A)->type_name));
2822: PetscCall(MatSeqAIJSetPreallocation_SeqAIJ(*B, 0, lens));
2823: PetscCall(PetscFree(lens));
2825: PetscCall(PetscMalloc1(n, &cnew));
2826: for (i = 0; i < m; i++) {
2827: PetscCall(MatGetRow_SeqAIJ(A, i, &nz, &cwork, &vwork));
2828: for (j = 0; j < nz; j++) cnew[j] = col[cwork[j]];
2829: PetscCall(MatSetValues_SeqAIJ(*B, 1, &row[i], nz, cnew, vwork, INSERT_VALUES));
2830: PetscCall(MatRestoreRow_SeqAIJ(A, i, &nz, &cwork, &vwork));
2831: }
2832: PetscCall(PetscFree(cnew));
2834: (*B)->assembled = PETSC_FALSE;
2836: #if PetscDefined(HAVE_DEVICE)
2837: PetscCall(MatBindToCPU(*B, A->boundtocpu));
2838: #endif
2839: PetscCall(MatAssemblyBegin(*B, MAT_FINAL_ASSEMBLY));
2840: PetscCall(MatAssemblyEnd(*B, MAT_FINAL_ASSEMBLY));
2841: PetscCall(ISRestoreIndices(irowp, &row));
2842: PetscCall(ISRestoreIndices(icolp, &col));
2843: PetscCall(ISDestroy(&irowp));
2844: PetscCall(ISDestroy(&icolp));
2845: if (rowp == colp) PetscCall(MatPropagateSymmetryOptions(A, *B));
2846: PetscFunctionReturn(PETSC_SUCCESS);
2847: }
2849: PetscErrorCode MatCopy_SeqAIJ(Mat A, Mat B, MatStructure str)
2850: {
2851: PetscFunctionBegin;
2852: /* If the two matrices have the same copy implementation, use fast copy. */
2853: if (str == SAME_NONZERO_PATTERN && (A->ops->copy == B->ops->copy)) {
2854: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
2855: Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data;
2856: const PetscScalar *aa;
2857: PetscScalar *bb;
2859: PetscCall(MatSeqAIJGetArrayRead(A, &aa));
2860: PetscCall(MatSeqAIJGetArrayWrite(B, &bb));
2862: 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]);
2863: PetscCall(PetscArraycpy(bb, aa, a->i[A->rmap->n]));
2864: PetscCall(PetscObjectStateIncrease((PetscObject)B));
2865: PetscCall(MatSeqAIJRestoreArrayRead(A, &aa));
2866: PetscCall(MatSeqAIJRestoreArrayWrite(B, &bb));
2867: } else {
2868: PetscCall(MatCopy_Basic(A, B, str));
2869: }
2870: PetscFunctionReturn(PETSC_SUCCESS);
2871: }
2873: PETSC_INTERN PetscErrorCode MatSeqAIJGetArray_SeqAIJ(Mat A, PetscScalar *array[])
2874: {
2875: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
2877: PetscFunctionBegin;
2878: *array = a->a;
2879: PetscFunctionReturn(PETSC_SUCCESS);
2880: }
2882: PETSC_INTERN PetscErrorCode MatSeqAIJRestoreArray_SeqAIJ(Mat A, PetscScalar *array[])
2883: {
2884: PetscFunctionBegin;
2885: *array = NULL;
2886: PetscFunctionReturn(PETSC_SUCCESS);
2887: }
2889: /*
2890: Computes the number of nonzeros per row needed for preallocation when X and Y
2891: have different nonzero structure.
2892: */
2893: PetscErrorCode MatAXPYGetPreallocation_SeqX_private(PetscInt m, const PetscInt *xi, const PetscInt *xj, const PetscInt *yi, const PetscInt *yj, PetscInt *nnz)
2894: {
2895: PetscInt i, j, k, nzx, nzy;
2897: PetscFunctionBegin;
2898: /* Set the number of nonzeros in the new matrix */
2899: for (i = 0; i < m; i++) {
2900: const PetscInt *xjj = PetscSafePointerPlusOffset(xj, xi[i]), *yjj = PetscSafePointerPlusOffset(yj, yi[i]);
2901: nzx = xi[i + 1] - xi[i];
2902: nzy = yi[i + 1] - yi[i];
2903: nnz[i] = 0;
2904: for (j = 0, k = 0; j < nzx; j++) { /* Point in X */
2905: for (; k < nzy && yjj[k] < xjj[j]; k++) nnz[i]++; /* Catch up to X */
2906: if (k < nzy && yjj[k] == xjj[j]) k++; /* Skip duplicate */
2907: nnz[i]++;
2908: }
2909: for (; k < nzy; k++) nnz[i]++;
2910: }
2911: PetscFunctionReturn(PETSC_SUCCESS);
2912: }
2914: PetscErrorCode MatAXPYGetPreallocation_SeqAIJ(Mat Y, Mat X, PetscInt *nnz)
2915: {
2916: PetscInt m = Y->rmap->N;
2917: Mat_SeqAIJ *x = (Mat_SeqAIJ *)X->data;
2918: Mat_SeqAIJ *y = (Mat_SeqAIJ *)Y->data;
2920: PetscFunctionBegin;
2921: /* Set the number of nonzeros in the new matrix */
2922: PetscCall(MatAXPYGetPreallocation_SeqX_private(m, x->i, x->j, y->i, y->j, nnz));
2923: PetscFunctionReturn(PETSC_SUCCESS);
2924: }
2926: PetscErrorCode MatAXPY_SeqAIJ(Mat Y, PetscScalar a, Mat X, MatStructure str)
2927: {
2928: Mat_SeqAIJ *x = (Mat_SeqAIJ *)X->data, *y = (Mat_SeqAIJ *)Y->data;
2930: PetscFunctionBegin;
2931: if (str == UNKNOWN_NONZERO_PATTERN || (PetscDefined(USE_DEBUG) && str == SAME_NONZERO_PATTERN)) {
2932: PetscBool e = x->nz == y->nz ? PETSC_TRUE : PETSC_FALSE;
2933: if (e) {
2934: PetscCall(PetscArraycmp(x->i, y->i, Y->rmap->n + 1, &e));
2935: if (e) {
2936: PetscCall(PetscArraycmp(x->j, y->j, y->nz, &e));
2937: if (e) str = SAME_NONZERO_PATTERN;
2938: }
2939: }
2940: if (!e) PetscCheck(str != SAME_NONZERO_PATTERN, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "MatStructure is not SAME_NONZERO_PATTERN");
2941: }
2942: if (str == SAME_NONZERO_PATTERN) {
2943: const PetscScalar *xa;
2944: PetscScalar *ya, alpha = a;
2945: PetscBLASInt one = 1, bnz;
2947: PetscCall(PetscBLASIntCast(x->nz, &bnz));
2948: PetscCall(MatSeqAIJGetArray(Y, &ya));
2949: PetscCall(MatSeqAIJGetArrayRead(X, &xa));
2950: PetscCallBLAS("BLASaxpy", BLASaxpy_(&bnz, &alpha, xa, &one, ya, &one));
2951: PetscCall(MatSeqAIJRestoreArrayRead(X, &xa));
2952: PetscCall(MatSeqAIJRestoreArray(Y, &ya));
2953: PetscCall(PetscLogFlops(2.0 * bnz));
2954: PetscCall(PetscObjectStateIncrease((PetscObject)Y));
2955: } else if (str == SUBSET_NONZERO_PATTERN) { /* nonzeros of X is a subset of Y's */
2956: PetscCall(MatAXPY_Basic(Y, a, X, str));
2957: } else {
2958: Mat B;
2959: PetscInt *nnz;
2960: PetscCall(PetscMalloc1(Y->rmap->N, &nnz));
2961: PetscCall(MatCreate(PetscObjectComm((PetscObject)Y), &B));
2962: PetscCall(PetscObjectSetName((PetscObject)B, ((PetscObject)Y)->name));
2963: PetscCall(MatSetLayouts(B, Y->rmap, Y->cmap));
2964: PetscCall(MatSetType(B, ((PetscObject)Y)->type_name));
2965: PetscCall(MatAXPYGetPreallocation_SeqAIJ(Y, X, nnz));
2966: PetscCall(MatSeqAIJSetPreallocation(B, 0, nnz));
2967: PetscCall(MatAXPY_BasicWithPreallocation(B, Y, a, X, str));
2968: PetscCall(MatHeaderMerge(Y, &B));
2969: PetscCall(MatSeqAIJCheckInode(Y));
2970: PetscCall(PetscFree(nnz));
2971: }
2972: PetscFunctionReturn(PETSC_SUCCESS);
2973: }
2975: PETSC_INTERN PetscErrorCode MatConjugate_SeqAIJ(Mat mat)
2976: {
2977: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)mat->data;
2978: PetscInt i, nz = aij->nz;
2979: PetscScalar *a;
2981: PetscFunctionBegin;
2982: PetscCall(MatSeqAIJGetArray(mat, &a));
2983: for (i = 0; i < nz; i++) a[i] = PetscConj(a[i]);
2984: PetscCall(MatSeqAIJRestoreArray(mat, &a));
2985: PetscFunctionReturn(PETSC_SUCCESS);
2986: }
2988: static PetscErrorCode MatGetRowMaxAbs_SeqAIJ(Mat A, Vec v, PetscInt idx[])
2989: {
2990: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
2991: PetscInt i, j, m = A->rmap->n, *ai, *aj, ncols, n;
2992: PetscReal atmp;
2993: PetscScalar *x;
2994: const MatScalar *aa, *av;
2996: PetscFunctionBegin;
2997: PetscCheck(!A->factortype, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Not for factored matrix");
2998: PetscCall(MatSeqAIJGetArrayRead(A, &av));
2999: aa = av;
3000: ai = a->i;
3001: aj = a->j;
3003: PetscCall(VecGetArrayWrite(v, &x));
3004: PetscCall(VecGetLocalSize(v, &n));
3005: PetscCheck(n == A->rmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Nonconforming matrix and vector");
3006: for (i = 0; i < m; i++) {
3007: ncols = ai[1] - ai[0];
3008: ai++;
3009: x[i] = 0;
3010: for (j = 0; j < ncols; j++) {
3011: atmp = PetscAbsScalar(*aa);
3012: if (PetscAbsScalar(x[i]) < atmp) {
3013: x[i] = atmp;
3014: if (idx) idx[i] = *aj;
3015: }
3016: aa++;
3017: aj++;
3018: }
3019: }
3020: PetscCall(VecRestoreArrayWrite(v, &x));
3021: PetscCall(MatSeqAIJRestoreArrayRead(A, &av));
3022: PetscFunctionReturn(PETSC_SUCCESS);
3023: }
3025: static PetscErrorCode MatGetRowSumAbs_SeqAIJ(Mat A, Vec v)
3026: {
3027: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
3028: PetscInt i, j, m = A->rmap->n, *ai, ncols, n;
3029: PetscScalar *x;
3030: const MatScalar *aa, *av;
3032: PetscFunctionBegin;
3033: PetscCheck(!A->factortype, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Not for factored matrix");
3034: PetscCall(MatSeqAIJGetArrayRead(A, &av));
3035: aa = av;
3036: ai = a->i;
3038: PetscCall(VecGetArrayWrite(v, &x));
3039: PetscCall(VecGetLocalSize(v, &n));
3040: PetscCheck(n == A->rmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Nonconforming matrix and vector");
3041: for (i = 0; i < m; i++) {
3042: ncols = ai[1] - ai[0];
3043: ai++;
3044: x[i] = 0;
3045: for (j = 0; j < ncols; j++) {
3046: x[i] += PetscAbsScalar(*aa);
3047: aa++;
3048: }
3049: }
3050: PetscCall(VecRestoreArrayWrite(v, &x));
3051: PetscCall(MatSeqAIJRestoreArrayRead(A, &av));
3052: PetscFunctionReturn(PETSC_SUCCESS);
3053: }
3055: static PetscErrorCode MatGetRowMax_SeqAIJ(Mat A, Vec v, PetscInt idx[])
3056: {
3057: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
3058: PetscInt i, j, m = A->rmap->n, *ai, *aj, ncols, n;
3059: PetscScalar *x;
3060: const MatScalar *aa, *av;
3062: PetscFunctionBegin;
3063: PetscCheck(!A->factortype, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Not for factored matrix");
3064: PetscCall(MatSeqAIJGetArrayRead(A, &av));
3065: aa = av;
3066: ai = a->i;
3067: aj = a->j;
3069: PetscCall(VecGetArrayWrite(v, &x));
3070: PetscCall(VecGetLocalSize(v, &n));
3071: PetscCheck(n == A->rmap->n, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Nonconforming matrix and vector");
3072: for (i = 0; i < m; i++) {
3073: ncols = ai[1] - ai[0];
3074: ai++;
3075: if (ncols == A->cmap->n) { /* row is dense */
3076: x[i] = *aa;
3077: if (idx) idx[i] = 0;
3078: } else { /* row is sparse so already KNOW maximum is 0.0 or higher */
3079: x[i] = 0.0;
3080: if (idx) {
3081: for (j = 0; j < ncols; j++) { /* find first implicit 0.0 in the row */
3082: if (aj[j] > j) {
3083: idx[i] = j;
3084: break;
3085: }
3086: }
3087: /* in case first implicit 0.0 in the row occurs at ncols-th column */
3088: if (j == ncols && j < A->cmap->n) idx[i] = j;
3089: }
3090: }
3091: for (j = 0; j < ncols; j++) {
3092: if (PetscRealPart(x[i]) < PetscRealPart(*aa)) {
3093: x[i] = *aa;
3094: if (idx) idx[i] = *aj;
3095: }
3096: aa++;
3097: aj++;
3098: }
3099: }
3100: PetscCall(VecRestoreArrayWrite(v, &x));
3101: PetscCall(MatSeqAIJRestoreArrayRead(A, &av));
3102: PetscFunctionReturn(PETSC_SUCCESS);
3103: }
3105: static PetscErrorCode MatGetRowMinAbs_SeqAIJ(Mat A, Vec v, PetscInt idx[])
3106: {
3107: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
3108: PetscInt i, j, m = A->rmap->n, *ai, *aj, ncols, n;
3109: PetscScalar *x;
3110: const MatScalar *aa, *av;
3112: PetscFunctionBegin;
3113: PetscCall(MatSeqAIJGetArrayRead(A, &av));
3114: aa = av;
3115: ai = a->i;
3116: aj = a->j;
3118: PetscCall(VecGetArrayWrite(v, &x));
3119: PetscCall(VecGetLocalSize(v, &n));
3120: PetscCheck(n == m, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Nonconforming matrix and vector, %" PetscInt_FMT " vs. %" PetscInt_FMT " rows", m, n);
3121: for (i = 0; i < m; i++) {
3122: ncols = ai[1] - ai[0];
3123: ai++;
3124: if (ncols == A->cmap->n) { /* row is dense */
3125: x[i] = *aa;
3126: if (idx) idx[i] = 0;
3127: } else { /* row is sparse so already KNOW minimum is 0.0 or higher */
3128: x[i] = 0.0;
3129: if (idx) { /* find first implicit 0.0 in the row */
3130: for (j = 0; j < ncols; j++) {
3131: if (aj[j] > j) {
3132: idx[i] = j;
3133: break;
3134: }
3135: }
3136: /* in case first implicit 0.0 in the row occurs at ncols-th column */
3137: if (j == ncols && j < A->cmap->n) idx[i] = j;
3138: }
3139: }
3140: for (j = 0; j < ncols; j++) {
3141: if (PetscAbsScalar(x[i]) > PetscAbsScalar(*aa)) {
3142: x[i] = *aa;
3143: if (idx) idx[i] = *aj;
3144: }
3145: aa++;
3146: aj++;
3147: }
3148: }
3149: PetscCall(VecRestoreArrayWrite(v, &x));
3150: PetscCall(MatSeqAIJRestoreArrayRead(A, &av));
3151: PetscFunctionReturn(PETSC_SUCCESS);
3152: }
3154: static PetscErrorCode MatGetRowMin_SeqAIJ(Mat A, Vec v, PetscInt idx[])
3155: {
3156: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
3157: PetscInt i, j, m = A->rmap->n, ncols, n;
3158: const PetscInt *ai, *aj;
3159: PetscScalar *x;
3160: const MatScalar *aa, *av;
3162: PetscFunctionBegin;
3163: PetscCheck(!A->factortype, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Not for factored matrix");
3164: PetscCall(MatSeqAIJGetArrayRead(A, &av));
3165: aa = av;
3166: ai = a->i;
3167: aj = a->j;
3169: PetscCall(VecGetArrayWrite(v, &x));
3170: PetscCall(VecGetLocalSize(v, &n));
3171: PetscCheck(n == m, PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "Nonconforming matrix and vector");
3172: for (i = 0; i < m; i++) {
3173: ncols = ai[1] - ai[0];
3174: ai++;
3175: if (ncols == A->cmap->n) { /* row is dense */
3176: x[i] = *aa;
3177: if (idx) idx[i] = 0;
3178: } else { /* row is sparse so already KNOW minimum is 0.0 or lower */
3179: x[i] = 0.0;
3180: if (idx) { /* find first implicit 0.0 in the row */
3181: for (j = 0; j < ncols; j++) {
3182: if (aj[j] > j) {
3183: idx[i] = j;
3184: break;
3185: }
3186: }
3187: /* in case first implicit 0.0 in the row occurs at ncols-th column */
3188: if (j == ncols && j < A->cmap->n) idx[i] = j;
3189: }
3190: }
3191: for (j = 0; j < ncols; j++) {
3192: if (PetscRealPart(x[i]) > PetscRealPart(*aa)) {
3193: x[i] = *aa;
3194: if (idx) idx[i] = *aj;
3195: }
3196: aa++;
3197: aj++;
3198: }
3199: }
3200: PetscCall(VecRestoreArrayWrite(v, &x));
3201: PetscCall(MatSeqAIJRestoreArrayRead(A, &av));
3202: PetscFunctionReturn(PETSC_SUCCESS);
3203: }
3205: static PetscErrorCode MatInvertBlockDiagonal_SeqAIJ(Mat A, const PetscScalar **values)
3206: {
3207: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
3208: PetscInt i, bs = A->rmap->bs, mbs = A->rmap->n / bs, ipvt[5], bs2 = bs * bs, *v_pivots, ij[7], *IJ, j;
3209: MatScalar *diag, work[25], *v_work;
3210: const PetscReal shift = 0.0;
3211: PetscBool allowzeropivot, zeropivotdetected = PETSC_FALSE;
3213: PetscFunctionBegin;
3214: allowzeropivot = PetscNot(A->erroriffailure);
3215: if (a->ibdiagvalid) {
3216: if (values) *values = a->ibdiag;
3217: PetscFunctionReturn(PETSC_SUCCESS);
3218: }
3219: if (!a->ibdiag) PetscCall(PetscMalloc1(bs2 * mbs, &a->ibdiag));
3220: diag = a->ibdiag;
3221: if (values) *values = a->ibdiag;
3222: /* factor and invert each block */
3223: switch (bs) {
3224: case 1:
3225: for (i = 0; i < mbs; i++) {
3226: PetscCall(MatGetValues(A, 1, &i, 1, &i, diag + i));
3227: if (PetscAbsScalar(diag[i] + shift) < PETSC_MACHINE_EPSILON) {
3228: 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);
3229: A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
3230: A->factorerror_zeropivot_value = PetscAbsScalar(diag[i]);
3231: A->factorerror_zeropivot_row = i;
3232: PetscCall(PetscInfo(A, "Zero pivot, row %" PetscInt_FMT " pivot %g tolerance %g\n", i, (double)PetscAbsScalar(diag[i]), (double)PETSC_MACHINE_EPSILON));
3233: }
3234: diag[i] = (PetscScalar)1.0 / (diag[i] + shift);
3235: }
3236: break;
3237: case 2:
3238: for (i = 0; i < mbs; i++) {
3239: ij[0] = 2 * i;
3240: ij[1] = 2 * i + 1;
3241: PetscCall(MatGetValues(A, 2, ij, 2, ij, diag));
3242: PetscCall(PetscKernel_A_gets_inverse_A_2(diag, shift, allowzeropivot, &zeropivotdetected));
3243: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
3244: PetscCall(PetscKernel_A_gets_transpose_A_2(diag));
3245: diag += 4;
3246: }
3247: break;
3248: case 3:
3249: for (i = 0; i < mbs; i++) {
3250: ij[0] = 3 * i;
3251: ij[1] = 3 * i + 1;
3252: ij[2] = 3 * i + 2;
3253: PetscCall(MatGetValues(A, 3, ij, 3, ij, diag));
3254: PetscCall(PetscKernel_A_gets_inverse_A_3(diag, shift, allowzeropivot, &zeropivotdetected));
3255: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
3256: PetscCall(PetscKernel_A_gets_transpose_A_3(diag));
3257: diag += 9;
3258: }
3259: break;
3260: case 4:
3261: for (i = 0; i < mbs; i++) {
3262: ij[0] = 4 * i;
3263: ij[1] = 4 * i + 1;
3264: ij[2] = 4 * i + 2;
3265: ij[3] = 4 * i + 3;
3266: PetscCall(MatGetValues(A, 4, ij, 4, ij, diag));
3267: PetscCall(PetscKernel_A_gets_inverse_A_4(diag, shift, allowzeropivot, &zeropivotdetected));
3268: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
3269: PetscCall(PetscKernel_A_gets_transpose_A_4(diag));
3270: diag += 16;
3271: }
3272: break;
3273: case 5:
3274: for (i = 0; i < mbs; i++) {
3275: ij[0] = 5 * i;
3276: ij[1] = 5 * i + 1;
3277: ij[2] = 5 * i + 2;
3278: ij[3] = 5 * i + 3;
3279: ij[4] = 5 * i + 4;
3280: PetscCall(MatGetValues(A, 5, ij, 5, ij, diag));
3281: PetscCall(PetscKernel_A_gets_inverse_A_5(diag, ipvt, work, shift, allowzeropivot, &zeropivotdetected));
3282: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
3283: PetscCall(PetscKernel_A_gets_transpose_A_5(diag));
3284: diag += 25;
3285: }
3286: break;
3287: case 6:
3288: for (i = 0; i < mbs; i++) {
3289: ij[0] = 6 * i;
3290: ij[1] = 6 * i + 1;
3291: ij[2] = 6 * i + 2;
3292: ij[3] = 6 * i + 3;
3293: ij[4] = 6 * i + 4;
3294: ij[5] = 6 * i + 5;
3295: PetscCall(MatGetValues(A, 6, ij, 6, ij, diag));
3296: PetscCall(PetscKernel_A_gets_inverse_A_6(diag, shift, allowzeropivot, &zeropivotdetected));
3297: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
3298: PetscCall(PetscKernel_A_gets_transpose_A_6(diag));
3299: diag += 36;
3300: }
3301: break;
3302: case 7:
3303: for (i = 0; i < mbs; i++) {
3304: ij[0] = 7 * i;
3305: ij[1] = 7 * i + 1;
3306: ij[2] = 7 * i + 2;
3307: ij[3] = 7 * i + 3;
3308: ij[4] = 7 * i + 4;
3309: ij[5] = 7 * i + 5;
3310: ij[6] = 7 * i + 6;
3311: PetscCall(MatGetValues(A, 7, ij, 7, ij, diag));
3312: PetscCall(PetscKernel_A_gets_inverse_A_7(diag, shift, allowzeropivot, &zeropivotdetected));
3313: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
3314: PetscCall(PetscKernel_A_gets_transpose_A_7(diag));
3315: diag += 49;
3316: }
3317: break;
3318: default:
3319: PetscCall(PetscMalloc3(bs, &v_work, bs, &v_pivots, bs, &IJ));
3320: for (i = 0; i < mbs; i++) {
3321: for (j = 0; j < bs; j++) IJ[j] = bs * i + j;
3322: PetscCall(MatGetValues(A, bs, IJ, bs, IJ, diag));
3323: PetscCall(PetscKernel_A_gets_inverse_A(bs, diag, v_pivots, v_work, allowzeropivot, &zeropivotdetected));
3324: if (zeropivotdetected) A->factorerrortype = MAT_FACTOR_NUMERIC_ZEROPIVOT;
3325: PetscCall(PetscKernel_A_gets_transpose_A_N(diag, bs));
3326: diag += bs2;
3327: }
3328: PetscCall(PetscFree3(v_work, v_pivots, IJ));
3329: }
3330: a->ibdiagvalid = PETSC_TRUE;
3331: PetscFunctionReturn(PETSC_SUCCESS);
3332: }
3334: static PetscErrorCode MatSetRandom_SeqAIJ(Mat x, PetscRandom rctx)
3335: {
3336: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)x->data;
3337: PetscScalar a, *aa;
3338: PetscInt m, n, i, j, col;
3340: PetscFunctionBegin;
3341: if (!x->assembled) {
3342: PetscCall(MatGetSize(x, &m, &n));
3343: for (i = 0; i < m; i++) {
3344: for (j = 0; j < aij->imax[i]; j++) {
3345: PetscCall(PetscRandomGetValue(rctx, &a));
3346: col = (PetscInt)(n * PetscRealPart(a));
3347: PetscCall(MatSetValues(x, 1, &i, 1, &col, &a, ADD_VALUES));
3348: }
3349: }
3350: } else {
3351: PetscCall(MatSeqAIJGetArrayWrite(x, &aa));
3352: for (i = 0; i < aij->nz; i++) PetscCall(PetscRandomGetValue(rctx, aa + i));
3353: PetscCall(MatSeqAIJRestoreArrayWrite(x, &aa));
3354: }
3355: PetscCall(MatAssemblyBegin(x, MAT_FINAL_ASSEMBLY));
3356: PetscCall(MatAssemblyEnd(x, MAT_FINAL_ASSEMBLY));
3357: PetscFunctionReturn(PETSC_SUCCESS);
3358: }
3360: /* Like MatSetRandom_SeqAIJ, but do not set values on columns in range of [low, high) */
3361: PetscErrorCode MatSetRandomSkipColumnRange_SeqAIJ_Private(Mat x, PetscInt low, PetscInt high, PetscRandom rctx)
3362: {
3363: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)x->data;
3364: PetscScalar a;
3365: PetscInt m, n, i, j, col, nskip;
3367: PetscFunctionBegin;
3368: nskip = high - low;
3369: PetscCall(MatGetSize(x, &m, &n));
3370: n -= nskip; /* shrink number of columns where nonzeros can be set */
3371: for (i = 0; i < m; i++) {
3372: for (j = 0; j < aij->imax[i]; j++) {
3373: PetscCall(PetscRandomGetValue(rctx, &a));
3374: col = (PetscInt)(n * PetscRealPart(a));
3375: if (col >= low) col += nskip; /* shift col rightward to skip the hole */
3376: PetscCall(MatSetValues(x, 1, &i, 1, &col, &a, ADD_VALUES));
3377: }
3378: }
3379: PetscCall(MatAssemblyBegin(x, MAT_FINAL_ASSEMBLY));
3380: PetscCall(MatAssemblyEnd(x, MAT_FINAL_ASSEMBLY));
3381: PetscFunctionReturn(PETSC_SUCCESS);
3382: }
3384: static struct _MatOps MatOps_Values = {MatSetValues_SeqAIJ,
3385: MatGetRow_SeqAIJ,
3386: MatRestoreRow_SeqAIJ,
3387: MatMult_SeqAIJ,
3388: /* 4*/ MatMultAdd_SeqAIJ,
3389: MatMultTranspose_SeqAIJ,
3390: MatMultTransposeAdd_SeqAIJ,
3391: NULL,
3392: NULL,
3393: NULL,
3394: /* 10*/ NULL,
3395: MatLUFactor_SeqAIJ,
3396: NULL,
3397: MatSOR_SeqAIJ,
3398: MatTranspose_SeqAIJ,
3399: /* 15*/ MatGetInfo_SeqAIJ,
3400: MatEqual_SeqAIJ,
3401: MatGetDiagonal_SeqAIJ,
3402: MatDiagonalScale_SeqAIJ,
3403: MatNorm_SeqAIJ,
3404: /* 20*/ NULL,
3405: MatAssemblyEnd_SeqAIJ,
3406: MatSetOption_SeqAIJ,
3407: MatZeroEntries_SeqAIJ,
3408: /* 24*/ MatZeroRows_SeqAIJ,
3409: NULL,
3410: NULL,
3411: NULL,
3412: NULL,
3413: /* 29*/ MatSetUp_Seq_Hash,
3414: NULL,
3415: NULL,
3416: NULL,
3417: NULL,
3418: /* 34*/ MatDuplicate_SeqAIJ,
3419: NULL,
3420: NULL,
3421: MatILUFactor_SeqAIJ,
3422: NULL,
3423: /* 39*/ MatAXPY_SeqAIJ,
3424: MatCreateSubMatrices_SeqAIJ,
3425: MatIncreaseOverlap_SeqAIJ,
3426: MatGetValues_SeqAIJ,
3427: MatCopy_SeqAIJ,
3428: /* 44*/ MatGetRowMax_SeqAIJ,
3429: MatScale_SeqAIJ,
3430: MatShift_SeqAIJ,
3431: MatDiagonalSet_SeqAIJ,
3432: MatZeroRowsColumns_SeqAIJ,
3433: /* 49*/ MatSetRandom_SeqAIJ,
3434: MatGetRowIJ_SeqAIJ,
3435: MatRestoreRowIJ_SeqAIJ,
3436: MatGetColumnIJ_SeqAIJ,
3437: MatRestoreColumnIJ_SeqAIJ,
3438: /* 54*/ MatFDColoringCreate_SeqXAIJ,
3439: NULL,
3440: NULL,
3441: MatPermute_SeqAIJ,
3442: NULL,
3443: /* 59*/ NULL,
3444: MatDestroy_SeqAIJ,
3445: MatView_SeqAIJ,
3446: NULL,
3447: NULL,
3448: /* 64*/ MatMatMatMultNumeric_SeqAIJ_SeqAIJ_SeqAIJ,
3449: NULL,
3450: NULL,
3451: NULL,
3452: MatGetRowMaxAbs_SeqAIJ,
3453: /* 69*/ MatGetRowMinAbs_SeqAIJ,
3454: NULL,
3455: NULL,
3456: MatFDColoringApply_AIJ,
3457: NULL,
3458: /* 74*/ MatFindZeroDiagonals_SeqAIJ,
3459: NULL,
3460: NULL,
3461: NULL,
3462: MatLoad_SeqAIJ,
3463: /* 79*/ NULL,
3464: NULL,
3465: NULL,
3466: NULL,
3467: NULL,
3468: /* 84*/ NULL,
3469: MatMatMultNumeric_SeqAIJ_SeqAIJ,
3470: MatPtAPNumeric_SeqAIJ_SeqAIJ_SparseAxpy,
3471: NULL,
3472: MatMatTransposeMultNumeric_SeqAIJ_SeqAIJ,
3473: /* 90*/ NULL,
3474: MatProductSetFromOptions_SeqAIJ,
3475: NULL,
3476: NULL,
3477: MatConjugate_SeqAIJ,
3478: /* 94*/ NULL,
3479: MatSetValuesRow_SeqAIJ,
3480: MatRealPart_SeqAIJ,
3481: MatImaginaryPart_SeqAIJ,
3482: NULL,
3483: /* 99*/ NULL,
3484: MatMatSolve_SeqAIJ,
3485: NULL,
3486: MatGetRowMin_SeqAIJ,
3487: NULL,
3488: /*104*/ NULL,
3489: NULL,
3490: NULL,
3491: NULL,
3492: NULL,
3493: /*109*/ NULL,
3494: NULL,
3495: NULL,
3496: NULL,
3497: MatGetMultiProcBlock_SeqAIJ,
3498: /*114*/ MatFindNonzeroRows_SeqAIJ,
3499: MatGetColumnReductions_SeqAIJ,
3500: MatInvertBlockDiagonal_SeqAIJ,
3501: MatInvertVariableBlockDiagonal_SeqAIJ,
3502: NULL,
3503: /*119*/ NULL,
3504: MatTransposeMatMultNumeric_SeqAIJ_SeqAIJ,
3505: MatTransposeColoringCreate_SeqAIJ,
3506: MatTransColoringApplySpToDen_SeqAIJ,
3507: MatTransColoringApplyDenToSp_SeqAIJ,
3508: /*124*/ MatRARtNumeric_SeqAIJ_SeqAIJ,
3509: NULL,
3510: NULL,
3511: MatFDColoringSetUp_SeqXAIJ,
3512: MatFindOffBlockDiagonalEntries_SeqAIJ,
3513: /*129*/ MatCreateMPIMatConcatenateSeqMat_SeqAIJ,
3514: MatDestroySubMatrices_SeqAIJ,
3515: NULL,
3516: NULL,
3517: MatCreateGraph_Simple_AIJ,
3518: /*134*/ MatTransposeSymbolic_SeqAIJ,
3519: MatEliminateZeros_SeqAIJ,
3520: MatGetRowSumAbs_SeqAIJ,
3521: NULL,
3522: NULL,
3523: /*139*/ NULL,
3524: MatCopyHashToXAIJ_Seq_Hash,
3525: NULL,
3526: NULL,
3527: MatADot_Default,
3528: /*144*/ MatANorm_Default,
3529: NULL,
3530: NULL,
3531: NULL};
3533: static PetscErrorCode MatSeqAIJSetColumnIndices_SeqAIJ(Mat mat, PetscInt *indices)
3534: {
3535: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)mat->data;
3536: PetscInt i, nz, n;
3538: PetscFunctionBegin;
3539: nz = aij->maxnz;
3540: n = mat->rmap->n;
3541: for (i = 0; i < nz; i++) aij->j[i] = indices[i];
3542: aij->nz = nz;
3543: for (i = 0; i < n; i++) aij->ilen[i] = aij->imax[i];
3544: PetscFunctionReturn(PETSC_SUCCESS);
3545: }
3547: /*
3548: * Given a sparse matrix with global column indices, compact it by using a local column space.
3549: * The result matrix helps saving memory in other algorithms, such as MatPtAPSymbolic_MPIAIJ_MPIAIJ_scalable()
3550: */
3551: PetscErrorCode MatSeqAIJCompactOutExtraColumns_SeqAIJ(Mat mat, ISLocalToGlobalMapping *mapping)
3552: {
3553: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)mat->data;
3554: PetscHMapI gid1_lid1;
3555: PetscHashIter tpos;
3556: PetscInt gid, lid, i, ec, nz = aij->nz;
3557: PetscInt *garray, *jj = aij->j;
3559: PetscFunctionBegin;
3561: PetscAssertPointer(mapping, 2);
3562: /* use a table */
3563: PetscCall(PetscHMapICreateWithSize(mat->rmap->n, &gid1_lid1));
3564: ec = 0;
3565: for (i = 0; i < nz; i++) {
3566: PetscInt data, gid1 = jj[i] + 1;
3567: PetscCall(PetscHMapIGetWithDefault(gid1_lid1, gid1, 0, &data));
3568: if (!data) {
3569: /* one based table */
3570: PetscCall(PetscHMapISet(gid1_lid1, gid1, ++ec));
3571: }
3572: }
3573: /* form array of columns we need */
3574: PetscCall(PetscMalloc1(ec, &garray));
3575: PetscHashIterBegin(gid1_lid1, tpos);
3576: while (!PetscHashIterAtEnd(gid1_lid1, tpos)) {
3577: PetscHashIterGetKey(gid1_lid1, tpos, gid);
3578: PetscHashIterGetVal(gid1_lid1, tpos, lid);
3579: PetscHashIterNext(gid1_lid1, tpos);
3580: gid--;
3581: lid--;
3582: garray[lid] = gid;
3583: }
3584: PetscCall(PetscSortInt(ec, garray)); /* sort, and rebuild */
3585: PetscCall(PetscHMapIClear(gid1_lid1));
3586: for (i = 0; i < ec; i++) PetscCall(PetscHMapISet(gid1_lid1, garray[i] + 1, i + 1));
3587: /* compact out the extra columns in B */
3588: for (i = 0; i < nz; i++) {
3589: PetscInt gid1 = jj[i] + 1;
3590: PetscCall(PetscHMapIGetWithDefault(gid1_lid1, gid1, 0, &lid));
3591: lid--;
3592: jj[i] = lid;
3593: }
3594: PetscCall(PetscLayoutDestroy(&mat->cmap));
3595: PetscCall(PetscHMapIDestroy(&gid1_lid1));
3596: PetscCall(PetscLayoutCreateFromSizes(PetscObjectComm((PetscObject)mat), ec, ec, 1, &mat->cmap));
3597: PetscCall(ISLocalToGlobalMappingCreate(PETSC_COMM_SELF, mat->cmap->bs, mat->cmap->n, garray, PETSC_OWN_POINTER, mapping));
3598: PetscCall(ISLocalToGlobalMappingSetType(*mapping, ISLOCALTOGLOBALMAPPINGHASH));
3599: PetscFunctionReturn(PETSC_SUCCESS);
3600: }
3602: /*@
3603: MatSeqAIJSetColumnIndices - Set the column indices for all the rows
3604: in the matrix.
3606: Input Parameters:
3607: + mat - the `MATSEQAIJ` matrix
3608: - indices - the column indices
3610: Level: advanced
3612: Notes:
3613: This can be called if you have precomputed the nonzero structure of the
3614: matrix and want to provide it to the matrix object to improve the performance
3615: of the `MatSetValues()` operation.
3617: You MUST have set the correct numbers of nonzeros per row in the call to
3618: `MatCreateSeqAIJ()`, and the columns indices MUST be sorted.
3620: MUST be called before any calls to `MatSetValues()`
3622: The indices should start with zero, not one.
3624: .seealso: [](ch_matrices), `Mat`, `MATSEQAIJ`
3625: @*/
3626: PetscErrorCode MatSeqAIJSetColumnIndices(Mat mat, PetscInt *indices)
3627: {
3628: PetscFunctionBegin;
3630: PetscAssertPointer(indices, 2);
3631: PetscUseMethod(mat, "MatSeqAIJSetColumnIndices_C", (Mat, PetscInt *), (mat, indices));
3632: PetscFunctionReturn(PETSC_SUCCESS);
3633: }
3635: static PetscErrorCode MatStoreValues_SeqAIJ(Mat mat)
3636: {
3637: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)mat->data;
3638: size_t nz = aij->i[mat->rmap->n];
3640: PetscFunctionBegin;
3641: PetscCheck(aij->nonew, PETSC_COMM_SELF, PETSC_ERR_ORDER, "Must call MatSetOption(A,MAT_NEW_NONZERO_LOCATIONS,PETSC_FALSE);first");
3643: /* allocate space for values if not already there */
3644: if (!aij->saved_values) PetscCall(PetscMalloc1(nz + 1, &aij->saved_values));
3646: /* copy values over */
3647: PetscCall(PetscArraycpy(aij->saved_values, aij->a, nz));
3648: PetscFunctionReturn(PETSC_SUCCESS);
3649: }
3651: /*@
3652: MatStoreValues - Stashes a copy of the matrix values; this allows reusing of the linear part of a Jacobian, while recomputing only the
3653: nonlinear portion.
3655: Logically Collect
3657: Input Parameter:
3658: . mat - the matrix (currently only `MATAIJ` matrices support this option)
3660: Level: advanced
3662: Example Usage:
3663: .vb
3664: Using SNES
3665: Create Jacobian matrix
3666: Set linear terms into matrix
3667: Apply boundary conditions to matrix, at this time matrix must have
3668: final nonzero structure (i.e. setting the nonlinear terms and applying
3669: boundary conditions again will not change the nonzero structure
3670: MatSetOption(mat, MAT_NEW_NONZERO_LOCATIONS, PETSC_FALSE);
3671: MatStoreValues(mat);
3672: Call SNESSetJacobian() with matrix
3673: In your Jacobian routine
3674: MatRetrieveValues(mat);
3675: Set nonlinear terms in matrix
3677: Without `SNESSolve()`, i.e. when you handle nonlinear solve yourself:
3678: // build linear portion of Jacobian
3679: MatSetOption(mat, MAT_NEW_NONZERO_LOCATIONS, PETSC_FALSE);
3680: MatStoreValues(mat);
3681: loop over nonlinear iterations
3682: MatRetrieveValues(mat);
3683: // call MatSetValues(mat,...) to set nonliner portion of Jacobian
3684: // call MatAssemblyBegin/End() on matrix
3685: Solve linear system with Jacobian
3686: endloop
3687: .ve
3689: Notes:
3690: Matrix must already be assembled before calling this routine
3691: Must set the matrix option `MatSetOption`(mat,`MAT_NEW_NONZERO_LOCATIONS`,`PETSC_FALSE`); before
3692: calling this routine.
3694: When this is called multiple times it overwrites the previous set of stored values
3695: and does not allocated additional space.
3697: .seealso: [](ch_matrices), `Mat`, `MatRetrieveValues()`
3698: @*/
3699: PetscErrorCode MatStoreValues(Mat mat)
3700: {
3701: PetscFunctionBegin;
3703: PetscCheck(mat->assembled, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Not for unassembled matrix");
3704: PetscCheck(!mat->factortype, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Not for factored matrix");
3705: PetscUseMethod(mat, "MatStoreValues_C", (Mat), (mat));
3706: PetscFunctionReturn(PETSC_SUCCESS);
3707: }
3709: static PetscErrorCode MatRetrieveValues_SeqAIJ(Mat mat)
3710: {
3711: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)mat->data;
3712: PetscInt nz = aij->i[mat->rmap->n];
3714: PetscFunctionBegin;
3715: PetscCheck(aij->nonew, PETSC_COMM_SELF, PETSC_ERR_ORDER, "Must call MatSetOption(A,MAT_NEW_NONZERO_LOCATIONS,PETSC_FALSE);first");
3716: PetscCheck(aij->saved_values, PETSC_COMM_SELF, PETSC_ERR_ORDER, "Must call MatStoreValues(A);first");
3717: /* copy values over */
3718: PetscCall(PetscArraycpy(aij->a, aij->saved_values, nz));
3719: PetscFunctionReturn(PETSC_SUCCESS);
3720: }
3722: /*@
3723: MatRetrieveValues - Retrieves the copy of the matrix values that was stored with `MatStoreValues()`
3725: Logically Collect
3727: Input Parameter:
3728: . mat - the matrix (currently only `MATAIJ` matrices support this option)
3730: Level: advanced
3732: .seealso: [](ch_matrices), `Mat`, `MatStoreValues()`
3733: @*/
3734: PetscErrorCode MatRetrieveValues(Mat mat)
3735: {
3736: PetscFunctionBegin;
3738: PetscCheck(mat->assembled, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Not for unassembled matrix");
3739: PetscCheck(!mat->factortype, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Not for factored matrix");
3740: PetscUseMethod(mat, "MatRetrieveValues_C", (Mat), (mat));
3741: PetscFunctionReturn(PETSC_SUCCESS);
3742: }
3744: /*@
3745: MatCreateSeqAIJ - Creates a sparse matrix in `MATSEQAIJ` (compressed row) format
3746: (the default parallel PETSc format). For good matrix assembly performance
3747: the user should preallocate the matrix storage by setting the parameter `nz`
3748: (or the array `nnz`).
3750: Collective
3752: Input Parameters:
3753: + comm - MPI communicator, set to `PETSC_COMM_SELF`
3754: . m - number of rows
3755: . n - number of columns
3756: . nz - number of nonzeros per row (same for all rows)
3757: - nnz - array containing the number of nonzeros in the various rows
3758: (possibly different for each row) or NULL
3760: Output Parameter:
3761: . A - the matrix
3763: Options Database Keys:
3764: + -mat_no_inode - Do not use inodes
3765: - -mat_inode_limit limit - Sets inode limit (max limit=5)
3767: Level: intermediate
3769: Notes:
3770: It is recommend to use `MatCreateFromOptions()` instead of this routine
3772: If `nnz` is given then `nz` is ignored
3774: The `MATSEQAIJ` format, also called
3775: compressed row storage, is fully compatible with standard Fortran
3776: storage. That is, the stored row and column indices can begin at
3777: either one (as in Fortran) or zero.
3779: Specify the preallocated storage with either `nz` or `nnz` (not both).
3780: Set `nz` = `PETSC_DEFAULT` and `nnz` = `NULL` for PETSc to control dynamic memory
3781: allocation.
3783: By default, this format uses inodes (identical nodes) when possible, to
3784: improve numerical efficiency of matrix-vector products and solves. We
3785: search for consecutive rows with the same nonzero structure, thereby
3786: reusing matrix information to achieve increased efficiency.
3788: .seealso: [](ch_matrices), `Mat`, [Sparse Matrix Creation](sec_matsparse), `MatCreate()`, `MatCreateAIJ()`, `MatSetValues()`, `MatSeqAIJSetColumnIndices()`, `MatCreateSeqAIJWithArrays()`
3789: @*/
3790: PetscErrorCode MatCreateSeqAIJ(MPI_Comm comm, PetscInt m, PetscInt n, PetscInt nz, const PetscInt nnz[], Mat *A)
3791: {
3792: PetscFunctionBegin;
3793: PetscCall(MatCreate(comm, A));
3794: PetscCall(MatSetSizes(*A, m, n, m, n));
3795: PetscCall(MatSetType(*A, MATSEQAIJ));
3796: PetscCall(MatSeqAIJSetPreallocation_SeqAIJ(*A, nz, nnz));
3797: PetscFunctionReturn(PETSC_SUCCESS);
3798: }
3800: /*@
3801: MatSeqAIJSetPreallocation - For good matrix assembly performance
3802: the user should preallocate the matrix storage by setting the parameter nz
3803: (or the array nnz). By setting these parameters accurately, performance
3804: during matrix assembly can be increased by more than a factor of 50.
3806: Collective
3808: Input Parameters:
3809: + B - The matrix
3810: . nz - number of nonzeros per row (same for all rows)
3811: - nnz - array containing the number of nonzeros in the various rows
3812: (possibly different for each row) or NULL
3814: Options Database Keys:
3815: + -mat_no_inode - Do not use inodes
3816: - -mat_inode_limit limit - Sets inode limit (max limit=5)
3818: Level: intermediate
3820: Notes:
3821: If `nnz` is given then `nz` is ignored
3823: The `MATSEQAIJ` format also called
3824: compressed row storage, is fully compatible with standard Fortran
3825: storage. That is, the stored row and column indices can begin at
3826: either one (as in Fortran) or zero. See the users' manual for details.
3828: Specify the preallocated storage with either `nz` or `nnz` (not both).
3829: Set nz = `PETSC_DEFAULT` and `nnz` = `NULL` for PETSc to control dynamic memory
3830: allocation.
3832: You can call `MatGetInfo()` to get information on how effective the preallocation was;
3833: for example the fields mallocs,nz_allocated,nz_used,nz_unneeded;
3834: You can also run with the option -info and look for messages with the string
3835: malloc in them to see if additional memory allocation was needed.
3837: Developer Notes:
3838: Use nz of `MAT_SKIP_ALLOCATION` to not allocate any space for the matrix
3839: entries or columns indices
3841: By default, this format uses inodes (identical nodes) when possible, to
3842: improve numerical efficiency of matrix-vector products and solves. We
3843: search for consecutive rows with the same nonzero structure, thereby
3844: reusing matrix information to achieve increased efficiency.
3846: .seealso: [](ch_matrices), `Mat`, `MatCreate()`, `MatCreateAIJ()`, `MatSetValues()`, `MatSeqAIJSetColumnIndices()`, `MatCreateSeqAIJWithArrays()`, `MatGetInfo()`,
3847: `MatSeqAIJSetTotalPreallocation()`
3848: @*/
3849: PetscErrorCode MatSeqAIJSetPreallocation(Mat B, PetscInt nz, const PetscInt nnz[])
3850: {
3851: PetscFunctionBegin;
3854: PetscTryMethod(B, "MatSeqAIJSetPreallocation_C", (Mat, PetscInt, const PetscInt[]), (B, nz, nnz));
3855: PetscFunctionReturn(PETSC_SUCCESS);
3856: }
3858: PetscErrorCode MatSeqAIJSetPreallocation_SeqAIJ(Mat B, PetscInt nz, const PetscInt *nnz)
3859: {
3860: Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data;
3861: PetscBool skipallocation = PETSC_FALSE, realalloc = PETSC_FALSE;
3862: PetscInt i;
3864: PetscFunctionBegin;
3865: if (B->hash_active) {
3866: B->ops[0] = b->cops;
3867: PetscCall(PetscHMapIJVDestroy(&b->ht));
3868: PetscCall(PetscFree(b->dnz));
3869: B->hash_active = PETSC_FALSE;
3870: }
3871: if (nz >= 0 || nnz) realalloc = PETSC_TRUE;
3872: if (nz == MAT_SKIP_ALLOCATION) {
3873: skipallocation = PETSC_TRUE;
3874: nz = 0;
3875: }
3876: PetscCall(PetscLayoutSetUp(B->rmap));
3877: PetscCall(PetscLayoutSetUp(B->cmap));
3879: if (nz == PETSC_DEFAULT || nz == PETSC_DECIDE) nz = 5;
3880: PetscCheck(nz >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "nz cannot be less than 0: value %" PetscInt_FMT, nz);
3881: if (nnz) {
3882: for (i = 0; i < B->rmap->n; i++) {
3883: 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]);
3884: 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);
3885: }
3886: }
3888: B->preallocated = PETSC_TRUE;
3889: if (!skipallocation) {
3890: if (!b->imax) PetscCall(PetscMalloc1(B->rmap->n, &b->imax));
3891: if (!b->ilen) {
3892: /* b->ilen will count nonzeros in each row so far. */
3893: PetscCall(PetscCalloc1(B->rmap->n, &b->ilen));
3894: } else {
3895: PetscCall(PetscMemzero(b->ilen, B->rmap->n * sizeof(PetscInt)));
3896: }
3897: if (!b->ipre) PetscCall(PetscMalloc1(B->rmap->n, &b->ipre));
3898: if (!nnz) {
3899: if (nz == PETSC_DEFAULT || nz == PETSC_DECIDE) nz = 10;
3900: else if (nz < 0) nz = 1;
3901: nz = PetscMin(nz, B->cmap->n);
3902: for (i = 0; i < B->rmap->n; i++) b->imax[i] = nz;
3903: PetscCall(PetscIntMultError(nz, B->rmap->n, &nz));
3904: } else {
3905: PetscInt64 nz64 = 0;
3906: for (i = 0; i < B->rmap->n; i++) {
3907: b->imax[i] = nnz[i];
3908: nz64 += nnz[i];
3909: }
3910: PetscCall(PetscIntCast(nz64, &nz));
3911: }
3913: /* allocate the matrix space */
3914: PetscCall(MatSeqXAIJFreeAIJ(B, &b->a, &b->j, &b->i));
3915: PetscCall(PetscShmgetAllocateArray(nz, sizeof(PetscInt), (void **)&b->j));
3916: PetscCall(PetscShmgetAllocateArray(B->rmap->n + 1, sizeof(PetscInt), (void **)&b->i));
3917: b->free_ij = PETSC_TRUE;
3918: if (B->structure_only) {
3919: b->free_a = PETSC_FALSE;
3920: } else {
3921: PetscCall(PetscShmgetAllocateArray(nz, sizeof(PetscScalar), (void **)&b->a));
3922: b->free_a = PETSC_TRUE;
3923: }
3924: b->i[0] = 0;
3925: for (i = 1; i < B->rmap->n + 1; i++) b->i[i] = b->i[i - 1] + b->imax[i - 1];
3926: } else {
3927: b->free_a = PETSC_FALSE;
3928: b->free_ij = PETSC_FALSE;
3929: }
3931: if (b->ipre && nnz != b->ipre && b->imax) {
3932: /* reserve user-requested sparsity */
3933: PetscCall(PetscArraycpy(b->ipre, b->imax, B->rmap->n));
3934: }
3936: b->nz = 0;
3937: b->maxnz = nz;
3938: B->info.nz_unneeded = (double)b->maxnz;
3939: if (realalloc) PetscCall(MatSetOption(B, MAT_NEW_NONZERO_ALLOCATION_ERR, PETSC_TRUE));
3940: B->was_assembled = PETSC_FALSE;
3941: B->assembled = PETSC_FALSE;
3942: /* We simply deem preallocation has changed nonzero state. Updating the state
3943: will give clients (like AIJKokkos) a chance to know something has happened.
3944: */
3945: B->nonzerostate++;
3946: PetscFunctionReturn(PETSC_SUCCESS);
3947: }
3949: PetscErrorCode MatResetPreallocation_SeqAIJ_Private(Mat A, PetscBool *memoryreset)
3950: {
3951: Mat_SeqAIJ *a;
3952: PetscInt i;
3953: PetscBool skipreset;
3955: PetscFunctionBegin;
3958: 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()");
3959: if (A->num_ass == 0) PetscFunctionReturn(PETSC_SUCCESS);
3961: /* Check local size. If zero, then return */
3962: if (!A->rmap->n) PetscFunctionReturn(PETSC_SUCCESS);
3964: a = (Mat_SeqAIJ *)A->data;
3965: /* if no saved info, we error out */
3966: PetscCheck(a->ipre, PETSC_COMM_SELF, PETSC_ERR_ARG_NULL, "No saved preallocation info ");
3968: PetscCheck(a->i && a->imax && a->ilen, PETSC_COMM_SELF, PETSC_ERR_ARG_NULL, "Memory info is incomplete, and cannot reset preallocation ");
3970: PetscCall(PetscArraycmp(a->ipre, a->ilen, A->rmap->n, &skipreset));
3971: if (skipreset) PetscCall(MatZeroEntries(A));
3972: else {
3973: PetscCall(PetscArraycpy(a->imax, a->ipre, A->rmap->n));
3974: PetscCall(PetscArrayzero(a->ilen, A->rmap->n));
3975: a->i[0] = 0;
3976: for (i = 1; i < A->rmap->n + 1; i++) a->i[i] = a->i[i - 1] + a->imax[i - 1];
3977: A->preallocated = PETSC_TRUE;
3978: a->nz = 0;
3979: a->maxnz = a->i[A->rmap->n];
3980: A->info.nz_unneeded = (double)a->maxnz;
3981: A->was_assembled = PETSC_FALSE;
3982: A->assembled = PETSC_FALSE;
3983: A->nonzerostate++;
3984: /* Log that the state of this object has changed; this will help guarantee that preconditioners get re-setup */
3985: PetscCall(PetscObjectStateIncrease((PetscObject)A));
3986: }
3987: if (memoryreset) *memoryreset = (PetscBool)!skipreset;
3988: PetscFunctionReturn(PETSC_SUCCESS);
3989: }
3991: static PetscErrorCode MatResetPreallocation_SeqAIJ(Mat A)
3992: {
3993: PetscFunctionBegin;
3994: PetscCall(MatResetPreallocation_SeqAIJ_Private(A, NULL));
3995: PetscFunctionReturn(PETSC_SUCCESS);
3996: }
3998: /*@
3999: MatSeqAIJSetPreallocationCSR - Allocates memory for a sparse sequential matrix in `MATSEQAIJ` format.
4001: Input Parameters:
4002: + B - the matrix
4003: . i - the indices into `j` for the start of each row (indices start with zero)
4004: . j - the column indices for each row (indices start with zero) these must be sorted for each row
4005: - v - optional values in the matrix, use `NULL` if not provided
4007: Level: developer
4009: Notes:
4010: The `i`,`j`,`v` values are COPIED with this routine; to avoid the copy use `MatCreateSeqAIJWithArrays()`
4012: This routine may be called multiple times with different nonzero patterns (or the same nonzero pattern). The nonzero
4013: structure will be the union of all the previous nonzero structures.
4015: Developer Notes:
4016: 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
4017: then just copies the `v` values directly with `PetscMemcpy()`.
4019: This routine could also take a `PetscCopyMode` argument to allow sharing the values instead of always copying them.
4021: .seealso: [](ch_matrices), `Mat`, `MatCreate()`, `MatCreateSeqAIJ()`, `MatSetValues()`, `MatSeqAIJSetPreallocation()`, `MATSEQAIJ`, `MatResetPreallocation()`
4022: @*/
4023: PetscErrorCode MatSeqAIJSetPreallocationCSR(Mat B, const PetscInt i[], const PetscInt j[], const PetscScalar v[])
4024: {
4025: PetscFunctionBegin;
4028: PetscTryMethod(B, "MatSeqAIJSetPreallocationCSR_C", (Mat, const PetscInt[], const PetscInt[], const PetscScalar[]), (B, i, j, v));
4029: PetscFunctionReturn(PETSC_SUCCESS);
4030: }
4032: static PetscErrorCode MatSeqAIJSetPreallocationCSR_SeqAIJ(Mat B, const PetscInt Ii[], const PetscInt J[], const PetscScalar v[])
4033: {
4034: PetscInt i;
4035: PetscInt m, n;
4036: PetscInt nz;
4037: PetscInt *nnz;
4039: PetscFunctionBegin;
4040: PetscCheck(Ii[0] == 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Ii[0] must be 0 it is %" PetscInt_FMT, Ii[0]);
4042: PetscCall(PetscLayoutSetUp(B->rmap));
4043: PetscCall(PetscLayoutSetUp(B->cmap));
4045: PetscCall(MatGetSize(B, &m, &n));
4046: PetscCall(PetscMalloc1(m + 1, &nnz));
4047: for (i = 0; i < m; i++) {
4048: nz = Ii[i + 1] - Ii[i];
4049: PetscCheck(nz >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Local row %" PetscInt_FMT " has a negative number of columns %" PetscInt_FMT, i, nz);
4050: nnz[i] = nz;
4051: }
4052: PetscCall(MatSeqAIJSetPreallocation(B, 0, nnz));
4053: PetscCall(PetscFree(nnz));
4055: 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));
4057: PetscCall(MatAssemblyBegin(B, MAT_FINAL_ASSEMBLY));
4058: PetscCall(MatAssemblyEnd(B, MAT_FINAL_ASSEMBLY));
4060: PetscCall(MatSetOption(B, MAT_NEW_NONZERO_LOCATION_ERR, PETSC_TRUE));
4061: PetscFunctionReturn(PETSC_SUCCESS);
4062: }
4064: /*@
4065: MatSeqAIJKron - Computes `C`, the Kronecker product of `A` and `B`.
4067: Input Parameters:
4068: + A - left-hand side matrix
4069: . B - right-hand side matrix
4070: - reuse - either `MAT_INITIAL_MATRIX` or `MAT_REUSE_MATRIX`
4072: Output Parameter:
4073: . C - Kronecker product of `A` and `B`
4075: Level: intermediate
4077: Note:
4078: `MAT_REUSE_MATRIX` can only be used when the nonzero structure of the product matrix has not changed from that last call to `MatSeqAIJKron()`.
4080: .seealso: [](ch_matrices), `Mat`, `MatCreateSeqAIJ()`, `MATSEQAIJ`, `MATKAIJ`, `MatReuse`
4081: @*/
4082: PetscErrorCode MatSeqAIJKron(Mat A, Mat B, MatReuse reuse, Mat *C)
4083: {
4084: PetscFunctionBegin;
4089: PetscAssertPointer(C, 4);
4090: if (reuse == MAT_REUSE_MATRIX) {
4093: }
4094: PetscTryMethod(A, "MatSeqAIJKron_C", (Mat, Mat, MatReuse, Mat *), (A, B, reuse, C));
4095: PetscFunctionReturn(PETSC_SUCCESS);
4096: }
4098: static PetscErrorCode MatSeqAIJKron_SeqAIJ(Mat A, Mat B, MatReuse reuse, Mat *C)
4099: {
4100: Mat newmat;
4101: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
4102: Mat_SeqAIJ *b = (Mat_SeqAIJ *)B->data;
4103: PetscScalar *v;
4104: const PetscScalar *aa, *ba;
4105: 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;
4106: PetscBool flg;
4108: PetscFunctionBegin;
4109: PetscCheck(!A->factortype, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Not for factored matrix");
4110: PetscCheck(A->assembled, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Not for unassembled matrix");
4111: PetscCheck(!B->factortype, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Not for factored matrix");
4112: PetscCheck(B->assembled, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Not for unassembled matrix");
4113: PetscCall(PetscObjectTypeCompare((PetscObject)B, MATSEQAIJ, &flg));
4114: PetscCheck(flg, PETSC_COMM_SELF, PETSC_ERR_SUP, "MatType %s", ((PetscObject)B)->type_name);
4115: PetscCheck(reuse == MAT_INITIAL_MATRIX || reuse == MAT_REUSE_MATRIX, PETSC_COMM_SELF, PETSC_ERR_SUP, "MatReuse %d", (int)reuse);
4116: if (reuse == MAT_INITIAL_MATRIX) {
4117: PetscCall(PetscMalloc2(am * bm + 1, &i, a->i[am] * b->i[bm], &j));
4118: PetscCall(MatCreate(PETSC_COMM_SELF, &newmat));
4119: PetscCall(MatSetSizes(newmat, am * bm, an * bn, am * bm, an * bn));
4120: PetscCall(MatSetType(newmat, MATAIJ));
4121: i[0] = 0;
4122: for (m = 0; m < am; ++m) {
4123: for (p = 0; p < bm; ++p) {
4124: i[m * bm + p + 1] = i[m * bm + p] + (a->i[m + 1] - a->i[m]) * (b->i[p + 1] - b->i[p]);
4125: for (n = a->i[m]; n < a->i[m + 1]; ++n) {
4126: for (q = b->i[p]; q < b->i[p + 1]; ++q) j[nnz++] = a->j[n] * bn + b->j[q];
4127: }
4128: }
4129: }
4130: PetscCall(MatSeqAIJSetPreallocationCSR(newmat, i, j, NULL));
4131: *C = newmat;
4132: PetscCall(PetscFree2(i, j));
4133: nnz = 0;
4134: }
4135: PetscCall(MatSeqAIJGetArray(*C, &v));
4136: PetscCall(MatSeqAIJGetArrayRead(A, &aa));
4137: PetscCall(MatSeqAIJGetArrayRead(B, &ba));
4138: for (m = 0; m < am; ++m) {
4139: for (p = 0; p < bm; ++p) {
4140: for (n = a->i[m]; n < a->i[m + 1]; ++n) {
4141: for (q = b->i[p]; q < b->i[p + 1]; ++q) v[nnz++] = aa[n] * ba[q];
4142: }
4143: }
4144: }
4145: PetscCall(MatSeqAIJRestoreArray(*C, &v));
4146: PetscCall(MatSeqAIJRestoreArrayRead(A, &aa));
4147: PetscCall(MatSeqAIJRestoreArrayRead(B, &ba));
4148: PetscFunctionReturn(PETSC_SUCCESS);
4149: }
4151: #include <../src/mat/impls/dense/seq/dense.h>
4152: #include <petsc/private/kernels/petscaxpy.h>
4154: /*
4155: Computes (B'*A')' since computing B*A directly is untenable
4157: n p p
4158: [ ] [ ] [ ]
4159: m [ A ] * n [ B ] = m [ C ]
4160: [ ] [ ] [ ]
4162: */
4163: PetscErrorCode MatMatMultNumeric_SeqDense_SeqAIJ(Mat A, Mat B, Mat C)
4164: {
4165: Mat_SeqDense *sub_a = (Mat_SeqDense *)A->data;
4166: Mat_SeqAIJ *sub_b = (Mat_SeqAIJ *)B->data;
4167: Mat_SeqDense *sub_c = (Mat_SeqDense *)C->data;
4168: PetscInt i, j, n, m, q, p;
4169: const PetscInt *ii, *idx;
4170: const PetscScalar *b, *a, *a_q;
4171: PetscScalar *c, *c_q;
4172: PetscInt clda = sub_c->lda;
4173: PetscInt alda = sub_a->lda;
4175: PetscFunctionBegin;
4176: m = A->rmap->n;
4177: n = A->cmap->n;
4178: p = B->cmap->n;
4179: a = sub_a->v;
4180: b = sub_b->a;
4181: c = sub_c->v;
4182: if (clda == m) {
4183: PetscCall(PetscArrayzero(c, m * p));
4184: } else {
4185: for (j = 0; j < p; j++)
4186: for (i = 0; i < m; i++) c[j * clda + i] = 0.0;
4187: }
4188: ii = sub_b->i;
4189: idx = sub_b->j;
4190: for (i = 0; i < n; i++) {
4191: q = ii[i + 1] - ii[i];
4192: while (q-- > 0) {
4193: c_q = c + clda * (*idx);
4194: a_q = a + alda * i;
4195: PetscKernelAXPY(c_q, *b, a_q, m);
4196: idx++;
4197: b++;
4198: }
4199: }
4200: PetscFunctionReturn(PETSC_SUCCESS);
4201: }
4203: PetscErrorCode MatMatMultSymbolic_SeqDense_SeqAIJ(Mat A, Mat B, PetscReal fill, Mat C)
4204: {
4205: PetscInt m = A->rmap->n, n = B->cmap->n;
4206: PetscBool cisdense;
4208: PetscFunctionBegin;
4209: 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);
4210: PetscCall(MatSetSizes(C, m, n, m, n));
4211: PetscCall(MatSetBlockSizesFromMats(C, A, B));
4212: PetscCall(PetscObjectTypeCompareAny((PetscObject)C, &cisdense, MATSEQDENSE, MATSEQDENSECUDA, MATSEQDENSEHIP, ""));
4213: if (!cisdense) PetscCall(MatSetType(C, MATDENSE));
4214: PetscCall(MatSetUp(C));
4216: C->ops->matmultnumeric = MatMatMultNumeric_SeqDense_SeqAIJ;
4217: PetscFunctionReturn(PETSC_SUCCESS);
4218: }
4220: /*MC
4221: MATSEQAIJ - MATSEQAIJ = "seqaij" - A matrix type to be used for sequential sparse matrices,
4222: based on compressed sparse row format.
4224: Options Database Key:
4225: . -mat_type seqaij - sets the matrix type to "seqaij" during a call to MatSetFromOptions()
4227: Level: beginner
4229: Notes:
4230: `MatSetValues()` may be called for this matrix type with a `NULL` argument for the numerical values,
4231: in this case the values associated with the rows and columns one passes in are set to zero
4232: in the matrix
4234: `MatSetOptions`(,`MAT_STRUCTURE_ONLY`,`PETSC_TRUE`) may be called for this matrix type. In this no
4235: space is allocated for the nonzero entries and any entries passed with `MatSetValues()` are ignored
4237: Developer Note:
4238: It would be nice if all matrix formats supported passing `NULL` in for the numerical values
4240: .seealso: [](ch_matrices), `Mat`, `MatCreateSeqAIJ()`, `MatSetFromOptions()`, `MatSetType()`, `MatCreate()`, `MatType`, `MATSELL`, `MATSEQSELL`, `MATMPISELL`
4241: M*/
4243: /*MC
4244: MATAIJ - MATAIJ = "aij" - A matrix type to be used for sparse matrices.
4246: This matrix type is identical to `MATSEQAIJ` when constructed with a single process communicator,
4247: and `MATMPIAIJ` otherwise. As a result, for single process communicators,
4248: `MatSeqAIJSetPreallocation()` is supported, and similarly `MatMPIAIJSetPreallocation()` is supported
4249: for communicators controlling multiple processes. It is recommended that you call both of
4250: the above preallocation routines for simplicity.
4252: Options Database Key:
4253: . -mat_type aij - sets the matrix type to "aij" during a call to `MatSetFromOptions()`
4255: Level: beginner
4257: Note:
4258: Subclasses include `MATAIJCUSPARSE`, `MATAIJPERM`, `MATAIJSELL`, `MATAIJMKL`, `MATAIJCRL`, and also automatically switches over to use inodes when
4259: enough exist.
4261: .seealso: [](ch_matrices), `Mat`, `MatCreateAIJ()`, `MatCreateSeqAIJ()`, `MATSEQAIJ`, `MATMPIAIJ`, `MATSELL`, `MATSEQSELL`, `MATMPISELL`
4262: M*/
4264: PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqAIJCRL(Mat, MatType, MatReuse, Mat *);
4265: #if PetscDefined(HAVE_ELEMENTAL)
4266: PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_Elemental(Mat, MatType, MatReuse, Mat *);
4267: #endif
4268: #if PetscDefined(HAVE_SCALAPACK) && (PetscDefined(USE_REAL_SINGLE) || PetscDefined(USE_REAL_DOUBLE))
4269: PETSC_INTERN PetscErrorCode MatConvert_AIJ_ScaLAPACK(Mat, MatType, MatReuse, Mat *);
4270: #endif
4271: #if PetscDefined(HAVE_HYPRE)
4272: PETSC_INTERN PetscErrorCode MatConvert_AIJ_HYPRE(Mat A, MatType, MatReuse, Mat *);
4273: #endif
4275: PETSC_EXTERN PetscErrorCode MatConvert_SeqAIJ_SeqSELL(Mat, MatType, MatReuse, Mat *);
4276: PETSC_INTERN PetscErrorCode MatConvert_XAIJ_IS(Mat, MatType, MatReuse, Mat *);
4277: PETSC_INTERN PetscErrorCode MatProductSetFromOptions_IS_XAIJ(Mat);
4279: /*@C
4280: MatSeqAIJGetArray - gives read/write access to the array where the data for a `MATSEQAIJ` matrix is stored
4282: Not Collective
4284: Input Parameter:
4285: . A - a `MATSEQAIJ` matrix
4287: Output Parameter:
4288: . array - pointer to the data
4290: Level: intermediate
4292: .seealso: [](ch_matrices), `Mat`, `MatSeqAIJRestoreArray()`
4293: @*/
4294: PetscErrorCode MatSeqAIJGetArray(Mat A, PetscScalar *array[])
4295: {
4296: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)A->data;
4298: PetscFunctionBegin;
4299: if (aij->ops->getarray) {
4300: PetscCall((*aij->ops->getarray)(A, array));
4301: } else {
4302: *array = aij->a;
4303: }
4304: PetscFunctionReturn(PETSC_SUCCESS);
4305: }
4307: /*@C
4308: MatSeqAIJRestoreArray - returns access to the array where the data for a `MATSEQAIJ` matrix is stored obtained by `MatSeqAIJGetArray()`
4310: Not Collective
4312: Input Parameters:
4313: + A - a `MATSEQAIJ` matrix
4314: - array - pointer to the data
4316: Level: intermediate
4318: .seealso: [](ch_matrices), `Mat`, `MatSeqAIJGetArray()`
4319: @*/
4320: PetscErrorCode MatSeqAIJRestoreArray(Mat A, PetscScalar *array[])
4321: {
4322: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)A->data;
4324: PetscFunctionBegin;
4325: if (aij->ops->restorearray) {
4326: PetscCall((*aij->ops->restorearray)(A, array));
4327: } else {
4328: *array = NULL;
4329: }
4330: PetscCall(PetscObjectStateIncrease((PetscObject)A));
4331: PetscFunctionReturn(PETSC_SUCCESS);
4332: }
4334: /*@C
4335: MatSeqAIJGetArrayRead - gives read-only access to the array where the data for a `MATSEQAIJ` matrix is stored
4337: Not Collective
4339: Input Parameter:
4340: . A - a `MATSEQAIJ` matrix
4342: Output Parameter:
4343: . array - pointer to the data
4345: Level: intermediate
4347: .seealso: [](ch_matrices), `Mat`, `MatSeqAIJGetArray()`, `MatSeqAIJRestoreArrayRead()`
4348: @*/
4349: PetscErrorCode MatSeqAIJGetArrayRead(Mat A, const PetscScalar *array[])
4350: {
4351: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)A->data;
4353: PetscFunctionBegin;
4354: if (aij->ops->getarrayread) {
4355: PetscCall((*aij->ops->getarrayread)(A, array));
4356: } else {
4357: *array = aij->a;
4358: }
4359: PetscFunctionReturn(PETSC_SUCCESS);
4360: }
4362: /*@C
4363: MatSeqAIJRestoreArrayRead - restore the read-only access array obtained from `MatSeqAIJGetArrayRead()`
4365: Not Collective
4367: Input Parameter:
4368: . A - a `MATSEQAIJ` matrix
4370: Output Parameter:
4371: . array - pointer to the data
4373: Level: intermediate
4375: .seealso: [](ch_matrices), `Mat`, `MatSeqAIJGetArray()`, `MatSeqAIJGetArrayRead()`
4376: @*/
4377: PetscErrorCode MatSeqAIJRestoreArrayRead(Mat A, const PetscScalar *array[])
4378: {
4379: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)A->data;
4381: PetscFunctionBegin;
4382: if (aij->ops->restorearrayread) {
4383: PetscCall((*aij->ops->restorearrayread)(A, array));
4384: } else {
4385: *array = NULL;
4386: }
4387: PetscFunctionReturn(PETSC_SUCCESS);
4388: }
4390: /*@C
4391: MatSeqAIJGetArrayWrite - gives write-only access to the array where the data for a `MATSEQAIJ` matrix is stored
4393: Not Collective
4395: Input Parameter:
4396: . A - a `MATSEQAIJ` matrix
4398: Output Parameter:
4399: . array - pointer to the data
4401: Level: intermediate
4403: .seealso: [](ch_matrices), `Mat`, `MatSeqAIJGetArray()`, `MatSeqAIJRestoreArrayWrite()`
4404: @*/
4405: PetscErrorCode MatSeqAIJGetArrayWrite(Mat A, PetscScalar *array[])
4406: {
4407: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)A->data;
4409: PetscFunctionBegin;
4410: if (aij->ops->getarraywrite) {
4411: PetscCall((*aij->ops->getarraywrite)(A, array));
4412: } else {
4413: *array = aij->a;
4414: }
4415: PetscCall(PetscObjectStateIncrease((PetscObject)A));
4416: PetscFunctionReturn(PETSC_SUCCESS);
4417: }
4419: /*@C
4420: MatSeqAIJRestoreArrayWrite - restore the write-only access array obtained from `MatSeqAIJGetArrayWrite()`
4422: Not Collective
4424: Input Parameter:
4425: . A - a `MATSEQAIJ` matrix
4427: Output Parameter:
4428: . array - pointer to the data
4430: Level: intermediate
4432: .seealso: [](ch_matrices), `Mat`, `MatSeqAIJGetArray()`, `MatSeqAIJGetArrayWrite()`
4433: @*/
4434: PetscErrorCode MatSeqAIJRestoreArrayWrite(Mat A, PetscScalar *array[])
4435: {
4436: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)A->data;
4438: PetscFunctionBegin;
4439: if (aij->ops->restorearraywrite) {
4440: PetscCall((*aij->ops->restorearraywrite)(A, array));
4441: } else {
4442: *array = NULL;
4443: }
4444: PetscFunctionReturn(PETSC_SUCCESS);
4445: }
4447: /*@C
4448: MatSeqAIJGetCSRAndMemType - Get the CSR arrays and the memory type of the `MATSEQAIJ` matrix
4450: Not Collective; No Fortran Support
4452: Input Parameter:
4453: . mat - a matrix of type `MATSEQAIJ` or its subclasses
4455: Output Parameters:
4456: + i - row map array of the matrix
4457: . j - column index array of the matrix
4458: . a - data array of the matrix
4459: - mtype - memory type of the arrays
4461: Level: developer
4463: Notes:
4464: Any of the output parameters can be `NULL`, in which case the corresponding value is not returned.
4465: If mat is a device matrix, the arrays are on the device. Otherwise, they are on the host.
4467: One can call this routine on a preallocated but not assembled matrix to just get the memory of the CSR underneath the matrix.
4468: If the matrix is assembled, the data array `a` is guaranteed to have the latest values of the matrix.
4470: .seealso: [](ch_matrices), `Mat`, `MatSeqAIJGetArray()`, `MatSeqAIJGetArrayRead()`
4471: @*/
4472: PetscErrorCode MatSeqAIJGetCSRAndMemType(Mat mat, const PetscInt *i[], const PetscInt *j[], PetscScalar *a[], PetscMemType *mtype)
4473: {
4474: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)mat->data;
4476: PetscFunctionBegin;
4477: PetscCheck(mat->preallocated, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "matrix is not preallocated");
4478: if (aij->ops->getcsrandmemtype) {
4479: PetscCall((*aij->ops->getcsrandmemtype)(mat, i, j, a, mtype));
4480: } else {
4481: if (i) *i = aij->i;
4482: if (j) *j = aij->j;
4483: if (a) *a = aij->a;
4484: if (mtype) *mtype = PETSC_MEMTYPE_HOST;
4485: }
4486: PetscFunctionReturn(PETSC_SUCCESS);
4487: }
4489: /*@
4490: MatSeqAIJGetMaxRowNonzeros - returns the maximum number of nonzeros in any row
4492: Not Collective
4494: Input Parameter:
4495: . A - a `MATSEQAIJ` matrix
4497: Output Parameter:
4498: . nz - the maximum number of nonzeros in any row
4500: Level: intermediate
4502: .seealso: [](ch_matrices), `Mat`, `MatSeqAIJRestoreArray()`
4503: @*/
4504: PetscErrorCode MatSeqAIJGetMaxRowNonzeros(Mat A, PetscInt *nz)
4505: {
4506: Mat_SeqAIJ *aij = (Mat_SeqAIJ *)A->data;
4508: PetscFunctionBegin;
4509: *nz = aij->rmax;
4510: PetscFunctionReturn(PETSC_SUCCESS);
4511: }
4513: static PetscErrorCode MatCOOStructDestroy_SeqAIJ(PetscCtxRt data)
4514: {
4515: MatCOOStruct_SeqAIJ *coo = *(MatCOOStruct_SeqAIJ **)data;
4517: PetscFunctionBegin;
4518: PetscCall(PetscFree(coo->perm));
4519: PetscCall(PetscFree(coo->jmap));
4520: PetscCall(PetscFree(coo));
4521: PetscFunctionReturn(PETSC_SUCCESS);
4522: }
4524: PetscErrorCode MatSetPreallocationCOO_SeqAIJ(Mat mat, PetscCount coo_n, PetscInt coo_i[], PetscInt coo_j[])
4525: {
4526: MPI_Comm comm;
4527: PetscInt *i, *j;
4528: PetscInt M, N, row, iprev;
4529: PetscCount k, p, q, nneg, nnz, start, end; /* Index the coo array, so use PetscCount as their type */
4530: PetscInt *Ai; /* Change to PetscCount once we use it for row pointers */
4531: PetscInt *Aj;
4532: PetscScalar *Aa;
4533: Mat_SeqAIJ *seqaij = (Mat_SeqAIJ *)mat->data;
4534: MatType rtype;
4535: PetscCount *perm, *jmap;
4536: MatCOOStruct_SeqAIJ *coo;
4537: PetscBool isorted;
4538: PetscBool hypre;
4540: PetscFunctionBegin;
4541: PetscCall(PetscObjectGetComm((PetscObject)mat, &comm));
4542: PetscCall(MatGetSize(mat, &M, &N));
4543: i = coo_i;
4544: j = coo_j;
4545: PetscCall(PetscMalloc1(coo_n, &perm));
4547: /* 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) */
4548: isorted = PETSC_TRUE;
4549: iprev = PETSC_INT_MIN;
4550: for (k = 0; k < coo_n; k++) {
4551: if (j[k] < 0) i[k] = -1;
4552: if (isorted) {
4553: if (i[k] < iprev) isorted = PETSC_FALSE;
4554: else iprev = i[k];
4555: }
4556: perm[k] = k;
4557: }
4559: /* Sort by row if not already */
4560: if (!isorted) PetscCall(PetscSortIntWithIntCountArrayPair(coo_n, i, j, perm));
4561: 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);
4563: /* Advance k to the first row with a non-negative index */
4564: for (k = 0; k < coo_n; k++)
4565: if (i[k] >= 0) break;
4566: nneg = k;
4567: 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 */
4568: nnz = 0; /* Total number of unique nonzeros to be counted */
4569: jmap++; /* Inc jmap by 1 for convenience */
4571: PetscCall(PetscShmgetAllocateArray(M + 1, sizeof(PetscInt), (void **)&Ai)); /* CSR of A */
4572: PetscCall(PetscArrayzero(Ai, M + 1));
4573: PetscCall(PetscShmgetAllocateArray(coo_n - nneg, sizeof(PetscInt), (void **)&Aj)); /* We have at most coo_n-nneg unique nonzeros */
4575: PetscCall(PetscStrcmp("_internal_COO_mat_for_hypre", ((PetscObject)mat)->name, &hypre));
4577: /* In each row, sort by column, then unique column indices to get row length */
4578: Ai++; /* Inc by 1 for convenience */
4579: q = 0; /* q-th unique nonzero, with q starting from 0 */
4580: while (k < coo_n) {
4581: PetscBool strictly_sorted; // this row is strictly sorted?
4582: PetscInt jprev;
4584: /* get [start,end) indices for this row; also check if cols in this row are strictly sorted */
4585: row = i[k];
4586: start = k;
4587: jprev = PETSC_INT_MIN;
4588: strictly_sorted = PETSC_TRUE;
4589: while (k < coo_n && i[k] == row) {
4590: if (strictly_sorted) {
4591: if (j[k] <= jprev) strictly_sorted = PETSC_FALSE;
4592: else jprev = j[k];
4593: }
4594: k++;
4595: }
4596: end = k;
4598: /* hack for HYPRE: swap min column to diag so that diagonal values will go first */
4599: if (hypre) {
4600: PetscInt minj = PETSC_INT_MAX;
4601: PetscBool hasdiag = PETSC_FALSE;
4603: if (strictly_sorted) { // fast path to swap the first and the diag
4604: PetscCount tmp;
4605: for (p = start; p < end; p++) {
4606: if (j[p] == row && p != start) {
4607: j[p] = j[start]; // swap j[], so that the diagonal value will go first (manipulated by perm[])
4608: j[start] = row;
4609: tmp = perm[start];
4610: perm[start] = perm[p]; // also swap perm[] so we can save the call to PetscSortIntWithCountArray() below
4611: perm[p] = tmp;
4612: break;
4613: }
4614: }
4615: } else {
4616: for (p = start; p < end; p++) {
4617: hasdiag = (PetscBool)(hasdiag || (j[p] == row));
4618: minj = PetscMin(minj, j[p]);
4619: }
4621: if (hasdiag) {
4622: for (p = start; p < end; p++) {
4623: if (j[p] == minj) j[p] = row;
4624: else if (j[p] == row) j[p] = minj;
4625: }
4626: }
4627: }
4628: }
4629: // sort by columns in a row. perm[] indicates their original order
4630: if (!strictly_sorted) PetscCall(PetscSortIntWithCountArray(end - start, j + start, perm + start));
4631: 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);
4633: if (strictly_sorted) { // fast path to set Aj[], jmap[], Ai[], nnz, q
4634: for (p = start; p < end; p++, q++) {
4635: Aj[q] = j[p];
4636: jmap[q] = 1;
4637: }
4638: PetscCall(PetscIntCast(end - start, Ai + row));
4639: nnz += Ai[row]; // q is already advanced
4640: } else {
4641: /* Find number of unique col entries in this row */
4642: Aj[q] = j[start]; /* Log the first nonzero in this row */
4643: jmap[q] = 1; /* Number of repeats of this nonzero entry */
4644: Ai[row] = 1;
4645: nnz++;
4647: for (p = start + 1; p < end; p++) { /* Scan remaining nonzero in this row */
4648: if (j[p] != j[p - 1]) { /* Meet a new nonzero */
4649: q++;
4650: jmap[q] = 1;
4651: Aj[q] = j[p];
4652: Ai[row]++;
4653: nnz++;
4654: } else {
4655: jmap[q]++;
4656: }
4657: }
4658: q++; /* Move to next row and thus next unique nonzero */
4659: }
4660: }
4662: Ai--; /* Back to the beginning of Ai[] */
4663: for (k = 0; k < M; k++) Ai[k + 1] += Ai[k];
4664: jmap--; // Back to the beginning of jmap[]
4665: jmap[0] = 0;
4666: for (k = 0; k < nnz; k++) jmap[k + 1] += jmap[k];
4668: if (nnz < coo_n - nneg) { /* Reallocate with actual number of unique nonzeros */
4669: PetscCount *jmap_new;
4670: PetscInt *Aj_new;
4672: PetscCall(PetscMalloc1(nnz + 1, &jmap_new));
4673: PetscCall(PetscArraycpy(jmap_new, jmap, nnz + 1));
4674: PetscCall(PetscFree(jmap));
4675: jmap = jmap_new;
4677: PetscCall(PetscShmgetAllocateArray(nnz, sizeof(PetscInt), (void **)&Aj_new));
4678: PetscCall(PetscArraycpy(Aj_new, Aj, nnz));
4679: PetscCall(PetscShmgetDeallocateArray((void **)&Aj));
4680: Aj = Aj_new;
4681: }
4683: if (nneg) { /* Discard heading entries with negative indices in perm[], as we'll access it from index 0 in MatSetValuesCOO */
4684: PetscCount *perm_new;
4686: PetscCall(PetscMalloc1(coo_n - nneg, &perm_new));
4687: PetscCall(PetscArraycpy(perm_new, perm + nneg, coo_n - nneg));
4688: PetscCall(PetscFree(perm));
4689: perm = perm_new;
4690: }
4692: PetscCall(MatGetRootType_Private(mat, &rtype));
4693: PetscCall(PetscShmgetAllocateArray(nnz, sizeof(PetscScalar), (void **)&Aa));
4694: PetscCall(PetscArrayzero(Aa, nnz));
4695: PetscCall(MatSetSeqAIJWithArrays_private(PETSC_COMM_SELF, M, N, Ai, Aj, Aa, rtype, mat));
4697: seqaij->free_a = seqaij->free_ij = PETSC_TRUE; /* Let newmat own Ai, Aj and Aa */
4699: // Put the COO struct in a container and then attach that to the matrix
4700: PetscCall(PetscMalloc1(1, &coo));
4701: PetscCall(PetscIntCast(nnz, &coo->nz));
4702: coo->n = coo_n;
4703: coo->Atot = coo_n - nneg; // Annz is seqaij->nz, so no need to record that again
4704: coo->jmap = jmap; // of length nnz+1
4705: coo->perm = perm;
4706: PetscCall(PetscObjectContainerCompose((PetscObject)mat, "__PETSc_MatCOOStruct_Host", coo, MatCOOStructDestroy_SeqAIJ));
4707: PetscFunctionReturn(PETSC_SUCCESS);
4708: }
4710: static PetscErrorCode MatSetValuesCOO_SeqAIJ(Mat A, const PetscScalar v[], InsertMode imode)
4711: {
4712: Mat_SeqAIJ *aseq = (Mat_SeqAIJ *)A->data;
4713: PetscCount i, j, Annz = aseq->nz;
4714: PetscCount *perm, *jmap;
4715: PetscScalar *Aa;
4716: PetscContainer container;
4717: MatCOOStruct_SeqAIJ *coo;
4719: PetscFunctionBegin;
4720: PetscCall(PetscObjectQuery((PetscObject)A, "__PETSc_MatCOOStruct_Host", (PetscObject *)&container));
4721: PetscCheck(container, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Not found MatCOOStruct on this matrix");
4722: PetscCall(PetscContainerGetPointer(container, &coo));
4723: perm = coo->perm;
4724: jmap = coo->jmap;
4725: PetscCall(MatSeqAIJGetArray(A, &Aa));
4726: for (i = 0; i < Annz; i++) {
4727: PetscScalar sum = 0.0;
4728: for (j = jmap[i]; j < jmap[i + 1]; j++) sum += v[perm[j]];
4729: Aa[i] = (imode == INSERT_VALUES ? 0.0 : Aa[i]) + sum;
4730: }
4731: PetscCall(MatSeqAIJRestoreArray(A, &Aa));
4732: PetscFunctionReturn(PETSC_SUCCESS);
4733: }
4735: #if PetscDefined(HAVE_CUDA)
4736: PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqAIJCUSPARSE(Mat, MatType, MatReuse, Mat *);
4737: #endif
4738: #if PetscDefined(HAVE_HIP)
4739: PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqAIJHIPSPARSE(Mat, MatType, MatReuse, Mat *);
4740: #endif
4741: #if PetscDefined(HAVE_KOKKOS_KERNELS)
4742: PETSC_INTERN PetscErrorCode MatConvert_SeqAIJ_SeqAIJKokkos(Mat, MatType, MatReuse, Mat *);
4743: #endif
4745: PETSC_EXTERN PetscErrorCode MatCreate_SeqAIJ(Mat B)
4746: {
4747: Mat_SeqAIJ *b;
4748: PetscMPIInt size;
4750: PetscFunctionBegin;
4751: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)B), &size));
4752: PetscCheck(size <= 1, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Comm must be of size 1");
4754: PetscCall(PetscNew(&b));
4756: B->data = (void *)b;
4757: B->ops[0] = MatOps_Values;
4758: if (B->sortedfull) B->ops->setvalues = MatSetValues_SeqAIJ_SortedFull;
4760: b->row = NULL;
4761: b->col = NULL;
4762: b->icol = NULL;
4763: b->reallocs = 0;
4764: b->ignorezeroentries = PETSC_FALSE;
4765: b->roworiented = PETSC_TRUE;
4766: b->nonew = 0;
4767: b->diag = NULL;
4768: b->solve_work = NULL;
4769: B->spptr = NULL;
4770: b->saved_values = NULL;
4771: b->idiag = NULL;
4772: b->mdiag = NULL;
4773: b->ssor_work = NULL;
4774: b->omega = 1.0;
4775: b->fshift = 0.0;
4776: b->ibdiagvalid = PETSC_FALSE;
4777: b->keepnonzeropattern = PETSC_FALSE;
4779: PetscCall(PetscObjectChangeTypeName((PetscObject)B, MATSEQAIJ));
4780: #if PetscDefined(HAVE_MATLAB)
4781: PetscCall(PetscObjectComposeFunction((PetscObject)B, "PetscMatlabEnginePut_C", MatlabEnginePut_SeqAIJ));
4782: PetscCall(PetscObjectComposeFunction((PetscObject)B, "PetscMatlabEngineGet_C", MatlabEngineGet_SeqAIJ));
4783: #endif
4784: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatSeqAIJSetColumnIndices_C", MatSeqAIJSetColumnIndices_SeqAIJ));
4785: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatStoreValues_C", MatStoreValues_SeqAIJ));
4786: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatRetrieveValues_C", MatRetrieveValues_SeqAIJ));
4787: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_seqsbaij_C", MatConvert_SeqAIJ_SeqSBAIJ));
4788: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_seqbaij_C", MatConvert_SeqAIJ_SeqBAIJ));
4789: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_seqaijperm_C", MatConvert_SeqAIJ_SeqAIJPERM));
4790: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_seqaijsell_C", MatConvert_SeqAIJ_SeqAIJSELL));
4791: #if PetscDefined(HAVE_MKL_SPARSE)
4792: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_seqaijmkl_C", MatConvert_SeqAIJ_SeqAIJMKL));
4793: #endif
4794: #if PetscDefined(HAVE_CUDA)
4795: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_seqaijcusparse_C", MatConvert_SeqAIJ_SeqAIJCUSPARSE));
4796: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_seqaijcusparse_seqaij_C", MatProductSetFromOptions_SeqAIJ));
4797: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_seqaij_seqaijcusparse_C", MatProductSetFromOptions_SeqAIJ));
4798: #endif
4799: #if PetscDefined(HAVE_HIP)
4800: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_seqaijhipsparse_C", MatConvert_SeqAIJ_SeqAIJHIPSPARSE));
4801: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_seqaijhipsparse_seqaij_C", MatProductSetFromOptions_SeqAIJ));
4802: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_seqaij_seqaijhipsparse_C", MatProductSetFromOptions_SeqAIJ));
4803: #endif
4804: #if PetscDefined(HAVE_KOKKOS_KERNELS)
4805: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_seqaijkokkos_C", MatConvert_SeqAIJ_SeqAIJKokkos));
4806: #endif
4807: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_seqaijcrl_C", MatConvert_SeqAIJ_SeqAIJCRL));
4808: #if PetscDefined(HAVE_ELEMENTAL)
4809: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_elemental_C", MatConvert_SeqAIJ_Elemental));
4810: #endif
4811: #if PetscDefined(HAVE_SCALAPACK) && (PetscDefined(USE_REAL_SINGLE) || PetscDefined(USE_REAL_DOUBLE))
4812: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_scalapack_C", MatConvert_AIJ_ScaLAPACK));
4813: #endif
4814: #if PetscDefined(HAVE_HYPRE)
4815: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_hypre_C", MatConvert_AIJ_HYPRE));
4816: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_transpose_seqaij_seqaij_C", MatProductSetFromOptions_Transpose_AIJ_AIJ));
4817: #endif
4818: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_seqdense_C", MatConvert_SeqAIJ_SeqDense));
4819: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_seqsell_C", MatConvert_SeqAIJ_SeqSELL));
4820: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatConvert_seqaij_is_C", MatConvert_XAIJ_IS));
4821: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatIsTranspose_C", MatIsTranspose_SeqAIJ));
4822: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatIsHermitianTranspose_C", MatIsHermitianTranspose_SeqAIJ));
4823: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatSeqAIJSetPreallocation_C", MatSeqAIJSetPreallocation_SeqAIJ));
4824: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatResetPreallocation_C", MatResetPreallocation_SeqAIJ));
4825: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatResetHash_C", MatResetHash_SeqAIJ));
4826: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatSeqAIJSetPreallocationCSR_C", MatSeqAIJSetPreallocationCSR_SeqAIJ));
4827: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatReorderForNonzeroDiagonal_C", MatReorderForNonzeroDiagonal_SeqAIJ));
4828: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_is_seqaij_C", MatProductSetFromOptions_IS_XAIJ));
4829: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_seqdense_seqaij_C", MatProductSetFromOptions_SeqDense_SeqAIJ));
4830: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatProductSetFromOptions_seqaij_seqaij_C", MatProductSetFromOptions_SeqAIJ));
4831: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatSeqAIJKron_C", MatSeqAIJKron_SeqAIJ));
4832: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatSetPreallocationCOO_C", MatSetPreallocationCOO_SeqAIJ));
4833: PetscCall(PetscObjectComposeFunction((PetscObject)B, "MatSetValuesCOO_C", MatSetValuesCOO_SeqAIJ));
4834: PetscCall(MatCreate_SeqAIJ_Inode(B));
4835: PetscCall(PetscObjectChangeTypeName((PetscObject)B, MATSEQAIJ));
4836: PetscCall(MatSeqAIJSetTypeFromOptions(B)); /* this allows changing the matrix subtype to say MATSEQAIJPERM */
4837: PetscFunctionReturn(PETSC_SUCCESS);
4838: }
4840: /*
4841: Given a matrix generated with MatGetFactor() duplicates all the information in A into C
4842: */
4843: PetscErrorCode MatDuplicateNoCreate_SeqAIJ(Mat C, Mat A, MatDuplicateOption cpvalues, PetscBool mallocmatspace)
4844: {
4845: Mat_SeqAIJ *c = (Mat_SeqAIJ *)C->data, *a = (Mat_SeqAIJ *)A->data;
4846: PetscInt m = A->rmap->n, i;
4848: PetscFunctionBegin;
4849: PetscCheck(A->assembled || cpvalues == MAT_DO_NOT_COPY_VALUES, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Cannot duplicate unassembled matrix");
4851: C->factortype = A->factortype;
4852: c->row = NULL;
4853: c->col = NULL;
4854: c->icol = NULL;
4855: c->reallocs = 0;
4856: C->assembled = A->assembled;
4858: if (A->preallocated) {
4859: PetscCall(PetscLayoutReference(A->rmap, &C->rmap));
4860: PetscCall(PetscLayoutReference(A->cmap, &C->cmap));
4862: if (!A->hash_active) {
4863: PetscCall(PetscMalloc1(m, &c->imax));
4864: PetscCall(PetscArraycpy(c->imax, a->imax, m));
4865: PetscCall(PetscMalloc1(m, &c->ilen));
4866: PetscCall(PetscArraycpy(c->ilen, a->ilen, m));
4868: /* allocate the matrix space */
4869: if (mallocmatspace) {
4870: PetscCall(PetscShmgetAllocateArray(a->i[m], sizeof(PetscScalar), (void **)&c->a));
4871: PetscCall(PetscShmgetAllocateArray(a->i[m], sizeof(PetscInt), (void **)&c->j));
4872: PetscCall(PetscShmgetAllocateArray(m + 1, sizeof(PetscInt), (void **)&c->i));
4873: PetscCall(PetscArraycpy(c->i, a->i, m + 1));
4874: c->free_a = PETSC_TRUE;
4875: c->free_ij = PETSC_TRUE;
4876: if (m > 0) {
4877: PetscCall(PetscArraycpy(c->j, a->j, a->i[m]));
4878: if (cpvalues == MAT_COPY_VALUES) {
4879: const PetscScalar *aa;
4881: PetscCall(MatSeqAIJGetArrayRead(A, &aa));
4882: PetscCall(PetscArraycpy(c->a, aa, a->i[m]));
4883: PetscCall(MatSeqAIJGetArrayRead(A, &aa));
4884: } else {
4885: PetscCall(PetscArrayzero(c->a, a->i[m]));
4886: }
4887: }
4888: }
4889: C->preallocated = PETSC_TRUE;
4890: } else {
4891: PetscCheck(mallocmatspace, PetscObjectComm((PetscObject)A), PETSC_ERR_ARG_WRONGSTATE, "Cannot malloc matrix memory from a non-preallocated matrix");
4892: PetscCall(MatSetUp(C));
4893: }
4895: c->ignorezeroentries = a->ignorezeroentries;
4896: c->roworiented = a->roworiented;
4897: c->nonew = a->nonew;
4898: c->solve_work = NULL;
4899: c->saved_values = NULL;
4900: c->idiag = NULL;
4901: c->ssor_work = NULL;
4902: c->keepnonzeropattern = a->keepnonzeropattern;
4904: c->rmax = a->rmax;
4905: c->nz = a->nz;
4906: c->maxnz = a->nz; /* Since we allocate exactly the right amount */
4908: c->compressedrow.use = a->compressedrow.use;
4909: c->compressedrow.nrows = a->compressedrow.nrows;
4910: if (a->compressedrow.use) {
4911: i = a->compressedrow.nrows;
4912: PetscCall(PetscMalloc2(i + 1, &c->compressedrow.i, i, &c->compressedrow.rindex));
4913: PetscCall(PetscArraycpy(c->compressedrow.i, a->compressedrow.i, i + 1));
4914: PetscCall(PetscArraycpy(c->compressedrow.rindex, a->compressedrow.rindex, i));
4915: } else {
4916: c->compressedrow.use = PETSC_FALSE;
4917: c->compressedrow.i = NULL;
4918: c->compressedrow.rindex = NULL;
4919: }
4920: c->nonzerorowcnt = a->nonzerorowcnt;
4921: C->nonzerostate = A->nonzerostate;
4923: PetscCall(MatDuplicate_SeqAIJ_Inode(A, cpvalues, &C));
4924: }
4925: PetscCall(PetscFunctionListDuplicate(((PetscObject)A)->qlist, &((PetscObject)C)->qlist));
4926: PetscFunctionReturn(PETSC_SUCCESS);
4927: }
4929: PetscErrorCode MatDuplicate_SeqAIJ(Mat A, MatDuplicateOption cpvalues, Mat *B)
4930: {
4931: PetscFunctionBegin;
4932: PetscCall(MatCreate(PetscObjectComm((PetscObject)A), B));
4933: PetscCall(MatSetSizes(*B, A->rmap->n, A->cmap->n, A->rmap->n, A->cmap->n));
4934: if (!(A->rmap->n % A->rmap->bs) && !(A->cmap->n % A->cmap->bs)) PetscCall(MatSetBlockSizesFromMats(*B, A, A));
4935: PetscCall(MatSetType(*B, ((PetscObject)A)->type_name));
4936: PetscCall(MatDuplicateNoCreate_SeqAIJ(*B, A, cpvalues, PETSC_TRUE));
4937: PetscFunctionReturn(PETSC_SUCCESS);
4938: }
4940: PetscErrorCode MatLoad_SeqAIJ(Mat newMat, PetscViewer viewer)
4941: {
4942: PetscBool isbinary, ishdf5;
4944: PetscFunctionBegin;
4947: /* force binary viewer to load .info file if it has not yet done so */
4948: PetscCall(PetscViewerSetUp(viewer));
4949: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERBINARY, &isbinary));
4950: PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERHDF5, &ishdf5));
4951: if (isbinary) {
4952: PetscCall(MatLoad_SeqAIJ_Binary(newMat, viewer));
4953: } else if (ishdf5) {
4954: #if PetscDefined(HAVE_HDF5)
4955: PetscCall(MatLoad_AIJ_HDF5(newMat, viewer));
4956: #else
4957: SETERRQ(PetscObjectComm((PetscObject)newMat), PETSC_ERR_SUP, "HDF5 not supported in this build.\nPlease reconfigure using --download-hdf5");
4958: #endif
4959: } else {
4960: 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);
4961: }
4962: PetscFunctionReturn(PETSC_SUCCESS);
4963: }
4965: PetscErrorCode MatLoad_SeqAIJ_Binary(Mat mat, PetscViewer viewer)
4966: {
4967: Mat_SeqAIJ *a = (Mat_SeqAIJ *)mat->data;
4968: PetscInt header[4], *rowlens, M, N, nz, sum, rows, cols, i;
4970: PetscFunctionBegin;
4971: PetscCall(PetscViewerSetUp(viewer));
4973: /* read in matrix header */
4974: PetscCall(PetscViewerBinaryRead(viewer, header, 4, NULL, PETSC_INT));
4975: PetscCheck(header[0] == MAT_FILE_CLASSID, PETSC_COMM_SELF, PETSC_ERR_FILE_UNEXPECTED, "Not a matrix object in file");
4976: M = header[1];
4977: N = header[2];
4978: nz = header[3];
4979: PetscCheck(M >= 0, PetscObjectComm((PetscObject)viewer), PETSC_ERR_FILE_UNEXPECTED, "Matrix row size (%" PetscInt_FMT ") in file is negative", M);
4980: PetscCheck(N >= 0, PetscObjectComm((PetscObject)viewer), PETSC_ERR_FILE_UNEXPECTED, "Matrix column size (%" PetscInt_FMT ") in file is negative", N);
4981: PetscCheck(nz >= 0, PETSC_COMM_SELF, PETSC_ERR_FILE_UNEXPECTED, "Matrix stored in special format on disk, cannot load as SeqAIJ");
4983: /* set block sizes from the viewer's .info file */
4984: PetscCall(MatLoad_Binary_BlockSizes(mat, viewer));
4985: /* set local and global sizes if not set already */
4986: if (mat->rmap->n < 0) mat->rmap->n = M;
4987: if (mat->cmap->n < 0) mat->cmap->n = N;
4988: if (mat->rmap->N < 0) mat->rmap->N = M;
4989: if (mat->cmap->N < 0) mat->cmap->N = N;
4990: PetscCall(PetscLayoutSetUp(mat->rmap));
4991: PetscCall(PetscLayoutSetUp(mat->cmap));
4993: /* check if the matrix sizes are correct */
4994: PetscCall(MatGetSize(mat, &rows, &cols));
4995: 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);
4997: /* read in row lengths */
4998: PetscCall(PetscMalloc1(M, &rowlens));
4999: PetscCall(PetscViewerBinaryRead(viewer, rowlens, M, NULL, PETSC_INT));
5000: /* check if sum(rowlens) is same as nz */
5001: sum = 0;
5002: for (i = 0; i < M; i++) sum += rowlens[i];
5003: 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);
5004: /* preallocate and check sizes */
5005: PetscCall(MatSeqAIJSetPreallocation_SeqAIJ(mat, 0, rowlens));
5006: PetscCall(MatGetSize(mat, &rows, &cols));
5007: 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);
5008: /* store row lengths */
5009: PetscCall(PetscArraycpy(a->ilen, rowlens, M));
5010: PetscCall(PetscFree(rowlens));
5012: /* fill in "i" row pointers */
5013: a->i[0] = 0;
5014: for (i = 0; i < M; i++) a->i[i + 1] = a->i[i] + a->ilen[i];
5015: /* read in "j" column indices */
5016: PetscCall(PetscViewerBinaryRead(viewer, a->j, nz, NULL, PETSC_INT));
5017: /* read in "a" nonzero values */
5018: PetscCall(PetscViewerBinaryRead(viewer, a->a, nz, NULL, PETSC_SCALAR));
5020: PetscCall(MatAssemblyBegin(mat, MAT_FINAL_ASSEMBLY));
5021: PetscCall(MatAssemblyEnd(mat, MAT_FINAL_ASSEMBLY));
5022: PetscFunctionReturn(PETSC_SUCCESS);
5023: }
5025: PetscErrorCode MatEqual_SeqAIJ(Mat A, Mat B, PetscBool *flg)
5026: {
5027: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data, *b = (Mat_SeqAIJ *)B->data;
5028: const PetscScalar *aa, *ba;
5030: PetscFunctionBegin;
5031: /* If the matrix dimensions are not equal,or no of nonzeros */
5032: if ((A->rmap->n != B->rmap->n) || (A->cmap->n != B->cmap->n) || (a->nz != b->nz)) {
5033: *flg = PETSC_FALSE;
5034: PetscFunctionReturn(PETSC_SUCCESS);
5035: }
5037: /* if the a->i are the same */
5038: PetscCall(PetscArraycmp(a->i, b->i, A->rmap->n + 1, flg));
5039: if (!*flg) PetscFunctionReturn(PETSC_SUCCESS);
5041: /* if a->j are the same */
5042: PetscCall(PetscArraycmp(a->j, b->j, a->nz, flg));
5043: if (!*flg) PetscFunctionReturn(PETSC_SUCCESS);
5045: PetscCall(MatSeqAIJGetArrayRead(A, &aa));
5046: PetscCall(MatSeqAIJGetArrayRead(B, &ba));
5047: /* if a->a are the same */
5048: PetscCall(PetscArraycmp(aa, ba, a->nz, flg));
5049: PetscCall(MatSeqAIJRestoreArrayRead(A, &aa));
5050: PetscCall(MatSeqAIJRestoreArrayRead(B, &ba));
5051: PetscFunctionReturn(PETSC_SUCCESS);
5052: }
5054: /*@
5055: MatCreateSeqAIJWithArrays - Creates an sequential `MATSEQAIJ` matrix using matrix elements (in CSR format)
5056: provided by the user.
5058: Collective
5060: Input Parameters:
5061: + comm - must be an MPI communicator of size 1
5062: . m - number of rows
5063: . n - number of columns
5064: . i - row indices; that is i[0] = 0, i[row] = i[row-1] + number of elements in that row of the matrix
5065: . j - column indices
5066: - a - matrix values
5068: Output Parameter:
5069: . mat - the matrix
5071: Level: intermediate
5073: Notes:
5074: The `i`, `j`, and `a` arrays are not copied by this routine, the user must free these arrays
5075: once the matrix is destroyed and not before
5077: You cannot set new nonzero locations into this matrix, that will generate an error.
5079: The `i` and `j` indices are 0 based
5081: The format which is used for the sparse matrix input, is equivalent to a
5082: row-major ordering.. i.e for the following matrix, the input data expected is
5083: as shown
5084: .vb
5085: 1 0 0
5086: 2 0 3
5087: 4 5 6
5089: i = {0,1,3,6} [size = nrow+1 = 3+1]
5090: j = {0,0,2,0,1,2} [size = 6]; values must be sorted for each row
5091: v = {1,2,3,4,5,6} [size = 6]
5092: .ve
5094: .seealso: [](ch_matrices), `Mat`, `MatCreate()`, `MatCreateAIJ()`, `MatCreateSeqAIJ()`, `MatCreateMPIAIJWithArrays()`, `MatMPIAIJSetPreallocationCSR()`
5095: @*/
5096: PetscErrorCode MatCreateSeqAIJWithArrays(MPI_Comm comm, PetscInt m, PetscInt n, PetscInt i[], PetscInt j[], PetscScalar a[], Mat *mat)
5097: {
5098: PetscInt ii;
5099: Mat_SeqAIJ *aij;
5101: PetscFunctionBegin;
5102: PetscCheck(m <= 0 || i[0] == 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "i (row indices) must start with 0");
5103: PetscCall(MatCreate(comm, mat));
5104: PetscCall(MatSetSizes(*mat, m, n, m, n));
5105: /* PetscCall(MatSetBlockSizes(*mat,,)); */
5106: PetscCall(MatSetType(*mat, MATSEQAIJ));
5107: PetscCall(MatSeqAIJSetPreallocation_SeqAIJ(*mat, MAT_SKIP_ALLOCATION, NULL));
5108: aij = (Mat_SeqAIJ *)(*mat)->data;
5109: PetscCall(PetscMalloc1(m, &aij->imax));
5110: PetscCall(PetscMalloc1(m, &aij->ilen));
5112: aij->i = i;
5113: aij->j = j;
5114: aij->a = a;
5115: aij->nonew = -1; /*this indicates that inserting a new value in the matrix that generates a new nonzero is an error*/
5116: aij->free_a = PETSC_FALSE;
5117: aij->free_ij = PETSC_FALSE;
5119: for (ii = 0, aij->nonzerorowcnt = 0, aij->rmax = 0; ii < m; ii++) {
5120: aij->ilen[ii] = aij->imax[ii] = i[ii + 1] - i[ii];
5121: if (PetscDefined(USE_DEBUG)) {
5122: 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]);
5123: for (PetscInt jj = i[ii] + 1; jj < i[ii + 1]; jj++) {
5124: 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);
5125: 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);
5126: }
5127: }
5128: }
5129: if (PetscDefined(USE_DEBUG)) {
5130: for (ii = 0; ii < aij->i[m]; ii++) {
5131: PetscCheck(j[ii] >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Negative column index at location = %" PetscInt_FMT " index = %" PetscInt_FMT, ii, j[ii]);
5132: 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);
5133: }
5134: }
5136: PetscCall(MatAssemblyBegin(*mat, MAT_FINAL_ASSEMBLY));
5137: PetscCall(MatAssemblyEnd(*mat, MAT_FINAL_ASSEMBLY));
5138: PetscFunctionReturn(PETSC_SUCCESS);
5139: }
5141: /*@
5142: MatCreateSeqAIJFromTriple - Creates an sequential `MATSEQAIJ` matrix using matrix elements (in COO format)
5143: provided by the user.
5145: Collective
5147: Input Parameters:
5148: + comm - must be an MPI communicator of size 1
5149: . m - number of rows
5150: . n - number of columns
5151: . i - row indices
5152: . j - column indices
5153: . a - matrix values
5154: . nz - number of nonzeros
5155: - idx - if the `i` and `j` indices start with 1 use `PETSC_TRUE` otherwise use `PETSC_FALSE`
5157: Output Parameter:
5158: . mat - the matrix
5160: Level: intermediate
5162: Example:
5163: For the following matrix, the input data expected is as shown (using 0 based indexing)
5164: .vb
5165: 1 0 0
5166: 2 0 3
5167: 4 5 6
5169: i = {0,1,1,2,2,2}
5170: j = {0,0,2,0,1,2}
5171: v = {1,2,3,4,5,6}
5172: .ve
5174: Note:
5175: Instead of using this function, users should also consider `MatSetPreallocationCOO()` and `MatSetValuesCOO()`, which allow repeated or remote entries,
5176: and are particularly useful in iterative applications.
5178: .seealso: [](ch_matrices), `Mat`, `MatCreate()`, `MatCreateAIJ()`, `MatCreateSeqAIJ()`, `MatCreateSeqAIJWithArrays()`, `MatMPIAIJSetPreallocationCSR()`, `MatSetValuesCOO()`, `MatSetPreallocationCOO()`
5179: @*/
5180: PetscErrorCode MatCreateSeqAIJFromTriple(MPI_Comm comm, PetscInt m, PetscInt n, PetscInt i[], PetscInt j[], PetscScalar a[], Mat *mat, PetscCount nz, PetscBool idx)
5181: {
5182: PetscInt ii, *nnz, one = 1, row, col;
5184: PetscFunctionBegin;
5185: PetscCall(PetscCalloc1(m, &nnz));
5186: for (ii = 0; ii < nz; ii++) nnz[i[ii] - !!idx] += 1;
5187: PetscCall(MatCreate(comm, mat));
5188: PetscCall(MatSetSizes(*mat, m, n, m, n));
5189: PetscCall(MatSetType(*mat, MATSEQAIJ));
5190: PetscCall(MatSeqAIJSetPreallocation_SeqAIJ(*mat, 0, nnz));
5191: for (ii = 0; ii < nz; ii++) {
5192: if (idx) {
5193: row = i[ii] - 1;
5194: col = j[ii] - 1;
5195: } else {
5196: row = i[ii];
5197: col = j[ii];
5198: }
5199: PetscCall(MatSetValues(*mat, one, &row, one, &col, &a[ii], ADD_VALUES));
5200: }
5201: PetscCall(MatAssemblyBegin(*mat, MAT_FINAL_ASSEMBLY));
5202: PetscCall(MatAssemblyEnd(*mat, MAT_FINAL_ASSEMBLY));
5203: PetscCall(PetscFree(nnz));
5204: PetscFunctionReturn(PETSC_SUCCESS);
5205: }
5207: PetscErrorCode MatCreateMPIMatConcatenateSeqMat_SeqAIJ(MPI_Comm comm, Mat inmat, PetscInt n, MatReuse scall, Mat *outmat)
5208: {
5209: PetscFunctionBegin;
5210: PetscCall(MatCreateMPIMatConcatenateSeqMat_MPIAIJ(comm, inmat, n, scall, outmat));
5211: PetscFunctionReturn(PETSC_SUCCESS);
5212: }
5214: /*
5215: Permute A into C's *local* index space using rowemb,colemb.
5216: The embedding are supposed to be injections and the above implies that the range of rowemb is a subset
5217: of [0,m), colemb is in [0,n).
5218: If pattern == DIFFERENT_NONZERO_PATTERN, C is preallocated according to A.
5219: */
5220: PetscErrorCode MatSetSeqMat_SeqAIJ(Mat C, IS rowemb, IS colemb, MatStructure pattern, Mat B)
5221: {
5222: /* If making this function public, change the error returned in this function away from _PLIB. */
5223: Mat_SeqAIJ *Baij;
5224: PetscBool seqaij;
5225: PetscInt m, n, *nz, i, j, count;
5226: PetscScalar v;
5227: const PetscInt *rowindices, *colindices;
5229: PetscFunctionBegin;
5230: if (!B) PetscFunctionReturn(PETSC_SUCCESS);
5231: /* Check to make sure the target matrix (and embeddings) are compatible with C and each other. */
5232: PetscCall(PetscObjectBaseTypeCompare((PetscObject)B, MATSEQAIJ, &seqaij));
5233: PetscCheck(seqaij, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Input matrix is of wrong type");
5234: if (rowemb) {
5235: PetscCall(ISGetLocalSize(rowemb, &m));
5236: 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);
5237: } else PetscCheck(C->rmap->n == B->rmap->n, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Input matrix is row-incompatible with the target matrix");
5238: if (colemb) {
5239: PetscCall(ISGetLocalSize(colemb, &n));
5240: 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);
5241: } else PetscCheck(C->cmap->n == B->cmap->n, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Input matrix is col-incompatible with the target matrix");
5243: Baij = (Mat_SeqAIJ *)B->data;
5244: rowindices = NULL;
5245: if (rowemb) PetscCall(ISGetIndices(rowemb, &rowindices));
5246: if (pattern == DIFFERENT_NONZERO_PATTERN) {
5247: PetscCall(PetscMalloc1(C->rmap->n, &nz));
5248: if (rowemb) {
5249: PetscCall(PetscArrayzero(nz, C->rmap->n));
5250: for (i = 0; i < B->rmap->n; i++) nz[rowindices[i]] = Baij->i[i + 1] - Baij->i[i];
5251: } else {
5252: for (i = 0; i < B->rmap->n; i++) nz[i] = Baij->i[i + 1] - Baij->i[i];
5253: }
5254: PetscCall(MatSeqAIJSetPreallocation(C, 0, nz));
5255: PetscCall(PetscFree(nz));
5256: }
5257: if (pattern == SUBSET_NONZERO_PATTERN) PetscCall(MatZeroEntries(C));
5258: count = 0;
5259: colindices = NULL;
5260: if (colemb) PetscCall(ISGetIndices(colemb, &colindices));
5261: for (i = 0; i < B->rmap->n; i++) {
5262: PetscInt row;
5263: row = i;
5264: if (rowindices) row = rowindices[i];
5265: for (j = Baij->i[i]; j < Baij->i[i + 1]; j++) {
5266: PetscInt col;
5267: col = Baij->j[count];
5268: if (colindices) col = colindices[col];
5269: v = Baij->a[count];
5270: PetscCall(MatSetValues(C, 1, &row, 1, &col, &v, INSERT_VALUES));
5271: ++count;
5272: }
5273: }
5274: if (colemb) PetscCall(ISRestoreIndices(colemb, &colindices));
5275: if (rowemb) PetscCall(ISRestoreIndices(rowemb, &rowindices));
5276: /* FIXME: set C's nonzerostate correctly. */
5277: /* Assembly for C is necessary. */
5278: C->preallocated = PETSC_TRUE;
5279: C->assembled = PETSC_TRUE;
5280: C->was_assembled = PETSC_FALSE;
5281: PetscFunctionReturn(PETSC_SUCCESS);
5282: }
5284: PetscErrorCode MatEliminateZeros_SeqAIJ(Mat A, PetscBool keep)
5285: {
5286: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
5287: MatScalar *aa = a->a;
5288: PetscInt m = A->rmap->n, fshift = 0, fshift_prev = 0, i, k;
5289: PetscInt *ailen = a->ilen, *imax = a->imax, *ai = a->i, *aj = a->j, rmax = 0;
5291: PetscFunctionBegin;
5292: PetscCheck(A->assembled, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Cannot eliminate zeros for unassembled matrix");
5293: if (m) rmax = ailen[0]; /* determine row with most nonzeros */
5294: for (i = 1, a->nonzerorowcnt = 0; i <= m; i++) {
5295: /* move each nonzero entry back by the amount of zero slots (fshift) before it*/
5296: for (k = ai[i - 1]; k < ai[i]; k++) {
5297: if (aa[k] == 0 && (aj[k] != i - 1 || !keep)) fshift++;
5298: else {
5299: if (aa[k] == 0 && aj[k] == i - 1) PetscCall(PetscInfo(A, "Keep the diagonal zero at row %" PetscInt_FMT "\n", i - 1));
5300: aa[k - fshift] = aa[k];
5301: aj[k - fshift] = aj[k];
5302: }
5303: }
5304: ai[i - 1] -= fshift_prev; // safe to update ai[i-1] now since it will not be used in the next iteration
5305: fshift_prev = fshift;
5306: /* reset ilen and imax for each row */
5307: ailen[i - 1] = imax[i - 1] = ai[i] - fshift - ai[i - 1];
5308: a->nonzerorowcnt += ((ai[i] - fshift - ai[i - 1]) > 0);
5309: rmax = PetscMax(rmax, ailen[i - 1]);
5310: }
5311: if (fshift) {
5312: if (m) {
5313: ai[m] -= fshift;
5314: a->nz = ai[m];
5315: }
5316: 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));
5317: A->nonzerostate++;
5318: A->info.nz_unneeded += (PetscReal)fshift;
5319: a->rmax = rmax;
5320: if (a->inode.use && a->inode.checked) PetscCall(MatSeqAIJCheckInode(A));
5321: PetscCall(MatAssemblyBegin(A, MAT_FINAL_ASSEMBLY));
5322: PetscCall(MatAssemblyEnd(A, MAT_FINAL_ASSEMBLY));
5323: }
5324: PetscFunctionReturn(PETSC_SUCCESS);
5325: }
5327: PetscFunctionList MatSeqAIJList = NULL;
5329: /*@
5330: MatSeqAIJSetType - Converts a `MATSEQAIJ` matrix to a subtype
5332: Collective
5334: Input Parameters:
5335: + mat - the matrix object
5336: - matype - matrix type
5338: Options Database Key:
5339: . -mat_seqaij_type method - for example seqaijcrl
5341: Level: intermediate
5343: .seealso: [](ch_matrices), `Mat`, `PCSetType()`, `VecSetType()`, `MatCreate()`, `MatType`
5344: @*/
5345: PetscErrorCode MatSeqAIJSetType(Mat mat, MatType matype)
5346: {
5347: PetscBool sametype;
5348: PetscErrorCode (*r)(Mat, MatType, MatReuse, Mat *);
5350: PetscFunctionBegin;
5352: PetscCall(PetscObjectTypeCompare((PetscObject)mat, matype, &sametype));
5353: if (sametype) PetscFunctionReturn(PETSC_SUCCESS);
5355: PetscCall(PetscFunctionListFind(MatSeqAIJList, matype, &r));
5356: PetscCheck(r, PetscObjectComm((PetscObject)mat), PETSC_ERR_ARG_UNKNOWN_TYPE, "Unknown Mat type given: %s", matype);
5357: PetscCall((*r)(mat, matype, MAT_INPLACE_MATRIX, &mat));
5358: PetscFunctionReturn(PETSC_SUCCESS);
5359: }
5361: /*@C
5362: MatSeqAIJRegister - - Adds a new sub-matrix type for sequential `MATSEQAIJ` matrices
5364: Not Collective, No Fortran Support
5366: Input Parameters:
5367: + sname - name of a new user-defined matrix type, for example `MATSEQAIJCRL`
5368: - function - routine to convert to subtype
5370: Level: advanced
5372: Notes:
5373: `MatSeqAIJRegister()` may be called multiple times to add several user-defined solvers.
5375: Then, your matrix can be chosen with the procedural interface at runtime via the option
5376: .vb
5377: -mat_seqaij_type my_mat
5378: .ve
5380: .seealso: [](ch_matrices), `Mat`, `MatSeqAIJRegisterAll()`
5381: @*/
5382: PetscErrorCode MatSeqAIJRegister(const char sname[], PetscErrorCode (*function)(Mat, MatType, MatReuse, Mat *))
5383: {
5384: PetscFunctionBegin;
5385: PetscCall(MatInitializePackage());
5386: PetscCall(PetscFunctionListAdd(&MatSeqAIJList, sname, function));
5387: PetscFunctionReturn(PETSC_SUCCESS);
5388: }
5390: PetscBool MatSeqAIJRegisterAllCalled = PETSC_FALSE;
5392: /*@C
5393: MatSeqAIJRegisterAll - Registers all of the matrix subtypes of `MATSSEQAIJ`
5395: Not Collective
5397: Level: advanced
5399: Note:
5400: This registers the versions of `MATSEQAIJ` for GPUs
5402: .seealso: [](ch_matrices), `Mat`, `MatRegisterAll()`, `MatSeqAIJRegister()`
5403: @*/
5404: PetscErrorCode MatSeqAIJRegisterAll(void)
5405: {
5406: PetscFunctionBegin;
5407: if (MatSeqAIJRegisterAllCalled) PetscFunctionReturn(PETSC_SUCCESS);
5408: MatSeqAIJRegisterAllCalled = PETSC_TRUE;
5410: PetscCall(MatSeqAIJRegister(MATSEQAIJCRL, MatConvert_SeqAIJ_SeqAIJCRL));
5411: PetscCall(MatSeqAIJRegister(MATSEQAIJPERM, MatConvert_SeqAIJ_SeqAIJPERM));
5412: PetscCall(MatSeqAIJRegister(MATSEQAIJSELL, MatConvert_SeqAIJ_SeqAIJSELL));
5413: #if PetscDefined(HAVE_MKL_SPARSE)
5414: PetscCall(MatSeqAIJRegister(MATSEQAIJMKL, MatConvert_SeqAIJ_SeqAIJMKL));
5415: #endif
5416: #if PetscDefined(HAVE_CUDA)
5417: PetscCall(MatSeqAIJRegister(MATSEQAIJCUSPARSE, MatConvert_SeqAIJ_SeqAIJCUSPARSE));
5418: #endif
5419: #if PetscDefined(HAVE_HIP)
5420: PetscCall(MatSeqAIJRegister(MATSEQAIJHIPSPARSE, MatConvert_SeqAIJ_SeqAIJHIPSPARSE));
5421: #endif
5422: #if PetscDefined(HAVE_KOKKOS_KERNELS)
5423: PetscCall(MatSeqAIJRegister(MATSEQAIJKOKKOS, MatConvert_SeqAIJ_SeqAIJKokkos));
5424: #endif
5425: #if PetscDefined(HAVE_VIENNACL) && PetscDefined(HAVE_VIENNACL_NO_CUDA)
5426: PetscCall(MatSeqAIJRegister(MATMPIAIJVIENNACL, MatConvert_SeqAIJ_SeqAIJViennaCL));
5427: #endif
5428: PetscFunctionReturn(PETSC_SUCCESS);
5429: }
5431: /*
5432: Special version for direct calls from Fortran
5433: */
5434: #if PetscDefined(HAVE_FORTRAN_CAPS)
5435: #define matsetvaluesseqaij_ MATSETVALUESSEQAIJ
5436: #elif !PetscDefined(HAVE_FORTRAN_UNDERSCORE)
5437: #define matsetvaluesseqaij_ matsetvaluesseqaij
5438: #endif
5440: /* Change these macros so can be used in void function */
5442: /* Change these macros so can be used in void function */
5443: /* Identical to PetscCallVoid, except it assigns to *_ierr */
5444: #undef PetscCall
5445: #define PetscCall(...) \
5446: do { \
5447: PetscErrorCode ierr_msv_mpiaij = __VA_ARGS__; \
5448: if (PetscUnlikely(ierr_msv_mpiaij)) { \
5449: *_ierr = PetscError(PETSC_COMM_SELF, __LINE__, PETSC_FUNCTION_NAME, __FILE__, ierr_msv_mpiaij, PETSC_ERROR_REPEAT, " "); \
5450: return; \
5451: } \
5452: } while (0)
5454: #undef SETERRQ
5455: #define SETERRQ(comm, ierr, ...) \
5456: do { \
5457: *_ierr = PetscError(comm, __LINE__, PETSC_FUNCTION_NAME, __FILE__, ierr, PETSC_ERROR_INITIAL, __VA_ARGS__); \
5458: return; \
5459: } while (0)
5461: PETSC_EXTERN void matsetvaluesseqaij_(Mat *AA, PetscInt *mm, const PetscInt im[], PetscInt *nn, const PetscInt in[], const PetscScalar v[], InsertMode *isis, PetscErrorCode *_ierr)
5462: {
5463: Mat A = *AA;
5464: PetscInt m = *mm, n = *nn;
5465: InsertMode is = *isis;
5466: Mat_SeqAIJ *a = (Mat_SeqAIJ *)A->data;
5467: PetscInt *rp, k, low, high, t, ii, row, nrow, i, col, l, rmax, N;
5468: PetscInt *imax, *ai, *ailen;
5469: PetscInt *aj, nonew = a->nonew, lastcol = -1;
5470: MatScalar *ap, value, *aa;
5471: PetscBool ignorezeroentries = a->ignorezeroentries;
5472: PetscBool roworiented = a->roworiented;
5474: PetscFunctionBegin;
5475: MatCheckPreallocated(A, 1);
5476: imax = a->imax;
5477: ai = a->i;
5478: ailen = a->ilen;
5479: aj = a->j;
5480: aa = a->a;
5482: for (k = 0; k < m; k++) { /* loop over added rows */
5483: row = im[k];
5484: if (row < 0) continue;
5485: PetscCheck(row < A->rmap->n, PetscObjectComm((PetscObject)A), PETSC_ERR_ARG_OUTOFRANGE, "Row too large");
5486: rp = aj + ai[row];
5487: ap = aa + ai[row];
5488: rmax = imax[row];
5489: nrow = ailen[row];
5490: low = 0;
5491: high = nrow;
5492: for (l = 0; l < n; l++) { /* loop over added columns */
5493: if (in[l] < 0) continue;
5494: PetscCheck(in[l] < A->cmap->n, PetscObjectComm((PetscObject)A), PETSC_ERR_ARG_OUTOFRANGE, "Column too large");
5495: col = in[l];
5496: if (roworiented) value = v[l + k * n];
5497: else value = v[k + l * m];
5499: if (value == 0.0 && ignorezeroentries && (is == ADD_VALUES)) continue;
5501: if (col <= lastcol) low = 0;
5502: else high = nrow;
5503: lastcol = col;
5504: while (high - low > 5) {
5505: t = (low + high) / 2;
5506: if (rp[t] > col) high = t;
5507: else low = t;
5508: }
5509: for (i = low; i < high; i++) {
5510: if (rp[i] > col) break;
5511: if (rp[i] == col) {
5512: if (is == ADD_VALUES) ap[i] += value;
5513: else ap[i] = value;
5514: goto noinsert;
5515: }
5516: }
5517: if (value == 0.0 && ignorezeroentries) goto noinsert;
5518: if (nonew == 1) goto noinsert;
5519: PetscCheck(nonew != -1, PetscObjectComm((PetscObject)A), PETSC_ERR_ARG_OUTOFRANGE, "Inserting a new nonzero in the matrix");
5520: MatSeqXAIJReallocateAIJ(A, A->rmap->n, 1, nrow, row, col, rmax, aa, ai, aj, rp, ap, imax, nonew, MatScalar);
5521: N = nrow++ - 1;
5522: a->nz++;
5523: high++;
5524: /* shift up all the later entries in this row */
5525: for (ii = N; ii >= i; ii--) {
5526: rp[ii + 1] = rp[ii];
5527: ap[ii + 1] = ap[ii];
5528: }
5529: rp[i] = col;
5530: ap[i] = value;
5531: noinsert:;
5532: low = i + 1;
5533: }
5534: ailen[row] = nrow;
5535: }
5536: PetscFunctionReturnVoid();
5537: }
5538: /* Undefining these here since they were redefined from their original definition above! No
5539: * other PETSc functions should be defined past this point, as it is impossible to recover the
5540: * original definitions */
5541: #undef PetscCall
5542: #undef SETERRQ